[Mono] Race in init_method when using LLVM AOT. (#75088)
[Mono] Race in init_method when using LLVM AOT. When using LLVM AOT codegen, init_method updates two GOT slots. These slots are initialized as part of init_method, but there is a race between initialization of the two slots. Current implementation can have two threads running init_method for the same method, but as soon as: [got_slots [pindex]] = addr store is visible, it will trigger other threads to return back from init_method, but since that could happen before the corresponding LLVM AOT const slot is set, second thread will return to method calling init_method, load the LLVM aot const, and crash when trying to use it (since its still NULL). This crash is very rare but have been identified on x86/x64 CPU's, when one thread is either preempted between updating regular GOT slot and LLVM GOT slot or store into LLVM GOT slot gets delayed in store buffer. I have also been able to emulate the scenario in debugger, triggering the issue and crashing in the method loading from LLVM aot const slot. Fix change order of updates and make sure the update of LLVM aot const slot happens before memory_barrier, since: got [got_slots [pindex]] = addr; have release semantics in relation to addr and update of LLVM aot const slot. Fix also add acquire/release semantics for ji->type in init_method since it is used to guard if a thread ignores a patch or not and it should not be re-ordered with previous stores, since it can cause similar race conditions with updated slots.
Showing
想要评论请 注册 或 登录