[LU-8846] lfs_migrate() should try to unlink volatile file

• Status: Resolved
• Project: Lustre
• Fix Version/s: Lustre 2.10.0
• Type: Bug
• Priority: Minor
• Reporter: John Hammond
• Assignee: John Hammond
• Resolution: Fixed
• Labels: lfs
• Severity: 3

Description

lfs_migrate() bypasses llapi_create_volatile_idx() and creates a volatile file directly. (It does this because it needs to create the volatile file with a specific striping and llapi_create_volatile_idx() does not allow us to pass in striping.) It should do as llapi_create_volatile_idx() does and try to unlink the volatile file after opening it (in case volatile file creation is not supported).

Or we could add O_LOV_DELAY_CREATE to the open flags and use llapi_create_volatile_idx() but then we would need a function to apply the striping described by a struct llapi_stripe_param to an open file handle.

Issue Links

• is related to: LU-8559 sanity test_56xa: cannot swap layouts...

Comments

[Gerrit Updater]

John L. Hammond ([e-mail]) uploaded a new patch: http://review.whamcloud.com/23842
Subject: LU-8846 lfs: try to unlink volatile file in lfs_migrate()
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 222f9b93d7675a53d0b3d5d5e047ffb9f3f50dca

[Robert Read (Inactive)]

I'm curious, what is the situation where volatile files are not supported and we're doing lfs migrate?

[Andreas Dilger]

Robert, I think John is saying that it isn't possible to specify the file layout when creating a volatile file, so that it can't be used when restriping a file during migration.

I'm thinking it makes sense to create a new API to allow specifying the layout for a volatile file. Probably passing O_LOV_DELAY_CREATE and setting the layout explicitly (within the same llapi_create_volatile_layout() function) would be most straight forward.

I don't think the current patch is wrong, however, since it shouldn't leave files around if the volatile file create API isn't available.

[Robert Read (Inactive)]

The existing API already allows passing O_LOV_DELAY_CREATE), and the copytool uses this so it can set striping on the volatile file.

But my question was based on the patch comment: "In lfs_migrate() try to unlink the volatile file in case the MDT does not support volatile file creation."

[John Hammond]

It's also totally conceivable that upstream reviewers will notice volatile file handling and decide that it's a security issue or something.

[Robert Read (Inactive)]

Ah, that's a good point.

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/23842/
Subject: LU-8846 lfs: try to unlink volatile file in lfs_migrate()
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: d15fbc4fd9fbf5b892ea4320771ad40963bfdc69

[Peter Jones]

Landed for 2.10