[lkml]   [2018]   [Aug]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [BUG][BISECT] NFSv4 root failures after "fs/locks: allow a lock request to block other requests."
On Wed, Aug 15 2018, Jeff Layton wrote:

> On Wed, 2018-08-15 at 14:28 +0200, Krzysztof Kozlowski wrote:
>> Hi,
>> Bisect pointed commit ce3147990450a68b3f549088b30f087742a08b5d
>> ("fs/locks: allow a lock request to block other requests.") to failure
>> boot of NFSv4 with root on several boards.
>> Log is here:
>> With several errors:
>> kernel BUG at ../fs/locks.c:336!
>> Unable to handle kernel NULL pointer dereference at virtual address 00000004
>> Configuration:
>> 1. exynos_defconfig
>> 2. Arch ARM Linux
>> 3. Boards:
>> a. Odroid family (ARMv7, octa-core (Cortex-A7+A15), Exynos5422 SoC)
>> b. Toradex Colibri VF50 (ARMv7, UP, Cortex-A5)
>> 4. Systemd: v236, 238
>> 5. All boards boot from TFTP with NFS root (NFSv4)
>> On Colibri VF50 I got slightly different errors:
>> [ 11.663204] Internal error: Oops - undefined instruction: 0 [#1] ARM
>> [ 12.455273] Unable to handle kernel NULL pointer dereference at
>> virtual address 00000004
>> and only with some specific GCC (v6.3) or with other conditions which
>> I did not bisect yet. Maybe Colibri's failure is unrelated to that
>> commit.
>> Best regards,
>> Krzysztof

Thanks a lot for the report Krzysztof!!

> The BUG is due to a lock being freed when the fl_blocked list wasn't
> empty (implying that there were still blocked locks waiting on it).
> There are a number of calls to locks_delete_lock_ctx in posix_lock_inode
> and I don't think the fl_blocked list is being handled properly with all
> of them. It only transplants the blocked locks to a new lock when there
> are surviving locks on the list, and that may not be the case when the
> whole file is being unlocked.

locks_delete_lock_ctx() calls locks_unlink_lock_ctx() which calls
locks_wake_up_block() which doesn't only wake_up the blocks, but also
detached them. When that function completes, ->fl_blocked must be empty.

The trace shows the locks_free_lock() call at the end of fcntl_setlk64()
as the problematic call.
This suggests that do_lock_file_wait() exited with ->fl_blocked
non-empty, which it shouldn't.

I think we need to insert a call to locks_wake_up_block() in
do_lock_file_wait() before it returns.
I cannot find a sequence that would make this necessary, but
it isn't surprising that there might be one.

I'll dig through the code a bit more later and make sure I understand
what is happening.

[unhandled content-type:application/pgp-signature]
 \ /
  Last update: 2018-08-16 07:20    [W:0.052 / U:0.116 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site