diff options
| author | Dmitry Vyukov <dvyukov@google.com> | 2021-07-06 12:27:19 +0200 |
|---|---|---|
| committer | Dmitry Vyukov <dvyukov@google.com> | 2021-07-06 13:26:46 +0200 |
| commit | 22ff12bdbb3509fbf00cd979af4629b50aaff7a3 (patch) | |
| tree | 8bca828bde1c003c8c1975f16e71716de6a959b0 /pkg | |
| parent | 9cac178ab43e4a07f6ba76c029a02417fe5f55c8 (diff) | |
dashboard/config/linux: increase printk buffer size
There is a number of one-off "BUG: stack guard page was hit" bugs:
https://syzkaller.appspot.com/bug?extid=b9419ddbdb57ce0e7f36
https://syzkaller.appspot.com/bug?extid=1ea34900b9a6fb8526c4
https://syzkaller.appspot.com/bug?extid=cd2009ad04934b665765
https://syzkaller.appspot.com/bug?extid=4e1ccdc40f48e600d960
Most likely these have the same root cause (recursion via call_netdevice_notifiers),
and we should attribute them to one of top frames as we do for stalls.
But we can't do this because in all these cases the stack is truncated and ends with:
Lost 408 message(s)!
Lost 394 message(s)!
Lost 519 message(s)!
These messages come from kernel/printk/printk_safe.c and happen when
a "safe" buffer overflows. Increasing CONFIG_PRINTK_SAFE_LOG_BUF_SHIFT
should help to get complete parsable stack traces.
The default value is 13 (8K).
The largest number of lost lines I found is 519 and the longest
line in the stack trace is 67 bytes. So that's 67*519+8K = 42965.
Increase the config to 16 (64K).
There are 2 such buffers per CPU (safe and nmi), so this will
increase memory consumption to 128K per CPU. Should be fine.
Diffstat (limited to 'pkg')
0 files changed, 0 insertions, 0 deletions