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 dev_err(dev->adev->dev, "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 dev_err(dev->adev->dev, "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 dev_err(dev->adev->dev, "Failed to init pq queues size %d\n", 89 queue_size); 90 goto err_pq_allocate_vidmem; 91 } 92 93 kq->pq_kernel_addr = kq->pq->cpu_ptr; 94 kq->pq_gpu_addr = kq->pq->gpu_addr; 95 96 /* For CIK family asics, kq->eop_mem is not needed */ 97 if (dev->adev->asic_type > CHIP_MULLINS) { 98 retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem); 99 if (retval != 0) 100 goto err_eop_allocate_vidmem; 101 102 kq->eop_gpu_addr = kq->eop_mem->gpu_addr; 103 kq->eop_kernel_addr = kq->eop_mem->cpu_ptr; 104 105 memset(kq->eop_kernel_addr, 0, PAGE_SIZE); 106 } 107 108 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel), 109 &kq->rptr_mem); 110 111 if (retval != 0) 112 goto err_rptr_allocate_vidmem; 113 114 kq->rptr_kernel = kq->rptr_mem->cpu_ptr; 115 kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr; 116 117 retval = kfd_gtt_sa_allocate(dev, dev->kfd->device_info.doorbell_size, 118 &kq->wptr_mem); 119 120 if (retval != 0) 121 goto err_wptr_allocate_vidmem; 122 123 kq->wptr_kernel = kq->wptr_mem->cpu_ptr; 124 kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr; 125 126 memset(kq->pq_kernel_addr, 0, queue_size); 127 memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel)); 128 memset(kq->wptr_kernel, 0, sizeof(*kq->wptr_kernel)); 129 130 prop.queue_size = queue_size; 131 prop.is_interop = false; 132 prop.is_gws = false; 133 prop.priority = 1; 134 prop.queue_percent = 100; 135 prop.type = type; 136 prop.vmid = 0; 137 prop.queue_address = kq->pq_gpu_addr; 138 prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr; 139 prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr; 140 prop.eop_ring_buffer_address = kq->eop_gpu_addr; 141 prop.eop_ring_buffer_size = PAGE_SIZE; 142 143 if (init_queue(&kq->queue, &prop) != 0) 144 goto err_init_queue; 145 146 kq->queue->device = dev; 147 kq->queue->process = kfd_get_process(current); 148 149 kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev, 150 &kq->queue->properties); 151 if (!kq->queue->mqd_mem_obj) 152 goto err_allocate_mqd; 153 kq->mqd_mgr->init_mqd(kq->mqd_mgr, &kq->queue->mqd, 154 kq->queue->mqd_mem_obj, 155 &kq->queue->gart_mqd_addr, 156 &kq->queue->properties); 157 /* assign HIQ to HQD */ 158 if (type == KFD_QUEUE_TYPE_HIQ) { 159 pr_debug("Assigning hiq to hqd\n"); 160 kq->queue->pipe = KFD_CIK_HIQ_PIPE; 161 kq->queue->queue = KFD_CIK_HIQ_QUEUE; 162 kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd, 163 kq->queue->pipe, kq->queue->queue, 164 &kq->queue->properties, NULL); 165 } else { 166 /* allocate fence for DIQ */ 167 168 retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t), 169 &kq->fence_mem_obj); 170 171 if (retval != 0) 172 goto err_alloc_fence; 173 174 kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr; 175 kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr; 176 } 177 178 print_queue(kq->queue); 179 180 return true; 181 err_alloc_fence: 182 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, kq->queue->mqd_mem_obj); 183 err_allocate_mqd: 184 uninit_queue(kq->queue); 185 err_init_queue: 186 kfd_gtt_sa_free(dev, kq->wptr_mem); 187 err_wptr_allocate_vidmem: 188 kfd_gtt_sa_free(dev, kq->rptr_mem); 189 err_rptr_allocate_vidmem: 190 kfd_gtt_sa_free(dev, kq->eop_mem); 191 err_eop_allocate_vidmem: 192 kfd_gtt_sa_free(dev, kq->pq); 193 err_pq_allocate_vidmem: 194 kfd_release_kernel_doorbell(dev->kfd, prop.doorbell_ptr); 195 err_get_kernel_doorbell: 196 return false; 197 198 } 199 200 /* Uninitialize a kernel queue and free all its memory usages. */ 201 static void kq_uninitialize(struct kernel_queue *kq) 202 { 203 if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ && down_read_trylock(&kq->dev->adev->reset_domain->sem)) { 204 kq->mqd_mgr->destroy_mqd(kq->mqd_mgr, 205 kq->queue->mqd, 206 KFD_PREEMPT_TYPE_WAVEFRONT_RESET, 207 KFD_UNMAP_LATENCY_MS, 208 kq->queue->pipe, 209 kq->queue->queue); 210 up_read(&kq->dev->adev->reset_domain->sem); 211 } 212 else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ) 213 kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj); 214 215 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, 216 kq->queue->mqd_mem_obj); 217 218 kfd_gtt_sa_free(kq->dev, kq->rptr_mem); 219 kfd_gtt_sa_free(kq->dev, kq->wptr_mem); 220 221 /* For CIK family asics, kq->eop_mem is Null, kfd_gtt_sa_free() 222 * is able to handle NULL properly. 223 */ 224 kfd_gtt_sa_free(kq->dev, kq->eop_mem); 225 226 kfd_gtt_sa_free(kq->dev, kq->pq); 227 kfd_release_kernel_doorbell(kq->dev->kfd, 228 kq->queue->properties.doorbell_ptr); 229 uninit_queue(kq->queue); 230 } 231 232 int kq_acquire_packet_buffer(struct kernel_queue *kq, 233 size_t packet_size_in_dwords, unsigned int **buffer_ptr) 234 { 235 size_t available_size; 236 size_t queue_size_dwords; 237 uint32_t wptr, rptr; 238 uint64_t wptr64; 239 unsigned int *queue_address; 240 241 /* When rptr == wptr, the buffer is empty. 242 * When rptr == wptr + 1, the buffer is full. 243 * It is always rptr that advances to the position of wptr, rather than 244 * the opposite. So we can only use up to queue_size_dwords - 1 dwords. 245 */ 246 rptr = *kq->rptr_kernel; 247 wptr = kq->pending_wptr; 248 wptr64 = kq->pending_wptr64; 249 queue_address = (unsigned int *)kq->pq_kernel_addr; 250 queue_size_dwords = kq->queue->properties.queue_size / 4; 251 252 pr_debug("rptr: %d\n", rptr); 253 pr_debug("wptr: %d\n", wptr); 254 pr_debug("queue_address 0x%p\n", queue_address); 255 256 available_size = (rptr + queue_size_dwords - 1 - wptr) % 257 queue_size_dwords; 258 259 if (packet_size_in_dwords > available_size) { 260 /* 261 * make sure calling functions know 262 * acquire_packet_buffer() failed 263 */ 264 goto err_no_space; 265 } 266 267 if (wptr + packet_size_in_dwords >= queue_size_dwords) { 268 /* make sure after rolling back to position 0, there is 269 * still enough space. 270 */ 271 if (packet_size_in_dwords >= rptr) 272 goto err_no_space; 273 274 /* fill nops, roll back and start at position 0 */ 275 while (wptr > 0) { 276 queue_address[wptr] = kq->nop_packet; 277 wptr = (wptr + 1) % queue_size_dwords; 278 wptr64++; 279 } 280 } 281 282 *buffer_ptr = &queue_address[wptr]; 283 kq->pending_wptr = wptr + packet_size_in_dwords; 284 kq->pending_wptr64 = wptr64 + packet_size_in_dwords; 285 286 return 0; 287 288 err_no_space: 289 *buffer_ptr = NULL; 290 return -ENOMEM; 291 } 292 293 int kq_submit_packet(struct kernel_queue *kq) 294 { 295 #ifdef DEBUG 296 int i; 297 298 for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) { 299 pr_debug("0x%2X ", kq->pq_kernel_addr[i]); 300 if (i % 15 == 0) 301 pr_debug("\n"); 302 } 303 pr_debug("\n"); 304 #endif 305 /* Fatal err detected, packet submission won't go through */ 306 if (amdgpu_amdkfd_is_fed(kq->dev->adev)) 307 return -EIO; 308 309 if (kq->dev->kfd->device_info.doorbell_size == 8) { 310 *kq->wptr64_kernel = kq->pending_wptr64; 311 write_kernel_doorbell64(kq->queue->properties.doorbell_ptr, 312 kq->pending_wptr64); 313 } else { 314 *kq->wptr_kernel = kq->pending_wptr; 315 write_kernel_doorbell(kq->queue->properties.doorbell_ptr, 316 kq->pending_wptr); 317 } 318 319 return 0; 320 } 321 322 void kq_rollback_packet(struct kernel_queue *kq) 323 { 324 if (kq->dev->kfd->device_info.doorbell_size == 8) { 325 kq->pending_wptr64 = *kq->wptr64_kernel; 326 kq->pending_wptr = *kq->wptr_kernel % 327 (kq->queue->properties.queue_size / 4); 328 } else { 329 kq->pending_wptr = *kq->wptr_kernel; 330 } 331 } 332 333 struct kernel_queue *kernel_queue_init(struct kfd_node *dev, 334 enum kfd_queue_type type) 335 { 336 struct kernel_queue *kq; 337 338 kq = kzalloc(sizeof(*kq), GFP_KERNEL); 339 if (!kq) 340 return NULL; 341 342 if (kq_initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE)) 343 return kq; 344 345 dev_err(dev->adev->dev, "Failed to init kernel queue\n"); 346 347 kfree(kq); 348 return NULL; 349 } 350 351 void kernel_queue_uninit(struct kernel_queue *kq) 352 { 353 kq_uninitialize(kq); 354 kfree(kq); 355 } 356 357 /* FIXME: Can this test be removed? */ 358 static __attribute__((unused)) void test_kq(struct kfd_node *dev) 359 { 360 struct kernel_queue *kq; 361 uint32_t *buffer, i; 362 int retval; 363 364 dev_err(dev->adev->dev, "Starting kernel queue test\n"); 365 366 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ); 367 if (unlikely(!kq)) { 368 dev_err(dev->adev->dev, " Failed to initialize HIQ\n"); 369 dev_err(dev->adev->dev, "Kernel queue test failed\n"); 370 return; 371 } 372 373 retval = kq_acquire_packet_buffer(kq, 5, &buffer); 374 if (unlikely(retval != 0)) { 375 dev_err(dev->adev->dev, " Failed to acquire packet buffer\n"); 376 dev_err(dev->adev->dev, "Kernel queue test failed\n"); 377 return; 378 } 379 for (i = 0; i < 5; i++) 380 buffer[i] = kq->nop_packet; 381 kq_submit_packet(kq); 382 383 dev_err(dev->adev->dev, "Ending kernel queue test\n"); 384 } 385 386 387