Skip to contentSkip to conversation

your patch "crypto: sha1 - SSSE3 based SHA1 implementation for x86-64" vs xsave

DISCUSSION Jan Beulich 2 participants 4 messages Nov 17, 2011

lore
Share & tools
lore
#0 lore raw

Mathias,

in avx_usable() you use cpu_has_osxsave to determine whether AVX can actually be used, but I can't see how that conditional would ever evaluate to true. Checking CPUID.OSXSAVE is actually a user land requirement, whereas in the kernel you should really look at kernel internal variables to determine whether xsave was enabled (or, if you really want to stay with using cpu_has_osxsave you'd have to force re-execution of get_cpu_cap() after xsave got enabled on individual CPUs), namely mmu_cr4_features having X86_CR4_OSXSAVE set.

Additionally, under a hypervisor, CPUID.OSXSAVE may be set (due to the hypervisor having enabled xsave), while the kernel may be running with xsave disabled (e.g. due to a command line option saying so).

Thanks, Jan

#1 lore raw
Quoted 5 lines · Jan Beulich

Mathias,

in avx_usable() you use cpu_has_osxsave to determine whether AVX can actually be used, but I can't see how that conditional would ever evaluate to true.

Well, it does, at least in my setup, a Core i7-2620M:

  [   44.902613] sha1_ssse3: Using AVX optimized SHA-1 implementation
Quoted 3 lines

Checking CPUID.OSXSAVE is actually a user land requirement, whereas in the kernel you should really look at kernel internal variables to determine whether xsave was enabled

That's a valid point. I actually ported the check from the initial assembler implementation which was written with userland code in mind. Albeit this test can also be used in kernel code because the kernel only sets this flag after xsave was enabled in arch/x86/kernel/xsave.c:xstate_enable_boot_cpu/xstate_enable

Quoted 4 lines

(or, if you really want to stay
with using cpu_has_osxsave you'd have to force re-execution of get_cpu_cap() after xsave got enabled on individual CPUs), namely mmu_cr4_features having X86_CR4_OSXSAVE set.

I don't see where X86_CR4_OSXSAVE would get reset after xsave init. Can you give me a pointer to that code?

Quoted 3 lines

Additionally, under a hypervisor, CPUID.OSXSAVE may be set (due to the hypervisor having enabled xsave), while the kernel may be running with xsave disabled (e.g. due to a command line option saying so).

When noxsave is given on the kernel command line, X86_FEATURE_XSAVE will be cleared and xsave_init() won't call xstate_enable(), so not setting X86_CR4_OSXSAVE. All fine.

Thanks,
Mathias

#2 lore raw
Quoted 16 lines
Quoted 1 line
Quoted 1 line · Mathias Krause
Quoted 5 lines · Jan Beulich

Mathias,

in avx_usable() you use cpu_has_osxsave to determine whether AVX can actually be used, but I can't see how that conditional would ever evaluate to true.

Well, it does, at least in my setup, a Core i7-2620M:

  [   44.902613] sha1_ssse3: Using AVX optimized SHA-1 implementation
Quoted 3 lines

Checking CPUID.OSXSAVE is actually a user land requirement, whereas in the kernel you should really look at kernel internal variables to determine whether xsave was enabled

Quoted 5 lines

That's a valid point. I actually ported the check from the initial assembler implementation which was written with userland code in mind. Albeit this test can also be used in kernel code because the kernel only sets this flag after xsave was enabled in arch/x86/kernel/xsave.c:xstate_enable_boot_cpu/xstate_enable

Indeed, I got this the wrong way round.

Quoted 7 lines
Quoted 4 lines

(or, if you really want to stay
with using cpu_has_osxsave you'd have to force re-execution of get_cpu_cap() after xsave got enabled on individual CPUs), namely mmu_cr4_features having X86_CR4_OSXSAVE set.

I don't see where X86_CR4_OSXSAVE would get reset after xsave init. Can you give me a pointer to that code?

I didn't say it would get cleared. I said it should be used in favor of cpu_has_osxsave.

Quoted 7 lines
Quoted 3 lines

Additionally, under a hypervisor, CPUID.OSXSAVE may be set (due to the hypervisor having enabled xsave), while the kernel may be running with xsave disabled (e.g. due to a command line option saying so).

When noxsave is given on the kernel command line, X86_FEATURE_XSAVE will be cleared and xsave_init() won't call xstate_enable(), so not setting X86_CR4_OSXSAVE. All fine.

But you're not looking at this bit, you're looking at cpu_has_osxsave, which is a CPUID bit (not a CR4 one).

Jan

#3 lore raw
Quoted 10 lines · Jan Beulich
Quoted 2 lines

I don't see where X86_CR4_OSXSAVE would get reset after xsave init. Can you give me a pointer to that code?

I didn't say it would get cleared. I said it should be used in favor of cpu_has_osxsave.

Quoted 4 lines
Quoted 3 lines

Additionally, under a hypervisor, CPUID.OSXSAVE may be set (due to the hypervisor having enabled xsave), while the kernel may be running with xsave disabled (e.g. due to a command line option saying so).

Quoted 6 lines
Quoted 3 lines

When noxsave is given on the kernel command line, X86_FEATURE_XSAVE will be cleared and xsave_init() won't call xstate_enable(), so not setting X86_CR4_OSXSAVE. All fine.

But you're not looking at this bit, you're looking at cpu_has_osxsave, which is a CPUID bit (not a CR4 one).

You're right. But that's no problem at all, as the cpu_has_* tests are only to error out early. They also ensure it is save to execute the next instruction -- xsetbv(). This instruction actually does the xsave runtime support check by looking at the extended control register XCR0 and the enabled features in there.
So you're right that the cpu_has_osxsave might be superfluous here but it doesn't hurt either.

Mathias