[LU-2657] Shouldn't deleting objects in mds_lov_update_objids()

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 1.8.8
• Fix Version/s: Lustre 2.1.5, Lustre 1.8.9
• Type: Bug
• Priority: Minor
• Reporter: Niu Yawei (Inactive)
• Assignee: Niu Yawei (Inactive)
• Resolution: Fixed
• Severity: 3

Description

There is a code snippet from The code in mds_lov_update_objids():

                if (id > data[idx]) {
#ifndef HAVE_DELAYED_RECOVERY
                        int lost = id - data[idx] - 1;
                        /* we might have lost precreated objects due to VBR */
                        if (lost > 0 && obd->obd_recovering) {
                                CDEBUG(D_HA, "GAP in objids is %u\n", lost);
                                if (!obd->obd_version_recov)
                                        CWARN("Unexpected gap in objids\n");
                                /* lsm is allocated if NULL */
                                mds_log_lost_precreated(obd, &lsm, &stripes,
                                                        data[idx] + 1, lost, i);
                        }
#endif
                        data[idx] = id;
                        cfs_bitmap_set(mds->mds_lov_page_dirty, page);
                }
        }

Which is not quite right, because we can't guarantee the object id is always increased in transno order, and if the lov_objid file was removed by administrator (to fix the corrupted lov_objid file), many objects could be deleted mistakenly then.

Comments

[Niu Yawei (Inactive)]

http://review.whamcloud.com/5137

[Andreas Dilger]

Niu, is this still a problem for 2.1 or 2.4?

[Niu Yawei (Inactive)]

Yes, it's still a problem for 2.1, but looks it has been fixed in ORI-692 for 2.4.

[Niu Yawei (Inactive)]

patch for b2_1: http://review.whamcloud.com/5206

[Niu Yawei (Inactive)]

patches landed.