Re: [PATCH v7 3/3] KVM: perf: kvm events analysis tool

[David Ahern <dsahern@gmail.com>]

On 9/3/12 2:48 AM, don wrote:
> 于 2012年08月31日 02:29, David Ahern 写道:
>> In addition to Andrew's comment about making the stats struct and
>> functions generic...
> Yes. :-)
>>
>> On 8/27/12 3:51 AM, Dong Hao wrote:
>> ---8<---
>>
>>> +static void exit_event_decode_key(struct event_key *key, char
>>> decode[20])
>>> +{
>>> +       const char *exit_reason = get_exit_reason(key->key);
>>> +
>>> +       snprintf(decode, 20, "%s", exit_reason);
>>> +}
>>
>> Use scnprintf rather than snprintf.
> Why? Since we don't care about the return value, what's the difference?

1. consistency (scnprintf is preferred).
2. what if a new exit reason is added in the future that is larger than 
19 characters? Better to be safe now.

>>
>> ---8<---
>>
>>> +static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
>>> +{
>>> +    int old_max_vcpu = event->max_vcpu;
>>> +
>>> +    if (vcpu_id < event->max_vcpu)
>>> +        return true;
>>> +
>>> +    while (event->max_vcpu <= vcpu_id)
>>> +        event->max_vcpu += DEFAULT_VCPU_NUM;
>>> +
>>> +    event->vcpu = realloc(event->vcpu,
>>> +                  event->max_vcpu * sizeof(*event->vcpu));
>>> +    if (!event->vcpu) {
>>
>> If realloc fails you leak memory by overwriting the current pointer.
> Thanks for pointing it out, we will terminate the running instance in
> our next
> version.

Ok. Make sure to free the memory of the previous pointer on failure 
before cleaning up. The idea is that all allocations are properly freed 
on exit.

>>
>> ---8<---
>>
>>> +static double event_stats_stddev(int vcpu_id, struct kvm_event *event)
>>> +{
>>> +    struct event_stats *stats = &event->total;
>>> +    double variance, variance_mean, stddev;
>>> +
>>> +    if (vcpu_id != -1)
>>> +        stats = &event->vcpu[vcpu_id];
>>> +
>>> +    BUG_ON(!stats->count);
>>> +
>>> +    variance = stats->M2 / (stats->count - 1);
>>> +    variance_mean = variance / stats->count;
>>> +    stddev = sqrt(variance_mean);
>>> +
>>> +    return stddev * 100 / stats->mean;
>>> +}
>>
>> perf should be consistent in the stddev it shows the user. Any reason
>> to dump relative stddev versus stddev used by perf-stat?
> Since 'perf stat' uses relative standard deviation rather than stddev,
> 'perf kvm stat'
> just follows the style of 'perf stat'.

Yes, I've been working on an idea motivated by Andi Kleen. I have 
noticed that perf stat also uses relative stddev just in a non-direct 
way. I suggest moving common stats code from perf-stat to 
utils/stat.[ch], add a rel_stddev_stats function that returns the above 
and have perf-kvm use that.


>> ---8<---
>>
>>> +    /*
>>> +     * Append "-a" only if "-p"/"--pid" is not specified since they
>>> +     * are mutually exclusive.
>>> +     */
>>> +    if (!kvm_record_specified_guest(argc, argv))
>>> +        rec_argv[i++] = STRDUP_FAIL_EXIT("-a");
>>
>> Other perf-kvm commands rely on perf-record semantics -- i.e., for
>> user to add the -a or -p option.
> You mean, remove '-a' from the default options, then:
> if a user wants to record all guest he will use 'perf stat record -a';
> and if a user wants to record the specified guest, he should use
> 'perf stat record -p xxx'?

Yes. Well, 'perf kvm stat record' with a -a or -p. That makes the new 
stat subcommand consistent with just 'perf kvm record'.

>
> Well, as the style of other subcommand, e.g., perf lock/perf sched, the
> 'perf xxx record' record all events on all cpus, no need to use '-a'.
>
> Based on mentioned above, I prefer the original way. ;)

Yes, I noticed that some commands add -a before calling cmd_record(). 
Can't change that but we can keep perf-kvm consistent with its own 
variants which means not automagically adding -a.

David