Re: [PATCHv2 13/13] x86/acpi: Add support for CPU offlining for ACPI MADT wakeup method

[Kirill A. Shutemov]

On Sun, Oct 29, 2023 at 06:31:36PM +0100, Thomas Gleixner wrote:
> On Fri, Oct 20 2023 at 18:12, Kirill A. Shutemov wrote:
> 
> > MADT Multiprocessor Wakeup structure version 1 brings support of CPU
> > offlining: BIOS provides a reset vector where the CPU has to jump to
> > offline itself. The new TEST mailbox command can be used to test the CPU
> > offlined successfully and BIOS has control over it.
> >
> > Add CPU offling support for ACPI MADT wakeup method by implementing
> > custom cpu_die, play_dead and stop_other_cpus SMP operations.
> >
> > CPU offlining makes possible to hand over secondary CPUs over kexec, not
> 
> makes it possible
> 
> > limiting the second kernel with single CPU.
> 
> s/with/to/

Okay.

> > The change conforms to the approved ACPI spec change proposal. See the
> > +SYM_FUNC_START(asm_acpi_mp_play_dead)
> > +	/* Load address of reset vector into RCX to jump when kernel is ready */
> > +	movq	acpi_mp_reset_vector_paddr(%rip), %rcx
> > +
> > +	/* Turn off global entries. Following CR3 write will flush them. */
> > +	movq	%cr4, %rdx
> > +	andq	$~(X86_CR4_PGE), %rdx
> > +	movq	%rdx, %cr4
> > +
> > +	/* Switch to identity mapping */
> > +	movq	acpi_mp_pgd(%rip), %rax
> > +	movq	%rax, %cr3
> 
> You can just make this function:
> 
>     asm_acpi_mp_play_dead(u64 reset_vector, u64 pgd_pa);
> 
> then you have the reset vector in RDI and the pgd in RSI and spare the
> global variables.

Yeah, it is better. Thanks.

> >  /* Physical address of the Multiprocessor Wakeup Structure mailbox */
> > @@ -11,6 +16,150 @@ static u64 acpi_mp_wake_mailbox_paddr;
> >  /* Virtual address of the Multiprocessor Wakeup Structure mailbox */
> >  static struct acpi_madt_multiproc_wakeup_mailbox *acpi_mp_wake_mailbox;
> >  
> > +u64 acpi_mp_pgd;
> > +u64 acpi_mp_reset_vector_paddr;
> 
> See above (static) and __ro_after_init please
> 
> > +
> > +void asm_acpi_mp_play_dead(void);
> > +
> > +static void __init *alloc_pgt_page(void *context)
> > +{
> 
> What's the purpose of the context argument?

To conform to x86_mapping_info::alloc_pgt_page type.

I will rename the argument to 'dummy' and add comment.

> > +	return memblock_alloc(PAGE_SIZE, PAGE_SIZE);
> > +}
> > +
> > +/*
> > + * Make sure asm_acpi_mp_play_dead() is present in the identity mapping at
> > + * the same place as in the kernel page tables. The function switches to
> > + * the identity mapping and has be present at the same spot in before and
> > + * after transition.
> 
> Why does it need to be there after the CPU jumped to the reset vector?

After transition to the identity mapping, not after jumping to reset
vector. I will adjust the comment.

> > + */
> > +static int __init init_transition_pgtable(pgd_t *pgd)
> > +{
> > +	pgprot_t prot = PAGE_KERNEL_EXEC_NOENC;
> > +	unsigned long vaddr, paddr;
> > +	int result = -ENOMEM;
> > +	p4d_t *p4d;
> > +	pud_t *pud;
> > +	pmd_t *pmd;
> > +	pte_t *pte;
> > +
> > +	vaddr = (unsigned long)asm_acpi_mp_play_dead;
> > +	pgd += pgd_index(vaddr);
> > +	if (!pgd_present(*pgd)) {
> > +		p4d = (p4d_t *)alloc_pgt_page(NULL);
> > +		if (!p4d)
> > +			goto err;
> 
>         return -ENOMEM?
> 
> the error labels is pretty silly without an actual cleanup, right?

Right.

> > +		set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
> > +	}
> > +	p4d = p4d_offset(pgd, vaddr);
> > +	if (!p4d_present(*p4d)) {
> > +		pud = (pud_t *)alloc_pgt_page(NULL);
> > +		if (!pud)
> > +			goto err;
> 
> Ditto. But what mops up the already allocated page above?

Oops. I will add cleanup in acpi_mp_setup_reset() if
kernel_ident_mapping_init() or init_transition_pgtable() fails.

> > +	set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
> > +
> > +	return 0;
> > +err:
> > +	return result;
> > +}
> 
> Can you please move that function to the place where it is used?

Sure.

> > +
> > +static void acpi_mp_play_dead(void)
> > +{
> > +	play_dead_common();
> > +	asm_acpi_mp_play_dead();
> > +}
> > +
> > +static void acpi_mp_cpu_die(unsigned int cpu)
> > +{
> > +	int apicid = per_cpu(x86_cpu_to_apicid, cpu);
> 
> u32 apicid

Okay.

> > +	unsigned long timeout;
> > +
> > +	/*
> > +	 * Use TEST mailbox command to prove that BIOS got control over
> > +	 * the CPU before declaring it dead.
> > +	 *
> > +	 * BIOS has to clear 'command' field of the mailbox.
> > +	 */
> > +	acpi_mp_wake_mailbox->apic_id = apicid;
> > +	smp_store_release(&acpi_mp_wake_mailbox->command,
> > +			  ACPI_MP_WAKE_COMMAND_TEST);
> > +
> > +	/* Don't wait longer than a second. */
> > +	timeout = USEC_PER_SEC;
> > +	while (READ_ONCE(acpi_mp_wake_mailbox->command) && timeout--)
> > +		udelay(1);
> > +}
> > +
> > +static void acpi_mp_stop_other_cpus(int wait)
> > +{
> > +	smp_shutdown_nonboot_cpus(smp_processor_id());
> 
> This clearly was never tested with lockdep. At the point where
> stop_other_cpus() is invoked the invoking CPU has interrupts disabled...

Hm. I do have lockdep enabled and there's no reported.

And I am not sure why it should complain. At this point we are running on
reboot cpu (after migrate_to_reboot_cpu()).

Could you elaborate on what the problem here?

> > +}
> > +
> > +static void acpi_mp_crash_stop_other_cpus(void)
> > +{
> > +	smp_shutdown_nonboot_cpus(smp_processor_id());
> 
> Yuck. Crash can happen at arbitrary places. So you really cannot invoke
> the whole CPU hotplug state machine from here.
> 
> There is a reason why the other implementation just kick CPUs into some
> "safe" state.

Yeah, fair enough.

What about the implementation below? It is heavily inspired by
nmi_shootdown_cpus().

BTW, I don't understand why nmi_shootdown_cpus() needs 'crashing_cpu' if
it uses apic_send_IPI_allbutself(). 'crashing_cpu' == self, no?


static atomic_t waiting_for_crash_ipi;
static bool crash_ipi_issued;

static int crash_nmi_callback(unsigned int val, struct pt_regs *regs)
{
	local_irq_disable();

	crash_save_cpu(regs, raw_smp_processor_id());

	cpu_emergency_stop_pt();

	disable_local_APIC();

	/*
	 * Prepare the CPU for reboot _after_ invoking the callback so that the
	 * callback can safely use virtualization instructions, e.g. VMCLEAR.
	 */
	cpu_emergency_disable_virtualization();

	atomic_dec(&waiting_for_crash_ipi);

	asm_acpi_mp_play_dead(acpi_mp_reset_vector_paddr,
			      acpi_mp_pgd);

	return NMI_HANDLED;
}

static void acpi_mp_crash_stop_other_cpus(void)
{
	unsigned long timeout;

	/* The kernel is broken so disable interrupts */
	local_irq_disable();

	/*
	 * Avoid certain doom if a shootdown already occurred; re-registering
	 * the NMI handler will cause list corruption, modifying the callback
	 * will do who knows what, etc...
	 */
	if (WARN_ON_ONCE(crash_ipi_issued))
		return;

	atomic_set(&waiting_for_crash_ipi, num_online_cpus() - 1);

	/* Would it be better to replace the trap vector here? */
	if (register_nmi_handler(NMI_LOCAL, crash_nmi_callback,
				 NMI_FLAG_FIRST, "crash"))
		return;		/* Return what? */

	apic_send_IPI_allbutself(NMI_VECTOR);

	WRITE_ONCE(crash_ipi_issued, 1);

	/* Don't wait longer than a second. */
	timeout = USEC_PER_SEC;
	while (atomic_read(&waiting_for_crash_ipi) && timeout--)
		udelay(1);
}

> > +	/* The kernel is broken so disable interrupts */
> > +	local_irq_disable();
> > +}
> > +
> > +static int __init acpi_mp_setup_reset(u64 reset_vector)
> > +{
> > +	pgd_t *pgd;
> > +	struct x86_mapping_info info = {
> > +		.alloc_pgt_page = alloc_pgt_page,
> > +		.page_flag      = __PAGE_KERNEL_LARGE_EXEC,
> > +		.kernpg_flag    = _KERNPG_TABLE_NOENC,
> > +	};
> > +
> > +	pgd = alloc_pgt_page(NULL);
> > +
> > +	for (int i = 0; i < nr_pfn_mapped; i++) {
> > +		unsigned long mstart, mend;
> 
> Missing newline
> 

Okay.

> >  
> > -	/*
> > -	 * ACPI MADT doesn't allow to offline CPU after it got woke up.
> > -	 * It limits kexec: the second kernel won't be able to use more than
> > -	 * one CPU.
> > -	 *
> > -	 * Now acpi_mp_wake_mailbox_paddr already has the mailbox address.
> > -	 * The acpi_wakeup_cpu() will use it to bring up secondary cpus.
> > -	 *
> > -	 * Zero out mailbox address in the ACPI MADT wakeup structure to
> > -	 * indicate that the mailbox is not usable.  This prevents the
> > -	 * kexec()-ed kernel from reading a vaild mailbox, which in turn
> > -	 * makes the kexec()-ed kernel only be able to use the boot CPU.
> > -	 *
> > -	 * This is Linux-specific protocol and not reflected in ACPI spec.
> > -	 */
> > -	mp_wake->mailbox_address = 0;
> > +		/*
> > +		 * ACPI MADT doesn't allow to offline CPU after it got woke up.
> 
> This is not longer accurate as V1 allows that ....
> 

Will fix.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov