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.
13 #include <linux/types.h>
14 #include <linux/ctype.h>
15 #include <linux/agp_backend.h>
16 #include <linux/init.h>
18 #include <linux/string.h>
19 #include <linux/spinlock.h>
20 #include <linux/pci.h>
21 #include <linux/module.h>
22 #include <linux/topology.h>
23 #include <linux/interrupt.h>
24 #include <linux/bitops.h>
25 #include <asm/atomic.h>
28 #include <asm/pgtable.h>
29 #include <asm/proto.h>
30 #include <asm/cacheflush.h>
31 #include <asm/kdebug.h>
32 #include <asm/swiotlb.h>
36 unsigned long iommu_bus_base; /* GART remapping area (physical) */
37 static unsigned long iommu_size; /* size of remapping area bytes */
38 static unsigned long iommu_pages; /* .. and in pages */
40 u32 *iommu_gatt_base; /* Remapping table */
42 /* If this is disabled the IOMMU will use an optimized flushing strategy
43 of only flushing when an mapping is reused. With it true the GART is flushed
44 for every mapping. Problem is that doing the lazy flush seems to trigger
45 bugs with some popular PCI cards, in particular 3ware (but has been also
46 also seen with Qlogic at least). */
47 int iommu_fullflush = 1;
49 /* Allocation bitmap for the remapping area */
50 static DEFINE_SPINLOCK(iommu_bitmap_lock);
51 static unsigned long *iommu_gart_bitmap; /* guarded by iommu_bitmap_lock */
53 static u32 gart_unmapped_entry;
56 #define GPTE_COHERENT 2
57 #define GPTE_ENCODE(x) \
58 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
59 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
61 #define to_pages(addr,size) \
62 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
64 #define EMERGENCY_PAGES 32 /* = 128KB */
67 #define AGPEXTERN extern
72 /* backdoor interface to AGP driver */
73 AGPEXTERN int agp_memory_reserved;
74 AGPEXTERN __u32 *agp_gatt_table;
76 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
77 static int need_flush; /* global flush state. set for each gart wrap */
79 static unsigned long alloc_iommu(int size)
81 unsigned long offset, flags;
83 spin_lock_irqsave(&iommu_bitmap_lock, flags);
84 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
87 offset = find_next_zero_string(iommu_gart_bitmap,0,iommu_pages,size);
90 set_bit_string(iommu_gart_bitmap, offset, size);
91 next_bit = offset+size;
92 if (next_bit >= iommu_pages) {
99 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
103 static void free_iommu(unsigned long offset, int size)
106 spin_lock_irqsave(&iommu_bitmap_lock, flags);
107 __clear_bit_string(iommu_gart_bitmap, offset, size);
108 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
112 * Use global flush state to avoid races with multiple flushers.
114 static void flush_gart(void)
117 spin_lock_irqsave(&iommu_bitmap_lock, flags);
122 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
125 #ifdef CONFIG_IOMMU_LEAK
127 #define SET_LEAK(x) if (iommu_leak_tab) \
128 iommu_leak_tab[x] = __builtin_return_address(0);
129 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
130 iommu_leak_tab[x] = NULL;
132 /* Debugging aid for drivers that don't free their IOMMU tables */
133 static void **iommu_leak_tab;
134 static int leak_trace;
135 int iommu_leak_pages = 20;
140 if (dump || !iommu_leak_tab) return;
142 show_stack(NULL,NULL);
143 /* Very crude. dump some from the end of the table too */
144 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
145 for (i = 0; i < iommu_leak_pages; i+=2) {
146 printk("%lu: ", iommu_pages-i);
147 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
148 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
154 #define CLEAR_LEAK(x)
157 static void iommu_full(struct device *dev, size_t size, int dir)
160 * Ran out of IOMMU space for this operation. This is very bad.
161 * Unfortunately the drivers cannot handle this operation properly.
162 * Return some non mapped prereserved space in the aperture and
163 * let the Northbridge deal with it. This will result in garbage
164 * in the IO operation. When the size exceeds the prereserved space
165 * memory corruption will occur or random memory will be DMAed
166 * out. Hopefully no network devices use single mappings that big.
170 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
173 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
174 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
175 panic("PCI-DMA: Memory would be corrupted\n");
176 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
177 panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
180 #ifdef CONFIG_IOMMU_LEAK
185 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
187 u64 mask = *dev->dma_mask;
188 int high = addr + size >= mask;
195 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
197 u64 mask = *dev->dma_mask;
198 int high = addr + size >= mask;
203 /* Map a single continuous physical area into the IOMMU.
204 * Caller needs to check if the iommu is needed and flush.
206 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
207 size_t size, int dir)
209 unsigned long npages = to_pages(phys_mem, size);
210 unsigned long iommu_page = alloc_iommu(npages);
212 if (iommu_page == -1) {
213 if (!nonforced_iommu(dev, phys_mem, size))
215 if (panic_on_overflow)
216 panic("dma_map_area overflow %lu bytes\n", size);
217 iommu_full(dev, size, dir);
218 return bad_dma_address;
221 for (i = 0; i < npages; i++) {
222 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
223 SET_LEAK(iommu_page + i);
224 phys_mem += PAGE_SIZE;
226 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
229 static dma_addr_t gart_map_simple(struct device *dev, char *buf,
230 size_t size, int dir)
232 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
237 /* Map a single area into the IOMMU */
238 dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
240 unsigned long phys_mem, bus;
242 BUG_ON(dir == DMA_NONE);
247 phys_mem = virt_to_phys(addr);
248 if (!need_iommu(dev, phys_mem, size))
251 bus = gart_map_simple(dev, addr, size, dir);
256 * Free a DMA mapping.
258 void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
259 size_t size, int direction)
261 unsigned long iommu_page;
265 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
266 dma_addr >= iommu_bus_base + iommu_size)
268 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
269 npages = to_pages(dma_addr, size);
270 for (i = 0; i < npages; i++) {
271 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
272 CLEAR_LEAK(iommu_page + i);
274 free_iommu(iommu_page, npages);
278 * Wrapper for pci_unmap_single working with scatterlists.
280 void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
284 for (i = 0; i < nents; i++) {
285 struct scatterlist *s = &sg[i];
286 if (!s->dma_length || !s->length)
288 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
292 /* Fallback for dma_map_sg in case of overflow */
293 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
298 #ifdef CONFIG_IOMMU_DEBUG
299 printk(KERN_DEBUG "dma_map_sg overflow\n");
302 for (i = 0; i < nents; i++ ) {
303 struct scatterlist *s = &sg[i];
304 unsigned long addr = page_to_phys(s->page) + s->offset;
305 if (nonforced_iommu(dev, addr, s->length)) {
306 addr = dma_map_area(dev, addr, s->length, dir);
307 if (addr == bad_dma_address) {
309 gart_unmap_sg(dev, sg, i, dir);
311 sg[0].dma_length = 0;
315 s->dma_address = addr;
316 s->dma_length = s->length;
322 /* Map multiple scatterlist entries continuous into the first. */
323 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
324 struct scatterlist *sout, unsigned long pages)
326 unsigned long iommu_start = alloc_iommu(pages);
327 unsigned long iommu_page = iommu_start;
330 if (iommu_start == -1)
333 for (i = start; i < stopat; i++) {
334 struct scatterlist *s = &sg[i];
335 unsigned long pages, addr;
336 unsigned long phys_addr = s->dma_address;
338 BUG_ON(i > 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 *sg, int start, int stopat,
362 struct scatterlist *sout,
363 unsigned long pages, int need)
366 BUG_ON(stopat - start != 1);
368 sout->dma_length = sg[start].length;
371 return __dma_map_cont(sg, start, stopat, sout, pages);
375 * DMA map all entries in a scatterlist.
376 * Merge chunks that have page aligned sizes into a continuous mapping.
378 int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
383 unsigned long pages = 0;
384 int need = 0, nextneed;
386 BUG_ON(dir == DMA_NONE);
395 for (i = 0; i < nents; i++) {
396 struct scatterlist *s = &sg[i];
397 dma_addr_t addr = page_to_phys(s->page) + s->offset;
398 s->dma_address = addr;
399 BUG_ON(s->length == 0);
401 nextneed = need_iommu(dev, addr, s->length);
403 /* Handle the previous not yet processed entries */
405 struct scatterlist *ps = &sg[i-1];
406 /* Can only merge when the last chunk ends on a page
407 boundary and the new one doesn't have an offset. */
408 if (!iommu_merge || !nextneed || !need || s->offset ||
409 (ps->offset + ps->length) % PAGE_SIZE) {
410 if (dma_map_cont(sg, start, i, sg+out, pages,
420 pages += to_pages(s->offset, s->length);
422 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
427 sg[out].dma_length = 0;
432 gart_unmap_sg(dev, sg, nents, dir);
433 /* When it was forced or merged try again in a dumb way */
434 if (force_iommu || iommu_merge) {
435 out = dma_map_sg_nonforce(dev, sg, nents, dir);
439 if (panic_on_overflow)
440 panic("dma_map_sg: overflow on %lu pages\n", pages);
441 iommu_full(dev, pages << PAGE_SHIFT, dir);
442 for (i = 0; i < nents; i++)
443 sg[i].dma_address = bad_dma_address;
449 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
453 iommu_size = aper_size;
458 a = aper + iommu_size;
459 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
461 if (iommu_size < 64*1024*1024)
463 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
468 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
470 unsigned aper_size = 0, aper_base_32;
474 pci_read_config_dword(dev, 0x94, &aper_base_32);
475 pci_read_config_dword(dev, 0x90, &aper_order);
476 aper_order = (aper_order >> 1) & 7;
478 aper_base = aper_base_32 & 0x7fff;
481 aper_size = (32 * 1024 * 1024) << aper_order;
482 if (aper_base + aper_size >= 0xffffffff || !aper_size)
490 * Private Northbridge GATT initialization in case we cannot use the
491 * AGP driver for some reason.
493 static __init int init_k8_gatt(struct agp_kern_info *info)
497 unsigned aper_base, new_aper_base;
498 unsigned aper_size, gatt_size, new_aper_size;
501 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
502 aper_size = aper_base = info->aper_size = 0;
504 for (i = 0; i < num_k8_northbridges; i++) {
505 dev = k8_northbridges[i];
506 new_aper_base = read_aperture(dev, &new_aper_size);
511 aper_size = new_aper_size;
512 aper_base = new_aper_base;
514 if (aper_size != new_aper_size || aper_base != new_aper_base)
519 info->aper_base = aper_base;
520 info->aper_size = aper_size>>20;
522 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
523 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
525 panic("Cannot allocate GATT table");
526 memset(gatt, 0, gatt_size);
527 agp_gatt_table = gatt;
529 for (i = 0; i < num_k8_northbridges; i++) {
533 dev = k8_northbridges[i];
534 gatt_reg = __pa(gatt) >> 12;
536 pci_write_config_dword(dev, 0x98, gatt_reg);
537 pci_read_config_dword(dev, 0x90, &ctl);
540 ctl &= ~((1<<4) | (1<<5));
542 pci_write_config_dword(dev, 0x90, ctl);
546 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
550 /* Should not happen anymore */
551 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
552 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
556 extern int agp_amd64_init(void);
558 static struct dma_mapping_ops gart_dma_ops = {
559 .mapping_error = NULL,
560 .map_single = gart_map_single,
561 .map_simple = gart_map_simple,
562 .unmap_single = gart_unmap_single,
563 .sync_single_for_cpu = NULL,
564 .sync_single_for_device = NULL,
565 .sync_single_range_for_cpu = NULL,
566 .sync_single_range_for_device = NULL,
567 .sync_sg_for_cpu = NULL,
568 .sync_sg_for_device = NULL,
569 .map_sg = gart_map_sg,
570 .unmap_sg = gart_unmap_sg,
573 void __init gart_iommu_init(void)
575 struct agp_kern_info info;
576 unsigned long aper_size;
577 unsigned long iommu_start;
578 unsigned long scratch;
581 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
582 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
586 #ifndef CONFIG_AGP_AMD64
589 /* Makefile puts PCI initialization via subsys_initcall first. */
590 /* Add other K8 AGP bridge drivers here */
592 (agp_amd64_init() < 0) ||
593 (agp_copy_info(agp_bridge, &info) < 0);
599 /* Did we detect a different HW IOMMU? */
600 if (iommu_detected && !iommu_aperture)
604 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
606 (no_agp && init_k8_gatt(&info) < 0)) {
607 printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n");
608 if (end_pfn > MAX_DMA32_PFN) {
609 printk(KERN_ERR "WARNING more than 4GB of memory "
610 "but IOMMU not available.\n"
611 KERN_ERR "WARNING 32bit PCI may malfunction.\n");
616 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
617 aper_size = info.aper_size * 1024 * 1024;
618 iommu_size = check_iommu_size(info.aper_base, aper_size);
619 iommu_pages = iommu_size >> PAGE_SHIFT;
621 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
622 get_order(iommu_pages/8));
623 if (!iommu_gart_bitmap)
624 panic("Cannot allocate iommu bitmap\n");
625 memset(iommu_gart_bitmap, 0, iommu_pages/8);
627 #ifdef CONFIG_IOMMU_LEAK
629 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
630 get_order(iommu_pages*sizeof(void *)));
632 memset(iommu_leak_tab, 0, iommu_pages * 8);
634 printk("PCI-DMA: Cannot allocate leak trace area\n");
639 * Out of IOMMU space handling.
640 * Reserve some invalid pages at the beginning of the GART.
642 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
644 agp_memory_reserved = iommu_size;
646 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
649 iommu_start = aper_size - iommu_size;
650 iommu_bus_base = info.aper_base + iommu_start;
651 bad_dma_address = iommu_bus_base;
652 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
655 * Unmap the IOMMU part of the GART. The alias of the page is
656 * always mapped with cache enabled and there is no full cache
657 * coherency across the GART remapping. The unmapping avoids
658 * automatic prefetches from the CPU allocating cache lines in
659 * there. All CPU accesses are done via the direct mapping to
660 * the backing memory. The GART address is only used by PCI
663 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
666 * Try to workaround a bug (thanks to BenH)
667 * Set unmapped entries to a scratch page instead of 0.
668 * Any prefetches that hit unmapped entries won't get an bus abort
671 scratch = get_zeroed_page(GFP_KERNEL);
673 panic("Cannot allocate iommu scratch page");
674 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
675 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
676 iommu_gatt_base[i] = gart_unmapped_entry;
679 dma_ops = &gart_dma_ops;
682 void gart_parse_options(char *p)
686 #ifdef CONFIG_IOMMU_LEAK
687 if (!strncmp(p,"leak",4)) {
691 if (isdigit(*p) && get_option(&p, &arg))
692 iommu_leak_pages = arg;
695 if (isdigit(*p) && get_option(&p, &arg))
697 if (!strncmp(p, "fullflush",8))
699 if (!strncmp(p, "nofullflush",11))
701 if (!strncmp(p,"noagp",5))
703 if (!strncmp(p, "noaperture",10))
705 /* duplicated from pci-dma.c */
706 if (!strncmp(p,"force",5))
707 iommu_aperture_allowed = 1;
708 if (!strncmp(p,"allowed",7))
709 iommu_aperture_allowed = 1;
710 if (!strncmp(p, "memaper", 7)) {
711 fallback_aper_force = 1;
715 if (get_option(&p, &arg))
716 fallback_aper_order = arg;