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