| Message ID | 1401464642-33890-7-git-send-email-Waiman.Long@hp.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, May 30, 2014 at 11:43:52AM -0400, Waiman Long wrote: > --- > kernel/locking/qspinlock.c | 18 ++++++++++++++++-- > 1 files changed, 16 insertions(+), 2 deletions(-) > > diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c > index fc7fd8c..7f10758 100644 > --- a/kernel/locking/qspinlock.c > +++ b/kernel/locking/qspinlock.c > @@ -233,11 +233,25 @@ void queue_spin_lock_slowpath(struct qspinlock *lock, u32 val) > */ > for (;;) { > /* > - * If we observe any contention; queue. > + * If we observe that the queue is not empty or both > + * the pending and lock bits are set, queue > */ > - if (val & ~_Q_LOCKED_MASK) > + if ((val & _Q_TAIL_MASK) || > + (val == (_Q_LOCKED_VAL|_Q_PENDING_VAL))) > goto queue; > > + if (val == _Q_PENDING_VAL) { > + /* > + * Pending bit is set, but not the lock bit. > + * Assuming that the pending bit holder is going to > + * set the lock bit and clear the pending bit soon, > + * it is better to wait than to exit at this point. > + */ > + cpu_relax(); > + val = atomic_read(&lock->val); > + continue; > + } > + > new = _Q_LOCKED_VAL; > if (val == new) > new |= _Q_PENDING_VAL; So, again, you just posted a new version without replying to the previous discussion; so let me try again, what's wrong with the proposal here: lkml.kernel.org/r/20140417163640.GT11096@twins.programming.kicks-ass.net
On 6/11/2014 6:26 AM, Peter Zijlstra wrote: > On Fri, May 30, 2014 at 11:43:52AM -0400, Waiman Long wrote: >> --- >> kernel/locking/qspinlock.c | 18 ++++++++++++++++-- >> 1 files changed, 16 insertions(+), 2 deletions(-) >> >> diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c >> index fc7fd8c..7f10758 100644 >> --- a/kernel/locking/qspinlock.c >> +++ b/kernel/locking/qspinlock.c >> @@ -233,11 +233,25 @@ void queue_spin_lock_slowpath(struct qspinlock *lock, u32 val) >> */ >> for (;;) { >> /* >> - * If we observe any contention; queue. >> + * If we observe that the queue is not empty or both >> + * the pending and lock bits are set, queue >> */ >> - if (val & ~_Q_LOCKED_MASK) >> + if ((val & _Q_TAIL_MASK) || >> + (val == (_Q_LOCKED_VAL|_Q_PENDING_VAL))) >> goto queue; >> >> + if (val == _Q_PENDING_VAL) { >> + /* >> + * Pending bit is set, but not the lock bit. >> + * Assuming that the pending bit holder is going to >> + * set the lock bit and clear the pending bit soon, >> + * it is better to wait than to exit at this point. >> + */ >> + cpu_relax(); >> + val = atomic_read(&lock->val); >> + continue; >> + } >> + >> new = _Q_LOCKED_VAL; >> if (val == new) >> new |= _Q_PENDING_VAL; > > So, again, you just posted a new version without replying to the > previous discussion; so let me try again, what's wrong with the proposal > here: > > lkml.kernel.org/r/20140417163640.GT11096@twins.programming.kicks-ass.net > > I thought I had answered you before, maybe the message was lost or the answer was not complete. Anyway, I will try to response to your question again here. > Wouldn't something like: > > while (atomic_read(&lock->val) == _Q_PENDING_VAL) > cpu_relax(); > > before the cmpxchg loop have gotten you all this? That is not exactly the same. The loop will exit if other bits are set or the pending bit cleared. In the case, we will need to do the same check at the beginning of the for loop in order to avoid doing an extra cmpxchg that is not necessary. > I just tried this on my code and I cannot see a difference. As I said before, I did see a difference with that change. I think it depends on the CPU chip that we used for testing. I ran my test on a 10-core Westmere-EX chip. I run my microbench on different pairs of core within the same chip. It produces different results that varies from 779.5ms to up to 1192ms. Without that patch, the lowest value I can get is still close to 800ms, but the highest can be up to 1800ms or so. So I believe it is just a matter of timing that you did not observed in your test machine. -Longman -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Jun 11, 2014 at 05:22:28PM -0400, Long, Wai Man wrote: > >>@@ -233,11 +233,25 @@ void queue_spin_lock_slowpath(struct qspinlock *lock, u32 val) > >> */ > >> for (;;) { > >> /* > >>- * If we observe any contention; queue. > >>+ * If we observe that the queue is not empty or both > >>+ * the pending and lock bits are set, queue > >> */ > >>- if (val & ~_Q_LOCKED_MASK) > >>+ if ((val & _Q_TAIL_MASK) || > >>+ (val == (_Q_LOCKED_VAL|_Q_PENDING_VAL))) > >> goto queue; > >>+ if (val == _Q_PENDING_VAL) { > >>+ /* > >>+ * Pending bit is set, but not the lock bit. > >>+ * Assuming that the pending bit holder is going to > >>+ * set the lock bit and clear the pending bit soon, > >>+ * it is better to wait than to exit at this point. > >>+ */ > >>+ cpu_relax(); > >>+ val = atomic_read(&lock->val); > >>+ continue; > >>+ } > >>+ > >> new = _Q_LOCKED_VAL; > >> if (val == new) > >> new |= _Q_PENDING_VAL; > >Wouldn't something like: > > > > while (atomic_read(&lock->val) == _Q_PENDING_VAL) > > cpu_relax(); > > > >before the cmpxchg loop have gotten you all this? > > That is not exactly the same. The loop will exit if other bits are set or the pending > bit cleared. In the case, we will need to do the same check at the beginning of the > for loop in order to avoid doing an extra cmpxchg that is not necessary. If other bits get set we should stop poking at the pending bit and get queued. The only transition we want to wait for is: 0,1,0 -> 0,0,1. What extra unneeded cmpxchg() is there? If we have two cpus waiting in this loop for the pending bit to go away then both will attempt to grab the now free pending bit, one will loose and get queued? There's no avoiding that contention.
On 06/12/2014 02:00 AM, Peter Zijlstra wrote: > On Wed, Jun 11, 2014 at 05:22:28PM -0400, Long, Wai Man wrote: > >>>> @@ -233,11 +233,25 @@ void queue_spin_lock_slowpath(struct qspinlock *lock, u32 val) >>>> */ >>>> for (;;) { >>>> /* >>>> - * If we observe any contention; queue. >>>> + * If we observe that the queue is not empty or both >>>> + * the pending and lock bits are set, queue >>>> */ >>>> - if (val& ~_Q_LOCKED_MASK) >>>> + if ((val& _Q_TAIL_MASK) || >>>> + (val == (_Q_LOCKED_VAL|_Q_PENDING_VAL))) >>>> goto queue; >>>> + if (val == _Q_PENDING_VAL) { >>>> + /* >>>> + * Pending bit is set, but not the lock bit. >>>> + * Assuming that the pending bit holder is going to >>>> + * set the lock bit and clear the pending bit soon, >>>> + * it is better to wait than to exit at this point. >>>> + */ >>>> + cpu_relax(); >>>> + val = atomic_read(&lock->val); >>>> + continue; >>>> + } >>>> + >>>> new = _Q_LOCKED_VAL; >>>> if (val == new) >>>> new |= _Q_PENDING_VAL; >>> Wouldn't something like: >>> >>> while (atomic_read(&lock->val) == _Q_PENDING_VAL) >>> cpu_relax(); >>> >>> before the cmpxchg loop have gotten you all this? >> That is not exactly the same. The loop will exit if other bits are set or the pending >> bit cleared. In the case, we will need to do the same check at the beginning of the >> for loop in order to avoid doing an extra cmpxchg that is not necessary. > If other bits get set we should stop poking at the pending bit and get > queued. The only transition we want to wait for is: 0,1,0 -> 0,0,1. > > What extra unneeded cmpxchg() is there? If we have two cpus waiting in > this loop for the pending bit to go away then both will attempt to grab > the now free pending bit, one will loose and get queued? > > There's no avoiding that contention. If two tasks see the pending bit goes away and try to grab it with cmpxchg, there is no way we can avoid the contention. However, if some how the pending bit holder get the lock and another task set the pending bit before the current task, the spinlock value will become _Q_PENDING_VAL|_Q_LOCKED_VAL. The while loop will end and the code will blindly try to do a cmpxchg unless we check for this case before hand. This is what my code does by going back to the beginning of the for loop. -Longman -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Jun 12, 2014 at 04:54:52PM -0400, Waiman Long wrote: > If two tasks see the pending bit goes away and try to grab it with cmpxchg, > there is no way we can avoid the contention. However, if some how the > pending bit holder get the lock and another task set the pending bit before > the current task, the spinlock value will become > _Q_PENDING_VAL|_Q_LOCKED_VAL. The while loop will end and the code will > blindly try to do a cmpxchg unless we check for this case before hand. This > is what my code does by going back to the beginning of the for loop. There is already a test for that; see the goto queue; --- /* * wait for in-progress pending->locked hand-overs * * 0,1,0 -> 0,0,1 */ if (val == _Q_PENDING_VAL) { while ((val = atomic_read(&lock->val)) == _Q_PENDING_VAL) cpu_relax(); } /* * trylock || pending * * 0,0,0 -> 0,0,1 ; trylock * 0,0,1 -> 0,1,1 ; pending */ for (;;) { /* * If we observe any contention; queue. */ if (val & ~_Q_LOCKED_MASK) goto queue; new = _Q_LOCKED_VAL; if (val == new) new |= _Q_PENDING_VAL; old = atomic_cmpxchg(&lock->val, val, new); if (old == val) break; val = old; } -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c index fc7fd8c..7f10758 100644 --- a/kernel/locking/qspinlock.c +++ b/kernel/locking/qspinlock.c @@ -233,11 +233,25 @@ void queue_spin_lock_slowpath(struct qspinlock *lock, u32 val) */ for (;;) { /* - * If we observe any contention; queue. + * If we observe that the queue is not empty or both + * the pending and lock bits are set, queue */ - if (val & ~_Q_LOCKED_MASK) + if ((val & _Q_TAIL_MASK) || + (val == (_Q_LOCKED_VAL|_Q_PENDING_VAL))) goto queue; + if (val == _Q_PENDING_VAL) { + /* + * Pending bit is set, but not the lock bit. + * Assuming that the pending bit holder is going to + * set the lock bit and clear the pending bit soon, + * it is better to wait than to exit at this point. + */ + cpu_relax(); + val = atomic_read(&lock->val); + continue; + } + new = _Q_LOCKED_VAL; if (val == new) new |= _Q_PENDING_VAL;
There is a problem in the current pending bit spinning code. When the lock is free, but the pending bit holder hasn't grabbed the lock & cleared the pending bit yet, the spinning code will not be run. As a result, the regular queuing code path might be used most of the time even when there is only 2 tasks contending for the lock. Assuming that the pending bit holder is going to get the lock and clear the pending bit soon, it is actually better to wait than to be queued up which has a higher overhead. The following tables show the before-patch execution time (in ms) of a micro-benchmark where 5M iterations of the lock/unlock cycles were run on a 10-core Westere-EX x86-64 CPU with 2 different types of loads - standalone (lock and protected data in different cachelines) and embedded (lock and protected data in the same cacheline). [Standalone/Embedded - same node] # of tasks Ticket lock Queue lock %Change ---------- ----------- ---------- ------- 1 135/ 111 135/ 101 0%/ -9% 2 890/ 779 1885/1990 +112%/+156% 3 1932/1859 2333/2341 +21%/ +26% 4 2829/2726 2900/2923 +3%/ +7% 5 3834/3761 3655/3648 -5%/ -3% 6 4963/4976 4336/4326 -13%/ -13% 7 6299/6269 5057/5064 -20%/ -19% 8 7691/7569 5786/5798 -25%/ -23% Of course, the results will varies depending on what kind of test machine is used. With 1 task per NUMA node, the execution times are: [Standalone - different nodes] # of nodes Ticket lock Queue lock %Change ---------- ----------- ---------- ------- 1 135 135 0% 2 4604 5087 +10% 3 10940 12224 +12% 4 21555 10555 -51% It can be seen that the queue spinlock is slower than the ticket spinlock when there are 2 or 3 contending tasks. In all the other case, the queue spinlock is either equal or faster than the ticket spinlock. With this patch, the performance data for 2 contending tasks are: [Standalone/Embedded] # of tasks Ticket lock Queue lock %Change ---------- ----------- ---------- ------- 2 890/779 984/871 +11%/+12% [Standalone - different nodes] # of nodes Ticket lock Queue lock %Change ---------- ----------- ---------- ------- 2 4604 1364 -70% It can be seen that the queue spinlock performance for 2 contending tasks is now comparable to ticket spinlock on the same node, but much faster when in different nodes. With 3 contending tasks, however, the ticket spinlock is still quite a bit faster. Signed-off-by: Waiman Long <Waiman.Long@hp.com> --- kernel/locking/qspinlock.c | 18 ++++++++++++++++-- 1 files changed, 16 insertions(+), 2 deletions(-)