[PATCH] D54889: Fiber support for thread sanitizer

[Dmitry Vyukov via Phabricator via llvm-commits]

dvyukov added a comment.

In D54889#1327234 <https://reviews.llvm.org/D54889#1327234>, @741g wrote:

> In D54889#1326849 <https://reviews.llvm.org/D54889#1326849>, @dvyukov wrote:
>
> > Lingfeng, thanks for extensive feedback.
> >
> > > This should allow that logic to work for QEMU at least, and it feels like it should work for all possible other uses (?)
> >
> > I am not sure there are no programs that use setjmp/longjmp to switch fibers across threads. Yes, man says behavior is undefined. But strictly saying, longjmp shouldn't be used for fiber switching too (it's solely for jumping upwards on the same stack). So I feel people are just relying on implementation details here (let's test it, works, ok, let's use it).
> >
> > You said that QEMU uses sigsetjmp/siglongjmp for coroutine switches. How does a coroutine always stay on the same thread then? Say, if they have a thread pool servicing all coroutines, I would expect that coroutines are randomly intermixed across threads. So it can happen that we setjmp on one thread, but then reschedule the coroutine with longjmp on another thread.
>
>
> Thanks for pointing that out. I have to check more closely, but I haven't observed migration of fibers across physical threads at all in qemu. The tendency is to both create and switch the coroutine on a dedicated thread. Swapcontext is done to a trampoline function, to which one can return again later on. Kind of like a fiber pool, where reuse of a fiber means longjmping to the trampoline instead of another swapcontext.
>
> I don't think there is a thread pool that mixes around the fibers. I guess this means we would have to annotate in the more general use case, but for QEMU, if we make the assumption that setjmp/longjmp follow the spec, while only allowing for physical thread switches in swapcontext, we should be OK (And physical thread switching in swapcontext is somethign QEMU does not do).


Makes sense.

If somebody will want to do longjmp to switch fibers across threads, we will need some additional support in tsan for that. Let's stick with Yuri and your use cases for now.


Repository:
  rCRT Compiler Runtime

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D54889/new/

https://reviews.llvm.org/D54889