2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published by
7 * the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * You should have received a copy of the GNU General Public License along with
15 * this program. If not, see <http://www.gnu.org/licenses/>.
18 #include <linux/spinlock.h>
19 #include <linux/shmem_fs.h>
20 #include <linux/dma-buf.h>
27 static dma_addr_t physaddr(struct drm_gem_object *obj)
29 struct msm_gem_object *msm_obj = to_msm_bo(obj);
30 struct msm_drm_private *priv = obj->dev->dev_private;
31 return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) +
35 static bool use_pages(struct drm_gem_object *obj)
37 struct msm_gem_object *msm_obj = to_msm_bo(obj);
38 return !msm_obj->vram_node;
41 /* allocate pages from VRAM carveout, used when no IOMMU: */
42 static struct page **get_pages_vram(struct drm_gem_object *obj,
45 struct msm_gem_object *msm_obj = to_msm_bo(obj);
46 struct msm_drm_private *priv = obj->dev->dev_private;
51 p = drm_malloc_ab(npages, sizeof(struct page *));
53 return ERR_PTR(-ENOMEM);
55 ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node,
56 npages, 0, DRM_MM_SEARCH_DEFAULT);
62 paddr = physaddr(obj);
63 for (i = 0; i < npages; i++) {
64 p[i] = phys_to_page(paddr);
71 /* called with dev->struct_mutex held */
72 static struct page **get_pages(struct drm_gem_object *obj)
74 struct msm_gem_object *msm_obj = to_msm_bo(obj);
76 if (!msm_obj->pages) {
77 struct drm_device *dev = obj->dev;
79 int npages = obj->size >> PAGE_SHIFT;
82 p = drm_gem_get_pages(obj);
84 p = get_pages_vram(obj, npages);
87 dev_err(dev->dev, "could not get pages: %ld\n",
94 msm_obj->sgt = drm_prime_pages_to_sg(p, npages);
95 if (IS_ERR(msm_obj->sgt)) {
96 void *ptr = ERR_CAST(msm_obj->sgt);
98 dev_err(dev->dev, "failed to allocate sgt\n");
103 /* For non-cached buffers, ensure the new pages are clean
104 * because display controller, GPU, etc. are not coherent:
106 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
107 dma_map_sg(dev->dev, msm_obj->sgt->sgl,
108 msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
111 return msm_obj->pages;
114 static void put_pages(struct drm_gem_object *obj)
116 struct msm_gem_object *msm_obj = to_msm_bo(obj);
118 if (msm_obj->pages) {
120 /* For non-cached buffers, ensure the new
121 * pages are clean because display controller,
122 * GPU, etc. are not coherent:
124 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
125 dma_unmap_sg(obj->dev->dev, msm_obj->sgt->sgl,
129 sg_free_table(msm_obj->sgt);
134 drm_gem_put_pages(obj, msm_obj->pages, true, false);
136 drm_mm_remove_node(msm_obj->vram_node);
137 drm_free_large(msm_obj->pages);
140 msm_obj->pages = NULL;
144 struct page **msm_gem_get_pages(struct drm_gem_object *obj)
146 struct drm_device *dev = obj->dev;
148 mutex_lock(&dev->struct_mutex);
150 mutex_unlock(&dev->struct_mutex);
154 void msm_gem_put_pages(struct drm_gem_object *obj)
156 /* when we start tracking the pin count, then do something here */
159 int msm_gem_mmap_obj(struct drm_gem_object *obj,
160 struct vm_area_struct *vma)
162 struct msm_gem_object *msm_obj = to_msm_bo(obj);
164 vma->vm_flags &= ~VM_PFNMAP;
165 vma->vm_flags |= VM_MIXEDMAP;
167 if (msm_obj->flags & MSM_BO_WC) {
168 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
169 } else if (msm_obj->flags & MSM_BO_UNCACHED) {
170 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
173 * Shunt off cached objs to shmem file so they have their own
174 * address_space (so unmap_mapping_range does what we want,
175 * in particular in the case of mmap'd dmabufs)
180 vma->vm_file = obj->filp;
182 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
188 int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
192 ret = drm_gem_mmap(filp, vma);
194 DBG("mmap failed: %d", ret);
198 return msm_gem_mmap_obj(vma->vm_private_data, vma);
201 int msm_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
203 struct drm_gem_object *obj = vma->vm_private_data;
204 struct drm_device *dev = obj->dev;
210 /* Make sure we don't parallel update on a fault, nor move or remove
211 * something from beneath our feet
213 ret = mutex_lock_interruptible(&dev->struct_mutex);
217 /* make sure we have pages attached now */
218 pages = get_pages(obj);
220 ret = PTR_ERR(pages);
224 /* We don't use vmf->pgoff since that has the fake offset: */
225 pgoff = ((unsigned long)vmf->virtual_address -
226 vma->vm_start) >> PAGE_SHIFT;
228 pfn = page_to_pfn(pages[pgoff]);
230 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
231 pfn, pfn << PAGE_SHIFT);
233 ret = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, pfn);
236 mutex_unlock(&dev->struct_mutex);
245 * EBUSY is ok: this just means that another thread
246 * already did the job.
248 return VM_FAULT_NOPAGE;
252 return VM_FAULT_SIGBUS;
256 /** get mmap offset */
257 static uint64_t mmap_offset(struct drm_gem_object *obj)
259 struct drm_device *dev = obj->dev;
262 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
264 /* Make it mmapable */
265 ret = drm_gem_create_mmap_offset(obj);
268 dev_err(dev->dev, "could not allocate mmap offset\n");
272 return drm_vma_node_offset_addr(&obj->vma_node);
275 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
278 mutex_lock(&obj->dev->struct_mutex);
279 offset = mmap_offset(obj);
280 mutex_unlock(&obj->dev->struct_mutex);
284 /* should be called under struct_mutex.. although it can be called
285 * from atomic context without struct_mutex to acquire an extra
286 * iova ref if you know one is already held.
288 * That means when I do eventually need to add support for unpinning
289 * the refcnt counter needs to be atomic_t.
291 int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id,
294 struct msm_gem_object *msm_obj = to_msm_bo(obj);
297 if (!msm_obj->domain[id].iova) {
298 struct msm_drm_private *priv = obj->dev->dev_private;
299 struct page **pages = get_pages(obj);
302 return PTR_ERR(pages);
304 if (iommu_present(&platform_bus_type)) {
305 struct msm_mmu *mmu = priv->mmus[id];
311 offset = (uint32_t)mmap_offset(obj);
312 ret = mmu->funcs->map(mmu, offset, msm_obj->sgt,
313 obj->size, IOMMU_READ | IOMMU_WRITE);
314 msm_obj->domain[id].iova = offset;
316 msm_obj->domain[id].iova = physaddr(obj);
321 *iova = msm_obj->domain[id].iova;
326 /* get iova, taking a reference. Should have a matching put */
327 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova)
329 struct msm_gem_object *msm_obj = to_msm_bo(obj);
332 /* this is safe right now because we don't unmap until the
335 if (msm_obj->domain[id].iova) {
336 *iova = msm_obj->domain[id].iova;
340 mutex_lock(&obj->dev->struct_mutex);
341 ret = msm_gem_get_iova_locked(obj, id, iova);
342 mutex_unlock(&obj->dev->struct_mutex);
346 /* get iova without taking a reference, used in places where you have
347 * already done a 'msm_gem_get_iova()'.
349 uint32_t msm_gem_iova(struct drm_gem_object *obj, int id)
351 struct msm_gem_object *msm_obj = to_msm_bo(obj);
352 WARN_ON(!msm_obj->domain[id].iova);
353 return msm_obj->domain[id].iova;
356 void msm_gem_put_iova(struct drm_gem_object *obj, int id)
359 // NOTE: probably don't need a _locked() version.. we wouldn't
360 // normally unmap here, but instead just mark that it could be
361 // unmapped (if the iova refcnt drops to zero), but then later
362 // if another _get_iova_locked() fails we can start unmapping
363 // things that are no longer needed..
366 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
367 struct drm_mode_create_dumb *args)
369 args->pitch = align_pitch(args->width, args->bpp);
370 args->size = PAGE_ALIGN(args->pitch * args->height);
371 return msm_gem_new_handle(dev, file, args->size,
372 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle);
375 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
376 uint32_t handle, uint64_t *offset)
378 struct drm_gem_object *obj;
381 /* GEM does all our handle to object mapping */
382 obj = drm_gem_object_lookup(dev, file, handle);
388 *offset = msm_gem_mmap_offset(obj);
390 drm_gem_object_unreference_unlocked(obj);
396 void *msm_gem_vaddr_locked(struct drm_gem_object *obj)
398 struct msm_gem_object *msm_obj = to_msm_bo(obj);
399 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
400 if (!msm_obj->vaddr) {
401 struct page **pages = get_pages(obj);
403 return ERR_CAST(pages);
404 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
405 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
407 return msm_obj->vaddr;
410 void *msm_gem_vaddr(struct drm_gem_object *obj)
413 mutex_lock(&obj->dev->struct_mutex);
414 ret = msm_gem_vaddr_locked(obj);
415 mutex_unlock(&obj->dev->struct_mutex);
419 /* setup callback for when bo is no longer busy..
420 * TODO probably want to differentiate read vs write..
422 int msm_gem_queue_inactive_cb(struct drm_gem_object *obj,
423 struct msm_fence_cb *cb)
425 struct msm_gem_object *msm_obj = to_msm_bo(obj);
426 uint32_t fence = msm_gem_fence(msm_obj,
427 MSM_PREP_READ | MSM_PREP_WRITE);
428 return msm_queue_fence_cb(obj->dev, cb, fence);
431 void msm_gem_move_to_active(struct drm_gem_object *obj,
432 struct msm_gpu *gpu, bool write, uint32_t fence)
434 struct msm_gem_object *msm_obj = to_msm_bo(obj);
437 msm_obj->write_fence = fence;
439 msm_obj->read_fence = fence;
440 list_del_init(&msm_obj->mm_list);
441 list_add_tail(&msm_obj->mm_list, &gpu->active_list);
444 void msm_gem_move_to_inactive(struct drm_gem_object *obj)
446 struct drm_device *dev = obj->dev;
447 struct msm_drm_private *priv = dev->dev_private;
448 struct msm_gem_object *msm_obj = to_msm_bo(obj);
450 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
453 msm_obj->read_fence = 0;
454 msm_obj->write_fence = 0;
455 list_del_init(&msm_obj->mm_list);
456 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
459 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout)
461 struct drm_device *dev = obj->dev;
462 struct msm_gem_object *msm_obj = to_msm_bo(obj);
465 if (is_active(msm_obj)) {
466 uint32_t fence = msm_gem_fence(msm_obj, op);
468 if (op & MSM_PREP_NOSYNC)
471 ret = msm_wait_fence(dev, fence, timeout, true);
474 /* TODO cache maintenance */
479 int msm_gem_cpu_fini(struct drm_gem_object *obj)
481 /* TODO cache maintenance */
485 #ifdef CONFIG_DEBUG_FS
486 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
488 struct drm_device *dev = obj->dev;
489 struct msm_gem_object *msm_obj = to_msm_bo(obj);
490 uint64_t off = drm_vma_node_start(&obj->vma_node);
492 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
493 seq_printf(m, "%08x: %c(r=%u,w=%u) %2d (%2d) %08llx %p %zu\n",
494 msm_obj->flags, is_active(msm_obj) ? 'A' : 'I',
495 msm_obj->read_fence, msm_obj->write_fence,
496 obj->name, obj->refcount.refcount.counter,
497 off, msm_obj->vaddr, obj->size);
500 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
502 struct msm_gem_object *msm_obj;
506 list_for_each_entry(msm_obj, list, mm_list) {
507 struct drm_gem_object *obj = &msm_obj->base;
509 msm_gem_describe(obj, m);
514 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
518 void msm_gem_free_object(struct drm_gem_object *obj)
520 struct drm_device *dev = obj->dev;
521 struct msm_drm_private *priv = obj->dev->dev_private;
522 struct msm_gem_object *msm_obj = to_msm_bo(obj);
525 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
527 /* object should not be on active list: */
528 WARN_ON(is_active(msm_obj));
530 list_del(&msm_obj->mm_list);
532 for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) {
533 struct msm_mmu *mmu = priv->mmus[id];
534 if (mmu && msm_obj->domain[id].iova) {
535 uint32_t offset = msm_obj->domain[id].iova;
536 mmu->funcs->unmap(mmu, offset, msm_obj->sgt, obj->size);
540 if (obj->import_attach) {
542 dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr);
544 /* Don't drop the pages for imported dmabuf, as they are not
545 * ours, just free the array we allocated:
548 drm_free_large(msm_obj->pages);
550 drm_prime_gem_destroy(obj, msm_obj->sgt);
552 vunmap(msm_obj->vaddr);
556 if (msm_obj->resv == &msm_obj->_resv)
557 reservation_object_fini(msm_obj->resv);
559 drm_gem_object_release(obj);
564 /* convenience method to construct a GEM buffer object, and userspace handle */
565 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
566 uint32_t size, uint32_t flags, uint32_t *handle)
568 struct drm_gem_object *obj;
571 ret = mutex_lock_interruptible(&dev->struct_mutex);
575 obj = msm_gem_new(dev, size, flags);
577 mutex_unlock(&dev->struct_mutex);
582 ret = drm_gem_handle_create(file, obj, handle);
584 /* drop reference from allocate - handle holds it now */
585 drm_gem_object_unreference_unlocked(obj);
590 static int msm_gem_new_impl(struct drm_device *dev,
591 uint32_t size, uint32_t flags,
592 struct drm_gem_object **obj)
594 struct msm_drm_private *priv = dev->dev_private;
595 struct msm_gem_object *msm_obj;
597 bool use_vram = false;
599 switch (flags & MSM_BO_CACHE_MASK) {
600 case MSM_BO_UNCACHED:
605 dev_err(dev->dev, "invalid cache flag: %x\n",
606 (flags & MSM_BO_CACHE_MASK));
610 if (!iommu_present(&platform_bus_type))
612 else if ((flags & MSM_BO_STOLEN) && priv->vram.size)
615 if (WARN_ON(use_vram && !priv->vram.size))
618 sz = sizeof(*msm_obj);
620 sz += sizeof(struct drm_mm_node);
622 msm_obj = kzalloc(sz, GFP_KERNEL);
627 msm_obj->vram_node = (void *)&msm_obj[1];
629 msm_obj->flags = flags;
631 msm_obj->resv = &msm_obj->_resv;
632 reservation_object_init(msm_obj->resv);
634 INIT_LIST_HEAD(&msm_obj->submit_entry);
635 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
637 *obj = &msm_obj->base;
642 struct drm_gem_object *msm_gem_new(struct drm_device *dev,
643 uint32_t size, uint32_t flags)
645 struct drm_gem_object *obj = NULL;
648 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
650 size = PAGE_ALIGN(size);
652 ret = msm_gem_new_impl(dev, size, flags, &obj);
656 if (use_pages(obj)) {
657 ret = drm_gem_object_init(dev, obj, size);
661 drm_gem_private_object_init(dev, obj, size);
668 drm_gem_object_unreference(obj);
673 struct drm_gem_object *msm_gem_import(struct drm_device *dev,
674 uint32_t size, struct sg_table *sgt)
676 struct msm_gem_object *msm_obj;
677 struct drm_gem_object *obj;
680 /* if we don't have IOMMU, don't bother pretending we can import: */
681 if (!iommu_present(&platform_bus_type)) {
682 dev_err(dev->dev, "cannot import without IOMMU\n");
683 return ERR_PTR(-EINVAL);
686 size = PAGE_ALIGN(size);
688 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, &obj);
692 drm_gem_private_object_init(dev, obj, size);
694 npages = size / PAGE_SIZE;
696 msm_obj = to_msm_bo(obj);
698 msm_obj->pages = drm_malloc_ab(npages, sizeof(struct page *));
699 if (!msm_obj->pages) {
704 ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages);
712 drm_gem_object_unreference_unlocked(obj);