[LU-3992] Fix NUMA emulated mode

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.6.0, Lustre 2.5.2
• Labels: patch
• Severity: 3
• Assignee: Liang Zhen (Inactive)
• Reporter: Andriy Skulysh

Description

Kernel commit c1c3443c9c5e9be92641029ed229a41563e44506
assigns all allowed cpus to emulated node.

Comments

[Wally Wang (Inactive)]

Is there any concern or 'side effect' for the fix? We'd like to adopt the fix over the cpu-pattern workaround.

[Wally Wang (Inactive)]

We run into this problem and with the patch it works fine for us. I think at least the user/admin prefer a fix instead of using cpu_pattern as a workaround. Is there a probelm with the fix?

[Andriy Skulysh]

ncpt is always > 0 and we do only cpt++, so "if" can be only "=="
What errors can be here ? Only with strange cfs_node_to_cpumask() like in emulated NUMA case.
Resulting topology can be easily examined via /proc/sys/lnet/cpu_partition_table.
Is new parameter really needed ? We will force user to create dummy cpu_pattern.

[Liang Zhen (Inactive)]

I think module parameter cpu_pattern can work around this, would it be OK if you just use this parameter instead of adding a patch? I'd like user to see these errors when situation like this happened.

[Andriy Skulysh]

each emulated node has all cpus in cpumask (cpumask_of_node()).
So each cpu exists in each node's mask
We need to stop loop when all cpus are assigned. It doesn't matter which cpu is chosen for node.

[Liang Zhen (Inactive)]

sorry I failed to understand how this can happen, could you give an example of this?

[Andriy Skulysh]

without the fix insmod libcfs.ko fails with numa=fake=16 in kernel boot cmdline.

    LNetError: 4055:0:(linux-cpu.c:881:cfs_cpt_table_create()) Failed to setup CPU-partition-table with 4 CPU-partitions, online HW nodes: 16, HW cpus: 32.
    LNetError: 4055:0:(linux-cpu.c:1093:cfs_cpu_init()) Failed to create ptable with npartitions 0

[Andriy Skulysh]

patch: http://review.whamcloud.com/7724