On 2018-11-02, Steven Rostedt wrote: > > As an aside, I just tested with the frame unwinder and it isn't thrown > > off-course by kretprobe_trampoline (though obviously the stack is still > > wrong). So I think we just need to hook into the ORC unwinder to get it > > to continue skipping up the stack, as well as add the rewriting code for > > the stack traces (for all unwinders I guess -- though ideally we should > > I agree that this is the right solution. > > > do this without having to add the same code to every architecture). > > True, and there's an art to consolidating the code between > architectures. > > I'm currently looking at function graph and seeing if I can consolidate > it too. And I'm also trying to get multiple uses to hook into its > infrastructure. I think I finally figured out a way to do so. > > The reason it is difficult, is that you need to maintain state between > the entry of a function and the exit for each task and callback that is > registered. Hence, it's a 3x tuple (function stack, task, callbacks). > And this must be maintained with preemption. A task may sleep for > minutes, and the state needs to be retained. > > The only state that must be retained is the function stack with the > task, because if that gets out of sync, the system crashes. But the > callback state can be removed. > > Here's what is there now: > > When something is registered with the function graph tracer, every > task gets a shadowed stack. A hook is added to fork to add shadow > stacks to new tasks. Once a shadow stack is added to a task, that > shadow stack is never removed until the task exits. > > When the function is entered, the real return code is stored in the > shadow stack and the trampoline address is put in its place. > > On return, the trampoline is called, and it will pop off the return > code from the shadow stack and return to that. I was playing with a toy version of this using the existing kretprobe architecture (by borrowing the shadow stack concept of storing the stack_addr() -- though it's actually stored inside the existing kretprobe_instance and thus doesn't need separate push-pop code). The rewriting of kretprobe_trampoline in the stack unwinder *works* on x86 now except for the regs handling (the current function is still incorrectly shown as kretprobe_trampoline). This is actually a general bug in ftrace as well, because (for instance) while the unwinder calls into ftrace_graph_ret_addr() while walking up the stack this isn't used to correct regs->ip in perf_callchain_kernel(). I think this is the cause of the bug in ftrace-graph (if you switch to the "frame" unwinder you might be able to see it more clearly) -- but when trying to fix it I'm having trouble figuring out what retp should be (stack_addr(regs) doesn't give the right result for some reason). I will try to fix it up and attach it, but if you're already working on a prototype the unifies all the users that works too. The patch I have at the moment duplicates each of the key ftrace_graph_return_addr invocations with a matching kretprobe_return_addr (though there's only three of these). -- Aleksa Sarai Senior Software Engineer (Containers) SUSE Linux GmbH