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