[PATCH 5/5] arm/perfevents: implement perf event support for ARMv6

[Will Deacon]

Hi Jamie,

I've been trying to test your patches with a quad-core ARM 11MPCore on a
Realview PB11MP board.

Unfortunately, I occasionally experience a complete system hang during some
profiling runs. I don't think it's your fault however, as it can occur even
when monitoring only software events. I've managed to reproduce this on the 
tip/master branch and got the following information [I enabled lock debugging]:

=================================
[ INFO: inconsistent lock state ]
2.6.33-rc2-tip+ #5
---------------------------------
inconsistent {IN-HARDIRQ-W} -> {HARDIRQ-ON-W} usage.
perf/1463 [HC0[0]:SC0[0]:HE1:SE1] takes:
 (&ctx->lock){?.....}, at: [<c009e538>] __perf_event_sched_in+0x20/0x168
{IN-HARDIRQ-W} state was registered at:
  [<c00718b0>] __lock_acquire+0x5c8/0x17b4
  [<c0072b70>] lock_acquire+0xd4/0xec
  [<c0315700>] _raw_spin_lock+0x2c/0x3c
  [<c00a20f8>] perf_ctx_adjust_freq+0xc/0x1dc
  [<c00a230c>] perf_event_task_tick+0x44/0xb4
  [<c0041f4c>] scheduler_tick+0xec/0x144
  [<c0053eb8>] update_process_times+0x40/0x4c
  [<c006d91c>] tick_periodic+0xdc/0x108
  [<c006d96c>] tick_handle_periodic+0x24/0xf0
  [<c0030124>] ipi_timer+0x34/0x44
  [<c002a3cc>] do_local_timer+0x50/0x80
  [<c002aca4>] __irq_svc+0x44/0xe0
  [<c002c388>] default_idle+0x28/0x2c
  [<c002c8ac>] cpu_idle+0x8c/0xe4
  [<70008080>] 0x70008080
irq event stamp: 454
hardirqs last  enabled at (454): [<c0315d30>] _raw_spin_unlock_irq+0x24/0x2c
hardirqs last disabled at (453): [<c0315784>] _raw_spin_lock_irq+0x18/0x50
softirqs last  enabled at (0): [<c0045650>] copy_process+0x328/0xf70
softirqs last disabled at (0): [<(null)>] (null)
 
other info that might help us debug this:
no locks held by perf/1463.
 
stack backtrace:
[<c0031694>] (unwind_backtrace+0x0/0xd4) from [<c006ff88>] (print_usage_bug+0x16c/0x1ac)
[<c006ff88>] (print_usage_bug+0x16c/0x1ac) from [<c00702f0>] (mark_lock+0x328/0x5f0)
[<c00702f0>] (mark_lock+0x328/0x5f0) from [<c007193c>] (__lock_acquire+0x654/0x17b4)
[<c007193c>] (__lock_acquire+0x654/0x17b4) from [<c0072b70>] (lock_acquire+0xd4/0xec)
[<c0072b70>] (lock_acquire+0xd4/0xec) from [<c0315700>] (_raw_spin_lock+0x2c/0x3c)
[<c0315700>] (_raw_spin_lock+0x2c/0x3c) from [<c009e538>] (__perf_event_sched_in+0x20/0x168)
[<c009e538>] (__perf_event_sched_in+0x20/0x168) from [<c009e6c8>] (perf_event_task_sched_in+0x48/0x58)
[<c009e6c8>] (perf_event_task_sched_in+0x48/0x58) from [<c003f2dc>] (finish_task_switch+0x34/0xb4)
[<c003f2dc>] (finish_task_switch+0x34/0xb4) from [<c03133e0>] (schedule+0x728/0x834)
[<c03133e0>] (schedule+0x728/0x834) from [<c00d5190>] (pipe_wait+0x64/0x84)
[<c00d5190>] (pipe_wait+0x64/0x84) from [<c00d59d8>] (pipe_read+0x3ac/0x428)
[<c00d59d8>] (pipe_read+0x3ac/0x428) from [<c00cdd34>] (do_sync_read+0x94/0xe0)
[<c00cdd34>] (do_sync_read+0x94/0xe0) from [<c00ce868>] (vfs_read+0xa8/0x150)
[<c00ce868>] (vfs_read+0xa8/0x150) from [<c00ce9bc>] (sys_read+0x3c/0x68)
[<c00ce9bc>] (sys_read+0x3c/0x68) from [<c002b140>] (ret_fast_syscall+0x0/0x38)

This occurs more frequently when profiling code that spawns child processes
[e.g. gcc] but I've see it happen with sleep too. It also happens if I pass
maxcpus=1 as a bootarg.

Have you seen this happen on your board? I've failed to reproduce it on my
x86 machine.

Will


* Jamie Iles wrote:

> This patch implements support for ARMv6 performance counters in the
> Linux performance events subsystem. ARMv6 architectures that have the
> performance counters should enable HW_PERF_EVENTS and define the
> interrupts for the counters in arch/arm/kernel/perf_event.c
> 
> This implementation also provides an ARM PMU abstraction layer to allow
> ARMv7 and others to be supported in the future by adding new a
> 'struct arm_pmu'.
> 
> Signed-off-by: Jamie Iles <jamie.iles at picochip.com>
> Cc: Peter Zijlstra <peterz at infradead.org>
> Cc: Ingo Molnar <mingo at elte.hu>
> Cc: Jean Pihet <jpihet at mvista.com>
> Cc: Will Deacon <will.deacon at arm.com>
> ---
>  arch/arm/Kconfig             |    8 +
>  arch/arm/kernel/Makefile     |    1 +
>  arch/arm/kernel/perf_event.c | 1338 ++++++++++++++++++++++++++++++++++++++++++
>  3 files changed, 1347 insertions(+), 0 deletions(-)
>  create mode 100644 arch/arm/kernel/perf_event.c

<truncated>