[LU-13974] Wrong behavior for out operations create+write especially for update log

Related — Fri, 14 Oct 2022 20:00:23 +0000

I've catch a next problem with update log.

When MDT writes to a new update log, after failover, a writer creates two update request. A first update with llog CREATE and catalog modification. A second update with WRITE header and record with index 1. Both are sending through a out queue.

After first update request osp_create_interpreter set zero to obj->opo_attr.la_valid. Any new adding llog record would call dt_attr_get(cahe is invalid) for llog before dt_write. A size is 0, and llog record with index 2 would be written right after a llog header.

So any new llog add between end of osp_create_interpreter(1 update req) and osp_write_interpreter(2 update req) would lead to update log corruption.

Here is a log of the problem

40000000:00000040:1.0:1600183643.158176:0:15651:0:(fid_request.c:400:seq_client_alloc_fid()) cli-cli-lustre-MDT0000-osp-MDT0001: Allocated FID [0x2000061c1:0x1:0x0]
00000004:00000040:1.0:1600183643.158181:0:15651:0:(mdt_handler.c:5793:mdt_object_init()) object init, fid = [0x2000061c1:0x1:0x0]
00000040:00001000:1.0:1600183643.158557:0:15651:0:(llog_osd.c:400:llog_osd_write_rec()) new record 106a0000 to [0x2000061c1:0x1:0x0]
00000004:00000040:1.0:1600183643.158559:0:15651:0:(osp_md_object.c:1177:osp_md_write()) write [0x2000061c1:0x1:0x0] offset = 0 length = 32768
00000004:00000040:1.0:1600183643.158597:0:15651:0:(osp_md_object.c:1177:osp_md_write()) write [0x2000061c1:0x1:0x0] offset = 32768 length = 464
00000040:00080000:1.0:1600183643.158600:0:15651:0:(llog_osd.c:708:llog_osd_write_rec()) added record [0x2000061c1:0x1:0x0].1, 464 off33232
00000004:00000040:1.0:1600183643.158614:0:29981:0:(osp_trans.c:266:object_update_request_dump()) i = 0 fid = [0x2000061c1:0x1:0x0] op = create params = 1 batchid = 0 size = 256 repsize 0
00000004:00000040:0.0:1600183643.158743:0:15657:0:(mdt_handler.c:5793:mdt_object_init()) object init, fid = [0x2000061c1:0x1:0x0]
00000020:00000040:0.0:1600183643.158763:0:15657:0:(out_handler.c:574:out_index_insert()) lustre-MDT0000: [0x20000000a:0x0:0x0] index insert [0x2000061c1:0x1:0x0]: rc = 0
00000020:00001000:0.0:1600183643.158777:0:15657:0:(out_lib.c:548:out_tx_create_exec()) lustre-MDT0000-osd: create [0x2000061c1:0x1:0x0]: dof 0, mode 100000
00000020:00000040:0.0:1600183643.158819:0:15657:0:(out_lib.c:1092:out_obj_index_insert()) lustre-MDT0000-osd: index insert [0x20000000a:0x0:0x0] name: [0x2000061c1:0x1:0x0] fid [0x2000061c1:0x1:0x0], type 32768
00000040:00001000:1.0:1600183643.163358:0:16581:0:(llog_osd.c:400:llog_osd_write_rec()) new record 106a0000 to [0x2000061c1:0x1:0x0]
00000004:00000040:1.0:1600183643.163374:0:16581:0:(osp_trans.c:266:object_update_request_dump()) i = 0 fid = [0x2000061c1:0x1:0x0] op = attr_get params = 0 batchid = 0 size = 40 repsize 208
00000004:00000040:1.0:1600183643.163641:0:16581:0:(osp_md_object.c:1177:osp_md_write()) write [0x2000061c1:0x1:0x0] offset = 0 length = 32768
00000004:00000040:1.0:1600183643.163686:0:16581:0:(osp_md_object.c:1177:osp_md_write()) write [0x2000061c1:0x1:0x0] offset = 32768 length = 424
00000040:00080000:1.0:1600183643.163690:0:16581:0:(llog_osd.c:708:llog_osd_write_rec()) added record [0x2000061c1:0x1:0x0].2, 424 off33192
00000004:00000040:1.0:1600183643.163790:0:29981:0:(osp_trans.c:266:object_update_request_dump()) i = 0 fid = [0x2000061c1:0x1:0x0] op = write params = 2 batchid = 0 size = 32832 repsize 0
00000004:00000040:1.0:1600183643.163793:0:29981:0:(osp_trans.c:266:object_update_request_dump()) i = 1 fid = [0x2000061c1:0x1:0x0] op = write params = 2 batchid = 0 size = 528 repsize 0
00000020:00000040:0.0:1600183643.164058:0:15657:0:(out_lib.c:695:out_tx_write_exec()) write [0x2000061c1:0x1:0x0] pos 0 buf ffff9bdfea590038, len 32768
00000020:00000040:0.0:1600183643.164174:0:15657:0:(out_lib.c:695:out_tx_write_exec()) write [0x2000061c1:0x1:0x0] pos 32768 buf ffff9bdfea598078, len 464
00000004:00000040:1.0:1600183643.164381:0:29981:0:(osp_trans.c:266:object_update_request_dump()) i = 0 fid = [0x2000061c1:0x1:0x0] op = write params = 2 batchid = 0 size = 32832 repsize 0
00000004:00000040:1.0:1600183643.164383:0:29981:0:(osp_trans.c:266:object_update_request_dump()) i = 1 fid = [0x2000061c1:0x1:0x0] op = write params = 2 batchid = 0 size = 488 repsize 0
00000020:00000040:0.0:1600183643.164533:0:15657:0:(out_lib.c:695:out_tx_write_exec()) write [0x2000061c1:0x1:0x0] pos 0 buf ffff9bdfea590038, len 32768
00000020:00000040:0.0:1600183643.164545:0:15657:0:(out_lib.c:695:out_tx_write_exec()) write [0x2000061c1:0x1:0x0] pos 32768 buf ffff9bdfea598078, len 424
00000020:00080000:0.0:1600183643.164547:0:15657:0:(out_lib.c:718:out_tx_write_exec()) llog wrong idx [0x2000061c1:0x1:0x0] pos 32768, len 424, size 32768 idx 2->275382272

At log record 1 and 2 was added at the same offset.

llog_osd_write_rec()) added record [0x2000061c1:0x1:0x0].1, 464 off33232
llog_osd_write_rec()) added record [0x2000061c1:0x1:0x0].2, 424 off33192

So our with index 2 would rewrite the 1.
Any thoughts on how better to fix it?

Whamcloud Community JIRA

[LU-13974] Wrong behavior for out operations create+write especially for update log