]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/commitdiff
rds: deliver zerocopy completion notification with data
authorSowmini Varadhan <sowmini.varadhan@oracle.com>
Tue, 27 Feb 2018 17:52:43 +0000 (09:52 -0800)
committerDavid S. Miller <davem@davemloft.net>
Tue, 27 Feb 2018 19:19:11 +0000 (14:19 -0500)
This commit is an optimization over commit 01883eda72bd
("rds: support for zcopy completion notification") for PF_RDS sockets.

RDS applications are predominantly request-response transactions, so
it is more efficient to reduce the number of system calls and have
zerocopy completion notification delivered as ancillary data on the
POLLIN channel.

Cookies are passed up as ancillary data (at level SOL_RDS) in a
struct rds_zcopy_cookies when the returned value of recvmsg() is
greater than, or equal to, 0. A max of RDS_MAX_ZCOOKIES may be passed
with each message.

This commit removes support for zerocopy completion notification on
MSG_ERRQUEUE for PF_RDS sockets.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/uapi/linux/errqueue.h
include/uapi/linux/rds.h
net/rds/af_rds.c
net/rds/message.c
net/rds/rds.h
net/rds/recv.c

index 28812eda420908ebf73008525d90c1abb3907536..dc64cfaf13da08564a8271e50a4edb89d221b148 100644 (file)
@@ -20,13 +20,11 @@ struct sock_extended_err {
 #define SO_EE_ORIGIN_ICMP6     3
 #define SO_EE_ORIGIN_TXSTATUS  4
 #define SO_EE_ORIGIN_ZEROCOPY  5
-#define SO_EE_ORIGIN_ZCOOKIE   6
 #define SO_EE_ORIGIN_TIMESTAMPING SO_EE_ORIGIN_TXSTATUS
 
 #define SO_EE_OFFENDER(ee)     ((struct sockaddr*)((ee)+1))
 
 #define SO_EE_CODE_ZEROCOPY_COPIED     1
-#define        SO_EE_ORIGIN_MAX_ZCOOKIES       8
 
 /**
  *     struct scm_timestamping - timestamps exposed through cmsg
index 12e3bca32cadb4cdf1e67bf7c53baff3a4165541..a66b213de3d7a40ee13cb40d900bfc1a18818692 100644 (file)
 #define RDS_CMSG_MASKED_ATOMIC_CSWP    9
 #define RDS_CMSG_RXPATH_LATENCY                11
 #define        RDS_CMSG_ZCOPY_COOKIE           12
+#define        RDS_CMSG_ZCOPY_COMPLETION       13
 
 #define RDS_INFO_FIRST                 10000
 #define RDS_INFO_COUNTERS              10000
@@ -317,6 +318,12 @@ struct rds_rdma_notify {
 #define RDS_RDMA_DROPPED       3
 #define RDS_RDMA_OTHER_ERROR   4
 
+#define        RDS_MAX_ZCOOKIES        8
+struct rds_zcopy_cookies {
+       __u32 num;
+       __u32 cookies[RDS_MAX_ZCOOKIES];
+};
+
 /*
  * Common set of flags for all RDMA related structs
  */
index a937f18896aee6c0fce229b528d77b9ed33ae05c..f7126108a811561da9aca7db4d27739be1c32efb 100644 (file)
@@ -77,6 +77,7 @@ static int rds_release(struct socket *sock)
        rds_send_drop_to(rs, NULL);
        rds_rdma_drop_keys(rs);
        rds_notify_queue_get(rs, NULL);
+       __skb_queue_purge(&rs->rs_zcookie_queue);
 
        spin_lock_bh(&rds_sock_lock);
        list_del_init(&rs->rs_item);
@@ -144,7 +145,7 @@ static int rds_getname(struct socket *sock, struct sockaddr *uaddr,
  *  -  to signal that a previously congested destination may have become
  *     uncongested
  *  -  A notification has been queued to the socket (this can be a congestion
- *     update, or a RDMA completion).
+ *     update, or a RDMA completion, or a MSG_ZEROCOPY completion).
  *
  * EPOLLOUT is asserted if there is room on the send queue. This does not mean
  * however, that the next sendmsg() call will succeed. If the application tries
@@ -178,7 +179,8 @@ static __poll_t rds_poll(struct file *file, struct socket *sock,
                spin_unlock(&rs->rs_lock);
        }
        if (!list_empty(&rs->rs_recv_queue) ||
-           !list_empty(&rs->rs_notify_queue))
+           !list_empty(&rs->rs_notify_queue) ||
+           !skb_queue_empty(&rs->rs_zcookie_queue))
                mask |= (EPOLLIN | EPOLLRDNORM);
        if (rs->rs_snd_bytes < rds_sk_sndbuf(rs))
                mask |= (EPOLLOUT | EPOLLWRNORM);
@@ -513,6 +515,7 @@ static int __rds_create(struct socket *sock, struct sock *sk, int protocol)
        INIT_LIST_HEAD(&rs->rs_recv_queue);
        INIT_LIST_HEAD(&rs->rs_notify_queue);
        INIT_LIST_HEAD(&rs->rs_cong_list);
+       skb_queue_head_init(&rs->rs_zcookie_queue);
        spin_lock_init(&rs->rs_rdma_lock);
        rs->rs_rdma_keys = RB_ROOT;
        rs->rs_rx_traces = 0;
index 651834513481a3db69f62e884f1b2e9f84d833f8..116cf87ccb898081e031e07347a32bd907b843c5 100644 (file)
@@ -58,32 +58,26 @@ EXPORT_SYMBOL_GPL(rds_message_addref);
 
 static inline bool skb_zcookie_add(struct sk_buff *skb, u32 cookie)
 {
-       struct sock_exterr_skb *serr = SKB_EXT_ERR(skb);
-       int ncookies;
-       u32 *ptr;
+       struct rds_zcopy_cookies *ck = (struct rds_zcopy_cookies *)skb->cb;
+       int ncookies = ck->num;
 
-       if (serr->ee.ee_origin != SO_EE_ORIGIN_ZCOOKIE)
+       if (ncookies == RDS_MAX_ZCOOKIES)
                return false;
-       ncookies = serr->ee.ee_data;
-       if (ncookies == SO_EE_ORIGIN_MAX_ZCOOKIES)
-               return false;
-       ptr = skb_put(skb, sizeof(u32));
-       *ptr = cookie;
-       serr->ee.ee_data = ++ncookies;
+       ck->cookies[ncookies] = cookie;
+       ck->num =  ++ncookies;
        return true;
 }
 
 static void rds_rm_zerocopy_callback(struct rds_sock *rs,
                                     struct rds_znotifier *znotif)
 {
-       struct sock *sk = rds_rs_to_sk(rs);
        struct sk_buff *skb, *tail;
-       struct sock_exterr_skb *serr;
        unsigned long flags;
        struct sk_buff_head *q;
        u32 cookie = znotif->z_cookie;
+       struct rds_zcopy_cookies *ck;
 
-       q = &sk->sk_error_queue;
+       q = &rs->rs_zcookie_queue;
        spin_lock_irqsave(&q->lock, flags);
        tail = skb_peek_tail(q);
 
@@ -91,22 +85,19 @@ static void rds_rm_zerocopy_callback(struct rds_sock *rs,
                spin_unlock_irqrestore(&q->lock, flags);
                mm_unaccount_pinned_pages(&znotif->z_mmp);
                consume_skb(rds_skb_from_znotifier(znotif));
-               sk->sk_error_report(sk);
+               /* caller invokes rds_wake_sk_sleep() */
                return;
        }
 
        skb = rds_skb_from_znotifier(znotif);
-       serr = SKB_EXT_ERR(skb);
-       memset(&serr->ee, 0, sizeof(serr->ee));
-       serr->ee.ee_errno = 0;
-       serr->ee.ee_origin = SO_EE_ORIGIN_ZCOOKIE;
-       serr->ee.ee_info = 0;
+       ck = (struct rds_zcopy_cookies *)skb->cb;
+       memset(ck, 0, sizeof(*ck));
        WARN_ON(!skb_zcookie_add(skb, cookie));
 
        __skb_queue_tail(q, skb);
 
        spin_unlock_irqrestore(&q->lock, flags);
-       sk->sk_error_report(sk);
+       /* caller invokes rds_wake_sk_sleep() */
 
        mm_unaccount_pinned_pages(&znotif->z_mmp);
 }
@@ -129,6 +120,7 @@ static void rds_message_purge(struct rds_message *rm)
                if (rm->data.op_mmp_znotifier) {
                        zcopy = true;
                        rds_rm_zerocopy_callback(rs, rm->data.op_mmp_znotifier);
+                       rds_wake_sk_sleep(rs);
                        rm->data.op_mmp_znotifier = NULL;
                }
                sock_put(rds_rs_to_sk(rs));
@@ -362,10 +354,12 @@ int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
                int total_copied = 0;
                struct sk_buff *skb;
 
-               skb = alloc_skb(SO_EE_ORIGIN_MAX_ZCOOKIES * sizeof(u32),
-                               GFP_KERNEL);
+               skb = alloc_skb(0, GFP_KERNEL);
                if (!skb)
                        return -ENOMEM;
+               BUILD_BUG_ON(sizeof(skb->cb) <
+                            max_t(int, sizeof(struct rds_znotifier),
+                                  sizeof(struct rds_zcopy_cookies)));
                rm->data.op_mmp_znotifier = RDS_ZCOPY_SKB(skb);
                if (mm_account_pinned_pages(&rm->data.op_mmp_znotifier->z_mmp,
                                            length)) {
index 31cd388520502c5bbbc8b331953a38528a059b6a..33b16353d8f396da9ef2648907886c8e95800d05 100644 (file)
@@ -603,6 +603,8 @@ struct rds_sock {
        /* Socket receive path trace points*/
        u8                      rs_rx_traces;
        u8                      rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
+
+       struct sk_buff_head     rs_zcookie_queue;
 };
 
 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
index b080961464df71ac048e971a544b32f3be5e7b48..d50747725221f87ae18438eb6e2461bacbcca89d 100644 (file)
@@ -577,6 +577,32 @@ out:
        return ret;
 }
 
+static bool rds_recvmsg_zcookie(struct rds_sock *rs, struct msghdr *msg)
+{
+       struct sk_buff *skb;
+       struct sk_buff_head *q = &rs->rs_zcookie_queue;
+       struct rds_zcopy_cookies *done;
+
+       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;
+
+       skb = skb_dequeue(q);
+       if (!skb)
+               return false;
+       done = (struct rds_zcopy_cookies *)skb->cb;
+       if (put_cmsg(msg, SOL_RDS, RDS_CMSG_ZCOPY_COMPLETION, sizeof(*done),
+                    done)) {
+               skb_queue_head(q, skb);
+               return false;
+       }
+       consume_skb(skb);
+       return true;
+}
+
 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
                int msg_flags)
 {
@@ -611,7 +637,9 @@ int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
 
                if (!rds_next_incoming(rs, &inc)) {
                        if (nonblock) {
-                               ret = -EAGAIN;
+                               bool reaped = rds_recvmsg_zcookie(rs, msg);
+
+                               ret = reaped ?  0 : -EAGAIN;
                                break;
                        }
 
@@ -660,6 +688,7 @@ int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
                        ret = -EFAULT;
                        goto out;
                }
+               rds_recvmsg_zcookie(rs, msg);
 
                rds_stats_inc(s_recv_delivered);