Re: [PATCH] jbd/2[stable only]: Use WRITE_SYNC_PLUG in journal_commit_transaction.

[Tao Ma <tm@tao.ma>]

Hi Vivek and Jeff,
On 07/15/2011 05:38 AM, Jan Kara wrote:
> On Thu 14-07-11 16:08:24, Jeff Moyer wrote:
>> Jan Kara <jack@suse.cz> writes:
>>
>>> On Thu 14-07-11 12:30:32, Jeff Moyer wrote:
>>>> Tao Ma <tm@tao.ma> writes:
>>>>>> - WRITE_SYNC_PLUG will plug the queue and expects explicity unplug. Who
>>>>>>   is doing unplug in this case?
>>>>> See the comments I removed, "we rely on sync_buffer() doing the unplug
>>>>> for us". I removed them cause we all use pluged write now.
>>>>
>>>> Your logic is upside-down.  The code currently only uses the _PLUG
>>>> variant when t_synchronous_commit is set, meaning somebody *will* call
>>>> sync_buffer.  Simply setting WRITE_SYNC_PLUG doens't mean the upper
>>>> layer is going to issue the unplug.  Of course, I'm not 100% sure of the
>>>> journaling process, so it may very well be that there always is an
>>>> unplug.  Can Jan or someone comment on that?  Anyway, you could test
>>>> this theory by seeing if your kernel generates any timer unplugs in the
>>>> blktrace output.
>>>   So I'm not expert in plugging code but from what I understand when we do
>>> wait_on_buffer() (which calls io_schedule()) which will do
>>> blk_flush_plug()), the queue will get unplugged and IO starts. And we wait
>>> for all buffers we submit so we are guaranteed wait_on_buffer() will be
>>> called...
>>
>> Sorry, I should have been more specific.  As Vivek mentioned, we're
>> talking about older kernels (pre the blk plugging series).  So, the
>> question is, if journal_commit_transaction is called with
>> t_synchronous_commit not set, will the underlying device ever be
>> unplugged by the journal code?  My guess is there's no explicit unplug,
>> so it's not correct to replace a WRITE_SYNC with a WRITE_SYNC_PLUG.
>   There are no explicit unplugs in journalling code. But checking the code
> in 2.6.37, I still see wait_on_buffer() calls sync_bh() which calls
> blk_run_address_space() which ends up calling bdi->unplug_io_fn() so I
> would say unplug is called anyway.
yeah, jbd2 works like what Jan described.
And what's more, if you looked at the commit I mentioned(749ef9f8423),
this commit just changed WRITE to WRITE_SYNC, so in the old times(before
that commit is merged), WRITE has been used in jbd/2 for many years. It
also doesn't do a explicit unplug, let the scheduler do the merge and
reply on sync_bh to unplug the device. So change WRITE to
WRITE_SYNC_PLUG is welcomed, but change WRITE to WRITE_SYNC is broken
since it splits the sequential write to several i/o requests.

Thanks
Tao