locking/rwsem: Scan the wait_list for readers only once
authorDavidlohr Bueso <dave@stgolabs.net>
Fri, 5 Aug 2016 08:04:45 +0000 (01:04 -0700)
committerIngo Molnar <mingo@kernel.org>
Thu, 18 Aug 2016 13:37:11 +0000 (15:37 +0200)
When wanting to wakeup readers, __rwsem_mark_wakeup() currently
iterates the wait_list twice while looking to wakeup the first N
queued reader-tasks. While this can be quite inefficient, it was
there such that a awoken reader would be first and foremost
acknowledged by the lock counter.

Keeping the same logic, we can further benefit from the use of
wake_qs and avoid entirely the first wait_list iteration that sets
the counter as wake_up_process() isn't going to occur right away,
and therefore we maintain the counter->list order of going about
things.

Other than saving cycles with O(n) "scanning", this change also
nicely cleans up a good chunk of __rwsem_mark_wakeup(); both
visually and less tedious to read.

For example, the following improvements where seen on some will
it scale microbenchmarks, on a 48-core Haswell:

                                       v4.7              v4.7-rwsem-v1
  Hmean    signal1-processes-8    5792691.42 (  0.00%)  5771971.04 ( -0.36%)
  Hmean    signal1-processes-12   6081199.96 (  0.00%)  6072174.38 ( -0.15%)
  Hmean    signal1-processes-21   3071137.71 (  0.00%)  3041336.72 ( -0.97%)
  Hmean    signal1-processes-48   3712039.98 (  0.00%)  3708113.59 ( -0.11%)
  Hmean    signal1-processes-79   4464573.45 (  0.00%)  4682798.66 (  4.89%)
  Hmean    signal1-processes-110  4486842.01 (  0.00%)  4633781.71 (  3.27%)
  Hmean    signal1-processes-141  4611816.83 (  0.00%)  4692725.38 (  1.75%)
  Hmean    signal1-processes-172  4638157.05 (  0.00%)  4714387.86 (  1.64%)
  Hmean    signal1-processes-203  4465077.80 (  0.00%)  4690348.07 (  5.05%)
  Hmean    signal1-processes-224  4410433.74 (  0.00%)  4687534.43 (  6.28%)

  Stddev   signal1-processes-8       6360.47 (  0.00%)     8455.31 ( 32.94%)
  Stddev   signal1-processes-12      4004.98 (  0.00%)     9156.13 (128.62%)
  Stddev   signal1-processes-21      3273.14 (  0.00%)     5016.80 ( 53.27%)
  Stddev   signal1-processes-48     28420.25 (  0.00%)    26576.22 ( -6.49%)
  Stddev   signal1-processes-79     22038.34 (  0.00%)    18992.70 (-13.82%)
  Stddev   signal1-processes-110    23226.93 (  0.00%)    17245.79 (-25.75%)
  Stddev   signal1-processes-141     6358.98 (  0.00%)     7636.14 ( 20.08%)
  Stddev   signal1-processes-172     9523.70 (  0.00%)     4824.75 (-49.34%)
  Stddev   signal1-processes-203    13915.33 (  0.00%)     9326.33 (-32.98%)
  Stddev   signal1-processes-224    15573.94 (  0.00%)    10613.82 (-31.85%)

Other runs that saw improvements include context_switch and pipe; and
as expected, this is particularly highlighted on larger thread counts
as it becomes more expensive to walk the list twice.

No change in wakeup ordering or semantics.

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman.Long@hp.com
Cc: dave@stgolabs.net
Cc: jason.low2@hpe.com
Cc: wanpeng.li@hotmail.com
Link: http://lkml.kernel.org/r/1470384285-32163-4-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/locking/rwsem-xadd.c

index e02fe3289b5a9e43afa661196c247126c3434391..2337b4bb2366df1c2e21ada2ddb5d4b024aae890 100644 (file)
@@ -125,12 +125,14 @@ static void __rwsem_mark_wake(struct rw_semaphore *sem,
                              enum rwsem_wake_type wake_type,
                              struct wake_q_head *wake_q)
 {
-       struct rwsem_waiter *waiter;
-       struct task_struct *tsk;
-       struct list_head *next;
-       long loop, oldcount, woken = 0, adjustment = 0;
+       struct rwsem_waiter *waiter, *tmp;
+       long oldcount, woken = 0, adjustment = 0;
 
-       waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
+       /*
+        * Take a peek at the queue head waiter such that we can determine
+        * the wakeup(s) to perform.
+        */
+       waiter = list_first_entry(&sem->wait_list, struct rwsem_waiter, list);
 
        if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
                if (wake_type == RWSEM_WAKE_ANY) {
@@ -180,36 +182,21 @@ static void __rwsem_mark_wake(struct rw_semaphore *sem,
 
        /*
         * 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.
+        * of the queue. We know that woken will be at least 1 as we accounted
+        * for above. Note we increment the 'active part' of the count by the
+        * number of readers before waking any processes up.
         */
-       do {
-               woken++;
+       list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
+               struct task_struct *tsk;
 
-               if (waiter->list.next == &sem->wait_list)
+               if (waiter->type == RWSEM_WAITING_FOR_WRITE)
                        break;
 
-               waiter = list_entry(waiter->list.next,
-                                       struct rwsem_waiter, list);
-
-       } while (waiter->type != RWSEM_WAITING_FOR_WRITE);
-
-       adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
-       if (waiter->type != RWSEM_WAITING_FOR_WRITE)
-               /* hit end of list above */
-               adjustment -= RWSEM_WAITING_BIAS;
-
-       if (adjustment)
-               atomic_long_add(adjustment, &sem->count);
-
-       next = sem->wait_list.next;
-       loop = woken;
-       do {
-               waiter = list_entry(next, struct rwsem_waiter, list);
-               next = waiter->list.next;
+               woken++;
                tsk = waiter->task;
 
                wake_q_add(wake_q, tsk);
+               list_del(&waiter->list);
                /*
                 * Ensure that the last operation is setting the reader
                 * waiter to nil such that rwsem_down_read_failed() cannot
@@ -217,10 +204,16 @@ static void __rwsem_mark_wake(struct rw_semaphore *sem,
                 * to the task to wakeup.
                 */
                smp_store_release(&waiter->task, NULL);
-       } while (--loop);
+       }
 
-       sem->wait_list.next = next;
-       next->prev = &sem->wait_list;
+       adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
+       if (list_empty(&sem->wait_list)) {
+               /* hit end of list above */
+               adjustment -= RWSEM_WAITING_BIAS;
+       }
+
+       if (adjustment)
+               atomic_long_add(adjustment, &sem->count);
 }
 
 /*
@@ -245,7 +238,8 @@ struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
        /* we're now waiting on the lock, but no longer actively locking */
        count = atomic_long_add_return(adjustment, &sem->count);
 
-       /* If there are no active locks, wake the front queued process(es).
+       /*
+        * 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 !