From mboxrd@z Thu Jan 1 00:00:00 1970 From: don Subject: Re: [PATCH v7 3/3] KVM: perf: kvm events analysis tool Date: Mon, 03 Sep 2012 16:48:42 +0800 Message-ID: <50446EEA.5060306@linux.vnet.ibm.com> References: <1346061106-5364-1-git-send-email-haodong@linux.vnet.ibm.com> <1346061106-5364-4-git-send-email-haodong@linux.vnet.ibm.com> <503FB105.9000205@gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: avi@redhat.com, acme@infradead.org, mtosatti@redhat.com, mingo@elte.hu, xiaoguangrong@linux.vnet.ibm.com, linux-kernel@vger.kernel.org, kvm@vger.kernel.org To: David Ahern Return-path: In-Reply-To: <503FB105.9000205@gmail.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: kvm.vger.kernel.org =E4=BA=8E 2012=E5=B9=B408=E6=9C=8831=E6=97=A5 02:29, David Ahern =E5=86= =99=E9=81=93: > In addition to Andrew's comment about making the stats struct and=20 > 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=20 >> decode[20]) >> +{ >> + const char *exit_reason =3D 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? > > ---8<--- > >> +static bool kvm_event_expand(struct kvm_event *event, int vcpu_id) >> +{ >> + int old_max_vcpu =3D event->max_vcpu; >> + >> + if (vcpu_id < event->max_vcpu) >> + return true; >> + >> + while (event->max_vcpu <=3D vcpu_id) >> + event->max_vcpu +=3D DEFAULT_VCPU_NUM; >> + >> + event->vcpu =3D 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=20 our next version. > > ---8<--- > >> +static double event_stats_stddev(int vcpu_id, struct kvm_event *eve= nt) >> +{ >> + struct event_stats *stats =3D &event->total; >> + double variance, variance_mean, stddev; >> + >> + if (vcpu_id !=3D -1) >> + stats =3D &event->vcpu[vcpu_id]; >> + >> + BUG_ON(!stats->count); >> + >> + variance =3D stats->M2 / (stats->count - 1); >> + variance_mean =3D variance / stats->count; >> + stddev =3D sqrt(variance_mean); >> + >> + return stddev * 100 / stats->mean; >> +} > > perf should be consistent in the stddev it shows the user. Any reason= =20 > to dump relative stddev versus stddev used by perf-stat? Since 'perf stat' uses relative standard deviation rather than stddev,=20 'perf kvm stat' just follows the style of 'perf stat'. > > ---8<--- > >> +/* returns left most element of result, and erase it */ >> +static struct kvm_event *pop_from_result(void) >> +{ >> + struct rb_node *node =3D result.rb_node; >> + >> + if (!node) >> + return NULL; >> + >> + while (node->rb_left) >> + node =3D node->rb_left; > > Use rb_first(). OK, we will use it in the next version. > > ---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++] =3D STRDUP_FAIL_EXIT("-a"); > > Other perf-kvm commands rely on perf-record semantics -- i.e., for=20 > 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'? 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. ;) > > ---8<--- > >> +static const char * const kvm_events_report_usage[] =3D { >> + "perf kvm stat report []", >> + NULL >> +}; >> + >> +static const struct option kvm_events_report_options[] =3D { >> + OPT_STRING(0, "event", &report_event, "report event", >> + "event for reporting: vmexit, mmio, ioport"), >> + OPT_INTEGER(0, "vcpu", &trace_vcpu, >> + "vcpu id to report"), >> + OPT_STRING('k', "key", &sort_key, "sort-key", >> + "key for sorting: sample(sort by samples number)" >> + " time (sort by avg time)"), >> + OPT_END() >> +}; >> + >> +static int kvm_events_report(int argc, const char **argv) >> +{ >> + symbol__init(); >> + >> + if (argc) { >> + argc =3D parse_options(argc, argv, >> + kvm_events_report_options, >> + kvm_events_report_usage, 0); >> + if (argc) >> + usage_with_options(kvm_events_report_usage, >> + kvm_events_report_options); >> + } >> + >> + return kvm_events_report_vcpu(trace_vcpu); >> +} >> + >> +static int kvm_cmd_stat(int argc, const char **argv) >> +{ >> + if (argc > 1) { >> + if (!strncmp(argv[1], "rec", 3)) >> + return kvm_events_record(argc - 1, argv + 1); >> + >> + if (!strncmp(argv[1], "rep", 3)) >> + return kvm_events_report(argc - 1 , argv + 1); >> + } >> + >> + return cmd_stat(argc, argv, NULL); >> +} >> + >> static char name_buffer[256]; >> >> static const char * const kvm_usage[] =3D { >> - "perf kvm [] {top|record|report|diff|buildid-list}", >> + "perf kvm [] {top|record|report|diff|buildid-list|stat= }", >> NULL >> }; >> > > The usage for the report/record sub commands of stat is never shown.=20 > e.g., > $ perf kvm stat > --> shows help for perf-stat > > $ perf kvm > --> shows the above and perf-kvm's usage > > To get help for the record/report subcommands you have to know that=20 > record and report are subcommands. Okay, we will improve this. > > ---8<--- > >> +static int perf_file_section__read_feature(struct perf_file_section= =20 >> *section, >> + struct perf_header *ph, >> + int feat, int fd, void *data) >> +{ >> + struct header_read_data *hd =3D data; >> + >> + if (feat !=3D hd->feat) >> + return 0; >> + >> + if (lseek(fd, section->offset, SEEK_SET) =3D=3D (off_t)-1) { >> + pr_debug("Failed to lseek to %" PRIu64 " offset for feature= " >> + "%d, continuing...\n", section->offset, feat); >> + return 0; >> + } >> + >> + if (feat >=3D HEADER_LAST_FEATURE) { >> + pr_warning("unknown feature %d\n", feat); >> + return 0; >> + } >> + >> + hd->result =3D feat_ops[feat].read(ph, fd); > > you should verify the read() function is implemented for the requeste= d=20 > feature. OK, thank you for your comments, we will correct it. :-)