1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (c) 2007, Intel Corporation.
4 * All Rights Reserved.
5 *
6 * Authors: Thomas Hellstrom <thomas-at-tungstengraphics.com>
7 * Alan Cox <alan@linux.intel.com>
8 */
9
10 #include <linux/shmem_fs.h>
11
12 #include <asm/set_memory.h>
13
14 #include "blitter.h"
15 #include "psb_drv.h"
16
17
18 /*
19 * GTT resource allocator - manage page mappings in GTT space
20 */
21
22 /**
23 * psb_gtt_mask_pte - generate GTT pte entry
24 * @pfn: page number to encode
25 * @type: type of memory in the GTT
26 *
27 * Set the GTT entry for the appropriate memory type.
28 */
psb_gtt_mask_pte(uint32_t pfn,int type)29 static inline uint32_t psb_gtt_mask_pte(uint32_t pfn, int type)
30 {
31 uint32_t mask = PSB_PTE_VALID;
32
33 /* Ensure we explode rather than put an invalid low mapping of
34 a high mapping page into the gtt */
35 BUG_ON(pfn & ~(0xFFFFFFFF >> PAGE_SHIFT));
36
37 if (type & PSB_MMU_CACHED_MEMORY)
38 mask |= PSB_PTE_CACHED;
39 if (type & PSB_MMU_RO_MEMORY)
40 mask |= PSB_PTE_RO;
41 if (type & PSB_MMU_WO_MEMORY)
42 mask |= PSB_PTE_WO;
43
44 return (pfn << PAGE_SHIFT) | mask;
45 }
46
47 /**
48 * psb_gtt_entry - find the GTT entries for a gtt_range
49 * @dev: our DRM device
50 * @r: our GTT range
51 *
52 * Given a gtt_range object return the GTT offset of the page table
53 * entries for this gtt_range
54 */
psb_gtt_entry(struct drm_device * dev,struct gtt_range * r)55 static u32 __iomem *psb_gtt_entry(struct drm_device *dev, struct gtt_range *r)
56 {
57 struct drm_psb_private *dev_priv = dev->dev_private;
58 unsigned long offset;
59
60 offset = r->resource.start - dev_priv->gtt_mem->start;
61
62 return dev_priv->gtt_map + (offset >> PAGE_SHIFT);
63 }
64
65 /**
66 * psb_gtt_insert - put an object into the GTT
67 * @dev: our DRM device
68 * @r: our GTT range
69 * @resume: on resume
70 *
71 * Take our preallocated GTT range and insert the GEM object into
72 * the GTT. This is protected via the gtt mutex which the caller
73 * must hold.
74 */
psb_gtt_insert(struct drm_device * dev,struct gtt_range * r,int resume)75 static int psb_gtt_insert(struct drm_device *dev, struct gtt_range *r,
76 int resume)
77 {
78 u32 __iomem *gtt_slot;
79 u32 pte;
80 struct page **pages;
81 int i;
82
83 if (r->pages == NULL) {
84 WARN_ON(1);
85 return -EINVAL;
86 }
87
88 WARN_ON(r->stolen); /* refcount these maybe ? */
89
90 gtt_slot = psb_gtt_entry(dev, r);
91 pages = r->pages;
92
93 if (!resume) {
94 /* Make sure changes are visible to the GPU */
95 set_pages_array_wc(pages, r->npage);
96 }
97
98 /* Write our page entries into the GTT itself */
99 for (i = r->roll; i < r->npage; i++) {
100 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
101 PSB_MMU_CACHED_MEMORY);
102 iowrite32(pte, gtt_slot++);
103 }
104 for (i = 0; i < r->roll; i++) {
105 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
106 PSB_MMU_CACHED_MEMORY);
107 iowrite32(pte, gtt_slot++);
108 }
109 /* Make sure all the entries are set before we return */
110 ioread32(gtt_slot - 1);
111
112 return 0;
113 }
114
115 /**
116 * psb_gtt_remove - remove an object from the GTT
117 * @dev: our DRM device
118 * @r: our GTT range
119 *
120 * Remove a preallocated GTT range from the GTT. Overwrite all the
121 * page table entries with the dummy page. This is protected via the gtt
122 * mutex which the caller must hold.
123 */
psb_gtt_remove(struct drm_device * dev,struct gtt_range * r)124 static void psb_gtt_remove(struct drm_device *dev, struct gtt_range *r)
125 {
126 struct drm_psb_private *dev_priv = dev->dev_private;
127 u32 __iomem *gtt_slot;
128 u32 pte;
129 int i;
130
131 WARN_ON(r->stolen);
132
133 gtt_slot = psb_gtt_entry(dev, r);
134 pte = psb_gtt_mask_pte(page_to_pfn(dev_priv->scratch_page),
135 PSB_MMU_CACHED_MEMORY);
136
137 for (i = 0; i < r->npage; i++)
138 iowrite32(pte, gtt_slot++);
139 ioread32(gtt_slot - 1);
140 set_pages_array_wb(r->pages, r->npage);
141 }
142
143 /**
144 * psb_gtt_roll - set scrolling position
145 * @dev: our DRM device
146 * @r: the gtt mapping we are using
147 * @roll: roll offset
148 *
149 * Roll an existing pinned mapping by moving the pages through the GTT.
150 * This allows us to implement hardware scrolling on the consoles without
151 * a 2D engine
152 */
psb_gtt_roll(struct drm_device * dev,struct gtt_range * r,int roll)153 void psb_gtt_roll(struct drm_device *dev, struct gtt_range *r, int roll)
154 {
155 u32 __iomem *gtt_slot;
156 u32 pte;
157 int i;
158
159 if (roll >= r->npage) {
160 WARN_ON(1);
161 return;
162 }
163
164 r->roll = roll;
165
166 /* Not currently in the GTT - no worry we will write the mapping at
167 the right position when it gets pinned */
168 if (!r->stolen && !r->in_gart)
169 return;
170
171 gtt_slot = psb_gtt_entry(dev, r);
172
173 for (i = r->roll; i < r->npage; i++) {
174 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
175 PSB_MMU_CACHED_MEMORY);
176 iowrite32(pte, gtt_slot++);
177 }
178 for (i = 0; i < r->roll; i++) {
179 pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
180 PSB_MMU_CACHED_MEMORY);
181 iowrite32(pte, gtt_slot++);
182 }
183 ioread32(gtt_slot - 1);
184 }
185
186 /**
187 * psb_gtt_attach_pages - attach and pin GEM pages
188 * @gt: the gtt range
189 *
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.
193 */
psb_gtt_attach_pages(struct gtt_range * gt)194 static int psb_gtt_attach_pages(struct gtt_range *gt)
195 {
196 struct page **pages;
197
198 WARN_ON(gt->pages);
199
200 pages = drm_gem_get_pages(>->gem);
201 if (IS_ERR(pages))
202 return PTR_ERR(pages);
203
204 gt->npage = gt->gem.size / PAGE_SIZE;
205 gt->pages = pages;
206
207 return 0;
208 }
209
210 /**
211 * psb_gtt_detach_pages - attach and pin GEM pages
212 * @gt: the gtt range
213 *
214 * Undo the effect of psb_gtt_attach_pages. At this point the pages
215 * must have been removed from the GTT as they could now be paged out
216 * and move bus address. This is protected via the gtt mutex which the
217 * caller must hold.
218 */
psb_gtt_detach_pages(struct gtt_range * gt)219 static void psb_gtt_detach_pages(struct gtt_range *gt)
220 {
221 drm_gem_put_pages(>->gem, gt->pages, true, false);
222 gt->pages = NULL;
223 }
224
225 /**
226 * psb_gtt_pin - pin pages into the GTT
227 * @gt: range to pin
228 *
229 * Pin a set of pages into the GTT. The pins are refcounted so that
230 * multiple pins need multiple unpins to undo.
231 *
232 * Non GEM backed objects treat this as a no-op as they are always GTT
233 * backed objects.
234 */
psb_gtt_pin(struct gtt_range * gt)235 int psb_gtt_pin(struct gtt_range *gt)
236 {
237 int ret = 0;
238 struct drm_device *dev = gt->gem.dev;
239 struct drm_psb_private *dev_priv = dev->dev_private;
240 u32 gpu_base = dev_priv->gtt.gatt_start;
241
242 mutex_lock(&dev_priv->gtt_mutex);
243
244 if (gt->in_gart == 0 && gt->stolen == 0) {
245 ret = psb_gtt_attach_pages(gt);
246 if (ret < 0)
247 goto out;
248 ret = psb_gtt_insert(dev, gt, 0);
249 if (ret < 0) {
250 psb_gtt_detach_pages(gt);
251 goto out;
252 }
253 psb_mmu_insert_pages(psb_mmu_get_default_pd(dev_priv->mmu),
254 gt->pages, (gpu_base + gt->offset),
255 gt->npage, 0, 0, PSB_MMU_CACHED_MEMORY);
256 }
257 gt->in_gart++;
258 out:
259 mutex_unlock(&dev_priv->gtt_mutex);
260 return ret;
261 }
262
263 /**
264 * psb_gtt_unpin - Drop a GTT pin requirement
265 * @gt: range to pin
266 *
267 * Undoes the effect of psb_gtt_pin. On the last drop the GEM object
268 * will be removed from the GTT which will also drop the page references
269 * and allow the VM to clean up or page stuff.
270 *
271 * Non GEM backed objects treat this as a no-op as they are always GTT
272 * backed objects.
273 */
psb_gtt_unpin(struct gtt_range * gt)274 void psb_gtt_unpin(struct gtt_range *gt)
275 {
276 struct drm_device *dev = gt->gem.dev;
277 struct drm_psb_private *dev_priv = dev->dev_private;
278 u32 gpu_base = dev_priv->gtt.gatt_start;
279 int ret;
280
281 /* While holding the gtt_mutex no new blits can be initiated */
282 mutex_lock(&dev_priv->gtt_mutex);
283
284 /* Wait for any possible usage of the memory to be finished */
285 ret = gma_blt_wait_idle(dev_priv);
286 if (ret) {
287 DRM_ERROR("Failed to idle the blitter, unpin failed!");
288 goto out;
289 }
290
291 WARN_ON(!gt->in_gart);
292
293 gt->in_gart--;
294 if (gt->in_gart == 0 && gt->stolen == 0) {
295 psb_mmu_remove_pages(psb_mmu_get_default_pd(dev_priv->mmu),
296 (gpu_base + gt->offset), gt->npage, 0, 0);
297 psb_gtt_remove(dev, gt);
298 psb_gtt_detach_pages(gt);
299 }
300
301 out:
302 mutex_unlock(&dev_priv->gtt_mutex);
303 }
304
305 /*
306 * GTT resource allocator - allocate and manage GTT address space
307 */
308
309 /**
310 * psb_gtt_alloc_range - allocate GTT address space
311 * @dev: Our DRM device
312 * @len: length (bytes) of address space required
313 * @name: resource name
314 * @backed: resource should be backed by stolen pages
315 * @align: requested alignment
316 *
317 * Ask the kernel core to find us a suitable range of addresses
318 * to use for a GTT mapping.
319 *
320 * Returns a gtt_range structure describing the object, or NULL on
321 * error. On successful return the resource is both allocated and marked
322 * as in use.
323 */
psb_gtt_alloc_range(struct drm_device * dev,int len,const char * name,int backed,u32 align)324 struct gtt_range *psb_gtt_alloc_range(struct drm_device *dev, int len,
325 const char *name, int backed, u32 align)
326 {
327 struct drm_psb_private *dev_priv = dev->dev_private;
328 struct gtt_range *gt;
329 struct resource *r = dev_priv->gtt_mem;
330 int ret;
331 unsigned long start, end;
332
333 if (backed) {
334 /* The start of the GTT is the stolen pages */
335 start = r->start;
336 end = r->start + dev_priv->gtt.stolen_size - 1;
337 } else {
338 /* The rest we will use for GEM backed objects */
339 start = r->start + dev_priv->gtt.stolen_size;
340 end = r->end;
341 }
342
343 gt = kzalloc(sizeof(struct gtt_range), GFP_KERNEL);
344 if (gt == NULL)
345 return NULL;
346 gt->resource.name = name;
347 gt->stolen = backed;
348 gt->in_gart = backed;
349 gt->roll = 0;
350 /* Ensure this is set for non GEM objects */
351 gt->gem.dev = dev;
352 ret = allocate_resource(dev_priv->gtt_mem, >->resource,
353 len, start, end, align, NULL, NULL);
354 if (ret == 0) {
355 gt->offset = gt->resource.start - r->start;
356 return gt;
357 }
358 kfree(gt);
359 return NULL;
360 }
361
362 /**
363 * psb_gtt_free_range - release GTT address space
364 * @dev: our DRM device
365 * @gt: a mapping created with psb_gtt_alloc_range
366 *
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.
369 */
psb_gtt_free_range(struct drm_device * dev,struct gtt_range * gt)370 void psb_gtt_free_range(struct drm_device *dev, struct gtt_range *gt)
371 {
372 /* Undo the mmap pin if we are destroying the object */
373 if (gt->mmapping) {
374 psb_gtt_unpin(gt);
375 gt->mmapping = 0;
376 }
377 WARN_ON(gt->in_gart && !gt->stolen);
378 release_resource(>->resource);
379 kfree(gt);
380 }
381
psb_gtt_alloc(struct drm_device * dev)382 static void psb_gtt_alloc(struct drm_device *dev)
383 {
384 struct drm_psb_private *dev_priv = dev->dev_private;
385 init_rwsem(&dev_priv->gtt.sem);
386 }
387
psb_gtt_takedown(struct drm_device * dev)388 void psb_gtt_takedown(struct drm_device *dev)
389 {
390 struct drm_psb_private *dev_priv = dev->dev_private;
391
392 if (dev_priv->gtt_map) {
393 iounmap(dev_priv->gtt_map);
394 dev_priv->gtt_map = NULL;
395 }
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);
401 }
402 if (dev_priv->vram_addr)
403 iounmap(dev_priv->gtt_map);
404 }
405
psb_gtt_init(struct drm_device * dev,int resume)406 int psb_gtt_init(struct drm_device *dev, int resume)
407 {
408 struct drm_psb_private *dev_priv = dev->dev_private;
409 unsigned gtt_pages;
410 unsigned long stolen_size, vram_stolen_size;
411 unsigned i, num_pages;
412 unsigned pfn_base;
413 struct psb_gtt *pg;
414
415 int ret = 0;
416 uint32_t pte;
417
418 if (!resume) {
419 mutex_init(&dev_priv->gtt_mutex);
420 mutex_init(&dev_priv->mmap_mutex);
421 psb_gtt_alloc(dev);
422 }
423
424 pg = &dev_priv->gtt;
425
426 /* Enable the GTT */
427 pci_read_config_word(dev->pdev, PSB_GMCH_CTRL, &dev_priv->gmch_ctrl);
428 pci_write_config_word(dev->pdev, PSB_GMCH_CTRL,
429 dev_priv->gmch_ctrl | _PSB_GMCH_ENABLED);
430
431 dev_priv->pge_ctl = PSB_RVDC32(PSB_PGETBL_CTL);
432 PSB_WVDC32(dev_priv->pge_ctl | _PSB_PGETBL_ENABLED, PSB_PGETBL_CTL);
433 (void) PSB_RVDC32(PSB_PGETBL_CTL);
434
435 /* The root resource we allocate address space from */
436 dev_priv->gtt_initialized = 1;
437
438 pg->gtt_phys_start = dev_priv->pge_ctl & PAGE_MASK;
439
440 /*
441 * The video mmu has a hw bug when accessing 0x0D0000000.
442 * Make gatt start at 0x0e000,0000. This doesn't actually
443 * matter for us but may do if the video acceleration ever
444 * gets opened up.
445 */
446 pg->mmu_gatt_start = 0xE0000000;
447
448 pg->gtt_start = pci_resource_start(dev->pdev, PSB_GTT_RESOURCE);
449 gtt_pages = pci_resource_len(dev->pdev, PSB_GTT_RESOURCE)
450 >> PAGE_SHIFT;
451 /* CDV doesn't report this. In which case the system has 64 gtt pages */
452 if (pg->gtt_start == 0 || gtt_pages == 0) {
453 dev_dbg(dev->dev, "GTT PCI BAR not initialized.\n");
454 gtt_pages = 64;
455 pg->gtt_start = dev_priv->pge_ctl;
456 }
457
458 pg->gatt_start = pci_resource_start(dev->pdev, PSB_GATT_RESOURCE);
459 pg->gatt_pages = pci_resource_len(dev->pdev, PSB_GATT_RESOURCE)
460 >> PAGE_SHIFT;
461 dev_priv->gtt_mem = &dev->pdev->resource[PSB_GATT_RESOURCE];
462
463 if (pg->gatt_pages == 0 || pg->gatt_start == 0) {
464 static struct resource fudge; /* Preferably peppermint */
465 /* This can occur on CDV systems. Fudge it in this case.
466 We really don't care what imaginary space is being allocated
467 at this point */
468 dev_dbg(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;
480 }
481
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
484 - PAGE_SIZE;
485
486 stolen_size = vram_stolen_size;
487
488 dev_dbg(dev->dev, "Stolen memory base 0x%x, size %luK\n",
489 dev_priv->stolen_base, vram_stolen_size / 1024);
490
491 if (resume && (gtt_pages != pg->gtt_pages) &&
492 (stolen_size != pg->stolen_size)) {
493 dev_err(dev->dev, "GTT resume error.\n");
494 ret = -EINVAL;
495 goto out_err;
496 }
497
498 pg->gtt_pages = gtt_pages;
499 pg->stolen_size = stolen_size;
500 dev_priv->vram_stolen_size = vram_stolen_size;
501
502 /*
503 * Map the GTT and the stolen memory area
504 */
505 if (!resume)
506 dev_priv->gtt_map = ioremap(pg->gtt_phys_start,
507 gtt_pages << PAGE_SHIFT);
508 if (!dev_priv->gtt_map) {
509 dev_err(dev->dev, "Failure to map gtt.\n");
510 ret = -ENOMEM;
511 goto out_err;
512 }
513
514 if (!resume)
515 dev_priv->vram_addr = ioremap_wc(dev_priv->stolen_base,
516 stolen_size);
517
518 if (!dev_priv->vram_addr) {
519 dev_err(dev->dev, "Failure to map stolen base.\n");
520 ret = -ENOMEM;
521 goto out_err;
522 }
523
524 /*
525 * Insert vram stolen pages into the GTT
526 */
527
528 pfn_base = dev_priv->stolen_base >> PAGE_SHIFT;
529 num_pages = vram_stolen_size >> PAGE_SHIFT;
530 dev_dbg(dev->dev, "Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n",
531 num_pages, pfn_base << PAGE_SHIFT, 0);
532 for (i = 0; i < num_pages; ++i) {
533 pte = psb_gtt_mask_pte(pfn_base + i, PSB_MMU_CACHED_MEMORY);
534 iowrite32(pte, dev_priv->gtt_map + i);
535 }
536
537 /*
538 * Init rest of GTT to the scratch page to avoid accidents or scribbles
539 */
540
541 pfn_base = page_to_pfn(dev_priv->scratch_page);
542 pte = psb_gtt_mask_pte(pfn_base, PSB_MMU_CACHED_MEMORY);
543 for (; i < gtt_pages; ++i)
544 iowrite32(pte, dev_priv->gtt_map + i);
545
546 (void) ioread32(dev_priv->gtt_map + i - 1);
547 return 0;
548
549 out_err:
550 psb_gtt_takedown(dev);
551 return ret;
552 }
553
psb_gtt_restore(struct drm_device * dev)554 int psb_gtt_restore(struct drm_device *dev)
555 {
556 struct drm_psb_private *dev_priv = dev->dev_private;
557 struct resource *r = dev_priv->gtt_mem->child;
558 struct gtt_range *range;
559 unsigned int restored = 0, total = 0, size = 0;
560
561 /* On resume, the gtt_mutex is already initialized */
562 mutex_lock(&dev_priv->gtt_mutex);
563 psb_gtt_init(dev, 1);
564
565 while (r != NULL) {
566 range = container_of(r, struct gtt_range, resource);
567 if (range->pages) {
568 psb_gtt_insert(dev, range, 1);
569 size += range->resource.end - range->resource.start;
570 restored++;
571 }
572 r = r->sibling;
573 total++;
574 }
575 mutex_unlock(&dev_priv->gtt_mutex);
576 DRM_DEBUG_DRIVER("Restored %u of %u gtt ranges (%u KB)", restored,
577 total, (size / 1024));
578
579 return 0;
580 }
581
