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/config.h>
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 <asm/atomic.h>
29 #include <asm/pgtable.h>
30 #include <asm/proto.h>
31 #include <asm/cacheflush.h>
32 #include <asm/kdebug.h>
33 #include <asm/swiotlb.h>
36 unsigned long iommu_bus_base; /* GART remapping area (physical) */
37 static unsigned long iommu_size; /* size of remapping area bytes */
38 static unsigned long iommu_pages; /* .. and in pages */
40 u32 *iommu_gatt_base; /* Remapping table */
42 /* If this is disabled the IOMMU will use an optimized flushing strategy
43 of only flushing when an mapping is reused. With it true the GART is flushed
44 for every mapping. Problem is that doing the lazy flush seems to trigger
45 bugs with some popular PCI cards, in particular 3ware (but has been also
46 also seen with Qlogic at least). */
47 int iommu_fullflush = 1;
51 /* Allocation bitmap for the remapping area */
52 static DEFINE_SPINLOCK(iommu_bitmap_lock);
53 static unsigned long *iommu_gart_bitmap; /* guarded by iommu_bitmap_lock */
55 static u32 gart_unmapped_entry;
58 #define GPTE_COHERENT 2
59 #define GPTE_ENCODE(x) \
60 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
61 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
63 #define to_pages(addr,size) \
64 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
66 #define for_all_nb(dev) \
68 while ((dev = pci_get_device(PCI_VENDOR_ID_AMD, 0x1103, dev))!=NULL)\
69 if (dev->bus->number == 0 && \
70 (PCI_SLOT(dev->devfn) >= 24) && (PCI_SLOT(dev->devfn) <= 31))
72 static struct pci_dev *northbridges[MAX_NB];
73 static u32 northbridge_flush_word[MAX_NB];
75 #define EMERGENCY_PAGES 32 /* = 128KB */
78 #define AGPEXTERN extern
83 /* backdoor interface to AGP driver */
84 AGPEXTERN int agp_memory_reserved;
85 AGPEXTERN __u32 *agp_gatt_table;
87 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
88 static int need_flush; /* global flush state. set for each gart wrap */
90 static unsigned long alloc_iommu(int size)
92 unsigned long offset, flags;
94 spin_lock_irqsave(&iommu_bitmap_lock, flags);
95 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
98 offset = find_next_zero_string(iommu_gart_bitmap,0,next_bit,size);
101 set_bit_string(iommu_gart_bitmap, offset, size);
102 next_bit = offset+size;
103 if (next_bit >= iommu_pages) {
110 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
114 static void free_iommu(unsigned long offset, int size)
118 clear_bit(offset, iommu_gart_bitmap);
121 spin_lock_irqsave(&iommu_bitmap_lock, flags);
122 __clear_bit_string(iommu_gart_bitmap, offset, size);
123 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
127 * Use global flush state to avoid races with multiple flushers.
129 static void flush_gart(struct device *dev)
135 spin_lock_irqsave(&iommu_bitmap_lock, flags);
138 for (i = 0; i < MAX_NB; i++) {
139 if (!northbridges[i])
141 pci_write_config_dword(northbridges[i], 0x9c,
142 northbridge_flush_word[i] | 1);
146 for (i = 0; i <= max; i++) {
148 if (!northbridges[i])
150 /* Make sure the hardware actually executed the flush. */
152 pci_read_config_dword(northbridges[i], 0x9c, &w);
156 printk("nothing to flush?\n");
159 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
164 #ifdef CONFIG_IOMMU_LEAK
166 #define SET_LEAK(x) if (iommu_leak_tab) \
167 iommu_leak_tab[x] = __builtin_return_address(0);
168 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
169 iommu_leak_tab[x] = NULL;
171 /* Debugging aid for drivers that don't free their IOMMU tables */
172 static void **iommu_leak_tab;
173 static int leak_trace;
174 int iommu_leak_pages = 20;
179 if (dump || !iommu_leak_tab) return;
181 show_stack(NULL,NULL);
182 /* Very crude. dump some from the end of the table too */
183 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
184 for (i = 0; i < iommu_leak_pages; i+=2) {
185 printk("%lu: ", iommu_pages-i);
186 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
187 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
193 #define CLEAR_LEAK(x)
196 static void iommu_full(struct device *dev, size_t size, int dir)
199 * Ran out of IOMMU space for this operation. This is very bad.
200 * Unfortunately the drivers cannot handle this operation properly.
201 * Return some non mapped prereserved space in the aperture and
202 * let the Northbridge deal with it. This will result in garbage
203 * in the IO operation. When the size exceeds the prereserved space
204 * memory corruption will occur or random memory will be DMAed
205 * out. Hopefully no network devices use single mappings that big.
209 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
212 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
213 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
214 panic("PCI-DMA: Memory would be corrupted\n");
215 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
216 panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
219 #ifdef CONFIG_IOMMU_LEAK
224 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
226 u64 mask = *dev->dma_mask;
227 int high = addr + size >= mask;
233 panic("PCI-DMA: high address but no IOMMU.\n");
239 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
241 u64 mask = *dev->dma_mask;
242 int high = addr + size >= mask;
246 panic("PCI-DMA: high address but no IOMMU.\n");
252 /* Map a single continuous physical area into the IOMMU.
253 * Caller needs to check if the iommu is needed and flush.
255 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
256 size_t size, int dir)
258 unsigned long npages = to_pages(phys_mem, size);
259 unsigned long iommu_page = alloc_iommu(npages);
261 if (iommu_page == -1) {
262 if (!nonforced_iommu(dev, phys_mem, size))
264 if (panic_on_overflow)
265 panic("dma_map_area overflow %lu bytes\n", size);
266 iommu_full(dev, size, dir);
267 return bad_dma_address;
270 for (i = 0; i < npages; i++) {
271 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
272 SET_LEAK(iommu_page + i);
273 phys_mem += PAGE_SIZE;
275 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
278 static dma_addr_t gart_map_simple(struct device *dev, char *buf,
279 size_t size, int dir)
281 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
286 /* Map a single area into the IOMMU */
287 dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
289 unsigned long phys_mem, bus;
291 BUG_ON(dir == DMA_NONE);
296 phys_mem = virt_to_phys(addr);
297 if (!need_iommu(dev, phys_mem, size))
300 bus = gart_map_simple(dev, addr, size, dir);
305 * Wrapper for pci_unmap_single working with scatterlists.
307 void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
311 for (i = 0; i < nents; i++) {
312 struct scatterlist *s = &sg[i];
313 if (!s->dma_length || !s->length)
315 dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
319 /* Fallback for dma_map_sg in case of overflow */
320 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
325 #ifdef CONFIG_IOMMU_DEBUG
326 printk(KERN_DEBUG "dma_map_sg overflow\n");
329 for (i = 0; i < nents; i++ ) {
330 struct scatterlist *s = &sg[i];
331 unsigned long addr = page_to_phys(s->page) + s->offset;
332 if (nonforced_iommu(dev, addr, s->length)) {
333 addr = dma_map_area(dev, addr, s->length, dir);
334 if (addr == bad_dma_address) {
336 gart_unmap_sg(dev, sg, i, dir);
338 sg[0].dma_length = 0;
342 s->dma_address = addr;
343 s->dma_length = s->length;
349 /* Map multiple scatterlist entries continuous into the first. */
350 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
351 struct scatterlist *sout, unsigned long pages)
353 unsigned long iommu_start = alloc_iommu(pages);
354 unsigned long iommu_page = iommu_start;
357 if (iommu_start == -1)
360 for (i = start; i < stopat; i++) {
361 struct scatterlist *s = &sg[i];
362 unsigned long pages, addr;
363 unsigned long phys_addr = s->dma_address;
365 BUG_ON(i > start && s->offset);
368 sout->dma_address = iommu_bus_base;
369 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
370 sout->dma_length = s->length;
372 sout->dma_length += s->length;
376 pages = to_pages(s->offset, s->length);
378 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
379 SET_LEAK(iommu_page);
384 BUG_ON(iommu_page - iommu_start != pages);
388 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
389 struct scatterlist *sout,
390 unsigned long pages, int need)
393 BUG_ON(stopat - start != 1);
395 sout->dma_length = sg[start].length;
398 return __dma_map_cont(sg, start, stopat, sout, pages);
402 * DMA map all entries in a scatterlist.
403 * Merge chunks that have page aligned sizes into a continuous mapping.
405 int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
410 unsigned long pages = 0;
411 int need = 0, nextneed;
413 BUG_ON(dir == DMA_NONE);
422 for (i = 0; i < nents; i++) {
423 struct scatterlist *s = &sg[i];
424 dma_addr_t addr = page_to_phys(s->page) + s->offset;
425 s->dma_address = addr;
426 BUG_ON(s->length == 0);
428 nextneed = need_iommu(dev, addr, s->length);
430 /* Handle the previous not yet processed entries */
432 struct scatterlist *ps = &sg[i-1];
433 /* Can only merge when the last chunk ends on a page
434 boundary and the new one doesn't have an offset. */
435 if (!iommu_merge || !nextneed || !need || s->offset ||
436 (ps->offset + ps->length) % PAGE_SIZE) {
437 if (dma_map_cont(sg, start, i, sg+out, pages,
447 pages += to_pages(s->offset, s->length);
449 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
454 sg[out].dma_length = 0;
459 gart_unmap_sg(dev, sg, nents, dir);
460 /* When it was forced try again unforced */
462 return dma_map_sg_nonforce(dev, sg, nents, dir);
463 if (panic_on_overflow)
464 panic("dma_map_sg: overflow on %lu pages\n", pages);
465 iommu_full(dev, pages << PAGE_SHIFT, dir);
466 for (i = 0; i < nents; i++)
467 sg[i].dma_address = bad_dma_address;
472 * Free a DMA mapping.
474 void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
475 size_t size, int direction)
477 unsigned long iommu_page;
481 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
482 dma_addr >= iommu_bus_base + iommu_size)
484 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
485 npages = to_pages(dma_addr, size);
486 for (i = 0; i < npages; i++) {
487 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
488 CLEAR_LEAK(iommu_page + i);
490 free_iommu(iommu_page, npages);
495 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
499 iommu_size = aper_size;
504 a = aper + iommu_size;
505 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
507 if (iommu_size < 64*1024*1024)
509 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
514 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
516 unsigned aper_size = 0, aper_base_32;
520 pci_read_config_dword(dev, 0x94, &aper_base_32);
521 pci_read_config_dword(dev, 0x90, &aper_order);
522 aper_order = (aper_order >> 1) & 7;
524 aper_base = aper_base_32 & 0x7fff;
527 aper_size = (32 * 1024 * 1024) << aper_order;
528 if (aper_base + aper_size >= 0xffffffff || !aper_size)
536 * Private Northbridge GATT initialization in case we cannot use the
537 * AGP driver for some reason.
539 static __init int init_k8_gatt(struct agp_kern_info *info)
543 unsigned aper_base, new_aper_base;
544 unsigned aper_size, gatt_size, new_aper_size;
546 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
547 aper_size = aper_base = info->aper_size = 0;
549 new_aper_base = read_aperture(dev, &new_aper_size);
554 aper_size = new_aper_size;
555 aper_base = new_aper_base;
557 if (aper_size != new_aper_size || aper_base != new_aper_base)
562 info->aper_base = aper_base;
563 info->aper_size = aper_size>>20;
565 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
566 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
568 panic("Cannot allocate GATT table");
569 memset(gatt, 0, gatt_size);
570 agp_gatt_table = gatt;
576 gatt_reg = __pa(gatt) >> 12;
578 pci_write_config_dword(dev, 0x98, gatt_reg);
579 pci_read_config_dword(dev, 0x90, &ctl);
582 ctl &= ~((1<<4) | (1<<5));
584 pci_write_config_dword(dev, 0x90, ctl);
588 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
592 /* Should not happen anymore */
593 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
594 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
598 extern int agp_amd64_init(void);
600 static struct dma_mapping_ops gart_dma_ops = {
601 .mapping_error = NULL,
602 .map_single = gart_map_single,
603 .map_simple = gart_map_simple,
604 .unmap_single = gart_unmap_single,
605 .sync_single_for_cpu = NULL,
606 .sync_single_for_device = NULL,
607 .sync_single_range_for_cpu = NULL,
608 .sync_single_range_for_device = NULL,
609 .sync_sg_for_cpu = NULL,
610 .sync_sg_for_device = NULL,
611 .map_sg = gart_map_sg,
612 .unmap_sg = gart_unmap_sg,
615 static int __init pci_iommu_init(void)
617 struct agp_kern_info info;
618 unsigned long aper_size;
619 unsigned long iommu_start;
621 unsigned long scratch;
624 #ifndef CONFIG_AGP_AMD64
627 /* Makefile puts PCI initialization via subsys_initcall first. */
628 /* Add other K8 AGP bridge drivers here */
630 (agp_amd64_init() < 0) ||
631 (agp_copy_info(agp_bridge, &info) < 0);
640 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
642 (no_agp && init_k8_gatt(&info) < 0)) {
648 aper_size = info.aper_size * 1024 * 1024;
649 iommu_size = check_iommu_size(info.aper_base, aper_size);
650 iommu_pages = iommu_size >> PAGE_SHIFT;
652 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
653 get_order(iommu_pages/8));
654 if (!iommu_gart_bitmap)
655 panic("Cannot allocate iommu bitmap\n");
656 memset(iommu_gart_bitmap, 0, iommu_pages/8);
658 #ifdef CONFIG_IOMMU_LEAK
660 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
661 get_order(iommu_pages*sizeof(void *)));
663 memset(iommu_leak_tab, 0, iommu_pages * 8);
665 printk("PCI-DMA: Cannot allocate leak trace area\n");
670 * Out of IOMMU space handling.
671 * Reserve some invalid pages at the beginning of the GART.
673 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
675 agp_memory_reserved = iommu_size;
677 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
680 iommu_start = aper_size - iommu_size;
681 iommu_bus_base = info.aper_base + iommu_start;
682 bad_dma_address = iommu_bus_base;
683 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
686 * Unmap the IOMMU part of the GART. The alias of the page is
687 * always mapped with cache enabled and there is no full cache
688 * coherency across the GART remapping. The unmapping avoids
689 * automatic prefetches from the CPU allocating cache lines in
690 * there. All CPU accesses are done via the direct mapping to
691 * the backing memory. The GART address is only used by PCI
694 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
697 * Try to workaround a bug (thanks to BenH)
698 * Set unmapped entries to a scratch page instead of 0.
699 * Any prefetches that hit unmapped entries won't get an bus abort
702 scratch = get_zeroed_page(GFP_KERNEL);
704 panic("Cannot allocate iommu scratch page");
705 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
706 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
707 iommu_gatt_base[i] = gart_unmapped_entry;
711 int cpu = PCI_SLOT(dev->devfn) - 24;
714 northbridges[cpu] = dev;
715 pci_read_config_dword(dev, 0x9c, &flag); /* cache flush word */
716 northbridge_flush_word[cpu] = flag;
721 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
722 dma_ops = &gart_dma_ops;
727 /* Must execute after PCI subsystem */
728 fs_initcall(pci_iommu_init);
730 void gart_parse_options(char *p)
734 #ifdef CONFIG_IOMMU_LEAK
735 if (!strncmp(p,"leak",4)) {
739 if (isdigit(*p) && get_option(&p, &arg))
740 iommu_leak_pages = arg;
743 if (isdigit(*p) && get_option(&p, &arg))
745 if (!strncmp(p, "fullflush",8))
747 if (!strncmp(p, "nofullflush",11))
749 if (!strncmp(p,"noagp",5))
751 if (!strncmp(p, "noaperture",10))
753 /* duplicated from pci-dma.c */
754 if (!strncmp(p,"force",5))
755 iommu_aperture_allowed = 1;
756 if (!strncmp(p,"allowed",7))
757 iommu_aperture_allowed = 1;
758 if (!strncmp(p, "memaper", 7)) {
759 fallback_aper_force = 1;
763 if (get_option(&p, &arg))
764 fallback_aper_order = arg;