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 /* allocate pages from VRAM carveout, used when no IOMMU: */
36 static struct page
**get_pages_vram(struct drm_gem_object
*obj
,
39 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
40 struct msm_drm_private
*priv
= obj
->dev
->dev_private
;
45 p
= drm_malloc_ab(npages
, sizeof(struct page
*));
47 return ERR_PTR(-ENOMEM
);
49 ret
= drm_mm_insert_node(&priv
->vram
.mm
, msm_obj
->vram_node
,
50 npages
, 0, DRM_MM_SEARCH_DEFAULT
);
56 paddr
= physaddr(obj
);
57 for (i
= 0; i
< npages
; i
++) {
58 p
[i
] = phys_to_page(paddr
);
65 /* called with dev->struct_mutex held */
66 static struct page
**get_pages(struct drm_gem_object
*obj
)
68 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
70 if (!msm_obj
->pages
) {
71 struct drm_device
*dev
= obj
->dev
;
73 int npages
= obj
->size
>> PAGE_SHIFT
;
75 if (iommu_present(&platform_bus_type
))
76 p
= drm_gem_get_pages(obj
, 0);
78 p
= get_pages_vram(obj
, npages
);
81 dev_err(dev
->dev
, "could not get pages: %ld\n",
86 msm_obj
->sgt
= drm_prime_pages_to_sg(p
, npages
);
87 if (IS_ERR(msm_obj
->sgt
)) {
88 dev_err(dev
->dev
, "failed to allocate sgt\n");
89 return ERR_CAST(msm_obj
->sgt
);
94 /* For non-cached buffers, ensure the new pages are clean
95 * because display controller, GPU, etc. are not coherent:
97 if (msm_obj
->flags
& (MSM_BO_WC
|MSM_BO_UNCACHED
))
98 dma_map_sg(dev
->dev
, msm_obj
->sgt
->sgl
,
99 msm_obj
->sgt
->nents
, DMA_BIDIRECTIONAL
);
102 return msm_obj
->pages
;
105 static void put_pages(struct drm_gem_object
*obj
)
107 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
109 if (msm_obj
->pages
) {
110 /* For non-cached buffers, ensure the new pages are clean
111 * because display controller, GPU, etc. are not coherent:
113 if (msm_obj
->flags
& (MSM_BO_WC
|MSM_BO_UNCACHED
))
114 dma_unmap_sg(obj
->dev
->dev
, msm_obj
->sgt
->sgl
,
115 msm_obj
->sgt
->nents
, DMA_BIDIRECTIONAL
);
116 sg_free_table(msm_obj
->sgt
);
119 if (iommu_present(&platform_bus_type
))
120 drm_gem_put_pages(obj
, msm_obj
->pages
, true, false);
122 drm_mm_remove_node(msm_obj
->vram_node
);
124 msm_obj
->pages
= NULL
;
128 struct page
**msm_gem_get_pages(struct drm_gem_object
*obj
)
130 struct drm_device
*dev
= obj
->dev
;
132 mutex_lock(&dev
->struct_mutex
);
134 mutex_unlock(&dev
->struct_mutex
);
138 void msm_gem_put_pages(struct drm_gem_object
*obj
)
140 /* when we start tracking the pin count, then do something here */
143 int msm_gem_mmap_obj(struct drm_gem_object
*obj
,
144 struct vm_area_struct
*vma
)
146 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
148 vma
->vm_flags
&= ~VM_PFNMAP
;
149 vma
->vm_flags
|= VM_MIXEDMAP
;
151 if (msm_obj
->flags
& MSM_BO_WC
) {
152 vma
->vm_page_prot
= pgprot_writecombine(vm_get_page_prot(vma
->vm_flags
));
153 } else if (msm_obj
->flags
& MSM_BO_UNCACHED
) {
154 vma
->vm_page_prot
= pgprot_noncached(vm_get_page_prot(vma
->vm_flags
));
157 * Shunt off cached objs to shmem file so they have their own
158 * address_space (so unmap_mapping_range does what we want,
159 * in particular in the case of mmap'd dmabufs)
164 vma
->vm_file
= obj
->filp
;
166 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
172 int msm_gem_mmap(struct file
*filp
, struct vm_area_struct
*vma
)
176 ret
= drm_gem_mmap(filp
, vma
);
178 DBG("mmap failed: %d", ret
);
182 return msm_gem_mmap_obj(vma
->vm_private_data
, vma
);
185 int msm_gem_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
187 struct drm_gem_object
*obj
= vma
->vm_private_data
;
188 struct drm_device
*dev
= obj
->dev
;
194 /* Make sure we don't parallel update on a fault, nor move or remove
195 * something from beneath our feet
197 ret
= mutex_lock_interruptible(&dev
->struct_mutex
);
201 /* make sure we have pages attached now */
202 pages
= get_pages(obj
);
204 ret
= PTR_ERR(pages
);
208 /* We don't use vmf->pgoff since that has the fake offset: */
209 pgoff
= ((unsigned long)vmf
->virtual_address
-
210 vma
->vm_start
) >> PAGE_SHIFT
;
212 pfn
= page_to_pfn(pages
[pgoff
]);
214 VERB("Inserting %p pfn %lx, pa %lx", vmf
->virtual_address
,
215 pfn
, pfn
<< PAGE_SHIFT
);
217 ret
= vm_insert_mixed(vma
, (unsigned long)vmf
->virtual_address
, pfn
);
220 mutex_unlock(&dev
->struct_mutex
);
229 * EBUSY is ok: this just means that another thread
230 * already did the job.
232 return VM_FAULT_NOPAGE
;
236 return VM_FAULT_SIGBUS
;
240 /** get mmap offset */
241 static uint64_t mmap_offset(struct drm_gem_object
*obj
)
243 struct drm_device
*dev
= obj
->dev
;
246 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
248 /* Make it mmapable */
249 ret
= drm_gem_create_mmap_offset(obj
);
252 dev_err(dev
->dev
, "could not allocate mmap offset\n");
256 return drm_vma_node_offset_addr(&obj
->vma_node
);
259 uint64_t msm_gem_mmap_offset(struct drm_gem_object
*obj
)
262 mutex_lock(&obj
->dev
->struct_mutex
);
263 offset
= mmap_offset(obj
);
264 mutex_unlock(&obj
->dev
->struct_mutex
);
268 /* should be called under struct_mutex.. although it can be called
269 * from atomic context without struct_mutex to acquire an extra
270 * iova ref if you know one is already held.
272 * That means when I do eventually need to add support for unpinning
273 * the refcnt counter needs to be atomic_t.
275 int msm_gem_get_iova_locked(struct drm_gem_object
*obj
, int id
,
278 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
281 if (!msm_obj
->domain
[id
].iova
) {
282 struct msm_drm_private
*priv
= obj
->dev
->dev_private
;
283 struct msm_mmu
*mmu
= priv
->mmus
[id
];
284 struct page
**pages
= get_pages(obj
);
287 return PTR_ERR(pages
);
289 if (iommu_present(&platform_bus_type
)) {
290 uint32_t offset
= (uint32_t)mmap_offset(obj
);
291 ret
= mmu
->funcs
->map(mmu
, offset
, msm_obj
->sgt
,
292 obj
->size
, IOMMU_READ
| IOMMU_WRITE
);
293 msm_obj
->domain
[id
].iova
= offset
;
295 msm_obj
->domain
[id
].iova
= physaddr(obj
);
300 *iova
= msm_obj
->domain
[id
].iova
;
305 int msm_gem_get_iova(struct drm_gem_object
*obj
, int id
, uint32_t *iova
)
307 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
310 /* this is safe right now because we don't unmap until the
313 if (msm_obj
->domain
[id
].iova
) {
314 *iova
= msm_obj
->domain
[id
].iova
;
318 mutex_lock(&obj
->dev
->struct_mutex
);
319 ret
= msm_gem_get_iova_locked(obj
, id
, iova
);
320 mutex_unlock(&obj
->dev
->struct_mutex
);
324 void msm_gem_put_iova(struct drm_gem_object
*obj
, int id
)
327 // NOTE: probably don't need a _locked() version.. we wouldn't
328 // normally unmap here, but instead just mark that it could be
329 // unmapped (if the iova refcnt drops to zero), but then later
330 // if another _get_iova_locked() fails we can start unmapping
331 // things that are no longer needed..
334 int msm_gem_dumb_create(struct drm_file
*file
, struct drm_device
*dev
,
335 struct drm_mode_create_dumb
*args
)
337 args
->pitch
= align_pitch(args
->width
, args
->bpp
);
338 args
->size
= PAGE_ALIGN(args
->pitch
* args
->height
);
339 return msm_gem_new_handle(dev
, file
, args
->size
,
340 MSM_BO_SCANOUT
| MSM_BO_WC
, &args
->handle
);
343 int msm_gem_dumb_map_offset(struct drm_file
*file
, struct drm_device
*dev
,
344 uint32_t handle
, uint64_t *offset
)
346 struct drm_gem_object
*obj
;
349 /* GEM does all our handle to object mapping */
350 obj
= drm_gem_object_lookup(dev
, file
, handle
);
356 *offset
= msm_gem_mmap_offset(obj
);
358 drm_gem_object_unreference_unlocked(obj
);
364 void *msm_gem_vaddr_locked(struct drm_gem_object
*obj
)
366 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
367 WARN_ON(!mutex_is_locked(&obj
->dev
->struct_mutex
));
368 if (!msm_obj
->vaddr
) {
369 struct page
**pages
= get_pages(obj
);
371 return ERR_CAST(pages
);
372 msm_obj
->vaddr
= vmap(pages
, obj
->size
>> PAGE_SHIFT
,
373 VM_MAP
, pgprot_writecombine(PAGE_KERNEL
));
375 return msm_obj
->vaddr
;
378 void *msm_gem_vaddr(struct drm_gem_object
*obj
)
381 mutex_lock(&obj
->dev
->struct_mutex
);
382 ret
= msm_gem_vaddr_locked(obj
);
383 mutex_unlock(&obj
->dev
->struct_mutex
);
387 /* setup callback for when bo is no longer busy..
388 * TODO probably want to differentiate read vs write..
390 int msm_gem_queue_inactive_cb(struct drm_gem_object
*obj
,
391 struct msm_fence_cb
*cb
)
393 struct drm_device
*dev
= obj
->dev
;
394 struct msm_drm_private
*priv
= dev
->dev_private
;
395 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
398 mutex_lock(&dev
->struct_mutex
);
399 if (!list_empty(&cb
->work
.entry
)) {
401 } else if (is_active(msm_obj
)) {
402 cb
->fence
= max(msm_obj
->read_fence
, msm_obj
->write_fence
);
403 list_add_tail(&cb
->work
.entry
, &priv
->fence_cbs
);
405 queue_work(priv
->wq
, &cb
->work
);
407 mutex_unlock(&dev
->struct_mutex
);
412 void msm_gem_move_to_active(struct drm_gem_object
*obj
,
413 struct msm_gpu
*gpu
, bool write
, uint32_t fence
)
415 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
418 msm_obj
->write_fence
= fence
;
420 msm_obj
->read_fence
= fence
;
421 list_del_init(&msm_obj
->mm_list
);
422 list_add_tail(&msm_obj
->mm_list
, &gpu
->active_list
);
425 void msm_gem_move_to_inactive(struct drm_gem_object
*obj
)
427 struct drm_device
*dev
= obj
->dev
;
428 struct msm_drm_private
*priv
= dev
->dev_private
;
429 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
431 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
434 msm_obj
->read_fence
= 0;
435 msm_obj
->write_fence
= 0;
436 list_del_init(&msm_obj
->mm_list
);
437 list_add_tail(&msm_obj
->mm_list
, &priv
->inactive_list
);
440 int msm_gem_cpu_prep(struct drm_gem_object
*obj
, uint32_t op
,
441 struct timespec
*timeout
)
443 struct drm_device
*dev
= obj
->dev
;
444 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
447 if (is_active(msm_obj
)) {
450 if (op
& MSM_PREP_READ
)
451 fence
= msm_obj
->write_fence
;
452 if (op
& MSM_PREP_WRITE
)
453 fence
= max(fence
, msm_obj
->read_fence
);
454 if (op
& MSM_PREP_NOSYNC
)
457 ret
= msm_wait_fence_interruptable(dev
, fence
, timeout
);
460 /* TODO cache maintenance */
465 int msm_gem_cpu_fini(struct drm_gem_object
*obj
)
467 /* TODO cache maintenance */
471 #ifdef CONFIG_DEBUG_FS
472 void msm_gem_describe(struct drm_gem_object
*obj
, struct seq_file
*m
)
474 struct drm_device
*dev
= obj
->dev
;
475 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
476 uint64_t off
= drm_vma_node_start(&obj
->vma_node
);
478 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
479 seq_printf(m
, "%08x: %c(r=%u,w=%u) %2d (%2d) %08llx %p %d\n",
480 msm_obj
->flags
, is_active(msm_obj
) ? 'A' : 'I',
481 msm_obj
->read_fence
, msm_obj
->write_fence
,
482 obj
->name
, obj
->refcount
.refcount
.counter
,
483 off
, msm_obj
->vaddr
, obj
->size
);
486 void msm_gem_describe_objects(struct list_head
*list
, struct seq_file
*m
)
488 struct msm_gem_object
*msm_obj
;
492 list_for_each_entry(msm_obj
, list
, mm_list
) {
493 struct drm_gem_object
*obj
= &msm_obj
->base
;
495 msm_gem_describe(obj
, m
);
500 seq_printf(m
, "Total %d objects, %zu bytes\n", count
, size
);
504 void msm_gem_free_object(struct drm_gem_object
*obj
)
506 struct drm_device
*dev
= obj
->dev
;
507 struct msm_drm_private
*priv
= obj
->dev
->dev_private
;
508 struct msm_gem_object
*msm_obj
= to_msm_bo(obj
);
511 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
513 /* object should not be on active list: */
514 WARN_ON(is_active(msm_obj
));
516 list_del(&msm_obj
->mm_list
);
518 for (id
= 0; id
< ARRAY_SIZE(msm_obj
->domain
); id
++) {
519 struct msm_mmu
*mmu
= priv
->mmus
[id
];
520 if (mmu
&& msm_obj
->domain
[id
].iova
) {
521 uint32_t offset
= (uint32_t)mmap_offset(obj
);
522 mmu
->funcs
->unmap(mmu
, offset
, msm_obj
->sgt
, obj
->size
);
526 drm_gem_free_mmap_offset(obj
);
528 if (obj
->import_attach
) {
530 dma_buf_vunmap(obj
->import_attach
->dmabuf
, msm_obj
->vaddr
);
532 /* Don't drop the pages for imported dmabuf, as they are not
533 * ours, just free the array we allocated:
536 drm_free_large(msm_obj
->pages
);
540 vunmap(msm_obj
->vaddr
);
544 if (msm_obj
->resv
== &msm_obj
->_resv
)
545 reservation_object_fini(msm_obj
->resv
);
547 drm_gem_object_release(obj
);
552 /* convenience method to construct a GEM buffer object, and userspace handle */
553 int msm_gem_new_handle(struct drm_device
*dev
, struct drm_file
*file
,
554 uint32_t size
, uint32_t flags
, uint32_t *handle
)
556 struct drm_gem_object
*obj
;
559 ret
= mutex_lock_interruptible(&dev
->struct_mutex
);
563 obj
= msm_gem_new(dev
, size
, flags
);
565 mutex_unlock(&dev
->struct_mutex
);
570 ret
= drm_gem_handle_create(file
, obj
, handle
);
572 /* drop reference from allocate - handle holds it now */
573 drm_gem_object_unreference_unlocked(obj
);
578 static int msm_gem_new_impl(struct drm_device
*dev
,
579 uint32_t size
, uint32_t flags
,
580 struct drm_gem_object
**obj
)
582 struct msm_drm_private
*priv
= dev
->dev_private
;
583 struct msm_gem_object
*msm_obj
;
586 switch (flags
& MSM_BO_CACHE_MASK
) {
587 case MSM_BO_UNCACHED
:
592 dev_err(dev
->dev
, "invalid cache flag: %x\n",
593 (flags
& MSM_BO_CACHE_MASK
));
597 sz
= sizeof(*msm_obj
);
598 if (!iommu_present(&platform_bus_type
))
599 sz
+= sizeof(struct drm_mm_node
);
601 msm_obj
= kzalloc(sz
, GFP_KERNEL
);
605 if (!iommu_present(&platform_bus_type
))
606 msm_obj
->vram_node
= (void *)&msm_obj
[1];
608 msm_obj
->flags
= flags
;
610 msm_obj
->resv
= &msm_obj
->_resv
;
611 reservation_object_init(msm_obj
->resv
);
613 INIT_LIST_HEAD(&msm_obj
->submit_entry
);
614 list_add_tail(&msm_obj
->mm_list
, &priv
->inactive_list
);
616 *obj
= &msm_obj
->base
;
621 struct drm_gem_object
*msm_gem_new(struct drm_device
*dev
,
622 uint32_t size
, uint32_t flags
)
624 struct drm_gem_object
*obj
= NULL
;
627 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
629 size
= PAGE_ALIGN(size
);
631 ret
= msm_gem_new_impl(dev
, size
, flags
, &obj
);
635 if (iommu_present(&platform_bus_type
)) {
636 ret
= drm_gem_object_init(dev
, obj
, size
);
640 drm_gem_private_object_init(dev
, obj
, size
);
647 drm_gem_object_unreference(obj
);
652 struct drm_gem_object
*msm_gem_import(struct drm_device
*dev
,
653 uint32_t size
, struct sg_table
*sgt
)
655 struct msm_gem_object
*msm_obj
;
656 struct drm_gem_object
*obj
;
659 /* if we don't have IOMMU, don't bother pretending we can import: */
660 if (!iommu_present(&platform_bus_type
)) {
661 dev_err(dev
->dev
, "cannot import without IOMMU\n");
662 return ERR_PTR(-EINVAL
);
665 size
= PAGE_ALIGN(size
);
667 ret
= msm_gem_new_impl(dev
, size
, MSM_BO_WC
, &obj
);
671 drm_gem_private_object_init(dev
, obj
, size
);
673 npages
= size
/ PAGE_SIZE
;
675 msm_obj
= to_msm_bo(obj
);
677 msm_obj
->pages
= drm_malloc_ab(npages
, sizeof(struct page
*));
678 if (!msm_obj
->pages
) {
683 ret
= drm_prime_sg_to_page_addr_arrays(sgt
, msm_obj
->pages
, NULL
, npages
);
691 drm_gem_object_unreference_unlocked(obj
);