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 #define BO_INVALID_FLAGS ~(MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE) 27 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */ 28 #define BO_VALID 0x8000 29 #define BO_LOCKED 0x4000 30 #define BO_PINNED 0x2000 31 32 static inline void __user *to_user_ptr(u64 address) 33 { 34 return (void __user *)(uintptr_t)address; 35 } 36 37 static struct msm_gem_submit *submit_create(struct drm_device *dev, 38 struct msm_gpu *gpu, int nr) 39 { 40 struct msm_gem_submit *submit; 41 int sz = sizeof(*submit) + (nr * sizeof(submit->bos[0])); 42 43 submit = kmalloc(sz, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); 44 if (submit) { 45 submit->dev = dev; 46 submit->gpu = gpu; 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->bo_list); 53 ww_acquire_init(&submit->ticket, &reservation_ww_class); 54 } 55 56 return submit; 57 } 58 59 static int submit_lookup_objects(struct msm_gem_submit *submit, 60 struct drm_msm_gem_submit *args, struct drm_file *file) 61 { 62 unsigned i; 63 int ret = 0; 64 65 spin_lock(&file->table_lock); 66 67 for (i = 0; i < args->nr_bos; i++) { 68 struct drm_msm_gem_submit_bo submit_bo; 69 struct drm_gem_object *obj; 70 struct msm_gem_object *msm_obj; 71 void __user *userptr = 72 to_user_ptr(args->bos + (i * sizeof(submit_bo))); 73 74 ret = copy_from_user(&submit_bo, userptr, sizeof(submit_bo)); 75 if (ret) { 76 ret = -EFAULT; 77 goto out_unlock; 78 } 79 80 if (submit_bo.flags & BO_INVALID_FLAGS) { 81 DRM_ERROR("invalid flags: %x\n", submit_bo.flags); 82 ret = -EINVAL; 83 goto out_unlock; 84 } 85 86 submit->bos[i].flags = submit_bo.flags; 87 /* in validate_objects() we figure out if this is true: */ 88 submit->bos[i].iova = submit_bo.presumed; 89 90 /* normally use drm_gem_object_lookup(), but for bulk lookup 91 * all under single table_lock just hit object_idr directly: 92 */ 93 obj = idr_find(&file->object_idr, submit_bo.handle); 94 if (!obj) { 95 DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i); 96 ret = -EINVAL; 97 goto out_unlock; 98 } 99 100 msm_obj = to_msm_bo(obj); 101 102 if (!list_empty(&msm_obj->submit_entry)) { 103 DRM_ERROR("handle %u at index %u already on submit list\n", 104 submit_bo.handle, i); 105 ret = -EINVAL; 106 goto out_unlock; 107 } 108 109 drm_gem_object_reference(obj); 110 111 submit->bos[i].obj = msm_obj; 112 113 list_add_tail(&msm_obj->submit_entry, &submit->bo_list); 114 } 115 116 out_unlock: 117 submit->nr_bos = i; 118 spin_unlock(&file->table_lock); 119 120 return ret; 121 } 122 123 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i) 124 { 125 struct msm_gem_object *msm_obj = submit->bos[i].obj; 126 127 if (submit->bos[i].flags & BO_PINNED) 128 msm_gem_put_iova(&msm_obj->base, submit->gpu->id); 129 130 if (submit->bos[i].flags & BO_LOCKED) 131 ww_mutex_unlock(&msm_obj->resv->lock); 132 133 if (!(submit->bos[i].flags & BO_VALID)) 134 submit->bos[i].iova = 0; 135 136 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED); 137 } 138 139 /* This is where we make sure all the bo's are reserved and pin'd: */ 140 static int submit_validate_objects(struct msm_gem_submit *submit) 141 { 142 int contended, slow_locked = -1, i, ret = 0; 143 144 retry: 145 submit->valid = true; 146 147 for (i = 0; i < submit->nr_bos; i++) { 148 struct msm_gem_object *msm_obj = submit->bos[i].obj; 149 uint32_t iova; 150 151 if (slow_locked == i) 152 slow_locked = -1; 153 154 contended = i; 155 156 if (!(submit->bos[i].flags & BO_LOCKED)) { 157 ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock, 158 &submit->ticket); 159 if (ret) 160 goto fail; 161 submit->bos[i].flags |= BO_LOCKED; 162 } 163 164 165 /* if locking succeeded, pin bo: */ 166 ret = msm_gem_get_iova(&msm_obj->base, 167 submit->gpu->id, &iova); 168 169 /* this would break the logic in the fail path.. there is no 170 * reason for this to happen, but just to be on the safe side 171 * let's notice if this starts happening in the future: 172 */ 173 WARN_ON(ret == -EDEADLK); 174 175 if (ret) 176 goto fail; 177 178 submit->bos[i].flags |= BO_PINNED; 179 180 if (iova == submit->bos[i].iova) { 181 submit->bos[i].flags |= BO_VALID; 182 } else { 183 submit->bos[i].iova = iova; 184 submit->bos[i].flags &= ~BO_VALID; 185 submit->valid = false; 186 } 187 } 188 189 ww_acquire_done(&submit->ticket); 190 191 return 0; 192 193 fail: 194 for (; i >= 0; i--) 195 submit_unlock_unpin_bo(submit, i); 196 197 if (slow_locked > 0) 198 submit_unlock_unpin_bo(submit, slow_locked); 199 200 if (ret == -EDEADLK) { 201 struct msm_gem_object *msm_obj = submit->bos[contended].obj; 202 /* we lost out in a seqno race, lock and retry.. */ 203 ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock, 204 &submit->ticket); 205 if (!ret) { 206 submit->bos[contended].flags |= BO_LOCKED; 207 slow_locked = contended; 208 goto retry; 209 } 210 } 211 212 return ret; 213 } 214 215 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx, 216 struct msm_gem_object **obj, uint32_t *iova, bool *valid) 217 { 218 if (idx >= submit->nr_bos) { 219 DRM_ERROR("invalid buffer index: %u (out of %u)\n", 220 idx, submit->nr_bos); 221 return -EINVAL; 222 } 223 224 if (obj) 225 *obj = submit->bos[idx].obj; 226 if (iova) 227 *iova = submit->bos[idx].iova; 228 if (valid) 229 *valid = !!(submit->bos[idx].flags & BO_VALID); 230 231 return 0; 232 } 233 234 /* process the reloc's and patch up the cmdstream as needed: */ 235 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj, 236 uint32_t offset, uint32_t nr_relocs, uint64_t relocs) 237 { 238 uint32_t i, last_offset = 0; 239 uint32_t *ptr; 240 int ret; 241 242 if (offset % 4) { 243 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset); 244 return -EINVAL; 245 } 246 247 /* For now, just map the entire thing. Eventually we probably 248 * to do it page-by-page, w/ kmap() if not vmap()d.. 249 */ 250 ptr = msm_gem_vaddr(&obj->base); 251 252 if (IS_ERR(ptr)) { 253 ret = PTR_ERR(ptr); 254 DBG("failed to map: %d", ret); 255 return ret; 256 } 257 258 for (i = 0; i < nr_relocs; i++) { 259 struct drm_msm_gem_submit_reloc submit_reloc; 260 void __user *userptr = 261 to_user_ptr(relocs + (i * sizeof(submit_reloc))); 262 uint32_t iova, off; 263 bool valid; 264 265 ret = copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc)); 266 if (ret) 267 return -EFAULT; 268 269 if (submit_reloc.submit_offset % 4) { 270 DRM_ERROR("non-aligned reloc offset: %u\n", 271 submit_reloc.submit_offset); 272 return -EINVAL; 273 } 274 275 /* offset in dwords: */ 276 off = submit_reloc.submit_offset / 4; 277 278 if ((off >= (obj->base.size / 4)) || 279 (off < last_offset)) { 280 DRM_ERROR("invalid offset %u at reloc %u\n", off, i); 281 return -EINVAL; 282 } 283 284 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid); 285 if (ret) 286 return ret; 287 288 if (valid) 289 continue; 290 291 iova += submit_reloc.reloc_offset; 292 293 if (submit_reloc.shift < 0) 294 iova >>= -submit_reloc.shift; 295 else 296 iova <<= submit_reloc.shift; 297 298 ptr[off] = iova | submit_reloc.or; 299 300 last_offset = off; 301 } 302 303 return 0; 304 } 305 306 static void submit_cleanup(struct msm_gem_submit *submit, bool fail) 307 { 308 unsigned i; 309 310 mutex_lock(&submit->dev->struct_mutex); 311 for (i = 0; i < submit->nr_bos; i++) { 312 struct msm_gem_object *msm_obj = submit->bos[i].obj; 313 submit_unlock_unpin_bo(submit, i); 314 list_del_init(&msm_obj->submit_entry); 315 drm_gem_object_unreference(&msm_obj->base); 316 } 317 mutex_unlock(&submit->dev->struct_mutex); 318 319 ww_acquire_fini(&submit->ticket); 320 kfree(submit); 321 } 322 323 int msm_ioctl_gem_submit(struct drm_device *dev, void *data, 324 struct drm_file *file) 325 { 326 struct msm_drm_private *priv = dev->dev_private; 327 struct drm_msm_gem_submit *args = data; 328 struct msm_file_private *ctx = file->driver_priv; 329 struct msm_gem_submit *submit; 330 struct msm_gpu *gpu; 331 unsigned i; 332 int ret; 333 334 /* for now, we just have 3d pipe.. eventually this would need to 335 * be more clever to dispatch to appropriate gpu module: 336 */ 337 if (args->pipe != MSM_PIPE_3D0) 338 return -EINVAL; 339 340 gpu = priv->gpu; 341 342 if (args->nr_cmds > MAX_CMDS) 343 return -EINVAL; 344 345 submit = submit_create(dev, gpu, args->nr_bos); 346 if (!submit) { 347 ret = -ENOMEM; 348 goto out; 349 } 350 351 ret = submit_lookup_objects(submit, args, file); 352 if (ret) 353 goto out; 354 355 ret = submit_validate_objects(submit); 356 if (ret) 357 goto out; 358 359 for (i = 0; i < args->nr_cmds; i++) { 360 struct drm_msm_gem_submit_cmd submit_cmd; 361 void __user *userptr = 362 to_user_ptr(args->cmds + (i * sizeof(submit_cmd))); 363 struct msm_gem_object *msm_obj; 364 uint32_t iova; 365 366 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd)); 367 if (ret) { 368 ret = -EFAULT; 369 goto out; 370 } 371 372 ret = submit_bo(submit, submit_cmd.submit_idx, 373 &msm_obj, &iova, NULL); 374 if (ret) 375 goto out; 376 377 if (submit_cmd.size % 4) { 378 DRM_ERROR("non-aligned cmdstream buffer size: %u\n", 379 submit_cmd.size); 380 ret = -EINVAL; 381 goto out; 382 } 383 384 if ((submit_cmd.size + submit_cmd.submit_offset) >= 385 msm_obj->base.size) { 386 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size); 387 ret = -EINVAL; 388 goto out; 389 } 390 391 submit->cmd[i].type = submit_cmd.type; 392 submit->cmd[i].size = submit_cmd.size / 4; 393 submit->cmd[i].iova = iova + submit_cmd.submit_offset; 394 395 if (submit->valid) 396 continue; 397 398 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset, 399 submit_cmd.nr_relocs, submit_cmd.relocs); 400 if (ret) 401 goto out; 402 } 403 404 submit->nr_cmds = i; 405 406 ret = msm_gpu_submit(gpu, submit, ctx); 407 408 args->fence = submit->fence; 409 410 out: 411 if (submit) 412 submit_cleanup(submit, !!ret); 413 return ret; 414 } 415