Bug #36171
openmds: ctime should not use client provided ctime/mtime
0%
Description
Otherwise, you can set a ctime that is far in the future and it cannot be rolled back.
pdonnell@senta02 ~/ceph$ cd /home/pdonnell/mnt/tmp.9cYgpRTGXD pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ mkdir foo pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ touch foo/bar pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ getfattr -n ceph.dir.rctime foo # file: foo ceph.dir.rctime="1537838662.09897717617" pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ stat foo/bar File: 'foo/bar' Size: 0 Blocks: 0 IO Block: 4194304 regular empty file Device: 2ch/44d Inode: 1099511627777 Links: 1 Access: (0640/-rw-r-----) Uid: ( 1163/pdonnell) Gid: ( 1163/pdonnell) Access: 2018-09-24 21:24:22.893455885 -0400 Modify: 2018-09-24 21:24:22.893455885 -0400 Change: 2018-09-24 21:24:22.897717617 -0400 Birth: - pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ touch foo/derp pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ date +%s 1537838726 pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ getfattr -n ceph.dir.rctime foo # file: foo ceph.dir.rctime="1537838725.0981013223" pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ touch -m --date='Jan 01 2022' foo/derp pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ getfattr -n ceph.dir.rctime foo # file: foo ceph.dir.rctime="1641013200.090" pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ rm foo/derp pdonnell@senta02 ~/mnt/tmp.9cYgpRTGXD$ getfattr -n ceph.dir.rctime foo # file: foo ceph.dir.rctime="1641013200.090"
Partially caused by 60f73cde4e04be2685eaf7cd04e03c045e1f9977 and #35945.
I think the right approach here is to have the client set its own ctime so that it's locally useful but not to transmit it to the MDS. The MDS always uses its own clock to update the ctime/rctime. This means from the client's perspective that the ctime changes twice for a single metadata operation but I think that's acceptable. The standards provide no guarantees or guidance on this.
Updated by Patrick Donnelly over 5 years ago
- Subject changed from mds: ctime should not use mtime value to mds: ctime should not client provided ctime/mtime
Updated by Patrick Donnelly over 5 years ago
- Subject changed from mds: ctime should not client provided ctime/mtime to mds: ctime should not use client provided ctime/mtime
- Status changed from New to In Progress
- Assignee set to Patrick Donnelly
Updated by Patrick Donnelly over 5 years ago
- Status changed from In Progress to Fix Under Review
Updated by Patrick Donnelly about 5 years ago
- Target version changed from v14.0.0 to v15.0.0
Updated by Patrick Donnelly about 5 years ago
- Status changed from Fix Under Review to In Progress
- Pull request ID set to 24277
Updated by Malcolm Haak about 4 years ago
I think the right approach here is to have the client set its own ctime so that it's locally useful but not to transmit it to the MDS. The MDS always uses its own clock to update the ctime/rctime. This means from the client's perspective that the ctime changes twice for a single metadata operation but I think that's acceptable. The standards provide no guarantees or guidance on this.
I totally disagree for the mtime case. This breaks rsync and lots of other applications expected behaviour. At least for mtime anyway. Ctime this is the sane behaviour.
mtime should be arbitrary. I've had issues with applications freak out when its in the future but it is technically 'valid'. But I've also encountered workloads where it gets set into the future so it gets left alone until after that date. (I work in HPC and scientists are odd sometimes). But the real issue is if I can't update mtime, rsync can't cheat and use mtimes to detect file changes. That makes many backups sad.
Anyway long story short, mtime should ALWAYS be whatever weird thing the client sets it to. Ctime is less important.
Thanks.
Updated by Greg Farnum about 4 years ago
Yeah, we previously assigned mtimes from the MDS and it was a disaster. That's not something we'll be going back to!
I'm less sure about ctimes but I don't think applications outside of samba tend to care about those much...
Updated by Zheng Yan about 4 years ago
see https://github.com/ceph/ceph/pull/32126. that commit add 'dirty_from' field to rstat, which is mds time. mds can use 'dirty_from' to decide if rctime should be set to a smaller value.
Updated by Dan van der Ster over 3 years ago
IMHO ctime should always be `ceph_clock_now()` rather than any time from the client.
Here's an XFS demo. Note how Change time is the clock when a change is made, not related to the arbitrary Modify time of the file:
# touch asdf # stat asdf File: ‘asdf’ Size: 0 Blocks: 0 IO Block: 4096 regular empty file Device: fc01h/64513d Inode: 25166519 Links: 1 Access: (0644/-rw-r--r--) Uid: ( 0/ root) Gid: ( 0/ root) Context: unconfined_u:object_r:user_tmp_t:s0 Access: 2020-10-15 14:33:51.874524960 +0200 Modify: 2020-10-15 14:33:51.874524960 +0200 Change: 2020-10-15 14:33:51.874524960 +0200 Birth: - # touch -t 202501010000 asdf # stat asdf File: ‘asdf’ Size: 0 Blocks: 0 IO Block: 4096 regular empty file Device: fc01h/64513d Inode: 25166519 Links: 1 Access: (0644/-rw-r--r--) Uid: ( 0/ root) Gid: ( 0/ root) Context: unconfined_u:object_r:user_tmp_t:s0 Access: 2025-01-01 00:00:00.000000000 +0100 Modify: 2025-01-01 00:00:00.000000000 +0100 Change: 2020-10-15 14:34:00.520475964 +0200 Birth: -
Updated by Patrick Donnelly 6 months ago
- Status changed from In Progress to New
- Assignee deleted (
Patrick Donnelly) - Backport deleted (
mimic,luminous) - Labels (FS) task(intern), task(medium) added
- Labels (FS) deleted (
task(easy))