tcp: annotate sk->sk_wmem_queued lockless reads
authorEric Dumazet <edumazet@google.com>
Fri, 11 Oct 2019 03:17:46 +0000 (20:17 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sun, 13 Oct 2019 17:13:08 +0000 (10:13 -0700)
For the sake of tcp_poll(), there are few places where we fetch
sk->sk_wmem_queued while this field can change from IRQ or other cpu.

We need to add READ_ONCE() annotations, and also make sure write
sides use corresponding WRITE_ONCE() to avoid store-tearing.

sk_wmem_queued_add() helper is added so that we can in
the future convert to ADD_ONCE() or equivalent if/when
available.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/sock.h
include/trace/events/sock.h
net/core/datagram.c
net/core/sock.c
net/ipv4/inet_diag.c
net/ipv4/tcp.c
net/ipv4/tcp_output.c
net/sched/em_meta.c

index 3d1e7502333e7ea0ff866f8982048800193caf33..f69b58bff7e5c69537444d99d68db5afc3abfa27 100644 (file)
@@ -878,12 +878,17 @@ static inline bool sk_acceptq_is_full(const struct sock *sk)
  */
 static inline int sk_stream_min_wspace(const struct sock *sk)
 {
-       return sk->sk_wmem_queued >> 1;
+       return READ_ONCE(sk->sk_wmem_queued) >> 1;
 }
 
 static inline int sk_stream_wspace(const struct sock *sk)
 {
-       return READ_ONCE(sk->sk_sndbuf) - sk->sk_wmem_queued;
+       return READ_ONCE(sk->sk_sndbuf) - READ_ONCE(sk->sk_wmem_queued);
+}
+
+static inline void sk_wmem_queued_add(struct sock *sk, int val)
+{
+       WRITE_ONCE(sk->sk_wmem_queued, sk->sk_wmem_queued + val);
 }
 
 void sk_stream_write_space(struct sock *sk);
@@ -1207,7 +1212,7 @@ static inline void sk_refcnt_debug_release(const struct sock *sk)
 
 static inline bool __sk_stream_memory_free(const struct sock *sk, int wake)
 {
-       if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
+       if (READ_ONCE(sk->sk_wmem_queued) >= READ_ONCE(sk->sk_sndbuf))
                return false;
 
        return sk->sk_prot->stream_memory_free ?
@@ -1467,7 +1472,7 @@ DECLARE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
 static inline void sk_wmem_free_skb(struct sock *sk, struct sk_buff *skb)
 {
        sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
-       sk->sk_wmem_queued -= skb->truesize;
+       sk_wmem_queued_add(sk, -skb->truesize);
        sk_mem_uncharge(sk, skb->truesize);
        if (static_branch_unlikely(&tcp_tx_skb_cache_key) &&
            !sk->sk_tx_skb_cache && !skb_cloned(skb)) {
@@ -2014,7 +2019,7 @@ static inline int skb_copy_to_page_nocache(struct sock *sk, struct iov_iter *fro
        skb->len             += copy;
        skb->data_len        += copy;
        skb->truesize        += copy;
-       sk->sk_wmem_queued   += copy;
+       sk_wmem_queued_add(sk, copy);
        sk_mem_charge(sk, copy);
        return 0;
 }
index f720c32e7dfd6f41c04194318d6c3f2e68b821cb..51fe9f6719eb13d872054676078df87e293fcd01 100644 (file)
@@ -115,7 +115,7 @@ TRACE_EVENT(sock_exceed_buf_limit,
                __entry->rmem_alloc = atomic_read(&sk->sk_rmem_alloc);
                __entry->sysctl_wmem = sk_get_wmem0(sk, prot);
                __entry->wmem_alloc = refcount_read(&sk->sk_wmem_alloc);
-               __entry->wmem_queued = sk->sk_wmem_queued;
+               __entry->wmem_queued = READ_ONCE(sk->sk_wmem_queued);
                __entry->kind = kind;
        ),
 
index 4cc8dc5db2b73471ae3a15fda753912d5e869624..c210fc116103d9915a2a4abc5225e0eb75825b0b 100644 (file)
@@ -640,7 +640,7 @@ int __zerocopy_sg_from_iter(struct sock *sk, struct sk_buff *skb,
                skb->len += copied;
                skb->truesize += truesize;
                if (sk && sk->sk_type == SOCK_STREAM) {
-                       sk->sk_wmem_queued += truesize;
+                       sk_wmem_queued_add(sk, truesize);
                        sk_mem_charge(sk, truesize);
                } else {
                        refcount_add(truesize, &skb->sk->sk_wmem_alloc);
index cd075bc86407a5816bd448521955525ebe941694..a515392ba84b67b2bf5400e0cfb7c3454fa87af8 100644 (file)
@@ -3212,7 +3212,7 @@ void sk_get_meminfo(const struct sock *sk, u32 *mem)
        mem[SK_MEMINFO_WMEM_ALLOC] = sk_wmem_alloc_get(sk);
        mem[SK_MEMINFO_SNDBUF] = READ_ONCE(sk->sk_sndbuf);
        mem[SK_MEMINFO_FWD_ALLOC] = sk->sk_forward_alloc;
-       mem[SK_MEMINFO_WMEM_QUEUED] = sk->sk_wmem_queued;
+       mem[SK_MEMINFO_WMEM_QUEUED] = READ_ONCE(sk->sk_wmem_queued);
        mem[SK_MEMINFO_OPTMEM] = atomic_read(&sk->sk_omem_alloc);
        mem[SK_MEMINFO_BACKLOG] = READ_ONCE(sk->sk_backlog.len);
        mem[SK_MEMINFO_DROPS] = atomic_read(&sk->sk_drops);
index bbb005eb5218c2765567b1d14ef564d2332479cc..7dc79b973e6edcc64e668e14c71c732ca1187e8f 100644 (file)
@@ -193,7 +193,7 @@ int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
        if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
                struct inet_diag_meminfo minfo = {
                        .idiag_rmem = sk_rmem_alloc_get(sk),
-                       .idiag_wmem = sk->sk_wmem_queued,
+                       .idiag_wmem = READ_ONCE(sk->sk_wmem_queued),
                        .idiag_fmem = sk->sk_forward_alloc,
                        .idiag_tmem = sk_wmem_alloc_get(sk),
                };
index 11185326297211bee3746edd725d106643b78720..b2ac4f074e2da21db57923fda722b6d23f170de9 100644 (file)
@@ -659,7 +659,7 @@ static void skb_entail(struct sock *sk, struct sk_buff *skb)
        tcb->sacked  = 0;
        __skb_header_release(skb);
        tcp_add_write_queue_tail(sk, skb);
-       sk->sk_wmem_queued += skb->truesize;
+       sk_wmem_queued_add(sk, skb->truesize);
        sk_mem_charge(sk, skb->truesize);
        if (tp->nonagle & TCP_NAGLE_PUSH)
                tp->nonagle &= ~TCP_NAGLE_PUSH;
@@ -1034,7 +1034,7 @@ new_segment:
                skb->len += copy;
                skb->data_len += copy;
                skb->truesize += copy;
-               sk->sk_wmem_queued += copy;
+               sk_wmem_queued_add(sk, copy);
                sk_mem_charge(sk, copy);
                skb->ip_summed = CHECKSUM_PARTIAL;
                WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
index a115a991dfb5b36c5b3dafd8c9ad94d07685f3a0..0488607c5cd3615633af207f0bb41bea0c0176ce 100644 (file)
@@ -1199,7 +1199,7 @@ static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
        WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
        __skb_header_release(skb);
        tcp_add_write_queue_tail(sk, skb);
-       sk->sk_wmem_queued += skb->truesize;
+       sk_wmem_queued_add(sk, skb->truesize);
        sk_mem_charge(sk, skb->truesize);
 }
 
@@ -1333,7 +1333,7 @@ int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
                return -ENOMEM; /* We'll just try again later. */
        skb_copy_decrypted(buff, skb);
 
-       sk->sk_wmem_queued += buff->truesize;
+       sk_wmem_queued_add(sk, buff->truesize);
        sk_mem_charge(sk, buff->truesize);
        nlen = skb->len - len - nsize;
        buff->truesize += nlen;
@@ -1443,7 +1443,7 @@ int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
 
        if (delta_truesize) {
                skb->truesize      -= delta_truesize;
-               sk->sk_wmem_queued -= delta_truesize;
+               sk_wmem_queued_add(sk, -delta_truesize);
                sk_mem_uncharge(sk, delta_truesize);
                sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
        }
@@ -1888,7 +1888,7 @@ static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
                return -ENOMEM;
        skb_copy_decrypted(buff, skb);
 
-       sk->sk_wmem_queued += buff->truesize;
+       sk_wmem_queued_add(sk, buff->truesize);
        sk_mem_charge(sk, buff->truesize);
        buff->truesize += nlen;
        skb->truesize -= nlen;
@@ -2152,7 +2152,7 @@ static int tcp_mtu_probe(struct sock *sk)
        nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
        if (!nskb)
                return -1;
-       sk->sk_wmem_queued += nskb->truesize;
+       sk_wmem_queued_add(sk, nskb->truesize);
        sk_mem_charge(sk, nskb->truesize);
 
        skb = tcp_send_head(sk);
@@ -3222,7 +3222,7 @@ int tcp_send_synack(struct sock *sk)
                        tcp_rtx_queue_unlink_and_free(skb, sk);
                        __skb_header_release(nskb);
                        tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
-                       sk->sk_wmem_queued += nskb->truesize;
+                       sk_wmem_queued_add(sk, nskb->truesize);
                        sk_mem_charge(sk, nskb->truesize);
                        skb = nskb;
                }
@@ -3447,7 +3447,7 @@ static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
 
        tcb->end_seq += skb->len;
        __skb_header_release(skb);
-       sk->sk_wmem_queued += skb->truesize;
+       sk_wmem_queued_add(sk, skb->truesize);
        sk_mem_charge(sk, skb->truesize);
        WRITE_ONCE(tp->write_seq, tcb->end_seq);
        tp->packets_out += tcp_skb_pcount(skb);
index 4c9122fc35c9d5f86ed60bc03427da1cde57b636..3177dcb173161629a801278db38fabeb6fcdbdd9 100644 (file)
@@ -446,7 +446,7 @@ META_COLLECTOR(int_sk_wmem_queued)
                *err = -1;
                return;
        }
-       dst->value = sk->sk_wmem_queued;
+       dst->value = READ_ONCE(sk->sk_wmem_queued);
 }
 
 META_COLLECTOR(int_sk_fwd_alloc)