[LU-16806] mkdir: cannot create directory ‘test5’: Object is remote

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Affects Version/s: Lustre 2.14.0, Lustre 2.15.2
• Labels: llnl
• Environment: server (garter): ; toss 4.5-6rc6 ; 4.18.0-425.19.2.1toss.t4.x86_64 ; 2.14.0_19.llnl ; ; client (mutt): ; toss 4.5-6 ; 4.18.0-425.19.2.1toss.t4.x86_64 ; 2.15.2_5.llnl
• Severity: 3
• Assignee: Lai Siyao
• Reporter: Gian-Carlo Defazio

Description

Attempts to create directories are intermittently failing when using "mkdir".

So far, "lfs mkdir" succeeds every time, so does file creation using "touch".

 

Comments

[Peter Jones]

Is there any further work outstanding here? AFAIK the b2_14 fix has been proven to work and this issue does not exist on more current releases, so I would think that we can close out the ticket...

[Lai Siyao]

Olaf, yes, it works with both 2.12 and 2.15 clients.

[Lai Siyao]

Gian, these code is to check whether client sends mkdir request to wrong MDT. If parent has default LMV (2.14 code doesn't check this), and this default LMV is not space balanced, it's not allowed to mkdir on remote MDT.

				if (hint->dah_parent &&
				    dt_object_remote(hint->dah_parent) && lds &&
				    lds->lds_dir_def_striping_set &&
				    lds->lds_dir_def_stripe_offset !=
				    LMV_OFFSET_DEFAULT)
					GOTO(out, rc = -EREMOTE);

As for your question, yes, if parent doesn't have default LMV, the filesystem default LMV will be applied.

[Olaf Faaland]

Hi Lai,

Please also confirm whether you believe 2.14 servers with this patch will work properly with both Lustre 2.12 and 2.15 clients.  As we work to switch to Lustre 2.15, we will have

Clients: Lustre 2.12, 2.15
Routers: Lustre 2.12, 2.15

thanks

[Gian-Carlo Defazio]

The patch fixed the issue on 2.14.

I'm a bit confused by what the patch is doing.

Is the idea that when (lds->lds_dir_def_striping_set == 0) that means that the parent directory doesn't have default striping set for its sub-directories, so when the sub-directory is created it should instead inherit from the root of the whole filesystem?

 

[Lai Siyao]

Yes, upgrading to 2.15.x can get around this issue since related code has been removed.

[Peter Jones]

So LLNL's exposure to this issue will disappear once they upgrade to 2.15.x?

[Lai Siyao]

This is caused by a bug in 2.14 code, but it's not an issue in the following release because related code has been removed.

[Gerrit Updater]

"Lai Siyao <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/51045
Subject: LU-16806 lod: access DMV when it's valid
Project: fs/lustre-release
Branch: b2_14
Current Patch Set: 1
Commit: 65887f6728d68d8b2e10969df25e5303603a4797

[Gian-Carlo Defazio]

I've added mkdir-attempt-client-and-mds.tar.gz which has debug logs for the client and mds's with debug=-1.

[Lai Siyao]

Could you collect MDS debug logs upon failure? I'm afraid some patch for backward compatibility is missing, maybe it's on server side.