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, unsigned gfp, unsigned order)
194 if (dev->bus == &pci_bus_type) {
196 mask = pcibus_to_cpumask(to_pci_dev(dev)->bus);
197 node = cpu_to_node(first_cpu(mask));
199 node = numa_node_id();
200 page = alloc_pages_node(node, gfp, order);
201 return page ? page_address(page) : NULL;
205 * Allocate memory for a coherent mapping.
208 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
212 unsigned long dma_mask = 0;
217 dma_mask = dev->coherent_dma_mask;
219 dma_mask = 0xffffffff;
221 /* Kludge to make it bug-to-bug compatible with i386. i386
222 uses the normal dma_mask for alloc_coherent. */
223 dma_mask &= *dev->dma_mask;
226 memory = dma_alloc_pages(dev, gfp, get_order(size));
232 bus = virt_to_bus(memory);
233 high = (bus + size) >= dma_mask;
235 if (force_iommu && !(gfp & GFP_DMA))
237 if (no_iommu || dma_mask < 0xffffffffUL) {
239 free_pages((unsigned long)memory,
244 swiotlb_alloc_coherent(dev, size,
249 if (!(gfp & GFP_DMA)) {
257 memset(memory, 0, size);
259 *dma_handle = virt_to_bus(memory);
264 *dma_handle = dma_map_area(dev, bus, size, PCI_DMA_BIDIRECTIONAL, 0);
265 if (*dma_handle == bad_dma_address)
271 if (panic_on_overflow)
272 panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n", size);
273 free_pages((unsigned long)memory, get_order(size));
278 * Unmap coherent memory.
279 * The caller must ensure that the device has finished accessing the mapping.
281 void dma_free_coherent(struct device *dev, size_t size,
282 void *vaddr, dma_addr_t bus)
285 swiotlb_free_coherent(dev, size, vaddr, bus);
289 dma_unmap_single(dev, bus, size, 0);
290 free_pages((unsigned long)vaddr, get_order(size));
293 #ifdef CONFIG_IOMMU_LEAK
295 #define SET_LEAK(x) if (iommu_leak_tab) \
296 iommu_leak_tab[x] = __builtin_return_address(0);
297 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
298 iommu_leak_tab[x] = NULL;
300 /* Debugging aid for drivers that don't free their IOMMU tables */
301 static void **iommu_leak_tab;
302 static int leak_trace;
303 int iommu_leak_pages = 20;
308 if (dump || !iommu_leak_tab) return;
310 show_stack(NULL,NULL);
311 /* Very crude. dump some from the end of the table too */
312 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
313 for (i = 0; i < iommu_leak_pages; i+=2) {
314 printk("%lu: ", iommu_pages-i);
315 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
316 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
322 #define CLEAR_LEAK(x)
325 static void iommu_full(struct device *dev, size_t size, int dir, int do_panic)
328 * Ran out of IOMMU space for this operation. This is very bad.
329 * Unfortunately the drivers cannot handle this operation properly.
330 * Return some non mapped prereserved space in the aperture and
331 * let the Northbridge deal with it. This will result in garbage
332 * in the IO operation. When the size exceeds the prereserved space
333 * memory corruption will occur or random memory will be DMAed
334 * out. Hopefully no network devices use single mappings that big.
338 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
341 if (size > PAGE_SIZE*EMERGENCY_PAGES && do_panic) {
342 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
343 panic("PCI-DMA: Memory would be corrupted\n");
344 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
345 panic("PCI-DMA: Random memory would be DMAed\n");
348 #ifdef CONFIG_IOMMU_LEAK
353 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
355 u64 mask = *dev->dma_mask;
356 int high = addr + size >= mask;
362 panic("PCI-DMA: high address but no IOMMU.\n");
368 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
370 u64 mask = *dev->dma_mask;
371 int high = addr + size >= mask;
375 panic("PCI-DMA: high address but no IOMMU.\n");
381 /* Map a single continuous physical area into the IOMMU.
382 * Caller needs to check if the iommu is needed and flush.
384 static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
385 size_t size, int dir, int do_panic)
387 unsigned long npages = to_pages(phys_mem, size);
388 unsigned long iommu_page = alloc_iommu(npages);
390 if (iommu_page == -1) {
391 if (!nonforced_iommu(dev, phys_mem, size))
393 if (panic_on_overflow)
394 panic("dma_map_area overflow %lu bytes\n", size);
395 iommu_full(dev, size, dir, do_panic);
396 return bad_dma_address;
399 for (i = 0; i < npages; i++) {
400 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
401 SET_LEAK(iommu_page + i);
402 phys_mem += PAGE_SIZE;
404 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
407 /* Map a single area into the IOMMU */
408 dma_addr_t dma_map_single(struct device *dev, void *addr, size_t size, int dir)
410 unsigned long phys_mem, bus;
412 BUG_ON(dir == DMA_NONE);
415 return swiotlb_map_single(dev,addr,size,dir);
419 phys_mem = virt_to_phys(addr);
420 if (!need_iommu(dev, phys_mem, size))
423 bus = dma_map_area(dev, phys_mem, size, dir, 1);
428 /* Fallback for dma_map_sg in case of overflow */
429 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
434 #ifdef CONFIG_IOMMU_DEBUG
435 printk(KERN_DEBUG "dma_map_sg overflow\n");
438 for (i = 0; i < nents; i++ ) {
439 struct scatterlist *s = &sg[i];
440 unsigned long addr = page_to_phys(s->page) + s->offset;
441 if (nonforced_iommu(dev, addr, s->length)) {
442 addr = dma_map_area(dev, addr, s->length, dir, 0);
443 if (addr == bad_dma_address) {
445 dma_unmap_sg(dev, sg, i, dir);
447 sg[0].dma_length = 0;
451 s->dma_address = addr;
452 s->dma_length = s->length;
458 /* Map multiple scatterlist entries continuous into the first. */
459 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
460 struct scatterlist *sout, unsigned long pages)
462 unsigned long iommu_start = alloc_iommu(pages);
463 unsigned long iommu_page = iommu_start;
466 if (iommu_start == -1)
469 for (i = start; i < stopat; i++) {
470 struct scatterlist *s = &sg[i];
471 unsigned long pages, addr;
472 unsigned long phys_addr = s->dma_address;
474 BUG_ON(i > start && s->offset);
477 sout->dma_address = iommu_bus_base;
478 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
479 sout->dma_length = s->length;
481 sout->dma_length += s->length;
485 pages = to_pages(s->offset, s->length);
487 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
488 SET_LEAK(iommu_page);
493 BUG_ON(iommu_page - iommu_start != pages);
497 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
498 struct scatterlist *sout,
499 unsigned long pages, int need)
502 BUG_ON(stopat - start != 1);
504 sout->dma_length = sg[start].length;
507 return __dma_map_cont(sg, start, stopat, sout, pages);
511 * DMA map all entries in a scatterlist.
512 * Merge chunks that have page aligned sizes into a continuous mapping.
514 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
519 unsigned long pages = 0;
520 int need = 0, nextneed;
522 BUG_ON(dir == DMA_NONE);
527 return swiotlb_map_sg(dev,sg,nents,dir);
533 for (i = 0; i < nents; i++) {
534 struct scatterlist *s = &sg[i];
535 dma_addr_t addr = page_to_phys(s->page) + s->offset;
536 s->dma_address = addr;
537 BUG_ON(s->length == 0);
539 nextneed = need_iommu(dev, addr, s->length);
541 /* Handle the previous not yet processed entries */
543 struct scatterlist *ps = &sg[i-1];
544 /* Can only merge when the last chunk ends on a page
545 boundary and the new one doesn't have an offset. */
546 if (!iommu_merge || !nextneed || !need || s->offset ||
547 (ps->offset + ps->length) % PAGE_SIZE) {
548 if (dma_map_cont(sg, start, i, sg+out, pages,
558 pages += to_pages(s->offset, s->length);
560 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
565 sg[out].dma_length = 0;
570 dma_unmap_sg(dev, sg, nents, dir);
571 /* When it was forced try again unforced */
573 return dma_map_sg_nonforce(dev, sg, nents, dir);
574 if (panic_on_overflow)
575 panic("dma_map_sg: overflow on %lu pages\n", pages);
576 iommu_full(dev, pages << PAGE_SHIFT, dir, 0);
577 for (i = 0; i < nents; i++)
578 sg[i].dma_address = bad_dma_address;
583 * Free a DMA mapping.
585 void dma_unmap_single(struct device *dev, dma_addr_t dma_addr,
586 size_t size, int direction)
588 unsigned long iommu_page;
593 swiotlb_unmap_single(dev,dma_addr,size,direction);
597 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
598 dma_addr >= iommu_bus_base + iommu_size)
600 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
601 npages = to_pages(dma_addr, size);
602 for (i = 0; i < npages; i++) {
603 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
604 CLEAR_LEAK(iommu_page + i);
606 free_iommu(iommu_page, npages);
610 * Wrapper for pci_unmap_single working with scatterlists.
612 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
616 swiotlb_unmap_sg(dev,sg,nents,dir);
619 for (i = 0; i < nents; i++) {
620 struct scatterlist *s = &sg[i];
621 if (!s->dma_length || !s->length)
623 dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
627 int dma_supported(struct device *dev, u64 mask)
629 /* Copied from i386. Doesn't make much sense, because it will
630 only work for pci_alloc_coherent.
631 The caller just has to use GFP_DMA in this case. */
632 if (mask < 0x00ffffff)
635 /* Tell the device to use SAC when IOMMU force is on.
636 This allows the driver to use cheaper accesses in some cases.
638 Problem with this is that if we overflow the IOMMU area
639 and return DAC as fallback address the device may not handle it correctly.
641 As a special case some controllers have a 39bit address mode
642 that is as efficient as 32bit (aic79xx). Don't force SAC for these.
643 Assume all masks <= 40 bits are of this type. Normally this doesn't
644 make any difference, but gives more gentle handling of IOMMU overflow. */
645 if (iommu_sac_force && (mask >= 0xffffffffffULL)) {
646 printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
653 int dma_get_cache_alignment(void)
655 return boot_cpu_data.x86_clflush_size;
658 EXPORT_SYMBOL(dma_unmap_sg);
659 EXPORT_SYMBOL(dma_map_sg);
660 EXPORT_SYMBOL(dma_map_single);
661 EXPORT_SYMBOL(dma_unmap_single);
662 EXPORT_SYMBOL(dma_supported);
663 EXPORT_SYMBOL(no_iommu);
664 EXPORT_SYMBOL(force_iommu);
665 EXPORT_SYMBOL(bad_dma_address);
666 EXPORT_SYMBOL(iommu_bio_merge);
667 EXPORT_SYMBOL(iommu_sac_force);
668 EXPORT_SYMBOL(dma_get_cache_alignment);
669 EXPORT_SYMBOL(dma_alloc_coherent);
670 EXPORT_SYMBOL(dma_free_coherent);
672 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
676 iommu_size = aper_size;
681 a = aper + iommu_size;
682 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
684 if (iommu_size < 64*1024*1024)
686 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
691 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
693 unsigned aper_size = 0, aper_base_32;
697 pci_read_config_dword(dev, 0x94, &aper_base_32);
698 pci_read_config_dword(dev, 0x90, &aper_order);
699 aper_order = (aper_order >> 1) & 7;
701 aper_base = aper_base_32 & 0x7fff;
704 aper_size = (32 * 1024 * 1024) << aper_order;
705 if (aper_base + aper_size >= 0xffffffff || !aper_size)
713 * Private Northbridge GATT initialization in case we cannot use the
714 * AGP driver for some reason.
716 static __init int init_k8_gatt(struct agp_kern_info *info)
720 unsigned aper_base, new_aper_base;
721 unsigned aper_size, gatt_size, new_aper_size;
723 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
724 aper_size = aper_base = info->aper_size = 0;
726 new_aper_base = read_aperture(dev, &new_aper_size);
731 aper_size = new_aper_size;
732 aper_base = new_aper_base;
734 if (aper_size != new_aper_size || aper_base != new_aper_base)
739 info->aper_base = aper_base;
740 info->aper_size = aper_size>>20;
742 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
743 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
745 panic("Cannot allocate GATT table");
746 memset(gatt, 0, gatt_size);
747 agp_gatt_table = gatt;
753 gatt_reg = __pa(gatt) >> 12;
755 pci_write_config_dword(dev, 0x98, gatt_reg);
756 pci_read_config_dword(dev, 0x90, &ctl);
759 ctl &= ~((1<<4) | (1<<5));
761 pci_write_config_dword(dev, 0x90, ctl);
765 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
769 /* Should not happen anymore */
770 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
771 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.");
775 extern int agp_amd64_init(void);
777 static int __init pci_iommu_init(void)
779 struct agp_kern_info info;
780 unsigned long aper_size;
781 unsigned long iommu_start;
783 unsigned long scratch;
786 #ifndef CONFIG_AGP_AMD64
789 /* Makefile puts PCI initialization via subsys_initcall first. */
790 /* Add other K8 AGP bridge drivers here */
792 (agp_amd64_init() < 0) ||
793 (agp_copy_info(agp_bridge, &info) < 0);
798 printk(KERN_INFO "PCI-DMA: Using software bounce buffering for IO (SWIOTLB)\n");
803 (!force_iommu && end_pfn < 0xffffffff>>PAGE_SHIFT) ||
805 (no_agp && init_k8_gatt(&info) < 0)) {
806 printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n");
811 aper_size = info.aper_size * 1024 * 1024;
812 iommu_size = check_iommu_size(info.aper_base, aper_size);
813 iommu_pages = iommu_size >> PAGE_SHIFT;
815 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
816 get_order(iommu_pages/8));
817 if (!iommu_gart_bitmap)
818 panic("Cannot allocate iommu bitmap\n");
819 memset(iommu_gart_bitmap, 0, iommu_pages/8);
821 #ifdef CONFIG_IOMMU_LEAK
823 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
824 get_order(iommu_pages*sizeof(void *)));
826 memset(iommu_leak_tab, 0, iommu_pages * 8);
828 printk("PCI-DMA: Cannot allocate leak trace area\n");
833 * Out of IOMMU space handling.
834 * Reserve some invalid pages at the beginning of the GART.
836 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
838 agp_memory_reserved = iommu_size;
840 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
843 iommu_start = aper_size - iommu_size;
844 iommu_bus_base = info.aper_base + iommu_start;
845 bad_dma_address = iommu_bus_base;
846 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
849 * Unmap the IOMMU part of the GART. The alias of the page is
850 * always mapped with cache enabled and there is no full cache
851 * coherency across the GART remapping. The unmapping avoids
852 * automatic prefetches from the CPU allocating cache lines in
853 * there. All CPU accesses are done via the direct mapping to
854 * the backing memory. The GART address is only used by PCI
857 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
860 * Try to workaround a bug (thanks to BenH)
861 * Set unmapped entries to a scratch page instead of 0.
862 * Any prefetches that hit unmapped entries won't get an bus abort
865 scratch = get_zeroed_page(GFP_KERNEL);
867 panic("Cannot allocate iommu scratch page");
868 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
869 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
870 iommu_gatt_base[i] = gart_unmapped_entry;
874 int cpu = PCI_SLOT(dev->devfn) - 24;
877 northbridges[cpu] = dev;
878 pci_read_config_dword(dev, 0x9c, &flag); /* cache flush word */
879 northbridge_flush_word[cpu] = flag;
887 /* Must execute after PCI subsystem */
888 fs_initcall(pci_iommu_init);
890 /* iommu=[size][,noagp][,off][,force][,noforce][,leak][,memaper[=order]][,merge]
891 [,forcesac][,fullflush][,nomerge][,biomerge]
892 size set size of iommu (in bytes)
893 noagp don't initialize the AGP driver and use full aperture.
894 off don't use the IOMMU
895 leak turn on simple iommu leak tracing (only when CONFIG_IOMMU_LEAK is on)
896 memaper[=order] allocate an own aperture over RAM with size 32MB^order.
897 noforce don't force IOMMU usage. Default.
899 merge Do lazy merging. This may improve performance on some block devices.
900 Implies force (experimental)
901 biomerge Do merging at the BIO layer. This is more efficient than merge,
902 but should be only done with very big IOMMUs. Implies merge,force.
903 nomerge Don't do SG merging.
904 forcesac For SAC mode for masks <40bits (experimental)
905 fullflush Flush IOMMU on each allocation (default)
906 nofullflush Don't use IOMMU fullflush
907 allowed overwrite iommu off workarounds for specific chipsets.
908 soft Use software bounce buffering (default for Intel machines)
909 noaperture Don't touch the aperture for AGP.
911 __init int iommu_setup(char *p)
916 if (!strncmp(p,"noagp",5))
918 if (!strncmp(p,"off",3))
920 if (!strncmp(p,"force",5)) {
922 iommu_aperture_allowed = 1;
924 if (!strncmp(p,"allowed",7))
925 iommu_aperture_allowed = 1;
926 if (!strncmp(p,"noforce",7)) {
930 if (!strncmp(p, "memaper", 7)) {
931 fallback_aper_force = 1;
935 if (get_option(&p, &arg))
936 fallback_aper_order = arg;
939 if (!strncmp(p, "biomerge",8)) {
940 iommu_bio_merge = 4096;
944 if (!strncmp(p, "panic",5))
945 panic_on_overflow = 1;
946 if (!strncmp(p, "nopanic",7))
947 panic_on_overflow = 0;
948 if (!strncmp(p, "merge",5)) {
952 if (!strncmp(p, "nomerge",7))
954 if (!strncmp(p, "forcesac",8))
956 if (!strncmp(p, "fullflush",8))
958 if (!strncmp(p, "nofullflush",11))
960 if (!strncmp(p, "soft",4))
962 if (!strncmp(p, "noaperture",10))
964 #ifdef CONFIG_IOMMU_LEAK
965 if (!strncmp(p,"leak",4)) {
969 if (isdigit(*p) && get_option(&p, &arg))
970 iommu_leak_pages = arg;
973 if (isdigit(*p) && get_option(&p, &arg))
975 p += strcspn(p, ",");