blob: 1f99b62ff616ee34f48723897af2cf187ad18b12 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Dynamic DMA mapping support for AMD Hammer.
Ingo Molnar05fccb02008-01-30 13:30:12 +01003 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 * Use the integrated AGP GART in the Hammer northbridge as an IOMMU for PCI.
5 * This allows to use PCI devices that only support 32bit addresses on systems
Ingo Molnar05fccb02008-01-30 13:30:12 +01006 * with more than 4GB.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007 *
8 * See Documentation/DMA-mapping.txt for the interface specification.
Ingo Molnar05fccb02008-01-30 13:30:12 +01009 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 * Copyright 2002 Andi Kleen, SuSE Labs.
Andi Kleenff7f3642007-10-17 18:04:37 +020011 * Subject to the GNU General Public License v2 only.
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 */
13
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/types.h>
15#include <linux/ctype.h>
16#include <linux/agp_backend.h>
17#include <linux/init.h>
18#include <linux/mm.h>
19#include <linux/string.h>
20#include <linux/spinlock.h>
21#include <linux/pci.h>
22#include <linux/module.h>
23#include <linux/topology.h>
24#include <linux/interrupt.h>
25#include <linux/bitops.h>
Christoph Hellwig1eeb66a2007-05-08 00:27:03 -070026#include <linux/kdebug.h>
Jens Axboe9ee1bea2007-10-04 09:35:37 +020027#include <linux/scatterlist.h>
FUJITA Tomonorifde9a102008-02-04 22:28:11 -080028#include <linux/iommu-helper.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <asm/atomic.h>
30#include <asm/io.h>
31#include <asm/mtrr.h>
32#include <asm/pgtable.h>
33#include <asm/proto.h>
Joerg Roedel395624f2007-10-24 12:49:47 +020034#include <asm/gart.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include <asm/cacheflush.h>
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +010036#include <asm/swiotlb.h>
37#include <asm/dma.h>
Andi Kleena32073b2006-06-26 13:56:40 +020038#include <asm/k8.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039
Joerg Roedel79da0872007-10-24 12:49:49 +020040static unsigned long iommu_bus_base; /* GART remapping area (physical) */
Ingo Molnar05fccb02008-01-30 13:30:12 +010041static unsigned long iommu_size; /* size of remapping area bytes */
Linus Torvalds1da177e2005-04-16 15:20:36 -070042static unsigned long iommu_pages; /* .. and in pages */
43
Ingo Molnar05fccb02008-01-30 13:30:12 +010044static u32 *iommu_gatt_base; /* Remapping table */
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Ingo Molnar05fccb02008-01-30 13:30:12 +010046/*
47 * If this is disabled the IOMMU will use an optimized flushing strategy
48 * of only flushing when an mapping is reused. With it true the GART is
49 * flushed for every mapping. Problem is that doing the lazy flush seems
50 * to trigger bugs with some popular PCI cards, in particular 3ware (but
51 * has been also also seen with Qlogic at least).
52 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070053int iommu_fullflush = 1;
54
Ingo Molnar05fccb02008-01-30 13:30:12 +010055/* Allocation bitmap for the remapping area: */
Linus Torvalds1da177e2005-04-16 15:20:36 -070056static DEFINE_SPINLOCK(iommu_bitmap_lock);
Ingo Molnar05fccb02008-01-30 13:30:12 +010057/* Guarded by iommu_bitmap_lock: */
58static unsigned long *iommu_gart_bitmap;
Linus Torvalds1da177e2005-04-16 15:20:36 -070059
Ingo Molnar05fccb02008-01-30 13:30:12 +010060static u32 gart_unmapped_entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62#define GPTE_VALID 1
63#define GPTE_COHERENT 2
64#define GPTE_ENCODE(x) \
65 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
66#define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
67
Ingo Molnar05fccb02008-01-30 13:30:12 +010068#define to_pages(addr, size) \
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
70
Ingo Molnar05fccb02008-01-30 13:30:12 +010071#define EMERGENCY_PAGES 32 /* = 128KB */
Linus Torvalds1da177e2005-04-16 15:20:36 -070072
73#ifdef CONFIG_AGP
74#define AGPEXTERN extern
75#else
76#define AGPEXTERN
77#endif
78
79/* backdoor interface to AGP driver */
80AGPEXTERN int agp_memory_reserved;
81AGPEXTERN __u32 *agp_gatt_table;
82
83static unsigned long next_bit; /* protected by iommu_bitmap_lock */
Ingo Molnar05fccb02008-01-30 13:30:12 +010084static int need_flush; /* global flush state. set for each gart wrap */
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
FUJITA Tomonorifde9a102008-02-04 22:28:11 -080086static unsigned long alloc_iommu(struct device *dev, int size)
Ingo Molnar05fccb02008-01-30 13:30:12 +010087{
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 unsigned long offset, flags;
FUJITA Tomonorifde9a102008-02-04 22:28:11 -080089 unsigned long boundary_size;
90 unsigned long base_index;
91
92 base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
93 PAGE_SIZE) >> PAGE_SHIFT;
94 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
95 PAGE_SIZE) >> PAGE_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
Ingo Molnar05fccb02008-01-30 13:30:12 +010097 spin_lock_irqsave(&iommu_bitmap_lock, flags);
FUJITA Tomonorifde9a102008-02-04 22:28:11 -080098 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
99 size, base_index, boundary_size, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 if (offset == -1) {
101 need_flush = 1;
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800102 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
103 size, base_index, boundary_size, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 }
Ingo Molnar05fccb02008-01-30 13:30:12 +0100105 if (offset != -1) {
106 set_bit_string(iommu_gart_bitmap, offset, size);
107 next_bit = offset+size;
108 if (next_bit >= iommu_pages) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109 next_bit = 0;
110 need_flush = 1;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100111 }
112 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 if (iommu_fullflush)
114 need_flush = 1;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100115 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
116
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 return offset;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100118}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
120static void free_iommu(unsigned long offset, int size)
Ingo Molnar05fccb02008-01-30 13:30:12 +0100121{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 unsigned long flags;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100123
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 spin_lock_irqsave(&iommu_bitmap_lock, flags);
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800125 iommu_area_free(iommu_gart_bitmap, offset, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100127}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128
Ingo Molnar05fccb02008-01-30 13:30:12 +0100129/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 * Use global flush state to avoid races with multiple flushers.
131 */
Andi Kleena32073b2006-06-26 13:56:40 +0200132static void flush_gart(void)
Ingo Molnar05fccb02008-01-30 13:30:12 +0100133{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 unsigned long flags;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100135
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 spin_lock_irqsave(&iommu_bitmap_lock, flags);
Andi Kleena32073b2006-06-26 13:56:40 +0200137 if (need_flush) {
138 k8_flush_garts();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 need_flush = 0;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100140 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100142}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144#ifdef CONFIG_IOMMU_LEAK
145
Ingo Molnar05fccb02008-01-30 13:30:12 +0100146#define SET_LEAK(x) \
147 do { \
148 if (iommu_leak_tab) \
149 iommu_leak_tab[x] = __builtin_return_address(0);\
150 } while (0)
151
152#define CLEAR_LEAK(x) \
153 do { \
154 if (iommu_leak_tab) \
155 iommu_leak_tab[x] = NULL; \
156 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
158/* Debugging aid for drivers that don't free their IOMMU tables */
Ingo Molnar05fccb02008-01-30 13:30:12 +0100159static void **iommu_leak_tab;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160static int leak_trace;
Joerg Roedel79da0872007-10-24 12:49:49 +0200161static int iommu_leak_pages = 20;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100162
Joerg Roedel79da0872007-10-24 12:49:49 +0200163static void dump_leak(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164{
165 int i;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100166 static int dump;
167
168 if (dump || !iommu_leak_tab)
169 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 dump = 1;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100171 show_stack(NULL, NULL);
172
173 /* Very crude. dump some from the end of the table too */
174 printk(KERN_DEBUG "Dumping %d pages from end of IOMMU:\n",
175 iommu_leak_pages);
176 for (i = 0; i < iommu_leak_pages; i += 2) {
177 printk(KERN_DEBUG "%lu: ", iommu_pages-i);
Arjan van de Venbc850d62008-01-30 13:33:07 +0100178 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i], 0);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100179 printk(KERN_CONT "%c", (i+1)%2 == 0 ? '\n' : ' ');
180 }
181 printk(KERN_DEBUG "\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182}
183#else
Ingo Molnar05fccb02008-01-30 13:30:12 +0100184# define SET_LEAK(x)
185# define CLEAR_LEAK(x)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186#endif
187
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100188static void iommu_full(struct device *dev, size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189{
Ingo Molnar05fccb02008-01-30 13:30:12 +0100190 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 * Ran out of IOMMU space for this operation. This is very bad.
192 * Unfortunately the drivers cannot handle this operation properly.
Ingo Molnar05fccb02008-01-30 13:30:12 +0100193 * Return some non mapped prereserved space in the aperture and
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 * let the Northbridge deal with it. This will result in garbage
195 * in the IO operation. When the size exceeds the prereserved space
Ingo Molnar05fccb02008-01-30 13:30:12 +0100196 * memory corruption will occur or random memory will be DMAed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 * out. Hopefully no network devices use single mappings that big.
Ingo Molnar05fccb02008-01-30 13:30:12 +0100198 */
199
200 printk(KERN_ERR
201 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
202 size, dev->bus_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100204 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
206 panic("PCI-DMA: Memory would be corrupted\n");
Ingo Molnar05fccb02008-01-30 13:30:12 +0100207 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
208 panic(KERN_ERR
209 "PCI-DMA: Random memory would be DMAed\n");
210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211#ifdef CONFIG_IOMMU_LEAK
Ingo Molnar05fccb02008-01-30 13:30:12 +0100212 dump_leak();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214}
215
Ingo Molnar05fccb02008-01-30 13:30:12 +0100216static inline int
217need_iommu(struct device *dev, unsigned long addr, size_t size)
218{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 u64 mask = *dev->dma_mask;
Andi Kleen00edefa2007-02-13 13:26:24 +0100220 int high = addr + size > mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 int mmu = high;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100222
223 if (force_iommu)
224 mmu = 1;
225
226 return mmu;
227}
228
229static inline int
230nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
231{
232 u64 mask = *dev->dma_mask;
233 int high = addr + size > mask;
234 int mmu = high;
235
236 return mmu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237}
238
239/* Map a single continuous physical area into the IOMMU.
240 * Caller needs to check if the iommu is needed and flush.
241 */
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100242static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
243 size_t size, int dir)
Ingo Molnar05fccb02008-01-30 13:30:12 +0100244{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 unsigned long npages = to_pages(phys_mem, size);
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800246 unsigned long iommu_page = alloc_iommu(dev, npages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 int i;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100248
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 if (iommu_page == -1) {
250 if (!nonforced_iommu(dev, phys_mem, size))
Ingo Molnar05fccb02008-01-30 13:30:12 +0100251 return phys_mem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 if (panic_on_overflow)
253 panic("dma_map_area overflow %lu bytes\n", size);
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100254 iommu_full(dev, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 return bad_dma_address;
256 }
257
258 for (i = 0; i < npages; i++) {
259 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
260 SET_LEAK(iommu_page + i);
261 phys_mem += PAGE_SIZE;
262 }
263 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
264}
265
Ingo Molnar05fccb02008-01-30 13:30:12 +0100266static dma_addr_t
Ingo Molnar2be62142008-04-19 19:19:56 +0200267gart_map_simple(struct device *dev, phys_addr_t paddr, size_t size, int dir)
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100268{
Ingo Molnar2be62142008-04-19 19:19:56 +0200269 dma_addr_t map = dma_map_area(dev, paddr, size, dir);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100270
Andi Kleena32073b2006-06-26 13:56:40 +0200271 flush_gart();
Ingo Molnar05fccb02008-01-30 13:30:12 +0100272
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100273 return map;
274}
275
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276/* Map a single area into the IOMMU */
Ingo Molnar05fccb02008-01-30 13:30:12 +0100277static dma_addr_t
Ingo Molnar2be62142008-04-19 19:19:56 +0200278gart_map_single(struct device *dev, phys_addr_t paddr, size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279{
Ingo Molnar2be62142008-04-19 19:19:56 +0200280 unsigned long bus;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 if (!dev)
283 dev = &fallback_dev;
284
Ingo Molnar2be62142008-04-19 19:19:56 +0200285 if (!need_iommu(dev, paddr, size))
286 return paddr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287
Ingo Molnar2be62142008-04-19 19:19:56 +0200288 bus = gart_map_simple(dev, paddr, size, dir);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100289
290 return bus;
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100291}
292
293/*
Jon Mason7c2d9cd2006-06-26 13:56:37 +0200294 * Free a DMA mapping.
295 */
Yinghai Lu1048fa52007-07-21 17:11:23 +0200296static void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
Ingo Molnar05fccb02008-01-30 13:30:12 +0100297 size_t size, int direction)
Jon Mason7c2d9cd2006-06-26 13:56:37 +0200298{
299 unsigned long iommu_page;
300 int npages;
301 int i;
302
303 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
304 dma_addr >= iommu_bus_base + iommu_size)
305 return;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100306
Jon Mason7c2d9cd2006-06-26 13:56:37 +0200307 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
308 npages = to_pages(dma_addr, size);
309 for (i = 0; i < npages; i++) {
310 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
311 CLEAR_LEAK(iommu_page + i);
312 }
313 free_iommu(iommu_page, npages);
314}
315
316/*
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100317 * Wrapper for pci_unmap_single working with scatterlists.
318 */
Ingo Molnar05fccb02008-01-30 13:30:12 +0100319static void
320gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100321{
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200322 struct scatterlist *s;
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100323 int i;
324
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200325 for_each_sg(sg, s, nents, i) {
Jon Mason60b08c62006-02-26 04:18:22 +0100326 if (!s->dma_length || !s->length)
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100327 break;
Jon Mason7c2d9cd2006-06-26 13:56:37 +0200328 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100329 }
330}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331
332/* Fallback for dma_map_sg in case of overflow */
333static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
334 int nents, int dir)
335{
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200336 struct scatterlist *s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 int i;
338
339#ifdef CONFIG_IOMMU_DEBUG
340 printk(KERN_DEBUG "dma_map_sg overflow\n");
341#endif
342
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200343 for_each_sg(sg, s, nents, i) {
Jens Axboe58b053e2007-10-22 20:02:46 +0200344 unsigned long addr = sg_phys(s);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100345
346 if (nonforced_iommu(dev, addr, s->length)) {
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100347 addr = dma_map_area(dev, addr, s->length, dir);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100348 if (addr == bad_dma_address) {
349 if (i > 0)
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100350 gart_unmap_sg(dev, sg, i, dir);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100351 nents = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 sg[0].dma_length = 0;
353 break;
354 }
355 }
356 s->dma_address = addr;
357 s->dma_length = s->length;
358 }
Andi Kleena32073b2006-06-26 13:56:40 +0200359 flush_gart();
Ingo Molnar05fccb02008-01-30 13:30:12 +0100360
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 return nents;
362}
363
364/* Map multiple scatterlist entries continuous into the first. */
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800365static int __dma_map_cont(struct device *dev, struct scatterlist *start,
366 int nelems, struct scatterlist *sout,
367 unsigned long pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368{
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800369 unsigned long iommu_start = alloc_iommu(dev, pages);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100370 unsigned long iommu_page = iommu_start;
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200371 struct scatterlist *s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 int i;
373
374 if (iommu_start == -1)
375 return -1;
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200376
377 for_each_sg(start, s, nelems, i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 unsigned long pages, addr;
379 unsigned long phys_addr = s->dma_address;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100380
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200381 BUG_ON(s != start && s->offset);
382 if (s == start) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 sout->dma_address = iommu_bus_base;
384 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
385 sout->dma_length = s->length;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100386 } else {
387 sout->dma_length += s->length;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 }
389
390 addr = phys_addr;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100391 pages = to_pages(s->offset, s->length);
392 while (pages--) {
393 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 SET_LEAK(iommu_page);
395 addr += PAGE_SIZE;
396 iommu_page++;
Andi Kleen0d5410642006-02-12 14:34:59 -0800397 }
Ingo Molnar05fccb02008-01-30 13:30:12 +0100398 }
399 BUG_ON(iommu_page - iommu_start != pages);
400
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 return 0;
402}
403
Ingo Molnar05fccb02008-01-30 13:30:12 +0100404static inline int
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800405dma_map_cont(struct device *dev, struct scatterlist *start, int nelems,
406 struct scatterlist *sout, unsigned long pages, int need)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407{
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200408 if (!need) {
409 BUG_ON(nelems != 1);
FUJITA Tomonorie88a39d2007-10-25 09:13:32 +0200410 sout->dma_address = start->dma_address;
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200411 sout->dma_length = start->length;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 return 0;
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200413 }
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800414 return __dma_map_cont(dev, start, nelems, sout, pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415}
Ingo Molnar05fccb02008-01-30 13:30:12 +0100416
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417/*
418 * DMA map all entries in a scatterlist.
Ingo Molnar05fccb02008-01-30 13:30:12 +0100419 * Merge chunks that have page aligned sizes into a continuous mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 */
Ingo Molnar05fccb02008-01-30 13:30:12 +0100421static int
422gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423{
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200424 struct scatterlist *s, *ps, *start_sg, *sgmap;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100425 int need = 0, nextneed, i, out, start;
426 unsigned long pages = 0;
FUJITA Tomonori42d00282008-02-04 22:27:56 -0800427 unsigned int seg_size;
428 unsigned int max_seg_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
Ingo Molnar05fccb02008-01-30 13:30:12 +0100430 if (nents == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 return 0;
432
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 if (!dev)
434 dev = &fallback_dev;
435
436 out = 0;
437 start = 0;
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200438 start_sg = sgmap = sg;
FUJITA Tomonori42d00282008-02-04 22:27:56 -0800439 seg_size = 0;
440 max_seg_size = dma_get_max_seg_size(dev);
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200441 ps = NULL; /* shut up gcc */
442 for_each_sg(sg, s, nents, i) {
Jens Axboe58b053e2007-10-22 20:02:46 +0200443 dma_addr_t addr = sg_phys(s);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
Ingo Molnar05fccb02008-01-30 13:30:12 +0100445 s->dma_address = addr;
446 BUG_ON(s->length == 0);
447
448 nextneed = need_iommu(dev, addr, s->length);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
450 /* Handle the previous not yet processed entries */
451 if (i > start) {
Ingo Molnar05fccb02008-01-30 13:30:12 +0100452 /*
453 * Can only merge when the last chunk ends on a
454 * page boundary and the new one doesn't have an
455 * offset.
456 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 if (!iommu_merge || !nextneed || !need || s->offset ||
FUJITA Tomonori42d00282008-02-04 22:27:56 -0800458 (s->length + seg_size > max_seg_size) ||
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200459 (ps->offset + ps->length) % PAGE_SIZE) {
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800460 if (dma_map_cont(dev, start_sg, i - start,
461 sgmap, pages, need) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 goto error;
463 out++;
FUJITA Tomonori42d00282008-02-04 22:27:56 -0800464 seg_size = 0;
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200465 sgmap = sg_next(sgmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 pages = 0;
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200467 start = i;
468 start_sg = s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 }
470 }
471
FUJITA Tomonori42d00282008-02-04 22:27:56 -0800472 seg_size += s->length;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 need = nextneed;
474 pages += to_pages(s->offset, s->length);
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200475 ps = s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 }
FUJITA Tomonorifde9a102008-02-04 22:28:11 -0800477 if (dma_map_cont(dev, start_sg, i - start, sgmap, pages, need) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 goto error;
479 out++;
Andi Kleena32073b2006-06-26 13:56:40 +0200480 flush_gart();
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200481 if (out < nents) {
482 sgmap = sg_next(sgmap);
483 sgmap->dma_length = 0;
484 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 return out;
486
487error:
Andi Kleena32073b2006-06-26 13:56:40 +0200488 flush_gart();
FUJITA Tomonori53369402007-10-26 13:56:24 +0200489 gart_unmap_sg(dev, sg, out, dir);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100490
Kevin VanMarena1002a42006-02-03 21:51:32 +0100491 /* When it was forced or merged try again in a dumb way */
492 if (force_iommu || iommu_merge) {
493 out = dma_map_sg_nonforce(dev, sg, nents, dir);
494 if (out > 0)
495 return out;
496 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 if (panic_on_overflow)
498 panic("dma_map_sg: overflow on %lu pages\n", pages);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100499
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100500 iommu_full(dev, pages << PAGE_SHIFT, dir);
Jens Axboe9ee1bea2007-10-04 09:35:37 +0200501 for_each_sg(sg, s, nents, i)
502 s->dma_address = bad_dma_address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 return 0;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100504}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100506static int no_agp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507
508static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
Ingo Molnar05fccb02008-01-30 13:30:12 +0100509{
510 unsigned long a;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511
Ingo Molnar05fccb02008-01-30 13:30:12 +0100512 if (!iommu_size) {
513 iommu_size = aper_size;
514 if (!no_agp)
515 iommu_size /= 2;
516 }
517
518 a = aper + iommu_size;
Andi Kleen31422c52008-02-04 16:48:08 +0100519 iommu_size -= round_up(a, PMD_PAGE_SIZE) - a;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520
Ingo Molnar05fccb02008-01-30 13:30:12 +0100521 if (iommu_size < 64*1024*1024) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 printk(KERN_WARNING
Ingo Molnar05fccb02008-01-30 13:30:12 +0100523 "PCI-DMA: Warning: Small IOMMU %luMB."
524 " Consider increasing the AGP aperture in BIOS\n",
525 iommu_size >> 20);
526 }
527
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 return iommu_size;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100529}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530
Ingo Molnar05fccb02008-01-30 13:30:12 +0100531static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
532{
533 unsigned aper_size = 0, aper_base_32, aper_order;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 u64 aper_base;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535
Pavel Machek3bb6fbf2008-04-15 12:43:57 +0200536 pci_read_config_dword(dev, AMD64_GARTAPERTUREBASE, &aper_base_32);
537 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &aper_order);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100538 aper_order = (aper_order >> 1) & 7;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539
Ingo Molnar05fccb02008-01-30 13:30:12 +0100540 aper_base = aper_base_32 & 0x7fff;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 aper_base <<= 25;
542
Ingo Molnar05fccb02008-01-30 13:30:12 +0100543 aper_size = (32 * 1024 * 1024) << aper_order;
544 if (aper_base + aper_size > 0x100000000UL || !aper_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 aper_base = 0;
546
547 *size = aper_size;
548 return aper_base;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100549}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550
Ingo Molnar05fccb02008-01-30 13:30:12 +0100551/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 * Private Northbridge GATT initialization in case we cannot use the
Ingo Molnar05fccb02008-01-30 13:30:12 +0100553 * AGP driver for some reason.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 */
555static __init int init_k8_gatt(struct agp_kern_info *info)
Ingo Molnar05fccb02008-01-30 13:30:12 +0100556{
557 unsigned aper_size, gatt_size, new_aper_size;
558 unsigned aper_base, new_aper_base;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 struct pci_dev *dev;
560 void *gatt;
Andi Kleena32073b2006-06-26 13:56:40 +0200561 int i;
562
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
564 aper_size = aper_base = info->aper_size = 0;
Andi Kleena32073b2006-06-26 13:56:40 +0200565 dev = NULL;
566 for (i = 0; i < num_k8_northbridges; i++) {
567 dev = k8_northbridges[i];
Ingo Molnar05fccb02008-01-30 13:30:12 +0100568 new_aper_base = read_aperture(dev, &new_aper_size);
569 if (!new_aper_base)
570 goto nommu;
571
572 if (!aper_base) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 aper_size = new_aper_size;
574 aper_base = new_aper_base;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100575 }
576 if (aper_size != new_aper_size || aper_base != new_aper_base)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 goto nommu;
578 }
579 if (!aper_base)
Ingo Molnar05fccb02008-01-30 13:30:12 +0100580 goto nommu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 info->aper_base = aper_base;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100582 info->aper_size = aper_size >> 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583
Ingo Molnar05fccb02008-01-30 13:30:12 +0100584 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
585 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
586 if (!gatt)
Joachim Deguaracf6387d2007-04-24 13:05:36 +0200587 panic("Cannot allocate GATT table");
Arjan van de Ven6d238cc2008-01-30 13:34:06 +0100588 if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
Joachim Deguaracf6387d2007-04-24 13:05:36 +0200589 panic("Could not set GART PTEs to uncacheable pages");
Joachim Deguaracf6387d2007-04-24 13:05:36 +0200590
Ingo Molnar05fccb02008-01-30 13:30:12 +0100591 memset(gatt, 0, gatt_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 agp_gatt_table = gatt;
Andi Kleena32073b2006-06-26 13:56:40 +0200593
594 for (i = 0; i < num_k8_northbridges; i++) {
Andi Kleena32073b2006-06-26 13:56:40 +0200595 dev = k8_northbridges[i];
Pavel Machek3bb6fbf2008-04-15 12:43:57 +0200596 enable_gart_translation(dev, __pa(gatt));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 }
Andi Kleena32073b2006-06-26 13:56:40 +0200598 flush_gart();
Ingo Molnar05fccb02008-01-30 13:30:12 +0100599
600 printk(KERN_INFO "PCI-DMA: aperture base @ %x size %u KB\n",
601 aper_base, aper_size>>10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 return 0;
603
604 nommu:
Ingo Molnar05fccb02008-01-30 13:30:12 +0100605 /* Should not happen anymore */
Pavel Machek8f596102008-04-01 14:24:03 +0200606 printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
607 KERN_WARNING "falling back to iommu=soft.\n");
Ingo Molnar05fccb02008-01-30 13:30:12 +0100608 return -1;
609}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
611extern int agp_amd64_init(void);
612
Stephen Hemmingere6584502007-05-02 19:27:06 +0200613static const struct dma_mapping_ops gart_dma_ops = {
Ingo Molnar05fccb02008-01-30 13:30:12 +0100614 .mapping_error = NULL,
615 .map_single = gart_map_single,
616 .map_simple = gart_map_simple,
617 .unmap_single = gart_unmap_single,
618 .sync_single_for_cpu = NULL,
619 .sync_single_for_device = NULL,
620 .sync_single_range_for_cpu = NULL,
621 .sync_single_range_for_device = NULL,
622 .sync_sg_for_cpu = NULL,
623 .sync_sg_for_device = NULL,
624 .map_sg = gart_map_sg,
625 .unmap_sg = gart_unmap_sg,
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100626};
627
Yinghai Lubc2cea62007-07-21 17:11:28 +0200628void gart_iommu_shutdown(void)
629{
630 struct pci_dev *dev;
631 int i;
632
633 if (no_agp && (dma_ops != &gart_dma_ops))
634 return;
635
Ingo Molnar05fccb02008-01-30 13:30:12 +0100636 for (i = 0; i < num_k8_northbridges; i++) {
637 u32 ctl;
Yinghai Lubc2cea62007-07-21 17:11:28 +0200638
Ingo Molnar05fccb02008-01-30 13:30:12 +0100639 dev = k8_northbridges[i];
Pavel Machek3bb6fbf2008-04-15 12:43:57 +0200640 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &ctl);
Yinghai Lubc2cea62007-07-21 17:11:28 +0200641
Pavel Machek3bb6fbf2008-04-15 12:43:57 +0200642 ctl &= ~GARTEN;
Yinghai Lubc2cea62007-07-21 17:11:28 +0200643
Pavel Machek3bb6fbf2008-04-15 12:43:57 +0200644 pci_write_config_dword(dev, AMD64_GARTAPERTURECTL, ctl);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100645 }
Yinghai Lubc2cea62007-07-21 17:11:28 +0200646}
647
Jon Mason0dc243a2006-06-26 13:58:11 +0200648void __init gart_iommu_init(void)
Ingo Molnar05fccb02008-01-30 13:30:12 +0100649{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 struct agp_kern_info info;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 unsigned long iommu_start;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100652 unsigned long aper_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 unsigned long scratch;
654 long i;
655
Andi Kleena32073b2006-06-26 13:56:40 +0200656 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
657 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
Jon Mason0dc243a2006-06-26 13:58:11 +0200658 return;
Andi Kleena32073b2006-06-26 13:56:40 +0200659 }
660
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661#ifndef CONFIG_AGP_AMD64
Ingo Molnar05fccb02008-01-30 13:30:12 +0100662 no_agp = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663#else
664 /* Makefile puts PCI initialization via subsys_initcall first. */
665 /* Add other K8 AGP bridge drivers here */
Ingo Molnar05fccb02008-01-30 13:30:12 +0100666 no_agp = no_agp ||
667 (agp_amd64_init() < 0) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 (agp_copy_info(agp_bridge, &info) < 0);
Ingo Molnar05fccb02008-01-30 13:30:12 +0100669#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670
Jon Mason60b08c62006-02-26 04:18:22 +0100671 if (swiotlb)
Jon Mason0dc243a2006-06-26 13:58:11 +0200672 return;
Jon Mason60b08c62006-02-26 04:18:22 +0100673
Jon Mason8d4f6b92006-06-26 13:58:05 +0200674 /* Did we detect a different HW IOMMU? */
Joerg Roedel0440d4c2007-10-24 12:49:50 +0200675 if (iommu_detected && !gart_iommu_aperture)
Jon Mason0dc243a2006-06-26 13:58:11 +0200676 return;
Jon Mason8d4f6b92006-06-26 13:58:05 +0200677
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 if (no_iommu ||
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100679 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
Joerg Roedel0440d4c2007-10-24 12:49:50 +0200680 !gart_iommu_aperture ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 (no_agp && init_k8_gatt(&info) < 0)) {
Jon Mason5b7b6442006-02-03 21:51:59 +0100682 if (end_pfn > MAX_DMA32_PFN) {
Pavel Machek8f596102008-04-01 14:24:03 +0200683 printk(KERN_WARNING "More than 4GB of memory "
684 "but GART IOMMU not available.\n"
685 KERN_WARNING "falling back to iommu=soft.\n");
Jon Mason5b7b6442006-02-03 21:51:59 +0100686 }
Jon Mason0dc243a2006-06-26 13:58:11 +0200687 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 }
689
Jon Mason5b7b6442006-02-03 21:51:59 +0100690 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
Ingo Molnar05fccb02008-01-30 13:30:12 +0100691 aper_size = info.aper_size * 1024 * 1024;
692 iommu_size = check_iommu_size(info.aper_base, aper_size);
693 iommu_pages = iommu_size >> PAGE_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694
Ingo Molnar05fccb02008-01-30 13:30:12 +0100695 iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL,
696 get_order(iommu_pages/8));
697 if (!iommu_gart_bitmap)
698 panic("Cannot allocate iommu bitmap\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 memset(iommu_gart_bitmap, 0, iommu_pages/8);
700
701#ifdef CONFIG_IOMMU_LEAK
Ingo Molnar05fccb02008-01-30 13:30:12 +0100702 if (leak_trace) {
703 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 get_order(iommu_pages*sizeof(void *)));
Ingo Molnar05fccb02008-01-30 13:30:12 +0100705 if (iommu_leak_tab)
706 memset(iommu_leak_tab, 0, iommu_pages * 8);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 else
Ingo Molnar05fccb02008-01-30 13:30:12 +0100708 printk(KERN_DEBUG
709 "PCI-DMA: Cannot allocate leak trace area\n");
710 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711#endif
712
Ingo Molnar05fccb02008-01-30 13:30:12 +0100713 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 * Out of IOMMU space handling.
Ingo Molnar05fccb02008-01-30 13:30:12 +0100715 * Reserve some invalid pages at the beginning of the GART.
716 */
717 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718
Ingo Molnar05fccb02008-01-30 13:30:12 +0100719 agp_memory_reserved = iommu_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 printk(KERN_INFO
721 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
Ingo Molnar05fccb02008-01-30 13:30:12 +0100722 iommu_size >> 20);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723
Ingo Molnar05fccb02008-01-30 13:30:12 +0100724 iommu_start = aper_size - iommu_size;
725 iommu_bus_base = info.aper_base + iommu_start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 bad_dma_address = iommu_bus_base;
727 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
728
Ingo Molnar05fccb02008-01-30 13:30:12 +0100729 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 * Unmap the IOMMU part of the GART. The alias of the page is
731 * always mapped with cache enabled and there is no full cache
732 * coherency across the GART remapping. The unmapping avoids
733 * automatic prefetches from the CPU allocating cache lines in
734 * there. All CPU accesses are done via the direct mapping to
735 * the backing memory. The GART address is only used by PCI
Ingo Molnar05fccb02008-01-30 13:30:12 +0100736 * devices.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 */
Andi Kleen28d6ee42008-02-04 16:48:08 +0100738 set_memory_np((unsigned long)__va(iommu_bus_base),
739 iommu_size >> PAGE_SHIFT);
Ingo Molnar184652e2008-02-14 23:30:20 +0100740 /*
741 * Tricky. The GART table remaps the physical memory range,
742 * so the CPU wont notice potential aliases and if the memory
743 * is remapped to UC later on, we might surprise the PCI devices
744 * with a stray writeout of a cacheline. So play it sure and
745 * do an explicit, full-scale wbinvd() _after_ having marked all
746 * the pages as Not-Present:
747 */
748 wbinvd();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749
Ingo Molnar05fccb02008-01-30 13:30:12 +0100750 /*
751 * Try to workaround a bug (thanks to BenH)
752 * Set unmapped entries to a scratch page instead of 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 * Any prefetches that hit unmapped entries won't get an bus abort
754 * then.
755 */
Ingo Molnar05fccb02008-01-30 13:30:12 +0100756 scratch = get_zeroed_page(GFP_KERNEL);
757 if (!scratch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 panic("Cannot allocate iommu scratch page");
759 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
Ingo Molnar05fccb02008-01-30 13:30:12 +0100760 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 iommu_gatt_base[i] = gart_unmapped_entry;
762
Andi Kleena32073b2006-06-26 13:56:40 +0200763 flush_gart();
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100764 dma_ops = &gart_dma_ops;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100765}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766
Sam Ravnborg43999d92007-03-16 21:07:36 +0100767void __init gart_parse_options(char *p)
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100768{
769 int arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771#ifdef CONFIG_IOMMU_LEAK
Ingo Molnar05fccb02008-01-30 13:30:12 +0100772 if (!strncmp(p, "leak", 4)) {
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100773 leak_trace = 1;
774 p += 4;
775 if (*p == '=') ++p;
776 if (isdigit(*p) && get_option(&p, &arg))
777 iommu_leak_pages = arg;
778 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779#endif
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100780 if (isdigit(*p) && get_option(&p, &arg))
781 iommu_size = arg;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100782 if (!strncmp(p, "fullflush", 8))
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100783 iommu_fullflush = 1;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100784 if (!strncmp(p, "nofullflush", 11))
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100785 iommu_fullflush = 0;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100786 if (!strncmp(p, "noagp", 5))
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100787 no_agp = 1;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100788 if (!strncmp(p, "noaperture", 10))
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100789 fix_aperture = 0;
790 /* duplicated from pci-dma.c */
Ingo Molnar05fccb02008-01-30 13:30:12 +0100791 if (!strncmp(p, "force", 5))
Joerg Roedel0440d4c2007-10-24 12:49:50 +0200792 gart_iommu_aperture_allowed = 1;
Ingo Molnar05fccb02008-01-30 13:30:12 +0100793 if (!strncmp(p, "allowed", 7))
Joerg Roedel0440d4c2007-10-24 12:49:50 +0200794 gart_iommu_aperture_allowed = 1;
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100795 if (!strncmp(p, "memaper", 7)) {
796 fallback_aper_force = 1;
797 p += 7;
798 if (*p == '=') {
799 ++p;
800 if (get_option(&p, &arg))
801 fallback_aper_order = arg;
802 }
803 }
804}