[llvm-dev] XRay TID mismatch when forking

Dean Michael Berris via llvm-dev llvm-dev at lists.llvm.org
Fri Jun 8 02:02:42 PDT 2018

Hi Henry,
On Thu, Jun 7, 2018 at 11:51 PM Henry Zhu via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> Hello,
> There seems to be a bug with the value of the thread ID (TID) that is generated by XRay when calling fork(). The value of the TID of the child has the same TID as its parent (this is seen in the “thread: “ field of the YAML output produced by the llvm-xray tool). This is a problem as the trace generated contains the same TID for the parent and children processes, making it difficult to distinguish between threads running in different processes.
> I am running linux on x86_64 machine and am using clang 7.0.
> Is this an issue that can be fixed? If not, is there any advice on how to make xray generate different TIDs when forking on my end?

XRay's basic mode doesn't get the thread ID for every event, and
actually caches the thread ID the first time it creates a thread-local
buffer it uses. This means because you effectively have a
single-threaded application, then forks a number of times, it will get
a copy of the thread ID in the parent process' thread.

There's ways to fix this but will make basic mode a tad more expensive
by getting the thread ID every time we write an event.

We can also attempt to write different files per-process, which means
it's also going to be a bit more expensive to check from the XRay
handler whether the process' ID is different from the cached data we
get from the thread-local structure.

The implementation of the Basic Mode handlers are in
-- in particular that's 'getThreadLocalData()' which caches the thread

Note that we haven't considered the multi-process use-case very much
with XRay instrumentation -- and that we may need to make some
adjustments on how to handle multi-process programming effectively.

We *could* have multi-process detection done somehow as an option. If
you have time to give that a shot, then I'd be happy to review the
changes to make that happen.



More information about the llvm-dev mailing list