linux-kernel - Re: [PATCH v1 2/2] perf annotation: Switch lock from a mutex to a sharded

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <CAP-5=fXnqZxWVbgMcainpTPc6mSHjUu3y1tELjdRk2zfLM4bOw@mail.gmail.com>
Date:   Wed, 14 Jun 2023 18:49:40 -0700
From:   Ian Rogers <irogers@...gle.com>
To:     Namhyung Kim <namhyung@...nel.org>
Cc:     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@...nel.org>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Yuan Can <yuancan@...wei.com>,
        Kan Liang <kan.liang@...ux.intel.com>,
        Masami Hiramatsu <mhiramat@...nel.org>,
        Huacai Chen <chenhuacai@...nel.org>,
        Andres Freund <andres@...razel.de>,
        linux-kernel@...r.kernel.org, linux-perf-users@...r.kernel.org
Subject: Re: [PATCH v1 2/2] perf annotation: Switch lock from a mutex to a sharded_mutex

On Wed, Jun 14, 2023 at 5:34 PM Namhyung Kim <namhyung@...nel.org> wrote:
>
> Hi Ian,
>
> On Sun, Jun 11, 2023 at 12:28 AM Ian Rogers <irogers@...gle.com> wrote:
> >
> > Remove the "struct mutex lock" variable from annotation that is
> > allocated per symbol. This removes in the region of 40 bytes per
> > symbol allocation. Use a sharded mutex where the number of shards is
> > set to the number of CPUs. Assuming good hashing of the annotation
> > (done based on the pointer), this means in order to contend there
> > needs to be more threads than CPUs, which is not currently true in any
> > perf command. Were contention an issue it is straightforward to
> > increase the number of shards in the mutex.
> >
> > On my Debian/glibc based machine, this reduces the size of struct
> > annotation from 136 bytes to 96 bytes, or nearly 30%.
>
> That's quite a good improvement given the number of symbols
> we can have in a report session!
>
> >
> > Signed-off-by: Ian Rogers <irogers@...gle.com>
> > ---
>
> [SNIP]
> > @@ -1291,17 +1292,64 @@ int disasm_line__scnprintf(struct disasm_line *dl, char *bf, size_t size, bool r
> >         return ins__scnprintf(&dl->ins, bf, size, &dl->ops, max_ins_name);
> >  }
> >
> > -void annotation__init(struct annotation *notes)
> > +void annotation__exit(struct annotation *notes)
> >  {
> > -       mutex_init(&notes->lock);
> > +       annotated_source__delete(notes->src);
> >  }
> >
> > -void annotation__exit(struct annotation *notes)
> > +static struct sharded_mutex *sharded_mutex;
> > +
> > +static void annotation__init_sharded_mutex(void)
> >  {
> > -       annotated_source__delete(notes->src);
> > -       mutex_destroy(&notes->lock);
> > +       /* As many mutexes as there are CPUs. */
> > +       sharded_mutex = sharded_mutex__new(cpu__max_present_cpu().cpu);
> > +}
> > +
> > +static size_t annotation__hash(const struct annotation *notes)
> > +{
> > +       return ((size_t)notes) >> 4;
>
> But I'm afraid it might create more contention depending on the
> malloc implementation.  If it always returns 128-byte (or 256)
> aligned memory for this struct then it could always collide in the
> slot 0 if the number of CPU is 8 or less, right?

Right, I think we can use a secondary hash and hashmap.h has one lying
around for us:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/hashmap.h?h=tmp.perf-tools-next#n15
it will mean that the sharded locks will need to be a power of 2
capacity. I'll work on a v2. Fwiw, the hash of a pointer for
collection like those in abseil is just the pointer value, so I'll
drop the shift to remove the low-bits once I'm using a more expensive
hash function.

Thanks,
Ian

> Thanks,
> Namhyung
>
>
> >  }
> >
> > +static struct mutex *annotation__get_mutex(const struct annotation *notes)
> > +{
> > +       static pthread_once_t once = PTHREAD_ONCE_INIT;
> > +
> > +       pthread_once(&once, annotation__init_sharded_mutex);
> > +       if (!sharded_mutex)
> > +               return NULL;
> > +
> > +       return sharded_mutex__get_mutex(sharded_mutex, annotation__hash(notes));
> > +}
> > +
> > +void annotation__lock(struct annotation *notes)
> > +       NO_THREAD_SAFETY_ANALYSIS
> > +{
> > +       struct mutex *mutex = annotation__get_mutex(notes);
> > +
> > +       if (mutex)
> > +               mutex_lock(mutex);
> > +}
> > +
> > +void annotation__unlock(struct annotation *notes)
> > +       NO_THREAD_SAFETY_ANALYSIS
> > +{
> > +       struct mutex *mutex = annotation__get_mutex(notes);
> > +
> > +       if (mutex)
> > +               mutex_unlock(mutex);
> > +}
> > +
> > +bool annotation__trylock(struct annotation *notes)
> > +{
> > +       struct mutex *mutex = annotation__get_mutex(notes);
> > +
> > +       if (!mutex)
> > +               return false;
> > +
> > +       return mutex_trylock(mutex);
> > +}
> > +
> > +
> >  static void annotation_line__add(struct annotation_line *al, struct list_head *head)
> >  {
> >         list_add_tail(&al->node, head);