linux/lib/rwsem.c

294 lines
7.9 KiB
C
Raw Normal View History

/* rwsem.c: R/W semaphores: contention handling functions
*
* Written by David Howells (dhowells@redhat.com).
* Derived from arch/i386/kernel/semaphore.c
rwsem: Implement writer lock-stealing for better scalability Commit 5a505085f043 ("mm/rmap: Convert the struct anon_vma::mutex to an rwsem") changed struct anon_vma::mutex to an rwsem, which caused aim7 fork_test performance to drop by 50%. Yuanhan Liu did the following excellent analysis: https://lkml.org/lkml/2013/1/29/84 and found that the regression is caused by strict, serialized, FIFO sequential write-ownership of rwsems. Ingo suggested implementing opportunistic lock-stealing for the front writer task in the waitqueue. Yuanhan Liu implemented lock-stealing for spinlock-rwsems, which indeed recovered much of the regression - confirming the analysis that the main factor in the regression was the FIFO writer-fairness of rwsems. In this patch we allow lock-stealing to happen when the first waiter is also writer. With that change in place the aim7 fork_test performance is fully recovered on my Intel NHM EP, NHM EX, SNB EP 2S and 4S test-machines. Reported-by: lkp@linux.intel.com Reported-by: Yuanhan Liu <yuanhan.liu@linux.intel.com> Signed-off-by: Alex Shi <alex.shi@intel.com> Cc: David Howells <dhowells@redhat.com> Cc: Michel Lespinasse <walken@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Anton Blanchard <anton@samba.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: paul.gortmaker@windriver.com Link: https://lkml.org/lkml/2013/1/29/84 Link: http://lkml.kernel.org/r/1360069915-31619-1-git-send-email-alex.shi@intel.com [ Small stylistic fixes, updated changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-02-05 13:11:55 +00:00
*
* Writer lock-stealing by Alex Shi <alex.shi@intel.com>
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
* and Michel Lespinasse <walken@google.com>
*/
#include <linux/rwsem.h>
#include <linux/sched.h>
#include <linux/init.h>
#include <linux/export.h>
/*
* Initialize an rwsem:
*/
void __init_rwsem(struct rw_semaphore *sem, const char *name,
struct lock_class_key *key)
{
#ifdef CONFIG_DEBUG_LOCK_ALLOC
/*
* Make sure we are not reinitializing a held semaphore:
*/
debug_check_no_locks_freed((void *)sem, sizeof(*sem));
lockdep_init_map(&sem->dep_map, name, key, 0);
#endif
sem->count = RWSEM_UNLOCKED_VALUE;
raw_spin_lock_init(&sem->wait_lock);
INIT_LIST_HEAD(&sem->wait_list);
}
EXPORT_SYMBOL(__init_rwsem);
enum rwsem_waiter_type {
RWSEM_WAITING_FOR_WRITE,
RWSEM_WAITING_FOR_READ
};
struct rwsem_waiter {
struct list_head list;
struct task_struct *task;
enum rwsem_waiter_type type;
};
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
enum rwsem_wake_type {
RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */
RWSEM_WAKE_READERS, /* Wake readers only */
RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */
};
/*
* handle the lock release when processes blocked on it that can now run
* - if we come here from up_xxxx(), then:
* - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed)
* - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so)
* - there must be someone on the queue
* - the spinlock must be held by the caller
* - woken process blocks are discarded from the list after having task zeroed
* - writers are only woken if downgrading is false
*/
static struct rw_semaphore *
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
__rwsem_do_wake(struct rw_semaphore *sem, enum rwsem_wake_type wake_type)
{
struct rwsem_waiter *waiter;
struct task_struct *tsk;
struct list_head *next;
long oldcount, woken, loop, adjustment;
waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
if (wake_type == RWSEM_WAKE_ANY)
/* Wake writer at the front of the queue, but do not
* grant it the lock yet as we want other writers
* to be able to steal it. Readers, on the other hand,
* will block as they will notice the queued writer.
*/
wake_up_process(waiter->task);
goto out;
}
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
/* Writers might steal the lock before we grant it to the next reader.
* We prefer to do the first reader grant before counting readers
* so we can bail out early if a writer stole the lock.
*/
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
adjustment = 0;
if (wake_type != RWSEM_WAKE_READ_OWNED) {
adjustment = RWSEM_ACTIVE_READ_BIAS;
try_reader_grant:
oldcount = rwsem_atomic_update(adjustment, sem) - adjustment;
if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
/* A writer stole the lock. Undo our reader grant. */
if (rwsem_atomic_update(-adjustment, sem) &
RWSEM_ACTIVE_MASK)
goto out;
/* Last active locker left. Retry waking readers. */
goto try_reader_grant;
}
}
/* Grant an infinite number of read locks to the readers at the front
* of the queue. Note we increment the 'active part' of the count by
* the number of readers before waking any processes up.
*/
woken = 0;
do {
woken++;
if (waiter->list.next == &sem->wait_list)
break;
waiter = list_entry(waiter->list.next,
struct rwsem_waiter, list);
} while (waiter->type != RWSEM_WAITING_FOR_WRITE);
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
if (waiter->type != RWSEM_WAITING_FOR_WRITE)
/* hit end of list above */
adjustment -= RWSEM_WAITING_BIAS;
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
if (adjustment)
rwsem_atomic_add(adjustment, sem);
next = sem->wait_list.next;
loop = woken;
do {
waiter = list_entry(next, struct rwsem_waiter, list);
next = waiter->list.next;
tsk = waiter->task;
smp_mb();
waiter->task = NULL;
wake_up_process(tsk);
put_task_struct(tsk);
} while (--loop);
sem->wait_list.next = next;
next->prev = &sem->wait_list;
out:
return sem;
rwsem: Implement writer lock-stealing for better scalability Commit 5a505085f043 ("mm/rmap: Convert the struct anon_vma::mutex to an rwsem") changed struct anon_vma::mutex to an rwsem, which caused aim7 fork_test performance to drop by 50%. Yuanhan Liu did the following excellent analysis: https://lkml.org/lkml/2013/1/29/84 and found that the regression is caused by strict, serialized, FIFO sequential write-ownership of rwsems. Ingo suggested implementing opportunistic lock-stealing for the front writer task in the waitqueue. Yuanhan Liu implemented lock-stealing for spinlock-rwsems, which indeed recovered much of the regression - confirming the analysis that the main factor in the regression was the FIFO writer-fairness of rwsems. In this patch we allow lock-stealing to happen when the first waiter is also writer. With that change in place the aim7 fork_test performance is fully recovered on my Intel NHM EP, NHM EX, SNB EP 2S and 4S test-machines. Reported-by: lkp@linux.intel.com Reported-by: Yuanhan Liu <yuanhan.liu@linux.intel.com> Signed-off-by: Alex Shi <alex.shi@intel.com> Cc: David Howells <dhowells@redhat.com> Cc: Michel Lespinasse <walken@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Anton Blanchard <anton@samba.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: paul.gortmaker@windriver.com Link: https://lkml.org/lkml/2013/1/29/84 Link: http://lkml.kernel.org/r/1360069915-31619-1-git-send-email-alex.shi@intel.com [ Small stylistic fixes, updated changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-02-05 13:11:55 +00:00
}
/*
* wait for the read lock to be granted
*/
struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
{
long count, adjustment = -RWSEM_ACTIVE_READ_BIAS;
struct rwsem_waiter waiter;
struct task_struct *tsk = current;
/* set up my own style of waitqueue */
waiter.task = tsk;
waiter.type = RWSEM_WAITING_FOR_READ;
get_task_struct(tsk);
raw_spin_lock_irq(&sem->wait_lock);
if (list_empty(&sem->wait_list))
adjustment += RWSEM_WAITING_BIAS;
list_add_tail(&waiter.list, &sem->wait_list);
/* we're now waiting on the lock, but no longer actively locking */
count = rwsem_atomic_update(adjustment, sem);
/* If there are no active locks, wake the front queued process(es).
*
* If there are no writers and we are first in the queue,
* wake our own waiter to join the existing active readers !
*/
if (count == RWSEM_WAITING_BIAS ||
(count > RWSEM_WAITING_BIAS &&
adjustment != -RWSEM_ACTIVE_READ_BIAS))
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
raw_spin_unlock_irq(&sem->wait_lock);
/* wait to be given the lock */
while (true) {
set_task_state(tsk, TASK_UNINTERRUPTIBLE);
if (!waiter.task)
break;
schedule();
}
tsk->state = TASK_RUNNING;
return sem;
}
/*
* wait until we successfully acquire the write lock
*/
struct rw_semaphore __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
{
long count, adjustment = -RWSEM_ACTIVE_WRITE_BIAS;
struct rwsem_waiter waiter;
struct task_struct *tsk = current;
/* set up my own style of waitqueue */
waiter.task = tsk;
waiter.type = RWSEM_WAITING_FOR_WRITE;
raw_spin_lock_irq(&sem->wait_lock);
if (list_empty(&sem->wait_list))
adjustment += RWSEM_WAITING_BIAS;
list_add_tail(&waiter.list, &sem->wait_list);
/* we're now waiting on the lock, but no longer actively locking */
count = rwsem_atomic_update(adjustment, sem);
/* If there were already threads queued before us and there are no
* active writers, the lock must be read owned; so we try to wake
* any read locks that were queued ahead of us. */
if (count > RWSEM_WAITING_BIAS &&
adjustment == -RWSEM_ACTIVE_WRITE_BIAS)
rwsem: implement support for write lock stealing on the fastpath When we decide to wake up readers, we must first grant them as many read locks as necessary, and then actually wake up all these readers. But in order to know how many read shares to grant, we must first count the readers at the head of the queue. This might take a while if there are many readers, and we want to be protected against a writer stealing the lock while we're counting. To that end, we grant the first reader lock before counting how many more readers are queued. We also require some adjustments to the wake_type semantics. RWSEM_WAKE_NO_ACTIVE used to mean that we had found the count to be RWSEM_WAITING_BIAS, in which case the rwsem was known to be free as nobody could steal it while we hold the wait_lock. This doesn't make sense once we implement fastpath write lock stealing, so we now use RWSEM_WAKE_ANY in that case. Similarly, when rwsem_down_write_failed found that a read lock was active, it would use RWSEM_WAKE_READ_OWNED which signalled that new readers could be woken without checking first that the rwsem was available. We can't do that anymore since the existing readers might release their read locks, and a writer could steal the lock before we wake up additional readers. So, we have to use a new RWSEM_WAKE_READERS value to indicate we only want to wake readers, but we don't currently hold any read lock. Signed-off-by: Michel Lespinasse <walken@google.com> Reviewed-by: Peter Hurley <peter@hurleysoftware.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 13:45:59 +00:00
sem = __rwsem_do_wake(sem, RWSEM_WAKE_READERS);
/* wait until we successfully acquire the lock */
set_task_state(tsk, TASK_UNINTERRUPTIBLE);
while (true) {
if (!(count & RWSEM_ACTIVE_MASK)) {
/* Try acquiring the write lock. */
count = RWSEM_ACTIVE_WRITE_BIAS;
if (!list_is_singular(&sem->wait_list))
count += RWSEM_WAITING_BIAS;
rwsem: check counter to avoid cmpxchg calls This patch tries to reduce the amount of cmpxchg calls in the writer failed path by checking the counter value first before issuing the instruction. If ->count is not set to RWSEM_WAITING_BIAS then there is no point wasting a cmpxchg call. Furthermore, Michel states "I suppose it helps due to the case where someone else steals the lock while we're trying to acquire sem->wait_lock." Two very different workloads and machines were used to see how this patch improves throughput: pgbench on a quad-core laptop and aim7 on a large 8 socket box with 80 cores. Some results comparing Michel's fast-path write lock stealing (tps-rwsem) on a quad-core laptop running pgbench: | db_size | clients | tps-rwsem | tps-patch | +---------+----------+----------------+--------------+ | 160 MB | 1 | 6906 | 9153 | + 32.5 | 160 MB | 2 | 15931 | 22487 | + 41.1% | 160 MB | 4 | 33021 | 32503 | | 160 MB | 8 | 34626 | 34695 | | 160 MB | 16 | 33098 | 34003 | | 160 MB | 20 | 31343 | 31440 | | 160 MB | 30 | 28961 | 28987 | | 160 MB | 40 | 26902 | 26970 | | 160 MB | 50 | 25760 | 25810 | ------------------------------------------------------ | 1.6 GB | 1 | 7729 | 7537 | | 1.6 GB | 2 | 19009 | 23508 | + 23.7% | 1.6 GB | 4 | 33185 | 32666 | | 1.6 GB | 8 | 34550 | 34318 | | 1.6 GB | 16 | 33079 | 32689 | | 1.6 GB | 20 | 31494 | 31702 | | 1.6 GB | 30 | 28535 | 28755 | | 1.6 GB | 40 | 27054 | 27017 | | 1.6 GB | 50 | 25591 | 25560 | ------------------------------------------------------ | 7.6 GB | 1 | 6224 | 7469 | + 20.0% | 7.6 GB | 2 | 13611 | 12778 | | 7.6 GB | 4 | 33108 | 32927 | | 7.6 GB | 8 | 34712 | 34878 | | 7.6 GB | 16 | 32895 | 33003 | | 7.6 GB | 20 | 31689 | 31974 | | 7.6 GB | 30 | 29003 | 28806 | | 7.6 GB | 40 | 26683 | 26976 | | 7.6 GB | 50 | 25925 | 25652 | ------------------------------------------------------ For the aim7 worloads, they overall improved on top of Michel's patchset. For full graphs on how the rwsem series plus this patch behaves on a large 8 socket machine against a vanilla kernel: http://stgolabs.net/rwsem-aim7-results.tar.gz Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 22:39:03 +00:00
if (sem->count == RWSEM_WAITING_BIAS &&
cmpxchg(&sem->count, RWSEM_WAITING_BIAS, count) ==
RWSEM_WAITING_BIAS)
break;
}
raw_spin_unlock_irq(&sem->wait_lock);
/* Block until there are no active lockers. */
do {
schedule();
set_task_state(tsk, TASK_UNINTERRUPTIBLE);
} while ((count = sem->count) & RWSEM_ACTIVE_MASK);
raw_spin_lock_irq(&sem->wait_lock);
}
list_del(&waiter.list);
raw_spin_unlock_irq(&sem->wait_lock);
tsk->state = TASK_RUNNING;
return sem;
}
/*
* handle waking up a waiter on the semaphore
* - up_read/up_write has decremented the active part of count if we come here
*/
struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
{
unsigned long flags;
raw_spin_lock_irqsave(&sem->wait_lock, flags);
/* do nothing if list empty */
if (!list_empty(&sem->wait_list))
sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
return sem;
}
/*
* downgrade a write lock into a read lock
* - caller incremented waiting part of count and discovered it still negative
* - just wake up any readers at the front of the queue
*/
struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
{
unsigned long flags;
raw_spin_lock_irqsave(&sem->wait_lock, flags);
/* do nothing if list empty */
if (!list_empty(&sem->wait_list))
sem = __rwsem_do_wake(sem, RWSEM_WAKE_READ_OWNED);
raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
return sem;
}
EXPORT_SYMBOL(rwsem_down_read_failed);
EXPORT_SYMBOL(rwsem_down_write_failed);
EXPORT_SYMBOL(rwsem_wake);
EXPORT_SYMBOL(rwsem_downgrade_wake);