]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
locking/rwsem: Make handoff writer optimistically spin on owner
authorWaiman Long <longman@redhat.com>
Tue, 25 Jun 2019 14:39:13 +0000 (10:39 -0400)
committerPeter Zijlstra <peterz@infradead.org>
Tue, 6 Aug 2019 10:49:15 +0000 (12:49 +0200)
When the handoff bit is set by a writer, no other tasks other than
the setting writer itself is allowed to acquire the lock. If the
to-be-handoff'ed writer goes to sleep, there will be a wakeup latency
period where the lock is free, but no one can acquire it. That is less
than ideal.

To reduce that latency, the handoff writer will now optimistically spin
on the owner if it happens to be a on-cpu writer. It will spin until
it releases the lock and the to-be-handoff'ed writer can then acquire
the lock immediately without any delay. Of course, if the owner is not
a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway.

The optimistic spinning code is also modified to not stop spinning
when the handoff bit is set. This will prevent an occasional setting of
handoff bit from causing a bunch of optimistic spinners from entering
into the wait queue causing significant reduction in throughput.

On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory
workload was run with 7000 users. The throughput (jobs/min) of the
following kernels were as follows:

 1) 5.2-rc6
    - 8,092,486
 2) 5.2-rc6 + tip's rwsem patches
    - 7,567,568
 3) 5.2-rc6 + tip's rwsem patches + this patch
    - 7,954,545

Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(),
rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%,
5.46% and 2.08% respectively.

Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: x86@kernel.org
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: huang ying <huang.ying.caritas@gmail.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Link: https://lkml.kernel.org/r/20190625143913.24154-1-longman@redhat.com
kernel/locking/rwsem.c

index bd0f0d05724ca943649788f5a72bd1c36f0bb567..354238a08b7aea46fcd3394ae7a9239a8634fe43 100644 (file)
@@ -724,11 +724,12 @@ rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
 
        rcu_read_lock();
        for (;;) {
-               if (atomic_long_read(&sem->count) & RWSEM_FLAG_HANDOFF) {
-                       state = OWNER_NONSPINNABLE;
-                       break;
-               }
-
+               /*
+                * When a waiting writer set the handoff flag, it may spin
+                * on the owner as well. Once that writer acquires the lock,
+                * we can spin on it. So we don't need to quit even when the
+                * handoff bit is set.
+                */
                new = rwsem_owner_flags(sem, &new_flags);
                if ((new != owner) || (new_flags != flags)) {
                        state = rwsem_owner_state(new, new_flags, nonspinnable);
@@ -974,6 +975,13 @@ static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
 {
        return false;
 }
+
+static inline int
+rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
+{
+       return 0;
+}
+#define OWNER_NULL     1
 #endif
 
 /*
@@ -1206,6 +1214,18 @@ wait:
 
                raw_spin_unlock_irq(&sem->wait_lock);
 
+               /*
+                * After setting the handoff bit and failing to acquire
+                * the lock, attempt to spin on owner to accelerate lock
+                * transfer. If the previous owner is a on-cpu writer and it
+                * has just released the lock, OWNER_NULL will be returned.
+                * In this case, we attempt to acquire the lock again
+                * without sleeping.
+                */
+               if ((wstate == WRITER_HANDOFF) &&
+                   (rwsem_spin_on_owner(sem, 0) == OWNER_NULL))
+                       goto trylock_again;
+
                /* Block until there are no active lockers. */
                for (;;) {
                        if (signal_pending_state(state, current))
@@ -1240,7 +1260,7 @@ wait:
                                break;
                        }
                }
-
+trylock_again:
                raw_spin_lock_irq(&sem->wait_lock);
        }
        __set_current_state(TASK_RUNNING);