1 /*
2 * Copyright (c) 2007, Intel Corporation.
3 * All Rights Reserved.
4 *
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.
8 *
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
12 * more details.
13 *
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.
17 *
18 * Authors: Thomas Hellstrom <thomas-at-tungstengraphics.com>
19 * Alan Cox <alan@linux.intel.com>
20 */
21
22 #include <drm/drmP.h>
23 #include <linux/shmem_fs.h>
24 #include "psb_drv.h"
25 #include "blitter.h"
26
27
28 /*
29 * GTT resource allocator - manage page mappings in GTT space
30 */
31
32 /**
33 * psb_gtt_mask_pte - generate GTT pte entry
34 * @pfn: page number to encode
35 * @type: type of memory in the GTT
36 *
37 * Set the GTT entry for the appropriate memory type.
38 */
psb_gtt_mask_pte(uint32_t pfn,int type)39 static inline uint32_t psb_gtt_mask_pte(uint32_t pfn, int type)
40 {
41 uint32_t mask = PSB_PTE_VALID;
42
43 /* Ensure we explode rather than put an invalid low mapping of
44 a high mapping page into the gtt */
45 BUG_ON(pfn & ~(0xFFFFFFFF >> PAGE_SHIFT));
46
47 if (type & PSB_MMU_CACHED_MEMORY)
48 mask |= PSB_PTE_CACHED;
49 if (type & PSB_MMU_RO_MEMORY)
50 mask |= PSB_PTE_RO;
51 if (type & PSB_MMU_WO_MEMORY)
52 mask |= PSB_PTE_WO;
53
54 return (pfn << PAGE_SHIFT) | mask;
55 }
56
57 /**
58 * psb_gtt_entry - find the GTT entries for a gtt_range
59 * @dev: our DRM device
60 * @r: our GTT range
61 *
62 * Given a gtt_range object return the GTT offset of the page table
63 * entries for this gtt_range
64 */
psb_gtt_entry(struct drm_device * dev,struct gtt_range * r)65 static u32 __iomem *psb_gtt_entry(struct drm_device *dev, struct gtt_range *r)
66 {
67 struct drm_psb_private *dev_priv = dev->dev_private;
68 unsigned long offset;
69
70 offset = r->resource.start - dev_priv->gtt_mem->start;
71
72 return dev_priv->gtt_map + (offset >> PAGE_SHIFT);
73 }
74
75 /**
76 * psb_gtt_insert - put an object into the GTT
77 * @dev: our DRM device
78 * @r: our GTT range
79 *
80 * Take our preallocated GTT range and insert the GEM object into
81 * the GTT. This is protected via the gtt mutex which the caller
82 * must hold.
83 */
psb_gtt_insert(struct drm_device * dev,struct gtt_range * r,int resume)84 static int psb_gtt_insert(struct drm_device *dev, struct gtt_range *r,
85 int resume)
86 {
87 u32 __iomem *gtt_slot;
88 u32 pte;
89 struct page **pages;
90 int i;
91
92 if (r->pages == NULL) {
93 WARN_ON(1);
94 return -EINVAL;
95 }
96
97 WARN_ON(r->stolen); /* refcount these maybe ? */
98
99 gtt_slot = psb_gtt_entry(dev, r);
100 pages = r->pages;
101
102 if (!resume) {
103 /* Make sure changes are visible to the GPU */
104 set_pages_array_wc(pages, r->npage);
105 }
106
107 /* Write our page entries into the GTT itself */
108 for (i = r->roll; i < r->npage; i++) {
109 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
110 PSB_MMU_CACHED_MEMORY);
111 iowrite32(pte, gtt_slot++);
112 }
113 for (i = 0; i < r->roll; i++) {
114 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
115 PSB_MMU_CACHED_MEMORY);
116 iowrite32(pte, gtt_slot++);
117 }
118 /* Make sure all the entries are set before we return */
119 ioread32(gtt_slot - 1);
120
121 return 0;
122 }
123
124 /**
125 * psb_gtt_remove - remove an object from the GTT
126 * @dev: our DRM device
127 * @r: our GTT range
128 *
129 * Remove a preallocated GTT range from the GTT. Overwrite all the
130 * page table entries with the dummy page. This is protected via the gtt
131 * mutex which the caller must hold.
132 */
psb_gtt_remove(struct drm_device * dev,struct gtt_range * r)133 void psb_gtt_remove(struct drm_device *dev, struct gtt_range *r)
134 {
135 struct drm_psb_private *dev_priv = dev->dev_private;
136 u32 __iomem *gtt_slot;
137 u32 pte;
138 int i;
139
140 WARN_ON(r->stolen);
141
142 gtt_slot = psb_gtt_entry(dev, r);
143 pte = psb_gtt_mask_pte(page_to_pfn(dev_priv->scratch_page),
144 PSB_MMU_CACHED_MEMORY);
145
146 for (i = 0; i < r->npage; i++)
147 iowrite32(pte, gtt_slot++);
148 ioread32(gtt_slot - 1);
149 set_pages_array_wb(r->pages, r->npage);
150 }
151
152 /**
153 * psb_gtt_roll - set scrolling position
154 * @dev: our DRM device
155 * @r: the gtt mapping we are using
156 * @roll: roll offset
157 *
158 * Roll an existing pinned mapping by moving the pages through the GTT.
159 * This allows us to implement hardware scrolling on the consoles without
160 * a 2D engine
161 */
psb_gtt_roll(struct drm_device * dev,struct gtt_range * r,int roll)162 void psb_gtt_roll(struct drm_device *dev, struct gtt_range *r, int roll)
163 {
164 u32 __iomem *gtt_slot;
165 u32 pte;
166 int i;
167
168 if (roll >= r->npage) {
169 WARN_ON(1);
170 return;
171 }
172
173 r->roll = roll;
174
175 /* Not currently in the GTT - no worry we will write the mapping at
176 the right position when it gets pinned */
177 if (!r->stolen && !r->in_gart)
178 return;
179
180 gtt_slot = psb_gtt_entry(dev, r);
181
182 for (i = r->roll; i < r->npage; i++) {
183 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
184 PSB_MMU_CACHED_MEMORY);
185 iowrite32(pte, gtt_slot++);
186 }
187 for (i = 0; i < r->roll; i++) {
188 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
189 PSB_MMU_CACHED_MEMORY);
190 iowrite32(pte, gtt_slot++);
191 }
192 ioread32(gtt_slot - 1);
193 }
194
195 /**
196 * psb_gtt_attach_pages - attach and pin GEM pages
197 * @gt: the gtt range
198 *
199 * Pin and build an in kernel list of the pages that back our GEM object.
200 * While we hold this the pages cannot be swapped out. This is protected
201 * via the gtt mutex which the caller must hold.
202 */
psb_gtt_attach_pages(struct gtt_range * gt)203 static int psb_gtt_attach_pages(struct gtt_range *gt)
204 {
205 struct page **pages;
206
207 WARN_ON(gt->pages);
208
209 pages = drm_gem_get_pages(>->gem);
210 if (IS_ERR(pages))
211 return PTR_ERR(pages);
212
213 gt->npage = gt->gem.size / PAGE_SIZE;
214 gt->pages = pages;
215
216 return 0;
217 }
218
219 /**
220 * psb_gtt_detach_pages - attach and pin GEM pages
221 * @gt: the gtt range
222 *
223 * Undo the effect of psb_gtt_attach_pages. At this point the pages
224 * must have been removed from the GTT as they could now be paged out
225 * and move bus address. This is protected via the gtt mutex which the
226 * caller must hold.
227 */
psb_gtt_detach_pages(struct gtt_range * gt)228 static void psb_gtt_detach_pages(struct gtt_range *gt)
229 {
230 drm_gem_put_pages(>->gem, gt->pages, true, false);
231 gt->pages = NULL;
232 }
233
234 /**
235 * psb_gtt_pin - pin pages into the GTT
236 * @gt: range to pin
237 *
238 * Pin a set of pages into the GTT. The pins are refcounted so that
239 * multiple pins need multiple unpins to undo.
240 *
241 * Non GEM backed objects treat this as a no-op as they are always GTT
242 * backed objects.
243 */
psb_gtt_pin(struct gtt_range * gt)244 int psb_gtt_pin(struct gtt_range *gt)
245 {
246 int ret = 0;
247 struct drm_device *dev = gt->gem.dev;
248 struct drm_psb_private *dev_priv = dev->dev_private;
249 u32 gpu_base = dev_priv->gtt.gatt_start;
250
251 mutex_lock(&dev_priv->gtt_mutex);
252
253 if (gt->in_gart == 0 && gt->stolen == 0) {
254 ret = psb_gtt_attach_pages(gt);
255 if (ret < 0)
256 goto out;
257 ret = psb_gtt_insert(dev, gt, 0);
258 if (ret < 0) {
259 psb_gtt_detach_pages(gt);
260 goto out;
261 }
262 psb_mmu_insert_pages(psb_mmu_get_default_pd(dev_priv->mmu),
263 gt->pages, (gpu_base + gt->offset),
264 gt->npage, 0, 0, PSB_MMU_CACHED_MEMORY);
265 }
266 gt->in_gart++;
267 out:
268 mutex_unlock(&dev_priv->gtt_mutex);
269 return ret;
270 }
271
272 /**
273 * psb_gtt_unpin - Drop a GTT pin requirement
274 * @gt: range to pin
275 *
276 * Undoes the effect of psb_gtt_pin. On the last drop the GEM object
277 * will be removed from the GTT which will also drop the page references
278 * and allow the VM to clean up or page stuff.
279 *
280 * Non GEM backed objects treat this as a no-op as they are always GTT
281 * backed objects.
282 */
psb_gtt_unpin(struct gtt_range * gt)283 void psb_gtt_unpin(struct gtt_range *gt)
284 {
285 struct drm_device *dev = gt->gem.dev;
286 struct drm_psb_private *dev_priv = dev->dev_private;
287 u32 gpu_base = dev_priv->gtt.gatt_start;
288 int ret;
289
290 /* While holding the gtt_mutex no new blits can be initiated */
291 mutex_lock(&dev_priv->gtt_mutex);
292
293 /* Wait for any possible usage of the memory to be finished */
294 ret = gma_blt_wait_idle(dev_priv);
295 if (ret) {
296 DRM_ERROR("Failed to idle the blitter, unpin failed!");
297 goto out;
298 }
299
300 WARN_ON(!gt->in_gart);
301
302 gt->in_gart--;
303 if (gt->in_gart == 0 && gt->stolen == 0) {
304 psb_mmu_remove_pages(psb_mmu_get_default_pd(dev_priv->mmu),
305 (gpu_base + gt->offset), gt->npage, 0, 0);
306 psb_gtt_remove(dev, gt);
307 psb_gtt_detach_pages(gt);
308 }
309
310 out:
311 mutex_unlock(&dev_priv->gtt_mutex);
312 }
313
314 /*
315 * GTT resource allocator - allocate and manage GTT address space
316 */
317
318 /**
319 * psb_gtt_alloc_range - allocate GTT address space
320 * @dev: Our DRM device
321 * @len: length (bytes) of address space required
322 * @name: resource name
323 * @backed: resource should be backed by stolen pages
324 *
325 * Ask the kernel core to find us a suitable range of addresses
326 * to use for a GTT mapping.
327 *
328 * Returns a gtt_range structure describing the object, or NULL on
329 * error. On successful return the resource is both allocated and marked
330 * as in use.
331 */
psb_gtt_alloc_range(struct drm_device * dev,int len,const char * name,int backed,u32 align)332 struct gtt_range *psb_gtt_alloc_range(struct drm_device *dev, int len,
333 const char *name, int backed, u32 align)
334 {
335 struct drm_psb_private *dev_priv = dev->dev_private;
336 struct gtt_range *gt;
337 struct resource *r = dev_priv->gtt_mem;
338 int ret;
339 unsigned long start, end;
340
341 if (backed) {
342 /* The start of the GTT is the stolen pages */
343 start = r->start;
344 end = r->start + dev_priv->gtt.stolen_size - 1;
345 } else {
346 /* The rest we will use for GEM backed objects */
347 start = r->start + dev_priv->gtt.stolen_size;
348 end = r->end;
349 }
350
351 gt = kzalloc(sizeof(struct gtt_range), GFP_KERNEL);
352 if (gt == NULL)
353 return NULL;
354 gt->resource.name = name;
355 gt->stolen = backed;
356 gt->in_gart = backed;
357 gt->roll = 0;
358 /* Ensure this is set for non GEM objects */
359 gt->gem.dev = dev;
360 ret = allocate_resource(dev_priv->gtt_mem, >->resource,
361 len, start, end, align, NULL, NULL);
362 if (ret == 0) {
363 gt->offset = gt->resource.start - r->start;
364 return gt;
365 }
366 kfree(gt);
367 return NULL;
368 }
369
370 /**
371 * psb_gtt_free_range - release GTT address space
372 * @dev: our DRM device
373 * @gt: a mapping created with psb_gtt_alloc_range
374 *
375 * Release a resource that was allocated with psb_gtt_alloc_range. If the
376 * object has been pinned by mmap users we clean this up here currently.
377 */
psb_gtt_free_range(struct drm_device * dev,struct gtt_range * gt)378 void psb_gtt_free_range(struct drm_device *dev, struct gtt_range *gt)
379 {
380 /* Undo the mmap pin if we are destroying the object */
381 if (gt->mmapping) {
382 psb_gtt_unpin(gt);
383 gt->mmapping = 0;
384 }
385 WARN_ON(gt->in_gart && !gt->stolen);
386 release_resource(>->resource);
387 kfree(gt);
388 }
389
psb_gtt_alloc(struct drm_device * dev)390 static void psb_gtt_alloc(struct drm_device *dev)
391 {
392 struct drm_psb_private *dev_priv = dev->dev_private;
393 init_rwsem(&dev_priv->gtt.sem);
394 }
395
psb_gtt_takedown(struct drm_device * dev)396 void psb_gtt_takedown(struct drm_device *dev)
397 {
398 struct drm_psb_private *dev_priv = dev->dev_private;
399
400 if (dev_priv->gtt_map) {
401 iounmap(dev_priv->gtt_map);
402 dev_priv->gtt_map = NULL;
403 }
404 if (dev_priv->gtt_initialized) {
405 pci_write_config_word(dev->pdev, PSB_GMCH_CTRL,
406 dev_priv->gmch_ctrl);
407 PSB_WVDC32(dev_priv->pge_ctl, PSB_PGETBL_CTL);
408 (void) PSB_RVDC32(PSB_PGETBL_CTL);
409 }
410 if (dev_priv->vram_addr)
411 iounmap(dev_priv->gtt_map);
412 }
413
psb_gtt_init(struct drm_device * dev,int resume)414 int psb_gtt_init(struct drm_device *dev, int resume)
415 {
416 struct drm_psb_private *dev_priv = dev->dev_private;
417 unsigned gtt_pages;
418 unsigned long stolen_size, vram_stolen_size;
419 unsigned i, num_pages;
420 unsigned pfn_base;
421 struct psb_gtt *pg;
422
423 int ret = 0;
424 uint32_t pte;
425
426 if (!resume) {
427 mutex_init(&dev_priv->gtt_mutex);
428 psb_gtt_alloc(dev);
429 }
430
431 pg = &dev_priv->gtt;
432
433 /* Enable the GTT */
434 pci_read_config_word(dev->pdev, PSB_GMCH_CTRL, &dev_priv->gmch_ctrl);
435 pci_write_config_word(dev->pdev, PSB_GMCH_CTRL,
436 dev_priv->gmch_ctrl | _PSB_GMCH_ENABLED);
437
438 dev_priv->pge_ctl = PSB_RVDC32(PSB_PGETBL_CTL);
439 PSB_WVDC32(dev_priv->pge_ctl | _PSB_PGETBL_ENABLED, PSB_PGETBL_CTL);
440 (void) PSB_RVDC32(PSB_PGETBL_CTL);
441
442 /* The root resource we allocate address space from */
443 dev_priv->gtt_initialized = 1;
444
445 pg->gtt_phys_start = dev_priv->pge_ctl & PAGE_MASK;
446
447 /*
448 * The video mmu has a hw bug when accessing 0x0D0000000.
449 * Make gatt start at 0x0e000,0000. This doesn't actually
450 * matter for us but may do if the video acceleration ever
451 * gets opened up.
452 */
453 pg->mmu_gatt_start = 0xE0000000;
454
455 pg->gtt_start = pci_resource_start(dev->pdev, PSB_GTT_RESOURCE);
456 gtt_pages = pci_resource_len(dev->pdev, PSB_GTT_RESOURCE)
457 >> PAGE_SHIFT;
458 /* CDV doesn't report this. In which case the system has 64 gtt pages */
459 if (pg->gtt_start == 0 || gtt_pages == 0) {
460 dev_dbg(dev->dev, "GTT PCI BAR not initialized.\n");
461 gtt_pages = 64;
462 pg->gtt_start = dev_priv->pge_ctl;
463 }
464
465 pg->gatt_start = pci_resource_start(dev->pdev, PSB_GATT_RESOURCE);
466 pg->gatt_pages = pci_resource_len(dev->pdev, PSB_GATT_RESOURCE)
467 >> PAGE_SHIFT;
468 dev_priv->gtt_mem = &dev->pdev->resource[PSB_GATT_RESOURCE];
469
470 if (pg->gatt_pages == 0 || pg->gatt_start == 0) {
471 static struct resource fudge; /* Preferably peppermint */
472 /* This can occur on CDV systems. Fudge it in this case.
473 We really don't care what imaginary space is being allocated
474 at this point */
475 dev_dbg(dev->dev, "GATT PCI BAR not initialized.\n");
476 pg->gatt_start = 0x40000000;
477 pg->gatt_pages = (128 * 1024 * 1024) >> PAGE_SHIFT;
478 /* This is a little confusing but in fact the GTT is providing
479 a view from the GPU into memory and not vice versa. As such
480 this is really allocating space that is not the same as the
481 CPU address space on CDV */
482 fudge.start = 0x40000000;
483 fudge.end = 0x40000000 + 128 * 1024 * 1024 - 1;
484 fudge.name = "fudge";
485 fudge.flags = IORESOURCE_MEM;
486 dev_priv->gtt_mem = &fudge;
487 }
488
489 pci_read_config_dword(dev->pdev, PSB_BSM, &dev_priv->stolen_base);
490 vram_stolen_size = pg->gtt_phys_start - dev_priv->stolen_base
491 - PAGE_SIZE;
492
493 stolen_size = vram_stolen_size;
494
495 dev_dbg(dev->dev, "Stolen memory base 0x%x, size %luK\n",
496 dev_priv->stolen_base, vram_stolen_size / 1024);
497
498 if (resume && (gtt_pages != pg->gtt_pages) &&
499 (stolen_size != pg->stolen_size)) {
500 dev_err(dev->dev, "GTT resume error.\n");
501 ret = -EINVAL;
502 goto out_err;
503 }
504
505 pg->gtt_pages = gtt_pages;
506 pg->stolen_size = stolen_size;
507 dev_priv->vram_stolen_size = vram_stolen_size;
508
509 /*
510 * Map the GTT and the stolen memory area
511 */
512 if (!resume)
513 dev_priv->gtt_map = ioremap_nocache(pg->gtt_phys_start,
514 gtt_pages << PAGE_SHIFT);
515 if (!dev_priv->gtt_map) {
516 dev_err(dev->dev, "Failure to map gtt.\n");
517 ret = -ENOMEM;
518 goto out_err;
519 }
520
521 if (!resume)
522 dev_priv->vram_addr = ioremap_wc(dev_priv->stolen_base,
523 stolen_size);
524
525 if (!dev_priv->vram_addr) {
526 dev_err(dev->dev, "Failure to map stolen base.\n");
527 ret = -ENOMEM;
528 goto out_err;
529 }
530
531 /*
532 * Insert vram stolen pages into the GTT
533 */
534
535 pfn_base = dev_priv->stolen_base >> PAGE_SHIFT;
536 num_pages = vram_stolen_size >> PAGE_SHIFT;
537 dev_dbg(dev->dev, "Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n",
538 num_pages, pfn_base << PAGE_SHIFT, 0);
539 for (i = 0; i < num_pages; ++i) {
540 pte = psb_gtt_mask_pte(pfn_base + i, PSB_MMU_CACHED_MEMORY);
541 iowrite32(pte, dev_priv->gtt_map + i);
542 }
543
544 /*
545 * Init rest of GTT to the scratch page to avoid accidents or scribbles
546 */
547
548 pfn_base = page_to_pfn(dev_priv->scratch_page);
549 pte = psb_gtt_mask_pte(pfn_base, PSB_MMU_CACHED_MEMORY);
550 for (; i < gtt_pages; ++i)
551 iowrite32(pte, dev_priv->gtt_map + i);
552
553 (void) ioread32(dev_priv->gtt_map + i - 1);
554 return 0;
555
556 out_err:
557 psb_gtt_takedown(dev);
558 return ret;
559 }
560
psb_gtt_restore(struct drm_device * dev)561 int psb_gtt_restore(struct drm_device *dev)
562 {
563 struct drm_psb_private *dev_priv = dev->dev_private;
564 struct resource *r = dev_priv->gtt_mem->child;
565 struct gtt_range *range;
566 unsigned int restored = 0, total = 0, size = 0;
567
568 /* On resume, the gtt_mutex is already initialized */
569 mutex_lock(&dev_priv->gtt_mutex);
570 psb_gtt_init(dev, 1);
571
572 while (r != NULL) {
573 range = container_of(r, struct gtt_range, resource);
574 if (range->pages) {
575 psb_gtt_insert(dev, range, 1);
576 size += range->resource.end - range->resource.start;
577 restored++;
578 }
579 r = r->sibling;
580 total++;
581 }
582 mutex_unlock(&dev_priv->gtt_mutex);
583 DRM_DEBUG_DRIVER("Restored %u of %u gtt ranges (%u KB)", restored,
584 total, (size / 1024));
585
586 return 0;
587 }
588