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, 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 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) { 408 in_fence = sync_file_get_fence(args->fence_fd); 409 410 if (!in_fence) 411 return -EINVAL; 412 413 /* TODO if we get an array-fence due to userspace merging multiple 414 * fences, we need a way to determine if all the backing fences 415 * are from our own context.. 416 */ 417 418 if (in_fence->context != gpu->fctx->context) { 419 ret = dma_fence_wait(in_fence, true); 420 if (ret) 421 return ret; 422 } 423 } 424 425 ret = mutex_lock_interruptible(&dev->struct_mutex); 426 if (ret) 427 return ret; 428 429 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 430 out_fence_fd = get_unused_fd_flags(O_CLOEXEC); 431 if (out_fence_fd < 0) { 432 ret = out_fence_fd; 433 goto out_unlock; 434 } 435 } 436 priv->struct_mutex_task = current; 437 438 submit = submit_create(dev, gpu, args->nr_bos, args->nr_cmds); 439 if (!submit) { 440 ret = -ENOMEM; 441 goto out_unlock; 442 } 443 444 ret = submit_lookup_objects(submit, args, file); 445 if (ret) 446 goto out; 447 448 ret = submit_lock_objects(submit); 449 if (ret) 450 goto out; 451 452 if (!(args->fence & MSM_SUBMIT_NO_IMPLICIT)) { 453 ret = submit_fence_sync(submit); 454 if (ret) 455 goto out; 456 } 457 458 ret = submit_pin_objects(submit); 459 if (ret) 460 goto out; 461 462 for (i = 0; i < args->nr_cmds; i++) { 463 struct drm_msm_gem_submit_cmd submit_cmd; 464 void __user *userptr = 465 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd))); 466 struct msm_gem_object *msm_obj; 467 uint64_t iova; 468 469 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd)); 470 if (ret) { 471 ret = -EFAULT; 472 goto out; 473 } 474 475 /* validate input from userspace: */ 476 switch (submit_cmd.type) { 477 case MSM_SUBMIT_CMD_BUF: 478 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 479 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 480 break; 481 default: 482 DRM_ERROR("invalid type: %08x\n", submit_cmd.type); 483 ret = -EINVAL; 484 goto out; 485 } 486 487 ret = submit_bo(submit, submit_cmd.submit_idx, 488 &msm_obj, &iova, NULL); 489 if (ret) 490 goto out; 491 492 if (submit_cmd.size % 4) { 493 DRM_ERROR("non-aligned cmdstream buffer size: %u\n", 494 submit_cmd.size); 495 ret = -EINVAL; 496 goto out; 497 } 498 499 if ((submit_cmd.size + submit_cmd.submit_offset) >= 500 msm_obj->base.size) { 501 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size); 502 ret = -EINVAL; 503 goto out; 504 } 505 506 submit->cmd[i].type = submit_cmd.type; 507 submit->cmd[i].size = submit_cmd.size / 4; 508 submit->cmd[i].iova = iova + submit_cmd.submit_offset; 509 submit->cmd[i].idx = submit_cmd.submit_idx; 510 511 if (submit->valid) 512 continue; 513 514 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset, 515 submit_cmd.nr_relocs, submit_cmd.relocs); 516 if (ret) 517 goto out; 518 } 519 520 submit->nr_cmds = i; 521 522 submit->fence = msm_fence_alloc(gpu->fctx); 523 if (IS_ERR(submit->fence)) { 524 ret = PTR_ERR(submit->fence); 525 submit->fence = NULL; 526 goto out; 527 } 528 529 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 530 sync_file = sync_file_create(submit->fence); 531 if (!sync_file) { 532 ret = -ENOMEM; 533 goto out; 534 } 535 } 536 537 msm_gpu_submit(gpu, submit, ctx); 538 539 args->fence = submit->fence->seqno; 540 541 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 542 fd_install(out_fence_fd, sync_file->file); 543 args->fence_fd = out_fence_fd; 544 } 545 546 out: 547 if (in_fence) 548 dma_fence_put(in_fence); 549 submit_cleanup(submit); 550 if (ret) 551 msm_gem_submit_free(submit); 552 out_unlock: 553 if (ret && (out_fence_fd >= 0)) 554 put_unused_fd(out_fence_fd); 555 priv->struct_mutex_task = NULL; 556 mutex_unlock(&dev->struct_mutex); 557 return ret; 558 } 559