xref: /openbmc/linux/drivers/gpu/drm/omapdrm/omap_gem.c (revision 31b90347)
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 */
29 struct page **_drm_gem_get_pages(struct drm_gem_object *obj, gfp_t gfpmask);
30 void _drm_gem_put_pages(struct drm_gem_object *obj, struct page **pages,
31 		bool dirty, bool accessed);
32 int _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 
46 struct 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 
122 static int get_pages(struct drm_gem_object *obj, struct page ***pages);
123 static 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
137 struct 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 };
144 static 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 
153 static void evict_entry(struct drm_gem_object *obj,
154 		enum tiler_fmt fmt, struct usergart_entry *entry)
155 {
156 	if (obj->dev->dev_mapping) {
157 		struct omap_gem_object *omap_obj = to_omap_bo(obj);
158 		int n = usergart[fmt].height;
159 		size_t size = PAGE_SIZE * n;
160 		loff_t off = mmap_offset(obj) +
161 				(entry->obj_pgoff << PAGE_SHIFT);
162 		const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
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->dev_mapping,
168 						off, PAGE_SIZE, 1);
169 				off += PAGE_SIZE * m;
170 			}
171 		} else {
172 			unmap_mapping_range(obj->dev->dev_mapping, off, size, 1);
173 		}
174 	}
175 
176 	entry->obj = NULL;
177 }
178 
179 /* Evict a buffer from usergart, if it is mapped there */
180 static 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  */
206 static 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  */
215 static 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 
222 static DEFINE_SPINLOCK(sync_lock);
223 
224 /** ensure backing pages are allocated */
225 static 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 
273 free_pages:
274 	drm_gem_put_pages(obj, pages, true, false);
275 
276 	return ret;
277 }
278 
279 /** release backing pages */
280 static 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 */
303 uint32_t omap_gem_flags(struct drm_gem_object *obj)
304 {
305 	return to_omap_bo(obj)->flags;
306 }
307 
308 /** get mmap offset */
309 static 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 
328 uint64_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 */
338 size_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 */
358 int 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 */
370 static 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 */
396 static 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  */
509 int 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 
539 fail:
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.. */
554 int 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 
567 int 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  */
614 int 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  */
640 int 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 
657 fail:
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  */
666 int 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 
692 fail:
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  */
701 void 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 */
714 void 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  */
744 int 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 
807 fail:
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  */
816 int 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 	}
839 fail:
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  */
848 int 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) */
865 int 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  */
877 static 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  */
906 int 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 */
924 int 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  */
937 void *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: */
954 int 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
981 void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
982 {
983 	struct drm_device *dev = obj->dev;
984 	struct omap_gem_object *omap_obj = to_omap_bo(obj);
985 	uint64_t off;
986 
987 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
988 
989 	off = drm_vma_node_start(&obj->vma_node);
990 
991 	seq_printf(m, "%08x: %2d (%2d) %08llx %08Zx (%2d) %p %4d",
992 			omap_obj->flags, obj->name, obj->refcount.refcount.counter,
993 			off, omap_obj->paddr, omap_obj->paddr_cnt,
994 			omap_obj->vaddr, omap_obj->roll);
995 
996 	if (omap_obj->flags & OMAP_BO_TILED) {
997 		seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
998 		if (omap_obj->block) {
999 			struct tcm_area *area = &omap_obj->block->area;
1000 			seq_printf(m, " (%dx%d, %dx%d)",
1001 					area->p0.x, area->p0.y,
1002 					area->p1.x, area->p1.y);
1003 		}
1004 	} else {
1005 		seq_printf(m, " %d", obj->size);
1006 	}
1007 
1008 	seq_printf(m, "\n");
1009 }
1010 
1011 void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1012 {
1013 	struct omap_gem_object *omap_obj;
1014 	int count = 0;
1015 	size_t size = 0;
1016 
1017 	list_for_each_entry(omap_obj, list, mm_list) {
1018 		struct drm_gem_object *obj = &omap_obj->base;
1019 		seq_printf(m, "   ");
1020 		omap_gem_describe(obj, m);
1021 		count++;
1022 		size += obj->size;
1023 	}
1024 
1025 	seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1026 }
1027 #endif
1028 
1029 /* Buffer Synchronization:
1030  */
1031 
1032 struct omap_gem_sync_waiter {
1033 	struct list_head list;
1034 	struct omap_gem_object *omap_obj;
1035 	enum omap_gem_op op;
1036 	uint32_t read_target, write_target;
1037 	/* notify called w/ sync_lock held */
1038 	void (*notify)(void *arg);
1039 	void *arg;
1040 };
1041 
1042 /* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1043  * the read and/or write target count is achieved which can call a user
1044  * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1045  * cpu access), etc.
1046  */
1047 static LIST_HEAD(waiters);
1048 
1049 static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1050 {
1051 	struct omap_gem_object *omap_obj = waiter->omap_obj;
1052 	if ((waiter->op & OMAP_GEM_READ) &&
1053 			(omap_obj->sync->read_complete < waiter->read_target))
1054 		return true;
1055 	if ((waiter->op & OMAP_GEM_WRITE) &&
1056 			(omap_obj->sync->write_complete < waiter->write_target))
1057 		return true;
1058 	return false;
1059 }
1060 
1061 /* macro for sync debug.. */
1062 #define SYNCDBG 0
1063 #define SYNC(fmt, ...) do { if (SYNCDBG) \
1064 		printk(KERN_ERR "%s:%d: "fmt"\n", \
1065 				__func__, __LINE__, ##__VA_ARGS__); \
1066 	} while (0)
1067 
1068 
1069 static void sync_op_update(void)
1070 {
1071 	struct omap_gem_sync_waiter *waiter, *n;
1072 	list_for_each_entry_safe(waiter, n, &waiters, list) {
1073 		if (!is_waiting(waiter)) {
1074 			list_del(&waiter->list);
1075 			SYNC("notify: %p", waiter);
1076 			waiter->notify(waiter->arg);
1077 			kfree(waiter);
1078 		}
1079 	}
1080 }
1081 
1082 static inline int sync_op(struct drm_gem_object *obj,
1083 		enum omap_gem_op op, bool start)
1084 {
1085 	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1086 	int ret = 0;
1087 
1088 	spin_lock(&sync_lock);
1089 
1090 	if (!omap_obj->sync) {
1091 		omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1092 		if (!omap_obj->sync) {
1093 			ret = -ENOMEM;
1094 			goto unlock;
1095 		}
1096 	}
1097 
1098 	if (start) {
1099 		if (op & OMAP_GEM_READ)
1100 			omap_obj->sync->read_pending++;
1101 		if (op & OMAP_GEM_WRITE)
1102 			omap_obj->sync->write_pending++;
1103 	} else {
1104 		if (op & OMAP_GEM_READ)
1105 			omap_obj->sync->read_complete++;
1106 		if (op & OMAP_GEM_WRITE)
1107 			omap_obj->sync->write_complete++;
1108 		sync_op_update();
1109 	}
1110 
1111 unlock:
1112 	spin_unlock(&sync_lock);
1113 
1114 	return ret;
1115 }
1116 
1117 /* it is a bit lame to handle updates in this sort of polling way, but
1118  * in case of PVR, the GPU can directly update read/write complete
1119  * values, and not really tell us which ones it updated.. this also
1120  * means that sync_lock is not quite sufficient.  So we'll need to
1121  * do something a bit better when it comes time to add support for
1122  * separate 2d hw..
1123  */
1124 void omap_gem_op_update(void)
1125 {
1126 	spin_lock(&sync_lock);
1127 	sync_op_update();
1128 	spin_unlock(&sync_lock);
1129 }
1130 
1131 /* mark the start of read and/or write operation */
1132 int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1133 {
1134 	return sync_op(obj, op, true);
1135 }
1136 
1137 int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1138 {
1139 	return sync_op(obj, op, false);
1140 }
1141 
1142 static DECLARE_WAIT_QUEUE_HEAD(sync_event);
1143 
1144 static void sync_notify(void *arg)
1145 {
1146 	struct task_struct **waiter_task = arg;
1147 	*waiter_task = NULL;
1148 	wake_up_all(&sync_event);
1149 }
1150 
1151 int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1152 {
1153 	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1154 	int ret = 0;
1155 	if (omap_obj->sync) {
1156 		struct task_struct *waiter_task = current;
1157 		struct omap_gem_sync_waiter *waiter =
1158 				kzalloc(sizeof(*waiter), GFP_KERNEL);
1159 
1160 		if (!waiter)
1161 			return -ENOMEM;
1162 
1163 		waiter->omap_obj = omap_obj;
1164 		waiter->op = op;
1165 		waiter->read_target = omap_obj->sync->read_pending;
1166 		waiter->write_target = omap_obj->sync->write_pending;
1167 		waiter->notify = sync_notify;
1168 		waiter->arg = &waiter_task;
1169 
1170 		spin_lock(&sync_lock);
1171 		if (is_waiting(waiter)) {
1172 			SYNC("waited: %p", waiter);
1173 			list_add_tail(&waiter->list, &waiters);
1174 			spin_unlock(&sync_lock);
1175 			ret = wait_event_interruptible(sync_event,
1176 					(waiter_task == NULL));
1177 			spin_lock(&sync_lock);
1178 			if (waiter_task) {
1179 				SYNC("interrupted: %p", waiter);
1180 				/* we were interrupted */
1181 				list_del(&waiter->list);
1182 				waiter_task = NULL;
1183 			} else {
1184 				/* freed in sync_op_update() */
1185 				waiter = NULL;
1186 			}
1187 		}
1188 		spin_unlock(&sync_lock);
1189 
1190 		if (waiter)
1191 			kfree(waiter);
1192 	}
1193 	return ret;
1194 }
1195 
1196 /* call fxn(arg), either synchronously or asynchronously if the op
1197  * is currently blocked..  fxn() can be called from any context
1198  *
1199  * (TODO for now fxn is called back from whichever context calls
1200  * omap_gem_op_update().. but this could be better defined later
1201  * if needed)
1202  *
1203  * TODO more code in common w/ _sync()..
1204  */
1205 int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1206 		void (*fxn)(void *arg), void *arg)
1207 {
1208 	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1209 	if (omap_obj->sync) {
1210 		struct omap_gem_sync_waiter *waiter =
1211 				kzalloc(sizeof(*waiter), GFP_ATOMIC);
1212 
1213 		if (!waiter)
1214 			return -ENOMEM;
1215 
1216 		waiter->omap_obj = omap_obj;
1217 		waiter->op = op;
1218 		waiter->read_target = omap_obj->sync->read_pending;
1219 		waiter->write_target = omap_obj->sync->write_pending;
1220 		waiter->notify = fxn;
1221 		waiter->arg = arg;
1222 
1223 		spin_lock(&sync_lock);
1224 		if (is_waiting(waiter)) {
1225 			SYNC("waited: %p", waiter);
1226 			list_add_tail(&waiter->list, &waiters);
1227 			spin_unlock(&sync_lock);
1228 			return 0;
1229 		}
1230 
1231 		spin_unlock(&sync_lock);
1232 	}
1233 
1234 	/* no waiting.. */
1235 	fxn(arg);
1236 
1237 	return 0;
1238 }
1239 
1240 /* special API so PVR can update the buffer to use a sync-object allocated
1241  * from it's sync-obj heap.  Only used for a newly allocated (from PVR's
1242  * perspective) sync-object, so we overwrite the new syncobj w/ values
1243  * from the already allocated syncobj (if there is one)
1244  */
1245 int omap_gem_set_sync_object(struct drm_gem_object *obj, void *syncobj)
1246 {
1247 	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1248 	int ret = 0;
1249 
1250 	spin_lock(&sync_lock);
1251 
1252 	if ((omap_obj->flags & OMAP_BO_EXT_SYNC) && !syncobj) {
1253 		/* clearing a previously set syncobj */
1254 		syncobj = kmemdup(omap_obj->sync, sizeof(*omap_obj->sync),
1255 				  GFP_ATOMIC);
1256 		if (!syncobj) {
1257 			ret = -ENOMEM;
1258 			goto unlock;
1259 		}
1260 		omap_obj->flags &= ~OMAP_BO_EXT_SYNC;
1261 		omap_obj->sync = syncobj;
1262 	} else if (syncobj && !(omap_obj->flags & OMAP_BO_EXT_SYNC)) {
1263 		/* replacing an existing syncobj */
1264 		if (omap_obj->sync) {
1265 			memcpy(syncobj, omap_obj->sync, sizeof(*omap_obj->sync));
1266 			kfree(omap_obj->sync);
1267 		}
1268 		omap_obj->flags |= OMAP_BO_EXT_SYNC;
1269 		omap_obj->sync = syncobj;
1270 	}
1271 
1272 unlock:
1273 	spin_unlock(&sync_lock);
1274 	return ret;
1275 }
1276 
1277 /* don't call directly.. called from GEM core when it is time to actually
1278  * free the object..
1279  */
1280 void omap_gem_free_object(struct drm_gem_object *obj)
1281 {
1282 	struct drm_device *dev = obj->dev;
1283 	struct omap_gem_object *omap_obj = to_omap_bo(obj);
1284 
1285 	evict(obj);
1286 
1287 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1288 
1289 	list_del(&omap_obj->mm_list);
1290 
1291 	drm_gem_free_mmap_offset(obj);
1292 
1293 	/* this means the object is still pinned.. which really should
1294 	 * not happen.  I think..
1295 	 */
1296 	WARN_ON(omap_obj->paddr_cnt > 0);
1297 
1298 	/* don't free externally allocated backing memory */
1299 	if (!(omap_obj->flags & OMAP_BO_EXT_MEM)) {
1300 		if (omap_obj->pages)
1301 			omap_gem_detach_pages(obj);
1302 
1303 		if (!is_shmem(obj)) {
1304 			dma_free_writecombine(dev->dev, obj->size,
1305 					omap_obj->vaddr, omap_obj->paddr);
1306 		} else if (omap_obj->vaddr) {
1307 			vunmap(omap_obj->vaddr);
1308 		}
1309 	}
1310 
1311 	/* don't free externally allocated syncobj */
1312 	if (!(omap_obj->flags & OMAP_BO_EXT_SYNC))
1313 		kfree(omap_obj->sync);
1314 
1315 	drm_gem_object_release(obj);
1316 
1317 	kfree(obj);
1318 }
1319 
1320 /* convenience method to construct a GEM buffer object, and userspace handle */
1321 int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1322 		union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1323 {
1324 	struct drm_gem_object *obj;
1325 	int ret;
1326 
1327 	obj = omap_gem_new(dev, gsize, flags);
1328 	if (!obj)
1329 		return -ENOMEM;
1330 
1331 	ret = drm_gem_handle_create(file, obj, handle);
1332 	if (ret) {
1333 		drm_gem_object_release(obj);
1334 		kfree(obj); /* TODO isn't there a dtor to call? just copying i915 */
1335 		return ret;
1336 	}
1337 
1338 	/* drop reference from allocate - handle holds it now */
1339 	drm_gem_object_unreference_unlocked(obj);
1340 
1341 	return 0;
1342 }
1343 
1344 /* GEM buffer object constructor */
1345 struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1346 		union omap_gem_size gsize, uint32_t flags)
1347 {
1348 	struct omap_drm_private *priv = dev->dev_private;
1349 	struct omap_gem_object *omap_obj;
1350 	struct drm_gem_object *obj = NULL;
1351 	size_t size;
1352 	int ret;
1353 
1354 	if (flags & OMAP_BO_TILED) {
1355 		if (!usergart) {
1356 			dev_err(dev->dev, "Tiled buffers require DMM\n");
1357 			goto fail;
1358 		}
1359 
1360 		/* tiled buffers are always shmem paged backed.. when they are
1361 		 * scanned out, they are remapped into DMM/TILER
1362 		 */
1363 		flags &= ~OMAP_BO_SCANOUT;
1364 
1365 		/* currently don't allow cached buffers.. there is some caching
1366 		 * stuff that needs to be handled better
1367 		 */
1368 		flags &= ~(OMAP_BO_CACHED|OMAP_BO_UNCACHED);
1369 		flags |= OMAP_BO_WC;
1370 
1371 		/* align dimensions to slot boundaries... */
1372 		tiler_align(gem2fmt(flags),
1373 				&gsize.tiled.width, &gsize.tiled.height);
1374 
1375 		/* ...and calculate size based on aligned dimensions */
1376 		size = tiler_size(gem2fmt(flags),
1377 				gsize.tiled.width, gsize.tiled.height);
1378 	} else {
1379 		size = PAGE_ALIGN(gsize.bytes);
1380 	}
1381 
1382 	omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1383 	if (!omap_obj)
1384 		goto fail;
1385 
1386 	list_add(&omap_obj->mm_list, &priv->obj_list);
1387 
1388 	obj = &omap_obj->base;
1389 
1390 	if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1391 		/* attempt to allocate contiguous memory if we don't
1392 		 * have DMM for remappign discontiguous buffers
1393 		 */
1394 		omap_obj->vaddr =  dma_alloc_writecombine(dev->dev, size,
1395 				&omap_obj->paddr, GFP_KERNEL);
1396 		if (omap_obj->vaddr)
1397 			flags |= OMAP_BO_DMA;
1398 
1399 	}
1400 
1401 	omap_obj->flags = flags;
1402 
1403 	if (flags & OMAP_BO_TILED) {
1404 		omap_obj->width = gsize.tiled.width;
1405 		omap_obj->height = gsize.tiled.height;
1406 	}
1407 
1408 	ret = 0;
1409 	if (flags & (OMAP_BO_DMA|OMAP_BO_EXT_MEM))
1410 		drm_gem_private_object_init(dev, obj, size);
1411 	else
1412 		ret = drm_gem_object_init(dev, obj, size);
1413 
1414 	if (ret)
1415 		goto fail;
1416 
1417 	return obj;
1418 
1419 fail:
1420 	if (obj)
1421 		omap_gem_free_object(obj);
1422 
1423 	return NULL;
1424 }
1425 
1426 /* init/cleanup.. if DMM is used, we need to set some stuff up.. */
1427 void omap_gem_init(struct drm_device *dev)
1428 {
1429 	struct omap_drm_private *priv = dev->dev_private;
1430 	const enum tiler_fmt fmts[] = {
1431 			TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1432 	};
1433 	int i, j;
1434 
1435 	if (!dmm_is_available()) {
1436 		/* DMM only supported on OMAP4 and later, so this isn't fatal */
1437 		dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1438 		return;
1439 	}
1440 
1441 	usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1442 	if (!usergart)
1443 		return;
1444 
1445 	/* reserve 4k aligned/wide regions for userspace mappings: */
1446 	for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1447 		uint16_t h = 1, w = PAGE_SIZE >> i;
1448 		tiler_align(fmts[i], &w, &h);
1449 		/* note: since each region is 1 4kb page wide, and minimum
1450 		 * number of rows, the height ends up being the same as the
1451 		 * # of pages in the region
1452 		 */
1453 		usergart[i].height = h;
1454 		usergart[i].height_shift = ilog2(h);
1455 		usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1456 		usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1457 		for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1458 			struct usergart_entry *entry = &usergart[i].entry[j];
1459 			struct tiler_block *block =
1460 					tiler_reserve_2d(fmts[i], w, h,
1461 							PAGE_SIZE);
1462 			if (IS_ERR(block)) {
1463 				dev_err(dev->dev,
1464 						"reserve failed: %d, %d, %ld\n",
1465 						i, j, PTR_ERR(block));
1466 				return;
1467 			}
1468 			entry->paddr = tiler_ssptr(block);
1469 			entry->block = block;
1470 
1471 			DBG("%d:%d: %dx%d: paddr=%08x stride=%d", i, j, w, h,
1472 					entry->paddr,
1473 					usergart[i].stride_pfn << PAGE_SHIFT);
1474 		}
1475 	}
1476 
1477 	priv->has_dmm = true;
1478 }
1479 
1480 void omap_gem_deinit(struct drm_device *dev)
1481 {
1482 	/* I believe we can rely on there being no more outstanding GEM
1483 	 * objects which could depend on usergart/dmm at this point.
1484 	 */
1485 	kfree(usergart);
1486 }
1487