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;
223 /* Why <=? Even when the mask is smaller than 4GB it is often larger
224 than 16MB and in this case we have a chance of finding fitting memory
225 in the next higher zone first. If not retry with true GFP_DMA. -AK */
226 if (dma_mask <= 0xffffffff)
230 memory = dma_alloc_pages(dev, gfp, get_order(size));
236 bus = virt_to_bus(memory);
237 high = (bus + size) >= dma_mask;
239 if (force_iommu && !(gfp & GFP_DMA))
241 if (no_iommu || dma_mask < 0xffffffffUL) {
243 free_pages((unsigned long)memory,
248 swiotlb_alloc_coherent(dev, size,
253 if (!(gfp & GFP_DMA)) {
254 gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
261 memset(memory, 0, size);
263 *dma_handle = virt_to_bus(memory);
268 *dma_handle = dma_map_area(dev, bus, size, PCI_DMA_BIDIRECTIONAL, 0);
269 if (*dma_handle == bad_dma_address)
275 if (panic_on_overflow)
276 panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n", size);
277 free_pages((unsigned long)memory, get_order(size));
282 * Unmap coherent memory.
283 * The caller must ensure that the device has finished accessing the mapping.
285 void dma_free_coherent(struct device *dev, size_t size,
286 void *vaddr, dma_addr_t bus)
289 swiotlb_free_coherent(dev, size, vaddr, bus);
293 dma_unmap_single(dev, bus, size, 0);
294 free_pages((unsigned long)vaddr, get_order(size));
297 #ifdef CONFIG_IOMMU_LEAK
299 #define SET_LEAK(x) if (iommu_leak_tab) \
300 iommu_leak_tab[x] = __builtin_return_address(0);
301 #define CLEAR_LEAK(x) if (iommu_leak_tab) \
302 iommu_leak_tab[x] = NULL;
304 /* Debugging aid for drivers that don't free their IOMMU tables */
305 static void **iommu_leak_tab;
306 static int leak_trace;
307 int iommu_leak_pages = 20;
312 if (dump || !iommu_leak_tab) return;
314 show_stack(NULL,NULL);
315 /* Very crude. dump some from the end of the table too */
316 printk("Dumping %d pages from end of IOMMU:\n", iommu_leak_pages);
317 for (i = 0; i < iommu_leak_pages; i+=2) {
318 printk("%lu: ", iommu_pages-i);
319 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i]);
320 printk("%c", (i+1)%2 == 0 ? '\n' : ' ');
326 #define CLEAR_LEAK(x)
329 static void iommu_full(struct device *dev, size_t size, int dir, int do_panic)
332 * Ran out of IOMMU space for this operation. This is very bad.
333 * Unfortunately the drivers cannot handle this operation properly.
334 * Return some non mapped prereserved space in the aperture and
335 * let the Northbridge deal with it. This will result in garbage
336 * in the IO operation. When the size exceeds the prereserved space
337 * memory corruption will occur or random memory will be DMAed
338 * out. Hopefully no network devices use single mappings that big.
342 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
345 if (size > PAGE_SIZE*EMERGENCY_PAGES && do_panic) {
346 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
347 panic("PCI-DMA: Memory would be corrupted\n");
348 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
349 panic("PCI-DMA: Random memory would be DMAed\n");
352 #ifdef CONFIG_IOMMU_LEAK
357 static inline int need_iommu(struct device *dev, unsigned long addr, size_t size)
359 u64 mask = *dev->dma_mask;
360 int high = addr + size >= mask;
366 panic("PCI-DMA: high address but no IOMMU.\n");
372 static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
374 u64 mask = *dev->dma_mask;
375 int high = addr + size >= mask;
379 panic("PCI-DMA: high address but no IOMMU.\n");
385 /* Map a single continuous physical area into the IOMMU.
386 * Caller needs to check if the iommu is needed and flush.
388 static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
389 size_t size, int dir, int do_panic)
391 unsigned long npages = to_pages(phys_mem, size);
392 unsigned long iommu_page = alloc_iommu(npages);
394 if (iommu_page == -1) {
395 if (!nonforced_iommu(dev, phys_mem, size))
397 if (panic_on_overflow)
398 panic("dma_map_area overflow %lu bytes\n", size);
399 iommu_full(dev, size, dir, do_panic);
400 return bad_dma_address;
403 for (i = 0; i < npages; i++) {
404 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
405 SET_LEAK(iommu_page + i);
406 phys_mem += PAGE_SIZE;
408 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
411 /* Map a single area into the IOMMU */
412 dma_addr_t dma_map_single(struct device *dev, void *addr, size_t size, int dir)
414 unsigned long phys_mem, bus;
416 BUG_ON(dir == DMA_NONE);
419 return swiotlb_map_single(dev,addr,size,dir);
423 phys_mem = virt_to_phys(addr);
424 if (!need_iommu(dev, phys_mem, size))
427 bus = dma_map_area(dev, phys_mem, size, dir, 1);
432 /* Fallback for dma_map_sg in case of overflow */
433 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
438 #ifdef CONFIG_IOMMU_DEBUG
439 printk(KERN_DEBUG "dma_map_sg overflow\n");
442 for (i = 0; i < nents; i++ ) {
443 struct scatterlist *s = &sg[i];
444 unsigned long addr = page_to_phys(s->page) + s->offset;
445 if (nonforced_iommu(dev, addr, s->length)) {
446 addr = dma_map_area(dev, addr, s->length, dir, 0);
447 if (addr == bad_dma_address) {
449 dma_unmap_sg(dev, sg, i, dir);
451 sg[0].dma_length = 0;
455 s->dma_address = addr;
456 s->dma_length = s->length;
462 /* Map multiple scatterlist entries continuous into the first. */
463 static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
464 struct scatterlist *sout, unsigned long pages)
466 unsigned long iommu_start = alloc_iommu(pages);
467 unsigned long iommu_page = iommu_start;
470 if (iommu_start == -1)
473 for (i = start; i < stopat; i++) {
474 struct scatterlist *s = &sg[i];
475 unsigned long pages, addr;
476 unsigned long phys_addr = s->dma_address;
478 BUG_ON(i > start && s->offset);
481 sout->dma_address = iommu_bus_base;
482 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
483 sout->dma_length = s->length;
485 sout->dma_length += s->length;
489 pages = to_pages(s->offset, s->length);
491 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
492 SET_LEAK(iommu_page);
497 BUG_ON(iommu_page - iommu_start != pages);
501 static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
502 struct scatterlist *sout,
503 unsigned long pages, int need)
506 BUG_ON(stopat - start != 1);
508 sout->dma_length = sg[start].length;
511 return __dma_map_cont(sg, start, stopat, sout, pages);
515 * DMA map all entries in a scatterlist.
516 * Merge chunks that have page aligned sizes into a continuous mapping.
518 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
523 unsigned long pages = 0;
524 int need = 0, nextneed;
526 BUG_ON(dir == DMA_NONE);
531 return swiotlb_map_sg(dev,sg,nents,dir);
537 for (i = 0; i < nents; i++) {
538 struct scatterlist *s = &sg[i];
539 dma_addr_t addr = page_to_phys(s->page) + s->offset;
540 s->dma_address = addr;
541 BUG_ON(s->length == 0);
543 nextneed = need_iommu(dev, addr, s->length);
545 /* Handle the previous not yet processed entries */
547 struct scatterlist *ps = &sg[i-1];
548 /* Can only merge when the last chunk ends on a page
549 boundary and the new one doesn't have an offset. */
550 if (!iommu_merge || !nextneed || !need || s->offset ||
551 (ps->offset + ps->length) % PAGE_SIZE) {
552 if (dma_map_cont(sg, start, i, sg+out, pages,
562 pages += to_pages(s->offset, s->length);
564 if (dma_map_cont(sg, start, i, sg+out, pages, need) < 0)
569 sg[out].dma_length = 0;
574 dma_unmap_sg(dev, sg, nents, dir);
575 /* When it was forced try again unforced */
577 return dma_map_sg_nonforce(dev, sg, nents, dir);
578 if (panic_on_overflow)
579 panic("dma_map_sg: overflow on %lu pages\n", pages);
580 iommu_full(dev, pages << PAGE_SHIFT, dir, 0);
581 for (i = 0; i < nents; i++)
582 sg[i].dma_address = bad_dma_address;
587 * Free a DMA mapping.
589 void dma_unmap_single(struct device *dev, dma_addr_t dma_addr,
590 size_t size, int direction)
592 unsigned long iommu_page;
597 swiotlb_unmap_single(dev,dma_addr,size,direction);
601 if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
602 dma_addr >= iommu_bus_base + iommu_size)
604 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
605 npages = to_pages(dma_addr, size);
606 for (i = 0; i < npages; i++) {
607 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
608 CLEAR_LEAK(iommu_page + i);
610 free_iommu(iommu_page, npages);
614 * Wrapper for pci_unmap_single working with scatterlists.
616 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
620 swiotlb_unmap_sg(dev,sg,nents,dir);
623 for (i = 0; i < nents; i++) {
624 struct scatterlist *s = &sg[i];
625 if (!s->dma_length || !s->length)
627 dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
631 int dma_supported(struct device *dev, u64 mask)
633 /* Copied from i386. Doesn't make much sense, because it will
634 only work for pci_alloc_coherent.
635 The caller just has to use GFP_DMA in this case. */
636 if (mask < 0x00ffffff)
639 /* Tell the device to use SAC when IOMMU force is on.
640 This allows the driver to use cheaper accesses in some cases.
642 Problem with this is that if we overflow the IOMMU area
643 and return DAC as fallback address the device may not handle it correctly.
645 As a special case some controllers have a 39bit address mode
646 that is as efficient as 32bit (aic79xx). Don't force SAC for these.
647 Assume all masks <= 40 bits are of this type. Normally this doesn't
648 make any difference, but gives more gentle handling of IOMMU overflow. */
649 if (iommu_sac_force && (mask >= 0xffffffffffULL)) {
650 printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
657 int dma_get_cache_alignment(void)
659 return boot_cpu_data.x86_clflush_size;
662 EXPORT_SYMBOL(dma_unmap_sg);
663 EXPORT_SYMBOL(dma_map_sg);
664 EXPORT_SYMBOL(dma_map_single);
665 EXPORT_SYMBOL(dma_unmap_single);
666 EXPORT_SYMBOL(dma_supported);
667 EXPORT_SYMBOL(no_iommu);
668 EXPORT_SYMBOL(force_iommu);
669 EXPORT_SYMBOL(bad_dma_address);
670 EXPORT_SYMBOL(iommu_bio_merge);
671 EXPORT_SYMBOL(iommu_sac_force);
672 EXPORT_SYMBOL(dma_get_cache_alignment);
673 EXPORT_SYMBOL(dma_alloc_coherent);
674 EXPORT_SYMBOL(dma_free_coherent);
676 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
680 iommu_size = aper_size;
685 a = aper + iommu_size;
686 iommu_size -= round_up(a, LARGE_PAGE_SIZE) - a;
688 if (iommu_size < 64*1024*1024)
690 "PCI-DMA: Warning: Small IOMMU %luMB. Consider increasing the AGP aperture in BIOS\n",iommu_size>>20);
695 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
697 unsigned aper_size = 0, aper_base_32;
701 pci_read_config_dword(dev, 0x94, &aper_base_32);
702 pci_read_config_dword(dev, 0x90, &aper_order);
703 aper_order = (aper_order >> 1) & 7;
705 aper_base = aper_base_32 & 0x7fff;
708 aper_size = (32 * 1024 * 1024) << aper_order;
709 if (aper_base + aper_size >= 0xffffffff || !aper_size)
717 * Private Northbridge GATT initialization in case we cannot use the
718 * AGP driver for some reason.
720 static __init int init_k8_gatt(struct agp_kern_info *info)
724 unsigned aper_base, new_aper_base;
725 unsigned aper_size, gatt_size, new_aper_size;
727 printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
728 aper_size = aper_base = info->aper_size = 0;
730 new_aper_base = read_aperture(dev, &new_aper_size);
735 aper_size = new_aper_size;
736 aper_base = new_aper_base;
738 if (aper_size != new_aper_size || aper_base != new_aper_base)
743 info->aper_base = aper_base;
744 info->aper_size = aper_size>>20;
746 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
747 gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
749 panic("Cannot allocate GATT table");
750 memset(gatt, 0, gatt_size);
751 agp_gatt_table = gatt;
757 gatt_reg = __pa(gatt) >> 12;
759 pci_write_config_dword(dev, 0x98, gatt_reg);
760 pci_read_config_dword(dev, 0x90, &ctl);
763 ctl &= ~((1<<4) | (1<<5));
765 pci_write_config_dword(dev, 0x90, ctl);
769 printk("PCI-DMA: aperture base @ %x size %u KB\n",aper_base, aper_size>>10);
773 /* Should not happen anymore */
774 printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
775 KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.");
779 extern int agp_amd64_init(void);
781 static int __init pci_iommu_init(void)
783 struct agp_kern_info info;
784 unsigned long aper_size;
785 unsigned long iommu_start;
787 unsigned long scratch;
790 #ifndef CONFIG_AGP_AMD64
793 /* Makefile puts PCI initialization via subsys_initcall first. */
794 /* Add other K8 AGP bridge drivers here */
796 (agp_amd64_init() < 0) ||
797 (agp_copy_info(agp_bridge, &info) < 0);
802 printk(KERN_INFO "PCI-DMA: Using software bounce buffering for IO (SWIOTLB)\n");
807 (!force_iommu && end_pfn < 0xffffffff>>PAGE_SHIFT) ||
809 (no_agp && init_k8_gatt(&info) < 0)) {
810 printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n");
815 aper_size = info.aper_size * 1024 * 1024;
816 iommu_size = check_iommu_size(info.aper_base, aper_size);
817 iommu_pages = iommu_size >> PAGE_SHIFT;
819 iommu_gart_bitmap = (void*)__get_free_pages(GFP_KERNEL,
820 get_order(iommu_pages/8));
821 if (!iommu_gart_bitmap)
822 panic("Cannot allocate iommu bitmap\n");
823 memset(iommu_gart_bitmap, 0, iommu_pages/8);
825 #ifdef CONFIG_IOMMU_LEAK
827 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
828 get_order(iommu_pages*sizeof(void *)));
830 memset(iommu_leak_tab, 0, iommu_pages * 8);
832 printk("PCI-DMA: Cannot allocate leak trace area\n");
837 * Out of IOMMU space handling.
838 * Reserve some invalid pages at the beginning of the GART.
840 set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
842 agp_memory_reserved = iommu_size;
844 "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
847 iommu_start = aper_size - iommu_size;
848 iommu_bus_base = info.aper_base + iommu_start;
849 bad_dma_address = iommu_bus_base;
850 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
853 * Unmap the IOMMU part of the GART. The alias of the page is
854 * always mapped with cache enabled and there is no full cache
855 * coherency across the GART remapping. The unmapping avoids
856 * automatic prefetches from the CPU allocating cache lines in
857 * there. All CPU accesses are done via the direct mapping to
858 * the backing memory. The GART address is only used by PCI
861 clear_kernel_mapping((unsigned long)__va(iommu_bus_base), iommu_size);
864 * Try to workaround a bug (thanks to BenH)
865 * Set unmapped entries to a scratch page instead of 0.
866 * Any prefetches that hit unmapped entries won't get an bus abort
869 scratch = get_zeroed_page(GFP_KERNEL);
871 panic("Cannot allocate iommu scratch page");
872 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
873 for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
874 iommu_gatt_base[i] = gart_unmapped_entry;
878 int cpu = PCI_SLOT(dev->devfn) - 24;
881 northbridges[cpu] = dev;
882 pci_read_config_dword(dev, 0x9c, &flag); /* cache flush word */
883 northbridge_flush_word[cpu] = flag;
891 /* Must execute after PCI subsystem */
892 fs_initcall(pci_iommu_init);
894 /* iommu=[size][,noagp][,off][,force][,noforce][,leak][,memaper[=order]][,merge]
895 [,forcesac][,fullflush][,nomerge][,biomerge]
896 size set size of iommu (in bytes)
897 noagp don't initialize the AGP driver and use full aperture.
898 off don't use the IOMMU
899 leak turn on simple iommu leak tracing (only when CONFIG_IOMMU_LEAK is on)
900 memaper[=order] allocate an own aperture over RAM with size 32MB^order.
901 noforce don't force IOMMU usage. Default.
903 merge Do lazy merging. This may improve performance on some block devices.
904 Implies force (experimental)
905 biomerge Do merging at the BIO layer. This is more efficient than merge,
906 but should be only done with very big IOMMUs. Implies merge,force.
907 nomerge Don't do SG merging.
908 forcesac For SAC mode for masks <40bits (experimental)
909 fullflush Flush IOMMU on each allocation (default)
910 nofullflush Don't use IOMMU fullflush
911 allowed overwrite iommu off workarounds for specific chipsets.
912 soft Use software bounce buffering (default for Intel machines)
913 noaperture Don't touch the aperture for AGP.
915 __init int iommu_setup(char *p)
920 if (!strncmp(p,"noagp",5))
922 if (!strncmp(p,"off",3))
924 if (!strncmp(p,"force",5)) {
926 iommu_aperture_allowed = 1;
928 if (!strncmp(p,"allowed",7))
929 iommu_aperture_allowed = 1;
930 if (!strncmp(p,"noforce",7)) {
934 if (!strncmp(p, "memaper", 7)) {
935 fallback_aper_force = 1;
939 if (get_option(&p, &arg))
940 fallback_aper_order = arg;
943 if (!strncmp(p, "biomerge",8)) {
944 iommu_bio_merge = 4096;
948 if (!strncmp(p, "panic",5))
949 panic_on_overflow = 1;
950 if (!strncmp(p, "nopanic",7))
951 panic_on_overflow = 0;
952 if (!strncmp(p, "merge",5)) {
956 if (!strncmp(p, "nomerge",7))
958 if (!strncmp(p, "forcesac",8))
960 if (!strncmp(p, "fullflush",8))
962 if (!strncmp(p, "nofullflush",11))
964 if (!strncmp(p, "soft",4))
966 if (!strncmp(p, "noaperture",10))
968 #ifdef CONFIG_IOMMU_LEAK
969 if (!strncmp(p,"leak",4)) {
973 if (isdigit(*p) && get_option(&p, &arg))
974 iommu_leak_pages = arg;
977 if (isdigit(*p) && get_option(&p, &arg))
979 p += strcspn(p, ",");