Re: [PATCH] perf/x86/intel/cqm: Move WARN_ONs from intel_cqm_cpu_prepare to cqm_pick_event_reader

[Matt Fleming <matt@codeblueprint.co.uk>]

On Tue, 11 Aug, at 01:31:05PM, Yasuaki Ishimatsu wrote:
> When hot adding a CPU and onlining it, the following WARN_ON() messages
> are shown:
> 
> [  772.891448] ------------[ cut here ]------------
> [  772.896624] WARNING: CPU: 58 PID: 15169 at arch/x86/kernel/cpu/perf_event_intel_cqm.c:1268 intel_cqm_cpu_prepare+0x88/0x90()
> [  772.909167] Modules linked in:
> [  772.995134] CPU: 58 PID: 15169
> [  773.016633]  0000000000000000 0000000092fb60ed ffff88104febbba8 ffffffff8167b5fa
> [  773.024789]  0000000000000000 0000000000000000 ffff88104febbbe8 ffffffff810819ea
> [  773.033119]  ffff88103be60000 ffff8c0fbc7ca020 ffffffff819fadf0 000000000000008f
> [  773.041461] Call Trace:
> [  773.044402]  [<ffffffff8167b5fa>] dump_stack+0x45/0x57
> [  773.050160]  [<ffffffff810819ea>] warn_slowpath_common+0x8a/0xc0
> [  773.056888]  [<ffffffff81081b1a>] warn_slowpath_null+0x1a/0x20
> [  773.063426]  [<ffffffff810365f8>] intel_cqm_cpu_prepare+0x88/0x90
> [  773.070253]  [<ffffffff81036732>] intel_cqm_cpu_notifier+0x42/0x160
> [  773.077271]  [<ffffffff810a0d3d>] notifier_call_chain+0x4d/0x80
> [  773.083901]  [<ffffffff810a0e4e>] __raw_notifier_call_chain+0xe/0x10
> [  773.091007]  [<ffffffff81081ef8>] _cpu_up+0xe8/0x190
> [  773.096555]  [<ffffffff8108201a>] cpu_up+0x7a/0xa0
> [  773.101910]  [<ffffffff816701b0>] cpu_subsys_online+0x40/0x90
> [  773.108332]  [<ffffffff8143d777>] device_online+0x67/0x90
> [  773.114368]  [<ffffffff8143d82a>] online_store+0x8a/0xa0
> [  773.120305]  [<ffffffff8143aab8>] dev_attr_store+0x18/0x30
> [  773.126437]  [<ffffffff8127224a>] sysfs_kf_write+0x3a/0x50
> [  773.132560]  [<ffffffff812718d0>] kernfs_fop_write+0x120/0x170
> [  773.139078]  [<ffffffff811f7657>] __vfs_write+0x37/0x100
> [  773.145019]  [<ffffffff811fa398>] ? __sb_start_write+0x58/0x110
> [  773.151635]  [<ffffffff8129d7ed>] ? security_file_permission+0x3d/0xc0
> [  773.158932]  [<ffffffff811f7d59>] vfs_write+0xa9/0x190
> [  773.164674]  [<ffffffff810234e6>] ? do_audit_syscall_entry+0x66/0x70
> [  773.171776]  [<ffffffff811f8b55>] SyS_write+0x55/0xc0
> [  773.177423]  [<ffffffff810672f0>] ? do_page_fault+0x30/0x80
> [  773.183654]  [<ffffffff8168232e>] entry_SYSCALL_64_fastpath+0x12/0x71
> [  773.190843] ---[ end trace e6219d24386873bd ]---
> [  773.196573] smpboot: Booting Node 7 Processor 143 APIC 0x1f7
> [  773.221241] microcode: CPU143 sig=0x306f3, pf=0x80, revision=0x9
> [  773.228005] Will online and init hotplugged CPU: 143
> 
> Here is the root cause of the issue:
> When calling intel_cqm_cpu_prepare() at CPU_UP_PREPARE notification,
> the function checks that c86_chache_max_rmid is same as cqm_max_rmid
> as follows:
> 
> static void intel_cqm_cpu_prepare(unsigned int cpu)
> {
> ...
>         WARN_ON(c->x86_cache_max_rmid != cqm_max_rmid);
> 
> But x86_cache_max_rmid of hot added CPU is not set yet, because it will 
> set in get_cpu_cap() which is called after CPU_UP_PREPARE notification.
> 
> So when onlining a hot added CPU, the WARN_ON() are always shown:

Thanks for the report! I submitted the following patch on August 6th to
fix this issue,

  https://lkml.kernel.org/r/1438863163-14083-1-git-send-email-matt@codeblueprint.co.uk

> To fix the issue, the patch moves WARN_ON()s from intel_cqm_cpu_prepare() to
> cqm_pick_event_reader() which is called at CPU_STARTING notification.

I took a different approach and moved the notifier from CPU_UP_PREPARE
to CPU_STARTING since there's no reason to call the handler as early as
CPU_UP_PREPARE time at all.

Yes, we could collapse cqm_pick_event_reader() and
intel_cqm_cpu_prepare() together but I have a patch to add a
CPU_DOWN_FAILED call that invokes cqm_pick_event_reader() but not
intel_cqm_cpu_prepare(). I actually need to mail that out.

-- 
Matt Fleming, Intel Open Source Technology Center