[PATCH] riscv: kprobe: Optimize kprobe with accurate atomicity

[Guo Ren]

On Thu, Feb 16, 2023 at 11:42 PM Masami Hiramatsu <mhiramat at kernel.org> wrote:
>
> On Thu, 26 Jan 2023 11:15:59 -0500
> guoren at kernel.org wrote:
>
> > From: Guo Ren <guoren at linux.alibaba.com>
> >
> > The previous implementation was based on the stop_matchine mechanism,
> > which reduced the speed of arm/disarm_kprobe. Using minimum ebreak
> > instruction would get accurate atomicity.
> >
> > This patch removes the patch_text of riscv, which is based on
> > stop_machine. Then riscv only reserved patch_text_nosync, and developers
> > need to be more careful in dealing with patch_text atomicity.
> >
> > When CONFIG_RISCV_ISA_C=n, the ebreak could replace the whole
> > instruction. When CONFIG_RISCV_ISA_C=y, the patch uses 16-bit length
> > c.ebreak instruction, which may occupy the first part of the 32-bit
> > instruction and leave half the rest of the broken instruction. Because
> > ebreak could detour the flow to skip it, leaving it in the kernel text
> > memory is okay.
> >
> > Signed-off-by: Guo Ren <guoren at linux.alibaba.com>
> > Signed-off-by: Guo Ren <guoren at kernel.org>
>
> I'm not sure how the RISCV specification ensures this type of self
> code modification. But if you think calling the stop_machine() for
> *each* probe arm/disarm is slow, there may be another way to avoid
> ot by introducing a batch arming interface too. (reserve-commit way)
Could you give out "reserve-commit way" link? I'm not making sense of
that, thank you.

>
> BTW, for the BPF usecase which is usually only for function
> entry/exit, we will use Fprobes. Since that will use ftrace batch
> text patching, I think we already avoid such slowdown.
>
> FYI, for ftrace dynamic event usecase, there is another reason to slow
> down the enable/disable dynamic event itself (to sync up the event
> enabled status to ensure all event handler has been done, it waits
> for rcu-sync for each operation.)
If it's done, the ftrace common code will guarantee all events are
done. Do you know if anyone has started this work? - "sync up the
event enabled status to ensure all event handler has been done".

>
> Thank you,
>
> > ---
> >  arch/riscv/include/asm/patch.h     |  1 -
> >  arch/riscv/kernel/patch.c          | 33 ------------------------------
> >  arch/riscv/kernel/probes/kprobes.c | 29 ++++++++++++++++++--------
> >  3 files changed, 21 insertions(+), 42 deletions(-)
> >
> > diff --git a/arch/riscv/include/asm/patch.h b/arch/riscv/include/asm/patch.h
> > index 9a7d7346001e..2500782e6f5b 100644
> > --- a/arch/riscv/include/asm/patch.h
> > +++ b/arch/riscv/include/asm/patch.h
> > @@ -7,6 +7,5 @@
> >  #define _ASM_RISCV_PATCH_H
> >
> >  int patch_text_nosync(void *addr, const void *insns, size_t len);
> > -int patch_text(void *addr, u32 insn);
> >
> >  #endif /* _ASM_RISCV_PATCH_H */
> > diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c
> > index 765004b60513..8bd51ed8b806 100644
> > --- a/arch/riscv/kernel/patch.c
> > +++ b/arch/riscv/kernel/patch.c
> > @@ -98,36 +98,3 @@ int patch_text_nosync(void *addr, const void *insns, size_t len)
> >       return ret;
> >  }
> >  NOKPROBE_SYMBOL(patch_text_nosync);
> > -
> > -static int patch_text_cb(void *data)
> > -{
> > -     struct patch_insn *patch = data;
> > -     int ret = 0;
> > -
> > -     if (atomic_inc_return(&patch->cpu_count) == num_online_cpus()) {
> > -             ret =
> > -                 patch_text_nosync(patch->addr, &patch->insn,
> > -                                         GET_INSN_LENGTH(patch->insn));
> > -             atomic_inc(&patch->cpu_count);
> > -     } else {
> > -             while (atomic_read(&patch->cpu_count) <= num_online_cpus())
> > -                     cpu_relax();
> > -             smp_mb();
> > -     }
> > -
> > -     return ret;
> > -}
> > -NOKPROBE_SYMBOL(patch_text_cb);
> > -
> > -int patch_text(void *addr, u32 insn)
> > -{
> > -     struct patch_insn patch = {
> > -             .addr = addr,
> > -             .insn = insn,
> > -             .cpu_count = ATOMIC_INIT(0),
> > -     };
> > -
> > -     return stop_machine_cpuslocked(patch_text_cb,
> > -                                    &patch, cpu_online_mask);
> > -}
> > -NOKPROBE_SYMBOL(patch_text);
> > diff --git a/arch/riscv/kernel/probes/kprobes.c b/arch/riscv/kernel/probes/kprobes.c
> > index 475989f06d6d..27f8960c321c 100644
> > --- a/arch/riscv/kernel/probes/kprobes.c
> > +++ b/arch/riscv/kernel/probes/kprobes.c
> > @@ -24,12 +24,18 @@ post_kprobe_handler(struct kprobe *, struct kprobe_ctlblk *, struct pt_regs *);
> >  static void __kprobes arch_prepare_ss_slot(struct kprobe *p)
> >  {
> >       unsigned long offset = GET_INSN_LENGTH(p->opcode);
> > +#ifdef CONFIG_RISCV_ISA_C
> > +     u32 opcode = __BUG_INSN_16;
> > +#else
> > +     u32 opcode = __BUG_INSN_32;
> > +#endif
> >
> >       p->ainsn.api.restore = (unsigned long)p->addr + offset;
> >
> > -     patch_text(p->ainsn.api.insn, p->opcode);
> > -     patch_text((void *)((unsigned long)(p->ainsn.api.insn) + offset),
> > -                __BUG_INSN_32);
> > +     patch_text_nosync(p->ainsn.api.insn, &p->opcode, offset);
> > +     patch_text_nosync((void *)((unsigned long)(p->ainsn.api.insn) + offset),
> > +                       &opcode, GET_INSN_LENGTH(opcode));
> > +
> >  }
> >
> >  static void __kprobes arch_prepare_simulate(struct kprobe *p)
> > @@ -114,16 +120,23 @@ void *alloc_insn_page(void)
> >  /* install breakpoint in text */
> >  void __kprobes arch_arm_kprobe(struct kprobe *p)
> >  {
> > -     if ((p->opcode & __INSN_LENGTH_MASK) == __INSN_LENGTH_32)
> > -             patch_text(p->addr, __BUG_INSN_32);
> > -     else
> > -             patch_text(p->addr, __BUG_INSN_16);
> > +#ifdef CONFIG_RISCV_ISA_C
> > +     u32 opcode = __BUG_INSN_16;
> > +#else
> > +     u32 opcode = __BUG_INSN_32;
> > +#endif
> > +     patch_text_nosync(p->addr, &opcode, GET_INSN_LENGTH(opcode));
> >  }
> >
> >  /* remove breakpoint from text */
> >  void __kprobes arch_disarm_kprobe(struct kprobe *p)
> >  {
> > -     patch_text(p->addr, p->opcode);
> > +#ifdef CONFIG_RISCV_ISA_C
> > +     u32 opcode = __BUG_INSN_16;
> > +#else
> > +     u32 opcode = __BUG_INSN_32;
> > +#endif
> > +     patch_text_nosync(p->addr, &p->opcode, GET_INSN_LENGTH(opcode));
> >  }
> >
> >  void __kprobes arch_remove_kprobe(struct kprobe *p)
> > --
> > 2.36.1
> >
>
>
> --
> Masami Hiramatsu (Google) <mhiramat at kernel.org>



-- 
Best Regards
 Guo Ren