-
The original atomic implementation in pulseaudio based on libatomic stated that the intent was to use full memory barriers. According to [1], the load and store implementation based on gcc builtins matches sequential consistent (i.e. full memory barrier) load and store ordering only for x86. I observed random crashes in client applications using memfd srbchannel transport on an armv8-aarch64 platform (cortex-a57). In all those crashes the first read on the pstream descriptor (the size field) was wrong and looked like it contained old data. I boiled the relevant parts of the srbchannel implementation down to a simple test case and could observe random test failures. So I figured that the atomic implementation was broken for armv8 with respect to cross-cpu memory access ordering consistency. In order to come up with a minimal fix, I used the newer __atomic_load_n/__atomic_store_n builtins from gcc. With aarch64-linux-gnu-gcc (Linaro GCC 7.3-2018.05) 7.3.1 20180425 they compile to ldar and stlxr on arm64, which is correct according to [1] and [2]. The other atomic operations based on __sync builtins don't need to be touched since they already are of the full memory barrier variety. [1] https://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html [2] <https://community.arm.com/developer/ip-products/processors /b/processors-ip-blog/posts/armv8-a-architecture-2016-additions>
d4ff4adc
