drivers/gpu/drm/msm/msm_gem_submit.c

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