1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2007, Intel Corporation.
6 * Authors: Thomas Hellstrom <thomas-at-tungstengraphics.com>
10 #include <linux/shmem_fs.h>
12 #include <asm/set_memory.h>
19 * GTT resource allocator - manage page mappings in GTT space
23 * psb_gtt_mask_pte - generate GTT pte entry
24 * @pfn: page number to encode
25 * @type: type of memory in the GTT
27 * Set the GTT entry for the appropriate memory type.
29 static inline uint32_t psb_gtt_mask_pte(uint32_t pfn, int type)
31 uint32_t mask = PSB_PTE_VALID;
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));
37 if (type & PSB_MMU_CACHED_MEMORY)
38 mask |= PSB_PTE_CACHED;
39 if (type & PSB_MMU_RO_MEMORY)
41 if (type & PSB_MMU_WO_MEMORY)
44 return (pfn << PAGE_SHIFT) | mask;
48 * psb_gtt_entry - find the GTT entries for a gtt_range
49 * @dev: our DRM device
52 * Given a gtt_range object return the GTT offset of the page table
53 * entries for this gtt_range
55 static u32 __iomem *psb_gtt_entry(struct drm_device *dev, struct gtt_range *r)
57 struct drm_psb_private *dev_priv = dev->dev_private;
60 offset = r->resource.start - dev_priv->gtt_mem->start;
62 return dev_priv->gtt_map + (offset >> PAGE_SHIFT);
66 * psb_gtt_insert - put an object into the GTT
67 * @dev: our DRM device
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
75 static int psb_gtt_insert(struct drm_device *dev, struct gtt_range *r,
78 u32 __iomem *gtt_slot;
83 if (r->pages == NULL) {
88 WARN_ON(r->stolen); /* refcount these maybe ? */
90 gtt_slot = psb_gtt_entry(dev, r);
94 /* Make sure changes are visible to the GPU */
95 set_pages_array_wc(pages, r->npage);
98 /* Write our page entries into the GTT itself */
99 for (i = 0; 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++);
105 /* Make sure all the entries are set before we return */
106 ioread32(gtt_slot - 1);
112 * psb_gtt_remove - remove an object from the GTT
113 * @dev: our DRM device
116 * Remove a preallocated GTT range from the GTT. Overwrite all the
117 * page table entries with the dummy page. This is protected via the gtt
118 * mutex which the caller must hold.
120 static void psb_gtt_remove(struct drm_device *dev, struct gtt_range *r)
122 struct drm_psb_private *dev_priv = dev->dev_private;
123 u32 __iomem *gtt_slot;
129 gtt_slot = psb_gtt_entry(dev, r);
130 pte = psb_gtt_mask_pte(page_to_pfn(dev_priv->scratch_page),
131 PSB_MMU_CACHED_MEMORY);
133 for (i = 0; i < r->npage; i++)
134 iowrite32(pte, gtt_slot++);
135 ioread32(gtt_slot - 1);
136 set_pages_array_wb(r->pages, r->npage);
140 * psb_gtt_attach_pages - attach and pin GEM pages
143 * Pin and build an in kernel list of the pages that back our GEM object.
144 * While we hold this the pages cannot be swapped out. This is protected
145 * via the gtt mutex which the caller must hold.
147 static int psb_gtt_attach_pages(struct gtt_range *gt)
153 pages = drm_gem_get_pages(>->gem);
155 return PTR_ERR(pages);
157 gt->npage = gt->gem.size / PAGE_SIZE;
164 * psb_gtt_detach_pages - attach and pin GEM pages
167 * Undo the effect of psb_gtt_attach_pages. At this point the pages
168 * must have been removed from the GTT as they could now be paged out
169 * and move bus address. This is protected via the gtt mutex which the
172 static void psb_gtt_detach_pages(struct gtt_range *gt)
174 drm_gem_put_pages(>->gem, gt->pages, true, false);
179 * psb_gtt_pin - pin pages into the GTT
182 * Pin a set of pages into the GTT. The pins are refcounted so that
183 * multiple pins need multiple unpins to undo.
185 * Non GEM backed objects treat this as a no-op as they are always GTT
188 int psb_gtt_pin(struct gtt_range *gt)
191 struct drm_device *dev = gt->gem.dev;
192 struct drm_psb_private *dev_priv = dev->dev_private;
193 u32 gpu_base = dev_priv->gtt.gatt_start;
195 mutex_lock(&dev_priv->gtt_mutex);
197 if (gt->in_gart == 0 && gt->stolen == 0) {
198 ret = psb_gtt_attach_pages(gt);
201 ret = psb_gtt_insert(dev, gt, 0);
203 psb_gtt_detach_pages(gt);
206 psb_mmu_insert_pages(psb_mmu_get_default_pd(dev_priv->mmu),
207 gt->pages, (gpu_base + gt->offset),
208 gt->npage, 0, 0, PSB_MMU_CACHED_MEMORY);
212 mutex_unlock(&dev_priv->gtt_mutex);
217 * psb_gtt_unpin - Drop a GTT pin requirement
220 * Undoes the effect of psb_gtt_pin. On the last drop the GEM object
221 * will be removed from the GTT which will also drop the page references
222 * and allow the VM to clean up or page stuff.
224 * Non GEM backed objects treat this as a no-op as they are always GTT
227 void psb_gtt_unpin(struct gtt_range *gt)
229 struct drm_device *dev = gt->gem.dev;
230 struct drm_psb_private *dev_priv = dev->dev_private;
231 u32 gpu_base = dev_priv->gtt.gatt_start;
234 /* While holding the gtt_mutex no new blits can be initiated */
235 mutex_lock(&dev_priv->gtt_mutex);
237 /* Wait for any possible usage of the memory to be finished */
238 ret = gma_blt_wait_idle(dev_priv);
240 DRM_ERROR("Failed to idle the blitter, unpin failed!");
244 WARN_ON(!gt->in_gart);
247 if (gt->in_gart == 0 && gt->stolen == 0) {
248 psb_mmu_remove_pages(psb_mmu_get_default_pd(dev_priv->mmu),
249 (gpu_base + gt->offset), gt->npage, 0, 0);
250 psb_gtt_remove(dev, gt);
251 psb_gtt_detach_pages(gt);
255 mutex_unlock(&dev_priv->gtt_mutex);
259 * GTT resource allocator - allocate and manage GTT address space
263 * psb_gtt_alloc_range - allocate GTT address space
264 * @dev: Our DRM device
265 * @len: length (bytes) of address space required
266 * @name: resource name
267 * @backed: resource should be backed by stolen pages
268 * @align: requested alignment
270 * Ask the kernel core to find us a suitable range of addresses
271 * to use for a GTT mapping.
273 * Returns a gtt_range structure describing the object, or NULL on
274 * error. On successful return the resource is both allocated and marked
277 struct gtt_range *psb_gtt_alloc_range(struct drm_device *dev, int len,
278 const char *name, int backed, u32 align)
280 struct drm_psb_private *dev_priv = dev->dev_private;
281 struct gtt_range *gt;
282 struct resource *r = dev_priv->gtt_mem;
284 unsigned long start, end;
287 /* The start of the GTT is the stolen pages */
289 end = r->start + dev_priv->gtt.stolen_size - 1;
291 /* The rest we will use for GEM backed objects */
292 start = r->start + dev_priv->gtt.stolen_size;
296 gt = kzalloc(sizeof(struct gtt_range), GFP_KERNEL);
299 gt->resource.name = name;
301 gt->in_gart = backed;
302 /* Ensure this is set for non GEM objects */
304 ret = allocate_resource(dev_priv->gtt_mem, >->resource,
305 len, start, end, align, NULL, NULL);
307 gt->offset = gt->resource.start - r->start;
315 * psb_gtt_free_range - release GTT address space
316 * @dev: our DRM device
317 * @gt: a mapping created with psb_gtt_alloc_range
319 * Release a resource that was allocated with psb_gtt_alloc_range. If the
320 * object has been pinned by mmap users we clean this up here currently.
322 void psb_gtt_free_range(struct drm_device *dev, struct gtt_range *gt)
324 /* Undo the mmap pin if we are destroying the object */
329 WARN_ON(gt->in_gart && !gt->stolen);
330 release_resource(>->resource);
334 static void psb_gtt_alloc(struct drm_device *dev)
336 struct drm_psb_private *dev_priv = dev->dev_private;
337 init_rwsem(&dev_priv->gtt.sem);
340 void psb_gtt_takedown(struct drm_device *dev)
342 struct drm_psb_private *dev_priv = dev->dev_private;
343 struct pci_dev *pdev = to_pci_dev(dev->dev);
345 if (dev_priv->gtt_map) {
346 iounmap(dev_priv->gtt_map);
347 dev_priv->gtt_map = NULL;
349 if (dev_priv->gtt_initialized) {
350 pci_write_config_word(pdev, PSB_GMCH_CTRL,
351 dev_priv->gmch_ctrl);
352 PSB_WVDC32(dev_priv->pge_ctl, PSB_PGETBL_CTL);
353 (void) PSB_RVDC32(PSB_PGETBL_CTL);
355 if (dev_priv->vram_addr)
356 iounmap(dev_priv->gtt_map);
359 int psb_gtt_init(struct drm_device *dev, int resume)
361 struct drm_psb_private *dev_priv = dev->dev_private;
362 struct pci_dev *pdev = to_pci_dev(dev->dev);
364 unsigned long stolen_size, vram_stolen_size;
365 unsigned i, num_pages;
373 mutex_init(&dev_priv->gtt_mutex);
374 mutex_init(&dev_priv->mmap_mutex);
381 pci_read_config_word(pdev, PSB_GMCH_CTRL, &dev_priv->gmch_ctrl);
382 pci_write_config_word(pdev, PSB_GMCH_CTRL,
383 dev_priv->gmch_ctrl | _PSB_GMCH_ENABLED);
385 dev_priv->pge_ctl = PSB_RVDC32(PSB_PGETBL_CTL);
386 PSB_WVDC32(dev_priv->pge_ctl | _PSB_PGETBL_ENABLED, PSB_PGETBL_CTL);
387 (void) PSB_RVDC32(PSB_PGETBL_CTL);
389 /* The root resource we allocate address space from */
390 dev_priv->gtt_initialized = 1;
392 pg->gtt_phys_start = dev_priv->pge_ctl & PAGE_MASK;
395 * The video mmu has a hw bug when accessing 0x0D0000000.
396 * Make gatt start at 0x0e000,0000. This doesn't actually
397 * matter for us but may do if the video acceleration ever
400 pg->mmu_gatt_start = 0xE0000000;
402 pg->gtt_start = pci_resource_start(pdev, PSB_GTT_RESOURCE);
403 gtt_pages = pci_resource_len(pdev, PSB_GTT_RESOURCE)
405 /* CDV doesn't report this. In which case the system has 64 gtt pages */
406 if (pg->gtt_start == 0 || gtt_pages == 0) {
407 dev_dbg(dev->dev, "GTT PCI BAR not initialized.\n");
409 pg->gtt_start = dev_priv->pge_ctl;
412 pg->gatt_start = pci_resource_start(pdev, PSB_GATT_RESOURCE);
413 pg->gatt_pages = pci_resource_len(pdev, PSB_GATT_RESOURCE)
415 dev_priv->gtt_mem = &pdev->resource[PSB_GATT_RESOURCE];
417 if (pg->gatt_pages == 0 || pg->gatt_start == 0) {
418 static struct resource fudge; /* Preferably peppermint */
419 /* This can occur on CDV systems. Fudge it in this case.
420 We really don't care what imaginary space is being allocated
422 dev_dbg(dev->dev, "GATT PCI BAR not initialized.\n");
423 pg->gatt_start = 0x40000000;
424 pg->gatt_pages = (128 * 1024 * 1024) >> PAGE_SHIFT;
425 /* This is a little confusing but in fact the GTT is providing
426 a view from the GPU into memory and not vice versa. As such
427 this is really allocating space that is not the same as the
428 CPU address space on CDV */
429 fudge.start = 0x40000000;
430 fudge.end = 0x40000000 + 128 * 1024 * 1024 - 1;
431 fudge.name = "fudge";
432 fudge.flags = IORESOURCE_MEM;
433 dev_priv->gtt_mem = &fudge;
436 pci_read_config_dword(pdev, PSB_BSM, &dev_priv->stolen_base);
437 vram_stolen_size = pg->gtt_phys_start - dev_priv->stolen_base
440 stolen_size = vram_stolen_size;
442 dev_dbg(dev->dev, "Stolen memory base 0x%x, size %luK\n",
443 dev_priv->stolen_base, vram_stolen_size / 1024);
445 if (resume && (gtt_pages != pg->gtt_pages) &&
446 (stolen_size != pg->stolen_size)) {
447 dev_err(dev->dev, "GTT resume error.\n");
452 pg->gtt_pages = gtt_pages;
453 pg->stolen_size = stolen_size;
454 dev_priv->vram_stolen_size = vram_stolen_size;
457 * Map the GTT and the stolen memory area
460 dev_priv->gtt_map = ioremap(pg->gtt_phys_start,
461 gtt_pages << PAGE_SHIFT);
462 if (!dev_priv->gtt_map) {
463 dev_err(dev->dev, "Failure to map gtt.\n");
469 dev_priv->vram_addr = ioremap_wc(dev_priv->stolen_base,
472 if (!dev_priv->vram_addr) {
473 dev_err(dev->dev, "Failure to map stolen base.\n");
479 * Insert vram stolen pages into the GTT
482 pfn_base = dev_priv->stolen_base >> PAGE_SHIFT;
483 num_pages = vram_stolen_size >> PAGE_SHIFT;
484 dev_dbg(dev->dev, "Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n",
485 num_pages, pfn_base << PAGE_SHIFT, 0);
486 for (i = 0; i < num_pages; ++i) {
487 pte = psb_gtt_mask_pte(pfn_base + i, PSB_MMU_CACHED_MEMORY);
488 iowrite32(pte, dev_priv->gtt_map + i);
492 * Init rest of GTT to the scratch page to avoid accidents or scribbles
495 pfn_base = page_to_pfn(dev_priv->scratch_page);
496 pte = psb_gtt_mask_pte(pfn_base, PSB_MMU_CACHED_MEMORY);
497 for (; i < gtt_pages; ++i)
498 iowrite32(pte, dev_priv->gtt_map + i);
500 (void) ioread32(dev_priv->gtt_map + i - 1);
504 psb_gtt_takedown(dev);
508 int psb_gtt_restore(struct drm_device *dev)
510 struct drm_psb_private *dev_priv = dev->dev_private;
511 struct resource *r = dev_priv->gtt_mem->child;
512 struct gtt_range *range;
513 unsigned int restored = 0, total = 0, size = 0;
515 /* On resume, the gtt_mutex is already initialized */
516 mutex_lock(&dev_priv->gtt_mutex);
517 psb_gtt_init(dev, 1);
520 range = container_of(r, struct gtt_range, resource);
522 psb_gtt_insert(dev, range, 1);
523 size += range->resource.end - range->resource.start;
529 mutex_unlock(&dev_priv->gtt_mutex);
530 DRM_DEBUG_DRIVER("Restored %u of %u gtt ranges (%u KB)", restored,
531 total, (size / 1024));