Message-ID: <20220829124719.675715-1-elver@google.com>
Date:   Mon, 29 Aug 2022 14:47:05 +0200
From:   Marco Elver <elver@...gle.com>
To:     elver@...gle.com, Peter Zijlstra <peterz@...radead.org>,
        Frederic Weisbecker <frederic@...nel.org>,
        Ingo Molnar <mingo@...nel.org>
Cc:     Thomas Gleixner <tglx@...utronix.de>,
        Arnaldo Carvalho de Melo <acme@...nel.org>,
        Mark Rutland <mark.rutland@....com>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Jiri Olsa <jolsa@...hat.com>,
        Namhyung Kim <namhyung@...nel.org>,
        Dmitry Vyukov <dvyukov@...gle.com>,
        Michael Ellerman <mpe@...erman.id.au>,
        linuxppc-dev@...ts.ozlabs.org, linux-perf-users@...r.kernel.org,
        x86@...nel.org, linux-sh@...r.kernel.org,
        kasan-dev@...glegroups.com, linux-kernel@...r.kernel.org
Subject: [PATCH v4 00/14] perf/hw_breakpoint: Optimize for thousands of tasks

The hw_breakpoint subsystem's code has seen little change in over 10
years. In that time, systems with >100s of CPUs have become common,
along with improvements to the perf subsystem: using breakpoints on
thousands of concurrent tasks should be a supported usecase.

The breakpoint constraints accounting algorithm is the major bottleneck
in doing so:

  1. toggle_bp_slot() and fetch_bp_busy_slots() are O(#cpus * #tasks):
     Both iterate through all CPUs and call task_bp_pinned(), which is
     O(#tasks).

  2. Everything is serialized on a global mutex, 'nr_bp_mutex'.

The series progresses with the simpler optimizations and finishes with
the more complex optimizations:

 1. We first optimize task_bp_pinned() to only take O(1) on average.

 2. Rework synchronization to allow concurrency when checking and
    updating breakpoint constraints for tasks.

 3. Eliminate the O(#cpus) loops in the CPU-independent case.

Along the way, smaller micro-optimizations and cleanups are done as they
seemed obvious when staring at the code (but likely insignificant).

The result is (on a system with 256 CPUs) that we go from:

 | $> perf bench -r 30 breakpoint thread -b 4 -p 64 -t 64
	 	[ ^ more aggressive benchmark parameters took too long ]
 | # Running 'breakpoint/thread' benchmark:
 | # Created/joined 30 threads with 4 breakpoints and 64 parallelism
 |      Total time: 236.418 [sec]
 |
 |   123134.794271 usecs/op
 |  7880626.833333 usecs/op/cpu

... to the following with all optimizations:

 | $> perf bench -r 30 breakpoint thread -b 4 -p 64 -t 64
 | # Running 'breakpoint/thread' benchmark:
 | # Created/joined 30 threads with 4 breakpoints and 64 parallelism
 |      Total time: 0.067 [sec]
 |
 |       35.292187 usecs/op
 |     2258.700000 usecs/op/cpu

On the used test system, that's an effective speedup of ~3490x per op.

Which is on par with the theoretical ideal performance through
optimizations in hw_breakpoint.c (constraints accounting disabled), and
only 12% slower than no breakpoints at all.

Changelog
---------

v4:
* Fix percpu_is_read_locked(): Due to spurious read_count increments in
  __percpu_down_read_trylock() if sem->block != 0, check that
  !sem->block (reported by Peter).
* Apply Reviewed/Acked-by.

v3: https://lkml.kernel.org/r/20220704150514.48816-1-elver@google.com
* Fix typos.
* Introduce hw_breakpoint_is_used() for the test.
* Add WARN_ON in bp_blots_histogram_add().
* Don't use raw_smp_processor_id() in test.
* Apply Acked-by/Reviewed-by given in v2 for mostly unchanged patches.


v2: https://lkml.kernel.org/r/20220628095833.2579903-1-elver@google.com
 * Add KUnit test suite.
 * Remove struct bp_busy_slots and simplify functions.
 * Add "powerpc/hw_breakpoint: Avoid relying on caller synchronization".
 * Add "locking/percpu-rwsem: Add percpu_is_write_locked() and percpu_is_read_locked()".
 * Use percpu-rwsem instead of rwlock.
 * Use task_struct::perf_event_mutex instead of sharded mutex.
 * Drop v1 "perf/hw_breakpoint: Optimize task_bp_pinned() if CPU-independent".
 * Add "perf/hw_breakpoint: Introduce bp_slots_histogram".
 * Add "perf/hw_breakpoint: Optimize max_bp_pinned_slots() for CPU-independent task targets".
 * Add "perf/hw_breakpoint: Optimize toggle_bp_slot() for CPU-independent task targets".
 * Apply Acked-by/Reviewed-by given in v1 for unchanged patches.
==> Speedup of ~3490x (vs. ~3315x in v1).

v1: https://lore.kernel.org/all/20220609113046.780504-1-elver@google.com/

Marco Elver (14):
  perf/hw_breakpoint: Add KUnit test for constraints accounting
  perf/hw_breakpoint: Provide hw_breakpoint_is_used() and use in test
  perf/hw_breakpoint: Clean up headers
  perf/hw_breakpoint: Optimize list of per-task breakpoints
  perf/hw_breakpoint: Mark data __ro_after_init
  perf/hw_breakpoint: Optimize constant number of breakpoint slots
  perf/hw_breakpoint: Make hw_breakpoint_weight() inlinable
  perf/hw_breakpoint: Remove useless code related to flexible
    breakpoints
  powerpc/hw_breakpoint: Avoid relying on caller synchronization
  locking/percpu-rwsem: Add percpu_is_write_locked() and
    percpu_is_read_locked()
  perf/hw_breakpoint: Reduce contention with large number of tasks
  perf/hw_breakpoint: Introduce bp_slots_histogram
  perf/hw_breakpoint: Optimize max_bp_pinned_slots() for CPU-independent
    task targets
  perf/hw_breakpoint: Optimize toggle_bp_slot() for CPU-independent task
    targets

 arch/powerpc/kernel/hw_breakpoint.c  |  53 ++-
 arch/sh/include/asm/hw_breakpoint.h  |   5 +-
 arch/x86/include/asm/hw_breakpoint.h |   5 +-
 include/linux/hw_breakpoint.h        |   4 +-
 include/linux/percpu-rwsem.h         |   6 +
 include/linux/perf_event.h           |   3 +-
 kernel/events/Makefile               |   1 +
 kernel/events/hw_breakpoint.c        | 638 ++++++++++++++++++++-------
 kernel/events/hw_breakpoint_test.c   | 333 ++++++++++++++
 kernel/locking/percpu-rwsem.c        |   6 +
 lib/Kconfig.debug                    |  10 +
 11 files changed, 885 insertions(+), 179 deletions(-)
 create mode 100644 kernel/events/hw_breakpoint_test.c

-- 
2.37.2.672.g94769d06f0-goog

Powered by blists - more mailing lists