2 * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
5 * Permission is hereby granted, free of charge, to any person obtaining
6 * a copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
14 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
15 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
16 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
17 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
20 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 * The above copyright notice and this permission notice (including the
23 * next paragraph) shall be included in all copies or substantial portions
27 #define _FILE_OFFSET_BITS 64
28 #include "radeon_drm_cs.h"
30 #include "util/u_hash_table.h"
31 #include "util/u_memory.h"
32 #include "util/u_simple_list.h"
33 #include "os/os_thread.h"
35 #include "state_tracker/drm_driver.h"
37 #include <sys/ioctl.h>
42 #define RADEON_BO_FLAGS_MACRO_TILE 1
43 #define RADEON_BO_FLAGS_MICRO_TILE 2
44 #define RADEON_BO_FLAGS_MICRO_TILE_SQUARE 0x20
46 extern const struct pb_vtbl radeon_bo_vtbl
;
49 static INLINE
struct radeon_bo
*radeon_bo(struct pb_buffer
*bo
)
51 assert(bo
->vtbl
== &radeon_bo_vtbl
);
52 return (struct radeon_bo
*)bo
;
57 struct pb_manager base
;
60 struct radeon_drm_winsys
*rws
;
62 /* List of buffer handles and its mutex. */
63 struct util_hash_table
*bo_handles
;
64 pipe_mutex bo_handles_mutex
;
67 static INLINE
struct radeon_bomgr
*radeon_bomgr(struct pb_manager
*mgr
)
69 return (struct radeon_bomgr
*)mgr
;
72 static struct radeon_bo
*get_radeon_bo(struct pb_buffer
*_buf
)
74 struct radeon_bo
*bo
= NULL
;
76 if (_buf
->vtbl
== &radeon_bo_vtbl
) {
79 struct pb_buffer
*base_buf
;
81 pb_get_base_buffer(_buf
, &base_buf
, &offset
);
83 if (base_buf
->vtbl
== &radeon_bo_vtbl
)
84 bo
= radeon_bo(base_buf
);
90 static void radeon_bo_wait(struct pb_buffer
*_buf
)
92 struct radeon_bo
*bo
= get_radeon_bo(_buf
);
93 struct drm_radeon_gem_wait_idle args
= {};
95 while (p_atomic_read(&bo
->num_active_ioctls
)) {
99 args
.handle
= bo
->handle
;
100 while (drmCommandWriteRead(bo
->rws
->fd
, DRM_RADEON_GEM_WAIT_IDLE
,
101 &args
, sizeof(args
)) == -EBUSY
);
104 static boolean
radeon_bo_is_busy(struct pb_buffer
*_buf
)
106 struct radeon_bo
*bo
= get_radeon_bo(_buf
);
107 struct drm_radeon_gem_busy args
= {};
110 if (p_atomic_read(&bo
->num_active_ioctls
)) {
114 args
.handle
= bo
->handle
;
115 busy
= drmCommandWriteRead(bo
->rws
->fd
, DRM_RADEON_GEM_BUSY
,
116 &args
, sizeof(args
)) != 0;
121 static void radeon_bo_destroy(struct pb_buffer
*_buf
)
123 struct radeon_bo
*bo
= radeon_bo(_buf
);
124 struct drm_gem_close args
= {};
127 pipe_mutex_lock(bo
->mgr
->bo_handles_mutex
);
128 util_hash_table_remove(bo
->mgr
->bo_handles
,
129 (void*)(uintptr_t)bo
->name
);
130 pipe_mutex_unlock(bo
->mgr
->bo_handles_mutex
);
134 munmap(bo
->ptr
, bo
->size
);
137 args
.handle
= bo
->handle
;
138 drmIoctl(bo
->rws
->fd
, DRM_IOCTL_GEM_CLOSE
, &args
);
139 pipe_mutex_destroy(bo
->map_mutex
);
143 static unsigned get_pb_usage_from_transfer_flags(enum pipe_transfer_usage usage
)
147 if (usage
& PIPE_TRANSFER_WRITE
)
148 res
|= PB_USAGE_CPU_WRITE
;
150 if (usage
& PIPE_TRANSFER_DONTBLOCK
)
151 res
|= PB_USAGE_DONTBLOCK
;
153 if (usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)
154 res
|= PB_USAGE_UNSYNCHRONIZED
;
159 static void *radeon_bo_map_internal(struct pb_buffer
*_buf
,
160 unsigned flags
, void *flush_ctx
)
162 struct radeon_bo
*bo
= radeon_bo(_buf
);
163 struct radeon_drm_cs
*cs
= flush_ctx
;
164 struct drm_radeon_gem_mmap args
= {};
167 /* If it's not unsynchronized bo_map, flush CS if needed and then wait. */
168 if (!(flags
& PB_USAGE_UNSYNCHRONIZED
)) {
169 /* DONTBLOCK doesn't make sense with UNSYNCHRONIZED. */
170 if (flags
& PB_USAGE_DONTBLOCK
) {
171 if (radeon_bo_is_referenced_by_cs(cs
, bo
)) {
172 cs
->flush_cs(cs
->flush_data
, RADEON_FLUSH_ASYNC
);
176 if (radeon_bo_is_busy((struct pb_buffer
*)bo
)) {
180 if (!(flags
& PB_USAGE_CPU_WRITE
)) {
183 * Since we are mapping for read, we don't need to wait
184 * if the GPU is using the buffer for read too
185 * (neither one is changing it).
187 * Only check whether the buffer is being used for write. */
188 if (radeon_bo_is_referenced_by_cs_for_write(cs
, bo
)) {
189 cs
->flush_cs(cs
->flush_data
, 0);
190 radeon_bo_wait((struct pb_buffer
*)bo
);
192 /* XXX We could check whether the buffer is busy for write here. */
193 radeon_bo_wait((struct pb_buffer
*)bo
);
196 /* Mapping for write. */
197 if (radeon_bo_is_referenced_by_cs(cs
, bo
)) {
198 cs
->flush_cs(cs
->flush_data
, 0);
200 /* Try to avoid busy-waiting in radeon_bo_wait. */
201 if (p_atomic_read(&bo
->num_active_ioctls
))
202 radeon_drm_cs_sync_flush(cs
);
205 radeon_bo_wait((struct pb_buffer
*)bo
);
210 /* Return the pointer if it's already mapped. */
214 /* Map the buffer. */
215 pipe_mutex_lock(bo
->map_mutex
);
216 /* Return the pointer if it's already mapped (in case of a race). */
218 pipe_mutex_unlock(bo
->map_mutex
);
221 args
.handle
= bo
->handle
;
223 args
.size
= (uint64_t)bo
->size
;
224 if (drmCommandWriteRead(bo
->rws
->fd
,
228 pipe_mutex_unlock(bo
->map_mutex
);
229 fprintf(stderr
, "radeon: gem_mmap failed: %p 0x%08X\n",
234 ptr
= mmap(0, args
.size
, PROT_READ
|PROT_WRITE
, MAP_SHARED
,
235 bo
->rws
->fd
, args
.addr_ptr
);
236 if (ptr
== MAP_FAILED
) {
237 pipe_mutex_unlock(bo
->map_mutex
);
238 fprintf(stderr
, "radeon: mmap failed, errno: %i\n", errno
);
242 pipe_mutex_unlock(bo
->map_mutex
);
247 static void radeon_bo_unmap_internal(struct pb_buffer
*_buf
)
252 static void radeon_bo_get_base_buffer(struct pb_buffer
*buf
,
253 struct pb_buffer
**base_buf
,
260 static enum pipe_error
radeon_bo_validate(struct pb_buffer
*_buf
,
261 struct pb_validate
*vl
,
268 static void radeon_bo_fence(struct pb_buffer
*buf
,
269 struct pipe_fence_handle
*fence
)
273 const struct pb_vtbl radeon_bo_vtbl
= {
275 radeon_bo_map_internal
,
276 radeon_bo_unmap_internal
,
279 radeon_bo_get_base_buffer
,
282 static struct pb_buffer
*radeon_bomgr_create_bo(struct pb_manager
*_mgr
,
284 const struct pb_desc
*desc
)
286 struct radeon_bomgr
*mgr
= radeon_bomgr(_mgr
);
287 struct radeon_drm_winsys
*rws
= mgr
->rws
;
288 struct radeon_bo
*bo
;
289 struct drm_radeon_gem_create args
= {};
292 args
.alignment
= desc
->alignment
;
293 args
.initial_domain
=
294 (desc
->usage
& RADEON_PB_USAGE_DOMAIN_GTT
?
295 RADEON_GEM_DOMAIN_GTT
: 0) |
296 (desc
->usage
& RADEON_PB_USAGE_DOMAIN_VRAM
?
297 RADEON_GEM_DOMAIN_VRAM
: 0);
299 if (drmCommandWriteRead(rws
->fd
, DRM_RADEON_GEM_CREATE
,
300 &args
, sizeof(args
))) {
301 fprintf(stderr
, "radeon: Failed to allocate a buffer:\n");
302 fprintf(stderr
, "radeon: size : %d bytes\n", size
);
303 fprintf(stderr
, "radeon: alignment : %d bytes\n", desc
->alignment
);
304 fprintf(stderr
, "radeon: domains : %d\n", args
.initial_domain
);
308 bo
= CALLOC_STRUCT(radeon_bo
);
312 pipe_reference_init(&bo
->base
.base
.reference
, 1);
313 bo
->base
.base
.alignment
= desc
->alignment
;
314 bo
->base
.base
.usage
= desc
->usage
;
315 bo
->base
.base
.size
= size
;
316 bo
->base
.vtbl
= &radeon_bo_vtbl
;
319 bo
->handle
= args
.handle
;
321 pipe_mutex_init(bo
->map_mutex
);
326 static void radeon_bomgr_flush(struct pb_manager
*mgr
)
331 /* This is for the cache bufmgr. */
332 static boolean
radeon_bomgr_is_buffer_busy(struct pb_manager
*_mgr
,
333 struct pb_buffer
*_buf
)
335 struct radeon_bo
*bo
= radeon_bo(_buf
);
337 if (radeon_bo_is_referenced_by_any_cs(bo
)) {
341 if (radeon_bo_is_busy((struct pb_buffer
*)bo
)) {
348 static void radeon_bomgr_destroy(struct pb_manager
*_mgr
)
350 struct radeon_bomgr
*mgr
= radeon_bomgr(_mgr
);
351 util_hash_table_destroy(mgr
->bo_handles
);
352 pipe_mutex_destroy(mgr
->bo_handles_mutex
);
356 #define PTR_TO_UINT(x) ((unsigned)((intptr_t)(x)))
358 static unsigned handle_hash(void *key
)
360 return PTR_TO_UINT(key
);
363 static int handle_compare(void *key1
, void *key2
)
365 return PTR_TO_UINT(key1
) != PTR_TO_UINT(key2
);
368 struct pb_manager
*radeon_bomgr_create(struct radeon_drm_winsys
*rws
)
370 struct radeon_bomgr
*mgr
;
372 mgr
= CALLOC_STRUCT(radeon_bomgr
);
376 mgr
->base
.destroy
= radeon_bomgr_destroy
;
377 mgr
->base
.create_buffer
= radeon_bomgr_create_bo
;
378 mgr
->base
.flush
= radeon_bomgr_flush
;
379 mgr
->base
.is_buffer_busy
= radeon_bomgr_is_buffer_busy
;
382 mgr
->bo_handles
= util_hash_table_create(handle_hash
, handle_compare
);
383 pipe_mutex_init(mgr
->bo_handles_mutex
);
387 static void *radeon_bo_map(struct pb_buffer
*buf
,
388 struct radeon_winsys_cs
*cs
,
389 enum pipe_transfer_usage usage
)
391 return pb_map(buf
, get_pb_usage_from_transfer_flags(usage
), cs
);
394 static void radeon_bo_get_tiling(struct pb_buffer
*_buf
,
395 enum radeon_bo_layout
*microtiled
,
396 enum radeon_bo_layout
*macrotiled
)
398 struct radeon_bo
*bo
= get_radeon_bo(_buf
);
399 struct drm_radeon_gem_set_tiling args
= {};
401 args
.handle
= bo
->handle
;
403 drmCommandWriteRead(bo
->rws
->fd
,
404 DRM_RADEON_GEM_GET_TILING
,
408 *microtiled
= RADEON_LAYOUT_LINEAR
;
409 *macrotiled
= RADEON_LAYOUT_LINEAR
;
410 if (args
.tiling_flags
& RADEON_BO_FLAGS_MICRO_TILE
)
411 *microtiled
= RADEON_LAYOUT_TILED
;
413 if (args
.tiling_flags
& RADEON_BO_FLAGS_MACRO_TILE
)
414 *macrotiled
= RADEON_LAYOUT_TILED
;
417 static void radeon_bo_set_tiling(struct pb_buffer
*_buf
,
418 struct radeon_winsys_cs
*rcs
,
419 enum radeon_bo_layout microtiled
,
420 enum radeon_bo_layout macrotiled
,
423 struct radeon_bo
*bo
= get_radeon_bo(_buf
);
424 struct radeon_drm_cs
*cs
= radeon_drm_cs(rcs
);
425 struct drm_radeon_gem_set_tiling args
= {};
427 /* Tiling determines how DRM treats the buffer data.
428 * We must flush CS when changing it if the buffer is referenced. */
429 if (cs
&& radeon_bo_is_referenced_by_cs(cs
, bo
)) {
430 cs
->flush_cs(cs
->flush_data
, 0);
433 while (p_atomic_read(&bo
->num_active_ioctls
)) {
437 if (microtiled
== RADEON_LAYOUT_TILED
)
438 args
.tiling_flags
|= RADEON_BO_FLAGS_MICRO_TILE
;
439 else if (microtiled
== RADEON_LAYOUT_SQUARETILED
)
440 args
.tiling_flags
|= RADEON_BO_FLAGS_MICRO_TILE_SQUARE
;
442 if (macrotiled
== RADEON_LAYOUT_TILED
)
443 args
.tiling_flags
|= RADEON_BO_FLAGS_MACRO_TILE
;
445 args
.handle
= bo
->handle
;
448 drmCommandWriteRead(bo
->rws
->fd
,
449 DRM_RADEON_GEM_SET_TILING
,
454 static struct radeon_winsys_cs_handle
*radeon_drm_get_cs_handle(
455 struct pb_buffer
*_buf
)
457 /* return radeon_bo. */
458 return (struct radeon_winsys_cs_handle
*)get_radeon_bo(_buf
);
461 static unsigned get_pb_usage_from_create_flags(enum radeon_bo_domain domain
)
465 if (domain
& RADEON_DOMAIN_GTT
)
466 res
|= RADEON_PB_USAGE_DOMAIN_GTT
;
468 if (domain
& RADEON_DOMAIN_VRAM
)
469 res
|= RADEON_PB_USAGE_DOMAIN_VRAM
;
474 static struct pb_buffer
*
475 radeon_winsys_bo_create(struct radeon_winsys
*rws
,
479 enum radeon_bo_domain domain
)
481 struct radeon_drm_winsys
*ws
= radeon_drm_winsys(rws
);
483 struct pb_manager
*provider
;
484 struct pb_buffer
*buffer
;
486 memset(&desc
, 0, sizeof(desc
));
487 desc
.alignment
= alignment
;
488 desc
.usage
= get_pb_usage_from_create_flags(domain
);
490 /* Assign a buffer manager. */
491 if (bind
& (PIPE_BIND_VERTEX_BUFFER
| PIPE_BIND_INDEX_BUFFER
|
492 PIPE_BIND_CONSTANT_BUFFER
))
497 buffer
= provider
->create_buffer(provider
, size
, &desc
);
501 return (struct pb_buffer
*)buffer
;
504 static struct pb_buffer
*radeon_winsys_bo_from_handle(struct radeon_winsys
*rws
,
505 struct winsys_handle
*whandle
,
509 struct radeon_drm_winsys
*ws
= radeon_drm_winsys(rws
);
510 struct radeon_bo
*bo
;
511 struct radeon_bomgr
*mgr
= radeon_bomgr(ws
->kman
);
512 struct drm_gem_open open_arg
= {};
514 /* We must maintain a list of pairs <handle, bo>, so that we always return
515 * the same BO for one particular handle. If we didn't do that and created
516 * more than one BO for the same handle and then relocated them in a CS,
517 * we would hit a deadlock in the kernel.
519 * The list of pairs is guarded by a mutex, of course. */
520 pipe_mutex_lock(mgr
->bo_handles_mutex
);
522 /* First check if there already is an existing bo for the handle. */
523 bo
= util_hash_table_get(mgr
->bo_handles
, (void*)(uintptr_t)whandle
->handle
);
525 /* Increase the refcount. */
526 struct pb_buffer
*b
= NULL
;
527 pb_reference(&b
, &bo
->base
);
531 /* There isn't, create a new one. */
532 bo
= CALLOC_STRUCT(radeon_bo
);
538 open_arg
.name
= whandle
->handle
;
539 if (drmIoctl(ws
->fd
, DRM_IOCTL_GEM_OPEN
, &open_arg
)) {
543 bo
->handle
= open_arg
.handle
;
544 bo
->size
= open_arg
.size
;
545 bo
->name
= whandle
->handle
;
548 pipe_reference_init(&bo
->base
.base
.reference
, 1);
549 bo
->base
.base
.alignment
= 0;
550 bo
->base
.base
.usage
= PB_USAGE_GPU_WRITE
| PB_USAGE_GPU_READ
;
551 bo
->base
.base
.size
= bo
->size
;
552 bo
->base
.vtbl
= &radeon_bo_vtbl
;
555 pipe_mutex_init(bo
->map_mutex
);
557 util_hash_table_set(mgr
->bo_handles
, (void*)(uintptr_t)whandle
->handle
, bo
);
560 pipe_mutex_unlock(mgr
->bo_handles_mutex
);
563 *stride
= whandle
->stride
;
565 *size
= bo
->base
.base
.size
;
567 return (struct pb_buffer
*)bo
;
570 pipe_mutex_unlock(mgr
->bo_handles_mutex
);
574 static boolean
radeon_winsys_bo_get_handle(struct pb_buffer
*buffer
,
576 struct winsys_handle
*whandle
)
578 struct drm_gem_flink flink
= {};
579 struct radeon_bo
*bo
= get_radeon_bo(buffer
);
581 if (whandle
->type
== DRM_API_HANDLE_TYPE_SHARED
) {
583 flink
.handle
= bo
->handle
;
585 if (ioctl(bo
->rws
->fd
, DRM_IOCTL_GEM_FLINK
, &flink
)) {
590 bo
->flink
= flink
.name
;
592 whandle
->handle
= bo
->flink
;
593 } else if (whandle
->type
== DRM_API_HANDLE_TYPE_KMS
) {
594 whandle
->handle
= bo
->handle
;
597 whandle
->stride
= stride
;
601 void radeon_bomgr_init_functions(struct radeon_drm_winsys
*ws
)
603 ws
->base
.buffer_get_cs_handle
= radeon_drm_get_cs_handle
;
604 ws
->base
.buffer_set_tiling
= radeon_bo_set_tiling
;
605 ws
->base
.buffer_get_tiling
= radeon_bo_get_tiling
;
606 ws
->base
.buffer_map
= radeon_bo_map
;
607 ws
->base
.buffer_unmap
= pb_unmap
;
608 ws
->base
.buffer_wait
= radeon_bo_wait
;
609 ws
->base
.buffer_is_busy
= radeon_bo_is_busy
;
610 ws
->base
.buffer_create
= radeon_winsys_bo_create
;
611 ws
->base
.buffer_from_handle
= radeon_winsys_bo_from_handle
;
612 ws
->base
.buffer_get_handle
= radeon_winsys_bo_get_handle
;