src/patches/suse-2.6.27.25/patches.suse/SoN-23-netvm-tcp-deadlock.patch

   1 From: Peter Zijlstra <a.p.zijlstra@chello.nl>
   2 Subject: netvm: prevent a stream specific deadlock
   3 Patch-mainline: No
   4 References: FATE#303834
   5
   6 It could happen that all !SOCK_MEMALLOC sockets have buffered so much data
   7 that we're over the global rmem limit. This will prevent SOCK_MEMALLOC buffers
   8 from receiving data, which will prevent userspace from running, which is needed
   9 to reduce the buffered data.
  10
  11 Fix this by exempting the SOCK_MEMALLOC sockets from the rmem limit.
  12
  13 Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
  14 Acked-by: Neil Brown <neilb@suse.de>
  15 Acked-by: Suresh Jayaraman <sjayaraman@suse.de>
  16
  17 ---
  18  include/net/sock.h   |    7 ++++---
  19  net/core/sock.c      |    2 +-
  20  net/ipv4/tcp_input.c |   12 ++++++------
  21  net/sctp/ulpevent.c  |    2 +-
  22  4 files changed, 12 insertions(+), 11 deletions(-)
  23
  24 --- a/include/net/sock.h
  25 +++ b/include/net/sock.h
  26 @@ -788,12 +788,13 @@ static inline int sk_wmem_schedule(struc
  27                 __sk_mem_schedule(sk, size, SK_MEM_SEND);
  28  }
  29
  30 -static inline int sk_rmem_schedule(struct sock *sk, int size)
  31 +static inline int sk_rmem_schedule(struct sock *sk, struct sk_buff *skb)
  32  {
  33         if (!sk_has_account(sk))
  34                 return 1;
  35 -       return size <= sk->sk_forward_alloc ||
  36 -               __sk_mem_schedule(sk, size, SK_MEM_RECV);
  37 +       return skb->truesize <= sk->sk_forward_alloc ||
  38 +               __sk_mem_schedule(sk, skb->truesize, SK_MEM_RECV) ||
  39 +               skb_emergency(skb);
  40  }
  41
  42  static inline void sk_mem_reclaim(struct sock *sk)
  43 --- a/net/core/sock.c
  44 +++ b/net/core/sock.c
  45 @@ -381,7 +381,7 @@ int sock_queue_rcv_skb(struct sock *sk,
  46         if (err)
  47                 goto out;
  48
  49 -       if (!sk_rmem_schedule(sk, skb->truesize)) {
  50 +       if (!sk_rmem_schedule(sk, skb)) {
  51                 err = -ENOBUFS;
  52                 goto out;
  53         }
  54 --- a/net/ipv4/tcp_input.c
  55 +++ b/net/ipv4/tcp_input.c
  56 @@ -3939,19 +3939,19 @@ static void tcp_ofo_queue(struct sock *s
  57  static int tcp_prune_ofo_queue(struct sock *sk);
  58  static int tcp_prune_queue(struct sock *sk);
  59
  60 -static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
  61 +static inline int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb)
  62  {
  63         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  64 -           !sk_rmem_schedule(sk, size)) {
  65 +           !sk_rmem_schedule(sk, skb)) {
  66
  67                 if (tcp_prune_queue(sk) < 0)
  68                         return -1;
  69
  70 -               if (!sk_rmem_schedule(sk, size)) {
  71 +               if (!sk_rmem_schedule(sk, skb)) {
  72                         if (!tcp_prune_ofo_queue(sk))
  73                                 return -1;
  74
  75 -                       if (!sk_rmem_schedule(sk, size))
  76 +                       if (!sk_rmem_schedule(sk, skb))
  77                                 return -1;
  78                 }
  79         }
  80 @@ -4006,7 +4006,7 @@ static void tcp_data_queue(struct sock *
  81                 if (eaten <= 0) {
  82  queue_and_out:
  83                         if (eaten < 0 &&
  84 -                           tcp_try_rmem_schedule(sk, skb->truesize))
  85 +                           tcp_try_rmem_schedule(sk, skb))
  86                                 goto drop;
  87
  88                         skb_set_owner_r(skb, sk);
  89 @@ -4077,7 +4077,7 @@ drop:
  90
  91         TCP_ECN_check_ce(tp, skb);
  92
  93 -       if (tcp_try_rmem_schedule(sk, skb->truesize))
  94 +       if (tcp_try_rmem_schedule(sk, skb))
  95                 goto drop;
  96
  97         /* Disable header prediction. */
  98 --- a/net/sctp/ulpevent.c
  99 +++ b/net/sctp/ulpevent.c
 100 @@ -701,7 +701,7 @@ struct sctp_ulpevent *sctp_ulpevent_make
 101         if (rx_count >= asoc->base.sk->sk_rcvbuf) {
 102
 103                 if ((asoc->base.sk->sk_userlocks & SOCK_RCVBUF_LOCK) ||
 104 -                   (!sk_rmem_schedule(asoc->base.sk, chunk->skb->truesize)))
 105 +                   (!sk_rmem_schedule(asoc->base.sk, chunk->skb)))
 106                         goto fail;
 107         }
 108