]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
md/raid1: clean up read_balance.
authorNeilBrown <neilb@suse.de>
Wed, 11 May 2011 04:34:56 +0000 (14:34 +1000)
committerNeilBrown <neilb@suse.de>
Wed, 11 May 2011 04:34:56 +0000 (14:34 +1000)
read_balance has two loops which both look for a 'best'
device based on slightly different criteria.
This is clumsy and makes is hard to add extra criteria.

So replace it all with a single loop that combines everything.

Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid1.c

index 2b7a7ff401dc6fba50047bbc3f0fb1980055da13..f0b0c79b3899bce3d4ef0d3c8ee084e6d9acc414 100644 (file)
@@ -411,10 +411,10 @@ static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 {
        const sector_t this_sector = r1_bio->sector;
        const int sectors = r1_bio->sectors;
-       int new_disk = -1;
        int start_disk;
+       int best_disk;
        int i;
-       sector_t new_distance, current_distance;
+       sector_t best_dist;
        mdk_rdev_t *rdev;
        int choose_first;
 
@@ -425,6 +425,8 @@ static int read_balance(conf_t *conf, r1bio_t *r1_bio)
         * We take the first readable disk when above the resync window.
         */
  retry:
+       best_disk = -1;
+       best_dist = MaxSector;
        if (conf->mddev->recovery_cp < MaxSector &&
            (this_sector + sectors >= conf->next_resync)) {
                choose_first = 1;
@@ -434,8 +436,8 @@ static int read_balance(conf_t *conf, r1bio_t *r1_bio)
                start_disk = conf->last_used;
        }
 
-       /* make sure the disk is operational */
        for (i = 0 ; i < conf->raid_disks ; i++) {
+               sector_t dist;
                int disk = start_disk + i;
                if (disk >= conf->raid_disks)
                        disk -= conf->raid_disks;
@@ -443,60 +445,43 @@ static int read_balance(conf_t *conf, r1bio_t *r1_bio)
                rdev = rcu_dereference(conf->mirrors[disk].rdev);
                if (r1_bio->bios[disk] == IO_BLOCKED
                    || rdev == NULL
-                   || !test_bit(In_sync, &rdev->flags))
+                   || test_bit(Faulty, &rdev->flags))
                        continue;
-
-               new_disk = disk;
-               if (!test_bit(WriteMostly, &rdev->flags))
-                       break;
-       }
-
-       if (new_disk < 0 || choose_first)
-               goto rb_out;
-
-       /*
-        * Don't change to another disk for sequential reads:
-        */
-       if (conf->next_seq_sect == this_sector)
-               goto rb_out;
-       if (this_sector == conf->mirrors[new_disk].head_position)
-               goto rb_out;
-
-       current_distance = abs(this_sector 
-                              - conf->mirrors[new_disk].head_position);
-
-       /* look for a better disk - i.e. head is closer */
-       start_disk = new_disk;
-       for (i = 1; i < conf->raid_disks; i++) {
-               int disk = start_disk + 1;
-               if (disk >= conf->raid_disks)
-                       disk -= conf->raid_disks;
-
-               rdev = rcu_dereference(conf->mirrors[disk].rdev);
-               if (r1_bio->bios[disk] == IO_BLOCKED
-                   || rdev == NULL
-                   || !test_bit(In_sync, &rdev->flags)
-                   || test_bit(WriteMostly, &rdev->flags))
+               if (!test_bit(In_sync, &rdev->flags) &&
+                   rdev->recovery_offset < this_sector + sectors)
                        continue;
-
-               if (!atomic_read(&rdev->nr_pending)) {
-                       new_disk = disk;
+               if (test_bit(WriteMostly, &rdev->flags)) {
+                       /* Don't balance among write-mostly, just
+                        * use the first as a last resort */
+                       if (best_disk < 0)
+                               best_disk = disk;
+                       continue;
+               }
+               /* This is a reasonable device to use.  It might
+                * even be best.
+                */
+               dist = abs(this_sector - conf->mirrors[disk].head_position);
+               if (choose_first
+                   /* Don't change to another disk for sequential reads */
+                   || conf->next_seq_sect == this_sector
+                   || dist == 0
+                   /* If device is idle, use it */
+                   || atomic_read(&rdev->nr_pending) == 0) {
+                       best_disk = disk;
                        break;
                }
-               new_distance = abs(this_sector - conf->mirrors[disk].head_position);
-               if (new_distance < current_distance) {
-                       current_distance = new_distance;
-                       new_disk = disk;
+               if (dist < best_dist) {
+                       best_dist = dist;
+                       best_disk = disk;
                }
        }
 
- rb_out:
-       if (new_disk >= 0) {
-               rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
+       if (best_disk >= 0) {
+               rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
                if (!rdev)
                        goto retry;
                atomic_inc(&rdev->nr_pending);
-               if (!test_bit(In_sync, &rdev->flags)) {
+               if (test_bit(Faulty, &rdev->flags)) {
                        /* cannot risk returning a device that failed
                         * before we inc'ed nr_pending
                         */
@@ -504,11 +489,11 @@ static int read_balance(conf_t *conf, r1bio_t *r1_bio)
                        goto retry;
                }
                conf->next_seq_sect = this_sector + sectors;
-               conf->last_used = new_disk;
+               conf->last_used = best_disk;
        }
        rcu_read_unlock();
 
-       return new_disk;
+       return best_disk;
 }
 
 static int raid1_congested(void *data, int bits)