lkml.org 
[lkml]   [2021]   [Oct]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [GIT PULL] arm64 fixes for 5.15-rc5
Hi Linus, Thomas,

A couple of minor comments below -- I'll have a go at the rework Linus
suggested and see what blows up.

On Tue, Oct 12, 2021 at 03:18:16PM +0200, Thomas Gleixner wrote:
> On Mon, Oct 11 2021 at 12:54, Linus Torvalds wrote:
> > On Mon, Oct 11, 2021 at 3:47 AM Mark Rutland <mark.rutland@arm.com> wrote:
> > And so the reason I really hate that patch is that it introduces a new
> > "different architectures randomly and inexplicably do different
> > things, and the generic behavior is very different on arm64 than it is
> > elsewhere".
> >
> > That's just the worst kind of hack to me.
> >
> > And in this case, it's really *horribly* hard to see what the call
> > chain is. It all ends up being actively obfuscated and obscured
> > through that 'handle_arch_irq' function pointer, that is sometimes set
> > through set_handle_irq(), and sometimes set directly.
> >
> > I really think that if the rule is "we can't do accounting in
> > handle_domain_irq(), because it's too late for arm64", then the fix
> > really should be to just not do that.
> >
> > Move the irq_enter()/irq_exit() to the callers - quite possibly far up
> > the call chain to the root of it all, and just say "architecture code
> > needs to do this in the low-level code before calling
> > handle_arch_irq".
>
> That's where it belongs. It's mandatory to have it there for NOHZ_FULL
> to work correctly vs. instrumentation etc. I've pointed that out back
> then after we fed the X86 entry code into the mincer and added noinstr
> sections to keep tracers, BPF and kprobes away from it.
>
> Looking at the architectures which "support" that by selecting
> HAVE_CONTEXT_TRACKING:
>
> arch/arm/Kconfig: select HAVE_CONTEXT_TRACKING
> arch/arm64/Kconfig: select HAVE_CONTEXT_TRACKING
> arch/csky/Kconfig: select HAVE_CONTEXT_TRACKING
> arch/mips/Kconfig: select HAVE_CONTEXT_TRACKING
> arch/powerpc/Kconfig: select HAVE_CONTEXT_TRACKING if PPC64
> arch/riscv/Kconfig: select HAVE_CONTEXT_TRACKING
> arch/sparc/Kconfig: select HAVE_CONTEXT_TRACKING
> arch/x86/Kconfig: select HAVE_CONTEXT_TRACKING if X86_64
>
> S390 and X86 are (mostly) complete and use the generic entry code. S390
> does not even select HAVE_CONTEXT_TRACKING!
>
> PPC64 has done quite some work to fix that, but it looks not yet complete.
>
> Mark is working on ARM64.
>
> There is some effort underway to convert MIPS over to generic entry.
>
> The rest needs all the fundamental architecture side changes.
>
> > Anyway, it _looks_ to me like the pattern is very simple:
> >
> > Step 1:
> > - remove irq_enter/irq_exit from handle_domain_irq(), move it to all callers.
> >
> > This clearly doesn't change anything at all, but also doesn't fix the
> > problem you have. But it's easy to verify that the code is the same
> > before-and-after.

I'm happy with this in principle. The only reason we didn't go down that
route initially is because the callers are (typically) in the bowels of
arch asm or platform code, they all need to be fixed in one go to avoid
breaking anything, and it's a headache if we collide with any rework
(e.g. MIPS moving to generic entry).

As above, I'll have a go and see what blows up.

It should be possible to have C wrappers for the common cases, and have
the asm call that instead of branching directly to whichever irqchip
handler handle_arch_irq points at.

> > Step 2 is the pattern matching step:
> >
> > - if the caller of handle_domain_irq() ends up being a function that
> > is registered with set_handle_irq(), then we
> > (a) remove the irq_enter/irq_exit from it
> > (b) add it to the architectures that use handle_arch_irq.
> > (c) make sure that if there are other callers of it (not through
> > handle_arch_irq) we move that irq_enter/irq_exit into them too
> >
> > I _suspect_ - but didn't check - that Step 2(c) doesn't actually
> > exist. But who knows.
>
> It only exists with chained handlers, but they do not need that at all
> because:
>
> irq_enter()
> arch_handle_irq()
> handle_domain_irq()
> chained_handler()
> handle_domain_irq()

... and this is exactly the shape of case where we need to avoid the
nested calls so as to not break RCU's view of nesting.

> which is still the same interrupt context and not a nested interrupt.
> > It really looks like there is a very tight connection between "uses
> > handle_domain_irq()" and "uses handle_arch_irq/set_handle_irq()". No?
>
> Looks like. That might conflict with the MIPS rework though. I don't
> know how far that came already. Cc'ed the MIPS people.

There's also a bunch of old platforms on arch/arm which have a
hard-coded handler (so not using handle_arch_irq/set_handle_irq()) which
calls handle_domain_irq() -- those can be fixed up.

Thanks,
Mark.

\
 
 \ /
  Last update: 2021-10-12 16:03    [W:0.153 / U:0.228 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site