[LU-1423] 16K pagesize clients error during ls

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.1.1, Lustre 1.8.6
• Fix Version/s: Lustre 2.3.0, Lustre 2.1.3
• Type: Bug
• Priority: Minor
• Reporter: Mahmoud Hanafi
• Assignee: Yang Sheng
• Resolution: Fixed

Environment

clients = ia64 (16Kpagesize) lustre-1.8.6.81-5.1nas
server = x86(4kpagesize) lustre-2.1.1

• Severity: 2

Description

Mounting a lustre2.1.1 filesystem on a 1.8.6 client doing a ls will fail with the following error on the client logs.

LustreError: 2551436:0:(mdc_request.c:983:mdc_readpage()) Unexpected # bytes transferred: 4096 (16384 expected)
LustreError: 2556239:0:(mdc_request.c:983:mdc_readpage()) Unexpected # bytes transferred: 4096 (16384 expected)
LustreError: 2556239:0:(dir.c:949:ll_readdir_20()) error reading dir [0x12480001:0xc1bc1b0c:0x0] at 0: rc -71

Comments

[Peter Jones]

YangSheng

Could you please look into this one?

Thanks

Peter

[Mahmoud Hanafi]

Could I get an update on this.
Thanks,
Mahmoud

[Jay Lan (Inactive)]

The kernel in the IA64 system is a sles10sp2 kernel 2.6.16.60-0.77.1-default.
The kernel config set CONFIG_IA64_PAGE_SIZE_16KB=y.

[Jay Lan (Inactive)]

We also mounted a lustre 1.8.6 server on the IA64 client and it worked. So, it appears the problem was at the 2.1.1 server.

[Andreas Dilger]

This likely relates to the changes made for 1MB readdir RPCs on 2.x, but the compatibility code on 1.8 was not updated correctly. Possibly the server needs to reply with the requested RPC size for old clients that do not handle OBD_CONNECT_BRW_SIZE, rather than hard-coding the 4096-byte lu_page size.

[Jay Lan (Inactive)]

In the case that server and client have different page size, should they be settled with the higher value or lower value?

[Yang Sheng]

Hi, Jay, Could you tell me what net type is used between client&server.

[Jay Lan (Inactive)]

We use IB. ipoib on ib0 and rdma on ib1.

[Yang Sheng]

This issue related to change: http://review.whamcloud.com/#change,604

For the fix, We need to check OBD_CONNECT_BRW_SIZE on server and return single page for old client. So looks like we need restore some codes that removed from above change.

[Yang Sheng]

Patch unload to: http://review.whamcloud.com/3014

[Mahmoud Hanafi]

We installed the patch. Got a LBUG!

Lustre: 3692:0:(sec.c:1474:sptlrpc_import_sec_adapt()) import MGS->NET_0x500000a9719cf_UUID netid 50000: select flavor null^M
Lustre: 3692:0:(sec.c:1474:sptlrpc_import_sec_adapt()) Skipped 37 previous similar messages^M
LustreError: 3761:0:(mdt_handler.c:1247:mdt_sendpage()) ASSERTION(desc->bd_nob == nob) failed^M
LustreError: 3761:0:(mdt_handler.c:1247:mdt_sendpage()) LBUG^M
Pid: 3761, comm: mdt_rdpg_01^M
^M
Call Trace:^M
[<ffffffffa05e3855>] libcfs_debug_dumpstack+0x55/0x80 [libcfs]^M
[<ffffffffa05e3e95>] lbug_with_loc+0x75/0xe0 [libcfs]^M
^M
Entering kdb (current=0xffff880c1f4fa0c0, pid 3761) on processor 0 Oops: (null)^M

[Yang Sheng]

Patch update, Please retry, sorry for the inconvenient.

[Jay Lan (Inactive)]

The latest patch seemed working for us. Mahmoud tested it.

[Yang Sheng]

Hi, Jay, Cloud we close this ticket? Please advice.

[Jay Lan (Inactive)]

The patch was landed to b2_1 tree 3 days ago. So, please close this ticket.

[Yang Sheng]

Patch landed. close bug.