Re: [RFC PATCH 06/13] x86/uintr: Introduce uintr receiver syscalls

[Sohil Mehta]

On 9/23/2021 4:52 PM, Thomas Gleixner wrote:
> On Mon, Sep 13 2021 at 13:01, Sohil Mehta wrote:
>
> +/* UPID Notification control status */
> +#define UPID_ON		0x0	/* Outstanding notification */
> +#define UPID_SN		0x1	/* Suppressed notification */
> Come on. This are bits in upid.status, right? So why can't the comment
> above these defines says so and why can't the names not reflect that?
I'll fix this.
> > +struct uintr_upid_ctx {
> > +	struct uintr_upid *upid;
> > +	refcount_t refs;
> Please use tabular format for struct members.
Will do.
> > +};
> > +
> > +struct uintr_receiver {
> > +	struct uintr_upid_ctx *upid_ctx;
> > +};
> So we need a struct to wrap a pointer to another struct. Why?

The struct will have more members added later.  Should the wrapper be 
created then?

I didn't want to add members that are not used in this patch.

> > +inline bool uintr_arch_enabled(void)
> What's this arch_enabled indirection for? Is this used anywhere in
> non-architecture code?


I'll remove this indirection.

It is a remnant of some older code that had uintr_fd managed outside of 
the x86 code.

> > +{
> > +	return static_cpu_has(X86_FEATURE_UINTR);
> > +}
> > +
> > +static inline bool is_uintr_receiver(struct task_struct *t)
> > +{
> > +	return !!t->thread.ui_recv;
> > +}
> > +
> > +static inline u32 cpu_to_ndst(int cpu)
> > +{
> > +	u32 apicid = (u32)apic->cpu_present_to_apicid(cpu);
> > +
> > +	WARN_ON_ONCE(apicid == BAD_APICID);
> Brilliant. If x2apic is not enabled then this case returns


I'll fix this.

> > +	if (!x2apic_enabled())
> > +		return (apicid << 8) & 0xFF00;
>    (BAD_APICID << 8) & 0xFF00 == 0xFF ....
>
> > +int do_uintr_unregister_handler(void)
> > +{
> > +	struct task_struct *t = current;
> > +	struct fpu *fpu = &t->thread.fpu;
> > +	struct uintr_receiver *ui_recv;
> > +	u64 msr64;
> > +
> > +	if (!is_uintr_receiver(t))
> > +		return -EINVAL;
> > +
> > +	pr_debug("recv: Unregister handler and clear MSRs for task=%d\n",
> > +		 t->pid);
> > +
> > +	/*
> > +	 * TODO: Evaluate usage of fpregs_lock() and get_xsave_addr(). Bugs
> > +	 * have been reported recently for PASID and WRPKRU.
> Again. Which bugs and why haven't they been evaluated before posting?
I apologize again. This comment is no longer valid.
> > +	 * UPID and ui_recv will be referenced during context switch. Need to
> > +	 * disable preemption while modifying the MSRs, UPID and ui_recv thread
> > +	 * struct.
> > +	 */
> > +	fpregs_lock();
> And because you need to disable preemption you need to use
> fpregs_lock(), right? That's not what fpregs_lock() is about.
Got it. I'll evaluate the use of fpregs_lock() at all places.
> > +		wrmsrl(MSR_IA32_UINTR_MISC, msr64);
> > +		wrmsrl(MSR_IA32_UINTR_PD, 0ULL);
> > +		wrmsrl(MSR_IA32_UINTR_RR, 0ULL);
> > +		wrmsrl(MSR_IA32_UINTR_STACKADJUST, 0ULL);
> > +		wrmsrl(MSR_IA32_UINTR_HANDLER, 0ULL);
> > +	} else {
> > +		struct uintr_state *p;
> > +
> > +		p = get_xsave_addr(&fpu->state.xsave, XFEATURE_UINTR);
> > +		if (p) {
> > +			p->handler = 0;
> > +			p->stack_adjust = 0;
> > +			p->upid_addr = 0;
> > +			p->uinv = 0;
> > +			p->uirr = 0;
> > +		}
> So p == NULL is expected here?
I'll fix this and other usages of get_xsave_addr().

Thanks,

Sohil