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 if (WARN_ON(type != KFD_QUEUE_TYPE_DIQ && type != KFD_QUEUE_TYPE_HIQ)) 45 return false; 46 47 pr_debug("Initializing queue type %d size %d\n", KFD_QUEUE_TYPE_HIQ, 48 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 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_DIQ]; 62 break; 63 case KFD_QUEUE_TYPE_HIQ: 64 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]; 65 break; 66 default: 67 pr_err("Invalid queue type %d\n", type); 68 return false; 69 } 70 71 if (!kq->mqd_mgr) 72 return false; 73 74 prop.doorbell_ptr = kfd_get_kernel_doorbell(dev, &prop.doorbell_off); 75 76 if (!prop.doorbell_ptr) { 77 pr_err("Failed to initialize doorbell"); 78 goto err_get_kernel_doorbell; 79 } 80 81 retval = kfd_gtt_sa_allocate(dev, queue_size, &kq->pq); 82 if (retval != 0) { 83 pr_err("Failed to init pq queues size %d\n", queue_size); 84 goto err_pq_allocate_vidmem; 85 } 86 87 kq->pq_kernel_addr = kq->pq->cpu_ptr; 88 kq->pq_gpu_addr = kq->pq->gpu_addr; 89 90 retval = kq->ops_asic_specific.initialize(kq, dev, type, queue_size); 91 if (!retval) 92 goto err_eop_allocate_vidmem; 93 94 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel), 95 &kq->rptr_mem); 96 97 if (retval != 0) 98 goto err_rptr_allocate_vidmem; 99 100 kq->rptr_kernel = kq->rptr_mem->cpu_ptr; 101 kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr; 102 103 retval = kfd_gtt_sa_allocate(dev, dev->device_info->doorbell_size, 104 &kq->wptr_mem); 105 106 if (retval != 0) 107 goto err_wptr_allocate_vidmem; 108 109 kq->wptr_kernel = kq->wptr_mem->cpu_ptr; 110 kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr; 111 112 memset(kq->pq_kernel_addr, 0, queue_size); 113 memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel)); 114 memset(kq->wptr_kernel, 0, sizeof(*kq->wptr_kernel)); 115 116 prop.queue_size = queue_size; 117 prop.is_interop = false; 118 prop.priority = 1; 119 prop.queue_percent = 100; 120 prop.type = type; 121 prop.vmid = 0; 122 prop.queue_address = kq->pq_gpu_addr; 123 prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr; 124 prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr; 125 prop.eop_ring_buffer_address = kq->eop_gpu_addr; 126 prop.eop_ring_buffer_size = PAGE_SIZE; 127 prop.cu_mask = NULL; 128 129 if (init_queue(&kq->queue, &prop) != 0) 130 goto err_init_queue; 131 132 kq->queue->device = dev; 133 kq->queue->process = kfd_get_process(current); 134 135 retval = kq->mqd_mgr->init_mqd(kq->mqd_mgr, &kq->queue->mqd, 136 &kq->queue->mqd_mem_obj, 137 &kq->queue->gart_mqd_addr, 138 &kq->queue->properties); 139 if (retval != 0) 140 goto err_init_mqd; 141 142 /* assign HIQ to HQD */ 143 if (type == KFD_QUEUE_TYPE_HIQ) { 144 pr_debug("Assigning hiq to hqd\n"); 145 kq->queue->pipe = KFD_CIK_HIQ_PIPE; 146 kq->queue->queue = KFD_CIK_HIQ_QUEUE; 147 kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd, 148 kq->queue->pipe, kq->queue->queue, 149 &kq->queue->properties, NULL); 150 } else { 151 /* allocate fence for DIQ */ 152 153 retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t), 154 &kq->fence_mem_obj); 155 156 if (retval != 0) 157 goto err_alloc_fence; 158 159 kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr; 160 kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr; 161 } 162 163 print_queue(kq->queue); 164 165 return true; 166 err_alloc_fence: 167 err_init_mqd: 168 uninit_queue(kq->queue); 169 err_init_queue: 170 kfd_gtt_sa_free(dev, kq->wptr_mem); 171 err_wptr_allocate_vidmem: 172 kfd_gtt_sa_free(dev, kq->rptr_mem); 173 err_rptr_allocate_vidmem: 174 kfd_gtt_sa_free(dev, kq->eop_mem); 175 err_eop_allocate_vidmem: 176 kfd_gtt_sa_free(dev, kq->pq); 177 err_pq_allocate_vidmem: 178 kfd_release_kernel_doorbell(dev, prop.doorbell_ptr); 179 err_get_kernel_doorbell: 180 return false; 181 182 } 183 184 static void uninitialize(struct kernel_queue *kq) 185 { 186 if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ) 187 kq->mqd_mgr->destroy_mqd(kq->mqd_mgr, 188 kq->queue->mqd, 189 KFD_PREEMPT_TYPE_WAVEFRONT_RESET, 190 KFD_UNMAP_LATENCY_MS, 191 kq->queue->pipe, 192 kq->queue->queue); 193 else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ) 194 kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj); 195 196 kq->mqd_mgr->uninit_mqd(kq->mqd_mgr, kq->queue->mqd, 197 kq->queue->mqd_mem_obj); 198 199 kfd_gtt_sa_free(kq->dev, kq->rptr_mem); 200 kfd_gtt_sa_free(kq->dev, kq->wptr_mem); 201 kq->ops_asic_specific.uninitialize(kq); 202 kfd_gtt_sa_free(kq->dev, kq->pq); 203 kfd_release_kernel_doorbell(kq->dev, 204 kq->queue->properties.doorbell_ptr); 205 uninit_queue(kq->queue); 206 } 207 208 static int acquire_packet_buffer(struct kernel_queue *kq, 209 size_t packet_size_in_dwords, unsigned int **buffer_ptr) 210 { 211 size_t available_size; 212 size_t queue_size_dwords; 213 uint32_t wptr, rptr; 214 uint64_t wptr64; 215 unsigned int *queue_address; 216 217 /* When rptr == wptr, the buffer is empty. 218 * When rptr == wptr + 1, the buffer is full. 219 * It is always rptr that advances to the position of wptr, rather than 220 * the opposite. So we can only use up to queue_size_dwords - 1 dwords. 221 */ 222 rptr = *kq->rptr_kernel; 223 wptr = kq->pending_wptr; 224 wptr64 = kq->pending_wptr64; 225 queue_address = (unsigned int *)kq->pq_kernel_addr; 226 queue_size_dwords = kq->queue->properties.queue_size / 4; 227 228 pr_debug("rptr: %d\n", rptr); 229 pr_debug("wptr: %d\n", wptr); 230 pr_debug("queue_address 0x%p\n", queue_address); 231 232 available_size = (rptr + queue_size_dwords - 1 - wptr) % 233 queue_size_dwords; 234 235 if (packet_size_in_dwords > available_size) { 236 /* 237 * make sure calling functions know 238 * acquire_packet_buffer() failed 239 */ 240 goto err_no_space; 241 } 242 243 if (wptr + packet_size_in_dwords >= queue_size_dwords) { 244 /* make sure after rolling back to position 0, there is 245 * still enough space. 246 */ 247 if (packet_size_in_dwords >= rptr) 248 goto err_no_space; 249 250 /* fill nops, roll back and start at position 0 */ 251 while (wptr > 0) { 252 queue_address[wptr] = kq->nop_packet; 253 wptr = (wptr + 1) % queue_size_dwords; 254 wptr64++; 255 } 256 } 257 258 *buffer_ptr = &queue_address[wptr]; 259 kq->pending_wptr = wptr + packet_size_in_dwords; 260 kq->pending_wptr64 = wptr64 + packet_size_in_dwords; 261 262 return 0; 263 264 err_no_space: 265 *buffer_ptr = NULL; 266 return -ENOMEM; 267 } 268 269 static void submit_packet(struct kernel_queue *kq) 270 { 271 #ifdef DEBUG 272 int i; 273 274 for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) { 275 pr_debug("0x%2X ", kq->pq_kernel_addr[i]); 276 if (i % 15 == 0) 277 pr_debug("\n"); 278 } 279 pr_debug("\n"); 280 #endif 281 282 kq->ops_asic_specific.submit_packet(kq); 283 } 284 285 static void rollback_packet(struct kernel_queue *kq) 286 { 287 if (kq->dev->device_info->doorbell_size == 8) { 288 kq->pending_wptr64 = *kq->wptr64_kernel; 289 kq->pending_wptr = *kq->wptr_kernel % 290 (kq->queue->properties.queue_size / 4); 291 } else { 292 kq->pending_wptr = *kq->wptr_kernel; 293 } 294 } 295 296 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 297 enum kfd_queue_type type) 298 { 299 struct kernel_queue *kq; 300 301 kq = kzalloc(sizeof(*kq), GFP_KERNEL); 302 if (!kq) 303 return NULL; 304 305 kq->ops.initialize = initialize; 306 kq->ops.uninitialize = uninitialize; 307 kq->ops.acquire_packet_buffer = acquire_packet_buffer; 308 kq->ops.submit_packet = submit_packet; 309 kq->ops.rollback_packet = rollback_packet; 310 311 switch (dev->device_info->asic_family) { 312 case CHIP_CARRIZO: 313 case CHIP_TONGA: 314 case CHIP_FIJI: 315 case CHIP_POLARIS10: 316 case CHIP_POLARIS11: 317 case CHIP_POLARIS12: 318 case CHIP_VEGAM: 319 kernel_queue_init_vi(&kq->ops_asic_specific); 320 break; 321 322 case CHIP_KAVERI: 323 case CHIP_HAWAII: 324 kernel_queue_init_cik(&kq->ops_asic_specific); 325 break; 326 327 case CHIP_VEGA10: 328 case CHIP_VEGA12: 329 case CHIP_VEGA20: 330 case CHIP_RAVEN: 331 kernel_queue_init_v9(&kq->ops_asic_specific); 332 break; 333 default: 334 WARN(1, "Unexpected ASIC family %u", 335 dev->device_info->asic_family); 336 goto out_free; 337 } 338 339 if (kq->ops.initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE)) 340 return kq; 341 342 pr_err("Failed to init kernel queue\n"); 343 344 out_free: 345 kfree(kq); 346 return NULL; 347 } 348 349 void kernel_queue_uninit(struct kernel_queue *kq) 350 { 351 kq->ops.uninitialize(kq); 352 kfree(kq); 353 } 354 355 /* FIXME: Can this test be removed? */ 356 static __attribute__((unused)) void test_kq(struct kfd_dev *dev) 357 { 358 struct kernel_queue *kq; 359 uint32_t *buffer, i; 360 int retval; 361 362 pr_err("Starting kernel queue test\n"); 363 364 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ); 365 if (unlikely(!kq)) { 366 pr_err(" Failed to initialize HIQ\n"); 367 pr_err("Kernel queue test failed\n"); 368 return; 369 } 370 371 retval = kq->ops.acquire_packet_buffer(kq, 5, &buffer); 372 if (unlikely(retval != 0)) { 373 pr_err(" Failed to acquire packet buffer\n"); 374 pr_err("Kernel queue test failed\n"); 375 return; 376 } 377 for (i = 0; i < 5; i++) 378 buffer[i] = kq->nop_packet; 379 kq->ops.submit_packet(kq); 380 381 pr_err("Ending kernel queue test\n"); 382 } 383 384 385