Re: [PATCH 3/4] x86/ftrace: make ftrace_int3_handler() not to skip fops invocation

[Nicolai Stange]

Steven Rostedt <rostedt@xxxxxxxxxxx> writes:

> On Mon, 29 Apr 2019 14:38:35 -0700
> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
>> On Mon, Apr 29, 2019 at 1:30 PM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>> >
>> > The update from "call custom_trampoline" to "call iterator_trampoline"
>> > is where we have an issue.  
>> 
>> So it has never worked. Just tell people that they have two chocies:
>
> The custom call to iterator translation never worked. One option is to
> always call the iterator, and just take the hit. There's another
> solution to to make permanent updates that would handle the live
> patching case, but not for the tracing case. It is more critical for
> live patching than tracing (missed traced function is not as critical
> as running buggy code unexpectedly). I could look to work on that
> instead.

Making the live patching case permanent would disable tracing of live
patched functions though?

For unconditionally calling the iterator: I don't have numbers, but
would expect that always searching through something like 50-70 live
patching ftrace_ops from some hot path will be noticeable.


> If Nicolai has time, perhaps he can try out the method you suggest and
> see if we can move this forward.

You mean making ftrace handle call->call transitions one by one in
non-batch mode, right? Sure, I can do that.

Another question though: is there anything that prevents us from calling
ftrace_ops_list_func() directly from ftrace_int3_handler()? We don't
have parent_ip, but if that is used for reporting only, maybe setting it
to some ftrace_is_in_int3_and_doesnt_now_the_parent() will work?
Graph tracing entries would still be lost, but well...

Thanks,

Nicolai