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>
27 * GTT resource allocator - manage page mappings in GTT space
31 * psb_gtt_mask_pte - generate GTT pte entry
32 * @pfn: page number to encode
33 * @type: type of memory in the GTT
35 * Set the GTT entry for the appropriate memory type.
37 static inline uint32_t psb_gtt_mask_pte(uint32_t pfn
, int type
)
39 uint32_t mask
= PSB_PTE_VALID
;
41 if (type
& PSB_MMU_CACHED_MEMORY
)
42 mask
|= PSB_PTE_CACHED
;
43 if (type
& PSB_MMU_RO_MEMORY
)
45 if (type
& PSB_MMU_WO_MEMORY
)
48 return (pfn
<< PAGE_SHIFT
) | mask
;
52 * psb_gtt_entry - find the GTT entries for a gtt_range
53 * @dev: our DRM device
56 * Given a gtt_range object return the GTT offset of the page table
57 * entries for this gtt_range
59 u32
*psb_gtt_entry(struct drm_device
*dev
, struct gtt_range
*r
)
61 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
64 offset
= r
->resource
.start
- dev_priv
->gtt_mem
->start
;
66 return dev_priv
->gtt_map
+ (offset
>> PAGE_SHIFT
);
70 * psb_gtt_insert - put an object into the GTT
71 * @dev: our DRM device
74 * Take our preallocated GTT range and insert the GEM object into
75 * the GTT. This is protected via the gtt mutex which the caller
78 static int psb_gtt_insert(struct drm_device
*dev
, struct gtt_range
*r
)
84 if (r
->pages
== NULL
) {
89 WARN_ON(r
->stolen
); /* refcount these maybe ? */
91 gtt_slot
= psb_gtt_entry(dev
, r
);
94 /* Make sure changes are visible to the GPU */
95 set_pages_array_uc(pages
, r
->npage
);
97 /* Write our page entries into the GTT itself */
98 for (i
= r
->roll
; i
< r
->npage
; i
++) {
99 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
100 iowrite32(pte
, gtt_slot
++);
102 for (i
= 0; i
< r
->roll
; i
++) {
103 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
104 iowrite32(pte
, gtt_slot
++);
106 /* Make sure all the entries are set before we return */
107 ioread32(gtt_slot
- 1);
113 * psb_gtt_remove - remove an object from the GTT
114 * @dev: our DRM device
117 * Remove a preallocated GTT range from the GTT. Overwrite all the
118 * page table entries with the dummy page. This is protected via the gtt
119 * mutex which the caller must hold.
121 static void psb_gtt_remove(struct drm_device
*dev
, struct gtt_range
*r
)
123 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
129 gtt_slot
= psb_gtt_entry(dev
, r
);
130 pte
= psb_gtt_mask_pte(page_to_pfn(dev_priv
->scratch_page
), 0);
132 for (i
= 0; i
< r
->npage
; i
++)
133 iowrite32(pte
, gtt_slot
++);
134 ioread32(gtt_slot
- 1);
135 set_pages_array_wb(r
->pages
, r
->npage
);
139 * psb_gtt_roll - set scrolling position
140 * @dev: our DRM device
141 * @r: the gtt mapping we are using
144 * Roll an existing pinned mapping by moving the pages through the GTT.
145 * This allows us to implement hardware scrolling on the consoles without
148 void psb_gtt_roll(struct drm_device
*dev
, struct gtt_range
*r
, int roll
)
153 if (roll
>= r
->npage
) {
160 /* Not currently in the GTT - no worry we will write the mapping at
161 the right position when it gets pinned */
162 if (!r
->stolen
&& !r
->in_gart
)
165 gtt_slot
= psb_gtt_entry(dev
, r
);
167 for (i
= r
->roll
; i
< r
->npage
; i
++) {
168 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
169 iowrite32(pte
, gtt_slot
++);
171 for (i
= 0; i
< r
->roll
; i
++) {
172 pte
= psb_gtt_mask_pte(page_to_pfn(r
->pages
[i
]), 0);
173 iowrite32(pte
, gtt_slot
++);
175 ioread32(gtt_slot
- 1);
179 * psb_gtt_attach_pages - attach and pin GEM pages
182 * Pin and build an in kernel list of the pages that back our GEM object.
183 * While we hold this the pages cannot be swapped out. This is protected
184 * via the gtt mutex which the caller must hold.
186 static int psb_gtt_attach_pages(struct gtt_range
*gt
)
189 struct address_space
*mapping
;
192 int pages
= gt
->gem
.size
/ PAGE_SIZE
;
196 /* This is the shared memory object that backs the GEM resource */
197 inode
= gt
->gem
.filp
->f_path
.dentry
->d_inode
;
198 mapping
= inode
->i_mapping
;
200 gt
->pages
= kmalloc(pages
* sizeof(struct page
*), GFP_KERNEL
);
201 if (gt
->pages
== NULL
)
205 for (i
= 0; i
< pages
; i
++) {
206 /* FIXME: needs updating as per mail from Hugh Dickins */
207 p
= read_cache_page_gfp(mapping
, i
,
208 __GFP_COLD
| GFP_KERNEL
);
217 page_cache_release(gt
->pages
[i
]);
224 * psb_gtt_detach_pages - attach and pin GEM pages
227 * Undo the effect of psb_gtt_attach_pages. At this point the pages
228 * must have been removed from the GTT as they could now be paged out
229 * and move bus address. This is protected via the gtt mutex which the
232 static void psb_gtt_detach_pages(struct gtt_range
*gt
)
235 for (i
= 0; i
< gt
->npage
; i
++) {
236 /* FIXME: do we need to force dirty */
237 set_page_dirty(gt
->pages
[i
]);
238 page_cache_release(gt
->pages
[i
]);
245 * psb_gtt_pin - pin pages into the GTT
248 * Pin a set of pages into the GTT. The pins are refcounted so that
249 * multiple pins need multiple unpins to undo.
251 * Non GEM backed objects treat this as a no-op as they are always GTT
254 int psb_gtt_pin(struct gtt_range
*gt
)
257 struct drm_device
*dev
= gt
->gem
.dev
;
258 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
260 mutex_lock(&dev_priv
->gtt_mutex
);
262 if (gt
->in_gart
== 0 && gt
->stolen
== 0) {
263 ret
= psb_gtt_attach_pages(gt
);
266 ret
= psb_gtt_insert(dev
, gt
);
268 psb_gtt_detach_pages(gt
);
274 mutex_unlock(&dev_priv
->gtt_mutex
);
279 * psb_gtt_unpin - Drop a GTT pin requirement
282 * Undoes the effect of psb_gtt_pin. On the last drop the GEM object
283 * will be removed from the GTT which will also drop the page references
284 * and allow the VM to clean up or page stuff.
286 * Non GEM backed objects treat this as a no-op as they are always GTT
289 void psb_gtt_unpin(struct gtt_range
*gt
)
291 struct drm_device
*dev
= gt
->gem
.dev
;
292 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
294 mutex_lock(&dev_priv
->gtt_mutex
);
296 WARN_ON(!gt
->in_gart
);
299 if (gt
->in_gart
== 0 && gt
->stolen
== 0) {
300 psb_gtt_remove(dev
, gt
);
301 psb_gtt_detach_pages(gt
);
303 mutex_unlock(&dev_priv
->gtt_mutex
);
307 * GTT resource allocator - allocate and manage GTT address space
311 * psb_gtt_alloc_range - allocate GTT address space
312 * @dev: Our DRM device
313 * @len: length (bytes) of address space required
314 * @name: resource name
315 * @backed: resource should be backed by stolen pages
317 * Ask the kernel core to find us a suitable range of addresses
318 * to use for a GTT mapping.
320 * Returns a gtt_range structure describing the object, or NULL on
321 * error. On successful return the resource is both allocated and marked
324 struct gtt_range
*psb_gtt_alloc_range(struct drm_device
*dev
, int len
,
325 const char *name
, int backed
)
327 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
328 struct gtt_range
*gt
;
329 struct resource
*r
= dev_priv
->gtt_mem
;
331 unsigned long start
, end
;
334 /* The start of the GTT is the stolen pages */
336 end
= r
->start
+ dev_priv
->gtt
.stolen_size
- 1;
338 /* The rest we will use for GEM backed objects */
339 start
= r
->start
+ dev_priv
->gtt
.stolen_size
;
343 gt
= kzalloc(sizeof(struct gtt_range
), GFP_KERNEL
);
346 gt
->resource
.name
= name
;
348 gt
->in_gart
= backed
;
350 /* Ensure this is set for non GEM objects */
352 ret
= allocate_resource(dev_priv
->gtt_mem
, >
->resource
,
353 len
, start
, end
, PAGE_SIZE
, NULL
, NULL
);
355 gt
->offset
= gt
->resource
.start
- r
->start
;
363 * psb_gtt_free_range - release GTT address space
364 * @dev: our DRM device
365 * @gt: a mapping created with psb_gtt_alloc_range
367 * Release a resource that was allocated with psb_gtt_alloc_range. If the
368 * object has been pinned by mmap users we clean this up here currently.
370 void psb_gtt_free_range(struct drm_device
*dev
, struct gtt_range
*gt
)
372 /* Undo the mmap pin if we are destroying the object */
377 WARN_ON(gt
->in_gart
&& !gt
->stolen
);
378 release_resource(>
->resource
);
382 void psb_gtt_alloc(struct drm_device
*dev
)
384 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
385 init_rwsem(&dev_priv
->gtt
.sem
);
388 void psb_gtt_takedown(struct drm_device
*dev
)
390 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
392 if (dev_priv
->gtt_map
) {
393 iounmap(dev_priv
->gtt_map
);
394 dev_priv
->gtt_map
= NULL
;
396 if (dev_priv
->gtt_initialized
) {
397 pci_write_config_word(dev
->pdev
, PSB_GMCH_CTRL
,
398 dev_priv
->gmch_ctrl
);
399 PSB_WVDC32(dev_priv
->pge_ctl
, PSB_PGETBL_CTL
);
400 (void) PSB_RVDC32(PSB_PGETBL_CTL
);
402 if (dev_priv
->vram_addr
)
403 iounmap(dev_priv
->gtt_map
);
406 int psb_gtt_init(struct drm_device
*dev
, int resume
)
408 struct drm_psb_private
*dev_priv
= dev
->dev_private
;
410 unsigned long stolen_size
, vram_stolen_size
;
411 unsigned i
, num_pages
;
414 uint32_t dvmt_mode
= 0;
420 mutex_init(&dev_priv
->gtt_mutex
);
426 pci_read_config_word(dev
->pdev
, PSB_GMCH_CTRL
, &dev_priv
->gmch_ctrl
);
427 pci_write_config_word(dev
->pdev
, PSB_GMCH_CTRL
,
428 dev_priv
->gmch_ctrl
| _PSB_GMCH_ENABLED
);
430 dev_priv
->pge_ctl
= PSB_RVDC32(PSB_PGETBL_CTL
);
431 PSB_WVDC32(dev_priv
->pge_ctl
| _PSB_PGETBL_ENABLED
, PSB_PGETBL_CTL
);
432 (void) PSB_RVDC32(PSB_PGETBL_CTL
);
434 /* The root resource we allocate address space from */
435 dev_priv
->gtt_initialized
= 1;
437 pg
->gtt_phys_start
= dev_priv
->pge_ctl
& PAGE_MASK
;
440 * The video mmu has a hw bug when accessing 0x0D0000000.
441 * Make gatt start at 0x0e000,0000. This doesn't actually
442 * matter for us but may do if the video acceleration ever
445 pg
->mmu_gatt_start
= 0xE0000000;
447 pg
->gtt_start
= pci_resource_start(dev
->pdev
, PSB_GTT_RESOURCE
);
448 gtt_pages
= pci_resource_len(dev
->pdev
, PSB_GTT_RESOURCE
)
450 /* Some CDV firmware doesn't report this currently. In which case the
451 system has 64 gtt pages */
452 if (pg
->gtt_start
== 0 || gtt_pages
== 0) {
453 dev_err(dev
->dev
, "GTT PCI BAR not initialized.\n");
455 pg
->gtt_start
= dev_priv
->pge_ctl
;
458 pg
->gatt_start
= pci_resource_start(dev
->pdev
, PSB_GATT_RESOURCE
);
459 pg
->gatt_pages
= pci_resource_len(dev
->pdev
, PSB_GATT_RESOURCE
)
461 dev_priv
->gtt_mem
= &dev
->pdev
->resource
[PSB_GATT_RESOURCE
];
463 if (pg
->gatt_pages
== 0 || pg
->gatt_start
== 0) {
464 static struct resource fudge
; /* Preferably peppermint */
465 /* This can occur on CDV SDV systems. Fudge it in this case.
466 We really don't care what imaginary space is being allocated
468 dev_err(dev
->dev
, "GATT PCI BAR not initialized.\n");
469 pg
->gatt_start
= 0x40000000;
470 pg
->gatt_pages
= (128 * 1024 * 1024) >> PAGE_SHIFT
;
471 /* This is a little confusing but in fact the GTT is providing
472 a view from the GPU into memory and not vice versa. As such
473 this is really allocating space that is not the same as the
474 CPU address space on CDV */
475 fudge
.start
= 0x40000000;
476 fudge
.end
= 0x40000000 + 128 * 1024 * 1024 - 1;
477 fudge
.name
= "fudge";
478 fudge
.flags
= IORESOURCE_MEM
;
479 dev_priv
->gtt_mem
= &fudge
;
482 pci_read_config_dword(dev
->pdev
, PSB_BSM
, &dev_priv
->stolen_base
);
483 vram_stolen_size
= pg
->gtt_phys_start
- dev_priv
->stolen_base
486 stolen_size
= vram_stolen_size
;
488 printk(KERN_INFO
"Stolen memory information\n");
489 printk(KERN_INFO
" base in RAM: 0x%x\n", dev_priv
->stolen_base
);
490 printk(KERN_INFO
" size: %luK, calculated by (GTT RAM base) - (Stolen base), seems wrong\n",
491 vram_stolen_size
/1024);
492 dvmt_mode
= (dev_priv
->gmch_ctrl
>> 4) & 0x7;
493 printk(KERN_INFO
" the correct size should be: %dM(dvmt mode=%d)\n",
494 (dvmt_mode
== 1) ? 1 : (2 << (dvmt_mode
- 1)), dvmt_mode
);
496 if (resume
&& (gtt_pages
!= pg
->gtt_pages
) &&
497 (stolen_size
!= pg
->stolen_size
)) {
498 dev_err(dev
->dev
, "GTT resume error.\n");
503 pg
->gtt_pages
= gtt_pages
;
504 pg
->stolen_size
= stolen_size
;
505 dev_priv
->vram_stolen_size
= vram_stolen_size
;
508 * Map the GTT and the stolen memory area
510 dev_priv
->gtt_map
= ioremap_nocache(pg
->gtt_phys_start
,
511 gtt_pages
<< PAGE_SHIFT
);
512 if (!dev_priv
->gtt_map
) {
513 dev_err(dev
->dev
, "Failure to map gtt.\n");
518 dev_priv
->vram_addr
= ioremap_wc(dev_priv
->stolen_base
, stolen_size
);
519 if (!dev_priv
->vram_addr
) {
520 dev_err(dev
->dev
, "Failure to map stolen base.\n");
526 * Insert vram stolen pages into the GTT
529 pfn_base
= dev_priv
->stolen_base
>> PAGE_SHIFT
;
530 vram_pages
= num_pages
= vram_stolen_size
>> PAGE_SHIFT
;
531 printk(KERN_INFO
"Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n",
532 num_pages
, pfn_base
<< PAGE_SHIFT
, 0);
533 for (i
= 0; i
< num_pages
; ++i
) {
534 pte
= psb_gtt_mask_pte(pfn_base
+ i
, 0);
535 iowrite32(pte
, dev_priv
->gtt_map
+ i
);
539 * Init rest of GTT to the scratch page to avoid accidents or scribbles
542 pfn_base
= page_to_pfn(dev_priv
->scratch_page
);
543 pte
= psb_gtt_mask_pte(pfn_base
, 0);
544 for (; i
< gtt_pages
; ++i
)
545 iowrite32(pte
, dev_priv
->gtt_map
+ i
);
547 (void) ioread32(dev_priv
->gtt_map
+ i
- 1);
551 psb_gtt_takedown(dev
);