md/raid5: per hash value and exclusive wait_for_stripe
authorYuanhan Liu <yuanhan.liu@linux.intel.com>
Fri, 8 May 2015 08:19:07 +0000 (18:19 +1000)
committerNeilBrown <neilb@suse.de>
Wed, 17 Jun 2015 00:00:27 +0000 (10:00 +1000)
I noticed heavy spin lock contention at get_active_stripe() with fsmark
multiple thread write workloads.

Here is how this hot contention comes from. We have limited stripes, and
it's a multiple thread write workload. Hence, those stripes will be taken
soon, which puts later processes to sleep for waiting free stripes. When
enough stripes(>= 1/4 total stripes) are released, all process are woken,
trying to get the lock. But there is one only being able to get this lock
for each hash lock, making other processes spinning out there for acquiring
the lock.

Thus, it's effectiveless to wakeup all processes and let them battle for
a lock that permits one to access only each time. Instead, we could make
it be a exclusive wake up: wake up one process only. That avoids the heavy
spin lock contention naturally.

To do the exclusive wake up, we've to split wait_for_stripe into multiple
wait queues, to make it per hash value, just like the hash lock.

Here are some test results I have got with this patch applied(all test run
3 times):

`fsmark.files_per_sec'
=====================

next-20150317                 this patch
-------------------------     -------------------------
metric_value     ±stddev      metric_value     ±stddev     change      testbox/benchmark/testcase-params
-------------------------     -------------------------   --------     ------------------------------
      25.600     ±0.0              92.700     ±2.5          262.1%     ivb44/fsmark/1x-64t-4BRD_12G-RAID5-btrfs-4M-30G-fsyncBeforeClose
      25.600     ±0.0              77.800     ±0.6          203.9%     ivb44/fsmark/1x-64t-9BRD_6G-RAID5-btrfs-4M-30G-fsyncBeforeClose
      32.000     ±0.0              93.800     ±1.7          193.1%     ivb44/fsmark/1x-64t-4BRD_12G-RAID5-ext4-4M-30G-fsyncBeforeClose
      32.000     ±0.0              81.233     ±1.7          153.9%     ivb44/fsmark/1x-64t-9BRD_6G-RAID5-ext4-4M-30G-fsyncBeforeClose
      48.800     ±14.5             99.667     ±2.0          104.2%     ivb44/fsmark/1x-64t-4BRD_12G-RAID5-xfs-4M-30G-fsyncBeforeClose
       6.400     ±0.0              12.800     ±0.0          100.0%     ivb44/fsmark/1x-64t-3HDD-RAID5-btrfs-4M-40G-fsyncBeforeClose
      63.133     ±8.2              82.800     ±0.7           31.2%     ivb44/fsmark/1x-64t-9BRD_6G-RAID5-xfs-4M-30G-fsyncBeforeClose
     245.067     ±0.7             306.567     ±7.9           25.1%     ivb44/fsmark/1x-64t-4BRD_12G-RAID5-f2fs-4M-30G-fsyncBeforeClose
      17.533     ±0.3              21.000     ±0.8           19.8%     ivb44/fsmark/1x-1t-3HDD-RAID5-xfs-4M-40G-fsyncBeforeClose
     188.167     ±1.9             215.033     ±3.1           14.3%     ivb44/fsmark/1x-1t-4BRD_12G-RAID5-btrfs-4M-30G-NoSync
     254.500     ±1.8             290.733     ±2.4           14.2%     ivb44/fsmark/1x-1t-9BRD_6G-RAID5-btrfs-4M-30G-NoSync

`time.system_time'
=====================

next-20150317                 this patch
-------------------------    -------------------------
metric_value     ±stddev     metric_value     ±stddev     change       testbox/benchmark/testcase-params
-------------------------    -------------------------    --------     ------------------------------
    7235.603     ±1.2             185.163     ±1.9          -97.4%     ivb44/fsmark/1x-64t-4BRD_12G-RAID5-btrfs-4M-30G-fsyncBeforeClose
    7666.883     ±2.9             202.750     ±1.0          -97.4%     ivb44/fsmark/1x-64t-9BRD_6G-RAID5-btrfs-4M-30G-fsyncBeforeClose
   14567.893     ±0.7             421.230     ±0.4          -97.1%     ivb44/fsmark/1x-64t-3HDD-RAID5-btrfs-4M-40G-fsyncBeforeClose
    3697.667     ±14.0            148.190     ±1.7          -96.0%     ivb44/fsmark/1x-64t-4BRD_12G-RAID5-xfs-4M-30G-fsyncBeforeClose
    5572.867     ±3.8             310.717     ±1.4          -94.4%     ivb44/fsmark/1x-64t-9BRD_6G-RAID5-ext4-4M-30G-fsyncBeforeClose
    5565.050     ±0.5             313.277     ±1.5          -94.4%     ivb44/fsmark/1x-64t-4BRD_12G-RAID5-ext4-4M-30G-fsyncBeforeClose
    2420.707     ±17.1            171.043     ±2.7          -92.9%     ivb44/fsmark/1x-64t-9BRD_6G-RAID5-xfs-4M-30G-fsyncBeforeClose
    3743.300     ±4.6             379.827     ±3.5          -89.9%     ivb44/fsmark/1x-64t-3HDD-RAID5-ext4-4M-40G-fsyncBeforeClose
    3308.687     ±6.3             363.050     ±2.0          -89.0%     ivb44/fsmark/1x-64t-3HDD-RAID5-xfs-4M-40G-fsyncBeforeClose

Where,

     1x: where 'x' means iterations or loop, corresponding to the 'L' option of fsmark

     1t, 64t: where 't' means thread

     4M: means the single file size, corresponding to the '-s' option of fsmark
     40G, 30G, 120G: means the total test size

     4BRD_12G: BRD is the ramdisk, where '4' means 4 ramdisk, and where '12G' means
               the size of one ramdisk. So, it would be 48G in total. And we made a
               raid on those ramdisk

As you can see, though there are no much performance gain for hard disk
workload, the system time is dropped heavily, up to 97%. And as expected,
the performance increased a lot, up to 260%, for fast device(ram disk).

v2: use bits instead of array to note down wait queue need to wake up.

Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid5.c
drivers/md/raid5.h

index a9112b39afee834eea5afae388eb167ec07527f4..9a3b143b0b68a7caf956f9033944f9c46dccdc7b 100644 (file)
@@ -344,7 +344,8 @@ static void release_inactive_stripe_list(struct r5conf *conf,
                                         int hash)
 {
        int size;
-       bool do_wakeup = false;
+       unsigned long do_wakeup = 0;
+       int i = 0;
        unsigned long flags;
 
        if (hash == NR_STRIPE_HASH_LOCKS) {
@@ -365,15 +366,19 @@ static void release_inactive_stripe_list(struct r5conf *conf,
                            !list_empty(list))
                                atomic_dec(&conf->empty_inactive_list_nr);
                        list_splice_tail_init(list, conf->inactive_list + hash);
-                       do_wakeup = true;
+                       do_wakeup |= 1 << hash;
                        spin_unlock_irqrestore(conf->hash_locks + hash, flags);
                }
                size--;
                hash--;
        }
 
+       for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
+               if (do_wakeup & (1 << i))
+                       wake_up(&conf->wait_for_stripe[i]);
+       }
+
        if (do_wakeup) {
-               wake_up(&conf->wait_for_stripe);
                if (atomic_read(&conf->active_stripes) == 0)
                        wake_up(&conf->wait_for_quiescent);
                if (conf->retry_read_aligned)
@@ -686,14 +691,15 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
                        if (!sh) {
                                set_bit(R5_INACTIVE_BLOCKED,
                                        &conf->cache_state);
-                               wait_event_lock_irq(
-                                       conf->wait_for_stripe,
+                               wait_event_exclusive_cmd(
+                                       conf->wait_for_stripe[hash],
                                        !list_empty(conf->inactive_list + hash) &&
                                        (atomic_read(&conf->active_stripes)
                                         < (conf->max_nr_stripes * 3 / 4)
                                         || !test_bit(R5_INACTIVE_BLOCKED,
                                                      &conf->cache_state)),
-                                       *(conf->hash_locks + hash));
+                                       spin_unlock_irq(conf->hash_locks + hash),
+                                       spin_lock_irq(conf->hash_locks + hash));
                                clear_bit(R5_INACTIVE_BLOCKED,
                                          &conf->cache_state);
                        } else {
@@ -718,6 +724,9 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
                }
        } while (sh == NULL);
 
+       if (!list_empty(conf->inactive_list + hash))
+               wake_up(&conf->wait_for_stripe[hash]);
+
        spin_unlock_irq(conf->hash_locks + hash);
        return sh;
 }
@@ -2179,7 +2188,7 @@ static int resize_stripes(struct r5conf *conf, int newsize)
        cnt = 0;
        list_for_each_entry(nsh, &newstripes, lru) {
                lock_device_hash_lock(conf, hash);
-               wait_event_cmd(conf->wait_for_stripe,
+               wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
                                    !list_empty(conf->inactive_list + hash),
                                    unlock_device_hash_lock(conf, hash),
                                    lock_device_hash_lock(conf, hash));
@@ -6436,7 +6445,9 @@ static struct r5conf *setup_conf(struct mddev *mddev)
        spin_lock_init(&conf->device_lock);
        seqcount_init(&conf->gen_lock);
        init_waitqueue_head(&conf->wait_for_quiescent);
-       init_waitqueue_head(&conf->wait_for_stripe);
+       for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
+               init_waitqueue_head(&conf->wait_for_stripe[i]);
+       }
        init_waitqueue_head(&conf->wait_for_overlap);
        INIT_LIST_HEAD(&conf->handle_list);
        INIT_LIST_HEAD(&conf->hold_list);
index 9b84b8820fc566c2c2029f55dba242014e8779b5..02c3bf8fbfe7aa1b0a1c568605393dc3769c6b18 100644 (file)
@@ -512,7 +512,7 @@ struct r5conf {
        atomic_t                empty_inactive_list_nr;
        struct llist_head       released_stripes;
        wait_queue_head_t       wait_for_quiescent;
-       wait_queue_head_t       wait_for_stripe;
+       wait_queue_head_t       wait_for_stripe[NR_STRIPE_HASH_LOCKS];
        wait_queue_head_t       wait_for_overlap;
        unsigned long           cache_state;
 #define R5_INACTIVE_BLOCKED    1       /* release of inactive stripes blocked,