drivers/gpu/drm/msm/msm_gem_submit.c

   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   /* obj lock is held */
  27 #define BO_ACTIVE   0x2000   /* active refcnt is held */
  28 #define BO_PINNED   0x1000   /* obj is pinned and on active list */
  29
  30 static struct msm_gem_submit *submit_create(struct drm_device *dev,
  31                 struct msm_gpu *gpu,
  32                 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
  33                 uint32_t nr_cmds)
  34 {
  35         struct msm_gem_submit *submit;
  36         uint64_t sz;
  37         int ret;
  38
  39         sz = struct_size(submit, bos, nr_bos) +
  40                         ((u64)nr_cmds * sizeof(submit->cmd[0]));
  41
  42         if (sz > SIZE_MAX)
  43                 return ERR_PTR(-ENOMEM);
  44
  45         submit = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
  46         if (!submit)
  47                 return ERR_PTR(-ENOMEM);
  48
  49         ret = drm_sched_job_init(&submit->base, queue->entity, queue);
  50         if (ret) {
  51                 kfree(submit);
  52                 return ERR_PTR(ret);
  53         }
  54
  55         kref_init(&submit->ref);
  56         submit->dev = dev;
  57         submit->aspace = queue->ctx->aspace;
  58         submit->gpu = gpu;
  59         submit->cmd = (void *)&submit->bos[nr_bos];
  60         submit->queue = queue;
  61         submit->ring = gpu->rb[queue->ring_nr];
  62         submit->fault_dumped = false;
  63
  64         INIT_LIST_HEAD(&submit->node);
  65
  66         return submit;
  67 }
  68
  69 void __msm_gem_submit_destroy(struct kref *kref)
  70 {
  71         struct msm_gem_submit *submit =
  72                         container_of(kref, struct msm_gem_submit, ref);
  73         unsigned i;
  74
  75         if (submit->fence_id) {
  76                 mutex_lock(&submit->queue->lock);
  77                 idr_remove(&submit->queue->fence_idr, submit->fence_id);
  78                 mutex_unlock(&submit->queue->lock);
  79         }
  80
  81         dma_fence_put(submit->user_fence);
  82         dma_fence_put(submit->hw_fence);
  83
  84         put_pid(submit->pid);
  85         msm_submitqueue_put(submit->queue);
  86
  87         for (i = 0; i < submit->nr_cmds; i++)
  88                 kfree(submit->cmd[i].relocs);
  89
  90         kfree(submit);
  91 }
  92
  93 static int submit_lookup_objects(struct msm_gem_submit *submit,
  94                 struct drm_msm_gem_submit *args, struct drm_file *file)
  95 {
  96         unsigned i;
  97         int ret = 0;
  98
  99         for (i = 0; i < args->nr_bos; i++) {
 100                 struct drm_msm_gem_submit_bo submit_bo;
 101                 void __user *userptr =
 102                         u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
 103
 104                 /* make sure we don't have garbage flags, in case we hit
 105                  * error path before flags is initialized:
 106                  */
 107                 submit->bos[i].flags = 0;
 108
 109                 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
 110                         ret = -EFAULT;
 111                         i = 0;
 112                         goto out;
 113                 }
 114
 115 /* at least one of READ and/or WRITE flags should be set: */
 116 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
 117
 118                 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
 119                         !(submit_bo.flags & MANDATORY_FLAGS)) {
 120                         DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
 121                         ret = -EINVAL;
 122                         i = 0;
 123                         goto out;
 124                 }
 125
 126                 submit->bos[i].handle = submit_bo.handle;
 127                 submit->bos[i].flags = submit_bo.flags;
 128                 /* in validate_objects() we figure out if this is true: */
 129                 submit->bos[i].iova  = submit_bo.presumed;
 130         }
 131
 132         spin_lock(&file->table_lock);
 133
 134         for (i = 0; i < args->nr_bos; i++) {
 135                 struct drm_gem_object *obj;
 136
 137                 /* normally use drm_gem_object_lookup(), but for bulk lookup
 138                  * all under single table_lock just hit object_idr directly:
 139                  */
 140                 obj = idr_find(&file->object_idr, submit->bos[i].handle);
 141                 if (!obj) {
 142                         DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
 143                         ret = -EINVAL;
 144                         goto out_unlock;
 145                 }
 146
 147                 drm_gem_object_get(obj);
 148
 149                 submit->bos[i].obj = to_msm_bo(obj);
 150         }
 151
 152 out_unlock:
 153         spin_unlock(&file->table_lock);
 154
 155 out:
 156         submit->nr_bos = i;
 157
 158         return ret;
 159 }
 160
 161 static int submit_lookup_cmds(struct msm_gem_submit *submit,
 162                 struct drm_msm_gem_submit *args, struct drm_file *file)
 163 {
 164         unsigned i;
 165         size_t sz;
 166         int ret = 0;
 167
 168         for (i = 0; i < args->nr_cmds; i++) {
 169                 struct drm_msm_gem_submit_cmd submit_cmd;
 170                 void __user *userptr =
 171                         u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
 172
 173                 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
 174                 if (ret) {
 175                         ret = -EFAULT;
 176                         goto out;
 177                 }
 178
 179                 /* validate input from userspace: */
 180                 switch (submit_cmd.type) {
 181                 case MSM_SUBMIT_CMD_BUF:
 182                 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
 183                 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
 184                         break;
 185                 default:
 186                         DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
 187                         return -EINVAL;
 188                 }
 189
 190                 if (submit_cmd.size % 4) {
 191                         DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
 192                                         submit_cmd.size);
 193                         ret = -EINVAL;
 194                         goto out;
 195                 }
 196
 197                 submit->cmd[i].type = submit_cmd.type;
 198                 submit->cmd[i].size = submit_cmd.size / 4;
 199                 submit->cmd[i].offset = submit_cmd.submit_offset / 4;
 200                 submit->cmd[i].idx  = submit_cmd.submit_idx;
 201                 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
 202
 203                 userptr = u64_to_user_ptr(submit_cmd.relocs);
 204
 205                 sz = array_size(submit_cmd.nr_relocs,
 206                                 sizeof(struct drm_msm_gem_submit_reloc));
 207                 /* check for overflow: */
 208                 if (sz == SIZE_MAX) {
 209                         ret = -ENOMEM;
 210                         goto out;
 211                 }
 212                 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL);
 213                 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
 214                 if (ret) {
 215                         ret = -EFAULT;
 216                         goto out;
 217                 }
 218         }
 219
 220 out:
 221         return ret;
 222 }
 223
 224 /* Unwind bo state, according to cleanup_flags.  In the success case, only
 225  * the lock is dropped at the end of the submit (and active/pin ref is dropped
 226  * later when the submit is retired).
 227  */
 228 static void submit_cleanup_bo(struct msm_gem_submit *submit, int i,
 229                 unsigned cleanup_flags)
 230 {
 231         struct drm_gem_object *obj = &submit->bos[i].obj->base;
 232         unsigned flags = submit->bos[i].flags & cleanup_flags;
 233
 234         if (flags & BO_PINNED)
 235                 msm_gem_unpin_iova_locked(obj, submit->aspace);
 236
 237         if (flags & BO_ACTIVE)
 238                 msm_gem_active_put(obj);
 239
 240         if (flags & BO_LOCKED)
 241                 dma_resv_unlock(obj->resv);
 242
 243         submit->bos[i].flags &= ~cleanup_flags;
 244 }
 245
 246 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
 247 {
 248         submit_cleanup_bo(submit, i, BO_PINNED | BO_ACTIVE | BO_LOCKED);
 249
 250         if (!(submit->bos[i].flags & BO_VALID))
 251                 submit->bos[i].iova = 0;
 252 }
 253
 254 /* This is where we make sure all the bo's are reserved and pin'd: */
 255 static int submit_lock_objects(struct msm_gem_submit *submit)
 256 {
 257         int contended, slow_locked = -1, i, ret = 0;
 258
 259 retry:
 260         for (i = 0; i < submit->nr_bos; i++) {
 261                 struct msm_gem_object *msm_obj = submit->bos[i].obj;
 262
 263                 if (slow_locked == i)
 264                         slow_locked = -1;
 265
 266                 contended = i;
 267
 268                 if (!(submit->bos[i].flags & BO_LOCKED)) {
 269                         ret = dma_resv_lock_interruptible(msm_obj->base.resv,
 270                                                           &submit->ticket);
 271                         if (ret)
 272                                 goto fail;
 273                         submit->bos[i].flags |= BO_LOCKED;
 274                 }
 275         }
 276
 277         ww_acquire_done(&submit->ticket);
 278
 279         return 0;
 280
 281 fail:
 282         if (ret == -EALREADY) {
 283                 DRM_ERROR("handle %u at index %u already on submit list\n",
 284                                 submit->bos[i].handle, i);
 285                 ret = -EINVAL;
 286         }
 287
 288         for (; i >= 0; i--)
 289                 submit_unlock_unpin_bo(submit, i);
 290
 291         if (slow_locked > 0)
 292                 submit_unlock_unpin_bo(submit, slow_locked);
 293
 294         if (ret == -EDEADLK) {
 295                 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
 296                 /* we lost out in a seqno race, lock and retry.. */
 297                 ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv,
 298                                                        &submit->ticket);
 299                 if (!ret) {
 300                         submit->bos[contended].flags |= BO_LOCKED;
 301                         slow_locked = contended;
 302                         goto retry;
 303                 }
 304
 305                 /* Not expecting -EALREADY here, if the bo was already
 306                  * locked, we should have gotten -EALREADY already from
 307                  * the dma_resv_lock_interruptable() call.
 308                  */
 309                 WARN_ON_ONCE(ret == -EALREADY);
 310         }
 311
 312         return ret;
 313 }
 314
 315 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
 316 {
 317         int i, ret = 0;
 318
 319         for (i = 0; i < submit->nr_bos; i++) {
 320                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
 321                 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
 322
 323                 if (!write) {
 324                         /* NOTE: _reserve_shared() must happen before
 325                          * _add_shared_fence(), which makes this a slightly
 326                          * strange place to call it.  OTOH this is a
 327                          * convenient can-fail point to hook it in.
 328                          */
 329                         ret = dma_resv_reserve_shared(obj->resv, 1);
 330                         if (ret)
 331                                 return ret;
 332                 }
 333
 334                 /* exclusive fences must be ordered */
 335                 if (no_implicit && !write)
 336                         continue;
 337
 338                 ret = drm_sched_job_add_implicit_dependencies(&submit->base,
 339                                                               obj,
 340                                                               write);
 341                 if (ret)
 342                         break;
 343         }
 344
 345         return ret;
 346 }
 347
 348 static int submit_pin_objects(struct msm_gem_submit *submit)
 349 {
 350         int i, ret = 0;
 351
 352         submit->valid = true;
 353
 354         /*
 355          * Increment active_count first, so if under memory pressure, we
 356          * don't inadvertently evict a bo needed by the submit in order
 357          * to pin an earlier bo in the same submit.
 358          */
 359         for (i = 0; i < submit->nr_bos; i++) {
 360                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
 361
 362                 msm_gem_active_get(obj, submit->gpu);
 363                 submit->bos[i].flags |= BO_ACTIVE;
 364         }
 365
 366         for (i = 0; i < submit->nr_bos; i++) {
 367                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
 368                 uint64_t iova;
 369
 370                 /* if locking succeeded, pin bo: */
 371                 ret = msm_gem_get_and_pin_iova_locked(obj,
 372                                 submit->aspace, &iova);
 373
 374                 if (ret)
 375                         break;
 376
 377                 submit->bos[i].flags |= BO_PINNED;
 378
 379                 if (iova == submit->bos[i].iova) {
 380                         submit->bos[i].flags |= BO_VALID;
 381                 } else {
 382                         submit->bos[i].iova = iova;
 383                         /* iova changed, so address in cmdstream is not valid: */
 384                         submit->bos[i].flags &= ~BO_VALID;
 385                         submit->valid = false;
 386                 }
 387         }
 388
 389         return ret;
 390 }
 391
 392 static void submit_attach_object_fences(struct msm_gem_submit *submit)
 393 {
 394         int i;
 395
 396         for (i = 0; i < submit->nr_bos; i++) {
 397                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
 398
 399                 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
 400                         dma_resv_add_excl_fence(obj->resv, submit->user_fence);
 401                 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
 402                         dma_resv_add_shared_fence(obj->resv, submit->user_fence);
 403         }
 404 }
 405
 406 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
 407                 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
 408 {
 409         if (idx >= submit->nr_bos) {
 410                 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
 411                                 idx, submit->nr_bos);
 412                 return -EINVAL;
 413         }
 414
 415         if (obj)
 416                 *obj = submit->bos[idx].obj;
 417         if (iova)
 418                 *iova = submit->bos[idx].iova;
 419         if (valid)
 420                 *valid = !!(submit->bos[idx].flags & BO_VALID);
 421
 422         return 0;
 423 }
 424
 425 /* process the reloc's and patch up the cmdstream as needed: */
 426 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
 427                 uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
 428 {
 429         uint32_t i, last_offset = 0;
 430         uint32_t *ptr;
 431         int ret = 0;
 432
 433         if (!nr_relocs)
 434                 return 0;
 435
 436         if (offset % 4) {
 437                 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
 438                 return -EINVAL;
 439         }
 440
 441         /* For now, just map the entire thing.  Eventually we probably
 442          * to do it page-by-page, w/ kmap() if not vmap()d..
 443          */
 444         ptr = msm_gem_get_vaddr_locked(&obj->base);
 445
 446         if (IS_ERR(ptr)) {
 447                 ret = PTR_ERR(ptr);
 448                 DBG("failed to map: %d", ret);
 449                 return ret;
 450         }
 451
 452         for (i = 0; i < nr_relocs; i++) {
 453                 struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
 454                 uint32_t off;
 455                 uint64_t iova;
 456                 bool valid;
 457
 458                 if (submit_reloc.submit_offset % 4) {
 459                         DRM_ERROR("non-aligned reloc offset: %u\n",
 460                                         submit_reloc.submit_offset);
 461                         ret = -EINVAL;
 462                         goto out;
 463                 }
 464
 465                 /* offset in dwords: */
 466                 off = submit_reloc.submit_offset / 4;
 467
 468                 if ((off >= (obj->base.size / 4)) ||
 469                                 (off < last_offset)) {
 470                         DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
 471                         ret = -EINVAL;
 472                         goto out;
 473                 }
 474
 475                 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
 476                 if (ret)
 477                         goto out;
 478
 479                 if (valid)
 480                         continue;
 481
 482                 iova += submit_reloc.reloc_offset;
 483
 484                 if (submit_reloc.shift < 0)
 485                         iova >>= -submit_reloc.shift;
 486                 else
 487                         iova <<= submit_reloc.shift;
 488
 489                 ptr[off] = iova | submit_reloc.or;
 490
 491                 last_offset = off;
 492         }
 493
 494 out:
 495         msm_gem_put_vaddr_locked(&obj->base);
 496
 497         return ret;
 498 }
 499
 500 /* Cleanup submit at end of ioctl.  In the error case, this also drops
 501  * references, unpins, and drops active refcnt.  In the non-error case,
 502  * this is done when the submit is retired.
 503  */
 504 static void submit_cleanup(struct msm_gem_submit *submit, bool error)
 505 {
 506         unsigned cleanup_flags = BO_LOCKED;
 507         unsigned i;
 508
 509         if (error)
 510                 cleanup_flags |= BO_PINNED | BO_ACTIVE;
 511
 512         for (i = 0; i < submit->nr_bos; i++) {
 513                 struct msm_gem_object *msm_obj = submit->bos[i].obj;
 514                 submit_cleanup_bo(submit, i, cleanup_flags);
 515                 if (error)
 516                         drm_gem_object_put(&msm_obj->base);
 517         }
 518 }
 519
 520 void msm_submit_retire(struct msm_gem_submit *submit)
 521 {
 522         int i;
 523
 524         for (i = 0; i < submit->nr_bos; i++) {
 525                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
 526
 527                 msm_gem_lock(obj);
 528                 submit_cleanup_bo(submit, i, BO_PINNED | BO_ACTIVE);
 529                 msm_gem_unlock(obj);
 530                 drm_gem_object_put(obj);
 531         }
 532 }
 533
 534 struct msm_submit_post_dep {
 535         struct drm_syncobj *syncobj;
 536         uint64_t point;
 537         struct dma_fence_chain *chain;
 538 };
 539
 540 static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
 541                                            struct drm_file *file,
 542                                            uint64_t in_syncobjs_addr,
 543                                            uint32_t nr_in_syncobjs,
 544                                            size_t syncobj_stride,
 545                                            struct msm_ringbuffer *ring)
 546 {
 547         struct drm_syncobj **syncobjs = NULL;
 548         struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
 549         int ret = 0;
 550         uint32_t i, j;
 551
 552         syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
 553                            GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
 554         if (!syncobjs)
 555                 return ERR_PTR(-ENOMEM);
 556
 557         for (i = 0; i < nr_in_syncobjs; ++i) {
 558                 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
 559                 struct dma_fence *fence;
 560
 561                 if (copy_from_user(&syncobj_desc,
 562                                    u64_to_user_ptr(address),
 563                                    min(syncobj_stride, sizeof(syncobj_desc)))) {
 564                         ret = -EFAULT;
 565                         break;
 566                 }
 567
 568                 if (syncobj_desc.point &&
 569                     !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
 570                         ret = -EOPNOTSUPP;
 571                         break;
 572                 }
 573
 574                 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
 575                         ret = -EINVAL;
 576                         break;
 577                 }
 578
 579                 ret = drm_syncobj_find_fence(file, syncobj_desc.handle,
 580                                              syncobj_desc.point, 0, &fence);
 581                 if (ret)
 582                         break;
 583
 584                 ret = drm_sched_job_add_dependency(&submit->base, fence);
 585                 if (ret)
 586                         break;
 587
 588                 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
 589                         syncobjs[i] =
 590                                 drm_syncobj_find(file, syncobj_desc.handle);
 591                         if (!syncobjs[i]) {
 592                                 ret = -EINVAL;
 593                                 break;
 594                         }
 595                 }
 596         }
 597
 598         if (ret) {
 599                 for (j = 0; j <= i; ++j) {
 600                         if (syncobjs[j])
 601                                 drm_syncobj_put(syncobjs[j]);
 602                 }
 603                 kfree(syncobjs);
 604                 return ERR_PTR(ret);
 605         }
 606         return syncobjs;
 607 }
 608
 609 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
 610                                uint32_t nr_syncobjs)
 611 {
 612         uint32_t i;
 613
 614         for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
 615                 if (syncobjs[i])
 616                         drm_syncobj_replace_fence(syncobjs[i], NULL);
 617         }
 618 }
 619
 620 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
 621                                                        struct drm_file *file,
 622                                                        uint64_t syncobjs_addr,
 623                                                        uint32_t nr_syncobjs,
 624                                                        size_t syncobj_stride)
 625 {
 626         struct msm_submit_post_dep *post_deps;
 627         struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
 628         int ret = 0;
 629         uint32_t i, j;
 630
 631         post_deps = kmalloc_array(nr_syncobjs, sizeof(*post_deps),
 632                                   GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
 633         if (!post_deps)
 634                 return ERR_PTR(-ENOMEM);
 635
 636         for (i = 0; i < nr_syncobjs; ++i) {
 637                 uint64_t address = syncobjs_addr + i * syncobj_stride;
 638
 639                 if (copy_from_user(&syncobj_desc,
 640                                    u64_to_user_ptr(address),
 641                                    min(syncobj_stride, sizeof(syncobj_desc)))) {
 642                         ret = -EFAULT;
 643                         break;
 644                 }
 645
 646                 post_deps[i].point = syncobj_desc.point;
 647                 post_deps[i].chain = NULL;
 648
 649                 if (syncobj_desc.flags) {
 650                         ret = -EINVAL;
 651                         break;
 652                 }
 653
 654                 if (syncobj_desc.point) {
 655                         if (!drm_core_check_feature(dev,
 656                                                     DRIVER_SYNCOBJ_TIMELINE)) {
 657                                 ret = -EOPNOTSUPP;
 658                                 break;
 659                         }
 660
 661                         post_deps[i].chain = dma_fence_chain_alloc();
 662                         if (!post_deps[i].chain) {
 663                                 ret = -ENOMEM;
 664                                 break;
 665                         }
 666                 }
 667
 668                 post_deps[i].syncobj =
 669                         drm_syncobj_find(file, syncobj_desc.handle);
 670                 if (!post_deps[i].syncobj) {
 671                         ret = -EINVAL;
 672                         break;
 673                 }
 674         }
 675
 676         if (ret) {
 677                 for (j = 0; j <= i; ++j) {
 678                         dma_fence_chain_free(post_deps[j].chain);
 679                         if (post_deps[j].syncobj)
 680                                 drm_syncobj_put(post_deps[j].syncobj);
 681                 }
 682
 683                 kfree(post_deps);
 684                 return ERR_PTR(ret);
 685         }
 686
 687         return post_deps;
 688 }
 689
 690 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
 691                                   uint32_t count, struct dma_fence *fence)
 692 {
 693         uint32_t i;
 694
 695         for (i = 0; post_deps && i < count; ++i) {
 696                 if (post_deps[i].chain) {
 697                         drm_syncobj_add_point(post_deps[i].syncobj,
 698                                               post_deps[i].chain,
 699                                               fence, post_deps[i].point);
 700                         post_deps[i].chain = NULL;
 701                 } else {
 702                         drm_syncobj_replace_fence(post_deps[i].syncobj,
 703                                                   fence);
 704                 }
 705         }
 706 }
 707
 708 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 709                 struct drm_file *file)
 710 {
 711         static atomic_t ident = ATOMIC_INIT(0);
 712         struct msm_drm_private *priv = dev->dev_private;
 713         struct drm_msm_gem_submit *args = data;
 714         struct msm_file_private *ctx = file->driver_priv;
 715         struct msm_gem_submit *submit = NULL;
 716         struct msm_gpu *gpu = priv->gpu;
 717         struct msm_gpu_submitqueue *queue;
 718         struct msm_ringbuffer *ring;
 719         struct msm_submit_post_dep *post_deps = NULL;
 720         struct drm_syncobj **syncobjs_to_reset = NULL;
 721         int out_fence_fd = -1;
 722         struct pid *pid = get_pid(task_pid(current));
 723         bool has_ww_ticket = false;
 724         unsigned i;
 725         int ret, submitid;
 726
 727         if (!gpu)
 728                 return -ENXIO;
 729
 730         if (args->pad)
 731                 return -EINVAL;
 732
 733         /* for now, we just have 3d pipe.. eventually this would need to
 734          * be more clever to dispatch to appropriate gpu module:
 735          */
 736         if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
 737                 return -EINVAL;
 738
 739         if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
 740                 return -EINVAL;
 741
 742         if (args->flags & MSM_SUBMIT_SUDO) {
 743                 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
 744                     !capable(CAP_SYS_RAWIO))
 745                         return -EINVAL;
 746         }
 747
 748         queue = msm_submitqueue_get(ctx, args->queueid);
 749         if (!queue)
 750                 return -ENOENT;
 751
 752         /* Get a unique identifier for the submission for logging purposes */
 753         submitid = atomic_inc_return(&ident) - 1;
 754
 755         ring = gpu->rb[queue->ring_nr];
 756         trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid,
 757                 args->nr_bos, args->nr_cmds);
 758
 759         ret = mutex_lock_interruptible(&queue->lock);
 760         if (ret)
 761                 goto out_post_unlock;
 762
 763         if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
 764                 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
 765                 if (out_fence_fd < 0) {
 766                         ret = out_fence_fd;
 767                         goto out_unlock;
 768                 }
 769         }
 770
 771         submit = submit_create(dev, gpu, queue, args->nr_bos,
 772                 args->nr_cmds);
 773         if (IS_ERR(submit)) {
 774                 ret = PTR_ERR(submit);
 775                 submit = NULL;
 776                 goto out_unlock;
 777         }
 778
 779         submit->pid = pid;
 780         submit->ident = submitid;
 781
 782         if (args->flags & MSM_SUBMIT_SUDO)
 783                 submit->in_rb = true;
 784
 785         if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
 786                 struct dma_fence *in_fence;
 787
 788                 in_fence = sync_file_get_fence(args->fence_fd);
 789
 790                 if (!in_fence) {
 791                         ret = -EINVAL;
 792                         goto out_unlock;
 793                 }
 794
 795                 ret = drm_sched_job_add_dependency(&submit->base, in_fence);
 796                 if (ret)
 797                         goto out_unlock;
 798         }
 799
 800         if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
 801                 syncobjs_to_reset = msm_parse_deps(submit, file,
 802                                                    args->in_syncobjs,
 803                                                    args->nr_in_syncobjs,
 804                                                    args->syncobj_stride, ring);
 805                 if (IS_ERR(syncobjs_to_reset)) {
 806                         ret = PTR_ERR(syncobjs_to_reset);
 807                         goto out_unlock;
 808                 }
 809         }
 810
 811         if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
 812                 post_deps = msm_parse_post_deps(dev, file,
 813                                                 args->out_syncobjs,
 814                                                 args->nr_out_syncobjs,
 815                                                 args->syncobj_stride);
 816                 if (IS_ERR(post_deps)) {
 817                         ret = PTR_ERR(post_deps);
 818                         goto out_unlock;
 819                 }
 820         }
 821
 822         ret = submit_lookup_objects(submit, args, file);
 823         if (ret)
 824                 goto out;
 825
 826         ret = submit_lookup_cmds(submit, args, file);
 827         if (ret)
 828                 goto out;
 829
 830         /* copy_*_user while holding a ww ticket upsets lockdep */
 831         ww_acquire_init(&submit->ticket, &reservation_ww_class);
 832         has_ww_ticket = true;
 833         ret = submit_lock_objects(submit);
 834         if (ret)
 835                 goto out;
 836
 837         ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
 838         if (ret)
 839                 goto out;
 840
 841         ret = submit_pin_objects(submit);
 842         if (ret)
 843                 goto out;
 844
 845         for (i = 0; i < args->nr_cmds; i++) {
 846                 struct msm_gem_object *msm_obj;
 847                 uint64_t iova;
 848
 849                 ret = submit_bo(submit, submit->cmd[i].idx,
 850                                 &msm_obj, &iova, NULL);
 851                 if (ret)
 852                         goto out;
 853
 854                 if (!submit->cmd[i].size ||
 855                         ((submit->cmd[i].size + submit->cmd[i].offset) >
 856                                 msm_obj->base.size / 4)) {
 857                         DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
 858                         ret = -EINVAL;
 859                         goto out;
 860                 }
 861
 862                 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
 863
 864                 if (submit->valid)
 865                         continue;
 866
 867                 ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4,
 868                                 submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
 869                 if (ret)
 870                         goto out;
 871         }
 872
 873         submit->nr_cmds = i;
 874
 875         /*
 876          * If using userspace provided seqno fence, validate that the id
 877          * is available before arming sched job.  Since access to fence_idr
 878          * is serialized on the queue lock, the slot should be still avail
 879          * after the job is armed
 880          */
 881         if ((args->flags & MSM_SUBMIT_FENCE_SN_IN) &&
 882                         idr_find(&queue->fence_idr, args->fence)) {
 883                 ret = -EINVAL;
 884                 goto out;
 885         }
 886
 887         drm_sched_job_arm(&submit->base);
 888
 889         submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
 890
 891         if (args->flags & MSM_SUBMIT_FENCE_SN_IN) {
 892                 /*
 893                  * Userspace has assigned the seqno fence that it wants
 894                  * us to use.  It is an error to pick a fence sequence
 895                  * number that is not available.
 896                  */
 897                 submit->fence_id = args->fence;
 898                 ret = idr_alloc_u32(&queue->fence_idr, submit->user_fence,
 899                                     &submit->fence_id, submit->fence_id,
 900                                     GFP_KERNEL);
 901                 /*
 902                  * We've already validated that the fence_id slot is valid,
 903                  * so if idr_alloc_u32 failed, it is a kernel bug
 904                  */
 905                 WARN_ON(ret);
 906         } else {
 907                 /*
 908                  * Allocate an id which can be used by WAIT_FENCE ioctl to map
 909                  * back to the underlying fence.
 910                  */
 911                 submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
 912                                                     submit->user_fence, 1,
 913                                                     INT_MAX, GFP_KERNEL);
 914         }
 915         if (submit->fence_id < 0) {
 916                 ret = submit->fence_id = 0;
 917                 submit->fence_id = 0;
 918         }
 919
 920         if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
 921                 struct sync_file *sync_file = sync_file_create(submit->user_fence);
 922                 if (!sync_file) {
 923                         ret = -ENOMEM;
 924                 } else {
 925                         fd_install(out_fence_fd, sync_file->file);
 926                         args->fence_fd = out_fence_fd;
 927                 }
 928         }
 929
 930         submit_attach_object_fences(submit);
 931
 932         /* The scheduler owns a ref now: */
 933         msm_gem_submit_get(submit);
 934
 935         drm_sched_entity_push_job(&submit->base);
 936
 937         args->fence = submit->fence_id;
 938         queue->last_fence = submit->fence_id;
 939
 940         msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
 941         msm_process_post_deps(post_deps, args->nr_out_syncobjs,
 942                               submit->user_fence);
 943
 944
 945 out:
 946         submit_cleanup(submit, !!ret);
 947         if (has_ww_ticket)
 948                 ww_acquire_fini(&submit->ticket);
 949 out_unlock:
 950         if (ret && (out_fence_fd >= 0))
 951                 put_unused_fd(out_fence_fd);
 952         mutex_unlock(&queue->lock);
 953         if (submit)
 954                 msm_gem_submit_put(submit);
 955 out_post_unlock:
 956         if (!IS_ERR_OR_NULL(post_deps)) {
 957                 for (i = 0; i < args->nr_out_syncobjs; ++i) {
 958                         kfree(post_deps[i].chain);
 959                         drm_syncobj_put(post_deps[i].syncobj);
 960                 }
 961                 kfree(post_deps);
 962         }
 963
 964         if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
 965                 for (i = 0; i < args->nr_in_syncobjs; ++i) {
 966                         if (syncobjs_to_reset[i])
 967                                 drm_syncobj_put(syncobjs_to_reset[i]);
 968                 }
 969                 kfree(syncobjs_to_reset);
 970         }
 971
 972         return ret;
 973 }