Re: [PATCH v3 6/7] riscv: Add KPROBES_ON_FTRACE supported

[Masami Hiramatsu <mhiramat@kernel.org>]

On Wed, 22 Jul 2020 21:31:20 +0800
Guo Ren <guoren@kernel.org> wrote:

> Hi Masami,
> 
> Current riscv ftrace_caller utilize fp(s0) - 8 in stack to get ra of
> function, eg:
> foo:
>    2bb0:       7119                    addi    sp,sp,-128
>     2bb2:       f8a2                    sd      s0,112(sp)
>     2bb4:       fc86                    sd      ra,120(sp)
> ...
>     2bc4:       0100                    addi    s0,sp,128
> ...
> 0000000000002bca <.LVL828>:
>     2bca:       00000097                auipc   ra,0x0
>     2bce:       000080e7                jalr    ra # 2bca <.LVL828> //_mcount
> 
> So just put two nops before prologue of function isn't enough, because
> riscv don't like arm64 which could use x9-x18 reserved regs to pass
> ra(x30).
>     | mov       x9, x30
>     | bl        <ftrace-entry>
> If the benefit is just making a kprobe on function symbol address to
> prevent disassembling, I'll delay this feature.

I recommend that. This feature has to involve ftrace and gcc, so
it is better to split it from this series.

> 
> 
> I also have a look at HAVE_FENTRY & HAVE_NOP_MCOUNT. Seems it just
> avoid using scripts/recordmcount.pl script and directly generate nops
> for _mcount.

Right.

> It's different from -fpatchable-function-entry=2 which generating nops
> before function prologue in arm64, isn't it?

Yes, fentry is for x86, but -fpatchable-function-entry=2 is making a
placeholder with nop at the entry of the functions for direct patching.

Thank you,

> 
> On Tue, Jul 21, 2020 at 9:27 PM Masami Hiramatsu <mhiramat@kernel.org> wrote:
> >
> > On Wed, 15 Jul 2020 00:24:54 +0800
> > Guo Ren <guoren@kernel.org> wrote:
> >
> > > Thx Masami,
> > >
> > > On Tue, Jul 14, 2020 at 7:38 PM Masami Hiramatsu <mhiramat@kernel.org> wrote:
> > > >
> > > > On Mon, 13 Jul 2020 23:39:21 +0000
> > > > guoren@kernel.org wrote:
> > > >
> > > > > From: Guo Ren <guoren@linux.alibaba.com>
> > > > >
> > > > > This patch adds support for kprobes on ftrace call sites to avoids
> > > > > much of the overhead with regular kprobes. Try it with simple
> > > > > steps:
> > > > >
> > > > > 1. Get _do_fork ftrace call site.
> > > > > Dump of assembler code for function _do_fork:
> > > > >    0xffffffe00020af64 <+0>:     addi    sp,sp,-128
> > > > >    0xffffffe00020af66 <+2>:     sd      s0,112(sp)
> > > > >    0xffffffe00020af68 <+4>:     sd      ra,120(sp)
> > > > >    0xffffffe00020af6a <+6>:     addi    s0,sp,128
> > > > >    0xffffffe00020af6c <+8>:     sd      s1,104(sp)
> > > > >    0xffffffe00020af6e <+10>:    sd      s2,96(sp)
> > > > >    0xffffffe00020af70 <+12>:    sd      s3,88(sp)
> > > > >    0xffffffe00020af72 <+14>:    sd      s4,80(sp)
> > > > >    0xffffffe00020af74 <+16>:    sd      s5,72(sp)
> > > > >    0xffffffe00020af76 <+18>:    sd      s6,64(sp)
> > > > >    0xffffffe00020af78 <+20>:    sd      s7,56(sp)
> > > > >    0xffffffe00020af7a <+22>:    mv      s4,a0
> > > > >    0xffffffe00020af7c <+24>:    mv      a0,ra
> > > > >    0xffffffe00020af7e <+26>:    nop   <<<<<<<< here!
> > > > >    0xffffffe00020af82 <+30>:    nop
> > > > >    0xffffffe00020af86 <+34>:    ld      s3,0(s4)
> > > > >
> > > > > 2. Set _do_fork+26 as the kprobe.
> > > > >   echo 'p:myprobe _do_fork+26 dfd=%a0 filename=%a1 flags=%a2 mode=+4($stack)' > /sys/kernel/debug/tracing/kprobe_events
> > > > >   echo 1 > /sys/kernel/debug/tracing/events/kprobes/enable
> > > > >   cat /sys/kernel/debug/tracing/trace
> > > > >   tracer: nop
> > > > >
> > > > >   entries-in-buffer/entries-written: 3/3   #P:1
> > > > >
> > > > >                                _-----=> irqs-off
> > > > >                               / _----=> need-resched
> > > > >                              | / _---=> hardirq/softirq
> > > > >                              || / _--=> preempt-depth
> > > > >                              ||| /     delay
> > > > >             TASK-PID   CPU#  ||||    TIMESTAMP  FUNCTION
> > > > >                | |       |   ||||       |         |
> > > > >               sh-87    [000] ....   551.557031: myprobe: (_do_fork+0x1a/0x2e6) dfd=0xffffffe00020af7e filename=0xffffffe00020b34e flags=0xffffffe00101e7c0 mode=0x20af86ffffffe0
> > > > >
> > > > >   cat /sys/kernel/debug/kprobes/list
> > > > > ffffffe00020af7e  k  _do_fork+0x1a    [FTRACE]
> > > > >                                        ^^^^^^
> > > >
> > > > Hmm, this seems fentry is not supported on RISC-V yet. But anyway,
> > > > it will be useful for users (if they can find the offset).
> > >
> > > Seems only x86 & ⬆️90 use fentry，can you elaborate more about fentry's
> > > benefit and how the user could set kprobe on ftrace call site without
> > > disassemble?
> >
> > On x86, the fentry replaces the mcount with just one call instruction, without
> > saving any arguments. This means all probes which are puts on the address of
> > target symbol, are automatically using ftrace. IOW, all probes on _do_fork+0
> > will use ftrace. We don't need any disassembling.
> >
> > I think if RISC-V already support "-fpatchable-function-entry=2" option on
> > GCC, you can easily enable it as same as arm64. See https://lkml.org/lkml/2019/6/18/648
> >
> > Thank you,
> >
> > --
> > Masami Hiramatsu <mhiramat@kernel.org>
> 
> 
> 
> -- 
> Best Regards
>  Guo Ren
> 
> ML: https://lore.kernel.org/linux-csky/


-- 
Masami Hiramatsu <mhiramat@kernel.org>