per-tagset SRCU struct and quiesce
Subsystems (5)
Christoph Hellwig
maintainerHi all,
this series moves the SRCU struct used for quiescing to the tag_set as the SRCU critical sections for dispatch take about the same time on all queues anyway, and then adopts the series from Chao that provides tagset-wide quiesce to use that infrastructure.
Christoph Hellwig
maintainerPATCH 1/8 block: set the disk capacity to 0 in blk_mark_disk_dead
nvme and xen-blkfront are already doing this to stop buffered writes from creating dirty pages that can't be written out later. Move it to the common code. Note that this follows the xen-blkfront version that does not send and uevent as the uevent is a bit confusing when the device is about to go away a little later, and the the size change is just to stop buffered writes faster.
This also removes the comment about the ordering from nvme, as bd_mutex not only is gone entirely, but also hasn't been used for locking updates to the disk size long before that, and thus the ordering requirement documented there doesn't apply any more.
--- block/genhd.c | 3 +++ drivers/block/xen-blkfront.c | 1 - drivers/nvme/host/core.c | 7 +------ 3 files changed, 4 insertions(+), 7 deletions(-)
block/genhd.c +3 -0
@@ -555,6 +555,9 @@ void blk_mark_disk_dead(struct gendisk *disk) {set_bit (GD_DEAD , &disk->state);blk_queue_start_drain (disk->queue); + +/* stop buffered writers from dirtying pages that can't written out */ +set_capacity (disk,0 ); }EXPORT_SYMBOL_GPL (blk_mark_disk_dead);
drivers/block/xen-blkfront.c +0 -1
@@ -2129,7 +2129,6 @@ static void blkfront_closing(struct blkfront_info *info)if (info->rq && info->gd) {blk_mq_stop_hw_queues (info->rq);blk_mark_disk_dead (info->gd); -set_capacity (info->gd,0 ); }for_each_rinfo (info, rinfo, i) {
drivers/nvme/host/core.c +1 -6
@@ -5106,10 +5106,7 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)/* * Prepare a queue for teardown. * -* This must forcibly unquiesce queues to avoid blocking dispatch, and only set -* the capacity to 0 after that to avoid blocking dispatchers that may be -* holding bd_butex. This will end buffered writers dirtying pages that can't -* be synced. +* This must forcibly unquiesce queues to avoid blocking dispatch. */ static void nvme_set_queue_dying (struct nvme_ns *ns) {
@@ -5118,8 +5115,6 @@ static void nvme_set_queue_dying(struct nvme_ns *ns)blk_mark_disk_dead (ns->disk);nvme_start_ns_queue (ns); - -set_capacity_and_notify (ns->disk,0 ); }/**
Signature · 1 line
2.30.2
Re: [PATCH 1/8] block: set the disk capacity to 0 in blk_mark_disk_dead
Re: [PATCH 1/8] block: set the disk capacity to 0 in blk_mark_disk_dead
In reply to Christoph HellwigQuoted 24 lines · Christoph Hellwig
nvme and xen-blkfront are already doing this to stop buffered writes from creating dirty pages that can't be written out later. Move it to the common code. Note that this follows the xen-blkfront version that does not send and uevent as the uevent is a bit confusing when the device is about to go away a little later, and the the size change is just to stop buffered writes faster.
This also removes the comment about the ordering from nvme, as
bd_mutexnot only is gone entirely, but also hasn't been used for locking updates to the disk size long before that, and thus the ordering requirement documented there doesn't apply any more.Signed-off-by Christoph Hellwig <[email protected]>--- block/genhd.c | 3 +++ drivers/block/xen-blkfront.c | 1 - drivers/nvme/host/core.c | 7 +------ 3 files changed, 4 insertions(+), 7 deletions(-)block/genhd.c +0
-0
block/genhd.c void blk_mark_disk_dead(struct gendisk *disk) · line 560
Quoted 6 lines
block/genhd.c +3
-0@@ -555,6 +555,9 @@ void blk_mark_disk_dead(struct gendisk *disk) {set_bit (GD_DEAD , &disk->state);blk_queue_start_drain (disk->queue); + +/* stop buffered writers from dirtying pages that can't written out */ +set_capacity (disk,0 );
The idea makes sense:
Just one small issue on mtip32xx, which may call blk_mark_disk_dead() in irq context, and ->bd_size_lock is actually not irq safe.
But mtip32xx is already broken since blk_queue_start_drain() need mutex, maybe mtip32xx isn't actively used at all.
Thanks,
Ming
Christoph Hellwig
maintainerRe: [PATCH 1/8] block: set the disk capacity to 0 in blk_mark_disk_dead
Quoted 5 lines · Ming Lei
Just one small issue on mtip32xx, which may call
blk_mark_disk_dead()in irq context, and ->bd_size_lock is actually not irq safe.But mtip32xx is already broken since
blk_queue_start_drain()need mutex, maybe mtip32xx isn't actively used at all.
A while ago active users were still reported. But I bet no one regularly tests hot removals with these cards.
Re: [PATCH 1/8] block: set the disk capacity to 0 in blk_mark_disk_dead
In reply to Christoph HellwigQuoted 71 lines · Christoph Hellwig
nvme and xen-blkfront are already doing this to stop buffered writes from creating dirty pages that can't be written out later. Move it to the common code. Note that this follows the xen-blkfront version that does not send and uevent as the uevent is a bit confusing when the device is about to go away a little later, and the the size change is just to stop buffered writes faster.
This also removes the comment about the ordering from nvme, as
bd_mutexnot only is gone entirely, but also hasn't been used for locking updates to the disk size long before that, and thus the ordering requirement documented there doesn't apply any more.Signed-off-by Christoph Hellwig <[email protected]>--- block/genhd.c | 3 +++ drivers/block/xen-blkfront.c | 1 - drivers/nvme/host/core.c | 7 +------ 3 files changed, 4 insertions(+), 7 deletions(-)block/genhd.c +3
-0@@ -555,6 +555,9 @@ void blk_mark_disk_dead(struct gendisk *disk) {set_bit (GD_DEAD , &disk->state);blk_queue_start_drain (disk->queue); + +/* stop buffered writers from dirtying pages that can't written out */ +set_capacity (disk,0 ); }EXPORT_SYMBOL_GPL (blk_mark_disk_dead);drivers/block/xen-blkfront.c +0
-1@@ -2129,7 +2129,6 @@ static void blkfront_closing(struct blkfront_info *info)if (info->rq && info->gd) {blk_mq_stop_hw_queues (info->rq);blk_mark_disk_dead (info->gd); -set_capacity (info->gd,0 ); }for_each_rinfo (info, rinfo, i) {drivers/nvme/host/core.c +1
-6@@ -5106,10 +5106,7 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)/* * Prepare a queue for teardown. * -* This must forcibly unquiesce queues to avoid blocking dispatch, and only set -* the capacity to 0 after that to avoid blocking dispatchers that may be -* holding bd_butex. This will end buffered writers dirtying pages that can't -* be synced. +* This must forcibly unquiesce queues to avoid blocking dispatch. */ static void nvme_set_queue_dying (struct nvme_ns *ns) {@@ -5118,8 +5115,6 @@ static void nvme_set_queue_dying(struct nvme_ns *ns)blk_mark_disk_dead (ns->disk);nvme_start_ns_queue (ns); - -set_capacity_and_notify (ns->disk,0 ); }/**
Re: [PATCH 1/8] block: set the disk capacity to 0 in blk_mark_disk_dead
In reply to Christoph HellwigQuoted 62 lines · Christoph Hellwig
nvme and xen-blkfront are already doing this to stop buffered writes from creating dirty pages that can't be written out later. Move it to the common code. Note that this follows the xen-blkfront version that does not send and uevent as the uevent is a bit confusing when the device is about to go away a little later, and the the size change is just to stop buffered writes faster.
This also removes the comment about the ordering from nvme, as
bd_mutexnot only is gone entirely, but also hasn't been used for locking updates to the disk size long before that, and thus the ordering requirement documented there doesn't apply any more.Signed-off-by Christoph Hellwig <[email protected]>--- block/genhd.c | 3 +++ drivers/block/xen-blkfront.c | 1 - drivers/nvme/host/core.c | 7 +------ 3 files changed, 4 insertions(+), 7 deletions(-)block/genhd.c +3
-0@@ -555,6 +555,9 @@ void blk_mark_disk_dead(struct gendisk *disk) {set_bit (GD_DEAD , &disk->state);blk_queue_start_drain (disk->queue); + +/* stop buffered writers from dirtying pages that can't written out */ +set_capacity (disk,0 ); }EXPORT_SYMBOL_GPL (blk_mark_disk_dead);drivers/block/xen-blkfront.c +0
-1@@ -2129,7 +2129,6 @@ static void blkfront_closing(struct blkfront_info *info)if (info->rq && info->gd) {blk_mq_stop_hw_queues (info->rq);blk_mark_disk_dead (info->gd); -set_capacity (info->gd,0 ); }for_each_rinfo (info, rinfo, i) {drivers/nvme/host/core.c +1
-4@@ -5106,10 +5106,7 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)/* * Prepare a queue for teardown. * -* This must forcibly unquiesce queues to avoid blocking dispatch, and only set -* the capacity to 0 after that to avoid blocking dispatchers that may be -* holding bd_butex. This will end buffered writers dirtying pages that can't -* be synced. +* This must forcibly unquiesce queues to avoid blocking dispatch. */ static void nvme_set_queue_dying (struct nvme_ns *ns) {
drivers/nvme/host/core.c static void nvme_set_queue_dying(struct nvme_ns *ns) · line 5120
Quoted 8 lines
drivers/nvme/host/core.c +0
-2@@ -5118,8 +5115,6 @@ static void nvme_set_queue_dying(struct nvme_ns *ns)blk_mark_disk_dead (ns->disk);nvme_start_ns_queue (ns); - -set_capacity_and_notify (ns->disk,0 ); }/**
I'm ever so slightly concerned about not sending the uevent anymore; MD for one relies on that event to figure out if a device is down. And I'm also relatively sure that testing with MD on Xen had been relatively few.
What do we lose by using the 'notify' version instead?
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Christoph Hellwig
maintainerRe: [PATCH 1/8] block: set the disk capacity to 0 in blk_mark_disk_dead
Quoted 2 lines · Hannes Reinecke
I'm ever so slightly concerned about not sending the uevent anymore; MD for one relies on that event to figure out if a device is down.
Hmm, where? I actually just had customer reports about md not noticing nvme devices going down properly and thus looking into in-kernel delivery of notifications..
Quoted 3 lines
And I'm also relatively sure that testing with MD on Xen had been relatively few.
What do we lose by using the 'notify' version instead?
Mostly that we get incorrect resize events to userspace, but maybe because nvme is more common that might still be better overall.
Re: [PATCH 1/8] block: set the disk capacity to 0 in blk_mark_disk_dead
In reply to Christoph Hellwigblock/genhd.c void blk_mark_disk_dead(struct gendisk *disk) · line 559
Quoted 1 line · Christoph Hellwig
block/genhd.c +1
-0+/* stop buffered writers from dirtying pages that can't written out */
Hi Christoph,
If this series is reposted, please insert the missing verb "be" in the above comment.
Thanks,
Bart.
Christoph Hellwig
maintainerPATCH 2/8 blk-mq: skip non-mq queues in blk_mq_quiesce_queue
In reply to Christoph HellwigFor submit_bio based queues there is no (S)RCU critical section during I/O submission and thus nothing to wait for in blk_mq_wait_quiesce_done, so skip doing any synchronization.
--- block/blk-mq.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
block/blk-mq.c +3 -1
@@ -280,7 +280,9 @@ EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);void blk_mq_quiesce_queue (struct request_queue *q) {blk_mq_quiesce_queue_nowait (q); -blk_mq_wait_quiesce_done (q); +/* nothing to wait for non-mq queues */ +if (queue_is_mq (q)) +blk_mq_wait_quiesce_done (q); }EXPORT_SYMBOL_GPL (blk_mq_quiesce_queue);
Signature · 1 line
2.30.2
Re: [PATCH 2/8] blk-mq: skip non-mq queues in blk_mq_quiesce_queue
Re: [PATCH 2/8] blk-mq: skip non-mq queues in blk_mq_quiesce_queue
In reply to Christoph HellwigQuoted 14 lines · Christoph Hellwig
For
submit_biobased queues there is no (S)RCU critical section during I/O submission and thus nothing to wait for inblk_mq_wait_quiesce_done, so skip doing any synchronization.Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)block/blk-mq.c +0
-0
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); · line 285
Quoted 7 lines
block/blk-mq.c +3
-1@@ -280,7 +280,9 @@ EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);void blk_mq_quiesce_queue (struct request_queue *q) {blk_mq_quiesce_queue_nowait (q); -blk_mq_wait_quiesce_done (q); +/* nothing to wait for non-mq queues */ +if (queue_is_mq (q)) +blk_mq_wait_quiesce_done (q);
This interface can't work as expected for bio drivers, the only user should be del_gendisk(), but anyway the patch is fine:
thanks,
Ming
Christoph Hellwig
maintainerRe: [PATCH 2/8] blk-mq: skip non-mq queues in blk_mq_quiesce_queue
Quoted 7 lines · Ming Lei
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); · line 285
Quoted 4 lines
block/blk-mq.c +3
-1-blk_mq_wait_quiesce_done (q); +/* nothing to wait for non-mq queues */ +if (queue_is_mq (q)) +blk_mq_wait_quiesce_done (q);This interface can't work as expected for bio drivers, the only user should be
del_gendisk(), but anyway the patch is fine:
Another one is the wb_lat_usec sysfs attribute. But maybe it is better do just do this in the callers and WARN?
Re: [PATCH 2/8] blk-mq: skip non-mq queues in blk_mq_quiesce_queue
Quoted 11 lines · Christoph Hellwig
Quoted 7 lines · Ming Lei
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); · line 285
Quoted 4 lines
block/blk-mq.c +3
-1-blk_mq_wait_quiesce_done (q); +/* nothing to wait for non-mq queues */ +if (queue_is_mq (q)) +blk_mq_wait_quiesce_done (q);This interface can't work as expected for bio drivers, the only user should be
del_gendisk(), but anyway the patch is fine:Another one is the
wb_lat_usecsysfs attribute. But maybe it is better do just do this in the callers and WARN?
wbt is only available for blk-mq.
Thanks,
Ming
Re: [PATCH 2/8] blk-mq: skip non-mq queues in blk_mq_quiesce_queue
In reply to Christoph HellwigQuoted 14 lines · Christoph Hellwig
For
submit_biobased queues there is no (S)RCU critical section during I/O submission and thus nothing to wait for inblk_mq_wait_quiesce_done, so skip doing any synchronization.Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)block/blk-mq.c +0
-0
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); · line 285
Quoted 7 lines
block/blk-mq.c +3
-1@@ -280,7 +280,9 @@ EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);void blk_mq_quiesce_queue (struct request_queue *q) {blk_mq_quiesce_queue_nowait (q); -blk_mq_wait_quiesce_done (q); +/* nothing to wait for non-mq queues */ +if (queue_is_mq (q)) +blk_mq_wait_quiesce_done (q);
For the non-mq queues, maybe we should add a synchronize_rcu.
if (queue_is_mq(q)) blk_mq_wait_quiesce_done(q); else synchronize_rcu();
Quoted 4 lines
block/blk-mq.c +0
-0}EXPORT_SYMBOL_GPL (blk_mq_quiesce_queue);
Re: [PATCH 2/8] blk-mq: skip non-mq queues in blk_mq_quiesce_queue
Quoted 21 lines · Chao Leng
Quoted 18 lines · Christoph Hellwig
For
submit_biobased queues there is no (S)RCU critical section during I/O submission and thus nothing to wait for inblk_mq_wait_quiesce_done, so skip doing any synchronization.Signed-off-by Christoph Hellwig <[email protected]>---
block/blk-mq.c| 4 +++-1 file changed, 3 insertions(+), 1 deletion(-)block/blk-mq.c +0
-1@@ -280,7 +280,9 @@ EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); void blk_mq_quiesce_queue(struct request_queue *q) { blk_mq_quiesce_queue_nowait(q); - blk_mq_wait_quiesce_done(q);
Quoted 8 lines
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); · line 285
Quoted 3 lines
block/blk-mq.c +3
-0+ /* nothing to wait for non-mq queues */ + if (queue_is_mq(q)) + blk_mq_wait_quiesce_done(q);For the non-mq queues, maybe we should add a
synchronize_rcu.
if (queue_is_mq(q))blk_mq_wait_quiesce_done(q);
else
synchronize_rcu();
Please ignore this, synchronize_rcu is not needed.
Quoted 5 lines
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); · line 288
Quoted 3 lines
block/blk-mq.c +0
-0} EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);.
Re: [PATCH 2/8] blk-mq: skip non-mq queues in blk_mq_quiesce_queue
In reply to Christoph HellwigQuoted 14 lines · Christoph Hellwig
For
submit_biobased queues there is no (S)RCU critical section during I/O submission and thus nothing to wait for inblk_mq_wait_quiesce_done, so skip doing any synchronization.Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)block/blk-mq.c +0
-0
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); · line 288
Quoted 10 lines
block/blk-mq.c +3
-1@@ -280,7 +280,9 @@ EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);void blk_mq_quiesce_queue (struct request_queue *q) {blk_mq_quiesce_queue_nowait (q); -blk_mq_wait_quiesce_done (q); +/* nothing to wait for non-mq queues */ +if (queue_is_mq (q)) +blk_mq_wait_quiesce_done (q); }EXPORT_SYMBOL_GPL (blk_mq_quiesce_queue);
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Christoph Hellwig
maintainerPATCH 3/8 blk-mq: move the srcu_struct used for quiescing to the tagset
In reply to Christoph HellwigAll I/O submissions have fairly similar latencies, and a tagset-wide quiesce is a fairly common operation. Becuase there are a lot less tagsets there is also no need for the variable size allocation trick.
--- block/blk-core.c | 27 +++++---------------------- block/blk-mq.c | 25 +++++++++++++++++-------- block/blk-mq.h | 14 +++++++------- block/blk-sysfs.c | 9 ++------- block/blk.h | 9 +-------- block/genhd.c | 2 +- include/linux/blk-mq.h | 4 ++++ include/linux/blkdev.h | 9 --------- 8 files changed, 37 insertions(+), 62 deletions(-)
block/blk-core.c +5 -22
@@ -65,7 +65,6 @@ DEFINE_IDA(blk_queue_ida);* For queue allocation */ struct kmem_cache *blk_requestq_cachep; -struct kmem_cache *blk_requestq_srcu_cachep;/* * Controlling structure to kblockd
@@ -373,26 +372,20 @@ static void blk_timeout_work(struct work_struct *work) { } -struct request_queue *blk_alloc_queue (int node_id,bool alloc_srcu) +struct request_queue *blk_alloc_queue (int node_id) {struct request_queue *q; - q =kmem_cache_alloc_node (blk_get_queue_kmem_cache (alloc_srcu), -GFP_KERNEL | __GFP_ZERO, node_id); + q =kmem_cache_alloc_node (blk_requestq_cachep,GFP_KERNEL | __GFP_ZERO, + node_id);if (!q)return NULL ; -if (alloc_srcu) { -blk_queue_flag_set (QUEUE_FLAG_HAS_SRCU , q); -if (init_srcu_struct (q->srcu) !=0 ) - goto fail_q; - } - q->last_merge =NULL ; q->id =ida_alloc (&blk_queue_ida,GFP_KERNEL );if (q->id <0 ) - goto fail_srcu; + goto fail_q; q->stats =blk_alloc_queue_stats ();if (!q->stats)
@@ -435,11 +428,8 @@ struct request_queue *blk_alloc_queue(int node_id, bool alloc_srcu)blk_free_queue_stats (q->stats); fail_id:ida_free (&blk_queue_ida, q->id); -fail_srcu: -if (alloc_srcu) -cleanup_srcu_struct (q->srcu); fail_q: -kmem_cache_free (blk_get_queue_kmem_cache (alloc_srcu), q); +kmem_cache_free (blk_requestq_cachep, q);return NULL ; }
@@ -1184,9 +1174,6 @@ int __init blk_dev_init(void) sizeof_field(struct request, cmd_flags));BUILD_BUG_ON (REQ_OP_BITS +REQ_FLAG_BITS >8 * sizeof_field(struct bio, bi_opf)); -BUILD_BUG_ON (ALIGN (offsetof(struct request_queue , srcu), - __alignof__(struct request_queue )) != -sizeof (struct request_queue ));/* used for unplugging and affects IO latency/throughput - HIGHPRI */ kblockd_workqueue = alloc_workqueue("kblockd",
@@ -1197,10 +1184,6 @@ int __init blk_dev_init(void) blk_requestq_cachep =kmem_cache_create ("request_queue" ,sizeof (struct request_queue ),0 ,SLAB_PANIC ,NULL ); - blk_requestq_srcu_cachep =kmem_cache_create ("request_queue_srcu" , -sizeof (struct request_queue ) + -sizeof (struct srcu_struct ),0 ,SLAB_PANIC ,NULL ); - blk_debugfs_root =debugfs_create_dir ("block" ,NULL );return 0 ;
block/blk-mq.c +17 -8
@@ -261,8 +261,8 @@ EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);*/ void blk_mq_wait_quiesce_done (struct request_queue *q) { -if (blk_queue_has_srcu (q)) -synchronize_srcu (q->srcu); +if (q->tag_set->flags &BLK_MQ_F_BLOCKING ) +synchronize_srcu (&q->tag_set->srcu);else synchronize_rcu (); }
@@ -3971,7 +3971,7 @@ static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,struct request_queue *q;int ret; - q =blk_alloc_queue (set->numa_node, set->flags &BLK_MQ_F_BLOCKING ); + q =blk_alloc_queue (set->numa_node);if (!q)return ERR_PTR (-ENOMEM ); q->queuedata = queuedata;
@@ -4138,9 +4138,6 @@ static void blk_mq_update_poll_flag(struct request_queue *q)int blk_mq_init_allocated_queue (struct blk_mq_tag_set *set,struct request_queue *q) { -WARN_ON_ONCE (blk_queue_has_srcu (q) != - !!(set->flags &BLK_MQ_F_BLOCKING )); -/* mark the queue as mq asap */ q->mq_ops = set->ops;
@@ -4398,9 +4395,16 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)*/ if (set->nr_maps ==1 && set->nr_hw_queues > nr_cpu_ids) set->nr_hw_queues = nr_cpu_ids; + +if (set->flags &BLK_MQ_F_BLOCKING ) { + ret =init_srcu_struct (&set->srcu); +if (ret) +return ret; + } -if (blk_mq_alloc_tag_set_tags (set, set->nr_hw_queues) <0 ) -return -ENOMEM ; + ret =blk_mq_alloc_tag_set_tags (set, set->nr_hw_queues); +if (ret) + goto out_free_srcu; ret = -ENOMEM ;for (i =0 ; i < set->nr_maps; i++) {
@@ -4430,6 +4434,9 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) }kfree (set->tags); set->tags =NULL ; +out_free_srcu: +if (set->flags &BLK_MQ_F_BLOCKING ) +cleanup_srcu_struct (&set->srcu);return ret; }EXPORT_SYMBOL (blk_mq_alloc_tag_set);
@@ -4469,6 +4476,8 @@ void blk_mq_free_tag_set(struct blk_mq_tag_set *set)kfree (set->tags); set->tags =NULL ; +if (set->flags &BLK_MQ_F_BLOCKING ) +cleanup_srcu_struct (&set->srcu); }EXPORT_SYMBOL (blk_mq_free_tag_set);
block/blk-mq.h +7 -7
@@ -377,17 +377,17 @@ static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,/* run the code block in @dispatch_ops with rcu/srcu read lock held */ #define __blk_mq_run_dispatch_ops (q, check_sleep, dispatch_ops) \ do { \ - if (!blk_queue_has_srcu(q)) { \ - rcu_read_lock(); \ - (dispatch_ops); \ - rcu_read_unlock(); \ - } else { \ + if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) { \ int srcu_idx; \ \ might_sleep_if(check_sleep); \ - srcu_idx = srcu_read_lock((q)->srcu); \ + srcu_idx = srcu_read_lock(&((q)->tag_set->srcu)); \ (dispatch_ops); \ - srcu_read_unlock((q)->srcu, srcu_idx); \ + srcu_read_unlock(&((q)->tag_set->srcu), srcu_idx); \ + } else { \ + rcu_read_lock(); \ + (dispatch_ops); \ + rcu_read_unlock(); \ } \ } while (0)
block/blk-sysfs.c +2 -7
@@ -739,10 +739,8 @@ queue_attr_store(struct kobject *kobj, struct attribute *attr,static void blk_free_queue_rcu (struct rcu_head *rcu_head) { - struct request_queue *q = container_of(rcu_head, struct request_queue, - rcu_head); - -kmem_cache_free (blk_get_queue_kmem_cache (blk_queue_has_srcu (q)), q); +kmem_cache_free (blk_requestq_cachep, + container_of(rcu_head, struct request_queue, rcu_head)); }/**
@@ -779,9 +777,6 @@ static void blk_release_queue(struct kobject *kobj)if (queue_is_mq (q))blk_mq_release (q); -if (blk_queue_has_srcu (q)) -cleanup_srcu_struct (q->srcu); -ida_free (&blk_queue_ida, q->id);call_rcu (&q->rcu_head, blk_free_queue_rcu); }
block/blk.h +1 -8
@@ -27,7 +27,6 @@ struct blk_flush_queue { };extern struct kmem_cache *blk_requestq_cachep; -extern struct kmem_cache *blk_requestq_srcu_cachep;extern struct kobj_type blk_queue_ktype;extern struct ida blk_queue_ida;
@@ -420,13 +419,7 @@ int bio_add_hw_page(struct request_queue *q, struct bio *bio, struct page *page, unsigned int len, unsigned int offset, unsigned int max_sectors, bool *same_page); -static inline struct kmem_cache *blk_get_queue_kmem_cache (bool srcu) -{ -if (srcu) -return blk_requestq_srcu_cachep; -return blk_requestq_cachep; -} -struct request_queue *blk_alloc_queue (int node_id,bool alloc_srcu); +struct request_queue *blk_alloc_queue (int node_id);int disk_scan_partitions (struct gendisk *disk,fmode_t mode);
block/genhd.c +1 -1
@@ -1410,7 +1410,7 @@ struct gendisk *__blk_alloc_disk(int node, struct lock_class_key *lkclass)struct request_queue *q;struct gendisk *disk; - q =blk_alloc_queue (node, false); + q =blk_alloc_queue (node);if (!q)return NULL ;
include/linux/blk-mq.h +4 -0
@@ -7,6 +7,7 @@#include <linux/lockdep.h> #include <linux/scatterlist.h> #include <linux/prefetch.h> +#include <linux/srcu.h> struct blk_mq_tags ;struct blk_flush_queue ;
@@ -501,6 +502,8 @@ enum hctx_type {* @tag_list_lock: Serializes tag_list accesses. * @tag_list: List of the request queues that use this tag set. See also * request_queue.tag_set_list. +* @srcu: Use as lock when type of the request queue is blocking +* (BLK_MQ_F_BLOCKING). Must be the last member */ struct blk_mq_tag_set {struct blk_mq_queue_map map[HCTX_MAX_TYPES ];
@@ -521,6 +524,7 @@ struct blk_mq_tag_set {struct mutex tag_list_lock;struct list_head tag_list; +struct srcu_struct srcu; };/**
include/linux/blkdev.h +0 -9
@@ -22,7 +22,6 @@#include <linux/blkzoned.h> #include <linux/sched.h> #include <linux/sbitmap.h> -#include <linux/srcu.h> #include <linux/uuid.h> #include <linux/xarray.h>
@@ -543,18 +542,11 @@ struct request_queue {struct mutex debugfs_mutex;bool mq_sysfs_init_done; - -/** -* @srcu: Sleepable RCU. Use as lock when type of the request queue -* is blocking (BLK_MQ_F_BLOCKING). Must be the last member -*/ -struct srcu_struct srcu[]; };/* Keep blk_queue_flag_name[] in sync with the definitions below */ #define QUEUE_FLAG_STOPPED 0/* queue is stopped */ #define QUEUE_FLAG_DYING 1/* queue being torn down */ -#define QUEUE_FLAG_HAS_SRCU 2/* SRCU is allocated */ #define QUEUE_FLAG_NOMERGES 3/* disable merge attempts */ #define QUEUE_FLAG_SAME_COMP 4/* complete on same CPU-group */ #define QUEUE_FLAG_FAIL_IO 5/* fake timeout */
@@ -590,7 +582,6 @@ bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q);#define blk_queue_stopped (q) test_bit(QUEUE_FLAG_STOPPED, &(q)->queue_flags)#define blk_queue_dying (q) test_bit(QUEUE_FLAG_DYING, &(q)->queue_flags) -#define blk_queue_has_srcu (q) test_bit(QUEUE_FLAG_HAS_SRCU, &(q)->queue_flags)#define blk_queue_init_done (q) test_bit(QUEUE_FLAG_INIT_DONE, &(q)->queue_flags)#define blk_queue_nomerges (q) test_bit(QUEUE_FLAG_NOMERGES, &(q)->queue_flags)#define blk_queue_noxmerges (q) \
Signature · 1 line
2.30.2
Re: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
Re: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
In reply to Christoph HellwigQuoted 2 lines · Christoph Hellwig
All I/O submissions have fairly similar latencies, and a tagset-wide quiesce is a fairly common operation. Becuase there are a lot less
s/Becuase/Because
include/linux/blk-mq.h enum hctx_type { · line 506
Quoted 6 lines
include/linux/blk-mq.h +2
-0@@ -501,6 +502,8 @@ enum hctx_type {* @tag_list_lock: Serializes tag_list accesses. * @tag_list: List of the request queues that use this tag set. See also * request_queue.tag_set_list. +* @srcu: Use as lock when type of the request queue is blocking +* (BLK_MQ_F_BLOCKING). Must be the last member
Since you're not dealing with flexible arrays anymore, I don't think srcu strictly needs to be the last member.
The code looks great, though!
Quoted 9 lines
include/linux/blk-mq.h +1
-0*/ struct blk_mq_tag_set {struct blk_mq_queue_map map[HCTX_MAX_TYPES ];@@ -521,6 +524,7 @@ struct blk_mq_tag_set {struct mutex tag_list_lock;struct list_head tag_list; +struct srcu_struct srcu; };
Christoph Hellwig
maintainerRe: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
Quoted 2 lines · Keith Busch
Since you're not dealing with flexible arrays anymore, I don't think srcu strictly needs to be the last member.
True. I thought I had removed this sentence when copying the comment over, but it looks like I messed that up somewhow.
Re: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
In reply to Christoph HellwigQuoted 5 lines · Christoph Hellwig
All I/O submissions have fairly similar latencies, and a tagset-wide quiesce is a fairly common operation. Becuase there are a lot less tagsets there is also no need for the variable size allocation trick.
Signed-off-by Christoph Hellwig <[email protected]>
Quoted 10 lines
--- block/blk-core.c | 27 +++++---------------------- block/blk-mq.c | 25 +++++++++++++++++-------- block/blk-mq.h | 14 +++++++------- block/blk-sysfs.c | 9 ++------- block/blk.h | 9 +-------- block/genhd.c | 2 +- include/linux/blk-mq.h | 4 ++++ include/linux/blkdev.h | 9 --------- 8 files changed, 37 insertions(+), 62 deletions(-)
q->tag_set is supposed to be live when calling blk_mq_quiesce_queue and blk_mq_unquiesce_queue, especially del_gendisk() quieses request queue, so looks fine:
Thanks,
Ming
Re: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
In reply to Christoph HellwigQuoted 330 lines · Christoph Hellwig
All I/O submissions have fairly similar latencies, and a tagset-wide quiesce is a fairly common operation. Becuase there are a lot less tagsets there is also no need for the variable size allocation trick.
Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-core.c | 27 +++++---------------------- block/blk-mq.c | 25 +++++++++++++++++-------- block/blk-mq.h | 14 +++++++------- block/blk-sysfs.c | 9 ++------- block/blk.h | 9 +-------- block/genhd.c | 2 +- include/linux/blk-mq.h | 4 ++++ include/linux/blkdev.h | 9 --------- 8 files changed, 37 insertions(+), 62 deletions(-)block/blk-core.c +5
-22@@ -65,7 +65,6 @@ DEFINE_IDA(blk_queue_ida);* For queue allocation */ struct kmem_cache *blk_requestq_cachep; -struct kmem_cache *blk_requestq_srcu_cachep;/* * Controlling structure to kblockd @@ -373,26 +372,20 @@ static void blk_timeout_work(struct work_struct *work) { } -struct request_queue *blk_alloc_queue (int node_id,bool alloc_srcu) +struct request_queue *blk_alloc_queue (int node_id) {struct request_queue *q; - q =kmem_cache_alloc_node (blk_get_queue_kmem_cache (alloc_srcu), -GFP_KERNEL | __GFP_ZERO, node_id); + q =kmem_cache_alloc_node (blk_requestq_cachep,GFP_KERNEL | __GFP_ZERO, + node_id);if (!q)return NULL ; -if (alloc_srcu) { -blk_queue_flag_set (QUEUE_FLAG_HAS_SRCU , q); -if (init_srcu_struct (q->srcu) !=0 ) - goto fail_q; - } - q->last_merge =NULL ; q->id =ida_alloc (&blk_queue_ida,GFP_KERNEL );if (q->id <0 ) - goto fail_srcu; + goto fail_q; q->stats =blk_alloc_queue_stats ();if (!q->stats)@@ -435,11 +428,8 @@ struct request_queue *blk_alloc_queue(int node_id, bool alloc_srcu)blk_free_queue_stats (q->stats); fail_id:ida_free (&blk_queue_ida, q->id); -fail_srcu: -if (alloc_srcu) -cleanup_srcu_struct (q->srcu); fail_q: -kmem_cache_free (blk_get_queue_kmem_cache (alloc_srcu), q); +kmem_cache_free (blk_requestq_cachep, q);return NULL ; }@@ -1184,9 +1174,6 @@ int __init blk_dev_init(void) sizeof_field(struct request, cmd_flags));BUILD_BUG_ON (REQ_OP_BITS +REQ_FLAG_BITS >8 * sizeof_field(struct bio, bi_opf)); -BUILD_BUG_ON (ALIGN (offsetof(struct request_queue , srcu), - __alignof__(struct request_queue )) != -sizeof (struct request_queue ));/* used for unplugging and affects IO latency/throughput - HIGHPRI */ kblockd_workqueue = alloc_workqueue("kblockd",@@ -1197,10 +1184,6 @@ int __init blk_dev_init(void) blk_requestq_cachep =kmem_cache_create ("request_queue" ,sizeof (struct request_queue ),0 ,SLAB_PANIC ,NULL ); - blk_requestq_srcu_cachep =kmem_cache_create ("request_queue_srcu" , -sizeof (struct request_queue ) + -sizeof (struct srcu_struct ),0 ,SLAB_PANIC ,NULL ); - blk_debugfs_root =debugfs_create_dir ("block" ,NULL );return 0 ;block/blk-mq.c +17
-8@@ -261,8 +261,8 @@ EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);*/ void blk_mq_wait_quiesce_done (struct request_queue *q) { -if (blk_queue_has_srcu (q)) -synchronize_srcu (q->srcu); +if (q->tag_set->flags &BLK_MQ_F_BLOCKING ) +synchronize_srcu (&q->tag_set->srcu);else synchronize_rcu (); }@@ -3971,7 +3971,7 @@ static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,struct request_queue *q;int ret; - q =blk_alloc_queue (set->numa_node, set->flags &BLK_MQ_F_BLOCKING ); + q =blk_alloc_queue (set->numa_node);if (!q)return ERR_PTR (-ENOMEM ); q->queuedata = queuedata;@@ -4138,9 +4138,6 @@ static void blk_mq_update_poll_flag(struct request_queue *q)int blk_mq_init_allocated_queue (struct blk_mq_tag_set *set,struct request_queue *q) { -WARN_ON_ONCE (blk_queue_has_srcu (q) != - !!(set->flags &BLK_MQ_F_BLOCKING )); -/* mark the queue as mq asap */ q->mq_ops = set->ops;@@ -4398,9 +4395,16 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)*/ if (set->nr_maps ==1 && set->nr_hw_queues > nr_cpu_ids) set->nr_hw_queues = nr_cpu_ids; + +if (set->flags &BLK_MQ_F_BLOCKING ) { + ret =init_srcu_struct (&set->srcu); +if (ret) +return ret; + } -if (blk_mq_alloc_tag_set_tags (set, set->nr_hw_queues) <0 ) -return -ENOMEM ; + ret =blk_mq_alloc_tag_set_tags (set, set->nr_hw_queues); +if (ret) + goto out_free_srcu; ret = -ENOMEM ;for (i =0 ; i < set->nr_maps; i++) {@@ -4430,6 +4434,9 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) }kfree (set->tags); set->tags =NULL ; +out_free_srcu: +if (set->flags &BLK_MQ_F_BLOCKING ) +cleanup_srcu_struct (&set->srcu);return ret; }EXPORT_SYMBOL (blk_mq_alloc_tag_set);@@ -4469,6 +4476,8 @@ void blk_mq_free_tag_set(struct blk_mq_tag_set *set)kfree (set->tags); set->tags =NULL ; +if (set->flags &BLK_MQ_F_BLOCKING ) +cleanup_srcu_struct (&set->srcu); }EXPORT_SYMBOL (blk_mq_free_tag_set);block/blk-mq.h +7
-7@@ -377,17 +377,17 @@ static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,/* run the code block in @dispatch_ops with rcu/srcu read lock held */ #define __blk_mq_run_dispatch_ops (q, check_sleep, dispatch_ops) \ do { \ - if (!blk_queue_has_srcu(q)) { \ - rcu_read_lock(); \ - (dispatch_ops); \ - rcu_read_unlock(); \ - } else { \ + if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) { \ int srcu_idx; \ \ might_sleep_if(check_sleep); \ - srcu_idx = srcu_read_lock((q)->srcu); \ + srcu_idx = srcu_read_lock(&((q)->tag_set->srcu)); \ (dispatch_ops); \ - srcu_read_unlock((q)->srcu, srcu_idx); \ + srcu_read_unlock(&((q)->tag_set->srcu), srcu_idx); \ + } else { \ + rcu_read_lock(); \ + (dispatch_ops); \ + rcu_read_unlock(); \ } \ } while (0)block/blk-sysfs.c +2
-7@@ -739,10 +739,8 @@ queue_attr_store(struct kobject *kobj, struct attribute *attr,static void blk_free_queue_rcu (struct rcu_head *rcu_head) { - struct request_queue *q = container_of(rcu_head, struct request_queue, - rcu_head); - -kmem_cache_free (blk_get_queue_kmem_cache (blk_queue_has_srcu (q)), q); +kmem_cache_free (blk_requestq_cachep, + container_of(rcu_head, struct request_queue, rcu_head)); }/** @@ -779,9 +777,6 @@ static void blk_release_queue(struct kobject *kobj)if (queue_is_mq (q))blk_mq_release (q); -if (blk_queue_has_srcu (q)) -cleanup_srcu_struct (q->srcu); -ida_free (&blk_queue_ida, q->id);call_rcu (&q->rcu_head, blk_free_queue_rcu); }block/blk.h +1
-8@@ -27,7 +27,6 @@ struct blk_flush_queue { };extern struct kmem_cache *blk_requestq_cachep; -extern struct kmem_cache *blk_requestq_srcu_cachep;extern struct kobj_type blk_queue_ktype;extern struct ida blk_queue_ida;@@ -420,13 +419,7 @@ int bio_add_hw_page(struct request_queue *q, struct bio *bio, struct page *page, unsigned int len, unsigned int offset, unsigned int max_sectors, bool *same_page); -static inline struct kmem_cache *blk_get_queue_kmem_cache (bool srcu) -{ -if (srcu) -return blk_requestq_srcu_cachep; -return blk_requestq_cachep; -} -struct request_queue *blk_alloc_queue (int node_id,bool alloc_srcu); +struct request_queue *blk_alloc_queue (int node_id);int disk_scan_partitions (struct gendisk *disk,fmode_t mode);block/genhd.c +1
-1@@ -1410,7 +1410,7 @@ struct gendisk *__blk_alloc_disk(int node, struct lock_class_key *lkclass)struct request_queue *q;struct gendisk *disk; - q =blk_alloc_queue (node, false); + q =blk_alloc_queue (node);if (!q)return NULL ;include/linux/blk-mq.h +4
-0@@ -7,6 +7,7 @@#include <linux/lockdep.h> #include <linux/scatterlist.h> #include <linux/prefetch.h> +#include <linux/srcu.h> struct blk_mq_tags ;struct blk_flush_queue ;@@ -501,6 +502,8 @@ enum hctx_type {* @tag_list_lock: Serializes tag_list accesses. * @tag_list: List of the request queues that use this tag set. See also * request_queue.tag_set_list. +* @srcu: Use as lock when type of the request queue is blocking +* (BLK_MQ_F_BLOCKING). Must be the last member */ struct blk_mq_tag_set {struct blk_mq_queue_map map[HCTX_MAX_TYPES ];@@ -521,6 +524,7 @@ struct blk_mq_tag_set {struct mutex tag_list_lock;struct list_head tag_list; +struct srcu_struct srcu; };/** include/linux/blkdev.h +0
-9@@ -22,7 +22,6 @@#include <linux/blkzoned.h> #include <linux/sched.h> #include <linux/sbitmap.h> -#include <linux/srcu.h> #include <linux/uuid.h> #include <linux/xarray.h> @@ -543,18 +542,11 @@ struct request_queue {struct mutex debugfs_mutex;bool mq_sysfs_init_done; - -/** -* @srcu: Sleepable RCU. Use as lock when type of the request queue -* is blocking (BLK_MQ_F_BLOCKING). Must be the last member -*/ -struct srcu_struct srcu[]; };/* Keep blk_queue_flag_name[] in sync with the definitions below */ #define QUEUE_FLAG_STOPPED 0/* queue is stopped */ #define QUEUE_FLAG_DYING 1/* queue being torn down */ -#define QUEUE_FLAG_HAS_SRCU 2/* SRCU is allocated */ #define QUEUE_FLAG_NOMERGES 3/* disable merge attempts */ #define QUEUE_FLAG_SAME_COMP 4/* complete on same CPU-group */ #define QUEUE_FLAG_FAIL_IO 5/* fake timeout */ @@ -590,7 +582,6 @@ bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q);#define blk_queue_stopped (q) test_bit(QUEUE_FLAG_STOPPED, &(q)->queue_flags)#define blk_queue_dying (q) test_bit(QUEUE_FLAG_DYING, &(q)->queue_flags) -#define blk_queue_has_srcu (q) test_bit(QUEUE_FLAG_HAS_SRCU, &(q)->queue_flags)#define blk_queue_init_done (q) test_bit(QUEUE_FLAG_INIT_DONE, &(q)->queue_flags)#define blk_queue_nomerges (q) test_bit(QUEUE_FLAG_NOMERGES, &(q)->queue_flags)#define blk_queue_noxmerges (q) \
Re: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
In reply to Christoph HellwigQuoted 5 lines · Christoph Hellwig
All I/O submissions have fairly similar latencies, and a tagset-wide quiesce is a fairly common operation. Becuase there are a lot less tagsets there is also no need for the variable size allocation trick.
Signed-off-by Christoph Hellwig <[email protected]>
Quoted 11 lines
--- block/blk-core.c | 27 +++++---------------------- block/blk-mq.c | 25 +++++++++++++++++-------- block/blk-mq.h | 14 +++++++------- block/blk-sysfs.c | 9 ++------- block/blk.h | 9 +-------- block/genhd.c | 2 +- include/linux/blk-mq.h | 4 ++++ include/linux/blkdev.h | 9 --------- 8 files changed, 37 insertions(+), 62 deletions(-)
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Re: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
In reply to Christoph HellwigQuoted 283 lines · Christoph Hellwig
All I/O submissions have fairly similar latencies, and a tagset-wide quiesce is a fairly common operation. Becuase there are a lot less tagsets there is also no need for the variable size allocation trick.
Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-core.c | 27 +++++---------------------- block/blk-mq.c | 25 +++++++++++++++++-------- block/blk-mq.h | 14 +++++++------- block/blk-sysfs.c | 9 ++------- block/blk.h | 9 +-------- block/genhd.c | 2 +- include/linux/blk-mq.h | 4 ++++ include/linux/blkdev.h | 9 --------- 8 files changed, 37 insertions(+), 62 deletions(-)block/blk-core.c +5
-22@@ -65,7 +65,6 @@ DEFINE_IDA(blk_queue_ida);* For queue allocation */ struct kmem_cache *blk_requestq_cachep; -struct kmem_cache *blk_requestq_srcu_cachep;/* * Controlling structure to kblockd @@ -373,26 +372,20 @@ static void blk_timeout_work(struct work_struct *work) { } -struct request_queue *blk_alloc_queue (int node_id,bool alloc_srcu) +struct request_queue *blk_alloc_queue (int node_id) {struct request_queue *q; - q =kmem_cache_alloc_node (blk_get_queue_kmem_cache (alloc_srcu), -GFP_KERNEL | __GFP_ZERO, node_id); + q =kmem_cache_alloc_node (blk_requestq_cachep,GFP_KERNEL | __GFP_ZERO, + node_id);if (!q)return NULL ; -if (alloc_srcu) { -blk_queue_flag_set (QUEUE_FLAG_HAS_SRCU , q); -if (init_srcu_struct (q->srcu) !=0 ) - goto fail_q; - } - q->last_merge =NULL ; q->id =ida_alloc (&blk_queue_ida,GFP_KERNEL );if (q->id <0 ) - goto fail_srcu; + goto fail_q; q->stats =blk_alloc_queue_stats ();if (!q->stats)@@ -435,11 +428,8 @@ struct request_queue *blk_alloc_queue(int node_id, bool alloc_srcu)blk_free_queue_stats (q->stats); fail_id:ida_free (&blk_queue_ida, q->id); -fail_srcu: -if (alloc_srcu) -cleanup_srcu_struct (q->srcu); fail_q: -kmem_cache_free (blk_get_queue_kmem_cache (alloc_srcu), q); +kmem_cache_free (blk_requestq_cachep, q);return NULL ; }@@ -1184,9 +1174,6 @@ int __init blk_dev_init(void) sizeof_field(struct request, cmd_flags));BUILD_BUG_ON (REQ_OP_BITS +REQ_FLAG_BITS >8 * sizeof_field(struct bio, bi_opf)); -BUILD_BUG_ON (ALIGN (offsetof(struct request_queue , srcu), - __alignof__(struct request_queue )) != -sizeof (struct request_queue ));/* used for unplugging and affects IO latency/throughput - HIGHPRI */ kblockd_workqueue = alloc_workqueue("kblockd",@@ -1197,10 +1184,6 @@ int __init blk_dev_init(void) blk_requestq_cachep =kmem_cache_create ("request_queue" ,sizeof (struct request_queue ),0 ,SLAB_PANIC ,NULL ); - blk_requestq_srcu_cachep =kmem_cache_create ("request_queue_srcu" , -sizeof (struct request_queue ) + -sizeof (struct srcu_struct ),0 ,SLAB_PANIC ,NULL ); - blk_debugfs_root =debugfs_create_dir ("block" ,NULL );return 0 ;block/blk-mq.c +17
-8@@ -261,8 +261,8 @@ EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);*/ void blk_mq_wait_quiesce_done (struct request_queue *q) { -if (blk_queue_has_srcu (q)) -synchronize_srcu (q->srcu); +if (q->tag_set->flags &BLK_MQ_F_BLOCKING ) +synchronize_srcu (&q->tag_set->srcu);else synchronize_rcu (); }@@ -3971,7 +3971,7 @@ static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,struct request_queue *q;int ret; - q =blk_alloc_queue (set->numa_node, set->flags &BLK_MQ_F_BLOCKING ); + q =blk_alloc_queue (set->numa_node);if (!q)return ERR_PTR (-ENOMEM ); q->queuedata = queuedata;@@ -4138,9 +4138,6 @@ static void blk_mq_update_poll_flag(struct request_queue *q)int blk_mq_init_allocated_queue (struct blk_mq_tag_set *set,struct request_queue *q) { -WARN_ON_ONCE (blk_queue_has_srcu (q) != - !!(set->flags &BLK_MQ_F_BLOCKING )); -/* mark the queue as mq asap */ q->mq_ops = set->ops;@@ -4398,9 +4395,16 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)*/ if (set->nr_maps ==1 && set->nr_hw_queues > nr_cpu_ids) set->nr_hw_queues = nr_cpu_ids; + +if (set->flags &BLK_MQ_F_BLOCKING ) { + ret =init_srcu_struct (&set->srcu); +if (ret) +return ret; + } -if (blk_mq_alloc_tag_set_tags (set, set->nr_hw_queues) <0 ) -return -ENOMEM ; + ret =blk_mq_alloc_tag_set_tags (set, set->nr_hw_queues); +if (ret) + goto out_free_srcu; ret = -ENOMEM ;for (i =0 ; i < set->nr_maps; i++) {@@ -4430,6 +4434,9 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) }kfree (set->tags); set->tags =NULL ; +out_free_srcu: +if (set->flags &BLK_MQ_F_BLOCKING ) +cleanup_srcu_struct (&set->srcu);return ret; }EXPORT_SYMBOL (blk_mq_alloc_tag_set);@@ -4469,6 +4476,8 @@ void blk_mq_free_tag_set(struct blk_mq_tag_set *set)kfree (set->tags); set->tags =NULL ; +if (set->flags &BLK_MQ_F_BLOCKING ) +cleanup_srcu_struct (&set->srcu); }EXPORT_SYMBOL (blk_mq_free_tag_set);block/blk-mq.h +7
-7@@ -377,17 +377,17 @@ static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,/* run the code block in @dispatch_ops with rcu/srcu read lock held */ #define __blk_mq_run_dispatch_ops (q, check_sleep, dispatch_ops) \ do { \ - if (!blk_queue_has_srcu(q)) { \ - rcu_read_lock(); \ - (dispatch_ops); \ - rcu_read_unlock(); \ - } else { \ + if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) { \ int srcu_idx; \ \ might_sleep_if(check_sleep); \ - srcu_idx = srcu_read_lock((q)->srcu); \ + srcu_idx = srcu_read_lock(&((q)->tag_set->srcu)); \ (dispatch_ops); \ - srcu_read_unlock((q)->srcu, srcu_idx); \ + srcu_read_unlock(&((q)->tag_set->srcu), srcu_idx); \ + } else { \ + rcu_read_lock(); \ + (dispatch_ops); \ + rcu_read_unlock(); \ } \ } while (0)block/blk-sysfs.c +2
-7@@ -739,10 +739,8 @@ queue_attr_store(struct kobject *kobj, struct attribute *attr,static void blk_free_queue_rcu (struct rcu_head *rcu_head) { - struct request_queue *q = container_of(rcu_head, struct request_queue, - rcu_head); - -kmem_cache_free (blk_get_queue_kmem_cache (blk_queue_has_srcu (q)), q); +kmem_cache_free (blk_requestq_cachep, + container_of(rcu_head, struct request_queue, rcu_head)); }/** @@ -779,9 +777,6 @@ static void blk_release_queue(struct kobject *kobj)if (queue_is_mq (q))blk_mq_release (q); -if (blk_queue_has_srcu (q)) -cleanup_srcu_struct (q->srcu); -ida_free (&blk_queue_ida, q->id);call_rcu (&q->rcu_head, blk_free_queue_rcu); }block/blk.h +1
-8@@ -27,7 +27,6 @@ struct blk_flush_queue { };extern struct kmem_cache *blk_requestq_cachep; -extern struct kmem_cache *blk_requestq_srcu_cachep;extern struct kobj_type blk_queue_ktype;extern struct ida blk_queue_ida;@@ -420,13 +419,7 @@ int bio_add_hw_page(struct request_queue *q, struct bio *bio, struct page *page, unsigned int len, unsigned int offset, unsigned int max_sectors, bool *same_page); -static inline struct kmem_cache *blk_get_queue_kmem_cache (bool srcu) -{ -if (srcu) -return blk_requestq_srcu_cachep; -return blk_requestq_cachep; -} -struct request_queue *blk_alloc_queue (int node_id,bool alloc_srcu); +struct request_queue *blk_alloc_queue (int node_id);int disk_scan_partitions (struct gendisk *disk,fmode_t mode);block/genhd.c +1
-1@@ -1410,7 +1410,7 @@ struct gendisk *__blk_alloc_disk(int node, struct lock_class_key *lkclass)struct request_queue *q;struct gendisk *disk; - q =blk_alloc_queue (node, false); + q =blk_alloc_queue (node);if (!q)return NULL ;include/linux/blk-mq.h +3
-0@@ -7,6 +7,7 @@#include <linux/lockdep.h> #include <linux/scatterlist.h> #include <linux/prefetch.h> +#include <linux/srcu.h> struct blk_mq_tags ;struct blk_flush_queue ;@@ -501,6 +502,8 @@ enum hctx_type {* @tag_list_lock: Serializes tag_list accesses. * @tag_list: List of the request queues that use this tag set. See also * request_queue.tag_set_list. +* @srcu: Use as lock when type of the request queue is blocking +* (BLK_MQ_F_BLOCKING). Must be the last member */ struct blk_mq_tag_set {struct blk_mq_queue_map map[HCTX_MAX_TYPES ];
include/linux/blk-mq.h struct blk_mq_tag_set { · line 527
Quoted 4 lines
include/linux/blk-mq.h +1
-0@@ -521,6 +524,7 @@ struct blk_mq_tag_set {struct mutex tag_list_lock;struct list_head tag_list; +struct srcu_struct srcu;
srcu_struct size is more than 50+ KB, it is waste for the tagset which do not set the BLK_MQ_F_BLOCKING, and most tagsets do not set the BLK_MQ_F_BLOCKING. Maybe we can define "srcu" as a pointer, and allocate the memory in blk_mq_alloc_tag_set just for tagset which set the BLK_MQ_F_BLOCKING.
Quoted 43 lines
include/linux/blk-mq.h +0
-0};/** include/linux/blkdev.h +0
-9@@ -22,7 +22,6 @@#include <linux/blkzoned.h> #include <linux/sched.h> #include <linux/sbitmap.h> -#include <linux/srcu.h> #include <linux/uuid.h> #include <linux/xarray.h> @@ -543,18 +542,11 @@ struct request_queue {struct mutex debugfs_mutex;bool mq_sysfs_init_done; - -/** -* @srcu: Sleepable RCU. Use as lock when type of the request queue -* is blocking (BLK_MQ_F_BLOCKING). Must be the last member -*/ -struct srcu_struct srcu[]; };/* Keep blk_queue_flag_name[] in sync with the definitions below */ #define QUEUE_FLAG_STOPPED 0/* queue is stopped */ #define QUEUE_FLAG_DYING 1/* queue being torn down */ -#define QUEUE_FLAG_HAS_SRCU 2/* SRCU is allocated */ #define QUEUE_FLAG_NOMERGES 3/* disable merge attempts */ #define QUEUE_FLAG_SAME_COMP 4/* complete on same CPU-group */ #define QUEUE_FLAG_FAIL_IO 5/* fake timeout */ @@ -590,7 +582,6 @@ bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q);#define blk_queue_stopped (q) test_bit(QUEUE_FLAG_STOPPED, &(q)->queue_flags)#define blk_queue_dying (q) test_bit(QUEUE_FLAG_DYING, &(q)->queue_flags) -#define blk_queue_has_srcu (q) test_bit(QUEUE_FLAG_HAS_SRCU, &(q)->queue_flags)#define blk_queue_init_done (q) test_bit(QUEUE_FLAG_INIT_DONE, &(q)->queue_flags)#define blk_queue_nomerges (q) test_bit(QUEUE_FLAG_NOMERGES, &(q)->queue_flags)#define blk_queue_noxmerges (q) \
Christoph Hellwig
maintainerRe: [PATCH 3/8] blk-mq: move the srcu_struct used for quiescing to the tagset
Quoted 6 lines · Chao Leng
include/linux/blk-mq.h struct blk_mq_tag_set { · line 527
Quoted 2 lines
struct
list_headtag_list;include/linux/blk-mq.h +1
-0+struct srcu_struct srcu;
srcu_structsize is more than 50+ KB, it is waste for the tagset which do not set theBLK_MQ_F_BLOCKING, and most tagsets do not set theBLK_MQ_F_BLOCKING. Maybe we can define "srcu" as a pointer, and allocate the memory inblk_mq_alloc_tag_setjust for tagset which set theBLK_MQ_F_BLOCKING.
I guess we could do that. Still better than the variable sized struct.
p.s. any chance you could properly trim your replies? I took me a while to find your comment in the sea of quotes.
Christoph Hellwig
maintainerPATCH 4/8 blk-mq: pass a tagset to blk_mq_wait_quiesce_done
In reply to Christoph HellwigNoting in blk_mq_wait_quiesce_done needs the request_queue now, so just pass the tagset, and move the non-mq check into the only caller that needs it.
--- block/blk-mq.c | 10 +++++----- drivers/nvme/host/core.c | 4 ++-- drivers/scsi/scsi_lib.c | 2 +- include/linux/blk-mq.h | 2 +- 4 files changed, 9 insertions(+), 9 deletions(-)
block/blk-mq.c +5 -5
@@ -254,15 +254,15 @@ EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);/** * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done -* @q: request queue. +* @set: tag_set to wait on * * Note: it is driver's responsibility for making sure that quiesce has * been started. */ -void blk_mq_wait_quiesce_done (struct request_queue *q) +void blk_mq_wait_quiesce_done (struct blk_mq_tag_set *set) { -if (q->tag_set->flags &BLK_MQ_F_BLOCKING ) -synchronize_srcu (&q->tag_set->srcu); +if (set->flags &BLK_MQ_F_BLOCKING ) +synchronize_srcu (&set->srcu);else synchronize_rcu (); }
@@ -282,7 +282,7 @@ void blk_mq_quiesce_queue(struct request_queue *q)blk_mq_quiesce_queue_nowait (q);/* nothing to wait for non-mq queues */ if (queue_is_mq (q)) -blk_mq_wait_quiesce_done (q); +blk_mq_wait_quiesce_done (q->tag_set); }EXPORT_SYMBOL_GPL (blk_mq_quiesce_queue);
drivers/nvme/host/core.c +2 -2
@@ -5100,7 +5100,7 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)if (!test_and_set_bit (NVME_NS_STOPPED , &ns->flags))blk_mq_quiesce_queue (ns->queue);else -blk_mq_wait_quiesce_done (ns->queue); +blk_mq_wait_quiesce_done (ns->queue->tag_set); }/*
@@ -5216,7 +5216,7 @@ void nvme_stop_admin_queue(struct nvme_ctrl *ctrl)if (!test_and_set_bit (NVME_CTRL_ADMIN_Q_STOPPED , &ctrl->flags))blk_mq_quiesce_queue (ctrl->admin_q);else -blk_mq_wait_quiesce_done (ctrl->admin_q); +blk_mq_wait_quiesce_done (ctrl->admin_q->tag_set); }EXPORT_SYMBOL_GPL (nvme_stop_admin_queue);
drivers/scsi/scsi_lib.c +1 -1
@@ -2735,7 +2735,7 @@ static void scsi_stop_queue(struct scsi_device *sdev, bool nowait)blk_mq_quiesce_queue (sdev->request_queue); }else {if (!nowait) -blk_mq_wait_quiesce_done (sdev->request_queue); +blk_mq_wait_quiesce_done (sdev->request_queue->tag_set); } }
include/linux/blk-mq.h +1 -1
@@ -881,7 +881,7 @@ void blk_mq_start_hw_queues(struct request_queue *q);void blk_mq_start_stopped_hw_queue (struct blk_mq_hw_ctx *hctx,bool async);void blk_mq_start_stopped_hw_queues (struct request_queue *q,bool async);void blk_mq_quiesce_queue (struct request_queue *q); -void blk_mq_wait_quiesce_done (struct request_queue *q); +void blk_mq_wait_quiesce_done (struct blk_mq_tag_set *set);void blk_mq_unquiesce_queue (struct request_queue *q);void blk_mq_delay_run_hw_queue (struct blk_mq_hw_ctx *hctx,unsigned long msecs);void blk_mq_run_hw_queue (struct blk_mq_hw_ctx *hctx,bool async);
Signature · 1 line
2.30.2
Re: [PATCH 4/8] blk-mq: pass a tagset to blk_mq_wait_quiesce_done
Re: [PATCH 4/8] blk-mq: pass a tagset to blk_mq_wait_quiesce_done
In reply to Christoph HellwigQuoted 18 lines · Christoph Hellwig
Noting in
blk_mq_wait_quiesce_doneneeds therequest_queuenow, so just pass the tagset, and move the non-mq check into the only caller that needs it.Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 10 +++++----- drivers/nvme/host/core.c | 4 ++-- drivers/scsi/scsi_lib.c | 2 +- include/linux/blk-mq.h | 2 +- 4 files changed, 9 insertions(+), 9 deletions(-)block/blk-mq.c +0
-0@@ -254,15 +254,15 @@ EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
block/blk-mq.c EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); · line 262
Quoted 10 lines
block/blk-mq.c +2
-2/** * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done -* @q: request queue. +* @set: tag_set to wait on * * Note: it is driver's responsibility for making sure that quiesce has * been started. */ -void blk_mq_wait_quiesce_done (struct request_queue *q) +void blk_mq_wait_quiesce_done (struct blk_mq_tag_set *set)
The change is fine, but the interface could confuse people, it looks like it is waiting for whole tagset quiesced, but it needs to mark all request queues as quiesced first, otherwise it is just wait for one specific queue's quiesce.
So suggest to document such thing.
Thanks,
Ming
Christoph Hellwig
maintainerRe: [PATCH 4/8] blk-mq: pass a tagset to blk_mq_wait_quiesce_done
Quoted 6 lines · Ming Lei
The change is fine, but the interface could confuse people, it looks like it is waiting for whole tagset quiesced, but it needs to mark all request queues as quiesced first, otherwise it is just wait for one specific queue's quiesce.
So suggest to document such thing.
Yes, that's probably a good idea. Still better would be to make this API purely internal, as the pure wait callers in NVMe and SCSI are a bit sketchy.
Re: [PATCH 4/8] blk-mq: pass a tagset to blk_mq_wait_quiesce_done
In reply to Christoph HellwigQuoted 94 lines · Christoph Hellwig
Noting in
blk_mq_wait_quiesce_doneneeds therequest_queuenow, so just pass the tagset, and move the non-mq check into the only caller that needs it.Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 10 +++++----- drivers/nvme/host/core.c | 4 ++-- drivers/scsi/scsi_lib.c | 2 +- include/linux/blk-mq.h | 2 +- 4 files changed, 9 insertions(+), 9 deletions(-)block/blk-mq.c +5
-5@@ -254,15 +254,15 @@ EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);/** * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done -* @q: request queue. +* @set: tag_set to wait on * * Note: it is driver's responsibility for making sure that quiesce has * been started. */ -void blk_mq_wait_quiesce_done (struct request_queue *q) +void blk_mq_wait_quiesce_done (struct blk_mq_tag_set *set) { -if (q->tag_set->flags &BLK_MQ_F_BLOCKING ) -synchronize_srcu (&q->tag_set->srcu); +if (set->flags &BLK_MQ_F_BLOCKING ) +synchronize_srcu (&set->srcu);else synchronize_rcu (); }@@ -282,7 +282,7 @@ void blk_mq_quiesce_queue(struct request_queue *q)blk_mq_quiesce_queue_nowait (q);/* nothing to wait for non-mq queues */ if (queue_is_mq (q)) -blk_mq_wait_quiesce_done (q); +blk_mq_wait_quiesce_done (q->tag_set); }EXPORT_SYMBOL_GPL (blk_mq_quiesce_queue);drivers/nvme/host/core.c +2
-2@@ -5100,7 +5100,7 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)if (!test_and_set_bit (NVME_NS_STOPPED , &ns->flags))blk_mq_quiesce_queue (ns->queue);else -blk_mq_wait_quiesce_done (ns->queue); +blk_mq_wait_quiesce_done (ns->queue->tag_set); }/* @@ -5216,7 +5216,7 @@ void nvme_stop_admin_queue(struct nvme_ctrl *ctrl)if (!test_and_set_bit (NVME_CTRL_ADMIN_Q_STOPPED , &ctrl->flags))blk_mq_quiesce_queue (ctrl->admin_q);else -blk_mq_wait_quiesce_done (ctrl->admin_q); +blk_mq_wait_quiesce_done (ctrl->admin_q->tag_set); }EXPORT_SYMBOL_GPL (nvme_stop_admin_queue);drivers/scsi/scsi_lib.c +1
-1@@ -2735,7 +2735,7 @@ static void scsi_stop_queue(struct scsi_device *sdev, bool nowait)blk_mq_quiesce_queue (sdev->request_queue); }else {if (!nowait) -blk_mq_wait_quiesce_done (sdev->request_queue); +blk_mq_wait_quiesce_done (sdev->request_queue->tag_set); } }include/linux/blk-mq.h +1
-1@@ -881,7 +881,7 @@ void blk_mq_start_hw_queues(struct request_queue *q);void blk_mq_start_stopped_hw_queue (struct blk_mq_hw_ctx *hctx,bool async);void blk_mq_start_stopped_hw_queues (struct request_queue *q,bool async);void blk_mq_quiesce_queue (struct request_queue *q); -void blk_mq_wait_quiesce_done (struct request_queue *q); +void blk_mq_wait_quiesce_done (struct blk_mq_tag_set *set);void blk_mq_unquiesce_queue (struct request_queue *q);void blk_mq_delay_run_hw_queue (struct blk_mq_hw_ctx *hctx,unsigned long msecs);void blk_mq_run_hw_queue (struct blk_mq_hw_ctx *hctx,bool async);
Re: [PATCH 4/8] blk-mq: pass a tagset to blk_mq_wait_quiesce_done
In reply to Christoph HellwigQuoted 4 lines · Christoph Hellwig
Noting in
blk_mq_wait_quiesce_doneneeds therequest_queuenow, so just pass the tagset, and move the non-mq check into the only caller that needs it.
Quoted 8 lines
Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 10 +++++----- drivers/nvme/host/core.c | 4 ++-- drivers/scsi/scsi_lib.c | 2 +- include/linux/blk-mq.h | 2 +- 4 files changed, 9 insertions(+), 9 deletions(-)
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Re: [PATCH 4/8] blk-mq: pass a tagset to blk_mq_wait_quiesce_done
In reply to Christoph HellwigQuoted 1 line · Christoph Hellwig
Noting in
blk_mq_wait_quiesce_doneneeds therequest_queuenow, so just
Noting -> Nothing
Thanks,
Bart.
Christoph Hellwig
maintainerPATCH 5/8 blk-mq: add tagset quiesce interface
In reply to Christoph HellwigFrom: Chao Leng <[email protected]>
Drivers that have shared tagsets may need to quiesce potentially a lot of request queues that all share a single tagset (e.g. nvme). Add an interface to quiesce all the queues on a given tagset. This interface is useful because it can speedup the quiesce by doing it in parallel.
Because some queues should not need to be quiesced(e.g. nvme connect_q) when quiesce the tagset. So introduce QUEUE_FLAG_SKIP_TAGSET_QUIESCE to tagset quiesce interface to skip the queue.
--- block/blk-mq.c | 25 +++++++++++++++++++++++++ include/linux/blk-mq.h | 2 ++ include/linux/blkdev.h | 3 +++ 3 files changed, 30 insertions(+)
block/blk-mq.c +25 -0
@@ -313,6 +313,31 @@ void blk_mq_unquiesce_queue(struct request_queue *q) }EXPORT_SYMBOL_GPL (blk_mq_unquiesce_queue); +void blk_mq_quiesce_tagset (struct blk_mq_tag_set *set) +{ +struct request_queue *q; + +mutex_lock (&set->tag_list_lock); +list_for_each_entry (q, &set->tag_list, tag_set_list) { +if (!blk_queue_skip_tagset_quiesce (q)) +blk_mq_quiesce_queue_nowait (q); + } +blk_mq_wait_quiesce_done (set); +mutex_unlock (&set->tag_list_lock); +} +EXPORT_SYMBOL_GPL (blk_mq_quiesce_tagset); + +void blk_mq_unquiesce_tagset (struct blk_mq_tag_set *set) +{ +struct request_queue *q; + +mutex_lock (&set->tag_list_lock); +list_for_each_entry (q, &set->tag_list, tag_set_list) +blk_mq_unquiesce_queue (q); +mutex_unlock (&set->tag_list_lock); +} +EXPORT_SYMBOL_GPL (blk_mq_unquiesce_tagset); +void blk_mq_wake_waiters (struct request_queue *q) {struct blk_mq_hw_ctx *hctx;
include/linux/blk-mq.h +2 -0
@@ -882,6 +882,8 @@ void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);void blk_mq_start_stopped_hw_queues (struct request_queue *q,bool async);void blk_mq_quiesce_queue (struct request_queue *q);void blk_mq_wait_quiesce_done (struct blk_mq_tag_set *set); +void blk_mq_quiesce_tagset (struct blk_mq_tag_set *set); +void blk_mq_unquiesce_tagset (struct blk_mq_tag_set *set);void blk_mq_unquiesce_queue (struct request_queue *q);void blk_mq_delay_run_hw_queue (struct blk_mq_hw_ctx *hctx,unsigned long msecs);void blk_mq_run_hw_queue (struct blk_mq_hw_ctx *hctx,bool async);
include/linux/blkdev.h +3 -0
@@ -571,6 +571,7 @@ struct request_queue {#define QUEUE_FLAG_HCTX_ACTIVE 28/* at least one blk-mq hctx is active */ #define QUEUE_FLAG_NOWAIT 29/* device supports NOWAIT */ #define QUEUE_FLAG_SQ_SCHED 30/* single queue style io dispatch */ +#define QUEUE_FLAG_SKIP_TAGSET_QUIESCE 31/* quiesce_tagset skip the queue*/ #define QUEUE_FLAG_MQ_DEFAULT ((1UL << QUEUE_FLAG_IO_STAT) | \ (1UL << QUEUE_FLAG_SAME_COMP) | \
@@ -610,6 +611,8 @@ bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q);#define blk_queue_pm_only (q) atomic_read(&(q)->pm_only)#define blk_queue_registered (q) test_bit(QUEUE_FLAG_REGISTERED, &(q)->queue_flags)#define blk_queue_sq_sched (q) test_bit(QUEUE_FLAG_SQ_SCHED, &(q)->queue_flags) +#define blk_queue_skip_tagset_quiesce (q) \ + test_bit(QUEUE_FLAG_SKIP_TAGSET_QUIESCE, &(q)->queue_flags)extern void blk_set_pm_only (struct request_queue *q);extern void blk_clear_pm_only (struct request_queue *q);
Signature · 1 line
2.30.2
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
In reply to Christoph HellwigQuoted 7 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
Drivers that have shared tagsets may need to quiesce potentially a lot of request queues that all share a single tagset (e.g. nvme). Add an interface to quiesce all the queues on a given tagset. This interface is useful because it can speedup the quiesce by doing it in parallel.
Quoted 7 lines
Because some queues should not need to be quiesced(e.g. nvme
connect_q) when quiesce the tagset. So introduceQUEUE_FLAG_SKIP_TAGSET_QUIESCEto tagset quiesce interface to skip the queue.Signed-off-by Chao Leng <[email protected]>[hch: simplify for the per-tag_set srcu_struct]Signed-off-by Christoph Hellwig <[email protected]>
Thanks,
Ming
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
In reply to Christoph HellwigQuoted 91 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
Drivers that have shared tagsets may need to quiesce potentially a lot of request queues that all share a single tagset (e.g. nvme). Add an interface to quiesce all the queues on a given tagset. This interface is useful because it can speedup the quiesce by doing it in parallel.
Because some queues should not need to be quiesced(e.g. nvme
connect_q) when quiesce the tagset. So introduceQUEUE_FLAG_SKIP_TAGSET_QUIESCEto tagset quiesce interface to skip the queue.Signed-off-by Chao Leng <[email protected]>[hch: simplify for the per-tag_set srcu_struct]Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 25 +++++++++++++++++++++++++ include/linux/blk-mq.h | 2 ++ include/linux/blkdev.h | 3 +++ 3 files changed, 30 insertions(+)block/blk-mq.c +25
-0@@ -313,6 +313,31 @@ void blk_mq_unquiesce_queue(struct request_queue *q) }EXPORT_SYMBOL_GPL (blk_mq_unquiesce_queue); +void blk_mq_quiesce_tagset (struct blk_mq_tag_set *set) +{ +struct request_queue *q; + +mutex_lock (&set->tag_list_lock); +list_for_each_entry (q, &set->tag_list, tag_set_list) { +if (!blk_queue_skip_tagset_quiesce (q)) +blk_mq_quiesce_queue_nowait (q); + } +blk_mq_wait_quiesce_done (set); +mutex_unlock (&set->tag_list_lock); +} +EXPORT_SYMBOL_GPL (blk_mq_quiesce_tagset); + +void blk_mq_unquiesce_tagset (struct blk_mq_tag_set *set) +{ +struct request_queue *q; + +mutex_lock (&set->tag_list_lock); +list_for_each_entry (q, &set->tag_list, tag_set_list) +blk_mq_unquiesce_queue (q); +mutex_unlock (&set->tag_list_lock); +} +EXPORT_SYMBOL_GPL (blk_mq_unquiesce_tagset); +void blk_mq_wake_waiters (struct request_queue *q) {struct blk_mq_hw_ctx *hctx;include/linux/blk-mq.h +2
-0@@ -882,6 +882,8 @@ void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);void blk_mq_start_stopped_hw_queues (struct request_queue *q,bool async);void blk_mq_quiesce_queue (struct request_queue *q);void blk_mq_wait_quiesce_done (struct blk_mq_tag_set *set); +void blk_mq_quiesce_tagset (struct blk_mq_tag_set *set); +void blk_mq_unquiesce_tagset (struct blk_mq_tag_set *set);void blk_mq_unquiesce_queue (struct request_queue *q);void blk_mq_delay_run_hw_queue (struct blk_mq_hw_ctx *hctx,unsigned long msecs);void blk_mq_run_hw_queue (struct blk_mq_hw_ctx *hctx,bool async);include/linux/blkdev.h +3
-0@@ -571,6 +571,7 @@ struct request_queue {#define QUEUE_FLAG_HCTX_ACTIVE 28/* at least one blk-mq hctx is active */ #define QUEUE_FLAG_NOWAIT 29/* device supports NOWAIT */ #define QUEUE_FLAG_SQ_SCHED 30/* single queue style io dispatch */ +#define QUEUE_FLAG_SKIP_TAGSET_QUIESCE 31/* quiesce_tagset skip the queue*/ #define QUEUE_FLAG_MQ_DEFAULT ((1UL << QUEUE_FLAG_IO_STAT) | \ (1UL << QUEUE_FLAG_SAME_COMP) | \@@ -610,6 +611,8 @@ bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q);#define blk_queue_pm_only (q) atomic_read(&(q)->pm_only)#define blk_queue_registered (q) test_bit(QUEUE_FLAG_REGISTERED, &(q)->queue_flags)#define blk_queue_sq_sched (q) test_bit(QUEUE_FLAG_SQ_SCHED, &(q)->queue_flags) +#define blk_queue_skip_tagset_quiesce (q) \ + test_bit(QUEUE_FLAG_SKIP_TAGSET_QUIESCE, &(q)->queue_flags)extern void blk_set_pm_only (struct request_queue *q);extern void blk_clear_pm_only (struct request_queue *q);
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
In reply to Christoph HellwigQuoted 11 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
Drivers that have shared tagsets may need to quiesce potentially a lot of request queues that all share a single tagset (e.g. nvme). Add an interface to quiesce all the queues on a given tagset. This interface is useful because it can speedup the quiesce by doing it in parallel.
Because some queues should not need to be quiesced(e.g. nvme
connect_q) when quiesce the tagset. So introduceQUEUE_FLAG_SKIP_TAGSET_QUIESCEto tagset quiesce interface to skip the queue.
Quoted 9 lines
Signed-off-by Chao Leng <[email protected]>[hch: simplify for the per-tag_set srcu_struct]Signed-off-by Christoph Hellwig <[email protected]>--- block/blk-mq.c | 25 +++++++++++++++++++++++++ include/linux/blk-mq.h | 2 ++ include/linux/blkdev.h | 3 +++ 3 files changed, 30 insertions(+)
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
In reply to Christoph HellwigQuoted 8 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
Drivers that have shared tagsets may need to quiesce potentially a lot of request queues that all share a single tagset (e.g. nvme). Add an interface to quiesce all the queues on a given tagset. This interface is useful because it can speedup the quiesce by doing it in parallel.
Because some queues should not need to be quiesced(e.g. nvme
connect_q)
should not need -> do not have
Quoted 1 line
when quiesce the tagset. So introduce
QUEUE_FLAG_SKIP_TAGSET_QUIESCEto
quiesce -> quiescing
. So -> ,
Quoted 1 line
tagset quiesce interface to skip the queue.
Hi Christoph,
Are there any drivers that may call blk_mq_quiesce_tagset() concurrently from two different threads? blk_mq_quiesce_queue() supports this but blk_mq_quiesce_tagset() cannot support this because of the mechanism used to mark request queues that should be skipped (queue flag).
Thanks,
Bart.
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
Quoted 16 lines · Bart Van Assche
Quoted 8 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
Drivers that have shared tagsets may need to quiesce potentially a lot of request queues that all share a single tagset (e.g. nvme). Add an interface to quiesce all the queues on a given tagset. This interface is useful because it can speedup the quiesce by doing it in parallel.
Because some queues should not need to be quiesced(e.g. nvme
connect_q)should not need -> do not have
Quoted 1 line
when quiesce the tagset. So introduce
QUEUE_FLAG_SKIP_TAGSET_QUIESCEtoquiesce -> quiescing
Quoted 8 lines
. So -> ,
Quoted 1 line
tagset quiesce interface to skip the queue.
Hi Christoph,
Are there any drivers that may call
blk_mq_quiesce_tagset()concurrently from two different threads?blk_mq_quiesce_queue()supports this butblk_mq_quiesce_tagset()cannot support this because of the mechanism used to mark request queues that should be skipped (queue flag).
blk_mq_quiesce_tagset() support concurrency. blk_mq_quiesce_tagset() just check the flag(QUEUE_FLAG_SKIP_TAGSET_QUIESCE), it has no impact on concurrency.
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
Quoted 2 lines · Chao Leng
blk_mq_quiesce_tagset()support concurrency.blk_mq_quiesce_tagset()just check theflag(QUEUE_FLAG_SKIP_TAGSET_QUIESCE), it has no impact on concurrency.
Hi Chao,
I think it depends on how the QUEUE_FLAG_SKIP_TAGSET_QUIESCE flag is set. I agree if that flag is set once and never modified that there is no race. What I'm wondering about is whether there could be a need for block drivers to set the QUEUE_FLAG_SKIP_TAGSET_QUIESCE flag just before blk_mq_quiesce_tagset() is called and cleared immediately after blk_mq_quiesce_tagset() returns? In that case I think there is a race condition.
Thanks,
Bart.
Re: [PATCH 5/8] blk-mq: add tagset quiesce interface
Quoted 8 lines · Bart Van Assche
Quoted 2 lines · Chao Leng
blk_mq_quiesce_tagset()support concurrency.blk_mq_quiesce_tagset()just check theflag(QUEUE_FLAG_SKIP_TAGSET_QUIESCE), it has no impact on concurrency.Hi Chao,
I think it depends on how the
QUEUE_FLAG_SKIP_TAGSET_QUIESCEflag is set. I agree if that flag is set once and never modified that there is no race. What I'm wondering about is whether there could be a need for block drivers to set theQUEUE_FLAG_SKIP_TAGSET_QUIESCEflag just beforeblk_mq_quiesce_tagset()is called and cleared immediately afterblk_mq_quiesce_tagset()returns? In that case I think there is a race condition.
Even if dynamically modify the QUEUE_FLAG_SKIP_TAGSET_QUIESCE, it still has no impact on concurrency of blk_mq_quiesce_tagset(). It may affect atomicity of blk_mq_quiesce_tagset(), blk_mq_quiesce_tagset() do not ensure atomicity, the caller should ensure it if needed.
Christoph Hellwig
maintainerPATCH 6/8 nvme: move the NS_DEAD flag to the controller
In reply to Christoph HellwigThe NVME_NS_DEAD flag is only set in nvme_set_queue_dying, which is called in a loop over all namespaces in nvme_kill_queues. Switch it to a controller flag checked and set outside said loop.
--- drivers/nvme/host/core.c | 16 +++++++--------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 8 insertions(+), 10 deletions(-)
drivers/nvme/host/core.c +7 -9
@@ -4330,7 +4330,7 @@ static void nvme_validate_ns(struct nvme_ns *ns, struct nvme_ns_info *info) {int ret =NVME_SC_INVALID_NS |NVME_SC_DNR ; -if (test_bit (NVME_NS_DEAD , &ns->flags)) +if (test_bit (NVME_CTRL_NS_DEAD , &ns->ctrl->flags)) goto out; ret =NVME_SC_INVALID_NS |NVME_SC_DNR ;
@@ -4404,7 +4404,8 @@ static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,down_write (&ctrl->namespaces_rwsem);list_for_each_entry_safe (ns, next, &ctrl->namespaces, list) { -if (ns->head->ns_id > nsid ||test_bit (NVME_NS_DEAD , &ns->flags)) +if (ns->head->ns_id > nsid || +test_bit (NVME_CTRL_NS_DEAD , &ns->ctrl->flags))list_move_tail (&ns->list, &rm_list); }up_write (&ctrl->namespaces_rwsem);
@@ -5110,9 +5111,6 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)*/ static void nvme_set_queue_dying (struct nvme_ns *ns) { -if (test_and_set_bit (NVME_NS_DEAD , &ns->flags)) -return ; -blk_mark_disk_dead (ns->disk);nvme_start_ns_queue (ns); }
@@ -5129,14 +5127,14 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)struct nvme_ns *ns;down_read (&ctrl->namespaces_rwsem); -/* Forcibly unquiesce queues to avoid blocking dispatch */ if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_set_queue_dying (ns); - +if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { +list_for_each_entry (ns, &ctrl->namespaces, list) +nvme_set_queue_dying (ns); + }up_read (&ctrl->namespaces_rwsem); }EXPORT_SYMBOL_GPL (nvme_kill_queues);
drivers/nvme/host/nvme.h +1 -1
@@ -237,6 +237,7 @@ enum nvme_ctrl_flags {NVME_CTRL_FAILFAST_EXPIRED =0 ,NVME_CTRL_ADMIN_Q_STOPPED =1 ,NVME_CTRL_STARTED_ONCE =2 , + NVME_CTRL_NS_DEAD = 3, }; struct nvme_ctrl {
@@ -483,7 +484,6 @@ struct nvme_ns {unsigned long features;unsigned long flags;#define NVME_NS_REMOVING 0 -#define NVME_NS_DEAD 1#define NVME_NS_ANA_PENDING 2#define NVME_NS_FORCE_RO 3#define NVME_NS_READY 4
Signature · 1 line
2.30.2
Sagi Grimberg
maintainerRe: [PATCH 6/8] nvme: move the NS_DEAD flag to the controller
Quoted 49 lines · Christoph Hellwig
The
NVME_NS_DEADflag is only set innvme_set_queue_dying, which is called in a loop over all namespaces innvme_kill_queues. Switch it to a controller flag checked and set outside said loop.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 16 +++++++--------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 8 insertions(+), 10 deletions(-)drivers/nvme/host/core.c +3
-6@@ -4330,7 +4330,7 @@ static void nvme_validate_ns(struct nvme_ns *ns, struct nvme_ns_info *info) {int ret =NVME_SC_INVALID_NS |NVME_SC_DNR ; -if (test_bit (NVME_NS_DEAD , &ns->flags)) +if (test_bit (NVME_CTRL_NS_DEAD , &ns->ctrl->flags)) goto out; ret =NVME_SC_INVALID_NS |NVME_SC_DNR ;@@ -4404,7 +4404,8 @@ static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,down_write (&ctrl->namespaces_rwsem);list_for_each_entry_safe (ns, next, &ctrl->namespaces, list) { -if (ns->head->ns_id > nsid ||test_bit (NVME_NS_DEAD , &ns->flags)) +if (ns->head->ns_id > nsid || +test_bit (NVME_CTRL_NS_DEAD , &ns->ctrl->flags))list_move_tail (&ns->list, &rm_list); }up_write (&ctrl->namespaces_rwsem);@@ -5110,9 +5111,6 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)*/ static void nvme_set_queue_dying (struct nvme_ns *ns) { -if (test_and_set_bit (NVME_NS_DEAD , &ns->flags)) -return ; -blk_mark_disk_dead (ns->disk);nvme_start_ns_queue (ns); }@@ -5129,14 +5127,14 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)struct nvme_ns *ns;down_read (&ctrl->namespaces_rwsem); -/* Forcibly unquiesce queues to avoid blocking dispatch */
drivers/nvme/host/core.c void nvme_kill_queues(struct nvme_ctrl *ctrl) · line 5137
Quoted 10 lines
drivers/nvme/host/core.c +4
-3if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_set_queue_dying (ns); - +if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { +list_for_each_entry (ns, &ctrl->namespaces, list) +nvme_set_queue_dying (ns); + }
Looking at it now, I'm not sure I understand the need for this flag. It seems to make nvme_kill_queues reentrant safe, but the admin queue unquiesce can still end up unbalanced under reentrance?
How is this not broken today (or ever since quiesce/unquiesce started accounting)? Maybe I lost some context on the exact subtlety of how nvme-pci uses this interface...
Quoted 23 lines
drivers/nvme/host/core.c +0
-0up_read (&ctrl->namespaces_rwsem); }EXPORT_SYMBOL_GPL (nvme_kill_queues);drivers/nvme/host/nvme.h +1
-1@@ -237,6 +237,7 @@ enum nvme_ctrl_flags {NVME_CTRL_FAILFAST_EXPIRED =0 ,NVME_CTRL_ADMIN_Q_STOPPED =1 ,NVME_CTRL_STARTED_ONCE =2 , + NVME_CTRL_NS_DEAD = 3, }; struct nvme_ctrl {@@ -483,7 +484,6 @@ struct nvme_ns {unsigned long features;unsigned long flags;#define NVME_NS_REMOVING 0 -#define NVME_NS_DEAD 1#define NVME_NS_ANA_PENDING 2#define NVME_NS_FORCE_RO 3#define NVME_NS_READY 4
Christoph Hellwig
maintainerRe: [PATCH 6/8] nvme: move the NS_DEAD flag to the controller
Quoted 10 lines · Sagi Grimberg
drivers/nvme/host/core.c void nvme_kill_queues(struct nvme_ctrl *ctrl) · line 5137
Quoted 5 lines
drivers/nvme/host/core.c +4
-1- +if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { +list_for_each_entry (ns, &ctrl->namespaces, list) +nvme_set_queue_dying (ns); + }Looking at it now, I'm not sure I understand the need for this flag. It seems to make
nvme_kill_queuesreentrant safe, but the admin queue unquiesce can still end up unbalanced under reentrance?
Quoted 3 lines
How is this not broken today (or ever since quiesce/unquiesce started accounting)? Maybe I lost some context on the exact subtlety of how nvme-pci uses this interface...
Yes, this also looks weird and I had a TODO list entry for myself to look into what is going on here. The whole interaction with nvme_remove_namespaces is pretty weird to start with, and then the code in PCIe is even more weird. But to feel confident to touch this I'd need real hot removal testing, for which I don't have a good rig right now.
Sagi Grimberg
maintainerRe: [PATCH 6/8] nvme: move the NS_DEAD flag to the controller
Quoted 15 lines
Quoted 11 lines
drivers/nvme/host/core.c void nvme_kill_queues(struct nvme_ctrl *ctrl) · line 5137
Quoted 6 lines
drivers/nvme/host/core.c +4
-1- +if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { +list_for_each_entry (ns, &ctrl->namespaces, list) +nvme_set_queue_dying (ns); + }Looking at it now, I'm not sure I understand the need for this flag. It seems to make
nvme_kill_queuesreentrant safe, but the admin queue unquiesce can still end up unbalanced under reentrance?Quoted 3 lines
How is this not broken today (or ever since quiesce/unquiesce started accounting)? Maybe I lost some context on the exact subtlety of how nvme-pci uses this interface...
Quoted 6 lines
Yes, this also looks weird and I had a TODO list entry for myself to look into what is going on here. The whole interaction with
nvme_remove_namespacesis pretty weird to start with, and then the code in PCIe is even more weird. But to feel confident to touch this I'd need real hot removal testing, for which I don't have a good rig right now.
Lets for start move the bit check up in the function and reverse the polarity to return if it is set. Unless someone can make sense of why this is OK.
Sagi Grimberg
maintainerRe: [PATCH 6/8] nvme: move the NS_DEAD flag to the controller
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 6 lines
-
+ if (!test_and_set_bit(NVME_CTRL_NS_DEAD,&ctrl->flags)) { +list_for_each_entry(ns, &ctrl->namespaces, list)+nvme_set_queue_dying(ns);
+ }Looking at it now, I'm not sure I understand the need for this flag. It seems to make
nvme_kill_queuesreentrant safe, but the admin queue unquiesce can still end up unbalanced under reentrance?
Quoted 7 lines
Quoted 7 lines
Quoted 3 lines
How is this not broken today (or ever since quiesce/unquiesce started accounting)? Maybe I lost some context on the exact subtlety of how nvme-pci uses this interface...
Yes, this also looks weird and I had a TODO list entry for myself to look into what is going on here. The whole interaction with
nvme_remove_namespacesis pretty weird to start with,
git blame points to:0ff9d4e1a284 ("NVMe: Short-cut removal on surprise hot-unplug")
But it is unclear as to why it sits in nvme_remove_namespaces instead of somewhere before that in nvme_remove()...
Keith?
Re: [PATCH 6/8] nvme: move the NS_DEAD flag to the controller
In reply to Christoph HellwigQuoted 83 lines · Christoph Hellwig
The
NVME_NS_DEADflag is only set innvme_set_queue_dying, which is called in a loop over all namespaces innvme_kill_queues. Switch it to a controller flag checked and set outside said loop.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 16 +++++++--------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 8 insertions(+), 10 deletions(-)drivers/nvme/host/core.c +7
-9@@ -4330,7 +4330,7 @@ static void nvme_validate_ns(struct nvme_ns *ns, struct nvme_ns_info *info) {int ret =NVME_SC_INVALID_NS |NVME_SC_DNR ; -if (test_bit (NVME_NS_DEAD , &ns->flags)) +if (test_bit (NVME_CTRL_NS_DEAD , &ns->ctrl->flags)) goto out; ret =NVME_SC_INVALID_NS |NVME_SC_DNR ;@@ -4404,7 +4404,8 @@ static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,down_write (&ctrl->namespaces_rwsem);list_for_each_entry_safe (ns, next, &ctrl->namespaces, list) { -if (ns->head->ns_id > nsid ||test_bit (NVME_NS_DEAD , &ns->flags)) +if (ns->head->ns_id > nsid || +test_bit (NVME_CTRL_NS_DEAD , &ns->ctrl->flags))list_move_tail (&ns->list, &rm_list); }up_write (&ctrl->namespaces_rwsem);@@ -5110,9 +5111,6 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)*/ static void nvme_set_queue_dying (struct nvme_ns *ns) { -if (test_and_set_bit (NVME_NS_DEAD , &ns->flags)) -return ; -blk_mark_disk_dead (ns->disk);nvme_start_ns_queue (ns); }@@ -5129,14 +5127,14 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)struct nvme_ns *ns;down_read (&ctrl->namespaces_rwsem); -/* Forcibly unquiesce queues to avoid blocking dispatch */ if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_set_queue_dying (ns); - +if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { +list_for_each_entry (ns, &ctrl->namespaces, list) +nvme_set_queue_dying (ns); + }up_read (&ctrl->namespaces_rwsem); }EXPORT_SYMBOL_GPL (nvme_kill_queues);drivers/nvme/host/nvme.h +1
-1@@ -237,6 +237,7 @@ enum nvme_ctrl_flags {NVME_CTRL_FAILFAST_EXPIRED =0 ,NVME_CTRL_ADMIN_Q_STOPPED =1 ,NVME_CTRL_STARTED_ONCE =2 , + NVME_CTRL_NS_DEAD = 3, }; struct nvme_ctrl {@@ -483,7 +484,6 @@ struct nvme_ns {unsigned long features;unsigned long flags;#define NVME_NS_REMOVING 0 -#define NVME_NS_DEAD 1#define NVME_NS_ANA_PENDING 2#define NVME_NS_FORCE_RO 3#define NVME_NS_READY 4
Re: [PATCH 6/8] nvme: move the NS_DEAD flag to the controller
In reply to Christoph HellwigQuoted 10 lines · Christoph Hellwig
The
NVME_NS_DEADflag is only set innvme_set_queue_dying, which is called in a loop over all namespaces innvme_kill_queues. Switch it to a controller flag checked and set outside said loop.Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 16 +++++++--------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 8 insertions(+), 10 deletions(-)
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Christoph Hellwig
maintainerPATCH 7/8 nvme: remove nvme_set_queue_dying
In reply to Christoph HellwigThis helper is pretty pointless now, and also in the way of per-tagset quiesce.
--- drivers/nvme/host/core.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
drivers/nvme/host/core.c +4 -14
@@ -5104,17 +5104,6 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)blk_mq_wait_quiesce_done (ns->queue->tag_set); } -/* -* Prepare a queue for teardown. -* -* This must forcibly unquiesce queues to avoid blocking dispatch. -*/ -static void nvme_set_queue_dying (struct nvme_ns *ns) -{ -blk_mark_disk_dead (ns->disk); -nvme_start_ns_queue (ns); -} -/** * nvme_kill_queues(): Ends all namespace queues * @ctrl: the dead controller that needs to end
@@ -5130,10 +5119,11 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)/* Forcibly unquiesce queues to avoid blocking dispatch */ if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl); -if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_set_queue_dying (ns); +list_for_each_entry (ns, &ctrl->namespaces, list) { +blk_mark_disk_dead (ns->disk); +nvme_start_ns_queue (ns); + } }up_read (&ctrl->namespaces_rwsem); }
Signature · 1 line
2.30.2
Sagi Grimberg
maintainerRe: [PATCH 7/8] nvme: remove nvme_set_queue_dying
Quoted 33 lines
This helper is pretty pointless now, and also in the way of per-tagset quiesce.
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)drivers/nvme/host/core.c +0
-11@@ -5104,17 +5104,6 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)blk_mq_wait_quiesce_done (ns->queue->tag_set); } -/* -* Prepare a queue for teardown. -* -* This must forcibly unquiesce queues to avoid blocking dispatch. -*/ -static void nvme_set_queue_dying (struct nvme_ns *ns) -{ -blk_mark_disk_dead (ns->disk); -nvme_start_ns_queue (ns); -} -/** * nvme_kill_queues(): Ends all namespace queues * @ctrl: the dead controller that needs to end @@ -5130,10 +5119,11 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)/* Forcibly unquiesce queues to avoid blocking dispatch */
drivers/nvme/host/core.c void nvme_kill_queues(struct nvme_ctrl *ctrl) · line 5126
Quoted 10 lines
drivers/nvme/host/core.c +4
-3if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl); -if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_set_queue_dying (ns); +list_for_each_entry (ns, &ctrl->namespaces, list) { +blk_mark_disk_dead (ns->disk); +nvme_start_ns_queue (ns); + }
I have to say that I always found nvme_kill_queues interface somewhat odd. its a core function that unquiesces the admin/io queues assuming that they were stopped at some point by the driver.
If now there is no dependency between unquiesce and blk_mark_disk_dead, maybe it would be a good idea to move the unquiescing to the drivers which can pair with the quiesce itself, and rename it to nvme_mark_namespaces_dead() or something?
Christoph Hellwig
maintainerRe: [PATCH 7/8] nvme: remove nvme_set_queue_dying
Quoted 2 lines · Sagi Grimberg
I have to say that I always found
nvme_kill_queuesinterface somewhat odd.
It is..
Quoted 7 lines
its a core function that unquiesces the admin/io queues assuming that they were stopped at some point by the driver.
If now there is no dependency between unquiesce and
blk_mark_disk_dead, maybe it would be a good idea to move the unquiescing to the drivers which can pair with the quiesce itself, and rename it tonvme_mark_namespaces_dead()or something?
Yes, something like that is probably a good idea. This entire area is a bit of a mess to say it mildly, so maybe we need to sort that out first.
Sagi Grimberg
maintainerRe: [PATCH 7/8] nvme: remove nvme_set_queue_dying
Quoted 9 lines
Quoted 3 lines
I have to say that I always found
nvme_kill_queuesinterface somewhat odd.It is..
Quoted 3 lines
its a core function that unquiesces the admin/io queues assuming that they were stopped at some point by the driver.
Quoted 8 lines
Quoted 4 lines
If now there is no dependency between unquiesce and
blk_mark_disk_dead, maybe it would be a good idea to move the unquiescing to the drivers which can pair with the quiesce itself, and rename it tonvme_mark_namespaces_dead()or something?Yes, something like that is probably a good idea. This entire area is a bit of a mess to say it mildly, so maybe we need to sort that out first.
I agree.
Re: [PATCH 7/8] nvme: remove nvme_set_queue_dying
In reply to Christoph HellwigQuoted 46 lines · Christoph Hellwig
This helper is pretty pointless now, and also in the way of per-tagset quiesce.
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)drivers/nvme/host/core.c +4
-14@@ -5104,17 +5104,6 @@ static void nvme_stop_ns_queue(struct nvme_ns *ns)blk_mq_wait_quiesce_done (ns->queue->tag_set); } -/* -* Prepare a queue for teardown. -* -* This must forcibly unquiesce queues to avoid blocking dispatch. -*/ -static void nvme_set_queue_dying (struct nvme_ns *ns) -{ -blk_mark_disk_dead (ns->disk); -nvme_start_ns_queue (ns); -} -/** * nvme_kill_queues(): Ends all namespace queues * @ctrl: the dead controller that needs to end @@ -5130,10 +5119,11 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)/* Forcibly unquiesce queues to avoid blocking dispatch */ if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl); -if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_set_queue_dying (ns); +list_for_each_entry (ns, &ctrl->namespaces, list) { +blk_mark_disk_dead (ns->disk); +nvme_start_ns_queue (ns); + } }up_read (&ctrl->namespaces_rwsem); }
Re: [PATCH 7/8] nvme: remove nvme_set_queue_dying
In reply to Christoph HellwigQuoted 8 lines · Christoph Hellwig
This helper is pretty pointless now, and also in the way of per-tagset quiesce.
Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-)
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Christoph Hellwig
maintainerPATCH 8/8 nvme: use blk_mq_[un]quiesce_tagset
In reply to Christoph HellwigFrom: Chao Leng <[email protected]>
All controller namespaces share the same tagset, so we can use this interface which does the optimal operation for parallel quiesce based on the tagset type(e.g. blocking tagsets and non-blocking tagsets).
nvme connect_q should not be quiesced when quiesce tagset, so set the QUEUE_FLAG_SKIP_TAGSET_QUIESCE to skip it when init connect_q.
Currently we use NVME_NS_STOPPED to ensure pairing quiescing and unquiescing. If use blk_mq_[un]quiesce_tagset, NVME_NS_STOPPED will be invalided, so introduce NVME_CTRL_STOPPED to replace NVME_NS_STOPPED. In addition, we never really quiesce a single namespace. It is a better choice to move the flag from ns to ctrl.
--- drivers/nvme/host/core.c | 38 +++++++++----------------------------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 10 insertions(+), 30 deletions(-)
drivers/nvme/host/core.c +9 -29
@@ -4891,6 +4891,7 @@ int nvme_alloc_io_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set, ret =PTR_ERR (ctrl->connect_q); goto out_free_tag_set; } +blk_queue_flag_set (QUEUE_FLAG_SKIP_TAGSET_QUIESCE , ctrl->connect_q); } ctrl->tagset = set;
@@ -5090,20 +5091,6 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, }EXPORT_SYMBOL_GPL (nvme_init_ctrl); -static void nvme_start_ns_queue (struct nvme_ns *ns) -{ -if (test_and_clear_bit (NVME_NS_STOPPED , &ns->flags)) -blk_mq_unquiesce_queue (ns->queue); -} - -static void nvme_stop_ns_queue (struct nvme_ns *ns) -{ -if (!test_and_set_bit (NVME_NS_STOPPED , &ns->flags)) -blk_mq_quiesce_queue (ns->queue); -else -blk_mq_wait_quiesce_done (ns->queue->tag_set); -} -/** * nvme_kill_queues(): Ends all namespace queues * @ctrl: the dead controller that needs to end
@@ -5120,10 +5107,9 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl);if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { -list_for_each_entry (ns, &ctrl->namespaces, list) { +list_for_each_entry (ns, &ctrl->namespaces, list)blk_mark_disk_dead (ns->disk); -nvme_start_ns_queue (ns); - } +nvme_start_queues (ctrl); }up_read (&ctrl->namespaces_rwsem); }
@@ -5179,23 +5165,17 @@ EXPORT_SYMBOL_GPL(nvme_start_freeze);void nvme_stop_queues (struct nvme_ctrl *ctrl) { -struct nvme_ns *ns; - -down_read (&ctrl->namespaces_rwsem); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_stop_ns_queue (ns); -up_read (&ctrl->namespaces_rwsem); +if (!test_and_set_bit (NVME_CTRL_STOPPED , &ctrl->flags)) +blk_mq_quiesce_tagset (ctrl->tagset); +else +blk_mq_wait_quiesce_done (ctrl->tagset); }EXPORT_SYMBOL_GPL (nvme_stop_queues);void nvme_start_queues (struct nvme_ctrl *ctrl) { -struct nvme_ns *ns; - -down_read (&ctrl->namespaces_rwsem); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_start_ns_queue (ns); -up_read (&ctrl->namespaces_rwsem); +if (test_and_clear_bit (NVME_CTRL_STOPPED , &ctrl->flags)) +blk_mq_unquiesce_tagset (ctrl->tagset); }EXPORT_SYMBOL_GPL (nvme_start_queues);
drivers/nvme/host/nvme.h +1 -1
@@ -238,6 +238,7 @@ enum nvme_ctrl_flags {NVME_CTRL_ADMIN_Q_STOPPED =1 ,NVME_CTRL_STARTED_ONCE =2 ,NVME_CTRL_NS_DEAD =3 , + NVME_CTRL_STOPPED = 4, }; struct nvme_ctrl {
@@ -487,7 +488,6 @@ struct nvme_ns {#define NVME_NS_ANA_PENDING 2#define NVME_NS_FORCE_RO 3#define NVME_NS_READY 4 -#define NVME_NS_STOPPED 5struct cdev cdev;struct device cdev_device;
Signature · 1 line
2.30.2
Sagi Grimberg
maintainerRe: [PATCH 8/8] nvme: use blk_mq_[un]quiesce_tagset
Quoted 73 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
All controller namespaces share the same tagset, so we can use this interface which does the optimal operation for parallel quiesce based on the tagset type(e.g. blocking tagsets and non-blocking tagsets).
nvme
connect_qshould not be quiesced when quiesce tagset, so set theQUEUE_FLAG_SKIP_TAGSET_QUIESCEto skip it when initconnect_q.Currently we use
NVME_NS_STOPPEDto ensure pairing quiescing and unquiescing. If use blk_mq_[un]quiesce_tagset,NVME_NS_STOPPEDwill be invalided, so introduceNVME_CTRL_STOPPEDto replaceNVME_NS_STOPPED. In addition, we never really quiesce a single namespace. It is a better choice to move the flag from ns to ctrl.Signed-off-by Chao Leng <[email protected]>[hch: rebased on top of prep patches]Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 38 +++++++++----------------------------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 10 insertions(+), 30 deletions(-)drivers/nvme/host/core.c +3
-17@@ -4891,6 +4891,7 @@ int nvme_alloc_io_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set, ret =PTR_ERR (ctrl->connect_q); goto out_free_tag_set; } +blk_queue_flag_set (QUEUE_FLAG_SKIP_TAGSET_QUIESCE , ctrl->connect_q); } ctrl->tagset = set;@@ -5090,20 +5091,6 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, }EXPORT_SYMBOL_GPL (nvme_init_ctrl); -static void nvme_start_ns_queue (struct nvme_ns *ns) -{ -if (test_and_clear_bit (NVME_NS_STOPPED , &ns->flags)) -blk_mq_unquiesce_queue (ns->queue); -} - -static void nvme_stop_ns_queue (struct nvme_ns *ns) -{ -if (!test_and_set_bit (NVME_NS_STOPPED , &ns->flags)) -blk_mq_quiesce_queue (ns->queue); -else -blk_mq_wait_quiesce_done (ns->queue->tag_set); -} -/** * nvme_kill_queues(): Ends all namespace queues * @ctrl: the dead controller that needs to end @@ -5120,10 +5107,9 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl);if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { -list_for_each_entry (ns, &ctrl->namespaces, list) { +list_for_each_entry (ns, &ctrl->namespaces, list)blk_mark_disk_dead (ns->disk); -nvme_start_ns_queue (ns); - } +nvme_start_queues (ctrl); }up_read (&ctrl->namespaces_rwsem); }@@ -5179,23 +5165,17 @@ EXPORT_SYMBOL_GPL(nvme_start_freeze);void nvme_stop_queues (struct nvme_ctrl *ctrl) {
drivers/nvme/host/core.c EXPORT_SYMBOL_GPL(nvme_start_freeze); · line 5171
Quoted 10 lines
drivers/nvme/host/core.c +4
-6-struct nvme_ns *ns; - -down_read (&ctrl->namespaces_rwsem); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_stop_ns_queue (ns); -up_read (&ctrl->namespaces_rwsem); +if (!test_and_set_bit (NVME_CTRL_STOPPED , &ctrl->flags)) +blk_mq_quiesce_tagset (ctrl->tagset); +else +blk_mq_wait_quiesce_done (ctrl->tagset);
Isn't blk_mq_quiesce_tagset already waits for the (s)rcu grace? Is this to make a concurrent caller also wait?
I wish we would sort out all the concurrency issues in the driver(s) instead of making core functions reentrant safe...
Re: [PATCH 8/8] nvme: use blk_mq_[un]quiesce_tagset
In reply to Christoph HellwigQuoted 120 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
All controller namespaces share the same tagset, so we can use this interface which does the optimal operation for parallel quiesce based on the tagset type(e.g. blocking tagsets and non-blocking tagsets).
nvme
connect_qshould not be quiesced when quiesce tagset, so set theQUEUE_FLAG_SKIP_TAGSET_QUIESCEto skip it when initconnect_q.Currently we use
NVME_NS_STOPPEDto ensure pairing quiescing and unquiescing. If use blk_mq_[un]quiesce_tagset,NVME_NS_STOPPEDwill be invalided, so introduceNVME_CTRL_STOPPEDto replaceNVME_NS_STOPPED. In addition, we never really quiesce a single namespace. It is a better choice to move the flag from ns to ctrl.Signed-off-by Chao Leng <[email protected]>[hch: rebased on top of prep patches]Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 38 +++++++++----------------------------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 10 insertions(+), 30 deletions(-)drivers/nvme/host/core.c +9
-29@@ -4891,6 +4891,7 @@ int nvme_alloc_io_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set, ret =PTR_ERR (ctrl->connect_q); goto out_free_tag_set; } +blk_queue_flag_set (QUEUE_FLAG_SKIP_TAGSET_QUIESCE , ctrl->connect_q); } ctrl->tagset = set;@@ -5090,20 +5091,6 @@ int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, }EXPORT_SYMBOL_GPL (nvme_init_ctrl); -static void nvme_start_ns_queue (struct nvme_ns *ns) -{ -if (test_and_clear_bit (NVME_NS_STOPPED , &ns->flags)) -blk_mq_unquiesce_queue (ns->queue); -} - -static void nvme_stop_ns_queue (struct nvme_ns *ns) -{ -if (!test_and_set_bit (NVME_NS_STOPPED , &ns->flags)) -blk_mq_quiesce_queue (ns->queue); -else -blk_mq_wait_quiesce_done (ns->queue->tag_set); -} -/** * nvme_kill_queues(): Ends all namespace queues * @ctrl: the dead controller that needs to end @@ -5120,10 +5107,9 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl)if (ctrl->admin_q && !blk_queue_dying (ctrl->admin_q))nvme_start_admin_queue (ctrl);if (!test_and_set_bit (NVME_CTRL_NS_DEAD , &ctrl->flags)) { -list_for_each_entry (ns, &ctrl->namespaces, list) { +list_for_each_entry (ns, &ctrl->namespaces, list)blk_mark_disk_dead (ns->disk); -nvme_start_ns_queue (ns); - } +nvme_start_queues (ctrl); }up_read (&ctrl->namespaces_rwsem); }@@ -5179,23 +5165,17 @@ EXPORT_SYMBOL_GPL(nvme_start_freeze);void nvme_stop_queues (struct nvme_ctrl *ctrl) { -struct nvme_ns *ns; - -down_read (&ctrl->namespaces_rwsem); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_stop_ns_queue (ns); -up_read (&ctrl->namespaces_rwsem); +if (!test_and_set_bit (NVME_CTRL_STOPPED , &ctrl->flags)) +blk_mq_quiesce_tagset (ctrl->tagset); +else +blk_mq_wait_quiesce_done (ctrl->tagset); }EXPORT_SYMBOL_GPL (nvme_stop_queues);void nvme_start_queues (struct nvme_ctrl *ctrl) { -struct nvme_ns *ns; - -down_read (&ctrl->namespaces_rwsem); -list_for_each_entry (ns, &ctrl->namespaces, list) -nvme_start_ns_queue (ns); -up_read (&ctrl->namespaces_rwsem); +if (test_and_clear_bit (NVME_CTRL_STOPPED , &ctrl->flags)) +blk_mq_unquiesce_tagset (ctrl->tagset); }EXPORT_SYMBOL_GPL (nvme_start_queues);drivers/nvme/host/nvme.h +1
-1@@ -238,6 +238,7 @@ enum nvme_ctrl_flags {NVME_CTRL_ADMIN_Q_STOPPED =1 ,NVME_CTRL_STARTED_ONCE =2 ,NVME_CTRL_NS_DEAD =3 , + NVME_CTRL_STOPPED = 4, }; struct nvme_ctrl {@@ -487,7 +488,6 @@ struct nvme_ns {#define NVME_NS_ANA_PENDING 2#define NVME_NS_FORCE_RO 3#define NVME_NS_READY 4 -#define NVME_NS_STOPPED 5struct cdev cdev;struct device cdev_device;
Re: [PATCH 8/8] nvme: use blk_mq_[un]quiesce_tagset
In reply to Christoph HellwigQuoted 15 lines · Christoph Hellwig
From: Chao Leng <[email protected]>
All controller namespaces share the same tagset, so we can use this interface which does the optimal operation for parallel quiesce based on the tagset type(e.g. blocking tagsets and non-blocking tagsets).
nvme
connect_qshould not be quiesced when quiesce tagset, so set theQUEUE_FLAG_SKIP_TAGSET_QUIESCEto skip it when initconnect_q.Currently we use
NVME_NS_STOPPEDto ensure pairing quiescing and unquiescing. If use blk_mq_[un]quiesce_tagset,NVME_NS_STOPPEDwill be invalided, so introduceNVME_CTRL_STOPPEDto replaceNVME_NS_STOPPED. In addition, we never really quiesce a single namespace. It is a better choice to move the flag from ns to ctrl.
Quoted 8 lines
Signed-off-by Chao Leng <[email protected]>[hch: rebased on top of prep patches]Signed-off-by Christoph Hellwig <[email protected]>--- drivers/nvme/host/core.c | 38 +++++++++----------------------------- drivers/nvme/host/nvme.h | 2 +- 2 files changed, 10 insertions(+), 30 deletions(-)
Cheers,
Hannes
Signature · 5 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew Myers, Andrew McDonald, Martje Boudien Moerman
Sagi Grimberg
maintainerQuoted 7 lines
Hi all,
this series moves the SRCU struct used for quiescing to the
tag_setas the SRCU critical sections for dispatch take about the same time on all queues anyway, and then adopts the series from Chao that provides tagset-wide quiesce to use that infrastructure.
Looks nice, should be easy enough to modify scsi_host_block() to use this as well.
Looks good; for the series: