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