Re: [RFC PATCH V4 10/17] x86/hyperv: Add smp support for sev-snp guest

[Tianyu Lan]

On 4/12/2023 10:59 PM, Michael Kelley (LINUX) wrote:
> >   #ifdef CONFIG_AMD_MEM_ENCRYPT
> >
> >   #define GHCB_USAGE_HYPERV_CALL	1
> >
> > +static u8 ap_start_input_arg[PAGE_SIZE] __bss_decrypted __aligned(PAGE_SIZE);
> > +static u8 ap_start_stack[PAGE_SIZE] __aligned(PAGE_SIZE);
> Just a question:  ap_start_stack is a static variable that gets used as the
> starting stack for every AP.  So obviously, once each AP is started, we must
> be sure that the AP moves off the ap_start_stack before the next AP is
> started.  How is that synchronization done?  I see that do_boot_cpu() is
> where the wakeup_secondary_cpu() function is called.  Then there's
> some waiting until the AP completes "initial initialization" per the
> comment in the code.  Is there where we know that the AP is no
> longer using ap_start_stack?

Hi Micahel:
	secondary_startup_64_no_verify() in the head_64.S initializes
a boot time stack to replace the old stack. It's very begining stage of
starting AP. The initial_stack was initialized with idle->thread.sp in
the do_boot_cpu(). The AP is started one by one in current code and so
It's safe to reuse the stack for all APs to boot up.

278        /*
279         * Setup a boot time stack - Any secondary CPU will have lost 
its stack
280         * by now because the cr3-switch above unmaps the real-mode stack
281         */
282        movq initial_stack(%rip), %rsp
283