Re: [PATCH 1/2] virtio_pci: Don't make an extra copy of cpu affinity mask

[Jakub Sitnicki]

On Tue, Oct 24, 2023 at 07:46 PM +08, Xuan Zhuo wrote:
> On Tue, 24 Oct 2023 13:26:49 +0200, Jakub Sitnicki <jakub@xxxxxxxxxxxxxx> wrote:
>> On Tue, Oct 24, 2023 at 06:53 PM +08, Xuan Zhuo wrote:
>> > On Tue, 24 Oct 2023 10:17:19 +0200, Jakub Sitnicki <jakub@xxxxxxxxxxxxxx> wrote:
>> >> On Tue, Oct 24, 2023 at 10:31 AM +08, Xuan Zhuo wrote:
>> >> > On Mon, 23 Oct 2023 18:52:45 +0200, Jakub Sitnicki <jakub@xxxxxxxxxxxxxx> wrote:
>> >> >> On Thu, Oct 19, 2023 at 08:55 PM +08, Xuan Zhuo wrote:
>> >> >> > On Thu, 19 Oct 2023 12:16:24 +0200, Jakub Sitnicki <jakub@xxxxxxxxxxxxxx> wrote:
>> >> >> >> Since commit 19e226e8cc5d ("virtio: Make vp_set_vq_affinity() take a
>> >> >> >> mask.") it is actually not needed to have a local copy of the cpu mask.
>> >> >> >
>> >> >> >
>> >> >> > Could you give more info to prove this?
>> >> >
>> >> >
>> >> > Actually, my question is that can we pass a val on the stack(or temp value) to
>> >> > the irq_set_affinity_hint()?
>> >> >
>> >> > Such as the virtio-net uses zalloc_cpumask_var to alloc a cpu_mask, and
>> >> > that will be released.
>> >> >
>> >> >
>> >> >
>> >> > 	int __irq_apply_affinity_hint(unsigned int irq, const struct cpumask *m,
>> >> > 				      bool setaffinity)
>> >> > 	{
>> >> > 		unsigned long flags;
>> >> > 		struct irq_desc *desc = irq_get_desc_lock(irq, &flags, IRQ_GET_DESC_CHECK_GLOBAL);
>> >> >
>> >> > 		if (!desc)
>> >> > 			return -EINVAL;
>> >> > ->		desc->affinity_hint = m;
>> >> > 		irq_put_desc_unlock(desc, flags);
>> >> > 		if (m && setaffinity)
>> >> > 			__irq_set_affinity(irq, m, false);
>> >> > 		return 0;
>> >> > 	}
>> >> > 	EXPORT_SYMBOL_GPL(__irq_apply_affinity_hint);
>> >> >
>> >> > The above code directly refers the mask pointer. If the mask is a temp value, I
>> >> > think that is a bug.
>> >>
>> >> You are completely right. irq_set_affinity_hint stores the mask pointer.
>> >> irq_affinity_hint_proc_show later dereferences it when user reads out
>> >> /proc/irq/*/affinity_hint.
>> >>
>> >> I have failed to notice that. That's why we need cpumask_copy to stay.
>> >>
>> >> My patch is buggy. Please disregard.
>> >>
>> >> I will send a v2 to only migrate from deprecated irq_set_affinity_hint.
>> >>
>> >> > And I notice that many places directly pass the temp value to this API.
>> >> > And I am a little confused. ^_^ Or I missed something.
>> >>
>> >> There seem two be two gropus of callers:
>> >>
>> >> 1. Those that use get_cpu_mask/cpumask_of/cpumask_of_node to produce a
>> >>    cpumask pointer which is a preallocated constant.
>> >>
>> >>    $ weggli 'irq_set_affinity_hint(_, $func(_));' ~/src/linux
>> >>
>> >> 2. Those that pass a pointer to memory somewhere.
>> >>
>> >>    $ weggli 'irq_set_affinity_hint(_, $mask);' ~/src/linux
>> >>
>> >> (weggli tool can be found at https://github.com/weggli-rs/weggli)
>> >>
>> >> I've looked over the callers from group #2 but I couldn't find any
>> >> passing a pointer memory on stack :-)
>> >
>> > Pls check stmmac_request_irq_multi_msi()
>>
>> Good catch. That one looks buggy.
>>
>> I should also checked for callers that take an address of a var/field:
>>
>>   $ weggli 'irq_set_affinity_hint(_, &$mask);' ~/src/linux
>
> Do you find more?

No, just the one you pointed out. Unless I missed something.

I ran an improved query. Shows everything but the non-interesting cases:

$ weggli '{
	NOT: irq_set_affinity_hint(_, NULL);
	NOT: irq_set_affinity_hint(_, get_cpu_mask(_));
	NOT: irq_set_affinity_hint(_, cpumask_of(_));
	irq_set_affinity_hint(_, _);
}' ~/src/linux

And repeated it for irq_set_affinity_and_hint and irq_update_affinity.

The calls where it was not obvious at first sight that we're passing a
pointer to some heap memory, turned out to use a temporary variable to
either store address to heap memory or return value from cpumask_of*().