Re: [PATCH v5 0/7] arm64: perf: Support for chained counters
From: Suzuki K Poulose
Date: Tue Jul 10 2018 - 11:01:16 EST
On 07/10/2018 03:31 PM, Will Deacon wrote:
Hi Suzuki,
On Tue, Jul 10, 2018 at 09:57:57AM +0100, Suzuki K Poulose wrote:
This series adds support for counting PMU events using 64bit counters
for arm64 PMU.
The Arm v8 PMUv3 supports combining two adjacent 32bit counters
(low even and hig odd counters) to count a given "event" in 64bit mode.
This series adds the support for 64bit events in the core arm_pmu driver
infrastructure and adds the support for armv8 64bit kernel PMU to use
chained counters to count in 64bit mode. For CPU cycles, we use the cycle
counter in 64bit mode, only when requested. If the cycle counter is not
available, we fall back to chaining the counters.
Tested on Juno, Fast models. Applies on 4.18-rc4
Thanks, this looks pretty good to me. How far did you get with the perf
fuzzer?
Hi Will,
I ran perf_fuzzer on Juno and the fuzzer hung with RCU stalls.
However, this was reproducible without the series. I am yet to
investigate the RCU stall.
Log below (from 4.18-rc3, without the series).
[445885.968489] perf_fuzzer R running task 0 11468 4397
0x00000003
[445885.975568] Call trace:
[445885.978084] ret_from_fork+0x0/0x1c
[445963.841606] INFO: rcu_preempt self-detected stall on CPU
[445963.845814] INFO: rcu_preempt detected stalls on CPUs/tasks:
[445963.846957] 2-....: (1 GPs behind) idle=376/1/4611686018427387908
softirq=57145/57145 fqs=54519307
[445963.852649]
[445963.861784] 2-....: (1 GPs behind) idle=376/1/4611686018427387908
softirq=57145/57145 fqs=54519307
[445963.861787] (t=109254007 jiffies g=61092 c=61091 q=1517144)
[445963.878244]
[445963.878245] Task dump for CPU 2:
[445963.878250] perf_fuzzer R
[445963.879897] (detected by 0, t=109254007 jiffies, g=61092, c=61091,
q=1517144)
[445963.883167] running task 0 11468 4397 0x00000003
[445963.898932] Call trace:
[445963.901439] dump_backtrace+0x0/0x1e0
[445963.905149] show_stack+0x24/0x30
[445963.908516] sched_show_task+0x220/0x2c0
[445963.912484] dump_cpu_task+0x48/0x58
[445963.916109] rcu_dump_cpu_stacks+0xa0/0xe0
[445963.920250] rcu_check_callbacks+0xa40/0xc20
[445963.924564] update_process_times+0x34/0x60
[445963.928791] tick_sched_handle.isra.5+0x44/0x68
[445963.933361] tick_sched_timer+0x50/0xa0
[445963.937244] __hrtimer_run_queues+0x19c/0x6d0
[445963.941644] hrtimer_interrupt+0xec/0x248
[445963.945699] arch_timer_handler_phys+0x3c/0x50
[445963.950184] handle_percpu_devid_irq+0xcc/0x4a0
[445963.954754] generic_handle_irq+0x34/0x50
[445963.958808] __handle_domain_irq+0x68/0xc0
[445963.962948] gic_handle_irq+0x60/0xb8
[445963.966657] el1_irq+0xb4/0x130
[445963.969851] __do_softirq+0xcc/0x5c4
[445963.973475] irq_exit+0x12c/0x138
[445963.976841] __handle_domain_irq+0x6c/0xc0
[445963.980981] gic_handle_irq+0x60/0xb8
[445963.984691] el0_irq_naked+0x50/0x5c
[445963.988316] Task dump for CPU 2:
[445963.991601] perf_fuzzer R running task 0 11468 4397
0x00000003
[445963.998679] Call trace:
[445964.001194] ret_from_fork+0x0/0x1c
root@localhost:~# uname -a
Linux localhost.localdomain 4.18.0-rc3 #117 SMP PREEMPT Thu Jul 5
11:47:12 BST 2018 aarch64 aarch64 aarch64 GNU/Linux
Suzuki