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 <linux/iommu-helper.h>
29 #include <linux/sysdev.h>
30 #include <asm/atomic.h>
33 #include <asm/pgtable.h>
34 #include <asm/proto.h>
35 #include <asm/iommu.h>
37 #include <asm/cacheflush.h>
38 #include <asm/swiotlb.h>
42 static unsigned long iommu_bus_base; /* GART remapping area (physical) */
43 static unsigned long iommu_size; /* size of remapping area bytes */
44 static unsigned long iommu_pages; /* .. and in pages */
46 static u32 *iommu_gatt_base; /* Remapping table */
49 * If this is disabled the IOMMU will use an optimized flushing strategy
50 * of only flushing when an mapping is reused. With it true the GART is
51 * flushed for every mapping. Problem is that doing the lazy flush seems
52 * to trigger bugs with some popular PCI cards, in particular 3ware (but
53 * has been also also seen with Qlogic at least).
55 int iommu_fullflush = 1;
57 /* Allocation bitmap for the remapping area: */
58 static DEFINE_SPINLOCK(iommu_bitmap_lock);
59 /* Guarded by iommu_bitmap_lock: */
60 static unsigned long *iommu_gart_bitmap;
62 static u32 gart_unmapped_entry;
65 #define GPTE_COHERENT 2
66 #define GPTE_ENCODE(x) \
67 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
68 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
70 #define to_pages(addr, size) \
71 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
73 #define EMERGENCY_PAGES 32 /* = 128KB */
76 #define AGPEXTERN extern
81 /* backdoor interface to AGP driver */
82 AGPEXTERN int agp_memory_reserved;
83 AGPEXTERN __u32 *agp_gatt_table;
85 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
86 static int need_flush; /* global flush state. set for each gart wrap */
88 static unsigned long alloc_iommu(struct device *dev, int size)
90 unsigned long offset, flags;
91 unsigned long boundary_size;
92 unsigned long base_index;
94 base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
95 PAGE_SIZE) >> PAGE_SHIFT;
96 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
97 PAGE_SIZE) >> PAGE_SHIFT;
99 spin_lock_irqsave(&iommu_bitmap_lock, flags);
100 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
101 size, base_index, boundary_size, 0);
104 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
105 size, base_index, boundary_size, 0);
108 next_bit = offset+size;
109 if (next_bit >= iommu_pages) {
116 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
121 static void free_iommu(unsigned long offset, int size)
125 spin_lock_irqsave(&iommu_bitmap_lock, flags);
126 iommu_area_free(iommu_gart_bitmap, offset, size);
127 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
131 * Use global flush state to avoid races with multiple flushers.
133 static void flush_gart(void)
137 spin_lock_irqsave(&iommu_bitmap_lock, flags);
142 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
145 #ifdef CONFIG_IOMMU_LEAK
147 #define SET_LEAK(x) \
149 if (iommu_leak_tab) \
150 iommu_leak_tab[x] = __builtin_return_address(0);\
153 #define CLEAR_LEAK(x) \
155 if (iommu_leak_tab) \
156 iommu_leak_tab[x] = NULL; \
159 /* Debugging aid for drivers that don't free their IOMMU tables */
160 static void **iommu_leak_tab;
161 static int leak_trace;
162 static int iommu_leak_pages = 20;
164 static void dump_leak(void)
169 if (dump || !iommu_leak_tab)
172 show_stack(NULL, NULL);
174 /* Very crude. dump some from the end of the table too */
175 printk(KERN_DEBUG "Dumping %d pages from end of IOMMU:\n",
177 for (i = 0; i < iommu_leak_pages; i += 2) {
178 printk(KERN_DEBUG "%lu: ", iommu_pages-i);
179 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i], 0);
180 printk(KERN_CONT "%c", (i+1)%2 == 0 ? '\n' : ' ');
182 printk(KERN_DEBUG "\n");
186 # define CLEAR_LEAK(x)
189 static void iommu_full(struct device *dev, size_t size, int dir)
192 * Ran out of IOMMU space for this operation. This is very bad.
193 * Unfortunately the drivers cannot handle this operation properly.
194 * Return some non mapped prereserved space in the aperture and
195 * let the Northbridge deal with it. This will result in garbage
196 * in the IO operation. When the size exceeds the prereserved space
197 * memory corruption will occur or random memory will be DMAed
198 * out. Hopefully no network devices use single mappings that big.
201 dev_err(dev, "PCI-DMA: Out of IOMMU space for %lu bytes\n", size);
203 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
204 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
205 panic("PCI-DMA: Memory would be corrupted\n");
206 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
208 "PCI-DMA: Random memory would be DMAed\n");
210 #ifdef CONFIG_IOMMU_LEAK
216 need_iommu(struct device *dev, unsigned long addr, size_t size)
218 u64 mask = *dev->dma_mask;
219 int high = addr + size > mask;
229 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
231 u64 mask = *dev->dma_mask;
232 int high = addr + size > mask;
238 /* Map a single continuous physical area into the IOMMU.
239 * Caller needs to check if the iommu is needed and flush.
241 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
242 size_t size, int dir)
244 unsigned long npages = to_pages(phys_mem, size);
245 unsigned long iommu_page = alloc_iommu(dev, npages);
248 if (iommu_page == -1) {
249 if (!nonforced_iommu(dev, phys_mem, size))
251 if (panic_on_overflow)
252 panic("dma_map_area overflow %lu bytes\n", size);
253 iommu_full(dev, size, dir);
254 return bad_dma_address;
257 for (i = 0; i < npages; i++) {
258 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
259 SET_LEAK(iommu_page + i);
260 phys_mem += PAGE_SIZE;
262 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
266 gart_map_simple(struct device *dev, phys_addr_t paddr, size_t size, int dir)
268 dma_addr_t map = dma_map_area(dev, paddr, size, dir);
275 /* Map a single area into the IOMMU */
277 gart_map_single(struct device *dev, phys_addr_t paddr, size_t size, int dir)
284 if (!need_iommu(dev, paddr, size))
287 bus = gart_map_simple(dev, paddr, size, dir);
293 * Free a DMA mapping.
295 static void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
296 size_t size, int direction)
298 unsigned long iommu_page;
302 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
303 dma_addr >= iommu_bus_base + iommu_size)
306 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
307 npages = to_pages(dma_addr, size);
308 for (i = 0; i < npages; i++) {
309 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
310 CLEAR_LEAK(iommu_page + i);
312 free_iommu(iommu_page, npages);
316 * Wrapper for pci_unmap_single working with scatterlists.
319 gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
321 struct scatterlist *s;
324 for_each_sg(sg, s, nents, i) {
325 if (!s->dma_length || !s->length)
327 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
331 /* Fallback for dma_map_sg in case of overflow */
332 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
335 struct scatterlist *s;
338 #ifdef CONFIG_IOMMU_DEBUG
339 printk(KERN_DEBUG "dma_map_sg overflow\n");
342 for_each_sg(sg, s, nents, i) {
343 unsigned long addr = sg_phys(s);
345 if (nonforced_iommu(dev, addr, s->length)) {
346 addr = dma_map_area(dev, addr, s->length, dir);
347 if (addr == bad_dma_address) {
349 gart_unmap_sg(dev, sg, i, dir);
351 sg[0].dma_length = 0;
355 s->dma_address = addr;
356 s->dma_length = s->length;
363 /* Map multiple scatterlist entries continuous into the first. */
364 static int __dma_map_cont(struct device *dev, struct scatterlist *start,
365 int nelems, struct scatterlist *sout,
368 unsigned long iommu_start = alloc_iommu(dev, pages);
369 unsigned long iommu_page = iommu_start;
370 struct scatterlist *s;
373 if (iommu_start == -1)
376 for_each_sg(start, s, nelems, i) {
377 unsigned long pages, addr;
378 unsigned long phys_addr = s->dma_address;
380 BUG_ON(s != start && s->offset);
382 sout->dma_address = iommu_bus_base;
383 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
384 sout->dma_length = s->length;
386 sout->dma_length += s->length;
390 pages = to_pages(s->offset, s->length);
392 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
393 SET_LEAK(iommu_page);
398 BUG_ON(iommu_page - iommu_start != pages);
404 dma_map_cont(struct device *dev, struct scatterlist *start, int nelems,
405 struct scatterlist *sout, unsigned long pages, int need)
409 sout->dma_address = start->dma_address;
410 sout->dma_length = start->length;
413 return __dma_map_cont(dev, start, nelems, sout, pages);
417 * DMA map all entries in a scatterlist.
418 * Merge chunks that have page aligned sizes into a continuous mapping.
421 gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
423 struct scatterlist *s, *ps, *start_sg, *sgmap;
424 int need = 0, nextneed, i, out, start;
425 unsigned long pages = 0;
426 unsigned int seg_size;
427 unsigned int max_seg_size;
437 start_sg = sgmap = sg;
439 max_seg_size = dma_get_max_seg_size(dev);
440 ps = NULL; /* shut up gcc */
441 for_each_sg(sg, s, nents, i) {
442 dma_addr_t addr = sg_phys(s);
444 s->dma_address = addr;
445 BUG_ON(s->length == 0);
447 nextneed = need_iommu(dev, addr, s->length);
449 /* Handle the previous not yet processed entries */
452 * Can only merge when the last chunk ends on a
453 * page boundary and the new one doesn't have an
456 if (!iommu_merge || !nextneed || !need || s->offset ||
457 (s->length + seg_size > max_seg_size) ||
458 (ps->offset + ps->length) % PAGE_SIZE) {
459 if (dma_map_cont(dev, start_sg, i - start,
460 sgmap, pages, need) < 0)
464 sgmap = sg_next(sgmap);
471 seg_size += s->length;
473 pages += to_pages(s->offset, s->length);
476 if (dma_map_cont(dev, start_sg, i - start, sgmap, pages, need) < 0)
481 sgmap = sg_next(sgmap);
482 sgmap->dma_length = 0;
488 gart_unmap_sg(dev, sg, out, dir);
490 /* When it was forced or merged try again in a dumb way */
491 if (force_iommu || iommu_merge) {
492 out = dma_map_sg_nonforce(dev, sg, nents, dir);
496 if (panic_on_overflow)
497 panic("dma_map_sg: overflow on %lu pages\n", pages);
499 iommu_full(dev, pages << PAGE_SHIFT, dir);
500 for_each_sg(sg, s, nents, i)
501 s->dma_address = bad_dma_address;
507 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
512 iommu_size = aper_size;
517 a = aper + iommu_size;
518 iommu_size -= round_up(a, PMD_PAGE_SIZE) - a;
520 if (iommu_size < 64*1024*1024) {
522 "PCI-DMA: Warning: Small IOMMU %luMB."
523 " Consider increasing the AGP aperture in BIOS\n",
530 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
532 unsigned aper_size = 0, aper_base_32, aper_order;
535 pci_read_config_dword(dev, AMD64_GARTAPERTUREBASE, &aper_base_32);
536 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &aper_order);
537 aper_order = (aper_order >> 1) & 7;
539 aper_base = aper_base_32 & 0x7fff;
542 aper_size = (32 * 1024 * 1024) << aper_order;
543 if (aper_base + aper_size > 0x100000000UL || !aper_size)
550 static void enable_gart_translations(void)
554 for (i = 0; i < num_k8_northbridges; i++) {
555 struct pci_dev *dev = k8_northbridges[i];
557 enable_gart_translation(dev, __pa(agp_gatt_table));
562 * If fix_up_north_bridges is set, the north bridges have to be fixed up on
563 * resume in the same way as they are handled in gart_iommu_hole_init().
565 static bool fix_up_north_bridges;
566 static u32 aperture_order;
567 static u32 aperture_alloc;
569 void set_up_gart_resume(u32 aper_order, u32 aper_alloc)
571 fix_up_north_bridges = true;
572 aperture_order = aper_order;
573 aperture_alloc = aper_alloc;
576 static int gart_resume(struct sys_device *dev)
578 printk(KERN_INFO "PCI-DMA: Resuming GART IOMMU\n");
580 if (fix_up_north_bridges) {
583 printk(KERN_INFO "PCI-DMA: Restoring GART aperture settings\n");
585 for (i = 0; i < num_k8_northbridges; i++) {
586 struct pci_dev *dev = k8_northbridges[i];
589 * Don't enable translations just yet. That is the next
590 * step. Restore the pre-suspend aperture settings.
592 pci_write_config_dword(dev, AMD64_GARTAPERTURECTL,
593 aperture_order << 1);
594 pci_write_config_dword(dev, AMD64_GARTAPERTUREBASE,
595 aperture_alloc >> 25);
599 enable_gart_translations();
604 static int gart_suspend(struct sys_device *dev, pm_message_t state)
609 static struct sysdev_class gart_sysdev_class = {
611 .suspend = gart_suspend,
612 .resume = gart_resume,
616 static struct sys_device device_gart = {
618 .cls = &gart_sysdev_class,
622 * Private Northbridge GATT initialization in case we cannot use the
623 * AGP driver for some reason.
625 static __init int init_k8_gatt(struct agp_kern_info *info)
627 unsigned aper_size, gatt_size, new_aper_size;
628 unsigned aper_base, new_aper_base;
632 unsigned long start_pfn, end_pfn;
634 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
635 aper_size = aper_base = info->aper_size = 0;
637 for (i = 0; i < num_k8_northbridges; i++) {
638 dev = k8_northbridges[i];
639 new_aper_base = read_aperture(dev, &new_aper_size);
644 aper_size = new_aper_size;
645 aper_base = new_aper_base;
647 if (aper_size != new_aper_size || aper_base != new_aper_base)
652 info->aper_base = aper_base;
653 info->aper_size = aper_size >> 20;
655 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
656 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
658 panic("Cannot allocate GATT table");
659 if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
660 panic("Could not set GART PTEs to uncacheable pages");
662 memset(gatt, 0, gatt_size);
663 agp_gatt_table = gatt;
665 enable_gart_translations();
667 error = sysdev_class_register(&gart_sysdev_class);
669 error = sysdev_register(&device_gart);
671 panic("Could not register gart_sysdev -- would corrupt data on next suspend");
675 printk(KERN_INFO "PCI-DMA: aperture base @ %x size %u KB\n",
676 aper_base, aper_size>>10);
678 /* need to map that range */
679 end_pfn = (aper_base>>PAGE_SHIFT) + (aper_size>>PAGE_SHIFT);
680 if (end_pfn > max_low_pfn_mapped) {
681 start_pfn = (aper_base>>PAGE_SHIFT);
682 init_memory_mapping(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
687 /* Should not happen anymore */
688 printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
689 KERN_WARNING "falling back to iommu=soft.\n");
693 extern int agp_amd64_init(void);
695 static const struct dma_mapping_ops gart_dma_ops = {
696 .mapping_error = NULL,
697 .map_single = gart_map_single,
698 .map_simple = gart_map_simple,
699 .unmap_single = gart_unmap_single,
700 .sync_single_for_cpu = NULL,
701 .sync_single_for_device = NULL,
702 .sync_single_range_for_cpu = NULL,
703 .sync_single_range_for_device = NULL,
704 .sync_sg_for_cpu = NULL,
705 .sync_sg_for_device = NULL,
706 .map_sg = gart_map_sg,
707 .unmap_sg = gart_unmap_sg,
710 void gart_iommu_shutdown(void)
715 if (no_agp && (dma_ops != &gart_dma_ops))
718 for (i = 0; i < num_k8_northbridges; i++) {
721 dev = k8_northbridges[i];
722 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &ctl);
726 pci_write_config_dword(dev, AMD64_GARTAPERTURECTL, ctl);
730 void __init gart_iommu_init(void)
732 struct agp_kern_info info;
733 unsigned long iommu_start;
734 unsigned long aper_size;
735 unsigned long scratch;
738 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
739 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
743 #ifndef CONFIG_AGP_AMD64
746 /* Makefile puts PCI initialization via subsys_initcall first. */
747 /* Add other K8 AGP bridge drivers here */
749 (agp_amd64_init() < 0) ||
750 (agp_copy_info(agp_bridge, &info) < 0);
756 /* Did we detect a different HW IOMMU? */
757 if (iommu_detected && !gart_iommu_aperture)
761 (!force_iommu && max_pfn <= MAX_DMA32_PFN) ||
762 !gart_iommu_aperture ||
763 (no_agp && init_k8_gatt(&info) < 0)) {
764 if (max_pfn > MAX_DMA32_PFN) {
765 printk(KERN_WARNING "More than 4GB of memory "
766 "but GART IOMMU not available.\n"
767 KERN_WARNING "falling back to iommu=soft.\n");
772 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
773 aper_size = info.aper_size * 1024 * 1024;
774 iommu_size = check_iommu_size(info.aper_base, aper_size);
775 iommu_pages = iommu_size >> PAGE_SHIFT;
777 iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL,
778 get_order(iommu_pages/8));
779 if (!iommu_gart_bitmap)
780 panic("Cannot allocate iommu bitmap\n");
781 memset(iommu_gart_bitmap, 0, iommu_pages/8);
783 #ifdef CONFIG_IOMMU_LEAK
785 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
786 get_order(iommu_pages*sizeof(void *)));
788 memset(iommu_leak_tab, 0, iommu_pages * 8);
791 "PCI-DMA: Cannot allocate leak trace area\n");
796 * Out of IOMMU space handling.
797 * Reserve some invalid pages at the beginning of the GART.
799 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
801 agp_memory_reserved = iommu_size;
803 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
806 iommu_start = aper_size - iommu_size;
807 iommu_bus_base = info.aper_base + iommu_start;
808 bad_dma_address = iommu_bus_base;
809 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
812 * Unmap the IOMMU part of the GART. The alias of the page is
813 * always mapped with cache enabled and there is no full cache
814 * coherency across the GART remapping. The unmapping avoids
815 * automatic prefetches from the CPU allocating cache lines in
816 * there. All CPU accesses are done via the direct mapping to
817 * the backing memory. The GART address is only used by PCI
820 set_memory_np((unsigned long)__va(iommu_bus_base),
821 iommu_size >> PAGE_SHIFT);
823 * Tricky. The GART table remaps the physical memory range,
824 * so the CPU wont notice potential aliases and if the memory
825 * is remapped to UC later on, we might surprise the PCI devices
826 * with a stray writeout of a cacheline. So play it sure and
827 * do an explicit, full-scale wbinvd() _after_ having marked all
828 * the pages as Not-Present:
833 * Try to workaround a bug (thanks to BenH):
834 * Set unmapped entries to a scratch page instead of 0.
835 * Any prefetches that hit unmapped entries won't get an bus abort
836 * then. (P2P bridge may be prefetching on DMA reads).
838 scratch = get_zeroed_page(GFP_KERNEL);
840 panic("Cannot allocate iommu scratch page");
841 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
842 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
843 iommu_gatt_base[i] = gart_unmapped_entry;
846 dma_ops = &gart_dma_ops;
849 void __init gart_parse_options(char *p)
853 #ifdef CONFIG_IOMMU_LEAK
854 if (!strncmp(p, "leak", 4)) {
858 if (isdigit(*p) && get_option(&p, &arg))
859 iommu_leak_pages = arg;
862 if (isdigit(*p) && get_option(&p, &arg))
864 if (!strncmp(p, "fullflush", 8))
866 if (!strncmp(p, "nofullflush", 11))
868 if (!strncmp(p, "noagp", 5))
870 if (!strncmp(p, "noaperture", 10))
872 /* duplicated from pci-dma.c */
873 if (!strncmp(p, "force", 5))
874 gart_iommu_aperture_allowed = 1;
875 if (!strncmp(p, "allowed", 7))
876 gart_iommu_aperture_allowed = 1;
877 if (!strncmp(p, "memaper", 7)) {
878 fallback_aper_force = 1;
882 if (get_option(&p, &arg))
883 fallback_aper_order = arg;