[LU-18758] 'lfs migrate --block' does not block modifications

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.17.0
• Affects Version/s: Lustre 2.14.0, Lustre 2.16.1
• Severity: 3
• Assignee: Feng Lei
• Reporter: Andreas Dilger

Description

The "lfs migrate --block FILE" command appears to be broken, since it does not block other processes from modifying a file during migration, and instead returns an error:

$ dd if=/dev/urandom of=/myth/tmp/random bs=1 oflag=sync &
$ lfs migrate -c2 --block /myth/tmp/random
error: lfs migrate: /myth/tmp/random: file changed: Resource temporarily unavailable
$ lfs migrate -c2 --block /myth/tmp/random
error: lfs migrate: /myth/tmp/random: file changed: Resource temporarily unavailable

This is likely due to the use of a lease to be notified of the file modification during migration. Previously, "lfs migrate --block" was using a group lock to block other clients from accessing the file during migration.

The "lfs migrate --block" command is important for OST evacuation, because it is the "last resort" mechanism to move files off an OST before it is removed from service. This needs to complete even if there is an application that is writing to the file, though the application writing to the file will block until the migration is complete.

Issue Links

• is related to: LU-16571 "lfs migrate -b" can cause thread starvation on OSS
• is related to: LU-14084 change 'lfs migrate' to use 'MIGRATION_NONBLOCK' by default

Comments

[Peter Jones]

Merged for 2.17

[Gerrit Updater]

"Oleg Drokin <[e-mail]>" merged in patch https://review.whamcloud.com/c/fs/lustre-release/+/58299/
Subject: LU-18758 tests: test lfs migrate --block on changing file
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 4794fcd61a25fe32f0c6f162049983961d01307a

[Gerrit Updater]

"Feng Lei <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58299
Subject: LU-18758 tests: test lfs migrate --block on changing file
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 0950d59a3e40a20ee47c49f9711112572df4a347

[Feng Lei]

The test passes too with oflag=sync flag:

# dd if=/dev/zero of=random bs=1 oflag=sync &
[1] 2692411
# lfs migrate -o 0 --block random
# lfs migrate -c 2 --block random
# kill 2692411
# 
[1]+  Terminated              dd if=/dev/zero of=random bs=1 oflag=sync 

Please correct me if there is any mistake.

[Feng Lei]

Looks like it has been fixed in latest version.

# lctl version 
lt-lctl 2.16.52_30_g8cd8f3f
# dd if=/dev/zero of=random bs=1 &
[1] 2692265
# lfs migrate -o 0 --block random
# kill 2692265
# 
[1]+  Terminated              dd if=/dev/zero of=random bs=1
# lfs getstripe random
random
lmm_stripe_count:  1
lmm_stripe_size:   4194304
lmm_pattern:       raid0
lmm_layout_gen:    1
lmm_stripe_offset: 0
	obdidx		 objid		 objid		 group
	     0	          1869	        0x74d	   0x240000400

# ls -l random
-rw-r--r-- 1 root root 80966 Mar  4 14:09 random

[Andreas Dilger]

It would also be useful to add an alias "--nonblock" in addition to "--non-block" for ease of use.