blk-mq: allow to unfreeze queue when io isn't drained
Subsystems (4)
cover blk-mq: allow to unfreeze queue when io isn't drained
Hi,
The 1st patch is one cleanup and prepares for introducing percpu_ref_resurge().
The 2nd patch introduces percpu_ref_resurge() for implementing blk_mq_unfreeze_queue_no_drain_io().
The 3rd patch introdces blk_mq_unfreeze_queue_no_drain_io() for cases in which queue can be unfreeze without draininig IO.
The 4th patch applies blk_mq_unfreeze_queue_no_drain_io() to NVMe PCI timeout, so that IO hang may be avoided inside nvme_reset_work() when new IO timeout is triggered.
Part of idea is from Jianchao's early work:
https://marc.info/?l=linux-kernel&m=153612052611020&w=2
Ming Lei (4):
percpu-refcount: move zeroing of percpu part into
percpu_ref_switch_to_atomic_rcu
lib/percpu-refcount: introduce percpu_ref_resurge()
blk-mq: introduce blk_mq_unfreeze_queue_no_drain_io
nvme: don't drain IO in nvme_reset_work()
block/blk-mq.c | 25 +++++++++++++-- drivers/nvme/host/core.c | 12 ++++--- drivers/nvme/host/nvme.h | 2 +- drivers/nvme/host/pci.c | 3 +- include/linux/blk-mq.h | 1 + include/linux/percpu-refcount.h | 1 + lib/percpu-refcount.c | 71 +++++++++++++++++++++++++++++++---------- 7 files changed, 90 insertions(+), 25 deletions(-)
Signature · 1 line
2.9.5
PATCH 1/4 percpu-refcount: move zeroing of percpu part into percpu_ref_switch_to_atomic_rcu
The percpu parts are added to atomic part in percpu_ref_switch_to_atomic_rcu(), so it is readable & reasonable to clear them there.
--- lib/percpu-refcount.c | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-)
lib/percpu-refcount.c +3 -5
@@ -130,8 +130,10 @@ static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)unsigned long count =0 ;int cpu; -for_each_possible_cpu (cpu) +for_each_possible_cpu (cpu) { count += *per_cpu_ptr (percpu_count, cpu); + *per_cpu_ptr (percpu_count, cpu) =0 ; + }pr_debug ("global %ld percpu %ld" ,atomic_long_read (&ref->count), (long )count);
@@ -187,7 +189,6 @@ static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,static void __percpu_ref_switch_to_percpu (struct percpu_ref *ref) {unsigned long __percpu *percpu_count =percpu_count_ptr (ref); -int cpu;BUG_ON (!percpu_count);
@@ -202,9 +203,6 @@ static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)* zeroing is visible to all percpu accesses which can see the * following __PERCPU_REF_ATOMIC clearing. */ -for_each_possible_cpu (cpu) - *per_cpu_ptr (percpu_count, cpu) =0 ; -smp_store_release (&ref->percpu_count_ptr, ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC); }
Signature · 1 line
2.9.5
PATCH 2/4 lib/percpu-refcount: introduce percpu_ref_resurge()
In reply to Ming LeiNow percpu_ref_reinit() can only be done on one percpu refcounter when it drops zero. And the limit shouldn't be so strict, and it is quite straightforward that we can do it when the refcount doesn't drop zero because it is at atomic mode.
This patch introduces percpu_ref_resurge() in which the above limit is relaxed, so we may avoid extra change[1] for NVMe timeout's requirement.
[1] https://marc.info/?l=linux-kernel&m=153612052611020&w=2
--- include/linux/percpu-refcount.h | 1 + lib/percpu-refcount.c | 63 ++++++++++++++++++++++++++++++++++------- 2 files changed, 53 insertions(+), 11 deletions(-)
include/linux/percpu-refcount.h +1 -0
@@ -109,6 +109,7 @@ void percpu_ref_switch_to_percpu(struct percpu_ref *ref);void percpu_ref_kill_and_confirm (struct percpu_ref *ref,percpu_ref_func_t *confirm_kill);void percpu_ref_reinit (struct percpu_ref *ref); +void percpu_ref_resurge (struct percpu_ref *ref);/** * percpu_ref_kill - drop the initial ref
lib/percpu-refcount.c +52 -11
74 lines not on this page · raw message
Signature · 1 line
2.9.5
Re: [PATCH 2/4] lib/percpu-refcount: introduce percpu_ref_resurge()
Quoted 7 lines · Ming Lei
Now
percpu_ref_reinit()can only be done on one percpu refcounter when it drops zero. And the limit shouldn't be so strict, and it is quite straightforward that we can do it when the refcount doesn't drop zero because it is at atomic mode.This patch introduces
percpu_ref_resurge()in which the above limit is relaxed, so we may avoid extra change[1] for NVMe timeout's requirement.
For now,
Please see the original discussion thread.
Thanks.
Signature · 1 line
tejun
Re: [PATCH 2/4] lib/percpu-refcount: introduce percpu_ref_resurge()
Quoted 6 lines · Tejun Heo
Quoted 5 lines · Ming Lei
Now
percpu_ref_reinit()can only be done on one percpu refcounter when it drops zero. And the limit shouldn't be so strict, and it is quite straightforward that we can do it when the refcount doesn't drop zero because it is at atomic mode.
Quoted 8 lines
Quoted 2 lines
This patch introduces
percpu_ref_resurge()in which the above limit is relaxed, so we may avoid extra change[1] for NVMe timeout's requirement.For now,
Nacked-by Tejun Heo <[email protected]>Please see the original discussion thread.
Your comment in that thread supposes that synchronize_rcu() is used for avoiding race with .release().
But this patchset doesn't use that approach at all for avoiding this race.
Thanks,
Ming
Re: [PATCH 2/4] lib/percpu-refcount: introduce percpu_ref_resurge()
In reply to Ming LeiHi Ming
lib/percpu-refcount.c void percpu_ref_kill_and_confirm(struct percpu_ref *ref, · line 365
Quoted 9 lines · Ming Lei
lib/percpu-refcount.c +9
-0+unsigned long __percpu *percpu_count; + +WARN_ON_ONCE (__ref_is_percpu (ref, &percpu_count)); + +/* get one extra ref for avoiding race with .release */ +rcu_read_lock_sched (); +atomic_long_add (1 , &ref->count); +rcu_read_unlock_sched (); + }
The rcu_read_lock_sched here is redundant. We have been in the critical section of a spin_lock_irqsave.
The atomic_long_add(1, &ref->count) may have two result.
1. ref->count > 1
it will not drop to zero any more.
2. ref->count == 1
it has dropped to zero and .release may be running.
Thanks
Jianchao
Re: [PATCH 2/4] lib/percpu-refcount: introduce percpu_ref_resurge()
Quoted 15 lines · jianchao.wang
Hi Ming
lib/percpu-refcount.c void percpu_ref_kill_and_confirm(struct percpu_ref *ref, · line 365
Quoted 9 lines · Ming Lei
lib/percpu-refcount.c +9
-0+unsigned long __percpu *percpu_count; + +WARN_ON_ONCE (__ref_is_percpu (ref, &percpu_count)); + +/* get one extra ref for avoiding race with .release */ +rcu_read_lock_sched (); +atomic_long_add (1 , &ref->count); +rcu_read_unlock_sched (); + }The
rcu_read_lock_schedhere is redundant. We have been in the critical section of aspin_lock_irqsave.
Right.
Quoted 6 lines
The
atomic_long_add(1, &ref->count)may have two result.
1.ref->count> 1
it will not drop to zero any more.
2.ref->count== 1
it has dropped to zero and .release may be running.
IMO, both the two cases are fine and supported, or do you have other concern about this way?
thanks,
Ming
Re: [PATCH 2/4] lib/percpu-refcount: introduce percpu_ref_resurge()
Quoted 20 lines · Ming Lei
Quoted 15 lines · jianchao.wang
Hi Ming
lib/percpu-refcount.c void percpu_ref_kill_and_confirm(struct percpu_ref *ref, · line 365
Quoted 9 lines · Ming Lei
lib/percpu-refcount.c +9
-0+unsigned long __percpu *percpu_count; + +WARN_ON_ONCE (__ref_is_percpu (ref, &percpu_count)); + +/* get one extra ref for avoiding race with .release */ +rcu_read_lock_sched (); +atomic_long_add (1 , &ref->count); +rcu_read_unlock_sched (); + }The
rcu_read_lock_schedhere is redundant. We have been in the critical section of aspin_lock_irqsave.Right.
Quoted 1 line
Quoted 8 lines
Quoted 5 lines
The
atomic_long_add(1, &ref->count)may have two result.
1.ref->count> 1
it will not drop to zero any more.
2.ref->count== 1
it has dropped to zero and .release may be running.IMO, both the two cases are fine and supported, or do you have other concern about this way?
It is too quick, :-)
Yeah, the .release() may be running.
For blk-mq/NVMe's use case, it won't be an issue. We may comment on this race and let user handle it if it is a problem.
thanks,
Ming
PATCH 3/4 blk-mq: introduce blk_mq_unfreeze_queue_no_drain_io
In reply to Ming LeiThis patch introduces blk_mq_unfreeze_queue_no_drain_io() so that it can be used when no necessary to check if IO is drained, such as nvme pci resetting(nvme_reset_work).
--- block/blk-mq.c | 25 +++++++++++++++++++++++-- include/linux/blk-mq.h | 1 + 2 files changed, 24 insertions(+), 2 deletions(-)
block/blk-mq.c +23 -2
@@ -191,20 +191,41 @@ void blk_mq_freeze_queue(struct request_queue *q) }EXPORT_SYMBOL_GPL (blk_mq_freeze_queue); -void blk_mq_unfreeze_queue (struct request_queue *q) +static void __blk_mq_unfreeze_queue (struct request_queue *q, +bool need_drop_zero) {int freeze_depth; freeze_depth =atomic_dec_return (&q->mq_freeze_depth);WARN_ON_ONCE (freeze_depth <0 );if (!freeze_depth) { -percpu_ref_reinit (&q->q_usage_counter); +if (need_drop_zero) +percpu_ref_reinit (&q->q_usage_counter); +else +percpu_ref_resurge (&q->q_usage_counter);wake_up_all (&q->mq_freeze_wq); } } + +void blk_mq_unfreeze_queue (struct request_queue *q) +{ +__blk_mq_unfreeze_queue (q, true); +}EXPORT_SYMBOL_GPL (blk_mq_unfreeze_queue);/* +* Compared with blk_mq_unfreeze_queue(), the verion of _no_drain_io +* doesn't require the queue is really frozen, and it is useful in +* case of timeout handling in which IO can't be drained and has to +* be retried after controler is recovered. +*/ +void blk_mq_unfreeze_queue_no_drain_io (struct request_queue *q) +{ +__blk_mq_unfreeze_queue (q, false); +} +EXPORT_SYMBOL_GPL (blk_mq_unfreeze_queue_no_drain_io); + +/* * FIXME: replace the scsi_internal_device_*block_nowait() calls in the * mpt3sas driver such that this function can be removed. */
include/linux/blk-mq.h +1 -0
@@ -279,6 +279,7 @@ void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, busy_tag_iter_fn *fn, void *priv);void blk_mq_freeze_queue (struct request_queue *q);void blk_mq_unfreeze_queue (struct request_queue *q); +void blk_mq_unfreeze_queue_no_drain_io (struct request_queue *q);void blk_freeze_queue_start (struct request_queue *q);void blk_mq_freeze_queue_wait (struct request_queue *q); int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
Signature · 1 line
2.9.5
PATCH 4/4 nvme: don't drain IO in nvme_reset_work()
In reply to Ming LeiAfter the controller is recovered, it isn't necessary to wait for completion of all in-flight IO. More importantly, it is easy to trigger deadlock if there is new IO timeout triggered.
--- drivers/nvme/host/core.c | 12 ++++++++---- drivers/nvme/host/nvme.h | 2 +- drivers/nvme/host/pci.c | 3 +-- 3 files changed, 10 insertions(+), 7 deletions(-)
drivers/nvme/host/core.c +8 -4
@@ -1203,7 +1203,7 @@ static void nvme_passthru_end(struct nvme_ctrl *ctrl, u32 effects)if (effects &NVME_CMD_EFFECTS_LBCC )nvme_update_formats (ctrl);if (effects & (NVME_CMD_EFFECTS_LBCC |NVME_CMD_EFFECTS_CSE_MASK )) -nvme_unfreeze (ctrl); +nvme_unfreeze (ctrl, true);if (effects &NVME_CMD_EFFECTS_CCC )nvme_init_identify (ctrl);if (effects & (NVME_CMD_EFFECTS_NIC |NVME_CMD_EFFECTS_NCC ))
@@ -3602,13 +3602,17 @@ void nvme_kill_queues(struct nvme_ctrl *ctrl) }EXPORT_SYMBOL_GPL (nvme_kill_queues); -void nvme_unfreeze (struct nvme_ctrl *ctrl) +void nvme_unfreeze (struct nvme_ctrl *ctrl,bool check_io_drained) {struct nvme_ns *ns;down_read (&ctrl->namespaces_rwsem); -list_for_each_entry (ns, &ctrl->namespaces, list) -blk_mq_unfreeze_queue (ns->queue); +list_for_each_entry (ns, &ctrl->namespaces, list) { +if (check_io_drained) +blk_mq_unfreeze_queue (ns->queue); +else +blk_mq_unfreeze_queue_no_drain_io (ns->queue); + }up_read (&ctrl->namespaces_rwsem); }EXPORT_SYMBOL_GPL (nvme_unfreeze);
drivers/nvme/host/nvme.h +1 -1
@@ -432,7 +432,7 @@ void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,void nvme_stop_queues (struct nvme_ctrl *ctrl);void nvme_start_queues (struct nvme_ctrl *ctrl);void nvme_kill_queues (struct nvme_ctrl *ctrl); -void nvme_unfreeze (struct nvme_ctrl *ctrl); +void nvme_unfreeze (struct nvme_ctrl *ctrl,bool check_io_drained);void nvme_wait_freeze (struct nvme_ctrl *ctrl);void nvme_wait_freeze_timeout (struct nvme_ctrl *ctrl,long timeout);void nvme_start_freeze (struct nvme_ctrl *ctrl);
drivers/nvme/host/pci.c +1 -2
@@ -2328,11 +2328,10 @@ static void nvme_reset_work(struct work_struct *work) new_state =NVME_CTRL_ADMIN_ONLY ; }else {nvme_start_queues (&dev->ctrl); -nvme_wait_freeze (&dev->ctrl);/* hit this only when allocate tagset fails */ if (nvme_dev_add (dev)) new_state =NVME_CTRL_ADMIN_ONLY ; -nvme_unfreeze (&dev->ctrl); +nvme_unfreeze (&dev->ctrl, false); }/*
Signature · 1 line
2.9.5