blk-cgroup: Flush stats before releasing blkcg_gq
Subsystems (4)
PATCH v3 1/1 blk-cgroup: Flush stats before releasing blkcg_gq
As noted by Michal, the blkg_iostat_set's in the lockless list hold reference to blkg's to protect against their removal. Those blkg's hold reference to blkcg. When a cgroup is being destroyed, cgroup_rstat_flush() is only called at css_release_work_fn() which is called when the blkcg reference count reaches 0. This circular dependency will prevent blkcg and some blkgs from being freed after they are made offline.
It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call cgroup_rstat_flush() which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.
To prevent this potential memory leak:
- flush blkcg per-cpu stats list in __blkg_release(), when no new stat can be added
- add global blkg_stat_lock for covering concurrent parent blkg stat update
- don't grab bio->bi_blkg reference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patch
Based on Waiman's patch:
https://lore.kernel.org/linux-block/[email protected]/
---
V3:
- add one global blkg_stat_lock for avoiding concurrent update on blkg stat; this way is easier for backport, also won't cause contention;
V2:
- remove kernel/cgroup change, and call blkcg_rstat_flush() to flush stat directly
block/blk-cgroup.c | 40 +++++++++++++++++++++++++++++++--------- 1 file changed, 31 insertions(+), 9 deletions(-)
block/blk-cgroup.c +31 -9
@@ -34,6 +34,8 @@#include "blk-ioprio.h" #include "blk-throttle.h" +static void __blkcg_rstat_flush (struct blkcg *blkcg,int cpu); +/* * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation. * blkcg_pol_register_mutex nests outside of it and synchronizes entire
@@ -56,6 +58,8 @@ static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */bool blkcg_debug_stats = false; +static DEFINE_RAW_SPINLOCK (blkg_stat_lock); +#define BLKG_DESTROY_BATCH_SIZE 64/*
@@ -163,10 +167,20 @@ static void blkg_free(struct blkcg_gq *blkg)static void __blkg_release (struct rcu_head *rcu) { struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head); +struct blkcg *blkcg = blkg->blkcg; +int cpu;#ifdef CONFIG_BLK_CGROUP_PUNT_BIO WARN_ON (!bio_list_empty (&blkg->async_bios));#endif +/* +* Flush all the non-empty percpu lockless lists before releasing +* us, given these stat belongs to us. +* +* blkg_stat_lock is for serializing blkg stat update +*/ +for_each_possible_cpu (cpu) +__blkcg_rstat_flush (blkcg, cpu);/* release the blkcg and parent blkg refs this blkg has been holding */ css_put (&blkg->blkcg->css);
@@ -951,23 +965,26 @@ static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,u64_stats_update_end_irqrestore (&blkg->iostat.sync, flags); } -static void blkcg_rstat_flush (struct cgroup_subsys_state *css,int cpu) +static void __blkcg_rstat_flush (struct blkcg *blkcg,int cpu) { -struct blkcg *blkcg =css_to_blkcg (css);struct llist_head *lhead =per_cpu_ptr (blkcg->lhead, cpu);struct llist_node *lnode;struct blkg_iostat_set *bisc, *next_bisc; -/* Root-level stats are sourced from system-wide IO stats */ -if (!cgroup_parent (css->cgroup)) -return ; -rcu_read_lock (); lnode =llist_del_all (lhead);if (!lnode) goto out; +/* +* For covering concurrent parent blkg update from blkg_release(). +* +* When flushing from cgroup, cgroup_rstat_lock is always held, so +* this lock won't cause contention most of time. +*/ +raw_spin_lock (&blkg_stat_lock); +/* * Iterate only the iostat_cpu's queued in the lockless list. */
@@ -991,13 +1008,19 @@ static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)if (parent && parent->parent)blkcg_iostat_update (parent, &blkg->iostat.cur, &blkg->iostat.last); -percpu_ref_put (&blkg->refcnt); } - +raw_spin_unlock (&blkg_stat_lock); out:rcu_read_unlock (); } +static void blkcg_rstat_flush (struct cgroup_subsys_state *css,int cpu) +{ +/* Root-level stats are sourced from system-wide IO stats */ +if (cgroup_parent (css->cgroup)) +__blkcg_rstat_flush (css_to_blkcg (css), cpu); +} +/* * We source root cgroup stats from the system-wide stats to avoid * tracking the same information twice and incurring overhead when no
@@ -2075,7 +2098,6 @@ void blk_cgroup_bio_start(struct bio *bio)llist_add (&bis->lnode, lhead);WRITE_ONCE (bis->lqueued, true); -percpu_ref_get (&bis->blkg->refcnt); }u64_stats_update_end_irqrestore (&bis->sync, flags);
Signature · 1 line
2.40.1
Michal Koutný
maintainerQuoted 5 lines · Ming Lei
It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call
cgroup_rstat_flush()which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.
On v2, io implies memory too.
Do you observe the leak on the v2 system too?
(Beware that (not only) dirty pages would pin offlined memcg, so the actual mem_cgroup_css_release and cgroup_rstat_flush would be further delayed.)
Quoted 7 lines
To prevent this potential memory leak:
- flush blkcg per-cpu stats list in
__blkg_release(), when no new stat can be added- add global
blkg_stat_lockfor covering concurrent parent blkg stat update
It's bit unfortunate yet another lock is added :-/
IIUC, even Waiman's patch (flush in blkcg_destroy_blkcfs) would need synchronization for different CPU replicas flushes in blkcg_iostat_update, right?
Quoted 4 lines
- don't grab
bio->bi_blkgreference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patch
At one moment, the lhead -> blkcg_gq reference seemed alright to me and consequently blkcg_gq -> blkcg is the one that looks reversed (forming the cycle). But changing its direction would be much more fundamental change, it'd need also kind of blkcg_gq reparenting -- similarly to memcg.
Thanks,
Michal
Quoted 6 lines · Michal Koutný
Quoted 5 lines · Ming Lei
It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call
cgroup_rstat_flush()which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.
Quoted 6 lines
On v2, io implies memory too.
Do you observe the leak on the v2 system too?(Beware that (not only) dirty pages would pin offlined memcg, so the actual
mem_cgroup_css_releaseandcgroup_rstat_flushwould be further delayed.)
The memory leak isn't observed on cgroup v2 as the cgroup_rstat_flush() call by the memory controller will help to flush the extra references holding back blkcgs. It is only happening with a cgroup v1 configuration where blkcg is standalone in a cgroup.
Quoted 6 lines
Quoted 6 lines
To prevent this potential memory leak:
- flush blkcg per-cpu stats list in
__blkg_release(), when no new stat can be added
Quoted 7 lines
Quoted 2 lines
- add global
blkg_stat_lockfor covering concurrent parent blkg stat updateIt's bit unfortunate yet another lock is added :-/
IIUC, even Waiman's patch (flush in
blkcg_destroy_blkcfs) would need synchronization for different CPU replicas flushes inblkcg_iostat_update, right?
Right, the goal is to prevent concurrent call of blkcg_iostat_update() to the same blkg.
Cheers,
Longman
Quoted 9 lines · Michal Koutný
Quoted 5 lines · Ming Lei
It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call
cgroup_rstat_flush()which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.On v2, io implies memory too.
Do you observe the leak on the v2 system too?
blkg leak is only observed on v1.
Quoted 7 lines
(Beware that (not only) dirty pages would pin offlined memcg, so the actual
mem_cgroup_css_releaseandcgroup_rstat_flushwould be further delayed.)Quoted 2 lines
To prevent this potential memory leak:
Quoted 7 lines
Quoted 5 lines
- flush blkcg per-cpu stats list in
__blkg_release(), when no new stat can be added- add global
blkg_stat_lockfor covering concurrent parent blkg stat updateIt's bit unfortunate yet another lock is added :-/
Cause it should be the simplest one for backport, :-)
Quoted 5 lines
IIUC, even Waiman's patch (flush in
blkcg_destroy_blkcfs) would need synchronization for different CPU replicas flushes inblkcg_iostat_update, right?
Quoted 7 lines
Quoted 4 lines
- don't grab
bio->bi_blkgreference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patchAt one moment, the lhead ->
blkcg_gqreference seemed alright to me and
But bio->bi_blkg has grabbed one reference in blk_cgroup_bio_start already.
Quoted 1 line
consequently
blkcg_gq-> blkcg is the one that looks reversed (forming
IMO, this one is correct, cause blkcg_gq depends on blkcg.
Quoted 3 lines
the cycle). But changing its direction would be much more fundamental change, it'd need also kind of
blkcg_gqreparenting -- similarly to memcg.
Thanks,
Ming
PATCH 1/1 blk-cgroup: Flush stats before releasing blkcg_gq
In reply to Ming LeiAs noted by Michal, the blkg_iostat_set's in the lockless list hold reference to blkg's to protect against their removal. Those blkg's hold reference to blkcg. When a cgroup is being destroyed, cgroup_rstat_flush() is only called at css_release_work_fn() which is called when the blkcg reference count reaches 0. This circular dependency will prevent blkcg and some blkgs from being freed after they are made offline.
It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call cgroup_rstat_flush() which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.
To prevent this potential memory leak:
- flush blkcg per-cpu stats list in __blkg_release(), when no new stat can be added to avoid use-after-free of the percpu blkg_iostat_set in futue cgroup_rstat_flush*() calls.
- add a cgroup_rstat_flush_acquire() helper and call it to acquire cgroup_rstat_lock to block concurrent execution of other cgroup_rstat_flush*() calls
- don't grab bio->bi_blkg reference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patch
Based on Waiman's patch:
https://lore.kernel.org/linux-block/[email protected]/
--- block/blk-cgroup.c | 57 +++++++++++++++++++++++++++++++----------- include/linux/cgroup.h | 1 + kernel/cgroup/rstat.c | 15 ++++++++++- 3 files changed, 57 insertions(+), 16 deletions(-)
block/blk-cgroup.c +42 -15
@@ -160,13 +160,39 @@ static void blkg_free(struct blkcg_gq *blkg)schedule_work (&blkg->free_work); } +static void __blkcg_rstat_flush (struct llist_node *lnode); +static void __blkg_release (struct rcu_head *rcu) { struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head); +struct blkcg *blkcg = blkg->blkcg; +int cpu;#ifdef CONFIG_BLK_CGROUP_PUNT_BIO WARN_ON (!bio_list_empty (&blkg->async_bios));#endif +/* +* Flush all the non-empty percpu lockless lists before releasing +* us, given these stat belongs to us. +* +* Hold the cgroup_rstat_lock before calling __blkcg_rstat_flush() +* to block concurrent cgroup_rstat_flush*() calls. +*/ +for_each_possible_cpu (cpu) { +struct llist_head *lhead =per_cpu_ptr (blkcg->lhead, cpu); +struct llist_node *lnode; + +if (llist_empty (lhead)) +continue ; + + lnode =llist_del_all (lhead); +if (!lnode) +continue ; + +cgroup_rstat_flush_acquire (); +__blkcg_rstat_flush (lnode); +cgroup_rstat_flush_release (); + }/* release the blkcg and parent blkg refs this blkg has been holding */ css_put (&blkg->blkcg->css);
@@ -951,23 +977,12 @@ static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,u64_stats_update_end_irqrestore (&blkg->iostat.sync, flags); } -static void blkcg_rstat_flush (struct cgroup_subsys_state *css,int cpu) +static void __blkcg_rstat_flush (struct llist_node *lnode) { -struct blkcg *blkcg =css_to_blkcg (css); -struct llist_head *lhead =per_cpu_ptr (blkcg->lhead, cpu); -struct llist_node *lnode;struct blkg_iostat_set *bisc, *next_bisc; -/* Root-level stats are sourced from system-wide IO stats */ -if (!cgroup_parent (css->cgroup)) -return ; -rcu_read_lock (); - lnode =llist_del_all (lhead); -if (!lnode) - goto out; -/* * Iterate only the iostat_cpu's queued in the lockless list. */
@@ -991,13 +1006,26 @@ static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)if (parent && parent->parent)blkcg_iostat_update (parent, &blkg->iostat.cur, &blkg->iostat.last); -percpu_ref_put (&blkg->refcnt); } -out:rcu_read_unlock (); } +static void blkcg_rstat_flush (struct cgroup_subsys_state *css,int cpu) +{ +struct blkcg *blkcg =css_to_blkcg (css); +struct llist_head *lhead =per_cpu_ptr (blkcg->lhead, cpu); +struct llist_node *lnode; + +/* Root-level stats are sourced from system-wide IO stats */ +if (!cgroup_parent (css->cgroup)) +return ; + + lnode =llist_del_all (lhead); +if (lnode) +__blkcg_rstat_flush (lnode); +} +/* * We source root cgroup stats from the system-wide stats to avoid * tracking the same information twice and incurring overhead when no
@@ -2075,7 +2103,6 @@ void blk_cgroup_bio_start(struct bio *bio)llist_add (&bis->lnode, lhead);WRITE_ONCE (bis->lqueued, true); -percpu_ref_get (&bis->blkg->refcnt); }u64_stats_update_end_irqrestore (&bis->sync, flags);
include/linux/cgroup.h +1 -0
@@ -694,6 +694,7 @@ void cgroup_rstat_updated(struct cgroup *cgrp, int cpu);void cgroup_rstat_flush (struct cgroup *cgrp);void cgroup_rstat_flush_atomic (struct cgroup *cgrp);void cgroup_rstat_flush_hold (struct cgroup *cgrp); +void cgroup_rstat_flush_acquire (void );void cgroup_rstat_flush_release (void );/*
kernel/cgroup/rstat.c +14 -1
@@ -273,7 +273,20 @@ void cgroup_rstat_flush_hold(struct cgroup *cgrp) }/** -* cgroup_rstat_flush_release - release cgroup_rstat_flush_hold() +* cgroup_rstat_flush_acquire - acquire cgroup_rstat_lock +* +* Callers can acquire the internal cgroup_rstat_lock to prevent concurrent +* execution of cgroup_rstat_flush*() and the controller callbacks. +*/ +void cgroup_rstat_flush_acquire (void ) + __acquires(&cgroup_rstat_lock) +{ +spin_lock_irq (&cgroup_rstat_lock); +} + +/** +* cgroup_rstat_flush_release - release cgroup_rstat_flush_hold() or +* cgroup_rstat_flush_acquire() */ void cgroup_rstat_flush_release(void) __releases(&cgroup_rstat_lock)
Signature · 1 line
2.31.1
Quoted 43 lines · Waiman Long
As noted by Michal, the blkg_iostat_set's in the lockless list hold reference to blkg's to protect against their removal. Those blkg's hold reference to blkcg. When a cgroup is being destroyed,
cgroup_rstat_flush()is only called atcss_release_work_fn()which is called when the blkcg reference count reaches 0. This circular dependency will prevent blkcg and some blkgs from being freed after they are made offline.It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call
cgroup_rstat_flush()which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.To prevent this potential memory leak:
- flush blkcg per-cpu stats list in
__blkg_release(), when no new stat can be added to avoid use-after-free of the percpublkg_iostat_setin futue cgroup_rstat_flush*() calls.- add a
cgroup_rstat_flush_acquire()helper and call it to acquirecgroup_rstat_lockto block concurrent execution of other cgroup_rstat_flush*() calls- don't grab
bio->bi_blkgreference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patchBased on Waiman's patch:
https://lore.kernel.org/linux-block/[email protected]/
Fixes 3b8cc6298724 ("blk-cgroup: Optimize blkcg_rstat_flush()")Reported-by Jay Shin <[email protected]>Cc Waiman Long <[email protected]>Cc Tejun Heo <[email protected]>Cc Yosry Ahmed <[email protected]>Co-developed-by Ming Lei <[email protected]>Signed-off-by Ming Lei <[email protected]>Signed-off-by Waiman Long <[email protected]>
Quoted 5 lines
--- block/blk-cgroup.c | 57 +++++++++++++++++++++++++++++++----------- include/linux/cgroup.h | 1 + kernel/cgroup/rstat.c | 15 ++++++++++- 3 files changed, 57 insertions(+), 16 deletions(-)
This is my counter-proposal to Ming's v3 patch. The major difference is that I used the existing cgroup_rstat_lock instead of adding a new internal lock. This minimizes performance impact to existing cgroup_rstat_flush*() call while achieving the same objective. I am fine with Ming current v3 patch if we decide to go that way.
Thanks,
Longman
Quoted 145 lines
block/blk-cgroup.c +42
-15@@ -160,13 +160,39 @@ static void blkg_free(struct blkcg_gq *blkg)schedule_work (&blkg->free_work); } +static void __blkcg_rstat_flush (struct llist_node *lnode); +static void __blkg_release (struct rcu_head *rcu) { struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head); +struct blkcg *blkcg = blkg->blkcg; +int cpu;#ifdef CONFIG_BLK_CGROUP_PUNT_BIO WARN_ON (!bio_list_empty (&blkg->async_bios));#endif +/* +* Flush all the non-empty percpu lockless lists before releasing +* us, given these stat belongs to us. +* +* Hold the cgroup_rstat_lock before calling __blkcg_rstat_flush() +* to block concurrent cgroup_rstat_flush*() calls. +*/ +for_each_possible_cpu (cpu) { +struct llist_head *lhead =per_cpu_ptr (blkcg->lhead, cpu); +struct llist_node *lnode; + +if (llist_empty (lhead)) +continue ; + + lnode =llist_del_all (lhead); +if (!lnode) +continue ; + +cgroup_rstat_flush_acquire (); +__blkcg_rstat_flush (lnode); +cgroup_rstat_flush_release (); + }/* release the blkcg and parent blkg refs this blkg has been holding */ css_put (&blkg->blkcg->css);@@ -951,23 +977,12 @@ static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,u64_stats_update_end_irqrestore (&blkg->iostat.sync, flags); } -static void blkcg_rstat_flush (struct cgroup_subsys_state *css,int cpu) +static void __blkcg_rstat_flush (struct llist_node *lnode) { -struct blkcg *blkcg =css_to_blkcg (css); -struct llist_head *lhead =per_cpu_ptr (blkcg->lhead, cpu); -struct llist_node *lnode;struct blkg_iostat_set *bisc, *next_bisc; -/* Root-level stats are sourced from system-wide IO stats */ -if (!cgroup_parent (css->cgroup)) -return ; -rcu_read_lock (); - lnode =llist_del_all (lhead); -if (!lnode) - goto out; -/* * Iterate only the iostat_cpu's queued in the lockless list. */ @@ -991,13 +1006,26 @@ static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)if (parent && parent->parent)blkcg_iostat_update (parent, &blkg->iostat.cur, &blkg->iostat.last); -percpu_ref_put (&blkg->refcnt); } -out:rcu_read_unlock (); } +static void blkcg_rstat_flush (struct cgroup_subsys_state *css,int cpu) +{ +struct blkcg *blkcg =css_to_blkcg (css); +struct llist_head *lhead =per_cpu_ptr (blkcg->lhead, cpu); + struct llist_node *lnode;Jay Shin <[email protected]> + +/* Root-level stats are sourced from system-wide IO stats */ +if (!cgroup_parent (css->cgroup)) +return ; + + lnode =llist_del_all (lhead); +if (lnode) +__blkcg_rstat_flush (lnode); +} +/* * We source root cgroup stats from the system-wide stats to avoid * tracking the same information twice and incurring overhead when no @@ -2075,7 +2103,6 @@ void blk_cgroup_bio_start(struct bio *bio)llist_add (&bis->lnode, lhead);WRITE_ONCE (bis->lqueued, true); -percpu_ref_get (&bis->blkg->refcnt); }u64_stats_update_end_irqrestore (&bis->sync, flags);include/linux/cgroup.h +1
-0@@ -694,6 +694,7 @@ void cgroup_rstat_updated(struct cgroup *cgrp, int cpu);void cgroup_rstat_flush (struct cgroup *cgrp);void cgroup_rstat_flush_atomic (struct cgroup *cgrp);void cgroup_rstat_flush_hold (struct cgroup *cgrp); +void cgroup_rstat_flush_acquire (void );void cgroup_rstat_flush_release (void );/* kernel/cgroup/rstat.c +14
-1@@ -273,7 +273,20 @@ void cgroup_rstat_flush_hold(struct cgroup *cgrp) }/** -* cgroup_rstat_flush_release - release cgroup_rstat_flush_hold() +* cgroup_rstat_flush_acquire - acquire cgroup_rstat_lock +* +* Callers can acquire the internal cgroup_rstat_lock to prevent concurrent +* execution of cgroup_rstat_flush*() and the controller callbacks. +*/ +void cgroup_rstat_flush_acquire (void ) + __acquires(&cgroup_rstat_lock) +{ +spin_lock_irq (&cgroup_rstat_lock); +} + +/** +* cgroup_rstat_flush_release - release cgroup_rstat_flush_hold() or +* cgroup_rstat_flush_acquire() */ void cgroup_rstat_flush_release(void) __releases(&cgroup_rstat_lock)
Quoted 50 lines · Waiman Long
Quoted 43 lines · Waiman Long
As noted by Michal, the blkg_iostat_set's in the lockless list hold reference to blkg's to protect against their removal. Those blkg's hold reference to blkcg. When a cgroup is being destroyed,
cgroup_rstat_flush()is only called atcss_release_work_fn()which is called when the blkcg reference count reaches 0. This circular dependency will prevent blkcg and some blkgs from being freed after they are made offline.It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call
cgroup_rstat_flush()which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.To prevent this potential memory leak:
- flush blkcg per-cpu stats list in
__blkg_release(), when no new stat can be added to avoid use-after-free of the percpublkg_iostat_setin futue cgroup_rstat_flush*() calls.- add a
cgroup_rstat_flush_acquire()helper and call it to acquirecgroup_rstat_lockto block concurrent execution of other cgroup_rstat_flush*() calls- don't grab
bio->bi_blkgreference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patchBased on Waiman's patch:
https://lore.kernel.org/linux-block/[email protected]/
Fixes 3b8cc6298724 ("blk-cgroup: Optimize blkcg_rstat_flush()")Reported-by Jay Shin <[email protected]>Cc Waiman Long <[email protected]>Cc Tejun Heo <[email protected]>Cc Yosry Ahmed <[email protected]>Co-developed-by Ming Lei <[email protected]>Signed-off-by Ming Lei <[email protected]>Signed-off-by Waiman Long <[email protected]>Quoted 5 lines
--- block/blk-cgroup.c | 57 +++++++++++++++++++++++++++++++----------- include/linux/cgroup.h | 1 + kernel/cgroup/rstat.c | 15 ++++++++++- 3 files changed, 57 insertions(+), 16 deletions(-)
Quoted 3 lines
This is my counter-proposal to Ming's v3 patch. The major difference is that I used the existing
cgroup_rstat_lockinstead of adding a new internal lock. This minimizes performance impact to existing cgroup_rstat_flush*() call
The added internal lock has ~zero perf impact on rstat flush cause the lock won't be contended basically.
Quoted 2 lines
while achieving the same objective. I am fine with Ming current v3 patch if we decide to go that way.
As I mentioned, the main motivation with internal lock is to make the fix as simple as possible since cross-subsystem change isn't involved, and I am fine with any following cleanup or improvement on current blkg rstat flush.
Another benefit with this internal lock is that race in blkcg_reset_stats() can be avoided.
Thanks,
Ming
Quoted 39 lines · Ming Lei
As noted by Michal, the blkg_iostat_set's in the lockless list hold reference to blkg's to protect against their removal. Those blkg's hold reference to blkcg. When a cgroup is being destroyed,
cgroup_rstat_flush()is only called atcss_release_work_fn()which is called when the blkcg reference count reaches 0. This circular dependency will prevent blkcg and some blkgs from being freed after they are made offline.It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call
cgroup_rstat_flush()which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.To prevent this potential memory leak:
- flush blkcg per-cpu stats list in
__blkg_release(), when no new stat can be added- add global
blkg_stat_lockfor covering concurrent parent blkg stat update- don't grab
bio->bi_blkgreference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patchBased on Waiman's patch:
https://lore.kernel.org/linux-block/[email protected]/
Fixes 3b8cc6298724 ("blk-cgroup: Optimize blkcg_rstat_flush()")Reported-by Jay Shin <[email protected]>Cc Waiman Long <[email protected]>Cc Tejun Heo <[email protected]>Cc Yosry Ahmed <[email protected]>Signed-off-by Ming Lei <[email protected]>
Quoted 4 lines
---
V3:- add one global blkg_stat_lock for avoiding concurrent update on blkg stat; this way is easier for backport, also won't cause contention;
Hello Jens and Tejun,
Can we move on with this patch or Waiman's version[1]?
I am fine with either one.
[1] https://lore.kernel.org/linux-block/[email protected]/
Thanks,
Ming
Tejun Heo
maintainerQuoted 11 lines · Ming Lei
Quoted 3 lines
V3:
- add one global blkg_stat_lock for avoiding concurrent update on blkg stat; this way is easier for backport, also won't cause contention;Hello Jens and Tejun,
Can we move on with this patch or Waiman's version[1]?
I am fine with either one.
I personally like Ming's version with a separate lock as it's a bit simpler. Let's go with that one.
Thanks.
Signature · 1 line
tejun
Quoted 31 lines · Ming Lei
As noted by Michal, the blkg_iostat_set's in the lockless list hold reference to blkg's to protect against their removal. Those blkg's hold reference to blkcg. When a cgroup is being destroyed,
cgroup_rstat_flush()is only called atcss_release_work_fn()which is called when the blkcg reference count reaches 0. This circular dependency will prevent blkcg and some blkgs from being freed after they are made offline.It is less a problem if the cgroup to be destroyed also has other controllers like memory that will call
cgroup_rstat_flush()which will clean up the reference count. If block is the only controller that uses rstat, these offline blkcg and blkgs may never be freed leaking more and more memory over time.To prevent this potential memory leak:
- flush blkcg per-cpu stats list in
__blkg_release(), when no new stat can be added- add global
blkg_stat_lockfor covering concurrent parent blkg stat update- don't grab
bio->bi_blkgreference when adding the stats into blkcg's per-cpu stat list since all stats are guaranteed to be consumed before releasing blkg instance, and grabbing blkg reference for stats was the most fragile part of original patchBased on Waiman's patch:
Quoted 8 lines
Fixes 3b8cc6298724 ("blk-cgroup: Optimize blkcg_rstat_flush()")Reported-by Jay Shin <[email protected]>Cc Waiman Long <[email protected]>Cc Tejun Heo <[email protected]>Cc Yosry Ahmed <[email protected]>Signed-off-by Ming Lei <[email protected]>
Thanks.
Signature · 1 line
tejun