On Tue, 22 Mar 2022 22:42:36 -0400 Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: > On Wed, 23 Mar 2022 11:23:23 +0900 > Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote: > > > I see the __fexit__ is needed, but why __ftail__ is needed? I guess because > > func_B is notrace, in that case the __fexit__ will not be in the func_B. > > Am I correct? > > I believe Peter and I agreed that the "best" solution so far, that has the > least amount of regressions (doesn't remove anything currently being > function graph traced, nor removes current tail calls) is: > > > At that point giving us something like: > > > > 1: > > pushsection __ftail_loc > > .long 1b - . > > popsection > > > > jmp.d32 func_B > > call __fexit__ > > ret > > > Functions with a tail call will not have a __fexit__ and we can not rely on > the function that is the tail call to do the __fexit__ for the parent > function. Thus, the compromise is to add a label where the jmp to the > tail-call function is, and when we want to trace the return of that > function, we first have to patch the jmp into a call, which will then > return back to the call to __fexit__. Got it. So the tail call "jump" will be replaced with a normal call when we trace it. That's a good idea :) > > -- Steve -- Masami Hiramatsu <mhiramat@xxxxxxxxxx>
