diff mbox

[05/10] block: remove per-queue plugging

Message ID 4DA44C86.3090305@fusionio.com (mailing list archive)
State Not Applicable, archived
Headers show

Commit Message

Jens Axboe April 12, 2011, 12:58 p.m. UTC
On 2011-04-12 14:41, Dave Chinner wrote:
> On Tue, Apr 12, 2011 at 02:28:31PM +0200, Jens Axboe wrote:
>> On 2011-04-12 14:22, Dave Chinner wrote:
>>> On Tue, Apr 12, 2011 at 10:36:30AM +0200, Jens Axboe wrote:
>>>> On 2011-04-12 03:12, hch@infradead.org wrote:
>>>>> On Mon, Apr 11, 2011 at 02:48:45PM +0200, Jens Axboe wrote:
>>>>>> Great, once you do that and XFS kills the blk_flush_plug() calls too,
>>>>>> then we can remove that export and make it internal only.
>>>>>
>>>>> Linus pulled the tree, so they are gone now.  Btw, there's still some
>>>>> bits in the area that confuse me:
>>>>
>>>> Great!
>>>>
>>>>>  - what's the point of the queue_sync_plugs?  It has a lot of comment
>>>>>    that seem to pre-data the onstack plugging, but except for that
>>>>>    it's trivial wrapper around blk_flush_plug, with an argument
>>>>>    that is not used.
>>>>
>>>> There's really no point to it anymore. It's existance was due to the
>>>> older revision that had to track write requests for serializaing around
>>>> a barrier. I'll kill it, since we don't do that anymore.
>>>>
>>>>>  - is there a good reason for the existance of __blk_flush_plug?  You'd
>>>>>    get one additional instruction in the inlined version of
>>>>>    blk_flush_plug when opencoding, but avoid the need for chained
>>>>>    function calls.
>>>>>  - Why is having a plug in blk_flush_plug marked unlikely?  Note that
>>>>>    unlikely is the static branch prediction hint to mark the case
>>>>>    extremly unlikely and is even used for hot/cold partitioning.  But
>>>>>    when we call it we usually check beforehand if we actually have
>>>>>    plugs, so it's actually likely to happen.
>>>>
>>>> The existance and out-of-line is for the scheduler() hook. It should be
>>>> an unlikely event to schedule with a plug held, normally the plug should
>>>> have been explicitly unplugged before that happens.
>>>
>>> Though if it does, haven't you just added a significant amount of
>>> depth to the worst case stack usage? I'm seeing this sort of thing
>>> from io_schedule():
>>>
>>>         Depth    Size   Location    (40 entries)
>>>         -----    ----   --------
>>>   0)     4256      16   mempool_alloc_slab+0x15/0x20
>>>   1)     4240     144   mempool_alloc+0x63/0x160
>>>   2)     4096      16   scsi_sg_alloc+0x4c/0x60
>>>   3)     4080     112   __sg_alloc_table+0x66/0x140
>>>   4)     3968      32   scsi_init_sgtable+0x33/0x90
>>>   5)     3936      48   scsi_init_io+0x31/0xc0
>>>   6)     3888      32   scsi_setup_fs_cmnd+0x79/0xe0
>>>   7)     3856     112   sd_prep_fn+0x150/0xa90
>>>   8)     3744      48   blk_peek_request+0x6a/0x1f0
>>>   9)     3696      96   scsi_request_fn+0x60/0x510
>>>  10)     3600      32   __blk_run_queue+0x57/0x100
>>>  11)     3568      80   flush_plug_list+0x133/0x1d0
>>>  12)     3488      32   __blk_flush_plug+0x24/0x50
>>>  13)     3456      32   io_schedule+0x79/0x80
>>>
>>> (This is from a page fault on ext3 that is doing page cache
>>> readahead and blocking on a locked buffer.)
>>>
>>> I've seen traces where mempool_alloc_slab enters direct reclaim
>>> which adds another 1.5k of stack usage to this path. So I'm
>>> extremely concerned that you've just reduced the stack available to
>>> every thread by at least 2.5k of space...
>>
>> Yeah, that does not look great. If this turns out to be problematic, we
>> can turn the queue runs from the unlikely case into out-of-line from
>> kblockd.
>>
>> But this really isn't that new, you could enter the IO dispatch path
>> when doing IO already (when submitting it). So we better be able to
>> handle that.
> 
> The problem I see is that IO is submitted when there's plenty of
> stack available whould have previously been fine. However now it
> hits the plug, and then later on after the thread consumes a lot
> more stack it, say, waits for a completion. We then schedule, it
> unplugs the queue and we add the IO stack to a place where there
> isn't much space available.
>
> So effectively we are moving the places where stack is consumed
> about, and it's complete unpredictable where that stack is going to
> land now.

Isn't that example fairly contrived? If we ended up doing the IO
dispatch before, then the only difference now is the stack usage of
schedule() itself. Apart from that, as far as I can tell, there should
not be much difference.

 
>> If it's a problem from the schedule()/io_schedule() path, then
>> lets ensure that those are truly unlikely events so we can punt
>> them to kblockd.
> 
> Rather than wait for an explosion to be reported before doing this,
> why not just punt unplugs to kblockd unconditionally?

Supposedly it's faster to do it inline rather than punt the dispatch.
But that may actually not be true, if you have multiple plugs going (and
thus multiple contenders for the queue lock on dispatch). So lets play
it safe and punt to kblockd, we can always revisit this later.

Comments

Dave Chinner April 12, 2011, 1:31 p.m. UTC | #1
On Tue, Apr 12, 2011 at 02:58:46PM +0200, Jens Axboe wrote:
> On 2011-04-12 14:41, Dave Chinner wrote:
> > On Tue, Apr 12, 2011 at 02:28:31PM +0200, Jens Axboe wrote:
> >> On 2011-04-12 14:22, Dave Chinner wrote:
> >>> On Tue, Apr 12, 2011 at 10:36:30AM +0200, Jens Axboe wrote:
> >>>> On 2011-04-12 03:12, hch@infradead.org wrote:
> >>>>> On Mon, Apr 11, 2011 at 02:48:45PM +0200, Jens Axboe wrote:
> >>>>>> Great, once you do that and XFS kills the blk_flush_plug() calls too,
> >>>>>> then we can remove that export and make it internal only.
> >>>>>
> >>>>> Linus pulled the tree, so they are gone now.  Btw, there's still some
> >>>>> bits in the area that confuse me:
> >>>>
> >>>> Great!
> >>>>
> >>>>>  - what's the point of the queue_sync_plugs?  It has a lot of comment
> >>>>>    that seem to pre-data the onstack plugging, but except for that
> >>>>>    it's trivial wrapper around blk_flush_plug, with an argument
> >>>>>    that is not used.
> >>>>
> >>>> There's really no point to it anymore. It's existance was due to the
> >>>> older revision that had to track write requests for serializaing around
> >>>> a barrier. I'll kill it, since we don't do that anymore.
> >>>>
> >>>>>  - is there a good reason for the existance of __blk_flush_plug?  You'd
> >>>>>    get one additional instruction in the inlined version of
> >>>>>    blk_flush_plug when opencoding, but avoid the need for chained
> >>>>>    function calls.
> >>>>>  - Why is having a plug in blk_flush_plug marked unlikely?  Note that
> >>>>>    unlikely is the static branch prediction hint to mark the case
> >>>>>    extremly unlikely and is even used for hot/cold partitioning.  But
> >>>>>    when we call it we usually check beforehand if we actually have
> >>>>>    plugs, so it's actually likely to happen.
> >>>>
> >>>> The existance and out-of-line is for the scheduler() hook. It should be
> >>>> an unlikely event to schedule with a plug held, normally the plug should
> >>>> have been explicitly unplugged before that happens.
> >>>
> >>> Though if it does, haven't you just added a significant amount of
> >>> depth to the worst case stack usage? I'm seeing this sort of thing
> >>> from io_schedule():
> >>>
> >>>         Depth    Size   Location    (40 entries)
> >>>         -----    ----   --------
> >>>   0)     4256      16   mempool_alloc_slab+0x15/0x20
> >>>   1)     4240     144   mempool_alloc+0x63/0x160
> >>>   2)     4096      16   scsi_sg_alloc+0x4c/0x60
> >>>   3)     4080     112   __sg_alloc_table+0x66/0x140
> >>>   4)     3968      32   scsi_init_sgtable+0x33/0x90
> >>>   5)     3936      48   scsi_init_io+0x31/0xc0
> >>>   6)     3888      32   scsi_setup_fs_cmnd+0x79/0xe0
> >>>   7)     3856     112   sd_prep_fn+0x150/0xa90
> >>>   8)     3744      48   blk_peek_request+0x6a/0x1f0
> >>>   9)     3696      96   scsi_request_fn+0x60/0x510
> >>>  10)     3600      32   __blk_run_queue+0x57/0x100
> >>>  11)     3568      80   flush_plug_list+0x133/0x1d0
> >>>  12)     3488      32   __blk_flush_plug+0x24/0x50
> >>>  13)     3456      32   io_schedule+0x79/0x80
> >>>
> >>> (This is from a page fault on ext3 that is doing page cache
> >>> readahead and blocking on a locked buffer.)
> >>>
> >>> I've seen traces where mempool_alloc_slab enters direct reclaim
> >>> which adds another 1.5k of stack usage to this path. So I'm
> >>> extremely concerned that you've just reduced the stack available to
> >>> every thread by at least 2.5k of space...
> >>
> >> Yeah, that does not look great. If this turns out to be problematic, we
> >> can turn the queue runs from the unlikely case into out-of-line from
> >> kblockd.
> >>
> >> But this really isn't that new, you could enter the IO dispatch path
> >> when doing IO already (when submitting it). So we better be able to
> >> handle that.
> > 
> > The problem I see is that IO is submitted when there's plenty of
> > stack available whould have previously been fine. However now it
> > hits the plug, and then later on after the thread consumes a lot
> > more stack it, say, waits for a completion. We then schedule, it
> > unplugs the queue and we add the IO stack to a place where there
> > isn't much space available.
> >
> > So effectively we are moving the places where stack is consumed
> > about, and it's complete unpredictable where that stack is going to
> > land now.
> 
> Isn't that example fairly contrived?

I don't think so. e.g. in the XFS allocation path we do btree block
readahead, then go do the real work. The real work can end up with a
deeper stack before blocking on locks or completions unrelated to
the readahead, leading to schedule() being called and an unplug
being issued at that point.  You might think it contrived, but if
you can't provide a guarantee that it can't happen then I have to
assume it will happen.

My concern is that we're already under stack space stress in the
writeback path, so anything that has the potential to increase it
significantly is a major worry from my point of view...

> If we ended up doing the IO
> dispatch before, then the only difference now is the stack usage of
> schedule() itself. Apart from that, as far as I can tell, there should
> not be much difference.

There's a difference between IO submission and IO dispatch. IO
submission is submit_bio thru to the plug; IO dispatch is from the
plug down to the disk. If they happen at the same place, there's no
problem. If IO dispatch is moved to schedule() via a plug....

> >> If it's a problem from the schedule()/io_schedule() path, then
> >> lets ensure that those are truly unlikely events so we can punt
> >> them to kblockd.
> > 
> > Rather than wait for an explosion to be reported before doing this,
> > why not just punt unplugs to kblockd unconditionally?
> 
> Supposedly it's faster to do it inline rather than punt the dispatch.
> But that may actually not be true, if you have multiple plugs going (and
> thus multiple contenders for the queue lock on dispatch). So lets play
> it safe and punt to kblockd, we can always revisit this later.

It's always best to play it safe when it comes to other peoples
data....

Cheers,

Dave.
Jens Axboe April 12, 2011, 1:45 p.m. UTC | #2
On 2011-04-12 15:31, Dave Chinner wrote:
> On Tue, Apr 12, 2011 at 02:58:46PM +0200, Jens Axboe wrote:
>> On 2011-04-12 14:41, Dave Chinner wrote:
>>> On Tue, Apr 12, 2011 at 02:28:31PM +0200, Jens Axboe wrote:
>>>> On 2011-04-12 14:22, Dave Chinner wrote:
>>>>> On Tue, Apr 12, 2011 at 10:36:30AM +0200, Jens Axboe wrote:
>>>>>> On 2011-04-12 03:12, hch@infradead.org wrote:
>>>>>>> On Mon, Apr 11, 2011 at 02:48:45PM +0200, Jens Axboe wrote:
>>>>>>>> Great, once you do that and XFS kills the blk_flush_plug() calls too,
>>>>>>>> then we can remove that export and make it internal only.
>>>>>>>
>>>>>>> Linus pulled the tree, so they are gone now.  Btw, there's still some
>>>>>>> bits in the area that confuse me:
>>>>>>
>>>>>> Great!
>>>>>>
>>>>>>>  - what's the point of the queue_sync_plugs?  It has a lot of comment
>>>>>>>    that seem to pre-data the onstack plugging, but except for that
>>>>>>>    it's trivial wrapper around blk_flush_plug, with an argument
>>>>>>>    that is not used.
>>>>>>
>>>>>> There's really no point to it anymore. It's existance was due to the
>>>>>> older revision that had to track write requests for serializaing around
>>>>>> a barrier. I'll kill it, since we don't do that anymore.
>>>>>>
>>>>>>>  - is there a good reason for the existance of __blk_flush_plug?  You'd
>>>>>>>    get one additional instruction in the inlined version of
>>>>>>>    blk_flush_plug when opencoding, but avoid the need for chained
>>>>>>>    function calls.
>>>>>>>  - Why is having a plug in blk_flush_plug marked unlikely?  Note that
>>>>>>>    unlikely is the static branch prediction hint to mark the case
>>>>>>>    extremly unlikely and is even used for hot/cold partitioning.  But
>>>>>>>    when we call it we usually check beforehand if we actually have
>>>>>>>    plugs, so it's actually likely to happen.
>>>>>>
>>>>>> The existance and out-of-line is for the scheduler() hook. It should be
>>>>>> an unlikely event to schedule with a plug held, normally the plug should
>>>>>> have been explicitly unplugged before that happens.
>>>>>
>>>>> Though if it does, haven't you just added a significant amount of
>>>>> depth to the worst case stack usage? I'm seeing this sort of thing
>>>>> from io_schedule():
>>>>>
>>>>>         Depth    Size   Location    (40 entries)
>>>>>         -----    ----   --------
>>>>>   0)     4256      16   mempool_alloc_slab+0x15/0x20
>>>>>   1)     4240     144   mempool_alloc+0x63/0x160
>>>>>   2)     4096      16   scsi_sg_alloc+0x4c/0x60
>>>>>   3)     4080     112   __sg_alloc_table+0x66/0x140
>>>>>   4)     3968      32   scsi_init_sgtable+0x33/0x90
>>>>>   5)     3936      48   scsi_init_io+0x31/0xc0
>>>>>   6)     3888      32   scsi_setup_fs_cmnd+0x79/0xe0
>>>>>   7)     3856     112   sd_prep_fn+0x150/0xa90
>>>>>   8)     3744      48   blk_peek_request+0x6a/0x1f0
>>>>>   9)     3696      96   scsi_request_fn+0x60/0x510
>>>>>  10)     3600      32   __blk_run_queue+0x57/0x100
>>>>>  11)     3568      80   flush_plug_list+0x133/0x1d0
>>>>>  12)     3488      32   __blk_flush_plug+0x24/0x50
>>>>>  13)     3456      32   io_schedule+0x79/0x80
>>>>>
>>>>> (This is from a page fault on ext3 that is doing page cache
>>>>> readahead and blocking on a locked buffer.)
>>>>>
>>>>> I've seen traces where mempool_alloc_slab enters direct reclaim
>>>>> which adds another 1.5k of stack usage to this path. So I'm
>>>>> extremely concerned that you've just reduced the stack available to
>>>>> every thread by at least 2.5k of space...
>>>>
>>>> Yeah, that does not look great. If this turns out to be problematic, we
>>>> can turn the queue runs from the unlikely case into out-of-line from
>>>> kblockd.
>>>>
>>>> But this really isn't that new, you could enter the IO dispatch path
>>>> when doing IO already (when submitting it). So we better be able to
>>>> handle that.
>>>
>>> The problem I see is that IO is submitted when there's plenty of
>>> stack available whould have previously been fine. However now it
>>> hits the plug, and then later on after the thread consumes a lot
>>> more stack it, say, waits for a completion. We then schedule, it
>>> unplugs the queue and we add the IO stack to a place where there
>>> isn't much space available.
>>>
>>> So effectively we are moving the places where stack is consumed
>>> about, and it's complete unpredictable where that stack is going to
>>> land now.
>>
>> Isn't that example fairly contrived?
> 
> I don't think so. e.g. in the XFS allocation path we do btree block
> readahead, then go do the real work. The real work can end up with a
> deeper stack before blocking on locks or completions unrelated to
> the readahead, leading to schedule() being called and an unplug
> being issued at that point.  You might think it contrived, but if
> you can't provide a guarantee that it can't happen then I have to
> assume it will happen.

If you ended up in lock_page() somewhere along the way, the path would
have been pretty much the same as it is now:

lock_page()
        __lock_page()
                __wait_on_bit_lock()
                        sync_page()
                                aops->sync_page();
                                        block_sync_page()
                                                __blk_run_backing_dev()

and the dispatch follows after that. If your schedules are only due to,
say, blocking on a mutex, then yes it'll be different. But is that
really the case?

I bet that worst case stack usage is exactly the same as before, and
that's the only metric we really care about.

> My concern is that we're already under stack space stress in the
> writeback path, so anything that has the potential to increase it
> significantly is a major worry from my point of view...

I agree on writeback being a worry, and that's why I made the change
(since it makes sense for other reasons, too). I just don't think we are
worse of than before.

>> If we ended up doing the IO
>> dispatch before, then the only difference now is the stack usage of
>> schedule() itself. Apart from that, as far as I can tell, there should
>> not be much difference.
> 
> There's a difference between IO submission and IO dispatch. IO
> submission is submit_bio thru to the plug; IO dispatch is from the
> plug down to the disk. If they happen at the same place, there's no
> problem. If IO dispatch is moved to schedule() via a plug....

The IO submission can easily and non-deterministically turn into an IO
dispatch, so there's no real difference for the submitter. That was the
case before. With the explicit plug now, you _know_ that the IO
submission is only that and doesn't include IO dispatch. Not until you
schedule() or call blk_finish_plug(), both of which are events that you
can control.

>>>> If it's a problem from the schedule()/io_schedule() path, then
>>>> lets ensure that those are truly unlikely events so we can punt
>>>> them to kblockd.
>>>
>>> Rather than wait for an explosion to be reported before doing this,
>>> why not just punt unplugs to kblockd unconditionally?
>>
>> Supposedly it's faster to do it inline rather than punt the dispatch.
>> But that may actually not be true, if you have multiple plugs going (and
>> thus multiple contenders for the queue lock on dispatch). So lets play
>> it safe and punt to kblockd, we can always revisit this later.
> 
> It's always best to play it safe when it comes to other peoples
> data....

Certainly, but so far I see no real evidence that this is in fact any
safer.
Dave Chinner April 12, 2011, 2:34 p.m. UTC | #3
On Tue, Apr 12, 2011 at 03:45:52PM +0200, Jens Axboe wrote:
> On 2011-04-12 15:31, Dave Chinner wrote:
> > On Tue, Apr 12, 2011 at 02:58:46PM +0200, Jens Axboe wrote:
> >> On 2011-04-12 14:41, Dave Chinner wrote:
> >> Isn't that example fairly contrived?
> > 
> > I don't think so. e.g. in the XFS allocation path we do btree block
> > readahead, then go do the real work. The real work can end up with a
> > deeper stack before blocking on locks or completions unrelated to
> > the readahead, leading to schedule() being called and an unplug
> > being issued at that point.  You might think it contrived, but if
> > you can't provide a guarantee that it can't happen then I have to
> > assume it will happen.
> 
> If you ended up in lock_page() somewhere along the way, the path would
> have been pretty much the same as it is now:
> 
> lock_page()
>         __lock_page()
>                 __wait_on_bit_lock()
>                         sync_page()
>                                 aops->sync_page();
>                                         block_sync_page()
>                                                 __blk_run_backing_dev()
> 
> and the dispatch follows after that. If your schedules are only due to,
> say, blocking on a mutex, then yes it'll be different. But is that
> really the case?

XFS metadata IO does not use the page cache anymore, so won't take
that path - no page locks are taken during read or write. Even
before that change contending on page locks was extremely rare as
XFs uses the buffer container for synchronisation.

AFAICT, we have nothing that will cause plugs to be flushed until
scheduling occurs. In many cases it will be at the same points as
before (the explicit flushes XFS had), but there are going to be new
ones....

Like this:

  0)     5360      40   zone_statistics+0xad/0xc0
  1)     5320     288   get_page_from_freelist+0x2cf/0x840
  2)     5032     304   __alloc_pages_nodemask+0x121/0x930
  3)     4728      48   kmem_getpages+0x62/0x160
  4)     4680      96   cache_grow+0x308/0x330
  5)     4584      80   cache_alloc_refill+0x21c/0x260
  6)     4504      16   __kmalloc+0x230/0x240
  7)     4488     176   virtqueue_add_buf_gfp+0x1f9/0x3e0
  8)     4312     144   do_virtblk_request+0x1f3/0x400
  9)     4168      32   __blk_run_queue+0x57/0x100
 10)     4136      80   flush_plug_list+0x133/0x1d0
 11)     4056      32   __blk_flush_plug+0x24/0x50
 12)     4024     160   schedule+0x867/0x9f0
 13)     3864     208   schedule_timeout+0x1f5/0x2c0
 14)     3656     144   wait_for_common+0xe7/0x190
 15)     3512      16   wait_for_completion+0x1d/0x20
 16)     3496      48   xfs_buf_iowait+0x36/0xb0
 17)     3448      32   _xfs_buf_read+0x98/0xa0
 18)     3416      48   xfs_buf_read+0xa2/0x100
 19)     3368      80   xfs_trans_read_buf+0x1db/0x680
......

This path adds roughly 500 bytes to the previous case of
immediate dispatch of the IO down through _xfs_buf_read()...

> I bet that worst case stack usage is exactly the same as before, and
> that's the only metric we really care about.

I've already demonstrated much worse stack usage with ext3 through
the page fault path via io_schedule(). io_schedule() never used to
dispatch IO and now it does. Similarly there are changes and
increases in XFS stack usage like above. IMO, worst case stack
usage is definitely increased by these changes.

> > My concern is that we're already under stack space stress in the
> > writeback path, so anything that has the potential to increase it
> > significantly is a major worry from my point of view...
> 
> I agree on writeback being a worry, and that's why I made the change
> (since it makes sense for other reasons, too). I just don't think we are
> worse of than before.

We certainly are.

Hmmm, I just noticed a new cumulative stack usage path through
direct reclaim - via congestion_wait() -> io_schedule()....

> >> If we ended up doing the IO
> >> dispatch before, then the only difference now is the stack usage of
> >> schedule() itself. Apart from that, as far as I can tell, there should
> >> not be much difference.
> > 
> > There's a difference between IO submission and IO dispatch. IO
> > submission is submit_bio thru to the plug; IO dispatch is from the
> > plug down to the disk. If they happen at the same place, there's no
> > problem. If IO dispatch is moved to schedule() via a plug....
> 
> The IO submission can easily and non-deterministically turn into an IO
> dispatch, so there's no real difference for the submitter. That was the
> case before. With the explicit plug now, you _know_ that the IO
> submission is only that and doesn't include IO dispatch.

You're violently agreeing with me that you've changed where the IO
dispatch path is run from. ;)

> Not until you
> schedule() or call blk_finish_plug(), both of which are events that you
> can control.

Well, not really - now taking any sleeping lock or waiting on
anything can trigger a plug flush where previously you had to
explicitly issue them. I'm not saying what we had is better, just
that there are implicit flushes with your changes that are
inherently uncontrollable...

Cheers,

Dave.
Christoph Hellwig April 12, 2011, 4:44 p.m. UTC | #4
On Tue, Apr 12, 2011 at 02:58:46PM +0200, Jens Axboe wrote:
> Supposedly it's faster to do it inline rather than punt the dispatch.
> But that may actually not be true, if you have multiple plugs going (and
> thus multiple contenders for the queue lock on dispatch). So lets play
> it safe and punt to kblockd, we can always revisit this later.

Note that this can be optimized further by adding a new helper that just
queues up work on kblockd without taking the queue lock, e.g. adding a
new

void blk_run_queue_async(struct request_queue *q)
{
	if (likely(!blk_queue_stopped(q)))
		queue_delayed_work(kblockd_workqueue, &q->delay_work, 0);
}

And replacing all

	__blk_run_queue(q, true);

callers with that, at which point they won't need the queuelock any
more.

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel
Jens Axboe April 12, 2011, 4:49 p.m. UTC | #5
On 2011-04-12 18:44, hch@infradead.org wrote:
> On Tue, Apr 12, 2011 at 02:58:46PM +0200, Jens Axboe wrote:
>> Supposedly it's faster to do it inline rather than punt the dispatch.
>> But that may actually not be true, if you have multiple plugs going (and
>> thus multiple contenders for the queue lock on dispatch). So lets play
>> it safe and punt to kblockd, we can always revisit this later.
> 
> Note that this can be optimized further by adding a new helper that just
> queues up work on kblockd without taking the queue lock, e.g. adding a
> new
> 
> void blk_run_queue_async(struct request_queue *q)
> {
> 	if (likely(!blk_queue_stopped(q)))
> 		queue_delayed_work(kblockd_workqueue, &q->delay_work, 0);
> }
> 
> And replacing all
> 
> 	__blk_run_queue(q, true);
> 
> callers with that, at which point they won't need the queuelock any
> more.

I realize that, in fact it's already safe as long as you pass in 'true'
for __blk_run_queue(). Before I had rewritten it to move the running
out, so that makes the trick a little difficult. This afternoon I also
tested it and saw no noticable difference, but I'll probably just do it
anyway as it makes sense.
Christoph Hellwig April 12, 2011, 4:54 p.m. UTC | #6
On Tue, Apr 12, 2011 at 06:49:53PM +0200, Jens Axboe wrote:
> I realize that, in fact it's already safe as long as you pass in 'true'
> for __blk_run_queue(). Before I had rewritten it to move the running
> out, so that makes the trick a little difficult. This afternoon I also
> tested it and saw no noticable difference, but I'll probably just do it
> anyway as it makes sense.

We still need the lock for __elv_add_request, so we'll need to keep the
logic anyway.  But splitting out the just queue to kblockd case from
__blk_run_queue and giving the latter a sane prototype still sounds
like a good idea to me.

Btw, now that we don't call the request_fn directly any more and thus
can't block, can the unplugging be moved into the preempt notifiers?

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel
Christoph Hellwig April 12, 2011, 4:58 p.m. UTC | #7
On Tue, Apr 12, 2011 at 11:31:17PM +1000, Dave Chinner wrote:
> I don't think so. e.g. in the XFS allocation path we do btree block
> readahead, then go do the real work. The real work can end up with a
> deeper stack before blocking on locks or completions unrelated to
> the readahead, leading to schedule() being called and an unplug
> being issued at that point.  You might think it contrived, but if
> you can't provide a guarantee that it can't happen then I have to
> assume it will happen.

In addition to the stack issue, which is a killer to this also has
latency implications.  Before we could submit a synchronous metadata
read request inside readpage or writepage and kick it off to the disk
immediately, while now it won't get submitted until we block the next
time, i.e. have done some more work that could have been used for
doing I/O in the background.  With the kblockd offload not only have
we spent more time but at the point where we finally kick it we
also need another context switch.  It seem like we really need to
go through the filesystems and explicitly flush the plugging queue
for such cases.  In fact a bio flag marking things as synchronous
metadata reads would help, but then again we need to clean up our
existing bio flags first..

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel
Jens Axboe April 12, 2011, 5:24 p.m. UTC | #8
On 2011-04-12 18:54, hch@infradead.org wrote:
> On Tue, Apr 12, 2011 at 06:49:53PM +0200, Jens Axboe wrote:
>> I realize that, in fact it's already safe as long as you pass in 'true'
>> for __blk_run_queue(). Before I had rewritten it to move the running
>> out, so that makes the trick a little difficult. This afternoon I also
>> tested it and saw no noticable difference, but I'll probably just do it
>> anyway as it makes sense.
> 
> We still need the lock for __elv_add_request, so we'll need to keep the
> logic anyway.  But splitting out the just queue to kblockd case from
> __blk_run_queue and giving the latter a sane prototype still sounds
> like a good idea to me.
> 
> Btw, now that we don't call the request_fn directly any more and thus
> can't block, can the unplugging be moved into the preempt notifiers?

It was only partly the reason, there's still the notice on preempt
(instead of schedule) and the runqueue lock problem. And if we allow
preempt, then we need to do disable preempt around all the plug logic.
Jens Axboe April 12, 2011, 5:29 p.m. UTC | #9
On 2011-04-12 18:58, hch@infradead.org wrote:
> On Tue, Apr 12, 2011 at 11:31:17PM +1000, Dave Chinner wrote:
>> I don't think so. e.g. in the XFS allocation path we do btree block
>> readahead, then go do the real work. The real work can end up with a
>> deeper stack before blocking on locks or completions unrelated to
>> the readahead, leading to schedule() being called and an unplug
>> being issued at that point.  You might think it contrived, but if
>> you can't provide a guarantee that it can't happen then I have to
>> assume it will happen.
> 
> In addition to the stack issue, which is a killer to this also has
> latency implications.  Before we could submit a synchronous metadata
> read request inside readpage or writepage and kick it off to the disk
> immediately, while now it won't get submitted until we block the next
> time, i.e. have done some more work that could have been used for
> doing I/O in the background.  With the kblockd offload not only have
> we spent more time but at the point where we finally kick it we
> also need another context switch.  It seem like we really need to
> go through the filesystems and explicitly flush the plugging queue
> for such cases.  In fact a bio flag marking things as synchronous
> metadata reads would help, but then again we need to clean up our
> existing bio flags first..

I think it would be a good idea to audit the SYNC cases, and if feasible
let that retain the 'immediate kick off' logic. If not, have some way to
signal that at least. Essentially allow some fine grained control of
what goes into the plug and what does not.
NeilBrown April 12, 2011, 9:08 p.m. UTC | #10
On Wed, 13 Apr 2011 00:34:52 +1000 Dave Chinner <david@fromorbit.com> wrote:

> On Tue, Apr 12, 2011 at 03:45:52PM +0200, Jens Axboe wrote:
> Not until you
> > schedule() or call blk_finish_plug(), both of which are events that you
> > can control.
> 
> Well, not really - now taking any sleeping lock or waiting on
> anything can trigger a plug flush where previously you had to
> explicitly issue them. I'm not saying what we had is better, just
> that there are implicit flushes with your changes that are
> inherently uncontrollable...

It's not just sleeping locks - if preempt is enabled a schedule can happen at
any time - at any depth.  I've seen a spin_unlock do it.

NeilBrown

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel
diff mbox

Patch

diff --git a/block/blk-core.c b/block/blk-core.c
index c6eaa1f..36b1a75 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -2665,7 +2665,7 @@  static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
 static void queue_unplugged(struct request_queue *q, unsigned int depth)
 {
 	trace_block_unplug_io(q, depth);
-	__blk_run_queue(q, false);
+	__blk_run_queue(q, true);
 
 	if (q->unplugged_fn)
 		q->unplugged_fn(q);