Re: [PATCH v2 2/2] powerpc: Large user copy aware of full:rt:lazy preemption

[Shrikanth Hegde]

On 11/20/24 13:30, Sebastian Andrzej Siewior wrote:
> On 2024-11-17 00:53:06 [+0530], Shrikanth Hegde wrote:
> > Large user copy_to/from (more than 16 bytes) uses vmx instructions to
> > speed things up. Once the copy is done, it makes sense to try schedule
> > as soon as possible for preemptible kernels. So do this for
> > preempt=full/lazy and rt kernel.
> >
> > Not checking for lazy bit here, since it could lead to unnecessary
> > context switches.
> >
> > Suggested-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > Signed-off-by: Shrikanth Hegde <sshegde@xxxxxxxxxxxxx>
> > ---
> >   arch/powerpc/lib/vmx-helper.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/arch/powerpc/lib/vmx-helper.c b/arch/powerpc/lib/vmx-helper.c
> > index d491da8d1838..58ed6bd613a6 100644
> > --- a/arch/powerpc/lib/vmx-helper.c
> > +++ b/arch/powerpc/lib/vmx-helper.c
> > @@ -45,7 +45,7 @@ int exit_vmx_usercopy(void)
> >   	 * set and we are preemptible. The hack here is to schedule a
> >   	 * decrementer to fire here and reschedule for us if necessary.
> >   	 */
> > -	if (IS_ENABLED(CONFIG_PREEMPT) && need_resched())
> > +	if (IS_ENABLED(CONFIG_PREEMPTION) && need_resched())
> >   		set_dec(1);
>
> Now looking at this again there is a comment why preempt_enable() is
> bad. An interrupt between preempt_enable_no_resched() and set_dec() is
> fine because irq-exit would preempt properly?

I think so. AFAIU the comment says issue lies with amr register not being saved across
context switch. interrupt_exit_kernel_prepare saves it and restore using kuap_kernel_restore.

 Regular preemption works
> again once copy_to_user() is done? So if you copy 1GiB, you are blocked
> for that 1GiB?


yes, regular preemption would work on exit of copy_to_user. Since the preempt_disable was done
before copy starts, i think yes, it would be blocked until it is complete.

> >   	return 0;
> >   }
>
> Sebastian

nick, mpe; please correct me if i am wrong.