2 * Copyright (c) 2007, Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope 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, write to the Free Software Foundation, Inc.,
16 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
18 * Authors: Thomas Hellstrom <thomas-at-tungstengraphics.com>
19 * Alan Cox <alan@linux.intel.com>
23 #include <linux/shmem_fs.h>
28 * GTT resource allocator - manage page mappings in GTT space
32 * psb_gtt_mask_pte - generate GTT pte entry
33 * @pfn: page number to encode
34 * @type: type of memory in the GTT
36 * Set the GTT entry for the appropriate memory type.
38 static inline uint32_t psb_gtt_mask_pte(uint32_t pfn
, int type
)
40 uint32_t mask
= PSB_PTE_VALID
;
42 /* Ensure we explode rather than put an invalid low mapping of
43 a high mapping page into the gtt */
44 BUG_ON(pfn
& ~(0xFFFFFFFF >> PAGE_SHIFT
));
46 if (type
& PSB_MMU_CACHED_MEMORY
)
47 mask
|= PSB_PTE_CACHED
;
48 if (type
& PSB_MMU_RO_MEMORY
)
50 if (type
& PSB_MMU_WO_MEMORY
)
53 return (pfn
<< PAGE_SHIFT
) | mask
;
57 * psb_gtt_entry - find the GTT entries for a gtt_range
58 * @dev: our DRM device
61 * Given a gtt_range object return the GTT offset of the page table
62 * entries for this gtt_range
64 static u32 __iomem
*psb_gtt_entry(struct drm_device
*dev
, struct gtt_range
*r
)
66 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
69 offset
= r
->resource
.start
- dev_priv
->gtt_mem
->start
;
71 return dev_priv
->gtt_map
+ (offset
>> PAGE_SHIFT
);
75 * psb_gtt_insert - put an object into the GTT
76 * @dev: our DRM device
79 * Take our preallocated GTT range and insert the GEM object into
80 * the GTT. This is protected via the gtt mutex which the caller
83 static int psb_gtt_insert(struct drm_device
*dev
, struct gtt_range
*r
)
85 u32 __iomem
*gtt_slot
;
90 if (r
->pages
== NULL
) {
95 WARN_ON(r
->stolen
); /* refcount these maybe ? */
97 gtt_slot
= psb_gtt_entry(dev
, r
);
100 /* Make sure changes are visible to the GPU */
101 set_pages_array_wc(pages
, r
->npage
);
103 /* Write our page entries into the GTT itself */
104 for (i
= r
->roll
; i
< r
->npage
; i
++) {
105 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
106 iowrite32(pte
, gtt_slot
++);
108 for (i
= 0; i
< r
->roll
; i
++) {
109 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
110 iowrite32(pte
, gtt_slot
++);
112 /* Make sure all the entries are set before we return */
113 ioread32(gtt_slot
- 1);
119 * psb_gtt_remove - remove an object from the GTT
120 * @dev: our DRM device
123 * Remove a preallocated GTT range from the GTT. Overwrite all the
124 * page table entries with the dummy page. This is protected via the gtt
125 * mutex which the caller must hold.
127 static void psb_gtt_remove(struct drm_device
*dev
, struct gtt_range
*r
)
129 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
130 u32 __iomem
*gtt_slot
;
136 gtt_slot
= psb_gtt_entry(dev
, r
);
137 pte
= psb_gtt_mask_pte(page_to_pfn(dev_priv
->scratch_page
), 0);
139 for (i
= 0; i
< r
->npage
; i
++)
140 iowrite32(pte
, gtt_slot
++);
141 ioread32(gtt_slot
- 1);
142 set_pages_array_wb(r
->pages
, r
->npage
);
146 * psb_gtt_roll - set scrolling position
147 * @dev: our DRM device
148 * @r: the gtt mapping we are using
151 * Roll an existing pinned mapping by moving the pages through the GTT.
152 * This allows us to implement hardware scrolling on the consoles without
155 void psb_gtt_roll(struct drm_device
*dev
, struct gtt_range
*r
, int roll
)
157 u32 __iomem
*gtt_slot
;
161 if (roll
>= r
->npage
) {
168 /* Not currently in the GTT - no worry we will write the mapping at
169 the right position when it gets pinned */
170 if (!r
->stolen
&& !r
->in_gart
)
173 gtt_slot
= psb_gtt_entry(dev
, r
);
175 for (i
= r
->roll
; i
< r
->npage
; i
++) {
176 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
177 iowrite32(pte
, gtt_slot
++);
179 for (i
= 0; i
< r
->roll
; i
++) {
180 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
181 iowrite32(pte
, gtt_slot
++);
183 ioread32(gtt_slot
- 1);
187 * psb_gtt_attach_pages - attach and pin GEM pages
190 * Pin and build an in kernel list of the pages that back our GEM object.
191 * While we hold this the pages cannot be swapped out. This is protected
192 * via the gtt mutex which the caller must hold.
194 static int psb_gtt_attach_pages(struct gtt_range
*gt
)
197 struct address_space
*mapping
;
200 int pages
= gt
->gem
.size
/ PAGE_SIZE
;
204 /* This is the shared memory object that backs the GEM resource */
205 inode
= gt
->gem
.filp
->f_path
.dentry
->d_inode
;
206 mapping
= inode
->i_mapping
;
208 gt
->pages
= kmalloc(pages
* sizeof(struct page
*), GFP_KERNEL
);
209 if (gt
->pages
== NULL
)
213 for (i
= 0; i
< pages
; i
++) {
214 p
= shmem_read_mapping_page(mapping
, i
);
223 page_cache_release(gt
->pages
[i
]);
230 * psb_gtt_detach_pages - attach and pin GEM pages
233 * Undo the effect of psb_gtt_attach_pages. At this point the pages
234 * must have been removed from the GTT as they could now be paged out
235 * and move bus address. This is protected via the gtt mutex which the
238 static void psb_gtt_detach_pages(struct gtt_range
*gt
)
241 for (i
= 0; i
< gt
->npage
; i
++) {
242 /* FIXME: do we need to force dirty */
243 set_page_dirty(gt
->pages
[i
]);
244 page_cache_release(gt
->pages
[i
]);
251 * psb_gtt_pin - pin pages into the GTT
254 * Pin a set of pages into the GTT. The pins are refcounted so that
255 * multiple pins need multiple unpins to undo.
257 * Non GEM backed objects treat this as a no-op as they are always GTT
260 int psb_gtt_pin(struct gtt_range
*gt
)
263 struct drm_device
*dev
= gt
->gem
.dev
;
264 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
266 mutex_lock(&dev_priv
->gtt_mutex
);
268 if (gt
->in_gart
== 0 && gt
->stolen
== 0) {
269 ret
= psb_gtt_attach_pages(gt
);
272 ret
= psb_gtt_insert(dev
, gt
);
274 psb_gtt_detach_pages(gt
);
280 mutex_unlock(&dev_priv
->gtt_mutex
);
285 * psb_gtt_unpin - Drop a GTT pin requirement
288 * Undoes the effect of psb_gtt_pin. On the last drop the GEM object
289 * will be removed from the GTT which will also drop the page references
290 * and allow the VM to clean up or page stuff.
292 * Non GEM backed objects treat this as a no-op as they are always GTT
295 void psb_gtt_unpin(struct gtt_range
*gt
)
297 struct drm_device
*dev
= gt
->gem
.dev
;
298 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
300 mutex_lock(&dev_priv
->gtt_mutex
);
302 WARN_ON(!gt
->in_gart
);
305 if (gt
->in_gart
== 0 && gt
->stolen
== 0) {
306 psb_gtt_remove(dev
, gt
);
307 psb_gtt_detach_pages(gt
);
309 mutex_unlock(&dev_priv
->gtt_mutex
);
313 * GTT resource allocator - allocate and manage GTT address space
317 * psb_gtt_alloc_range - allocate GTT address space
318 * @dev: Our DRM device
319 * @len: length (bytes) of address space required
320 * @name: resource name
321 * @backed: resource should be backed by stolen pages
323 * Ask the kernel core to find us a suitable range of addresses
324 * to use for a GTT mapping.
326 * Returns a gtt_range structure describing the object, or NULL on
327 * error. On successful return the resource is both allocated and marked
330 struct gtt_range
*psb_gtt_alloc_range(struct drm_device
*dev
, int len
,
331 const char *name
, int backed
)
333 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
334 struct gtt_range
*gt
;
335 struct resource
*r
= dev_priv
->gtt_mem
;
337 unsigned long start
, end
;
340 /* The start of the GTT is the stolen pages */
342 end
= r
->start
+ dev_priv
->gtt
.stolen_size
- 1;
344 /* The rest we will use for GEM backed objects */
345 start
= r
->start
+ dev_priv
->gtt
.stolen_size
;
349 gt
= kzalloc(sizeof(struct gtt_range
), GFP_KERNEL
);
352 gt
->resource
.name
= name
;
354 gt
->in_gart
= backed
;
356 /* Ensure this is set for non GEM objects */
358 ret
= allocate_resource(dev_priv
->gtt_mem
, >
->resource
,
359 len
, start
, end
, PAGE_SIZE
, NULL
, NULL
);
361 gt
->offset
= gt
->resource
.start
- r
->start
;
369 * psb_gtt_free_range - release GTT address space
370 * @dev: our DRM device
371 * @gt: a mapping created with psb_gtt_alloc_range
373 * Release a resource that was allocated with psb_gtt_alloc_range. If the
374 * object has been pinned by mmap users we clean this up here currently.
376 void psb_gtt_free_range(struct drm_device
*dev
, struct gtt_range
*gt
)
378 /* Undo the mmap pin if we are destroying the object */
383 WARN_ON(gt
->in_gart
&& !gt
->stolen
);
384 release_resource(>
->resource
);
388 static void psb_gtt_alloc(struct drm_device
*dev
)
390 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
391 init_rwsem(&dev_priv
->gtt
.sem
);
394 void psb_gtt_takedown(struct drm_device
*dev
)
396 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
398 if (dev_priv
->gtt_map
) {
399 iounmap(dev_priv
->gtt_map
);
400 dev_priv
->gtt_map
= NULL
;
402 if (dev_priv
->gtt_initialized
) {
403 pci_write_config_word(dev
->pdev
, PSB_GMCH_CTRL
,
404 dev_priv
->gmch_ctrl
);
405 PSB_WVDC32(dev_priv
->pge_ctl
, PSB_PGETBL_CTL
);
406 (void) PSB_RVDC32(PSB_PGETBL_CTL
);
408 if (dev_priv
->vram_addr
)
409 iounmap(dev_priv
->gtt_map
);
412 int psb_gtt_init(struct drm_device
*dev
, int resume
)
414 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
416 unsigned long stolen_size
, vram_stolen_size
;
417 unsigned i
, num_pages
;
424 mutex_init(&dev_priv
->gtt_mutex
);
430 pci_read_config_word(dev
->pdev
, PSB_GMCH_CTRL
, &dev_priv
->gmch_ctrl
);
431 pci_write_config_word(dev
->pdev
, PSB_GMCH_CTRL
,
432 dev_priv
->gmch_ctrl
| _PSB_GMCH_ENABLED
);
434 dev_priv
->pge_ctl
= PSB_RVDC32(PSB_PGETBL_CTL
);
435 PSB_WVDC32(dev_priv
->pge_ctl
| _PSB_PGETBL_ENABLED
, PSB_PGETBL_CTL
);
436 (void) PSB_RVDC32(PSB_PGETBL_CTL
);
438 /* The root resource we allocate address space from */
439 dev_priv
->gtt_initialized
= 1;
441 pg
->gtt_phys_start
= dev_priv
->pge_ctl
& PAGE_MASK
;
444 * The video mmu has a hw bug when accessing 0x0D0000000.
445 * Make gatt start at 0x0e000,0000. This doesn't actually
446 * matter for us but may do if the video acceleration ever
449 pg
->mmu_gatt_start
= 0xE0000000;
451 pg
->gtt_start
= pci_resource_start(dev
->pdev
, PSB_GTT_RESOURCE
);
452 gtt_pages
= pci_resource_len(dev
->pdev
, PSB_GTT_RESOURCE
)
454 /* CDV doesn't report this. In which case the system has 64 gtt pages */
455 if (pg
->gtt_start
== 0 || gtt_pages
== 0) {
456 dev_dbg(dev
->dev
, "GTT PCI BAR not initialized.\n");
458 pg
->gtt_start
= dev_priv
->pge_ctl
;
461 pg
->gatt_start
= pci_resource_start(dev
->pdev
, PSB_GATT_RESOURCE
);
462 pg
->gatt_pages
= pci_resource_len(dev
->pdev
, PSB_GATT_RESOURCE
)
464 dev_priv
->gtt_mem
= &dev
->pdev
->resource
[PSB_GATT_RESOURCE
];
466 if (pg
->gatt_pages
== 0 || pg
->gatt_start
== 0) {
467 static struct resource fudge
; /* Preferably peppermint */
468 /* This can occur on CDV systems. Fudge it in this case.
469 We really don't care what imaginary space is being allocated
471 dev_dbg(dev
->dev
, "GATT PCI BAR not initialized.\n");
472 pg
->gatt_start
= 0x40000000;
473 pg
->gatt_pages
= (128 * 1024 * 1024) >> PAGE_SHIFT
;
474 /* This is a little confusing but in fact the GTT is providing
475 a view from the GPU into memory and not vice versa. As such
476 this is really allocating space that is not the same as the
477 CPU address space on CDV */
478 fudge
.start
= 0x40000000;
479 fudge
.end
= 0x40000000 + 128 * 1024 * 1024 - 1;
480 fudge
.name
= "fudge";
481 fudge
.flags
= IORESOURCE_MEM
;
482 dev_priv
->gtt_mem
= &fudge
;
485 pci_read_config_dword(dev
->pdev
, PSB_BSM
, &dev_priv
->stolen_base
);
486 vram_stolen_size
= pg
->gtt_phys_start
- dev_priv
->stolen_base
489 stolen_size
= vram_stolen_size
;
491 dev_dbg(dev
->dev
, "Stolen memory base 0x%x, size %luK\n",
492 dev_priv
->stolen_base
, vram_stolen_size
/ 1024);
494 if (resume
&& (gtt_pages
!= pg
->gtt_pages
) &&
495 (stolen_size
!= pg
->stolen_size
)) {
496 dev_err(dev
->dev
, "GTT resume error.\n");
501 pg
->gtt_pages
= gtt_pages
;
502 pg
->stolen_size
= stolen_size
;
503 dev_priv
->vram_stolen_size
= vram_stolen_size
;
506 * Map the GTT and the stolen memory area
508 dev_priv
->gtt_map
= ioremap_nocache(pg
->gtt_phys_start
,
509 gtt_pages
<< PAGE_SHIFT
);
510 if (!dev_priv
->gtt_map
) {
511 dev_err(dev
->dev
, "Failure to map gtt.\n");
516 dev_priv
->vram_addr
= ioremap_wc(dev_priv
->stolen_base
, stolen_size
);
517 if (!dev_priv
->vram_addr
) {
518 dev_err(dev
->dev
, "Failure to map stolen base.\n");
524 * Insert vram stolen pages into the GTT
527 pfn_base
= dev_priv
->stolen_base
>> PAGE_SHIFT
;
528 num_pages
= vram_stolen_size
>> PAGE_SHIFT
;
529 dev_dbg(dev
->dev
, "Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n",
530 num_pages
, pfn_base
<< PAGE_SHIFT
, 0);
531 for (i
= 0; i
< num_pages
; ++i
) {
532 pte
= psb_gtt_mask_pte(pfn_base
+ i
, 0);
533 iowrite32(pte
, dev_priv
->gtt_map
+ i
);
537 * Init rest of GTT to the scratch page to avoid accidents or scribbles
540 pfn_base
= page_to_pfn(dev_priv
->scratch_page
);
541 pte
= psb_gtt_mask_pte(pfn_base
, 0);
542 for (; i
< gtt_pages
; ++i
)
543 iowrite32(pte
, dev_priv
->gtt_map
+ i
);
545 (void) ioread32(dev_priv
->gtt_map
+ i
- 1);
549 psb_gtt_takedown(dev
);