Re: INFO: task hung in __sb_start_write

[Dmitry Vyukov <dvyukov@google.com>]

On Tue, Jun 19, 2018 at 1:10 PM, Tetsuo Handa
<penguin-kernel@i-love.sakura.ne.jp> wrote:
> On 2018/06/16 4:40, Tetsuo Handa wrote:
>> Hmm, there might be other locations calling percpu_rwsem_release() ?
>
> There are other locations calling percpu_rwsem_release(), but quite few.
>
> include/linux/fs.h:1494:#define __sb_writers_release(sb, lev)   \
> include/linux/fs.h-1495-        percpu_rwsem_release(&(sb)->s_writers.rw_sem[(lev)-1], 1, _THIS_IP_)
>
> fs/btrfs/transaction.c:1821:            __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
> fs/aio.c:1566:                  __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
> fs/xfs/xfs_aops.c:211:  __sb_writers_release(ioend->io_inode->i_sb, SB_FREEZE_FS);
>
>
>
> I'd like to check what atomic_long_read(&sem->rw_sem.count) says
> when hung task is reported.
>
> Dmitry, if you succeed to reproduce khungtaskd messages, can you try this patch?
>
> ---
>  include/linux/sched.h         |  1 +
>  kernel/hung_task.c            | 25 +++++++++++++++++++++++++
>  kernel/locking/percpu-rwsem.c |  2 ++
>  3 files changed, 28 insertions(+)
>
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 87bf02d..af95251 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1179,6 +1179,7 @@ struct task_struct {
>         /* Used by LSM modules for access restriction: */
>         void                            *security;
>  #endif
> +       struct percpu_rw_semaphore      *pcpu_rwsem_read_waiting;
>
>         /*
>          * New fields for task_struct should be added above here, so that
> diff --git a/kernel/hung_task.c b/kernel/hung_task.c
> index 32b4794..409cc82 100644
> --- a/kernel/hung_task.c
> +++ b/kernel/hung_task.c
> @@ -18,6 +18,7 @@
>  #include <linux/utsname.h>
>  #include <linux/sched/signal.h>
>  #include <linux/sched/debug.h>
> +#include <linux/percpu-rwsem.h>
>
>  #include <trace/events/sched.h>
>
> @@ -77,6 +78,29 @@ static int __init hung_task_panic_setup(char *str)
>         .notifier_call = hung_task_panic,
>  };
>
> +static void dump_percpu_rwsem_state(struct percpu_rw_semaphore *sem)
> +{
> +       unsigned int sum = 0;
> +       int cpu;
> +
> +       if (!sem)
> +               return;
> +       pr_info("percpu_rw_semaphore(%p)\n", sem);
> +       pr_info("  ->rw_sem.count=0x%lx\n",
> +               atomic_long_read(&sem->rw_sem.count));
> +       pr_info("  ->rss.gp_state=%d\n", sem->rss.gp_state);
> +       pr_info("  ->rss.gp_count=%d\n", sem->rss.gp_count);
> +       pr_info("  ->rss.cb_state=%d\n", sem->rss.cb_state);
> +       pr_info("  ->rss.gp_type=%d\n", sem->rss.gp_type);
> +       pr_info("  ->readers_block=%d\n", sem->readers_block);
> +       for_each_possible_cpu(cpu)
> +               sum += per_cpu(*sem->read_count, cpu);
> +       pr_info("  ->read_count=%d\n", sum);
> +       pr_info("  ->list_empty(rw_sem.wait_list)=%d\n",
> +              list_empty(&sem->rw_sem.wait_list));
> +       pr_info("  ->writer.task=%p\n", sem->writer.task);
> +}
> +
>  static void check_hung_task(struct task_struct *t, unsigned long timeout)
>  {
>         unsigned long switch_count = t->nvcsw + t->nivcsw;
> @@ -122,6 +146,7 @@ static void check_hung_task(struct task_struct *t, unsigned long timeout)
>                 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
>                         " disables this message.\n");
>                 sched_show_task(t);
> +               dump_percpu_rwsem_state(t->pcpu_rwsem_read_waiting);
>                 hung_task_show_lock = true;
>         }
>
> diff --git a/kernel/locking/percpu-rwsem.c b/kernel/locking/percpu-rwsem.c
> index 883cf1b..b3654ab 100644
> --- a/kernel/locking/percpu-rwsem.c
> +++ b/kernel/locking/percpu-rwsem.c
> @@ -82,7 +82,9 @@ int __percpu_down_read(struct percpu_rw_semaphore *sem, int try)
>         /*
>          * Avoid lockdep for the down/up_read() we already have them.
>          */
> +       current->pcpu_rwsem_read_waiting = sem;
>         __down_read(&sem->rw_sem);
> +       current->pcpu_rwsem_read_waiting = NULL;
>         this_cpu_inc(*sem->read_count);
>         __up_read(&sem->rw_sem);
>
> --
> 1.8.3.1



I wasn't able to reproduce this locally yet.

> If no, we would want a git tree for testing under syzbot.

Thinking of this, we could setup a sandbox instance that won't report
anything over email at all, but crashes will be available on the web.
We could point this instance to a custom git tree, where additional
patches can be applied as necessary.  The main question is: who and
how will manage this tree? The tree needs to be rebased periodically,
patches applied, old patches taken out.