[LU-12997] getcwd() returns ENOENT on RHEL7

• Type: Bug
• Resolution: Not a Bug
• Priority: Minor
• Affects Version/s: Lustre 2.12.2
• Environment: RHEL7.7 as well as other minor versions of RHEL7 on x86_64.
• Severity: 3
• Assignee: WC Triage
• Reporter: K. Scott Rowe

Description

We have been seeing getcwd() return ENOENT on directories that are, in
fact, always there. We can reliably reproduce this problem with the
attached test-getcwd.c code on Lustre Server 2.12.2 and Lustre Client
2.12.3 on RHEL7.7 as well as many other Lustre version and RHEL7
version combinations.

We see reports in LU-9735 about RHEL7 clients getting an ENOENT return
from getcwd(), but I don't understand if a solution is in the works or
not. We are also not sure if this is a Lustre problem, an RHEL kernel
problem, or both.

The LD_PRELOAD workaround from LU-9735 is working for us, but I am
wondering if there is a proper solution pending. Is there anything we
can do to help?

Issue Links

• is related to: LU-9868 dcache/namei fixes for lustre

Comments

[James A Simmons]

Sigh. RedHat claimed this was fixed. Its going to take some push to get them to resolve this. I don't have the power to resolve this. Some one with greater influence with RedHat will have to discuss a fix.

[K. Scott Rowe]

Perhaps I wasn't clear enough.

This is still a problem.

I can still reproduce this error with RHEL7.8 using the test-getcwd.c program above.

 

[James A Simmons]

RHEL7.8 contains a fix so this can be closed. If people encounter this issue please move to RHEL7.8

[K. Scott Rowe]

W installed a machine with RHEL-7.8 using kernel 3.10.0-1127.el7.x86_64 and while I see a bug fix for a getcwd problem

 

 

$ rpm -qi --changelog kernel|grep getcwd
- [fs] vfs: close race between getcwd() and d_move() (Miklos Szeredi) [1631631]

 

 

1631631 is a different bug id than 1811124 that James A Simmons reported.  And, I can still reproduce the problem on our Lustre-2.10.8 filesystem using the 2.12.4 client on RHEL-7.8.

 

$ ./test-getcwd /lustre/aoc/sciops/krowe/tmp
test-getcwd: test-getcwd.c:44: main: Assertion `rc == 0' failed.
Aborted

 

 

[James A Simmons]

I was told by RedHat that a fix was landed to RHEL7.8

[K. Scott Rowe]

Thanks for continuing to work on this.  I have some good news.  We upgraded our servers from Lustre-2.5.5 to Lustre-2.10.8 on Feb. 20, 2020 and the problem has been greatly reduced.

The test program (test-getcwd.c) still fails when run on our Lustre filesystem but the number of times a user has run into this getcwd() problem with other code since we upgraded our Lustre servers has dropped to almost zero.

 

[James A Simmons]

I submitted a bugzilla:

https://bugzilla.redhat.com/show_bug.cgi?id=1811124

[Neil Brown]

This sounds like the bug fixed upstream by

Commit 61647823aa92 ("VFS: close race between getcwd() and d_move()")

Fixed in v4.16

Probably the best approach for lustre supporting older kernels is to copy d_drop() from a newer kernel into libcfs, and use that instead of the exported d_drop.

 

[James A Simmons]

This is a race in ll_splice_alias() due to the use of d_move() when the inode is for a directories. Their are fixes for this but it corrects the way Lustre handles it dcache which causes other types of breakage. I tried some ideas to fix this but its a work in progress. Basically the bug is $MOUNT/.lustre/fid/$fid_for_mount means that dentry $fid_for_mount == dentry $MOUNT which causes a circular loop that crashes the node. I might be able to handle this special case using d_real() that is in newer kernels.