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: %m\n");
  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 %m\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         /* To maximize BO cache usage, don't allocate tiny BOs */
  92         size = MAX2(size, 4096);
  93
  94         /* GROWABLE BOs cannot be mmapped */
  95         if (flags & PAN_ALLOCATE_GROWABLE)
  96                 assert(flags & PAN_ALLOCATE_INVISIBLE);
  97
  98         unsigned translated_flags = 0;
  99
 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;
 106         }
 107
 108         struct drm_panfrost_create_bo create_bo = {
 109                 .size = size,
 110                 .flags = translated_flags,
 111         };
 112
 113         /* Before creating a BO, we first want to check the cache */
 114
 115         bo = panfrost_bo_cache_fetch(screen, size, flags);
 116
 117         if (bo == NULL) {
 118                 /* Otherwise, the cache misses and we need to allocate a BO fresh from
 119                  * the kernel */
 120
 121                 int ret;
 122
 123                 ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
 124                 if (ret) {
 125                         fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n");
 126                         assert(0);
 127                 }
 128
 129                 /* We have a BO allocated from the kernel; fill in the userspace
 130                  * version */
 131
 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;
 136                 bo->flags = flags;
 137         }
 138
 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. */
 142
 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);
 148         }
 149
 150         pipe_reference_init(&bo->reference, 1);
 151         return bo;
 152 }
 153
 154 void
 155 panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable)
 156 {
 157         struct drm_gem_close gem_close = { .handle = bo->gem_handle };
 158         int ret;
 159
 160         if (!bo)
 161                 return;
 162
 163         /* Rather than freeing the BO now, we'll cache the BO for later
 164          * allocations if we're allowed to */
 165
 166         if (cacheable) {
 167                 bool cached = panfrost_bo_cache_put(screen, bo);
 168
 169                 if (cached)
 170                         return;
 171         }
 172
 173         /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */
 174
 175         panfrost_drm_munmap_bo(screen, bo);
 176
 177         ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
 178         if (ret) {
 179                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %m\n");
 180                 assert(0);
 181         }
 182
 183         ralloc_free(bo);
 184 }
 185
 186 void
 187 panfrost_drm_allocate_slab(struct panfrost_screen *screen,
 188                            struct panfrost_memory *mem,
 189                            size_t pages,
 190                            bool same_va,
 191                            int extra_flags,
 192                            int commit_count,
 193                            int extent)
 194 {
 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;
 200 }
 201
 202 void
 203 panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
 204 {
 205         panfrost_bo_unreference(&screen->base, mem->bo);
 206         mem->bo = NULL;
 207 }
 208
 209 struct panfrost_bo *
 210 panfrost_drm_import_bo(struct panfrost_screen *screen, int fd)
 211 {
 212         struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
 213         struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
 214         ASSERTED int ret;
 215         unsigned gem_handle;
 216
 217         ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle);
 218         assert(!ret);
 219
 220         get_bo_offset.handle = gem_handle;
 221         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
 222         assert(!ret);
 223
 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);
 229
 230         // TODO map and unmap on demand?
 231         panfrost_drm_mmap_bo(screen, bo);
 232         return bo;
 233 }
 234
 235 int
 236 panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo)
 237 {
 238         struct drm_prime_handle args = {
 239                 .handle = bo->gem_handle,
 240                 .flags = DRM_CLOEXEC,
 241         };
 242
 243         int ret = drmIoctl(screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 244         if (ret == -1)
 245                 return -1;
 246
 247         return args.fd;
 248 }
 249
 250 static int
 251 panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs)
 252 {
 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;
 258
 259         submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync;
 260         submit.in_sync_count = 1;
 261
 262         submit.out_sync = ctx->out_sync;
 263
 264         submit.jc = job_desc;
 265         submit.requirements = reqs;
 266
 267         bo_handles = calloc(job->bos->entries, sizeof(*bo_handles));
 268         assert(bo_handles);
 269
 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;
 274         }
 275
 276         submit.bo_handles = (u64) (uintptr_t) bo_handles;
 277         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
 278         free(bo_handles);
 279         if (ret) {
 280                 fprintf(stderr, "Error submitting: %m\n");
 281                 return errno;
 282         }
 283
 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);
 289         }
 290
 291         return 0;
 292 }
 293
 294 int
 295 panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout)
 296 {
 297         int ret = 0;
 298
 299         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 300
 301         /* TODO: Add here the transient pools */
 302         panfrost_job_add_bo(job, ctx->scratchpad.bo);
 303         panfrost_job_add_bo(job, ctx->tiler_heap.bo);
 304         panfrost_job_add_bo(job, job->polygon_list);
 305
 306         if (job->first_job.gpu) {
 307                 ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0);
 308                 assert(!ret);
 309         }
 310
 311         if (job->first_tiler.gpu || job->clear) {
 312                 ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx, has_draws), PANFROST_JD_REQ_FS);
 313                 assert(!ret);
 314         }
 315
 316         return ret;
 317 }
 318
 319 static struct panfrost_fence *
 320 panfrost_fence_create(struct panfrost_context *ctx)
 321 {
 322         struct pipe_context *gallium = (struct pipe_context *) ctx;
 323         struct panfrost_screen *screen = pan_screen(gallium->screen);
 324         struct panfrost_fence *f = calloc(1, sizeof(*f));
 325         if (!f)
 326                 return NULL;
 327
 328         /* Snapshot the last Panfrost's rendering's out fence.  We'd rather have
 329          * another syncobj instead of a sync file, but this is all we get.
 330          * (HandleToFD/FDToHandle just gives you another syncobj ID for the
 331          * same syncobj).
 332          */
 333         drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd);
 334         if (f->fd == -1) {
 335                 fprintf(stderr, "export failed: %m\n");
 336                 free(f);
 337                 return NULL;
 338         }
 339
 340         pipe_reference_init(&f->reference, 1);
 341
 342         return f;
 343 }
 344
 345 void
 346 panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
 347                                   struct pipe_fence_handle **fence)
 348 {
 349         struct pipe_context *gallium = (struct pipe_context *) ctx;
 350         struct panfrost_screen *screen = pan_screen(gallium->screen);
 351
 352         if (!screen->last_fragment_flushed) {
 353                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 354                 screen->last_fragment_flushed = true;
 355
 356                 /* The job finished up, so we're safe to clean it up now */
 357                 panfrost_free_job(ctx, screen->last_job);
 358         }
 359
 360         if (fence) {
 361                 struct panfrost_fence *f = panfrost_fence_create(ctx);
 362                 gallium->screen->fence_reference(gallium->screen, fence, NULL);
 363                 *fence = (struct pipe_fence_handle *)f;
 364         }
 365 }
 366
 367 unsigned
 368 panfrost_drm_query_gpu_version(struct panfrost_screen *screen)
 369 {
 370         struct drm_panfrost_get_param get_param = {0,};
 371         ASSERTED int ret;
 372
 373         get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
 374         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
 375         assert(!ret);
 376
 377         return get_param.value;
 378 }
 379
 380 int
 381 panfrost_drm_init_context(struct panfrost_context *ctx)
 382 {
 383         struct pipe_context *gallium = (struct pipe_context *) ctx;
 384         struct panfrost_screen *screen = pan_screen(gallium->screen);
 385
 386         return drmSyncobjCreate(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED,
 387                                 &ctx->out_sync);
 388 }
 389
 390 void
 391 panfrost_drm_fence_reference(struct pipe_screen *screen,
 392                              struct pipe_fence_handle **ptr,
 393                              struct pipe_fence_handle *fence)
 394 {
 395         struct panfrost_fence **p = (struct panfrost_fence **)ptr;
 396         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 397         struct panfrost_fence *old = *p;
 398
 399         if (pipe_reference(&(*p)->reference, &f->reference)) {
 400                 close(old->fd);
 401                 free(old);
 402         }
 403         *p = f;
 404 }
 405
 406 boolean
 407 panfrost_drm_fence_finish(struct pipe_screen *pscreen,
 408                           struct pipe_context *ctx,
 409                           struct pipe_fence_handle *fence,
 410                           uint64_t timeout)
 411 {
 412         struct panfrost_screen *screen = pan_screen(pscreen);
 413         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 414         int ret;
 415
 416         unsigned syncobj;
 417         ret = drmSyncobjCreate(screen->fd, 0, &syncobj);
 418         if (ret) {
 419                 fprintf(stderr, "Failed to create syncobj to wait on: %m\n");
 420                 return false;
 421         }
 422
 423         drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd);
 424         if (ret) {
 425                 fprintf(stderr, "Failed to import fence to syncobj: %m\n");
 426                 return false;
 427         }
 428
 429         uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
 430         if (abs_timeout == OS_TIMEOUT_INFINITE)
 431                 abs_timeout = INT64_MAX;
 432
 433         ret = drmSyncobjWait(screen->fd, &syncobj, 1, abs_timeout, 0, NULL);
 434
 435         drmSyncobjDestroy(screen->fd, syncobj);
 436
 437         return ret >= 0;
 438 }