[PATCH] xfs: prevent spurious "head behind tail" warnings

[Mark Tinguely]

On 11/19/13 17:24, Eric Sandeen wrote:
> On 11/19/13, 5:08 PM, Mark Tinguely wrote:
>> On 11/19/13 16:37, Dave Chinner wrote:
>>> From: Dave Chinner<dchinner at redhat.com>
>>>
>>> When xlog_space_left() cracks the grant head and the log tail, it
>>> does so without locking to synchronise the sampling of the
>>> variables. It samples the grant head first, so if there is a delay
>>> before it smaples the log tail, there is a window where the log tail
>>> could have moved onwards and be moved past the sampled value of the
>>> grant head. This then leads to the "xlog_space_left: head behind
>>> tail" warning message.
>>>
>>> To avoid spurious output in this situation, swap the order in which
>>> the variables are cracked. This means that the head may grant head
>>> may move if there is a delay, but the log tail will be stable, hence
>>> ensure the tail does not jump the head accidentally.
>>>
>>> While this avoids the spurious head behind tail problem, it
>>> introduces the opposite problem - the head can move more than a full
>>> cycle past the tail. The code already handles this case by
>>> indicating that the log is full (i.e. zero space available) but
>>> that's still (generally) a spurious situation.
>>>
>>> Hence, if we detect that the head is more than a cycle ahead of the
>>> tail or the head is behind the tail, start the calculation again by
>>> resampling the variables and trying again. If we get too many
>>> resamples, then throw a warning and return a full or empty log
>>> appropriately.
>>>
>>> Signed-off-by: Dave Chinner<dchinner at redhat.com>
>>> ---
>>
>> I am still getting the debug message:
>>
>>    xlog_verify_grant_tail: space>  BBTOB(tail_blocks)
>>
>> This is a real over grant. It has been a while since I did all the tests, but basically the only way to stop it is to have a lock between checking for xlog_space_left() and actually reserving the space.
>>
>> I am not a fan of another band-aid on a problem that is caused because we are granting space without locks.
>
> Mark, can you remind us of your testcase that produces this?
> (sorry, I guess I should search for that old thread...)
>
> Thanks,
> -Eric
>
>> --Mark.

xfstest 273 hits it 100% of the time for me, as does 32+ process 
fsstress, pretty much any high log usage test.

I know Brian hit this with xfstest 273 when he was testing for commit 
9a3a5dab.

Using xfstest 273, I was seeing ten of thousand of bytes of over commit. 
 From what I recall, I tried a separate lock for the write/reserve grant 
heads, put locks to make sure the verifier was not getting stale 
information, ordered the write/reserve ungrants relative to the grants, 
put in cache smp_mb() call. Some attempts were more successful than 
others, but the only way I could prevent the overgrant completely was to 
put back the global lock between the checking for space and the granting 
of space.

--Mark.