Re: [PATCH v6] arm64: implement KPROBES_ON_FTRACE

From: Masami Hiramatsu
Date: Tue Dec 24 2019 - 05:09:25 EST


Hi Jisheng,

On Mon, 23 Dec 2019 07:47:24 +0000
Jisheng Zhang <Jisheng.Zhang@xxxxxxxxxxxxx> wrote:

> Hi Masami,
>
> On Wed, 18 Dec 2019 22:25:50 +0900 Masami Hiramatsu wrote:
>
>
> >
> >
> > On Wed, 18 Dec 2019 06:21:35 +0000
> > Jisheng Zhang <Jisheng.Zhang@xxxxxxxxxxxxx> wrote:
> >
> > > KPROBES_ON_FTRACE avoids much of the overhead with regular kprobes as it
> > > eliminates the need for a trap, as well as the need to emulate or
> > > single-step instructions.
> > >
> > > Tested on berlin arm64 platform.
> > >
> > > ~ # mount -t debugfs debugfs /sys/kernel/debug/
> > > ~ # cd /sys/kernel/debug/
> > > /sys/kernel/debug # echo 'p _do_fork' > tracing/kprobe_events
> > >
> > > before the patch:
> > >
> > > /sys/kernel/debug # cat kprobes/list
> > > ffffff801009fe28 k _do_fork+0x0 [DISABLED]
> > >
> > > after the patch:
> > >
> > > /sys/kernel/debug # cat kprobes/list
> > > ffffff801009ff54 k _do_fork+0x4 [DISABLED][FTRACE]
> >
> > BTW, it seems this automatically changes the offset without
> > user's intention or any warnings. How would you manage if the user
> > pass a new probe on _do_fork+0x4?
>
> In current implementation, two probes at the same address _do_fork+0x4

OK, that is my point.

> > IOW, it is still the question who really wants to probe on
> > the _do_fork+"0", if kprobes modifies it automatically,
> > no one can do that anymore.
> > This can be happen if the user want to record LR or SP value
> > at the function call for debug. If kprobe always modifies it,
> > we will lose the way to do it.
>
> arm64's DYNAMIC_FTRACE_WITH_REGS implementation makes use of GCC
> -fpatchable-function-entry=2 option to insert two nops. When the function
> is traced, the first nop will be modified to the LR saver, then the
> second nop to "bl <ftrace-entry>", commit 3b23e4991fb6("
> arm64: implement ftrace with regs") explains the mechanism.

So both of the instruction at func+0 and func+4 are replaced.

> So on arm64(in fact any arch makes use of -fpatchable-function-entry will
> behave similarly), when DYNAMIC_FTRACE_WITH_REGS is enabled, the ftrace location
> is always on the first 4 bytes offset.
>
> I think when users want to register a kprobe on _do_fork+0, what he really want
> is to kprobe on the patched(by -fpatchable-function-entry)_do_fork+4

OK, in this case, kprobe should treat the first 2 instructions as a
single virtual instruction. This means,

- kprobes can probe func+0, but not func+4 if the function is ftraced.
(-EILSEQ must be returned)
- both debugfs/kprobes/list and tracefs/kprobe_events should show the
probed address as func+0. Not func+4.

Then, user can use kprobes as if there is one big (8-byte) instruction
at the entry of the function. Since probing on func+4 is rejected and
the call-site LR/SP is restored in ftrace, there should be no
contradiction. It should work as if we put a breakpoint on the func + 0.

>
> PS: per my understanding, powerpc's kprobes_on_ftrace also introduces an
> extra automatic offset due to its implementation.

Uh, that is also ugly.... must be fixed.


> >
> > Could you remove below function at this moment?
> >
> > > +kprobe_opcode_t *kprobe_lookup_name(const char *name, unsigned int offset)
> > > +{
> > > + unsigned long addr = kallsyms_lookup_name(name);
> > > +
> > > + if (addr && !offset) {
> > > + unsigned long faddr;
> > > + /*
> > > + * with -fpatchable-function-entry=2, the first 4 bytes is the
> > > + * LR saver, then the actual call insn. So ftrace location is
> > > + * always on the first 4 bytes offset.
> > > + */
> > > + faddr = ftrace_location_range(addr,
> > > + addr + AARCH64_INSN_SIZE);
> > > + if (faddr)
> > > + return (kprobe_opcode_t *)faddr;
> > > + }
> > > + return (kprobe_opcode_t *)addr;
> > > +}
> > > +
> > > +bool arch_kprobe_on_func_entry(unsigned long offset)
> > > +{
> > > + return offset <= AARCH64_INSN_SIZE;
> > > +}
> >
> >
> > Without this automatic change, we still can change the offset
> > in upper layer.
>
> If remove the two functions, kprobe on _do_fork can't ride on
> ftrace infrastructure, but kprobe on _do_fork+4 can. I'm not sure
> whether this is reasonable. Every kprobe users on arm64 will need to
> remember to pass an extra offset +4 to make use of kprobe_on_ftrace, could
> we hide the "+4"?

Yes, OK, as I said above, please hide +4. We will see the virtual
"call" instruction (= "mov x9, lr; br <addr>") at the entry of ftraced
function. :)

Thank you,


--
Masami Hiramatsu <mhiramat@xxxxxxxxxx>