2 * drivers/staging/omapdrm/omap_gem.c
4 * Copyright (C) 2011 Texas Instruments
5 * Author: Rob Clark <rob.clark@linaro.org>
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.
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
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/>.
21 #include <linux/spinlock.h>
22 #include <linux/shmem_fs.h>
25 #include "omap_dmm_tiler.h"
27 /* remove these once drm core helpers are merged */
28 struct page ** _drm_gem_get_pages(struct drm_gem_object *obj, gfp_t gfpmask);
29 void _drm_gem_put_pages(struct drm_gem_object *obj, struct page **pages,
30 bool dirty, bool accessed);
31 int _drm_gem_create_mmap_offset_size(struct drm_gem_object *obj, size_t size);
34 * GEM buffer object implementation.
37 #define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
39 /* note: we use upper 8 bits of flags for driver-internal flags: */
40 #define OMAP_BO_DMA 0x01000000 /* actually is physically contiguous */
41 #define OMAP_BO_EXT_SYNC 0x02000000 /* externally allocated sync object */
42 #define OMAP_BO_EXT_MEM 0x04000000 /* externally allocated memory */
45 struct omap_gem_object {
46 struct drm_gem_object base;
48 struct list_head mm_list;
52 /** width/height for tiled formats (rounded up to slot boundaries) */
53 uint16_t width, height;
55 /** roll applied when mapping to DMM */
59 * If buffer is allocated physically contiguous, the OMAP_BO_DMA flag
60 * is set and the paddr is valid. Also if the buffer is remapped in
61 * TILER and paddr_cnt > 0, then paddr is valid. But if you are using
62 * the physical address and OMAP_BO_DMA is not set, then you should
63 * be going thru omap_gem_{get,put}_paddr() to ensure the mapping is
64 * not removed from under your feet.
66 * Note that OMAP_BO_SCANOUT is a hint from userspace that DMA capable
67 * buffer is requested, but doesn't mean that it is. Use the
68 * OMAP_BO_DMA flag to determine if the buffer has a DMA capable
79 * tiler block used when buffer is remapped in DMM/TILER.
81 struct tiler_block *block;
84 * Array of backing pages, if allocated. Note that pages are never
85 * allocated for buffers originally allocated from contiguous memory
89 /** addresses corresponding to pages in above array */
93 * Virtual address, if mapped.
98 * sync-object allocated on demand (if needed)
100 * Per-buffer sync-object for tracking pending and completed hw/dma
101 * read and write operations. The layout in memory is dictated by
102 * the SGX firmware, which uses this information to stall the command
103 * stream if a surface is not ready yet.
105 * Note that when buffer is used by SGX, the sync-object needs to be
106 * allocated from a special heap of sync-objects. This way many sync
107 * objects can be packed in a page, and not waste GPU virtual address
108 * space. Because of this we have to have a omap_gem_set_sync_object()
109 * API to allow replacement of the syncobj after it has (potentially)
110 * already been allocated. A bit ugly but I haven't thought of a
111 * better alternative.
114 uint32_t write_pending;
115 uint32_t write_complete;
116 uint32_t read_pending;
117 uint32_t read_complete;
121 static int get_pages(struct drm_gem_object *obj, struct page ***pages);
122 static uint64_t mmap_offset(struct drm_gem_object *obj);
124 /* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
125 * not necessarily pinned in TILER all the time, and (b) when they are
126 * they are not necessarily page aligned, we reserve one or more small
127 * regions in each of the 2d containers to use as a user-GART where we
128 * can create a second page-aligned mapping of parts of the buffer
129 * being accessed from userspace.
131 * Note that we could optimize slightly when we know that multiple
132 * tiler containers are backed by the same PAT.. but I'll leave that
135 #define NUM_USERGART_ENTRIES 2
136 struct usergart_entry {
137 struct tiler_block *block; /* the reserved tiler block */
139 struct drm_gem_object *obj; /* the current pinned obj */
140 pgoff_t obj_pgoff; /* page offset of obj currently
144 struct usergart_entry entry[NUM_USERGART_ENTRIES];
145 int height; /* height in rows */
146 int height_shift; /* ilog2(height in rows) */
147 int slot_shift; /* ilog2(width per slot) */
148 int stride_pfn; /* stride in pages */
149 int last; /* index of last used entry */
152 static void evict_entry(struct drm_gem_object *obj,
153 enum tiler_fmt fmt, struct usergart_entry *entry)
155 if (obj->dev->dev_mapping) {
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);
164 /* if stride > than PAGE_SIZE then sparse mapping: */
165 for (i = n; i > 0; i--) {
166 unmap_mapping_range(obj->dev->dev_mapping,
168 off += PAGE_SIZE * m;
171 unmap_mapping_range(obj->dev->dev_mapping, off, size, 1);
178 /* Evict a buffer from usergart, if it is mapped there */
179 static void evict(struct drm_gem_object *obj)
181 struct omap_gem_object *omap_obj = to_omap_bo(obj);
183 if (omap_obj->flags & OMAP_BO_TILED) {
184 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
190 for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
191 struct usergart_entry *entry = &usergart[fmt].entry[i];
192 if (entry->obj == obj)
193 evict_entry(obj, fmt, entry);
198 /* GEM objects can either be allocated from contiguous memory (in which
199 * case obj->filp==NULL), or w/ shmem backing (obj->filp!=NULL). But non
200 * contiguous buffers can be remapped in TILER/DMM if they need to be
201 * contiguous... but we don't do this all the time to reduce pressure
202 * on TILER/DMM space when we know at allocation time that the buffer
203 * will need to be scanned out.
205 static inline bool is_shmem(struct drm_gem_object *obj)
207 return obj->filp != NULL;
211 * shmem buffers that are mapped cached can simulate coherency via using
212 * page faulting to keep track of dirty pages
214 static inline bool is_cached_coherent(struct drm_gem_object *obj)
216 struct omap_gem_object *omap_obj = to_omap_bo(obj);
217 return is_shmem(obj) &&
218 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED);
221 static DEFINE_SPINLOCK(sync_lock);
223 /** ensure backing pages are allocated */
224 static int omap_gem_attach_pages(struct drm_gem_object *obj)
226 struct drm_device *dev = obj->dev;
227 struct omap_gem_object *omap_obj = to_omap_bo(obj);
229 int npages = obj->size >> PAGE_SHIFT;
233 WARN_ON(omap_obj->pages);
235 /* TODO: __GFP_DMA32 .. but somehow GFP_HIGHMEM is coming from the
236 * mapping_gfp_mask(mapping) which conflicts w/ GFP_DMA32.. probably
237 * we actually want CMA memory for it all anyways..
239 pages = _drm_gem_get_pages(obj, GFP_KERNEL);
241 dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
242 return PTR_ERR(pages);
245 /* for non-cached buffers, ensure the new pages are clean because
246 * DSS, GPU, etc. are not cache coherent:
248 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
249 addrs = kmalloc(npages * sizeof(addrs), GFP_KERNEL);
255 for (i = 0; i < npages; i++) {
256 addrs[i] = dma_map_page(dev->dev, pages[i],
257 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
260 addrs = kzalloc(npages * sizeof(addrs), GFP_KERNEL);
267 omap_obj->addrs = addrs;
268 omap_obj->pages = pages;
273 _drm_gem_put_pages(obj, pages, true, false);
278 /** release backing pages */
279 static void omap_gem_detach_pages(struct drm_gem_object *obj)
281 struct omap_gem_object *omap_obj = to_omap_bo(obj);
283 /* for non-cached buffers, ensure the new pages are clean because
284 * DSS, GPU, etc. are not cache coherent:
286 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
287 int i, npages = obj->size >> PAGE_SHIFT;
288 for (i = 0; i < npages; i++) {
289 dma_unmap_page(obj->dev->dev, omap_obj->addrs[i],
290 PAGE_SIZE, DMA_BIDIRECTIONAL);
294 kfree(omap_obj->addrs);
295 omap_obj->addrs = NULL;
297 _drm_gem_put_pages(obj, omap_obj->pages, true, false);
298 omap_obj->pages = NULL;
301 /* get buffer flags */
302 uint32_t omap_gem_flags(struct drm_gem_object *obj)
304 return to_omap_bo(obj)->flags;
307 /** get mmap offset */
308 static uint64_t mmap_offset(struct drm_gem_object *obj)
310 struct drm_device *dev = obj->dev;
312 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
314 if (!obj->map_list.map) {
315 /* Make it mmapable */
316 size_t size = omap_gem_mmap_size(obj);
317 int ret = _drm_gem_create_mmap_offset_size(obj, size);
320 dev_err(dev->dev, "could not allocate mmap offset\n");
325 return (uint64_t)obj->map_list.hash.key << PAGE_SHIFT;
328 uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj)
331 mutex_lock(&obj->dev->struct_mutex);
332 offset = mmap_offset(obj);
333 mutex_unlock(&obj->dev->struct_mutex);
338 size_t omap_gem_mmap_size(struct drm_gem_object *obj)
340 struct omap_gem_object *omap_obj = to_omap_bo(obj);
341 size_t size = obj->size;
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
350 size = tiler_vsize(gem2fmt(omap_obj->flags),
351 omap_obj->width, omap_obj->height);
357 /* get tiled size, returns -EINVAL if not tiled buffer */
358 int omap_gem_tiled_size(struct drm_gem_object *obj, uint16_t *w, uint16_t *h)
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;
369 /* Normal handling for the case of faulting in non-tiled buffers */
370 static int fault_1d(struct drm_gem_object *obj,
371 struct vm_area_struct *vma, struct vm_fault *vmf)
373 struct omap_gem_object *omap_obj = to_omap_bo(obj);
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;
381 if (omap_obj->pages) {
382 omap_gem_cpu_sync(obj, pgoff);
383 pfn = page_to_pfn(omap_obj->pages[pgoff]);
385 BUG_ON(!(omap_obj->flags & OMAP_BO_DMA));
386 pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff;
389 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
390 pfn, pfn << PAGE_SHIFT);
392 return vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, pfn);
395 /* Special handling for the case of faulting in 2d tiled buffers */
396 static int fault_2d(struct drm_gem_object *obj,
397 struct vm_area_struct *vma, struct vm_fault *vmf)
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? */
404 pgoff_t pgoff, base_pgoff;
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.
413 const int n = usergart[fmt].height;
414 const int n_shift = usergart[fmt].height_shift;
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
422 const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
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;
429 * Actual address we start mapping at is rounded down to previous slot
430 * boundary in the y direction:
432 base_pgoff = round_down(pgoff, m << n_shift);
434 /* figure out buffer width in slots */
435 slots = omap_obj->width >> usergart[fmt].slot_shift;
437 vaddr = vmf->virtual_address - ((pgoff - base_pgoff) << PAGE_SHIFT);
439 entry = &usergart[fmt].entry[usergart[fmt].last];
441 /* evict previous buffer using this usergart entry, if any: */
443 evict_entry(entry->obj, fmt, entry);
446 entry->obj_pgoff = base_pgoff;
448 /* now convert base_pgoff to phys offset from virt offset: */
449 base_pgoff = (base_pgoff >> n_shift) * slots;
451 /* for wider-than 4k.. figure out which part of the slot-row we want: */
454 entry->obj_pgoff += off;
456 slots = min(slots - (off << n_shift), n);
457 base_pgoff += off << n_shift;
458 vaddr += off << PAGE_SHIFT;
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.
468 memcpy(pages, &omap_obj->pages[base_pgoff],
469 sizeof(struct page *) * slots);
470 memset(pages + slots, 0,
471 sizeof(struct page *) * (n - slots));
473 ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
475 dev_err(obj->dev->dev, "failed to pin: %d\n", ret);
479 pfn = entry->paddr >> PAGE_SHIFT;
481 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
482 pfn, pfn << PAGE_SHIFT);
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;
490 /* simple round-robin: */
491 usergart[fmt].last = (usergart[fmt].last + 1) % NUM_USERGART_ENTRIES;
497 * omap_gem_fault - pagefault handler for GEM objects
498 * @vma: the VMA of the GEM object
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.
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
509 int omap_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
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;
517 /* Make sure we don't parallel update on a fault, nor move or remove
518 * something from beneath our feet
520 mutex_lock(&dev->struct_mutex);
522 /* if a shmem backed object, make sure we have pages attached now */
523 ret = get_pages(obj, &pages);
528 /* where should we do corresponding put_pages().. we are mapping
529 * the original page, rather than thru a GART, so we can't rely
530 * on eviction to trigger this. But munmap() or all mappings should
531 * probably trigger put_pages()?
534 if (omap_obj->flags & OMAP_BO_TILED)
535 ret = fault_2d(obj, vma, vmf);
537 ret = fault_1d(obj, vma, vmf);
541 mutex_unlock(&dev->struct_mutex);
546 return VM_FAULT_NOPAGE;
550 return VM_FAULT_SIGBUS;
554 /** We override mainly to fix up some of the vm mapping flags.. */
555 int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
559 ret = drm_gem_mmap(filp, vma);
561 DBG("mmap failed: %d", ret);
565 return omap_gem_mmap_obj(vma->vm_private_data, vma);
568 int omap_gem_mmap_obj(struct drm_gem_object *obj,
569 struct vm_area_struct *vma)
571 struct omap_gem_object *omap_obj = to_omap_bo(obj);
573 vma->vm_flags &= ~VM_PFNMAP;
574 vma->vm_flags |= VM_MIXEDMAP;
576 if (omap_obj->flags & OMAP_BO_WC) {
577 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
578 } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
579 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
582 * We do have some private objects, at least for scanout buffers
583 * on hardware without DMM/TILER. But these are allocated write-
586 if (WARN_ON(!obj->filp))
590 * Shunt off cached objs to shmem file so they have their own
591 * address_space (so unmap_mapping_range does what we want,
592 * in particular in the case of mmap'd dmabufs)
596 vma->vm_file = get_file(obj->filp);
598 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
606 * omap_gem_dumb_create - create a dumb buffer
607 * @drm_file: our client file
609 * @args: the requested arguments copied from userspace
611 * Allocate a buffer suitable for use for a frame buffer of the
612 * form described by user space. Give userspace a handle by which
615 int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
616 struct drm_mode_create_dumb *args)
618 union omap_gem_size gsize;
620 /* in case someone tries to feed us a completely bogus stride: */
621 args->pitch = align_pitch(args->pitch, args->width, args->bpp);
622 args->size = PAGE_ALIGN(args->pitch * args->height);
624 gsize = (union omap_gem_size){
628 return omap_gem_new_handle(dev, file, gsize,
629 OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
633 * omap_gem_dumb_destroy - destroy a dumb buffer
635 * @dev: our DRM device
636 * @handle: the object handle
638 * Destroy a handle that was created via omap_gem_dumb_create.
640 int omap_gem_dumb_destroy(struct drm_file *file, struct drm_device *dev,
643 /* No special work needed, drop the reference and see what falls out */
644 return drm_gem_handle_delete(file, handle);
648 * omap_gem_dumb_map - buffer mapping for dumb interface
649 * @file: our drm client file
651 * @handle: GEM handle to the object (from dumb_create)
653 * Do the necessary setup to allow the mapping of the frame buffer
654 * into user memory. We don't have to do much here at the moment.
656 int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
657 uint32_t handle, uint64_t *offset)
659 struct drm_gem_object *obj;
662 /* GEM does all our handle to object mapping */
663 obj = drm_gem_object_lookup(dev, file, handle);
669 *offset = omap_gem_mmap_offset(obj);
671 drm_gem_object_unreference_unlocked(obj);
677 /* Set scrolling position. This allows us to implement fast scrolling
680 * Call only from non-atomic contexts.
682 int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll)
684 struct omap_gem_object *omap_obj = to_omap_bo(obj);
685 uint32_t npages = obj->size >> PAGE_SHIFT;
689 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
693 omap_obj->roll = roll;
695 mutex_lock(&obj->dev->struct_mutex);
697 /* if we aren't mapped yet, we don't need to do anything */
698 if (omap_obj->block) {
700 ret = get_pages(obj, &pages);
703 ret = tiler_pin(omap_obj->block, pages, npages, roll, true);
705 dev_err(obj->dev->dev, "could not repin: %d\n", ret);
709 mutex_unlock(&obj->dev->struct_mutex);
714 /* Sync the buffer for CPU access.. note pages should already be
715 * attached, ie. omap_gem_get_pages()
717 void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff)
719 struct drm_device *dev = obj->dev;
720 struct omap_gem_object *omap_obj = to_omap_bo(obj);
722 if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) {
723 dma_unmap_page(dev->dev, omap_obj->addrs[pgoff],
724 PAGE_SIZE, DMA_BIDIRECTIONAL);
725 omap_obj->addrs[pgoff] = 0;
729 /* sync the buffer for DMA access */
730 void omap_gem_dma_sync(struct drm_gem_object *obj,
731 enum dma_data_direction dir)
733 struct drm_device *dev = obj->dev;
734 struct omap_gem_object *omap_obj = to_omap_bo(obj);
736 if (is_cached_coherent(obj)) {
737 int i, npages = obj->size >> PAGE_SHIFT;
738 struct page **pages = omap_obj->pages;
741 for (i = 0; i < npages; i++) {
742 if (!omap_obj->addrs[i]) {
743 omap_obj->addrs[i] = dma_map_page(dev->dev, pages[i], 0,
744 PAGE_SIZE, DMA_BIDIRECTIONAL);
750 unmap_mapping_range(obj->filp->f_mapping, 0,
751 omap_gem_mmap_size(obj), 1);
756 /* Get physical address for DMA.. if 'remap' is true, and the buffer is not
757 * already contiguous, remap it to pin in physically contiguous memory.. (ie.
760 int omap_gem_get_paddr(struct drm_gem_object *obj,
761 dma_addr_t *paddr, bool remap)
763 struct omap_drm_private *priv = obj->dev->dev_private;
764 struct omap_gem_object *omap_obj = to_omap_bo(obj);
767 mutex_lock(&obj->dev->struct_mutex);
769 if (remap && is_shmem(obj) && priv->has_dmm) {
770 if (omap_obj->paddr_cnt == 0) {
772 uint32_t npages = obj->size >> PAGE_SHIFT;
773 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
774 struct tiler_block *block;
776 BUG_ON(omap_obj->block);
778 ret = get_pages(obj, &pages);
782 if (omap_obj->flags & OMAP_BO_TILED) {
783 block = tiler_reserve_2d(fmt,
785 omap_obj->height, 0);
787 block = tiler_reserve_1d(obj->size);
791 ret = PTR_ERR(block);
792 dev_err(obj->dev->dev,
793 "could not remap: %d (%d)\n", ret, fmt);
797 /* TODO: enable async refill.. */
798 ret = tiler_pin(block, pages, npages,
799 omap_obj->roll, true);
801 tiler_release(block);
802 dev_err(obj->dev->dev,
803 "could not pin: %d\n", ret);
807 omap_obj->paddr = tiler_ssptr(block);
808 omap_obj->block = block;
810 DBG("got paddr: %08x", omap_obj->paddr);
813 omap_obj->paddr_cnt++;
815 *paddr = omap_obj->paddr;
816 } else if (omap_obj->flags & OMAP_BO_DMA) {
817 *paddr = omap_obj->paddr;
824 mutex_unlock(&obj->dev->struct_mutex);
829 /* Release physical address, when DMA is no longer being performed.. this
830 * could potentially unpin and unmap buffers from TILER
832 int omap_gem_put_paddr(struct drm_gem_object *obj)
834 struct omap_gem_object *omap_obj = to_omap_bo(obj);
837 mutex_lock(&obj->dev->struct_mutex);
838 if (omap_obj->paddr_cnt > 0) {
839 omap_obj->paddr_cnt--;
840 if (omap_obj->paddr_cnt == 0) {
841 ret = tiler_unpin(omap_obj->block);
843 dev_err(obj->dev->dev,
844 "could not unpin pages: %d\n", ret);
847 ret = tiler_release(omap_obj->block);
849 dev_err(obj->dev->dev,
850 "could not release unmap: %d\n", ret);
852 omap_obj->block = NULL;
856 mutex_unlock(&obj->dev->struct_mutex);
860 /* Get rotated scanout address (only valid if already pinned), at the
861 * specified orientation and x,y offset from top-left corner of buffer
862 * (only valid for tiled 2d buffers)
864 int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient,
865 int x, int y, dma_addr_t *paddr)
867 struct omap_gem_object *omap_obj = to_omap_bo(obj);
870 mutex_lock(&obj->dev->struct_mutex);
871 if ((omap_obj->paddr_cnt > 0) && omap_obj->block &&
872 (omap_obj->flags & OMAP_BO_TILED)) {
873 *paddr = tiler_tsptr(omap_obj->block, orient, x, y);
876 mutex_unlock(&obj->dev->struct_mutex);
880 /* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
881 int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient)
883 struct omap_gem_object *omap_obj = to_omap_bo(obj);
885 if (omap_obj->flags & OMAP_BO_TILED)
886 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
890 /* acquire pages when needed (for example, for DMA where physically
891 * contiguous buffer is not required
893 static int get_pages(struct drm_gem_object *obj, struct page ***pages)
895 struct omap_gem_object *omap_obj = to_omap_bo(obj);
898 if (is_shmem(obj) && !omap_obj->pages) {
899 ret = omap_gem_attach_pages(obj);
901 dev_err(obj->dev->dev, "could not attach pages\n");
906 /* TODO: even phys-contig.. we should have a list of pages? */
907 *pages = omap_obj->pages;
912 /* if !remap, and we don't have pages backing, then fail, rather than
913 * increasing the pin count (which we don't really do yet anyways,
914 * because we don't support swapping pages back out). And 'remap'
915 * might not be quite the right name, but I wanted to keep it working
916 * similarly to omap_gem_get_paddr(). Note though that mutex is not
917 * aquired if !remap (because this can be called in atomic ctxt),
918 * but probably omap_gem_get_paddr() should be changed to work in the
919 * same way. If !remap, a matching omap_gem_put_pages() call is not
920 * required (and should not be made).
922 int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
927 struct omap_gem_object *omap_obj = to_omap_bo(obj);
928 if (!omap_obj->pages)
930 *pages = omap_obj->pages;
933 mutex_lock(&obj->dev->struct_mutex);
934 ret = get_pages(obj, pages);
935 mutex_unlock(&obj->dev->struct_mutex);
939 /* release pages when DMA no longer being performed */
940 int omap_gem_put_pages(struct drm_gem_object *obj)
942 /* do something here if we dynamically attach/detach pages.. at
943 * least they would no longer need to be pinned if everyone has
944 * released the pages..
949 /* Get kernel virtual address for CPU access.. this more or less only
950 * exists for omap_fbdev. This should be called with struct_mutex
953 void *omap_gem_vaddr(struct drm_gem_object *obj)
955 struct omap_gem_object *omap_obj = to_omap_bo(obj);
956 WARN_ON(! mutex_is_locked(&obj->dev->struct_mutex));
957 if (!omap_obj->vaddr) {
959 int ret = get_pages(obj, &pages);
962 omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
963 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
965 return omap_obj->vaddr;
968 #ifdef CONFIG_DEBUG_FS
969 void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
971 struct drm_device *dev = obj->dev;
972 struct omap_gem_object *omap_obj = to_omap_bo(obj);
975 WARN_ON(! mutex_is_locked(&dev->struct_mutex));
977 if (obj->map_list.map)
978 off = (uint64_t)obj->map_list.hash.key;
980 seq_printf(m, "%08x: %2d (%2d) %08llx %08Zx (%2d) %p %4d",
981 omap_obj->flags, obj->name, obj->refcount.refcount.counter,
982 off, omap_obj->paddr, omap_obj->paddr_cnt,
983 omap_obj->vaddr, omap_obj->roll);
985 if (omap_obj->flags & OMAP_BO_TILED) {
986 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
987 if (omap_obj->block) {
988 struct tcm_area *area = &omap_obj->block->area;
989 seq_printf(m, " (%dx%d, %dx%d)",
990 area->p0.x, area->p0.y,
991 area->p1.x, area->p1.y);
994 seq_printf(m, " %d", obj->size);
1000 void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1002 struct omap_gem_object *omap_obj;
1006 list_for_each_entry(omap_obj, list, mm_list) {
1007 struct drm_gem_object *obj = &omap_obj->base;
1009 omap_gem_describe(obj, m);
1014 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1018 /* Buffer Synchronization:
1021 struct omap_gem_sync_waiter {
1022 struct list_head list;
1023 struct omap_gem_object *omap_obj;
1024 enum omap_gem_op op;
1025 uint32_t read_target, write_target;
1026 /* notify called w/ sync_lock held */
1027 void (*notify)(void *arg);
1031 /* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1032 * the read and/or write target count is achieved which can call a user
1033 * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1036 static LIST_HEAD(waiters);
1038 static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1040 struct omap_gem_object *omap_obj = waiter->omap_obj;
1041 if ((waiter->op & OMAP_GEM_READ) &&
1042 (omap_obj->sync->read_complete < waiter->read_target))
1044 if ((waiter->op & OMAP_GEM_WRITE) &&
1045 (omap_obj->sync->write_complete < waiter->write_target))
1050 /* macro for sync debug.. */
1052 #define SYNC(fmt, ...) do { if (SYNCDBG) \
1053 printk(KERN_ERR "%s:%d: "fmt"\n", \
1054 __func__, __LINE__, ##__VA_ARGS__); \
1058 static void sync_op_update(void)
1060 struct omap_gem_sync_waiter *waiter, *n;
1061 list_for_each_entry_safe(waiter, n, &waiters, list) {
1062 if (!is_waiting(waiter)) {
1063 list_del(&waiter->list);
1064 SYNC("notify: %p", waiter);
1065 waiter->notify(waiter->arg);
1071 static inline int sync_op(struct drm_gem_object *obj,
1072 enum omap_gem_op op, bool start)
1074 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1077 spin_lock(&sync_lock);
1079 if (!omap_obj->sync) {
1080 omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1081 if (!omap_obj->sync) {
1088 if (op & OMAP_GEM_READ)
1089 omap_obj->sync->read_pending++;
1090 if (op & OMAP_GEM_WRITE)
1091 omap_obj->sync->write_pending++;
1093 if (op & OMAP_GEM_READ)
1094 omap_obj->sync->read_complete++;
1095 if (op & OMAP_GEM_WRITE)
1096 omap_obj->sync->write_complete++;
1101 spin_unlock(&sync_lock);
1106 /* it is a bit lame to handle updates in this sort of polling way, but
1107 * in case of PVR, the GPU can directly update read/write complete
1108 * values, and not really tell us which ones it updated.. this also
1109 * means that sync_lock is not quite sufficient. So we'll need to
1110 * do something a bit better when it comes time to add support for
1113 void omap_gem_op_update(void)
1115 spin_lock(&sync_lock);
1117 spin_unlock(&sync_lock);
1120 /* mark the start of read and/or write operation */
1121 int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1123 return sync_op(obj, op, true);
1126 int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1128 return sync_op(obj, op, false);
1131 static DECLARE_WAIT_QUEUE_HEAD(sync_event);
1133 static void sync_notify(void *arg)
1135 struct task_struct **waiter_task = arg;
1136 *waiter_task = NULL;
1137 wake_up_all(&sync_event);
1140 int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1142 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1144 if (omap_obj->sync) {
1145 struct task_struct *waiter_task = current;
1146 struct omap_gem_sync_waiter *waiter =
1147 kzalloc(sizeof(*waiter), GFP_KERNEL);
1153 waiter->omap_obj = omap_obj;
1155 waiter->read_target = omap_obj->sync->read_pending;
1156 waiter->write_target = omap_obj->sync->write_pending;
1157 waiter->notify = sync_notify;
1158 waiter->arg = &waiter_task;
1160 spin_lock(&sync_lock);
1161 if (is_waiting(waiter)) {
1162 SYNC("waited: %p", waiter);
1163 list_add_tail(&waiter->list, &waiters);
1164 spin_unlock(&sync_lock);
1165 ret = wait_event_interruptible(sync_event,
1166 (waiter_task == NULL));
1167 spin_lock(&sync_lock);
1169 SYNC("interrupted: %p", waiter);
1170 /* we were interrupted */
1171 list_del(&waiter->list);
1174 /* freed in sync_op_update() */
1178 spin_unlock(&sync_lock);
1187 /* call fxn(arg), either synchronously or asynchronously if the op
1188 * is currently blocked.. fxn() can be called from any context
1190 * (TODO for now fxn is called back from whichever context calls
1191 * omap_gem_op_update().. but this could be better defined later
1194 * TODO more code in common w/ _sync()..
1196 int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1197 void (*fxn)(void *arg), void *arg)
1199 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1200 if (omap_obj->sync) {
1201 struct omap_gem_sync_waiter *waiter =
1202 kzalloc(sizeof(*waiter), GFP_ATOMIC);
1208 waiter->omap_obj = omap_obj;
1210 waiter->read_target = omap_obj->sync->read_pending;
1211 waiter->write_target = omap_obj->sync->write_pending;
1212 waiter->notify = fxn;
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);
1223 spin_unlock(&sync_lock);
1232 /* special API so PVR can update the buffer to use a sync-object allocated
1233 * from it's sync-obj heap. Only used for a newly allocated (from PVR's
1234 * perspective) sync-object, so we overwrite the new syncobj w/ values
1235 * from the already allocated syncobj (if there is one)
1237 int omap_gem_set_sync_object(struct drm_gem_object *obj, void *syncobj)
1239 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1242 spin_lock(&sync_lock);
1244 if ((omap_obj->flags & OMAP_BO_EXT_SYNC) && !syncobj) {
1245 /* clearing a previously set syncobj */
1246 syncobj = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1251 memcpy(syncobj, omap_obj->sync, sizeof(*omap_obj->sync));
1252 omap_obj->flags &= ~OMAP_BO_EXT_SYNC;
1253 omap_obj->sync = syncobj;
1254 } else if (syncobj && !(omap_obj->flags & OMAP_BO_EXT_SYNC)) {
1255 /* replacing an existing syncobj */
1256 if (omap_obj->sync) {
1257 memcpy(syncobj, omap_obj->sync, sizeof(*omap_obj->sync));
1258 kfree(omap_obj->sync);
1260 omap_obj->flags |= OMAP_BO_EXT_SYNC;
1261 omap_obj->sync = syncobj;
1265 spin_unlock(&sync_lock);
1269 int omap_gem_init_object(struct drm_gem_object *obj)
1271 return -EINVAL; /* unused */
1274 /* don't call directly.. called from GEM core when it is time to actually
1277 void omap_gem_free_object(struct drm_gem_object *obj)
1279 struct drm_device *dev = obj->dev;
1280 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1284 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1286 list_del(&omap_obj->mm_list);
1288 if (obj->map_list.map) {
1289 drm_gem_free_mmap_offset(obj);
1292 /* this means the object is still pinned.. which really should
1293 * not happen. I think..
1295 WARN_ON(omap_obj->paddr_cnt > 0);
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);
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);
1310 /* don't free externally allocated syncobj */
1311 if (!(omap_obj->flags & OMAP_BO_EXT_SYNC)) {
1312 kfree(omap_obj->sync);
1315 drm_gem_object_release(obj);
1320 /* convenience method to construct a GEM buffer object, and userspace handle */
1321 int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1322 union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1324 struct drm_gem_object *obj;
1327 obj = omap_gem_new(dev, gsize, flags);
1331 ret = drm_gem_handle_create(file, obj, handle);
1333 drm_gem_object_release(obj);
1334 kfree(obj); /* TODO isn't there a dtor to call? just copying i915 */
1338 /* drop reference from allocate - handle holds it now */
1339 drm_gem_object_unreference_unlocked(obj);
1344 /* GEM buffer object constructor */
1345 struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1346 union omap_gem_size gsize, uint32_t flags)
1348 struct omap_drm_private *priv = dev->dev_private;
1349 struct omap_gem_object *omap_obj;
1350 struct drm_gem_object *obj = NULL;
1354 if (flags & OMAP_BO_TILED) {
1356 dev_err(dev->dev, "Tiled buffers require DMM\n");
1360 /* tiled buffers are always shmem paged backed.. when they are
1361 * scanned out, they are remapped into DMM/TILER
1363 flags &= ~OMAP_BO_SCANOUT;
1365 /* currently don't allow cached buffers.. there is some caching
1366 * stuff that needs to be handled better
1368 flags &= ~(OMAP_BO_CACHED|OMAP_BO_UNCACHED);
1369 flags |= OMAP_BO_WC;
1371 /* align dimensions to slot boundaries... */
1372 tiler_align(gem2fmt(flags),
1373 &gsize.tiled.width, &gsize.tiled.height);
1375 /* ...and calculate size based on aligned dimensions */
1376 size = tiler_size(gem2fmt(flags),
1377 gsize.tiled.width, gsize.tiled.height);
1379 size = PAGE_ALIGN(gsize.bytes);
1382 omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1384 dev_err(dev->dev, "could not allocate GEM object\n");
1388 list_add(&omap_obj->mm_list, &priv->obj_list);
1390 obj = &omap_obj->base;
1392 if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1393 /* attempt to allocate contiguous memory if we don't
1394 * have DMM for remappign discontiguous buffers
1396 omap_obj->vaddr = dma_alloc_writecombine(dev->dev, size,
1397 &omap_obj->paddr, GFP_KERNEL);
1398 if (omap_obj->vaddr) {
1399 flags |= OMAP_BO_DMA;
1403 omap_obj->flags = flags;
1405 if (flags & OMAP_BO_TILED) {
1406 omap_obj->width = gsize.tiled.width;
1407 omap_obj->height = gsize.tiled.height;
1410 if (flags & (OMAP_BO_DMA|OMAP_BO_EXT_MEM)) {
1411 ret = drm_gem_private_object_init(dev, obj, size);
1413 ret = drm_gem_object_init(dev, obj, size);
1424 omap_gem_free_object(obj);
1429 /* init/cleanup.. if DMM is used, we need to set some stuff up.. */
1430 void omap_gem_init(struct drm_device *dev)
1432 struct omap_drm_private *priv = dev->dev_private;
1433 const enum tiler_fmt fmts[] = {
1434 TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1438 if (!dmm_is_initialized()) {
1439 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1440 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1444 usergart = kzalloc(3 * sizeof(*usergart), GFP_KERNEL);
1446 dev_warn(dev->dev, "could not allocate usergart\n");
1450 /* reserve 4k aligned/wide regions for userspace mappings: */
1451 for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1452 uint16_t h = 1, w = PAGE_SIZE >> i;
1453 tiler_align(fmts[i], &w, &h);
1454 /* note: since each region is 1 4kb page wide, and minimum
1455 * number of rows, the height ends up being the same as the
1456 * # of pages in the region
1458 usergart[i].height = h;
1459 usergart[i].height_shift = ilog2(h);
1460 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1461 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1462 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1463 struct usergart_entry *entry = &usergart[i].entry[j];
1464 struct tiler_block *block =
1465 tiler_reserve_2d(fmts[i], w, h,
1467 if (IS_ERR(block)) {
1469 "reserve failed: %d, %d, %ld\n",
1470 i, j, PTR_ERR(block));
1473 entry->paddr = tiler_ssptr(block);
1474 entry->block = block;
1476 DBG("%d:%d: %dx%d: paddr=%08x stride=%d", i, j, w, h,
1478 usergart[i].stride_pfn << PAGE_SHIFT);
1482 priv->has_dmm = true;
1485 void omap_gem_deinit(struct drm_device *dev)
1487 /* I believe we can rely on there being no more outstanding GEM
1488 * objects which could depend on usergart/dmm at this point.