On 1/13/17 9:44 PM, Brian Foster wrote:
> On Fri, Jan 13, 2017 at 08:25:47PM -0600, Eric Sandeen wrote:
>> On 1/13/17 7:35 AM, Brian Foster wrote:
>>>> + fprintf(stderr,
>>>> +_("%s: device %s AG count is insane, but could be %u. Limiting reads to AG 0.\n"),
>>>> + progname, fsdevice, dblocks / sbp->sb_agblocks);
>>>> + } else {
>>>> + fprintf(stderr,
>>>> +_("%s: device %s AG count is insane. Limiting reads to AG 0.\n"),
>>>> + progname, fsdevice);
>>>> + }
>>> For reasons like the above, I think xfs_db shouldn't be in the business
>>> of repair like validation (xfs_check notwithstanding). That said,
>>> dropping into a fixed single AG mode seems less risky than trying to
>>> surmise a valid geometry. I'd get rid of the "this might be your
>>> agcount" messaging entirely though and just replace it with something
>>> that explicitly states the filesystem is corrupted, the runtime geometry
>>> is invalid and that the user should probably run xfs_repair before doing
>>> anything.
>>
>> So keep in mind that xfs_db is for people with super xfs powers. (*)
>>
>> I wouldn't suggest repair, I'd start with 1 ag to avoid the OOM, state
>> that clearly, and punt the problem to the admin with no other specific
>> suggestions.
>>
>>> I still like the idea of the single AG mode thing as a command line flag
>>> rather than default behavior because it requires user acknowledgement,
>>> but this is a debug tool after all, so I'll defer to Eric on that. I do
>>> think that if we create this kind of invalid runtime mode, this should
>>> be split into two patches. First, a bugfix patch for the core OOM
>>> problem (i.e., detect a wacky superblock and exit). Second, replace the
>>> exit with the single AG runtime mode thing.
>>
>> Well, the problem with a flag, I think, is that you might have already
>> unwittingly OOMed your box to find out that you need it.
>> Rebooting to try again with a flag sucks.
>>
>
> I don't see how that is relevant. I'm not suggesting a
> --please-don't-oom-in-case-of-corruption flag. :) As mentioned
> previously, I think the bug fix here is a simple patch to detect the
> bogus superblock and exit gracefully rather than go off the rails and
> end up OOM killed.
sorry, misunderstood the
"idea of the single AG mode thing as a command line flag" idea, I guess.
> From there the OOM is irrelevant and we can optionally enhance xfs_db to
> try and allow it to run in such situations. To be honest, I'm perfectly
> happy for xfs_db to exit gracefully in this situation and to leave it at
> that. I think the majority of cases where this problem occurs, the next
> logical step is to run xfs_repair. I suggested the flag approach more
> because I think it's more appropriate to do things like fabricate fs
> geometry behind a flag rather than by default. The larger point is that
> if we want this kind of enhancement, it should probably be driven more
> by a use case than an unfortunate (and probably rare) bug. I don't see
> why we need to complicate the bug fix with the fancy enhancement.
*nod* ok, I had understood the flag idea backwards-ly I guess.
I do think that mere mortal invocations via xfs_admin need to be
handled in this "ignore agcount" case, though...
-Eric
> Brian
>
>> (*) unless you are invoking it via xfs_admin.sh, dammit. We sure wouldn't
>> want xfs_admin to exit happily, having updated only one AG. Dammit!
>>
>> Perhaps it should set exitcode, and then xfs_admin could do something
>> like:
>>
>> xfs_db -c quit $DEV
>>
>> first, and check that db is able to initialize sanely before using it again
>> to perform normal admin functions.
>>
>> -Eric
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
>> the body of a message to majordomo@xxxxxxxxxxxxxxx
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
