https://bugs.freedesktop.org/show_bug.cgi?id=89366
Bug ID: 89366 Summary: DisplayPort MST (multi-stream transport) "atomic sleep" Linux kernel bug Product: DRI Version: unspecified Hardware: x86 (IA32) OS: Linux (All) Status: NEW Severity: normal Priority: medium Component: General Assignee: dri-devel@lists.freedesktop.org Reporter: adam_richter2004@yahoo.com
[This is a slightly edited version of an email that I attempted to send to the dri-devel mailing list.]
CONFIG_DEBUG_ATOMIC_SLEEP complains about the following locking problem in linux-4.0-rc1/drivers/gpu/drm/drm_dp_mst_topology.c:
drm_dp_mst_wait_tx_reply --> wait_event_timeout --> check_txmsg_state --> mutex_lock
I believe that any function called in the "condition" argument in the wait_event_timeout macro (in this case, check_txmsg_state) is not allowed to block when the condition is being evalutated to determine whether to unblock the process.
I think the problem is real. On two different computers and three different DisplayPort MST hubs, plugging in a DisplayPort hub or having it plugged in from boot time results in a hang within a few minutes of doing a few "xrandr" commands.
At first glance, it looked to me like it might be safe to remove the mutex_{,un}lock calls from check_txmsg_state (which is not called from anywhere else), and change the integer field txmsg->state to be an atomic_t (although I'd be surprised if there is existing hardware that supports an MST hub where the accessing that field is not atomic. However, altough removing those mutex calls eliminated the complaint from CONFIG_DEBUG_ATOMIC_SLEEP, it also resulted in the system sometimes seeming to ignore the MST hub and otherwise eventually getting a kernel memory fault in the DisplayPort MST code or another spontaneous reset (possibly deadlock follwed by a watchdog reset).
Advice is welcome, although I am not blocked in my own efforts to analyze this further. I plan to post updates when I have more news.