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>
34 dma_addr_t bad_dma_address;
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 */
44 #ifdef CONFIG_IOMMU_DEBUG
45 int panic_on_overflow = 1;
48 int panic_on_overflow = 0;
52 int iommu_sac_force = 0;
54 /* If this is disabled the IOMMU will use an optimized flushing strategy
55 of only flushing when an mapping is reused. With it true the GART is flushed
56 for every mapping. Problem is that doing the lazy flush seems to trigger
57 bugs with some popular PCI cards, in particular 3ware (but has been also
58 also seen with Qlogic at least). */
59 int iommu_fullflush = 1;
61 /* This tells the BIO block layer to assume merging. Default to off
62 because we cannot guarantee merging later. */
63 int iommu_bio_merge = 0;
67 /* Allocation bitmap for the remapping area */
68 static DEFINE_SPINLOCK(iommu_bitmap_lock);
69 static unsigned long *iommu_gart_bitmap; /* guarded by iommu_bitmap_lock */
71 static u32 gart_unmapped_entry;
74 #define GPTE_COHERENT 2
75 #define GPTE_ENCODE(x) \
76 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
77 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
79 #define to_pages(addr,size) \
80 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
82 #define for_all_nb(dev) \
84 while ((dev = pci_get_device(PCI_VENDOR_ID_AMD, 0x1103, dev))!=NULL)\
85 if (dev->bus->number == 0 && \
86 (PCI_SLOT(dev->devfn) >= 24) && (PCI_SLOT(dev->devfn) <= 31))
88 static struct pci_dev *northbridges[MAX_NB];
89 static u32 northbridge_flush_word[MAX_NB];
91 #define EMERGENCY_PAGES 32 /* = 128KB */
94 #define AGPEXTERN extern
99 /* backdoor interface to AGP driver */
100 AGPEXTERN int agp_memory_reserved;
101 AGPEXTERN __u32 *agp_gatt_table;
103 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
104 static int need_flush; /* global flush state. set for each gart wrap */
105 static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
106 size_t size, int dir, int do_panic);
108 /* Dummy device used for NULL arguments (normally ISA). Better would
109 be probably a smaller DMA mask, but this is bug-to-bug compatible to i386. */
110 static struct device fallback_dev = {
111 .bus_id = "fallback device",
112 .coherent_dma_mask = 0xffffffff,
113 .dma_mask = &fallback_dev.coherent_dma_mask,
116 static unsigned long alloc_iommu(int size)
118 unsigned long offset, flags;
120 spin_lock_irqsave(&iommu_bitmap_lock, flags);
121 offset = find_next_zero_string(iommu_gart_bitmap,next_bit,iommu_pages,size);
124 offset = find_next_zero_string(iommu_gart_bitmap,0,next_bit,size);
127 set_bit_string(iommu_gart_bitmap, offset, size);
128 next_bit = offset+size;
129 if (next_bit >= iommu_pages) {
136 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
140 static void free_iommu(unsigned long offset, int size)
144 clear_bit(offset, iommu_gart_bitmap);
147 spin_lock_irqsave(&iommu_bitmap_lock, flags);
148 __clear_bit_string(iommu_gart_bitmap, offset, size);
149 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
153 * Use global flush state to avoid races with multiple flushers.
155 static void flush_gart(struct device *dev)
161 spin_lock_irqsave(&iommu_bitmap_lock, flags);
164 for (i = 0; i < MAX_NB; i++) {
165 if (!northbridges[i])
167 pci_write_config_dword(northbridges[i], 0x9c,
168 northbridge_flush_word[i] | 1);
172 for (i = 0; i <= max; i++) {
174 if (!northbridges[i])
176 /* Make sure the hardware actually executed the flush. */
178 pci_read_config_dword(northbridges[i], 0x9c, &w);
182 printk("nothing to flush?\n");
185 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
188 /* Allocate DMA memory on node near device */
190 static void *dma_alloc_pages(struct device *dev, gfp_t gfp, unsigned order)
194 if (dev->bus == &pci_bus_type)
195 node = pcibus_to_node(to_pci_dev(dev)->bus);
197 node = numa_node_id();
198 page = alloc_pages_node(node, gfp, order);
199 return page ? page_address(page) : NULL;
203 * Allocate memory for a coherent mapping.
206 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
210 unsigned long dma_mask = 0;
215 dma_mask = dev->coherent_dma_mask;
217 dma_mask = 0xffffffff;
219 /* Kludge to make it bug-to-bug compatible with i386. i386
220 uses the normal dma_mask for alloc_coherent. */
221 dma_mask &= *dev->dma_mask;
224 memory = dma_alloc_pages(dev, gfp, get_order(size));
230 bus = virt_to_bus(memory);
231 high = (bus + size) >= dma_mask;
233 if (force_iommu && !(gfp & GFP_DMA))
235 if (no_iommu || dma_mask < 0xffffffffUL) {
237 free_pages((unsigned long)memory,
242 swiotlb_alloc_coherent(dev, size,
247 if (!(gfp & GFP_DMA)) {
255 memset(memory, 0, size);
257 *dma_handle = virt_to_bus(memory);
262 *dma_handle = dma_map_area(dev, bus, size, PCI_DMA_BIDIRECTIONAL, 0);
263 if (*dma_handle == bad_dma_address)
269 if (panic_on_overflow)
270 panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n", size);
271 free_pages((unsigned long)memory, get_order(size));
276 * Unmap coherent memory.
277 * The caller must ensure that the device has finished accessing the mapping.
279 void dma_free_coherent(struct device *dev, size_t size,
280 void *vaddr, dma_addr_t bus)
283 swiotlb_free_coherent(dev, size, vaddr, bus);
287 dma_unmap_single(dev, bus, size, 0);
288 free_pages((unsigned long)vaddr, get_order(size));
291 #ifdef CONFIG_IOMMU_LEAK
293 #define SET_LEAK(x) if (iommu_leak_tab) \
294 iommu_leak_tab[x] = __builtin_return_address(0);
295 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
296 iommu_leak_tab[x] = NULL;
298 /* Debugging aid for drivers that don't free their IOMMU tables */
299 static void **iommu_leak_tab;
300 static int leak_trace;
301 int iommu_leak_pages = 20;
306 if (dump || !iommu_leak_tab) return;
308 show_stack(NULL,NULL);
309 /* Very crude. dump some from the end of the table too */
310 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
311 for (i = 0; i < iommu_leak_pages; i+=2) {
312 printk("%lu: ", iommu_pages-i);
313 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
314 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
320 #define CLEAR_LEAK(x)
323 static void iommu_full(struct device *dev, size_t size, int dir, int do_panic)
326 * Ran out of IOMMU space for this operation. This is very bad.
327 * Unfortunately the drivers cannot handle this operation properly.
328 * Return some non mapped prereserved space in the aperture and
329 * let the Northbridge deal with it. This will result in garbage
330 * in the IO operation. When the size exceeds the prereserved space
331 * memory corruption will occur or random memory will be DMAed
332 * out. Hopefully no network devices use single mappings that big.
336 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
339 if (size > PAGE_SIZE*EMERGENCY_PAGES && do_panic) {
340 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
341 panic("PCI-DMA: Memory would be corrupted\n");
342 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
343 panic("PCI-DMA: Random memory would be DMAed\n");
346 #ifdef CONFIG_IOMMU_LEAK
351 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
353 u64 mask = *dev->dma_mask;
354 int high = addr + size >= mask;
360 panic("PCI-DMA: high address but no IOMMU.\n");
366 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
368 u64 mask = *dev->dma_mask;
369 int high = addr + size >= mask;
373 panic("PCI-DMA: high address but no IOMMU.\n");
379 /* Map a single continuous physical area into the IOMMU.
380 * Caller needs to check if the iommu is needed and flush.
382 static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
383 size_t size, int dir, int do_panic)
385 unsigned long npages = to_pages(phys_mem, size);
386 unsigned long iommu_page = alloc_iommu(npages);
388 if (iommu_page == -1) {
389 if (!nonforced_iommu(dev, phys_mem, size))
391 if (panic_on_overflow)
392 panic("dma_map_area overflow %lu bytes\n", size);
393 iommu_full(dev, size, dir, do_panic);
394 return bad_dma_address;
397 for (i = 0; i < npages; i++) {
398 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
399 SET_LEAK(iommu_page + i);
400 phys_mem += PAGE_SIZE;
402 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
405 /* Map a single area into the IOMMU */
406 dma_addr_t dma_map_single(struct device *dev, void *addr, size_t size, int dir)
408 unsigned long phys_mem, bus;
410 BUG_ON(dir == DMA_NONE);
413 return swiotlb_map_single(dev,addr,size,dir);
417 phys_mem = virt_to_phys(addr);
418 if (!need_iommu(dev, phys_mem, size))
421 bus = dma_map_area(dev, phys_mem, size, dir, 1);
426 /* Fallback for dma_map_sg in case of overflow */
427 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
432 #ifdef CONFIG_IOMMU_DEBUG
433 printk(KERN_DEBUG "dma_map_sg overflow\n");
436 for (i = 0; i < nents; i++ ) {
437 struct scatterlist *s = &sg[i];
438 unsigned long addr = page_to_phys(s->page) + s->offset;
439 if (nonforced_iommu(dev, addr, s->length)) {
440 addr = dma_map_area(dev, addr, s->length, dir, 0);
441 if (addr == bad_dma_address) {
443 dma_unmap_sg(dev, sg, i, dir);
445 sg[0].dma_length = 0;
449 s->dma_address = addr;
450 s->dma_length = s->length;
456 /* Map multiple scatterlist entries continuous into the first. */
457 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
458 struct scatterlist *sout, unsigned long pages)
460 unsigned long iommu_start = alloc_iommu(pages);
461 unsigned long iommu_page = iommu_start;
464 if (iommu_start == -1)
467 for (i = start; i < stopat; i++) {
468 struct scatterlist *s = &sg[i];
469 unsigned long pages, addr;
470 unsigned long phys_addr = s->dma_address;
472 BUG_ON(i > start && s->offset);
475 sout->dma_address = iommu_bus_base;
476 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
477 sout->dma_length = s->length;
479 sout->dma_length += s->length;
483 pages = to_pages(s->offset, s->length);
485 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
486 SET_LEAK(iommu_page);
491 BUG_ON(iommu_page - iommu_start != pages);
495 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
496 struct scatterlist *sout,
497 unsigned long pages, int need)
500 BUG_ON(stopat - start != 1);
502 sout->dma_length = sg[start].length;
505 return __dma_map_cont(sg, start, stopat, sout, pages);
509 * DMA map all entries in a scatterlist.
510 * Merge chunks that have page aligned sizes into a continuous mapping.
512 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
517 unsigned long pages = 0;
518 int need = 0, nextneed;
520 BUG_ON(dir == DMA_NONE);
525 return swiotlb_map_sg(dev,sg,nents,dir);
531 for (i = 0; i < nents; i++) {
532 struct scatterlist *s = &sg[i];
533 dma_addr_t addr = page_to_phys(s->page) + s->offset;
534 s->dma_address = addr;
535 BUG_ON(s->length == 0);
537 nextneed = need_iommu(dev, addr, s->length);
539 /* Handle the previous not yet processed entries */
541 struct scatterlist *ps = &sg[i-1];
542 /* Can only merge when the last chunk ends on a page
543 boundary and the new one doesn't have an offset. */
544 if (!iommu_merge || !nextneed || !need || s->offset ||
545 (ps->offset + ps->length) % PAGE_SIZE) {
546 if (dma_map_cont(sg, start, i, sg+out, pages,
556 pages += to_pages(s->offset, s->length);
558 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
563 sg[out].dma_length = 0;
568 dma_unmap_sg(dev, sg, nents, dir);
569 /* When it was forced try again unforced */
571 return dma_map_sg_nonforce(dev, sg, nents, dir);
572 if (panic_on_overflow)
573 panic("dma_map_sg: overflow on %lu pages\n", pages);
574 iommu_full(dev, pages << PAGE_SHIFT, dir, 0);
575 for (i = 0; i < nents; i++)
576 sg[i].dma_address = bad_dma_address;
581 * Free a DMA mapping.
583 void dma_unmap_single(struct device *dev, dma_addr_t dma_addr,
584 size_t size, int direction)
586 unsigned long iommu_page;
591 swiotlb_unmap_single(dev,dma_addr,size,direction);
595 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
596 dma_addr >= iommu_bus_base + iommu_size)
598 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
599 npages = to_pages(dma_addr, size);
600 for (i = 0; i < npages; i++) {
601 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
602 CLEAR_LEAK(iommu_page + i);
604 free_iommu(iommu_page, npages);
608 * Wrapper for pci_unmap_single working with scatterlists.
610 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
614 swiotlb_unmap_sg(dev,sg,nents,dir);
617 for (i = 0; i < nents; i++) {
618 struct scatterlist *s = &sg[i];
619 if (!s->dma_length || !s->length)
621 dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
625 int dma_supported(struct device *dev, u64 mask)
627 /* Copied from i386. Doesn't make much sense, because it will
628 only work for pci_alloc_coherent.
629 The caller just has to use GFP_DMA in this case. */
630 if (mask < 0x00ffffff)
633 /* Tell the device to use SAC when IOMMU force is on.
634 This allows the driver to use cheaper accesses in some cases.
636 Problem with this is that if we overflow the IOMMU area
637 and return DAC as fallback address the device may not handle it correctly.
639 As a special case some controllers have a 39bit address mode
640 that is as efficient as 32bit (aic79xx). Don't force SAC for these.
641 Assume all masks <= 40 bits are of this type. Normally this doesn't
642 make any difference, but gives more gentle handling of IOMMU overflow. */
643 if (iommu_sac_force && (mask >= 0xffffffffffULL)) {
644 printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
651 int dma_get_cache_alignment(void)
653 return boot_cpu_data.x86_clflush_size;
656 EXPORT_SYMBOL(dma_unmap_sg);
657 EXPORT_SYMBOL(dma_map_sg);
658 EXPORT_SYMBOL(dma_map_single);
659 EXPORT_SYMBOL(dma_unmap_single);
660 EXPORT_SYMBOL(dma_supported);
661 EXPORT_SYMBOL(no_iommu);
662 EXPORT_SYMBOL(force_iommu);
663 EXPORT_SYMBOL(bad_dma_address);
664 EXPORT_SYMBOL(iommu_bio_merge);
665 EXPORT_SYMBOL(iommu_sac_force);
666 EXPORT_SYMBOL(dma_get_cache_alignment);
667 EXPORT_SYMBOL(dma_alloc_coherent);
668 EXPORT_SYMBOL(dma_free_coherent);
670 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
674 iommu_size = aper_size;
679 a = aper + iommu_size;
680 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
682 if (iommu_size < 64*1024*1024)
684 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
689 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
691 unsigned aper_size = 0, aper_base_32;
695 pci_read_config_dword(dev, 0x94, &aper_base_32);
696 pci_read_config_dword(dev, 0x90, &aper_order);
697 aper_order = (aper_order >> 1) & 7;
699 aper_base = aper_base_32 & 0x7fff;
702 aper_size = (32 * 1024 * 1024) << aper_order;
703 if (aper_base + aper_size >= 0xffffffff || !aper_size)
711 * Private Northbridge GATT initialization in case we cannot use the
712 * AGP driver for some reason.
714 static __init int init_k8_gatt(struct agp_kern_info *info)
718 unsigned aper_base, new_aper_base;
719 unsigned aper_size, gatt_size, new_aper_size;
721 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
722 aper_size = aper_base = info->aper_size = 0;
724 new_aper_base = read_aperture(dev, &new_aper_size);
729 aper_size = new_aper_size;
730 aper_base = new_aper_base;
732 if (aper_size != new_aper_size || aper_base != new_aper_base)
737 info->aper_base = aper_base;
738 info->aper_size = aper_size>>20;
740 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
741 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
743 panic("Cannot allocate GATT table");
744 memset(gatt, 0, gatt_size);
745 agp_gatt_table = gatt;
751 gatt_reg = __pa(gatt) >> 12;
753 pci_write_config_dword(dev, 0x98, gatt_reg);
754 pci_read_config_dword(dev, 0x90, &ctl);
757 ctl &= ~((1<<4) | (1<<5));
759 pci_write_config_dword(dev, 0x90, ctl);
763 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
767 /* Should not happen anymore */
768 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
769 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.");
773 extern int agp_amd64_init(void);
775 static int __init pci_iommu_init(void)
777 struct agp_kern_info info;
778 unsigned long aper_size;
779 unsigned long iommu_start;
781 unsigned long scratch;
784 #ifndef CONFIG_AGP_AMD64
787 /* Makefile puts PCI initialization via subsys_initcall first. */
788 /* Add other K8 AGP bridge drivers here */
790 (agp_amd64_init() < 0) ||
791 (agp_copy_info(agp_bridge, &info) < 0);
796 printk(KERN_INFO "PCI-DMA: Using software bounce buffering for IO (SWIOTLB)\n");
801 (!force_iommu && end_pfn < 0xffffffff>>PAGE_SHIFT) ||
803 (no_agp && init_k8_gatt(&info) < 0)) {
804 printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n");
809 aper_size = info.aper_size * 1024 * 1024;
810 iommu_size = check_iommu_size(info.aper_base, aper_size);
811 iommu_pages = iommu_size >> PAGE_SHIFT;
813 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
814 get_order(iommu_pages/8));
815 if (!iommu_gart_bitmap)
816 panic("Cannot allocate iommu bitmap\n");
817 memset(iommu_gart_bitmap, 0, iommu_pages/8);
819 #ifdef CONFIG_IOMMU_LEAK
821 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
822 get_order(iommu_pages*sizeof(void *)));
824 memset(iommu_leak_tab, 0, iommu_pages * 8);
826 printk("PCI-DMA: Cannot allocate leak trace area\n");
831 * Out of IOMMU space handling.
832 * Reserve some invalid pages at the beginning of the GART.
834 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
836 agp_memory_reserved = iommu_size;
838 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
841 iommu_start = aper_size - iommu_size;
842 iommu_bus_base = info.aper_base + iommu_start;
843 bad_dma_address = iommu_bus_base;
844 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
847 * Unmap the IOMMU part of the GART. The alias of the page is
848 * always mapped with cache enabled and there is no full cache
849 * coherency across the GART remapping. The unmapping avoids
850 * automatic prefetches from the CPU allocating cache lines in
851 * there. All CPU accesses are done via the direct mapping to
852 * the backing memory. The GART address is only used by PCI
855 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
858 * Try to workaround a bug (thanks to BenH)
859 * Set unmapped entries to a scratch page instead of 0.
860 * Any prefetches that hit unmapped entries won't get an bus abort
863 scratch = get_zeroed_page(GFP_KERNEL);
865 panic("Cannot allocate iommu scratch page");
866 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
867 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
868 iommu_gatt_base[i] = gart_unmapped_entry;
872 int cpu = PCI_SLOT(dev->devfn) - 24;
875 northbridges[cpu] = dev;
876 pci_read_config_dword(dev, 0x9c, &flag); /* cache flush word */
877 northbridge_flush_word[cpu] = flag;
885 /* Must execute after PCI subsystem */
886 fs_initcall(pci_iommu_init);
888 /* iommu=[size][,noagp][,off][,force][,noforce][,leak][,memaper[=order]][,merge]
889 [,forcesac][,fullflush][,nomerge][,biomerge]
890 size set size of iommu (in bytes)
891 noagp don't initialize the AGP driver and use full aperture.
892 off don't use the IOMMU
893 leak turn on simple iommu leak tracing (only when CONFIG_IOMMU_LEAK is on)
894 memaper[=order] allocate an own aperture over RAM with size 32MB^order.
895 noforce don't force IOMMU usage. Default.
897 merge Do lazy merging. This may improve performance on some block devices.
898 Implies force (experimental)
899 biomerge Do merging at the BIO layer. This is more efficient than merge,
900 but should be only done with very big IOMMUs. Implies merge,force.
901 nomerge Don't do SG merging.
902 forcesac For SAC mode for masks <40bits (experimental)
903 fullflush Flush IOMMU on each allocation (default)
904 nofullflush Don't use IOMMU fullflush
905 allowed overwrite iommu off workarounds for specific chipsets.
906 soft Use software bounce buffering (default for Intel machines)
907 noaperture Don't touch the aperture for AGP.
909 __init int iommu_setup(char *p)
914 if (!strncmp(p,"noagp",5))
916 if (!strncmp(p,"off",3))
918 if (!strncmp(p,"force",5)) {
920 iommu_aperture_allowed = 1;
922 if (!strncmp(p,"allowed",7))
923 iommu_aperture_allowed = 1;
924 if (!strncmp(p,"noforce",7)) {
928 if (!strncmp(p, "memaper", 7)) {
929 fallback_aper_force = 1;
933 if (get_option(&p, &arg))
934 fallback_aper_order = arg;
937 if (!strncmp(p, "biomerge",8)) {
938 iommu_bio_merge = 4096;
942 if (!strncmp(p, "panic",5))
943 panic_on_overflow = 1;
944 if (!strncmp(p, "nopanic",7))
945 panic_on_overflow = 0;
946 if (!strncmp(p, "merge",5)) {
950 if (!strncmp(p, "nomerge",7))
952 if (!strncmp(p, "forcesac",8))
954 if (!strncmp(p, "fullflush",8))
956 if (!strncmp(p, "nofullflush",11))
958 if (!strncmp(p, "soft",4))
960 if (!strncmp(p, "noaperture",10))
962 #ifdef CONFIG_IOMMU_LEAK
963 if (!strncmp(p,"leak",4)) {
967 if (isdigit(*p) && get_option(&p, &arg))
968 iommu_leak_pages = arg;
971 if (isdigit(*p) && get_option(&p, &arg))
973 p += strcspn(p, ",");