A thin-p over 256 GiB fails with I/O errors with non-power-of-two chunk

Why do I get the following error, and what should I do about it? When I create a raid0 md with a non-power-of-two chunk size (e.g. 1152K instead of 512K), then create a thinly-provisioned volume that is over 256 GiB, I get the following dmesg error when I try to create a file system on it:

"make_request bug: can't convert block across chunks or bigger than 1152k 4384 127"

This bubbles up to mkfs.xfs as

"libxfs_device_zero write failed: Input/output error"

What I find interesting is that it seems to require all three conditions (chunk size, thin-p, and >256 GiB) in order to fail. Without those, it seems to work fine:

* Power-of-two chunk (e.g. 512K), thin-p vol, >256 GiB? Works.
* Non-power-of-two chunk (e.g. 1152K), thin-p vol, <256 GiB? Works.
* Non-power-of-two chunk (e.g. 1152K), regular vol, >256 GiB? Works.
* Non-power-of-two chunk (e.g. 1152K), thin-p vol, >256 GiB? FAIL.

Attached is a self-contained test case to reproduce the error, version numbers, and an strace. Thank you in advance,
--
Daniel Browning
Kavod Technologies

Appendix A. Self-contained reproduce script
===========================================================
dd if=/dev/zero of=loop0.img bs=1G count=150; losetup /dev/loop0 loop0.img dd if=/dev/zero of=loop1.img bs=1G count=150; losetup /dev/loop1 loop1.img

mdadm --create /dev/md99 --verbose --level=0 --raid-devices=2 \
      --chunk=1152K /dev/loop0 /dev/loop1

pvcreate /dev/md99
vgcreate test_vg /dev/md99
lvcreate --size 257G --type thin-pool --thinpool test_thin_pool test_vg lvcreate --virtualsize 257G --thin test_vg/test_thin_pool --name test_lv mkfs.xfs /dev/test_vg/test_lv

# That is where the error occurs. Next is cleanup. lvremove -f /dev/test_vg/test_lv
lvremove -f /dev/mapper/test_vg-test_thin_pool vgremove -f test_vg
pvremove /dev/md99
mdadm --stop /dev/md99
mdadm --zero-superblock /dev/loop0 /dev/loop1 losetup -d /dev/loop0 /dev/loop1
rm loop*.img

Appendix B. Versions
===========================================================

Distro:          CentOS 6.3
Kernel:          3.7.2-1.el6xen.x86_64 from dev.crc.id.au
LVM version:     2.02.99(2)-git (2012-10-22)
Library version: 1.02.78-git (2012-10-22)
Driver version:  4.23.0
XFS userspace:   xfsprogs-3.1.1-7.el6.x86_64

Appendix C. strace of mkfs.xfs
===========================================================
See http://pastebin.com/raw.php?i=hLLm0jVC for the full strace. An excerpt:

lseek(4, 137975840768, SEEK_SET) = 137975840768
write(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 262144) = 262144 write(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 262144) = 262144 write(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 262144) = -1 EIO (Input/output error)

Limits of the raid0 device (/dev/md99):
cat /sys/block/md99/queue/minimum_io_size
1179648
cat /sys/block/md99/queue/optimal_io_size
2359296

Limits of the thin-pool device (/dev/test_vg/test_thin_pool):
cat /sys/block/dm-9/queue/minimum_io_size
512
cat /sys/block/dm-9/queue/optimal_io_size
262144

Limits of the thin-device device (/dev/test_vg/test_lv):
cat /sys/block/dm-10/queue/minimum_io_size
512
cat /sys/block/dm-10/queue/optimal_io_size 262144

I notice that lvcreate is not using a thin-pool chunksize that matches the raid0's chunksize (just uses the lvm2 default of 256K).

Switching the thin-pool lvcreate to use --chunksize 1152K at least enables me to format the filesystem.

And both the thin-pool and thin device have an optimal_io_size that matches the chunk_size of the underlying raid volume:

cat /sys/block/dm-9/queue/optimal_io_size
1179648
cat /sys/block/dm-10/queue/optimal_io_size
1179648

I'm still investigating the limits issue when --chunksize 1152K isn't used for the thin-pool lvcreate.

Dne 21.1.2013 19:49, Mike Snitzer napsal(a):

Just a comment for the selection of thin chunksize here -

I think it has couple aspects here - by default (unless changed via lvm.conf {allocation/thin_pool_chunk_size}) it is targeting for 64K and scales chunksize up to fit thin metadata within 128MB. (compiled in as DEFAULT_THIN_POOL_OPTIMAL_SIZE)
So lvm2 here scaled from 64k to 256k in multiTB case.

lvcreate currently doesn't look out for geometry of underlying PV(s) during its allocation (somewhat chicken-egg problem) - yet there are possible ways to try to put this into equation - thought it might not be actually wanted by the user - since for snapshots the smaller chunksize is more usable (>1MB is quite a lot here IMHO) - but it probably worth some thinking.

Zdenek

Not quite sure what you mean by "to fit thin metadata within 128MB". Why is fitting within 128MB the goal? I recall Joe helping to establish the rule of thumb for lvm2 but I don't recall specifics at this point.

I've found that the mkfs.xfs (which uses direct IO) will work if the thinp chunksize is a factor of the raid0 chunksize. So all of the following thinp chunksizes "work" given that the raid0 chunksize is 1152K: 64K, 128K, 384K, 576K, 1152K

I haven't done extensive IO testing on the result XFS filesystem though. So I don't want to get too far into shaping lvm2's chunksize selection algorithm until I can dive into the kernel's limits stacking further (which I'm doing now).

When a thin-pool uses an MD device for the data device a thin device from the thin-pool must respect MD's constraints about disallowing a bio to span multiple chunks. Otherwise we can see problems. If the raid0 chunksize is 1152K and thin-pool chunksize is 256K I see the following md/radi0 error (with extra debug tracing added to thin_endio) when mkfs.xfs is executed against the thin device:

md/raid0:md99: make_request bug: can't convert block across chunks or bigger than 1152k 6688 127
device-mapper: thin: bio sector=2080 err=-5 bi_size=130560 bi_rw=17 bi_vcnt=32 bi_idx=0

This extra DM debugging shows that the failing bio is spanning across the first and second logical 1152K chunk (sector 2080 + 255 takes the bio beyond the first chunk's boundary of sector 2304). So the bio splitting that DM is doing clearly isn't respecting the MD limits.

max_hw_sectors_kb is 127 for both the thin-pool and thin device (queue_max_hw_sectors returns 255 so we'll excuse sysfs's lack of precision). So this explains why bi_size is 130560.

But the thin device's max_hw_sectors_kb should be 4 (PAGE_SIZE) given that it doesn't have a .merge function (for bio_add_page to consult indirectly via dm_merge_bvec) yet the thin-pool does sit above an MD device that has a compulsory merge_bvec_fn. This scenario is exactly why DM must resort to sending single PAGE_SIZE bios to the underlying layer -- some additional context for this is available in the header for commit 8cbeb67a.

Look story short, the reason a thin device doesn't properly get configured to have a max_hw_sectors_kb of 4 (PAGE_SIZE) is that thin_io_hints() is blindly copying the queue limits from the thin-pool device directly to the thin device's queue limits.

Fix this by eliminating thin_io_hints. Doing so is safe because the block layer's queue limits stacking already enables the upper level thin device to inherit the thin-pool device's discard and minimum_io_size and optimal_io_size limits that get set in pool_io_hints. But avoiding the queue limits copy allows the thin and thin-pool limits to be different where it is important, namely max_hw_sectors_kb.

---
 drivers/md/dm-thin.c |   11 -----------
 1 files changed, 0 insertions(+), 11 deletions(-)

index 074e570..9bd59ae 100644
@@ -2819,16 +2819,6 @@ static int thin_iterate_devices(struct dm_target *ti,
 	return 0;
 }
 
-/*
- * A thin device always inherits its queue limits from its pool.
- */
-static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
-{
-	struct thin_c *tc = ti->private;
-
-	*limits = bdev_get_queue(tc->pool_dev->bdev)->limits;
-}
-
 static struct target_type thin_target = {
 	.name = "thin",
 	.version = {1, 6, 0},

@@ -2840,7 +2830,6 @@ static struct target_type thin_target = {
 	.postsuspend = thin_postsuspend,
 	.status = thin_status,
 	.iterate_devices = thin_iterate_devices,
-	.io_hints = thin_io_hints,
 };
 
 /*----------------------------------------------------------------*/