[ASSERT failure] transaction reservations changes bad?

[Jeff Liu]

On 03/12/2013 07:05 PM, Jeff Liu wrote:
> On 03/12/2013 06:31 PM, Dave Chinner wrote:
>> On Tue, Mar 12, 2013 at 04:08:20PM +0800, Jeff Liu wrote:
>>> Hi Dave,
>>>
>>> On 03/12/2013 02:25 PM, Dave Chinner wrote:
>>>> On Tue, Mar 12, 2013 at 05:20:02PM +1100, Dave Chinner wrote:
>>>>> Folks,
>>>>>
>>>>> I just got this ASSERT failure running xfstests on a 3.1.8 xfsprogs
>>>>> and a 3.9-rc1 kernel running test 297:
>>>>
>>>> FYI, it's 100% reproducable here with:
>>>>
>>>> # sudo MKFS_OPTIONS="-b size=512" ./check 297
>>>>
>>>> (reproduced on multiple VMs now with the same command line)
>> ....
>>>>> This implies that the permanent transaction reservation ended up
>>>>> larger than the log itself:
>>>>>
>>>>> $ sudo xfs_info /mnt/scratch/
>>>>> [sudo] password for dave: 
>>>>> meta-data=/dev/vdb               isize=256    agcount=16, agsize=1441792 blks
>>>>>          =                       sectsz=512   attr=2
>>>>> data     =                       bsize=512    blocks=23068672, imaxpct=25
>>>>>          =                       sunit=512    swidth=6144 blks
>>>>> naming   =version 2              bsize=4096   ascii-ci=0
>>>>> log      =internal               bsize=512    blocks=2560, version=2
>>>>>          =                       sectsz=512   sunit=512 blks, lazy-count=1
>>>>> realtime =none                   extsz=4096   blocks=0, rtextents=0
>>>>>
>>>>> Can someone please check that the before/after mkdir transaction
>>>>> reservation sizes are unchanged for such a configuration?
>>> I just did a quick verification. 
>>>
>>> # mkfs.xfs -V
>>> mkfs.xfs version 3.1.8
>>>
>>> # uname -a
>>> Linux koala 3.9.0-rc1 #80 SMP Tue Mar 12 15:06:39 CST 2013 x86_64 x86_64 x86_64 GNU/Linux
>>>
>>> # mkfs.xfs -f -b size=512 /dev/sda6
>>> meta-data=/dev/sda6              isize=256    agcount=4, agsize=5242880 blks
>>>          =                       sectsz=512   attr=2, projid32bit=0
>>> data     =                       bsize=512    blocks=20971520, imaxpct=25
>>>          =                       sunit=0      swidth=0 blks
>>> naming   =version 2              bsize=4096   ascii-ci=0
>>> log      =internal log           bsize=512    blocks=20480, version=2
>>>          =                       sectsz=512   sunit=0 blks, lazy-count=1
>>> realtime =none                   extsz=4096   blocks=0, rtextents=0
>>
>> That's a different mkfs.xfs config to what test 297 is using.
>> Different log size, different AG count, no log stripe unit, etc.
>> 297 is using:
>>
>> scratch_mkfs_xfs -d agcount=16,su=256k,sw=12 -l su=256k,size=2560b <dev>
>>
>> And when I add the extra MKFS_OPTIONS in it actually is:
>>
>> # mkfs.xfs -b size=512 -d agcount=16,su=256k,sw=12 -l su=256k,size=2560b <dev>
>>
>>> The reservation size does not changed, both are 70072 bytes:
>>>
>>> [  230.905092] xfs_calc_mkdir_reservation: res=70072 bytes.
>>
>> And it's not just the calculation that I'm worried about here - it's
>> the actual reservation that ends up in the ticket that matters as
>> that is fed into the code that has triggered the assert. The value
>> in the ticket takes into account log stripe units and other
>> roundings, so it's typically much larger than just the reservation
>> calculation itself...
>>
>>> However, I can always reproducing this issue with
>>> '"MKFS_OPTIONS=-b size=512" ./check 297' as well.
>>
>> Can you check that it also fails on kernels prior to the reservation
>> changes? That will rule out it being a recent regression...
> Sure, verified against a new built 3.8.0 kernel, so this should be a
> regression issue.
> $ uname -a
> Linux koala 3.8.0 #81 SMP Tue Mar 12 18:03:44 CST 2013 x86_64 x86_64
> x86_64 GNU/Linux
> 
> commit 19f949f52599ba7c3f67a5897ac6be14bfcb1200
> Author: Linus Torvalds <torvalds at linux-foundation.org>
> Date:   Mon Feb 18 15:58:34 2013 -0800
More info, 3.7.0 is the oldest kernel on my environment, I ran into the
same problem.

Thanks,
-Jeff