On Tue, 2 May 2023 10:40:28 -0700 Song Liu <song@xxxxxxxxxx> wrote: > We hit the following hiccup a couple times in the past few months: > > A function is marked as "inline", but the compiler decided not to inline it. > Since "inline" implies "notrace" since [1], this function doesn't have > fentry/mcount. When we built livepatch for this function, kpatch-build > failed with: > > xxx.o: function yyy has no fentry/mcount call, unable to patch > > This is not a deal breaker, as we can usually modify the patch to work > around it. But I wonder whether we still need "inline" to imply "notrace". > Can we remove this to make livepatch a little easier? The history behind this is that there were cases that functions that were inlined, were in critical paths that could cause crashes if they were traced. In testing they never triggered because the developer's compiler inlined them. Then on someone else's machine, the compiler decided not to inline the function and the system crashed. It was hell to debug because I was not able to reproduce the issue, as my compiler would always keep the function inlined! But a lot has changed since then. I've implemented the "ftrace_test_recursion_trylock()" that catches pretty much all recursive bugs (and can ever report when they happen). So I may be open to removing the "notrace" from "inline". -- Steve
