md/linear: remove rcu protections in favour of suspend/resume
authorNeilBrown <neilb@suse.de>
Mon, 15 Dec 2014 01:56:57 +0000 (12:56 +1100)
committerNeilBrown <neilb@suse.de>
Tue, 3 Feb 2015 21:35:52 +0000 (08:35 +1100)
The use of 'rcu' to protect accesses to ->private_data so that
the ->private_data could be updated predates the introduction
of mddev_suspend/mddev_resume.
These are a cleaner mechanism for providing stability while
swapping in a new ->private data - it is used by level_store()
to support changing of raid levels.

So get rid of the RCU stuff and just use mddev_suspend, mddev_resume.

As these function call ->quiesce(), we add an empty function for
linear just like for raid0.

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

index 4c2a92ce2b0b96553f9cbbb27d7a5b7db8999bb2..b3e717adbc9b480ad9bf69e2142ce44b545f3bc7 100644 (file)
@@ -34,7 +34,7 @@ static inline struct dev_info *which_dev(struct mddev *mddev, sector_t sector)
 
        lo = 0;
        hi = mddev->raid_disks - 1;
-       conf = rcu_dereference(mddev->private);
+       conf = mddev->private;
 
        /*
         * Binary Search
@@ -70,7 +70,6 @@ static int linear_mergeable_bvec(struct mddev *mddev,
        int maxbytes = biovec->bv_len;
        struct request_queue *subq;
 
-       rcu_read_lock();
        dev0 = which_dev(mddev, sector);
        maxsectors = dev0->end_sector - sector;
        subq = bdev_get_queue(dev0->rdev->bdev);
@@ -80,7 +79,6 @@ static int linear_mergeable_bvec(struct mddev *mddev,
                maxbytes = min(maxbytes, subq->merge_bvec_fn(subq, bvm,
                                                             biovec));
        }
-       rcu_read_unlock();
 
        if (maxsectors < bio_sectors)
                maxsectors = 0;
@@ -101,15 +99,13 @@ static int linear_congested(struct mddev *mddev, int bits)
        struct linear_conf *conf;
        int i, ret = 0;
 
-       rcu_read_lock();
-       conf = rcu_dereference(mddev->private);
+       conf = mddev->private;
 
        for (i = 0; i < mddev->raid_disks && !ret ; i++) {
                struct request_queue *q = bdev_get_queue(conf->disks[i].rdev->bdev);
                ret |= bdi_congested(&q->backing_dev_info, bits);
        }
 
-       rcu_read_unlock();
        return ret;
 }
 
@@ -118,12 +114,10 @@ static sector_t linear_size(struct mddev *mddev, sector_t sectors, int raid_disk
        struct linear_conf *conf;
        sector_t array_sectors;
 
-       rcu_read_lock();
-       conf = rcu_dereference(mddev->private);
+       conf = mddev->private;
        WARN_ONCE(sectors || raid_disks,
                  "%s does not support generic reshape\n", __func__);
        array_sectors = conf->array_sectors;
-       rcu_read_unlock();
 
        return array_sectors;
 }
@@ -243,33 +237,22 @@ static int linear_add(struct mddev *mddev, struct md_rdev *rdev)
        if (!newconf)
                return -ENOMEM;
 
-       oldconf = rcu_dereference_protected(mddev->private,
-                                           lockdep_is_held(
-                                                   &mddev->reconfig_mutex));
+       mddev_suspend(mddev);
+       oldconf = mddev->private;
        mddev->raid_disks++;
-       rcu_assign_pointer(mddev->private, newconf);
+       mddev->private = newconf;
        md_set_array_sectors(mddev, linear_size(mddev, 0, 0));
        set_capacity(mddev->gendisk, mddev->array_sectors);
+       mddev_resume(mddev);
        revalidate_disk(mddev->gendisk);
-       kfree_rcu(oldconf, rcu);
+       kfree(oldconf);
        return 0;
 }
 
 static int linear_stop (struct mddev *mddev)
 {
-       struct linear_conf *conf =
-               rcu_dereference_protected(mddev->private,
-                                         lockdep_is_held(
-                                                 &mddev->reconfig_mutex));
+       struct linear_conf *conf = mddev->private;
 
-       /*
-        * We do not require rcu protection here since
-        * we hold reconfig_mutex for both linear_add and
-        * linear_stop, so they cannot race.
-        * We should make sure any old 'conf's are properly
-        * freed though.
-        */
-       rcu_barrier();
        blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
        kfree(conf);
        mddev->private = NULL;
@@ -290,16 +273,12 @@ static void linear_make_request(struct mddev *mddev, struct bio *bio)
        }
 
        do {
-               rcu_read_lock();
-
                tmp_dev = which_dev(mddev, bio->bi_iter.bi_sector);
                start_sector = tmp_dev->end_sector - tmp_dev->rdev->sectors;
                end_sector = tmp_dev->end_sector;
                data_offset = tmp_dev->rdev->data_offset;
                bio->bi_bdev = tmp_dev->rdev->bdev;
 
-               rcu_read_unlock();
-
                if (unlikely(bio->bi_iter.bi_sector >= end_sector ||
                             bio->bi_iter.bi_sector < start_sector))
                        goto out_of_bounds;
@@ -346,6 +325,10 @@ static void linear_status (struct seq_file *seq, struct mddev *mddev)
        seq_printf(seq, " %dk rounding", mddev->chunk_sectors / 2);
 }
 
+static void linear_quiesce(struct mddev *mddev, int state)
+{
+}
+
 static struct md_personality linear_personality =
 {
        .name           = "linear",
@@ -357,6 +340,7 @@ static struct md_personality linear_personality =
        .status         = linear_status,
        .hot_add_disk   = linear_add,
        .size           = linear_size,
+       .quiesce        = linear_quiesce,
        .congested      = linear_congested,
        .mergeable_bvec = linear_mergeable_bvec,
 };