Re: [CAUTION - External Sender] Re: [PATCH] riscv: Improve exception and system call latency

[Cyril Bur]

On Mon, Jun 3, 2024 at 4:39 PM Guo Ren <guoren@xxxxxxxxxx> wrote:
>
> On Mon, Jun 3, 2024 at 12:38 PM Cyril Bur <cyrilbur@xxxxxxxxxxxxxxx> wrote:
> >
> > [ apologies, I think my mailer is going to mess up the formatting ]
> >
> > On 26 Dec 2023, at 2:56 PM, Guo Ren <guoren@xxxxxxxxxx> wrote:
> >
> > On Sun, Dec 24, 2023 at 08:00:18PM -0800, Anton Blanchard wrote:
> >
> > Many CPUs implement return address branch prediction as a stack. The
> > RISCV architecture refers to this as a return address stack (RAS). If
> > this gets corrupted then the CPU will mispredict at least one but
> > potentally many function returns.
> >
> > There are two issues with the current RISCV exception code:
> >
> > - We are using the alternate link stack (x5/t0) for the indirect branch
> >  which makes the hardware think this is a function return. This will
> >  corrupt the RAS.
> >
> > - We modify the return address of handle_exception to point to
> >  ret_from_exception. This will also corrupt the RAS.
> >
> > Testing the null system call latency before and after the patch:
> >
> > Visionfive2 (StarFive JH7110 / U74)
> > baseline: 189.87 ns
> > patched:  176.76 ns
> >
> > Lichee pi 4a (T-Head TH1520 / C910)
> > baseline: 666.58 ns
> > patched:  636.90 ns
> >
> > Just over 7% on the U74 and just over 4% on the C910.
> >
> >
> > Yes, the wrong "jalr zero, t0/ra" would pop RAS and destroy the RAS
> > layout of the hardware for the userspace. How about giving a fake push
> > for the RAS to connect "jalr zero, ra" of sub-function call return? I'm
> > curious if you could measure the difference with only one RAS
> > misprediction.
> >
> > diff --git a/arch/riscv/kernel/entry.S b/arch/riscv/kernel/entry.S
> > index 54ca4564a926..94c7d2be35d0 100644
> > --- a/arch/riscv/kernel/entry.S
> > +++ b/arch/riscv/kernel/entry.S
> > @@ -93,7 +93,8 @@ SYM_CODE_START(handle_exception)
> >        bge s4, zero, 1f
> >
> >        /* Handle interrupts */
> > -       tail do_irq
> > +       auipc t0, do_irq
> > +       jalr  t0, t0
> > 1:
> >        /* Handle other exceptions */
> >        slli t0, s4, RISCV_LGPTR
> > @@ -103,9 +104,10 @@ SYM_CODE_START(handle_exception)
> >        /* Check if exception code lies within bounds */
> >        bgeu t0, t2, 1f
> >        REG_L t0, 0(t0)
> > -       jr t0
> > +       jalr t0, t0
> > 1:
> > -       tail do_trap_unknown
> > +       auipc t0, do_trap_unknown
> > +       jalr  t0, t0
> > SYM_CODE_END(handle_exception)
> >
> > You could prepare a deeper userspace stack calling if you want better
> > measurement results.
> >
> >
> > Signed-off-by: Anton Blanchard <antonb@xxxxxxxxxxxxxxx>
> > Reviewed-by: Jisheng Zhang <jszhang@xxxxxxxxxx>
> > ---
> >
> > This introduces some complexity in the stackframe walk code. PowerPC
> > resolves the multiple exception exit paths issue by placing a value into
> > the exception stack frame (basically the word "REGS") that the stack frame
> > code can look for. Perhaps something to look at.
> >
> > arch/riscv/kernel/entry.S      | 21 ++++++++++++++-------
> > arch/riscv/kernel/stacktrace.c | 14 +++++++++++++-
> > 2 files changed, 27 insertions(+), 8 deletions(-)
> >
> > diff --git a/arch/riscv/kernel/entry.S b/arch/riscv/kernel/entry.S
> > index 54ca4564a926..89af35edbf6c 100644
> > --- a/arch/riscv/kernel/entry.S
> > +++ b/arch/riscv/kernel/entry.S
> > @@ -84,7 +84,6 @@ SYM_CODE_START(handle_exception)
> > scs_load_current_if_task_changed s5
> >
> > move a0, sp /* pt_regs */
> > - la ra, ret_from_exception
> >
> > /*
> > * MSB of cause differentiates between
> > @@ -93,7 +92,10 @@ SYM_CODE_START(handle_exception)
> > bge s4, zero, 1f
> >
> > /* Handle interrupts */
> > - tail do_irq
> > + call do_irq
> > +.globl ret_from_irq_exception
> > +ret_from_irq_exception:
> > + j ret_from_exception
> > 1:
> > /* Handle other exceptions */
> > slli t0, s4, RISCV_LGPTR
> > @@ -101,11 +103,16 @@ SYM_CODE_START(handle_exception)
> > la t2, excp_vect_table_end
> > add t0, t1, t0
> > /* Check if exception code lies within bounds */
> > - bgeu t0, t2, 1f
> > - REG_L t0, 0(t0)
> > - jr t0
> > -1:
> > - tail do_trap_unknown
> > + bgeu t0, t2, 3f
> > + REG_L t1, 0(t0)
> > +2: jalr ra,t1
> > +.globl ret_from_other_exception
> > +ret_from_other_exception:
> > + j ret_from_exception
> > +3:
> > +
> > + la t1, do_trap_unknown
> > + j 2b
> > SYM_CODE_END(handle_exception)
> >
> > /*
> > diff --git a/arch/riscv/kernel/stacktrace.c b/arch/riscv/kernel/stacktrace.c
> > index 64a9c093aef9..b9cd131bbc4c 100644
> > --- a/arch/riscv/kernel/stacktrace.c
> > +++ b/arch/riscv/kernel/stacktrace.c
> > @@ -17,6 +17,18 @@
> > #ifdef CONFIG_FRAME_POINTER
> >
> > extern asmlinkage void ret_from_exception(void);
> > +extern asmlinkage void ret_from_irq_exception(void);
> > +extern asmlinkage void ret_from_other_exception(void);
> > +
> > +static inline bool is_exception_frame(unsigned long pc)
> > +{
> > + if ((pc == (unsigned long)ret_from_exception) ||
> > +    (pc == (unsigned long)ret_from_irq_exception) ||
> > +    (pc == (unsigned long)ret_from_other_exception))
> > + return true;
> > +
> > + return false;
> > +}
> >
> > We needn't put too many .globl in the entry.S, and just check that pc is
> > in SYM_CODE_START/END(handle_exception), then entry.S would be cleaner:
> >
> > Hi Guo,
> >
> > I've taken this patch over from Anton, mostly just to tidy it up. I'd
> > like to incorporate
> > what you mention here but I'm not sure how to achieve it. Have I
> > missed something
> > obvious? As things currently stand there doesn't seem to be a way to get the end
> > (or size) of handle_exception in C code.
> "just check that pc is in SYM_CODE_START/END(handle_exception)."
> Sorry, I think my previous description is wrong.
>
> Instead, "We needn't modify anything in stacktrace.c because we keep
> ra = ret_from_exception."
>
> I want only cleaner and smaller modifications to the entry.S to
> satisfy RAS prediction performance requirements.
>

I completely agree with keeping entry.S as clean as possible.

I'm trying to understand what you mean.
Isn't the point of the patch to remove ra = ret_from_exception?

I'm not sure but maybe we can leave entry.S as it is and the check in
stacktrace.c can become a check for pc == handle_exception?

>
> >
> > Your advice is greatly appreciated,
> >
> > Thanks,
> >
> > Cyril
> >
> > diff --git a/arch/riscv/kernel/entry.S b/arch/riscv/kernel/entry.S
> > index 54ca4564a926..d452d5f12b1b 100644
> > --- a/arch/riscv/kernel/entry.S
> > +++ b/arch/riscv/kernel/entry.S
> > @@ -84,7 +84,6 @@ SYM_CODE_START(handle_exception)
> >        scs_load_current_if_task_changed s5
> >
> >        move a0, sp /* pt_regs */
> >
> >        /*
> >         * MSB of cause differentiates between
> > @@ -93,7 +92,8 @@ SYM_CODE_START(handle_exception)
> >        bge s4, zero, 1f
> >
> >        /* Handle interrupts */
> >        call do_irq
> >        j ret_from_exception
> > 1:
> >        /* Handle other exceptions */
> >        slli t0, s4, RISCV_LGPTR
> > @@ -102,10 +102,12 @@ SYM_CODE_START(handle_exception)
> >        add t0, t1, t0
> >        /* Check if exception code lies within bounds */
> >        bgeu t0, t2, 1f
> >        REG_L ra, 0(t0)
> >        jalr ra, ra
> >        j ret_from_exception
> > 1:
> >        call do_trap_unknown
> >        j ret_from_exception
> > SYM_CODE_END(handle_exception)
> >
> >
> >
> > void notrace walk_stackframe(struct task_struct *task, struct pt_regs *regs,
> >     bool (*fn)(void *, unsigned long), void *arg)
> > @@ -62,7 +74,7 @@ void notrace walk_stackframe(struct task_struct
> > *task, struct pt_regs *regs,
> > fp = frame->fp;
> > pc = ftrace_graph_ret_addr(current, NULL, frame->ra,
> >   &frame->ra);
> > - if (pc == (unsigned long)ret_from_exception) {
> > + if (is_exception_frame(pc)) {
> > if (unlikely(!__kernel_text_address(pc) || !fn(arg, pc)))
> > break;
> >
> > --
> > 2.25.1
> >
> >
> > _______________________________________________
> > linux-riscv mailing list
> > linux-riscv@xxxxxxxxxxxxxxxxxxx
> > http://lists.infradead.org/mailman/listinfo/linux-riscv
>
>
>
> --
> Best Regards
>  Guo Ren