linux-kernel - Re: [MD] Crash with 4.12+ kernel and high disk load -- bisected to 4ad23a976413: MD: use per-cpu counter for writes

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Tue, 08 Aug 2017 17:04:07 +1000
From:   NeilBrown <neilb@...e.com>
To:     David R <david@...olicited.net>, Shaohua Li <shli@...nel.org>
Cc:     Dominik Brodowski <linux@...inikbrodowski.net>,
        Shaohua Li <shli@...com>, linux-kernel@...r.kernel.org,
        linux-raid@...r.kernel.org
Subject: Re: [MD] Crash with 4.12+ kernel and high disk load -- bisected to 4ad23a976413: MD: use per-cpu counter for writes_pending

On Tue, Aug 08 2017, David R wrote:

> Quoting Shaohua Li <shli@...nel.org>:
>
>> Spent some time to check this one, unfortunately I can't find how that patch
>> makes rcu stall. the percpu part looks good to me too. Can you  
>> double check if
>> reverting 4ad23a976413aa57 makes the issue go away? When the rcu  
>> stall happens,
>> what the /sys/block/md/md0/array_state? please also attach /proc/mdstat. When
>> you say the mdx_raid1 threads are in 'R' state, can you double check if the
>> /proc/pid/stack always 0xffffffffff?
>>
>> Thanks,
>> Shaohua
>
> I confess to knowing absolutely nothing about the md code, so please  
> don't be too hard on me. However
> :-
>
> static bool set_in_sync(struct mddev *mddev)
> {
>      WARN_ON_ONCE(!spin_is_locked(&mddev->lock));
>      if (!mddev->in_sync) {
>          mddev->sync_checkers++;
>          spin_unlock(&mddev->lock);
>          percpu_ref_switch_to_atomic_sync(&mddev->writes_pending);
>          spin_lock(&mddev->lock);
>          if (!mddev->in_sync &&
>              percpu_ref_is_zero(&mddev->writes_pending)) {
>              mddev->in_sync = 1;
>              /*
>               * Ensure ->in_sync is visible before we clear
>               * ->sync_checkers.
>               */
>              smp_mb();
>              set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
>              sysfs_notify_dirent_safe(mddev->sysfs_state);
>          }
>          if (--mddev->sync_checkers == 0)
>              percpu_ref_switch_to_percpu(&mddev->writes_pending);
>
>
> The switch_to_percpu() takes place under mddev->lock however  
> switch_to_atomic_sync() does not. A thread can be in the middle of (or  
> about to execute) switch_to_atomic_sync() at the same time as another  
> is calling switch_to_percpu(). This can't be correct surely?

No, that wouldn't be correct.

When switch_to_atomic_sync is called, ->sync_checkers > 0.
When switch_to_percpu is called ->sync_checkers == 0.

So they cannot happen at the same time.

Thanks for looking at the code!

NeilBrown

>
> Cheers
> David

Download attachment "signature.asc" of type "application/pgp-signature" (833 bytes)