[LU-3660] Can't disable ACL support with ZFS MDT Created: 29/Jul/13 Updated: 19/Sep/14 Resolved: 26/Aug/14 |
|
| Status: | Resolved |
| Project: | Lustre |
| Component/s: | None |
| Affects Version/s: | Lustre 2.4.0 |
| Fix Version/s: | Lustre 2.7.0, Lustre 2.5.4 |
| Type: | Bug | Priority: | Critical |
| Reporter: | Ned Bass | Assignee: | Lai Siyao |
| Resolution: | Fixed | Votes: | 0 |
| Labels: | llnl, prz, zfs | ||
| Issue Links: |
|
||||||||
| Severity: | 3 | ||||||||
| Rank (Obsolete): | 9444 | ||||||||
| Description |
|
With ldiskfs, the MDT can be mounted with -o noacl to disable POSIX ACL support. With ZFS, that option seems to have no effect, and clients are still able to set and honor ACLs. It would be desirable to be able to turn off ACLs, as they can be detrimental to performance. For example, the cp -a command tries to write the xattr system.posix_acl_access on every destination file, regardless of whether the source file has that attribute. The ACL update is handled synchronously on the MDT, so the request handler has to wait for a ZFS transaction group to sync. This can introduce significant latency on a busy MDS, effectively limiting the per-file copy rate to the txg sync rate. On ldiskfs, cp -a performance is significantly improved by disabling ACLs. (This performance problem is a separate issue that may be helped by integrating ZIL support in Lustre. Turning off ACLs is a temporary workaround.) LLNL-bug-ID: TOSS-2207 |
| Comments |
| Comment by Andreas Dilger [ 29/Jul/13 ] |
|
Is there a reason that ACL updates are done synchronously on the MDS? I don't think there is a valid security rationale for this, since (AFAIK) chmod(), chown(), or chgrp() operations are not synchronous either, and they can have the same security impact as any ACL change. It would be far more beneficial to disable synchronous updates on the MDS entirely for this operation than to make them somewhat more efficient with the ZIL. |
| Comment by Ned Bass [ 30/Jul/13 ] |
|
I don't know why they're done synchronously. In fact my claim that it is synchronous is based on the observation that the request handler seems to block until the txg syncs, not on a close reading of the code. It think it's related to this comment, but it doesn't explain why. mdt_xattr.c:mdt_reint_setxattr() 334 /* Revoke all clients' lookup lock, since the access 335 * permissions for this inode is changed when ACL_ACCESS is 336 * set. This isn't needed for ACL_DEFAULT, since that does 337 * not change the access permissions of this inode, nor any 338 * other existing inodes. It is setting the ACLs inherited 339 * by new directories/files at create time. */ 340 /* We need revoke both LOOKUP|PERM lock here, see mdt_attr_set. */ 341 if (!strcmp(xattr_name, XATTR_NAME_ACL_ACCESS)) 342 lockpart |= MDS_INODELOCK_PERM | MDS_INODELOCK_LOOKUP; |
| Comment by Peter Jones [ 30/Jul/13 ] |
|
Lai Could you please advise on this one? Thanks Peter |
| Comment by Ned Bass [ 30/Jul/13 ] |
|
Created |
| Comment by Ned Bass [ 30/Jul/13 ] |
|
Lowered priority to minor because disabling ACL support is not really an effective workaround to the performance issue described here. But, we should still decide if this should be fixed and document ACL functionality as it relates to ZFS backends. |
| Comment by Lai Siyao [ 12/Aug/13 ] |
|
osd_conf_get() for zfs osd doesn't really get mount options from super_block like ldiskfs, but enables ACL by default. I'll see how to get it work. |
| Comment by Emoly Liu [ 31/Jul/14 ] |
|
Lai's patch for this problem is at http://review.whamcloud.com/#/c/10850/ . |
| Comment by Peter Jones [ 26/Aug/14 ] |
|
Landed for 2.7 |