183 lines · c
1// SPDX-License-Identifier: GPL-2.0 OR MIT2/*3 * Copyright 2014-2022 Advanced Micro Devices, Inc.4 *5 * Permission is hereby granted, free of charge, to any person obtaining a6 * copy of this software and associated documentation files (the "Software"),7 * to deal in the Software without restriction, including without limitation8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,9 * and/or sell copies of the Software, and to permit persons to whom the10 * Software is furnished to do so, subject to the following conditions:11 *12 * The above copyright notice and this permission notice shall be included in13 * all copies or substantial portions of the Software.14 *15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR21 * OTHER DEALINGS IN THE SOFTWARE.22 */23 24/*25 * KFD Interrupts.26 *27 * AMD GPUs deliver interrupts by pushing an interrupt description onto the28 * interrupt ring and then sending an interrupt. KGD receives the interrupt29 * in ISR and sends us a pointer to each new entry on the interrupt ring.30 *31 * We generally can't process interrupt-signaled events from ISR, so we call32 * out to each interrupt client module (currently only the scheduler) to ask if33 * each interrupt is interesting. If they return true, then it requires further34 * processing so we copy it to an internal interrupt ring and call each35 * interrupt client again from a work-queue.36 *37 * There's no acknowledgment for the interrupts we use. The hardware simply38 * queues a new interrupt each time without waiting.39 *40 * The fixed-size internal queue means that it's possible for us to lose41 * interrupts because we have no back-pressure to the hardware.42 */43 44#include <linux/slab.h>45#include <linux/device.h>46#include <linux/kfifo.h>47#include "kfd_priv.h"48 49#define KFD_IH_NUM_ENTRIES 819250 51static void interrupt_wq(struct work_struct *);52 53int kfd_interrupt_init(struct kfd_node *node)54{55 int r;56 57 r = kfifo_alloc(&node->ih_fifo,58 KFD_IH_NUM_ENTRIES * node->kfd->device_info.ih_ring_entry_size,59 GFP_KERNEL);60 if (r) {61 dev_err(node->adev->dev, "Failed to allocate IH fifo\n");62 return r;63 }64 65 node->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1);66 if (unlikely(!node->ih_wq)) {67 kfifo_free(&node->ih_fifo);68 dev_err(node->adev->dev, "Failed to allocate KFD IH workqueue\n");69 return -ENOMEM;70 }71 spin_lock_init(&node->interrupt_lock);72 73 INIT_WORK(&node->interrupt_work, interrupt_wq);74 75 node->interrupts_active = true;76 77 /*78 * After this function returns, the interrupt will be enabled. This79 * barrier ensures that the interrupt running on a different processor80 * sees all the above writes.81 */82 smp_wmb();83 84 return 0;85}86 87void kfd_interrupt_exit(struct kfd_node *node)88{89 /*90 * Stop the interrupt handler from writing to the ring and scheduling91 * workqueue items. The spinlock ensures that any interrupt running92 * after we have unlocked sees interrupts_active = false.93 */94 unsigned long flags;95 96 spin_lock_irqsave(&node->interrupt_lock, flags);97 node->interrupts_active = false;98 spin_unlock_irqrestore(&node->interrupt_lock, flags);99 100 /*101 * flush_work ensures that there are no outstanding102 * work-queue items that will access interrupt_ring. New work items103 * can't be created because we stopped interrupt handling above.104 */105 flush_workqueue(node->ih_wq);106 107 destroy_workqueue(node->ih_wq);108 109 kfifo_free(&node->ih_fifo);110}111 112/*113 * Assumption: single reader/writer. This function is not re-entrant114 */115bool enqueue_ih_ring_entry(struct kfd_node *node, const void *ih_ring_entry)116{117 int count;118 119 count = kfifo_in(&node->ih_fifo, ih_ring_entry,120 node->kfd->device_info.ih_ring_entry_size);121 if (count != node->kfd->device_info.ih_ring_entry_size) {122 dev_dbg_ratelimited(node->adev->dev,123 "Interrupt ring overflow, dropping interrupt %d\n",124 count);125 return false;126 }127 128 return true;129}130 131/*132 * Assumption: single reader/writer. This function is not re-entrant133 */134static bool dequeue_ih_ring_entry(struct kfd_node *node, void *ih_ring_entry)135{136 int count;137 138 count = kfifo_out(&node->ih_fifo, ih_ring_entry,139 node->kfd->device_info.ih_ring_entry_size);140 141 WARN_ON(count && count != node->kfd->device_info.ih_ring_entry_size);142 143 return count == node->kfd->device_info.ih_ring_entry_size;144}145 146static void interrupt_wq(struct work_struct *work)147{148 struct kfd_node *dev = container_of(work, struct kfd_node,149 interrupt_work);150 uint32_t ih_ring_entry[KFD_MAX_RING_ENTRY_SIZE];151 unsigned long start_jiffies = jiffies;152 153 if (dev->kfd->device_info.ih_ring_entry_size > sizeof(ih_ring_entry)) {154 dev_err_once(dev->adev->dev, "Ring entry too small\n");155 return;156 }157 158 while (dequeue_ih_ring_entry(dev, ih_ring_entry)) {159 dev->kfd->device_info.event_interrupt_class->interrupt_wq(dev,160 ih_ring_entry);161 if (time_is_before_jiffies(start_jiffies + HZ)) {162 /* If we spent more than a second processing signals,163 * reschedule the worker to avoid soft-lockup warnings164 */165 queue_work(dev->ih_wq, &dev->interrupt_work);166 break;167 }168 }169}170 171bool interrupt_is_wanted(struct kfd_node *dev,172 const uint32_t *ih_ring_entry,173 uint32_t *patched_ihre, bool *flag)174{175 /* integer and bitwise OR so there is no boolean short-circuiting */176 unsigned int wanted = 0;177 178 wanted |= dev->kfd->device_info.event_interrupt_class->interrupt_isr(dev,179 ih_ring_entry, patched_ihre, flag);180 181 return wanted != 0;182}183