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