Re: [PATCH 3/4] ftrace: abstract DYNAMIC_FTRACE_WITH_ARGS accesses

[Google]

On Tue, 25 Oct 2022 15:20:57 +0200
Florent Revest <revest@xxxxxxxxxxxx> wrote:

> On Tue, Oct 25, 2022 at 12:30 PM Mark Rutland <mark.rutland@xxxxxxx> wrote:
> >
> > On Tue, Oct 25, 2022 at 05:40:01PM +0900, Masami Hiramatsu wrote:
> > > Hi Mark,
> > >
> > > On Mon, 24 Oct 2022 15:08:45 +0100
> > > Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > >
> > > > In subsequent patches we'll arrange for architectures to have an
> > > > ftrace_regs which is entirely distinct from pt_regs. In preparation for
> > > > this, we need to minimize the use of pt_regs to where strictly necessary
> > > > in the core ftrace code.
> > > >
> > > > This patch adds new ftrace_regs_{get,set}_*() helpers which can be used
> > > > to manipulate ftrace_regs. When CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS=y,
> > > > these can always be used on any ftrace_regs, and when
> > > > CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS=n these can be used when regs are
> > > > available. A new ftrace_regs_has_args(fregs) helper is added which code
> > > > can use to check when these are usable.
> > >
> > > Can you also add the ftrace_regs_query_register_offset() as a wrapper of
> > > regs_query_register_offset()? I would like to use it for fprobe_events.
> >
> > Sure!
> >
> > Just to check, with FTRACE_WITH_REGS, does fprobe always sample the full
> > pt_regs, or do callers also need to check ftrace_regs_has_args(fregs)?

No, if the register is NOT saved, just return -ENOENT.

> 
> Currently, we have:
> config FPROBE
>     depends on DYNAMIC_FTRACE_WITH_REGS
> 
> Because fprobe registers its ftrace ops with FTRACE_OPS_FL_SAVE_REGS
> and calls ftrace_get_regs() to give pt_regs to registered fprobe
> callbacks.
> 
> We'll need to refactor fprobe a bit to support ||
> DYNAMIC_FTRACE_WITH_ARGS and therefore work on arm64.

Yeah, that is what I will do.

> > I ask because if neither of those are the case, with FTRACE_WITH_REGS,
> > ftrace_regs_query_register_offset() would accept names of registers which might
> > not have been sampled, and could give offsets to uninitialized memory.
> 
> Indeed, if one were to call the regs_query_register_offset() on a
> pt_regs that comes out of a ftrace trampoline with FTRACE_WITH_REGS,
> for example in a fprobe callback, one could get offsets to
> uninitialized memory already (for the registers we can't get outside
> of an exception handler on arm64 for example iiuc)
> 
> And it'd get way worse with FTRACE_WITH_ARGS if we implement
> ftrace_regs_query_register_offset() by calling
> regs_query_register_offset() for ftrace_regs that contain sparse
> pt_regs.

Ah, I got what you meant. If you consider that case, it can return
-ENOTSUPPORT for FTRACE_WITH_ARGS case at this moment. I'll fill it
afterwards.

> 
> > Atop that, I'm not exactly sure what to implement for powerpc/s390/x86 here. If
> > those might be used without a full pt_regs, I think
> > ftrace_regs_query_register_offset() should also take the fregs as a parameter
> > and use that to check which registers are available.
> 
> I think it would make sense for a ftrace_regs_query_register_offset()
> to only return offsets to the registers that are actually saved by the
> arch in a ftrace_regs (whether that's WITH_ARGS or WITH_REGS).
> 
> But that also means that if we introduce "fprobe_events" in the
> tracing sysfs interface, we can't have it support a %REG syntax
> compatible with the one in "kprobe_events" anyway.
> 
> Masami, how about having "fprobe_events" only support $argN, $retval
> etc but no %REG, from the beginning ? Then it would be clear to users
> that fprobe can not guarantee registers and we'd never have to fake
> registers when we don't have them. Users would have to make a decision
> between using fprobe which is fast but only has arguments and return
> value and kprobe which is slow but has all registers.

Hmm, for the first implementation, it is OK. But later I need to
implement the register access, because, debuginfo maps it differently.
I would like to arrow perf-probe to set it up eventually.

> I realize this has consequences for the kretprobe and rethook
> unification plan but if we have fprobe_events support %REG at the
> beginning, we'd have to break it at some point down the line anyway
> right ?

No, because %REG support doesn't mean we guarantee to access
all registers. We can just use %REG as alias of $argN :)

Thank you,

> 
> > ... does that make sense to you?
> >
> > Thanks,
> > Mark.


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