Am 08.10.2018 um 21:53 hat Eric Blake geschrieben: > On 10/8/18 11:40 AM, Kevin Wolf wrote: > > Am 08.10.2018 um 17:43 hat Peter Maydell geschrieben: > > > Looking at the backtraces I'm wondering if this is the result of > > > an implicit reliance on the order in which per-thread destructors > > > are called (which is left unspecified by POSIX) -- the destructor > > > function qemu_thread_atexit_run() is called after some other > > > destructor, but accesses its memory. > > > > > > Specifically, the memory it's trying to read looks like > > > the __thread local variable pollfds_cleanup_notifier in > > > util/aio-posix.c. So I think what is happening is: > > > * util/aio-posix.c calls qemu_thread_atexit_add(), passing > > > it a pointer to a thread-local variable pollfds_cleanup_notifier > > > * qemu_thread_atexit_add() works by arranging to run the > > > notifiers when its 'exit_key' variable's destructor is called > > > * the destructor for pollfds_cleanup_notifier runs before that > > > for exit_key, and so the qemu_thread_atexit_run() function > > > ends up touching freed memory > > > > > > I'm pretty confident this analysis of the problem is correct: > > > unfortunately I have no idea what the right way to fix it is... > > > > Yes, I agree with your analysis. If __thread variables can be destructed > > before pthread_key_create() destructors are called (and in particular if > > the former are implemented in terms of the latter), this implies at > > least two rules: > > > > 1. The Notfier itself can't be a TLS variable > > > > 2. The notifier callback can't access any TLS variables > > > > Of course, with these restrictions, qemu_thread_atexit_*() with its > > existing API is as useless as it could be. > > > > The best I can think of at the moment would be to use a separate > > pthread_key_create() (and therefore a separate destructor) for > > registering each TLS variable, so that the destructor always gets a > > valid pointer. Maybe move all __thread variables of a file into a single > > malloced struct to make it more managable (we could then keep a __thread > > pointer to it for convenience, but only free the struct with the pointer > > passed by the pthread_key destructor so that we don't have to access > > __thread variables in the destructor). > > pthread_key_create() says that a when a destructor is triggered, it sets the > value of the key to NULL; but that you can once again set the key back to a > non-NULL value, and that the implementation will loop at least > PTHREAD_DESTRUCTOR_ITERATIONS over all destructors or until it has convinced > the destructors to leave values at NULL. Thus, while you cannot guarantee > ordering between destructors within a single iteration of the cleanup loop, > you CAN do some sort of witness locking or down-counter where a destructor > purposefully calls pthread_setspecific() to revive the value to survive into > the next iteration of destructor calls, for variables which are known to be > referenced by other destructors while the witness count is still high > enough, as a way of imposing order between loops.
Yes, everything that is explicitly managed with pthread_key_create() is fine. The point is that the destructors for __thread variables aren't controlled by QEMU, but by the system libraries, so their memory can go away earlier than other destructors are called. Kevin