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_trace.h" 40 41struct panfrost_drm { 42 struct panfrost_driver base; 43 int fd; 44}; 45 46static void 47panfrost_drm_allocate_slab(struct panfrost_screen *screen, 48 struct panfrost_memory *mem, 49 size_t pages, 50 bool same_va, 51 int extra_flags, 52 int commit_count, 53 int extent) 54{ 55 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 56 struct drm_panfrost_create_bo create_bo = { 57 .size = pages * 4096, 58 .flags = 0, // TODO figure out proper flags.. 59 }; 60 struct drm_panfrost_mmap_bo mmap_bo = {0,}; 61 int ret; 62 63 // TODO cache allocations 64 // TODO properly handle errors 65 // TODO take into account extra_flags 66 67 ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo); 68 if (ret) { 69 fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %d\n", ret); 70 assert(0); 71 } 72 73 mem->gpu = create_bo.offset; 74 mem->gem_handle = create_bo.handle; 75 mem->stack_bottom = 0; 76 mem->size = create_bo.size; 77 78 // TODO map and unmap on demand? 79 mmap_bo.handle = create_bo.handle; 80 ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo); 81 if (ret) { 82 fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret); 83 assert(0); 84 } 85 86 mem->cpu = os_mmap(NULL, mem->size, PROT_READ | PROT_WRITE, MAP_SHARED, 87 drm->fd, mmap_bo.offset); 88 if (mem->cpu == MAP_FAILED) { 89 fprintf(stderr, "mmap failed: %p\n", mem->cpu); 90 assert(0); 91 } 92 93 /* Record the mmap if we're tracing */ 94 if (!(extra_flags & PAN_ALLOCATE_GROWABLE)) 95 pantrace_mmap(mem->gpu, mem->cpu, mem->size, NULL); 96} 97 98static void 99panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem) 100{ 101 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 102 struct drm_gem_close gem_close = { 103 .handle = mem->gem_handle, 104 }; 105 int ret; 106 107 if (os_munmap((void *) (uintptr_t) mem->cpu, mem->size)) { 108 perror("munmap"); 109 abort(); 110 } 111 112 mem->cpu = NULL; 113 114 ret = drmIoctl(drm->fd, DRM_IOCTL_GEM_CLOSE, &gem_close); 115 if (ret) { 116 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret); 117 assert(0); 118 } 119 120 mem->gem_handle = -1; 121} 122 123static struct panfrost_bo * 124panfrost_drm_import_bo(struct panfrost_screen *screen, struct winsys_handle *whandle) 125{ 126 struct panfrost_bo *bo = CALLOC_STRUCT(panfrost_bo); 127 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 128 struct drm_panfrost_get_bo_offset get_bo_offset = {0,}; 129 struct drm_panfrost_mmap_bo mmap_bo = {0,}; 130 int ret; 131 unsigned gem_handle; 132 133 ret = drmPrimeFDToHandle(drm->fd, whandle->handle, &gem_handle); 134 assert(!ret); 135 136 get_bo_offset.handle = gem_handle; 137 ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset); 138 assert(!ret); 139 140 bo->gem_handle = gem_handle; 141 bo->gpu = (mali_ptr) get_bo_offset.offset; 142 pipe_reference_init(&bo->reference, 1); 143 144 // TODO map and unmap on demand? 145 mmap_bo.handle = gem_handle; 146 ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo); 147 if (ret) { 148 fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret); 149 assert(0); 150 } 151 152 bo->size = lseek(whandle->handle, 0, SEEK_END); 153 assert(bo->size > 0); 154 bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED, 155 drm->fd, mmap_bo.offset); 156 if (bo->cpu == MAP_FAILED) { 157 fprintf(stderr, "mmap failed: %p\n", bo->cpu); 158 assert(0); 159 } 160 161 /* Record the mmap if we're tracing */ 162 pantrace_mmap(bo->gpu, bo->cpu, bo->size, NULL); 163 164 return bo; 165} 166 167static int 168panfrost_drm_export_bo(struct panfrost_screen *screen, int gem_handle, unsigned int stride, struct winsys_handle *whandle) 169{ 170 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 171 struct drm_prime_handle args = { 172 .handle = gem_handle, 173 .flags = DRM_CLOEXEC, 174 }; 175 176 int ret = drmIoctl(drm->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args); 177 if (ret == -1) 178 return FALSE; 179 180 whandle->handle = args.fd; 181 whandle->stride = stride; 182 183 return TRUE; 184} 185 186static void 187panfrost_drm_free_imported_bo(struct panfrost_screen *screen, struct panfrost_bo *bo) 188{ 189 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 190 struct drm_gem_close gem_close = { 191 .handle = bo->gem_handle, 192 }; 193 int ret; 194 195 ret = drmIoctl(drm->fd, DRM_IOCTL_GEM_CLOSE, &gem_close); 196 if (ret) { 197 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret); 198 assert(0); 199 } 200 201 bo->gem_handle = -1; 202 bo->gpu = (mali_ptr)NULL; 203} 204 205static int 206panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs, struct pipe_surface *surf) 207{ 208 struct pipe_context *gallium = (struct pipe_context *) ctx; 209 struct panfrost_screen *screen = pan_screen(gallium->screen); 210 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 211 struct drm_panfrost_submit submit = {0,}; 212 int bo_handles[7]; 213 214 submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync; 215 submit.in_sync_count = 1; 216 217 submit.out_sync = ctx->out_sync; 218 219 submit.jc = job_desc; 220 submit.requirements = reqs; 221 222 if (surf) { 223 struct panfrost_resource *res = pan_resource(surf->texture); 224 assert(res->bo->gem_handle > 0); 225 bo_handles[submit.bo_handle_count++] = res->bo->gem_handle; 226 227 if (res->bo->checksum_slab.gem_handle) 228 bo_handles[submit.bo_handle_count++] = res->bo->checksum_slab.gem_handle; 229 } 230 231 /* TODO: Add here the transient pools */ 232 /* TODO: Add here the BOs listed in the panfrost_job */ 233 bo_handles[submit.bo_handle_count++] = ctx->shaders.gem_handle; 234 bo_handles[submit.bo_handle_count++] = ctx->scratchpad.gem_handle; 235 bo_handles[submit.bo_handle_count++] = ctx->tiler_heap.gem_handle; 236 bo_handles[submit.bo_handle_count++] = ctx->varying_mem.gem_handle; 237 bo_handles[submit.bo_handle_count++] = ctx->misc_0.gem_handle; 238 submit.bo_handles = (u64) (uintptr_t) bo_handles; 239 240 /* Dump memory _before_ submitting so we're not corrupted with actual GPU results */ 241 pantrace_dump_memory(); 242 243 if (drmIoctl(drm->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit)) { 244 fprintf(stderr, "Error submitting: %m\n"); 245 return errno; 246 } 247 248 /* Trace the job if we're doing that and do a memory dump. We may 249 * want to adjust this logic once we're ready to trace FBOs */ 250 pantrace_submit_job(submit.jc, submit.requirements, FALSE); 251 252 return 0; 253} 254 255static int 256panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout) 257{ 258 struct pipe_surface *surf = ctx->pipe_framebuffer.cbufs[0]; 259 int ret; 260 261 if (has_draws) { 262 ret = panfrost_drm_submit_job(ctx, ctx->set_value_job, 0, NULL); 263 assert(!ret); 264 } 265 266 ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx), PANFROST_JD_REQ_FS, surf); 267 268 return ret; 269} 270 271static struct panfrost_fence * 272panfrost_fence_create(struct panfrost_context *ctx) 273{ 274 struct pipe_context *gallium = (struct pipe_context *) ctx; 275 struct panfrost_screen *screen = pan_screen(gallium->screen); 276 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 277 struct panfrost_fence *f = calloc(1, sizeof(*f)); 278 if (!f) 279 return NULL; 280 281 /* Snapshot the last Panfrost's rendering's out fence. We'd rather have 282 * another syncobj instead of a sync file, but this is all we get. 283 * (HandleToFD/FDToHandle just gives you another syncobj ID for the 284 * same syncobj). 285 */ 286 drmSyncobjExportSyncFile(drm->fd, ctx->out_sync, &f->fd); 287 if (f->fd == -1) { 288 fprintf(stderr, "export failed\n"); 289 free(f); 290 return NULL; 291 } 292 293 pipe_reference_init(&f->reference, 1); 294 295 return f; 296} 297 298static void 299panfrost_drm_force_flush_fragment(struct panfrost_context *ctx, 300 struct pipe_fence_handle **fence) 301{ 302 struct pipe_context *gallium = (struct pipe_context *) ctx; 303 struct panfrost_screen *screen = pan_screen(gallium->screen); 304 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 305 306 if (!screen->last_fragment_flushed) { 307 drmSyncobjWait(drm->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL); 308 screen->last_fragment_flushed = true; 309 310 /* The job finished up, so we're safe to clean it up now */ 311 panfrost_free_job(ctx, screen->last_job); 312 } 313 314 if (fence) { 315 struct panfrost_fence *f = panfrost_fence_create(ctx); 316 gallium->screen->fence_reference(gallium->screen, fence, NULL); 317 *fence = (struct pipe_fence_handle *)f; 318 } 319} 320 321static void 322panfrost_drm_enable_counters(struct panfrost_screen *screen) 323{ 324 fprintf(stderr, "unimplemented: %s\n", __func__); 325} 326 327static void 328panfrost_drm_dump_counters(struct panfrost_screen *screen) 329{ 330 fprintf(stderr, "unimplemented: %s\n", __func__); 331} 332 333static unsigned 334panfrost_drm_query_gpu_version(struct panfrost_screen *screen) 335{ 336 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 337 struct drm_panfrost_get_param get_param = {0,}; 338 int ret; 339 340 get_param.param = DRM_PANFROST_PARAM_GPU_ID; 341 ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param); 342 assert(!ret); 343 344 return get_param.value; 345} 346 347static int 348panfrost_drm_init_context(struct panfrost_context *ctx) 349{ 350 struct pipe_context *gallium = (struct pipe_context *) ctx; 351 struct panfrost_screen *screen = pan_screen(gallium->screen); 352 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 353 354 return drmSyncobjCreate(drm->fd, DRM_SYNCOBJ_CREATE_SIGNALED, 355 &ctx->out_sync); 356} 357 358static void 359panfrost_drm_fence_reference(struct pipe_screen *screen, 360 struct pipe_fence_handle **ptr, 361 struct pipe_fence_handle *fence) 362{ 363 struct panfrost_fence **p = (struct panfrost_fence **)ptr; 364 struct panfrost_fence *f = (struct panfrost_fence *)fence; 365 struct panfrost_fence *old = *p; 366 367 if (pipe_reference(&(*p)->reference, &f->reference)) { 368 close(old->fd); 369 free(old); 370 } 371 *p = f; 372} 373 374static boolean 375panfrost_drm_fence_finish(struct pipe_screen *pscreen, 376 struct pipe_context *ctx, 377 struct pipe_fence_handle *fence, 378 uint64_t timeout) 379{ 380 struct panfrost_screen *screen = pan_screen(pscreen); 381 struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver; 382 struct panfrost_fence *f = (struct panfrost_fence *)fence; 383 int ret; 384 385 unsigned syncobj; 386 ret = drmSyncobjCreate(drm->fd, 0, &syncobj); 387 if (ret) { 388 fprintf(stderr, "Failed to create syncobj to wait on: %m\n"); 389 return false; 390 } 391 392 drmSyncobjImportSyncFile(drm->fd, syncobj, f->fd); 393 if (ret) { 394 fprintf(stderr, "Failed to import fence to syncobj: %m\n"); 395 return false; 396 } 397 398 uint64_t abs_timeout = os_time_get_absolute_timeout(timeout); 399 if (abs_timeout == OS_TIMEOUT_INFINITE) 400 abs_timeout = INT64_MAX; 401 402 ret = drmSyncobjWait(drm->fd, &syncobj, 1, abs_timeout, 0, NULL); 403 404 drmSyncobjDestroy(drm->fd, syncobj); 405 406 return ret >= 0; 407} 408 409struct panfrost_driver * 410panfrost_create_drm_driver(int fd) 411{ 412 struct panfrost_drm *driver = CALLOC_STRUCT(panfrost_drm); 413 414 driver->fd = fd; 415 416 driver->base.import_bo = panfrost_drm_import_bo; 417 driver->base.export_bo = panfrost_drm_export_bo; 418 driver->base.free_imported_bo = panfrost_drm_free_imported_bo; 419 driver->base.submit_vs_fs_job = panfrost_drm_submit_vs_fs_job; 420 driver->base.force_flush_fragment = panfrost_drm_force_flush_fragment; 421 driver->base.allocate_slab = panfrost_drm_allocate_slab; 422 driver->base.free_slab = panfrost_drm_free_slab; 423 driver->base.enable_counters = panfrost_drm_enable_counters; 424 driver->base.query_gpu_version = panfrost_drm_query_gpu_version; 425 driver->base.init_context = panfrost_drm_init_context; 426 driver->base.fence_reference = panfrost_drm_fence_reference; 427 driver->base.fence_finish = panfrost_drm_fence_finish; 428 driver->base.dump_counters = panfrost_drm_dump_counters; 429 430 return &driver->base; 431} 432