Resguardo inicial

linux-7.2.3/net/rds/Kconfig

@@ -0,0 +1,37 @@
+# SPDX-License-Identifier: GPL-2.0-only
+
+config RDS
+	tristate "The Reliable Datagram Sockets Protocol"
+	depends on INET
+	help
+	  The RDS (Reliable Datagram Sockets) protocol provides reliable,
+	  sequenced delivery of datagrams over Infiniband or TCP.
+
+config RDS_RDMA
+	tristate "RDS over Infiniband"
+	depends on RDS && INFINIBAND && INFINIBAND_ADDR_TRANS
+	help
+	  Allow RDS to use Infiniband as a transport.
+	  This transport supports RDMA operations.
+
+config RDS_TCP
+	tristate "RDS over TCP"
+	depends on RDS
+	depends on IPV6 || !IPV6
+	help
+	  Allow RDS to use TCP as a transport.
+	  This transport does not support RDMA operations.
+
+config RDS_DEBUG
+	bool "RDS debugging messages"
+	depends on RDS
+	default n
+
+config GCOV_PROFILE_RDS
+	bool "Enable GCOV profiling on RDS"
+	depends on GCOV_KERNEL
+	help
+	  Enable GCOV profiling on RDS for checking which functions/lines
+	  are executed.
+
+	  If unsure, say N.

linux-7.2.3/net/rds/Makefile

@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: GPL-2.0
+obj-$(CONFIG_RDS) += rds.o
+rds-y :=	af_rds.o bind.o cong.o connection.o info.o message.o   \
+			recv.o send.o stats.o sysctl.o threads.o transport.o \
+			loop.o page.o rdma.o
+
+obj-$(CONFIG_RDS_RDMA) += rds_rdma.o
+rds_rdma-y :=	rdma_transport.o \
+			ib.o ib_cm.o ib_recv.o ib_ring.o ib_send.o ib_stats.o \
+			ib_sysctl.o ib_rdma.o ib_frmr.o
+
+
+obj-$(CONFIG_RDS_TCP) += rds_tcp.o
+rds_tcp-y :=		tcp.o tcp_connect.o tcp_listen.o tcp_recv.o \
+			tcp_send.o tcp_stats.o
+
+ccflags-$(CONFIG_RDS_DEBUG)	:=	-DRDS_DEBUG
+
+# for GCOV coverage profiling
+ifdef CONFIG_GCOV_PROFILE_RDS
+GCOV_PROFILE := y
+endif

linux-7.2.3/net/rds/bind.c

@@ -0,0 +1,283 @@
+/*
+ * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <net/sock.h>
+#include <linux/in.h>
+#include <linux/ipv6.h>
+#include <linux/if_arp.h>
+#include <linux/jhash.h>
+#include <linux/ratelimit.h>
+#include "rds.h"
+
+static struct rhashtable bind_hash_table;
+
+static const struct rhashtable_params ht_parms = {
+	.nelem_hint = 768,
+	.key_len = RDS_BOUND_KEY_LEN,
+	.key_offset = offsetof(struct rds_sock, rs_bound_key),
+	.head_offset = offsetof(struct rds_sock, rs_bound_node),
+	.max_size = 16384,
+	.min_size = 1024,
+};
+
+/* Create a key for the bind hash table manipulation.  Port is in network byte
+ * order.
+ */
+static inline void __rds_create_bind_key(u8 *key, const struct in6_addr *addr,
+					 __be16 port, __u32 scope_id)
+{
+	memcpy(key, addr, sizeof(*addr));
+	key += sizeof(*addr);
+	memcpy(key, &port, sizeof(port));
+	key += sizeof(port);
+	memcpy(key, &scope_id, sizeof(scope_id));
+}
+
+/*
+ * Return the rds_sock bound at the given local address.
+ *
+ * The rx path can race with rds_release.  We notice if rds_release() has
+ * marked this socket and don't return a rs ref to the rx path.
+ */
+struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port,
+				__u32 scope_id)
+{
+	u8 key[RDS_BOUND_KEY_LEN];
+	struct rds_sock *rs;
+
+	__rds_create_bind_key(key, addr, port, scope_id);
+	rcu_read_lock();
+	rs = rhashtable_lookup(&bind_hash_table, key, ht_parms);
+	if (rs && (sock_flag(rds_rs_to_sk(rs), SOCK_DEAD) ||
+		   !refcount_inc_not_zero(&rds_rs_to_sk(rs)->sk_refcnt)))
+		rs = NULL;
+
+	rcu_read_unlock();
+
+	rdsdebug("returning rs %p for %pI6c:%u\n", rs, addr,
+		 ntohs(port));
+
+	return rs;
+}
+
+/* returns -ve errno or +ve port */
+static int rds_add_bound(struct rds_sock *rs, const struct in6_addr *addr,
+			 __be16 *port, __u32 scope_id)
+{
+	int ret = -EADDRINUSE;
+	u16 rover, last;
+	u8 key[RDS_BOUND_KEY_LEN];
+
+	if (*port != 0) {
+		rover = be16_to_cpu(*port);
+		if (rover == RDS_FLAG_PROBE_PORT)
+			return -EINVAL;
+		last = rover;
+	} else {
+		rover = max_t(u16, get_random_u16(), 2);
+		last = rover - 1;
+	}
+
+	do {
+		if (rover == 0)
+			rover++;
+
+		if (rover == RDS_FLAG_PROBE_PORT)
+			continue;
+		__rds_create_bind_key(key, addr, cpu_to_be16(rover),
+				      scope_id);
+		if (rhashtable_lookup_fast(&bind_hash_table, key, ht_parms))
+			continue;
+
+		memcpy(rs->rs_bound_key, key, sizeof(rs->rs_bound_key));
+		rs->rs_bound_addr = *addr;
+		net_get_random_once(&rs->rs_hash_initval,
+				    sizeof(rs->rs_hash_initval));
+		rs->rs_bound_port = cpu_to_be16(rover);
+		rs->rs_bound_node.next = NULL;
+		rds_sock_addref(rs);
+		if (!rhashtable_insert_fast(&bind_hash_table,
+					    &rs->rs_bound_node, ht_parms)) {
+			*port = rs->rs_bound_port;
+			rs->rs_bound_scope_id = scope_id;
+			ret = 0;
+			rdsdebug("rs %p binding to %pI6c:%d\n",
+				 rs, addr, (int)ntohs(*port));
+			break;
+		} else {
+			rs->rs_bound_addr = in6addr_any;
+			rds_sock_put(rs);
+			ret = -ENOMEM;
+			break;
+		}
+	} while (rover++ != last);
+
+	return ret;
+}
+
+void rds_remove_bound(struct rds_sock *rs)
+{
+
+	if (ipv6_addr_any(&rs->rs_bound_addr))
+		return;
+
+	rdsdebug("rs %p unbinding from %pI6c:%d\n",
+		 rs, &rs->rs_bound_addr,
+		 ntohs(rs->rs_bound_port));
+
+	rhashtable_remove_fast(&bind_hash_table, &rs->rs_bound_node, ht_parms);
+	rds_sock_put(rs);
+	rs->rs_bound_addr = in6addr_any;
+}
+
+int rds_bind(struct socket *sock, struct sockaddr_unsized *uaddr, int addr_len)
+{
+	struct sock *sk = sock->sk;
+	struct rds_sock *rs = rds_sk_to_rs(sk);
+	struct in6_addr v6addr, *binding_addr;
+	struct rds_transport *trans;
+	__u32 scope_id = 0;
+	int ret = 0;
+	__be16 port;
+
+	/* We allow an RDS socket to be bound to either IPv4 or IPv6
+	 * address.
+	 */
+	if (addr_len < offsetofend(struct sockaddr, sa_family))
+		return -EINVAL;
+	if (uaddr->sa_family == AF_INET) {
+		struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
+
+		if (addr_len < sizeof(struct sockaddr_in) ||
+		    sin->sin_addr.s_addr == htonl(INADDR_ANY) ||
+		    sin->sin_addr.s_addr == htonl(INADDR_BROADCAST) ||
+		    ipv4_is_multicast(sin->sin_addr.s_addr))
+			return -EINVAL;
+		ipv6_addr_set_v4mapped(sin->sin_addr.s_addr, &v6addr);
+		binding_addr = &v6addr;
+		port = sin->sin_port;
+#if IS_ENABLED(CONFIG_IPV6)
+	} else if (uaddr->sa_family == AF_INET6) {
+		struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)uaddr;
+		int addr_type;
+
+		if (addr_len < sizeof(struct sockaddr_in6))
+			return -EINVAL;
+		addr_type = ipv6_addr_type(&sin6->sin6_addr);
+		if (!(addr_type & IPV6_ADDR_UNICAST)) {
+			__be32 addr4;
+
+			if (!(addr_type & IPV6_ADDR_MAPPED))
+				return -EINVAL;
+
+			/* It is a mapped address.  Need to do some sanity
+			 * checks.
+			 */
+			addr4 = sin6->sin6_addr.s6_addr32[3];
+			if (addr4 == htonl(INADDR_ANY) ||
+			    addr4 == htonl(INADDR_BROADCAST) ||
+			    ipv4_is_multicast(addr4))
+				return -EINVAL;
+		}
+		/* The scope ID must be specified for link local address. */
+		if (addr_type & IPV6_ADDR_LINKLOCAL) {
+			if (sin6->sin6_scope_id == 0)
+				return -EINVAL;
+			scope_id = sin6->sin6_scope_id;
+		}
+		binding_addr = &sin6->sin6_addr;
+		port = sin6->sin6_port;
+#endif
+	} else {
+		return -EINVAL;
+	}
+	lock_sock(sk);
+
+	/* RDS socket does not allow re-binding. */
+	if (!ipv6_addr_any(&rs->rs_bound_addr)) {
+		ret = -EINVAL;
+		goto out;
+	}
+	/* Socket is connected.  The binding address should have the same
+	 * scope ID as the connected address, except the case when one is
+	 * non-link local address (scope_id is 0).
+	 */
+	if (!ipv6_addr_any(&rs->rs_conn_addr) && scope_id &&
+	    rs->rs_bound_scope_id &&
+	    scope_id != rs->rs_bound_scope_id) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	/* The transport can be set using SO_RDS_TRANSPORT option before the
+	 * socket is bound.
+	 */
+	if (rs->rs_transport) {
+		trans = rs->rs_transport;
+		if (!trans->laddr_check ||
+		    trans->laddr_check(sock_net(sock->sk),
+				       binding_addr, scope_id) != 0) {
+			ret = -ENOPROTOOPT;
+			goto out;
+		}
+	} else {
+		trans = rds_trans_get_preferred(sock_net(sock->sk),
+						binding_addr, scope_id);
+		if (!trans) {
+			ret = -EADDRNOTAVAIL;
+			pr_info_ratelimited("RDS: %s could not find a transport for %pI6c, load rds_tcp or rds_rdma?\n",
+					    __func__, binding_addr);
+			goto out;
+		}
+		rs->rs_transport = trans;
+	}
+
+	sock_set_flag(sk, SOCK_RCU_FREE);
+	ret = rds_add_bound(rs, binding_addr, &port, scope_id);
+	if (ret)
+		rs->rs_transport = NULL;
+
+out:
+	release_sock(sk);
+	return ret;
+}
+
+void rds_bind_lock_destroy(void)
+{
+	rhashtable_destroy(&bind_hash_table);
+}
+
+int rds_bind_lock_init(void)
+{
+	return rhashtable_init(&bind_hash_table, &ht_parms);
+}

linux-7.2.3/net/rds/cong.c

@@ -0,0 +1,428 @@
+/*
+ * Copyright (c) 2007, 2017 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/slab.h>
+#include <linux/types.h>
+#include <linux/rbtree.h>
+#include <linux/bitops.h>
+#include <linux/export.h>
+
+#include "rds.h"
+
+/*
+ * This file implements the receive side of the unconventional congestion
+ * management in RDS.
+ *
+ * Messages waiting in the receive queue on the receiving socket are accounted
+ * against the sockets SO_RCVBUF option value.  Only the payload bytes in the
+ * message are accounted for.  If the number of bytes queued equals or exceeds
+ * rcvbuf then the socket is congested.  All sends attempted to this socket's
+ * address should return block or return -EWOULDBLOCK.
+ *
+ * Applications are expected to be reasonably tuned such that this situation
+ * very rarely occurs.  An application encountering this "back-pressure" is
+ * considered a bug.
+ *
+ * This is implemented by having each node maintain bitmaps which indicate
+ * which ports on bound addresses are congested.  As the bitmap changes it is
+ * sent through all the connections which terminate in the local address of the
+ * bitmap which changed.
+ *
+ * The bitmaps are allocated as connections are brought up.  This avoids
+ * allocation in the interrupt handling path which queues messages on sockets.
+ * The dense bitmaps let transports send the entire bitmap on any bitmap change
+ * reasonably efficiently.  This is much easier to implement than some
+ * finer-grained communication of per-port congestion.  The sender does a very
+ * inexpensive bit test to test if the port it's about to send to is congested
+ * or not.
+ */
+
+/*
+ * Interaction with poll is a tad tricky. We want all processes stuck in
+ * poll to wake up and check whether a congested destination became uncongested.
+ * The really sad thing is we have no idea which destinations the application
+ * wants to send to - we don't even know which rds_connections are involved.
+ * So until we implement a more flexible rds poll interface, we have to make
+ * do with this:
+ * We maintain a global counter that is incremented each time a congestion map
+ * update is received. Each rds socket tracks this value, and if rds_poll
+ * finds that the saved generation number is smaller than the global generation
+ * number, it wakes up the process.
+ */
+static atomic_t		rds_cong_generation = ATOMIC_INIT(0);
+
+/*
+ * Congestion monitoring
+ */
+static LIST_HEAD(rds_cong_monitor);
+static DEFINE_RWLOCK(rds_cong_monitor_lock);
+
+/*
+ * Yes, a global lock.  It's used so infrequently that it's worth keeping it
+ * global to simplify the locking.  It's only used in the following
+ * circumstances:
+ *
+ *  - on connection buildup to associate a conn with its maps
+ *  - on map changes to inform conns of a new map to send
+ *
+ *  It's sadly ordered under the socket callback lock and the connection lock.
+ *  Receive paths can mark ports congested from interrupt context so the
+ *  lock masks interrupts.
+ */
+static DEFINE_SPINLOCK(rds_cong_lock);
+static struct rb_root rds_cong_tree = RB_ROOT;
+
+static struct rds_cong_map *rds_cong_tree_walk(const struct in6_addr *addr,
+					       struct rds_cong_map *insert)
+{
+	struct rb_node **p = &rds_cong_tree.rb_node;
+	struct rb_node *parent = NULL;
+	struct rds_cong_map *map;
+
+	while (*p) {
+		int diff;
+
+		parent = *p;
+		map = rb_entry(parent, struct rds_cong_map, m_rb_node);
+
+		diff = rds_addr_cmp(addr, &map->m_addr);
+		if (diff < 0)
+			p = &(*p)->rb_left;
+		else if (diff > 0)
+			p = &(*p)->rb_right;
+		else
+			return map;
+	}
+
+	if (insert) {
+		rb_link_node(&insert->m_rb_node, parent, p);
+		rb_insert_color(&insert->m_rb_node, &rds_cong_tree);
+	}
+	return NULL;
+}
+
+/*
+ * There is only ever one bitmap for any address.  Connections try and allocate
+ * these bitmaps in the process getting pointers to them.  The bitmaps are only
+ * ever freed as the module is removed after all connections have been freed.
+ */
+static struct rds_cong_map *rds_cong_from_addr(const struct in6_addr *addr)
+{
+	struct rds_cong_map *map;
+	struct rds_cong_map *ret = NULL;
+	unsigned long zp;
+	unsigned long i;
+	unsigned long flags;
+
+	map = kzalloc_obj(struct rds_cong_map);
+	if (!map)
+		return NULL;
+
+	map->m_addr = *addr;
+	init_waitqueue_head(&map->m_waitq);
+	INIT_LIST_HEAD(&map->m_conn_list);
+
+	for (i = 0; i < RDS_CONG_MAP_PAGES; i++) {
+		zp = get_zeroed_page(GFP_KERNEL);
+		if (zp == 0)
+			goto out;
+		map->m_page_addrs[i] = zp;
+	}
+
+	spin_lock_irqsave(&rds_cong_lock, flags);
+	ret = rds_cong_tree_walk(addr, map);
+	spin_unlock_irqrestore(&rds_cong_lock, flags);
+
+	if (!ret) {
+		ret = map;
+		map = NULL;
+	}
+
+out:
+	if (map) {
+		for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
+			free_page(map->m_page_addrs[i]);
+		kfree(map);
+	}
+
+	rdsdebug("map %p for addr %pI6c\n", ret, addr);
+
+	return ret;
+}
+
+/*
+ * Put the conn on its local map's list.  This is called when the conn is
+ * really added to the hash.  It's nested under the rds_conn_lock, sadly.
+ */
+void rds_cong_add_conn(struct rds_connection *conn)
+{
+	unsigned long flags;
+
+	rdsdebug("conn %p now on map %p\n", conn, conn->c_lcong);
+	spin_lock_irqsave(&rds_cong_lock, flags);
+	list_add_tail(&conn->c_map_item, &conn->c_lcong->m_conn_list);
+	spin_unlock_irqrestore(&rds_cong_lock, flags);
+}
+
+void rds_cong_remove_conn(struct rds_connection *conn)
+{
+	unsigned long flags;
+
+	rdsdebug("removing conn %p from map %p\n", conn, conn->c_lcong);
+	spin_lock_irqsave(&rds_cong_lock, flags);
+	list_del_init(&conn->c_map_item);
+	spin_unlock_irqrestore(&rds_cong_lock, flags);
+}
+
+int rds_cong_get_maps(struct rds_connection *conn)
+{
+	conn->c_lcong = rds_cong_from_addr(&conn->c_laddr);
+	conn->c_fcong = rds_cong_from_addr(&conn->c_faddr);
+
+	if (!(conn->c_lcong && conn->c_fcong))
+		return -ENOMEM;
+
+	return 0;
+}
+
+void rds_cong_queue_updates(struct rds_cong_map *map)
+{
+	struct rds_connection *conn;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rds_cong_lock, flags);
+
+	list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
+		struct rds_conn_path *cp = &conn->c_path[0];
+
+		rcu_read_lock();
+		if (!test_and_set_bit(0, &conn->c_map_queued) &&
+		    !rds_destroy_pending(cp->cp_conn)) {
+			rds_stats_inc(s_cong_update_queued);
+			/* We cannot inline the call to rds_send_xmit() here
+			 * for two reasons (both pertaining to a TCP transport):
+			 * 1. When we get here from the receive path, we
+			 *    are already holding the sock_lock (held by
+			 *    tcp_v4_rcv()). So inlining calls to
+			 *    tcp_setsockopt and/or tcp_sendmsg will deadlock
+			 *    when it tries to get the sock_lock())
+			 * 2. Interrupts are masked so that we can mark the
+			 *    port congested from both send and recv paths.
+			 *    (See comment around declaration of rdc_cong_lock).
+			 *    An attempt to get the sock_lock() here will
+			 *    therefore trigger warnings.
+			 * Defer the xmit to rds_send_worker() instead.
+			 */
+			queue_delayed_work(cp->cp_wq, &cp->cp_send_w, 0);
+		}
+		rcu_read_unlock();
+	}
+
+	spin_unlock_irqrestore(&rds_cong_lock, flags);
+}
+
+void rds_cong_map_updated(struct rds_cong_map *map, uint64_t portmask)
+{
+	rdsdebug("waking map %p for %pI4\n",
+	  map, &map->m_addr);
+	rds_stats_inc(s_cong_update_received);
+	atomic_inc(&rds_cong_generation);
+	if (waitqueue_active(&map->m_waitq))
+		wake_up(&map->m_waitq);
+	if (waitqueue_active(&rds_poll_waitq))
+		wake_up_all(&rds_poll_waitq);
+
+	if (portmask && !list_empty(&rds_cong_monitor)) {
+		unsigned long flags;
+		struct rds_sock *rs;
+
+		read_lock_irqsave(&rds_cong_monitor_lock, flags);
+		list_for_each_entry(rs, &rds_cong_monitor, rs_cong_list) {
+			spin_lock(&rs->rs_lock);
+			rs->rs_cong_notify |= (rs->rs_cong_mask & portmask);
+			rs->rs_cong_mask &= ~portmask;
+			spin_unlock(&rs->rs_lock);
+			if (rs->rs_cong_notify)
+				rds_wake_sk_sleep(rs);
+		}
+		read_unlock_irqrestore(&rds_cong_monitor_lock, flags);
+	}
+}
+EXPORT_SYMBOL_GPL(rds_cong_map_updated);
+
+int rds_cong_updated_since(unsigned long *recent)
+{
+	unsigned long gen = atomic_read(&rds_cong_generation);
+
+	if (likely(*recent == gen))
+		return 0;
+	*recent = gen;
+	return 1;
+}
+
+/*
+ * We're called under the locking that protects the sockets receive buffer
+ * consumption.  This makes it a lot easier for the caller to only call us
+ * when it knows that an existing set bit needs to be cleared, and vice versa.
+ * We can't block and we need to deal with concurrent sockets working against
+ * the same per-address map.
+ */
+void rds_cong_set_bit(struct rds_cong_map *map, __be16 port)
+{
+	unsigned long i;
+	unsigned long off;
+
+	rdsdebug("setting congestion for %pI4:%u in map %p\n",
+	  &map->m_addr, ntohs(port), map);
+
+	i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
+	off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
+
+	set_bit_le(off, (void *)map->m_page_addrs[i]);
+}
+
+void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port)
+{
+	unsigned long i;
+	unsigned long off;
+
+	rdsdebug("clearing congestion for %pI4:%u in map %p\n",
+	  &map->m_addr, ntohs(port), map);
+
+	i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
+	off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
+
+	clear_bit_le(off, (void *)map->m_page_addrs[i]);
+}
+
+static int rds_cong_test_bit(struct rds_cong_map *map, __be16 port)
+{
+	unsigned long i;
+	unsigned long off;
+
+	i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
+	off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
+
+	return test_bit_le(off, (void *)map->m_page_addrs[i]);
+}
+
+void rds_cong_add_socket(struct rds_sock *rs)
+{
+	unsigned long flags;
+
+	write_lock_irqsave(&rds_cong_monitor_lock, flags);
+	if (list_empty(&rs->rs_cong_list))
+		list_add(&rs->rs_cong_list, &rds_cong_monitor);
+	write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
+}
+
+void rds_cong_remove_socket(struct rds_sock *rs)
+{
+	unsigned long flags;
+	struct rds_cong_map *map;
+
+	write_lock_irqsave(&rds_cong_monitor_lock, flags);
+	list_del_init(&rs->rs_cong_list);
+	write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
+
+	/* update congestion map for now-closed port */
+	spin_lock_irqsave(&rds_cong_lock, flags);
+	map = rds_cong_tree_walk(&rs->rs_bound_addr, NULL);
+	spin_unlock_irqrestore(&rds_cong_lock, flags);
+
+	if (map && rds_cong_test_bit(map, rs->rs_bound_port)) {
+		rds_cong_clear_bit(map, rs->rs_bound_port);
+		rds_cong_queue_updates(map);
+	}
+}
+
+int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock,
+		  struct rds_sock *rs)
+{
+	if (!rds_cong_test_bit(map, port))
+		return 0;
+	if (nonblock) {
+		if (rs && rs->rs_cong_monitor) {
+			unsigned long flags;
+
+			/* It would have been nice to have an atomic set_bit on
+			 * a uint64_t. */
+			spin_lock_irqsave(&rs->rs_lock, flags);
+			rs->rs_cong_mask |= RDS_CONG_MONITOR_MASK(ntohs(port));
+			spin_unlock_irqrestore(&rs->rs_lock, flags);
+
+			/* Test again - a congestion update may have arrived in
+			 * the meantime. */
+			if (!rds_cong_test_bit(map, port))
+				return 0;
+		}
+		rds_stats_inc(s_cong_send_error);
+		return -ENOBUFS;
+	}
+
+	rds_stats_inc(s_cong_send_blocked);
+	rdsdebug("waiting on map %p for port %u\n", map, be16_to_cpu(port));
+
+	return wait_event_interruptible(map->m_waitq,
+					!rds_cong_test_bit(map, port));
+}
+
+void rds_cong_exit(void)
+{
+	struct rb_node *node;
+	struct rds_cong_map *map;
+	unsigned long i;
+
+	while ((node = rb_first(&rds_cong_tree))) {
+		map = rb_entry(node, struct rds_cong_map, m_rb_node);
+		rdsdebug("freeing map %p\n", map);
+		rb_erase(&map->m_rb_node, &rds_cong_tree);
+		for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
+			free_page(map->m_page_addrs[i]);
+		kfree(map);
+	}
+}
+
+/*
+ * Allocate a RDS message containing a congestion update.
+ */
+struct rds_message *rds_cong_update_alloc(struct rds_connection *conn)
+{
+	struct rds_cong_map *map = conn->c_lcong;
+	struct rds_message *rm;
+
+	rm = rds_message_map_pages(map->m_page_addrs, RDS_CONG_MAP_BYTES);
+	if (!IS_ERR(rm))
+		rm->m_inc.i_hdr.h_flags = RDS_FLAG_CONG_BITMAP;
+
+	return rm;
+}

linux-7.2.3/net/rds/ib.c

@@ -0,0 +1,629 @@
+/*
+ * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <linux/if.h>
+#include <linux/netdevice.h>
+#include <linux/inetdevice.h>
+#include <linux/if_arp.h>
+#include <linux/delay.h>
+#include <linux/slab.h>
+#include <linux/module.h>
+#include <net/addrconf.h>
+
+#include "rds_single_path.h"
+#include "rds.h"
+#include "ib.h"
+#include "ib_mr.h"
+
+static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
+static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
+unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
+static atomic_t rds_ib_unloading;
+
+module_param(rds_ib_mr_1m_pool_size, int, 0444);
+MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
+module_param(rds_ib_mr_8k_pool_size, int, 0444);
+MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
+module_param(rds_ib_retry_count, int, 0444);
+MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
+
+/*
+ * we have a clumsy combination of RCU and a rwsem protecting this list
+ * because it is used both in the get_mr fast path and while blocking in
+ * the FMR flushing path.
+ */
+DECLARE_RWSEM(rds_ib_devices_lock);
+struct list_head rds_ib_devices;
+
+/* NOTE: if also grabbing ibdev lock, grab this first */
+DEFINE_SPINLOCK(ib_nodev_conns_lock);
+LIST_HEAD(ib_nodev_conns);
+
+static void rds_ib_nodev_connect(void)
+{
+	struct rds_ib_connection *ic;
+
+	spin_lock(&ib_nodev_conns_lock);
+	list_for_each_entry(ic, &ib_nodev_conns, ib_node)
+		rds_conn_connect_if_down(ic->conn);
+	spin_unlock(&ib_nodev_conns_lock);
+}
+
+static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
+{
+	struct rds_ib_connection *ic;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rds_ibdev->spinlock, flags);
+	list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
+		rds_conn_path_drop(&ic->conn->c_path[0], true);
+	spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
+}
+
+/*
+ * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
+ * from interrupt context so we push freing off into a work struct in krdsd.
+ */
+static void rds_ib_dev_free(struct work_struct *work)
+{
+	struct rds_ib_ipaddr *i_ipaddr, *i_next;
+	struct rds_ib_device *rds_ibdev = container_of(work,
+					struct rds_ib_device, free_work);
+
+	if (rds_ibdev->mr_8k_pool)
+		rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
+	if (rds_ibdev->mr_1m_pool)
+		rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
+	if (rds_ibdev->pd)
+		ib_dealloc_pd(rds_ibdev->pd);
+
+	list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
+		list_del(&i_ipaddr->list);
+		kfree(i_ipaddr);
+	}
+
+	kfree(rds_ibdev->vector_load);
+
+	kfree(rds_ibdev);
+}
+
+void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
+{
+	BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
+	if (refcount_dec_and_test(&rds_ibdev->refcount))
+		queue_work(rds_wq, &rds_ibdev->free_work);
+}
+
+static int rds_ib_add_one(struct ib_device *device)
+{
+	struct rds_ib_device *rds_ibdev;
+	int ret;
+
+	/* Only handle IB (no iWARP) devices */
+	if (device->node_type != RDMA_NODE_IB_CA)
+		return -EOPNOTSUPP;
+
+	/* Device must support FRWR */
+	if (!(device->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
+		return -EOPNOTSUPP;
+
+	rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
+				 ibdev_to_node(device));
+	if (!rds_ibdev)
+		return -ENOMEM;
+
+	spin_lock_init(&rds_ibdev->spinlock);
+	refcount_set(&rds_ibdev->refcount, 1);
+	INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
+
+	INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
+	INIT_LIST_HEAD(&rds_ibdev->conn_list);
+
+	rds_ibdev->max_wrs = device->attrs.max_qp_wr;
+	rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
+
+	rds_ibdev->odp_capable =
+		!!(device->attrs.kernel_cap_flags &
+		   IBK_ON_DEMAND_PAGING) &&
+		!!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
+		   IB_ODP_SUPPORT_WRITE) &&
+		!!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
+		   IB_ODP_SUPPORT_READ);
+
+	rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
+		min_t(unsigned int, (device->attrs.max_mr / 2),
+		      rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
+
+	rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
+		min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
+		      rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
+
+	rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
+	rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
+
+	rds_ibdev->vector_load = kzalloc_objs(int, device->num_comp_vectors);
+	if (!rds_ibdev->vector_load) {
+		pr_err("RDS/IB: %s failed to allocate vector memory\n",
+			__func__);
+		ret = -ENOMEM;
+		goto put_dev;
+	}
+
+	rds_ibdev->dev = device;
+	rds_ibdev->pd = ib_alloc_pd(device, 0);
+	if (IS_ERR(rds_ibdev->pd)) {
+		ret = PTR_ERR(rds_ibdev->pd);
+		rds_ibdev->pd = NULL;
+		goto put_dev;
+	}
+
+	rds_ibdev->mr_1m_pool =
+		rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
+	if (IS_ERR(rds_ibdev->mr_1m_pool)) {
+		ret = PTR_ERR(rds_ibdev->mr_1m_pool);
+		rds_ibdev->mr_1m_pool = NULL;
+		goto put_dev;
+	}
+
+	rds_ibdev->mr_8k_pool =
+		rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
+	if (IS_ERR(rds_ibdev->mr_8k_pool)) {
+		ret = PTR_ERR(rds_ibdev->mr_8k_pool);
+		rds_ibdev->mr_8k_pool = NULL;
+		goto put_dev;
+	}
+
+	rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, max_1m_mrs = %d, max_8k_mrs = %d\n",
+		 device->attrs.max_mr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
+		 rds_ibdev->max_1m_mrs, rds_ibdev->max_8k_mrs);
+
+	pr_info("RDS/IB: %s: added\n", device->name);
+
+	down_write(&rds_ib_devices_lock);
+	list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
+	up_write(&rds_ib_devices_lock);
+	refcount_inc(&rds_ibdev->refcount);
+
+	ib_set_client_data(device, &rds_ib_client, rds_ibdev);
+
+	rds_ib_nodev_connect();
+	return 0;
+
+put_dev:
+	rds_ib_dev_put(rds_ibdev);
+	return ret;
+}
+
+/*
+ * New connections use this to find the device to associate with the
+ * connection.  It's not in the fast path so we're not concerned about the
+ * performance of the IB call.  (As of this writing, it uses an interrupt
+ * blocking spinlock to serialize walking a per-device list of all registered
+ * clients.)
+ *
+ * RCU is used to handle incoming connections racing with device teardown.
+ * Rather than use a lock to serialize removal from the client_data and
+ * getting a new reference, we use an RCU grace period.  The destruction
+ * path removes the device from client_data and then waits for all RCU
+ * readers to finish.
+ *
+ * A new connection can get NULL from this if its arriving on a
+ * device that is in the process of being removed.
+ */
+struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
+{
+	struct rds_ib_device *rds_ibdev;
+
+	rcu_read_lock();
+	rds_ibdev = ib_get_client_data(device, &rds_ib_client);
+	if (rds_ibdev)
+		refcount_inc(&rds_ibdev->refcount);
+	rcu_read_unlock();
+	return rds_ibdev;
+}
+
+/*
+ * The IB stack is letting us know that a device is going away.  This can
+ * happen if the underlying HCA driver is removed or if PCI hotplug is removing
+ * the pci function, for example.
+ *
+ * This can be called at any time and can be racing with any other RDS path.
+ */
+static void rds_ib_remove_one(struct ib_device *device, void *client_data)
+{
+	struct rds_ib_device *rds_ibdev = client_data;
+
+	rds_ib_dev_shutdown(rds_ibdev);
+
+	/* stop connection attempts from getting a reference to this device. */
+	ib_set_client_data(device, &rds_ib_client, NULL);
+
+	down_write(&rds_ib_devices_lock);
+	list_del_rcu(&rds_ibdev->list);
+	up_write(&rds_ib_devices_lock);
+
+	/*
+	 * This synchronize rcu is waiting for readers of both the ib
+	 * client data and the devices list to finish before we drop
+	 * both of those references.
+	 */
+	synchronize_rcu();
+	rds_ib_dev_put(rds_ibdev);
+	rds_ib_dev_put(rds_ibdev);
+}
+
+struct ib_client rds_ib_client = {
+	.name   = "rds_ib",
+	.add    = rds_ib_add_one,
+	.remove = rds_ib_remove_one
+};
+
+static int rds_ib_conn_info_visitor(struct rds_connection *conn,
+				    void *buffer)
+{
+	struct rds_info_rdma_connection *iinfo = buffer;
+	struct rds_ib_connection *ic = conn->c_transport_data;
+
+	/* We will only ever look at IB transports */
+	if (conn->c_trans != &rds_ib_transport)
+		return 0;
+	if (conn->c_isv6)
+		return 0;
+
+	iinfo->src_addr = conn->c_laddr.s6_addr32[3];
+	iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
+	if (ic) {
+		iinfo->tos = conn->c_tos;
+		iinfo->sl = ic->i_sl;
+	}
+
+	memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
+	memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
+	if (rds_conn_state(conn) == RDS_CONN_UP) {
+		struct rds_ib_device *rds_ibdev;
+
+		rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
+			       (union ib_gid *)&iinfo->dst_gid);
+
+		rds_ibdev = ic->rds_ibdev;
+		iinfo->max_send_wr = ic->i_send_ring.w_nr;
+		iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
+		iinfo->max_send_sge = rds_ibdev->max_sge;
+		rds_ib_get_mr_info(rds_ibdev, iinfo);
+		iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
+	}
+	return 1;
+}
+
+#if IS_ENABLED(CONFIG_IPV6)
+/* IPv6 version of rds_ib_conn_info_visitor(). */
+static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
+				     void *buffer)
+{
+	struct rds6_info_rdma_connection *iinfo6 = buffer;
+	struct rds_ib_connection *ic = conn->c_transport_data;
+
+	/* We will only ever look at IB transports */
+	if (conn->c_trans != &rds_ib_transport)
+		return 0;
+
+	iinfo6->src_addr = conn->c_laddr;
+	iinfo6->dst_addr = conn->c_faddr;
+	if (ic) {
+		iinfo6->tos = conn->c_tos;
+		iinfo6->sl = ic->i_sl;
+	}
+
+	memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
+	memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
+
+	if (rds_conn_state(conn) == RDS_CONN_UP) {
+		struct rds_ib_device *rds_ibdev;
+
+		rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
+			       (union ib_gid *)&iinfo6->dst_gid);
+		rds_ibdev = ic->rds_ibdev;
+		iinfo6->max_send_wr = ic->i_send_ring.w_nr;
+		iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
+		iinfo6->max_send_sge = rds_ibdev->max_sge;
+		rds6_ib_get_mr_info(rds_ibdev, iinfo6);
+		iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
+	}
+	return 1;
+}
+#endif
+
+static void rds_ib_ic_info(struct socket *sock, unsigned int len,
+			   struct rds_info_iterator *iter,
+			   struct rds_info_lengths *lens)
+{
+	u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
+
+	rds_for_each_conn_info(sock, len, iter, lens,
+				rds_ib_conn_info_visitor,
+				buffer,
+				sizeof(struct rds_info_rdma_connection));
+}
+
+#if IS_ENABLED(CONFIG_IPV6)
+/* IPv6 version of rds_ib_ic_info(). */
+static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
+			    struct rds_info_iterator *iter,
+			    struct rds_info_lengths *lens)
+{
+	u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
+
+	rds_for_each_conn_info(sock, len, iter, lens,
+			       rds6_ib_conn_info_visitor,
+			       buffer,
+			       sizeof(struct rds6_info_rdma_connection));
+}
+#endif
+
+/*
+ * Early RDS/IB was built to only bind to an address if there is an IPoIB
+ * device with that address set.
+ *
+ * If it were me, I'd advocate for something more flexible.  Sending and
+ * receiving should be device-agnostic.  Transports would try and maintain
+ * connections between peers who have messages queued.  Userspace would be
+ * allowed to influence which paths have priority.  We could call userspace
+ * asserting this policy "routing".
+ */
+static int rds_ib_laddr_check_cm(struct net *net, const struct in6_addr *addr,
+				 __u32 scope_id)
+{
+	int ret;
+	struct rdma_cm_id *cm_id;
+#if IS_ENABLED(CONFIG_IPV6)
+	struct sockaddr_in6 sin6;
+#endif
+	struct sockaddr_in sin;
+	struct sockaddr *sa;
+	bool isv4;
+
+	isv4 = ipv6_addr_v4mapped(addr);
+	/* Create a CMA ID and try to bind it. This catches both
+	 * IB and iWARP capable NICs.
+	 */
+	cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
+			       NULL, RDMA_PS_TCP, IB_QPT_RC);
+	if (IS_ERR(cm_id))
+		return PTR_ERR(cm_id);
+
+	if (isv4) {
+		memset(&sin, 0, sizeof(sin));
+		sin.sin_family = AF_INET;
+		sin.sin_addr.s_addr = addr->s6_addr32[3];
+		sa = (struct sockaddr *)&sin;
+	} else {
+#if IS_ENABLED(CONFIG_IPV6)
+		if (!ipv6_mod_enabled()) {
+			ret = -EADDRNOTAVAIL;
+			goto out;
+		}
+		memset(&sin6, 0, sizeof(sin6));
+		sin6.sin6_family = AF_INET6;
+		sin6.sin6_addr = *addr;
+		sin6.sin6_scope_id = scope_id;
+		sa = (struct sockaddr *)&sin6;
+
+		/* XXX Do a special IPv6 link local address check here.  The
+		 * reason is that rdma_bind_addr() always succeeds with IPv6
+		 * link local address regardless it is indeed configured in a
+		 * system.
+		 */
+		if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
+			struct net_device *dev;
+
+			if (scope_id == 0) {
+				ret = -EADDRNOTAVAIL;
+				goto out;
+			}
+
+			/* Use init_net for now as RDS is not network
+			 * name space aware.
+			 */
+			dev = dev_get_by_index(&init_net, scope_id);
+			if (!dev) {
+				ret = -EADDRNOTAVAIL;
+				goto out;
+			}
+			if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
+				dev_put(dev);
+				ret = -EADDRNOTAVAIL;
+				goto out;
+			}
+			dev_put(dev);
+		}
+#else
+		ret = -EADDRNOTAVAIL;
+		goto out;
+#endif
+	}
+
+	/* rdma_bind_addr will only succeed for IB & iWARP devices */
+	ret = rdma_bind_addr(cm_id, sa);
+	/* due to this, we will claim to support iWARP devices unless we
+	   check node_type. */
+	if (ret || !cm_id->device ||
+	    cm_id->device->node_type != RDMA_NODE_IB_CA)
+		ret = -EADDRNOTAVAIL;
+
+	rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
+		 addr, scope_id, ret,
+		 cm_id->device ? cm_id->device->node_type : -1);
+
+out:
+	rdma_destroy_id(cm_id);
+
+	return ret;
+}
+
+static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
+			      __u32 scope_id)
+{
+	struct rds_ib_device *rds_ibdev = NULL;
+
+	/* RDS/IB is restricted to the initial network namespace */
+	if (!net_eq(net, &init_net))
+		return -EPROTOTYPE;
+
+	if (ipv6_addr_v4mapped(addr)) {
+		rds_ibdev = rds_ib_get_device(addr->s6_addr32[3]);
+		if (rds_ibdev) {
+			rds_ib_dev_put(rds_ibdev);
+			return 0;
+		}
+	}
+
+	return rds_ib_laddr_check_cm(net, addr, scope_id);
+}
+
+static void rds_ib_unregister_client(void)
+{
+	ib_unregister_client(&rds_ib_client);
+	/* wait for rds_ib_dev_free() to complete */
+	flush_workqueue(rds_wq);
+}
+
+static void rds_ib_set_unloading(void)
+{
+	atomic_set(&rds_ib_unloading, 1);
+}
+
+static bool rds_ib_is_unloading(struct rds_connection *conn)
+{
+	struct rds_conn_path *cp = &conn->c_path[0];
+
+	return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
+		atomic_read(&rds_ib_unloading) != 0);
+}
+
+void rds_ib_exit(void)
+{
+	rds_ib_set_unloading();
+	synchronize_rcu();
+	rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
+#if IS_ENABLED(CONFIG_IPV6)
+	rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
+#endif
+	rds_ib_unregister_client();
+	rds_ib_destroy_nodev_conns();
+	rds_ib_sysctl_exit();
+	rds_ib_recv_exit();
+	rds_trans_unregister(&rds_ib_transport);
+	rds_ib_mr_exit();
+}
+
+static u8 rds_ib_get_tos_map(u8 tos)
+{
+	/* 1:1 user to transport map for RDMA transport.
+	 * In future, if custom map is desired, hook can export
+	 * user configurable map.
+	 */
+	return tos;
+}
+
+struct rds_transport rds_ib_transport = {
+	.laddr_check		= rds_ib_laddr_check,
+	.xmit_path_complete	= rds_ib_xmit_path_complete,
+	.xmit			= rds_ib_xmit,
+	.xmit_rdma		= rds_ib_xmit_rdma,
+	.xmit_atomic		= rds_ib_xmit_atomic,
+	.recv_path		= rds_ib_recv_path,
+	.conn_alloc		= rds_ib_conn_alloc,
+	.conn_free		= rds_ib_conn_free,
+	.conn_path_connect	= rds_ib_conn_path_connect,
+	.conn_path_shutdown	= rds_ib_conn_path_shutdown,
+	.inc_copy_to_user	= rds_ib_inc_copy_to_user,
+	.inc_free		= rds_ib_inc_free,
+	.cm_initiate_connect	= rds_ib_cm_initiate_connect,
+	.cm_handle_connect	= rds_ib_cm_handle_connect,
+	.cm_connect_complete	= rds_ib_cm_connect_complete,
+	.stats_info_copy	= rds_ib_stats_info_copy,
+	.exit			= rds_ib_exit,
+	.get_mr			= rds_ib_get_mr,
+	.sync_mr		= rds_ib_sync_mr,
+	.free_mr		= rds_ib_free_mr,
+	.flush_mrs		= rds_ib_flush_mrs,
+	.get_tos_map		= rds_ib_get_tos_map,
+	.t_owner		= THIS_MODULE,
+	.t_name			= "infiniband",
+	.t_unloading		= rds_ib_is_unloading,
+	.t_type			= RDS_TRANS_IB
+};
+
+int rds_ib_init(void)
+{
+	int ret;
+
+	INIT_LIST_HEAD(&rds_ib_devices);
+
+	ret = rds_ib_mr_init();
+	if (ret)
+		goto out;
+
+	ret = ib_register_client(&rds_ib_client);
+	if (ret)
+		goto out_mr_exit;
+
+	ret = rds_ib_sysctl_init();
+	if (ret)
+		goto out_ibreg;
+
+	ret = rds_ib_recv_init();
+	if (ret)
+		goto out_sysctl;
+
+	rds_trans_register(&rds_ib_transport);
+
+	rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
+#if IS_ENABLED(CONFIG_IPV6)
+	rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
+#endif
+
+	goto out;
+
+out_sysctl:
+	rds_ib_sysctl_exit();
+out_ibreg:
+	rds_ib_unregister_client();
+out_mr_exit:
+	rds_ib_mr_exit();
+out:
+	return ret;
+}
+
+MODULE_LICENSE("GPL");

linux-7.2.3/net/rds/ib.h

@@ -0,0 +1,455 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _RDS_IB_H
+#define _RDS_IB_H
+
+#include <rdma/ib_verbs.h>
+#include <rdma/rdma_cm.h>
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/slab.h>
+#include "rds.h"
+#include "rdma_transport.h"
+
+#define RDS_IB_MAX_SGE			8
+#define RDS_IB_RECV_SGE 		2
+
+#define RDS_IB_DEFAULT_RECV_WR		1024
+#define RDS_IB_DEFAULT_SEND_WR		256
+#define RDS_IB_DEFAULT_FR_WR		512
+
+#define RDS_IB_DEFAULT_RETRY_COUNT	1
+
+#define RDS_IB_SUPPORTED_PROTOCOLS	0x00000003	/* minor versions supported */
+
+#define RDS_IB_RECYCLE_BATCH_COUNT	32
+
+#define RDS_IB_WC_MAX			32
+
+extern struct rw_semaphore rds_ib_devices_lock;
+extern struct list_head rds_ib_devices;
+
+/*
+ * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
+ * try and minimize the amount of memory tied up both the device and
+ * socket receive queues.
+ */
+struct rds_page_frag {
+	struct list_head	f_item;
+	struct list_head	f_cache_entry;
+	struct scatterlist	f_sg;
+};
+
+struct rds_ib_incoming {
+	struct list_head	ii_frags;
+	struct list_head	ii_cache_entry;
+	struct rds_incoming	ii_inc;
+};
+
+struct rds_ib_cache_head {
+	struct list_head *first;
+	unsigned long count;
+};
+
+struct rds_ib_refill_cache {
+	struct rds_ib_cache_head __percpu *percpu;
+	struct list_head	 *xfer;
+	struct list_head	 *ready;
+};
+
+/* This is the common structure for the IB private data exchange in setting up
+ * an RDS connection.  The exchange is different for IPv4 and IPv6 connections.
+ * The reason is that the address size is different and the addresses
+ * exchanged are in the beginning of the structure.  Hence it is not possible
+ * for interoperability if same structure is used.
+ */
+struct rds_ib_conn_priv_cmn {
+	u8			ricpc_protocol_major;
+	u8			ricpc_protocol_minor;
+	__be16			ricpc_protocol_minor_mask;	/* bitmask */
+	u8			ricpc_dp_toss;
+	u8			ripc_reserved1;
+	__be16			ripc_reserved2;
+	__be64			ricpc_ack_seq;
+	__be32			ricpc_credit;	/* non-zero enables flow ctl */
+};
+
+struct rds_ib_connect_private {
+	/* Add new fields at the end, and don't permute existing fields. */
+	__be32				dp_saddr;
+	__be32				dp_daddr;
+	struct rds_ib_conn_priv_cmn	dp_cmn;
+};
+
+struct rds6_ib_connect_private {
+	/* Add new fields at the end, and don't permute existing fields. */
+	struct in6_addr			dp_saddr;
+	struct in6_addr			dp_daddr;
+	struct rds_ib_conn_priv_cmn	dp_cmn;
+};
+
+#define dp_protocol_major	dp_cmn.ricpc_protocol_major
+#define dp_protocol_minor	dp_cmn.ricpc_protocol_minor
+#define dp_protocol_minor_mask	dp_cmn.ricpc_protocol_minor_mask
+#define dp_ack_seq		dp_cmn.ricpc_ack_seq
+#define dp_credit		dp_cmn.ricpc_credit
+
+union rds_ib_conn_priv {
+	struct rds_ib_connect_private	ricp_v4;
+	struct rds6_ib_connect_private	ricp_v6;
+};
+
+struct rds_ib_send_work {
+	void			*s_op;
+	union {
+		struct ib_send_wr	s_wr;
+		struct ib_rdma_wr	s_rdma_wr;
+		struct ib_atomic_wr	s_atomic_wr;
+	};
+	struct ib_sge		s_sge[RDS_IB_MAX_SGE];
+	unsigned long		s_queued;
+};
+
+struct rds_ib_recv_work {
+	struct rds_ib_incoming 	*r_ibinc;
+	struct rds_page_frag	*r_frag;
+	struct ib_recv_wr	r_wr;
+	struct ib_sge		r_sge[2];
+};
+
+struct rds_ib_work_ring {
+	u32		w_nr;
+	u32		w_alloc_ptr;
+	u32		w_alloc_ctr;
+	u32		w_free_ptr;
+	atomic_t	w_free_ctr;
+};
+
+/* Rings are posted with all the allocations they'll need to queue the
+ * incoming message to the receiving socket so this can't fail.
+ * All fragments start with a header, so we can make sure we're not receiving
+ * garbage, and we can tell a small 8 byte fragment from an ACK frame.
+ */
+struct rds_ib_ack_state {
+	u64		ack_next;
+	u64		ack_recv;
+	unsigned int	ack_required:1;
+	unsigned int	ack_next_valid:1;
+	unsigned int	ack_recv_valid:1;
+};
+
+
+struct rds_ib_device;
+
+struct rds_ib_connection {
+
+	struct list_head	ib_node;
+	struct rds_ib_device	*rds_ibdev;
+	struct rds_connection	*conn;
+
+	/* alphabet soup, IBTA style */
+	struct rdma_cm_id	*i_cm_id;
+	struct ib_pd		*i_pd;
+	struct ib_cq		*i_send_cq;
+	struct ib_cq		*i_recv_cq;
+	struct ib_wc		i_send_wc[RDS_IB_WC_MAX];
+	struct ib_wc		i_recv_wc[RDS_IB_WC_MAX];
+
+	/* To control the number of wrs from fastreg */
+	atomic_t		i_fastreg_wrs;
+	atomic_t		i_fastreg_inuse_count;
+
+	/* interrupt handling */
+	struct tasklet_struct	i_send_tasklet;
+	struct tasklet_struct	i_recv_tasklet;
+
+	/* tx */
+	struct rds_ib_work_ring	i_send_ring;
+	struct rm_data_op	*i_data_op;
+	struct rds_header	**i_send_hdrs;
+	dma_addr_t		*i_send_hdrs_dma;
+	struct rds_ib_send_work *i_sends;
+	atomic_t		i_signaled_sends;
+
+	/* rx */
+	struct mutex		i_recv_mutex;
+	struct rds_ib_work_ring	i_recv_ring;
+	struct rds_ib_incoming	*i_ibinc;
+	u32			i_recv_data_rem;
+	struct rds_header	**i_recv_hdrs;
+	dma_addr_t		*i_recv_hdrs_dma;
+	struct rds_ib_recv_work *i_recvs;
+	u64			i_ack_recv;	/* last ACK received */
+	struct rds_ib_refill_cache i_cache_incs;
+	struct rds_ib_refill_cache i_cache_frags;
+	atomic_t		i_cache_allocs;
+
+	/* sending acks */
+	unsigned long		i_ack_flags;
+#ifdef KERNEL_HAS_ATOMIC64
+	atomic64_t		i_ack_next;	/* next ACK to send */
+#else
+	spinlock_t		i_ack_lock;	/* protect i_ack_next */
+	u64			i_ack_next;	/* next ACK to send */
+#endif
+	struct rds_header	*i_ack;
+	struct ib_send_wr	i_ack_wr;
+	struct ib_sge		i_ack_sge;
+	dma_addr_t		i_ack_dma;
+	unsigned long		i_ack_queued;
+
+	/* Flow control related information
+	 *
+	 * Our algorithm uses a pair variables that we need to access
+	 * atomically - one for the send credits, and one posted
+	 * recv credits we need to transfer to remote.
+	 * Rather than protect them using a slow spinlock, we put both into
+	 * a single atomic_t and update it using cmpxchg
+	 */
+	atomic_t		i_credits;
+
+	/* Protocol version specific information */
+	unsigned int		i_flowctl:1;	/* enable/disable flow ctl */
+
+	/* Batched completions */
+	unsigned int		i_unsignaled_wrs;
+
+	/* Endpoint role in connection */
+	bool			i_active_side;
+	atomic_t		i_cq_quiesce;
+
+	/* Send/Recv vectors */
+	int			i_scq_vector;
+	int			i_rcq_vector;
+	u8			i_sl;
+};
+
+/* This assumes that atomic_t is at least 32 bits */
+#define IB_GET_SEND_CREDITS(v)	((v) & 0xffff)
+#define IB_GET_POST_CREDITS(v)	((v) >> 16)
+#define IB_SET_SEND_CREDITS(v)	((v) & 0xffff)
+#define IB_SET_POST_CREDITS(v)	((v) << 16)
+
+struct rds_ib_ipaddr {
+	struct list_head	list;
+	__be32			ipaddr;
+	struct rcu_head		rcu;
+};
+
+enum {
+	RDS_IB_MR_8K_POOL,
+	RDS_IB_MR_1M_POOL,
+};
+
+struct rds_ib_device {
+	struct list_head	list;
+	struct list_head	ipaddr_list;
+	struct list_head	conn_list;
+	struct ib_device	*dev;
+	struct ib_pd		*pd;
+	u8			odp_capable:1;
+
+	unsigned int		max_mrs;
+	struct rds_ib_mr_pool	*mr_1m_pool;
+	struct rds_ib_mr_pool   *mr_8k_pool;
+	unsigned int		max_8k_mrs;
+	unsigned int		max_1m_mrs;
+	int			max_sge;
+	unsigned int		max_wrs;
+	unsigned int		max_initiator_depth;
+	unsigned int		max_responder_resources;
+	spinlock_t		spinlock;	/* protect the above */
+	refcount_t		refcount;
+	struct work_struct	free_work;
+	int			*vector_load;
+};
+
+#define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
+
+/* bits for i_ack_flags */
+#define IB_ACK_IN_FLIGHT	0
+#define IB_ACK_REQUESTED	1
+
+/* Magic WR_ID for ACKs */
+#define RDS_IB_ACK_WR_ID	(~(u64) 0)
+
+struct rds_ib_statistics {
+	uint64_t	s_ib_connect_raced;
+	uint64_t	s_ib_listen_closed_stale;
+	uint64_t	s_ib_evt_handler_call;
+	uint64_t	s_ib_tasklet_call;
+	uint64_t	s_ib_tx_cq_event;
+	uint64_t	s_ib_tx_ring_full;
+	uint64_t	s_ib_tx_throttle;
+	uint64_t	s_ib_tx_sg_mapping_failure;
+	uint64_t	s_ib_tx_stalled;
+	uint64_t	s_ib_tx_credit_updates;
+	uint64_t	s_ib_rx_cq_event;
+	uint64_t	s_ib_rx_ring_empty;
+	uint64_t	s_ib_rx_refill_from_cq;
+	uint64_t	s_ib_rx_refill_from_thread;
+	uint64_t	s_ib_rx_alloc_limit;
+	uint64_t	s_ib_rx_total_frags;
+	uint64_t	s_ib_rx_total_incs;
+	uint64_t	s_ib_rx_credit_updates;
+	uint64_t	s_ib_ack_sent;
+	uint64_t	s_ib_ack_send_failure;
+	uint64_t	s_ib_ack_send_delayed;
+	uint64_t	s_ib_ack_send_piggybacked;
+	uint64_t	s_ib_ack_received;
+	uint64_t	s_ib_rdma_mr_8k_alloc;
+	uint64_t	s_ib_rdma_mr_8k_free;
+	uint64_t	s_ib_rdma_mr_8k_used;
+	uint64_t	s_ib_rdma_mr_8k_pool_flush;
+	uint64_t	s_ib_rdma_mr_8k_pool_wait;
+	uint64_t	s_ib_rdma_mr_8k_pool_depleted;
+	uint64_t	s_ib_rdma_mr_1m_alloc;
+	uint64_t	s_ib_rdma_mr_1m_free;
+	uint64_t	s_ib_rdma_mr_1m_used;
+	uint64_t	s_ib_rdma_mr_1m_pool_flush;
+	uint64_t	s_ib_rdma_mr_1m_pool_wait;
+	uint64_t	s_ib_rdma_mr_1m_pool_depleted;
+	uint64_t	s_ib_rdma_mr_8k_reused;
+	uint64_t	s_ib_rdma_mr_1m_reused;
+	uint64_t	s_ib_atomic_cswp;
+	uint64_t	s_ib_atomic_fadd;
+	uint64_t	s_ib_recv_added_to_cache;
+	uint64_t	s_ib_recv_removed_from_cache;
+};
+
+extern struct workqueue_struct *rds_ib_wq;
+
+/*
+ * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
+ * doesn't define it.
+ */
+static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
+					      struct scatterlist *sglist,
+					      unsigned int sg_dma_len,
+					      int direction)
+{
+	struct scatterlist *sg;
+	unsigned int i;
+
+	for_each_sg(sglist, sg, sg_dma_len, i) {
+		ib_dma_sync_single_for_cpu(dev, sg_dma_address(sg),
+					   sg_dma_len(sg), direction);
+	}
+}
+#define ib_dma_sync_sg_for_cpu	rds_ib_dma_sync_sg_for_cpu
+
+static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
+						 struct scatterlist *sglist,
+						 unsigned int sg_dma_len,
+						 int direction)
+{
+	struct scatterlist *sg;
+	unsigned int i;
+
+	for_each_sg(sglist, sg, sg_dma_len, i) {
+		ib_dma_sync_single_for_device(dev, sg_dma_address(sg),
+					      sg_dma_len(sg), direction);
+	}
+}
+#define ib_dma_sync_sg_for_device	rds_ib_dma_sync_sg_for_device
+
+
+/* ib.c */
+extern struct rds_transport rds_ib_transport;
+struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
+void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
+extern struct ib_client rds_ib_client;
+
+extern unsigned int rds_ib_retry_count;
+
+extern spinlock_t ib_nodev_conns_lock;
+extern struct list_head ib_nodev_conns;
+
+/* ib_cm.c */
+int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
+void rds_ib_conn_free(void *arg);
+int rds_ib_conn_path_connect(struct rds_conn_path *cp);
+void rds_ib_conn_path_shutdown(struct rds_conn_path *cp);
+__printf(2, 3)
+void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
+int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
+			     struct rdma_cm_event *event, bool isv6);
+int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6);
+void rds_ib_cm_connect_complete(struct rds_connection *conn,
+				struct rdma_cm_event *event);
+
+#define rds_ib_conn_error(conn, fmt...) \
+	__rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
+
+/* ib_rdma.c */
+struct rds_ib_device *rds_ib_get_device(__be32 ipaddr);
+int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
+			 struct in6_addr *ipaddr);
+void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
+void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
+void rds_ib_destroy_nodev_conns(void);
+void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
+
+/* ib_recv.c */
+int rds_ib_recv_init(void);
+void rds_ib_recv_exit(void);
+int rds_ib_recv_path(struct rds_conn_path *conn);
+int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp);
+void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
+void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp);
+void rds_ib_inc_free(struct rds_incoming *inc);
+int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
+void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc,
+			     struct rds_ib_ack_state *state);
+void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
+void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
+void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
+void rds_ib_attempt_ack(struct rds_ib_connection *ic);
+void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
+u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
+void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required);
+
+/* ib_ring.c */
+void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
+void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
+u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
+void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
+void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
+int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
+int rds_ib_ring_low(struct rds_ib_work_ring *ring);
+u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
+u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
+extern wait_queue_head_t rds_ib_ring_empty_wait;
+
+/* ib_send.c */
+void rds_ib_xmit_path_complete(struct rds_conn_path *cp);
+int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
+		unsigned int hdr_off, unsigned int sg, unsigned int off);
+void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
+void rds_ib_send_init_ring(struct rds_ib_connection *ic);
+void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
+int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
+void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
+void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
+int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
+			     u32 *adv_credits, int need_posted, int max_posted);
+int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
+
+/* ib_stats.c */
+DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics, rds_ib_stats);
+#define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
+#define rds_ib_stats_add(member, count) \
+		rds_stats_add_which(rds_ib_stats, member, count)
+unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
+				    unsigned int avail);
+
+/* ib_sysctl.c */
+int rds_ib_sysctl_init(void);
+void rds_ib_sysctl_exit(void);
+extern unsigned long rds_ib_sysctl_max_send_wr;
+extern unsigned long rds_ib_sysctl_max_recv_wr;
+extern unsigned long rds_ib_sysctl_max_unsig_wrs;
+extern unsigned long rds_ib_sysctl_max_unsig_bytes;
+extern unsigned long rds_ib_sysctl_max_recv_allocation;
+extern unsigned int rds_ib_sysctl_flow_control;
+
+#endif

linux-7.2.3/net/rds/ib_frmr.c

@@ -0,0 +1,450 @@
+/*
+ * Copyright (c) 2016 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ib_mr.h"
+
+static inline void
+rds_transition_frwr_state(struct rds_ib_mr *ibmr,
+			  enum rds_ib_fr_state old_state,
+			  enum rds_ib_fr_state new_state)
+{
+	if (cmpxchg(&ibmr->u.frmr.fr_state,
+		    old_state, new_state) == old_state &&
+	    old_state == FRMR_IS_INUSE) {
+		/* enforce order of ibmr->u.frmr.fr_state update
+		 * before decrementing i_fastreg_inuse_count
+		 */
+		smp_mb__before_atomic();
+		atomic_dec(&ibmr->ic->i_fastreg_inuse_count);
+		if (waitqueue_active(&rds_ib_ring_empty_wait))
+			wake_up(&rds_ib_ring_empty_wait);
+	}
+}
+
+static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev,
+					   int npages)
+{
+	struct rds_ib_mr_pool *pool;
+	struct rds_ib_mr *ibmr = NULL;
+	struct rds_ib_frmr *frmr;
+	int err = 0;
+
+	if (npages <= RDS_MR_8K_MSG_SIZE)
+		pool = rds_ibdev->mr_8k_pool;
+	else
+		pool = rds_ibdev->mr_1m_pool;
+
+	ibmr = rds_ib_try_reuse_ibmr(pool);
+	if (ibmr)
+		return ibmr;
+
+	ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL,
+			    rdsibdev_to_node(rds_ibdev));
+	if (!ibmr) {
+		err = -ENOMEM;
+		goto out_no_cigar;
+	}
+
+	frmr = &ibmr->u.frmr;
+	frmr->mr = ib_alloc_mr(rds_ibdev->pd, IB_MR_TYPE_MEM_REG,
+			 pool->max_pages);
+	if (IS_ERR(frmr->mr)) {
+		pr_warn("RDS/IB: %s failed to allocate MR", __func__);
+		err = PTR_ERR(frmr->mr);
+		goto out_no_cigar;
+	}
+
+	ibmr->pool = pool;
+	if (pool->pool_type == RDS_IB_MR_8K_POOL)
+		rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
+	else
+		rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
+
+	if (atomic_read(&pool->item_count) > pool->max_items_soft)
+		pool->max_items_soft = pool->max_items;
+
+	frmr->fr_state = FRMR_IS_FREE;
+	init_waitqueue_head(&frmr->fr_inv_done);
+	init_waitqueue_head(&frmr->fr_reg_done);
+	return ibmr;
+
+out_no_cigar:
+	kfree(ibmr);
+	atomic_dec(&pool->item_count);
+	return ERR_PTR(err);
+}
+
+static void rds_ib_free_frmr(struct rds_ib_mr *ibmr, bool drop)
+{
+	struct rds_ib_mr_pool *pool = ibmr->pool;
+
+	if (drop)
+		llist_add(&ibmr->llnode, &pool->drop_list);
+	else
+		llist_add(&ibmr->llnode, &pool->free_list);
+	atomic_add(ibmr->sg_len, &pool->free_pinned);
+	atomic_inc(&pool->dirty_count);
+
+	/* If we've pinned too many pages, request a flush */
+	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
+	    atomic_read(&pool->dirty_count) >= pool->max_items / 5)
+		queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
+}
+
+static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr)
+{
+	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+	struct ib_reg_wr reg_wr;
+	int ret, off = 0;
+
+	while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
+		atomic_inc(&ibmr->ic->i_fastreg_wrs);
+		cpu_relax();
+	}
+
+	ret = ib_map_mr_sg_zbva(frmr->mr, ibmr->sg, ibmr->sg_dma_len,
+				&off, PAGE_SIZE);
+	if (unlikely(ret != ibmr->sg_dma_len)) {
+		ret = ret < 0 ? ret : -EINVAL;
+		goto out_inc;
+	}
+
+	if (cmpxchg(&frmr->fr_state, FRMR_IS_FREE, FRMR_IS_INUSE) != FRMR_IS_FREE) {
+		ret = -EBUSY;
+		goto out_inc;
+	}
+
+	atomic_inc(&ibmr->ic->i_fastreg_inuse_count);
+
+	/* Perform a WR for the fast_reg_mr. Each individual page
+	 * in the sg list is added to the fast reg page list and placed
+	 * inside the fast_reg_mr WR.  The key used is a rolling 8bit
+	 * counter, which should guarantee uniqueness.
+	 */
+	ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++);
+	frmr->fr_reg = true;
+
+	memset(&reg_wr, 0, sizeof(reg_wr));
+	reg_wr.wr.wr_id = (unsigned long)(void *)ibmr;
+	reg_wr.wr.opcode = IB_WR_REG_MR;
+	reg_wr.wr.num_sge = 0;
+	reg_wr.mr = frmr->mr;
+	reg_wr.key = frmr->mr->rkey;
+	reg_wr.access = IB_ACCESS_LOCAL_WRITE |
+			IB_ACCESS_REMOTE_READ |
+			IB_ACCESS_REMOTE_WRITE;
+	reg_wr.wr.send_flags = IB_SEND_SIGNALED;
+
+	ret = ib_post_send(ibmr->ic->i_cm_id->qp, &reg_wr.wr, NULL);
+	if (unlikely(ret)) {
+		/* Failure here can be because of -ENOMEM as well */
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE);
+
+		if (printk_ratelimit())
+			pr_warn("RDS/IB: %s returned error(%d)\n",
+				__func__, ret);
+		goto out_inc;
+	}
+
+	/* Wait for the registration to complete in order to prevent an invalid
+	 * access error resulting from a race between the memory region already
+	 * being accessed while registration is still pending.
+	 */
+	wait_event(frmr->fr_reg_done, !frmr->fr_reg);
+
+	return ret;
+
+out_inc:
+	atomic_inc(&ibmr->ic->i_fastreg_wrs);
+	return ret;
+}
+
+static int rds_ib_map_frmr(struct rds_ib_device *rds_ibdev,
+			   struct rds_ib_mr_pool *pool,
+			   struct rds_ib_mr *ibmr,
+			   struct scatterlist *sg, unsigned int sg_len)
+{
+	struct ib_device *dev = rds_ibdev->dev;
+	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+	int i;
+	u32 len;
+	int ret = 0;
+
+	/* We want to teardown old ibmr values here and fill it up with
+	 * new sg values
+	 */
+	rds_ib_teardown_mr(ibmr);
+
+	ibmr->sg = sg;
+	ibmr->sg_len = sg_len;
+	ibmr->sg_dma_len = 0;
+	frmr->sg_byte_len = 0;
+	WARN_ON(ibmr->sg_dma_len);
+	ibmr->sg_dma_len = ib_dma_map_sg(dev, ibmr->sg, ibmr->sg_len,
+					 DMA_BIDIRECTIONAL);
+	if (unlikely(!ibmr->sg_dma_len)) {
+		pr_warn("RDS/IB: %s failed!\n", __func__);
+		return -EBUSY;
+	}
+
+	frmr->sg_byte_len = 0;
+	frmr->dma_npages = 0;
+	len = 0;
+
+	ret = -EINVAL;
+	for (i = 0; i < ibmr->sg_dma_len; ++i) {
+		unsigned int dma_len = sg_dma_len(&ibmr->sg[i]);
+		u64 dma_addr = sg_dma_address(&ibmr->sg[i]);
+
+		frmr->sg_byte_len += dma_len;
+		if (dma_addr & ~PAGE_MASK) {
+			if (i > 0)
+				goto out_unmap;
+			else
+				++frmr->dma_npages;
+		}
+
+		if ((dma_addr + dma_len) & ~PAGE_MASK) {
+			if (i < ibmr->sg_dma_len - 1)
+				goto out_unmap;
+			else
+				++frmr->dma_npages;
+		}
+
+		len += dma_len;
+	}
+	frmr->dma_npages += len >> PAGE_SHIFT;
+
+	if (frmr->dma_npages > ibmr->pool->max_pages) {
+		ret = -EMSGSIZE;
+		goto out_unmap;
+	}
+
+	ret = rds_ib_post_reg_frmr(ibmr);
+	if (ret)
+		goto out_unmap;
+
+	if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
+		rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
+	else
+		rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
+
+	return ret;
+
+out_unmap:
+	ib_dma_unmap_sg(rds_ibdev->dev, ibmr->sg, ibmr->sg_len,
+			DMA_BIDIRECTIONAL);
+	ibmr->sg_dma_len = 0;
+	return ret;
+}
+
+static int rds_ib_post_inv(struct rds_ib_mr *ibmr)
+{
+	struct ib_send_wr *s_wr;
+	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+	struct rdma_cm_id *i_cm_id = ibmr->ic->i_cm_id;
+	int ret = -EINVAL;
+
+	if (!i_cm_id || !i_cm_id->qp || !frmr->mr)
+		goto out;
+
+	if (frmr->fr_state != FRMR_IS_INUSE)
+		goto out;
+
+	while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
+		atomic_inc(&ibmr->ic->i_fastreg_wrs);
+		cpu_relax();
+	}
+
+	frmr->fr_inv = true;
+	s_wr = &frmr->fr_wr;
+
+	memset(s_wr, 0, sizeof(*s_wr));
+	s_wr->wr_id = (unsigned long)(void *)ibmr;
+	s_wr->opcode = IB_WR_LOCAL_INV;
+	s_wr->ex.invalidate_rkey = frmr->mr->rkey;
+	s_wr->send_flags = IB_SEND_SIGNALED;
+
+	ret = ib_post_send(i_cm_id->qp, s_wr, NULL);
+	if (unlikely(ret)) {
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE);
+		frmr->fr_inv = false;
+		/* enforce order of frmr->fr_inv update
+		 * before incrementing i_fastreg_wrs
+		 */
+		smp_mb__before_atomic();
+		atomic_inc(&ibmr->ic->i_fastreg_wrs);
+		pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret);
+		goto out;
+	}
+
+	/* Wait for the FRMR_IS_FREE (or FRMR_IS_STALE) transition in order to
+	 * 1) avoid a silly bouncing between "clean_list" and "drop_list"
+	 *    triggered by function "rds_ib_reg_frmr" as it is releases frmr
+	 *    regions whose state is not "FRMR_IS_FREE" right away.
+	 * 2) prevents an invalid access error in a race
+	 *    from a pending "IB_WR_LOCAL_INV" operation
+	 *    with a teardown ("dma_unmap_sg", "put_page")
+	 *    and de-registration ("ib_dereg_mr") of the corresponding
+	 *    memory region.
+	 */
+	wait_event(frmr->fr_inv_done, frmr->fr_state != FRMR_IS_INUSE);
+
+out:
+	return ret;
+}
+
+void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
+{
+	struct rds_ib_mr *ibmr = (void *)(unsigned long)wc->wr_id;
+	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+
+	if (wc->status != IB_WC_SUCCESS) {
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE);
+		if (rds_conn_up(ic->conn))
+			rds_ib_conn_error(ic->conn,
+					  "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n",
+					  &ic->conn->c_laddr,
+					  &ic->conn->c_faddr,
+					  wc->status,
+					  ib_wc_status_msg(wc->status),
+					  wc->vendor_err);
+	}
+
+	if (frmr->fr_inv) {
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_FREE);
+		frmr->fr_inv = false;
+		wake_up(&frmr->fr_inv_done);
+	}
+
+	if (frmr->fr_reg) {
+		frmr->fr_reg = false;
+		wake_up(&frmr->fr_reg_done);
+	}
+
+	/* enforce order of frmr->{fr_reg,fr_inv} update
+	 * before incrementing i_fastreg_wrs
+	 */
+	smp_mb__before_atomic();
+	atomic_inc(&ic->i_fastreg_wrs);
+}
+
+void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
+		       unsigned long *unpinned, unsigned int goal)
+{
+	struct rds_ib_mr *ibmr, *next;
+	struct rds_ib_frmr *frmr;
+	int ret = 0, ret2;
+	unsigned int freed = *nfreed;
+
+	/* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
+	list_for_each_entry(ibmr, list, unmap_list) {
+		if (ibmr->sg_dma_len) {
+			ret2 = rds_ib_post_inv(ibmr);
+			if (ret2 && !ret)
+				ret = ret2;
+		}
+	}
+
+	if (ret)
+		pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret);
+
+	/* Now we can destroy the DMA mapping and unpin any pages */
+	list_for_each_entry_safe(ibmr, next, list, unmap_list) {
+		*unpinned += ibmr->sg_len;
+		frmr = &ibmr->u.frmr;
+		__rds_ib_teardown_mr(ibmr);
+		if (freed < goal || frmr->fr_state == FRMR_IS_STALE) {
+			/* Don't de-allocate if the MR is not free yet */
+			if (frmr->fr_state == FRMR_IS_INUSE)
+				continue;
+
+			if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
+				rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
+			else
+				rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
+			list_del(&ibmr->unmap_list);
+			if (frmr->mr)
+				ib_dereg_mr(frmr->mr);
+			kfree(ibmr);
+			freed++;
+		}
+	}
+	*nfreed = freed;
+}
+
+struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev,
+				  struct rds_ib_connection *ic,
+				  struct scatterlist *sg,
+				  unsigned long nents, u32 *key)
+{
+	struct rds_ib_mr *ibmr = NULL;
+	struct rds_ib_frmr *frmr;
+	int ret;
+
+	if (!ic) {
+		/* TODO: Add FRWR support for RDS_GET_MR using proxy qp*/
+		return ERR_PTR(-EOPNOTSUPP);
+	}
+
+	do {
+		if (ibmr)
+			rds_ib_free_frmr(ibmr, true);
+		ibmr = rds_ib_alloc_frmr(rds_ibdev, nents);
+		if (IS_ERR(ibmr))
+			return ibmr;
+		frmr = &ibmr->u.frmr;
+	} while (frmr->fr_state != FRMR_IS_FREE);
+
+	ibmr->ic = ic;
+	ibmr->device = rds_ibdev;
+	ret = rds_ib_map_frmr(rds_ibdev, ibmr->pool, ibmr, sg, nents);
+	if (ret == 0) {
+		*key = frmr->mr->rkey;
+	} else {
+		rds_ib_free_frmr(ibmr, false);
+		ibmr = ERR_PTR(ret);
+	}
+
+	return ibmr;
+}
+
+void rds_ib_free_frmr_list(struct rds_ib_mr *ibmr)
+{
+	struct rds_ib_mr_pool *pool = ibmr->pool;
+	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+
+	if (frmr->fr_state == FRMR_IS_STALE)
+		llist_add(&ibmr->llnode, &pool->drop_list);
+	else
+		llist_add(&ibmr->llnode, &pool->free_list);
+}

linux-7.2.3/net/rds/ib_mr.h

@@ -0,0 +1,142 @@
+/*
+ * Copyright (c) 2016 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef _RDS_IB_MR_H
+#define _RDS_IB_MR_H
+
+#include <linux/kernel.h>
+
+#include "rds.h"
+#include "ib.h"
+
+#define RDS_MR_1M_POOL_SIZE		(8192 / 2)
+#define RDS_MR_1M_MSG_SIZE		256
+#define RDS_MR_8K_MSG_SIZE		2
+#define RDS_MR_8K_SCALE			(256 / (RDS_MR_8K_MSG_SIZE + 1))
+#define RDS_MR_8K_POOL_SIZE		(RDS_MR_8K_SCALE * (8192 / 2))
+
+enum rds_ib_fr_state {
+	FRMR_IS_FREE,	/* mr invalidated & ready for use */
+	FRMR_IS_INUSE,	/* mr is in use or used & can be invalidated */
+	FRMR_IS_STALE,	/* Stale MR and needs to be dropped  */
+};
+
+struct rds_ib_frmr {
+	struct ib_mr		*mr;
+	enum rds_ib_fr_state	fr_state;
+	bool			fr_inv;
+	wait_queue_head_t	fr_inv_done;
+	bool			fr_reg;
+	wait_queue_head_t	fr_reg_done;
+	struct ib_send_wr	fr_wr;
+	unsigned int		dma_npages;
+	unsigned int		sg_byte_len;
+};
+
+/* This is stored as mr->r_trans_private. */
+struct rds_ib_mr {
+	struct delayed_work		work;
+	struct rds_ib_device		*device;
+	struct rds_ib_mr_pool		*pool;
+	struct rds_ib_connection	*ic;
+
+	struct llist_node		llnode;
+
+	/* unmap_list is for freeing */
+	struct list_head		unmap_list;
+	unsigned int			remap_count;
+
+	struct scatterlist		*sg;
+	unsigned int			sg_len;
+	int				sg_dma_len;
+
+	u8				odp:1;
+	union {
+		struct rds_ib_frmr	frmr;
+		struct ib_mr		*mr;
+	} u;
+};
+
+/* Our own little MR pool */
+struct rds_ib_mr_pool {
+	unsigned int            pool_type;
+	struct mutex		flush_lock;	/* serialize fmr invalidate */
+	struct delayed_work	flush_worker;	/* flush worker */
+
+	atomic_t		item_count;	/* total # of MRs */
+	atomic_t		dirty_count;	/* # dirty of MRs */
+
+	struct llist_head	drop_list;	/* MRs not reached max_maps */
+	struct llist_head	free_list;	/* unused MRs */
+	struct llist_head	clean_list;	/* unused & unmapped MRs */
+	wait_queue_head_t	flush_wait;
+	spinlock_t		clean_lock;	/* "clean_list" concurrency */
+
+	atomic_t		free_pinned;	/* memory pinned by free MRs */
+	unsigned long		max_items;
+	unsigned long		max_items_soft;
+	unsigned long		max_free_pinned;
+	unsigned int		max_pages;
+};
+
+extern struct workqueue_struct *rds_ib_mr_wq;
+
+struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_dev,
+					     int npages);
+void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
+			struct rds_info_rdma_connection *iinfo);
+void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
+			 struct rds6_info_rdma_connection *iinfo6);
+void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
+void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
+		    struct rds_sock *rs, u32 *key_ret,
+		    struct rds_connection *conn, u64 start, u64 length,
+		    int need_odp);
+void rds_ib_sync_mr(void *trans_private, int dir);
+void rds_ib_free_mr(void *trans_private, int invalidate);
+void rds_ib_flush_mrs(void);
+int rds_ib_mr_init(void);
+void rds_ib_mr_exit(void);
+u32 rds_ib_get_lkey(void *trans_private);
+
+void __rds_ib_teardown_mr(struct rds_ib_mr *);
+void rds_ib_teardown_mr(struct rds_ib_mr *);
+struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *);
+int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *, int, struct rds_ib_mr **);
+struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *);
+struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev,
+				  struct rds_ib_connection *ic,
+				  struct scatterlist *sg,
+				  unsigned long nents, u32 *key);
+void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
+		       unsigned long *unpinned, unsigned int goal);
+void rds_ib_free_frmr_list(struct rds_ib_mr *);
+#endif

linux-7.2.3/net/rds/ib_rdma.c

@@ -0,0 +1,703 @@
+/*
+ * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/rculist.h>
+#include <linux/llist.h>
+
+#include "rds_single_path.h"
+#include "ib_mr.h"
+#include "rds.h"
+
+struct workqueue_struct *rds_ib_mr_wq;
+
+static void rds_ib_odp_mr_worker(struct work_struct *work);
+
+struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
+{
+	struct rds_ib_device *rds_ibdev;
+	struct rds_ib_ipaddr *i_ipaddr;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
+		list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
+			if (i_ipaddr->ipaddr == ipaddr) {
+				refcount_inc(&rds_ibdev->refcount);
+				rcu_read_unlock();
+				return rds_ibdev;
+			}
+		}
+	}
+	rcu_read_unlock();
+
+	return NULL;
+}
+
+static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
+{
+	struct rds_ib_ipaddr *i_ipaddr;
+
+	i_ipaddr = kmalloc_obj(*i_ipaddr);
+	if (!i_ipaddr)
+		return -ENOMEM;
+
+	i_ipaddr->ipaddr = ipaddr;
+
+	spin_lock_irq(&rds_ibdev->spinlock);
+	list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
+	spin_unlock_irq(&rds_ibdev->spinlock);
+
+	return 0;
+}
+
+static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
+{
+	struct rds_ib_ipaddr *i_ipaddr;
+	struct rds_ib_ipaddr *to_free = NULL;
+
+
+	spin_lock_irq(&rds_ibdev->spinlock);
+	list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
+		if (i_ipaddr->ipaddr == ipaddr) {
+			list_del_rcu(&i_ipaddr->list);
+			to_free = i_ipaddr;
+			break;
+		}
+	}
+	spin_unlock_irq(&rds_ibdev->spinlock);
+
+	if (to_free)
+		kfree_rcu(to_free, rcu);
+}
+
+int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
+			 struct in6_addr *ipaddr)
+{
+	struct rds_ib_device *rds_ibdev_old;
+
+	rds_ibdev_old = rds_ib_get_device(ipaddr->s6_addr32[3]);
+	if (!rds_ibdev_old)
+		return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
+
+	if (rds_ibdev_old != rds_ibdev) {
+		rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr->s6_addr32[3]);
+		rds_ib_dev_put(rds_ibdev_old);
+		return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
+	}
+	rds_ib_dev_put(rds_ibdev_old);
+
+	return 0;
+}
+
+void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
+{
+	struct rds_ib_connection *ic = conn->c_transport_data;
+
+	/* conn was previously on the nodev_conns_list */
+	spin_lock_irq(&ib_nodev_conns_lock);
+	BUG_ON(list_empty(&ib_nodev_conns));
+	BUG_ON(list_empty(&ic->ib_node));
+	list_del(&ic->ib_node);
+
+	spin_lock(&rds_ibdev->spinlock);
+	list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
+	spin_unlock(&rds_ibdev->spinlock);
+	spin_unlock_irq(&ib_nodev_conns_lock);
+
+	ic->rds_ibdev = rds_ibdev;
+	refcount_inc(&rds_ibdev->refcount);
+}
+
+void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
+{
+	struct rds_ib_connection *ic = conn->c_transport_data;
+
+	/* place conn on nodev_conns_list */
+	spin_lock(&ib_nodev_conns_lock);
+
+	spin_lock_irq(&rds_ibdev->spinlock);
+	BUG_ON(list_empty(&ic->ib_node));
+	list_del(&ic->ib_node);
+	spin_unlock_irq(&rds_ibdev->spinlock);
+
+	list_add_tail(&ic->ib_node, &ib_nodev_conns);
+
+	spin_unlock(&ib_nodev_conns_lock);
+
+	ic->rds_ibdev = NULL;
+	rds_ib_dev_put(rds_ibdev);
+}
+
+void rds_ib_destroy_nodev_conns(void)
+{
+	struct rds_ib_connection *ic, *_ic;
+	LIST_HEAD(tmp_list);
+
+	/* avoid calling conn_destroy with irqs off */
+	spin_lock_irq(&ib_nodev_conns_lock);
+	list_splice(&ib_nodev_conns, &tmp_list);
+	spin_unlock_irq(&ib_nodev_conns_lock);
+
+	list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
+		rds_conn_destroy(ic->conn);
+}
+
+void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
+{
+	struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
+
+	iinfo->rdma_mr_max = pool_1m->max_items;
+	iinfo->rdma_mr_size = pool_1m->max_pages;
+}
+
+#if IS_ENABLED(CONFIG_IPV6)
+void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
+			 struct rds6_info_rdma_connection *iinfo6)
+{
+	struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
+
+	iinfo6->rdma_mr_max = pool_1m->max_items;
+	iinfo6->rdma_mr_size = pool_1m->max_pages;
+}
+#endif
+
+struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
+{
+	struct rds_ib_mr *ibmr = NULL;
+	struct llist_node *ret;
+	unsigned long flags;
+
+	spin_lock_irqsave(&pool->clean_lock, flags);
+	ret = llist_del_first(&pool->clean_list);
+	spin_unlock_irqrestore(&pool->clean_lock, flags);
+	if (ret) {
+		ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
+		if (pool->pool_type == RDS_IB_MR_8K_POOL)
+			rds_ib_stats_inc(s_ib_rdma_mr_8k_reused);
+		else
+			rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
+	}
+
+	return ibmr;
+}
+
+void rds_ib_sync_mr(void *trans_private, int direction)
+{
+	struct rds_ib_mr *ibmr = trans_private;
+	struct rds_ib_device *rds_ibdev = ibmr->device;
+
+	if (ibmr->odp)
+		return;
+
+	switch (direction) {
+	case DMA_FROM_DEVICE:
+		ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
+			ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
+		break;
+	case DMA_TO_DEVICE:
+		ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
+			ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
+		break;
+	}
+}
+
+void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
+{
+	struct rds_ib_device *rds_ibdev = ibmr->device;
+
+	if (ibmr->sg_dma_len) {
+		ib_dma_unmap_sg(rds_ibdev->dev,
+				ibmr->sg, ibmr->sg_len,
+				DMA_BIDIRECTIONAL);
+		ibmr->sg_dma_len = 0;
+	}
+
+	/* Release the s/g list */
+	if (ibmr->sg_len) {
+		unsigned int i;
+
+		for (i = 0; i < ibmr->sg_len; ++i) {
+			struct page *page = sg_page(&ibmr->sg[i]);
+
+			/* FIXME we need a way to tell a r/w MR
+			 * from a r/o MR */
+			WARN_ON(!page->mapping && irqs_disabled());
+			set_page_dirty(page);
+			put_page(page);
+		}
+		kfree(ibmr->sg);
+
+		ibmr->sg = NULL;
+		ibmr->sg_len = 0;
+	}
+}
+
+void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
+{
+	unsigned int pinned = ibmr->sg_len;
+
+	__rds_ib_teardown_mr(ibmr);
+	if (pinned) {
+		struct rds_ib_mr_pool *pool = ibmr->pool;
+
+		atomic_sub(pinned, &pool->free_pinned);
+	}
+}
+
+static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
+{
+	unsigned int item_count;
+
+	item_count = atomic_read(&pool->item_count);
+	if (free_all)
+		return item_count;
+
+	return 0;
+}
+
+/*
+ * given an llist of mrs, put them all into the list_head for more processing
+ */
+static unsigned int llist_append_to_list(struct llist_head *llist,
+					 struct list_head *list)
+{
+	struct rds_ib_mr *ibmr;
+	struct llist_node *node;
+	struct llist_node *next;
+	unsigned int count = 0;
+
+	node = llist_del_all(llist);
+	while (node) {
+		next = node->next;
+		ibmr = llist_entry(node, struct rds_ib_mr, llnode);
+		list_add_tail(&ibmr->unmap_list, list);
+		node = next;
+		count++;
+	}
+	return count;
+}
+
+/*
+ * this takes a list head of mrs and turns it into linked llist nodes
+ * of clusters.  Each cluster has linked llist nodes of
+ * MR_CLUSTER_SIZE mrs that are ready for reuse.
+ */
+static void list_to_llist_nodes(struct list_head *list,
+				struct llist_node **nodes_head,
+				struct llist_node **nodes_tail)
+{
+	struct rds_ib_mr *ibmr;
+	struct llist_node *cur = NULL;
+	struct llist_node **next = nodes_head;
+
+	list_for_each_entry(ibmr, list, unmap_list) {
+		cur = &ibmr->llnode;
+		*next = cur;
+		next = &cur->next;
+	}
+	*next = NULL;
+	*nodes_tail = cur;
+}
+
+/*
+ * Flush our pool of MRs.
+ * At a minimum, all currently unused MRs are unmapped.
+ * If the number of MRs allocated exceeds the limit, we also try
+ * to free as many MRs as needed to get back to this limit.
+ */
+int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
+			 int free_all, struct rds_ib_mr **ibmr_ret)
+{
+	struct rds_ib_mr *ibmr;
+	struct llist_node *clean_nodes;
+	struct llist_node *clean_tail;
+	LIST_HEAD(unmap_list);
+	unsigned long unpinned = 0;
+	unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
+
+	if (pool->pool_type == RDS_IB_MR_8K_POOL)
+		rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
+	else
+		rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
+
+	if (ibmr_ret) {
+		DEFINE_WAIT(wait);
+		while (!mutex_trylock(&pool->flush_lock)) {
+			ibmr = rds_ib_reuse_mr(pool);
+			if (ibmr) {
+				*ibmr_ret = ibmr;
+				finish_wait(&pool->flush_wait, &wait);
+				goto out_nolock;
+			}
+
+			prepare_to_wait(&pool->flush_wait, &wait,
+					TASK_UNINTERRUPTIBLE);
+			if (llist_empty(&pool->clean_list))
+				schedule();
+
+			ibmr = rds_ib_reuse_mr(pool);
+			if (ibmr) {
+				*ibmr_ret = ibmr;
+				finish_wait(&pool->flush_wait, &wait);
+				goto out_nolock;
+			}
+		}
+		finish_wait(&pool->flush_wait, &wait);
+	} else
+		mutex_lock(&pool->flush_lock);
+
+	if (ibmr_ret) {
+		ibmr = rds_ib_reuse_mr(pool);
+		if (ibmr) {
+			*ibmr_ret = ibmr;
+			goto out;
+		}
+	}
+
+	/* Get the list of all MRs to be dropped. Ordering matters -
+	 * we want to put drop_list ahead of free_list.
+	 */
+	dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
+	dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
+	if (free_all) {
+		unsigned long flags;
+
+		spin_lock_irqsave(&pool->clean_lock, flags);
+		llist_append_to_list(&pool->clean_list, &unmap_list);
+		spin_unlock_irqrestore(&pool->clean_lock, flags);
+	}
+
+	free_goal = rds_ib_flush_goal(pool, free_all);
+
+	if (list_empty(&unmap_list))
+		goto out;
+
+	rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
+
+	if (!list_empty(&unmap_list)) {
+		unsigned long flags;
+
+		list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
+		if (ibmr_ret) {
+			*ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
+			clean_nodes = clean_nodes->next;
+		}
+		/* more than one entry in llist nodes */
+		if (clean_nodes) {
+			spin_lock_irqsave(&pool->clean_lock, flags);
+			llist_add_batch(clean_nodes, clean_tail,
+					&pool->clean_list);
+			spin_unlock_irqrestore(&pool->clean_lock, flags);
+		}
+	}
+
+	atomic_sub(unpinned, &pool->free_pinned);
+	atomic_sub(dirty_to_clean, &pool->dirty_count);
+	atomic_sub(nfreed, &pool->item_count);
+
+out:
+	mutex_unlock(&pool->flush_lock);
+	if (waitqueue_active(&pool->flush_wait))
+		wake_up(&pool->flush_wait);
+out_nolock:
+	return 0;
+}
+
+struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
+{
+	struct rds_ib_mr *ibmr = NULL;
+	int iter = 0;
+
+	while (1) {
+		ibmr = rds_ib_reuse_mr(pool);
+		if (ibmr)
+			return ibmr;
+
+		if (atomic_inc_return(&pool->item_count) <= pool->max_items)
+			break;
+
+		atomic_dec(&pool->item_count);
+
+		if (++iter > 2) {
+			if (pool->pool_type == RDS_IB_MR_8K_POOL)
+				rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
+			else
+				rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
+			break;
+		}
+
+		/* We do have some empty MRs. Flush them out. */
+		if (pool->pool_type == RDS_IB_MR_8K_POOL)
+			rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
+		else
+			rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
+
+		rds_ib_flush_mr_pool(pool, 0, &ibmr);
+		if (ibmr)
+			return ibmr;
+	}
+
+	return NULL;
+}
+
+static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
+{
+	struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
+
+	rds_ib_flush_mr_pool(pool, 0, NULL);
+}
+
+void rds_ib_free_mr(void *trans_private, int invalidate)
+{
+	struct rds_ib_mr *ibmr = trans_private;
+	struct rds_ib_mr_pool *pool = ibmr->pool;
+	struct rds_ib_device *rds_ibdev = ibmr->device;
+
+	rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
+
+	if (ibmr->odp) {
+		/* A MR created and marked as use_once. We use delayed work,
+		 * because there is a change that we are in interrupt and can't
+		 * call to ib_dereg_mr() directly.
+		 */
+		INIT_DELAYED_WORK(&ibmr->work, rds_ib_odp_mr_worker);
+		queue_delayed_work(rds_ib_mr_wq, &ibmr->work, 0);
+		return;
+	}
+
+	/* Return it to the pool's free list */
+	rds_ib_free_frmr_list(ibmr);
+
+	atomic_add(ibmr->sg_len, &pool->free_pinned);
+	atomic_inc(&pool->dirty_count);
+
+	/* If we've pinned too many pages, request a flush */
+	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
+	    atomic_read(&pool->dirty_count) >= pool->max_items / 5)
+		queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
+
+	if (invalidate) {
+		if (likely(!in_interrupt())) {
+			rds_ib_flush_mr_pool(pool, 0, NULL);
+		} else {
+			/* We get here if the user created a MR marked
+			 * as use_once and invalidate at the same time.
+			 */
+			queue_delayed_work(rds_ib_mr_wq,
+					   &pool->flush_worker, 10);
+		}
+	}
+
+	rds_ib_dev_put(rds_ibdev);
+}
+
+void rds_ib_flush_mrs(void)
+{
+	struct rds_ib_device *rds_ibdev;
+
+	down_read(&rds_ib_devices_lock);
+	list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
+		if (rds_ibdev->mr_8k_pool)
+			rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
+
+		if (rds_ibdev->mr_1m_pool)
+			rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
+	}
+	up_read(&rds_ib_devices_lock);
+}
+
+u32 rds_ib_get_lkey(void *trans_private)
+{
+	struct rds_ib_mr *ibmr = trans_private;
+
+	return ibmr->u.mr->lkey;
+}
+
+void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
+		    struct rds_sock *rs, u32 *key_ret,
+		    struct rds_connection *conn,
+		    u64 start, u64 length, int need_odp)
+{
+	struct rds_ib_device *rds_ibdev;
+	struct rds_ib_mr *ibmr = NULL;
+	struct rds_ib_connection *ic = NULL;
+	int ret;
+
+	rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]);
+	if (!rds_ibdev) {
+		ret = -ENODEV;
+		goto out;
+	}
+
+	if (need_odp == ODP_ZEROBASED || need_odp == ODP_VIRTUAL) {
+		u64 virt_addr = need_odp == ODP_ZEROBASED ? 0 : start;
+		int access_flags =
+			(IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
+			 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_ATOMIC |
+			 IB_ACCESS_ON_DEMAND);
+		struct ib_sge sge = {};
+		struct ib_mr *ib_mr;
+
+		if (!rds_ibdev->odp_capable) {
+			ret = -EOPNOTSUPP;
+			goto out;
+		}
+
+		ib_mr = ib_reg_user_mr(rds_ibdev->pd, start, length, virt_addr,
+				       access_flags);
+
+		if (IS_ERR(ib_mr)) {
+			rdsdebug("rds_ib_get_user_mr returned %d\n",
+				 IS_ERR(ib_mr));
+			ret = PTR_ERR(ib_mr);
+			goto out;
+		}
+		if (key_ret)
+			*key_ret = ib_mr->rkey;
+
+		ibmr = kzalloc_obj(*ibmr);
+		if (!ibmr) {
+			ib_dereg_mr(ib_mr);
+			ret = -ENOMEM;
+			goto out;
+		}
+		ibmr->u.mr = ib_mr;
+		ibmr->odp = 1;
+
+		sge.addr = virt_addr;
+		sge.length = length;
+		sge.lkey = ib_mr->lkey;
+
+		ib_advise_mr(rds_ibdev->pd,
+			     IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE,
+			     IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sge, 1);
+		return ibmr;
+	}
+
+	if (conn) {
+		ic = conn->c_transport_data;
+		if (!ic || !ic->i_cm_id || !ic->i_cm_id->qp) {
+			ret = -ENODEV;
+			goto out;
+		}
+	}
+
+	if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
+		ret = -ENODEV;
+		goto out;
+	}
+
+	ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
+	if (IS_ERR(ibmr)) {
+		ret = PTR_ERR(ibmr);
+		pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
+	} else {
+		return ibmr;
+	}
+
+ out:
+	if (rds_ibdev)
+		rds_ib_dev_put(rds_ibdev);
+
+	return ERR_PTR(ret);
+}
+
+void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
+{
+	cancel_delayed_work_sync(&pool->flush_worker);
+	rds_ib_flush_mr_pool(pool, 1, NULL);
+	WARN_ON(atomic_read(&pool->item_count));
+	WARN_ON(atomic_read(&pool->free_pinned));
+	kfree(pool);
+}
+
+struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
+					     int pool_type)
+{
+	struct rds_ib_mr_pool *pool;
+
+	pool = kzalloc_obj(*pool);
+	if (!pool)
+		return ERR_PTR(-ENOMEM);
+
+	pool->pool_type = pool_type;
+	init_llist_head(&pool->free_list);
+	init_llist_head(&pool->drop_list);
+	init_llist_head(&pool->clean_list);
+	spin_lock_init(&pool->clean_lock);
+	mutex_init(&pool->flush_lock);
+	init_waitqueue_head(&pool->flush_wait);
+	INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
+
+	if (pool_type == RDS_IB_MR_1M_POOL) {
+		/* +1 allows for unaligned MRs */
+		pool->max_pages = RDS_MR_1M_MSG_SIZE + 1;
+		pool->max_items = rds_ibdev->max_1m_mrs;
+	} else {
+		/* pool_type == RDS_IB_MR_8K_POOL */
+		pool->max_pages = RDS_MR_8K_MSG_SIZE + 1;
+		pool->max_items = rds_ibdev->max_8k_mrs;
+	}
+
+	pool->max_free_pinned = pool->max_items * pool->max_pages / 4;
+	pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
+
+	return pool;
+}
+
+int rds_ib_mr_init(void)
+{
+	rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd",
+				       WQ_MEM_RECLAIM | WQ_PERCPU, 0);
+	if (!rds_ib_mr_wq)
+		return -ENOMEM;
+	return 0;
+}
+
+/* By the time this is called all the IB devices should have been torn down and
+ * had their pools freed.  As each pool is freed its work struct is waited on,
+ * so the pool flushing work queue should be idle by the time we get here.
+ */
+void rds_ib_mr_exit(void)
+{
+	destroy_workqueue(rds_ib_mr_wq);
+}
+
+static void rds_ib_odp_mr_worker(struct work_struct  *work)
+{
+	struct rds_ib_mr *ibmr;
+
+	ibmr = container_of(work, struct rds_ib_mr, work.work);
+	ib_dereg_mr(ibmr->u.mr);
+	kfree(ibmr);
+}

linux-7.2.3/net/rds/ib_ring.c

@@ -0,0 +1,168 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+
+#include "rds.h"
+#include "ib.h"
+
+/*
+ * Locking for IB rings.
+ * We assume that allocation is always protected by a mutex
+ * in the caller (this is a valid assumption for the current
+ * implementation).
+ *
+ * Freeing always happens in an interrupt, and hence only
+ * races with allocations, but not with other free()s.
+ *
+ * The interaction between allocation and freeing is that
+ * the alloc code has to determine the number of free entries.
+ * To this end, we maintain two counters; an allocation counter
+ * and a free counter. Both are allowed to run freely, and wrap
+ * around.
+ * The number of used entries is always (alloc_ctr - free_ctr) % NR.
+ *
+ * The current implementation makes free_ctr atomic. When the
+ * caller finds an allocation fails, it should set an "alloc fail"
+ * bit and retry the allocation. The "alloc fail" bit essentially tells
+ * the CQ completion handlers to wake it up after freeing some
+ * more entries.
+ */
+
+/*
+ * This only happens on shutdown.
+ */
+DECLARE_WAIT_QUEUE_HEAD(rds_ib_ring_empty_wait);
+
+void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr)
+{
+	memset(ring, 0, sizeof(*ring));
+	ring->w_nr = nr;
+	rdsdebug("ring %p nr %u\n", ring, ring->w_nr);
+}
+
+static inline u32 __rds_ib_ring_used(struct rds_ib_work_ring *ring)
+{
+	u32 diff;
+
+	/* This assumes that atomic_t has at least as many bits as u32 */
+	diff = ring->w_alloc_ctr - (u32) atomic_read(&ring->w_free_ctr);
+	BUG_ON(diff > ring->w_nr);
+
+	return diff;
+}
+
+void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr)
+{
+	/* We only ever get called from the connection setup code,
+	 * prior to creating the QP. */
+	BUG_ON(__rds_ib_ring_used(ring));
+	ring->w_nr = nr;
+}
+
+static int __rds_ib_ring_empty(struct rds_ib_work_ring *ring)
+{
+	return __rds_ib_ring_used(ring) == 0;
+}
+
+u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos)
+{
+	u32 ret = 0, avail;
+
+	avail = ring->w_nr - __rds_ib_ring_used(ring);
+
+	rdsdebug("ring %p val %u next %u free %u\n", ring, val,
+		 ring->w_alloc_ptr, avail);
+
+	if (val && avail) {
+		ret = min(val, avail);
+		*pos = ring->w_alloc_ptr;
+
+		ring->w_alloc_ptr = (ring->w_alloc_ptr + ret) % ring->w_nr;
+		ring->w_alloc_ctr += ret;
+	}
+
+	return ret;
+}
+
+void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val)
+{
+	ring->w_free_ptr = (ring->w_free_ptr + val) % ring->w_nr;
+	atomic_add(val, &ring->w_free_ctr);
+
+	if (__rds_ib_ring_empty(ring) &&
+	    waitqueue_active(&rds_ib_ring_empty_wait))
+		wake_up(&rds_ib_ring_empty_wait);
+}
+
+void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val)
+{
+	ring->w_alloc_ptr = (ring->w_alloc_ptr - val) % ring->w_nr;
+	ring->w_alloc_ctr -= val;
+}
+
+int rds_ib_ring_empty(struct rds_ib_work_ring *ring)
+{
+	return __rds_ib_ring_empty(ring);
+}
+
+int rds_ib_ring_low(struct rds_ib_work_ring *ring)
+{
+	return __rds_ib_ring_used(ring) <= (ring->w_nr >> 1);
+}
+
+/*
+ * returns the oldest allocated ring entry.  This will be the next one
+ * freed.  This can't be called if there are none allocated.
+ */
+u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring)
+{
+	return ring->w_free_ptr;
+}
+
+/*
+ * returns the number of completed work requests.
+ */
+
+u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest)
+{
+	u32 ret;
+
+	if (oldest <= (unsigned long long)wr_id)
+		ret = (unsigned long long)wr_id - oldest + 1;
+	else
+		ret = ring->w_nr - oldest + (unsigned long long)wr_id + 1;
+
+	rdsdebug("ring %p ret %u wr_id %u oldest %u\n", ring, ret,
+		 wr_id, oldest);
+	return ret;
+}

linux-7.2.3/net/rds/ib_stats.c

@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/percpu.h>
+#include <linux/seq_file.h>
+#include <linux/proc_fs.h>
+
+#include "rds.h"
+#include "ib.h"
+
+DEFINE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics, rds_ib_stats);
+
+static const char *const rds_ib_stat_names[] = {
+	"ib_connect_raced",
+	"ib_listen_closed_stale",
+	"ib_evt_handler_call",
+	"ib_tasklet_call",
+	"ib_tx_cq_event",
+	"ib_tx_ring_full",
+	"ib_tx_throttle",
+	"ib_tx_sg_mapping_failure",
+	"ib_tx_stalled",
+	"ib_tx_credit_updates",
+	"ib_rx_cq_event",
+	"ib_rx_ring_empty",
+	"ib_rx_refill_from_cq",
+	"ib_rx_refill_from_thread",
+	"ib_rx_alloc_limit",
+	"ib_rx_total_frags",
+	"ib_rx_total_incs",
+	"ib_rx_credit_updates",
+	"ib_ack_sent",
+	"ib_ack_send_failure",
+	"ib_ack_send_delayed",
+	"ib_ack_send_piggybacked",
+	"ib_ack_received",
+	"ib_rdma_mr_8k_alloc",
+	"ib_rdma_mr_8k_free",
+	"ib_rdma_mr_8k_used",
+	"ib_rdma_mr_8k_pool_flush",
+	"ib_rdma_mr_8k_pool_wait",
+	"ib_rdma_mr_8k_pool_depleted",
+	"ib_rdma_mr_1m_alloc",
+	"ib_rdma_mr_1m_free",
+	"ib_rdma_mr_1m_used",
+	"ib_rdma_mr_1m_pool_flush",
+	"ib_rdma_mr_1m_pool_wait",
+	"ib_rdma_mr_1m_pool_depleted",
+	"ib_rdma_mr_8k_reused",
+	"ib_rdma_mr_1m_reused",
+	"ib_atomic_cswp",
+	"ib_atomic_fadd",
+};
+
+unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
+				    unsigned int avail)
+{
+	struct rds_ib_statistics stats = {0, };
+	uint64_t *src;
+	uint64_t *sum;
+	size_t i;
+	int cpu;
+
+	if (avail < ARRAY_SIZE(rds_ib_stat_names))
+		goto out;
+
+	for_each_online_cpu(cpu) {
+		src = (uint64_t *)&(per_cpu(rds_ib_stats, cpu));
+		sum = (uint64_t *)&stats;
+		for (i = 0; i < sizeof(stats) / sizeof(uint64_t); i++)
+			*(sum++) += *(src++);
+	}
+
+	rds_stats_info_copy(iter, (uint64_t *)&stats, rds_ib_stat_names,
+			    ARRAY_SIZE(rds_ib_stat_names));
+out:
+	return ARRAY_SIZE(rds_ib_stat_names);
+}

linux-7.2.3/net/rds/ib_sysctl.c

@@ -0,0 +1,120 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/sysctl.h>
+#include <linux/proc_fs.h>
+
+#include "ib.h"
+
+static struct ctl_table_header *rds_ib_sysctl_hdr;
+
+unsigned long rds_ib_sysctl_max_send_wr = RDS_IB_DEFAULT_SEND_WR;
+unsigned long rds_ib_sysctl_max_recv_wr = RDS_IB_DEFAULT_RECV_WR;
+unsigned long rds_ib_sysctl_max_recv_allocation = (128 * 1024 * 1024) / RDS_FRAG_SIZE;
+static unsigned long rds_ib_sysctl_max_wr_min = 1;
+/* hardware will fail CQ creation long before this */
+static unsigned long rds_ib_sysctl_max_wr_max = (u32)~0;
+
+unsigned long rds_ib_sysctl_max_unsig_wrs = 16;
+static unsigned long rds_ib_sysctl_max_unsig_wr_min = 1;
+static unsigned long rds_ib_sysctl_max_unsig_wr_max = 64;
+
+/*
+ * This sysctl does nothing.
+ *
+ * Backwards compatibility with RDS 3.0 wire protocol
+ * disables initial FC credit exchange.
+ * If it's ever possible to drop 3.0 support,
+ * setting this to 1 and moving init/refill of send/recv
+ * rings from ib_cm_connect_complete() back into ib_setup_qp()
+ * will cause credits to be added before protocol negotiation.
+ */
+unsigned int rds_ib_sysctl_flow_control = 0;
+
+static struct ctl_table rds_ib_sysctl_table[] = {
+	{
+		.procname       = "max_send_wr",
+		.data		= &rds_ib_sysctl_max_send_wr,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = proc_doulongvec_minmax,
+		.extra1		= &rds_ib_sysctl_max_wr_min,
+		.extra2		= &rds_ib_sysctl_max_wr_max,
+	},
+	{
+		.procname       = "max_recv_wr",
+		.data		= &rds_ib_sysctl_max_recv_wr,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = proc_doulongvec_minmax,
+		.extra1		= &rds_ib_sysctl_max_wr_min,
+		.extra2		= &rds_ib_sysctl_max_wr_max,
+	},
+	{
+		.procname       = "max_unsignaled_wr",
+		.data		= &rds_ib_sysctl_max_unsig_wrs,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = proc_doulongvec_minmax,
+		.extra1		= &rds_ib_sysctl_max_unsig_wr_min,
+		.extra2		= &rds_ib_sysctl_max_unsig_wr_max,
+	},
+	{
+		.procname       = "max_recv_allocation",
+		.data		= &rds_ib_sysctl_max_recv_allocation,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = proc_doulongvec_minmax,
+	},
+	{
+		.procname	= "flow_control",
+		.data		= &rds_ib_sysctl_flow_control,
+		.maxlen		= sizeof(rds_ib_sysctl_flow_control),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec,
+	},
+};
+
+void rds_ib_sysctl_exit(void)
+{
+	if (rds_ib_sysctl_hdr)
+		unregister_net_sysctl_table(rds_ib_sysctl_hdr);
+}
+
+int rds_ib_sysctl_init(void)
+{
+	rds_ib_sysctl_hdr = register_net_sysctl(&init_net, "net/rds/ib", rds_ib_sysctl_table);
+	if (!rds_ib_sysctl_hdr)
+		return -ENOMEM;
+	return 0;
+}

linux-7.2.3/net/rds/info.c

@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/percpu.h>
+#include <linux/seq_file.h>
+#include <linux/slab.h>
+#include <linux/proc_fs.h>
+#include <linux/export.h>
+#include <linux/uio.h>
+
+#include "rds.h"
+
+/*
+ * This file implements a getsockopt() call which copies a set of fixed
+ * sized structs into a user-specified buffer as a means of providing
+ * read-only information about RDS.
+ *
+ * For a given information source there are a given number of fixed sized
+ * structs at a given time.  The structs are only copied if the user-specified
+ * buffer is big enough.  The destination pages that make up the buffer
+ * are pinned for the duration of the copy.
+ *
+ * This gives us the following benefits:
+ *
+ * - simple implementation, no copy "position" across multiple calls
+ * - consistent snapshot of an info source
+ * - atomic copy works well with whatever locking info source has
+ * - one portable tool to get rds info across implementations
+ * - long-lived tool can get info without allocating
+ *
+ * at the following costs:
+ *
+ * - info source copy must be pinned, may be "large"
+ */
+
+struct rds_info_iterator {
+	struct page **pages;
+	void *addr;
+	unsigned long offset;
+};
+
+static DEFINE_SPINLOCK(rds_info_lock);
+static rds_info_func rds_info_funcs[RDS_INFO_LAST - RDS_INFO_FIRST + 1];
+
+void rds_info_register_func(int optname, rds_info_func func)
+{
+	int offset = optname - RDS_INFO_FIRST;
+
+	BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
+
+	spin_lock(&rds_info_lock);
+	BUG_ON(rds_info_funcs[offset]);
+	rds_info_funcs[offset] = func;
+	spin_unlock(&rds_info_lock);
+}
+EXPORT_SYMBOL_GPL(rds_info_register_func);
+
+void rds_info_deregister_func(int optname, rds_info_func func)
+{
+	int offset = optname - RDS_INFO_FIRST;
+
+	BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
+
+	spin_lock(&rds_info_lock);
+	BUG_ON(rds_info_funcs[offset] != func);
+	rds_info_funcs[offset] = NULL;
+	spin_unlock(&rds_info_lock);
+}
+EXPORT_SYMBOL_GPL(rds_info_deregister_func);
+
+/*
+ * Typically we hold an atomic kmap across multiple rds_info_copy() calls
+ * because the kmap is so expensive.  This must be called before using blocking
+ * operations while holding the mapping and as the iterator is torn down.
+ */
+void rds_info_iter_unmap(struct rds_info_iterator *iter)
+{
+	if (iter->addr) {
+		kunmap_atomic(iter->addr);
+		iter->addr = NULL;
+	}
+}
+
+/*
+ * get_user_pages() called flush_dcache_page() on the pages for us.
+ */
+void rds_info_copy(struct rds_info_iterator *iter, void *data,
+		   unsigned long bytes)
+{
+	unsigned long this;
+
+	while (bytes) {
+		if (!iter->addr)
+			iter->addr = kmap_atomic(*iter->pages);
+
+		this = min(bytes, PAGE_SIZE - iter->offset);
+
+		rdsdebug("page %p addr %p offset %lu this %lu data %p "
+			  "bytes %lu\n", *iter->pages, iter->addr,
+			  iter->offset, this, data, bytes);
+
+		memcpy(iter->addr + iter->offset, data, this);
+
+		data += this;
+		bytes -= this;
+		iter->offset += this;
+
+		if (iter->offset == PAGE_SIZE) {
+			kunmap_atomic(iter->addr);
+			iter->addr = NULL;
+			iter->offset = 0;
+			iter->pages++;
+		}
+	}
+}
+EXPORT_SYMBOL_GPL(rds_info_copy);
+
+/*
+ * @opt->iter_out describes the buffer that the information snapshot will be
+ * copied into, and @opt->optlen is the size of that buffer on input.  On
+ * output @opt->optlen is set to the size of the requested snapshot in bytes.
+ *
+ * This function returns -errno if there is a failure, particularly -ENOSPC
+ * if the given buffer was not large enough to fit the snapshot.  On success
+ * it returns the positive number of bytes of each array element in the
+ * snapshot.
+ */
+int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
+{
+	struct rds_info_iterator iter;
+	struct rds_info_lengths lens;
+	unsigned long nr_pages = 0;
+	rds_info_func func;
+	struct page **pages = NULL;
+	size_t offset0 = 0;
+	int npages = 0;
+	int ret;
+	int len;
+	int total;
+
+	len = opt->optlen;
+
+	/* check for all kinds of wrapping and the like */
+	if (len < 0 || len > INT_MAX - PAGE_SIZE + 1) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	/* The info producers write into the pages with kmap_atomic() while
+	 * holding a spinlock, so they need a genuine page-backed user buffer.
+	 */
+	if (!user_backed_iter(&opt->iter_out)) {
+		ret = -EOPNOTSUPP;
+		goto out;
+	}
+
+	/* a 0 len call is just trying to probe its length */
+	if (len == 0)
+		goto call_func;
+
+	/*
+	 * Preallocate the page array and pass it in so that
+	 * iov_iter_extract_pages() fills it in place rather than allocating
+	 * one for us.  Handing it a non-NULL array keeps ownership of the
+	 * array with us on every return path, instead of depending on the
+	 * iterator code to allocate and hand it back.
+	 */
+	npages = iov_iter_npages(&opt->iter_out, INT_MAX);
+	pages = kvmalloc_array(npages, sizeof(*pages), GFP_KERNEL);
+	if (!pages) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	ret = iov_iter_extract_pages(&opt->iter_out, &pages, len, npages,
+				     0, &offset0);
+	if (ret < 0)
+		goto out;
+	nr_pages = DIV_ROUND_UP(offset0 + ret, PAGE_SIZE);
+	if (ret != len) {
+		ret = -EAGAIN; /* XXX ? */
+		goto out;
+	}
+
+	rdsdebug("len %d nr_pages %lu\n", len, nr_pages);
+
+call_func:
+	func = rds_info_funcs[optname - RDS_INFO_FIRST];
+	if (!func) {
+		ret = -ENOPROTOOPT;
+		goto out;
+	}
+
+	iter.pages = pages;
+	iter.addr = NULL;
+	iter.offset = offset0;
+
+	func(sock, len, &iter, &lens);
+	BUG_ON(lens.each == 0);
+
+	total = lens.nr * lens.each;
+
+	rds_info_iter_unmap(&iter);
+
+	if (total > len) {
+		len = total;
+		ret = -ENOSPC;
+	} else {
+		len = total;
+		ret = lens.each;
+	}
+
+	opt->optlen = len;
+
+out:
+	/*
+	 * iov_iter_extract_pages() pins only user-backed (ubuf) iters;
+	 * iov_iter_extract_will_pin() reports whether an unpin is owed here.
+	 */
+	if (pages && iov_iter_extract_will_pin(&opt->iter_out))
+		unpin_user_pages_dirty_lock(pages, nr_pages, true);
+	kvfree(pages);
+
+	return ret;
+}

linux-7.2.3/net/rds/info.h

@@ -0,0 +1,30 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _RDS_INFO_H
+#define _RDS_INFO_H
+
+struct rds_info_lengths {
+	unsigned int	nr;
+	unsigned int	each;
+};
+
+struct rds_info_iterator;
+
+/*
+ * These functions must fill in the fields of @lens to reflect the size
+ * of the available info source.  If the snapshot fits in @len then it
+ * should be copied using @iter.  The caller will deduce if it was copied
+ * or not by comparing the lengths.
+ */
+typedef void (*rds_info_func)(struct socket *sock, unsigned int len,
+			      struct rds_info_iterator *iter,
+			      struct rds_info_lengths *lens);
+
+void rds_info_register_func(int optname, rds_info_func func);
+void rds_info_deregister_func(int optname, rds_info_func func);
+int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt);
+void rds_info_copy(struct rds_info_iterator *iter, void *data,
+		   unsigned long bytes);
+void rds_info_iter_unmap(struct rds_info_iterator *iter);
+
+
+#endif

linux-7.2.3/net/rds/loop.c

@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/in.h>
+#include <net/net_namespace.h>
+#include <net/netns/generic.h>
+#include <linux/ipv6.h>
+
+#include "rds_single_path.h"
+#include "rds.h"
+#include "loop.h"
+
+static DEFINE_SPINLOCK(loop_conns_lock);
+static LIST_HEAD(loop_conns);
+static atomic_t rds_loop_unloading = ATOMIC_INIT(0);
+
+static void rds_loop_set_unloading(void)
+{
+	atomic_set(&rds_loop_unloading, 1);
+}
+
+static bool rds_loop_is_unloading(struct rds_connection *conn)
+{
+	return atomic_read(&rds_loop_unloading) != 0;
+}
+
+/*
+ * This 'loopback' transport is a special case for flows that originate
+ * and terminate on the same machine.
+ *
+ * Connection build-up notices if the destination address is thought of
+ * as a local address by a transport.  At that time it decides to use the
+ * loopback transport instead of the bound transport of the sending socket.
+ *
+ * The loopback transport's sending path just hands the sent rds_message
+ * straight to the receiving path via an embedded rds_incoming.
+ */
+
+/*
+ * Usually a message transits both the sender and receiver's conns as it
+ * flows to the receiver.  In the loopback case, though, the receive path
+ * is handed the sending conn so the sense of the addresses is reversed.
+ */
+static int rds_loop_xmit(struct rds_connection *conn, struct rds_message *rm,
+			 unsigned int hdr_off, unsigned int sg,
+			 unsigned int off)
+{
+	struct scatterlist *sgp = &rm->data.op_sg[sg];
+	int ret = sizeof(struct rds_header) +
+			be32_to_cpu(rm->m_inc.i_hdr.h_len);
+
+	/* Do not send cong updates to loopback */
+	if (rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
+		rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
+		ret = min_t(int, ret, sgp->length - conn->c_xmit_data_off);
+		goto out;
+	}
+
+	BUG_ON(hdr_off || sg || off);
+
+	rds_inc_init(&rm->m_inc, conn, &conn->c_laddr);
+	/* For the embedded inc. Matching put is in loop_inc_free() */
+	rds_message_addref(rm);
+
+	rds_recv_incoming(conn, &conn->c_laddr, &conn->c_faddr, &rm->m_inc,
+			  GFP_KERNEL);
+
+	rds_send_drop_acked(conn, be64_to_cpu(rm->m_inc.i_hdr.h_sequence),
+			    NULL);
+
+	rds_inc_put(&rm->m_inc);
+out:
+	return ret;
+}
+
+/*
+ * See rds_loop_xmit(). Since our inc is embedded in the rm, we
+ * make sure the rm lives at least until the inc is done.
+ */
+static void rds_loop_inc_free(struct rds_incoming *inc)
+{
+	struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
+
+	rds_message_put(rm);
+}
+
+/* we need to at least give the thread something to succeed */
+static int rds_loop_recv_path(struct rds_conn_path *cp)
+{
+	return 0;
+}
+
+struct rds_loop_connection {
+	struct list_head loop_node;
+	struct rds_connection *conn;
+};
+
+/*
+ * Even the loopback transport needs to keep track of its connections,
+ * so it can call rds_conn_destroy() on them on exit. N.B. there are
+ * 1+ loopback addresses (127.*.*.*) so it's not a bug to have
+ * multiple loopback conns allocated, although rather useless.
+ */
+static int rds_loop_conn_alloc(struct rds_connection *conn, gfp_t gfp)
+{
+	struct rds_loop_connection *lc;
+	unsigned long flags;
+
+	lc = kzalloc_obj(struct rds_loop_connection, gfp);
+	if (!lc)
+		return -ENOMEM;
+
+	INIT_LIST_HEAD(&lc->loop_node);
+	lc->conn = conn;
+	conn->c_transport_data = lc;
+
+	spin_lock_irqsave(&loop_conns_lock, flags);
+	list_add_tail(&lc->loop_node, &loop_conns);
+	spin_unlock_irqrestore(&loop_conns_lock, flags);
+
+	return 0;
+}
+
+static void rds_loop_conn_free(void *arg)
+{
+	struct rds_loop_connection *lc = arg;
+	unsigned long flags;
+
+	rdsdebug("lc %p\n", lc);
+	spin_lock_irqsave(&loop_conns_lock, flags);
+	list_del(&lc->loop_node);
+	spin_unlock_irqrestore(&loop_conns_lock, flags);
+	kfree(lc);
+}
+
+static int rds_loop_conn_path_connect(struct rds_conn_path *cp)
+{
+	rds_connect_complete(cp->cp_conn);
+	return 0;
+}
+
+static void rds_loop_conn_path_shutdown(struct rds_conn_path *cp)
+{
+}
+
+void rds_loop_exit(void)
+{
+	struct rds_loop_connection *lc, *_lc;
+	LIST_HEAD(tmp_list);
+
+	rds_loop_set_unloading();
+	synchronize_rcu();
+	/* avoid calling conn_destroy with irqs off */
+	spin_lock_irq(&loop_conns_lock);
+	list_splice(&loop_conns, &tmp_list);
+	INIT_LIST_HEAD(&loop_conns);
+	spin_unlock_irq(&loop_conns_lock);
+
+	list_for_each_entry_safe(lc, _lc, &tmp_list, loop_node) {
+		WARN_ON(lc->conn->c_passive);
+		rds_conn_destroy(lc->conn);
+	}
+}
+
+static void rds_loop_kill_conns(struct net *net)
+{
+	struct rds_loop_connection *lc, *_lc;
+	LIST_HEAD(tmp_list);
+
+	spin_lock_irq(&loop_conns_lock);
+	list_for_each_entry_safe(lc, _lc, &loop_conns, loop_node)  {
+		struct net *c_net = read_pnet(&lc->conn->c_net);
+
+		if (net != c_net)
+			continue;
+		list_move_tail(&lc->loop_node, &tmp_list);
+	}
+	spin_unlock_irq(&loop_conns_lock);
+
+	list_for_each_entry_safe(lc, _lc, &tmp_list, loop_node) {
+		WARN_ON(lc->conn->c_passive);
+		rds_conn_destroy(lc->conn);
+	}
+}
+
+static void __net_exit rds_loop_exit_net(struct net *net)
+{
+	rds_loop_kill_conns(net);
+}
+
+static struct pernet_operations rds_loop_net_ops = {
+	.exit = rds_loop_exit_net,
+};
+
+int rds_loop_net_init(void)
+{
+	return register_pernet_device(&rds_loop_net_ops);
+}
+
+void rds_loop_net_exit(void)
+{
+	unregister_pernet_device(&rds_loop_net_ops);
+}
+
+/*
+ * This is missing .xmit_* because loop doesn't go through generic
+ * rds_send_xmit() and doesn't call rds_recv_incoming().  .listen_stop and
+ * .laddr_check are missing because transport.c doesn't iterate over
+ * rds_loop_transport.
+ */
+struct rds_transport rds_loop_transport = {
+	.xmit			= rds_loop_xmit,
+	.recv_path		= rds_loop_recv_path,
+	.conn_alloc		= rds_loop_conn_alloc,
+	.conn_free		= rds_loop_conn_free,
+	.conn_path_connect	= rds_loop_conn_path_connect,
+	.conn_path_shutdown	= rds_loop_conn_path_shutdown,
+	.inc_copy_to_user	= rds_message_inc_copy_to_user,
+	.inc_free		= rds_loop_inc_free,
+	.t_name			= "loopback",
+	.t_type			= RDS_TRANS_LOOP,
+	.t_unloading		= rds_loop_is_unloading,
+};

linux-7.2.3/net/rds/loop.h

@@ -0,0 +1,12 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _RDS_LOOP_H
+#define _RDS_LOOP_H
+
+/* loop.c */
+extern struct rds_transport rds_loop_transport;
+
+int rds_loop_net_init(void);
+void rds_loop_net_exit(void);
+void rds_loop_exit(void);
+
+#endif

linux-7.2.3/net/rds/message.c

@@ -0,0 +1,570 @@
+/*
+ * Copyright (c) 2006, 2020 Oracle and/or its affiliates.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/export.h>
+#include <linux/skbuff.h>
+#include <linux/list.h>
+#include <linux/errqueue.h>
+
+#include "rds.h"
+
+static unsigned int	rds_exthdr_size[__RDS_EXTHDR_MAX] = {
+[RDS_EXTHDR_NONE]	= 0,
+[RDS_EXTHDR_VERSION]	= sizeof(struct rds_ext_header_version),
+[RDS_EXTHDR_RDMA]	= sizeof(struct rds_ext_header_rdma),
+[RDS_EXTHDR_RDMA_DEST]	= sizeof(struct rds_ext_header_rdma_dest),
+[RDS_EXTHDR_RDMA_BYTES] = sizeof(struct rds_ext_header_rdma_bytes),
+[RDS_EXTHDR_NPATHS]	= sizeof(__be16),
+[RDS_EXTHDR_GEN_NUM]	= sizeof(__be32),
+[RDS_EXTHDR_SPORT_IDX]	= 1,
+};
+
+void rds_message_addref(struct rds_message *rm)
+{
+	rdsdebug("addref rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
+	refcount_inc(&rm->m_refcount);
+}
+EXPORT_SYMBOL_GPL(rds_message_addref);
+
+static inline bool rds_zcookie_add(struct rds_msg_zcopy_info *info, u32 cookie)
+{
+	struct rds_zcopy_cookies *ck = &info->zcookies;
+	int ncookies = ck->num;
+
+	if (ncookies == RDS_MAX_ZCOOKIES)
+		return false;
+	ck->cookies[ncookies] = cookie;
+	ck->num =  ++ncookies;
+	return true;
+}
+
+static struct rds_msg_zcopy_info *rds_info_from_znotifier(struct rds_znotifier *znotif)
+{
+	return container_of(znotif, struct rds_msg_zcopy_info, znotif);
+}
+
+void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *q)
+{
+	unsigned long flags;
+	LIST_HEAD(copy);
+	struct rds_msg_zcopy_info *info, *tmp;
+
+	spin_lock_irqsave(&q->lock, flags);
+	list_splice(&q->zcookie_head, &copy);
+	INIT_LIST_HEAD(&q->zcookie_head);
+	spin_unlock_irqrestore(&q->lock, flags);
+
+	list_for_each_entry_safe(info, tmp, &copy, rs_zcookie_next) {
+		list_del(&info->rs_zcookie_next);
+		kfree(info);
+	}
+}
+
+static void rds_rm_zerocopy_callback(struct rds_sock *rs,
+				     struct rds_znotifier *znotif)
+{
+	struct rds_msg_zcopy_info *info;
+	struct rds_msg_zcopy_queue *q;
+	u32 cookie = znotif->z_cookie;
+	struct rds_zcopy_cookies *ck;
+	struct list_head *head;
+	unsigned long flags;
+
+	mm_unaccount_pinned_pages(&znotif->z_mmp);
+	q = &rs->rs_zcookie_queue;
+	spin_lock_irqsave(&q->lock, flags);
+	head = &q->zcookie_head;
+	if (!list_empty(head)) {
+		info = list_first_entry(head, struct rds_msg_zcopy_info,
+					rs_zcookie_next);
+		if (rds_zcookie_add(info, cookie)) {
+			spin_unlock_irqrestore(&q->lock, flags);
+			kfree(rds_info_from_znotifier(znotif));
+			/* caller invokes rds_wake_sk_sleep() */
+			return;
+		}
+	}
+
+	info = rds_info_from_znotifier(znotif);
+	ck = &info->zcookies;
+	memset(ck, 0, sizeof(*ck));
+	WARN_ON(!rds_zcookie_add(info, cookie));
+	list_add_tail(&info->rs_zcookie_next, &q->zcookie_head);
+
+	spin_unlock_irqrestore(&q->lock, flags);
+	/* caller invokes rds_wake_sk_sleep() */
+}
+
+/*
+ * This relies on dma_map_sg() not touching sg[].page during merging.
+ */
+static void rds_message_purge(struct rds_message *rm)
+{
+	struct rds_znotifier *znotifier;
+	unsigned long i, flags;
+	bool zcopy;
+
+	if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
+		return;
+
+	spin_lock_irqsave(&rm->m_rs_lock, flags);
+	znotifier = rm->data.op_mmp_znotifier;
+	rm->data.op_mmp_znotifier = NULL;
+	zcopy = !!znotifier;
+
+	if (rm->m_rs) {
+		struct rds_sock *rs = rm->m_rs;
+
+		if (znotifier) {
+			rds_rm_zerocopy_callback(rs, znotifier);
+			rds_wake_sk_sleep(rs);
+		}
+		sock_put(rds_rs_to_sk(rs));
+		rm->m_rs = NULL;
+	} else if (znotifier) {
+		/*
+		 * Zerocopy can fail before the message is queued on the
+		 * socket, so there is no rs to carry the notification.
+		 */
+		mm_unaccount_pinned_pages(&znotifier->z_mmp);
+		kfree(rds_info_from_znotifier(znotifier));
+	}
+	spin_unlock_irqrestore(&rm->m_rs_lock, flags);
+
+	for (i = 0; i < rm->data.op_nents; i++) {
+		/* XXX will have to put_page for page refs */
+		if (!zcopy)
+			__free_page(sg_page(&rm->data.op_sg[i]));
+		else
+			put_page(sg_page(&rm->data.op_sg[i]));
+	}
+	rm->data.op_nents = 0;
+
+	if (rm->rdma.op_active)
+		rds_rdma_free_op(&rm->rdma);
+	if (rm->rdma.op_rdma_mr)
+		kref_put(&rm->rdma.op_rdma_mr->r_kref, __rds_put_mr_final);
+
+	if (rm->atomic.op_active)
+		rds_atomic_free_op(&rm->atomic);
+	if (rm->atomic.op_rdma_mr)
+		kref_put(&rm->atomic.op_rdma_mr->r_kref, __rds_put_mr_final);
+}
+
+void rds_message_put(struct rds_message *rm)
+{
+	rdsdebug("put rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
+	WARN(!refcount_read(&rm->m_refcount), "danger refcount zero on %p\n", rm);
+	if (refcount_dec_and_test(&rm->m_refcount)) {
+		BUG_ON(!list_empty(&rm->m_sock_item));
+		BUG_ON(!list_empty(&rm->m_conn_item));
+		rds_message_purge(rm);
+
+		kfree(rm);
+	}
+}
+EXPORT_SYMBOL_GPL(rds_message_put);
+
+void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
+				 __be16 dport, u64 seq)
+{
+	hdr->h_flags = 0;
+	hdr->h_sport = sport;
+	hdr->h_dport = dport;
+	hdr->h_sequence = cpu_to_be64(seq);
+	/* see rds_find_next_ext_space for reason why we memset the
+	 * ext header
+	 */
+	memset(hdr->h_exthdr, RDS_EXTHDR_NONE, RDS_HEADER_EXT_SPACE);
+}
+EXPORT_SYMBOL_GPL(rds_message_populate_header);
+
+/*
+ * Find the next place we can add an RDS header extension with
+ * specific length. Extension headers are pushed one after the
+ * other. In the following, the number after the colon is the number
+ * of bytes:
+ *
+ * [ type1:1 dta1:len1 [ type2:1 dta2:len2 ] ... ] RDS_EXTHDR_NONE
+ *
+ * If the extension headers fill the complete extension header space
+ * (16 bytes), the trailing RDS_EXTHDR_NONE is omitted.
+ */
+static int rds_find_next_ext_space(struct rds_header *hdr, unsigned int len,
+				   u8 **ext_start)
+{
+	unsigned int ext_len;
+	unsigned int type;
+	int ind = 0;
+
+	while ((ind + 1 + len) <= RDS_HEADER_EXT_SPACE) {
+		if (hdr->h_exthdr[ind] == RDS_EXTHDR_NONE) {
+			*ext_start = hdr->h_exthdr + ind;
+			return 0;
+		}
+
+		type = hdr->h_exthdr[ind];
+
+		ext_len = (type < __RDS_EXTHDR_MAX) ? rds_exthdr_size[type] : 0;
+		WARN_ONCE(!ext_len, "Unknown ext hdr type %d\n", type);
+		if (!ext_len)
+			return -EINVAL;
+
+		/* ind points to a valid ext hdr with known length */
+		ind += 1 + ext_len;
+	}
+
+	/* no room for extension */
+	return -ENOSPC;
+}
+
+/* The ext hdr space is prefilled with zero from the kzalloc() */
+int rds_message_add_extension(struct rds_header *hdr,
+			      unsigned int type, const void *data)
+{
+	unsigned char *dst;
+	unsigned int len;
+
+	len = (type < __RDS_EXTHDR_MAX) ? rds_exthdr_size[type] : 0;
+	if (!len)
+		return 0;
+
+	if (rds_find_next_ext_space(hdr, len, &dst))
+		return 0;
+
+	*dst++ = type;
+	memcpy(dst, data, len);
+
+	return 1;
+}
+EXPORT_SYMBOL_GPL(rds_message_add_extension);
+
+/*
+ * If a message has extension headers, retrieve them here.
+ * Call like this:
+ *
+ * unsigned int pos = 0;
+ *
+ * while (1) {
+ *	buflen = sizeof(buffer);
+ *	type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
+ *	if (type == RDS_EXTHDR_NONE)
+ *		break;
+ *	...
+ * }
+ */
+int rds_message_next_extension(struct rds_header *hdr,
+		unsigned int *pos, void *buf, unsigned int *buflen)
+{
+	unsigned int offset, ext_type, ext_len;
+	u8 *src = hdr->h_exthdr;
+
+	offset = *pos;
+	if (offset >= RDS_HEADER_EXT_SPACE)
+		goto none;
+
+	/* Get the extension type and length. For now, the
+	 * length is implied by the extension type. */
+	ext_type = src[offset++];
+
+	if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
+		goto none;
+	ext_len = rds_exthdr_size[ext_type];
+	if (offset + ext_len > RDS_HEADER_EXT_SPACE)
+		goto none;
+
+	*pos = offset + ext_len;
+	if (ext_len < *buflen)
+		*buflen = ext_len;
+	memcpy(buf, src + offset, *buflen);
+	return ext_type;
+
+none:
+	*pos = RDS_HEADER_EXT_SPACE;
+	*buflen = 0;
+	return RDS_EXTHDR_NONE;
+}
+
+int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
+{
+	struct rds_ext_header_rdma_dest ext_hdr;
+
+	ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
+	ext_hdr.h_rdma_offset = cpu_to_be32(offset);
+	return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr);
+}
+EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
+
+/*
+ * Each rds_message is allocated with extra space for the scatterlist entries
+ * rds ops will need. This is to minimize memory allocation count. Then, each rds op
+ * can grab SGs when initializing its part of the rds_message.
+ */
+struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
+{
+	struct rds_message *rm;
+
+	if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message))
+		return NULL;
+
+	rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
+	if (!rm)
+		goto out;
+
+	rm->m_used_sgs = 0;
+	rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
+
+	refcount_set(&rm->m_refcount, 1);
+	INIT_LIST_HEAD(&rm->m_sock_item);
+	INIT_LIST_HEAD(&rm->m_conn_item);
+	spin_lock_init(&rm->m_rs_lock);
+	init_waitqueue_head(&rm->m_flush_wait);
+
+out:
+	return rm;
+}
+
+/*
+ * RDS ops use this to grab SG entries from the rm's sg pool.
+ */
+struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
+{
+	struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
+	struct scatterlist *sg_ret;
+
+	if (nents <= 0) {
+		pr_warn("rds: alloc sgs failed! nents <= 0\n");
+		return ERR_PTR(-EINVAL);
+	}
+
+	if (rm->m_used_sgs + nents > rm->m_total_sgs) {
+		pr_warn("rds: alloc sgs failed! total %d used %d nents %d\n",
+			rm->m_total_sgs, rm->m_used_sgs, nents);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	sg_ret = &sg_first[rm->m_used_sgs];
+	sg_init_table(sg_ret, nents);
+	rm->m_used_sgs += nents;
+
+	return sg_ret;
+}
+
+struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
+{
+	struct rds_message *rm;
+	unsigned int i;
+	int num_sgs = DIV_ROUND_UP(total_len, PAGE_SIZE);
+	int extra_bytes = num_sgs * sizeof(struct scatterlist);
+
+	rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
+	if (!rm)
+		return ERR_PTR(-ENOMEM);
+
+	set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
+	rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
+	rm->data.op_nents = DIV_ROUND_UP(total_len, PAGE_SIZE);
+	rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
+	if (IS_ERR(rm->data.op_sg)) {
+		void *err = ERR_CAST(rm->data.op_sg);
+		rds_message_put(rm);
+		return err;
+	}
+
+	for (i = 0; i < rm->data.op_nents; ++i) {
+		sg_set_page(&rm->data.op_sg[i],
+				virt_to_page((void *)page_addrs[i]),
+				PAGE_SIZE, 0);
+	}
+
+	return rm;
+}
+
+static int rds_message_zcopy_from_user(struct rds_message *rm, struct iov_iter *from)
+{
+	struct scatterlist *sg;
+	int ret = 0;
+	int length = iov_iter_count(from);
+	struct rds_msg_zcopy_info *info;
+
+	rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
+
+	/*
+	 * now allocate and copy in the data payload.
+	 */
+	sg = rm->data.op_sg;
+
+	info = kzalloc_obj(*info);
+	if (!info)
+		return -ENOMEM;
+	INIT_LIST_HEAD(&info->rs_zcookie_next);
+	rm->data.op_mmp_znotifier = &info->znotif;
+	if (mm_account_pinned_pages(&rm->data.op_mmp_znotifier->z_mmp,
+				    length)) {
+		ret = -ENOMEM;
+		goto err;
+	}
+	while (iov_iter_count(from)) {
+		struct page *pages;
+		size_t start;
+		ssize_t copied;
+
+		copied = iov_iter_get_pages2(from, &pages, PAGE_SIZE,
+					    1, &start);
+		if (copied < 0) {
+			struct mmpin *mmp;
+			int i;
+
+			for (i = 0; i < rm->data.op_nents; i++)
+				put_page(sg_page(&rm->data.op_sg[i]));
+			rm->data.op_nents = 0;
+			mmp = &rm->data.op_mmp_znotifier->z_mmp;
+			mm_unaccount_pinned_pages(mmp);
+			ret = -EFAULT;
+			goto err;
+		}
+		length -= copied;
+		sg_set_page(sg, pages, copied, start);
+		rm->data.op_nents++;
+		sg++;
+	}
+	WARN_ON_ONCE(length != 0);
+	return ret;
+err:
+	kfree(info);
+	rm->data.op_mmp_znotifier = NULL;
+	return ret;
+}
+
+int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
+			       bool zcopy)
+{
+	unsigned long to_copy, nbytes;
+	unsigned long sg_off;
+	struct scatterlist *sg;
+	int ret = 0;
+
+	rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
+
+	/* now allocate and copy in the data payload.  */
+	sg = rm->data.op_sg;
+	sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
+
+	if (zcopy)
+		return rds_message_zcopy_from_user(rm, from);
+
+	while (iov_iter_count(from)) {
+		if (!sg_page(sg)) {
+			ret = rds_page_remainder_alloc(sg, iov_iter_count(from),
+						       GFP_HIGHUSER);
+			if (ret)
+				return ret;
+			rm->data.op_nents++;
+			sg_off = 0;
+		}
+
+		to_copy = min_t(unsigned long, iov_iter_count(from),
+				sg->length - sg_off);
+
+		rds_stats_add(s_copy_from_user, to_copy);
+		nbytes = copy_page_from_iter(sg_page(sg), sg->offset + sg_off,
+					     to_copy, from);
+		if (nbytes != to_copy)
+			return -EFAULT;
+
+		sg_off += to_copy;
+
+		if (sg_off == sg->length)
+			sg++;
+	}
+
+	return ret;
+}
+
+int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
+{
+	struct rds_message *rm;
+	struct scatterlist *sg;
+	unsigned long to_copy;
+	unsigned long vec_off;
+	int copied;
+	int ret;
+	u32 len;
+
+	rm = container_of(inc, struct rds_message, m_inc);
+	len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
+
+	sg = rm->data.op_sg;
+	vec_off = 0;
+	copied = 0;
+
+	while (iov_iter_count(to) && copied < len) {
+		to_copy = min_t(unsigned long, iov_iter_count(to),
+				sg->length - vec_off);
+		to_copy = min_t(unsigned long, to_copy, len - copied);
+
+		rds_stats_add(s_copy_to_user, to_copy);
+		ret = copy_page_to_iter(sg_page(sg), sg->offset + vec_off,
+					to_copy, to);
+		if (ret != to_copy)
+			return -EFAULT;
+
+		vec_off += to_copy;
+		copied += to_copy;
+
+		if (vec_off == sg->length) {
+			vec_off = 0;
+			sg++;
+		}
+	}
+
+	return copied;
+}
+
+/*
+ * If the message is still on the send queue, wait until the transport
+ * is done with it. This is particularly important for RDMA operations.
+ */
+void rds_message_wait(struct rds_message *rm)
+{
+	wait_event_interruptible(rm->m_flush_wait,
+			!test_bit(RDS_MSG_MAPPED, &rm->m_flags));
+}
+
+void rds_message_unmapped(struct rds_message *rm)
+{
+	clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
+	wake_up_interruptible(&rm->m_flush_wait);
+}
+EXPORT_SYMBOL_GPL(rds_message_unmapped);

linux-7.2.3/net/rds/page.c

@@ -0,0 +1,170 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/highmem.h>
+#include <linux/gfp.h>
+#include <linux/cpu.h>
+#include <linux/export.h>
+
+#include "rds.h"
+
+struct rds_page_remainder {
+	struct page	*r_page;
+	unsigned long	r_offset;
+	local_lock_t	bh_lock;
+};
+
+static DEFINE_PER_CPU_SHARED_ALIGNED(struct rds_page_remainder, rds_page_remainders) = {
+	.bh_lock = INIT_LOCAL_LOCK(bh_lock),
+};
+
+/**
+ * rds_page_remainder_alloc - build up regions of a message.
+ *
+ * @scat: Scatter list for message
+ * @bytes: the number of bytes needed.
+ * @gfp: the waiting behaviour of the allocation
+ *
+ * @gfp is always ored with __GFP_HIGHMEM.  Callers must be prepared to
+ * kmap the pages, etc.
+ *
+ * If @bytes is at least a full page then this just returns a page from
+ * alloc_page().
+ *
+ * If @bytes is a partial page then this stores the unused region of the
+ * page in a per-cpu structure.  Future partial-page allocations may be
+ * satisfied from that cached region.  This lets us waste less memory on
+ * small allocations with minimal complexity.  It works because the transmit
+ * path passes read-only page regions down to devices.  They hold a page
+ * reference until they are done with the region.
+ */
+int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
+			     gfp_t gfp)
+{
+	struct rds_page_remainder *rem;
+	struct page *page;
+	int ret;
+
+	gfp |= __GFP_HIGHMEM;
+
+	/* jump straight to allocation if we're trying for a huge page */
+	if (bytes >= PAGE_SIZE) {
+		page = alloc_page(gfp);
+		if (!page) {
+			ret = -ENOMEM;
+		} else {
+			sg_set_page(scat, page, PAGE_SIZE, 0);
+			ret = 0;
+		}
+		goto out;
+	}
+
+	local_bh_disable();
+	local_lock_nested_bh(&rds_page_remainders.bh_lock);
+	rem = this_cpu_ptr(&rds_page_remainders);
+
+	while (1) {
+		/* avoid a tiny region getting stuck by tossing it */
+		if (rem->r_page && bytes > (PAGE_SIZE - rem->r_offset)) {
+			rds_stats_inc(s_page_remainder_miss);
+			__free_page(rem->r_page);
+			rem->r_page = NULL;
+		}
+
+		/* hand out a fragment from the cached page */
+		if (rem->r_page && bytes <= (PAGE_SIZE - rem->r_offset)) {
+			sg_set_page(scat, rem->r_page, bytes, rem->r_offset);
+			get_page(sg_page(scat));
+
+			if (rem->r_offset != 0)
+				rds_stats_inc(s_page_remainder_hit);
+
+			rem->r_offset += ALIGN(bytes, 8);
+			if (rem->r_offset >= PAGE_SIZE) {
+				__free_page(rem->r_page);
+				rem->r_page = NULL;
+			}
+			ret = 0;
+			break;
+		}
+
+		/* alloc if there is nothing for us to use */
+		local_unlock_nested_bh(&rds_page_remainders.bh_lock);
+		local_bh_enable();
+
+		page = alloc_page(gfp);
+
+		local_bh_disable();
+		local_lock_nested_bh(&rds_page_remainders.bh_lock);
+		rem = this_cpu_ptr(&rds_page_remainders);
+
+		if (!page) {
+			ret = -ENOMEM;
+			break;
+		}
+
+		/* did someone race to fill the remainder before us? */
+		if (rem->r_page) {
+			__free_page(page);
+			continue;
+		}
+
+		/* otherwise install our page and loop around to alloc */
+		rem->r_page = page;
+		rem->r_offset = 0;
+	}
+
+	local_unlock_nested_bh(&rds_page_remainders.bh_lock);
+	local_bh_enable();
+out:
+	rdsdebug("bytes %lu ret %d %p %u %u\n", bytes, ret,
+		 ret ? NULL : sg_page(scat), ret ? 0 : scat->offset,
+		 ret ? 0 : scat->length);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(rds_page_remainder_alloc);
+
+void rds_page_exit(void)
+{
+	unsigned int cpu;
+
+	for_each_possible_cpu(cpu) {
+		struct rds_page_remainder *rem;
+
+		rem = &per_cpu(rds_page_remainders, cpu);
+		rdsdebug("cpu %u\n", cpu);
+
+		if (rem->r_page)
+			__free_page(rem->r_page);
+		rem->r_page = NULL;
+	}
+}

linux-7.2.3/net/rds/rdma.c

@@ -0,0 +1,954 @@
+/*
+ * Copyright (c) 2007, 2020 Oracle and/or its affiliates.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/pagemap.h>
+#include <linux/slab.h>
+#include <linux/rbtree.h>
+#include <linux/dma-mapping.h> /* for DMA_*_DEVICE */
+
+#include "rds.h"
+
+/*
+ * XXX
+ *  - build with sparse
+ *  - should we detect duplicate keys on a socket?  hmm.
+ *  - an rdma is an mlock, apply rlimit?
+ */
+
+/*
+ * get the number of pages by looking at the page indices that the start and
+ * end addresses fall in.
+ *
+ * Returns 0 if the vec is invalid.  It is invalid if the number of bytes
+ * causes the address to wrap or overflows an unsigned int.  This comes
+ * from being stored in the 'length' member of 'struct scatterlist'.
+ */
+static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
+{
+	if ((vec->addr + vec->bytes <= vec->addr) ||
+	    (vec->bytes > (u64)UINT_MAX))
+		return 0;
+
+	return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
+		(vec->addr >> PAGE_SHIFT);
+}
+
+static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
+				       struct rds_mr *insert)
+{
+	struct rb_node **p = &root->rb_node;
+	struct rb_node *parent = NULL;
+	struct rds_mr *mr;
+
+	while (*p) {
+		parent = *p;
+		mr = rb_entry(parent, struct rds_mr, r_rb_node);
+
+		if (key < mr->r_key)
+			p = &(*p)->rb_left;
+		else if (key > mr->r_key)
+			p = &(*p)->rb_right;
+		else
+			return mr;
+	}
+
+	if (insert) {
+		rb_link_node(&insert->r_rb_node, parent, p);
+		rb_insert_color(&insert->r_rb_node, root);
+		kref_get(&insert->r_kref);
+	}
+	return NULL;
+}
+
+/*
+ * Destroy the transport-specific part of a MR.
+ */
+static void rds_destroy_mr(struct rds_mr *mr)
+{
+	struct rds_sock *rs = mr->r_sock;
+	void *trans_private = NULL;
+	unsigned long flags;
+
+	rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
+		 mr->r_key, kref_read(&mr->r_kref));
+
+	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
+	if (!RB_EMPTY_NODE(&mr->r_rb_node))
+		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
+	trans_private = mr->r_trans_private;
+	mr->r_trans_private = NULL;
+	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+
+	if (trans_private)
+		mr->r_trans->free_mr(trans_private, mr->r_invalidate);
+}
+
+void __rds_put_mr_final(struct kref *kref)
+{
+	struct rds_mr *mr = container_of(kref, struct rds_mr, r_kref);
+
+	rds_destroy_mr(mr);
+	kfree(mr);
+}
+
+/*
+ * By the time this is called we can't have any more ioctls called on
+ * the socket so we don't need to worry about racing with others.
+ */
+void rds_rdma_drop_keys(struct rds_sock *rs)
+{
+	struct rds_mr *mr;
+	struct rb_node *node;
+	unsigned long flags;
+
+	/* Release any MRs associated with this socket */
+	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
+	while ((node = rb_first(&rs->rs_rdma_keys))) {
+		mr = rb_entry(node, struct rds_mr, r_rb_node);
+		if (mr->r_trans == rs->rs_transport)
+			mr->r_invalidate = 0;
+		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
+		RB_CLEAR_NODE(&mr->r_rb_node);
+		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+		kref_put(&mr->r_kref, __rds_put_mr_final);
+		spin_lock_irqsave(&rs->rs_rdma_lock, flags);
+	}
+	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+
+	if (rs->rs_transport && rs->rs_transport->flush_mrs)
+		rs->rs_transport->flush_mrs();
+}
+
+/*
+ * Helper function to pin user pages.
+ */
+static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
+			struct page **pages, int write)
+{
+	unsigned int gup_flags = FOLL_LONGTERM;
+	int ret;
+
+	if (write)
+		gup_flags |= FOLL_WRITE;
+
+	ret = pin_user_pages_fast(user_addr, nr_pages, gup_flags, pages);
+	if (ret >= 0 && ret < nr_pages) {
+		unpin_user_pages(pages, ret);
+		ret = -EFAULT;
+	}
+
+	return ret;
+}
+
+static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
+			  u64 *cookie_ret, struct rds_mr **mr_ret,
+			  struct rds_conn_path *cp)
+{
+	struct rds_mr *mr = NULL, *found;
+	struct scatterlist *sg = NULL;
+	unsigned int nr_pages;
+	struct page **pages = NULL;
+	void *trans_private;
+	unsigned long flags;
+	rds_rdma_cookie_t cookie;
+	unsigned int nents = 0;
+	int need_odp = 0;
+	long i;
+	int ret;
+
+	if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) {
+		ret = -ENOTCONN; /* XXX not a great errno */
+		goto out;
+	}
+
+	if (!rs->rs_transport->get_mr) {
+		ret = -EOPNOTSUPP;
+		goto out;
+	}
+
+	/* If the combination of the addr and size requested for this memory
+	 * region causes an integer overflow, return error.
+	 */
+	if (((args->vec.addr + args->vec.bytes) < args->vec.addr) ||
+	    PAGE_ALIGN(args->vec.addr + args->vec.bytes) <
+		    (args->vec.addr + args->vec.bytes)) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	if (!can_do_mlock()) {
+		ret = -EPERM;
+		goto out;
+	}
+
+	nr_pages = rds_pages_in_vec(&args->vec);
+	if (nr_pages == 0) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	/* Restrict the size of mr irrespective of underlying transport
+	 * To account for unaligned mr regions, subtract one from nr_pages
+	 */
+	if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) {
+		ret = -EMSGSIZE;
+		goto out;
+	}
+
+	rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
+		args->vec.addr, args->vec.bytes, nr_pages);
+
+	/* XXX clamp nr_pages to limit the size of this alloc? */
+	pages = kzalloc_objs(struct page *, nr_pages);
+	if (!pages) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	mr = kzalloc_obj(struct rds_mr);
+	if (!mr) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	kref_init(&mr->r_kref);
+	RB_CLEAR_NODE(&mr->r_rb_node);
+	mr->r_trans = rs->rs_transport;
+	mr->r_sock = rs;
+
+	if (args->flags & RDS_RDMA_USE_ONCE)
+		mr->r_use_once = 1;
+	if (args->flags & RDS_RDMA_INVALIDATE)
+		mr->r_invalidate = 1;
+	if (args->flags & RDS_RDMA_READWRITE)
+		mr->r_write = 1;
+
+	/*
+	 * Pin the pages that make up the user buffer and transfer the page
+	 * pointers to the mr's sg array.  We check to see if we've mapped
+	 * the whole region after transferring the partial page references
+	 * to the sg array so that we can have one page ref cleanup path.
+	 *
+	 * For now we have no flag that tells us whether the mapping is
+	 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
+	 * the zero page.
+	 */
+	ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
+	if (ret == -EOPNOTSUPP) {
+		need_odp = 1;
+	} else if (ret <= 0) {
+		goto out;
+	} else {
+		nents = ret;
+		sg = kmalloc_objs(*sg, nents);
+		if (!sg) {
+			ret = -ENOMEM;
+			goto out;
+		}
+		WARN_ON(!nents);
+		sg_init_table(sg, nents);
+
+		/* Stick all pages into the scatterlist */
+		for (i = 0 ; i < nents; i++)
+			sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
+
+		rdsdebug("RDS: trans_private nents is %u\n", nents);
+	}
+	/* Obtain a transport specific MR. If this succeeds, the
+	 * s/g list is now owned by the MR.
+	 * Note that dma_map() implies that pending writes are
+	 * flushed to RAM, so no dma_sync is needed here. */
+	trans_private = rs->rs_transport->get_mr(
+		sg, nents, rs, &mr->r_key, cp ? cp->cp_conn : NULL,
+		args->vec.addr, args->vec.bytes,
+		need_odp ? ODP_ZEROBASED : ODP_NOT_NEEDED);
+
+	if (IS_ERR(trans_private)) {
+		/* In ODP case, we don't GUP pages, so don't need
+		 * to release anything.
+		 */
+		if (!need_odp) {
+			unpin_user_pages(pages, nr_pages);
+			kfree(sg);
+		}
+		ret = PTR_ERR(trans_private);
+		/* Trigger connection so that its ready for the next retry */
+		if (ret == -ENODEV && cp)
+			rds_conn_connect_if_down(cp->cp_conn);
+		goto out;
+	}
+
+	mr->r_trans_private = trans_private;
+
+	rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
+	       mr->r_key, (void *)(unsigned long) args->cookie_addr);
+
+	/* The user may pass us an unaligned address, but we can only
+	 * map page aligned regions. So we keep the offset, and build
+	 * a 64bit cookie containing <R_Key, offset> and pass that
+	 * around. */
+	if (need_odp)
+		cookie = rds_rdma_make_cookie(mr->r_key, 0);
+	else
+		cookie = rds_rdma_make_cookie(mr->r_key,
+					      args->vec.addr & ~PAGE_MASK);
+	if (cookie_ret)
+		*cookie_ret = cookie;
+
+	if (args->cookie_addr &&
+	    put_user(cookie, (u64 __user *)(unsigned long)args->cookie_addr)) {
+		ret = -EFAULT;
+		goto out;
+	}
+
+	/* Inserting the new MR into the rbtree bumps its
+	 * reference count. */
+	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
+	found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
+	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+
+	BUG_ON(found && found != mr);
+
+	rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
+	if (mr_ret) {
+		kref_get(&mr->r_kref);
+		*mr_ret = mr;
+	}
+
+	ret = 0;
+out:
+	kfree(pages);
+	if (mr)
+		kref_put(&mr->r_kref, __rds_put_mr_final);
+	return ret;
+}
+
+int rds_get_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
+{
+	struct rds_get_mr_args args;
+
+	if (optlen != sizeof(struct rds_get_mr_args))
+		return -EINVAL;
+
+	if (copy_from_sockptr(&args, optval, sizeof(struct rds_get_mr_args)))
+		return -EFAULT;
+
+	return __rds_rdma_map(rs, &args, NULL, NULL, NULL);
+}
+
+int rds_get_mr_for_dest(struct rds_sock *rs, sockptr_t optval, int optlen)
+{
+	struct rds_get_mr_for_dest_args args;
+	struct rds_get_mr_args new_args;
+
+	if (optlen != sizeof(struct rds_get_mr_for_dest_args))
+		return -EINVAL;
+
+	if (copy_from_sockptr(&args, optval,
+			   sizeof(struct rds_get_mr_for_dest_args)))
+		return -EFAULT;
+
+	/*
+	 * Initially, just behave like get_mr().
+	 * TODO: Implement get_mr as wrapper around this
+	 *	 and deprecate it.
+	 */
+	new_args.vec = args.vec;
+	new_args.cookie_addr = args.cookie_addr;
+	new_args.flags = args.flags;
+
+	return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL);
+}
+
+/*
+ * Free the MR indicated by the given R_Key
+ */
+int rds_free_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
+{
+	struct rds_free_mr_args args;
+	struct rds_mr *mr;
+	unsigned long flags;
+
+	if (optlen != sizeof(struct rds_free_mr_args))
+		return -EINVAL;
+
+	if (copy_from_sockptr(&args, optval, sizeof(struct rds_free_mr_args)))
+		return -EFAULT;
+
+	/* Special case - a null cookie means flush all unused MRs */
+	if (args.cookie == 0) {
+		if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
+			return -EINVAL;
+		rs->rs_transport->flush_mrs();
+		return 0;
+	}
+
+	/* Look up the MR given its R_key and remove it from the rbtree
+	 * so nobody else finds it.
+	 * This should also prevent races with rds_rdma_unuse.
+	 */
+	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
+	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
+	if (mr) {
+		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
+		RB_CLEAR_NODE(&mr->r_rb_node);
+		if (args.flags & RDS_RDMA_INVALIDATE)
+			mr->r_invalidate = 1;
+	}
+	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+
+	if (!mr)
+		return -EINVAL;
+
+	kref_put(&mr->r_kref, __rds_put_mr_final);
+	return 0;
+}
+
+/*
+ * This is called when we receive an extension header that
+ * tells us this MR was used. It allows us to implement
+ * use_once semantics
+ */
+void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
+{
+	struct rds_mr *mr;
+	unsigned long flags;
+	int zot_me = 0;
+
+	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
+	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
+	if (!mr) {
+		pr_debug("rds: trying to unuse MR with unknown r_key %u!\n",
+			 r_key);
+		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+		return;
+	}
+
+	/* Get a reference so that the MR won't go away before calling
+	 * sync_mr() below.
+	 */
+	kref_get(&mr->r_kref);
+
+	/* If it is going to be freed, remove it from the tree now so
+	 * that no other thread can find it and free it.
+	 */
+	if (mr->r_use_once || force) {
+		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
+		RB_CLEAR_NODE(&mr->r_rb_node);
+		zot_me = 1;
+	}
+	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+
+	/* May have to issue a dma_sync on this memory region.
+	 * Note we could avoid this if the operation was a RDMA READ,
+	 * but at this point we can't tell. */
+	if (mr->r_trans->sync_mr)
+		mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
+
+	/* Release the reference held above. */
+	kref_put(&mr->r_kref, __rds_put_mr_final);
+
+	/* If the MR was marked as invalidate, this will
+	 * trigger an async flush. */
+	if (zot_me)
+		kref_put(&mr->r_kref, __rds_put_mr_final);
+}
+
+void rds_rdma_free_op(struct rm_rdma_op *ro)
+{
+	unsigned int i;
+
+	if (ro->op_odp_mr) {
+		kref_put(&ro->op_odp_mr->r_kref, __rds_put_mr_final);
+	} else {
+		for (i = 0; i < ro->op_nents; i++) {
+			struct page *page = sg_page(&ro->op_sg[i]);
+
+			/* Mark page dirty if it was possibly modified, which
+			 * is the case for a RDMA_READ which copies from remote
+			 * to local memory
+			 */
+			unpin_user_pages_dirty_lock(&page, 1, !ro->op_write);
+		}
+	}
+
+	kfree(ro->op_notifier);
+	ro->op_notifier = NULL;
+	ro->op_active = 0;
+	ro->op_odp_mr = NULL;
+}
+
+void rds_atomic_free_op(struct rm_atomic_op *ao)
+{
+	struct page *page = sg_page(ao->op_sg);
+
+	/* Mark page dirty if it was possibly modified, which
+	 * is the case for a RDMA_READ which copies from remote
+	 * to local memory */
+	unpin_user_pages_dirty_lock(&page, 1, true);
+
+	kfree(ao->op_notifier);
+	ao->op_notifier = NULL;
+	ao->op_active = 0;
+}
+
+
+/*
+ * Count the number of pages needed to describe an incoming iovec array.
+ */
+static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
+{
+	int tot_pages = 0;
+	unsigned int nr_pages;
+	unsigned int i;
+
+	/* figure out the number of pages in the vector */
+	for (i = 0; i < nr_iovecs; i++) {
+		nr_pages = rds_pages_in_vec(&iov[i]);
+		if (nr_pages == 0)
+			return -EINVAL;
+
+		tot_pages += nr_pages;
+
+		/*
+		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
+		 * so tot_pages cannot overflow without first going negative.
+		 */
+		if (tot_pages < 0)
+			return -EINVAL;
+	}
+
+	return tot_pages;
+}
+
+int rds_rdma_extra_size(struct rds_rdma_args *args,
+			struct rds_iov_vector *iov)
+{
+	struct rds_iovec *vec;
+	struct rds_iovec __user *local_vec;
+	int tot_pages = 0;
+	unsigned int nr_pages;
+	unsigned int i;
+
+	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
+
+	if (args->nr_local == 0)
+		return -EINVAL;
+
+	if (args->nr_local > UIO_MAXIOV)
+		return -EMSGSIZE;
+
+	iov->iov = kzalloc_objs(struct rds_iovec, args->nr_local);
+	if (!iov->iov)
+		return -ENOMEM;
+
+	vec = &iov->iov[0];
+
+	if (copy_from_user(vec, local_vec, args->nr_local *
+			   sizeof(struct rds_iovec)))
+		return -EFAULT;
+	iov->len = args->nr_local;
+
+	/* figure out the number of pages in the vector */
+	for (i = 0; i < args->nr_local; i++, vec++) {
+
+		nr_pages = rds_pages_in_vec(vec);
+		if (nr_pages == 0)
+			return -EINVAL;
+
+		tot_pages += nr_pages;
+
+		/*
+		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
+		 * so tot_pages cannot overflow without first going negative.
+		 */
+		if (tot_pages < 0)
+			return -EINVAL;
+	}
+
+	return tot_pages * sizeof(struct scatterlist);
+}
+
+/*
+ * The application asks for a RDMA transfer.
+ * Extract all arguments and set up the rdma_op
+ */
+int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
+		       struct cmsghdr *cmsg,
+		       struct rds_iov_vector *vec)
+{
+	struct rds_rdma_args *args;
+	struct rm_rdma_op *op = &rm->rdma;
+	int nr_pages;
+	unsigned int nr_bytes;
+	struct page **pages = NULL;
+	struct rds_iovec *iovs;
+	unsigned int i, j;
+	int ret = 0;
+	bool odp_supported = true;
+
+	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
+	    || rm->rdma.op_active)
+		return -EINVAL;
+
+	args = CMSG_DATA(cmsg);
+
+	if (ipv6_addr_any(&rs->rs_bound_addr)) {
+		ret = -ENOTCONN; /* XXX not a great errno */
+		goto out_ret;
+	}
+
+	if (args->nr_local > UIO_MAXIOV) {
+		ret = -EMSGSIZE;
+		goto out_ret;
+	}
+
+	if (vec->len != args->nr_local) {
+		ret = -EINVAL;
+		goto out_ret;
+	}
+	/* odp-mr is not supported for multiple requests within one message */
+	if (args->nr_local != 1)
+		odp_supported = false;
+
+	iovs = vec->iov;
+
+	nr_pages = rds_rdma_pages(iovs, args->nr_local);
+	if (nr_pages < 0) {
+		ret = -EINVAL;
+		goto out_ret;
+	}
+
+	pages = kzalloc_objs(struct page *, nr_pages);
+	if (!pages) {
+		ret = -ENOMEM;
+		goto out_ret;
+	}
+
+	op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
+	op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
+	op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
+	op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
+	op->op_active = 1;
+	op->op_recverr = rs->rs_recverr;
+	op->op_odp_mr = NULL;
+
+	WARN_ON(!nr_pages);
+	op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
+	if (IS_ERR(op->op_sg)) {
+		ret = PTR_ERR(op->op_sg);
+		goto out_pages;
+	}
+
+	if (op->op_notify || op->op_recverr) {
+		/* We allocate an uninitialized notifier here, because
+		 * we don't want to do that in the completion handler. We
+		 * would have to use GFP_ATOMIC there, and don't want to deal
+		 * with failed allocations.
+		 */
+		op->op_notifier = kmalloc_obj(struct rds_notifier);
+		if (!op->op_notifier) {
+			ret = -ENOMEM;
+			goto out_pages;
+		}
+		op->op_notifier->n_user_token = args->user_token;
+		op->op_notifier->n_status = RDS_RDMA_SUCCESS;
+	}
+
+	/* The cookie contains the R_Key of the remote memory region, and
+	 * optionally an offset into it. This is how we implement RDMA into
+	 * unaligned memory.
+	 * When setting up the RDMA, we need to add that offset to the
+	 * destination address (which is really an offset into the MR)
+	 * FIXME: We may want to move this into ib_rdma.c
+	 */
+	op->op_rkey = rds_rdma_cookie_key(args->cookie);
+	op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
+
+	nr_bytes = 0;
+
+	rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
+	       (unsigned long long)args->nr_local,
+	       (unsigned long long)args->remote_vec.addr,
+	       op->op_rkey);
+
+	for (i = 0; i < args->nr_local; i++) {
+		struct rds_iovec *iov = &iovs[i];
+		/* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
+		unsigned int nr = rds_pages_in_vec(iov);
+
+		rs->rs_user_addr = iov->addr;
+		rs->rs_user_bytes = iov->bytes;
+
+		/* If it's a WRITE operation, we want to pin the pages for reading.
+		 * If it's a READ operation, we need to pin the pages for writing.
+		 */
+		ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
+		if ((!odp_supported && ret <= 0) ||
+		    (odp_supported && ret <= 0 && ret != -EOPNOTSUPP))
+			goto out_pages;
+
+		if (ret == -EOPNOTSUPP) {
+			struct rds_mr *local_odp_mr;
+
+			if (!rs->rs_transport->get_mr) {
+				ret = -EOPNOTSUPP;
+				goto out_pages;
+			}
+			local_odp_mr = kzalloc_obj(*local_odp_mr);
+			if (!local_odp_mr) {
+				ret = -ENOMEM;
+				goto out_pages;
+			}
+			RB_CLEAR_NODE(&local_odp_mr->r_rb_node);
+			kref_init(&local_odp_mr->r_kref);
+			local_odp_mr->r_trans = rs->rs_transport;
+			local_odp_mr->r_sock = rs;
+			local_odp_mr->r_trans_private =
+				rs->rs_transport->get_mr(
+					NULL, 0, rs, &local_odp_mr->r_key, NULL,
+					iov->addr, iov->bytes, ODP_VIRTUAL);
+			if (IS_ERR(local_odp_mr->r_trans_private)) {
+				ret = PTR_ERR(local_odp_mr->r_trans_private);
+				rdsdebug("get_mr ret %d %p\"", ret,
+					 local_odp_mr->r_trans_private);
+				kfree(local_odp_mr);
+				ret = -EOPNOTSUPP;
+				goto out_pages;
+			}
+			rdsdebug("Need odp; local_odp_mr %p trans_private %p\n",
+				 local_odp_mr, local_odp_mr->r_trans_private);
+			op->op_odp_mr = local_odp_mr;
+			op->op_odp_addr = iov->addr;
+		}
+
+		rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
+			 nr_bytes, nr, iov->bytes, iov->addr);
+
+		nr_bytes += iov->bytes;
+
+		for (j = 0; j < nr; j++) {
+			unsigned int offset = iov->addr & ~PAGE_MASK;
+			struct scatterlist *sg;
+
+			sg = &op->op_sg[op->op_nents + j];
+			sg_set_page(sg, pages[j],
+					min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
+					offset);
+
+			sg_dma_len(sg) = sg->length;
+			rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
+			       sg->offset, sg->length, iov->addr, iov->bytes);
+
+			iov->addr += sg->length;
+			iov->bytes -= sg->length;
+		}
+
+		op->op_nents += nr;
+	}
+
+	if (nr_bytes > args->remote_vec.bytes) {
+		rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
+				nr_bytes,
+				(unsigned int) args->remote_vec.bytes);
+		ret = -EINVAL;
+		goto out_pages;
+	}
+	op->op_bytes = nr_bytes;
+	ret = 0;
+
+out_pages:
+	kfree(pages);
+out_ret:
+	if (ret)
+		rds_rdma_free_op(op);
+	else
+		rds_stats_inc(s_send_rdma);
+
+	return ret;
+}
+
+/*
+ * The application wants us to pass an RDMA destination (aka MR)
+ * to the remote
+ */
+int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
+			  struct cmsghdr *cmsg)
+{
+	unsigned long flags;
+	struct rds_mr *mr;
+	u32 r_key;
+	int err = 0;
+
+	if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
+	    rm->m_rdma_cookie != 0)
+		return -EINVAL;
+
+	memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
+
+	/* We are reusing a previously mapped MR here. Most likely, the
+	 * application has written to the buffer, so we need to explicitly
+	 * flush those writes to RAM. Otherwise the HCA may not see them
+	 * when doing a DMA from that buffer.
+	 */
+	r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
+
+	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
+	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
+	if (!mr)
+		err = -EINVAL;	/* invalid r_key */
+	else
+		kref_get(&mr->r_kref);
+	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+
+	if (mr) {
+		mr->r_trans->sync_mr(mr->r_trans_private,
+				     DMA_TO_DEVICE);
+		rm->rdma.op_rdma_mr = mr;
+	}
+	return err;
+}
+
+/*
+ * The application passes us an address range it wants to enable RDMA
+ * to/from. We map the area, and save the <R_Key,offset> pair
+ * in rm->m_rdma_cookie. This causes it to be sent along to the peer
+ * in an extension header.
+ */
+int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
+			  struct cmsghdr *cmsg)
+{
+	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
+	    rm->m_rdma_cookie != 0)
+		return -EINVAL;
+
+	return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie,
+			      &rm->rdma.op_rdma_mr, rm->m_conn_path);
+}
+
+/*
+ * Fill in rds_message for an atomic request.
+ */
+int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
+		    struct cmsghdr *cmsg)
+{
+	struct page *page = NULL;
+	struct rds_atomic_args *args;
+	int ret = 0;
+
+	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
+	 || rm->atomic.op_active)
+		return -EINVAL;
+
+	args = CMSG_DATA(cmsg);
+
+	/* Nonmasked & masked cmsg ops converted to masked hw ops */
+	switch (cmsg->cmsg_type) {
+	case RDS_CMSG_ATOMIC_FADD:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
+		rm->atomic.op_m_fadd.add = args->fadd.add;
+		rm->atomic.op_m_fadd.nocarry_mask = 0;
+		break;
+	case RDS_CMSG_MASKED_ATOMIC_FADD:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
+		rm->atomic.op_m_fadd.add = args->m_fadd.add;
+		rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
+		break;
+	case RDS_CMSG_ATOMIC_CSWP:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
+		rm->atomic.op_m_cswp.compare = args->cswp.compare;
+		rm->atomic.op_m_cswp.swap = args->cswp.swap;
+		rm->atomic.op_m_cswp.compare_mask = ~0;
+		rm->atomic.op_m_cswp.swap_mask = ~0;
+		break;
+	case RDS_CMSG_MASKED_ATOMIC_CSWP:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
+		rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
+		rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
+		rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
+		rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
+		break;
+	default:
+		BUG(); /* should never happen */
+	}
+
+	rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
+	rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
+	rm->atomic.op_active = 1;
+	rm->atomic.op_recverr = rs->rs_recverr;
+	rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
+	if (IS_ERR(rm->atomic.op_sg)) {
+		ret = PTR_ERR(rm->atomic.op_sg);
+		goto err;
+	}
+
+	/* verify 8 byte-aligned */
+	if (args->local_addr & 0x7) {
+		ret = -EFAULT;
+		goto err;
+	}
+
+	ret = rds_pin_pages(args->local_addr, 1, &page, 1);
+	if (ret != 1)
+		goto err;
+	ret = 0;
+
+	sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
+
+	if (rm->atomic.op_notify || rm->atomic.op_recverr) {
+		/* We allocate an uninitialized notifier here, because
+		 * we don't want to do that in the completion handler. We
+		 * would have to use GFP_ATOMIC there, and don't want to deal
+		 * with failed allocations.
+		 */
+		rm->atomic.op_notifier = kmalloc_obj(*rm->atomic.op_notifier);
+		if (!rm->atomic.op_notifier) {
+			ret = -ENOMEM;
+			goto err;
+		}
+
+		rm->atomic.op_notifier->n_user_token = args->user_token;
+		rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
+	}
+
+	rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
+	rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
+
+	return ret;
+err:
+	if (page)
+		unpin_user_page(page);
+	rm->atomic.op_active = 0;
+	kfree(rm->atomic.op_notifier);
+
+	return ret;
+}

linux-7.2.3/net/rds/rdma_transport.c

@@ -0,0 +1,322 @@
+/*
+ * Copyright (c) 2009, 2018 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/module.h>
+#include <rdma/rdma_cm.h>
+
+#include "rds_single_path.h"
+#include "rdma_transport.h"
+#include "ib.h"
+
+/* Global IPv4 and IPv6 RDS RDMA listener cm_id */
+static struct rdma_cm_id *rds_rdma_listen_id;
+#if IS_ENABLED(CONFIG_IPV6)
+static struct rdma_cm_id *rds6_rdma_listen_id;
+#endif
+
+/* Per IB specification 7.7.3, service level is a 4-bit field. */
+#define TOS_TO_SL(tos)		((tos) & 0xF)
+
+static int rds_rdma_cm_event_handler_cmn(struct rdma_cm_id *cm_id,
+					 struct rdma_cm_event *event,
+					 bool isv6)
+{
+	/* this can be null in the listening path */
+	struct rds_connection *conn = cm_id->context;
+	struct rds_transport *trans;
+	int ret = 0;
+	int *err;
+	u8 len;
+
+	rdsdebug("conn %p id %p handling event %u (%s)\n", conn, cm_id,
+		 event->event, rdma_event_msg(event->event));
+
+	if (cm_id->device->node_type == RDMA_NODE_IB_CA)
+		trans = &rds_ib_transport;
+
+	/* Prevent shutdown from tearing down the connection
+	 * while we're executing. */
+	if (conn) {
+		mutex_lock(&conn->c_cm_lock);
+
+		/* If the connection is being shut down, bail out
+		 * right away. We return 0 so cm_id doesn't get
+		 * destroyed prematurely */
+		if (rds_conn_state(conn) == RDS_CONN_DISCONNECTING) {
+			/* Reject incoming connections while we're tearing
+			 * down an existing one. */
+			if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST)
+				ret = 1;
+			goto out;
+		}
+	}
+
+	switch (event->event) {
+	case RDMA_CM_EVENT_CONNECT_REQUEST:
+		ret = trans->cm_handle_connect(cm_id, event, isv6);
+		break;
+
+	case RDMA_CM_EVENT_ADDR_RESOLVED:
+		if (conn) {
+			rdma_set_service_type(cm_id, conn->c_tos);
+			rdma_set_min_rnr_timer(cm_id, IB_RNR_TIMER_000_32);
+			/* XXX do we need to clean up if this fails? */
+			ret = rdma_resolve_route(cm_id,
+						 RDS_RDMA_RESOLVE_TIMEOUT_MS);
+		}
+		break;
+
+	case RDMA_CM_EVENT_ROUTE_RESOLVED:
+		/* Connection could have been dropped so make sure the
+		 * cm_id is valid before proceeding
+		 */
+		if (conn) {
+			struct rds_ib_connection *ibic;
+
+			ibic = conn->c_transport_data;
+			if (ibic && ibic->i_cm_id == cm_id) {
+				cm_id->route.path_rec[0].sl =
+					TOS_TO_SL(conn->c_tos);
+				ret = trans->cm_initiate_connect(cm_id, isv6);
+			} else {
+				rds_conn_drop(conn);
+			}
+		}
+		break;
+
+	case RDMA_CM_EVENT_ESTABLISHED:
+		if (conn)
+			trans->cm_connect_complete(conn, event);
+		break;
+
+	case RDMA_CM_EVENT_REJECTED:
+		if (!conn)
+			break;
+		err = (int *)rdma_consumer_reject_data(cm_id, event, &len);
+		if (!err ||
+		    (err && len >= sizeof(*err) &&
+		     ((*err) <= RDS_RDMA_REJ_INCOMPAT))) {
+			pr_warn("RDS/RDMA: conn <%pI6c, %pI6c> rejected, dropping connection\n",
+				&conn->c_laddr, &conn->c_faddr);
+
+			if (!conn->c_tos)
+				conn->c_proposed_version = RDS_PROTOCOL_COMPAT_VERSION;
+
+			rds_conn_drop(conn);
+		}
+		rdsdebug("Connection rejected: %s\n",
+			 rdma_reject_msg(cm_id, event->status));
+		break;
+	case RDMA_CM_EVENT_ADDR_ERROR:
+	case RDMA_CM_EVENT_ROUTE_ERROR:
+	case RDMA_CM_EVENT_CONNECT_ERROR:
+	case RDMA_CM_EVENT_UNREACHABLE:
+	case RDMA_CM_EVENT_DEVICE_REMOVAL:
+	case RDMA_CM_EVENT_ADDR_CHANGE:
+		if (conn)
+			rds_conn_drop(conn);
+		break;
+
+	case RDMA_CM_EVENT_DISCONNECTED:
+		if (!conn)
+			break;
+		rdsdebug("DISCONNECT event - dropping connection "
+			 "%pI6c->%pI6c\n", &conn->c_laddr,
+			 &conn->c_faddr);
+		rds_conn_drop(conn);
+		break;
+
+	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
+		if (conn) {
+			pr_info("RDS: RDMA_CM_EVENT_TIMEWAIT_EXIT event: dropping connection %pI6c->%pI6c\n",
+				&conn->c_laddr, &conn->c_faddr);
+			rds_conn_drop(conn);
+		}
+		break;
+
+	default:
+		/* things like device disconnect? */
+		printk(KERN_ERR "RDS: unknown event %u (%s)!\n",
+		       event->event, rdma_event_msg(event->event));
+		break;
+	}
+
+out:
+	if (conn)
+		mutex_unlock(&conn->c_cm_lock);
+
+	rdsdebug("id %p event %u (%s) handling ret %d\n", cm_id, event->event,
+		 rdma_event_msg(event->event), ret);
+
+	return ret;
+}
+
+int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
+			      struct rdma_cm_event *event)
+{
+	return rds_rdma_cm_event_handler_cmn(cm_id, event, false);
+}
+
+#if IS_ENABLED(CONFIG_IPV6)
+int rds6_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
+			       struct rdma_cm_event *event)
+{
+	return rds_rdma_cm_event_handler_cmn(cm_id, event, true);
+}
+#endif
+
+static int rds_rdma_listen_init_common(rdma_cm_event_handler handler,
+				       struct sockaddr *sa,
+				       struct rdma_cm_id **ret_cm_id)
+{
+	struct rdma_cm_id *cm_id;
+	int ret;
+
+	cm_id = rdma_create_id(&init_net, handler, NULL,
+			       RDMA_PS_TCP, IB_QPT_RC);
+	if (IS_ERR(cm_id)) {
+		ret = PTR_ERR(cm_id);
+		printk(KERN_ERR "RDS/RDMA: failed to setup listener, "
+		       "rdma_create_id() returned %d\n", ret);
+		return ret;
+	}
+
+	/*
+	 * XXX I bet this binds the cm_id to a device.  If we want to support
+	 * fail-over we'll have to take this into consideration.
+	 */
+	ret = rdma_bind_addr(cm_id, sa);
+	if (ret) {
+		printk(KERN_ERR "RDS/RDMA: failed to setup listener, "
+		       "rdma_bind_addr() returned %d\n", ret);
+		goto out;
+	}
+
+	ret = rdma_listen(cm_id, 128);
+	if (ret) {
+		printk(KERN_ERR "RDS/RDMA: failed to setup listener, "
+		       "rdma_listen() returned %d\n", ret);
+		goto out;
+	}
+
+	rdsdebug("cm %p listening on port %u\n", cm_id, RDS_PORT);
+
+	*ret_cm_id = cm_id;
+	cm_id = NULL;
+out:
+	if (cm_id)
+		rdma_destroy_id(cm_id);
+	return ret;
+}
+
+/* Initialize the RDS RDMA listeners.  We create two listeners for
+ * compatibility reason.  The one on RDS_PORT is used for IPv4
+ * requests only.  The one on RDS_CM_PORT is used for IPv6 requests
+ * only.  So only IPv6 enabled RDS module will communicate using this
+ * port.
+ */
+static int rds_rdma_listen_init(void)
+{
+	int ret;
+#if IS_ENABLED(CONFIG_IPV6)
+	struct sockaddr_in6 sin6;
+#endif
+	struct sockaddr_in sin;
+
+	sin.sin_family = PF_INET;
+	sin.sin_addr.s_addr = htonl(INADDR_ANY);
+	sin.sin_port = htons(RDS_PORT);
+	ret = rds_rdma_listen_init_common(rds_rdma_cm_event_handler,
+					  (struct sockaddr *)&sin,
+					  &rds_rdma_listen_id);
+	if (ret != 0)
+		return ret;
+
+#if IS_ENABLED(CONFIG_IPV6)
+	sin6.sin6_family = PF_INET6;
+	sin6.sin6_addr = in6addr_any;
+	sin6.sin6_port = htons(RDS_CM_PORT);
+	sin6.sin6_scope_id = 0;
+	sin6.sin6_flowinfo = 0;
+	ret = rds_rdma_listen_init_common(rds6_rdma_cm_event_handler,
+					  (struct sockaddr *)&sin6,
+					  &rds6_rdma_listen_id);
+	/* Keep going even when IPv6 is not enabled in the system. */
+	if (ret != 0)
+		rdsdebug("Cannot set up IPv6 RDMA listener\n");
+#endif
+	return 0;
+}
+
+static void rds_rdma_listen_stop(void)
+{
+	if (rds_rdma_listen_id) {
+		rdsdebug("cm %p\n", rds_rdma_listen_id);
+		rdma_destroy_id(rds_rdma_listen_id);
+		rds_rdma_listen_id = NULL;
+	}
+#if IS_ENABLED(CONFIG_IPV6)
+	if (rds6_rdma_listen_id) {
+		rdsdebug("cm %p\n", rds6_rdma_listen_id);
+		rdma_destroy_id(rds6_rdma_listen_id);
+		rds6_rdma_listen_id = NULL;
+	}
+#endif
+}
+
+static int __init rds_rdma_init(void)
+{
+	int ret;
+
+	ret = rds_ib_init();
+	if (ret)
+		goto out;
+
+	ret = rds_rdma_listen_init();
+	if (ret)
+		rds_ib_exit();
+out:
+	return ret;
+}
+module_init(rds_rdma_init);
+
+static void __exit rds_rdma_exit(void)
+{
+	/* stop listening first to ensure no new connections are attempted */
+	rds_rdma_listen_stop();
+	rds_ib_exit();
+}
+module_exit(rds_rdma_exit);
+
+MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
+MODULE_DESCRIPTION("RDS: IB transport");
+MODULE_LICENSE("Dual BSD/GPL");

linux-7.2.3/net/rds/rdma_transport.h

@@ -0,0 +1,30 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _RDMA_TRANSPORT_H
+#define _RDMA_TRANSPORT_H
+
+#include <rdma/ib_verbs.h>
+#include <rdma/rdma_cm.h>
+#include "rds.h"
+
+/* RDMA_CM also uses 16385 as the listener port. */
+#define RDS_CM_PORT	16385
+
+#define RDS_RDMA_RESOLVE_TIMEOUT_MS     5000
+
+/* Below reject reason is for legacy interoperability issue with non-linux
+ * RDS endpoints where older version incompatibility is conveyed via value 1.
+ * For future version(s), proper encoded reject reason should be used.
+ */
+#define RDS_RDMA_REJ_INCOMPAT		1
+
+int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
+			      struct rdma_cm_event *event);
+int rds6_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
+			       struct rdma_cm_event *event);
+
+/* from ib.c */
+extern struct rds_transport rds_ib_transport;
+int rds_ib_init(void);
+void rds_ib_exit(void);
+
+#endif

linux-7.2.3/net/rds/rds_single_path.h

@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _RDS_RDS_SINGLE_H
+#define _RDS_RDS_SINGLE_H
+
+#define	c_xmit_rm		c_path[0].cp_xmit_rm
+#define	c_xmit_sg		c_path[0].cp_xmit_sg
+#define	c_xmit_hdr_off		c_path[0].cp_xmit_hdr_off
+#define	c_xmit_data_off		c_path[0].cp_xmit_data_off
+#define	c_xmit_atomic_sent	c_path[0].cp_xmit_atomic_sent
+#define	c_xmit_rdma_sent	c_path[0].cp_xmit_rdma_sent
+#define	c_xmit_data_sent	c_path[0].cp_xmit_data_sent
+#define	c_lock			c_path[0].cp_lock
+#define c_next_tx_seq		c_path[0].cp_next_tx_seq
+#define c_send_queue		c_path[0].cp_send_queue
+#define c_retrans		c_path[0].cp_retrans
+#define c_next_rx_seq		c_path[0].cp_next_rx_seq
+#define c_transport_data	c_path[0].cp_transport_data
+#define c_state			c_path[0].cp_state
+#define c_send_gen		c_path[0].cp_send_gen
+#define c_flags			c_path[0].cp_flags
+#define c_reconnect_jiffies	c_path[0].cp_reconnect_jiffies
+#define c_send_w		c_path[0].cp_send_w
+#define c_recv_w		c_path[0].cp_recv_w
+#define c_conn_w		c_path[0].cp_conn_w
+#define c_down_w		c_path[0].cp_down_w
+#define c_cm_lock		c_path[0].cp_cm_lock
+#define c_waitq			c_path[0].cp_waitq
+#define c_unacked_packets	c_path[0].cp_unacked_packets
+#define c_unacked_bytes		c_path[0].cp_unacked_bytes
+
+#endif /* _RDS_RDS_SINGLE_H */

linux-7.2.3/net/rds/recv.c

@@ -0,0 +1,889 @@
+/*
+ * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <net/sock.h>
+#include <linux/in.h>
+#include <linux/export.h>
+#include <linux/sched/clock.h>
+#include <linux/time.h>
+#include <linux/rds.h>
+
+#include "rds.h"
+
+void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
+		 struct in6_addr *saddr)
+{
+	refcount_set(&inc->i_refcount, 1);
+	INIT_LIST_HEAD(&inc->i_item);
+	inc->i_conn = conn;
+	inc->i_saddr = *saddr;
+	inc->i_usercopy.rdma_cookie = 0;
+	inc->i_usercopy.rx_tstamp = ktime_set(0, 0);
+
+	memset(inc->i_rx_lat_trace, 0, sizeof(inc->i_rx_lat_trace));
+}
+EXPORT_SYMBOL_GPL(rds_inc_init);
+
+void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *cp,
+		       struct in6_addr  *saddr)
+{
+	refcount_set(&inc->i_refcount, 1);
+	INIT_LIST_HEAD(&inc->i_item);
+	inc->i_conn = cp->cp_conn;
+	inc->i_conn_path = cp;
+	inc->i_saddr = *saddr;
+	inc->i_usercopy.rdma_cookie = 0;
+	inc->i_usercopy.rx_tstamp = ktime_set(0, 0);
+}
+EXPORT_SYMBOL_GPL(rds_inc_path_init);
+
+static void rds_inc_addref(struct rds_incoming *inc)
+{
+	rdsdebug("addref inc %p ref %d\n", inc, refcount_read(&inc->i_refcount));
+	refcount_inc(&inc->i_refcount);
+}
+
+void rds_inc_put(struct rds_incoming *inc)
+{
+	rdsdebug("put inc %p ref %d\n", inc, refcount_read(&inc->i_refcount));
+	if (refcount_dec_and_test(&inc->i_refcount)) {
+		BUG_ON(!list_empty(&inc->i_item));
+
+		inc->i_conn->c_trans->inc_free(inc);
+	}
+}
+EXPORT_SYMBOL_GPL(rds_inc_put);
+
+static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk,
+				  struct rds_cong_map *map,
+				  int delta, __be16 port)
+{
+	int now_congested;
+
+	if (delta == 0)
+		return;
+
+	rs->rs_rcv_bytes += delta;
+	if (delta > 0)
+		rds_stats_add(s_recv_bytes_added_to_socket, delta);
+	else
+		rds_stats_add(s_recv_bytes_removed_from_socket, -delta);
+
+	/* loop transport doesn't send/recv congestion updates */
+	if (rs->rs_transport->t_type == RDS_TRANS_LOOP)
+		return;
+
+	now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
+
+	rdsdebug("rs %p (%pI6c:%u) recv bytes %d buf %d "
+	  "now_cong %d delta %d\n",
+	  rs, &rs->rs_bound_addr,
+	  ntohs(rs->rs_bound_port), rs->rs_rcv_bytes,
+	  rds_sk_rcvbuf(rs), now_congested, delta);
+
+	/* wasn't -> am congested */
+	if (!rs->rs_congested && now_congested) {
+		rs->rs_congested = 1;
+		rds_cong_set_bit(map, port);
+		rds_cong_queue_updates(map);
+	}
+	/* was -> aren't congested */
+	/* Require more free space before reporting uncongested to prevent
+	   bouncing cong/uncong state too often */
+	else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) {
+		rs->rs_congested = 0;
+		rds_cong_clear_bit(map, port);
+		rds_cong_queue_updates(map);
+	}
+
+	/* do nothing if no change in cong state */
+}
+
+static void rds_conn_peer_gen_update(struct rds_connection *conn,
+				     u32 peer_gen_num)
+{
+	int i;
+	struct rds_message *rm, *tmp;
+	unsigned long flags;
+
+	WARN_ON(conn->c_trans->t_type != RDS_TRANS_TCP);
+	if (peer_gen_num != 0) {
+		if (conn->c_peer_gen_num != 0 &&
+		    peer_gen_num != conn->c_peer_gen_num) {
+			for (i = 0; i < RDS_MPATH_WORKERS; i++) {
+				struct rds_conn_path *cp;
+
+				cp = &conn->c_path[i];
+				spin_lock_irqsave(&cp->cp_lock, flags);
+				cp->cp_next_tx_seq = 1;
+				cp->cp_next_rx_seq = 0;
+				list_for_each_entry_safe(rm, tmp,
+							 &cp->cp_retrans,
+							 m_conn_item) {
+					set_bit(RDS_MSG_FLUSH, &rm->m_flags);
+				}
+				spin_unlock_irqrestore(&cp->cp_lock, flags);
+			}
+		}
+		conn->c_peer_gen_num = peer_gen_num;
+	}
+}
+
+/*
+ * Process all extension headers that come with this message.
+ */
+static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs)
+{
+	struct rds_header *hdr = &inc->i_hdr;
+	unsigned int pos = 0, type, len;
+	union {
+		struct rds_ext_header_version version;
+		struct rds_ext_header_rdma rdma;
+		struct rds_ext_header_rdma_dest rdma_dest;
+	} buffer;
+
+	while (1) {
+		len = sizeof(buffer);
+		type = rds_message_next_extension(hdr, &pos, &buffer, &len);
+		if (type == RDS_EXTHDR_NONE)
+			break;
+		/* Process extension header here */
+		switch (type) {
+		case RDS_EXTHDR_RDMA:
+			rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0);
+			break;
+
+		case RDS_EXTHDR_RDMA_DEST:
+			/* We ignore the size for now. We could stash it
+			 * somewhere and use it for error checking. */
+			inc->i_usercopy.rdma_cookie = rds_rdma_make_cookie(
+					be32_to_cpu(buffer.rdma_dest.h_rdma_rkey),
+					be32_to_cpu(buffer.rdma_dest.h_rdma_offset));
+
+			break;
+		}
+	}
+}
+
+static void rds_recv_hs_exthdrs(struct rds_header *hdr,
+				struct rds_connection *conn)
+{
+	unsigned int pos = 0, type, len;
+	union {
+		struct rds_ext_header_version version;
+		__be16 rds_npaths;
+		__be32 rds_gen_num;
+		u8 dummy;
+	} buffer;
+	bool new_with_sport_idx = false;
+	u32 new_peer_gen_num = 0;
+	int new_npaths;
+	bool fan_out;
+
+	new_npaths = conn->c_npaths;
+
+	while (1) {
+		len = sizeof(buffer);
+		type = rds_message_next_extension(hdr, &pos, &buffer, &len);
+		if (type == RDS_EXTHDR_NONE)
+			break;
+		/* Process extension header here */
+		switch (type) {
+		case RDS_EXTHDR_NPATHS:
+			new_npaths = min_t(int, RDS_MPATH_WORKERS,
+					   be16_to_cpu(buffer.rds_npaths));
+			break;
+		case RDS_EXTHDR_GEN_NUM:
+			new_peer_gen_num = be32_to_cpu(buffer.rds_gen_num);
+			break;
+		case RDS_EXTHDR_SPORT_IDX:
+			new_with_sport_idx = true;
+			break;
+		default:
+			pr_warn_ratelimited("ignoring unknown exthdr type "
+					     "0x%x\n", type);
+		}
+	}
+
+	conn->c_with_sport_idx = new_with_sport_idx;
+
+	if (new_npaths > 1 && new_npaths != conn->c_npaths) {
+		/* We're about to fan-out.
+		 * Make sure that messages from cp_index#0
+		 * are sent prior to handling other lanes.
+		 */
+		struct rds_conn_path *cp0 = conn->c_path;
+		unsigned long flags;
+
+		spin_lock_irqsave(&cp0->cp_lock, flags);
+		conn->c_cp0_mprds_catchup_tx_seq = cp0->cp_next_tx_seq;
+		spin_unlock_irqrestore(&cp0->cp_lock, flags);
+		fan_out = true;
+	} else {
+		fan_out = false;
+	}
+
+	/* if RDS_EXTHDR_NPATHS was not found, default to a single-path */
+	conn->c_npaths = max_t(int, new_npaths, 1);
+
+	conn->c_ping_triggered = 0;
+	rds_conn_peer_gen_update(conn, new_peer_gen_num);
+
+	if (conn->c_npaths > 1 &&
+	    conn->c_trans->conn_slots_available)
+		conn->c_trans->conn_slots_available(conn, fan_out);
+}
+
+/* rds_start_mprds() will synchronously start multiple paths when appropriate.
+ * The scheme is based on the following rules:
+ *
+ * 1. rds_sendmsg on first connect attempt sends the probe ping, with the
+ *    sender's npaths (s_npaths)
+ * 2. rcvr of probe-ping knows the mprds_paths = min(s_npaths, r_npaths). It
+ *    sends back a probe-pong with r_npaths. After that, if rcvr is the
+ *    smaller ip addr, it starts rds_conn_path_connect_if_down on all
+ *    mprds_paths.
+ * 3. sender gets woken up, and can move to rds_conn_path_connect_if_down.
+ *    If it is the smaller ipaddr, rds_conn_path_connect_if_down can be
+ *    called after reception of the probe-pong on all mprds_paths.
+ *    Otherwise (sender of probe-ping is not the smaller ip addr): just call
+ *    rds_conn_path_connect_if_down on the hashed path. (see rule 4)
+ * 4. rds_connect_worker must only trigger a connection if laddr < faddr.
+ * 5. sender may end up queuing the packet on the cp. will get sent out later.
+ *    when connection is completed.
+ */
+static void rds_start_mprds(struct rds_connection *conn)
+{
+	int i;
+	struct rds_conn_path *cp;
+
+	if (conn->c_npaths > 1 &&
+	    rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) < 0) {
+		for (i = 0; i < conn->c_npaths; i++) {
+			cp = &conn->c_path[i];
+			rds_conn_path_connect_if_down(cp);
+		}
+	}
+}
+
+/*
+ * The transport must make sure that this is serialized against other
+ * rx and conn reset on this specific conn.
+ *
+ * We currently assert that only one fragmented message will be sent
+ * down a connection at a time.  This lets us reassemble in the conn
+ * instead of per-flow which means that we don't have to go digging through
+ * flows to tear down partial reassembly progress on conn failure and
+ * we save flow lookup and locking for each frag arrival.  It does mean
+ * that small messages will wait behind large ones.  Fragmenting at all
+ * is only to reduce the memory consumption of pre-posted buffers.
+ *
+ * The caller passes in saddr and daddr instead of us getting it from the
+ * conn.  This lets loopback, who only has one conn for both directions,
+ * tell us which roles the addrs in the conn are playing for this message.
+ */
+void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
+		       struct in6_addr *daddr,
+		       struct rds_incoming *inc, gfp_t gfp)
+{
+	struct rds_sock *rs = NULL;
+	struct sock *sk;
+	unsigned long flags;
+	struct rds_conn_path *cp;
+
+	inc->i_conn = conn;
+	inc->i_rx_jiffies = jiffies;
+	if (conn->c_trans->t_mp_capable)
+		cp = inc->i_conn_path;
+	else
+		cp = &conn->c_path[0];
+
+	rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
+		 "flags 0x%x rx_jiffies %lu\n", conn,
+		 (unsigned long long)cp->cp_next_rx_seq,
+		 inc,
+		 (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
+		 be32_to_cpu(inc->i_hdr.h_len),
+		 be16_to_cpu(inc->i_hdr.h_sport),
+		 be16_to_cpu(inc->i_hdr.h_dport),
+		 inc->i_hdr.h_flags,
+		 inc->i_rx_jiffies);
+
+	/*
+	 * Sequence numbers should only increase.  Messages get their
+	 * sequence number as they're queued in a sending conn.  They
+	 * can be dropped, though, if the sending socket is closed before
+	 * they hit the wire.  So sequence numbers can skip forward
+	 * under normal operation.  They can also drop back in the conn
+	 * failover case as previously sent messages are resent down the
+	 * new instance of a conn.  We drop those, otherwise we have
+	 * to assume that the next valid seq does not come after a
+	 * hole in the fragment stream.
+	 *
+	 * The headers don't give us a way to realize if fragments of
+	 * a message have been dropped.  We assume that frags that arrive
+	 * to a flow are part of the current message on the flow that is
+	 * being reassembled.  This means that senders can't drop messages
+	 * from the sending conn until all their frags are sent.
+	 *
+	 * XXX we could spend more on the wire to get more robust failure
+	 * detection, arguably worth it to avoid data corruption.
+	 */
+	if (be64_to_cpu(inc->i_hdr.h_sequence) < cp->cp_next_rx_seq &&
+	    (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
+		rds_stats_inc(s_recv_drop_old_seq);
+		goto out;
+	}
+	cp->cp_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
+
+	if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
+		if (inc->i_hdr.h_sport == 0) {
+			rdsdebug("ignore ping with 0 sport from %pI6c\n",
+				 saddr);
+			goto out;
+		}
+		rds_stats_inc(s_recv_ping);
+		rds_send_pong(cp, inc->i_hdr.h_sport);
+		/* if this is a handshake ping, start multipath if necessary */
+		if (RDS_HS_PROBE(be16_to_cpu(inc->i_hdr.h_sport),
+				 be16_to_cpu(inc->i_hdr.h_dport))) {
+			rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
+			rds_start_mprds(cp->cp_conn);
+		}
+		goto out;
+	}
+
+	if (be16_to_cpu(inc->i_hdr.h_dport) ==  RDS_FLAG_PROBE_PORT &&
+	    inc->i_hdr.h_sport == 0) {
+		rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
+		/* if this is a handshake pong, start multipath if necessary */
+		rds_start_mprds(cp->cp_conn);
+		wake_up(&cp->cp_conn->c_hs_waitq);
+		goto out;
+	}
+
+	rs = rds_find_bound(daddr, inc->i_hdr.h_dport, conn->c_bound_if);
+	if (!rs) {
+		rds_stats_inc(s_recv_drop_no_sock);
+		goto out;
+	}
+
+	/*
+	 * rds_find_bound() uses a global (netns-agnostic) hash table.
+	 * An RDS connection created in netns A can match a socket bound
+	 * in the init netns, delivering inc cross-netns with inc->i_conn
+	 * pointing into netns A.  When cleanup_net() then frees that conn,
+	 * any subsequent dereference of inc->i_conn is a use-after-free.
+	 * Drop the inc if the receiving socket lives in a different netns.
+	 */
+	if (!net_eq(sock_net(rds_rs_to_sk(rs)), rds_conn_net(conn))) {
+		rds_stats_inc(s_recv_drop_no_sock);
+		rds_sock_put(rs);
+		rs = NULL;
+		goto out;
+	}
+
+	/* Process extension headers */
+	rds_recv_incoming_exthdrs(inc, rs);
+
+	/* We can be racing with rds_release() which marks the socket dead. */
+	sk = rds_rs_to_sk(rs);
+
+	/* serialize with rds_release -> sock_orphan */
+	write_lock_irqsave(&rs->rs_recv_lock, flags);
+	if (!sock_flag(sk, SOCK_DEAD)) {
+		rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
+		rds_stats_inc(s_recv_queued);
+		rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
+				      be32_to_cpu(inc->i_hdr.h_len),
+				      inc->i_hdr.h_dport);
+		if (sock_flag(sk, SOCK_RCVTSTAMP))
+			inc->i_usercopy.rx_tstamp = ktime_get_real();
+		rds_inc_addref(inc);
+		inc->i_rx_lat_trace[RDS_MSG_RX_END] = local_clock();
+		list_add_tail(&inc->i_item, &rs->rs_recv_queue);
+		__rds_wake_sk_sleep(sk);
+	} else {
+		rds_stats_inc(s_recv_drop_dead_sock);
+	}
+	write_unlock_irqrestore(&rs->rs_recv_lock, flags);
+
+out:
+	if (rs)
+		rds_sock_put(rs);
+}
+EXPORT_SYMBOL_GPL(rds_recv_incoming);
+
+/*
+ * be very careful here.  This is being called as the condition in
+ * wait_event_*() needs to cope with being called many times.
+ */
+static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
+{
+	unsigned long flags;
+
+	if (!*inc) {
+		read_lock_irqsave(&rs->rs_recv_lock, flags);
+		if (!list_empty(&rs->rs_recv_queue)) {
+			*inc = list_entry(rs->rs_recv_queue.next,
+					  struct rds_incoming,
+					  i_item);
+			rds_inc_addref(*inc);
+		}
+		read_unlock_irqrestore(&rs->rs_recv_lock, flags);
+	}
+
+	return *inc != NULL;
+}
+
+static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc,
+			    int drop)
+{
+	struct sock *sk = rds_rs_to_sk(rs);
+	int ret = 0;
+	unsigned long flags;
+	struct rds_incoming *to_drop = NULL;
+
+	write_lock_irqsave(&rs->rs_recv_lock, flags);
+	if (!list_empty(&inc->i_item)) {
+		ret = 1;
+		if (drop) {
+			/* XXX make sure this i_conn is reliable */
+			rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
+					      -be32_to_cpu(inc->i_hdr.h_len),
+					      inc->i_hdr.h_dport);
+			list_del_init(&inc->i_item);
+			to_drop = inc;
+		}
+	}
+	write_unlock_irqrestore(&rs->rs_recv_lock, flags);
+
+	if (to_drop)
+		rds_inc_put(to_drop);
+
+	rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop);
+	return ret;
+}
+
+/*
+ * Pull errors off the error queue.
+ * If msghdr is NULL, we will just purge the error queue.
+ */
+int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
+{
+	struct rds_notifier *notifier;
+	struct rds_rdma_notify cmsg;
+	unsigned int count = 0, max_messages = ~0U;
+	unsigned long flags;
+	LIST_HEAD(copy);
+	int err = 0;
+
+	memset(&cmsg, 0, sizeof(cmsg));	/* fill holes with zero */
+
+	/* put_cmsg copies to user space and thus may sleep. We can't do this
+	 * with rs_lock held, so first grab as many notifications as we can stuff
+	 * in the user provided cmsg buffer. We don't try to copy more, to avoid
+	 * losing notifications - except when the buffer is so small that it wouldn't
+	 * even hold a single notification. Then we give him as much of this single
+	 * msg as we can squeeze in, and set MSG_CTRUNC.
+	 */
+	if (msghdr) {
+		max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg));
+		if (!max_messages)
+			max_messages = 1;
+	}
+
+	spin_lock_irqsave(&rs->rs_lock, flags);
+	while (!list_empty(&rs->rs_notify_queue) && count < max_messages) {
+		notifier = list_entry(rs->rs_notify_queue.next,
+				struct rds_notifier, n_list);
+		list_move(&notifier->n_list, &copy);
+		count++;
+	}
+	spin_unlock_irqrestore(&rs->rs_lock, flags);
+
+	if (!count)
+		return 0;
+
+	while (!list_empty(&copy)) {
+		notifier = list_entry(copy.next, struct rds_notifier, n_list);
+
+		if (msghdr) {
+			cmsg.user_token = notifier->n_user_token;
+			cmsg.status = notifier->n_status;
+
+			err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
+				       sizeof(cmsg), &cmsg);
+			if (err)
+				break;
+		}
+
+		list_del_init(&notifier->n_list);
+		kfree(notifier);
+	}
+
+	/* If we bailed out because of an error in put_cmsg,
+	 * we may be left with one or more notifications that we
+	 * didn't process. Return them to the head of the list. */
+	if (!list_empty(&copy)) {
+		spin_lock_irqsave(&rs->rs_lock, flags);
+		list_splice(&copy, &rs->rs_notify_queue);
+		spin_unlock_irqrestore(&rs->rs_lock, flags);
+	}
+
+	return err;
+}
+
+/*
+ * Queue a congestion notification
+ */
+static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr)
+{
+	uint64_t notify = rs->rs_cong_notify;
+	unsigned long flags;
+	int err;
+
+	err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE,
+			sizeof(notify), &notify);
+	if (err)
+		return err;
+
+	spin_lock_irqsave(&rs->rs_lock, flags);
+	rs->rs_cong_notify &= ~notify;
+	spin_unlock_irqrestore(&rs->rs_lock, flags);
+
+	return 0;
+}
+
+/*
+ * Receive any control messages.
+ */
+static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg,
+			 struct rds_sock *rs)
+{
+	int ret = 0;
+
+	if (inc->i_usercopy.rdma_cookie) {
+		ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
+				sizeof(inc->i_usercopy.rdma_cookie),
+				&inc->i_usercopy.rdma_cookie);
+		if (ret)
+			goto out;
+	}
+
+	if ((inc->i_usercopy.rx_tstamp != 0) &&
+	    sock_flag(rds_rs_to_sk(rs), SOCK_RCVTSTAMP)) {
+		struct __kernel_old_timeval tv =
+			ns_to_kernel_old_timeval(inc->i_usercopy.rx_tstamp);
+
+		if (!sock_flag(rds_rs_to_sk(rs), SOCK_TSTAMP_NEW)) {
+			ret = put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
+				       sizeof(tv), &tv);
+		} else {
+			struct __kernel_sock_timeval sk_tv;
+
+			sk_tv.tv_sec = tv.tv_sec;
+			sk_tv.tv_usec = tv.tv_usec;
+
+			ret = put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
+				       sizeof(sk_tv), &sk_tv);
+		}
+
+		if (ret)
+			goto out;
+	}
+
+	if (rs->rs_rx_traces) {
+		struct rds_cmsg_rx_trace t;
+		int i, j;
+
+		memset(&t, 0, sizeof(t));
+		inc->i_rx_lat_trace[RDS_MSG_RX_CMSG] = local_clock();
+		t.rx_traces =  rs->rs_rx_traces;
+		for (i = 0; i < rs->rs_rx_traces; i++) {
+			j = rs->rs_rx_trace[i];
+			t.rx_trace_pos[i] = j;
+			t.rx_trace[i] = inc->i_rx_lat_trace[j + 1] -
+					  inc->i_rx_lat_trace[j];
+		}
+
+		ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RXPATH_LATENCY,
+			       sizeof(t), &t);
+		if (ret)
+			goto out;
+	}
+
+out:
+	return ret;
+}
+
+static bool rds_recvmsg_zcookie(struct rds_sock *rs, struct msghdr *msg)
+{
+	struct rds_msg_zcopy_queue *q = &rs->rs_zcookie_queue;
+	struct rds_msg_zcopy_info *info = NULL;
+	struct rds_zcopy_cookies *done;
+	unsigned long flags;
+
+	if (!msg->msg_control)
+		return false;
+
+	if (!sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY) ||
+	    msg->msg_controllen < CMSG_SPACE(sizeof(*done)))
+		return false;
+
+	spin_lock_irqsave(&q->lock, flags);
+	if (!list_empty(&q->zcookie_head)) {
+		info = list_entry(q->zcookie_head.next,
+				  struct rds_msg_zcopy_info, rs_zcookie_next);
+		list_del(&info->rs_zcookie_next);
+	}
+	spin_unlock_irqrestore(&q->lock, flags);
+	if (!info)
+		return false;
+	done = &info->zcookies;
+	if (put_cmsg(msg, SOL_RDS, RDS_CMSG_ZCOPY_COMPLETION, sizeof(*done),
+		     done)) {
+		spin_lock_irqsave(&q->lock, flags);
+		list_add(&info->rs_zcookie_next, &q->zcookie_head);
+		spin_unlock_irqrestore(&q->lock, flags);
+		return false;
+	}
+	kfree(info);
+	return true;
+}
+
+int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
+		int msg_flags)
+{
+	struct sock *sk = sock->sk;
+	struct rds_sock *rs = rds_sk_to_rs(sk);
+	long timeo;
+	int ret = 0, nonblock = msg_flags & MSG_DONTWAIT;
+	DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
+	DECLARE_SOCKADDR(struct sockaddr_in *, sin, msg->msg_name);
+	struct rds_incoming *inc = NULL;
+
+	/* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */
+	timeo = sock_rcvtimeo(sk, nonblock);
+
+	rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo);
+
+	if (msg_flags & MSG_OOB)
+		goto out;
+	if (msg_flags & MSG_ERRQUEUE)
+		return sock_recv_errqueue(sk, msg, size, SOL_IP, IP_RECVERR);
+
+	while (1) {
+		/* If there are pending notifications, do those - and nothing else */
+		if (!list_empty(&rs->rs_notify_queue)) {
+			ret = rds_notify_queue_get(rs, msg);
+			break;
+		}
+
+		if (rs->rs_cong_notify) {
+			ret = rds_notify_cong(rs, msg);
+			break;
+		}
+
+		if (!rds_next_incoming(rs, &inc)) {
+			if (nonblock) {
+				bool reaped = rds_recvmsg_zcookie(rs, msg);
+
+				ret = reaped ?  0 : -EAGAIN;
+				break;
+			}
+
+			timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
+					(!list_empty(&rs->rs_notify_queue) ||
+					 rs->rs_cong_notify ||
+					 rds_next_incoming(rs, &inc)), timeo);
+			rdsdebug("recvmsg woke inc %p timeo %ld\n", inc,
+				 timeo);
+			if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
+				continue;
+
+			ret = timeo;
+			if (ret == 0)
+				ret = -ETIMEDOUT;
+			break;
+		}
+
+		rdsdebug("copying inc %p from %pI6c:%u to user\n", inc,
+			 &inc->i_conn->c_faddr,
+			 ntohs(inc->i_hdr.h_sport));
+		ret = inc->i_conn->c_trans->inc_copy_to_user(inc, &msg->msg_iter);
+		if (ret < 0)
+			break;
+
+		/*
+		 * if the message we just copied isn't at the head of the
+		 * recv queue then someone else raced us to return it, try
+		 * to get the next message.
+		 */
+		if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) {
+			rds_inc_put(inc);
+			inc = NULL;
+			rds_stats_inc(s_recv_deliver_raced);
+			iov_iter_revert(&msg->msg_iter, ret);
+			continue;
+		}
+
+		if (ret < be32_to_cpu(inc->i_hdr.h_len)) {
+			if (msg_flags & MSG_TRUNC)
+				ret = be32_to_cpu(inc->i_hdr.h_len);
+			msg->msg_flags |= MSG_TRUNC;
+		}
+
+		if (rds_cmsg_recv(inc, msg, rs)) {
+			ret = -EFAULT;
+			break;
+		}
+		rds_recvmsg_zcookie(rs, msg);
+
+		rds_stats_inc(s_recv_delivered);
+
+		if (msg->msg_name) {
+			if (ipv6_addr_v4mapped(&inc->i_saddr)) {
+				sin->sin_family = AF_INET;
+				sin->sin_port = inc->i_hdr.h_sport;
+				sin->sin_addr.s_addr =
+				    inc->i_saddr.s6_addr32[3];
+				memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
+				msg->msg_namelen = sizeof(*sin);
+			} else {
+				sin6->sin6_family = AF_INET6;
+				sin6->sin6_port = inc->i_hdr.h_sport;
+				sin6->sin6_addr = inc->i_saddr;
+				sin6->sin6_flowinfo = 0;
+				sin6->sin6_scope_id = rs->rs_bound_scope_id;
+				msg->msg_namelen = sizeof(*sin6);
+			}
+		}
+		break;
+	}
+
+	if (inc)
+		rds_inc_put(inc);
+
+out:
+	return ret;
+}
+
+/*
+ * The socket is being shut down and we're asked to drop messages that were
+ * queued for recvmsg.  The caller has unbound the socket so the receive path
+ * won't queue any more incoming fragments or messages on the socket.
+ */
+void rds_clear_recv_queue(struct rds_sock *rs)
+{
+	struct sock *sk = rds_rs_to_sk(rs);
+	struct rds_incoming *inc, *tmp;
+	unsigned long flags;
+	LIST_HEAD(to_drop);
+
+	write_lock_irqsave(&rs->rs_recv_lock, flags);
+	list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) {
+		rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
+				      -be32_to_cpu(inc->i_hdr.h_len),
+				      inc->i_hdr.h_dport);
+		list_move(&inc->i_item, &to_drop);
+	}
+	write_unlock_irqrestore(&rs->rs_recv_lock, flags);
+
+	list_for_each_entry_safe(inc, tmp, &to_drop, i_item) {
+		list_del_init(&inc->i_item);
+		rds_inc_put(inc);
+	}
+}
+
+/*
+ * inc->i_saddr isn't used here because it is only set in the receive
+ * path.
+ */
+void rds_inc_info_copy(struct rds_incoming *inc,
+		       struct rds_info_iterator *iter,
+		       __be32 saddr, __be32 daddr, int flip)
+{
+	struct rds_info_message minfo;
+
+	minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence);
+	minfo.len = be32_to_cpu(inc->i_hdr.h_len);
+	minfo.tos = inc->i_conn->c_tos;
+
+	if (flip) {
+		minfo.laddr = daddr;
+		minfo.faddr = saddr;
+		minfo.lport = inc->i_hdr.h_dport;
+		minfo.fport = inc->i_hdr.h_sport;
+	} else {
+		minfo.laddr = saddr;
+		minfo.faddr = daddr;
+		minfo.lport = inc->i_hdr.h_sport;
+		minfo.fport = inc->i_hdr.h_dport;
+	}
+
+	minfo.flags = 0;
+
+	rds_info_copy(iter, &minfo, sizeof(minfo));
+}
+
+#if IS_ENABLED(CONFIG_IPV6)
+void rds6_inc_info_copy(struct rds_incoming *inc,
+			struct rds_info_iterator *iter,
+			struct in6_addr *saddr, struct in6_addr *daddr,
+			int flip)
+{
+	struct rds6_info_message minfo6;
+
+	minfo6.seq = be64_to_cpu(inc->i_hdr.h_sequence);
+	minfo6.len = be32_to_cpu(inc->i_hdr.h_len);
+	minfo6.tos = inc->i_conn->c_tos;
+
+	if (flip) {
+		minfo6.laddr = *daddr;
+		minfo6.faddr = *saddr;
+		minfo6.lport = inc->i_hdr.h_dport;
+		minfo6.fport = inc->i_hdr.h_sport;
+	} else {
+		minfo6.laddr = *saddr;
+		minfo6.faddr = *daddr;
+		minfo6.lport = inc->i_hdr.h_sport;
+		minfo6.fport = inc->i_hdr.h_dport;
+	}
+
+	minfo6.flags = 0;
+
+	rds_info_copy(iter, &minfo6, sizeof(minfo6));
+}
+#endif

linux-7.2.3/net/rds/stats.c

@@ -0,0 +1,155 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/percpu.h>
+#include <linux/seq_file.h>
+#include <linux/proc_fs.h>
+#include <linux/export.h>
+
+#include "rds.h"
+
+DEFINE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
+EXPORT_PER_CPU_SYMBOL_GPL(rds_stats);
+
+/* :.,$s/unsigned long\>.*\<s_\(.*\);/"\1",/g */
+
+static const char *const rds_stat_names[] = {
+	"conn_reset",
+	"recv_drop_bad_checksum",
+	"recv_drop_old_seq",
+	"recv_drop_no_sock",
+	"recv_drop_dead_sock",
+	"recv_deliver_raced",
+	"recv_delivered",
+	"recv_queued",
+	"recv_immediate_retry",
+	"recv_delayed_retry",
+	"recv_ack_required",
+	"recv_rdma_bytes",
+	"recv_ping",
+	"send_queue_empty",
+	"send_queue_full",
+	"send_lock_contention",
+	"send_lock_queue_raced",
+	"send_immediate_retry",
+	"send_delayed_retry",
+	"send_drop_acked",
+	"send_ack_required",
+	"send_queued",
+	"send_rdma",
+	"send_rdma_bytes",
+	"send_pong",
+	"page_remainder_hit",
+	"page_remainder_miss",
+	"copy_to_user",
+	"copy_from_user",
+	"cong_update_queued",
+	"cong_update_received",
+	"cong_send_error",
+	"cong_send_blocked",
+	"recv_bytes_added_to_sock",
+	"recv_bytes_freed_fromsock",
+	"send_stuck_rm",
+	"mprds_catchup_tx0_retries",
+};
+
+void rds_stats_info_copy(struct rds_info_iterator *iter,
+			 uint64_t *values, const char *const *names, size_t nr)
+{
+	struct rds_info_counter ctr;
+	size_t i;
+
+	for (i = 0; i < nr; i++) {
+		BUG_ON(strlen(names[i]) >= sizeof(ctr.name));
+		strscpy_pad(ctr.name, names[i]);
+		ctr.value = values[i];
+
+		rds_info_copy(iter, &ctr, sizeof(ctr));
+	}
+}
+EXPORT_SYMBOL_GPL(rds_stats_info_copy);
+
+/*
+ * This gives global counters across all the transports.  The strings
+ * are copied in so that the tool doesn't need knowledge of the specific
+ * stats that we're exporting.  Some are pretty implementation dependent
+ * and may change over time.  That doesn't stop them from being useful.
+ *
+ * This is the only function in the chain that knows about the byte granular
+ * length in userspace.  It converts it to number of stat entries that the
+ * rest of the functions operate in.
+ */
+static void rds_stats_info(struct socket *sock, unsigned int len,
+			   struct rds_info_iterator *iter,
+			   struct rds_info_lengths *lens)
+{
+	struct rds_statistics stats = {0, };
+	uint64_t *src;
+	uint64_t *sum;
+	size_t i;
+	int cpu;
+	unsigned int avail;
+
+	avail = len / sizeof(struct rds_info_counter);
+
+	if (avail < ARRAY_SIZE(rds_stat_names)) {
+		avail = 0;
+		goto trans;
+	}
+
+	for_each_online_cpu(cpu) {
+		src = (uint64_t *)&(per_cpu(rds_stats, cpu));
+		sum = (uint64_t *)&stats;
+		for (i = 0; i < sizeof(stats) / sizeof(uint64_t); i++)
+			*(sum++) += *(src++);
+	}
+
+	rds_stats_info_copy(iter, (uint64_t *)&stats, rds_stat_names,
+			    ARRAY_SIZE(rds_stat_names));
+	avail -= ARRAY_SIZE(rds_stat_names);
+
+trans:
+	lens->each = sizeof(struct rds_info_counter);
+	lens->nr = rds_trans_stats_info_copy(iter, avail) +
+		   ARRAY_SIZE(rds_stat_names);
+}
+
+void rds_stats_exit(void)
+{
+	rds_info_deregister_func(RDS_INFO_COUNTERS, rds_stats_info);
+}
+
+int rds_stats_init(void)
+{
+	rds_info_register_func(RDS_INFO_COUNTERS, rds_stats_info);
+	return 0;
+}

linux-7.2.3/net/rds/sysctl.c

@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/sysctl.h>
+#include <linux/proc_fs.h>
+
+#include "rds.h"
+
+static struct ctl_table_header *rds_sysctl_reg_table;
+
+static unsigned long rds_sysctl_reconnect_min = 1;
+static unsigned long rds_sysctl_reconnect_max = ~0UL;
+
+unsigned long rds_sysctl_reconnect_min_jiffies;
+unsigned long rds_sysctl_reconnect_max_jiffies = HZ;
+
+unsigned int  rds_sysctl_max_unacked_packets = 8;
+unsigned int  rds_sysctl_max_unacked_bytes = (16 << 20);
+
+unsigned int rds_sysctl_ping_enable = 1;
+
+static struct ctl_table rds_sysctl_rds_table[] = {
+	{
+		.procname       = "reconnect_min_delay_ms",
+		.data		= &rds_sysctl_reconnect_min_jiffies,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = proc_doulongvec_ms_jiffies_minmax,
+		.extra1		= &rds_sysctl_reconnect_min,
+		.extra2		= &rds_sysctl_reconnect_max_jiffies,
+	},
+	{
+		.procname       = "reconnect_max_delay_ms",
+		.data		= &rds_sysctl_reconnect_max_jiffies,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = proc_doulongvec_ms_jiffies_minmax,
+		.extra1		= &rds_sysctl_reconnect_min_jiffies,
+		.extra2		= &rds_sysctl_reconnect_max,
+	},
+	{
+		.procname	= "max_unacked_packets",
+		.data		= &rds_sysctl_max_unacked_packets,
+		.maxlen         = sizeof(int),
+		.mode           = 0644,
+		.proc_handler   = proc_dointvec,
+	},
+	{
+		.procname	= "max_unacked_bytes",
+		.data		= &rds_sysctl_max_unacked_bytes,
+		.maxlen         = sizeof(int),
+		.mode           = 0644,
+		.proc_handler   = proc_dointvec,
+	},
+	{
+		.procname	= "ping_enable",
+		.data		= &rds_sysctl_ping_enable,
+		.maxlen         = sizeof(int),
+		.mode           = 0644,
+		.proc_handler   = proc_dointvec,
+	},
+};
+
+void rds_sysctl_exit(void)
+{
+	unregister_net_sysctl_table(rds_sysctl_reg_table);
+}
+
+int rds_sysctl_init(void)
+{
+	rds_sysctl_reconnect_min = msecs_to_jiffies(1);
+	rds_sysctl_reconnect_min_jiffies = rds_sysctl_reconnect_min;
+
+	rds_sysctl_reg_table =
+		register_net_sysctl(&init_net, "net/rds", rds_sysctl_rds_table);
+	if (!rds_sysctl_reg_table)
+		return -ENOMEM;
+	return 0;
+}

linux-7.2.3/net/rds/tcp.c

@@ -0,0 +1,804 @@
+/*
+ * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/in.h>
+#include <linux/module.h>
+#include <net/tcp.h>
+#include <net/net_namespace.h>
+#include <net/netns/generic.h>
+#include <net/addrconf.h>
+
+#include "rds.h"
+#include "tcp.h"
+
+/* only for info exporting */
+static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
+static LIST_HEAD(rds_tcp_tc_list);
+
+/* Track rds_tcp_connection structs so they can be cleaned up */
+static DEFINE_SPINLOCK(rds_tcp_conn_lock);
+static LIST_HEAD(rds_tcp_conn_list);
+static atomic_t rds_tcp_unloading = ATOMIC_INIT(0);
+
+static struct kmem_cache *rds_tcp_conn_slab;
+
+static int rds_tcp_sndbuf_handler(const struct ctl_table *ctl, int write,
+				  void *buffer, size_t *lenp, loff_t *fpos);
+static int rds_tcp_rcvbuf_handler(const struct ctl_table *ctl, int write,
+				  void *buffer, size_t *lenp, loff_t *fpos);
+
+static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
+static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
+
+static struct ctl_table rds_tcp_sysctl_table[] = {
+#define	RDS_TCP_SNDBUF	0
+	{
+		.procname       = "rds_tcp_sndbuf",
+		/* data is per-net pointer */
+		.maxlen         = sizeof(int),
+		.mode           = 0644,
+		.proc_handler   = rds_tcp_sndbuf_handler,
+		.extra1		= &rds_tcp_min_sndbuf,
+	},
+#define	RDS_TCP_RCVBUF	1
+	{
+		.procname       = "rds_tcp_rcvbuf",
+		/* data is per-net pointer */
+		.maxlen         = sizeof(int),
+		.mode           = 0644,
+		.proc_handler   = rds_tcp_rcvbuf_handler,
+		.extra1		= &rds_tcp_min_rcvbuf,
+	},
+};
+
+u32 rds_tcp_write_seq(struct rds_tcp_connection *tc)
+{
+	/* seq# of the last byte of data in tcp send buffer */
+	return tcp_sk(tc->t_sock->sk)->write_seq;
+}
+
+u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
+{
+	return tcp_sk(tc->t_sock->sk)->snd_una;
+}
+
+void rds_tcp_restore_callbacks(struct socket *sock,
+			       struct rds_tcp_connection *tc)
+{
+	rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
+	write_lock_bh(&sock->sk->sk_callback_lock);
+
+	/* done under the callback_lock to serialize with write_space */
+	spin_lock(&rds_tcp_tc_list_lock);
+	list_del_init(&tc->t_list_item);
+	spin_unlock(&rds_tcp_tc_list_lock);
+
+	tc->t_sock = NULL;
+
+	sock->sk->sk_write_space = tc->t_orig_write_space;
+	sock->sk->sk_data_ready = tc->t_orig_data_ready;
+	sock->sk->sk_state_change = tc->t_orig_state_change;
+	sock->sk->sk_user_data = NULL;
+
+	write_unlock_bh(&sock->sk->sk_callback_lock);
+}
+
+/*
+ * rds_tcp_reset_callbacks() switches the to the new sock and
+ * returns the existing tc->t_sock.
+ *
+ * The only functions that set tc->t_sock are rds_tcp_set_callbacks
+ * and rds_tcp_reset_callbacks.  Send and receive trust that
+ * it is set.  The absence of RDS_CONN_UP bit protects those paths
+ * from being called while it isn't set.
+ */
+void rds_tcp_reset_callbacks(struct socket *sock,
+			     struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+	struct socket *osock = tc->t_sock;
+
+	if (!osock)
+		goto newsock;
+
+	/* Need to resolve a duelling SYN between peers.
+	 * We have an outstanding SYN to this peer, which may
+	 * potentially have transitioned to the RDS_CONN_UP state,
+	 * so we must quiesce any send threads before resetting
+	 * cp_transport_data. We quiesce these threads by setting
+	 * cp_state to something other than RDS_CONN_UP, and then
+	 * waiting for any existing threads in rds_send_xmit to
+	 * complete release_in_xmit(). (Subsequent threads entering
+	 * rds_send_xmit() will bail on !rds_conn_up().
+	 *
+	 * However an incoming syn-ack at this point would end up
+	 * marking the conn as RDS_CONN_UP, and would again permit
+	 * rds_send_xmi() threads through, so ideally we would
+	 * synchronize on RDS_CONN_UP after lock_sock(), but cannot
+	 * do that: waiting on !RDS_IN_XMIT after lock_sock() may
+	 * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
+	 * would not get set. As a result, we set c_state to
+	 * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
+	 * cannot mark rds_conn_path_up() in the window before lock_sock()
+	 */
+	atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
+	wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
+	/* reset receive side state for rds_tcp_data_recv() for osock  */
+	cancel_delayed_work_sync(&cp->cp_send_w);
+	cancel_delayed_work_sync(&cp->cp_recv_w);
+	lock_sock(osock->sk);
+	if (tc->t_tinc) {
+		rds_inc_put(&tc->t_tinc->ti_inc);
+		tc->t_tinc = NULL;
+	}
+	tc->t_tinc_hdr_rem = sizeof(struct rds_header);
+	tc->t_tinc_data_rem = 0;
+	rds_tcp_restore_callbacks(osock, tc);
+	release_sock(osock->sk);
+	sock_release(osock);
+newsock:
+	rds_send_path_reset(cp);
+	lock_sock(sock->sk);
+	rds_tcp_set_callbacks(sock, cp);
+	release_sock(sock->sk);
+}
+
+/* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
+ * above rds_tcp_reset_callbacks for notes about synchronization
+ * with data path
+ */
+void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+
+	rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
+	write_lock_bh(&sock->sk->sk_callback_lock);
+
+	/* done under the callback_lock to serialize with write_space.
+	 * Set t_sock inside rds_tcp_tc_list_lock so readers walking
+	 * rds_tcp_tc_list under the same lock cannot observe an
+	 * entry whose t_sock is NULL.
+	 */
+	spin_lock(&rds_tcp_tc_list_lock);
+	tc->t_sock = sock;
+	list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
+	spin_unlock(&rds_tcp_tc_list_lock);
+
+	/* accepted sockets need our listen data ready undone */
+	if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
+		sock->sk->sk_data_ready = sock->sk->sk_user_data;
+	if (!tc->t_rtn)
+		tc->t_rtn = net_generic(sock_net(sock->sk), rds_tcp_netid);
+	tc->t_cpath = cp;
+	tc->t_orig_data_ready = sock->sk->sk_data_ready;
+	tc->t_orig_write_space = sock->sk->sk_write_space;
+	tc->t_orig_state_change = sock->sk->sk_state_change;
+
+	sock->sk->sk_user_data = cp;
+	sock->sk->sk_data_ready = rds_tcp_data_ready;
+	sock->sk->sk_write_space = rds_tcp_write_space;
+	sock->sk->sk_state_change = rds_tcp_state_change;
+
+	write_unlock_bh(&sock->sk->sk_callback_lock);
+}
+
+/* Handle RDS_INFO_TCP_SOCKETS socket option.  It only returns IPv4
+ * connections for backward compatibility.
+ */
+static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
+			    struct rds_info_iterator *iter,
+			    struct rds_info_lengths *lens)
+{
+	struct net *net = sock_net(rds_sock->sk);
+	struct rds_info_tcp_socket tsinfo;
+	struct rds_tcp_connection *tc;
+	unsigned int copied = 0;
+	unsigned int cnt = 0;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
+
+	/* First pass: count entries visible in the caller's netns. */
+	list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
+		if (tc->t_cpath->cp_conn->c_isv6)
+			continue;
+		if (!net_eq(rds_conn_net(tc->t_cpath->cp_conn), net))
+			continue;
+		cnt++;
+	}
+
+	if (len / sizeof(tsinfo) < cnt)
+		goto out;
+
+	list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
+		struct inet_sock *inet = inet_sk(tc->t_sock->sk);
+
+		if (copied >= cnt)
+			break;
+		if (tc->t_cpath->cp_conn->c_isv6)
+			continue;
+		/* Only show connections in the caller's netns. */
+		if (!net_eq(rds_conn_net(tc->t_cpath->cp_conn), net))
+			continue;
+
+		tsinfo.local_addr = inet->inet_saddr;
+		tsinfo.local_port = inet->inet_sport;
+		tsinfo.peer_addr = inet->inet_daddr;
+		tsinfo.peer_port = inet->inet_dport;
+
+		tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
+		tsinfo.data_rem = tc->t_tinc_data_rem;
+		tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
+		tsinfo.last_expected_una = tc->t_last_expected_una;
+		tsinfo.last_seen_una = tc->t_last_seen_una;
+		tsinfo.tos = tc->t_cpath->cp_conn->c_tos;
+
+		rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
+		copied++;
+	}
+	cnt = copied;
+
+out:
+	lens->nr = cnt;
+	lens->each = sizeof(tsinfo);
+
+	spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
+}
+
+#if IS_ENABLED(CONFIG_IPV6)
+/* Handle RDS6_INFO_TCP_SOCKETS socket option. It returns both IPv4 and
+ * IPv6 connections. IPv4 connection address is returned in an IPv4 mapped
+ * address.
+ */
+static void rds6_tcp_tc_info(struct socket *sock, unsigned int len,
+			     struct rds_info_iterator *iter,
+			     struct rds_info_lengths *lens)
+{
+	struct net *net = sock_net(sock->sk);
+	struct rds6_info_tcp_socket tsinfo6;
+	struct rds_tcp_connection *tc;
+	unsigned int copied = 0;
+	unsigned int cnt = 0;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
+
+	/* First pass: count entries visible in the caller's netns. */
+	list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
+		if (!net_eq(rds_conn_net(tc->t_cpath->cp_conn), net))
+			continue;
+		cnt++;
+	}
+
+	if (len / sizeof(tsinfo6) < cnt)
+		goto out;
+
+	list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
+		struct sock *sk = tc->t_sock->sk;
+		struct inet_sock *inet = inet_sk(sk);
+
+		if (copied >= cnt)
+			break;
+		/* Only show connections in the caller's netns. */
+		if (!net_eq(rds_conn_net(tc->t_cpath->cp_conn), net))
+			continue;
+
+		tsinfo6.local_addr = sk->sk_v6_rcv_saddr;
+		tsinfo6.local_port = inet->inet_sport;
+		tsinfo6.peer_addr = sk->sk_v6_daddr;
+		tsinfo6.peer_port = inet->inet_dport;
+
+		tsinfo6.hdr_rem = tc->t_tinc_hdr_rem;
+		tsinfo6.data_rem = tc->t_tinc_data_rem;
+		tsinfo6.last_sent_nxt = tc->t_last_sent_nxt;
+		tsinfo6.last_expected_una = tc->t_last_expected_una;
+		tsinfo6.last_seen_una = tc->t_last_seen_una;
+
+		rds_info_copy(iter, &tsinfo6, sizeof(tsinfo6));
+		copied++;
+	}
+	cnt = copied;
+
+out:
+	lens->nr = cnt;
+	lens->each = sizeof(tsinfo6);
+
+	spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
+}
+#endif
+
+int rds_tcp_laddr_check(struct net *net, const struct in6_addr *addr,
+			__u32 scope_id)
+{
+	struct net_device *dev = NULL;
+#if IS_ENABLED(CONFIG_IPV6)
+	int ret;
+#endif
+
+	if (ipv6_addr_v4mapped(addr)) {
+		if (inet_addr_type(net, addr->s6_addr32[3]) == RTN_LOCAL)
+			return 0;
+		return -EADDRNOTAVAIL;
+	}
+
+	/* If the scope_id is specified, check only those addresses
+	 * hosted on the specified interface.
+	 */
+	rcu_read_lock();
+	if (scope_id != 0) {
+		dev = dev_get_by_index_rcu(net, scope_id);
+		/* scope_id is not valid... */
+		if (!dev) {
+			rcu_read_unlock();
+			return -EADDRNOTAVAIL;
+		}
+	}
+#if IS_ENABLED(CONFIG_IPV6)
+	if (ipv6_mod_enabled()) {
+		ret = ipv6_chk_addr(net, addr, dev, 0);
+		if (ret) {
+			rcu_read_unlock();
+			return 0;
+		}
+	}
+#endif
+	rcu_read_unlock();
+	return -EADDRNOTAVAIL;
+}
+
+static void rds_tcp_conn_free(void *arg)
+{
+	struct rds_tcp_connection *tc = arg;
+	unsigned long flags;
+
+	rdsdebug("freeing tc %p\n", tc);
+
+	spin_lock_irqsave(&rds_tcp_conn_lock, flags);
+	if (!tc->t_tcp_node_detached)
+		list_del(&tc->t_tcp_node);
+	spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
+
+	kmem_cache_free(rds_tcp_conn_slab, tc);
+}
+
+static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
+{
+	struct rds_tcp_connection *tc;
+	int i, j;
+	int ret = 0;
+
+	for (i = 0; i < RDS_MPATH_WORKERS; i++) {
+		tc = kmem_cache_zalloc(rds_tcp_conn_slab, gfp);
+		if (!tc) {
+			ret = -ENOMEM;
+			goto fail;
+		}
+		mutex_init(&tc->t_conn_path_lock);
+		tc->t_sock = NULL;
+		tc->t_rtn = NULL;
+		tc->t_tinc = NULL;
+		tc->t_tinc_hdr_rem = sizeof(struct rds_header);
+		tc->t_tinc_data_rem = 0;
+		init_waitqueue_head(&tc->t_recv_done_waitq);
+
+		conn->c_path[i].cp_transport_data = tc;
+		tc->t_cpath = &conn->c_path[i];
+		tc->t_tcp_node_detached = true;
+
+		rdsdebug("rds_conn_path [%d] tc %p\n", i,
+			 conn->c_path[i].cp_transport_data);
+	}
+	spin_lock_irq(&rds_tcp_conn_lock);
+	for (i = 0; i < RDS_MPATH_WORKERS; i++) {
+		tc = conn->c_path[i].cp_transport_data;
+		tc->t_tcp_node_detached = false;
+		list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
+	}
+	spin_unlock_irq(&rds_tcp_conn_lock);
+fail:
+	if (ret) {
+		for (j = 0; j < i; j++)
+			rds_tcp_conn_free(conn->c_path[j].cp_transport_data);
+	}
+	return ret;
+}
+
+static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
+{
+	struct rds_tcp_connection *tc, *_tc;
+
+	list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
+		if (tc->t_cpath->cp_conn == conn)
+			return true;
+	}
+	return false;
+}
+
+static void rds_tcp_set_unloading(void)
+{
+	atomic_set(&rds_tcp_unloading, 1);
+}
+
+static bool rds_tcp_is_unloading(struct rds_connection *conn)
+{
+	return atomic_read(&rds_tcp_unloading) != 0;
+}
+
+static void rds_tcp_destroy_conns(void)
+{
+	struct rds_tcp_connection *tc, *_tc;
+	LIST_HEAD(tmp_list);
+
+	/* avoid calling conn_destroy with irqs off */
+	spin_lock_irq(&rds_tcp_conn_lock);
+	list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
+		if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
+			list_move_tail(&tc->t_tcp_node, &tmp_list);
+	}
+	spin_unlock_irq(&rds_tcp_conn_lock);
+
+	list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
+		rds_conn_destroy(tc->t_cpath->cp_conn);
+}
+
+static void rds_tcp_exit(void);
+
+static u8 rds_tcp_get_tos_map(u8 tos)
+{
+	/* all user tos mapped to default 0 for TCP transport */
+	return 0;
+}
+
+struct rds_transport rds_tcp_transport = {
+	.laddr_check		= rds_tcp_laddr_check,
+	.xmit_path_prepare	= rds_tcp_xmit_path_prepare,
+	.xmit_path_complete	= rds_tcp_xmit_path_complete,
+	.xmit			= rds_tcp_xmit,
+	.recv_path		= rds_tcp_recv_path,
+	.conn_alloc		= rds_tcp_conn_alloc,
+	.conn_free		= rds_tcp_conn_free,
+	.conn_slots_available	= rds_tcp_conn_slots_available,
+	.conn_path_connect	= rds_tcp_conn_path_connect,
+	.conn_path_shutdown	= rds_tcp_conn_path_shutdown,
+	.inc_copy_to_user	= rds_tcp_inc_copy_to_user,
+	.inc_free		= rds_tcp_inc_free,
+	.stats_info_copy	= rds_tcp_stats_info_copy,
+	.exit			= rds_tcp_exit,
+	.get_tos_map		= rds_tcp_get_tos_map,
+	.t_owner		= THIS_MODULE,
+	.t_name			= "tcp",
+	.t_type			= RDS_TRANS_TCP,
+	.t_prefer_loopback	= 1,
+	.t_mp_capable		= 1,
+	.t_unloading		= rds_tcp_is_unloading,
+};
+
+int rds_tcp_netid;
+
+/* All module specific customizations to the RDS-TCP socket should be done in
+ * rds_tcp_tune() and applied after socket creation.
+ */
+bool rds_tcp_tune(struct socket *sock)
+{
+	struct sock *sk = sock->sk;
+	struct net *net = sock_net(sk);
+	struct rds_tcp_net *rtn;
+
+	tcp_sock_set_nodelay(sock->sk);
+	/* TCP timer functions might access net namespace even after
+	 * a process which created this net namespace terminated.
+	 */
+	if (!sk->sk_net_refcnt) {
+		if (!maybe_get_net(net))
+			return false;
+		/*
+		 * sk_net_refcnt_upgrade() must be called before lock_sock()
+		 * because it does a GFP_KERNEL allocation, which can trigger
+		 * fs_reclaim and create a circular lock dependency with the
+		 * socket lock.  The fields it modifies (sk_net_refcnt,
+		 * ns_tracker) are not accessed by any concurrent code path
+		 * at this point.
+		 */
+		sk_net_refcnt_upgrade(sk);
+		put_net(net);
+	}
+	lock_sock(sk);
+	rtn = net_generic(net, rds_tcp_netid);
+	if (rtn->sndbuf_size > 0) {
+		sk->sk_sndbuf = rtn->sndbuf_size;
+		sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
+	}
+	if (rtn->rcvbuf_size > 0) {
+		sk->sk_rcvbuf = rtn->rcvbuf_size;
+		sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
+	}
+	release_sock(sk);
+	return true;
+}
+
+static void rds_tcp_accept_worker(struct work_struct *work)
+{
+	struct rds_tcp_net *rtn = container_of(work,
+					       struct rds_tcp_net,
+					       rds_tcp_accept_w);
+
+	while (rds_tcp_accept_one(rtn) == 0)
+		cond_resched();
+}
+
+void rds_tcp_accept_work(struct rds_tcp_net *rtn)
+{
+	queue_work(rds_wq, &rtn->rds_tcp_accept_w);
+}
+
+static __net_init int rds_tcp_init_net(struct net *net)
+{
+	struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
+	struct ctl_table *tbl;
+	int err = 0;
+
+	memset(rtn, 0, sizeof(*rtn));
+
+	mutex_init(&rtn->rds_tcp_accept_lock);
+
+	/* {snd, rcv}buf_size default to 0, which implies we let the
+	 * stack pick the value, and permit auto-tuning of buffer size.
+	 */
+	if (net == &init_net) {
+		tbl = rds_tcp_sysctl_table;
+	} else {
+		tbl = kmemdup(rds_tcp_sysctl_table,
+			      sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
+		if (!tbl) {
+			pr_warn("could not set allocate sysctl table\n");
+			return -ENOMEM;
+		}
+		rtn->ctl_table = tbl;
+	}
+	tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
+	tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
+	rtn->rds_tcp_sysctl = register_net_sysctl_sz(net, "net/rds/tcp", tbl,
+						     ARRAY_SIZE(rds_tcp_sysctl_table));
+	if (!rtn->rds_tcp_sysctl) {
+		pr_warn("could not register sysctl\n");
+		err = -ENOMEM;
+		goto fail;
+	}
+
+#if IS_ENABLED(CONFIG_IPV6)
+	rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, true);
+#else
+	rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false);
+#endif
+	if (!rtn->rds_tcp_listen_sock) {
+		pr_warn("could not set up IPv6 listen sock\n");
+
+#if IS_ENABLED(CONFIG_IPV6)
+		/* Try IPv4 as some systems disable IPv6 */
+		rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false);
+		if (!rtn->rds_tcp_listen_sock) {
+#endif
+			unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
+			rtn->rds_tcp_sysctl = NULL;
+			err = -EAFNOSUPPORT;
+			goto fail;
+#if IS_ENABLED(CONFIG_IPV6)
+		}
+#endif
+	}
+	INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
+	return 0;
+
+fail:
+	if (net != &init_net)
+		kfree(tbl);
+	return err;
+}
+
+static void rds_tcp_kill_sock(struct net *net)
+{
+	struct rds_tcp_connection *tc, *_tc;
+	LIST_HEAD(tmp_list);
+	struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
+	struct socket *lsock = rtn->rds_tcp_listen_sock;
+
+	rtn->rds_tcp_listen_sock = NULL;
+	rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
+	if (rtn->rds_tcp_accepted_sock)
+		sock_release(rtn->rds_tcp_accepted_sock);
+	spin_lock_irq(&rds_tcp_conn_lock);
+	list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
+		struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
+
+		if (net != c_net)
+			continue;
+		if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) {
+			list_move_tail(&tc->t_tcp_node, &tmp_list);
+		} else {
+			list_del(&tc->t_tcp_node);
+			tc->t_tcp_node_detached = true;
+		}
+	}
+	spin_unlock_irq(&rds_tcp_conn_lock);
+	list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
+		rds_conn_destroy(tc->t_cpath->cp_conn);
+}
+
+static void __net_exit rds_tcp_exit_net(struct net *net)
+{
+	struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
+
+	if (rtn->rds_tcp_sysctl)
+		unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
+
+	if (net != &init_net)
+		kfree(rtn->ctl_table);
+
+	rds_tcp_kill_sock(net);
+}
+
+static struct pernet_operations rds_tcp_net_ops = {
+	.init = rds_tcp_init_net,
+	.exit = rds_tcp_exit_net,
+	.id = &rds_tcp_netid,
+	.size = sizeof(struct rds_tcp_net),
+};
+
+void *rds_tcp_listen_sock_def_readable(struct net *net)
+{
+	struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
+	struct socket *lsock = rtn->rds_tcp_listen_sock;
+
+	if (!lsock)
+		return NULL;
+
+	return lsock->sk->sk_user_data;
+}
+
+/* when sysctl is used to modify some kernel socket parameters,this
+ * function  resets the RDS connections in that netns  so that we can
+ * restart with new parameters.  The assumption is that such reset
+ * events are few and far-between.
+ */
+static void rds_tcp_sysctl_reset(struct net *net)
+{
+	struct rds_tcp_connection *tc, *_tc;
+
+	spin_lock_irq(&rds_tcp_conn_lock);
+	list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
+		struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
+
+		if (net != c_net || !tc->t_sock)
+			continue;
+
+		/* reconnect with new parameters */
+		rds_conn_path_drop(tc->t_cpath, false);
+	}
+	spin_unlock_irq(&rds_tcp_conn_lock);
+}
+
+static int rds_tcp_skbuf_handler(struct rds_tcp_net *rtn,
+				 const struct ctl_table *ctl, int write,
+				 void *buffer, size_t *lenp, loff_t *fpos)
+{
+	int err;
+
+	err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
+	if (err < 0) {
+		pr_warn("Invalid input. Must be >= %d\n",
+			*(int *)(ctl->extra1));
+		return err;
+	}
+
+	if (write && rtn->rds_tcp_listen_sock && rtn->rds_tcp_listen_sock->sk) {
+		struct net *net = sock_net(rtn->rds_tcp_listen_sock->sk);
+
+		rds_tcp_sysctl_reset(net);
+	}
+
+	return 0;
+}
+
+static int rds_tcp_sndbuf_handler(const struct ctl_table *ctl, int write,
+				  void *buffer, size_t *lenp, loff_t *fpos)
+{
+	struct rds_tcp_net *rtn = container_of(ctl->data, struct rds_tcp_net,
+					       sndbuf_size);
+
+	return rds_tcp_skbuf_handler(rtn, ctl, write, buffer, lenp, fpos);
+}
+
+static int rds_tcp_rcvbuf_handler(const struct ctl_table *ctl, int write,
+				  void *buffer, size_t *lenp, loff_t *fpos)
+{
+	struct rds_tcp_net *rtn = container_of(ctl->data, struct rds_tcp_net,
+					       rcvbuf_size);
+
+	return rds_tcp_skbuf_handler(rtn, ctl, write, buffer, lenp, fpos);
+}
+
+static void rds_tcp_exit(void)
+{
+	rds_tcp_set_unloading();
+	synchronize_rcu();
+	rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
+#if IS_ENABLED(CONFIG_IPV6)
+	rds_info_deregister_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info);
+#endif
+	unregister_pernet_device(&rds_tcp_net_ops);
+	rds_tcp_destroy_conns();
+	rds_trans_unregister(&rds_tcp_transport);
+	rds_tcp_recv_exit();
+	kmem_cache_destroy(rds_tcp_conn_slab);
+}
+module_exit(rds_tcp_exit);
+
+static int __init rds_tcp_init(void)
+{
+	int ret;
+
+	rds_tcp_conn_slab = KMEM_CACHE(rds_tcp_connection, 0);
+	if (!rds_tcp_conn_slab) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	ret = rds_tcp_recv_init();
+	if (ret)
+		goto out_slab;
+
+	ret = register_pernet_device(&rds_tcp_net_ops);
+	if (ret)
+		goto out_recv;
+
+	rds_trans_register(&rds_tcp_transport);
+
+	rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
+#if IS_ENABLED(CONFIG_IPV6)
+	rds_info_register_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info);
+#endif
+
+	goto out;
+out_recv:
+	rds_tcp_recv_exit();
+out_slab:
+	kmem_cache_destroy(rds_tcp_conn_slab);
+out:
+	return ret;
+}
+module_init(rds_tcp_init);
+
+MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
+MODULE_DESCRIPTION("RDS: TCP transport");
+MODULE_LICENSE("Dual BSD/GPL");

linux-7.2.3/net/rds/tcp.h

@@ -0,0 +1,120 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _RDS_TCP_H
+#define _RDS_TCP_H
+
+#define RDS_TCP_PORT	16385
+
+/* per-network namespace private data for this module */
+struct rds_tcp_net {
+	/* serialize "rds_tcp_accept_one" with "rds_tcp_accept_lock"
+	 * to protect "rds_tcp_accepted_sock"
+	 */
+	struct mutex		rds_tcp_accept_lock;
+	struct socket		*rds_tcp_listen_sock;
+	struct socket		*rds_tcp_accepted_sock;
+	struct work_struct	rds_tcp_accept_w;
+	struct ctl_table_header	*rds_tcp_sysctl;
+	const struct ctl_table	*ctl_table;
+	int			sndbuf_size;
+	int			rcvbuf_size;
+};
+
+struct rds_tcp_incoming {
+	struct rds_incoming	ti_inc;
+	struct sk_buff_head	ti_skb_list;
+};
+
+struct rds_tcp_connection {
+
+	struct list_head	t_tcp_node;
+	bool			t_tcp_node_detached;
+	struct rds_conn_path	*t_cpath;
+	/* t_conn_path_lock synchronizes the connection establishment between
+	 * rds_tcp_accept_one and rds_tcp_conn_path_connect
+	 */
+	struct mutex		t_conn_path_lock;
+	struct socket		*t_sock;
+	u32			t_client_port_group;
+	struct rds_tcp_net	*t_rtn;
+	void			*t_orig_write_space;
+	void			*t_orig_data_ready;
+	void			*t_orig_state_change;
+
+	struct rds_tcp_incoming	*t_tinc;
+	size_t			t_tinc_hdr_rem;
+	size_t			t_tinc_data_rem;
+
+	/* XXX error report? */
+	struct work_struct	t_conn_w;
+	struct work_struct	t_send_w;
+	struct work_struct	t_down_w;
+	struct work_struct	t_recv_w;
+
+	/* for info exporting only */
+	struct list_head	t_list_item;
+	u32			t_last_sent_nxt;
+	u32			t_last_expected_una;
+	u32			t_last_seen_una;
+
+	/* for rds_tcp_conn_path_shutdown */
+	wait_queue_head_t	t_recv_done_waitq;
+};
+
+struct rds_tcp_statistics {
+	uint64_t	s_tcp_data_ready_calls;
+	uint64_t	s_tcp_write_space_calls;
+	uint64_t	s_tcp_sndbuf_full;
+	uint64_t	s_tcp_connect_raced;
+	uint64_t	s_tcp_listen_closed_stale;
+};
+
+/* tcp.c */
+extern int rds_tcp_netid;
+bool rds_tcp_tune(struct socket *sock);
+void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp);
+void rds_tcp_reset_callbacks(struct socket *sock, struct rds_conn_path *cp);
+void rds_tcp_restore_callbacks(struct socket *sock,
+			       struct rds_tcp_connection *tc);
+u32 rds_tcp_write_seq(struct rds_tcp_connection *tc);
+u32 rds_tcp_snd_una(struct rds_tcp_connection *tc);
+extern struct rds_transport rds_tcp_transport;
+void rds_tcp_accept_work(struct rds_tcp_net *rtn);
+int rds_tcp_laddr_check(struct net *net, const struct in6_addr *addr,
+			__u32 scope_id);
+/* tcp_connect.c */
+int rds_tcp_conn_path_connect(struct rds_conn_path *cp);
+void rds_tcp_conn_path_shutdown(struct rds_conn_path *conn);
+void rds_tcp_state_change(struct sock *sk);
+
+/* tcp_listen.c */
+struct socket *rds_tcp_listen_init(struct net *net, bool isv6);
+void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor);
+void rds_tcp_listen_data_ready(struct sock *sk);
+void rds_tcp_conn_slots_available(struct rds_connection *conn, bool fan_out);
+int rds_tcp_accept_one(struct rds_tcp_net *rtn);
+void rds_tcp_keepalive(struct socket *sock);
+void *rds_tcp_listen_sock_def_readable(struct net *net);
+
+/* tcp_recv.c */
+int rds_tcp_recv_init(void);
+void rds_tcp_recv_exit(void);
+void rds_tcp_data_ready(struct sock *sk);
+int rds_tcp_recv_path(struct rds_conn_path *cp);
+void rds_tcp_inc_free(struct rds_incoming *inc);
+int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
+
+/* tcp_send.c */
+void rds_tcp_xmit_path_prepare(struct rds_conn_path *cp);
+void rds_tcp_xmit_path_complete(struct rds_conn_path *cp);
+int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
+		 unsigned int hdr_off, unsigned int sg, unsigned int off);
+int rds_tcp_is_acked(struct rds_message *rm, uint64_t ack);
+void rds_tcp_write_space(struct sock *sk);
+
+/* tcp_stats.c */
+DECLARE_PER_CPU(struct rds_tcp_statistics, rds_tcp_stats);
+#define rds_tcp_stats_inc(member) rds_stats_inc_which(rds_tcp_stats, member)
+unsigned int rds_tcp_stats_info_copy(struct rds_info_iterator *iter,
+				     unsigned int avail);
+
+#endif

linux-7.2.3/net/rds/tcp_connect.c

@@ -0,0 +1,298 @@
+/*
+ * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <net/tcp.h>
+
+#include "rds.h"
+#include "tcp.h"
+
+void rds_tcp_state_change(struct sock *sk)
+{
+	void (*state_change)(struct sock *sk);
+	struct rds_conn_path *cp;
+	struct rds_tcp_connection *tc;
+
+	read_lock_bh(&sk->sk_callback_lock);
+	cp = sk->sk_user_data;
+	if (!cp) {
+		state_change = sk->sk_state_change;
+		goto out;
+	}
+	tc = cp->cp_transport_data;
+	state_change = tc->t_orig_state_change;
+
+	rdsdebug("sock %p state_change to %d\n", tc->t_sock, sk->sk_state);
+
+	switch (sk->sk_state) {
+	/* ignore connecting sockets as they make progress */
+	case TCP_SYN_SENT:
+	case TCP_SYN_RECV:
+		break;
+	case TCP_ESTABLISHED:
+		/* Force the peer to reconnect so that we have the
+		 * TCP ports going from <smaller-ip>.<transient> to
+		 * <larger-ip>.<RDS_TCP_PORT>. We avoid marking the
+		 * RDS connection as RDS_CONN_UP until the reconnect,
+		 * to avoid RDS datagram loss.
+		 */
+		if (rds_addr_cmp(&cp->cp_conn->c_laddr,
+				 &cp->cp_conn->c_faddr) >= 0 &&
+		    rds_conn_path_transition(cp, RDS_CONN_CONNECTING,
+					     RDS_CONN_ERROR)) {
+			rds_conn_path_drop(cp, false);
+		} else {
+			rds_connect_path_complete(cp, RDS_CONN_CONNECTING);
+		}
+		break;
+	case TCP_CLOSING:
+	case TCP_TIME_WAIT:
+		if (wq_has_sleeper(&tc->t_recv_done_waitq))
+			wake_up(&tc->t_recv_done_waitq);
+		break;
+	case TCP_CLOSE_WAIT:
+	case TCP_LAST_ACK:
+	case TCP_CLOSE:
+		if (wq_has_sleeper(&tc->t_recv_done_waitq))
+			wake_up(&tc->t_recv_done_waitq);
+		rds_conn_path_drop(cp, false);
+		break;
+	default:
+		break;
+	}
+out:
+	read_unlock_bh(&sk->sk_callback_lock);
+	state_change(sk);
+}
+
+int rds_tcp_conn_path_connect(struct rds_conn_path *cp)
+{
+	struct socket *sock = NULL;
+	struct sockaddr_in6 sin6;
+	struct sockaddr_in sin;
+	struct sockaddr *addr;
+	int port_low, port_high, port;
+	int port_groups, groups_left;
+	int addrlen;
+	bool isv6;
+	int ret;
+	struct rds_connection *conn = cp->cp_conn;
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+
+	/* for multipath rds,we only trigger the connection after
+	 * the handshake probe has determined the number of paths.
+	 */
+	if (cp->cp_index > 0 && cp->cp_conn->c_npaths < 2)
+		return -EAGAIN;
+
+	mutex_lock(&tc->t_conn_path_lock);
+
+	if (rds_conn_path_up(cp)) {
+		mutex_unlock(&tc->t_conn_path_lock);
+		return 0;
+	}
+	if (ipv6_addr_v4mapped(&conn->c_laddr)) {
+		ret = sock_create_kern(rds_conn_net(conn), PF_INET,
+				       SOCK_STREAM, IPPROTO_TCP, &sock);
+		isv6 = false;
+	} else {
+		ret = sock_create_kern(rds_conn_net(conn), PF_INET6,
+				       SOCK_STREAM, IPPROTO_TCP, &sock);
+		isv6 = true;
+	}
+
+	if (ret < 0)
+		goto out;
+
+	if (!rds_tcp_tune(sock)) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	if (isv6) {
+		sin6.sin6_family = AF_INET6;
+		sin6.sin6_addr = conn->c_laddr;
+		sin6.sin6_port = 0;
+		sin6.sin6_flowinfo = 0;
+		sin6.sin6_scope_id = conn->c_dev_if;
+		addr = (struct sockaddr *)&sin6;
+		addrlen = sizeof(sin6);
+	} else {
+		sin.sin_family = AF_INET;
+		sin.sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
+		sin.sin_port = 0;
+		addr = (struct sockaddr *)&sin;
+		addrlen = sizeof(sin);
+	}
+
+	/* encode cp->cp_index in lowest bits of source-port */
+	inet_get_local_port_range(rds_conn_net(conn), &port_low, &port_high);
+	port_low = ALIGN(port_low, RDS_MPATH_WORKERS);
+	port_groups = (port_high - port_low + 1) / RDS_MPATH_WORKERS;
+	ret = -EADDRINUSE;
+	groups_left = port_groups;
+	while (groups_left-- > 0 && ret) {
+		if (++tc->t_client_port_group >= port_groups)
+			tc->t_client_port_group = 0;
+		port =  port_low +
+			tc->t_client_port_group * RDS_MPATH_WORKERS +
+			cp->cp_index;
+
+		if (isv6)
+			sin6.sin6_port = htons(port);
+		else
+			sin.sin_port = htons(port);
+		ret = kernel_bind(sock, (struct sockaddr_unsized *)addr,
+				  addrlen);
+	}
+	if (ret) {
+		rdsdebug("bind failed with %d at address %pI6c\n",
+			 ret, &conn->c_laddr);
+		goto out;
+	}
+
+	if (isv6) {
+		sin6.sin6_family = AF_INET6;
+		sin6.sin6_addr = conn->c_faddr;
+		sin6.sin6_port = htons(RDS_TCP_PORT);
+		sin6.sin6_flowinfo = 0;
+		sin6.sin6_scope_id = conn->c_dev_if;
+		addr = (struct sockaddr *)&sin6;
+		addrlen = sizeof(sin6);
+	} else {
+		sin.sin_family = AF_INET;
+		sin.sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
+		sin.sin_port = htons(RDS_TCP_PORT);
+		addr = (struct sockaddr *)&sin;
+		addrlen = sizeof(sin);
+	}
+
+	/*
+	 * once we call connect() we can start getting callbacks and they
+	 * own the socket
+	 */
+	rds_tcp_set_callbacks(sock, cp);
+	ret = kernel_connect(sock, (struct sockaddr_unsized *)addr, addrlen, O_NONBLOCK);
+
+	rdsdebug("connect to address %pI6c returned %d\n", &conn->c_faddr, ret);
+	if (ret == -EINPROGRESS)
+		ret = 0;
+	if (ret == 0) {
+		rds_tcp_keepalive(sock);
+		sock = NULL;
+	} else {
+		rds_tcp_restore_callbacks(sock, cp->cp_transport_data);
+	}
+
+out:
+	mutex_unlock(&tc->t_conn_path_lock);
+	if (sock)
+		sock_release(sock);
+	return ret;
+}
+
+/*
+ * Before killing the tcp socket this needs to serialize with callbacks.  The
+ * caller has already grabbed the sending sem so we're serialized with other
+ * senders.
+ *
+ * TCP calls the callbacks with the sock lock so we hold it while we reset the
+ * callbacks to those set by TCP.  Our callbacks won't execute again once we
+ * hold the sock lock.
+ */
+void rds_tcp_conn_path_shutdown(struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+	struct socket *sock = tc->t_sock;
+	struct sock *sk;
+	unsigned int rounds;
+
+	rdsdebug("shutting down conn %p tc %p sock %p\n",
+		 cp->cp_conn, tc, sock);
+
+	if (sock) {
+		sk = sock->sk;
+		if (rds_destroy_pending(cp->cp_conn))
+			sock_no_linger(sk);
+
+		sock->ops->shutdown(sock, SHUT_WR);
+
+		/* after sending FIN,
+		 * wait until we processed all incoming messages
+		 * and we're sure that there won't be any more:
+		 * i.e. state CLOSING, TIME_WAIT, CLOSE_WAIT,
+		 * LAST_ACK, or CLOSE (RFC 793).
+		 *
+		 * Give up waiting after 5 seconds and allow messages
+		 * to theoretically get dropped, if the TCP transition
+		 * didn't happen.
+		 */
+		rounds = 0;
+		do {
+			/* we need to ensure messages are dequeued here
+			 * since "rds_recv_worker" only dispatches messages
+			 * while the connection is still in RDS_CONN_UP
+			 * and there is no guarantee that "rds_tcp_data_ready"
+			 * was called nor that "sk_data_ready" still points to
+			 * it.
+			 */
+			rds_tcp_recv_path(cp);
+		} while (!wait_event_timeout(tc->t_recv_done_waitq,
+				(sk->sk_state == TCP_CLOSING ||
+				 sk->sk_state == TCP_TIME_WAIT ||
+				 sk->sk_state == TCP_CLOSE_WAIT ||
+				 sk->sk_state == TCP_LAST_ACK ||
+				 sk->sk_state == TCP_CLOSE) &&
+				skb_queue_empty_lockless(&sk->sk_receive_queue),
+				msecs_to_jiffies(100)) &&
+			 ++rounds < 50);
+		lock_sock(sk);
+
+		/* discard messages that the peer received already */
+		tc->t_last_seen_una = rds_tcp_snd_una(tc);
+		rds_send_path_drop_acked(cp, rds_tcp_snd_una(tc),
+					 rds_tcp_is_acked);
+
+		rds_tcp_restore_callbacks(sock, tc); /* tc->tc_sock = NULL */
+
+		release_sock(sk);
+		sock_release(sock);
+	}
+
+	if (tc->t_tinc) {
+		rds_inc_put(&tc->t_tinc->ti_inc);
+		tc->t_tinc = NULL;
+	}
+	tc->t_tinc_hdr_rem = sizeof(struct rds_header);
+	tc->t_tinc_data_rem = 0;
+}

linux-7.2.3/net/rds/tcp_listen.c

@@ -0,0 +1,464 @@
+/*
+ * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/gfp.h>
+#include <linux/in.h>
+#include <net/tcp.h>
+#include <trace/events/sock.h>
+#include <net/net_namespace.h>
+#include <net/netns/generic.h>
+
+#include "rds.h"
+#include "tcp.h"
+
+void rds_tcp_keepalive(struct socket *sock)
+{
+	/* values below based on xs_udp_default_timeout */
+	int keepidle = 5; /* send a probe 'keepidle' secs after last data */
+	int keepcnt = 5; /* number of unack'ed probes before declaring dead */
+
+	sock_set_keepalive(sock->sk);
+	tcp_sock_set_keepcnt(sock->sk, keepcnt);
+	tcp_sock_set_keepidle(sock->sk, keepidle);
+	/* KEEPINTVL is the interval between successive probes. We follow
+	 * the model in xs_tcp_finish_connecting() and re-use keepidle.
+	 */
+	tcp_sock_set_keepintvl(sock->sk, keepidle);
+}
+
+static int
+rds_tcp_get_peer_sport(struct socket *sock)
+{
+	struct sock *sk = sock->sk;
+
+	if (!sk)
+		return -1;
+
+	return ntohs(READ_ONCE(inet_sk(sk)->inet_dport));
+}
+
+/* rds_tcp_accept_one_path(): if accepting on cp_index > 0, make sure the
+ * client's ipaddr < server's ipaddr. Otherwise, close the accepted
+ * socket and force a reconnect from smaller -> larger ip addr. The reason
+ * we special case cp_index 0 is to allow the rds probe ping itself to itself
+ * get through efficiently.
+ */
+static struct rds_tcp_connection *
+rds_tcp_accept_one_path(struct rds_connection *conn, struct socket *sock)
+{
+	int sport, npaths, i_min, i_max, i;
+
+	if (conn->c_with_sport_idx)
+		/* cp->cp_index is encoded in lowest bits of source-port */
+		sport = rds_tcp_get_peer_sport(sock);
+	else
+		sport = -1;
+
+	npaths = max_t(int, 1, conn->c_npaths);
+
+	if (sport >= 0) {
+		i_min = sport % npaths;
+		i_max = i_min;
+	} else {
+		i_min = 0;
+		i_max = npaths - 1;
+	}
+
+	for (i = i_min; i <= i_max; i++) {
+		struct rds_conn_path *cp = &conn->c_path[i];
+
+		if (rds_conn_path_transition(cp, RDS_CONN_DOWN,
+					     RDS_CONN_CONNECTING))
+			return cp->cp_transport_data;
+	}
+
+	return NULL;
+}
+
+void rds_tcp_conn_slots_available(struct rds_connection *conn, bool fan_out)
+{
+	struct rds_tcp_connection *tc;
+	struct rds_tcp_net *rtn;
+	struct socket *sock;
+	int sport, npaths;
+
+	if (rds_destroy_pending(conn))
+		return;
+
+	tc = conn->c_path->cp_transport_data;
+	rtn = tc->t_rtn;
+	if (!rtn)
+		return;
+
+	sock = tc->t_sock;
+
+	/* During fan-out, check that the connection we already
+	 * accepted in slot#0 carried the proper source port modulo.
+	 */
+	if (fan_out && conn->c_with_sport_idx && sock &&
+	    rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) > 0) {
+		/* cp->cp_index is encoded in lowest bits of source-port */
+		sport = rds_tcp_get_peer_sport(sock);
+		npaths = max_t(int, 1, conn->c_npaths);
+		if (sport >= 0 && sport % npaths != 0)
+			/* peer initiated with a non-#0 lane first */
+			rds_conn_path_drop(conn->c_path, 0);
+	}
+
+	/* As soon as a connection went down,
+	 * it is safe to schedule a "rds_tcp_accept_one"
+	 * attempt even if there are no connections pending:
+	 * Function "rds_tcp_accept_one" won't block
+	 * but simply return -EAGAIN in that case.
+	 *
+	 * Doing so is necessary to address the case where an
+	 * incoming connection on "rds_tcp_listen_sock" is ready
+	 * to be accepted prior to a free slot being available:
+	 * the -ENOBUFS case in "rds_tcp_accept_one".
+	 */
+	rds_tcp_accept_work(rtn);
+}
+
+int rds_tcp_accept_one(struct rds_tcp_net *rtn)
+{
+	struct socket *listen_sock = rtn->rds_tcp_listen_sock;
+	struct socket *new_sock = NULL;
+	struct rds_connection *conn;
+	int ret;
+	struct inet_sock *inet;
+	struct rds_tcp_connection *rs_tcp = NULL;
+	int conn_state;
+	struct rds_conn_path *cp;
+	struct sock *sk;
+	struct in6_addr *my_addr, *peer_addr;
+#if !IS_ENABLED(CONFIG_IPV6)
+	struct in6_addr saddr, daddr;
+#endif
+	int dev_if = 0;
+
+	if (!listen_sock) /* module unload or netns delete in progress */
+		return -ENETUNREACH;
+
+	mutex_lock(&rtn->rds_tcp_accept_lock);
+	new_sock = rtn->rds_tcp_accepted_sock;
+	rtn->rds_tcp_accepted_sock = NULL;
+
+	if (!new_sock) {
+		ret = kernel_accept(listen_sock, &new_sock, O_NONBLOCK);
+		if (ret)
+			goto out;
+
+		rds_tcp_keepalive(new_sock);
+		if (!rds_tcp_tune(new_sock)) {
+			ret = -EINVAL;
+			goto out;
+		}
+	}
+
+	inet = inet_sk(new_sock->sk);
+
+#if IS_ENABLED(CONFIG_IPV6)
+	my_addr = &new_sock->sk->sk_v6_rcv_saddr;
+	peer_addr = &new_sock->sk->sk_v6_daddr;
+#else
+	ipv6_addr_set_v4mapped(inet->inet_saddr, &saddr);
+	ipv6_addr_set_v4mapped(inet->inet_daddr, &daddr);
+	my_addr = &saddr;
+	peer_addr = &daddr;
+#endif
+	rdsdebug("accepted family %d tcp %pI6c:%u -> %pI6c:%u\n",
+		 listen_sock->sk->sk_family,
+		 my_addr, ntohs(inet->inet_sport),
+		 peer_addr, ntohs(inet->inet_dport));
+
+#if IS_ENABLED(CONFIG_IPV6)
+	/* sk_bound_dev_if is not set if the peer address is not link local
+	 * address.  In this case, it happens that mcast_oif is set.  So
+	 * just use it.
+	 */
+	if ((ipv6_addr_type(my_addr) & IPV6_ADDR_LINKLOCAL) &&
+	    !(ipv6_addr_type(peer_addr) & IPV6_ADDR_LINKLOCAL)) {
+		struct ipv6_pinfo *inet6;
+
+		inet6 = inet6_sk(new_sock->sk);
+		dev_if = READ_ONCE(inet6->mcast_oif);
+	} else {
+		dev_if = new_sock->sk->sk_bound_dev_if;
+	}
+#endif
+
+	if (!rds_tcp_laddr_check(sock_net(listen_sock->sk), peer_addr, dev_if)) {
+		/* local address connection is only allowed via loopback */
+		ret = -EOPNOTSUPP;
+		goto out;
+	}
+
+	conn = rds_conn_create(sock_net(listen_sock->sk),
+			       my_addr, peer_addr,
+			       &rds_tcp_transport, 0, GFP_KERNEL, dev_if);
+
+	if (IS_ERR(conn)) {
+		ret = PTR_ERR(conn);
+		goto out;
+	}
+	/* An incoming SYN request came in, and TCP just accepted it.
+	 *
+	 * If the client reboots, this conn will need to be cleaned up.
+	 * rds_tcp_state_change() will do that cleanup
+	 */
+	if (rds_addr_cmp(&conn->c_faddr, &conn->c_laddr) < 0) {
+		/* Try to obtain a free connection slot.
+		 * If unsuccessful, we need to preserve "new_sock"
+		 * that we just accepted, since its "sk_receive_queue"
+		 * may contain messages already that have been acknowledged
+		 * to and discarded by the sender.
+		 * We must not throw those away!
+		 */
+		rs_tcp = rds_tcp_accept_one_path(conn, new_sock);
+		if (!rs_tcp) {
+			/* It's okay to stash "new_sock", since
+			 * "rds_tcp_conn_slots_available" triggers
+			 * "rds_tcp_accept_one" again as soon as one of the
+			 * connection slots becomes available again
+			 */
+			rtn->rds_tcp_accepted_sock = new_sock;
+			new_sock = NULL;
+			ret = -ENOBUFS;
+			goto out;
+		}
+	} else {
+		/* This connection request came from a peer with
+		 * a larger address.
+		 * Function "rds_tcp_state_change" makes sure
+		 * that the connection doesn't transition
+		 * to state "RDS_CONN_UP", and therefore
+		 * we should not have received any messages
+		 * on this socket yet.
+		 * This is the only case where it's okay to
+		 * not dequeue messages from "sk_receive_queue".
+		 */
+		if (conn->c_npaths <= 1)
+			rds_conn_path_connect_if_down(&conn->c_path[0]);
+		rs_tcp = NULL;
+		goto rst_nsk;
+	}
+
+	mutex_lock(&rs_tcp->t_conn_path_lock);
+	cp = rs_tcp->t_cpath;
+	conn_state = rds_conn_path_state(cp);
+	WARN_ON(conn_state == RDS_CONN_UP);
+	if (conn_state != RDS_CONN_CONNECTING && conn_state != RDS_CONN_ERROR) {
+		rds_conn_path_drop(cp, 0);
+		goto rst_nsk;
+	}
+	/* Save a local pointer to sk and hold a reference before setting
+	 * callbacks. Once callbacks are set, a concurrent
+	 * rds_tcp_conn_path_shutdown() may call sock_release(), which
+	 * sets new_sock->sk to NULL and drops a reference on sk.
+	 * The local pointer lets us safely access sk_state below even
+	 * if new_sock->sk has been nulled, and sock_hold() keeps sk
+	 * itself valid until we are done.
+	 */
+	sk = new_sock->sk;
+	sock_hold(sk);
+
+	if (rs_tcp->t_sock) {
+		/* Duelling SYN has been handled in rds_tcp_accept_one() */
+		rds_tcp_reset_callbacks(new_sock, cp);
+		/* rds_connect_path_complete() marks RDS_CONN_UP */
+		rds_connect_path_complete(cp, RDS_CONN_RESETTING);
+	} else {
+		rds_tcp_set_callbacks(new_sock, cp);
+		rds_connect_path_complete(cp, RDS_CONN_CONNECTING);
+	}
+
+	/* Since "rds_tcp_set_callbacks" happens this late
+	 * the connection may already have been closed without
+	 * "rds_tcp_state_change" doing its due diligence.
+	 *
+	 * If that's the case, we simply drop the path,
+	 * knowing that "rds_tcp_conn_path_shutdown" will
+	 * dequeue pending messages.
+	 */
+	if (READ_ONCE(sk->sk_state) == TCP_CLOSE_WAIT ||
+	    READ_ONCE(sk->sk_state) == TCP_LAST_ACK ||
+	    READ_ONCE(sk->sk_state) == TCP_CLOSE)
+		rds_conn_path_drop(cp, 0);
+	else
+		queue_delayed_work(cp->cp_wq, &cp->cp_recv_w, 0);
+
+	sock_put(sk);
+
+	new_sock = NULL;
+	ret = 0;
+	if (conn->c_npaths == 0)
+		rds_send_ping(cp->cp_conn, cp->cp_index);
+	goto out;
+rst_nsk:
+	/* reset the newly returned accept sock and bail.
+	 * It is safe to set linger on new_sock because the RDS connection
+	 * has not been brought up on new_sock, so no RDS-level data could
+	 * be pending on it. By setting linger, we achieve the side-effect
+	 * of avoiding TIME_WAIT state on new_sock.
+	 */
+	sock_no_linger(new_sock->sk);
+	kernel_sock_shutdown(new_sock, SHUT_RDWR);
+	ret = 0;
+out:
+	if (rs_tcp)
+		mutex_unlock(&rs_tcp->t_conn_path_lock);
+	if (new_sock)
+		sock_release(new_sock);
+
+	mutex_unlock(&rtn->rds_tcp_accept_lock);
+
+	return ret;
+}
+
+void rds_tcp_listen_data_ready(struct sock *sk)
+{
+	void (*ready)(struct sock *sk);
+
+	trace_sk_data_ready(sk);
+	rdsdebug("listen data ready sk %p\n", sk);
+
+	read_lock_bh(&sk->sk_callback_lock);
+	ready = sk->sk_user_data;
+	if (!ready) { /* check for teardown race */
+		ready = sk->sk_data_ready;
+		goto out;
+	}
+
+	/*
+	 * ->sk_data_ready is also called for a newly established child socket
+	 * before it has been accepted and the accepter has set up their
+	 * data_ready.. we only want to queue listen work for our listening
+	 * socket
+	 *
+	 * (*ready)() may be null if we are racing with netns delete, and
+	 * the listen socket is being torn down.
+	 */
+	if (sk->sk_state == TCP_LISTEN)
+		rds_tcp_accept_work(net_generic(sock_net(sk), rds_tcp_netid));
+	else
+		ready = rds_tcp_listen_sock_def_readable(sock_net(sk));
+
+out:
+	read_unlock_bh(&sk->sk_callback_lock);
+	if (ready)
+		ready(sk);
+}
+
+struct socket *rds_tcp_listen_init(struct net *net, bool isv6)
+{
+	struct socket *sock = NULL;
+	struct sockaddr_storage ss;
+	struct sockaddr_in6 *sin6;
+	struct sockaddr_in *sin;
+	int addr_len;
+	int ret;
+
+	ret = sock_create_kern(net, isv6 ? PF_INET6 : PF_INET, SOCK_STREAM,
+			       IPPROTO_TCP, &sock);
+	if (ret < 0) {
+		rdsdebug("could not create %s listener socket: %d\n",
+			 isv6 ? "IPv6" : "IPv4", ret);
+		goto out;
+	}
+
+	sock->sk->sk_reuse = SK_CAN_REUSE;
+	tcp_sock_set_nodelay(sock->sk);
+
+	write_lock_bh(&sock->sk->sk_callback_lock);
+	sock->sk->sk_user_data = sock->sk->sk_data_ready;
+	sock->sk->sk_data_ready = rds_tcp_listen_data_ready;
+	write_unlock_bh(&sock->sk->sk_callback_lock);
+
+	if (isv6) {
+		sin6 = (struct sockaddr_in6 *)&ss;
+		sin6->sin6_family = PF_INET6;
+		sin6->sin6_addr = in6addr_any;
+		sin6->sin6_port = htons(RDS_TCP_PORT);
+		sin6->sin6_scope_id = 0;
+		sin6->sin6_flowinfo = 0;
+		addr_len = sizeof(*sin6);
+	} else {
+		sin = (struct sockaddr_in *)&ss;
+		sin->sin_family = PF_INET;
+		sin->sin_addr.s_addr = htonl(INADDR_ANY);
+		sin->sin_port = htons(RDS_TCP_PORT);
+		addr_len = sizeof(*sin);
+	}
+
+	ret = kernel_bind(sock, (struct sockaddr_unsized *)&ss, addr_len);
+	if (ret < 0) {
+		rdsdebug("could not bind %s listener socket: %d\n",
+			 isv6 ? "IPv6" : "IPv4", ret);
+		goto out;
+	}
+
+	ret = sock->ops->listen(sock, 64);
+	if (ret < 0)
+		goto out;
+
+	return sock;
+out:
+	if (sock)
+		sock_release(sock);
+	return NULL;
+}
+
+void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor)
+{
+	struct sock *sk;
+
+	if (!sock)
+		return;
+
+	sk = sock->sk;
+
+	/* serialize with and prevent further callbacks */
+	lock_sock(sk);
+	write_lock_bh(&sk->sk_callback_lock);
+	if (sk->sk_user_data) {
+		sk->sk_data_ready = sk->sk_user_data;
+		sk->sk_user_data = NULL;
+	}
+	write_unlock_bh(&sk->sk_callback_lock);
+	release_sock(sk);
+
+	/* wait for accepts to stop and close the socket */
+	flush_workqueue(rds_wq);
+	flush_work(acceptor);
+	sock_release(sock);
+}

linux-7.2.3/net/rds/tcp_recv.c

@@ -0,0 +1,353 @@
+/*
+ * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <net/tcp.h>
+#include <trace/events/sock.h>
+
+#include "rds.h"
+#include "tcp.h"
+
+static struct kmem_cache *rds_tcp_incoming_slab;
+
+static void rds_tcp_inc_purge(struct rds_incoming *inc)
+{
+	struct rds_tcp_incoming *tinc;
+	tinc = container_of(inc, struct rds_tcp_incoming, ti_inc);
+	rdsdebug("purging tinc %p inc %p\n", tinc, inc);
+	skb_queue_purge(&tinc->ti_skb_list);
+}
+
+void rds_tcp_inc_free(struct rds_incoming *inc)
+{
+	struct rds_tcp_incoming *tinc;
+	tinc = container_of(inc, struct rds_tcp_incoming, ti_inc);
+	rds_tcp_inc_purge(inc);
+	rdsdebug("freeing tinc %p inc %p\n", tinc, inc);
+	kmem_cache_free(rds_tcp_incoming_slab, tinc);
+}
+
+/*
+ * this is pretty lame, but, whatever.
+ */
+int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
+{
+	struct rds_tcp_incoming *tinc;
+	struct sk_buff *skb;
+	int ret = 0;
+
+	if (!iov_iter_count(to))
+		goto out;
+
+	tinc = container_of(inc, struct rds_tcp_incoming, ti_inc);
+
+	skb_queue_walk(&tinc->ti_skb_list, skb) {
+		unsigned long to_copy, skb_off;
+		for (skb_off = 0; skb_off < skb->len; skb_off += to_copy) {
+			to_copy = iov_iter_count(to);
+			to_copy = min(to_copy, skb->len - skb_off);
+
+			if (skb_copy_datagram_iter(skb, skb_off, to, to_copy))
+				return -EFAULT;
+
+			rds_stats_add(s_copy_to_user, to_copy);
+			ret += to_copy;
+
+			if (!iov_iter_count(to))
+				goto out;
+		}
+	}
+out:
+	return ret;
+}
+
+/*
+ * We have a series of skbs that have fragmented pieces of the congestion
+ * bitmap.  They must add up to the exact size of the congestion bitmap.  We
+ * use the skb helpers to copy those into the pages that make up the in-memory
+ * congestion bitmap for the remote address of this connection.  We then tell
+ * the congestion core that the bitmap has been changed so that it can wake up
+ * sleepers.
+ *
+ * This is racing with sending paths which are using test_bit to see if the
+ * bitmap indicates that their recipient is congested.
+ */
+
+static void rds_tcp_cong_recv(struct rds_connection *conn,
+			      struct rds_tcp_incoming *tinc)
+{
+	struct sk_buff *skb;
+	unsigned int to_copy, skb_off;
+	unsigned int map_off;
+	unsigned int map_page;
+	struct rds_cong_map *map;
+	int ret;
+
+	/* catch completely corrupt packets */
+	if (be32_to_cpu(tinc->ti_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
+		return;
+
+	map_page = 0;
+	map_off = 0;
+	map = conn->c_fcong;
+
+	skb_queue_walk(&tinc->ti_skb_list, skb) {
+		skb_off = 0;
+		while (skb_off < skb->len) {
+			to_copy = min_t(unsigned int, PAGE_SIZE - map_off,
+					skb->len - skb_off);
+
+			BUG_ON(map_page >= RDS_CONG_MAP_PAGES);
+
+			/* only returns 0 or -error */
+			ret = skb_copy_bits(skb, skb_off,
+				(void *)map->m_page_addrs[map_page] + map_off,
+				to_copy);
+			BUG_ON(ret != 0);
+
+			skb_off += to_copy;
+			map_off += to_copy;
+			if (map_off == PAGE_SIZE) {
+				map_off = 0;
+				map_page++;
+			}
+		}
+	}
+
+	rds_cong_map_updated(map, ~(u64) 0);
+}
+
+struct rds_tcp_desc_arg {
+	struct rds_conn_path *conn_path;
+	gfp_t gfp;
+};
+
+static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb,
+			     unsigned int offset, size_t len)
+{
+	struct rds_tcp_desc_arg *arg = desc->arg.data;
+	struct rds_conn_path *cp = arg->conn_path;
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+	struct rds_tcp_incoming *tinc = tc->t_tinc;
+	struct sk_buff *clone;
+	size_t left = len, to_copy;
+
+	rdsdebug("tcp data tc %p skb %p offset %u len %zu\n", tc, skb, offset,
+		 len);
+
+	/*
+	 * tcp_read_sock() interprets partial progress as an indication to stop
+	 * processing.
+	 */
+	while (left) {
+		if (!tinc) {
+			tinc = kmem_cache_alloc(rds_tcp_incoming_slab,
+						arg->gfp);
+			if (!tinc) {
+				desc->error = -ENOMEM;
+				goto out;
+			}
+			tc->t_tinc = tinc;
+			rdsdebug("allocated tinc %p\n", tinc);
+			rds_inc_path_init(&tinc->ti_inc, cp,
+					  &cp->cp_conn->c_faddr);
+			tinc->ti_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
+					local_clock();
+
+			/*
+			 * XXX * we might be able to use the __ variants when
+			 * we've already serialized at a higher level.
+			 */
+			skb_queue_head_init(&tinc->ti_skb_list);
+		}
+
+		if (left && tc->t_tinc_hdr_rem) {
+			to_copy = min(tc->t_tinc_hdr_rem, left);
+			rdsdebug("copying %zu header from skb %p\n", to_copy,
+				 skb);
+			skb_copy_bits(skb, offset,
+				      (char *)&tinc->ti_inc.i_hdr +
+						sizeof(struct rds_header) -
+						tc->t_tinc_hdr_rem,
+				      to_copy);
+			tc->t_tinc_hdr_rem -= to_copy;
+			left -= to_copy;
+			offset += to_copy;
+
+			if (tc->t_tinc_hdr_rem == 0) {
+				/* could be 0 for a 0 len message */
+				tc->t_tinc_data_rem =
+					be32_to_cpu(tinc->ti_inc.i_hdr.h_len);
+				tinc->ti_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
+					local_clock();
+			}
+		}
+
+		if (left && tc->t_tinc_data_rem) {
+			to_copy = min(tc->t_tinc_data_rem, left);
+
+			clone = pskb_extract(skb, offset, to_copy, arg->gfp);
+			if (!clone) {
+				desc->error = -ENOMEM;
+				goto out;
+			}
+
+			skb_queue_tail(&tinc->ti_skb_list, clone);
+
+			rdsdebug("skb %p data %p len %d off %u to_copy %zu -> "
+				 "clone %p data %p len %d\n",
+				 skb, skb->data, skb->len, offset, to_copy,
+				 clone, clone->data, clone->len);
+
+			tc->t_tinc_data_rem -= to_copy;
+			left -= to_copy;
+			offset += to_copy;
+		}
+
+		if (tc->t_tinc_hdr_rem == 0 && tc->t_tinc_data_rem == 0) {
+			struct rds_connection *conn = cp->cp_conn;
+
+			if (tinc->ti_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
+				rds_tcp_cong_recv(conn, tinc);
+			else
+				rds_recv_incoming(conn, &conn->c_faddr,
+						  &conn->c_laddr,
+						  &tinc->ti_inc,
+						  arg->gfp);
+
+			tc->t_tinc_hdr_rem = sizeof(struct rds_header);
+			tc->t_tinc_data_rem = 0;
+			tc->t_tinc = NULL;
+			rds_inc_put(&tinc->ti_inc);
+			tinc = NULL;
+		}
+	}
+out:
+	rdsdebug("returning len %zu left %zu skb len %d rx queue depth %d\n",
+		 len, left, skb->len,
+		 skb_queue_len(&tc->t_sock->sk->sk_receive_queue));
+	return len - left;
+}
+
+/* the caller has to hold the sock lock */
+static int rds_tcp_read_sock(struct rds_conn_path *cp, gfp_t gfp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+	struct socket *sock = tc->t_sock;
+	read_descriptor_t desc;
+	struct rds_tcp_desc_arg arg;
+
+	/* It's like glib in the kernel! */
+	arg.conn_path = cp;
+	arg.gfp = gfp;
+	desc.arg.data = &arg;
+	desc.error = 0;
+	desc.count = 1; /* give more than one skb per call */
+
+	tcp_read_sock(sock->sk, &desc, rds_tcp_data_recv);
+	rdsdebug("tcp_read_sock for tc %p gfp %pGg returned %d\n", tc, &gfp,
+		 desc.error);
+
+	if (skb_queue_empty_lockless(&sock->sk->sk_receive_queue) &&
+	    wq_has_sleeper(&tc->t_recv_done_waitq))
+		wake_up(&tc->t_recv_done_waitq);
+
+	return desc.error;
+}
+
+/*
+ * We hold the sock lock to serialize our rds_tcp_recv->tcp_read_sock from
+ * data_ready.
+ *
+ * if we fail to allocate we're in trouble.. blindly wait some time before
+ * trying again to see if the VM can free up something for us.
+ */
+int rds_tcp_recv_path(struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+	struct socket *sock = tc->t_sock;
+	int ret = 0;
+
+	rdsdebug("recv worker path [%d] tc %p sock %p\n",
+		 cp->cp_index, tc, sock);
+
+	lock_sock(sock->sk);
+	ret = rds_tcp_read_sock(cp, GFP_KERNEL);
+	release_sock(sock->sk);
+
+	return ret;
+}
+
+void rds_tcp_data_ready(struct sock *sk)
+{
+	void (*ready)(struct sock *sk);
+	struct rds_conn_path *cp;
+	struct rds_tcp_connection *tc;
+
+	trace_sk_data_ready(sk);
+	rdsdebug("data ready sk %p\n", sk);
+
+	read_lock_bh(&sk->sk_callback_lock);
+	cp = sk->sk_user_data;
+	if (!cp) { /* check for teardown race */
+		ready = sk->sk_data_ready;
+		goto out;
+	}
+
+	tc = cp->cp_transport_data;
+	ready = tc->t_orig_data_ready;
+	rds_tcp_stats_inc(s_tcp_data_ready_calls);
+
+	if (rds_tcp_read_sock(cp, GFP_ATOMIC) == -ENOMEM) {
+		rcu_read_lock();
+		if (!rds_destroy_pending(cp->cp_conn))
+			queue_delayed_work(cp->cp_wq, &cp->cp_recv_w, 0);
+		rcu_read_unlock();
+	}
+out:
+	read_unlock_bh(&sk->sk_callback_lock);
+	ready(sk);
+}
+
+int rds_tcp_recv_init(void)
+{
+	rds_tcp_incoming_slab = KMEM_CACHE(rds_tcp_incoming, 0);
+	if (!rds_tcp_incoming_slab)
+		return -ENOMEM;
+	return 0;
+}
+
+void rds_tcp_recv_exit(void)
+{
+	kmem_cache_destroy(rds_tcp_incoming_slab);
+}

linux-7.2.3/net/rds/tcp_send.c

@@ -0,0 +1,226 @@
+/*
+ * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <net/tcp.h>
+
+#include "rds_single_path.h"
+#include "rds.h"
+#include "tcp.h"
+
+void rds_tcp_xmit_path_prepare(struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+
+	tcp_sock_set_cork(tc->t_sock->sk, true);
+}
+
+void rds_tcp_xmit_path_complete(struct rds_conn_path *cp)
+{
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+
+	tcp_sock_set_cork(tc->t_sock->sk, false);
+}
+
+/* the core send_sem serializes this with other xmit and shutdown */
+static int rds_tcp_sendmsg(struct socket *sock, void *data, unsigned int len)
+{
+	struct kvec vec = {
+		.iov_base = data,
+		.iov_len = len,
+	};
+	struct msghdr msg = {
+		.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL,
+	};
+
+	return kernel_sendmsg(sock, &msg, &vec, 1, vec.iov_len);
+}
+
+/* the core send_sem serializes this with other xmit and shutdown */
+int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
+		 unsigned int hdr_off, unsigned int sg, unsigned int off)
+{
+	struct rds_conn_path *cp = rm->m_inc.i_conn_path;
+	struct rds_tcp_connection *tc = cp->cp_transport_data;
+	struct msghdr msg = {};
+	struct bio_vec bvec;
+	int done = 0;
+	int ret = 0;
+
+	if (hdr_off == 0) {
+		/*
+		 * m_ack_seq is set to the sequence number of the last byte of
+		 * header and data.  see rds_tcp_is_acked().
+		 */
+		tc->t_last_sent_nxt = rds_tcp_write_seq(tc);
+		rm->m_ack_seq = tc->t_last_sent_nxt +
+				sizeof(struct rds_header) +
+				be32_to_cpu(rm->m_inc.i_hdr.h_len) - 1;
+		smp_mb__before_atomic();
+		set_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags);
+		tc->t_last_expected_una = rm->m_ack_seq + 1;
+
+		if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
+			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
+
+		rdsdebug("rm %p tcp nxt %u ack_seq %llu\n",
+			 rm, rds_tcp_write_seq(tc),
+			 (unsigned long long)rm->m_ack_seq);
+	}
+
+	if (hdr_off < sizeof(struct rds_header)) {
+		/* see rds_tcp_write_space() */
+		set_bit(SOCK_NOSPACE, &tc->t_sock->sk->sk_socket->flags);
+
+		ret = rds_tcp_sendmsg(tc->t_sock,
+				      (void *)&rm->m_inc.i_hdr + hdr_off,
+				      sizeof(rm->m_inc.i_hdr) - hdr_off);
+		if (ret < 0)
+			goto out;
+		done += ret;
+		if (hdr_off + done != sizeof(struct rds_header))
+			goto out;
+	}
+
+	while (sg < rm->data.op_nents) {
+		msg.msg_flags = MSG_SPLICE_PAGES | MSG_DONTWAIT | MSG_NOSIGNAL;
+		if (sg + 1 < rm->data.op_nents)
+			msg.msg_flags |= MSG_MORE;
+
+		bvec_set_page(&bvec, sg_page(&rm->data.op_sg[sg]),
+			      rm->data.op_sg[sg].length - off,
+			      rm->data.op_sg[sg].offset + off);
+		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1,
+			      rm->data.op_sg[sg].length - off);
+		ret = sock_sendmsg(tc->t_sock, &msg);
+		rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->data.op_sg[sg]),
+			 rm->data.op_sg[sg].offset + off, rm->data.op_sg[sg].length - off,
+			 ret);
+		if (ret <= 0)
+			break;
+
+		off += ret;
+		done += ret;
+		if (off == rm->data.op_sg[sg].length) {
+			off = 0;
+			sg++;
+		}
+	}
+
+out:
+	if (ret <= 0) {
+		/* write_space will hit after EAGAIN, all else fatal */
+		if (ret == -EAGAIN) {
+			rds_tcp_stats_inc(s_tcp_sndbuf_full);
+			ret = 0;
+		} else {
+			/* No need to disconnect/reconnect if path_drop
+			 * has already been triggered, because, e.g., of
+			 * an incoming RST.
+			 */
+			if (rds_conn_path_up(cp)) {
+				pr_warn("RDS/tcp: send to %pI6c on cp [%d]"
+					"returned %d, "
+					"disconnecting and reconnecting\n",
+					&conn->c_faddr, cp->cp_index, ret);
+				rds_conn_path_drop(cp, false);
+			}
+		}
+	}
+	if (done == 0)
+		done = ret;
+	return done;
+}
+
+/*
+ * rm->m_ack_seq is set to the tcp sequence number that corresponds to the
+ * last byte of the message, including the header.  This means that the
+ * entire message has been received if rm->m_ack_seq is "before" the next
+ * unacked byte of the TCP sequence space.  We have to do very careful
+ * wrapping 32bit comparisons here.
+ */
+int rds_tcp_is_acked(struct rds_message *rm, uint64_t ack)
+{
+	if (!test_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags))
+		return 0;
+	return (__s32)((u32)rm->m_ack_seq - (u32)ack) < 0;
+}
+
+void rds_tcp_write_space(struct sock *sk)
+{
+	void (*write_space)(struct sock *sk);
+	struct rds_conn_path *cp;
+	struct rds_tcp_connection *tc;
+
+	read_lock_bh(&sk->sk_callback_lock);
+	cp = sk->sk_user_data;
+	if (!cp) {
+		write_space = sk->sk_write_space;
+		goto out;
+	}
+
+	tc = cp->cp_transport_data;
+	rdsdebug("write_space for tc %p\n", tc);
+	write_space = tc->t_orig_write_space;
+	rds_tcp_stats_inc(s_tcp_write_space_calls);
+
+	rdsdebug("tcp una %u\n", rds_tcp_snd_una(tc));
+	tc->t_last_seen_una = rds_tcp_snd_una(tc);
+	rds_send_path_drop_acked(cp, rds_tcp_snd_una(tc), rds_tcp_is_acked);
+
+	rcu_read_lock();
+	if ((refcount_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf &&
+	    !rds_destroy_pending(cp->cp_conn))
+		queue_delayed_work(cp->cp_wq, &cp->cp_send_w, 0);
+	rcu_read_unlock();
+
+out:
+	read_unlock_bh(&sk->sk_callback_lock);
+
+	/*
+	 * write_space is only called when data leaves tcp's send queue if
+	 * SOCK_NOSPACE is set.  We set SOCK_NOSPACE every time we put
+	 * data in tcp's send queue because we use write_space to parse the
+	 * sequence numbers and notice that rds messages have been fully
+	 * received.
+	 *
+	 * tcp's write_space clears SOCK_NOSPACE if the send queue has more
+	 * than a certain amount of space. So we need to set it again *after*
+	 * we call tcp's write_space or else we might only get called on the
+	 * first of a series of incoming tcp acks.
+	 */
+	write_space(sk);
+
+	if (sk->sk_socket)
+		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
+}

linux-7.2.3/net/rds/tcp_stats.c

@@ -0,0 +1,74 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/percpu.h>
+#include <linux/seq_file.h>
+#include <linux/proc_fs.h>
+
+#include "rds.h"
+#include "tcp.h"
+
+DEFINE_PER_CPU(struct rds_tcp_statistics, rds_tcp_stats)
+	____cacheline_aligned;
+
+static const char * const rds_tcp_stat_names[] = {
+	"tcp_data_ready_calls",
+	"tcp_write_space_calls",
+	"tcp_sndbuf_full",
+	"tcp_connect_raced",
+	"tcp_listen_closed_stale",
+};
+
+unsigned int rds_tcp_stats_info_copy(struct rds_info_iterator *iter,
+				     unsigned int avail)
+{
+	struct rds_tcp_statistics stats = {0, };
+	uint64_t *src;
+	uint64_t *sum;
+	size_t i;
+	int cpu;
+
+	if (avail < ARRAY_SIZE(rds_tcp_stat_names))
+		goto out;
+
+	for_each_online_cpu(cpu) {
+		src = (uint64_t *)&(per_cpu(rds_tcp_stats, cpu));
+		sum = (uint64_t *)&stats;
+		for (i = 0; i < sizeof(stats) / sizeof(uint64_t); i++)
+			*(sum++) += *(src++);
+	}
+
+	rds_stats_info_copy(iter, (uint64_t *)&stats, rds_tcp_stat_names,
+			    ARRAY_SIZE(rds_tcp_stat_names));
+out:
+	return ARRAY_SIZE(rds_tcp_stat_names);
+}

linux-7.2.3/net/rds/threads.c

@@ -0,0 +1,311 @@
+/*
+ * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/random.h>
+#include <linux/export.h>
+
+#include "rds.h"
+
+/*
+ * All of connection management is simplified by serializing it through
+ * work queues that execute in a connection managing thread.
+ *
+ * TCP wants to send acks through sendpage() in response to data_ready(),
+ * but it needs a process context to do so.
+ *
+ * The receive paths need to allocate but can't drop packets (!) so we have
+ * a thread around to block allocating if the receive fast path sees an
+ * allocation failure.
+ */
+
+/* Grand Unified Theory of connection life cycle:
+ * At any point in time, the connection can be in one of these states:
+ * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
+ *
+ * The following transitions are possible:
+ *  ANY		  -> ERROR
+ *  UP		  -> DISCONNECTING
+ *  ERROR	  -> DISCONNECTING
+ *  DISCONNECTING -> DOWN
+ *  DOWN	  -> CONNECTING
+ *  CONNECTING	  -> UP
+ *
+ * Transition to state DISCONNECTING/DOWN:
+ *  -	Inside the shutdown worker; synchronizes with xmit path
+ *	through RDS_IN_XMIT, and with connection management callbacks
+ *	via c_cm_lock.
+ *
+ *	For receive callbacks, we rely on the underlying transport
+ *	(TCP, IB/RDMA) to provide the necessary synchronisation.
+ */
+struct workqueue_struct *rds_wq;
+EXPORT_SYMBOL_GPL(rds_wq);
+
+void rds_connect_path_complete(struct rds_conn_path *cp, int curr)
+{
+	if (!rds_conn_path_transition(cp, curr, RDS_CONN_UP)) {
+		printk(KERN_WARNING "%s: Cannot transition to state UP, "
+				"current state is %d\n",
+				__func__,
+				atomic_read(&cp->cp_state));
+		rds_conn_path_drop(cp, false);
+		return;
+	}
+
+	rdsdebug("conn %p for %pI6c to %pI6c complete\n",
+		 cp->cp_conn, &cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr);
+
+	cp->cp_reconnect_jiffies = 0;
+	set_bit(0, &cp->cp_conn->c_map_queued);
+	rcu_read_lock();
+	if (!rds_destroy_pending(cp->cp_conn)) {
+		queue_delayed_work(cp->cp_wq, &cp->cp_send_w, 0);
+		queue_delayed_work(cp->cp_wq, &cp->cp_recv_w, 0);
+	}
+	rcu_read_unlock();
+	cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
+}
+EXPORT_SYMBOL_GPL(rds_connect_path_complete);
+
+void rds_connect_complete(struct rds_connection *conn)
+{
+	rds_connect_path_complete(&conn->c_path[0], RDS_CONN_CONNECTING);
+}
+EXPORT_SYMBOL_GPL(rds_connect_complete);
+
+/*
+ * This random exponential backoff is relied on to eventually resolve racing
+ * connects.
+ *
+ * If connect attempts race then both parties drop both connections and come
+ * here to wait for a random amount of time before trying again.  Eventually
+ * the backoff range will be so much greater than the time it takes to
+ * establish a connection that one of the pair will establish the connection
+ * before the other's random delay fires.
+ *
+ * Connection attempts that arrive while a connection is already established
+ * are also considered to be racing connects.  This lets a connection from
+ * a rebooted machine replace an existing stale connection before the transport
+ * notices that the connection has failed.
+ *
+ * We should *always* start with a random backoff; otherwise a broken connection
+ * will always take several iterations to be re-established.
+ */
+void rds_queue_reconnect(struct rds_conn_path *cp)
+{
+	unsigned long rand;
+	struct rds_connection *conn = cp->cp_conn;
+
+	rdsdebug("conn %p for %pI6c to %pI6c reconnect jiffies %lu\n",
+		 conn, &conn->c_laddr, &conn->c_faddr,
+		 cp->cp_reconnect_jiffies);
+
+	/* let peer with smaller addr initiate reconnect, to avoid duels */
+	if (conn->c_trans->t_type == RDS_TRANS_TCP &&
+	    rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) >= 0)
+		return;
+
+	set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
+	if (cp->cp_reconnect_jiffies == 0) {
+		cp->cp_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
+		rcu_read_lock();
+		if (!rds_destroy_pending(cp->cp_conn))
+			queue_delayed_work(cp->cp_wq, &cp->cp_conn_w, 0);
+		rcu_read_unlock();
+		return;
+	}
+
+	get_random_bytes(&rand, sizeof(rand));
+	rdsdebug("%lu delay %lu ceil conn %p for %pI6c -> %pI6c\n",
+		 rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies,
+		 conn, &conn->c_laddr, &conn->c_faddr);
+	rcu_read_lock();
+	if (!rds_destroy_pending(cp->cp_conn))
+		queue_delayed_work(cp->cp_wq, &cp->cp_conn_w,
+				   rand % cp->cp_reconnect_jiffies);
+	rcu_read_unlock();
+
+	cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2,
+					rds_sysctl_reconnect_max_jiffies);
+}
+
+void rds_connect_worker(struct work_struct *work)
+{
+	struct rds_conn_path *cp = container_of(work,
+						struct rds_conn_path,
+						cp_conn_w.work);
+	struct rds_connection *conn = cp->cp_conn;
+	int ret;
+
+	if (cp->cp_index > 0 &&
+	    rds_addr_cmp(&cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr) >= 0)
+		return;
+	clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
+	ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
+	if (ret) {
+		ret = conn->c_trans->conn_path_connect(cp);
+		rdsdebug("conn %p for %pI6c to %pI6c dispatched, ret %d\n",
+			 conn, &conn->c_laddr, &conn->c_faddr, ret);
+
+		if (ret) {
+			if (rds_conn_path_transition(cp,
+						     RDS_CONN_CONNECTING,
+						     RDS_CONN_DOWN))
+				rds_queue_reconnect(cp);
+			else
+				rds_conn_path_error(cp, "connect failed\n");
+		}
+	}
+}
+
+void rds_send_worker(struct work_struct *work)
+{
+	struct rds_conn_path *cp = container_of(work,
+						struct rds_conn_path,
+						cp_send_w.work);
+	int ret;
+
+	if (rds_conn_path_state(cp) == RDS_CONN_UP) {
+		clear_bit(RDS_LL_SEND_FULL, &cp->cp_flags);
+		ret = rds_send_xmit(cp);
+		cond_resched();
+		rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
+		switch (ret) {
+		case -EAGAIN:
+			rds_stats_inc(s_send_immediate_retry);
+			queue_delayed_work(cp->cp_wq, &cp->cp_send_w, 0);
+			break;
+		case -ENOMEM:
+			rds_stats_inc(s_send_delayed_retry);
+			queue_delayed_work(cp->cp_wq, &cp->cp_send_w, 2);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
+void rds_recv_worker(struct work_struct *work)
+{
+	struct rds_conn_path *cp = container_of(work,
+						struct rds_conn_path,
+						cp_recv_w.work);
+	int ret;
+
+	if (rds_conn_path_state(cp) == RDS_CONN_UP) {
+		ret = cp->cp_conn->c_trans->recv_path(cp);
+		rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
+		switch (ret) {
+		case -EAGAIN:
+			rds_stats_inc(s_recv_immediate_retry);
+			queue_delayed_work(cp->cp_wq, &cp->cp_recv_w, 0);
+			break;
+		case -ENOMEM:
+			rds_stats_inc(s_recv_delayed_retry);
+			queue_delayed_work(cp->cp_wq, &cp->cp_recv_w, 2);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
+void rds_shutdown_worker(struct work_struct *work)
+{
+	struct rds_conn_path *cp = container_of(work,
+						struct rds_conn_path,
+						cp_down_w);
+
+	rds_conn_shutdown(cp);
+}
+
+void rds_threads_exit(void)
+{
+	destroy_workqueue(rds_wq);
+}
+
+int rds_threads_init(void)
+{
+	rds_wq = create_singlethread_workqueue("krdsd");
+	if (!rds_wq)
+		return -ENOMEM;
+
+	return 0;
+}
+
+/* Compare two IPv6 addresses.  Return 0 if the two addresses are equal.
+ * Return 1 if the first is greater.  Return -1 if the second is greater.
+ */
+int rds_addr_cmp(const struct in6_addr *addr1,
+		 const struct in6_addr *addr2)
+{
+#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && BITS_PER_LONG == 64
+	const __be64 *a1, *a2;
+	u64 x, y;
+
+	a1 = (__be64 *)addr1;
+	a2 = (__be64 *)addr2;
+
+	if (*a1 != *a2) {
+		if (be64_to_cpu(*a1) < be64_to_cpu(*a2))
+			return -1;
+		else
+			return 1;
+	} else {
+		x = be64_to_cpu(*++a1);
+		y = be64_to_cpu(*++a2);
+		if (x < y)
+			return -1;
+		else if (x > y)
+			return 1;
+		else
+			return 0;
+	}
+#else
+	u32 a, b;
+	int i;
+
+	for (i = 0; i < 4; i++) {
+		if (addr1->s6_addr32[i] != addr2->s6_addr32[i]) {
+			a = ntohl(addr1->s6_addr32[i]);
+			b = ntohl(addr2->s6_addr32[i]);
+			if (a < b)
+				return -1;
+			else if (a > b)
+				return 1;
+		}
+	}
+	return 0;
+#endif
+}
+EXPORT_SYMBOL_GPL(rds_addr_cmp);

linux-7.2.3/net/rds/transport.c

@@ -0,0 +1,169 @@
+/*
+ * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/in.h>
+#include <linux/ipv6.h>
+
+#include "rds.h"
+#include "loop.h"
+
+static char * const rds_trans_modules[] = {
+	[RDS_TRANS_IB] = "rds_rdma",
+	[RDS_TRANS_GAP] = NULL,
+	[RDS_TRANS_TCP] = "rds_tcp",
+};
+
+static struct rds_transport *transports[RDS_TRANS_COUNT];
+static DECLARE_RWSEM(rds_trans_sem);
+
+void rds_trans_register(struct rds_transport *trans)
+{
+	BUG_ON(strlen(trans->t_name) + 1 > TRANSNAMSIZ);
+
+	down_write(&rds_trans_sem);
+
+	if (transports[trans->t_type])
+		printk(KERN_ERR "RDS Transport type %d already registered\n",
+			trans->t_type);
+	else {
+		transports[trans->t_type] = trans;
+		printk(KERN_INFO "Registered RDS/%s transport\n", trans->t_name);
+	}
+
+	up_write(&rds_trans_sem);
+}
+EXPORT_SYMBOL_GPL(rds_trans_register);
+
+void rds_trans_unregister(struct rds_transport *trans)
+{
+	down_write(&rds_trans_sem);
+
+	transports[trans->t_type] = NULL;
+	printk(KERN_INFO "Unregistered RDS/%s transport\n", trans->t_name);
+
+	up_write(&rds_trans_sem);
+}
+EXPORT_SYMBOL_GPL(rds_trans_unregister);
+
+void rds_trans_put(struct rds_transport *trans)
+{
+	if (trans)
+		module_put(trans->t_owner);
+}
+
+struct rds_transport *rds_trans_get_preferred(struct net *net,
+					      const struct in6_addr *addr,
+					      __u32 scope_id)
+{
+	struct rds_transport *ret = NULL;
+	struct rds_transport *trans;
+	unsigned int i;
+
+	if (ipv6_addr_v4mapped(addr)) {
+		if (*(u_int8_t *)&addr->s6_addr32[3] == IN_LOOPBACKNET)
+			return &rds_loop_transport;
+	} else if (ipv6_addr_loopback(addr)) {
+		return &rds_loop_transport;
+	}
+
+	down_read(&rds_trans_sem);
+	for (i = 0; i < RDS_TRANS_COUNT; i++) {
+		trans = transports[i];
+
+		if (trans && (trans->laddr_check(net, addr, scope_id) == 0) &&
+		    (!trans->t_owner || try_module_get(trans->t_owner))) {
+			ret = trans;
+			break;
+		}
+	}
+	up_read(&rds_trans_sem);
+
+	return ret;
+}
+
+struct rds_transport *rds_trans_get(int t_type)
+{
+	struct rds_transport *ret = NULL;
+	struct rds_transport *trans;
+
+	down_read(&rds_trans_sem);
+	trans = transports[t_type];
+	if (!trans) {
+		up_read(&rds_trans_sem);
+		if (rds_trans_modules[t_type])
+			request_module(rds_trans_modules[t_type]);
+		down_read(&rds_trans_sem);
+		trans = transports[t_type];
+	}
+	if (trans && trans->t_type == t_type &&
+	    (!trans->t_owner || try_module_get(trans->t_owner)))
+		ret = trans;
+
+	up_read(&rds_trans_sem);
+
+	return ret;
+}
+
+/*
+ * This returns the number of stats entries in the snapshot and only
+ * copies them using the iter if there is enough space for them.  The
+ * caller passes in the global stats so that we can size and copy while
+ * holding the lock.
+ */
+unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
+				       unsigned int avail)
+
+{
+	struct rds_transport *trans;
+	unsigned int total = 0;
+	unsigned int part;
+	int i;
+
+	rds_info_iter_unmap(iter);
+	down_read(&rds_trans_sem);
+
+	for (i = 0; i < RDS_TRANS_COUNT; i++) {
+		trans = transports[i];
+		if (!trans || !trans->stats_info_copy)
+			continue;
+
+		part = trans->stats_info_copy(iter, avail);
+		avail -= min(avail, part);
+		total += part;
+	}
+
+	up_read(&rds_trans_sem);
+
+	return total;
+}