[PATCH] md: improve handing of read errors with raid6
NeilBrown [Fri, 6 Jan 2006 08:20:18 +0000 (00:20 -0800)]
This is a simple port of match functionality across from raid5.  If we get a
read error, we don't kick the drive straight away, but try to over-write with
good data first.

Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

drivers/md/raid6main.c

index 52e8796..7a51553 100644 (file)
@@ -367,8 +367,8 @@ static void shrink_stripes(raid6_conf_t *conf)
        conf->slab_cache = NULL;
 }
 
-static int raid6_end_read_request (struct bio * bi, unsigned int bytes_done,
-                                  int error)
+static int raid6_end_read_request(struct bio * bi, unsigned int bytes_done,
+                                 int error)
 {
        struct stripe_head *sh = bi->bi_private;
        raid6_conf_t *conf = sh->raid_conf;
@@ -420,9 +420,35 @@ static int raid6_end_read_request (struct bio * bi, unsigned int bytes_done,
 #else
                set_bit(R5_UPTODATE, &sh->dev[i].flags);
 #endif
+               if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
+                       printk(KERN_INFO "raid6: read error corrected!!\n");
+                       clear_bit(R5_ReadError, &sh->dev[i].flags);
+                       clear_bit(R5_ReWrite, &sh->dev[i].flags);
+               }
+               if (atomic_read(&conf->disks[i].rdev->read_errors))
+                       atomic_set(&conf->disks[i].rdev->read_errors, 0);
        } else {
-               md_error(conf->mddev, conf->disks[i].rdev);
+               int retry = 0;
                clear_bit(R5_UPTODATE, &sh->dev[i].flags);
+               atomic_inc(&conf->disks[i].rdev->read_errors);
+               if (conf->mddev->degraded)
+                       printk(KERN_WARNING "raid6: read error not correctable.\n");
+               else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
+                       /* Oh, no!!! */
+                       printk(KERN_WARNING "raid6: read error NOT corrected!!\n");
+               else if (atomic_read(&conf->disks[i].rdev->read_errors)
+                        > conf->max_nr_stripes)
+                       printk(KERN_WARNING
+                              "raid6: Too many read errors, failing device.\n");
+               else
+                       retry = 1;
+               if (retry)
+                       set_bit(R5_ReadError, &sh->dev[i].flags);
+               else {
+                       clear_bit(R5_ReadError, &sh->dev[i].flags);
+                       clear_bit(R5_ReWrite, &sh->dev[i].flags);
+                       md_error(conf->mddev, conf->disks[i].rdev);
+               }
        }
        rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 #if 0
@@ -1079,6 +1105,12 @@ static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
                if (dev->written) written++;
                rdev = conf->disks[i].rdev; /* FIXME, should I be looking rdev */
                if (!rdev || !test_bit(In_sync, &rdev->flags)) {
+                       /* The ReadError flag will just be confusing now */
+                       clear_bit(R5_ReadError, &dev->flags);
+                       clear_bit(R5_ReWrite, &dev->flags);
+               }
+               if (!rdev || !test_bit(In_sync, &rdev->flags)
+                   || test_bit(R5_ReadError, &dev->flags)) {
                        if ( failed < 2 )
                                failed_num[failed] = i;
                        failed++;
@@ -1095,6 +1127,14 @@ static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
        if (failed > 2 && to_read+to_write+written) {
                for (i=disks; i--; ) {
                        int bitmap_end = 0;
+
+                       if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
+                               mdk_rdev_t *rdev = conf->disks[i].rdev;
+                               if (rdev && test_bit(In_sync, &rdev->flags))
+                                       /* multiple read failures in one stripe */
+                                       md_error(conf->mddev, rdev);
+                       }
+
                        spin_lock_irq(&conf->device_lock);
                        /* fail all writes first */
                        bi = sh->dev[i].towrite;
@@ -1130,7 +1170,8 @@ static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
                        }
 
                        /* fail any reads if this device is non-operational */
-                       if (!test_bit(R5_Insync, &sh->dev[i].flags)) {
+                       if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
+                           test_bit(R5_ReadError, &sh->dev[i].flags)) {
                                bi = sh->dev[i].toread;
                                sh->dev[i].toread = NULL;
                                if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
@@ -1457,6 +1498,27 @@ static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
                clear_bit(STRIPE_SYNCING, &sh->state);
        }
 
+       /* If the failed drives are just a ReadError, then we might need
+        * to progress the repair/check process
+        */
+       if (failed <= 2 && ! conf->mddev->ro)
+               for (i=0; i<failed;i++) {
+                       dev = &sh->dev[failed_num[i]];
+                       if (test_bit(R5_ReadError, &dev->flags)
+                           && !test_bit(R5_LOCKED, &dev->flags)
+                           && test_bit(R5_UPTODATE, &dev->flags)
+                               ) {
+                               if (!test_bit(R5_ReWrite, &dev->flags)) {
+                                       set_bit(R5_Wantwrite, &dev->flags);
+                                       set_bit(R5_ReWrite, &dev->flags);
+                                       set_bit(R5_LOCKED, &dev->flags);
+                               } else {
+                                       /* let's read it back */
+                                       set_bit(R5_Wantread, &dev->flags);
+                                       set_bit(R5_LOCKED, &dev->flags);
+                               }
+                       }
+               }
        spin_unlock(&sh->lock);
 
        while ((bi=return_bi)) {