[LU-11485] MDS allows "lfs setstripe" to mark last mirror as stale

Related — Fri, 26 Jun 2020 18:34:06 +0000

A trivial modification to sanity-flr test_0h shows a serious problem.

If we write to the file before changing the layout flags, the file metadata is corrupted:

 lfs mirror create -N -E 1M -S 1M --flags=prefer -E eof -N2 tfile
 echo 1 > tfile
#set flags to the first component
 lfs setstripe --comp-set -I 0x10001 --comp-flags=^prefer,stale tfile
ls -la 
 ls: cannot access tfile: Input/output error
 total 16
 drwxr-xr-x 3 root root 4096 Oct 9 13:00 .
 drwxr-xr-x. 43 root root 122880 Sep 27 22:49 ..
 -?????????? ? ? ? ? ? tfile

So this corrupts the layout. I tried adding https://review.whamcloud.com/#/c/32847/ (~~LU-11158~~ mdt: grow lvb buffer to hold layout) and it did not resolve this issue.

But additionally, I'm not sure what the correct behavior is here - should it be possible to remove the prefer flag in this situation, while the other replicas are stale?

And to set stale on the write replica? (That, at least, seems clearly incorrect.)

The effect of this (... if it worked ...) would presumably be that all the replicas are stale, which would be unresolvable, and so should not be allowed. (Separate from the corruption)

There are various implications here.

Whamcloud Community JIRA

[LU-11485] MDS allows "lfs setstripe" to mark last mirror as stale