md: resolve external metadata handling deadlock in md_allow_write

md_allow_write() marks the metadata dirty while holding mddev->lock and then waits for the write to complete. For externally managed metadata this causes a deadlock as userspace needs to take the lock to communicate that the metadata update has completed. Change md_allow_write() in the 'external' case to start the 'mark active' operation and then return -EAGAIN. The expected side effects while waiting for userspace to write 'active' to 'array_state' are holding off reshape (code currently handles -ENOMEM), cause some 'stripe_cache_size' change requests to fail, cause some GET_BITMAP_FILE ioctl requests to fall back to GFP_NOIO, and cause updates to 'raid_disks' to fail. Except for 'stripe_cache_size' changes these failures can be mitigated by coordinating with mdmon. md_write_start() still prevents writes from occurring until the metadata handler has had a chance to take action as it unconditionally waits for MD_CHANGE_CLEAN to be cleared. [neilb@suse.de: return -EAGAIN, try GFP_NOIO] Signed-off-by: Dan Williams <dan.j.williams@intel.com>

Dan Williams committed 18 years ago

Commit b5470dc1 parent 1fe797e4 files · +30 −17Browse files

Unified · Split

4 files

drivers/md/md.c +16 −11
drivers/md/raid1.c +4 −2
drivers/md/raid5.c +9 −3
include/linux/raid/md.h +1 −1

drivers/md/md.c+16−11

@@ -4172,9 +4172,11 @@ static int get_bitmap_file(mddev_t * mddev, void __user * arg) · 4171 unmodified lines
4172	4172	char ptr, buf = NULL;
4173	4173	int err = -ENOMEM;
4174	4174
4175		− md_allow_write(mddev);
	4175	+ if (md_allow_write(mddev))
	4176	+ file = kmalloc(sizeof(*file), GFP_NOIO);
	4177	+ else
	4178	+ file = kmalloc(sizeof(*file), GFP_KERNEL);
4176	4179
4177		− file = kmalloc(sizeof(*file), GFP_KERNEL);
4178	4180	if (!file)
4179	4181	goto out;
4180	4182
@@ -5667,15 +5669,18 @@ void md_write_end(mddev_t *mddev) · 1486 unmodified lines
5667	5669	* may proceed without blocking. It is important to call this before
5668	5670	* attempting a GFP_KERNEL allocation while holding the mddev lock.
5669	5671	* Must be called with mddev_lock held.
	5672	+ *
	5673	+ * In the ->external case MD_CHANGE_CLEAN can not be cleared until mddev->lock
	5674	+ * is dropped, so return -EAGAIN after notifying userspace.
5670	5675	*/
5671		−void md_allow_write(mddev_t *mddev)
	5676	+int md_allow_write(mddev_t *mddev)
5672	5677	{
5673	5678	if (!mddev->pers)
5674		− return;
	5679	+ return 0;
5675	5680	if (mddev->ro)
5676		− return;
	5681	+ return 0;
5677	5682	if (!mddev->pers->sync_request)
5678		− return;
	5683	+ return 0;
5679	5684
5680	5685	spin_lock_irq(&mddev->write_lock);
5681	5686	if (mddev->in_sync) {
@@ -5686,14 +5691,14 @@ void md_allow_write(mddev_t *mddev) · 4 unmodified lines
5686	5691	mddev->safemode = 1;
5687	5692	spin_unlock_irq(&mddev->write_lock);
5688	5693	md_update_sb(mddev, 0);
5689		−
5690	5694	sysfs_notify(&mddev->kobj, NULL, "array_state");
5691		− /* wait for the dirty state to be recorded in the metadata */
5692		− wait_event(mddev->sb_wait,
5693		− !test_bit(MD_CHANGE_CLEAN, &mddev->flags) &&
5694		− !test_bit(MD_CHANGE_PENDING, &mddev->flags));
5695	5695	} else
5696	5696	spin_unlock_irq(&mddev->write_lock);
	5697	+
	5698	+ if (test_bit(MD_CHANGE_CLEAN, &mddev->flags))
	5699	+ return -EAGAIN;
	5700	+ else
	5701	+ return 0;
5697	5702	}
5698	5703	EXPORT_SYMBOL_GPL(md_allow_write);
5699	5704

drivers/md/raid1.c+4−2

@@ -2136,7 +2136,7 @@ static int raid1_reshape(mddev_t *mddev) · 2135 unmodified lines
2136	2136	conf_t *conf = mddev_to_conf(mddev);
2137	2137	int cnt, raid_disks;
2138	2138	unsigned long flags;
2139		− int d, d2;
	2139	+ int d, d2, err;
2140	2140
2141	2141	/* Cannot change chunk_size, layout, or level */
2142	2142	if (mddev->chunk_size != mddev->new_chunk \|\|
@@ -2148,7 +2148,9 @@ static int raid1_reshape(mddev_t *mddev) · 5 unmodified lines
2148	2148	return -EINVAL;
2149	2149	}
2150	2150
2151		− md_allow_write(mddev);
	2151	+ err = md_allow_write(mddev);
	2152	+ if (err)
	2153	+ return err;
2152	2154
2153	2155	raid_disks = mddev->raid_disks + mddev->delta_disks;
2154	2156

drivers/md/raid5.c+9−3

@@ -911,14 +911,16 @@ static int resize_stripes(raid5_conf_t *conf, int newsize) · 910 unmodified lines
911	911	struct stripe_head osh, nsh;
912	912	LIST_HEAD(newstripes);
913	913	struct disk_info *ndisks;
914		− int err = 0;
	914	+ int err;
915	915	struct kmem_cache *sc;
916	916	int i;
917	917
918	918	if (newsize <= conf->pool_size)
919	919	return 0; /* never bother to shrink */
920	920
921		− md_allow_write(conf->mddev);
	921	+ err = md_allow_write(conf->mddev);
	922	+ if (err)
	923	+ return err;
922	924
923	925	/* Step 1 */
924	926	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
@@ -3843,6 +3845,8 @@ raid5_store_stripe_cache_size(mddev_t mddev, const char page, size_t len) · 2918 unmodified lines
3843	3845	{
3844	3846	raid5_conf_t *conf = mddev_to_conf(mddev);
3845	3847	unsigned long new;
	3848	+ int err;
	3849	+
3846	3850	if (len >= PAGE_SIZE)
3847	3851	return -EINVAL;
3848	3852	if (!conf)
@@ -3858,7 +3862,9 @@ raid5_store_stripe_cache_size(mddev_t mddev, const char page, size_t len) · 9 unmodified lines
3858	3862	else
3859	3863	break;
3860	3864	}
3861		− md_allow_write(mddev);
	3865	+ err = md_allow_write(mddev);
	3866	+ if (err)
	3867	+ return err;
3862	3868	while (new > conf->max_nr_stripes) {
3863	3869	if (grow_one_stripe(conf))
3864	3870	conf->max_nr_stripes++;

include/linux/raid/md.h+1−1

@@ -95,7 +95,7 @@ extern int sync_page_io(struct block_device *bdev, sector_t sector, int size, · 94 unmodified lines
95	95	struct page *page, int rw);
96	96	extern void md_do_sync(mddev_t *mddev);
97	97	extern void md_new_event(mddev_t *mddev);
98		−extern void md_allow_write(mddev_t *mddev);
	98	+extern int md_allow_write(mddev_t *mddev);
99	99	extern void md_wait_for_blocked_rdev(mdk_rdev_t rdev, mddev_t mddev);
100	100
101	101	#endif /* CONFIG_MD */

md: resolve external metadata handling deadlock in md_allow_write

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code