RDS: recv.c
Andy Grover [Tue, 24 Feb 2009 15:30:28 +0000 (15:30 +0000)]
Upon receiving a datagram from the transport, RDS parses the
headers and potentially queues an ACK.

Signed-off-by: Andy Grover <andy.grover@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/rds/recv.c [new file with mode: 0644]

diff --git a/net/rds/recv.c b/net/rds/recv.c
new file mode 100644 (file)
index 0000000..f2118c5
--- /dev/null
@@ -0,0 +1,542 @@
+/*
+ * 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 <net/sock.h>
+#include <linux/in.h>
+
+#include "rds.h"
+#include "rdma.h"
+
+void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
+                 __be32 saddr)
+{
+       atomic_set(&inc->i_refcount, 1);
+       INIT_LIST_HEAD(&inc->i_item);
+       inc->i_conn = conn;
+       inc->i_saddr = saddr;
+       inc->i_rdma_cookie = 0;
+}
+
+void rds_inc_addref(struct rds_incoming *inc)
+{
+       rdsdebug("addref inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
+       atomic_inc(&inc->i_refcount);
+}
+
+void rds_inc_put(struct rds_incoming *inc)
+{
+       rdsdebug("put inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
+       if (atomic_dec_and_test(&inc->i_refcount)) {
+               BUG_ON(!list_empty(&inc->i_item));
+
+               inc->i_conn->c_trans->inc_free(inc);
+       }
+}
+
+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;
+       now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
+
+       rdsdebug("rs %p (%pI4:%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 */
+}
+
+/*
+ * 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_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;
+               }
+       }
+}
+
+/*
+ * 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, __be32 saddr, __be32 daddr,
+                      struct rds_incoming *inc, gfp_t gfp, enum km_type km)
+{
+       struct rds_sock *rs = NULL;
+       struct sock *sk;
+       unsigned long flags;
+
+       inc->i_conn = conn;
+       inc->i_rx_jiffies = jiffies;
+
+       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)conn->c_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) < conn->c_next_rx_seq
+        && (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
+               rds_stats_inc(s_recv_drop_old_seq);
+               goto out;
+       }
+       conn->c_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
+
+       if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
+               rds_stats_inc(s_recv_ping);
+               rds_send_pong(conn, inc->i_hdr.h_sport);
+               goto out;
+       }
+
+       rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
+       if (rs == NULL) {
+               rds_stats_inc(s_recv_drop_no_sock);
+               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);
+               rds_inc_addref(inc);
+               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);
+}
+
+/*
+ * 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 == NULL) {
+               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;
+
+       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);
+                       rds_inc_put(inc);
+               }
+       }
+       write_unlock_irqrestore(&rs->rs_recv_lock, flags);
+
+       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;
+
+
+       /* 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)
+{
+       int ret = 0;
+
+       if (inc->i_rdma_cookie) {
+               ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
+                               sizeof(inc->i_rdma_cookie), &inc->i_rdma_cookie);
+               if (ret)
+                       return ret;
+       }
+
+       return 0;
+}
+
+int rds_recvmsg(struct kiocb *iocb, 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;
+       struct sockaddr_in *sin;
+       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 there are pending notifications, do those - and nothing else */
+       if (!list_empty(&rs->rs_notify_queue)) {
+               ret = rds_notify_queue_get(rs, msg);
+               goto out;
+       }
+
+       if (rs->rs_cong_notify) {
+               ret = rds_notify_cong(rs, msg);
+               goto out;
+       }
+
+       while (1) {
+               if (!rds_next_incoming(rs, &inc)) {
+                       if (nonblock) {
+                               ret = -EAGAIN;
+                               break;
+                       }
+
+                       timeo = wait_event_interruptible_timeout(*sk->sk_sleep,
+                                               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 %pI4:%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_iov,
+                                                            size);
+               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);
+                       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)) {
+                       ret = -EFAULT;
+                       goto out;
+               }
+
+               rds_stats_inc(s_recv_delivered);
+
+               sin = (struct sockaddr_in *)msg->msg_name;
+               if (sin) {
+                       sin->sin_family = AF_INET;
+                       sin->sin_port = inc->i_hdr.h_sport;
+                       sin->sin_addr.s_addr = inc->i_saddr;
+                       memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
+               }
+               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;
+
+       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_del_init(&inc->i_item);
+               rds_inc_put(inc);
+       }
+       write_unlock_irqrestore(&rs->rs_recv_lock, flags);
+}
+
+/*
+ * 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);
+
+       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;
+       }
+
+       rds_info_copy(iter, &minfo, sizeof(minfo));
+}