Re: [PATCH v2 1/6] powerpc/smp: Cache CPU has Asymmetric SMP

[Srikar Dronamraju <srikar@linux.vnet.ibm.com>]

* Michael Ellerman <mpe@ellerman.id.au> [2023-10-19 15:33:16]:

> Srikar Dronamraju <srikar@linux.vnet.ibm.com> writes:
> > Currently cpu feature flag is checked whenever powerpc_smt_flags gets
> > called. This is an unnecessary overhead. CPU_FTR_ASYM_SMT is set based
> > on the processor and all processors will either have this set or will
> > have it unset.
> 
> The cpu_has_feature() test is implemented with a static key.
> 
> So AFAICS this is just replacing one static key with another?
> 

> I see that you use the new static key in subsequent patches. But
> couldn't those just use the existing cpu feature test?
> 

Yes, we can use the existing cpu feature test itself.

> Anyway I'd be interested to see how the generated code differs
> before/after this.
> 
---------------------------->8----------------------------------------------8<------------
Before this change
0000000000000500 <powerpc_smt_flags>:
{
     500:	00 00 4c 3c 	addis   r2,r12,0
     504:	00 00 42 38 	addi    r2,r2,0
     508:	a6 02 08 7c 	mflr    r0
     50c:	01 00 00 48 	bl      50c <powerpc_smt_flags+0xc>
     510:	f8 ff e1 fb 	std     r31,-8(r1)
     514:	91 ff 21 f8 	stdu    r1,-112(r1)
#define JUMP_ENTRY_TYPE		stringify_in_c(FTR_ENTRY_LONG)
#define JUMP_LABEL_NOP_SIZE	4

static __always_inline bool arch_static_branch(struct static_key *key, bool branch)
{
	asm_volatile_goto("1:\n\t"
     518:	00 00 00 60 	nop
		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
     51c:	00 00 22 3d 	addis   r9,r2,0
		flags |= SD_ASYM_PACKING;
     520:	80 05 e0 3b 	li      r31,1408
		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
     524:	00 00 29 89 	lbz     r9,0(r9)
     528:	00 00 09 2c 	cmpwi   r9,0
     52c:	28 00 82 41 	beq     554 <powerpc_smt_flags+0x54>
}
     530:	70 00 21 38 	addi    r1,r1,112
     534:	b4 07 e3 7f 	extsw   r3,r31
     538:	f8 ff e1 eb 	ld      r31,-8(r1)
     53c:	20 00 80 4e 	blr
	int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
     540:	80 01 e0 3b 	li      r31,384
}
     544:	70 00 21 38 	addi    r1,r1,112
     548:	b4 07 e3 7f 	extsw   r3,r31
     54c:	f8 ff e1 eb 	ld      r31,-8(r1)
     550:	20 00 80 4e 	blr
		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
     554:	a6 02 08 7c 	mflr    r0
     558:	00 00 62 3c 	addis   r3,r2,0
     55c:	01 00 20 39 	li      r9,1
     560:	00 00 42 3d 	addis   r10,r2,0
     564:	00 00 63 38 	addi    r3,r3,0
     568:	00 00 2a 99 	stb     r9,0(r10)
     56c:	80 00 01 f8 	std     r0,128(r1)
     570:	01 00 00 48 	bl      570 <powerpc_smt_flags+0x70>
     574:	00 00 00 60 	nop
     578:	80 00 01 e8 	ld      r0,128(r1)
     57c:	a6 03 08 7c 	mtlr    r0
     580:	b0 ff ff 4b 	b       530 <powerpc_smt_flags+0x30>
     584:	00 00 00 60 	nop
     588:	00 00 00 60 	nop
     58c:	00 00 00 60 	nop


post this change.
0000000000000340 <powerpc_smt_flags>:
{
     340:	a6 02 08 7c 	mflr    r0
     344:	01 00 00 48 	bl      344 <powerpc_smt_flags+0x4>
#define JUMP_ENTRY_TYPE		stringify_in_c(FTR_ENTRY_LONG)
#define JUMP_LABEL_NOP_SIZE	4

static __always_inline bool arch_static_branch(struct static_key *key, bool branch)
{
	asm_volatile_goto("1:\n\t"
     348:	00 00 00 60 	nop
	return SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
     34c:	80 01 60 38 	li      r3,384
}
     350:	b4 07 63 7c 	extsw   r3,r3
     354:	20 00 80 4e 	blr
     358:	00 00 00 60 	nop
     35c:	00 00 00 60 	nop
		return SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING;
     360:	80 05 60 38 	li      r3,1408
}
     364:	b4 07 63 7c 	extsw   r3,r3
     368:	20 00 80 4e 	blr
     36c:	00 00 00 60 	nop

---------------------------->8----------------------------------------------8<------------

I think the most of the difference is due to moving pr_info_once to
fixup_topology. Does it make sense to move the pr_info_once to
fixup_topology (which is called less often) from powerpc_smt_flags?

Even though the pr_info_once would probably translate to load + cmp + branch
we could avoid that for each smt_flag call.

So something like

diff --git a/arch/powerpc/kernel/smp.c b/arch/powerpc/kernel/smp.c
index 5826f5108a12..bc22f775425b 100644
--- a/arch/powerpc/kernel/smp.c
+++ b/arch/powerpc/kernel/smp.c
@@ -993,13 +993,10 @@ static bool shared_caches;
 /* cpumask of CPUs with asymmetric SMT dependency */
 static int powerpc_smt_flags(void)
 {
-	int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
+	if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
+		return SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING;
 
-	if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
-		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
-		flags |= SD_ASYM_PACKING;
-	}
-	return flags;
+	return SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
 }
 #endif
 
@@ -1687,6 +1684,9 @@ static void __init fixup_topology(void)
 	int i;
 
 #ifdef CONFIG_SCHED_SMT
+	if (cpu_has_feature(CPU_FTR_ASYM_SMT))
+		pr_info_once("Enabling Asymmetric SMT scheduling\n");
+
 	if (has_big_cores) {
 		pr_info("Big cores detected but using small core scheduling\n");
 		powerpc_topology[smt_idx].mask = smallcore_smt_mask;
-- 
Thanks and Regards
Srikar Dronamraju