[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAJpZYjXY4R=t1Wn2bkjEuThyCAC8_HTZ6JwtWPuQEs=weshbGQ@mail.gmail.com>
Date: Fri, 1 Nov 2024 14:32:39 -0700
From: Chun-Tse Shao <ctshao@...gle.com>
To: Ian Rogers <irogers@...gle.com>
Cc: linux-kernel@...r.kernel.org, Peter Zijlstra <peterz@...radead.org>,
Ingo Molnar <mingo@...hat.com>, Arnaldo Carvalho de Melo <acme@...nel.org>, Namhyung Kim <namhyung@...nel.org>,
Mark Rutland <mark.rutland@....com>,
Alexander Shishkin <alexander.shishkin@...ux.intel.com>, Jiri Olsa <jolsa@...nel.org>,
Adrian Hunter <adrian.hunter@...el.com>, Kan <kan.liang@...ux.intel.com>,
Ze Gao <zegao2021@...il.com>, Weilin Wang <weilin.wang@...el.com>,
linux-perf-users@...r.kernel.org
Subject: Re: [PATCH 3/3] perf evsel: Find process with busy PMUs for EBUSY
Thank you for your review! Here is the link to the v2 patch with
fixes: https://lore.kernel.org/all/20241101211757.824743-3-ctshao@google.com/
On Fri, Nov 1, 2024 at 9:14 AM Ian Rogers <irogers@...gle.com> wrote:
>
> On Thu, Oct 31, 2024 at 3:39 PM Chun-Tse Shao <ctshao@...gle.com> wrote:
> >
> > It parses fdinfo with PMU type, comparing with the event which failed to
> > open, and report the processes causing EBUSY error.
> >
> > ```
> > Testing cycles and intel_pt//
> > $ ./perf stat -e cycles &
> > [1] 55569
> > $ ./perf stat -e intel_pt// &
> > [2] 55683
> > $ ./perf stat -e intel_pt//
> > Error:
> > The PMU intel_pt counters are busy and in use by another process.
> > Possible processes:
> > 55683 ./perf stat -e intel_pt//
> > ```
> > Only perf with intel_pt was reported.
>
> I think this is a very nice addition. It is a shame there is a race
> with the existing process exiting, between the perf_event_open and the
> /proc scanning. A PMU may return EBUSY just because say perf list is
> probing features on the PMU so we probably have some extra retries for
> EBUSY too.
>
> > Signed-off-by: Chun-Tse Shao <ctshao@...gle.com>
> > ---
> > tools/perf/util/evsel.c | 79 +++++++++++++++++++++++++++++------------
> > 1 file changed, 57 insertions(+), 22 deletions(-)
> >
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index 9a5b6a6f8d2e5..d2f7c19e023ec 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -3286,7 +3286,8 @@ static bool find_process(const char *name)
> > return ret ? false : true;
> > }
> >
> > -static int dump_perf_event_processes(char *msg, size_t size)
> > +static int dump_perf_event_processes(const struct perf_event_attr *failed_attr,
> > + char *msg, size_t size)
> > {
> > DIR *proc_dir;
> > struct dirent *proc_entry;
> > @@ -3327,29 +3328,61 @@ static int dump_perf_event_processes(char *msg, size_t size)
> > continue;
> > /* Take care as readlink doesn't null terminate the string. */
> > if (!strncmp(path, "anon_inode:[perf_event]", link_size)) {
> > - int cmdline_fd;
> > - ssize_t cmdline_size;
> > -
> > - scnprintf(path, sizeof(path), "%s/cmdline", proc_entry->d_name);
> > - cmdline_fd = openat(dirfd(proc_dir), path, O_RDONLY);
> > - if (cmdline_fd == -1)
> > - continue;
> > - cmdline_size = read(cmdline_fd, path, sizeof(path) - 1);
> > - close(cmdline_fd);
> > - if (cmdline_size < 0)
> > + int fdinfo_fd;
> > + ssize_t fdinfo_size;
> > + char *line;
> > + u32 perf_event_type = PERF_TYPE_MAX;
>
> PERF_TYPE_MAX is beyond the pre-defined perf PMU types at 6 but PMU
> drivers loaded by the kernel may use this number - I think task-clock
> may use this PMU type number but it shouldn't return EBUSY. Anyway, I
> think -1 would be a better marker to use here and in the corresponding
> check below.
>
> Thanks,
> Ian
In this case I am using UINT32_MAX instead to avoid the overflow. If
kernel would still use this value, I can make it u64 or add a variable
`found` instead.
>
> > +
> > + /* Let's check the PMU type reserved by this process */
> > + scnprintf(path, sizeof(path), "%s/fdinfo/%s",
> > + proc_entry->d_name, fd_entry->d_name);
> > + fdinfo_fd = openat(dirfd(proc_dir), path, O_RDONLY);
> > + fdinfo_size = read(fdinfo_fd, path, sizeof(path) - 1);
> > + if (fdinfo_size < 0)
> > continue;
> > - path[cmdline_size] = '\0';
> > - for (ssize_t i = 0; i < cmdline_size; i++) {
> > - if (path[i] == '\0')
> > - path[i] = ' ';
> > + path[fdinfo_size] = '\0';
> > +
> > + line = strtok(path, "\n");
> > + while (line != NULL) {
> > + if (sscanf(line,
> > + "perf_event-attr.type:\t%u",
> > + &perf_event_type) == 1)
> > + break;
> > + line = strtok(NULL, "\n");
> > }
> >
> > - if (printed == 0)
> > - printed += scnprintf(msg, size, "Possible processes:\n");
> > -
> > - printed += scnprintf(msg + printed, size - printed,
> > - "%s %s\n", proc_entry->d_name, path);
> > - break;
> > + /* Report the process which reserves the conflicted PMU. */
> > + /* If fdinfo does not contain PMU type, report it too. */
> > + if (perf_event_type == failed_attr->type ||
> > + perf_event_type == PERF_TYPE_MAX) {
> > + int cmdline_fd;
> > + ssize_t cmdline_size;
> > +
> > + scnprintf(path, sizeof(path),
> > + "%s/cmdline",
> > + proc_entry->d_name);
> > + cmdline_fd = openat(dirfd(proc_dir), path, O_RDONLY);
> > + if (cmdline_fd == -1)
> > + continue;
> > + cmdline_size = read(cmdline_fd, path, sizeof(path) - 1);
> > + close(cmdline_fd);
> > + if (cmdline_size < 0)
> > + continue;
> > + path[cmdline_size] = '\0';
> > + for (ssize_t i = 0; i < cmdline_size; i++) {
> > + if (path[i] == '\0')
> > + path[i] = ' ';
> > + }
> > +
> > + if (printed == 0)
> > + printed += scnprintf(
> > + msg, size,
> > + "Possible processes:\n");
> > +
> > + printed += scnprintf(msg + printed, size - printed,
> > + "%s %s\n", proc_entry->d_name, path);
> > + break;
> > + }
> > }
> > }
> > closedir(fd_dir);
> > @@ -3458,7 +3491,9 @@ int evsel__open_strerror(struct evsel *evsel, struct target *target,
> > msg, size,
> > "The PMU %s counters are busy and in use by another process.\n",
> > evsel->pmu ? evsel->pmu->name : "");
> > - return printed + dump_perf_event_processes(msg + printed, size - printed);
> > + return printed + dump_perf_event_processes(&evsel->core.attr,
> > + msg + printed,
> > + size - printed);
> > break;
> > case EINVAL:
> > if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
> > --
> > 2.47.0.163.g1226f6d8fa-goog
> >
Powered by blists - more mailing lists