[PATCH 44/49] xfs: Reduce allocations during CIL insertion

[Mark Tinguely]

On 07/29/13 19:30, Dave Chinner wrote:
> On Mon, Jul 29, 2013 at 09:15:16AM -0500, Mark Tinguely wrote:
>> On 07/27/13 20:12, Dave Chinner wrote:
>>> On Sat, Jul 27, 2013 at 01:32:48PM -0500, Mark Tinguely wrote:
>>>> On 07/26/13 20:58, Dave Chinner wrote:
>>>>> On Fri, Jul 26, 2013 at 04:06:37PM -0500, Mark Tinguely wrote:
>>>>>>
>>>>>> I can reproduce a problem in patch 44 too. It takes my copy test 20
>>>>>> minutes to deplete the log space with patch 44, Same test with patch
>>>>>> 43 has been running a day and a half. I do not think that patch 44
>>>>>> is 100 times faster than patch 43 but I will let patch 43 spin all
>>>>>> weekend on a couple machines to verify that patch 43 does not hang.
>>>>>
>>>>> Details, please. What's your "copy test"?
> ....
>>>> Micheal found the problem using a simple copy, so I am using copy-like test.
>>
>> BTW, the long term run of the copy.pl from bug 922 with patch 43 results:
>>   tail            0x601000055d7
>>   grant/reserve   0x60100abb200
>>   ctx t_unit_res  0x834
>>
>> My log math may be off, tail/reserve diff is 1024, but the CTX holds
>> more than that (2100 bytes).
>>
>> Looking at patch 44, it is the first time we use the calculation for
>> the number of bytes in patch 43. So I am looking at where the new
>> calculation in iop_size differs from the previous len calculation in
>> xlog_cil_prepare_log_vecs(). So far, I am that inode entry is 140
>> bytes larger with the new calculation (former len 164 vrs new nbytes
>> 304 type 123b - non-crc filesystem).
>
> Which size calculation is wrong? t~he one used to size the buffer
> being allocated - which is intentionally oversized for the inode
> forks - or the actual size formatted into the buffer, which was
> unchanged?
>
> I mean, 164 bytes is an inode core (96 bytes) plus a inode log
> format structure. The increase of 140 bytes indicates that we are
> logging roughly the entire 256 byte inode - i.e. both forks.
>
> But are you looking at the size returned by iop_size or iop_format?
> iop_size will be new, iop_format is unchanged by this patchset.
> indeed, what iop_format returns as vectors is *unchanged* by this
> patchset, so I think that you are chasing down the wrong path here.
>

164 is the variable len. It is the sum of the xfs_log_iovec.i_len after
format.

304 is the result of the iop_size(). starting in patch 44 that is used
insteead of len. When we go do the format, we only use 164.

It is the XFS_DINODE_FMT_EXTENTS inode type that generates the extra
140 bytes in xfs_inode_item_size().

In patch 43, use nbytes instead of len like is done in patch 44 and it
hangs at the same time and manner than patch 44.

>> I will see if I can substitute the nbytes for len (without
>> triggering the ptr being exceeded assert) to see that will cause the
>> hang.
>
> I know what the problem is will be, and that won't fix it. Hint: the
> buffer is over sized.
>
> Cheers,
>
> Dave.

--Mark.