]> err.no Git - linux-2.6/commitdiff
md: give userspace control over removing failed devices when external metdata in use
authorNeilBrown <neilb@suse.de>
Wed, 6 Feb 2008 09:39:51 +0000 (01:39 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Wed, 6 Feb 2008 18:41:18 +0000 (10:41 -0800)
When a device fails, we must not allow an further writes to the array until
the device failure has been recorded in array metadata.  When metadata is
managed externally, this requires some synchronisation...

Allow/require userspace to explicitly remove failed devices from active
service in the array by writing 'none' to the 'slot' attribute.  If this
reduces the number of failed devices to 0, the write block will automatically
be lowered.

Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/md/md.c

index e2782a04012da1e2869a2351dcbb689d88537524..00788c56276ff1f5db5b2250faf0abeadabb16c1 100644 (file)
@@ -1891,20 +1891,44 @@ static ssize_t
 slot_store(mdk_rdev_t *rdev, const char *buf, size_t len)
 {
        char *e;
+       int err;
+       char nm[20];
        int slot = simple_strtoul(buf, &e, 10);
        if (strncmp(buf, "none", 4)==0)
                slot = -1;
        else if (e==buf || (*e && *e!= '\n'))
                return -EINVAL;
-       if (rdev->mddev->pers)
-               /* Cannot set slot in active array (yet) */
-               return -EBUSY;
-       if (slot >= rdev->mddev->raid_disks)
-               return -ENOSPC;
-       rdev->raid_disk = slot;
-       /* assume it is working */
-       rdev->flags = 0;
-       set_bit(In_sync, &rdev->flags);
+       if (rdev->mddev->pers) {
+               /* Setting 'slot' on an active array requires also
+                * updating the 'rd%d' link, and communicating
+                * with the personality with ->hot_*_disk.
+                * For now we only support removing
+                * failed/spare devices.  This normally happens automatically,
+                * but not when the metadata is externally managed.
+                */
+               if (slot != -1)
+                       return -EBUSY;
+               if (rdev->raid_disk == -1)
+                       return -EEXIST;
+               /* personality does all needed checks */
+               if (rdev->mddev->pers->hot_add_disk == NULL)
+                       return -EINVAL;
+               err = rdev->mddev->pers->
+                       hot_remove_disk(rdev->mddev, rdev->raid_disk);
+               if (err)
+                       return err;
+               sprintf(nm, "rd%d", rdev->raid_disk);
+               sysfs_remove_link(&rdev->mddev->kobj, nm);
+               set_bit(MD_RECOVERY_NEEDED, &rdev->mddev->recovery);
+               md_wakeup_thread(rdev->mddev->thread);
+       } else {
+               if (slot >= rdev->mddev->raid_disks)
+                       return -ENOSPC;
+               rdev->raid_disk = slot;
+               /* assume it is working */
+               rdev->flags = 0;
+               set_bit(In_sync, &rdev->flags);
+       }
        return len;
 }
 
@@ -5549,6 +5573,7 @@ static int remove_and_add_spares(mddev_t *mddev)
 
        ITERATE_RDEV(mddev,rdev,rtmp)
                if (rdev->raid_disk >= 0 &&
+                   !mddev->external &&
                    (test_bit(Faulty, &rdev->flags) ||
                     ! test_bit(In_sync, &rdev->flags)) &&
                    atomic_read(&rdev->nr_pending)==0) {