Re: [PATCH v8 1/4] perf,kvm/{x86,s390}: Remove dependency on uapi/kvm_perf.h

[David Ahern <dsahern@gmail.com>]

On 9/28/15 9:16 AM, Scott Wood wrote:
> On Mon, 2015-09-28 at 08:31 -0600, David Ahern wrote:
>> On 9/28/15 7:00 AM, Alexander Yarygin wrote:
>>>> diff --git a/tools/perf/builtin-kvm.c b/tools/perf/builtin-kvm.c
>>>> index fc1cffb..ef25fcf 100644
>>>> --- a/tools/perf/builtin-kvm.c
>>>> +++ b/tools/perf/builtin-kvm.c
>>>> @@ -31,20 +31,18 @@
>>>>    #include <math.h>
>>>>
>>>>    #ifdef HAVE_KVM_STAT_SUPPORT
>>>> -#include <asm/kvm_perf.h>
>>>>    #include "util/kvm-stat.h"
>>>>
>>>> -void exit_event_get_key(struct perf_evsel *evsel,
>>>> -                 struct perf_sample *sample,
>>>> +void exit_event_get_key(struct perf_evsel *evsel, struct perf_sample
>>>> *sample,
>>>>                            struct event_key *key)
>>>>    {
>>>>            key->info = 0;
>>>> - key->key = perf_evsel__intval(evsel, sample, KVM_EXIT_REASON);
>>>> + key->key = perf_evsel__intval(evsel, sample, exit_reason_code);
>>>>    }
>>>>
>>>>    bool kvm_exit_event(struct perf_evsel *evsel)
>>>>    {
>>>> - return !strcmp(evsel->name, KVM_EXIT_TRACE);
>>>> + return !strncmp(evsel->name, kvm_events_tp[1], strlen(evsel->name));
>>>>    }
>>>
>>> Hmm, direct access to kvm_events_tp? Maybe add a getter for this or
>>> something like extern char *kvm_exit_trace;?
>>> /* why strncmp? */
>>>
>>>>
>>>>    bool exit_event_begin(struct perf_evsel *evsel,
>>>> @@ -60,7 +58,7 @@ bool exit_event_begin(struct perf_evsel *evsel,
>>>>
>>>>    bool kvm_entry_event(struct perf_evsel *evsel)
>>>>    {
>>>> - return !strcmp(evsel->name, KVM_ENTRY_TRACE);
>>>> + return !strncmp(evsel->name, kvm_events_tp[0], strlen(evsel->name));
>>>>    }
>>>>
>>>>    bool exit_event_end(struct perf_evsel *evsel,
>>
>> I agree; don't rely on kvm_events_tp. Define KVM_ENTRY_TRACE and
>> KVM_EXIT_TRACE like x86.
>
> If you mean defining them in uapi, that doesn't work for arches that have
> multiple subarches that may have different trace events.  This patchset
> doesn't actually implement dynamic support for the subarches, but it avoids
> adding constants to uapi headers that only apply to one of the subarches.

I don't agree on relying on kvm_events_tp[0] and [1]. If you need that 
to be a runtime definition then change KVM_ENTRY_TRACE to const char 
*kvm_entry_trace and s390 and other arches can have code to set 
kvm_{entry,exit}_trace at runtime.

David