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 <asm/atomic.h>
28 #include <asm/pgtable.h>
29 #include <asm/proto.h>
30 #include <asm/cacheflush.h>
31 #include <asm/kdebug.h>
32 #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;
49 /* Allocation bitmap for the remapping area */
50 static DEFINE_SPINLOCK(iommu_bitmap_lock);
51 static unsigned long *iommu_gart_bitmap; /* guarded by iommu_bitmap_lock */
53 static u32 gart_unmapped_entry;
56 #define GPTE_COHERENT 2
57 #define GPTE_ENCODE(x) \
58 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
59 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
61 #define to_pages(addr,size) \
62 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
64 #define EMERGENCY_PAGES 32 /* = 128KB */
67 #define AGPEXTERN extern
72 /* backdoor interface to AGP driver */
73 AGPEXTERN int agp_memory_reserved;
74 AGPEXTERN __u32 *agp_gatt_table;
76 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
77 static int need_flush; /* global flush state. set for each gart wrap */
79 static unsigned long alloc_iommu(int size)
81 unsigned long offset, flags;
83 spin_lock_irqsave(&iommu_bitmap_lock, flags);
84 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
87 offset = find_next_zero_string(iommu_gart_bitmap,0,iommu_pages,size);
90 set_bit_string(iommu_gart_bitmap, offset, size);
91 next_bit = offset+size;
92 if (next_bit >= iommu_pages) {
99 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
103 static void free_iommu(unsigned long offset, int size)
106 spin_lock_irqsave(&iommu_bitmap_lock, flags);
107 __clear_bit_string(iommu_gart_bitmap, offset, size);
108 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
112 * Use global flush state to avoid races with multiple flushers.
114 static void flush_gart(void)
117 spin_lock_irqsave(&iommu_bitmap_lock, flags);
122 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
125 #ifdef CONFIG_IOMMU_LEAK
127 #define SET_LEAK(x) if (iommu_leak_tab) \
128 iommu_leak_tab[x] = __builtin_return_address(0);
129 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
130 iommu_leak_tab[x] = NULL;
132 /* Debugging aid for drivers that don't free their IOMMU tables */
133 static void **iommu_leak_tab;
134 static int leak_trace;
135 int iommu_leak_pages = 20;
140 if (dump || !iommu_leak_tab) return;
142 show_stack(NULL,NULL);
143 /* Very crude. dump some from the end of the table too */
144 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
145 for (i = 0; i < iommu_leak_pages; i+=2) {
146 printk("%lu: ", iommu_pages-i);
147 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
148 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
154 #define CLEAR_LEAK(x)
157 static void iommu_full(struct device *dev, size_t size, int dir)
160 * Ran out of IOMMU space for this operation. This is very bad.
161 * Unfortunately the drivers cannot handle this operation properly.
162 * Return some non mapped prereserved space in the aperture and
163 * let the Northbridge deal with it. This will result in garbage
164 * in the IO operation. When the size exceeds the prereserved space
165 * memory corruption will occur or random memory will be DMAed
166 * out. Hopefully no network devices use single mappings that big.
170 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
173 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
174 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
175 panic("PCI-DMA: Memory would be corrupted\n");
176 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
177 panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
180 #ifdef CONFIG_IOMMU_LEAK
185 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
187 u64 mask = *dev->dma_mask;
188 int high = addr + size >= mask;
195 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
197 u64 mask = *dev->dma_mask;
198 int high = addr + size >= mask;
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)
209 unsigned long npages = to_pages(phys_mem, size);
210 unsigned long iommu_page = alloc_iommu(npages);
212 if (iommu_page == -1) {
213 if (!nonforced_iommu(dev, phys_mem, size))
215 if (panic_on_overflow)
216 panic("dma_map_area overflow %lu bytes\n", size);
217 iommu_full(dev, size, dir);
218 return bad_dma_address;
221 for (i = 0; i < npages; i++) {
222 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
223 SET_LEAK(iommu_page + i);
224 phys_mem += PAGE_SIZE;
226 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
229 static dma_addr_t gart_map_simple(struct device *dev, char *buf,
230 size_t size, int dir)
232 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
237 /* Map a single area into the IOMMU */
238 dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
240 unsigned long phys_mem, bus;
245 phys_mem = virt_to_phys(addr);
246 if (!need_iommu(dev, phys_mem, size))
249 bus = gart_map_simple(dev, addr, size, dir);
254 * Free a DMA mapping.
256 void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
257 size_t size, int direction)
259 unsigned long iommu_page;
263 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
264 dma_addr >= iommu_bus_base + iommu_size)
266 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
267 npages = to_pages(dma_addr, size);
268 for (i = 0; i < npages; i++) {
269 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
270 CLEAR_LEAK(iommu_page + i);
272 free_iommu(iommu_page, npages);
276 * Wrapper for pci_unmap_single working with scatterlists.
278 void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
282 for (i = 0; i < nents; i++) {
283 struct scatterlist *s = &sg[i];
284 if (!s->dma_length || !s->length)
286 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
290 /* Fallback for dma_map_sg in case of overflow */
291 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
296 #ifdef CONFIG_IOMMU_DEBUG
297 printk(KERN_DEBUG "dma_map_sg overflow\n");
300 for (i = 0; i < nents; i++ ) {
301 struct scatterlist *s = &sg[i];
302 unsigned long addr = page_to_phys(s->page) + s->offset;
303 if (nonforced_iommu(dev, addr, s->length)) {
304 addr = dma_map_area(dev, addr, s->length, dir);
305 if (addr == bad_dma_address) {
307 gart_unmap_sg(dev, sg, i, dir);
309 sg[0].dma_length = 0;
313 s->dma_address = addr;
314 s->dma_length = s->length;
320 /* Map multiple scatterlist entries continuous into the first. */
321 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
322 struct scatterlist *sout, unsigned long pages)
324 unsigned long iommu_start = alloc_iommu(pages);
325 unsigned long iommu_page = iommu_start;
328 if (iommu_start == -1)
331 for (i = start; i < stopat; i++) {
332 struct scatterlist *s = &sg[i];
333 unsigned long pages, addr;
334 unsigned long phys_addr = s->dma_address;
336 BUG_ON(i > start && s->offset);
339 sout->dma_address = iommu_bus_base;
340 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
341 sout->dma_length = s->length;
343 sout->dma_length += s->length;
347 pages = to_pages(s->offset, s->length);
349 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
350 SET_LEAK(iommu_page);
355 BUG_ON(iommu_page - iommu_start != pages);
359 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
360 struct scatterlist *sout,
361 unsigned long pages, int need)
364 BUG_ON(stopat - start != 1);
366 sout->dma_length = sg[start].length;
369 return __dma_map_cont(sg, start, stopat, sout, pages);
373 * DMA map all entries in a scatterlist.
374 * Merge chunks that have page aligned sizes into a continuous mapping.
376 int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
381 unsigned long pages = 0;
382 int need = 0, nextneed;
392 for (i = 0; i < nents; i++) {
393 struct scatterlist *s = &sg[i];
394 dma_addr_t addr = page_to_phys(s->page) + s->offset;
395 s->dma_address = addr;
396 BUG_ON(s->length == 0);
398 nextneed = need_iommu(dev, addr, s->length);
400 /* Handle the previous not yet processed entries */
402 struct scatterlist *ps = &sg[i-1];
403 /* Can only merge when the last chunk ends on a page
404 boundary and the new one doesn't have an offset. */
405 if (!iommu_merge || !nextneed || !need || s->offset ||
406 (ps->offset + ps->length) % PAGE_SIZE) {
407 if (dma_map_cont(sg, start, i, sg+out, pages,
417 pages += to_pages(s->offset, s->length);
419 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
424 sg[out].dma_length = 0;
429 gart_unmap_sg(dev, sg, nents, dir);
430 /* When it was forced or merged try again in a dumb way */
431 if (force_iommu || iommu_merge) {
432 out = dma_map_sg_nonforce(dev, sg, nents, dir);
436 if (panic_on_overflow)
437 panic("dma_map_sg: overflow on %lu pages\n", pages);
438 iommu_full(dev, pages << PAGE_SHIFT, dir);
439 for (i = 0; i < nents; i++)
440 sg[i].dma_address = bad_dma_address;
446 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
450 iommu_size = aper_size;
455 a = aper + iommu_size;
456 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
458 if (iommu_size < 64*1024*1024)
460 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
465 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
467 unsigned aper_size = 0, aper_base_32;
471 pci_read_config_dword(dev, 0x94, &aper_base_32);
472 pci_read_config_dword(dev, 0x90, &aper_order);
473 aper_order = (aper_order >> 1) & 7;
475 aper_base = aper_base_32 & 0x7fff;
478 aper_size = (32 * 1024 * 1024) << aper_order;
479 if (aper_base + aper_size >= 0xffffffff || !aper_size)
487 * Private Northbridge GATT initialization in case we cannot use the
488 * AGP driver for some reason.
490 static __init int init_k8_gatt(struct agp_kern_info *info)
494 unsigned aper_base, new_aper_base;
495 unsigned aper_size, gatt_size, new_aper_size;
498 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
499 aper_size = aper_base = info->aper_size = 0;
501 for (i = 0; i < num_k8_northbridges; i++) {
502 dev = k8_northbridges[i];
503 new_aper_base = read_aperture(dev, &new_aper_size);
508 aper_size = new_aper_size;
509 aper_base = new_aper_base;
511 if (aper_size != new_aper_size || aper_base != new_aper_base)
516 info->aper_base = aper_base;
517 info->aper_size = aper_size>>20;
519 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
520 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
522 panic("Cannot allocate GATT table");
523 memset(gatt, 0, gatt_size);
524 agp_gatt_table = gatt;
526 for (i = 0; i < num_k8_northbridges; i++) {
530 dev = k8_northbridges[i];
531 gatt_reg = __pa(gatt) >> 12;
533 pci_write_config_dword(dev, 0x98, gatt_reg);
534 pci_read_config_dword(dev, 0x90, &ctl);
537 ctl &= ~((1<<4) | (1<<5));
539 pci_write_config_dword(dev, 0x90, ctl);
543 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
547 /* Should not happen anymore */
548 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
549 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
553 extern int agp_amd64_init(void);
555 static struct dma_mapping_ops gart_dma_ops = {
556 .mapping_error = NULL,
557 .map_single = gart_map_single,
558 .map_simple = gart_map_simple,
559 .unmap_single = gart_unmap_single,
560 .sync_single_for_cpu = NULL,
561 .sync_single_for_device = NULL,
562 .sync_single_range_for_cpu = NULL,
563 .sync_single_range_for_device = NULL,
564 .sync_sg_for_cpu = NULL,
565 .sync_sg_for_device = NULL,
566 .map_sg = gart_map_sg,
567 .unmap_sg = gart_unmap_sg,
570 void __init gart_iommu_init(void)
572 struct agp_kern_info info;
573 unsigned long aper_size;
574 unsigned long iommu_start;
575 unsigned long scratch;
578 if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
579 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
583 #ifndef CONFIG_AGP_AMD64
586 /* Makefile puts PCI initialization via subsys_initcall first. */
587 /* Add other K8 AGP bridge drivers here */
589 (agp_amd64_init() < 0) ||
590 (agp_copy_info(agp_bridge, &info) < 0);
596 /* Did we detect a different HW IOMMU? */
597 if (iommu_detected && !iommu_aperture)
601 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
603 (no_agp && init_k8_gatt(&info) < 0)) {
604 if (end_pfn > MAX_DMA32_PFN) {
605 printk(KERN_ERR "WARNING more than 4GB of memory "
606 "but GART IOMMU not available.\n"
607 KERN_ERR "WARNING 32bit PCI may malfunction.\n");
612 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
613 aper_size = info.aper_size * 1024 * 1024;
614 iommu_size = check_iommu_size(info.aper_base, aper_size);
615 iommu_pages = iommu_size >> PAGE_SHIFT;
617 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
618 get_order(iommu_pages/8));
619 if (!iommu_gart_bitmap)
620 panic("Cannot allocate iommu bitmap\n");
621 memset(iommu_gart_bitmap, 0, iommu_pages/8);
623 #ifdef CONFIG_IOMMU_LEAK
625 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
626 get_order(iommu_pages*sizeof(void *)));
628 memset(iommu_leak_tab, 0, iommu_pages * 8);
630 printk("PCI-DMA: Cannot allocate leak trace area\n");
635 * Out of IOMMU space handling.
636 * Reserve some invalid pages at the beginning of the GART.
638 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
640 agp_memory_reserved = iommu_size;
642 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
645 iommu_start = aper_size - iommu_size;
646 iommu_bus_base = info.aper_base + iommu_start;
647 bad_dma_address = iommu_bus_base;
648 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
651 * Unmap the IOMMU part of the GART. The alias of the page is
652 * always mapped with cache enabled and there is no full cache
653 * coherency across the GART remapping. The unmapping avoids
654 * automatic prefetches from the CPU allocating cache lines in
655 * there. All CPU accesses are done via the direct mapping to
656 * the backing memory. The GART address is only used by PCI
659 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
662 * Try to workaround a bug (thanks to BenH)
663 * Set unmapped entries to a scratch page instead of 0.
664 * Any prefetches that hit unmapped entries won't get an bus abort
667 scratch = get_zeroed_page(GFP_KERNEL);
669 panic("Cannot allocate iommu scratch page");
670 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
671 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
672 iommu_gatt_base[i] = gart_unmapped_entry;
675 dma_ops = &gart_dma_ops;
678 void gart_parse_options(char *p)
682 #ifdef CONFIG_IOMMU_LEAK
683 if (!strncmp(p,"leak",4)) {
687 if (isdigit(*p) && get_option(&p, &arg))
688 iommu_leak_pages = arg;
691 if (isdigit(*p) && get_option(&p, &arg))
693 if (!strncmp(p, "fullflush",8))
695 if (!strncmp(p, "nofullflush",11))
697 if (!strncmp(p,"noagp",5))
699 if (!strncmp(p, "noaperture",10))
701 /* duplicated from pci-dma.c */
702 if (!strncmp(p,"force",5))
703 iommu_aperture_allowed = 1;
704 if (!strncmp(p,"allowed",7))
705 iommu_aperture_allowed = 1;
706 if (!strncmp(p, "memaper", 7)) {
707 fallback_aper_force = 1;
711 if (get_option(&p, &arg))
712 fallback_aper_order = arg;