+36
−19
Loading
ACPICA: Update use of acpi_os_wait_events_complete interface.
This patch cleans up all of the acpi_os_wait_events_complete() invocations to make it to be invoked inside of ACPICA in the way to accommodate Linux's work queue implementation. According to the report, current Linux kernel code is facing a boot time race issue in the acpi_remove_notify_handler(). This is because: Linux is using work queues to implement a deferred handler call environment while ACPICA expects OSPM to implement acpi_os_wait_events_complete() using wait queues. The position to invoke a "waiter" is not suitable for a "flusher" as new invocations can be scheduled after this position and before the deletion of the handler from its management container. Since the following commit has deleted acpi_os_wait_events_complete() parameters, it thus might not be possible for OSPM to achieve a safe removal using wait queues. This requires ACPICA to be changed accordingly to "flush" handlers rather than "wait" them to be drain up: Commit: 5ff986a2a9db11858247b71fe242fe17617229aa Date: Wed, 16 May 2012 13:36:07 -0700 Subject: Introduce acpi_os_wait_events_complete interface. This interface will block until asynchronous events like notifies and GPEs are complete. Within ACPICA, it is called before a notify or GPE handler is removed. ACPICA BZ 868. This patch fixes this issue by invoking acpi_os_wait_events_complete() in the way to "flush" things - it thus should be put to the position after handler is removed from its management container but before it is destructed. The technical concerns are: 1. MTX_NAMESPACE is used to protect things that acpi_os_wait_events_complete() might be waiting for, thus MTX_NAMESPACE must be unlocked before invoking acpi_os_wait_events_complete(). 2. MTX_NAMESPACE is also used to implement the serialization of acpi_install_notify_handler() and acpi_remove_notify_handler(). This patch changes this logic, thus if there are many acpi_install/remove_notify_handler() invoked in parallel, the acpi_os_wait_events_complete() might face the races which could cause it never running to an end. Normally this will require additional code to implement a separate locking facility which is not implemented due to 3. 3. Given ACPICA users will always invoke acpi_install_notify_handler() once during Linux module/device initialization and invoke acpi_remove_notify_handler() once during module/device finalization, problem stated in 2 will not happen in Linux environment due to the mutual exclusive module/device existence, this fix thus is sufficient. Same concerns can apply to acpi_install/remove_gpe_handler(). Reported and tested: Ronald Vink. Fixed: Lv Zheng. References: https://bugzilla.kernel.org/show_bug.cgi?id=60583 Signed-off-by:Lv Zheng <lv.zheng@intel.com> Reported-and-Tested-by: