Message-ID:
 <SN6PR02MB41577A6B0E5898B68E2CD3C9D4842@SN6PR02MB4157.namprd02.prod.outlook.com>
Date: Fri, 25 Apr 2025 14:32:14 +0000
From: Michael Kelley <mhklinux@...look.com>
To: Peter Zijlstra <peterz@...radead.org>
CC: "x86@...nel.org" <x86@...nel.org>, "kys@...rosoft.com"
	<kys@...rosoft.com>, "haiyangz@...rosoft.com" <haiyangz@...rosoft.com>,
	"wei.liu@...nel.org" <wei.liu@...nel.org>, "decui@...rosoft.com"
	<decui@...rosoft.com>, "tglx@...utronix.de" <tglx@...utronix.de>,
	"mingo@...hat.com" <mingo@...hat.com>, "bp@...en8.de" <bp@...en8.de>,
	"dave.hansen@...ux.intel.com" <dave.hansen@...ux.intel.com>, "hpa@...or.com"
	<hpa@...or.com>, "jpoimboe@...nel.org" <jpoimboe@...nel.org>,
	"pawan.kumar.gupta@...ux.intel.com" <pawan.kumar.gupta@...ux.intel.com>,
	"seanjc@...gle.com" <seanjc@...gle.com>, "pbonzini@...hat.com"
	<pbonzini@...hat.com>, "ardb@...nel.org" <ardb@...nel.org>, "kees@...nel.org"
	<kees@...nel.org>, Arnd Bergmann <arnd@...db.de>,
	"gregkh@...uxfoundation.org" <gregkh@...uxfoundation.org>,
	"linux-hyperv@...r.kernel.org" <linux-hyperv@...r.kernel.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"kvm@...r.kernel.org" <kvm@...r.kernel.org>, "linux-efi@...r.kernel.org"
	<linux-efi@...r.kernel.org>, "samitolvanen@...gle.com"
	<samitolvanen@...gle.com>, "ojeda@...nel.org" <ojeda@...nel.org>
Subject: RE: [PATCH 5/6] x86_64,hyperv: Use direct call to hypercall-page

From: Peter Zijlstra <peterz@...radead.org> Sent: Friday, April 25, 2025 7:04 AM
> 
> On Mon, Apr 21, 2025 at 06:28:42PM +0000, Michael Kelley wrote:
> 
> > >  #ifdef CONFIG_X86_64
> > > +static u64 __hv_hyperfail(u64 control, u64 param1, u64 param2)
> > > +{
> > > +	return U64_MAX;
> > > +}
> > > +
> > > +DEFINE_STATIC_CALL(__hv_hypercall, __hv_hyperfail);
> > > +
> > >  u64 hv_pg_hypercall(u64 control, u64 param1, u64 param2)
> > >  {
> > >  	u64 hv_status;
> > >
> > > +	asm volatile ("call " STATIC_CALL_TRAMP_STR(__hv_hypercall)
> > >  		      : "=a" (hv_status), ASM_CALL_CONSTRAINT,
> > >  		        "+c" (control), "+d" (param1)
> > > +		      : "r" (__r8)
> > >  		      : "cc", "memory", "r9", "r10", "r11");
> > >
> > >  	return hv_status;
> > >  }
> > > +
> > > +typedef u64 (*hv_hypercall_f)(u64 control, u64 param1, u64 param2);
> > > +
> > > +static inline void hv_set_hypercall_pg(void *ptr)
> > > +{
> > > +	hv_hypercall_pg = ptr;
> > > +
> > > +	if (!ptr)
> > > +		ptr = &__hv_hyperfail;
> > > +	static_call_update(__hv_hypercall, (hv_hypercall_f)ptr);
> > > +}
> 
> ^ kept for reference, as I try and explain how static_call() works
> below.
> 
> > > -skip_hypercall_pg_init:
> > > -	/*
> > > -	 * Some versions of Hyper-V that provide IBT in guest VMs have a bug
> > > -	 * in that there's no ENDBR64 instruction at the entry to the
> > > -	 * hypercall page. Because hypercalls are invoked via an indirect call
> > > -	 * to the hypercall page, all hypercall attempts fail when IBT is
> > > -	 * enabled, and Linux panics. For such buggy versions, disable IBT.
> > > -	 *
> > > -	 * Fixed versions of Hyper-V always provide ENDBR64 on the hypercall
> > > -	 * page, so if future Linux kernel versions enable IBT for 32-bit
> > > -	 * builds, additional hypercall page hackery will be required here
> > > -	 * to provide an ENDBR32.
> > > -	 */
> > > -#ifdef CONFIG_X86_KERNEL_IBT
> > > -	if (cpu_feature_enabled(X86_FEATURE_IBT) &&
> > > -	    *(u32 *)hv_hypercall_pg != gen_endbr()) {
> > > -		setup_clear_cpu_cap(X86_FEATURE_IBT);
> > > -		pr_warn("Disabling IBT because of Hyper-V bug\n");
> > > -	}
> > > -#endif
> >
> > With this patch set, it's nice to see IBT working in a Hyper-V guest!
> > I had previously tested IBT with some hackery to the hypercall page
> > to add the missing ENDBR64, and didn't see any problems. Same
> > after these changes -- no complaints from IBT.
> 
> No indirect calls left, no IBT complaints ;-)
> 
> > > +	hv_set_hypercall_pg(hv_hypercall_pg);
> > >
> > > +skip_hypercall_pg_init:
> > >  	/*
> > >  	 * hyperv_init() is called before LAPIC is initialized: see
> > >  	 * apic_intr_mode_init() -> x86_platform.apic_post_init() and
> > > @@ -658,7 +658,7 @@ void hyperv_cleanup(void)
> > >  	 * let hypercall operations fail safely rather than
> > >  	 * panic the kernel for using invalid hypercall page
> > >  	 */
> > > -	hv_hypercall_pg = NULL;
> > > +	hv_set_hypercall_pg(NULL);
> >
> > This causes a hang getting into the kdump kernel after a panic.
> > hyperv_cleanup() is called after native_machine_crash_shutdown()
> > has done crash_smp_send_stop() on all the other CPUs. I don't know
> > the details of how static_call_update() works,
> 
> Right, so let me try and explain this :-)
> 
> So we get the compiler to emit direct calls (CALL/JMP) to symbols
> prefixed with "__SCT__", in this case from asm, but more usually by
> means of the static_call() macro mess.
> 
> Meanwhile DEFINE_STATIC_CALL() ensures such a symbol actually exists.
> This symbol is a little trampoline that redirects to the actual
> target function given to DEFINE_STATIC_CALL() -- __hv_hyperfail() in the
> above case.
> 
> Then objtool runs through the resulting object file and stores the
> location of every call to these __STC__ prefixed symbols in a custom
> section.
> 
> This enables static_call init (boot time) to go through the section and
> rewrite all the trampoline calls to direct calls to the target.
> Subsequent static_call_update() calls will again rewrite the direct call
> to point elsewhere.
> 
> So very much how static_branch() does a NOP/JMP rewrite to toggle
> branches, static_call() rewrites (direct) call targets.
> 
> > but it's easy to imagine that
> > it wouldn't work when the kernel is in such a state.
> >
> > The original code setting hv_hypercall_pg to NULL is just tidiness.
> > Other CPUs are stopped and can't be making hypercalls, and this CPU
> > shouldn't be making hypercalls either, so setting it to NULL more
> > cleanly catches some erroneous hypercall (vs. accessing the hypercall
> > page after Hyper-V has been told to reset it).
> 
> So if you look at (retained above) hv_set_hypercall_pg(), when given
> NULL, the call target is set to __hv_hyperfail(), which does an
> unconditional U64_MAX return.
> 
> Combined with the fact that the thing *should* not be doing hypercalls
> anymore at this point, something is iffy.
> 
> I can easily remove it, but it *should* be equivalent to before, where
> it dynamicall checked for hv_hypercall_pg being NULL.

I agree that setting the call target to __hv_hyperfail() should be good.
But my theory is that static_call_update() is hanging when trying to
do the rewrite, because of the state of the other CPUs. I don't think
control is ever returning from static_call_update() when invoked
through hyperv_cleanup(). Wouldn't static_call_update() need to park
the other CPUs temporarily and/or flush instruction caches to make
everything consistent?

But that's just my theory. I'll run a few more experiments to confirm
if control ever returns from static_call_update() in this case.

Michael

Powered by blists - more mailing lists