1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2014-2022 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * 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 OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 * 23 */ 24 25 #include <linux/types.h> 26 #include <linux/mutex.h> 27 #include <linux/slab.h> 28 #include <linux/printk.h> 29 #include <linux/sched.h> 30 #include "kfd_kernel_queue.h" 31 #include "kfd_priv.h" 32 #include "kfd_device_queue_manager.h" 33 #include "kfd_pm4_headers.h" 34 #include "kfd_pm4_opcodes.h" 35 #include "amdgpu_reset.h" 36 37 #define PM4_COUNT_ZERO (((1 << 15) - 1) << 16) 38 39 /* Initialize a kernel queue, including allocations of GART memory 40 * needed for the queue. 41 */ 42 static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev, 43 enum kfd_queue_type type, unsigned int queue_size) 44 { 45 struct queue_properties prop; 46 int retval; 47 union PM4_MES_TYPE_3_HEADER nop; 48 49 if (WARN_ON(type != KFD_QUEUE_TYPE_DIQ && type != KFD_QUEUE_TYPE_HIQ)) 50 return false; 51 52 pr_debug("Initializing queue type %d size %d\n", KFD_QUEUE_TYPE_HIQ, 53 queue_size); 54 55 memset(&prop, 0, sizeof(prop)); 56 memset(&nop, 0, sizeof(nop)); 57 58 nop.opcode = IT_NOP; 59 nop.type = PM4_TYPE_3; 60 nop.u32all |= PM4_COUNT_ZERO; 61 62 kq->dev = dev; 63 kq->nop_packet = nop.u32all; 64 switch (type) { 65 case KFD_QUEUE_TYPE_DIQ: 66 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_DIQ]; 67 break; 68 case KFD_QUEUE_TYPE_HIQ: 69 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]; 70 break; 71 default: 72 pr_err("Invalid queue type %d\n", type); 73 return false; 74 } 75 76 if (!kq->mqd_mgr) 77 return false; 78 79 prop.doorbell_ptr = kfd_get_kernel_doorbell(dev->kfd, &prop.doorbell_off); 80 81 if (!prop.doorbell_ptr) { 82 pr_err("Failed to initialize doorbell"); 83 goto err_get_kernel_doorbell; 84 } 85 86 retval = kfd_gtt_sa_allocate(dev, queue_size, &kq->pq); 87 if (retval != 0) { 88 pr_err("Failed to init pq queues size %d\n", queue_size); 89 goto err_pq_allocate_vidmem; 90 } 91 92 kq->pq_kernel_addr = kq->pq->cpu_ptr; 93 kq->pq_gpu_addr = kq->pq->gpu_addr; 94 95 /* For CIK family asics, kq->eop_mem is not needed */ 96 if (dev->adev->asic_type > CHIP_MULLINS) { 97 retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem); 98 if (retval != 0) 99 goto err_eop_allocate_vidmem; 100 101 kq->eop_gpu_addr = kq->eop_mem->gpu_addr; 102 kq->eop_kernel_addr = kq->eop_mem->cpu_ptr; 103 104 memset(kq->eop_kernel_addr, 0, PAGE_SIZE); 105 } 106 107 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel), 108 &kq->rptr_mem); 109 110 if (retval != 0) 111 goto err_rptr_allocate_vidmem; 112 113 kq->rptr_kernel = kq->rptr_mem->cpu_ptr; 114 kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr; 115 116 retval = kfd_gtt_sa_allocate(dev, dev->kfd->device_info.doorbell_size, 117 &kq->wptr_mem); 118 119 if (retval != 0) 120 goto err_wptr_allocate_vidmem; 121 122 kq->wptr_kernel = kq->wptr_mem->cpu_ptr; 123 kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr; 124 125 memset(kq->pq_kernel_addr, 0, queue_size); 126 memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel)); 127 memset(kq->wptr_kernel, 0, sizeof(*kq->wptr_kernel)); 128 129 prop.queue_size = queue_size; 130 prop.is_interop = false; 131 prop.is_gws = false; 132 prop.priority = 1; 133 prop.queue_percent = 100; 134 prop.type = type; 135 prop.vmid = 0; 136 prop.queue_address = kq->pq_gpu_addr; 137 prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr; 138 prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr; 139 prop.eop_ring_buffer_address = kq->eop_gpu_addr; 140 prop.eop_ring_buffer_size = PAGE_SIZE; 141 142 if (init_queue(&kq->queue, &prop) != 0) 143 goto err_init_queue; 144 145 kq->queue->device = dev; 146 kq->queue->process = kfd_get_process(current); 147 148 kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev, 149 &kq->queue->properties); 150 if (!kq->queue->mqd_mem_obj) 151 goto err_allocate_mqd; 152 kq->mqd_mgr->init_mqd(kq->mqd_mgr, &kq->queue->mqd, 153 kq->queue->mqd_mem_obj, 154 &kq->queue->gart_mqd_addr, 155 &kq->queue->properties); 156 /* assign HIQ to HQD */ 157 if (type == KFD_QUEUE_TYPE_HIQ) { 158 pr_debug("Assigning hiq to hqd\n"); 159 kq->queue->pipe = KFD_CIK_HIQ_PIPE; 160 kq->queue->queue = KFD_CIK_HIQ_QUEUE; 161 kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd, 162 kq->queue->pipe, kq->queue->queue, 163 &kq->queue->properties, NULL); 164 } else { 165 /* allocate fence for DIQ */ 166 167 retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t), 168 &kq->fence_mem_obj); 169 170 if (retval != 0) 171 goto err_alloc_fence; 172 173 kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr; 174 kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr; 175 } 176 177 print_queue(kq->queue); 178 179 return true; 180 err_alloc_fence: 181 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, kq->queue->mqd_mem_obj); 182 err_allocate_mqd: 183 uninit_queue(kq->queue); 184 err_init_queue: 185 kfd_gtt_sa_free(dev, kq->wptr_mem); 186 err_wptr_allocate_vidmem: 187 kfd_gtt_sa_free(dev, kq->rptr_mem); 188 err_rptr_allocate_vidmem: 189 kfd_gtt_sa_free(dev, kq->eop_mem); 190 err_eop_allocate_vidmem: 191 kfd_gtt_sa_free(dev, kq->pq); 192 err_pq_allocate_vidmem: 193 kfd_release_kernel_doorbell(dev->kfd, prop.doorbell_ptr); 194 err_get_kernel_doorbell: 195 return false; 196 197 } 198 199 /* Uninitialize a kernel queue and free all its memory usages. */ 200 static void kq_uninitialize(struct kernel_queue *kq) 201 { 202 if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ && down_read_trylock(&kq->dev->adev->reset_domain->sem)) { 203 kq->mqd_mgr->destroy_mqd(kq->mqd_mgr, 204 kq->queue->mqd, 205 KFD_PREEMPT_TYPE_WAVEFRONT_RESET, 206 KFD_UNMAP_LATENCY_MS, 207 kq->queue->pipe, 208 kq->queue->queue); 209 up_read(&kq->dev->adev->reset_domain->sem); 210 } 211 else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ) 212 kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj); 213 214 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, 215 kq->queue->mqd_mem_obj); 216 217 kfd_gtt_sa_free(kq->dev, kq->rptr_mem); 218 kfd_gtt_sa_free(kq->dev, kq->wptr_mem); 219 220 /* For CIK family asics, kq->eop_mem is Null, kfd_gtt_sa_free() 221 * is able to handle NULL properly. 222 */ 223 kfd_gtt_sa_free(kq->dev, kq->eop_mem); 224 225 kfd_gtt_sa_free(kq->dev, kq->pq); 226 kfd_release_kernel_doorbell(kq->dev->kfd, 227 kq->queue->properties.doorbell_ptr); 228 uninit_queue(kq->queue); 229 } 230 231 int kq_acquire_packet_buffer(struct kernel_queue *kq, 232 size_t packet_size_in_dwords, unsigned int **buffer_ptr) 233 { 234 size_t available_size; 235 size_t queue_size_dwords; 236 uint32_t wptr, rptr; 237 uint64_t wptr64; 238 unsigned int *queue_address; 239 240 /* When rptr == wptr, the buffer is empty. 241 * When rptr == wptr + 1, the buffer is full. 242 * It is always rptr that advances to the position of wptr, rather than 243 * the opposite. So we can only use up to queue_size_dwords - 1 dwords. 244 */ 245 rptr = *kq->rptr_kernel; 246 wptr = kq->pending_wptr; 247 wptr64 = kq->pending_wptr64; 248 queue_address = (unsigned int *)kq->pq_kernel_addr; 249 queue_size_dwords = kq->queue->properties.queue_size / 4; 250 251 pr_debug("rptr: %d\n", rptr); 252 pr_debug("wptr: %d\n", wptr); 253 pr_debug("queue_address 0x%p\n", queue_address); 254 255 available_size = (rptr + queue_size_dwords - 1 - wptr) % 256 queue_size_dwords; 257 258 if (packet_size_in_dwords > available_size) { 259 /* 260 * make sure calling functions know 261 * acquire_packet_buffer() failed 262 */ 263 goto err_no_space; 264 } 265 266 if (wptr + packet_size_in_dwords >= queue_size_dwords) { 267 /* make sure after rolling back to position 0, there is 268 * still enough space. 269 */ 270 if (packet_size_in_dwords >= rptr) 271 goto err_no_space; 272 273 /* fill nops, roll back and start at position 0 */ 274 while (wptr > 0) { 275 queue_address[wptr] = kq->nop_packet; 276 wptr = (wptr + 1) % queue_size_dwords; 277 wptr64++; 278 } 279 } 280 281 *buffer_ptr = &queue_address[wptr]; 282 kq->pending_wptr = wptr + packet_size_in_dwords; 283 kq->pending_wptr64 = wptr64 + packet_size_in_dwords; 284 285 return 0; 286 287 err_no_space: 288 *buffer_ptr = NULL; 289 return -ENOMEM; 290 } 291 292 int kq_submit_packet(struct kernel_queue *kq) 293 { 294 #ifdef DEBUG 295 int i; 296 297 for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) { 298 pr_debug("0x%2X ", kq->pq_kernel_addr[i]); 299 if (i % 15 == 0) 300 pr_debug("\n"); 301 } 302 pr_debug("\n"); 303 #endif 304 /* Fatal err detected, packet submission won't go through */ 305 if (amdgpu_amdkfd_is_fed(kq->dev->adev)) 306 return -EIO; 307 308 if (kq->dev->kfd->device_info.doorbell_size == 8) { 309 *kq->wptr64_kernel = kq->pending_wptr64; 310 write_kernel_doorbell64(kq->queue->properties.doorbell_ptr, 311 kq->pending_wptr64); 312 } else { 313 *kq->wptr_kernel = kq->pending_wptr; 314 write_kernel_doorbell(kq->queue->properties.doorbell_ptr, 315 kq->pending_wptr); 316 } 317 318 return 0; 319 } 320 321 void kq_rollback_packet(struct kernel_queue *kq) 322 { 323 if (kq->dev->kfd->device_info.doorbell_size == 8) { 324 kq->pending_wptr64 = *kq->wptr64_kernel; 325 kq->pending_wptr = *kq->wptr_kernel % 326 (kq->queue->properties.queue_size / 4); 327 } else { 328 kq->pending_wptr = *kq->wptr_kernel; 329 } 330 } 331 332 struct kernel_queue *kernel_queue_init(struct kfd_node *dev, 333 enum kfd_queue_type type) 334 { 335 struct kernel_queue *kq; 336 337 kq = kzalloc(sizeof(*kq), GFP_KERNEL); 338 if (!kq) 339 return NULL; 340 341 if (kq_initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE)) 342 return kq; 343 344 pr_err("Failed to init kernel queue\n"); 345 346 kfree(kq); 347 return NULL; 348 } 349 350 void kernel_queue_uninit(struct kernel_queue *kq) 351 { 352 kq_uninitialize(kq); 353 kfree(kq); 354 } 355 356 /* FIXME: Can this test be removed? */ 357 static __attribute__((unused)) void test_kq(struct kfd_node *dev) 358 { 359 struct kernel_queue *kq; 360 uint32_t *buffer, i; 361 int retval; 362 363 pr_err("Starting kernel queue test\n"); 364 365 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ); 366 if (unlikely(!kq)) { 367 pr_err(" Failed to initialize HIQ\n"); 368 pr_err("Kernel queue test failed\n"); 369 return; 370 } 371 372 retval = kq_acquire_packet_buffer(kq, 5, &buffer); 373 if (unlikely(retval != 0)) { 374 pr_err(" Failed to acquire packet buffer\n"); 375 pr_err("Kernel queue test failed\n"); 376 return; 377 } 378 for (i = 0; i < 5; i++) 379 buffer[i] = kq->nop_packet; 380 kq_submit_packet(kq); 381 382 pr_err("Ending kernel queue test\n"); 383 } 384 385 386