drivers/gpu/drm/omapdrm/omap_gem.c

   1 /*
   2  * drivers/gpu/drm/omapdrm/omap_gem.c
   3  *
   4  * Copyright (C) 2011 Texas Instruments
   5  * Author: Rob Clark <rob.clark@linaro.org>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms of the GNU General Public License version 2 as published by
   9  * the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along with
  17  * this program.  If not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #include <linux/seq_file.h>
  21 #include <linux/shmem_fs.h>
  22 #include <linux/spinlock.h>
  23 #include <linux/pfn_t.h>
  24
  25 #include <drm/drm_vma_manager.h>
  26
  27 #include "omap_drv.h"
  28 #include "omap_dmm_tiler.h"
  29
  30 /*
  31  * GEM buffer object implementation.
  32  */
  33
  34 /* note: we use upper 8 bits of flags for driver-internal flags: */
  35 #define OMAP_BO_MEM_DMA_API     0x01000000      /* memory allocated with the dma_alloc_* API */
  36 #define OMAP_BO_MEM_SHMEM       0x02000000      /* memory allocated through shmem backing */
  37 #define OMAP_BO_MEM_DMABUF      0x08000000      /* memory imported from a dmabuf */
  38
  39 struct omap_gem_object {
  40         struct drm_gem_object base;
  41
  42         struct list_head mm_list;
  43
  44         uint32_t flags;
  45
  46         /** width/height for tiled formats (rounded up to slot boundaries) */
  47         uint16_t width, height;
  48
  49         /** roll applied when mapping to DMM */
  50         uint32_t roll;
  51
  52         /**
  53          * paddr contains the buffer DMA address. It is valid for
  54          *
  55          * - buffers allocated through the DMA mapping API (with the
  56          *   OMAP_BO_MEM_DMA_API flag set)
  57          *
  58          * - buffers imported from dmabuf (with the OMAP_BO_MEM_DMABUF flag set)
  59          *   if they are physically contiguous (when sgt->orig_nents == 1)
  60          *
  61          * - buffers mapped through the TILER when paddr_cnt is not zero, in
  62          *   which case the DMA address points to the TILER aperture
  63          *
  64          * Physically contiguous buffers have their DMA address equal to the
  65          * physical address as we don't remap those buffers through the TILER.
  66          *
  67          * Buffers mapped to the TILER have their DMA address pointing to the
  68          * TILER aperture. As TILER mappings are refcounted (through paddr_cnt)
  69          * the DMA address must be accessed through omap_get_get_paddr() to
  70          * ensure that the mapping won't disappear unexpectedly. References must
  71          * be released with omap_gem_put_paddr().
  72          */
  73         dma_addr_t paddr;
  74
  75         /**
  76          * # of users of paddr
  77          */
  78         uint32_t paddr_cnt;
  79
  80         /**
  81          * If the buffer has been imported from a dmabuf the OMAP_DB_DMABUF flag
  82          * is set and the sgt field is valid.
  83          */
  84         struct sg_table *sgt;
  85
  86         /**
  87          * tiler block used when buffer is remapped in DMM/TILER.
  88          */
  89         struct tiler_block *block;
  90
  91         /**
  92          * Array of backing pages, if allocated.  Note that pages are never
  93          * allocated for buffers originally allocated from contiguous memory
  94          */
  95         struct page **pages;
  96
  97         /** addresses corresponding to pages in above array */
  98         dma_addr_t *addrs;
  99
 100         /**
 101          * Virtual address, if mapped.
 102          */
 103         void *vaddr;
 104
 105         /**
 106          * sync-object allocated on demand (if needed)
 107          *
 108          * Per-buffer sync-object for tracking pending and completed hw/dma
 109          * read and write operations.
 110          */
 111         struct {
 112                 uint32_t write_pending;
 113                 uint32_t write_complete;
 114                 uint32_t read_pending;
 115                 uint32_t read_complete;
 116         } *sync;
 117 };
 118
 119 #define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
 120
 121 /* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
 122  * not necessarily pinned in TILER all the time, and (b) when they are
 123  * they are not necessarily page aligned, we reserve one or more small
 124  * regions in each of the 2d containers to use as a user-GART where we
 125  * can create a second page-aligned mapping of parts of the buffer
 126  * being accessed from userspace.
 127  *
 128  * Note that we could optimize slightly when we know that multiple
 129  * tiler containers are backed by the same PAT.. but I'll leave that
 130  * for later..
 131  */
 132 #define NUM_USERGART_ENTRIES 2
 133 struct omap_drm_usergart_entry {
 134         struct tiler_block *block;      /* the reserved tiler block */
 135         dma_addr_t paddr;
 136         struct drm_gem_object *obj;     /* the current pinned obj */
 137         pgoff_t obj_pgoff;              /* page offset of obj currently
 138                                            mapped in */
 139 };
 140
 141 struct omap_drm_usergart {
 142         struct omap_drm_usergart_entry entry[NUM_USERGART_ENTRIES];
 143         int height;                             /* height in rows */
 144         int height_shift;               /* ilog2(height in rows) */
 145         int slot_shift;                 /* ilog2(width per slot) */
 146         int stride_pfn;                 /* stride in pages */
 147         int last;                               /* index of last used entry */
 148 };
 149
 150 /* -----------------------------------------------------------------------------
 151  * Helpers
 152  */
 153
 154 /** get mmap offset */
 155 static uint64_t mmap_offset(struct drm_gem_object *obj)
 156 {
 157         struct drm_device *dev = obj->dev;
 158         int ret;
 159         size_t size;
 160
 161         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 162
 163         /* Make it mmapable */
 164         size = omap_gem_mmap_size(obj);
 165         ret = drm_gem_create_mmap_offset_size(obj, size);
 166         if (ret) {
 167                 dev_err(dev->dev, "could not allocate mmap offset\n");
 168                 return 0;
 169         }
 170
 171         return drm_vma_node_offset_addr(&obj->vma_node);
 172 }
 173
 174 static bool is_contiguous(struct omap_gem_object *omap_obj)
 175 {
 176         if (omap_obj->flags & OMAP_BO_MEM_DMA_API)
 177                 return true;
 178
 179         if ((omap_obj->flags & OMAP_BO_MEM_DMABUF) && omap_obj->sgt->nents == 1)
 180                 return true;
 181
 182         return false;
 183 }
 184
 185 /* -----------------------------------------------------------------------------
 186  * Eviction
 187  */
 188
 189 static void evict_entry(struct drm_gem_object *obj,
 190                 enum tiler_fmt fmt, struct omap_drm_usergart_entry *entry)
 191 {
 192         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 193         struct omap_drm_private *priv = obj->dev->dev_private;
 194         int n = priv->usergart[fmt].height;
 195         size_t size = PAGE_SIZE * n;
 196         loff_t off = mmap_offset(obj) +
 197                         (entry->obj_pgoff << PAGE_SHIFT);
 198         const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
 199
 200         if (m > 1) {
 201                 int i;
 202                 /* if stride > than PAGE_SIZE then sparse mapping: */
 203                 for (i = n; i > 0; i--) {
 204                         unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 205                                             off, PAGE_SIZE, 1);
 206                         off += PAGE_SIZE * m;
 207                 }
 208         } else {
 209                 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 210                                     off, size, 1);
 211         }
 212
 213         entry->obj = NULL;
 214 }
 215
 216 /* Evict a buffer from usergart, if it is mapped there */
 217 static void evict(struct drm_gem_object *obj)
 218 {
 219         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 220         struct omap_drm_private *priv = obj->dev->dev_private;
 221
 222         if (omap_obj->flags & OMAP_BO_TILED) {
 223                 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 224                 int i;
 225
 226                 for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
 227                         struct omap_drm_usergart_entry *entry =
 228                                 &priv->usergart[fmt].entry[i];
 229
 230                         if (entry->obj == obj)
 231                                 evict_entry(obj, fmt, entry);
 232                 }
 233         }
 234 }
 235
 236 /* -----------------------------------------------------------------------------
 237  * Page Management
 238  */
 239
 240 /** ensure backing pages are allocated */
 241 static int omap_gem_attach_pages(struct drm_gem_object *obj)
 242 {
 243         struct drm_device *dev = obj->dev;
 244         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 245         struct page **pages;
 246         int npages = obj->size >> PAGE_SHIFT;
 247         int i, ret;
 248         dma_addr_t *addrs;
 249
 250         WARN_ON(omap_obj->pages);
 251
 252         pages = drm_gem_get_pages(obj);
 253         if (IS_ERR(pages)) {
 254                 dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
 255                 return PTR_ERR(pages);
 256         }
 257
 258         /* for non-cached buffers, ensure the new pages are clean because
 259          * DSS, GPU, etc. are not cache coherent:
 260          */
 261         if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
 262                 addrs = kmalloc(npages * sizeof(*addrs), GFP_KERNEL);
 263                 if (!addrs) {
 264                         ret = -ENOMEM;
 265                         goto free_pages;
 266                 }
 267
 268                 for (i = 0; i < npages; i++) {
 269                         addrs[i] = dma_map_page(dev->dev, pages[i],
 270                                         0, PAGE_SIZE, DMA_BIDIRECTIONAL);
 271
 272                         if (dma_mapping_error(dev->dev, addrs[i])) {
 273                                 dev_warn(dev->dev,
 274                                         "%s: failed to map page\n", __func__);
 275
 276                                 for (i = i - 1; i >= 0; --i) {
 277                                         dma_unmap_page(dev->dev, addrs[i],
 278                                                 PAGE_SIZE, DMA_BIDIRECTIONAL);
 279                                 }
 280
 281                                 ret = -ENOMEM;
 282                                 goto free_addrs;
 283                         }
 284                 }
 285         } else {
 286                 addrs = kzalloc(npages * sizeof(*addrs), GFP_KERNEL);
 287                 if (!addrs) {
 288                         ret = -ENOMEM;
 289                         goto free_pages;
 290                 }
 291         }
 292
 293         omap_obj->addrs = addrs;
 294         omap_obj->pages = pages;
 295
 296         return 0;
 297
 298 free_addrs:
 299         kfree(addrs);
 300 free_pages:
 301         drm_gem_put_pages(obj, pages, true, false);
 302
 303         return ret;
 304 }
 305
 306 /* acquire pages when needed (for example, for DMA where physically
 307  * contiguous buffer is not required
 308  */
 309 static int get_pages(struct drm_gem_object *obj, struct page ***pages)
 310 {
 311         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 312         int ret = 0;
 313
 314         if ((omap_obj->flags & OMAP_BO_MEM_SHMEM) && !omap_obj->pages) {
 315                 ret = omap_gem_attach_pages(obj);
 316                 if (ret) {
 317                         dev_err(obj->dev->dev, "could not attach pages\n");
 318                         return ret;
 319                 }
 320         }
 321
 322         /* TODO: even phys-contig.. we should have a list of pages? */
 323         *pages = omap_obj->pages;
 324
 325         return 0;
 326 }
 327
 328 /** release backing pages */
 329 static void omap_gem_detach_pages(struct drm_gem_object *obj)
 330 {
 331         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 332
 333         /* for non-cached buffers, ensure the new pages are clean because
 334          * DSS, GPU, etc. are not cache coherent:
 335          */
 336         if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
 337                 int i, npages = obj->size >> PAGE_SHIFT;
 338                 for (i = 0; i < npages; i++) {
 339                         dma_unmap_page(obj->dev->dev, omap_obj->addrs[i],
 340                                         PAGE_SIZE, DMA_BIDIRECTIONAL);
 341                 }
 342         }
 343
 344         kfree(omap_obj->addrs);
 345         omap_obj->addrs = NULL;
 346
 347         drm_gem_put_pages(obj, omap_obj->pages, true, false);
 348         omap_obj->pages = NULL;
 349 }
 350
 351 /* get buffer flags */
 352 uint32_t omap_gem_flags(struct drm_gem_object *obj)
 353 {
 354         return to_omap_bo(obj)->flags;
 355 }
 356
 357 uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj)
 358 {
 359         uint64_t offset;
 360         mutex_lock(&obj->dev->struct_mutex);
 361         offset = mmap_offset(obj);
 362         mutex_unlock(&obj->dev->struct_mutex);
 363         return offset;
 364 }
 365
 366 /** get mmap size */
 367 size_t omap_gem_mmap_size(struct drm_gem_object *obj)
 368 {
 369         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 370         size_t size = obj->size;
 371
 372         if (omap_obj->flags & OMAP_BO_TILED) {
 373                 /* for tiled buffers, the virtual size has stride rounded up
 374                  * to 4kb.. (to hide the fact that row n+1 might start 16kb or
 375                  * 32kb later!).  But we don't back the entire buffer with
 376                  * pages, only the valid picture part.. so need to adjust for
 377                  * this in the size used to mmap and generate mmap offset
 378                  */
 379                 size = tiler_vsize(gem2fmt(omap_obj->flags),
 380                                 omap_obj->width, omap_obj->height);
 381         }
 382
 383         return size;
 384 }
 385
 386 /* -----------------------------------------------------------------------------
 387  * Fault Handling
 388  */
 389
 390 /* Normal handling for the case of faulting in non-tiled buffers */
 391 static int fault_1d(struct drm_gem_object *obj,
 392                 struct vm_area_struct *vma, struct vm_fault *vmf)
 393 {
 394         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 395         unsigned long pfn;
 396         pgoff_t pgoff;
 397
 398         /* We don't use vmf->pgoff since that has the fake offset: */
 399         pgoff = ((unsigned long)vmf->virtual_address -
 400                         vma->vm_start) >> PAGE_SHIFT;
 401
 402         if (omap_obj->pages) {
 403                 omap_gem_cpu_sync(obj, pgoff);
 404                 pfn = page_to_pfn(omap_obj->pages[pgoff]);
 405         } else {
 406                 BUG_ON(!is_contiguous(omap_obj));
 407                 pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff;
 408         }
 409
 410         VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
 411                         pfn, pfn << PAGE_SHIFT);
 412
 413         return vm_insert_mixed(vma, (unsigned long)vmf->virtual_address,
 414                         __pfn_to_pfn_t(pfn, PFN_DEV));
 415 }
 416
 417 /* Special handling for the case of faulting in 2d tiled buffers */
 418 static int fault_2d(struct drm_gem_object *obj,
 419                 struct vm_area_struct *vma, struct vm_fault *vmf)
 420 {
 421         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 422         struct omap_drm_private *priv = obj->dev->dev_private;
 423         struct omap_drm_usergart_entry *entry;
 424         enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 425         struct page *pages[64];  /* XXX is this too much to have on stack? */
 426         unsigned long pfn;
 427         pgoff_t pgoff, base_pgoff;
 428         void __user *vaddr;
 429         int i, ret, slots;
 430
 431         /*
 432          * Note the height of the slot is also equal to the number of pages
 433          * that need to be mapped in to fill 4kb wide CPU page.  If the slot
 434          * height is 64, then 64 pages fill a 4kb wide by 64 row region.
 435          */
 436         const int n = priv->usergart[fmt].height;
 437         const int n_shift = priv->usergart[fmt].height_shift;
 438
 439         /*
 440          * If buffer width in bytes > PAGE_SIZE then the virtual stride is
 441          * rounded up to next multiple of PAGE_SIZE.. this need to be taken
 442          * into account in some of the math, so figure out virtual stride
 443          * in pages
 444          */
 445         const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
 446
 447         /* We don't use vmf->pgoff since that has the fake offset: */
 448         pgoff = ((unsigned long)vmf->virtual_address -
 449                         vma->vm_start) >> PAGE_SHIFT;
 450
 451         /*
 452          * Actual address we start mapping at is rounded down to previous slot
 453          * boundary in the y direction:
 454          */
 455         base_pgoff = round_down(pgoff, m << n_shift);
 456
 457         /* figure out buffer width in slots */
 458         slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
 459
 460         vaddr = vmf->virtual_address - ((pgoff - base_pgoff) << PAGE_SHIFT);
 461
 462         entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
 463
 464         /* evict previous buffer using this usergart entry, if any: */
 465         if (entry->obj)
 466                 evict_entry(entry->obj, fmt, entry);
 467
 468         entry->obj = obj;
 469         entry->obj_pgoff = base_pgoff;
 470
 471         /* now convert base_pgoff to phys offset from virt offset: */
 472         base_pgoff = (base_pgoff >> n_shift) * slots;
 473
 474         /* for wider-than 4k.. figure out which part of the slot-row we want: */
 475         if (m > 1) {
 476                 int off = pgoff % m;
 477                 entry->obj_pgoff += off;
 478                 base_pgoff /= m;
 479                 slots = min(slots - (off << n_shift), n);
 480                 base_pgoff += off << n_shift;
 481                 vaddr += off << PAGE_SHIFT;
 482         }
 483
 484         /*
 485          * Map in pages. Beyond the valid pixel part of the buffer, we set
 486          * pages[i] to NULL to get a dummy page mapped in.. if someone
 487          * reads/writes it they will get random/undefined content, but at
 488          * least it won't be corrupting whatever other random page used to
 489          * be mapped in, or other undefined behavior.
 490          */
 491         memcpy(pages, &omap_obj->pages[base_pgoff],
 492                         sizeof(struct page *) * slots);
 493         memset(pages + slots, 0,
 494                         sizeof(struct page *) * (n - slots));
 495
 496         ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
 497         if (ret) {
 498                 dev_err(obj->dev->dev, "failed to pin: %d\n", ret);
 499                 return ret;
 500         }
 501
 502         pfn = entry->paddr >> PAGE_SHIFT;
 503
 504         VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
 505                         pfn, pfn << PAGE_SHIFT);
 506
 507         for (i = n; i > 0; i--) {
 508                 vm_insert_mixed(vma, (unsigned long)vaddr,
 509                                 __pfn_to_pfn_t(pfn, PFN_DEV));
 510                 pfn += priv->usergart[fmt].stride_pfn;
 511                 vaddr += PAGE_SIZE * m;
 512         }
 513
 514         /* simple round-robin: */
 515         priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
 516                                  % NUM_USERGART_ENTRIES;
 517
 518         return 0;
 519 }
 520
 521 /**
 522  * omap_gem_fault               -       pagefault handler for GEM objects
 523  * @vma: the VMA of the GEM object
 524  * @vmf: fault detail
 525  *
 526  * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
 527  * does most of the work for us including the actual map/unmap calls
 528  * but we need to do the actual page work.
 529  *
 530  * The VMA was set up by GEM. In doing so it also ensured that the
 531  * vma->vm_private_data points to the GEM object that is backing this
 532  * mapping.
 533  */
 534 int omap_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 535 {
 536         struct drm_gem_object *obj = vma->vm_private_data;
 537         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 538         struct drm_device *dev = obj->dev;
 539         struct page **pages;
 540         int ret;
 541
 542         /* Make sure we don't parallel update on a fault, nor move or remove
 543          * something from beneath our feet
 544          */
 545         mutex_lock(&dev->struct_mutex);
 546
 547         /* if a shmem backed object, make sure we have pages attached now */
 548         ret = get_pages(obj, &pages);
 549         if (ret)
 550                 goto fail;
 551
 552         /* where should we do corresponding put_pages().. we are mapping
 553          * the original page, rather than thru a GART, so we can't rely
 554          * on eviction to trigger this.  But munmap() or all mappings should
 555          * probably trigger put_pages()?
 556          */
 557
 558         if (omap_obj->flags & OMAP_BO_TILED)
 559                 ret = fault_2d(obj, vma, vmf);
 560         else
 561                 ret = fault_1d(obj, vma, vmf);
 562
 563
 564 fail:
 565         mutex_unlock(&dev->struct_mutex);
 566         switch (ret) {
 567         case 0:
 568         case -ERESTARTSYS:
 569         case -EINTR:
 570         case -EBUSY:
 571                 /*
 572                  * EBUSY is ok: this just means that another thread
 573                  * already did the job.
 574                  */
 575                 return VM_FAULT_NOPAGE;
 576         case -ENOMEM:
 577                 return VM_FAULT_OOM;
 578         default:
 579                 return VM_FAULT_SIGBUS;
 580         }
 581 }
 582
 583 /** We override mainly to fix up some of the vm mapping flags.. */
 584 int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
 585 {
 586         int ret;
 587
 588         ret = drm_gem_mmap(filp, vma);
 589         if (ret) {
 590                 DBG("mmap failed: %d", ret);
 591                 return ret;
 592         }
 593
 594         return omap_gem_mmap_obj(vma->vm_private_data, vma);
 595 }
 596
 597 int omap_gem_mmap_obj(struct drm_gem_object *obj,
 598                 struct vm_area_struct *vma)
 599 {
 600         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 601
 602         vma->vm_flags &= ~VM_PFNMAP;
 603         vma->vm_flags |= VM_MIXEDMAP;
 604
 605         if (omap_obj->flags & OMAP_BO_WC) {
 606                 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
 607         } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
 608                 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
 609         } else {
 610                 /*
 611                  * We do have some private objects, at least for scanout buffers
 612                  * on hardware without DMM/TILER.  But these are allocated write-
 613                  * combine
 614                  */
 615                 if (WARN_ON(!obj->filp))
 616                         return -EINVAL;
 617
 618                 /*
 619                  * Shunt off cached objs to shmem file so they have their own
 620                  * address_space (so unmap_mapping_range does what we want,
 621                  * in particular in the case of mmap'd dmabufs)
 622                  */
 623                 fput(vma->vm_file);
 624                 vma->vm_pgoff = 0;
 625                 vma->vm_file  = get_file(obj->filp);
 626
 627                 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 628         }
 629
 630         return 0;
 631 }
 632
 633 /* -----------------------------------------------------------------------------
 634  * Dumb Buffers
 635  */
 636
 637 /**
 638  * omap_gem_dumb_create -       create a dumb buffer
 639  * @drm_file: our client file
 640  * @dev: our device
 641  * @args: the requested arguments copied from userspace
 642  *
 643  * Allocate a buffer suitable for use for a frame buffer of the
 644  * form described by user space. Give userspace a handle by which
 645  * to reference it.
 646  */
 647 int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
 648                 struct drm_mode_create_dumb *args)
 649 {
 650         union omap_gem_size gsize;
 651
 652         args->pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
 653
 654         args->size = PAGE_ALIGN(args->pitch * args->height);
 655
 656         gsize = (union omap_gem_size){
 657                 .bytes = args->size,
 658         };
 659
 660         return omap_gem_new_handle(dev, file, gsize,
 661                         OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
 662 }
 663
 664 /**
 665  * omap_gem_dumb_map    -       buffer mapping for dumb interface
 666  * @file: our drm client file
 667  * @dev: drm device
 668  * @handle: GEM handle to the object (from dumb_create)
 669  *
 670  * Do the necessary setup to allow the mapping of the frame buffer
 671  * into user memory. We don't have to do much here at the moment.
 672  */
 673 int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 674                 uint32_t handle, uint64_t *offset)
 675 {
 676         struct drm_gem_object *obj;
 677         int ret = 0;
 678
 679         /* GEM does all our handle to object mapping */
 680         obj = drm_gem_object_lookup(file, handle);
 681         if (obj == NULL) {
 682                 ret = -ENOENT;
 683                 goto fail;
 684         }
 685
 686         *offset = omap_gem_mmap_offset(obj);
 687
 688         drm_gem_object_unreference_unlocked(obj);
 689
 690 fail:
 691         return ret;
 692 }
 693
 694 #ifdef CONFIG_DRM_FBDEV_EMULATION
 695 /* Set scrolling position.  This allows us to implement fast scrolling
 696  * for console.
 697  *
 698  * Call only from non-atomic contexts.
 699  */
 700 int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll)
 701 {
 702         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 703         uint32_t npages = obj->size >> PAGE_SHIFT;
 704         int ret = 0;
 705
 706         if (roll > npages) {
 707                 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
 708                 return -EINVAL;
 709         }
 710
 711         omap_obj->roll = roll;
 712
 713         mutex_lock(&obj->dev->struct_mutex);
 714
 715         /* if we aren't mapped yet, we don't need to do anything */
 716         if (omap_obj->block) {
 717                 struct page **pages;
 718                 ret = get_pages(obj, &pages);
 719                 if (ret)
 720                         goto fail;
 721                 ret = tiler_pin(omap_obj->block, pages, npages, roll, true);
 722                 if (ret)
 723                         dev_err(obj->dev->dev, "could not repin: %d\n", ret);
 724         }
 725
 726 fail:
 727         mutex_unlock(&obj->dev->struct_mutex);
 728
 729         return ret;
 730 }
 731 #endif
 732
 733 /* -----------------------------------------------------------------------------
 734  * Memory Management & DMA Sync
 735  */
 736
 737 /**
 738  * shmem buffers that are mapped cached can simulate coherency via using
 739  * page faulting to keep track of dirty pages
 740  */
 741 static inline bool is_cached_coherent(struct drm_gem_object *obj)
 742 {
 743         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 744
 745         return (omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
 746                 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED);
 747 }
 748
 749 /* Sync the buffer for CPU access.. note pages should already be
 750  * attached, ie. omap_gem_get_pages()
 751  */
 752 void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff)
 753 {
 754         struct drm_device *dev = obj->dev;
 755         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 756
 757         if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) {
 758                 dma_unmap_page(dev->dev, omap_obj->addrs[pgoff],
 759                                 PAGE_SIZE, DMA_BIDIRECTIONAL);
 760                 omap_obj->addrs[pgoff] = 0;
 761         }
 762 }
 763
 764 /* sync the buffer for DMA access */
 765 void omap_gem_dma_sync(struct drm_gem_object *obj,
 766                 enum dma_data_direction dir)
 767 {
 768         struct drm_device *dev = obj->dev;
 769         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 770
 771         if (is_cached_coherent(obj)) {
 772                 int i, npages = obj->size >> PAGE_SHIFT;
 773                 struct page **pages = omap_obj->pages;
 774                 bool dirty = false;
 775
 776                 for (i = 0; i < npages; i++) {
 777                         if (!omap_obj->addrs[i]) {
 778                                 dma_addr_t addr;
 779
 780                                 addr = dma_map_page(dev->dev, pages[i], 0,
 781                                                 PAGE_SIZE, DMA_BIDIRECTIONAL);
 782
 783                                 if (dma_mapping_error(dev->dev, addr)) {
 784                                         dev_warn(dev->dev,
 785                                                 "%s: failed to map page\n",
 786                                                 __func__);
 787                                         break;
 788                                 }
 789
 790                                 dirty = true;
 791                                 omap_obj->addrs[i] = addr;
 792                         }
 793                 }
 794
 795                 if (dirty) {
 796                         unmap_mapping_range(obj->filp->f_mapping, 0,
 797                                         omap_gem_mmap_size(obj), 1);
 798                 }
 799         }
 800 }
 801
 802 /* Get physical address for DMA.. if 'remap' is true, and the buffer is not
 803  * already contiguous, remap it to pin in physically contiguous memory.. (ie.
 804  * map in TILER)
 805  */
 806 int omap_gem_get_paddr(struct drm_gem_object *obj,
 807                 dma_addr_t *paddr, bool remap)
 808 {
 809         struct omap_drm_private *priv = obj->dev->dev_private;
 810         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 811         int ret = 0;
 812
 813         mutex_lock(&obj->dev->struct_mutex);
 814
 815         if (!is_contiguous(omap_obj) && remap && priv->has_dmm) {
 816                 if (omap_obj->paddr_cnt == 0) {
 817                         struct page **pages;
 818                         uint32_t npages = obj->size >> PAGE_SHIFT;
 819                         enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 820                         struct tiler_block *block;
 821
 822                         BUG_ON(omap_obj->block);
 823
 824                         ret = get_pages(obj, &pages);
 825                         if (ret)
 826                                 goto fail;
 827
 828                         if (omap_obj->flags & OMAP_BO_TILED) {
 829                                 block = tiler_reserve_2d(fmt,
 830                                                 omap_obj->width,
 831                                                 omap_obj->height, 0);
 832                         } else {
 833                                 block = tiler_reserve_1d(obj->size);
 834                         }
 835
 836                         if (IS_ERR(block)) {
 837                                 ret = PTR_ERR(block);
 838                                 dev_err(obj->dev->dev,
 839                                         "could not remap: %d (%d)\n", ret, fmt);
 840                                 goto fail;
 841                         }
 842
 843                         /* TODO: enable async refill.. */
 844                         ret = tiler_pin(block, pages, npages,
 845                                         omap_obj->roll, true);
 846                         if (ret) {
 847                                 tiler_release(block);
 848                                 dev_err(obj->dev->dev,
 849                                                 "could not pin: %d\n", ret);
 850                                 goto fail;
 851                         }
 852
 853                         omap_obj->paddr = tiler_ssptr(block);
 854                         omap_obj->block = block;
 855
 856                         DBG("got paddr: %pad", &omap_obj->paddr);
 857                 }
 858
 859                 omap_obj->paddr_cnt++;
 860
 861                 *paddr = omap_obj->paddr;
 862         } else if (is_contiguous(omap_obj)) {
 863                 *paddr = omap_obj->paddr;
 864         } else {
 865                 ret = -EINVAL;
 866                 goto fail;
 867         }
 868
 869 fail:
 870         mutex_unlock(&obj->dev->struct_mutex);
 871
 872         return ret;
 873 }
 874
 875 /* Release physical address, when DMA is no longer being performed.. this
 876  * could potentially unpin and unmap buffers from TILER
 877  */
 878 void omap_gem_put_paddr(struct drm_gem_object *obj)
 879 {
 880         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 881         int ret;
 882
 883         mutex_lock(&obj->dev->struct_mutex);
 884         if (omap_obj->paddr_cnt > 0) {
 885                 omap_obj->paddr_cnt--;
 886                 if (omap_obj->paddr_cnt == 0) {
 887                         ret = tiler_unpin(omap_obj->block);
 888                         if (ret) {
 889                                 dev_err(obj->dev->dev,
 890                                         "could not unpin pages: %d\n", ret);
 891                         }
 892                         ret = tiler_release(omap_obj->block);
 893                         if (ret) {
 894                                 dev_err(obj->dev->dev,
 895                                         "could not release unmap: %d\n", ret);
 896                         }
 897                         omap_obj->paddr = 0;
 898                         omap_obj->block = NULL;
 899                 }
 900         }
 901
 902         mutex_unlock(&obj->dev->struct_mutex);
 903 }
 904
 905 /* Get rotated scanout address (only valid if already pinned), at the
 906  * specified orientation and x,y offset from top-left corner of buffer
 907  * (only valid for tiled 2d buffers)
 908  */
 909 int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient,
 910                 int x, int y, dma_addr_t *paddr)
 911 {
 912         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 913         int ret = -EINVAL;
 914
 915         mutex_lock(&obj->dev->struct_mutex);
 916         if ((omap_obj->paddr_cnt > 0) && omap_obj->block &&
 917                         (omap_obj->flags & OMAP_BO_TILED)) {
 918                 *paddr = tiler_tsptr(omap_obj->block, orient, x, y);
 919                 ret = 0;
 920         }
 921         mutex_unlock(&obj->dev->struct_mutex);
 922         return ret;
 923 }
 924
 925 /* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
 926 int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient)
 927 {
 928         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 929         int ret = -EINVAL;
 930         if (omap_obj->flags & OMAP_BO_TILED)
 931                 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
 932         return ret;
 933 }
 934
 935 /* if !remap, and we don't have pages backing, then fail, rather than
 936  * increasing the pin count (which we don't really do yet anyways,
 937  * because we don't support swapping pages back out).  And 'remap'
 938  * might not be quite the right name, but I wanted to keep it working
 939  * similarly to omap_gem_get_paddr().  Note though that mutex is not
 940  * aquired if !remap (because this can be called in atomic ctxt),
 941  * but probably omap_gem_get_paddr() should be changed to work in the
 942  * same way.  If !remap, a matching omap_gem_put_pages() call is not
 943  * required (and should not be made).
 944  */
 945 int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
 946                 bool remap)
 947 {
 948         int ret;
 949         if (!remap) {
 950                 struct omap_gem_object *omap_obj = to_omap_bo(obj);
 951                 if (!omap_obj->pages)
 952                         return -ENOMEM;
 953                 *pages = omap_obj->pages;
 954                 return 0;
 955         }
 956         mutex_lock(&obj->dev->struct_mutex);
 957         ret = get_pages(obj, pages);
 958         mutex_unlock(&obj->dev->struct_mutex);
 959         return ret;
 960 }
 961
 962 /* release pages when DMA no longer being performed */
 963 int omap_gem_put_pages(struct drm_gem_object *obj)
 964 {
 965         /* do something here if we dynamically attach/detach pages.. at
 966          * least they would no longer need to be pinned if everyone has
 967          * released the pages..
 968          */
 969         return 0;
 970 }
 971
 972 #ifdef CONFIG_DRM_FBDEV_EMULATION
 973 /* Get kernel virtual address for CPU access.. this more or less only
 974  * exists for omap_fbdev.  This should be called with struct_mutex
 975  * held.
 976  */
 977 void *omap_gem_vaddr(struct drm_gem_object *obj)
 978 {
 979         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 980         WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
 981         if (!omap_obj->vaddr) {
 982                 struct page **pages;
 983                 int ret = get_pages(obj, &pages);
 984                 if (ret)
 985                         return ERR_PTR(ret);
 986                 omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
 987                                 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
 988         }
 989         return omap_obj->vaddr;
 990 }
 991 #endif
 992
 993 /* -----------------------------------------------------------------------------
 994  * Power Management
 995  */
 996
 997 #ifdef CONFIG_PM
 998 /* re-pin objects in DMM in resume path: */
 999 int omap_gem_resume(struct device *dev)
1000 {
1001         struct drm_device *drm_dev = dev_get_drvdata(dev);
1002         struct omap_drm_private *priv = drm_dev->dev_private;
1003         struct omap_gem_object *omap_obj;
1004         int ret = 0;
1005
1006         list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
1007                 if (omap_obj->block) {
1008                         struct drm_gem_object *obj = &omap_obj->base;
1009                         uint32_t npages = obj->size >> PAGE_SHIFT;
1010                         WARN_ON(!omap_obj->pages);  /* this can't happen */
1011                         ret = tiler_pin(omap_obj->block,
1012                                         omap_obj->pages, npages,
1013                                         omap_obj->roll, true);
1014                         if (ret) {
1015                                 dev_err(dev, "could not repin: %d\n", ret);
1016                                 return ret;
1017                         }
1018                 }
1019         }
1020
1021         return 0;
1022 }
1023 #endif
1024
1025 /* -----------------------------------------------------------------------------
1026  * DebugFS
1027  */
1028
1029 #ifdef CONFIG_DEBUG_FS
1030 void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1031 {
1032         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1033         uint64_t off;
1034
1035         off = drm_vma_node_start(&obj->vma_node);
1036
1037         seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1038                         omap_obj->flags, obj->name, obj->refcount.refcount.counter,
1039                         off, &omap_obj->paddr, omap_obj->paddr_cnt,
1040                         omap_obj->vaddr, omap_obj->roll);
1041
1042         if (omap_obj->flags & OMAP_BO_TILED) {
1043                 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1044                 if (omap_obj->block) {
1045                         struct tcm_area *area = &omap_obj->block->area;
1046                         seq_printf(m, " (%dx%d, %dx%d)",
1047                                         area->p0.x, area->p0.y,
1048                                         area->p1.x, area->p1.y);
1049                 }
1050         } else {
1051                 seq_printf(m, " %d", obj->size);
1052         }
1053
1054         seq_printf(m, "\n");
1055 }
1056
1057 void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1058 {
1059         struct omap_gem_object *omap_obj;
1060         int count = 0;
1061         size_t size = 0;
1062
1063         list_for_each_entry(omap_obj, list, mm_list) {
1064                 struct drm_gem_object *obj = &omap_obj->base;
1065                 seq_printf(m, "   ");
1066                 omap_gem_describe(obj, m);
1067                 count++;
1068                 size += obj->size;
1069         }
1070
1071         seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1072 }
1073 #endif
1074
1075 /* -----------------------------------------------------------------------------
1076  * Buffer Synchronization
1077  */
1078
1079 static DEFINE_SPINLOCK(sync_lock);
1080
1081 struct omap_gem_sync_waiter {
1082         struct list_head list;
1083         struct omap_gem_object *omap_obj;
1084         enum omap_gem_op op;
1085         uint32_t read_target, write_target;
1086         /* notify called w/ sync_lock held */
1087         void (*notify)(void *arg);
1088         void *arg;
1089 };
1090
1091 /* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1092  * the read and/or write target count is achieved which can call a user
1093  * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1094  * cpu access), etc.
1095  */
1096 static LIST_HEAD(waiters);
1097
1098 static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1099 {
1100         struct omap_gem_object *omap_obj = waiter->omap_obj;
1101         if ((waiter->op & OMAP_GEM_READ) &&
1102                         (omap_obj->sync->write_complete < waiter->write_target))
1103                 return true;
1104         if ((waiter->op & OMAP_GEM_WRITE) &&
1105                         (omap_obj->sync->read_complete < waiter->read_target))
1106                 return true;
1107         return false;
1108 }
1109
1110 /* macro for sync debug.. */
1111 #define SYNCDBG 0
1112 #define SYNC(fmt, ...) do { if (SYNCDBG) \
1113                 printk(KERN_ERR "%s:%d: "fmt"\n", \
1114                                 __func__, __LINE__, ##__VA_ARGS__); \
1115         } while (0)
1116
1117
1118 static void sync_op_update(void)
1119 {
1120         struct omap_gem_sync_waiter *waiter, *n;
1121         list_for_each_entry_safe(waiter, n, &waiters, list) {
1122                 if (!is_waiting(waiter)) {
1123                         list_del(&waiter->list);
1124                         SYNC("notify: %p", waiter);
1125                         waiter->notify(waiter->arg);
1126                         kfree(waiter);
1127                 }
1128         }
1129 }
1130
1131 static inline int sync_op(struct drm_gem_object *obj,
1132                 enum omap_gem_op op, bool start)
1133 {
1134         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1135         int ret = 0;
1136
1137         spin_lock(&sync_lock);
1138
1139         if (!omap_obj->sync) {
1140                 omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1141                 if (!omap_obj->sync) {
1142                         ret = -ENOMEM;
1143                         goto unlock;
1144                 }
1145         }
1146
1147         if (start) {
1148                 if (op & OMAP_GEM_READ)
1149                         omap_obj->sync->read_pending++;
1150                 if (op & OMAP_GEM_WRITE)
1151                         omap_obj->sync->write_pending++;
1152         } else {
1153                 if (op & OMAP_GEM_READ)
1154                         omap_obj->sync->read_complete++;
1155                 if (op & OMAP_GEM_WRITE)
1156                         omap_obj->sync->write_complete++;
1157                 sync_op_update();
1158         }
1159
1160 unlock:
1161         spin_unlock(&sync_lock);
1162
1163         return ret;
1164 }
1165
1166 /* mark the start of read and/or write operation */
1167 int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1168 {
1169         return sync_op(obj, op, true);
1170 }
1171
1172 int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1173 {
1174         return sync_op(obj, op, false);
1175 }
1176
1177 static DECLARE_WAIT_QUEUE_HEAD(sync_event);
1178
1179 static void sync_notify(void *arg)
1180 {
1181         struct task_struct **waiter_task = arg;
1182         *waiter_task = NULL;
1183         wake_up_all(&sync_event);
1184 }
1185
1186 int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1187 {
1188         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1189         int ret = 0;
1190         if (omap_obj->sync) {
1191                 struct task_struct *waiter_task = current;
1192                 struct omap_gem_sync_waiter *waiter =
1193                                 kzalloc(sizeof(*waiter), GFP_KERNEL);
1194
1195                 if (!waiter)
1196                         return -ENOMEM;
1197
1198                 waiter->omap_obj = omap_obj;
1199                 waiter->op = op;
1200                 waiter->read_target = omap_obj->sync->read_pending;
1201                 waiter->write_target = omap_obj->sync->write_pending;
1202                 waiter->notify = sync_notify;
1203                 waiter->arg = &waiter_task;
1204
1205                 spin_lock(&sync_lock);
1206                 if (is_waiting(waiter)) {
1207                         SYNC("waited: %p", waiter);
1208                         list_add_tail(&waiter->list, &waiters);
1209                         spin_unlock(&sync_lock);
1210                         ret = wait_event_interruptible(sync_event,
1211                                         (waiter_task == NULL));
1212                         spin_lock(&sync_lock);
1213                         if (waiter_task) {
1214                                 SYNC("interrupted: %p", waiter);
1215                                 /* we were interrupted */
1216                                 list_del(&waiter->list);
1217                                 waiter_task = NULL;
1218                         } else {
1219                                 /* freed in sync_op_update() */
1220                                 waiter = NULL;
1221                         }
1222                 }
1223                 spin_unlock(&sync_lock);
1224                 kfree(waiter);
1225         }
1226         return ret;
1227 }
1228
1229 /* call fxn(arg), either synchronously or asynchronously if the op
1230  * is currently blocked..  fxn() can be called from any context
1231  *
1232  * (TODO for now fxn is called back from whichever context calls
1233  * omap_gem_op_finish().. but this could be better defined later
1234  * if needed)
1235  *
1236  * TODO more code in common w/ _sync()..
1237  */
1238 int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1239                 void (*fxn)(void *arg), void *arg)
1240 {
1241         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1242         if (omap_obj->sync) {
1243                 struct omap_gem_sync_waiter *waiter =
1244                                 kzalloc(sizeof(*waiter), GFP_ATOMIC);
1245
1246                 if (!waiter)
1247                         return -ENOMEM;
1248
1249                 waiter->omap_obj = omap_obj;
1250                 waiter->op = op;
1251                 waiter->read_target = omap_obj->sync->read_pending;
1252                 waiter->write_target = omap_obj->sync->write_pending;
1253                 waiter->notify = fxn;
1254                 waiter->arg = arg;
1255
1256                 spin_lock(&sync_lock);
1257                 if (is_waiting(waiter)) {
1258                         SYNC("waited: %p", waiter);
1259                         list_add_tail(&waiter->list, &waiters);
1260                         spin_unlock(&sync_lock);
1261                         return 0;
1262                 }
1263
1264                 spin_unlock(&sync_lock);
1265
1266                 kfree(waiter);
1267         }
1268
1269         /* no waiting.. */
1270         fxn(arg);
1271
1272         return 0;
1273 }
1274
1275 /* -----------------------------------------------------------------------------
1276  * Constructor & Destructor
1277  */
1278
1279 void omap_gem_free_object(struct drm_gem_object *obj)
1280 {
1281         struct drm_device *dev = obj->dev;
1282         struct omap_drm_private *priv = dev->dev_private;
1283         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1284
1285         evict(obj);
1286
1287         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1288
1289         spin_lock(&priv->list_lock);
1290         list_del(&omap_obj->mm_list);
1291         spin_unlock(&priv->list_lock);
1292
1293         /* this means the object is still pinned.. which really should
1294          * not happen.  I think..
1295          */
1296         WARN_ON(omap_obj->paddr_cnt > 0);
1297
1298         if (omap_obj->pages) {
1299                 if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1300                         kfree(omap_obj->pages);
1301                 else
1302                         omap_gem_detach_pages(obj);
1303         }
1304
1305         if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1306                 dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1307                             omap_obj->paddr);
1308         } else if (omap_obj->vaddr) {
1309                 vunmap(omap_obj->vaddr);
1310         } else if (obj->import_attach) {
1311                 drm_prime_gem_destroy(obj, omap_obj->sgt);
1312         }
1313
1314         kfree(omap_obj->sync);
1315
1316         drm_gem_object_release(obj);
1317
1318         kfree(omap_obj);
1319 }
1320
1321 /* GEM buffer object constructor */
1322 struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1323                 union omap_gem_size gsize, uint32_t flags)
1324 {
1325         struct omap_drm_private *priv = dev->dev_private;
1326         struct omap_gem_object *omap_obj;
1327         struct drm_gem_object *obj;
1328         struct address_space *mapping;
1329         size_t size;
1330         int ret;
1331
1332         /* Validate the flags and compute the memory and cache flags. */
1333         if (flags & OMAP_BO_TILED) {
1334                 if (!priv->usergart) {
1335                         dev_err(dev->dev, "Tiled buffers require DMM\n");
1336                         return NULL;
1337                 }
1338
1339                 /*
1340                  * Tiled buffers are always shmem paged backed. When they are
1341                  * scanned out, they are remapped into DMM/TILER.
1342                  */
1343                 flags &= ~OMAP_BO_SCANOUT;
1344                 flags |= OMAP_BO_MEM_SHMEM;
1345
1346                 /*
1347                  * Currently don't allow cached buffers. There is some caching
1348                  * stuff that needs to be handled better.
1349                  */
1350                 flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1351                 flags |= tiler_get_cpu_cache_flags();
1352         } else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1353                 /*
1354                  * OMAP_BO_SCANOUT hints that the buffer doesn't need to be
1355                  * tiled. However, to lower the pressure on memory allocation,
1356                  * use contiguous memory only if no TILER is available.
1357                  */
1358                 flags |= OMAP_BO_MEM_DMA_API;
1359         } else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1360                 /*
1361                  * All other buffers not backed by dma_buf are shmem-backed.
1362                  */
1363                 flags |= OMAP_BO_MEM_SHMEM;
1364         }
1365
1366         /* Allocate the initialize the OMAP GEM object. */
1367         omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1368         if (!omap_obj)
1369                 return NULL;
1370
1371         obj = &omap_obj->base;
1372         omap_obj->flags = flags;
1373
1374         if (flags & OMAP_BO_TILED) {
1375                 /*
1376                  * For tiled buffers align dimensions to slot boundaries and
1377                  * calculate size based on aligned dimensions.
1378                  */
1379                 tiler_align(gem2fmt(flags), &gsize.tiled.width,
1380                             &gsize.tiled.height);
1381
1382                 size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1383                                   gsize.tiled.height);
1384
1385                 omap_obj->width = gsize.tiled.width;
1386                 omap_obj->height = gsize.tiled.height;
1387         } else {
1388                 size = PAGE_ALIGN(gsize.bytes);
1389         }
1390
1391         /* Initialize the GEM object. */
1392         if (!(flags & OMAP_BO_MEM_SHMEM)) {
1393                 drm_gem_private_object_init(dev, obj, size);
1394         } else {
1395                 ret = drm_gem_object_init(dev, obj, size);
1396                 if (ret)
1397                         goto err_free;
1398
1399                 mapping = obj->filp->f_mapping;
1400                 mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1401         }
1402
1403         /* Allocate memory if needed. */
1404         if (flags & OMAP_BO_MEM_DMA_API) {
1405                 omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1406                                                &omap_obj->paddr,
1407                                                GFP_KERNEL);
1408                 if (!omap_obj->vaddr)
1409                         goto err_release;
1410         }
1411
1412         spin_lock(&priv->list_lock);
1413         list_add(&omap_obj->mm_list, &priv->obj_list);
1414         spin_unlock(&priv->list_lock);
1415
1416         return obj;
1417
1418 err_release:
1419         drm_gem_object_release(obj);
1420 err_free:
1421         kfree(omap_obj);
1422         return NULL;
1423 }
1424
1425 struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1426                                            struct sg_table *sgt)
1427 {
1428         struct omap_drm_private *priv = dev->dev_private;
1429         struct omap_gem_object *omap_obj;
1430         struct drm_gem_object *obj;
1431         union omap_gem_size gsize;
1432
1433         /* Without a DMM only physically contiguous buffers can be supported. */
1434         if (sgt->orig_nents != 1 && !priv->has_dmm)
1435                 return ERR_PTR(-EINVAL);
1436
1437         mutex_lock(&dev->struct_mutex);
1438
1439         gsize.bytes = PAGE_ALIGN(size);
1440         obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1441         if (!obj) {
1442                 obj = ERR_PTR(-ENOMEM);
1443                 goto done;
1444         }
1445
1446         omap_obj = to_omap_bo(obj);
1447         omap_obj->sgt = sgt;
1448
1449         if (sgt->orig_nents == 1) {
1450                 omap_obj->paddr = sg_dma_address(sgt->sgl);
1451         } else {
1452                 /* Create pages list from sgt */
1453                 struct sg_page_iter iter;
1454                 struct page **pages;
1455                 unsigned int npages;
1456                 unsigned int i = 0;
1457
1458                 npages = DIV_ROUND_UP(size, PAGE_SIZE);
1459                 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1460                 if (!pages) {
1461                         omap_gem_free_object(obj);
1462                         obj = ERR_PTR(-ENOMEM);
1463                         goto done;
1464                 }
1465
1466                 omap_obj->pages = pages;
1467
1468                 for_each_sg_page(sgt->sgl, &iter, sgt->orig_nents, 0) {
1469                         pages[i++] = sg_page_iter_page(&iter);
1470                         if (i > npages)
1471                                 break;
1472                 }
1473
1474                 if (WARN_ON(i != npages)) {
1475                         omap_gem_free_object(obj);
1476                         obj = ERR_PTR(-ENOMEM);
1477                         goto done;
1478                 }
1479         }
1480
1481 done:
1482         mutex_unlock(&dev->struct_mutex);
1483         return obj;
1484 }
1485
1486 /* convenience method to construct a GEM buffer object, and userspace handle */
1487 int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1488                 union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1489 {
1490         struct drm_gem_object *obj;
1491         int ret;
1492
1493         obj = omap_gem_new(dev, gsize, flags);
1494         if (!obj)
1495                 return -ENOMEM;
1496
1497         ret = drm_gem_handle_create(file, obj, handle);
1498         if (ret) {
1499                 omap_gem_free_object(obj);
1500                 return ret;
1501         }
1502
1503         /* drop reference from allocate - handle holds it now */
1504         drm_gem_object_unreference_unlocked(obj);
1505
1506         return 0;
1507 }
1508
1509 /* -----------------------------------------------------------------------------
1510  * Init & Cleanup
1511  */
1512
1513 /* If DMM is used, we need to set some stuff up.. */
1514 void omap_gem_init(struct drm_device *dev)
1515 {
1516         struct omap_drm_private *priv = dev->dev_private;
1517         struct omap_drm_usergart *usergart;
1518         const enum tiler_fmt fmts[] = {
1519                         TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1520         };
1521         int i, j;
1522
1523         if (!dmm_is_available()) {
1524                 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1525                 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1526                 return;
1527         }
1528
1529         usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1530         if (!usergart)
1531                 return;
1532
1533         /* reserve 4k aligned/wide regions for userspace mappings: */
1534         for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1535                 uint16_t h = 1, w = PAGE_SIZE >> i;
1536                 tiler_align(fmts[i], &w, &h);
1537                 /* note: since each region is 1 4kb page wide, and minimum
1538                  * number of rows, the height ends up being the same as the
1539                  * # of pages in the region
1540                  */
1541                 usergart[i].height = h;
1542                 usergart[i].height_shift = ilog2(h);
1543                 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1544                 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1545                 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1546                         struct omap_drm_usergart_entry *entry;
1547                         struct tiler_block *block;
1548
1549                         entry = &usergart[i].entry[j];
1550                         block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1551                         if (IS_ERR(block)) {
1552                                 dev_err(dev->dev,
1553                                                 "reserve failed: %d, %d, %ld\n",
1554                                                 i, j, PTR_ERR(block));
1555                                 return;
1556                         }
1557                         entry->paddr = tiler_ssptr(block);
1558                         entry->block = block;
1559
1560                         DBG("%d:%d: %dx%d: paddr=%pad stride=%d", i, j, w, h,
1561                                         &entry->paddr,
1562                                         usergart[i].stride_pfn << PAGE_SHIFT);
1563                 }
1564         }
1565
1566         priv->usergart = usergart;
1567         priv->has_dmm = true;
1568 }
1569
1570 void omap_gem_deinit(struct drm_device *dev)
1571 {
1572         struct omap_drm_private *priv = dev->dev_private;
1573
1574         /* I believe we can rely on there being no more outstanding GEM
1575          * objects which could depend on usergart/dmm at this point.
1576          */
1577         kfree(priv->usergart);
1578 }