Re: [PATCH v4 5/9] ftrace: Add ftrace_partial_regs() for converting ftrace_regs to pt_regs

[Google]

On Tue, 5 Sep 2023 12:50:28 -0700
Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote:

> On Fri, Aug 25, 2023 at 6:56 PM Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> >
> > On Fri, 25 Aug 2023 14:49:48 -0700
> > Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote:
> >
> > > On Wed, Aug 23, 2023 at 8:16 AM Masami Hiramatsu (Google)
> > > <mhiramat@xxxxxxxxxx> wrote:
> > > >
> > > > From: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> > > >
> > > > Add ftrace_partial_regs() which converts the ftrace_regs to pt_regs.
> > > > If the architecture defines its own ftrace_regs, this copies partial
> > > > registers to pt_regs and returns it. If not, ftrace_regs is the same as
> > > > pt_regs and ftrace_partial_regs() will return ftrace_regs::regs.
> > > >
> > > > Signed-off-by: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> > > > Acked-by: Florent Revest <revest@xxxxxxxxxxxx>
> > > > ---
> > > >  Changes in v3:
> > > >   - Fix to use pt_regs::regs instead of x.
> > > >   - Return ftrace_regs::regs forcibly if HAVE_PT_REGS_COMPAT_FTRACE_REGS=y.
> > > >   - Fix typo.
> > > >   - Fix to copy correct registers to the pt_regs on arm64.
> > > >  Changes in v4:
> > > >   - Change the patch order in the series so that fprobe event can use this.
> > > > ---
> > > >  arch/arm64/include/asm/ftrace.h |   11 +++++++++++
> > > >  include/linux/ftrace.h          |   17 +++++++++++++++++
> > > >  2 files changed, 28 insertions(+)
> > > >
> > > > diff --git a/arch/arm64/include/asm/ftrace.h b/arch/arm64/include/asm/ftrace.h
> > > > index ab158196480c..5ad24f315d52 100644
> > > > --- a/arch/arm64/include/asm/ftrace.h
> > > > +++ b/arch/arm64/include/asm/ftrace.h
> > > > @@ -137,6 +137,17 @@ ftrace_override_function_with_return(struct ftrace_regs *fregs)
> > > >         fregs->pc = fregs->lr;
> > > >  }
> > > >
> > > > +static __always_inline struct pt_regs *
> > > > +ftrace_partial_regs(const struct ftrace_regs *fregs, struct pt_regs *regs)
> > > > +{
> > > > +       memcpy(regs->regs, fregs->regs, sizeof(u64) * 9);
> > > > +       regs->sp = fregs->sp;
> > > > +       regs->pc = fregs->pc;
> > > > +       regs->regs[29] = fregs->fp;
> > > > +       regs->regs[30] = fregs->lr;
> > >
> > > I see that orig_x0 from pt_regs is used on arm64 to get syscall's
> > > first parameter. And it seems like this ftrace_regs to pt_regs
> > > conversion doesn't touch orig_x0 at all. Is it maintained/preserved
> > > somewhere else, or will we lose the ability to get the first syscall's
> > > argument?
> >
> > Thanks for checking it!
> >
> > Does BPF uses kprobe probe to trace syscalls? Since we have raw_syscall
> > trace events, no need to use kprobe to do that. (and I don't recommend to
> > use kprobe to do such fixed event)
> 
> Yeah, lots of tools and projects actually trace syscalls with kprobes.
> I don't think there is anything we can do to quickly change that, so
> we should avoid breaking all of them.

Yes, ah, but anyway, this is the fprobe case, not kprobe. Do you use
multi_kprobes for tracing syscalls? 

Jiri, do you know when the multi-kprobe feature is used? Is that used
implicitly or explicitly?

> 
> So getting back to my original question, is it possible to preserve orig_x0?

I'm curious that the orig_x0 is stored to pt_regs of the kprobes itself
because it is not a real register. There is no way to access it. You can
use regs->x0 instead of that.

I think the orig_x0 is stored when the syscall happened because it has
another user pt_regs on the stack, right?

If so, we don't need to save orig_x0 on the pt_regs for kprobes, but user can
dig the stack to find the orig_x0. Here the arm64, syscall entry handler,

static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
                           const syscall_fn_t syscall_table[])
{
        unsigned long flags = read_thread_flags();

        regs->orig_x0 = regs->regs[0];
        regs->syscallno = scno;

(BTW, it seems syscall number is saved on regs->syscallno.)

It seems that if you probe the el0_svc_common() for tracing syscall,
all you need is tracing $arg1 == pt_regs and $arg2 == syscall number.

Thank you,

> 
> >
> > >
> > > Looking at libbpf's bpf_tracing.h, other than orig_x0, I think all the
> > > other registers are still preserved, so this seems to be the only
> > > potential problem.
> >
> > Great!
> >
> > Thank you,
> >
> > >
> > >
> > > > +       return regs;
> > > > +}
> > > > +
> > > >  int ftrace_regs_query_register_offset(const char *name);
> > > >
> > > >  int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec);
> > > > diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h
> > > > index c0a42d0860b8..a6ed2aa71efc 100644
> > > > --- a/include/linux/ftrace.h
> > > > +++ b/include/linux/ftrace.h
> > > > @@ -165,6 +165,23 @@ static __always_inline struct pt_regs *ftrace_get_regs(struct ftrace_regs *fregs
> > > >         return arch_ftrace_get_regs(fregs);
> > > >  }
> > > >
> > > > +#if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) || \
> > > > +       defined(CONFIG_HAVE_PT_REGS_TO_FTRACE_REGS_CAST)
> > > > +
> > > > +static __always_inline struct pt_regs *
> > > > +ftrace_partial_regs(struct ftrace_regs *fregs, struct pt_regs *regs)
> > > > +{
> > > > +       /*
> > > > +        * If CONFIG_HAVE_PT_REGS_TO_FTRACE_REGS_CAST=y, ftrace_regs memory
> > > > +        * layout is the same as pt_regs. So always returns that address.
> > > > +        * Since arch_ftrace_get_regs() will check some members and may return
> > > > +        * NULL, we can not use it.
> > > > +        */
> > > > +       return &fregs->regs;
> > > > +}
> > > > +
> > > > +#endif /* !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS || CONFIG_HAVE_PT_REGS_TO_FTRACE_REGS_CAST */
> > > > +
> > > >  /*
> > > >   * When true, the ftrace_regs_{get,set}_*() functions may be used on fregs.
> > > >   * Note: this can be true even when ftrace_get_regs() cannot provide a pt_regs.
> > > >
> > > >
> >
> >
> > --
> > Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>


-- 
Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>