1 1.11 mrg /* $NetBSD: amdgpu_fence.c,v 1.11 2023/07/20 18:02:45 mrg Exp $ */ 2 1.1 riastrad 3 1.1 riastrad /* 4 1.1 riastrad * Copyright 2009 Jerome Glisse. 5 1.1 riastrad * All Rights Reserved. 6 1.1 riastrad * 7 1.1 riastrad * Permission is hereby granted, free of charge, to any person obtaining a 8 1.1 riastrad * copy of this software and associated documentation files (the 9 1.1 riastrad * "Software"), to deal in the Software without restriction, including 10 1.1 riastrad * without limitation the rights to use, copy, modify, merge, publish, 11 1.1 riastrad * distribute, sub license, and/or sell copies of the Software, and to 12 1.1 riastrad * permit persons to whom the Software is furnished to do so, subject to 13 1.1 riastrad * the following conditions: 14 1.1 riastrad * 15 1.1 riastrad * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 1.1 riastrad * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 1.1 riastrad * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 18 1.1 riastrad * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 19 1.1 riastrad * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 20 1.1 riastrad * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 21 1.1 riastrad * USE OR OTHER DEALINGS IN THE SOFTWARE. 22 1.1 riastrad * 23 1.1 riastrad * The above copyright notice and this permission notice (including the 24 1.1 riastrad * next paragraph) shall be included in all copies or substantial portions 25 1.1 riastrad * of the Software. 26 1.1 riastrad * 27 1.1 riastrad */ 28 1.1 riastrad /* 29 1.1 riastrad * Authors: 30 1.1 riastrad * Jerome Glisse <glisse (at) freedesktop.org> 31 1.1 riastrad * Dave Airlie 32 1.1 riastrad */ 33 1.1 riastrad #include <sys/cdefs.h> 34 1.11 mrg __KERNEL_RCSID(0, "$NetBSD: amdgpu_fence.c,v 1.11 2023/07/20 18:02:45 mrg Exp $"); 35 1.1 riastrad 36 1.1 riastrad #include <linux/seq_file.h> 37 1.1 riastrad #include <linux/atomic.h> 38 1.1 riastrad #include <linux/wait.h> 39 1.1 riastrad #include <linux/kref.h> 40 1.1 riastrad #include <linux/slab.h> 41 1.1 riastrad #include <linux/firmware.h> 42 1.7 riastrad #include <linux/pm_runtime.h> 43 1.7 riastrad 44 1.7 riastrad #include <drm/drm_debugfs.h> 45 1.7 riastrad 46 1.1 riastrad #include "amdgpu.h" 47 1.1 riastrad #include "amdgpu_trace.h" 48 1.1 riastrad 49 1.5 riastrad #include <linux/nbsd-namespace.h> 50 1.5 riastrad 51 1.1 riastrad /* 52 1.1 riastrad * Fences 53 1.1 riastrad * Fences mark an event in the GPUs pipeline and are used 54 1.1 riastrad * for GPU/CPU synchronization. When the fence is written, 55 1.1 riastrad * it is expected that all buffers associated with that fence 56 1.1 riastrad * are no longer in use by the associated ring on the GPU and 57 1.1 riastrad * that the the relevant GPU caches have been flushed. 58 1.1 riastrad */ 59 1.1 riastrad 60 1.7 riastrad struct amdgpu_fence { 61 1.7 riastrad struct dma_fence base; 62 1.7 riastrad 63 1.7 riastrad /* RB, DMA, etc. */ 64 1.7 riastrad struct amdgpu_ring *ring; 65 1.7 riastrad }; 66 1.7 riastrad 67 1.1 riastrad static struct kmem_cache *amdgpu_fence_slab; 68 1.7 riastrad 69 1.7 riastrad int amdgpu_fence_slab_init(void) 70 1.7 riastrad { 71 1.7 riastrad amdgpu_fence_slab = kmem_cache_create( 72 1.7 riastrad "amdgpu_fence", sizeof(struct amdgpu_fence), 0, 73 1.7 riastrad SLAB_HWCACHE_ALIGN, NULL); 74 1.7 riastrad if (!amdgpu_fence_slab) 75 1.7 riastrad return -ENOMEM; 76 1.7 riastrad return 0; 77 1.7 riastrad } 78 1.7 riastrad 79 1.7 riastrad void amdgpu_fence_slab_fini(void) 80 1.7 riastrad { 81 1.7 riastrad rcu_barrier(); 82 1.7 riastrad kmem_cache_destroy(amdgpu_fence_slab); 83 1.7 riastrad } 84 1.7 riastrad /* 85 1.7 riastrad * Cast helper 86 1.7 riastrad */ 87 1.7 riastrad static const struct dma_fence_ops amdgpu_fence_ops; 88 1.7 riastrad static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f) 89 1.7 riastrad { 90 1.7 riastrad struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base); 91 1.7 riastrad 92 1.7 riastrad if (__f->base.ops == &amdgpu_fence_ops) 93 1.7 riastrad return __f; 94 1.7 riastrad 95 1.7 riastrad return NULL; 96 1.7 riastrad } 97 1.1 riastrad 98 1.1 riastrad /** 99 1.1 riastrad * amdgpu_fence_write - write a fence value 100 1.1 riastrad * 101 1.1 riastrad * @ring: ring the fence is associated with 102 1.1 riastrad * @seq: sequence number to write 103 1.1 riastrad * 104 1.1 riastrad * Writes a fence value to memory (all asics). 105 1.1 riastrad */ 106 1.1 riastrad static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq) 107 1.1 riastrad { 108 1.1 riastrad struct amdgpu_fence_driver *drv = &ring->fence_drv; 109 1.1 riastrad 110 1.1 riastrad if (drv->cpu_addr) 111 1.1 riastrad *drv->cpu_addr = cpu_to_le32(seq); 112 1.1 riastrad } 113 1.1 riastrad 114 1.1 riastrad /** 115 1.1 riastrad * amdgpu_fence_read - read a fence value 116 1.1 riastrad * 117 1.1 riastrad * @ring: ring the fence is associated with 118 1.1 riastrad * 119 1.1 riastrad * Reads a fence value from memory (all asics). 120 1.1 riastrad * Returns the value of the fence read from memory. 121 1.1 riastrad */ 122 1.1 riastrad static u32 amdgpu_fence_read(struct amdgpu_ring *ring) 123 1.1 riastrad { 124 1.1 riastrad struct amdgpu_fence_driver *drv = &ring->fence_drv; 125 1.1 riastrad u32 seq = 0; 126 1.1 riastrad 127 1.1 riastrad if (drv->cpu_addr) 128 1.1 riastrad seq = le32_to_cpu(*drv->cpu_addr); 129 1.1 riastrad else 130 1.7 riastrad seq = atomic_read(&drv->last_seq); 131 1.1 riastrad 132 1.1 riastrad return seq; 133 1.1 riastrad } 134 1.1 riastrad 135 1.1 riastrad /** 136 1.1 riastrad * amdgpu_fence_emit - emit a fence on the requested ring 137 1.1 riastrad * 138 1.1 riastrad * @ring: ring the fence is associated with 139 1.7 riastrad * @f: resulting fence object 140 1.1 riastrad * 141 1.1 riastrad * Emits a fence command on the requested ring (all asics). 142 1.1 riastrad * Returns 0 on success, -ENOMEM on failure. 143 1.1 riastrad */ 144 1.7 riastrad int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, 145 1.7 riastrad unsigned flags) 146 1.1 riastrad { 147 1.1 riastrad struct amdgpu_device *adev = ring->adev; 148 1.7 riastrad struct amdgpu_fence *fence; 149 1.7 riastrad struct dma_fence __rcu **ptr; 150 1.7 riastrad uint32_t seq; 151 1.7 riastrad int r; 152 1.1 riastrad 153 1.7 riastrad fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); 154 1.7 riastrad if (fence == NULL) 155 1.1 riastrad return -ENOMEM; 156 1.7 riastrad 157 1.7 riastrad seq = ++ring->fence_drv.sync_seq; 158 1.7 riastrad fence->ring = ring; 159 1.7 riastrad dma_fence_init(&fence->base, &amdgpu_fence_ops, 160 1.7 riastrad &ring->fence_drv.lock, 161 1.7 riastrad adev->fence_context + ring->idx, 162 1.7 riastrad seq); 163 1.7 riastrad amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, 164 1.7 riastrad seq, flags | AMDGPU_FENCE_FLAG_INT); 165 1.7 riastrad pm_runtime_get_noresume(adev->ddev->dev); 166 1.7 riastrad ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask]; 167 1.7 riastrad if (unlikely(rcu_dereference_protected(*ptr, 1))) { 168 1.7 riastrad struct dma_fence *old; 169 1.7 riastrad 170 1.7 riastrad rcu_read_lock(); 171 1.7 riastrad old = dma_fence_get_rcu_safe(ptr); 172 1.7 riastrad rcu_read_unlock(); 173 1.7 riastrad 174 1.7 riastrad if (old) { 175 1.7 riastrad r = dma_fence_wait(old, false); 176 1.7 riastrad dma_fence_put(old); 177 1.7 riastrad if (r) 178 1.7 riastrad return r; 179 1.7 riastrad } 180 1.1 riastrad } 181 1.7 riastrad 182 1.7 riastrad /* This function can't be called concurrently anyway, otherwise 183 1.7 riastrad * emitting the fence would mess up the hardware ring buffer. 184 1.7 riastrad */ 185 1.7 riastrad rcu_assign_pointer(*ptr, dma_fence_get(&fence->base)); 186 1.7 riastrad 187 1.7 riastrad *f = &fence->base; 188 1.7 riastrad 189 1.7 riastrad return 0; 190 1.7 riastrad } 191 1.7 riastrad 192 1.7 riastrad /** 193 1.7 riastrad * amdgpu_fence_emit_polling - emit a fence on the requeste ring 194 1.7 riastrad * 195 1.7 riastrad * @ring: ring the fence is associated with 196 1.7 riastrad * @s: resulting sequence number 197 1.7 riastrad * 198 1.7 riastrad * Emits a fence command on the requested ring (all asics). 199 1.7 riastrad * Used For polling fence. 200 1.7 riastrad * Returns 0 on success, -ENOMEM on failure. 201 1.7 riastrad */ 202 1.7 riastrad int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s) 203 1.7 riastrad { 204 1.7 riastrad uint32_t seq; 205 1.7 riastrad 206 1.7 riastrad if (!s) 207 1.7 riastrad return -EINVAL; 208 1.7 riastrad 209 1.7 riastrad seq = ++ring->fence_drv.sync_seq; 210 1.1 riastrad amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, 211 1.7 riastrad seq, 0); 212 1.7 riastrad 213 1.7 riastrad *s = seq; 214 1.7 riastrad 215 1.1 riastrad return 0; 216 1.1 riastrad } 217 1.1 riastrad 218 1.1 riastrad /** 219 1.1 riastrad * amdgpu_fence_schedule_fallback - schedule fallback check 220 1.1 riastrad * 221 1.1 riastrad * @ring: pointer to struct amdgpu_ring 222 1.1 riastrad * 223 1.1 riastrad * Start a timer as fallback to our interrupts. 224 1.1 riastrad */ 225 1.1 riastrad static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring) 226 1.1 riastrad { 227 1.1 riastrad mod_timer(&ring->fence_drv.fallback_timer, 228 1.1 riastrad jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT); 229 1.1 riastrad } 230 1.1 riastrad 231 1.1 riastrad /** 232 1.7 riastrad * amdgpu_fence_process - check for fence activity 233 1.1 riastrad * 234 1.1 riastrad * @ring: pointer to struct amdgpu_ring 235 1.1 riastrad * 236 1.1 riastrad * Checks the current fence value and calculates the last 237 1.7 riastrad * signalled fence value. Wakes the fence queue if the 238 1.7 riastrad * sequence number has increased. 239 1.7 riastrad * 240 1.7 riastrad * Returns true if fence was processed 241 1.1 riastrad */ 242 1.7 riastrad bool amdgpu_fence_process(struct amdgpu_ring *ring) 243 1.1 riastrad { 244 1.7 riastrad struct amdgpu_fence_driver *drv = &ring->fence_drv; 245 1.7 riastrad struct amdgpu_device *adev = ring->adev; 246 1.7 riastrad uint32_t seq, last_seq; 247 1.7 riastrad int r; 248 1.1 riastrad 249 1.1 riastrad do { 250 1.7 riastrad last_seq = atomic_read(&ring->fence_drv.last_seq); 251 1.1 riastrad seq = amdgpu_fence_read(ring); 252 1.1 riastrad 253 1.7 riastrad } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq); 254 1.1 riastrad 255 1.7 riastrad if (del_timer(&ring->fence_drv.fallback_timer) && 256 1.7 riastrad seq != ring->fence_drv.sync_seq) 257 1.1 riastrad amdgpu_fence_schedule_fallback(ring); 258 1.1 riastrad 259 1.7 riastrad if (unlikely(seq == last_seq)) 260 1.7 riastrad return false; 261 1.7 riastrad 262 1.7 riastrad last_seq &= drv->num_fences_mask; 263 1.7 riastrad seq &= drv->num_fences_mask; 264 1.7 riastrad 265 1.7 riastrad do { 266 1.7 riastrad struct dma_fence *fence, **ptr; 267 1.7 riastrad 268 1.7 riastrad ++last_seq; 269 1.7 riastrad last_seq &= drv->num_fences_mask; 270 1.7 riastrad ptr = &drv->fences[last_seq]; 271 1.7 riastrad 272 1.7 riastrad /* There is always exactly one thread signaling this fence slot */ 273 1.7 riastrad fence = rcu_dereference_protected(*ptr, 1); 274 1.7 riastrad RCU_INIT_POINTER(*ptr, NULL); 275 1.7 riastrad 276 1.7 riastrad if (!fence) 277 1.7 riastrad continue; 278 1.7 riastrad 279 1.7 riastrad r = dma_fence_signal(fence); 280 1.7 riastrad if (!r) 281 1.7 riastrad DMA_FENCE_TRACE(fence, "signaled from irq context\n"); 282 1.7 riastrad else 283 1.7 riastrad BUG(); 284 1.7 riastrad 285 1.7 riastrad dma_fence_put(fence); 286 1.7 riastrad pm_runtime_mark_last_busy(adev->ddev->dev); 287 1.7 riastrad pm_runtime_put_autosuspend(adev->ddev->dev); 288 1.7 riastrad } while (last_seq != seq); 289 1.7 riastrad 290 1.7 riastrad return true; 291 1.1 riastrad } 292 1.1 riastrad 293 1.1 riastrad /** 294 1.1 riastrad * amdgpu_fence_fallback - fallback for hardware interrupts 295 1.1 riastrad * 296 1.1 riastrad * @work: delayed work item 297 1.1 riastrad * 298 1.1 riastrad * Checks for fence activity. 299 1.1 riastrad */ 300 1.7 riastrad static void amdgpu_fence_fallback(struct timer_list *t) 301 1.1 riastrad { 302 1.7 riastrad struct amdgpu_ring *ring = from_timer(ring, t, 303 1.7 riastrad fence_drv.fallback_timer); 304 1.1 riastrad 305 1.7 riastrad if (amdgpu_fence_process(ring)) 306 1.7 riastrad DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name); 307 1.1 riastrad } 308 1.1 riastrad 309 1.1 riastrad /** 310 1.7 riastrad * amdgpu_fence_wait_empty - wait for all fences to signal 311 1.1 riastrad * 312 1.7 riastrad * @adev: amdgpu device pointer 313 1.7 riastrad * @ring: ring index the fence is associated with 314 1.1 riastrad * 315 1.7 riastrad * Wait for all fences on the requested ring to signal (all asics). 316 1.7 riastrad * Returns 0 if the fences have passed, error for all other cases. 317 1.1 riastrad */ 318 1.7 riastrad int amdgpu_fence_wait_empty(struct amdgpu_ring *ring) 319 1.1 riastrad { 320 1.7 riastrad uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq); 321 1.7 riastrad struct dma_fence *fence, **ptr; 322 1.7 riastrad int r; 323 1.1 riastrad 324 1.7 riastrad if (!seq) 325 1.1 riastrad return 0; 326 1.1 riastrad 327 1.7 riastrad ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask]; 328 1.7 riastrad rcu_read_lock(); 329 1.7 riastrad fence = rcu_dereference(*ptr); 330 1.7 riastrad if (!fence || !dma_fence_get_rcu(fence)) { 331 1.7 riastrad rcu_read_unlock(); 332 1.1 riastrad return 0; 333 1.7 riastrad } 334 1.7 riastrad rcu_read_unlock(); 335 1.1 riastrad 336 1.7 riastrad r = dma_fence_wait(fence, false); 337 1.7 riastrad dma_fence_put(fence); 338 1.7 riastrad return r; 339 1.1 riastrad } 340 1.1 riastrad 341 1.1 riastrad /** 342 1.7 riastrad * amdgpu_fence_wait_polling - busy wait for givn sequence number 343 1.1 riastrad * 344 1.1 riastrad * @ring: ring index the fence is associated with 345 1.7 riastrad * @wait_seq: sequence number to wait 346 1.7 riastrad * @timeout: the timeout for waiting in usecs 347 1.1 riastrad * 348 1.1 riastrad * Wait for all fences on the requested ring to signal (all asics). 349 1.7 riastrad * Returns left time if no timeout, 0 or minus if timeout. 350 1.1 riastrad */ 351 1.7 riastrad signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring, 352 1.7 riastrad uint32_t wait_seq, 353 1.7 riastrad signed long timeout) 354 1.1 riastrad { 355 1.7 riastrad uint32_t seq; 356 1.1 riastrad 357 1.7 riastrad do { 358 1.7 riastrad seq = amdgpu_fence_read(ring); 359 1.7 riastrad udelay(5); 360 1.7 riastrad timeout -= 5; 361 1.7 riastrad } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0); 362 1.1 riastrad 363 1.7 riastrad return timeout > 0 ? timeout : 0; 364 1.1 riastrad } 365 1.1 riastrad /** 366 1.1 riastrad * amdgpu_fence_count_emitted - get the count of emitted fences 367 1.1 riastrad * 368 1.1 riastrad * @ring: ring the fence is associated with 369 1.1 riastrad * 370 1.1 riastrad * Get the number of fences emitted on the requested ring (all asics). 371 1.1 riastrad * Returns the number of emitted fences on the ring. Used by the 372 1.1 riastrad * dynpm code to ring track activity. 373 1.1 riastrad */ 374 1.1 riastrad unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) 375 1.1 riastrad { 376 1.1 riastrad uint64_t emitted; 377 1.1 riastrad 378 1.1 riastrad /* We are not protected by ring lock when reading the last sequence 379 1.1 riastrad * but it's ok to report slightly wrong fence count here. 380 1.1 riastrad */ 381 1.1 riastrad amdgpu_fence_process(ring); 382 1.7 riastrad emitted = 0x100000000ull; 383 1.7 riastrad emitted -= atomic_read(&ring->fence_drv.last_seq); 384 1.7 riastrad emitted += READ_ONCE(ring->fence_drv.sync_seq); 385 1.7 riastrad return lower_32_bits(emitted); 386 1.1 riastrad } 387 1.1 riastrad 388 1.1 riastrad /** 389 1.1 riastrad * amdgpu_fence_driver_start_ring - make the fence driver 390 1.1 riastrad * ready for use on the requested ring. 391 1.1 riastrad * 392 1.1 riastrad * @ring: ring to start the fence driver on 393 1.1 riastrad * @irq_src: interrupt source to use for this ring 394 1.1 riastrad * @irq_type: interrupt type to use for this ring 395 1.1 riastrad * 396 1.1 riastrad * Make the fence driver ready for processing (all asics). 397 1.1 riastrad * Not all asics have all rings, so each asic will only 398 1.1 riastrad * start the fence driver on the rings it has. 399 1.1 riastrad * Returns 0 for success, errors for failure. 400 1.1 riastrad */ 401 1.1 riastrad int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, 402 1.1 riastrad struct amdgpu_irq_src *irq_src, 403 1.1 riastrad unsigned irq_type) 404 1.1 riastrad { 405 1.1 riastrad struct amdgpu_device *adev = ring->adev; 406 1.1 riastrad uint64_t index; 407 1.1 riastrad 408 1.7 riastrad if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) { 409 1.1 riastrad ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs]; 410 1.1 riastrad ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4); 411 1.1 riastrad } else { 412 1.1 riastrad /* put fence directly behind firmware */ 413 1.1 riastrad index = ALIGN(adev->uvd.fw->size, 8); 414 1.7 riastrad ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index; 415 1.7 riastrad ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index; 416 1.1 riastrad } 417 1.7 riastrad amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq)); 418 1.1 riastrad amdgpu_irq_get(adev, irq_src, irq_type); 419 1.1 riastrad 420 1.1 riastrad ring->fence_drv.irq_src = irq_src; 421 1.1 riastrad ring->fence_drv.irq_type = irq_type; 422 1.1 riastrad ring->fence_drv.initialized = true; 423 1.1 riastrad 424 1.7 riastrad DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr " 425 1.10 riastrad "0x%016"PRIx64", cpu addr 0x%p\n", ring->name, 426 1.7 riastrad ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr); 427 1.1 riastrad return 0; 428 1.1 riastrad } 429 1.1 riastrad 430 1.1 riastrad /** 431 1.1 riastrad * amdgpu_fence_driver_init_ring - init the fence driver 432 1.1 riastrad * for the requested ring. 433 1.1 riastrad * 434 1.1 riastrad * @ring: ring to init the fence driver on 435 1.7 riastrad * @num_hw_submission: number of entries on the hardware queue 436 1.1 riastrad * 437 1.1 riastrad * Init the fence driver for the requested ring (all asics). 438 1.1 riastrad * Helper function for amdgpu_fence_driver_init(). 439 1.1 riastrad */ 440 1.7 riastrad int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring, 441 1.7 riastrad unsigned num_hw_submission) 442 1.1 riastrad { 443 1.7 riastrad struct amdgpu_device *adev = ring->adev; 444 1.7 riastrad long timeout; 445 1.7 riastrad int r; 446 1.7 riastrad 447 1.7 riastrad if (!adev) 448 1.7 riastrad return -EINVAL; 449 1.7 riastrad 450 1.7 riastrad /* Check that num_hw_submission is a power of two */ 451 1.7 riastrad if ((num_hw_submission & (num_hw_submission - 1)) != 0) 452 1.7 riastrad return -EINVAL; 453 1.1 riastrad 454 1.1 riastrad ring->fence_drv.cpu_addr = NULL; 455 1.1 riastrad ring->fence_drv.gpu_addr = 0; 456 1.7 riastrad ring->fence_drv.sync_seq = 0; 457 1.7 riastrad atomic_set(&ring->fence_drv.last_seq, 0); 458 1.1 riastrad ring->fence_drv.initialized = false; 459 1.1 riastrad 460 1.7 riastrad timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0); 461 1.1 riastrad 462 1.7 riastrad ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1; 463 1.7 riastrad spin_lock_init(&ring->fence_drv.lock); 464 1.7 riastrad ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *), 465 1.7 riastrad GFP_KERNEL); 466 1.7 riastrad if (!ring->fence_drv.fences) 467 1.7 riastrad return -ENOMEM; 468 1.1 riastrad 469 1.7 riastrad /* No need to setup the GPU scheduler for KIQ ring */ 470 1.7 riastrad if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) { 471 1.7 riastrad switch (ring->funcs->type) { 472 1.7 riastrad case AMDGPU_RING_TYPE_GFX: 473 1.7 riastrad timeout = adev->gfx_timeout; 474 1.7 riastrad break; 475 1.7 riastrad case AMDGPU_RING_TYPE_COMPUTE: 476 1.7 riastrad timeout = adev->compute_timeout; 477 1.7 riastrad break; 478 1.7 riastrad case AMDGPU_RING_TYPE_SDMA: 479 1.7 riastrad timeout = adev->sdma_timeout; 480 1.7 riastrad break; 481 1.7 riastrad default: 482 1.7 riastrad timeout = adev->video_timeout; 483 1.7 riastrad break; 484 1.1 riastrad } 485 1.7 riastrad 486 1.7 riastrad r = drm_sched_init(&ring->sched, &amdgpu_sched_ops, 487 1.7 riastrad num_hw_submission, amdgpu_job_hang_limit, 488 1.1 riastrad timeout, ring->name); 489 1.1 riastrad if (r) { 490 1.1 riastrad DRM_ERROR("Failed to create scheduler on ring %s.\n", 491 1.1 riastrad ring->name); 492 1.1 riastrad return r; 493 1.1 riastrad } 494 1.1 riastrad } 495 1.1 riastrad 496 1.1 riastrad return 0; 497 1.1 riastrad } 498 1.1 riastrad 499 1.1 riastrad /** 500 1.1 riastrad * amdgpu_fence_driver_init - init the fence driver 501 1.1 riastrad * for all possible rings. 502 1.1 riastrad * 503 1.1 riastrad * @adev: amdgpu device pointer 504 1.1 riastrad * 505 1.1 riastrad * Init the fence driver for all possible rings (all asics). 506 1.1 riastrad * Not all asics have all rings, so each asic will only 507 1.1 riastrad * start the fence driver on the rings it has using 508 1.1 riastrad * amdgpu_fence_driver_start_ring(). 509 1.1 riastrad * Returns 0 for success. 510 1.1 riastrad */ 511 1.1 riastrad int amdgpu_fence_driver_init(struct amdgpu_device *adev) 512 1.1 riastrad { 513 1.1 riastrad if (amdgpu_debugfs_fence_init(adev)) 514 1.1 riastrad dev_err(adev->dev, "fence debugfs file creation failed\n"); 515 1.1 riastrad 516 1.1 riastrad return 0; 517 1.1 riastrad } 518 1.1 riastrad 519 1.1 riastrad /** 520 1.1 riastrad * amdgpu_fence_driver_fini - tear down the fence driver 521 1.1 riastrad * for all possible rings. 522 1.1 riastrad * 523 1.1 riastrad * @adev: amdgpu device pointer 524 1.1 riastrad * 525 1.1 riastrad * Tear down the fence driver for all possible rings (all asics). 526 1.1 riastrad */ 527 1.1 riastrad void amdgpu_fence_driver_fini(struct amdgpu_device *adev) 528 1.1 riastrad { 529 1.7 riastrad unsigned i, j; 530 1.7 riastrad int r; 531 1.1 riastrad 532 1.1 riastrad for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 533 1.1 riastrad struct amdgpu_ring *ring = adev->rings[i]; 534 1.1 riastrad 535 1.1 riastrad if (!ring || !ring->fence_drv.initialized) 536 1.1 riastrad continue; 537 1.1 riastrad r = amdgpu_fence_wait_empty(ring); 538 1.1 riastrad if (r) { 539 1.1 riastrad /* no need to trigger GPU reset as we are unloading */ 540 1.7 riastrad amdgpu_fence_driver_force_completion(ring); 541 1.1 riastrad } 542 1.1 riastrad amdgpu_irq_put(adev, ring->fence_drv.irq_src, 543 1.1 riastrad ring->fence_drv.irq_type); 544 1.11 mrg if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) { 545 1.11 mrg drm_sched_fini(&ring->sched); 546 1.11 mrg } 547 1.1 riastrad del_timer_sync(&ring->fence_drv.fallback_timer); 548 1.7 riastrad for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j) 549 1.7 riastrad dma_fence_put(ring->fence_drv.fences[j]); 550 1.7 riastrad kfree(ring->fence_drv.fences); 551 1.7 riastrad ring->fence_drv.fences = NULL; 552 1.1 riastrad ring->fence_drv.initialized = false; 553 1.8 riastrad spin_lock_destroy(&ring->fence_drv.lock); 554 1.1 riastrad } 555 1.1 riastrad } 556 1.1 riastrad 557 1.1 riastrad /** 558 1.1 riastrad * amdgpu_fence_driver_suspend - suspend the fence driver 559 1.1 riastrad * for all possible rings. 560 1.1 riastrad * 561 1.1 riastrad * @adev: amdgpu device pointer 562 1.1 riastrad * 563 1.1 riastrad * Suspend the fence driver for all possible rings (all asics). 564 1.1 riastrad */ 565 1.1 riastrad void amdgpu_fence_driver_suspend(struct amdgpu_device *adev) 566 1.1 riastrad { 567 1.1 riastrad int i, r; 568 1.1 riastrad 569 1.1 riastrad for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 570 1.1 riastrad struct amdgpu_ring *ring = adev->rings[i]; 571 1.1 riastrad if (!ring || !ring->fence_drv.initialized) 572 1.1 riastrad continue; 573 1.1 riastrad 574 1.1 riastrad /* wait for gpu to finish processing current batch */ 575 1.1 riastrad r = amdgpu_fence_wait_empty(ring); 576 1.1 riastrad if (r) { 577 1.1 riastrad /* delay GPU reset to resume */ 578 1.7 riastrad amdgpu_fence_driver_force_completion(ring); 579 1.1 riastrad } 580 1.1 riastrad 581 1.1 riastrad /* disable the interrupt */ 582 1.1 riastrad amdgpu_irq_put(adev, ring->fence_drv.irq_src, 583 1.1 riastrad ring->fence_drv.irq_type); 584 1.1 riastrad } 585 1.1 riastrad } 586 1.1 riastrad 587 1.1 riastrad /** 588 1.1 riastrad * amdgpu_fence_driver_resume - resume the fence driver 589 1.1 riastrad * for all possible rings. 590 1.1 riastrad * 591 1.1 riastrad * @adev: amdgpu device pointer 592 1.1 riastrad * 593 1.1 riastrad * Resume the fence driver for all possible rings (all asics). 594 1.1 riastrad * Not all asics have all rings, so each asic will only 595 1.1 riastrad * start the fence driver on the rings it has using 596 1.1 riastrad * amdgpu_fence_driver_start_ring(). 597 1.1 riastrad * Returns 0 for success. 598 1.1 riastrad */ 599 1.1 riastrad void amdgpu_fence_driver_resume(struct amdgpu_device *adev) 600 1.1 riastrad { 601 1.1 riastrad int i; 602 1.1 riastrad 603 1.1 riastrad for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 604 1.1 riastrad struct amdgpu_ring *ring = adev->rings[i]; 605 1.1 riastrad if (!ring || !ring->fence_drv.initialized) 606 1.1 riastrad continue; 607 1.1 riastrad 608 1.1 riastrad /* enable the interrupt */ 609 1.1 riastrad amdgpu_irq_get(adev, ring->fence_drv.irq_src, 610 1.1 riastrad ring->fence_drv.irq_type); 611 1.1 riastrad } 612 1.1 riastrad } 613 1.1 riastrad 614 1.1 riastrad /** 615 1.7 riastrad * amdgpu_fence_driver_force_completion - force signal latest fence of ring 616 1.1 riastrad * 617 1.7 riastrad * @ring: fence of the ring to signal 618 1.1 riastrad * 619 1.1 riastrad */ 620 1.7 riastrad void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring) 621 1.1 riastrad { 622 1.7 riastrad amdgpu_fence_write(ring, ring->fence_drv.sync_seq); 623 1.7 riastrad amdgpu_fence_process(ring); 624 1.1 riastrad } 625 1.1 riastrad 626 1.1 riastrad /* 627 1.1 riastrad * Common fence implementation 628 1.1 riastrad */ 629 1.1 riastrad 630 1.7 riastrad static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence) 631 1.1 riastrad { 632 1.1 riastrad return "amdgpu"; 633 1.1 riastrad } 634 1.1 riastrad 635 1.7 riastrad static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f) 636 1.1 riastrad { 637 1.1 riastrad struct amdgpu_fence *fence = to_amdgpu_fence(f); 638 1.1 riastrad return (const char *)fence->ring->name; 639 1.1 riastrad } 640 1.1 riastrad 641 1.1 riastrad /** 642 1.7 riastrad * amdgpu_fence_enable_signaling - enable signalling on fence 643 1.7 riastrad * @fence: fence 644 1.1 riastrad * 645 1.7 riastrad * This function is called with fence_queue lock held, and adds a callback 646 1.7 riastrad * to fence_queue that checks if this fence is signaled, and if so it 647 1.7 riastrad * signals the fence and removes itself. 648 1.1 riastrad */ 649 1.7 riastrad static bool amdgpu_fence_enable_signaling(struct dma_fence *f) 650 1.1 riastrad { 651 1.1 riastrad struct amdgpu_fence *fence = to_amdgpu_fence(f); 652 1.1 riastrad struct amdgpu_ring *ring = fence->ring; 653 1.1 riastrad 654 1.7 riastrad if (!timer_pending(&ring->fence_drv.fallback_timer)) 655 1.7 riastrad amdgpu_fence_schedule_fallback(ring); 656 1.1 riastrad 657 1.7 riastrad DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx); 658 1.1 riastrad 659 1.7 riastrad return true; 660 1.1 riastrad } 661 1.1 riastrad 662 1.1 riastrad /** 663 1.7 riastrad * amdgpu_fence_free - free up the fence memory 664 1.1 riastrad * 665 1.7 riastrad * @rcu: RCU callback head 666 1.7 riastrad * 667 1.7 riastrad * Free up the fence memory after the RCU grace period. 668 1.1 riastrad */ 669 1.7 riastrad static void amdgpu_fence_free(struct rcu_head *rcu) 670 1.1 riastrad { 671 1.7 riastrad struct dma_fence *f = container_of(rcu, struct dma_fence, rcu); 672 1.7 riastrad struct amdgpu_fence *fence = to_amdgpu_fence(f); 673 1.10 riastrad dma_fence_destroy(f); 674 1.7 riastrad kmem_cache_free(amdgpu_fence_slab, fence); 675 1.1 riastrad } 676 1.1 riastrad 677 1.1 riastrad /** 678 1.7 riastrad * amdgpu_fence_release - callback that fence can be freed 679 1.7 riastrad * 680 1.1 riastrad * @fence: fence 681 1.1 riastrad * 682 1.7 riastrad * This function is called when the reference count becomes zero. 683 1.7 riastrad * It just RCU schedules freeing up the fence. 684 1.1 riastrad */ 685 1.7 riastrad static void amdgpu_fence_release(struct dma_fence *f) 686 1.1 riastrad { 687 1.7 riastrad call_rcu(&f->rcu, amdgpu_fence_free); 688 1.1 riastrad } 689 1.1 riastrad 690 1.7 riastrad static const struct dma_fence_ops amdgpu_fence_ops = { 691 1.1 riastrad .get_driver_name = amdgpu_fence_get_driver_name, 692 1.1 riastrad .get_timeline_name = amdgpu_fence_get_timeline_name, 693 1.1 riastrad .enable_signaling = amdgpu_fence_enable_signaling, 694 1.1 riastrad .release = amdgpu_fence_release, 695 1.1 riastrad }; 696 1.1 riastrad 697 1.1 riastrad /* 698 1.1 riastrad * Fence debugfs 699 1.1 riastrad */ 700 1.1 riastrad #if defined(CONFIG_DEBUG_FS) 701 1.1 riastrad static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) 702 1.1 riastrad { 703 1.1 riastrad struct drm_info_node *node = (struct drm_info_node *)m->private; 704 1.1 riastrad struct drm_device *dev = node->minor->dev; 705 1.1 riastrad struct amdgpu_device *adev = dev->dev_private; 706 1.7 riastrad int i; 707 1.1 riastrad 708 1.1 riastrad for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 709 1.1 riastrad struct amdgpu_ring *ring = adev->rings[i]; 710 1.1 riastrad if (!ring || !ring->fence_drv.initialized) 711 1.1 riastrad continue; 712 1.1 riastrad 713 1.1 riastrad amdgpu_fence_process(ring); 714 1.1 riastrad 715 1.1 riastrad seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name); 716 1.7 riastrad seq_printf(m, "Last signaled fence 0x%08x\n", 717 1.7 riastrad atomic_read(&ring->fence_drv.last_seq)); 718 1.7 riastrad seq_printf(m, "Last emitted 0x%08x\n", 719 1.7 riastrad ring->fence_drv.sync_seq); 720 1.7 riastrad 721 1.7 riastrad if (ring->funcs->type == AMDGPU_RING_TYPE_GFX || 722 1.7 riastrad ring->funcs->type == AMDGPU_RING_TYPE_SDMA) { 723 1.7 riastrad seq_printf(m, "Last signaled trailing fence 0x%08x\n", 724 1.7 riastrad le32_to_cpu(*ring->trail_fence_cpu_addr)); 725 1.7 riastrad seq_printf(m, "Last emitted 0x%08x\n", 726 1.7 riastrad ring->trail_seq); 727 1.1 riastrad } 728 1.7 riastrad 729 1.7 riastrad if (ring->funcs->type != AMDGPU_RING_TYPE_GFX) 730 1.7 riastrad continue; 731 1.7 riastrad 732 1.7 riastrad /* set in CP_VMID_PREEMPT and preemption occurred */ 733 1.7 riastrad seq_printf(m, "Last preempted 0x%08x\n", 734 1.7 riastrad le32_to_cpu(*(ring->fence_drv.cpu_addr + 2))); 735 1.7 riastrad /* set in CP_VMID_RESET and reset occurred */ 736 1.7 riastrad seq_printf(m, "Last reset 0x%08x\n", 737 1.7 riastrad le32_to_cpu(*(ring->fence_drv.cpu_addr + 4))); 738 1.7 riastrad /* Both preemption and reset occurred */ 739 1.7 riastrad seq_printf(m, "Last both 0x%08x\n", 740 1.7 riastrad le32_to_cpu(*(ring->fence_drv.cpu_addr + 6))); 741 1.1 riastrad } 742 1.1 riastrad return 0; 743 1.1 riastrad } 744 1.1 riastrad 745 1.7 riastrad /** 746 1.7 riastrad * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover 747 1.7 riastrad * 748 1.7 riastrad * Manually trigger a gpu reset at the next fence wait. 749 1.7 riastrad */ 750 1.7 riastrad static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data) 751 1.7 riastrad { 752 1.7 riastrad struct drm_info_node *node = (struct drm_info_node *) m->private; 753 1.7 riastrad struct drm_device *dev = node->minor->dev; 754 1.7 riastrad struct amdgpu_device *adev = dev->dev_private; 755 1.7 riastrad int r; 756 1.7 riastrad 757 1.7 riastrad r = pm_runtime_get_sync(dev->dev); 758 1.7 riastrad if (r < 0) 759 1.7 riastrad return 0; 760 1.7 riastrad 761 1.7 riastrad seq_printf(m, "gpu recover\n"); 762 1.7 riastrad amdgpu_device_gpu_recover(adev, NULL); 763 1.7 riastrad 764 1.7 riastrad pm_runtime_mark_last_busy(dev->dev); 765 1.7 riastrad pm_runtime_put_autosuspend(dev->dev); 766 1.7 riastrad 767 1.7 riastrad return 0; 768 1.7 riastrad } 769 1.7 riastrad 770 1.7 riastrad static const struct drm_info_list amdgpu_debugfs_fence_list[] = { 771 1.7 riastrad {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL}, 772 1.7 riastrad {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL} 773 1.7 riastrad }; 774 1.7 riastrad 775 1.7 riastrad static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = { 776 1.1 riastrad {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL}, 777 1.1 riastrad }; 778 1.1 riastrad #endif 779 1.1 riastrad 780 1.1 riastrad int amdgpu_debugfs_fence_init(struct amdgpu_device *adev) 781 1.1 riastrad { 782 1.1 riastrad #if defined(CONFIG_DEBUG_FS) 783 1.7 riastrad if (amdgpu_sriov_vf(adev)) 784 1.7 riastrad return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1); 785 1.7 riastrad return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2); 786 1.1 riastrad #else 787 1.1 riastrad return 0; 788 1.1 riastrad #endif 789 1.1 riastrad } 790 1.1 riastrad 791