drivers/gpu/drm/msm/msm_gem.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 <linux/spinlock.h>
  19 #include <linux/shmem_fs.h>
  20
  21 #include "msm_drv.h"
  22 #include "msm_gem.h"
  23 #include "msm_gpu.h"
  24
  25
  26 /* called with dev->struct_mutex held */
  27 static struct page **get_pages(struct drm_gem_object *obj)
  28 {
  29         struct msm_gem_object *msm_obj = to_msm_bo(obj);
  30
  31         if (!msm_obj->pages) {
  32                 struct drm_device *dev = obj->dev;
  33                 struct page **p = drm_gem_get_pages(obj, 0);
  34                 int npages = obj->size >> PAGE_SHIFT;
  35
  36                 if (IS_ERR(p)) {
  37                         dev_err(dev->dev, "could not get pages: %ld\n",
  38                                         PTR_ERR(p));
  39                         return p;
  40                 }
  41
  42                 msm_obj->sgt = drm_prime_pages_to_sg(p, npages);
  43                 if (IS_ERR(msm_obj->sgt)) {
  44                         dev_err(dev->dev, "failed to allocate sgt\n");
  45                         return ERR_CAST(msm_obj->sgt);
  46                 }
  47
  48                 msm_obj->pages = p;
  49
  50                 /* For non-cached buffers, ensure the new pages are clean
  51                  * because display controller, GPU, etc. are not coherent:
  52                  */
  53                 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
  54                         dma_map_sg(dev->dev, msm_obj->sgt->sgl,
  55                                         msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
  56         }
  57
  58         return msm_obj->pages;
  59 }
  60
  61 static void put_pages(struct drm_gem_object *obj)
  62 {
  63         struct msm_gem_object *msm_obj = to_msm_bo(obj);
  64
  65         if (msm_obj->pages) {
  66                 /* For non-cached buffers, ensure the new pages are clean
  67                  * because display controller, GPU, etc. are not coherent:
  68                  */
  69                 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
  70                         dma_unmap_sg(obj->dev->dev, msm_obj->sgt->sgl,
  71                                         msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
  72                 sg_free_table(msm_obj->sgt);
  73                 kfree(msm_obj->sgt);
  74
  75                 drm_gem_put_pages(obj, msm_obj->pages, true, false);
  76                 msm_obj->pages = NULL;
  77         }
  78 }
  79
  80 int msm_gem_mmap_obj(struct drm_gem_object *obj,
  81                 struct vm_area_struct *vma)
  82 {
  83         struct msm_gem_object *msm_obj = to_msm_bo(obj);
  84
  85         vma->vm_flags &= ~VM_PFNMAP;
  86         vma->vm_flags |= VM_MIXEDMAP;
  87
  88         if (msm_obj->flags & MSM_BO_WC) {
  89                 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
  90         } else if (msm_obj->flags & MSM_BO_UNCACHED) {
  91                 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
  92         } else {
  93                 /*
  94                  * Shunt off cached objs to shmem file so they have their own
  95                  * address_space (so unmap_mapping_range does what we want,
  96                  * in particular in the case of mmap'd dmabufs)
  97                  */
  98                 fput(vma->vm_file);
  99                 get_file(obj->filp);
 100                 vma->vm_pgoff = 0;
 101                 vma->vm_file  = obj->filp;
 102
 103                 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 104         }
 105
 106         return 0;
 107 }
 108
 109 int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
 110 {
 111         int ret;
 112
 113         ret = drm_gem_mmap(filp, vma);
 114         if (ret) {
 115                 DBG("mmap failed: %d", ret);
 116                 return ret;
 117         }
 118
 119         return msm_gem_mmap_obj(vma->vm_private_data, vma);
 120 }
 121
 122 int msm_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 123 {
 124         struct drm_gem_object *obj = vma->vm_private_data;
 125         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 126         struct drm_device *dev = obj->dev;
 127         struct page **pages;
 128         unsigned long pfn;
 129         pgoff_t pgoff;
 130         int ret;
 131
 132         /* Make sure we don't parallel update on a fault, nor move or remove
 133          * something from beneath our feet
 134          */
 135         ret = mutex_lock_interruptible(&dev->struct_mutex);
 136         if (ret)
 137                 goto out;
 138
 139         /* make sure we have pages attached now */
 140         pages = get_pages(obj);
 141         if (IS_ERR(pages)) {
 142                 ret = PTR_ERR(pages);
 143                 goto out_unlock;
 144         }
 145
 146         /* We don't use vmf->pgoff since that has the fake offset: */
 147         pgoff = ((unsigned long)vmf->virtual_address -
 148                         vma->vm_start) >> PAGE_SHIFT;
 149
 150         pfn = page_to_pfn(msm_obj->pages[pgoff]);
 151
 152         VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
 153                         pfn, pfn << PAGE_SHIFT);
 154
 155         ret = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, pfn);
 156
 157 out_unlock:
 158         mutex_unlock(&dev->struct_mutex);
 159 out:
 160         switch (ret) {
 161         case -EAGAIN:
 162         case 0:
 163         case -ERESTARTSYS:
 164         case -EINTR:
 165                 return VM_FAULT_NOPAGE;
 166         case -ENOMEM:
 167                 return VM_FAULT_OOM;
 168         default:
 169                 return VM_FAULT_SIGBUS;
 170         }
 171 }
 172
 173 /** get mmap offset */
 174 static uint64_t mmap_offset(struct drm_gem_object *obj)
 175 {
 176         struct drm_device *dev = obj->dev;
 177         int ret;
 178
 179         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 180
 181         /* Make it mmapable */
 182         ret = drm_gem_create_mmap_offset(obj);
 183
 184         if (ret) {
 185                 dev_err(dev->dev, "could not allocate mmap offset\n");
 186                 return 0;
 187         }
 188
 189         return drm_vma_node_offset_addr(&obj->vma_node);
 190 }
 191
 192 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
 193 {
 194         uint64_t offset;
 195         mutex_lock(&obj->dev->struct_mutex);
 196         offset = mmap_offset(obj);
 197         mutex_unlock(&obj->dev->struct_mutex);
 198         return offset;
 199 }
 200
 201 /* helpers for dealing w/ iommu: */
 202 static int map_range(struct iommu_domain *domain, unsigned int iova,
 203                 struct sg_table *sgt, unsigned int len, int prot)
 204 {
 205         struct scatterlist *sg;
 206         unsigned int da = iova;
 207         unsigned int i, j;
 208         int ret;
 209
 210         if (!domain || !sgt)
 211                 return -EINVAL;
 212
 213         for_each_sg(sgt->sgl, sg, sgt->nents, i) {
 214                 u32 pa = sg_phys(sg) - sg->offset;
 215                 size_t bytes = sg->length + sg->offset;
 216
 217                 VERB("map[%d]: %08x %08x(%x)", i, iova, pa, bytes);
 218
 219                 ret = iommu_map(domain, da, pa, bytes, prot);
 220                 if (ret)
 221                         goto fail;
 222
 223                 da += bytes;
 224         }
 225
 226         return 0;
 227
 228 fail:
 229         da = iova;
 230
 231         for_each_sg(sgt->sgl, sg, i, j) {
 232                 size_t bytes = sg->length + sg->offset;
 233                 iommu_unmap(domain, da, bytes);
 234                 da += bytes;
 235         }
 236         return ret;
 237 }
 238
 239 static void unmap_range(struct iommu_domain *domain, unsigned int iova,
 240                 struct sg_table *sgt, unsigned int len)
 241 {
 242         struct scatterlist *sg;
 243         unsigned int da = iova;
 244         int i;
 245
 246         for_each_sg(sgt->sgl, sg, sgt->nents, i) {
 247                 size_t bytes = sg->length + sg->offset;
 248                 size_t unmapped;
 249
 250                 unmapped = iommu_unmap(domain, da, bytes);
 251                 if (unmapped < bytes)
 252                         break;
 253
 254                 VERB("unmap[%d]: %08x(%x)", i, iova, bytes);
 255
 256                 BUG_ON(!IS_ALIGNED(bytes, PAGE_SIZE));
 257
 258                 da += bytes;
 259         }
 260 }
 261
 262 /* should be called under struct_mutex.. although it can be called
 263  * from atomic context without struct_mutex to acquire an extra
 264  * iova ref if you know one is already held.
 265  *
 266  * That means when I do eventually need to add support for unpinning
 267  * the refcnt counter needs to be atomic_t.
 268  */
 269 int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id,
 270                 uint32_t *iova)
 271 {
 272         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 273         int ret = 0;
 274
 275         if (!msm_obj->domain[id].iova) {
 276                 struct msm_drm_private *priv = obj->dev->dev_private;
 277                 uint32_t offset = (uint32_t)mmap_offset(obj);
 278                 struct page **pages;
 279                 pages = get_pages(obj);
 280                 if (IS_ERR(pages))
 281                         return PTR_ERR(pages);
 282                 // XXX ideally we would not map buffers writable when not needed...
 283                 ret = map_range(priv->iommus[id], offset, msm_obj->sgt,
 284                                 obj->size, IOMMU_READ | IOMMU_WRITE);
 285                 msm_obj->domain[id].iova = offset;
 286         }
 287
 288         if (!ret)
 289                 *iova = msm_obj->domain[id].iova;
 290
 291         return ret;
 292 }
 293
 294 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova)
 295 {
 296         int ret;
 297         mutex_lock(&obj->dev->struct_mutex);
 298         ret = msm_gem_get_iova_locked(obj, id, iova);
 299         mutex_unlock(&obj->dev->struct_mutex);
 300         return ret;
 301 }
 302
 303 void msm_gem_put_iova(struct drm_gem_object *obj, int id)
 304 {
 305         // XXX TODO ..
 306         // NOTE: probably don't need a _locked() version.. we wouldn't
 307         // normally unmap here, but instead just mark that it could be
 308         // unmapped (if the iova refcnt drops to zero), but then later
 309         // if another _get_iova_locked() fails we can start unmapping
 310         // things that are no longer needed..
 311 }
 312
 313 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
 314                 struct drm_mode_create_dumb *args)
 315 {
 316         args->pitch = align_pitch(args->width, args->bpp);
 317         args->size  = PAGE_ALIGN(args->pitch * args->height);
 318         return msm_gem_new_handle(dev, file, args->size,
 319                         MSM_BO_SCANOUT | MSM_BO_WC, &args->handle);
 320 }
 321
 322 int msm_gem_dumb_destroy(struct drm_file *file, struct drm_device *dev,
 323                 uint32_t handle)
 324 {
 325         /* No special work needed, drop the reference and see what falls out */
 326         return drm_gem_handle_delete(file, handle);
 327 }
 328
 329 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 330                 uint32_t handle, uint64_t *offset)
 331 {
 332         struct drm_gem_object *obj;
 333         int ret = 0;
 334
 335         /* GEM does all our handle to object mapping */
 336         obj = drm_gem_object_lookup(dev, file, handle);
 337         if (obj == NULL) {
 338                 ret = -ENOENT;
 339                 goto fail;
 340         }
 341
 342         *offset = msm_gem_mmap_offset(obj);
 343
 344         drm_gem_object_unreference_unlocked(obj);
 345
 346 fail:
 347         return ret;
 348 }
 349
 350 void *msm_gem_vaddr_locked(struct drm_gem_object *obj)
 351 {
 352         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 353         WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
 354         if (!msm_obj->vaddr) {
 355                 struct page **pages = get_pages(obj);
 356                 if (IS_ERR(pages))
 357                         return ERR_CAST(pages);
 358                 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
 359                                 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
 360         }
 361         return msm_obj->vaddr;
 362 }
 363
 364 void *msm_gem_vaddr(struct drm_gem_object *obj)
 365 {
 366         void *ret;
 367         mutex_lock(&obj->dev->struct_mutex);
 368         ret = msm_gem_vaddr_locked(obj);
 369         mutex_unlock(&obj->dev->struct_mutex);
 370         return ret;
 371 }
 372
 373 int msm_gem_queue_inactive_work(struct drm_gem_object *obj,
 374                 struct work_struct *work)
 375 {
 376         struct drm_device *dev = obj->dev;
 377         struct msm_drm_private *priv = dev->dev_private;
 378         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 379         int ret = 0;
 380
 381         mutex_lock(&dev->struct_mutex);
 382         if (!list_empty(&work->entry)) {
 383                 ret = -EINVAL;
 384         } else if (is_active(msm_obj)) {
 385                 list_add_tail(&work->entry, &msm_obj->inactive_work);
 386         } else {
 387                 queue_work(priv->wq, work);
 388         }
 389         mutex_unlock(&dev->struct_mutex);
 390
 391         return ret;
 392 }
 393
 394 void msm_gem_move_to_active(struct drm_gem_object *obj,
 395                 struct msm_gpu *gpu, bool write, uint32_t fence)
 396 {
 397         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 398         msm_obj->gpu = gpu;
 399         if (write)
 400                 msm_obj->write_fence = fence;
 401         else
 402                 msm_obj->read_fence = fence;
 403         list_del_init(&msm_obj->mm_list);
 404         list_add_tail(&msm_obj->mm_list, &gpu->active_list);
 405 }
 406
 407 void msm_gem_move_to_inactive(struct drm_gem_object *obj)
 408 {
 409         struct drm_device *dev = obj->dev;
 410         struct msm_drm_private *priv = dev->dev_private;
 411         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 412
 413         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 414
 415         msm_obj->gpu = NULL;
 416         msm_obj->read_fence = 0;
 417         msm_obj->write_fence = 0;
 418         list_del_init(&msm_obj->mm_list);
 419         list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
 420
 421         while (!list_empty(&msm_obj->inactive_work)) {
 422                 struct work_struct *work;
 423
 424                 work = list_first_entry(&msm_obj->inactive_work,
 425                                 struct work_struct, entry);
 426
 427                 list_del_init(&work->entry);
 428                 queue_work(priv->wq, work);
 429         }
 430 }
 431
 432 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op,
 433                 struct timespec *timeout)
 434 {
 435         struct drm_device *dev = obj->dev;
 436         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 437         int ret = 0;
 438
 439         if (is_active(msm_obj)) {
 440                 uint32_t fence = 0;
 441
 442                 if (op & MSM_PREP_READ)
 443                         fence = msm_obj->write_fence;
 444                 if (op & MSM_PREP_WRITE)
 445                         fence = max(fence, msm_obj->read_fence);
 446                 if (op & MSM_PREP_NOSYNC)
 447                         timeout = NULL;
 448
 449                 ret = msm_wait_fence_interruptable(dev, fence, timeout);
 450         }
 451
 452         /* TODO cache maintenance */
 453
 454         return ret;
 455 }
 456
 457 int msm_gem_cpu_fini(struct drm_gem_object *obj)
 458 {
 459         /* TODO cache maintenance */
 460         return 0;
 461 }
 462
 463 #ifdef CONFIG_DEBUG_FS
 464 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
 465 {
 466         struct drm_device *dev = obj->dev;
 467         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 468         uint64_t off = drm_vma_node_start(&obj->vma_node);
 469
 470         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 471         seq_printf(m, "%08x: %c(r=%u,w=%u) %2d (%2d) %08llx %p %d\n",
 472                         msm_obj->flags, is_active(msm_obj) ? 'A' : 'I',
 473                         msm_obj->read_fence, msm_obj->write_fence,
 474                         obj->name, obj->refcount.refcount.counter,
 475                         off, msm_obj->vaddr, obj->size);
 476 }
 477
 478 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
 479 {
 480         struct msm_gem_object *msm_obj;
 481         int count = 0;
 482         size_t size = 0;
 483
 484         list_for_each_entry(msm_obj, list, mm_list) {
 485                 struct drm_gem_object *obj = &msm_obj->base;
 486                 seq_printf(m, "   ");
 487                 msm_gem_describe(obj, m);
 488                 count++;
 489                 size += obj->size;
 490         }
 491
 492         seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
 493 }
 494 #endif
 495
 496 void msm_gem_free_object(struct drm_gem_object *obj)
 497 {
 498         struct drm_device *dev = obj->dev;
 499         struct msm_gem_object *msm_obj = to_msm_bo(obj);
 500         int id;
 501
 502         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 503
 504         /* object should not be on active list: */
 505         WARN_ON(is_active(msm_obj));
 506
 507         list_del(&msm_obj->mm_list);
 508
 509         for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) {
 510                 if (msm_obj->domain[id].iova) {
 511                         struct msm_drm_private *priv = obj->dev->dev_private;
 512                         uint32_t offset = (uint32_t)mmap_offset(obj);
 513                         unmap_range(priv->iommus[id], offset,
 514                                         msm_obj->sgt, obj->size);
 515                 }
 516         }
 517
 518         drm_gem_free_mmap_offset(obj);
 519
 520         if (msm_obj->vaddr)
 521                 vunmap(msm_obj->vaddr);
 522
 523         put_pages(obj);
 524
 525         if (msm_obj->resv == &msm_obj->_resv)
 526                 reservation_object_fini(msm_obj->resv);
 527
 528         drm_gem_object_release(obj);
 529
 530         kfree(msm_obj);
 531 }
 532
 533 /* convenience method to construct a GEM buffer object, and userspace handle */
 534 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 535                 uint32_t size, uint32_t flags, uint32_t *handle)
 536 {
 537         struct drm_gem_object *obj;
 538         int ret;
 539
 540         ret = mutex_lock_interruptible(&dev->struct_mutex);
 541         if (ret)
 542                 return ret;
 543
 544         obj = msm_gem_new(dev, size, flags);
 545
 546         mutex_unlock(&dev->struct_mutex);
 547
 548         if (IS_ERR(obj))
 549                 return PTR_ERR(obj);
 550
 551         ret = drm_gem_handle_create(file, obj, handle);
 552
 553         /* drop reference from allocate - handle holds it now */
 554         drm_gem_object_unreference_unlocked(obj);
 555
 556         return ret;
 557 }
 558
 559 struct drm_gem_object *msm_gem_new(struct drm_device *dev,
 560                 uint32_t size, uint32_t flags)
 561 {
 562         struct msm_drm_private *priv = dev->dev_private;
 563         struct msm_gem_object *msm_obj;
 564         struct drm_gem_object *obj = NULL;
 565         int ret;
 566
 567         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 568
 569         size = PAGE_ALIGN(size);
 570
 571         switch (flags & MSM_BO_CACHE_MASK) {
 572         case MSM_BO_UNCACHED:
 573         case MSM_BO_CACHED:
 574         case MSM_BO_WC:
 575                 break;
 576         default:
 577                 dev_err(dev->dev, "invalid cache flag: %x\n",
 578                                 (flags & MSM_BO_CACHE_MASK));
 579                 ret = -EINVAL;
 580                 goto fail;
 581         }
 582
 583         msm_obj = kzalloc(sizeof(*msm_obj), GFP_KERNEL);
 584         if (!msm_obj) {
 585                 ret = -ENOMEM;
 586                 goto fail;
 587         }
 588
 589         obj = &msm_obj->base;
 590
 591         ret = drm_gem_object_init(dev, obj, size);
 592         if (ret)
 593                 goto fail;
 594
 595         msm_obj->flags = flags;
 596
 597         msm_obj->resv = &msm_obj->_resv;
 598         reservation_object_init(msm_obj->resv);
 599
 600         INIT_LIST_HEAD(&msm_obj->submit_entry);
 601         INIT_LIST_HEAD(&msm_obj->inactive_work);
 602         list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
 603
 604         return obj;
 605
 606 fail:
 607         if (obj)
 608                 drm_gem_object_unreference_unlocked(obj);
 609
 610         return ERR_PTR(ret);
 611 }