src/gallium/drivers/panfrost/pan_drm.c

   1 /*
   2  * © Copyright 2019 Collabora, Ltd.
   3  * Copyright 2019 Alyssa Rosenzweig
   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 (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  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
  22  * SOFTWARE.
  23  *
  24  */
  25
  26 #include <fcntl.h>
  27 #include <xf86drm.h>
  28
  29 #include "drm-uapi/panfrost_drm.h"
  30
  31 #include "util/u_memory.h"
  32 #include "util/os_time.h"
  33 #include "os/os_mman.h"
  34
  35 #include "pan_screen.h"
  36 #include "pan_resource.h"
  37 #include "pan_context.h"
  38 #include "pan_util.h"
  39 #include "pandecode/decode.h"
  40
  41 void
  42 panfrost_drm_mmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo)
  43 {
  44         struct drm_panfrost_mmap_bo mmap_bo = { .handle = bo->gem_handle };
  45         int ret;
  46
  47         if (bo->cpu)
  48                 return;
  49
  50         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
  51         if (ret) {
  52                 fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret);
  53                 assert(0);
  54         }
  55
  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\n", bo->cpu);
  60                 assert(0);
  61         }
  62
  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);
  66 }
  67
  68 static void
  69 panfrost_drm_munmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo)
  70 {
  71         if (!bo->cpu)
  72                 return;
  73
  74         if (os_munmap((void *) (uintptr_t)bo->cpu, bo->size)) {
  75                 perror("munmap");
  76                 abort();
  77         }
  78
  79         bo->cpu = NULL;
  80 }
  81
  82 struct panfrost_bo *
  83 panfrost_drm_create_bo(struct panfrost_screen *screen, size_t size,
  84                        uint32_t flags)
  85 {
  86         struct panfrost_bo *bo;
  87
  88         /* Kernel will fail (confusingly) with EPERM otherwise */
  89         assert(size > 0);
  90
  91         unsigned translated_flags = 0;
  92
  93         /* TODO: translate flags to kernel flags, if the kernel supports */
  94
  95         struct drm_panfrost_create_bo create_bo = {
  96                 .size = size,
  97                 .flags = translated_flags,
  98         };
  99
 100         /* Before creating a BO, we first want to check the cache */
 101
 102         bo = panfrost_bo_cache_fetch(screen, size, flags);
 103
 104         if (bo == NULL) {
 105                 /* Otherwise, the cache misses and we need to allocate a BO fresh from
 106                  * the kernel */
 107
 108                 int ret;
 109
 110                 ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
 111                 if (ret) {
 112                         fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %d\n", ret);
 113                         assert(0);
 114                 }
 115
 116                 /* We have a BO allocated from the kernel; fill in the userspace
 117                  * version */
 118
 119                 bo = rzalloc(screen, struct panfrost_bo);
 120                 bo->size = create_bo.size;
 121                 bo->gpu = create_bo.offset;
 122                 bo->gem_handle = create_bo.handle;
 123         }
 124
 125         /* Only mmap now if we know we need to. For CPU-invisible buffers, we
 126          * never map since we don't care about their contents; they're purely
 127          * for GPU-internal use. But we do trace them anyway. */
 128
 129         if (!(flags & (PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_DELAY_MMAP)))
 130                 panfrost_drm_mmap_bo(screen, bo);
 131         else if (flags & PAN_ALLOCATE_INVISIBLE) {
 132                 if (pan_debug & PAN_DBG_TRACE)
 133                         pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL);
 134         }
 135
 136         pipe_reference_init(&bo->reference, 1);
 137         return bo;
 138 }
 139
 140 void
 141 panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable)
 142 {
 143         struct drm_gem_close gem_close = { .handle = bo->gem_handle };
 144         int ret;
 145
 146         if (!bo)
 147                 return;
 148
 149         /* Rather than freeing the BO now, we'll cache the BO for later
 150          * allocations if we're allowed to */
 151
 152         if (cacheable) {
 153                 bool cached = panfrost_bo_cache_put(screen, bo);
 154
 155                 if (cached)
 156                         return;
 157         }
 158
 159         /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */
 160
 161         panfrost_drm_munmap_bo(screen, bo);
 162
 163         ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
 164         if (ret) {
 165                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
 166                 assert(0);
 167         }
 168
 169         ralloc_free(bo);
 170 }
 171
 172 void
 173 panfrost_drm_allocate_slab(struct panfrost_screen *screen,
 174                            struct panfrost_memory *mem,
 175                            size_t pages,
 176                            bool same_va,
 177                            int extra_flags,
 178                            int commit_count,
 179                            int extent)
 180 {
 181         // TODO cache allocations
 182         // TODO properly handle errors
 183         // TODO take into account extra_flags
 184         mem->bo = panfrost_drm_create_bo(screen, pages * 4096, extra_flags);
 185         mem->stack_bottom = 0;
 186 }
 187
 188 void
 189 panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
 190 {
 191         panfrost_bo_unreference(&screen->base, mem->bo);
 192         mem->bo = NULL;
 193 }
 194
 195 struct panfrost_bo *
 196 panfrost_drm_import_bo(struct panfrost_screen *screen, int fd)
 197 {
 198         struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
 199         struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
 200         ASSERTED int ret;
 201         unsigned gem_handle;
 202
 203         ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle);
 204         assert(!ret);
 205
 206         get_bo_offset.handle = gem_handle;
 207         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
 208         assert(!ret);
 209
 210         bo->gem_handle = gem_handle;
 211         bo->gpu = (mali_ptr) get_bo_offset.offset;
 212         bo->size = lseek(fd, 0, SEEK_END);
 213         assert(bo->size > 0);
 214         pipe_reference_init(&bo->reference, 1);
 215
 216         // TODO map and unmap on demand?
 217         panfrost_drm_mmap_bo(screen, bo);
 218         return bo;
 219 }
 220
 221 int
 222 panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo)
 223 {
 224         struct drm_prime_handle args = {
 225                 .handle = bo->gem_handle,
 226                 .flags = DRM_CLOEXEC,
 227         };
 228
 229         int ret = drmIoctl(screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 230         if (ret == -1)
 231                 return -1;
 232
 233         return args.fd;
 234 }
 235
 236 static int
 237 panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs)
 238 {
 239         struct pipe_context *gallium = (struct pipe_context *) ctx;
 240         struct panfrost_screen *screen = pan_screen(gallium->screen);
 241         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 242         struct drm_panfrost_submit submit = {0,};
 243         int *bo_handles, ret;
 244
 245         submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync;
 246         submit.in_sync_count = 1;
 247
 248         submit.out_sync = ctx->out_sync;
 249
 250         submit.jc = job_desc;
 251         submit.requirements = reqs;
 252
 253         bo_handles = calloc(job->bos->entries, sizeof(*bo_handles));
 254         assert(bo_handles);
 255
 256         set_foreach(job->bos, entry) {
 257                 struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
 258                 assert(bo->gem_handle > 0);
 259                 bo_handles[submit.bo_handle_count++] = bo->gem_handle;
 260         }
 261
 262         submit.bo_handles = (u64) (uintptr_t) bo_handles;
 263         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
 264         free(bo_handles);
 265         if (ret) {
 266                 fprintf(stderr, "Error submitting: %m\n");
 267                 return errno;
 268         }
 269
 270         /* Trace the job if we're doing that */
 271         if (pan_debug & PAN_DBG_TRACE) {
 272                 /* Wait so we can get errors reported back */
 273                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 274                 pandecode_jc(submit.jc, FALSE);
 275         }
 276
 277         return 0;
 278 }
 279
 280 int
 281 panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout)
 282 {
 283         int ret = 0;
 284
 285         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 286
 287         /* TODO: Add here the transient pools */
 288         panfrost_job_add_bo(job, ctx->shaders.bo);
 289         panfrost_job_add_bo(job, ctx->scratchpad.bo);
 290         panfrost_job_add_bo(job, ctx->tiler_heap.bo);
 291         panfrost_job_add_bo(job, ctx->tiler_polygon_list.bo);
 292
 293         if (job->first_job.gpu) {
 294                 ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0);
 295                 assert(!ret);
 296         }
 297
 298         if (job->first_tiler.gpu || job->clear) {
 299                 ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx, has_draws), PANFROST_JD_REQ_FS);
 300                 assert(!ret);
 301         }
 302
 303         return ret;
 304 }
 305
 306 static struct panfrost_fence *
 307 panfrost_fence_create(struct panfrost_context *ctx)
 308 {
 309         struct pipe_context *gallium = (struct pipe_context *) ctx;
 310         struct panfrost_screen *screen = pan_screen(gallium->screen);
 311         struct panfrost_fence *f = calloc(1, sizeof(*f));
 312         if (!f)
 313                 return NULL;
 314
 315         /* Snapshot the last Panfrost's rendering's out fence.  We'd rather have
 316          * another syncobj instead of a sync file, but this is all we get.
 317          * (HandleToFD/FDToHandle just gives you another syncobj ID for the
 318          * same syncobj).
 319          */
 320         drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd);
 321         if (f->fd == -1) {
 322                 fprintf(stderr, "export failed\n");
 323                 free(f);
 324                 return NULL;
 325         }
 326
 327         pipe_reference_init(&f->reference, 1);
 328
 329         return f;
 330 }
 331
 332 void
 333 panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
 334                                   struct pipe_fence_handle **fence)
 335 {
 336         struct pipe_context *gallium = (struct pipe_context *) ctx;
 337         struct panfrost_screen *screen = pan_screen(gallium->screen);
 338
 339         if (!screen->last_fragment_flushed) {
 340                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 341                 screen->last_fragment_flushed = true;
 342
 343                 /* The job finished up, so we're safe to clean it up now */
 344                 panfrost_free_job(ctx, screen->last_job);
 345         }
 346
 347         if (fence) {
 348                 struct panfrost_fence *f = panfrost_fence_create(ctx);
 349                 gallium->screen->fence_reference(gallium->screen, fence, NULL);
 350                 *fence = (struct pipe_fence_handle *)f;
 351         }
 352 }
 353
 354 unsigned
 355 panfrost_drm_query_gpu_version(struct panfrost_screen *screen)
 356 {
 357         struct drm_panfrost_get_param get_param = {0,};
 358         ASSERTED int ret;
 359
 360         get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
 361         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
 362         assert(!ret);
 363
 364         return get_param.value;
 365 }
 366
 367 int
 368 panfrost_drm_init_context(struct panfrost_context *ctx)
 369 {
 370         struct pipe_context *gallium = (struct pipe_context *) ctx;
 371         struct panfrost_screen *screen = pan_screen(gallium->screen);
 372
 373         return drmSyncobjCreate(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED,
 374                                 &ctx->out_sync);
 375 }
 376
 377 void
 378 panfrost_drm_fence_reference(struct pipe_screen *screen,
 379                              struct pipe_fence_handle **ptr,
 380                              struct pipe_fence_handle *fence)
 381 {
 382         struct panfrost_fence **p = (struct panfrost_fence **)ptr;
 383         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 384         struct panfrost_fence *old = *p;
 385
 386         if (pipe_reference(&(*p)->reference, &f->reference)) {
 387                 close(old->fd);
 388                 free(old);
 389         }
 390         *p = f;
 391 }
 392
 393 boolean
 394 panfrost_drm_fence_finish(struct pipe_screen *pscreen,
 395                           struct pipe_context *ctx,
 396                           struct pipe_fence_handle *fence,
 397                           uint64_t timeout)
 398 {
 399         struct panfrost_screen *screen = pan_screen(pscreen);
 400         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 401         int ret;
 402
 403         unsigned syncobj;
 404         ret = drmSyncobjCreate(screen->fd, 0, &syncobj);
 405         if (ret) {
 406                 fprintf(stderr, "Failed to create syncobj to wait on: %m\n");
 407                 return false;
 408         }
 409
 410         drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd);
 411         if (ret) {
 412                 fprintf(stderr, "Failed to import fence to syncobj: %m\n");
 413                 return false;
 414         }
 415
 416         uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
 417         if (abs_timeout == OS_TIMEOUT_INFINITE)
 418                 abs_timeout = INT64_MAX;
 419
 420         ret = drmSyncobjWait(screen->fd, &syncobj, 1, abs_timeout, 0, NULL);
 421
 422         drmSyncobjDestroy(screen->fd, syncobj);
 423
 424         return ret >= 0;
 425 }