summaryrefslogtreecommitdiff
path: root/arch/x86/kernel/pci-gart_64.c
blob: 9c557c0c928c32e2ae6ddcd4c775db6fed4e6bd9 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
/*
 * Dynamic DMA mapping support for AMD Hammer.
 *
 * Use the integrated AGP GART in the Hammer northbridge as an IOMMU for PCI.
 * This allows to use PCI devices that only support 32bit addresses on systems
 * with more than 4GB.
 *
 * See Documentation/DMA-mapping.txt for the interface specification.
 *
 * Copyright 2002 Andi Kleen, SuSE Labs.
 * Subject to the GNU General Public License v2 only.
 */

#include <linux/types.h>
#include <linux/ctype.h>
#include <linux/agp_backend.h>
#include <linux/init.h>
#include <linux/mm.h>
#include <linux/string.h>
#include <linux/spinlock.h>
#include <linux/pci.h>
#include <linux/module.h>
#include <linux/topology.h>
#include <linux/interrupt.h>
#include <linux/bitops.h>
#include <linux/kdebug.h>
#include <linux/scatterlist.h>
#include <linux/iommu-helper.h>
#include <linux/sysdev.h>
#include <linux/io.h>
#include <asm/atomic.h>
#include <asm/mtrr.h>
#include <asm/pgtable.h>
#include <asm/proto.h>
#include <asm/iommu.h>
#include <asm/gart.h>
#include <asm/cacheflush.h>
#include <asm/swiotlb.h>
#include <asm/dma.h>
#include <asm/k8.h>

static unsigned long iommu_bus_base;	/* GART remapping area (physical) */
static unsigned long iommu_size;	/* size of remapping area bytes */
static unsigned long iommu_pages;	/* .. and in pages */

static u32 *iommu_gatt_base;		/* Remapping table */

/*
 * If this is disabled the IOMMU will use an optimized flushing strategy
 * of only flushing when an mapping is reused. With it true the GART is
 * flushed for every mapping. Problem is that doing the lazy flush seems
 * to trigger bugs with some popular PCI cards, in particular 3ware (but
 * has been also also seen with Qlogic at least).
 */
static int iommu_fullflush = 1;

/* Allocation bitmap for the remapping area: */
static DEFINE_SPINLOCK(iommu_bitmap_lock);
/* Guarded by iommu_bitmap_lock: */
static unsigned long *iommu_gart_bitmap;

static u32 gart_unmapped_entry;

#define GPTE_VALID    1
#define GPTE_COHERENT 2
#define GPTE_ENCODE(x) \
	(((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
#define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))

#define EMERGENCY_PAGES 32 /* = 128KB */

#ifdef CONFIG_AGP
#define AGPEXTERN extern
#else
#define AGPEXTERN
#endif

/* backdoor interface to AGP driver */
AGPEXTERN int agp_memory_reserved;
AGPEXTERN __u32 *agp_gatt_table;

static unsigned long next_bit;  /* protected by iommu_bitmap_lock */
static bool need_flush;		/* global flush state. set for each gart wrap */

static unsigned long alloc_iommu(struct device *dev, int size,
				 unsigned long align_mask)
{
	unsigned long offset, flags;
	unsigned long boundary_size;
	unsigned long base_index;

	base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
			   PAGE_SIZE) >> PAGE_SHIFT;
	boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1,
			      PAGE_SIZE) >> PAGE_SHIFT;

	spin_lock_irqsave(&iommu_bitmap_lock, flags);
	offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
				  size, base_index, boundary_size, align_mask);
	if (offset == -1) {
		need_flush = true;
		offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
					  size, base_index, boundary_size,
					  align_mask);
	}
	if (offset != -1) {
		next_bit = offset+size;
		if (next_bit >= iommu_pages) {
			next_bit = 0;
			need_flush = true;
		}
	}
	if (iommu_fullflush)
		need_flush = true;
	spin_unlock_irqrestore(&iommu_bitmap_lock, flags);

	return offset;
}

static void free_iommu(unsigned long offset, int size)
{
	unsigned long flags;

	spin_lock_irqsave(&iommu_bitmap_lock, flags);
	iommu_area_free(iommu_gart_bitmap, offset, size);
	if (offset >= next_bit)
		next_bit = offset + size;
	spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
}

/*
 * Use global flush state to avoid races with multiple flushers.
 */
static void flush_gart(void)
{
	unsigned long flags;

	spin_lock_irqsave(&iommu_bitmap_lock, flags);
	if (need_flush) {
		k8_flush_garts();
		need_flush = false;
	}
	spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
}

#ifdef CONFIG_IOMMU_LEAK

#define SET_LEAK(x)							\
	do {								\
		if (iommu_leak_tab)					\
			iommu_leak_tab[x] = __builtin_return_address(0);\
	} while (0)

#define CLEAR_LEAK(x)							\
	do {								\
		if (iommu_leak_tab)					\
			iommu_leak_tab[x] = NULL;			\
	} while (0)

/* Debugging aid for drivers that don't free their IOMMU tables */
static void **iommu_leak_tab;
static int leak_trace;
static int iommu_leak_pages = 20;

static void dump_leak(void)
{
	int i;
	static int dump;

	if (dump || !iommu_leak_tab)
		return;
	dump = 1;
	show_stack(NULL, NULL);

	/* Very crude. dump some from the end of the table too */
	printk(KERN_DEBUG "Dumping %d pages from end of IOMMU:\n",
	       iommu_leak_pages);
	for (i = 0; i < iommu_leak_pages; i += 2) {
		printk(KERN_DEBUG "%lu: ", iommu_pages-i);
		printk_address((unsigned long) iommu_leak_tab[iommu_pages-i],
				0);
		printk(KERN_CONT "%c", (i+1)%2 == 0 ? '\n' : ' ');
	}
	printk(KERN_DEBUG "\n");
}
#else
# define SET_LEAK(x)
# define CLEAR_LEAK(x)
#endif

static void iommu_full(struct device *dev, size_t size, int dir)
{
	/*
	 * Ran out of IOMMU space for this operation. This is very bad.
	 * Unfortunately the drivers cannot handle this operation properly.
	 * Return some non mapped prereserved space in the aperture and
	 * let the Northbridge deal with it. This will result in garbage
	 * in the IO operation. When the size exceeds the prereserved space
	 * memory corruption will occur or random memory will be DMAed
	 * out. Hopefully no network devices use single mappings that big.
	 */

	dev_err(dev, "PCI-DMA: Out of IOMMU space for %lu bytes\n", size);

	if (size > PAGE_SIZE*EMERGENCY_PAGES) {
		if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
			panic("PCI-DMA: Memory would be corrupted\n");
		if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
			panic(KERN_ERR
				"PCI-DMA: Random memory would be DMAed\n");
	}
#ifdef CONFIG_IOMMU_LEAK
	dump_leak();
#endif
}

static inline int
need_iommu(struct device *dev, unsigned long addr, size_t size)
{
	return force_iommu ||
		!is_buffer_dma_capable(*dev->dma_mask, addr, size);
}

static inline int
nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
{
	return !is_buffer_dma_capable(*dev->dma_mask, addr, size);
}

/* Map a single continuous physical area into the IOMMU.
 * Caller needs to check if the iommu is needed and flush.
 */
static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
				size_t size, int dir, unsigned long align_mask)
{
	unsigned long npages = iommu_num_pages(phys_mem, size, PAGE_SIZE);
	unsigned long iommu_page = alloc_iommu(dev, npages, align_mask);
	int i;

	if (iommu_page == -1) {
		if (!nonforced_iommu(dev, phys_mem, size))
			return phys_mem;
		if (panic_on_overflow)
			panic("dma_map_area overflow %lu bytes\n", size);
		iommu_full(dev, size, dir);
		return bad_dma_address;
	}

	for (i = 0; i < npages; i++) {
		iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
		SET_LEAK(iommu_page + i);
		phys_mem += PAGE_SIZE;
	}
	return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
}

/* Map a single area into the IOMMU */
static dma_addr_t gart_map_page(struct device *dev, struct page *page,
				unsigned long offset, size_t size,
				enum dma_data_direction dir,
				struct dma_attrs *attrs)
{
	unsigned long bus;
	phys_addr_t paddr = page_to_phys(page) + offset;

	if (!dev)
		dev = &x86_dma_fallback_dev;

	if (!need_iommu(dev, paddr, size))
		return paddr;

	bus = dma_map_area(dev, paddr, size, dir, 0);
	flush_gart();

	return bus;
}

/*
 * Free a DMA mapping.
 */
static void gart_unmap_page(struct device *dev, dma_addr_t dma_addr,
			    size_t size, enum dma_data_direction dir,
			    struct dma_attrs *attrs)
{
	unsigned long iommu_page;
	int npages;
	int i;

	if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
	    dma_addr >= iommu_bus_base + iommu_size)
		return;

	iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
	npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
	for (i = 0; i < npages; i++) {
		iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
		CLEAR_LEAK(iommu_page + i);
	}
	free_iommu(iommu_page, npages);
}

/*
 * Wrapper for pci_unmap_single working with scatterlists.
 */
static void
gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
{
	struct scatterlist *s;
	int i;

	for_each_sg(sg, s, nents, i) {
		if (!s->dma_length || !s->length)
			break;
		gart_unmap_page(dev, s->dma_address, s->dma_length, dir, NULL);
	}
}

/* Fallback for dma_map_sg in case of overflow */
static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
			       int nents, int dir)
{
	struct scatterlist *s;
	int i;

#ifdef CONFIG_IOMMU_DEBUG
	printk(KERN_DEBUG "dma_map_sg overflow\n");
#endif

	for_each_sg(sg, s, nents, i) {
		unsigned long addr = sg_phys(s);

		if (nonforced_iommu(dev, addr, s->length)) {
			addr = dma_map_area(dev, addr, s->length, dir, 0);
			if (addr == bad_dma_address) {
				if (i > 0)
					gart_unmap_sg(dev, sg, i, dir);
				nents = 0;
				sg[0].dma_length = 0;
				break;
			}
		}
		s->dma_address = addr;
		s->dma_length = s->length;
	}
	flush_gart();

	return nents;
}

/* Map multiple scatterlist entries continuous into the first. */
static int __dma_map_cont(struct device *dev, struct scatterlist *start,
			  int nelems, struct scatterlist *sout,
			  unsigned long pages)
{
	unsigned long iommu_start = alloc_iommu(dev, pages, 0);
	unsigned long iommu_page = iommu_start;
	struct scatterlist *s;
	int i;

	if (iommu_start == -1)
		return -1;

	for_each_sg(start, s, nelems, i) {
		unsigned long pages, addr;
		unsigned long phys_addr = s->dma_address;

		BUG_ON(s != start && s->offset);
		if (s == start) {
			sout->dma_address = iommu_bus_base;
			sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
			sout->dma_length = s->length;
		} else {
			sout->dma_length += s->length;
		}

		addr = phys_addr;
		pages = iommu_num_pages(s->offset, s->length, PAGE_SIZE);
		while (pages--) {
			iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
			SET_LEAK(iommu_page);
			addr += PAGE_SIZE;
			iommu_page++;
		}
	}
	BUG_ON(iommu_page - iommu_start != pages);

	return 0;
}

static inline int
dma_map_cont(struct device *dev, struct scatterlist *start, int nelems,
	     struct scatterlist *sout, unsigned long pages, int need)
{
	if (!need) {
		BUG_ON(nelems != 1);
		sout->dma_address = start->dma_address;
		sout->dma_length = start->length;
		return 0;
	}
	return __dma_map_cont(dev, start, nelems, sout, pages);
}

/*
 * DMA map all entries in a scatterlist.
 * Merge chunks that have page aligned sizes into a continuous mapping.
 */
static int
gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
{
	struct scatterlist *s, *ps, *start_sg, *sgmap;
	int need = 0, nextneed, i, out, start;
	unsigned long pages = 0;
	unsigned int seg_size;
	unsigned int max_seg_size;

	if (nents == 0)
		return 0;

	if (!dev)
		dev = &x86_dma_fallback_dev;

	out = 0;
	start = 0;
	start_sg = sgmap = sg;
	seg_size = 0;
	max_seg_size = dma_get_max_seg_size(dev);
	ps = NULL; /* shut up gcc */
	for_each_sg(sg, s, nents, i) {
		dma_addr_t addr = sg_phys(s);

		s->dma_address = addr;
		BUG_ON(s->length == 0);

		nextneed = need_iommu(dev, addr, s->length);

		/* Handle the previous not yet processed entries */
		if (i > start) {
			/*
			 * Can only merge when the last chunk ends on a
			 * page boundary and the new one doesn't have an
			 * offset.
			 */
			if (!iommu_merge || !nextneed || !need || s->offset ||
			    (s->length + seg_size > max_seg_size) ||
			    (ps->offset + ps->length) % PAGE_SIZE) {
				if (dma_map_cont(dev, start_sg, i - start,
						 sgmap, pages, need) < 0)
					goto error;
				out++;
				seg_size = 0;
				sgmap = sg_next(sgmap);
				pages = 0;
				start = i;
				start_sg = s;
			}
		}

		seg_size += s->length;
		need = nextneed;
		pages += iommu_num_pages(s->offset, s->length, PAGE_SIZE);
		ps = s;
	}
	if (dma_map_cont(dev, start_sg, i - start, sgmap, pages, need) < 0)
		goto error;
	out++;
	flush_gart();
	if (out < nents) {
		sgmap = sg_next(sgmap);
		sgmap->dma_length = 0;
	}
	return out;

error:
	flush_gart();
	gart_unmap_sg(dev, sg, out, dir);

	/* When it was forced or merged try again in a dumb way */
	if (force_iommu || iommu_merge) {
		out = dma_map_sg_nonforce(dev, sg, nents, dir);
		if (out > 0)
			return out;
	}
	if (panic_on_overflow)
		panic("dma_map_sg: overflow on %lu pages\n", pages);

	iommu_full(dev, pages << PAGE_SHIFT, dir);
	for_each_sg(sg, s, nents, i)
		s->dma_address = bad_dma_address;
	return 0;
}

/* allocate and map a coherent mapping */
static void *
gart_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_addr,
		    gfp_t flag)
{
	dma_addr_t paddr;
	unsigned long align_mask;
	struct page *page;

	if (force_iommu && !(flag & GFP_DMA)) {
		flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
		page = alloc_pages(flag | __GFP_ZERO, get_order(size));
		if (!page)
			return NULL;

		align_mask = (1UL << get_order(size)) - 1;
		paddr = dma_map_area(dev, page_to_phys(page), size,
				     DMA_BIDIRECTIONAL, align_mask);

		flush_gart();
		if (paddr != bad_dma_address) {
			*dma_addr = paddr;
			return page_address(page);
		}
		__free_pages(page, get_order(size));
	} else
		return dma_generic_alloc_coherent(dev, size, dma_addr, flag);

	return NULL;
}

/* free a coherent mapping */
static void
gart_free_coherent(struct device *dev, size_t size, void *vaddr,
		   dma_addr_t dma_addr)
{
	gart_unmap_page(dev, dma_addr, size, DMA_BIDIRECTIONAL, NULL);
	free_pages((unsigned long)vaddr, get_order(size));
}

static int no_agp;

static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
{
	unsigned long a;

	if (!iommu_size) {
		iommu_size = aper_size;
		if (!no_agp)
			iommu_size /= 2;
	}

	a = aper + iommu_size;
	iommu_size -= round_up(a, PMD_PAGE_SIZE) - a;

	if (iommu_size < 64*1024*1024) {
		printk(KERN_WARNING
			"PCI-DMA: Warning: Small IOMMU %luMB."
			" Consider increasing the AGP aperture in BIOS\n",
				iommu_size >> 20);
	}

	return iommu_size;
}

static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
{
	unsigned aper_size = 0, aper_base_32, aper_order;
	u64 aper_base;

	pci_read_config_dword(dev, AMD64_GARTAPERTUREBASE, &aper_base_32);
	pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &aper_order);
	aper_order = (aper_order >> 1) & 7;

	aper_base = aper_base_32 & 0x7fff;
	aper_base <<= 25;

	aper_size = (32 * 1024 * 1024) << aper_order;
	if (aper_base + aper_size > 0x100000000UL || !aper_size)
		aper_base = 0;

	*size = aper_size;
	return aper_base;
}

static void enable_gart_translations(void)
{
	int i;

	for (i = 0; i < num_k8_northbridges; i++) {
		struct pci_dev *dev = k8_northbridges[i];

		enable_gart_translation(dev, __pa(agp_gatt_table));
	}
}

/*
 * If fix_up_north_bridges is set, the north bridges have to be fixed up on
 * resume in the same way as they are handled in gart_iommu_hole_init().
 */
static bool fix_up_north_bridges;
static u32 aperture_order;
static u32 aperture_alloc;

void set_up_gart_resume(u32 aper_order, u32 aper_alloc)
{
	fix_up_north_bridges = true;
	aperture_order = aper_order;
	aperture_alloc = aper_alloc;
}

static int gart_resume(struct sys_device *dev)
{
	printk(KERN_INFO "PCI-DMA: Resuming GART IOMMU\n");

	if (fix_up_north_bridges) {
		int i;

		printk(KERN_INFO "PCI-DMA: Restoring GART aperture settings\n");

		for (i = 0; i < num_k8_northbridges; i++) {
			struct pci_dev *dev = k8_northbridges[i];

			/*
			 * Don't enable translations just yet.  That is the next
			 * step.  Restore the pre-suspend aperture settings.
			 */
			pci_write_config_dword(dev, AMD64_GARTAPERTURECTL,
						aperture_order << 1);
			pci_write_config_dword(dev, AMD64_GARTAPERTUREBASE,
						aperture_alloc >> 25);
		}
	}

	enable_gart_translations();

	return 0;
}

static int gart_suspend(struct sys_device *dev, pm_message_t state)
{
	return 0;
}

static struct sysdev_class gart_sysdev_class = {
	.name = "gart",
	.suspend = gart_suspend,
	.resume = gart_resume,

};

static struct sys_device device_gart = {
	.id	= 0,
	.cls	= &gart_sysdev_class,
};

/*
 * Private Northbridge GATT initialization in case we cannot use the
 * AGP driver for some reason.
 */
static __init int init_k8_gatt(struct agp_kern_info *info)
{
	unsigned aper_size, gatt_size, new_aper_size;
	unsigned aper_base, new_aper_base;
	struct pci_dev *dev;
	void *gatt;
	int i, error;

	printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
	aper_size = aper_base = info->aper_size = 0;
	dev = NULL;
	for (i = 0; i < num_k8_northbridges; i++) {
		dev = k8_northbridges[i];
		new_aper_base = read_aperture(dev, &new_aper_size);
		if (!new_aper_base)
			goto nommu;

		if (!aper_base) {
			aper_size = new_aper_size;
			aper_base = new_aper_base;
		}
		if (aper_size != new_aper_size || aper_base != new_aper_base)
			goto nommu;
	}
	if (!aper_base)
		goto nommu;
	info->aper_base = aper_base;
	info->aper_size = aper_size >> 20;

	gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
	gatt = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
					get_order(gatt_size));
	if (!gatt)
		panic("Cannot allocate GATT table");
	if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
		panic("Could not set GART PTEs to uncacheable pages");

	agp_gatt_table = gatt;

	enable_gart_translations();

	error = sysdev_class_register(&gart_sysdev_class);
	if (!error)
		error = sysdev_register(&device_gart);
	if (error)
		panic("Could not register gart_sysdev -- "
		      "would corrupt data on next suspend");

	flush_gart();

	printk(KERN_INFO "PCI-DMA: aperture base @ %x size %u KB\n",
	       aper_base, aper_size>>10);

	return 0;

 nommu:
	/* Should not happen anymore */
	printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
	       KERN_WARNING "falling back to iommu=soft.\n");
	return -1;
}

static struct dma_mapping_ops gart_dma_ops = {
	.map_sg				= gart_map_sg,
	.unmap_sg			= gart_unmap_sg,
	.map_page			= gart_map_page,
	.unmap_page			= gart_unmap_page,
	.alloc_coherent			= gart_alloc_coherent,
	.free_coherent			= gart_free_coherent,
};

void gart_iommu_shutdown(void)
{
	struct pci_dev *dev;
	int i;

	if (no_agp && (dma_ops != &gart_dma_ops))
		return;

	for (i = 0; i < num_k8_northbridges; i++) {
		u32 ctl;

		dev = k8_northbridges[i];
		pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &ctl);

		ctl &= ~GARTEN;

		pci_write_config_dword(dev, AMD64_GARTAPERTURECTL, ctl);
	}
}

void __init gart_iommu_init(void)
{
	struct agp_kern_info info;
	unsigned long iommu_start;
	unsigned long aper_base, aper_size;
	unsigned long start_pfn, end_pfn;
	unsigned long scratch;
	long i;

	if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0)
		return;

#ifndef CONFIG_AGP_AMD64
	no_agp = 1;
#else
	/* Makefile puts PCI initialization via subsys_initcall first. */
	/* Add other K8 AGP bridge drivers here */
	no_agp = no_agp ||
		(agp_amd64_init() < 0) ||
		(agp_copy_info(agp_bridge, &info) < 0);
#endif

	if (swiotlb)
		return;

	/* Did we detect a different HW IOMMU? */
	if (iommu_detected && !gart_iommu_aperture)
		return;

	if (no_iommu ||
	    (!force_iommu && max_pfn <= MAX_DMA32_PFN) ||
	    !gart_iommu_aperture ||
	    (no_agp && init_k8_gatt(&info) < 0)) {
		if (max_pfn > MAX_DMA32_PFN) {
			printk(KERN_WARNING "More than 4GB of memory "
			       "but GART IOMMU not available.\n");
			printk(KERN_WARNING "falling back to iommu=soft.\n");
		}
		return;
	}

	/* need to map that range */
	aper_size = info.aper_size << 20;
	aper_base = info.aper_base;
	end_pfn = (aper_base>>PAGE_SHIFT) + (aper_size>>PAGE_SHIFT);
	if (end_pfn > max_low_pfn_mapped) {
		start_pfn = (aper_base>>PAGE_SHIFT);
		init_memory_mapping(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
	}

	printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
	iommu_size = check_iommu_size(info.aper_base, aper_size);
	iommu_pages = iommu_size >> PAGE_SHIFT;

	iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO,
						      get_order(iommu_pages/8));
	if (!iommu_gart_bitmap)
		panic("Cannot allocate iommu bitmap\n");

#ifdef CONFIG_IOMMU_LEAK
	if (leak_trace) {
		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
				  get_order(iommu_pages*sizeof(void *)));
		if (!iommu_leak_tab)
			printk(KERN_DEBUG
			       "PCI-DMA: Cannot allocate leak trace area\n");
	}
#endif

	/*
	 * Out of IOMMU space handling.
	 * Reserve some invalid pages at the beginning of the GART.
	 */
	iommu_area_reserve(iommu_gart_bitmap, 0, EMERGENCY_PAGES);

	agp_memory_reserved = iommu_size;
	printk(KERN_INFO
	       "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
	       iommu_size >> 20);

	iommu_start = aper_size - iommu_size;
	iommu_bus_base = info.aper_base + iommu_start;
	bad_dma_address = iommu_bus_base;
	iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);

	/*
	 * Unmap the IOMMU part of the GART. The alias of the page is
	 * always mapped with cache enabled and there is no full cache
	 * coherency across the GART remapping. The unmapping avoids
	 * automatic prefetches from the CPU allocating cache lines in
	 * there. All CPU accesses are done via the direct mapping to
	 * the backing memory. The GART address is only used by PCI
	 * devices.
	 */
	set_memory_np((unsigned long)__va(iommu_bus_base),
				iommu_size >> PAGE_SHIFT);
	/*
	 * Tricky. The GART table remaps the physical memory range,
	 * so the CPU wont notice potential aliases and if the memory
	 * is remapped to UC later on, we might surprise the PCI devices
	 * with a stray writeout of a cacheline. So play it sure and
	 * do an explicit, full-scale wbinvd() _after_ having marked all
	 * the pages as Not-Present:
	 */
	wbinvd();

	/*
	 * Try to workaround a bug (thanks to BenH):
	 * Set unmapped entries to a scratch page instead of 0.
	 * Any prefetches that hit unmapped entries won't get an bus abort
	 * then. (P2P bridge may be prefetching on DMA reads).
	 */
	scratch = get_zeroed_page(GFP_KERNEL);
	if (!scratch)
		panic("Cannot allocate iommu scratch page");
	gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
	for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
		iommu_gatt_base[i] = gart_unmapped_entry;

	flush_gart();
	dma_ops = &gart_dma_ops;
}

void __init gart_parse_options(char *p)
{
	int arg;

#ifdef CONFIG_IOMMU_LEAK
	if (!strncmp(p, "leak", 4)) {
		leak_trace = 1;
		p += 4;
		if (*p == '=')
			++p;
		if (isdigit(*p) && get_option(&p, &arg))
			iommu_leak_pages = arg;
	}
#endif
	if (isdigit(*p) && get_option(&p, &arg))
		iommu_size = arg;
	if (!strncmp(p, "fullflush", 8))
		iommu_fullflush = 1;
	if (!strncmp(p, "nofullflush", 11))
		iommu_fullflush = 0;
	if (!strncmp(p, "noagp", 5))
		no_agp = 1;
	if (!strncmp(p, "noaperture", 10))
		fix_aperture = 0;
	/* duplicated from pci-dma.c */
	if (!strncmp(p, "force", 5))
		gart_iommu_aperture_allowed = 1;
	if (!strncmp(p, "allowed", 7))
		gart_iommu_aperture_allowed = 1;
	if (!strncmp(p, "memaper", 7)) {
		fallback_aper_force = 1;
		p += 7;
		if (*p == '=') {
			++p;
			if (get_option(&p, &arg))
				fallback_aper_order = arg;
		}
	}
}