1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2013 Red Hat 4 * Author: Rob Clark <robdclark@gmail.com> 5 */ 6 7 #include <linux/file.h> 8 #include <linux/sync_file.h> 9 #include <linux/uaccess.h> 10 11 #include <drm/drm_drv.h> 12 #include <drm/drm_file.h> 13 #include <drm/drm_syncobj.h> 14 15 #include "msm_drv.h" 16 #include "msm_gpu.h" 17 #include "msm_gem.h" 18 #include "msm_gpu_trace.h" 19 20 /* 21 * Cmdstream submission: 22 */ 23 24 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */ 25 #define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */ 26 #define BO_LOCKED 0x4000 27 #define BO_PINNED 0x2000 28 29 static struct msm_gem_submit *submit_create(struct drm_device *dev, 30 struct msm_gpu *gpu, 31 struct msm_gpu_submitqueue *queue, uint32_t nr_bos, 32 uint32_t nr_cmds) 33 { 34 struct msm_gem_submit *submit; 35 uint64_t sz = struct_size(submit, bos, nr_bos) + 36 ((u64)nr_cmds * sizeof(submit->cmd[0])); 37 38 if (sz > SIZE_MAX) 39 return NULL; 40 41 submit = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); 42 if (!submit) 43 return NULL; 44 45 kref_init(&submit->ref); 46 submit->dev = dev; 47 submit->aspace = queue->ctx->aspace; 48 submit->gpu = gpu; 49 submit->fence = NULL; 50 submit->cmd = (void *)&submit->bos[nr_bos]; 51 submit->queue = queue; 52 submit->ring = gpu->rb[queue->prio]; 53 54 /* initially, until copy_from_user() and bo lookup succeeds: */ 55 submit->nr_bos = 0; 56 submit->nr_cmds = 0; 57 58 INIT_LIST_HEAD(&submit->node); 59 INIT_LIST_HEAD(&submit->bo_list); 60 61 return submit; 62 } 63 64 void __msm_gem_submit_destroy(struct kref *kref) 65 { 66 struct msm_gem_submit *submit = 67 container_of(kref, struct msm_gem_submit, ref); 68 unsigned i; 69 70 dma_fence_put(submit->fence); 71 put_pid(submit->pid); 72 msm_submitqueue_put(submit->queue); 73 74 for (i = 0; i < submit->nr_cmds; i++) 75 kfree(submit->cmd[i].relocs); 76 77 kfree(submit); 78 } 79 80 static int submit_lookup_objects(struct msm_gem_submit *submit, 81 struct drm_msm_gem_submit *args, struct drm_file *file) 82 { 83 unsigned i; 84 int ret = 0; 85 86 for (i = 0; i < args->nr_bos; i++) { 87 struct drm_msm_gem_submit_bo submit_bo; 88 void __user *userptr = 89 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo))); 90 91 /* make sure we don't have garbage flags, in case we hit 92 * error path before flags is initialized: 93 */ 94 submit->bos[i].flags = 0; 95 96 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) { 97 ret = -EFAULT; 98 i = 0; 99 goto out; 100 } 101 102 /* at least one of READ and/or WRITE flags should be set: */ 103 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE) 104 105 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) || 106 !(submit_bo.flags & MANDATORY_FLAGS)) { 107 DRM_ERROR("invalid flags: %x\n", submit_bo.flags); 108 ret = -EINVAL; 109 i = 0; 110 goto out; 111 } 112 113 submit->bos[i].handle = submit_bo.handle; 114 submit->bos[i].flags = submit_bo.flags; 115 /* in validate_objects() we figure out if this is true: */ 116 submit->bos[i].iova = submit_bo.presumed; 117 } 118 119 spin_lock(&file->table_lock); 120 121 for (i = 0; i < args->nr_bos; i++) { 122 struct drm_gem_object *obj; 123 struct msm_gem_object *msm_obj; 124 125 /* normally use drm_gem_object_lookup(), but for bulk lookup 126 * all under single table_lock just hit object_idr directly: 127 */ 128 obj = idr_find(&file->object_idr, submit->bos[i].handle); 129 if (!obj) { 130 DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i); 131 ret = -EINVAL; 132 goto out_unlock; 133 } 134 135 msm_obj = to_msm_bo(obj); 136 137 if (!list_empty(&msm_obj->submit_entry)) { 138 DRM_ERROR("handle %u at index %u already on submit list\n", 139 submit->bos[i].handle, i); 140 ret = -EINVAL; 141 goto out_unlock; 142 } 143 144 drm_gem_object_get(obj); 145 146 submit->bos[i].obj = msm_obj; 147 148 list_add_tail(&msm_obj->submit_entry, &submit->bo_list); 149 } 150 151 out_unlock: 152 spin_unlock(&file->table_lock); 153 154 out: 155 submit->nr_bos = i; 156 157 return ret; 158 } 159 160 static int submit_lookup_cmds(struct msm_gem_submit *submit, 161 struct drm_msm_gem_submit *args, struct drm_file *file) 162 { 163 unsigned i, sz; 164 int ret = 0; 165 166 for (i = 0; i < args->nr_cmds; i++) { 167 struct drm_msm_gem_submit_cmd submit_cmd; 168 void __user *userptr = 169 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd))); 170 171 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd)); 172 if (ret) { 173 ret = -EFAULT; 174 goto out; 175 } 176 177 /* validate input from userspace: */ 178 switch (submit_cmd.type) { 179 case MSM_SUBMIT_CMD_BUF: 180 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 181 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 182 break; 183 default: 184 DRM_ERROR("invalid type: %08x\n", submit_cmd.type); 185 return -EINVAL; 186 } 187 188 if (submit_cmd.size % 4) { 189 DRM_ERROR("non-aligned cmdstream buffer size: %u\n", 190 submit_cmd.size); 191 ret = -EINVAL; 192 goto out; 193 } 194 195 submit->cmd[i].type = submit_cmd.type; 196 submit->cmd[i].size = submit_cmd.size / 4; 197 submit->cmd[i].offset = submit_cmd.submit_offset / 4; 198 submit->cmd[i].idx = submit_cmd.submit_idx; 199 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs; 200 201 userptr = u64_to_user_ptr(submit_cmd.relocs); 202 203 sz = array_size(submit_cmd.nr_relocs, 204 sizeof(struct drm_msm_gem_submit_reloc)); 205 /* check for overflow: */ 206 if (sz == SIZE_MAX) { 207 ret = -ENOMEM; 208 goto out; 209 } 210 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL); 211 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz); 212 if (ret) { 213 ret = -EFAULT; 214 goto out; 215 } 216 } 217 218 out: 219 return ret; 220 } 221 222 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, 223 int i, bool backoff) 224 { 225 struct msm_gem_object *msm_obj = submit->bos[i].obj; 226 227 if (submit->bos[i].flags & BO_PINNED) 228 msm_gem_unpin_iova_locked(&msm_obj->base, submit->aspace); 229 230 if (submit->bos[i].flags & BO_LOCKED) 231 dma_resv_unlock(msm_obj->base.resv); 232 233 if (backoff && !(submit->bos[i].flags & BO_VALID)) 234 submit->bos[i].iova = 0; 235 236 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED); 237 } 238 239 /* This is where we make sure all the bo's are reserved and pin'd: */ 240 static int submit_lock_objects(struct msm_gem_submit *submit) 241 { 242 int contended, slow_locked = -1, i, ret = 0; 243 244 retry: 245 for (i = 0; i < submit->nr_bos; i++) { 246 struct msm_gem_object *msm_obj = submit->bos[i].obj; 247 248 if (slow_locked == i) 249 slow_locked = -1; 250 251 contended = i; 252 253 if (!(submit->bos[i].flags & BO_LOCKED)) { 254 ret = dma_resv_lock_interruptible(msm_obj->base.resv, 255 &submit->ticket); 256 if (ret) 257 goto fail; 258 submit->bos[i].flags |= BO_LOCKED; 259 } 260 } 261 262 ww_acquire_done(&submit->ticket); 263 264 return 0; 265 266 fail: 267 for (; i >= 0; i--) 268 submit_unlock_unpin_bo(submit, i, true); 269 270 if (slow_locked > 0) 271 submit_unlock_unpin_bo(submit, slow_locked, true); 272 273 if (ret == -EDEADLK) { 274 struct msm_gem_object *msm_obj = submit->bos[contended].obj; 275 /* we lost out in a seqno race, lock and retry.. */ 276 ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv, 277 &submit->ticket); 278 if (!ret) { 279 submit->bos[contended].flags |= BO_LOCKED; 280 slow_locked = contended; 281 goto retry; 282 } 283 } 284 285 return ret; 286 } 287 288 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit) 289 { 290 int i, ret = 0; 291 292 for (i = 0; i < submit->nr_bos; i++) { 293 struct msm_gem_object *msm_obj = submit->bos[i].obj; 294 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE; 295 296 if (!write) { 297 /* NOTE: _reserve_shared() must happen before 298 * _add_shared_fence(), which makes this a slightly 299 * strange place to call it. OTOH this is a 300 * convenient can-fail point to hook it in. 301 */ 302 ret = dma_resv_reserve_shared(msm_obj->base.resv, 303 1); 304 if (ret) 305 return ret; 306 } 307 308 if (no_implicit) 309 continue; 310 311 ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx, 312 write); 313 if (ret) 314 break; 315 } 316 317 return ret; 318 } 319 320 static int submit_pin_objects(struct msm_gem_submit *submit) 321 { 322 int i, ret = 0; 323 324 submit->valid = true; 325 326 for (i = 0; i < submit->nr_bos; i++) { 327 struct msm_gem_object *msm_obj = submit->bos[i].obj; 328 uint64_t iova; 329 330 /* if locking succeeded, pin bo: */ 331 ret = msm_gem_get_and_pin_iova_locked(&msm_obj->base, 332 submit->aspace, &iova); 333 334 if (ret) 335 break; 336 337 submit->bos[i].flags |= BO_PINNED; 338 339 if (iova == submit->bos[i].iova) { 340 submit->bos[i].flags |= BO_VALID; 341 } else { 342 submit->bos[i].iova = iova; 343 /* iova changed, so address in cmdstream is not valid: */ 344 submit->bos[i].flags &= ~BO_VALID; 345 submit->valid = false; 346 } 347 } 348 349 return ret; 350 } 351 352 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx, 353 struct msm_gem_object **obj, uint64_t *iova, bool *valid) 354 { 355 if (idx >= submit->nr_bos) { 356 DRM_ERROR("invalid buffer index: %u (out of %u)\n", 357 idx, submit->nr_bos); 358 return -EINVAL; 359 } 360 361 if (obj) 362 *obj = submit->bos[idx].obj; 363 if (iova) 364 *iova = submit->bos[idx].iova; 365 if (valid) 366 *valid = !!(submit->bos[idx].flags & BO_VALID); 367 368 return 0; 369 } 370 371 /* process the reloc's and patch up the cmdstream as needed: */ 372 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj, 373 uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs) 374 { 375 uint32_t i, last_offset = 0; 376 uint32_t *ptr; 377 int ret = 0; 378 379 if (!nr_relocs) 380 return 0; 381 382 if (offset % 4) { 383 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset); 384 return -EINVAL; 385 } 386 387 /* For now, just map the entire thing. Eventually we probably 388 * to do it page-by-page, w/ kmap() if not vmap()d.. 389 */ 390 ptr = msm_gem_get_vaddr_locked(&obj->base); 391 392 if (IS_ERR(ptr)) { 393 ret = PTR_ERR(ptr); 394 DBG("failed to map: %d", ret); 395 return ret; 396 } 397 398 for (i = 0; i < nr_relocs; i++) { 399 struct drm_msm_gem_submit_reloc submit_reloc = relocs[i]; 400 uint32_t off; 401 uint64_t iova; 402 bool valid; 403 404 if (submit_reloc.submit_offset % 4) { 405 DRM_ERROR("non-aligned reloc offset: %u\n", 406 submit_reloc.submit_offset); 407 ret = -EINVAL; 408 goto out; 409 } 410 411 /* offset in dwords: */ 412 off = submit_reloc.submit_offset / 4; 413 414 if ((off >= (obj->base.size / 4)) || 415 (off < last_offset)) { 416 DRM_ERROR("invalid offset %u at reloc %u\n", off, i); 417 ret = -EINVAL; 418 goto out; 419 } 420 421 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid); 422 if (ret) 423 goto out; 424 425 if (valid) 426 continue; 427 428 iova += submit_reloc.reloc_offset; 429 430 if (submit_reloc.shift < 0) 431 iova >>= -submit_reloc.shift; 432 else 433 iova <<= submit_reloc.shift; 434 435 ptr[off] = iova | submit_reloc.or; 436 437 last_offset = off; 438 } 439 440 out: 441 msm_gem_put_vaddr_locked(&obj->base); 442 443 return ret; 444 } 445 446 static void submit_cleanup(struct msm_gem_submit *submit) 447 { 448 unsigned i; 449 450 for (i = 0; i < submit->nr_bos; i++) { 451 struct msm_gem_object *msm_obj = submit->bos[i].obj; 452 submit_unlock_unpin_bo(submit, i, false); 453 list_del_init(&msm_obj->submit_entry); 454 drm_gem_object_put_locked(&msm_obj->base); 455 } 456 } 457 458 459 struct msm_submit_post_dep { 460 struct drm_syncobj *syncobj; 461 uint64_t point; 462 struct dma_fence_chain *chain; 463 }; 464 465 static struct drm_syncobj **msm_wait_deps(struct drm_device *dev, 466 struct drm_file *file, 467 uint64_t in_syncobjs_addr, 468 uint32_t nr_in_syncobjs, 469 size_t syncobj_stride, 470 struct msm_ringbuffer *ring) 471 { 472 struct drm_syncobj **syncobjs = NULL; 473 struct drm_msm_gem_submit_syncobj syncobj_desc = {0}; 474 int ret = 0; 475 uint32_t i, j; 476 477 syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs), 478 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); 479 if (!syncobjs) 480 return ERR_PTR(-ENOMEM); 481 482 for (i = 0; i < nr_in_syncobjs; ++i) { 483 uint64_t address = in_syncobjs_addr + i * syncobj_stride; 484 struct dma_fence *fence; 485 486 if (copy_from_user(&syncobj_desc, 487 u64_to_user_ptr(address), 488 min(syncobj_stride, sizeof(syncobj_desc)))) { 489 ret = -EFAULT; 490 break; 491 } 492 493 if (syncobj_desc.point && 494 !drm_core_check_feature(dev, DRIVER_SYNCOBJ_TIMELINE)) { 495 ret = -EOPNOTSUPP; 496 break; 497 } 498 499 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) { 500 ret = -EINVAL; 501 break; 502 } 503 504 ret = drm_syncobj_find_fence(file, syncobj_desc.handle, 505 syncobj_desc.point, 0, &fence); 506 if (ret) 507 break; 508 509 if (!dma_fence_match_context(fence, ring->fctx->context)) 510 ret = dma_fence_wait(fence, true); 511 512 dma_fence_put(fence); 513 if (ret) 514 break; 515 516 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) { 517 syncobjs[i] = 518 drm_syncobj_find(file, syncobj_desc.handle); 519 if (!syncobjs[i]) { 520 ret = -EINVAL; 521 break; 522 } 523 } 524 } 525 526 if (ret) { 527 for (j = 0; j <= i; ++j) { 528 if (syncobjs[j]) 529 drm_syncobj_put(syncobjs[j]); 530 } 531 kfree(syncobjs); 532 return ERR_PTR(ret); 533 } 534 return syncobjs; 535 } 536 537 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs, 538 uint32_t nr_syncobjs) 539 { 540 uint32_t i; 541 542 for (i = 0; syncobjs && i < nr_syncobjs; ++i) { 543 if (syncobjs[i]) 544 drm_syncobj_replace_fence(syncobjs[i], NULL); 545 } 546 } 547 548 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev, 549 struct drm_file *file, 550 uint64_t syncobjs_addr, 551 uint32_t nr_syncobjs, 552 size_t syncobj_stride) 553 { 554 struct msm_submit_post_dep *post_deps; 555 struct drm_msm_gem_submit_syncobj syncobj_desc = {0}; 556 int ret = 0; 557 uint32_t i, j; 558 559 post_deps = kmalloc_array(nr_syncobjs, sizeof(*post_deps), 560 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); 561 if (!post_deps) 562 return ERR_PTR(-ENOMEM); 563 564 for (i = 0; i < nr_syncobjs; ++i) { 565 uint64_t address = syncobjs_addr + i * syncobj_stride; 566 567 if (copy_from_user(&syncobj_desc, 568 u64_to_user_ptr(address), 569 min(syncobj_stride, sizeof(syncobj_desc)))) { 570 ret = -EFAULT; 571 break; 572 } 573 574 post_deps[i].point = syncobj_desc.point; 575 post_deps[i].chain = NULL; 576 577 if (syncobj_desc.flags) { 578 ret = -EINVAL; 579 break; 580 } 581 582 if (syncobj_desc.point) { 583 if (!drm_core_check_feature(dev, 584 DRIVER_SYNCOBJ_TIMELINE)) { 585 ret = -EOPNOTSUPP; 586 break; 587 } 588 589 post_deps[i].chain = 590 kmalloc(sizeof(*post_deps[i].chain), 591 GFP_KERNEL); 592 if (!post_deps[i].chain) { 593 ret = -ENOMEM; 594 break; 595 } 596 } 597 598 post_deps[i].syncobj = 599 drm_syncobj_find(file, syncobj_desc.handle); 600 if (!post_deps[i].syncobj) { 601 ret = -EINVAL; 602 break; 603 } 604 } 605 606 if (ret) { 607 for (j = 0; j <= i; ++j) { 608 kfree(post_deps[j].chain); 609 if (post_deps[j].syncobj) 610 drm_syncobj_put(post_deps[j].syncobj); 611 } 612 613 kfree(post_deps); 614 return ERR_PTR(ret); 615 } 616 617 return post_deps; 618 } 619 620 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps, 621 uint32_t count, struct dma_fence *fence) 622 { 623 uint32_t i; 624 625 for (i = 0; post_deps && i < count; ++i) { 626 if (post_deps[i].chain) { 627 drm_syncobj_add_point(post_deps[i].syncobj, 628 post_deps[i].chain, 629 fence, post_deps[i].point); 630 post_deps[i].chain = NULL; 631 } else { 632 drm_syncobj_replace_fence(post_deps[i].syncobj, 633 fence); 634 } 635 } 636 } 637 638 int msm_ioctl_gem_submit(struct drm_device *dev, void *data, 639 struct drm_file *file) 640 { 641 static atomic_t ident = ATOMIC_INIT(0); 642 struct msm_drm_private *priv = dev->dev_private; 643 struct drm_msm_gem_submit *args = data; 644 struct msm_file_private *ctx = file->driver_priv; 645 struct msm_gem_submit *submit; 646 struct msm_gpu *gpu = priv->gpu; 647 struct sync_file *sync_file = NULL; 648 struct msm_gpu_submitqueue *queue; 649 struct msm_ringbuffer *ring; 650 struct msm_submit_post_dep *post_deps = NULL; 651 struct drm_syncobj **syncobjs_to_reset = NULL; 652 int out_fence_fd = -1; 653 struct pid *pid = get_pid(task_pid(current)); 654 bool has_ww_ticket = false; 655 unsigned i; 656 int ret, submitid; 657 if (!gpu) 658 return -ENXIO; 659 660 if (args->pad) 661 return -EINVAL; 662 663 /* for now, we just have 3d pipe.. eventually this would need to 664 * be more clever to dispatch to appropriate gpu module: 665 */ 666 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0) 667 return -EINVAL; 668 669 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS) 670 return -EINVAL; 671 672 if (args->flags & MSM_SUBMIT_SUDO) { 673 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) || 674 !capable(CAP_SYS_RAWIO)) 675 return -EINVAL; 676 } 677 678 queue = msm_submitqueue_get(ctx, args->queueid); 679 if (!queue) 680 return -ENOENT; 681 682 /* Get a unique identifier for the submission for logging purposes */ 683 submitid = atomic_inc_return(&ident) - 1; 684 685 ring = gpu->rb[queue->prio]; 686 trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid, 687 args->nr_bos, args->nr_cmds); 688 689 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) { 690 struct dma_fence *in_fence; 691 692 in_fence = sync_file_get_fence(args->fence_fd); 693 694 if (!in_fence) 695 return -EINVAL; 696 697 /* 698 * Wait if the fence is from a foreign context, or if the fence 699 * array contains any fence from a foreign context. 700 */ 701 ret = 0; 702 if (!dma_fence_match_context(in_fence, ring->fctx->context)) 703 ret = dma_fence_wait(in_fence, true); 704 705 dma_fence_put(in_fence); 706 if (ret) 707 return ret; 708 } 709 710 if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) { 711 syncobjs_to_reset = msm_wait_deps(dev, file, 712 args->in_syncobjs, 713 args->nr_in_syncobjs, 714 args->syncobj_stride, ring); 715 if (IS_ERR(syncobjs_to_reset)) 716 return PTR_ERR(syncobjs_to_reset); 717 } 718 719 if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) { 720 post_deps = msm_parse_post_deps(dev, file, 721 args->out_syncobjs, 722 args->nr_out_syncobjs, 723 args->syncobj_stride); 724 if (IS_ERR(post_deps)) { 725 ret = PTR_ERR(post_deps); 726 goto out_post_unlock; 727 } 728 } 729 730 ret = mutex_lock_interruptible(&dev->struct_mutex); 731 if (ret) 732 goto out_post_unlock; 733 734 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 735 out_fence_fd = get_unused_fd_flags(O_CLOEXEC); 736 if (out_fence_fd < 0) { 737 ret = out_fence_fd; 738 goto out_unlock; 739 } 740 } 741 742 submit = submit_create(dev, gpu, queue, args->nr_bos, 743 args->nr_cmds); 744 if (!submit) { 745 ret = -ENOMEM; 746 goto out_unlock; 747 } 748 749 submit->pid = pid; 750 submit->ident = submitid; 751 752 if (args->flags & MSM_SUBMIT_SUDO) 753 submit->in_rb = true; 754 755 ret = submit_lookup_objects(submit, args, file); 756 if (ret) 757 goto out_pre_pm; 758 759 ret = submit_lookup_cmds(submit, args, file); 760 if (ret) 761 goto out_pre_pm; 762 763 /* 764 * Thanks to dev_pm_opp opp_table_lock interactions with mm->mmap_sem 765 * in the resume path, we need to to rpm get before we lock objs. 766 * Which unfortunately might involve powering up the GPU sooner than 767 * is necessary. But at least in the explicit fencing case, we will 768 * have already done all the fence waiting. 769 */ 770 pm_runtime_get_sync(&gpu->pdev->dev); 771 772 /* copy_*_user while holding a ww ticket upsets lockdep */ 773 ww_acquire_init(&submit->ticket, &reservation_ww_class); 774 has_ww_ticket = true; 775 ret = submit_lock_objects(submit); 776 if (ret) 777 goto out; 778 779 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT)); 780 if (ret) 781 goto out; 782 783 ret = submit_pin_objects(submit); 784 if (ret) 785 goto out; 786 787 for (i = 0; i < args->nr_cmds; i++) { 788 struct msm_gem_object *msm_obj; 789 uint64_t iova; 790 791 ret = submit_bo(submit, submit->cmd[i].idx, 792 &msm_obj, &iova, NULL); 793 if (ret) 794 goto out; 795 796 if (!submit->cmd[i].size || 797 ((submit->cmd[i].size + submit->cmd[i].offset) > 798 msm_obj->base.size / 4)) { 799 DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4); 800 ret = -EINVAL; 801 goto out; 802 } 803 804 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4); 805 806 if (submit->valid) 807 continue; 808 809 ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4, 810 submit->cmd[i].nr_relocs, submit->cmd[i].relocs); 811 if (ret) 812 goto out; 813 } 814 815 submit->nr_cmds = i; 816 817 submit->fence = msm_fence_alloc(ring->fctx); 818 if (IS_ERR(submit->fence)) { 819 ret = PTR_ERR(submit->fence); 820 submit->fence = NULL; 821 goto out; 822 } 823 824 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 825 sync_file = sync_file_create(submit->fence); 826 if (!sync_file) { 827 ret = -ENOMEM; 828 goto out; 829 } 830 } 831 832 msm_gpu_submit(gpu, submit); 833 834 args->fence = submit->fence->seqno; 835 836 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 837 fd_install(out_fence_fd, sync_file->file); 838 args->fence_fd = out_fence_fd; 839 } 840 841 msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs); 842 msm_process_post_deps(post_deps, args->nr_out_syncobjs, 843 submit->fence); 844 845 846 out: 847 pm_runtime_put(&gpu->pdev->dev); 848 out_pre_pm: 849 submit_cleanup(submit); 850 if (has_ww_ticket) 851 ww_acquire_fini(&submit->ticket); 852 msm_gem_submit_put(submit); 853 out_unlock: 854 if (ret && (out_fence_fd >= 0)) 855 put_unused_fd(out_fence_fd); 856 mutex_unlock(&dev->struct_mutex); 857 858 out_post_unlock: 859 if (!IS_ERR_OR_NULL(post_deps)) { 860 for (i = 0; i < args->nr_out_syncobjs; ++i) { 861 kfree(post_deps[i].chain); 862 drm_syncobj_put(post_deps[i].syncobj); 863 } 864 kfree(post_deps); 865 } 866 867 if (!IS_ERR_OR_NULL(syncobjs_to_reset)) { 868 for (i = 0; i < args->nr_in_syncobjs; ++i) { 869 if (syncobjs_to_reset[i]) 870 drm_syncobj_put(syncobjs_to_reset[i]); 871 } 872 kfree(syncobjs_to_reset); 873 } 874 875 return ret; 876 } 877