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                         if (omap_obj->addrs[i])
 340                                 dma_unmap_page(obj->dev->dev,
 341                                                omap_obj->addrs[i],
 342                                                PAGE_SIZE, DMA_BIDIRECTIONAL);
 343                 }
 344         }
 345
 346         kfree(omap_obj->addrs);
 347         omap_obj->addrs = NULL;
 348
 349         drm_gem_put_pages(obj, omap_obj->pages, true, false);
 350         omap_obj->pages = NULL;
 351 }
 352
 353 /* get buffer flags */
 354 uint32_t omap_gem_flags(struct drm_gem_object *obj)
 355 {
 356         return to_omap_bo(obj)->flags;
 357 }
 358
 359 uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj)
 360 {
 361         uint64_t offset;
 362         mutex_lock(&obj->dev->struct_mutex);
 363         offset = mmap_offset(obj);
 364         mutex_unlock(&obj->dev->struct_mutex);
 365         return offset;
 366 }
 367
 368 /** get mmap size */
 369 size_t omap_gem_mmap_size(struct drm_gem_object *obj)
 370 {
 371         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 372         size_t size = obj->size;
 373
 374         if (omap_obj->flags & OMAP_BO_TILED) {
 375                 /* for tiled buffers, the virtual size has stride rounded up
 376                  * to 4kb.. (to hide the fact that row n+1 might start 16kb or
 377                  * 32kb later!).  But we don't back the entire buffer with
 378                  * pages, only the valid picture part.. so need to adjust for
 379                  * this in the size used to mmap and generate mmap offset
 380                  */
 381                 size = tiler_vsize(gem2fmt(omap_obj->flags),
 382                                 omap_obj->width, omap_obj->height);
 383         }
 384
 385         return size;
 386 }
 387
 388 /* -----------------------------------------------------------------------------
 389  * Fault Handling
 390  */
 391
 392 /* Normal handling for the case of faulting in non-tiled buffers */
 393 static int fault_1d(struct drm_gem_object *obj,
 394                 struct vm_area_struct *vma, struct vm_fault *vmf)
 395 {
 396         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 397         unsigned long pfn;
 398         pgoff_t pgoff;
 399
 400         /* We don't use vmf->pgoff since that has the fake offset: */
 401         pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
 402
 403         if (omap_obj->pages) {
 404                 omap_gem_cpu_sync(obj, pgoff);
 405                 pfn = page_to_pfn(omap_obj->pages[pgoff]);
 406         } else {
 407                 BUG_ON(!is_contiguous(omap_obj));
 408                 pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff;
 409         }
 410
 411         VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
 412                         pfn, pfn << PAGE_SHIFT);
 413
 414         return vm_insert_mixed(vma, vmf->address, __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         unsigned long 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 = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
 449
 450         /*
 451          * Actual address we start mapping at is rounded down to previous slot
 452          * boundary in the y direction:
 453          */
 454         base_pgoff = round_down(pgoff, m << n_shift);
 455
 456         /* figure out buffer width in slots */
 457         slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
 458
 459         vaddr = vmf->address - ((pgoff - base_pgoff) << PAGE_SHIFT);
 460
 461         entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
 462
 463         /* evict previous buffer using this usergart entry, if any: */
 464         if (entry->obj)
 465                 evict_entry(entry->obj, fmt, entry);
 466
 467         entry->obj = obj;
 468         entry->obj_pgoff = base_pgoff;
 469
 470         /* now convert base_pgoff to phys offset from virt offset: */
 471         base_pgoff = (base_pgoff >> n_shift) * slots;
 472
 473         /* for wider-than 4k.. figure out which part of the slot-row we want: */
 474         if (m > 1) {
 475                 int off = pgoff % m;
 476                 entry->obj_pgoff += off;
 477                 base_pgoff /= m;
 478                 slots = min(slots - (off << n_shift), n);
 479                 base_pgoff += off << n_shift;
 480                 vaddr += off << PAGE_SHIFT;
 481         }
 482
 483         /*
 484          * Map in pages. Beyond the valid pixel part of the buffer, we set
 485          * pages[i] to NULL to get a dummy page mapped in.. if someone
 486          * reads/writes it they will get random/undefined content, but at
 487          * least it won't be corrupting whatever other random page used to
 488          * be mapped in, or other undefined behavior.
 489          */
 490         memcpy(pages, &omap_obj->pages[base_pgoff],
 491                         sizeof(struct page *) * slots);
 492         memset(pages + slots, 0,
 493                         sizeof(struct page *) * (n - slots));
 494
 495         ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
 496         if (ret) {
 497                 dev_err(obj->dev->dev, "failed to pin: %d\n", ret);
 498                 return ret;
 499         }
 500
 501         pfn = entry->paddr >> PAGE_SHIFT;
 502
 503         VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
 504                         pfn, pfn << PAGE_SHIFT);
 505
 506         for (i = n; i > 0; i--) {
 507                 vm_insert_mixed(vma, vaddr, __pfn_to_pfn_t(pfn, PFN_DEV));
 508                 pfn += priv->usergart[fmt].stride_pfn;
 509                 vaddr += PAGE_SIZE * m;
 510         }
 511
 512         /* simple round-robin: */
 513         priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
 514                                  % NUM_USERGART_ENTRIES;
 515
 516         return 0;
 517 }
 518
 519 /**
 520  * omap_gem_fault               -       pagefault handler for GEM objects
 521  * @vmf: fault detail
 522  *
 523  * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
 524  * does most of the work for us including the actual map/unmap calls
 525  * but we need to do the actual page work.
 526  *
 527  * The VMA was set up by GEM. In doing so it also ensured that the
 528  * vma->vm_private_data points to the GEM object that is backing this
 529  * mapping.
 530  */
 531 int omap_gem_fault(struct vm_fault *vmf)
 532 {
 533         struct vm_area_struct *vma = vmf->vma;
 534         struct drm_gem_object *obj = vma->vm_private_data;
 535         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 536         struct drm_device *dev = obj->dev;
 537         struct page **pages;
 538         int ret;
 539
 540         /* Make sure we don't parallel update on a fault, nor move or remove
 541          * something from beneath our feet
 542          */
 543         mutex_lock(&dev->struct_mutex);
 544
 545         /* if a shmem backed object, make sure we have pages attached now */
 546         ret = get_pages(obj, &pages);
 547         if (ret)
 548                 goto fail;
 549
 550         /* where should we do corresponding put_pages().. we are mapping
 551          * the original page, rather than thru a GART, so we can't rely
 552          * on eviction to trigger this.  But munmap() or all mappings should
 553          * probably trigger put_pages()?
 554          */
 555
 556         if (omap_obj->flags & OMAP_BO_TILED)
 557                 ret = fault_2d(obj, vma, vmf);
 558         else
 559                 ret = fault_1d(obj, vma, vmf);
 560
 561
 562 fail:
 563         mutex_unlock(&dev->struct_mutex);
 564         switch (ret) {
 565         case 0:
 566         case -ERESTARTSYS:
 567         case -EINTR:
 568         case -EBUSY:
 569                 /*
 570                  * EBUSY is ok: this just means that another thread
 571                  * already did the job.
 572                  */
 573                 return VM_FAULT_NOPAGE;
 574         case -ENOMEM:
 575                 return VM_FAULT_OOM;
 576         default:
 577                 return VM_FAULT_SIGBUS;
 578         }
 579 }
 580
 581 /** We override mainly to fix up some of the vm mapping flags.. */
 582 int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
 583 {
 584         int ret;
 585
 586         ret = drm_gem_mmap(filp, vma);
 587         if (ret) {
 588                 DBG("mmap failed: %d", ret);
 589                 return ret;
 590         }
 591
 592         return omap_gem_mmap_obj(vma->vm_private_data, vma);
 593 }
 594
 595 int omap_gem_mmap_obj(struct drm_gem_object *obj,
 596                 struct vm_area_struct *vma)
 597 {
 598         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 599
 600         vma->vm_flags &= ~VM_PFNMAP;
 601         vma->vm_flags |= VM_MIXEDMAP;
 602
 603         if (omap_obj->flags & OMAP_BO_WC) {
 604                 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
 605         } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
 606                 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
 607         } else {
 608                 /*
 609                  * We do have some private objects, at least for scanout buffers
 610                  * on hardware without DMM/TILER.  But these are allocated write-
 611                  * combine
 612                  */
 613                 if (WARN_ON(!obj->filp))
 614                         return -EINVAL;
 615
 616                 /*
 617                  * Shunt off cached objs to shmem file so they have their own
 618                  * address_space (so unmap_mapping_range does what we want,
 619                  * in particular in the case of mmap'd dmabufs)
 620                  */
 621                 fput(vma->vm_file);
 622                 vma->vm_pgoff = 0;
 623                 vma->vm_file  = get_file(obj->filp);
 624
 625                 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 626         }
 627
 628         return 0;
 629 }
 630
 631 /* -----------------------------------------------------------------------------
 632  * Dumb Buffers
 633  */
 634
 635 /**
 636  * omap_gem_dumb_create -       create a dumb buffer
 637  * @drm_file: our client file
 638  * @dev: our device
 639  * @args: the requested arguments copied from userspace
 640  *
 641  * Allocate a buffer suitable for use for a frame buffer of the
 642  * form described by user space. Give userspace a handle by which
 643  * to reference it.
 644  */
 645 int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
 646                 struct drm_mode_create_dumb *args)
 647 {
 648         union omap_gem_size gsize;
 649
 650         args->pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
 651
 652         args->size = PAGE_ALIGN(args->pitch * args->height);
 653
 654         gsize = (union omap_gem_size){
 655                 .bytes = args->size,
 656         };
 657
 658         return omap_gem_new_handle(dev, file, gsize,
 659                         OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
 660 }
 661
 662 /**
 663  * omap_gem_dumb_map    -       buffer mapping for dumb interface
 664  * @file: our drm client file
 665  * @dev: drm device
 666  * @handle: GEM handle to the object (from dumb_create)
 667  *
 668  * Do the necessary setup to allow the mapping of the frame buffer
 669  * into user memory. We don't have to do much here at the moment.
 670  */
 671 int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 672                 uint32_t handle, uint64_t *offset)
 673 {
 674         struct drm_gem_object *obj;
 675         int ret = 0;
 676
 677         /* GEM does all our handle to object mapping */
 678         obj = drm_gem_object_lookup(file, handle);
 679         if (obj == NULL) {
 680                 ret = -ENOENT;
 681                 goto fail;
 682         }
 683
 684         *offset = omap_gem_mmap_offset(obj);
 685
 686         drm_gem_object_unreference_unlocked(obj);
 687
 688 fail:
 689         return ret;
 690 }
 691
 692 #ifdef CONFIG_DRM_FBDEV_EMULATION
 693 /* Set scrolling position.  This allows us to implement fast scrolling
 694  * for console.
 695  *
 696  * Call only from non-atomic contexts.
 697  */
 698 int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll)
 699 {
 700         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 701         uint32_t npages = obj->size >> PAGE_SHIFT;
 702         int ret = 0;
 703
 704         if (roll > npages) {
 705                 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
 706                 return -EINVAL;
 707         }
 708
 709         omap_obj->roll = roll;
 710
 711         mutex_lock(&obj->dev->struct_mutex);
 712
 713         /* if we aren't mapped yet, we don't need to do anything */
 714         if (omap_obj->block) {
 715                 struct page **pages;
 716                 ret = get_pages(obj, &pages);
 717                 if (ret)
 718                         goto fail;
 719                 ret = tiler_pin(omap_obj->block, pages, npages, roll, true);
 720                 if (ret)
 721                         dev_err(obj->dev->dev, "could not repin: %d\n", ret);
 722         }
 723
 724 fail:
 725         mutex_unlock(&obj->dev->struct_mutex);
 726
 727         return ret;
 728 }
 729 #endif
 730
 731 /* -----------------------------------------------------------------------------
 732  * Memory Management & DMA Sync
 733  */
 734
 735 /**
 736  * shmem buffers that are mapped cached can simulate coherency via using
 737  * page faulting to keep track of dirty pages
 738  */
 739 static inline bool is_cached_coherent(struct drm_gem_object *obj)
 740 {
 741         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 742
 743         return (omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
 744                 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED);
 745 }
 746
 747 /* Sync the buffer for CPU access.. note pages should already be
 748  * attached, ie. omap_gem_get_pages()
 749  */
 750 void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff)
 751 {
 752         struct drm_device *dev = obj->dev;
 753         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 754
 755         if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) {
 756                 dma_unmap_page(dev->dev, omap_obj->addrs[pgoff],
 757                                 PAGE_SIZE, DMA_BIDIRECTIONAL);
 758                 omap_obj->addrs[pgoff] = 0;
 759         }
 760 }
 761
 762 /* sync the buffer for DMA access */
 763 void omap_gem_dma_sync(struct drm_gem_object *obj,
 764                 enum dma_data_direction dir)
 765 {
 766         struct drm_device *dev = obj->dev;
 767         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 768
 769         if (is_cached_coherent(obj)) {
 770                 int i, npages = obj->size >> PAGE_SHIFT;
 771                 struct page **pages = omap_obj->pages;
 772                 bool dirty = false;
 773
 774                 for (i = 0; i < npages; i++) {
 775                         if (!omap_obj->addrs[i]) {
 776                                 dma_addr_t addr;
 777
 778                                 addr = dma_map_page(dev->dev, pages[i], 0,
 779                                                 PAGE_SIZE, DMA_BIDIRECTIONAL);
 780
 781                                 if (dma_mapping_error(dev->dev, addr)) {
 782                                         dev_warn(dev->dev,
 783                                                 "%s: failed to map page\n",
 784                                                 __func__);
 785                                         break;
 786                                 }
 787
 788                                 dirty = true;
 789                                 omap_obj->addrs[i] = addr;
 790                         }
 791                 }
 792
 793                 if (dirty) {
 794                         unmap_mapping_range(obj->filp->f_mapping, 0,
 795                                         omap_gem_mmap_size(obj), 1);
 796                 }
 797         }
 798 }
 799
 800 /* Get physical address for DMA.. if 'remap' is true, and the buffer is not
 801  * already contiguous, remap it to pin in physically contiguous memory.. (ie.
 802  * map in TILER)
 803  */
 804 int omap_gem_get_paddr(struct drm_gem_object *obj,
 805                 dma_addr_t *paddr, bool remap)
 806 {
 807         struct omap_drm_private *priv = obj->dev->dev_private;
 808         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 809         int ret = 0;
 810
 811         mutex_lock(&obj->dev->struct_mutex);
 812
 813         if (!is_contiguous(omap_obj) && remap && priv->has_dmm) {
 814                 if (omap_obj->paddr_cnt == 0) {
 815                         struct page **pages;
 816                         uint32_t npages = obj->size >> PAGE_SHIFT;
 817                         enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 818                         struct tiler_block *block;
 819
 820                         BUG_ON(omap_obj->block);
 821
 822                         ret = get_pages(obj, &pages);
 823                         if (ret)
 824                                 goto fail;
 825
 826                         if (omap_obj->flags & OMAP_BO_TILED) {
 827                                 block = tiler_reserve_2d(fmt,
 828                                                 omap_obj->width,
 829                                                 omap_obj->height, 0);
 830                         } else {
 831                                 block = tiler_reserve_1d(obj->size);
 832                         }
 833
 834                         if (IS_ERR(block)) {
 835                                 ret = PTR_ERR(block);
 836                                 dev_err(obj->dev->dev,
 837                                         "could not remap: %d (%d)\n", ret, fmt);
 838                                 goto fail;
 839                         }
 840
 841                         /* TODO: enable async refill.. */
 842                         ret = tiler_pin(block, pages, npages,
 843                                         omap_obj->roll, true);
 844                         if (ret) {
 845                                 tiler_release(block);
 846                                 dev_err(obj->dev->dev,
 847                                                 "could not pin: %d\n", ret);
 848                                 goto fail;
 849                         }
 850
 851                         omap_obj->paddr = tiler_ssptr(block);
 852                         omap_obj->block = block;
 853
 854                         DBG("got paddr: %pad", &omap_obj->paddr);
 855                 }
 856
 857                 omap_obj->paddr_cnt++;
 858
 859                 *paddr = omap_obj->paddr;
 860         } else if (is_contiguous(omap_obj)) {
 861                 *paddr = omap_obj->paddr;
 862         } else {
 863                 ret = -EINVAL;
 864                 goto fail;
 865         }
 866
 867 fail:
 868         mutex_unlock(&obj->dev->struct_mutex);
 869
 870         return ret;
 871 }
 872
 873 /* Release physical address, when DMA is no longer being performed.. this
 874  * could potentially unpin and unmap buffers from TILER
 875  */
 876 void omap_gem_put_paddr(struct drm_gem_object *obj)
 877 {
 878         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 879         int ret;
 880
 881         mutex_lock(&obj->dev->struct_mutex);
 882         if (omap_obj->paddr_cnt > 0) {
 883                 omap_obj->paddr_cnt--;
 884                 if (omap_obj->paddr_cnt == 0) {
 885                         ret = tiler_unpin(omap_obj->block);
 886                         if (ret) {
 887                                 dev_err(obj->dev->dev,
 888                                         "could not unpin pages: %d\n", ret);
 889                         }
 890                         ret = tiler_release(omap_obj->block);
 891                         if (ret) {
 892                                 dev_err(obj->dev->dev,
 893                                         "could not release unmap: %d\n", ret);
 894                         }
 895                         omap_obj->paddr = 0;
 896                         omap_obj->block = NULL;
 897                 }
 898         }
 899
 900         mutex_unlock(&obj->dev->struct_mutex);
 901 }
 902
 903 /* Get rotated scanout address (only valid if already pinned), at the
 904  * specified orientation and x,y offset from top-left corner of buffer
 905  * (only valid for tiled 2d buffers)
 906  */
 907 int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient,
 908                 int x, int y, dma_addr_t *paddr)
 909 {
 910         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 911         int ret = -EINVAL;
 912
 913         mutex_lock(&obj->dev->struct_mutex);
 914         if ((omap_obj->paddr_cnt > 0) && omap_obj->block &&
 915                         (omap_obj->flags & OMAP_BO_TILED)) {
 916                 *paddr = tiler_tsptr(omap_obj->block, orient, x, y);
 917                 ret = 0;
 918         }
 919         mutex_unlock(&obj->dev->struct_mutex);
 920         return ret;
 921 }
 922
 923 /* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
 924 int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient)
 925 {
 926         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 927         int ret = -EINVAL;
 928         if (omap_obj->flags & OMAP_BO_TILED)
 929                 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
 930         return ret;
 931 }
 932
 933 /* if !remap, and we don't have pages backing, then fail, rather than
 934  * increasing the pin count (which we don't really do yet anyways,
 935  * because we don't support swapping pages back out).  And 'remap'
 936  * might not be quite the right name, but I wanted to keep it working
 937  * similarly to omap_gem_get_paddr().  Note though that mutex is not
 938  * aquired if !remap (because this can be called in atomic ctxt),
 939  * but probably omap_gem_get_paddr() should be changed to work in the
 940  * same way.  If !remap, a matching omap_gem_put_pages() call is not
 941  * required (and should not be made).
 942  */
 943 int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
 944                 bool remap)
 945 {
 946         int ret;
 947         if (!remap) {
 948                 struct omap_gem_object *omap_obj = to_omap_bo(obj);
 949                 if (!omap_obj->pages)
 950                         return -ENOMEM;
 951                 *pages = omap_obj->pages;
 952                 return 0;
 953         }
 954         mutex_lock(&obj->dev->struct_mutex);
 955         ret = get_pages(obj, pages);
 956         mutex_unlock(&obj->dev->struct_mutex);
 957         return ret;
 958 }
 959
 960 /* release pages when DMA no longer being performed */
 961 int omap_gem_put_pages(struct drm_gem_object *obj)
 962 {
 963         /* do something here if we dynamically attach/detach pages.. at
 964          * least they would no longer need to be pinned if everyone has
 965          * released the pages..
 966          */
 967         return 0;
 968 }
 969
 970 #ifdef CONFIG_DRM_FBDEV_EMULATION
 971 /* Get kernel virtual address for CPU access.. this more or less only
 972  * exists for omap_fbdev.  This should be called with struct_mutex
 973  * held.
 974  */
 975 void *omap_gem_vaddr(struct drm_gem_object *obj)
 976 {
 977         struct omap_gem_object *omap_obj = to_omap_bo(obj);
 978         WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
 979         if (!omap_obj->vaddr) {
 980                 struct page **pages;
 981                 int ret = get_pages(obj, &pages);
 982                 if (ret)
 983                         return ERR_PTR(ret);
 984                 omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
 985                                 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
 986         }
 987         return omap_obj->vaddr;
 988 }
 989 #endif
 990
 991 /* -----------------------------------------------------------------------------
 992  * Power Management
 993  */
 994
 995 #ifdef CONFIG_PM
 996 /* re-pin objects in DMM in resume path: */
 997 int omap_gem_resume(struct device *dev)
 998 {
 999         struct drm_device *drm_dev = dev_get_drvdata(dev);
1000         struct omap_drm_private *priv = drm_dev->dev_private;
1001         struct omap_gem_object *omap_obj;
1002         int ret = 0;
1003
1004         list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
1005                 if (omap_obj->block) {
1006                         struct drm_gem_object *obj = &omap_obj->base;
1007                         uint32_t npages = obj->size >> PAGE_SHIFT;
1008                         WARN_ON(!omap_obj->pages);  /* this can't happen */
1009                         ret = tiler_pin(omap_obj->block,
1010                                         omap_obj->pages, npages,
1011                                         omap_obj->roll, true);
1012                         if (ret) {
1013                                 dev_err(dev, "could not repin: %d\n", ret);
1014                                 return ret;
1015                         }
1016                 }
1017         }
1018
1019         return 0;
1020 }
1021 #endif
1022
1023 /* -----------------------------------------------------------------------------
1024  * DebugFS
1025  */
1026
1027 #ifdef CONFIG_DEBUG_FS
1028 void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1029 {
1030         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1031         uint64_t off;
1032
1033         off = drm_vma_node_start(&obj->vma_node);
1034
1035         seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1036                         omap_obj->flags, obj->name, kref_read(&obj->refcount),
1037                         off, &omap_obj->paddr, omap_obj->paddr_cnt,
1038                         omap_obj->vaddr, omap_obj->roll);
1039
1040         if (omap_obj->flags & OMAP_BO_TILED) {
1041                 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1042                 if (omap_obj->block) {
1043                         struct tcm_area *area = &omap_obj->block->area;
1044                         seq_printf(m, " (%dx%d, %dx%d)",
1045                                         area->p0.x, area->p0.y,
1046                                         area->p1.x, area->p1.y);
1047                 }
1048         } else {
1049                 seq_printf(m, " %d", obj->size);
1050         }
1051
1052         seq_printf(m, "\n");
1053 }
1054
1055 void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1056 {
1057         struct omap_gem_object *omap_obj;
1058         int count = 0;
1059         size_t size = 0;
1060
1061         list_for_each_entry(omap_obj, list, mm_list) {
1062                 struct drm_gem_object *obj = &omap_obj->base;
1063                 seq_printf(m, "   ");
1064                 omap_gem_describe(obj, m);
1065                 count++;
1066                 size += obj->size;
1067         }
1068
1069         seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1070 }
1071 #endif
1072
1073 /* -----------------------------------------------------------------------------
1074  * Buffer Synchronization
1075  */
1076
1077 static DEFINE_SPINLOCK(sync_lock);
1078
1079 struct omap_gem_sync_waiter {
1080         struct list_head list;
1081         struct omap_gem_object *omap_obj;
1082         enum omap_gem_op op;
1083         uint32_t read_target, write_target;
1084         /* notify called w/ sync_lock held */
1085         void (*notify)(void *arg);
1086         void *arg;
1087 };
1088
1089 /* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1090  * the read and/or write target count is achieved which can call a user
1091  * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1092  * cpu access), etc.
1093  */
1094 static LIST_HEAD(waiters);
1095
1096 static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1097 {
1098         struct omap_gem_object *omap_obj = waiter->omap_obj;
1099         if ((waiter->op & OMAP_GEM_READ) &&
1100                         (omap_obj->sync->write_complete < waiter->write_target))
1101                 return true;
1102         if ((waiter->op & OMAP_GEM_WRITE) &&
1103                         (omap_obj->sync->read_complete < waiter->read_target))
1104                 return true;
1105         return false;
1106 }
1107
1108 /* macro for sync debug.. */
1109 #define SYNCDBG 0
1110 #define SYNC(fmt, ...) do { if (SYNCDBG)                                \
1111                 pr_err("%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__); \
1112         } while (0)
1113
1114
1115 static void sync_op_update(void)
1116 {
1117         struct omap_gem_sync_waiter *waiter, *n;
1118         list_for_each_entry_safe(waiter, n, &waiters, list) {
1119                 if (!is_waiting(waiter)) {
1120                         list_del(&waiter->list);
1121                         SYNC("notify: %p", waiter);
1122                         waiter->notify(waiter->arg);
1123                         kfree(waiter);
1124                 }
1125         }
1126 }
1127
1128 static inline int sync_op(struct drm_gem_object *obj,
1129                 enum omap_gem_op op, bool start)
1130 {
1131         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1132         int ret = 0;
1133
1134         spin_lock(&sync_lock);
1135
1136         if (!omap_obj->sync) {
1137                 omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1138                 if (!omap_obj->sync) {
1139                         ret = -ENOMEM;
1140                         goto unlock;
1141                 }
1142         }
1143
1144         if (start) {
1145                 if (op & OMAP_GEM_READ)
1146                         omap_obj->sync->read_pending++;
1147                 if (op & OMAP_GEM_WRITE)
1148                         omap_obj->sync->write_pending++;
1149         } else {
1150                 if (op & OMAP_GEM_READ)
1151                         omap_obj->sync->read_complete++;
1152                 if (op & OMAP_GEM_WRITE)
1153                         omap_obj->sync->write_complete++;
1154                 sync_op_update();
1155         }
1156
1157 unlock:
1158         spin_unlock(&sync_lock);
1159
1160         return ret;
1161 }
1162
1163 /* mark the start of read and/or write operation */
1164 int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1165 {
1166         return sync_op(obj, op, true);
1167 }
1168
1169 int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1170 {
1171         return sync_op(obj, op, false);
1172 }
1173
1174 static DECLARE_WAIT_QUEUE_HEAD(sync_event);
1175
1176 static void sync_notify(void *arg)
1177 {
1178         struct task_struct **waiter_task = arg;
1179         *waiter_task = NULL;
1180         wake_up_all(&sync_event);
1181 }
1182
1183 int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1184 {
1185         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1186         int ret = 0;
1187         if (omap_obj->sync) {
1188                 struct task_struct *waiter_task = current;
1189                 struct omap_gem_sync_waiter *waiter =
1190                                 kzalloc(sizeof(*waiter), GFP_KERNEL);
1191
1192                 if (!waiter)
1193                         return -ENOMEM;
1194
1195                 waiter->omap_obj = omap_obj;
1196                 waiter->op = op;
1197                 waiter->read_target = omap_obj->sync->read_pending;
1198                 waiter->write_target = omap_obj->sync->write_pending;
1199                 waiter->notify = sync_notify;
1200                 waiter->arg = &waiter_task;
1201
1202                 spin_lock(&sync_lock);
1203                 if (is_waiting(waiter)) {
1204                         SYNC("waited: %p", waiter);
1205                         list_add_tail(&waiter->list, &waiters);
1206                         spin_unlock(&sync_lock);
1207                         ret = wait_event_interruptible(sync_event,
1208                                         (waiter_task == NULL));
1209                         spin_lock(&sync_lock);
1210                         if (waiter_task) {
1211                                 SYNC("interrupted: %p", waiter);
1212                                 /* we were interrupted */
1213                                 list_del(&waiter->list);
1214                                 waiter_task = NULL;
1215                         } else {
1216                                 /* freed in sync_op_update() */
1217                                 waiter = NULL;
1218                         }
1219                 }
1220                 spin_unlock(&sync_lock);
1221                 kfree(waiter);
1222         }
1223         return ret;
1224 }
1225
1226 /* call fxn(arg), either synchronously or asynchronously if the op
1227  * is currently blocked..  fxn() can be called from any context
1228  *
1229  * (TODO for now fxn is called back from whichever context calls
1230  * omap_gem_op_finish().. but this could be better defined later
1231  * if needed)
1232  *
1233  * TODO more code in common w/ _sync()..
1234  */
1235 int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1236                 void (*fxn)(void *arg), void *arg)
1237 {
1238         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1239         if (omap_obj->sync) {
1240                 struct omap_gem_sync_waiter *waiter =
1241                                 kzalloc(sizeof(*waiter), GFP_ATOMIC);
1242
1243                 if (!waiter)
1244                         return -ENOMEM;
1245
1246                 waiter->omap_obj = omap_obj;
1247                 waiter->op = op;
1248                 waiter->read_target = omap_obj->sync->read_pending;
1249                 waiter->write_target = omap_obj->sync->write_pending;
1250                 waiter->notify = fxn;
1251                 waiter->arg = arg;
1252
1253                 spin_lock(&sync_lock);
1254                 if (is_waiting(waiter)) {
1255                         SYNC("waited: %p", waiter);
1256                         list_add_tail(&waiter->list, &waiters);
1257                         spin_unlock(&sync_lock);
1258                         return 0;
1259                 }
1260
1261                 spin_unlock(&sync_lock);
1262
1263                 kfree(waiter);
1264         }
1265
1266         /* no waiting.. */
1267         fxn(arg);
1268
1269         return 0;
1270 }
1271
1272 /* -----------------------------------------------------------------------------
1273  * Constructor & Destructor
1274  */
1275
1276 void omap_gem_free_object(struct drm_gem_object *obj)
1277 {
1278         struct drm_device *dev = obj->dev;
1279         struct omap_drm_private *priv = dev->dev_private;
1280         struct omap_gem_object *omap_obj = to_omap_bo(obj);
1281
1282         evict(obj);
1283
1284         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1285
1286         spin_lock(&priv->list_lock);
1287         list_del(&omap_obj->mm_list);
1288         spin_unlock(&priv->list_lock);
1289
1290         /* this means the object is still pinned.. which really should
1291          * not happen.  I think..
1292          */
1293         WARN_ON(omap_obj->paddr_cnt > 0);
1294
1295         if (omap_obj->pages) {
1296                 if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1297                         kfree(omap_obj->pages);
1298                 else
1299                         omap_gem_detach_pages(obj);
1300         }
1301
1302         if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1303                 dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1304                             omap_obj->paddr);
1305         } else if (omap_obj->vaddr) {
1306                 vunmap(omap_obj->vaddr);
1307         } else if (obj->import_attach) {
1308                 drm_prime_gem_destroy(obj, omap_obj->sgt);
1309         }
1310
1311         kfree(omap_obj->sync);
1312
1313         drm_gem_object_release(obj);
1314
1315         kfree(omap_obj);
1316 }
1317
1318 /* GEM buffer object constructor */
1319 struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1320                 union omap_gem_size gsize, uint32_t flags)
1321 {
1322         struct omap_drm_private *priv = dev->dev_private;
1323         struct omap_gem_object *omap_obj;
1324         struct drm_gem_object *obj;
1325         struct address_space *mapping;
1326         size_t size;
1327         int ret;
1328
1329         /* Validate the flags and compute the memory and cache flags. */
1330         if (flags & OMAP_BO_TILED) {
1331                 if (!priv->usergart) {
1332                         dev_err(dev->dev, "Tiled buffers require DMM\n");
1333                         return NULL;
1334                 }
1335
1336                 /*
1337                  * Tiled buffers are always shmem paged backed. When they are
1338                  * scanned out, they are remapped into DMM/TILER.
1339                  */
1340                 flags &= ~OMAP_BO_SCANOUT;
1341                 flags |= OMAP_BO_MEM_SHMEM;
1342
1343                 /*
1344                  * Currently don't allow cached buffers. There is some caching
1345                  * stuff that needs to be handled better.
1346                  */
1347                 flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1348                 flags |= tiler_get_cpu_cache_flags();
1349         } else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1350                 /*
1351                  * OMAP_BO_SCANOUT hints that the buffer doesn't need to be
1352                  * tiled. However, to lower the pressure on memory allocation,
1353                  * use contiguous memory only if no TILER is available.
1354                  */
1355                 flags |= OMAP_BO_MEM_DMA_API;
1356         } else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1357                 /*
1358                  * All other buffers not backed by dma_buf are shmem-backed.
1359                  */
1360                 flags |= OMAP_BO_MEM_SHMEM;
1361         }
1362
1363         /* Allocate the initialize the OMAP GEM object. */
1364         omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1365         if (!omap_obj)
1366                 return NULL;
1367
1368         obj = &omap_obj->base;
1369         omap_obj->flags = flags;
1370
1371         if (flags & OMAP_BO_TILED) {
1372                 /*
1373                  * For tiled buffers align dimensions to slot boundaries and
1374                  * calculate size based on aligned dimensions.
1375                  */
1376                 tiler_align(gem2fmt(flags), &gsize.tiled.width,
1377                             &gsize.tiled.height);
1378
1379                 size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1380                                   gsize.tiled.height);
1381
1382                 omap_obj->width = gsize.tiled.width;
1383                 omap_obj->height = gsize.tiled.height;
1384         } else {
1385                 size = PAGE_ALIGN(gsize.bytes);
1386         }
1387
1388         /* Initialize the GEM object. */
1389         if (!(flags & OMAP_BO_MEM_SHMEM)) {
1390                 drm_gem_private_object_init(dev, obj, size);
1391         } else {
1392                 ret = drm_gem_object_init(dev, obj, size);
1393                 if (ret)
1394                         goto err_free;
1395
1396                 mapping = obj->filp->f_mapping;
1397                 mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1398         }
1399
1400         /* Allocate memory if needed. */
1401         if (flags & OMAP_BO_MEM_DMA_API) {
1402                 omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1403                                                &omap_obj->paddr,
1404                                                GFP_KERNEL);
1405                 if (!omap_obj->vaddr)
1406                         goto err_release;
1407         }
1408
1409         spin_lock(&priv->list_lock);
1410         list_add(&omap_obj->mm_list, &priv->obj_list);
1411         spin_unlock(&priv->list_lock);
1412
1413         return obj;
1414
1415 err_release:
1416         drm_gem_object_release(obj);
1417 err_free:
1418         kfree(omap_obj);
1419         return NULL;
1420 }
1421
1422 struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1423                                            struct sg_table *sgt)
1424 {
1425         struct omap_drm_private *priv = dev->dev_private;
1426         struct omap_gem_object *omap_obj;
1427         struct drm_gem_object *obj;
1428         union omap_gem_size gsize;
1429
1430         /* Without a DMM only physically contiguous buffers can be supported. */
1431         if (sgt->orig_nents != 1 && !priv->has_dmm)
1432                 return ERR_PTR(-EINVAL);
1433
1434         mutex_lock(&dev->struct_mutex);
1435
1436         gsize.bytes = PAGE_ALIGN(size);
1437         obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1438         if (!obj) {
1439                 obj = ERR_PTR(-ENOMEM);
1440                 goto done;
1441         }
1442
1443         omap_obj = to_omap_bo(obj);
1444         omap_obj->sgt = sgt;
1445
1446         if (sgt->orig_nents == 1) {
1447                 omap_obj->paddr = sg_dma_address(sgt->sgl);
1448         } else {
1449                 /* Create pages list from sgt */
1450                 struct sg_page_iter iter;
1451                 struct page **pages;
1452                 unsigned int npages;
1453                 unsigned int i = 0;
1454
1455                 npages = DIV_ROUND_UP(size, PAGE_SIZE);
1456                 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1457                 if (!pages) {
1458                         omap_gem_free_object(obj);
1459                         obj = ERR_PTR(-ENOMEM);
1460                         goto done;
1461                 }
1462
1463                 omap_obj->pages = pages;
1464
1465                 for_each_sg_page(sgt->sgl, &iter, sgt->orig_nents, 0) {
1466                         pages[i++] = sg_page_iter_page(&iter);
1467                         if (i > npages)
1468                                 break;
1469                 }
1470
1471                 if (WARN_ON(i != npages)) {
1472                         omap_gem_free_object(obj);
1473                         obj = ERR_PTR(-ENOMEM);
1474                         goto done;
1475                 }
1476         }
1477
1478 done:
1479         mutex_unlock(&dev->struct_mutex);
1480         return obj;
1481 }
1482
1483 /* convenience method to construct a GEM buffer object, and userspace handle */
1484 int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1485                 union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1486 {
1487         struct drm_gem_object *obj;
1488         int ret;
1489
1490         obj = omap_gem_new(dev, gsize, flags);
1491         if (!obj)
1492                 return -ENOMEM;
1493
1494         ret = drm_gem_handle_create(file, obj, handle);
1495         if (ret) {
1496                 omap_gem_free_object(obj);
1497                 return ret;
1498         }
1499
1500         /* drop reference from allocate - handle holds it now */
1501         drm_gem_object_unreference_unlocked(obj);
1502
1503         return 0;
1504 }
1505
1506 /* -----------------------------------------------------------------------------
1507  * Init & Cleanup
1508  */
1509
1510 /* If DMM is used, we need to set some stuff up.. */
1511 void omap_gem_init(struct drm_device *dev)
1512 {
1513         struct omap_drm_private *priv = dev->dev_private;
1514         struct omap_drm_usergart *usergart;
1515         const enum tiler_fmt fmts[] = {
1516                         TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1517         };
1518         int i, j;
1519
1520         if (!dmm_is_available()) {
1521                 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1522                 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1523                 return;
1524         }
1525
1526         usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1527         if (!usergart)
1528                 return;
1529
1530         /* reserve 4k aligned/wide regions for userspace mappings: */
1531         for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1532                 uint16_t h = 1, w = PAGE_SIZE >> i;
1533                 tiler_align(fmts[i], &w, &h);
1534                 /* note: since each region is 1 4kb page wide, and minimum
1535                  * number of rows, the height ends up being the same as the
1536                  * # of pages in the region
1537                  */
1538                 usergart[i].height = h;
1539                 usergart[i].height_shift = ilog2(h);
1540                 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1541                 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1542                 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1543                         struct omap_drm_usergart_entry *entry;
1544                         struct tiler_block *block;
1545
1546                         entry = &usergart[i].entry[j];
1547                         block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1548                         if (IS_ERR(block)) {
1549                                 dev_err(dev->dev,
1550                                                 "reserve failed: %d, %d, %ld\n",
1551                                                 i, j, PTR_ERR(block));
1552                                 return;
1553                         }
1554                         entry->paddr = tiler_ssptr(block);
1555                         entry->block = block;
1556
1557                         DBG("%d:%d: %dx%d: paddr=%pad stride=%d", i, j, w, h,
1558                                         &entry->paddr,
1559                                         usergart[i].stride_pfn << PAGE_SHIFT);
1560                 }
1561         }
1562
1563         priv->usergart = usergart;
1564         priv->has_dmm = true;
1565 }
1566
1567 void omap_gem_deinit(struct drm_device *dev)
1568 {
1569         struct omap_drm_private *priv = dev->dev_private;
1570
1571         /* I believe we can rely on there being no more outstanding GEM
1572          * objects which could depend on usergart/dmm at this point.
1573          */
1574         kfree(priv->usergart);
1575 }