1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2011 Texas Instruments Incorporated - http://www.ti.com/
7 #include <linux/seq_file.h>
8 #include <linux/shmem_fs.h>
9 #include <linux/spinlock.h>
10 #include <linux/pfn_t.h>
12 #include <drm/drm_vma_manager.h>
15 #include "omap_dmm_tiler.h"
18 * GEM buffer object implementation.
21 /* note: we use upper 8 bits of flags for driver-internal flags: */
22 #define OMAP_BO_MEM_DMA_API 0x01000000 /* memory allocated with the dma_alloc_* API */
23 #define OMAP_BO_MEM_SHMEM 0x02000000 /* memory allocated through shmem backing */
24 #define OMAP_BO_MEM_DMABUF 0x08000000 /* memory imported from a dmabuf */
26 struct omap_gem_object {
27 struct drm_gem_object base;
29 struct list_head mm_list;
33 /** width/height for tiled formats (rounded up to slot boundaries) */
36 /** roll applied when mapping to DMM */
39 /** protects dma_addr_cnt, block, pages, dma_addrs and vaddr */
43 * dma_addr contains the buffer DMA address. It is valid for
45 * - buffers allocated through the DMA mapping API (with the
46 * OMAP_BO_MEM_DMA_API flag set)
48 * - buffers imported from dmabuf (with the OMAP_BO_MEM_DMABUF flag set)
49 * if they are physically contiguous (when sgt->orig_nents == 1)
51 * - buffers mapped through the TILER when dma_addr_cnt is not zero, in
52 * which case the DMA address points to the TILER aperture
54 * Physically contiguous buffers have their DMA address equal to the
55 * physical address as we don't remap those buffers through the TILER.
57 * Buffers mapped to the TILER have their DMA address pointing to the
58 * TILER aperture. As TILER mappings are refcounted (through
59 * dma_addr_cnt) the DMA address must be accessed through omap_gem_pin()
60 * to ensure that the mapping won't disappear unexpectedly. References
61 * must be released with omap_gem_unpin().
66 * # of users of dma_addr
71 * If the buffer has been imported from a dmabuf the OMAP_DB_DMABUF flag
72 * is set and the sgt field is valid.
77 * tiler block used when buffer is remapped in DMM/TILER.
79 struct tiler_block *block;
82 * Array of backing pages, if allocated. Note that pages are never
83 * allocated for buffers originally allocated from contiguous memory
87 /** addresses corresponding to pages in above array */
88 dma_addr_t *dma_addrs;
91 * Virtual address, if mapped.
96 #define to_omap_bo(x) container_of(x, struct omap_gem_object, base)
98 /* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are
99 * not necessarily pinned in TILER all the time, and (b) when they are
100 * they are not necessarily page aligned, we reserve one or more small
101 * regions in each of the 2d containers to use as a user-GART where we
102 * can create a second page-aligned mapping of parts of the buffer
103 * being accessed from userspace.
105 * Note that we could optimize slightly when we know that multiple
106 * tiler containers are backed by the same PAT.. but I'll leave that
109 #define NUM_USERGART_ENTRIES 2
110 struct omap_drm_usergart_entry {
111 struct tiler_block *block; /* the reserved tiler block */
113 struct drm_gem_object *obj; /* the current pinned obj */
114 pgoff_t obj_pgoff; /* page offset of obj currently
118 struct omap_drm_usergart {
119 struct omap_drm_usergart_entry entry[NUM_USERGART_ENTRIES];
120 int height; /* height in rows */
121 int height_shift; /* ilog2(height in rows) */
122 int slot_shift; /* ilog2(width per slot) */
123 int stride_pfn; /* stride in pages */
124 int last; /* index of last used entry */
127 /* -----------------------------------------------------------------------------
131 /** get mmap offset */
132 u64 omap_gem_mmap_offset(struct drm_gem_object *obj)
134 struct drm_device *dev = obj->dev;
138 /* Make it mmapable */
139 size = omap_gem_mmap_size(obj);
140 ret = drm_gem_create_mmap_offset_size(obj, size);
142 dev_err(dev->dev, "could not allocate mmap offset\n");
146 return drm_vma_node_offset_addr(&obj->vma_node);
149 static bool omap_gem_is_contiguous(struct omap_gem_object *omap_obj)
151 if (omap_obj->flags & OMAP_BO_MEM_DMA_API)
154 if ((omap_obj->flags & OMAP_BO_MEM_DMABUF) && omap_obj->sgt->nents == 1)
160 /* -----------------------------------------------------------------------------
164 static void omap_gem_evict_entry(struct drm_gem_object *obj,
165 enum tiler_fmt fmt, struct omap_drm_usergart_entry *entry)
167 struct omap_gem_object *omap_obj = to_omap_bo(obj);
168 struct omap_drm_private *priv = obj->dev->dev_private;
169 int n = priv->usergart[fmt].height;
170 size_t size = PAGE_SIZE * n;
171 loff_t off = omap_gem_mmap_offset(obj) +
172 (entry->obj_pgoff << PAGE_SHIFT);
173 const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
177 /* if stride > than PAGE_SIZE then sparse mapping: */
178 for (i = n; i > 0; i--) {
179 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
181 off += PAGE_SIZE * m;
184 unmap_mapping_range(obj->dev->anon_inode->i_mapping,
191 /* Evict a buffer from usergart, if it is mapped there */
192 static void omap_gem_evict(struct drm_gem_object *obj)
194 struct omap_gem_object *omap_obj = to_omap_bo(obj);
195 struct omap_drm_private *priv = obj->dev->dev_private;
197 if (omap_obj->flags & OMAP_BO_TILED) {
198 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
201 for (i = 0; i < NUM_USERGART_ENTRIES; i++) {
202 struct omap_drm_usergart_entry *entry =
203 &priv->usergart[fmt].entry[i];
205 if (entry->obj == obj)
206 omap_gem_evict_entry(obj, fmt, entry);
211 /* -----------------------------------------------------------------------------
216 * Ensure backing pages are allocated. Must be called with the omap_obj.lock
219 static int omap_gem_attach_pages(struct drm_gem_object *obj)
221 struct drm_device *dev = obj->dev;
222 struct omap_gem_object *omap_obj = to_omap_bo(obj);
224 int npages = obj->size >> PAGE_SHIFT;
228 lockdep_assert_held(&omap_obj->lock);
231 * If not using shmem (in which case backing pages don't need to be
232 * allocated) or if pages are already allocated we're done.
234 if (!(omap_obj->flags & OMAP_BO_MEM_SHMEM) || omap_obj->pages)
237 pages = drm_gem_get_pages(obj);
239 dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages));
240 return PTR_ERR(pages);
243 /* for non-cached buffers, ensure the new pages are clean because
244 * DSS, GPU, etc. are not cache coherent:
246 if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) {
247 addrs = kmalloc_array(npages, sizeof(*addrs), GFP_KERNEL);
253 for (i = 0; i < npages; i++) {
254 addrs[i] = dma_map_page(dev->dev, pages[i],
255 0, PAGE_SIZE, DMA_TO_DEVICE);
257 if (dma_mapping_error(dev->dev, addrs[i])) {
259 "%s: failed to map page\n", __func__);
261 for (i = i - 1; i >= 0; --i) {
262 dma_unmap_page(dev->dev, addrs[i],
263 PAGE_SIZE, DMA_TO_DEVICE);
271 addrs = kcalloc(npages, sizeof(*addrs), GFP_KERNEL);
278 omap_obj->dma_addrs = addrs;
279 omap_obj->pages = pages;
286 drm_gem_put_pages(obj, pages, true, false);
291 /* Release backing pages. Must be called with the omap_obj.lock held. */
292 static void omap_gem_detach_pages(struct drm_gem_object *obj)
294 struct omap_gem_object *omap_obj = to_omap_bo(obj);
295 unsigned int npages = obj->size >> PAGE_SHIFT;
298 lockdep_assert_held(&omap_obj->lock);
300 for (i = 0; i < npages; i++) {
301 if (omap_obj->dma_addrs[i])
302 dma_unmap_page(obj->dev->dev, omap_obj->dma_addrs[i],
303 PAGE_SIZE, DMA_TO_DEVICE);
306 kfree(omap_obj->dma_addrs);
307 omap_obj->dma_addrs = NULL;
309 drm_gem_put_pages(obj, omap_obj->pages, true, false);
310 omap_obj->pages = NULL;
313 /* get buffer flags */
314 u32 omap_gem_flags(struct drm_gem_object *obj)
316 return to_omap_bo(obj)->flags;
320 size_t omap_gem_mmap_size(struct drm_gem_object *obj)
322 struct omap_gem_object *omap_obj = to_omap_bo(obj);
323 size_t size = obj->size;
325 if (omap_obj->flags & OMAP_BO_TILED) {
326 /* for tiled buffers, the virtual size has stride rounded up
327 * to 4kb.. (to hide the fact that row n+1 might start 16kb or
328 * 32kb later!). But we don't back the entire buffer with
329 * pages, only the valid picture part.. so need to adjust for
330 * this in the size used to mmap and generate mmap offset
332 size = tiler_vsize(gem2fmt(omap_obj->flags),
333 omap_obj->width, omap_obj->height);
339 /* -----------------------------------------------------------------------------
343 /* Normal handling for the case of faulting in non-tiled buffers */
344 static vm_fault_t omap_gem_fault_1d(struct drm_gem_object *obj,
345 struct vm_area_struct *vma, struct vm_fault *vmf)
347 struct omap_gem_object *omap_obj = to_omap_bo(obj);
351 /* We don't use vmf->pgoff since that has the fake offset: */
352 pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
354 if (omap_obj->pages) {
355 omap_gem_cpu_sync_page(obj, pgoff);
356 pfn = page_to_pfn(omap_obj->pages[pgoff]);
358 BUG_ON(!omap_gem_is_contiguous(omap_obj));
359 pfn = (omap_obj->dma_addr >> PAGE_SHIFT) + pgoff;
362 VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
363 pfn, pfn << PAGE_SHIFT);
365 return vmf_insert_mixed(vma, vmf->address,
366 __pfn_to_pfn_t(pfn, PFN_DEV));
369 /* Special handling for the case of faulting in 2d tiled buffers */
370 static vm_fault_t omap_gem_fault_2d(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);
374 struct omap_drm_private *priv = obj->dev->dev_private;
375 struct omap_drm_usergart_entry *entry;
376 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
377 struct page *pages[64]; /* XXX is this too much to have on stack? */
379 pgoff_t pgoff, base_pgoff;
382 vm_fault_t ret = VM_FAULT_NOPAGE;
385 * Note the height of the slot is also equal to the number of pages
386 * that need to be mapped in to fill 4kb wide CPU page. If the slot
387 * height is 64, then 64 pages fill a 4kb wide by 64 row region.
389 const int n = priv->usergart[fmt].height;
390 const int n_shift = priv->usergart[fmt].height_shift;
393 * If buffer width in bytes > PAGE_SIZE then the virtual stride is
394 * rounded up to next multiple of PAGE_SIZE.. this need to be taken
395 * into account in some of the math, so figure out virtual stride
398 const int m = DIV_ROUND_UP(omap_obj->width << fmt, PAGE_SIZE);
400 /* We don't use vmf->pgoff since that has the fake offset: */
401 pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
404 * Actual address we start mapping at is rounded down to previous slot
405 * boundary in the y direction:
407 base_pgoff = round_down(pgoff, m << n_shift);
409 /* figure out buffer width in slots */
410 slots = omap_obj->width >> priv->usergart[fmt].slot_shift;
412 vaddr = vmf->address - ((pgoff - base_pgoff) << PAGE_SHIFT);
414 entry = &priv->usergart[fmt].entry[priv->usergart[fmt].last];
416 /* evict previous buffer using this usergart entry, if any: */
418 omap_gem_evict_entry(entry->obj, fmt, entry);
421 entry->obj_pgoff = base_pgoff;
423 /* now convert base_pgoff to phys offset from virt offset: */
424 base_pgoff = (base_pgoff >> n_shift) * slots;
426 /* for wider-than 4k.. figure out which part of the slot-row we want: */
429 entry->obj_pgoff += off;
431 slots = min(slots - (off << n_shift), n);
432 base_pgoff += off << n_shift;
433 vaddr += off << PAGE_SHIFT;
437 * Map in pages. Beyond the valid pixel part of the buffer, we set
438 * pages[i] to NULL to get a dummy page mapped in.. if someone
439 * reads/writes it they will get random/undefined content, but at
440 * least it won't be corrupting whatever other random page used to
441 * be mapped in, or other undefined behavior.
443 memcpy(pages, &omap_obj->pages[base_pgoff],
444 sizeof(struct page *) * slots);
445 memset(pages + slots, 0,
446 sizeof(struct page *) * (n - slots));
448 err = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true);
450 ret = vmf_error(err);
451 dev_err(obj->dev->dev, "failed to pin: %d\n", err);
455 pfn = entry->dma_addr >> PAGE_SHIFT;
457 VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
458 pfn, pfn << PAGE_SHIFT);
460 for (i = n; i > 0; i--) {
461 ret = vmf_insert_mixed(vma,
462 vaddr, __pfn_to_pfn_t(pfn, PFN_DEV));
463 if (ret & VM_FAULT_ERROR)
465 pfn += priv->usergart[fmt].stride_pfn;
466 vaddr += PAGE_SIZE * m;
469 /* simple round-robin: */
470 priv->usergart[fmt].last = (priv->usergart[fmt].last + 1)
471 % NUM_USERGART_ENTRIES;
477 * omap_gem_fault - pagefault handler for GEM objects
480 * Invoked when a fault occurs on an mmap of a GEM managed area. GEM
481 * does most of the work for us including the actual map/unmap calls
482 * but we need to do the actual page work.
484 * The VMA was set up by GEM. In doing so it also ensured that the
485 * vma->vm_private_data points to the GEM object that is backing this
488 vm_fault_t omap_gem_fault(struct vm_fault *vmf)
490 struct vm_area_struct *vma = vmf->vma;
491 struct drm_gem_object *obj = vma->vm_private_data;
492 struct omap_gem_object *omap_obj = to_omap_bo(obj);
496 /* Make sure we don't parallel update on a fault, nor move or remove
497 * something from beneath our feet
499 mutex_lock(&omap_obj->lock);
501 /* if a shmem backed object, make sure we have pages attached now */
502 err = omap_gem_attach_pages(obj);
504 ret = vmf_error(err);
508 /* where should we do corresponding put_pages().. we are mapping
509 * the original page, rather than thru a GART, so we can't rely
510 * on eviction to trigger this. But munmap() or all mappings should
511 * probably trigger put_pages()?
514 if (omap_obj->flags & OMAP_BO_TILED)
515 ret = omap_gem_fault_2d(obj, vma, vmf);
517 ret = omap_gem_fault_1d(obj, vma, vmf);
521 mutex_unlock(&omap_obj->lock);
525 /** We override mainly to fix up some of the vm mapping flags.. */
526 int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma)
530 ret = drm_gem_mmap(filp, vma);
532 DBG("mmap failed: %d", ret);
536 return omap_gem_mmap_obj(vma->vm_private_data, vma);
539 int omap_gem_mmap_obj(struct drm_gem_object *obj,
540 struct vm_area_struct *vma)
542 struct omap_gem_object *omap_obj = to_omap_bo(obj);
544 vma->vm_flags &= ~VM_PFNMAP;
545 vma->vm_flags |= VM_MIXEDMAP;
547 if (omap_obj->flags & OMAP_BO_WC) {
548 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
549 } else if (omap_obj->flags & OMAP_BO_UNCACHED) {
550 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
553 * We do have some private objects, at least for scanout buffers
554 * on hardware without DMM/TILER. But these are allocated write-
557 if (WARN_ON(!obj->filp))
561 * Shunt off cached objs to shmem file so they have their own
562 * address_space (so unmap_mapping_range does what we want,
563 * in particular in the case of mmap'd dmabufs)
567 vma->vm_file = get_file(obj->filp);
569 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
575 /* -----------------------------------------------------------------------------
580 * omap_gem_dumb_create - create a dumb buffer
581 * @drm_file: our client file
583 * @args: the requested arguments copied from userspace
585 * Allocate a buffer suitable for use for a frame buffer of the
586 * form described by user space. Give userspace a handle by which
589 int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
590 struct drm_mode_create_dumb *args)
592 union omap_gem_size gsize;
594 args->pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
596 args->size = PAGE_ALIGN(args->pitch * args->height);
598 gsize = (union omap_gem_size){
602 return omap_gem_new_handle(dev, file, gsize,
603 OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle);
607 * omap_gem_dumb_map - buffer mapping for dumb interface
608 * @file: our drm client file
610 * @handle: GEM handle to the object (from dumb_create)
612 * Do the necessary setup to allow the mapping of the frame buffer
613 * into user memory. We don't have to do much here at the moment.
615 int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
616 u32 handle, u64 *offset)
618 struct drm_gem_object *obj;
621 /* GEM does all our handle to object mapping */
622 obj = drm_gem_object_lookup(file, handle);
628 *offset = omap_gem_mmap_offset(obj);
630 drm_gem_object_put_unlocked(obj);
636 #ifdef CONFIG_DRM_FBDEV_EMULATION
637 /* Set scrolling position. This allows us to implement fast scrolling
640 * Call only from non-atomic contexts.
642 int omap_gem_roll(struct drm_gem_object *obj, u32 roll)
644 struct omap_gem_object *omap_obj = to_omap_bo(obj);
645 u32 npages = obj->size >> PAGE_SHIFT;
649 dev_err(obj->dev->dev, "invalid roll: %d\n", roll);
653 omap_obj->roll = roll;
655 mutex_lock(&omap_obj->lock);
657 /* if we aren't mapped yet, we don't need to do anything */
658 if (omap_obj->block) {
659 ret = omap_gem_attach_pages(obj);
663 ret = tiler_pin(omap_obj->block, omap_obj->pages, npages,
666 dev_err(obj->dev->dev, "could not repin: %d\n", ret);
670 mutex_unlock(&omap_obj->lock);
676 /* -----------------------------------------------------------------------------
677 * Memory Management & DMA Sync
681 * shmem buffers that are mapped cached are not coherent.
683 * We keep track of dirty pages using page faulting to perform cache management.
684 * When a page is mapped to the CPU in read/write mode the device can't access
685 * it and omap_obj->dma_addrs[i] is NULL. When a page is mapped to the device
686 * the omap_obj->dma_addrs[i] is set to the DMA address, and the page is
687 * unmapped from the CPU.
689 static inline bool omap_gem_is_cached_coherent(struct drm_gem_object *obj)
691 struct omap_gem_object *omap_obj = to_omap_bo(obj);
693 return !((omap_obj->flags & OMAP_BO_MEM_SHMEM) &&
694 ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED));
697 /* Sync the buffer for CPU access.. note pages should already be
698 * attached, ie. omap_gem_get_pages()
700 void omap_gem_cpu_sync_page(struct drm_gem_object *obj, int pgoff)
702 struct drm_device *dev = obj->dev;
703 struct omap_gem_object *omap_obj = to_omap_bo(obj);
705 if (omap_gem_is_cached_coherent(obj))
708 if (omap_obj->dma_addrs[pgoff]) {
709 dma_unmap_page(dev->dev, omap_obj->dma_addrs[pgoff],
710 PAGE_SIZE, DMA_TO_DEVICE);
711 omap_obj->dma_addrs[pgoff] = 0;
715 /* sync the buffer for DMA access */
716 void omap_gem_dma_sync_buffer(struct drm_gem_object *obj,
717 enum dma_data_direction dir)
719 struct drm_device *dev = obj->dev;
720 struct omap_gem_object *omap_obj = to_omap_bo(obj);
721 int i, npages = obj->size >> PAGE_SHIFT;
722 struct page **pages = omap_obj->pages;
725 if (omap_gem_is_cached_coherent(obj))
728 for (i = 0; i < npages; i++) {
729 if (!omap_obj->dma_addrs[i]) {
732 addr = dma_map_page(dev->dev, pages[i], 0,
734 if (dma_mapping_error(dev->dev, addr)) {
735 dev_warn(dev->dev, "%s: failed to map page\n",
741 omap_obj->dma_addrs[i] = addr;
746 unmap_mapping_range(obj->filp->f_mapping, 0,
747 omap_gem_mmap_size(obj), 1);
752 * omap_gem_pin() - Pin a GEM object in memory
753 * @obj: the GEM object
754 * @dma_addr: the DMA address
756 * Pin the given GEM object in memory and fill the dma_addr pointer with the
757 * object's DMA address. If the buffer is not physically contiguous it will be
758 * remapped through the TILER to provide a contiguous view.
760 * Pins are reference-counted, calling this function multiple times is allowed
761 * as long the corresponding omap_gem_unpin() calls are balanced.
763 * Return 0 on success or a negative error code otherwise.
765 int omap_gem_pin(struct drm_gem_object *obj, dma_addr_t *dma_addr)
767 struct omap_drm_private *priv = obj->dev->dev_private;
768 struct omap_gem_object *omap_obj = to_omap_bo(obj);
771 mutex_lock(&omap_obj->lock);
773 if (!omap_gem_is_contiguous(omap_obj) && priv->has_dmm) {
774 if (omap_obj->dma_addr_cnt == 0) {
775 u32 npages = obj->size >> PAGE_SHIFT;
776 enum tiler_fmt fmt = gem2fmt(omap_obj->flags);
777 struct tiler_block *block;
779 BUG_ON(omap_obj->block);
781 ret = omap_gem_attach_pages(obj);
785 if (omap_obj->flags & OMAP_BO_TILED) {
786 block = tiler_reserve_2d(fmt,
788 omap_obj->height, 0);
790 block = tiler_reserve_1d(obj->size);
794 ret = PTR_ERR(block);
795 dev_err(obj->dev->dev,
796 "could not remap: %d (%d)\n", ret, fmt);
800 /* TODO: enable async refill.. */
801 ret = tiler_pin(block, omap_obj->pages, npages,
802 omap_obj->roll, true);
804 tiler_release(block);
805 dev_err(obj->dev->dev,
806 "could not pin: %d\n", ret);
810 omap_obj->dma_addr = tiler_ssptr(block);
811 omap_obj->block = block;
813 DBG("got dma address: %pad", &omap_obj->dma_addr);
816 omap_obj->dma_addr_cnt++;
818 *dma_addr = omap_obj->dma_addr;
819 } else if (omap_gem_is_contiguous(omap_obj)) {
820 *dma_addr = omap_obj->dma_addr;
827 mutex_unlock(&omap_obj->lock);
833 * omap_gem_unpin() - Unpin a GEM object from memory
834 * @obj: the GEM object
836 * Unpin the given GEM object previously pinned with omap_gem_pin(). Pins are
837 * reference-counted, the actualy unpin will only be performed when the number
838 * of calls to this function matches the number of calls to omap_gem_pin().
840 void omap_gem_unpin(struct drm_gem_object *obj)
842 struct omap_gem_object *omap_obj = to_omap_bo(obj);
845 mutex_lock(&omap_obj->lock);
847 if (omap_obj->dma_addr_cnt > 0) {
848 omap_obj->dma_addr_cnt--;
849 if (omap_obj->dma_addr_cnt == 0) {
850 ret = tiler_unpin(omap_obj->block);
852 dev_err(obj->dev->dev,
853 "could not unpin pages: %d\n", ret);
855 ret = tiler_release(omap_obj->block);
857 dev_err(obj->dev->dev,
858 "could not release unmap: %d\n", ret);
860 omap_obj->dma_addr = 0;
861 omap_obj->block = NULL;
865 mutex_unlock(&omap_obj->lock);
868 /* Get rotated scanout address (only valid if already pinned), at the
869 * specified orientation and x,y offset from top-left corner of buffer
870 * (only valid for tiled 2d buffers)
872 int omap_gem_rotated_dma_addr(struct drm_gem_object *obj, u32 orient,
873 int x, int y, dma_addr_t *dma_addr)
875 struct omap_gem_object *omap_obj = to_omap_bo(obj);
878 mutex_lock(&omap_obj->lock);
880 if ((omap_obj->dma_addr_cnt > 0) && omap_obj->block &&
881 (omap_obj->flags & OMAP_BO_TILED)) {
882 *dma_addr = tiler_tsptr(omap_obj->block, orient, x, y);
886 mutex_unlock(&omap_obj->lock);
891 /* Get tiler stride for the buffer (only valid for 2d tiled buffers) */
892 int omap_gem_tiled_stride(struct drm_gem_object *obj, u32 orient)
894 struct omap_gem_object *omap_obj = to_omap_bo(obj);
896 if (omap_obj->flags & OMAP_BO_TILED)
897 ret = tiler_stride(gem2fmt(omap_obj->flags), orient);
901 /* if !remap, and we don't have pages backing, then fail, rather than
902 * increasing the pin count (which we don't really do yet anyways,
903 * because we don't support swapping pages back out). And 'remap'
904 * might not be quite the right name, but I wanted to keep it working
905 * similarly to omap_gem_pin(). Note though that mutex is not
906 * aquired if !remap (because this can be called in atomic ctxt),
907 * but probably omap_gem_unpin() should be changed to work in the
908 * same way. If !remap, a matching omap_gem_put_pages() call is not
909 * required (and should not be made).
911 int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages,
914 struct omap_gem_object *omap_obj = to_omap_bo(obj);
917 mutex_lock(&omap_obj->lock);
920 ret = omap_gem_attach_pages(obj);
925 if (!omap_obj->pages) {
930 *pages = omap_obj->pages;
933 mutex_unlock(&omap_obj->lock);
938 /* release pages when DMA no longer being performed */
939 int omap_gem_put_pages(struct drm_gem_object *obj)
941 /* do something here if we dynamically attach/detach pages.. at
942 * least they would no longer need to be pinned if everyone has
943 * released the pages..
948 #ifdef CONFIG_DRM_FBDEV_EMULATION
950 * Get kernel virtual address for CPU access.. this more or less only
951 * exists for omap_fbdev.
953 void *omap_gem_vaddr(struct drm_gem_object *obj)
955 struct omap_gem_object *omap_obj = to_omap_bo(obj);
959 mutex_lock(&omap_obj->lock);
961 if (!omap_obj->vaddr) {
962 ret = omap_gem_attach_pages(obj);
964 vaddr = ERR_PTR(ret);
968 omap_obj->vaddr = vmap(omap_obj->pages, obj->size >> PAGE_SHIFT,
969 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
972 vaddr = omap_obj->vaddr;
975 mutex_unlock(&omap_obj->lock);
980 /* -----------------------------------------------------------------------------
985 /* re-pin objects in DMM in resume path: */
986 int omap_gem_resume(struct drm_device *dev)
988 struct omap_drm_private *priv = dev->dev_private;
989 struct omap_gem_object *omap_obj;
992 mutex_lock(&priv->list_lock);
993 list_for_each_entry(omap_obj, &priv->obj_list, mm_list) {
994 if (omap_obj->block) {
995 struct drm_gem_object *obj = &omap_obj->base;
996 u32 npages = obj->size >> PAGE_SHIFT;
998 WARN_ON(!omap_obj->pages); /* this can't happen */
999 ret = tiler_pin(omap_obj->block,
1000 omap_obj->pages, npages,
1001 omap_obj->roll, true);
1003 dev_err(dev->dev, "could not repin: %d\n", ret);
1010 mutex_unlock(&priv->list_lock);
1015 /* -----------------------------------------------------------------------------
1019 #ifdef CONFIG_DEBUG_FS
1020 void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
1022 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1025 off = drm_vma_node_start(&obj->vma_node);
1027 mutex_lock(&omap_obj->lock);
1029 seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d",
1030 omap_obj->flags, obj->name, kref_read(&obj->refcount),
1031 off, &omap_obj->dma_addr, omap_obj->dma_addr_cnt,
1032 omap_obj->vaddr, omap_obj->roll);
1034 if (omap_obj->flags & OMAP_BO_TILED) {
1035 seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height);
1036 if (omap_obj->block) {
1037 struct tcm_area *area = &omap_obj->block->area;
1038 seq_printf(m, " (%dx%d, %dx%d)",
1039 area->p0.x, area->p0.y,
1040 area->p1.x, area->p1.y);
1043 seq_printf(m, " %zu", obj->size);
1046 mutex_unlock(&omap_obj->lock);
1048 seq_printf(m, "\n");
1051 void omap_gem_describe_objects(struct list_head *list, struct seq_file *m)
1053 struct omap_gem_object *omap_obj;
1057 list_for_each_entry(omap_obj, list, mm_list) {
1058 struct drm_gem_object *obj = &omap_obj->base;
1060 omap_gem_describe(obj, m);
1065 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
1069 /* -----------------------------------------------------------------------------
1070 * Constructor & Destructor
1073 void omap_gem_free_object(struct drm_gem_object *obj)
1075 struct drm_device *dev = obj->dev;
1076 struct omap_drm_private *priv = dev->dev_private;
1077 struct omap_gem_object *omap_obj = to_omap_bo(obj);
1079 omap_gem_evict(obj);
1081 mutex_lock(&priv->list_lock);
1082 list_del(&omap_obj->mm_list);
1083 mutex_unlock(&priv->list_lock);
1086 * We own the sole reference to the object at this point, but to keep
1087 * lockdep happy, we must still take the omap_obj_lock to call
1088 * omap_gem_detach_pages(). This should hardly make any difference as
1089 * there can't be any lock contention.
1091 mutex_lock(&omap_obj->lock);
1093 /* The object should not be pinned. */
1094 WARN_ON(omap_obj->dma_addr_cnt > 0);
1096 if (omap_obj->pages) {
1097 if (omap_obj->flags & OMAP_BO_MEM_DMABUF)
1098 kfree(omap_obj->pages);
1100 omap_gem_detach_pages(obj);
1103 if (omap_obj->flags & OMAP_BO_MEM_DMA_API) {
1104 dma_free_wc(dev->dev, obj->size, omap_obj->vaddr,
1105 omap_obj->dma_addr);
1106 } else if (omap_obj->vaddr) {
1107 vunmap(omap_obj->vaddr);
1108 } else if (obj->import_attach) {
1109 drm_prime_gem_destroy(obj, omap_obj->sgt);
1112 mutex_unlock(&omap_obj->lock);
1114 drm_gem_object_release(obj);
1116 mutex_destroy(&omap_obj->lock);
1121 /* GEM buffer object constructor */
1122 struct drm_gem_object *omap_gem_new(struct drm_device *dev,
1123 union omap_gem_size gsize, u32 flags)
1125 struct omap_drm_private *priv = dev->dev_private;
1126 struct omap_gem_object *omap_obj;
1127 struct drm_gem_object *obj;
1128 struct address_space *mapping;
1132 /* Validate the flags and compute the memory and cache flags. */
1133 if (flags & OMAP_BO_TILED) {
1134 if (!priv->usergart) {
1135 dev_err(dev->dev, "Tiled buffers require DMM\n");
1140 * Tiled buffers are always shmem paged backed. When they are
1141 * scanned out, they are remapped into DMM/TILER.
1143 flags &= ~OMAP_BO_SCANOUT;
1144 flags |= OMAP_BO_MEM_SHMEM;
1147 * Currently don't allow cached buffers. There is some caching
1148 * stuff that needs to be handled better.
1150 flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
1151 flags |= tiler_get_cpu_cache_flags();
1152 } else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
1154 * OMAP_BO_SCANOUT hints that the buffer doesn't need to be
1155 * tiled. However, to lower the pressure on memory allocation,
1156 * use contiguous memory only if no TILER is available.
1158 flags |= OMAP_BO_MEM_DMA_API;
1159 } else if (!(flags & OMAP_BO_MEM_DMABUF)) {
1161 * All other buffers not backed by dma_buf are shmem-backed.
1163 flags |= OMAP_BO_MEM_SHMEM;
1166 /* Allocate the initialize the OMAP GEM object. */
1167 omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL);
1171 obj = &omap_obj->base;
1172 omap_obj->flags = flags;
1173 mutex_init(&omap_obj->lock);
1175 if (flags & OMAP_BO_TILED) {
1177 * For tiled buffers align dimensions to slot boundaries and
1178 * calculate size based on aligned dimensions.
1180 tiler_align(gem2fmt(flags), &gsize.tiled.width,
1181 &gsize.tiled.height);
1183 size = tiler_size(gem2fmt(flags), gsize.tiled.width,
1184 gsize.tiled.height);
1186 omap_obj->width = gsize.tiled.width;
1187 omap_obj->height = gsize.tiled.height;
1189 size = PAGE_ALIGN(gsize.bytes);
1192 /* Initialize the GEM object. */
1193 if (!(flags & OMAP_BO_MEM_SHMEM)) {
1194 drm_gem_private_object_init(dev, obj, size);
1196 ret = drm_gem_object_init(dev, obj, size);
1200 mapping = obj->filp->f_mapping;
1201 mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32);
1204 /* Allocate memory if needed. */
1205 if (flags & OMAP_BO_MEM_DMA_API) {
1206 omap_obj->vaddr = dma_alloc_wc(dev->dev, size,
1207 &omap_obj->dma_addr,
1209 if (!omap_obj->vaddr)
1213 mutex_lock(&priv->list_lock);
1214 list_add(&omap_obj->mm_list, &priv->obj_list);
1215 mutex_unlock(&priv->list_lock);
1220 drm_gem_object_release(obj);
1226 struct drm_gem_object *omap_gem_new_dmabuf(struct drm_device *dev, size_t size,
1227 struct sg_table *sgt)
1229 struct omap_drm_private *priv = dev->dev_private;
1230 struct omap_gem_object *omap_obj;
1231 struct drm_gem_object *obj;
1232 union omap_gem_size gsize;
1234 /* Without a DMM only physically contiguous buffers can be supported. */
1235 if (sgt->orig_nents != 1 && !priv->has_dmm)
1236 return ERR_PTR(-EINVAL);
1238 gsize.bytes = PAGE_ALIGN(size);
1239 obj = omap_gem_new(dev, gsize, OMAP_BO_MEM_DMABUF | OMAP_BO_WC);
1241 return ERR_PTR(-ENOMEM);
1243 omap_obj = to_omap_bo(obj);
1245 mutex_lock(&omap_obj->lock);
1247 omap_obj->sgt = sgt;
1249 if (sgt->orig_nents == 1) {
1250 omap_obj->dma_addr = sg_dma_address(sgt->sgl);
1252 /* Create pages list from sgt */
1253 struct sg_page_iter iter;
1254 struct page **pages;
1255 unsigned int npages;
1258 npages = DIV_ROUND_UP(size, PAGE_SIZE);
1259 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1261 omap_gem_free_object(obj);
1262 obj = ERR_PTR(-ENOMEM);
1266 omap_obj->pages = pages;
1268 for_each_sg_page(sgt->sgl, &iter, sgt->orig_nents, 0) {
1269 pages[i++] = sg_page_iter_page(&iter);
1274 if (WARN_ON(i != npages)) {
1275 omap_gem_free_object(obj);
1276 obj = ERR_PTR(-ENOMEM);
1282 mutex_unlock(&omap_obj->lock);
1286 /* convenience method to construct a GEM buffer object, and userspace handle */
1287 int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1288 union omap_gem_size gsize, u32 flags, u32 *handle)
1290 struct drm_gem_object *obj;
1293 obj = omap_gem_new(dev, gsize, flags);
1297 ret = drm_gem_handle_create(file, obj, handle);
1299 omap_gem_free_object(obj);
1303 /* drop reference from allocate - handle holds it now */
1304 drm_gem_object_put_unlocked(obj);
1309 /* -----------------------------------------------------------------------------
1313 /* If DMM is used, we need to set some stuff up.. */
1314 void omap_gem_init(struct drm_device *dev)
1316 struct omap_drm_private *priv = dev->dev_private;
1317 struct omap_drm_usergart *usergart;
1318 const enum tiler_fmt fmts[] = {
1319 TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT
1323 if (!dmm_is_available()) {
1324 /* DMM only supported on OMAP4 and later, so this isn't fatal */
1325 dev_warn(dev->dev, "DMM not available, disable DMM support\n");
1329 usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL);
1333 /* reserve 4k aligned/wide regions for userspace mappings: */
1334 for (i = 0; i < ARRAY_SIZE(fmts); i++) {
1335 u16 h = 1, w = PAGE_SIZE >> i;
1337 tiler_align(fmts[i], &w, &h);
1338 /* note: since each region is 1 4kb page wide, and minimum
1339 * number of rows, the height ends up being the same as the
1340 * # of pages in the region
1342 usergart[i].height = h;
1343 usergart[i].height_shift = ilog2(h);
1344 usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT;
1345 usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i);
1346 for (j = 0; j < NUM_USERGART_ENTRIES; j++) {
1347 struct omap_drm_usergart_entry *entry;
1348 struct tiler_block *block;
1350 entry = &usergart[i].entry[j];
1351 block = tiler_reserve_2d(fmts[i], w, h, PAGE_SIZE);
1352 if (IS_ERR(block)) {
1354 "reserve failed: %d, %d, %ld\n",
1355 i, j, PTR_ERR(block));
1358 entry->dma_addr = tiler_ssptr(block);
1359 entry->block = block;
1361 DBG("%d:%d: %dx%d: dma_addr=%pad stride=%d", i, j, w, h,
1363 usergart[i].stride_pfn << PAGE_SHIFT);
1367 priv->usergart = usergart;
1368 priv->has_dmm = true;
1371 void omap_gem_deinit(struct drm_device *dev)
1373 struct omap_drm_private *priv = dev->dev_private;
1375 /* I believe we can rely on there being no more outstanding GEM
1376 * objects which could depend on usergart/dmm at this point.
1378 kfree(priv->usergart);