Am 22.02.2022 um 15:01 hat Stefan Hajnoczi geschrieben: > v5: > - Added explicit "#include "qemu/coroutine-tls.h" in patch 4 [Philippe] > - Updated patch 1 commit description and comments to describe the current > noinline plus asm volatile approach [Peter] > v4: > - Dropped '[RFC]'. > - Dropped inline asm for now. -fPIC versions of the code are missing and I > hit several issues including a clang LTO bug where thread local variables > are > incorrectly discarded because inline asm is not analyzed to find symbol > dependencies (Serge Guelton is aware). > - Fixed CI failures. > v3: > - Added __attribute__((weak)) to get_ptr_*() [Florian] > - Replace rdfsbase with mov %%fs:0,%0 [Florian] > > This patch series solves the coroutines TLS problem. Coroutines re-entered > from > another thread sometimes see stale TLS values. This happens because compilers > may cache values across yield points, so a value from the previous thread will > be used when the coroutine is re-entered in another thread. > > Serge Guelton developed a portable technique, see the first patch for details. > > I have audited all __thread variables in QEMU and converted those that can be > used from coroutines. Most actually look safe to me.
Hm, what about the ones in the coroutine implementation itself? static __thread CoroutineUContext leader; static __thread Coroutine *current; Both of them are used in qemu_coroutine_self(), which is a coroutine_fn, and in qemu_in_coroutine(), which may be called from coroutine context. And I seem to remember I've seen crashes related to this in one of the bug reports we got, where the stack trace clearly showed that one of these functions had returned a wrong result. I'm applying this series anyway, it doesn't make the patches incorrect. But it feels incomplete, so we may need a follow-up patch. Kevin
