1 /* 2 * Copyright (C) 2013 Red Hat 3 * Author: Rob Clark <robdclark@gmail.com> 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 as published by 7 * the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 12 * more details. 13 * 14 * You should have received a copy of the GNU General Public License along with 15 * this program. If not, see <http://www.gnu.org/licenses/>. 16 */ 17 18 #include <linux/sync_file.h> 19 20 #include "msm_drv.h" 21 #include "msm_gpu.h" 22 #include "msm_gem.h" 23 24 /* 25 * Cmdstream submission: 26 */ 27 28 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */ 29 #define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */ 30 #define BO_LOCKED 0x4000 31 #define BO_PINNED 0x2000 32 33 static struct msm_gem_submit *submit_create(struct drm_device *dev, 34 struct msm_gpu *gpu, struct msm_gpu_submitqueue *queue, 35 uint32_t nr_bos, uint32_t nr_cmds) 36 { 37 struct msm_gem_submit *submit; 38 uint64_t sz = sizeof(*submit) + ((u64)nr_bos * sizeof(submit->bos[0])) + 39 ((u64)nr_cmds * sizeof(submit->cmd[0])); 40 41 if (sz > SIZE_MAX) 42 return NULL; 43 44 submit = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); 45 if (!submit) 46 return NULL; 47 48 submit->dev = dev; 49 submit->gpu = gpu; 50 submit->fence = NULL; 51 submit->pid = get_pid(task_pid(current)); 52 submit->cmd = (void *)&submit->bos[nr_bos]; 53 submit->queue = queue; 54 submit->ring = gpu->rb[queue->prio]; 55 56 /* initially, until copy_from_user() and bo lookup succeeds: */ 57 submit->nr_bos = 0; 58 submit->nr_cmds = 0; 59 60 INIT_LIST_HEAD(&submit->node); 61 INIT_LIST_HEAD(&submit->bo_list); 62 ww_acquire_init(&submit->ticket, &reservation_ww_class); 63 64 return submit; 65 } 66 67 void msm_gem_submit_free(struct msm_gem_submit *submit) 68 { 69 dma_fence_put(submit->fence); 70 list_del(&submit->node); 71 put_pid(submit->pid); 72 msm_submitqueue_put(submit->queue); 73 74 kfree(submit); 75 } 76 77 static inline unsigned long __must_check 78 copy_from_user_inatomic(void *to, const void __user *from, unsigned long n) 79 { 80 if (access_ok(VERIFY_READ, from, n)) 81 return __copy_from_user_inatomic(to, from, n); 82 return -EFAULT; 83 } 84 85 static int submit_lookup_objects(struct msm_gem_submit *submit, 86 struct drm_msm_gem_submit *args, struct drm_file *file) 87 { 88 unsigned i; 89 int ret = 0; 90 91 spin_lock(&file->table_lock); 92 pagefault_disable(); 93 94 for (i = 0; i < args->nr_bos; i++) { 95 struct drm_msm_gem_submit_bo submit_bo; 96 struct drm_gem_object *obj; 97 struct msm_gem_object *msm_obj; 98 void __user *userptr = 99 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo))); 100 101 /* make sure we don't have garbage flags, in case we hit 102 * error path before flags is initialized: 103 */ 104 submit->bos[i].flags = 0; 105 106 if (copy_from_user_inatomic(&submit_bo, userptr, sizeof(submit_bo))) { 107 pagefault_enable(); 108 spin_unlock(&file->table_lock); 109 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) { 110 ret = -EFAULT; 111 goto out; 112 } 113 spin_lock(&file->table_lock); 114 pagefault_disable(); 115 } 116 117 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) || 118 !(submit_bo.flags & MSM_SUBMIT_BO_FLAGS)) { 119 DRM_ERROR("invalid flags: %x\n", submit_bo.flags); 120 ret = -EINVAL; 121 goto out_unlock; 122 } 123 124 submit->bos[i].flags = submit_bo.flags; 125 /* in validate_objects() we figure out if this is true: */ 126 submit->bos[i].iova = submit_bo.presumed; 127 128 /* normally use drm_gem_object_lookup(), but for bulk lookup 129 * all under single table_lock just hit object_idr directly: 130 */ 131 obj = idr_find(&file->object_idr, submit_bo.handle); 132 if (!obj) { 133 DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i); 134 ret = -EINVAL; 135 goto out_unlock; 136 } 137 138 msm_obj = to_msm_bo(obj); 139 140 if (!list_empty(&msm_obj->submit_entry)) { 141 DRM_ERROR("handle %u at index %u already on submit list\n", 142 submit_bo.handle, i); 143 ret = -EINVAL; 144 goto out_unlock; 145 } 146 147 drm_gem_object_get(obj); 148 149 submit->bos[i].obj = msm_obj; 150 151 list_add_tail(&msm_obj->submit_entry, &submit->bo_list); 152 } 153 154 out_unlock: 155 pagefault_enable(); 156 spin_unlock(&file->table_lock); 157 158 out: 159 submit->nr_bos = i; 160 161 return ret; 162 } 163 164 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, 165 int i, bool backoff) 166 { 167 struct msm_gem_object *msm_obj = submit->bos[i].obj; 168 169 if (submit->bos[i].flags & BO_PINNED) 170 msm_gem_put_iova(&msm_obj->base, submit->gpu->aspace); 171 172 if (submit->bos[i].flags & BO_LOCKED) 173 ww_mutex_unlock(&msm_obj->resv->lock); 174 175 if (backoff && !(submit->bos[i].flags & BO_VALID)) 176 submit->bos[i].iova = 0; 177 178 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED); 179 } 180 181 /* This is where we make sure all the bo's are reserved and pin'd: */ 182 static int submit_lock_objects(struct msm_gem_submit *submit) 183 { 184 int contended, slow_locked = -1, i, ret = 0; 185 186 retry: 187 for (i = 0; i < submit->nr_bos; i++) { 188 struct msm_gem_object *msm_obj = submit->bos[i].obj; 189 190 if (slow_locked == i) 191 slow_locked = -1; 192 193 contended = i; 194 195 if (!(submit->bos[i].flags & BO_LOCKED)) { 196 ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock, 197 &submit->ticket); 198 if (ret) 199 goto fail; 200 submit->bos[i].flags |= BO_LOCKED; 201 } 202 } 203 204 ww_acquire_done(&submit->ticket); 205 206 return 0; 207 208 fail: 209 for (; i >= 0; i--) 210 submit_unlock_unpin_bo(submit, i, true); 211 212 if (slow_locked > 0) 213 submit_unlock_unpin_bo(submit, slow_locked, true); 214 215 if (ret == -EDEADLK) { 216 struct msm_gem_object *msm_obj = submit->bos[contended].obj; 217 /* we lost out in a seqno race, lock and retry.. */ 218 ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock, 219 &submit->ticket); 220 if (!ret) { 221 submit->bos[contended].flags |= BO_LOCKED; 222 slow_locked = contended; 223 goto retry; 224 } 225 } 226 227 return ret; 228 } 229 230 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit) 231 { 232 int i, ret = 0; 233 234 for (i = 0; i < submit->nr_bos; i++) { 235 struct msm_gem_object *msm_obj = submit->bos[i].obj; 236 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE; 237 238 if (!write) { 239 /* NOTE: _reserve_shared() must happen before 240 * _add_shared_fence(), which makes this a slightly 241 * strange place to call it. OTOH this is a 242 * convenient can-fail point to hook it in. 243 */ 244 ret = reservation_object_reserve_shared(msm_obj->resv, 245 1); 246 if (ret) 247 return ret; 248 } 249 250 if (no_implicit) 251 continue; 252 253 ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx, 254 write); 255 if (ret) 256 break; 257 } 258 259 return ret; 260 } 261 262 static int submit_pin_objects(struct msm_gem_submit *submit) 263 { 264 int i, ret = 0; 265 266 submit->valid = true; 267 268 for (i = 0; i < submit->nr_bos; i++) { 269 struct msm_gem_object *msm_obj = submit->bos[i].obj; 270 uint64_t iova; 271 272 /* if locking succeeded, pin bo: */ 273 ret = msm_gem_get_iova(&msm_obj->base, 274 submit->gpu->aspace, &iova); 275 276 if (ret) 277 break; 278 279 submit->bos[i].flags |= BO_PINNED; 280 281 if (iova == submit->bos[i].iova) { 282 submit->bos[i].flags |= BO_VALID; 283 } else { 284 submit->bos[i].iova = iova; 285 /* iova changed, so address in cmdstream is not valid: */ 286 submit->bos[i].flags &= ~BO_VALID; 287 submit->valid = false; 288 } 289 } 290 291 return ret; 292 } 293 294 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx, 295 struct msm_gem_object **obj, uint64_t *iova, bool *valid) 296 { 297 if (idx >= submit->nr_bos) { 298 DRM_ERROR("invalid buffer index: %u (out of %u)\n", 299 idx, submit->nr_bos); 300 return -EINVAL; 301 } 302 303 if (obj) 304 *obj = submit->bos[idx].obj; 305 if (iova) 306 *iova = submit->bos[idx].iova; 307 if (valid) 308 *valid = !!(submit->bos[idx].flags & BO_VALID); 309 310 return 0; 311 } 312 313 /* process the reloc's and patch up the cmdstream as needed: */ 314 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj, 315 uint32_t offset, uint32_t nr_relocs, uint64_t relocs) 316 { 317 uint32_t i, last_offset = 0; 318 uint32_t *ptr; 319 int ret = 0; 320 321 if (offset % 4) { 322 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset); 323 return -EINVAL; 324 } 325 326 /* For now, just map the entire thing. Eventually we probably 327 * to do it page-by-page, w/ kmap() if not vmap()d.. 328 */ 329 ptr = msm_gem_get_vaddr(&obj->base); 330 331 if (IS_ERR(ptr)) { 332 ret = PTR_ERR(ptr); 333 DBG("failed to map: %d", ret); 334 return ret; 335 } 336 337 for (i = 0; i < nr_relocs; i++) { 338 struct drm_msm_gem_submit_reloc submit_reloc; 339 void __user *userptr = 340 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc))); 341 uint32_t off; 342 uint64_t iova; 343 bool valid; 344 345 if (copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc))) { 346 ret = -EFAULT; 347 goto out; 348 } 349 350 if (submit_reloc.submit_offset % 4) { 351 DRM_ERROR("non-aligned reloc offset: %u\n", 352 submit_reloc.submit_offset); 353 ret = -EINVAL; 354 goto out; 355 } 356 357 /* offset in dwords: */ 358 off = submit_reloc.submit_offset / 4; 359 360 if ((off >= (obj->base.size / 4)) || 361 (off < last_offset)) { 362 DRM_ERROR("invalid offset %u at reloc %u\n", off, i); 363 ret = -EINVAL; 364 goto out; 365 } 366 367 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid); 368 if (ret) 369 goto out; 370 371 if (valid) 372 continue; 373 374 iova += submit_reloc.reloc_offset; 375 376 if (submit_reloc.shift < 0) 377 iova >>= -submit_reloc.shift; 378 else 379 iova <<= submit_reloc.shift; 380 381 ptr[off] = iova | submit_reloc.or; 382 383 last_offset = off; 384 } 385 386 out: 387 msm_gem_put_vaddr(&obj->base); 388 389 return ret; 390 } 391 392 static void submit_cleanup(struct msm_gem_submit *submit) 393 { 394 unsigned i; 395 396 for (i = 0; i < submit->nr_bos; i++) { 397 struct msm_gem_object *msm_obj = submit->bos[i].obj; 398 submit_unlock_unpin_bo(submit, i, false); 399 list_del_init(&msm_obj->submit_entry); 400 drm_gem_object_put(&msm_obj->base); 401 } 402 403 ww_acquire_fini(&submit->ticket); 404 } 405 406 int msm_ioctl_gem_submit(struct drm_device *dev, void *data, 407 struct drm_file *file) 408 { 409 struct msm_drm_private *priv = dev->dev_private; 410 struct drm_msm_gem_submit *args = data; 411 struct msm_file_private *ctx = file->driver_priv; 412 struct msm_gem_submit *submit; 413 struct msm_gpu *gpu = priv->gpu; 414 struct dma_fence *in_fence = NULL; 415 struct sync_file *sync_file = NULL; 416 struct msm_gpu_submitqueue *queue; 417 struct msm_ringbuffer *ring; 418 int out_fence_fd = -1; 419 unsigned i; 420 int ret; 421 422 if (!gpu) 423 return -ENXIO; 424 425 /* for now, we just have 3d pipe.. eventually this would need to 426 * be more clever to dispatch to appropriate gpu module: 427 */ 428 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0) 429 return -EINVAL; 430 431 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS) 432 return -EINVAL; 433 434 if (args->flags & MSM_SUBMIT_SUDO) { 435 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) || 436 !capable(CAP_SYS_RAWIO)) 437 return -EINVAL; 438 } 439 440 queue = msm_submitqueue_get(ctx, args->queueid); 441 if (!queue) 442 return -ENOENT; 443 444 ring = gpu->rb[queue->prio]; 445 446 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) { 447 in_fence = sync_file_get_fence(args->fence_fd); 448 449 if (!in_fence) 450 return -EINVAL; 451 452 /* 453 * Wait if the fence is from a foreign context, or if the fence 454 * array contains any fence from a foreign context. 455 */ 456 if (!dma_fence_match_context(in_fence, ring->fctx->context)) { 457 ret = dma_fence_wait(in_fence, true); 458 if (ret) 459 return ret; 460 } 461 } 462 463 ret = mutex_lock_interruptible(&dev->struct_mutex); 464 if (ret) 465 return ret; 466 467 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 468 out_fence_fd = get_unused_fd_flags(O_CLOEXEC); 469 if (out_fence_fd < 0) { 470 ret = out_fence_fd; 471 goto out_unlock; 472 } 473 } 474 475 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds); 476 if (!submit) { 477 ret = -ENOMEM; 478 goto out_unlock; 479 } 480 481 if (args->flags & MSM_SUBMIT_SUDO) 482 submit->in_rb = true; 483 484 ret = submit_lookup_objects(submit, args, file); 485 if (ret) 486 goto out; 487 488 ret = submit_lock_objects(submit); 489 if (ret) 490 goto out; 491 492 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT)); 493 if (ret) 494 goto out; 495 496 ret = submit_pin_objects(submit); 497 if (ret) 498 goto out; 499 500 for (i = 0; i < args->nr_cmds; i++) { 501 struct drm_msm_gem_submit_cmd submit_cmd; 502 void __user *userptr = 503 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd))); 504 struct msm_gem_object *msm_obj; 505 uint64_t iova; 506 507 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd)); 508 if (ret) { 509 ret = -EFAULT; 510 goto out; 511 } 512 513 /* validate input from userspace: */ 514 switch (submit_cmd.type) { 515 case MSM_SUBMIT_CMD_BUF: 516 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 517 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 518 break; 519 default: 520 DRM_ERROR("invalid type: %08x\n", submit_cmd.type); 521 ret = -EINVAL; 522 goto out; 523 } 524 525 ret = submit_bo(submit, submit_cmd.submit_idx, 526 &msm_obj, &iova, NULL); 527 if (ret) 528 goto out; 529 530 if (submit_cmd.size % 4) { 531 DRM_ERROR("non-aligned cmdstream buffer size: %u\n", 532 submit_cmd.size); 533 ret = -EINVAL; 534 goto out; 535 } 536 537 if (!submit_cmd.size || 538 ((submit_cmd.size + submit_cmd.submit_offset) > 539 msm_obj->base.size)) { 540 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size); 541 ret = -EINVAL; 542 goto out; 543 } 544 545 submit->cmd[i].type = submit_cmd.type; 546 submit->cmd[i].size = submit_cmd.size / 4; 547 submit->cmd[i].iova = iova + submit_cmd.submit_offset; 548 submit->cmd[i].idx = submit_cmd.submit_idx; 549 550 if (submit->valid) 551 continue; 552 553 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset, 554 submit_cmd.nr_relocs, submit_cmd.relocs); 555 if (ret) 556 goto out; 557 } 558 559 submit->nr_cmds = i; 560 561 submit->fence = msm_fence_alloc(ring->fctx); 562 if (IS_ERR(submit->fence)) { 563 ret = PTR_ERR(submit->fence); 564 submit->fence = NULL; 565 goto out; 566 } 567 568 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 569 sync_file = sync_file_create(submit->fence); 570 if (!sync_file) { 571 ret = -ENOMEM; 572 goto out; 573 } 574 } 575 576 msm_gpu_submit(gpu, submit, ctx); 577 578 args->fence = submit->fence->seqno; 579 580 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 581 fd_install(out_fence_fd, sync_file->file); 582 args->fence_fd = out_fence_fd; 583 } 584 585 out: 586 if (in_fence) 587 dma_fence_put(in_fence); 588 submit_cleanup(submit); 589 if (ret) 590 msm_gem_submit_free(submit); 591 out_unlock: 592 if (ret && (out_fence_fd >= 0)) 593 put_unused_fd(out_fence_fd); 594 mutex_unlock(&dev->struct_mutex); 595 return ret; 596 } 597