[LU-5050] cpu partitioning oddities

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.9.0
• Labels: llnl
• Severity: 3
• Assignee: Dmitry Eremin (Inactive)
• Reporter: Christopher Morrone (Inactive)

Description

In tracking down the root cause of LU-5043, I went through the cpu partitioning code in Lustre and found some of it rather difficult to follow. A number of not-necessarily-safe assumptions seem to made, with no code comments explaining what was assumed and why. I think we have some technical debt to deal with here.

Issue Links

• is related to: LU-8610 Better handle of HW dependency of DLC configuration
• is related to: LU-7553 Lustre cpu_npartitions default value breaks memory allocation on clients
• is related to: LU-8395 Limit lnet-selftest Threads
• is related to: LU-11163 CPT-bound allocations can fail if NUMA node is OOM

Comments

[Peter Jones]

Landed for 2.9

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch http://review.whamcloud.com/22377/
Subject: LU-5050 libcfs: default CPT matches NUMA topology
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 4fb513d0a635ce749ddb2173e9841814622ba4a2

[Gerrit Updater]

Andreas Dilger ([e-mail]) uploaded a new patch: http://review.whamcloud.com/22507
Subject: LU-5050 libcfs: default CPT matches NUMA topology
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: cf67b83eb74bc84e90d0a02ea306405a0fc76b43

[Gerrit Updater]

Dmitry Eremin ([e-mail]) uploaded a new patch: http://review.whamcloud.com/22377
Subject: LU-5050 libcfs: default CPT matches NUMA topology
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 1916e280b1f0ac579d17a6cb9134923f1481176c

[Andreas Dilger]

Dmitry, could you please work on an updated version of the http://review.whamcloud.com/17824 patch. We'd like to get it included into 2.9.0.

[James A Simmons]

That is one old bug. I have a question. How did it ever pass testing on RHEL7 in the first place?

[Olaf Weber (Inactive)]

Quick guess as to the cause is that cfs_trimwhite() modifies the input string, while the string literal "N" was stored in read-only memory. The easiest solution is likely for cfs_cpt_table_create_pattern() to make a copy of the pattern string.