| 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 <stdlib.h> |
| 25 |
#include <stdio.h> |
| 26 |
#include <string.h> |
| 27 |
#include <errno.h> |
| 28 |
#include <pthread.h> |
| 29 |
#include <sched.h> |
| 30 |
#include <sys/ioctl.h> |
| 31 |
#if HAVE_ALLOCA_H |
| 32 |
# include <alloca.h> |
| 33 |
#endif |
| 34 |
|
| 35 |
#include "xf86drm.h" |
| 36 |
#include "amdgpu_drm.h" |
| 37 |
#include "amdgpu_internal.h" |
| 38 |
|
| 39 |
static int amdgpu_cs_unreference_sem(amdgpu_semaphore_handle sem); |
| 40 |
static int amdgpu_cs_reset_sem(amdgpu_semaphore_handle sem); |
| 41 |
|
| 42 |
/** |
| 43 |
* Create command submission context |
| 44 |
* |
| 45 |
* \param dev - \c [in] Device handle. See #amdgpu_device_initialize() |
| 46 |
* \param priority - \c [in] Context creation flags. See AMDGPU_CTX_PRIORITY_* |
| 47 |
* \param context - \c [out] GPU Context handle |
| 48 |
* |
| 49 |
* \return 0 on success otherwise POSIX Error code |
| 50 |
*/ |
| 51 |
drm_public int amdgpu_cs_ctx_create2(amdgpu_device_handle dev, |
| 52 |
uint32_t priority, |
| 53 |
amdgpu_context_handle *context) |
| 54 |
{ |
| 55 |
struct amdgpu_context *gpu_context; |
| 56 |
union drm_amdgpu_ctx args; |
| 57 |
int i, j, k; |
| 58 |
int r; |
| 59 |
char *override_priority; |
| 60 |
|
| 61 |
if (!dev || !context) |
| 62 |
return -EINVAL; |
| 63 |
|
| 64 |
override_priority = getenv("AMD_PRIORITY"); |
| 65 |
if (override_priority) { |
| 66 |
/* The priority is a signed integer. The variable type is |
| 67 |
* wrong. If parsing fails, priority is unchanged. |
| 68 |
*/ |
| 69 |
if (sscanf(override_priority, "%i", &priority) == 1) { |
| 70 |
printf("amdgpu: context priority changed to %i\n", |
| 71 |
priority); |
| 72 |
} |
| 73 |
} |
| 74 |
|
| 75 |
gpu_context = calloc(1, sizeof(struct amdgpu_context)); |
| 76 |
if (!gpu_context) |
| 77 |
return -ENOMEM; |
| 78 |
|
| 79 |
gpu_context->dev = dev; |
| 80 |
|
| 81 |
r = pthread_mutex_init(&gpu_context->sequence_mutex, NULL); |
| 82 |
if (r) |
| 83 |
goto error; |
| 84 |
|
| 85 |
/* Create the context */ |
| 86 |
memset(&args, 0, sizeof(args)); |
| 87 |
args.in.op = AMDGPU_CTX_OP_ALLOC_CTX; |
| 88 |
args.in.priority = priority; |
| 89 |
|
| 90 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_CTX, &args, sizeof(args)); |
| 91 |
if (r) |
| 92 |
goto error; |
| 93 |
|
| 94 |
gpu_context->id = args.out.alloc.ctx_id; |
| 95 |
for (i = 0; i < AMDGPU_HW_IP_NUM; i++) |
| 96 |
for (j = 0; j < AMDGPU_HW_IP_INSTANCE_MAX_COUNT; j++) |
| 97 |
for (k = 0; k < AMDGPU_CS_MAX_RINGS; k++) |
| 98 |
list_inithead(&gpu_context->sem_list[i][j][k]); |
| 99 |
*context = (amdgpu_context_handle)gpu_context; |
| 100 |
|
| 101 |
return 0; |
| 102 |
|
| 103 |
error: |
| 104 |
pthread_mutex_destroy(&gpu_context->sequence_mutex); |
| 105 |
free(gpu_context); |
| 106 |
return r; |
| 107 |
} |
| 108 |
|
| 109 |
drm_public int amdgpu_cs_ctx_create(amdgpu_device_handle dev, |
| 110 |
amdgpu_context_handle *context) |
| 111 |
{ |
| 112 |
return amdgpu_cs_ctx_create2(dev, AMDGPU_CTX_PRIORITY_NORMAL, context); |
| 113 |
} |
| 114 |
|
| 115 |
/** |
| 116 |
* Release command submission context |
| 117 |
* |
| 118 |
* \param dev - \c [in] amdgpu device handle |
| 119 |
* \param context - \c [in] amdgpu context handle |
| 120 |
* |
| 121 |
* \return 0 on success otherwise POSIX Error code |
| 122 |
*/ |
| 123 |
drm_public int amdgpu_cs_ctx_free(amdgpu_context_handle context) |
| 124 |
{ |
| 125 |
union drm_amdgpu_ctx args; |
| 126 |
int i, j, k; |
| 127 |
int r; |
| 128 |
|
| 129 |
if (!context) |
| 130 |
return -EINVAL; |
| 131 |
|
| 132 |
pthread_mutex_destroy(&context->sequence_mutex); |
| 133 |
|
| 134 |
/* now deal with kernel side */ |
| 135 |
memset(&args, 0, sizeof(args)); |
| 136 |
args.in.op = AMDGPU_CTX_OP_FREE_CTX; |
| 137 |
args.in.ctx_id = context->id; |
| 138 |
r = drmCommandWriteRead(context->dev->fd, DRM_AMDGPU_CTX, |
| 139 |
&args, sizeof(args)); |
| 140 |
for (i = 0; i < AMDGPU_HW_IP_NUM; i++) { |
| 141 |
for (j = 0; j < AMDGPU_HW_IP_INSTANCE_MAX_COUNT; j++) { |
| 142 |
for (k = 0; k < AMDGPU_CS_MAX_RINGS; k++) { |
| 143 |
amdgpu_semaphore_handle sem, tmp; |
| 144 |
LIST_FOR_EACH_ENTRY_SAFE(sem, tmp, &context->sem_list[i][j][k], list) { |
| 145 |
list_del(&sem->list); |
| 146 |
amdgpu_cs_reset_sem(sem); |
| 147 |
amdgpu_cs_unreference_sem(sem); |
| 148 |
} |
| 149 |
} |
| 150 |
} |
| 151 |
} |
| 152 |
free(context); |
| 153 |
|
| 154 |
return r; |
| 155 |
} |
| 156 |
|
| 157 |
drm_public int amdgpu_cs_ctx_override_priority(amdgpu_device_handle dev, |
| 158 |
amdgpu_context_handle context, |
| 159 |
int master_fd, |
| 160 |
unsigned priority) |
| 161 |
{ |
| 162 |
union drm_amdgpu_sched args; |
| 163 |
int r; |
| 164 |
|
| 165 |
if (!dev || !context || master_fd < 0) |
| 166 |
return -EINVAL; |
| 167 |
|
| 168 |
memset(&args, 0, sizeof(args)); |
| 169 |
|
| 170 |
args.in.op = AMDGPU_SCHED_OP_CONTEXT_PRIORITY_OVERRIDE; |
| 171 |
args.in.fd = dev->fd; |
| 172 |
args.in.priority = priority; |
| 173 |
args.in.ctx_id = context->id; |
| 174 |
|
| 175 |
r = drmCommandWrite(master_fd, DRM_AMDGPU_SCHED, &args, sizeof(args)); |
| 176 |
if (r) |
| 177 |
return r; |
| 178 |
|
| 179 |
return 0; |
| 180 |
} |
| 181 |
|
| 182 |
drm_public int amdgpu_cs_ctx_stable_pstate(amdgpu_context_handle context, |
| 183 |
uint32_t op, |
| 184 |
uint32_t flags, |
| 185 |
uint32_t *out_flags) |
| 186 |
{ |
| 187 |
union drm_amdgpu_ctx args; |
| 188 |
int r; |
| 189 |
|
| 190 |
if (!context) |
| 191 |
return -EINVAL; |
| 192 |
|
| 193 |
memset(&args, 0, sizeof(args)); |
| 194 |
args.in.op = op; |
| 195 |
args.in.ctx_id = context->id; |
| 196 |
args.in.flags = flags; |
| 197 |
r = drmCommandWriteRead(context->dev->fd, DRM_AMDGPU_CTX, |
| 198 |
&args, sizeof(args)); |
| 199 |
if (!r && out_flags) |
| 200 |
*out_flags = args.out.pstate.flags; |
| 201 |
return r; |
| 202 |
} |
| 203 |
|
| 204 |
drm_public int amdgpu_cs_query_reset_state(amdgpu_context_handle context, |
| 205 |
uint32_t *state, uint32_t *hangs) |
| 206 |
{ |
| 207 |
union drm_amdgpu_ctx args; |
| 208 |
int r; |
| 209 |
|
| 210 |
if (!context) |
| 211 |
return -EINVAL; |
| 212 |
|
| 213 |
memset(&args, 0, sizeof(args)); |
| 214 |
args.in.op = AMDGPU_CTX_OP_QUERY_STATE; |
| 215 |
args.in.ctx_id = context->id; |
| 216 |
r = drmCommandWriteRead(context->dev->fd, DRM_AMDGPU_CTX, |
| 217 |
&args, sizeof(args)); |
| 218 |
if (!r) { |
| 219 |
*state = args.out.state.reset_status; |
| 220 |
*hangs = args.out.state.hangs; |
| 221 |
} |
| 222 |
return r; |
| 223 |
} |
| 224 |
|
| 225 |
drm_public int amdgpu_cs_query_reset_state2(amdgpu_context_handle context, |
| 226 |
uint64_t *flags) |
| 227 |
{ |
| 228 |
union drm_amdgpu_ctx args; |
| 229 |
int r; |
| 230 |
|
| 231 |
if (!context) |
| 232 |
return -EINVAL; |
| 233 |
|
| 234 |
memset(&args, 0, sizeof(args)); |
| 235 |
args.in.op = AMDGPU_CTX_OP_QUERY_STATE2; |
| 236 |
args.in.ctx_id = context->id; |
| 237 |
r = drmCommandWriteRead(context->dev->fd, DRM_AMDGPU_CTX, |
| 238 |
&args, sizeof(args)); |
| 239 |
if (!r) |
| 240 |
*flags = args.out.state.flags; |
| 241 |
return r; |
| 242 |
} |
| 243 |
|
| 244 |
/** |
| 245 |
* Submit command to kernel DRM |
| 246 |
* \param dev - \c [in] Device handle |
| 247 |
* \param context - \c [in] GPU Context |
| 248 |
* \param ibs_request - \c [in] Pointer to submission requests |
| 249 |
* \param fence - \c [out] return fence for this submission |
| 250 |
* |
| 251 |
* \return 0 on success otherwise POSIX Error code |
| 252 |
* \sa amdgpu_cs_submit() |
| 253 |
*/ |
| 254 |
static int amdgpu_cs_submit_one(amdgpu_context_handle context, |
| 255 |
struct amdgpu_cs_request *ibs_request) |
| 256 |
{ |
| 257 |
struct drm_amdgpu_cs_chunk *chunks; |
| 258 |
struct drm_amdgpu_cs_chunk_data *chunk_data; |
| 259 |
struct drm_amdgpu_cs_chunk_dep *dependencies = NULL; |
| 260 |
struct drm_amdgpu_cs_chunk_dep *sem_dependencies = NULL; |
| 261 |
amdgpu_device_handle dev = context->dev; |
| 262 |
struct list_head *sem_list; |
| 263 |
amdgpu_semaphore_handle sem, tmp; |
| 264 |
uint32_t i, size, num_chunks, bo_list_handle = 0, sem_count = 0; |
| 265 |
uint64_t seq_no; |
| 266 |
bool user_fence; |
| 267 |
int r = 0; |
| 268 |
|
| 269 |
if (ibs_request->ip_type >= AMDGPU_HW_IP_NUM) |
| 270 |
return -EINVAL; |
| 271 |
if (ibs_request->ring >= AMDGPU_CS_MAX_RINGS) |
| 272 |
return -EINVAL; |
| 273 |
if (ibs_request->number_of_ibs == 0) { |
| 274 |
ibs_request->seq_no = AMDGPU_NULL_SUBMIT_SEQ; |
| 275 |
return 0; |
| 276 |
} |
| 277 |
user_fence = (ibs_request->fence_info.handle != NULL); |
| 278 |
|
| 279 |
size = ibs_request->number_of_ibs + (user_fence ? 2 : 1) + 1; |
| 280 |
|
| 281 |
chunks = alloca(sizeof(struct drm_amdgpu_cs_chunk) * size); |
| 282 |
|
| 283 |
size = ibs_request->number_of_ibs + (user_fence ? 1 : 0); |
| 284 |
|
| 285 |
chunk_data = alloca(sizeof(struct drm_amdgpu_cs_chunk_data) * size); |
| 286 |
|
| 287 |
if (ibs_request->resources) |
| 288 |
bo_list_handle = ibs_request->resources->handle; |
| 289 |
num_chunks = ibs_request->number_of_ibs; |
| 290 |
/* IB chunks */ |
| 291 |
for (i = 0; i < ibs_request->number_of_ibs; i++) { |
| 292 |
struct amdgpu_cs_ib_info *ib; |
| 293 |
chunks[i].chunk_id = AMDGPU_CHUNK_ID_IB; |
| 294 |
chunks[i].length_dw = sizeof(struct drm_amdgpu_cs_chunk_ib) / 4; |
| 295 |
chunks[i].chunk_data = (uint64_t)(uintptr_t)&chunk_data[i]; |
| 296 |
|
| 297 |
ib = &ibs_request->ibs[i]; |
| 298 |
|
| 299 |
chunk_data[i].ib_data._pad = 0; |
| 300 |
chunk_data[i].ib_data.va_start = ib->ib_mc_address; |
| 301 |
chunk_data[i].ib_data.ib_bytes = ib->size * 4; |
| 302 |
chunk_data[i].ib_data.ip_type = ibs_request->ip_type; |
| 303 |
chunk_data[i].ib_data.ip_instance = ibs_request->ip_instance; |
| 304 |
chunk_data[i].ib_data.ring = ibs_request->ring; |
| 305 |
chunk_data[i].ib_data.flags = ib->flags; |
| 306 |
} |
| 307 |
|
| 308 |
pthread_mutex_lock(&context->sequence_mutex); |
| 309 |
|
| 310 |
if (user_fence) { |
| 311 |
i = num_chunks++; |
| 312 |
|
| 313 |
/* fence chunk */ |
| 314 |
chunks[i].chunk_id = AMDGPU_CHUNK_ID_FENCE; |
| 315 |
chunks[i].length_dw = sizeof(struct drm_amdgpu_cs_chunk_fence) / 4; |
| 316 |
chunks[i].chunk_data = (uint64_t)(uintptr_t)&chunk_data[i]; |
| 317 |
|
| 318 |
/* fence bo handle */ |
| 319 |
chunk_data[i].fence_data.handle = ibs_request->fence_info.handle->handle; |
| 320 |
/* offset */ |
| 321 |
chunk_data[i].fence_data.offset = |
| 322 |
ibs_request->fence_info.offset * sizeof(uint64_t); |
| 323 |
} |
| 324 |
|
| 325 |
if (ibs_request->number_of_dependencies) { |
| 326 |
dependencies = alloca(sizeof(struct drm_amdgpu_cs_chunk_dep) * |
| 327 |
ibs_request->number_of_dependencies); |
| 328 |
if (!dependencies) { |
| 329 |
r = -ENOMEM; |
| 330 |
goto error_unlock; |
| 331 |
} |
| 332 |
|
| 333 |
for (i = 0; i < ibs_request->number_of_dependencies; ++i) { |
| 334 |
struct amdgpu_cs_fence *info = &ibs_request->dependencies[i]; |
| 335 |
struct drm_amdgpu_cs_chunk_dep *dep = &dependencies[i]; |
| 336 |
dep->ip_type = info->ip_type; |
| 337 |
dep->ip_instance = info->ip_instance; |
| 338 |
dep->ring = info->ring; |
| 339 |
dep->ctx_id = info->context->id; |
| 340 |
dep->handle = info->fence; |
| 341 |
} |
| 342 |
|
| 343 |
i = num_chunks++; |
| 344 |
|
| 345 |
/* dependencies chunk */ |
| 346 |
chunks[i].chunk_id = AMDGPU_CHUNK_ID_DEPENDENCIES; |
| 347 |
chunks[i].length_dw = sizeof(struct drm_amdgpu_cs_chunk_dep) / 4 |
| 348 |
* ibs_request->number_of_dependencies; |
| 349 |
chunks[i].chunk_data = (uint64_t)(uintptr_t)dependencies; |
| 350 |
} |
| 351 |
|
| 352 |
sem_list = &context->sem_list[ibs_request->ip_type][ibs_request->ip_instance][ibs_request->ring]; |
| 353 |
LIST_FOR_EACH_ENTRY(sem, sem_list, list) |
| 354 |
sem_count++; |
| 355 |
if (sem_count) { |
| 356 |
sem_dependencies = alloca(sizeof(struct drm_amdgpu_cs_chunk_dep) * sem_count); |
| 357 |
if (!sem_dependencies) { |
| 358 |
r = -ENOMEM; |
| 359 |
goto error_unlock; |
| 360 |
} |
| 361 |
sem_count = 0; |
| 362 |
LIST_FOR_EACH_ENTRY_SAFE(sem, tmp, sem_list, list) { |
| 363 |
struct amdgpu_cs_fence *info = &sem->signal_fence; |
| 364 |
struct drm_amdgpu_cs_chunk_dep *dep = &sem_dependencies[sem_count++]; |
| 365 |
dep->ip_type = info->ip_type; |
| 366 |
dep->ip_instance = info->ip_instance; |
| 367 |
dep->ring = info->ring; |
| 368 |
dep->ctx_id = info->context->id; |
| 369 |
dep->handle = info->fence; |
| 370 |
|
| 371 |
list_del(&sem->list); |
| 372 |
amdgpu_cs_reset_sem(sem); |
| 373 |
amdgpu_cs_unreference_sem(sem); |
| 374 |
} |
| 375 |
i = num_chunks++; |
| 376 |
|
| 377 |
/* dependencies chunk */ |
| 378 |
chunks[i].chunk_id = AMDGPU_CHUNK_ID_DEPENDENCIES; |
| 379 |
chunks[i].length_dw = sizeof(struct drm_amdgpu_cs_chunk_dep) / 4 * sem_count; |
| 380 |
chunks[i].chunk_data = (uint64_t)(uintptr_t)sem_dependencies; |
| 381 |
} |
| 382 |
|
| 383 |
r = amdgpu_cs_submit_raw2(dev, context, bo_list_handle, num_chunks, |
| 384 |
chunks, &seq_no); |
| 385 |
if (r) |
| 386 |
goto error_unlock; |
| 387 |
|
| 388 |
ibs_request->seq_no = seq_no; |
| 389 |
context->last_seq[ibs_request->ip_type][ibs_request->ip_instance][ibs_request->ring] = ibs_request->seq_no; |
| 390 |
error_unlock: |
| 391 |
pthread_mutex_unlock(&context->sequence_mutex); |
| 392 |
return r; |
| 393 |
} |
| 394 |
|
| 395 |
drm_public int amdgpu_cs_submit(amdgpu_context_handle context, |
| 396 |
uint64_t flags, |
| 397 |
struct amdgpu_cs_request *ibs_request, |
| 398 |
uint32_t number_of_requests) |
| 399 |
{ |
| 400 |
uint32_t i; |
| 401 |
int r; |
| 402 |
|
| 403 |
if (!context || !ibs_request) |
| 404 |
return -EINVAL; |
| 405 |
|
| 406 |
r = 0; |
| 407 |
for (i = 0; i < number_of_requests; i++) { |
| 408 |
r = amdgpu_cs_submit_one(context, ibs_request); |
| 409 |
if (r) |
| 410 |
break; |
| 411 |
ibs_request++; |
| 412 |
} |
| 413 |
|
| 414 |
return r; |
| 415 |
} |
| 416 |
|
| 417 |
/** |
| 418 |
* Calculate absolute timeout. |
| 419 |
* |
| 420 |
* \param timeout - \c [in] timeout in nanoseconds. |
| 421 |
* |
| 422 |
* \return absolute timeout in nanoseconds |
| 423 |
*/ |
| 424 |
drm_private uint64_t amdgpu_cs_calculate_timeout(uint64_t timeout) |
| 425 |
{ |
| 426 |
int r; |
| 427 |
|
| 428 |
if (timeout != AMDGPU_TIMEOUT_INFINITE) { |
| 429 |
struct timespec current; |
| 430 |
uint64_t current_ns; |
| 431 |
r = clock_gettime(CLOCK_MONOTONIC, ¤t); |
| 432 |
if (r) { |
| 433 |
fprintf(stderr, "clock_gettime() returned error (%d)!", errno); |
| 434 |
return AMDGPU_TIMEOUT_INFINITE; |
| 435 |
} |
| 436 |
|
| 437 |
current_ns = ((uint64_t)current.tv_sec) * 1000000000ull; |
| 438 |
current_ns += current.tv_nsec; |
| 439 |
timeout += current_ns; |
| 440 |
if (timeout < current_ns) |
| 441 |
timeout = AMDGPU_TIMEOUT_INFINITE; |
| 442 |
} |
| 443 |
return timeout; |
| 444 |
} |
| 445 |
|
| 446 |
static int amdgpu_ioctl_wait_cs(amdgpu_context_handle context, |
| 447 |
unsigned ip, |
| 448 |
unsigned ip_instance, |
| 449 |
uint32_t ring, |
| 450 |
uint64_t handle, |
| 451 |
uint64_t timeout_ns, |
| 452 |
uint64_t flags, |
| 453 |
bool *busy) |
| 454 |
{ |
| 455 |
amdgpu_device_handle dev = context->dev; |
| 456 |
union drm_amdgpu_wait_cs args; |
| 457 |
int r; |
| 458 |
|
| 459 |
memset(&args, 0, sizeof(args)); |
| 460 |
args.in.handle = handle; |
| 461 |
args.in.ip_type = ip; |
| 462 |
args.in.ip_instance = ip_instance; |
| 463 |
args.in.ring = ring; |
| 464 |
args.in.ctx_id = context->id; |
| 465 |
|
| 466 |
if (flags & AMDGPU_QUERY_FENCE_TIMEOUT_IS_ABSOLUTE) |
| 467 |
args.in.timeout = timeout_ns; |
| 468 |
else |
| 469 |
args.in.timeout = amdgpu_cs_calculate_timeout(timeout_ns); |
| 470 |
|
| 471 |
r = drmIoctl(dev->fd, DRM_IOCTL_AMDGPU_WAIT_CS, &args); |
| 472 |
if (r) |
| 473 |
return -errno; |
| 474 |
|
| 475 |
*busy = args.out.status; |
| 476 |
return 0; |
| 477 |
} |
| 478 |
|
| 479 |
drm_public int amdgpu_cs_query_fence_status(struct amdgpu_cs_fence *fence, |
| 480 |
uint64_t timeout_ns, |
| 481 |
uint64_t flags, |
| 482 |
uint32_t *expired) |
| 483 |
{ |
| 484 |
bool busy = true; |
| 485 |
int r; |
| 486 |
|
| 487 |
if (!fence || !expired || !fence->context) |
| 488 |
return -EINVAL; |
| 489 |
if (fence->ip_type >= AMDGPU_HW_IP_NUM) |
| 490 |
return -EINVAL; |
| 491 |
if (fence->ring >= AMDGPU_CS_MAX_RINGS) |
| 492 |
return -EINVAL; |
| 493 |
if (fence->fence == AMDGPU_NULL_SUBMIT_SEQ) { |
| 494 |
*expired = true; |
| 495 |
return 0; |
| 496 |
} |
| 497 |
|
| 498 |
*expired = false; |
| 499 |
|
| 500 |
r = amdgpu_ioctl_wait_cs(fence->context, fence->ip_type, |
| 501 |
fence->ip_instance, fence->ring, |
| 502 |
fence->fence, timeout_ns, flags, &busy); |
| 503 |
|
| 504 |
if (!r && !busy) |
| 505 |
*expired = true; |
| 506 |
|
| 507 |
return r; |
| 508 |
} |
| 509 |
|
| 510 |
static int amdgpu_ioctl_wait_fences(struct amdgpu_cs_fence *fences, |
| 511 |
uint32_t fence_count, |
| 512 |
bool wait_all, |
| 513 |
uint64_t timeout_ns, |
| 514 |
uint32_t *status, |
| 515 |
uint32_t *first) |
| 516 |
{ |
| 517 |
struct drm_amdgpu_fence *drm_fences; |
| 518 |
amdgpu_device_handle dev = fences[0].context->dev; |
| 519 |
union drm_amdgpu_wait_fences args; |
| 520 |
int r; |
| 521 |
uint32_t i; |
| 522 |
|
| 523 |
drm_fences = alloca(sizeof(struct drm_amdgpu_fence) * fence_count); |
| 524 |
for (i = 0; i < fence_count; i++) { |
| 525 |
drm_fences[i].ctx_id = fences[i].context->id; |
| 526 |
drm_fences[i].ip_type = fences[i].ip_type; |
| 527 |
drm_fences[i].ip_instance = fences[i].ip_instance; |
| 528 |
drm_fences[i].ring = fences[i].ring; |
| 529 |
drm_fences[i].seq_no = fences[i].fence; |
| 530 |
} |
| 531 |
|
| 532 |
memset(&args, 0, sizeof(args)); |
| 533 |
args.in.fences = (uint64_t)(uintptr_t)drm_fences; |
| 534 |
args.in.fence_count = fence_count; |
| 535 |
args.in.wait_all = wait_all; |
| 536 |
args.in.timeout_ns = amdgpu_cs_calculate_timeout(timeout_ns); |
| 537 |
|
| 538 |
r = drmIoctl(dev->fd, DRM_IOCTL_AMDGPU_WAIT_FENCES, &args); |
| 539 |
if (r) |
| 540 |
return -errno; |
| 541 |
|
| 542 |
*status = args.out.status; |
| 543 |
|
| 544 |
if (first) |
| 545 |
*first = args.out.first_signaled; |
| 546 |
|
| 547 |
return 0; |
| 548 |
} |
| 549 |
|
| 550 |
drm_public int amdgpu_cs_wait_fences(struct amdgpu_cs_fence *fences, |
| 551 |
uint32_t fence_count, |
| 552 |
bool wait_all, |
| 553 |
uint64_t timeout_ns, |
| 554 |
uint32_t *status, |
| 555 |
uint32_t *first) |
| 556 |
{ |
| 557 |
uint32_t i; |
| 558 |
|
| 559 |
/* Sanity check */ |
| 560 |
if (!fences || !status || !fence_count) |
| 561 |
return -EINVAL; |
| 562 |
|
| 563 |
for (i = 0; i < fence_count; i++) { |
| 564 |
if (NULL == fences[i].context) |
| 565 |
return -EINVAL; |
| 566 |
if (fences[i].ip_type >= AMDGPU_HW_IP_NUM) |
| 567 |
return -EINVAL; |
| 568 |
if (fences[i].ring >= AMDGPU_CS_MAX_RINGS) |
| 569 |
return -EINVAL; |
| 570 |
} |
| 571 |
|
| 572 |
*status = 0; |
| 573 |
|
| 574 |
return amdgpu_ioctl_wait_fences(fences, fence_count, wait_all, |
| 575 |
timeout_ns, status, first); |
| 576 |
} |
| 577 |
|
| 578 |
drm_public int amdgpu_cs_create_semaphore(amdgpu_semaphore_handle *sem) |
| 579 |
{ |
| 580 |
struct amdgpu_semaphore *gpu_semaphore; |
| 581 |
|
| 582 |
if (!sem) |
| 583 |
return -EINVAL; |
| 584 |
|
| 585 |
gpu_semaphore = calloc(1, sizeof(struct amdgpu_semaphore)); |
| 586 |
if (!gpu_semaphore) |
| 587 |
return -ENOMEM; |
| 588 |
|
| 589 |
atomic_set(&gpu_semaphore->refcount, 1); |
| 590 |
*sem = gpu_semaphore; |
| 591 |
|
| 592 |
return 0; |
| 593 |
} |
| 594 |
|
| 595 |
drm_public int amdgpu_cs_signal_semaphore(amdgpu_context_handle ctx, |
| 596 |
uint32_t ip_type, |
| 597 |
uint32_t ip_instance, |
| 598 |
uint32_t ring, |
| 599 |
amdgpu_semaphore_handle sem) |
| 600 |
{ |
| 601 |
int ret; |
| 602 |
|
| 603 |
if (!ctx || !sem) |
| 604 |
return -EINVAL; |
| 605 |
if (ip_type >= AMDGPU_HW_IP_NUM) |
| 606 |
return -EINVAL; |
| 607 |
if (ring >= AMDGPU_CS_MAX_RINGS) |
| 608 |
return -EINVAL; |
| 609 |
|
| 610 |
pthread_mutex_lock(&ctx->sequence_mutex); |
| 611 |
/* sem has been signaled */ |
| 612 |
if (sem->signal_fence.context) { |
| 613 |
ret = -EINVAL; |
| 614 |
goto unlock; |
| 615 |
} |
| 616 |
sem->signal_fence.context = ctx; |
| 617 |
sem->signal_fence.ip_type = ip_type; |
| 618 |
sem->signal_fence.ip_instance = ip_instance; |
| 619 |
sem->signal_fence.ring = ring; |
| 620 |
sem->signal_fence.fence = ctx->last_seq[ip_type][ip_instance][ring]; |
| 621 |
update_references(NULL, &sem->refcount); |
| 622 |
ret = 0; |
| 623 |
unlock: |
| 624 |
pthread_mutex_unlock(&ctx->sequence_mutex); |
| 625 |
return ret; |
| 626 |
} |
| 627 |
|
| 628 |
drm_public int amdgpu_cs_wait_semaphore(amdgpu_context_handle ctx, |
| 629 |
uint32_t ip_type, |
| 630 |
uint32_t ip_instance, |
| 631 |
uint32_t ring, |
| 632 |
amdgpu_semaphore_handle sem) |
| 633 |
{ |
| 634 |
if (!ctx || !sem) |
| 635 |
return -EINVAL; |
| 636 |
if (ip_type >= AMDGPU_HW_IP_NUM) |
| 637 |
return -EINVAL; |
| 638 |
if (ring >= AMDGPU_CS_MAX_RINGS) |
| 639 |
return -EINVAL; |
| 640 |
/* must signal first */ |
| 641 |
if (!sem->signal_fence.context) |
| 642 |
return -EINVAL; |
| 643 |
|
| 644 |
pthread_mutex_lock(&ctx->sequence_mutex); |
| 645 |
list_add(&sem->list, &ctx->sem_list[ip_type][ip_instance][ring]); |
| 646 |
pthread_mutex_unlock(&ctx->sequence_mutex); |
| 647 |
return 0; |
| 648 |
} |
| 649 |
|
| 650 |
static int amdgpu_cs_reset_sem(amdgpu_semaphore_handle sem) |
| 651 |
{ |
| 652 |
if (!sem || !sem->signal_fence.context) |
| 653 |
return -EINVAL; |
| 654 |
|
| 655 |
sem->signal_fence.context = NULL; |
| 656 |
sem->signal_fence.ip_type = 0; |
| 657 |
sem->signal_fence.ip_instance = 0; |
| 658 |
sem->signal_fence.ring = 0; |
| 659 |
sem->signal_fence.fence = 0; |
| 660 |
|
| 661 |
return 0; |
| 662 |
} |
| 663 |
|
| 664 |
static int amdgpu_cs_unreference_sem(amdgpu_semaphore_handle sem) |
| 665 |
{ |
| 666 |
if (!sem) |
| 667 |
return -EINVAL; |
| 668 |
|
| 669 |
if (update_references(&sem->refcount, NULL)) |
| 670 |
free(sem); |
| 671 |
return 0; |
| 672 |
} |
| 673 |
|
| 674 |
drm_public int amdgpu_cs_destroy_semaphore(amdgpu_semaphore_handle sem) |
| 675 |
{ |
| 676 |
return amdgpu_cs_unreference_sem(sem); |
| 677 |
} |
| 678 |
|
| 679 |
drm_public int amdgpu_cs_create_syncobj2(amdgpu_device_handle dev, |
| 680 |
uint32_t flags, |
| 681 |
uint32_t *handle) |
| 682 |
{ |
| 683 |
if (NULL == dev) |
| 684 |
return -EINVAL; |
| 685 |
|
| 686 |
return drmSyncobjCreate(dev->fd, flags, handle); |
| 687 |
} |
| 688 |
|
| 689 |
drm_public int amdgpu_cs_create_syncobj(amdgpu_device_handle dev, |
| 690 |
uint32_t *handle) |
| 691 |
{ |
| 692 |
if (NULL == dev) |
| 693 |
return -EINVAL; |
| 694 |
|
| 695 |
return drmSyncobjCreate(dev->fd, 0, handle); |
| 696 |
} |
| 697 |
|
| 698 |
drm_public int amdgpu_cs_destroy_syncobj(amdgpu_device_handle dev, |
| 699 |
uint32_t handle) |
| 700 |
{ |
| 701 |
if (NULL == dev) |
| 702 |
return -EINVAL; |
| 703 |
|
| 704 |
return drmSyncobjDestroy(dev->fd, handle); |
| 705 |
} |
| 706 |
|
| 707 |
drm_public int amdgpu_cs_syncobj_reset(amdgpu_device_handle dev, |
| 708 |
const uint32_t *syncobjs, |
| 709 |
uint32_t syncobj_count) |
| 710 |
{ |
| 711 |
if (NULL == dev) |
| 712 |
return -EINVAL; |
| 713 |
|
| 714 |
return drmSyncobjReset(dev->fd, syncobjs, syncobj_count); |
| 715 |
} |
| 716 |
|
| 717 |
drm_public int amdgpu_cs_syncobj_signal(amdgpu_device_handle dev, |
| 718 |
const uint32_t *syncobjs, |
| 719 |
uint32_t syncobj_count) |
| 720 |
{ |
| 721 |
if (NULL == dev) |
| 722 |
return -EINVAL; |
| 723 |
|
| 724 |
return drmSyncobjSignal(dev->fd, syncobjs, syncobj_count); |
| 725 |
} |
| 726 |
|
| 727 |
drm_public int amdgpu_cs_syncobj_timeline_signal(amdgpu_device_handle dev, |
| 728 |
const uint32_t *syncobjs, |
| 729 |
uint64_t *points, |
| 730 |
uint32_t syncobj_count) |
| 731 |
{ |
| 732 |
if (NULL == dev) |
| 733 |
return -EINVAL; |
| 734 |
|
| 735 |
return drmSyncobjTimelineSignal(dev->fd, syncobjs, |
| 736 |
points, syncobj_count); |
| 737 |
} |
| 738 |
|
| 739 |
drm_public int amdgpu_cs_syncobj_wait(amdgpu_device_handle dev, |
| 740 |
uint32_t *handles, unsigned num_handles, |
| 741 |
int64_t timeout_nsec, unsigned flags, |
| 742 |
uint32_t *first_signaled) |
| 743 |
{ |
| 744 |
if (NULL == dev) |
| 745 |
return -EINVAL; |
| 746 |
|
| 747 |
return drmSyncobjWait(dev->fd, handles, num_handles, timeout_nsec, |
| 748 |
flags, first_signaled); |
| 749 |
} |
| 750 |
|
| 751 |
drm_public int amdgpu_cs_syncobj_timeline_wait(amdgpu_device_handle dev, |
| 752 |
uint32_t *handles, uint64_t *points, |
| 753 |
unsigned num_handles, |
| 754 |
int64_t timeout_nsec, unsigned flags, |
| 755 |
uint32_t *first_signaled) |
| 756 |
{ |
| 757 |
if (NULL == dev) |
| 758 |
return -EINVAL; |
| 759 |
|
| 760 |
return drmSyncobjTimelineWait(dev->fd, handles, points, num_handles, |
| 761 |
timeout_nsec, flags, first_signaled); |
| 762 |
} |
| 763 |
|
| 764 |
drm_public int amdgpu_cs_syncobj_query(amdgpu_device_handle dev, |
| 765 |
uint32_t *handles, uint64_t *points, |
| 766 |
unsigned num_handles) |
| 767 |
{ |
| 768 |
if (NULL == dev) |
| 769 |
return -EINVAL; |
| 770 |
|
| 771 |
return drmSyncobjQuery(dev->fd, handles, points, num_handles); |
| 772 |
} |
| 773 |
|
| 774 |
drm_public int amdgpu_cs_syncobj_query2(amdgpu_device_handle dev, |
| 775 |
uint32_t *handles, uint64_t *points, |
| 776 |
unsigned num_handles, uint32_t flags) |
| 777 |
{ |
| 778 |
if (!dev) |
| 779 |
return -EINVAL; |
| 780 |
|
| 781 |
return drmSyncobjQuery2(dev->fd, handles, points, num_handles, flags); |
| 782 |
} |
| 783 |
|
| 784 |
drm_public int amdgpu_cs_export_syncobj(amdgpu_device_handle dev, |
| 785 |
uint32_t handle, |
| 786 |
int *shared_fd) |
| 787 |
{ |
| 788 |
if (NULL == dev) |
| 789 |
return -EINVAL; |
| 790 |
|
| 791 |
return drmSyncobjHandleToFD(dev->fd, handle, shared_fd); |
| 792 |
} |
| 793 |
|
| 794 |
drm_public int amdgpu_cs_import_syncobj(amdgpu_device_handle dev, |
| 795 |
int shared_fd, |
| 796 |
uint32_t *handle) |
| 797 |
{ |
| 798 |
if (NULL == dev) |
| 799 |
return -EINVAL; |
| 800 |
|
| 801 |
return drmSyncobjFDToHandle(dev->fd, shared_fd, handle); |
| 802 |
} |
| 803 |
|
| 804 |
drm_public int amdgpu_cs_syncobj_export_sync_file(amdgpu_device_handle dev, |
| 805 |
uint32_t syncobj, |
| 806 |
int *sync_file_fd) |
| 807 |
{ |
| 808 |
if (NULL == dev) |
| 809 |
return -EINVAL; |
| 810 |
|
| 811 |
return drmSyncobjExportSyncFile(dev->fd, syncobj, sync_file_fd); |
| 812 |
} |
| 813 |
|
| 814 |
drm_public int amdgpu_cs_syncobj_import_sync_file(amdgpu_device_handle dev, |
| 815 |
uint32_t syncobj, |
| 816 |
int sync_file_fd) |
| 817 |
{ |
| 818 |
if (NULL == dev) |
| 819 |
return -EINVAL; |
| 820 |
|
| 821 |
return drmSyncobjImportSyncFile(dev->fd, syncobj, sync_file_fd); |
| 822 |
} |
| 823 |
|
| 824 |
drm_public int amdgpu_cs_syncobj_export_sync_file2(amdgpu_device_handle dev, |
| 825 |
uint32_t syncobj, |
| 826 |
uint64_t point, |
| 827 |
uint32_t flags, |
| 828 |
int *sync_file_fd) |
| 829 |
{ |
| 830 |
uint32_t binary_handle; |
| 831 |
int ret; |
| 832 |
|
| 833 |
if (NULL == dev) |
| 834 |
return -EINVAL; |
| 835 |
|
| 836 |
if (!point) |
| 837 |
return drmSyncobjExportSyncFile(dev->fd, syncobj, sync_file_fd); |
| 838 |
|
| 839 |
ret = drmSyncobjCreate(dev->fd, 0, &binary_handle); |
| 840 |
if (ret) |
| 841 |
return ret; |
| 842 |
|
| 843 |
ret = drmSyncobjTransfer(dev->fd, binary_handle, 0, |
| 844 |
syncobj, point, flags); |
| 845 |
if (ret) |
| 846 |
goto out; |
| 847 |
ret = drmSyncobjExportSyncFile(dev->fd, binary_handle, sync_file_fd); |
| 848 |
out: |
| 849 |
drmSyncobjDestroy(dev->fd, binary_handle); |
| 850 |
return ret; |
| 851 |
} |
| 852 |
|
| 853 |
drm_public int amdgpu_cs_syncobj_import_sync_file2(amdgpu_device_handle dev, |
| 854 |
uint32_t syncobj, |
| 855 |
uint64_t point, |
| 856 |
int sync_file_fd) |
| 857 |
{ |
| 858 |
uint32_t binary_handle; |
| 859 |
int ret; |
| 860 |
|
| 861 |
if (NULL == dev) |
| 862 |
return -EINVAL; |
| 863 |
|
| 864 |
if (!point) |
| 865 |
return drmSyncobjImportSyncFile(dev->fd, syncobj, sync_file_fd); |
| 866 |
|
| 867 |
ret = drmSyncobjCreate(dev->fd, 0, &binary_handle); |
| 868 |
if (ret) |
| 869 |
return ret; |
| 870 |
ret = drmSyncobjImportSyncFile(dev->fd, binary_handle, sync_file_fd); |
| 871 |
if (ret) |
| 872 |
goto out; |
| 873 |
ret = drmSyncobjTransfer(dev->fd, syncobj, point, |
| 874 |
binary_handle, 0, 0); |
| 875 |
out: |
| 876 |
drmSyncobjDestroy(dev->fd, binary_handle); |
| 877 |
return ret; |
| 878 |
} |
| 879 |
|
| 880 |
drm_public int amdgpu_cs_syncobj_transfer(amdgpu_device_handle dev, |
| 881 |
uint32_t dst_handle, |
| 882 |
uint64_t dst_point, |
| 883 |
uint32_t src_handle, |
| 884 |
uint64_t src_point, |
| 885 |
uint32_t flags) |
| 886 |
{ |
| 887 |
if (NULL == dev) |
| 888 |
return -EINVAL; |
| 889 |
|
| 890 |
return drmSyncobjTransfer(dev->fd, |
| 891 |
dst_handle, dst_point, |
| 892 |
src_handle, src_point, |
| 893 |
flags); |
| 894 |
} |
| 895 |
|
| 896 |
drm_public int amdgpu_cs_submit_raw(amdgpu_device_handle dev, |
| 897 |
amdgpu_context_handle context, |
| 898 |
amdgpu_bo_list_handle bo_list_handle, |
| 899 |
int num_chunks, |
| 900 |
struct drm_amdgpu_cs_chunk *chunks, |
| 901 |
uint64_t *seq_no) |
| 902 |
{ |
| 903 |
union drm_amdgpu_cs cs; |
| 904 |
uint64_t *chunk_array; |
| 905 |
int i, r; |
| 906 |
if (num_chunks == 0) |
| 907 |
return -EINVAL; |
| 908 |
|
| 909 |
memset(&cs, 0, sizeof(cs)); |
| 910 |
chunk_array = alloca(sizeof(uint64_t) * num_chunks); |
| 911 |
for (i = 0; i < num_chunks; i++) |
| 912 |
chunk_array[i] = (uint64_t)(uintptr_t)&chunks[i]; |
| 913 |
cs.in.chunks = (uint64_t)(uintptr_t)chunk_array; |
| 914 |
cs.in.ctx_id = context->id; |
| 915 |
cs.in.bo_list_handle = bo_list_handle ? bo_list_handle->handle : 0; |
| 916 |
cs.in.num_chunks = num_chunks; |
| 917 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_CS, |
| 918 |
&cs, sizeof(cs)); |
| 919 |
if (r) |
| 920 |
return r; |
| 921 |
|
| 922 |
if (seq_no) |
| 923 |
*seq_no = cs.out.handle; |
| 924 |
return 0; |
| 925 |
} |
| 926 |
|
| 927 |
drm_public int amdgpu_cs_submit_raw2(amdgpu_device_handle dev, |
| 928 |
amdgpu_context_handle context, |
| 929 |
uint32_t bo_list_handle, |
| 930 |
int num_chunks, |
| 931 |
struct drm_amdgpu_cs_chunk *chunks, |
| 932 |
uint64_t *seq_no) |
| 933 |
{ |
| 934 |
union drm_amdgpu_cs cs; |
| 935 |
uint64_t *chunk_array; |
| 936 |
int i, r; |
| 937 |
|
| 938 |
memset(&cs, 0, sizeof(cs)); |
| 939 |
chunk_array = alloca(sizeof(uint64_t) * num_chunks); |
| 940 |
for (i = 0; i < num_chunks; i++) |
| 941 |
chunk_array[i] = (uint64_t)(uintptr_t)&chunks[i]; |
| 942 |
cs.in.chunks = (uint64_t)(uintptr_t)chunk_array; |
| 943 |
cs.in.ctx_id = context->id; |
| 944 |
cs.in.bo_list_handle = bo_list_handle; |
| 945 |
cs.in.num_chunks = num_chunks; |
| 946 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_CS, |
| 947 |
&cs, sizeof(cs)); |
| 948 |
if (!r && seq_no) |
| 949 |
*seq_no = cs.out.handle; |
| 950 |
return r; |
| 951 |
} |
| 952 |
|
| 953 |
drm_public void amdgpu_cs_chunk_fence_info_to_data(struct amdgpu_cs_fence_info *fence_info, |
| 954 |
struct drm_amdgpu_cs_chunk_data *data) |
| 955 |
{ |
| 956 |
data->fence_data.handle = fence_info->handle->handle; |
| 957 |
data->fence_data.offset = fence_info->offset * sizeof(uint64_t); |
| 958 |
} |
| 959 |
|
| 960 |
drm_public void amdgpu_cs_chunk_fence_to_dep(struct amdgpu_cs_fence *fence, |
| 961 |
struct drm_amdgpu_cs_chunk_dep *dep) |
| 962 |
{ |
| 963 |
dep->ip_type = fence->ip_type; |
| 964 |
dep->ip_instance = fence->ip_instance; |
| 965 |
dep->ring = fence->ring; |
| 966 |
dep->ctx_id = fence->context->id; |
| 967 |
dep->handle = fence->fence; |
| 968 |
} |
| 969 |
|
| 970 |
drm_public int amdgpu_cs_fence_to_handle(amdgpu_device_handle dev, |
| 971 |
struct amdgpu_cs_fence *fence, |
| 972 |
uint32_t what, |
| 973 |
uint32_t *out_handle) |
| 974 |
{ |
| 975 |
union drm_amdgpu_fence_to_handle fth; |
| 976 |
int r; |
| 977 |
|
| 978 |
memset(&fth, 0, sizeof(fth)); |
| 979 |
fth.in.fence.ctx_id = fence->context->id; |
| 980 |
fth.in.fence.ip_type = fence->ip_type; |
| 981 |
fth.in.fence.ip_instance = fence->ip_instance; |
| 982 |
fth.in.fence.ring = fence->ring; |
| 983 |
fth.in.fence.seq_no = fence->fence; |
| 984 |
fth.in.what = what; |
| 985 |
|
| 986 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_FENCE_TO_HANDLE, |
| 987 |
&fth, sizeof(fth)); |
| 988 |
if (r == 0) |
| 989 |
*out_handle = fth.out.handle; |
| 990 |
return r; |
| 991 |
} |