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, npages);
63 paddr = physaddr(obj);
64 for (i = 0; i < npages; i++) {
65 p[i] = phys_to_page(paddr);
72 /* called with dev->struct_mutex held */
73 static struct page **get_pages(struct drm_gem_object *obj)
75 struct msm_gem_object *msm_obj = to_msm_bo(obj);
77 if (!msm_obj->pages) {
78 struct drm_device *dev = obj->dev;
80 int npages = obj->size >> PAGE_SHIFT;
83 p = drm_gem_get_pages(obj);
85 p = get_pages_vram(obj, npages);
88 dev_err(dev->dev, "could not get pages: %ld\n",
93 msm_obj->sgt = drm_prime_pages_to_sg(p, npages);
94 if (IS_ERR(msm_obj->sgt)) {
95 dev_err(dev->dev, "failed to allocate sgt\n");
96 return ERR_CAST(msm_obj->sgt);
101 /* For non-cached buffers, ensure the new pages are clean
102 * because display controller, GPU, etc. are not coherent:
104 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
105 dma_map_sg(dev->dev, msm_obj->sgt->sgl,
106 msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
109 return msm_obj->pages;
112 static void put_pages(struct drm_gem_object *obj)
114 struct msm_gem_object *msm_obj = to_msm_bo(obj);
116 if (msm_obj->pages) {
117 /* For non-cached buffers, ensure the new pages are clean
118 * because display controller, GPU, etc. are not coherent:
120 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
121 dma_unmap_sg(obj->dev->dev, msm_obj->sgt->sgl,
122 msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
123 sg_free_table(msm_obj->sgt);
127 drm_gem_put_pages(obj, msm_obj->pages, true, false);
129 drm_mm_remove_node(msm_obj->vram_node);
130 drm_free_large(msm_obj->pages);
133 msm_obj->pages = NULL;
137 struct page **msm_gem_get_pages(struct drm_gem_object *obj)
139 struct drm_device *dev = obj->dev;
141 mutex_lock(&dev->struct_mutex);
143 mutex_unlock(&dev->struct_mutex);
147 void msm_gem_put_pages(struct drm_gem_object *obj)
149 /* when we start tracking the pin count, then do something here */
152 int msm_gem_mmap_obj(struct drm_gem_object *obj,
153 struct vm_area_struct *vma)
155 struct msm_gem_object *msm_obj = to_msm_bo(obj);
157 vma->vm_flags &= ~VM_PFNMAP;
158 vma->vm_flags |= VM_MIXEDMAP;
160 if (msm_obj->flags & MSM_BO_WC) {
161 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
162 } else if (msm_obj->flags & MSM_BO_UNCACHED) {
163 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
166 * Shunt off cached objs to shmem file so they have their own
167 * address_space (so unmap_mapping_range does what we want,
168 * in particular in the case of mmap'd dmabufs)
173 vma->vm_file = obj->filp;
175 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
181 int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
185 ret = drm_gem_mmap(filp, vma);
187 DBG("mmap failed: %d", ret);
191 return msm_gem_mmap_obj(vma->vm_private_data, vma);
194 int msm_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
196 struct drm_gem_object *obj = vma->vm_private_data;
197 struct drm_device *dev = obj->dev;
198 struct msm_drm_private *priv = dev->dev_private;
204 /* This should only happen if userspace tries to pass a mmap'd
205 * but unfaulted gem bo vaddr into submit ioctl, triggering
206 * a page fault while struct_mutex is already held. This is
207 * not a valid use-case so just bail.
209 if (priv->struct_mutex_task == current)
210 return VM_FAULT_SIGBUS;
212 /* Make sure we don't parallel update on a fault, nor move or remove
213 * something from beneath our feet
215 ret = mutex_lock_interruptible(&dev->struct_mutex);
219 /* make sure we have pages attached now */
220 pages = get_pages(obj);
222 ret = PTR_ERR(pages);
226 /* We don't use vmf->pgoff since that has the fake offset: */
227 pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
229 pfn = page_to_pfn(pages[pgoff]);
231 VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
232 pfn, pfn << PAGE_SHIFT);
234 ret = vm_insert_mixed(vma, vmf->address, __pfn_to_pfn_t(pfn, PFN_DEV));
237 mutex_unlock(&dev->struct_mutex);
246 * EBUSY is ok: this just means that another thread
247 * already did the job.
249 return VM_FAULT_NOPAGE;
253 return VM_FAULT_SIGBUS;
257 /** get mmap offset */
258 static uint64_t mmap_offset(struct drm_gem_object *obj)
260 struct drm_device *dev = obj->dev;
263 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
265 /* Make it mmapable */
266 ret = drm_gem_create_mmap_offset(obj);
269 dev_err(dev->dev, "could not allocate mmap offset\n");
273 return drm_vma_node_offset_addr(&obj->vma_node);
276 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
279 mutex_lock(&obj->dev->struct_mutex);
280 offset = mmap_offset(obj);
281 mutex_unlock(&obj->dev->struct_mutex);
286 put_iova(struct drm_gem_object *obj)
288 struct drm_device *dev = obj->dev;
289 struct msm_drm_private *priv = obj->dev->dev_private;
290 struct msm_gem_object *msm_obj = to_msm_bo(obj);
293 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
295 for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) {
296 if (!priv->aspace[id])
298 msm_gem_unmap_vma(priv->aspace[id],
299 &msm_obj->domain[id], msm_obj->sgt);
303 /* should be called under struct_mutex.. although it can be called
304 * from atomic context without struct_mutex to acquire an extra
305 * iova ref if you know one is already held.
307 * That means when I do eventually need to add support for unpinning
308 * the refcnt counter needs to be atomic_t.
310 int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id,
313 struct msm_gem_object *msm_obj = to_msm_bo(obj);
316 if (!msm_obj->domain[id].iova) {
317 struct msm_drm_private *priv = obj->dev->dev_private;
318 struct page **pages = get_pages(obj);
321 return PTR_ERR(pages);
323 if (iommu_present(&platform_bus_type)) {
324 ret = msm_gem_map_vma(priv->aspace[id], &msm_obj->domain[id],
325 msm_obj->sgt, obj->size >> PAGE_SHIFT);
327 msm_obj->domain[id].iova = physaddr(obj);
332 *iova = msm_obj->domain[id].iova;
337 /* get iova, taking a reference. Should have a matching put */
338 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint64_t *iova)
340 struct msm_gem_object *msm_obj = to_msm_bo(obj);
343 /* this is safe right now because we don't unmap until the
346 if (msm_obj->domain[id].iova) {
347 *iova = msm_obj->domain[id].iova;
351 mutex_lock(&obj->dev->struct_mutex);
352 ret = msm_gem_get_iova_locked(obj, id, iova);
353 mutex_unlock(&obj->dev->struct_mutex);
357 /* get iova without taking a reference, used in places where you have
358 * already done a 'msm_gem_get_iova()'.
360 uint64_t msm_gem_iova(struct drm_gem_object *obj, int id)
362 struct msm_gem_object *msm_obj = to_msm_bo(obj);
363 WARN_ON(!msm_obj->domain[id].iova);
364 return msm_obj->domain[id].iova;
367 void msm_gem_put_iova(struct drm_gem_object *obj, int id)
370 // NOTE: probably don't need a _locked() version.. we wouldn't
371 // normally unmap here, but instead just mark that it could be
372 // unmapped (if the iova refcnt drops to zero), but then later
373 // if another _get_iova_locked() fails we can start unmapping
374 // things that are no longer needed..
377 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
378 struct drm_mode_create_dumb *args)
380 args->pitch = align_pitch(args->width, args->bpp);
381 args->size = PAGE_ALIGN(args->pitch * args->height);
382 return msm_gem_new_handle(dev, file, args->size,
383 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle);
386 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
387 uint32_t handle, uint64_t *offset)
389 struct drm_gem_object *obj;
392 /* GEM does all our handle to object mapping */
393 obj = drm_gem_object_lookup(file, handle);
399 *offset = msm_gem_mmap_offset(obj);
401 drm_gem_object_unreference_unlocked(obj);
407 void *msm_gem_get_vaddr_locked(struct drm_gem_object *obj)
409 struct msm_gem_object *msm_obj = to_msm_bo(obj);
410 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
411 if (!msm_obj->vaddr) {
412 struct page **pages = get_pages(obj);
414 return ERR_CAST(pages);
415 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
416 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
417 if (msm_obj->vaddr == NULL)
418 return ERR_PTR(-ENOMEM);
420 msm_obj->vmap_count++;
421 return msm_obj->vaddr;
424 void *msm_gem_get_vaddr(struct drm_gem_object *obj)
427 mutex_lock(&obj->dev->struct_mutex);
428 ret = msm_gem_get_vaddr_locked(obj);
429 mutex_unlock(&obj->dev->struct_mutex);
433 void msm_gem_put_vaddr_locked(struct drm_gem_object *obj)
435 struct msm_gem_object *msm_obj = to_msm_bo(obj);
436 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
437 WARN_ON(msm_obj->vmap_count < 1);
438 msm_obj->vmap_count--;
441 void msm_gem_put_vaddr(struct drm_gem_object *obj)
443 mutex_lock(&obj->dev->struct_mutex);
444 msm_gem_put_vaddr_locked(obj);
445 mutex_unlock(&obj->dev->struct_mutex);
448 /* Update madvise status, returns true if not purged, else
451 int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
453 struct msm_gem_object *msm_obj = to_msm_bo(obj);
455 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
457 if (msm_obj->madv != __MSM_MADV_PURGED)
458 msm_obj->madv = madv;
460 return (msm_obj->madv != __MSM_MADV_PURGED);
463 void msm_gem_purge(struct drm_gem_object *obj)
465 struct drm_device *dev = obj->dev;
466 struct msm_gem_object *msm_obj = to_msm_bo(obj);
468 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
469 WARN_ON(!is_purgeable(msm_obj));
470 WARN_ON(obj->import_attach);
478 msm_obj->madv = __MSM_MADV_PURGED;
480 drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
481 drm_gem_free_mmap_offset(obj);
483 /* Our goal here is to return as much of the memory as
484 * is possible back to the system as we are called from OOM.
485 * To do this we must instruct the shmfs to drop all of its
486 * backing pages, *now*.
488 shmem_truncate_range(file_inode(obj->filp), 0, (loff_t)-1);
490 invalidate_mapping_pages(file_inode(obj->filp)->i_mapping,
494 void msm_gem_vunmap(struct drm_gem_object *obj)
496 struct msm_gem_object *msm_obj = to_msm_bo(obj);
498 if (!msm_obj->vaddr || WARN_ON(!is_vunmapable(msm_obj)))
501 vunmap(msm_obj->vaddr);
502 msm_obj->vaddr = NULL;
505 /* must be called before _move_to_active().. */
506 int msm_gem_sync_object(struct drm_gem_object *obj,
507 struct msm_fence_context *fctx, bool exclusive)
509 struct msm_gem_object *msm_obj = to_msm_bo(obj);
510 struct reservation_object_list *fobj;
511 struct dma_fence *fence;
515 /* NOTE: _reserve_shared() must happen before _add_shared_fence(),
516 * which makes this a slightly strange place to call it. OTOH this
517 * is a convenient can-fail point to hook it in. (And similar to
518 * how etnaviv and nouveau handle this.)
520 ret = reservation_object_reserve_shared(msm_obj->resv);
525 fobj = reservation_object_get_list(msm_obj->resv);
526 if (!fobj || (fobj->shared_count == 0)) {
527 fence = reservation_object_get_excl(msm_obj->resv);
528 /* don't need to wait on our own fences, since ring is fifo */
529 if (fence && (fence->context != fctx->context)) {
530 ret = dma_fence_wait(fence, true);
536 if (!exclusive || !fobj)
539 for (i = 0; i < fobj->shared_count; i++) {
540 fence = rcu_dereference_protected(fobj->shared[i],
541 reservation_object_held(msm_obj->resv));
542 if (fence->context != fctx->context) {
543 ret = dma_fence_wait(fence, true);
552 void msm_gem_move_to_active(struct drm_gem_object *obj,
553 struct msm_gpu *gpu, bool exclusive, struct dma_fence *fence)
555 struct msm_gem_object *msm_obj = to_msm_bo(obj);
556 WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED);
559 reservation_object_add_excl_fence(msm_obj->resv, fence);
561 reservation_object_add_shared_fence(msm_obj->resv, fence);
562 list_del_init(&msm_obj->mm_list);
563 list_add_tail(&msm_obj->mm_list, &gpu->active_list);
566 void msm_gem_move_to_inactive(struct drm_gem_object *obj)
568 struct drm_device *dev = obj->dev;
569 struct msm_drm_private *priv = dev->dev_private;
570 struct msm_gem_object *msm_obj = to_msm_bo(obj);
572 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
575 list_del_init(&msm_obj->mm_list);
576 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
579 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout)
581 struct msm_gem_object *msm_obj = to_msm_bo(obj);
582 bool write = !!(op & MSM_PREP_WRITE);
583 unsigned long remain =
584 op & MSM_PREP_NOSYNC ? 0 : timeout_to_jiffies(timeout);
587 ret = reservation_object_wait_timeout_rcu(msm_obj->resv, write,
590 return remain == 0 ? -EBUSY : -ETIMEDOUT;
594 /* TODO cache maintenance */
599 int msm_gem_cpu_fini(struct drm_gem_object *obj)
601 /* TODO cache maintenance */
605 #ifdef CONFIG_DEBUG_FS
606 static void describe_fence(struct dma_fence *fence, const char *type,
609 if (!dma_fence_is_signaled(fence))
610 seq_printf(m, "\t%9s: %s %s seq %u\n", type,
611 fence->ops->get_driver_name(fence),
612 fence->ops->get_timeline_name(fence),
616 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
618 struct msm_gem_object *msm_obj = to_msm_bo(obj);
619 struct reservation_object *robj = msm_obj->resv;
620 struct reservation_object_list *fobj;
621 struct msm_drm_private *priv = obj->dev->dev_private;
622 struct dma_fence *fence;
623 uint64_t off = drm_vma_node_start(&obj->vma_node);
627 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
629 switch (msm_obj->madv) {
630 case __MSM_MADV_PURGED:
633 case MSM_MADV_DONTNEED:
636 case MSM_MADV_WILLNEED:
642 seq_printf(m, "%08x: %c %2d (%2d) %08llx %p\t",
643 msm_obj->flags, is_active(msm_obj) ? 'A' : 'I',
644 obj->name, kref_read(&obj->refcount),
645 off, msm_obj->vaddr);
647 for (id = 0; id < priv->num_aspaces; id++)
648 seq_printf(m, " %08llx", msm_obj->domain[id].iova);
650 seq_printf(m, " %zu%s\n", obj->size, madv);
653 fobj = rcu_dereference(robj->fence);
655 unsigned int i, shared_count = fobj->shared_count;
657 for (i = 0; i < shared_count; i++) {
658 fence = rcu_dereference(fobj->shared[i]);
659 describe_fence(fence, "Shared", m);
663 fence = rcu_dereference(robj->fence_excl);
665 describe_fence(fence, "Exclusive", m);
669 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
671 struct msm_gem_object *msm_obj;
675 list_for_each_entry(msm_obj, list, mm_list) {
676 struct drm_gem_object *obj = &msm_obj->base;
678 msm_gem_describe(obj, m);
683 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
687 void msm_gem_free_object(struct drm_gem_object *obj)
689 struct drm_device *dev = obj->dev;
690 struct msm_gem_object *msm_obj = to_msm_bo(obj);
692 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
694 /* object should not be on active list: */
695 WARN_ON(is_active(msm_obj));
697 list_del(&msm_obj->mm_list);
701 if (obj->import_attach) {
703 dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr);
705 /* Don't drop the pages for imported dmabuf, as they are not
706 * ours, just free the array we allocated:
709 drm_free_large(msm_obj->pages);
711 drm_prime_gem_destroy(obj, msm_obj->sgt);
717 if (msm_obj->resv == &msm_obj->_resv)
718 reservation_object_fini(msm_obj->resv);
720 drm_gem_object_release(obj);
725 /* convenience method to construct a GEM buffer object, and userspace handle */
726 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
727 uint32_t size, uint32_t flags, uint32_t *handle)
729 struct drm_gem_object *obj;
732 ret = mutex_lock_interruptible(&dev->struct_mutex);
736 obj = msm_gem_new(dev, size, flags);
738 mutex_unlock(&dev->struct_mutex);
743 ret = drm_gem_handle_create(file, obj, handle);
745 /* drop reference from allocate - handle holds it now */
746 drm_gem_object_unreference_unlocked(obj);
751 static int msm_gem_new_impl(struct drm_device *dev,
752 uint32_t size, uint32_t flags,
753 struct reservation_object *resv,
754 struct drm_gem_object **obj)
756 struct msm_drm_private *priv = dev->dev_private;
757 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 msm_obj = kzalloc(sizeof(*msm_obj), GFP_KERNEL);
784 msm_obj->vram_node = &msm_obj->domain[0].node;
786 msm_obj->flags = flags;
787 msm_obj->madv = MSM_MADV_WILLNEED;
790 msm_obj->resv = resv;
792 msm_obj->resv = &msm_obj->_resv;
793 reservation_object_init(msm_obj->resv);
796 INIT_LIST_HEAD(&msm_obj->submit_entry);
797 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
799 *obj = &msm_obj->base;
804 struct drm_gem_object *msm_gem_new(struct drm_device *dev,
805 uint32_t size, uint32_t flags)
807 struct drm_gem_object *obj = NULL;
810 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
812 size = PAGE_ALIGN(size);
814 ret = msm_gem_new_impl(dev, size, flags, NULL, &obj);
818 if (use_pages(obj)) {
819 ret = drm_gem_object_init(dev, obj, size);
823 drm_gem_private_object_init(dev, obj, size);
829 drm_gem_object_unreference(obj);
833 struct drm_gem_object *msm_gem_import(struct drm_device *dev,
834 struct dma_buf *dmabuf, struct sg_table *sgt)
836 struct msm_gem_object *msm_obj;
837 struct drm_gem_object *obj;
841 /* if we don't have IOMMU, don't bother pretending we can import: */
842 if (!iommu_present(&platform_bus_type)) {
843 dev_err(dev->dev, "cannot import without IOMMU\n");
844 return ERR_PTR(-EINVAL);
847 size = PAGE_ALIGN(dmabuf->size);
849 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, dmabuf->resv, &obj);
853 drm_gem_private_object_init(dev, obj, size);
855 npages = size / PAGE_SIZE;
857 msm_obj = to_msm_bo(obj);
859 msm_obj->pages = drm_malloc_ab(npages, sizeof(struct page *));
860 if (!msm_obj->pages) {
865 ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages);
872 drm_gem_object_unreference_unlocked(obj);