sbitmap: Fix two issues in the per-bitmap wakeup counter code
Subsystems (3)
cover sbitmap: Fix two issues in the per-bitmap wakeup counter code
Jan reported two issues in the original thread.
The first is that wake_index was not updated after returning from sbq_wake_ptr which meant we'd have to empty the wq before moving to the next one. Patch 1/3 in this series reorders the code to avoid this condition, increasing fairness of queue selection and preventing starvation. I sent this patch already on the other thread and Jan reviewed it, but since it is a small one, and a dependency for the other, I'm resending it a along this series.
The second issue is trickier. When the selected queue is emptied after the waitqueue_active check and before wake_up_nr, there is no waiters to be awaken in that queue, even if other queues might have it. This causes us to loose one too many wakeups, and there might not be enough requests in flight to wake up every queued request.
The proposed fix, is to walk through every queue after doing the atomic update, such that we ensure any waiters already queued are candidates for awakening, and that we awake at least 1 waiter in any of the queues. The patch is a bit more complex than the suggestion since it avoids partial updates to wake_index, which measurably hurt performance unnecessarily.
It survived the same tests done on the original patch.
btw, I'm still missing the latency and utilisation reports. I haven't forgotten about it, but I didn't have a chance to collect them yet. Sorry. I will follow up with them for completeness, even if the original patch is already queued.
Gabriel Krisman Bertazi (3):
sbitmap: Advance the queue index before waking up a queue
wait: Return number of exclusive waiters awaken
sbitmap: Try each queue to wake up at least one waiter
include/linux/wait.h | 2 +- kernel/sched/wait.c | 18 +++++++++++------- lib/sbitmap.c | 36 +++++++++++++++++++----------------- 3 files changed, 31 insertions(+), 25 deletions(-)
Signature · 1 line
2.35.3
PATCH 1/3 sbitmap: Advance the queue index before waking up a queue
When a queue is awaken, the wake_index written by sbq_wake_ptr currently keeps pointing to the same queue. On the next wake up, it will thus retry the same queue, which is unfair to other queues, and can lead to starvation. This patch, moves the index update to happen before the queue is returned, such that it will now try a different queue first on the next wake up, improving fairness.
--- lib/sbitmap.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)
lib/sbitmap.c +8 -2
@@ -571,13 +571,19 @@ static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)for (i =0 ; i <SBQ_WAIT_QUEUES ; i++) {struct sbq_wait_state *ws = &sbq->ws[wake_index]; +/* +* Advance the index before checking the current queue. +* It improves fairness, by ensuring the queue doesn't +* need to be fully emptied before trying to wake up +* from the next one. +*/ + wake_index =sbq_index_inc (wake_index); +if (waitqueue_active (&ws->wait)) {if (wake_index !=atomic_read (&sbq->wake_index))atomic_set (&sbq->wake_index, wake_index);return ws; } - - wake_index =sbq_index_inc (wake_index); }return NULL ;
Signature · 1 line
2.35.3
PATCH 2/3 wait: Return number of exclusive waiters awaken
In reply to Gabriel Krisman BertaziSbitmap code will need to know how many waiters were actually woken for its batched wakeups implementation. Return the number of woken exclusive waiters from __wake_up() to facilitate that.
--- include/linux/wait.h | 2 +- kernel/sched/wait.c | 18 +++++++++++------- 2 files changed, 12 insertions(+), 8 deletions(-)
include/linux/wait.h +1 -1
@@ -209,7 +209,7 @@ __remove_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wqlist_del (&wq_entry->entry); } -void __wake_up (struct wait_queue_head *wq_head,unsigned int mode,int nr,void *key); +int __wake_up (struct wait_queue_head *wq_head,unsigned int mode,int nr,void *key);void __wake_up_locked_key (struct wait_queue_head *wq_head,unsigned int mode,void *key);void __wake_up_locked_key_bookmark (struct wait_queue_head *wq_head,unsigned int mode,void *key,wait_queue_entry_t *bookmark);
kernel/sched/wait.c +11 -7
@@ -121,11 +121,12 @@ static int __wake_up_common(struct wait_queue_head *wq_head, unsigned int mode,return nr_exclusive; } -static void __wake_up_common_lock (struct wait_queue_head *wq_head,unsigned int mode, +static int __wake_up_common_lock (struct wait_queue_head *wq_head,unsigned int mode,int nr_exclusive,int wake_flags,void *key) {unsigned long flags;wait_queue_entry_t bookmark; +int remaining = nr_exclusive; bookmark.flags =0 ; bookmark.private =NULL ;
@@ -134,10 +135,12 @@ static void __wake_up_common_lock(struct wait_queue_head *wq_head, unsigned intdo {spin_lock_irqsave (&wq_head->lock, flags); - nr_exclusive =__wake_up_common (wq_head, mode, nr_exclusive, + remaining =__wake_up_common (wq_head, mode, remaining, wake_flags, key, &bookmark);spin_unlock_irqrestore (&wq_head->lock, flags); }while (bookmark.flags &WQ_FLAG_BOOKMARK ); + +return nr_exclusive - remaining; }/**
@@ -147,13 +150,14 @@ static void __wake_up_common_lock(struct wait_queue_head *wq_head, unsigned int* @nr_exclusive: how many wake-one or wake-many threads to wake up * @key: is directly passed to the wakeup function * -* If this function wakes up a task, it executes a full memory barrier before -* accessing the task state. +* If this function wakes up a task, it executes a full memory barrier +* before accessing the task state. Returns the number of exclusive +* tasks that were awaken. */ -void __wake_up (struct wait_queue_head *wq_head,unsigned int mode, -int nr_exclusive,void *key) +int __wake_up (struct wait_queue_head *wq_head,unsigned int mode, +int nr_exclusive,void *key) { -__wake_up_common_lock (wq_head, mode, nr_exclusive,0 , key); +return __wake_up_common_lock (wq_head, mode, nr_exclusive,0 , key); }EXPORT_SYMBOL (__wake_up);
Signature · 1 line
2.35.3
Re: [PATCH 2/3] wait: Return number of exclusive waiters awaken
Quoted 6 lines · Gabriel Krisman Bertazi
Sbitmap code will need to know how many waiters were actually woken for its batched wakeups implementation. Return the number of woken exclusive waiters from
__wake_up()to facilitate that.Suggested-by Jan Kara <[email protected]>Signed-off-by Gabriel Krisman Bertazi <[email protected]>
Looks good. Feel free to add:
Honza
Quoted 73 lines
--- include/linux/wait.h | 2 +- kernel/sched/wait.c | 18 +++++++++++------- 2 files changed, 12 insertions(+), 8 deletions(-)include/linux/wait.h +1
-1@@ -209,7 +209,7 @@ __remove_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wqlist_del (&wq_entry->entry); } -void __wake_up (struct wait_queue_head *wq_head,unsigned int mode,int nr,void *key); +int __wake_up (struct wait_queue_head *wq_head,unsigned int mode,int nr,void *key);void __wake_up_locked_key (struct wait_queue_head *wq_head,unsigned int mode,void *key);void __wake_up_locked_key_bookmark (struct wait_queue_head *wq_head,unsigned int mode,void *key,wait_queue_entry_t *bookmark);kernel/sched/wait.c +11
-7@@ -121,11 +121,12 @@ static int __wake_up_common(struct wait_queue_head *wq_head, unsigned int mode,return nr_exclusive; } -static void __wake_up_common_lock (struct wait_queue_head *wq_head,unsigned int mode, +static int __wake_up_common_lock (struct wait_queue_head *wq_head,unsigned int mode,int nr_exclusive,int wake_flags,void *key) {unsigned long flags;wait_queue_entry_t bookmark; +int remaining = nr_exclusive; bookmark.flags =0 ; bookmark.private =NULL ;@@ -134,10 +135,12 @@ static void __wake_up_common_lock(struct wait_queue_head *wq_head, unsigned intdo {spin_lock_irqsave (&wq_head->lock, flags); - nr_exclusive =__wake_up_common (wq_head, mode, nr_exclusive, + remaining =__wake_up_common (wq_head, mode, remaining, wake_flags, key, &bookmark);spin_unlock_irqrestore (&wq_head->lock, flags); }while (bookmark.flags &WQ_FLAG_BOOKMARK ); + +return nr_exclusive - remaining; }/** @@ -147,13 +150,14 @@ static void __wake_up_common_lock(struct wait_queue_head *wq_head, unsigned int* @nr_exclusive: how many wake-one or wake-many threads to wake up * @key: is directly passed to the wakeup function * -* If this function wakes up a task, it executes a full memory barrier before -* accessing the task state. +* If this function wakes up a task, it executes a full memory barrier +* before accessing the task state. Returns the number of exclusive +* tasks that were awaken. */ -void __wake_up (struct wait_queue_head *wq_head,unsigned int mode, -int nr_exclusive,void *key) +int __wake_up (struct wait_queue_head *wq_head,unsigned int mode, +int nr_exclusive,void *key) { -__wake_up_common_lock (wq_head, mode, nr_exclusive,0 , key); +return __wake_up_common_lock (wq_head, mode, nr_exclusive,0 , key); }EXPORT_SYMBOL (__wake_up);Signature · 1 line
2.35.3
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Re: [PATCH 2/3] wait: Return number of exclusive waiters awaken
In reply to Gabriel Krisman BertaziQuoted 6 lines · Gabriel Krisman Bertazi
Sbitmap code will need to know how many waiters were actually woken for its batched wakeups implementation. Return the number of woken exclusive waiters from
__wake_up()to facilitate that.Suggested-by Jan Kara <[email protected]>Signed-off-by Gabriel Krisman Bertazi <[email protected]>
Quoted 4 lines
--- include/linux/wait.h | 2 +- kernel/sched/wait.c | 18 +++++++++++------- 2 files changed, 12 insertions(+), 8 deletions(-)
PATCH 3/3 sbitmap: Try each queue to wake up at least one waiter
In reply to Gabriel Krisman BertaziJan reported the new algorithm as merged might be problematic if the queue being awaken becomes empty between the waitqueue_active inside sbq_wake_ptr check and the wake up. If that happens, wake_up_nr will not wake up any waiter and we loose too many wake ups. In order to guarantee progress, we need to wake up at least one waiter here, if there are any. This now requires trying to wake up from every queue.
Instead of walking through all the queues with sbq_wake_ptr, this call moves the wake up inside that function. In a previous version of the patch, I found that updating wake_index several times when walking through queues had a measurable overhead. This ensures we only update it once, at the end.
--- lib/sbitmap.c | 28 ++++++++++++---------------- 1 file changed, 12 insertions(+), 16 deletions(-)
lib/sbitmap.c +12 -16
@@ -560,12 +560,12 @@ void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq, }EXPORT_SYMBOL_GPL (sbitmap_queue_min_shallow_depth); -static struct sbq_wait_state *sbq_wake_ptr (struct sbitmap_queue *sbq) +static void __sbitmap_queue_wake_up (struct sbitmap_queue *sbq,int nr) {int i, wake_index;if (!atomic_read (&sbq->ws_active)) -return NULL ; +return ; wake_index =atomic_read (&sbq->wake_index);for (i =0 ; i <SBQ_WAIT_QUEUES ; i++) {
@@ -579,20 +579,22 @@ static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)*/ wake_index =sbq_index_inc (wake_index); -if (waitqueue_active (&ws->wait)) { -if (wake_index !=atomic_read (&sbq->wake_index)) -atomic_set (&sbq->wake_index, wake_index); -return ws; - } +/* +* It is sufficient to wake up at least one waiter to +* guarantee forward progress. +*/ +if (waitqueue_active (&ws->wait) && +wake_up_nr (&ws->wait, nr)) +break ; } -return NULL ; +if (wake_index !=atomic_read (&sbq->wake_index)) +atomic_set (&sbq->wake_index, wake_index); }void sbitmap_queue_wake_up (struct sbitmap_queue *sbq,int nr) {unsigned int wake_batch =READ_ONCE (sbq->wake_batch); -struct sbq_wait_state *ws =NULL ;unsigned int wakeups;if (!atomic_read (&sbq->ws_active))
@@ -604,16 +606,10 @@ void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)do {if (atomic_read (&sbq->completion_cnt) - wakeups < wake_batch)return ; - -if (!ws) { - ws =sbq_wake_ptr (sbq); -if (!ws) -return ; - } }while (!atomic_try_cmpxchg (&sbq->wakeup_cnt, &wakeups, wakeups + wake_batch)); -wake_up_nr (&ws->wait, wake_batch); +__sbitmap_queue_wake_up (sbq, wake_batch); }EXPORT_SYMBOL_GPL (sbitmap_queue_wake_up);
Signature · 1 line
2.35.3
Re: [PATCH 3/3] sbitmap: Try each queue to wake up at least one waiter
Quoted 13 lines · Gabriel Krisman Bertazi
Jan reported the new algorithm as merged might be problematic if the queue being awaken becomes empty between the
waitqueue_activeinsidesbq_wake_ptrcheck and the wake up. If that happens,wake_up_nrwill not wake up any waiter and we loose too many wake ups. In order to guarantee progress, we need to wake up at least one waiter here, if there are any. This now requires trying to wake up from every queue.Instead of walking through all the queues with
sbq_wake_ptr, this call moves the wake up inside that function. In a previous version of the patch, I found that updatingwake_indexseveral times when walking through queues had a measurable overhead. This ensures we only update it once, at the end.
Quoted 3 lines
Fixes 4f8126bb2308 ("sbitmap: Use single per-bitmap counting to wake up queued tags")Reported-by Jan Kara <[email protected]>Signed-off-by Gabriel Krisman Bertazi <[email protected]>
Elegant and looks good to me! Feel free to add:
Honza
Quoted 74 lines
--- lib/sbitmap.c | 28 ++++++++++++---------------- 1 file changed, 12 insertions(+), 16 deletions(-)lib/sbitmap.c +12
-16@@ -560,12 +560,12 @@ void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq, }EXPORT_SYMBOL_GPL (sbitmap_queue_min_shallow_depth); -static struct sbq_wait_state *sbq_wake_ptr (struct sbitmap_queue *sbq) +static void __sbitmap_queue_wake_up (struct sbitmap_queue *sbq,int nr) {int i, wake_index;if (!atomic_read (&sbq->ws_active)) -return NULL ; +return ; wake_index =atomic_read (&sbq->wake_index);for (i =0 ; i <SBQ_WAIT_QUEUES ; i++) {@@ -579,20 +579,22 @@ static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)*/ wake_index =sbq_index_inc (wake_index); -if (waitqueue_active (&ws->wait)) { -if (wake_index !=atomic_read (&sbq->wake_index)) -atomic_set (&sbq->wake_index, wake_index); -return ws; - } +/* +* It is sufficient to wake up at least one waiter to +* guarantee forward progress. +*/ +if (waitqueue_active (&ws->wait) && +wake_up_nr (&ws->wait, nr)) +break ; } -return NULL ; +if (wake_index !=atomic_read (&sbq->wake_index)) +atomic_set (&sbq->wake_index, wake_index); }void sbitmap_queue_wake_up (struct sbitmap_queue *sbq,int nr) {unsigned int wake_batch =READ_ONCE (sbq->wake_batch); -struct sbq_wait_state *ws =NULL ;unsigned int wakeups;if (!atomic_read (&sbq->ws_active))@@ -604,16 +606,10 @@ void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)do {if (atomic_read (&sbq->completion_cnt) - wakeups < wake_batch)return ; - -if (!ws) { - ws =sbq_wake_ptr (sbq); -if (!ws) -return ; - } }while (!atomic_try_cmpxchg (&sbq->wakeup_cnt, &wakeups, wakeups + wake_batch)); -wake_up_nr (&ws->wait, wake_batch); +__sbitmap_queue_wake_up (sbq, wake_batch); }EXPORT_SYMBOL_GPL (sbitmap_queue_wake_up);Signature · 1 line
2.35.3
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Quoted 11 lines · Gabriel Krisman Bertazi
Jan reported two issues in the original thread.
The first is that
wake_indexwas not updated after returning fromsbq_wake_ptrwhich meant we'd have to empty the wq before moving to the next one. Patch 1/3 in this series reorders the code to avoid this condition, increasing fairness of queue selection and preventing starvation. I sent this patch already on the other thread and Jan reviewed it, but since it is a small one, and a dependency for the other, I'm resending it a along this series.[...]
Applied, thanks!
[1/3] sbitmap: Advance the queue index before waking up a queue
commit: 976570b4ecd30d3ec6e1b0910da8e5edc591f2b6
[2/3] wait: Return number of exclusive waiters awaken
commit: ee7dc86b6d3e3b86c2c487f713eda657850de238
[3/3] sbitmap: Try each queue to wake up at least one waiter
commit: 26edb30dd1c0c9be11fa676b4f330ada7b794ba6Best regards,
Signature · 1 line
Jens Axboe