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