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