page_pool: check for dma_sync_size earlier
Subsystems (3)
PATCH v3 1/1 page_pool: check for dma_sync_size earlier
Setting dma_sync_size to 0 is not illegal, fec_main.c and ravb_main.c already did.
We can save a couple of function calls if check for dma_sync_size earlier.
This is a micro optimization, about 0.6% PPS performance improvement has been observed on a single Cortex-A53 CPU core with 64 bytes UDP RX traffic test.
Before this patch:
The average of packets per second is 234026 in one minute.
After this patch:
The average of packets per second is 235537 in one minute.
---
V2 -> V3: Add more details about measurement in commit message
V2: https://lore.kernel.org/r/[email protected]
V1 -> V2: Add measurement data about performance improvement in commit message
V1: https://lore.kernel.org/r/[email protected]
--- net/core/page_pool.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
net/core/page_pool.c +1 -1
@@ -458,7 +458,7 @@ page_pool_dma_sync_for_device(const struct page_pool *pool, netmem_ref netmem, u32 dma_sync_size) { -if (pool->dma_sync &&dma_dev_need_sync (pool->p.dev)) +if (pool->dma_sync &&dma_dev_need_sync (pool->p.dev) && dma_sync_size)__page_pool_dma_sync_for_device (pool, netmem, dma_sync_size); }
Signature · 1 line
2.34.1
Quoted 9 lines · Furong Xu
Setting
dma_sync_sizeto 0 is not illegal,fec_main.candravb_main.calready did.
We can save a couple of function calls if check fordma_sync_sizeearlier.This is a micro optimization, about 0.6% PPS performance improvement has been observed on a single Cortex-A53 CPU core with 64 bytes UDP RX traffic test.
Quoted 5 lines
Before this patch:
The average of packets per second is 234026 in one minute.After this patch:
The average of packets per second is 235537 in one minute.
Sorry, I keep skeptical that this small improvement can be statically observed? What exact tool or benchmark are you using, I wonder?
Thanks,
Jason
Quoted 10 lines · Jason Xing
Quoted 9 lines · Furong Xu
Setting
dma_sync_sizeto 0 is not illegal,fec_main.candravb_main.calready did.
We can save a couple of function calls if check fordma_sync_sizeearlier.This is a micro optimization, about 0.6% PPS performance improvement has been observed on a single Cortex-A53 CPU core with 64 bytes UDP RX traffic test.
Quoted 8 lines
Quoted 5 lines
Before this patch:
The average of packets per second is 234026 in one minute.After this patch:
The average of packets per second is 235537 in one minute.Sorry, I keep skeptical that this small improvement can be statically observed? What exact tool or benchmark are you using, I wonder?
A x86 PC send out UDP packet and the sar cmd from Sysstat package to report the PPS on RX side:
sar -n DEV 60 1
Paolo Abeni
maintainerQuoted 18 lines · Furong Xu
Quoted 16 lines · Jason Xing
Quoted 14 lines · Furong Xu
Setting
dma_sync_sizeto 0 is not illegal,fec_main.candravb_main.calready did.
We can save a couple of function calls if check fordma_sync_sizeearlier.This is a micro optimization, about 0.6% PPS performance improvement has been observed on a single Cortex-A53 CPU core with 64 bytes UDP RX traffic test.
Before this patch:
The average of packets per second is 234026 in one minute.After this patch:
The average of packets per second is 235537 in one minute.
Quoted 6 lines
Quoted 2 lines
Sorry, I keep skeptical that this small improvement can be statically observed? What exact tool or benchmark are you using, I wonder?
A x86 PC send out UDP packet and the sar cmd from Sysstat package to report the PPS on RX side:
sar -n DEV 60 1
I agree with Jason: in my experience this kind of delta on UDP pps tests is quite below the noise level.
I suggest to do a micro-benchmarking, measuring the CPU cycles required for whole page_pool_dma_sync_for_device() call via get_cycles(), on vanilla and with your patch - assuming the arch you have handy supports it.
The delta in such testing should be significant.
Thanks,
Paolo
From: Furong Xu <[email protected]>
Date: Mon, 6 Jan 2025 11:02:25 +0800
Quoted 30 lines
Setting
dma_sync_sizeto 0 is not illegal,fec_main.candravb_main.calready did.
We can save a couple of function calls if check fordma_sync_sizeearlier.This is a micro optimization, about 0.6% PPS performance improvement has been observed on a single Cortex-A53 CPU core with 64 bytes UDP RX traffic test.
Before this patch:
The average of packets per second is 234026 in one minute.After this patch:
The average of packets per second is 235537 in one minute.Signed-off-by Furong Xu <[email protected]>---
V2 -> V3: Add more details about measurement in commit message
V2: https://lore.kernel.org/r/[email protected]V1 -> V2: Add measurement data about performance improvement in commit message
V1: https://lore.kernel.org/r/[email protected]--- net/core/page_pool.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)net/core/page_pool.c +0
-0
net/core/page_pool.c page_pool_dma_sync_for_device(const struct page_pool *pool, · line 461
Quoted 5 lines
net/core/page_pool.c +1
-1@@ -458,7 +458,7 @@ page_pool_dma_sync_for_device(const struct page_pool *pool, netmem_ref netmem, u32 dma_sync_size) { -if (pool->dma_sync &&dma_dev_need_sync (pool->p.dev)) +if (pool->dma_sync &&dma_dev_need_sync (pool->p.dev) && dma_sync_size)
page_pool_dma_sync_for_device() with dma_sync_size == 0, but with pool->dma_sync set is VERY uncommon case. In general, this would happen only when the device didn't write anything to the buffer. IOW, this "shortcut" would only help *slowpath* code a bit, but potentially harming really hot functions. Such hot inline helpers are designed to make code paths which get executed in 99.999% times faster, while we don't care about the rest 0.001%.
I dunno how did you get this +0.6%, but if your driver makes Page Pool call sync_for_device(0) too often, the problem is in your driver.
net/core/page_pool.c page_pool_dma_sync_for_device(const struct page_pool *pool, · line 463
Quoted 2 lines
net/core/page_pool.c +0
-0__page_pool_dma_sync_for_device (pool, netmem, dma_sync_size); }
Thanks,
Olek