Re: [PATCH] x86/idt: Keep spurious entries unset in system_vectors

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Thomas Gleixner <tglx@linutronix.de> writes:

> Vitaly Kuznetsov <vkuznets@redhat.com> writes:
>> Commit dc20b2d52653 ("x86/idt: Move interrupt gate initialization to IDT
>> code") had a side-effect of 'set_bit(i, used_vectors)' for unused entries
>
> That was not a side effect. That was intentional.
>
>> which are being mapped to spurious entries. (user_vectors were later
>
> user_vectors?

*used*

>
>> renamed to system_vectors).
>>
>> Previously, we used to count on system_vectors in arch_show_interrupts()
>
> We used to? Maybe you used to. I did not.

I was referring to 

commit 9d87cd61a6b71ee00b7576a3ebc10208becdbea1
Author: Vitaly Kuznetsov <vkuznets@redhat.com>
Date:   Tue Jul 7 18:26:13 2015 +0200

    x86/irq: Hide 'HYP:' line in /proc/interrupts when not on Xen/Hyper-V

and right you are, it lacks your Signed-off-by :-)

>
>> to not print unexisting entries in /proc/interrupts. E.g. 'Hypervisor
>> callback interrupts' should not be printed on bare metal. This is
>> currently broken.
>
> That was definitely unintended, but that's a purely cosmetic issue.
>
>> Setting bits in system_vectors for all unused entries also makes
>> alloc_intr_gate() fail in case someone decides to do it later. It seems
>> this is not currently an issue because all alloc_intr_gate() users are
>> calling it early, before we call idt_setup_apic_and_irq_gates() but
>> this also seems wrong.
>
> No, that's not wrong. There is absolutely no reason to allocate an
> interrupt gate later.
>
> The only thing what's wrong is that alloc_intr_gate() lacks an __init
> annotation and a sanity check to reject attempts which come in late
> after idt_setup_apic_and_irq_gates() was invoked.
>
> With that addressed we can remove the set_bit() for unused vectors to
> cure the /proc/interrupts cosmetics.
>
> Talking about side effects. This also reflects the actual number of
> system vectors which are assigned in the debugfs interface as that was
> 'broken' as well.

Ok, I was only aiming at fixing the cosmetics but making
alloc_intr_gate() __init seems to make sense too.

>
> Thanks,
>
>         tglx
>
> 8<--------------------
>  arch/x86/kernel/idt.c            |   20 ++++++++++++++++----
>  drivers/xen/events/events_base.c |   24 +++++++++++++++---------
>  2 files changed, 31 insertions(+), 13 deletions(-)
>
> --- a/arch/x86/kernel/idt.c
> +++ b/arch/x86/kernel/idt.c
> @@ -51,6 +51,9 @@ struct idt_data {
>  #define TSKG(_vector, _gdt)				\
>  	G(_vector, NULL, DEFAULT_STACK, GATE_TASK, DPL0, _gdt << 3)
>  
> +
> +static __initdata bool idt_setup_done;
> +
>  /*
>   * Early traps running on the DEFAULT_STACK because the other interrupt
>   * stacks work only after cpu_init().
> @@ -318,11 +321,16 @@ void __init idt_setup_apic_and_irq_gates
>  
>  #ifdef CONFIG_X86_LOCAL_APIC
>  	for_each_clear_bit_from(i, system_vectors, NR_VECTORS) {
> -		set_bit(i, system_vectors);
> +		/*
> +		 * Don't set the non assigned system vectors in the
> +		 * system_vectors bitmap. Otherwise they show up in
> +		 * /proc/interrupts.
> +		 */
>  		entry = spurious_entries_start + 8 * (i - FIRST_SYSTEM_VECTOR);
>  		set_intr_gate(i, entry);
>  	}
>  #endif
> +	idt_setup_done = true;
>  }
>  
>  /**
> @@ -352,6 +360,7 @@ void idt_invalidate(void *addr)
>  	load_idt(&idt);
>  }
>  
> +/* This goes away once ASYNC_PF is sanitized */

Rumor has it that the work is ongoing...

>  void __init update_intr_gate(unsigned int n, const void *addr)
>  {
>  	if (WARN_ON_ONCE(!test_bit(n, system_vectors)))
> @@ -359,9 +368,12 @@ void __init update_intr_gate(unsigned in
>  	set_intr_gate(n, addr);
>  }
>  
> -void alloc_intr_gate(unsigned int n, const void *addr)
> +void __init alloc_intr_gate(unsigned int n, const void *addr)
>  {
> -	BUG_ON(n < FIRST_SYSTEM_VECTOR);
> -	if (!test_and_set_bit(n, system_vectors))
> +	if (WARN_ON(n < FIRST_SYSTEM_VECTOR))
> +		return;
> +	if (WARN_ON(idt_setup_done))
> +		return;
> +	if (!WARN_ON(test_and_set_bit(n, system_vectors)))
>  		set_intr_gate(n, addr);
>  }
> --- a/drivers/xen/events/events_base.c
> +++ b/drivers/xen/events/events_base.c
> @@ -1639,26 +1639,32 @@ EXPORT_SYMBOL_GPL(xen_set_callback_via);
>  /* Vector callbacks are better than PCI interrupts to receive event
>   * channel notifications because we can receive vector callbacks on any
>   * vcpu and we don't need PCI support or APIC interactions. */
> -void xen_callback_vector(void)
> +static void xen_callback_vector(void)
>  {
> -	int rc;
>  	uint64_t callback_via;
>  
>  	if (xen_have_vector_callback) {
>  		callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
> -		rc = xen_set_callback_via(callback_via);
> -		if (rc) {
> +		if (xen_set_callback_via(callback_via)) {
>  			pr_err("Request for Xen HVM callback vector failed\n");
>  			xen_have_vector_callback = 0;
> -			return;
>  		}
> -		pr_info_once("Xen HVM callback vector for event delivery is enabled\n");
> -		alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR,
> -				xen_hvm_callback_vector);
>  	}
>  }
> +
> +static __init void xen_init_callback_vector(void)
> +{
> +	xen_callback_vector();

I don't quite like the 'xen_callback_vector()' name, it sounds like it
is the irq handler but it's not. If we are to split the
alloc_intr_gate() part I'd suggest the following:

xen_alloc_callback_vector() -> alloc_intr_gate()
xen_setup_callback_vector() -> hypercall

Juergen, what do you think?

> +	if (!xen_have_vector_callback)
> +		return;
> +
> +	alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, xen_hvm_callback_vector);
> +	pr_info("Xen HVM callback vector for event delivery is enabled\n");
> +}
> +
>  #else
>  void xen_callback_vector(void) {}
> +static inline void xen_init_callback_vector(void) {}
>  #endif
>  
>  #undef MODULE_PARAM_PREFIX
> @@ -1693,7 +1699,7 @@ void __init xen_init_IRQ(void)
>  			pci_xen_initial_domain();
>  	}
>  	if (xen_feature(XENFEAT_hvm_callback_vector))
> -		xen_callback_vector();
> +		xen_init_callback_vector();
>  
>  	if (xen_hvm_domain()) {
>  		native_init_IRQ();
>

Ok, I'll split this into several patches, write 'to-be-massaged'
changelogs and send v2 out. Thanks!

-- 
Vitaly