+133
−28
Loading
perf/hw_breakpoint: Reduce contention with large number of tasks
While optimizing task_bp_pinned()'s runtime complexity to O(1) on
average helps reduce time spent in the critical section, we still suffer
due to serializing everything via 'nr_bp_mutex'. Indeed, a profile shows
that now contention is the biggest issue:
95.93% [kernel] [k] osq_lock
0.70% [kernel] [k] mutex_spin_on_owner
0.22% [kernel] [k] smp_cfm_core_cond
0.18% [kernel] [k] task_bp_pinned
0.18% [kernel] [k] rhashtable_jhash2
0.15% [kernel] [k] queued_spin_lock_slowpath
when running the breakpoint benchmark with (system with 256 CPUs):
| $> 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: 0.207 [sec]
|
| 108.267188 usecs/op
| 6929.100000 usecs/op/cpu
The main concern for synchronizing the breakpoint constraints data is
that a consistent snapshot of the per-CPU and per-task data is observed.
The access pattern is as follows:
1. If the target is a task: the task's pinned breakpoints are counted,
checked for space, and then appended to; only bp_cpuinfo::cpu_pinned
is used to check for conflicts with CPU-only breakpoints;
bp_cpuinfo::tsk_pinned are incremented/decremented, but otherwise
unused.
2. If the target is a CPU: bp_cpuinfo::cpu_pinned are counted, along
with bp_cpuinfo::tsk_pinned; after a successful check, cpu_pinned is
incremented. No per-task breakpoints are checked.
Since rhltable safely synchronizes insertions/deletions, we can allow
concurrency as follows:
1. If the target is a task: independent tasks may update and check the
constraints concurrently, but same-task target calls need to be
serialized; since bp_cpuinfo::tsk_pinned is only updated, but not
checked, these modifications can happen concurrently by switching
tsk_pinned to atomic_t.
2. If the target is a CPU: access to the per-CPU constraints needs to
be serialized with other CPU-target and task-target callers (to
stabilize the bp_cpuinfo::tsk_pinned snapshot).
We can allow the above concurrency by introducing a per-CPU constraints
data reader-writer lock (bp_cpuinfo_sem), and per-task mutexes (reuses
task_struct::perf_event_mutex):
1. If the target is a task: acquires perf_event_mutex, and acquires
bp_cpuinfo_sem as a reader. The choice of percpu-rwsem minimizes
contention in the presence of many read-lock but few write-lock
acquisitions: we assume many orders of magnitude more task target
breakpoints creations/destructions than CPU target breakpoints.
2. If the target is a CPU: acquires bp_cpuinfo_sem as a writer.
With these changes, contention with thousands of tasks is reduced to the
point where waiting on locking no longer dominates the profile:
| $> 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: 0.077 [sec]
|
| 40.201563 usecs/op
| 2572.900000 usecs/op/cpu
21.54% [kernel] [k] task_bp_pinned
20.18% [kernel] [k] rhashtable_jhash2
6.81% [kernel] [k] toggle_bp_slot
5.47% [kernel] [k] queued_spin_lock_slowpath
3.75% [kernel] [k] smp_cfm_core_cond
3.48% [kernel] [k] bcmp
On this particular setup that's a speedup of 2.7x.
We're also getting closer to the theoretical ideal performance through
optimizations in hw_breakpoint.c -- constraints accounting disabled:
| 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]
|
| 35.286458 usecs/op
| 2258.333333 usecs/op/cpu
Which means the current implementation is ~12% slower than the
theoretical ideal.
For reference, performance without any breakpoints:
| $> bench -r 30 breakpoint thread -b 0 -p 64 -t 64
| # Running 'breakpoint/thread' benchmark:
| # Created/joined 30 threads with 0 breakpoints and 64 parallelism
| Total time: 0.060 [sec]
|
| 31.365625 usecs/op
| 2007.400000 usecs/op/cpu
On a system with 256 CPUs, the theoretical ideal is only ~12% slower
than no breakpoints at all; the current implementation is ~28% slower.
Signed-off-by: 
Marco Elver <elver@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20220829124719.675715-12-elver@google.com