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>
21 #include <linux/pfn_t.h>
24 #include "msm_fence.h"
29 static dma_addr_t physaddr(struct drm_gem_object *obj)
31 struct msm_gem_object *msm_obj = to_msm_bo(obj);
32 struct msm_drm_private *priv = obj->dev->dev_private;
33 return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) +
37 static bool use_pages(struct drm_gem_object *obj)
39 struct msm_gem_object *msm_obj = to_msm_bo(obj);
40 return !msm_obj->vram_node;
43 /* allocate pages from VRAM carveout, used when no IOMMU: */
44 static struct page **get_pages_vram(struct drm_gem_object *obj,
47 struct msm_gem_object *msm_obj = to_msm_bo(obj);
48 struct msm_drm_private *priv = obj->dev->dev_private;
53 p = drm_malloc_ab(npages, sizeof(struct page *));
55 return ERR_PTR(-ENOMEM);
57 ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node,
58 npages, 0, DRM_MM_SEARCH_DEFAULT);
64 paddr = physaddr(obj);
65 for (i = 0; i < npages; i++) {
66 p[i] = phys_to_page(paddr);
73 /* called with dev->struct_mutex held */
74 static struct page **get_pages(struct drm_gem_object *obj)
76 struct msm_gem_object *msm_obj = to_msm_bo(obj);
78 if (!msm_obj->pages) {
79 struct drm_device *dev = obj->dev;
81 int npages = obj->size >> PAGE_SHIFT;
84 p = drm_gem_get_pages(obj);
86 p = get_pages_vram(obj, npages);
89 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 dev_err(dev->dev, "failed to allocate sgt\n");
97 return ERR_CAST(msm_obj->sgt);
102 /* For non-cached buffers, ensure the new pages are clean
103 * because display controller, GPU, etc. are not coherent:
105 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
106 dma_map_sg(dev->dev, msm_obj->sgt->sgl,
107 msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
110 return msm_obj->pages;
113 static void put_pages(struct drm_gem_object *obj)
115 struct msm_gem_object *msm_obj = to_msm_bo(obj);
117 if (msm_obj->pages) {
118 /* For non-cached buffers, ensure the new pages are clean
119 * because display controller, GPU, etc. are not coherent:
121 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
122 dma_unmap_sg(obj->dev->dev, msm_obj->sgt->sgl,
123 msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
124 sg_free_table(msm_obj->sgt);
128 drm_gem_put_pages(obj, msm_obj->pages, true, false);
130 drm_mm_remove_node(msm_obj->vram_node);
131 drm_free_large(msm_obj->pages);
134 msm_obj->pages = NULL;
138 struct page **msm_gem_get_pages(struct drm_gem_object *obj)
140 struct drm_device *dev = obj->dev;
142 mutex_lock(&dev->struct_mutex);
144 mutex_unlock(&dev->struct_mutex);
148 void msm_gem_put_pages(struct drm_gem_object *obj)
150 /* when we start tracking the pin count, then do something here */
153 int msm_gem_mmap_obj(struct drm_gem_object *obj,
154 struct vm_area_struct *vma)
156 struct msm_gem_object *msm_obj = to_msm_bo(obj);
158 vma->vm_flags &= ~VM_PFNMAP;
159 vma->vm_flags |= VM_MIXEDMAP;
161 if (msm_obj->flags & MSM_BO_WC) {
162 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
163 } else if (msm_obj->flags & MSM_BO_UNCACHED) {
164 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
167 * Shunt off cached objs to shmem file so they have their own
168 * address_space (so unmap_mapping_range does what we want,
169 * in particular in the case of mmap'd dmabufs)
174 vma->vm_file = obj->filp;
176 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
182 int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
186 ret = drm_gem_mmap(filp, vma);
188 DBG("mmap failed: %d", ret);
192 return msm_gem_mmap_obj(vma->vm_private_data, vma);
195 int msm_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
197 struct drm_gem_object *obj = vma->vm_private_data;
198 struct drm_device *dev = obj->dev;
204 /* Make sure we don't parallel update on a fault, nor move or remove
205 * something from beneath our feet
207 ret = mutex_lock_interruptible(&dev->struct_mutex);
211 /* make sure we have pages attached now */
212 pages = get_pages(obj);
214 ret = PTR_ERR(pages);
218 /* We don't use vmf->pgoff since that has the fake offset: */
219 pgoff = ((unsigned long)vmf->virtual_address -
220 vma->vm_start) >> PAGE_SHIFT;
222 pfn = page_to_pfn(pages[pgoff]);
224 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
225 pfn, pfn << PAGE_SHIFT);
227 ret = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address,
228 __pfn_to_pfn_t(pfn, PFN_DEV));
231 mutex_unlock(&dev->struct_mutex);
240 * EBUSY is ok: this just means that another thread
241 * already did the job.
243 return VM_FAULT_NOPAGE;
247 return VM_FAULT_SIGBUS;
251 /** get mmap offset */
252 static uint64_t mmap_offset(struct drm_gem_object *obj)
254 struct drm_device *dev = obj->dev;
257 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
259 /* Make it mmapable */
260 ret = drm_gem_create_mmap_offset(obj);
263 dev_err(dev->dev, "could not allocate mmap offset\n");
267 return drm_vma_node_offset_addr(&obj->vma_node);
270 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
273 mutex_lock(&obj->dev->struct_mutex);
274 offset = mmap_offset(obj);
275 mutex_unlock(&obj->dev->struct_mutex);
280 put_iova(struct drm_gem_object *obj)
282 struct drm_device *dev = obj->dev;
283 struct msm_drm_private *priv = obj->dev->dev_private;
284 struct msm_gem_object *msm_obj = to_msm_bo(obj);
287 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
289 for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) {
290 struct msm_mmu *mmu = priv->mmus[id];
291 if (mmu && msm_obj->domain[id].iova) {
292 uint32_t offset = msm_obj->domain[id].iova;
293 mmu->funcs->unmap(mmu, offset, msm_obj->sgt, obj->size);
294 msm_obj->domain[id].iova = 0;
299 /* should be called under struct_mutex.. although it can be called
300 * from atomic context without struct_mutex to acquire an extra
301 * iova ref if you know one is already held.
303 * That means when I do eventually need to add support for unpinning
304 * the refcnt counter needs to be atomic_t.
306 int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id,
309 struct msm_gem_object *msm_obj = to_msm_bo(obj);
312 if (!msm_obj->domain[id].iova) {
313 struct msm_drm_private *priv = obj->dev->dev_private;
314 struct page **pages = get_pages(obj);
317 return PTR_ERR(pages);
319 if (iommu_present(&platform_bus_type)) {
320 struct msm_mmu *mmu = priv->mmus[id];
326 offset = (uint32_t)mmap_offset(obj);
327 ret = mmu->funcs->map(mmu, offset, msm_obj->sgt,
328 obj->size, IOMMU_READ | IOMMU_WRITE);
329 msm_obj->domain[id].iova = offset;
331 msm_obj->domain[id].iova = physaddr(obj);
336 *iova = msm_obj->domain[id].iova;
341 /* get iova, taking a reference. Should have a matching put */
342 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova)
344 struct msm_gem_object *msm_obj = to_msm_bo(obj);
347 /* this is safe right now because we don't unmap until the
350 if (msm_obj->domain[id].iova) {
351 *iova = msm_obj->domain[id].iova;
355 mutex_lock(&obj->dev->struct_mutex);
356 ret = msm_gem_get_iova_locked(obj, id, iova);
357 mutex_unlock(&obj->dev->struct_mutex);
361 /* get iova without taking a reference, used in places where you have
362 * already done a 'msm_gem_get_iova()'.
364 uint32_t msm_gem_iova(struct drm_gem_object *obj, int id)
366 struct msm_gem_object *msm_obj = to_msm_bo(obj);
367 WARN_ON(!msm_obj->domain[id].iova);
368 return msm_obj->domain[id].iova;
371 void msm_gem_put_iova(struct drm_gem_object *obj, int id)
374 // NOTE: probably don't need a _locked() version.. we wouldn't
375 // normally unmap here, but instead just mark that it could be
376 // unmapped (if the iova refcnt drops to zero), but then later
377 // if another _get_iova_locked() fails we can start unmapping
378 // things that are no longer needed..
381 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
382 struct drm_mode_create_dumb *args)
384 args->pitch = align_pitch(args->width, args->bpp);
385 args->size = PAGE_ALIGN(args->pitch * args->height);
386 return msm_gem_new_handle(dev, file, args->size,
387 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle);
390 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
391 uint32_t handle, uint64_t *offset)
393 struct drm_gem_object *obj;
396 /* GEM does all our handle to object mapping */
397 obj = drm_gem_object_lookup(file, handle);
403 *offset = msm_gem_mmap_offset(obj);
405 drm_gem_object_unreference_unlocked(obj);
411 void *msm_gem_get_vaddr_locked(struct drm_gem_object *obj)
413 struct msm_gem_object *msm_obj = to_msm_bo(obj);
414 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
415 if (!msm_obj->vaddr) {
416 struct page **pages = get_pages(obj);
418 return ERR_CAST(pages);
419 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
420 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
421 if (msm_obj->vaddr == NULL)
422 return ERR_PTR(-ENOMEM);
424 msm_obj->vmap_count++;
425 return msm_obj->vaddr;
428 void *msm_gem_get_vaddr(struct drm_gem_object *obj)
431 mutex_lock(&obj->dev->struct_mutex);
432 ret = msm_gem_get_vaddr_locked(obj);
433 mutex_unlock(&obj->dev->struct_mutex);
437 void msm_gem_put_vaddr_locked(struct drm_gem_object *obj)
439 struct msm_gem_object *msm_obj = to_msm_bo(obj);
440 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
441 WARN_ON(msm_obj->vmap_count < 1);
442 msm_obj->vmap_count--;
445 void msm_gem_put_vaddr(struct drm_gem_object *obj)
447 mutex_lock(&obj->dev->struct_mutex);
448 msm_gem_put_vaddr_locked(obj);
449 mutex_unlock(&obj->dev->struct_mutex);
452 /* Update madvise status, returns true if not purged, else
455 int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
457 struct msm_gem_object *msm_obj = to_msm_bo(obj);
459 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
461 if (msm_obj->madv != __MSM_MADV_PURGED)
462 msm_obj->madv = madv;
464 return (msm_obj->madv != __MSM_MADV_PURGED);
467 void msm_gem_purge(struct drm_gem_object *obj)
469 struct drm_device *dev = obj->dev;
470 struct msm_gem_object *msm_obj = to_msm_bo(obj);
472 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
473 WARN_ON(!is_purgeable(msm_obj));
474 WARN_ON(obj->import_attach);
482 msm_obj->madv = __MSM_MADV_PURGED;
484 drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
485 drm_gem_free_mmap_offset(obj);
487 /* Our goal here is to return as much of the memory as
488 * is possible back to the system as we are called from OOM.
489 * To do this we must instruct the shmfs to drop all of its
490 * backing pages, *now*.
492 shmem_truncate_range(file_inode(obj->filp), 0, (loff_t)-1);
494 invalidate_mapping_pages(file_inode(obj->filp)->i_mapping,
498 void msm_gem_vunmap(struct drm_gem_object *obj)
500 struct msm_gem_object *msm_obj = to_msm_bo(obj);
502 if (!msm_obj->vaddr || WARN_ON(!is_vunmapable(msm_obj)))
505 vunmap(msm_obj->vaddr);
506 msm_obj->vaddr = NULL;
509 /* must be called before _move_to_active().. */
510 int msm_gem_sync_object(struct drm_gem_object *obj,
511 struct msm_fence_context *fctx, bool exclusive)
513 struct msm_gem_object *msm_obj = to_msm_bo(obj);
514 struct reservation_object_list *fobj;
519 /* NOTE: _reserve_shared() must happen before _add_shared_fence(),
520 * which makes this a slightly strange place to call it. OTOH this
521 * is a convenient can-fail point to hook it in. (And similar to
522 * how etnaviv and nouveau handle this.)
524 ret = reservation_object_reserve_shared(msm_obj->resv);
529 fobj = reservation_object_get_list(msm_obj->resv);
530 if (!fobj || (fobj->shared_count == 0)) {
531 fence = reservation_object_get_excl(msm_obj->resv);
532 /* don't need to wait on our own fences, since ring is fifo */
533 if (fence && (fence->context != fctx->context)) {
534 ret = fence_wait(fence, true);
540 if (!exclusive || !fobj)
543 for (i = 0; i < fobj->shared_count; i++) {
544 fence = rcu_dereference_protected(fobj->shared[i],
545 reservation_object_held(msm_obj->resv));
546 if (fence->context != fctx->context) {
547 ret = fence_wait(fence, true);
556 void msm_gem_move_to_active(struct drm_gem_object *obj,
557 struct msm_gpu *gpu, bool exclusive, struct fence *fence)
559 struct msm_gem_object *msm_obj = to_msm_bo(obj);
560 WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED);
563 reservation_object_add_excl_fence(msm_obj->resv, fence);
565 reservation_object_add_shared_fence(msm_obj->resv, fence);
566 list_del_init(&msm_obj->mm_list);
567 list_add_tail(&msm_obj->mm_list, &gpu->active_list);
570 void msm_gem_move_to_inactive(struct drm_gem_object *obj)
572 struct drm_device *dev = obj->dev;
573 struct msm_drm_private *priv = dev->dev_private;
574 struct msm_gem_object *msm_obj = to_msm_bo(obj);
576 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
579 list_del_init(&msm_obj->mm_list);
580 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
583 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout)
585 struct msm_gem_object *msm_obj = to_msm_bo(obj);
586 bool write = !!(op & MSM_PREP_WRITE);
588 if (op & MSM_PREP_NOSYNC) {
589 if (!reservation_object_test_signaled_rcu(msm_obj->resv, write))
594 ret = reservation_object_wait_timeout_rcu(msm_obj->resv, write,
595 true, timeout_to_jiffies(timeout));
597 return ret == 0 ? -ETIMEDOUT : ret;
600 /* TODO cache maintenance */
605 int msm_gem_cpu_fini(struct drm_gem_object *obj)
607 /* TODO cache maintenance */
611 #ifdef CONFIG_DEBUG_FS
612 static void describe_fence(struct fence *fence, const char *type,
615 if (!fence_is_signaled(fence))
616 seq_printf(m, "\t%9s: %s %s seq %u\n", type,
617 fence->ops->get_driver_name(fence),
618 fence->ops->get_timeline_name(fence),
622 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
624 struct msm_gem_object *msm_obj = to_msm_bo(obj);
625 struct reservation_object *robj = msm_obj->resv;
626 struct reservation_object_list *fobj;
628 uint64_t off = drm_vma_node_start(&obj->vma_node);
631 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
633 switch (msm_obj->madv) {
634 case __MSM_MADV_PURGED:
637 case MSM_MADV_DONTNEED:
640 case MSM_MADV_WILLNEED:
646 seq_printf(m, "%08x: %c %2d (%2d) %08llx %p %zu%s\n",
647 msm_obj->flags, is_active(msm_obj) ? 'A' : 'I',
648 obj->name, obj->refcount.refcount.counter,
649 off, msm_obj->vaddr, obj->size, madv);
652 fobj = rcu_dereference(robj->fence);
654 unsigned int i, shared_count = fobj->shared_count;
656 for (i = 0; i < shared_count; i++) {
657 fence = rcu_dereference(fobj->shared[i]);
658 describe_fence(fence, "Shared", m);
662 fence = rcu_dereference(robj->fence_excl);
664 describe_fence(fence, "Exclusive", m);
668 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
670 struct msm_gem_object *msm_obj;
674 list_for_each_entry(msm_obj, list, mm_list) {
675 struct drm_gem_object *obj = &msm_obj->base;
677 msm_gem_describe(obj, m);
682 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
686 void msm_gem_free_object(struct drm_gem_object *obj)
688 struct drm_device *dev = obj->dev;
689 struct msm_gem_object *msm_obj = to_msm_bo(obj);
691 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
693 /* object should not be on active list: */
694 WARN_ON(is_active(msm_obj));
696 list_del(&msm_obj->mm_list);
700 if (obj->import_attach) {
702 dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr);
704 /* Don't drop the pages for imported dmabuf, as they are not
705 * ours, just free the array we allocated:
708 drm_free_large(msm_obj->pages);
710 drm_prime_gem_destroy(obj, msm_obj->sgt);
716 if (msm_obj->resv == &msm_obj->_resv)
717 reservation_object_fini(msm_obj->resv);
719 drm_gem_object_release(obj);
724 /* convenience method to construct a GEM buffer object, and userspace handle */
725 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
726 uint32_t size, uint32_t flags, uint32_t *handle)
728 struct drm_gem_object *obj;
731 ret = mutex_lock_interruptible(&dev->struct_mutex);
735 obj = msm_gem_new(dev, size, flags);
737 mutex_unlock(&dev->struct_mutex);
742 ret = drm_gem_handle_create(file, obj, handle);
744 /* drop reference from allocate - handle holds it now */
745 drm_gem_object_unreference_unlocked(obj);
750 static int msm_gem_new_impl(struct drm_device *dev,
751 uint32_t size, uint32_t flags,
752 struct reservation_object *resv,
753 struct drm_gem_object **obj)
755 struct msm_drm_private *priv = dev->dev_private;
756 struct msm_gem_object *msm_obj;
758 bool use_vram = false;
760 switch (flags & MSM_BO_CACHE_MASK) {
761 case MSM_BO_UNCACHED:
766 dev_err(dev->dev, "invalid cache flag: %x\n",
767 (flags & MSM_BO_CACHE_MASK));
771 if (!iommu_present(&platform_bus_type))
773 else if ((flags & MSM_BO_STOLEN) && priv->vram.size)
776 if (WARN_ON(use_vram && !priv->vram.size))
779 sz = sizeof(*msm_obj);
781 sz += sizeof(struct drm_mm_node);
783 msm_obj = kzalloc(sz, GFP_KERNEL);
788 msm_obj->vram_node = (void *)&msm_obj[1];
790 msm_obj->flags = flags;
791 msm_obj->madv = MSM_MADV_WILLNEED;
794 msm_obj->resv = resv;
796 msm_obj->resv = &msm_obj->_resv;
797 reservation_object_init(msm_obj->resv);
800 INIT_LIST_HEAD(&msm_obj->submit_entry);
801 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
803 *obj = &msm_obj->base;
808 struct drm_gem_object *msm_gem_new(struct drm_device *dev,
809 uint32_t size, uint32_t flags)
811 struct drm_gem_object *obj = NULL;
814 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
816 size = PAGE_ALIGN(size);
818 ret = msm_gem_new_impl(dev, size, flags, NULL, &obj);
822 if (use_pages(obj)) {
823 ret = drm_gem_object_init(dev, obj, size);
827 drm_gem_private_object_init(dev, obj, size);
833 drm_gem_object_unreference(obj);
837 struct drm_gem_object *msm_gem_import(struct drm_device *dev,
838 struct dma_buf *dmabuf, struct sg_table *sgt)
840 struct msm_gem_object *msm_obj;
841 struct drm_gem_object *obj;
845 /* if we don't have IOMMU, don't bother pretending we can import: */
846 if (!iommu_present(&platform_bus_type)) {
847 dev_err(dev->dev, "cannot import without IOMMU\n");
848 return ERR_PTR(-EINVAL);
851 size = PAGE_ALIGN(dmabuf->size);
853 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, dmabuf->resv, &obj);
857 drm_gem_private_object_init(dev, obj, size);
859 npages = size / PAGE_SIZE;
861 msm_obj = to_msm_bo(obj);
863 msm_obj->pages = drm_malloc_ab(npages, sizeof(struct page *));
864 if (!msm_obj->pages) {
869 ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages);
876 drm_gem_object_unreference_unlocked(obj);