locking/rwsem: Allow slowpath writer to ignore handoff bit if not set by first waiter
Waiman Long
reviewerPATCH 1/1 locking/rwsem: Allow slowpath writer to ignore handoff bit if not set by first waiter
With commit d257cc8cb8d5 ("locking/rwsem: Make handoff bit handling more consistent"), the writer that sets the handoff bit can be interrupted out without clearing the bit if the wait queue isn't empty. This disables reader and writer optimistic lock spinning and stealing.
Now if a non-first writer in the queue is somehow woken up or first entering the waiting loop, it can't acquire the lock. This is not the case before that commit as the writer that set the handoff bit will clear it when exiting out via the out_nolock path. This is less efficient as the busy rwsem stays in an unlock state for a longer time.
This patch allows a non-first writer to ignore the handoff bit if it is not originally set or initiated by the first waiter.
--- kernel/locking/rwsem.c | 30 ++++++++++++++++++++---------- 1 file changed, 20 insertions(+), 10 deletions(-)
kernel/locking/rwsem.c +20 -10
@@ -335,8 +335,6 @@ struct rwsem_waiter {struct task_struct *task;enum rwsem_waiter_type type;unsigned long timeout; - -/* Writer only, not initialized in reader */ bool handoff_set; };#define rwsem_first_waiter (sem) \
@@ -459,10 +457,12 @@ static void rwsem_mark_wake(struct rw_semaphore *sem,* to give up the lock), request a HANDOFF to * force the issue. */ -if (!(oldcount &RWSEM_FLAG_HANDOFF ) && -time_after (jiffies, waiter->timeout)) { - adjustment -=RWSEM_FLAG_HANDOFF ; -lockevent_inc (rwsem_rlock_handoff); +if (time_after (jiffies, waiter->timeout)) { +if (!(oldcount &RWSEM_FLAG_HANDOFF )) { + adjustment -=RWSEM_FLAG_HANDOFF ; +lockevent_inc (rwsem_rlock_handoff); + } + waiter->handoff_set = true; }atomic_long_add (-adjustment, &sem->count);
@@ -599,7 +599,7 @@ rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,static inline bool rwsem_try_write_lock (struct rw_semaphore *sem,struct rwsem_waiter *waiter) { -bool first =rwsem_first_waiter (sem) == waiter; +struct rwsem_waiter *first =rwsem_first_waiter (sem);long count, new;lockdep_assert_held (&sem->wait_lock);
@@ -609,11 +609,20 @@ static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,bool has_handoff = !!(count &RWSEM_FLAG_HANDOFF );if (has_handoff) { -if (!first) +/* +* Honor handoff bit and yield only when the first +* waiter is the one that set it. Otherwisee, we +* still try to acquire the rwsem. +*/ +if (first->handoff_set && (waiter != first))return false; -/* First waiter inherits a previously set handoff bit */ - waiter->handoff_set = true; +/* +* First waiter can inherit a previously set handoff +* bit and spin on rwsem if lock acquisition fails. +*/ +if (waiter == first) + waiter->handoff_set = true; } new = count;
@@ -1027,6 +1036,7 @@ rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int stat waiter.task = current; waiter.type =RWSEM_WAITING_FOR_READ ; waiter.timeout = jiffies +RWSEM_WAIT_TIMEOUT ; + waiter.handoff_set = false;raw_spin_lock_irq (&sem->wait_lock);if (list_empty (&sem->wait_list)) {
Signature · 1 line
2.27.0
Quoted 11 lines · Waiman Long
With commit
d257cc8cb8d5("locking/rwsem: Make handoff bit handling more consistent"), the writer that sets the handoff bit can be interrupted out without clearing the bit if the wait queue isn't empty. This disables reader and writer optimistic lock spinning and stealing.Now if a non-first writer in the queue is somehow woken up or first entering the waiting loop, it can't acquire the lock. This is not the case before that commit as the writer that set the handoff bit will clear it when exiting out via the
out_nolockpath. This is less efficient as the busy rwsem stays in an unlock state for a longer time.
Quoted 4 lines
This patch allows a non-first writer to ignore the handoff bit if it is not originally set or initiated by the first waiter.
Fixes d257cc8cb8d5 ("locking/rwsem: Make handoff bit handling more consistent")
Hi Waiman,
1. You likely need :
Since d257cc8cb8d5 has been migrated to other LTS threads (5.15.y for sure) .
2. I had posted this earlier:
[PATCH 5.15 1/1] Revert "locking/rwsem: Make handoff bit handling more consistent"
That is likely not needed now.
3. Please add :
We can likely test this, but I can't give a ETA when that will happen.
Quoted 1 line
Signed-off-by Waiman Long <[email protected]>
Quoted 75 lines
--- kernel/locking/rwsem.c | 30 ++++++++++++++++++++---------- 1 file changed, 20 insertions(+), 10 deletions(-)kernel/locking/rwsem.c +20
-10@@ -335,8 +335,6 @@ struct rwsem_waiter {struct task_struct *task;enum rwsem_waiter_type type;unsigned long timeout; - -/* Writer only, not initialized in reader */ bool handoff_set; };#define rwsem_first_waiter (sem) \@@ -459,10 +457,12 @@ static void rwsem_mark_wake(struct rw_semaphore *sem,* to give up the lock), request a HANDOFF to * force the issue. */ -if (!(oldcount &RWSEM_FLAG_HANDOFF ) && -time_after (jiffies, waiter->timeout)) { - adjustment -=RWSEM_FLAG_HANDOFF ; -lockevent_inc (rwsem_rlock_handoff); +if (time_after (jiffies, waiter->timeout)) { +if (!(oldcount &RWSEM_FLAG_HANDOFF )) { + adjustment -=RWSEM_FLAG_HANDOFF ; +lockevent_inc (rwsem_rlock_handoff); + } + waiter->handoff_set = true; }atomic_long_add (-adjustment, &sem->count);@@ -599,7 +599,7 @@ rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,static inline bool rwsem_try_write_lock (struct rw_semaphore *sem,struct rwsem_waiter *waiter) { -bool first =rwsem_first_waiter (sem) == waiter; +struct rwsem_waiter *first =rwsem_first_waiter (sem);long count, new;lockdep_assert_held (&sem->wait_lock);@@ -609,11 +609,20 @@ static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,bool has_handoff = !!(count &RWSEM_FLAG_HANDOFF );if (has_handoff) { -if (!first) +/* +* Honor handoff bit and yield only when the first +* waiter is the one that set it. Otherwisee, we +* still try to acquire the rwsem. +*/ +if (first->handoff_set && (waiter != first))return false; -/* First waiter inherits a previously set handoff bit */ - waiter->handoff_set = true; +/* +* First waiter can inherit a previously set handoff +* bit and spin on rwsem if lock acquisition fails. +*/ +if (waiter == first) + waiter->handoff_set = true; } new = count;@@ -1027,6 +1036,7 @@ rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int stat waiter.task = current; waiter.type =RWSEM_WAITING_FOR_READ ; waiter.timeout = jiffies +RWSEM_WAIT_TIMEOUT ; + waiter.handoff_set = false;raw_spin_lock_irq (&sem->wait_lock);if (list_empty (&sem->wait_list)) {
Waiman Long
reviewerQuoted 17 lines · John Donnelly
Quoted 11 lines · Waiman Long
With commit
d257cc8cb8d5("locking/rwsem: Make handoff bit handling more consistent"), the writer that sets the handoff bit can be interrupted out without clearing the bit if the wait queue isn't empty. This disables reader and writer optimistic lock spinning and stealing.Now if a non-first writer in the queue is somehow woken up or first entering the waiting loop, it can't acquire the lock. This is not the case before that commit as the writer that set the handoff bit will clear it when exiting out via the
out_nolockpath. This is less efficient as the busy rwsem stays in an unlock state for a longer time.Quoted 4 lines
This patch allows a non-first writer to ignore the handoff bit if it is not originally set or initiated by the first waiter.
Fixes d257cc8cb8d5 ("locking/rwsem: Make handoff bit handling more consistent")
Quoted 7 lines
Hi Waiman,
1. You likely need :
Cc <[email protected]>Since
d257cc8cb8d5has been migrated to other LTS threads (5.15.y for sure) .
Since the commit has a fixes tag, the stable will automatically pick it up.
Quoted 6 lines
2. I had posted this earlier:
[PATCH 5.15 1/1] Revert "locking/rwsem: Make handoff bit handling more consistent"
That is likely not needed now.
Right. The patch was created to fix a problem reported by another user. So reverting it may not the right move.
Quoted 4 lines
3. Please add :
Quoted 8 lines
Reported-by Jorge Lopez <[email protected]>We can likely test this, but I can't give a ETA when that will happen.
Quoted 1 line
Signed-off-by Waiman Long <[email protected]>Acked-by John Donnelly <[email protected]>
Will add the that when I send out the next version.
Cheers,
Longman
Quoted 29 lines · Waiman Long
Quoted 17 lines · John Donnelly
Quoted 11 lines · Waiman Long
With commit
d257cc8cb8d5("locking/rwsem: Make handoff bit handling more consistent"), the writer that sets the handoff bit can be interrupted out without clearing the bit if the wait queue isn't empty. This disables reader and writer optimistic lock spinning and stealing.Now if a non-first writer in the queue is somehow woken up or first entering the waiting loop, it can't acquire the lock. This is not the case before that commit as the writer that set the handoff bit will clear it when exiting out via the
out_nolockpath. This is less efficient as the busy rwsem stays in an unlock state for a longer time.Quoted 4 lines
This patch allows a non-first writer to ignore the handoff bit if it is not originally set or initiated by the first waiter.
Fixes d257cc8cb8d5 ("locking/rwsem: Make handoff bit handling more consistent")Quoted 7 lines
Hi Waiman,
1. You likely need :
Cc <[email protected]>Since
d257cc8cb8d5has been migrated to other LTS threads (5.15.y for sure) .Since the commit has a fixes tag, the stable will automatically pick it up.
Quoted 1 line
Quoted 6 lines
Quoted 5 lines
2. I had posted this earlier:
[PATCH 5.15 1/1] Revert "locking/rwsem: Make handoff bit handling more consistent"
That is likely not needed now.
Right. The patch was created to fix a problem reported by another user. So reverting it may not the right move.
Thanks for the quick follow up.
Quoted 18 lines
Quoted 4 lines
3. Please add :
Quoted 8 lines
Reported-by Jorge Lopez <[email protected]>We can likely test this, but I can't give a ETA when that will happen.
Quoted 1 line
Signed-off-by Waiman Long <[email protected]>Acked-by John Donnelly <[email protected]>Will add the that when I send out the next version.
Cheers,
Longman