Re: x86/paravirt: Detect over-sized patching bugs in paravirt_patch_call()

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Apr 25, 2019 at 12:57:45PM +0200, Ingo Molnar wrote:
> * Peter Zijlstra <peterz@infradead.org> wrote:
> > On Thu, Apr 25, 2019 at 11:50:39AM +0200, Ingo Molnar wrote:
> > > * Peter Zijlstra <peterz@infradead.org> wrote:
> > > > On Thu, Apr 25, 2019 at 11:17:17AM +0200, Ingo Molnar wrote:
> > > > > It basically means that we silently won't do any patching and the kernel 
> > > > > will crash later on in mysterious ways, because paravirt patching is 
> > > > > usually relied on.
> > > > 
> > > > That's OK. The compiler emits an indirect CALL/JMP to the pv_ops
> > > > structure contents. That _should_ stay valid and function correctly at
> > > > all times.
> > > 
> > > It might result in a correctly executing kernel in terms of code 
> > > generation, but it doesn't result in a viable kernel: some of the places 
> > > rely on the patching going through and don't know what to do when it 
> > > doesn't and misbehave or crash in interesting ways.
> > > 
> > > Guess how I know this. ;-)
> > 
> > What sites would that be? It really should work AFAIK.
> 
> So for example I tried to increasing the size of one of the struct 
> patch_xxl members:
> 
> --- a/arch/x86/kernel/paravirt_patch.c
> +++ b/arch/x86/kernel/paravirt_patch.c
> @@ -28,7 +28,7 @@ struct patch_xxl {
>  	const unsigned char	irq_restore_fl[2];
>  # ifdef CONFIG_X86_64
>  	const unsigned char	cpu_wbinvd[2];
> -	const unsigned char	cpu_usergs_sysret64[6];
> +	const unsigned char	cpu_usergs_sysret64[60];
>  	const unsigned char	cpu_swapgs[3];
>  	const unsigned char	mov64[3];
>  # else

So this then fails to patch the immediate; but the compiler emitted:

175:       ff 25 00 00 00 00       jmpq   *0x0(%rip)        # 17b <syscall_return_via_sysret+0x75>
		177: R_X86_64_PC32      pv_ops+0xfc

and pv_ops+0xfc is (+4 because of reloc magic):

void               (*usergs_sysret64)(void);                             /* 0x100   0x8 */

which defaults to:

arch/x86/kernel/paravirt.c:     .cpu.usergs_sysret64    = native_usergs_sysret64,

which in turn reads like:

0000000000000000 <native_usergs_sysret64>:
0:       0f 01 f8                swapgs
3:       48 0f 07                sysretq

So I _really_ don't understand how:

> Which with the vanilla kernel crashes on boot much, much later:
> 
> [    2.478026] PANIC: double fault, error_code: 0x0

happens.

> But in any case, even if many of the others will work if the patching 
> fails silently, is there any case where we'd treat patching failure as an 
> acceptable case?

It really should just work. And we need to figure out why it comes
unstuck. Can you print the code when it fails patching?

> BUG_ON() in paravirt kernels is an easily debuggable condition and beats 
> the above kinds of symptoms. But I can turn it into a WARN_ON_ONCE() if 
> you think that's better?

Not patching should be a performance issue; not a correctness issue, as
per the above. So WARN is the right thing.