Re: [PATCH] block: fix a crash in do_task_dead()

From: Jens Axboe <axboe@kernel.dk>
To: Qian Cai <cai@lca.pw>
Cc: akpm@linux-foundation.org, hch@lst.de, peterz@infradead.org,
	oleg@redhat.com, gkohli@codeaurora.org, mingo@redhat.com,
	linux-mm@kvack.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] block: fix a crash in do_task_dead()
Date: Wed, 29 May 2019 14:31:58 -0600	[thread overview]
Message-ID: <ad24b8de-c1dc-f52a-06af-103ceda891a6@kernel.dk> (raw)
In-Reply-To: <1559161526-618-1-git-send-email-cai@lca.pw>

On 5/29/19 2:25 PM, Qian Cai wrote:
> The commit 0619317ff8ba ("block: add polled wakeup task helper")
> replaced wake_up_process() with blk_wake_io_task() in
> end_swap_bio_read() which triggers a crash when running heavy swapping
> workloads.
> 
> [T114538] kernel BUG at kernel/sched/core.c:3462!
> [T114538] Process oom01 (pid: 114538, stack limit = 0x000000004f40e0c1)
> [T114538] Call trace:
> [T114538]  do_task_dead+0xf0/0xf8
> [T114538]  do_exit+0xd5c/0x10fc
> [T114538]  do_group_exit+0xf4/0x110
> [T114538]  get_signal+0x280/0xdd8
> [T114538]  do_notify_resume+0x720/0x968
> [T114538]  work_pending+0x8/0x10
> 
> This is because shortly after set_special_state(TASK_DEAD),
> end_swap_bio_read() is called from an interrupt handler that revive the
> task state to TASK_RUNNING causes __schedule() to return and trip the
> BUG() later.
> 
> [  C206] Call trace:
> [  C206]  dump_backtrace+0x0/0x268
> [  C206]  show_stack+0x20/0x2c
> [  C206]  dump_stack+0xb4/0x108
> [  C206]  blk_wake_io_task+0x7c/0x80
> [  C206]  end_swap_bio_read+0x22c/0x31c
> [  C206]  bio_endio+0x3d8/0x414
> [  C206]  dec_pending+0x280/0x378 [dm_mod]
> [  C206]  clone_endio+0x128/0x2ac [dm_mod]
> [  C206]  bio_endio+0x3d8/0x414
> [  C206]  blk_update_request+0x3ac/0x924
> [  C206]  scsi_end_request+0x54/0x350
> [  C206]  scsi_io_completion+0xf0/0x6f4
> [  C206]  scsi_finish_command+0x214/0x228
> [  C206]  scsi_softirq_done+0x170/0x1a4
> [  C206]  blk_done_softirq+0x100/0x194
> [  C206]  __do_softirq+0x350/0x790
> [  C206]  irq_exit+0x200/0x26c
> [  C206]  handle_IPI+0x2e8/0x514
> [  C206]  gic_handle_irq+0x224/0x228
> [  C206]  el1_irq+0xb8/0x140
> [  C206]  _raw_spin_unlock_irqrestore+0x3c/0x74
> [  C206]  do_task_dead+0x88/0xf8
> [  C206]  do_exit+0xd5c/0x10fc
> [  C206]  do_group_exit+0xf4/0x110
> [  C206]  get_signal+0x280/0xdd8
> [  C206]  do_notify_resume+0x720/0x968
> [  C206]  work_pending+0x8/0x10
> 
> Before the offensive commit, wake_up_process() will prevent this from
> happening by taking the pi_lock and bail out immediately if TASK_DEAD is
> set.
> 
> if (!(p->state & TASK_NORMAL))
> 	goto out;
> 
> Fix it by calling wake_up_process() if it is in a non-task context.

I like this one a lot better than the previous fix. Unless folks
object, I'll queue this up for 5.2, thanks.

-- 
Jens Axboe