On 16/01/2025 13.52, Yunsheng Lin wrote:
On 2025/1/16 1:40, Jesper Dangaard Brouer wrote:
On 15/01/2025 12.33, Yunsheng Lin wrote:
On 2025/1/14 22:31, Jesper Dangaard Brouer wrote:
On 10/01/2025 14.06, Yunsheng Lin wrote:
This patchset fix a possible time window problem for page_pool and
the dma API misuse problem as mentioned in [1], and try to avoid the
overhead of the fixing using some optimization.
From the below performance data, the overhead is not so obvious
due to performance variations for time_bench_page_pool01_fast_path()
and time_bench_page_pool02_ptr_ring, and there is about 20ns overhead
for time_bench_page_pool03_slow() for fixing the bug.
My benchmarking on x86_64 CPUs looks significantly different.
- CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz
Benchmark (bench_page_pool_simple) results from before and after patchset:
| Test name | Cycles | | |Nanosec | | | % |
| (tasklet_*)| Before | After |diff| Before | After | diff | change |
|------------+--------+-------+----+--------+--------+-------+--------|
| fast_path | 19 | 24 | 5| 5.399 | 6.928 | 1.529 | 28.3 |
| ptr_ring | 54 | 79 | 25| 15.090 | 21.976 | 6.886 | 45.6 |
| slow | 238 | 299 | 61| 66.134 | 83.298 |17.164 | 26.0 |
#+TBLFM: $4=$3-$2::$7=$6-$5::$8=(($7/$5)*100);%.1f
My above testing show a clear performance regressions across three
different page_pool operating modes.
I retested it on arm64 server patch by patch as the raw performance
data in the attachment, it seems the result seemed similar as before.
Before this patchset:
fast_path ptr_ring slow
1. 31.171 ns 60.980 ns 164.917 ns
2. 28.824 ns 60.891 ns 170.241 ns
3. 14.236 ns 60.583 ns 164.355 ns
With patch 1-4:
4. 31.443 ns 53.242 ns 210.148 ns
5. 31.406 ns 53.270 ns 210.189 ns
With patch 1-5:
6. 26.163 ns 53.781 ns 189.450 ns
7. 26.189 ns 53.798 ns 189.466 ns
With patch 1-8:
8. 28.108 ns 68.199 ns 202.516 ns
9. 16.128 ns 55.904 ns 202.711 ns
I am not able to get hold of a x86 server yet, I might be able
to get one during weekend.
Theoretically, patch 1-4 or 1-5 should not have much performance
impact for fast_path and ptr_ring except for the rcu_lock mentioned
in page_pool_napi_local(), so it would be good if patch 1-5 is also
tested in your testlab with the rcu_lock removing in
page_pool_napi_local().
What are you saying?
- (1) test patch 1-5
- or (2) test patch 1-5 but revert patch 2 with page_pool_napi_local()
patch 1-5 with below applied.
--- a/net/core/page_pool.c
+++ b/net/core/page_pool.c
@@ -1207,10 +1207,8 @@ static bool page_pool_napi_local(const struct page_pool *pool)
/* Synchronizated with page_pool_destory() to avoid use-after-free
* for 'napi'.
*/
- rcu_read_lock();
napi = READ_ONCE(pool->p.napi);
napi_local = napi && READ_ONCE(napi->list_owner) == cpuid;
- rcu_read_unlock();
return napi_local;
}
Benchmark (bench_page_pool_simple) results from before and after
patchset with patches 1-5m and rcu lock removal as requested.
| Test name |Cycles | 1-5 | | Nanosec | 1-5 | | % |
| (tasklet_*)|Before | After |diff| Before | After | diff | change |
|------------+-------+-------+----+---------+--------+--------+--------|
| fast_path | 19 | 19 | 0| 5.399 | 5.492 | 0.093 | 1.7 |
| ptr_ring | 54 | 57 | 3| 15.090 | 15.849 | 0.759 | 5.0 |
| slow | 238 | 284 | 46| 66.134 | 78.909 | 12.775 | 19.3 |
#+TBLFM: $4=$3-$2::$7=$6-$5::$8=(($7/$5)*100);%.1f
This test with patches 1-5 looks much better regarding performance.
--Jesper
https://github.com/xdp-project/xdp-project/blob/main/areas/mem/page_pool07_bench_DMA_fix.org#e5-1650-pp01-dma-fix-v7-p1-5
Kernel:
- 6.13.0-rc6-pp01-DMA-fix-v7-p1-5+ #5 SMP PREEMPT_DYNAMIC Thu Jan 16
18:06:53 CET 2025 x86_64 GNU/Linux
Machine: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz
modprobe bench_page_pool_simple loops=100000000
Raw data:
[ 187.309423] bench_page_pool_simple:
time_bench_page_pool01_fast_path(): Cannot use page_pool fast-path
[ 187.872849] time_bench: Type:no-softirq-page_pool01 Per elem: 19
cycles(tsc) 5.539 ns (step:0) - (measurement period time:0.553906443 sec
time_interval:553906443) - (invoke count:100000000 tsc_interval:1994123064)
[ 187.892023] bench_page_pool_simple:
time_bench_page_pool02_ptr_ring(): Cannot use page_pool fast-path
[ 189.611070] time_bench: Type:no-softirq-page_pool02 Per elem: 61
cycles(tsc) 17.095 ns (step:0) - (measurement period time:1.709580367
sec time_interval:1709580367) - (invoke count:100000000
tsc_interval:6154679394)
[ 189.630414] bench_page_pool_simple: time_bench_page_pool03_slow():
Cannot use page_pool fast-path
[ 197.222387] time_bench: Type:no-softirq-page_pool03 Per elem: 272
cycles(tsc) 75.826 ns (step:0) - (measurement period time:7.582681388
sec time_interval:7582681388) - (invoke count:100000000
tsc_interval:27298499214)
[ 197.241926] bench_page_pool_simple: pp_tasklet_handler():
in_serving_softirq fast-path
[ 197.249968] bench_page_pool_simple:
time_bench_page_pool01_fast_path(): in_serving_softirq fast-path
[ 197.808470] time_bench: Type:tasklet_page_pool01_fast_path Per elem:
19 cycles(tsc) 5.492 ns (step:0) - (measurement period time:0.549225541
sec time_interval:549225541) - (invoke count:100000000
tsc_interval:1977272238)
[ 197.828174] bench_page_pool_simple:
time_bench_page_pool02_ptr_ring(): in_serving_softirq fast-path
[ 199.422305] time_bench: Type:tasklet_page_pool02_ptr_ring Per elem:
57 cycles(tsc) 15.849 ns (step:0) - (measurement period time:1.584920736
sec time_interval:1584920736) - (invoke count:100000000
tsc_interval:5705890830)
[ 199.442087] bench_page_pool_simple: time_bench_page_pool03_slow():
in_serving_softirq fast-path
[ 207.342120] time_bench: Type:tasklet_page_pool03_slow Per elem: 284
cycles(tsc) 78.909 ns (step:0) - (measurement period time:7.890955151
sec time_interval:7890955151) - (invoke count:100000000
tsc_interval:28408319289)
