Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
  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 "psb_drv.h"
 15
 16
 17/*
 18 *	GTT resource allocator - manage page mappings in GTT space
 19 */
 20
 21/**
 22 *	psb_gtt_mask_pte	-	generate GTT pte entry
 23 *	@pfn: page number to encode
 24 *	@type: type of memory in the GTT
 25 *
 26 *	Set the GTT entry for the appropriate memory type.
 27 */
 28static inline uint32_t psb_gtt_mask_pte(uint32_t pfn, int type)
 29{
 30	uint32_t mask = PSB_PTE_VALID;
 31
 32	/* Ensure we explode rather than put an invalid low mapping of
 33	   a high mapping page into the gtt */
 34	BUG_ON(pfn & ~(0xFFFFFFFF >> PAGE_SHIFT));
 35
 36	if (type & PSB_MMU_CACHED_MEMORY)
 37		mask |= PSB_PTE_CACHED;
 38	if (type & PSB_MMU_RO_MEMORY)
 39		mask |= PSB_PTE_RO;
 40	if (type & PSB_MMU_WO_MEMORY)
 41		mask |= PSB_PTE_WO;
 42
 43	return (pfn << PAGE_SHIFT) | mask;
 44}
 45
 46/**
 47 *	psb_gtt_entry		-	find the GTT entries for a gtt_range
 48 *	@dev: our DRM device
 49 *	@r: our GTT range
 50 *
 51 *	Given a gtt_range object return the GTT offset of the page table
 52 *	entries for this gtt_range
 53 */
 54static u32 __iomem *psb_gtt_entry(struct drm_device *dev, struct gtt_range *r)
 55{
 56	struct drm_psb_private *dev_priv = dev->dev_private;
 57	unsigned long offset;
 58
 59	offset = r->resource.start - dev_priv->gtt_mem->start;
 60
 61	return dev_priv->gtt_map + (offset >> PAGE_SHIFT);
 62}
 63
 64/**
 65 *	psb_gtt_insert	-	put an object into the GTT
 66 *	@dev: our DRM device
 67 *	@r: our GTT range
 68 *	@resume: on resume
 69 *
 70 *	Take our preallocated GTT range and insert the GEM object into
 71 *	the GTT. This is protected via the gtt mutex which the caller
 72 *	must hold.
 73 */
 74static int psb_gtt_insert(struct drm_device *dev, struct gtt_range *r,
 75			  int resume)
 76{
 77	u32 __iomem *gtt_slot;
 78	u32 pte;
 79	struct page **pages;
 80	int i;
 81
 82	if (r->pages == NULL) {
 83		WARN_ON(1);
 84		return -EINVAL;
 85	}
 86
 87	WARN_ON(r->stolen);	/* refcount these maybe ? */
 88
 89	gtt_slot = psb_gtt_entry(dev, r);
 90	pages = r->pages;
 91
 92	if (!resume) {
 93		/* Make sure changes are visible to the GPU */
 94		set_pages_array_wc(pages, r->npage);
 95	}
 96
 97	/* Write our page entries into the GTT itself */
 98	for (i = 0; i < r->npage; i++) {
 99		pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
100				       PSB_MMU_CACHED_MEMORY);
101		iowrite32(pte, gtt_slot++);
102	}
103
104	/* Make sure all the entries are set before we return */
105	ioread32(gtt_slot - 1);
106
107	return 0;
108}
109
110/**
111 *	psb_gtt_remove	-	remove an object from the GTT
112 *	@dev: our DRM device
113 *	@r: our GTT range
114 *
115 *	Remove a preallocated GTT range from the GTT. Overwrite all the
116 *	page table entries with the dummy page. This is protected via the gtt
117 *	mutex which the caller must hold.
118 */
119static void psb_gtt_remove(struct drm_device *dev, struct gtt_range *r)
120{
121	struct drm_psb_private *dev_priv = dev->dev_private;
122	u32 __iomem *gtt_slot;
123	u32 pte;
124	int i;
125
126	WARN_ON(r->stolen);
127
128	gtt_slot = psb_gtt_entry(dev, r);
129	pte = psb_gtt_mask_pte(page_to_pfn(dev_priv->scratch_page),
130			       PSB_MMU_CACHED_MEMORY);
131
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);
136}
137
138/**
139 *	psb_gtt_attach_pages	-	attach and pin GEM pages
140 *	@gt: the gtt range
141 *
142 *	Pin and build an in kernel list of the pages that back our GEM object.
143 *	While we hold this the pages cannot be swapped out. This is protected
144 *	via the gtt mutex which the caller must hold.
145 */
146static int psb_gtt_attach_pages(struct gtt_range *gt)
147{
148	struct page **pages;
149
150	WARN_ON(gt->pages);
151
152	pages = drm_gem_get_pages(&gt->gem);
153	if (IS_ERR(pages))
154		return PTR_ERR(pages);
155
156	gt->npage = gt->gem.size / PAGE_SIZE;
157	gt->pages = pages;
158
159	return 0;
160}
161
162/**
163 *	psb_gtt_detach_pages	-	attach and pin GEM pages
164 *	@gt: the gtt range
165 *
166 *	Undo the effect of psb_gtt_attach_pages. At this point the pages
167 *	must have been removed from the GTT as they could now be paged out
168 *	and move bus address. This is protected via the gtt mutex which the
169 *	caller must hold.
170 */
171static void psb_gtt_detach_pages(struct gtt_range *gt)
172{
173	drm_gem_put_pages(&gt->gem, gt->pages, true, false);
174	gt->pages = NULL;
175}
176
177/**
178 *	psb_gtt_pin		-	pin pages into the GTT
179 *	@gt: range to pin
180 *
181 *	Pin a set of pages into the GTT. The pins are refcounted so that
182 *	multiple pins need multiple unpins to undo.
183 *
184 *	Non GEM backed objects treat this as a no-op as they are always GTT
185 *	backed objects.
186 */
187int psb_gtt_pin(struct gtt_range *gt)
188{
189	int ret = 0;
190	struct drm_device *dev = gt->gem.dev;
191	struct drm_psb_private *dev_priv = dev->dev_private;
192	u32 gpu_base = dev_priv->gtt.gatt_start;
193
194	mutex_lock(&dev_priv->gtt_mutex);
195
196	if (gt->in_gart == 0 && gt->stolen == 0) {
197		ret = psb_gtt_attach_pages(gt);
198		if (ret < 0)
199			goto out;
200		ret = psb_gtt_insert(dev, gt, 0);
201		if (ret < 0) {
202			psb_gtt_detach_pages(gt);
203			goto out;
204		}
205		psb_mmu_insert_pages(psb_mmu_get_default_pd(dev_priv->mmu),
206				     gt->pages, (gpu_base + gt->offset),
207				     gt->npage, 0, 0, PSB_MMU_CACHED_MEMORY);
208	}
209	gt->in_gart++;
210out:
211	mutex_unlock(&dev_priv->gtt_mutex);
212	return ret;
213}
214
215/**
216 *	psb_gtt_unpin		-	Drop a GTT pin requirement
217 *	@gt: range to pin
218 *
219 *	Undoes the effect of psb_gtt_pin. On the last drop the GEM object
220 *	will be removed from the GTT which will also drop the page references
221 *	and allow the VM to clean up or page stuff.
222 *
223 *	Non GEM backed objects treat this as a no-op as they are always GTT
224 *	backed objects.
225 */
226void psb_gtt_unpin(struct gtt_range *gt)
227{
228	struct drm_device *dev = gt->gem.dev;
229	struct drm_psb_private *dev_priv = dev->dev_private;
230	u32 gpu_base = dev_priv->gtt.gatt_start;
231
232	mutex_lock(&dev_priv->gtt_mutex);
233
234	WARN_ON(!gt->in_gart);
235
236	gt->in_gart--;
237	if (gt->in_gart == 0 && gt->stolen == 0) {
238		psb_mmu_remove_pages(psb_mmu_get_default_pd(dev_priv->mmu),
239				     (gpu_base + gt->offset), gt->npage, 0, 0);
240		psb_gtt_remove(dev, gt);
241		psb_gtt_detach_pages(gt);
242	}
243
244	mutex_unlock(&dev_priv->gtt_mutex);
245}
246
247/*
248 *	GTT resource allocator - allocate and manage GTT address space
249 */
250
251/**
252 *	psb_gtt_alloc_range	-	allocate GTT address space
253 *	@dev: Our DRM device
254 *	@len: length (bytes) of address space required
255 *	@name: resource name
256 *	@backed: resource should be backed by stolen pages
257 *	@align: requested alignment
258 *
259 *	Ask the kernel core to find us a suitable range of addresses
260 *	to use for a GTT mapping.
261 *
262 *	Returns a gtt_range structure describing the object, or NULL on
263 *	error. On successful return the resource is both allocated and marked
264 *	as in use.
265 */
266struct gtt_range *psb_gtt_alloc_range(struct drm_device *dev, int len,
267				      const char *name, int backed, u32 align)
268{
269	struct drm_psb_private *dev_priv = dev->dev_private;
270	struct gtt_range *gt;
271	struct resource *r = dev_priv->gtt_mem;
272	int ret;
273	unsigned long start, end;
274
275	if (backed) {
276		/* The start of the GTT is the stolen pages */
277		start = r->start;
278		end = r->start + dev_priv->gtt.stolen_size - 1;
279	} else {
280		/* The rest we will use for GEM backed objects */
281		start = r->start + dev_priv->gtt.stolen_size;
282		end = r->end;
283	}
284
285	gt = kzalloc(sizeof(struct gtt_range), GFP_KERNEL);
286	if (gt == NULL)
287		return NULL;
288	gt->resource.name = name;
289	gt->stolen = backed;
290	gt->in_gart = backed;
291	/* Ensure this is set for non GEM objects */
292	gt->gem.dev = dev;
293	ret = allocate_resource(dev_priv->gtt_mem, &gt->resource,
294				len, start, end, align, NULL, NULL);
295	if (ret == 0) {
296		gt->offset = gt->resource.start - r->start;
297		return gt;
298	}
299	kfree(gt);
300	return NULL;
301}
302
303/**
304 *	psb_gtt_free_range	-	release GTT address space
305 *	@dev: our DRM device
306 *	@gt: a mapping created with psb_gtt_alloc_range
307 *
308 *	Release a resource that was allocated with psb_gtt_alloc_range. If the
309 *	object has been pinned by mmap users we clean this up here currently.
310 */
311void psb_gtt_free_range(struct drm_device *dev, struct gtt_range *gt)
312{
313	/* Undo the mmap pin if we are destroying the object */
314	if (gt->mmapping) {
315		psb_gtt_unpin(gt);
316		gt->mmapping = 0;
317	}
318	WARN_ON(gt->in_gart && !gt->stolen);
319	release_resource(&gt->resource);
320	kfree(gt);
321}
322
323static void psb_gtt_alloc(struct drm_device *dev)
324{
325	struct drm_psb_private *dev_priv = dev->dev_private;
326	init_rwsem(&dev_priv->gtt.sem);
327}
328
329void psb_gtt_takedown(struct drm_device *dev)
330{
331	struct drm_psb_private *dev_priv = dev->dev_private;
332	struct pci_dev *pdev = to_pci_dev(dev->dev);
333
334	if (dev_priv->gtt_map) {
335		iounmap(dev_priv->gtt_map);
336		dev_priv->gtt_map = NULL;
337	}
338	if (dev_priv->gtt_initialized) {
339		pci_write_config_word(pdev, PSB_GMCH_CTRL,
340				      dev_priv->gmch_ctrl);
341		PSB_WVDC32(dev_priv->pge_ctl, PSB_PGETBL_CTL);
342		(void) PSB_RVDC32(PSB_PGETBL_CTL);
343	}
344	if (dev_priv->vram_addr)
345		iounmap(dev_priv->gtt_map);
346}
347
348int psb_gtt_init(struct drm_device *dev, int resume)
349{
350	struct drm_psb_private *dev_priv = dev->dev_private;
351	struct pci_dev *pdev = to_pci_dev(dev->dev);
352	unsigned gtt_pages;
353	unsigned long stolen_size, vram_stolen_size;
354	unsigned i, num_pages;
355	unsigned pfn_base;
356	struct psb_gtt *pg;
357
358	int ret = 0;
359	uint32_t pte;
360
361	if (!resume) {
362		mutex_init(&dev_priv->gtt_mutex);
363		mutex_init(&dev_priv->mmap_mutex);
364		psb_gtt_alloc(dev);
365	}
366
367	pg = &dev_priv->gtt;
368
369	/* Enable the GTT */
370	pci_read_config_word(pdev, PSB_GMCH_CTRL, &dev_priv->gmch_ctrl);
371	pci_write_config_word(pdev, PSB_GMCH_CTRL,
372			      dev_priv->gmch_ctrl | _PSB_GMCH_ENABLED);
373
374	dev_priv->pge_ctl = PSB_RVDC32(PSB_PGETBL_CTL);
375	PSB_WVDC32(dev_priv->pge_ctl | _PSB_PGETBL_ENABLED, PSB_PGETBL_CTL);
376	(void) PSB_RVDC32(PSB_PGETBL_CTL);
377
378	/* The root resource we allocate address space from */
379	dev_priv->gtt_initialized = 1;
380
381	pg->gtt_phys_start = dev_priv->pge_ctl & PAGE_MASK;
382
383	/*
384	 *	The video mmu has a hw bug when accessing 0x0D0000000.
385	 *	Make gatt start at 0x0e000,0000. This doesn't actually
386	 *	matter for us but may do if the video acceleration ever
387	 *	gets opened up.
388	 */
389	pg->mmu_gatt_start = 0xE0000000;
390
391	pg->gtt_start = pci_resource_start(pdev, PSB_GTT_RESOURCE);
392	gtt_pages = pci_resource_len(pdev, PSB_GTT_RESOURCE)
393								>> PAGE_SHIFT;
394	/* CDV doesn't report this. In which case the system has 64 gtt pages */
395	if (pg->gtt_start == 0 || gtt_pages == 0) {
396		dev_dbg(dev->dev, "GTT PCI BAR not initialized.\n");
397		gtt_pages = 64;
398		pg->gtt_start = dev_priv->pge_ctl;
399	}
400
401	pg->gatt_start = pci_resource_start(pdev, PSB_GATT_RESOURCE);
402	pg->gatt_pages = pci_resource_len(pdev, PSB_GATT_RESOURCE)
403								>> PAGE_SHIFT;
404	dev_priv->gtt_mem = &pdev->resource[PSB_GATT_RESOURCE];
405
406	if (pg->gatt_pages == 0 || pg->gatt_start == 0) {
407		static struct resource fudge;	/* Preferably peppermint */
408		/* This can occur on CDV systems. Fudge it in this case.
409		   We really don't care what imaginary space is being allocated
410		   at this point */
411		dev_dbg(dev->dev, "GATT PCI BAR not initialized.\n");
412		pg->gatt_start = 0x40000000;
413		pg->gatt_pages = (128 * 1024 * 1024) >> PAGE_SHIFT;
414		/* This is a little confusing but in fact the GTT is providing
415		   a view from the GPU into memory and not vice versa. As such
416		   this is really allocating space that is not the same as the
417		   CPU address space on CDV */
418		fudge.start = 0x40000000;
419		fudge.end = 0x40000000 + 128 * 1024 * 1024 - 1;
420		fudge.name = "fudge";
421		fudge.flags = IORESOURCE_MEM;
422		dev_priv->gtt_mem = &fudge;
423	}
424
425	pci_read_config_dword(pdev, PSB_BSM, &dev_priv->stolen_base);
426	vram_stolen_size = pg->gtt_phys_start - dev_priv->stolen_base
427								- PAGE_SIZE;
428
429	stolen_size = vram_stolen_size;
430
431	dev_dbg(dev->dev, "Stolen memory base 0x%x, size %luK\n",
432			dev_priv->stolen_base, vram_stolen_size / 1024);
433
434	if (resume && (gtt_pages != pg->gtt_pages) &&
435	    (stolen_size != pg->stolen_size)) {
436		dev_err(dev->dev, "GTT resume error.\n");
437		ret = -EINVAL;
438		goto out_err;
439	}
440
441	pg->gtt_pages = gtt_pages;
442	pg->stolen_size = stolen_size;
443	dev_priv->vram_stolen_size = vram_stolen_size;
444
445	/*
446	 *	Map the GTT and the stolen memory area
447	 */
448	if (!resume)
449		dev_priv->gtt_map = ioremap(pg->gtt_phys_start,
450						gtt_pages << PAGE_SHIFT);
451	if (!dev_priv->gtt_map) {
452		dev_err(dev->dev, "Failure to map gtt.\n");
453		ret = -ENOMEM;
454		goto out_err;
455	}
456
457	if (!resume)
458		dev_priv->vram_addr = ioremap_wc(dev_priv->stolen_base,
459						 stolen_size);
460
461	if (!dev_priv->vram_addr) {
462		dev_err(dev->dev, "Failure to map stolen base.\n");
463		ret = -ENOMEM;
464		goto out_err;
465	}
466
467	/*
468	 * Insert vram stolen pages into the GTT
469	 */
470
471	pfn_base = dev_priv->stolen_base >> PAGE_SHIFT;
472	num_pages = vram_stolen_size >> PAGE_SHIFT;
473	dev_dbg(dev->dev, "Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n",
474		num_pages, pfn_base << PAGE_SHIFT, 0);
475	for (i = 0; i < num_pages; ++i) {
476		pte = psb_gtt_mask_pte(pfn_base + i, PSB_MMU_CACHED_MEMORY);
477		iowrite32(pte, dev_priv->gtt_map + i);
478	}
479
480	/*
481	 * Init rest of GTT to the scratch page to avoid accidents or scribbles
482	 */
483
484	pfn_base = page_to_pfn(dev_priv->scratch_page);
485	pte = psb_gtt_mask_pte(pfn_base, PSB_MMU_CACHED_MEMORY);
486	for (; i < gtt_pages; ++i)
487		iowrite32(pte, dev_priv->gtt_map + i);
488
489	(void) ioread32(dev_priv->gtt_map + i - 1);
490	return 0;
491
492out_err:
493	psb_gtt_takedown(dev);
494	return ret;
495}
496
497int psb_gtt_restore(struct drm_device *dev)
498{
499	struct drm_psb_private *dev_priv = dev->dev_private;
500	struct resource *r = dev_priv->gtt_mem->child;
501	struct gtt_range *range;
502	unsigned int restored = 0, total = 0, size = 0;
503
504	/* On resume, the gtt_mutex is already initialized */
505	mutex_lock(&dev_priv->gtt_mutex);
506	psb_gtt_init(dev, 1);
507
508	while (r != NULL) {
509		range = container_of(r, struct gtt_range, resource);
510		if (range->pages) {
511			psb_gtt_insert(dev, range, 1);
512			size += range->resource.end - range->resource.start;
513			restored++;
514		}
515		r = r->sibling;
516		total++;
517	}
518	mutex_unlock(&dev_priv->gtt_mutex);
519	DRM_DEBUG_DRIVER("Restored %u of %u gtt ranges (%u KB)", restored,
520			 total, (size / 1024));
521
522	return 0;
523}