drivers/gpu/drm/msm/msm_gem_submit.c

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