md: introduce get_priority_stripe() to improve raid456 write performance
Dan Williams [Mon, 28 Apr 2008 09:15:53 +0000 (02:15 -0700)]
Improve write performance by preventing the delayed_list from dumping all its
stripes onto the handle_list in one shot.  Delayed stripes are now further
delayed by being held on the 'hold_list'.  The 'hold_list' is bypassed when:

  * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list'
  * 'handle_list' is empty and i/o is being done to satisfy full stripe-width
    write requests
  * 'bypass_count' is less than 'bypass_threshold'.  By default the threshold
    is 1, i.e. every other stripe handled is a preread stripe provided the
    top two conditions are false.

Benchmark data:
System: 2x Xeon 5150, 4x SATA, mem=1GB
Baseline: 2.6.24-rc7
Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean
Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048
  * patched:  +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512)

Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS)
  * patched: +13%
  * patched + preread_bypass_threshold = 0: +37%

Changes since v1:
* reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and
  make it configurable.  This defaults to fairness and modest performance
  gains out of the box.
Changes since v2:
* [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not
  necessary, the important change was clearing STRIPE_DELAYED in
  add_stripe_bio and this has been moved out to make_request for the hang
  fix.
* [neilb@suse.de]: simplify get_priority_stripe
* [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is
  sampled empty (+11%)
* [dan.j.williams@intel.com]: decrement the bypass_count at the detection
  of stripes being naturally promoted off of hold_list +2%.  Note, resetting
  bypass_count instead of decrementing on these events yields +4% but that is
  probably too aggressive.
Changes since v3:
* cosmetic fixups

Tested-by: James W. Laferriere <babydr@baby-dragons.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Documentation/md.txt
drivers/md/raid5.c
include/linux/raid/raid5.h

index 396cdd9..a8b4306 100644 (file)
@@ -450,3 +450,9 @@ These currently include
       there are upper and lower limits (32768, 16).  Default is 128.
   strip_cache_active (currently raid5 only)
       number of active entries in the stripe cache
+  preread_bypass_threshold (currently raid5 only)
+      number of times a stripe requiring preread will be bypassed by
+      a stripe that does not require preread.  For fairness defaults
+      to 1.  Setting this to 0 disables bypass accounting and
+      requires preread stripes to wait until all full-width stripe-
+      writes are complete.  Valid values are 0 to stripe_cache_size.
index 4efec46..45eead6 100644 (file)
@@ -63,6 +63,7 @@
 #define STRIPE_SHIFT           (PAGE_SHIFT - 9)
 #define STRIPE_SECTORS         (STRIPE_SIZE>>9)
 #define        IO_THRESHOLD            1
+#define BYPASS_THRESHOLD       1
 #define NR_HASH                        (PAGE_SIZE / sizeof(struct hlist_head))
 #define HASH_MASK              (NR_HASH - 1)
 
@@ -398,6 +399,7 @@ static void ops_run_io(struct stripe_head *sh)
 
        might_sleep();
 
+       set_bit(STRIPE_IO_STARTED, &sh->state);
        for (i = disks; i--; ) {
                int rw;
                struct bio *bi;
@@ -1720,6 +1722,9 @@ handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
                                locked++;
                        }
                }
+               if (locked + 1 == disks)
+                       if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
+                               atomic_inc(&sh->raid_conf->pending_full_writes);
        } else {
                BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
                        test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
@@ -1947,6 +1952,9 @@ handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
                                        STRIPE_SECTORS, 0, 0);
        }
 
+       if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
+               if (atomic_dec_and_test(&conf->pending_full_writes))
+                       md_wakeup_thread(conf->mddev->thread);
 }
 
 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
@@ -2149,6 +2157,10 @@ static void handle_completed_write_requests(raid5_conf_t *conf,
                                                        0);
                        }
                }
+
+       if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
+               if (atomic_dec_and_test(&conf->pending_full_writes))
+                       md_wakeup_thread(conf->mddev->thread);
 }
 
 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
@@ -2333,6 +2345,9 @@ static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
                                s->locked++;
                                set_bit(R5_Wantwrite, &sh->dev[i].flags);
                        }
+               if (s->locked == disks)
+                       if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
+                               atomic_inc(&conf->pending_full_writes);
                /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
                set_bit(STRIPE_INSYNC, &sh->state);
 
@@ -3094,6 +3109,8 @@ static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
                else
                        continue;
 
+               set_bit(STRIPE_IO_STARTED, &sh->state);
+
                bi = &sh->dev[i].req;
 
                bi->bi_rw = rw;
@@ -3164,7 +3181,7 @@ static void raid5_activate_delayed(raid5_conf_t *conf)
                        clear_bit(STRIPE_DELAYED, &sh->state);
                        if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
                                atomic_inc(&conf->preread_active_stripes);
-                       list_add_tail(&sh->lru, &conf->handle_list);
+                       list_add_tail(&sh->lru, &conf->hold_list);
                }
        } else
                blk_plug_device(conf->mddev->queue);
@@ -3442,6 +3459,58 @@ static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
        }
 }
 
+/* __get_priority_stripe - get the next stripe to process
+ *
+ * Full stripe writes are allowed to pass preread active stripes up until
+ * the bypass_threshold is exceeded.  In general the bypass_count
+ * increments when the handle_list is handled before the hold_list; however, it
+ * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
+ * stripe with in flight i/o.  The bypass_count will be reset when the
+ * head of the hold_list has changed, i.e. the head was promoted to the
+ * handle_list.
+ */
+static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
+{
+       struct stripe_head *sh;
+
+       pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
+                 __func__,
+                 list_empty(&conf->handle_list) ? "empty" : "busy",
+                 list_empty(&conf->hold_list) ? "empty" : "busy",
+                 atomic_read(&conf->pending_full_writes), conf->bypass_count);
+
+       if (!list_empty(&conf->handle_list)) {
+               sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
+
+               if (list_empty(&conf->hold_list))
+                       conf->bypass_count = 0;
+               else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
+                       if (conf->hold_list.next == conf->last_hold)
+                               conf->bypass_count++;
+                       else {
+                               conf->last_hold = conf->hold_list.next;
+                               conf->bypass_count -= conf->bypass_threshold;
+                               if (conf->bypass_count < 0)
+                                       conf->bypass_count = 0;
+                       }
+               }
+       } else if (!list_empty(&conf->hold_list) &&
+                  ((conf->bypass_threshold &&
+                    conf->bypass_count > conf->bypass_threshold) ||
+                   atomic_read(&conf->pending_full_writes) == 0)) {
+               sh = list_entry(conf->hold_list.next,
+                               typeof(*sh), lru);
+               conf->bypass_count -= conf->bypass_threshold;
+               if (conf->bypass_count < 0)
+                       conf->bypass_count = 0;
+       } else
+               return NULL;
+
+       list_del_init(&sh->lru);
+       atomic_inc(&sh->count);
+       BUG_ON(atomic_read(&sh->count) != 1);
+       return sh;
+}
 
 static int make_request(struct request_queue *q, struct bio * bi)
 {
@@ -3914,7 +3983,6 @@ static void raid5d(mddev_t *mddev)
        handled = 0;
        spin_lock_irq(&conf->device_lock);
        while (1) {
-               struct list_head *first;
                struct bio *bio;
 
                if (conf->seq_flush != conf->seq_write) {
@@ -3936,17 +4004,12 @@ static void raid5d(mddev_t *mddev)
                        handled++;
                }
 
-               if (list_empty(&conf->handle_list)) {
+               sh = __get_priority_stripe(conf);
+
+               if (!sh) {
                        async_tx_issue_pending_all();
                        break;
                }
-
-               first = conf->handle_list.next;
-               sh = list_entry(first, struct stripe_head, lru);
-
-               list_del_init(first);
-               atomic_inc(&sh->count);
-               BUG_ON(atomic_read(&sh->count)!= 1);
                spin_unlock_irq(&conf->device_lock);
                
                handled++;
@@ -4011,6 +4074,42 @@ raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
                                raid5_store_stripe_cache_size);
 
 static ssize_t
+raid5_show_preread_threshold(mddev_t *mddev, char *page)
+{
+       raid5_conf_t *conf = mddev_to_conf(mddev);
+       if (conf)
+               return sprintf(page, "%d\n", conf->bypass_threshold);
+       else
+               return 0;
+}
+
+static ssize_t
+raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
+{
+       raid5_conf_t *conf = mddev_to_conf(mddev);
+       char *end;
+       int new;
+       if (len >= PAGE_SIZE)
+               return -EINVAL;
+       if (!conf)
+               return -ENODEV;
+
+       new = simple_strtoul(page, &end, 10);
+       if (!*page || (*end && *end != '\n'))
+               return -EINVAL;
+       if (new > conf->max_nr_stripes || new < 0)
+               return -EINVAL;
+       conf->bypass_threshold = new;
+       return len;
+}
+
+static struct md_sysfs_entry
+raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
+                                       S_IRUGO | S_IWUSR,
+                                       raid5_show_preread_threshold,
+                                       raid5_store_preread_threshold);
+
+static ssize_t
 stripe_cache_active_show(mddev_t *mddev, char *page)
 {
        raid5_conf_t *conf = mddev_to_conf(mddev);
@@ -4026,6 +4125,7 @@ raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
 static struct attribute *raid5_attrs[] =  {
        &raid5_stripecache_size.attr,
        &raid5_stripecache_active.attr,
+       &raid5_preread_bypass_threshold.attr,
        NULL,
 };
 static struct attribute_group raid5_attrs_group = {
@@ -4130,12 +4230,14 @@ static int run(mddev_t *mddev)
        init_waitqueue_head(&conf->wait_for_stripe);
        init_waitqueue_head(&conf->wait_for_overlap);
        INIT_LIST_HEAD(&conf->handle_list);
+       INIT_LIST_HEAD(&conf->hold_list);
        INIT_LIST_HEAD(&conf->delayed_list);
        INIT_LIST_HEAD(&conf->bitmap_list);
        INIT_LIST_HEAD(&conf->inactive_list);
        atomic_set(&conf->active_stripes, 0);
        atomic_set(&conf->preread_active_stripes, 0);
        atomic_set(&conf->active_aligned_reads, 0);
+       conf->bypass_threshold = BYPASS_THRESHOLD;
 
        pr_debug("raid5: run(%s) called.\n", mdname(mddev));
 
index 93678f5..f0827d3 100644 (file)
@@ -252,6 +252,8 @@ struct r6_state {
 #define        STRIPE_EXPANDING        9
 #define        STRIPE_EXPAND_SOURCE    10
 #define        STRIPE_EXPAND_READY     11
+#define        STRIPE_IO_STARTED       12 /* do not count towards 'bypass_count' */
+#define        STRIPE_FULL_WRITE       13 /* all blocks are set to be overwritten */
 /*
  * Operations flags (in issue order)
  */
@@ -316,12 +318,17 @@ struct raid5_private_data {
        int                     previous_raid_disks;
 
        struct list_head        handle_list; /* stripes needing handling */
+       struct list_head        hold_list; /* preread ready stripes */
        struct list_head        delayed_list; /* stripes that have plugged requests */
        struct list_head        bitmap_list; /* stripes delaying awaiting bitmap update */
        struct bio              *retry_read_aligned; /* currently retrying aligned bios   */
        struct bio              *retry_read_aligned_list; /* aligned bios retry list  */
        atomic_t                preread_active_stripes; /* stripes with scheduled io */
        atomic_t                active_aligned_reads;
+       atomic_t                pending_full_writes; /* full write backlog */
+       int                     bypass_count; /* bypassed prereads */
+       int                     bypass_threshold; /* preread nice */
+       struct list_head        *last_hold; /* detect hold_list promotions */
 
        atomic_t                reshape_stripes; /* stripes with pending writes for reshape */
        /* unfortunately we need two cache names as we temporarily have