Loading...
1/*
2 * drivers/gpu/drm/omapdrm/omap_gem.c
3 *
4 * Copyright (C) 2011 Texas Instruments
5 * Author: Rob Clark <rob.clark@linaro.org>
6 *
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 as published by
9 * the Free Software Foundation.
10 *
11 * This program is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * more details.
15 *
16 * You should have received a copy of the GNU General Public License along with
17 * this program. If not, see <http://www.gnu.org/licenses/>.
18 */
19
20#include <linux/shmem_fs.h>
21#include <linux/spinlock.h>
22#include <linux/pfn_t.h>
23
24#include <drm/drm_vma_manager.h>
25
26#include "omap_drv.h"
27#include "omap_dmm_tiler.h"
28
29/*
30 * GEM buffer object implementation.
31 */
32
33/* note: we use upper 8 bits of flags for driver-internal flags: */
34#define OMAP_BO_MEM_DMA_API 0x01000000 /* memory allocated with the dma_alloc_* API */
35#define OMAP_BO_MEM_SHMEM 0x02000000 /* memory allocated through shmem backing */
36#define OMAP_BO_MEM_DMABUF 0x08000000 /* memory imported from a dmabuf */
37
38struct omap_gem_object {
39 struct drm_gem_object base;
40
41 struct list_head mm_list;
42
43 uint32_t flags;
44
45 /** width/height for tiled formats (rounded up to slot boundaries) */
46 uint16_t width, height;
47
48 /** roll applied when mapping to DMM */
49 uint32_t roll;
50
51 /**
52 * paddr contains the buffer DMA address. It is valid for
53 *
54 * - buffers allocated through the DMA mapping API (with the
55 * OMAP_BO_MEM_DMA_API flag set)
56 *
57 * - buffers imported from dmabuf (with the OMAP_BO_MEM_DMABUF flag set)
58 * if they are physically contiguous (when sgt->orig_nents == 1)
59 *
60 * - buffers mapped through the TILER when paddr_cnt is not zero, in
61 * which case the DMA address points to the TILER aperture
62 *
63 * Physically contiguous buffers have their DMA address equal to the
64 * physical address as we don't remap those buffers through the TILER.
65 *
66 * Buffers mapped to the TILER have their DMA address pointing to the
67 * TILER aperture. As TILER mappings are refcounted (through paddr_cnt)
68 * the DMA address must be accessed through omap_get_get_paddr() to
69 * ensure that the mapping won't disappear unexpectedly. References must
70 * be released with omap_gem_put_paddr().
71 */
72 dma_addr_t paddr;
73
74 /**
75 * # of users of paddr
76 */
77 uint32_t paddr_cnt;
78
79 /**
80 * If the buffer has been imported from a dmabuf the OMAP_DB_DMABUF flag
81 * is set and the sgt field is valid.
82 */
83 struct sg_table *sgt;
84
85 /**
86 * tiler block used when buffer is remapped in DMM/TILER.
87 */
88 struct tiler_block *block;
89
90 /**
91 * Array of backing pages, if allocated. Note that pages are never
92 * allocated for buffers originally allocated from contiguous memory
93 */
94 struct page **pages;
95
96 /** addresses corresponding to pages in above array */
97 dma_addr_t *addrs;
98
99 /**
100 * Virtual address, if mapped.
101 */
102 void *vaddr;
103
104 /**
105 * sync-object allocated on demand (if needed)
106 *
107 * Per-buffer sync-object for tracking pending and completed hw/dma
108 * read and write operations.
109 */
110 struct {
111 uint32_t write_pending;
112 uint32_t write_complete;
113 uint32_t read_pending;
114 uint32_t read_complete;
115 } *sync;
116};
117
118#define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
119
120/* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
121 * not necessarily pinned in TILER all the time, and (b) when they are
122 * they are not necessarily page aligned, we reserve one or more small
123 * regions in each of the 2d containers to use as a user-GART where we
124 * can create a second page-aligned mapping of parts of the buffer
125 * being accessed from userspace.
126 *
127 * Note that we could optimize slightly when we know that multiple
128 * tiler containers are backed by the same PAT.. but I'll leave that
129 * for later..
130 */
131#define NUM_USERGART_ENTRIES 2
132struct omap_drm_usergart_entry {
133 struct tiler_block *block; /* the reserved tiler block */
134 dma_addr_t paddr;
135 struct drm_gem_object *obj; /* the current pinned obj */
136 pgoff_t obj_pgoff; /* page offset of obj currently
137 mapped in */
138};
139
140struct omap_drm_usergart {
141 struct omap_drm_usergart_entry entry[NUM_USERGART_ENTRIES];
142 int height; /* height in rows */
143 int height_shift; /* ilog2(height in rows) */
144 int slot_shift; /* ilog2(width per slot) */
145 int stride_pfn; /* stride in pages */
146 int last; /* index of last used entry */
147};
148
149/* -----------------------------------------------------------------------------
150 * Helpers
151 */
152
153/** get mmap offset */
154static uint64_t mmap_offset(struct drm_gem_object *obj)
155{
156 struct drm_device *dev = obj->dev;
157 int ret;
158 size_t size;
159
160 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
161
162 /* Make it mmapable */
163 size = omap_gem_mmap_size(obj);
164 ret = drm_gem_create_mmap_offset_size(obj, size);
165 if (ret) {
166 dev_err(dev->dev, "could not allocate mmap offset\n");
167 return 0;
168 }
169
170 return drm_vma_node_offset_addr(&obj->vma_node);
171}
172
173static bool is_contiguous(struct omap_gem_object *omap_obj)
174{
175 if (omap_obj->flags & OMAP_BO_MEM_DMA_API)
176 return true;
177
178 if ((omap_obj->flags & OMAP_BO_MEM_DMABUF) && omap_obj->sgt->nents == 1)
179 return true;
180
181 return false;
182}
183
184/* -----------------------------------------------------------------------------
185 * Eviction
186 */
187
188static void evict_entry(struct drm_gem_object *obj,
189 enum tiler_fmt fmt, struct omap_drm_usergart_entry *entry)
190{
191 struct omap_gem_object *omap_obj = to_omap_bo(obj);
192 struct omap_drm_private *priv = obj->dev->dev_private;
193 int n = priv->usergart[fmt].height;
194 size_t size = PAGE_SIZE * n;
195 loff_t off = mmap_offset(obj) +
196 (entry->obj_pgoff << PAGE_SHIFT);
197 const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
198
199 if (m > 1) {
200 int i;
201 /* if stride > than PAGE_SIZE then sparse mapping: */
202 for (i = n; i > 0; i--) {
203 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
204 off, PAGE_SIZE, 1);
205 off += PAGE_SIZE * m;
206 }
207 } else {
208 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
209 off, size, 1);
210 }
211
212 entry->obj = NULL;
213}
214
215/* Evict a buffer from usergart, if it is mapped there */
216static void evict(struct drm_gem_object *obj)
217{
218 struct omap_gem_object *omap_obj = to_omap_bo(obj);
219 struct omap_drm_private *priv = obj->dev->dev_private;
220
221 if (omap_obj->flags & OMAP_BO_TILED) {
222 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
223 int i;
224
225 for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
226 struct omap_drm_usergart_entry *entry =
227 &priv->usergart[fmt].entry[i];
228
229 if (entry->obj == obj)
230 evict_entry(obj, fmt, entry);
231 }
232 }
233}
234
235/* -----------------------------------------------------------------------------
236 * Page Management
237 */
238
239/** ensure backing pages are allocated */
240static int omap_gem_attach_pages(struct drm_gem_object *obj)
241{
242 struct drm_device *dev = obj->dev;
243 struct omap_gem_object *omap_obj = to_omap_bo(obj);
244 struct page **pages;
245 int npages = obj->size >> PAGE_SHIFT;
246 int i, ret;
247 dma_addr_t *addrs;
248
249 WARN_ON(omap_obj->pages);
250
251 pages = drm_gem_get_pages(obj);
252 if (IS_ERR(pages)) {
253 dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
254 return PTR_ERR(pages);
255 }
256
257 /* for non-cached buffers, ensure the new pages are clean because
258 * DSS, GPU, etc. are not cache coherent:
259 */
260 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
261 addrs = kmalloc(npages * sizeof(*addrs), GFP_KERNEL);
262 if (!addrs) {
263 ret = -ENOMEM;
264 goto free_pages;
265 }
266
267 for (i = 0; i < npages; i++) {
268 addrs[i] = dma_map_page(dev->dev, pages[i],
269 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
270
271 if (dma_mapping_error(dev->dev, addrs[i])) {
272 dev_warn(dev->dev,
273 "%s: failed to map page\n", __func__);
274
275 for (i = i - 1; i >= 0; --i) {
276 dma_unmap_page(dev->dev, addrs[i],
277 PAGE_SIZE, DMA_BIDIRECTIONAL);
278 }
279
280 ret = -ENOMEM;
281 goto free_addrs;
282 }
283 }
284 } else {
285 addrs = kzalloc(npages * sizeof(*addrs), GFP_KERNEL);
286 if (!addrs) {
287 ret = -ENOMEM;
288 goto free_pages;
289 }
290 }
291
292 omap_obj->addrs = addrs;
293 omap_obj->pages = pages;
294
295 return 0;
296
297free_addrs:
298 kfree(addrs);
299free_pages:
300 drm_gem_put_pages(obj, pages, true, false);
301
302 return ret;
303}
304
305/* acquire pages when needed (for example, for DMA where physically
306 * contiguous buffer is not required
307 */
308static int get_pages(struct drm_gem_object *obj, struct page ***pages)
309{
310 struct omap_gem_object *omap_obj = to_omap_bo(obj);
311 int ret = 0;
312
313 if ((omap_obj->flags & OMAP_BO_MEM_SHMEM) && !omap_obj->pages) {
314 ret = omap_gem_attach_pages(obj);
315 if (ret) {
316 dev_err(obj->dev->dev, "could not attach pages\n");
317 return ret;
318 }
319 }
320
321 /* TODO: even phys-contig.. we should have a list of pages? */
322 *pages = omap_obj->pages;
323
324 return 0;
325}
326
327/** release backing pages */
328static void omap_gem_detach_pages(struct drm_gem_object *obj)
329{
330 struct omap_gem_object *omap_obj = to_omap_bo(obj);
331
332 /* for non-cached buffers, ensure the new pages are clean because
333 * DSS, GPU, etc. are not cache coherent:
334 */
335 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
336 int i, npages = obj->size >> PAGE_SHIFT;
337 for (i = 0; i < npages; i++) {
338 dma_unmap_page(obj->dev->dev, omap_obj->addrs[i],
339 PAGE_SIZE, DMA_BIDIRECTIONAL);
340 }
341 }
342
343 kfree(omap_obj->addrs);
344 omap_obj->addrs = NULL;
345
346 drm_gem_put_pages(obj, omap_obj->pages, true, false);
347 omap_obj->pages = NULL;
348}
349
350/* get buffer flags */
351uint32_t omap_gem_flags(struct drm_gem_object *obj)
352{
353 return to_omap_bo(obj)->flags;
354}
355
356uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj)
357{
358 uint64_t offset;
359 mutex_lock(&obj->dev->struct_mutex);
360 offset = mmap_offset(obj);
361 mutex_unlock(&obj->dev->struct_mutex);
362 return offset;
363}
364
365/** get mmap size */
366size_t omap_gem_mmap_size(struct drm_gem_object *obj)
367{
368 struct omap_gem_object *omap_obj = to_omap_bo(obj);
369 size_t size = obj->size;
370
371 if (omap_obj->flags & OMAP_BO_TILED) {
372 /* for tiled buffers, the virtual size has stride rounded up
373 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
374 * 32kb later!). But we don't back the entire buffer with
375 * pages, only the valid picture part.. so need to adjust for
376 * this in the size used to mmap and generate mmap offset
377 */
378 size = tiler_vsize(gem2fmt(omap_obj->flags),
379 omap_obj->width, omap_obj->height);
380 }
381
382 return size;
383}
384
385/* get tiled size, returns -EINVAL if not tiled buffer */
386int omap_gem_tiled_size(struct drm_gem_object *obj, uint16_t *w, uint16_t *h)
387{
388 struct omap_gem_object *omap_obj = to_omap_bo(obj);
389 if (omap_obj->flags & OMAP_BO_TILED) {
390 *w = omap_obj->width;
391 *h = omap_obj->height;
392 return 0;
393 }
394 return -EINVAL;
395}
396
397/* -----------------------------------------------------------------------------
398 * Fault Handling
399 */
400
401/* Normal handling for the case of faulting in non-tiled buffers */
402static int fault_1d(struct drm_gem_object *obj,
403 struct vm_area_struct *vma, struct vm_fault *vmf)
404{
405 struct omap_gem_object *omap_obj = to_omap_bo(obj);
406 unsigned long pfn;
407 pgoff_t pgoff;
408
409 /* We don't use vmf->pgoff since that has the fake offset: */
410 pgoff = ((unsigned long)vmf->virtual_address -
411 vma->vm_start) >> PAGE_SHIFT;
412
413 if (omap_obj->pages) {
414 omap_gem_cpu_sync(obj, pgoff);
415 pfn = page_to_pfn(omap_obj->pages[pgoff]);
416 } else {
417 BUG_ON(!is_contiguous(omap_obj));
418 pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff;
419 }
420
421 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
422 pfn, pfn << PAGE_SHIFT);
423
424 return vm_insert_mixed(vma, (unsigned long)vmf->virtual_address,
425 __pfn_to_pfn_t(pfn, PFN_DEV));
426}
427
428/* Special handling for the case of faulting in 2d tiled buffers */
429static int fault_2d(struct drm_gem_object *obj,
430 struct vm_area_struct *vma, struct vm_fault *vmf)
431{
432 struct omap_gem_object *omap_obj = to_omap_bo(obj);
433 struct omap_drm_private *priv = obj->dev->dev_private;
434 struct omap_drm_usergart_entry *entry;
435 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
436 struct page *pages[64]; /* XXX is this too much to have on stack? */
437 unsigned long pfn;
438 pgoff_t pgoff, base_pgoff;
439 void __user *vaddr;
440 int i, ret, slots;
441
442 /*
443 * Note the height of the slot is also equal to the number of pages
444 * that need to be mapped in to fill 4kb wide CPU page. If the slot
445 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
446 */
447 const int n = priv->usergart[fmt].height;
448 const int n_shift = priv->usergart[fmt].height_shift;
449
450 /*
451 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
452 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
453 * into account in some of the math, so figure out virtual stride
454 * in pages
455 */
456 const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
457
458 /* We don't use vmf->pgoff since that has the fake offset: */
459 pgoff = ((unsigned long)vmf->virtual_address -
460 vma->vm_start) >> PAGE_SHIFT;
461
462 /*
463 * Actual address we start mapping at is rounded down to previous slot
464 * boundary in the y direction:
465 */
466 base_pgoff = round_down(pgoff, m << n_shift);
467
468 /* figure out buffer width in slots */
469 slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
470
471 vaddr = vmf->virtual_address - ((pgoff - base_pgoff) << PAGE_SHIFT);
472
473 entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
474
475 /* evict previous buffer using this usergart entry, if any: */
476 if (entry->obj)
477 evict_entry(entry->obj, fmt, entry);
478
479 entry->obj = obj;
480 entry->obj_pgoff = base_pgoff;
481
482 /* now convert base_pgoff to phys offset from virt offset: */
483 base_pgoff = (base_pgoff >> n_shift) * slots;
484
485 /* for wider-than 4k.. figure out which part of the slot-row we want: */
486 if (m > 1) {
487 int off = pgoff % m;
488 entry->obj_pgoff += off;
489 base_pgoff /= m;
490 slots = min(slots - (off << n_shift), n);
491 base_pgoff += off << n_shift;
492 vaddr += off << PAGE_SHIFT;
493 }
494
495 /*
496 * Map in pages. Beyond the valid pixel part of the buffer, we set
497 * pages[i] to NULL to get a dummy page mapped in.. if someone
498 * reads/writes it they will get random/undefined content, but at
499 * least it won't be corrupting whatever other random page used to
500 * be mapped in, or other undefined behavior.
501 */
502 memcpy(pages, &omap_obj->pages[base_pgoff],
503 sizeof(struct page *) * slots);
504 memset(pages + slots, 0,
505 sizeof(struct page *) * (n - slots));
506
507 ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
508 if (ret) {
509 dev_err(obj->dev->dev, "failed to pin: %d\n", ret);
510 return ret;
511 }
512
513 pfn = entry->paddr >> PAGE_SHIFT;
514
515 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
516 pfn, pfn << PAGE_SHIFT);
517
518 for (i = n; i > 0; i--) {
519 vm_insert_mixed(vma, (unsigned long)vaddr,
520 __pfn_to_pfn_t(pfn, PFN_DEV));
521 pfn += priv->usergart[fmt].stride_pfn;
522 vaddr += PAGE_SIZE * m;
523 }
524
525 /* simple round-robin: */
526 priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
527 % NUM_USERGART_ENTRIES;
528
529 return 0;
530}
531
532/**
533 * omap_gem_fault - pagefault handler for GEM objects
534 * @vma: the VMA of the GEM object
535 * @vmf: fault detail
536 *
537 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
538 * does most of the work for us including the actual map/unmap calls
539 * but we need to do the actual page work.
540 *
541 * The VMA was set up by GEM. In doing so it also ensured that the
542 * vma->vm_private_data points to the GEM object that is backing this
543 * mapping.
544 */
545int omap_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
546{
547 struct drm_gem_object *obj = vma->vm_private_data;
548 struct omap_gem_object *omap_obj = to_omap_bo(obj);
549 struct drm_device *dev = obj->dev;
550 struct page **pages;
551 int ret;
552
553 /* Make sure we don't parallel update on a fault, nor move or remove
554 * something from beneath our feet
555 */
556 mutex_lock(&dev->struct_mutex);
557
558 /* if a shmem backed object, make sure we have pages attached now */
559 ret = get_pages(obj, &pages);
560 if (ret)
561 goto fail;
562
563 /* where should we do corresponding put_pages().. we are mapping
564 * the original page, rather than thru a GART, so we can't rely
565 * on eviction to trigger this. But munmap() or all mappings should
566 * probably trigger put_pages()?
567 */
568
569 if (omap_obj->flags & OMAP_BO_TILED)
570 ret = fault_2d(obj, vma, vmf);
571 else
572 ret = fault_1d(obj, vma, vmf);
573
574
575fail:
576 mutex_unlock(&dev->struct_mutex);
577 switch (ret) {
578 case 0:
579 case -ERESTARTSYS:
580 case -EINTR:
581 case -EBUSY:
582 /*
583 * EBUSY is ok: this just means that another thread
584 * already did the job.
585 */
586 return VM_FAULT_NOPAGE;
587 case -ENOMEM:
588 return VM_FAULT_OOM;
589 default:
590 return VM_FAULT_SIGBUS;
591 }
592}
593
594/** We override mainly to fix up some of the vm mapping flags.. */
595int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
596{
597 int ret;
598
599 ret = drm_gem_mmap(filp, vma);
600 if (ret) {
601 DBG("mmap failed: %d", ret);
602 return ret;
603 }
604
605 return omap_gem_mmap_obj(vma->vm_private_data, vma);
606}
607
608int omap_gem_mmap_obj(struct drm_gem_object *obj,
609 struct vm_area_struct *vma)
610{
611 struct omap_gem_object *omap_obj = to_omap_bo(obj);
612
613 vma->vm_flags &= ~VM_PFNMAP;
614 vma->vm_flags |= VM_MIXEDMAP;
615
616 if (omap_obj->flags & OMAP_BO_WC) {
617 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
618 } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
619 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
620 } else {
621 /*
622 * We do have some private objects, at least for scanout buffers
623 * on hardware without DMM/TILER. But these are allocated write-
624 * combine
625 */
626 if (WARN_ON(!obj->filp))
627 return -EINVAL;
628
629 /*
630 * Shunt off cached objs to shmem file so they have their own
631 * address_space (so unmap_mapping_range does what we want,
632 * in particular in the case of mmap'd dmabufs)
633 */
634 fput(vma->vm_file);
635 vma->vm_pgoff = 0;
636 vma->vm_file = get_file(obj->filp);
637
638 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
639 }
640
641 return 0;
642}
643
644/* -----------------------------------------------------------------------------
645 * Dumb Buffers
646 */
647
648/**
649 * omap_gem_dumb_create - create a dumb buffer
650 * @drm_file: our client file
651 * @dev: our device
652 * @args: the requested arguments copied from userspace
653 *
654 * Allocate a buffer suitable for use for a frame buffer of the
655 * form described by user space. Give userspace a handle by which
656 * to reference it.
657 */
658int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
659 struct drm_mode_create_dumb *args)
660{
661 union omap_gem_size gsize;
662
663 args->pitch = align_pitch(0, args->width, args->bpp);
664 args->size = PAGE_ALIGN(args->pitch * args->height);
665
666 gsize = (union omap_gem_size){
667 .bytes = args->size,
668 };
669
670 return omap_gem_new_handle(dev, file, gsize,
671 OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
672}
673
674/**
675 * omap_gem_dumb_map - buffer mapping for dumb interface
676 * @file: our drm client file
677 * @dev: drm device
678 * @handle: GEM handle to the object (from dumb_create)
679 *
680 * Do the necessary setup to allow the mapping of the frame buffer
681 * into user memory. We don't have to do much here at the moment.
682 */
683int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
684 uint32_t handle, uint64_t *offset)
685{
686 struct drm_gem_object *obj;
687 int ret = 0;
688
689 /* GEM does all our handle to object mapping */
690 obj = drm_gem_object_lookup(dev, file, handle);
691 if (obj == NULL) {
692 ret = -ENOENT;
693 goto fail;
694 }
695
696 *offset = omap_gem_mmap_offset(obj);
697
698 drm_gem_object_unreference_unlocked(obj);
699
700fail:
701 return ret;
702}
703
704#ifdef CONFIG_DRM_FBDEV_EMULATION
705/* Set scrolling position. This allows us to implement fast scrolling
706 * for console.
707 *
708 * Call only from non-atomic contexts.
709 */
710int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll)
711{
712 struct omap_gem_object *omap_obj = to_omap_bo(obj);
713 uint32_t npages = obj->size >> PAGE_SHIFT;
714 int ret = 0;
715
716 if (roll > npages) {
717 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
718 return -EINVAL;
719 }
720
721 omap_obj->roll = roll;
722
723 mutex_lock(&obj->dev->struct_mutex);
724
725 /* if we aren't mapped yet, we don't need to do anything */
726 if (omap_obj->block) {
727 struct page **pages;
728 ret = get_pages(obj, &pages);
729 if (ret)
730 goto fail;
731 ret = tiler_pin(omap_obj->block, pages, npages, roll, true);
732 if (ret)
733 dev_err(obj->dev->dev, "could not repin: %d\n", ret);
734 }
735
736fail:
737 mutex_unlock(&obj->dev->struct_mutex);
738
739 return ret;
740}
741#endif
742
743/* -----------------------------------------------------------------------------
744 * Memory Management & DMA Sync
745 */
746
747/**
748 * shmem buffers that are mapped cached can simulate coherency via using
749 * page faulting to keep track of dirty pages
750 */
751static inline bool is_cached_coherent(struct drm_gem_object *obj)
752{
753 struct omap_gem_object *omap_obj = to_omap_bo(obj);
754
755 return (omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
756 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED);
757}
758
759/* Sync the buffer for CPU access.. note pages should already be
760 * attached, ie. omap_gem_get_pages()
761 */
762void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff)
763{
764 struct drm_device *dev = obj->dev;
765 struct omap_gem_object *omap_obj = to_omap_bo(obj);
766
767 if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) {
768 dma_unmap_page(dev->dev, omap_obj->addrs[pgoff],
769 PAGE_SIZE, DMA_BIDIRECTIONAL);
770 omap_obj->addrs[pgoff] = 0;
771 }
772}
773
774/* sync the buffer for DMA access */
775void omap_gem_dma_sync(struct drm_gem_object *obj,
776 enum dma_data_direction dir)
777{
778 struct drm_device *dev = obj->dev;
779 struct omap_gem_object *omap_obj = to_omap_bo(obj);
780
781 if (is_cached_coherent(obj)) {
782 int i, npages = obj->size >> PAGE_SHIFT;
783 struct page **pages = omap_obj->pages;
784 bool dirty = false;
785
786 for (i = 0; i < npages; i++) {
787 if (!omap_obj->addrs[i]) {
788 dma_addr_t addr;
789
790 addr = dma_map_page(dev->dev, pages[i], 0,
791 PAGE_SIZE, DMA_BIDIRECTIONAL);
792
793 if (dma_mapping_error(dev->dev, addr)) {
794 dev_warn(dev->dev,
795 "%s: failed to map page\n",
796 __func__);
797 break;
798 }
799
800 dirty = true;
801 omap_obj->addrs[i] = addr;
802 }
803 }
804
805 if (dirty) {
806 unmap_mapping_range(obj->filp->f_mapping, 0,
807 omap_gem_mmap_size(obj), 1);
808 }
809 }
810}
811
812/* Get physical address for DMA.. if 'remap' is true, and the buffer is not
813 * already contiguous, remap it to pin in physically contiguous memory.. (ie.
814 * map in TILER)
815 */
816int omap_gem_get_paddr(struct drm_gem_object *obj,
817 dma_addr_t *paddr, bool remap)
818{
819 struct omap_drm_private *priv = obj->dev->dev_private;
820 struct omap_gem_object *omap_obj = to_omap_bo(obj);
821 int ret = 0;
822
823 mutex_lock(&obj->dev->struct_mutex);
824
825 if (!is_contiguous(omap_obj) && remap && priv->has_dmm) {
826 if (omap_obj->paddr_cnt == 0) {
827 struct page **pages;
828 uint32_t npages = obj->size >> PAGE_SHIFT;
829 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
830 struct tiler_block *block;
831
832 BUG_ON(omap_obj->block);
833
834 ret = get_pages(obj, &pages);
835 if (ret)
836 goto fail;
837
838 if (omap_obj->flags & OMAP_BO_TILED) {
839 block = tiler_reserve_2d(fmt,
840 omap_obj->width,
841 omap_obj->height, 0);
842 } else {
843 block = tiler_reserve_1d(obj->size);
844 }
845
846 if (IS_ERR(block)) {
847 ret = PTR_ERR(block);
848 dev_err(obj->dev->dev,
849 "could not remap: %d (%d)\n", ret, fmt);
850 goto fail;
851 }
852
853 /* TODO: enable async refill.. */
854 ret = tiler_pin(block, pages, npages,
855 omap_obj->roll, true);
856 if (ret) {
857 tiler_release(block);
858 dev_err(obj->dev->dev,
859 "could not pin: %d\n", ret);
860 goto fail;
861 }
862
863 omap_obj->paddr = tiler_ssptr(block);
864 omap_obj->block = block;
865
866 DBG("got paddr: %pad", &omap_obj->paddr);
867 }
868
869 omap_obj->paddr_cnt++;
870
871 *paddr = omap_obj->paddr;
872 } else if (is_contiguous(omap_obj)) {
873 *paddr = omap_obj->paddr;
874 } else {
875 ret = -EINVAL;
876 goto fail;
877 }
878
879fail:
880 mutex_unlock(&obj->dev->struct_mutex);
881
882 return ret;
883}
884
885/* Release physical address, when DMA is no longer being performed.. this
886 * could potentially unpin and unmap buffers from TILER
887 */
888void omap_gem_put_paddr(struct drm_gem_object *obj)
889{
890 struct omap_gem_object *omap_obj = to_omap_bo(obj);
891 int ret;
892
893 mutex_lock(&obj->dev->struct_mutex);
894 if (omap_obj->paddr_cnt > 0) {
895 omap_obj->paddr_cnt--;
896 if (omap_obj->paddr_cnt == 0) {
897 ret = tiler_unpin(omap_obj->block);
898 if (ret) {
899 dev_err(obj->dev->dev,
900 "could not unpin pages: %d\n", ret);
901 }
902 ret = tiler_release(omap_obj->block);
903 if (ret) {
904 dev_err(obj->dev->dev,
905 "could not release unmap: %d\n", ret);
906 }
907 omap_obj->paddr = 0;
908 omap_obj->block = NULL;
909 }
910 }
911
912 mutex_unlock(&obj->dev->struct_mutex);
913}
914
915/* Get rotated scanout address (only valid if already pinned), at the
916 * specified orientation and x,y offset from top-left corner of buffer
917 * (only valid for tiled 2d buffers)
918 */
919int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient,
920 int x, int y, dma_addr_t *paddr)
921{
922 struct omap_gem_object *omap_obj = to_omap_bo(obj);
923 int ret = -EINVAL;
924
925 mutex_lock(&obj->dev->struct_mutex);
926 if ((omap_obj->paddr_cnt > 0) && omap_obj->block &&
927 (omap_obj->flags & OMAP_BO_TILED)) {
928 *paddr = tiler_tsptr(omap_obj->block, orient, x, y);
929 ret = 0;
930 }
931 mutex_unlock(&obj->dev->struct_mutex);
932 return ret;
933}
934
935/* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
936int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient)
937{
938 struct omap_gem_object *omap_obj = to_omap_bo(obj);
939 int ret = -EINVAL;
940 if (omap_obj->flags & OMAP_BO_TILED)
941 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
942 return ret;
943}
944
945/* if !remap, and we don't have pages backing, then fail, rather than
946 * increasing the pin count (which we don't really do yet anyways,
947 * because we don't support swapping pages back out). And 'remap'
948 * might not be quite the right name, but I wanted to keep it working
949 * similarly to omap_gem_get_paddr(). Note though that mutex is not
950 * aquired if !remap (because this can be called in atomic ctxt),
951 * but probably omap_gem_get_paddr() should be changed to work in the
952 * same way. If !remap, a matching omap_gem_put_pages() call is not
953 * required (and should not be made).
954 */
955int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
956 bool remap)
957{
958 int ret;
959 if (!remap) {
960 struct omap_gem_object *omap_obj = to_omap_bo(obj);
961 if (!omap_obj->pages)
962 return -ENOMEM;
963 *pages = omap_obj->pages;
964 return 0;
965 }
966 mutex_lock(&obj->dev->struct_mutex);
967 ret = get_pages(obj, pages);
968 mutex_unlock(&obj->dev->struct_mutex);
969 return ret;
970}
971
972/* release pages when DMA no longer being performed */
973int omap_gem_put_pages(struct drm_gem_object *obj)
974{
975 /* do something here if we dynamically attach/detach pages.. at
976 * least they would no longer need to be pinned if everyone has
977 * released the pages..
978 */
979 return 0;
980}
981
982#ifdef CONFIG_DRM_FBDEV_EMULATION
983/* Get kernel virtual address for CPU access.. this more or less only
984 * exists for omap_fbdev. This should be called with struct_mutex
985 * held.
986 */
987void *omap_gem_vaddr(struct drm_gem_object *obj)
988{
989 struct omap_gem_object *omap_obj = to_omap_bo(obj);
990 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
991 if (!omap_obj->vaddr) {
992 struct page **pages;
993 int ret = get_pages(obj, &pages);
994 if (ret)
995 return ERR_PTR(ret);
996 omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
997 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
998 }
999 return omap_obj->vaddr;
1000}
1001#endif
1002
1003/* -----------------------------------------------------------------------------
1004 * Power Management
1005 */
1006
1007#ifdef CONFIG_PM
1008/* re-pin objects in DMM in resume path: */
1009int omap_gem_resume(struct device *dev)
1010{
1011 struct drm_device *drm_dev = dev_get_drvdata(dev);
1012 struct omap_drm_private *priv = drm_dev->dev_private;
1013 struct omap_gem_object *omap_obj;
1014 int ret = 0;
1015
1016 list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
1017 if (omap_obj->block) {
1018 struct drm_gem_object *obj = &omap_obj->base;
1019 uint32_t npages = obj->size >> PAGE_SHIFT;
1020 WARN_ON(!omap_obj->pages); /* this can't happen */
1021 ret = tiler_pin(omap_obj->block,
1022 omap_obj->pages, npages,
1023 omap_obj->roll, true);
1024 if (ret) {
1025 dev_err(dev, "could not repin: %d\n", ret);
1026 return ret;
1027 }
1028 }
1029 }
1030
1031 return 0;
1032}
1033#endif
1034
1035/* -----------------------------------------------------------------------------
1036 * DebugFS
1037 */
1038
1039#ifdef CONFIG_DEBUG_FS
1040void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1041{
1042 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1043 uint64_t off;
1044
1045 off = drm_vma_node_start(&obj->vma_node);
1046
1047 seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1048 omap_obj->flags, obj->name, obj->refcount.refcount.counter,
1049 off, &omap_obj->paddr, omap_obj->paddr_cnt,
1050 omap_obj->vaddr, omap_obj->roll);
1051
1052 if (omap_obj->flags & OMAP_BO_TILED) {
1053 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1054 if (omap_obj->block) {
1055 struct tcm_area *area = &omap_obj->block->area;
1056 seq_printf(m, " (%dx%d, %dx%d)",
1057 area->p0.x, area->p0.y,
1058 area->p1.x, area->p1.y);
1059 }
1060 } else {
1061 seq_printf(m, " %d", obj->size);
1062 }
1063
1064 seq_printf(m, "\n");
1065}
1066
1067void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1068{
1069 struct omap_gem_object *omap_obj;
1070 int count = 0;
1071 size_t size = 0;
1072
1073 list_for_each_entry(omap_obj, list, mm_list) {
1074 struct drm_gem_object *obj = &omap_obj->base;
1075 seq_printf(m, " ");
1076 omap_gem_describe(obj, m);
1077 count++;
1078 size += obj->size;
1079 }
1080
1081 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1082}
1083#endif
1084
1085/* -----------------------------------------------------------------------------
1086 * Buffer Synchronization
1087 */
1088
1089static DEFINE_SPINLOCK(sync_lock);
1090
1091struct omap_gem_sync_waiter {
1092 struct list_head list;
1093 struct omap_gem_object *omap_obj;
1094 enum omap_gem_op op;
1095 uint32_t read_target, write_target;
1096 /* notify called w/ sync_lock held */
1097 void (*notify)(void *arg);
1098 void *arg;
1099};
1100
1101/* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1102 * the read and/or write target count is achieved which can call a user
1103 * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1104 * cpu access), etc.
1105 */
1106static LIST_HEAD(waiters);
1107
1108static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1109{
1110 struct omap_gem_object *omap_obj = waiter->omap_obj;
1111 if ((waiter->op & OMAP_GEM_READ) &&
1112 (omap_obj->sync->write_complete < waiter->write_target))
1113 return true;
1114 if ((waiter->op & OMAP_GEM_WRITE) &&
1115 (omap_obj->sync->read_complete < waiter->read_target))
1116 return true;
1117 return false;
1118}
1119
1120/* macro for sync debug.. */
1121#define SYNCDBG 0
1122#define SYNC(fmt, ...) do { if (SYNCDBG) \
1123 printk(KERN_ERR "%s:%d: "fmt"\n", \
1124 __func__, __LINE__, ##__VA_ARGS__); \
1125 } while (0)
1126
1127
1128static void sync_op_update(void)
1129{
1130 struct omap_gem_sync_waiter *waiter, *n;
1131 list_for_each_entry_safe(waiter, n, &waiters, list) {
1132 if (!is_waiting(waiter)) {
1133 list_del(&waiter->list);
1134 SYNC("notify: %p", waiter);
1135 waiter->notify(waiter->arg);
1136 kfree(waiter);
1137 }
1138 }
1139}
1140
1141static inline int sync_op(struct drm_gem_object *obj,
1142 enum omap_gem_op op, bool start)
1143{
1144 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1145 int ret = 0;
1146
1147 spin_lock(&sync_lock);
1148
1149 if (!omap_obj->sync) {
1150 omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1151 if (!omap_obj->sync) {
1152 ret = -ENOMEM;
1153 goto unlock;
1154 }
1155 }
1156
1157 if (start) {
1158 if (op & OMAP_GEM_READ)
1159 omap_obj->sync->read_pending++;
1160 if (op & OMAP_GEM_WRITE)
1161 omap_obj->sync->write_pending++;
1162 } else {
1163 if (op & OMAP_GEM_READ)
1164 omap_obj->sync->read_complete++;
1165 if (op & OMAP_GEM_WRITE)
1166 omap_obj->sync->write_complete++;
1167 sync_op_update();
1168 }
1169
1170unlock:
1171 spin_unlock(&sync_lock);
1172
1173 return ret;
1174}
1175
1176/* mark the start of read and/or write operation */
1177int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1178{
1179 return sync_op(obj, op, true);
1180}
1181
1182int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1183{
1184 return sync_op(obj, op, false);
1185}
1186
1187static DECLARE_WAIT_QUEUE_HEAD(sync_event);
1188
1189static void sync_notify(void *arg)
1190{
1191 struct task_struct **waiter_task = arg;
1192 *waiter_task = NULL;
1193 wake_up_all(&sync_event);
1194}
1195
1196int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1197{
1198 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1199 int ret = 0;
1200 if (omap_obj->sync) {
1201 struct task_struct *waiter_task = current;
1202 struct omap_gem_sync_waiter *waiter =
1203 kzalloc(sizeof(*waiter), GFP_KERNEL);
1204
1205 if (!waiter)
1206 return -ENOMEM;
1207
1208 waiter->omap_obj = omap_obj;
1209 waiter->op = op;
1210 waiter->read_target = omap_obj->sync->read_pending;
1211 waiter->write_target = omap_obj->sync->write_pending;
1212 waiter->notify = sync_notify;
1213 waiter->arg = &waiter_task;
1214
1215 spin_lock(&sync_lock);
1216 if (is_waiting(waiter)) {
1217 SYNC("waited: %p", waiter);
1218 list_add_tail(&waiter->list, &waiters);
1219 spin_unlock(&sync_lock);
1220 ret = wait_event_interruptible(sync_event,
1221 (waiter_task == NULL));
1222 spin_lock(&sync_lock);
1223 if (waiter_task) {
1224 SYNC("interrupted: %p", waiter);
1225 /* we were interrupted */
1226 list_del(&waiter->list);
1227 waiter_task = NULL;
1228 } else {
1229 /* freed in sync_op_update() */
1230 waiter = NULL;
1231 }
1232 }
1233 spin_unlock(&sync_lock);
1234 kfree(waiter);
1235 }
1236 return ret;
1237}
1238
1239/* call fxn(arg), either synchronously or asynchronously if the op
1240 * is currently blocked.. fxn() can be called from any context
1241 *
1242 * (TODO for now fxn is called back from whichever context calls
1243 * omap_gem_op_finish().. but this could be better defined later
1244 * if needed)
1245 *
1246 * TODO more code in common w/ _sync()..
1247 */
1248int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1249 void (*fxn)(void *arg), void *arg)
1250{
1251 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1252 if (omap_obj->sync) {
1253 struct omap_gem_sync_waiter *waiter =
1254 kzalloc(sizeof(*waiter), GFP_ATOMIC);
1255
1256 if (!waiter)
1257 return -ENOMEM;
1258
1259 waiter->omap_obj = omap_obj;
1260 waiter->op = op;
1261 waiter->read_target = omap_obj->sync->read_pending;
1262 waiter->write_target = omap_obj->sync->write_pending;
1263 waiter->notify = fxn;
1264 waiter->arg = arg;
1265
1266 spin_lock(&sync_lock);
1267 if (is_waiting(waiter)) {
1268 SYNC("waited: %p", waiter);
1269 list_add_tail(&waiter->list, &waiters);
1270 spin_unlock(&sync_lock);
1271 return 0;
1272 }
1273
1274 spin_unlock(&sync_lock);
1275
1276 kfree(waiter);
1277 }
1278
1279 /* no waiting.. */
1280 fxn(arg);
1281
1282 return 0;
1283}
1284
1285/* -----------------------------------------------------------------------------
1286 * Constructor & Destructor
1287 */
1288
1289void omap_gem_free_object(struct drm_gem_object *obj)
1290{
1291 struct drm_device *dev = obj->dev;
1292 struct omap_drm_private *priv = dev->dev_private;
1293 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1294
1295 evict(obj);
1296
1297 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1298
1299 spin_lock(&priv->list_lock);
1300 list_del(&omap_obj->mm_list);
1301 spin_unlock(&priv->list_lock);
1302
1303 /* this means the object is still pinned.. which really should
1304 * not happen. I think..
1305 */
1306 WARN_ON(omap_obj->paddr_cnt > 0);
1307
1308 if (omap_obj->pages) {
1309 if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1310 kfree(omap_obj->pages);
1311 else
1312 omap_gem_detach_pages(obj);
1313 }
1314
1315 if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1316 dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1317 omap_obj->paddr);
1318 } else if (omap_obj->vaddr) {
1319 vunmap(omap_obj->vaddr);
1320 } else if (obj->import_attach) {
1321 drm_prime_gem_destroy(obj, omap_obj->sgt);
1322 }
1323
1324 kfree(omap_obj->sync);
1325
1326 drm_gem_object_release(obj);
1327
1328 kfree(omap_obj);
1329}
1330
1331/* GEM buffer object constructor */
1332struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1333 union omap_gem_size gsize, uint32_t flags)
1334{
1335 struct omap_drm_private *priv = dev->dev_private;
1336 struct omap_gem_object *omap_obj;
1337 struct drm_gem_object *obj;
1338 struct address_space *mapping;
1339 size_t size;
1340 int ret;
1341
1342 /* Validate the flags and compute the memory and cache flags. */
1343 if (flags & OMAP_BO_TILED) {
1344 if (!priv->usergart) {
1345 dev_err(dev->dev, "Tiled buffers require DMM\n");
1346 return NULL;
1347 }
1348
1349 /*
1350 * Tiled buffers are always shmem paged backed. When they are
1351 * scanned out, they are remapped into DMM/TILER.
1352 */
1353 flags &= ~OMAP_BO_SCANOUT;
1354 flags |= OMAP_BO_MEM_SHMEM;
1355
1356 /*
1357 * Currently don't allow cached buffers. There is some caching
1358 * stuff that needs to be handled better.
1359 */
1360 flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1361 flags |= tiler_get_cpu_cache_flags();
1362 } else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1363 /*
1364 * OMAP_BO_SCANOUT hints that the buffer doesn't need to be
1365 * tiled. However, to lower the pressure on memory allocation,
1366 * use contiguous memory only if no TILER is available.
1367 */
1368 flags |= OMAP_BO_MEM_DMA_API;
1369 } else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1370 /*
1371 * All other buffers not backed by dma_buf are shmem-backed.
1372 */
1373 flags |= OMAP_BO_MEM_SHMEM;
1374 }
1375
1376 /* Allocate the initialize the OMAP GEM object. */
1377 omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1378 if (!omap_obj)
1379 return NULL;
1380
1381 obj = &omap_obj->base;
1382 omap_obj->flags = flags;
1383
1384 if (flags & OMAP_BO_TILED) {
1385 /*
1386 * For tiled buffers align dimensions to slot boundaries and
1387 * calculate size based on aligned dimensions.
1388 */
1389 tiler_align(gem2fmt(flags), &gsize.tiled.width,
1390 &gsize.tiled.height);
1391
1392 size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1393 gsize.tiled.height);
1394
1395 omap_obj->width = gsize.tiled.width;
1396 omap_obj->height = gsize.tiled.height;
1397 } else {
1398 size = PAGE_ALIGN(gsize.bytes);
1399 }
1400
1401 /* Initialize the GEM object. */
1402 if (!(flags & OMAP_BO_MEM_SHMEM)) {
1403 drm_gem_private_object_init(dev, obj, size);
1404 } else {
1405 ret = drm_gem_object_init(dev, obj, size);
1406 if (ret)
1407 goto err_free;
1408
1409 mapping = file_inode(obj->filp)->i_mapping;
1410 mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1411 }
1412
1413 /* Allocate memory if needed. */
1414 if (flags & OMAP_BO_MEM_DMA_API) {
1415 omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1416 &omap_obj->paddr,
1417 GFP_KERNEL);
1418 if (!omap_obj->vaddr)
1419 goto err_release;
1420 }
1421
1422 spin_lock(&priv->list_lock);
1423 list_add(&omap_obj->mm_list, &priv->obj_list);
1424 spin_unlock(&priv->list_lock);
1425
1426 return obj;
1427
1428err_release:
1429 drm_gem_object_release(obj);
1430err_free:
1431 kfree(omap_obj);
1432 return NULL;
1433}
1434
1435struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1436 struct sg_table *sgt)
1437{
1438 struct omap_drm_private *priv = dev->dev_private;
1439 struct omap_gem_object *omap_obj;
1440 struct drm_gem_object *obj;
1441 union omap_gem_size gsize;
1442
1443 /* Without a DMM only physically contiguous buffers can be supported. */
1444 if (sgt->orig_nents != 1 && !priv->has_dmm)
1445 return ERR_PTR(-EINVAL);
1446
1447 mutex_lock(&dev->struct_mutex);
1448
1449 gsize.bytes = PAGE_ALIGN(size);
1450 obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1451 if (!obj) {
1452 obj = ERR_PTR(-ENOMEM);
1453 goto done;
1454 }
1455
1456 omap_obj = to_omap_bo(obj);
1457 omap_obj->sgt = sgt;
1458
1459 if (sgt->orig_nents == 1) {
1460 omap_obj->paddr = sg_dma_address(sgt->sgl);
1461 } else {
1462 /* Create pages list from sgt */
1463 struct sg_page_iter iter;
1464 struct page **pages;
1465 unsigned int npages;
1466 unsigned int i = 0;
1467
1468 npages = DIV_ROUND_UP(size, PAGE_SIZE);
1469 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1470 if (!pages) {
1471 omap_gem_free_object(obj);
1472 obj = ERR_PTR(-ENOMEM);
1473 goto done;
1474 }
1475
1476 omap_obj->pages = pages;
1477
1478 for_each_sg_page(sgt->sgl, &iter, sgt->orig_nents, 0) {
1479 pages[i++] = sg_page_iter_page(&iter);
1480 if (i > npages)
1481 break;
1482 }
1483
1484 if (WARN_ON(i != npages)) {
1485 omap_gem_free_object(obj);
1486 obj = ERR_PTR(-ENOMEM);
1487 goto done;
1488 }
1489 }
1490
1491done:
1492 mutex_unlock(&dev->struct_mutex);
1493 return obj;
1494}
1495
1496/* convenience method to construct a GEM buffer object, and userspace handle */
1497int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1498 union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1499{
1500 struct drm_gem_object *obj;
1501 int ret;
1502
1503 obj = omap_gem_new(dev, gsize, flags);
1504 if (!obj)
1505 return -ENOMEM;
1506
1507 ret = drm_gem_handle_create(file, obj, handle);
1508 if (ret) {
1509 omap_gem_free_object(obj);
1510 return ret;
1511 }
1512
1513 /* drop reference from allocate - handle holds it now */
1514 drm_gem_object_unreference_unlocked(obj);
1515
1516 return 0;
1517}
1518
1519/* -----------------------------------------------------------------------------
1520 * Init & Cleanup
1521 */
1522
1523/* If DMM is used, we need to set some stuff up.. */
1524void omap_gem_init(struct drm_device *dev)
1525{
1526 struct omap_drm_private *priv = dev->dev_private;
1527 struct omap_drm_usergart *usergart;
1528 const enum tiler_fmt fmts[] = {
1529 TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1530 };
1531 int i, j;
1532
1533 if (!dmm_is_available()) {
1534 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1535 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1536 return;
1537 }
1538
1539 usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1540 if (!usergart)
1541 return;
1542
1543 /* reserve 4k aligned/wide regions for userspace mappings: */
1544 for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1545 uint16_t h = 1, w = PAGE_SIZE >> i;
1546 tiler_align(fmts[i], &w, &h);
1547 /* note: since each region is 1 4kb page wide, and minimum
1548 * number of rows, the height ends up being the same as the
1549 * # of pages in the region
1550 */
1551 usergart[i].height = h;
1552 usergart[i].height_shift = ilog2(h);
1553 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1554 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1555 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1556 struct omap_drm_usergart_entry *entry;
1557 struct tiler_block *block;
1558
1559 entry = &usergart[i].entry[j];
1560 block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1561 if (IS_ERR(block)) {
1562 dev_err(dev->dev,
1563 "reserve failed: %d, %d, %ld\n",
1564 i, j, PTR_ERR(block));
1565 return;
1566 }
1567 entry->paddr = tiler_ssptr(block);
1568 entry->block = block;
1569
1570 DBG("%d:%d: %dx%d: paddr=%pad stride=%d", i, j, w, h,
1571 &entry->paddr,
1572 usergart[i].stride_pfn << PAGE_SHIFT);
1573 }
1574 }
1575
1576 priv->usergart = usergart;
1577 priv->has_dmm = true;
1578}
1579
1580void omap_gem_deinit(struct drm_device *dev)
1581{
1582 struct omap_drm_private *priv = dev->dev_private;
1583
1584 /* I believe we can rely on there being no more outstanding GEM
1585 * objects which could depend on usergart/dmm at this point.
1586 */
1587 kfree(priv->usergart);
1588}
1/*
2 * drivers/gpu/drm/omapdrm/omap_gem.c
3 *
4 * Copyright (C) 2011 Texas Instruments
5 * Author: Rob Clark <rob.clark@linaro.org>
6 *
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 as published by
9 * the Free Software Foundation.
10 *
11 * This program is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * more details.
15 *
16 * You should have received a copy of the GNU General Public License along with
17 * this program. If not, see <http://www.gnu.org/licenses/>.
18 */
19
20
21#include <linux/spinlock.h>
22#include <linux/shmem_fs.h>
23#include <drm/drm_vma_manager.h>
24
25#include "omap_drv.h"
26#include "omap_dmm_tiler.h"
27
28/* remove these once drm core helpers are merged */
29struct page **_drm_gem_get_pages(struct drm_gem_object *obj, gfp_t gfpmask);
30void _drm_gem_put_pages(struct drm_gem_object *obj, struct page **pages,
31 bool dirty, bool accessed);
32int _drm_gem_create_mmap_offset_size(struct drm_gem_object *obj, size_t size);
33
34/*
35 * GEM buffer object implementation.
36 */
37
38#define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
39
40/* note: we use upper 8 bits of flags for driver-internal flags: */
41#define OMAP_BO_DMA 0x01000000 /* actually is physically contiguous */
42#define OMAP_BO_EXT_SYNC 0x02000000 /* externally allocated sync object */
43#define OMAP_BO_EXT_MEM 0x04000000 /* externally allocated memory */
44
45
46struct omap_gem_object {
47 struct drm_gem_object base;
48
49 struct list_head mm_list;
50
51 uint32_t flags;
52
53 /** width/height for tiled formats (rounded up to slot boundaries) */
54 uint16_t width, height;
55
56 /** roll applied when mapping to DMM */
57 uint32_t roll;
58
59 /**
60 * If buffer is allocated physically contiguous, the OMAP_BO_DMA flag
61 * is set and the paddr is valid. Also if the buffer is remapped in
62 * TILER and paddr_cnt > 0, then paddr is valid. But if you are using
63 * the physical address and OMAP_BO_DMA is not set, then you should
64 * be going thru omap_gem_{get,put}_paddr() to ensure the mapping is
65 * not removed from under your feet.
66 *
67 * Note that OMAP_BO_SCANOUT is a hint from userspace that DMA capable
68 * buffer is requested, but doesn't mean that it is. Use the
69 * OMAP_BO_DMA flag to determine if the buffer has a DMA capable
70 * physical address.
71 */
72 dma_addr_t paddr;
73
74 /**
75 * # of users of paddr
76 */
77 uint32_t paddr_cnt;
78
79 /**
80 * tiler block used when buffer is remapped in DMM/TILER.
81 */
82 struct tiler_block *block;
83
84 /**
85 * Array of backing pages, if allocated. Note that pages are never
86 * allocated for buffers originally allocated from contiguous memory
87 */
88 struct page **pages;
89
90 /** addresses corresponding to pages in above array */
91 dma_addr_t *addrs;
92
93 /**
94 * Virtual address, if mapped.
95 */
96 void *vaddr;
97
98 /**
99 * sync-object allocated on demand (if needed)
100 *
101 * Per-buffer sync-object for tracking pending and completed hw/dma
102 * read and write operations. The layout in memory is dictated by
103 * the SGX firmware, which uses this information to stall the command
104 * stream if a surface is not ready yet.
105 *
106 * Note that when buffer is used by SGX, the sync-object needs to be
107 * allocated from a special heap of sync-objects. This way many sync
108 * objects can be packed in a page, and not waste GPU virtual address
109 * space. Because of this we have to have a omap_gem_set_sync_object()
110 * API to allow replacement of the syncobj after it has (potentially)
111 * already been allocated. A bit ugly but I haven't thought of a
112 * better alternative.
113 */
114 struct {
115 uint32_t write_pending;
116 uint32_t write_complete;
117 uint32_t read_pending;
118 uint32_t read_complete;
119 } *sync;
120};
121
122static int get_pages(struct drm_gem_object *obj, struct page ***pages);
123static uint64_t mmap_offset(struct drm_gem_object *obj);
124
125/* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
126 * not necessarily pinned in TILER all the time, and (b) when they are
127 * they are not necessarily page aligned, we reserve one or more small
128 * regions in each of the 2d containers to use as a user-GART where we
129 * can create a second page-aligned mapping of parts of the buffer
130 * being accessed from userspace.
131 *
132 * Note that we could optimize slightly when we know that multiple
133 * tiler containers are backed by the same PAT.. but I'll leave that
134 * for later..
135 */
136#define NUM_USERGART_ENTRIES 2
137struct usergart_entry {
138 struct tiler_block *block; /* the reserved tiler block */
139 dma_addr_t paddr;
140 struct drm_gem_object *obj; /* the current pinned obj */
141 pgoff_t obj_pgoff; /* page offset of obj currently
142 mapped in */
143};
144static struct {
145 struct usergart_entry entry[NUM_USERGART_ENTRIES];
146 int height; /* height in rows */
147 int height_shift; /* ilog2(height in rows) */
148 int slot_shift; /* ilog2(width per slot) */
149 int stride_pfn; /* stride in pages */
150 int last; /* index of last used entry */
151} *usergart;
152
153static void evict_entry(struct drm_gem_object *obj,
154 enum tiler_fmt fmt, struct usergart_entry *entry)
155{
156 struct omap_gem_object *omap_obj = to_omap_bo(obj);
157 int n = usergart[fmt].height;
158 size_t size = PAGE_SIZE * n;
159 loff_t off = mmap_offset(obj) +
160 (entry->obj_pgoff << PAGE_SHIFT);
161 const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
162
163 if (m > 1) {
164 int i;
165 /* if stride > than PAGE_SIZE then sparse mapping: */
166 for (i = n; i > 0; i--) {
167 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
168 off, PAGE_SIZE, 1);
169 off += PAGE_SIZE * m;
170 }
171 } else {
172 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
173 off, size, 1);
174 }
175
176 entry->obj = NULL;
177}
178
179/* Evict a buffer from usergart, if it is mapped there */
180static void evict(struct drm_gem_object *obj)
181{
182 struct omap_gem_object *omap_obj = to_omap_bo(obj);
183
184 if (omap_obj->flags & OMAP_BO_TILED) {
185 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
186 int i;
187
188 if (!usergart)
189 return;
190
191 for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
192 struct usergart_entry *entry = &usergart[fmt].entry[i];
193 if (entry->obj == obj)
194 evict_entry(obj, fmt, entry);
195 }
196 }
197}
198
199/* GEM objects can either be allocated from contiguous memory (in which
200 * case obj->filp==NULL), or w/ shmem backing (obj->filp!=NULL). But non
201 * contiguous buffers can be remapped in TILER/DMM if they need to be
202 * contiguous... but we don't do this all the time to reduce pressure
203 * on TILER/DMM space when we know at allocation time that the buffer
204 * will need to be scanned out.
205 */
206static inline bool is_shmem(struct drm_gem_object *obj)
207{
208 return obj->filp != NULL;
209}
210
211/**
212 * shmem buffers that are mapped cached can simulate coherency via using
213 * page faulting to keep track of dirty pages
214 */
215static inline bool is_cached_coherent(struct drm_gem_object *obj)
216{
217 struct omap_gem_object *omap_obj = to_omap_bo(obj);
218 return is_shmem(obj) &&
219 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED);
220}
221
222static DEFINE_SPINLOCK(sync_lock);
223
224/** ensure backing pages are allocated */
225static int omap_gem_attach_pages(struct drm_gem_object *obj)
226{
227 struct drm_device *dev = obj->dev;
228 struct omap_gem_object *omap_obj = to_omap_bo(obj);
229 struct page **pages;
230 int npages = obj->size >> PAGE_SHIFT;
231 int i, ret;
232 dma_addr_t *addrs;
233
234 WARN_ON(omap_obj->pages);
235
236 /* TODO: __GFP_DMA32 .. but somehow GFP_HIGHMEM is coming from the
237 * mapping_gfp_mask(mapping) which conflicts w/ GFP_DMA32.. probably
238 * we actually want CMA memory for it all anyways..
239 */
240 pages = drm_gem_get_pages(obj, GFP_KERNEL);
241 if (IS_ERR(pages)) {
242 dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
243 return PTR_ERR(pages);
244 }
245
246 /* for non-cached buffers, ensure the new pages are clean because
247 * DSS, GPU, etc. are not cache coherent:
248 */
249 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
250 addrs = kmalloc(npages * sizeof(*addrs), GFP_KERNEL);
251 if (!addrs) {
252 ret = -ENOMEM;
253 goto free_pages;
254 }
255
256 for (i = 0; i < npages; i++) {
257 addrs[i] = dma_map_page(dev->dev, pages[i],
258 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
259 }
260 } else {
261 addrs = kzalloc(npages * sizeof(*addrs), GFP_KERNEL);
262 if (!addrs) {
263 ret = -ENOMEM;
264 goto free_pages;
265 }
266 }
267
268 omap_obj->addrs = addrs;
269 omap_obj->pages = pages;
270
271 return 0;
272
273free_pages:
274 drm_gem_put_pages(obj, pages, true, false);
275
276 return ret;
277}
278
279/** release backing pages */
280static void omap_gem_detach_pages(struct drm_gem_object *obj)
281{
282 struct omap_gem_object *omap_obj = to_omap_bo(obj);
283
284 /* for non-cached buffers, ensure the new pages are clean because
285 * DSS, GPU, etc. are not cache coherent:
286 */
287 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
288 int i, npages = obj->size >> PAGE_SHIFT;
289 for (i = 0; i < npages; i++) {
290 dma_unmap_page(obj->dev->dev, omap_obj->addrs[i],
291 PAGE_SIZE, DMA_BIDIRECTIONAL);
292 }
293 }
294
295 kfree(omap_obj->addrs);
296 omap_obj->addrs = NULL;
297
298 drm_gem_put_pages(obj, omap_obj->pages, true, false);
299 omap_obj->pages = NULL;
300}
301
302/* get buffer flags */
303uint32_t omap_gem_flags(struct drm_gem_object *obj)
304{
305 return to_omap_bo(obj)->flags;
306}
307
308/** get mmap offset */
309static uint64_t mmap_offset(struct drm_gem_object *obj)
310{
311 struct drm_device *dev = obj->dev;
312 int ret;
313 size_t size;
314
315 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
316
317 /* Make it mmapable */
318 size = omap_gem_mmap_size(obj);
319 ret = drm_gem_create_mmap_offset_size(obj, size);
320 if (ret) {
321 dev_err(dev->dev, "could not allocate mmap offset\n");
322 return 0;
323 }
324
325 return drm_vma_node_offset_addr(&obj->vma_node);
326}
327
328uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj)
329{
330 uint64_t offset;
331 mutex_lock(&obj->dev->struct_mutex);
332 offset = mmap_offset(obj);
333 mutex_unlock(&obj->dev->struct_mutex);
334 return offset;
335}
336
337/** get mmap size */
338size_t omap_gem_mmap_size(struct drm_gem_object *obj)
339{
340 struct omap_gem_object *omap_obj = to_omap_bo(obj);
341 size_t size = obj->size;
342
343 if (omap_obj->flags & OMAP_BO_TILED) {
344 /* for tiled buffers, the virtual size has stride rounded up
345 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
346 * 32kb later!). But we don't back the entire buffer with
347 * pages, only the valid picture part.. so need to adjust for
348 * this in the size used to mmap and generate mmap offset
349 */
350 size = tiler_vsize(gem2fmt(omap_obj->flags),
351 omap_obj->width, omap_obj->height);
352 }
353
354 return size;
355}
356
357/* get tiled size, returns -EINVAL if not tiled buffer */
358int omap_gem_tiled_size(struct drm_gem_object *obj, uint16_t *w, uint16_t *h)
359{
360 struct omap_gem_object *omap_obj = to_omap_bo(obj);
361 if (omap_obj->flags & OMAP_BO_TILED) {
362 *w = omap_obj->width;
363 *h = omap_obj->height;
364 return 0;
365 }
366 return -EINVAL;
367}
368
369/* Normal handling for the case of faulting in non-tiled buffers */
370static int fault_1d(struct drm_gem_object *obj,
371 struct vm_area_struct *vma, struct vm_fault *vmf)
372{
373 struct omap_gem_object *omap_obj = to_omap_bo(obj);
374 unsigned long pfn;
375 pgoff_t pgoff;
376
377 /* We don't use vmf->pgoff since that has the fake offset: */
378 pgoff = ((unsigned long)vmf->virtual_address -
379 vma->vm_start) >> PAGE_SHIFT;
380
381 if (omap_obj->pages) {
382 omap_gem_cpu_sync(obj, pgoff);
383 pfn = page_to_pfn(omap_obj->pages[pgoff]);
384 } else {
385 BUG_ON(!(omap_obj->flags & OMAP_BO_DMA));
386 pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff;
387 }
388
389 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
390 pfn, pfn << PAGE_SHIFT);
391
392 return vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, pfn);
393}
394
395/* Special handling for the case of faulting in 2d tiled buffers */
396static int fault_2d(struct drm_gem_object *obj,
397 struct vm_area_struct *vma, struct vm_fault *vmf)
398{
399 struct omap_gem_object *omap_obj = to_omap_bo(obj);
400 struct usergart_entry *entry;
401 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
402 struct page *pages[64]; /* XXX is this too much to have on stack? */
403 unsigned long pfn;
404 pgoff_t pgoff, base_pgoff;
405 void __user *vaddr;
406 int i, ret, slots;
407
408 /*
409 * Note the height of the slot is also equal to the number of pages
410 * that need to be mapped in to fill 4kb wide CPU page. If the slot
411 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
412 */
413 const int n = usergart[fmt].height;
414 const int n_shift = usergart[fmt].height_shift;
415
416 /*
417 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
418 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
419 * into account in some of the math, so figure out virtual stride
420 * in pages
421 */
422 const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
423
424 /* We don't use vmf->pgoff since that has the fake offset: */
425 pgoff = ((unsigned long)vmf->virtual_address -
426 vma->vm_start) >> PAGE_SHIFT;
427
428 /*
429 * Actual address we start mapping at is rounded down to previous slot
430 * boundary in the y direction:
431 */
432 base_pgoff = round_down(pgoff, m << n_shift);
433
434 /* figure out buffer width in slots */
435 slots = omap_obj->width >> usergart[fmt].slot_shift;
436
437 vaddr = vmf->virtual_address - ((pgoff - base_pgoff) << PAGE_SHIFT);
438
439 entry = &usergart[fmt].entry[usergart[fmt].last];
440
441 /* evict previous buffer using this usergart entry, if any: */
442 if (entry->obj)
443 evict_entry(entry->obj, fmt, entry);
444
445 entry->obj = obj;
446 entry->obj_pgoff = base_pgoff;
447
448 /* now convert base_pgoff to phys offset from virt offset: */
449 base_pgoff = (base_pgoff >> n_shift) * slots;
450
451 /* for wider-than 4k.. figure out which part of the slot-row we want: */
452 if (m > 1) {
453 int off = pgoff % m;
454 entry->obj_pgoff += off;
455 base_pgoff /= m;
456 slots = min(slots - (off << n_shift), n);
457 base_pgoff += off << n_shift;
458 vaddr += off << PAGE_SHIFT;
459 }
460
461 /*
462 * Map in pages. Beyond the valid pixel part of the buffer, we set
463 * pages[i] to NULL to get a dummy page mapped in.. if someone
464 * reads/writes it they will get random/undefined content, but at
465 * least it won't be corrupting whatever other random page used to
466 * be mapped in, or other undefined behavior.
467 */
468 memcpy(pages, &omap_obj->pages[base_pgoff],
469 sizeof(struct page *) * slots);
470 memset(pages + slots, 0,
471 sizeof(struct page *) * (n - slots));
472
473 ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
474 if (ret) {
475 dev_err(obj->dev->dev, "failed to pin: %d\n", ret);
476 return ret;
477 }
478
479 pfn = entry->paddr >> PAGE_SHIFT;
480
481 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
482 pfn, pfn << PAGE_SHIFT);
483
484 for (i = n; i > 0; i--) {
485 vm_insert_mixed(vma, (unsigned long)vaddr, pfn);
486 pfn += usergart[fmt].stride_pfn;
487 vaddr += PAGE_SIZE * m;
488 }
489
490 /* simple round-robin: */
491 usergart[fmt].last = (usergart[fmt].last + 1) % NUM_USERGART_ENTRIES;
492
493 return 0;
494}
495
496/**
497 * omap_gem_fault - pagefault handler for GEM objects
498 * @vma: the VMA of the GEM object
499 * @vmf: fault detail
500 *
501 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
502 * does most of the work for us including the actual map/unmap calls
503 * but we need to do the actual page work.
504 *
505 * The VMA was set up by GEM. In doing so it also ensured that the
506 * vma->vm_private_data points to the GEM object that is backing this
507 * mapping.
508 */
509int omap_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
510{
511 struct drm_gem_object *obj = vma->vm_private_data;
512 struct omap_gem_object *omap_obj = to_omap_bo(obj);
513 struct drm_device *dev = obj->dev;
514 struct page **pages;
515 int ret;
516
517 /* Make sure we don't parallel update on a fault, nor move or remove
518 * something from beneath our feet
519 */
520 mutex_lock(&dev->struct_mutex);
521
522 /* if a shmem backed object, make sure we have pages attached now */
523 ret = get_pages(obj, &pages);
524 if (ret)
525 goto fail;
526
527 /* where should we do corresponding put_pages().. we are mapping
528 * the original page, rather than thru a GART, so we can't rely
529 * on eviction to trigger this. But munmap() or all mappings should
530 * probably trigger put_pages()?
531 */
532
533 if (omap_obj->flags & OMAP_BO_TILED)
534 ret = fault_2d(obj, vma, vmf);
535 else
536 ret = fault_1d(obj, vma, vmf);
537
538
539fail:
540 mutex_unlock(&dev->struct_mutex);
541 switch (ret) {
542 case 0:
543 case -ERESTARTSYS:
544 case -EINTR:
545 return VM_FAULT_NOPAGE;
546 case -ENOMEM:
547 return VM_FAULT_OOM;
548 default:
549 return VM_FAULT_SIGBUS;
550 }
551}
552
553/** We override mainly to fix up some of the vm mapping flags.. */
554int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
555{
556 int ret;
557
558 ret = drm_gem_mmap(filp, vma);
559 if (ret) {
560 DBG("mmap failed: %d", ret);
561 return ret;
562 }
563
564 return omap_gem_mmap_obj(vma->vm_private_data, vma);
565}
566
567int omap_gem_mmap_obj(struct drm_gem_object *obj,
568 struct vm_area_struct *vma)
569{
570 struct omap_gem_object *omap_obj = to_omap_bo(obj);
571
572 vma->vm_flags &= ~VM_PFNMAP;
573 vma->vm_flags |= VM_MIXEDMAP;
574
575 if (omap_obj->flags & OMAP_BO_WC) {
576 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
577 } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
578 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
579 } else {
580 /*
581 * We do have some private objects, at least for scanout buffers
582 * on hardware without DMM/TILER. But these are allocated write-
583 * combine
584 */
585 if (WARN_ON(!obj->filp))
586 return -EINVAL;
587
588 /*
589 * Shunt off cached objs to shmem file so they have their own
590 * address_space (so unmap_mapping_range does what we want,
591 * in particular in the case of mmap'd dmabufs)
592 */
593 fput(vma->vm_file);
594 vma->vm_pgoff = 0;
595 vma->vm_file = get_file(obj->filp);
596
597 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
598 }
599
600 return 0;
601}
602
603
604/**
605 * omap_gem_dumb_create - create a dumb buffer
606 * @drm_file: our client file
607 * @dev: our device
608 * @args: the requested arguments copied from userspace
609 *
610 * Allocate a buffer suitable for use for a frame buffer of the
611 * form described by user space. Give userspace a handle by which
612 * to reference it.
613 */
614int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
615 struct drm_mode_create_dumb *args)
616{
617 union omap_gem_size gsize;
618
619 /* in case someone tries to feed us a completely bogus stride: */
620 args->pitch = align_pitch(args->pitch, args->width, args->bpp);
621 args->size = PAGE_ALIGN(args->pitch * args->height);
622
623 gsize = (union omap_gem_size){
624 .bytes = args->size,
625 };
626
627 return omap_gem_new_handle(dev, file, gsize,
628 OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
629}
630
631/**
632 * omap_gem_dumb_map - buffer mapping for dumb interface
633 * @file: our drm client file
634 * @dev: drm device
635 * @handle: GEM handle to the object (from dumb_create)
636 *
637 * Do the necessary setup to allow the mapping of the frame buffer
638 * into user memory. We don't have to do much here at the moment.
639 */
640int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
641 uint32_t handle, uint64_t *offset)
642{
643 struct drm_gem_object *obj;
644 int ret = 0;
645
646 /* GEM does all our handle to object mapping */
647 obj = drm_gem_object_lookup(dev, file, handle);
648 if (obj == NULL) {
649 ret = -ENOENT;
650 goto fail;
651 }
652
653 *offset = omap_gem_mmap_offset(obj);
654
655 drm_gem_object_unreference_unlocked(obj);
656
657fail:
658 return ret;
659}
660
661/* Set scrolling position. This allows us to implement fast scrolling
662 * for console.
663 *
664 * Call only from non-atomic contexts.
665 */
666int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll)
667{
668 struct omap_gem_object *omap_obj = to_omap_bo(obj);
669 uint32_t npages = obj->size >> PAGE_SHIFT;
670 int ret = 0;
671
672 if (roll > npages) {
673 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
674 return -EINVAL;
675 }
676
677 omap_obj->roll = roll;
678
679 mutex_lock(&obj->dev->struct_mutex);
680
681 /* if we aren't mapped yet, we don't need to do anything */
682 if (omap_obj->block) {
683 struct page **pages;
684 ret = get_pages(obj, &pages);
685 if (ret)
686 goto fail;
687 ret = tiler_pin(omap_obj->block, pages, npages, roll, true);
688 if (ret)
689 dev_err(obj->dev->dev, "could not repin: %d\n", ret);
690 }
691
692fail:
693 mutex_unlock(&obj->dev->struct_mutex);
694
695 return ret;
696}
697
698/* Sync the buffer for CPU access.. note pages should already be
699 * attached, ie. omap_gem_get_pages()
700 */
701void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff)
702{
703 struct drm_device *dev = obj->dev;
704 struct omap_gem_object *omap_obj = to_omap_bo(obj);
705
706 if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) {
707 dma_unmap_page(dev->dev, omap_obj->addrs[pgoff],
708 PAGE_SIZE, DMA_BIDIRECTIONAL);
709 omap_obj->addrs[pgoff] = 0;
710 }
711}
712
713/* sync the buffer for DMA access */
714void omap_gem_dma_sync(struct drm_gem_object *obj,
715 enum dma_data_direction dir)
716{
717 struct drm_device *dev = obj->dev;
718 struct omap_gem_object *omap_obj = to_omap_bo(obj);
719
720 if (is_cached_coherent(obj)) {
721 int i, npages = obj->size >> PAGE_SHIFT;
722 struct page **pages = omap_obj->pages;
723 bool dirty = false;
724
725 for (i = 0; i < npages; i++) {
726 if (!omap_obj->addrs[i]) {
727 omap_obj->addrs[i] = dma_map_page(dev->dev, pages[i], 0,
728 PAGE_SIZE, DMA_BIDIRECTIONAL);
729 dirty = true;
730 }
731 }
732
733 if (dirty) {
734 unmap_mapping_range(obj->filp->f_mapping, 0,
735 omap_gem_mmap_size(obj), 1);
736 }
737 }
738}
739
740/* Get physical address for DMA.. if 'remap' is true, and the buffer is not
741 * already contiguous, remap it to pin in physically contiguous memory.. (ie.
742 * map in TILER)
743 */
744int omap_gem_get_paddr(struct drm_gem_object *obj,
745 dma_addr_t *paddr, bool remap)
746{
747 struct omap_drm_private *priv = obj->dev->dev_private;
748 struct omap_gem_object *omap_obj = to_omap_bo(obj);
749 int ret = 0;
750
751 mutex_lock(&obj->dev->struct_mutex);
752
753 if (remap && is_shmem(obj) && priv->has_dmm) {
754 if (omap_obj->paddr_cnt == 0) {
755 struct page **pages;
756 uint32_t npages = obj->size >> PAGE_SHIFT;
757 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
758 struct tiler_block *block;
759
760 BUG_ON(omap_obj->block);
761
762 ret = get_pages(obj, &pages);
763 if (ret)
764 goto fail;
765
766 if (omap_obj->flags & OMAP_BO_TILED) {
767 block = tiler_reserve_2d(fmt,
768 omap_obj->width,
769 omap_obj->height, 0);
770 } else {
771 block = tiler_reserve_1d(obj->size);
772 }
773
774 if (IS_ERR(block)) {
775 ret = PTR_ERR(block);
776 dev_err(obj->dev->dev,
777 "could not remap: %d (%d)\n", ret, fmt);
778 goto fail;
779 }
780
781 /* TODO: enable async refill.. */
782 ret = tiler_pin(block, pages, npages,
783 omap_obj->roll, true);
784 if (ret) {
785 tiler_release(block);
786 dev_err(obj->dev->dev,
787 "could not pin: %d\n", ret);
788 goto fail;
789 }
790
791 omap_obj->paddr = tiler_ssptr(block);
792 omap_obj->block = block;
793
794 DBG("got paddr: %08x", omap_obj->paddr);
795 }
796
797 omap_obj->paddr_cnt++;
798
799 *paddr = omap_obj->paddr;
800 } else if (omap_obj->flags & OMAP_BO_DMA) {
801 *paddr = omap_obj->paddr;
802 } else {
803 ret = -EINVAL;
804 goto fail;
805 }
806
807fail:
808 mutex_unlock(&obj->dev->struct_mutex);
809
810 return ret;
811}
812
813/* Release physical address, when DMA is no longer being performed.. this
814 * could potentially unpin and unmap buffers from TILER
815 */
816int omap_gem_put_paddr(struct drm_gem_object *obj)
817{
818 struct omap_gem_object *omap_obj = to_omap_bo(obj);
819 int ret = 0;
820
821 mutex_lock(&obj->dev->struct_mutex);
822 if (omap_obj->paddr_cnt > 0) {
823 omap_obj->paddr_cnt--;
824 if (omap_obj->paddr_cnt == 0) {
825 ret = tiler_unpin(omap_obj->block);
826 if (ret) {
827 dev_err(obj->dev->dev,
828 "could not unpin pages: %d\n", ret);
829 goto fail;
830 }
831 ret = tiler_release(omap_obj->block);
832 if (ret) {
833 dev_err(obj->dev->dev,
834 "could not release unmap: %d\n", ret);
835 }
836 omap_obj->block = NULL;
837 }
838 }
839fail:
840 mutex_unlock(&obj->dev->struct_mutex);
841 return ret;
842}
843
844/* Get rotated scanout address (only valid if already pinned), at the
845 * specified orientation and x,y offset from top-left corner of buffer
846 * (only valid for tiled 2d buffers)
847 */
848int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient,
849 int x, int y, dma_addr_t *paddr)
850{
851 struct omap_gem_object *omap_obj = to_omap_bo(obj);
852 int ret = -EINVAL;
853
854 mutex_lock(&obj->dev->struct_mutex);
855 if ((omap_obj->paddr_cnt > 0) && omap_obj->block &&
856 (omap_obj->flags & OMAP_BO_TILED)) {
857 *paddr = tiler_tsptr(omap_obj->block, orient, x, y);
858 ret = 0;
859 }
860 mutex_unlock(&obj->dev->struct_mutex);
861 return ret;
862}
863
864/* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
865int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient)
866{
867 struct omap_gem_object *omap_obj = to_omap_bo(obj);
868 int ret = -EINVAL;
869 if (omap_obj->flags & OMAP_BO_TILED)
870 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
871 return ret;
872}
873
874/* acquire pages when needed (for example, for DMA where physically
875 * contiguous buffer is not required
876 */
877static int get_pages(struct drm_gem_object *obj, struct page ***pages)
878{
879 struct omap_gem_object *omap_obj = to_omap_bo(obj);
880 int ret = 0;
881
882 if (is_shmem(obj) && !omap_obj->pages) {
883 ret = omap_gem_attach_pages(obj);
884 if (ret) {
885 dev_err(obj->dev->dev, "could not attach pages\n");
886 return ret;
887 }
888 }
889
890 /* TODO: even phys-contig.. we should have a list of pages? */
891 *pages = omap_obj->pages;
892
893 return 0;
894}
895
896/* if !remap, and we don't have pages backing, then fail, rather than
897 * increasing the pin count (which we don't really do yet anyways,
898 * because we don't support swapping pages back out). And 'remap'
899 * might not be quite the right name, but I wanted to keep it working
900 * similarly to omap_gem_get_paddr(). Note though that mutex is not
901 * aquired if !remap (because this can be called in atomic ctxt),
902 * but probably omap_gem_get_paddr() should be changed to work in the
903 * same way. If !remap, a matching omap_gem_put_pages() call is not
904 * required (and should not be made).
905 */
906int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
907 bool remap)
908{
909 int ret;
910 if (!remap) {
911 struct omap_gem_object *omap_obj = to_omap_bo(obj);
912 if (!omap_obj->pages)
913 return -ENOMEM;
914 *pages = omap_obj->pages;
915 return 0;
916 }
917 mutex_lock(&obj->dev->struct_mutex);
918 ret = get_pages(obj, pages);
919 mutex_unlock(&obj->dev->struct_mutex);
920 return ret;
921}
922
923/* release pages when DMA no longer being performed */
924int omap_gem_put_pages(struct drm_gem_object *obj)
925{
926 /* do something here if we dynamically attach/detach pages.. at
927 * least they would no longer need to be pinned if everyone has
928 * released the pages..
929 */
930 return 0;
931}
932
933/* Get kernel virtual address for CPU access.. this more or less only
934 * exists for omap_fbdev. This should be called with struct_mutex
935 * held.
936 */
937void *omap_gem_vaddr(struct drm_gem_object *obj)
938{
939 struct omap_gem_object *omap_obj = to_omap_bo(obj);
940 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
941 if (!omap_obj->vaddr) {
942 struct page **pages;
943 int ret = get_pages(obj, &pages);
944 if (ret)
945 return ERR_PTR(ret);
946 omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
947 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
948 }
949 return omap_obj->vaddr;
950}
951
952#ifdef CONFIG_PM
953/* re-pin objects in DMM in resume path: */
954int omap_gem_resume(struct device *dev)
955{
956 struct drm_device *drm_dev = dev_get_drvdata(dev);
957 struct omap_drm_private *priv = drm_dev->dev_private;
958 struct omap_gem_object *omap_obj;
959 int ret = 0;
960
961 list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
962 if (omap_obj->block) {
963 struct drm_gem_object *obj = &omap_obj->base;
964 uint32_t npages = obj->size >> PAGE_SHIFT;
965 WARN_ON(!omap_obj->pages); /* this can't happen */
966 ret = tiler_pin(omap_obj->block,
967 omap_obj->pages, npages,
968 omap_obj->roll, true);
969 if (ret) {
970 dev_err(dev, "could not repin: %d\n", ret);
971 return ret;
972 }
973 }
974 }
975
976 return 0;
977}
978#endif
979
980#ifdef CONFIG_DEBUG_FS
981void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
982{
983 struct omap_gem_object *omap_obj = to_omap_bo(obj);
984 uint64_t off;
985
986 off = drm_vma_node_start(&obj->vma_node);
987
988 seq_printf(m, "%08x: %2d (%2d) %08llx %08Zx (%2d) %p %4d",
989 omap_obj->flags, obj->name, obj->refcount.refcount.counter,
990 off, omap_obj->paddr, omap_obj->paddr_cnt,
991 omap_obj->vaddr, omap_obj->roll);
992
993 if (omap_obj->flags & OMAP_BO_TILED) {
994 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
995 if (omap_obj->block) {
996 struct tcm_area *area = &omap_obj->block->area;
997 seq_printf(m, " (%dx%d, %dx%d)",
998 area->p0.x, area->p0.y,
999 area->p1.x, area->p1.y);
1000 }
1001 } else {
1002 seq_printf(m, " %d", obj->size);
1003 }
1004
1005 seq_printf(m, "\n");
1006}
1007
1008void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1009{
1010 struct omap_gem_object *omap_obj;
1011 int count = 0;
1012 size_t size = 0;
1013
1014 list_for_each_entry(omap_obj, list, mm_list) {
1015 struct drm_gem_object *obj = &omap_obj->base;
1016 seq_printf(m, " ");
1017 omap_gem_describe(obj, m);
1018 count++;
1019 size += obj->size;
1020 }
1021
1022 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1023}
1024#endif
1025
1026/* Buffer Synchronization:
1027 */
1028
1029struct omap_gem_sync_waiter {
1030 struct list_head list;
1031 struct omap_gem_object *omap_obj;
1032 enum omap_gem_op op;
1033 uint32_t read_target, write_target;
1034 /* notify called w/ sync_lock held */
1035 void (*notify)(void *arg);
1036 void *arg;
1037};
1038
1039/* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1040 * the read and/or write target count is achieved which can call a user
1041 * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1042 * cpu access), etc.
1043 */
1044static LIST_HEAD(waiters);
1045
1046static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1047{
1048 struct omap_gem_object *omap_obj = waiter->omap_obj;
1049 if ((waiter->op & OMAP_GEM_READ) &&
1050 (omap_obj->sync->write_complete < waiter->write_target))
1051 return true;
1052 if ((waiter->op & OMAP_GEM_WRITE) &&
1053 (omap_obj->sync->read_complete < waiter->read_target))
1054 return true;
1055 return false;
1056}
1057
1058/* macro for sync debug.. */
1059#define SYNCDBG 0
1060#define SYNC(fmt, ...) do { if (SYNCDBG) \
1061 printk(KERN_ERR "%s:%d: "fmt"\n", \
1062 __func__, __LINE__, ##__VA_ARGS__); \
1063 } while (0)
1064
1065
1066static void sync_op_update(void)
1067{
1068 struct omap_gem_sync_waiter *waiter, *n;
1069 list_for_each_entry_safe(waiter, n, &waiters, list) {
1070 if (!is_waiting(waiter)) {
1071 list_del(&waiter->list);
1072 SYNC("notify: %p", waiter);
1073 waiter->notify(waiter->arg);
1074 kfree(waiter);
1075 }
1076 }
1077}
1078
1079static inline int sync_op(struct drm_gem_object *obj,
1080 enum omap_gem_op op, bool start)
1081{
1082 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1083 int ret = 0;
1084
1085 spin_lock(&sync_lock);
1086
1087 if (!omap_obj->sync) {
1088 omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1089 if (!omap_obj->sync) {
1090 ret = -ENOMEM;
1091 goto unlock;
1092 }
1093 }
1094
1095 if (start) {
1096 if (op & OMAP_GEM_READ)
1097 omap_obj->sync->read_pending++;
1098 if (op & OMAP_GEM_WRITE)
1099 omap_obj->sync->write_pending++;
1100 } else {
1101 if (op & OMAP_GEM_READ)
1102 omap_obj->sync->read_complete++;
1103 if (op & OMAP_GEM_WRITE)
1104 omap_obj->sync->write_complete++;
1105 sync_op_update();
1106 }
1107
1108unlock:
1109 spin_unlock(&sync_lock);
1110
1111 return ret;
1112}
1113
1114/* it is a bit lame to handle updates in this sort of polling way, but
1115 * in case of PVR, the GPU can directly update read/write complete
1116 * values, and not really tell us which ones it updated.. this also
1117 * means that sync_lock is not quite sufficient. So we'll need to
1118 * do something a bit better when it comes time to add support for
1119 * separate 2d hw..
1120 */
1121void omap_gem_op_update(void)
1122{
1123 spin_lock(&sync_lock);
1124 sync_op_update();
1125 spin_unlock(&sync_lock);
1126}
1127
1128/* mark the start of read and/or write operation */
1129int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1130{
1131 return sync_op(obj, op, true);
1132}
1133
1134int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1135{
1136 return sync_op(obj, op, false);
1137}
1138
1139static DECLARE_WAIT_QUEUE_HEAD(sync_event);
1140
1141static void sync_notify(void *arg)
1142{
1143 struct task_struct **waiter_task = arg;
1144 *waiter_task = NULL;
1145 wake_up_all(&sync_event);
1146}
1147
1148int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1149{
1150 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1151 int ret = 0;
1152 if (omap_obj->sync) {
1153 struct task_struct *waiter_task = current;
1154 struct omap_gem_sync_waiter *waiter =
1155 kzalloc(sizeof(*waiter), GFP_KERNEL);
1156
1157 if (!waiter)
1158 return -ENOMEM;
1159
1160 waiter->omap_obj = omap_obj;
1161 waiter->op = op;
1162 waiter->read_target = omap_obj->sync->read_pending;
1163 waiter->write_target = omap_obj->sync->write_pending;
1164 waiter->notify = sync_notify;
1165 waiter->arg = &waiter_task;
1166
1167 spin_lock(&sync_lock);
1168 if (is_waiting(waiter)) {
1169 SYNC("waited: %p", waiter);
1170 list_add_tail(&waiter->list, &waiters);
1171 spin_unlock(&sync_lock);
1172 ret = wait_event_interruptible(sync_event,
1173 (waiter_task == NULL));
1174 spin_lock(&sync_lock);
1175 if (waiter_task) {
1176 SYNC("interrupted: %p", waiter);
1177 /* we were interrupted */
1178 list_del(&waiter->list);
1179 waiter_task = NULL;
1180 } else {
1181 /* freed in sync_op_update() */
1182 waiter = NULL;
1183 }
1184 }
1185 spin_unlock(&sync_lock);
1186
1187 if (waiter)
1188 kfree(waiter);
1189 }
1190 return ret;
1191}
1192
1193/* call fxn(arg), either synchronously or asynchronously if the op
1194 * is currently blocked.. fxn() can be called from any context
1195 *
1196 * (TODO for now fxn is called back from whichever context calls
1197 * omap_gem_op_update().. but this could be better defined later
1198 * if needed)
1199 *
1200 * TODO more code in common w/ _sync()..
1201 */
1202int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1203 void (*fxn)(void *arg), void *arg)
1204{
1205 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1206 if (omap_obj->sync) {
1207 struct omap_gem_sync_waiter *waiter =
1208 kzalloc(sizeof(*waiter), GFP_ATOMIC);
1209
1210 if (!waiter)
1211 return -ENOMEM;
1212
1213 waiter->omap_obj = omap_obj;
1214 waiter->op = op;
1215 waiter->read_target = omap_obj->sync->read_pending;
1216 waiter->write_target = omap_obj->sync->write_pending;
1217 waiter->notify = fxn;
1218 waiter->arg = arg;
1219
1220 spin_lock(&sync_lock);
1221 if (is_waiting(waiter)) {
1222 SYNC("waited: %p", waiter);
1223 list_add_tail(&waiter->list, &waiters);
1224 spin_unlock(&sync_lock);
1225 return 0;
1226 }
1227
1228 spin_unlock(&sync_lock);
1229
1230 kfree(waiter);
1231 }
1232
1233 /* no waiting.. */
1234 fxn(arg);
1235
1236 return 0;
1237}
1238
1239/* special API so PVR can update the buffer to use a sync-object allocated
1240 * from it's sync-obj heap. Only used for a newly allocated (from PVR's
1241 * perspective) sync-object, so we overwrite the new syncobj w/ values
1242 * from the already allocated syncobj (if there is one)
1243 */
1244int omap_gem_set_sync_object(struct drm_gem_object *obj, void *syncobj)
1245{
1246 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1247 int ret = 0;
1248
1249 spin_lock(&sync_lock);
1250
1251 if ((omap_obj->flags & OMAP_BO_EXT_SYNC) && !syncobj) {
1252 /* clearing a previously set syncobj */
1253 syncobj = kmemdup(omap_obj->sync, sizeof(*omap_obj->sync),
1254 GFP_ATOMIC);
1255 if (!syncobj) {
1256 ret = -ENOMEM;
1257 goto unlock;
1258 }
1259 omap_obj->flags &= ~OMAP_BO_EXT_SYNC;
1260 omap_obj->sync = syncobj;
1261 } else if (syncobj && !(omap_obj->flags & OMAP_BO_EXT_SYNC)) {
1262 /* replacing an existing syncobj */
1263 if (omap_obj->sync) {
1264 memcpy(syncobj, omap_obj->sync, sizeof(*omap_obj->sync));
1265 kfree(omap_obj->sync);
1266 }
1267 omap_obj->flags |= OMAP_BO_EXT_SYNC;
1268 omap_obj->sync = syncobj;
1269 }
1270
1271unlock:
1272 spin_unlock(&sync_lock);
1273 return ret;
1274}
1275
1276/* don't call directly.. called from GEM core when it is time to actually
1277 * free the object..
1278 */
1279void omap_gem_free_object(struct drm_gem_object *obj)
1280{
1281 struct drm_device *dev = obj->dev;
1282 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1283
1284 evict(obj);
1285
1286 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1287
1288 list_del(&omap_obj->mm_list);
1289
1290 drm_gem_free_mmap_offset(obj);
1291
1292 /* this means the object is still pinned.. which really should
1293 * not happen. I think..
1294 */
1295 WARN_ON(omap_obj->paddr_cnt > 0);
1296
1297 /* don't free externally allocated backing memory */
1298 if (!(omap_obj->flags & OMAP_BO_EXT_MEM)) {
1299 if (omap_obj->pages)
1300 omap_gem_detach_pages(obj);
1301
1302 if (!is_shmem(obj)) {
1303 dma_free_writecombine(dev->dev, obj->size,
1304 omap_obj->vaddr, omap_obj->paddr);
1305 } else if (omap_obj->vaddr) {
1306 vunmap(omap_obj->vaddr);
1307 }
1308 }
1309
1310 /* don't free externally allocated syncobj */
1311 if (!(omap_obj->flags & OMAP_BO_EXT_SYNC))
1312 kfree(omap_obj->sync);
1313
1314 drm_gem_object_release(obj);
1315
1316 kfree(obj);
1317}
1318
1319/* convenience method to construct a GEM buffer object, and userspace handle */
1320int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1321 union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1322{
1323 struct drm_gem_object *obj;
1324 int ret;
1325
1326 obj = omap_gem_new(dev, gsize, flags);
1327 if (!obj)
1328 return -ENOMEM;
1329
1330 ret = drm_gem_handle_create(file, obj, handle);
1331 if (ret) {
1332 drm_gem_object_release(obj);
1333 kfree(obj); /* TODO isn't there a dtor to call? just copying i915 */
1334 return ret;
1335 }
1336
1337 /* drop reference from allocate - handle holds it now */
1338 drm_gem_object_unreference_unlocked(obj);
1339
1340 return 0;
1341}
1342
1343/* GEM buffer object constructor */
1344struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1345 union omap_gem_size gsize, uint32_t flags)
1346{
1347 struct omap_drm_private *priv = dev->dev_private;
1348 struct omap_gem_object *omap_obj;
1349 struct drm_gem_object *obj = NULL;
1350 size_t size;
1351 int ret;
1352
1353 if (flags & OMAP_BO_TILED) {
1354 if (!usergart) {
1355 dev_err(dev->dev, "Tiled buffers require DMM\n");
1356 goto fail;
1357 }
1358
1359 /* tiled buffers are always shmem paged backed.. when they are
1360 * scanned out, they are remapped into DMM/TILER
1361 */
1362 flags &= ~OMAP_BO_SCANOUT;
1363
1364 /* currently don't allow cached buffers.. there is some caching
1365 * stuff that needs to be handled better
1366 */
1367 flags &= ~(OMAP_BO_CACHED|OMAP_BO_UNCACHED);
1368 flags |= OMAP_BO_WC;
1369
1370 /* align dimensions to slot boundaries... */
1371 tiler_align(gem2fmt(flags),
1372 &gsize.tiled.width, &gsize.tiled.height);
1373
1374 /* ...and calculate size based on aligned dimensions */
1375 size = tiler_size(gem2fmt(flags),
1376 gsize.tiled.width, gsize.tiled.height);
1377 } else {
1378 size = PAGE_ALIGN(gsize.bytes);
1379 }
1380
1381 omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1382 if (!omap_obj)
1383 goto fail;
1384
1385 list_add(&omap_obj->mm_list, &priv->obj_list);
1386
1387 obj = &omap_obj->base;
1388
1389 if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1390 /* attempt to allocate contiguous memory if we don't
1391 * have DMM for remappign discontiguous buffers
1392 */
1393 omap_obj->vaddr = dma_alloc_writecombine(dev->dev, size,
1394 &omap_obj->paddr, GFP_KERNEL);
1395 if (omap_obj->vaddr)
1396 flags |= OMAP_BO_DMA;
1397
1398 }
1399
1400 omap_obj->flags = flags;
1401
1402 if (flags & OMAP_BO_TILED) {
1403 omap_obj->width = gsize.tiled.width;
1404 omap_obj->height = gsize.tiled.height;
1405 }
1406
1407 ret = 0;
1408 if (flags & (OMAP_BO_DMA|OMAP_BO_EXT_MEM))
1409 drm_gem_private_object_init(dev, obj, size);
1410 else
1411 ret = drm_gem_object_init(dev, obj, size);
1412
1413 if (ret)
1414 goto fail;
1415
1416 return obj;
1417
1418fail:
1419 if (obj)
1420 omap_gem_free_object(obj);
1421
1422 return NULL;
1423}
1424
1425/* init/cleanup.. if DMM is used, we need to set some stuff up.. */
1426void omap_gem_init(struct drm_device *dev)
1427{
1428 struct omap_drm_private *priv = dev->dev_private;
1429 const enum tiler_fmt fmts[] = {
1430 TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1431 };
1432 int i, j;
1433
1434 if (!dmm_is_available()) {
1435 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1436 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1437 return;
1438 }
1439
1440 usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1441 if (!usergart)
1442 return;
1443
1444 /* reserve 4k aligned/wide regions for userspace mappings: */
1445 for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1446 uint16_t h = 1, w = PAGE_SIZE >> i;
1447 tiler_align(fmts[i], &w, &h);
1448 /* note: since each region is 1 4kb page wide, and minimum
1449 * number of rows, the height ends up being the same as the
1450 * # of pages in the region
1451 */
1452 usergart[i].height = h;
1453 usergart[i].height_shift = ilog2(h);
1454 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1455 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1456 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1457 struct usergart_entry *entry = &usergart[i].entry[j];
1458 struct tiler_block *block =
1459 tiler_reserve_2d(fmts[i], w, h,
1460 PAGE_SIZE);
1461 if (IS_ERR(block)) {
1462 dev_err(dev->dev,
1463 "reserve failed: %d, %d, %ld\n",
1464 i, j, PTR_ERR(block));
1465 return;
1466 }
1467 entry->paddr = tiler_ssptr(block);
1468 entry->block = block;
1469
1470 DBG("%d:%d: %dx%d: paddr=%08x stride=%d", i, j, w, h,
1471 entry->paddr,
1472 usergart[i].stride_pfn << PAGE_SHIFT);
1473 }
1474 }
1475
1476 priv->has_dmm = true;
1477}
1478
1479void omap_gem_deinit(struct drm_device *dev)
1480{
1481 /* I believe we can rely on there being no more outstanding GEM
1482 * objects which could depend on usergart/dmm at this point.
1483 */
1484 kfree(usergart);
1485}