]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
RAID1: avoid unnecessary spin locks in I/O barrier code
authorcolyli@suse.de <colyli@suse.de>
Fri, 17 Feb 2017 19:05:57 +0000 (03:05 +0800)
committerShaohua Li <shli@fb.com>
Mon, 20 Feb 2017 06:04:25 +0000 (22:04 -0800)
When I run a parallel reading performan testing on a md raid1 device with
two NVMe SSDs, I observe very bad throughput in supprise: by fio with 64KB
block size, 40 seq read I/O jobs, 128 iodepth, overall throughput is
only 2.7GB/s, this is around 50% of the idea performance number.

The perf reports locking contention happens at allow_barrier() and
wait_barrier() code,
 - 41.41%  fio [kernel.kallsyms]     [k] _raw_spin_lock_irqsave
   - _raw_spin_lock_irqsave
         + 89.92% allow_barrier
         + 9.34% __wake_up
 - 37.30%  fio [kernel.kallsyms]     [k] _raw_spin_lock_irq
   - _raw_spin_lock_irq
         - 100.00% wait_barrier

The reason is, in these I/O barrier related functions,
 - raise_barrier()
 - lower_barrier()
 - wait_barrier()
 - allow_barrier()
They always hold conf->resync_lock firstly, even there are only regular
reading I/Os and no resync I/O at all. This is a huge performance penalty.

The solution is a lockless-like algorithm in I/O barrier code, and only
holding conf->resync_lock when it has to.

The original idea is from Hannes Reinecke, and Neil Brown provides
comments to improve it. I continue to work on it, and make the patch into
current form.

In the new simpler raid1 I/O barrier implementation, there are two
wait barrier functions,
 - wait_barrier()
   Which calls _wait_barrier(), is used for regular write I/O. If there is
   resync I/O happening on the same I/O barrier bucket, or the whole
   array is frozen, task will wait until no barrier on same barrier bucket,
   or the whold array is unfreezed.
 - wait_read_barrier()
   Since regular read I/O won't interfere with resync I/O (read_balance()
   will make sure only uptodate data will be read out), it is unnecessary
   to wait for barrier in regular read I/Os, waiting in only necessary
   when the whole array is frozen.

The operations on conf->nr_pending[idx], conf->nr_waiting[idx], conf->
barrier[idx] are very carefully designed in raise_barrier(),
lower_barrier(), _wait_barrier() and wait_read_barrier(), in order to
avoid unnecessary spin locks in these functions. Once conf->
nr_pengding[idx] is increased, a resync I/O with same barrier bucket index
has to wait in raise_barrier(). Then in _wait_barrier() if no barrier
raised in same barrier bucket index and array is not frozen, the regular
I/O doesn't need to hold conf->resync_lock, it can just increase
conf->nr_pending[idx], and return to its caller. wait_read_barrier() is
very similar to _wait_barrier(), the only difference is it only waits when
array is frozen. For heavy parallel reading I/Os, the lockless I/O barrier
code almostly gets rid of all spin lock cost.

This patch significantly improves raid1 reading peroformance. From my
testing, a raid1 device built by two NVMe SSD, runs fio with 64KB
blocksize, 40 seq read I/O jobs, 128 iodepth, overall throughput
increases from 2.7GB/s to 4.6GB/s (+70%).

Changelog
V4:
- Change conf->nr_queued[] to atomic_t.
- Define BARRIER_BUCKETS_NR_BITS by (PAGE_SHIFT - ilog2(sizeof(atomic_t)))
V3:
- Add smp_mb__after_atomic() as Shaohua and Neil suggested.
- Change conf->nr_queued[] from atomic_t to int.
- Change conf->array_frozen from atomic_t back to int, and use
  READ_ONCE(conf->array_frozen) to check value of conf->array_frozen
  in _wait_barrier() and wait_read_barrier().
- In _wait_barrier() and wait_read_barrier(), add a call to
  wake_up(&conf->wait_barrier) after atomic_dec(&conf->nr_pending[idx]),
  to fix a deadlock between  _wait_barrier()/wait_read_barrier and
  freeze_array().
V2:
- Remove a spin_lock/unlock pair in raid1d().
- Add more code comments to explain why there is no racy when checking two
  atomic_t variables at same time.
V1:
- Original RFC patch for comments.

Signed-off-by: Coly Li <colyli@suse.de>
Cc: Shaohua Li <shli@fb.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Cc: Guoqing Jiang <gqjiang@suse.com>
Reviewed-by: Neil Brown <neilb@suse.de>
Signed-off-by: Shaohua Li <shli@fb.com>
drivers/md/raid1.c
drivers/md/raid1.h

index 40297fd17f7ea5b1efa707e8c8d44d96259490ab..fefbbfdb440b4199ecd2a79fb90517df0693156c 100644 (file)
@@ -226,7 +226,7 @@ static void reschedule_retry(struct r1bio *r1_bio)
        idx = sector_to_idx(r1_bio->sector);
        spin_lock_irqsave(&conf->device_lock, flags);
        list_add(&r1_bio->retry_list, &conf->retry_list);
-       conf->nr_queued[idx]++;
+       atomic_inc(&conf->nr_queued[idx]);
        spin_unlock_irqrestore(&conf->device_lock, flags);
 
        wake_up(&conf->wait_barrier);
@@ -836,11 +836,21 @@ static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
        spin_lock_irq(&conf->resync_lock);
 
        /* Wait until no block IO is waiting */
-       wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting[idx],
+       wait_event_lock_irq(conf->wait_barrier,
+                           !atomic_read(&conf->nr_waiting[idx]),
                            conf->resync_lock);
 
        /* block any new IO from starting */
-       conf->barrier[idx]++;
+       atomic_inc(&conf->barrier[idx]);
+       /*
+        * In raise_barrier() we firstly increase conf->barrier[idx] then
+        * check conf->nr_pending[idx]. In _wait_barrier() we firstly
+        * increase conf->nr_pending[idx] then check conf->barrier[idx].
+        * A memory barrier here to make sure conf->nr_pending[idx] won't
+        * be fetched before conf->barrier[idx] is increased. Otherwise
+        * there will be a race between raise_barrier() and _wait_barrier().
+        */
+       smp_mb__after_atomic();
 
        /* For these conditions we must wait:
         * A: while the array is in frozen state
@@ -851,42 +861,81 @@ static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
         */
        wait_event_lock_irq(conf->wait_barrier,
                            !conf->array_frozen &&
-                            !conf->nr_pending[idx] &&
-                            conf->barrier[idx] < RESYNC_DEPTH,
+                            !atomic_read(&conf->nr_pending[idx]) &&
+                            atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
                            conf->resync_lock);
 
-       conf->nr_pending[idx]++;
+       atomic_inc(&conf->nr_pending[idx]);
        spin_unlock_irq(&conf->resync_lock);
 }
 
 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 {
-       unsigned long flags;
        int idx = sector_to_idx(sector_nr);
 
-       BUG_ON(conf->barrier[idx] <= 0);
+       BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 
-       spin_lock_irqsave(&conf->resync_lock, flags);
-       conf->barrier[idx]--;
-       conf->nr_pending[idx]--;
-       spin_unlock_irqrestore(&conf->resync_lock, flags);
+       atomic_dec(&conf->barrier[idx]);
+       atomic_dec(&conf->nr_pending[idx]);
        wake_up(&conf->wait_barrier);
 }
 
 static void _wait_barrier(struct r1conf *conf, int idx)
 {
-       spin_lock_irq(&conf->resync_lock);
-       if (conf->array_frozen || conf->barrier[idx]) {
-               conf->nr_waiting[idx]++;
-               /* Wait for the barrier to drop. */
-               wait_event_lock_irq(
-                       conf->wait_barrier,
-                       !conf->array_frozen && !conf->barrier[idx],
-                       conf->resync_lock);
-               conf->nr_waiting[idx]--;
-       }
+       /*
+        * We need to increase conf->nr_pending[idx] very early here,
+        * then raise_barrier() can be blocked when it waits for
+        * conf->nr_pending[idx] to be 0. Then we can avoid holding
+        * conf->resync_lock when there is no barrier raised in same
+        * barrier unit bucket. Also if the array is frozen, I/O
+        * should be blocked until array is unfrozen.
+        */
+       atomic_inc(&conf->nr_pending[idx]);
+       /*
+        * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
+        * check conf->barrier[idx]. In raise_barrier() we firstly increase
+        * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
+        * barrier is necessary here to make sure conf->barrier[idx] won't be
+        * fetched before conf->nr_pending[idx] is increased. Otherwise there
+        * will be a race between _wait_barrier() and raise_barrier().
+        */
+       smp_mb__after_atomic();
+
+       /*
+        * Don't worry about checking two atomic_t variables at same time
+        * here. If during we check conf->barrier[idx], the array is
+        * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
+        * 0, it is safe to return and make the I/O continue. Because the
+        * array is frozen, all I/O returned here will eventually complete
+        * or be queued, no race will happen. See code comment in
+        * frozen_array().
+        */
+       if (!READ_ONCE(conf->array_frozen) &&
+           !atomic_read(&conf->barrier[idx]))
+               return;
 
-       conf->nr_pending[idx]++;
+       /*
+        * After holding conf->resync_lock, conf->nr_pending[idx]
+        * should be decreased before waiting for barrier to drop.
+        * Otherwise, we may encounter a race condition because
+        * raise_barrer() might be waiting for conf->nr_pending[idx]
+        * to be 0 at same time.
+        */
+       spin_lock_irq(&conf->resync_lock);
+       atomic_inc(&conf->nr_waiting[idx]);
+       atomic_dec(&conf->nr_pending[idx]);
+       /*
+        * In case freeze_array() is waiting for
+        * get_unqueued_pending() == extra
+        */
+       wake_up(&conf->wait_barrier);
+       /* Wait for the barrier in same barrier unit bucket to drop. */
+       wait_event_lock_irq(conf->wait_barrier,
+                           !conf->array_frozen &&
+                            !atomic_read(&conf->barrier[idx]),
+                           conf->resync_lock);
+       atomic_inc(&conf->nr_pending[idx]);
+       atomic_dec(&conf->nr_waiting[idx]);
        spin_unlock_irq(&conf->resync_lock);
 }
 
@@ -894,18 +943,32 @@ static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 {
        int idx = sector_to_idx(sector_nr);
 
-       spin_lock_irq(&conf->resync_lock);
-       if (conf->array_frozen) {
-               conf->nr_waiting[idx]++;
-               /* Wait for array to unfreeze */
-               wait_event_lock_irq(
-                       conf->wait_barrier,
-                       !conf->array_frozen,
-                       conf->resync_lock);
-               conf->nr_waiting[idx]--;
-       }
+       /*
+        * Very similar to _wait_barrier(). The difference is, for read
+        * I/O we don't need wait for sync I/O, but if the whole array
+        * is frozen, the read I/O still has to wait until the array is
+        * unfrozen. Since there is no ordering requirement with
+        * conf->barrier[idx] here, memory barrier is unnecessary as well.
+        */
+       atomic_inc(&conf->nr_pending[idx]);
 
-       conf->nr_pending[idx]++;
+       if (!READ_ONCE(conf->array_frozen))
+               return;
+
+       spin_lock_irq(&conf->resync_lock);
+       atomic_inc(&conf->nr_waiting[idx]);
+       atomic_dec(&conf->nr_pending[idx]);
+       /*
+        * In case freeze_array() is waiting for
+        * get_unqueued_pending() == extra
+        */
+       wake_up(&conf->wait_barrier);
+       /* Wait for array to be unfrozen */
+       wait_event_lock_irq(conf->wait_barrier,
+                           !conf->array_frozen,
+                           conf->resync_lock);
+       atomic_inc(&conf->nr_pending[idx]);
+       atomic_dec(&conf->nr_waiting[idx]);
        spin_unlock_irq(&conf->resync_lock);
 }
 
@@ -926,11 +989,7 @@ static void wait_all_barriers(struct r1conf *conf)
 
 static void _allow_barrier(struct r1conf *conf, int idx)
 {
-       unsigned long flags;
-
-       spin_lock_irqsave(&conf->resync_lock, flags);
-       conf->nr_pending[idx]--;
-       spin_unlock_irqrestore(&conf->resync_lock, flags);
+       atomic_dec(&conf->nr_pending[idx]);
        wake_up(&conf->wait_barrier);
 }
 
@@ -955,7 +1014,8 @@ static int get_unqueued_pending(struct r1conf *conf)
        int idx, ret;
 
        for (ret = 0, idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
-               ret += conf->nr_pending[idx] - conf->nr_queued[idx];
+               ret += atomic_read(&conf->nr_pending[idx]) -
+                       atomic_read(&conf->nr_queued[idx]);
 
        return ret;
 }
@@ -1000,8 +1060,8 @@ static void unfreeze_array(struct r1conf *conf)
        /* reverse the effect of the freeze */
        spin_lock_irq(&conf->resync_lock);
        conf->array_frozen = 0;
-       wake_up(&conf->wait_barrier);
        spin_unlock_irq(&conf->resync_lock);
+       wake_up(&conf->wait_barrier);
 }
 
 /* duplicate the data pages for behind I/O
@@ -2391,8 +2451,13 @@ static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
                spin_lock_irq(&conf->device_lock);
                list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
                idx = sector_to_idx(r1_bio->sector);
-               conf->nr_queued[idx]++;
+               atomic_inc(&conf->nr_queued[idx]);
                spin_unlock_irq(&conf->device_lock);
+               /*
+                * In case freeze_array() is waiting for condition
+                * get_unqueued_pending() == extra to be true.
+                */
+               wake_up(&conf->wait_barrier);
                md_wakeup_thread(conf->mddev->thread);
        } else {
                if (test_bit(R1BIO_WriteError, &r1_bio->state))
@@ -2523,9 +2588,7 @@ static void raid1d(struct md_thread *thread)
                                                  retry_list);
                        list_del(&r1_bio->retry_list);
                        idx = sector_to_idx(r1_bio->sector);
-                       spin_lock_irqsave(&conf->device_lock, flags);
-                       conf->nr_queued[idx]--;
-                       spin_unlock_irqrestore(&conf->device_lock, flags);
+                       atomic_dec(&conf->nr_queued[idx]);
                        if (mddev->degraded)
                                set_bit(R1BIO_Degraded, &r1_bio->state);
                        if (test_bit(R1BIO_WriteError, &r1_bio->state))
@@ -2547,7 +2610,7 @@ static void raid1d(struct md_thread *thread)
                r1_bio = list_entry(head->prev, struct r1bio, retry_list);
                list_del(head->prev);
                idx = sector_to_idx(r1_bio->sector);
-               conf->nr_queued[idx]--;
+               atomic_dec(&conf->nr_queued[idx]);
                spin_unlock_irqrestore(&conf->device_lock, flags);
 
                mddev = r1_bio->mddev;
@@ -2664,7 +2727,7 @@ static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
         * If there is non-resync activity waiting for a turn, then let it
         * though before starting on this new sync request.
         */
-       if (conf->nr_waiting[idx])
+       if (atomic_read(&conf->nr_waiting[idx]))
                schedule_timeout_uninterruptible(1);
 
        /* we are incrementing sector_nr below. To be safe, we check against
@@ -2924,22 +2987,22 @@ static struct r1conf *setup_conf(struct mddev *mddev)
                goto abort;
 
        conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
-                                  sizeof(int), GFP_KERNEL);
+                                  sizeof(atomic_t), GFP_KERNEL);
        if (!conf->nr_pending)
                goto abort;
 
        conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
-                                  sizeof(int), GFP_KERNEL);
+                                  sizeof(atomic_t), GFP_KERNEL);
        if (!conf->nr_waiting)
                goto abort;
 
        conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
-                                 sizeof(int), GFP_KERNEL);
+                                 sizeof(atomic_t), GFP_KERNEL);
        if (!conf->nr_queued)
                goto abort;
 
        conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
-                               sizeof(int), GFP_KERNEL);
+                               sizeof(atomic_t), GFP_KERNEL);
        if (!conf->barrier)
                goto abort;
 
index 3442e8fe3fcd0a113c4e0ec367482754ec2d61fe..dd22a37d0d8332e12785b9c270445aba09cce576 100644 (file)
 /*
  * In struct r1conf, the following members are related to I/O barrier
  * buckets,
- *     int     *nr_pending;
- *     int     *nr_waiting;
- *     int     *nr_queued;
- *     int     *barrier;
- * Each of them points to array of integers, each array is designed to
- * have BARRIER_BUCKETS_NR elements and occupy a single memory page. The
- * data width of integer variables is 4, equal to 1<<(ilog2(sizeof(int))),
- * BARRIER_BUCKETS_NR_BITS is defined as (PAGE_SHIFT - ilog2(sizeof(int)))
- * to make sure an array of integers with BARRIER_BUCKETS_NR elements just
- * exactly occupies a single memory page.
+ *     atomic_t        *nr_pending;
+ *     atomic_t        *nr_waiting;
+ *     atomic_t        *nr_queued;
+ *     atomic_t        *barrier;
+ * Each of them points to array of atomic_t variables, each array is
+ * designed to have BARRIER_BUCKETS_NR elements and occupy a single
+ * memory page. The data width of atomic_t variables is 4 bytes, equal
+ * to 1<<(ilog2(sizeof(atomic_t))), BARRIER_BUCKETS_NR_BITS is defined
+ * as (PAGE_SHIFT - ilog2(sizeof(int))) to make sure an array of
+ * atomic_t variables with BARRIER_BUCKETS_NR elements just exactly
+ * occupies a single memory page.
  */
-#define BARRIER_BUCKETS_NR_BITS                (PAGE_SHIFT - ilog2(sizeof(int)))
+#define BARRIER_BUCKETS_NR_BITS                (PAGE_SHIFT - ilog2(sizeof(atomic_t)))
 #define BARRIER_BUCKETS_NR             (1<<BARRIER_BUCKETS_NR_BITS)
 
 struct raid1_info {
@@ -83,10 +84,10 @@ struct r1conf {
         */
        wait_queue_head_t       wait_barrier;
        spinlock_t              resync_lock;
-       int                     *nr_pending;
-       int                     *nr_waiting;
-       int                     *nr_queued;
-       int                     *barrier;
+       atomic_t                *nr_pending;
+       atomic_t                *nr_waiting;
+       atomic_t                *nr_queued;
+       atomic_t                *barrier;
        int                     array_frozen;
 
        /* Set to 1 if a full sync is needed, (fresh device added).