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_drm.h"
  39 #include "pan_util.h"
  40 #include "pandecode/decode.h"
  41
  42 void
  43 panfrost_drm_allocate_slab(struct panfrost_screen *screen,
  44                            struct panfrost_memory *mem,
  45                            size_t pages,
  46                            bool same_va,
  47                            int extra_flags,
  48                            int commit_count,
  49                            int extent)
  50 {
  51         struct drm_panfrost_create_bo create_bo = {
  52                         .size = pages * 4096,
  53                         .flags = 0,  // TODO figure out proper flags..
  54         };
  55         struct drm_panfrost_mmap_bo mmap_bo = {0,};
  56         int ret;
  57
  58         // TODO cache allocations
  59         // TODO properly handle errors
  60         // TODO take into account extra_flags
  61
  62         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
  63         if (ret) {
  64                 fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %d\n", ret);
  65                 assert(0);
  66         }
  67
  68         mem->gpu = create_bo.offset;
  69         mem->gem_handle = create_bo.handle;
  70         mem->stack_bottom = 0;
  71         mem->size = create_bo.size;
  72
  73         // TODO map and unmap on demand?
  74         mmap_bo.handle = create_bo.handle;
  75         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
  76         if (ret) {
  77                 fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret);
  78                 assert(0);
  79         }
  80
  81         mem->cpu = os_mmap(NULL, mem->size, PROT_READ | PROT_WRITE, MAP_SHARED,
  82                        screen->fd, mmap_bo.offset);
  83         if (mem->cpu == MAP_FAILED) {
  84                 fprintf(stderr, "mmap failed: %p\n", mem->cpu);
  85                 assert(0);
  86         }
  87
  88         /* Record the mmap if we're tracing */
  89         if (pan_debug & PAN_DBG_TRACE)
  90                 pandecode_inject_mmap(mem->gpu, mem->cpu, mem->size, NULL);
  91 }
  92
  93 void
  94 panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
  95 {
  96         struct drm_gem_close gem_close = {
  97                 .handle = mem->gem_handle,
  98         };
  99         int ret;
 100
 101         if (os_munmap((void *) (uintptr_t) mem->cpu, mem->size)) {
 102                 perror("munmap");
 103                 abort();
 104         }
 105
 106         mem->cpu = NULL;
 107
 108         ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
 109         if (ret) {
 110                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
 111                 assert(0);
 112         }
 113
 114         mem->gem_handle = -1;
 115 }
 116
 117 struct panfrost_bo *
 118 panfrost_drm_import_bo(struct panfrost_screen *screen, struct winsys_handle *whandle)
 119 {
 120         struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
 121         struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
 122         struct drm_panfrost_mmap_bo mmap_bo = {0,};
 123         int ret;
 124         unsigned gem_handle;
 125
 126         ret = drmPrimeFDToHandle(screen->fd, whandle->handle, &gem_handle);
 127         assert(!ret);
 128
 129         get_bo_offset.handle = gem_handle;
 130         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
 131         assert(!ret);
 132
 133         bo->gem_handle = gem_handle;
 134         bo->gpu = (mali_ptr) get_bo_offset.offset;
 135         pipe_reference_init(&bo->reference, 1);
 136
 137         // TODO map and unmap on demand?
 138         mmap_bo.handle = gem_handle;
 139         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
 140         if (ret) {
 141                 fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret);
 142                 assert(0);
 143         }
 144
 145         bo->size = lseek(whandle->handle, 0, SEEK_END);
 146         assert(bo->size > 0);
 147         bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
 148                        screen->fd, mmap_bo.offset);
 149         if (bo->cpu == MAP_FAILED) {
 150                 fprintf(stderr, "mmap failed: %p\n", bo->cpu);
 151                 assert(0);
 152         }
 153
 154         /* Record the mmap if we're tracing */
 155         if (pan_debug & PAN_DBG_TRACE)
 156                 pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL);
 157
 158         return bo;
 159 }
 160
 161 int
 162 panfrost_drm_export_bo(struct panfrost_screen *screen, int gem_handle, unsigned int stride, struct winsys_handle *whandle)
 163 {
 164         struct drm_prime_handle args = {
 165                 .handle = gem_handle,
 166                 .flags = DRM_CLOEXEC,
 167         };
 168
 169         int ret = drmIoctl(screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 170         if (ret == -1)
 171                 return FALSE;
 172
 173         whandle->handle = args.fd;
 174         whandle->stride = stride;
 175
 176         return TRUE;
 177 }
 178
 179 void
 180 panfrost_drm_free_imported_bo(struct panfrost_screen *screen, struct panfrost_bo *bo)
 181 {
 182         struct drm_gem_close gem_close = {
 183                 .handle = bo->gem_handle,
 184         };
 185         int ret;
 186
 187         ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
 188         if (ret) {
 189                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
 190                 assert(0);
 191         }
 192
 193         bo->gem_handle = -1;
 194         bo->gpu = (mali_ptr)NULL;
 195 }
 196
 197 int
 198 panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs, struct pipe_surface *surf)
 199 {
 200         struct pipe_context *gallium = (struct pipe_context *) ctx;
 201         struct panfrost_screen *screen = pan_screen(gallium->screen);
 202         struct drm_panfrost_submit submit = {0,};
 203         int bo_handles[7];
 204
 205         submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync;
 206         submit.in_sync_count = 1;
 207
 208         submit.out_sync = ctx->out_sync;
 209
 210         submit.jc = job_desc;
 211         submit.requirements = reqs;
 212
 213         if (surf) {
 214                 struct panfrost_resource *res = pan_resource(surf->texture);
 215                 assert(res->bo->gem_handle > 0);
 216                 bo_handles[submit.bo_handle_count++] = res->bo->gem_handle;
 217
 218                 if (res->bo->checksum_slab.gem_handle)
 219                         bo_handles[submit.bo_handle_count++] = res->bo->checksum_slab.gem_handle;
 220         }
 221
 222         /* TODO: Add here the transient pools */
 223         /* TODO: Add here the BOs listed in the panfrost_job */
 224         bo_handles[submit.bo_handle_count++] = ctx->shaders.gem_handle;
 225         bo_handles[submit.bo_handle_count++] = ctx->scratchpad.gem_handle;
 226         bo_handles[submit.bo_handle_count++] = ctx->tiler_heap.gem_handle;
 227         bo_handles[submit.bo_handle_count++] = ctx->varying_mem.gem_handle;
 228         bo_handles[submit.bo_handle_count++] = ctx->tiler_polygon_list.gem_handle;
 229         submit.bo_handles = (u64) (uintptr_t) bo_handles;
 230
 231         if (drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit)) {
 232                 fprintf(stderr, "Error submitting: %m\n");
 233                 return errno;
 234         }
 235
 236         /* Trace the job if we're doing that */
 237
 238         if (pan_debug & PAN_DBG_TRACE)
 239                 pandecode_replay_jc(submit.jc, FALSE);
 240
 241         return 0;
 242 }
 243
 244 int
 245 panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout)
 246 {
 247         struct pipe_surface *surf = ctx->pipe_framebuffer.cbufs[0];
 248         int ret;
 249
 250         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 251
 252         if (job->first_job.gpu) {
 253                 ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0, NULL);
 254                 assert(!ret);
 255         }
 256
 257         if (job->first_tiler.gpu || job->clear) {
 258                 ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx, has_draws), PANFROST_JD_REQ_FS, surf);
 259                 assert(!ret);
 260         }
 261
 262         return ret;
 263 }
 264
 265 struct panfrost_fence *
 266 panfrost_fence_create(struct panfrost_context *ctx)
 267 {
 268         struct pipe_context *gallium = (struct pipe_context *) ctx;
 269         struct panfrost_screen *screen = pan_screen(gallium->screen);
 270         struct panfrost_fence *f = calloc(1, sizeof(*f));
 271         if (!f)
 272                 return NULL;
 273
 274         /* Snapshot the last Panfrost's rendering's out fence.  We'd rather have
 275          * another syncobj instead of a sync file, but this is all we get.
 276          * (HandleToFD/FDToHandle just gives you another syncobj ID for the
 277          * same syncobj).
 278          */
 279         drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd);
 280         if (f->fd == -1) {
 281                 fprintf(stderr, "export failed\n");
 282                 free(f);
 283                 return NULL;
 284         }
 285
 286         pipe_reference_init(&f->reference, 1);
 287
 288         return f;
 289 }
 290
 291 void
 292 panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
 293                                   struct pipe_fence_handle **fence)
 294 {
 295         struct pipe_context *gallium = (struct pipe_context *) ctx;
 296         struct panfrost_screen *screen = pan_screen(gallium->screen);
 297
 298         if (!screen->last_fragment_flushed) {
 299                 drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
 300                 screen->last_fragment_flushed = true;
 301
 302                 /* The job finished up, so we're safe to clean it up now */
 303                 panfrost_free_job(ctx, screen->last_job);
 304         }
 305
 306         if (fence) {
 307                 struct panfrost_fence *f = panfrost_fence_create(ctx);
 308                 gallium->screen->fence_reference(gallium->screen, fence, NULL);
 309                 *fence = (struct pipe_fence_handle *)f;
 310         }
 311 }
 312
 313 unsigned
 314 panfrost_drm_query_gpu_version(struct panfrost_screen *screen)
 315 {
 316         struct drm_panfrost_get_param get_param = {0,};
 317         int ret;
 318
 319         get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
 320         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
 321         assert(!ret);
 322
 323         return get_param.value;
 324 }
 325
 326 int
 327 panfrost_drm_init_context(struct panfrost_context *ctx)
 328 {
 329         struct pipe_context *gallium = (struct pipe_context *) ctx;
 330         struct panfrost_screen *screen = pan_screen(gallium->screen);
 331
 332         return drmSyncobjCreate(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED,
 333                                 &ctx->out_sync);
 334 }
 335
 336 void
 337 panfrost_drm_fence_reference(struct pipe_screen *screen,
 338                          struct pipe_fence_handle **ptr,
 339                          struct pipe_fence_handle *fence)
 340 {
 341         struct panfrost_fence **p = (struct panfrost_fence **)ptr;
 342         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 343         struct panfrost_fence *old = *p;
 344
 345         if (pipe_reference(&(*p)->reference, &f->reference)) {
 346                 close(old->fd);
 347                 free(old);
 348         }
 349         *p = f;
 350 }
 351
 352 boolean
 353 panfrost_drm_fence_finish(struct pipe_screen *pscreen,
 354                       struct pipe_context *ctx,
 355                       struct pipe_fence_handle *fence,
 356                       uint64_t timeout)
 357 {
 358         struct panfrost_screen *screen = pan_screen(pscreen);
 359         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 360         int ret;
 361
 362         unsigned syncobj;
 363         ret = drmSyncobjCreate(screen->fd, 0, &syncobj);
 364         if (ret) {
 365                 fprintf(stderr, "Failed to create syncobj to wait on: %m\n");
 366                 return false;
 367         }
 368
 369         drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd);
 370         if (ret) {
 371                 fprintf(stderr, "Failed to import fence to syncobj: %m\n");
 372                 return false;
 373         }
 374
 375         uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
 376         if (abs_timeout == OS_TIMEOUT_INFINITE)
 377                 abs_timeout = INT64_MAX;
 378
 379         ret = drmSyncobjWait(screen->fd, &syncobj, 1, abs_timeout, 0, NULL);
 380
 381         drmSyncobjDestroy(screen->fd, syncobj);
 382
 383         return ret >= 0;
 384 }