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/config.h>
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 <asm/atomic.h>
29 #include <asm/pgtable.h>
30 #include <asm/proto.h>
31 #include <asm/cacheflush.h>
32 #include <asm/kdebug.h>
33 #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;
51 /* Allocation bitmap for the remapping area */
52 static DEFINE_SPINLOCK(iommu_bitmap_lock);
53 static unsigned long *iommu_gart_bitmap; /* guarded by iommu_bitmap_lock */
55 static u32 gart_unmapped_entry;
58 #define GPTE_COHERENT 2
59 #define GPTE_ENCODE(x) \
60 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
61 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
63 #define to_pages(addr,size) \
64 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
66 #define for_all_nb(dev) \
68 while ((dev = pci_get_device(PCI_VENDOR_ID_AMD, 0x1103, dev))!=NULL)\
69 if (dev->bus->number == 0 && \
70 (PCI_SLOT(dev->devfn) >= 24) && (PCI_SLOT(dev->devfn) <= 31))
72 static struct pci_dev *northbridges[MAX_NB];
73 static u32 northbridge_flush_word[MAX_NB];
75 #define EMERGENCY_PAGES 32 /* = 128KB */
78 #define AGPEXTERN extern
83 /* backdoor interface to AGP driver */
84 AGPEXTERN int agp_memory_reserved;
85 AGPEXTERN __u32 *agp_gatt_table;
87 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
88 static int need_flush; /* global flush state. set for each gart wrap */
90 static unsigned long alloc_iommu(int size)
92 unsigned long offset, flags;
94 spin_lock_irqsave(&iommu_bitmap_lock, flags);
95 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
98 offset = find_next_zero_string(iommu_gart_bitmap,0,next_bit,size);
101 set_bit_string(iommu_gart_bitmap, offset, size);
102 next_bit = offset+size;
103 if (next_bit >= iommu_pages) {
110 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
114 static void free_iommu(unsigned long offset, int size)
118 clear_bit(offset, iommu_gart_bitmap);
121 spin_lock_irqsave(&iommu_bitmap_lock, flags);
122 __clear_bit_string(iommu_gart_bitmap, offset, size);
123 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
127 * Use global flush state to avoid races with multiple flushers.
129 static void flush_gart(struct device *dev)
135 spin_lock_irqsave(&iommu_bitmap_lock, flags);
138 for (i = 0; i < MAX_NB; i++) {
139 if (!northbridges[i])
141 pci_write_config_dword(northbridges[i], 0x9c,
142 northbridge_flush_word[i] | 1);
146 for (i = 0; i <= max; i++) {
148 if (!northbridges[i])
150 /* Make sure the hardware actually executed the flush. */
152 pci_read_config_dword(northbridges[i], 0x9c, &w);
159 printk("nothing to flush?\n");
162 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
167 #ifdef CONFIG_IOMMU_LEAK
169 #define SET_LEAK(x) if (iommu_leak_tab) \
170 iommu_leak_tab[x] = __builtin_return_address(0);
171 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
172 iommu_leak_tab[x] = NULL;
174 /* Debugging aid for drivers that don't free their IOMMU tables */
175 static void **iommu_leak_tab;
176 static int leak_trace;
177 int iommu_leak_pages = 20;
182 if (dump || !iommu_leak_tab) return;
184 show_stack(NULL,NULL);
185 /* Very crude. dump some from the end of the table too */
186 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
187 for (i = 0; i < iommu_leak_pages; i+=2) {
188 printk("%lu: ", iommu_pages-i);
189 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
190 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
196 #define CLEAR_LEAK(x)
199 static void iommu_full(struct device *dev, size_t size, int dir)
202 * Ran out of IOMMU space for this operation. This is very bad.
203 * Unfortunately the drivers cannot handle this operation properly.
204 * Return some non mapped prereserved space in the aperture and
205 * let the Northbridge deal with it. This will result in garbage
206 * in the IO operation. When the size exceeds the prereserved space
207 * memory corruption will occur or random memory will be DMAed
208 * out. Hopefully no network devices use single mappings that big.
212 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
215 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
216 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
217 panic("PCI-DMA: Memory would be corrupted\n");
218 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
219 panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
222 #ifdef CONFIG_IOMMU_LEAK
227 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
229 u64 mask = *dev->dma_mask;
230 int high = addr + size >= mask;
237 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
239 u64 mask = *dev->dma_mask;
240 int high = addr + size >= mask;
245 /* Map a single continuous physical area into the IOMMU.
246 * Caller needs to check if the iommu is needed and flush.
248 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
249 size_t size, int dir)
251 unsigned long npages = to_pages(phys_mem, size);
252 unsigned long iommu_page = alloc_iommu(npages);
254 if (iommu_page == -1) {
255 if (!nonforced_iommu(dev, phys_mem, size))
257 if (panic_on_overflow)
258 panic("dma_map_area overflow %lu bytes\n", size);
259 iommu_full(dev, size, dir);
260 return bad_dma_address;
263 for (i = 0; i < npages; i++) {
264 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
265 SET_LEAK(iommu_page + i);
266 phys_mem += PAGE_SIZE;
268 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
271 static dma_addr_t gart_map_simple(struct device *dev, char *buf,
272 size_t size, int dir)
274 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
279 /* Map a single area into the IOMMU */
280 dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
282 unsigned long phys_mem, bus;
284 BUG_ON(dir == DMA_NONE);
289 phys_mem = virt_to_phys(addr);
290 if (!need_iommu(dev, phys_mem, size))
293 bus = gart_map_simple(dev, addr, size, dir);
298 * Wrapper for pci_unmap_single working with scatterlists.
300 void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
304 for (i = 0; i < nents; i++) {
305 struct scatterlist *s = &sg[i];
306 if (!s->dma_length || !s->length)
308 dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
312 /* Fallback for dma_map_sg in case of overflow */
313 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
318 #ifdef CONFIG_IOMMU_DEBUG
319 printk(KERN_DEBUG "dma_map_sg overflow\n");
322 for (i = 0; i < nents; i++ ) {
323 struct scatterlist *s = &sg[i];
324 unsigned long addr = page_to_phys(s->page) + s->offset;
325 if (nonforced_iommu(dev, addr, s->length)) {
326 addr = dma_map_area(dev, addr, s->length, dir);
327 if (addr == bad_dma_address) {
329 gart_unmap_sg(dev, sg, i, dir);
331 sg[0].dma_length = 0;
335 s->dma_address = addr;
336 s->dma_length = s->length;
342 /* Map multiple scatterlist entries continuous into the first. */
343 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
344 struct scatterlist *sout, unsigned long pages)
346 unsigned long iommu_start = alloc_iommu(pages);
347 unsigned long iommu_page = iommu_start;
350 if (iommu_start == -1)
353 for (i = start; i < stopat; i++) {
354 struct scatterlist *s = &sg[i];
355 unsigned long pages, addr;
356 unsigned long phys_addr = s->dma_address;
358 BUG_ON(i > start && s->offset);
361 sout->dma_address = iommu_bus_base;
362 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
363 sout->dma_length = s->length;
365 sout->dma_length += s->length;
369 pages = to_pages(s->offset, s->length);
371 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
372 SET_LEAK(iommu_page);
377 BUG_ON(iommu_page - iommu_start != pages);
381 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
382 struct scatterlist *sout,
383 unsigned long pages, int need)
386 BUG_ON(stopat - start != 1);
388 sout->dma_length = sg[start].length;
391 return __dma_map_cont(sg, start, stopat, sout, pages);
395 * DMA map all entries in a scatterlist.
396 * Merge chunks that have page aligned sizes into a continuous mapping.
398 int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
403 unsigned long pages = 0;
404 int need = 0, nextneed;
406 BUG_ON(dir == DMA_NONE);
415 for (i = 0; i < nents; i++) {
416 struct scatterlist *s = &sg[i];
417 dma_addr_t addr = page_to_phys(s->page) + s->offset;
418 s->dma_address = addr;
419 BUG_ON(s->length == 0);
421 nextneed = need_iommu(dev, addr, s->length);
423 /* Handle the previous not yet processed entries */
425 struct scatterlist *ps = &sg[i-1];
426 /* Can only merge when the last chunk ends on a page
427 boundary and the new one doesn't have an offset. */
428 if (!iommu_merge || !nextneed || !need || s->offset ||
429 (ps->offset + ps->length) % PAGE_SIZE) {
430 if (dma_map_cont(sg, start, i, sg+out, pages,
440 pages += to_pages(s->offset, s->length);
442 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
447 sg[out].dma_length = 0;
452 gart_unmap_sg(dev, sg, nents, dir);
453 /* When it was forced or merged try again in a dumb way */
454 if (force_iommu || iommu_merge) {
455 out = dma_map_sg_nonforce(dev, sg, nents, dir);
459 if (panic_on_overflow)
460 panic("dma_map_sg: overflow on %lu pages\n", pages);
461 iommu_full(dev, pages << PAGE_SHIFT, dir);
462 for (i = 0; i < nents; i++)
463 sg[i].dma_address = bad_dma_address;
468 * Free a DMA mapping.
470 void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
471 size_t size, int direction)
473 unsigned long iommu_page;
477 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
478 dma_addr >= iommu_bus_base + iommu_size)
480 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
481 npages = to_pages(dma_addr, size);
482 for (i = 0; i < npages; i++) {
483 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
484 CLEAR_LEAK(iommu_page + i);
486 free_iommu(iommu_page, npages);
491 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
495 iommu_size = aper_size;
500 a = aper + iommu_size;
501 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
503 if (iommu_size < 64*1024*1024)
505 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
510 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
512 unsigned aper_size = 0, aper_base_32;
516 pci_read_config_dword(dev, 0x94, &aper_base_32);
517 pci_read_config_dword(dev, 0x90, &aper_order);
518 aper_order = (aper_order >> 1) & 7;
520 aper_base = aper_base_32 & 0x7fff;
523 aper_size = (32 * 1024 * 1024) << aper_order;
524 if (aper_base + aper_size >= 0xffffffff || !aper_size)
532 * Private Northbridge GATT initialization in case we cannot use the
533 * AGP driver for some reason.
535 static __init int init_k8_gatt(struct agp_kern_info *info)
539 unsigned aper_base, new_aper_base;
540 unsigned aper_size, gatt_size, new_aper_size;
542 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
543 aper_size = aper_base = info->aper_size = 0;
545 new_aper_base = read_aperture(dev, &new_aper_size);
550 aper_size = new_aper_size;
551 aper_base = new_aper_base;
553 if (aper_size != new_aper_size || aper_base != new_aper_base)
558 info->aper_base = aper_base;
559 info->aper_size = aper_size>>20;
561 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
562 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
564 panic("Cannot allocate GATT table");
565 memset(gatt, 0, gatt_size);
566 agp_gatt_table = gatt;
572 gatt_reg = __pa(gatt) >> 12;
574 pci_write_config_dword(dev, 0x98, gatt_reg);
575 pci_read_config_dword(dev, 0x90, &ctl);
578 ctl &= ~((1<<4) | (1<<5));
580 pci_write_config_dword(dev, 0x90, ctl);
584 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
588 /* Should not happen anymore */
589 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
590 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
594 extern int agp_amd64_init(void);
596 static struct dma_mapping_ops gart_dma_ops = {
597 .mapping_error = NULL,
598 .map_single = gart_map_single,
599 .map_simple = gart_map_simple,
600 .unmap_single = gart_unmap_single,
601 .sync_single_for_cpu = NULL,
602 .sync_single_for_device = NULL,
603 .sync_single_range_for_cpu = NULL,
604 .sync_single_range_for_device = NULL,
605 .sync_sg_for_cpu = NULL,
606 .sync_sg_for_device = NULL,
607 .map_sg = gart_map_sg,
608 .unmap_sg = gart_unmap_sg,
611 static int __init pci_iommu_init(void)
613 struct agp_kern_info info;
614 unsigned long aper_size;
615 unsigned long iommu_start;
617 unsigned long scratch;
620 #ifndef CONFIG_AGP_AMD64
623 /* Makefile puts PCI initialization via subsys_initcall first. */
624 /* Add other K8 AGP bridge drivers here */
626 (agp_amd64_init() < 0) ||
627 (agp_copy_info(agp_bridge, &info) < 0);
634 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
636 (no_agp && init_k8_gatt(&info) < 0)) {
637 printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n");
638 if (end_pfn > MAX_DMA32_PFN) {
639 printk(KERN_ERR "WARNING more than 4GB of memory "
640 "but IOMMU not compiled in.\n"
641 KERN_ERR "WARNING 32bit PCI may malfunction.\n"
642 KERN_ERR "You might want to enable "
643 "CONFIG_GART_IOMMU\n");
648 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
649 aper_size = info.aper_size * 1024 * 1024;
650 iommu_size = check_iommu_size(info.aper_base, aper_size);
651 iommu_pages = iommu_size >> PAGE_SHIFT;
653 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
654 get_order(iommu_pages/8));
655 if (!iommu_gart_bitmap)
656 panic("Cannot allocate iommu bitmap\n");
657 memset(iommu_gart_bitmap, 0, iommu_pages/8);
659 #ifdef CONFIG_IOMMU_LEAK
661 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
662 get_order(iommu_pages*sizeof(void *)));
664 memset(iommu_leak_tab, 0, iommu_pages * 8);
666 printk("PCI-DMA: Cannot allocate leak trace area\n");
671 * Out of IOMMU space handling.
672 * Reserve some invalid pages at the beginning of the GART.
674 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
676 agp_memory_reserved = iommu_size;
678 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
681 iommu_start = aper_size - iommu_size;
682 iommu_bus_base = info.aper_base + iommu_start;
683 bad_dma_address = iommu_bus_base;
684 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
687 * Unmap the IOMMU part of the GART. The alias of the page is
688 * always mapped with cache enabled and there is no full cache
689 * coherency across the GART remapping. The unmapping avoids
690 * automatic prefetches from the CPU allocating cache lines in
691 * there. All CPU accesses are done via the direct mapping to
692 * the backing memory. The GART address is only used by PCI
695 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
698 * Try to workaround a bug (thanks to BenH)
699 * Set unmapped entries to a scratch page instead of 0.
700 * Any prefetches that hit unmapped entries won't get an bus abort
703 scratch = get_zeroed_page(GFP_KERNEL);
705 panic("Cannot allocate iommu scratch page");
706 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
707 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
708 iommu_gatt_base[i] = gart_unmapped_entry;
712 int cpu = PCI_SLOT(dev->devfn) - 24;
715 northbridges[cpu] = dev;
716 pci_read_config_dword(dev, 0x9c, &flag); /* cache flush word */
717 northbridge_flush_word[cpu] = flag;
722 dma_ops = &gart_dma_ops;
727 /* Must execute after PCI subsystem */
728 fs_initcall(pci_iommu_init);
730 void gart_parse_options(char *p)
734 #ifdef CONFIG_IOMMU_LEAK
735 if (!strncmp(p,"leak",4)) {
739 if (isdigit(*p) && get_option(&p, &arg))
740 iommu_leak_pages = arg;
743 if (isdigit(*p) && get_option(&p, &arg))
745 if (!strncmp(p, "fullflush",8))
747 if (!strncmp(p, "nofullflush",11))
749 if (!strncmp(p,"noagp",5))
751 if (!strncmp(p, "noaperture",10))
753 /* duplicated from pci-dma.c */
754 if (!strncmp(p,"force",5))
755 iommu_aperture_allowed = 1;
756 if (!strncmp(p,"allowed",7))
757 iommu_aperture_allowed = 1;
758 if (!strncmp(p, "memaper", 7)) {
759 fallback_aper_force = 1;
763 if (get_option(&p, &arg))
764 fallback_aper_order = arg;