Re: [PATCH V5 07/11] riscv: convert to generic entry

From: Peter Zijlstra <peterz@infradead.org>
To: Guo Ren <guoren@kernel.org>
Cc: arnd@arndb.de, palmer@rivosinc.com, tglx@linutronix.de,
	luto@kernel.org, conor.dooley@microchip.com, heiko@sntech.de,
	jszhang@kernel.org, lazyparser@gmail.com, falcon@tinylab.org,
	chenhuacai@kernel.org, apatel@ventanamicro.com,
	atishp@atishpatra.org, palmer@dabbelt.com,
	paul.walmsley@sifive.com, mark.rutland@arm.com,
	zouyipeng@huawei.com, bigeasy@linutronix.de,
	David.Laight@aculab.com, linux-arch@vger.kernel.org,
	linux-kernel@vger.kernel.org, linux-riscv@lists.infradead.org,
	Guo Ren <guoren@linux.alibaba.com>
Subject: Re: [PATCH V5 07/11] riscv: convert to generic entry
Date: Tue, 20 Sep 2022 09:22:50 +0200	[thread overview]
Message-ID: <YylqSsL6bdhIOMte@hirez.programming.kicks-ass.net> (raw)
In-Reply-To: <CAJF2gTRdkmemEWsDYhVXb8KD0PS6b1VAPu_MfeZ+Rmf2qEGa6Q@mail.gmail.com>

On Tue, Sep 20, 2022 at 02:36:33PM +0800, Guo Ren wrote:
> On Mon, Sep 19, 2022 at 9:34 PM Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > On Sun, Sep 18, 2022 at 11:52:42AM -0400, guoren@kernel.org wrote:
> >
> > > @@ -123,18 +126,22 @@ int handle_misaligned_store(struct pt_regs *regs);
> > >
> > >  asmlinkage void __trap_section do_trap_load_misaligned(struct pt_regs *regs)
> > >  {
> > > +     irqentry_state_t state = irqentry_enter(regs);
> > >       if (!handle_misaligned_load(regs))
> > >               return;
> > >       do_trap_error(regs, SIGBUS, BUS_ADRALN, regs->epc,
> > >                     "Oops - load address misaligned");
> > > +     irqentry_exit(regs, state);
> > >  }
> > >
> > >  asmlinkage void __trap_section do_trap_store_misaligned(struct pt_regs *regs)
> > >  {
> > > +     irqentry_state_t state = irqentry_enter(regs);
> > >       if (!handle_misaligned_store(regs))
> > >               return;
> > >       do_trap_error(regs, SIGBUS, BUS_ADRALN, regs->epc,
> > >                     "Oops - store (or AMO) address misaligned");
> > > +     irqentry_exit(regs, state);
> > >  }
> > >  #endif
> > >  DO_ERROR_INFO(do_trap_store_fault,
> > > @@ -158,6 +165,8 @@ static inline unsigned long get_break_insn_length(unsigned long pc)
> > >
> > >  asmlinkage __visible __trap_section void do_trap_break(struct pt_regs *regs)
> > >  {
> > > +     irqentry_state_t state = irqentry_enter(regs);
> > > +
> > >  #ifdef CONFIG_KPROBES
> > >       if (kprobe_single_step_handler(regs))
> > >               return;
> >
> > FWIW; on x86 I've classified many of the 'from-kernel' traps as
> > NMI-like, since those traps can happen from any context, including with
> > IRQs disabled.
> The do_trap_break is for ebreak instruction, not NMI. RISC-V NMI has
> separate CSR. ref:
> 
> This proposal adds support for resumable non-maskable interrupts
> (RNMIs) to RISC-V. The extension adds four new CSRs (`mnepc`,
> `mncause`, `mnstatus`, and `mnscratch`) to hold the interrupted state,
> and a new instruction to resume from the RNMI handler.

Yes, but that's not what I'm saying. I'm saying I've classified
'from-kernel' traps as NMI-like.

Consider:

	raw_spin_lock_irq(&foo);
	...
	<trap>

Then you want the trap to behave as if it were an NMI; that is abide by
the rules of NMI (strictly wait-free code).

So yes, they are not NMI, but they nest just like it, so we want the
handlers to abide by the same rules.

Does that make sense?

> >
> > The basic shape of the trap handlers looks a little like:
> >
> >         if (user_mode(regs)) {
> If nmi comes from user_mode, why we using
> irqenrty_enter/exit_from/to_user_mode instead of
> irqentry_nmi_enter/exit?

s/nmi/trap/ because the 'from-user' trap never nests inside kernel code.

Additionally, many 'from-user' traps want to do 'silly' things like send
signals, which is something that requires scheduling.

They're fundamentally different from 'from-kernel' traps, which per the
above, nest most dangerously.

> >                 irqenrty_enter_from_user_mode(regs);
> >                 do_user_trap();
> >                 irqentry_exit_to_user_mode(regs);
> >         } else {
> >                 irqentry_state_t state = irqentry_nmi_enter(regs);
> >                 do_kernel_trap();
> >                 irqentry_nmi_exit(regs, state);
> >         }
> >
> > Not saying you have to match Risc-V in this patch-set, just something to
> > consider.
> I think the shape of the riscv NMI handler looks a little like this:
> 
> asmlinkage __visible __trap_section void do_trap_nmi(struct pt_regs *regs)
> {
>                  irqentry_state_t state = irqentry_nmi_enter(regs);
>                  do_nmi_trap();
>                  irqentry_nmi_exit(regs, state);
> }

That is correct for the NMI handler; but here I'm specifically talking
about traps, like the unalign trap, break trap etc. Those that can
happen *anywhere* in kernel code and nest most unfortunate.