[PATCH] x86/cpuid: Fewer CPUID invocations in init_scattered_cpuid_features()

[PATCH] x86/cpuid: Fewer CPUID invocations in init_scattered_cpuid_features()

[David Woodhouse]

We were doing a fresh CPUID for every single bit in every single output
register. Do it once and then harvest *all* the bits we want.

We were also doing the max_level check with a new CPUID invocation for
every single bit. Stop that too.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
Spotted this in my travels; it offended me.

 arch/x86/kernel/cpu/scattered.c | 27 +++++++++++++++++++++------
 1 file changed, 21 insertions(+), 6 deletions(-)

diff --git a/arch/x86/kernel/cpu/scattered.c b/arch/x86/kernel/cpu/scattered.c
index df11f5d..26c15fb 100644
--- a/arch/x86/kernel/cpu/scattered.c
+++ b/arch/x86/kernel/cpu/scattered.c
@@ -23,8 +23,8 @@ static const struct cpuid_bit cpuid_bits[] = {
 	{ X86_FEATURE_EPB,		CPUID_ECX,  3, 0x00000006, 0 },
 	{ X86_FEATURE_CAT_L3,		CPUID_EBX,  1, 0x00000010, 0 },
 	{ X86_FEATURE_CAT_L2,		CPUID_EBX,  2, 0x00000010, 0 },
-	{ X86_FEATURE_CDP_L3,		CPUID_ECX,  2, 0x00000010, 1 },
 	{ X86_FEATURE_MBA,		CPUID_EBX,  3, 0x00000010, 0 },
+	{ X86_FEATURE_CDP_L3,		CPUID_ECX,  2, 0x00000010, 1 },
 	{ X86_FEATURE_HW_PSTATE,	CPUID_EDX,  7, 0x80000007, 0 },
 	{ X86_FEATURE_CPB,		CPUID_EDX,  9, 0x80000007, 0 },
 	{ X86_FEATURE_PROC_FEEDBACK,    CPUID_EDX, 11, 0x80000007, 0 },
@@ -38,20 +38,35 @@ void init_scattered_cpuid_features(struct cpuinfo_x86 *c)
 	u32 regs[4];
 	const struct cpuid_bit *cb;
 
+	/* Max level for basic CPUID leaves */
+	max_level = cpuid_eax(0);
+
 	for (cb = cpuid_bits; cb->feature; cb++) {
 
+		/* Are we on to extended CPUID leaves yet? */
+		if (cb->level >= 0x80000000 && max_level < 0x80000000) {
+			max_level = cpuid_eax(0x80000000);
+			if (max_level < 0x80000000 || max_level > 0x8000ffff) {
+				/* No extended leaves are supported */
+				break;
+			}
+		}
+
 		/* Verify that the level is valid */
-		max_level = cpuid_eax(cb->level & 0xffff0000);
-		if (max_level < cb->level ||
-		    max_level > (cb->level | 0xffff))
+		if (cb->level > max_level)
 			continue;
 
 		cpuid_count(cb->level, cb->sub_leaf, &regs[CPUID_EAX],
 			    &regs[CPUID_EBX], &regs[CPUID_ECX],
 			    &regs[CPUID_EDX]);
 
-		if (regs[cb->reg] & (1 << cb->bit))
-			set_cpu_cap(c, cb->feature);
+		do {
+			if (regs[cb->reg] & (1 << cb->bit))
+				set_cpu_cap(c, cb->feature);
+
+		} while (cb[1].level == cb[0].level &&
+			 cb[1].sub_leaf == cb[0].sub_leaf &&
+			 cb++);
 	}
 }
 
-- 
2.7.4

Re: [PATCH] x86/cpuid: Fewer CPUID invocations in init_scattered_cpuid_features()

[Borislav Petkov]

On Tue, Jan 23, 2018 at 02:53:33PM +0000, David Woodhouse wrote:
> We were doing a fresh CPUID for every single bit in every single output
> register. Do it once and then harvest *all* the bits we want.
> 
> We were also doing the max_level check with a new CPUID invocation for
> every single bit. Stop that too.
> 
> Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
> ---
> Spotted this in my travels; it offended me.

Ok, I see that it's itching so let's scratch it properly :-)

If we're going to optimize scattered.c, let's do something like this:

* do CPUID for each function once.
* for each set bit in there, set feature flag

which means we'd have to change the data structure.

struct cpuid_leaf {
	u32 level;
	u32 sub_leaf;
	struct cpuid_bit bits[];
};

and that last thing is:

struct cpuid_bit {
        u16 feature;
        u8 reg;
        u8 bit;
};

So that you have something like (for example with leaf 0x10):

struct cpuid_leaf leafs[] = {
...
{
	.level = 0x00000010,
	.sub_leaf = 0,
	.bits = {
		{ X86_FEATURE_CAT_L3, CPUID_EBX, 1 },
		{ X86_FEATURE_CAT_L2, CPUID_EBX, 2 },
		{ X86_FEATURE_MBA   , CPUID_EBX, 3 },
		{ 0 }
	}
}
...
}

This way you get the CPUID only once and then iterate over bits[] and
you can do the cleaner max level computation

	cpuid_eax(level & 0xffff0000);

without having to do the extended level checks.

Anyway, something like that.

It's probably not even worth doing anything though - I doubt the speedup
is visible at all.

But I certainly understand the intent to fix an annoying thing like that. :-))

Thx.

-- 
Regards/Gruss,
    Boris.

Good mailing practices for 400: avoid top-posting and trim the reply.

Re: [PATCH] x86/cpuid: Fewer CPUID invocations in init_scattered_cpuid_features()

[Borislav Petkov]

On Tue, Jan 23, 2018 at 06:32:16PM +0100, Borislav Petkov wrote:
> It's probably not even worth doing anything though - I doubt the speedup
> is visible at all.

One more thing I forgot to mention yesterday: I'm working on changing
the CPUID parsing we do now and we'll probably end up simply reading in
all CPUID leafs so scattered.c will disappear eventually. So I wouldn't
waste my energy on it.

:-)

-- 
Regards/Gruss,
    Boris.

Good mailing practices for 400: avoid top-posting and trim the reply.