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>
33 #include <asm/iommu.h>
34 #include <asm/cacheflush.h>
35 #include <asm/swiotlb.h>
39 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 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 int iommu_leak_pages = 20;
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 = page_to_phys(s->page) + s->offset;
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);
342 sout->dma_address = iommu_bus_base;
343 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
344 sout->dma_length = s->length;
346 sout->dma_length += s->length;
350 pages = to_pages(s->offset, s->length);
352 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
353 SET_LEAK(iommu_page);
358 BUG_ON(iommu_page - iommu_start != pages);
362 static inline int dma_map_cont(struct scatterlist *start, int nelems,
363 struct scatterlist *sout,
364 unsigned long pages, int need)
369 sout->dma_length = start->length;
372 return __dma_map_cont(start, nelems, sout, pages);
376 * DMA map all entries in a scatterlist.
377 * Merge chunks that have page aligned sizes into a continuous mapping.
379 static int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents,
385 unsigned long pages = 0;
386 int need = 0, nextneed;
387 struct scatterlist *s, *ps, *start_sg, *sgmap;
397 start_sg = sgmap = sg;
398 ps = NULL; /* shut up gcc */
399 for_each_sg(sg, s, nents, i) {
400 dma_addr_t addr = page_to_phys(s->page) + s->offset;
401 s->dma_address = addr;
402 BUG_ON(s->length == 0);
404 nextneed = need_iommu(dev, addr, s->length);
406 /* Handle the previous not yet processed entries */
408 /* Can only merge when the last chunk ends on a page
409 boundary and the new one doesn't have an offset. */
410 if (!iommu_merge || !nextneed || !need || s->offset ||
411 (ps->offset + ps->length) % PAGE_SIZE) {
412 if (dma_map_cont(start_sg, i - start, sgmap,
416 sgmap = sg_next(sgmap);
424 pages += to_pages(s->offset, s->length);
427 if (dma_map_cont(start_sg, i - start, sgmap, pages, need) < 0)
432 sgmap = sg_next(sgmap);
433 sgmap->dma_length = 0;
439 gart_unmap_sg(dev, sg, nents, dir);
440 /* When it was forced or merged try again in a dumb way */
441 if (force_iommu || iommu_merge) {
442 out = dma_map_sg_nonforce(dev, sg, nents, dir);
446 if (panic_on_overflow)
447 panic("dma_map_sg: overflow on %lu pages\n", pages);
448 iommu_full(dev, pages << PAGE_SHIFT, dir);
449 for_each_sg(sg, s, nents, i)
450 s->dma_address = bad_dma_address;
456 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
460 iommu_size = aper_size;
465 a = aper + iommu_size;
466 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
468 if (iommu_size < 64*1024*1024)
470 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
475 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
477 unsigned aper_size = 0, aper_base_32;
481 pci_read_config_dword(dev, 0x94, &aper_base_32);
482 pci_read_config_dword(dev, 0x90, &aper_order);
483 aper_order = (aper_order >> 1) & 7;
485 aper_base = aper_base_32 & 0x7fff;
488 aper_size = (32 * 1024 * 1024) << aper_order;
489 if (aper_base + aper_size > 0x100000000UL || !aper_size)
497 * Private Northbridge GATT initialization in case we cannot use the
498 * AGP driver for some reason.
500 static __init int init_k8_gatt(struct agp_kern_info *info)
504 unsigned aper_base, new_aper_base;
505 unsigned aper_size, gatt_size, new_aper_size;
508 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
509 aper_size = aper_base = info->aper_size = 0;
511 for (i = 0; i < num_k8_northbridges; i++) {
512 dev = k8_northbridges[i];
513 new_aper_base = read_aperture(dev, &new_aper_size);
518 aper_size = new_aper_size;
519 aper_base = new_aper_base;
521 if (aper_size != new_aper_size || aper_base != new_aper_base)
526 info->aper_base = aper_base;
527 info->aper_size = aper_size>>20;
529 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
530 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
532 panic("Cannot allocate GATT table");
533 if (change_page_attr_addr((unsigned long)gatt, gatt_size >> PAGE_SHIFT, PAGE_KERNEL_NOCACHE))
534 panic("Could not set GART PTEs to uncacheable pages");
537 memset(gatt, 0, gatt_size);
538 agp_gatt_table = gatt;
540 for (i = 0; i < num_k8_northbridges; i++) {
544 dev = k8_northbridges[i];
545 gatt_reg = __pa(gatt) >> 12;
547 pci_write_config_dword(dev, 0x98, gatt_reg);
548 pci_read_config_dword(dev, 0x90, &ctl);
551 ctl &= ~((1<<4) | (1<<5));
553 pci_write_config_dword(dev, 0x90, ctl);
557 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
561 /* Should not happen anymore */
562 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
563 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
567 extern int agp_amd64_init(void);
569 static const struct dma_mapping_ops gart_dma_ops = {
570 .mapping_error = NULL,
571 .map_single = gart_map_single,
572 .map_simple = gart_map_simple,
573 .unmap_single = gart_unmap_single,
574 .sync_single_for_cpu = NULL,
575 .sync_single_for_device = NULL,
576 .sync_single_range_for_cpu = NULL,
577 .sync_single_range_for_device = NULL,
578 .sync_sg_for_cpu = NULL,
579 .sync_sg_for_device = NULL,
580 .map_sg = gart_map_sg,
581 .unmap_sg = gart_unmap_sg,
584 void gart_iommu_shutdown(void)
589 if (no_agp && (dma_ops != &gart_dma_ops))
592 for (i = 0; i < num_k8_northbridges; i++) {
595 dev = k8_northbridges[i];
596 pci_read_config_dword(dev, 0x90, &ctl);
600 pci_write_config_dword(dev, 0x90, ctl);
604 void __init gart_iommu_init(void)
606 struct agp_kern_info info;
607 unsigned long aper_size;
608 unsigned long iommu_start;
609 unsigned long scratch;
612 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
613 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
617 #ifndef CONFIG_AGP_AMD64
620 /* Makefile puts PCI initialization via subsys_initcall first. */
621 /* Add other K8 AGP bridge drivers here */
623 (agp_amd64_init() < 0) ||
624 (agp_copy_info(agp_bridge, &info) < 0);
630 /* Did we detect a different HW IOMMU? */
631 if (iommu_detected && !iommu_aperture)
635 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
637 (no_agp && init_k8_gatt(&info) < 0)) {
638 if (end_pfn > MAX_DMA32_PFN) {
639 printk(KERN_ERR "WARNING more than 4GB of memory "
640 "but GART IOMMU not available.\n"
641 KERN_ERR "WARNING 32bit PCI may malfunction.\n");
646 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
647 aper_size = info.aper_size * 1024 * 1024;
648 iommu_size = check_iommu_size(info.aper_base, aper_size);
649 iommu_pages = iommu_size >> PAGE_SHIFT;
651 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
652 get_order(iommu_pages/8));
653 if (!iommu_gart_bitmap)
654 panic("Cannot allocate iommu bitmap\n");
655 memset(iommu_gart_bitmap, 0, iommu_pages/8);
657 #ifdef CONFIG_IOMMU_LEAK
659 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
660 get_order(iommu_pages*sizeof(void *)));
662 memset(iommu_leak_tab, 0, iommu_pages * 8);
664 printk("PCI-DMA: Cannot allocate leak trace area\n");
669 * Out of IOMMU space handling.
670 * Reserve some invalid pages at the beginning of the GART.
672 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
674 agp_memory_reserved = iommu_size;
676 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
679 iommu_start = aper_size - iommu_size;
680 iommu_bus_base = info.aper_base + iommu_start;
681 bad_dma_address = iommu_bus_base;
682 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
685 * Unmap the IOMMU part of the GART. The alias of the page is
686 * always mapped with cache enabled and there is no full cache
687 * coherency across the GART remapping. The unmapping avoids
688 * automatic prefetches from the CPU allocating cache lines in
689 * there. All CPU accesses are done via the direct mapping to
690 * the backing memory. The GART address is only used by PCI
693 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
696 * Try to workaround a bug (thanks to BenH)
697 * Set unmapped entries to a scratch page instead of 0.
698 * Any prefetches that hit unmapped entries won't get an bus abort
701 scratch = get_zeroed_page(GFP_KERNEL);
703 panic("Cannot allocate iommu scratch page");
704 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
705 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
706 iommu_gatt_base[i] = gart_unmapped_entry;
709 dma_ops = &gart_dma_ops;
712 void __init gart_parse_options(char *p)
716 #ifdef CONFIG_IOMMU_LEAK
717 if (!strncmp(p,"leak",4)) {
721 if (isdigit(*p) && get_option(&p, &arg))
722 iommu_leak_pages = arg;
725 if (isdigit(*p) && get_option(&p, &arg))
727 if (!strncmp(p, "fullflush",8))
729 if (!strncmp(p, "nofullflush",11))
731 if (!strncmp(p,"noagp",5))
733 if (!strncmp(p, "noaperture",10))
735 /* duplicated from pci-dma.c */
736 if (!strncmp(p,"force",5))
737 iommu_aperture_allowed = 1;
738 if (!strncmp(p,"allowed",7))
739 iommu_aperture_allowed = 1;
740 if (!strncmp(p, "memaper", 7)) {
741 fallback_aper_force = 1;
745 if (get_option(&p, &arg))
746 fallback_aper_order = arg;