2 * Dynamic DMA mapping support for AMD Hammer.
4 * 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
8 * See Documentation/DMA-mapping.txt for the interface specification.
10 * Copyright 2002 Andi Kleen, SuSE Labs.
11 * Subject to the GNU General Public License v2 only.
14 #include <linux/types.h>
15 #include <linux/ctype.h>
16 #include <linux/agp_backend.h>
17 #include <linux/init.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>
26 #include <linux/kdebug.h>
27 #include <linux/scatterlist.h>
28 #include <asm/atomic.h>
31 #include <asm/pgtable.h>
32 #include <asm/proto.h>
34 #include <asm/cacheflush.h>
35 #include <asm/swiotlb.h>
39 static unsigned long iommu_bus_base; /* GART remapping area (physical) */
40 static unsigned long iommu_size; /* size of remapping area bytes */
41 static unsigned long iommu_pages; /* .. and in pages */
43 static u32 *iommu_gatt_base; /* Remapping table */
45 /* If this is disabled the IOMMU will use an optimized flushing strategy
46 of only flushing when an mapping is reused. With it true the GART is flushed
47 for every mapping. Problem is that doing the lazy flush seems to trigger
48 bugs with some popular PCI cards, in particular 3ware (but has been also
49 also seen with Qlogic at least). */
50 int iommu_fullflush = 1;
52 /* Allocation bitmap for the remapping area */
53 static DEFINE_SPINLOCK(iommu_bitmap_lock);
54 static unsigned long *iommu_gart_bitmap; /* guarded by iommu_bitmap_lock */
56 static u32 gart_unmapped_entry;
59 #define GPTE_COHERENT 2
60 #define GPTE_ENCODE(x) \
61 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
62 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
64 #define to_pages(addr,size) \
65 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
67 #define EMERGENCY_PAGES 32 /* = 128KB */
70 #define AGPEXTERN extern
75 /* backdoor interface to AGP driver */
76 AGPEXTERN int agp_memory_reserved;
77 AGPEXTERN __u32 *agp_gatt_table;
79 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
80 static int need_flush; /* global flush state. set for each gart wrap */
82 static unsigned long alloc_iommu(int size)
84 unsigned long offset, flags;
86 spin_lock_irqsave(&iommu_bitmap_lock, flags);
87 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
90 offset = find_next_zero_string(iommu_gart_bitmap,0,iommu_pages,size);
93 set_bit_string(iommu_gart_bitmap, offset, size);
94 next_bit = offset+size;
95 if (next_bit >= iommu_pages) {
102 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
106 static void free_iommu(unsigned long offset, int size)
109 spin_lock_irqsave(&iommu_bitmap_lock, flags);
110 __clear_bit_string(iommu_gart_bitmap, offset, size);
111 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
115 * Use global flush state to avoid races with multiple flushers.
117 static void flush_gart(void)
120 spin_lock_irqsave(&iommu_bitmap_lock, flags);
125 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
128 #ifdef CONFIG_IOMMU_LEAK
130 #define SET_LEAK(x) if (iommu_leak_tab) \
131 iommu_leak_tab[x] = __builtin_return_address(0);
132 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
133 iommu_leak_tab[x] = NULL;
135 /* Debugging aid for drivers that don't free their IOMMU tables */
136 static void **iommu_leak_tab;
137 static int leak_trace;
138 static int iommu_leak_pages = 20;
139 static void dump_leak(void)
143 if (dump || !iommu_leak_tab) return;
145 show_stack(NULL,NULL);
146 /* Very crude. dump some from the end of the table too */
147 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
148 for (i = 0; i < iommu_leak_pages; i+=2) {
149 printk("%lu: ", iommu_pages-i);
150 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
151 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
157 #define CLEAR_LEAK(x)
160 static void iommu_full(struct device *dev, size_t size, int dir)
163 * Ran out of IOMMU space for this operation. This is very bad.
164 * Unfortunately the drivers cannot handle this operation properly.
165 * Return some non mapped prereserved space in the aperture and
166 * let the Northbridge deal with it. This will result in garbage
167 * in the IO operation. When the size exceeds the prereserved space
168 * memory corruption will occur or random memory will be DMAed
169 * out. Hopefully no network devices use single mappings that big.
173 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
176 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
177 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
178 panic("PCI-DMA: Memory would be corrupted\n");
179 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
180 panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
183 #ifdef CONFIG_IOMMU_LEAK
188 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
190 u64 mask = *dev->dma_mask;
191 int high = addr + size > mask;
198 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
200 u64 mask = *dev->dma_mask;
201 int high = addr + size > mask;
206 /* Map a single continuous physical area into the IOMMU.
207 * Caller needs to check if the iommu is needed and flush.
209 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
210 size_t size, int dir)
212 unsigned long npages = to_pages(phys_mem, size);
213 unsigned long iommu_page = alloc_iommu(npages);
215 if (iommu_page == -1) {
216 if (!nonforced_iommu(dev, phys_mem, size))
218 if (panic_on_overflow)
219 panic("dma_map_area overflow %lu bytes\n", size);
220 iommu_full(dev, size, dir);
221 return bad_dma_address;
224 for (i = 0; i < npages; i++) {
225 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
226 SET_LEAK(iommu_page + i);
227 phys_mem += PAGE_SIZE;
229 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
232 static dma_addr_t gart_map_simple(struct device *dev, char *buf,
233 size_t size, int dir)
235 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
240 /* Map a single area into the IOMMU */
241 static dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
243 unsigned long phys_mem, bus;
248 phys_mem = virt_to_phys(addr);
249 if (!need_iommu(dev, phys_mem, size))
252 bus = gart_map_simple(dev, addr, size, dir);
257 * Free a DMA mapping.
259 static void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
260 size_t size, int direction)
262 unsigned long iommu_page;
266 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
267 dma_addr >= iommu_bus_base + iommu_size)
269 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
270 npages = to_pages(dma_addr, size);
271 for (i = 0; i < npages; i++) {
272 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
273 CLEAR_LEAK(iommu_page + i);
275 free_iommu(iommu_page, npages);
279 * Wrapper for pci_unmap_single working with scatterlists.
281 static void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
283 struct scatterlist *s;
286 for_each_sg(sg, s, nents, i) {
287 if (!s->dma_length || !s->length)
289 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
293 /* Fallback for dma_map_sg in case of overflow */
294 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
297 struct scatterlist *s;
300 #ifdef CONFIG_IOMMU_DEBUG
301 printk(KERN_DEBUG "dma_map_sg overflow\n");
304 for_each_sg(sg, s, nents, i) {
305 unsigned long addr = sg_phys(s);
306 if (nonforced_iommu(dev, addr, s->length)) {
307 addr = dma_map_area(dev, addr, s->length, dir);
308 if (addr == bad_dma_address) {
310 gart_unmap_sg(dev, sg, i, dir);
312 sg[0].dma_length = 0;
316 s->dma_address = addr;
317 s->dma_length = s->length;
323 /* Map multiple scatterlist entries continuous into the first. */
324 static int __dma_map_cont(struct scatterlist *start, int nelems,
325 struct scatterlist *sout, unsigned long pages)
327 unsigned long iommu_start = alloc_iommu(pages);
328 unsigned long iommu_page = iommu_start;
329 struct scatterlist *s;
332 if (iommu_start == -1)
335 for_each_sg(start, s, nelems, i) {
336 unsigned long pages, addr;
337 unsigned long phys_addr = s->dma_address;
339 BUG_ON(s != start && s->offset);
341 sout->dma_address = iommu_bus_base;
342 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
343 sout->dma_length = s->length;
345 sout->dma_length += s->length;
349 pages = to_pages(s->offset, s->length);
351 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
352 SET_LEAK(iommu_page);
357 BUG_ON(iommu_page - iommu_start != pages);
361 static inline int dma_map_cont(struct scatterlist *start, int nelems,
362 struct scatterlist *sout,
363 unsigned long pages, int need)
367 sout->dma_address = start->dma_address;
368 sout->dma_length = start->length;
371 return __dma_map_cont(start, nelems, sout, pages);
375 * DMA map all entries in a scatterlist.
376 * Merge chunks that have page aligned sizes into a continuous mapping.
378 static int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents,
384 unsigned long pages = 0;
385 int need = 0, nextneed;
386 struct scatterlist *s, *ps, *start_sg, *sgmap;
396 start_sg = sgmap = sg;
397 ps = NULL; /* shut up gcc */
398 for_each_sg(sg, s, nents, i) {
399 dma_addr_t addr = sg_phys(s);
400 s->dma_address = addr;
401 BUG_ON(s->length == 0);
403 nextneed = need_iommu(dev, addr, s->length);
405 /* Handle the previous not yet processed entries */
407 /* Can only merge when the last chunk ends on a page
408 boundary and the new one doesn't have an offset. */
409 if (!iommu_merge || !nextneed || !need || s->offset ||
410 (ps->offset + ps->length) % PAGE_SIZE) {
411 if (dma_map_cont(start_sg, i - start, sgmap,
415 sgmap = sg_next(sgmap);
423 pages += to_pages(s->offset, s->length);
426 if (dma_map_cont(start_sg, i - start, sgmap, pages, need) < 0)
431 sgmap = sg_next(sgmap);
432 sgmap->dma_length = 0;
438 gart_unmap_sg(dev, sg, out, dir);
439 /* When it was forced or merged try again in a dumb way */
440 if (force_iommu || iommu_merge) {
441 out = dma_map_sg_nonforce(dev, sg, nents, dir);
445 if (panic_on_overflow)
446 panic("dma_map_sg: overflow on %lu pages\n", pages);
447 iommu_full(dev, pages << PAGE_SHIFT, dir);
448 for_each_sg(sg, s, nents, i)
449 s->dma_address = bad_dma_address;
455 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
459 iommu_size = aper_size;
464 a = aper + iommu_size;
465 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
467 if (iommu_size < 64*1024*1024)
469 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
474 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
476 unsigned aper_size = 0, aper_base_32;
480 pci_read_config_dword(dev, 0x94, &aper_base_32);
481 pci_read_config_dword(dev, 0x90, &aper_order);
482 aper_order = (aper_order >> 1) & 7;
484 aper_base = aper_base_32 & 0x7fff;
487 aper_size = (32 * 1024 * 1024) << aper_order;
488 if (aper_base + aper_size > 0x100000000UL || !aper_size)
496 * Private Northbridge GATT initialization in case we cannot use the
497 * AGP driver for some reason.
499 static __init int init_k8_gatt(struct agp_kern_info *info)
503 unsigned aper_base, new_aper_base;
504 unsigned aper_size, gatt_size, new_aper_size;
507 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
508 aper_size = aper_base = info->aper_size = 0;
510 for (i = 0; i < num_k8_northbridges; i++) {
511 dev = k8_northbridges[i];
512 new_aper_base = read_aperture(dev, &new_aper_size);
517 aper_size = new_aper_size;
518 aper_base = new_aper_base;
520 if (aper_size != new_aper_size || aper_base != new_aper_base)
525 info->aper_base = aper_base;
526 info->aper_size = aper_size>>20;
528 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
529 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
531 panic("Cannot allocate GATT table");
532 if (change_page_attr_addr((unsigned long)gatt, gatt_size >> PAGE_SHIFT, PAGE_KERNEL_NOCACHE))
533 panic("Could not set GART PTEs to uncacheable pages");
536 memset(gatt, 0, gatt_size);
537 agp_gatt_table = gatt;
539 for (i = 0; i < num_k8_northbridges; i++) {
543 dev = k8_northbridges[i];
544 gatt_reg = __pa(gatt) >> 12;
546 pci_write_config_dword(dev, 0x98, gatt_reg);
547 pci_read_config_dword(dev, 0x90, &ctl);
550 ctl &= ~((1<<4) | (1<<5));
552 pci_write_config_dword(dev, 0x90, ctl);
556 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
560 /* Should not happen anymore */
561 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
562 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
566 extern int agp_amd64_init(void);
568 static const struct dma_mapping_ops gart_dma_ops = {
569 .mapping_error = NULL,
570 .map_single = gart_map_single,
571 .map_simple = gart_map_simple,
572 .unmap_single = gart_unmap_single,
573 .sync_single_for_cpu = NULL,
574 .sync_single_for_device = NULL,
575 .sync_single_range_for_cpu = NULL,
576 .sync_single_range_for_device = NULL,
577 .sync_sg_for_cpu = NULL,
578 .sync_sg_for_device = NULL,
579 .map_sg = gart_map_sg,
580 .unmap_sg = gart_unmap_sg,
583 void gart_iommu_shutdown(void)
588 if (no_agp && (dma_ops != &gart_dma_ops))
591 for (i = 0; i < num_k8_northbridges; i++) {
594 dev = k8_northbridges[i];
595 pci_read_config_dword(dev, 0x90, &ctl);
599 pci_write_config_dword(dev, 0x90, ctl);
603 void __init gart_iommu_init(void)
605 struct agp_kern_info info;
606 unsigned long aper_size;
607 unsigned long iommu_start;
608 unsigned long scratch;
611 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
612 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
616 #ifndef CONFIG_AGP_AMD64
619 /* Makefile puts PCI initialization via subsys_initcall first. */
620 /* Add other K8 AGP bridge drivers here */
622 (agp_amd64_init() < 0) ||
623 (agp_copy_info(agp_bridge, &info) < 0);
629 /* Did we detect a different HW IOMMU? */
630 if (iommu_detected && !gart_iommu_aperture)
634 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
635 !gart_iommu_aperture ||
636 (no_agp && init_k8_gatt(&info) < 0)) {
637 if (end_pfn > MAX_DMA32_PFN) {
638 printk(KERN_ERR "WARNING more than 4GB of memory "
639 "but GART IOMMU not available.\n"
640 KERN_ERR "WARNING 32bit PCI may malfunction.\n");
645 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
646 aper_size = info.aper_size * 1024 * 1024;
647 iommu_size = check_iommu_size(info.aper_base, aper_size);
648 iommu_pages = iommu_size >> PAGE_SHIFT;
650 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
651 get_order(iommu_pages/8));
652 if (!iommu_gart_bitmap)
653 panic("Cannot allocate iommu bitmap\n");
654 memset(iommu_gart_bitmap, 0, iommu_pages/8);
656 #ifdef CONFIG_IOMMU_LEAK
658 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
659 get_order(iommu_pages*sizeof(void *)));
661 memset(iommu_leak_tab, 0, iommu_pages * 8);
663 printk("PCI-DMA: Cannot allocate leak trace area\n");
668 * Out of IOMMU space handling.
669 * Reserve some invalid pages at the beginning of the GART.
671 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
673 agp_memory_reserved = iommu_size;
675 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
678 iommu_start = aper_size - iommu_size;
679 iommu_bus_base = info.aper_base + iommu_start;
680 bad_dma_address = iommu_bus_base;
681 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
684 * Unmap the IOMMU part of the GART. The alias of the page is
685 * always mapped with cache enabled and there is no full cache
686 * coherency across the GART remapping. The unmapping avoids
687 * automatic prefetches from the CPU allocating cache lines in
688 * there. All CPU accesses are done via the direct mapping to
689 * the backing memory. The GART address is only used by PCI
692 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
695 * Try to workaround a bug (thanks to BenH)
696 * Set unmapped entries to a scratch page instead of 0.
697 * Any prefetches that hit unmapped entries won't get an bus abort
700 scratch = get_zeroed_page(GFP_KERNEL);
702 panic("Cannot allocate iommu scratch page");
703 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
704 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
705 iommu_gatt_base[i] = gart_unmapped_entry;
708 dma_ops = &gart_dma_ops;
711 void __init gart_parse_options(char *p)
715 #ifdef CONFIG_IOMMU_LEAK
716 if (!strncmp(p,"leak",4)) {
720 if (isdigit(*p) && get_option(&p, &arg))
721 iommu_leak_pages = arg;
724 if (isdigit(*p) && get_option(&p, &arg))
726 if (!strncmp(p, "fullflush",8))
728 if (!strncmp(p, "nofullflush",11))
730 if (!strncmp(p,"noagp",5))
732 if (!strncmp(p, "noaperture",10))
734 /* duplicated from pci-dma.c */
735 if (!strncmp(p,"force",5))
736 gart_iommu_aperture_allowed = 1;
737 if (!strncmp(p,"allowed",7))
738 gart_iommu_aperture_allowed = 1;
739 if (!strncmp(p, "memaper", 7)) {
740 fallback_aper_force = 1;
744 if (get_option(&p, &arg))
745 fallback_aper_order = arg;