]> err.no Git - linux-2.6/commitdiff
md: allow devices to be shared between md arrays
authorNeilBrown <neilb@suse.de>
Wed, 6 Feb 2008 09:39:54 +0000 (01:39 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Wed, 6 Feb 2008 18:41:18 +0000 (10:41 -0800)
Currently, a given device is "claimed" by a particular array so that it cannot
be used by other arrays.

This is not ideal for DDF and other metadata schemes which have their own
partitioning concept.

So for externally managed metadata, just claim the device for md in general,
require that "offset" and "size" are set properly for each device, and make
sure that if a device is included in different arrays then the active sections
do not overlap.

This involves adding another flag to the rdev which makes it awkward to set
"->flags = 0" to clear certain flags.  So now clear flags explicitly by name
when we want to clear things.

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
include/linux/raid/md_k.h

index 78fe3e97ff993293c1d018431c8f5bf1e2531f16..7c9a87b02e7784259d751ed9a57ce02f4a7409a1 100644 (file)
@@ -774,7 +774,11 @@ static int super_90_validate(mddev_t *mddev, mdk_rdev_t *rdev)
        __u64 ev1 = md_event(sb);
 
        rdev->raid_disk = -1;
-       rdev->flags = 0;
+       clear_bit(Faulty, &rdev->flags);
+       clear_bit(In_sync, &rdev->flags);
+       clear_bit(WriteMostly, &rdev->flags);
+       clear_bit(BarriersNotsupp, &rdev->flags);
+
        if (mddev->raid_disks == 0) {
                mddev->major_version = 0;
                mddev->minor_version = sb->minor_version;
@@ -1154,7 +1158,11 @@ static int super_1_validate(mddev_t *mddev, mdk_rdev_t *rdev)
        __u64 ev1 = le64_to_cpu(sb->events);
 
        rdev->raid_disk = -1;
-       rdev->flags = 0;
+       clear_bit(Faulty, &rdev->flags);
+       clear_bit(In_sync, &rdev->flags);
+       clear_bit(WriteMostly, &rdev->flags);
+       clear_bit(BarriersNotsupp, &rdev->flags);
+
        if (mddev->raid_disks == 0) {
                mddev->major_version = 1;
                mddev->patch_version = 0;
@@ -1402,7 +1410,7 @@ static int bind_rdev_to_array(mdk_rdev_t * rdev, mddev_t * mddev)
                goto fail;
        }
        list_add(&rdev->same_set, &mddev->disks);
-       bd_claim_by_disk(rdev->bdev, rdev, mddev->gendisk);
+       bd_claim_by_disk(rdev->bdev, rdev->bdev->bd_holder, mddev->gendisk);
        return 0;
 
  fail:
@@ -1442,7 +1450,7 @@ static void unbind_rdev_from_array(mdk_rdev_t * rdev)
  * otherwise reused by a RAID array (or any other kernel
  * subsystem), by bd_claiming the device.
  */
-static int lock_rdev(mdk_rdev_t *rdev, dev_t dev)
+static int lock_rdev(mdk_rdev_t *rdev, dev_t dev, int shared)
 {
        int err = 0;
        struct block_device *bdev;
@@ -1454,13 +1462,15 @@ static int lock_rdev(mdk_rdev_t *rdev, dev_t dev)
                        __bdevname(dev, b));
                return PTR_ERR(bdev);
        }
-       err = bd_claim(bdev, rdev);
+       err = bd_claim(bdev, shared ? (mdk_rdev_t *)lock_rdev : rdev);
        if (err) {
                printk(KERN_ERR "md: could not bd_claim %s.\n",
                        bdevname(bdev, b));
                blkdev_put(bdev);
                return err;
        }
+       if (!shared)
+               set_bit(AllReserved, &rdev->flags);
        rdev->bdev = bdev;
        return err;
 }
@@ -1925,7 +1935,8 @@ slot_store(mdk_rdev_t *rdev, const char *buf, size_t len)
                        return -ENOSPC;
                rdev->raid_disk = slot;
                /* assume it is working */
-               rdev->flags = 0;
+               clear_bit(Faulty, &rdev->flags);
+               clear_bit(WriteMostly, &rdev->flags);
                set_bit(In_sync, &rdev->flags);
        }
        return len;
@@ -1950,6 +1961,10 @@ offset_store(mdk_rdev_t *rdev, const char *buf, size_t len)
                return -EINVAL;
        if (rdev->mddev->pers)
                return -EBUSY;
+       if (rdev->size && rdev->mddev->external)
+               /* Must set offset before size, so overlap checks
+                * can be sane */
+               return -EBUSY;
        rdev->data_offset = offset;
        return len;
 }
@@ -1963,16 +1978,69 @@ rdev_size_show(mdk_rdev_t *rdev, char *page)
        return sprintf(page, "%llu\n", (unsigned long long)rdev->size);
 }
 
+static int overlaps(sector_t s1, sector_t l1, sector_t s2, sector_t l2)
+{
+       /* check if two start/length pairs overlap */
+       if (s1+l1 <= s2)
+               return 0;
+       if (s2+l2 <= s1)
+               return 0;
+       return 1;
+}
+
 static ssize_t
 rdev_size_store(mdk_rdev_t *rdev, const char *buf, size_t len)
 {
        char *e;
        unsigned long long size = simple_strtoull(buf, &e, 10);
+       unsigned long long oldsize = rdev->size;
        if (e==buf || (*e && *e != '\n'))
                return -EINVAL;
        if (rdev->mddev->pers)
                return -EBUSY;
        rdev->size = size;
+       if (size > oldsize && rdev->mddev->external) {
+               /* need to check that all other rdevs with the same ->bdev
+                * do not overlap.  We need to unlock the mddev to avoid
+                * a deadlock.  We have already changed rdev->size, and if
+                * we have to change it back, we will have the lock again.
+                */
+               mddev_t *mddev;
+               int overlap = 0;
+               struct list_head *tmp, *tmp2;
+
+               mddev_unlock(rdev->mddev);
+               ITERATE_MDDEV(mddev, tmp) {
+                       mdk_rdev_t *rdev2;
+
+                       mddev_lock(mddev);
+                       ITERATE_RDEV(mddev, rdev2, tmp2)
+                               if (test_bit(AllReserved, &rdev2->flags) ||
+                                   (rdev->bdev == rdev2->bdev &&
+                                    rdev != rdev2 &&
+                                    overlaps(rdev->data_offset, rdev->size,
+                                           rdev2->data_offset, rdev2->size))) {
+                                       overlap = 1;
+                                       break;
+                               }
+                       mddev_unlock(mddev);
+                       if (overlap) {
+                               mddev_put(mddev);
+                               break;
+                       }
+               }
+               mddev_lock(rdev->mddev);
+               if (overlap) {
+                       /* Someone else could have slipped in a size
+                        * change here, but doing so is just silly.
+                        * We put oldsize back because we *know* it is
+                        * safe, and trust userspace not to race with
+                        * itself
+                        */
+                       rdev->size = oldsize;
+                       return -EBUSY;
+               }
+       }
        if (size < rdev->mddev->size || rdev->mddev->size == 0)
                rdev->mddev->size = size;
        return len;
@@ -2056,7 +2124,7 @@ static mdk_rdev_t *md_import_device(dev_t newdev, int super_format, int super_mi
        if ((err = alloc_disk_sb(rdev)))
                goto abort_free;
 
-       err = lock_rdev(rdev, newdev);
+       err = lock_rdev(rdev, newdev, super_format == -2);
        if (err)
                goto abort_free;
 
@@ -2609,7 +2677,9 @@ new_dev_store(mddev_t *mddev, const char *buf, size_t len)
                        if (err < 0)
                                goto out;
                }
-       } else
+       } else if (mddev->external)
+               rdev = md_import_device(dev, -2, -1);
+       else
                rdev = md_import_device(dev, -1, -1);
 
        if (IS_ERR(rdev))
@@ -4019,8 +4089,6 @@ static int add_new_disk(mddev_t * mddev, mdu_disk_info_t *info)
                else
                        rdev->raid_disk = -1;
 
-               rdev->flags = 0;
-
                if (rdev->raid_disk < mddev->raid_disks)
                        if (info->state & (1<<MD_DISK_SYNC))
                                set_bit(In_sync, &rdev->flags);
index c77dca3221ed5891dcb6cae277b8470f487ab5a5..5b2102e40286dd40eb3f469774743778a1b301ac 100644 (file)
@@ -81,6 +81,8 @@ struct mdk_rdev_s
 #define        In_sync         2               /* device is in_sync with rest of array */
 #define        WriteMostly     4               /* Avoid reading if at all possible */
 #define        BarriersNotsupp 5               /* BIO_RW_BARRIER is not supported */
+#define        AllReserved     6               /* If whole device is reserved for
+                                        * one array */
 
        int desc_nr;                    /* descriptor index in the superblock */
        int raid_disk;                  /* role of device in array */