2 * © Copyright 2019 Collabora, Ltd.
3 * Copyright 2019 Alyssa Rosenzweig
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 (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
29 #include "drm-uapi/panfrost_drm.h"
31 #include "util/u_memory.h"
32 #include "util/os_time.h"
33 #include "os/os_mman.h"
35 #include "pan_screen.h"
36 #include "pan_resource.h"
37 #include "pan_context.h"
39 #include "pandecode/decode.h"
42 panfrost_drm_mmap_bo(struct panfrost_screen
*screen
, struct panfrost_bo
*bo
)
44 struct drm_panfrost_mmap_bo mmap_bo
= { .handle
= bo
->gem_handle
};
50 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_MMAP_BO
, &mmap_bo
);
52 fprintf(stderr
, "DRM_IOCTL_PANFROST_MMAP_BO failed: %m\n");
56 bo
->cpu
= os_mmap(NULL
, bo
->size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
57 screen
->fd
, mmap_bo
.offset
);
58 if (bo
->cpu
== MAP_FAILED
) {
59 fprintf(stderr
, "mmap failed: %p %m\n", bo
->cpu
);
63 /* Record the mmap if we're tracing */
64 if (pan_debug
& PAN_DBG_TRACE
)
65 pandecode_inject_mmap(bo
->gpu
, bo
->cpu
, bo
->size
, NULL
);
69 panfrost_drm_munmap_bo(struct panfrost_screen
*screen
, struct panfrost_bo
*bo
)
74 if (os_munmap((void *) (uintptr_t)bo
->cpu
, bo
->size
)) {
83 panfrost_drm_create_bo(struct panfrost_screen
*screen
, size_t size
,
86 struct panfrost_bo
*bo
;
88 /* Kernel will fail (confusingly) with EPERM otherwise */
91 /* To maximize BO cache usage, don't allocate tiny BOs */
92 size
= MAX2(size
, 4096);
94 /* GROWABLE BOs cannot be mmapped */
95 if (flags
& PAN_ALLOCATE_GROWABLE
)
96 assert(flags
& PAN_ALLOCATE_INVISIBLE
);
98 unsigned translated_flags
= 0;
100 if (screen
->kernel_version
->version_major
> 1 ||
101 screen
->kernel_version
->version_minor
>= 1) {
102 if (flags
& PAN_ALLOCATE_GROWABLE
)
103 translated_flags
|= PANFROST_BO_HEAP
;
104 if (!(flags
& PAN_ALLOCATE_EXECUTE
))
105 translated_flags
|= PANFROST_BO_NOEXEC
;
108 struct drm_panfrost_create_bo create_bo
= {
110 .flags
= translated_flags
,
113 /* Before creating a BO, we first want to check the cache */
115 bo
= panfrost_bo_cache_fetch(screen
, size
, flags
);
118 /* Otherwise, the cache misses and we need to allocate a BO fresh from
123 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_CREATE_BO
, &create_bo
);
125 fprintf(stderr
, "DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n");
129 /* We have a BO allocated from the kernel; fill in the userspace
132 bo
= rzalloc(screen
, struct panfrost_bo
);
133 bo
->size
= create_bo
.size
;
134 bo
->gpu
= create_bo
.offset
;
135 bo
->gem_handle
= create_bo
.handle
;
139 /* Only mmap now if we know we need to. For CPU-invisible buffers, we
140 * never map since we don't care about their contents; they're purely
141 * for GPU-internal use. But we do trace them anyway. */
143 if (!(flags
& (PAN_ALLOCATE_INVISIBLE
| PAN_ALLOCATE_DELAY_MMAP
)))
144 panfrost_drm_mmap_bo(screen
, bo
);
145 else if (flags
& PAN_ALLOCATE_INVISIBLE
) {
146 if (pan_debug
& PAN_DBG_TRACE
)
147 pandecode_inject_mmap(bo
->gpu
, NULL
, bo
->size
, NULL
);
150 pipe_reference_init(&bo
->reference
, 1);
155 panfrost_drm_release_bo(struct panfrost_screen
*screen
, struct panfrost_bo
*bo
, bool cacheable
)
160 struct drm_gem_close gem_close
= { .handle
= bo
->gem_handle
};
163 /* Rather than freeing the BO now, we'll cache the BO for later
164 * allocations if we're allowed to */
166 panfrost_drm_munmap_bo(screen
, bo
);
169 bool cached
= panfrost_bo_cache_put(screen
, bo
);
175 /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */
177 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_GEM_CLOSE
, &gem_close
);
179 fprintf(stderr
, "DRM_IOCTL_GEM_CLOSE failed: %m\n");
187 panfrost_drm_allocate_slab(struct panfrost_screen
*screen
,
188 struct panfrost_memory
*mem
,
195 // TODO cache allocations
196 // TODO properly handle errors
197 // TODO take into account extra_flags
198 mem
->bo
= panfrost_drm_create_bo(screen
, pages
* 4096, extra_flags
);
199 mem
->stack_bottom
= 0;
203 panfrost_drm_free_slab(struct panfrost_screen
*screen
, struct panfrost_memory
*mem
)
205 panfrost_bo_unreference(&screen
->base
, mem
->bo
);
210 panfrost_drm_import_bo(struct panfrost_screen
*screen
, int fd
)
212 struct panfrost_bo
*bo
= rzalloc(screen
, struct panfrost_bo
);
213 struct drm_panfrost_get_bo_offset get_bo_offset
= {0,};
217 ret
= drmPrimeFDToHandle(screen
->fd
, fd
, &gem_handle
);
220 get_bo_offset
.handle
= gem_handle
;
221 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_GET_BO_OFFSET
, &get_bo_offset
);
224 bo
->gem_handle
= gem_handle
;
225 bo
->gpu
= (mali_ptr
) get_bo_offset
.offset
;
226 bo
->size
= lseek(fd
, 0, SEEK_END
);
227 assert(bo
->size
> 0);
228 pipe_reference_init(&bo
->reference
, 1);
230 // TODO map and unmap on demand?
231 panfrost_drm_mmap_bo(screen
, bo
);
236 panfrost_drm_export_bo(struct panfrost_screen
*screen
, const struct panfrost_bo
*bo
)
238 struct drm_prime_handle args
= {
239 .handle
= bo
->gem_handle
,
240 .flags
= DRM_CLOEXEC
,
243 int ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PRIME_HANDLE_TO_FD
, &args
);
251 panfrost_drm_submit_job(struct panfrost_context
*ctx
, u64 job_desc
, int reqs
)
253 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
254 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
255 struct panfrost_job
*job
= panfrost_get_job_for_fbo(ctx
);
256 struct drm_panfrost_submit submit
= {0,};
257 int *bo_handles
, ret
;
259 submit
.in_syncs
= (u64
) (uintptr_t) &ctx
->out_sync
;
260 submit
.in_sync_count
= 1;
262 submit
.out_sync
= ctx
->out_sync
;
264 submit
.jc
= job_desc
;
265 submit
.requirements
= reqs
;
267 bo_handles
= calloc(job
->bos
->entries
, sizeof(*bo_handles
));
270 set_foreach(job
->bos
, entry
) {
271 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
272 assert(bo
->gem_handle
> 0);
273 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
276 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
277 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
280 fprintf(stderr
, "Error submitting: %m\n");
284 /* Trace the job if we're doing that */
285 if (pan_debug
& PAN_DBG_TRACE
) {
286 /* Wait so we can get errors reported back */
287 drmSyncobjWait(screen
->fd
, &ctx
->out_sync
, 1, INT64_MAX
, 0, NULL
);
288 pandecode_jc(submit
.jc
, FALSE
);
295 panfrost_drm_submit_vs_fs_job(struct panfrost_context
*ctx
, bool has_draws
)
299 struct panfrost_job
*job
= panfrost_get_job_for_fbo(ctx
);
301 panfrost_job_add_bo(job
, ctx
->scratchpad
.bo
);
302 panfrost_job_add_bo(job
, ctx
->tiler_heap
.bo
);
303 panfrost_job_add_bo(job
, job
->polygon_list
);
305 if (job
->first_job
.gpu
) {
306 ret
= panfrost_drm_submit_job(ctx
, job
->first_job
.gpu
, 0);
310 if (job
->first_tiler
.gpu
|| job
->clear
) {
311 ret
= panfrost_drm_submit_job(ctx
, panfrost_fragment_job(ctx
, has_draws
), PANFROST_JD_REQ_FS
);
318 static struct panfrost_fence
*
319 panfrost_fence_create(struct panfrost_context
*ctx
)
321 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
322 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
323 struct panfrost_fence
*f
= calloc(1, sizeof(*f
));
327 /* Snapshot the last Panfrost's rendering's out fence. We'd rather have
328 * another syncobj instead of a sync file, but this is all we get.
329 * (HandleToFD/FDToHandle just gives you another syncobj ID for the
332 drmSyncobjExportSyncFile(screen
->fd
, ctx
->out_sync
, &f
->fd
);
334 fprintf(stderr
, "export failed: %m\n");
339 pipe_reference_init(&f
->reference
, 1);
345 panfrost_drm_force_flush_fragment(struct panfrost_context
*ctx
,
346 struct pipe_fence_handle
**fence
)
348 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
349 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
351 if (!ctx
->last_fragment_flushed
) {
352 drmSyncobjWait(screen
->fd
, &ctx
->out_sync
, 1, INT64_MAX
, 0, NULL
);
353 ctx
->last_fragment_flushed
= true;
355 /* The job finished up, so we're safe to clean it up now */
356 panfrost_free_job(ctx
, ctx
->last_job
);
360 struct panfrost_fence
*f
= panfrost_fence_create(ctx
);
361 gallium
->screen
->fence_reference(gallium
->screen
, fence
, NULL
);
362 *fence
= (struct pipe_fence_handle
*)f
;
367 panfrost_drm_query_gpu_version(struct panfrost_screen
*screen
)
369 struct drm_panfrost_get_param get_param
= {0,};
372 get_param
.param
= DRM_PANFROST_PARAM_GPU_PROD_ID
;
373 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_GET_PARAM
, &get_param
);
376 return get_param
.value
;
380 panfrost_drm_init_context(struct panfrost_context
*ctx
)
382 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
383 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
385 return drmSyncobjCreate(screen
->fd
, DRM_SYNCOBJ_CREATE_SIGNALED
,
390 panfrost_drm_fence_reference(struct pipe_screen
*screen
,
391 struct pipe_fence_handle
**ptr
,
392 struct pipe_fence_handle
*fence
)
394 struct panfrost_fence
**p
= (struct panfrost_fence
**)ptr
;
395 struct panfrost_fence
*f
= (struct panfrost_fence
*)fence
;
396 struct panfrost_fence
*old
= *p
;
398 if (pipe_reference(&(*p
)->reference
, &f
->reference
)) {
406 panfrost_drm_fence_finish(struct pipe_screen
*pscreen
,
407 struct pipe_context
*ctx
,
408 struct pipe_fence_handle
*fence
,
411 struct panfrost_screen
*screen
= pan_screen(pscreen
);
412 struct panfrost_fence
*f
= (struct panfrost_fence
*)fence
;
416 ret
= drmSyncobjCreate(screen
->fd
, 0, &syncobj
);
418 fprintf(stderr
, "Failed to create syncobj to wait on: %m\n");
422 ret
= drmSyncobjImportSyncFile(screen
->fd
, syncobj
, f
->fd
);
424 fprintf(stderr
, "Failed to import fence to syncobj: %m\n");
428 uint64_t abs_timeout
= os_time_get_absolute_timeout(timeout
);
429 if (abs_timeout
== OS_TIMEOUT_INFINITE
)
430 abs_timeout
= INT64_MAX
;
432 ret
= drmSyncobjWait(screen
->fd
, &syncobj
, 1, abs_timeout
, 0, NULL
);
434 drmSyncobjDestroy(screen
->fd
, syncobj
);