| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Profiling shows that calling nr_possible_cpus() in objpool_pop() takes
a noticeable amount of CPU (when profiled on 80-core machine), as we
need to recalculate number of set bits in a CPU bit mask. This number
can't change, so there is no point in paying the price for recalculating
it. As such, cache this value in struct objpool_head and use it in
objpool_pop().
On the other hand, cached pool->nr_cpus isn't necessary, as it's not
used in hot path and is also a pretty trivial value to retrieve. So drop
pool->nr_cpus in favor of using nr_cpu_ids everywhere. This way the size
of struct objpool_head remains the same, which is a nice bonus.
Same BPF selftests benchmarks were used to evaluate the effect. Using
changes in previous patch (inlining of objpool_pop/objpool_push) as
baseline, here are the differences:
BASELINE
========
kretprobe : 9.937 ± 0.174M/s
kretprobe-multi: 10.440 ± 0.108M/s
AFTER
=====
kretprobe : 10.106 ± 0.120M/s (+1.7%)
kretprobe-multi: 10.515 ± 0.180M/s (+0.7%)
Link: https://lore.kernel.org/all/20240424215214.3956041-3-andrii@kernel.org/
Cc: Matt (Qiang) Wu <wuqiang.matt@bytedance.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
|
|
objpool_push() and objpool_pop() are very performance-critical functions
and can be called very frequently in kretprobe triggering path.
As such, it makes sense to allow compiler to inline them completely to
eliminate function calls overhead. Luckily, their logic is quite well
isolated and doesn't have any sprawling dependencies.
This patch moves both objpool_push() and objpool_pop() into
include/linux/objpool.h and marks them as static inline functions,
enabling inlining. To avoid anyone using internal helpers
(objpool_try_get_slot, objpool_try_add_slot), rename them to use leading
underscores.
We used kretprobe microbenchmark from BPF selftests (bench trig-kprobe
and trig-kprobe-multi benchmarks) running no-op BPF kretprobe/kretprobe.multi
programs in a tight loop to evaluate the effect. BPF own overhead in
this case is minimal and it mostly stresses the rest of in-kernel
kretprobe infrastructure overhead. Results are in millions of calls per
second. This is not super scientific, but shows the trend nevertheless.
BEFORE
======
kretprobe : 9.794 ± 0.086M/s
kretprobe-multi: 10.219 ± 0.032M/s
AFTER
=====
kretprobe : 9.937 ± 0.174M/s (+1.5%)
kretprobe-multi: 10.440 ± 0.108M/s (+2.2%)
Link: https://lore.kernel.org/all/20240424215214.3956041-2-andrii@kernel.org/
Cc: Matt (Qiang) Wu <wuqiang.matt@bytedance.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
|
|
objpool is a scalable implementation of high performance queue for
object allocation and reclamation, such as kretprobe instances.
With leveraging percpu ring-array to mitigate hot spots of memory
contention, it delivers near-linear scalability for high parallel
scenarios. The objpool is best suited for the following cases:
1) Memory allocation or reclamation are prohibited or too expensive
2) Consumers are of different priorities, such as irqs and threads
Limitations:
1) Maximum objects (capacity) is fixed after objpool creation
2) All pre-allocated objects are managed in percpu ring array,
which consumes more memory than linked lists
Link: https://lore.kernel.org/all/20231017135654.82270-2-wuqiang.matt@bytedance.com/
Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
|
