Re: [PATCH v4 05/10] mm: Return faster for non-fatal signals in user mode faults

From: Matthew Wilcox
Date: Tue Sep 24 2019 - 11:45:25 EST


On Tue, Sep 24, 2019 at 11:19:08AM +0800, Peter Xu wrote:
> On Mon, Sep 23, 2019 at 07:54:47PM -0700, Matthew Wilcox wrote:
> > On Tue, Sep 24, 2019 at 10:47:21AM +0800, Peter Xu wrote:
> > > On Mon, Sep 23, 2019 at 11:03:49AM -0700, Linus Torvalds wrote:
> > > > On Sun, Sep 22, 2019 at 9:26 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
> > > > >
> > > > > This patch is a preparation of removing that special path by allowing
> > > > > the page fault to return even faster if we were interrupted by a
> > > > > non-fatal signal during a user-mode page fault handling routine.
> > > >
> > > > So I really wish saome other vm person would also review these things,
> > > > but looking over this series once more, this is the patch I probably
> > > > like the least.
> > > >
> > > > And the reason I like it the least is that I have a hard time
> > > > explaining to myself what the code does and why, and why it's so full
> > > > of this pattern:
> > > >
> > > > > - if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current))
> > > > > + if ((fault & VM_FAULT_RETRY) &&
> > > > > + fault_should_check_signal(user_mode(regs)))
> > > > > return;
> > > >
> > > > which isn't all that pretty.
> > > >
> > > > Why isn't this just
> > > >
> > > > static bool fault_signal_pending(unsigned int fault_flags, struct
> > > > pt_regs *regs)
> > > > {
> > > > return (fault_flags & VM_FAULT_RETRY) &&
> > > > (fatal_signal_pending(current) ||
> > > > (user_mode(regs) && signal_pending(current)));
> > > > }
> > > >
> > > > and then most of the users would be something like
> > > >
> > > > if (fault_signal_pending(fault, regs))
> > > > return;
> > > >
> > > > and the exceptions could do their own thing.
> > > >
> > > > Now the code is prettier and more understandable, I feel.
> > > >
> > > > And if something doesn't follow this pattern, maybe it either _should_
> > > > follow that pattern or it should just not use the helper but explain
> > > > why it has an unusual pattern.
> >
> > > +++ b/arch/alpha/mm/fault.c
> > > @@ -150,7 +150,7 @@ do_page_fault(unsigned long address, unsigned long mmcsr,
> > > the fault. */
> > > fault = handle_mm_fault(vma, address, flags);
> > >
> > > - if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current))
> > > + if (fault_signal_pending(fault, regs))
> > > return;
> > >
> > > if (unlikely(fault & VM_FAULT_ERROR)) {
> >
> > > +++ b/arch/arm/mm/fault.c
> > > @@ -301,6 +301,11 @@ do_page_fault(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
> > > return 0;
> > > }
> > >
> > > + /* Fast path to handle user mode signals */
> > > + if ((fault & VM_FAULT_RETRY) && user_mode(regs) &&
> > > + signal_pending(current))
> > > + return 0;
> >
> > But _why_ are they different? This is a good opportunity to make more
> > code the same between architectures.
>
> (Thanks for joining the discussion)
>
> I'd like to do these - my only worry is that I can't really test them
> well simply because I don't have all the hardwares. For now the
> changes are mostly straightforward so I'm relatively confident (not to
> mention the code needs proper reviews too, and of course I would
> appreciate much if anyone wants to smoke test it). If I change it in
> a drastic way, I won't be that confident without some tests at least
> on multiple archs (not to mention that even smoke testing across major
> archs will be a huge amount of work...). So IMHO those might be more
> suitable as follow-up for per-arch developers if we can at least reach
> a consensus on the whole idea of this patchset.

I think the way to do this is to introduce fault_signal_pending(),
converting the architectures to it that match that pattern. Then one
patch per architecture to convert the ones which use a different pattern
to the same pattern.

Oh, and while you're looking at the callers of handle_mm_fault(), a
lot of them don't check conditions in the right order. x86, at least,
handles FAULT_RETRY before handling FAULT_ERROR, which is clearly wrong.

Kirill and I recently discussed it here:
https://lore.kernel.org/linux-mm/20190911152338.gqqgxrmqycodfocb@box/T/