Was examining why xfstest 614 failed (smb3.1.1 mount to Windows), and
noticed a couple of problems:
1) we don't requery the allocation size (number of blocks) in all
cases we should. This small fix should address that
--- a/fs/cifs/inode.c
+++ b/fs/cifs/inode.c
@@ -2390,15 +2390,16 @@ int cifs_getattr(struct user_namespace
*mnt_userns, const struct path *path,
struct cifs_tcon *tcon = cifs_sb_master_tcon(cifs_sb);
struct inode *inode = d_inode(dentry);
int rc;
/*
* We need to be sure that all dirty pages are written and the server
* has actual ctime, mtime and file length.
*/
- if ((request_mask & (STATX_CTIME | STATX_MTIME | STATX_SIZE)) &&
+ if ((request_mask & (STATX_CTIME | STATX_MTIME | STATX_SIZE |
STATX_BLOCKS)) &&
and 2) we don't set the allocation size on the case where a cached
file on this client is written, and to make it worse if we queried
(post query attributes flag) at SMB3 close for compounded operations -
the Windows server (not sure about others) apparently doesn't update
the allocation size until the next open/queryinfo so we still end up
with an allocation size of 0 for a 64K file which breaks the test.
What the test is doing is quite simple:
xfs_io -f -c "truncate 64K" -c "mmap -w 0 64K" -c "mwrite -S 0xab 0
64K" -c "munmap" foo1 ; stat -c %b foo1
And it fails - due to seeing a number of blocks 0 rather than the
correct value (128). With actimeo=0 we do a subsequent open/query
operation which would cause it to pass since the second open/query
does show the correct allocation size.
Any ideas?
--
Thanks,
Steve
