On Tue, 22 Mar 2022 16:48:02 +0100 Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > On Tue, Mar 22, 2022 at 11:04:38AM -0400, Steven Rostedt wrote: > > > > In recap: > > > > > > __fentry__ -- push on trace-stack > > > __ftail__ -- mark top-most entry complete > > > __fexit__ -- mark top-most entry complete; > > > pop all completed entries > > > > Again, this would require that the tail-calls are also being traced. > > Which is why we should inhibit tail-calls if the function is notrace. > > > > inhibit tail-calls to notrace. > > > > Just inhibiting tail-calls to notrace would work without any of the above. > > I'm lost again; what? Without any of the above you got nothing because > return-trampoline will not work. I think this got "lost in translation". "Inhibiting tail-calls to notrace" Is a bit ambiguous because of the "to notrace" which would be different if I had said "on notrace" which I may have screwed up the grammar here. Let me be more precise. "Limiting tail-calls to only notrace functions" That I think is a bit less ambiguous. English sucks. > > > But my fear is that will cause a noticeable performance impact. > > Most code isn't in fact notrace, and call+ret aren't *that* expensive. "isn't in fact notrace" Ug! Double negatives! This gets even more confusing when we are saying "notrace" which is a negative. We should probably just say "traced" functions which makes communication a bit more straight forward. > > > > It's function graph tracing, kretprobes and whatever else this rethook > > > stuff is about that needs this because return trampolines will stop > > > working somewhere in the not too distant future. > > > > Another crazy solution is to have: > > > > func_A: > > call __fentry__ > > ... > > tail: jmp 1f > > call 1f > > > call __fexit__ > > ret > > 1: jmp func_B > > > > > > where the compiler tells us about "tail:" and that we know that func_A ends > > with a tail call, and if we want to trace the end of func_A we convert that > > jmp 1f into a nop. And then we call the func_B and it's return comes back > > to where we call __fexit__ and then return normally. > > At that point giving us something like: > > 1: > pushsection __ftail_loc > .long 1b - . > popsection > > jmp.d32 func_B > call __fexit__ > ret > > is smaller and simpler, we can patch the jmp.d32 to call when tracing. > The only problem is SLS, that might wants an int3 after jmp too > ( https://www.amd.com/en/corporate/product-security/bulletin/amd-sb-1026 ). > > That does avoid the need for __ftail__ I suppose. Which is basically what I said earlier ;-) https://lore.kernel.org/all/20220321122259.28146a7a@xxxxxxxxxxxxxxxxxx/ > Or maybe another solution is: > > funcA: > [..] > jmp funcB > call __fexit__ > ret > > And if funcA is being traced, we change jmp to a call. > > [..] > call funcB > call __fexit__ > ret > > Such that we only remove the tail calls if we enable tracing on the > function with the tail call. -- Steve
