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>
36 #include <asm/cacheflush.h>
37 #include <asm/swiotlb.h>
41 static unsigned long iommu_bus_base; /* GART remapping area (physical) */
42 static unsigned long iommu_size; /* size of remapping area bytes */
43 static unsigned long iommu_pages; /* .. and in pages */
45 static u32 *iommu_gatt_base; /* Remapping table */
48 * If this is disabled the IOMMU will use an optimized flushing strategy
49 * of only flushing when an mapping is reused. With it true the GART is
50 * flushed for every mapping. Problem is that doing the lazy flush seems
51 * to trigger bugs with some popular PCI cards, in particular 3ware (but
52 * has been also also seen with Qlogic at least).
54 int iommu_fullflush = 1;
56 /* Allocation bitmap for the remapping area: */
57 static DEFINE_SPINLOCK(iommu_bitmap_lock);
58 /* Guarded by iommu_bitmap_lock: */
59 static unsigned long *iommu_gart_bitmap;
61 static u32 gart_unmapped_entry;
64 #define GPTE_COHERENT 2
65 #define GPTE_ENCODE(x) \
66 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
67 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
69 #define to_pages(addr, size) \
70 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
72 #define EMERGENCY_PAGES 32 /* = 128KB */
75 #define AGPEXTERN extern
80 /* backdoor interface to AGP driver */
81 AGPEXTERN int agp_memory_reserved;
82 AGPEXTERN __u32 *agp_gatt_table;
84 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
85 static int need_flush; /* global flush state. set for each gart wrap */
87 static unsigned long alloc_iommu(struct device *dev, int size)
89 unsigned long offset, flags;
90 unsigned long boundary_size;
91 unsigned long base_index;
93 base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
94 PAGE_SIZE) >> PAGE_SHIFT;
95 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
96 PAGE_SIZE) >> PAGE_SHIFT;
98 spin_lock_irqsave(&iommu_bitmap_lock, flags);
99 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
100 size, base_index, boundary_size, 0);
103 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
104 size, base_index, boundary_size, 0);
107 set_bit_string(iommu_gart_bitmap, offset, size);
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.
202 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
205 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
206 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
207 panic("PCI-DMA: Memory would be corrupted\n");
208 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
210 "PCI-DMA: Random memory would be DMAed\n");
212 #ifdef CONFIG_IOMMU_LEAK
218 need_iommu(struct device *dev, unsigned long addr, size_t size)
220 u64 mask = *dev->dma_mask;
221 int high = addr + size > mask;
231 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
233 u64 mask = *dev->dma_mask;
234 int high = addr + size > mask;
240 /* Map a single continuous physical area into the IOMMU.
241 * Caller needs to check if the iommu is needed and flush.
243 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
244 size_t size, int dir)
246 unsigned long npages = to_pages(phys_mem, size);
247 unsigned long iommu_page = alloc_iommu(dev, npages);
250 if (iommu_page == -1) {
251 if (!nonforced_iommu(dev, phys_mem, size))
253 if (panic_on_overflow)
254 panic("dma_map_area overflow %lu bytes\n", size);
255 iommu_full(dev, size, dir);
256 return bad_dma_address;
259 for (i = 0; i < npages; i++) {
260 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
261 SET_LEAK(iommu_page + i);
262 phys_mem += PAGE_SIZE;
264 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
268 gart_map_simple(struct device *dev, phys_addr_t paddr, size_t size, int dir)
270 dma_addr_t map = dma_map_area(dev, paddr, size, dir);
277 /* Map a single area into the IOMMU */
279 gart_map_single(struct device *dev, phys_addr_t paddr, size_t size, int dir)
286 if (!need_iommu(dev, paddr, size))
289 bus = gart_map_simple(dev, paddr, size, dir);
295 * Free a DMA mapping.
297 static void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
298 size_t size, int direction)
300 unsigned long iommu_page;
304 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
305 dma_addr >= iommu_bus_base + iommu_size)
308 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
309 npages = to_pages(dma_addr, size);
310 for (i = 0; i < npages; i++) {
311 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
312 CLEAR_LEAK(iommu_page + i);
314 free_iommu(iommu_page, npages);
318 * Wrapper for pci_unmap_single working with scatterlists.
321 gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
323 struct scatterlist *s;
326 for_each_sg(sg, s, nents, i) {
327 if (!s->dma_length || !s->length)
329 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
333 /* Fallback for dma_map_sg in case of overflow */
334 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
337 struct scatterlist *s;
340 #ifdef CONFIG_IOMMU_DEBUG
341 printk(KERN_DEBUG "dma_map_sg overflow\n");
344 for_each_sg(sg, s, nents, i) {
345 unsigned long addr = sg_phys(s);
347 if (nonforced_iommu(dev, addr, s->length)) {
348 addr = dma_map_area(dev, addr, s->length, dir);
349 if (addr == bad_dma_address) {
351 gart_unmap_sg(dev, sg, i, dir);
353 sg[0].dma_length = 0;
357 s->dma_address = addr;
358 s->dma_length = s->length;
365 /* Map multiple scatterlist entries continuous into the first. */
366 static int __dma_map_cont(struct device *dev, struct scatterlist *start,
367 int nelems, struct scatterlist *sout,
370 unsigned long iommu_start = alloc_iommu(dev, pages);
371 unsigned long iommu_page = iommu_start;
372 struct scatterlist *s;
375 if (iommu_start == -1)
378 for_each_sg(start, s, nelems, i) {
379 unsigned long pages, addr;
380 unsigned long phys_addr = s->dma_address;
382 BUG_ON(s != start && s->offset);
384 sout->dma_address = iommu_bus_base;
385 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
386 sout->dma_length = s->length;
388 sout->dma_length += s->length;
392 pages = to_pages(s->offset, s->length);
394 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
395 SET_LEAK(iommu_page);
400 BUG_ON(iommu_page - iommu_start != pages);
406 dma_map_cont(struct device *dev, struct scatterlist *start, int nelems,
407 struct scatterlist *sout, unsigned long pages, int need)
411 sout->dma_address = start->dma_address;
412 sout->dma_length = start->length;
415 return __dma_map_cont(dev, start, nelems, sout, pages);
419 * DMA map all entries in a scatterlist.
420 * Merge chunks that have page aligned sizes into a continuous mapping.
423 gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
425 struct scatterlist *s, *ps, *start_sg, *sgmap;
426 int need = 0, nextneed, i, out, start;
427 unsigned long pages = 0;
428 unsigned int seg_size;
429 unsigned int max_seg_size;
439 start_sg = sgmap = sg;
441 max_seg_size = dma_get_max_seg_size(dev);
442 ps = NULL; /* shut up gcc */
443 for_each_sg(sg, s, nents, i) {
444 dma_addr_t addr = sg_phys(s);
446 s->dma_address = addr;
447 BUG_ON(s->length == 0);
449 nextneed = need_iommu(dev, addr, s->length);
451 /* Handle the previous not yet processed entries */
454 * Can only merge when the last chunk ends on a
455 * page boundary and the new one doesn't have an
458 if (!iommu_merge || !nextneed || !need || s->offset ||
459 (s->length + seg_size > max_seg_size) ||
460 (ps->offset + ps->length) % PAGE_SIZE) {
461 if (dma_map_cont(dev, start_sg, i - start,
462 sgmap, pages, need) < 0)
466 sgmap = sg_next(sgmap);
473 seg_size += s->length;
475 pages += to_pages(s->offset, s->length);
478 if (dma_map_cont(dev, start_sg, i - start, sgmap, pages, need) < 0)
483 sgmap = sg_next(sgmap);
484 sgmap->dma_length = 0;
490 gart_unmap_sg(dev, sg, out, dir);
492 /* When it was forced or merged try again in a dumb way */
493 if (force_iommu || iommu_merge) {
494 out = dma_map_sg_nonforce(dev, sg, nents, dir);
498 if (panic_on_overflow)
499 panic("dma_map_sg: overflow on %lu pages\n", pages);
501 iommu_full(dev, pages << PAGE_SHIFT, dir);
502 for_each_sg(sg, s, nents, i)
503 s->dma_address = bad_dma_address;
509 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
514 iommu_size = aper_size;
519 a = aper + iommu_size;
520 iommu_size -= round_up(a, PMD_PAGE_SIZE) - a;
522 if (iommu_size < 64*1024*1024) {
524 "PCI-DMA: Warning: Small IOMMU %luMB."
525 " Consider increasing the AGP aperture in BIOS\n",
532 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
534 unsigned aper_size = 0, aper_base_32, aper_order;
537 pci_read_config_dword(dev, 0x94, &aper_base_32);
538 pci_read_config_dword(dev, 0x90, &aper_order);
539 aper_order = (aper_order >> 1) & 7;
541 aper_base = aper_base_32 & 0x7fff;
544 aper_size = (32 * 1024 * 1024) << aper_order;
545 if (aper_base + aper_size > 0x100000000UL || !aper_size)
552 static int gart_resume(struct sys_device *dev)
557 static int gart_suspend(struct sys_device *dev, pm_message_t state)
562 static struct sysdev_class gart_sysdev_class = {
564 .suspend = gart_suspend,
565 .resume = gart_resume,
569 static struct sys_device device_gart = {
571 .cls = &gart_sysdev_class,
575 * Private Northbridge GATT initialization in case we cannot use the
576 * AGP driver for some reason.
578 static __init int init_k8_gatt(struct agp_kern_info *info)
580 unsigned aper_size, gatt_size, new_aper_size;
581 unsigned aper_base, new_aper_base;
586 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
587 aper_size = aper_base = info->aper_size = 0;
589 for (i = 0; i < num_k8_northbridges; i++) {
590 dev = k8_northbridges[i];
591 new_aper_base = read_aperture(dev, &new_aper_size);
596 aper_size = new_aper_size;
597 aper_base = new_aper_base;
599 if (aper_size != new_aper_size || aper_base != new_aper_base)
604 info->aper_base = aper_base;
605 info->aper_size = aper_size >> 20;
607 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
608 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
610 panic("Cannot allocate GATT table");
611 if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
612 panic("Could not set GART PTEs to uncacheable pages");
614 memset(gatt, 0, gatt_size);
615 agp_gatt_table = gatt;
617 for (i = 0; i < num_k8_northbridges; i++) {
621 dev = k8_northbridges[i];
622 gatt_reg = __pa(gatt) >> 12;
624 pci_write_config_dword(dev, 0x98, gatt_reg);
625 pci_read_config_dword(dev, 0x90, &ctl);
628 ctl &= ~((1<<4) | (1<<5));
630 pci_write_config_dword(dev, 0x90, ctl);
633 error = sysdev_class_register(&gart_sysdev_class);
635 error = sysdev_register(&device_gart);
637 panic("Could not register gart_sysdev -- would corrupt data on next suspend");
640 printk(KERN_INFO "PCI-DMA: aperture base @ %x size %u KB\n",
641 aper_base, aper_size>>10);
645 /* Should not happen anymore */
646 printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
647 KERN_WARNING "falling back to iommu=soft.\n");
651 extern int agp_amd64_init(void);
653 static const struct dma_mapping_ops gart_dma_ops = {
654 .mapping_error = NULL,
655 .map_single = gart_map_single,
656 .map_simple = gart_map_simple,
657 .unmap_single = gart_unmap_single,
658 .sync_single_for_cpu = NULL,
659 .sync_single_for_device = NULL,
660 .sync_single_range_for_cpu = NULL,
661 .sync_single_range_for_device = NULL,
662 .sync_sg_for_cpu = NULL,
663 .sync_sg_for_device = NULL,
664 .map_sg = gart_map_sg,
665 .unmap_sg = gart_unmap_sg,
668 void gart_iommu_shutdown(void)
673 if (no_agp && (dma_ops != &gart_dma_ops))
676 for (i = 0; i < num_k8_northbridges; i++) {
679 dev = k8_northbridges[i];
680 pci_read_config_dword(dev, 0x90, &ctl);
684 pci_write_config_dword(dev, 0x90, ctl);
688 void __init gart_iommu_init(void)
690 struct agp_kern_info info;
691 unsigned long iommu_start;
692 unsigned long aper_size;
693 unsigned long scratch;
696 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
697 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
701 #ifndef CONFIG_AGP_AMD64
704 /* Makefile puts PCI initialization via subsys_initcall first. */
705 /* Add other K8 AGP bridge drivers here */
707 (agp_amd64_init() < 0) ||
708 (agp_copy_info(agp_bridge, &info) < 0);
714 /* Did we detect a different HW IOMMU? */
715 if (iommu_detected && !gart_iommu_aperture)
719 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
720 !gart_iommu_aperture ||
721 (no_agp && init_k8_gatt(&info) < 0)) {
722 if (end_pfn > MAX_DMA32_PFN) {
723 printk(KERN_WARNING "More than 4GB of memory "
724 "but GART IOMMU not available.\n"
725 KERN_WARNING "falling back to iommu=soft.\n");
730 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
731 aper_size = info.aper_size * 1024 * 1024;
732 iommu_size = check_iommu_size(info.aper_base, aper_size);
733 iommu_pages = iommu_size >> PAGE_SHIFT;
735 iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL,
736 get_order(iommu_pages/8));
737 if (!iommu_gart_bitmap)
738 panic("Cannot allocate iommu bitmap\n");
739 memset(iommu_gart_bitmap, 0, iommu_pages/8);
741 #ifdef CONFIG_IOMMU_LEAK
743 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
744 get_order(iommu_pages*sizeof(void *)));
746 memset(iommu_leak_tab, 0, iommu_pages * 8);
749 "PCI-DMA: Cannot allocate leak trace area\n");
754 * Out of IOMMU space handling.
755 * Reserve some invalid pages at the beginning of the GART.
757 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
759 agp_memory_reserved = iommu_size;
761 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
764 iommu_start = aper_size - iommu_size;
765 iommu_bus_base = info.aper_base + iommu_start;
766 bad_dma_address = iommu_bus_base;
767 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
770 * Unmap the IOMMU part of the GART. The alias of the page is
771 * always mapped with cache enabled and there is no full cache
772 * coherency across the GART remapping. The unmapping avoids
773 * automatic prefetches from the CPU allocating cache lines in
774 * there. All CPU accesses are done via the direct mapping to
775 * the backing memory. The GART address is only used by PCI
778 set_memory_np((unsigned long)__va(iommu_bus_base),
779 iommu_size >> PAGE_SHIFT);
781 * Tricky. The GART table remaps the physical memory range,
782 * so the CPU wont notice potential aliases and if the memory
783 * is remapped to UC later on, we might surprise the PCI devices
784 * with a stray writeout of a cacheline. So play it sure and
785 * do an explicit, full-scale wbinvd() _after_ having marked all
786 * the pages as Not-Present:
791 * Try to workaround a bug (thanks to BenH)
792 * Set unmapped entries to a scratch page instead of 0.
793 * Any prefetches that hit unmapped entries won't get an bus abort
796 scratch = get_zeroed_page(GFP_KERNEL);
798 panic("Cannot allocate iommu scratch page");
799 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
800 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
801 iommu_gatt_base[i] = gart_unmapped_entry;
804 dma_ops = &gart_dma_ops;
807 void __init gart_parse_options(char *p)
811 #ifdef CONFIG_IOMMU_LEAK
812 if (!strncmp(p, "leak", 4)) {
816 if (isdigit(*p) && get_option(&p, &arg))
817 iommu_leak_pages = arg;
820 if (isdigit(*p) && get_option(&p, &arg))
822 if (!strncmp(p, "fullflush", 8))
824 if (!strncmp(p, "nofullflush", 11))
826 if (!strncmp(p, "noagp", 5))
828 if (!strncmp(p, "noaperture", 10))
830 /* duplicated from pci-dma.c */
831 if (!strncmp(p, "force", 5))
832 gart_iommu_aperture_allowed = 1;
833 if (!strncmp(p, "allowed", 7))
834 gart_iommu_aperture_allowed = 1;
835 if (!strncmp(p, "memaper", 7)) {
836 fallback_aper_force = 1;
840 if (get_option(&p, &arg))
841 fallback_aper_order = arg;