Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2011 Texas Instruments Incorporated - https://www.ti.com/
 
 
   4 * Author: Rob Clark <rob.clark@linaro.org>
 
 
 
 
 
 
 
 
 
 
 
 
   5 */
   6
   7#include <linux/dma-mapping.h>
   8#include <linux/seq_file.h>
   9#include <linux/shmem_fs.h>
  10#include <linux/spinlock.h>
  11#include <linux/pfn_t.h>
  12
  13#include <drm/drm_prime.h>
 
  14#include <drm/drm_vma_manager.h>
  15
  16#include "omap_drv.h"
  17#include "omap_dmm_tiler.h"
  18
 
 
 
 
 
 
  19/*
  20 * GEM buffer object implementation.
  21 */
  22
 
 
  23/* note: we use upper 8 bits of flags for driver-internal flags: */
  24#define OMAP_BO_MEM_DMA_API	0x01000000	/* memory allocated with the dma_alloc_* API */
  25#define OMAP_BO_MEM_SHMEM	0x02000000	/* memory allocated through shmem backing */
  26#define OMAP_BO_MEM_DMABUF	0x08000000	/* memory imported from a dmabuf */
 
  27
  28struct omap_gem_object {
  29	struct drm_gem_object base;
  30
  31	struct list_head mm_list;
  32
  33	u32 flags;
  34
  35	/** width/height for tiled formats (rounded up to slot boundaries) */
  36	u16 width, height;
  37
  38	/** roll applied when mapping to DMM */
  39	u32 roll;
  40
  41	/** protects pin_cnt, block, pages, dma_addrs and vaddr */
  42	struct mutex lock;
  43
  44	/**
  45	 * dma_addr contains the buffer DMA address. It is valid for
  46	 *
  47	 * - buffers allocated through the DMA mapping API (with the
  48	 *   OMAP_BO_MEM_DMA_API flag set)
  49	 *
  50	 * - buffers imported from dmabuf (with the OMAP_BO_MEM_DMABUF flag set)
  51	 *   if they are physically contiguous
  52	 *
  53	 * - buffers mapped through the TILER when pin_cnt is not zero, in which
  54	 *   case the DMA address points to the TILER aperture
  55	 *
  56	 * Physically contiguous buffers have their DMA address equal to the
  57	 * physical address as we don't remap those buffers through the TILER.
  58	 *
  59	 * Buffers mapped to the TILER have their DMA address pointing to the
  60	 * TILER aperture. As TILER mappings are refcounted (through pin_cnt)
  61	 * the DMA address must be accessed through omap_gem_pin() to ensure
  62	 * that the mapping won't disappear unexpectedly. References must be
  63	 * released with omap_gem_unpin().
  64	 */
  65	dma_addr_t dma_addr;
  66
  67	/**
  68	 * # of users
  69	 */
  70	refcount_t pin_cnt;
  71
  72	/**
  73	 * If the buffer has been imported from a dmabuf the OMAP_DB_DMABUF flag
  74	 * is set and the sgt field is valid.
  75	 */
  76	struct sg_table *sgt;
  77
  78	/**
  79	 * tiler block used when buffer is remapped in DMM/TILER.
  80	 */
  81	struct tiler_block *block;
  82
  83	/**
  84	 * Array of backing pages, if allocated.  Note that pages are never
  85	 * allocated for buffers originally allocated from contiguous memory
  86	 */
  87	struct page **pages;
  88
  89	/** addresses corresponding to pages in above array */
  90	dma_addr_t *dma_addrs;
  91
  92	/**
  93	 * Virtual address, if mapped.
  94	 */
  95	void *vaddr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  96};
  97
  98#define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
 
  99
 100/* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
 101 * not necessarily pinned in TILER all the time, and (b) when they are
 102 * they are not necessarily page aligned, we reserve one or more small
 103 * regions in each of the 2d containers to use as a user-GART where we
 104 * can create a second page-aligned mapping of parts of the buffer
 105 * being accessed from userspace.
 106 *
 107 * Note that we could optimize slightly when we know that multiple
 108 * tiler containers are backed by the same PAT.. but I'll leave that
 109 * for later..
 110 */
 111#define NUM_USERGART_ENTRIES 2
 112struct omap_drm_usergart_entry {
 113	struct tiler_block *block;	/* the reserved tiler block */
 114	dma_addr_t dma_addr;
 115	struct drm_gem_object *obj;	/* the current pinned obj */
 116	pgoff_t obj_pgoff;		/* page offset of obj currently
 117					   mapped in */
 118};
 119
 120struct omap_drm_usergart {
 121	struct omap_drm_usergart_entry entry[NUM_USERGART_ENTRIES];
 122	int height;				/* height in rows */
 123	int height_shift;		/* ilog2(height in rows) */
 124	int slot_shift;			/* ilog2(width per slot) */
 125	int stride_pfn;			/* stride in pages */
 126	int last;				/* index of last used entry */
 127};
 128
 129/* -----------------------------------------------------------------------------
 130 * Helpers
 131 */
 132
 133/** get mmap offset */
 134u64 omap_gem_mmap_offset(struct drm_gem_object *obj)
 135{
 136	struct drm_device *dev = obj->dev;
 137	int ret;
 138	size_t size;
 139
 140	/* Make it mmapable */
 141	size = omap_gem_mmap_size(obj);
 142	ret = drm_gem_create_mmap_offset_size(obj, size);
 143	if (ret) {
 144		dev_err(dev->dev, "could not allocate mmap offset\n");
 145		return 0;
 146	}
 147
 148	return drm_vma_node_offset_addr(&obj->vma_node);
 149}
 150
 151static bool omap_gem_sgt_is_contiguous(struct sg_table *sgt, size_t size)
 152{
 153	return !(drm_prime_get_contiguous_size(sgt) < size);
 154}
 155
 156static bool omap_gem_is_contiguous(struct omap_gem_object *omap_obj)
 157{
 158	if (omap_obj->flags & OMAP_BO_MEM_DMA_API)
 159		return true;
 160
 161	if ((omap_obj->flags & OMAP_BO_MEM_DMABUF) &&
 162	    omap_gem_sgt_is_contiguous(omap_obj->sgt, omap_obj->base.size))
 163		return true;
 164
 165	return false;
 166}
 167
 168/* -----------------------------------------------------------------------------
 169 * Eviction
 170 */
 171
 172static void omap_gem_evict_entry(struct drm_gem_object *obj,
 173		enum tiler_fmt fmt, struct omap_drm_usergart_entry *entry)
 174{
 175	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 176	struct omap_drm_private *priv = obj->dev->dev_private;
 177	int n = priv->usergart[fmt].height;
 178	size_t size = PAGE_SIZE * n;
 179	loff_t off = omap_gem_mmap_offset(obj) +
 180			(entry->obj_pgoff << PAGE_SHIFT);
 181	const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
 182
 183	if (m > 1) {
 184		int i;
 185		/* if stride > than PAGE_SIZE then sparse mapping: */
 186		for (i = n; i > 0; i--) {
 187			unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 188					    off, PAGE_SIZE, 1);
 189			off += PAGE_SIZE * m;
 190		}
 191	} else {
 192		unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 193				    off, size, 1);
 194	}
 195
 196	entry->obj = NULL;
 197}
 198
 199/* Evict a buffer from usergart, if it is mapped there */
 200static void omap_gem_evict(struct drm_gem_object *obj)
 201{
 202	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 203	struct omap_drm_private *priv = obj->dev->dev_private;
 204
 205	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
 206		enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 207		int i;
 208
 209		for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
 210			struct omap_drm_usergart_entry *entry =
 211				&priv->usergart[fmt].entry[i];
 212
 
 
 213			if (entry->obj == obj)
 214				omap_gem_evict_entry(obj, fmt, entry);
 215		}
 216	}
 217}
 218
 219/* -----------------------------------------------------------------------------
 220 * Page Management
 
 
 
 
 221 */
 
 
 
 
 222
 223/*
 224 * Ensure backing pages are allocated. Must be called with the omap_obj.lock
 225 * held.
 226 */
 
 
 
 
 
 
 
 
 
 
 227static int omap_gem_attach_pages(struct drm_gem_object *obj)
 228{
 229	struct drm_device *dev = obj->dev;
 230	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 231	struct page **pages;
 232	int npages = obj->size >> PAGE_SHIFT;
 233	int i, ret;
 234	dma_addr_t *addrs;
 235
 236	lockdep_assert_held(&omap_obj->lock);
 237
 238	/*
 239	 * If not using shmem (in which case backing pages don't need to be
 240	 * allocated) or if pages are already allocated we're done.
 241	 */
 242	if (!(omap_obj->flags & OMAP_BO_MEM_SHMEM) || omap_obj->pages)
 243		return 0;
 244
 245	pages = drm_gem_get_pages(obj);
 246	if (IS_ERR(pages)) {
 247		dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
 248		return PTR_ERR(pages);
 249	}
 250
 251	/* for non-cached buffers, ensure the new pages are clean because
 252	 * DSS, GPU, etc. are not cache coherent:
 253	 */
 254	if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
 255		addrs = kmalloc_array(npages, sizeof(*addrs), GFP_KERNEL);
 256		if (!addrs) {
 257			ret = -ENOMEM;
 258			goto free_pages;
 259		}
 260
 261		for (i = 0; i < npages; i++) {
 262			addrs[i] = dma_map_page(dev->dev, pages[i],
 263					0, PAGE_SIZE, DMA_TO_DEVICE);
 264
 265			if (dma_mapping_error(dev->dev, addrs[i])) {
 266				dev_warn(dev->dev,
 267					"%s: failed to map page\n", __func__);
 268
 269				for (i = i - 1; i >= 0; --i) {
 270					dma_unmap_page(dev->dev, addrs[i],
 271						PAGE_SIZE, DMA_TO_DEVICE);
 272				}
 273
 274				ret = -ENOMEM;
 275				goto free_addrs;
 276			}
 277		}
 278	} else {
 279		addrs = kcalloc(npages, sizeof(*addrs), GFP_KERNEL);
 280		if (!addrs) {
 281			ret = -ENOMEM;
 282			goto free_pages;
 283		}
 284	}
 285
 286	omap_obj->dma_addrs = addrs;
 287	omap_obj->pages = pages;
 288
 289	return 0;
 290
 291free_addrs:
 292	kfree(addrs);
 293free_pages:
 294	drm_gem_put_pages(obj, pages, true, false);
 295
 296	return ret;
 297}
 298
 299/* Release backing pages. Must be called with the omap_obj.lock held. */
 300static void omap_gem_detach_pages(struct drm_gem_object *obj)
 301{
 302	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 303	unsigned int npages = obj->size >> PAGE_SHIFT;
 304	unsigned int i;
 305
 306	lockdep_assert_held(&omap_obj->lock);
 307
 308	for (i = 0; i < npages; i++) {
 309		if (omap_obj->dma_addrs[i])
 310			dma_unmap_page(obj->dev->dev, omap_obj->dma_addrs[i],
 311				       PAGE_SIZE, DMA_TO_DEVICE);
 
 
 
 
 
 312	}
 313
 314	kfree(omap_obj->dma_addrs);
 315	omap_obj->dma_addrs = NULL;
 316
 317	drm_gem_put_pages(obj, omap_obj->pages, true, false);
 318	omap_obj->pages = NULL;
 319}
 320
 321/* get buffer flags */
 322u32 omap_gem_flags(struct drm_gem_object *obj)
 323{
 324	return to_omap_bo(obj)->flags;
 325}
 326
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 327/** get mmap size */
 328size_t omap_gem_mmap_size(struct drm_gem_object *obj)
 329{
 330	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 331	size_t size = obj->size;
 332
 333	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
 334		/* for tiled buffers, the virtual size has stride rounded up
 335		 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
 336		 * 32kb later!).  But we don't back the entire buffer with
 337		 * pages, only the valid picture part.. so need to adjust for
 338		 * this in the size used to mmap and generate mmap offset
 339		 */
 340		size = tiler_vsize(gem2fmt(omap_obj->flags),
 341				omap_obj->width, omap_obj->height);
 342	}
 343
 344	return size;
 345}
 346
 347/* -----------------------------------------------------------------------------
 348 * Fault Handling
 349 */
 
 
 
 
 
 
 
 
 350
 351/* Normal handling for the case of faulting in non-tiled buffers */
 352static vm_fault_t omap_gem_fault_1d(struct drm_gem_object *obj,
 353		struct vm_area_struct *vma, struct vm_fault *vmf)
 354{
 355	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 356	unsigned long pfn;
 357	pgoff_t pgoff;
 358
 359	/* We don't use vmf->pgoff since that has the fake offset: */
 360	pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
 
 361
 362	if (omap_obj->pages) {
 363		omap_gem_cpu_sync_page(obj, pgoff);
 364		pfn = page_to_pfn(omap_obj->pages[pgoff]);
 365	} else {
 366		BUG_ON(!omap_gem_is_contiguous(omap_obj));
 367		pfn = (omap_obj->dma_addr >> PAGE_SHIFT) + pgoff;
 368	}
 369
 370	VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
 371			pfn, pfn << PAGE_SHIFT);
 372
 373	return vmf_insert_mixed(vma, vmf->address,
 374			__pfn_to_pfn_t(pfn, PFN_DEV));
 375}
 376
 377/* Special handling for the case of faulting in 2d tiled buffers */
 378static vm_fault_t omap_gem_fault_2d(struct drm_gem_object *obj,
 379		struct vm_area_struct *vma, struct vm_fault *vmf)
 380{
 381	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 382	struct omap_drm_private *priv = obj->dev->dev_private;
 383	struct omap_drm_usergart_entry *entry;
 384	enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 385	struct page *pages[64];  /* XXX is this too much to have on stack? */
 386	unsigned long pfn;
 387	pgoff_t pgoff, base_pgoff;
 388	unsigned long vaddr;
 389	int i, err, slots;
 390	vm_fault_t ret = VM_FAULT_NOPAGE;
 391
 392	/*
 393	 * Note the height of the slot is also equal to the number of pages
 394	 * that need to be mapped in to fill 4kb wide CPU page.  If the slot
 395	 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
 396	 */
 397	const int n = priv->usergart[fmt].height;
 398	const int n_shift = priv->usergart[fmt].height_shift;
 399
 400	/*
 401	 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
 402	 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
 403	 * into account in some of the math, so figure out virtual stride
 404	 * in pages
 405	 */
 406	const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
 407
 408	/* We don't use vmf->pgoff since that has the fake offset: */
 409	pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
 
 410
 411	/*
 412	 * Actual address we start mapping at is rounded down to previous slot
 413	 * boundary in the y direction:
 414	 */
 415	base_pgoff = round_down(pgoff, m << n_shift);
 416
 417	/* figure out buffer width in slots */
 418	slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
 419
 420	vaddr = vmf->address - ((pgoff - base_pgoff) << PAGE_SHIFT);
 421
 422	entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
 423
 424	/* evict previous buffer using this usergart entry, if any: */
 425	if (entry->obj)
 426		omap_gem_evict_entry(entry->obj, fmt, entry);
 427
 428	entry->obj = obj;
 429	entry->obj_pgoff = base_pgoff;
 430
 431	/* now convert base_pgoff to phys offset from virt offset: */
 432	base_pgoff = (base_pgoff >> n_shift) * slots;
 433
 434	/* for wider-than 4k.. figure out which part of the slot-row we want: */
 435	if (m > 1) {
 436		int off = pgoff % m;
 437		entry->obj_pgoff += off;
 438		base_pgoff /= m;
 439		slots = min(slots - (off << n_shift), n);
 440		base_pgoff += off << n_shift;
 441		vaddr += off << PAGE_SHIFT;
 442	}
 443
 444	/*
 445	 * Map in pages. Beyond the valid pixel part of the buffer, we set
 446	 * pages[i] to NULL to get a dummy page mapped in.. if someone
 447	 * reads/writes it they will get random/undefined content, but at
 448	 * least it won't be corrupting whatever other random page used to
 449	 * be mapped in, or other undefined behavior.
 450	 */
 451	memcpy(pages, &omap_obj->pages[base_pgoff],
 452			sizeof(struct page *) * slots);
 453	memset(pages + slots, 0,
 454			sizeof(struct page *) * (n - slots));
 455
 456	err = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
 457	if (err) {
 458		ret = vmf_error(err);
 459		dev_err(obj->dev->dev, "failed to pin: %d\n", err);
 460		return ret;
 461	}
 462
 463	pfn = entry->dma_addr >> PAGE_SHIFT;
 464
 465	VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
 466			pfn, pfn << PAGE_SHIFT);
 467
 468	for (i = n; i > 0; i--) {
 469		ret = vmf_insert_mixed(vma,
 470			vaddr, __pfn_to_pfn_t(pfn, PFN_DEV));
 471		if (ret & VM_FAULT_ERROR)
 472			break;
 473		pfn += priv->usergart[fmt].stride_pfn;
 474		vaddr += PAGE_SIZE * m;
 475	}
 476
 477	/* simple round-robin: */
 478	priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
 479				 % NUM_USERGART_ENTRIES;
 480
 481	return ret;
 482}
 483
 484/**
 485 * omap_gem_fault		-	pagefault handler for GEM objects
 
 486 * @vmf: fault detail
 487 *
 488 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
 489 * does most of the work for us including the actual map/unmap calls
 490 * but we need to do the actual page work.
 491 *
 492 * The VMA was set up by GEM. In doing so it also ensured that the
 493 * vma->vm_private_data points to the GEM object that is backing this
 494 * mapping.
 495 */
 496static vm_fault_t omap_gem_fault(struct vm_fault *vmf)
 497{
 498	struct vm_area_struct *vma = vmf->vma;
 499	struct drm_gem_object *obj = vma->vm_private_data;
 500	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 501	int err;
 502	vm_fault_t ret;
 
 503
 504	/* Make sure we don't parallel update on a fault, nor move or remove
 505	 * something from beneath our feet
 506	 */
 507	mutex_lock(&omap_obj->lock);
 508
 509	/* if a shmem backed object, make sure we have pages attached now */
 510	err = omap_gem_attach_pages(obj);
 511	if (err) {
 512		ret = vmf_error(err);
 513		goto fail;
 514	}
 515
 516	/* where should we do corresponding put_pages().. we are mapping
 517	 * the original page, rather than thru a GART, so we can't rely
 518	 * on eviction to trigger this.  But munmap() or all mappings should
 519	 * probably trigger put_pages()?
 520	 */
 521
 522	if (omap_obj->flags & OMAP_BO_TILED_MASK)
 523		ret = omap_gem_fault_2d(obj, vma, vmf);
 524	else
 525		ret = omap_gem_fault_1d(obj, vma, vmf);
 526
 527
 528fail:
 529	mutex_unlock(&omap_obj->lock);
 530	return ret;
 
 
 
 
 
 
 
 
 
 531}
 532
 533static int omap_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 534{
 535	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 536
 537	vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP | VM_IO | VM_MIXEDMAP);
 
 538
 539	if (omap_obj->flags & OMAP_BO_WC) {
 540		vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
 541	} else if (omap_obj->flags & OMAP_BO_UNCACHED) {
 542		vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
 543	} else {
 544		/*
 545		 * We do have some private objects, at least for scanout buffers
 546		 * on hardware without DMM/TILER.  But these are allocated write-
 547		 * combine
 548		 */
 549		if (WARN_ON(!obj->filp))
 550			return -EINVAL;
 551
 552		/*
 553		 * Shunt off cached objs to shmem file so they have their own
 554		 * address_space (so unmap_mapping_range does what we want,
 555		 * in particular in the case of mmap'd dmabufs)
 556		 */
 557		vma->vm_pgoff -= drm_vma_node_start(&obj->vma_node);
 558		vma_set_file(vma, obj->filp);
 
 559
 560		vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 561	}
 562
 563	vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
 564
 565	return 0;
 566}
 567
 568/* -----------------------------------------------------------------------------
 569 * Dumb Buffers
 570 */
 571
 572/**
 573 * omap_gem_dumb_create	-	create a dumb buffer
 574 * @file: our client file
 575 * @dev: our device
 576 * @args: the requested arguments copied from userspace
 577 *
 578 * Allocate a buffer suitable for use for a frame buffer of the
 579 * form described by user space. Give userspace a handle by which
 580 * to reference it.
 581 */
 582int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
 583		struct drm_mode_create_dumb *args)
 584{
 585	union omap_gem_size gsize;
 586
 587	args->pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
 588
 589	args->size = PAGE_ALIGN(args->pitch * args->height);
 590
 591	gsize = (union omap_gem_size){
 592		.bytes = args->size,
 593	};
 594
 595	return omap_gem_new_handle(dev, file, gsize,
 596			OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
 597}
 598
 599/**
 600 * omap_gem_dumb_map_offset - create an offset for a dumb buffer
 601 * @file: our drm client file
 602 * @dev: drm device
 603 * @handle: GEM handle to the object (from dumb_create)
 604 * @offset: memory map offset placeholder
 605 *
 606 * Do the necessary setup to allow the mapping of the frame buffer
 607 * into user memory. We don't have to do much here at the moment.
 608 */
 609int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 610		u32 handle, u64 *offset)
 611{
 612	struct drm_gem_object *obj;
 613	int ret = 0;
 614
 615	/* GEM does all our handle to object mapping */
 616	obj = drm_gem_object_lookup(file, handle);
 617	if (obj == NULL) {
 618		ret = -ENOENT;
 619		goto fail;
 620	}
 621
 622	*offset = omap_gem_mmap_offset(obj);
 623
 624	drm_gem_object_put(obj);
 625
 626fail:
 627	return ret;
 628}
 629
 630#ifdef CONFIG_DRM_FBDEV_EMULATION
 631/* Set scrolling position.  This allows us to implement fast scrolling
 632 * for console.
 633 *
 634 * Call only from non-atomic contexts.
 635 */
 636int omap_gem_roll(struct drm_gem_object *obj, u32 roll)
 637{
 638	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 639	u32 npages = obj->size >> PAGE_SHIFT;
 640	int ret = 0;
 641
 642	if (roll > npages) {
 643		dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
 644		return -EINVAL;
 645	}
 646
 647	omap_obj->roll = roll;
 648
 649	mutex_lock(&omap_obj->lock);
 650
 651	/* if we aren't mapped yet, we don't need to do anything */
 652	if (omap_obj->block) {
 653		ret = omap_gem_attach_pages(obj);
 
 654		if (ret)
 655			goto fail;
 656
 657		ret = tiler_pin(omap_obj->block, omap_obj->pages, npages,
 658				roll, true);
 659		if (ret)
 660			dev_err(obj->dev->dev, "could not repin: %d\n", ret);
 661	}
 662
 663fail:
 664	mutex_unlock(&omap_obj->lock);
 665
 666	return ret;
 667}
 668#endif
 669
 670/* -----------------------------------------------------------------------------
 671 * Memory Management & DMA Sync
 672 */
 673
 674/*
 675 * shmem buffers that are mapped cached are not coherent.
 676 *
 677 * We keep track of dirty pages using page faulting to perform cache management.
 678 * When a page is mapped to the CPU in read/write mode the device can't access
 679 * it and omap_obj->dma_addrs[i] is NULL. When a page is mapped to the device
 680 * the omap_obj->dma_addrs[i] is set to the DMA address, and the page is
 681 * unmapped from the CPU.
 682 */
 683static inline bool omap_gem_is_cached_coherent(struct drm_gem_object *obj)
 684{
 685	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 686
 687	return !((omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
 688		((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED));
 689}
 690
 691/* Sync the buffer for CPU access.. note pages should already be
 692 * attached, ie. omap_gem_get_pages()
 693 */
 694void omap_gem_cpu_sync_page(struct drm_gem_object *obj, int pgoff)
 695{
 696	struct drm_device *dev = obj->dev;
 697	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 698
 699	if (omap_gem_is_cached_coherent(obj))
 700		return;
 701
 702	if (omap_obj->dma_addrs[pgoff]) {
 703		dma_unmap_page(dev->dev, omap_obj->dma_addrs[pgoff],
 704				PAGE_SIZE, DMA_TO_DEVICE);
 705		omap_obj->dma_addrs[pgoff] = 0;
 706	}
 707}
 708
 709/* sync the buffer for DMA access */
 710void omap_gem_dma_sync_buffer(struct drm_gem_object *obj,
 711		enum dma_data_direction dir)
 712{
 713	struct drm_device *dev = obj->dev;
 714	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 715	int i, npages = obj->size >> PAGE_SHIFT;
 716	struct page **pages = omap_obj->pages;
 717	bool dirty = false;
 718
 719	if (omap_gem_is_cached_coherent(obj))
 720		return;
 
 
 721
 722	for (i = 0; i < npages; i++) {
 723		if (!omap_obj->dma_addrs[i]) {
 724			dma_addr_t addr;
 725
 726			addr = dma_map_page(dev->dev, pages[i], 0,
 727					    PAGE_SIZE, dir);
 728			if (dma_mapping_error(dev->dev, addr)) {
 729				dev_warn(dev->dev, "%s: failed to map page\n",
 730					__func__);
 731				break;
 732			}
 733
 734			dirty = true;
 735			omap_obj->dma_addrs[i] = addr;
 736		}
 737	}
 738
 739	if (dirty) {
 740		unmap_mapping_range(obj->filp->f_mapping, 0,
 741				    omap_gem_mmap_size(obj), 1);
 742	}
 743}
 744
 745static int omap_gem_pin_tiler(struct drm_gem_object *obj)
 746{
 747	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 748	u32 npages = obj->size >> PAGE_SHIFT;
 749	enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 750	struct tiler_block *block;
 751	int ret;
 752
 753	BUG_ON(omap_obj->block);
 754
 755	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
 756		block = tiler_reserve_2d(fmt, omap_obj->width, omap_obj->height,
 757					 PAGE_SIZE);
 758	} else {
 759		block = tiler_reserve_1d(obj->size);
 760	}
 761
 762	if (IS_ERR(block)) {
 763		ret = PTR_ERR(block);
 764		dev_err(obj->dev->dev, "could not remap: %d (%d)\n", ret, fmt);
 765		goto fail;
 766	}
 767
 768	/* TODO: enable async refill.. */
 769	ret = tiler_pin(block, omap_obj->pages, npages, omap_obj->roll, true);
 770	if (ret) {
 771		tiler_release(block);
 772		dev_err(obj->dev->dev, "could not pin: %d\n", ret);
 773		goto fail;
 774	}
 775
 776	omap_obj->dma_addr = tiler_ssptr(block);
 777	omap_obj->block = block;
 778
 779	DBG("got dma address: %pad", &omap_obj->dma_addr);
 780
 781fail:
 782	return ret;
 783}
 784
 785/**
 786 * omap_gem_pin() - Pin a GEM object in memory
 787 * @obj: the GEM object
 788 * @dma_addr: the DMA address
 789 *
 790 * Pin the given GEM object in memory and fill the dma_addr pointer with the
 791 * object's DMA address. If the buffer is not physically contiguous it will be
 792 * remapped through the TILER to provide a contiguous view.
 793 *
 794 * Pins are reference-counted, calling this function multiple times is allowed
 795 * as long the corresponding omap_gem_unpin() calls are balanced.
 796 *
 797 * Return 0 on success or a negative error code otherwise.
 798 */
 799int omap_gem_pin(struct drm_gem_object *obj, dma_addr_t *dma_addr)
 
 800{
 801	struct omap_drm_private *priv = obj->dev->dev_private;
 802	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 803	int ret = 0;
 804
 805	mutex_lock(&omap_obj->lock);
 806
 807	if (!omap_gem_is_contiguous(omap_obj)) {
 808		if (refcount_read(&omap_obj->pin_cnt) == 0) {
 
 
 
 
 809
 810			refcount_set(&omap_obj->pin_cnt, 1);
 811
 812			ret = omap_gem_attach_pages(obj);
 813			if (ret)
 814				goto fail;
 815
 816			if (omap_obj->flags & OMAP_BO_SCANOUT) {
 817				if (priv->has_dmm) {
 818					ret = omap_gem_pin_tiler(obj);
 819					if (ret)
 820						goto fail;
 821				}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 822			}
 823		} else {
 824			refcount_inc(&omap_obj->pin_cnt);
 
 
 
 825		}
 826	}
 827
 828	if (dma_addr)
 829		*dma_addr = omap_obj->dma_addr;
 
 
 
 
 
 
 
 830
 831fail:
 832	mutex_unlock(&omap_obj->lock);
 833
 834	return ret;
 835}
 836
 837/**
 838 * omap_gem_unpin_locked() - Unpin a GEM object from memory
 839 * @obj: the GEM object
 840 *
 841 * omap_gem_unpin() without locking.
 842 */
 843static void omap_gem_unpin_locked(struct drm_gem_object *obj)
 844{
 845	struct omap_drm_private *priv = obj->dev->dev_private;
 846	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 847	int ret;
 848
 849	if (omap_gem_is_contiguous(omap_obj))
 850		return;
 851
 852	if (refcount_dec_and_test(&omap_obj->pin_cnt)) {
 853		if (omap_obj->sgt) {
 854			sg_free_table(omap_obj->sgt);
 855			kfree(omap_obj->sgt);
 856			omap_obj->sgt = NULL;
 857		}
 858		if (!(omap_obj->flags & OMAP_BO_SCANOUT))
 859			return;
 860		if (priv->has_dmm) {
 861			ret = tiler_unpin(omap_obj->block);
 862			if (ret) {
 863				dev_err(obj->dev->dev,
 864					"could not unpin pages: %d\n", ret);
 
 865			}
 866			ret = tiler_release(omap_obj->block);
 867			if (ret) {
 868				dev_err(obj->dev->dev,
 869					"could not release unmap: %d\n", ret);
 870			}
 871			omap_obj->dma_addr = 0;
 872			omap_obj->block = NULL;
 873		}
 874	}
 875}
 876
 877/**
 878 * omap_gem_unpin() - Unpin a GEM object from memory
 879 * @obj: the GEM object
 880 *
 881 * Unpin the given GEM object previously pinned with omap_gem_pin(). Pins are
 882 * reference-counted, the actual unpin will only be performed when the number
 883 * of calls to this function matches the number of calls to omap_gem_pin().
 884 */
 885void omap_gem_unpin(struct drm_gem_object *obj)
 886{
 887	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 888
 889	mutex_lock(&omap_obj->lock);
 890	omap_gem_unpin_locked(obj);
 891	mutex_unlock(&omap_obj->lock);
 892}
 893
 894/* Get rotated scanout address (only valid if already pinned), at the
 895 * specified orientation and x,y offset from top-left corner of buffer
 896 * (only valid for tiled 2d buffers)
 897 */
 898int omap_gem_rotated_dma_addr(struct drm_gem_object *obj, u32 orient,
 899		int x, int y, dma_addr_t *dma_addr)
 900{
 901	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 902	int ret = -EINVAL;
 903
 904	mutex_lock(&omap_obj->lock);
 905
 906	if ((refcount_read(&omap_obj->pin_cnt) > 0) && omap_obj->block &&
 907			(omap_obj->flags & OMAP_BO_TILED_MASK)) {
 908		*dma_addr = tiler_tsptr(omap_obj->block, orient, x, y);
 909		ret = 0;
 910	}
 911
 912	mutex_unlock(&omap_obj->lock);
 913
 914	return ret;
 915}
 916
 917/* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
 918int omap_gem_tiled_stride(struct drm_gem_object *obj, u32 orient)
 919{
 920	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 921	int ret = -EINVAL;
 922	if (omap_obj->flags & OMAP_BO_TILED_MASK)
 923		ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
 924	return ret;
 925}
 926
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 927/* if !remap, and we don't have pages backing, then fail, rather than
 928 * increasing the pin count (which we don't really do yet anyways,
 929 * because we don't support swapping pages back out).  And 'remap'
 930 * might not be quite the right name, but I wanted to keep it working
 931 * similarly to omap_gem_pin().  Note though that mutex is not
 932 * aquired if !remap (because this can be called in atomic ctxt),
 933 * but probably omap_gem_unpin() should be changed to work in the
 934 * same way.  If !remap, a matching omap_gem_put_pages() call is not
 935 * required (and should not be made).
 936 */
 937int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
 938		bool remap)
 939{
 940	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 941	int ret = 0;
 942
 943	mutex_lock(&omap_obj->lock);
 944
 945	if (remap) {
 946		ret = omap_gem_attach_pages(obj);
 947		if (ret)
 948			goto unlock;
 949	}
 950
 951	if (!omap_obj->pages) {
 952		ret = -ENOMEM;
 953		goto unlock;
 954	}
 955
 956	*pages = omap_obj->pages;
 957
 958unlock:
 959	mutex_unlock(&omap_obj->lock);
 960
 961	return ret;
 962}
 963
 964/* release pages when DMA no longer being performed */
 965int omap_gem_put_pages(struct drm_gem_object *obj)
 966{
 967	/* do something here if we dynamically attach/detach pages.. at
 968	 * least they would no longer need to be pinned if everyone has
 969	 * released the pages..
 970	 */
 971	return 0;
 972}
 973
 974struct sg_table *omap_gem_get_sg(struct drm_gem_object *obj,
 975		enum dma_data_direction dir)
 976{
 977	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 978	dma_addr_t addr;
 979	struct sg_table *sgt;
 980	struct scatterlist *sg;
 981	unsigned int count, len, stride, i;
 982	int ret;
 983
 984	ret = omap_gem_pin(obj, &addr);
 985	if (ret)
 986		return ERR_PTR(ret);
 987
 988	mutex_lock(&omap_obj->lock);
 989
 990	sgt = omap_obj->sgt;
 991	if (sgt)
 992		goto out;
 993
 994	sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
 995	if (!sgt) {
 996		ret = -ENOMEM;
 997		goto err_unpin;
 998	}
 999
1000	if (addr) {
1001		if (omap_obj->flags & OMAP_BO_TILED_MASK) {
1002			enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
1003
1004			len = omap_obj->width << (int)fmt;
1005			count = omap_obj->height;
1006			stride = tiler_stride(fmt, 0);
1007		} else {
1008			len = obj->size;
1009			count = 1;
1010			stride = 0;
1011		}
1012	} else {
1013		count = obj->size >> PAGE_SHIFT;
1014	}
1015
1016	ret = sg_alloc_table(sgt, count, GFP_KERNEL);
1017	if (ret)
1018		goto err_free;
1019
1020	/* this must be after omap_gem_pin() to ensure we have pages attached */
1021	omap_gem_dma_sync_buffer(obj, dir);
1022
1023	if (addr) {
1024		for_each_sg(sgt->sgl, sg, count, i) {
1025			sg_set_page(sg, phys_to_page(addr), len,
1026				offset_in_page(addr));
1027			sg_dma_address(sg) = addr;
1028			sg_dma_len(sg) = len;
1029
1030			addr += stride;
1031		}
1032	} else {
1033		for_each_sg(sgt->sgl, sg, count, i) {
1034			sg_set_page(sg, omap_obj->pages[i], PAGE_SIZE, 0);
1035			sg_dma_address(sg) = omap_obj->dma_addrs[i];
1036			sg_dma_len(sg) =  PAGE_SIZE;
1037		}
1038	}
1039
1040	omap_obj->sgt = sgt;
1041out:
1042	mutex_unlock(&omap_obj->lock);
1043	return sgt;
1044
1045err_free:
1046	kfree(sgt);
1047err_unpin:
1048	mutex_unlock(&omap_obj->lock);
1049	omap_gem_unpin(obj);
1050	return ERR_PTR(ret);
1051}
1052
1053void omap_gem_put_sg(struct drm_gem_object *obj, struct sg_table *sgt)
1054{
1055	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1056
1057	if (WARN_ON(omap_obj->sgt != sgt))
1058		return;
1059
1060	omap_gem_unpin(obj);
1061}
1062
1063#ifdef CONFIG_DRM_FBDEV_EMULATION
1064/*
1065 * Get kernel virtual address for CPU access.. this more or less only
1066 * exists for omap_fbdev.
1067 */
1068void *omap_gem_vaddr(struct drm_gem_object *obj)
1069{
1070	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1071	void *vaddr;
1072	int ret;
1073
1074	mutex_lock(&omap_obj->lock);
1075
1076	if (!omap_obj->vaddr) {
1077		ret = omap_gem_attach_pages(obj);
1078		if (ret) {
1079			vaddr = ERR_PTR(ret);
1080			goto unlock;
1081		}
1082
1083		omap_obj->vaddr = vmap(omap_obj->pages, obj->size >> PAGE_SHIFT,
1084				VM_MAP, pgprot_writecombine(PAGE_KERNEL));
1085	}
1086
1087	vaddr = omap_obj->vaddr;
1088
1089unlock:
1090	mutex_unlock(&omap_obj->lock);
1091	return vaddr;
1092}
1093#endif
1094
1095/* -----------------------------------------------------------------------------
1096 * Power Management
1097 */
1098
1099#ifdef CONFIG_PM
1100/* re-pin objects in DMM in resume path: */
1101int omap_gem_resume(struct drm_device *dev)
1102{
1103	struct omap_drm_private *priv = dev->dev_private;
 
1104	struct omap_gem_object *omap_obj;
1105	int ret = 0;
1106
1107	mutex_lock(&priv->list_lock);
1108	list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
1109		if (omap_obj->block) {
1110			struct drm_gem_object *obj = &omap_obj->base;
1111			u32 npages = obj->size >> PAGE_SHIFT;
1112
1113			WARN_ON(!omap_obj->pages);  /* this can't happen */
1114			ret = tiler_pin(omap_obj->block,
1115					omap_obj->pages, npages,
1116					omap_obj->roll, true);
1117			if (ret) {
1118				dev_err(dev->dev, "could not repin: %d\n", ret);
1119				goto done;
1120			}
1121		}
1122	}
1123
1124done:
1125	mutex_unlock(&priv->list_lock);
1126	return ret;
1127}
1128#endif
1129
1130/* -----------------------------------------------------------------------------
1131 * DebugFS
1132 */
1133
1134#ifdef CONFIG_DEBUG_FS
1135void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1136{
1137	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1138	u64 off;
1139
1140	off = drm_vma_node_start(&obj->vma_node);
1141
1142	mutex_lock(&omap_obj->lock);
1143
1144	seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1145			omap_obj->flags, obj->name, kref_read(&obj->refcount),
1146			off, &omap_obj->dma_addr,
1147			refcount_read(&omap_obj->pin_cnt),
1148			omap_obj->vaddr, omap_obj->roll);
1149
1150	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
1151		seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1152		if (omap_obj->block) {
1153			struct tcm_area *area = &omap_obj->block->area;
1154			seq_printf(m, " (%dx%d, %dx%d)",
1155					area->p0.x, area->p0.y,
1156					area->p1.x, area->p1.y);
1157		}
1158	} else {
1159		seq_printf(m, " %zu", obj->size);
1160	}
1161
1162	mutex_unlock(&omap_obj->lock);
1163
1164	seq_printf(m, "\n");
1165}
1166
1167void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1168{
1169	struct omap_gem_object *omap_obj;
1170	int count = 0;
1171	size_t size = 0;
1172
1173	list_for_each_entry(omap_obj, list, mm_list) {
1174		struct drm_gem_object *obj = &omap_obj->base;
1175		seq_printf(m, "   ");
1176		omap_gem_describe(obj, m);
1177		count++;
1178		size += obj->size;
1179	}
1180
1181	seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1182}
1183#endif
1184
1185/* -----------------------------------------------------------------------------
1186 * Constructor & Destructor
1187 */
1188
1189static void omap_gem_free_object(struct drm_gem_object *obj)
1190{
1191	struct drm_device *dev = obj->dev;
1192	struct omap_drm_private *priv = dev->dev_private;
1193	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 
 
 
 
1194
1195	omap_gem_evict(obj);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1196
1197	mutex_lock(&priv->list_lock);
1198	list_del(&omap_obj->mm_list);
1199	mutex_unlock(&priv->list_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1200
1201	/*
1202	 * We own the sole reference to the object at this point, but to keep
1203	 * lockdep happy, we must still take the omap_obj_lock to call
1204	 * omap_gem_detach_pages(). This should hardly make any difference as
1205	 * there can't be any lock contention.
1206	 */
1207	mutex_lock(&omap_obj->lock);
1208
1209	/* The object should not be pinned. */
1210	WARN_ON(refcount_read(&omap_obj->pin_cnt) > 0);
1211
1212	if (omap_obj->pages) {
1213		if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1214			kfree(omap_obj->pages);
1215		else
1216			omap_gem_detach_pages(obj);
 
1217	}
1218
1219	if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1220		dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1221			    omap_obj->dma_addr);
1222	} else if (omap_obj->vaddr) {
1223		vunmap(omap_obj->vaddr);
1224	} else if (obj->import_attach) {
1225		drm_prime_gem_destroy(obj, omap_obj->sgt);
 
 
 
 
1226	}
1227
1228	mutex_unlock(&omap_obj->lock);
1229
1230	drm_gem_object_release(obj);
1231
1232	mutex_destroy(&omap_obj->lock);
1233
1234	kfree(omap_obj);
1235}
1236
1237static bool omap_gem_validate_flags(struct drm_device *dev, u32 flags)
 
 
 
 
 
 
 
1238{
1239	struct omap_drm_private *priv = dev->dev_private;
 
 
 
1240
1241	switch (flags & OMAP_BO_CACHE_MASK) {
1242	case OMAP_BO_CACHED:
1243	case OMAP_BO_WC:
1244	case OMAP_BO_CACHE_MASK:
1245		break;
1246
1247	default:
1248		return false;
1249	}
 
1250
1251	if (flags & OMAP_BO_TILED_MASK) {
1252		if (!priv->usergart)
1253			return false;
1254
1255		switch (flags & OMAP_BO_TILED_MASK) {
1256		case OMAP_BO_TILED_8:
1257		case OMAP_BO_TILED_16:
1258		case OMAP_BO_TILED_32:
1259			break;
 
1260
1261		default:
1262			return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1263		}
1264	}
1265
1266	return true;
 
 
 
1267}
1268
1269static const struct vm_operations_struct omap_gem_vm_ops = {
1270	.fault = omap_gem_fault,
1271	.open = drm_gem_vm_open,
1272	.close = drm_gem_vm_close,
1273};
1274
1275static const struct drm_gem_object_funcs omap_gem_object_funcs = {
1276	.free = omap_gem_free_object,
1277	.export = omap_gem_prime_export,
1278	.mmap = omap_gem_object_mmap,
1279	.vm_ops = &omap_gem_vm_ops,
1280};
1281
1282/* GEM buffer object constructor */
1283struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1284		union omap_gem_size gsize, u32 flags)
1285{
1286	struct omap_drm_private *priv = dev->dev_private;
1287	struct omap_gem_object *omap_obj;
1288	struct drm_gem_object *obj;
1289	struct address_space *mapping;
1290	size_t size;
1291	int ret;
1292
1293	if (!omap_gem_validate_flags(dev, flags))
1294		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
1295
1296	/* Validate the flags and compute the memory and cache flags. */
1297	if (flags & OMAP_BO_TILED_MASK) {
1298		/*
1299		 * Tiled buffers are always shmem paged backed. When they are
1300		 * scanned out, they are remapped into DMM/TILER.
1301		 */
1302		flags |= OMAP_BO_MEM_SHMEM;
1303
1304		/*
1305		 * Currently don't allow cached buffers. There is some caching
1306		 * stuff that needs to be handled better.
1307		 */
1308		flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1309		flags |= tiler_get_cpu_cache_flags();
1310	} else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1311		/*
1312		 * If we don't have DMM, we must allocate scanout buffers
1313		 * from contiguous DMA memory.
1314		 */
1315		flags |= OMAP_BO_MEM_DMA_API;
1316	} else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1317		/*
1318		 * All other buffers not backed by dma_buf are shmem-backed.
1319		 */
1320		flags |= OMAP_BO_MEM_SHMEM;
1321	}
1322
1323	/* Allocate the initialize the OMAP GEM object. */
1324	omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1325	if (!omap_obj)
1326		return NULL;
1327
1328	obj = &omap_obj->base;
1329	omap_obj->flags = flags;
1330	mutex_init(&omap_obj->lock);
1331
1332	if (flags & OMAP_BO_TILED_MASK) {
1333		/*
1334		 * For tiled buffers align dimensions to slot boundaries and
1335		 * calculate size based on aligned dimensions.
1336		 */
1337		tiler_align(gem2fmt(flags), &gsize.tiled.width,
1338			    &gsize.tiled.height);
1339
1340		size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1341				  gsize.tiled.height);
1342
1343		omap_obj->width = gsize.tiled.width;
1344		omap_obj->height = gsize.tiled.height;
1345	} else {
1346		size = PAGE_ALIGN(gsize.bytes);
1347	}
1348
1349	obj->funcs = &omap_gem_object_funcs;
 
1350
1351	/* Initialize the GEM object. */
1352	if (!(flags & OMAP_BO_MEM_SHMEM)) {
1353		drm_gem_private_object_init(dev, obj, size);
1354	} else {
1355		ret = drm_gem_object_init(dev, obj, size);
1356		if (ret)
1357			goto err_free;
 
 
1358
1359		mapping = obj->filp->f_mapping;
1360		mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1361	}
1362
1363	/* Allocate memory if needed. */
1364	if (flags & OMAP_BO_MEM_DMA_API) {
1365		omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1366					       &omap_obj->dma_addr,
1367					       GFP_KERNEL);
1368		if (!omap_obj->vaddr)
1369			goto err_release;
 
 
 
 
 
 
 
 
 
 
 
1370	}
1371
1372	mutex_lock(&priv->list_lock);
1373	list_add(&omap_obj->mm_list, &priv->obj_list);
1374	mutex_unlock(&priv->list_lock);
1375
1376	return obj;
1377
1378err_release:
1379	drm_gem_object_release(obj);
1380err_free:
1381	kfree(omap_obj);
1382	return NULL;
1383}
1384
1385struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1386					   struct sg_table *sgt)
 
 
1387{
1388	struct omap_drm_private *priv = dev->dev_private;
1389	struct omap_gem_object *omap_obj;
1390	struct drm_gem_object *obj;
1391	union omap_gem_size gsize;
1392
1393	/* Without a DMM only physically contiguous buffers can be supported. */
1394	if (!omap_gem_sgt_is_contiguous(sgt, size) && !priv->has_dmm)
1395		return ERR_PTR(-EINVAL);
1396
1397	gsize.bytes = PAGE_ALIGN(size);
1398	obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1399	if (!obj)
1400		return ERR_PTR(-ENOMEM);
1401
1402	omap_obj = to_omap_bo(obj);
1403
1404	mutex_lock(&omap_obj->lock);
1405
1406	omap_obj->sgt = sgt;
1407
1408	if (omap_gem_sgt_is_contiguous(sgt, size)) {
1409		omap_obj->dma_addr = sg_dma_address(sgt->sgl);
1410	} else {
1411		/* Create pages list from sgt */
1412		struct page **pages;
1413		unsigned int npages;
1414		unsigned int ret;
1415
1416		npages = DIV_ROUND_UP(size, PAGE_SIZE);
1417		pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1418		if (!pages) {
1419			omap_gem_free_object(obj);
1420			obj = ERR_PTR(-ENOMEM);
1421			goto done;
1422		}
1423
1424		omap_obj->pages = pages;
1425		ret = drm_prime_sg_to_page_array(sgt, pages, npages);
1426		if (ret) {
1427			omap_gem_free_object(obj);
1428			obj = ERR_PTR(-ENOMEM);
1429			goto done;
1430		}
1431	}
1432
1433done:
1434	mutex_unlock(&omap_obj->lock);
1435	return obj;
 
 
 
 
1436}
1437
1438/* convenience method to construct a GEM buffer object, and userspace handle */
1439int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1440		union omap_gem_size gsize, u32 flags, u32 *handle)
1441{
1442	struct drm_gem_object *obj;
1443	int ret;
1444
1445	obj = omap_gem_new(dev, gsize, flags);
1446	if (!obj)
1447		return -ENOMEM;
1448
1449	ret = drm_gem_handle_create(file, obj, handle);
1450	if (ret) {
1451		omap_gem_free_object(obj);
 
1452		return ret;
1453	}
1454
1455	/* drop reference from allocate - handle holds it now */
1456	drm_gem_object_put(obj);
1457
1458	return 0;
1459}
1460
1461/* -----------------------------------------------------------------------------
1462 * Init & Cleanup
1463 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1464
1465/* If DMM is used, we need to set some stuff up.. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1466void omap_gem_init(struct drm_device *dev)
1467{
1468	struct omap_drm_private *priv = dev->dev_private;
1469	struct omap_drm_usergart *usergart;
1470	const enum tiler_fmt fmts[] = {
1471			TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1472	};
1473	int i, j;
1474
1475	if (!dmm_is_available()) {
1476		/* DMM only supported on OMAP4 and later, so this isn't fatal */
1477		dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1478		return;
1479	}
1480
1481	usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1482	if (!usergart)
1483		return;
1484
1485	/* reserve 4k aligned/wide regions for userspace mappings: */
1486	for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1487		u16 h = 1, w = PAGE_SIZE >> i;
1488
1489		tiler_align(fmts[i], &w, &h);
1490		/* note: since each region is 1 4kb page wide, and minimum
1491		 * number of rows, the height ends up being the same as the
1492		 * # of pages in the region
1493		 */
1494		usergart[i].height = h;
1495		usergart[i].height_shift = ilog2(h);
1496		usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1497		usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1498		for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1499			struct omap_drm_usergart_entry *entry;
1500			struct tiler_block *block;
1501
1502			entry = &usergart[i].entry[j];
1503			block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1504			if (IS_ERR(block)) {
1505				dev_err(dev->dev,
1506						"reserve failed: %d, %d, %ld\n",
1507						i, j, PTR_ERR(block));
1508				return;
1509			}
1510			entry->dma_addr = tiler_ssptr(block);
1511			entry->block = block;
1512
1513			DBG("%d:%d: %dx%d: dma_addr=%pad stride=%d", i, j, w, h,
1514					&entry->dma_addr,
1515					usergart[i].stride_pfn << PAGE_SHIFT);
1516		}
1517	}
1518
1519	priv->usergart = usergart;
1520	priv->has_dmm = true;
1521}
1522
1523void omap_gem_deinit(struct drm_device *dev)
1524{
1525	struct omap_drm_private *priv = dev->dev_private;
1526
1527	/* I believe we can rely on there being no more outstanding GEM
1528	 * objects which could depend on usergart/dmm at this point.
1529	 */
1530	kfree(priv->usergart);
1531}
v3.15
 
   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
  21#include <linux/spinlock.h>
  22#include <linux/shmem_fs.h>
  23#include <drm/drm_vma_manager.h>
  24
  25#include "omap_drv.h"
  26#include "omap_dmm_tiler.h"
  27
  28/* remove these once drm core helpers are merged */
  29struct page **_drm_gem_get_pages(struct drm_gem_object *obj, gfp_t gfpmask);
  30void _drm_gem_put_pages(struct drm_gem_object *obj, struct page **pages,
  31		bool dirty, bool accessed);
  32int _drm_gem_create_mmap_offset_size(struct drm_gem_object *obj, size_t size);
  33
  34/*
  35 * GEM buffer object implementation.
  36 */
  37
  38#define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
  39
  40/* note: we use upper 8 bits of flags for driver-internal flags: */
  41#define OMAP_BO_DMA			0x01000000	/* actually is physically contiguous */
  42#define OMAP_BO_EXT_SYNC	0x02000000	/* externally allocated sync object */
  43#define OMAP_BO_EXT_MEM		0x04000000	/* externally allocated memory */
  44
  45
  46struct omap_gem_object {
  47	struct drm_gem_object base;
  48
  49	struct list_head mm_list;
  50
  51	uint32_t flags;
  52
  53	/** width/height for tiled formats (rounded up to slot boundaries) */
  54	uint16_t width, height;
  55
  56	/** roll applied when mapping to DMM */
  57	uint32_t roll;
 
 
 
  58
  59	/**
  60	 * If buffer is allocated physically contiguous, the OMAP_BO_DMA flag
  61	 * is set and the paddr is valid.  Also if the buffer is remapped in
  62	 * TILER and paddr_cnt > 0, then paddr is valid.  But if you are using
  63	 * the physical address and OMAP_BO_DMA is not set, then you should
  64	 * be going thru omap_gem_{get,put}_paddr() to ensure the mapping is
  65	 * not removed from under your feet.
 
 
 
 
 
 
 
  66	 *
  67	 * Note that OMAP_BO_SCANOUT is a hint from userspace that DMA capable
  68	 * buffer is requested, but doesn't mean that it is.  Use the
  69	 * OMAP_BO_DMA flag to determine if the buffer has a DMA capable
  70	 * physical address.
 
  71	 */
  72	dma_addr_t paddr;
  73
  74	/**
  75	 * # of users of paddr
  76	 */
  77	uint32_t paddr_cnt;
 
 
 
 
 
 
  78
  79	/**
  80	 * tiler block used when buffer is remapped in DMM/TILER.
  81	 */
  82	struct tiler_block *block;
  83
  84	/**
  85	 * Array of backing pages, if allocated.  Note that pages are never
  86	 * allocated for buffers originally allocated from contiguous memory
  87	 */
  88	struct page **pages;
  89
  90	/** addresses corresponding to pages in above array */
  91	dma_addr_t *addrs;
  92
  93	/**
  94	 * Virtual address, if mapped.
  95	 */
  96	void *vaddr;
  97
  98	/**
  99	 * sync-object allocated on demand (if needed)
 100	 *
 101	 * Per-buffer sync-object for tracking pending and completed hw/dma
 102	 * read and write operations.  The layout in memory is dictated by
 103	 * the SGX firmware, which uses this information to stall the command
 104	 * stream if a surface is not ready yet.
 105	 *
 106	 * Note that when buffer is used by SGX, the sync-object needs to be
 107	 * allocated from a special heap of sync-objects.  This way many sync
 108	 * objects can be packed in a page, and not waste GPU virtual address
 109	 * space.  Because of this we have to have a omap_gem_set_sync_object()
 110	 * API to allow replacement of the syncobj after it has (potentially)
 111	 * already been allocated.  A bit ugly but I haven't thought of a
 112	 * better alternative.
 113	 */
 114	struct {
 115		uint32_t write_pending;
 116		uint32_t write_complete;
 117		uint32_t read_pending;
 118		uint32_t read_complete;
 119	} *sync;
 120};
 121
 122static int get_pages(struct drm_gem_object *obj, struct page ***pages);
 123static uint64_t mmap_offset(struct drm_gem_object *obj);
 124
 125/* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
 126 * not necessarily pinned in TILER all the time, and (b) when they are
 127 * they are not necessarily page aligned, we reserve one or more small
 128 * regions in each of the 2d containers to use as a user-GART where we
 129 * can create a second page-aligned mapping of parts of the buffer
 130 * being accessed from userspace.
 131 *
 132 * Note that we could optimize slightly when we know that multiple
 133 * tiler containers are backed by the same PAT.. but I'll leave that
 134 * for later..
 135 */
 136#define NUM_USERGART_ENTRIES 2
 137struct usergart_entry {
 138	struct tiler_block *block;	/* the reserved tiler block */
 139	dma_addr_t paddr;
 140	struct drm_gem_object *obj;	/* the current pinned obj */
 141	pgoff_t obj_pgoff;		/* page offset of obj currently
 142					   mapped in */
 143};
 144static struct {
 145	struct usergart_entry entry[NUM_USERGART_ENTRIES];
 
 146	int height;				/* height in rows */
 147	int height_shift;		/* ilog2(height in rows) */
 148	int slot_shift;			/* ilog2(width per slot) */
 149	int stride_pfn;			/* stride in pages */
 150	int last;				/* index of last used entry */
 151} *usergart;
 152
 153static void evict_entry(struct drm_gem_object *obj,
 154		enum tiler_fmt fmt, struct usergart_entry *entry)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 155{
 156	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 157	int n = usergart[fmt].height;
 
 158	size_t size = PAGE_SIZE * n;
 159	loff_t off = mmap_offset(obj) +
 160			(entry->obj_pgoff << PAGE_SHIFT);
 161	const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
 162
 163	if (m > 1) {
 164		int i;
 165		/* if stride > than PAGE_SIZE then sparse mapping: */
 166		for (i = n; i > 0; i--) {
 167			unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 168					    off, PAGE_SIZE, 1);
 169			off += PAGE_SIZE * m;
 170		}
 171	} else {
 172		unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 173				    off, size, 1);
 174	}
 175
 176	entry->obj = NULL;
 177}
 178
 179/* Evict a buffer from usergart, if it is mapped there */
 180static void evict(struct drm_gem_object *obj)
 181{
 182	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 
 183
 184	if (omap_obj->flags & OMAP_BO_TILED) {
 185		enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 186		int i;
 187
 188		if (!usergart)
 189			return;
 
 190
 191		for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
 192			struct usergart_entry *entry = &usergart[fmt].entry[i];
 193			if (entry->obj == obj)
 194				evict_entry(obj, fmt, entry);
 195		}
 196	}
 197}
 198
 199/* GEM objects can either be allocated from contiguous memory (in which
 200 * case obj->filp==NULL), or w/ shmem backing (obj->filp!=NULL).  But non
 201 * contiguous buffers can be remapped in TILER/DMM if they need to be
 202 * contiguous... but we don't do this all the time to reduce pressure
 203 * on TILER/DMM space when we know at allocation time that the buffer
 204 * will need to be scanned out.
 205 */
 206static inline bool is_shmem(struct drm_gem_object *obj)
 207{
 208	return obj->filp != NULL;
 209}
 210
 211/**
 212 * shmem buffers that are mapped cached can simulate coherency via using
 213 * page faulting to keep track of dirty pages
 214 */
 215static inline bool is_cached_coherent(struct drm_gem_object *obj)
 216{
 217	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 218	return is_shmem(obj) &&
 219		((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED);
 220}
 221
 222static DEFINE_SPINLOCK(sync_lock);
 223
 224/** ensure backing pages are allocated */
 225static int omap_gem_attach_pages(struct drm_gem_object *obj)
 226{
 227	struct drm_device *dev = obj->dev;
 228	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 229	struct page **pages;
 230	int npages = obj->size >> PAGE_SHIFT;
 231	int i, ret;
 232	dma_addr_t *addrs;
 233
 234	WARN_ON(omap_obj->pages);
 235
 236	/* TODO: __GFP_DMA32 .. but somehow GFP_HIGHMEM is coming from the
 237	 * mapping_gfp_mask(mapping) which conflicts w/ GFP_DMA32.. probably
 238	 * we actually want CMA memory for it all anyways..
 239	 */
 240	pages = drm_gem_get_pages(obj, GFP_KERNEL);
 
 
 
 241	if (IS_ERR(pages)) {
 242		dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
 243		return PTR_ERR(pages);
 244	}
 245
 246	/* for non-cached buffers, ensure the new pages are clean because
 247	 * DSS, GPU, etc. are not cache coherent:
 248	 */
 249	if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
 250		addrs = kmalloc(npages * sizeof(*addrs), GFP_KERNEL);
 251		if (!addrs) {
 252			ret = -ENOMEM;
 253			goto free_pages;
 254		}
 255
 256		for (i = 0; i < npages; i++) {
 257			addrs[i] = dma_map_page(dev->dev, pages[i],
 258					0, PAGE_SIZE, DMA_BIDIRECTIONAL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 259		}
 260	} else {
 261		addrs = kzalloc(npages * sizeof(*addrs), GFP_KERNEL);
 262		if (!addrs) {
 263			ret = -ENOMEM;
 264			goto free_pages;
 265		}
 266	}
 267
 268	omap_obj->addrs = addrs;
 269	omap_obj->pages = pages;
 270
 271	return 0;
 272
 
 
 273free_pages:
 274	drm_gem_put_pages(obj, pages, true, false);
 275
 276	return ret;
 277}
 278
 279/** release backing pages */
 280static void omap_gem_detach_pages(struct drm_gem_object *obj)
 281{
 282	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 
 
 
 
 283
 284	/* for non-cached buffers, ensure the new pages are clean because
 285	 * DSS, GPU, etc. are not cache coherent:
 286	 */
 287	if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
 288		int i, npages = obj->size >> PAGE_SHIFT;
 289		for (i = 0; i < npages; i++) {
 290			dma_unmap_page(obj->dev->dev, omap_obj->addrs[i],
 291					PAGE_SIZE, DMA_BIDIRECTIONAL);
 292		}
 293	}
 294
 295	kfree(omap_obj->addrs);
 296	omap_obj->addrs = NULL;
 297
 298	drm_gem_put_pages(obj, omap_obj->pages, true, false);
 299	omap_obj->pages = NULL;
 300}
 301
 302/* get buffer flags */
 303uint32_t omap_gem_flags(struct drm_gem_object *obj)
 304{
 305	return to_omap_bo(obj)->flags;
 306}
 307
 308/** get mmap offset */
 309static uint64_t mmap_offset(struct drm_gem_object *obj)
 310{
 311	struct drm_device *dev = obj->dev;
 312	int ret;
 313	size_t size;
 314
 315	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 316
 317	/* Make it mmapable */
 318	size = omap_gem_mmap_size(obj);
 319	ret = drm_gem_create_mmap_offset_size(obj, size);
 320	if (ret) {
 321		dev_err(dev->dev, "could not allocate mmap offset\n");
 322		return 0;
 323	}
 324
 325	return drm_vma_node_offset_addr(&obj->vma_node);
 326}
 327
 328uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj)
 329{
 330	uint64_t offset;
 331	mutex_lock(&obj->dev->struct_mutex);
 332	offset = mmap_offset(obj);
 333	mutex_unlock(&obj->dev->struct_mutex);
 334	return offset;
 335}
 336
 337/** get mmap size */
 338size_t omap_gem_mmap_size(struct drm_gem_object *obj)
 339{
 340	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 341	size_t size = obj->size;
 342
 343	if (omap_obj->flags & OMAP_BO_TILED) {
 344		/* for tiled buffers, the virtual size has stride rounded up
 345		 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
 346		 * 32kb later!).  But we don't back the entire buffer with
 347		 * pages, only the valid picture part.. so need to adjust for
 348		 * this in the size used to mmap and generate mmap offset
 349		 */
 350		size = tiler_vsize(gem2fmt(omap_obj->flags),
 351				omap_obj->width, omap_obj->height);
 352	}
 353
 354	return size;
 355}
 356
 357/* get tiled size, returns -EINVAL if not tiled buffer */
 358int omap_gem_tiled_size(struct drm_gem_object *obj, uint16_t *w, uint16_t *h)
 359{
 360	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 361	if (omap_obj->flags & OMAP_BO_TILED) {
 362		*w = omap_obj->width;
 363		*h = omap_obj->height;
 364		return 0;
 365	}
 366	return -EINVAL;
 367}
 368
 369/* Normal handling for the case of faulting in non-tiled buffers */
 370static int fault_1d(struct drm_gem_object *obj,
 371		struct vm_area_struct *vma, struct vm_fault *vmf)
 372{
 373	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 374	unsigned long pfn;
 375	pgoff_t pgoff;
 376
 377	/* We don't use vmf->pgoff since that has the fake offset: */
 378	pgoff = ((unsigned long)vmf->virtual_address -
 379			vma->vm_start) >> PAGE_SHIFT;
 380
 381	if (omap_obj->pages) {
 382		omap_gem_cpu_sync(obj, pgoff);
 383		pfn = page_to_pfn(omap_obj->pages[pgoff]);
 384	} else {
 385		BUG_ON(!(omap_obj->flags & OMAP_BO_DMA));
 386		pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff;
 387	}
 388
 389	VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
 390			pfn, pfn << PAGE_SHIFT);
 391
 392	return vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, pfn);
 
 393}
 394
 395/* Special handling for the case of faulting in 2d tiled buffers */
 396static int fault_2d(struct drm_gem_object *obj,
 397		struct vm_area_struct *vma, struct vm_fault *vmf)
 398{
 399	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 400	struct usergart_entry *entry;
 
 401	enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 402	struct page *pages[64];  /* XXX is this too much to have on stack? */
 403	unsigned long pfn;
 404	pgoff_t pgoff, base_pgoff;
 405	void __user *vaddr;
 406	int i, ret, slots;
 
 407
 408	/*
 409	 * Note the height of the slot is also equal to the number of pages
 410	 * that need to be mapped in to fill 4kb wide CPU page.  If the slot
 411	 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
 412	 */
 413	const int n = usergart[fmt].height;
 414	const int n_shift = usergart[fmt].height_shift;
 415
 416	/*
 417	 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
 418	 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
 419	 * into account in some of the math, so figure out virtual stride
 420	 * in pages
 421	 */
 422	const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
 423
 424	/* We don't use vmf->pgoff since that has the fake offset: */
 425	pgoff = ((unsigned long)vmf->virtual_address -
 426			vma->vm_start) >> PAGE_SHIFT;
 427
 428	/*
 429	 * Actual address we start mapping at is rounded down to previous slot
 430	 * boundary in the y direction:
 431	 */
 432	base_pgoff = round_down(pgoff, m << n_shift);
 433
 434	/* figure out buffer width in slots */
 435	slots = omap_obj->width >> usergart[fmt].slot_shift;
 436
 437	vaddr = vmf->virtual_address - ((pgoff - base_pgoff) << PAGE_SHIFT);
 438
 439	entry = &usergart[fmt].entry[usergart[fmt].last];
 440
 441	/* evict previous buffer using this usergart entry, if any: */
 442	if (entry->obj)
 443		evict_entry(entry->obj, fmt, entry);
 444
 445	entry->obj = obj;
 446	entry->obj_pgoff = base_pgoff;
 447
 448	/* now convert base_pgoff to phys offset from virt offset: */
 449	base_pgoff = (base_pgoff >> n_shift) * slots;
 450
 451	/* for wider-than 4k.. figure out which part of the slot-row we want: */
 452	if (m > 1) {
 453		int off = pgoff % m;
 454		entry->obj_pgoff += off;
 455		base_pgoff /= m;
 456		slots = min(slots - (off << n_shift), n);
 457		base_pgoff += off << n_shift;
 458		vaddr += off << PAGE_SHIFT;
 459	}
 460
 461	/*
 462	 * Map in pages. Beyond the valid pixel part of the buffer, we set
 463	 * pages[i] to NULL to get a dummy page mapped in.. if someone
 464	 * reads/writes it they will get random/undefined content, but at
 465	 * least it won't be corrupting whatever other random page used to
 466	 * be mapped in, or other undefined behavior.
 467	 */
 468	memcpy(pages, &omap_obj->pages[base_pgoff],
 469			sizeof(struct page *) * slots);
 470	memset(pages + slots, 0,
 471			sizeof(struct page *) * (n - slots));
 472
 473	ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
 474	if (ret) {
 475		dev_err(obj->dev->dev, "failed to pin: %d\n", ret);
 
 476		return ret;
 477	}
 478
 479	pfn = entry->paddr >> PAGE_SHIFT;
 480
 481	VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
 482			pfn, pfn << PAGE_SHIFT);
 483
 484	for (i = n; i > 0; i--) {
 485		vm_insert_mixed(vma, (unsigned long)vaddr, pfn);
 486		pfn += usergart[fmt].stride_pfn;
 
 
 
 487		vaddr += PAGE_SIZE * m;
 488	}
 489
 490	/* simple round-robin: */
 491	usergart[fmt].last = (usergart[fmt].last + 1) % NUM_USERGART_ENTRIES;
 
 492
 493	return 0;
 494}
 495
 496/**
 497 * omap_gem_fault		-	pagefault handler for GEM objects
 498 * @vma: the VMA of the GEM object
 499 * @vmf: fault detail
 500 *
 501 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
 502 * does most of the work for us including the actual map/unmap calls
 503 * but we need to do the actual page work.
 504 *
 505 * The VMA was set up by GEM. In doing so it also ensured that the
 506 * vma->vm_private_data points to the GEM object that is backing this
 507 * mapping.
 508 */
 509int omap_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 510{
 
 511	struct drm_gem_object *obj = vma->vm_private_data;
 512	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 513	struct drm_device *dev = obj->dev;
 514	struct page **pages;
 515	int ret;
 516
 517	/* Make sure we don't parallel update on a fault, nor move or remove
 518	 * something from beneath our feet
 519	 */
 520	mutex_lock(&dev->struct_mutex);
 521
 522	/* if a shmem backed object, make sure we have pages attached now */
 523	ret = get_pages(obj, &pages);
 524	if (ret)
 
 525		goto fail;
 
 526
 527	/* where should we do corresponding put_pages().. we are mapping
 528	 * the original page, rather than thru a GART, so we can't rely
 529	 * on eviction to trigger this.  But munmap() or all mappings should
 530	 * probably trigger put_pages()?
 531	 */
 532
 533	if (omap_obj->flags & OMAP_BO_TILED)
 534		ret = fault_2d(obj, vma, vmf);
 535	else
 536		ret = fault_1d(obj, vma, vmf);
 537
 538
 539fail:
 540	mutex_unlock(&dev->struct_mutex);
 541	switch (ret) {
 542	case 0:
 543	case -ERESTARTSYS:
 544	case -EINTR:
 545		return VM_FAULT_NOPAGE;
 546	case -ENOMEM:
 547		return VM_FAULT_OOM;
 548	default:
 549		return VM_FAULT_SIGBUS;
 550	}
 551}
 552
 553/** We override mainly to fix up some of the vm mapping flags.. */
 554int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
 555{
 556	int ret;
 557
 558	ret = drm_gem_mmap(filp, vma);
 559	if (ret) {
 560		DBG("mmap failed: %d", ret);
 561		return ret;
 562	}
 563
 564	return omap_gem_mmap_obj(vma->vm_private_data, vma);
 565}
 566
 567int omap_gem_mmap_obj(struct drm_gem_object *obj,
 568		struct vm_area_struct *vma)
 569{
 570	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 571
 572	vma->vm_flags &= ~VM_PFNMAP;
 573	vma->vm_flags |= VM_MIXEDMAP;
 574
 575	if (omap_obj->flags & OMAP_BO_WC) {
 576		vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
 577	} else if (omap_obj->flags & OMAP_BO_UNCACHED) {
 578		vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
 579	} else {
 580		/*
 581		 * We do have some private objects, at least for scanout buffers
 582		 * on hardware without DMM/TILER.  But these are allocated write-
 583		 * combine
 584		 */
 585		if (WARN_ON(!obj->filp))
 586			return -EINVAL;
 587
 588		/*
 589		 * Shunt off cached objs to shmem file so they have their own
 590		 * address_space (so unmap_mapping_range does what we want,
 591		 * in particular in the case of mmap'd dmabufs)
 592		 */
 593		fput(vma->vm_file);
 594		vma->vm_pgoff = 0;
 595		vma->vm_file  = get_file(obj->filp);
 596
 597		vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 598	}
 599
 
 
 600	return 0;
 601}
 602
 
 
 
 603
 604/**
 605 * omap_gem_dumb_create	-	create a dumb buffer
 606 * @drm_file: our client file
 607 * @dev: our device
 608 * @args: the requested arguments copied from userspace
 609 *
 610 * Allocate a buffer suitable for use for a frame buffer of the
 611 * form described by user space. Give userspace a handle by which
 612 * to reference it.
 613 */
 614int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
 615		struct drm_mode_create_dumb *args)
 616{
 617	union omap_gem_size gsize;
 618
 619	/* in case someone tries to feed us a completely bogus stride: */
 620	args->pitch = align_pitch(args->pitch, args->width, args->bpp);
 621	args->size = PAGE_ALIGN(args->pitch * args->height);
 622
 623	gsize = (union omap_gem_size){
 624		.bytes = args->size,
 625	};
 626
 627	return omap_gem_new_handle(dev, file, gsize,
 628			OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
 629}
 630
 631/**
 632 * omap_gem_dumb_map	-	buffer mapping for dumb interface
 633 * @file: our drm client file
 634 * @dev: drm device
 635 * @handle: GEM handle to the object (from dumb_create)
 
 636 *
 637 * Do the necessary setup to allow the mapping of the frame buffer
 638 * into user memory. We don't have to do much here at the moment.
 639 */
 640int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 641		uint32_t handle, uint64_t *offset)
 642{
 643	struct drm_gem_object *obj;
 644	int ret = 0;
 645
 646	/* GEM does all our handle to object mapping */
 647	obj = drm_gem_object_lookup(dev, file, handle);
 648	if (obj == NULL) {
 649		ret = -ENOENT;
 650		goto fail;
 651	}
 652
 653	*offset = omap_gem_mmap_offset(obj);
 654
 655	drm_gem_object_unreference_unlocked(obj);
 656
 657fail:
 658	return ret;
 659}
 660
 
 661/* Set scrolling position.  This allows us to implement fast scrolling
 662 * for console.
 663 *
 664 * Call only from non-atomic contexts.
 665 */
 666int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll)
 667{
 668	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 669	uint32_t npages = obj->size >> PAGE_SHIFT;
 670	int ret = 0;
 671
 672	if (roll > npages) {
 673		dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
 674		return -EINVAL;
 675	}
 676
 677	omap_obj->roll = roll;
 678
 679	mutex_lock(&obj->dev->struct_mutex);
 680
 681	/* if we aren't mapped yet, we don't need to do anything */
 682	if (omap_obj->block) {
 683		struct page **pages;
 684		ret = get_pages(obj, &pages);
 685		if (ret)
 686			goto fail;
 687		ret = tiler_pin(omap_obj->block, pages, npages, roll, true);
 
 
 688		if (ret)
 689			dev_err(obj->dev->dev, "could not repin: %d\n", ret);
 690	}
 691
 692fail:
 693	mutex_unlock(&obj->dev->struct_mutex);
 694
 695	return ret;
 696}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 697
 698/* Sync the buffer for CPU access.. note pages should already be
 699 * attached, ie. omap_gem_get_pages()
 700 */
 701void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff)
 702{
 703	struct drm_device *dev = obj->dev;
 704	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 705
 706	if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) {
 707		dma_unmap_page(dev->dev, omap_obj->addrs[pgoff],
 708				PAGE_SIZE, DMA_BIDIRECTIONAL);
 709		omap_obj->addrs[pgoff] = 0;
 
 
 
 710	}
 711}
 712
 713/* sync the buffer for DMA access */
 714void omap_gem_dma_sync(struct drm_gem_object *obj,
 715		enum dma_data_direction dir)
 716{
 717	struct drm_device *dev = obj->dev;
 718	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 
 
 
 719
 720	if (is_cached_coherent(obj)) {
 721		int i, npages = obj->size >> PAGE_SHIFT;
 722		struct page **pages = omap_obj->pages;
 723		bool dirty = false;
 724
 725		for (i = 0; i < npages; i++) {
 726			if (!omap_obj->addrs[i]) {
 727				omap_obj->addrs[i] = dma_map_page(dev->dev, pages[i], 0,
 728						PAGE_SIZE, DMA_BIDIRECTIONAL);
 729				dirty = true;
 
 
 
 
 
 730			}
 
 
 
 731		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 732
 733		if (dirty) {
 734			unmap_mapping_range(obj->filp->f_mapping, 0,
 735					omap_gem_mmap_size(obj), 1);
 736		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 737	}
 
 
 
 
 
 
 
 
 738}
 739
 740/* Get physical address for DMA.. if 'remap' is true, and the buffer is not
 741 * already contiguous, remap it to pin in physically contiguous memory.. (ie.
 742 * map in TILER)
 
 
 
 
 
 
 
 
 
 
 743 */
 744int omap_gem_get_paddr(struct drm_gem_object *obj,
 745		dma_addr_t *paddr, bool remap)
 746{
 747	struct omap_drm_private *priv = obj->dev->dev_private;
 748	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 749	int ret = 0;
 750
 751	mutex_lock(&obj->dev->struct_mutex);
 752
 753	if (remap && is_shmem(obj) && priv->has_dmm) {
 754		if (omap_obj->paddr_cnt == 0) {
 755			struct page **pages;
 756			uint32_t npages = obj->size >> PAGE_SHIFT;
 757			enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 758			struct tiler_block *block;
 759
 760			BUG_ON(omap_obj->block);
 761
 762			ret = get_pages(obj, &pages);
 763			if (ret)
 764				goto fail;
 765
 766			if (omap_obj->flags & OMAP_BO_TILED) {
 767				block = tiler_reserve_2d(fmt,
 768						omap_obj->width,
 769						omap_obj->height, 0);
 770			} else {
 771				block = tiler_reserve_1d(obj->size);
 772			}
 773
 774			if (IS_ERR(block)) {
 775				ret = PTR_ERR(block);
 776				dev_err(obj->dev->dev,
 777					"could not remap: %d (%d)\n", ret, fmt);
 778				goto fail;
 779			}
 780
 781			/* TODO: enable async refill.. */
 782			ret = tiler_pin(block, pages, npages,
 783					omap_obj->roll, true);
 784			if (ret) {
 785				tiler_release(block);
 786				dev_err(obj->dev->dev,
 787						"could not pin: %d\n", ret);
 788				goto fail;
 789			}
 790
 791			omap_obj->paddr = tiler_ssptr(block);
 792			omap_obj->block = block;
 793
 794			DBG("got paddr: %08x", omap_obj->paddr);
 795		}
 
 796
 797		omap_obj->paddr_cnt++;
 798
 799		*paddr = omap_obj->paddr;
 800	} else if (omap_obj->flags & OMAP_BO_DMA) {
 801		*paddr = omap_obj->paddr;
 802	} else {
 803		ret = -EINVAL;
 804		goto fail;
 805	}
 806
 807fail:
 808	mutex_unlock(&obj->dev->struct_mutex);
 809
 810	return ret;
 811}
 812
 813/* Release physical address, when DMA is no longer being performed.. this
 814 * could potentially unpin and unmap buffers from TILER
 
 
 
 815 */
 816int omap_gem_put_paddr(struct drm_gem_object *obj)
 817{
 
 818	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 819	int ret = 0;
 
 
 
 820
 821	mutex_lock(&obj->dev->struct_mutex);
 822	if (omap_obj->paddr_cnt > 0) {
 823		omap_obj->paddr_cnt--;
 824		if (omap_obj->paddr_cnt == 0) {
 
 
 
 
 
 825			ret = tiler_unpin(omap_obj->block);
 826			if (ret) {
 827				dev_err(obj->dev->dev,
 828					"could not unpin pages: %d\n", ret);
 829				goto fail;
 830			}
 831			ret = tiler_release(omap_obj->block);
 832			if (ret) {
 833				dev_err(obj->dev->dev,
 834					"could not release unmap: %d\n", ret);
 835			}
 
 836			omap_obj->block = NULL;
 837		}
 838	}
 839fail:
 840	mutex_unlock(&obj->dev->struct_mutex);
 841	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 842}
 843
 844/* Get rotated scanout address (only valid if already pinned), at the
 845 * specified orientation and x,y offset from top-left corner of buffer
 846 * (only valid for tiled 2d buffers)
 847 */
 848int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient,
 849		int x, int y, dma_addr_t *paddr)
 850{
 851	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 852	int ret = -EINVAL;
 853
 854	mutex_lock(&obj->dev->struct_mutex);
 855	if ((omap_obj->paddr_cnt > 0) && omap_obj->block &&
 856			(omap_obj->flags & OMAP_BO_TILED)) {
 857		*paddr = tiler_tsptr(omap_obj->block, orient, x, y);
 
 858		ret = 0;
 859	}
 860	mutex_unlock(&obj->dev->struct_mutex);
 
 
 861	return ret;
 862}
 863
 864/* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
 865int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient)
 866{
 867	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 868	int ret = -EINVAL;
 869	if (omap_obj->flags & OMAP_BO_TILED)
 870		ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
 871	return ret;
 872}
 873
 874/* acquire pages when needed (for example, for DMA where physically
 875 * contiguous buffer is not required
 876 */
 877static int get_pages(struct drm_gem_object *obj, struct page ***pages)
 878{
 879	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 880	int ret = 0;
 881
 882	if (is_shmem(obj) && !omap_obj->pages) {
 883		ret = omap_gem_attach_pages(obj);
 884		if (ret) {
 885			dev_err(obj->dev->dev, "could not attach pages\n");
 886			return ret;
 887		}
 888	}
 889
 890	/* TODO: even phys-contig.. we should have a list of pages? */
 891	*pages = omap_obj->pages;
 892
 893	return 0;
 894}
 895
 896/* if !remap, and we don't have pages backing, then fail, rather than
 897 * increasing the pin count (which we don't really do yet anyways,
 898 * because we don't support swapping pages back out).  And 'remap'
 899 * might not be quite the right name, but I wanted to keep it working
 900 * similarly to omap_gem_get_paddr().  Note though that mutex is not
 901 * aquired if !remap (because this can be called in atomic ctxt),
 902 * but probably omap_gem_get_paddr() should be changed to work in the
 903 * same way.  If !remap, a matching omap_gem_put_pages() call is not
 904 * required (and should not be made).
 905 */
 906int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
 907		bool remap)
 908{
 909	int ret;
 910	if (!remap) {
 911		struct omap_gem_object *omap_obj = to_omap_bo(obj);
 912		if (!omap_obj->pages)
 913			return -ENOMEM;
 914		*pages = omap_obj->pages;
 915		return 0;
 
 
 
 
 
 
 
 916	}
 917	mutex_lock(&obj->dev->struct_mutex);
 918	ret = get_pages(obj, pages);
 919	mutex_unlock(&obj->dev->struct_mutex);
 
 
 
 920	return ret;
 921}
 922
 923/* release pages when DMA no longer being performed */
 924int omap_gem_put_pages(struct drm_gem_object *obj)
 925{
 926	/* do something here if we dynamically attach/detach pages.. at
 927	 * least they would no longer need to be pinned if everyone has
 928	 * released the pages..
 929	 */
 930	return 0;
 931}
 932
 933/* Get kernel virtual address for CPU access.. this more or less only
 934 * exists for omap_fbdev.  This should be called with struct_mutex
 935 * held.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 936 */
 937void *omap_gem_vaddr(struct drm_gem_object *obj)
 938{
 939	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 940	WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
 
 
 
 
 941	if (!omap_obj->vaddr) {
 942		struct page **pages;
 943		int ret = get_pages(obj, &pages);
 944		if (ret)
 945			return ERR_PTR(ret);
 946		omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
 
 
 947				VM_MAP, pgprot_writecombine(PAGE_KERNEL));
 948	}
 949	return omap_obj->vaddr;
 
 
 
 
 
 950}
 
 
 
 
 
 951
 952#ifdef CONFIG_PM
 953/* re-pin objects in DMM in resume path: */
 954int omap_gem_resume(struct device *dev)
 955{
 956	struct drm_device *drm_dev = dev_get_drvdata(dev);
 957	struct omap_drm_private *priv = drm_dev->dev_private;
 958	struct omap_gem_object *omap_obj;
 959	int ret = 0;
 960
 
 961	list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
 962		if (omap_obj->block) {
 963			struct drm_gem_object *obj = &omap_obj->base;
 964			uint32_t npages = obj->size >> PAGE_SHIFT;
 
 965			WARN_ON(!omap_obj->pages);  /* this can't happen */
 966			ret = tiler_pin(omap_obj->block,
 967					omap_obj->pages, npages,
 968					omap_obj->roll, true);
 969			if (ret) {
 970				dev_err(dev, "could not repin: %d\n", ret);
 971				return ret;
 972			}
 973		}
 974	}
 975
 976	return 0;
 
 
 977}
 978#endif
 979
 
 
 
 
 980#ifdef CONFIG_DEBUG_FS
 981void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
 982{
 983	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 984	uint64_t off;
 985
 986	off = drm_vma_node_start(&obj->vma_node);
 987
 988	seq_printf(m, "%08x: %2d (%2d) %08llx %08Zx (%2d) %p %4d",
 989			omap_obj->flags, obj->name, obj->refcount.refcount.counter,
 990			off, omap_obj->paddr, omap_obj->paddr_cnt,
 
 
 
 991			omap_obj->vaddr, omap_obj->roll);
 992
 993	if (omap_obj->flags & OMAP_BO_TILED) {
 994		seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
 995		if (omap_obj->block) {
 996			struct tcm_area *area = &omap_obj->block->area;
 997			seq_printf(m, " (%dx%d, %dx%d)",
 998					area->p0.x, area->p0.y,
 999					area->p1.x, area->p1.y);
1000		}
1001	} else {
1002		seq_printf(m, " %d", obj->size);
1003	}
1004
 
 
1005	seq_printf(m, "\n");
1006}
1007
1008void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1009{
1010	struct omap_gem_object *omap_obj;
1011	int count = 0;
1012	size_t size = 0;
1013
1014	list_for_each_entry(omap_obj, list, mm_list) {
1015		struct drm_gem_object *obj = &omap_obj->base;
1016		seq_printf(m, "   ");
1017		omap_gem_describe(obj, m);
1018		count++;
1019		size += obj->size;
1020	}
1021
1022	seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1023}
1024#endif
1025
1026/* Buffer Synchronization:
 
1027 */
1028
1029struct omap_gem_sync_waiter {
1030	struct list_head list;
1031	struct omap_gem_object *omap_obj;
1032	enum omap_gem_op op;
1033	uint32_t read_target, write_target;
1034	/* notify called w/ sync_lock held */
1035	void (*notify)(void *arg);
1036	void *arg;
1037};
1038
1039/* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1040 * the read and/or write target count is achieved which can call a user
1041 * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1042 * cpu access), etc.
1043 */
1044static LIST_HEAD(waiters);
1045
1046static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1047{
1048	struct omap_gem_object *omap_obj = waiter->omap_obj;
1049	if ((waiter->op & OMAP_GEM_READ) &&
1050			(omap_obj->sync->write_complete < waiter->write_target))
1051		return true;
1052	if ((waiter->op & OMAP_GEM_WRITE) &&
1053			(omap_obj->sync->read_complete < waiter->read_target))
1054		return true;
1055	return false;
1056}
1057
1058/* macro for sync debug.. */
1059#define SYNCDBG 0
1060#define SYNC(fmt, ...) do { if (SYNCDBG) \
1061		printk(KERN_ERR "%s:%d: "fmt"\n", \
1062				__func__, __LINE__, ##__VA_ARGS__); \
1063	} while (0)
1064
1065
1066static void sync_op_update(void)
1067{
1068	struct omap_gem_sync_waiter *waiter, *n;
1069	list_for_each_entry_safe(waiter, n, &waiters, list) {
1070		if (!is_waiting(waiter)) {
1071			list_del(&waiter->list);
1072			SYNC("notify: %p", waiter);
1073			waiter->notify(waiter->arg);
1074			kfree(waiter);
1075		}
1076	}
1077}
1078
1079static inline int sync_op(struct drm_gem_object *obj,
1080		enum omap_gem_op op, bool start)
1081{
1082	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1083	int ret = 0;
 
 
1084
1085	spin_lock(&sync_lock);
 
1086
1087	if (!omap_obj->sync) {
1088		omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1089		if (!omap_obj->sync) {
1090			ret = -ENOMEM;
1091			goto unlock;
1092		}
1093	}
1094
1095	if (start) {
1096		if (op & OMAP_GEM_READ)
1097			omap_obj->sync->read_pending++;
1098		if (op & OMAP_GEM_WRITE)
1099			omap_obj->sync->write_pending++;
1100	} else {
1101		if (op & OMAP_GEM_READ)
1102			omap_obj->sync->read_complete++;
1103		if (op & OMAP_GEM_WRITE)
1104			omap_obj->sync->write_complete++;
1105		sync_op_update();
1106	}
1107
1108unlock:
1109	spin_unlock(&sync_lock);
 
 
 
1110
1111	return ret;
1112}
1113
1114/* it is a bit lame to handle updates in this sort of polling way, but
1115 * in case of PVR, the GPU can directly update read/write complete
1116 * values, and not really tell us which ones it updated.. this also
1117 * means that sync_lock is not quite sufficient.  So we'll need to
1118 * do something a bit better when it comes time to add support for
1119 * separate 2d hw..
1120 */
1121void omap_gem_op_update(void)
1122{
1123	spin_lock(&sync_lock);
1124	sync_op_update();
1125	spin_unlock(&sync_lock);
1126}
1127
1128/* mark the start of read and/or write operation */
1129int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1130{
1131	return sync_op(obj, op, true);
1132}
1133
1134int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1135{
1136	return sync_op(obj, op, false);
1137}
1138
1139static DECLARE_WAIT_QUEUE_HEAD(sync_event);
 
 
1140
1141static void sync_notify(void *arg)
1142{
1143	struct task_struct **waiter_task = arg;
1144	*waiter_task = NULL;
1145	wake_up_all(&sync_event);
1146}
1147
1148int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1149{
1150	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1151	int ret = 0;
1152	if (omap_obj->sync) {
1153		struct task_struct *waiter_task = current;
1154		struct omap_gem_sync_waiter *waiter =
1155				kzalloc(sizeof(*waiter), GFP_KERNEL);
1156
1157		if (!waiter)
1158			return -ENOMEM;
1159
1160		waiter->omap_obj = omap_obj;
1161		waiter->op = op;
1162		waiter->read_target = omap_obj->sync->read_pending;
1163		waiter->write_target = omap_obj->sync->write_pending;
1164		waiter->notify = sync_notify;
1165		waiter->arg = &waiter_task;
1166
1167		spin_lock(&sync_lock);
1168		if (is_waiting(waiter)) {
1169			SYNC("waited: %p", waiter);
1170			list_add_tail(&waiter->list, &waiters);
1171			spin_unlock(&sync_lock);
1172			ret = wait_event_interruptible(sync_event,
1173					(waiter_task == NULL));
1174			spin_lock(&sync_lock);
1175			if (waiter_task) {
1176				SYNC("interrupted: %p", waiter);
1177				/* we were interrupted */
1178				list_del(&waiter->list);
1179				waiter_task = NULL;
1180			} else {
1181				/* freed in sync_op_update() */
1182				waiter = NULL;
1183			}
1184		}
1185		spin_unlock(&sync_lock);
1186
1187		if (waiter)
1188			kfree(waiter);
1189	}
1190	return ret;
1191}
1192
1193/* call fxn(arg), either synchronously or asynchronously if the op
1194 * is currently blocked..  fxn() can be called from any context
1195 *
1196 * (TODO for now fxn is called back from whichever context calls
1197 * omap_gem_op_update().. but this could be better defined later
1198 * if needed)
1199 *
1200 * TODO more code in common w/ _sync()..
1201 */
1202int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1203		void (*fxn)(void *arg), void *arg)
 
 
 
 
 
1204{
1205	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1206	if (omap_obj->sync) {
1207		struct omap_gem_sync_waiter *waiter =
1208				kzalloc(sizeof(*waiter), GFP_ATOMIC);
1209
1210		if (!waiter)
1211			return -ENOMEM;
1212
1213		waiter->omap_obj = omap_obj;
1214		waiter->op = op;
1215		waiter->read_target = omap_obj->sync->read_pending;
1216		waiter->write_target = omap_obj->sync->write_pending;
1217		waiter->notify = fxn;
1218		waiter->arg = arg;
1219
1220		spin_lock(&sync_lock);
1221		if (is_waiting(waiter)) {
1222			SYNC("waited: %p", waiter);
1223			list_add_tail(&waiter->list, &waiters);
1224			spin_unlock(&sync_lock);
1225			return 0;
1226		}
1227
1228		spin_unlock(&sync_lock);
 
 
 
 
 
 
1229
1230		kfree(waiter);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1231	}
1232
1233	/* no waiting.. */
1234	fxn(arg);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1235
1236	return 0;
1237}
1238
1239/* special API so PVR can update the buffer to use a sync-object allocated
1240 * from it's sync-obj heap.  Only used for a newly allocated (from PVR's
1241 * perspective) sync-object, so we overwrite the new syncobj w/ values
1242 * from the already allocated syncobj (if there is one)
1243 */
1244int omap_gem_set_sync_object(struct drm_gem_object *obj, void *syncobj)
1245{
1246	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1247	int ret = 0;
1248
1249	spin_lock(&sync_lock);
 
 
1250
1251	if ((omap_obj->flags & OMAP_BO_EXT_SYNC) && !syncobj) {
1252		/* clearing a previously set syncobj */
1253		syncobj = kmemdup(omap_obj->sync, sizeof(*omap_obj->sync),
1254				  GFP_ATOMIC);
1255		if (!syncobj) {
1256			ret = -ENOMEM;
1257			goto unlock;
1258		}
1259		omap_obj->flags &= ~OMAP_BO_EXT_SYNC;
1260		omap_obj->sync = syncobj;
1261	} else if (syncobj && !(omap_obj->flags & OMAP_BO_EXT_SYNC)) {
1262		/* replacing an existing syncobj */
1263		if (omap_obj->sync) {
1264			memcpy(syncobj, omap_obj->sync, sizeof(*omap_obj->sync));
1265			kfree(omap_obj->sync);
1266		}
1267		omap_obj->flags |= OMAP_BO_EXT_SYNC;
1268		omap_obj->sync = syncobj;
1269	}
1270
1271unlock:
1272	spin_unlock(&sync_lock);
1273	return ret;
 
 
 
 
 
 
 
 
1274}
1275
1276/* don't call directly.. called from GEM core when it is time to actually
1277 * free the object..
1278 */
1279void omap_gem_free_object(struct drm_gem_object *obj)
1280{
1281	struct drm_device *dev = obj->dev;
1282	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 
 
 
 
 
 
1283
1284	evict(obj);
 
 
 
1285
1286	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1287
1288	list_del(&omap_obj->mm_list);
1289
1290	drm_gem_free_mmap_offset(obj);
1291
1292	/* this means the object is still pinned.. which really should
1293	 * not happen.  I think..
1294	 */
1295	WARN_ON(omap_obj->paddr_cnt > 0);
 
 
 
1296
1297	/* don't free externally allocated backing memory */
1298	if (!(omap_obj->flags & OMAP_BO_EXT_MEM)) {
1299		if (omap_obj->pages)
1300			omap_gem_detach_pages(obj);
 
 
 
1301
1302		if (!is_shmem(obj)) {
1303			dma_free_writecombine(dev->dev, obj->size,
1304					omap_obj->vaddr, omap_obj->paddr);
1305		} else if (omap_obj->vaddr) {
1306			vunmap(omap_obj->vaddr);
 
1307		}
1308	}
1309
1310	/* don't free externally allocated syncobj */
1311	if (!(omap_obj->flags & OMAP_BO_EXT_SYNC))
1312		kfree(omap_obj->sync);
1313
1314	drm_gem_object_release(obj);
1315
1316	kfree(obj);
1317}
1318
1319/* convenience method to construct a GEM buffer object, and userspace handle */
1320int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1321		union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1322{
1323	struct drm_gem_object *obj;
1324	int ret;
1325
1326	obj = omap_gem_new(dev, gsize, flags);
1327	if (!obj)
1328		return -ENOMEM;
1329
1330	ret = drm_gem_handle_create(file, obj, handle);
1331	if (ret) {
1332		drm_gem_object_release(obj);
1333		kfree(obj); /* TODO isn't there a dtor to call? just copying i915 */
1334		return ret;
1335	}
1336
1337	/* drop reference from allocate - handle holds it now */
1338	drm_gem_object_unreference_unlocked(obj);
1339
1340	return 0;
1341}
1342
1343/* GEM buffer object constructor */
1344struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1345		union omap_gem_size gsize, uint32_t flags)
1346{
1347	struct omap_drm_private *priv = dev->dev_private;
1348	struct omap_gem_object *omap_obj;
1349	struct drm_gem_object *obj = NULL;
1350	size_t size;
1351	int ret;
1352
1353	if (flags & OMAP_BO_TILED) {
1354		if (!usergart) {
1355			dev_err(dev->dev, "Tiled buffers require DMM\n");
1356			goto fail;
1357		}
1358
1359		/* tiled buffers are always shmem paged backed.. when they are
1360		 * scanned out, they are remapped into DMM/TILER
1361		 */
1362		flags &= ~OMAP_BO_SCANOUT;
1363
1364		/* currently don't allow cached buffers.. there is some caching
1365		 * stuff that needs to be handled better
1366		 */
1367		flags &= ~(OMAP_BO_CACHED|OMAP_BO_UNCACHED);
1368		flags |= OMAP_BO_WC;
1369
1370		/* align dimensions to slot boundaries... */
1371		tiler_align(gem2fmt(flags),
1372				&gsize.tiled.width, &gsize.tiled.height);
1373
1374		/* ...and calculate size based on aligned dimensions */
1375		size = tiler_size(gem2fmt(flags),
1376				gsize.tiled.width, gsize.tiled.height);
1377	} else {
1378		size = PAGE_ALIGN(gsize.bytes);
1379	}
1380
1381	omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1382	if (!omap_obj)
1383		goto fail;
1384
1385	list_add(&omap_obj->mm_list, &priv->obj_list);
1386
1387	obj = &omap_obj->base;
1388
1389	if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1390		/* attempt to allocate contiguous memory if we don't
1391		 * have DMM for remappign discontiguous buffers
1392		 */
1393		omap_obj->vaddr =  dma_alloc_writecombine(dev->dev, size,
1394				&omap_obj->paddr, GFP_KERNEL);
1395		if (omap_obj->vaddr)
1396			flags |= OMAP_BO_DMA;
1397
1398	}
1399
1400	omap_obj->flags = flags;
1401
1402	if (flags & OMAP_BO_TILED) {
1403		omap_obj->width = gsize.tiled.width;
1404		omap_obj->height = gsize.tiled.height;
1405	}
1406
1407	ret = 0;
1408	if (flags & (OMAP_BO_DMA|OMAP_BO_EXT_MEM))
1409		drm_gem_private_object_init(dev, obj, size);
1410	else
1411		ret = drm_gem_object_init(dev, obj, size);
1412
1413	if (ret)
1414		goto fail;
1415
1416	return obj;
1417
1418fail:
1419	if (obj)
1420		omap_gem_free_object(obj);
1421
1422	return NULL;
1423}
1424
1425/* init/cleanup.. if DMM is used, we need to set some stuff up.. */
1426void omap_gem_init(struct drm_device *dev)
1427{
1428	struct omap_drm_private *priv = dev->dev_private;
 
1429	const enum tiler_fmt fmts[] = {
1430			TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1431	};
1432	int i, j;
1433
1434	if (!dmm_is_available()) {
1435		/* DMM only supported on OMAP4 and later, so this isn't fatal */
1436		dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1437		return;
1438	}
1439
1440	usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1441	if (!usergart)
1442		return;
1443
1444	/* reserve 4k aligned/wide regions for userspace mappings: */
1445	for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1446		uint16_t h = 1, w = PAGE_SIZE >> i;
 
1447		tiler_align(fmts[i], &w, &h);
1448		/* note: since each region is 1 4kb page wide, and minimum
1449		 * number of rows, the height ends up being the same as the
1450		 * # of pages in the region
1451		 */
1452		usergart[i].height = h;
1453		usergart[i].height_shift = ilog2(h);
1454		usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1455		usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1456		for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1457			struct usergart_entry *entry = &usergart[i].entry[j];
1458			struct tiler_block *block =
1459					tiler_reserve_2d(fmts[i], w, h,
1460							PAGE_SIZE);
 
1461			if (IS_ERR(block)) {
1462				dev_err(dev->dev,
1463						"reserve failed: %d, %d, %ld\n",
1464						i, j, PTR_ERR(block));
1465				return;
1466			}
1467			entry->paddr = tiler_ssptr(block);
1468			entry->block = block;
1469
1470			DBG("%d:%d: %dx%d: paddr=%08x stride=%d", i, j, w, h,
1471					entry->paddr,
1472					usergart[i].stride_pfn << PAGE_SHIFT);
1473		}
1474	}
1475
 
1476	priv->has_dmm = true;
1477}
1478
1479void omap_gem_deinit(struct drm_device *dev)
1480{
 
 
1481	/* I believe we can rely on there being no more outstanding GEM
1482	 * objects which could depend on usergart/dmm at this point.
1483	 */
1484	kfree(usergart);
1485}