Re: [PATCH v5 06/27] arm64: Delay daif masking for user return

[Julien Thierry]

Hi James,

On 12/09/18 11:31, James Morse wrote:
> Hi Julien,
>
> On 28/08/18 16:51, Julien Thierry wrote:
> > Masking daif flags is done very early before returning to EL0.
> >
> > Only toggle the interrupt masking while in the vector entry and mask daif
> > once in kernel_exit.
>
> I had an earlier version that did this, but it showed up as a performance
> problem. commit 8d66772e869e ("arm64: Mask all exceptions during kernel_exit")
> described it as:
> |    Adding a naked 'disable_daif' to kernel_exit causes a performance problem
> |    for micro-benchmarks that do no real work, (e.g. calling getpid() in a
> |    loop). This is because the ret_to_user loop has already masked IRQs so
> |    that the TIF_WORK_MASK thread flags can't change underneath it, adding
> |    disable_daif is an additional self-synchronising operation.
> |
> |    In the future, the RAS APEI code may need to modify the TIF_WORK_MASK
> |    flags from an SError, in which case the ret_to_user loop must mask SError
> |    while it examines the flags.
>
>
> We may decide that the benchmark is silly, and we don't care about this. (At the
> time it was easy enough to work around).
>
> We need regular-IRQs masked when we read the TIF flags, and to stay masked until
> we return to user-space.
> I assume you're changing this so that psuedo-NMI are unmasked for EL0 until
> kernel_exit.

Yes.

> I'd like to be able to change the TIF flags from the SError handlers for RAS,
> which means masking SError for do_notify_resume too. (The RAS code that does
> this doesn't exist today, so you can make this my problem to work out later!)
> I think we should have psuedo_NMI masked if SError is masked too.

Yes, my intention in the few daif changes was that PseudoNMI would have 
just a little bit more priority than interrupt:

Debug > Abort > FIQ (not used) > NMI (PMR masked, PSR.I == 0) > IRQ 
(daif + PMR cleared)

So if at any point I break this just shout. (I did that change because 
currently el0_error has everything enabled before returning).

> Is there a strong reason for having psuedo-NMI unmasked during
> do_notify_resume(), or is it just for having the maximum amount of code exposed?

As you suspected, this is to have the maximum amount of code exposed to 
Pseudo-NMIs.

Since it is not a strong requirement for Pseudo-NMI, if the perf issue 
is more important I can drop the patch for now. Although it would be 
useful to have other opinions to see what makes the most sense.

Thanks,

> Thanks,
>
> James
>
> > diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
> > index 09dbea22..85ce06ac 100644
> > --- a/arch/arm64/kernel/entry.S
> > +++ b/arch/arm64/kernel/entry.S
> > @@ -259,9 +259,9 @@ alternative_else_nop_endif
> >   	.endm
> >   
> >   	.macro	kernel_exit, el
> > -	.if	\el != 0
> >   	disable_daif
> >   
> > +	.if	\el != 0
> >   	/* Restore the task's original addr_limit. */
> >   	ldr	x20, [sp, #S_ORIG_ADDR_LIMIT]
> >   	str	x20, [tsk, #TSK_TI_ADDR_LIMIT]
> > @@ -896,7 +896,7 @@ work_pending:
> >    * "slow" syscall return path.
> >    */
> >   ret_to_user:
> > -	disable_daif
> > +	disable_irq				// disable interrupts
> >   	ldr	x1, [tsk, #TSK_TI_FLAGS]
> >   	and	x2, x1, #_TIF_WORK_MASK
> >   	cbnz	x2, work_pending

--
Julien Thierry