Re: [PATCH v13 00/11] Parallel CPU bringup for x86_64

[Tom Lendacky <thomas.lendacky@amd.com>]

On 3/8/23 03:04, David Woodhouse wrote:
> On Tue, 2023-03-07 at 16:55 -0600, Tom Lendacky wrote:
>> On 3/7/23 16:27, David Woodhouse wrote:
>>> On Tue, 2023-03-07 at 16:22 -0600, Tom Lendacky wrote:
>>>>
>>>> I did some Qemu/KVM testing. One thing I noticed is that on AMD, CPUID 0xB
>>>> EAX will be non-zero only if SMT is enabled. So just booting some guests
>>>> without CPU topology never did parallel booting ("smpboot: Disabling
>>>> parallel bringup because CPUID 0xb looks untrustworthy"). I would imagine
>>>> a bare-metal system that has diabled SMT will not do parallel booting, too
>>>> (but I haven't had time to test that).
>>>
>>> Interesting, thanks. Should I change to checking for *both* EAX and EBX
>>> being zero? That's what I did first, after reading only the Intel SDM.
>>> But I changed to only EAX because the AMD doc only says that EAX will
>>> be zero for unsupported leaves.
>>
>>   From a baremetal perspective, I think that works. Rome was the first
>> generation to support x2apic, and the PPR for Rome states that 0's are
>> returned in all 4 registers for undefined function numbers.
>>
>> For virtualization, at least Qemu/KVM, that also looks to be a safe test.
> 
> At Sean's suggestion, I've switched it to use the existing
> check_extended_topology_leaf() which checks for EBX being non-zero, and
> CH being 1 (SMT_TYPE).
> 
> I also made it work even if the kernel isn't using x2apic mode (is that
> even possible, or does SEV-ES require the MSR-based access anyway?)
> 
> It just looked odd handling SEV-ES in the CPUID 0x0B path but not the
> CPUID 0x01 case, and I certainly didn't want to implement the asm side
> for handling CPUID 0x01 via the GHCB protocol. And this way I can pull
> the check for CC_ATTR_GUEST_STATE_ENCRYPT up above. Which I've kept for
> now for the reason described in the comment, but I won't die on that
> hill.

You can boot an SEV-ES guest in apic mode, but that would be unusual, so I 
think this approach is fine.

Thanks,
Tom

> 
> https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/parallel-6.2-v14
> 
> Looks like this:
> 
> /*
>   * We can do 64-bit AP bringup in parallel if the CPU reports its APIC
>   * ID in CPUID (either leaf 0x0B if we need the full APIC ID in X2APIC
>   * mode, or leaf 0x01 if 8 bits are sufficient). Otherwise it's too
>   * hard.
>   */
> static bool prepare_parallel_bringup(void)
> {
> 	bool has_sev_es = IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT) &&
> 		static_branch_unlikely(&sev_es_enable_key);
> 
> 	if (IS_ENABLED(CONFIG_X86_32))
> 		return false;
> 
> 	/*
> 	 * Encrypted guests other than SEV-ES (in the future) will need to
> 	 * implement an early way of finding the APIC ID, since they will
> 	 * presumably block direct CPUID too. Be kind to our future selves
> 	 * by warning here instead of just letting them break. Parallel
> 	 * startup doesn't have to be in the first round of enabling patches
> 	 * for any such technology.
> 	 */
> 	if (cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT) || !has_sev_es) {
> 		pr_info("Disabling parallel bringup due to guest memory encryption\n");
> 		return false;
> 	}
> 
> 	if (x2apic_mode || has_sev_es) {
> 		if (boot_cpu_data.cpuid_level < 0x0b)
> 			return false;
> 
> 		if (check_extended_topology_leaf(0x0b) != 0) {
> 			pr_info("Disabling parallel bringup because CPUID 0xb looks untrustworthy\n");
> 			return false;
> 		}
> 
> 		if (has_sev_es) {
> 			pr_debug("Using SEV-ES CPUID 0xb for parallel CPU startup\n");
> 			smpboot_control = STARTUP_APICID_SEV_ES;
> 		} else {
> 			pr_debug("Using CPUID 0xb for parallel CPU startup\n");
> 			smpboot_control = STARTUP_APICID_CPUID_0B;
> 		}
> 	} else {
> 		/* Without X2APIC, what's in CPUID 0x01 should suffice. */
> 		if (boot_cpu_data.cpuid_level < 0x01)
> 			return false;
> 
> 		pr_debug("Using CPUID 0x1 for parallel CPU startup\n");
> 		smpboot_control = STARTUP_APICID_CPUID_01;
> 	}
> 
> 	cpuhp_setup_state_nocalls(CPUHP_BP_PARALLEL_DYN, "x86/cpu:kick",
> 				  native_cpu_kick, NULL);
> 	return true;
> }
>