-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
A client that reads the layout of a file with a component whose pattern it does not recognize oopses in lov_lsme_pack_v1v3(), e.g. "lfs getstripe" of such a file:
lov: unrecognized striping pattern: rc = 0 lov_dump_lmm_common()) objid 0x5:1026, magic 0x0bd10bd0, pattern 0x5 lov_init_composite()) lustre-clilov-...: unknown layout entry 1 pattern 0x5 could be an unrecognizable component set by other clients, skip to initialize the next component. BUG: kernel NULL pointer dereference, address: 0000000000000000 RIP: 0010:lov_lsme_pack_v1v3+0x90/0x120 [lov] lov_lsm_pack+0x1d8/0x450 [lov] lov_object_layout_get+0x157/0x2f0 [lov] ll_getxattr_lov+0x1de/0x410 [lustre] ll_xattr_get+0x129/0x1a0 [lustre] __vfs_getxattr+0x7e/0xb0 do_getxattr+0x129/0x180
LU-16837 (https://review.whamcloud.com/51060, "llite: handle unknown layout component") made lsme_unpack() allocate no lsme_oinfo[] for a component whose pattern is not lov_pattern_available(), but it keeps lsme_stripe_count and the INIT flag. lsme_free() and dump_lsm() skip the objects of such a component (and https://review.whamcloud.com/52826 fixed lov_delete_composite()), but lov_lsme_pack_v1v3() still packs lsme_stripe_count objects from lsme_oinfo[].
Seen on master v2_17_62 with an EC parity component that LFSCK rebuilt with the RAID0|PARITY pattern, which master no longer accepts since LOV_PATTERN_PARITY stands alone (LU-20690). The same applies to any pattern a newer server or client sets that this client does not know.
Parsing the objects of such a component regardless of its pattern would let getxattr return them, so that copying the layout (e.g. lfs migrate, mirror split) does not drop them.