Linux Audio

Check our new training course

Loading...
v6.13.7
   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#include <linux/vmalloc.h>
  13
  14#include <drm/drm_prime.h>
 
  15#include <drm/drm_vma_manager.h>
  16
  17#include "omap_drv.h"
  18#include "omap_dmm_tiler.h"
  19
 
 
 
 
 
 
  20/*
  21 * GEM buffer object implementation.
  22 */
  23
 
 
  24/* note: we use upper 8 bits of flags for driver-internal flags: */
  25#define OMAP_BO_MEM_DMA_API	0x01000000	/* memory allocated with the dma_alloc_* API */
  26#define OMAP_BO_MEM_SHMEM	0x02000000	/* memory allocated through shmem backing */
  27#define OMAP_BO_MEM_DMABUF	0x08000000	/* memory imported from a dmabuf */
 
  28
  29struct omap_gem_object {
  30	struct drm_gem_object base;
  31
  32	struct list_head mm_list;
  33
  34	u32 flags;
  35
  36	/** width/height for tiled formats (rounded up to slot boundaries) */
  37	u16 width, height;
  38
  39	/** roll applied when mapping to DMM */
  40	u32 roll;
  41
  42	/** protects pin_cnt, block, pages, dma_addrs and vaddr */
  43	struct mutex lock;
  44
  45	/**
  46	 * dma_addr contains the buffer DMA address. It is valid for
  47	 *
  48	 * - buffers allocated through the DMA mapping API (with the
  49	 *   OMAP_BO_MEM_DMA_API flag set)
  50	 *
  51	 * - buffers imported from dmabuf (with the OMAP_BO_MEM_DMABUF flag set)
  52	 *   if they are physically contiguous
  53	 *
  54	 * - buffers mapped through the TILER when pin_cnt is not zero, in which
  55	 *   case the DMA address points to the TILER aperture
  56	 *
  57	 * Physically contiguous buffers have their DMA address equal to the
  58	 * physical address as we don't remap those buffers through the TILER.
  59	 *
  60	 * Buffers mapped to the TILER have their DMA address pointing to the
  61	 * TILER aperture. As TILER mappings are refcounted (through pin_cnt)
  62	 * the DMA address must be accessed through omap_gem_pin() to ensure
  63	 * that the mapping won't disappear unexpectedly. References must be
  64	 * released with omap_gem_unpin().
  65	 */
  66	dma_addr_t dma_addr;
  67
  68	/**
  69	 * # of users
  70	 */
  71	refcount_t pin_cnt;
  72
  73	/**
  74	 * If the buffer has been imported from a dmabuf the OMAP_DB_DMABUF flag
  75	 * is set and the sgt field is valid.
  76	 */
  77	struct sg_table *sgt;
  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 *dma_addrs;
  92
  93	/**
  94	 * Virtual address, if mapped.
  95	 */
  96	void *vaddr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  97};
  98
  99#define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
 
 100
 101/* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
 102 * not necessarily pinned in TILER all the time, and (b) when they are
 103 * they are not necessarily page aligned, we reserve one or more small
 104 * regions in each of the 2d containers to use as a user-GART where we
 105 * can create a second page-aligned mapping of parts of the buffer
 106 * being accessed from userspace.
 107 *
 108 * Note that we could optimize slightly when we know that multiple
 109 * tiler containers are backed by the same PAT.. but I'll leave that
 110 * for later..
 111 */
 112#define NUM_USERGART_ENTRIES 2
 113struct omap_drm_usergart_entry {
 114	struct tiler_block *block;	/* the reserved tiler block */
 115	dma_addr_t dma_addr;
 116	struct drm_gem_object *obj;	/* the current pinned obj */
 117	pgoff_t obj_pgoff;		/* page offset of obj currently
 118					   mapped in */
 119};
 120
 121struct omap_drm_usergart {
 122	struct omap_drm_usergart_entry entry[NUM_USERGART_ENTRIES];
 123	int height;				/* height in rows */
 124	int height_shift;		/* ilog2(height in rows) */
 125	int slot_shift;			/* ilog2(width per slot) */
 126	int stride_pfn;			/* stride in pages */
 127	int last;				/* index of last used entry */
 128};
 129
 130/* -----------------------------------------------------------------------------
 131 * Helpers
 132 */
 133
 134/** get mmap offset */
 135u64 omap_gem_mmap_offset(struct drm_gem_object *obj)
 136{
 137	struct drm_device *dev = obj->dev;
 138	int ret;
 139	size_t size;
 140
 141	/* Make it mmapable */
 142	size = omap_gem_mmap_size(obj);
 143	ret = drm_gem_create_mmap_offset_size(obj, size);
 144	if (ret) {
 145		dev_err(dev->dev, "could not allocate mmap offset\n");
 146		return 0;
 147	}
 148
 149	return drm_vma_node_offset_addr(&obj->vma_node);
 150}
 151
 152static bool omap_gem_sgt_is_contiguous(struct sg_table *sgt, size_t size)
 153{
 154	return !(drm_prime_get_contiguous_size(sgt) < size);
 155}
 156
 157static bool omap_gem_is_contiguous(struct omap_gem_object *omap_obj)
 158{
 159	if (omap_obj->flags & OMAP_BO_MEM_DMA_API)
 160		return true;
 161
 162	if ((omap_obj->flags & OMAP_BO_MEM_DMABUF) &&
 163	    omap_gem_sgt_is_contiguous(omap_obj->sgt, omap_obj->base.size))
 164		return true;
 165
 166	return false;
 167}
 168
 169/* -----------------------------------------------------------------------------
 170 * Eviction
 171 */
 172
 173static void omap_gem_evict_entry(struct drm_gem_object *obj,
 174		enum tiler_fmt fmt, struct omap_drm_usergart_entry *entry)
 175{
 176	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 177	struct omap_drm_private *priv = obj->dev->dev_private;
 178	int n = priv->usergart[fmt].height;
 179	size_t size = PAGE_SIZE * n;
 180	loff_t off = omap_gem_mmap_offset(obj) +
 181			(entry->obj_pgoff << PAGE_SHIFT);
 182	const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
 183
 184	if (m > 1) {
 185		int i;
 186		/* if stride > than PAGE_SIZE then sparse mapping: */
 187		for (i = n; i > 0; i--) {
 188			unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 189					    off, PAGE_SIZE, 1);
 190			off += PAGE_SIZE * m;
 191		}
 192	} else {
 193		unmap_mapping_range(obj->dev->anon_inode->i_mapping,
 194				    off, size, 1);
 195	}
 196
 197	entry->obj = NULL;
 198}
 199
 200/* Evict a buffer from usergart, if it is mapped there */
 201static void omap_gem_evict(struct drm_gem_object *obj)
 202{
 203	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 204	struct omap_drm_private *priv = obj->dev->dev_private;
 205
 206	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
 207		enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 208		int i;
 209
 210		for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
 211			struct omap_drm_usergart_entry *entry =
 212				&priv->usergart[fmt].entry[i];
 213
 
 
 214			if (entry->obj == obj)
 215				omap_gem_evict_entry(obj, fmt, entry);
 216		}
 217	}
 218}
 219
 220/* -----------------------------------------------------------------------------
 221 * Page Management
 
 
 
 
 222 */
 
 
 
 
 223
 224/*
 225 * Ensure backing pages are allocated. Must be called with the omap_obj.lock
 226 * held.
 227 */
 
 
 
 
 
 
 
 
 
 
 228static int omap_gem_attach_pages(struct drm_gem_object *obj)
 229{
 230	struct drm_device *dev = obj->dev;
 231	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 232	struct page **pages;
 233	int npages = obj->size >> PAGE_SHIFT;
 234	int i, ret;
 235	dma_addr_t *addrs;
 236
 237	lockdep_assert_held(&omap_obj->lock);
 238
 239	/*
 240	 * If not using shmem (in which case backing pages don't need to be
 241	 * allocated) or if pages are already allocated we're done.
 242	 */
 243	if (!(omap_obj->flags & OMAP_BO_MEM_SHMEM) || omap_obj->pages)
 244		return 0;
 245
 246	pages = drm_gem_get_pages(obj);
 247	if (IS_ERR(pages)) {
 248		dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
 249		return PTR_ERR(pages);
 250	}
 251
 252	/* for non-cached buffers, ensure the new pages are clean because
 253	 * DSS, GPU, etc. are not cache coherent:
 254	 */
 255	if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
 256		addrs = kmalloc_array(npages, sizeof(*addrs), GFP_KERNEL);
 257		if (!addrs) {
 258			ret = -ENOMEM;
 259			goto free_pages;
 260		}
 261
 262		for (i = 0; i < npages; i++) {
 263			addrs[i] = dma_map_page(dev->dev, pages[i],
 264					0, PAGE_SIZE, DMA_TO_DEVICE);
 265
 266			if (dma_mapping_error(dev->dev, addrs[i])) {
 267				dev_warn(dev->dev,
 268					"%s: failed to map page\n", __func__);
 269
 270				for (i = i - 1; i >= 0; --i) {
 271					dma_unmap_page(dev->dev, addrs[i],
 272						PAGE_SIZE, DMA_TO_DEVICE);
 273				}
 274
 275				ret = -ENOMEM;
 276				goto free_addrs;
 277			}
 278		}
 279	} else {
 280		addrs = kcalloc(npages, sizeof(*addrs), GFP_KERNEL);
 281		if (!addrs) {
 282			ret = -ENOMEM;
 283			goto free_pages;
 284		}
 285	}
 286
 287	omap_obj->dma_addrs = addrs;
 288	omap_obj->pages = pages;
 289
 290	return 0;
 291
 292free_addrs:
 293	kfree(addrs);
 294free_pages:
 295	drm_gem_put_pages(obj, pages, true, false);
 296
 297	return ret;
 298}
 299
 300/* Release backing pages. Must be called with the omap_obj.lock held. */
 301static void omap_gem_detach_pages(struct drm_gem_object *obj)
 302{
 303	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 304	unsigned int npages = obj->size >> PAGE_SHIFT;
 305	unsigned int i;
 306
 307	lockdep_assert_held(&omap_obj->lock);
 308
 309	for (i = 0; i < npages; i++) {
 310		if (omap_obj->dma_addrs[i])
 311			dma_unmap_page(obj->dev->dev, omap_obj->dma_addrs[i],
 312				       PAGE_SIZE, DMA_TO_DEVICE);
 
 
 
 
 
 313	}
 314
 315	kfree(omap_obj->dma_addrs);
 316	omap_obj->dma_addrs = NULL;
 317
 318	drm_gem_put_pages(obj, omap_obj->pages, true, false);
 319	omap_obj->pages = NULL;
 320}
 321
 322/* get buffer flags */
 323u32 omap_gem_flags(struct drm_gem_object *obj)
 324{
 325	return to_omap_bo(obj)->flags;
 326}
 327
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 328/** get mmap size */
 329size_t omap_gem_mmap_size(struct drm_gem_object *obj)
 330{
 331	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 332	size_t size = obj->size;
 333
 334	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
 335		/* for tiled buffers, the virtual size has stride rounded up
 336		 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
 337		 * 32kb later!).  But we don't back the entire buffer with
 338		 * pages, only the valid picture part.. so need to adjust for
 339		 * this in the size used to mmap and generate mmap offset
 340		 */
 341		size = tiler_vsize(gem2fmt(omap_obj->flags),
 342				omap_obj->width, omap_obj->height);
 343	}
 344
 345	return size;
 346}
 347
 348/* -----------------------------------------------------------------------------
 349 * Fault Handling
 350 */
 
 
 
 
 
 
 
 
 351
 352/* Normal handling for the case of faulting in non-tiled buffers */
 353static vm_fault_t omap_gem_fault_1d(struct drm_gem_object *obj,
 354		struct vm_area_struct *vma, struct vm_fault *vmf)
 355{
 356	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 357	unsigned long pfn;
 358	pgoff_t pgoff;
 359
 360	/* We don't use vmf->pgoff since that has the fake offset: */
 361	pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
 
 362
 363	if (omap_obj->pages) {
 364		omap_gem_cpu_sync_page(obj, pgoff);
 365		pfn = page_to_pfn(omap_obj->pages[pgoff]);
 366	} else {
 367		BUG_ON(!omap_gem_is_contiguous(omap_obj));
 368		pfn = (omap_obj->dma_addr >> PAGE_SHIFT) + pgoff;
 369	}
 370
 371	VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
 372			pfn, pfn << PAGE_SHIFT);
 373
 374	return vmf_insert_mixed(vma, vmf->address,
 375			__pfn_to_pfn_t(pfn, PFN_DEV));
 376}
 377
 378/* Special handling for the case of faulting in 2d tiled buffers */
 379static vm_fault_t omap_gem_fault_2d(struct drm_gem_object *obj,
 380		struct vm_area_struct *vma, struct vm_fault *vmf)
 381{
 382	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 383	struct omap_drm_private *priv = obj->dev->dev_private;
 384	struct omap_drm_usergart_entry *entry;
 385	enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 386	struct page *pages[64];  /* XXX is this too much to have on stack? */
 387	unsigned long pfn;
 388	pgoff_t pgoff, base_pgoff;
 389	unsigned long vaddr;
 390	int i, err, slots;
 391	vm_fault_t ret = VM_FAULT_NOPAGE;
 392
 393	/*
 394	 * Note the height of the slot is also equal to the number of pages
 395	 * that need to be mapped in to fill 4kb wide CPU page.  If the slot
 396	 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
 397	 */
 398	const int n = priv->usergart[fmt].height;
 399	const int n_shift = priv->usergart[fmt].height_shift;
 400
 401	/*
 402	 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
 403	 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
 404	 * into account in some of the math, so figure out virtual stride
 405	 * in pages
 406	 */
 407	const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
 408
 409	/* We don't use vmf->pgoff since that has the fake offset: */
 410	pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
 
 411
 412	/*
 413	 * Actual address we start mapping at is rounded down to previous slot
 414	 * boundary in the y direction:
 415	 */
 416	base_pgoff = round_down(pgoff, m << n_shift);
 417
 418	/* figure out buffer width in slots */
 419	slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
 420
 421	vaddr = vmf->address - ((pgoff - base_pgoff) << PAGE_SHIFT);
 422
 423	entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
 424
 425	/* evict previous buffer using this usergart entry, if any: */
 426	if (entry->obj)
 427		omap_gem_evict_entry(entry->obj, fmt, entry);
 428
 429	entry->obj = obj;
 430	entry->obj_pgoff = base_pgoff;
 431
 432	/* now convert base_pgoff to phys offset from virt offset: */
 433	base_pgoff = (base_pgoff >> n_shift) * slots;
 434
 435	/* for wider-than 4k.. figure out which part of the slot-row we want: */
 436	if (m > 1) {
 437		int off = pgoff % m;
 438		entry->obj_pgoff += off;
 439		base_pgoff /= m;
 440		slots = min(slots - (off << n_shift), n);
 441		base_pgoff += off << n_shift;
 442		vaddr += off << PAGE_SHIFT;
 443	}
 444
 445	/*
 446	 * Map in pages. Beyond the valid pixel part of the buffer, we set
 447	 * pages[i] to NULL to get a dummy page mapped in.. if someone
 448	 * reads/writes it they will get random/undefined content, but at
 449	 * least it won't be corrupting whatever other random page used to
 450	 * be mapped in, or other undefined behavior.
 451	 */
 452	memcpy(pages, &omap_obj->pages[base_pgoff],
 453			sizeof(struct page *) * slots);
 454	memset(pages + slots, 0,
 455			sizeof(struct page *) * (n - slots));
 456
 457	err = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
 458	if (err) {
 459		ret = vmf_error(err);
 460		dev_err(obj->dev->dev, "failed to pin: %d\n", err);
 461		return ret;
 462	}
 463
 464	pfn = entry->dma_addr >> PAGE_SHIFT;
 465
 466	VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
 467			pfn, pfn << PAGE_SHIFT);
 468
 469	for (i = n; i > 0; i--) {
 470		ret = vmf_insert_mixed(vma,
 471			vaddr, __pfn_to_pfn_t(pfn, PFN_DEV));
 472		if (ret & VM_FAULT_ERROR)
 473			break;
 474		pfn += priv->usergart[fmt].stride_pfn;
 475		vaddr += PAGE_SIZE * m;
 476	}
 477
 478	/* simple round-robin: */
 479	priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
 480				 % NUM_USERGART_ENTRIES;
 481
 482	return ret;
 483}
 484
 485/**
 486 * omap_gem_fault		-	pagefault handler for GEM objects
 
 487 * @vmf: fault detail
 488 *
 489 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
 490 * does most of the work for us including the actual map/unmap calls
 491 * but we need to do the actual page work.
 492 *
 493 * The VMA was set up by GEM. In doing so it also ensured that the
 494 * vma->vm_private_data points to the GEM object that is backing this
 495 * mapping.
 496 */
 497static vm_fault_t omap_gem_fault(struct vm_fault *vmf)
 498{
 499	struct vm_area_struct *vma = vmf->vma;
 500	struct drm_gem_object *obj = vma->vm_private_data;
 501	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 502	int err;
 503	vm_fault_t ret;
 
 504
 505	/* Make sure we don't parallel update on a fault, nor move or remove
 506	 * something from beneath our feet
 507	 */
 508	mutex_lock(&omap_obj->lock);
 509
 510	/* if a shmem backed object, make sure we have pages attached now */
 511	err = omap_gem_attach_pages(obj);
 512	if (err) {
 513		ret = vmf_error(err);
 514		goto fail;
 515	}
 516
 517	/* where should we do corresponding put_pages().. we are mapping
 518	 * the original page, rather than thru a GART, so we can't rely
 519	 * on eviction to trigger this.  But munmap() or all mappings should
 520	 * probably trigger put_pages()?
 521	 */
 522
 523	if (omap_obj->flags & OMAP_BO_TILED_MASK)
 524		ret = omap_gem_fault_2d(obj, vma, vmf);
 525	else
 526		ret = omap_gem_fault_1d(obj, vma, vmf);
 527
 528
 529fail:
 530	mutex_unlock(&omap_obj->lock);
 531	return ret;
 
 
 
 
 
 
 
 
 
 532}
 533
 534static int omap_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 535{
 536	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 537
 538	vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP | VM_IO | VM_MIXEDMAP);
 
 539
 540	if (omap_obj->flags & OMAP_BO_WC) {
 541		vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
 542	} else if (omap_obj->flags & OMAP_BO_UNCACHED) {
 543		vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
 544	} else {
 545		/*
 546		 * We do have some private objects, at least for scanout buffers
 547		 * on hardware without DMM/TILER.  But these are allocated write-
 548		 * combine
 549		 */
 550		if (WARN_ON(!obj->filp))
 551			return -EINVAL;
 552
 553		/*
 554		 * Shunt off cached objs to shmem file so they have their own
 555		 * address_space (so unmap_mapping_range does what we want,
 556		 * in particular in the case of mmap'd dmabufs)
 557		 */
 558		vma->vm_pgoff -= drm_vma_node_start(&obj->vma_node);
 559		vma_set_file(vma, obj->filp);
 
 560
 561		vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 562	}
 563
 564	vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
 565
 566	return 0;
 567}
 568
 569/* -----------------------------------------------------------------------------
 570 * Dumb Buffers
 571 */
 572
 573/**
 574 * omap_gem_dumb_create	-	create a dumb buffer
 575 * @file: our client file
 576 * @dev: our device
 577 * @args: the requested arguments copied from userspace
 578 *
 579 * Allocate a buffer suitable for use for a frame buffer of the
 580 * form described by user space. Give userspace a handle by which
 581 * to reference it.
 582 */
 583int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
 584		struct drm_mode_create_dumb *args)
 585{
 586	union omap_gem_size gsize;
 587
 588	args->pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
 589
 590	args->size = PAGE_ALIGN(args->pitch * args->height);
 591
 592	gsize = (union omap_gem_size){
 593		.bytes = args->size,
 594	};
 595
 596	return omap_gem_new_handle(dev, file, gsize,
 597			OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
 598}
 599
 600/**
 601 * omap_gem_dumb_map_offset - create an offset for a dumb buffer
 602 * @file: our drm client file
 603 * @dev: drm device
 604 * @handle: GEM handle to the object (from dumb_create)
 605 * @offset: memory map offset placeholder
 606 *
 607 * Do the necessary setup to allow the mapping of the frame buffer
 608 * into user memory. We don't have to do much here at the moment.
 609 */
 610int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 611		u32 handle, u64 *offset)
 612{
 613	struct drm_gem_object *obj;
 614	int ret = 0;
 615
 616	/* GEM does all our handle to object mapping */
 617	obj = drm_gem_object_lookup(file, handle);
 618	if (obj == NULL) {
 619		ret = -ENOENT;
 620		goto fail;
 621	}
 622
 623	*offset = omap_gem_mmap_offset(obj);
 624
 625	drm_gem_object_put(obj);
 626
 627fail:
 628	return ret;
 629}
 630
 631#ifdef CONFIG_DRM_FBDEV_EMULATION
 632/* Set scrolling position.  This allows us to implement fast scrolling
 633 * for console.
 634 *
 635 * Call only from non-atomic contexts.
 636 */
 637int omap_gem_roll(struct drm_gem_object *obj, u32 roll)
 638{
 639	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 640	u32 npages = obj->size >> PAGE_SHIFT;
 641	int ret = 0;
 642
 643	if (roll > npages) {
 644		dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
 645		return -EINVAL;
 646	}
 647
 648	omap_obj->roll = roll;
 649
 650	mutex_lock(&omap_obj->lock);
 651
 652	/* if we aren't mapped yet, we don't need to do anything */
 653	if (omap_obj->block) {
 654		ret = omap_gem_attach_pages(obj);
 
 655		if (ret)
 656			goto fail;
 657
 658		ret = tiler_pin(omap_obj->block, omap_obj->pages, npages,
 659				roll, true);
 660		if (ret)
 661			dev_err(obj->dev->dev, "could not repin: %d\n", ret);
 662	}
 663
 664fail:
 665	mutex_unlock(&omap_obj->lock);
 666
 667	return ret;
 668}
 669#endif
 670
 671/* -----------------------------------------------------------------------------
 672 * Memory Management & DMA Sync
 673 */
 674
 675/*
 676 * shmem buffers that are mapped cached are not coherent.
 677 *
 678 * We keep track of dirty pages using page faulting to perform cache management.
 679 * When a page is mapped to the CPU in read/write mode the device can't access
 680 * it and omap_obj->dma_addrs[i] is NULL. When a page is mapped to the device
 681 * the omap_obj->dma_addrs[i] is set to the DMA address, and the page is
 682 * unmapped from the CPU.
 683 */
 684static inline bool omap_gem_is_cached_coherent(struct drm_gem_object *obj)
 685{
 686	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 687
 688	return !((omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
 689		((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED));
 690}
 691
 692/* Sync the buffer for CPU access.. note pages should already be
 693 * attached, ie. omap_gem_get_pages()
 694 */
 695void omap_gem_cpu_sync_page(struct drm_gem_object *obj, int pgoff)
 696{
 697	struct drm_device *dev = obj->dev;
 698	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 699
 700	if (omap_gem_is_cached_coherent(obj))
 701		return;
 702
 703	if (omap_obj->dma_addrs[pgoff]) {
 704		dma_unmap_page(dev->dev, omap_obj->dma_addrs[pgoff],
 705				PAGE_SIZE, DMA_TO_DEVICE);
 706		omap_obj->dma_addrs[pgoff] = 0;
 707	}
 708}
 709
 710/* sync the buffer for DMA access */
 711void omap_gem_dma_sync_buffer(struct drm_gem_object *obj,
 712		enum dma_data_direction dir)
 713{
 714	struct drm_device *dev = obj->dev;
 715	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 716	int i, npages = obj->size >> PAGE_SHIFT;
 717	struct page **pages = omap_obj->pages;
 718	bool dirty = false;
 719
 720	if (omap_gem_is_cached_coherent(obj))
 721		return;
 
 
 722
 723	for (i = 0; i < npages; i++) {
 724		if (!omap_obj->dma_addrs[i]) {
 725			dma_addr_t addr;
 726
 727			addr = dma_map_page(dev->dev, pages[i], 0,
 728					    PAGE_SIZE, dir);
 729			if (dma_mapping_error(dev->dev, addr)) {
 730				dev_warn(dev->dev, "%s: failed to map page\n",
 731					__func__);
 732				break;
 733			}
 734
 735			dirty = true;
 736			omap_obj->dma_addrs[i] = addr;
 737		}
 738	}
 739
 740	if (dirty) {
 741		unmap_mapping_range(obj->filp->f_mapping, 0,
 742				    omap_gem_mmap_size(obj), 1);
 743	}
 744}
 745
 746static int omap_gem_pin_tiler(struct drm_gem_object *obj)
 747{
 748	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 749	u32 npages = obj->size >> PAGE_SHIFT;
 750	enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
 751	struct tiler_block *block;
 752	int ret;
 753
 754	BUG_ON(omap_obj->block);
 755
 756	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
 757		block = tiler_reserve_2d(fmt, omap_obj->width, omap_obj->height,
 758					 PAGE_SIZE);
 759	} else {
 760		block = tiler_reserve_1d(obj->size);
 761	}
 762
 763	if (IS_ERR(block)) {
 764		ret = PTR_ERR(block);
 765		dev_err(obj->dev->dev, "could not remap: %d (%d)\n", ret, fmt);
 766		goto fail;
 767	}
 768
 769	/* TODO: enable async refill.. */
 770	ret = tiler_pin(block, omap_obj->pages, npages, omap_obj->roll, true);
 771	if (ret) {
 772		tiler_release(block);
 773		dev_err(obj->dev->dev, "could not pin: %d\n", ret);
 774		goto fail;
 775	}
 776
 777	omap_obj->dma_addr = tiler_ssptr(block);
 778	omap_obj->block = block;
 779
 780	DBG("got dma address: %pad", &omap_obj->dma_addr);
 781
 782fail:
 783	return ret;
 784}
 785
 786/**
 787 * omap_gem_pin() - Pin a GEM object in memory
 788 * @obj: the GEM object
 789 * @dma_addr: the DMA address
 790 *
 791 * Pin the given GEM object in memory and fill the dma_addr pointer with the
 792 * object's DMA address. If the buffer is not physically contiguous it will be
 793 * remapped through the TILER to provide a contiguous view.
 794 *
 795 * Pins are reference-counted, calling this function multiple times is allowed
 796 * as long the corresponding omap_gem_unpin() calls are balanced.
 797 *
 798 * Return 0 on success or a negative error code otherwise.
 799 */
 800int omap_gem_pin(struct drm_gem_object *obj, dma_addr_t *dma_addr)
 
 801{
 802	struct omap_drm_private *priv = obj->dev->dev_private;
 803	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 804	int ret = 0;
 805
 806	mutex_lock(&omap_obj->lock);
 807
 808	if (!omap_gem_is_contiguous(omap_obj)) {
 809		if (refcount_read(&omap_obj->pin_cnt) == 0) {
 
 
 
 
 810
 811			refcount_set(&omap_obj->pin_cnt, 1);
 812
 813			ret = omap_gem_attach_pages(obj);
 814			if (ret)
 815				goto fail;
 816
 817			if (omap_obj->flags & OMAP_BO_SCANOUT) {
 818				if (priv->has_dmm) {
 819					ret = omap_gem_pin_tiler(obj);
 820					if (ret)
 821						goto fail;
 822				}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 823			}
 824		} else {
 825			refcount_inc(&omap_obj->pin_cnt);
 
 
 
 826		}
 827	}
 828
 829	if (dma_addr)
 830		*dma_addr = omap_obj->dma_addr;
 
 
 
 
 
 
 
 831
 832fail:
 833	mutex_unlock(&omap_obj->lock);
 834
 835	return ret;
 836}
 837
 838/**
 839 * omap_gem_unpin_locked() - Unpin a GEM object from memory
 840 * @obj: the GEM object
 841 *
 842 * omap_gem_unpin() without locking.
 843 */
 844static void omap_gem_unpin_locked(struct drm_gem_object *obj)
 845{
 846	struct omap_drm_private *priv = obj->dev->dev_private;
 847	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 848	int ret;
 849
 850	if (omap_gem_is_contiguous(omap_obj))
 851		return;
 852
 853	if (refcount_dec_and_test(&omap_obj->pin_cnt)) {
 854		if (omap_obj->sgt) {
 855			sg_free_table(omap_obj->sgt);
 856			kfree(omap_obj->sgt);
 857			omap_obj->sgt = NULL;
 858		}
 859		if (!(omap_obj->flags & OMAP_BO_SCANOUT))
 860			return;
 861		if (priv->has_dmm) {
 862			ret = tiler_unpin(omap_obj->block);
 863			if (ret) {
 864				dev_err(obj->dev->dev,
 865					"could not unpin pages: %d\n", ret);
 
 866			}
 867			ret = tiler_release(omap_obj->block);
 868			if (ret) {
 869				dev_err(obj->dev->dev,
 870					"could not release unmap: %d\n", ret);
 871			}
 872			omap_obj->dma_addr = 0;
 873			omap_obj->block = NULL;
 874		}
 875	}
 876}
 877
 878/**
 879 * omap_gem_unpin() - Unpin a GEM object from memory
 880 * @obj: the GEM object
 881 *
 882 * Unpin the given GEM object previously pinned with omap_gem_pin(). Pins are
 883 * reference-counted, the actual unpin will only be performed when the number
 884 * of calls to this function matches the number of calls to omap_gem_pin().
 885 */
 886void omap_gem_unpin(struct drm_gem_object *obj)
 887{
 888	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 889
 890	mutex_lock(&omap_obj->lock);
 891	omap_gem_unpin_locked(obj);
 892	mutex_unlock(&omap_obj->lock);
 893}
 894
 895/* Get rotated scanout address (only valid if already pinned), at the
 896 * specified orientation and x,y offset from top-left corner of buffer
 897 * (only valid for tiled 2d buffers)
 898 */
 899int omap_gem_rotated_dma_addr(struct drm_gem_object *obj, u32 orient,
 900		int x, int y, dma_addr_t *dma_addr)
 901{
 902	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 903	int ret = -EINVAL;
 904
 905	mutex_lock(&omap_obj->lock);
 906
 907	if ((refcount_read(&omap_obj->pin_cnt) > 0) && omap_obj->block &&
 908			(omap_obj->flags & OMAP_BO_TILED_MASK)) {
 909		*dma_addr = tiler_tsptr(omap_obj->block, orient, x, y);
 910		ret = 0;
 911	}
 912
 913	mutex_unlock(&omap_obj->lock);
 914
 915	return ret;
 916}
 917
 918/* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
 919int omap_gem_tiled_stride(struct drm_gem_object *obj, u32 orient)
 920{
 921	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 922	int ret = -EINVAL;
 923	if (omap_obj->flags & OMAP_BO_TILED_MASK)
 924		ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
 925	return ret;
 926}
 927
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 928/* if !remap, and we don't have pages backing, then fail, rather than
 929 * increasing the pin count (which we don't really do yet anyways,
 930 * because we don't support swapping pages back out).  And 'remap'
 931 * might not be quite the right name, but I wanted to keep it working
 932 * similarly to omap_gem_pin().  Note though that mutex is not
 933 * aquired if !remap (because this can be called in atomic ctxt),
 934 * but probably omap_gem_unpin() should be changed to work in the
 935 * same way.  If !remap, a matching omap_gem_put_pages() call is not
 936 * required (and should not be made).
 937 */
 938int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
 939		bool remap)
 940{
 941	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 942	int ret = 0;
 943
 944	mutex_lock(&omap_obj->lock);
 945
 946	if (remap) {
 947		ret = omap_gem_attach_pages(obj);
 948		if (ret)
 949			goto unlock;
 950	}
 951
 952	if (!omap_obj->pages) {
 953		ret = -ENOMEM;
 954		goto unlock;
 955	}
 956
 957	*pages = omap_obj->pages;
 958
 959unlock:
 960	mutex_unlock(&omap_obj->lock);
 961
 962	return ret;
 963}
 964
 965/* release pages when DMA no longer being performed */
 966int omap_gem_put_pages(struct drm_gem_object *obj)
 967{
 968	/* do something here if we dynamically attach/detach pages.. at
 969	 * least they would no longer need to be pinned if everyone has
 970	 * released the pages..
 971	 */
 972	return 0;
 973}
 974
 975struct sg_table *omap_gem_get_sg(struct drm_gem_object *obj,
 976		enum dma_data_direction dir)
 977{
 978	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 979	dma_addr_t addr;
 980	struct sg_table *sgt;
 981	struct scatterlist *sg;
 982	unsigned int count, len, stride, i;
 983	int ret;
 984
 985	ret = omap_gem_pin(obj, &addr);
 986	if (ret)
 987		return ERR_PTR(ret);
 988
 989	mutex_lock(&omap_obj->lock);
 990
 991	sgt = omap_obj->sgt;
 992	if (sgt)
 993		goto out;
 994
 995	sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
 996	if (!sgt) {
 997		ret = -ENOMEM;
 998		goto err_unpin;
 999	}
1000
1001	if (addr) {
1002		if (omap_obj->flags & OMAP_BO_TILED_MASK) {
1003			enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
1004
1005			len = omap_obj->width << (int)fmt;
1006			count = omap_obj->height;
1007			stride = tiler_stride(fmt, 0);
1008		} else {
1009			len = obj->size;
1010			count = 1;
1011			stride = 0;
1012		}
1013	} else {
1014		count = obj->size >> PAGE_SHIFT;
1015	}
1016
1017	ret = sg_alloc_table(sgt, count, GFP_KERNEL);
1018	if (ret)
1019		goto err_free;
1020
1021	/* this must be after omap_gem_pin() to ensure we have pages attached */
1022	omap_gem_dma_sync_buffer(obj, dir);
1023
1024	if (addr) {
1025		for_each_sg(sgt->sgl, sg, count, i) {
1026			sg_set_page(sg, pfn_to_page(__phys_to_pfn(addr)),
1027				    len, offset_in_page(addr));
1028			sg_dma_address(sg) = addr;
1029			sg_dma_len(sg) = len;
1030
1031			addr += stride;
1032		}
1033	} else {
1034		for_each_sg(sgt->sgl, sg, count, i) {
1035			sg_set_page(sg, omap_obj->pages[i], PAGE_SIZE, 0);
1036			sg_dma_address(sg) = omap_obj->dma_addrs[i];
1037			sg_dma_len(sg) =  PAGE_SIZE;
1038		}
1039	}
1040
1041	omap_obj->sgt = sgt;
1042out:
1043	mutex_unlock(&omap_obj->lock);
1044	return sgt;
1045
1046err_free:
1047	kfree(sgt);
1048err_unpin:
1049	mutex_unlock(&omap_obj->lock);
1050	omap_gem_unpin(obj);
1051	return ERR_PTR(ret);
1052}
1053
1054void omap_gem_put_sg(struct drm_gem_object *obj, struct sg_table *sgt)
1055{
1056	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1057
1058	if (WARN_ON(omap_obj->sgt != sgt))
1059		return;
1060
1061	omap_gem_unpin(obj);
1062}
1063
1064#ifdef CONFIG_DRM_FBDEV_EMULATION
1065/*
1066 * Get kernel virtual address for CPU access.. this more or less only
1067 * exists for omap_fbdev.
1068 */
1069void *omap_gem_vaddr(struct drm_gem_object *obj)
1070{
1071	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1072	void *vaddr;
1073	int ret;
1074
1075	mutex_lock(&omap_obj->lock);
1076
1077	if (!omap_obj->vaddr) {
1078		ret = omap_gem_attach_pages(obj);
1079		if (ret) {
1080			vaddr = ERR_PTR(ret);
1081			goto unlock;
1082		}
1083
1084		omap_obj->vaddr = vmap(omap_obj->pages, obj->size >> PAGE_SHIFT,
1085				VM_MAP, pgprot_writecombine(PAGE_KERNEL));
1086	}
1087
1088	vaddr = omap_obj->vaddr;
1089
1090unlock:
1091	mutex_unlock(&omap_obj->lock);
1092	return vaddr;
1093}
1094#endif
1095
1096/* -----------------------------------------------------------------------------
1097 * Power Management
1098 */
1099
1100#ifdef CONFIG_PM
1101/* re-pin objects in DMM in resume path: */
1102int omap_gem_resume(struct drm_device *dev)
1103{
1104	struct omap_drm_private *priv = dev->dev_private;
 
1105	struct omap_gem_object *omap_obj;
1106	int ret = 0;
1107
1108	mutex_lock(&priv->list_lock);
1109	list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
1110		if (omap_obj->block) {
1111			struct drm_gem_object *obj = &omap_obj->base;
1112			u32 npages = obj->size >> PAGE_SHIFT;
1113
1114			WARN_ON(!omap_obj->pages);  /* this can't happen */
1115			ret = tiler_pin(omap_obj->block,
1116					omap_obj->pages, npages,
1117					omap_obj->roll, true);
1118			if (ret) {
1119				dev_err(dev->dev, "could not repin: %d\n", ret);
1120				goto done;
1121			}
1122		}
1123	}
1124
1125done:
1126	mutex_unlock(&priv->list_lock);
1127	return ret;
1128}
1129#endif
1130
1131/* -----------------------------------------------------------------------------
1132 * DebugFS
1133 */
1134
1135#ifdef CONFIG_DEBUG_FS
1136void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1137{
1138	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1139	u64 off;
1140
1141	off = drm_vma_node_start(&obj->vma_node);
1142
1143	mutex_lock(&omap_obj->lock);
1144
1145	seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1146			omap_obj->flags, obj->name, kref_read(&obj->refcount),
1147			off, &omap_obj->dma_addr,
1148			refcount_read(&omap_obj->pin_cnt),
1149			omap_obj->vaddr, omap_obj->roll);
1150
1151	if (omap_obj->flags & OMAP_BO_TILED_MASK) {
1152		seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1153		if (omap_obj->block) {
1154			struct tcm_area *area = &omap_obj->block->area;
1155			seq_printf(m, " (%dx%d, %dx%d)",
1156					area->p0.x, area->p0.y,
1157					area->p1.x, area->p1.y);
1158		}
1159	} else {
1160		seq_printf(m, " %zu", obj->size);
1161	}
1162
1163	mutex_unlock(&omap_obj->lock);
1164
1165	seq_printf(m, "\n");
1166}
1167
1168void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1169{
1170	struct omap_gem_object *omap_obj;
1171	int count = 0;
1172	size_t size = 0;
1173
1174	list_for_each_entry(omap_obj, list, mm_list) {
1175		struct drm_gem_object *obj = &omap_obj->base;
1176		seq_printf(m, "   ");
1177		omap_gem_describe(obj, m);
1178		count++;
1179		size += obj->size;
1180	}
1181
1182	seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1183}
1184#endif
1185
1186/* -----------------------------------------------------------------------------
1187 * Constructor & Destructor
1188 */
1189
1190static void omap_gem_free_object(struct drm_gem_object *obj)
1191{
1192	struct drm_device *dev = obj->dev;
1193	struct omap_drm_private *priv = dev->dev_private;
1194	struct omap_gem_object *omap_obj = to_omap_bo(obj);
 
 
 
 
1195
1196	omap_gem_evict(obj);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1197
1198	mutex_lock(&priv->list_lock);
1199	list_del(&omap_obj->mm_list);
1200	mutex_unlock(&priv->list_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1201
1202	/*
1203	 * We own the sole reference to the object at this point, but to keep
1204	 * lockdep happy, we must still take the omap_obj_lock to call
1205	 * omap_gem_detach_pages(). This should hardly make any difference as
1206	 * there can't be any lock contention.
1207	 */
1208	mutex_lock(&omap_obj->lock);
1209
1210	/* The object should not be pinned. */
1211	WARN_ON(refcount_read(&omap_obj->pin_cnt) > 0);
1212
1213	if (omap_obj->pages) {
1214		if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1215			kfree(omap_obj->pages);
1216		else
1217			omap_gem_detach_pages(obj);
 
1218	}
1219
1220	if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1221		dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1222			    omap_obj->dma_addr);
1223	} else if (omap_obj->vaddr) {
1224		vunmap(omap_obj->vaddr);
1225	} else if (obj->import_attach) {
1226		drm_prime_gem_destroy(obj, omap_obj->sgt);
 
 
 
 
1227	}
1228
1229	mutex_unlock(&omap_obj->lock);
1230
1231	drm_gem_object_release(obj);
1232
1233	mutex_destroy(&omap_obj->lock);
1234
1235	kfree(omap_obj);
1236}
1237
1238static bool omap_gem_validate_flags(struct drm_device *dev, u32 flags)
 
 
 
 
 
 
 
1239{
1240	struct omap_drm_private *priv = dev->dev_private;
 
 
 
1241
1242	switch (flags & OMAP_BO_CACHE_MASK) {
1243	case OMAP_BO_CACHED:
1244	case OMAP_BO_WC:
1245	case OMAP_BO_CACHE_MASK:
1246		break;
1247
1248	default:
1249		return false;
1250	}
 
1251
1252	if (flags & OMAP_BO_TILED_MASK) {
1253		if (!priv->usergart)
1254			return false;
1255
1256		switch (flags & OMAP_BO_TILED_MASK) {
1257		case OMAP_BO_TILED_8:
1258		case OMAP_BO_TILED_16:
1259		case OMAP_BO_TILED_32:
1260			break;
 
1261
1262		default:
1263			return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1264		}
1265	}
1266
1267	return true;
 
 
 
1268}
1269
1270static const struct vm_operations_struct omap_gem_vm_ops = {
1271	.fault = omap_gem_fault,
1272	.open = drm_gem_vm_open,
1273	.close = drm_gem_vm_close,
1274};
1275
1276static const struct drm_gem_object_funcs omap_gem_object_funcs = {
1277	.free = omap_gem_free_object,
1278	.export = omap_gem_prime_export,
1279	.mmap = omap_gem_object_mmap,
1280	.vm_ops = &omap_gem_vm_ops,
1281};
1282
1283/* GEM buffer object constructor */
1284struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1285		union omap_gem_size gsize, u32 flags)
1286{
1287	struct omap_drm_private *priv = dev->dev_private;
1288	struct omap_gem_object *omap_obj;
1289	struct drm_gem_object *obj;
1290	struct address_space *mapping;
1291	size_t size;
1292	int ret;
1293
1294	if (!omap_gem_validate_flags(dev, flags))
1295		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
1296
1297	/* Validate the flags and compute the memory and cache flags. */
1298	if (flags & OMAP_BO_TILED_MASK) {
1299		/*
1300		 * Tiled buffers are always shmem paged backed. When they are
1301		 * scanned out, they are remapped into DMM/TILER.
1302		 */
1303		flags |= OMAP_BO_MEM_SHMEM;
1304
1305		/*
1306		 * Currently don't allow cached buffers. There is some caching
1307		 * stuff that needs to be handled better.
1308		 */
1309		flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1310		flags |= tiler_get_cpu_cache_flags();
1311	} else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1312		/*
1313		 * If we don't have DMM, we must allocate scanout buffers
1314		 * from contiguous DMA memory.
1315		 */
1316		flags |= OMAP_BO_MEM_DMA_API;
1317	} else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1318		/*
1319		 * All other buffers not backed by dma_buf are shmem-backed.
1320		 */
1321		flags |= OMAP_BO_MEM_SHMEM;
1322	}
1323
1324	/* Allocate the initialize the OMAP GEM object. */
1325	omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1326	if (!omap_obj)
1327		return NULL;
1328
1329	obj = &omap_obj->base;
1330	omap_obj->flags = flags;
1331	mutex_init(&omap_obj->lock);
1332
1333	if (flags & OMAP_BO_TILED_MASK) {
1334		/*
1335		 * For tiled buffers align dimensions to slot boundaries and
1336		 * calculate size based on aligned dimensions.
1337		 */
1338		tiler_align(gem2fmt(flags), &gsize.tiled.width,
1339			    &gsize.tiled.height);
1340
1341		size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1342				  gsize.tiled.height);
1343
1344		omap_obj->width = gsize.tiled.width;
1345		omap_obj->height = gsize.tiled.height;
1346	} else {
1347		size = PAGE_ALIGN(gsize.bytes);
1348	}
1349
1350	obj->funcs = &omap_gem_object_funcs;
 
1351
1352	/* Initialize the GEM object. */
1353	if (!(flags & OMAP_BO_MEM_SHMEM)) {
1354		drm_gem_private_object_init(dev, obj, size);
1355	} else {
1356		ret = drm_gem_object_init(dev, obj, size);
1357		if (ret)
1358			goto err_free;
 
 
1359
1360		mapping = obj->filp->f_mapping;
1361		mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1362	}
1363
1364	/* Allocate memory if needed. */
1365	if (flags & OMAP_BO_MEM_DMA_API) {
1366		omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1367					       &omap_obj->dma_addr,
1368					       GFP_KERNEL);
1369		if (!omap_obj->vaddr)
1370			goto err_release;
 
 
 
 
 
 
 
 
 
 
 
1371	}
1372
1373	mutex_lock(&priv->list_lock);
1374	list_add(&omap_obj->mm_list, &priv->obj_list);
1375	mutex_unlock(&priv->list_lock);
1376
1377	return obj;
1378
1379err_release:
1380	drm_gem_object_release(obj);
1381err_free:
1382	kfree(omap_obj);
1383	return NULL;
1384}
1385
1386struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1387					   struct sg_table *sgt)
 
 
1388{
1389	struct omap_drm_private *priv = dev->dev_private;
1390	struct omap_gem_object *omap_obj;
1391	struct drm_gem_object *obj;
1392	union omap_gem_size gsize;
1393
1394	/* Without a DMM only physically contiguous buffers can be supported. */
1395	if (!omap_gem_sgt_is_contiguous(sgt, size) && !priv->has_dmm)
1396		return ERR_PTR(-EINVAL);
1397
1398	gsize.bytes = PAGE_ALIGN(size);
1399	obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1400	if (!obj)
1401		return ERR_PTR(-ENOMEM);
1402
1403	omap_obj = to_omap_bo(obj);
1404
1405	omap_obj->sgt = sgt;
1406
1407	if (omap_gem_sgt_is_contiguous(sgt, size)) {
1408		omap_obj->dma_addr = sg_dma_address(sgt->sgl);
1409	} else {
1410		/* Create pages list from sgt */
1411		struct page **pages;
1412		unsigned int npages;
1413		unsigned int ret;
1414
1415		npages = DIV_ROUND_UP(size, PAGE_SIZE);
1416		pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1417		if (!pages) {
1418			omap_gem_free_object(obj);
1419			return ERR_PTR(-ENOMEM);
1420		}
1421
1422		omap_obj->pages = pages;
1423		ret = drm_prime_sg_to_page_array(sgt, pages, npages);
1424		if (ret) {
1425			omap_gem_free_object(obj);
1426			return ERR_PTR(-ENOMEM);
1427		}
1428	}
1429
1430	return obj;
 
 
 
 
 
 
1431}
1432
1433/* convenience method to construct a GEM buffer object, and userspace handle */
1434int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1435		union omap_gem_size gsize, u32 flags, u32 *handle)
1436{
1437	struct drm_gem_object *obj;
1438	int ret;
1439
1440	obj = omap_gem_new(dev, gsize, flags);
1441	if (!obj)
1442		return -ENOMEM;
1443
1444	ret = drm_gem_handle_create(file, obj, handle);
1445	if (ret) {
1446		omap_gem_free_object(obj);
 
1447		return ret;
1448	}
1449
1450	/* drop reference from allocate - handle holds it now */
1451	drm_gem_object_put(obj);
1452
1453	return 0;
1454}
1455
1456/* -----------------------------------------------------------------------------
1457 * Init & Cleanup
1458 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1459
1460/* If DMM is used, we need to set some stuff up.. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1461void omap_gem_init(struct drm_device *dev)
1462{
1463	struct omap_drm_private *priv = dev->dev_private;
1464	struct omap_drm_usergart *usergart;
1465	const enum tiler_fmt fmts[] = {
1466			TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1467	};
1468	int i, j;
1469
1470	if (!dmm_is_available()) {
1471		/* DMM only supported on OMAP4 and later, so this isn't fatal */
1472		dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1473		return;
1474	}
1475
1476	usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1477	if (!usergart)
1478		return;
1479
1480	/* reserve 4k aligned/wide regions for userspace mappings: */
1481	for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1482		u16 h = 1, w = PAGE_SIZE >> i;
1483
1484		tiler_align(fmts[i], &w, &h);
1485		/* note: since each region is 1 4kb page wide, and minimum
1486		 * number of rows, the height ends up being the same as the
1487		 * # of pages in the region
1488		 */
1489		usergart[i].height = h;
1490		usergart[i].height_shift = ilog2(h);
1491		usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1492		usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1493		for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1494			struct omap_drm_usergart_entry *entry;
1495			struct tiler_block *block;
1496
1497			entry = &usergart[i].entry[j];
1498			block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1499			if (IS_ERR(block)) {
1500				dev_err(dev->dev,
1501						"reserve failed: %d, %d, %ld\n",
1502						i, j, PTR_ERR(block));
1503				return;
1504			}
1505			entry->dma_addr = tiler_ssptr(block);
1506			entry->block = block;
1507
1508			DBG("%d:%d: %dx%d: dma_addr=%pad stride=%d", i, j, w, h,
1509					&entry->dma_addr,
1510					usergart[i].stride_pfn << PAGE_SHIFT);
1511		}
1512	}
1513
1514	priv->usergart = usergart;
1515	priv->has_dmm = true;
1516}
1517
1518void omap_gem_deinit(struct drm_device *dev)
1519{
1520	struct omap_drm_private *priv = dev->dev_private;
1521
1522	/* I believe we can rely on there being no more outstanding GEM
1523	 * objects which could depend on usergart/dmm at this point.
1524	 */
1525	kfree(priv->usergart);
1526}
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}