On 10/09/2024 12.46, Yunsheng Lin wrote:
On 2024/9/10 1:28, Mina Almasry wrote:
On Mon, Sep 9, 2024 at 2:25 AM Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
The testing is done by ensuring that the page allocated from
the page_pool instance is pushed into a ptr_ring instance in
a kthread/napi binded to a specified cpu, and a kthread/napi
binded to a specified cpu will pop the page from the ptr_ring
and free it back to the page_pool.
Signed-off-by: Yunsheng Lin <linyunsheng@xxxxxxxxxx>
It seems this test is has a correctness part and a performance part.
For the performance test, Jesper has out of tree tests for the
page_pool:
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_simple.c
I have these rebased on top of net-next and use them to verify devmem
& memory-provider performance:
https://github.com/mina/linux/commit/07fd1c04591395d15d83c07298b4d37f6b56157f
Yes, I used that testing ko too when adding frag API support for
page_pool.
The main issue I remembered was that it only support x86:(
Yes, because I've added ASM code for reading TSC counter in a very
precise manor. Given we run many iterations, then I don't think we
need this precise reading. I guess it can simply be replaced with
get_cycles() or get_cycles64(). Then it should work on all archs.
The code already supports wall-clock time via ktime_get() (specifically
ktime_get_real_ts64()).
My preference here (for the performance part) is to upstream the
out-of-tree tests that Jesper (and probably others) are using, rather
than adding a new performance test that is not as battle-hardened.
I looked through the out-of-tree tests again, it seems we can take the
best of them.
For Jesper' ko:
It seems we can do prefill as something that pp_fill_ptr_ring() does
in bench_page_pool_simple.c to avoid the noise from the page allocator.
For the ko in this patch:
It uses NAPI instead of tasklet mimicking the NAPI context, support
PP_FLAG_DMA_MAP flag testing, and return '-EAGAIN' in module_init()
to use perf stat for collecting and calculating performance data.
My bench don't return minus-number on module load, because I used perf
record, and to see symbols decoded with perf report, I needed the module
to be loaded.
I started on reading the PMU counters[1] around the bench loop, it works
if enabling PMU counters yourself/manually, but I never finished that work.
[1]
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/include/linux/time_bench.h#L195-L209
Is there other testcase or better practicing that we can learn from
Jesper' out of tree ko?
I created a time_bench.c [2] module that other modules [3] can use to
easier reuse the benchmarking code in other modules.
[2]
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/time_bench.c
[3]
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_simple.c
--Jesper
