BUG: Files corrupt after moving LVM volume to USB disk

ho hum I'm getting rather bored of redirecting emails to the correct lists.

Does anyone there read linux-kernel?

Begin forwarded message:

Date: Tue, 20 Mar 2007 12:34:08 +0200
From: "Marti Raudsepp" <[email protected]>
To: "Kernel hackers" <[email protected]>, fsdevel <[email protected]>
Subject: BUG: Files corrupt after moving LVM volume to USB disk

Hello LKML,

Second repost of "BUG: Killing and reviving files with USB disks", this time also destined to linux-fsdevel.

This is a reproducible demonstration of the problem initially reported in my first e-mail, titled "PROBLEM: 'bio too big device' after moving to a USB disk" (http://lkml.org/lkml/2007/3/7/657)

Summary:
01. Create LVM volume; initialize dm-crypt; initialize reiserfs; mount
02. Populate the disk with files; sync; flush caches
03. Confirm that the files are still readable and non-corrupt (sha1sum)
04. Migrate logical volume to USB disk; sync; flush caches
05. Confirm that ALL PREVIOUSLY VERIFIED FILES ARE CORRUPT
06. Observe "bio too big device dm-0 (256 > 240)" messages in dmesg
07. Run reiserfsck to check for corruptions -- none.
08. Migrate logical volume back to SATA disk; sync; flush caches
09. Confirm that files are readable and non-corrupt again
10. Clean up

Environment/configuration:
* Linux non 2.6.19-gentoo-r6 #1 Wed Feb 28 20:55:24 EET 2007 x86_64 AMD Athlon(tm) 64 Processor 3000+ AuthenticAMD GNU/Linux
* USB disk 120GB Western Digital, model 00UE-00KVT0 (according to udev), serial DEF10CD7F64C
* Older SATA disk 200GB Seagate 7200.7, model ST3200822AS
* Motherboard Asus A8N5X, nForce4 chipset
* Offending file system: reiserfs v3.6, mounted with noatime,barrier=flush
* dm-crypt using aes-256 with cbc-essiv:sha256; using assembly-optimized AES on x86_64 (CONFIG_CRYPTO_AES_X86_64)
* Test partition is 68 extents times 16MiB = 1088 MiB large (that's the largest I could allocate while remaining in one segment -- otherwise lvmove wouldn't move the partition back to /dev/sda5 after "defragmenting" it.)
* LVM utilities version: 2.02.17 (2006-12-14)
* LVM library version: 1.02.12 (2006-10-13)
* LVM driver version: 4.10.0
* cryptsetup-luks 1.0.4 (user space interface to dm-crypt)

Involved block devices:
* /dev/sda: My old SATA disk.
* /dev/sda5: The LVM partition on the old disk.
* /dev/sdb: The new offending USB disk; whole disk is used as an LVM physical volume.
* /dev/primary: LVM volume group 'primary', consisting of /dev/sdb and

                /dev/sda5

* /dev/primary/punchbag: LVM volume 'punchbag' for demonstration purposes.
* /dev/mapper/crypt-punchbag: The dm-crypt "decrypted" loopback device.

----
00. PV information

[non]# pvdisplay /dev/sda5
--- Physical volume ---

  PV Name               /dev/sda5
  VG Name               primary
  PV Size               145.83 GB / not usable 2.73 MB
  Allocatable           yes
  PE Size (KByte)       16384
  Total PE              9333
  Free PE               117
  Allocated PE          9216
  PV UUID               YdrDuL-jw5z-J0SA-EEKU-NOC4-6gGR-T90YCA

[non]# pvdisplay /dev/sdb
--- Physical volume ---

  PV Name               /dev/sdb
  VG Name               primary
  PV Size               111.79 GB / not usable 9.46 MB
  Allocatable           yes
  PE Size (KByte)       16384
  Total PE              7154
  Free PE               7129
  Allocated PE          25
  PV UUID               nE8C5L-lfI1-VNgs-Q7ei-1Zz3-GeGT-UNhH4p

----
01. Create LVM volume; initialize dm-crypt; initialize reiserfs; mount

[non]# lvcreate -n punchbag --extents 68 primary /dev/sda5
Logical volume "punchbag" created
[non]# lvs --segments -o +devices
LV VG Attr #Str Type SSize Devices
[...]
punchbag primary -wi-a- 1 linear 1.06G /dev/sda5(0)
[...]
[non]# cryptsetup luksFormat /dev/primary/punchbag -c aes-cbc-essiv:sha256 -h sha1
[...]
Are you sure? (Type uppercase yes): YES
Enter LUKS passphrase:
Verify passphrase:
Command successful.
[non]# cryptsetup luksOpen /dev/primary/punchbag crypt-punchbag Enter LUKS passphrase:
key slot 0 unlocked.
Command successful.
[non]# mkfs.reiserfs /dev/mapper/crypt-punchbag
[...]
Guessing about desired format.. Kernel 2.6.19-gentoo-r6 is running. Format 3.6 with standard journal
Count of blocks on the device: 343920
Number of blocks consumed by mkreiserfs formatting process: 8222
Blocksize: 4096
Hash function used to sort names: "r5"
Journal Size 8193 blocks (first block 18)
Journal Max transaction length 1024
inode generation number: 0
UUID: c1857208-5090-49dd-9163-9fb002d96a88
ATTENTION: YOU SHOULD REBOOT AFTER FDISK!
ALL DATA WILL BE LOST ON '/dev/mapper/crypt-punchbag'!
Continue (y/n):y

Initializing journal - 0%....20%....40%....60%....80%....100% Syncing..ok
[...]

ReiserFS is successfully created on /dev/mapper/crypt-punchbag. [non]# mkdir /mnt/punchbag
[non]# mount /dev/mapper/crypt-punchbag /mnt/punchbag -o noatime,barrier=flush

----
02. Populate the disk with files; sync; flush caches

[non]# cp -r ~marti/mp3 /mnt/punchbag
[... yes, these are legal mp3s ;)]
cp: writing `/mnt/punchbag/mp3/...': No space left on device ^C
[non]# sync
[non]# echo 3 > /proc/sys/vm/drop_caches
[non]# cd /mnt/punchbag/mp3/mr\ Epic\ -\ Sideways

----
03. Confirm that the files are still readable and non-corrupt (sha1sum)

[non]# sha1sum -c *.sha1
mr Epic - Sideways - 01. Down Low.flac: OK
mr Epic - Sideways - 02. Blue Days.flac: OK
mr Epic - Sideways - 03. Sift.flac: OK
mr Epic - Sideways - 04. Slipping.flac: OK
mr Epic - Sideways - 05. Ruff and Tumble.flac: OK
mr Epic - Sideways - 06. In.flac: OK
mr Epic - Sideways - 07. Out.flac: OK

----
04. Migrate logical volume to USB disk; sync; flush caches

[non]# pvmove -n punchbag /dev/sda5 /dev/sdb
[...]
/dev/sda5: Moved: 100.0%
[non]# lvs --segments -o +devices
LV VG Attr #Str Type SSize Devices
[...]
punchbag primary -wi-ao 1 linear 1.06G /dev/sdb(25)
[non]# sync
[non]# echo 3 > /proc/sys/vm/drop_caches

----
05. Confirm that all previously verified files are corrupt

[non]# sha1sum -c *.sha1
sha1sum: mr Epic - Sideways - 01. Down Low.flac: Input/output error mr Epic - Sideways - 01. Down Low.flac: FAILED open or read
sha1sum: mr Epic - Sideways - 02. Blue Days.flac: Input/output error mr Epic - Sideways - 02. Blue Days.flac: FAILED open or read
sha1sum: mr Epic - Sideways - 03. Sift.flac: Input/output error mr Epic - Sideways - 03. Sift.flac: FAILED open or read
sha1sum: mr Epic - Sideways - 04. Slipping.flac: Input/output error mr Epic - Sideways - 04. Slipping.flac: FAILED open or read
sha1sum: mr Epic - Sideways - 05. Ruff and Tumble.flac: Input/output error mr Epic - Sideways - 05. Ruff and Tumble.flac: FAILED open or read
sha1sum: mr Epic - Sideways - 06. In.flac: Input/output error mr Epic - Sideways - 06. In.flac: FAILED open or read
sha1sum: mr Epic - Sideways - 07. Out.flac: Input/output error mr Epic - Sideways - 07. Out.flac: FAILED open or read
sha1sum: WARNING: 7 of 7 listed files could not be read

----
06. Observe "bio too big device dm-0 (256 > 240)" messages in dmesg

[non]# dmesg |tail -n7

[101403.106825] bio too big device dm-0 (256 > 240)
[101403.115228] bio too big device dm-0 (256 > 240)
[101403.115912] bio too big device dm-0 (256 > 240)
[101403.130529] bio too big device dm-0 (256 > 240)
[101403.131220] bio too big device dm-0 (256 > 240)
[101403.141577] bio too big device dm-0 (256 > 240)
[101403.142255] bio too big device dm-0 (256 > 240)

----
07. Run reiserfsck to check for corruptions -- none.

[non]# cd
[non]# umount /mnt/punchbag
[non]# reiserfsck --check /dev/mapper/crypt-punchbag
[...]
Will read-only check consistency of the filesystem on /dev/mapper/crypt-punchbag
Will put log info to 'stdout'

Do you want to run this program?[N/Yes] (note need to type Yes if you do):Yes
###########
reiserfsck --check started at Sat Mar 10 19:00:08 2007
###########
Replaying journal..
Reiserfs journal '/dev/mapper/crypt-punchbag' in blocks [18..8211]: 0 transactions replayed
Checking Internal tree...
0%....20%....40%....60%....80%....100%
Checking Semantic tree...
There are on the filesystem:
Leaves 282
Internal nodes 3
Directories 18
Other files 196
Data block pointers 268374 (0 of them are zero)
Safe links 0
No corruptions found
###########
reiserfsck finished at Sat Mar 10 19:00:10 2007
###########

----
08. Migrate logical volume back to SATA disk; sync; flush caches

[non]# mount /dev/mapper/crypt-punchbag /mnt/punchbag -o noatime,barrier=flush [non]# cd /mnt/punchbag/mp3/mr\ Epic\ -\ Sideways
[non]# pvmove -n punchbag /dev/sdb /dev/sda5
[...]
/dev/sdb: Moved: 100.0%
[non]# lvs --segments -o +devices
LV VG Attr #Str Type SSize Devices
[...]
punchbag primary -wi-a- 1 linear 1.06G /dev/sda5(0)
[...]
[non]# sync
[non]# echo 3 > /proc/sys/vm/drop_caches

----
09. Confirm that files are readable and non-corrupt again

[non]# sha1sum -c *.sha1
mr Epic - Sideways - 01. Down Low.flac: OK
mr Epic - Sideways - 02. Blue Days.flac: OK
mr Epic - Sideways - 03. Sift.flac: OK
mr Epic - Sideways - 04. Slipping.flac: OK
mr Epic - Sideways - 05. Ruff and Tumble.flac: OK
mr Epic - Sideways - 06. In.flac: OK
mr Epic - Sideways - 07. Out.flac: OK

----
10. Clean up

[non]# cd
[non]# umount /mnt/punchbag
[non]# cryptsetup remove crypt-punchbag
[non]# lvremove primary/punchbag
Do you really want to remove active logical volume "punchbag"? [y/n]: y
Logical volume "punchbag" successfully removed

I repeated this five times and each time was able to reproduce the problem. More details on the symptoms of the problem in my previous problem report:
http://lkml.org/lkml/2007/3/7/657

I'm not subscribed to the list, so please keep me in Cc:.

Regards,
Marti Raudsepp

I for one generally only spot things with 'lvm' or device-mapper/dm in the subject line. The original subject line included no clue this one had anything to do with device-mapper:-(

A couple of patches to try:
http://www.kernel.org/pub/linux/kernel/people/agk/patches/2.6/2.6.19/dm-io-fix-bi_max_vecs.patch
http://www.kernel.org/pub/linux/kernel/people/agk/patches/2.6/editing/dm-merge-max_hw_sector.patch

and perhaps these three:
http://www.kernel.org/pub/linux/kernel/people/agk/patches/2.6/editing/dm-crypt-fix-call-to-clone_init.patch
http://www.kernel.org/pub/linux/kernel/people/agk/patches/2.6/editing/dm-crypt-fix-avoid-cloned-bio-ref-after-free.patch
http://www.kernel.org/pub/linux/kernel/people/agk/patches/2.6/editing/dm-crypt-fix-remove-first_clone.patch

Alasdair

[email protected]

Summary of what I've managed to rule out so far:
1. this problem does not occur without dm-crypt
2. this problem does not occur when the file system hasn't been moved from the SATA disk to USB
3. Both reiserfs and ext3 are affected by this problem, but ext3 merely slows down.
4. Unmounting and re-mounting has no effect.

My initial report contains more details.

No luck. :(

And thanks for reminding me how annoying reboots are...

Doesn't occur if dm-crypt is not involved (e.g., when placing reiserfs or ext3 straight on top of /dev/primary/punchbag). I don't know if LVM is necessary, as I don't have any unallocated space for "oldschool" partitions left.

Nor do these problems occur when creating either file system directly on the USB disk -- so, only when they have been lvmoved from the SATA disk, or when the volume is physically dd'ed to the USB disk.

Tried with ext3 (LVM+dm-crypt+ext3), and the files didn't appear corrupt, however, I could still see those messages in my dmesg, and disk access was *extremely* slow, causing around 10k context switches per second (!?) reading below 5 MB/s. Normally, I can get read speeds around 12-22 MB/s with this particular USB disk. However, it's not CPU-bound, as over >80% of CPU is still spent in iowait.

When the ext3 file system is created directly on a volume that is located on the USB disk, however, these problems do not occur, just like with reiserfs.

However, when I copied the same files to a reiserfs volume when the LV had already been pvmoved to /dev/sdb, the files were readable (as pointed out in my initial post). Interestingly, it did not introduce the same slowdown as with ext3, reading around 13-14 MB/s with ~3k context switches per second.

To clarify (I double-checked all these scenarios):
(1) files written *before* pvmove are corrupt with reiserfs
(2) files written *after* pvmove read fast with reiserfs
(3) files written *before* pvmove read slowly with ext3
(4) files written *after* pvmove read slowly with ext3
(5) files written to an ext3 volume that was formatted on the USB disk and not pvmoved are fast and don't report these dmesg warnings.

This is purely my speculation, and I can't pretend to know much about the inner workings of device-mapper/file systems, but it appears that ext3 statically derives some attributes during the creation of a file system, while reiserfs gets them runtime. ext3 contains a slow workaround if these attributes don't match (and still posts a warning), but reiserfs returns outright I/O errors.

(P.S. does the LKML archive randomly omit some double-newlines in posts, or is this a problem on my end?)

Marti Raudsepp

Please could you test attached patch ?
Sorry for another annoying reboot :-)

Milan
--
[email protected]

From: Milan Broz <[email protected]>

Disable barriers in dm-crypt because of current workqueue processing can reorder requests.

This must be addresed later but for now disabling barriers is needed to prevent data corruption.

Index: linux-2.6.20.1/drivers/md/dm-crypt.c
===================================================================
@@ -925,6 +925,9 @@ static int crypt_map(struct dm_target *t
 	struct crypt_config *cc = ti->private;
 	struct crypt_io *io;
 
+	if (bio_barrier(bio))
+		return -EOPNOTSUPP;
+
 	io = mempool_alloc(cc->io_pool, GFP_NOIO);
 	io->target = ti;
 	io->base_bio = bio;