md/raid10: handle merge_bvec_fn in member devices.
authorNeilBrown <neilb@suse.de>
Mon, 19 Mar 2012 01:46:39 +0000 (12:46 +1100)
committerNeilBrown <neilb@suse.de>
Mon, 19 Mar 2012 01:46:39 +0000 (12:46 +1100)
Currently we don't honour merge_bvec_fn in member devices so if there
is one, we force all requests to be single-page at most.
This is not ideal.

So enhance the raid10 merge_bvec_fn to check that function in children
as well.

This introduces a small problem.  There is no locking around calls
the ->merge_bvec_fn and subsequent calls to ->make_request.  So a
device added between these could end up getting a request which
violates its merge_bvec_fn.

Currently the best we can do is synchronize_sched().  This will work
providing no preemption happens.  If there is preemption, we just
have to hope that new devices are largely consistent with old devices.

Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/md.c
drivers/md/md.h
drivers/md/raid10.c

index 119de175bf121bd8a4eeb67cb8ae64eb0378756b..4566b61373d50d9841337e9649869c7b9232b6d2 100644 (file)
@@ -5073,6 +5073,7 @@ static void md_clean(struct mddev *mddev)
        mddev->changed = 0;
        mddev->degraded = 0;
        mddev->safemode = 0;
+       mddev->merge_check_needed = 0;
        mddev->bitmap_info.offset = 0;
        mddev->bitmap_info.default_offset = 0;
        mddev->bitmap_info.chunksize = 0;
index 39acfe90cc26c380e6bbfd638e8054820951f2a8..1c2063ccf48ee6ddfd00a710231fcfecbd48a217 100644 (file)
@@ -128,6 +128,10 @@ struct md_rdev {
 enum flag_bits {
        Faulty,                 /* device is known to have a fault */
        In_sync,                /* device is in_sync with rest of array */
+       Unmerged,               /* device is being added to array and should
+                                * be considerred for bvec_merge_fn but not
+                                * yet for actual IO
+                                */
        WriteMostly,            /* Avoid reading if at all possible */
        AutoDetected,           /* added by auto-detect */
        Blocked,                /* An error occurred but has not yet
@@ -345,6 +349,10 @@ struct mddev {
        int                             degraded;       /* whether md should consider
                                                         * adding a spare
                                                         */
+       int                             merge_check_needed; /* at least one
+                                                            * member device
+                                                            * has a
+                                                            * merge_bvec_fn */
 
        atomic_t                        recovery_active; /* blocks scheduled, but not written */
        wait_queue_head_t               recovery_wait;
index 52bb37d4026dca5a362767bb8238d1bc9e755ad9..e4a66ab6b0fb6c7f7a00e996ef5ced1c3de22f40 100644 (file)
@@ -586,25 +586,68 @@ static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
  *     @biovec: the request that could be merged to it.
  *
  *     Return amount of bytes we can accept at this offset
- *      If near_copies == raid_disk, there are no striping issues,
- *      but in that case, the function isn't called at all.
+ *     This requires checking for end-of-chunk if near_copies != raid_disks,
+ *     and for subordinate merge_bvec_fns if merge_check_needed.
  */
 static int raid10_mergeable_bvec(struct request_queue *q,
                                 struct bvec_merge_data *bvm,
                                 struct bio_vec *biovec)
 {
        struct mddev *mddev = q->queuedata;
+       struct r10conf *conf = mddev->private;
        sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
        int max;
        unsigned int chunk_sectors = mddev->chunk_sectors;
        unsigned int bio_sectors = bvm->bi_size >> 9;
 
-       max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
-       if (max < 0) max = 0; /* bio_add cannot handle a negative return */
-       if (max <= biovec->bv_len && bio_sectors == 0)
-               return biovec->bv_len;
-       else
-               return max;
+       if (conf->near_copies < conf->raid_disks) {
+               max = (chunk_sectors - ((sector & (chunk_sectors - 1))
+                                       + bio_sectors)) << 9;
+               if (max < 0)
+                       /* bio_add cannot handle a negative return */
+                       max = 0;
+               if (max <= biovec->bv_len && bio_sectors == 0)
+                       return biovec->bv_len;
+       } else
+               max = biovec->bv_len;
+
+       if (mddev->merge_check_needed) {
+               struct r10bio r10_bio;
+               int s;
+               r10_bio.sector = sector;
+               raid10_find_phys(conf, &r10_bio);
+               rcu_read_lock();
+               for (s = 0; s < conf->copies; s++) {
+                       int disk = r10_bio.devs[s].devnum;
+                       struct md_rdev *rdev = rcu_dereference(
+                               conf->mirrors[disk].rdev);
+                       if (rdev && !test_bit(Faulty, &rdev->flags)) {
+                               struct request_queue *q =
+                                       bdev_get_queue(rdev->bdev);
+                               if (q->merge_bvec_fn) {
+                                       bvm->bi_sector = r10_bio.devs[s].addr
+                                               + rdev->data_offset;
+                                       bvm->bi_bdev = rdev->bdev;
+                                       max = min(max, q->merge_bvec_fn(
+                                                         q, bvm, biovec));
+                               }
+                       }
+                       rdev = rcu_dereference(conf->mirrors[disk].replacement);
+                       if (rdev && !test_bit(Faulty, &rdev->flags)) {
+                               struct request_queue *q =
+                                       bdev_get_queue(rdev->bdev);
+                               if (q->merge_bvec_fn) {
+                                       bvm->bi_sector = r10_bio.devs[s].addr
+                                               + rdev->data_offset;
+                                       bvm->bi_bdev = rdev->bdev;
+                                       max = min(max, q->merge_bvec_fn(
+                                                         q, bvm, biovec));
+                               }
+                       }
+               }
+               rcu_read_unlock();
+       }
+       return max;
 }
 
 /*
@@ -668,11 +711,12 @@ retry:
                disk = r10_bio->devs[slot].devnum;
                rdev = rcu_dereference(conf->mirrors[disk].replacement);
                if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
+                   test_bit(Unmerged, &rdev->flags) ||
                    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
                        rdev = rcu_dereference(conf->mirrors[disk].rdev);
-               if (rdev == NULL)
-                       continue;
-               if (test_bit(Faulty, &rdev->flags))
+               if (rdev == NULL ||
+                   test_bit(Faulty, &rdev->flags) ||
+                   test_bit(Unmerged, &rdev->flags))
                        continue;
                if (!test_bit(In_sync, &rdev->flags) &&
                    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
@@ -1134,12 +1178,14 @@ retry_write:
                        blocked_rdev = rrdev;
                        break;
                }
-               if (rrdev && test_bit(Faulty, &rrdev->flags))
+               if (rrdev && (test_bit(Faulty, &rrdev->flags)
+                             || test_bit(Unmerged, &rrdev->flags)))
                        rrdev = NULL;
 
                r10_bio->devs[i].bio = NULL;
                r10_bio->devs[i].repl_bio = NULL;
-               if (!rdev || test_bit(Faulty, &rdev->flags)) {
+               if (!rdev || test_bit(Faulty, &rdev->flags) ||
+                   test_bit(Unmerged, &rdev->flags)) {
                        set_bit(R10BIO_Degraded, &r10_bio->state);
                        continue;
                }
@@ -1490,6 +1536,7 @@ static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
        int mirror;
        int first = 0;
        int last = conf->raid_disks - 1;
+       struct request_queue *q = bdev_get_queue(rdev->bdev);
 
        if (mddev->recovery_cp < MaxSector)
                /* only hot-add to in-sync arrays, as recovery is
@@ -1502,6 +1549,11 @@ static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
        if (rdev->raid_disk >= 0)
                first = last = rdev->raid_disk;
 
+       if (q->merge_bvec_fn) {
+               set_bit(Unmerged, &rdev->flags);
+               mddev->merge_check_needed = 1;
+       }
+
        if (rdev->saved_raid_disk >= first &&
            conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
                mirror = rdev->saved_raid_disk;
@@ -1521,11 +1573,6 @@ static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
                        err = 0;
                        disk_stack_limits(mddev->gendisk, rdev->bdev,
                                          rdev->data_offset << 9);
-                       if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
-                               blk_queue_max_segments(mddev->queue, 1);
-                               blk_queue_segment_boundary(mddev->queue,
-                                                          PAGE_CACHE_SIZE - 1);
-                       }
                        conf->fullsync = 1;
                        rcu_assign_pointer(p->replacement, rdev);
                        break;
@@ -1533,17 +1580,6 @@ static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
 
                disk_stack_limits(mddev->gendisk, rdev->bdev,
                                  rdev->data_offset << 9);
-               /* as we don't honour merge_bvec_fn, we must
-                * never risk violating it, so limit
-                * ->max_segments to one lying with a single
-                * page, as a one page request is never in
-                * violation.
-                */
-               if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
-                       blk_queue_max_segments(mddev->queue, 1);
-                       blk_queue_segment_boundary(mddev->queue,
-                                                  PAGE_CACHE_SIZE - 1);
-               }
 
                p->head_position = 0;
                p->recovery_disabled = mddev->recovery_disabled - 1;
@@ -1554,7 +1590,19 @@ static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
                rcu_assign_pointer(p->rdev, rdev);
                break;
        }
-
+       if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
+               /* Some requests might not have seen this new
+                * merge_bvec_fn.  We must wait for them to complete
+                * before merging the device fully.
+                * First we make sure any code which has tested
+                * our function has submitted the request, then
+                * we wait for all outstanding requests to complete.
+                */
+               synchronize_sched();
+               raise_barrier(conf, 0);
+               lower_barrier(conf);
+               clear_bit(Unmerged, &rdev->flags);
+       }
        md_integrity_add_rdev(rdev, mddev);
        print_conf(conf);
        return err;
@@ -2098,6 +2146,7 @@ static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10
                        d = r10_bio->devs[sl].devnum;
                        rdev = rcu_dereference(conf->mirrors[d].rdev);
                        if (rdev &&
+                           !test_bit(Unmerged, &rdev->flags) &&
                            test_bit(In_sync, &rdev->flags) &&
                            is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
                                        &first_bad, &bad_sectors) == 0) {
@@ -2151,6 +2200,7 @@ static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10
                        d = r10_bio->devs[sl].devnum;
                        rdev = rcu_dereference(conf->mirrors[d].rdev);
                        if (!rdev ||
+                           test_bit(Unmerged, &rdev->flags) ||
                            !test_bit(In_sync, &rdev->flags))
                                continue;
 
@@ -3273,15 +3323,6 @@ static int run(struct mddev *mddev)
 
                disk_stack_limits(mddev->gendisk, rdev->bdev,
                                  rdev->data_offset << 9);
-               /* as we don't honour merge_bvec_fn, we must never risk
-                * violating it, so limit max_segments to 1 lying
-                * within a single page.
-                */
-               if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
-                       blk_queue_max_segments(mddev->queue, 1);
-                       blk_queue_segment_boundary(mddev->queue,
-                                                  PAGE_CACHE_SIZE - 1);
-               }
 
                disk->head_position = 0;
        }
@@ -3345,8 +3386,7 @@ static int run(struct mddev *mddev)
                        mddev->queue->backing_dev_info.ra_pages = 2* stripe;
        }
 
-       if (conf->near_copies < conf->raid_disks)
-               blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
+       blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
 
        if (md_integrity_register(mddev))
                goto out_free_conf;