1 /* $NetBSD: kfd_interrupt.c,v 1.2 2018/08/27 04:58:20 riastradh Exp $ */ 2 3 /* 4 * Copyright 2014 Advanced Micro Devices, Inc. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 */ 24 25 /* 26 * KFD Interrupts. 27 * 28 * AMD GPUs deliver interrupts by pushing an interrupt description onto the 29 * interrupt ring and then sending an interrupt. KGD receives the interrupt 30 * in ISR and sends us a pointer to each new entry on the interrupt ring. 31 * 32 * We generally can't process interrupt-signaled events from ISR, so we call 33 * out to each interrupt client module (currently only the scheduler) to ask if 34 * each interrupt is interesting. If they return true, then it requires further 35 * processing so we copy it to an internal interrupt ring and call each 36 * interrupt client again from a work-queue. 37 * 38 * There's no acknowledgment for the interrupts we use. The hardware simply 39 * queues a new interrupt each time without waiting. 40 * 41 * The fixed-size internal queue means that it's possible for us to lose 42 * interrupts because we have no back-pressure to the hardware. 43 */ 44 45 #include <sys/cdefs.h> 46 __KERNEL_RCSID(0, "$NetBSD: kfd_interrupt.c,v 1.2 2018/08/27 04:58:20 riastradh Exp $"); 47 48 #include <linux/slab.h> 49 #include <linux/device.h> 50 #include "kfd_priv.h" 51 52 #define KFD_INTERRUPT_RING_SIZE 1024 53 54 static void interrupt_wq(struct work_struct *); 55 56 int kfd_interrupt_init(struct kfd_dev *kfd) 57 { 58 void *interrupt_ring = kmalloc_array(KFD_INTERRUPT_RING_SIZE, 59 kfd->device_info->ih_ring_entry_size, 60 GFP_KERNEL); 61 if (!interrupt_ring) 62 return -ENOMEM; 63 64 kfd->interrupt_ring = interrupt_ring; 65 kfd->interrupt_ring_size = 66 KFD_INTERRUPT_RING_SIZE * kfd->device_info->ih_ring_entry_size; 67 atomic_set(&kfd->interrupt_ring_wptr, 0); 68 atomic_set(&kfd->interrupt_ring_rptr, 0); 69 70 spin_lock_init(&kfd->interrupt_lock); 71 72 INIT_WORK(&kfd->interrupt_work, interrupt_wq); 73 74 kfd->interrupts_active = true; 75 76 /* 77 * After this function returns, the interrupt will be enabled. This 78 * barrier ensures that the interrupt running on a different processor 79 * sees all the above writes. 80 */ 81 smp_wmb(); 82 83 return 0; 84 } 85 86 void kfd_interrupt_exit(struct kfd_dev *kfd) 87 { 88 /* 89 * Stop the interrupt handler from writing to the ring and scheduling 90 * workqueue items. The spinlock ensures that any interrupt running 91 * after we have unlocked sees interrupts_active = false. 92 */ 93 unsigned long flags; 94 95 spin_lock_irqsave(&kfd->interrupt_lock, flags); 96 kfd->interrupts_active = false; 97 spin_unlock_irqrestore(&kfd->interrupt_lock, flags); 98 99 /* 100 * Flush_scheduled_work ensures that there are no outstanding 101 * work-queue items that will access interrupt_ring. New work items 102 * can't be created because we stopped interrupt handling above. 103 */ 104 flush_scheduled_work(); 105 106 kfree(kfd->interrupt_ring); 107 } 108 109 /* 110 * This assumes that it can't be called concurrently with itself 111 * but only with dequeue_ih_ring_entry. 112 */ 113 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry) 114 { 115 unsigned int rptr = atomic_read(&kfd->interrupt_ring_rptr); 116 unsigned int wptr = atomic_read(&kfd->interrupt_ring_wptr); 117 118 if ((rptr - wptr) % kfd->interrupt_ring_size == 119 kfd->device_info->ih_ring_entry_size) { 120 /* This is very bad, the system is likely to hang. */ 121 dev_err_ratelimited(kfd_chardev(), 122 "Interrupt ring overflow, dropping interrupt.\n"); 123 return false; 124 } 125 126 memcpy(kfd->interrupt_ring + wptr, ih_ring_entry, 127 kfd->device_info->ih_ring_entry_size); 128 129 wptr = (wptr + kfd->device_info->ih_ring_entry_size) % 130 kfd->interrupt_ring_size; 131 smp_wmb(); /* Ensure memcpy'd data is visible before wptr update. */ 132 atomic_set(&kfd->interrupt_ring_wptr, wptr); 133 134 return true; 135 } 136 137 /* 138 * This assumes that it can't be called concurrently with itself 139 * but only with enqueue_ih_ring_entry. 140 */ 141 static bool dequeue_ih_ring_entry(struct kfd_dev *kfd, void *ih_ring_entry) 142 { 143 /* 144 * Assume that wait queues have an implicit barrier, i.e. anything that 145 * happened in the ISR before it queued work is visible. 146 */ 147 148 unsigned int wptr = atomic_read(&kfd->interrupt_ring_wptr); 149 unsigned int rptr = atomic_read(&kfd->interrupt_ring_rptr); 150 151 if (rptr == wptr) 152 return false; 153 154 memcpy(ih_ring_entry, kfd->interrupt_ring + rptr, 155 kfd->device_info->ih_ring_entry_size); 156 157 rptr = (rptr + kfd->device_info->ih_ring_entry_size) % 158 kfd->interrupt_ring_size; 159 160 /* 161 * Ensure the rptr write update is not visible until 162 * memcpy has finished reading. 163 */ 164 smp_mb(); 165 atomic_set(&kfd->interrupt_ring_rptr, rptr); 166 167 return true; 168 } 169 170 static void interrupt_wq(struct work_struct *work) 171 { 172 struct kfd_dev *dev = container_of(work, struct kfd_dev, 173 interrupt_work); 174 175 uint32_t ih_ring_entry[DIV_ROUND_UP( 176 dev->device_info->ih_ring_entry_size, 177 sizeof(uint32_t))]; 178 179 while (dequeue_ih_ring_entry(dev, ih_ring_entry)) 180 dev->device_info->event_interrupt_class->interrupt_wq(dev, 181 ih_ring_entry); 182 } 183 184 bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry) 185 { 186 /* integer and bitwise OR so there is no boolean short-circuiting */ 187 unsigned wanted = 0; 188 189 wanted |= dev->device_info->event_interrupt_class->interrupt_isr(dev, 190 ih_ring_entry); 191 192 return wanted != 0; 193 } 194