Re: [PATCH] arm64:align function __arch_clear_user

From: Robin Murphy
Date: Tue Apr 27 2021 - 12:50:21 EST


On 2021-04-25 03:07, Kai Shen wrote:
On 2021/4/23 23:37, Catalin Marinas wrote:
On Mon, Apr 19, 2021 at 10:05:16AM +0800, Kai Shen wrote:
On 2021/4/14 18:41, Catalin Marinas wrote:
On Wed, Apr 14, 2021 at 05:25:43PM +0800, Kai Shen wrote:
Performance decreases happen in __arch_clear_user when this
function is not correctly aligned on HISI-HIP08 arm64 SOC which
fetches 32 bytes (8 instructions) from icache with a 32-bytes
aligned end address. As a result, if the hot loop is not 32-bytes
aligned, it may take more icache fetches which leads to decrease
in performance.
Dump of assembler code for function __arch_clear_user:
         0xffff0000809e3f10 :    nop
         0xffff0000809e3f14 :    mov x2, x1
         0xffff0000809e3f18 :    subs x1, x1, #0x8
         0xffff0000809e3f1c :    b.mi 0xffff0000809e3f30 <__arch_clear_user+3
-----  0xffff0000809e3f20 :    str    xzr, [x0],#8
hot    0xffff0000809e3f24 :    nop
loop   0xffff0000809e3f28 :    subs x1, x1, #0x8
-----  0xffff0000809e3f2c :    b.pl  0xffff0000809e3f20 <__arch_clear_user+1
The hot loop above takes one icache fetch as the code is in one
32-bytes aligned area and the loop takes one more icache fetch
when it is not aligned like below.
         0xffff0000809e4178 :   str    xzr, [x0],#8
         0xffff0000809e417c :   nop
         0xffff0000809e4180 :   subs x1, x1, #0x8
         0xffff0000809e4184 :   b.pl  0xffff0000809e4178 <__arch_clear_user+
Data collected by perf:
                           aligned   not aligned
            instructions   57733790     57739065
         L1-dcache-store   14938070     13718242
L1-dcache-store-misses     349280       349869
         L1-icache-loads   15380895     28500665
As we can see, L1-icache-loads almost double when the loop is not
aligned.

Sure, if the function spans two cache lines, and I$ pressure is so high that it gets evicted between each call, then indeed one would expect about twice as many fetches in total compared to if the function fits in a single line. However, that's not necessarily indicative of visible performance - if the predictors and prefetchers can do their jobs well enough there could still be little to no impact on actual execution latency.

If you want to use perf to try to justify this, at least drill down into more specific events to show that you're really stalling on I$ misses. The most convincing argument, though, would be a measured difference in actual user-memory-zeroing bandwidth - the fact that you haven't shared that, even for your synthetic test, makes me suspect that it's probably not as significant and exciting as a supposed "2x" figure... ;)

This problem is found in linux 4.19 on HISI-HIP08 arm64 SOC.
Not sure what the case is on other arm64 SOC, but it should do
no harm.
Signed-off-by: Kai Shen <shenkai8@xxxxxxxxxx>

Do you have a real world workload that's affected by this function?

I'm against adding alignments and nops for specific hardware
implementations. What about lots of other loops that the compiler may
generate or that we wrote in asm?

The benchmark we used which suffer performance decrease:
     https://github.com/redhat-performance/libMicro
     pread $OPTS -N "pread_z1k"    -s 1k    -I 300  -f /dev/zero
     pread $OPTS -N "pread_z10k"    -s 10k    -I 1000 -f /dev/zero
     pread $OPTS -N "pread_z100k"    -s 100k    -I 2000 -f /dev/zero

Is there any real world use-case that would benefit from this
optimisation? Reading /dev/zero in a loop hardly counts as a practical
workload.

Operations like "dd if=/dev/zero of=/dev/sda1" ?

Surely the write() side of that operation is still going to dominate either way? What kind of device would sda be for the whole operation to be significantly bottlenecked on *instruction fetch bandwidth* in read()!?

FWIW I'm currently playing about with some changes to this code anyway - I'll make sure to check whether function alignment has any noticeable impact on the microarchitectures I have to hand.

Robin.