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