Message ID | 20240613001812.2141-1-jszhang@kernel.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | arm64/lib: copy_page: s/stnp/stp | expand |
On Thu, Jun 13, 2024 at 08:18:12AM +0800, Jisheng Zhang wrote: > stnp performs non-temporal store, give a hints to the memory system > that caching is not useful for this data. But the scenario where > copy_page() used may not have this implication, although I must admit > there's such case where stnp helps performance(good). In this good > case, we can rely on the HW write streaming mechanism in some > implementations such as cortex-a55 to detect the case and take actions. > > testing with https://github.com/apinski-cavium/copy_page_benchmark > this patch can reduce the time by about 3% on cortex-a55 platforms. What about other CPUs? I'm also not convinced by such microbenchmarks. It looks like it always copies to the same page, the stp may even benefit from some caching of the data which we wouldn't need in a real scenario. So, I'm not merging this unless it's backed by some solid data across several CPU implementations.
On Mon, Jun 24, 2024 at 06:56:33PM +0100, Catalin Marinas wrote: > On Thu, Jun 13, 2024 at 08:18:12AM +0800, Jisheng Zhang wrote: > > stnp performs non-temporal store, give a hints to the memory system > > that caching is not useful for this data. But the scenario where > > copy_page() used may not have this implication, although I must admit > > there's such case where stnp helps performance(good). In this good > > case, we can rely on the HW write streaming mechanism in some > > implementations such as cortex-a55 to detect the case and take actions. > > > > testing with https://github.com/apinski-cavium/copy_page_benchmark > > this patch can reduce the time by about 3% on cortex-a55 platforms. > > What about other CPUs? I'm also not convinced by such microbenchmarks. Per my test on CA53 and CA73, CA73 got similar improvements. As for CA53 there's no difference, maybe due to the follwoing commit in the ATF: 54035fc4672aa ("Disable non-temporal hint on Cortex-A53/57") > It looks like it always copies to the same page, the stp may even > benefit from some caching of the data which we wouldn't need in a real > scenario. Yep this is also my understanding where's the improvement from. And I must admit there's case where stnp helps performance. we can rely on the HW write streaming mechanism to detect and take actions. However, sometimes, the "cache" behavior can benefit the scenario. Then in this case, the stnp here would double lose. what do you think? > > So, I'm not merging this unless it's backed by some solid data across > several CPU implementations. > > -- > Catalin
On Wed, Jun 26, 2024 at 07:50:57PM +0800, Jisheng Zhang wrote: > On Mon, Jun 24, 2024 at 06:56:33PM +0100, Catalin Marinas wrote: > > On Thu, Jun 13, 2024 at 08:18:12AM +0800, Jisheng Zhang wrote: > > > stnp performs non-temporal store, give a hints to the memory system > > > that caching is not useful for this data. But the scenario where > > > copy_page() used may not have this implication, although I must admit > > > there's such case where stnp helps performance(good). In this good > > > case, we can rely on the HW write streaming mechanism in some > > > implementations such as cortex-a55 to detect the case and take actions. > > > > > > testing with https://github.com/apinski-cavium/copy_page_benchmark > > > this patch can reduce the time by about 3% on cortex-a55 platforms. [...] > > It looks like it always copies to the same page, the stp may even > > benefit from some caching of the data which we wouldn't need in a real > > scenario. > > Yep this is also my understanding where's the improvement from. And > I must admit there's case where stnp helps performance. we can rely > on the HW write streaming mechanism to detect and take actions. Well, is that case realistic? Can you show any improvement with some real-world uses? Most likely modern CPUs fall back to non-temporal stores after a series of STPs but it depends on how soon they do it, how much cache gets polluted. OTOH, page copying could be the result of a CoW and we'd expect subsequent accesses from the user where some caching may be beneficial. So, hard to tell but we should make a decision based on a microbenchmark that writes over the same page multiple times. If you have some real-world tests that exercise this path (e.g. CoW, Android app startup) and show an improvement, I'd be in favour of this. Otherwise, no. Thanks.
diff --git a/arch/arm64/lib/copy_page.S b/arch/arm64/lib/copy_page.S index 6a56d7cf309d..4c74fe2d8bd6 100644 --- a/arch/arm64/lib/copy_page.S +++ b/arch/arm64/lib/copy_page.S @@ -32,21 +32,21 @@ SYM_FUNC_START(__pi_copy_page) 1: tst x0, #(PAGE_SIZE - 1) - stnp x2, x3, [x0, #-256] + stp x2, x3, [x0, #-256] ldp x2, x3, [x1] - stnp x4, x5, [x0, #16 - 256] + stp x4, x5, [x0, #16 - 256] ldp x4, x5, [x1, #16] - stnp x6, x7, [x0, #32 - 256] + stp x6, x7, [x0, #32 - 256] ldp x6, x7, [x1, #32] - stnp x8, x9, [x0, #48 - 256] + stp x8, x9, [x0, #48 - 256] ldp x8, x9, [x1, #48] - stnp x10, x11, [x0, #64 - 256] + stp x10, x11, [x0, #64 - 256] ldp x10, x11, [x1, #64] - stnp x12, x13, [x0, #80 - 256] + stp x12, x13, [x0, #80 - 256] ldp x12, x13, [x1, #80] - stnp x14, x15, [x0, #96 - 256] + stp x14, x15, [x0, #96 - 256] ldp x14, x15, [x1, #96] - stnp x16, x17, [x0, #112 - 256] + stp x16, x17, [x0, #112 - 256] ldp x16, x17, [x1, #112] add x0, x0, #128 @@ -54,14 +54,14 @@ SYM_FUNC_START(__pi_copy_page) b.ne 1b - stnp x2, x3, [x0, #-256] - stnp x4, x5, [x0, #16 - 256] - stnp x6, x7, [x0, #32 - 256] - stnp x8, x9, [x0, #48 - 256] - stnp x10, x11, [x0, #64 - 256] - stnp x12, x13, [x0, #80 - 256] - stnp x14, x15, [x0, #96 - 256] - stnp x16, x17, [x0, #112 - 256] + stp x2, x3, [x0, #-256] + stp x4, x5, [x0, #16 - 256] + stp x6, x7, [x0, #32 - 256] + stp x8, x9, [x0, #48 - 256] + stp x10, x11, [x0, #64 - 256] + stp x12, x13, [x0, #80 - 256] + stp x14, x15, [x0, #96 - 256] + stp x16, x17, [x0, #112 - 256] ret SYM_FUNC_END(__pi_copy_page)
stnp performs non-temporal store, give a hints to the memory system that caching is not useful for this data. But the scenario where copy_page() used may not have this implication, although I must admit there's such case where stnp helps performance(good). In this good case, we can rely on the HW write streaming mechanism in some implementations such as cortex-a55 to detect the case and take actions. testing with https://github.com/apinski-cavium/copy_page_benchmark this patch can reduce the time by about 3% on cortex-a55 platforms. Signed-off-by: Jisheng Zhang <jszhang@kernel.org> --- arch/arm64/lib/copy_page.S | 32 ++++++++++++++++---------------- 1 file changed, 16 insertions(+), 16 deletions(-)