raid5: relieve lock contention in get_active_stripe()
authorShaohua Li <shli@kernel.org>
Thu, 14 Nov 2013 04:16:17 +0000 (15:16 +1100)
committerNeilBrown <neilb@suse.de>
Thu, 14 Nov 2013 04:20:58 +0000 (15:20 +1100)
get_active_stripe() is the last place we have lock contention. It has two
paths. One is stripe isn't found and new stripe is allocated, the other is
stripe is found.

The first path basically calls __find_stripe and init_stripe. It accesses
conf->generation, conf->previous_raid_disks, conf->raid_disks,
conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded,
conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except
stripe_hashtbl and inactive_list, other fields are changed very rarely.

With this patch, we split inactive_list and add new hash locks. Each free
stripe belongs to a specific inactive list. Which inactive list is determined
by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a
lock_hash assigned. Stripe's inactive list is protected by a hash lock, which
is determined by it's lock_hash too. The lock_hash is derivied from current
stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned
to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl
list too. The goal of the new hash locks introduced is we can only use the new
locks in the first path of get_active_stripe(). Since we have several hash
locks, lock contention is relieved significantly.

The first path of get_active_stripe() accesses other fields, since they are
changed rarely, changing them now need take conf->device_lock and all hash
locks. For a slow path, this isn't a problem.

If we need lock device_lock and hash lock, we always lock hash lock first. The
tricky part is release_stripe and friends. We need take device_lock first.
Neil's suggestion is we put inactive stripes to a temporary list and readd it
to inactive_list after device_lock is released. In this way, we add stripes to
temporary list with device_lock hold and remove stripes from the list with hash
lock hold. So we don't allow concurrent access to the temporary list, which
means we need allocate temporary list for all participants of release_stripe.

One downside is free stripes are maintained in their inactive list, they can't
across between the lists. By default, we have total 256 stripes and 8 lists, so
each list will have 32 stripes. It's possible one list has free stripe but
other list hasn't. The chance should be rare because stripes allocation are
even distributed. And we can always allocate more stripes for cache, several
mega bytes memory isn't a big deal.

This completely removes the lock contention of the first path of
get_active_stripe(). It slows down the second code path a little bit though
because we now need takes two locks, but since the hash lock isn't contended,
the overhead should be quite small (several atomic instructions). The second
path of get_active_stripe() (basically sequential write or big request size
randwrite) still has lock contentions.

Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid5.c
drivers/md/raid5.h

index 4bbcb7e26d128e5a6507d90b998bb3b086cc40c6..93090b2afab482afcf83b32322188ab8b141c340 100644 (file)
@@ -85,6 +85,42 @@ static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
        return &conf->stripe_hashtbl[hash];
 }
 
+static inline int stripe_hash_locks_hash(sector_t sect)
+{
+       return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
+}
+
+static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
+{
+       spin_lock_irq(conf->hash_locks + hash);
+       spin_lock(&conf->device_lock);
+}
+
+static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
+{
+       spin_unlock(&conf->device_lock);
+       spin_unlock_irq(conf->hash_locks + hash);
+}
+
+static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
+{
+       int i;
+       local_irq_disable();
+       spin_lock(conf->hash_locks);
+       for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
+               spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
+       spin_lock(&conf->device_lock);
+}
+
+static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
+{
+       int i;
+       spin_unlock(&conf->device_lock);
+       for (i = NR_STRIPE_HASH_LOCKS; i; i--)
+               spin_unlock(conf->hash_locks + i - 1);
+       local_irq_enable();
+}
+
 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  * order without overlap.  There may be several bio's per stripe+device, and
  * a bio could span several devices.
@@ -249,7 +285,8 @@ static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
        }
 }
 
-static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
+static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
+                             struct list_head *temp_inactive_list)
 {
        BUG_ON(!list_empty(&sh->lru));
        BUG_ON(atomic_read(&conf->active_stripes)==0);
@@ -278,19 +315,60 @@ static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
                            < IO_THRESHOLD)
                                md_wakeup_thread(conf->mddev->thread);
                atomic_dec(&conf->active_stripes);
-               if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
-                       list_add_tail(&sh->lru, &conf->inactive_list);
-                       wake_up(&conf->wait_for_stripe);
-                       if (conf->retry_read_aligned)
-                               md_wakeup_thread(conf->mddev->thread);
-               }
+               if (!test_bit(STRIPE_EXPANDING, &sh->state))
+                       list_add_tail(&sh->lru, temp_inactive_list);
        }
 }
 
-static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
+static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
+                            struct list_head *temp_inactive_list)
 {
        if (atomic_dec_and_test(&sh->count))
-               do_release_stripe(conf, sh);
+               do_release_stripe(conf, sh, temp_inactive_list);
+}
+
+/*
+ * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
+ *
+ * Be careful: Only one task can add/delete stripes from temp_inactive_list at
+ * given time. Adding stripes only takes device lock, while deleting stripes
+ * only takes hash lock.
+ */
+static void release_inactive_stripe_list(struct r5conf *conf,
+                                        struct list_head *temp_inactive_list,
+                                        int hash)
+{
+       int size;
+       bool do_wakeup = false;
+       unsigned long flags;
+
+       if (hash == NR_STRIPE_HASH_LOCKS) {
+               size = NR_STRIPE_HASH_LOCKS;
+               hash = NR_STRIPE_HASH_LOCKS - 1;
+       } else
+               size = 1;
+       while (size) {
+               struct list_head *list = &temp_inactive_list[size - 1];
+
+               /*
+                * We don't hold any lock here yet, get_active_stripe() might
+                * remove stripes from the list
+                */
+               if (!list_empty_careful(list)) {
+                       spin_lock_irqsave(conf->hash_locks + hash, flags);
+                       list_splice_tail_init(list, conf->inactive_list + hash);
+                       do_wakeup = true;
+                       spin_unlock_irqrestore(conf->hash_locks + hash, flags);
+               }
+               size--;
+               hash--;
+       }
+
+       if (do_wakeup) {
+               wake_up(&conf->wait_for_stripe);
+               if (conf->retry_read_aligned)
+                       md_wakeup_thread(conf->mddev->thread);
+       }
 }
 
 static struct llist_node *llist_reverse_order(struct llist_node *head)
@@ -308,7 +386,8 @@ static struct llist_node *llist_reverse_order(struct llist_node *head)
 }
 
 /* should hold conf->device_lock already */
-static int release_stripe_list(struct r5conf *conf)
+static int release_stripe_list(struct r5conf *conf,
+                              struct list_head *temp_inactive_list)
 {
        struct stripe_head *sh;
        int count = 0;
@@ -317,6 +396,8 @@ static int release_stripe_list(struct r5conf *conf)
        head = llist_del_all(&conf->released_stripes);
        head = llist_reverse_order(head);
        while (head) {
+               int hash;
+
                sh = llist_entry(head, struct stripe_head, release_list);
                head = llist_next(head);
                /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
@@ -327,7 +408,8 @@ static int release_stripe_list(struct r5conf *conf)
                 * again, the count is always > 1. This is true for
                 * STRIPE_ON_UNPLUG_LIST bit too.
                 */
-               __release_stripe(conf, sh);
+               hash = sh->hash_lock_index;
+               __release_stripe(conf, sh, &temp_inactive_list[hash]);
                count++;
        }
 
@@ -338,6 +420,8 @@ static void release_stripe(struct stripe_head *sh)
 {
        struct r5conf *conf = sh->raid_conf;
        unsigned long flags;
+       struct list_head list;
+       int hash;
        bool wakeup;
 
        if (unlikely(!conf->mddev->thread) ||
@@ -351,8 +435,11 @@ slow_path:
        local_irq_save(flags);
        /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
        if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
-               do_release_stripe(conf, sh);
+               INIT_LIST_HEAD(&list);
+               hash = sh->hash_lock_index;
+               do_release_stripe(conf, sh, &list);
                spin_unlock(&conf->device_lock);
+               release_inactive_stripe_list(conf, &list, hash);
        }
        local_irq_restore(flags);
 }
@@ -377,18 +464,19 @@ static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 
 
 /* find an idle stripe, make sure it is unhashed, and return it. */
-static struct stripe_head *get_free_stripe(struct r5conf *conf)
+static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 {
        struct stripe_head *sh = NULL;
        struct list_head *first;
 
-       if (list_empty(&conf->inactive_list))
+       if (list_empty(conf->inactive_list + hash))
                goto out;
-       first = conf->inactive_list.next;
+       first = (conf->inactive_list + hash)->next;
        sh = list_entry(first, struct stripe_head, lru);
        list_del_init(first);
        remove_hash(sh);
        atomic_inc(&conf->active_stripes);
+       BUG_ON(hash != sh->hash_lock_index);
 out:
        return sh;
 }
@@ -431,7 +519,7 @@ static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 {
        struct r5conf *conf = sh->raid_conf;
-       int i;
+       int i, seq;
 
        BUG_ON(atomic_read(&sh->count) != 0);
        BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
@@ -441,7 +529,8 @@ static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
                (unsigned long long)sh->sector);
 
        remove_hash(sh);
-
+retry:
+       seq = read_seqcount_begin(&conf->gen_lock);
        sh->generation = conf->generation - previous;
        sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
        sh->sector = sector;
@@ -463,6 +552,8 @@ static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
                dev->flags = 0;
                raid5_build_block(sh, i, previous);
        }
+       if (read_seqcount_retry(&conf->gen_lock, seq))
+               goto retry;
        insert_hash(conf, sh);
        sh->cpu = smp_processor_id();
 }
@@ -567,29 +658,31 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
                  int previous, int noblock, int noquiesce)
 {
        struct stripe_head *sh;
+       int hash = stripe_hash_locks_hash(sector);
 
        pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 
-       spin_lock_irq(&conf->device_lock);
+       spin_lock_irq(conf->hash_locks + hash);
 
        do {
                wait_event_lock_irq(conf->wait_for_stripe,
                                    conf->quiesce == 0 || noquiesce,
-                                   conf->device_lock);
+                                   *(conf->hash_locks + hash));
                sh = __find_stripe(conf, sector, conf->generation - previous);
                if (!sh) {
                        if (!conf->inactive_blocked)
-                               sh = get_free_stripe(conf);
+                               sh = get_free_stripe(conf, hash);
                        if (noblock && sh == NULL)
                                break;
                        if (!sh) {
                                conf->inactive_blocked = 1;
-                               wait_event_lock_irq(conf->wait_for_stripe,
-                                                   !list_empty(&conf->inactive_list) &&
-                                                   (atomic_read(&conf->active_stripes)
-                                                    < (conf->max_nr_stripes *3/4)
-                                                    || !conf->inactive_blocked),
-                                                   conf->device_lock);
+                               wait_event_lock_irq(
+                                       conf->wait_for_stripe,
+                                       !list_empty(conf->inactive_list + hash) &&
+                                       (atomic_read(&conf->active_stripes)
+                                        < (conf->max_nr_stripes * 3 / 4)
+                                        || !conf->inactive_blocked),
+                                       *(conf->hash_locks + hash));
                                conf->inactive_blocked = 0;
                        } else
                                init_stripe(sh, sector, previous);
@@ -600,9 +693,11 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
                                    && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)
                                    && !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
                        } else {
+                               spin_lock(&conf->device_lock);
                                if (!test_bit(STRIPE_HANDLE, &sh->state))
                                        atomic_inc(&conf->active_stripes);
                                if (list_empty(&sh->lru) &&
+                                   !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state) &&
                                    !test_bit(STRIPE_EXPANDING, &sh->state))
                                        BUG();
                                list_del_init(&sh->lru);
@@ -610,6 +705,7 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
                                        sh->group->stripes_cnt--;
                                        sh->group = NULL;
                                }
+                               spin_unlock(&conf->device_lock);
                        }
                }
        } while (sh == NULL);
@@ -617,7 +713,7 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
        if (sh)
                atomic_inc(&sh->count);
 
-       spin_unlock_irq(&conf->device_lock);
+       spin_unlock_irq(conf->hash_locks + hash);
        return sh;
 }
 
@@ -1597,7 +1693,7 @@ static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
        put_cpu();
 }
 
-static int grow_one_stripe(struct r5conf *conf)
+static int grow_one_stripe(struct r5conf *conf, int hash)
 {
        struct stripe_head *sh;
        sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
@@ -1613,6 +1709,7 @@ static int grow_one_stripe(struct r5conf *conf)
                kmem_cache_free(conf->slab_cache, sh);
                return 0;
        }
+       sh->hash_lock_index = hash;
        /* we just created an active stripe so... */
        atomic_set(&sh->count, 1);
        atomic_inc(&conf->active_stripes);
@@ -1625,6 +1722,7 @@ static int grow_stripes(struct r5conf *conf, int num)
 {
        struct kmem_cache *sc;
        int devs = max(conf->raid_disks, conf->previous_raid_disks);
+       int hash;
 
        if (conf->mddev->gendisk)
                sprintf(conf->cache_name[0],
@@ -1642,9 +1740,13 @@ static int grow_stripes(struct r5conf *conf, int num)
                return 1;
        conf->slab_cache = sc;
        conf->pool_size = devs;
-       while (num--)
-               if (!grow_one_stripe(conf))
+       hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
+       while (num--) {
+               if (!grow_one_stripe(conf, hash))
                        return 1;
+               conf->max_nr_stripes++;
+               hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
+       }
        return 0;
 }
 
@@ -1702,6 +1804,7 @@ static int resize_stripes(struct r5conf *conf, int newsize)
        int err;
        struct kmem_cache *sc;
        int i;
+       int hash, cnt;
 
        if (newsize <= conf->pool_size)
                return 0; /* never bother to shrink */
@@ -1741,19 +1844,29 @@ static int resize_stripes(struct r5conf *conf, int newsize)
         * OK, we have enough stripes, start collecting inactive
         * stripes and copying them over
         */
+       hash = 0;
+       cnt = 0;
        list_for_each_entry(nsh, &newstripes, lru) {
-               spin_lock_irq(&conf->device_lock);
-               wait_event_lock_irq(conf->wait_for_stripe,
-                                   !list_empty(&conf->inactive_list),
-                                   conf->device_lock);
-               osh = get_free_stripe(conf);
-               spin_unlock_irq(&conf->device_lock);
+               lock_device_hash_lock(conf, hash);
+               wait_event_cmd(conf->wait_for_stripe,
+                                   !list_empty(conf->inactive_list + hash),
+                                   unlock_device_hash_lock(conf, hash),
+                                   lock_device_hash_lock(conf, hash));
+               osh = get_free_stripe(conf, hash);
+               unlock_device_hash_lock(conf, hash);
                atomic_set(&nsh->count, 1);
                for(i=0; i<conf->pool_size; i++)
                        nsh->dev[i].page = osh->dev[i].page;
                for( ; i<newsize; i++)
                        nsh->dev[i].page = NULL;
+               nsh->hash_lock_index = hash;
                kmem_cache_free(conf->slab_cache, osh);
+               cnt++;
+               if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
+                   !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
+                       hash++;
+                       cnt = 0;
+               }
        }
        kmem_cache_destroy(conf->slab_cache);
 
@@ -1812,13 +1925,13 @@ static int resize_stripes(struct r5conf *conf, int newsize)
        return err;
 }
 
-static int drop_one_stripe(struct r5conf *conf)
+static int drop_one_stripe(struct r5conf *conf, int hash)
 {
        struct stripe_head *sh;
 
-       spin_lock_irq(&conf->device_lock);
-       sh = get_free_stripe(conf);
-       spin_unlock_irq(&conf->device_lock);
+       spin_lock_irq(conf->hash_locks + hash);
+       sh = get_free_stripe(conf, hash);
+       spin_unlock_irq(conf->hash_locks + hash);
        if (!sh)
                return 0;
        BUG_ON(atomic_read(&sh->count));
@@ -1830,8 +1943,10 @@ static int drop_one_stripe(struct r5conf *conf)
 
 static void shrink_stripes(struct r5conf *conf)
 {
-       while (drop_one_stripe(conf))
-               ;
+       int hash;
+       for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
+               while (drop_one_stripe(conf, hash))
+                       ;
 
        if (conf->slab_cache)
                kmem_cache_destroy(conf->slab_cache);
@@ -3915,7 +4030,8 @@ static void raid5_activate_delayed(struct r5conf *conf)
        }
 }
 
-static void activate_bit_delay(struct r5conf *conf)
+static void activate_bit_delay(struct r5conf *conf,
+       struct list_head *temp_inactive_list)
 {
        /* device_lock is held */
        struct list_head head;
@@ -3923,9 +4039,11 @@ static void activate_bit_delay(struct r5conf *conf)
        list_del_init(&conf->bitmap_list);
        while (!list_empty(&head)) {
                struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
+               int hash;
                list_del_init(&sh->lru);
                atomic_inc(&sh->count);
-               __release_stripe(conf, sh);
+               hash = sh->hash_lock_index;
+               __release_stripe(conf, sh, &temp_inactive_list[hash]);
        }
 }
 
@@ -3941,7 +4059,7 @@ int md_raid5_congested(struct mddev *mddev, int bits)
                return 1;
        if (conf->quiesce)
                return 1;
-       if (list_empty_careful(&conf->inactive_list))
+       if (atomic_read(&conf->active_stripes) == conf->max_nr_stripes)
                return 1;
 
        return 0;
@@ -4271,6 +4389,7 @@ static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
 struct raid5_plug_cb {
        struct blk_plug_cb      cb;
        struct list_head        list;
+       struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
 };
 
 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
@@ -4281,6 +4400,7 @@ static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
        struct mddev *mddev = cb->cb.data;
        struct r5conf *conf = mddev->private;
        int cnt = 0;
+       int hash;
 
        if (cb->list.next && !list_empty(&cb->list)) {
                spin_lock_irq(&conf->device_lock);
@@ -4298,11 +4418,14 @@ static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
                         * STRIPE_ON_RELEASE_LIST could be set here. In that
                         * case, the count is always > 1 here
                         */
-                       __release_stripe(conf, sh);
+                       hash = sh->hash_lock_index;
+                       __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
                        cnt++;
                }
                spin_unlock_irq(&conf->device_lock);
        }
+       release_inactive_stripe_list(conf, cb->temp_inactive_list,
+                                    NR_STRIPE_HASH_LOCKS);
        if (mddev->queue)
                trace_block_unplug(mddev->queue, cnt, !from_schedule);
        kfree(cb);
@@ -4323,8 +4446,12 @@ static void release_stripe_plug(struct mddev *mddev,
 
        cb = container_of(blk_cb, struct raid5_plug_cb, cb);
 
-       if (cb->list.next == NULL)
+       if (cb->list.next == NULL) {
+               int i;
                INIT_LIST_HEAD(&cb->list);
+               for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
+                       INIT_LIST_HEAD(cb->temp_inactive_list + i);
+       }
 
        if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
                list_add_tail(&sh->lru, &cb->list);
@@ -4969,27 +5096,45 @@ static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
 }
 
 static int handle_active_stripes(struct r5conf *conf, int group,
-                                struct r5worker *worker)
+                                struct r5worker *worker,
+                                struct list_head *temp_inactive_list)
 {
        struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
-       int i, batch_size = 0;
+       int i, batch_size = 0, hash;
+       bool release_inactive = false;
 
        while (batch_size < MAX_STRIPE_BATCH &&
                        (sh = __get_priority_stripe(conf, group)) != NULL)
                batch[batch_size++] = sh;
 
-       if (batch_size == 0)
-               return batch_size;
+       if (batch_size == 0) {
+               for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
+                       if (!list_empty(temp_inactive_list + i))
+                               break;
+               if (i == NR_STRIPE_HASH_LOCKS)
+                       return batch_size;
+               release_inactive = true;
+       }
        spin_unlock_irq(&conf->device_lock);
 
+       release_inactive_stripe_list(conf, temp_inactive_list,
+                                    NR_STRIPE_HASH_LOCKS);
+
+       if (release_inactive) {
+               spin_lock_irq(&conf->device_lock);
+               return 0;
+       }
+
        for (i = 0; i < batch_size; i++)
                handle_stripe(batch[i]);
 
        cond_resched();
 
        spin_lock_irq(&conf->device_lock);
-       for (i = 0; i < batch_size; i++)
-               __release_stripe(conf, batch[i]);
+       for (i = 0; i < batch_size; i++) {
+               hash = batch[i]->hash_lock_index;
+               __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
+       }
        return batch_size;
 }
 
@@ -5010,9 +5155,10 @@ static void raid5_do_work(struct work_struct *work)
        while (1) {
                int batch_size, released;
 
-               released = release_stripe_list(conf);
+               released = release_stripe_list(conf, worker->temp_inactive_list);
 
-               batch_size = handle_active_stripes(conf, group_id, worker);
+               batch_size = handle_active_stripes(conf, group_id, worker,
+                                                  worker->temp_inactive_list);
                worker->working = false;
                if (!batch_size && !released)
                        break;
@@ -5051,7 +5197,7 @@ static void raid5d(struct md_thread *thread)
                struct bio *bio;
                int batch_size, released;
 
-               released = release_stripe_list(conf);
+               released = release_stripe_list(conf, conf->temp_inactive_list);
 
                if (
                    !list_empty(&conf->bitmap_list)) {
@@ -5061,7 +5207,7 @@ static void raid5d(struct md_thread *thread)
                        bitmap_unplug(mddev->bitmap);
                        spin_lock_irq(&conf->device_lock);
                        conf->seq_write = conf->seq_flush;
-                       activate_bit_delay(conf);
+                       activate_bit_delay(conf, conf->temp_inactive_list);
                }
                raid5_activate_delayed(conf);
 
@@ -5075,7 +5221,8 @@ static void raid5d(struct md_thread *thread)
                        handled++;
                }
 
-               batch_size = handle_active_stripes(conf, ANY_GROUP, NULL);
+               batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
+                                                  conf->temp_inactive_list);
                if (!batch_size && !released)
                        break;
                handled += batch_size;
@@ -5111,22 +5258,29 @@ raid5_set_cache_size(struct mddev *mddev, int size)
 {
        struct r5conf *conf = mddev->private;
        int err;
+       int hash;
 
        if (size <= 16 || size > 32768)
                return -EINVAL;
+       hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
        while (size < conf->max_nr_stripes) {
-               if (drop_one_stripe(conf))
+               if (drop_one_stripe(conf, hash))
                        conf->max_nr_stripes--;
                else
                        break;
+               hash--;
+               if (hash < 0)
+                       hash = NR_STRIPE_HASH_LOCKS - 1;
        }
        err = md_allow_write(mddev);
        if (err)
                return err;
+       hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
        while (size > conf->max_nr_stripes) {
-               if (grow_one_stripe(conf))
+               if (grow_one_stripe(conf, hash))
                        conf->max_nr_stripes++;
                else break;
+               hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
        }
        return 0;
 }
@@ -5277,7 +5431,7 @@ static struct attribute_group raid5_attrs_group = {
 
 static int alloc_thread_groups(struct r5conf *conf, int cnt)
 {
-       int i, j;
+       int i, j, k;
        ssize_t size;
        struct r5worker *workers;
 
@@ -5307,8 +5461,12 @@ static int alloc_thread_groups(struct r5conf *conf, int cnt)
                group->workers = workers + i * cnt;
 
                for (j = 0; j < cnt; j++) {
-                       group->workers[j].group = group;
-                       INIT_WORK(&group->workers[j].work, raid5_do_work);
+                       struct r5worker *worker = group->workers + j;
+                       worker->group = group;
+                       INIT_WORK(&worker->work, raid5_do_work);
+
+                       for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
+                               INIT_LIST_HEAD(worker->temp_inactive_list + k);
                }
        }
 
@@ -5459,6 +5617,7 @@ static struct r5conf *setup_conf(struct mddev *mddev)
        struct md_rdev *rdev;
        struct disk_info *disk;
        char pers_name[6];
+       int i;
 
        if (mddev->new_level != 5
            && mddev->new_level != 4
@@ -5503,7 +5662,6 @@ static struct r5conf *setup_conf(struct mddev *mddev)
        INIT_LIST_HEAD(&conf->hold_list);
        INIT_LIST_HEAD(&conf->delayed_list);
        INIT_LIST_HEAD(&conf->bitmap_list);
-       INIT_LIST_HEAD(&conf->inactive_list);
        init_llist_head(&conf->released_stripes);
        atomic_set(&conf->active_stripes, 0);
        atomic_set(&conf->preread_active_stripes, 0);
@@ -5529,6 +5687,21 @@ static struct r5conf *setup_conf(struct mddev *mddev)
        if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
                goto abort;
 
+       /* We init hash_locks[0] separately to that it can be used
+        * as the reference lock in the spin_lock_nest_lock() call
+        * in lock_all_device_hash_locks_irq in order to convince
+        * lockdep that we know what we are doing.
+        */
+       spin_lock_init(conf->hash_locks);
+       for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
+               spin_lock_init(conf->hash_locks + i);
+
+       for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
+               INIT_LIST_HEAD(conf->inactive_list + i);
+
+       for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
+               INIT_LIST_HEAD(conf->temp_inactive_list + i);
+
        conf->level = mddev->new_level;
        if (raid5_alloc_percpu(conf) != 0)
                goto abort;
@@ -5569,7 +5742,6 @@ static struct r5conf *setup_conf(struct mddev *mddev)
        else
                conf->max_degraded = 1;
        conf->algorithm = mddev->new_layout;
-       conf->max_nr_stripes = NR_STRIPES;
        conf->reshape_progress = mddev->reshape_position;
        if (conf->reshape_progress != MaxSector) {
                conf->prev_chunk_sectors = mddev->chunk_sectors;
@@ -5578,7 +5750,7 @@ static struct r5conf *setup_conf(struct mddev *mddev)
 
        memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
                 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
-       if (grow_stripes(conf, conf->max_nr_stripes)) {
+       if (grow_stripes(conf, NR_STRIPES)) {
                printk(KERN_ERR
                       "md/raid:%s: couldn't allocate %dkB for buffers\n",
                       mdname(mddev), memory);
@@ -6483,27 +6655,28 @@ static void raid5_quiesce(struct mddev *mddev, int state)
                break;
 
        case 1: /* stop all writes */
-               spin_lock_irq(&conf->device_lock);
+               lock_all_device_hash_locks_irq(conf);
                /* '2' tells resync/reshape to pause so that all
                 * active stripes can drain
                 */
                conf->quiesce = 2;
-               wait_event_lock_irq(conf->wait_for_stripe,
+               wait_event_cmd(conf->wait_for_stripe,
                                    atomic_read(&conf->active_stripes) == 0 &&
                                    atomic_read(&conf->active_aligned_reads) == 0,
-                                   conf->device_lock);
+                                   unlock_all_device_hash_locks_irq(conf),
+                                   lock_all_device_hash_locks_irq(conf));
                conf->quiesce = 1;
-               spin_unlock_irq(&conf->device_lock);
+               unlock_all_device_hash_locks_irq(conf);
                /* allow reshape to continue */
                wake_up(&conf->wait_for_overlap);
                break;
 
        case 0: /* re-enable writes */
-               spin_lock_irq(&conf->device_lock);
+               lock_all_device_hash_locks_irq(conf);
                conf->quiesce = 0;
                wake_up(&conf->wait_for_stripe);
                wake_up(&conf->wait_for_overlap);
-               spin_unlock_irq(&conf->device_lock);
+               unlock_all_device_hash_locks_irq(conf);
                break;
        }
 }
index 2113ffa82c7a2506d21ce36076c3f6bd057d922b..a9e443a1116f193182c4f7e903f07c7609522fbf 100644 (file)
@@ -205,6 +205,7 @@ struct stripe_head {
        short                   pd_idx;         /* parity disk index */
        short                   qd_idx;         /* 'Q' disk index for raid6 */
        short                   ddf_layout;/* use DDF ordering to calculate Q */
+       short                   hash_lock_index;
        unsigned long           state;          /* state flags */
        atomic_t                count;        /* nr of active thread/requests */
        int                     bm_seq; /* sequence number for bitmap flushes */
@@ -367,9 +368,18 @@ struct disk_info {
        struct md_rdev  *rdev, *replacement;
 };
 
+/* NOTE NR_STRIPE_HASH_LOCKS must remain below 64.
+ * This is because we sometimes take all the spinlocks
+ * and creating that much locking depth can cause
+ * problems.
+ */
+#define NR_STRIPE_HASH_LOCKS 8
+#define STRIPE_HASH_LOCKS_MASK (NR_STRIPE_HASH_LOCKS - 1)
+
 struct r5worker {
        struct work_struct work;
        struct r5worker_group *group;
+       struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
        bool working;
 };
 
@@ -382,6 +392,8 @@ struct r5worker_group {
 
 struct r5conf {
        struct hlist_head       *stripe_hashtbl;
+       /* only protect corresponding hash list and inactive_list */
+       spinlock_t              hash_locks[NR_STRIPE_HASH_LOCKS];
        struct mddev            *mddev;
        int                     chunk_sectors;
        int                     level, algorithm;
@@ -462,7 +474,7 @@ struct r5conf {
         * Free stripes pool
         */
        atomic_t                active_stripes;
-       struct list_head        inactive_list;
+       struct list_head        inactive_list[NR_STRIPE_HASH_LOCKS];
        struct llist_head       released_stripes;
        wait_queue_head_t       wait_for_stripe;
        wait_queue_head_t       wait_for_overlap;
@@ -477,6 +489,7 @@ struct r5conf {
         * the new thread here until we fully activate the array.
         */
        struct md_thread        *thread;
+       struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
        struct r5worker_group   *worker_groups;
        int                     group_cnt;
        int                     worker_cnt_per_group;