Message ID | 20190709090219.8784-1-damien.lemoal@wdc.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | block: Disable write plugging for zoned block devices | expand |
Not sure if I like the new helper or I'd prefer another 'else' in
blk_mq_make_request().
But all variants I can come up with are ugly and disgusting, so let's got the
route you proposed.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
On 7/9/19 2:02 AM, Damien Le Moal wrote: > Simultaneously writing to a sequential zone of a zoned block device > from multiple contexts requires mutual exclusion for BIO issuing to > ensure that writes happen sequentially. However, even for a well > behaved user correctly implementing such synchronization, BIO plugging > may interfere and result in BIOs from the different contextx to be > reordered if plugging is done outside of the mutual exclusion section, > e.g. the plug was started by a function higher in the call chain than > the function issuing BIOs. > > Context A Context B > > | blk_start_plug() > | ... > | seq_write_zone() > | mutex_lock(zone) > | submit_bio(bio-0) > | submit_bio(bio-1) > | mutex_unlock(zone) > | return > | ------------------------------> | seq_write_zone() > | mutex_lock(zone) > | submit_bio(bio-2) > | mutex_unlock(zone) > | <------------------------------ | > | blk_finish_plug() > > In the above example, despite the mutex synchronization resulting in the > correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being > issued after BIO 2 when the plug is released with blk_finish_plug(). > > To fix this problem, introduce the internal helper function > blk_mq_plug() to access the current context plug, return the current > plug only if the target device is not a zoned block device or if the > BIO to be plugged not a write operation. Otherwise, ignore the plug and > return NULL, resulting is all writes to zoned block device to never be > plugged. Are there classes of zoned devices for which the plug list is useful? If so, have you considered any other approaches, e.g. one plug list per request queue instead of one plug list per task in case of zoned devices? Thanks, Bart.
On Tue, Jul 09, 2019 at 06:02:19PM +0900, Damien Le Moal wrote: > Simultaneously writing to a sequential zone of a zoned block device > from multiple contexts requires mutual exclusion for BIO issuing to > ensure that writes happen sequentially. However, even for a well > behaved user correctly implementing such synchronization, BIO plugging > may interfere and result in BIOs from the different contextx to be > reordered if plugging is done outside of the mutual exclusion section, > e.g. the plug was started by a function higher in the call chain than > the function issuing BIOs. > > Context A Context B > > | blk_start_plug() > | ... > | seq_write_zone() > | mutex_lock(zone) > | submit_bio(bio-0) > | submit_bio(bio-1) > | mutex_unlock(zone) > | return > | ------------------------------> | seq_write_zone() > | mutex_lock(zone) > | submit_bio(bio-2) > | mutex_unlock(zone) > | <------------------------------ | > | blk_finish_plug() > > In the above example, despite the mutex synchronization resulting in the > correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being > issued after BIO 2 when the plug is released with blk_finish_plug(). I am wondering how you guarantee that context B is always run after context A. > > To fix this problem, introduce the internal helper function > blk_mq_plug() to access the current context plug, return the current > plug only if the target device is not a zoned block device or if the > BIO to be plugged not a write operation. Otherwise, ignore the plug and > return NULL, resulting is all writes to zoned block device to never be > plugged. Another candidate approach is to run the following code before releasing 'zone' lock: if (current->plug) blk_finish_plug(context->plug) Then we can fix zone specific issue in zone code only, and avoid generic blk-core change for zone issue. Thanks, Ming
Hi Ming, On 2019/07/09 23:29, Ming Lei wrote: > On Tue, Jul 09, 2019 at 06:02:19PM +0900, Damien Le Moal wrote: >> Simultaneously writing to a sequential zone of a zoned block device >> from multiple contexts requires mutual exclusion for BIO issuing to >> ensure that writes happen sequentially. However, even for a well >> behaved user correctly implementing such synchronization, BIO plugging >> may interfere and result in BIOs from the different contextx to be >> reordered if plugging is done outside of the mutual exclusion section, >> e.g. the plug was started by a function higher in the call chain than >> the function issuing BIOs. >> >> Context A Context B >> >> | blk_start_plug() >> | ... >> | seq_write_zone() >> | mutex_lock(zone) >> | submit_bio(bio-0) >> | submit_bio(bio-1) >> | mutex_unlock(zone) >> | return >> | ------------------------------> | seq_write_zone() >> | mutex_lock(zone) >> | submit_bio(bio-2) >> | mutex_unlock(zone) >> | <------------------------------ | >> | blk_finish_plug() >> >> In the above example, despite the mutex synchronization resulting in the >> correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being >> issued after BIO 2 when the plug is released with blk_finish_plug(). > > I am wondering how you guarantee that context B is always run after > context A. My example was a little too oversimplified. Think of a file system allocating blocks sequentially and issuing page I/Os for the allocated blocks sequentially. The typical sequence is: mutex_lock(zone) alloc_block_extent(zone) for all blocks in the extent submit_bio() mutex_unlock(zone) This way, it does not matter which context gets the lock first, all write BIOs for the zone remain sequential. The problem with plugs as explained above is that if the plug start/finish is not within the zone lock, reordering can happen for the 2 sequences of BIOs issued by the 2 contexts. We hit this problem with btrfs writepages (page writeback) where plugging is done before the above sequence execution, in the caller function of the page writeback processing, resulting in unaligned write errors. >> To fix this problem, introduce the internal helper function >> blk_mq_plug() to access the current context plug, return the current >> plug only if the target device is not a zoned block device or if the >> BIO to be plugged not a write operation. Otherwise, ignore the plug and >> return NULL, resulting is all writes to zoned block device to never be >> plugged. > > Another candidate approach is to run the following code before > releasing 'zone' lock: > > if (current->plug) > blk_finish_plug(context->plug) > > Then we can fix zone specific issue in zone code only, and avoid generic > blk-core change for zone issue. Yes indeed, that would work too. But this patch is precisely to avoid having to add such code and simplify implementing support for zoned block device in existing code. Furthermore, plugging for writes to sequential zones has no real value because mq-deadline will dispatch at most one write per zone. So writes for a single zone tend to accumulate in the scheduler queue, and that creates plenty of opportunities for merging small sequential writes (e.g. file system page BIOs). If you think this patch is really not appropriate, we can still address the problem case by case in the support we add for zoned devices. But again, a generic solution makes things simpler I think. Best regards.
On 2019/07/09 22:51, Bart Van Assche wrote: > On 7/9/19 2:02 AM, Damien Le Moal wrote: >> Simultaneously writing to a sequential zone of a zoned block device >> from multiple contexts requires mutual exclusion for BIO issuing to >> ensure that writes happen sequentially. However, even for a well >> behaved user correctly implementing such synchronization, BIO plugging >> may interfere and result in BIOs from the different contextx to be >> reordered if plugging is done outside of the mutual exclusion section, >> e.g. the plug was started by a function higher in the call chain than >> the function issuing BIOs. >> >> Context A Context B >> >> | blk_start_plug() >> | ... >> | seq_write_zone() >> | mutex_lock(zone) >> | submit_bio(bio-0) >> | submit_bio(bio-1) >> | mutex_unlock(zone) >> | return >> | ------------------------------> | seq_write_zone() >> | mutex_lock(zone) >> | submit_bio(bio-2) >> | mutex_unlock(zone) >> | <------------------------------ | >> | blk_finish_plug() >> >> In the above example, despite the mutex synchronization resulting in the >> correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being >> issued after BIO 2 when the plug is released with blk_finish_plug(). >> >> To fix this problem, introduce the internal helper function >> blk_mq_plug() to access the current context plug, return the current >> plug only if the target device is not a zoned block device or if the >> BIO to be plugged not a write operation. Otherwise, ignore the plug and >> return NULL, resulting is all writes to zoned block device to never be >> plugged. > > Are there classes of zoned devices for which the plug list is useful? If > so, have you considered any other approaches, e.g. one plug list per > request queue instead of one plug list per task in case of zoned devices? Plugging for writes to zoned block devices is not really useful at all. The reason is that for any user of the disk executing requests at a queue depth larger than 1, to preserve write ordering, mq-deadline must be used. With this scheduler, zone write locking will prevent dispatching more than one write request per zone at any time, resulting in the accumulation of sequential writes for a zone in the scheduler queue. This creates plenty of opportunities for merging small (i.e. single page) write BIOs with preceding pending requests, which is exactly the intent of plugging in the first place. A per request queue plug list would work, but it would require a single lock, going against blk-mq design principle. Such method would also result in a lot more changes for no real gain at all (for the reason explained above). Performance-wise, simply disabling per context plugging for writes only has no measurable impact and is far simpler I think. Best regards.
On 7/9/19 2:02 AM, Damien Le Moal wrote: > +static inline struct blk_plug *blk_mq_plug(struct request_queue *q, > + struct bio *bio) > +{ > + struct blk_plug *plug = current->plug; > + > + if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio))) > + return plug; > + > + /* Zoned block device write case: do not plug the BIO */ > + return NULL; > +} Can the 'plug' variable be left out from this function and can 'return plug' be changed into 'return current->plug'? Anyway: Reviewed-by: Bart Van Assche <bvanassche@acm.org>
On 7/10/19 11:32 AM, Bart Van Assche wrote: > On 7/9/19 2:02 AM, Damien Le Moal wrote: >> +static inline struct blk_plug *blk_mq_plug(struct request_queue *q, >> + struct bio *bio) >> +{ >> + struct blk_plug *plug = current->plug; >> + >> + if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio))) >> + return plug; >> + >> + /* Zoned block device write case: do not plug the BIO */ >> + return NULL; >> +} > > Can the 'plug' variable be left out from this function and can 'return > plug' be changed into 'return current->plug'? Anyway: Sure, that would be cleaner. Will Send a V2. > > Reviewed-by: Bart Van Assche <bvanassche@acm.org> Thanks. Can I add this to the V2 or would you prefer to see the revised patch first ? Best regards.
On 7/9/19 8:29 AM, Ming Lei wrote: > On Tue, Jul 09, 2019 at 06:02:19PM +0900, Damien Le Moal wrote: >> Simultaneously writing to a sequential zone of a zoned block device >> from multiple contexts requires mutual exclusion for BIO issuing to >> ensure that writes happen sequentially. However, even for a well >> behaved user correctly implementing such synchronization, BIO plugging >> may interfere and result in BIOs from the different contextx to be >> reordered if plugging is done outside of the mutual exclusion section, >> e.g. the plug was started by a function higher in the call chain than >> the function issuing BIOs. >> >> Context A Context B >> >> | blk_start_plug() >> | ... >> | seq_write_zone() >> | mutex_lock(zone) >> | submit_bio(bio-0) >> | submit_bio(bio-1) >> | mutex_unlock(zone) >> | return >> | ------------------------------> | seq_write_zone() >> | mutex_lock(zone) >> | submit_bio(bio-2) >> | mutex_unlock(zone) >> | <------------------------------ | >> | blk_finish_plug() >> >> In the above example, despite the mutex synchronization resulting in the >> correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being >> issued after BIO 2 when the plug is released with blk_finish_plug(). > > I am wondering how you guarantee that context B is always run after > context A. > >> >> To fix this problem, introduce the internal helper function >> blk_mq_plug() to access the current context plug, return the current >> plug only if the target device is not a zoned block device or if the >> BIO to be plugged not a write operation. Otherwise, ignore the plug and >> return NULL, resulting is all writes to zoned block device to never be >> plugged. > > Another candidate approach is to run the following code before > releasing 'zone' lock: > > if (current->plug) > blk_finish_plug(context->plug) > > Then we can fix zone specific issue in zone code only, and avoid generic > blk-core change for zone issue. I prefer that to the existing solution as well.
On 7/10/19 11:55 AM, Jens Axboe wrote: > On 7/9/19 8:29 AM, Ming Lei wrote: >> On Tue, Jul 09, 2019 at 06:02:19PM +0900, Damien Le Moal wrote: >>> Simultaneously writing to a sequential zone of a zoned block device >>> from multiple contexts requires mutual exclusion for BIO issuing to >>> ensure that writes happen sequentially. However, even for a well >>> behaved user correctly implementing such synchronization, BIO plugging >>> may interfere and result in BIOs from the different contextx to be >>> reordered if plugging is done outside of the mutual exclusion section, >>> e.g. the plug was started by a function higher in the call chain than >>> the function issuing BIOs. >>> >>> Context A Context B >>> >>> | blk_start_plug() >>> | ... >>> | seq_write_zone() >>> | mutex_lock(zone) >>> | submit_bio(bio-0) >>> | submit_bio(bio-1) >>> | mutex_unlock(zone) >>> | return >>> | ------------------------------> | seq_write_zone() >>> | mutex_lock(zone) >>> | submit_bio(bio-2) >>> | mutex_unlock(zone) >>> | <------------------------------ | >>> | blk_finish_plug() >>> >>> In the above example, despite the mutex synchronization resulting in the >>> correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being >>> issued after BIO 2 when the plug is released with blk_finish_plug(). >> >> I am wondering how you guarantee that context B is always run after >> context A. >> >>> >>> To fix this problem, introduce the internal helper function >>> blk_mq_plug() to access the current context plug, return the current >>> plug only if the target device is not a zoned block device or if the >>> BIO to be plugged not a write operation. Otherwise, ignore the plug and >>> return NULL, resulting is all writes to zoned block device to never be >>> plugged. >> >> Another candidate approach is to run the following code before >> releasing 'zone' lock: >> >> if (current->plug) >> blk_finish_plug(context->plug) >> >> Then we can fix zone specific issue in zone code only, and avoid generic >> blk-core change for zone issue. > > I prefer that to the existing solution as well. My apologies, you lost me: do you mean that you prefer Ming's suggestion and force FS or dm users to manually unplug in the case of zoned block devices ? Thanks.
On Tue, Jul 09, 2019 at 02:47:12PM +0000, Damien Le Moal wrote: > Hi Ming, > > On 2019/07/09 23:29, Ming Lei wrote: > > On Tue, Jul 09, 2019 at 06:02:19PM +0900, Damien Le Moal wrote: > >> Simultaneously writing to a sequential zone of a zoned block device > >> from multiple contexts requires mutual exclusion for BIO issuing to > >> ensure that writes happen sequentially. However, even for a well > >> behaved user correctly implementing such synchronization, BIO plugging > >> may interfere and result in BIOs from the different contextx to be > >> reordered if plugging is done outside of the mutual exclusion section, > >> e.g. the plug was started by a function higher in the call chain than > >> the function issuing BIOs. > >> > >> Context A Context B > >> > >> | blk_start_plug() > >> | ... > >> | seq_write_zone() > >> | mutex_lock(zone) > >> | submit_bio(bio-0) > >> | submit_bio(bio-1) > >> | mutex_unlock(zone) > >> | return > >> | ------------------------------> | seq_write_zone() > >> | mutex_lock(zone) > >> | submit_bio(bio-2) > >> | mutex_unlock(zone) > >> | <------------------------------ | > >> | blk_finish_plug() > >> > >> In the above example, despite the mutex synchronization resulting in the > >> correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being > >> issued after BIO 2 when the plug is released with blk_finish_plug(). > > > > I am wondering how you guarantee that context B is always run after > > context A. > > My example was a little too oversimplified. Think of a file system allocating > blocks sequentially and issuing page I/Os for the allocated blocks sequentially. > The typical sequence is: > > mutex_lock(zone) > alloc_block_extent(zone) > for all blocks in the extent > submit_bio() > mutex_unlock(zone) > > This way, it does not matter which context gets the lock first, all write BIOs > for the zone remain sequential. The problem with plugs as explained above is But wrt. the example in the commit log, it does matter which context gets the lock first, and it implies that context A has to run seq_write_zone() first, because you mentioned bio-2 has to be issued after bio-0 and bio-1. If there is 3rd context which is holding the lock, then either context A or context B can win in getting the lock first. So looks the zone lock itself isn't enough for maintaining the IO order. But that may not be related with this patch. Also seems there is issue with REQ_NOWAIT for zone support, for example, context A may see out-of-request and return earlier, however context B may get request and move on. > that if the plug start/finish is not within the zone lock, reordering can happen > for the 2 sequences of BIOs issued by the 2 contexts. > > We hit this problem with btrfs writepages (page writeback) where plugging is > done before the above sequence execution, in the caller function of the page > writeback processing, resulting in unaligned write errors. > > >> To fix this problem, introduce the internal helper function > >> blk_mq_plug() to access the current context plug, return the current > >> plug only if the target device is not a zoned block device or if the > >> BIO to be plugged not a write operation. Otherwise, ignore the plug and > >> return NULL, resulting is all writes to zoned block device to never be > >> plugged. > > > > Another candidate approach is to run the following code before > > releasing 'zone' lock: > > > > if (current->plug) > > blk_finish_plug(context->plug) > > > > Then we can fix zone specific issue in zone code only, and avoid generic > > blk-core change for zone issue. > > Yes indeed, that would work too. But this patch is precisely to avoid having to > add such code and simplify implementing support for zoned block device in > existing code. Furthermore, plugging for writes to sequential zones has no real > value because mq-deadline will dispatch at most one write per zone. So writes > for a single zone tend to accumulate in the scheduler queue, and that creates > plenty of opportunities for merging small sequential writes (e.g. file system > page BIOs). > > If you think this patch is really not appropriate, we can still address the > problem case by case in the support we add for zoned devices. But again, a > generic solution makes things simpler I think. OK, then I am fine with this simple generic approach. Thanks, Ming
On 7/10/19 12:10 PM, Ming Lei wrote: > On Tue, Jul 09, 2019 at 02:47:12PM +0000, Damien Le Moal wrote: >> Hi Ming, >> >> On 2019/07/09 23:29, Ming Lei wrote: >>> On Tue, Jul 09, 2019 at 06:02:19PM +0900, Damien Le Moal wrote: >>>> Simultaneously writing to a sequential zone of a zoned block device >>>> from multiple contexts requires mutual exclusion for BIO issuing to >>>> ensure that writes happen sequentially. However, even for a well >>>> behaved user correctly implementing such synchronization, BIO plugging >>>> may interfere and result in BIOs from the different contextx to be >>>> reordered if plugging is done outside of the mutual exclusion section, >>>> e.g. the plug was started by a function higher in the call chain than >>>> the function issuing BIOs. >>>> >>>> Context A Context B >>>> >>>> | blk_start_plug() >>>> | ... >>>> | seq_write_zone() >>>> | mutex_lock(zone) >>>> | submit_bio(bio-0) >>>> | submit_bio(bio-1) >>>> | mutex_unlock(zone) >>>> | return >>>> | ------------------------------> | seq_write_zone() >>>> | mutex_lock(zone) >>>> | submit_bio(bio-2) >>>> | mutex_unlock(zone) >>>> | <------------------------------ | >>>> | blk_finish_plug() >>>> >>>> In the above example, despite the mutex synchronization resulting in the >>>> correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being >>>> issued after BIO 2 when the plug is released with blk_finish_plug(). >>> >>> I am wondering how you guarantee that context B is always run after >>> context A. >> >> My example was a little too oversimplified. Think of a file system allocating >> blocks sequentially and issuing page I/Os for the allocated blocks sequentially. >> The typical sequence is: >> >> mutex_lock(zone) >> alloc_block_extent(zone) >> for all blocks in the extent >> submit_bio() >> mutex_unlock(zone) >> >> This way, it does not matter which context gets the lock first, all write BIOs >> for the zone remain sequential. The problem with plugs as explained above is > > But wrt. the example in the commit log, it does matter which context gets the lock > first, and it implies that context A has to run seq_write_zone() first, > because you mentioned bio-2 has to be issued after bio-0 and bio-1. > > If there is 3rd context which is holding the lock, then either context A or > context B can win in getting the lock first. So looks the zone lock itself > isn't enough for maintaining the IO order. But that may not be related > with this patch. For a raw block device driver, the zone lock is enough to maintain sequential write sequence. This is not visible in my example, because it is too simplistic. My apologies for the confusion. The reason is that the target sector of any zone write BIO must always be set to the end sector of the last issued write BIO for the zone. A more detailed and correct typical sequence for writing to a zone for a raw block device driver (e.g. a dm target) is: seq_write_zone() { mutex_lock(zone) /* bio-0 */ bio = bio_alloc() bio->bi_iter.bi_sector = zone->wp zone->wp += bio_sectors(bio) submit_bio(bio) /* bio-1 */ bio = bio_alloc() bio->bi_iter.bi_sector = zone->wp zone->wp += bio_sectors(bio) submit_bio(bio) ... mutex_unlock(zone) } Doing so, multiple contexts serialized with the zone mutex can keep writing sequentially, no matter the number of BIOs they issue and no matter the order in which they grab the zone lock. Note that here, the zone write pointer is a "soft" write pointer, not the actual device managed write pointer, because this latter WP is updated only on completion of the write commands, so visible to the host only on completion of the write BIOs. The "soft" write pointer is thus always equal to or in advance of the device hard WP. The soft WP must be re-synced to the hard WP in case of failed writes. For a file system, the zone hard WP is used as a starting point for block allocation. BIO issuing can then simply use the allocated extent sector directly instead of the zone soft write pointer. The block allocation code will manage the zone soft WP and do the resync with the device hard WP in case of write error. > Also seems there is issue with REQ_NOWAIT for zone support, for example, > context A may see out-of-request and return earlier, however context B > may get request and move on. Yes, but context B will move on from the last successfully written sector so sequential writes can still go on. It is the responsibility of the user code to deal with failed writes and how to recover from them. If REQ_NOWAIT is used for a BIO and causes submit_bio() to fail ( BLK_QC_T_NONE returned) in one context, that context may retry until it succeeds and increment the soft WP or bail out without incrementing the zone soft WP. In both cases, other contexts may simply resume trying to write from the still valid soft WP. Any number of methods exist for dealing with this. All the responsibility of the user (fs or dm) because sequential write issuing must in the first place be guaranteed by the users. In this regard, the generic block layer is fine. >> that if the plug start/finish is not within the zone lock, reordering can happen >> for the 2 sequences of BIOs issued by the 2 contexts. >> >> We hit this problem with btrfs writepages (page writeback) where plugging is >> done before the above sequence execution, in the caller function of the page >> writeback processing, resulting in unaligned write errors. >> >>>> To fix this problem, introduce the internal helper function >>>> blk_mq_plug() to access the current context plug, return the current >>>> plug only if the target device is not a zoned block device or if the >>>> BIO to be plugged not a write operation. Otherwise, ignore the plug and >>>> return NULL, resulting is all writes to zoned block device to never be >>>> plugged. >>> >>> Another candidate approach is to run the following code before >>> releasing 'zone' lock: >>> >>> if (current->plug) >>> blk_finish_plug(context->plug) >>> >>> Then we can fix zone specific issue in zone code only, and avoid generic >>> blk-core change for zone issue. >> >> Yes indeed, that would work too. But this patch is precisely to avoid having to >> add such code and simplify implementing support for zoned block device in >> existing code. Furthermore, plugging for writes to sequential zones has no real >> value because mq-deadline will dispatch at most one write per zone. So writes >> for a single zone tend to accumulate in the scheduler queue, and that creates >> plenty of opportunities for merging small sequential writes (e.g. file system >> page BIOs). >> >> If you think this patch is really not appropriate, we can still address the >> problem case by case in the support we add for zoned devices. But again, a >> generic solution makes things simpler I think. > > OK, then I am fine with this simple generic approach. Thanks. Best regards.
On 7/9/19 8:59 PM, Damien Le Moal wrote: > On 7/10/19 11:55 AM, Jens Axboe wrote: >> On 7/9/19 8:29 AM, Ming Lei wrote: >>> On Tue, Jul 09, 2019 at 06:02:19PM +0900, Damien Le Moal wrote: >>>> Simultaneously writing to a sequential zone of a zoned block device >>>> from multiple contexts requires mutual exclusion for BIO issuing to >>>> ensure that writes happen sequentially. However, even for a well >>>> behaved user correctly implementing such synchronization, BIO plugging >>>> may interfere and result in BIOs from the different contextx to be >>>> reordered if plugging is done outside of the mutual exclusion section, >>>> e.g. the plug was started by a function higher in the call chain than >>>> the function issuing BIOs. >>>> >>>> Context A Context B >>>> >>>> | blk_start_plug() >>>> | ... >>>> | seq_write_zone() >>>> | mutex_lock(zone) >>>> | submit_bio(bio-0) >>>> | submit_bio(bio-1) >>>> | mutex_unlock(zone) >>>> | return >>>> | ------------------------------> | seq_write_zone() >>>> | mutex_lock(zone) >>>> | submit_bio(bio-2) >>>> | mutex_unlock(zone) >>>> | <------------------------------ | >>>> | blk_finish_plug() >>>> >>>> In the above example, despite the mutex synchronization resulting in the >>>> correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being >>>> issued after BIO 2 when the plug is released with blk_finish_plug(). >>> >>> I am wondering how you guarantee that context B is always run after >>> context A. >>> >>>> >>>> To fix this problem, introduce the internal helper function >>>> blk_mq_plug() to access the current context plug, return the current >>>> plug only if the target device is not a zoned block device or if the >>>> BIO to be plugged not a write operation. Otherwise, ignore the plug and >>>> return NULL, resulting is all writes to zoned block device to never be >>>> plugged. >>> >>> Another candidate approach is to run the following code before >>> releasing 'zone' lock: >>> >>> if (current->plug) >>> blk_finish_plug(context->plug) >>> >>> Then we can fix zone specific issue in zone code only, and avoid generic >>> blk-core change for zone issue. >> >> I prefer that to the existing solution as well. > > My apologies, you lost me: do you mean that you prefer Ming's suggestion > and force FS or dm users to manually unplug in the case of zoned block > devices ? I take that back, I thought we could do it manually in the zoned code while dealing with the locking, but I don't think that is feasible.
diff --git a/block/blk-core.c b/block/blk-core.c index 8340f69670d8..3957ea6811c3 100644 --- a/block/blk-core.c +++ b/block/blk-core.c @@ -645,7 +645,7 @@ bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, struct request *rq; struct list_head *plug_list; - plug = current->plug; + plug = blk_mq_plug(q, bio); if (!plug) return false; diff --git a/block/blk-mq.c b/block/blk-mq.c index ce0f5f4ede70..90be5bb6fa1b 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -1969,7 +1969,7 @@ static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) cookie = request_to_qc_t(data.hctx, rq); - plug = current->plug; + plug = blk_mq_plug(q, bio); if (unlikely(is_flush_fua)) { blk_mq_put_ctx(data.ctx); blk_mq_bio_to_request(rq, bio); diff --git a/block/blk-mq.h b/block/blk-mq.h index 633a5a77ee8b..d9b1e94b82a4 100644 --- a/block/blk-mq.h +++ b/block/blk-mq.h @@ -238,4 +238,16 @@ static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap) qmap->mq_map[cpu] = 0; } +static inline struct blk_plug *blk_mq_plug(struct request_queue *q, + struct bio *bio) +{ + struct blk_plug *plug = current->plug; + + if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio))) + return plug; + + /* Zoned block device write case: do not plug the BIO */ + return NULL; +} + #endif
Simultaneously writing to a sequential zone of a zoned block device from multiple contexts requires mutual exclusion for BIO issuing to ensure that writes happen sequentially. However, even for a well behaved user correctly implementing such synchronization, BIO plugging may interfere and result in BIOs from the different contextx to be reordered if plugging is done outside of the mutual exclusion section, e.g. the plug was started by a function higher in the call chain than the function issuing BIOs. Context A Context B | blk_start_plug() | ... | seq_write_zone() | mutex_lock(zone) | submit_bio(bio-0) | submit_bio(bio-1) | mutex_unlock(zone) | return | ------------------------------> | seq_write_zone() | mutex_lock(zone) | submit_bio(bio-2) | mutex_unlock(zone) | <------------------------------ | | blk_finish_plug() In the above example, despite the mutex synchronization resulting in the correct BIO issuing order 0, 1, 2, context A BIOs 0 and 1 end up being issued after BIO 2 when the plug is released with blk_finish_plug(). To fix this problem, introduce the internal helper function blk_mq_plug() to access the current context plug, return the current plug only if the target device is not a zoned block device or if the BIO to be plugged not a write operation. Otherwise, ignore the plug and return NULL, resulting is all writes to zoned block device to never be plugged. Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com> --- block/blk-core.c | 2 +- block/blk-mq.c | 2 +- block/blk-mq.h | 12 ++++++++++++ 3 files changed, 14 insertions(+), 2 deletions(-)