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         if (screen->kernel_version->version_major > 1 ||
  97             screen->kernel_version->version_minor >= 1) {
  98                 if (!(flags & PAN_ALLOCATE_EXECUTE))
  99                         translated_flags |= PANFROST_BO_NOEXEC;
 100         }
 101
 102         struct drm_panfrost_create_bo create_bo = {
 103                 .size = size,
 104                 .flags = translated_flags,
 105         };
 106
 107         /* Before creating a BO, we first want to check the cache */
 108
 109         bo = panfrost_bo_cache_fetch(screen, size, flags);
 110
 111         if (bo == NULL) {
 112                 /* Otherwise, the cache misses and we need to allocate a BO fresh from
 113                  * the kernel */
 114
 115                 int ret;
 116
 117                 ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
 118                 if (ret) {
 119                         fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %d\n", ret);
 120                         assert(0);
 121                 }
 122
 123                 /* We have a BO allocated from the kernel; fill in the userspace
 124                  * version */
 125
 126                 bo = rzalloc(screen, struct panfrost_bo);
 127                 bo->size = create_bo.size;
 128                 bo->gpu = create_bo.offset;
 129                 bo->gem_handle = create_bo.handle;
 130                 bo->flags = flags;
 131         }
 132
 133         /* Only mmap now if we know we need to. For CPU-invisible buffers, we
 134          * never map since we don't care about their contents; they're purely
 135          * for GPU-internal use. But we do trace them anyway. */
 136
 137         if (!(flags & (PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_DELAY_MMAP)))
 138                 panfrost_drm_mmap_bo(screen, bo);
 139         else if (flags & PAN_ALLOCATE_INVISIBLE) {
 140                 if (pan_debug & PAN_DBG_TRACE)
 141                         pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL);
 142         }
 143
 144         pipe_reference_init(&bo->reference, 1);
 145         return bo;
 146 }
 147
 148 void
 149 panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable)
 150 {
 151         struct drm_gem_close gem_close = { .handle = bo->gem_handle };
 152         int ret;
 153
 154         if (!bo)
 155                 return;
 156
 157         /* Rather than freeing the BO now, we'll cache the BO for later
 158          * allocations if we're allowed to */
 159
 160         if (cacheable) {
 161                 bool cached = panfrost_bo_cache_put(screen, bo);
 162
 163                 if (cached)
 164                         return;
 165         }
 166
 167         /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */
 168
 169         panfrost_drm_munmap_bo(screen, bo);
 170
 171         ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
 172         if (ret) {
 173                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
 174                 assert(0);
 175         }
 176
 177         ralloc_free(bo);
 178 }
 179
 180 void
 181 panfrost_drm_allocate_slab(struct panfrost_screen *screen,
 182                            struct panfrost_memory *mem,
 183                            size_t pages,
 184                            bool same_va,
 185                            int extra_flags,
 186                            int commit_count,
 187                            int extent)
 188 {
 189         // TODO cache allocations
 190         // TODO properly handle errors
 191         // TODO take into account extra_flags
 192         mem->bo = panfrost_drm_create_bo(screen, pages * 4096, extra_flags);
 193         mem->stack_bottom = 0;
 194 }
 195
 196 void
 197 panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
 198 {
 199         panfrost_bo_unreference(&screen->base, mem->bo);
 200         mem->bo = NULL;
 201 }
 202
 203 struct panfrost_bo *
 204 panfrost_drm_import_bo(struct panfrost_screen *screen, int fd)
 205 {
 206         struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
 207         struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
 208         ASSERTED int ret;
 209         unsigned gem_handle;
 210
 211         ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle);
 212         assert(!ret);
 213
 214         get_bo_offset.handle = gem_handle;
 215         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
 216         assert(!ret);
 217
 218         bo->gem_handle = gem_handle;
 219         bo->gpu = (mali_ptr) get_bo_offset.offset;
 220         bo->size = lseek(fd, 0, SEEK_END);
 221         assert(bo->size > 0);
 222         pipe_reference_init(&bo->reference, 1);
 223
 224         // TODO map and unmap on demand?
 225         panfrost_drm_mmap_bo(screen, bo);
 226         return bo;
 227 }
 228
 229 int
 230 panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo)
 231 {
 232         struct drm_prime_handle args = {
 233                 .handle = bo->gem_handle,
 234                 .flags = DRM_CLOEXEC,
 235         };
 236
 237         int ret = drmIoctl(screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 238         if (ret == -1)
 239                 return -1;
 240
 241         return args.fd;
 242 }
 243
 244 static int
 245 panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs)
 246 {
 247         struct pipe_context *gallium = (struct pipe_context *) ctx;
 248         struct panfrost_screen *screen = pan_screen(gallium->screen);
 249         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 250         struct drm_panfrost_submit submit = {0,};
 251         int *bo_handles, ret;
 252
 253         submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync;
 254         submit.in_sync_count = 1;
 255
 256         submit.out_sync = ctx->out_sync;
 257
 258         submit.jc = job_desc;
 259         submit.requirements = reqs;
 260
 261         bo_handles = calloc(job->bos->entries, sizeof(*bo_handles));
 262         assert(bo_handles);
 263
 264         set_foreach(job->bos, entry) {
 265                 struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
 266                 assert(bo->gem_handle > 0);
 267                 bo_handles[submit.bo_handle_count++] = bo->gem_handle;
 268         }
 269
 270         submit.bo_handles = (u64) (uintptr_t) bo_handles;
 271         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
 272         free(bo_handles);
 273         if (ret) {
 274                 fprintf(stderr, "Error submitting: %m\n");
 275                 return errno;
 276         }
 277
 278         /* Trace the job if we're doing that */
 279         if (pan_debug & PAN_DBG_TRACE) {
 280                 /* Wait so we can get errors reported back */
 281                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 282                 pandecode_jc(submit.jc, FALSE);
 283         }
 284
 285         return 0;
 286 }
 287
 288 int
 289 panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout)
 290 {
 291         int ret = 0;
 292
 293         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 294
 295         /* TODO: Add here the transient pools */
 296         panfrost_job_add_bo(job, ctx->scratchpad.bo);
 297         panfrost_job_add_bo(job, ctx->tiler_heap.bo);
 298         panfrost_job_add_bo(job, job->polygon_list);
 299
 300         if (job->first_job.gpu) {
 301                 ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0);
 302                 assert(!ret);
 303         }
 304
 305         if (job->first_tiler.gpu || job->clear) {
 306                 ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx, has_draws), PANFROST_JD_REQ_FS);
 307                 assert(!ret);
 308         }
 309
 310         return ret;
 311 }
 312
 313 static struct panfrost_fence *
 314 panfrost_fence_create(struct panfrost_context *ctx)
 315 {
 316         struct pipe_context *gallium = (struct pipe_context *) ctx;
 317         struct panfrost_screen *screen = pan_screen(gallium->screen);
 318         struct panfrost_fence *f = calloc(1, sizeof(*f));
 319         if (!f)
 320                 return NULL;
 321
 322         /* Snapshot the last Panfrost's rendering's out fence.  We'd rather have
 323          * another syncobj instead of a sync file, but this is all we get.
 324          * (HandleToFD/FDToHandle just gives you another syncobj ID for the
 325          * same syncobj).
 326          */
 327         drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd);
 328         if (f->fd == -1) {
 329                 fprintf(stderr, "export failed\n");
 330                 free(f);
 331                 return NULL;
 332         }
 333
 334         pipe_reference_init(&f->reference, 1);
 335
 336         return f;
 337 }
 338
 339 void
 340 panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
 341                                   struct pipe_fence_handle **fence)
 342 {
 343         struct pipe_context *gallium = (struct pipe_context *) ctx;
 344         struct panfrost_screen *screen = pan_screen(gallium->screen);
 345
 346         if (!screen->last_fragment_flushed) {
 347                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 348                 screen->last_fragment_flushed = true;
 349
 350                 /* The job finished up, so we're safe to clean it up now */
 351                 panfrost_free_job(ctx, screen->last_job);
 352         }
 353
 354         if (fence) {
 355                 struct panfrost_fence *f = panfrost_fence_create(ctx);
 356                 gallium->screen->fence_reference(gallium->screen, fence, NULL);
 357                 *fence = (struct pipe_fence_handle *)f;
 358         }
 359 }
 360
 361 unsigned
 362 panfrost_drm_query_gpu_version(struct panfrost_screen *screen)
 363 {
 364         struct drm_panfrost_get_param get_param = {0,};
 365         ASSERTED int ret;
 366
 367         get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
 368         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
 369         assert(!ret);
 370
 371         return get_param.value;
 372 }
 373
 374 int
 375 panfrost_drm_init_context(struct panfrost_context *ctx)
 376 {
 377         struct pipe_context *gallium = (struct pipe_context *) ctx;
 378         struct panfrost_screen *screen = pan_screen(gallium->screen);
 379
 380         return drmSyncobjCreate(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED,
 381                                 &ctx->out_sync);
 382 }
 383
 384 void
 385 panfrost_drm_fence_reference(struct pipe_screen *screen,
 386                              struct pipe_fence_handle **ptr,
 387                              struct pipe_fence_handle *fence)
 388 {
 389         struct panfrost_fence **p = (struct panfrost_fence **)ptr;
 390         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 391         struct panfrost_fence *old = *p;
 392
 393         if (pipe_reference(&(*p)->reference, &f->reference)) {
 394                 close(old->fd);
 395                 free(old);
 396         }
 397         *p = f;
 398 }
 399
 400 boolean
 401 panfrost_drm_fence_finish(struct pipe_screen *pscreen,
 402                           struct pipe_context *ctx,
 403                           struct pipe_fence_handle *fence,
 404                           uint64_t timeout)
 405 {
 406         struct panfrost_screen *screen = pan_screen(pscreen);
 407         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 408         int ret;
 409
 410         unsigned syncobj;
 411         ret = drmSyncobjCreate(screen->fd, 0, &syncobj);
 412         if (ret) {
 413                 fprintf(stderr, "Failed to create syncobj to wait on: %m\n");
 414                 return false;
 415         }
 416
 417         drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd);
 418         if (ret) {
 419                 fprintf(stderr, "Failed to import fence to syncobj: %m\n");
 420                 return false;
 421         }
 422
 423         uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
 424         if (abs_timeout == OS_TIMEOUT_INFINITE)
 425                 abs_timeout = INT64_MAX;
 426
 427         ret = drmSyncobjWait(screen->fd, &syncobj, 1, abs_timeout, 0, NULL);
 428
 429         drmSyncobjDestroy(screen->fd, syncobj);
 430
 431         return ret >= 0;
 432 }