Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
  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 */
 39static 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 */
 65static 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 *	@resume: on resume
 80 *
 81 *	Take our preallocated GTT range and insert the GEM object into
 82 *	the GTT. This is protected via the gtt mutex which the caller
 83 *	must hold.
 84 */
 85static int psb_gtt_insert(struct drm_device *dev, struct gtt_range *r,
 86			  int resume)
 87{
 88	u32 __iomem *gtt_slot;
 89	u32 pte;
 90	struct page **pages;
 91	int i;
 92
 93	if (r->pages == NULL) {
 94		WARN_ON(1);
 95		return -EINVAL;
 96	}
 97
 98	WARN_ON(r->stolen);	/* refcount these maybe ? */
 99
100	gtt_slot = psb_gtt_entry(dev, r);
101	pages = r->pages;
102
103	if (!resume) {
104		/* Make sure changes are visible to the GPU */
105		set_pages_array_wc(pages, r->npage);
106	}
107
108	/* Write our page entries into the GTT itself */
109	for (i = r->roll; i < r->npage; i++) {
110		pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
111				       PSB_MMU_CACHED_MEMORY);
112		iowrite32(pte, gtt_slot++);
113	}
114	for (i = 0; i < r->roll; i++) {
115		pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
116				       PSB_MMU_CACHED_MEMORY);
117		iowrite32(pte, gtt_slot++);
118	}
119	/* Make sure all the entries are set before we return */
120	ioread32(gtt_slot - 1);
121
122	return 0;
123}
124
125/**
126 *	psb_gtt_remove	-	remove an object from the GTT
127 *	@dev: our DRM device
128 *	@r: our GTT range
129 *
130 *	Remove a preallocated GTT range from the GTT. Overwrite all the
131 *	page table entries with the dummy page. This is protected via the gtt
132 *	mutex which the caller must hold.
133 */
134static void psb_gtt_remove(struct drm_device *dev, struct gtt_range *r)
135{
136	struct drm_psb_private *dev_priv = dev->dev_private;
137	u32 __iomem *gtt_slot;
138	u32 pte;
139	int i;
140
141	WARN_ON(r->stolen);
142
143	gtt_slot = psb_gtt_entry(dev, r);
144	pte = psb_gtt_mask_pte(page_to_pfn(dev_priv->scratch_page),
145			       PSB_MMU_CACHED_MEMORY);
146
147	for (i = 0; i < r->npage; i++)
148		iowrite32(pte, gtt_slot++);
149	ioread32(gtt_slot - 1);
150	set_pages_array_wb(r->pages, r->npage);
151}
152
153/**
154 *	psb_gtt_roll	-	set scrolling position
155 *	@dev: our DRM device
156 *	@r: the gtt mapping we are using
157 *	@roll: roll offset
158 *
159 *	Roll an existing pinned mapping by moving the pages through the GTT.
160 *	This allows us to implement hardware scrolling on the consoles without
161 *	a 2D engine
162 */
163void psb_gtt_roll(struct drm_device *dev, struct gtt_range *r, int roll)
164{
165	u32 __iomem *gtt_slot;
166	u32 pte;
167	int i;
168
169	if (roll >= r->npage) {
170		WARN_ON(1);
171		return;
172	}
173
174	r->roll = roll;
175
176	/* Not currently in the GTT - no worry we will write the mapping at
177	   the right position when it gets pinned */
178	if (!r->stolen && !r->in_gart)
179		return;
180
181	gtt_slot = psb_gtt_entry(dev, r);
182
183	for (i = r->roll; i < r->npage; i++) {
184		pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
185				       PSB_MMU_CACHED_MEMORY);
186		iowrite32(pte, gtt_slot++);
187	}
188	for (i = 0; i < r->roll; i++) {
189		pte = psb_gtt_mask_pte(page_to_pfn(r->pages[i]),
190				       PSB_MMU_CACHED_MEMORY);
191		iowrite32(pte, gtt_slot++);
192	}
193	ioread32(gtt_slot - 1);
194}
195
196/**
197 *	psb_gtt_attach_pages	-	attach and pin GEM pages
198 *	@gt: the gtt range
199 *
200 *	Pin and build an in kernel list of the pages that back our GEM object.
201 *	While we hold this the pages cannot be swapped out. This is protected
202 *	via the gtt mutex which the caller must hold.
203 */
204static int psb_gtt_attach_pages(struct gtt_range *gt)
205{
206	struct page **pages;
207
208	WARN_ON(gt->pages);
209
210	pages = drm_gem_get_pages(&gt->gem);
211	if (IS_ERR(pages))
212		return PTR_ERR(pages);
213
214	gt->npage = gt->gem.size / PAGE_SIZE;
215	gt->pages = pages;
216
217	return 0;
218}
219
220/**
221 *	psb_gtt_detach_pages	-	attach and pin GEM pages
222 *	@gt: the gtt range
223 *
224 *	Undo the effect of psb_gtt_attach_pages. At this point the pages
225 *	must have been removed from the GTT as they could now be paged out
226 *	and move bus address. This is protected via the gtt mutex which the
227 *	caller must hold.
228 */
229static void psb_gtt_detach_pages(struct gtt_range *gt)
230{
231	drm_gem_put_pages(&gt->gem, gt->pages, true, false);
232	gt->pages = NULL;
233}
234
235/**
236 *	psb_gtt_pin		-	pin pages into the GTT
237 *	@gt: range to pin
238 *
239 *	Pin a set of pages into the GTT. The pins are refcounted so that
240 *	multiple pins need multiple unpins to undo.
241 *
242 *	Non GEM backed objects treat this as a no-op as they are always GTT
243 *	backed objects.
244 */
245int psb_gtt_pin(struct gtt_range *gt)
246{
247	int ret = 0;
248	struct drm_device *dev = gt->gem.dev;
249	struct drm_psb_private *dev_priv = dev->dev_private;
250	u32 gpu_base = dev_priv->gtt.gatt_start;
251
252	mutex_lock(&dev_priv->gtt_mutex);
253
254	if (gt->in_gart == 0 && gt->stolen == 0) {
255		ret = psb_gtt_attach_pages(gt);
256		if (ret < 0)
257			goto out;
258		ret = psb_gtt_insert(dev, gt, 0);
259		if (ret < 0) {
260			psb_gtt_detach_pages(gt);
261			goto out;
262		}
263		psb_mmu_insert_pages(psb_mmu_get_default_pd(dev_priv->mmu),
264				     gt->pages, (gpu_base + gt->offset),
265				     gt->npage, 0, 0, PSB_MMU_CACHED_MEMORY);
266	}
267	gt->in_gart++;
268out:
269	mutex_unlock(&dev_priv->gtt_mutex);
270	return ret;
271}
272
273/**
274 *	psb_gtt_unpin		-	Drop a GTT pin requirement
275 *	@gt: range to pin
276 *
277 *	Undoes the effect of psb_gtt_pin. On the last drop the GEM object
278 *	will be removed from the GTT which will also drop the page references
279 *	and allow the VM to clean up or page stuff.
280 *
281 *	Non GEM backed objects treat this as a no-op as they are always GTT
282 *	backed objects.
283 */
284void psb_gtt_unpin(struct gtt_range *gt)
285{
286	struct drm_device *dev = gt->gem.dev;
287	struct drm_psb_private *dev_priv = dev->dev_private;
288	u32 gpu_base = dev_priv->gtt.gatt_start;
289	int ret;
290
291	/* While holding the gtt_mutex no new blits can be initiated */
292	mutex_lock(&dev_priv->gtt_mutex);
293
294	/* Wait for any possible usage of the memory to be finished */
295	ret = gma_blt_wait_idle(dev_priv);
296	if (ret) {
297		DRM_ERROR("Failed to idle the blitter, unpin failed!");
298		goto out;
299	}
300
301	WARN_ON(!gt->in_gart);
302
303	gt->in_gart--;
304	if (gt->in_gart == 0 && gt->stolen == 0) {
305		psb_mmu_remove_pages(psb_mmu_get_default_pd(dev_priv->mmu),
306				     (gpu_base + gt->offset), gt->npage, 0, 0);
307		psb_gtt_remove(dev, gt);
308		psb_gtt_detach_pages(gt);
309	}
310
311out:
312	mutex_unlock(&dev_priv->gtt_mutex);
313}
314
315/*
316 *	GTT resource allocator - allocate and manage GTT address space
317 */
318
319/**
320 *	psb_gtt_alloc_range	-	allocate GTT address space
321 *	@dev: Our DRM device
322 *	@len: length (bytes) of address space required
323 *	@name: resource name
324 *	@backed: resource should be backed by stolen pages
325 *	@align: requested alignment
326 *
327 *	Ask the kernel core to find us a suitable range of addresses
328 *	to use for a GTT mapping.
329 *
330 *	Returns a gtt_range structure describing the object, or NULL on
331 *	error. On successful return the resource is both allocated and marked
332 *	as in use.
333 */
334struct gtt_range *psb_gtt_alloc_range(struct drm_device *dev, int len,
335				      const char *name, int backed, u32 align)
336{
337	struct drm_psb_private *dev_priv = dev->dev_private;
338	struct gtt_range *gt;
339	struct resource *r = dev_priv->gtt_mem;
340	int ret;
341	unsigned long start, end;
342
343	if (backed) {
344		/* The start of the GTT is the stolen pages */
345		start = r->start;
346		end = r->start + dev_priv->gtt.stolen_size - 1;
347	} else {
348		/* The rest we will use for GEM backed objects */
349		start = r->start + dev_priv->gtt.stolen_size;
350		end = r->end;
351	}
352
353	gt = kzalloc(sizeof(struct gtt_range), GFP_KERNEL);
354	if (gt == NULL)
355		return NULL;
356	gt->resource.name = name;
357	gt->stolen = backed;
358	gt->in_gart = backed;
359	gt->roll = 0;
360	/* Ensure this is set for non GEM objects */
361	gt->gem.dev = dev;
362	ret = allocate_resource(dev_priv->gtt_mem, &gt->resource,
363				len, start, end, align, NULL, NULL);
364	if (ret == 0) {
365		gt->offset = gt->resource.start - r->start;
366		return gt;
367	}
368	kfree(gt);
369	return NULL;
370}
371
372/**
373 *	psb_gtt_free_range	-	release GTT address space
374 *	@dev: our DRM device
375 *	@gt: a mapping created with psb_gtt_alloc_range
376 *
377 *	Release a resource that was allocated with psb_gtt_alloc_range. If the
378 *	object has been pinned by mmap users we clean this up here currently.
379 */
380void psb_gtt_free_range(struct drm_device *dev, struct gtt_range *gt)
381{
382	/* Undo the mmap pin if we are destroying the object */
383	if (gt->mmapping) {
384		psb_gtt_unpin(gt);
385		gt->mmapping = 0;
386	}
387	WARN_ON(gt->in_gart && !gt->stolen);
388	release_resource(&gt->resource);
389	kfree(gt);
390}
391
392static void psb_gtt_alloc(struct drm_device *dev)
393{
394	struct drm_psb_private *dev_priv = dev->dev_private;
395	init_rwsem(&dev_priv->gtt.sem);
396}
397
398void psb_gtt_takedown(struct drm_device *dev)
399{
400	struct drm_psb_private *dev_priv = dev->dev_private;
401
402	if (dev_priv->gtt_map) {
403		iounmap(dev_priv->gtt_map);
404		dev_priv->gtt_map = NULL;
405	}
406	if (dev_priv->gtt_initialized) {
407		pci_write_config_word(dev->pdev, PSB_GMCH_CTRL,
408				      dev_priv->gmch_ctrl);
409		PSB_WVDC32(dev_priv->pge_ctl, PSB_PGETBL_CTL);
410		(void) PSB_RVDC32(PSB_PGETBL_CTL);
411	}
412	if (dev_priv->vram_addr)
413		iounmap(dev_priv->gtt_map);
414}
415
416int psb_gtt_init(struct drm_device *dev, int resume)
417{
418	struct drm_psb_private *dev_priv = dev->dev_private;
419	unsigned gtt_pages;
420	unsigned long stolen_size, vram_stolen_size;
421	unsigned i, num_pages;
422	unsigned pfn_base;
423	struct psb_gtt *pg;
424
425	int ret = 0;
426	uint32_t pte;
427
428	if (!resume) {
429		mutex_init(&dev_priv->gtt_mutex);
430		mutex_init(&dev_priv->mmap_mutex);
431		psb_gtt_alloc(dev);
432	}
433
434	pg = &dev_priv->gtt;
435
436	/* Enable the GTT */
437	pci_read_config_word(dev->pdev, PSB_GMCH_CTRL, &dev_priv->gmch_ctrl);
438	pci_write_config_word(dev->pdev, PSB_GMCH_CTRL,
439			      dev_priv->gmch_ctrl | _PSB_GMCH_ENABLED);
440
441	dev_priv->pge_ctl = PSB_RVDC32(PSB_PGETBL_CTL);
442	PSB_WVDC32(dev_priv->pge_ctl | _PSB_PGETBL_ENABLED, PSB_PGETBL_CTL);
443	(void) PSB_RVDC32(PSB_PGETBL_CTL);
444
445	/* The root resource we allocate address space from */
446	dev_priv->gtt_initialized = 1;
447
448	pg->gtt_phys_start = dev_priv->pge_ctl & PAGE_MASK;
449
450	/*
451	 *	The video mmu has a hw bug when accessing 0x0D0000000.
452	 *	Make gatt start at 0x0e000,0000. This doesn't actually
453	 *	matter for us but may do if the video acceleration ever
454	 *	gets opened up.
455	 */
456	pg->mmu_gatt_start = 0xE0000000;
457
458	pg->gtt_start = pci_resource_start(dev->pdev, PSB_GTT_RESOURCE);
459	gtt_pages = pci_resource_len(dev->pdev, PSB_GTT_RESOURCE)
460								>> PAGE_SHIFT;
461	/* CDV doesn't report this. In which case the system has 64 gtt pages */
462	if (pg->gtt_start == 0 || gtt_pages == 0) {
463		dev_dbg(dev->dev, "GTT PCI BAR not initialized.\n");
464		gtt_pages = 64;
465		pg->gtt_start = dev_priv->pge_ctl;
466	}
467
468	pg->gatt_start = pci_resource_start(dev->pdev, PSB_GATT_RESOURCE);
469	pg->gatt_pages = pci_resource_len(dev->pdev, PSB_GATT_RESOURCE)
470								>> PAGE_SHIFT;
471	dev_priv->gtt_mem = &dev->pdev->resource[PSB_GATT_RESOURCE];
472
473	if (pg->gatt_pages == 0 || pg->gatt_start == 0) {
474		static struct resource fudge;	/* Preferably peppermint */
475		/* This can occur on CDV systems. Fudge it in this case.
476		   We really don't care what imaginary space is being allocated
477		   at this point */
478		dev_dbg(dev->dev, "GATT PCI BAR not initialized.\n");
479		pg->gatt_start = 0x40000000;
480		pg->gatt_pages = (128 * 1024 * 1024) >> PAGE_SHIFT;
481		/* This is a little confusing but in fact the GTT is providing
482		   a view from the GPU into memory and not vice versa. As such
483		   this is really allocating space that is not the same as the
484		   CPU address space on CDV */
485		fudge.start = 0x40000000;
486		fudge.end = 0x40000000 + 128 * 1024 * 1024 - 1;
487		fudge.name = "fudge";
488		fudge.flags = IORESOURCE_MEM;
489		dev_priv->gtt_mem = &fudge;
490	}
491
492	pci_read_config_dword(dev->pdev, PSB_BSM, &dev_priv->stolen_base);
493	vram_stolen_size = pg->gtt_phys_start - dev_priv->stolen_base
494								- PAGE_SIZE;
495
496	stolen_size = vram_stolen_size;
497
498	dev_dbg(dev->dev, "Stolen memory base 0x%x, size %luK\n",
499			dev_priv->stolen_base, vram_stolen_size / 1024);
500
501	if (resume && (gtt_pages != pg->gtt_pages) &&
502	    (stolen_size != pg->stolen_size)) {
503		dev_err(dev->dev, "GTT resume error.\n");
504		ret = -EINVAL;
505		goto out_err;
506	}
507
508	pg->gtt_pages = gtt_pages;
509	pg->stolen_size = stolen_size;
510	dev_priv->vram_stolen_size = vram_stolen_size;
511
512	/*
513	 *	Map the GTT and the stolen memory area
514	 */
515	if (!resume)
516		dev_priv->gtt_map = ioremap_nocache(pg->gtt_phys_start,
517						gtt_pages << PAGE_SHIFT);
518	if (!dev_priv->gtt_map) {
519		dev_err(dev->dev, "Failure to map gtt.\n");
520		ret = -ENOMEM;
521		goto out_err;
522	}
523
524	if (!resume)
525		dev_priv->vram_addr = ioremap_wc(dev_priv->stolen_base,
526						 stolen_size);
527
528	if (!dev_priv->vram_addr) {
529		dev_err(dev->dev, "Failure to map stolen base.\n");
530		ret = -ENOMEM;
531		goto out_err;
532	}
533
534	/*
535	 * Insert vram stolen pages into the GTT
536	 */
537
538	pfn_base = dev_priv->stolen_base >> PAGE_SHIFT;
539	num_pages = vram_stolen_size >> PAGE_SHIFT;
540	dev_dbg(dev->dev, "Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n",
541		num_pages, pfn_base << PAGE_SHIFT, 0);
542	for (i = 0; i < num_pages; ++i) {
543		pte = psb_gtt_mask_pte(pfn_base + i, PSB_MMU_CACHED_MEMORY);
544		iowrite32(pte, dev_priv->gtt_map + i);
545	}
546
547	/*
548	 * Init rest of GTT to the scratch page to avoid accidents or scribbles
549	 */
550
551	pfn_base = page_to_pfn(dev_priv->scratch_page);
552	pte = psb_gtt_mask_pte(pfn_base, PSB_MMU_CACHED_MEMORY);
553	for (; i < gtt_pages; ++i)
554		iowrite32(pte, dev_priv->gtt_map + i);
555
556	(void) ioread32(dev_priv->gtt_map + i - 1);
557	return 0;
558
559out_err:
560	psb_gtt_takedown(dev);
561	return ret;
562}
563
564int psb_gtt_restore(struct drm_device *dev)
565{
566	struct drm_psb_private *dev_priv = dev->dev_private;
567	struct resource *r = dev_priv->gtt_mem->child;
568	struct gtt_range *range;
569	unsigned int restored = 0, total = 0, size = 0;
570
571	/* On resume, the gtt_mutex is already initialized */
572	mutex_lock(&dev_priv->gtt_mutex);
573	psb_gtt_init(dev, 1);
574
575	while (r != NULL) {
576		range = container_of(r, struct gtt_range, resource);
577		if (range->pages) {
578			psb_gtt_insert(dev, range, 1);
579			size += range->resource.end - range->resource.start;
580			restored++;
581		}
582		r = r->sibling;
583		total++;
584	}
585	mutex_unlock(&dev_priv->gtt_mutex);
586	DRM_DEBUG_DRIVER("Restored %u of %u gtt ranges (%u KB)", restored,
587			 total, (size / 1024));
588
589	return 0;
590}