[LU-3181] open by FID for write with O_LOV_DELAY_CREATE fails for files on MDT1

Related — Wed, 11 Feb 2015 16:58:18 +0000

mdt_cross_open() uses a slightly different test than is used elsewhere to detect nefarious attempts at creating striping by opening files for write using their FIDs.

For files on MDT0 this is handled "correctly":

# MDSCOUNT=2 llmount.sh
# cd /mnt/lustre
# multiop f0 m
# lfs path2fid f0
[0x280000400:0x1:0x0]
# multiop .lustre/fid/[0x280000400:0x1:0x0] oO_WRONLY:O_LOV_DELAY_CREATE
# lfs setstripe -c2 .lustre/fid/[0x280000400:0x1:0x0]
# lfs getstripe .lustre/fid/[0x280000400:0x1:0x0]

Now for a file on MDT1:

# lfs mkdir -i1 d1
# multiop d1/f1 m
# lfs path2fid d1/f1
[0x2c0000400:0x2:0x0]
# multiop .lustre/fid/[0x2c0000400:0x2:0x0] oO_WRONLY:O_LOV_DELAY_CREATE
open: Operation not permitted
# lfs setstripe -c2 .lustre/fid/[0x2c0000400:0x2:0x0]
unable to open '.lustre/fid/[0x2c0000400:0x2:0x0]': Operation not permitted (1)
error: setstripe: create stripe file '.lustre/fid/[0x2c0000400:0x2:0x0]' failed
# lfs getstripe .lustre/fid/[0x2c0000400:0x2:0x0]
.lustre/fid/[0x2c0000400:0x2:0x0] has no stripe info

I think that all these checks for fid_is_obf() are somewhat silly and that users of .lustre/fid should just be expected to use O_LOV_DELAY_CREATE.

Whamcloud Community JIRA

[LU-3181] open by FID for write with O_LOV_DELAY_CREATE fails for files on MDT1