2 * Copyright © 2006-2014 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 * Authors: David Woodhouse <dwmw2@infradead.org>,
14 * Ashok Raj <ashok.raj@intel.com>,
15 * Shaohua Li <shaohua.li@intel.com>,
16 * Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>,
17 * Fenghua Yu <fenghua.yu@intel.com>
20 #include <linux/init.h>
21 #include <linux/bitmap.h>
22 #include <linux/debugfs.h>
23 #include <linux/export.h>
24 #include <linux/slab.h>
25 #include <linux/irq.h>
26 #include <linux/interrupt.h>
27 #include <linux/spinlock.h>
28 #include <linux/pci.h>
29 #include <linux/dmar.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/mempool.h>
32 #include <linux/memory.h>
33 #include <linux/timer.h>
34 #include <linux/iova.h>
35 #include <linux/iommu.h>
36 #include <linux/intel-iommu.h>
37 #include <linux/syscore_ops.h>
38 #include <linux/tboot.h>
39 #include <linux/dmi.h>
40 #include <linux/pci-ats.h>
41 #include <linux/memblock.h>
42 #include <linux/dma-contiguous.h>
43 #include <asm/irq_remapping.h>
44 #include <asm/cacheflush.h>
45 #include <asm/iommu.h>
47 #include "irq_remapping.h"
50 #define ROOT_SIZE VTD_PAGE_SIZE
51 #define CONTEXT_SIZE VTD_PAGE_SIZE
53 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
54 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
55 #define IS_AZALIA(pdev) ((pdev)->vendor == 0x8086 && (pdev)->device == 0x3a3e)
57 #define IOAPIC_RANGE_START (0xfee00000)
58 #define IOAPIC_RANGE_END (0xfeefffff)
59 #define IOVA_START_ADDR (0x1000)
61 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
63 #define MAX_AGAW_WIDTH 64
64 #define MAX_AGAW_PFN_WIDTH (MAX_AGAW_WIDTH - VTD_PAGE_SHIFT)
66 #define __DOMAIN_MAX_PFN(gaw) ((((uint64_t)1) << (gaw-VTD_PAGE_SHIFT)) - 1)
67 #define __DOMAIN_MAX_ADDR(gaw) ((((uint64_t)1) << gaw) - 1)
69 /* We limit DOMAIN_MAX_PFN to fit in an unsigned long, and DOMAIN_MAX_ADDR
70 to match. That way, we can use 'unsigned long' for PFNs with impunity. */
71 #define DOMAIN_MAX_PFN(gaw) ((unsigned long) min_t(uint64_t, \
72 __DOMAIN_MAX_PFN(gaw), (unsigned long)-1))
73 #define DOMAIN_MAX_ADDR(gaw) (((uint64_t)__DOMAIN_MAX_PFN(gaw)) << VTD_PAGE_SHIFT)
75 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
76 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
77 #define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
79 /* page table handling */
80 #define LEVEL_STRIDE (9)
81 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
84 * This bitmap is used to advertise the page sizes our hardware support
85 * to the IOMMU core, which will then use this information to split
86 * physically contiguous memory regions it is mapping into page sizes
89 * Traditionally the IOMMU core just handed us the mappings directly,
90 * after making sure the size is an order of a 4KiB page and that the
91 * mapping has natural alignment.
93 * To retain this behavior, we currently advertise that we support
94 * all page sizes that are an order of 4KiB.
96 * If at some point we'd like to utilize the IOMMU core's new behavior,
97 * we could change this to advertise the real page sizes we support.
99 #define INTEL_IOMMU_PGSIZES (~0xFFFUL)
101 static inline int agaw_to_level(int agaw)
106 static inline int agaw_to_width(int agaw)
108 return min_t(int, 30 + agaw * LEVEL_STRIDE, MAX_AGAW_WIDTH);
111 static inline int width_to_agaw(int width)
113 return DIV_ROUND_UP(width - 30, LEVEL_STRIDE);
116 static inline unsigned int level_to_offset_bits(int level)
118 return (level - 1) * LEVEL_STRIDE;
121 static inline int pfn_level_offset(unsigned long pfn, int level)
123 return (pfn >> level_to_offset_bits(level)) & LEVEL_MASK;
126 static inline unsigned long level_mask(int level)
128 return -1UL << level_to_offset_bits(level);
131 static inline unsigned long level_size(int level)
133 return 1UL << level_to_offset_bits(level);
136 static inline unsigned long align_to_level(unsigned long pfn, int level)
138 return (pfn + level_size(level) - 1) & level_mask(level);
141 static inline unsigned long lvl_to_nr_pages(unsigned int lvl)
143 return 1 << min_t(int, (lvl - 1) * LEVEL_STRIDE, MAX_AGAW_PFN_WIDTH);
146 /* VT-d pages must always be _smaller_ than MM pages. Otherwise things
147 are never going to work. */
148 static inline unsigned long dma_to_mm_pfn(unsigned long dma_pfn)
150 return dma_pfn >> (PAGE_SHIFT - VTD_PAGE_SHIFT);
153 static inline unsigned long mm_to_dma_pfn(unsigned long mm_pfn)
155 return mm_pfn << (PAGE_SHIFT - VTD_PAGE_SHIFT);
157 static inline unsigned long page_to_dma_pfn(struct page *pg)
159 return mm_to_dma_pfn(page_to_pfn(pg));
161 static inline unsigned long virt_to_dma_pfn(void *p)
163 return page_to_dma_pfn(virt_to_page(p));
166 /* global iommu list, set NULL for ignored DMAR units */
167 static struct intel_iommu **g_iommus;
169 static void __init check_tylersburg_isoch(void);
170 static int rwbf_quirk;
173 * set to 1 to panic kernel if can't successfully enable VT-d
174 * (used when kernel is launched w/ TXT)
176 static int force_on = 0;
181 * 12-63: Context Ptr (12 - (haw-1))
188 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
189 static inline bool root_present(struct root_entry *root)
191 return (root->val & 1);
193 static inline void set_root_present(struct root_entry *root)
197 static inline void set_root_value(struct root_entry *root, unsigned long value)
199 root->val |= value & VTD_PAGE_MASK;
202 static inline struct context_entry *
203 get_context_addr_from_root(struct root_entry *root)
205 return (struct context_entry *)
206 (root_present(root)?phys_to_virt(
207 root->val & VTD_PAGE_MASK) :
214 * 1: fault processing disable
215 * 2-3: translation type
216 * 12-63: address space root
222 struct context_entry {
227 static inline bool context_present(struct context_entry *context)
229 return (context->lo & 1);
231 static inline void context_set_present(struct context_entry *context)
236 static inline void context_set_fault_enable(struct context_entry *context)
238 context->lo &= (((u64)-1) << 2) | 1;
241 static inline void context_set_translation_type(struct context_entry *context,
244 context->lo &= (((u64)-1) << 4) | 3;
245 context->lo |= (value & 3) << 2;
248 static inline void context_set_address_root(struct context_entry *context,
251 context->lo |= value & VTD_PAGE_MASK;
254 static inline void context_set_address_width(struct context_entry *context,
257 context->hi |= value & 7;
260 static inline void context_set_domain_id(struct context_entry *context,
263 context->hi |= (value & ((1 << 16) - 1)) << 8;
266 static inline void context_clear_entry(struct context_entry *context)
279 * 12-63: Host physcial address
285 static inline void dma_clear_pte(struct dma_pte *pte)
290 static inline u64 dma_pte_addr(struct dma_pte *pte)
293 return pte->val & VTD_PAGE_MASK;
295 /* Must have a full atomic 64-bit read */
296 return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
300 static inline bool dma_pte_present(struct dma_pte *pte)
302 return (pte->val & 3) != 0;
305 static inline bool dma_pte_superpage(struct dma_pte *pte)
307 return (pte->val & (1 << 7));
310 static inline int first_pte_in_page(struct dma_pte *pte)
312 return !((unsigned long)pte & ~VTD_PAGE_MASK);
316 * This domain is a statically identity mapping domain.
317 * 1. This domain creats a static 1:1 mapping to all usable memory.
318 * 2. It maps to each iommu if successful.
319 * 3. Each iommu mapps to this domain if successful.
321 static struct dmar_domain *si_domain;
322 static int hw_pass_through = 1;
324 /* devices under the same p2p bridge are owned in one domain */
325 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
327 /* domain represents a virtual machine, more than one devices
328 * across iommus may be owned in one domain, e.g. kvm guest.
330 #define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 1)
332 /* si_domain contains mulitple devices */
333 #define DOMAIN_FLAG_STATIC_IDENTITY (1 << 2)
335 /* define the limit of IOMMUs supported in each domain */
337 # define IOMMU_UNITS_SUPPORTED MAX_IO_APICS
339 # define IOMMU_UNITS_SUPPORTED 64
343 int id; /* domain id */
344 int nid; /* node id */
345 DECLARE_BITMAP(iommu_bmp, IOMMU_UNITS_SUPPORTED);
346 /* bitmap of iommus this domain uses*/
348 struct list_head devices; /* all devices' list */
349 struct iova_domain iovad; /* iova's that belong to this domain */
351 struct dma_pte *pgd; /* virtual address */
352 int gaw; /* max guest address width */
354 /* adjusted guest address width, 0 is level 2 30-bit */
357 int flags; /* flags to find out type of domain */
359 int iommu_coherency;/* indicate coherency of iommu access */
360 int iommu_snooping; /* indicate snooping control feature*/
361 int iommu_count; /* reference count of iommu */
362 int iommu_superpage;/* Level of superpages supported:
363 0 == 4KiB (no superpages), 1 == 2MiB,
364 2 == 1GiB, 3 == 512GiB, 4 == 1TiB */
365 spinlock_t iommu_lock; /* protect iommu set in domain */
366 u64 max_addr; /* maximum mapped address */
369 /* PCI domain-device relationship */
370 struct device_domain_info {
371 struct list_head link; /* link to domain siblings */
372 struct list_head global; /* link to global list */
373 u8 bus; /* PCI bus number */
374 u8 devfn; /* PCI devfn number */
375 struct device *dev; /* it's NULL for PCIe-to-PCI bridge */
376 struct intel_iommu *iommu; /* IOMMU used by this device */
377 struct dmar_domain *domain; /* pointer to domain */
380 struct dmar_rmrr_unit {
381 struct list_head list; /* list of rmrr units */
382 struct acpi_dmar_header *hdr; /* ACPI header */
383 u64 base_address; /* reserved base address*/
384 u64 end_address; /* reserved end address */
385 struct dmar_dev_scope *devices; /* target devices */
386 int devices_cnt; /* target device count */
389 struct dmar_atsr_unit {
390 struct list_head list; /* list of ATSR units */
391 struct acpi_dmar_header *hdr; /* ACPI header */
392 struct dmar_dev_scope *devices; /* target devices */
393 int devices_cnt; /* target device count */
394 u8 include_all:1; /* include all ports */
397 static LIST_HEAD(dmar_atsr_units);
398 static LIST_HEAD(dmar_rmrr_units);
400 #define for_each_rmrr_units(rmrr) \
401 list_for_each_entry(rmrr, &dmar_rmrr_units, list)
403 static void flush_unmaps_timeout(unsigned long data);
405 static DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
407 #define HIGH_WATER_MARK 250
408 struct deferred_flush_tables {
410 struct iova *iova[HIGH_WATER_MARK];
411 struct dmar_domain *domain[HIGH_WATER_MARK];
412 struct page *freelist[HIGH_WATER_MARK];
415 static struct deferred_flush_tables *deferred_flush;
417 /* bitmap for indexing intel_iommus */
418 static int g_num_of_iommus;
420 static DEFINE_SPINLOCK(async_umap_flush_lock);
421 static LIST_HEAD(unmaps_to_do);
424 static long list_size;
426 static void domain_exit(struct dmar_domain *domain);
427 static void domain_remove_dev_info(struct dmar_domain *domain);
428 static void domain_remove_one_dev_info(struct dmar_domain *domain,
430 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
433 #ifdef CONFIG_INTEL_IOMMU_DEFAULT_ON
434 int dmar_disabled = 0;
436 int dmar_disabled = 1;
437 #endif /*CONFIG_INTEL_IOMMU_DEFAULT_ON*/
439 int intel_iommu_enabled = 0;
440 EXPORT_SYMBOL_GPL(intel_iommu_enabled);
442 static int dmar_map_gfx = 1;
443 static int dmar_forcedac;
444 static int intel_iommu_strict;
445 static int intel_iommu_superpage = 1;
447 int intel_iommu_gfx_mapped;
448 EXPORT_SYMBOL_GPL(intel_iommu_gfx_mapped);
450 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
451 static DEFINE_SPINLOCK(device_domain_lock);
452 static LIST_HEAD(device_domain_list);
454 static struct iommu_ops intel_iommu_ops;
456 static int __init intel_iommu_setup(char *str)
461 if (!strncmp(str, "on", 2)) {
463 printk(KERN_INFO "Intel-IOMMU: enabled\n");
464 } else if (!strncmp(str, "off", 3)) {
466 printk(KERN_INFO "Intel-IOMMU: disabled\n");
467 } else if (!strncmp(str, "igfx_off", 8)) {
470 "Intel-IOMMU: disable GFX device mapping\n");
471 } else if (!strncmp(str, "forcedac", 8)) {
473 "Intel-IOMMU: Forcing DAC for PCI devices\n");
475 } else if (!strncmp(str, "strict", 6)) {
477 "Intel-IOMMU: disable batched IOTLB flush\n");
478 intel_iommu_strict = 1;
479 } else if (!strncmp(str, "sp_off", 6)) {
481 "Intel-IOMMU: disable supported super page\n");
482 intel_iommu_superpage = 0;
485 str += strcspn(str, ",");
491 __setup("intel_iommu=", intel_iommu_setup);
493 static struct kmem_cache *iommu_domain_cache;
494 static struct kmem_cache *iommu_devinfo_cache;
495 static struct kmem_cache *iommu_iova_cache;
497 static inline void *alloc_pgtable_page(int node)
502 page = alloc_pages_node(node, GFP_ATOMIC | __GFP_ZERO, 0);
504 vaddr = page_address(page);
508 static inline void free_pgtable_page(void *vaddr)
510 free_page((unsigned long)vaddr);
513 static inline void *alloc_domain_mem(void)
515 return kmem_cache_alloc(iommu_domain_cache, GFP_ATOMIC);
518 static void free_domain_mem(void *vaddr)
520 kmem_cache_free(iommu_domain_cache, vaddr);
523 static inline void * alloc_devinfo_mem(void)
525 return kmem_cache_alloc(iommu_devinfo_cache, GFP_ATOMIC);
528 static inline void free_devinfo_mem(void *vaddr)
530 kmem_cache_free(iommu_devinfo_cache, vaddr);
533 struct iova *alloc_iova_mem(void)
535 return kmem_cache_alloc(iommu_iova_cache, GFP_ATOMIC);
538 void free_iova_mem(struct iova *iova)
540 kmem_cache_free(iommu_iova_cache, iova);
544 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
549 sagaw = cap_sagaw(iommu->cap);
550 for (agaw = width_to_agaw(max_gaw);
552 if (test_bit(agaw, &sagaw))
560 * Calculate max SAGAW for each iommu.
562 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
564 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
568 * calculate agaw for each iommu.
569 * "SAGAW" may be different across iommus, use a default agaw, and
570 * get a supported less agaw for iommus that don't support the default agaw.
572 int iommu_calculate_agaw(struct intel_iommu *iommu)
574 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
577 /* This functionin only returns single iommu in a domain */
578 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
582 /* si_domain and vm domain should not get here. */
583 BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
584 BUG_ON(domain->flags & DOMAIN_FLAG_STATIC_IDENTITY);
586 iommu_id = find_first_bit(domain->iommu_bmp, g_num_of_iommus);
587 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
590 return g_iommus[iommu_id];
593 static void domain_update_iommu_coherency(struct dmar_domain *domain)
595 struct dmar_drhd_unit *drhd;
596 struct intel_iommu *iommu;
599 domain->iommu_coherency = 1;
601 for_each_set_bit(i, domain->iommu_bmp, g_num_of_iommus) {
603 if (!ecap_coherent(g_iommus[i]->ecap)) {
604 domain->iommu_coherency = 0;
611 /* No hardware attached; use lowest common denominator */
613 for_each_active_iommu(iommu, drhd) {
614 if (!ecap_coherent(iommu->ecap)) {
615 domain->iommu_coherency = 0;
622 static void domain_update_iommu_snooping(struct dmar_domain *domain)
626 domain->iommu_snooping = 1;
628 for_each_set_bit(i, domain->iommu_bmp, g_num_of_iommus) {
629 if (!ecap_sc_support(g_iommus[i]->ecap)) {
630 domain->iommu_snooping = 0;
636 static void domain_update_iommu_superpage(struct dmar_domain *domain)
638 struct dmar_drhd_unit *drhd;
639 struct intel_iommu *iommu = NULL;
642 if (!intel_iommu_superpage) {
643 domain->iommu_superpage = 0;
647 /* set iommu_superpage to the smallest common denominator */
649 for_each_active_iommu(iommu, drhd) {
650 mask &= cap_super_page_val(iommu->cap);
657 domain->iommu_superpage = fls(mask);
660 /* Some capabilities may be different across iommus */
661 static void domain_update_iommu_cap(struct dmar_domain *domain)
663 domain_update_iommu_coherency(domain);
664 domain_update_iommu_snooping(domain);
665 domain_update_iommu_superpage(domain);
668 static struct intel_iommu *device_to_iommu(struct device *dev, u8 *bus, u8 *devfn)
670 struct dmar_drhd_unit *drhd = NULL;
671 struct intel_iommu *iommu;
673 struct pci_dev *ptmp, *pdev = NULL;
677 if (dev_is_pci(dev)) {
678 pdev = to_pci_dev(dev);
679 segment = pci_domain_nr(pdev->bus);
680 } else if (ACPI_COMPANION(dev))
681 dev = &ACPI_COMPANION(dev)->dev;
684 for_each_active_iommu(iommu, drhd) {
685 if (pdev && segment != drhd->segment)
688 for_each_active_dev_scope(drhd->devices,
689 drhd->devices_cnt, i, tmp) {
691 *bus = drhd->devices[i].bus;
692 *devfn = drhd->devices[i].devfn;
696 if (!pdev || !dev_is_pci(tmp))
699 ptmp = to_pci_dev(tmp);
700 if (ptmp->subordinate &&
701 ptmp->subordinate->number <= pdev->bus->number &&
702 ptmp->subordinate->busn_res.end >= pdev->bus->number)
706 if (pdev && drhd->include_all) {
708 *bus = pdev->bus->number;
709 *devfn = pdev->devfn;
720 static void domain_flush_cache(struct dmar_domain *domain,
721 void *addr, int size)
723 if (!domain->iommu_coherency)
724 clflush_cache_range(addr, size);
727 /* Gets context entry for a given bus and devfn */
728 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
731 struct root_entry *root;
732 struct context_entry *context;
733 unsigned long phy_addr;
736 spin_lock_irqsave(&iommu->lock, flags);
737 root = &iommu->root_entry[bus];
738 context = get_context_addr_from_root(root);
740 context = (struct context_entry *)
741 alloc_pgtable_page(iommu->node);
743 spin_unlock_irqrestore(&iommu->lock, flags);
746 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
747 phy_addr = virt_to_phys((void *)context);
748 set_root_value(root, phy_addr);
749 set_root_present(root);
750 __iommu_flush_cache(iommu, root, sizeof(*root));
752 spin_unlock_irqrestore(&iommu->lock, flags);
753 return &context[devfn];
756 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
758 struct root_entry *root;
759 struct context_entry *context;
763 spin_lock_irqsave(&iommu->lock, flags);
764 root = &iommu->root_entry[bus];
765 context = get_context_addr_from_root(root);
770 ret = context_present(&context[devfn]);
772 spin_unlock_irqrestore(&iommu->lock, flags);
776 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
778 struct root_entry *root;
779 struct context_entry *context;
782 spin_lock_irqsave(&iommu->lock, flags);
783 root = &iommu->root_entry[bus];
784 context = get_context_addr_from_root(root);
786 context_clear_entry(&context[devfn]);
787 __iommu_flush_cache(iommu, &context[devfn], \
790 spin_unlock_irqrestore(&iommu->lock, flags);
793 static void free_context_table(struct intel_iommu *iommu)
795 struct root_entry *root;
798 struct context_entry *context;
800 spin_lock_irqsave(&iommu->lock, flags);
801 if (!iommu->root_entry) {
804 for (i = 0; i < ROOT_ENTRY_NR; i++) {
805 root = &iommu->root_entry[i];
806 context = get_context_addr_from_root(root);
808 free_pgtable_page(context);
810 free_pgtable_page(iommu->root_entry);
811 iommu->root_entry = NULL;
813 spin_unlock_irqrestore(&iommu->lock, flags);
816 static struct dma_pte *pfn_to_dma_pte(struct dmar_domain *domain,
817 unsigned long pfn, int *target_level)
819 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
820 struct dma_pte *parent, *pte = NULL;
821 int level = agaw_to_level(domain->agaw);
824 BUG_ON(!domain->pgd);
826 if (addr_width < BITS_PER_LONG && pfn >> addr_width)
827 /* Address beyond IOMMU's addressing capabilities. */
830 parent = domain->pgd;
835 offset = pfn_level_offset(pfn, level);
836 pte = &parent[offset];
837 if (!*target_level && (dma_pte_superpage(pte) || !dma_pte_present(pte)))
839 if (level == *target_level)
842 if (!dma_pte_present(pte)) {
845 tmp_page = alloc_pgtable_page(domain->nid);
850 domain_flush_cache(domain, tmp_page, VTD_PAGE_SIZE);
851 pteval = ((uint64_t)virt_to_dma_pfn(tmp_page) << VTD_PAGE_SHIFT) | DMA_PTE_READ | DMA_PTE_WRITE;
852 if (cmpxchg64(&pte->val, 0ULL, pteval)) {
853 /* Someone else set it while we were thinking; use theirs. */
854 free_pgtable_page(tmp_page);
857 domain_flush_cache(domain, pte, sizeof(*pte));
863 parent = phys_to_virt(dma_pte_addr(pte));
868 *target_level = level;
874 /* return address's pte at specific level */
875 static struct dma_pte *dma_pfn_level_pte(struct dmar_domain *domain,
877 int level, int *large_page)
879 struct dma_pte *parent, *pte = NULL;
880 int total = agaw_to_level(domain->agaw);
883 parent = domain->pgd;
884 while (level <= total) {
885 offset = pfn_level_offset(pfn, total);
886 pte = &parent[offset];
890 if (!dma_pte_present(pte)) {
895 if (dma_pte_superpage(pte)) {
900 parent = phys_to_virt(dma_pte_addr(pte));
906 /* clear last level pte, a tlb flush should be followed */
907 static void dma_pte_clear_range(struct dmar_domain *domain,
908 unsigned long start_pfn,
909 unsigned long last_pfn)
911 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
912 unsigned int large_page = 1;
913 struct dma_pte *first_pte, *pte;
915 BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
916 BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
917 BUG_ON(start_pfn > last_pfn);
919 /* we don't need lock here; nobody else touches the iova range */
922 first_pte = pte = dma_pfn_level_pte(domain, start_pfn, 1, &large_page);
924 start_pfn = align_to_level(start_pfn + 1, large_page + 1);
929 start_pfn += lvl_to_nr_pages(large_page);
931 } while (start_pfn <= last_pfn && !first_pte_in_page(pte));
933 domain_flush_cache(domain, first_pte,
934 (void *)pte - (void *)first_pte);
936 } while (start_pfn && start_pfn <= last_pfn);
939 static void dma_pte_free_level(struct dmar_domain *domain, int level,
940 struct dma_pte *pte, unsigned long pfn,
941 unsigned long start_pfn, unsigned long last_pfn)
943 pfn = max(start_pfn, pfn);
944 pte = &pte[pfn_level_offset(pfn, level)];
947 unsigned long level_pfn;
948 struct dma_pte *level_pte;
950 if (!dma_pte_present(pte) || dma_pte_superpage(pte))
953 level_pfn = pfn & level_mask(level - 1);
954 level_pte = phys_to_virt(dma_pte_addr(pte));
957 dma_pte_free_level(domain, level - 1, level_pte,
958 level_pfn, start_pfn, last_pfn);
960 /* If range covers entire pagetable, free it */
961 if (!(start_pfn > level_pfn ||
962 last_pfn < level_pfn + level_size(level) - 1)) {
964 domain_flush_cache(domain, pte, sizeof(*pte));
965 free_pgtable_page(level_pte);
968 pfn += level_size(level);
969 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
972 /* free page table pages. last level pte should already be cleared */
973 static void dma_pte_free_pagetable(struct dmar_domain *domain,
974 unsigned long start_pfn,
975 unsigned long last_pfn)
977 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
979 BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
980 BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
981 BUG_ON(start_pfn > last_pfn);
983 /* We don't need lock here; nobody else touches the iova range */
984 dma_pte_free_level(domain, agaw_to_level(domain->agaw),
985 domain->pgd, 0, start_pfn, last_pfn);
988 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
989 free_pgtable_page(domain->pgd);
994 /* When a page at a given level is being unlinked from its parent, we don't
995 need to *modify* it at all. All we need to do is make a list of all the
996 pages which can be freed just as soon as we've flushed the IOTLB and we
997 know the hardware page-walk will no longer touch them.
998 The 'pte' argument is the *parent* PTE, pointing to the page that is to
1000 static struct page *dma_pte_list_pagetables(struct dmar_domain *domain,
1001 int level, struct dma_pte *pte,
1002 struct page *freelist)
1006 pg = pfn_to_page(dma_pte_addr(pte) >> PAGE_SHIFT);
1007 pg->freelist = freelist;
1013 pte = page_address(pg);
1015 if (dma_pte_present(pte) && !dma_pte_superpage(pte))
1016 freelist = dma_pte_list_pagetables(domain, level - 1,
1019 } while (!first_pte_in_page(pte));
1024 static struct page *dma_pte_clear_level(struct dmar_domain *domain, int level,
1025 struct dma_pte *pte, unsigned long pfn,
1026 unsigned long start_pfn,
1027 unsigned long last_pfn,
1028 struct page *freelist)
1030 struct dma_pte *first_pte = NULL, *last_pte = NULL;
1032 pfn = max(start_pfn, pfn);
1033 pte = &pte[pfn_level_offset(pfn, level)];
1036 unsigned long level_pfn;
1038 if (!dma_pte_present(pte))
1041 level_pfn = pfn & level_mask(level);
1043 /* If range covers entire pagetable, free it */
1044 if (start_pfn <= level_pfn &&
1045 last_pfn >= level_pfn + level_size(level) - 1) {
1046 /* These suborbinate page tables are going away entirely. Don't
1047 bother to clear them; we're just going to *free* them. */
1048 if (level > 1 && !dma_pte_superpage(pte))
1049 freelist = dma_pte_list_pagetables(domain, level - 1, pte, freelist);
1055 } else if (level > 1) {
1056 /* Recurse down into a level that isn't *entirely* obsolete */
1057 freelist = dma_pte_clear_level(domain, level - 1,
1058 phys_to_virt(dma_pte_addr(pte)),
1059 level_pfn, start_pfn, last_pfn,
1063 pfn += level_size(level);
1064 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
1067 domain_flush_cache(domain, first_pte,
1068 (void *)++last_pte - (void *)first_pte);
1073 /* We can't just free the pages because the IOMMU may still be walking
1074 the page tables, and may have cached the intermediate levels. The
1075 pages can only be freed after the IOTLB flush has been done. */
1076 struct page *domain_unmap(struct dmar_domain *domain,
1077 unsigned long start_pfn,
1078 unsigned long last_pfn)
1080 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
1081 struct page *freelist = NULL;
1083 BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
1084 BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
1085 BUG_ON(start_pfn > last_pfn);
1087 /* we don't need lock here; nobody else touches the iova range */
1088 freelist = dma_pte_clear_level(domain, agaw_to_level(domain->agaw),
1089 domain->pgd, 0, start_pfn, last_pfn, NULL);
1092 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
1093 struct page *pgd_page = virt_to_page(domain->pgd);
1094 pgd_page->freelist = freelist;
1095 freelist = pgd_page;
1103 void dma_free_pagelist(struct page *freelist)
1107 while ((pg = freelist)) {
1108 freelist = pg->freelist;
1109 free_pgtable_page(page_address(pg));
1113 /* iommu handling */
1114 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
1116 struct root_entry *root;
1117 unsigned long flags;
1119 root = (struct root_entry *)alloc_pgtable_page(iommu->node);
1123 __iommu_flush_cache(iommu, root, ROOT_SIZE);
1125 spin_lock_irqsave(&iommu->lock, flags);
1126 iommu->root_entry = root;
1127 spin_unlock_irqrestore(&iommu->lock, flags);
1132 static void iommu_set_root_entry(struct intel_iommu *iommu)
1138 addr = iommu->root_entry;
1140 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1141 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
1143 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
1145 /* Make sure hardware complete it */
1146 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1147 readl, (sts & DMA_GSTS_RTPS), sts);
1149 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1152 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
1157 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
1160 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1161 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
1163 /* Make sure hardware complete it */
1164 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1165 readl, (!(val & DMA_GSTS_WBFS)), val);
1167 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1170 /* return value determine if we need a write buffer flush */
1171 static void __iommu_flush_context(struct intel_iommu *iommu,
1172 u16 did, u16 source_id, u8 function_mask,
1179 case DMA_CCMD_GLOBAL_INVL:
1180 val = DMA_CCMD_GLOBAL_INVL;
1182 case DMA_CCMD_DOMAIN_INVL:
1183 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
1185 case DMA_CCMD_DEVICE_INVL:
1186 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
1187 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
1192 val |= DMA_CCMD_ICC;
1194 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1195 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
1197 /* Make sure hardware complete it */
1198 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
1199 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
1201 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1204 /* return value determine if we need a write buffer flush */
1205 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
1206 u64 addr, unsigned int size_order, u64 type)
1208 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
1209 u64 val = 0, val_iva = 0;
1213 case DMA_TLB_GLOBAL_FLUSH:
1214 /* global flush doesn't need set IVA_REG */
1215 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
1217 case DMA_TLB_DSI_FLUSH:
1218 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1220 case DMA_TLB_PSI_FLUSH:
1221 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1222 /* IH bit is passed in as part of address */
1223 val_iva = size_order | addr;
1228 /* Note: set drain read/write */
1231 * This is probably to be super secure.. Looks like we can
1232 * ignore it without any impact.
1234 if (cap_read_drain(iommu->cap))
1235 val |= DMA_TLB_READ_DRAIN;
1237 if (cap_write_drain(iommu->cap))
1238 val |= DMA_TLB_WRITE_DRAIN;
1240 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1241 /* Note: Only uses first TLB reg currently */
1243 dmar_writeq(iommu->reg + tlb_offset, val_iva);
1244 dmar_writeq(iommu->reg + tlb_offset + 8, val);
1246 /* Make sure hardware complete it */
1247 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
1248 dmar_readq, (!(val & DMA_TLB_IVT)), val);
1250 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1252 /* check IOTLB invalidation granularity */
1253 if (DMA_TLB_IAIG(val) == 0)
1254 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
1255 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
1256 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
1257 (unsigned long long)DMA_TLB_IIRG(type),
1258 (unsigned long long)DMA_TLB_IAIG(val));
1261 static struct device_domain_info *
1262 iommu_support_dev_iotlb (struct dmar_domain *domain, struct intel_iommu *iommu,
1266 unsigned long flags;
1267 struct device_domain_info *info;
1268 struct pci_dev *pdev;
1270 if (!ecap_dev_iotlb_support(iommu->ecap))
1276 spin_lock_irqsave(&device_domain_lock, flags);
1277 list_for_each_entry(info, &domain->devices, link)
1278 if (info->bus == bus && info->devfn == devfn) {
1282 spin_unlock_irqrestore(&device_domain_lock, flags);
1284 if (!found || !info->dev || !dev_is_pci(info->dev))
1287 pdev = to_pci_dev(info->dev);
1289 if (!pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS))
1292 if (!dmar_find_matched_atsr_unit(pdev))
1298 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
1300 if (!info || !dev_is_pci(info->dev))
1303 pci_enable_ats(to_pci_dev(info->dev), VTD_PAGE_SHIFT);
1306 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1308 if (!info->dev || !dev_is_pci(info->dev) ||
1309 !pci_ats_enabled(to_pci_dev(info->dev)))
1312 pci_disable_ats(to_pci_dev(info->dev));
1315 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1316 u64 addr, unsigned mask)
1319 unsigned long flags;
1320 struct device_domain_info *info;
1322 spin_lock_irqsave(&device_domain_lock, flags);
1323 list_for_each_entry(info, &domain->devices, link) {
1324 struct pci_dev *pdev;
1325 if (!info->dev || !dev_is_pci(info->dev))
1328 pdev = to_pci_dev(info->dev);
1329 if (!pci_ats_enabled(pdev))
1332 sid = info->bus << 8 | info->devfn;
1333 qdep = pci_ats_queue_depth(pdev);
1334 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1336 spin_unlock_irqrestore(&device_domain_lock, flags);
1339 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
1340 unsigned long pfn, unsigned int pages, int ih, int map)
1342 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1343 uint64_t addr = (uint64_t)pfn << VTD_PAGE_SHIFT;
1350 * Fallback to domain selective flush if no PSI support or the size is
1352 * PSI requires page size to be 2 ^ x, and the base address is naturally
1353 * aligned to the size
1355 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1356 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1359 iommu->flush.flush_iotlb(iommu, did, addr | ih, mask,
1363 * In caching mode, changes of pages from non-present to present require
1364 * flush. However, device IOTLB doesn't need to be flushed in this case.
1366 if (!cap_caching_mode(iommu->cap) || !map)
1367 iommu_flush_dev_iotlb(iommu->domains[did], addr, mask);
1370 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1373 unsigned long flags;
1375 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1376 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1377 pmen &= ~DMA_PMEN_EPM;
1378 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1380 /* wait for the protected region status bit to clear */
1381 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1382 readl, !(pmen & DMA_PMEN_PRS), pmen);
1384 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1387 static int iommu_enable_translation(struct intel_iommu *iommu)
1390 unsigned long flags;
1392 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1393 iommu->gcmd |= DMA_GCMD_TE;
1394 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1396 /* Make sure hardware complete it */
1397 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1398 readl, (sts & DMA_GSTS_TES), sts);
1400 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1404 static int iommu_disable_translation(struct intel_iommu *iommu)
1409 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1410 iommu->gcmd &= ~DMA_GCMD_TE;
1411 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1413 /* Make sure hardware complete it */
1414 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1415 readl, (!(sts & DMA_GSTS_TES)), sts);
1417 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1422 static int iommu_init_domains(struct intel_iommu *iommu)
1424 unsigned long ndomains;
1425 unsigned long nlongs;
1427 ndomains = cap_ndoms(iommu->cap);
1428 pr_debug("IOMMU%d: Number of Domains supported <%ld>\n",
1429 iommu->seq_id, ndomains);
1430 nlongs = BITS_TO_LONGS(ndomains);
1432 spin_lock_init(&iommu->lock);
1434 /* TBD: there might be 64K domains,
1435 * consider other allocation for future chip
1437 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1438 if (!iommu->domain_ids) {
1439 pr_err("IOMMU%d: allocating domain id array failed\n",
1443 iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1445 if (!iommu->domains) {
1446 pr_err("IOMMU%d: allocating domain array failed\n",
1448 kfree(iommu->domain_ids);
1449 iommu->domain_ids = NULL;
1454 * if Caching mode is set, then invalid translations are tagged
1455 * with domainid 0. Hence we need to pre-allocate it.
1457 if (cap_caching_mode(iommu->cap))
1458 set_bit(0, iommu->domain_ids);
1462 static void free_dmar_iommu(struct intel_iommu *iommu)
1464 struct dmar_domain *domain;
1466 unsigned long flags;
1468 if ((iommu->domains) && (iommu->domain_ids)) {
1469 for_each_set_bit(i, iommu->domain_ids, cap_ndoms(iommu->cap)) {
1471 * Domain id 0 is reserved for invalid translation
1472 * if hardware supports caching mode.
1474 if (cap_caching_mode(iommu->cap) && i == 0)
1477 domain = iommu->domains[i];
1478 clear_bit(i, iommu->domain_ids);
1480 spin_lock_irqsave(&domain->iommu_lock, flags);
1481 count = --domain->iommu_count;
1482 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1484 domain_exit(domain);
1488 if (iommu->gcmd & DMA_GCMD_TE)
1489 iommu_disable_translation(iommu);
1491 kfree(iommu->domains);
1492 kfree(iommu->domain_ids);
1493 iommu->domains = NULL;
1494 iommu->domain_ids = NULL;
1496 g_iommus[iommu->seq_id] = NULL;
1498 /* free context mapping */
1499 free_context_table(iommu);
1502 static struct dmar_domain *alloc_domain(bool vm)
1504 /* domain id for virtual machine, it won't be set in context */
1505 static atomic_t vm_domid = ATOMIC_INIT(0);
1506 struct dmar_domain *domain;
1508 domain = alloc_domain_mem();
1513 domain->iommu_count = 0;
1514 memset(domain->iommu_bmp, 0, sizeof(domain->iommu_bmp));
1516 spin_lock_init(&domain->iommu_lock);
1517 INIT_LIST_HEAD(&domain->devices);
1519 domain->id = atomic_inc_return(&vm_domid);
1520 domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
1526 static int iommu_attach_domain(struct dmar_domain *domain,
1527 struct intel_iommu *iommu)
1530 unsigned long ndomains;
1531 unsigned long flags;
1533 ndomains = cap_ndoms(iommu->cap);
1535 spin_lock_irqsave(&iommu->lock, flags);
1537 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1538 if (num >= ndomains) {
1539 spin_unlock_irqrestore(&iommu->lock, flags);
1540 printk(KERN_ERR "IOMMU: no free domain ids\n");
1545 domain->iommu_count++;
1546 set_bit(num, iommu->domain_ids);
1547 set_bit(iommu->seq_id, domain->iommu_bmp);
1548 iommu->domains[num] = domain;
1549 spin_unlock_irqrestore(&iommu->lock, flags);
1554 static void iommu_detach_domain(struct dmar_domain *domain,
1555 struct intel_iommu *iommu)
1557 unsigned long flags;
1560 spin_lock_irqsave(&iommu->lock, flags);
1561 ndomains = cap_ndoms(iommu->cap);
1562 for_each_set_bit(num, iommu->domain_ids, ndomains) {
1563 if (iommu->domains[num] == domain) {
1564 clear_bit(num, iommu->domain_ids);
1565 iommu->domains[num] = NULL;
1569 spin_unlock_irqrestore(&iommu->lock, flags);
1572 static struct iova_domain reserved_iova_list;
1573 static struct lock_class_key reserved_rbtree_key;
1575 static int dmar_init_reserved_ranges(void)
1577 struct pci_dev *pdev = NULL;
1581 init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1583 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1584 &reserved_rbtree_key);
1586 /* IOAPIC ranges shouldn't be accessed by DMA */
1587 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1588 IOVA_PFN(IOAPIC_RANGE_END));
1590 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1594 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1595 for_each_pci_dev(pdev) {
1598 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1599 r = &pdev->resource[i];
1600 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1602 iova = reserve_iova(&reserved_iova_list,
1606 printk(KERN_ERR "Reserve iova failed\n");
1614 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1616 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1619 static inline int guestwidth_to_adjustwidth(int gaw)
1622 int r = (gaw - 12) % 9;
1633 static int domain_init(struct dmar_domain *domain, int guest_width)
1635 struct intel_iommu *iommu;
1636 int adjust_width, agaw;
1637 unsigned long sagaw;
1639 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1640 domain_reserve_special_ranges(domain);
1642 /* calculate AGAW */
1643 iommu = domain_get_iommu(domain);
1644 if (guest_width > cap_mgaw(iommu->cap))
1645 guest_width = cap_mgaw(iommu->cap);
1646 domain->gaw = guest_width;
1647 adjust_width = guestwidth_to_adjustwidth(guest_width);
1648 agaw = width_to_agaw(adjust_width);
1649 sagaw = cap_sagaw(iommu->cap);
1650 if (!test_bit(agaw, &sagaw)) {
1651 /* hardware doesn't support it, choose a bigger one */
1652 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1653 agaw = find_next_bit(&sagaw, 5, agaw);
1657 domain->agaw = agaw;
1659 if (ecap_coherent(iommu->ecap))
1660 domain->iommu_coherency = 1;
1662 domain->iommu_coherency = 0;
1664 if (ecap_sc_support(iommu->ecap))
1665 domain->iommu_snooping = 1;
1667 domain->iommu_snooping = 0;
1669 if (intel_iommu_superpage)
1670 domain->iommu_superpage = fls(cap_super_page_val(iommu->cap));
1672 domain->iommu_superpage = 0;
1674 domain->nid = iommu->node;
1676 /* always allocate the top pgd */
1677 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
1680 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1684 static void domain_exit(struct dmar_domain *domain)
1686 struct dmar_drhd_unit *drhd;
1687 struct intel_iommu *iommu;
1688 struct page *freelist = NULL;
1690 /* Domain 0 is reserved, so dont process it */
1694 /* Flush any lazy unmaps that may reference this domain */
1695 if (!intel_iommu_strict)
1696 flush_unmaps_timeout(0);
1698 /* remove associated devices */
1699 domain_remove_dev_info(domain);
1702 put_iova_domain(&domain->iovad);
1704 freelist = domain_unmap(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1706 /* clear attached or cached domains */
1708 for_each_active_iommu(iommu, drhd)
1709 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
1710 test_bit(iommu->seq_id, domain->iommu_bmp))
1711 iommu_detach_domain(domain, iommu);
1714 dma_free_pagelist(freelist);
1716 free_domain_mem(domain);
1719 static int domain_context_mapping_one(struct dmar_domain *domain,
1720 struct intel_iommu *iommu,
1721 u8 bus, u8 devfn, int translation)
1723 struct context_entry *context;
1724 unsigned long flags;
1725 struct dma_pte *pgd;
1727 unsigned long ndomains;
1730 struct device_domain_info *info = NULL;
1732 pr_debug("Set context mapping for %02x:%02x.%d\n",
1733 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1735 BUG_ON(!domain->pgd);
1736 BUG_ON(translation != CONTEXT_TT_PASS_THROUGH &&
1737 translation != CONTEXT_TT_MULTI_LEVEL);
1739 context = device_to_context_entry(iommu, bus, devfn);
1742 spin_lock_irqsave(&iommu->lock, flags);
1743 if (context_present(context)) {
1744 spin_unlock_irqrestore(&iommu->lock, flags);
1751 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
1752 domain->flags & DOMAIN_FLAG_STATIC_IDENTITY) {
1755 /* find an available domain id for this device in iommu */
1756 ndomains = cap_ndoms(iommu->cap);
1757 for_each_set_bit(num, iommu->domain_ids, ndomains) {
1758 if (iommu->domains[num] == domain) {
1766 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1767 if (num >= ndomains) {
1768 spin_unlock_irqrestore(&iommu->lock, flags);
1769 printk(KERN_ERR "IOMMU: no free domain ids\n");
1773 set_bit(num, iommu->domain_ids);
1774 iommu->domains[num] = domain;
1778 /* Skip top levels of page tables for
1779 * iommu which has less agaw than default.
1780 * Unnecessary for PT mode.
1782 if (translation != CONTEXT_TT_PASS_THROUGH) {
1783 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1784 pgd = phys_to_virt(dma_pte_addr(pgd));
1785 if (!dma_pte_present(pgd)) {
1786 spin_unlock_irqrestore(&iommu->lock, flags);
1793 context_set_domain_id(context, id);
1795 if (translation != CONTEXT_TT_PASS_THROUGH) {
1796 info = iommu_support_dev_iotlb(domain, iommu, bus, devfn);
1797 translation = info ? CONTEXT_TT_DEV_IOTLB :
1798 CONTEXT_TT_MULTI_LEVEL;
1801 * In pass through mode, AW must be programmed to indicate the largest
1802 * AGAW value supported by hardware. And ASR is ignored by hardware.
1804 if (unlikely(translation == CONTEXT_TT_PASS_THROUGH))
1805 context_set_address_width(context, iommu->msagaw);
1807 context_set_address_root(context, virt_to_phys(pgd));
1808 context_set_address_width(context, iommu->agaw);
1811 context_set_translation_type(context, translation);
1812 context_set_fault_enable(context);
1813 context_set_present(context);
1814 domain_flush_cache(domain, context, sizeof(*context));
1817 * It's a non-present to present mapping. If hardware doesn't cache
1818 * non-present entry we only need to flush the write-buffer. If the
1819 * _does_ cache non-present entries, then it does so in the special
1820 * domain #0, which we have to flush:
1822 if (cap_caching_mode(iommu->cap)) {
1823 iommu->flush.flush_context(iommu, 0,
1824 (((u16)bus) << 8) | devfn,
1825 DMA_CCMD_MASK_NOBIT,
1826 DMA_CCMD_DEVICE_INVL);
1827 iommu->flush.flush_iotlb(iommu, domain->id, 0, 0, DMA_TLB_DSI_FLUSH);
1829 iommu_flush_write_buffer(iommu);
1831 iommu_enable_dev_iotlb(info);
1832 spin_unlock_irqrestore(&iommu->lock, flags);
1834 spin_lock_irqsave(&domain->iommu_lock, flags);
1835 if (!test_and_set_bit(iommu->seq_id, domain->iommu_bmp)) {
1836 domain->iommu_count++;
1837 if (domain->iommu_count == 1)
1838 domain->nid = iommu->node;
1839 domain_update_iommu_cap(domain);
1841 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1846 domain_context_mapping(struct dmar_domain *domain, struct device *dev,
1850 struct pci_dev *pdev, *tmp, *parent;
1851 struct intel_iommu *iommu;
1854 iommu = device_to_iommu(dev, &bus, &devfn);
1858 ret = domain_context_mapping_one(domain, iommu, bus, devfn,
1860 if (ret || !dev_is_pci(dev))
1863 /* dependent device mapping */
1864 pdev = to_pci_dev(dev);
1865 tmp = pci_find_upstream_pcie_bridge(pdev);
1868 /* Secondary interface's bus number and devfn 0 */
1869 parent = pdev->bus->self;
1870 while (parent != tmp) {
1871 ret = domain_context_mapping_one(domain, iommu,
1872 parent->bus->number,
1873 parent->devfn, translation);
1876 parent = parent->bus->self;
1878 if (pci_is_pcie(tmp)) /* this is a PCIe-to-PCI bridge */
1879 return domain_context_mapping_one(domain, iommu,
1880 tmp->subordinate->number, 0,
1882 else /* this is a legacy PCI bridge */
1883 return domain_context_mapping_one(domain, iommu,
1889 static int domain_context_mapped(struct device *dev)
1892 struct pci_dev *pdev, *tmp, *parent;
1893 struct intel_iommu *iommu;
1896 iommu = device_to_iommu(dev, &bus, &devfn);
1900 ret = device_context_mapped(iommu, bus, devfn);
1901 if (!ret || !dev_is_pci(dev))
1904 /* dependent device mapping */
1905 pdev = to_pci_dev(dev);
1906 tmp = pci_find_upstream_pcie_bridge(pdev);
1909 /* Secondary interface's bus number and devfn 0 */
1910 parent = pdev->bus->self;
1911 while (parent != tmp) {
1912 ret = device_context_mapped(iommu, parent->bus->number,
1916 parent = parent->bus->self;
1918 if (pci_is_pcie(tmp))
1919 return device_context_mapped(iommu, tmp->subordinate->number,
1922 return device_context_mapped(iommu, tmp->bus->number,
1926 /* Returns a number of VTD pages, but aligned to MM page size */
1927 static inline unsigned long aligned_nrpages(unsigned long host_addr,
1930 host_addr &= ~PAGE_MASK;
1931 return PAGE_ALIGN(host_addr + size) >> VTD_PAGE_SHIFT;
1934 /* Return largest possible superpage level for a given mapping */
1935 static inline int hardware_largepage_caps(struct dmar_domain *domain,
1936 unsigned long iov_pfn,
1937 unsigned long phy_pfn,
1938 unsigned long pages)
1940 int support, level = 1;
1941 unsigned long pfnmerge;
1943 support = domain->iommu_superpage;
1945 /* To use a large page, the virtual *and* physical addresses
1946 must be aligned to 2MiB/1GiB/etc. Lower bits set in either
1947 of them will mean we have to use smaller pages. So just
1948 merge them and check both at once. */
1949 pfnmerge = iov_pfn | phy_pfn;
1951 while (support && !(pfnmerge & ~VTD_STRIDE_MASK)) {
1952 pages >>= VTD_STRIDE_SHIFT;
1955 pfnmerge >>= VTD_STRIDE_SHIFT;
1962 static int __domain_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1963 struct scatterlist *sg, unsigned long phys_pfn,
1964 unsigned long nr_pages, int prot)
1966 struct dma_pte *first_pte = NULL, *pte = NULL;
1967 phys_addr_t uninitialized_var(pteval);
1968 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
1969 unsigned long sg_res;
1970 unsigned int largepage_lvl = 0;
1971 unsigned long lvl_pages = 0;
1973 BUG_ON(addr_width < BITS_PER_LONG && (iov_pfn + nr_pages - 1) >> addr_width);
1975 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1978 prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
1983 sg_res = nr_pages + 1;
1984 pteval = ((phys_addr_t)phys_pfn << VTD_PAGE_SHIFT) | prot;
1987 while (nr_pages > 0) {
1991 sg_res = aligned_nrpages(sg->offset, sg->length);
1992 sg->dma_address = ((dma_addr_t)iov_pfn << VTD_PAGE_SHIFT) + sg->offset;
1993 sg->dma_length = sg->length;
1994 pteval = page_to_phys(sg_page(sg)) | prot;
1995 phys_pfn = pteval >> VTD_PAGE_SHIFT;
1999 largepage_lvl = hardware_largepage_caps(domain, iov_pfn, phys_pfn, sg_res);
2001 first_pte = pte = pfn_to_dma_pte(domain, iov_pfn, &largepage_lvl);
2004 /* It is large page*/
2005 if (largepage_lvl > 1) {
2006 pteval |= DMA_PTE_LARGE_PAGE;
2007 /* Ensure that old small page tables are removed to make room
2008 for superpage, if they exist. */
2009 dma_pte_clear_range(domain, iov_pfn,
2010 iov_pfn + lvl_to_nr_pages(largepage_lvl) - 1);
2011 dma_pte_free_pagetable(domain, iov_pfn,
2012 iov_pfn + lvl_to_nr_pages(largepage_lvl) - 1);
2014 pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;
2018 /* We don't need lock here, nobody else
2019 * touches the iova range
2021 tmp = cmpxchg64_local(&pte->val, 0ULL, pteval);
2023 static int dumps = 5;
2024 printk(KERN_CRIT "ERROR: DMA PTE for vPFN 0x%lx already set (to %llx not %llx)\n",
2025 iov_pfn, tmp, (unsigned long long)pteval);
2028 debug_dma_dump_mappings(NULL);
2033 lvl_pages = lvl_to_nr_pages(largepage_lvl);
2035 BUG_ON(nr_pages < lvl_pages);
2036 BUG_ON(sg_res < lvl_pages);
2038 nr_pages -= lvl_pages;
2039 iov_pfn += lvl_pages;
2040 phys_pfn += lvl_pages;
2041 pteval += lvl_pages * VTD_PAGE_SIZE;
2042 sg_res -= lvl_pages;
2044 /* If the next PTE would be the first in a new page, then we
2045 need to flush the cache on the entries we've just written.
2046 And then we'll need to recalculate 'pte', so clear it and
2047 let it get set again in the if (!pte) block above.
2049 If we're done (!nr_pages) we need to flush the cache too.
2051 Also if we've been setting superpages, we may need to
2052 recalculate 'pte' and switch back to smaller pages for the
2053 end of the mapping, if the trailing size is not enough to
2054 use another superpage (i.e. sg_res < lvl_pages). */
2056 if (!nr_pages || first_pte_in_page(pte) ||
2057 (largepage_lvl > 1 && sg_res < lvl_pages)) {
2058 domain_flush_cache(domain, first_pte,
2059 (void *)pte - (void *)first_pte);
2063 if (!sg_res && nr_pages)
2069 static inline int domain_sg_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2070 struct scatterlist *sg, unsigned long nr_pages,
2073 return __domain_mapping(domain, iov_pfn, sg, 0, nr_pages, prot);
2076 static inline int domain_pfn_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2077 unsigned long phys_pfn, unsigned long nr_pages,
2080 return __domain_mapping(domain, iov_pfn, NULL, phys_pfn, nr_pages, prot);
2083 static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
2088 clear_context_table(iommu, bus, devfn);
2089 iommu->flush.flush_context(iommu, 0, 0, 0,
2090 DMA_CCMD_GLOBAL_INVL);
2091 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2094 static inline void unlink_domain_info(struct device_domain_info *info)
2096 assert_spin_locked(&device_domain_lock);
2097 list_del(&info->link);
2098 list_del(&info->global);
2100 info->dev->archdata.iommu = NULL;
2103 static void domain_remove_dev_info(struct dmar_domain *domain)
2105 struct device_domain_info *info, *tmp;
2106 unsigned long flags, flags2;
2108 spin_lock_irqsave(&device_domain_lock, flags);
2109 list_for_each_entry_safe(info, tmp, &domain->devices, link) {
2110 unlink_domain_info(info);
2111 spin_unlock_irqrestore(&device_domain_lock, flags);
2113 iommu_disable_dev_iotlb(info);
2114 iommu_detach_dev(info->iommu, info->bus, info->devfn);
2116 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE) {
2117 iommu_detach_dependent_devices(info->iommu, info->dev);
2118 /* clear this iommu in iommu_bmp, update iommu count
2121 spin_lock_irqsave(&domain->iommu_lock, flags2);
2122 if (test_and_clear_bit(info->iommu->seq_id,
2123 domain->iommu_bmp)) {
2124 domain->iommu_count--;
2125 domain_update_iommu_cap(domain);
2127 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
2130 free_devinfo_mem(info);
2131 spin_lock_irqsave(&device_domain_lock, flags);
2133 spin_unlock_irqrestore(&device_domain_lock, flags);
2138 * Note: we use struct device->archdata.iommu stores the info
2140 static struct dmar_domain *find_domain(struct device *dev)
2142 struct device_domain_info *info;
2144 /* No lock here, assumes no domain exit in normal case */
2145 info = dev->archdata.iommu;
2147 return info->domain;
2151 static inline struct device_domain_info *
2152 dmar_search_domain_by_dev_info(int segment, int bus, int devfn)
2154 struct device_domain_info *info;
2156 list_for_each_entry(info, &device_domain_list, global)
2157 if (info->iommu->segment == segment && info->bus == bus &&
2158 info->devfn == devfn)
2164 static struct dmar_domain *dmar_insert_dev_info(struct intel_iommu *iommu,
2167 struct dmar_domain *domain)
2169 struct dmar_domain *found = NULL;
2170 struct device_domain_info *info;
2171 unsigned long flags;
2173 info = alloc_devinfo_mem();
2178 info->devfn = devfn;
2180 info->domain = domain;
2181 info->iommu = iommu;
2183 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
2185 spin_lock_irqsave(&device_domain_lock, flags);
2187 found = find_domain(dev);
2189 struct device_domain_info *info2;
2190 info2 = dmar_search_domain_by_dev_info(iommu->segment, bus, devfn);
2192 found = info2->domain;
2195 spin_unlock_irqrestore(&device_domain_lock, flags);
2196 free_devinfo_mem(info);
2197 /* Caller must free the original domain */
2201 list_add(&info->link, &domain->devices);
2202 list_add(&info->global, &device_domain_list);
2204 dev->archdata.iommu = info;
2205 spin_unlock_irqrestore(&device_domain_lock, flags);
2210 /* domain is initialized */
2211 static struct dmar_domain *get_domain_for_dev(struct device *dev, int gaw)
2213 struct dmar_domain *domain, *free = NULL;
2214 struct intel_iommu *iommu = NULL;
2215 struct device_domain_info *info;
2216 struct pci_dev *dev_tmp = NULL;
2217 unsigned long flags;
2218 u8 bus, devfn, bridge_bus, bridge_devfn;
2220 domain = find_domain(dev);
2224 if (dev_is_pci(dev)) {
2225 struct pci_dev *pdev = to_pci_dev(dev);
2228 segment = pci_domain_nr(pdev->bus);
2229 dev_tmp = pci_find_upstream_pcie_bridge(pdev);
2231 if (pci_is_pcie(dev_tmp)) {
2232 bridge_bus = dev_tmp->subordinate->number;
2235 bridge_bus = dev_tmp->bus->number;
2236 bridge_devfn = dev_tmp->devfn;
2238 spin_lock_irqsave(&device_domain_lock, flags);
2239 info = dmar_search_domain_by_dev_info(segment,
2243 iommu = info->iommu;
2244 domain = info->domain;
2246 spin_unlock_irqrestore(&device_domain_lock, flags);
2247 /* pcie-pci bridge already has a domain, uses it */
2253 iommu = device_to_iommu(dev, &bus, &devfn);
2257 /* Allocate and initialize new domain for the device */
2258 domain = alloc_domain(false);
2261 if (iommu_attach_domain(domain, iommu)) {
2262 free_domain_mem(domain);
2267 if (domain_init(domain, gaw))
2270 /* register pcie-to-pci device */
2272 domain = dmar_insert_dev_info(iommu, bridge_bus, bridge_devfn,
2279 domain = dmar_insert_dev_info(iommu, bus, devfn, dev, domain);
2287 static int iommu_identity_mapping;
2288 #define IDENTMAP_ALL 1
2289 #define IDENTMAP_GFX 2
2290 #define IDENTMAP_AZALIA 4
2292 static int iommu_domain_identity_map(struct dmar_domain *domain,
2293 unsigned long long start,
2294 unsigned long long end)
2296 unsigned long first_vpfn = start >> VTD_PAGE_SHIFT;
2297 unsigned long last_vpfn = end >> VTD_PAGE_SHIFT;
2299 if (!reserve_iova(&domain->iovad, dma_to_mm_pfn(first_vpfn),
2300 dma_to_mm_pfn(last_vpfn))) {
2301 printk(KERN_ERR "IOMMU: reserve iova failed\n");
2305 pr_debug("Mapping reserved region %llx-%llx for domain %d\n",
2306 start, end, domain->id);
2308 * RMRR range might have overlap with physical memory range,
2311 dma_pte_clear_range(domain, first_vpfn, last_vpfn);
2313 return domain_pfn_mapping(domain, first_vpfn, first_vpfn,
2314 last_vpfn - first_vpfn + 1,
2315 DMA_PTE_READ|DMA_PTE_WRITE);
2318 static int iommu_prepare_identity_map(struct device *dev,
2319 unsigned long long start,
2320 unsigned long long end)
2322 struct dmar_domain *domain;
2325 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2329 /* For _hardware_ passthrough, don't bother. But for software
2330 passthrough, we do it anyway -- it may indicate a memory
2331 range which is reserved in E820, so which didn't get set
2332 up to start with in si_domain */
2333 if (domain == si_domain && hw_pass_through) {
2334 printk("Ignoring identity map for HW passthrough device %s [0x%Lx - 0x%Lx]\n",
2335 dev_name(dev), start, end);
2340 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
2341 dev_name(dev), start, end);
2344 WARN(1, "Your BIOS is broken; RMRR ends before it starts!\n"
2345 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2346 dmi_get_system_info(DMI_BIOS_VENDOR),
2347 dmi_get_system_info(DMI_BIOS_VERSION),
2348 dmi_get_system_info(DMI_PRODUCT_VERSION));
2353 if (end >> agaw_to_width(domain->agaw)) {
2354 WARN(1, "Your BIOS is broken; RMRR exceeds permitted address width (%d bits)\n"
2355 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2356 agaw_to_width(domain->agaw),
2357 dmi_get_system_info(DMI_BIOS_VENDOR),
2358 dmi_get_system_info(DMI_BIOS_VERSION),
2359 dmi_get_system_info(DMI_PRODUCT_VERSION));
2364 ret = iommu_domain_identity_map(domain, start, end);
2368 /* context entry init */
2369 ret = domain_context_mapping(domain, dev, CONTEXT_TT_MULTI_LEVEL);
2376 domain_exit(domain);
2380 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
2383 if (dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2385 return iommu_prepare_identity_map(dev, rmrr->base_address,
2389 #ifdef CONFIG_INTEL_IOMMU_FLOPPY_WA
2390 static inline void iommu_prepare_isa(void)
2392 struct pci_dev *pdev;
2395 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
2399 printk(KERN_INFO "IOMMU: Prepare 0-16MiB unity mapping for LPC\n");
2400 ret = iommu_prepare_identity_map(&pdev->dev, 0, 16*1024*1024 - 1);
2403 printk(KERN_ERR "IOMMU: Failed to create 0-16MiB identity map; "
2404 "floppy might not work\n");
2409 static inline void iommu_prepare_isa(void)
2413 #endif /* !CONFIG_INTEL_IOMMU_FLPY_WA */
2415 static int md_domain_init(struct dmar_domain *domain, int guest_width);
2417 static int __init si_domain_init(int hw)
2419 struct dmar_drhd_unit *drhd;
2420 struct intel_iommu *iommu;
2423 si_domain = alloc_domain(false);
2427 si_domain->flags = DOMAIN_FLAG_STATIC_IDENTITY;
2429 for_each_active_iommu(iommu, drhd) {
2430 ret = iommu_attach_domain(si_domain, iommu);
2432 domain_exit(si_domain);
2437 if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2438 domain_exit(si_domain);
2442 pr_debug("IOMMU: identity mapping domain is domain %d\n",
2448 for_each_online_node(nid) {
2449 unsigned long start_pfn, end_pfn;
2452 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
2453 ret = iommu_domain_identity_map(si_domain,
2454 PFN_PHYS(start_pfn), PFN_PHYS(end_pfn));
2463 static int identity_mapping(struct device *dev)
2465 struct device_domain_info *info;
2467 if (likely(!iommu_identity_mapping))
2470 info = dev->archdata.iommu;
2471 if (info && info != DUMMY_DEVICE_DOMAIN_INFO)
2472 return (info->domain == si_domain);
2477 static int domain_add_dev_info(struct dmar_domain *domain,
2478 struct device *dev, int translation)
2480 struct dmar_domain *ndomain;
2481 struct intel_iommu *iommu;
2485 iommu = device_to_iommu(dev, &bus, &devfn);
2489 ndomain = dmar_insert_dev_info(iommu, bus, devfn, dev, domain);
2490 if (ndomain != domain)
2493 ret = domain_context_mapping(domain, dev, translation);
2495 domain_remove_one_dev_info(domain, dev);
2502 static bool device_has_rmrr(struct device *dev)
2504 struct dmar_rmrr_unit *rmrr;
2509 for_each_rmrr_units(rmrr) {
2511 * Return TRUE if this RMRR contains the device that
2514 for_each_active_dev_scope(rmrr->devices,
2515 rmrr->devices_cnt, i, tmp)
2525 static int iommu_should_identity_map(struct device *dev, int startup)
2528 if (dev_is_pci(dev)) {
2529 struct pci_dev *pdev = to_pci_dev(dev);
2532 * We want to prevent any device associated with an RMRR from
2533 * getting placed into the SI Domain. This is done because
2534 * problems exist when devices are moved in and out of domains
2535 * and their respective RMRR info is lost. We exempt USB devices
2536 * from this process due to their usage of RMRRs that are known
2537 * to not be needed after BIOS hand-off to OS.
2539 if (device_has_rmrr(dev) &&
2540 (pdev->class >> 8) != PCI_CLASS_SERIAL_USB)
2543 if ((iommu_identity_mapping & IDENTMAP_AZALIA) && IS_AZALIA(pdev))
2546 if ((iommu_identity_mapping & IDENTMAP_GFX) && IS_GFX_DEVICE(pdev))
2549 if (!(iommu_identity_mapping & IDENTMAP_ALL))
2553 * We want to start off with all devices in the 1:1 domain, and
2554 * take them out later if we find they can't access all of memory.
2556 * However, we can't do this for PCI devices behind bridges,
2557 * because all PCI devices behind the same bridge will end up
2558 * with the same source-id on their transactions.
2560 * Practically speaking, we can't change things around for these
2561 * devices at run-time, because we can't be sure there'll be no
2562 * DMA transactions in flight for any of their siblings.
2564 * So PCI devices (unless they're on the root bus) as well as
2565 * their parent PCI-PCI or PCIe-PCI bridges must be left _out_ of
2566 * the 1:1 domain, just in _case_ one of their siblings turns out
2567 * not to be able to map all of memory.
2569 if (!pci_is_pcie(pdev)) {
2570 if (!pci_is_root_bus(pdev->bus))
2572 if (pdev->class >> 8 == PCI_CLASS_BRIDGE_PCI)
2574 } else if (pci_pcie_type(pdev) == PCI_EXP_TYPE_PCI_BRIDGE)
2577 if (device_has_rmrr(dev))
2582 * At boot time, we don't yet know if devices will be 64-bit capable.
2583 * Assume that they will — if they turn out not to be, then we can
2584 * take them out of the 1:1 domain later.
2588 * If the device's dma_mask is less than the system's memory
2589 * size then this is not a candidate for identity mapping.
2591 u64 dma_mask = *dev->dma_mask;
2593 if (dev->coherent_dma_mask &&
2594 dev->coherent_dma_mask < dma_mask)
2595 dma_mask = dev->coherent_dma_mask;
2597 return dma_mask >= dma_get_required_mask(dev);
2603 static int __init dev_prepare_static_identity_mapping(struct device *dev, int hw)
2607 if (!iommu_should_identity_map(dev, 1))
2610 ret = domain_add_dev_info(si_domain, dev,
2611 hw ? CONTEXT_TT_PASS_THROUGH :
2612 CONTEXT_TT_MULTI_LEVEL);
2614 pr_info("IOMMU: %s identity mapping for device %s\n",
2615 hw ? "hardware" : "software", dev_name(dev));
2616 else if (ret == -ENODEV)
2617 /* device not associated with an iommu */
2624 static int __init iommu_prepare_static_identity_mapping(int hw)
2626 struct pci_dev *pdev = NULL;
2627 struct dmar_drhd_unit *drhd;
2628 struct intel_iommu *iommu;
2633 ret = si_domain_init(hw);
2637 for_each_pci_dev(pdev) {
2638 ret = dev_prepare_static_identity_mapping(&pdev->dev, hw);
2643 for_each_active_iommu(iommu, drhd)
2644 for_each_active_dev_scope(drhd->devices, drhd->devices_cnt, i, dev) {
2645 struct acpi_device_physical_node *pn;
2646 struct acpi_device *adev;
2648 if (dev->bus != &acpi_bus_type)
2651 adev= to_acpi_device(dev);
2652 mutex_lock(&adev->physical_node_lock);
2653 list_for_each_entry(pn, &adev->physical_node_list, node) {
2654 ret = dev_prepare_static_identity_mapping(pn->dev, hw);
2658 mutex_unlock(&adev->physical_node_lock);
2666 static int __init init_dmars(void)
2668 struct dmar_drhd_unit *drhd;
2669 struct dmar_rmrr_unit *rmrr;
2671 struct intel_iommu *iommu;
2677 * initialize and program root entry to not present
2680 for_each_drhd_unit(drhd) {
2682 * lock not needed as this is only incremented in the single
2683 * threaded kernel __init code path all other access are read
2686 if (g_num_of_iommus < IOMMU_UNITS_SUPPORTED) {
2690 printk_once(KERN_ERR "intel-iommu: exceeded %d IOMMUs\n",
2691 IOMMU_UNITS_SUPPORTED);
2694 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
2697 printk(KERN_ERR "Allocating global iommu array failed\n");
2702 deferred_flush = kzalloc(g_num_of_iommus *
2703 sizeof(struct deferred_flush_tables), GFP_KERNEL);
2704 if (!deferred_flush) {
2709 for_each_active_iommu(iommu, drhd) {
2710 g_iommus[iommu->seq_id] = iommu;
2712 ret = iommu_init_domains(iommu);
2718 * we could share the same root & context tables
2719 * among all IOMMU's. Need to Split it later.
2721 ret = iommu_alloc_root_entry(iommu);
2723 printk(KERN_ERR "IOMMU: allocate root entry failed\n");
2726 if (!ecap_pass_through(iommu->ecap))
2727 hw_pass_through = 0;
2731 * Start from the sane iommu hardware state.
2733 for_each_active_iommu(iommu, drhd) {
2735 * If the queued invalidation is already initialized by us
2736 * (for example, while enabling interrupt-remapping) then
2737 * we got the things already rolling from a sane state.
2743 * Clear any previous faults.
2745 dmar_fault(-1, iommu);
2747 * Disable queued invalidation if supported and already enabled
2748 * before OS handover.
2750 dmar_disable_qi(iommu);
2753 for_each_active_iommu(iommu, drhd) {
2754 if (dmar_enable_qi(iommu)) {
2756 * Queued Invalidate not enabled, use Register Based
2759 iommu->flush.flush_context = __iommu_flush_context;
2760 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2761 printk(KERN_INFO "IOMMU %d 0x%Lx: using Register based "
2764 (unsigned long long)drhd->reg_base_addr);
2766 iommu->flush.flush_context = qi_flush_context;
2767 iommu->flush.flush_iotlb = qi_flush_iotlb;
2768 printk(KERN_INFO "IOMMU %d 0x%Lx: using Queued "
2771 (unsigned long long)drhd->reg_base_addr);
2775 if (iommu_pass_through)
2776 iommu_identity_mapping |= IDENTMAP_ALL;
2778 #ifdef CONFIG_INTEL_IOMMU_BROKEN_GFX_WA
2779 iommu_identity_mapping |= IDENTMAP_GFX;
2782 check_tylersburg_isoch();
2785 * If pass through is not set or not enabled, setup context entries for
2786 * identity mappings for rmrr, gfx, and isa and may fall back to static
2787 * identity mapping if iommu_identity_mapping is set.
2789 if (iommu_identity_mapping) {
2790 ret = iommu_prepare_static_identity_mapping(hw_pass_through);
2792 printk(KERN_CRIT "Failed to setup IOMMU pass-through\n");
2798 * for each dev attached to rmrr
2800 * locate drhd for dev, alloc domain for dev
2801 * allocate free domain
2802 * allocate page table entries for rmrr
2803 * if context not allocated for bus
2804 * allocate and init context
2805 * set present in root table for this bus
2806 * init context with domain, translation etc
2810 printk(KERN_INFO "IOMMU: Setting RMRR:\n");
2811 for_each_rmrr_units(rmrr) {
2812 /* some BIOS lists non-exist devices in DMAR table. */
2813 for_each_active_dev_scope(rmrr->devices, rmrr->devices_cnt,
2815 ret = iommu_prepare_rmrr_dev(rmrr, dev);
2818 "IOMMU: mapping reserved region failed\n");
2822 iommu_prepare_isa();
2827 * global invalidate context cache
2828 * global invalidate iotlb
2829 * enable translation
2831 for_each_iommu(iommu, drhd) {
2832 if (drhd->ignored) {
2834 * we always have to disable PMRs or DMA may fail on
2838 iommu_disable_protect_mem_regions(iommu);
2842 iommu_flush_write_buffer(iommu);
2844 ret = dmar_set_interrupt(iommu);
2848 iommu_set_root_entry(iommu);
2850 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
2851 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2853 ret = iommu_enable_translation(iommu);
2857 iommu_disable_protect_mem_regions(iommu);
2863 for_each_active_iommu(iommu, drhd)
2864 free_dmar_iommu(iommu);
2865 kfree(deferred_flush);
2872 /* This takes a number of _MM_ pages, not VTD pages */
2873 static struct iova *intel_alloc_iova(struct device *dev,
2874 struct dmar_domain *domain,
2875 unsigned long nrpages, uint64_t dma_mask)
2877 struct iova *iova = NULL;
2879 /* Restrict dma_mask to the width that the iommu can handle */
2880 dma_mask = min_t(uint64_t, DOMAIN_MAX_ADDR(domain->gaw), dma_mask);
2882 if (!dmar_forcedac && dma_mask > DMA_BIT_MASK(32)) {
2884 * First try to allocate an io virtual address in
2885 * DMA_BIT_MASK(32) and if that fails then try allocating
2888 iova = alloc_iova(&domain->iovad, nrpages,
2889 IOVA_PFN(DMA_BIT_MASK(32)), 1);
2893 iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
2894 if (unlikely(!iova)) {
2895 printk(KERN_ERR "Allocating %ld-page iova for %s failed",
2896 nrpages, dev_name(dev));
2903 static struct dmar_domain *__get_valid_domain_for_dev(struct device *dev)
2905 struct dmar_domain *domain;
2908 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2910 printk(KERN_ERR "Allocating domain for %s failed",
2915 /* make sure context mapping is ok */
2916 if (unlikely(!domain_context_mapped(dev))) {
2917 ret = domain_context_mapping(domain, dev, CONTEXT_TT_MULTI_LEVEL);
2919 printk(KERN_ERR "Domain context map for %s failed",
2928 static inline struct dmar_domain *get_valid_domain_for_dev(struct device *dev)
2930 struct device_domain_info *info;
2932 /* No lock here, assumes no domain exit in normal case */
2933 info = dev->archdata.iommu;
2935 return info->domain;
2937 return __get_valid_domain_for_dev(dev);
2940 static int iommu_dummy(struct device *dev)
2942 return dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
2945 /* Check if the dev needs to go through non-identity map and unmap process.*/
2946 static int iommu_no_mapping(struct device *dev)
2950 if (iommu_dummy(dev))
2953 if (!iommu_identity_mapping)
2956 found = identity_mapping(dev);
2958 if (iommu_should_identity_map(dev, 0))
2962 * 32 bit DMA is removed from si_domain and fall back
2963 * to non-identity mapping.
2965 domain_remove_one_dev_info(si_domain, dev);
2966 printk(KERN_INFO "32bit %s uses non-identity mapping\n",
2972 * In case of a detached 64 bit DMA device from vm, the device
2973 * is put into si_domain for identity mapping.
2975 if (iommu_should_identity_map(dev, 0)) {
2977 ret = domain_add_dev_info(si_domain, dev,
2979 CONTEXT_TT_PASS_THROUGH :
2980 CONTEXT_TT_MULTI_LEVEL);
2982 printk(KERN_INFO "64bit %s uses identity mapping\n",
2992 static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
2993 size_t size, int dir, u64 dma_mask)
2995 struct dmar_domain *domain;
2996 phys_addr_t start_paddr;
3000 struct intel_iommu *iommu;
3001 unsigned long paddr_pfn = paddr >> PAGE_SHIFT;
3003 BUG_ON(dir == DMA_NONE);
3005 if (iommu_no_mapping(dev))
3008 domain = get_valid_domain_for_dev(dev);
3012 iommu = domain_get_iommu(domain);
3013 size = aligned_nrpages(paddr, size);
3015 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size), dma_mask);
3020 * Check if DMAR supports zero-length reads on write only
3023 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3024 !cap_zlr(iommu->cap))
3025 prot |= DMA_PTE_READ;
3026 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3027 prot |= DMA_PTE_WRITE;
3029 * paddr - (paddr + size) might be partial page, we should map the whole
3030 * page. Note: if two part of one page are separately mapped, we
3031 * might have two guest_addr mapping to the same host paddr, but this
3032 * is not a big problem
3034 ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
3035 mm_to_dma_pfn(paddr_pfn), size, prot);
3039 /* it's a non-present to present mapping. Only flush if caching mode */
3040 if (cap_caching_mode(iommu->cap))
3041 iommu_flush_iotlb_psi(iommu, domain->id, mm_to_dma_pfn(iova->pfn_lo), size, 0, 1);
3043 iommu_flush_write_buffer(iommu);
3045 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
3046 start_paddr += paddr & ~PAGE_MASK;
3051 __free_iova(&domain->iovad, iova);
3052 printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
3053 dev_name(dev), size, (unsigned long long)paddr, dir);
3057 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
3058 unsigned long offset, size_t size,
3059 enum dma_data_direction dir,
3060 struct dma_attrs *attrs)
3062 return __intel_map_single(dev, page_to_phys(page) + offset, size,
3063 dir, *dev->dma_mask);
3066 static void flush_unmaps(void)
3072 /* just flush them all */
3073 for (i = 0; i < g_num_of_iommus; i++) {
3074 struct intel_iommu *iommu = g_iommus[i];
3078 if (!deferred_flush[i].next)
3081 /* In caching mode, global flushes turn emulation expensive */
3082 if (!cap_caching_mode(iommu->cap))
3083 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3084 DMA_TLB_GLOBAL_FLUSH);
3085 for (j = 0; j < deferred_flush[i].next; j++) {
3087 struct iova *iova = deferred_flush[i].iova[j];
3088 struct dmar_domain *domain = deferred_flush[i].domain[j];
3090 /* On real hardware multiple invalidations are expensive */
3091 if (cap_caching_mode(iommu->cap))
3092 iommu_flush_iotlb_psi(iommu, domain->id,
3093 iova->pfn_lo, iova->pfn_hi - iova->pfn_lo + 1,
3094 !deferred_flush[i].freelist[j], 0);
3096 mask = ilog2(mm_to_dma_pfn(iova->pfn_hi - iova->pfn_lo + 1));
3097 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
3098 (uint64_t)iova->pfn_lo << PAGE_SHIFT, mask);
3100 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
3101 if (deferred_flush[i].freelist[j])
3102 dma_free_pagelist(deferred_flush[i].freelist[j]);
3104 deferred_flush[i].next = 0;
3110 static void flush_unmaps_timeout(unsigned long data)
3112 unsigned long flags;
3114 spin_lock_irqsave(&async_umap_flush_lock, flags);
3116 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3119 static void add_unmap(struct dmar_domain *dom, struct iova *iova, struct page *freelist)
3121 unsigned long flags;
3123 struct intel_iommu *iommu;
3125 spin_lock_irqsave(&async_umap_flush_lock, flags);
3126 if (list_size == HIGH_WATER_MARK)
3129 iommu = domain_get_iommu(dom);
3130 iommu_id = iommu->seq_id;
3132 next = deferred_flush[iommu_id].next;
3133 deferred_flush[iommu_id].domain[next] = dom;
3134 deferred_flush[iommu_id].iova[next] = iova;
3135 deferred_flush[iommu_id].freelist[next] = freelist;
3136 deferred_flush[iommu_id].next++;
3139 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
3143 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3146 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
3147 size_t size, enum dma_data_direction dir,
3148 struct dma_attrs *attrs)
3150 struct dmar_domain *domain;
3151 unsigned long start_pfn, last_pfn;
3153 struct intel_iommu *iommu;
3154 struct page *freelist;
3156 if (iommu_no_mapping(dev))
3159 domain = find_domain(dev);
3162 iommu = domain_get_iommu(domain);
3164 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
3165 if (WARN_ONCE(!iova, "Driver unmaps unmatched page at PFN %llx\n",
3166 (unsigned long long)dev_addr))
3169 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
3170 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
3172 pr_debug("Device %s unmapping: pfn %lx-%lx\n",
3173 dev_name(dev), start_pfn, last_pfn);
3175 freelist = domain_unmap(domain, start_pfn, last_pfn);
3177 if (intel_iommu_strict) {
3178 iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
3179 last_pfn - start_pfn + 1, !freelist, 0);
3181 __free_iova(&domain->iovad, iova);
3182 dma_free_pagelist(freelist);
3184 add_unmap(domain, iova, freelist);
3186 * queue up the release of the unmap to save the 1/6th of the
3187 * cpu used up by the iotlb flush operation...
3192 static void *intel_alloc_coherent(struct device *dev, size_t size,
3193 dma_addr_t *dma_handle, gfp_t flags,
3194 struct dma_attrs *attrs)
3196 struct page *page = NULL;
3199 size = PAGE_ALIGN(size);
3200 order = get_order(size);
3202 if (!iommu_no_mapping(dev))
3203 flags &= ~(GFP_DMA | GFP_DMA32);
3204 else if (dev->coherent_dma_mask < dma_get_required_mask(dev)) {
3205 if (dev->coherent_dma_mask < DMA_BIT_MASK(32))
3211 if (flags & __GFP_WAIT) {
3212 unsigned int count = size >> PAGE_SHIFT;
3214 page = dma_alloc_from_contiguous(dev, count, order);
3215 if (page && iommu_no_mapping(dev) &&
3216 page_to_phys(page) + size > dev->coherent_dma_mask) {
3217 dma_release_from_contiguous(dev, page, count);
3223 page = alloc_pages(flags, order);
3226 memset(page_address(page), 0, size);
3228 *dma_handle = __intel_map_single(dev, page_to_phys(page), size,
3230 dev->coherent_dma_mask);
3232 return page_address(page);
3233 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3234 __free_pages(page, order);
3239 static void intel_free_coherent(struct device *dev, size_t size, void *vaddr,
3240 dma_addr_t dma_handle, struct dma_attrs *attrs)
3243 struct page *page = virt_to_page(vaddr);
3245 size = PAGE_ALIGN(size);
3246 order = get_order(size);
3248 intel_unmap_page(dev, dma_handle, size, DMA_BIDIRECTIONAL, NULL);
3249 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3250 __free_pages(page, order);
3253 static void intel_unmap_sg(struct device *dev, struct scatterlist *sglist,
3254 int nelems, enum dma_data_direction dir,
3255 struct dma_attrs *attrs)
3257 struct dmar_domain *domain;
3258 unsigned long start_pfn, last_pfn;
3260 struct intel_iommu *iommu;
3261 struct page *freelist;
3263 if (iommu_no_mapping(dev))
3266 domain = find_domain(dev);
3269 iommu = domain_get_iommu(domain);
3271 iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
3272 if (WARN_ONCE(!iova, "Driver unmaps unmatched sglist at PFN %llx\n",
3273 (unsigned long long)sglist[0].dma_address))
3276 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
3277 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
3279 freelist = domain_unmap(domain, start_pfn, last_pfn);
3281 if (intel_iommu_strict) {
3282 iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
3283 last_pfn - start_pfn + 1, !freelist, 0);
3285 __free_iova(&domain->iovad, iova);
3286 dma_free_pagelist(freelist);
3288 add_unmap(domain, iova, freelist);
3290 * queue up the release of the unmap to save the 1/6th of the
3291 * cpu used up by the iotlb flush operation...
3296 static int intel_nontranslate_map_sg(struct device *hddev,
3297 struct scatterlist *sglist, int nelems, int dir)
3300 struct scatterlist *sg;
3302 for_each_sg(sglist, sg, nelems, i) {
3303 BUG_ON(!sg_page(sg));
3304 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
3305 sg->dma_length = sg->length;
3310 static int intel_map_sg(struct device *dev, struct scatterlist *sglist, int nelems,
3311 enum dma_data_direction dir, struct dma_attrs *attrs)
3314 struct dmar_domain *domain;
3317 struct iova *iova = NULL;
3319 struct scatterlist *sg;
3320 unsigned long start_vpfn;
3321 struct intel_iommu *iommu;
3323 BUG_ON(dir == DMA_NONE);
3324 if (iommu_no_mapping(dev))
3325 return intel_nontranslate_map_sg(dev, sglist, nelems, dir);
3327 domain = get_valid_domain_for_dev(dev);
3331 iommu = domain_get_iommu(domain);
3333 for_each_sg(sglist, sg, nelems, i)
3334 size += aligned_nrpages(sg->offset, sg->length);
3336 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size),
3339 sglist->dma_length = 0;
3344 * Check if DMAR supports zero-length reads on write only
3347 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3348 !cap_zlr(iommu->cap))
3349 prot |= DMA_PTE_READ;
3350 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3351 prot |= DMA_PTE_WRITE;
3353 start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
3355 ret = domain_sg_mapping(domain, start_vpfn, sglist, size, prot);
3356 if (unlikely(ret)) {
3357 /* clear the page */
3358 dma_pte_clear_range(domain, start_vpfn,
3359 start_vpfn + size - 1);
3360 /* free page tables */
3361 dma_pte_free_pagetable(domain, start_vpfn,
3362 start_vpfn + size - 1);
3364 __free_iova(&domain->iovad, iova);
3368 /* it's a non-present to present mapping. Only flush if caching mode */
3369 if (cap_caching_mode(iommu->cap))
3370 iommu_flush_iotlb_psi(iommu, domain->id, start_vpfn, size, 0, 1);
3372 iommu_flush_write_buffer(iommu);
3377 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
3382 struct dma_map_ops intel_dma_ops = {
3383 .alloc = intel_alloc_coherent,
3384 .free = intel_free_coherent,
3385 .map_sg = intel_map_sg,
3386 .unmap_sg = intel_unmap_sg,
3387 .map_page = intel_map_page,
3388 .unmap_page = intel_unmap_page,
3389 .mapping_error = intel_mapping_error,
3392 static inline int iommu_domain_cache_init(void)
3396 iommu_domain_cache = kmem_cache_create("iommu_domain",
3397 sizeof(struct dmar_domain),
3402 if (!iommu_domain_cache) {
3403 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
3410 static inline int iommu_devinfo_cache_init(void)
3414 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
3415 sizeof(struct device_domain_info),
3419 if (!iommu_devinfo_cache) {
3420 printk(KERN_ERR "Couldn't create devinfo cache\n");
3427 static inline int iommu_iova_cache_init(void)
3431 iommu_iova_cache = kmem_cache_create("iommu_iova",
3432 sizeof(struct iova),
3436 if (!iommu_iova_cache) {
3437 printk(KERN_ERR "Couldn't create iova cache\n");
3444 static int __init iommu_init_mempool(void)
3447 ret = iommu_iova_cache_init();
3451 ret = iommu_domain_cache_init();
3455 ret = iommu_devinfo_cache_init();
3459 kmem_cache_destroy(iommu_domain_cache);
3461 kmem_cache_destroy(iommu_iova_cache);
3466 static void __init iommu_exit_mempool(void)
3468 kmem_cache_destroy(iommu_devinfo_cache);
3469 kmem_cache_destroy(iommu_domain_cache);
3470 kmem_cache_destroy(iommu_iova_cache);
3474 static void quirk_ioat_snb_local_iommu(struct pci_dev *pdev)
3476 struct dmar_drhd_unit *drhd;
3480 /* We know that this device on this chipset has its own IOMMU.
3481 * If we find it under a different IOMMU, then the BIOS is lying
3482 * to us. Hope that the IOMMU for this device is actually
3483 * disabled, and it needs no translation...
3485 rc = pci_bus_read_config_dword(pdev->bus, PCI_DEVFN(0, 0), 0xb0, &vtbar);
3487 /* "can't" happen */
3488 dev_info(&pdev->dev, "failed to run vt-d quirk\n");
3491 vtbar &= 0xffff0000;
3493 /* we know that the this iommu should be at offset 0xa000 from vtbar */
3494 drhd = dmar_find_matched_drhd_unit(pdev);
3495 if (WARN_TAINT_ONCE(!drhd || drhd->reg_base_addr - vtbar != 0xa000,
3496 TAINT_FIRMWARE_WORKAROUND,
3497 "BIOS assigned incorrect VT-d unit for Intel(R) QuickData Technology device\n"))
3498 pdev->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3500 DECLARE_PCI_FIXUP_ENABLE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB, quirk_ioat_snb_local_iommu);
3502 static void __init init_no_remapping_devices(void)
3504 struct dmar_drhd_unit *drhd;
3508 for_each_drhd_unit(drhd) {
3509 if (!drhd->include_all) {
3510 for_each_active_dev_scope(drhd->devices,
3511 drhd->devices_cnt, i, dev)
3513 /* ignore DMAR unit if no devices exist */
3514 if (i == drhd->devices_cnt)
3519 for_each_active_drhd_unit(drhd) {
3520 if (drhd->include_all)
3523 for_each_active_dev_scope(drhd->devices,
3524 drhd->devices_cnt, i, dev)
3525 if (!dev_is_pci(dev) || !IS_GFX_DEVICE(to_pci_dev(dev)))
3527 if (i < drhd->devices_cnt)
3530 /* This IOMMU has *only* gfx devices. Either bypass it or
3531 set the gfx_mapped flag, as appropriate */
3533 intel_iommu_gfx_mapped = 1;
3536 for_each_active_dev_scope(drhd->devices,
3537 drhd->devices_cnt, i, dev)
3538 dev->archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3543 #ifdef CONFIG_SUSPEND
3544 static int init_iommu_hw(void)
3546 struct dmar_drhd_unit *drhd;
3547 struct intel_iommu *iommu = NULL;
3549 for_each_active_iommu(iommu, drhd)
3551 dmar_reenable_qi(iommu);
3553 for_each_iommu(iommu, drhd) {
3554 if (drhd->ignored) {
3556 * we always have to disable PMRs or DMA may fail on
3560 iommu_disable_protect_mem_regions(iommu);
3564 iommu_flush_write_buffer(iommu);
3566 iommu_set_root_entry(iommu);
3568 iommu->flush.flush_context(iommu, 0, 0, 0,
3569 DMA_CCMD_GLOBAL_INVL);
3570 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3571 DMA_TLB_GLOBAL_FLUSH);
3572 if (iommu_enable_translation(iommu))
3574 iommu_disable_protect_mem_regions(iommu);
3580 static void iommu_flush_all(void)
3582 struct dmar_drhd_unit *drhd;
3583 struct intel_iommu *iommu;
3585 for_each_active_iommu(iommu, drhd) {
3586 iommu->flush.flush_context(iommu, 0, 0, 0,
3587 DMA_CCMD_GLOBAL_INVL);
3588 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3589 DMA_TLB_GLOBAL_FLUSH);
3593 static int iommu_suspend(void)
3595 struct dmar_drhd_unit *drhd;
3596 struct intel_iommu *iommu = NULL;
3599 for_each_active_iommu(iommu, drhd) {
3600 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3602 if (!iommu->iommu_state)
3608 for_each_active_iommu(iommu, drhd) {
3609 iommu_disable_translation(iommu);
3611 raw_spin_lock_irqsave(&iommu->register_lock, flag);
3613 iommu->iommu_state[SR_DMAR_FECTL_REG] =
3614 readl(iommu->reg + DMAR_FECTL_REG);
3615 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
3616 readl(iommu->reg + DMAR_FEDATA_REG);
3617 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
3618 readl(iommu->reg + DMAR_FEADDR_REG);
3619 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
3620 readl(iommu->reg + DMAR_FEUADDR_REG);
3622 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
3627 for_each_active_iommu(iommu, drhd)
3628 kfree(iommu->iommu_state);
3633 static void iommu_resume(void)
3635 struct dmar_drhd_unit *drhd;
3636 struct intel_iommu *iommu = NULL;
3639 if (init_iommu_hw()) {
3641 panic("tboot: IOMMU setup failed, DMAR can not resume!\n");
3643 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
3647 for_each_active_iommu(iommu, drhd) {
3649 raw_spin_lock_irqsave(&iommu->register_lock, flag);
3651 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
3652 iommu->reg + DMAR_FECTL_REG);
3653 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
3654 iommu->reg + DMAR_FEDATA_REG);
3655 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
3656 iommu->reg + DMAR_FEADDR_REG);
3657 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
3658 iommu->reg + DMAR_FEUADDR_REG);
3660 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
3663 for_each_active_iommu(iommu, drhd)
3664 kfree(iommu->iommu_state);
3667 static struct syscore_ops iommu_syscore_ops = {
3668 .resume = iommu_resume,
3669 .suspend = iommu_suspend,
3672 static void __init init_iommu_pm_ops(void)
3674 register_syscore_ops(&iommu_syscore_ops);
3678 static inline void init_iommu_pm_ops(void) {}
3679 #endif /* CONFIG_PM */
3682 int __init dmar_parse_one_rmrr(struct acpi_dmar_header *header)
3684 struct acpi_dmar_reserved_memory *rmrr;
3685 struct dmar_rmrr_unit *rmrru;
3687 rmrru = kzalloc(sizeof(*rmrru), GFP_KERNEL);
3691 rmrru->hdr = header;
3692 rmrr = (struct acpi_dmar_reserved_memory *)header;
3693 rmrru->base_address = rmrr->base_address;
3694 rmrru->end_address = rmrr->end_address;
3695 rmrru->devices = dmar_alloc_dev_scope((void *)(rmrr + 1),
3696 ((void *)rmrr) + rmrr->header.length,
3697 &rmrru->devices_cnt);
3698 if (rmrru->devices_cnt && rmrru->devices == NULL) {
3703 list_add(&rmrru->list, &dmar_rmrr_units);
3708 int __init dmar_parse_one_atsr(struct acpi_dmar_header *hdr)
3710 struct acpi_dmar_atsr *atsr;
3711 struct dmar_atsr_unit *atsru;
3713 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
3714 atsru = kzalloc(sizeof(*atsru), GFP_KERNEL);
3719 atsru->include_all = atsr->flags & 0x1;
3720 if (!atsru->include_all) {
3721 atsru->devices = dmar_alloc_dev_scope((void *)(atsr + 1),
3722 (void *)atsr + atsr->header.length,
3723 &atsru->devices_cnt);
3724 if (atsru->devices_cnt && atsru->devices == NULL) {
3730 list_add_rcu(&atsru->list, &dmar_atsr_units);
3735 static void intel_iommu_free_atsr(struct dmar_atsr_unit *atsru)
3737 dmar_free_dev_scope(&atsru->devices, &atsru->devices_cnt);
3741 static void intel_iommu_free_dmars(void)
3743 struct dmar_rmrr_unit *rmrru, *rmrr_n;
3744 struct dmar_atsr_unit *atsru, *atsr_n;
3746 list_for_each_entry_safe(rmrru, rmrr_n, &dmar_rmrr_units, list) {
3747 list_del(&rmrru->list);
3748 dmar_free_dev_scope(&rmrru->devices, &rmrru->devices_cnt);
3752 list_for_each_entry_safe(atsru, atsr_n, &dmar_atsr_units, list) {
3753 list_del(&atsru->list);
3754 intel_iommu_free_atsr(atsru);
3758 int dmar_find_matched_atsr_unit(struct pci_dev *dev)
3761 struct pci_bus *bus;
3762 struct pci_dev *bridge = NULL;
3764 struct acpi_dmar_atsr *atsr;
3765 struct dmar_atsr_unit *atsru;
3767 dev = pci_physfn(dev);
3768 for (bus = dev->bus; bus; bus = bus->parent) {
3770 if (!bridge || !pci_is_pcie(bridge) ||
3771 pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE)
3773 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_ROOT_PORT)
3780 list_for_each_entry_rcu(atsru, &dmar_atsr_units, list) {
3781 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
3782 if (atsr->segment != pci_domain_nr(dev->bus))
3785 for_each_dev_scope(atsru->devices, atsru->devices_cnt, i, tmp)
3786 if (tmp == &bridge->dev)
3789 if (atsru->include_all)
3799 int dmar_iommu_notify_scope_dev(struct dmar_pci_notify_info *info)
3802 struct dmar_rmrr_unit *rmrru;
3803 struct dmar_atsr_unit *atsru;
3804 struct acpi_dmar_atsr *atsr;
3805 struct acpi_dmar_reserved_memory *rmrr;
3807 if (!intel_iommu_enabled && system_state != SYSTEM_BOOTING)
3810 list_for_each_entry(rmrru, &dmar_rmrr_units, list) {
3811 rmrr = container_of(rmrru->hdr,
3812 struct acpi_dmar_reserved_memory, header);
3813 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
3814 ret = dmar_insert_dev_scope(info, (void *)(rmrr + 1),
3815 ((void *)rmrr) + rmrr->header.length,
3816 rmrr->segment, rmrru->devices,
3817 rmrru->devices_cnt);
3820 } else if (info->event == BUS_NOTIFY_DEL_DEVICE) {
3821 dmar_remove_dev_scope(info, rmrr->segment,
3822 rmrru->devices, rmrru->devices_cnt);
3826 list_for_each_entry(atsru, &dmar_atsr_units, list) {
3827 if (atsru->include_all)
3830 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
3831 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
3832 ret = dmar_insert_dev_scope(info, (void *)(atsr + 1),
3833 (void *)atsr + atsr->header.length,
3834 atsr->segment, atsru->devices,
3835 atsru->devices_cnt);
3840 } else if (info->event == BUS_NOTIFY_DEL_DEVICE) {
3841 if (dmar_remove_dev_scope(info, atsr->segment,
3842 atsru->devices, atsru->devices_cnt))
3851 * Here we only respond to action of unbound device from driver.
3853 * Added device is not attached to its DMAR domain here yet. That will happen
3854 * when mapping the device to iova.
3856 static int device_notifier(struct notifier_block *nb,
3857 unsigned long action, void *data)
3859 struct device *dev = data;
3860 struct dmar_domain *domain;
3862 if (iommu_dummy(dev))
3865 if (action != BUS_NOTIFY_UNBOUND_DRIVER &&
3866 action != BUS_NOTIFY_DEL_DEVICE)
3869 domain = find_domain(dev);
3873 down_read(&dmar_global_lock);
3874 domain_remove_one_dev_info(domain, dev);
3875 if (!(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE) &&
3876 !(domain->flags & DOMAIN_FLAG_STATIC_IDENTITY) &&
3877 list_empty(&domain->devices))
3878 domain_exit(domain);
3879 up_read(&dmar_global_lock);
3884 static struct notifier_block device_nb = {
3885 .notifier_call = device_notifier,
3888 static int intel_iommu_memory_notifier(struct notifier_block *nb,
3889 unsigned long val, void *v)
3891 struct memory_notify *mhp = v;
3892 unsigned long long start, end;
3893 unsigned long start_vpfn, last_vpfn;
3896 case MEM_GOING_ONLINE:
3897 start = mhp->start_pfn << PAGE_SHIFT;
3898 end = ((mhp->start_pfn + mhp->nr_pages) << PAGE_SHIFT) - 1;
3899 if (iommu_domain_identity_map(si_domain, start, end)) {
3900 pr_warn("dmar: failed to build identity map for [%llx-%llx]\n",
3907 case MEM_CANCEL_ONLINE:
3908 start_vpfn = mm_to_dma_pfn(mhp->start_pfn);
3909 last_vpfn = mm_to_dma_pfn(mhp->start_pfn + mhp->nr_pages - 1);
3910 while (start_vpfn <= last_vpfn) {
3912 struct dmar_drhd_unit *drhd;
3913 struct intel_iommu *iommu;
3914 struct page *freelist;
3916 iova = find_iova(&si_domain->iovad, start_vpfn);
3918 pr_debug("dmar: failed get IOVA for PFN %lx\n",
3923 iova = split_and_remove_iova(&si_domain->iovad, iova,
3924 start_vpfn, last_vpfn);
3926 pr_warn("dmar: failed to split IOVA PFN [%lx-%lx]\n",
3927 start_vpfn, last_vpfn);
3931 freelist = domain_unmap(si_domain, iova->pfn_lo,
3935 for_each_active_iommu(iommu, drhd)
3936 iommu_flush_iotlb_psi(iommu, si_domain->id,
3938 iova->pfn_hi - iova->pfn_lo + 1,
3941 dma_free_pagelist(freelist);
3943 start_vpfn = iova->pfn_hi + 1;
3944 free_iova_mem(iova);
3952 static struct notifier_block intel_iommu_memory_nb = {
3953 .notifier_call = intel_iommu_memory_notifier,
3957 int __init intel_iommu_init(void)
3960 struct dmar_drhd_unit *drhd;
3961 struct intel_iommu *iommu;
3963 /* VT-d is required for a TXT/tboot launch, so enforce that */
3964 force_on = tboot_force_iommu();
3966 if (iommu_init_mempool()) {
3968 panic("tboot: Failed to initialize iommu memory\n");
3972 down_write(&dmar_global_lock);
3973 if (dmar_table_init()) {
3975 panic("tboot: Failed to initialize DMAR table\n");
3980 * Disable translation if already enabled prior to OS handover.
3982 for_each_active_iommu(iommu, drhd)
3983 if (iommu->gcmd & DMA_GCMD_TE)
3984 iommu_disable_translation(iommu);
3986 if (dmar_dev_scope_init() < 0) {
3988 panic("tboot: Failed to initialize DMAR device scope\n");
3992 if (no_iommu || dmar_disabled)
3995 if (list_empty(&dmar_rmrr_units))
3996 printk(KERN_INFO "DMAR: No RMRR found\n");
3998 if (list_empty(&dmar_atsr_units))
3999 printk(KERN_INFO "DMAR: No ATSR found\n");
4001 if (dmar_init_reserved_ranges()) {
4003 panic("tboot: Failed to reserve iommu ranges\n");
4004 goto out_free_reserved_range;
4007 init_no_remapping_devices();
4012 panic("tboot: Failed to initialize DMARs\n");
4013 printk(KERN_ERR "IOMMU: dmar init failed\n");
4014 goto out_free_reserved_range;
4016 up_write(&dmar_global_lock);
4018 "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
4020 init_timer(&unmap_timer);
4021 #ifdef CONFIG_SWIOTLB
4024 dma_ops = &intel_dma_ops;
4026 init_iommu_pm_ops();
4028 bus_set_iommu(&pci_bus_type, &intel_iommu_ops);
4029 bus_register_notifier(&pci_bus_type, &device_nb);
4030 if (si_domain && !hw_pass_through)
4031 register_memory_notifier(&intel_iommu_memory_nb);
4033 intel_iommu_enabled = 1;
4037 out_free_reserved_range:
4038 put_iova_domain(&reserved_iova_list);
4040 intel_iommu_free_dmars();
4041 up_write(&dmar_global_lock);
4042 iommu_exit_mempool();
4046 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
4049 struct pci_dev *tmp, *parent, *pdev;
4051 if (!iommu || !dev || !dev_is_pci(dev))
4054 pdev = to_pci_dev(dev);
4056 /* dependent device detach */
4057 tmp = pci_find_upstream_pcie_bridge(pdev);
4058 /* Secondary interface's bus number and devfn 0 */
4060 parent = pdev->bus->self;
4061 while (parent != tmp) {
4062 iommu_detach_dev(iommu, parent->bus->number,
4064 parent = parent->bus->self;
4066 if (pci_is_pcie(tmp)) /* this is a PCIe-to-PCI bridge */
4067 iommu_detach_dev(iommu,
4068 tmp->subordinate->number, 0);
4069 else /* this is a legacy PCI bridge */
4070 iommu_detach_dev(iommu, tmp->bus->number,
4075 static void domain_remove_one_dev_info(struct dmar_domain *domain,
4078 struct device_domain_info *info, *tmp;
4079 struct intel_iommu *iommu;
4080 unsigned long flags;
4084 iommu = device_to_iommu(dev, &bus, &devfn);
4088 spin_lock_irqsave(&device_domain_lock, flags);
4089 list_for_each_entry_safe(info, tmp, &domain->devices, link) {
4090 if (info->iommu == iommu && info->bus == bus &&
4091 info->devfn == devfn) {
4092 unlink_domain_info(info);
4093 spin_unlock_irqrestore(&device_domain_lock, flags);
4095 iommu_disable_dev_iotlb(info);
4096 iommu_detach_dev(iommu, info->bus, info->devfn);
4097 iommu_detach_dependent_devices(iommu, dev);
4098 free_devinfo_mem(info);
4100 spin_lock_irqsave(&device_domain_lock, flags);
4108 /* if there is no other devices under the same iommu
4109 * owned by this domain, clear this iommu in iommu_bmp
4110 * update iommu count and coherency
4112 if (info->iommu == iommu)
4116 spin_unlock_irqrestore(&device_domain_lock, flags);
4119 unsigned long tmp_flags;
4120 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
4121 clear_bit(iommu->seq_id, domain->iommu_bmp);
4122 domain->iommu_count--;
4123 domain_update_iommu_cap(domain);
4124 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
4126 if (!(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE) &&
4127 !(domain->flags & DOMAIN_FLAG_STATIC_IDENTITY)) {
4128 spin_lock_irqsave(&iommu->lock, tmp_flags);
4129 clear_bit(domain->id, iommu->domain_ids);
4130 iommu->domains[domain->id] = NULL;
4131 spin_unlock_irqrestore(&iommu->lock, tmp_flags);
4136 static int md_domain_init(struct dmar_domain *domain, int guest_width)
4140 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
4141 domain_reserve_special_ranges(domain);
4143 /* calculate AGAW */
4144 domain->gaw = guest_width;
4145 adjust_width = guestwidth_to_adjustwidth(guest_width);
4146 domain->agaw = width_to_agaw(adjust_width);
4148 domain->iommu_coherency = 0;
4149 domain->iommu_snooping = 0;
4150 domain->iommu_superpage = 0;
4151 domain->max_addr = 0;
4153 /* always allocate the top pgd */
4154 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
4157 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
4161 static int intel_iommu_domain_init(struct iommu_domain *domain)
4163 struct dmar_domain *dmar_domain;
4165 dmar_domain = alloc_domain(true);
4168 "intel_iommu_domain_init: dmar_domain == NULL\n");
4171 if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
4173 "intel_iommu_domain_init() failed\n");
4174 domain_exit(dmar_domain);
4177 domain_update_iommu_cap(dmar_domain);
4178 domain->priv = dmar_domain;
4180 domain->geometry.aperture_start = 0;
4181 domain->geometry.aperture_end = __DOMAIN_MAX_ADDR(dmar_domain->gaw);
4182 domain->geometry.force_aperture = true;
4187 static void intel_iommu_domain_destroy(struct iommu_domain *domain)
4189 struct dmar_domain *dmar_domain = domain->priv;
4191 domain->priv = NULL;
4192 domain_exit(dmar_domain);
4195 static int intel_iommu_attach_device(struct iommu_domain *domain,
4198 struct dmar_domain *dmar_domain = domain->priv;
4199 struct intel_iommu *iommu;
4203 /* normally dev is not mapped */
4204 if (unlikely(domain_context_mapped(dev))) {
4205 struct dmar_domain *old_domain;
4207 old_domain = find_domain(dev);
4209 if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
4210 dmar_domain->flags & DOMAIN_FLAG_STATIC_IDENTITY)
4211 domain_remove_one_dev_info(old_domain, dev);
4213 domain_remove_dev_info(old_domain);
4217 iommu = device_to_iommu(dev, &bus, &devfn);
4221 /* check if this iommu agaw is sufficient for max mapped address */
4222 addr_width = agaw_to_width(iommu->agaw);
4223 if (addr_width > cap_mgaw(iommu->cap))
4224 addr_width = cap_mgaw(iommu->cap);
4226 if (dmar_domain->max_addr > (1LL << addr_width)) {
4227 printk(KERN_ERR "%s: iommu width (%d) is not "
4228 "sufficient for the mapped address (%llx)\n",
4229 __func__, addr_width, dmar_domain->max_addr);
4232 dmar_domain->gaw = addr_width;
4235 * Knock out extra levels of page tables if necessary
4237 while (iommu->agaw < dmar_domain->agaw) {
4238 struct dma_pte *pte;
4240 pte = dmar_domain->pgd;
4241 if (dma_pte_present(pte)) {
4242 dmar_domain->pgd = (struct dma_pte *)
4243 phys_to_virt(dma_pte_addr(pte));
4244 free_pgtable_page(pte);
4246 dmar_domain->agaw--;
4249 return domain_add_dev_info(dmar_domain, dev, CONTEXT_TT_MULTI_LEVEL);
4252 static void intel_iommu_detach_device(struct iommu_domain *domain,
4255 struct dmar_domain *dmar_domain = domain->priv;
4257 domain_remove_one_dev_info(dmar_domain, dev);
4260 static int intel_iommu_map(struct iommu_domain *domain,
4261 unsigned long iova, phys_addr_t hpa,
4262 size_t size, int iommu_prot)
4264 struct dmar_domain *dmar_domain = domain->priv;
4269 if (iommu_prot & IOMMU_READ)
4270 prot |= DMA_PTE_READ;
4271 if (iommu_prot & IOMMU_WRITE)
4272 prot |= DMA_PTE_WRITE;
4273 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
4274 prot |= DMA_PTE_SNP;
4276 max_addr = iova + size;
4277 if (dmar_domain->max_addr < max_addr) {
4280 /* check if minimum agaw is sufficient for mapped address */
4281 end = __DOMAIN_MAX_ADDR(dmar_domain->gaw) + 1;
4282 if (end < max_addr) {
4283 printk(KERN_ERR "%s: iommu width (%d) is not "
4284 "sufficient for the mapped address (%llx)\n",
4285 __func__, dmar_domain->gaw, max_addr);
4288 dmar_domain->max_addr = max_addr;
4290 /* Round up size to next multiple of PAGE_SIZE, if it and
4291 the low bits of hpa would take us onto the next page */
4292 size = aligned_nrpages(hpa, size);
4293 ret = domain_pfn_mapping(dmar_domain, iova >> VTD_PAGE_SHIFT,
4294 hpa >> VTD_PAGE_SHIFT, size, prot);
4298 static size_t intel_iommu_unmap(struct iommu_domain *domain,
4299 unsigned long iova, size_t size)
4301 struct dmar_domain *dmar_domain = domain->priv;
4302 struct page *freelist = NULL;
4303 struct intel_iommu *iommu;
4304 unsigned long start_pfn, last_pfn;
4305 unsigned int npages;
4306 int iommu_id, num, ndomains, level = 0;
4308 /* Cope with horrid API which requires us to unmap more than the
4309 size argument if it happens to be a large-page mapping. */
4310 if (!pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level))
4313 if (size < VTD_PAGE_SIZE << level_to_offset_bits(level))
4314 size = VTD_PAGE_SIZE << level_to_offset_bits(level);
4316 start_pfn = iova >> VTD_PAGE_SHIFT;
4317 last_pfn = (iova + size - 1) >> VTD_PAGE_SHIFT;
4319 freelist = domain_unmap(dmar_domain, start_pfn, last_pfn);
4321 npages = last_pfn - start_pfn + 1;
4323 for_each_set_bit(iommu_id, dmar_domain->iommu_bmp, g_num_of_iommus) {
4324 iommu = g_iommus[iommu_id];
4327 * find bit position of dmar_domain
4329 ndomains = cap_ndoms(iommu->cap);
4330 for_each_set_bit(num, iommu->domain_ids, ndomains) {
4331 if (iommu->domains[num] == dmar_domain)
4332 iommu_flush_iotlb_psi(iommu, num, start_pfn,
4333 npages, !freelist, 0);
4338 dma_free_pagelist(freelist);
4340 if (dmar_domain->max_addr == iova + size)
4341 dmar_domain->max_addr = iova;
4346 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
4349 struct dmar_domain *dmar_domain = domain->priv;
4350 struct dma_pte *pte;
4354 pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level);
4356 phys = dma_pte_addr(pte);
4361 static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
4364 struct dmar_domain *dmar_domain = domain->priv;
4366 if (cap == IOMMU_CAP_CACHE_COHERENCY)
4367 return dmar_domain->iommu_snooping;
4368 if (cap == IOMMU_CAP_INTR_REMAP)
4369 return irq_remapping_enabled;
4374 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
4376 static int intel_iommu_add_device(struct device *dev)
4378 struct pci_dev *pdev = to_pci_dev(dev);
4379 struct pci_dev *bridge, *dma_pdev = NULL;
4380 struct iommu_group *group;
4384 if (!device_to_iommu(dev, &bus, &devfn))
4387 bridge = pci_find_upstream_pcie_bridge(pdev);
4389 if (pci_is_pcie(bridge))
4390 dma_pdev = pci_get_domain_bus_and_slot(
4391 pci_domain_nr(pdev->bus),
4392 bridge->subordinate->number, 0);
4394 dma_pdev = pci_dev_get(bridge);
4396 dma_pdev = pci_dev_get(pdev);
4398 /* Account for quirked devices */
4399 swap_pci_ref(&dma_pdev, pci_get_dma_source(dma_pdev));
4402 * If it's a multifunction device that does not support our
4403 * required ACS flags, add to the same group as lowest numbered
4404 * function that also does not suport the required ACS flags.
4406 if (dma_pdev->multifunction &&
4407 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS)) {
4408 u8 i, slot = PCI_SLOT(dma_pdev->devfn);
4410 for (i = 0; i < 8; i++) {
4411 struct pci_dev *tmp;
4413 tmp = pci_get_slot(dma_pdev->bus, PCI_DEVFN(slot, i));
4417 if (!pci_acs_enabled(tmp, REQ_ACS_FLAGS)) {
4418 swap_pci_ref(&dma_pdev, tmp);
4426 * Devices on the root bus go through the iommu. If that's not us,
4427 * find the next upstream device and test ACS up to the root bus.
4428 * Finding the next device may require skipping virtual buses.
4430 while (!pci_is_root_bus(dma_pdev->bus)) {
4431 struct pci_bus *bus = dma_pdev->bus;
4433 while (!bus->self) {
4434 if (!pci_is_root_bus(bus))
4440 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
4443 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
4447 group = iommu_group_get(&dma_pdev->dev);
4448 pci_dev_put(dma_pdev);
4450 group = iommu_group_alloc();
4452 return PTR_ERR(group);
4455 ret = iommu_group_add_device(group, dev);
4457 iommu_group_put(group);
4461 static void intel_iommu_remove_device(struct device *dev)
4463 iommu_group_remove_device(dev);
4466 static struct iommu_ops intel_iommu_ops = {
4467 .domain_init = intel_iommu_domain_init,
4468 .domain_destroy = intel_iommu_domain_destroy,
4469 .attach_dev = intel_iommu_attach_device,
4470 .detach_dev = intel_iommu_detach_device,
4471 .map = intel_iommu_map,
4472 .unmap = intel_iommu_unmap,
4473 .iova_to_phys = intel_iommu_iova_to_phys,
4474 .domain_has_cap = intel_iommu_domain_has_cap,
4475 .add_device = intel_iommu_add_device,
4476 .remove_device = intel_iommu_remove_device,
4477 .pgsize_bitmap = INTEL_IOMMU_PGSIZES,
4480 static void quirk_iommu_g4x_gfx(struct pci_dev *dev)
4482 /* G4x/GM45 integrated gfx dmar support is totally busted. */
4483 printk(KERN_INFO "DMAR: Disabling IOMMU for graphics on this chipset\n");
4487 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_g4x_gfx);
4488 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_g4x_gfx);
4489 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_g4x_gfx);
4490 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_g4x_gfx);
4491 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_g4x_gfx);
4492 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_g4x_gfx);
4493 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_g4x_gfx);
4495 static void quirk_iommu_rwbf(struct pci_dev *dev)
4498 * Mobile 4 Series Chipset neglects to set RWBF capability,
4499 * but needs it. Same seems to hold for the desktop versions.
4501 printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
4505 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);
4506 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_rwbf);
4507 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_rwbf);
4508 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_rwbf);
4509 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_rwbf);
4510 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_rwbf);
4511 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_rwbf);
4514 #define GGC_MEMORY_SIZE_MASK (0xf << 8)
4515 #define GGC_MEMORY_SIZE_NONE (0x0 << 8)
4516 #define GGC_MEMORY_SIZE_1M (0x1 << 8)
4517 #define GGC_MEMORY_SIZE_2M (0x3 << 8)
4518 #define GGC_MEMORY_VT_ENABLED (0x8 << 8)
4519 #define GGC_MEMORY_SIZE_2M_VT (0x9 << 8)
4520 #define GGC_MEMORY_SIZE_3M_VT (0xa << 8)
4521 #define GGC_MEMORY_SIZE_4M_VT (0xb << 8)
4523 static void quirk_calpella_no_shadow_gtt(struct pci_dev *dev)
4527 if (pci_read_config_word(dev, GGC, &ggc))
4530 if (!(ggc & GGC_MEMORY_VT_ENABLED)) {
4531 printk(KERN_INFO "DMAR: BIOS has allocated no shadow GTT; disabling IOMMU for graphics\n");
4533 } else if (dmar_map_gfx) {
4534 /* we have to ensure the gfx device is idle before we flush */
4535 printk(KERN_INFO "DMAR: Disabling batched IOTLB flush on Ironlake\n");
4536 intel_iommu_strict = 1;
4539 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0040, quirk_calpella_no_shadow_gtt);
4540 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0044, quirk_calpella_no_shadow_gtt);
4541 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0062, quirk_calpella_no_shadow_gtt);
4542 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x006a, quirk_calpella_no_shadow_gtt);
4544 /* On Tylersburg chipsets, some BIOSes have been known to enable the
4545 ISOCH DMAR unit for the Azalia sound device, but not give it any
4546 TLB entries, which causes it to deadlock. Check for that. We do
4547 this in a function called from init_dmars(), instead of in a PCI
4548 quirk, because we don't want to print the obnoxious "BIOS broken"
4549 message if VT-d is actually disabled.
4551 static void __init check_tylersburg_isoch(void)
4553 struct pci_dev *pdev;
4554 uint32_t vtisochctrl;
4556 /* If there's no Azalia in the system anyway, forget it. */
4557 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x3a3e, NULL);
4562 /* System Management Registers. Might be hidden, in which case
4563 we can't do the sanity check. But that's OK, because the
4564 known-broken BIOSes _don't_ actually hide it, so far. */
4565 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x342e, NULL);
4569 if (pci_read_config_dword(pdev, 0x188, &vtisochctrl)) {
4576 /* If Azalia DMA is routed to the non-isoch DMAR unit, fine. */
4577 if (vtisochctrl & 1)
4580 /* Drop all bits other than the number of TLB entries */
4581 vtisochctrl &= 0x1c;
4583 /* If we have the recommended number of TLB entries (16), fine. */
4584 if (vtisochctrl == 0x10)
4587 /* Zero TLB entries? You get to ride the short bus to school. */
4589 WARN(1, "Your BIOS is broken; DMA routed to ISOCH DMAR unit but no TLB space.\n"
4590 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
4591 dmi_get_system_info(DMI_BIOS_VENDOR),
4592 dmi_get_system_info(DMI_BIOS_VERSION),
4593 dmi_get_system_info(DMI_PRODUCT_VERSION));
4594 iommu_identity_mapping |= IDENTMAP_AZALIA;
4598 printk(KERN_WARNING "DMAR: Recommended TLB entries for ISOCH unit is 16; your BIOS set %d\n",