lkml.org 
[lkml]   [2014]   [Jun]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [Regression] 3.15 mmc related ext4 corruption with qemu-system-arm
From
On 12 June 2014 07:35, John Stultz <john.stultz@linaro.org> wrote:
> I've been seeing some ext4 corruption with recent kernels under qemu-system-arm.
>
> This issue seems to crop up after shutting down uncleanly (terminating
> qemu), shortly after booting about 50% of the time.
>
> ext4/mmc related dmesg details are:
> [ 3.206809] mmci-pl18x mb:mmci: mmc0: PL181 manf 41 rev0 at
> 0x10005000 irq 41,42 (pio)
> [ 3.268316] mmc0: new SDHC card at address 4567
> [ 3.281963] mmcblk0: mmc0:4567 QEMU! 2.00 GiB
> [ 3.315699] mmcblk0: p1 p2 p3 p4 < p5 p6 >
> ...
> [ 11.806169] EXT4-fs (mmcblk0p5): Ignoring removed nomblk_io_submit option
> [ 11.904714] EXT4-fs (mmcblk0p5): recovery complete
> [ 11.905854] EXT4-fs (mmcblk0p5): mounted filesystem with ordered
> data mode. Opts: nomblk_io_submit,errors=panic
> ...
> [ 91.558824] EXT4-fs error (device mmcblk0p5):
> ext4_mb_generate_buddy:756: group 1, 2252 clusters in bitmap, 2284 in
> gd; block bitmap corrupt.
> [ 91.560641] Aborting journal on device mmcblk0p5-8.
> [ 91.562589] Kernel panic - not syncing: EXT4-fs (device mmcblk0p5):
> panic forced after error
> [ 91.562589]
> [ 91.563486] CPU: 0 PID: 1 Comm: init Not tainted 3.15.0-rc1 #560
> [ 91.564616] [<c00116e5>] (unwind_backtrace) from [<c000f3b1>]
> (show_stack+0x11/0x14)
> [ 91.565154] [<c000f3b1>] (show_stack) from [<c04262b1>]
> (dump_stack+0x59/0x7c)
> [ 91.565666] [<c04262b1>] (dump_stack) from [<c0423297>] (panic+0x67/0x178)
> [ 91.566147] [<c0423297>] (panic) from [<c0134bb1>]
> (ext4_handle_error+0x69/0x74)
> [ 91.566659] [<c0134bb1>] (ext4_handle_error) from [<c0135437>]
> (__ext4_grp_locked_error+0x6b/0x160)
> [ 91.567223] [<c0135437>] (__ext4_grp_locked_error) from
> [<c0143079>] (ext4_mb_generate_buddy+0x1b1/0x29c)
> [ 91.567860] [<c0143079>] (ext4_mb_generate_buddy) from [<c01447e5>]
> (ext4_mb_init_cache+0x219/0x4e0)
> [ 91.568473] [<c01447e5>] (ext4_mb_init_cache) from [<c0144b67>]
> (ext4_mb_init_group+0xbb/0x138)
> [ 91.569021] [<c0144b67>] (ext4_mb_init_group) from [<c0144cd7>]
> (ext4_mb_good_group+0xf3/0xfc)
> [ 91.569659] [<c0144cd7>] (ext4_mb_good_group) from [<c0145c8f>]
> (ext4_mb_regular_allocator+0x153/0x2c4)
> [ 91.570250] [<c0145c8f>] (ext4_mb_regular_allocator) from
> [<c0148095>] (ext4_mb_new_blocks+0x2fd/0x4e4)
> [ 91.570868] [<c0148095>] (ext4_mb_new_blocks) from [<c013f931>]
> (ext4_ext_map_blocks+0x965/0x10bc)
> [ 91.571444] [<c013f931>] (ext4_ext_map_blocks) from [<c0122c8b>]
> (ext4_map_blocks+0xfb/0x36c)
> [ 91.571992] [<c0122c8b>] (ext4_map_blocks) from [<c01263b1>]
> (mpage_map_and_submit_extent+0x99/0x5f0)
> [ 91.572614] [<c01263b1>] (mpage_map_and_submit_extent) from
> [<c0126bc1>] (ext4_writepages+0x2b9/0x4e8)
> [ 91.573201] [<c0126bc1>] (ext4_writepages) from [<c0094ae9>]
> (do_writepages+0x19/0x28)
> [ 91.573709] [<c0094ae9>] (do_writepages) from [<c008c811>]
> (__filemap_fdatawrite_range+0x3d/0x44)
> [ 91.574265] [<c008c811>] (__filemap_fdatawrite_range) from
> [<c008c883>] (filemap_flush+0x23/0x28)
> [ 91.574854] [<c008c883>] (filemap_flush) from [<c012bf75>]
> (ext4_rename+0x2f9/0x3e4)
> [ 91.575360] [<c012bf75>] (ext4_rename) from [<c00c3363>]
> (vfs_rename+0x183/0x45c)
> [ 91.575911] [<c00c3363>] (vfs_rename) from [<c00c3867>]
> (SyS_renameat2+0x22b/0x26c)
> [ 91.576460] [<c00c3867>] (SyS_renameat2) from [<c00c38df>]
> (SyS_rename+0x1f/0x24)
> [ 91.576961] [<c00c38df>] (SyS_rename) from [<c000cd01>]
> (ret_fast_syscall+0x1/0x5c)
>
>
> Bisecting this points to: e7f3d22289e4307b3071cc18b1d8ecc6598c0be4
> (mmc: mmci: Handle CMD irq before DATA irq). Which I guess shouldn't
> be surprising, as I saw problems with that patch earlier in the
> 3.15-rc cycle:
> https://lkml.org/lkml/2014/4/14/824
>
> However that discussion petered out (possibly my fault for not
> following up) as to if it was an issue with the patch or a issue with
> qemu. Then the original issue disappeared for me, which I figured was
> due to a fix upstream, but now I'm guessing coincided with me updating
> my system and getting qemu v2.0 (where as previously I was on 1.5).
>
> $ qemu-system-arm -version
> QEMU emulator version 2.0.0 (Debian 2.0.0+dfsg-2ubuntu1.1), Copyright
> (c) 2003-2008 Fabrice Bellard
>
> While the previous behavior was annoying and kept my emulated
> environments from booting, this while a bit more rare and subtle eats
> the disks, which is much more painful for my testing.
>
> Unfortunately reverting the change (manually, as it doesn't revert
> cleanly anymore) doesn't seem to completely avoid the issue, so the
> bisection may have gone slightly astray (though it is interesting it
> landed on the same commit I earlier had trouble with). So I'll
> back-track and double check some of the last few "good" results to
> validate I didn't just luck into 3 good boots accidentally. I'll also
> review my revert in case I missed something subtle in doing it
> manually.
>
> Anyway, if there is any thoughts on how to better chase this down and
> debug it, I'd appreciate it! I can also provide reproduction
> instructions with a pre-built Linaro android disk image and hand built
> kernel if anyone wants to debug this themselves.

According to your log, the primecell-periphid is 0x00041181, which
means mmci will be using the arm_variant.

A simple fix; for the arm_variant, go back to use the old behaviour.

A quite simple fix; Invent a new primecell-periphid and a new
corresponding variant and use the old behaviour for this variant. The
new primecell-periphid then needs to be provided through DT for the
QEMU dtb.

Is there any of the above solution you see as the preferred one?

Kind regards
Uffe


\
 
 \ /
  Last update: 2014-06-12 14:21    [W:0.096 / U:0.236 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site