2 * drivers/gpu/drm/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/>.
20 #include <linux/seq_file.h>
21 #include <linux/shmem_fs.h>
22 #include <linux/spinlock.h>
23 #include <linux/pfn_t.h>
25 #include <drm/drm_vma_manager.h>
28 #include "omap_dmm_tiler.h"
31 * GEM buffer object implementation.
34 /* note: we use upper 8 bits of flags for driver-internal flags: */
35 #define OMAP_BO_MEM_DMA_API 0x01000000 /* memory allocated with the dma_alloc_* API */
36 #define OMAP_BO_MEM_SHMEM 0x02000000 /* memory allocated through shmem backing */
37 #define OMAP_BO_MEM_DMABUF 0x08000000 /* memory imported from a dmabuf */
39 struct omap_gem_object {
40 struct drm_gem_object base;
42 struct list_head mm_list;
46 /** width/height for tiled formats (rounded up to slot boundaries) */
47 uint16_t width, height;
49 /** roll applied when mapping to DMM */
53 * paddr contains the buffer DMA address. It is valid for
55 * - buffers allocated through the DMA mapping API (with the
56 * OMAP_BO_MEM_DMA_API flag set)
58 * - buffers imported from dmabuf (with the OMAP_BO_MEM_DMABUF flag set)
59 * if they are physically contiguous (when sgt->orig_nents == 1)
61 * - buffers mapped through the TILER when paddr_cnt is not zero, in
62 * which case the DMA address points to the TILER aperture
64 * Physically contiguous buffers have their DMA address equal to the
65 * physical address as we don't remap those buffers through the TILER.
67 * Buffers mapped to the TILER have their DMA address pointing to the
68 * TILER aperture. As TILER mappings are refcounted (through paddr_cnt)
69 * the DMA address must be accessed through omap_get_get_paddr() to
70 * ensure that the mapping won't disappear unexpectedly. References must
71 * be released with omap_gem_put_paddr().
81 * If the buffer has been imported from a dmabuf the OMAP_DB_DMABUF flag
82 * is set and the sgt field is valid.
87 * tiler block used when buffer is remapped in DMM/TILER.
89 struct tiler_block *block;
92 * Array of backing pages, if allocated. Note that pages are never
93 * allocated for buffers originally allocated from contiguous memory
97 /** addresses corresponding to pages in above array */
101 * Virtual address, if mapped.
106 * sync-object allocated on demand (if needed)
108 * Per-buffer sync-object for tracking pending and completed hw/dma
109 * read and write operations.
112 uint32_t write_pending;
113 uint32_t write_complete;
114 uint32_t read_pending;
115 uint32_t read_complete;
119 #define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
121 /* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
122 * not necessarily pinned in TILER all the time, and (b) when they are
123 * they are not necessarily page aligned, we reserve one or more small
124 * regions in each of the 2d containers to use as a user-GART where we
125 * can create a second page-aligned mapping of parts of the buffer
126 * being accessed from userspace.
128 * Note that we could optimize slightly when we know that multiple
129 * tiler containers are backed by the same PAT.. but I'll leave that
132 #define NUM_USERGART_ENTRIES 2
133 struct omap_drm_usergart_entry {
134 struct tiler_block *block; /* the reserved tiler block */
136 struct drm_gem_object *obj; /* the current pinned obj */
137 pgoff_t obj_pgoff; /* page offset of obj currently
141 struct omap_drm_usergart {
142 struct omap_drm_usergart_entry entry[NUM_USERGART_ENTRIES];
143 int height; /* height in rows */
144 int height_shift; /* ilog2(height in rows) */
145 int slot_shift; /* ilog2(width per slot) */
146 int stride_pfn; /* stride in pages */
147 int last; /* index of last used entry */
150 /* -----------------------------------------------------------------------------
154 /** get mmap offset */
155 static uint64_t mmap_offset(struct drm_gem_object *obj)
157 struct drm_device *dev = obj->dev;
161 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
163 /* Make it mmapable */
164 size = omap_gem_mmap_size(obj);
165 ret = drm_gem_create_mmap_offset_size(obj, size);
167 dev_err(dev->dev, "could not allocate mmap offset\n");
171 return drm_vma_node_offset_addr(&obj->vma_node);
174 static bool is_contiguous(struct omap_gem_object *omap_obj)
176 if (omap_obj->flags & OMAP_BO_MEM_DMA_API)
179 if ((omap_obj->flags & OMAP_BO_MEM_DMABUF) && omap_obj->sgt->nents == 1)
185 /* -----------------------------------------------------------------------------
189 static void evict_entry(struct drm_gem_object *obj,
190 enum tiler_fmt fmt, struct omap_drm_usergart_entry *entry)
192 struct omap_gem_object *omap_obj = to_omap_bo(obj);
193 struct omap_drm_private *priv = obj->dev->dev_private;
194 int n = priv->usergart[fmt].height;
195 size_t size = PAGE_SIZE * n;
196 loff_t off = mmap_offset(obj) +
197 (entry->obj_pgoff << PAGE_SHIFT);
198 const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
202 /* if stride > than PAGE_SIZE then sparse mapping: */
203 for (i = n; i > 0; i--) {
204 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
206 off += PAGE_SIZE * m;
209 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
216 /* Evict a buffer from usergart, if it is mapped there */
217 static void evict(struct drm_gem_object *obj)
219 struct omap_gem_object *omap_obj = to_omap_bo(obj);
220 struct omap_drm_private *priv = obj->dev->dev_private;
222 if (omap_obj->flags & OMAP_BO_TILED) {
223 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
226 for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
227 struct omap_drm_usergart_entry *entry =
228 &priv->usergart[fmt].entry[i];
230 if (entry->obj == obj)
231 evict_entry(obj, fmt, entry);
236 /* -----------------------------------------------------------------------------
240 /** ensure backing pages are allocated */
241 static int omap_gem_attach_pages(struct drm_gem_object *obj)
243 struct drm_device *dev = obj->dev;
244 struct omap_gem_object *omap_obj = to_omap_bo(obj);
246 int npages = obj->size >> PAGE_SHIFT;
250 WARN_ON(omap_obj->pages);
252 pages = drm_gem_get_pages(obj);
254 dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
255 return PTR_ERR(pages);
258 /* for non-cached buffers, ensure the new pages are clean because
259 * DSS, GPU, etc. are not cache coherent:
261 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
262 addrs = kmalloc(npages * sizeof(*addrs), GFP_KERNEL);
268 for (i = 0; i < npages; i++) {
269 addrs[i] = dma_map_page(dev->dev, pages[i],
270 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
272 if (dma_mapping_error(dev->dev, addrs[i])) {
274 "%s: failed to map page\n", __func__);
276 for (i = i - 1; i >= 0; --i) {
277 dma_unmap_page(dev->dev, addrs[i],
278 PAGE_SIZE, DMA_BIDIRECTIONAL);
286 addrs = kzalloc(npages * sizeof(*addrs), GFP_KERNEL);
293 omap_obj->addrs = addrs;
294 omap_obj->pages = pages;
301 drm_gem_put_pages(obj, pages, true, false);
306 /* acquire pages when needed (for example, for DMA where physically
307 * contiguous buffer is not required
309 static int get_pages(struct drm_gem_object *obj, struct page ***pages)
311 struct omap_gem_object *omap_obj = to_omap_bo(obj);
314 if ((omap_obj->flags & OMAP_BO_MEM_SHMEM) && !omap_obj->pages) {
315 ret = omap_gem_attach_pages(obj);
317 dev_err(obj->dev->dev, "could not attach pages\n");
322 /* TODO: even phys-contig.. we should have a list of pages? */
323 *pages = omap_obj->pages;
328 /** release backing pages */
329 static void omap_gem_detach_pages(struct drm_gem_object *obj)
331 struct omap_gem_object *omap_obj = to_omap_bo(obj);
333 /* for non-cached buffers, ensure the new pages are clean because
334 * DSS, GPU, etc. are not cache coherent:
336 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
337 int i, npages = obj->size >> PAGE_SHIFT;
338 for (i = 0; i < npages; i++) {
339 dma_unmap_page(obj->dev->dev, omap_obj->addrs[i],
340 PAGE_SIZE, DMA_BIDIRECTIONAL);
344 kfree(omap_obj->addrs);
345 omap_obj->addrs = NULL;
347 drm_gem_put_pages(obj, omap_obj->pages, true, false);
348 omap_obj->pages = NULL;
351 /* get buffer flags */
352 uint32_t omap_gem_flags(struct drm_gem_object *obj)
354 return to_omap_bo(obj)->flags;
357 uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj)
360 mutex_lock(&obj->dev->struct_mutex);
361 offset = mmap_offset(obj);
362 mutex_unlock(&obj->dev->struct_mutex);
367 size_t omap_gem_mmap_size(struct drm_gem_object *obj)
369 struct omap_gem_object *omap_obj = to_omap_bo(obj);
370 size_t size = obj->size;
372 if (omap_obj->flags & OMAP_BO_TILED) {
373 /* for tiled buffers, the virtual size has stride rounded up
374 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
375 * 32kb later!). But we don't back the entire buffer with
376 * pages, only the valid picture part.. so need to adjust for
377 * this in the size used to mmap and generate mmap offset
379 size = tiler_vsize(gem2fmt(omap_obj->flags),
380 omap_obj->width, omap_obj->height);
386 /* get tiled size, returns -EINVAL if not tiled buffer */
387 int omap_gem_tiled_size(struct drm_gem_object *obj, uint16_t *w, uint16_t *h)
389 struct omap_gem_object *omap_obj = to_omap_bo(obj);
390 if (omap_obj->flags & OMAP_BO_TILED) {
391 *w = omap_obj->width;
392 *h = omap_obj->height;
398 /* -----------------------------------------------------------------------------
402 /* Normal handling for the case of faulting in non-tiled buffers */
403 static int fault_1d(struct drm_gem_object *obj,
404 struct vm_area_struct *vma, struct vm_fault *vmf)
406 struct omap_gem_object *omap_obj = to_omap_bo(obj);
410 /* We don't use vmf->pgoff since that has the fake offset: */
411 pgoff = ((unsigned long)vmf->virtual_address -
412 vma->vm_start) >> PAGE_SHIFT;
414 if (omap_obj->pages) {
415 omap_gem_cpu_sync(obj, pgoff);
416 pfn = page_to_pfn(omap_obj->pages[pgoff]);
418 BUG_ON(!is_contiguous(omap_obj));
419 pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff;
422 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
423 pfn, pfn << PAGE_SHIFT);
425 return vm_insert_mixed(vma, (unsigned long)vmf->virtual_address,
426 __pfn_to_pfn_t(pfn, PFN_DEV));
429 /* Special handling for the case of faulting in 2d tiled buffers */
430 static int fault_2d(struct drm_gem_object *obj,
431 struct vm_area_struct *vma, struct vm_fault *vmf)
433 struct omap_gem_object *omap_obj = to_omap_bo(obj);
434 struct omap_drm_private *priv = obj->dev->dev_private;
435 struct omap_drm_usergart_entry *entry;
436 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
437 struct page *pages[64]; /* XXX is this too much to have on stack? */
439 pgoff_t pgoff, base_pgoff;
444 * Note the height of the slot is also equal to the number of pages
445 * that need to be mapped in to fill 4kb wide CPU page. If the slot
446 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
448 const int n = priv->usergart[fmt].height;
449 const int n_shift = priv->usergart[fmt].height_shift;
452 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
453 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
454 * into account in some of the math, so figure out virtual stride
457 const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE);
459 /* We don't use vmf->pgoff since that has the fake offset: */
460 pgoff = ((unsigned long)vmf->virtual_address -
461 vma->vm_start) >> PAGE_SHIFT;
464 * Actual address we start mapping at is rounded down to previous slot
465 * boundary in the y direction:
467 base_pgoff = round_down(pgoff, m << n_shift);
469 /* figure out buffer width in slots */
470 slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
472 vaddr = vmf->virtual_address - ((pgoff - base_pgoff) << PAGE_SHIFT);
474 entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
476 /* evict previous buffer using this usergart entry, if any: */
478 evict_entry(entry->obj, fmt, entry);
481 entry->obj_pgoff = base_pgoff;
483 /* now convert base_pgoff to phys offset from virt offset: */
484 base_pgoff = (base_pgoff >> n_shift) * slots;
486 /* for wider-than 4k.. figure out which part of the slot-row we want: */
489 entry->obj_pgoff += off;
491 slots = min(slots - (off << n_shift), n);
492 base_pgoff += off << n_shift;
493 vaddr += off << PAGE_SHIFT;
497 * Map in pages. Beyond the valid pixel part of the buffer, we set
498 * pages[i] to NULL to get a dummy page mapped in.. if someone
499 * reads/writes it they will get random/undefined content, but at
500 * least it won't be corrupting whatever other random page used to
501 * be mapped in, or other undefined behavior.
503 memcpy(pages, &omap_obj->pages[base_pgoff],
504 sizeof(struct page *) * slots);
505 memset(pages + slots, 0,
506 sizeof(struct page *) * (n - slots));
508 ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
510 dev_err(obj->dev->dev, "failed to pin: %d\n", ret);
514 pfn = entry->paddr >> PAGE_SHIFT;
516 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
517 pfn, pfn << PAGE_SHIFT);
519 for (i = n; i > 0; i--) {
520 vm_insert_mixed(vma, (unsigned long)vaddr,
521 __pfn_to_pfn_t(pfn, PFN_DEV));
522 pfn += priv->usergart[fmt].stride_pfn;
523 vaddr += PAGE_SIZE * m;
526 /* simple round-robin: */
527 priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
528 % NUM_USERGART_ENTRIES;
534 * omap_gem_fault - pagefault handler for GEM objects
535 * @vma: the VMA of the GEM object
538 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
539 * does most of the work for us including the actual map/unmap calls
540 * but we need to do the actual page work.
542 * The VMA was set up by GEM. In doing so it also ensured that the
543 * vma->vm_private_data points to the GEM object that is backing this
546 int omap_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
548 struct drm_gem_object *obj = vma->vm_private_data;
549 struct omap_gem_object *omap_obj = to_omap_bo(obj);
550 struct drm_device *dev = obj->dev;
554 /* Make sure we don't parallel update on a fault, nor move or remove
555 * something from beneath our feet
557 mutex_lock(&dev->struct_mutex);
559 /* if a shmem backed object, make sure we have pages attached now */
560 ret = get_pages(obj, &pages);
564 /* where should we do corresponding put_pages().. we are mapping
565 * the original page, rather than thru a GART, so we can't rely
566 * on eviction to trigger this. But munmap() or all mappings should
567 * probably trigger put_pages()?
570 if (omap_obj->flags & OMAP_BO_TILED)
571 ret = fault_2d(obj, vma, vmf);
573 ret = fault_1d(obj, vma, vmf);
577 mutex_unlock(&dev->struct_mutex);
584 * EBUSY is ok: this just means that another thread
585 * already did the job.
587 return VM_FAULT_NOPAGE;
591 return VM_FAULT_SIGBUS;
595 /** We override mainly to fix up some of the vm mapping flags.. */
596 int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
600 ret = drm_gem_mmap(filp, vma);
602 DBG("mmap failed: %d", ret);
606 return omap_gem_mmap_obj(vma->vm_private_data, vma);
609 int omap_gem_mmap_obj(struct drm_gem_object *obj,
610 struct vm_area_struct *vma)
612 struct omap_gem_object *omap_obj = to_omap_bo(obj);
614 vma->vm_flags &= ~VM_PFNMAP;
615 vma->vm_flags |= VM_MIXEDMAP;
617 if (omap_obj->flags & OMAP_BO_WC) {
618 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
619 } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
620 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
623 * We do have some private objects, at least for scanout buffers
624 * on hardware without DMM/TILER. But these are allocated write-
627 if (WARN_ON(!obj->filp))
631 * Shunt off cached objs to shmem file so they have their own
632 * address_space (so unmap_mapping_range does what we want,
633 * in particular in the case of mmap'd dmabufs)
637 vma->vm_file = get_file(obj->filp);
639 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
645 /* -----------------------------------------------------------------------------
650 * omap_gem_dumb_create - create a dumb buffer
651 * @drm_file: our client file
653 * @args: the requested arguments copied from userspace
655 * Allocate a buffer suitable for use for a frame buffer of the
656 * form described by user space. Give userspace a handle by which
659 int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
660 struct drm_mode_create_dumb *args)
662 union omap_gem_size gsize;
664 args->pitch = align_pitch(0, args->width, args->bpp);
665 args->size = PAGE_ALIGN(args->pitch * args->height);
667 gsize = (union omap_gem_size){
671 return omap_gem_new_handle(dev, file, gsize,
672 OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
676 * omap_gem_dumb_map - buffer mapping for dumb interface
677 * @file: our drm client file
679 * @handle: GEM handle to the object (from dumb_create)
681 * Do the necessary setup to allow the mapping of the frame buffer
682 * into user memory. We don't have to do much here at the moment.
684 int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
685 uint32_t handle, uint64_t *offset)
687 struct drm_gem_object *obj;
690 /* GEM does all our handle to object mapping */
691 obj = drm_gem_object_lookup(file, handle);
697 *offset = omap_gem_mmap_offset(obj);
699 drm_gem_object_unreference_unlocked(obj);
705 #ifdef CONFIG_DRM_FBDEV_EMULATION
706 /* Set scrolling position. This allows us to implement fast scrolling
709 * Call only from non-atomic contexts.
711 int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll)
713 struct omap_gem_object *omap_obj = to_omap_bo(obj);
714 uint32_t npages = obj->size >> PAGE_SHIFT;
718 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
722 omap_obj->roll = roll;
724 mutex_lock(&obj->dev->struct_mutex);
726 /* if we aren't mapped yet, we don't need to do anything */
727 if (omap_obj->block) {
729 ret = get_pages(obj, &pages);
732 ret = tiler_pin(omap_obj->block, pages, npages, roll, true);
734 dev_err(obj->dev->dev, "could not repin: %d\n", ret);
738 mutex_unlock(&obj->dev->struct_mutex);
744 /* -----------------------------------------------------------------------------
745 * Memory Management & DMA Sync
749 * shmem buffers that are mapped cached can simulate coherency via using
750 * page faulting to keep track of dirty pages
752 static inline bool is_cached_coherent(struct drm_gem_object *obj)
754 struct omap_gem_object *omap_obj = to_omap_bo(obj);
756 return (omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
757 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED);
760 /* Sync the buffer for CPU access.. note pages should already be
761 * attached, ie. omap_gem_get_pages()
763 void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff)
765 struct drm_device *dev = obj->dev;
766 struct omap_gem_object *omap_obj = to_omap_bo(obj);
768 if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) {
769 dma_unmap_page(dev->dev, omap_obj->addrs[pgoff],
770 PAGE_SIZE, DMA_BIDIRECTIONAL);
771 omap_obj->addrs[pgoff] = 0;
775 /* sync the buffer for DMA access */
776 void omap_gem_dma_sync(struct drm_gem_object *obj,
777 enum dma_data_direction dir)
779 struct drm_device *dev = obj->dev;
780 struct omap_gem_object *omap_obj = to_omap_bo(obj);
782 if (is_cached_coherent(obj)) {
783 int i, npages = obj->size >> PAGE_SHIFT;
784 struct page **pages = omap_obj->pages;
787 for (i = 0; i < npages; i++) {
788 if (!omap_obj->addrs[i]) {
791 addr = dma_map_page(dev->dev, pages[i], 0,
792 PAGE_SIZE, DMA_BIDIRECTIONAL);
794 if (dma_mapping_error(dev->dev, addr)) {
796 "%s: failed to map page\n",
802 omap_obj->addrs[i] = addr;
807 unmap_mapping_range(obj->filp->f_mapping, 0,
808 omap_gem_mmap_size(obj), 1);
813 /* Get physical address for DMA.. if 'remap' is true, and the buffer is not
814 * already contiguous, remap it to pin in physically contiguous memory.. (ie.
817 int omap_gem_get_paddr(struct drm_gem_object *obj,
818 dma_addr_t *paddr, bool remap)
820 struct omap_drm_private *priv = obj->dev->dev_private;
821 struct omap_gem_object *omap_obj = to_omap_bo(obj);
824 mutex_lock(&obj->dev->struct_mutex);
826 if (!is_contiguous(omap_obj) && remap && priv->has_dmm) {
827 if (omap_obj->paddr_cnt == 0) {
829 uint32_t npages = obj->size >> PAGE_SHIFT;
830 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
831 struct tiler_block *block;
833 BUG_ON(omap_obj->block);
835 ret = get_pages(obj, &pages);
839 if (omap_obj->flags & OMAP_BO_TILED) {
840 block = tiler_reserve_2d(fmt,
842 omap_obj->height, 0);
844 block = tiler_reserve_1d(obj->size);
848 ret = PTR_ERR(block);
849 dev_err(obj->dev->dev,
850 "could not remap: %d (%d)\n", ret, fmt);
854 /* TODO: enable async refill.. */
855 ret = tiler_pin(block, pages, npages,
856 omap_obj->roll, true);
858 tiler_release(block);
859 dev_err(obj->dev->dev,
860 "could not pin: %d\n", ret);
864 omap_obj->paddr = tiler_ssptr(block);
865 omap_obj->block = block;
867 DBG("got paddr: %pad", &omap_obj->paddr);
870 omap_obj->paddr_cnt++;
872 *paddr = omap_obj->paddr;
873 } else if (is_contiguous(omap_obj)) {
874 *paddr = omap_obj->paddr;
881 mutex_unlock(&obj->dev->struct_mutex);
886 /* Release physical address, when DMA is no longer being performed.. this
887 * could potentially unpin and unmap buffers from TILER
889 void omap_gem_put_paddr(struct drm_gem_object *obj)
891 struct omap_gem_object *omap_obj = to_omap_bo(obj);
894 mutex_lock(&obj->dev->struct_mutex);
895 if (omap_obj->paddr_cnt > 0) {
896 omap_obj->paddr_cnt--;
897 if (omap_obj->paddr_cnt == 0) {
898 ret = tiler_unpin(omap_obj->block);
900 dev_err(obj->dev->dev,
901 "could not unpin pages: %d\n", ret);
903 ret = tiler_release(omap_obj->block);
905 dev_err(obj->dev->dev,
906 "could not release unmap: %d\n", ret);
909 omap_obj->block = NULL;
913 mutex_unlock(&obj->dev->struct_mutex);
916 /* Get rotated scanout address (only valid if already pinned), at the
917 * specified orientation and x,y offset from top-left corner of buffer
918 * (only valid for tiled 2d buffers)
920 int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient,
921 int x, int y, dma_addr_t *paddr)
923 struct omap_gem_object *omap_obj = to_omap_bo(obj);
926 mutex_lock(&obj->dev->struct_mutex);
927 if ((omap_obj->paddr_cnt > 0) && omap_obj->block &&
928 (omap_obj->flags & OMAP_BO_TILED)) {
929 *paddr = tiler_tsptr(omap_obj->block, orient, x, y);
932 mutex_unlock(&obj->dev->struct_mutex);
936 /* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
937 int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient)
939 struct omap_gem_object *omap_obj = to_omap_bo(obj);
941 if (omap_obj->flags & OMAP_BO_TILED)
942 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
946 /* if !remap, and we don't have pages backing, then fail, rather than
947 * increasing the pin count (which we don't really do yet anyways,
948 * because we don't support swapping pages back out). And 'remap'
949 * might not be quite the right name, but I wanted to keep it working
950 * similarly to omap_gem_get_paddr(). Note though that mutex is not
951 * aquired if !remap (because this can be called in atomic ctxt),
952 * but probably omap_gem_get_paddr() should be changed to work in the
953 * same way. If !remap, a matching omap_gem_put_pages() call is not
954 * required (and should not be made).
956 int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
961 struct omap_gem_object *omap_obj = to_omap_bo(obj);
962 if (!omap_obj->pages)
964 *pages = omap_obj->pages;
967 mutex_lock(&obj->dev->struct_mutex);
968 ret = get_pages(obj, pages);
969 mutex_unlock(&obj->dev->struct_mutex);
973 /* release pages when DMA no longer being performed */
974 int omap_gem_put_pages(struct drm_gem_object *obj)
976 /* do something here if we dynamically attach/detach pages.. at
977 * least they would no longer need to be pinned if everyone has
978 * released the pages..
983 #ifdef CONFIG_DRM_FBDEV_EMULATION
984 /* Get kernel virtual address for CPU access.. this more or less only
985 * exists for omap_fbdev. This should be called with struct_mutex
988 void *omap_gem_vaddr(struct drm_gem_object *obj)
990 struct omap_gem_object *omap_obj = to_omap_bo(obj);
991 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
992 if (!omap_obj->vaddr) {
994 int ret = get_pages(obj, &pages);
997 omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
998 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
1000 return omap_obj->vaddr;
1004 /* -----------------------------------------------------------------------------
1009 /* re-pin objects in DMM in resume path: */
1010 int omap_gem_resume(struct device *dev)
1012 struct drm_device *drm_dev = dev_get_drvdata(dev);
1013 struct omap_drm_private *priv = drm_dev->dev_private;
1014 struct omap_gem_object *omap_obj;
1017 list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
1018 if (omap_obj->block) {
1019 struct drm_gem_object *obj = &omap_obj->base;
1020 uint32_t npages = obj->size >> PAGE_SHIFT;
1021 WARN_ON(!omap_obj->pages); /* this can't happen */
1022 ret = tiler_pin(omap_obj->block,
1023 omap_obj->pages, npages,
1024 omap_obj->roll, true);
1026 dev_err(dev, "could not repin: %d\n", ret);
1036 /* -----------------------------------------------------------------------------
1040 #ifdef CONFIG_DEBUG_FS
1041 void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1043 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1046 off = drm_vma_node_start(&obj->vma_node);
1048 seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1049 omap_obj->flags, obj->name, obj->refcount.refcount.counter,
1050 off, &omap_obj->paddr, omap_obj->paddr_cnt,
1051 omap_obj->vaddr, omap_obj->roll);
1053 if (omap_obj->flags & OMAP_BO_TILED) {
1054 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1055 if (omap_obj->block) {
1056 struct tcm_area *area = &omap_obj->block->area;
1057 seq_printf(m, " (%dx%d, %dx%d)",
1058 area->p0.x, area->p0.y,
1059 area->p1.x, area->p1.y);
1062 seq_printf(m, " %d", obj->size);
1065 seq_printf(m, "\n");
1068 void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1070 struct omap_gem_object *omap_obj;
1074 list_for_each_entry(omap_obj, list, mm_list) {
1075 struct drm_gem_object *obj = &omap_obj->base;
1077 omap_gem_describe(obj, m);
1082 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1086 /* -----------------------------------------------------------------------------
1087 * Buffer Synchronization
1090 static DEFINE_SPINLOCK(sync_lock);
1092 struct omap_gem_sync_waiter {
1093 struct list_head list;
1094 struct omap_gem_object *omap_obj;
1095 enum omap_gem_op op;
1096 uint32_t read_target, write_target;
1097 /* notify called w/ sync_lock held */
1098 void (*notify)(void *arg);
1102 /* list of omap_gem_sync_waiter.. the notify fxn gets called back when
1103 * the read and/or write target count is achieved which can call a user
1104 * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for
1107 static LIST_HEAD(waiters);
1109 static inline bool is_waiting(struct omap_gem_sync_waiter *waiter)
1111 struct omap_gem_object *omap_obj = waiter->omap_obj;
1112 if ((waiter->op & OMAP_GEM_READ) &&
1113 (omap_obj->sync->write_complete < waiter->write_target))
1115 if ((waiter->op & OMAP_GEM_WRITE) &&
1116 (omap_obj->sync->read_complete < waiter->read_target))
1121 /* macro for sync debug.. */
1123 #define SYNC(fmt, ...) do { if (SYNCDBG) \
1124 printk(KERN_ERR "%s:%d: "fmt"\n", \
1125 __func__, __LINE__, ##__VA_ARGS__); \
1129 static void sync_op_update(void)
1131 struct omap_gem_sync_waiter *waiter, *n;
1132 list_for_each_entry_safe(waiter, n, &waiters, list) {
1133 if (!is_waiting(waiter)) {
1134 list_del(&waiter->list);
1135 SYNC("notify: %p", waiter);
1136 waiter->notify(waiter->arg);
1142 static inline int sync_op(struct drm_gem_object *obj,
1143 enum omap_gem_op op, bool start)
1145 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1148 spin_lock(&sync_lock);
1150 if (!omap_obj->sync) {
1151 omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC);
1152 if (!omap_obj->sync) {
1159 if (op & OMAP_GEM_READ)
1160 omap_obj->sync->read_pending++;
1161 if (op & OMAP_GEM_WRITE)
1162 omap_obj->sync->write_pending++;
1164 if (op & OMAP_GEM_READ)
1165 omap_obj->sync->read_complete++;
1166 if (op & OMAP_GEM_WRITE)
1167 omap_obj->sync->write_complete++;
1172 spin_unlock(&sync_lock);
1177 /* mark the start of read and/or write operation */
1178 int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op)
1180 return sync_op(obj, op, true);
1183 int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op)
1185 return sync_op(obj, op, false);
1188 static DECLARE_WAIT_QUEUE_HEAD(sync_event);
1190 static void sync_notify(void *arg)
1192 struct task_struct **waiter_task = arg;
1193 *waiter_task = NULL;
1194 wake_up_all(&sync_event);
1197 int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op)
1199 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1201 if (omap_obj->sync) {
1202 struct task_struct *waiter_task = current;
1203 struct omap_gem_sync_waiter *waiter =
1204 kzalloc(sizeof(*waiter), GFP_KERNEL);
1209 waiter->omap_obj = omap_obj;
1211 waiter->read_target = omap_obj->sync->read_pending;
1212 waiter->write_target = omap_obj->sync->write_pending;
1213 waiter->notify = sync_notify;
1214 waiter->arg = &waiter_task;
1216 spin_lock(&sync_lock);
1217 if (is_waiting(waiter)) {
1218 SYNC("waited: %p", waiter);
1219 list_add_tail(&waiter->list, &waiters);
1220 spin_unlock(&sync_lock);
1221 ret = wait_event_interruptible(sync_event,
1222 (waiter_task == NULL));
1223 spin_lock(&sync_lock);
1225 SYNC("interrupted: %p", waiter);
1226 /* we were interrupted */
1227 list_del(&waiter->list);
1230 /* freed in sync_op_update() */
1234 spin_unlock(&sync_lock);
1240 /* call fxn(arg), either synchronously or asynchronously if the op
1241 * is currently blocked.. fxn() can be called from any context
1243 * (TODO for now fxn is called back from whichever context calls
1244 * omap_gem_op_finish().. but this could be better defined later
1247 * TODO more code in common w/ _sync()..
1249 int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op,
1250 void (*fxn)(void *arg), void *arg)
1252 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1253 if (omap_obj->sync) {
1254 struct omap_gem_sync_waiter *waiter =
1255 kzalloc(sizeof(*waiter), GFP_ATOMIC);
1260 waiter->omap_obj = omap_obj;
1262 waiter->read_target = omap_obj->sync->read_pending;
1263 waiter->write_target = omap_obj->sync->write_pending;
1264 waiter->notify = fxn;
1267 spin_lock(&sync_lock);
1268 if (is_waiting(waiter)) {
1269 SYNC("waited: %p", waiter);
1270 list_add_tail(&waiter->list, &waiters);
1271 spin_unlock(&sync_lock);
1275 spin_unlock(&sync_lock);
1286 /* -----------------------------------------------------------------------------
1287 * Constructor & Destructor
1290 void omap_gem_free_object(struct drm_gem_object *obj)
1292 struct drm_device *dev = obj->dev;
1293 struct omap_drm_private *priv = dev->dev_private;
1294 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1298 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1300 spin_lock(&priv->list_lock);
1301 list_del(&omap_obj->mm_list);
1302 spin_unlock(&priv->list_lock);
1304 /* this means the object is still pinned.. which really should
1305 * not happen. I think..
1307 WARN_ON(omap_obj->paddr_cnt > 0);
1309 if (omap_obj->pages) {
1310 if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1311 kfree(omap_obj->pages);
1313 omap_gem_detach_pages(obj);
1316 if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1317 dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1319 } else if (omap_obj->vaddr) {
1320 vunmap(omap_obj->vaddr);
1321 } else if (obj->import_attach) {
1322 drm_prime_gem_destroy(obj, omap_obj->sgt);
1325 kfree(omap_obj->sync);
1327 drm_gem_object_release(obj);
1332 /* GEM buffer object constructor */
1333 struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1334 union omap_gem_size gsize, uint32_t flags)
1336 struct omap_drm_private *priv = dev->dev_private;
1337 struct omap_gem_object *omap_obj;
1338 struct drm_gem_object *obj;
1339 struct address_space *mapping;
1343 /* Validate the flags and compute the memory and cache flags. */
1344 if (flags & OMAP_BO_TILED) {
1345 if (!priv->usergart) {
1346 dev_err(dev->dev, "Tiled buffers require DMM\n");
1351 * Tiled buffers are always shmem paged backed. When they are
1352 * scanned out, they are remapped into DMM/TILER.
1354 flags &= ~OMAP_BO_SCANOUT;
1355 flags |= OMAP_BO_MEM_SHMEM;
1358 * Currently don't allow cached buffers. There is some caching
1359 * stuff that needs to be handled better.
1361 flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1362 flags |= tiler_get_cpu_cache_flags();
1363 } else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1365 * OMAP_BO_SCANOUT hints that the buffer doesn't need to be
1366 * tiled. However, to lower the pressure on memory allocation,
1367 * use contiguous memory only if no TILER is available.
1369 flags |= OMAP_BO_MEM_DMA_API;
1370 } else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1372 * All other buffers not backed by dma_buf are shmem-backed.
1374 flags |= OMAP_BO_MEM_SHMEM;
1377 /* Allocate the initialize the OMAP GEM object. */
1378 omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1382 obj = &omap_obj->base;
1383 omap_obj->flags = flags;
1385 if (flags & OMAP_BO_TILED) {
1387 * For tiled buffers align dimensions to slot boundaries and
1388 * calculate size based on aligned dimensions.
1390 tiler_align(gem2fmt(flags), &gsize.tiled.width,
1391 &gsize.tiled.height);
1393 size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1394 gsize.tiled.height);
1396 omap_obj->width = gsize.tiled.width;
1397 omap_obj->height = gsize.tiled.height;
1399 size = PAGE_ALIGN(gsize.bytes);
1402 /* Initialize the GEM object. */
1403 if (!(flags & OMAP_BO_MEM_SHMEM)) {
1404 drm_gem_private_object_init(dev, obj, size);
1406 ret = drm_gem_object_init(dev, obj, size);
1410 mapping = file_inode(obj->filp)->i_mapping;
1411 mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1414 /* Allocate memory if needed. */
1415 if (flags & OMAP_BO_MEM_DMA_API) {
1416 omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1419 if (!omap_obj->vaddr)
1423 spin_lock(&priv->list_lock);
1424 list_add(&omap_obj->mm_list, &priv->obj_list);
1425 spin_unlock(&priv->list_lock);
1430 drm_gem_object_release(obj);
1436 struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1437 struct sg_table *sgt)
1439 struct omap_drm_private *priv = dev->dev_private;
1440 struct omap_gem_object *omap_obj;
1441 struct drm_gem_object *obj;
1442 union omap_gem_size gsize;
1444 /* Without a DMM only physically contiguous buffers can be supported. */
1445 if (sgt->orig_nents != 1 && !priv->has_dmm)
1446 return ERR_PTR(-EINVAL);
1448 mutex_lock(&dev->struct_mutex);
1450 gsize.bytes = PAGE_ALIGN(size);
1451 obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1453 obj = ERR_PTR(-ENOMEM);
1457 omap_obj = to_omap_bo(obj);
1458 omap_obj->sgt = sgt;
1460 if (sgt->orig_nents == 1) {
1461 omap_obj->paddr = sg_dma_address(sgt->sgl);
1463 /* Create pages list from sgt */
1464 struct sg_page_iter iter;
1465 struct page **pages;
1466 unsigned int npages;
1469 npages = DIV_ROUND_UP(size, PAGE_SIZE);
1470 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1472 omap_gem_free_object(obj);
1473 obj = ERR_PTR(-ENOMEM);
1477 omap_obj->pages = pages;
1479 for_each_sg_page(sgt->sgl, &iter, sgt->orig_nents, 0) {
1480 pages[i++] = sg_page_iter_page(&iter);
1485 if (WARN_ON(i != npages)) {
1486 omap_gem_free_object(obj);
1487 obj = ERR_PTR(-ENOMEM);
1493 mutex_unlock(&dev->struct_mutex);
1497 /* convenience method to construct a GEM buffer object, and userspace handle */
1498 int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1499 union omap_gem_size gsize, uint32_t flags, uint32_t *handle)
1501 struct drm_gem_object *obj;
1504 obj = omap_gem_new(dev, gsize, flags);
1508 ret = drm_gem_handle_create(file, obj, handle);
1510 omap_gem_free_object(obj);
1514 /* drop reference from allocate - handle holds it now */
1515 drm_gem_object_unreference_unlocked(obj);
1520 /* -----------------------------------------------------------------------------
1524 /* If DMM is used, we need to set some stuff up.. */
1525 void omap_gem_init(struct drm_device *dev)
1527 struct omap_drm_private *priv = dev->dev_private;
1528 struct omap_drm_usergart *usergart;
1529 const enum tiler_fmt fmts[] = {
1530 TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1534 if (!dmm_is_available()) {
1535 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1536 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1540 usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1544 /* reserve 4k aligned/wide regions for userspace mappings: */
1545 for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1546 uint16_t h = 1, w = PAGE_SIZE >> i;
1547 tiler_align(fmts[i], &w, &h);
1548 /* note: since each region is 1 4kb page wide, and minimum
1549 * number of rows, the height ends up being the same as the
1550 * # of pages in the region
1552 usergart[i].height = h;
1553 usergart[i].height_shift = ilog2(h);
1554 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1555 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1556 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1557 struct omap_drm_usergart_entry *entry;
1558 struct tiler_block *block;
1560 entry = &usergart[i].entry[j];
1561 block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1562 if (IS_ERR(block)) {
1564 "reserve failed: %d, %d, %ld\n",
1565 i, j, PTR_ERR(block));
1568 entry->paddr = tiler_ssptr(block);
1569 entry->block = block;
1571 DBG("%d:%d: %dx%d: paddr=%pad stride=%d", i, j, w, h,
1573 usergart[i].stride_pfn << PAGE_SHIFT);
1577 priv->usergart = usergart;
1578 priv->has_dmm = true;
1581 void omap_gem_deinit(struct drm_device *dev)
1583 struct omap_drm_private *priv = dev->dev_private;
1585 /* I believe we can rely on there being no more outstanding GEM
1586 * objects which could depend on usergart/dmm at this point.
1588 kfree(priv->usergart);