1 /* 2 * Copyright 2008 Advanced Micro Devices, Inc. 3 * Copyright 2008 Red Hat Inc. 4 * Copyright 2009 Jerome Glisse. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 * 24 * Authors: Dave Airlie 25 * Alex Deucher 26 * Jerome Glisse 27 * Christian König 28 */ 29 #include <linux/seq_file.h> 30 #include <linux/slab.h> 31 #include <linux/uaccess.h> 32 #include <linux/debugfs.h> 33 34 #include <drm/amdgpu_drm.h> 35 #include "amdgpu.h" 36 #include "atom.h" 37 38 /* 39 * Rings 40 * Most engines on the GPU are fed via ring buffers. Ring 41 * buffers are areas of GPU accessible memory that the host 42 * writes commands into and the GPU reads commands out of. 43 * There is a rptr (read pointer) that determines where the 44 * GPU is currently reading, and a wptr (write pointer) 45 * which determines where the host has written. When the 46 * pointers are equal, the ring is idle. When the host 47 * writes commands to the ring buffer, it increments the 48 * wptr. The GPU then starts fetching commands and executes 49 * them until the pointers are equal again. 50 */ 51 52 /** 53 * amdgpu_ring_alloc - allocate space on the ring buffer 54 * 55 * @adev: amdgpu_device pointer 56 * @ring: amdgpu_ring structure holding ring information 57 * @ndw: number of dwords to allocate in the ring buffer 58 * 59 * Allocate @ndw dwords in the ring buffer (all asics). 60 * Returns 0 on success, error on failure. 61 */ 62 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw) 63 { 64 /* Align requested size with padding so unlock_commit can 65 * pad safely */ 66 ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask; 67 68 /* Make sure we aren't trying to allocate more space 69 * than the maximum for one submission 70 */ 71 if (WARN_ON_ONCE(ndw > ring->max_dw)) 72 return -ENOMEM; 73 74 ring->count_dw = ndw; 75 ring->wptr_old = ring->wptr; 76 77 if (ring->funcs->begin_use) 78 ring->funcs->begin_use(ring); 79 80 return 0; 81 } 82 83 /** amdgpu_ring_insert_nop - insert NOP packets 84 * 85 * @ring: amdgpu_ring structure holding ring information 86 * @count: the number of NOP packets to insert 87 * 88 * This is the generic insert_nop function for rings except SDMA 89 */ 90 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) 91 { 92 int i; 93 94 for (i = 0; i < count; i++) 95 amdgpu_ring_write(ring, ring->funcs->nop); 96 } 97 98 /** amdgpu_ring_generic_pad_ib - pad IB with NOP packets 99 * 100 * @ring: amdgpu_ring structure holding ring information 101 * @ib: IB to add NOP packets to 102 * 103 * This is the generic pad_ib function for rings except SDMA 104 */ 105 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) 106 { 107 while (ib->length_dw & ring->funcs->align_mask) 108 ib->ptr[ib->length_dw++] = ring->funcs->nop; 109 } 110 111 /** 112 * amdgpu_ring_commit - tell the GPU to execute the new 113 * commands on the ring buffer 114 * 115 * @adev: amdgpu_device pointer 116 * @ring: amdgpu_ring structure holding ring information 117 * 118 * Update the wptr (write pointer) to tell the GPU to 119 * execute new commands on the ring buffer (all asics). 120 */ 121 void amdgpu_ring_commit(struct amdgpu_ring *ring) 122 { 123 uint32_t count; 124 125 /* We pad to match fetch size */ 126 count = ring->funcs->align_mask + 1 - 127 (ring->wptr & ring->funcs->align_mask); 128 count %= ring->funcs->align_mask + 1; 129 ring->funcs->insert_nop(ring, count); 130 131 mb(); 132 amdgpu_ring_set_wptr(ring); 133 134 if (ring->funcs->end_use) 135 ring->funcs->end_use(ring); 136 } 137 138 /** 139 * amdgpu_ring_undo - reset the wptr 140 * 141 * @ring: amdgpu_ring structure holding ring information 142 * 143 * Reset the driver's copy of the wptr (all asics). 144 */ 145 void amdgpu_ring_undo(struct amdgpu_ring *ring) 146 { 147 ring->wptr = ring->wptr_old; 148 149 if (ring->funcs->end_use) 150 ring->funcs->end_use(ring); 151 } 152 153 /** 154 * amdgpu_ring_init - init driver ring struct. 155 * 156 * @adev: amdgpu_device pointer 157 * @ring: amdgpu_ring structure holding ring information 158 * @max_ndw: maximum number of dw for ring alloc 159 * @nop: nop packet for this ring 160 * 161 * Initialize the driver information for the selected ring (all asics). 162 * Returns 0 on success, error on failure. 163 */ 164 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, 165 unsigned int max_dw, struct amdgpu_irq_src *irq_src, 166 unsigned int irq_type, unsigned int hw_prio) 167 { 168 int r, i; 169 int sched_hw_submission = amdgpu_sched_hw_submission; 170 u32 *num_sched; 171 u32 hw_ip; 172 173 /* Set the hw submission limit higher for KIQ because 174 * it's used for a number of gfx/compute tasks by both 175 * KFD and KGD which may have outstanding fences and 176 * it doesn't really use the gpu scheduler anyway; 177 * KIQ tasks get submitted directly to the ring. 178 */ 179 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ) 180 sched_hw_submission = max(sched_hw_submission, 256); 181 else if (ring == &adev->sdma.instance[0].page) 182 sched_hw_submission = 256; 183 184 if (ring->adev == NULL) { 185 if (adev->num_rings >= AMDGPU_MAX_RINGS) 186 return -EINVAL; 187 188 ring->adev = adev; 189 ring->idx = adev->num_rings++; 190 adev->rings[ring->idx] = ring; 191 r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission); 192 if (r) 193 return r; 194 } 195 196 r = amdgpu_device_wb_get(adev, &ring->rptr_offs); 197 if (r) { 198 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r); 199 return r; 200 } 201 202 r = amdgpu_device_wb_get(adev, &ring->wptr_offs); 203 if (r) { 204 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r); 205 return r; 206 } 207 208 r = amdgpu_device_wb_get(adev, &ring->fence_offs); 209 if (r) { 210 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r); 211 return r; 212 } 213 214 r = amdgpu_device_wb_get(adev, &ring->trail_fence_offs); 215 if (r) { 216 dev_err(adev->dev, 217 "(%d) ring trail_fence_offs wb alloc failed\n", r); 218 return r; 219 } 220 ring->trail_fence_gpu_addr = 221 adev->wb.gpu_addr + (ring->trail_fence_offs * 4); 222 ring->trail_fence_cpu_addr = &adev->wb.wb[ring->trail_fence_offs]; 223 224 r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs); 225 if (r) { 226 dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r); 227 return r; 228 } 229 ring->cond_exe_gpu_addr = adev->wb.gpu_addr + (ring->cond_exe_offs * 4); 230 ring->cond_exe_cpu_addr = &adev->wb.wb[ring->cond_exe_offs]; 231 /* always set cond_exec_polling to CONTINUE */ 232 *ring->cond_exe_cpu_addr = 1; 233 234 r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type); 235 if (r) { 236 dev_err(adev->dev, "failed initializing fences (%d).\n", r); 237 return r; 238 } 239 240 ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission); 241 242 ring->buf_mask = (ring->ring_size / 4) - 1; 243 ring->ptr_mask = ring->funcs->support_64bit_ptrs ? 244 0xffffffffffffffff : ring->buf_mask; 245 /* Allocate ring buffer */ 246 if (ring->ring_obj == NULL) { 247 r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE, 248 AMDGPU_GEM_DOMAIN_GTT, 249 &ring->ring_obj, 250 &ring->gpu_addr, 251 (void **)&ring->ring); 252 if (r) { 253 dev_err(adev->dev, "(%d) ring create failed\n", r); 254 return r; 255 } 256 amdgpu_ring_clear_ring(ring); 257 } 258 259 ring->max_dw = max_dw; 260 ring->priority = DRM_SCHED_PRIORITY_NORMAL; 261 mutex_init(&ring->priority_mutex); 262 263 if (!ring->no_scheduler) { 264 hw_ip = ring->funcs->type; 265 num_sched = &adev->gpu_sched[hw_ip][hw_prio].num_scheds; 266 adev->gpu_sched[hw_ip][hw_prio].sched[(*num_sched)++] = 267 &ring->sched; 268 } 269 270 for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_COUNT; ++i) 271 atomic_set(&ring->num_jobs[i], 0); 272 273 return 0; 274 } 275 276 /** 277 * amdgpu_ring_fini - tear down the driver ring struct. 278 * 279 * @adev: amdgpu_device pointer 280 * @ring: amdgpu_ring structure holding ring information 281 * 282 * Tear down the driver information for the selected ring (all asics). 283 */ 284 void amdgpu_ring_fini(struct amdgpu_ring *ring) 285 { 286 287 /* Not to finish a ring which is not initialized */ 288 if (!(ring->adev) || !(ring->adev->rings[ring->idx])) 289 return; 290 291 ring->sched.ready = false; 292 293 amdgpu_device_wb_free(ring->adev, ring->rptr_offs); 294 amdgpu_device_wb_free(ring->adev, ring->wptr_offs); 295 296 amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs); 297 amdgpu_device_wb_free(ring->adev, ring->fence_offs); 298 299 amdgpu_bo_free_kernel(&ring->ring_obj, 300 &ring->gpu_addr, 301 (void **)&ring->ring); 302 303 dma_fence_put(ring->vmid_wait); 304 ring->vmid_wait = NULL; 305 ring->me = 0; 306 307 ring->adev->rings[ring->idx] = NULL; 308 } 309 310 /** 311 * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper 312 * 313 * @adev: amdgpu_device pointer 314 * @reg0: register to write 315 * @reg1: register to wait on 316 * @ref: reference value to write/wait on 317 * @mask: mask to wait on 318 * 319 * Helper for rings that don't support write and wait in a 320 * single oneshot packet. 321 */ 322 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring, 323 uint32_t reg0, uint32_t reg1, 324 uint32_t ref, uint32_t mask) 325 { 326 amdgpu_ring_emit_wreg(ring, reg0, ref); 327 amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask); 328 } 329 330 /** 331 * amdgpu_ring_soft_recovery - try to soft recover a ring lockup 332 * 333 * @ring: ring to try the recovery on 334 * @vmid: VMID we try to get going again 335 * @fence: timedout fence 336 * 337 * Tries to get a ring proceeding again when it is stuck. 338 */ 339 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid, 340 struct dma_fence *fence) 341 { 342 ktime_t deadline = ktime_add_us(ktime_get(), 10000); 343 344 if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence) 345 return false; 346 347 atomic_inc(&ring->adev->gpu_reset_counter); 348 while (!dma_fence_is_signaled(fence) && 349 ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0) 350 ring->funcs->soft_recovery(ring, vmid); 351 352 return dma_fence_is_signaled(fence); 353 } 354 355 /* 356 * Debugfs info 357 */ 358 #if defined(CONFIG_DEBUG_FS) 359 360 /* Layout of file is 12 bytes consisting of 361 * - rptr 362 * - wptr 363 * - driver's copy of wptr 364 * 365 * followed by n-words of ring data 366 */ 367 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf, 368 size_t size, loff_t *pos) 369 { 370 struct amdgpu_ring *ring = file_inode(f)->i_private; 371 int r, i; 372 uint32_t value, result, early[3]; 373 374 if (*pos & 3 || size & 3) 375 return -EINVAL; 376 377 result = 0; 378 379 if (*pos < 12) { 380 early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask; 381 early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask; 382 early[2] = ring->wptr & ring->buf_mask; 383 for (i = *pos / 4; i < 3 && size; i++) { 384 r = put_user(early[i], (uint32_t *)buf); 385 if (r) 386 return r; 387 buf += 4; 388 result += 4; 389 size -= 4; 390 *pos += 4; 391 } 392 } 393 394 while (size) { 395 if (*pos >= (ring->ring_size + 12)) 396 return result; 397 398 value = ring->ring[(*pos - 12)/4]; 399 r = put_user(value, (uint32_t*)buf); 400 if (r) 401 return r; 402 buf += 4; 403 result += 4; 404 size -= 4; 405 *pos += 4; 406 } 407 408 return result; 409 } 410 411 static const struct file_operations amdgpu_debugfs_ring_fops = { 412 .owner = THIS_MODULE, 413 .read = amdgpu_debugfs_ring_read, 414 .llseek = default_llseek 415 }; 416 417 #endif 418 419 int amdgpu_debugfs_ring_init(struct amdgpu_device *adev, 420 struct amdgpu_ring *ring) 421 { 422 #if defined(CONFIG_DEBUG_FS) 423 struct drm_minor *minor = adev_to_drm(adev)->primary; 424 struct dentry *ent, *root = minor->debugfs_root; 425 char name[32]; 426 427 sprintf(name, "amdgpu_ring_%s", ring->name); 428 429 ent = debugfs_create_file(name, 430 S_IFREG | S_IRUGO, root, 431 ring, &amdgpu_debugfs_ring_fops); 432 if (!ent) 433 return -ENOMEM; 434 435 i_size_write(ent->d_inode, ring->ring_size + 12); 436 ring->ent = ent; 437 #endif 438 return 0; 439 } 440 441 /** 442 * amdgpu_ring_test_helper - tests ring and set sched readiness status 443 * 444 * @ring: ring to try the recovery on 445 * 446 * Tests ring and set sched readiness status 447 * 448 * Returns 0 on success, error on failure. 449 */ 450 int amdgpu_ring_test_helper(struct amdgpu_ring *ring) 451 { 452 struct amdgpu_device *adev = ring->adev; 453 int r; 454 455 r = amdgpu_ring_test_ring(ring); 456 if (r) 457 DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n", 458 ring->name, r); 459 else 460 DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n", 461 ring->name); 462 463 ring->sched.ready = !r; 464 return r; 465 } 466