[LU-591] 1.8->2.0 Lustre filesystem upgrade tool

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.1.0
• Fix Version/s: Lustre 2.4.0
• Type: New Feature
• Priority: Minor
• Reporter: Vitaly Fertman
• Assignee: nasf (Inactive)
• Resolution: Fixed

Issue Links

• is related to: LU-1866 LFSCK Phase 1.5 for FID-in-dirent and...

Comments

[Vitaly Fertman]

remote: http://review.whamcloud.com/1230
remote: http://review.whamcloud.com/1231
remote: http://review.whamcloud.com/1232

the code is not 100% ready, it is still under testing.

[Andreas Dilger]

I'm generally not in favour of the last patch (http://review.whamcloud.com/1232, automatically enabling the INCOMPAT_DIR_DATA feature at first mount) since this can cause interoperability issues, and may cause problems for downgrade to 1.8 if there are problems in that code. In upstream ext4 none of the new features are enabled automatically by the kernel, but rather depend on explicit user action (tune2fs). In this case, it makes sense that the enabling of this feature should be done either by tune2fs or some other explicit action that runs the filesystem upgrade tool.

As for the "compatible with 1 version", we are working to maintain protocol compatibility with 1.8.x clients and upgrade from 1.8.x -> 2.1, because almost nobody is running 2.0.0. I checked and it appears that there is "data_in_dirent" support going back to 1.8.2 (https://bugzilla.lustre.org/show_bug.cgi?id=21188 was also fixed in 1.8.2 after the initial patch landing), so this should be OK but I'd like feedback from Johann and Oleg as well.

[Vitaly Fertman]

HLD

[Vitaly Fertman]

code is rebased against b2_1 and moved to:
http://review.whamcloud.com/#change,1561
http://review.whamcloud.com/#change,1562
http://review.whamcloud.com/#change,1563

[Alex Zhuravlev]

1230 seem to be introduced a bug:

	rc = mdo_xattr_get(env, mdd_obj, ldata->ml_buf,
			   XATTR_NAME_LINK, capa);
	if (rc == -ERANGE) {
		/* Buf was too small, figure out what we need. */
		mdd_buf_put(ldata->ml_buf);
		rc = mdo_xattr_get(env, mdd_obj, ldata->ml_buf,
				   XATTR_NAME_LINK, capa);

in the latter mdo_xattr_get(), &LU_BUF_NULL should be used instead of ldata->ml_buf.
Currently it will return same -ERANGE, not actual size of the EA.

[Patrick Farrell (Inactive)]

In light of the changes to lfsck in phase 1.5(https://jira.hpdd.intel.com/browse/LU-1866), is this tool still needed?
If so, what does it do that the new lfsck does not?

Thanks.

[Andreas Dilger]

Patrick,
this code and the LFSCK 1.5 code are largely overlapping. The main differences are:

• this code changed all of the inode FIDs, which would cause problems with NFS exports and some backup tools
• LFSCK 1.5 keeps the 1.8 IGIF FIDs and treats them as normal 2.x FIDs, even after backup/restore
• this code must be run without mounted clients, due to above
• LFSCK 1.5 can be run while the filesystem is mounted and in use