On Wed, 2012-06-20 at 12:44 +0100, Mel Gorman wrote: > Commit [bad43ca8: net: introduce skb_try_coalesce()] introduced an > optimisation to coalesce skbs to reduce memory usage and cache line > misses. In the case where the socket is used for swapping this can result > in a warning like the following. > > [ 110.476565] nbd0: page allocation failure: order:0, mode:0x20 > [ 110.476568] Pid: 2714, comm: nbd0 Not tainted 3.5.0-rc2-swapnbd-v12r2-slab #3 > [ 110.476569] Call Trace: > [ 110.476573] [<ffffffff811042d3>] warn_alloc_failed+0xf3/0x160 > [ 110.476578] [<ffffffff81107c92>] __alloc_pages_nodemask+0x6e2/0x930 > [ 110.476582] [<ffffffff81107c92>] ? __alloc_pages_nodemask+0x6e2/0x930 > [ 110.476588] [<ffffffff81149f09>] kmem_getpages+0x59/0x1a0 > [ 110.476593] [<ffffffff8114ae5b>] fallback_alloc+0x17b/0x260 > [ 110.476597] [<ffffffff8114ac26>] ____cache_alloc_node+0x96/0x150 > [ 110.476602] [<ffffffff8114a458>] kmem_cache_alloc_node+0x78/0x1b0 > [ 110.476607] [<ffffffff8136c127>] __alloc_skb+0x57/0x1e0 > [ 110.476612] [<ffffffff813b9f81>] sk_stream_alloc_skb+0x41/0x120 > [ 110.476617] [<ffffffff813c8c72>] tcp_fragment+0x62/0x370 > [ 110.476622] [<ffffffff813c8fb9>] tso_fragment+0x39/0x180 > [ 110.476628] [<ffffffff813ca2a9>] tcp_write_xmit+0x1a9/0x3f0 > [ 110.476634] [<ffffffff813ca556>] __tcp_push_pending_frames+0x26/0xd0 > [ 110.476639] [<ffffffff813c61f5>] tcp_rcv_established+0x385/0x760 > [ 110.476644] [<ffffffff813ce671>] tcp_v4_do_rcv+0x111/0x1f0 > [ 110.476648] [<ffffffff81367259>] release_sock+0x99/0x140 > [ 110.476652] [<ffffffff813ba82b> tcp_sendmsg+0x7cb/0xe80 > [ 110.476657] [<ffffffff813df9b4>] inet_sendmsg+0x64/0xb0 > [ 110.476661] [<ffffffff811f0a00>] ? security_socket_sendmsg+0x10/0x20 > [ 110.476666] [<ffffffff81361dd8>] sock_sendmsg+0xf8/0x130 > [ 110.476672] [<ffffffff8124ba4c>] ? cpumask_next_and+0x3c/0x50 > [ 110.476677] [<ffffffff8107b053>] ? update_sd_lb_stats+0x123/0x620 > [ 110.476683] [<ffffffff8105164f>] ? recalc_sigpending+0x1f/0x70 > [ 110.476688] [<ffffffff81051e17>] ? __set_task_blocked+0x37/0x80 > [ 110.476693] [<ffffffff81361e51>] kernel_sendmsg+0x41/0x60 > [ 110.476698] [<ffffffffa048d417>] sock_xmit+0xb7/0x300 [nbd] > [ 110.476703] [<ffffffff8107bad7>] ? load_balance+0xd7/0x490 > [ 110.476710] [<ffffffffa048d7ac>] nbd_send_req+0x14c/0x270 [nbd] > [ 110.476716] [<ffffffffa048e21e>] nbd_handle_req+0x9e/0x180 [nbd] > [ 110.476721] [<ffffffffa048e4f2>] nbd_thread+0xb2/0x150 [nbd] > [ 110.476725] [<ffffffff81062580>] ? wake_up_bit+0x40/0x40 > [ 110.476730] [<ffffffffa048e440>] ? do_nbd_request+0x140/0x140 [nbd] > [ 110.476733] [<ffffffff81061d7e>] kthread+0x9e/0xb0 > [ 110.476739] [<ffffffff81439d64>] kernel_thread_helper+0x4/0x10 > [ 110.476743] [<ffffffff81061ce0>] ? flush_kthread_worker+0xc0/0xc0 > [ 110.476748] [<ffffffff81439d60>] ? gs_change+0x13/0x13 > > There were two ways this could be addressed. The first would be to > teach __tcp_push_pending_frames() to use __GFP_MEMALLOC if the socket > has SOCK_MEMALLOC set. This potentially defers the time of allocation > to a point where we are applying greater pressure on PFMEMALLOC reserves > which is undesirable. The second approach is to disable skb coalescing > for SOCK_MEMALLOC sockets and process them immediately. This patch takes > the second approach. > > Signed-off-by: Mel Gorman <mgorman@xxxxxxx> > --- > net/core/skbuff.c | 7 +++++++ > net/ipv4/tcp_input.c | 8 ++++++++ > 2 files changed, 15 insertions(+) > > diff --git a/net/core/skbuff.c b/net/core/skbuff.c > index d78671e..1d6ecc8 100644 > --- a/net/core/skbuff.c > +++ b/net/core/skbuff.c > @@ -3370,6 +3370,13 @@ bool skb_try_coalesce(struct sk_buff *to, struct sk_buff *from, > > *fragstolen = false; > > + /* > + * Avoid coalescing of SOCK_MEMALLOC socks are we do not want to defer > + * RX/TX to a time when pfmemallo reserves are under greater pressure > + */ > + if (sk_memalloc_socks() && sock_flag(to->sk, SOCK_MEMALLOC)) > + return false; > + > if (skb_cloned(to)) > return false; > > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c > index b224eb8..448f130 100644 > --- a/net/ipv4/tcp_input.c > +++ b/net/ipv4/tcp_input.c > @@ -4553,6 +4553,14 @@ static bool tcp_try_coalesce(struct sock *sk, > > *fragstolen = false; > > + /* > + * Do not attempt merging if the socket is used by the VM for swapping. > + * Attempts to defer can result in allocation failures during RX when > + * an attempt is made to push pending frames > + */ > + if (sk_memalloc_socks() && sock_flag(sk, SOCK_MEMALLOC)) > + return false; > + > if (tcp_hdr(from)->fin) > return false; > This makes absolutely no sense to me. This patch changes input path, while your stack trace is about output path and a packet being fragmented. If you really want to avoid this kind of thing, you'll need to disable TSO/GSO on the socket. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>