2 * Copyright (C) 2013 Red Hat
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>
28 static dma_addr_t physaddr(struct drm_gem_object *obj)
30 struct msm_gem_object *msm_obj = to_msm_bo(obj);
31 struct msm_drm_private *priv = obj->dev->dev_private;
32 return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) +
36 static bool use_pages(struct drm_gem_object *obj)
38 struct msm_gem_object *msm_obj = to_msm_bo(obj);
39 return !msm_obj->vram_node;
42 /* allocate pages from VRAM carveout, used when no IOMMU: */
43 static struct page **get_pages_vram(struct drm_gem_object *obj,
46 struct msm_gem_object *msm_obj = to_msm_bo(obj);
47 struct msm_drm_private *priv = obj->dev->dev_private;
52 p = drm_malloc_ab(npages, sizeof(struct page *));
54 return ERR_PTR(-ENOMEM);
56 ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node,
57 npages, 0, DRM_MM_SEARCH_DEFAULT);
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;
203 /* Make sure we don't parallel update on a fault, nor move or remove
204 * something from beneath our feet
206 ret = mutex_lock_interruptible(&dev->struct_mutex);
210 /* make sure we have pages attached now */
211 pages = get_pages(obj);
213 ret = PTR_ERR(pages);
217 /* We don't use vmf->pgoff since that has the fake offset: */
218 pgoff = ((unsigned long)vmf->virtual_address -
219 vma->vm_start) >> PAGE_SHIFT;
221 pfn = page_to_pfn(pages[pgoff]);
223 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
224 pfn, pfn << PAGE_SHIFT);
226 ret = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address,
227 __pfn_to_pfn_t(pfn, PFN_DEV));
230 mutex_unlock(&dev->struct_mutex);
239 * EBUSY is ok: this just means that another thread
240 * already did the job.
242 return VM_FAULT_NOPAGE;
246 return VM_FAULT_SIGBUS;
250 /** get mmap offset */
251 static uint64_t mmap_offset(struct drm_gem_object *obj)
253 struct drm_device *dev = obj->dev;
256 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
258 /* Make it mmapable */
259 ret = drm_gem_create_mmap_offset(obj);
262 dev_err(dev->dev, "could not allocate mmap offset\n");
266 return drm_vma_node_offset_addr(&obj->vma_node);
269 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
272 mutex_lock(&obj->dev->struct_mutex);
273 offset = mmap_offset(obj);
274 mutex_unlock(&obj->dev->struct_mutex);
278 /* should be called under struct_mutex.. although it can be called
279 * from atomic context without struct_mutex to acquire an extra
280 * iova ref if you know one is already held.
282 * That means when I do eventually need to add support for unpinning
283 * the refcnt counter needs to be atomic_t.
285 int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id,
288 struct msm_gem_object *msm_obj = to_msm_bo(obj);
291 if (!msm_obj->domain[id].iova) {
292 struct msm_drm_private *priv = obj->dev->dev_private;
293 struct page **pages = get_pages(obj);
296 return PTR_ERR(pages);
298 if (iommu_present(&platform_bus_type)) {
299 struct msm_mmu *mmu = priv->mmus[id];
305 offset = (uint32_t)mmap_offset(obj);
306 ret = mmu->funcs->map(mmu, offset, msm_obj->sgt,
307 obj->size, IOMMU_READ | IOMMU_WRITE);
308 msm_obj->domain[id].iova = offset;
310 msm_obj->domain[id].iova = physaddr(obj);
315 *iova = msm_obj->domain[id].iova;
320 /* get iova, taking a reference. Should have a matching put */
321 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova)
323 struct msm_gem_object *msm_obj = to_msm_bo(obj);
326 /* this is safe right now because we don't unmap until the
329 if (msm_obj->domain[id].iova) {
330 *iova = msm_obj->domain[id].iova;
334 mutex_lock(&obj->dev->struct_mutex);
335 ret = msm_gem_get_iova_locked(obj, id, iova);
336 mutex_unlock(&obj->dev->struct_mutex);
340 /* get iova without taking a reference, used in places where you have
341 * already done a 'msm_gem_get_iova()'.
343 uint32_t msm_gem_iova(struct drm_gem_object *obj, int id)
345 struct msm_gem_object *msm_obj = to_msm_bo(obj);
346 WARN_ON(!msm_obj->domain[id].iova);
347 return msm_obj->domain[id].iova;
350 void msm_gem_put_iova(struct drm_gem_object *obj, int id)
353 // NOTE: probably don't need a _locked() version.. we wouldn't
354 // normally unmap here, but instead just mark that it could be
355 // unmapped (if the iova refcnt drops to zero), but then later
356 // if another _get_iova_locked() fails we can start unmapping
357 // things that are no longer needed..
360 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
361 struct drm_mode_create_dumb *args)
363 args->pitch = align_pitch(args->width, args->bpp);
364 args->size = PAGE_ALIGN(args->pitch * args->height);
365 return msm_gem_new_handle(dev, file, args->size,
366 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle);
369 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
370 uint32_t handle, uint64_t *offset)
372 struct drm_gem_object *obj;
375 /* GEM does all our handle to object mapping */
376 obj = drm_gem_object_lookup(dev, file, handle);
382 *offset = msm_gem_mmap_offset(obj);
384 drm_gem_object_unreference_unlocked(obj);
390 void *msm_gem_vaddr_locked(struct drm_gem_object *obj)
392 struct msm_gem_object *msm_obj = to_msm_bo(obj);
393 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
394 if (!msm_obj->vaddr) {
395 struct page **pages = get_pages(obj);
397 return ERR_CAST(pages);
398 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
399 VM_MAP, pgprot_writecombine(PAGE_KERNEL));
401 return msm_obj->vaddr;
404 void *msm_gem_vaddr(struct drm_gem_object *obj)
407 mutex_lock(&obj->dev->struct_mutex);
408 ret = msm_gem_vaddr_locked(obj);
409 mutex_unlock(&obj->dev->struct_mutex);
413 /* setup callback for when bo is no longer busy..
414 * TODO probably want to differentiate read vs write..
416 int msm_gem_queue_inactive_cb(struct drm_gem_object *obj,
417 struct msm_fence_cb *cb)
419 struct msm_gem_object *msm_obj = to_msm_bo(obj);
420 uint32_t fence = msm_gem_fence(msm_obj,
421 MSM_PREP_READ | MSM_PREP_WRITE);
422 return msm_queue_fence_cb(obj->dev, cb, fence);
425 void msm_gem_move_to_active(struct drm_gem_object *obj,
426 struct msm_gpu *gpu, bool write, uint32_t fence)
428 struct msm_gem_object *msm_obj = to_msm_bo(obj);
431 msm_obj->write_fence = fence;
433 msm_obj->read_fence = fence;
434 list_del_init(&msm_obj->mm_list);
435 list_add_tail(&msm_obj->mm_list, &gpu->active_list);
438 void msm_gem_move_to_inactive(struct drm_gem_object *obj)
440 struct drm_device *dev = obj->dev;
441 struct msm_drm_private *priv = dev->dev_private;
442 struct msm_gem_object *msm_obj = to_msm_bo(obj);
444 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
447 msm_obj->read_fence = 0;
448 msm_obj->write_fence = 0;
449 list_del_init(&msm_obj->mm_list);
450 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
453 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout)
455 struct drm_device *dev = obj->dev;
456 struct msm_gem_object *msm_obj = to_msm_bo(obj);
459 if (is_active(msm_obj)) {
460 uint32_t fence = msm_gem_fence(msm_obj, op);
462 if (op & MSM_PREP_NOSYNC)
465 ret = msm_wait_fence(dev, fence, timeout, true);
468 /* TODO cache maintenance */
473 int msm_gem_cpu_fini(struct drm_gem_object *obj)
475 /* TODO cache maintenance */
479 #ifdef CONFIG_DEBUG_FS
480 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
482 struct drm_device *dev = obj->dev;
483 struct msm_gem_object *msm_obj = to_msm_bo(obj);
484 uint64_t off = drm_vma_node_start(&obj->vma_node);
486 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
487 seq_printf(m, "%08x: %c(r=%u,w=%u) %2d (%2d) %08llx %p %zu\n",
488 msm_obj->flags, is_active(msm_obj) ? 'A' : 'I',
489 msm_obj->read_fence, msm_obj->write_fence,
490 obj->name, obj->refcount.refcount.counter,
491 off, msm_obj->vaddr, obj->size);
494 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
496 struct msm_gem_object *msm_obj;
500 list_for_each_entry(msm_obj, list, mm_list) {
501 struct drm_gem_object *obj = &msm_obj->base;
503 msm_gem_describe(obj, m);
508 seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
512 void msm_gem_free_object(struct drm_gem_object *obj)
514 struct drm_device *dev = obj->dev;
515 struct msm_drm_private *priv = obj->dev->dev_private;
516 struct msm_gem_object *msm_obj = to_msm_bo(obj);
519 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
521 /* object should not be on active list: */
522 WARN_ON(is_active(msm_obj));
524 list_del(&msm_obj->mm_list);
526 for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) {
527 struct msm_mmu *mmu = priv->mmus[id];
528 if (mmu && msm_obj->domain[id].iova) {
529 uint32_t offset = msm_obj->domain[id].iova;
530 mmu->funcs->unmap(mmu, offset, msm_obj->sgt, obj->size);
534 if (obj->import_attach) {
536 dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr);
538 /* Don't drop the pages for imported dmabuf, as they are not
539 * ours, just free the array we allocated:
542 drm_free_large(msm_obj->pages);
544 drm_prime_gem_destroy(obj, msm_obj->sgt);
546 vunmap(msm_obj->vaddr);
550 if (msm_obj->resv == &msm_obj->_resv)
551 reservation_object_fini(msm_obj->resv);
553 drm_gem_object_release(obj);
558 /* convenience method to construct a GEM buffer object, and userspace handle */
559 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
560 uint32_t size, uint32_t flags, uint32_t *handle)
562 struct drm_gem_object *obj;
565 ret = mutex_lock_interruptible(&dev->struct_mutex);
569 obj = msm_gem_new(dev, size, flags);
571 mutex_unlock(&dev->struct_mutex);
576 ret = drm_gem_handle_create(file, obj, handle);
578 /* drop reference from allocate - handle holds it now */
579 drm_gem_object_unreference_unlocked(obj);
584 static int msm_gem_new_impl(struct drm_device *dev,
585 uint32_t size, uint32_t flags,
586 struct drm_gem_object **obj)
588 struct msm_drm_private *priv = dev->dev_private;
589 struct msm_gem_object *msm_obj;
591 bool use_vram = false;
593 switch (flags & MSM_BO_CACHE_MASK) {
594 case MSM_BO_UNCACHED:
599 dev_err(dev->dev, "invalid cache flag: %x\n",
600 (flags & MSM_BO_CACHE_MASK));
604 if (!iommu_present(&platform_bus_type))
606 else if ((flags & MSM_BO_STOLEN) && priv->vram.size)
609 if (WARN_ON(use_vram && !priv->vram.size))
612 sz = sizeof(*msm_obj);
614 sz += sizeof(struct drm_mm_node);
616 msm_obj = kzalloc(sz, GFP_KERNEL);
621 msm_obj->vram_node = (void *)&msm_obj[1];
623 msm_obj->flags = flags;
625 msm_obj->resv = &msm_obj->_resv;
626 reservation_object_init(msm_obj->resv);
628 INIT_LIST_HEAD(&msm_obj->submit_entry);
629 list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
631 *obj = &msm_obj->base;
636 struct drm_gem_object *msm_gem_new(struct drm_device *dev,
637 uint32_t size, uint32_t flags)
639 struct drm_gem_object *obj = NULL;
642 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
644 size = PAGE_ALIGN(size);
646 ret = msm_gem_new_impl(dev, size, flags, &obj);
650 if (use_pages(obj)) {
651 ret = drm_gem_object_init(dev, obj, size);
655 drm_gem_private_object_init(dev, obj, size);
662 drm_gem_object_unreference(obj);
667 struct drm_gem_object *msm_gem_import(struct drm_device *dev,
668 uint32_t size, struct sg_table *sgt)
670 struct msm_gem_object *msm_obj;
671 struct drm_gem_object *obj;
674 /* if we don't have IOMMU, don't bother pretending we can import: */
675 if (!iommu_present(&platform_bus_type)) {
676 dev_err(dev->dev, "cannot import without IOMMU\n");
677 return ERR_PTR(-EINVAL);
680 size = PAGE_ALIGN(size);
682 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, &obj);
686 drm_gem_private_object_init(dev, obj, size);
688 npages = size / PAGE_SIZE;
690 msm_obj = to_msm_bo(obj);
692 msm_obj->pages = drm_malloc_ab(npages, sizeof(struct page *));
693 if (!msm_obj->pages) {
698 ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages);
706 drm_gem_object_unreference_unlocked(obj);