Re: BISECT result: 6.0.0-RC kernels trigger Firefox snap bug with 6.0.0-rc3 through 6.0.0-rc7
From: Mirsad Todorovac
Date: Tue Oct 18 2022 - 05:23:46 EST
On 10/18/22 10:24, Hsin-Yi Wang wrote:
On Tue, Oct 18, 2022 at 2:52 PM Mirsad Todorovac
<mirsad.todorovac@xxxxxxxxxxxx> wrote:
On 10/18/22 04:15, Jintao Yin wrote:
On Sat, Oct 15, 2022 at 09:59:36PM +0100, Phillip Lougher wrote:
Thorsten Leemhuis <regressions@xxxxxxxxxxxxx> wrote:
Topposting, to make this easier to access for everyone.
@Mirsad, thx for bisecting.
@Phillip: if you want to see a problem description and the whole
backstory of the problem that apparently is caused by b09a7a036d20
("squashfs: support reading fragments in readahead call"), see this
thread (Mirsad sadly started a new one with the quoted mail below):
https://lore.kernel.org/all/b0c258c3-6dcf-aade-efc4-d62a8b3a1ce2@xxxxxxxxxxxx/
The above backstory tends to suggest data corruption which is happening
after a couple of hours especially on heavy loads, e.g. the comment
On 10/3/22 at 4:18 AM, Mirsad Goran Todorovac wrote:
The bug usually isn't showing immediately, but after a couple of hours
of running (especially with multimedia running inside Firefox).
Which is typically caused by double freed buffers or race conditions in
freeing and reusing.
Thanks Mirsad for the following
On Sat, 15 Oct 2022 16:59:44 +0200, Mirsad Goran Todorovac wrote:
Here are the results of the requested bisect on the bug involving the
Firefox snap build 104.x, 105.0.x, squashfs and which was manifested on
both Ubuntu snap and with snapd in AlmaLinux 8.6 (CentOS fork):
mtodorov@domac:~/linux/kernel/linux_stable$ git bisect log
git bisect start
# bad: [568035b01cfb107af8d2e4bd2fb9aea22cf5b868] Linux 6.0-rc1
git bisect bad 568035b01cfb107af8d2e4bd2fb9aea22cf5b868
# good: [51dd976781da8c0b47e106ed59a96d7c28972ce4] Linux 5.19.15
git bisect good 51dd976781da8c0b47e106ed59a96d7c28972ce4
# good: [3d7cb6b04c3f3115719235cc6866b10326de34cd] Linux 5.19
git bisect good 3d7cb6b04c3f3115719235cc6866b10326de34cd
# good: [b44f2fd87919b5ae6e1756d4c7ba2cbba22238e1] Merge tag
'drm-next-2022-08-03' of git://anongit.freedesktop.org/drm/drm
git bisect good b44f2fd87919b5ae6e1756d4c7ba2cbba22238e1
# good: [6614a3c3164a5df2b54abb0b3559f51041cf705b] Merge tag
'mm-stable-2022-08-03' of
git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm
git bisect good 6614a3c3164a5df2b54abb0b3559f51041cf705b
# bad: [eb5699ba31558bdb2cee6ebde3d0a68091e47dce] Merge tag
'mm-nonmm-stable-2022-08-06-2' of
git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm
git bisect bad eb5699ba31558bdb2cee6ebde3d0a68091e47dce
# good: [24df5428ef9d1ca1edd54eca7eb667110f2dfae3] ALSA: hda/realtek:
Add quirk for HP Spectre x360 15-eb0xxx
git bisect good 24df5428ef9d1ca1edd54eca7eb667110f2dfae3
# good: [c993e07be023acdeec8e84e2e0743c52adb5fc94] Merge tag
'dma-mapping-5.20-2022-08-06' of
git://git.infradead.org/users/hch/dma-mapping
git bisect good c993e07be023acdeec8e84e2e0743c52adb5fc94
# good: [4cfa6ff24a9744ba484521c38bea613134fbfcb3] powerpc/64e: Fix
kexec build error
git bisect good 4cfa6ff24a9744ba484521c38bea613134fbfcb3
# good: [24cb958695724ffb4488ef4f65892c0767bcd2f2] Merge tag
's390-5.20-1' of git://git.kernel.org/pub/scm/linux/kernel/git/s390/linux
git bisect good 24cb958695724ffb4488ef4f65892c0767bcd2f2
# good: [db98b43086275350294f5c6f797249b714d6316d] squashfs: always
build "file direct" version of page actor
git bisect good db98b43086275350294f5c6f797249b714d6316d
# good: [6ba592fa014f21f35a8ee8da4ca7b95a018f13e8] video: fbdev: s3fb:
Check the size of screen before memset_io()
git bisect good 6ba592fa014f21f35a8ee8da4ca7b95a018f13e8
# good: [b5a8466d37d30cfcc8015789f4a3f0c44b6c7bc6] Merge tag
'for-5.20/fbdev-1' of
git://git.kernel.org/pub/scm/linux/kernel/git/deller/linux-fbdev
git bisect good b5a8466d37d30cfcc8015789f4a3f0c44b6c7bc6
# bad: [97d3b2676fc6bc4865eb825037f4492f0fb804eb] ocfs2: remove some
useless functions
git bisect bad 97d3b2676fc6bc4865eb825037f4492f0fb804eb
# bad: [591c32bddbe20ba0e172d9def3c7f22b9c926ad9] kernel/hung_task: fix
address space of proc_dohung_task_timeout_secs
git bisect bad 591c32bddbe20ba0e172d9def3c7f22b9c926ad9
# bad: [b09a7a036d2035b14636cd4c4c69518d73770f65] squashfs: support
reading fragments in readahead call
git bisect bad b09a7a036d2035b14636cd4c4c69518d73770f65
mtodorov@domac:~/linux/kernel/linux_stable$
The git bisect stopped at the squashfs commit
b09a7a036d2035b14636cd4c4c69518d73770f65, so I included Phillip in Cc:,
according to the Code of Conduct.
Which identified the "squashfs: support reading fragments in readahead call"
patch.
There is a race-condition introduced in that patch, which involves cache
releasing and reuse.
The following diff will fix that race-condition. It would be great if
someone could test and verify before sending it out as a patch.
Thanks
Phillip
Hi Phillip,
There is a logical bug in commit 8fc78b6fe24c36b151ac98d7546591ed92083d4f
which is parent commit of commit b09a7a036d2035b14636cd4c4c69518d73770f65.
In function squashfs_readahead(...),
file_end is initialized with i_size_read(inode) >> msblk->block_log,
which means the last block index of the file.
But later in the logic to check if the page is last one or not the
code is
if (pages[nr_pages - 1]->index == file_end && bytes) {
...
}
, use file_end as the last page index of file but actually is the last
block index, so for the common setup of page and block size, the first
comparison is true only when pages[nr_pages - 1]->index is 0.
Otherwise, the trailing bytes of last page won't be zeroed.
Maybe it's the real cause of the snap bug in some way.
Hi Jintao,
Thanks for pointing out and sorry for missing this. Does the following
diff improve the issue?
diff --git a/fs/squashfs/file.c b/fs/squashfs/file.c
index e56510964b229..7759bd70dfbf2 100644
--- a/fs/squashfs/file.c
+++ b/fs/squashfs/file.c
@@ -600,7 +600,7 @@ static void squashfs_readahead(struct
readahead_control *ractl)
/* Last page (if present) may have trailing
bytes not filled */
bytes = res % PAGE_SIZE;
- if (pages[nr_pages - 1]->index == file_end && bytes)
+ if ((pages[nr_pages - 1]->index >> shift) ==
file_end && bytes)
memzero_page(pages[nr_pages - 1], bytes,
PAGE_SIZE - bytes);
readahead only handles the case that the first page and the last page
have the same block index:
index = pages[0]->index >> shift;
if ((pages[nr_pages - 1]->index >> shift) != index)
goto skip_pages;
The diff above makes a difference to SQUASHFS_INVALID_BLK case, which
will not be handled by squashfs_readahead_fragment() if
index==file_end.
With the above diff, it will now be memzero_page().
Hi Phillip,
Does the SQUASHFS_INVALID_BLK case handling with index==file_end
sounds reasonable to you?
Thanks.
Hi Hsin-Yi,
The patches were successfully applied (both yours and Phillip's) on v6.0.2 source
and the build is already in progress.
However, due to the stochastic nature of the bug, it may take time before saying
anything for sure.
I will test at around 6 PM UTC+02 on the machine with the same set of windows and
tabs in Firefox that triggered the bug so we'll know then if that case at least is
fixed.
Thanks.
--
Mirsad Goran Todorovac
Sistem inženjer
Grafički fakultet | Akademija likovnih umjetnosti
Sveučilište u Zagrebu
--
System engineer
Faculty of Graphic Arts | Academy of Fine Arts
University of Zagreb, Republic of Croatia