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)
70 static struct pci_dev *northbridges[MAX_NB];
71 static u32 northbridge_flush_word[MAX_NB];
73 #define EMERGENCY_PAGES 32 /* = 128KB */
76 #define AGPEXTERN extern
81 /* backdoor interface to AGP driver */
82 AGPEXTERN int agp_memory_reserved;
83 AGPEXTERN __u32 *agp_gatt_table;
85 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
86 static int need_flush; /* global flush state. set for each gart wrap */
88 static unsigned long alloc_iommu(int size)
90 unsigned long offset, flags;
92 spin_lock_irqsave(&iommu_bitmap_lock, flags);
93 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
96 offset = find_next_zero_string(iommu_gart_bitmap,0,next_bit,size);
99 set_bit_string(iommu_gart_bitmap, offset, size);
100 next_bit = offset+size;
101 if (next_bit >= iommu_pages) {
108 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
112 static void free_iommu(unsigned long offset, int size)
115 spin_lock_irqsave(&iommu_bitmap_lock, flags);
116 __clear_bit_string(iommu_gart_bitmap, offset, size);
117 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
121 * Use global flush state to avoid races with multiple flushers.
123 static void flush_gart(struct device *dev)
129 spin_lock_irqsave(&iommu_bitmap_lock, flags);
132 for (i = 0; i < MAX_NB; i++) {
133 if (!northbridges[i])
135 pci_write_config_dword(northbridges[i], 0x9c,
136 northbridge_flush_word[i] | 1);
140 for (i = 0; i <= max; i++) {
142 if (!northbridges[i])
144 /* Make sure the hardware actually executed the flush. */
146 pci_read_config_dword(northbridges[i], 0x9c, &w);
153 printk("nothing to flush?\n");
156 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
161 #ifdef CONFIG_IOMMU_LEAK
163 #define SET_LEAK(x) if (iommu_leak_tab) \
164 iommu_leak_tab[x] = __builtin_return_address(0);
165 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
166 iommu_leak_tab[x] = NULL;
168 /* Debugging aid for drivers that don't free their IOMMU tables */
169 static void **iommu_leak_tab;
170 static int leak_trace;
171 int iommu_leak_pages = 20;
176 if (dump || !iommu_leak_tab) return;
178 show_stack(NULL,NULL);
179 /* Very crude. dump some from the end of the table too */
180 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
181 for (i = 0; i < iommu_leak_pages; i+=2) {
182 printk("%lu: ", iommu_pages-i);
183 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
184 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
190 #define CLEAR_LEAK(x)
193 static void iommu_full(struct device *dev, size_t size, int dir)
196 * Ran out of IOMMU space for this operation. This is very bad.
197 * Unfortunately the drivers cannot handle this operation properly.
198 * Return some non mapped prereserved space in the aperture and
199 * let the Northbridge deal with it. This will result in garbage
200 * in the IO operation. When the size exceeds the prereserved space
201 * memory corruption will occur or random memory will be DMAed
202 * out. Hopefully no network devices use single mappings that big.
206 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
209 if (size > PAGE_SIZE*EMERGENCY_PAGES) {
210 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
211 panic("PCI-DMA: Memory would be corrupted\n");
212 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
213 panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
216 #ifdef CONFIG_IOMMU_LEAK
221 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
223 u64 mask = *dev->dma_mask;
224 int high = addr + size >= mask;
231 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
233 u64 mask = *dev->dma_mask;
234 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(npages);
248 if (iommu_page == -1) {
249 if (!nonforced_iommu(dev, phys_mem, size))
251 if (panic_on_overflow)
252 panic("dma_map_area overflow %lu bytes\n", size);
253 iommu_full(dev, size, dir);
254 return bad_dma_address;
257 for (i = 0; i < npages; i++) {
258 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
259 SET_LEAK(iommu_page + i);
260 phys_mem += PAGE_SIZE;
262 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
265 static dma_addr_t gart_map_simple(struct device *dev, char *buf,
266 size_t size, int dir)
268 dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
273 /* Map a single area into the IOMMU */
274 dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
276 unsigned long phys_mem, bus;
278 BUG_ON(dir == DMA_NONE);
283 phys_mem = virt_to_phys(addr);
284 if (!need_iommu(dev, phys_mem, size))
287 bus = gart_map_simple(dev, addr, size, dir);
292 * Wrapper for pci_unmap_single working with scatterlists.
294 void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
298 for (i = 0; i < nents; i++) {
299 struct scatterlist *s = &sg[i];
300 if (!s->dma_length || !s->length)
302 dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
306 /* Fallback for dma_map_sg in case of overflow */
307 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
312 #ifdef CONFIG_IOMMU_DEBUG
313 printk(KERN_DEBUG "dma_map_sg overflow\n");
316 for (i = 0; i < nents; i++ ) {
317 struct scatterlist *s = &sg[i];
318 unsigned long addr = page_to_phys(s->page) + s->offset;
319 if (nonforced_iommu(dev, addr, s->length)) {
320 addr = dma_map_area(dev, addr, s->length, dir);
321 if (addr == bad_dma_address) {
323 gart_unmap_sg(dev, sg, i, dir);
325 sg[0].dma_length = 0;
329 s->dma_address = addr;
330 s->dma_length = s->length;
336 /* Map multiple scatterlist entries continuous into the first. */
337 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
338 struct scatterlist *sout, unsigned long pages)
340 unsigned long iommu_start = alloc_iommu(pages);
341 unsigned long iommu_page = iommu_start;
344 if (iommu_start == -1)
347 for (i = start; i < stopat; i++) {
348 struct scatterlist *s = &sg[i];
349 unsigned long pages, addr;
350 unsigned long phys_addr = s->dma_address;
352 BUG_ON(i > start && s->offset);
355 sout->dma_address = iommu_bus_base;
356 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
357 sout->dma_length = s->length;
359 sout->dma_length += s->length;
363 pages = to_pages(s->offset, s->length);
365 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
366 SET_LEAK(iommu_page);
371 BUG_ON(iommu_page - iommu_start != pages);
375 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
376 struct scatterlist *sout,
377 unsigned long pages, int need)
380 BUG_ON(stopat - start != 1);
382 sout->dma_length = sg[start].length;
385 return __dma_map_cont(sg, start, stopat, sout, pages);
389 * DMA map all entries in a scatterlist.
390 * Merge chunks that have page aligned sizes into a continuous mapping.
392 int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
397 unsigned long pages = 0;
398 int need = 0, nextneed;
400 BUG_ON(dir == DMA_NONE);
409 for (i = 0; i < nents; i++) {
410 struct scatterlist *s = &sg[i];
411 dma_addr_t addr = page_to_phys(s->page) + s->offset;
412 s->dma_address = addr;
413 BUG_ON(s->length == 0);
415 nextneed = need_iommu(dev, addr, s->length);
417 /* Handle the previous not yet processed entries */
419 struct scatterlist *ps = &sg[i-1];
420 /* Can only merge when the last chunk ends on a page
421 boundary and the new one doesn't have an offset. */
422 if (!iommu_merge || !nextneed || !need || s->offset ||
423 (ps->offset + ps->length) % PAGE_SIZE) {
424 if (dma_map_cont(sg, start, i, sg+out, pages,
434 pages += to_pages(s->offset, s->length);
436 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
441 sg[out].dma_length = 0;
446 gart_unmap_sg(dev, sg, nents, dir);
447 /* When it was forced or merged try again in a dumb way */
448 if (force_iommu || iommu_merge) {
449 out = dma_map_sg_nonforce(dev, sg, nents, dir);
453 if (panic_on_overflow)
454 panic("dma_map_sg: overflow on %lu pages\n", pages);
455 iommu_full(dev, pages << PAGE_SHIFT, dir);
456 for (i = 0; i < nents; i++)
457 sg[i].dma_address = bad_dma_address;
462 * Free a DMA mapping.
464 void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
465 size_t size, int direction)
467 unsigned long iommu_page;
471 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
472 dma_addr >= iommu_bus_base + iommu_size)
474 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
475 npages = to_pages(dma_addr, size);
476 for (i = 0; i < npages; i++) {
477 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
478 CLEAR_LEAK(iommu_page + i);
480 free_iommu(iommu_page, npages);
485 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
489 iommu_size = aper_size;
494 a = aper + iommu_size;
495 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
497 if (iommu_size < 64*1024*1024)
499 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
504 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
506 unsigned aper_size = 0, aper_base_32;
510 pci_read_config_dword(dev, 0x94, &aper_base_32);
511 pci_read_config_dword(dev, 0x90, &aper_order);
512 aper_order = (aper_order >> 1) & 7;
514 aper_base = aper_base_32 & 0x7fff;
517 aper_size = (32 * 1024 * 1024) << aper_order;
518 if (aper_base + aper_size >= 0xffffffff || !aper_size)
526 * Private Northbridge GATT initialization in case we cannot use the
527 * AGP driver for some reason.
529 static __init int init_k8_gatt(struct agp_kern_info *info)
533 unsigned aper_base, new_aper_base;
534 unsigned aper_size, gatt_size, new_aper_size;
536 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
537 aper_size = aper_base = info->aper_size = 0;
539 new_aper_base = read_aperture(dev, &new_aper_size);
544 aper_size = new_aper_size;
545 aper_base = new_aper_base;
547 if (aper_size != new_aper_size || aper_base != new_aper_base)
552 info->aper_base = aper_base;
553 info->aper_size = aper_size>>20;
555 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
556 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
558 panic("Cannot allocate GATT table");
559 memset(gatt, 0, gatt_size);
560 agp_gatt_table = gatt;
566 gatt_reg = __pa(gatt) >> 12;
568 pci_write_config_dword(dev, 0x98, gatt_reg);
569 pci_read_config_dword(dev, 0x90, &ctl);
572 ctl &= ~((1<<4) | (1<<5));
574 pci_write_config_dword(dev, 0x90, ctl);
578 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
582 /* Should not happen anymore */
583 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
584 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
588 extern int agp_amd64_init(void);
590 static struct dma_mapping_ops gart_dma_ops = {
591 .mapping_error = NULL,
592 .map_single = gart_map_single,
593 .map_simple = gart_map_simple,
594 .unmap_single = gart_unmap_single,
595 .sync_single_for_cpu = NULL,
596 .sync_single_for_device = NULL,
597 .sync_single_range_for_cpu = NULL,
598 .sync_single_range_for_device = NULL,
599 .sync_sg_for_cpu = NULL,
600 .sync_sg_for_device = NULL,
601 .map_sg = gart_map_sg,
602 .unmap_sg = gart_unmap_sg,
605 static int __init pci_iommu_init(void)
607 struct agp_kern_info info;
608 unsigned long aper_size;
609 unsigned long iommu_start;
611 unsigned long scratch;
614 #ifndef CONFIG_AGP_AMD64
617 /* Makefile puts PCI initialization via subsys_initcall first. */
618 /* Add other K8 AGP bridge drivers here */
620 (agp_amd64_init() < 0) ||
621 (agp_copy_info(agp_bridge, &info) < 0);
628 (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
630 (no_agp && init_k8_gatt(&info) < 0)) {
631 printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n");
632 if (end_pfn > MAX_DMA32_PFN) {
633 printk(KERN_ERR "WARNING more than 4GB of memory "
634 "but IOMMU not compiled in.\n"
635 KERN_ERR "WARNING 32bit PCI may malfunction.\n"
636 KERN_ERR "You might want to enable "
637 "CONFIG_GART_IOMMU\n");
646 printk(KERN_ERR "PCI-GART: Too many northbridges (%ld). Disabled\n", i);
650 printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
651 aper_size = info.aper_size * 1024 * 1024;
652 iommu_size = check_iommu_size(info.aper_base, aper_size);
653 iommu_pages = iommu_size >> PAGE_SHIFT;
655 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
656 get_order(iommu_pages/8));
657 if (!iommu_gart_bitmap)
658 panic("Cannot allocate iommu bitmap\n");
659 memset(iommu_gart_bitmap, 0, iommu_pages/8);
661 #ifdef CONFIG_IOMMU_LEAK
663 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
664 get_order(iommu_pages*sizeof(void *)));
666 memset(iommu_leak_tab, 0, iommu_pages * 8);
668 printk("PCI-DMA: Cannot allocate leak trace area\n");
673 * Out of IOMMU space handling.
674 * Reserve some invalid pages at the beginning of the GART.
676 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
678 agp_memory_reserved = iommu_size;
680 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
683 iommu_start = aper_size - iommu_size;
684 iommu_bus_base = info.aper_base + iommu_start;
685 bad_dma_address = iommu_bus_base;
686 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
689 * Unmap the IOMMU part of the GART. The alias of the page is
690 * always mapped with cache enabled and there is no full cache
691 * coherency across the GART remapping. The unmapping avoids
692 * automatic prefetches from the CPU allocating cache lines in
693 * there. All CPU accesses are done via the direct mapping to
694 * the backing memory. The GART address is only used by PCI
697 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
700 * Try to workaround a bug (thanks to BenH)
701 * Set unmapped entries to a scratch page instead of 0.
702 * Any prefetches that hit unmapped entries won't get an bus abort
705 scratch = get_zeroed_page(GFP_KERNEL);
707 panic("Cannot allocate iommu scratch page");
708 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
709 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
710 iommu_gatt_base[i] = gart_unmapped_entry;
714 int cpu = PCI_SLOT(dev->devfn) - 24;
717 northbridges[cpu] = dev;
718 pci_read_config_dword(dev, 0x9c, &flag); /* cache flush word */
719 northbridge_flush_word[cpu] = flag;
724 dma_ops = &gart_dma_ops;
729 /* Must execute after PCI subsystem */
730 fs_initcall(pci_iommu_init);
732 void gart_parse_options(char *p)
736 #ifdef CONFIG_IOMMU_LEAK
737 if (!strncmp(p,"leak",4)) {
741 if (isdigit(*p) && get_option(&p, &arg))
742 iommu_leak_pages = arg;
745 if (isdigit(*p) && get_option(&p, &arg))
747 if (!strncmp(p, "fullflush",8))
749 if (!strncmp(p, "nofullflush",11))
751 if (!strncmp(p,"noagp",5))
753 if (!strncmp(p, "noaperture",10))
755 /* duplicated from pci-dma.c */
756 if (!strncmp(p,"force",5))
757 iommu_aperture_allowed = 1;
758 if (!strncmp(p,"allowed",7))
759 iommu_aperture_allowed = 1;
760 if (!strncmp(p, "memaper", 7)) {
761 fallback_aper_force = 1;
765 if (get_option(&p, &arg))
766 fallback_aper_order = arg;