From 8d5934952f26c25431dbc21aa8ae7614bb8f88c3 Mon Sep 17 00:00:00 2001 From: Erik Schmauss Date: Wed, 14 Mar 2018 16:12:58 -0700 Subject: ACPICA: Events: Add parallel GPE handling support to fix potential redundant _Exx evaluations There is a risk that a GPE method/handler may be invoked twice. Let's consider a case, both GPE0(RAW_HANDLER) and GPE1(_Exx) is triggered. =======================================+============================= IRQ handler (top-half) |IRQ polling =======================================+============================= acpi_ev_detect_gpe() | LOCK() | READ (GPE0-7 enable/status registers)| ^^^^^^^^^^^^ROOT CAUSE^^^^^^^^^^^^^^^| Walk GPE0 | UNLOCK() |LOCK() Invoke GPE0 RAW_HANDLER |READ (GPE1 enable/status bit) |acpi_ev_gpe_dispatch(irq=false) | CLEAR (GPE1 enable bit) | CLEAR (GPE1 status bit) LOCK() |UNLOCK() Walk GPE1 +============================= acpi_ev_gpe_dispatch(irq=true) |IRQ polling (defer) CLEAR (GPE1 enable bit) +============================= CLEAR (GPE1 status bit) |acpi_ev_async_execute_gpe_method() Walk others | Evaluate GPE1 _Exx fi | acpi_ev_async_enable_gpe() UNLOCK() | LOCK() =======================================+ SET (GPE enable bit) IRQ handler (bottom-half) | UNLOCK() =======================================+ acpi_ev_async_execute_gpe_method() | Evaluate GPE1 _Exx | acpi_ev_async_enable_gpe() | LOCK() | SET (GPE1 enable bit) | UNLOCK() | =======================================+============================= If acpi_ev_detect_gpe() is only invoked from the IRQ context, there won't be more than one _Lxx/_Exx evaluations for one status bit flagging if the IRQ handlers controlled by the underlying IRQ chip/driver (ex. APIC) are run in serial. Note that, this is a known potential gap and we had an approach, locking entire non-raw-handler processes in the top-half IRQ handler and handling all raw-handlers out of the locked loop to be friendly to those IRQ chip/driver. But the approach is too complicated while the issue is not so real, thus ACPICA treated such issue (if any) as a parallelism/quality issue of the underlying IRQ chip/driver to stop putting it on the radar. Bug in link #1 is suspiciously reflecting the same cause, and if so, it can also be fixed by this simpler approach. But it will be no excuse an ACPICA problem now if ACPICA starts to poll IRQs itself. In the changed scenario, _Exx will be evaluated from the task context due to new ACPICA provided "polling after enabling GPEs" mechanism. And the above figure uses edge-triggered GPEs demonstrating the possibility of evaluating _Exx twice for one status bit flagging. As a conclusion, there is now an increased chance of evaluating _Lxx/_Exx more than once for one status bit flagging. However this is still not a real problem if the _Lxx/_Exx checks the underlying hardware IRQ reasoning and finally just changes the 2nd and the follow-up evaluations into no-ops. Note that _Lxx should always be written in this way as a level-trigger GPE could have it's status wrongly duplicated by the underlying IRQ delivery mechanisms. But _Exx may have very low quality BIOS by BIOS to trigger real issues. For example, trigger duplicated button notifications. To solve this issue, we need to stop reading a bunch of enable/status register bits, but read only one GPE's enable/status bit. And GPE status register's W1C nature ensures that acknowledging one GPE won't affect another GPEs' status bits. Thus the hardware GPE architecture has already provided us with the mechanism of implementing such parallelism. So we can lock around one GPE handling process to achieve the parallelism: 1. If we can incorporate GPE enable bit check in detection and ensure the atomicity of the following process (top-half IRQ handler): READ (enable/status bit) if (enabled && raised) CLEAR (enable bit) and handle the GPE after this process, we can ensure that we will only invoke GPE handler once for one status bit flagging. 2. In addtion for edge-triggered GPEs, if we can ensure the atomicity of the following process (top-half IRQ handler): READ (enable/status bit) if (enabled && raised) CLEAR (enable bit) CLEAR (status bit) and handle the GPE after this process, we can ensure that we will only invoke GPE handler once for one status bit flagging. By doing a cleanup in this way, we can remove duplicate GPE handling code and ensure that all logics are collected in 1 function. And the function will be safe for both IRQ interrupt and IRQ polling, and will be safe for us to release and re-acquire acpi_gbl_gpe_lock at any time rather than raw handler only during the top-half IRQ handler. Lv Zheng. Link: https://bugzilla.kernel.org/show_bug.cgi?id=196703 [#1] Signed-off-by: Lv Zheng Signed-off-by: Erik Schmauss Signed-off-by: Rafael J. Wysocki --- drivers/acpi/acpica/acevents.h | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'drivers/acpi/acpica/acevents.h') diff --git a/drivers/acpi/acpica/acevents.h b/drivers/acpi/acpica/acevents.h index 1b0269f6ac2d..fa68544c7c07 100644 --- a/drivers/acpi/acpica/acevents.h +++ b/drivers/acpi/acpica/acevents.h @@ -103,6 +103,10 @@ struct acpi_gpe_event_info *acpi_ev_low_get_gpe_info(u32 gpe_number, acpi_status acpi_ev_finish_gpe(struct acpi_gpe_event_info *gpe_event_info); +u32 +acpi_ev_detect_gpe(struct acpi_namespace_node *gpe_device, + struct acpi_gpe_event_info *gpe_event_info, u32 gpe_number); + /* * evgpeblk - Upper-level GPE block support */ -- cgit v1.2.3