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 <asm/atomic.h>
32 #include <asm/pgtable.h>
33 #include <asm/proto.h>
35 #include <asm/cacheflush.h>
36 #include <asm/swiotlb.h>
40 static unsigned long iommu_bus_base; /* GART remapping area (physical) */
41 static unsigned long iommu_size; /* size of remapping area bytes */
42 static unsigned long iommu_pages; /* .. and in pages */
44 static u32 *iommu_gatt_base; /* Remapping table */
47 * If this is disabled the IOMMU will use an optimized flushing strategy
48 * of only flushing when an mapping is reused. With it true the GART is
49 * flushed for every mapping. Problem is that doing the lazy flush seems
50 * to trigger bugs with some popular PCI cards, in particular 3ware (but
51 * has been also also seen with Qlogic at least).
53 int iommu_fullflush = 1;
55 /* Allocation bitmap for the remapping area: */
56 static DEFINE_SPINLOCK(iommu_bitmap_lock);
57 /* Guarded by iommu_bitmap_lock: */
58 static unsigned long *iommu_gart_bitmap;
60 static u32 gart_unmapped_entry;
63 #define GPTE_COHERENT 2
64 #define GPTE_ENCODE(x) \
65 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
66 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
68 #define to_pages(addr, size) \
69 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
71 #define EMERGENCY_PAGES 32 /* = 128KB */
74 #define AGPEXTERN extern
79 /* backdoor interface to AGP driver */
80 AGPEXTERN int agp_memory_reserved;
81 AGPEXTERN __u32 *agp_gatt_table;
83 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
84 static int need_flush; /* global flush state. set for each gart wrap */
86 static unsigned long alloc_iommu(struct device *dev, int size)
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(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, 0);
102 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
103 size, base_index, boundary_size, 0);
106 set_bit_string(iommu_gart_bitmap, offset, 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 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
130 * Use global flush state to avoid races with multiple flushers.
132 static void flush_gart(void)
136 spin_lock_irqsave(&iommu_bitmap_lock, flags);
141 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
144 #ifdef CONFIG_IOMMU_LEAK
146 #define SET_LEAK(x) \
148 if (iommu_leak_tab) \
149 iommu_leak_tab[x] = __builtin_return_address(0);\
152 #define CLEAR_LEAK(x) \
154 if (iommu_leak_tab) \
155 iommu_leak_tab[x] = NULL; \
158 /* Debugging aid for drivers that don't free their IOMMU tables */
159 static void **iommu_leak_tab;
160 static int leak_trace;
161 static int iommu_leak_pages = 20;
163 static void dump_leak(void)
168 if (dump || !iommu_leak_tab)
171 show_stack(NULL, NULL);
173 /* Very crude. dump some from the end of the table too */
174 printk(KERN_DEBUG "Dumping %d pages from end of IOMMU:\n",
176 for (i = 0; i < iommu_leak_pages; i += 2) {
177 printk(KERN_DEBUG "%lu: ", iommu_pages-i);
178 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i], 0);
179 printk(KERN_CONT "%c", (i+1)%2 == 0 ? '\n' : ' ');
181 printk(KERN_DEBUG "\n");
185 # define CLEAR_LEAK(x)
188 static void iommu_full(struct device *dev, size_t size, int dir)
191 * Ran out of IOMMU space for this operation. This is very bad.
192 * Unfortunately the drivers cannot handle this operation properly.
193 * Return some non mapped prereserved space in the aperture and
194 * let the Northbridge deal with it. This will result in garbage
195 * in the IO operation. When the size exceeds the prereserved space
196 * memory corruption will occur or random memory will be DMAed
197 * out. Hopefully no network devices use single mappings that big.
201 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
204 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
205 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
206 panic("PCI-DMA: Memory would be corrupted\n");
207 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
209 "PCI-DMA: Random memory would be DMAed\n");
211 #ifdef CONFIG_IOMMU_LEAK
217 need_iommu(struct device *dev, unsigned long addr, size_t size)
219 u64 mask = *dev->dma_mask;
220 int high = addr + size > mask;
230 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
232 u64 mask = *dev->dma_mask;
233 int high = addr + size > mask;
239 /* Map a single continuous physical area into the IOMMU.
240 * Caller needs to check if the iommu is needed and flush.
242 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
243 size_t size, int dir)
245 unsigned long npages = to_pages(phys_mem, size);
246 unsigned long iommu_page = alloc_iommu(dev, npages);
249 if (iommu_page == -1) {
250 if (!nonforced_iommu(dev, phys_mem, size))
252 if (panic_on_overflow)
253 panic("dma_map_area overflow %lu bytes\n", size);
254 iommu_full(dev, size, dir);
255 return bad_dma_address;
258 for (i = 0; i < npages; i++) {
259 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
260 SET_LEAK(iommu_page + i);
261 phys_mem += PAGE_SIZE;
263 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
267 gart_map_simple(struct device *dev, char *buf, size_t size, int dir)
269 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
276 /* Map a single area into the IOMMU */
278 gart_map_single(struct device *dev, void *addr, size_t size, int dir)
280 unsigned long phys_mem, bus;
285 phys_mem = virt_to_phys(addr);
286 if (!need_iommu(dev, phys_mem, size))
289 bus = gart_map_simple(dev, addr, 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)
553 * Private Northbridge GATT initialization in case we cannot use the
554 * AGP driver for some reason.
556 static __init int init_k8_gatt(struct agp_kern_info *info)
558 unsigned aper_size, gatt_size, new_aper_size;
559 unsigned aper_base, new_aper_base;
564 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
565 aper_size = aper_base = info->aper_size = 0;
567 for (i = 0; i < num_k8_northbridges; i++) {
568 dev = k8_northbridges[i];
569 new_aper_base = read_aperture(dev, &new_aper_size);
574 aper_size = new_aper_size;
575 aper_base = new_aper_base;
577 if (aper_size != new_aper_size || aper_base != new_aper_base)
582 info->aper_base = aper_base;
583 info->aper_size = aper_size >> 20;
585 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
586 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
588 panic("Cannot allocate GATT table");
589 if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
590 panic("Could not set GART PTEs to uncacheable pages");
592 memset(gatt, 0, gatt_size);
593 agp_gatt_table = gatt;
595 for (i = 0; i < num_k8_northbridges; i++) {
599 dev = k8_northbridges[i];
600 gatt_reg = __pa(gatt) >> 12;
602 pci_write_config_dword(dev, 0x98, gatt_reg);
603 pci_read_config_dword(dev, 0x90, &ctl);
606 ctl &= ~((1<<4) | (1<<5));
608 pci_write_config_dword(dev, 0x90, ctl);
612 printk(KERN_INFO "PCI-DMA: aperture base @ %x size %u KB\n",
613 aper_base, aper_size>>10);
617 /* Should not happen anymore */
618 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
619 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
623 extern int agp_amd64_init(void);
625 static const struct dma_mapping_ops gart_dma_ops = {
626 .mapping_error = NULL,
627 .map_single = gart_map_single,
628 .map_simple = gart_map_simple,
629 .unmap_single = gart_unmap_single,
630 .sync_single_for_cpu = NULL,
631 .sync_single_for_device = NULL,
632 .sync_single_range_for_cpu = NULL,
633 .sync_single_range_for_device = NULL,
634 .sync_sg_for_cpu = NULL,
635 .sync_sg_for_device = NULL,
636 .map_sg = gart_map_sg,
637 .unmap_sg = gart_unmap_sg,
640 void gart_iommu_shutdown(void)
645 if (no_agp && (dma_ops != &gart_dma_ops))
648 for (i = 0; i < num_k8_northbridges; i++) {
651 dev = k8_northbridges[i];
652 pci_read_config_dword(dev, 0x90, &ctl);
656 pci_write_config_dword(dev, 0x90, ctl);
660 void __init gart_iommu_init(void)
662 struct agp_kern_info info;
663 unsigned long iommu_start;
664 unsigned long aper_size;
665 unsigned long scratch;
668 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
669 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
673 #ifndef CONFIG_AGP_AMD64
676 /* Makefile puts PCI initialization via subsys_initcall first. */
677 /* Add other K8 AGP bridge drivers here */
679 (agp_amd64_init() < 0) ||
680 (agp_copy_info(agp_bridge, &info) < 0);
686 /* Did we detect a different HW IOMMU? */
687 if (iommu_detected && !gart_iommu_aperture)
691 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
692 !gart_iommu_aperture ||
693 (no_agp && init_k8_gatt(&info) < 0)) {
694 if (end_pfn > MAX_DMA32_PFN) {
695 printk(KERN_ERR "WARNING more than 4GB of memory "
696 "but GART IOMMU not available.\n"
697 KERN_ERR "WARNING 32bit PCI may malfunction.\n");
702 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
703 aper_size = info.aper_size * 1024 * 1024;
704 iommu_size = check_iommu_size(info.aper_base, aper_size);
705 iommu_pages = iommu_size >> PAGE_SHIFT;
707 iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL,
708 get_order(iommu_pages/8));
709 if (!iommu_gart_bitmap)
710 panic("Cannot allocate iommu bitmap\n");
711 memset(iommu_gart_bitmap, 0, iommu_pages/8);
713 #ifdef CONFIG_IOMMU_LEAK
715 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
716 get_order(iommu_pages*sizeof(void *)));
718 memset(iommu_leak_tab, 0, iommu_pages * 8);
721 "PCI-DMA: Cannot allocate leak trace area\n");
726 * Out of IOMMU space handling.
727 * Reserve some invalid pages at the beginning of the GART.
729 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
731 agp_memory_reserved = iommu_size;
733 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
736 iommu_start = aper_size - iommu_size;
737 iommu_bus_base = info.aper_base + iommu_start;
738 bad_dma_address = iommu_bus_base;
739 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
742 * Unmap the IOMMU part of the GART. The alias of the page is
743 * always mapped with cache enabled and there is no full cache
744 * coherency across the GART remapping. The unmapping avoids
745 * automatic prefetches from the CPU allocating cache lines in
746 * there. All CPU accesses are done via the direct mapping to
747 * the backing memory. The GART address is only used by PCI
750 set_memory_np((unsigned long)__va(iommu_bus_base),
751 iommu_size >> PAGE_SHIFT);
754 * Try to workaround a bug (thanks to BenH)
755 * Set unmapped entries to a scratch page instead of 0.
756 * Any prefetches that hit unmapped entries won't get an bus abort
759 scratch = get_zeroed_page(GFP_KERNEL);
761 panic("Cannot allocate iommu scratch page");
762 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
763 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
764 iommu_gatt_base[i] = gart_unmapped_entry;
767 dma_ops = &gart_dma_ops;
770 void __init gart_parse_options(char *p)
774 #ifdef CONFIG_IOMMU_LEAK
775 if (!strncmp(p, "leak", 4)) {
779 if (isdigit(*p) && get_option(&p, &arg))
780 iommu_leak_pages = arg;
783 if (isdigit(*p) && get_option(&p, &arg))
785 if (!strncmp(p, "fullflush", 8))
787 if (!strncmp(p, "nofullflush", 11))
789 if (!strncmp(p, "noagp", 5))
791 if (!strncmp(p, "noaperture", 10))
793 /* duplicated from pci-dma.c */
794 if (!strncmp(p, "force", 5))
795 gart_iommu_aperture_allowed = 1;
796 if (!strncmp(p, "allowed", 7))
797 gart_iommu_aperture_allowed = 1;
798 if (!strncmp(p, "memaper", 7)) {
799 fallback_aper_force = 1;
803 if (get_option(&p, &arg))
804 fallback_aper_order = arg;