1 /* 2 * Copyright 2014 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 #include <linux/types.h> 25 #include <linux/mutex.h> 26 #include <linux/slab.h> 27 #include <linux/printk.h> 28 #include <linux/sched.h> 29 #include "kfd_kernel_queue.h" 30 #include "kfd_priv.h" 31 #include "kfd_device_queue_manager.h" 32 #include "kfd_pm4_headers.h" 33 #include "kfd_pm4_opcodes.h" 34 35 #define PM4_COUNT_ZERO (((1 << 15) - 1) << 16) 36 37 static bool initialize(struct kernel_queue *kq, struct kfd_dev *dev, 38 enum kfd_queue_type type, unsigned int queue_size) 39 { 40 struct queue_properties prop; 41 int retval; 42 union PM4_MES_TYPE_3_HEADER nop; 43 44 BUG_ON(!kq || !dev); 45 BUG_ON(type != KFD_QUEUE_TYPE_DIQ && type != KFD_QUEUE_TYPE_HIQ); 46 47 pr_debug("amdkfd: In func %s initializing queue type %d size %d\n", 48 __func__, KFD_QUEUE_TYPE_HIQ, queue_size); 49 50 memset(&prop, 0, sizeof(prop)); 51 memset(&nop, 0, sizeof(nop)); 52 53 nop.opcode = IT_NOP; 54 nop.type = PM4_TYPE_3; 55 nop.u32all |= PM4_COUNT_ZERO; 56 57 kq->dev = dev; 58 kq->nop_packet = nop.u32all; 59 switch (type) { 60 case KFD_QUEUE_TYPE_DIQ: 61 case KFD_QUEUE_TYPE_HIQ: 62 kq->mqd = dev->dqm->ops.get_mqd_manager(dev->dqm, 63 KFD_MQD_TYPE_HIQ); 64 break; 65 default: 66 BUG(); 67 break; 68 } 69 70 if (kq->mqd == NULL) 71 return false; 72 73 prop.doorbell_ptr = kfd_get_kernel_doorbell(dev, &prop.doorbell_off); 74 75 if (prop.doorbell_ptr == NULL) { 76 pr_err("amdkfd: error init doorbell"); 77 goto err_get_kernel_doorbell; 78 } 79 80 retval = kfd_gtt_sa_allocate(dev, queue_size, &kq->pq); 81 if (retval != 0) { 82 pr_err("amdkfd: error init pq queues size (%d)\n", queue_size); 83 goto err_pq_allocate_vidmem; 84 } 85 86 kq->pq_kernel_addr = kq->pq->cpu_ptr; 87 kq->pq_gpu_addr = kq->pq->gpu_addr; 88 89 retval = kq->ops_asic_specific.initialize(kq, dev, type, queue_size); 90 if (retval == false) 91 goto err_eop_allocate_vidmem; 92 93 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel), 94 &kq->rptr_mem); 95 96 if (retval != 0) 97 goto err_rptr_allocate_vidmem; 98 99 kq->rptr_kernel = kq->rptr_mem->cpu_ptr; 100 kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr; 101 102 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->wptr_kernel), 103 &kq->wptr_mem); 104 105 if (retval != 0) 106 goto err_wptr_allocate_vidmem; 107 108 kq->wptr_kernel = kq->wptr_mem->cpu_ptr; 109 kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr; 110 111 memset(kq->pq_kernel_addr, 0, queue_size); 112 memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel)); 113 memset(kq->wptr_kernel, 0, sizeof(*kq->wptr_kernel)); 114 115 prop.queue_size = queue_size; 116 prop.is_interop = false; 117 prop.priority = 1; 118 prop.queue_percent = 100; 119 prop.type = type; 120 prop.vmid = 0; 121 prop.queue_address = kq->pq_gpu_addr; 122 prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr; 123 prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr; 124 prop.eop_ring_buffer_address = kq->eop_gpu_addr; 125 prop.eop_ring_buffer_size = PAGE_SIZE; 126 127 if (init_queue(&kq->queue, &prop) != 0) 128 goto err_init_queue; 129 130 kq->queue->device = dev; 131 kq->queue->process = kfd_get_process(current); 132 133 retval = kq->mqd->init_mqd(kq->mqd, &kq->queue->mqd, 134 &kq->queue->mqd_mem_obj, 135 &kq->queue->gart_mqd_addr, 136 &kq->queue->properties); 137 if (retval != 0) 138 goto err_init_mqd; 139 140 /* assign HIQ to HQD */ 141 if (type == KFD_QUEUE_TYPE_HIQ) { 142 pr_debug("assigning hiq to hqd\n"); 143 kq->queue->pipe = KFD_CIK_HIQ_PIPE; 144 kq->queue->queue = KFD_CIK_HIQ_QUEUE; 145 kq->mqd->load_mqd(kq->mqd, kq->queue->mqd, kq->queue->pipe, 146 kq->queue->queue, NULL); 147 } else { 148 /* allocate fence for DIQ */ 149 150 retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t), 151 &kq->fence_mem_obj); 152 153 if (retval != 0) 154 goto err_alloc_fence; 155 156 kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr; 157 kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr; 158 } 159 160 print_queue(kq->queue); 161 162 return true; 163 err_alloc_fence: 164 err_init_mqd: 165 uninit_queue(kq->queue); 166 err_init_queue: 167 kfd_gtt_sa_free(dev, kq->wptr_mem); 168 err_wptr_allocate_vidmem: 169 kfd_gtt_sa_free(dev, kq->rptr_mem); 170 err_rptr_allocate_vidmem: 171 kfd_gtt_sa_free(dev, kq->eop_mem); 172 err_eop_allocate_vidmem: 173 kfd_gtt_sa_free(dev, kq->pq); 174 err_pq_allocate_vidmem: 175 kfd_release_kernel_doorbell(dev, prop.doorbell_ptr); 176 err_get_kernel_doorbell: 177 return false; 178 179 } 180 181 static void uninitialize(struct kernel_queue *kq) 182 { 183 BUG_ON(!kq); 184 185 if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ) 186 kq->mqd->destroy_mqd(kq->mqd, 187 NULL, 188 false, 189 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS, 190 kq->queue->pipe, 191 kq->queue->queue); 192 else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ) 193 kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj); 194 195 kq->mqd->uninit_mqd(kq->mqd, kq->queue->mqd, kq->queue->mqd_mem_obj); 196 197 kfd_gtt_sa_free(kq->dev, kq->rptr_mem); 198 kfd_gtt_sa_free(kq->dev, kq->wptr_mem); 199 kq->ops_asic_specific.uninitialize(kq); 200 kfd_gtt_sa_free(kq->dev, kq->pq); 201 kfd_release_kernel_doorbell(kq->dev, 202 kq->queue->properties.doorbell_ptr); 203 uninit_queue(kq->queue); 204 } 205 206 static int acquire_packet_buffer(struct kernel_queue *kq, 207 size_t packet_size_in_dwords, unsigned int **buffer_ptr) 208 { 209 size_t available_size; 210 size_t queue_size_dwords; 211 uint32_t wptr, rptr; 212 unsigned int *queue_address; 213 214 BUG_ON(!kq || !buffer_ptr); 215 216 rptr = *kq->rptr_kernel; 217 wptr = *kq->wptr_kernel; 218 queue_address = (unsigned int *)kq->pq_kernel_addr; 219 queue_size_dwords = kq->queue->properties.queue_size / sizeof(uint32_t); 220 221 pr_debug("rptr: %d\n", rptr); 222 pr_debug("wptr: %d\n", wptr); 223 pr_debug("queue_address 0x%p\n", queue_address); 224 225 available_size = (rptr - 1 - wptr + queue_size_dwords) % 226 queue_size_dwords; 227 228 if (packet_size_in_dwords >= queue_size_dwords || 229 packet_size_in_dwords >= available_size) { 230 /* 231 * make sure calling functions know 232 * acquire_packet_buffer() failed 233 */ 234 *buffer_ptr = NULL; 235 return -ENOMEM; 236 } 237 238 if (wptr + packet_size_in_dwords >= queue_size_dwords) { 239 while (wptr > 0) { 240 queue_address[wptr] = kq->nop_packet; 241 wptr = (wptr + 1) % queue_size_dwords; 242 } 243 } 244 245 *buffer_ptr = &queue_address[wptr]; 246 kq->pending_wptr = wptr + packet_size_in_dwords; 247 248 return 0; 249 } 250 251 static void submit_packet(struct kernel_queue *kq) 252 { 253 #ifdef DEBUG 254 int i; 255 #endif 256 257 BUG_ON(!kq); 258 259 #ifdef DEBUG 260 for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) { 261 pr_debug("0x%2X ", kq->pq_kernel_addr[i]); 262 if (i % 15 == 0) 263 pr_debug("\n"); 264 } 265 pr_debug("\n"); 266 #endif 267 268 *kq->wptr_kernel = kq->pending_wptr; 269 write_kernel_doorbell(kq->queue->properties.doorbell_ptr, 270 kq->pending_wptr); 271 } 272 273 static void rollback_packet(struct kernel_queue *kq) 274 { 275 BUG_ON(!kq); 276 kq->pending_wptr = *kq->queue->properties.write_ptr; 277 } 278 279 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 280 enum kfd_queue_type type) 281 { 282 struct kernel_queue *kq; 283 284 BUG_ON(!dev); 285 286 kq = kzalloc(sizeof(struct kernel_queue), GFP_KERNEL); 287 if (!kq) 288 return NULL; 289 290 kq->ops.initialize = initialize; 291 kq->ops.uninitialize = uninitialize; 292 kq->ops.acquire_packet_buffer = acquire_packet_buffer; 293 kq->ops.submit_packet = submit_packet; 294 kq->ops.rollback_packet = rollback_packet; 295 296 switch (dev->device_info->asic_family) { 297 case CHIP_CARRIZO: 298 kernel_queue_init_vi(&kq->ops_asic_specific); 299 break; 300 301 case CHIP_KAVERI: 302 kernel_queue_init_cik(&kq->ops_asic_specific); 303 break; 304 } 305 306 if (!kq->ops.initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE)) { 307 pr_err("amdkfd: failed to init kernel queue\n"); 308 kfree(kq); 309 return NULL; 310 } 311 return kq; 312 } 313 314 void kernel_queue_uninit(struct kernel_queue *kq) 315 { 316 BUG_ON(!kq); 317 318 kq->ops.uninitialize(kq); 319 kfree(kq); 320 } 321 322 static __attribute__((unused)) void test_kq(struct kfd_dev *dev) 323 { 324 struct kernel_queue *kq; 325 uint32_t *buffer, i; 326 int retval; 327 328 BUG_ON(!dev); 329 330 pr_err("amdkfd: starting kernel queue test\n"); 331 332 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ); 333 BUG_ON(!kq); 334 335 retval = kq->ops.acquire_packet_buffer(kq, 5, &buffer); 336 BUG_ON(retval != 0); 337 for (i = 0; i < 5; i++) 338 buffer[i] = kq->nop_packet; 339 kq->ops.submit_packet(kq); 340 341 pr_err("amdkfd: ending kernel queue test\n"); 342 } 343 344 345