2 * Copyright (c) 2006, Intel Corporation.
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * You should have received a copy of the GNU General Public License along with
14 * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
15 * Place - Suite 330, Boston, MA 02111-1307 USA.
17 * Copyright (C) 2006-2008 Intel Corporation
18 * Author: Ashok Raj <ashok.raj@intel.com>
19 * Author: Shaohua Li <shaohua.li@intel.com>
20 * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
21 * Author: Fenghua Yu <fenghua.yu@intel.com>
24 #include <linux/init.h>
25 #include <linux/bitmap.h>
26 #include <linux/debugfs.h>
27 #include <linux/slab.h>
28 #include <linux/irq.h>
29 #include <linux/interrupt.h>
30 #include <linux/spinlock.h>
31 #include <linux/pci.h>
32 #include <linux/dmar.h>
33 #include <linux/dma-mapping.h>
34 #include <linux/mempool.h>
35 #include <linux/timer.h>
36 #include <linux/iova.h>
37 #include <linux/iommu.h>
38 #include <linux/intel-iommu.h>
39 #include <linux/sysdev.h>
40 #include <asm/cacheflush.h>
41 #include <asm/iommu.h>
44 #define ROOT_SIZE VTD_PAGE_SIZE
45 #define CONTEXT_SIZE VTD_PAGE_SIZE
47 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
48 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
50 #define IOAPIC_RANGE_START (0xfee00000)
51 #define IOAPIC_RANGE_END (0xfeefffff)
52 #define IOVA_START_ADDR (0x1000)
54 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
56 #define MAX_AGAW_WIDTH 64
58 #define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
59 #define DOMAIN_MAX_PFN(gaw) ((((u64)1) << (gaw-VTD_PAGE_SHIFT)) - 1)
61 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
62 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
63 #define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
66 /* VT-d pages must always be _smaller_ than MM pages. Otherwise things
67 are never going to work. */
68 static inline unsigned long dma_to_mm_pfn(unsigned long dma_pfn)
70 return dma_pfn >> (PAGE_SHIFT - VTD_PAGE_SHIFT);
73 static inline unsigned long mm_to_dma_pfn(unsigned long mm_pfn)
75 return mm_pfn << (PAGE_SHIFT - VTD_PAGE_SHIFT);
77 static inline unsigned long page_to_dma_pfn(struct page *pg)
79 return mm_to_dma_pfn(page_to_pfn(pg));
81 static inline unsigned long virt_to_dma_pfn(void *p)
83 return page_to_dma_pfn(virt_to_page(p));
86 /* global iommu list, set NULL for ignored DMAR units */
87 static struct intel_iommu **g_iommus;
89 static int rwbf_quirk;
94 * 12-63: Context Ptr (12 - (haw-1))
101 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
102 static inline bool root_present(struct root_entry *root)
104 return (root->val & 1);
106 static inline void set_root_present(struct root_entry *root)
110 static inline void set_root_value(struct root_entry *root, unsigned long value)
112 root->val |= value & VTD_PAGE_MASK;
115 static inline struct context_entry *
116 get_context_addr_from_root(struct root_entry *root)
118 return (struct context_entry *)
119 (root_present(root)?phys_to_virt(
120 root->val & VTD_PAGE_MASK) :
127 * 1: fault processing disable
128 * 2-3: translation type
129 * 12-63: address space root
135 struct context_entry {
140 static inline bool context_present(struct context_entry *context)
142 return (context->lo & 1);
144 static inline void context_set_present(struct context_entry *context)
149 static inline void context_set_fault_enable(struct context_entry *context)
151 context->lo &= (((u64)-1) << 2) | 1;
154 static inline void context_set_translation_type(struct context_entry *context,
157 context->lo &= (((u64)-1) << 4) | 3;
158 context->lo |= (value & 3) << 2;
161 static inline void context_set_address_root(struct context_entry *context,
164 context->lo |= value & VTD_PAGE_MASK;
167 static inline void context_set_address_width(struct context_entry *context,
170 context->hi |= value & 7;
173 static inline void context_set_domain_id(struct context_entry *context,
176 context->hi |= (value & ((1 << 16) - 1)) << 8;
179 static inline void context_clear_entry(struct context_entry *context)
192 * 12-63: Host physcial address
198 static inline void dma_clear_pte(struct dma_pte *pte)
203 static inline void dma_set_pte_readable(struct dma_pte *pte)
205 pte->val |= DMA_PTE_READ;
208 static inline void dma_set_pte_writable(struct dma_pte *pte)
210 pte->val |= DMA_PTE_WRITE;
213 static inline void dma_set_pte_snp(struct dma_pte *pte)
215 pte->val |= DMA_PTE_SNP;
218 static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)
220 pte->val = (pte->val & ~3) | (prot & 3);
223 static inline u64 dma_pte_addr(struct dma_pte *pte)
226 return pte->val & VTD_PAGE_MASK;
228 /* Must have a full atomic 64-bit read */
229 return __cmpxchg64(pte, 0ULL, 0ULL) & VTD_PAGE_MASK;
233 static inline void dma_set_pte_pfn(struct dma_pte *pte, unsigned long pfn)
235 pte->val |= (uint64_t)pfn << VTD_PAGE_SHIFT;
238 static inline bool dma_pte_present(struct dma_pte *pte)
240 return (pte->val & 3) != 0;
243 static inline int first_pte_in_page(struct dma_pte *pte)
245 return !((unsigned long)pte & ~VTD_PAGE_MASK);
249 * This domain is a statically identity mapping domain.
250 * 1. This domain creats a static 1:1 mapping to all usable memory.
251 * 2. It maps to each iommu if successful.
252 * 3. Each iommu mapps to this domain if successful.
254 struct dmar_domain *si_domain;
256 /* devices under the same p2p bridge are owned in one domain */
257 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
259 /* domain represents a virtual machine, more than one devices
260 * across iommus may be owned in one domain, e.g. kvm guest.
262 #define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 1)
264 /* si_domain contains mulitple devices */
265 #define DOMAIN_FLAG_STATIC_IDENTITY (1 << 2)
268 int id; /* domain id */
269 unsigned long iommu_bmp; /* bitmap of iommus this domain uses*/
271 struct list_head devices; /* all devices' list */
272 struct iova_domain iovad; /* iova's that belong to this domain */
274 struct dma_pte *pgd; /* virtual address */
275 int gaw; /* max guest address width */
277 /* adjusted guest address width, 0 is level 2 30-bit */
280 int flags; /* flags to find out type of domain */
282 int iommu_coherency;/* indicate coherency of iommu access */
283 int iommu_snooping; /* indicate snooping control feature*/
284 int iommu_count; /* reference count of iommu */
285 spinlock_t iommu_lock; /* protect iommu set in domain */
286 u64 max_addr; /* maximum mapped address */
289 /* PCI domain-device relationship */
290 struct device_domain_info {
291 struct list_head link; /* link to domain siblings */
292 struct list_head global; /* link to global list */
293 int segment; /* PCI domain */
294 u8 bus; /* PCI bus number */
295 u8 devfn; /* PCI devfn number */
296 struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
297 struct intel_iommu *iommu; /* IOMMU used by this device */
298 struct dmar_domain *domain; /* pointer to domain */
301 static void flush_unmaps_timeout(unsigned long data);
303 DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
305 #define HIGH_WATER_MARK 250
306 struct deferred_flush_tables {
308 struct iova *iova[HIGH_WATER_MARK];
309 struct dmar_domain *domain[HIGH_WATER_MARK];
312 static struct deferred_flush_tables *deferred_flush;
314 /* bitmap for indexing intel_iommus */
315 static int g_num_of_iommus;
317 static DEFINE_SPINLOCK(async_umap_flush_lock);
318 static LIST_HEAD(unmaps_to_do);
321 static long list_size;
323 static void domain_remove_dev_info(struct dmar_domain *domain);
325 #ifdef CONFIG_DMAR_DEFAULT_ON
326 int dmar_disabled = 0;
328 int dmar_disabled = 1;
329 #endif /*CONFIG_DMAR_DEFAULT_ON*/
331 static int __initdata dmar_map_gfx = 1;
332 static int dmar_forcedac;
333 static int intel_iommu_strict;
335 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
336 static DEFINE_SPINLOCK(device_domain_lock);
337 static LIST_HEAD(device_domain_list);
339 static struct iommu_ops intel_iommu_ops;
341 static int __init intel_iommu_setup(char *str)
346 if (!strncmp(str, "on", 2)) {
348 printk(KERN_INFO "Intel-IOMMU: enabled\n");
349 } else if (!strncmp(str, "off", 3)) {
351 printk(KERN_INFO "Intel-IOMMU: disabled\n");
352 } else if (!strncmp(str, "igfx_off", 8)) {
355 "Intel-IOMMU: disable GFX device mapping\n");
356 } else if (!strncmp(str, "forcedac", 8)) {
358 "Intel-IOMMU: Forcing DAC for PCI devices\n");
360 } else if (!strncmp(str, "strict", 6)) {
362 "Intel-IOMMU: disable batched IOTLB flush\n");
363 intel_iommu_strict = 1;
366 str += strcspn(str, ",");
372 __setup("intel_iommu=", intel_iommu_setup);
374 static struct kmem_cache *iommu_domain_cache;
375 static struct kmem_cache *iommu_devinfo_cache;
376 static struct kmem_cache *iommu_iova_cache;
378 static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
383 /* trying to avoid low memory issues */
384 flags = current->flags & PF_MEMALLOC;
385 current->flags |= PF_MEMALLOC;
386 vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
387 current->flags &= (~PF_MEMALLOC | flags);
392 static inline void *alloc_pgtable_page(void)
397 /* trying to avoid low memory issues */
398 flags = current->flags & PF_MEMALLOC;
399 current->flags |= PF_MEMALLOC;
400 vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
401 current->flags &= (~PF_MEMALLOC | flags);
405 static inline void free_pgtable_page(void *vaddr)
407 free_page((unsigned long)vaddr);
410 static inline void *alloc_domain_mem(void)
412 return iommu_kmem_cache_alloc(iommu_domain_cache);
415 static void free_domain_mem(void *vaddr)
417 kmem_cache_free(iommu_domain_cache, vaddr);
420 static inline void * alloc_devinfo_mem(void)
422 return iommu_kmem_cache_alloc(iommu_devinfo_cache);
425 static inline void free_devinfo_mem(void *vaddr)
427 kmem_cache_free(iommu_devinfo_cache, vaddr);
430 struct iova *alloc_iova_mem(void)
432 return iommu_kmem_cache_alloc(iommu_iova_cache);
435 void free_iova_mem(struct iova *iova)
437 kmem_cache_free(iommu_iova_cache, iova);
441 static inline int width_to_agaw(int width);
443 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
448 sagaw = cap_sagaw(iommu->cap);
449 for (agaw = width_to_agaw(max_gaw);
451 if (test_bit(agaw, &sagaw))
459 * Calculate max SAGAW for each iommu.
461 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
463 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
467 * calculate agaw for each iommu.
468 * "SAGAW" may be different across iommus, use a default agaw, and
469 * get a supported less agaw for iommus that don't support the default agaw.
471 int iommu_calculate_agaw(struct intel_iommu *iommu)
473 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
476 /* This functionin only returns single iommu in a domain */
477 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
481 /* si_domain and vm domain should not get here. */
482 BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
483 BUG_ON(domain->flags & DOMAIN_FLAG_STATIC_IDENTITY);
485 iommu_id = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
486 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
489 return g_iommus[iommu_id];
492 static void domain_update_iommu_coherency(struct dmar_domain *domain)
496 domain->iommu_coherency = 1;
498 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
499 for (; i < g_num_of_iommus; ) {
500 if (!ecap_coherent(g_iommus[i]->ecap)) {
501 domain->iommu_coherency = 0;
504 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
508 static void domain_update_iommu_snooping(struct dmar_domain *domain)
512 domain->iommu_snooping = 1;
514 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
515 for (; i < g_num_of_iommus; ) {
516 if (!ecap_sc_support(g_iommus[i]->ecap)) {
517 domain->iommu_snooping = 0;
520 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
524 /* Some capabilities may be different across iommus */
525 static void domain_update_iommu_cap(struct dmar_domain *domain)
527 domain_update_iommu_coherency(domain);
528 domain_update_iommu_snooping(domain);
531 static struct intel_iommu *device_to_iommu(int segment, u8 bus, u8 devfn)
533 struct dmar_drhd_unit *drhd = NULL;
536 for_each_drhd_unit(drhd) {
539 if (segment != drhd->segment)
542 for (i = 0; i < drhd->devices_cnt; i++) {
543 if (drhd->devices[i] &&
544 drhd->devices[i]->bus->number == bus &&
545 drhd->devices[i]->devfn == devfn)
547 if (drhd->devices[i] &&
548 drhd->devices[i]->subordinate &&
549 drhd->devices[i]->subordinate->number <= bus &&
550 drhd->devices[i]->subordinate->subordinate >= bus)
554 if (drhd->include_all)
561 static void domain_flush_cache(struct dmar_domain *domain,
562 void *addr, int size)
564 if (!domain->iommu_coherency)
565 clflush_cache_range(addr, size);
568 /* Gets context entry for a given bus and devfn */
569 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
572 struct root_entry *root;
573 struct context_entry *context;
574 unsigned long phy_addr;
577 spin_lock_irqsave(&iommu->lock, flags);
578 root = &iommu->root_entry[bus];
579 context = get_context_addr_from_root(root);
581 context = (struct context_entry *)alloc_pgtable_page();
583 spin_unlock_irqrestore(&iommu->lock, flags);
586 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
587 phy_addr = virt_to_phys((void *)context);
588 set_root_value(root, phy_addr);
589 set_root_present(root);
590 __iommu_flush_cache(iommu, root, sizeof(*root));
592 spin_unlock_irqrestore(&iommu->lock, flags);
593 return &context[devfn];
596 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
598 struct root_entry *root;
599 struct context_entry *context;
603 spin_lock_irqsave(&iommu->lock, flags);
604 root = &iommu->root_entry[bus];
605 context = get_context_addr_from_root(root);
610 ret = context_present(&context[devfn]);
612 spin_unlock_irqrestore(&iommu->lock, flags);
616 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
618 struct root_entry *root;
619 struct context_entry *context;
622 spin_lock_irqsave(&iommu->lock, flags);
623 root = &iommu->root_entry[bus];
624 context = get_context_addr_from_root(root);
626 context_clear_entry(&context[devfn]);
627 __iommu_flush_cache(iommu, &context[devfn], \
630 spin_unlock_irqrestore(&iommu->lock, flags);
633 static void free_context_table(struct intel_iommu *iommu)
635 struct root_entry *root;
638 struct context_entry *context;
640 spin_lock_irqsave(&iommu->lock, flags);
641 if (!iommu->root_entry) {
644 for (i = 0; i < ROOT_ENTRY_NR; i++) {
645 root = &iommu->root_entry[i];
646 context = get_context_addr_from_root(root);
648 free_pgtable_page(context);
650 free_pgtable_page(iommu->root_entry);
651 iommu->root_entry = NULL;
653 spin_unlock_irqrestore(&iommu->lock, flags);
656 /* page table handling */
657 #define LEVEL_STRIDE (9)
658 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
660 static inline int agaw_to_level(int agaw)
665 static inline int agaw_to_width(int agaw)
667 return 30 + agaw * LEVEL_STRIDE;
671 static inline int width_to_agaw(int width)
673 return (width - 30) / LEVEL_STRIDE;
676 static inline unsigned int level_to_offset_bits(int level)
678 return (level - 1) * LEVEL_STRIDE;
681 static inline int pfn_level_offset(unsigned long pfn, int level)
683 return (pfn >> level_to_offset_bits(level)) & LEVEL_MASK;
686 static inline unsigned long level_mask(int level)
688 return -1UL << level_to_offset_bits(level);
691 static inline unsigned long level_size(int level)
693 return 1UL << level_to_offset_bits(level);
696 static inline unsigned long align_to_level(unsigned long pfn, int level)
698 return (pfn + level_size(level) - 1) & level_mask(level);
701 static struct dma_pte *pfn_to_dma_pte(struct dmar_domain *domain,
704 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
705 struct dma_pte *parent, *pte = NULL;
706 int level = agaw_to_level(domain->agaw);
709 BUG_ON(!domain->pgd);
710 BUG_ON(addr_width < BITS_PER_LONG && pfn >> addr_width);
711 parent = domain->pgd;
716 offset = pfn_level_offset(pfn, level);
717 pte = &parent[offset];
721 if (!dma_pte_present(pte)) {
724 tmp_page = alloc_pgtable_page();
729 domain_flush_cache(domain, tmp_page, VTD_PAGE_SIZE);
730 pteval = (virt_to_dma_pfn(tmp_page) << VTD_PAGE_SHIFT) | DMA_PTE_READ | DMA_PTE_WRITE;
731 if (cmpxchg64(&pte->val, 0ULL, pteval)) {
732 /* Someone else set it while we were thinking; use theirs. */
733 free_pgtable_page(tmp_page);
736 domain_flush_cache(domain, pte, sizeof(*pte));
739 parent = phys_to_virt(dma_pte_addr(pte));
746 /* return address's pte at specific level */
747 static struct dma_pte *dma_pfn_level_pte(struct dmar_domain *domain,
751 struct dma_pte *parent, *pte = NULL;
752 int total = agaw_to_level(domain->agaw);
755 parent = domain->pgd;
756 while (level <= total) {
757 offset = pfn_level_offset(pfn, total);
758 pte = &parent[offset];
762 if (!dma_pte_present(pte))
764 parent = phys_to_virt(dma_pte_addr(pte));
770 /* clear last level pte, a tlb flush should be followed */
771 static void dma_pte_clear_range(struct dmar_domain *domain,
772 unsigned long start_pfn,
773 unsigned long last_pfn)
775 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
776 struct dma_pte *first_pte, *pte;
778 BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
779 BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
781 /* we don't need lock here; nobody else touches the iova range */
782 while (start_pfn <= last_pfn) {
783 first_pte = pte = dma_pfn_level_pte(domain, start_pfn, 1);
785 start_pfn = align_to_level(start_pfn + 1, 2);
792 } while (start_pfn <= last_pfn && !first_pte_in_page(pte));
794 domain_flush_cache(domain, first_pte,
795 (void *)pte - (void *)first_pte);
799 /* free page table pages. last level pte should already be cleared */
800 static void dma_pte_free_pagetable(struct dmar_domain *domain,
801 unsigned long start_pfn,
802 unsigned long last_pfn)
804 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
805 struct dma_pte *first_pte, *pte;
806 int total = agaw_to_level(domain->agaw);
810 BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
811 BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
813 /* We don't need lock here; nobody else touches the iova range */
815 while (level <= total) {
816 tmp = align_to_level(start_pfn, level);
818 /* If we can't even clear one PTE at this level, we're done */
819 if (tmp + level_size(level) - 1 > last_pfn)
822 while (tmp + level_size(level) - 1 <= last_pfn) {
823 first_pte = pte = dma_pfn_level_pte(domain, tmp, level);
825 tmp = align_to_level(tmp + 1, level + 1);
829 if (dma_pte_present(pte)) {
830 free_pgtable_page(phys_to_virt(dma_pte_addr(pte)));
834 tmp += level_size(level);
835 } while (!first_pte_in_page(pte) &&
836 tmp + level_size(level) - 1 <= last_pfn);
838 domain_flush_cache(domain, first_pte,
839 (void *)pte - (void *)first_pte);
845 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
846 free_pgtable_page(domain->pgd);
852 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
854 struct root_entry *root;
857 root = (struct root_entry *)alloc_pgtable_page();
861 __iommu_flush_cache(iommu, root, ROOT_SIZE);
863 spin_lock_irqsave(&iommu->lock, flags);
864 iommu->root_entry = root;
865 spin_unlock_irqrestore(&iommu->lock, flags);
870 static void iommu_set_root_entry(struct intel_iommu *iommu)
876 addr = iommu->root_entry;
878 spin_lock_irqsave(&iommu->register_lock, flag);
879 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
881 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
883 /* Make sure hardware complete it */
884 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
885 readl, (sts & DMA_GSTS_RTPS), sts);
887 spin_unlock_irqrestore(&iommu->register_lock, flag);
890 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
895 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
898 spin_lock_irqsave(&iommu->register_lock, flag);
899 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
901 /* Make sure hardware complete it */
902 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
903 readl, (!(val & DMA_GSTS_WBFS)), val);
905 spin_unlock_irqrestore(&iommu->register_lock, flag);
908 /* return value determine if we need a write buffer flush */
909 static void __iommu_flush_context(struct intel_iommu *iommu,
910 u16 did, u16 source_id, u8 function_mask,
917 case DMA_CCMD_GLOBAL_INVL:
918 val = DMA_CCMD_GLOBAL_INVL;
920 case DMA_CCMD_DOMAIN_INVL:
921 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
923 case DMA_CCMD_DEVICE_INVL:
924 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
925 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
932 spin_lock_irqsave(&iommu->register_lock, flag);
933 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
935 /* Make sure hardware complete it */
936 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
937 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
939 spin_unlock_irqrestore(&iommu->register_lock, flag);
942 /* return value determine if we need a write buffer flush */
943 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
944 u64 addr, unsigned int size_order, u64 type)
946 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
947 u64 val = 0, val_iva = 0;
951 case DMA_TLB_GLOBAL_FLUSH:
952 /* global flush doesn't need set IVA_REG */
953 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
955 case DMA_TLB_DSI_FLUSH:
956 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
958 case DMA_TLB_PSI_FLUSH:
959 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
960 /* Note: always flush non-leaf currently */
961 val_iva = size_order | addr;
966 /* Note: set drain read/write */
969 * This is probably to be super secure.. Looks like we can
970 * ignore it without any impact.
972 if (cap_read_drain(iommu->cap))
973 val |= DMA_TLB_READ_DRAIN;
975 if (cap_write_drain(iommu->cap))
976 val |= DMA_TLB_WRITE_DRAIN;
978 spin_lock_irqsave(&iommu->register_lock, flag);
979 /* Note: Only uses first TLB reg currently */
981 dmar_writeq(iommu->reg + tlb_offset, val_iva);
982 dmar_writeq(iommu->reg + tlb_offset + 8, val);
984 /* Make sure hardware complete it */
985 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
986 dmar_readq, (!(val & DMA_TLB_IVT)), val);
988 spin_unlock_irqrestore(&iommu->register_lock, flag);
990 /* check IOTLB invalidation granularity */
991 if (DMA_TLB_IAIG(val) == 0)
992 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
993 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
994 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
995 (unsigned long long)DMA_TLB_IIRG(type),
996 (unsigned long long)DMA_TLB_IAIG(val));
999 static struct device_domain_info *iommu_support_dev_iotlb(
1000 struct dmar_domain *domain, int segment, u8 bus, u8 devfn)
1003 unsigned long flags;
1004 struct device_domain_info *info;
1005 struct intel_iommu *iommu = device_to_iommu(segment, bus, devfn);
1007 if (!ecap_dev_iotlb_support(iommu->ecap))
1013 spin_lock_irqsave(&device_domain_lock, flags);
1014 list_for_each_entry(info, &domain->devices, link)
1015 if (info->bus == bus && info->devfn == devfn) {
1019 spin_unlock_irqrestore(&device_domain_lock, flags);
1021 if (!found || !info->dev)
1024 if (!pci_find_ext_capability(info->dev, PCI_EXT_CAP_ID_ATS))
1027 if (!dmar_find_matched_atsr_unit(info->dev))
1030 info->iommu = iommu;
1035 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
1040 pci_enable_ats(info->dev, VTD_PAGE_SHIFT);
1043 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1045 if (!info->dev || !pci_ats_enabled(info->dev))
1048 pci_disable_ats(info->dev);
1051 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1052 u64 addr, unsigned mask)
1055 unsigned long flags;
1056 struct device_domain_info *info;
1058 spin_lock_irqsave(&device_domain_lock, flags);
1059 list_for_each_entry(info, &domain->devices, link) {
1060 if (!info->dev || !pci_ats_enabled(info->dev))
1063 sid = info->bus << 8 | info->devfn;
1064 qdep = pci_ats_queue_depth(info->dev);
1065 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1067 spin_unlock_irqrestore(&device_domain_lock, flags);
1070 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
1071 unsigned long pfn, unsigned int pages)
1073 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1074 uint64_t addr = (uint64_t)pfn << VTD_PAGE_SHIFT;
1079 * Fallback to domain selective flush if no PSI support or the size is
1081 * PSI requires page size to be 2 ^ x, and the base address is naturally
1082 * aligned to the size
1084 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1085 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1088 iommu->flush.flush_iotlb(iommu, did, addr, mask,
1092 * In caching mode, domain ID 0 is reserved for non-present to present
1093 * mapping flush. Device IOTLB doesn't need to be flushed in this case.
1095 if (!cap_caching_mode(iommu->cap) || did)
1096 iommu_flush_dev_iotlb(iommu->domains[did], addr, mask);
1099 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1102 unsigned long flags;
1104 spin_lock_irqsave(&iommu->register_lock, flags);
1105 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1106 pmen &= ~DMA_PMEN_EPM;
1107 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1109 /* wait for the protected region status bit to clear */
1110 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1111 readl, !(pmen & DMA_PMEN_PRS), pmen);
1113 spin_unlock_irqrestore(&iommu->register_lock, flags);
1116 static int iommu_enable_translation(struct intel_iommu *iommu)
1119 unsigned long flags;
1121 spin_lock_irqsave(&iommu->register_lock, flags);
1122 iommu->gcmd |= DMA_GCMD_TE;
1123 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1125 /* Make sure hardware complete it */
1126 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1127 readl, (sts & DMA_GSTS_TES), sts);
1129 spin_unlock_irqrestore(&iommu->register_lock, flags);
1133 static int iommu_disable_translation(struct intel_iommu *iommu)
1138 spin_lock_irqsave(&iommu->register_lock, flag);
1139 iommu->gcmd &= ~DMA_GCMD_TE;
1140 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1142 /* Make sure hardware complete it */
1143 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1144 readl, (!(sts & DMA_GSTS_TES)), sts);
1146 spin_unlock_irqrestore(&iommu->register_lock, flag);
1151 static int iommu_init_domains(struct intel_iommu *iommu)
1153 unsigned long ndomains;
1154 unsigned long nlongs;
1156 ndomains = cap_ndoms(iommu->cap);
1157 pr_debug("Number of Domains supportd <%ld>\n", ndomains);
1158 nlongs = BITS_TO_LONGS(ndomains);
1160 /* TBD: there might be 64K domains,
1161 * consider other allocation for future chip
1163 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1164 if (!iommu->domain_ids) {
1165 printk(KERN_ERR "Allocating domain id array failed\n");
1168 iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1170 if (!iommu->domains) {
1171 printk(KERN_ERR "Allocating domain array failed\n");
1172 kfree(iommu->domain_ids);
1176 spin_lock_init(&iommu->lock);
1179 * if Caching mode is set, then invalid translations are tagged
1180 * with domainid 0. Hence we need to pre-allocate it.
1182 if (cap_caching_mode(iommu->cap))
1183 set_bit(0, iommu->domain_ids);
1188 static void domain_exit(struct dmar_domain *domain);
1189 static void vm_domain_exit(struct dmar_domain *domain);
1191 void free_dmar_iommu(struct intel_iommu *iommu)
1193 struct dmar_domain *domain;
1195 unsigned long flags;
1197 i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1198 for (; i < cap_ndoms(iommu->cap); ) {
1199 domain = iommu->domains[i];
1200 clear_bit(i, iommu->domain_ids);
1202 spin_lock_irqsave(&domain->iommu_lock, flags);
1203 if (--domain->iommu_count == 0) {
1204 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
1205 vm_domain_exit(domain);
1207 domain_exit(domain);
1209 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1211 i = find_next_bit(iommu->domain_ids,
1212 cap_ndoms(iommu->cap), i+1);
1215 if (iommu->gcmd & DMA_GCMD_TE)
1216 iommu_disable_translation(iommu);
1219 set_irq_data(iommu->irq, NULL);
1220 /* This will mask the irq */
1221 free_irq(iommu->irq, iommu);
1222 destroy_irq(iommu->irq);
1225 kfree(iommu->domains);
1226 kfree(iommu->domain_ids);
1228 g_iommus[iommu->seq_id] = NULL;
1230 /* if all iommus are freed, free g_iommus */
1231 for (i = 0; i < g_num_of_iommus; i++) {
1236 if (i == g_num_of_iommus)
1239 /* free context mapping */
1240 free_context_table(iommu);
1243 static struct dmar_domain *alloc_domain(void)
1245 struct dmar_domain *domain;
1247 domain = alloc_domain_mem();
1251 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
1257 static int iommu_attach_domain(struct dmar_domain *domain,
1258 struct intel_iommu *iommu)
1261 unsigned long ndomains;
1262 unsigned long flags;
1264 ndomains = cap_ndoms(iommu->cap);
1266 spin_lock_irqsave(&iommu->lock, flags);
1268 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1269 if (num >= ndomains) {
1270 spin_unlock_irqrestore(&iommu->lock, flags);
1271 printk(KERN_ERR "IOMMU: no free domain ids\n");
1276 set_bit(num, iommu->domain_ids);
1277 set_bit(iommu->seq_id, &domain->iommu_bmp);
1278 iommu->domains[num] = domain;
1279 spin_unlock_irqrestore(&iommu->lock, flags);
1284 static void iommu_detach_domain(struct dmar_domain *domain,
1285 struct intel_iommu *iommu)
1287 unsigned long flags;
1291 spin_lock_irqsave(&iommu->lock, flags);
1292 ndomains = cap_ndoms(iommu->cap);
1293 num = find_first_bit(iommu->domain_ids, ndomains);
1294 for (; num < ndomains; ) {
1295 if (iommu->domains[num] == domain) {
1299 num = find_next_bit(iommu->domain_ids,
1300 cap_ndoms(iommu->cap), num+1);
1304 clear_bit(num, iommu->domain_ids);
1305 clear_bit(iommu->seq_id, &domain->iommu_bmp);
1306 iommu->domains[num] = NULL;
1308 spin_unlock_irqrestore(&iommu->lock, flags);
1311 static struct iova_domain reserved_iova_list;
1312 static struct lock_class_key reserved_rbtree_key;
1314 static void dmar_init_reserved_ranges(void)
1316 struct pci_dev *pdev = NULL;
1320 init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1322 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1323 &reserved_rbtree_key);
1325 /* IOAPIC ranges shouldn't be accessed by DMA */
1326 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1327 IOVA_PFN(IOAPIC_RANGE_END));
1329 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1331 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1332 for_each_pci_dev(pdev) {
1335 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1336 r = &pdev->resource[i];
1337 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1339 iova = reserve_iova(&reserved_iova_list,
1343 printk(KERN_ERR "Reserve iova failed\n");
1349 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1351 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1354 static inline int guestwidth_to_adjustwidth(int gaw)
1357 int r = (gaw - 12) % 9;
1368 static int domain_init(struct dmar_domain *domain, int guest_width)
1370 struct intel_iommu *iommu;
1371 int adjust_width, agaw;
1372 unsigned long sagaw;
1374 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1375 spin_lock_init(&domain->iommu_lock);
1377 domain_reserve_special_ranges(domain);
1379 /* calculate AGAW */
1380 iommu = domain_get_iommu(domain);
1381 if (guest_width > cap_mgaw(iommu->cap))
1382 guest_width = cap_mgaw(iommu->cap);
1383 domain->gaw = guest_width;
1384 adjust_width = guestwidth_to_adjustwidth(guest_width);
1385 agaw = width_to_agaw(adjust_width);
1386 sagaw = cap_sagaw(iommu->cap);
1387 if (!test_bit(agaw, &sagaw)) {
1388 /* hardware doesn't support it, choose a bigger one */
1389 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1390 agaw = find_next_bit(&sagaw, 5, agaw);
1394 domain->agaw = agaw;
1395 INIT_LIST_HEAD(&domain->devices);
1397 if (ecap_coherent(iommu->ecap))
1398 domain->iommu_coherency = 1;
1400 domain->iommu_coherency = 0;
1402 if (ecap_sc_support(iommu->ecap))
1403 domain->iommu_snooping = 1;
1405 domain->iommu_snooping = 0;
1407 domain->iommu_count = 1;
1409 /* always allocate the top pgd */
1410 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1413 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1417 static void domain_exit(struct dmar_domain *domain)
1419 struct dmar_drhd_unit *drhd;
1420 struct intel_iommu *iommu;
1422 /* Domain 0 is reserved, so dont process it */
1426 domain_remove_dev_info(domain);
1428 put_iova_domain(&domain->iovad);
1431 dma_pte_clear_range(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1433 /* free page tables */
1434 dma_pte_free_pagetable(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1436 for_each_active_iommu(iommu, drhd)
1437 if (test_bit(iommu->seq_id, &domain->iommu_bmp))
1438 iommu_detach_domain(domain, iommu);
1440 free_domain_mem(domain);
1443 static int domain_context_mapping_one(struct dmar_domain *domain, int segment,
1444 u8 bus, u8 devfn, int translation)
1446 struct context_entry *context;
1447 unsigned long flags;
1448 struct intel_iommu *iommu;
1449 struct dma_pte *pgd;
1451 unsigned long ndomains;
1454 struct device_domain_info *info = NULL;
1456 pr_debug("Set context mapping for %02x:%02x.%d\n",
1457 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1459 BUG_ON(!domain->pgd);
1460 BUG_ON(translation != CONTEXT_TT_PASS_THROUGH &&
1461 translation != CONTEXT_TT_MULTI_LEVEL);
1463 iommu = device_to_iommu(segment, bus, devfn);
1467 context = device_to_context_entry(iommu, bus, devfn);
1470 spin_lock_irqsave(&iommu->lock, flags);
1471 if (context_present(context)) {
1472 spin_unlock_irqrestore(&iommu->lock, flags);
1479 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
1480 domain->flags & DOMAIN_FLAG_STATIC_IDENTITY) {
1483 /* find an available domain id for this device in iommu */
1484 ndomains = cap_ndoms(iommu->cap);
1485 num = find_first_bit(iommu->domain_ids, ndomains);
1486 for (; num < ndomains; ) {
1487 if (iommu->domains[num] == domain) {
1492 num = find_next_bit(iommu->domain_ids,
1493 cap_ndoms(iommu->cap), num+1);
1497 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1498 if (num >= ndomains) {
1499 spin_unlock_irqrestore(&iommu->lock, flags);
1500 printk(KERN_ERR "IOMMU: no free domain ids\n");
1504 set_bit(num, iommu->domain_ids);
1505 set_bit(iommu->seq_id, &domain->iommu_bmp);
1506 iommu->domains[num] = domain;
1510 /* Skip top levels of page tables for
1511 * iommu which has less agaw than default.
1513 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1514 pgd = phys_to_virt(dma_pte_addr(pgd));
1515 if (!dma_pte_present(pgd)) {
1516 spin_unlock_irqrestore(&iommu->lock, flags);
1522 context_set_domain_id(context, id);
1524 if (translation != CONTEXT_TT_PASS_THROUGH) {
1525 info = iommu_support_dev_iotlb(domain, segment, bus, devfn);
1526 translation = info ? CONTEXT_TT_DEV_IOTLB :
1527 CONTEXT_TT_MULTI_LEVEL;
1530 * In pass through mode, AW must be programmed to indicate the largest
1531 * AGAW value supported by hardware. And ASR is ignored by hardware.
1533 if (unlikely(translation == CONTEXT_TT_PASS_THROUGH))
1534 context_set_address_width(context, iommu->msagaw);
1536 context_set_address_root(context, virt_to_phys(pgd));
1537 context_set_address_width(context, iommu->agaw);
1540 context_set_translation_type(context, translation);
1541 context_set_fault_enable(context);
1542 context_set_present(context);
1543 domain_flush_cache(domain, context, sizeof(*context));
1546 * It's a non-present to present mapping. If hardware doesn't cache
1547 * non-present entry we only need to flush the write-buffer. If the
1548 * _does_ cache non-present entries, then it does so in the special
1549 * domain #0, which we have to flush:
1551 if (cap_caching_mode(iommu->cap)) {
1552 iommu->flush.flush_context(iommu, 0,
1553 (((u16)bus) << 8) | devfn,
1554 DMA_CCMD_MASK_NOBIT,
1555 DMA_CCMD_DEVICE_INVL);
1556 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
1558 iommu_flush_write_buffer(iommu);
1560 iommu_enable_dev_iotlb(info);
1561 spin_unlock_irqrestore(&iommu->lock, flags);
1563 spin_lock_irqsave(&domain->iommu_lock, flags);
1564 if (!test_and_set_bit(iommu->seq_id, &domain->iommu_bmp)) {
1565 domain->iommu_count++;
1566 domain_update_iommu_cap(domain);
1568 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1573 domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev,
1577 struct pci_dev *tmp, *parent;
1579 ret = domain_context_mapping_one(domain, pci_domain_nr(pdev->bus),
1580 pdev->bus->number, pdev->devfn,
1585 /* dependent device mapping */
1586 tmp = pci_find_upstream_pcie_bridge(pdev);
1589 /* Secondary interface's bus number and devfn 0 */
1590 parent = pdev->bus->self;
1591 while (parent != tmp) {
1592 ret = domain_context_mapping_one(domain,
1593 pci_domain_nr(parent->bus),
1594 parent->bus->number,
1595 parent->devfn, translation);
1598 parent = parent->bus->self;
1600 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1601 return domain_context_mapping_one(domain,
1602 pci_domain_nr(tmp->subordinate),
1603 tmp->subordinate->number, 0,
1605 else /* this is a legacy PCI bridge */
1606 return domain_context_mapping_one(domain,
1607 pci_domain_nr(tmp->bus),
1613 static int domain_context_mapped(struct pci_dev *pdev)
1616 struct pci_dev *tmp, *parent;
1617 struct intel_iommu *iommu;
1619 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
1624 ret = device_context_mapped(iommu, pdev->bus->number, pdev->devfn);
1627 /* dependent device mapping */
1628 tmp = pci_find_upstream_pcie_bridge(pdev);
1631 /* Secondary interface's bus number and devfn 0 */
1632 parent = pdev->bus->self;
1633 while (parent != tmp) {
1634 ret = device_context_mapped(iommu, parent->bus->number,
1638 parent = parent->bus->self;
1641 return device_context_mapped(iommu, tmp->subordinate->number,
1644 return device_context_mapped(iommu, tmp->bus->number,
1648 static int __domain_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1649 struct scatterlist *sg, unsigned long phys_pfn,
1650 unsigned long nr_pages, int prot)
1652 struct dma_pte *first_pte = NULL, *pte = NULL;
1653 phys_addr_t uninitialized_var(pteval);
1654 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
1655 unsigned long sg_res;
1657 BUG_ON(addr_width < BITS_PER_LONG && (iov_pfn + nr_pages - 1) >> addr_width);
1659 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1662 prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
1667 sg_res = nr_pages + 1;
1668 pteval = ((phys_addr_t)phys_pfn << VTD_PAGE_SHIFT) | prot;
1671 while (nr_pages--) {
1675 sg_res = (sg->offset + sg->length + VTD_PAGE_SIZE - 1) >> VTD_PAGE_SHIFT;
1676 sg->dma_address = ((dma_addr_t)iov_pfn << VTD_PAGE_SHIFT) + sg->offset;
1677 sg->dma_length = sg->length;
1678 pteval = page_to_phys(sg_page(sg)) | prot;
1681 first_pte = pte = pfn_to_dma_pte(domain, iov_pfn);
1685 /* We don't need lock here, nobody else
1686 * touches the iova range
1688 tmp = cmpxchg64_local(&pte->val, 0ULL, pteval);
1690 static int dumps = 5;
1691 printk(KERN_CRIT "ERROR: DMA PTE for vPFN 0x%lx already set (to %llx not %llx)\n",
1692 iov_pfn, tmp, (unsigned long long)pteval);
1695 debug_dma_dump_mappings(NULL);
1700 if (!nr_pages || first_pte_in_page(pte)) {
1701 domain_flush_cache(domain, first_pte,
1702 (void *)pte - (void *)first_pte);
1706 pteval += VTD_PAGE_SIZE;
1714 static inline int domain_sg_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1715 struct scatterlist *sg, unsigned long nr_pages,
1718 return __domain_mapping(domain, iov_pfn, sg, 0, nr_pages, prot);
1721 static inline int domain_pfn_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1722 unsigned long phys_pfn, unsigned long nr_pages,
1725 return __domain_mapping(domain, iov_pfn, NULL, phys_pfn, nr_pages, prot);
1728 static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
1733 clear_context_table(iommu, bus, devfn);
1734 iommu->flush.flush_context(iommu, 0, 0, 0,
1735 DMA_CCMD_GLOBAL_INVL);
1736 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
1739 static void domain_remove_dev_info(struct dmar_domain *domain)
1741 struct device_domain_info *info;
1742 unsigned long flags;
1743 struct intel_iommu *iommu;
1745 spin_lock_irqsave(&device_domain_lock, flags);
1746 while (!list_empty(&domain->devices)) {
1747 info = list_entry(domain->devices.next,
1748 struct device_domain_info, link);
1749 list_del(&info->link);
1750 list_del(&info->global);
1752 info->dev->dev.archdata.iommu = NULL;
1753 spin_unlock_irqrestore(&device_domain_lock, flags);
1755 iommu_disable_dev_iotlb(info);
1756 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
1757 iommu_detach_dev(iommu, info->bus, info->devfn);
1758 free_devinfo_mem(info);
1760 spin_lock_irqsave(&device_domain_lock, flags);
1762 spin_unlock_irqrestore(&device_domain_lock, flags);
1767 * Note: we use struct pci_dev->dev.archdata.iommu stores the info
1769 static struct dmar_domain *
1770 find_domain(struct pci_dev *pdev)
1772 struct device_domain_info *info;
1774 /* No lock here, assumes no domain exit in normal case */
1775 info = pdev->dev.archdata.iommu;
1777 return info->domain;
1781 /* domain is initialized */
1782 static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1784 struct dmar_domain *domain, *found = NULL;
1785 struct intel_iommu *iommu;
1786 struct dmar_drhd_unit *drhd;
1787 struct device_domain_info *info, *tmp;
1788 struct pci_dev *dev_tmp;
1789 unsigned long flags;
1790 int bus = 0, devfn = 0;
1794 domain = find_domain(pdev);
1798 segment = pci_domain_nr(pdev->bus);
1800 dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1802 if (dev_tmp->is_pcie) {
1803 bus = dev_tmp->subordinate->number;
1806 bus = dev_tmp->bus->number;
1807 devfn = dev_tmp->devfn;
1809 spin_lock_irqsave(&device_domain_lock, flags);
1810 list_for_each_entry(info, &device_domain_list, global) {
1811 if (info->segment == segment &&
1812 info->bus == bus && info->devfn == devfn) {
1813 found = info->domain;
1817 spin_unlock_irqrestore(&device_domain_lock, flags);
1818 /* pcie-pci bridge already has a domain, uses it */
1825 domain = alloc_domain();
1829 /* Allocate new domain for the device */
1830 drhd = dmar_find_matched_drhd_unit(pdev);
1832 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1836 iommu = drhd->iommu;
1838 ret = iommu_attach_domain(domain, iommu);
1840 domain_exit(domain);
1844 if (domain_init(domain, gaw)) {
1845 domain_exit(domain);
1849 /* register pcie-to-pci device */
1851 info = alloc_devinfo_mem();
1853 domain_exit(domain);
1856 info->segment = segment;
1858 info->devfn = devfn;
1860 info->domain = domain;
1861 /* This domain is shared by devices under p2p bridge */
1862 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
1864 /* pcie-to-pci bridge already has a domain, uses it */
1866 spin_lock_irqsave(&device_domain_lock, flags);
1867 list_for_each_entry(tmp, &device_domain_list, global) {
1868 if (tmp->segment == segment &&
1869 tmp->bus == bus && tmp->devfn == devfn) {
1870 found = tmp->domain;
1875 free_devinfo_mem(info);
1876 domain_exit(domain);
1879 list_add(&info->link, &domain->devices);
1880 list_add(&info->global, &device_domain_list);
1882 spin_unlock_irqrestore(&device_domain_lock, flags);
1886 info = alloc_devinfo_mem();
1889 info->segment = segment;
1890 info->bus = pdev->bus->number;
1891 info->devfn = pdev->devfn;
1893 info->domain = domain;
1894 spin_lock_irqsave(&device_domain_lock, flags);
1895 /* somebody is fast */
1896 found = find_domain(pdev);
1897 if (found != NULL) {
1898 spin_unlock_irqrestore(&device_domain_lock, flags);
1899 if (found != domain) {
1900 domain_exit(domain);
1903 free_devinfo_mem(info);
1906 list_add(&info->link, &domain->devices);
1907 list_add(&info->global, &device_domain_list);
1908 pdev->dev.archdata.iommu = info;
1909 spin_unlock_irqrestore(&device_domain_lock, flags);
1912 /* recheck it here, maybe others set it */
1913 return find_domain(pdev);
1916 static int iommu_identity_mapping;
1918 static int iommu_domain_identity_map(struct dmar_domain *domain,
1919 unsigned long long start,
1920 unsigned long long end)
1922 unsigned long first_vpfn = start >> VTD_PAGE_SHIFT;
1923 unsigned long last_vpfn = end >> VTD_PAGE_SHIFT;
1925 if (!reserve_iova(&domain->iovad, dma_to_mm_pfn(first_vpfn),
1926 dma_to_mm_pfn(last_vpfn))) {
1927 printk(KERN_ERR "IOMMU: reserve iova failed\n");
1931 pr_debug("Mapping reserved region %llx-%llx for domain %d\n",
1932 start, end, domain->id);
1934 * RMRR range might have overlap with physical memory range,
1937 dma_pte_clear_range(domain, first_vpfn, last_vpfn);
1939 return domain_pfn_mapping(domain, first_vpfn, first_vpfn,
1940 last_vpfn - first_vpfn + 1,
1941 DMA_PTE_READ|DMA_PTE_WRITE);
1944 static int iommu_prepare_identity_map(struct pci_dev *pdev,
1945 unsigned long long start,
1946 unsigned long long end)
1948 struct dmar_domain *domain;
1952 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1953 pci_name(pdev), start, end);
1955 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1959 ret = iommu_domain_identity_map(domain, start, end);
1963 /* context entry init */
1964 ret = domain_context_mapping(domain, pdev, CONTEXT_TT_MULTI_LEVEL);
1971 domain_exit(domain);
1975 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1976 struct pci_dev *pdev)
1978 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1980 return iommu_prepare_identity_map(pdev, rmrr->base_address,
1981 rmrr->end_address + 1);
1984 #ifdef CONFIG_DMAR_FLOPPY_WA
1985 static inline void iommu_prepare_isa(void)
1987 struct pci_dev *pdev;
1990 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1994 printk(KERN_INFO "IOMMU: Prepare 0-16MiB unity mapping for LPC\n");
1995 ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1998 printk(KERN_ERR "IOMMU: Failed to create 0-16MiB identity map; "
1999 "floppy might not work\n");
2003 static inline void iommu_prepare_isa(void)
2007 #endif /* !CONFIG_DMAR_FLPY_WA */
2009 /* Initialize each context entry as pass through.*/
2010 static int __init init_context_pass_through(void)
2012 struct pci_dev *pdev = NULL;
2013 struct dmar_domain *domain;
2016 for_each_pci_dev(pdev) {
2017 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2018 ret = domain_context_mapping(domain, pdev,
2019 CONTEXT_TT_PASS_THROUGH);
2026 static int md_domain_init(struct dmar_domain *domain, int guest_width);
2028 static int __init si_domain_work_fn(unsigned long start_pfn,
2029 unsigned long end_pfn, void *datax)
2033 *ret = iommu_domain_identity_map(si_domain,
2034 (uint64_t)start_pfn << PAGE_SHIFT,
2035 (uint64_t)end_pfn << PAGE_SHIFT);
2040 static int si_domain_init(void)
2042 struct dmar_drhd_unit *drhd;
2043 struct intel_iommu *iommu;
2046 si_domain = alloc_domain();
2050 pr_debug("Identity mapping domain is domain %d\n", si_domain->id);
2052 for_each_active_iommu(iommu, drhd) {
2053 ret = iommu_attach_domain(si_domain, iommu);
2055 domain_exit(si_domain);
2060 if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2061 domain_exit(si_domain);
2065 si_domain->flags = DOMAIN_FLAG_STATIC_IDENTITY;
2067 for_each_online_node(nid) {
2068 work_with_active_regions(nid, si_domain_work_fn, &ret);
2076 static void domain_remove_one_dev_info(struct dmar_domain *domain,
2077 struct pci_dev *pdev);
2078 static int identity_mapping(struct pci_dev *pdev)
2080 struct device_domain_info *info;
2082 if (likely(!iommu_identity_mapping))
2086 list_for_each_entry(info, &si_domain->devices, link)
2087 if (info->dev == pdev)
2092 static int domain_add_dev_info(struct dmar_domain *domain,
2093 struct pci_dev *pdev)
2095 struct device_domain_info *info;
2096 unsigned long flags;
2098 info = alloc_devinfo_mem();
2102 info->segment = pci_domain_nr(pdev->bus);
2103 info->bus = pdev->bus->number;
2104 info->devfn = pdev->devfn;
2106 info->domain = domain;
2108 spin_lock_irqsave(&device_domain_lock, flags);
2109 list_add(&info->link, &domain->devices);
2110 list_add(&info->global, &device_domain_list);
2111 pdev->dev.archdata.iommu = info;
2112 spin_unlock_irqrestore(&device_domain_lock, flags);
2117 static int iommu_should_identity_map(struct pci_dev *pdev, int startup)
2119 if (iommu_identity_mapping == 2)
2120 return IS_GFX_DEVICE(pdev);
2123 * We want to start off with all devices in the 1:1 domain, and
2124 * take them out later if we find they can't access all of memory.
2126 * However, we can't do this for PCI devices behind bridges,
2127 * because all PCI devices behind the same bridge will end up
2128 * with the same source-id on their transactions.
2130 * Practically speaking, we can't change things around for these
2131 * devices at run-time, because we can't be sure there'll be no
2132 * DMA transactions in flight for any of their siblings.
2134 * So PCI devices (unless they're on the root bus) as well as
2135 * their parent PCI-PCI or PCIe-PCI bridges must be left _out_ of
2136 * the 1:1 domain, just in _case_ one of their siblings turns out
2137 * not to be able to map all of memory.
2139 if (!pdev->is_pcie) {
2140 if (!pci_is_root_bus(pdev->bus))
2142 if (pdev->class >> 8 == PCI_CLASS_BRIDGE_PCI)
2144 } else if (pdev->pcie_type == PCI_EXP_TYPE_PCI_BRIDGE)
2148 * At boot time, we don't yet know if devices will be 64-bit capable.
2149 * Assume that they will -- if they turn out not to be, then we can
2150 * take them out of the 1:1 domain later.
2153 return pdev->dma_mask > DMA_BIT_MASK(32);
2158 static int iommu_prepare_static_identity_mapping(void)
2160 struct pci_dev *pdev = NULL;
2163 ret = si_domain_init();
2167 for_each_pci_dev(pdev) {
2168 if (iommu_should_identity_map(pdev, 1)) {
2169 printk(KERN_INFO "IOMMU: identity mapping for device %s\n",
2172 ret = domain_context_mapping(si_domain, pdev,
2173 CONTEXT_TT_MULTI_LEVEL);
2176 ret = domain_add_dev_info(si_domain, pdev);
2185 int __init init_dmars(void)
2187 struct dmar_drhd_unit *drhd;
2188 struct dmar_rmrr_unit *rmrr;
2189 struct pci_dev *pdev;
2190 struct intel_iommu *iommu;
2192 int pass_through = 1;
2195 * In case pass through can not be enabled, iommu tries to use identity
2198 if (iommu_pass_through)
2199 iommu_identity_mapping = 1;
2204 * initialize and program root entry to not present
2207 for_each_drhd_unit(drhd) {
2210 * lock not needed as this is only incremented in the single
2211 * threaded kernel __init code path all other access are read
2216 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
2219 printk(KERN_ERR "Allocating global iommu array failed\n");
2224 deferred_flush = kzalloc(g_num_of_iommus *
2225 sizeof(struct deferred_flush_tables), GFP_KERNEL);
2226 if (!deferred_flush) {
2232 for_each_drhd_unit(drhd) {
2236 iommu = drhd->iommu;
2237 g_iommus[iommu->seq_id] = iommu;
2239 ret = iommu_init_domains(iommu);
2245 * we could share the same root & context tables
2246 * amoung all IOMMU's. Need to Split it later.
2248 ret = iommu_alloc_root_entry(iommu);
2250 printk(KERN_ERR "IOMMU: allocate root entry failed\n");
2253 if (!ecap_pass_through(iommu->ecap))
2256 if (iommu_pass_through)
2257 if (!pass_through) {
2259 "Pass Through is not supported by hardware.\n");
2260 iommu_pass_through = 0;
2264 * Start from the sane iommu hardware state.
2266 for_each_drhd_unit(drhd) {
2270 iommu = drhd->iommu;
2273 * If the queued invalidation is already initialized by us
2274 * (for example, while enabling interrupt-remapping) then
2275 * we got the things already rolling from a sane state.
2281 * Clear any previous faults.
2283 dmar_fault(-1, iommu);
2285 * Disable queued invalidation if supported and already enabled
2286 * before OS handover.
2288 dmar_disable_qi(iommu);
2291 for_each_drhd_unit(drhd) {
2295 iommu = drhd->iommu;
2297 if (dmar_enable_qi(iommu)) {
2299 * Queued Invalidate not enabled, use Register Based
2302 iommu->flush.flush_context = __iommu_flush_context;
2303 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2304 printk(KERN_INFO "IOMMU 0x%Lx: using Register based "
2306 (unsigned long long)drhd->reg_base_addr);
2308 iommu->flush.flush_context = qi_flush_context;
2309 iommu->flush.flush_iotlb = qi_flush_iotlb;
2310 printk(KERN_INFO "IOMMU 0x%Lx: using Queued "
2312 (unsigned long long)drhd->reg_base_addr);
2317 * If pass through is set and enabled, context entries of all pci
2318 * devices are intialized by pass through translation type.
2320 if (iommu_pass_through) {
2321 ret = init_context_pass_through();
2323 printk(KERN_ERR "IOMMU: Pass through init failed.\n");
2324 iommu_pass_through = 0;
2329 * If pass through is not set or not enabled, setup context entries for
2330 * identity mappings for rmrr, gfx, and isa and may fall back to static
2331 * identity mapping if iommu_identity_mapping is set.
2333 if (!iommu_pass_through) {
2334 #ifdef CONFIG_DMAR_BROKEN_GFX_WA
2335 if (!iommu_identity_mapping)
2336 iommu_identity_mapping = 2;
2338 if (iommu_identity_mapping)
2339 iommu_prepare_static_identity_mapping();
2342 * for each dev attached to rmrr
2344 * locate drhd for dev, alloc domain for dev
2345 * allocate free domain
2346 * allocate page table entries for rmrr
2347 * if context not allocated for bus
2348 * allocate and init context
2349 * set present in root table for this bus
2350 * init context with domain, translation etc
2354 printk(KERN_INFO "IOMMU: Setting RMRR:\n");
2355 for_each_rmrr_units(rmrr) {
2356 for (i = 0; i < rmrr->devices_cnt; i++) {
2357 pdev = rmrr->devices[i];
2359 * some BIOS lists non-exist devices in DMAR
2364 ret = iommu_prepare_rmrr_dev(rmrr, pdev);
2367 "IOMMU: mapping reserved region failed\n");
2371 iommu_prepare_isa();
2377 * global invalidate context cache
2378 * global invalidate iotlb
2379 * enable translation
2381 for_each_drhd_unit(drhd) {
2384 iommu = drhd->iommu;
2386 iommu_flush_write_buffer(iommu);
2388 ret = dmar_set_interrupt(iommu);
2392 iommu_set_root_entry(iommu);
2394 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
2395 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2396 iommu_disable_protect_mem_regions(iommu);
2398 ret = iommu_enable_translation(iommu);
2405 for_each_drhd_unit(drhd) {
2408 iommu = drhd->iommu;
2415 /* Returns a number of VTD pages, but aligned to MM page size */
2416 static inline unsigned long aligned_nrpages(unsigned long host_addr,
2419 host_addr &= ~PAGE_MASK;
2420 return PAGE_ALIGN(host_addr + size) >> VTD_PAGE_SHIFT;
2423 /* This takes a number of _MM_ pages, not VTD pages */
2424 static struct iova *intel_alloc_iova(struct device *dev,
2425 struct dmar_domain *domain,
2426 unsigned long nrpages, uint64_t dma_mask)
2428 struct pci_dev *pdev = to_pci_dev(dev);
2429 struct iova *iova = NULL;
2431 /* Restrict dma_mask to the width that the iommu can handle */
2432 dma_mask = min_t(uint64_t, DOMAIN_MAX_ADDR(domain->gaw), dma_mask);
2434 if (!dmar_forcedac && dma_mask > DMA_BIT_MASK(32)) {
2436 * First try to allocate an io virtual address in
2437 * DMA_BIT_MASK(32) and if that fails then try allocating
2440 iova = alloc_iova(&domain->iovad, nrpages,
2441 IOVA_PFN(DMA_BIT_MASK(32)), 1);
2445 iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
2446 if (unlikely(!iova)) {
2447 printk(KERN_ERR "Allocating %ld-page iova for %s failed",
2448 nrpages, pci_name(pdev));
2455 static struct dmar_domain *__get_valid_domain_for_dev(struct pci_dev *pdev)
2457 struct dmar_domain *domain;
2460 domain = get_domain_for_dev(pdev,
2461 DEFAULT_DOMAIN_ADDRESS_WIDTH);
2464 "Allocating domain for %s failed", pci_name(pdev));
2468 /* make sure context mapping is ok */
2469 if (unlikely(!domain_context_mapped(pdev))) {
2470 ret = domain_context_mapping(domain, pdev,
2471 CONTEXT_TT_MULTI_LEVEL);
2474 "Domain context map for %s failed",
2483 static inline struct dmar_domain *get_valid_domain_for_dev(struct pci_dev *dev)
2485 struct device_domain_info *info;
2487 /* No lock here, assumes no domain exit in normal case */
2488 info = dev->dev.archdata.iommu;
2490 return info->domain;
2492 return __get_valid_domain_for_dev(dev);
2495 static int iommu_dummy(struct pci_dev *pdev)
2497 return pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
2500 /* Check if the pdev needs to go through non-identity map and unmap process.*/
2501 static int iommu_no_mapping(struct device *dev)
2503 struct pci_dev *pdev;
2506 if (unlikely(dev->bus != &pci_bus_type))
2509 pdev = to_pci_dev(dev);
2510 if (iommu_dummy(pdev))
2513 if (!iommu_identity_mapping)
2516 found = identity_mapping(pdev);
2518 if (iommu_should_identity_map(pdev, 0))
2522 * 32 bit DMA is removed from si_domain and fall back
2523 * to non-identity mapping.
2525 domain_remove_one_dev_info(si_domain, pdev);
2526 printk(KERN_INFO "32bit %s uses non-identity mapping\n",
2532 * In case of a detached 64 bit DMA device from vm, the device
2533 * is put into si_domain for identity mapping.
2535 if (iommu_should_identity_map(pdev, 0)) {
2537 ret = domain_add_dev_info(si_domain, pdev);
2540 ret = domain_context_mapping(si_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
2542 printk(KERN_INFO "64bit %s uses identity mapping\n",
2552 static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
2553 size_t size, int dir, u64 dma_mask)
2555 struct pci_dev *pdev = to_pci_dev(hwdev);
2556 struct dmar_domain *domain;
2557 phys_addr_t start_paddr;
2561 struct intel_iommu *iommu;
2563 BUG_ON(dir == DMA_NONE);
2565 if (iommu_no_mapping(hwdev))
2568 domain = get_valid_domain_for_dev(pdev);
2572 iommu = domain_get_iommu(domain);
2573 size = aligned_nrpages(paddr, size);
2575 iova = intel_alloc_iova(hwdev, domain, dma_to_mm_pfn(size),
2581 * Check if DMAR supports zero-length reads on write only
2584 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2585 !cap_zlr(iommu->cap))
2586 prot |= DMA_PTE_READ;
2587 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2588 prot |= DMA_PTE_WRITE;
2590 * paddr - (paddr + size) might be partial page, we should map the whole
2591 * page. Note: if two part of one page are separately mapped, we
2592 * might have two guest_addr mapping to the same host paddr, but this
2593 * is not a big problem
2595 ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
2596 paddr >> VTD_PAGE_SHIFT, size, prot);
2600 /* it's a non-present to present mapping. Only flush if caching mode */
2601 if (cap_caching_mode(iommu->cap))
2602 iommu_flush_iotlb_psi(iommu, 0, mm_to_dma_pfn(iova->pfn_lo), size);
2604 iommu_flush_write_buffer(iommu);
2606 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
2607 start_paddr += paddr & ~PAGE_MASK;
2612 __free_iova(&domain->iovad, iova);
2613 printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
2614 pci_name(pdev), size, (unsigned long long)paddr, dir);
2618 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
2619 unsigned long offset, size_t size,
2620 enum dma_data_direction dir,
2621 struct dma_attrs *attrs)
2623 return __intel_map_single(dev, page_to_phys(page) + offset, size,
2624 dir, to_pci_dev(dev)->dma_mask);
2627 static void flush_unmaps(void)
2633 /* just flush them all */
2634 for (i = 0; i < g_num_of_iommus; i++) {
2635 struct intel_iommu *iommu = g_iommus[i];
2639 if (!deferred_flush[i].next)
2642 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2643 DMA_TLB_GLOBAL_FLUSH);
2644 for (j = 0; j < deferred_flush[i].next; j++) {
2646 struct iova *iova = deferred_flush[i].iova[j];
2648 mask = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT;
2649 mask = ilog2(mask >> VTD_PAGE_SHIFT);
2650 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
2651 iova->pfn_lo << PAGE_SHIFT, mask);
2652 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
2654 deferred_flush[i].next = 0;
2660 static void flush_unmaps_timeout(unsigned long data)
2662 unsigned long flags;
2664 spin_lock_irqsave(&async_umap_flush_lock, flags);
2666 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2669 static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2671 unsigned long flags;
2673 struct intel_iommu *iommu;
2675 spin_lock_irqsave(&async_umap_flush_lock, flags);
2676 if (list_size == HIGH_WATER_MARK)
2679 iommu = domain_get_iommu(dom);
2680 iommu_id = iommu->seq_id;
2682 next = deferred_flush[iommu_id].next;
2683 deferred_flush[iommu_id].domain[next] = dom;
2684 deferred_flush[iommu_id].iova[next] = iova;
2685 deferred_flush[iommu_id].next++;
2688 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2692 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2695 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
2696 size_t size, enum dma_data_direction dir,
2697 struct dma_attrs *attrs)
2699 struct pci_dev *pdev = to_pci_dev(dev);
2700 struct dmar_domain *domain;
2701 unsigned long start_pfn, last_pfn;
2703 struct intel_iommu *iommu;
2705 if (iommu_no_mapping(dev))
2708 domain = find_domain(pdev);
2711 iommu = domain_get_iommu(domain);
2713 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
2714 if (WARN_ONCE(!iova, "Driver unmaps unmatched page at PFN %llx\n",
2715 (unsigned long long)dev_addr))
2718 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
2719 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
2721 pr_debug("Device %s unmapping: pfn %lx-%lx\n",
2722 pci_name(pdev), start_pfn, last_pfn);
2724 /* clear the whole page */
2725 dma_pte_clear_range(domain, start_pfn, last_pfn);
2727 /* free page tables */
2728 dma_pte_free_pagetable(domain, start_pfn, last_pfn);
2730 if (intel_iommu_strict) {
2731 iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
2732 last_pfn - start_pfn + 1);
2734 __free_iova(&domain->iovad, iova);
2736 add_unmap(domain, iova);
2738 * queue up the release of the unmap to save the 1/6th of the
2739 * cpu used up by the iotlb flush operation...
2744 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr, size_t size,
2747 intel_unmap_page(dev, dev_addr, size, dir, NULL);
2750 static void *intel_alloc_coherent(struct device *hwdev, size_t size,
2751 dma_addr_t *dma_handle, gfp_t flags)
2756 size = PAGE_ALIGN(size);
2757 order = get_order(size);
2758 flags &= ~(GFP_DMA | GFP_DMA32);
2760 vaddr = (void *)__get_free_pages(flags, order);
2763 memset(vaddr, 0, size);
2765 *dma_handle = __intel_map_single(hwdev, virt_to_bus(vaddr), size,
2767 hwdev->coherent_dma_mask);
2770 free_pages((unsigned long)vaddr, order);
2774 static void intel_free_coherent(struct device *hwdev, size_t size, void *vaddr,
2775 dma_addr_t dma_handle)
2779 size = PAGE_ALIGN(size);
2780 order = get_order(size);
2782 intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2783 free_pages((unsigned long)vaddr, order);
2786 static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2787 int nelems, enum dma_data_direction dir,
2788 struct dma_attrs *attrs)
2790 struct pci_dev *pdev = to_pci_dev(hwdev);
2791 struct dmar_domain *domain;
2792 unsigned long start_pfn, last_pfn;
2794 struct intel_iommu *iommu;
2796 if (iommu_no_mapping(hwdev))
2799 domain = find_domain(pdev);
2802 iommu = domain_get_iommu(domain);
2804 iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
2805 if (WARN_ONCE(!iova, "Driver unmaps unmatched sglist at PFN %llx\n",
2806 (unsigned long long)sglist[0].dma_address))
2809 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
2810 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
2812 /* clear the whole page */
2813 dma_pte_clear_range(domain, start_pfn, last_pfn);
2815 /* free page tables */
2816 dma_pte_free_pagetable(domain, start_pfn, last_pfn);
2818 if (intel_iommu_strict) {
2819 iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
2820 last_pfn - start_pfn + 1);
2822 __free_iova(&domain->iovad, iova);
2824 add_unmap(domain, iova);
2826 * queue up the release of the unmap to save the 1/6th of the
2827 * cpu used up by the iotlb flush operation...
2832 static int intel_nontranslate_map_sg(struct device *hddev,
2833 struct scatterlist *sglist, int nelems, int dir)
2836 struct scatterlist *sg;
2838 for_each_sg(sglist, sg, nelems, i) {
2839 BUG_ON(!sg_page(sg));
2840 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
2841 sg->dma_length = sg->length;
2846 static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist, int nelems,
2847 enum dma_data_direction dir, struct dma_attrs *attrs)
2850 struct pci_dev *pdev = to_pci_dev(hwdev);
2851 struct dmar_domain *domain;
2854 size_t offset_pfn = 0;
2855 struct iova *iova = NULL;
2857 struct scatterlist *sg;
2858 unsigned long start_vpfn;
2859 struct intel_iommu *iommu;
2861 BUG_ON(dir == DMA_NONE);
2862 if (iommu_no_mapping(hwdev))
2863 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
2865 domain = get_valid_domain_for_dev(pdev);
2869 iommu = domain_get_iommu(domain);
2871 for_each_sg(sglist, sg, nelems, i)
2872 size += aligned_nrpages(sg->offset, sg->length);
2874 iova = intel_alloc_iova(hwdev, domain, dma_to_mm_pfn(size),
2877 sglist->dma_length = 0;
2882 * Check if DMAR supports zero-length reads on write only
2885 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2886 !cap_zlr(iommu->cap))
2887 prot |= DMA_PTE_READ;
2888 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2889 prot |= DMA_PTE_WRITE;
2891 start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
2893 ret = domain_sg_mapping(domain, start_vpfn, sglist, mm_to_dma_pfn(size), prot);
2894 if (unlikely(ret)) {
2895 /* clear the page */
2896 dma_pte_clear_range(domain, start_vpfn,
2897 start_vpfn + size - 1);
2898 /* free page tables */
2899 dma_pte_free_pagetable(domain, start_vpfn,
2900 start_vpfn + size - 1);
2902 __free_iova(&domain->iovad, iova);
2906 /* it's a non-present to present mapping. Only flush if caching mode */
2907 if (cap_caching_mode(iommu->cap))
2908 iommu_flush_iotlb_psi(iommu, 0, start_vpfn, offset_pfn);
2910 iommu_flush_write_buffer(iommu);
2915 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
2920 struct dma_map_ops intel_dma_ops = {
2921 .alloc_coherent = intel_alloc_coherent,
2922 .free_coherent = intel_free_coherent,
2923 .map_sg = intel_map_sg,
2924 .unmap_sg = intel_unmap_sg,
2925 .map_page = intel_map_page,
2926 .unmap_page = intel_unmap_page,
2927 .mapping_error = intel_mapping_error,
2930 static inline int iommu_domain_cache_init(void)
2934 iommu_domain_cache = kmem_cache_create("iommu_domain",
2935 sizeof(struct dmar_domain),
2940 if (!iommu_domain_cache) {
2941 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2948 static inline int iommu_devinfo_cache_init(void)
2952 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2953 sizeof(struct device_domain_info),
2957 if (!iommu_devinfo_cache) {
2958 printk(KERN_ERR "Couldn't create devinfo cache\n");
2965 static inline int iommu_iova_cache_init(void)
2969 iommu_iova_cache = kmem_cache_create("iommu_iova",
2970 sizeof(struct iova),
2974 if (!iommu_iova_cache) {
2975 printk(KERN_ERR "Couldn't create iova cache\n");
2982 static int __init iommu_init_mempool(void)
2985 ret = iommu_iova_cache_init();
2989 ret = iommu_domain_cache_init();
2993 ret = iommu_devinfo_cache_init();
2997 kmem_cache_destroy(iommu_domain_cache);
2999 kmem_cache_destroy(iommu_iova_cache);
3004 static void __init iommu_exit_mempool(void)
3006 kmem_cache_destroy(iommu_devinfo_cache);
3007 kmem_cache_destroy(iommu_domain_cache);
3008 kmem_cache_destroy(iommu_iova_cache);
3012 static void __init init_no_remapping_devices(void)
3014 struct dmar_drhd_unit *drhd;
3016 for_each_drhd_unit(drhd) {
3017 if (!drhd->include_all) {
3019 for (i = 0; i < drhd->devices_cnt; i++)
3020 if (drhd->devices[i] != NULL)
3022 /* ignore DMAR unit if no pci devices exist */
3023 if (i == drhd->devices_cnt)
3031 for_each_drhd_unit(drhd) {
3033 if (drhd->ignored || drhd->include_all)
3036 for (i = 0; i < drhd->devices_cnt; i++)
3037 if (drhd->devices[i] &&
3038 !IS_GFX_DEVICE(drhd->devices[i]))
3041 if (i < drhd->devices_cnt)
3044 /* bypass IOMMU if it is just for gfx devices */
3046 for (i = 0; i < drhd->devices_cnt; i++) {
3047 if (!drhd->devices[i])
3049 drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3054 #ifdef CONFIG_SUSPEND
3055 static int init_iommu_hw(void)
3057 struct dmar_drhd_unit *drhd;
3058 struct intel_iommu *iommu = NULL;
3060 for_each_active_iommu(iommu, drhd)
3062 dmar_reenable_qi(iommu);
3064 for_each_active_iommu(iommu, drhd) {
3065 iommu_flush_write_buffer(iommu);
3067 iommu_set_root_entry(iommu);
3069 iommu->flush.flush_context(iommu, 0, 0, 0,
3070 DMA_CCMD_GLOBAL_INVL);
3071 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3072 DMA_TLB_GLOBAL_FLUSH);
3073 iommu_disable_protect_mem_regions(iommu);
3074 iommu_enable_translation(iommu);
3080 static void iommu_flush_all(void)
3082 struct dmar_drhd_unit *drhd;
3083 struct intel_iommu *iommu;
3085 for_each_active_iommu(iommu, drhd) {
3086 iommu->flush.flush_context(iommu, 0, 0, 0,
3087 DMA_CCMD_GLOBAL_INVL);
3088 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3089 DMA_TLB_GLOBAL_FLUSH);
3093 static int iommu_suspend(struct sys_device *dev, pm_message_t state)
3095 struct dmar_drhd_unit *drhd;
3096 struct intel_iommu *iommu = NULL;
3099 for_each_active_iommu(iommu, drhd) {
3100 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3102 if (!iommu->iommu_state)
3108 for_each_active_iommu(iommu, drhd) {
3109 iommu_disable_translation(iommu);
3111 spin_lock_irqsave(&iommu->register_lock, flag);
3113 iommu->iommu_state[SR_DMAR_FECTL_REG] =
3114 readl(iommu->reg + DMAR_FECTL_REG);
3115 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
3116 readl(iommu->reg + DMAR_FEDATA_REG);
3117 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
3118 readl(iommu->reg + DMAR_FEADDR_REG);
3119 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
3120 readl(iommu->reg + DMAR_FEUADDR_REG);
3122 spin_unlock_irqrestore(&iommu->register_lock, flag);
3127 for_each_active_iommu(iommu, drhd)
3128 kfree(iommu->iommu_state);
3133 static int iommu_resume(struct sys_device *dev)
3135 struct dmar_drhd_unit *drhd;
3136 struct intel_iommu *iommu = NULL;
3139 if (init_iommu_hw()) {
3140 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
3144 for_each_active_iommu(iommu, drhd) {
3146 spin_lock_irqsave(&iommu->register_lock, flag);
3148 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
3149 iommu->reg + DMAR_FECTL_REG);
3150 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
3151 iommu->reg + DMAR_FEDATA_REG);
3152 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
3153 iommu->reg + DMAR_FEADDR_REG);
3154 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
3155 iommu->reg + DMAR_FEUADDR_REG);
3157 spin_unlock_irqrestore(&iommu->register_lock, flag);
3160 for_each_active_iommu(iommu, drhd)
3161 kfree(iommu->iommu_state);
3166 static struct sysdev_class iommu_sysclass = {
3168 .resume = iommu_resume,
3169 .suspend = iommu_suspend,
3172 static struct sys_device device_iommu = {
3173 .cls = &iommu_sysclass,
3176 static int __init init_iommu_sysfs(void)
3180 error = sysdev_class_register(&iommu_sysclass);
3184 error = sysdev_register(&device_iommu);
3186 sysdev_class_unregister(&iommu_sysclass);
3192 static int __init init_iommu_sysfs(void)
3196 #endif /* CONFIG_PM */
3198 int __init intel_iommu_init(void)
3202 if (dmar_table_init())
3205 if (dmar_dev_scope_init())
3209 * Check the need for DMA-remapping initialization now.
3210 * Above initialization will also be used by Interrupt-remapping.
3212 if (no_iommu || (swiotlb && !iommu_pass_through) || dmar_disabled)
3215 iommu_init_mempool();
3216 dmar_init_reserved_ranges();
3218 init_no_remapping_devices();
3222 printk(KERN_ERR "IOMMU: dmar init failed\n");
3223 put_iova_domain(&reserved_iova_list);
3224 iommu_exit_mempool();
3228 "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
3230 init_timer(&unmap_timer);
3233 if (!iommu_pass_through) {
3235 "Multi-level page-table translation for DMAR.\n");
3236 dma_ops = &intel_dma_ops;
3239 "DMAR: Pass through translation for DMAR.\n");
3243 register_iommu(&intel_iommu_ops);
3248 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
3249 struct pci_dev *pdev)
3251 struct pci_dev *tmp, *parent;
3253 if (!iommu || !pdev)
3256 /* dependent device detach */
3257 tmp = pci_find_upstream_pcie_bridge(pdev);
3258 /* Secondary interface's bus number and devfn 0 */
3260 parent = pdev->bus->self;
3261 while (parent != tmp) {
3262 iommu_detach_dev(iommu, parent->bus->number,
3264 parent = parent->bus->self;
3266 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
3267 iommu_detach_dev(iommu,
3268 tmp->subordinate->number, 0);
3269 else /* this is a legacy PCI bridge */
3270 iommu_detach_dev(iommu, tmp->bus->number,
3275 static void domain_remove_one_dev_info(struct dmar_domain *domain,
3276 struct pci_dev *pdev)
3278 struct device_domain_info *info;
3279 struct intel_iommu *iommu;
3280 unsigned long flags;
3282 struct list_head *entry, *tmp;
3284 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3289 spin_lock_irqsave(&device_domain_lock, flags);
3290 list_for_each_safe(entry, tmp, &domain->devices) {
3291 info = list_entry(entry, struct device_domain_info, link);
3292 /* No need to compare PCI domain; it has to be the same */
3293 if (info->bus == pdev->bus->number &&
3294 info->devfn == pdev->devfn) {
3295 list_del(&info->link);
3296 list_del(&info->global);
3298 info->dev->dev.archdata.iommu = NULL;
3299 spin_unlock_irqrestore(&device_domain_lock, flags);
3301 iommu_disable_dev_iotlb(info);
3302 iommu_detach_dev(iommu, info->bus, info->devfn);
3303 iommu_detach_dependent_devices(iommu, pdev);
3304 free_devinfo_mem(info);
3306 spin_lock_irqsave(&device_domain_lock, flags);
3314 /* if there is no other devices under the same iommu
3315 * owned by this domain, clear this iommu in iommu_bmp
3316 * update iommu count and coherency
3318 if (iommu == device_to_iommu(info->segment, info->bus,
3324 unsigned long tmp_flags;
3325 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
3326 clear_bit(iommu->seq_id, &domain->iommu_bmp);
3327 domain->iommu_count--;
3328 domain_update_iommu_cap(domain);
3329 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
3332 spin_unlock_irqrestore(&device_domain_lock, flags);
3335 static void vm_domain_remove_all_dev_info(struct dmar_domain *domain)
3337 struct device_domain_info *info;
3338 struct intel_iommu *iommu;
3339 unsigned long flags1, flags2;
3341 spin_lock_irqsave(&device_domain_lock, flags1);
3342 while (!list_empty(&domain->devices)) {
3343 info = list_entry(domain->devices.next,
3344 struct device_domain_info, link);
3345 list_del(&info->link);
3346 list_del(&info->global);
3348 info->dev->dev.archdata.iommu = NULL;
3350 spin_unlock_irqrestore(&device_domain_lock, flags1);
3352 iommu_disable_dev_iotlb(info);
3353 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
3354 iommu_detach_dev(iommu, info->bus, info->devfn);
3355 iommu_detach_dependent_devices(iommu, info->dev);
3357 /* clear this iommu in iommu_bmp, update iommu count
3360 spin_lock_irqsave(&domain->iommu_lock, flags2);
3361 if (test_and_clear_bit(iommu->seq_id,
3362 &domain->iommu_bmp)) {
3363 domain->iommu_count--;
3364 domain_update_iommu_cap(domain);
3366 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
3368 free_devinfo_mem(info);
3369 spin_lock_irqsave(&device_domain_lock, flags1);
3371 spin_unlock_irqrestore(&device_domain_lock, flags1);
3374 /* domain id for virtual machine, it won't be set in context */
3375 static unsigned long vm_domid;
3377 static int vm_domain_min_agaw(struct dmar_domain *domain)
3380 int min_agaw = domain->agaw;
3382 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
3383 for (; i < g_num_of_iommus; ) {
3384 if (min_agaw > g_iommus[i]->agaw)
3385 min_agaw = g_iommus[i]->agaw;
3387 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
3393 static struct dmar_domain *iommu_alloc_vm_domain(void)
3395 struct dmar_domain *domain;
3397 domain = alloc_domain_mem();
3401 domain->id = vm_domid++;
3402 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
3403 domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
3408 static int md_domain_init(struct dmar_domain *domain, int guest_width)
3412 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
3413 spin_lock_init(&domain->iommu_lock);
3415 domain_reserve_special_ranges(domain);
3417 /* calculate AGAW */
3418 domain->gaw = guest_width;
3419 adjust_width = guestwidth_to_adjustwidth(guest_width);
3420 domain->agaw = width_to_agaw(adjust_width);
3422 INIT_LIST_HEAD(&domain->devices);
3424 domain->iommu_count = 0;
3425 domain->iommu_coherency = 0;
3426 domain->max_addr = 0;
3428 /* always allocate the top pgd */
3429 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
3432 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
3436 static void iommu_free_vm_domain(struct dmar_domain *domain)
3438 unsigned long flags;
3439 struct dmar_drhd_unit *drhd;
3440 struct intel_iommu *iommu;
3442 unsigned long ndomains;
3444 for_each_drhd_unit(drhd) {
3447 iommu = drhd->iommu;
3449 ndomains = cap_ndoms(iommu->cap);
3450 i = find_first_bit(iommu->domain_ids, ndomains);
3451 for (; i < ndomains; ) {
3452 if (iommu->domains[i] == domain) {
3453 spin_lock_irqsave(&iommu->lock, flags);
3454 clear_bit(i, iommu->domain_ids);
3455 iommu->domains[i] = NULL;
3456 spin_unlock_irqrestore(&iommu->lock, flags);
3459 i = find_next_bit(iommu->domain_ids, ndomains, i+1);
3464 static void vm_domain_exit(struct dmar_domain *domain)
3466 /* Domain 0 is reserved, so dont process it */
3470 vm_domain_remove_all_dev_info(domain);
3472 put_iova_domain(&domain->iovad);
3475 dma_pte_clear_range(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
3477 /* free page tables */
3478 dma_pte_free_pagetable(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
3480 iommu_free_vm_domain(domain);
3481 free_domain_mem(domain);
3484 static int intel_iommu_domain_init(struct iommu_domain *domain)
3486 struct dmar_domain *dmar_domain;
3488 dmar_domain = iommu_alloc_vm_domain();
3491 "intel_iommu_domain_init: dmar_domain == NULL\n");
3494 if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
3496 "intel_iommu_domain_init() failed\n");
3497 vm_domain_exit(dmar_domain);
3500 domain->priv = dmar_domain;
3505 static void intel_iommu_domain_destroy(struct iommu_domain *domain)
3507 struct dmar_domain *dmar_domain = domain->priv;
3509 domain->priv = NULL;
3510 vm_domain_exit(dmar_domain);
3513 static int intel_iommu_attach_device(struct iommu_domain *domain,
3516 struct dmar_domain *dmar_domain = domain->priv;
3517 struct pci_dev *pdev = to_pci_dev(dev);
3518 struct intel_iommu *iommu;
3523 /* normally pdev is not mapped */
3524 if (unlikely(domain_context_mapped(pdev))) {
3525 struct dmar_domain *old_domain;
3527 old_domain = find_domain(pdev);
3529 if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
3530 dmar_domain->flags & DOMAIN_FLAG_STATIC_IDENTITY)
3531 domain_remove_one_dev_info(old_domain, pdev);
3533 domain_remove_dev_info(old_domain);
3537 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3542 /* check if this iommu agaw is sufficient for max mapped address */
3543 addr_width = agaw_to_width(iommu->agaw);
3544 end = DOMAIN_MAX_ADDR(addr_width);
3545 end = end & VTD_PAGE_MASK;
3546 if (end < dmar_domain->max_addr) {
3547 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3548 "sufficient for the mapped address (%llx)\n",
3549 __func__, iommu->agaw, dmar_domain->max_addr);
3553 ret = domain_add_dev_info(dmar_domain, pdev);
3557 ret = domain_context_mapping(dmar_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
3561 static void intel_iommu_detach_device(struct iommu_domain *domain,
3564 struct dmar_domain *dmar_domain = domain->priv;
3565 struct pci_dev *pdev = to_pci_dev(dev);
3567 domain_remove_one_dev_info(dmar_domain, pdev);
3570 static int intel_iommu_map_range(struct iommu_domain *domain,
3571 unsigned long iova, phys_addr_t hpa,
3572 size_t size, int iommu_prot)
3574 struct dmar_domain *dmar_domain = domain->priv;
3580 if (iommu_prot & IOMMU_READ)
3581 prot |= DMA_PTE_READ;
3582 if (iommu_prot & IOMMU_WRITE)
3583 prot |= DMA_PTE_WRITE;
3584 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
3585 prot |= DMA_PTE_SNP;
3587 max_addr = iova + size;
3588 if (dmar_domain->max_addr < max_addr) {
3592 /* check if minimum agaw is sufficient for mapped address */
3593 min_agaw = vm_domain_min_agaw(dmar_domain);
3594 addr_width = agaw_to_width(min_agaw);
3595 end = DOMAIN_MAX_ADDR(addr_width);
3596 end = end & VTD_PAGE_MASK;
3597 if (end < max_addr) {
3598 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3599 "sufficient for the mapped address (%llx)\n",
3600 __func__, min_agaw, max_addr);
3603 dmar_domain->max_addr = max_addr;
3605 /* Round up size to next multiple of PAGE_SIZE, if it and
3606 the low bits of hpa would take us onto the next page */
3607 size = aligned_nrpages(hpa, size);
3608 ret = domain_pfn_mapping(dmar_domain, iova >> VTD_PAGE_SHIFT,
3609 hpa >> VTD_PAGE_SHIFT, size, prot);
3613 static void intel_iommu_unmap_range(struct iommu_domain *domain,
3614 unsigned long iova, size_t size)
3616 struct dmar_domain *dmar_domain = domain->priv;
3618 dma_pte_clear_range(dmar_domain, iova >> VTD_PAGE_SHIFT,
3619 (iova + size - 1) >> VTD_PAGE_SHIFT);
3621 if (dmar_domain->max_addr == iova + size)
3622 dmar_domain->max_addr = iova;
3625 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
3628 struct dmar_domain *dmar_domain = domain->priv;
3629 struct dma_pte *pte;
3632 pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT);
3634 phys = dma_pte_addr(pte);
3639 static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
3642 struct dmar_domain *dmar_domain = domain->priv;
3644 if (cap == IOMMU_CAP_CACHE_COHERENCY)
3645 return dmar_domain->iommu_snooping;
3650 static struct iommu_ops intel_iommu_ops = {
3651 .domain_init = intel_iommu_domain_init,
3652 .domain_destroy = intel_iommu_domain_destroy,
3653 .attach_dev = intel_iommu_attach_device,
3654 .detach_dev = intel_iommu_detach_device,
3655 .map = intel_iommu_map_range,
3656 .unmap = intel_iommu_unmap_range,
3657 .iova_to_phys = intel_iommu_iova_to_phys,
3658 .domain_has_cap = intel_iommu_domain_has_cap,
3661 static void __devinit quirk_iommu_rwbf(struct pci_dev *dev)
3664 * Mobile 4 Series Chipset neglects to set RWBF capability,
3667 printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
3671 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);