From 74b51ee152b6d99e61ba329799a039453fb9438f Mon Sep 17 00:00:00 2001 From: Konstantin Khlebnikov Date: Sun, 9 Nov 2014 13:53:37 +0400 Subject: ACPI / osl: speedup grace period in acpi_os_map_cleanup ACPI maintains cache of ioremap regions to speed up operations and access to them from irq context where ioremap() calls aren't allowed. This code abuses synchronize_rcu() on unmap path for synchronization with fast-path in acpi_os_read/write_memory which uses this cache. Since v3.10 CPUs are allowed to enter idle state even if they have RCU callbacks queued, see commit c0f4dfd4f90f1667d234d21f15153ea09a2eaa66 ("rcu: Make RCU_FAST_NO_HZ take advantage of numbered callbacks"). That change caused problems with nvidia proprietary driver which calls acpi_os_map/unmap_generic_address several times during initialization. Each unmap calls synchronize_rcu and adds significant delay. Totally initialization is slowed for a couple of seconds and that is enough to trigger timeout in hardware, gpu decides to "fell off the bus". Widely spread workaround is reducing "rcu_idle_gp_delay" from 4 to 1 jiffy. This patch replaces synchronize_rcu() with synchronize_rcu_expedited() which is much faster. Link: https://devtalk.nvidia.com/default/topic/567297/linux/linux-3-10-driver-crash/ Signed-off-by: Konstantin Khlebnikov Reported-and-tested-by: Alexander Monakov Reviewed-by: Paul E. McKenney Signed-off-by: Rafael J. Wysocki --- drivers/acpi/osl.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/acpi/osl.c') diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c index 9964f70be98d..217713c11aaa 100644 --- a/drivers/acpi/osl.c +++ b/drivers/acpi/osl.c @@ -436,7 +436,7 @@ static void acpi_os_drop_map_ref(struct acpi_ioremap *map) static void acpi_os_map_cleanup(struct acpi_ioremap *map) { if (!map->refcount) { - synchronize_rcu(); + synchronize_rcu_expedited(); acpi_unmap(map->phys, map->virt); kfree(map); } -- cgit v1.2.2 From 90253a792ec23481402474704ef498ee81abe4e3 Mon Sep 17 00:00:00 2001 From: Lv Zheng Date: Wed, 5 Nov 2014 15:06:13 +0800 Subject: ACPI / OSL: Add IRQ handler flushing support in the OSL. It is possible that a GPE handler or a fixed event handler still accessed after removing the handlers by invoking acpi_remove_gpe_handler() or acpi_remove_fixed_event_handler(), this possibility can crash OPSM after a module removal. In the Linux kernel, though all other GPE drivers are not modules, since the IPMI_SI (ipmi_si_intf.c) can be compiled as a module, we still need to consider a solution for this issue when the driver switches to ACPI_GPE_RAW_HANDLER mode in order to invoke GPE APIs. ACPICA expects acpi_os_wait_events_complete() to be invoked after GPE disabling so that OSPM can ensure all running GPE handlers have exitted. But currently acpi_os_wait_events_complete() can only flush _Lxx/_Exx evaluation work queue and this philosophy cannot work for drivers that have installed a dedicated GPE handler. The only way to protect a callback is to perform some state holders (reference count, state machine) before invoking the callback. Then this issue can only be fixed by the following means: 1. Flush GPE in ACPICA before invoking the GPE handler. But currently, there is no such implementation in acpi_ev_gpe_dispatch(). 2. Flush GPE in ACPICA OSL before invoking the SCI handler. But currently, there is no such implementation in acpi_irq(). 3. Flush IRQ in OSPM IRQ layer before invoking the IRQ handler. In Linus kernel, this can be done by synchronize_irq(). 4. Flush scheduling in OSPM vector entry layer before invoking the vector. In Linux, this can be done by synchronize_sched(). Since ACPICA expects the GPE handlers to be flushed by the ACPICA OSL or the GPE drivers. If it is implemented by the GPE driver, we should see synchronize_irq()/synchronize_sched() invoked in such drivers. If it is implemented by the ACPICA OSL, ACPICA currently provides acpi_os_wait_events_complete() hook to achieve this. After the following commit: Commit: 69c841b6dd8313c9a673246cc0e2535174272cab Author: Lv Zheng Subject: ACPICA: Update use of acpi_os_wait_events_complete interface. The OSL acpi_os_wait_events_complete() is invoked after a GPE handler is removed from acpi_remove_gpe_handler() or a fixed event handler is removed from acpi_remove_fixed_event_handler(). Thus it is possible to implement GPE handler flushing using this ACPICA OSL now. So the solution 1 is currently not taken into account. By examining the IPMI_SI driver, we noticed that the IPMI_SI driver: 1. Uses free_irq() to flush non GPE based IRQ handlers, in free_irq(), synchronize_irq() is invoked, and 2. Uses acpi_remove_gpe_handler() to flush GPE based IRQ handlers, for such IRQ handlers, there is no synchronize_irq() invoked. Since there isn't synchronize_sched() implemented for this driver, from the driver's perspective, acpi_remove_gpe_handler() should have properly flushed the GPE handlers for it. Since the driver doesn't invoke synchronize_irq(), the solution 3 is not what the drivers expect. This patch implements solution 2. But since given the fact that the GPE is managed inside of ACPICA, and implementing the GPE flushing requires to implement the whole GPE management code again in the OSL, instead of flushing GPE, this patch flushes IRQ in acpi_os_wait_events_complete(). The flushing could last longer than expected as though the target GPE/fixed event that is removed can be fastly flushed, other GPEs/fix events can still be issued during the flushing period. This patch fixes this issue by invoking synchronize_hardirq() in acpi_os_wait_events_complete(). The reason why we don't invoke synchronize_irq() is: currently ACPICA is not threaded IRQ capable and the only difference between synchronize_irq() and synchronize_hardirq() is synchronize_irq() also flushes threaded IRQ handlers. Thus using synchronize_hardirq() can help to reduce the overall synchronization time for the current ACPICA implementation. Signed-off-by: Lv Zheng Cc: Rafael J. Wysocki Cc: Len Brown Cc: Robert Moore Cc: Corey Minyard Cc: linux-acpi@vger.kernel.org Cc: devel@acpica.org Cc: openipmi-developer@lists.sourceforge.net Signed-off-by: Rafael J. Wysocki --- drivers/acpi/osl.c | 6 ++++++ 1 file changed, 6 insertions(+) (limited to 'drivers/acpi/osl.c') diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c index 217713c11aaa..f9eeae871593 100644 --- a/drivers/acpi/osl.c +++ b/drivers/acpi/osl.c @@ -1188,6 +1188,12 @@ EXPORT_SYMBOL(acpi_os_execute); void acpi_os_wait_events_complete(void) { + /* + * Make sure the GPE handler or the fixed event handler is not used + * on another CPU after removal. + */ + if (acpi_irq_handler) + synchronize_hardirq(acpi_gbl_FADT.sci_interrupt); flush_workqueue(kacpid_wq); flush_workqueue(kacpi_notify_wq); } -- cgit v1.2.2