| 1 |
/* |
| 2 |
* Copyright © 2014 Advanced Micro Devices, Inc. |
| 3 |
* All Rights Reserved. |
| 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 <stdlib.h> |
| 26 |
#include <stdio.h> |
| 27 |
#include <stdint.h> |
| 28 |
#include <string.h> |
| 29 |
#include <errno.h> |
| 30 |
#include <fcntl.h> |
| 31 |
#include <unistd.h> |
| 32 |
#include <sys/ioctl.h> |
| 33 |
#include <sys/mman.h> |
| 34 |
#include <sys/time.h> |
| 35 |
|
| 36 |
#include "libdrm_macros.h" |
| 37 |
#include "xf86drm.h" |
| 38 |
#include "amdgpu_drm.h" |
| 39 |
#include "amdgpu_internal.h" |
| 40 |
#include "util_math.h" |
| 41 |
|
| 42 |
static int amdgpu_bo_create(amdgpu_device_handle dev, |
| 43 |
uint64_t size, |
| 44 |
uint32_t handle, |
| 45 |
amdgpu_bo_handle *buf_handle) |
| 46 |
{ |
| 47 |
struct amdgpu_bo *bo; |
| 48 |
int r; |
| 49 |
|
| 50 |
bo = calloc(1, sizeof(struct amdgpu_bo)); |
| 51 |
if (!bo) |
| 52 |
return -ENOMEM; |
| 53 |
|
| 54 |
r = handle_table_insert(&dev->bo_handles, handle, bo); |
| 55 |
if (r) { |
| 56 |
free(bo); |
| 57 |
return r; |
| 58 |
} |
| 59 |
|
| 60 |
atomic_set(&bo->refcount, 1); |
| 61 |
bo->dev = dev; |
| 62 |
bo->alloc_size = size; |
| 63 |
bo->handle = handle; |
| 64 |
pthread_mutex_init(&bo->cpu_access_mutex, NULL); |
| 65 |
|
| 66 |
*buf_handle = bo; |
| 67 |
return 0; |
| 68 |
} |
| 69 |
|
| 70 |
drm_public int amdgpu_bo_alloc(amdgpu_device_handle dev, |
| 71 |
struct amdgpu_bo_alloc_request *alloc_buffer, |
| 72 |
amdgpu_bo_handle *buf_handle) |
| 73 |
{ |
| 74 |
union drm_amdgpu_gem_create args; |
| 75 |
int r; |
| 76 |
|
| 77 |
if (!alloc_buffer || !buf_handle) |
| 78 |
return -EINVAL; |
| 79 |
|
| 80 |
memset(&args, 0, sizeof(args)); |
| 81 |
args.in.bo_size = alloc_buffer->alloc_size; |
| 82 |
args.in.alignment = alloc_buffer->phys_alignment; |
| 83 |
|
| 84 |
/* Set the placement. */ |
| 85 |
args.in.domains = alloc_buffer->preferred_heap; |
| 86 |
args.in.domain_flags = alloc_buffer->flags; |
| 87 |
|
| 88 |
/* Allocate the buffer with the preferred heap. */ |
| 89 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_CREATE, |
| 90 |
&args, sizeof(args)); |
| 91 |
if (r) |
| 92 |
goto out; |
| 93 |
|
| 94 |
pthread_mutex_lock(&dev->bo_table_mutex); |
| 95 |
r = amdgpu_bo_create(dev, alloc_buffer->alloc_size, args.out.handle, |
| 96 |
buf_handle); |
| 97 |
pthread_mutex_unlock(&dev->bo_table_mutex); |
| 98 |
if (r) { |
| 99 |
drmCloseBufferHandle(dev->fd, args.out.handle); |
| 100 |
} |
| 101 |
|
| 102 |
out: |
| 103 |
return r; |
| 104 |
} |
| 105 |
|
| 106 |
drm_public int amdgpu_bo_set_metadata(amdgpu_bo_handle bo, |
| 107 |
struct amdgpu_bo_metadata *info) |
| 108 |
{ |
| 109 |
struct drm_amdgpu_gem_metadata args = {}; |
| 110 |
|
| 111 |
if (!info) |
| 112 |
return -EINVAL; |
| 113 |
|
| 114 |
args.handle = bo->handle; |
| 115 |
args.op = AMDGPU_GEM_METADATA_OP_SET_METADATA; |
| 116 |
args.data.flags = info->flags; |
| 117 |
args.data.tiling_info = info->tiling_info; |
| 118 |
|
| 119 |
if (info->size_metadata > sizeof(args.data.data)) |
| 120 |
return -EINVAL; |
| 121 |
|
| 122 |
if (info->size_metadata) { |
| 123 |
args.data.data_size_bytes = info->size_metadata; |
| 124 |
memcpy(args.data.data, info->umd_metadata, info->size_metadata); |
| 125 |
} |
| 126 |
|
| 127 |
return drmCommandWriteRead(bo->dev->fd, |
| 128 |
DRM_AMDGPU_GEM_METADATA, |
| 129 |
&args, sizeof(args)); |
| 130 |
} |
| 131 |
|
| 132 |
drm_public int amdgpu_bo_query_info(amdgpu_bo_handle bo, |
| 133 |
struct amdgpu_bo_info *info) |
| 134 |
{ |
| 135 |
struct drm_amdgpu_gem_metadata metadata = {}; |
| 136 |
struct drm_amdgpu_gem_create_in bo_info = {}; |
| 137 |
struct drm_amdgpu_gem_op gem_op = {}; |
| 138 |
int r; |
| 139 |
|
| 140 |
/* Validate the BO passed in */ |
| 141 |
if (!bo->handle || !info) |
| 142 |
return -EINVAL; |
| 143 |
|
| 144 |
/* Query metadata. */ |
| 145 |
metadata.handle = bo->handle; |
| 146 |
metadata.op = AMDGPU_GEM_METADATA_OP_GET_METADATA; |
| 147 |
|
| 148 |
r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_METADATA, |
| 149 |
&metadata, sizeof(metadata)); |
| 150 |
if (r) |
| 151 |
return r; |
| 152 |
|
| 153 |
if (metadata.data.data_size_bytes > |
| 154 |
sizeof(info->metadata.umd_metadata)) |
| 155 |
return -EINVAL; |
| 156 |
|
| 157 |
/* Query buffer info. */ |
| 158 |
gem_op.handle = bo->handle; |
| 159 |
gem_op.op = AMDGPU_GEM_OP_GET_GEM_CREATE_INFO; |
| 160 |
gem_op.value = (uintptr_t)&bo_info; |
| 161 |
|
| 162 |
r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_OP, |
| 163 |
&gem_op, sizeof(gem_op)); |
| 164 |
if (r) |
| 165 |
return r; |
| 166 |
|
| 167 |
memset(info, 0, sizeof(*info)); |
| 168 |
info->alloc_size = bo_info.bo_size; |
| 169 |
info->phys_alignment = bo_info.alignment; |
| 170 |
info->preferred_heap = bo_info.domains; |
| 171 |
info->alloc_flags = bo_info.domain_flags; |
| 172 |
info->metadata.flags = metadata.data.flags; |
| 173 |
info->metadata.tiling_info = metadata.data.tiling_info; |
| 174 |
|
| 175 |
info->metadata.size_metadata = metadata.data.data_size_bytes; |
| 176 |
if (metadata.data.data_size_bytes > 0) |
| 177 |
memcpy(info->metadata.umd_metadata, metadata.data.data, |
| 178 |
metadata.data.data_size_bytes); |
| 179 |
|
| 180 |
return 0; |
| 181 |
} |
| 182 |
|
| 183 |
static int amdgpu_bo_export_flink(amdgpu_bo_handle bo) |
| 184 |
{ |
| 185 |
struct drm_gem_flink flink; |
| 186 |
int fd, dma_fd; |
| 187 |
uint32_t handle; |
| 188 |
int r; |
| 189 |
|
| 190 |
fd = bo->dev->fd; |
| 191 |
handle = bo->handle; |
| 192 |
if (bo->flink_name) |
| 193 |
return 0; |
| 194 |
|
| 195 |
|
| 196 |
if (bo->dev->flink_fd != bo->dev->fd) { |
| 197 |
r = drmPrimeHandleToFD(bo->dev->fd, bo->handle, DRM_CLOEXEC, |
| 198 |
&dma_fd); |
| 199 |
if (!r) { |
| 200 |
r = drmPrimeFDToHandle(bo->dev->flink_fd, dma_fd, &handle); |
| 201 |
close(dma_fd); |
| 202 |
} |
| 203 |
if (r) |
| 204 |
return r; |
| 205 |
fd = bo->dev->flink_fd; |
| 206 |
} |
| 207 |
memset(&flink, 0, sizeof(flink)); |
| 208 |
flink.handle = handle; |
| 209 |
|
| 210 |
r = drmIoctl(fd, DRM_IOCTL_GEM_FLINK, &flink); |
| 211 |
if (r) |
| 212 |
return r; |
| 213 |
|
| 214 |
bo->flink_name = flink.name; |
| 215 |
|
| 216 |
if (bo->dev->flink_fd != bo->dev->fd) |
| 217 |
drmCloseBufferHandle(bo->dev->flink_fd, handle); |
| 218 |
|
| 219 |
pthread_mutex_lock(&bo->dev->bo_table_mutex); |
| 220 |
r = handle_table_insert(&bo->dev->bo_flink_names, bo->flink_name, bo); |
| 221 |
pthread_mutex_unlock(&bo->dev->bo_table_mutex); |
| 222 |
|
| 223 |
return r; |
| 224 |
} |
| 225 |
|
| 226 |
drm_public int amdgpu_bo_export(amdgpu_bo_handle bo, |
| 227 |
enum amdgpu_bo_handle_type type, |
| 228 |
uint32_t *shared_handle) |
| 229 |
{ |
| 230 |
int r; |
| 231 |
|
| 232 |
switch (type) { |
| 233 |
case amdgpu_bo_handle_type_gem_flink_name: |
| 234 |
r = amdgpu_bo_export_flink(bo); |
| 235 |
if (r) |
| 236 |
return r; |
| 237 |
|
| 238 |
*shared_handle = bo->flink_name; |
| 239 |
return 0; |
| 240 |
|
| 241 |
case amdgpu_bo_handle_type_kms: |
| 242 |
case amdgpu_bo_handle_type_kms_noimport: |
| 243 |
*shared_handle = bo->handle; |
| 244 |
return 0; |
| 245 |
|
| 246 |
case amdgpu_bo_handle_type_dma_buf_fd: |
| 247 |
return drmPrimeHandleToFD(bo->dev->fd, bo->handle, |
| 248 |
DRM_CLOEXEC | DRM_RDWR, |
| 249 |
(int*)shared_handle); |
| 250 |
} |
| 251 |
return -EINVAL; |
| 252 |
} |
| 253 |
|
| 254 |
drm_public int amdgpu_bo_import(amdgpu_device_handle dev, |
| 255 |
enum amdgpu_bo_handle_type type, |
| 256 |
uint32_t shared_handle, |
| 257 |
struct amdgpu_bo_import_result *output) |
| 258 |
{ |
| 259 |
struct drm_gem_open open_arg = {}; |
| 260 |
struct amdgpu_bo *bo = NULL; |
| 261 |
uint32_t handle = 0, flink_name = 0; |
| 262 |
uint64_t alloc_size = 0; |
| 263 |
int r = 0; |
| 264 |
int dma_fd; |
| 265 |
uint64_t dma_buf_size = 0; |
| 266 |
|
| 267 |
/* We must maintain a list of pairs <handle, bo>, so that we always |
| 268 |
* return the same amdgpu_bo instance for the same handle. */ |
| 269 |
pthread_mutex_lock(&dev->bo_table_mutex); |
| 270 |
|
| 271 |
/* Convert a DMA buf handle to a KMS handle now. */ |
| 272 |
if (type == amdgpu_bo_handle_type_dma_buf_fd) { |
| 273 |
off_t size; |
| 274 |
|
| 275 |
/* Get a KMS handle. */ |
| 276 |
r = drmPrimeFDToHandle(dev->fd, shared_handle, &handle); |
| 277 |
if (r) |
| 278 |
goto unlock; |
| 279 |
|
| 280 |
/* Query the buffer size. */ |
| 281 |
size = lseek(shared_handle, 0, SEEK_END); |
| 282 |
if (size == (off_t)-1) { |
| 283 |
r = -errno; |
| 284 |
goto free_bo_handle; |
| 285 |
} |
| 286 |
lseek(shared_handle, 0, SEEK_SET); |
| 287 |
|
| 288 |
dma_buf_size = size; |
| 289 |
shared_handle = handle; |
| 290 |
} |
| 291 |
|
| 292 |
/* If we have already created a buffer with this handle, find it. */ |
| 293 |
switch (type) { |
| 294 |
case amdgpu_bo_handle_type_gem_flink_name: |
| 295 |
bo = handle_table_lookup(&dev->bo_flink_names, shared_handle); |
| 296 |
break; |
| 297 |
|
| 298 |
case amdgpu_bo_handle_type_dma_buf_fd: |
| 299 |
bo = handle_table_lookup(&dev->bo_handles, shared_handle); |
| 300 |
break; |
| 301 |
|
| 302 |
case amdgpu_bo_handle_type_kms: |
| 303 |
case amdgpu_bo_handle_type_kms_noimport: |
| 304 |
/* Importing a KMS handle in not allowed. */ |
| 305 |
r = -EPERM; |
| 306 |
goto unlock; |
| 307 |
|
| 308 |
default: |
| 309 |
r = -EINVAL; |
| 310 |
goto unlock; |
| 311 |
} |
| 312 |
|
| 313 |
if (bo) { |
| 314 |
/* The buffer already exists, just bump the refcount. */ |
| 315 |
atomic_inc(&bo->refcount); |
| 316 |
pthread_mutex_unlock(&dev->bo_table_mutex); |
| 317 |
|
| 318 |
output->buf_handle = bo; |
| 319 |
output->alloc_size = bo->alloc_size; |
| 320 |
return 0; |
| 321 |
} |
| 322 |
|
| 323 |
/* Open the handle. */ |
| 324 |
switch (type) { |
| 325 |
case amdgpu_bo_handle_type_gem_flink_name: |
| 326 |
open_arg.name = shared_handle; |
| 327 |
r = drmIoctl(dev->flink_fd, DRM_IOCTL_GEM_OPEN, &open_arg); |
| 328 |
if (r) |
| 329 |
goto unlock; |
| 330 |
|
| 331 |
flink_name = shared_handle; |
| 332 |
handle = open_arg.handle; |
| 333 |
alloc_size = open_arg.size; |
| 334 |
if (dev->flink_fd != dev->fd) { |
| 335 |
r = drmPrimeHandleToFD(dev->flink_fd, handle, |
| 336 |
DRM_CLOEXEC, &dma_fd); |
| 337 |
if (r) |
| 338 |
goto free_bo_handle; |
| 339 |
r = drmPrimeFDToHandle(dev->fd, dma_fd, &handle); |
| 340 |
close(dma_fd); |
| 341 |
if (r) |
| 342 |
goto free_bo_handle; |
| 343 |
r = drmCloseBufferHandle(dev->flink_fd, |
| 344 |
open_arg.handle); |
| 345 |
if (r) |
| 346 |
goto free_bo_handle; |
| 347 |
} |
| 348 |
open_arg.handle = 0; |
| 349 |
break; |
| 350 |
|
| 351 |
case amdgpu_bo_handle_type_dma_buf_fd: |
| 352 |
handle = shared_handle; |
| 353 |
alloc_size = dma_buf_size; |
| 354 |
break; |
| 355 |
|
| 356 |
case amdgpu_bo_handle_type_kms: |
| 357 |
case amdgpu_bo_handle_type_kms_noimport: |
| 358 |
assert(0); /* unreachable */ |
| 359 |
} |
| 360 |
|
| 361 |
/* Initialize it. */ |
| 362 |
r = amdgpu_bo_create(dev, alloc_size, handle, &bo); |
| 363 |
if (r) |
| 364 |
goto free_bo_handle; |
| 365 |
|
| 366 |
if (flink_name) { |
| 367 |
bo->flink_name = flink_name; |
| 368 |
r = handle_table_insert(&dev->bo_flink_names, flink_name, |
| 369 |
bo); |
| 370 |
if (r) |
| 371 |
goto free_bo_handle; |
| 372 |
|
| 373 |
} |
| 374 |
|
| 375 |
output->buf_handle = bo; |
| 376 |
output->alloc_size = bo->alloc_size; |
| 377 |
pthread_mutex_unlock(&dev->bo_table_mutex); |
| 378 |
return 0; |
| 379 |
|
| 380 |
free_bo_handle: |
| 381 |
if (flink_name && open_arg.handle) |
| 382 |
drmCloseBufferHandle(dev->flink_fd, open_arg.handle); |
| 383 |
|
| 384 |
if (bo) |
| 385 |
amdgpu_bo_free(bo); |
| 386 |
else |
| 387 |
drmCloseBufferHandle(dev->fd, handle); |
| 388 |
unlock: |
| 389 |
pthread_mutex_unlock(&dev->bo_table_mutex); |
| 390 |
return r; |
| 391 |
} |
| 392 |
|
| 393 |
drm_public int amdgpu_bo_free(amdgpu_bo_handle buf_handle) |
| 394 |
{ |
| 395 |
struct amdgpu_device *dev; |
| 396 |
struct amdgpu_bo *bo = buf_handle; |
| 397 |
|
| 398 |
assert(bo != NULL); |
| 399 |
dev = bo->dev; |
| 400 |
pthread_mutex_lock(&dev->bo_table_mutex); |
| 401 |
|
| 402 |
if (update_references(&bo->refcount, NULL)) { |
| 403 |
/* Remove the buffer from the hash tables. */ |
| 404 |
handle_table_remove(&dev->bo_handles, bo->handle); |
| 405 |
|
| 406 |
if (bo->flink_name) |
| 407 |
handle_table_remove(&dev->bo_flink_names, |
| 408 |
bo->flink_name); |
| 409 |
|
| 410 |
/* Release CPU access. */ |
| 411 |
if (bo->cpu_map_count > 0) { |
| 412 |
bo->cpu_map_count = 1; |
| 413 |
amdgpu_bo_cpu_unmap(bo); |
| 414 |
} |
| 415 |
|
| 416 |
drmCloseBufferHandle(dev->fd, bo->handle); |
| 417 |
pthread_mutex_destroy(&bo->cpu_access_mutex); |
| 418 |
free(bo); |
| 419 |
} |
| 420 |
|
| 421 |
pthread_mutex_unlock(&dev->bo_table_mutex); |
| 422 |
|
| 423 |
return 0; |
| 424 |
} |
| 425 |
|
| 426 |
drm_public void amdgpu_bo_inc_ref(amdgpu_bo_handle bo) |
| 427 |
{ |
| 428 |
atomic_inc(&bo->refcount); |
| 429 |
} |
| 430 |
|
| 431 |
drm_public int amdgpu_bo_cpu_map(amdgpu_bo_handle bo, void **cpu) |
| 432 |
{ |
| 433 |
union drm_amdgpu_gem_mmap args; |
| 434 |
void *ptr; |
| 435 |
int r; |
| 436 |
|
| 437 |
pthread_mutex_lock(&bo->cpu_access_mutex); |
| 438 |
|
| 439 |
if (bo->cpu_ptr) { |
| 440 |
/* already mapped */ |
| 441 |
assert(bo->cpu_map_count > 0); |
| 442 |
bo->cpu_map_count++; |
| 443 |
*cpu = bo->cpu_ptr; |
| 444 |
pthread_mutex_unlock(&bo->cpu_access_mutex); |
| 445 |
return 0; |
| 446 |
} |
| 447 |
|
| 448 |
assert(bo->cpu_map_count == 0); |
| 449 |
|
| 450 |
memset(&args, 0, sizeof(args)); |
| 451 |
|
| 452 |
/* Query the buffer address (args.addr_ptr). |
| 453 |
* The kernel driver ignores the offset and size parameters. */ |
| 454 |
args.in.handle = bo->handle; |
| 455 |
|
| 456 |
r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_MMAP, &args, |
| 457 |
sizeof(args)); |
| 458 |
if (r) { |
| 459 |
pthread_mutex_unlock(&bo->cpu_access_mutex); |
| 460 |
return r; |
| 461 |
} |
| 462 |
|
| 463 |
/* Map the buffer. */ |
| 464 |
ptr = drm_mmap(NULL, bo->alloc_size, PROT_READ | PROT_WRITE, MAP_SHARED, |
| 465 |
bo->dev->fd, args.out.addr_ptr); |
| 466 |
if (ptr == MAP_FAILED) { |
| 467 |
pthread_mutex_unlock(&bo->cpu_access_mutex); |
| 468 |
return -errno; |
| 469 |
} |
| 470 |
|
| 471 |
bo->cpu_ptr = ptr; |
| 472 |
bo->cpu_map_count = 1; |
| 473 |
pthread_mutex_unlock(&bo->cpu_access_mutex); |
| 474 |
|
| 475 |
*cpu = ptr; |
| 476 |
return 0; |
| 477 |
} |
| 478 |
|
| 479 |
drm_public int amdgpu_bo_cpu_unmap(amdgpu_bo_handle bo) |
| 480 |
{ |
| 481 |
int r; |
| 482 |
|
| 483 |
pthread_mutex_lock(&bo->cpu_access_mutex); |
| 484 |
assert(bo->cpu_map_count >= 0); |
| 485 |
|
| 486 |
if (bo->cpu_map_count == 0) { |
| 487 |
/* not mapped */ |
| 488 |
pthread_mutex_unlock(&bo->cpu_access_mutex); |
| 489 |
return -EINVAL; |
| 490 |
} |
| 491 |
|
| 492 |
bo->cpu_map_count--; |
| 493 |
if (bo->cpu_map_count > 0) { |
| 494 |
/* mapped multiple times */ |
| 495 |
pthread_mutex_unlock(&bo->cpu_access_mutex); |
| 496 |
return 0; |
| 497 |
} |
| 498 |
|
| 499 |
r = drm_munmap(bo->cpu_ptr, bo->alloc_size) == 0 ? 0 : -errno; |
| 500 |
bo->cpu_ptr = NULL; |
| 501 |
pthread_mutex_unlock(&bo->cpu_access_mutex); |
| 502 |
return r; |
| 503 |
} |
| 504 |
|
| 505 |
drm_public int amdgpu_query_buffer_size_alignment(amdgpu_device_handle dev, |
| 506 |
struct amdgpu_buffer_size_alignments *info) |
| 507 |
{ |
| 508 |
info->size_local = dev->dev_info.pte_fragment_size; |
| 509 |
info->size_remote = dev->dev_info.gart_page_size; |
| 510 |
return 0; |
| 511 |
} |
| 512 |
|
| 513 |
drm_public int amdgpu_bo_wait_for_idle(amdgpu_bo_handle bo, |
| 514 |
uint64_t timeout_ns, |
| 515 |
bool *busy) |
| 516 |
{ |
| 517 |
union drm_amdgpu_gem_wait_idle args; |
| 518 |
int r; |
| 519 |
|
| 520 |
memset(&args, 0, sizeof(args)); |
| 521 |
args.in.handle = bo->handle; |
| 522 |
args.in.timeout = amdgpu_cs_calculate_timeout(timeout_ns); |
| 523 |
|
| 524 |
r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_WAIT_IDLE, |
| 525 |
&args, sizeof(args)); |
| 526 |
|
| 527 |
if (r == 0) { |
| 528 |
*busy = args.out.status; |
| 529 |
return 0; |
| 530 |
} else { |
| 531 |
fprintf(stderr, "amdgpu: GEM_WAIT_IDLE failed with %i\n", r); |
| 532 |
return r; |
| 533 |
} |
| 534 |
} |
| 535 |
|
| 536 |
drm_public int amdgpu_find_bo_by_cpu_mapping(amdgpu_device_handle dev, |
| 537 |
void *cpu, |
| 538 |
uint64_t size, |
| 539 |
amdgpu_bo_handle *buf_handle, |
| 540 |
uint64_t *offset_in_bo) |
| 541 |
{ |
| 542 |
struct amdgpu_bo *bo = NULL; |
| 543 |
uint32_t i; |
| 544 |
int r = 0; |
| 545 |
|
| 546 |
if (cpu == NULL || size == 0) |
| 547 |
return -EINVAL; |
| 548 |
|
| 549 |
/* |
| 550 |
* Workaround for a buggy application which tries to import previously |
| 551 |
* exposed CPU pointers. If we find a real world use case we should |
| 552 |
* improve that by asking the kernel for the right handle. |
| 553 |
*/ |
| 554 |
pthread_mutex_lock(&dev->bo_table_mutex); |
| 555 |
for (i = 0; i < dev->bo_handles.max_key; i++) { |
| 556 |
bo = handle_table_lookup(&dev->bo_handles, i); |
| 557 |
if (!bo || !bo->cpu_ptr || size > bo->alloc_size) |
| 558 |
continue; |
| 559 |
if (cpu >= bo->cpu_ptr && |
| 560 |
cpu < (void*)((uintptr_t)bo->cpu_ptr + (size_t)bo->alloc_size)) |
| 561 |
break; |
| 562 |
} |
| 563 |
|
| 564 |
if (i < dev->bo_handles.max_key) { |
| 565 |
atomic_inc(&bo->refcount); |
| 566 |
*buf_handle = bo; |
| 567 |
*offset_in_bo = (uintptr_t)cpu - (uintptr_t)bo->cpu_ptr; |
| 568 |
} else { |
| 569 |
*buf_handle = NULL; |
| 570 |
*offset_in_bo = 0; |
| 571 |
r = -ENXIO; |
| 572 |
} |
| 573 |
pthread_mutex_unlock(&dev->bo_table_mutex); |
| 574 |
|
| 575 |
return r; |
| 576 |
} |
| 577 |
|
| 578 |
drm_public int amdgpu_create_bo_from_user_mem(amdgpu_device_handle dev, |
| 579 |
void *cpu, |
| 580 |
uint64_t size, |
| 581 |
amdgpu_bo_handle *buf_handle) |
| 582 |
{ |
| 583 |
int r; |
| 584 |
struct drm_amdgpu_gem_userptr args; |
| 585 |
|
| 586 |
args.addr = (uintptr_t)cpu; |
| 587 |
args.flags = AMDGPU_GEM_USERPTR_ANONONLY | AMDGPU_GEM_USERPTR_REGISTER | |
| 588 |
AMDGPU_GEM_USERPTR_VALIDATE; |
| 589 |
args.size = size; |
| 590 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_USERPTR, |
| 591 |
&args, sizeof(args)); |
| 592 |
if (r) |
| 593 |
goto out; |
| 594 |
|
| 595 |
pthread_mutex_lock(&dev->bo_table_mutex); |
| 596 |
r = amdgpu_bo_create(dev, size, args.handle, buf_handle); |
| 597 |
pthread_mutex_unlock(&dev->bo_table_mutex); |
| 598 |
if (r) { |
| 599 |
drmCloseBufferHandle(dev->fd, args.handle); |
| 600 |
} |
| 601 |
|
| 602 |
out: |
| 603 |
return r; |
| 604 |
} |
| 605 |
|
| 606 |
drm_public int amdgpu_bo_list_create_raw(amdgpu_device_handle dev, |
| 607 |
uint32_t number_of_buffers, |
| 608 |
struct drm_amdgpu_bo_list_entry *buffers, |
| 609 |
uint32_t *result) |
| 610 |
{ |
| 611 |
union drm_amdgpu_bo_list args; |
| 612 |
int r; |
| 613 |
|
| 614 |
memset(&args, 0, sizeof(args)); |
| 615 |
args.in.operation = AMDGPU_BO_LIST_OP_CREATE; |
| 616 |
args.in.bo_number = number_of_buffers; |
| 617 |
args.in.bo_info_size = sizeof(struct drm_amdgpu_bo_list_entry); |
| 618 |
args.in.bo_info_ptr = (uint64_t)(uintptr_t)buffers; |
| 619 |
|
| 620 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_BO_LIST, |
| 621 |
&args, sizeof(args)); |
| 622 |
if (!r) |
| 623 |
*result = args.out.list_handle; |
| 624 |
return r; |
| 625 |
} |
| 626 |
|
| 627 |
drm_public int amdgpu_bo_list_destroy_raw(amdgpu_device_handle dev, |
| 628 |
uint32_t bo_list) |
| 629 |
{ |
| 630 |
union drm_amdgpu_bo_list args; |
| 631 |
|
| 632 |
memset(&args, 0, sizeof(args)); |
| 633 |
args.in.operation = AMDGPU_BO_LIST_OP_DESTROY; |
| 634 |
args.in.list_handle = bo_list; |
| 635 |
|
| 636 |
return drmCommandWriteRead(dev->fd, DRM_AMDGPU_BO_LIST, |
| 637 |
&args, sizeof(args)); |
| 638 |
} |
| 639 |
|
| 640 |
drm_public int amdgpu_bo_list_create(amdgpu_device_handle dev, |
| 641 |
uint32_t number_of_resources, |
| 642 |
amdgpu_bo_handle *resources, |
| 643 |
uint8_t *resource_prios, |
| 644 |
amdgpu_bo_list_handle *result) |
| 645 |
{ |
| 646 |
struct drm_amdgpu_bo_list_entry *list; |
| 647 |
union drm_amdgpu_bo_list args; |
| 648 |
unsigned i; |
| 649 |
int r; |
| 650 |
|
| 651 |
if (!number_of_resources || !resources) |
| 652 |
return -EINVAL; |
| 653 |
|
| 654 |
/* overflow check for multiplication */ |
| 655 |
if (number_of_resources > UINT32_MAX / sizeof(struct drm_amdgpu_bo_list_entry)) |
| 656 |
return -EINVAL; |
| 657 |
|
| 658 |
list = malloc(number_of_resources * sizeof(struct drm_amdgpu_bo_list_entry)); |
| 659 |
if (!list) |
| 660 |
return -ENOMEM; |
| 661 |
|
| 662 |
*result = malloc(sizeof(struct amdgpu_bo_list)); |
| 663 |
if (!*result) { |
| 664 |
free(list); |
| 665 |
return -ENOMEM; |
| 666 |
} |
| 667 |
|
| 668 |
memset(&args, 0, sizeof(args)); |
| 669 |
args.in.operation = AMDGPU_BO_LIST_OP_CREATE; |
| 670 |
args.in.bo_number = number_of_resources; |
| 671 |
args.in.bo_info_size = sizeof(struct drm_amdgpu_bo_list_entry); |
| 672 |
args.in.bo_info_ptr = (uint64_t)(uintptr_t)list; |
| 673 |
|
| 674 |
for (i = 0; i < number_of_resources; i++) { |
| 675 |
list[i].bo_handle = resources[i]->handle; |
| 676 |
if (resource_prios) |
| 677 |
list[i].bo_priority = resource_prios[i]; |
| 678 |
else |
| 679 |
list[i].bo_priority = 0; |
| 680 |
} |
| 681 |
|
| 682 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_BO_LIST, |
| 683 |
&args, sizeof(args)); |
| 684 |
free(list); |
| 685 |
if (r) { |
| 686 |
free(*result); |
| 687 |
return r; |
| 688 |
} |
| 689 |
|
| 690 |
(*result)->dev = dev; |
| 691 |
(*result)->handle = args.out.list_handle; |
| 692 |
return 0; |
| 693 |
} |
| 694 |
|
| 695 |
drm_public int amdgpu_bo_list_destroy(amdgpu_bo_list_handle list) |
| 696 |
{ |
| 697 |
union drm_amdgpu_bo_list args; |
| 698 |
int r; |
| 699 |
|
| 700 |
memset(&args, 0, sizeof(args)); |
| 701 |
args.in.operation = AMDGPU_BO_LIST_OP_DESTROY; |
| 702 |
args.in.list_handle = list->handle; |
| 703 |
|
| 704 |
r = drmCommandWriteRead(list->dev->fd, DRM_AMDGPU_BO_LIST, |
| 705 |
&args, sizeof(args)); |
| 706 |
|
| 707 |
if (!r) |
| 708 |
free(list); |
| 709 |
|
| 710 |
return r; |
| 711 |
} |
| 712 |
|
| 713 |
drm_public int amdgpu_bo_list_update(amdgpu_bo_list_handle handle, |
| 714 |
uint32_t number_of_resources, |
| 715 |
amdgpu_bo_handle *resources, |
| 716 |
uint8_t *resource_prios) |
| 717 |
{ |
| 718 |
struct drm_amdgpu_bo_list_entry *list; |
| 719 |
union drm_amdgpu_bo_list args; |
| 720 |
unsigned i; |
| 721 |
int r; |
| 722 |
|
| 723 |
if (!number_of_resources) |
| 724 |
return -EINVAL; |
| 725 |
|
| 726 |
/* overflow check for multiplication */ |
| 727 |
if (number_of_resources > UINT32_MAX / sizeof(struct drm_amdgpu_bo_list_entry)) |
| 728 |
return -EINVAL; |
| 729 |
|
| 730 |
list = malloc(number_of_resources * sizeof(struct drm_amdgpu_bo_list_entry)); |
| 731 |
if (!list) |
| 732 |
return -ENOMEM; |
| 733 |
|
| 734 |
args.in.operation = AMDGPU_BO_LIST_OP_UPDATE; |
| 735 |
args.in.list_handle = handle->handle; |
| 736 |
args.in.bo_number = number_of_resources; |
| 737 |
args.in.bo_info_size = sizeof(struct drm_amdgpu_bo_list_entry); |
| 738 |
args.in.bo_info_ptr = (uintptr_t)list; |
| 739 |
|
| 740 |
for (i = 0; i < number_of_resources; i++) { |
| 741 |
list[i].bo_handle = resources[i]->handle; |
| 742 |
if (resource_prios) |
| 743 |
list[i].bo_priority = resource_prios[i]; |
| 744 |
else |
| 745 |
list[i].bo_priority = 0; |
| 746 |
} |
| 747 |
|
| 748 |
r = drmCommandWriteRead(handle->dev->fd, DRM_AMDGPU_BO_LIST, |
| 749 |
&args, sizeof(args)); |
| 750 |
free(list); |
| 751 |
return r; |
| 752 |
} |
| 753 |
|
| 754 |
drm_public int amdgpu_bo_va_op(amdgpu_bo_handle bo, |
| 755 |
uint64_t offset, |
| 756 |
uint64_t size, |
| 757 |
uint64_t addr, |
| 758 |
uint64_t flags, |
| 759 |
uint32_t ops) |
| 760 |
{ |
| 761 |
amdgpu_device_handle dev = bo->dev; |
| 762 |
|
| 763 |
size = ALIGN(size, getpagesize()); |
| 764 |
|
| 765 |
return amdgpu_bo_va_op_raw(dev, bo, offset, size, addr, |
| 766 |
AMDGPU_VM_PAGE_READABLE | |
| 767 |
AMDGPU_VM_PAGE_WRITEABLE | |
| 768 |
AMDGPU_VM_PAGE_EXECUTABLE, ops); |
| 769 |
} |
| 770 |
|
| 771 |
drm_public int amdgpu_bo_va_op_raw(amdgpu_device_handle dev, |
| 772 |
amdgpu_bo_handle bo, |
| 773 |
uint64_t offset, |
| 774 |
uint64_t size, |
| 775 |
uint64_t addr, |
| 776 |
uint64_t flags, |
| 777 |
uint32_t ops) |
| 778 |
{ |
| 779 |
struct drm_amdgpu_gem_va va; |
| 780 |
int r; |
| 781 |
|
| 782 |
if (ops != AMDGPU_VA_OP_MAP && ops != AMDGPU_VA_OP_UNMAP && |
| 783 |
ops != AMDGPU_VA_OP_REPLACE && ops != AMDGPU_VA_OP_CLEAR) |
| 784 |
return -EINVAL; |
| 785 |
|
| 786 |
memset(&va, 0, sizeof(va)); |
| 787 |
va.handle = bo ? bo->handle : 0; |
| 788 |
va.operation = ops; |
| 789 |
va.flags = flags; |
| 790 |
va.va_address = addr; |
| 791 |
va.offset_in_bo = offset; |
| 792 |
va.map_size = size; |
| 793 |
|
| 794 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_VA, &va, sizeof(va)); |
| 795 |
|
| 796 |
return r; |
| 797 |
} |
| 798 |
|
| 799 |
drm_public int amdgpu_bo_va_op_raw2(amdgpu_device_handle dev, |
| 800 |
amdgpu_bo_handle bo, |
| 801 |
uint64_t offset, |
| 802 |
uint64_t size, |
| 803 |
uint64_t addr, |
| 804 |
uint64_t flags, |
| 805 |
uint32_t ops, |
| 806 |
uint32_t vm_timeline_syncobj_out, |
| 807 |
uint64_t vm_timeline_point, |
| 808 |
uint64_t input_fence_syncobj_handles, |
| 809 |
uint32_t num_syncobj_handles) |
| 810 |
{ |
| 811 |
struct drm_amdgpu_gem_va va; |
| 812 |
int r; |
| 813 |
|
| 814 |
if (ops != AMDGPU_VA_OP_MAP && ops != AMDGPU_VA_OP_UNMAP && |
| 815 |
ops != AMDGPU_VA_OP_REPLACE && ops != AMDGPU_VA_OP_CLEAR) |
| 816 |
return -EINVAL; |
| 817 |
|
| 818 |
memset(&va, 0, sizeof(va)); |
| 819 |
va.handle = bo ? bo->handle : 0; |
| 820 |
va.operation = ops; |
| 821 |
va.flags = flags; |
| 822 |
va.va_address = addr; |
| 823 |
va.offset_in_bo = offset; |
| 824 |
va.map_size = size; |
| 825 |
va.vm_timeline_syncobj_out = vm_timeline_syncobj_out; |
| 826 |
va.vm_timeline_point = vm_timeline_point; |
| 827 |
va.input_fence_syncobj_handles = input_fence_syncobj_handles; |
| 828 |
va.num_syncobj_handles = num_syncobj_handles; |
| 829 |
|
| 830 |
r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_VA, &va, sizeof(va)); |
| 831 |
|
| 832 |
return r; |
| 833 |
} |