md/raid6: asynchronous handle_stripe6
authorYuri Tikhonov <yur@emcraft.com>
Sun, 30 Aug 2009 02:13:13 +0000 (19:13 -0700)
committerDan Williams <dan.j.williams@intel.com>
Sun, 30 Aug 2009 02:13:13 +0000 (19:13 -0700)
1/ Use STRIPE_OP_BIOFILL to offload completion of read requests to
   raid_run_ops
2/ Implement a handler for sh->reconstruct_state similar to the raid5 case
   (adds handling of Q parity)
3/ Prevent handle_parity_checks6 from running concurrently with 'compute'
   operations
4/ Hook up raid_run_ops

Signed-off-by: Yuri Tikhonov <yur@emcraft.com>
Signed-off-by: Ilya Yanok <yanok@emcraft.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
drivers/md/raid5.c

index 3c31f7f8aa6545cee53aa894b056d0ebb4d2bfbd..a833de189ca67ebee186c651ee0056c7f7a43ab9 100644 (file)
@@ -3424,9 +3424,10 @@ static bool handle_stripe6(struct stripe_head *sh)
        mdk_rdev_t *blocked_rdev = NULL;
 
        pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
-               "pd_idx=%d, qd_idx=%d\n",
+               "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
               (unsigned long long)sh->sector, sh->state,
-              atomic_read(&sh->count), pd_idx, qd_idx);
+              atomic_read(&sh->count), pd_idx, qd_idx,
+              sh->check_state, sh->reconstruct_state);
        memset(&s, 0, sizeof(s));
 
        spin_lock(&sh->lock);
@@ -3446,35 +3447,24 @@ static bool handle_stripe6(struct stripe_head *sh)
 
                pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
                        i, dev->flags, dev->toread, dev->towrite, dev->written);
-               /* maybe we can reply to a read */
-               if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
-                       struct bio *rbi, *rbi2;
-                       pr_debug("Return read for disc %d\n", i);
-                       spin_lock_irq(&conf->device_lock);
-                       rbi = dev->toread;
-                       dev->toread = NULL;
-                       if (test_and_clear_bit(R5_Overlap, &dev->flags))
-                               wake_up(&conf->wait_for_overlap);
-                       spin_unlock_irq(&conf->device_lock);
-                       while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
-                               copy_data(0, rbi, dev->page, dev->sector);
-                               rbi2 = r5_next_bio(rbi, dev->sector);
-                               spin_lock_irq(&conf->device_lock);
-                               if (!raid5_dec_bi_phys_segments(rbi)) {
-                                       rbi->bi_next = return_bi;
-                                       return_bi = rbi;
-                               }
-                               spin_unlock_irq(&conf->device_lock);
-                               rbi = rbi2;
-                       }
-               }
+               /* maybe we can reply to a read
+                *
+                * new wantfill requests are only permitted while
+                * ops_complete_biofill is guaranteed to be inactive
+                */
+               if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
+                   !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
+                       set_bit(R5_Wantfill, &dev->flags);
 
                /* now count some things */
                if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
                if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
+               if (test_bit(R5_Wantcompute, &dev->flags))
+                       BUG_ON(++s.compute > 2);
 
-
-               if (dev->toread)
+               if (test_bit(R5_Wantfill, &dev->flags)) {
+                       s.to_fill++;
+               } else if (dev->toread)
                        s.to_read++;
                if (dev->towrite) {
                        s.to_write++;
@@ -3515,6 +3505,11 @@ static bool handle_stripe6(struct stripe_head *sh)
                blocked_rdev = NULL;
        }
 
+       if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
+               set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
+               set_bit(STRIPE_BIOFILL_RUN, &sh->state);
+       }
+
        pr_debug("locked=%d uptodate=%d to_read=%d"
               " to_write=%d failed=%d failed_num=%d,%d\n",
               s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
@@ -3555,9 +3550,43 @@ static bool handle_stripe6(struct stripe_head *sh)
         * or to load a block that is being partially written.
         */
        if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
-           (s.syncing && (s.uptodate < disks)) || s.expanding)
+           (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
                handle_stripe_fill6(sh, &s, &r6s, disks);
 
+       /* Now we check to see if any write operations have recently
+        * completed
+        */
+       if (sh->reconstruct_state == reconstruct_state_drain_result) {
+               int qd_idx = sh->qd_idx;
+
+               sh->reconstruct_state = reconstruct_state_idle;
+               /* All the 'written' buffers and the parity blocks are ready to
+                * be written back to disk
+                */
+               BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
+               BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
+               for (i = disks; i--; ) {
+                       dev = &sh->dev[i];
+                       if (test_bit(R5_LOCKED, &dev->flags) &&
+                           (i == sh->pd_idx || i == qd_idx ||
+                            dev->written)) {
+                               pr_debug("Writing block %d\n", i);
+                               BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
+                               set_bit(R5_Wantwrite, &dev->flags);
+                               if (!test_bit(R5_Insync, &dev->flags) ||
+                                   ((i == sh->pd_idx || i == qd_idx) &&
+                                     s.failed == 0))
+                                       set_bit(STRIPE_INSYNC, &sh->state);
+                       }
+               }
+               if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
+                       atomic_dec(&conf->preread_active_stripes);
+                       if (atomic_read(&conf->preread_active_stripes) <
+                               IO_THRESHOLD)
+                               md_wakeup_thread(conf->mddev->thread);
+               }
+       }
+
        /* Now to consider new write requests and what else, if anything
         * should be read.  We do not handle new writes when:
         * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
@@ -3569,9 +3598,13 @@ static bool handle_stripe6(struct stripe_head *sh)
 
        /* maybe we need to check and possibly fix the parity for this stripe
         * Any reads will already have been scheduled, so we just see if enough
-        * data is available
+        * data is available.  The parity check is held off while parity
+        * dependent operations are in flight.
         */
-       if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
+       if (sh->check_state ||
+           (s.syncing && s.locked == 0 &&
+            !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
+            !test_bit(STRIPE_INSYNC, &sh->state)))
                handle_parity_checks6(conf, sh, &s, &r6s, disks);
 
        if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
@@ -3593,15 +3626,29 @@ static bool handle_stripe6(struct stripe_head *sh)
                                        set_bit(R5_Wantwrite, &dev->flags);
                                        set_bit(R5_ReWrite, &dev->flags);
                                        set_bit(R5_LOCKED, &dev->flags);
+                                       s.locked++;
                                } else {
                                        /* let's read it back */
                                        set_bit(R5_Wantread, &dev->flags);
                                        set_bit(R5_LOCKED, &dev->flags);
+                                       s.locked++;
                                }
                        }
                }
 
-       if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
+       /* Finish reconstruct operations initiated by the expansion process */
+       if (sh->reconstruct_state == reconstruct_state_result) {
+               sh->reconstruct_state = reconstruct_state_idle;
+               clear_bit(STRIPE_EXPANDING, &sh->state);
+               for (i = conf->raid_disks; i--; ) {
+                       set_bit(R5_Wantwrite, &sh->dev[i].flags);
+                       set_bit(R5_LOCKED, &sh->dev[i].flags);
+                       s.locked++;
+               }
+       }
+
+       if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
+           !sh->reconstruct_state) {
                struct stripe_head *sh2
                        = get_active_stripe(conf, sh->sector, 1, 1);
                if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
@@ -3622,14 +3669,8 @@ static bool handle_stripe6(struct stripe_head *sh)
                /* Need to write out all blocks after computing P&Q */
                sh->disks = conf->raid_disks;
                stripe_set_idx(sh->sector, conf, 0, sh);
-               compute_parity6(sh, RECONSTRUCT_WRITE);
-               for (i = conf->raid_disks ; i-- ;  ) {
-                       set_bit(R5_LOCKED, &sh->dev[i].flags);
-                       s.locked++;
-                       set_bit(R5_Wantwrite, &sh->dev[i].flags);
-               }
-               clear_bit(STRIPE_EXPANDING, &sh->state);
-       } else if (s.expanded) {
+               schedule_reconstruction(sh, &s, 1, 1);
+       } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
                clear_bit(STRIPE_EXPAND_READY, &sh->state);
                atomic_dec(&conf->reshape_stripes);
                wake_up(&conf->wait_for_overlap);
@@ -3647,6 +3688,9 @@ static bool handle_stripe6(struct stripe_head *sh)
        if (unlikely(blocked_rdev))
                md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
 
+       if (s.ops_request)
+               raid_run_ops(sh, s.ops_request);
+
        ops_run_io(sh, &s);
 
        return_io(return_bi);