]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
md/raid10: simplify handle_read_error()
authorNeilBrown <neilb@suse.com>
Wed, 5 Apr 2017 04:05:51 +0000 (14:05 +1000)
committerShaohua Li <shli@fb.com>
Tue, 11 Apr 2017 17:15:08 +0000 (10:15 -0700)
handle_read_error() duplicates a lot of the work that raid10_read_request()
does, so it makes sense to just use that function.

handle_read_error() relies on the same r10bio being re-used so that,
in the case of a read-only array, setting IO_BLOCKED in r1bio->devs[].bio
ensures read_balance() won't re-use that device.
So when called from raid10_make_request() we clear that array, but not
when called from handle_read_error().

Two parts of handle_read_error() that need to be preserved are the warning
message it prints, so they are conditionally added to
raid10_read_request().  If the failing rdev can be found, messages
are printed.  Otherwise they aren't.

Not that as rdev_dec_pending() has already been called on the failing
rdev, we need to use rcu_read_lock() to get a new reference from
the conf.  We only use this to get the name of the failing block device.

With this change, we no longer need inc_pending().

Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
drivers/md/raid10.c

index 41845bae67be76cdaa1c490df22293cfed23d6cf..4167091eff9a22ec9cab6a69be166d757b08e6b0 100644 (file)
@@ -1008,15 +1008,6 @@ static void wait_barrier(struct r10conf *conf)
        spin_unlock_irq(&conf->resync_lock);
 }
 
-static void inc_pending(struct r10conf *conf)
-{
-       /* The current request requires multiple r10_bio, so
-        * we need to increment the pending count.
-        */
-       WARN_ON(!atomic_read(&conf->nr_pending));
-       atomic_inc(&conf->nr_pending);
-}
-
 static void allow_barrier(struct r10conf *conf)
 {
        if ((atomic_dec_and_test(&conf->nr_pending)) ||
@@ -1130,8 +1121,38 @@ static void raid10_read_request(struct mddev *mddev, struct bio *bio,
        int max_sectors;
        sector_t sectors;
        struct md_rdev *rdev;
-       int slot;
+       char b[BDEVNAME_SIZE];
+       int slot = r10_bio->read_slot;
+       struct md_rdev *err_rdev = NULL;
+       gfp_t gfp = GFP_NOIO;
+
+       if (r10_bio->devs[slot].rdev) {
+               /*
+                * This is an error retry, but we cannot
+                * safely dereference the rdev in the r10_bio,
+                * we must use the one in conf.
+                * If it has already been disconnected (unlikely)
+                * we lose the device name in error messages.
+                */
+               int disk;
+               /*
+                * As we are blocking raid10, it is a little safer to
+                * use __GFP_HIGH.
+                */
+               gfp = GFP_NOIO | __GFP_HIGH;
 
+               rcu_read_lock();
+               disk = r10_bio->devs[slot].devnum;
+               err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
+               if (err_rdev)
+                       bdevname(err_rdev->bdev, b);
+               else {
+                       strcpy(b, "???");
+                       /* This never gets dereferenced */
+                       err_rdev = r10_bio->devs[slot].rdev;
+               }
+               rcu_read_unlock();
+       }
        /*
         * Register the new request and wait if the reconstruction
         * thread has put up a bar for new requests.
@@ -1158,12 +1179,22 @@ static void raid10_read_request(struct mddev *mddev, struct bio *bio,
 
        rdev = read_balance(conf, r10_bio, &max_sectors);
        if (!rdev) {
+               if (err_rdev) {
+                       pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
+                                           mdname(mddev), b,
+                                           (unsigned long long)r10_bio->sector);
+               }
                raid_end_bio_io(r10_bio);
                return;
        }
+       if (err_rdev)
+               pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
+                                  mdname(mddev),
+                                  bdevname(rdev->bdev, b),
+                                  (unsigned long long)r10_bio->sector);
        if (max_sectors < bio_sectors(bio)) {
                struct bio *split = bio_split(bio, max_sectors,
-                                             GFP_NOIO, conf->bio_split);
+                                             gfp, conf->bio_split);
                bio_chain(split, bio);
                generic_make_request(bio);
                bio = split;
@@ -1172,7 +1203,7 @@ static void raid10_read_request(struct mddev *mddev, struct bio *bio,
        }
        slot = r10_bio->read_slot;
 
-       read_bio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
+       read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
 
        r10_bio->devs[slot].bio = read_bio;
        r10_bio->devs[slot].rdev = rdev;
@@ -1487,6 +1518,7 @@ static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
        r10_bio->mddev = mddev;
        r10_bio->sector = bio->bi_iter.bi_sector;
        r10_bio->state = 0;
+       memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
 
        if (bio_data_dir(bio) == READ)
                raid10_read_request(mddev, bio, r10_bio);
@@ -2556,9 +2588,6 @@ static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
        struct bio *bio;
        struct r10conf *conf = mddev->private;
        struct md_rdev *rdev = r10_bio->devs[slot].rdev;
-       char b[BDEVNAME_SIZE];
-       unsigned long do_sync;
-       int max_sectors;
        dev_t bio_dev;
        sector_t bio_last_sector;
 
@@ -2571,7 +2600,6 @@ static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
         * frozen.
         */
        bio = r10_bio->devs[slot].bio;
-       bdevname(bio->bi_bdev, b);
        bio_dev = bio->bi_bdev->bd_dev;
        bio_last_sector = r10_bio->devs[slot].addr + rdev->data_offset + r10_bio->sectors;
        bio_put(bio);
@@ -2587,65 +2615,9 @@ static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
                md_error(mddev, rdev);
 
        rdev_dec_pending(rdev, mddev);
-
-read_more:
-       rdev = read_balance(conf, r10_bio, &max_sectors);
-       if (rdev == NULL) {
-               pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
-                                   mdname(mddev), b,
-                                   (unsigned long long)r10_bio->sector);
-               raid_end_bio_io(r10_bio);
-               return;
-       }
-
-       do_sync = (r10_bio->master_bio->bi_opf & REQ_SYNC);
-       slot = r10_bio->read_slot;
-       pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
-                          mdname(mddev),
-                          bdevname(rdev->bdev, b),
-                          (unsigned long long)r10_bio->sector);
-       bio = bio_clone_fast(r10_bio->master_bio, GFP_NOIO, mddev->bio_set);
-       bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
-       r10_bio->devs[slot].bio = bio;
-       r10_bio->devs[slot].rdev = rdev;
-       bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
-               + choose_data_offset(r10_bio, rdev);
-       bio->bi_bdev = rdev->bdev;
-       bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
-       if (test_bit(FailFast, &rdev->flags) &&
-           test_bit(R10BIO_FailFast, &r10_bio->state))
-               bio->bi_opf |= MD_FAILFAST;
-       bio->bi_private = r10_bio;
-       bio->bi_end_io = raid10_end_read_request;
-       trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
-                             bio, bio_dev,
-                             bio_last_sector - r10_bio->sectors);
-
-       if (max_sectors < r10_bio->sectors) {
-               /* Drat - have to split this up more */
-               struct bio *mbio = r10_bio->master_bio;
-               int sectors_handled =
-                       r10_bio->sector + max_sectors
-                       - mbio->bi_iter.bi_sector;
-               r10_bio->sectors = max_sectors;
-               bio_inc_remaining(mbio);
-               inc_pending(conf);
-               generic_make_request(bio);
-
-               r10_bio = mempool_alloc(conf->r10bio_pool,
-                                       GFP_NOIO);
-               r10_bio->master_bio = mbio;
-               r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
-               r10_bio->state = 0;
-               set_bit(R10BIO_ReadError,
-                       &r10_bio->state);
-               r10_bio->mddev = mddev;
-               r10_bio->sector = mbio->bi_iter.bi_sector
-                       + sectors_handled;
-
-               goto read_more;
-       } else
-               generic_make_request(bio);
+       allow_barrier(conf);
+       r10_bio->state = 0;
+       raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
 }
 
 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)