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/PCI/PCI-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>
31 #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 static 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 EMERGENCY_PAGES 32 /* = 128KB */
73 #define AGPEXTERN extern
78 /* backdoor interface to AGP driver */
79 AGPEXTERN int agp_memory_reserved;
80 AGPEXTERN __u32 *agp_gatt_table;
82 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
83 static bool need_flush; /* global flush state. set for each gart wrap */
85 static unsigned long alloc_iommu(struct device *dev, int size,
86 unsigned long align_mask)
88 unsigned long offset, flags;
89 unsigned long boundary_size;
90 unsigned long base_index;
92 base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
93 PAGE_SIZE) >> PAGE_SHIFT;
94 boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1,
95 PAGE_SIZE) >> PAGE_SHIFT;
97 spin_lock_irqsave(&iommu_bitmap_lock, flags);
98 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
99 size, base_index, boundary_size, align_mask);
102 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
103 size, base_index, boundary_size,
107 next_bit = offset+size;
108 if (next_bit >= iommu_pages) {
115 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
120 static void free_iommu(unsigned long offset, int size)
124 spin_lock_irqsave(&iommu_bitmap_lock, flags);
125 iommu_area_free(iommu_gart_bitmap, offset, size);
126 if (offset >= next_bit)
127 next_bit = offset + size;
128 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
132 * Use global flush state to avoid races with multiple flushers.
134 static void flush_gart(void)
138 spin_lock_irqsave(&iommu_bitmap_lock, flags);
143 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
146 #ifdef CONFIG_IOMMU_LEAK
147 /* Debugging aid for drivers that don't free their IOMMU tables */
148 static int leak_trace;
149 static int iommu_leak_pages = 20;
151 static void dump_leak(void)
159 show_stack(NULL, NULL);
160 debug_dma_dump_mappings(NULL);
164 static void iommu_full(struct device *dev, size_t size, int dir)
167 * Ran out of IOMMU space for this operation. This is very bad.
168 * Unfortunately the drivers cannot handle this operation properly.
169 * Return some non mapped prereserved space in the aperture and
170 * let the Northbridge deal with it. This will result in garbage
171 * in the IO operation. When the size exceeds the prereserved space
172 * memory corruption will occur or random memory will be DMAed
173 * out. Hopefully no network devices use single mappings that big.
176 dev_err(dev, "PCI-DMA: Out of IOMMU space for %lu bytes\n", size);
178 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
179 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
180 panic("PCI-DMA: Memory would be corrupted\n");
181 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
183 "PCI-DMA: Random memory would be DMAed\n");
185 #ifdef CONFIG_IOMMU_LEAK
191 need_iommu(struct device *dev, unsigned long addr, size_t size)
193 return force_iommu ||
194 !is_buffer_dma_capable(*dev->dma_mask, addr, size);
198 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
200 return !is_buffer_dma_capable(*dev->dma_mask, addr, size);
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, unsigned long align_mask)
209 unsigned long npages = iommu_num_pages(phys_mem, size, PAGE_SIZE);
210 unsigned long iommu_page = alloc_iommu(dev, npages, align_mask);
213 if (iommu_page == -1) {
214 if (!nonforced_iommu(dev, phys_mem, size))
216 if (panic_on_overflow)
217 panic("dma_map_area overflow %lu bytes\n", size);
218 iommu_full(dev, size, dir);
219 return bad_dma_address;
222 for (i = 0; i < npages; i++) {
223 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
224 phys_mem += PAGE_SIZE;
226 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
229 /* Map a single area into the IOMMU */
230 static dma_addr_t gart_map_page(struct device *dev, struct page *page,
231 unsigned long offset, size_t size,
232 enum dma_data_direction dir,
233 struct dma_attrs *attrs)
236 phys_addr_t paddr = page_to_phys(page) + offset;
239 dev = &x86_dma_fallback_dev;
241 if (!need_iommu(dev, paddr, size))
244 bus = dma_map_area(dev, paddr, size, dir, 0);
251 * Free a DMA mapping.
253 static void gart_unmap_page(struct device *dev, dma_addr_t dma_addr,
254 size_t size, enum dma_data_direction dir,
255 struct dma_attrs *attrs)
257 unsigned long iommu_page;
261 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
262 dma_addr >= iommu_bus_base + iommu_size)
265 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
266 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
267 for (i = 0; i < npages; i++) {
268 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
270 free_iommu(iommu_page, npages);
274 * Wrapper for pci_unmap_single working with scatterlists.
276 static void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
277 enum dma_data_direction dir, struct dma_attrs *attrs)
279 struct scatterlist *s;
282 for_each_sg(sg, s, nents, i) {
283 if (!s->dma_length || !s->length)
285 gart_unmap_page(dev, s->dma_address, s->dma_length, dir, NULL);
289 /* Fallback for dma_map_sg in case of overflow */
290 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
293 struct scatterlist *s;
296 #ifdef CONFIG_IOMMU_DEBUG
297 printk(KERN_DEBUG "dma_map_sg overflow\n");
300 for_each_sg(sg, s, nents, i) {
301 unsigned long addr = sg_phys(s);
303 if (nonforced_iommu(dev, addr, s->length)) {
304 addr = dma_map_area(dev, addr, s->length, dir, 0);
305 if (addr == bad_dma_address) {
307 gart_unmap_sg(dev, sg, i, dir, NULL);
309 sg[0].dma_length = 0;
313 s->dma_address = addr;
314 s->dma_length = s->length;
321 /* Map multiple scatterlist entries continuous into the first. */
322 static int __dma_map_cont(struct device *dev, struct scatterlist *start,
323 int nelems, struct scatterlist *sout,
326 unsigned long iommu_start = alloc_iommu(dev, pages, 0);
327 unsigned long iommu_page = iommu_start;
328 struct scatterlist *s;
331 if (iommu_start == -1)
334 for_each_sg(start, s, nelems, i) {
335 unsigned long pages, addr;
336 unsigned long phys_addr = s->dma_address;
338 BUG_ON(s != start && s->offset);
340 sout->dma_address = iommu_bus_base;
341 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
342 sout->dma_length = s->length;
344 sout->dma_length += s->length;
348 pages = iommu_num_pages(s->offset, s->length, PAGE_SIZE);
350 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
355 BUG_ON(iommu_page - iommu_start != pages);
361 dma_map_cont(struct device *dev, struct scatterlist *start, int nelems,
362 struct scatterlist *sout, unsigned long pages, int need)
366 sout->dma_address = start->dma_address;
367 sout->dma_length = start->length;
370 return __dma_map_cont(dev, start, nelems, sout, pages);
374 * DMA map all entries in a scatterlist.
375 * Merge chunks that have page aligned sizes into a continuous mapping.
377 static int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents,
378 enum dma_data_direction dir, struct dma_attrs *attrs)
380 struct scatterlist *s, *ps, *start_sg, *sgmap;
381 int need = 0, nextneed, i, out, start;
382 unsigned long pages = 0;
383 unsigned int seg_size;
384 unsigned int max_seg_size;
390 dev = &x86_dma_fallback_dev;
394 start_sg = sgmap = sg;
396 max_seg_size = dma_get_max_seg_size(dev);
397 ps = NULL; /* shut up gcc */
398 for_each_sg(sg, s, nents, i) {
399 dma_addr_t addr = sg_phys(s);
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 */
409 * Can only merge when the last chunk ends on a
410 * page boundary and the new one doesn't have an
413 if (!iommu_merge || !nextneed || !need || s->offset ||
414 (s->length + seg_size > max_seg_size) ||
415 (ps->offset + ps->length) % PAGE_SIZE) {
416 if (dma_map_cont(dev, start_sg, i - start,
417 sgmap, pages, need) < 0)
421 sgmap = sg_next(sgmap);
428 seg_size += s->length;
430 pages += iommu_num_pages(s->offset, s->length, PAGE_SIZE);
433 if (dma_map_cont(dev, start_sg, i - start, sgmap, pages, need) < 0)
438 sgmap = sg_next(sgmap);
439 sgmap->dma_length = 0;
445 gart_unmap_sg(dev, sg, out, dir, NULL);
447 /* When it was forced or merged try again in a dumb way */
448 if (force_iommu || iommu_merge) {
449 out = dma_map_sg_nonforce(dev, sg, nents, dir);
453 if (panic_on_overflow)
454 panic("dma_map_sg: overflow on %lu pages\n", pages);
456 iommu_full(dev, pages << PAGE_SHIFT, dir);
457 for_each_sg(sg, s, nents, i)
458 s->dma_address = bad_dma_address;
462 /* allocate and map a coherent mapping */
464 gart_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_addr,
468 unsigned long align_mask;
471 if (force_iommu && !(flag & GFP_DMA)) {
472 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
473 page = alloc_pages(flag | __GFP_ZERO, get_order(size));
477 align_mask = (1UL << get_order(size)) - 1;
478 paddr = dma_map_area(dev, page_to_phys(page), size,
479 DMA_BIDIRECTIONAL, align_mask);
482 if (paddr != bad_dma_address) {
484 return page_address(page);
486 __free_pages(page, get_order(size));
488 return dma_generic_alloc_coherent(dev, size, dma_addr, flag);
493 /* free a coherent mapping */
495 gart_free_coherent(struct device *dev, size_t size, void *vaddr,
498 gart_unmap_page(dev, dma_addr, size, DMA_BIDIRECTIONAL, NULL);
499 free_pages((unsigned long)vaddr, get_order(size));
504 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
509 iommu_size = aper_size;
514 a = aper + iommu_size;
515 iommu_size -= round_up(a, PMD_PAGE_SIZE) - a;
517 if (iommu_size < 64*1024*1024) {
519 "PCI-DMA: Warning: Small IOMMU %luMB."
520 " Consider increasing the AGP aperture in BIOS\n",
527 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
529 unsigned aper_size = 0, aper_base_32, aper_order;
532 pci_read_config_dword(dev, AMD64_GARTAPERTUREBASE, &aper_base_32);
533 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &aper_order);
534 aper_order = (aper_order >> 1) & 7;
536 aper_base = aper_base_32 & 0x7fff;
539 aper_size = (32 * 1024 * 1024) << aper_order;
540 if (aper_base + aper_size > 0x100000000UL || !aper_size)
547 static void enable_gart_translations(void)
551 for (i = 0; i < num_k8_northbridges; i++) {
552 struct pci_dev *dev = k8_northbridges[i];
554 enable_gart_translation(dev, __pa(agp_gatt_table));
559 * If fix_up_north_bridges is set, the north bridges have to be fixed up on
560 * resume in the same way as they are handled in gart_iommu_hole_init().
562 static bool fix_up_north_bridges;
563 static u32 aperture_order;
564 static u32 aperture_alloc;
566 void set_up_gart_resume(u32 aper_order, u32 aper_alloc)
568 fix_up_north_bridges = true;
569 aperture_order = aper_order;
570 aperture_alloc = aper_alloc;
573 static int gart_resume(struct sys_device *dev)
575 printk(KERN_INFO "PCI-DMA: Resuming GART IOMMU\n");
577 if (fix_up_north_bridges) {
580 printk(KERN_INFO "PCI-DMA: Restoring GART aperture settings\n");
582 for (i = 0; i < num_k8_northbridges; i++) {
583 struct pci_dev *dev = k8_northbridges[i];
586 * Don't enable translations just yet. That is the next
587 * step. Restore the pre-suspend aperture settings.
589 pci_write_config_dword(dev, AMD64_GARTAPERTURECTL,
590 aperture_order << 1);
591 pci_write_config_dword(dev, AMD64_GARTAPERTUREBASE,
592 aperture_alloc >> 25);
596 enable_gart_translations();
601 static int gart_suspend(struct sys_device *dev, pm_message_t state)
606 static struct sysdev_class gart_sysdev_class = {
608 .suspend = gart_suspend,
609 .resume = gart_resume,
613 static struct sys_device device_gart = {
615 .cls = &gart_sysdev_class,
619 * Private Northbridge GATT initialization in case we cannot use the
620 * AGP driver for some reason.
622 static __init int init_k8_gatt(struct agp_kern_info *info)
624 unsigned aper_size, gatt_size, new_aper_size;
625 unsigned aper_base, new_aper_base;
630 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
631 aper_size = aper_base = info->aper_size = 0;
633 for (i = 0; i < num_k8_northbridges; i++) {
634 dev = k8_northbridges[i];
635 new_aper_base = read_aperture(dev, &new_aper_size);
640 aper_size = new_aper_size;
641 aper_base = new_aper_base;
643 if (aper_size != new_aper_size || aper_base != new_aper_base)
648 info->aper_base = aper_base;
649 info->aper_size = aper_size >> 20;
651 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
652 gatt = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
653 get_order(gatt_size));
655 panic("Cannot allocate GATT table");
656 if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
657 panic("Could not set GART PTEs to uncacheable pages");
659 agp_gatt_table = gatt;
661 error = sysdev_class_register(&gart_sysdev_class);
663 error = sysdev_register(&device_gart);
665 panic("Could not register gart_sysdev -- "
666 "would corrupt data on next suspend");
670 printk(KERN_INFO "PCI-DMA: aperture base @ %x size %u KB\n",
671 aper_base, aper_size>>10);
676 /* Should not happen anymore */
677 printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
678 "falling back to iommu=soft.\n");
682 static struct dma_map_ops gart_dma_ops = {
683 .map_sg = gart_map_sg,
684 .unmap_sg = gart_unmap_sg,
685 .map_page = gart_map_page,
686 .unmap_page = gart_unmap_page,
687 .alloc_coherent = gart_alloc_coherent,
688 .free_coherent = gart_free_coherent,
691 void gart_iommu_shutdown(void)
696 if (no_agp && (dma_ops != &gart_dma_ops))
699 for (i = 0; i < num_k8_northbridges; i++) {
702 dev = k8_northbridges[i];
703 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &ctl);
707 pci_write_config_dword(dev, AMD64_GARTAPERTURECTL, ctl);
711 void __init gart_iommu_init(void)
713 struct agp_kern_info info;
714 unsigned long iommu_start;
715 unsigned long aper_base, aper_size;
716 unsigned long start_pfn, end_pfn;
717 unsigned long scratch;
720 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0)
723 #ifndef CONFIG_AGP_AMD64
726 /* Makefile puts PCI initialization via subsys_initcall first. */
727 /* Add other K8 AGP bridge drivers here */
729 (agp_amd64_init() < 0) ||
730 (agp_copy_info(agp_bridge, &info) < 0);
736 /* Did we detect a different HW IOMMU? */
737 if (iommu_detected && !gart_iommu_aperture)
741 (!force_iommu && max_pfn <= MAX_DMA32_PFN) ||
742 !gart_iommu_aperture ||
743 (no_agp && init_k8_gatt(&info) < 0)) {
744 if (max_pfn > MAX_DMA32_PFN) {
745 printk(KERN_WARNING "More than 4GB of memory "
746 "but GART IOMMU not available.\n");
747 printk(KERN_WARNING "falling back to iommu=soft.\n");
752 /* need to map that range */
753 aper_size = info.aper_size << 20;
754 aper_base = info.aper_base;
755 end_pfn = (aper_base>>PAGE_SHIFT) + (aper_size>>PAGE_SHIFT);
756 if (end_pfn > max_low_pfn_mapped) {
757 start_pfn = (aper_base>>PAGE_SHIFT);
758 init_memory_mapping(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
761 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
762 iommu_size = check_iommu_size(info.aper_base, aper_size);
763 iommu_pages = iommu_size >> PAGE_SHIFT;
765 iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO,
766 get_order(iommu_pages/8));
767 if (!iommu_gart_bitmap)
768 panic("Cannot allocate iommu bitmap\n");
770 #ifdef CONFIG_IOMMU_LEAK
774 ret = dma_debug_resize_entries(iommu_pages);
777 "PCI-DMA: Cannot trace all the entries\n");
782 * Out of IOMMU space handling.
783 * Reserve some invalid pages at the beginning of the GART.
785 iommu_area_reserve(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
787 agp_memory_reserved = iommu_size;
789 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
792 iommu_start = aper_size - iommu_size;
793 iommu_bus_base = info.aper_base + iommu_start;
794 bad_dma_address = iommu_bus_base;
795 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
798 * Unmap the IOMMU part of the GART. The alias of the page is
799 * always mapped with cache enabled and there is no full cache
800 * coherency across the GART remapping. The unmapping avoids
801 * automatic prefetches from the CPU allocating cache lines in
802 * there. All CPU accesses are done via the direct mapping to
803 * the backing memory. The GART address is only used by PCI
806 set_memory_np((unsigned long)__va(iommu_bus_base),
807 iommu_size >> PAGE_SHIFT);
809 * Tricky. The GART table remaps the physical memory range,
810 * so the CPU wont notice potential aliases and if the memory
811 * is remapped to UC later on, we might surprise the PCI devices
812 * with a stray writeout of a cacheline. So play it sure and
813 * do an explicit, full-scale wbinvd() _after_ having marked all
814 * the pages as Not-Present:
819 * Now all caches are flushed and we can safely enable
820 * GART hardware. Doing it early leaves the possibility
821 * of stale cache entries that can lead to GART PTE
824 enable_gart_translations();
827 * Try to workaround a bug (thanks to BenH):
828 * Set unmapped entries to a scratch page instead of 0.
829 * Any prefetches that hit unmapped entries won't get an bus abort
830 * then. (P2P bridge may be prefetching on DMA reads).
832 scratch = get_zeroed_page(GFP_KERNEL);
834 panic("Cannot allocate iommu scratch page");
835 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
836 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
837 iommu_gatt_base[i] = gart_unmapped_entry;
840 dma_ops = &gart_dma_ops;
843 void __init gart_parse_options(char *p)
847 #ifdef CONFIG_IOMMU_LEAK
848 if (!strncmp(p, "leak", 4)) {
853 if (isdigit(*p) && get_option(&p, &arg))
854 iommu_leak_pages = arg;
857 if (isdigit(*p) && get_option(&p, &arg))
859 if (!strncmp(p, "fullflush", 8))
861 if (!strncmp(p, "nofullflush", 11))
863 if (!strncmp(p, "noagp", 5))
865 if (!strncmp(p, "noaperture", 10))
867 /* duplicated from pci-dma.c */
868 if (!strncmp(p, "force", 5))
869 gart_iommu_aperture_allowed = 1;
870 if (!strncmp(p, "allowed", 7))
871 gart_iommu_aperture_allowed = 1;
872 if (!strncmp(p, "memaper", 7)) {
873 fallback_aper_force = 1;
877 if (get_option(&p, &arg))
878 fallback_aper_order = arg;