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