Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2011 Texas Instruments Incorporated - http://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 dma_addr_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 (when sgt->orig_nents == 1)
52 *
53 * - buffers mapped through the TILER when dma_addr_cnt is not zero, in
54 * which 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
61 * dma_addr_cnt) the DMA address must be accessed through omap_gem_pin()
62 * to ensure that the mapping won't disappear unexpectedly. References
63 * must be released with omap_gem_unpin().
64 */
65 dma_addr_t dma_addr;
66
67 /**
68 * # of users of dma_addr
69 */
70 u32 dma_addr_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_is_contiguous(struct omap_gem_object *omap_obj)
152{
153 if (omap_obj->flags & OMAP_BO_MEM_DMA_API)
154 return true;
155
156 if ((omap_obj->flags & OMAP_BO_MEM_DMABUF) && omap_obj->sgt->nents == 1)
157 return true;
158
159 return false;
160}
161
162/* -----------------------------------------------------------------------------
163 * Eviction
164 */
165
166static void omap_gem_evict_entry(struct drm_gem_object *obj,
167 enum tiler_fmt fmt, struct omap_drm_usergart_entry *entry)
168{
169 struct omap_gem_object *omap_obj = to_omap_bo(obj);
170 struct omap_drm_private *priv = obj->dev->dev_private;
171 int n = priv->usergart[fmt].height;
172 size_t size = PAGE_SIZE * n;
173 loff_t off = omap_gem_mmap_offset(obj) +
174 (entry->obj_pgoff << PAGE_SHIFT);
175 const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
176
177 if (m > 1) {
178 int i;
179 /* if stride > than PAGE_SIZE then sparse mapping: */
180 for (i = n; i > 0; i--) {
181 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
182 off, PAGE_SIZE, 1);
183 off += PAGE_SIZE * m;
184 }
185 } else {
186 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
187 off, size, 1);
188 }
189
190 entry->obj = NULL;
191}
192
193/* Evict a buffer from usergart, if it is mapped there */
194static void omap_gem_evict(struct drm_gem_object *obj)
195{
196 struct omap_gem_object *omap_obj = to_omap_bo(obj);
197 struct omap_drm_private *priv = obj->dev->dev_private;
198
199 if (omap_obj->flags & OMAP_BO_TILED) {
200 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
201 int i;
202
203 for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
204 struct omap_drm_usergart_entry *entry =
205 &priv->usergart[fmt].entry[i];
206
207 if (entry->obj == obj)
208 omap_gem_evict_entry(obj, fmt, entry);
209 }
210 }
211}
212
213/* -----------------------------------------------------------------------------
214 * Page Management
215 */
216
217/*
218 * Ensure backing pages are allocated. Must be called with the omap_obj.lock
219 * held.
220 */
221static int omap_gem_attach_pages(struct drm_gem_object *obj)
222{
223 struct drm_device *dev = obj->dev;
224 struct omap_gem_object *omap_obj = to_omap_bo(obj);
225 struct page **pages;
226 int npages = obj->size >> PAGE_SHIFT;
227 int i, ret;
228 dma_addr_t *addrs;
229
230 lockdep_assert_held(&omap_obj->lock);
231
232 /*
233 * If not using shmem (in which case backing pages don't need to be
234 * allocated) or if pages are already allocated we're done.
235 */
236 if (!(omap_obj->flags & OMAP_BO_MEM_SHMEM) || omap_obj->pages)
237 return 0;
238
239 pages = drm_gem_get_pages(obj);
240 if (IS_ERR(pages)) {
241 dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
242 return PTR_ERR(pages);
243 }
244
245 /* for non-cached buffers, ensure the new pages are clean because
246 * DSS, GPU, etc. are not cache coherent:
247 */
248 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
249 addrs = kmalloc_array(npages, sizeof(*addrs), GFP_KERNEL);
250 if (!addrs) {
251 ret = -ENOMEM;
252 goto free_pages;
253 }
254
255 for (i = 0; i < npages; i++) {
256 addrs[i] = dma_map_page(dev->dev, pages[i],
257 0, PAGE_SIZE, DMA_TO_DEVICE);
258
259 if (dma_mapping_error(dev->dev, addrs[i])) {
260 dev_warn(dev->dev,
261 "%s: failed to map page\n", __func__);
262
263 for (i = i - 1; i >= 0; --i) {
264 dma_unmap_page(dev->dev, addrs[i],
265 PAGE_SIZE, DMA_TO_DEVICE);
266 }
267
268 ret = -ENOMEM;
269 goto free_addrs;
270 }
271 }
272 } else {
273 addrs = kcalloc(npages, sizeof(*addrs), GFP_KERNEL);
274 if (!addrs) {
275 ret = -ENOMEM;
276 goto free_pages;
277 }
278 }
279
280 omap_obj->dma_addrs = addrs;
281 omap_obj->pages = pages;
282
283 return 0;
284
285free_addrs:
286 kfree(addrs);
287free_pages:
288 drm_gem_put_pages(obj, pages, true, false);
289
290 return ret;
291}
292
293/* Release backing pages. Must be called with the omap_obj.lock held. */
294static void omap_gem_detach_pages(struct drm_gem_object *obj)
295{
296 struct omap_gem_object *omap_obj = to_omap_bo(obj);
297 unsigned int npages = obj->size >> PAGE_SHIFT;
298 unsigned int i;
299
300 lockdep_assert_held(&omap_obj->lock);
301
302 for (i = 0; i < npages; i++) {
303 if (omap_obj->dma_addrs[i])
304 dma_unmap_page(obj->dev->dev, omap_obj->dma_addrs[i],
305 PAGE_SIZE, DMA_TO_DEVICE);
306 }
307
308 kfree(omap_obj->dma_addrs);
309 omap_obj->dma_addrs = NULL;
310
311 drm_gem_put_pages(obj, omap_obj->pages, true, false);
312 omap_obj->pages = NULL;
313}
314
315/* get buffer flags */
316u32 omap_gem_flags(struct drm_gem_object *obj)
317{
318 return to_omap_bo(obj)->flags;
319}
320
321/** get mmap size */
322size_t omap_gem_mmap_size(struct drm_gem_object *obj)
323{
324 struct omap_gem_object *omap_obj = to_omap_bo(obj);
325 size_t size = obj->size;
326
327 if (omap_obj->flags & OMAP_BO_TILED) {
328 /* for tiled buffers, the virtual size has stride rounded up
329 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
330 * 32kb later!). But we don't back the entire buffer with
331 * pages, only the valid picture part.. so need to adjust for
332 * this in the size used to mmap and generate mmap offset
333 */
334 size = tiler_vsize(gem2fmt(omap_obj->flags),
335 omap_obj->width, omap_obj->height);
336 }
337
338 return size;
339}
340
341/* -----------------------------------------------------------------------------
342 * Fault Handling
343 */
344
345/* Normal handling for the case of faulting in non-tiled buffers */
346static vm_fault_t omap_gem_fault_1d(struct drm_gem_object *obj,
347 struct vm_area_struct *vma, struct vm_fault *vmf)
348{
349 struct omap_gem_object *omap_obj = to_omap_bo(obj);
350 unsigned long pfn;
351 pgoff_t pgoff;
352
353 /* We don't use vmf->pgoff since that has the fake offset: */
354 pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
355
356 if (omap_obj->pages) {
357 omap_gem_cpu_sync_page(obj, pgoff);
358 pfn = page_to_pfn(omap_obj->pages[pgoff]);
359 } else {
360 BUG_ON(!omap_gem_is_contiguous(omap_obj));
361 pfn = (omap_obj->dma_addr >> PAGE_SHIFT) + pgoff;
362 }
363
364 VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
365 pfn, pfn << PAGE_SHIFT);
366
367 return vmf_insert_mixed(vma, vmf->address,
368 __pfn_to_pfn_t(pfn, PFN_DEV));
369}
370
371/* Special handling for the case of faulting in 2d tiled buffers */
372static vm_fault_t omap_gem_fault_2d(struct drm_gem_object *obj,
373 struct vm_area_struct *vma, struct vm_fault *vmf)
374{
375 struct omap_gem_object *omap_obj = to_omap_bo(obj);
376 struct omap_drm_private *priv = obj->dev->dev_private;
377 struct omap_drm_usergart_entry *entry;
378 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
379 struct page *pages[64]; /* XXX is this too much to have on stack? */
380 unsigned long pfn;
381 pgoff_t pgoff, base_pgoff;
382 unsigned long vaddr;
383 int i, err, slots;
384 vm_fault_t ret = VM_FAULT_NOPAGE;
385
386 /*
387 * Note the height of the slot is also equal to the number of pages
388 * that need to be mapped in to fill 4kb wide CPU page. If the slot
389 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
390 */
391 const int n = priv->usergart[fmt].height;
392 const int n_shift = priv->usergart[fmt].height_shift;
393
394 /*
395 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
396 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
397 * into account in some of the math, so figure out virtual stride
398 * in pages
399 */
400 const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
401
402 /* We don't use vmf->pgoff since that has the fake offset: */
403 pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
404
405 /*
406 * Actual address we start mapping at is rounded down to previous slot
407 * boundary in the y direction:
408 */
409 base_pgoff = round_down(pgoff, m << n_shift);
410
411 /* figure out buffer width in slots */
412 slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
413
414 vaddr = vmf->address - ((pgoff - base_pgoff) << PAGE_SHIFT);
415
416 entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
417
418 /* evict previous buffer using this usergart entry, if any: */
419 if (entry->obj)
420 omap_gem_evict_entry(entry->obj, fmt, entry);
421
422 entry->obj = obj;
423 entry->obj_pgoff = base_pgoff;
424
425 /* now convert base_pgoff to phys offset from virt offset: */
426 base_pgoff = (base_pgoff >> n_shift) * slots;
427
428 /* for wider-than 4k.. figure out which part of the slot-row we want: */
429 if (m > 1) {
430 int off = pgoff % m;
431 entry->obj_pgoff += off;
432 base_pgoff /= m;
433 slots = min(slots - (off << n_shift), n);
434 base_pgoff += off << n_shift;
435 vaddr += off << PAGE_SHIFT;
436 }
437
438 /*
439 * Map in pages. Beyond the valid pixel part of the buffer, we set
440 * pages[i] to NULL to get a dummy page mapped in.. if someone
441 * reads/writes it they will get random/undefined content, but at
442 * least it won't be corrupting whatever other random page used to
443 * be mapped in, or other undefined behavior.
444 */
445 memcpy(pages, &omap_obj->pages[base_pgoff],
446 sizeof(struct page *) * slots);
447 memset(pages + slots, 0,
448 sizeof(struct page *) * (n - slots));
449
450 err = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
451 if (err) {
452 ret = vmf_error(err);
453 dev_err(obj->dev->dev, "failed to pin: %d\n", err);
454 return ret;
455 }
456
457 pfn = entry->dma_addr >> PAGE_SHIFT;
458
459 VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
460 pfn, pfn << PAGE_SHIFT);
461
462 for (i = n; i > 0; i--) {
463 ret = vmf_insert_mixed(vma,
464 vaddr, __pfn_to_pfn_t(pfn, PFN_DEV));
465 if (ret & VM_FAULT_ERROR)
466 break;
467 pfn += priv->usergart[fmt].stride_pfn;
468 vaddr += PAGE_SIZE * m;
469 }
470
471 /* simple round-robin: */
472 priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
473 % NUM_USERGART_ENTRIES;
474
475 return ret;
476}
477
478/**
479 * omap_gem_fault - pagefault handler for GEM objects
480 * @vmf: fault detail
481 *
482 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
483 * does most of the work for us including the actual map/unmap calls
484 * but we need to do the actual page work.
485 *
486 * The VMA was set up by GEM. In doing so it also ensured that the
487 * vma->vm_private_data points to the GEM object that is backing this
488 * mapping.
489 */
490vm_fault_t omap_gem_fault(struct vm_fault *vmf)
491{
492 struct vm_area_struct *vma = vmf->vma;
493 struct drm_gem_object *obj = vma->vm_private_data;
494 struct omap_gem_object *omap_obj = to_omap_bo(obj);
495 int err;
496 vm_fault_t ret;
497
498 /* Make sure we don't parallel update on a fault, nor move or remove
499 * something from beneath our feet
500 */
501 mutex_lock(&omap_obj->lock);
502
503 /* if a shmem backed object, make sure we have pages attached now */
504 err = omap_gem_attach_pages(obj);
505 if (err) {
506 ret = vmf_error(err);
507 goto fail;
508 }
509
510 /* where should we do corresponding put_pages().. we are mapping
511 * the original page, rather than thru a GART, so we can't rely
512 * on eviction to trigger this. But munmap() or all mappings should
513 * probably trigger put_pages()?
514 */
515
516 if (omap_obj->flags & OMAP_BO_TILED)
517 ret = omap_gem_fault_2d(obj, vma, vmf);
518 else
519 ret = omap_gem_fault_1d(obj, vma, vmf);
520
521
522fail:
523 mutex_unlock(&omap_obj->lock);
524 return ret;
525}
526
527/** We override mainly to fix up some of the vm mapping flags.. */
528int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
529{
530 int ret;
531
532 ret = drm_gem_mmap(filp, vma);
533 if (ret) {
534 DBG("mmap failed: %d", ret);
535 return ret;
536 }
537
538 return omap_gem_mmap_obj(vma->vm_private_data, vma);
539}
540
541int omap_gem_mmap_obj(struct drm_gem_object *obj,
542 struct vm_area_struct *vma)
543{
544 struct omap_gem_object *omap_obj = to_omap_bo(obj);
545
546 vma->vm_flags &= ~VM_PFNMAP;
547 vma->vm_flags |= VM_MIXEDMAP;
548
549 if (omap_obj->flags & OMAP_BO_WC) {
550 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
551 } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
552 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
553 } else {
554 /*
555 * We do have some private objects, at least for scanout buffers
556 * on hardware without DMM/TILER. But these are allocated write-
557 * combine
558 */
559 if (WARN_ON(!obj->filp))
560 return -EINVAL;
561
562 /*
563 * Shunt off cached objs to shmem file so they have their own
564 * address_space (so unmap_mapping_range does what we want,
565 * in particular in the case of mmap'd dmabufs)
566 */
567 fput(vma->vm_file);
568 vma->vm_pgoff = 0;
569 vma->vm_file = get_file(obj->filp);
570
571 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
572 }
573
574 return 0;
575}
576
577/* -----------------------------------------------------------------------------
578 * Dumb Buffers
579 */
580
581/**
582 * omap_gem_dumb_create - create a dumb buffer
583 * @drm_file: our client file
584 * @dev: our device
585 * @args: the requested arguments copied from userspace
586 *
587 * Allocate a buffer suitable for use for a frame buffer of the
588 * form described by user space. Give userspace a handle by which
589 * to reference it.
590 */
591int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
592 struct drm_mode_create_dumb *args)
593{
594 union omap_gem_size gsize;
595
596 args->pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
597
598 args->size = PAGE_ALIGN(args->pitch * args->height);
599
600 gsize = (union omap_gem_size){
601 .bytes = args->size,
602 };
603
604 return omap_gem_new_handle(dev, file, gsize,
605 OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
606}
607
608/**
609 * omap_gem_dumb_map - buffer mapping for dumb interface
610 * @file: our drm client file
611 * @dev: drm device
612 * @handle: GEM handle to the object (from dumb_create)
613 *
614 * Do the necessary setup to allow the mapping of the frame buffer
615 * into user memory. We don't have to do much here at the moment.
616 */
617int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
618 u32 handle, u64 *offset)
619{
620 struct drm_gem_object *obj;
621 int ret = 0;
622
623 /* GEM does all our handle to object mapping */
624 obj = drm_gem_object_lookup(file, handle);
625 if (obj == NULL) {
626 ret = -ENOENT;
627 goto fail;
628 }
629
630 *offset = omap_gem_mmap_offset(obj);
631
632 drm_gem_object_put_unlocked(obj);
633
634fail:
635 return ret;
636}
637
638#ifdef CONFIG_DRM_FBDEV_EMULATION
639/* Set scrolling position. This allows us to implement fast scrolling
640 * for console.
641 *
642 * Call only from non-atomic contexts.
643 */
644int omap_gem_roll(struct drm_gem_object *obj, u32 roll)
645{
646 struct omap_gem_object *omap_obj = to_omap_bo(obj);
647 u32 npages = obj->size >> PAGE_SHIFT;
648 int ret = 0;
649
650 if (roll > npages) {
651 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
652 return -EINVAL;
653 }
654
655 omap_obj->roll = roll;
656
657 mutex_lock(&omap_obj->lock);
658
659 /* if we aren't mapped yet, we don't need to do anything */
660 if (omap_obj->block) {
661 ret = omap_gem_attach_pages(obj);
662 if (ret)
663 goto fail;
664
665 ret = tiler_pin(omap_obj->block, omap_obj->pages, npages,
666 roll, true);
667 if (ret)
668 dev_err(obj->dev->dev, "could not repin: %d\n", ret);
669 }
670
671fail:
672 mutex_unlock(&omap_obj->lock);
673
674 return ret;
675}
676#endif
677
678/* -----------------------------------------------------------------------------
679 * Memory Management & DMA Sync
680 */
681
682/*
683 * shmem buffers that are mapped cached are not coherent.
684 *
685 * We keep track of dirty pages using page faulting to perform cache management.
686 * When a page is mapped to the CPU in read/write mode the device can't access
687 * it and omap_obj->dma_addrs[i] is NULL. When a page is mapped to the device
688 * the omap_obj->dma_addrs[i] is set to the DMA address, and the page is
689 * unmapped from the CPU.
690 */
691static inline bool omap_gem_is_cached_coherent(struct drm_gem_object *obj)
692{
693 struct omap_gem_object *omap_obj = to_omap_bo(obj);
694
695 return !((omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
696 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED));
697}
698
699/* Sync the buffer for CPU access.. note pages should already be
700 * attached, ie. omap_gem_get_pages()
701 */
702void omap_gem_cpu_sync_page(struct drm_gem_object *obj, int pgoff)
703{
704 struct drm_device *dev = obj->dev;
705 struct omap_gem_object *omap_obj = to_omap_bo(obj);
706
707 if (omap_gem_is_cached_coherent(obj))
708 return;
709
710 if (omap_obj->dma_addrs[pgoff]) {
711 dma_unmap_page(dev->dev, omap_obj->dma_addrs[pgoff],
712 PAGE_SIZE, DMA_TO_DEVICE);
713 omap_obj->dma_addrs[pgoff] = 0;
714 }
715}
716
717/* sync the buffer for DMA access */
718void omap_gem_dma_sync_buffer(struct drm_gem_object *obj,
719 enum dma_data_direction dir)
720{
721 struct drm_device *dev = obj->dev;
722 struct omap_gem_object *omap_obj = to_omap_bo(obj);
723 int i, npages = obj->size >> PAGE_SHIFT;
724 struct page **pages = omap_obj->pages;
725 bool dirty = false;
726
727 if (omap_gem_is_cached_coherent(obj))
728 return;
729
730 for (i = 0; i < npages; i++) {
731 if (!omap_obj->dma_addrs[i]) {
732 dma_addr_t addr;
733
734 addr = dma_map_page(dev->dev, pages[i], 0,
735 PAGE_SIZE, dir);
736 if (dma_mapping_error(dev->dev, addr)) {
737 dev_warn(dev->dev, "%s: failed to map page\n",
738 __func__);
739 break;
740 }
741
742 dirty = true;
743 omap_obj->dma_addrs[i] = addr;
744 }
745 }
746
747 if (dirty) {
748 unmap_mapping_range(obj->filp->f_mapping, 0,
749 omap_gem_mmap_size(obj), 1);
750 }
751}
752
753/**
754 * omap_gem_pin() - Pin a GEM object in memory
755 * @obj: the GEM object
756 * @dma_addr: the DMA address
757 *
758 * Pin the given GEM object in memory and fill the dma_addr pointer with the
759 * object's DMA address. If the buffer is not physically contiguous it will be
760 * remapped through the TILER to provide a contiguous view.
761 *
762 * Pins are reference-counted, calling this function multiple times is allowed
763 * as long the corresponding omap_gem_unpin() calls are balanced.
764 *
765 * Return 0 on success or a negative error code otherwise.
766 */
767int omap_gem_pin(struct drm_gem_object *obj, dma_addr_t *dma_addr)
768{
769 struct omap_drm_private *priv = obj->dev->dev_private;
770 struct omap_gem_object *omap_obj = to_omap_bo(obj);
771 int ret = 0;
772
773 mutex_lock(&omap_obj->lock);
774
775 if (!omap_gem_is_contiguous(omap_obj) && priv->has_dmm) {
776 if (omap_obj->dma_addr_cnt == 0) {
777 u32 npages = obj->size >> PAGE_SHIFT;
778 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
779 struct tiler_block *block;
780
781 BUG_ON(omap_obj->block);
782
783 ret = omap_gem_attach_pages(obj);
784 if (ret)
785 goto fail;
786
787 if (omap_obj->flags & OMAP_BO_TILED) {
788 block = tiler_reserve_2d(fmt,
789 omap_obj->width,
790 omap_obj->height, 0);
791 } else {
792 block = tiler_reserve_1d(obj->size);
793 }
794
795 if (IS_ERR(block)) {
796 ret = PTR_ERR(block);
797 dev_err(obj->dev->dev,
798 "could not remap: %d (%d)\n", ret, fmt);
799 goto fail;
800 }
801
802 /* TODO: enable async refill.. */
803 ret = tiler_pin(block, omap_obj->pages, npages,
804 omap_obj->roll, true);
805 if (ret) {
806 tiler_release(block);
807 dev_err(obj->dev->dev,
808 "could not pin: %d\n", ret);
809 goto fail;
810 }
811
812 omap_obj->dma_addr = tiler_ssptr(block);
813 omap_obj->block = block;
814
815 DBG("got dma address: %pad", &omap_obj->dma_addr);
816 }
817
818 omap_obj->dma_addr_cnt++;
819
820 *dma_addr = omap_obj->dma_addr;
821 } else if (omap_gem_is_contiguous(omap_obj)) {
822 *dma_addr = omap_obj->dma_addr;
823 } else {
824 ret = -EINVAL;
825 goto fail;
826 }
827
828fail:
829 mutex_unlock(&omap_obj->lock);
830
831 return ret;
832}
833
834/**
835 * omap_gem_unpin() - Unpin a GEM object from memory
836 * @obj: the GEM object
837 *
838 * Unpin the given GEM object previously pinned with omap_gem_pin(). Pins are
839 * reference-counted, the actualy unpin will only be performed when the number
840 * of calls to this function matches the number of calls to omap_gem_pin().
841 */
842void omap_gem_unpin(struct drm_gem_object *obj)
843{
844 struct omap_gem_object *omap_obj = to_omap_bo(obj);
845 int ret;
846
847 mutex_lock(&omap_obj->lock);
848
849 if (omap_obj->dma_addr_cnt > 0) {
850 omap_obj->dma_addr_cnt--;
851 if (omap_obj->dma_addr_cnt == 0) {
852 ret = tiler_unpin(omap_obj->block);
853 if (ret) {
854 dev_err(obj->dev->dev,
855 "could not unpin pages: %d\n", ret);
856 }
857 ret = tiler_release(omap_obj->block);
858 if (ret) {
859 dev_err(obj->dev->dev,
860 "could not release unmap: %d\n", ret);
861 }
862 omap_obj->dma_addr = 0;
863 omap_obj->block = NULL;
864 }
865 }
866
867 mutex_unlock(&omap_obj->lock);
868}
869
870/* Get rotated scanout address (only valid if already pinned), at the
871 * specified orientation and x,y offset from top-left corner of buffer
872 * (only valid for tiled 2d buffers)
873 */
874int omap_gem_rotated_dma_addr(struct drm_gem_object *obj, u32 orient,
875 int x, int y, dma_addr_t *dma_addr)
876{
877 struct omap_gem_object *omap_obj = to_omap_bo(obj);
878 int ret = -EINVAL;
879
880 mutex_lock(&omap_obj->lock);
881
882 if ((omap_obj->dma_addr_cnt > 0) && omap_obj->block &&
883 (omap_obj->flags & OMAP_BO_TILED)) {
884 *dma_addr = tiler_tsptr(omap_obj->block, orient, x, y);
885 ret = 0;
886 }
887
888 mutex_unlock(&omap_obj->lock);
889
890 return ret;
891}
892
893/* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
894int omap_gem_tiled_stride(struct drm_gem_object *obj, u32 orient)
895{
896 struct omap_gem_object *omap_obj = to_omap_bo(obj);
897 int ret = -EINVAL;
898 if (omap_obj->flags & OMAP_BO_TILED)
899 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
900 return ret;
901}
902
903/* if !remap, and we don't have pages backing, then fail, rather than
904 * increasing the pin count (which we don't really do yet anyways,
905 * because we don't support swapping pages back out). And 'remap'
906 * might not be quite the right name, but I wanted to keep it working
907 * similarly to omap_gem_pin(). Note though that mutex is not
908 * aquired if !remap (because this can be called in atomic ctxt),
909 * but probably omap_gem_unpin() should be changed to work in the
910 * same way. If !remap, a matching omap_gem_put_pages() call is not
911 * required (and should not be made).
912 */
913int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
914 bool remap)
915{
916 struct omap_gem_object *omap_obj = to_omap_bo(obj);
917 int ret = 0;
918
919 mutex_lock(&omap_obj->lock);
920
921 if (remap) {
922 ret = omap_gem_attach_pages(obj);
923 if (ret)
924 goto unlock;
925 }
926
927 if (!omap_obj->pages) {
928 ret = -ENOMEM;
929 goto unlock;
930 }
931
932 *pages = omap_obj->pages;
933
934unlock:
935 mutex_unlock(&omap_obj->lock);
936
937 return ret;
938}
939
940/* release pages when DMA no longer being performed */
941int omap_gem_put_pages(struct drm_gem_object *obj)
942{
943 /* do something here if we dynamically attach/detach pages.. at
944 * least they would no longer need to be pinned if everyone has
945 * released the pages..
946 */
947 return 0;
948}
949
950#ifdef CONFIG_DRM_FBDEV_EMULATION
951/*
952 * Get kernel virtual address for CPU access.. this more or less only
953 * exists for omap_fbdev.
954 */
955void *omap_gem_vaddr(struct drm_gem_object *obj)
956{
957 struct omap_gem_object *omap_obj = to_omap_bo(obj);
958 void *vaddr;
959 int ret;
960
961 mutex_lock(&omap_obj->lock);
962
963 if (!omap_obj->vaddr) {
964 ret = omap_gem_attach_pages(obj);
965 if (ret) {
966 vaddr = ERR_PTR(ret);
967 goto unlock;
968 }
969
970 omap_obj->vaddr = vmap(omap_obj->pages, obj->size >> PAGE_SHIFT,
971 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
972 }
973
974 vaddr = omap_obj->vaddr;
975
976unlock:
977 mutex_unlock(&omap_obj->lock);
978 return vaddr;
979}
980#endif
981
982/* -----------------------------------------------------------------------------
983 * Power Management
984 */
985
986#ifdef CONFIG_PM
987/* re-pin objects in DMM in resume path: */
988int omap_gem_resume(struct drm_device *dev)
989{
990 struct omap_drm_private *priv = dev->dev_private;
991 struct omap_gem_object *omap_obj;
992 int ret = 0;
993
994 mutex_lock(&priv->list_lock);
995 list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
996 if (omap_obj->block) {
997 struct drm_gem_object *obj = &omap_obj->base;
998 u32 npages = obj->size >> PAGE_SHIFT;
999
1000 WARN_ON(!omap_obj->pages); /* this can't happen */
1001 ret = tiler_pin(omap_obj->block,
1002 omap_obj->pages, npages,
1003 omap_obj->roll, true);
1004 if (ret) {
1005 dev_err(dev->dev, "could not repin: %d\n", ret);
1006 goto done;
1007 }
1008 }
1009 }
1010
1011done:
1012 mutex_unlock(&priv->list_lock);
1013 return ret;
1014}
1015#endif
1016
1017/* -----------------------------------------------------------------------------
1018 * DebugFS
1019 */
1020
1021#ifdef CONFIG_DEBUG_FS
1022void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1023{
1024 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1025 u64 off;
1026
1027 off = drm_vma_node_start(&obj->vma_node);
1028
1029 mutex_lock(&omap_obj->lock);
1030
1031 seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1032 omap_obj->flags, obj->name, kref_read(&obj->refcount),
1033 off, &omap_obj->dma_addr, omap_obj->dma_addr_cnt,
1034 omap_obj->vaddr, omap_obj->roll);
1035
1036 if (omap_obj->flags & OMAP_BO_TILED) {
1037 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1038 if (omap_obj->block) {
1039 struct tcm_area *area = &omap_obj->block->area;
1040 seq_printf(m, " (%dx%d, %dx%d)",
1041 area->p0.x, area->p0.y,
1042 area->p1.x, area->p1.y);
1043 }
1044 } else {
1045 seq_printf(m, " %zu", obj->size);
1046 }
1047
1048 mutex_unlock(&omap_obj->lock);
1049
1050 seq_printf(m, "\n");
1051}
1052
1053void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1054{
1055 struct omap_gem_object *omap_obj;
1056 int count = 0;
1057 size_t size = 0;
1058
1059 list_for_each_entry(omap_obj, list, mm_list) {
1060 struct drm_gem_object *obj = &omap_obj->base;
1061 seq_printf(m, " ");
1062 omap_gem_describe(obj, m);
1063 count++;
1064 size += obj->size;
1065 }
1066
1067 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1068}
1069#endif
1070
1071/* -----------------------------------------------------------------------------
1072 * Constructor & Destructor
1073 */
1074
1075void omap_gem_free_object(struct drm_gem_object *obj)
1076{
1077 struct drm_device *dev = obj->dev;
1078 struct omap_drm_private *priv = dev->dev_private;
1079 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1080
1081 omap_gem_evict(obj);
1082
1083 mutex_lock(&priv->list_lock);
1084 list_del(&omap_obj->mm_list);
1085 mutex_unlock(&priv->list_lock);
1086
1087 /*
1088 * We own the sole reference to the object at this point, but to keep
1089 * lockdep happy, we must still take the omap_obj_lock to call
1090 * omap_gem_detach_pages(). This should hardly make any difference as
1091 * there can't be any lock contention.
1092 */
1093 mutex_lock(&omap_obj->lock);
1094
1095 /* The object should not be pinned. */
1096 WARN_ON(omap_obj->dma_addr_cnt > 0);
1097
1098 if (omap_obj->pages) {
1099 if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1100 kfree(omap_obj->pages);
1101 else
1102 omap_gem_detach_pages(obj);
1103 }
1104
1105 if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1106 dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1107 omap_obj->dma_addr);
1108 } else if (omap_obj->vaddr) {
1109 vunmap(omap_obj->vaddr);
1110 } else if (obj->import_attach) {
1111 drm_prime_gem_destroy(obj, omap_obj->sgt);
1112 }
1113
1114 mutex_unlock(&omap_obj->lock);
1115
1116 drm_gem_object_release(obj);
1117
1118 mutex_destroy(&omap_obj->lock);
1119
1120 kfree(omap_obj);
1121}
1122
1123/* GEM buffer object constructor */
1124struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1125 union omap_gem_size gsize, u32 flags)
1126{
1127 struct omap_drm_private *priv = dev->dev_private;
1128 struct omap_gem_object *omap_obj;
1129 struct drm_gem_object *obj;
1130 struct address_space *mapping;
1131 size_t size;
1132 int ret;
1133
1134 /* Validate the flags and compute the memory and cache flags. */
1135 if (flags & OMAP_BO_TILED) {
1136 if (!priv->usergart) {
1137 dev_err(dev->dev, "Tiled buffers require DMM\n");
1138 return NULL;
1139 }
1140
1141 /*
1142 * Tiled buffers are always shmem paged backed. When they are
1143 * scanned out, they are remapped into DMM/TILER.
1144 */
1145 flags &= ~OMAP_BO_SCANOUT;
1146 flags |= OMAP_BO_MEM_SHMEM;
1147
1148 /*
1149 * Currently don't allow cached buffers. There is some caching
1150 * stuff that needs to be handled better.
1151 */
1152 flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1153 flags |= tiler_get_cpu_cache_flags();
1154 } else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1155 /*
1156 * OMAP_BO_SCANOUT hints that the buffer doesn't need to be
1157 * tiled. However, to lower the pressure on memory allocation,
1158 * use contiguous memory only if no TILER is available.
1159 */
1160 flags |= OMAP_BO_MEM_DMA_API;
1161 } else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1162 /*
1163 * All other buffers not backed by dma_buf are shmem-backed.
1164 */
1165 flags |= OMAP_BO_MEM_SHMEM;
1166 }
1167
1168 /* Allocate the initialize the OMAP GEM object. */
1169 omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1170 if (!omap_obj)
1171 return NULL;
1172
1173 obj = &omap_obj->base;
1174 omap_obj->flags = flags;
1175 mutex_init(&omap_obj->lock);
1176
1177 if (flags & OMAP_BO_TILED) {
1178 /*
1179 * For tiled buffers align dimensions to slot boundaries and
1180 * calculate size based on aligned dimensions.
1181 */
1182 tiler_align(gem2fmt(flags), &gsize.tiled.width,
1183 &gsize.tiled.height);
1184
1185 size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1186 gsize.tiled.height);
1187
1188 omap_obj->width = gsize.tiled.width;
1189 omap_obj->height = gsize.tiled.height;
1190 } else {
1191 size = PAGE_ALIGN(gsize.bytes);
1192 }
1193
1194 /* Initialize the GEM object. */
1195 if (!(flags & OMAP_BO_MEM_SHMEM)) {
1196 drm_gem_private_object_init(dev, obj, size);
1197 } else {
1198 ret = drm_gem_object_init(dev, obj, size);
1199 if (ret)
1200 goto err_free;
1201
1202 mapping = obj->filp->f_mapping;
1203 mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1204 }
1205
1206 /* Allocate memory if needed. */
1207 if (flags & OMAP_BO_MEM_DMA_API) {
1208 omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1209 &omap_obj->dma_addr,
1210 GFP_KERNEL);
1211 if (!omap_obj->vaddr)
1212 goto err_release;
1213 }
1214
1215 mutex_lock(&priv->list_lock);
1216 list_add(&omap_obj->mm_list, &priv->obj_list);
1217 mutex_unlock(&priv->list_lock);
1218
1219 return obj;
1220
1221err_release:
1222 drm_gem_object_release(obj);
1223err_free:
1224 kfree(omap_obj);
1225 return NULL;
1226}
1227
1228struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1229 struct sg_table *sgt)
1230{
1231 struct omap_drm_private *priv = dev->dev_private;
1232 struct omap_gem_object *omap_obj;
1233 struct drm_gem_object *obj;
1234 union omap_gem_size gsize;
1235
1236 /* Without a DMM only physically contiguous buffers can be supported. */
1237 if (sgt->orig_nents != 1 && !priv->has_dmm)
1238 return ERR_PTR(-EINVAL);
1239
1240 gsize.bytes = PAGE_ALIGN(size);
1241 obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1242 if (!obj)
1243 return ERR_PTR(-ENOMEM);
1244
1245 omap_obj = to_omap_bo(obj);
1246
1247 mutex_lock(&omap_obj->lock);
1248
1249 omap_obj->sgt = sgt;
1250
1251 if (sgt->orig_nents == 1) {
1252 omap_obj->dma_addr = sg_dma_address(sgt->sgl);
1253 } else {
1254 /* Create pages list from sgt */
1255 struct sg_page_iter iter;
1256 struct page **pages;
1257 unsigned int npages;
1258 unsigned int i = 0;
1259
1260 npages = DIV_ROUND_UP(size, PAGE_SIZE);
1261 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1262 if (!pages) {
1263 omap_gem_free_object(obj);
1264 obj = ERR_PTR(-ENOMEM);
1265 goto done;
1266 }
1267
1268 omap_obj->pages = pages;
1269
1270 for_each_sg_page(sgt->sgl, &iter, sgt->orig_nents, 0) {
1271 pages[i++] = sg_page_iter_page(&iter);
1272 if (i > npages)
1273 break;
1274 }
1275
1276 if (WARN_ON(i != npages)) {
1277 omap_gem_free_object(obj);
1278 obj = ERR_PTR(-ENOMEM);
1279 goto done;
1280 }
1281 }
1282
1283done:
1284 mutex_unlock(&omap_obj->lock);
1285 return obj;
1286}
1287
1288/* convenience method to construct a GEM buffer object, and userspace handle */
1289int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1290 union omap_gem_size gsize, u32 flags, u32 *handle)
1291{
1292 struct drm_gem_object *obj;
1293 int ret;
1294
1295 obj = omap_gem_new(dev, gsize, flags);
1296 if (!obj)
1297 return -ENOMEM;
1298
1299 ret = drm_gem_handle_create(file, obj, handle);
1300 if (ret) {
1301 omap_gem_free_object(obj);
1302 return ret;
1303 }
1304
1305 /* drop reference from allocate - handle holds it now */
1306 drm_gem_object_put_unlocked(obj);
1307
1308 return 0;
1309}
1310
1311/* -----------------------------------------------------------------------------
1312 * Init & Cleanup
1313 */
1314
1315/* If DMM is used, we need to set some stuff up.. */
1316void omap_gem_init(struct drm_device *dev)
1317{
1318 struct omap_drm_private *priv = dev->dev_private;
1319 struct omap_drm_usergart *usergart;
1320 const enum tiler_fmt fmts[] = {
1321 TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1322 };
1323 int i, j;
1324
1325 if (!dmm_is_available()) {
1326 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1327 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1328 return;
1329 }
1330
1331 usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1332 if (!usergart)
1333 return;
1334
1335 /* reserve 4k aligned/wide regions for userspace mappings: */
1336 for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1337 u16 h = 1, w = PAGE_SIZE >> i;
1338
1339 tiler_align(fmts[i], &w, &h);
1340 /* note: since each region is 1 4kb page wide, and minimum
1341 * number of rows, the height ends up being the same as the
1342 * # of pages in the region
1343 */
1344 usergart[i].height = h;
1345 usergart[i].height_shift = ilog2(h);
1346 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1347 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1348 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1349 struct omap_drm_usergart_entry *entry;
1350 struct tiler_block *block;
1351
1352 entry = &usergart[i].entry[j];
1353 block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1354 if (IS_ERR(block)) {
1355 dev_err(dev->dev,
1356 "reserve failed: %d, %d, %ld\n",
1357 i, j, PTR_ERR(block));
1358 return;
1359 }
1360 entry->dma_addr = tiler_ssptr(block);
1361 entry->block = block;
1362
1363 DBG("%d:%d: %dx%d: dma_addr=%pad stride=%d", i, j, w, h,
1364 &entry->dma_addr,
1365 usergart[i].stride_pfn << PAGE_SHIFT);
1366 }
1367 }
1368
1369 priv->usergart = usergart;
1370 priv->has_dmm = true;
1371}
1372
1373void omap_gem_deinit(struct drm_device *dev)
1374{
1375 struct omap_drm_private *priv = dev->dev_private;
1376
1377 /* I believe we can rely on there being no more outstanding GEM
1378 * objects which could depend on usergart/dmm at this point.
1379 */
1380 kfree(priv->usergart);
1381}