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/types.h>
14 #include <linux/ctype.h>
15 #include <linux/agp_backend.h>
16 #include <linux/init.h>
18 #include <linux/string.h>
19 #include <linux/spinlock.h>
20 #include <linux/pci.h>
21 #include <linux/module.h>
22 #include <linux/topology.h>
23 #include <linux/interrupt.h>
24 #include <linux/bitops.h>
25 #include <linux/kdebug.h>
26 #include <asm/atomic.h>
29 #include <asm/pgtable.h>
30 #include <asm/proto.h>
31 #include <asm/iommu.h>
32 #include <asm/cacheflush.h>
33 #include <asm/swiotlb.h>
37 unsigned long iommu_bus_base; /* GART remapping area (physical) */
38 static unsigned long iommu_size; /* size of remapping area bytes */
39 static unsigned long iommu_pages; /* .. and in pages */
41 u32 *iommu_gatt_base; /* Remapping table */
43 /* If this is disabled the IOMMU will use an optimized flushing strategy
44 of only flushing when an mapping is reused. With it true the GART is flushed
45 for every mapping. Problem is that doing the lazy flush seems to trigger
46 bugs with some popular PCI cards, in particular 3ware (but has been also
47 also seen with Qlogic at least). */
48 int iommu_fullflush = 1;
50 /* Allocation bitmap for the remapping area */
51 static DEFINE_SPINLOCK(iommu_bitmap_lock);
52 static unsigned long *iommu_gart_bitmap; /* guarded by iommu_bitmap_lock */
54 static u32 gart_unmapped_entry;
57 #define GPTE_COHERENT 2
58 #define GPTE_ENCODE(x) \
59 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
60 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
62 #define to_pages(addr,size) \
63 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
65 #define EMERGENCY_PAGES 32 /* = 128KB */
68 #define AGPEXTERN extern
73 /* backdoor interface to AGP driver */
74 AGPEXTERN int agp_memory_reserved;
75 AGPEXTERN __u32 *agp_gatt_table;
77 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
78 static int need_flush; /* global flush state. set for each gart wrap */
80 static unsigned long alloc_iommu(int size)
82 unsigned long offset, flags;
84 spin_lock_irqsave(&iommu_bitmap_lock, flags);
85 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
88 offset = find_next_zero_string(iommu_gart_bitmap,0,iommu_pages,size);
91 set_bit_string(iommu_gart_bitmap, offset, size);
92 next_bit = offset+size;
93 if (next_bit >= iommu_pages) {
100 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
104 static void free_iommu(unsigned long offset, int size)
107 spin_lock_irqsave(&iommu_bitmap_lock, flags);
108 __clear_bit_string(iommu_gart_bitmap, offset, size);
109 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
113 * Use global flush state to avoid races with multiple flushers.
115 static void flush_gart(void)
118 spin_lock_irqsave(&iommu_bitmap_lock, flags);
123 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
126 #ifdef CONFIG_IOMMU_LEAK
128 #define SET_LEAK(x) if (iommu_leak_tab) \
129 iommu_leak_tab[x] = __builtin_return_address(0);
130 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
131 iommu_leak_tab[x] = NULL;
133 /* Debugging aid for drivers that don't free their IOMMU tables */
134 static void **iommu_leak_tab;
135 static int leak_trace;
136 int iommu_leak_pages = 20;
141 if (dump || !iommu_leak_tab) return;
143 show_stack(NULL,NULL);
144 /* Very crude. dump some from the end of the table too */
145 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
146 for (i = 0; i < iommu_leak_pages; i+=2) {
147 printk("%lu: ", iommu_pages-i);
148 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
149 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
155 #define CLEAR_LEAK(x)
158 static void iommu_full(struct device *dev, size_t size, int dir)
161 * Ran out of IOMMU space for this operation. This is very bad.
162 * Unfortunately the drivers cannot handle this operation properly.
163 * Return some non mapped prereserved space in the aperture and
164 * let the Northbridge deal with it. This will result in garbage
165 * in the IO operation. When the size exceeds the prereserved space
166 * memory corruption will occur or random memory will be DMAed
167 * out. Hopefully no network devices use single mappings that big.
171 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
174 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
175 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
176 panic("PCI-DMA: Memory would be corrupted\n");
177 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
178 panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
181 #ifdef CONFIG_IOMMU_LEAK
186 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
188 u64 mask = *dev->dma_mask;
189 int high = addr + size > mask;
196 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
198 u64 mask = *dev->dma_mask;
199 int high = addr + size > mask;
204 /* Map a single continuous physical area into the IOMMU.
205 * Caller needs to check if the iommu is needed and flush.
207 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
208 size_t size, int dir)
210 unsigned long npages = to_pages(phys_mem, size);
211 unsigned long iommu_page = alloc_iommu(npages);
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 SET_LEAK(iommu_page + i);
225 phys_mem += PAGE_SIZE;
227 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
230 static dma_addr_t gart_map_simple(struct device *dev, char *buf,
231 size_t size, int dir)
233 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
238 /* Map a single area into the IOMMU */
239 static dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
241 unsigned long phys_mem, bus;
246 phys_mem = virt_to_phys(addr);
247 if (!need_iommu(dev, phys_mem, size))
250 bus = gart_map_simple(dev, addr, size, dir);
255 * Free a DMA mapping.
257 static void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
258 size_t size, int direction)
260 unsigned long iommu_page;
264 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
265 dma_addr >= iommu_bus_base + iommu_size)
267 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
268 npages = to_pages(dma_addr, size);
269 for (i = 0; i < npages; i++) {
270 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
271 CLEAR_LEAK(iommu_page + i);
273 free_iommu(iommu_page, npages);
277 * Wrapper for pci_unmap_single working with scatterlists.
279 static void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
283 for (i = 0; i < nents; i++) {
284 struct scatterlist *s = &sg[i];
285 if (!s->dma_length || !s->length)
287 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
291 /* Fallback for dma_map_sg in case of overflow */
292 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
297 #ifdef CONFIG_IOMMU_DEBUG
298 printk(KERN_DEBUG "dma_map_sg overflow\n");
301 for (i = 0; i < nents; i++ ) {
302 struct scatterlist *s = &sg[i];
303 unsigned long addr = page_to_phys(s->page) + s->offset;
304 if (nonforced_iommu(dev, addr, s->length)) {
305 addr = dma_map_area(dev, addr, s->length, dir);
306 if (addr == bad_dma_address) {
308 gart_unmap_sg(dev, sg, i, dir);
310 sg[0].dma_length = 0;
314 s->dma_address = addr;
315 s->dma_length = s->length;
321 /* Map multiple scatterlist entries continuous into the first. */
322 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
323 struct scatterlist *sout, unsigned long pages)
325 unsigned long iommu_start = alloc_iommu(pages);
326 unsigned long iommu_page = iommu_start;
329 if (iommu_start == -1)
332 for (i = start; i < stopat; i++) {
333 struct scatterlist *s = &sg[i];
334 unsigned long pages, addr;
335 unsigned long phys_addr = s->dma_address;
337 BUG_ON(i > 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 = to_pages(s->offset, s->length);
350 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
351 SET_LEAK(iommu_page);
356 BUG_ON(iommu_page - iommu_start != pages);
360 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
361 struct scatterlist *sout,
362 unsigned long pages, int need)
365 BUG_ON(stopat - start != 1);
367 sout->dma_length = sg[start].length;
370 return __dma_map_cont(sg, start, stopat, sout, pages);
374 * DMA map all entries in a scatterlist.
375 * Merge chunks that have page aligned sizes into a continuous mapping.
377 int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
382 unsigned long pages = 0;
383 int need = 0, nextneed;
393 for (i = 0; i < nents; i++) {
394 struct scatterlist *s = &sg[i];
395 dma_addr_t addr = page_to_phys(s->page) + s->offset;
396 s->dma_address = addr;
397 BUG_ON(s->length == 0);
399 nextneed = need_iommu(dev, addr, s->length);
401 /* Handle the previous not yet processed entries */
403 struct scatterlist *ps = &sg[i-1];
404 /* Can only merge when the last chunk ends on a page
405 boundary and the new one doesn't have an offset. */
406 if (!iommu_merge || !nextneed || !need || s->offset ||
407 (ps->offset + ps->length) % PAGE_SIZE) {
408 if (dma_map_cont(sg, start, i, sg+out, pages,
418 pages += to_pages(s->offset, s->length);
420 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
425 sg[out].dma_length = 0;
430 gart_unmap_sg(dev, sg, nents, dir);
431 /* When it was forced or merged try again in a dumb way */
432 if (force_iommu || iommu_merge) {
433 out = dma_map_sg_nonforce(dev, sg, nents, dir);
437 if (panic_on_overflow)
438 panic("dma_map_sg: overflow on %lu pages\n", pages);
439 iommu_full(dev, pages << PAGE_SHIFT, dir);
440 for (i = 0; i < nents; i++)
441 sg[i].dma_address = bad_dma_address;
447 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
451 iommu_size = aper_size;
456 a = aper + iommu_size;
457 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
459 if (iommu_size < 64*1024*1024)
461 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
466 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
468 unsigned aper_size = 0, aper_base_32;
472 pci_read_config_dword(dev, 0x94, &aper_base_32);
473 pci_read_config_dword(dev, 0x90, &aper_order);
474 aper_order = (aper_order >> 1) & 7;
476 aper_base = aper_base_32 & 0x7fff;
479 aper_size = (32 * 1024 * 1024) << aper_order;
480 if (aper_base + aper_size > 0x100000000UL || !aper_size)
488 * Private Northbridge GATT initialization in case we cannot use the
489 * AGP driver for some reason.
491 static __init int init_k8_gatt(struct agp_kern_info *info)
495 unsigned aper_base, new_aper_base;
496 unsigned aper_size, gatt_size, new_aper_size;
499 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
500 aper_size = aper_base = info->aper_size = 0;
502 for (i = 0; i < num_k8_northbridges; i++) {
503 dev = k8_northbridges[i];
504 new_aper_base = read_aperture(dev, &new_aper_size);
509 aper_size = new_aper_size;
510 aper_base = new_aper_base;
512 if (aper_size != new_aper_size || aper_base != new_aper_base)
517 info->aper_base = aper_base;
518 info->aper_size = aper_size>>20;
520 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
521 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
523 panic("Cannot allocate GATT table");
524 if (change_page_attr_addr((unsigned long)gatt, gatt_size >> PAGE_SHIFT, PAGE_KERNEL_NOCACHE))
525 panic("Could not set GART PTEs to uncacheable pages");
528 memset(gatt, 0, gatt_size);
529 agp_gatt_table = gatt;
531 for (i = 0; i < num_k8_northbridges; i++) {
535 dev = k8_northbridges[i];
536 gatt_reg = __pa(gatt) >> 12;
538 pci_write_config_dword(dev, 0x98, gatt_reg);
539 pci_read_config_dword(dev, 0x90, &ctl);
542 ctl &= ~((1<<4) | (1<<5));
544 pci_write_config_dword(dev, 0x90, ctl);
548 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
552 /* Should not happen anymore */
553 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
554 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
558 extern int agp_amd64_init(void);
560 static const struct dma_mapping_ops gart_dma_ops = {
561 .mapping_error = NULL,
562 .map_single = gart_map_single,
563 .map_simple = gart_map_simple,
564 .unmap_single = gart_unmap_single,
565 .sync_single_for_cpu = NULL,
566 .sync_single_for_device = NULL,
567 .sync_single_range_for_cpu = NULL,
568 .sync_single_range_for_device = NULL,
569 .sync_sg_for_cpu = NULL,
570 .sync_sg_for_device = NULL,
571 .map_sg = gart_map_sg,
572 .unmap_sg = gart_unmap_sg,
575 void gart_iommu_shutdown(void)
580 if (no_agp && (dma_ops != &gart_dma_ops))
583 for (i = 0; i < num_k8_northbridges; i++) {
586 dev = k8_northbridges[i];
587 pci_read_config_dword(dev, 0x90, &ctl);
591 pci_write_config_dword(dev, 0x90, ctl);
595 void __init gart_iommu_init(void)
597 struct agp_kern_info info;
598 unsigned long aper_size;
599 unsigned long iommu_start;
600 unsigned long scratch;
603 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
604 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
608 #ifndef CONFIG_AGP_AMD64
611 /* Makefile puts PCI initialization via subsys_initcall first. */
612 /* Add other K8 AGP bridge drivers here */
614 (agp_amd64_init() < 0) ||
615 (agp_copy_info(agp_bridge, &info) < 0);
621 /* Did we detect a different HW IOMMU? */
622 if (iommu_detected && !iommu_aperture)
626 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
628 (no_agp && init_k8_gatt(&info) < 0)) {
629 if (end_pfn > MAX_DMA32_PFN) {
630 printk(KERN_ERR "WARNING more than 4GB of memory "
631 "but GART IOMMU not available.\n"
632 KERN_ERR "WARNING 32bit PCI may malfunction.\n");
637 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
638 aper_size = info.aper_size * 1024 * 1024;
639 iommu_size = check_iommu_size(info.aper_base, aper_size);
640 iommu_pages = iommu_size >> PAGE_SHIFT;
642 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
643 get_order(iommu_pages/8));
644 if (!iommu_gart_bitmap)
645 panic("Cannot allocate iommu bitmap\n");
646 memset(iommu_gart_bitmap, 0, iommu_pages/8);
648 #ifdef CONFIG_IOMMU_LEAK
650 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
651 get_order(iommu_pages*sizeof(void *)));
653 memset(iommu_leak_tab, 0, iommu_pages * 8);
655 printk("PCI-DMA: Cannot allocate leak trace area\n");
660 * Out of IOMMU space handling.
661 * Reserve some invalid pages at the beginning of the GART.
663 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
665 agp_memory_reserved = iommu_size;
667 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
670 iommu_start = aper_size - iommu_size;
671 iommu_bus_base = info.aper_base + iommu_start;
672 bad_dma_address = iommu_bus_base;
673 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
676 * Unmap the IOMMU part of the GART. The alias of the page is
677 * always mapped with cache enabled and there is no full cache
678 * coherency across the GART remapping. The unmapping avoids
679 * automatic prefetches from the CPU allocating cache lines in
680 * there. All CPU accesses are done via the direct mapping to
681 * the backing memory. The GART address is only used by PCI
684 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
687 * Try to workaround a bug (thanks to BenH)
688 * Set unmapped entries to a scratch page instead of 0.
689 * Any prefetches that hit unmapped entries won't get an bus abort
692 scratch = get_zeroed_page(GFP_KERNEL);
694 panic("Cannot allocate iommu scratch page");
695 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
696 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
697 iommu_gatt_base[i] = gart_unmapped_entry;
700 dma_ops = &gart_dma_ops;
703 void __init gart_parse_options(char *p)
707 #ifdef CONFIG_IOMMU_LEAK
708 if (!strncmp(p,"leak",4)) {
712 if (isdigit(*p) && get_option(&p, &arg))
713 iommu_leak_pages = arg;
716 if (isdigit(*p) && get_option(&p, &arg))
718 if (!strncmp(p, "fullflush",8))
720 if (!strncmp(p, "nofullflush",11))
722 if (!strncmp(p,"noagp",5))
724 if (!strncmp(p, "noaperture",10))
726 /* duplicated from pci-dma.c */
727 if (!strncmp(p,"force",5))
728 iommu_aperture_allowed = 1;
729 if (!strncmp(p,"allowed",7))
730 iommu_aperture_allowed = 1;
731 if (!strncmp(p, "memaper", 7)) {
732 fallback_aper_force = 1;
736 if (get_option(&p, &arg))
737 fallback_aper_order = arg;