Re: [PATCH] fstests: report: always save the dmesg as system-err if KEEP_DMESG is set

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

On 2022/12/20 02:56, Theodore Ts'o wrote:
> On Mon, Dec 19, 2022 at 09:39:33AM -0800, Darrick J. Wong wrote:
>> On Fri, Dec 16, 2022 at 02:51:21PM +0800, Qu Wenruo wrote:
>>> When KEEP_DMESG is set to "yes", we will always save the dmesg of any
>>> test case (no matter if it passed or not) into "$seqnum.dmesg".
>>>
>>> But this KEEP_DMESG behavior doesn't affect xunit report.
>>>
>>> This patch will make xunit report to follow KEEP_DMESG setting.
> 
> This may be dangerous; if the XML file is too large, the XML parser
> may end up rejecting the whole XML file because otherwlse a too-large
> XML file can trigger a denial of service attack[1].  (This is why I
> implemented "xunit-quiet".)

I guess your concern is correct.

Although during my auto run, single dmesg are not that large, the 
largest three are not MiB level yet:

108K    btrfs/187.dmesg
80K     btrfs/072.dmesg
64K     generic/311.dmesg

But the total dmesg sizes already go MiB levels, 2.6MiB.

> 
> [1] https://gitlab.com/gitlab-org/gitlab/-/issues/25357
> 
> So if you are running a large number of tests (e.g., "-g auto") it
> might very well that adding dmesg for all tests might very well end up
> bloating the XML file to the point where it will be unmangeable.  For
> example, this is the size for my syslog file after running "-g auto"
> on the "xfs/quota" config:
> 
> -rw-r----- 1 tytso primarygroup 10316684 Aug 25 10:35 ae/syslog
> 
> The syslog file for all of the xfs configs are 9-10 megabytes each.
> If I combined the 12 xfs configs that we run into a single xunit JML
> file with the dmesg output, this would be *guaranteed* to blow out
> most XML parsers.
> 
> Personally, I find that a better solution is to use the syslog daemon
> to save the dmesg output for all of the tests into a single file.  I
> prefer this for three reasons:
> 
>    * The single file is more compressibls compared to having it broken
>      out into separate $seqnum.dmesg files.
>    * By keeping dmesg and other test artifacts separate from the xml
>      file I can archive the xml file for a much larger period of time,
>      (perhaps indefinitely) while allowing the much more volunumous
>      test artifacts to be archived for a shorter time (say, 3-6 months).
>    * When there are test isolation issues, it's not uncommon for a
>      previous test to fail with some kind of global or cgroup-specific
>      OOM-kill, or when I'm testing on bare metal with real hardware
>      where hardware failures is a Thing, being able to look for unusual
>      kernel messages before the start of a particular test can often be
>      quite revealing.

On the other hand, separating dmesg from its test cases means extra 
parsing to bind the log section to certain test runs.

In fact, I'm not even sure if possible to do that.
E.g. on the same host with different kernel/fstests configs.

One reason I want to keep the dmesg bound to each test case is, we had 
cases in the past that some btrfs specific error messages are not (and 
should not) caught by fstests, but can still be some clue of bugs.

Thus if we have a proper paired dmesg with kernel/fstests configs and 
test case, we can determine how affected the bug is.

Any better idea to relate them? Or is this really too niche?

Thanks,
Qu
>      
> Cheers,
> 
> 						- Ted