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