[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <CAJpZYjVDVQ7ytJjBBjjK4On7bOdRN+8Q3wGeen3EeSmevyb=1Q@mail.gmail.com>
Date: Thu, 21 Nov 2024 16:34:27 -0800
From: Chun-Tse Shao <ctshao@...gle.com>
To: Namhyung Kim <namhyung@...nel.org>
Cc: linux-kernel@...r.kernel.org, Peter Zijlstra <peterz@...radead.org>,
Ingo Molnar <mingo@...hat.com>, Arnaldo Carvalho de Melo <acme@...nel.org>, Mark Rutland <mark.rutland@....com>,
Alexander Shishkin <alexander.shishkin@...ux.intel.com>, Jiri Olsa <jolsa@...hat.com>
Subject: Re: [PATCH] perf lock: Fix parse_lock_type which only retrieve one
lock flag
Thank you Namhyung! please check my patch v2:
https://lore.kernel.org/all/20241122002752.4179915-1-ctshao@google.com
On Wed, Nov 20, 2024 at 11:18 PM Namhyung Kim <namhyung@...nel.org> wrote:
>
> Hi CT,
>
> On Thu, Nov 21, 2024 at 04:45:51AM +0000, Chun-Tse Shao wrote:
> > `parse_lock_type` can only add the first lock flag in `lock_type_table`
> > given input `str`. For example, for `Y rwlock`, it only adds `rwlock:R`
> > into this perf session. Another example is for `-Y mutex`, it only adds
> > the mutex without `LCB_F_SPIN` flag. The patch fixes this issue, makes
> > sure both `rwlock:R` and `rwlock:W` will be added with `-Y rwlock`, and
> > so on.
>
> Thanks for working on this. I think I broke it at some point.
>
> >
> > Testing:
> > $ ./perf lock con -ab -Y rwlock,mutex -- perf bench sched pipe
> > # Running 'sched/pipe' benchmark:
> > # Executed 1000000 pipe operations between two processes
> >
> > Total time: 7.870 [sec]
> >
> > 7.870224 usecs/op
> > 127061 ops/sec
> > contended total wait max wait avg wait type caller
> >
> > 122 1.01 ms 18.73 us 8.32 us mutex pipe_read+0x57
> > 33 149.06 us 8.88 us 4.52 us mutex pipe_write+0x50
> > 5 56.06 us 15.07 us 11.21 us mutex do_epoll_wait+0x24d
> > 2 26.43 us 14.62 us 13.22 us mutex do_epoll_wait+0x24d
> > 2 21.51 us 15.68 us 10.76 us mutex pipe_read+0x282
> > 2 20.57 us 10.74 us 10.29 us mutex do_epoll_ctl+0x3d4
> > 1 8.58 us 8.58 us 8.58 us rwlock:W do_epoll_wait+0x255
> > 1 8.38 us 8.38 us 8.38 us rwlock:W do_epoll_ctl+0xb65
> > 1 8.11 us 8.11 us 8.11 us rwlock:R ep_poll_callback+0x35
> > 1 5.49 us 5.49 us 5.49 us mutex do_epoll_ctl+0x6c1
>
> It should have
>
> Fixes: d783ea8f62c4 ("perf lock contention: Simplify parse_lock_type()")
Done.
>
> >
> > Signed-off-by: Chun-Tse Shao <ctshao@...gle.com>
> > ---
> > tools/perf/builtin-lock.c | 118 +++++++++++++++++++++++---------------
> > 1 file changed, 71 insertions(+), 47 deletions(-)
> >
> > diff --git a/tools/perf/builtin-lock.c b/tools/perf/builtin-lock.c
> > index 062e2b56a2ab..2692b8e9eb9a 100644
> > --- a/tools/perf/builtin-lock.c
> > +++ b/tools/perf/builtin-lock.c
> > @@ -1575,61 +1575,57 @@ static void sort_result(void)
> >
> > static const struct {
> > unsigned int flags;
> > - const char *str;
> > + /* Name of the lock, it is unique and 1-1 mapping to flags */
> > const char *name;
> > + /*
> > + * Name of the group this lock belongs to.
> > + * For example, both rwlock:R and rwlock:W belong to rwlock.
> > + * This is used for reverse parsing while user specify the group name (ex. mutex/rwlock),
> > + * And for symbol name in LOCK_AGGR_ADDR mode.
> > + */
> > + const char *affiliated_group_name;
> > + /*
> > + * This is used for caller type in LOCK_AGGR_CALLER mode.
> > + */
> > + const char *caller_type;
>
> I admit that the 'str' and 'name' are not good but I'm not sure if we
> need those 3 different names because they are mostly the same. Maybe we
> have rename the original fields to 'name' and 'group'?
Done.
>
>
> > } lock_type_table[] = {
> > - { 0, "semaphore", "semaphore" },
> > - { LCB_F_SPIN, "spinlock", "spinlock" },
> > - { LCB_F_SPIN | LCB_F_READ, "rwlock:R", "rwlock" },
> > - { LCB_F_SPIN | LCB_F_WRITE, "rwlock:W", "rwlock" },
> > - { LCB_F_READ, "rwsem:R", "rwsem" },
> > - { LCB_F_WRITE, "rwsem:W", "rwsem" },
> > - { LCB_F_RT, "rt-mutex", "rt-mutex" },
> > - { LCB_F_RT | LCB_F_READ, "rwlock-rt:R", "rwlock-rt" },
> > - { LCB_F_RT | LCB_F_WRITE, "rwlock-rt:W", "rwlock-rt" },
> > - { LCB_F_PERCPU | LCB_F_READ, "pcpu-sem:R", "percpu-rwsem" },
> > - { LCB_F_PERCPU | LCB_F_WRITE, "pcpu-sem:W", "percpu-rwsem" },
> > - { LCB_F_MUTEX, "mutex", "mutex" },
> > - { LCB_F_MUTEX | LCB_F_SPIN, "mutex", "mutex" },
> > - /* alias for get_type_flag() */
> > - { LCB_F_MUTEX | LCB_F_SPIN, "mutex-spin", "mutex" },
> > + { 0, "semaphore", "semaphore", "semaphore" },
> > + { LCB_F_SPIN, "spinlock", "spinlock", "spinlock" },
> > + { LCB_F_SPIN | LCB_F_READ, "rwlock:R", "rwlock", "rwlock:R" },
> > + { LCB_F_SPIN | LCB_F_WRITE, "rwlock:W", "rwlock", "rwlock:W" },
> > + { LCB_F_READ, "rwsem:R", "rwsem", "rwsem:R" },
> > + { LCB_F_WRITE, "rwsem:W", "rwsem", "rwsem:W" },
> > + { LCB_F_RT, "rt-mutex", "rt-mutex", "rt-mutex" },
> > + { LCB_F_RT | LCB_F_READ, "rwlock-rt:R", "rwlock-rt", "rwlock-rt:R" },
> > + { LCB_F_RT | LCB_F_WRITE, "rwlock-rt:W", "rwlock-rt", "rwlock-rt:W" },
> > + { LCB_F_PERCPU | LCB_F_READ, "pcpu-sem:R", "percpu-rwsem", "pcpu-sem:R" },
> > + { LCB_F_PERCPU | LCB_F_WRITE, "pcpu-sem:W", "percpu-rwsem", "pcpu-sem:W" },
>
> I've realized that it doesn't match with the documentation. It should
> match with 'pcpu-sem'. Maybe we can replace the 'pcpu-sem' string in
> the parse_lock_type() to 'percpu-rwsem'.
Done.
>
> > + { LCB_F_MUTEX, "mutex-nospin", "mutex", "mutex" },
> > + { LCB_F_MUTEX | LCB_F_SPIN, "mutex-spin", "mutex", "mutex" },
>
> These are not for users and we don't track mutex-spinning precisely
> (yet). So I think we can name both of them simply as "mutex".
>
> { LCB_F_MUTEX, "mutex", "mutex" },
> { LCB_F_MUTEX | LCB_F_SPIN "mutex", "mutex" },
>
> And remove the last alias entry.
Done.
>
> > };
> >
> > -static const char *get_type_str(unsigned int flags)
> > +static const char *get_lock_caller_type(unsigned int flags)
> > {
> > flags &= LCB_F_MAX_FLAGS - 1;
> >
> > for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
> > if (lock_type_table[i].flags == flags)
> > - return lock_type_table[i].str;
> > + return lock_type_table[i].caller_type;
> > }
> > return "unknown";
> > }
> >
> > -static const char *get_type_name(unsigned int flags)
> > +static const char *get_lock_affiliated_group_name(unsigned int flags)
> > {
> > flags &= LCB_F_MAX_FLAGS - 1;
> >
> > for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
> > if (lock_type_table[i].flags == flags)
> > - return lock_type_table[i].name;
> > + return lock_type_table[i].affiliated_group_name;
> > }
> > return "unknown";
> > }
> >
> > -static unsigned int get_type_flag(const char *str)
> > -{
> > - for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
> > - if (!strcmp(lock_type_table[i].name, str))
> > - return lock_type_table[i].flags;
> > - }
> > - for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
> > - if (!strcmp(lock_type_table[i].str, str))
> > - return lock_type_table[i].flags;
> > - }
> > - return UINT_MAX;
> > -}
> > -
> > static void lock_filter_finish(void)
> > {
> > zfree(&filters.types);
> > @@ -1732,7 +1728,8 @@ static void print_lock_stat_stdio(struct lock_contention *con, struct lock_stat
> >
> > switch (aggr_mode) {
> > case LOCK_AGGR_CALLER:
> > - fprintf(lock_output, " %10s %s\n", get_type_str(st->flags), st->name);
> > + fprintf(lock_output, " %10s %s\n",
> > + get_lock_caller_type(st->flags), st->name);
> > break;
> > case LOCK_AGGR_TASK:
> > pid = st->addr;
> > @@ -1742,7 +1739,7 @@ static void print_lock_stat_stdio(struct lock_contention *con, struct lock_stat
> > break;
> > case LOCK_AGGR_ADDR:
> > fprintf(lock_output, " %016llx %s (%s)\n", (unsigned long long)st->addr,
> > - st->name, get_type_name(st->flags));
> > + st->name, get_lock_affiliated_group_name(st->flags));
> > break;
> > case LOCK_AGGR_CGROUP:
> > fprintf(lock_output, " %s\n", st->name);
> > @@ -1783,7 +1780,8 @@ static void print_lock_stat_csv(struct lock_contention *con, struct lock_stat *s
> >
> > switch (aggr_mode) {
> > case LOCK_AGGR_CALLER:
> > - fprintf(lock_output, "%s%s %s", get_type_str(st->flags), sep, st->name);
> > + fprintf(lock_output, "%s%s %s",
> > + get_lock_caller_type(st->flags), sep, st->name);
> > if (verbose <= 0)
> > fprintf(lock_output, "\n");
> > break;
> > @@ -1795,7 +1793,7 @@ static void print_lock_stat_csv(struct lock_contention *con, struct lock_stat *s
> > break;
> > case LOCK_AGGR_ADDR:
> > fprintf(lock_output, "%llx%s %s%s %s\n", (unsigned long long)st->addr, sep,
> > - st->name, sep, get_type_name(st->flags));
> > + st->name, sep, get_lock_affiliated_group_name(st->flags));
> > break;
> > case LOCK_AGGR_CGROUP:
> > fprintf(lock_output, "%s\n",st->name);
> > @@ -2333,6 +2331,7 @@ static int parse_max_stack(const struct option *opt, const char *str,
> > return 0;
> > }
> >
> > +
> > static bool add_lock_type(unsigned int flags)
> > {
> > unsigned int *tmp;
> > @@ -2350,29 +2349,54 @@ static int parse_lock_type(const struct option *opt __maybe_unused, const char *
> > int unset __maybe_unused)
> > {
> > char *s, *tmp, *tok;
> > - int ret = 0;
> >
> > s = strdup(str);
> > if (s == NULL)
> > return -1;
> >
> > for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
> > - unsigned int flags = get_type_flag(tok);
> > + bool found = false;
> >
> > - if (flags == -1U) {
> > - pr_err("Unknown lock flags: %s\n", tok);
> > - ret = -1;
> > - break;
> > + /* Traverse lock name first. */
> > + for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
> > + if (!strcmp(lock_type_table[i].name, tok)) {
> > + if (add_lock_type(lock_type_table[i].flags)) {
> > + found = true;
> > + break;
> > + }
> > + pr_err("Failed to alloc lock: %s\n", tok);
> > + free(s);
> > + return -1;
> > + }
> > }
> > + if (found)
> > + continue;
>
> Then I think the logic can be like this.
>
> if (strchr(tok, ':')) {
> flags = get_type_flag(tok);
> add_lock_type(flag);
> continue;
> }
>
> if (!strcmp(tok, "pcpu-sem"))
> tok = "percpu-rwsem";
>
> for (i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
> if (!strcmp(lock_type_table[i].group, tok))
> add_lock_type(...);
> }
Done. I think `get_type_flag` would be unnecessary since it's only called once.
>
> Thanks,
> Namhyung
>
> >
> > - if (!add_lock_type(flags)) {
> > - ret = -1;
> > - break;
> > + /*
> > + * If `tok` can not be found in lock name, look up the lock affiliated group
> > + * instead. A group would contain more than one lock flag.
> > + */
> > + for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
> > + if (!strcmp(lock_type_table[i].affiliated_group_name, tok)) {
> > + if (add_lock_type(lock_type_table[i].flags)) {
> > + found = true;
> > + } else {
> > + pr_err("Failed to alloc lock: %s\n", tok);
> > + free(s);
> > + return -1;
> > + }
> > + }
> > + }
> > +
> > + if (!found) {
> > + pr_err("Unknown lock flags: %s\n", tok);
> > + free(s);
> > + return -1;
> > }
> > }
> >
> > free(s);
> > - return ret;
> > + return 0;
> > }
> >
> > static bool add_lock_addr(unsigned long addr)
> > --
> > 2.47.0.371.ga323438b13-goog
> >
Powered by blists - more mailing lists