Re: [Qemu-devel] Status of multithreading in user mode

[Laurent Vivier]

Le 23/02/2016 17:40, Cecilia Gonzalez-Alvarez a écrit :
> Hi all,

Hi Cecilia,

> I'm planning to use Qemu as a functional simulation frontend for a
> multicore simulator.
> 
> Although later plans include full-system simulation, I'd like to start,
> for the sake of simplicity, with the individual simulation of
> applications in user mode.
> 
> I'm interested in targets x86_64 and AArch64, and my general question
> is: how is the status of multithreading in user mode for those targets?
> Concretely, can we trust that running multiple threads on qemu-x86_64
> and qemu-aarch64 will be reliable most of the time?
> 
> Going through the mailing list I have the feeling that multithreading
> and user mode is a recurring issue.
> 
> I've done some experiments with the latest Qemu from the repository
> (2.5.50) and although user mode for both architectures is mostly working
> well, it fails with this particular example for qemu-x86_64:
> http://users.elis.ugent.be/~cgonzlez/fft.c
> 
> Compiled with GCC 4.8.4:
>     gcc -mno-sse4 -mno-sse4.1 -mno-sse4.2 -mno-sse4a -mno-avx -mno-avx2
> -c -o fft.o fft.c
>     gcc fft.o -lm -static -pthread  -o fft
> and executed with 32 threads:
>     qemu-x86_64 fft -p32
> sometimes, randomly, stalls on a sys_futex.
> 
> I don't know if this is a very concrete bug that I should report
> somewhere else; my main concern is to know if threading support in user
> mode is almost fully implemented and stable with just some random and
> unimportant bugs.

ARM has some atomic operations implemented by stopping all other CPU
threads (see linux-user/main.c, do_strex_a64()). But I don't think
locked operations have been implemented for x86_64.

And Michael has found a bug in the futex implementation in linux-user mode.

You can find a fix here:

https://github.com/vivier/qemu-m68k/commit/236f512f1c020e2d774216a9beb1466c260a89e2

This is not the original fix from Michael, but the idea is here and he
did all the work.

Laurent