2 * Copyright © 2014 Advanced Micro Devices, Inc.
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:
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
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.
36 #include <sys/ioctl.h>
40 #include "libdrm_macros.h"
42 #include "amdgpu_drm.h"
43 #include "amdgpu_internal.h"
44 #include "util_hash_table.h"
45 #include "util_math.h"
47 static void amdgpu_close_kms_handle(amdgpu_device_handle dev,
50 struct drm_gem_close args = {};
53 drmIoctl(dev->fd, DRM_IOCTL_GEM_CLOSE, &args);
56 drm_private void amdgpu_bo_free_internal(amdgpu_bo_handle bo)
58 /* Remove the buffer from the hash tables. */
59 pthread_mutex_lock(&bo->dev->bo_table_mutex);
60 util_hash_table_remove(bo->dev->bo_handles,
61 (void*)(uintptr_t)bo->handle);
63 util_hash_table_remove(bo->dev->bo_flink_names,
64 (void*)(uintptr_t)bo->flink_name);
66 pthread_mutex_unlock(&bo->dev->bo_table_mutex);
68 /* Release CPU access. */
69 if (bo->cpu_map_count > 0) {
70 bo->cpu_map_count = 1;
71 amdgpu_bo_cpu_unmap(bo);
74 amdgpu_close_kms_handle(bo->dev, bo->handle);
75 pthread_mutex_destroy(&bo->cpu_access_mutex);
79 int amdgpu_bo_alloc(amdgpu_device_handle dev,
80 struct amdgpu_bo_alloc_request *alloc_buffer,
81 amdgpu_bo_handle *buf_handle)
84 union drm_amdgpu_gem_create args;
85 unsigned heap = alloc_buffer->preferred_heap;
88 /* It's an error if the heap is not specified */
89 if (!(heap & (AMDGPU_GEM_DOMAIN_GTT | AMDGPU_GEM_DOMAIN_VRAM)))
92 bo = calloc(1, sizeof(struct amdgpu_bo));
96 atomic_set(&bo->refcount, 1);
98 bo->alloc_size = alloc_buffer->alloc_size;
100 memset(&args, 0, sizeof(args));
101 args.in.bo_size = alloc_buffer->alloc_size;
102 args.in.alignment = alloc_buffer->phys_alignment;
104 /* Set the placement. */
105 args.in.domains = heap;
106 args.in.domain_flags = alloc_buffer->flags;
108 /* Allocate the buffer with the preferred heap. */
109 r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_CREATE,
110 &args, sizeof(args));
116 bo->handle = args.out.handle;
118 pthread_mutex_init(&bo->cpu_access_mutex, NULL);
124 int amdgpu_bo_set_metadata(amdgpu_bo_handle bo,
125 struct amdgpu_bo_metadata *info)
127 struct drm_amdgpu_gem_metadata args = {};
129 args.handle = bo->handle;
130 args.op = AMDGPU_GEM_METADATA_OP_SET_METADATA;
131 args.data.flags = info->flags;
132 args.data.tiling_info = info->tiling_info;
134 if (info->size_metadata > sizeof(args.data.data))
137 if (info->size_metadata) {
138 args.data.data_size_bytes = info->size_metadata;
139 memcpy(args.data.data, info->umd_metadata, info->size_metadata);
142 return drmCommandWriteRead(bo->dev->fd,
143 DRM_AMDGPU_GEM_METADATA,
144 &args, sizeof(args));
147 int amdgpu_bo_query_info(amdgpu_bo_handle bo,
148 struct amdgpu_bo_info *info)
150 struct drm_amdgpu_gem_metadata metadata = {};
151 struct drm_amdgpu_gem_create_in bo_info = {};
152 struct drm_amdgpu_gem_op gem_op = {};
155 /* Validate the BO passed in */
159 /* Query metadata. */
160 metadata.handle = bo->handle;
161 metadata.op = AMDGPU_GEM_METADATA_OP_GET_METADATA;
163 r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_METADATA,
164 &metadata, sizeof(metadata));
168 if (metadata.data.data_size_bytes >
169 sizeof(info->metadata.umd_metadata))
172 /* Query buffer info. */
173 gem_op.handle = bo->handle;
174 gem_op.op = AMDGPU_GEM_OP_GET_GEM_CREATE_INFO;
175 gem_op.value = (uintptr_t)&bo_info;
177 r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_OP,
178 &gem_op, sizeof(gem_op));
182 memset(info, 0, sizeof(*info));
183 info->alloc_size = bo_info.bo_size;
184 info->phys_alignment = bo_info.alignment;
185 info->preferred_heap = bo_info.domains;
186 info->alloc_flags = bo_info.domain_flags;
187 info->metadata.flags = metadata.data.flags;
188 info->metadata.tiling_info = metadata.data.tiling_info;
190 info->metadata.size_metadata = metadata.data.data_size_bytes;
191 if (metadata.data.data_size_bytes > 0)
192 memcpy(info->metadata.umd_metadata, metadata.data.data,
193 metadata.data.data_size_bytes);
198 static void amdgpu_add_handle_to_table(amdgpu_bo_handle bo)
200 pthread_mutex_lock(&bo->dev->bo_table_mutex);
201 util_hash_table_set(bo->dev->bo_handles,
202 (void*)(uintptr_t)bo->handle, bo);
203 pthread_mutex_unlock(&bo->dev->bo_table_mutex);
206 static int amdgpu_bo_export_flink(amdgpu_bo_handle bo)
208 struct drm_gem_flink flink;
219 if (bo->dev->flink_fd != bo->dev->fd) {
220 r = drmPrimeHandleToFD(bo->dev->fd, bo->handle, DRM_CLOEXEC,
223 r = drmPrimeFDToHandle(bo->dev->flink_fd, dma_fd, &handle);
228 fd = bo->dev->flink_fd;
230 memset(&flink, 0, sizeof(flink));
231 flink.handle = handle;
233 r = drmIoctl(fd, DRM_IOCTL_GEM_FLINK, &flink);
237 bo->flink_name = flink.name;
239 if (bo->dev->flink_fd != bo->dev->fd) {
240 struct drm_gem_close args = {};
241 args.handle = handle;
242 drmIoctl(bo->dev->flink_fd, DRM_IOCTL_GEM_CLOSE, &args);
245 pthread_mutex_lock(&bo->dev->bo_table_mutex);
246 util_hash_table_set(bo->dev->bo_flink_names,
247 (void*)(uintptr_t)bo->flink_name,
249 pthread_mutex_unlock(&bo->dev->bo_table_mutex);
254 int amdgpu_bo_export(amdgpu_bo_handle bo,
255 enum amdgpu_bo_handle_type type,
256 uint32_t *shared_handle)
261 case amdgpu_bo_handle_type_gem_flink_name:
262 r = amdgpu_bo_export_flink(bo);
266 *shared_handle = bo->flink_name;
269 case amdgpu_bo_handle_type_kms:
270 amdgpu_add_handle_to_table(bo);
271 *shared_handle = bo->handle;
274 case amdgpu_bo_handle_type_dma_buf_fd:
275 amdgpu_add_handle_to_table(bo);
276 return drmPrimeHandleToFD(bo->dev->fd, bo->handle, DRM_CLOEXEC,
277 (int*)shared_handle);
282 int amdgpu_bo_import(amdgpu_device_handle dev,
283 enum amdgpu_bo_handle_type type,
284 uint32_t shared_handle,
285 struct amdgpu_bo_import_result *output)
287 struct drm_gem_open open_arg = {};
288 struct amdgpu_bo *bo = NULL;
291 uint64_t dma_buf_size = 0;
293 /* Convert a DMA buf handle to a KMS handle now. */
294 if (type == amdgpu_bo_handle_type_dma_buf_fd) {
298 /* Get a KMS handle. */
299 r = drmPrimeFDToHandle(dev->fd, shared_handle, &handle);
304 /* Query the buffer size. */
305 size = lseek(shared_handle, 0, SEEK_END);
306 if (size == (off_t)-1) {
307 amdgpu_close_kms_handle(dev, handle);
310 lseek(shared_handle, 0, SEEK_SET);
313 shared_handle = handle;
316 /* We must maintain a list of pairs <handle, bo>, so that we always
317 * return the same amdgpu_bo instance for the same handle. */
318 pthread_mutex_lock(&dev->bo_table_mutex);
320 /* If we have already created a buffer with this handle, find it. */
322 case amdgpu_bo_handle_type_gem_flink_name:
323 bo = util_hash_table_get(dev->bo_flink_names,
324 (void*)(uintptr_t)shared_handle);
327 case amdgpu_bo_handle_type_dma_buf_fd:
328 bo = util_hash_table_get(dev->bo_handles,
329 (void*)(uintptr_t)shared_handle);
332 case amdgpu_bo_handle_type_kms:
333 /* Importing a KMS handle in not allowed. */
334 pthread_mutex_unlock(&dev->bo_table_mutex);
338 pthread_mutex_unlock(&dev->bo_table_mutex);
343 pthread_mutex_unlock(&dev->bo_table_mutex);
345 /* The buffer already exists, just bump the refcount. */
346 atomic_inc(&bo->refcount);
348 output->buf_handle = bo;
349 output->alloc_size = bo->alloc_size;
353 bo = calloc(1, sizeof(struct amdgpu_bo));
355 pthread_mutex_unlock(&dev->bo_table_mutex);
356 if (type == amdgpu_bo_handle_type_dma_buf_fd) {
357 amdgpu_close_kms_handle(dev, shared_handle);
362 /* Open the handle. */
364 case amdgpu_bo_handle_type_gem_flink_name:
365 open_arg.name = shared_handle;
366 r = drmIoctl(dev->flink_fd, DRM_IOCTL_GEM_OPEN, &open_arg);
369 pthread_mutex_unlock(&dev->bo_table_mutex);
373 bo->handle = open_arg.handle;
374 if (dev->flink_fd != dev->fd) {
375 r = drmPrimeHandleToFD(dev->flink_fd, bo->handle, DRM_CLOEXEC, &dma_fd);
378 pthread_mutex_unlock(&dev->bo_table_mutex);
381 r = drmPrimeFDToHandle(dev->fd, dma_fd, &bo->handle );
387 pthread_mutex_unlock(&dev->bo_table_mutex);
391 bo->flink_name = shared_handle;
392 bo->alloc_size = open_arg.size;
393 util_hash_table_set(dev->bo_flink_names,
394 (void*)(uintptr_t)bo->flink_name, bo);
397 case amdgpu_bo_handle_type_dma_buf_fd:
398 bo->handle = shared_handle;
399 bo->alloc_size = dma_buf_size;
402 case amdgpu_bo_handle_type_kms:
403 assert(0); /* unreachable */
407 atomic_set(&bo->refcount, 1);
409 pthread_mutex_init(&bo->cpu_access_mutex, NULL);
411 util_hash_table_set(dev->bo_handles, (void*)(uintptr_t)bo->handle, bo);
412 pthread_mutex_unlock(&dev->bo_table_mutex);
414 output->buf_handle = bo;
415 output->alloc_size = bo->alloc_size;
419 int amdgpu_bo_free(amdgpu_bo_handle buf_handle)
421 /* Just drop the reference. */
422 amdgpu_bo_reference(&buf_handle, NULL);
426 int amdgpu_bo_cpu_map(amdgpu_bo_handle bo, void **cpu)
428 union drm_amdgpu_gem_mmap args;
432 pthread_mutex_lock(&bo->cpu_access_mutex);
436 assert(bo->cpu_map_count > 0);
439 pthread_mutex_unlock(&bo->cpu_access_mutex);
443 assert(bo->cpu_map_count == 0);
445 memset(&args, 0, sizeof(args));
447 /* Query the buffer address (args.addr_ptr).
448 * The kernel driver ignores the offset and size parameters. */
449 args.in.handle = bo->handle;
451 r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_MMAP, &args,
454 pthread_mutex_unlock(&bo->cpu_access_mutex);
458 /* Map the buffer. */
459 ptr = drm_mmap(NULL, bo->alloc_size, PROT_READ | PROT_WRITE, MAP_SHARED,
460 bo->dev->fd, args.out.addr_ptr);
461 if (ptr == MAP_FAILED) {
462 pthread_mutex_unlock(&bo->cpu_access_mutex);
467 bo->cpu_map_count = 1;
468 pthread_mutex_unlock(&bo->cpu_access_mutex);
474 int amdgpu_bo_cpu_unmap(amdgpu_bo_handle bo)
478 pthread_mutex_lock(&bo->cpu_access_mutex);
479 assert(bo->cpu_map_count >= 0);
481 if (bo->cpu_map_count == 0) {
483 pthread_mutex_unlock(&bo->cpu_access_mutex);
488 if (bo->cpu_map_count > 0) {
489 /* mapped multiple times */
490 pthread_mutex_unlock(&bo->cpu_access_mutex);
494 r = drm_munmap(bo->cpu_ptr, bo->alloc_size) == 0 ? 0 : -errno;
496 pthread_mutex_unlock(&bo->cpu_access_mutex);
500 int amdgpu_query_buffer_size_alignment(amdgpu_device_handle dev,
501 struct amdgpu_buffer_size_alignments *info)
503 info->size_local = dev->dev_info.pte_fragment_size;
504 info->size_remote = dev->dev_info.gart_page_size;
508 int amdgpu_bo_wait_for_idle(amdgpu_bo_handle bo,
512 union drm_amdgpu_gem_wait_idle args;
515 memset(&args, 0, sizeof(args));
516 args.in.handle = bo->handle;
517 args.in.timeout = amdgpu_cs_calculate_timeout(timeout_ns);
519 r = drmCommandWriteRead(bo->dev->fd, DRM_AMDGPU_GEM_WAIT_IDLE,
520 &args, sizeof(args));
523 *busy = args.out.status;
526 fprintf(stderr, "amdgpu: GEM_WAIT_IDLE failed with %i\n", r);
531 int amdgpu_create_bo_from_user_mem(amdgpu_device_handle dev,
534 amdgpu_bo_handle *buf_handle)
537 struct amdgpu_bo *bo;
538 struct drm_amdgpu_gem_userptr args;
542 memset(&args, 0, sizeof(args));
545 cpu0 = ROUND_DOWN((uintptr_t)cpu, ps);
546 off = (uintptr_t)cpu - cpu0;
547 size = ROUND_UP(size + off, ps);
550 args.flags = AMDGPU_GEM_USERPTR_ANONONLY | AMDGPU_GEM_USERPTR_REGISTER;
552 r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_USERPTR,
553 &args, sizeof(args));
557 bo = calloc(1, sizeof(struct amdgpu_bo));
561 atomic_set(&bo->refcount, 1);
563 bo->alloc_size = size;
564 bo->handle = args.handle;
571 int amdgpu_bo_list_create(amdgpu_device_handle dev,
572 uint32_t number_of_resources,
573 amdgpu_bo_handle *resources,
574 uint8_t *resource_prios,
575 amdgpu_bo_list_handle *result)
577 struct drm_amdgpu_bo_list_entry *list;
578 union drm_amdgpu_bo_list args;
582 if (!number_of_resources)
585 /* overflow check for multiplication */
586 if (number_of_resources > UINT32_MAX / sizeof(struct drm_amdgpu_bo_list_entry))
589 list = malloc(number_of_resources * sizeof(struct drm_amdgpu_bo_list_entry));
593 memset(&args, 0, sizeof(args));
594 args.in.operation = AMDGPU_BO_LIST_OP_CREATE;
595 args.in.bo_number = number_of_resources;
596 args.in.bo_info_size = sizeof(struct drm_amdgpu_bo_list_entry);
597 args.in.bo_info_ptr = (uint64_t)(uintptr_t)list;
599 for (i = 0; i < number_of_resources; i++) {
600 list[i].bo_handle = resources[i]->handle;
602 list[i].bo_priority = resource_prios[i];
604 list[i].bo_priority = 0;
607 r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_BO_LIST,
608 &args, sizeof(args));
613 *result = malloc(sizeof(struct amdgpu_bo_list));
614 (*result)->dev = dev;
615 (*result)->handle = args.out.list_handle;
619 int amdgpu_bo_list_destroy(amdgpu_bo_list_handle list)
621 union drm_amdgpu_bo_list args;
624 memset(&args, 0, sizeof(args));
625 args.in.operation = AMDGPU_BO_LIST_OP_DESTROY;
626 args.in.list_handle = list->handle;
628 r = drmCommandWriteRead(list->dev->fd, DRM_AMDGPU_BO_LIST,
629 &args, sizeof(args));
637 int amdgpu_bo_list_update(amdgpu_bo_list_handle handle,
638 uint32_t number_of_resources,
639 amdgpu_bo_handle *resources,
640 uint8_t *resource_prios)
642 struct drm_amdgpu_bo_list_entry *list;
643 union drm_amdgpu_bo_list args;
647 if (!number_of_resources)
650 /* overflow check for multiplication */
651 if (number_of_resources > UINT32_MAX / sizeof(struct drm_amdgpu_bo_list_entry))
654 list = malloc(number_of_resources * sizeof(struct drm_amdgpu_bo_list_entry));
658 args.in.operation = AMDGPU_BO_LIST_OP_UPDATE;
659 args.in.list_handle = handle->handle;
660 args.in.bo_number = number_of_resources;
661 args.in.bo_info_size = sizeof(struct drm_amdgpu_bo_list_entry);
662 args.in.bo_info_ptr = (uintptr_t)list;
664 for (i = 0; i < number_of_resources; i++) {
665 list[i].bo_handle = resources[i]->handle;
667 list[i].bo_priority = resource_prios[i];
669 list[i].bo_priority = 0;
672 r = drmCommandWriteRead(handle->dev->fd, DRM_AMDGPU_BO_LIST,
673 &args, sizeof(args));
678 int amdgpu_bo_va_op(amdgpu_bo_handle bo,
685 amdgpu_device_handle dev = bo->dev;
686 struct drm_amdgpu_gem_va va;
689 if (ops != AMDGPU_VA_OP_MAP && ops != AMDGPU_VA_OP_UNMAP)
692 memset(&va, 0, sizeof(va));
693 va.handle = bo->handle;
695 va.flags = AMDGPU_VM_PAGE_READABLE |
696 AMDGPU_VM_PAGE_WRITEABLE |
697 AMDGPU_VM_PAGE_EXECUTABLE;
698 va.va_address = addr;
699 va.offset_in_bo = offset;
700 va.map_size = ALIGN(size, getpagesize());
702 r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_VA, &va, sizeof(va));