[llvm-dev] XRay: Demo on x86_64/Linux almost done; some questions.
Dean Michael Berris via llvm-dev
llvm-dev at lists.llvm.org
Thu Aug 4 00:57:41 PDT 2016
> On 4 Aug 2016, at 06:27, Serge Rogatch <serge.rogatch at gmail.com> wrote:
> Hi Dean,
> I have a question about the following piece of code in compiler-rt/trunk/lib/xray/xray_trampoline_x86.S :
> movq _ZN6__xray19XRayPatchedFunctionE(%rip), %rax
> testq %rax, %rax
> je .Ltmp0
> // assume that %r10d has the function id.
> movl %r10d, %edi
> xor %esi,%esi
> callq *%rax
> What happens if someone unsets the handler function (i.e. calls __xray_remove_handler() ) or changes the handler (i.e. calls __xray_set_handler() with a different pointer to function) between "movq _ZN6__xray19XRayPatchedFunctionE(%rip), %rax" and "callq *%rax" ? I understood that the old handler will still be called, but isn't it a problem? What if the code removing the handler starts destructing some objects after that, so that a handler call would result in a crash or other bad things?
That would be unfortunate. :)
Yes, you're right, it will be called despite having the handler be un-set from the global atomic. Though I'd suggest that log handler implementations really shouldn't be crashing, and should assume that it will be called at any given time (from multiple threads, so it should also be thread-safe). Consider also that we're taking a function pointer by design -- assuming that it will be a pointer to a function that will only deal with globals and the explicitly passed in arguments.
An implementer of the log handler should really be responsible for "internal signalling" -- i.e. implementing it so that any required synchronisation happens internally.
Does this help?
PS. I think more definitive documentation for this is due, let me think about putting something some documentation on how to use XRay as it is currently implemented once we have the naive in-memory logging implementation land (https://reviews.llvm.org/D21982).
More information about the llvm-dev