[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CALvZod6Sw-2Wh0KEBiMgGZ1c+2nFW0ueL_4TM4d=Z0JcbvSXrw@mail.gmail.com>
Date: Thu, 31 Oct 2019 11:30:57 -0700
From: Shakeel Butt <shakeelb@...gle.com>
To: Eric Dumazet <eric.dumazet@...il.com>
Cc: Tejun Heo <tj@...nel.org>, Michal Hocko <mhocko@...nel.org>,
"David S. Miller" <davem@...emloft.net>, netdev@...r.kernel.org,
Kernel Team <kernel-team@...com>,
LKML <linux-kernel@...r.kernel.org>,
Josef Bacik <josef@...icpanda.com>,
Jakub Kicinski <jakub.kicinski@...ronome.com>,
Johannes Weiner <hannes@...xchg.org>,
Linux MM <linux-mm@...ck.org>, Mel Gorman <mgorman@...e.de>,
Andrew Morton <akpm@...ux-foundation.org>
Subject: Re: [PATCH v2] net: fix sk_page_frag() recursion from memory reclaim
On Thu, Oct 31, 2019 at 10:47 AM Eric Dumazet <eric.dumazet@...il.com> wrote:
>
>
>
> On 10/31/19 10:35 AM, Shakeel Butt wrote:
> > +Michal Hocko
> >
> > On Thu, Oct 24, 2019 at 1:50 PM Tejun Heo <tj@...nel.org> wrote:
> >>
> >> sk_page_frag() optimizes skb_frag allocations by using per-task
> >> skb_frag cache when it knows it's the only user. The condition is
> >> determined by seeing whether the socket allocation mask allows
> >> blocking - if the allocation may block, it obviously owns the task's
> >> context and ergo exclusively owns current->task_frag.
> >>
> >> Unfortunately, this misses recursion through memory reclaim path.
> >> Please take a look at the following backtrace.
> >>
> >> [2] RIP: 0010:tcp_sendmsg_locked+0xccf/0xe10
> >> ...
> >> tcp_sendmsg+0x27/0x40
> >> sock_sendmsg+0x30/0x40
> >> sock_xmit.isra.24+0xa1/0x170 [nbd]
> >> nbd_send_cmd+0x1d2/0x690 [nbd]
> >> nbd_queue_rq+0x1b5/0x3b0 [nbd]
> >> __blk_mq_try_issue_directly+0x108/0x1b0
> >> blk_mq_request_issue_directly+0xbd/0xe0
> >> blk_mq_try_issue_list_directly+0x41/0xb0
> >> blk_mq_sched_insert_requests+0xa2/0xe0
> >> blk_mq_flush_plug_list+0x205/0x2a0
> >> blk_flush_plug_list+0xc3/0xf0
> >> [1] blk_finish_plug+0x21/0x2e
> >> _xfs_buf_ioapply+0x313/0x460
> >> __xfs_buf_submit+0x67/0x220
> >> xfs_buf_read_map+0x113/0x1a0
> >> xfs_trans_read_buf_map+0xbf/0x330
> >> xfs_btree_read_buf_block.constprop.42+0x95/0xd0
> >> xfs_btree_lookup_get_block+0x95/0x170
> >> xfs_btree_lookup+0xcc/0x470
> >> xfs_bmap_del_extent_real+0x254/0x9a0
> >> __xfs_bunmapi+0x45c/0xab0
> >> xfs_bunmapi+0x15/0x30
> >> xfs_itruncate_extents_flags+0xca/0x250
> >> xfs_free_eofblocks+0x181/0x1e0
> >> xfs_fs_destroy_inode+0xa8/0x1b0
> >> destroy_inode+0x38/0x70
> >> dispose_list+0x35/0x50
> >> prune_icache_sb+0x52/0x70
> >> super_cache_scan+0x120/0x1a0
> >> do_shrink_slab+0x120/0x290
> >> shrink_slab+0x216/0x2b0
> >> shrink_node+0x1b6/0x4a0
> >> do_try_to_free_pages+0xc6/0x370
> >> try_to_free_mem_cgroup_pages+0xe3/0x1e0
> >> try_charge+0x29e/0x790
> >> mem_cgroup_charge_skmem+0x6a/0x100
> >> __sk_mem_raise_allocated+0x18e/0x390
> >> __sk_mem_schedule+0x2a/0x40
> >> [0] tcp_sendmsg_locked+0x8eb/0xe10
> >> tcp_sendmsg+0x27/0x40
> >> sock_sendmsg+0x30/0x40
> >> ___sys_sendmsg+0x26d/0x2b0
> >> __sys_sendmsg+0x57/0xa0
> >> do_syscall_64+0x42/0x100
> >> entry_SYSCALL_64_after_hwframe+0x44/0xa9
> >>
> >> In [0], tcp_send_msg_locked() was using current->page_frag when it
> >> called sk_wmem_schedule(). It already calculated how many bytes can
> >> be fit into current->page_frag. Due to memory pressure,
> >> sk_wmem_schedule() called into memory reclaim path which called into
> >> xfs and then IO issue path. Because the filesystem in question is
> >> backed by nbd, the control goes back into the tcp layer - back into
> >> tcp_sendmsg_locked().
> >>
> >> nbd sets sk_allocation to (GFP_NOIO | __GFP_MEMALLOC) which makes
> >> sense - it's in the process of freeing memory and wants to be able to,
> >> e.g., drop clean pages to make forward progress. However, this
> >> confused sk_page_frag() called from [2]. Because it only tests
> >> whether the allocation allows blocking which it does, it now thinks
> >> current->page_frag can be used again although it already was being
> >> used in [0].
> >>
> >> After [2] used current->page_frag, the offset would be increased by
> >> the used amount. When the control returns to [0],
> >> current->page_frag's offset is increased and the previously calculated
> >> number of bytes now may overrun the end of allocated memory leading to
> >> silent memory corruptions.
> >>
> >> Fix it by adding gfpflags_normal_context() which tests sleepable &&
> >> !reclaim and use it to determine whether to use current->task_frag.
> >>
> >> v2: Eric didn't like gfp flags being tested twice. Introduce a new
> >> helper gfpflags_normal_context() and combine the two tests.
> >>
> >> Signed-off-by: Tejun Heo <tj@...nel.org>
> >> Cc: Josef Bacik <josef@...icpanda.com>
> >> Cc: Eric Dumazet <eric.dumazet@...il.com>
> >> Cc: stable@...r.kernel.org
> >> ---
> >> include/linux/gfp.h | 23 +++++++++++++++++++++++
> >> include/net/sock.h | 11 ++++++++---
> >> 2 files changed, 31 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> >> index fb07b503dc45..61f2f6ff9467 100644
> >> --- a/include/linux/gfp.h
> >> +++ b/include/linux/gfp.h
> >> @@ -325,6 +325,29 @@ static inline bool gfpflags_allow_blocking(const gfp_t gfp_flags)
> >> return !!(gfp_flags & __GFP_DIRECT_RECLAIM);
> >> }
> >>
> >> +/**
> >> + * gfpflags_normal_context - is gfp_flags a normal sleepable context?
> >> + * @gfp_flags: gfp_flags to test
> >> + *
> >> + * Test whether @gfp_flags indicates that the allocation is from the
> >> + * %current context and allowed to sleep.
> >> + *
> >> + * An allocation being allowed to block doesn't mean it owns the %current
> >> + * context. When direct reclaim path tries to allocate memory, the
> >> + * allocation context is nested inside whatever %current was doing at the
> >> + * time of the original allocation. The nested allocation may be allowed
> >> + * to block but modifying anything %current owns can corrupt the outer
> >> + * context's expectations.
> >> + *
> >> + * %true result from this function indicates that the allocation context
> >> + * can sleep and use anything that's associated with %current.
> >> + */
> >> +static inline bool gfpflags_normal_context(const gfp_t gfp_flags)
> >> +{
> >> + return (gfp_flags & (__GFP_DIRECT_RECLAIM | __GFP_MEMALLOC)) ==
> >> + __GFP_DIRECT_RECLAIM;
> >
> > I think we should be checking PF_MEMALLOC here instead. Something like:
> >
> > return gfpflags_allow_blocking(gfp_flags) && !(current->flags & PF_MEMALLOC);
> >
> > In my limited understanding, __GFP_MEMALLOC gives access to reserve
> > but we have overloaded PF_MEMALLOC to also define the reclaim context.
> > There are PF_MEMALLOC users which does not use __GFP_MEMALLOC like
> > iscsi_sw_tcp_pdu_xmit() which can call sock_sendmsg().
>
> Why would this layer not set sk->sk_allocation to GFP_ATOMIC ?
>
> And it also might call sk_set_memalloc() too.
>
> Please double check scsi layer, I am pretty sure it did well at some point.
>
Yes, you are right, quoted the wrong example. SCSI is indeed setting
sk->sk_allocation to GFP_ATOMIC and sk_set_memalloc() in
iscsi_sw_tcp_conn_bind().
Basically what I wanted to say that MM treats PF_MEMALLOC as the
reclaim context while __GFP_MEMALLOC just tells to give access to the
reserves. As gfpflags_allow_blocking() can be used beyond net
subsystem, my only concern is its potential usage under PF_MEMALLOC
context but without __GFP_MEMALLOC.
Powered by blists - more mailing lists