drm/amdkfd: Drop interrupt SW ring buffer
authorMichel Dänzer <michel.daenzer@amd.com>
Thu, 8 Jan 2015 04:27:15 +0000 (13:27 +0900)
committerOded Gabbay <oded.gabbay@amd.com>
Thu, 8 Jan 2015 04:27:15 +0000 (13:27 +0900)
The work queue couldn't reliably prevent the SW ring buffer from
overflowing, so dmesg was spammed by

 kfd kfd: Interrupt ring overflow, dropping interrupt.

messages when running e.g. the Atlantis Substance demo from
https://wiki.unrealengine.com/Linux_Demos on Kaveri.

Since the SW ring buffer doesn't actually do anything at this point, just
remove it for now. When actual interrupt processing code is added to
amdkfd, it should try to do things immediately and only defer to work
queues when necessary.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@amd.com>
drivers/gpu/drm/amd/amdkfd/Makefile
drivers/gpu/drm/amd/amdkfd/kfd_device.c
drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c [deleted file]
drivers/gpu/drm/amd/amdkfd/kfd_priv.h

index be6246de5091d0a589c25de25d147d8ed786af39..307a309110e60d5f473be86d26c24e3b0cb80a90 100644 (file)
@@ -8,7 +8,6 @@ amdkfd-y        := kfd_module.o kfd_device.o kfd_chardev.o kfd_topology.o \
                kfd_pasid.o kfd_doorbell.o kfd_flat_memory.o \
                kfd_process.o kfd_queue.o kfd_mqd_manager.o \
                kfd_kernel_queue.o kfd_packet_manager.o \
-               kfd_process_queue_manager.o kfd_device_queue_manager.o \
-               kfd_interrupt.o
+               kfd_process_queue_manager.o kfd_device_queue_manager.o
 
 obj-$(CONFIG_HSA_AMD)  += amdkfd.o
index 43884ebd4303fedd47b56431737a0f726852c552..633532a2e7ec875b968343975c2d49d69d220abc 100644 (file)
@@ -192,13 +192,6 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
                goto kfd_topology_add_device_error;
        }
 
-       if (kfd_interrupt_init(kfd)) {
-               dev_err(kfd_device,
-                       "Error initializing interrupts for device (%x:%x)\n",
-                       kfd->pdev->vendor, kfd->pdev->device);
-               goto kfd_interrupt_error;
-       }
-
        if (!device_iommu_pasid_init(kfd)) {
                dev_err(kfd_device,
                        "Error initializing iommuv2 for device (%x:%x)\n",
@@ -237,8 +230,6 @@ dqm_start_error:
 device_queue_manager_error:
        amd_iommu_free_device(kfd->pdev);
 device_iommu_pasid_error:
-       kfd_interrupt_exit(kfd);
-kfd_interrupt_error:
        kfd_topology_remove_device(kfd);
 kfd_topology_add_device_error:
        kfd2kgd->fini_sa_manager(kfd->kgd);
@@ -254,7 +245,6 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd)
        if (kfd->init_complete) {
                device_queue_manager_uninit(kfd->dqm);
                amd_iommu_free_device(kfd->pdev);
-               kfd_interrupt_exit(kfd);
                kfd_topology_remove_device(kfd);
        }
 
@@ -296,13 +286,5 @@ int kgd2kfd_resume(struct kfd_dev *kfd)
 /* This is called directly from KGD at ISR. */
 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 {
-       if (kfd->init_complete) {
-               spin_lock(&kfd->interrupt_lock);
-
-               if (kfd->interrupts_active
-                   && enqueue_ih_ring_entry(kfd, ih_ring_entry))
-                       schedule_work(&kfd->interrupt_work);
-
-               spin_unlock(&kfd->interrupt_lock);
-       }
+       /* Process interrupts / schedule work as necessary */
 }
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c
deleted file mode 100644 (file)
index 5b99909..0000000
+++ /dev/null
@@ -1,176 +0,0 @@
-/*
- * Copyright 2014 Advanced Micro Devices, Inc.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in
- * all copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
- * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
- * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
- * OTHER DEALINGS IN THE SOFTWARE.
- */
-
-/*
- * KFD Interrupts.
- *
- * AMD GPUs deliver interrupts by pushing an interrupt description onto the
- * interrupt ring and then sending an interrupt. KGD receives the interrupt
- * in ISR and sends us a pointer to each new entry on the interrupt ring.
- *
- * We generally can't process interrupt-signaled events from ISR, so we call
- * out to each interrupt client module (currently only the scheduler) to ask if
- * each interrupt is interesting. If they return true, then it requires further
- * processing so we copy it to an internal interrupt ring and call each
- * interrupt client again from a work-queue.
- *
- * There's no acknowledgment for the interrupts we use. The hardware simply
- * queues a new interrupt each time without waiting.
- *
- * The fixed-size internal queue means that it's possible for us to lose
- * interrupts because we have no back-pressure to the hardware.
- */
-
-#include <linux/slab.h>
-#include <linux/device.h>
-#include "kfd_priv.h"
-
-#define KFD_INTERRUPT_RING_SIZE 256
-
-static void interrupt_wq(struct work_struct *);
-
-int kfd_interrupt_init(struct kfd_dev *kfd)
-{
-       void *interrupt_ring = kmalloc_array(KFD_INTERRUPT_RING_SIZE,
-                                       kfd->device_info->ih_ring_entry_size,
-                                       GFP_KERNEL);
-       if (!interrupt_ring)
-               return -ENOMEM;
-
-       kfd->interrupt_ring = interrupt_ring;
-       kfd->interrupt_ring_size =
-               KFD_INTERRUPT_RING_SIZE * kfd->device_info->ih_ring_entry_size;
-       atomic_set(&kfd->interrupt_ring_wptr, 0);
-       atomic_set(&kfd->interrupt_ring_rptr, 0);
-
-       spin_lock_init(&kfd->interrupt_lock);
-
-       INIT_WORK(&kfd->interrupt_work, interrupt_wq);
-
-       kfd->interrupts_active = true;
-
-       /*
-        * After this function returns, the interrupt will be enabled. This
-        * barrier ensures that the interrupt running on a different processor
-        * sees all the above writes.
-        */
-       smp_wmb();
-
-       return 0;
-}
-
-void kfd_interrupt_exit(struct kfd_dev *kfd)
-{
-       /*
-        * Stop the interrupt handler from writing to the ring and scheduling
-        * workqueue items. The spinlock ensures that any interrupt running
-        * after we have unlocked sees interrupts_active = false.
-        */
-       unsigned long flags;
-
-       spin_lock_irqsave(&kfd->interrupt_lock, flags);
-       kfd->interrupts_active = false;
-       spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
-
-       /*
-        * Flush_scheduled_work ensures that there are no outstanding
-        * work-queue items that will access interrupt_ring. New work items
-        * can't be created because we stopped interrupt handling above.
-        */
-       flush_scheduled_work();
-
-       kfree(kfd->interrupt_ring);
-}
-
-/*
- * This assumes that it can't be called concurrently with itself
- * but only with dequeue_ih_ring_entry.
- */
-bool enqueue_ih_ring_entry(struct kfd_dev *kfd,        const void *ih_ring_entry)
-{
-       unsigned int rptr = atomic_read(&kfd->interrupt_ring_rptr);
-       unsigned int wptr = atomic_read(&kfd->interrupt_ring_wptr);
-
-       if ((rptr - wptr) % kfd->interrupt_ring_size ==
-                                       kfd->device_info->ih_ring_entry_size) {
-               /* This is very bad, the system is likely to hang. */
-               dev_err_ratelimited(kfd_chardev(),
-                       "Interrupt ring overflow, dropping interrupt.\n");
-               return false;
-       }
-
-       memcpy(kfd->interrupt_ring + wptr, ih_ring_entry,
-                       kfd->device_info->ih_ring_entry_size);
-
-       wptr = (wptr + kfd->device_info->ih_ring_entry_size) %
-                       kfd->interrupt_ring_size;
-       smp_wmb(); /* Ensure memcpy'd data is visible before wptr update. */
-       atomic_set(&kfd->interrupt_ring_wptr, wptr);
-
-       return true;
-}
-
-/*
- * This assumes that it can't be called concurrently with itself
- * but only with enqueue_ih_ring_entry.
- */
-static bool dequeue_ih_ring_entry(struct kfd_dev *kfd, void *ih_ring_entry)
-{
-       /*
-        * Assume that wait queues have an implicit barrier, i.e. anything that
-        * happened in the ISR before it queued work is visible.
-        */
-
-       unsigned int wptr = atomic_read(&kfd->interrupt_ring_wptr);
-       unsigned int rptr = atomic_read(&kfd->interrupt_ring_rptr);
-
-       if (rptr == wptr)
-               return false;
-
-       memcpy(ih_ring_entry, kfd->interrupt_ring + rptr,
-                       kfd->device_info->ih_ring_entry_size);
-
-       rptr = (rptr + kfd->device_info->ih_ring_entry_size) %
-                       kfd->interrupt_ring_size;
-
-       /*
-        * Ensure the rptr write update is not visible until
-        * memcpy has finished reading.
-        */
-       smp_mb();
-       atomic_set(&kfd->interrupt_ring_rptr, rptr);
-
-       return true;
-}
-
-static void interrupt_wq(struct work_struct *work)
-{
-       struct kfd_dev *dev = container_of(work, struct kfd_dev,
-                                               interrupt_work);
-
-       uint32_t ih_ring_entry[DIV_ROUND_UP(
-                               dev->device_info->ih_ring_entry_size,
-                               sizeof(uint32_t))];
-
-       while (dequeue_ih_ring_entry(dev, ih_ring_entry))
-               ;
-}
index a5edb29507e310af8ebe437ef2a22fb218687ad1..b3dc13c83169c1a3fb5cd54c9ae179aaf9508d36 100644 (file)
@@ -135,22 +135,10 @@ struct kfd_dev {
 
        struct kgd2kfd_shared_resources shared_resources;
 
-       void *interrupt_ring;
-       size_t interrupt_ring_size;
-       atomic_t interrupt_ring_rptr;
-       atomic_t interrupt_ring_wptr;
-       struct work_struct interrupt_work;
-       spinlock_t interrupt_lock;
-
        /* QCM Device instance */
        struct device_queue_manager *dqm;
 
        bool init_complete;
-       /*
-        * Interrupts of interest to KFD are copied
-        * from the HW ring into a SW ring.
-        */
-       bool interrupts_active;
 };
 
 /* KGD2KFD callbacks */
@@ -531,10 +519,7 @@ struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
 struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx);
 
 /* Interrupts */
-int kfd_interrupt_init(struct kfd_dev *dev);
-void kfd_interrupt_exit(struct kfd_dev *dev);
 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
-bool enqueue_ih_ring_entry(struct kfd_dev *kfd,        const void *ih_ring_entry);
 
 /* Power Management */
 void kgd2kfd_suspend(struct kfd_dev *kfd);