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 DBG("invalid flags: %x", 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 DBG("invalid handle %u at index %u", 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 DBG("handle %u at index %u already on submit list", 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 DBG("invalid buffer index: %u (out of %u)", idx, submit->nr_bos); 220 return EINVAL; 221 } 222 223 if (obj) 224 *obj = submit->bos[idx].obj; 225 if (iova) 226 *iova = submit->bos[idx].iova; 227 if (valid) 228 *valid = !!(submit->bos[idx].flags & BO_VALID); 229 230 return 0; 231 } 232 233 /* process the reloc's and patch up the cmdstream as needed: */ 234 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj, 235 uint32_t offset, uint32_t nr_relocs, uint64_t relocs) 236 { 237 uint32_t i, last_offset = 0; 238 uint32_t *ptr; 239 int ret; 240 241 if (offset % 4) { 242 DBG("non-aligned cmdstream buffer: %u", offset); 243 return -EINVAL; 244 } 245 246 /* For now, just map the entire thing. Eventually we probably 247 * to do it page-by-page, w/ kmap() if not vmap()d.. 248 */ 249 ptr = msm_gem_vaddr(&obj->base); 250 251 if (IS_ERR(ptr)) { 252 ret = PTR_ERR(ptr); 253 DBG("failed to map: %d", ret); 254 return ret; 255 } 256 257 for (i = 0; i < nr_relocs; i++) { 258 struct drm_msm_gem_submit_reloc submit_reloc; 259 void __user *userptr = 260 to_user_ptr(relocs + (i * sizeof(submit_reloc))); 261 uint32_t iova, off; 262 bool valid; 263 264 ret = copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc)); 265 if (ret) 266 return -EFAULT; 267 268 if (submit_reloc.submit_offset % 4) { 269 DBG("non-aligned reloc offset: %u", 270 submit_reloc.submit_offset); 271 return -EINVAL; 272 } 273 274 /* offset in dwords: */ 275 off = submit_reloc.submit_offset / 4; 276 277 if ((off >= (obj->base.size / 4)) || 278 (off < last_offset)) { 279 DBG("invalid offset %u at reloc %u", off, i); 280 return -EINVAL; 281 } 282 283 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid); 284 if (ret) 285 return ret; 286 287 if (valid) 288 continue; 289 290 iova += submit_reloc.reloc_offset; 291 292 if (submit_reloc.shift < 0) 293 iova >>= -submit_reloc.shift; 294 else 295 iova <<= submit_reloc.shift; 296 297 ptr[off] = iova | submit_reloc.or; 298 299 last_offset = off; 300 } 301 302 return 0; 303 } 304 305 static void submit_cleanup(struct msm_gem_submit *submit, bool fail) 306 { 307 unsigned i; 308 309 mutex_lock(&submit->dev->struct_mutex); 310 for (i = 0; i < submit->nr_bos; i++) { 311 struct msm_gem_object *msm_obj = submit->bos[i].obj; 312 submit_unlock_unpin_bo(submit, i); 313 list_del_init(&msm_obj->submit_entry); 314 drm_gem_object_unreference(&msm_obj->base); 315 } 316 mutex_unlock(&submit->dev->struct_mutex); 317 318 ww_acquire_fini(&submit->ticket); 319 kfree(submit); 320 } 321 322 int msm_ioctl_gem_submit(struct drm_device *dev, void *data, 323 struct drm_file *file) 324 { 325 struct msm_drm_private *priv = dev->dev_private; 326 struct drm_msm_gem_submit *args = data; 327 struct msm_file_private *ctx = file->driver_priv; 328 struct msm_gem_submit *submit; 329 struct msm_gpu *gpu; 330 unsigned i; 331 int ret; 332 333 /* for now, we just have 3d pipe.. eventually this would need to 334 * be more clever to dispatch to appropriate gpu module: 335 */ 336 if (args->pipe != MSM_PIPE_3D0) 337 return -EINVAL; 338 339 gpu = priv->gpu; 340 341 if (args->nr_cmds > MAX_CMDS) 342 return -EINVAL; 343 344 submit = submit_create(dev, gpu, args->nr_bos); 345 if (!submit) { 346 ret = -ENOMEM; 347 goto out; 348 } 349 350 ret = submit_lookup_objects(submit, args, file); 351 if (ret) 352 goto out; 353 354 ret = submit_validate_objects(submit); 355 if (ret) 356 goto out; 357 358 for (i = 0; i < args->nr_cmds; i++) { 359 struct drm_msm_gem_submit_cmd submit_cmd; 360 void __user *userptr = 361 to_user_ptr(args->cmds + (i * sizeof(submit_cmd))); 362 struct msm_gem_object *msm_obj; 363 uint32_t iova; 364 365 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd)); 366 if (ret) { 367 ret = -EFAULT; 368 goto out; 369 } 370 371 ret = submit_bo(submit, submit_cmd.submit_idx, 372 &msm_obj, &iova, NULL); 373 if (ret) 374 goto out; 375 376 if (submit_cmd.size % 4) { 377 DBG("non-aligned cmdstream buffer size: %u", 378 submit_cmd.size); 379 ret = -EINVAL; 380 goto out; 381 } 382 383 if (submit_cmd.size >= msm_obj->base.size) { 384 DBG("invalid cmdstream size: %u", submit_cmd.size); 385 ret = -EINVAL; 386 goto out; 387 } 388 389 submit->cmd[i].type = submit_cmd.type; 390 submit->cmd[i].size = submit_cmd.size / 4; 391 submit->cmd[i].iova = iova + submit_cmd.submit_offset; 392 393 if (submit->valid) 394 continue; 395 396 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset, 397 submit_cmd.nr_relocs, submit_cmd.relocs); 398 if (ret) 399 goto out; 400 } 401 402 submit->nr_cmds = i; 403 404 ret = msm_gpu_submit(gpu, submit, ctx); 405 406 args->fence = submit->fence; 407 408 out: 409 if (submit) 410 submit_cleanup(submit, !!ret); 411 return ret; 412 } 413