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 "msm_drv.h" 19 #include "msm_gpu.h" 20 #include "msm_gem.h" 21 22 /* 23 * Cmdstream submission: 24 */ 25 26 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */ 27 #define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */ 28 #define BO_LOCKED 0x4000 29 #define BO_PINNED 0x2000 30 31 static struct msm_gem_submit *submit_create(struct drm_device *dev, 32 struct msm_gpu *gpu, int nr_bos, int nr_cmds) 33 { 34 struct msm_gem_submit *submit; 35 int sz = sizeof(*submit) + (nr_bos * sizeof(submit->bos[0])) + 36 (nr_cmds * sizeof(*submit->cmd)); 37 38 submit = kmalloc(sz, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); 39 if (!submit) 40 return NULL; 41 42 submit->dev = dev; 43 submit->gpu = gpu; 44 submit->fence = NULL; 45 submit->pid = get_pid(task_pid(current)); 46 submit->cmd = (void *)&submit->bos[nr_bos]; 47 48 /* initially, until copy_from_user() and bo lookup succeeds: */ 49 submit->nr_bos = 0; 50 submit->nr_cmds = 0; 51 52 INIT_LIST_HEAD(&submit->node); 53 INIT_LIST_HEAD(&submit->bo_list); 54 ww_acquire_init(&submit->ticket, &reservation_ww_class); 55 56 return submit; 57 } 58 59 void msm_gem_submit_free(struct msm_gem_submit *submit) 60 { 61 fence_put(submit->fence); 62 list_del(&submit->node); 63 put_pid(submit->pid); 64 kfree(submit); 65 } 66 67 static inline unsigned long __must_check 68 copy_from_user_inatomic(void *to, const void __user *from, unsigned long n) 69 { 70 if (access_ok(VERIFY_READ, from, n)) 71 return __copy_from_user_inatomic(to, from, n); 72 return -EFAULT; 73 } 74 75 static int submit_lookup_objects(struct msm_gem_submit *submit, 76 struct drm_msm_gem_submit *args, struct drm_file *file) 77 { 78 unsigned i; 79 int ret = 0; 80 81 spin_lock(&file->table_lock); 82 pagefault_disable(); 83 84 for (i = 0; i < args->nr_bos; i++) { 85 struct drm_msm_gem_submit_bo submit_bo; 86 struct drm_gem_object *obj; 87 struct msm_gem_object *msm_obj; 88 void __user *userptr = 89 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo))); 90 91 /* make sure we don't have garbage flags, in case we hit 92 * error path before flags is initialized: 93 */ 94 submit->bos[i].flags = 0; 95 96 ret = copy_from_user_inatomic(&submit_bo, userptr, sizeof(submit_bo)); 97 if (unlikely(ret)) { 98 pagefault_enable(); 99 spin_unlock(&file->table_lock); 100 ret = copy_from_user(&submit_bo, userptr, sizeof(submit_bo)); 101 if (ret) 102 goto out; 103 spin_lock(&file->table_lock); 104 pagefault_disable(); 105 } 106 107 if (submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) { 108 DRM_ERROR("invalid flags: %x\n", submit_bo.flags); 109 ret = -EINVAL; 110 goto out_unlock; 111 } 112 113 submit->bos[i].flags = submit_bo.flags; 114 /* in validate_objects() we figure out if this is true: */ 115 submit->bos[i].iova = submit_bo.presumed; 116 117 /* normally use drm_gem_object_lookup(), but for bulk lookup 118 * all under single table_lock just hit object_idr directly: 119 */ 120 obj = idr_find(&file->object_idr, submit_bo.handle); 121 if (!obj) { 122 DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i); 123 ret = -EINVAL; 124 goto out_unlock; 125 } 126 127 msm_obj = to_msm_bo(obj); 128 129 if (!list_empty(&msm_obj->submit_entry)) { 130 DRM_ERROR("handle %u at index %u already on submit list\n", 131 submit_bo.handle, i); 132 ret = -EINVAL; 133 goto out_unlock; 134 } 135 136 drm_gem_object_reference(obj); 137 138 submit->bos[i].obj = msm_obj; 139 140 list_add_tail(&msm_obj->submit_entry, &submit->bo_list); 141 } 142 143 out_unlock: 144 pagefault_enable(); 145 spin_unlock(&file->table_lock); 146 147 out: 148 submit->nr_bos = i; 149 150 return ret; 151 } 152 153 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i) 154 { 155 struct msm_gem_object *msm_obj = submit->bos[i].obj; 156 157 if (submit->bos[i].flags & BO_PINNED) 158 msm_gem_put_iova(&msm_obj->base, submit->gpu->id); 159 160 if (submit->bos[i].flags & BO_LOCKED) 161 ww_mutex_unlock(&msm_obj->resv->lock); 162 163 if (!(submit->bos[i].flags & BO_VALID)) 164 submit->bos[i].iova = 0; 165 166 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED); 167 } 168 169 /* This is where we make sure all the bo's are reserved and pin'd: */ 170 static int submit_lock_objects(struct msm_gem_submit *submit) 171 { 172 int contended, slow_locked = -1, i, ret = 0; 173 174 retry: 175 for (i = 0; i < submit->nr_bos; i++) { 176 struct msm_gem_object *msm_obj = submit->bos[i].obj; 177 178 if (slow_locked == i) 179 slow_locked = -1; 180 181 contended = i; 182 183 if (!(submit->bos[i].flags & BO_LOCKED)) { 184 ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock, 185 &submit->ticket); 186 if (ret) 187 goto fail; 188 submit->bos[i].flags |= BO_LOCKED; 189 } 190 } 191 192 ww_acquire_done(&submit->ticket); 193 194 return 0; 195 196 fail: 197 for (; i >= 0; i--) 198 submit_unlock_unpin_bo(submit, i); 199 200 if (slow_locked > 0) 201 submit_unlock_unpin_bo(submit, slow_locked); 202 203 if (ret == -EDEADLK) { 204 struct msm_gem_object *msm_obj = submit->bos[contended].obj; 205 /* we lost out in a seqno race, lock and retry.. */ 206 ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock, 207 &submit->ticket); 208 if (!ret) { 209 submit->bos[contended].flags |= BO_LOCKED; 210 slow_locked = contended; 211 goto retry; 212 } 213 } 214 215 return ret; 216 } 217 218 static int submit_fence_sync(struct msm_gem_submit *submit) 219 { 220 int i, ret = 0; 221 222 for (i = 0; i < submit->nr_bos; i++) { 223 struct msm_gem_object *msm_obj = submit->bos[i].obj; 224 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE; 225 226 ret = msm_gem_sync_object(&msm_obj->base, submit->gpu->fctx, write); 227 if (ret) 228 break; 229 } 230 231 return ret; 232 } 233 234 static int submit_pin_objects(struct msm_gem_submit *submit) 235 { 236 int i, ret = 0; 237 238 submit->valid = true; 239 240 for (i = 0; i < submit->nr_bos; i++) { 241 struct msm_gem_object *msm_obj = submit->bos[i].obj; 242 uint32_t iova; 243 244 /* if locking succeeded, pin bo: */ 245 ret = msm_gem_get_iova_locked(&msm_obj->base, 246 submit->gpu->id, &iova); 247 248 if (ret) 249 break; 250 251 submit->bos[i].flags |= BO_PINNED; 252 253 if (iova == submit->bos[i].iova) { 254 submit->bos[i].flags |= BO_VALID; 255 } else { 256 submit->bos[i].iova = iova; 257 /* iova changed, so address in cmdstream is not valid: */ 258 submit->bos[i].flags &= ~BO_VALID; 259 submit->valid = false; 260 } 261 } 262 263 return ret; 264 } 265 266 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx, 267 struct msm_gem_object **obj, uint32_t *iova, bool *valid) 268 { 269 if (idx >= submit->nr_bos) { 270 DRM_ERROR("invalid buffer index: %u (out of %u)\n", 271 idx, submit->nr_bos); 272 return -EINVAL; 273 } 274 275 if (obj) 276 *obj = submit->bos[idx].obj; 277 if (iova) 278 *iova = submit->bos[idx].iova; 279 if (valid) 280 *valid = !!(submit->bos[idx].flags & BO_VALID); 281 282 return 0; 283 } 284 285 /* process the reloc's and patch up the cmdstream as needed: */ 286 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj, 287 uint32_t offset, uint32_t nr_relocs, uint64_t relocs) 288 { 289 uint32_t i, last_offset = 0; 290 uint32_t *ptr; 291 int ret; 292 293 if (offset % 4) { 294 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset); 295 return -EINVAL; 296 } 297 298 /* For now, just map the entire thing. Eventually we probably 299 * to do it page-by-page, w/ kmap() if not vmap()d.. 300 */ 301 ptr = msm_gem_get_vaddr_locked(&obj->base); 302 303 if (IS_ERR(ptr)) { 304 ret = PTR_ERR(ptr); 305 DBG("failed to map: %d", ret); 306 return ret; 307 } 308 309 for (i = 0; i < nr_relocs; i++) { 310 struct drm_msm_gem_submit_reloc submit_reloc; 311 void __user *userptr = 312 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc))); 313 uint32_t iova, off; 314 bool valid; 315 316 ret = copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc)); 317 if (ret) 318 return -EFAULT; 319 320 if (submit_reloc.submit_offset % 4) { 321 DRM_ERROR("non-aligned reloc offset: %u\n", 322 submit_reloc.submit_offset); 323 return -EINVAL; 324 } 325 326 /* offset in dwords: */ 327 off = submit_reloc.submit_offset / 4; 328 329 if ((off >= (obj->base.size / 4)) || 330 (off < last_offset)) { 331 DRM_ERROR("invalid offset %u at reloc %u\n", off, i); 332 return -EINVAL; 333 } 334 335 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid); 336 if (ret) 337 return ret; 338 339 if (valid) 340 continue; 341 342 iova += submit_reloc.reloc_offset; 343 344 if (submit_reloc.shift < 0) 345 iova >>= -submit_reloc.shift; 346 else 347 iova <<= submit_reloc.shift; 348 349 ptr[off] = iova | submit_reloc.or; 350 351 last_offset = off; 352 } 353 354 msm_gem_put_vaddr_locked(&obj->base); 355 356 return 0; 357 } 358 359 static void submit_cleanup(struct msm_gem_submit *submit) 360 { 361 unsigned i; 362 363 for (i = 0; i < submit->nr_bos; i++) { 364 struct msm_gem_object *msm_obj = submit->bos[i].obj; 365 submit_unlock_unpin_bo(submit, i); 366 list_del_init(&msm_obj->submit_entry); 367 drm_gem_object_unreference(&msm_obj->base); 368 } 369 370 ww_acquire_fini(&submit->ticket); 371 } 372 373 int msm_ioctl_gem_submit(struct drm_device *dev, void *data, 374 struct drm_file *file) 375 { 376 struct msm_drm_private *priv = dev->dev_private; 377 struct drm_msm_gem_submit *args = data; 378 struct msm_file_private *ctx = file->driver_priv; 379 struct msm_gem_submit *submit; 380 struct msm_gpu *gpu = priv->gpu; 381 unsigned i; 382 int ret; 383 384 if (!gpu) 385 return -ENXIO; 386 387 /* for now, we just have 3d pipe.. eventually this would need to 388 * be more clever to dispatch to appropriate gpu module: 389 */ 390 if (args->pipe != MSM_PIPE_3D0) 391 return -EINVAL; 392 393 ret = mutex_lock_interruptible(&dev->struct_mutex); 394 if (ret) 395 return ret; 396 397 priv->struct_mutex_task = current; 398 399 submit = submit_create(dev, gpu, args->nr_bos, args->nr_cmds); 400 if (!submit) { 401 ret = -ENOMEM; 402 goto out_unlock; 403 } 404 405 ret = submit_lookup_objects(submit, args, file); 406 if (ret) 407 goto out; 408 409 ret = submit_lock_objects(submit); 410 if (ret) 411 goto out; 412 413 ret = submit_fence_sync(submit); 414 if (ret) 415 goto out; 416 417 ret = submit_pin_objects(submit); 418 if (ret) 419 goto out; 420 421 for (i = 0; i < args->nr_cmds; i++) { 422 struct drm_msm_gem_submit_cmd submit_cmd; 423 void __user *userptr = 424 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd))); 425 struct msm_gem_object *msm_obj; 426 uint32_t iova; 427 428 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd)); 429 if (ret) { 430 ret = -EFAULT; 431 goto out; 432 } 433 434 /* validate input from userspace: */ 435 switch (submit_cmd.type) { 436 case MSM_SUBMIT_CMD_BUF: 437 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 438 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 439 break; 440 default: 441 DRM_ERROR("invalid type: %08x\n", submit_cmd.type); 442 ret = -EINVAL; 443 goto out; 444 } 445 446 ret = submit_bo(submit, submit_cmd.submit_idx, 447 &msm_obj, &iova, NULL); 448 if (ret) 449 goto out; 450 451 if (submit_cmd.size % 4) { 452 DRM_ERROR("non-aligned cmdstream buffer size: %u\n", 453 submit_cmd.size); 454 ret = -EINVAL; 455 goto out; 456 } 457 458 if ((submit_cmd.size + submit_cmd.submit_offset) >= 459 msm_obj->base.size) { 460 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size); 461 ret = -EINVAL; 462 goto out; 463 } 464 465 submit->cmd[i].type = submit_cmd.type; 466 submit->cmd[i].size = submit_cmd.size / 4; 467 submit->cmd[i].iova = iova + submit_cmd.submit_offset; 468 submit->cmd[i].idx = submit_cmd.submit_idx; 469 470 if (submit->valid) 471 continue; 472 473 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset, 474 submit_cmd.nr_relocs, submit_cmd.relocs); 475 if (ret) 476 goto out; 477 } 478 479 submit->nr_cmds = i; 480 481 ret = msm_gpu_submit(gpu, submit, ctx); 482 483 args->fence = submit->fence->seqno; 484 485 out: 486 submit_cleanup(submit); 487 if (ret) 488 msm_gem_submit_free(submit); 489 out_unlock: 490 priv->struct_mutex_task = NULL; 491 mutex_unlock(&dev->struct_mutex); 492 return ret; 493 } 494