On Tue, Jan 21, 2014 at 09:59:54PM +0000, Stephen Boyd wrote:

On 01/21/14 10:37, Stephen Boyd wrote:

quoted

On 01/21/14 10:07, Will Deacon wrote:

quoted

Finally, I'd really like to see this get some test coverage, but I don't
want to try running mainline on my phone :) Could you give your patches a
spin with Vince's perf fuzzer please?

  https://github.com/deater/perf_event_tests.git

(then build the contents of the fuzzer directory and run it for as long as
you can).

Ok. I'll see what I can do.

Neat. This quickly discovered a problem.

Yeah, the fuzzer is good at that!

BUG: using smp_processor_id() in preemptible [00000000] code: perf_fuzzer/70
caller is krait_pmu_get_event_idx+0x58/0xd8
CPU: 2 PID: 70 Comm: perf_fuzzer Not tainted 3.13.0-rc7-next-20140108-00041-gb038353c8516-dirty #58
[<c02165dc>] (unwind_backtrace) from [<c0212ffc>] (show_stack+0x10/0x14)
[<c0212ffc>] (show_stack) from [<c06ad5fc>] (dump_stack+0x6c/0xb8)
[<c06ad5fc>] (dump_stack) from [<c04b0928>] (debug_smp_processor_id+0xc4/0xe8)
[<c04b0928>] (debug_smp_processor_id) from [<c021a19c>] (krait_pmu_get_event_idx+0x58/0xd8)
[<c021a19c>] (krait_pmu_get_event_idx) from [<c021833c>] (validate_event+0x2c/0x54)
[<c021833c>] (validate_event) from [<c02186b4>] (armpmu_event_init+0x264/0x2dc)
[<c02186b4>] (armpmu_event_init) from [<c02b28e4>] (perf_init_event+0x138/0x194)
[<c02b28e4>] (perf_init_event) from [<c02b2c04>] (perf_event_alloc+0x2c4/0x348)
[<c02b2c04>] (perf_event_alloc) from [<c02b37e4>] (SyS_perf_event_open+0x7b8/0x9cc)
[<c02b37e4>] (SyS_perf_event_open) from [<c020ef80>] (ret_fast_syscall+0x0/0x48)

This is happening because the pmresrn_used mask is per-cpu but
validate_group() is called in preemptible context. It looks like we'll
need to add another unsigned long * field to struct pmu_hw_events just
for Krait purposes? Or we could use the upper 16 bits of the used_mask
bitmap to hold the pmresr_used values? Sounds sort of hacky but it
avoids adding another bitmap for every PMU user and there aren't any
Krait CPUs with more than 5 counters anyway.

I'm fine with that approach, but a comment saying what we're doing with
those upper bits wouldn't go amiss.

This is the diff for the latter version. I'll let the fuzzer run overnight.

Cheers,

Will