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         /* To maximize BO cache usage, don't allocate tiny BOs */
  92         size = MAX2(size, 4096);
  93
  94         unsigned translated_flags = 0;
  95
  96         /* TODO: translate flags to kernel flags, if the kernel supports */
  97
  98         struct drm_panfrost_create_bo create_bo = {
  99                 .size = size,
 100                 .flags = translated_flags,
 101         };
 102
 103         /* Before creating a BO, we first want to check the cache */
 104
 105         bo = panfrost_bo_cache_fetch(screen, size, flags);
 106
 107         if (bo == NULL) {
 108                 /* Otherwise, the cache misses and we need to allocate a BO fresh from
 109                  * the kernel */
 110
 111                 int ret;
 112
 113                 ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
 114                 if (ret) {
 115                         fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %d\n", ret);
 116                         assert(0);
 117                 }
 118
 119                 /* We have a BO allocated from the kernel; fill in the userspace
 120                  * version */
 121
 122                 bo = rzalloc(screen, struct panfrost_bo);
 123                 bo->size = create_bo.size;
 124                 bo->gpu = create_bo.offset;
 125                 bo->gem_handle = create_bo.handle;
 126                 bo->flags = flags;
 127         }
 128
 129         /* Only mmap now if we know we need to. For CPU-invisible buffers, we
 130          * never map since we don't care about their contents; they're purely
 131          * for GPU-internal use. But we do trace them anyway. */
 132
 133         if (!(flags & (PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_DELAY_MMAP)))
 134                 panfrost_drm_mmap_bo(screen, bo);
 135         else if (flags & PAN_ALLOCATE_INVISIBLE) {
 136                 if (pan_debug & PAN_DBG_TRACE)
 137                         pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL);
 138         }
 139
 140         pipe_reference_init(&bo->reference, 1);
 141         return bo;
 142 }
 143
 144 void
 145 panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable)
 146 {
 147         struct drm_gem_close gem_close = { .handle = bo->gem_handle };
 148         int ret;
 149
 150         if (!bo)
 151                 return;
 152
 153         /* Rather than freeing the BO now, we'll cache the BO for later
 154          * allocations if we're allowed to */
 155
 156         if (cacheable) {
 157                 bool cached = panfrost_bo_cache_put(screen, bo);
 158
 159                 if (cached)
 160                         return;
 161         }
 162
 163         /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */
 164
 165         panfrost_drm_munmap_bo(screen, bo);
 166
 167         ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
 168         if (ret) {
 169                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
 170                 assert(0);
 171         }
 172
 173         ralloc_free(bo);
 174 }
 175
 176 void
 177 panfrost_drm_allocate_slab(struct panfrost_screen *screen,
 178                            struct panfrost_memory *mem,
 179                            size_t pages,
 180                            bool same_va,
 181                            int extra_flags,
 182                            int commit_count,
 183                            int extent)
 184 {
 185         // TODO cache allocations
 186         // TODO properly handle errors
 187         // TODO take into account extra_flags
 188         mem->bo = panfrost_drm_create_bo(screen, pages * 4096, extra_flags);
 189         mem->stack_bottom = 0;
 190 }
 191
 192 void
 193 panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
 194 {
 195         panfrost_bo_unreference(&screen->base, mem->bo);
 196         mem->bo = NULL;
 197 }
 198
 199 struct panfrost_bo *
 200 panfrost_drm_import_bo(struct panfrost_screen *screen, int fd)
 201 {
 202         struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
 203         struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
 204         ASSERTED int ret;
 205         unsigned gem_handle;
 206
 207         ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle);
 208         assert(!ret);
 209
 210         get_bo_offset.handle = gem_handle;
 211         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
 212         assert(!ret);
 213
 214         bo->gem_handle = gem_handle;
 215         bo->gpu = (mali_ptr) get_bo_offset.offset;
 216         bo->size = lseek(fd, 0, SEEK_END);
 217         assert(bo->size > 0);
 218         pipe_reference_init(&bo->reference, 1);
 219
 220         // TODO map and unmap on demand?
 221         panfrost_drm_mmap_bo(screen, bo);
 222         return bo;
 223 }
 224
 225 int
 226 panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo)
 227 {
 228         struct drm_prime_handle args = {
 229                 .handle = bo->gem_handle,
 230                 .flags = DRM_CLOEXEC,
 231         };
 232
 233         int ret = drmIoctl(screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 234         if (ret == -1)
 235                 return -1;
 236
 237         return args.fd;
 238 }
 239
 240 static int
 241 panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs)
 242 {
 243         struct pipe_context *gallium = (struct pipe_context *) ctx;
 244         struct panfrost_screen *screen = pan_screen(gallium->screen);
 245         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 246         struct drm_panfrost_submit submit = {0,};
 247         int *bo_handles, ret;
 248
 249         submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync;
 250         submit.in_sync_count = 1;
 251
 252         submit.out_sync = ctx->out_sync;
 253
 254         submit.jc = job_desc;
 255         submit.requirements = reqs;
 256
 257         bo_handles = calloc(job->bos->entries, sizeof(*bo_handles));
 258         assert(bo_handles);
 259
 260         set_foreach(job->bos, entry) {
 261                 struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
 262                 assert(bo->gem_handle > 0);
 263                 bo_handles[submit.bo_handle_count++] = bo->gem_handle;
 264         }
 265
 266         submit.bo_handles = (u64) (uintptr_t) bo_handles;
 267         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
 268         free(bo_handles);
 269         if (ret) {
 270                 fprintf(stderr, "Error submitting: %m\n");
 271                 return errno;
 272         }
 273
 274         /* Trace the job if we're doing that */
 275         if (pan_debug & PAN_DBG_TRACE) {
 276                 /* Wait so we can get errors reported back */
 277                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 278                 pandecode_jc(submit.jc, FALSE);
 279         }
 280
 281         return 0;
 282 }
 283
 284 int
 285 panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout)
 286 {
 287         int ret = 0;
 288
 289         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 290
 291         /* TODO: Add here the transient pools */
 292         panfrost_job_add_bo(job, ctx->scratchpad.bo);
 293         panfrost_job_add_bo(job, ctx->tiler_heap.bo);
 294         panfrost_job_add_bo(job, job->polygon_list);
 295
 296         if (job->first_job.gpu) {
 297                 ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0);
 298                 assert(!ret);
 299         }
 300
 301         if (job->first_tiler.gpu || job->clear) {
 302                 ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx, has_draws), PANFROST_JD_REQ_FS);
 303                 assert(!ret);
 304         }
 305
 306         return ret;
 307 }
 308
 309 static struct panfrost_fence *
 310 panfrost_fence_create(struct panfrost_context *ctx)
 311 {
 312         struct pipe_context *gallium = (struct pipe_context *) ctx;
 313         struct panfrost_screen *screen = pan_screen(gallium->screen);
 314         struct panfrost_fence *f = calloc(1, sizeof(*f));
 315         if (!f)
 316                 return NULL;
 317
 318         /* Snapshot the last Panfrost's rendering's out fence.  We'd rather have
 319          * another syncobj instead of a sync file, but this is all we get.
 320          * (HandleToFD/FDToHandle just gives you another syncobj ID for the
 321          * same syncobj).
 322          */
 323         drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd);
 324         if (f->fd == -1) {
 325                 fprintf(stderr, "export failed\n");
 326                 free(f);
 327                 return NULL;
 328         }
 329
 330         pipe_reference_init(&f->reference, 1);
 331
 332         return f;
 333 }
 334
 335 void
 336 panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
 337                                   struct pipe_fence_handle **fence)
 338 {
 339         struct pipe_context *gallium = (struct pipe_context *) ctx;
 340         struct panfrost_screen *screen = pan_screen(gallium->screen);
 341
 342         if (!screen->last_fragment_flushed) {
 343                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 344                 screen->last_fragment_flushed = true;
 345
 346                 /* The job finished up, so we're safe to clean it up now */
 347                 panfrost_free_job(ctx, screen->last_job);
 348         }
 349
 350         if (fence) {
 351                 struct panfrost_fence *f = panfrost_fence_create(ctx);
 352                 gallium->screen->fence_reference(gallium->screen, fence, NULL);
 353                 *fence = (struct pipe_fence_handle *)f;
 354         }
 355 }
 356
 357 unsigned
 358 panfrost_drm_query_gpu_version(struct panfrost_screen *screen)
 359 {
 360         struct drm_panfrost_get_param get_param = {0,};
 361         ASSERTED int ret;
 362
 363         get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
 364         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
 365         assert(!ret);
 366
 367         return get_param.value;
 368 }
 369
 370 int
 371 panfrost_drm_init_context(struct panfrost_context *ctx)
 372 {
 373         struct pipe_context *gallium = (struct pipe_context *) ctx;
 374         struct panfrost_screen *screen = pan_screen(gallium->screen);
 375
 376         return drmSyncobjCreate(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED,
 377                                 &ctx->out_sync);
 378 }
 379
 380 void
 381 panfrost_drm_fence_reference(struct pipe_screen *screen,
 382                              struct pipe_fence_handle **ptr,
 383                              struct pipe_fence_handle *fence)
 384 {
 385         struct panfrost_fence **p = (struct panfrost_fence **)ptr;
 386         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 387         struct panfrost_fence *old = *p;
 388
 389         if (pipe_reference(&(*p)->reference, &f->reference)) {
 390                 close(old->fd);
 391                 free(old);
 392         }
 393         *p = f;
 394 }
 395
 396 boolean
 397 panfrost_drm_fence_finish(struct pipe_screen *pscreen,
 398                           struct pipe_context *ctx,
 399                           struct pipe_fence_handle *fence,
 400                           uint64_t timeout)
 401 {
 402         struct panfrost_screen *screen = pan_screen(pscreen);
 403         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 404         int ret;
 405
 406         unsigned syncobj;
 407         ret = drmSyncobjCreate(screen->fd, 0, &syncobj);
 408         if (ret) {
 409                 fprintf(stderr, "Failed to create syncobj to wait on: %m\n");
 410                 return false;
 411         }
 412
 413         drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd);
 414         if (ret) {
 415                 fprintf(stderr, "Failed to import fence to syncobj: %m\n");
 416                 return false;
 417         }
 418
 419         uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
 420         if (abs_timeout == OS_TIMEOUT_INFINITE)
 421                 abs_timeout = INT64_MAX;
 422
 423         ret = drmSyncobjWait(screen->fd, &syncobj, 1, abs_timeout, 0, NULL);
 424
 425         drmSyncobjDestroy(screen->fd, syncobj);
 426
 427         return ret >= 0;
 428 }