[LU-15268] lfs mirror extend error propagation

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.15.0
• Severity: 3
• Assignee: Zhenyu Xu
• Reporter: Andreas Dilger

Description

If a mirror extend operation fails on the MDT side then lfs mirror extend will generally indicate that the file was busy regardless of the cause of the failure. For example, if a file already has 16 mirrors then.

lfs mirror extend -N -p ddn_ssd f0
lfs mirror mirror: cannot get UNLOCK lease, ext 4: Device or resource busy (16)
error: lfs mirror extend: f0: cannot merge layout: Device or resource busy

In the RPC handler, we see -ERANGE returned by the lod handlers

00000004:00000001:3.0:1637071921.523456:0:24268:0:(lod_object.c:3388:lod_declare_layout_merge()) Process leaving (rc=18446744073709551582 : -34 : ffffffffffffffde)
00000004:00000001:3.0:1637071921.523500:0:24268:0:(mdt_open.c:2108:mdt_close_handle_layouts()) Process leaving via out_unlock2 (rc=18446744073709551582 : -34 : 0xffffffffffffffde)
00000004:00000002:3.0:1637071921.523805:0:24268:0:(mdt_open.c:2311:mdt_mfd_close()) lustre-MDT0000: cannot swap layout of [0x200000404:0x1:0x0]: rc = -34

But in mdt_mfd_close() the rc from the layout operation is clobbered. And in ll_lease_close_intent() we just check if OBD_MD_CLOSE_INTENT_EXECED is set and return -EBUSY if not.

This is misleading to users and makes support analysis difficult. We need to have the real rc returned up to userspace here.

Issue Links

• is related to: LU-15392 mdt_mfd_close() error proapgation
• is related to: LU-15572 Interop sanity-flr failing with "cannot get UNLOCK lease"
• is related to: LU-15552 Interop: sanity-flr test 0d fails with 'exceeded maximum mirror count returns 16 not ERANGE(34)'