dm: Start pr_preempt from the same starting path
authorMike Christie <michael.christie@oracle.com>
Sun, 17 Jul 2022 22:45:08 +0000 (17:45 -0500)
committerMike Snitzer <snitzer@kernel.org>
Thu, 28 Jul 2022 21:29:56 +0000 (17:29 -0400)
pr_preempt has a similar issue as reserve where for all the
reservation types except the All Registrants ones the preempt can
create a reservation. And a follow up reservation or release needs to
go down the same path the preempt did. This has the pr_preempt work
like reserve and release where we always start from the first path in
the first group.

This commit has been tested with windows failover clustering's
validation test and libiscsi's PGR tests to check for regressions.
They both don't have tests to verify this case, so I tested it
manually.

Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
drivers/md/dm.c

index 9ce87dd..47bcc50 100644 (file)
@@ -3076,6 +3076,7 @@ struct dm_pr {
        u64     old_key;
        u64     new_key;
        u32     flags;
+       bool    abort;
        bool    fail_early;
        int     ret;
        enum pr_type type;
@@ -3254,25 +3255,41 @@ static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
        return pr.ret;
 }
 
+static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
+                          sector_t start, sector_t len, void *data)
+{
+       struct dm_pr *pr = data;
+       const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
+
+       if (!ops || !ops->pr_preempt) {
+               pr->ret = -EOPNOTSUPP;
+               return -1;
+       }
+
+       pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
+                                 pr->abort);
+       if (!pr->ret)
+               return -1;
+
+       return 0;
+}
+
 static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
                         enum pr_type type, bool abort)
 {
-       struct mapped_device *md = bdev->bd_disk->private_data;
-       const struct pr_ops *ops;
-       int r, srcu_idx;
+       struct dm_pr pr = {
+               .new_key        = new_key,
+               .old_key        = old_key,
+               .type           = type,
+               .fail_early     = false,
+       };
+       int ret;
 
-       r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
-       if (r < 0)
-               goto out;
+       ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
+       if (ret)
+               return ret;
 
-       ops = bdev->bd_disk->fops->pr_ops;
-       if (ops && ops->pr_preempt)
-               r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
-       else
-               r = -EOPNOTSUPP;
-out:
-       dm_unprepare_ioctl(md, srcu_idx);
-       return r;
+       return pr.ret;
 }
 
 static int dm_pr_clear(struct block_device *bdev, u64 key)