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>
18 * Joerg Roedel <jroedel@suse.de>
21 #define pr_fmt(fmt) "DMAR: " fmt
23 #include <linux/init.h>
24 #include <linux/bitmap.h>
25 #include <linux/debugfs.h>
26 #include <linux/export.h>
27 #include <linux/slab.h>
28 #include <linux/irq.h>
29 #include <linux/interrupt.h>
30 #include <linux/spinlock.h>
31 #include <linux/pci.h>
32 #include <linux/dmar.h>
33 #include <linux/dma-mapping.h>
34 #include <linux/mempool.h>
35 #include <linux/memory.h>
36 #include <linux/timer.h>
37 #include <linux/iova.h>
38 #include <linux/iommu.h>
39 #include <linux/intel-iommu.h>
40 #include <linux/syscore_ops.h>
41 #include <linux/tboot.h>
42 #include <linux/dmi.h>
43 #include <linux/pci-ats.h>
44 #include <linux/memblock.h>
45 #include <linux/dma-contiguous.h>
46 #include <linux/crash_dump.h>
47 #include <asm/irq_remapping.h>
48 #include <asm/cacheflush.h>
49 #include <asm/iommu.h>
51 #include "irq_remapping.h"
53 #define ROOT_SIZE VTD_PAGE_SIZE
54 #define CONTEXT_SIZE VTD_PAGE_SIZE
56 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
57 #define IS_USB_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_SERIAL_USB)
58 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
59 #define IS_AZALIA(pdev) ((pdev)->vendor == 0x8086 && (pdev)->device == 0x3a3e)
61 #define IOAPIC_RANGE_START (0xfee00000)
62 #define IOAPIC_RANGE_END (0xfeefffff)
63 #define IOVA_START_ADDR (0x1000)
65 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
67 #define MAX_AGAW_WIDTH 64
68 #define MAX_AGAW_PFN_WIDTH (MAX_AGAW_WIDTH - VTD_PAGE_SHIFT)
70 #define __DOMAIN_MAX_PFN(gaw) ((((uint64_t)1) << (gaw-VTD_PAGE_SHIFT)) - 1)
71 #define __DOMAIN_MAX_ADDR(gaw) ((((uint64_t)1) << gaw) - 1)
73 /* We limit DOMAIN_MAX_PFN to fit in an unsigned long, and DOMAIN_MAX_ADDR
74 to match. That way, we can use 'unsigned long' for PFNs with impunity. */
75 #define DOMAIN_MAX_PFN(gaw) ((unsigned long) min_t(uint64_t, \
76 __DOMAIN_MAX_PFN(gaw), (unsigned long)-1))
77 #define DOMAIN_MAX_ADDR(gaw) (((uint64_t)__DOMAIN_MAX_PFN(gaw)) << VTD_PAGE_SHIFT)
79 /* IO virtual address start page frame number */
80 #define IOVA_START_PFN (1)
82 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
83 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
84 #define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
86 /* page table handling */
87 #define LEVEL_STRIDE (9)
88 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
91 * This bitmap is used to advertise the page sizes our hardware support
92 * to the IOMMU core, which will then use this information to split
93 * physically contiguous memory regions it is mapping into page sizes
96 * Traditionally the IOMMU core just handed us the mappings directly,
97 * after making sure the size is an order of a 4KiB page and that the
98 * mapping has natural alignment.
100 * To retain this behavior, we currently advertise that we support
101 * all page sizes that are an order of 4KiB.
103 * If at some point we'd like to utilize the IOMMU core's new behavior,
104 * we could change this to advertise the real page sizes we support.
106 #define INTEL_IOMMU_PGSIZES (~0xFFFUL)
108 static inline int agaw_to_level(int agaw)
113 static inline int agaw_to_width(int agaw)
115 return min_t(int, 30 + agaw * LEVEL_STRIDE, MAX_AGAW_WIDTH);
118 static inline int width_to_agaw(int width)
120 return DIV_ROUND_UP(width - 30, LEVEL_STRIDE);
123 static inline unsigned int level_to_offset_bits(int level)
125 return (level - 1) * LEVEL_STRIDE;
128 static inline int pfn_level_offset(unsigned long pfn, int level)
130 return (pfn >> level_to_offset_bits(level)) & LEVEL_MASK;
133 static inline unsigned long level_mask(int level)
135 return -1UL << level_to_offset_bits(level);
138 static inline unsigned long level_size(int level)
140 return 1UL << level_to_offset_bits(level);
143 static inline unsigned long align_to_level(unsigned long pfn, int level)
145 return (pfn + level_size(level) - 1) & level_mask(level);
148 static inline unsigned long lvl_to_nr_pages(unsigned int lvl)
150 return 1 << min_t(int, (lvl - 1) * LEVEL_STRIDE, MAX_AGAW_PFN_WIDTH);
153 /* VT-d pages must always be _smaller_ than MM pages. Otherwise things
154 are never going to work. */
155 static inline unsigned long dma_to_mm_pfn(unsigned long dma_pfn)
157 return dma_pfn >> (PAGE_SHIFT - VTD_PAGE_SHIFT);
160 static inline unsigned long mm_to_dma_pfn(unsigned long mm_pfn)
162 return mm_pfn << (PAGE_SHIFT - VTD_PAGE_SHIFT);
164 static inline unsigned long page_to_dma_pfn(struct page *pg)
166 return mm_to_dma_pfn(page_to_pfn(pg));
168 static inline unsigned long virt_to_dma_pfn(void *p)
170 return page_to_dma_pfn(virt_to_page(p));
173 /* global iommu list, set NULL for ignored DMAR units */
174 static struct intel_iommu **g_iommus;
176 static void __init check_tylersburg_isoch(void);
177 static int rwbf_quirk;
180 * set to 1 to panic kernel if can't successfully enable VT-d
181 * (used when kernel is launched w/ TXT)
183 static int force_on = 0;
188 * 12-63: Context Ptr (12 - (haw-1))
195 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
198 * Take a root_entry and return the Lower Context Table Pointer (LCTP)
201 static phys_addr_t root_entry_lctp(struct root_entry *re)
206 return re->lo & VTD_PAGE_MASK;
210 * Take a root_entry and return the Upper Context Table Pointer (UCTP)
213 static phys_addr_t root_entry_uctp(struct root_entry *re)
218 return re->hi & VTD_PAGE_MASK;
223 * 1: fault processing disable
224 * 2-3: translation type
225 * 12-63: address space root
231 struct context_entry {
236 static inline void context_clear_pasid_enable(struct context_entry *context)
238 context->lo &= ~(1ULL << 11);
241 static inline bool context_pasid_enabled(struct context_entry *context)
243 return !!(context->lo & (1ULL << 11));
246 static inline void context_set_copied(struct context_entry *context)
248 context->hi |= (1ull << 3);
251 static inline bool context_copied(struct context_entry *context)
253 return !!(context->hi & (1ULL << 3));
256 static inline bool __context_present(struct context_entry *context)
258 return (context->lo & 1);
261 static inline bool context_present(struct context_entry *context)
263 return context_pasid_enabled(context) ?
264 __context_present(context) :
265 __context_present(context) && !context_copied(context);
268 static inline void context_set_present(struct context_entry *context)
273 static inline void context_set_fault_enable(struct context_entry *context)
275 context->lo &= (((u64)-1) << 2) | 1;
278 static inline void context_set_translation_type(struct context_entry *context,
281 context->lo &= (((u64)-1) << 4) | 3;
282 context->lo |= (value & 3) << 2;
285 static inline void context_set_address_root(struct context_entry *context,
288 context->lo &= ~VTD_PAGE_MASK;
289 context->lo |= value & VTD_PAGE_MASK;
292 static inline void context_set_address_width(struct context_entry *context,
295 context->hi |= value & 7;
298 static inline void context_set_domain_id(struct context_entry *context,
301 context->hi |= (value & ((1 << 16) - 1)) << 8;
304 static inline int context_domain_id(struct context_entry *c)
306 return((c->hi >> 8) & 0xffff);
309 static inline void context_clear_entry(struct context_entry *context)
322 * 12-63: Host physcial address
328 static inline void dma_clear_pte(struct dma_pte *pte)
333 static inline u64 dma_pte_addr(struct dma_pte *pte)
336 return pte->val & VTD_PAGE_MASK;
338 /* Must have a full atomic 64-bit read */
339 return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
343 static inline bool dma_pte_present(struct dma_pte *pte)
345 return (pte->val & 3) != 0;
348 static inline bool dma_pte_superpage(struct dma_pte *pte)
350 return (pte->val & DMA_PTE_LARGE_PAGE);
353 static inline int first_pte_in_page(struct dma_pte *pte)
355 return !((unsigned long)pte & ~VTD_PAGE_MASK);
359 * This domain is a statically identity mapping domain.
360 * 1. This domain creats a static 1:1 mapping to all usable memory.
361 * 2. It maps to each iommu if successful.
362 * 3. Each iommu mapps to this domain if successful.
364 static struct dmar_domain *si_domain;
365 static int hw_pass_through = 1;
368 * Domain represents a virtual machine, more than one devices
369 * across iommus may be owned in one domain, e.g. kvm guest.
371 #define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 0)
373 /* si_domain contains mulitple devices */
374 #define DOMAIN_FLAG_STATIC_IDENTITY (1 << 1)
376 #define for_each_domain_iommu(idx, domain) \
377 for (idx = 0; idx < g_num_of_iommus; idx++) \
378 if (domain->iommu_refcnt[idx])
381 int nid; /* node id */
383 unsigned iommu_refcnt[DMAR_UNITS_SUPPORTED];
384 /* Refcount of devices per iommu */
387 u16 iommu_did[DMAR_UNITS_SUPPORTED];
388 /* Domain ids per IOMMU. Use u16 since
389 * domain ids are 16 bit wide according
390 * to VT-d spec, section 9.3 */
392 struct list_head devices; /* all devices' list */
393 struct iova_domain iovad; /* iova's that belong to this domain */
395 struct dma_pte *pgd; /* virtual address */
396 int gaw; /* max guest address width */
398 /* adjusted guest address width, 0 is level 2 30-bit */
401 int flags; /* flags to find out type of domain */
403 int iommu_coherency;/* indicate coherency of iommu access */
404 int iommu_snooping; /* indicate snooping control feature*/
405 int iommu_count; /* reference count of iommu */
406 int iommu_superpage;/* Level of superpages supported:
407 0 == 4KiB (no superpages), 1 == 2MiB,
408 2 == 1GiB, 3 == 512GiB, 4 == 1TiB */
409 spinlock_t iommu_lock; /* protect iommu set in domain */
410 u64 max_addr; /* maximum mapped address */
412 struct iommu_domain domain; /* generic domain data structure for
416 /* PCI domain-device relationship */
417 struct device_domain_info {
418 struct list_head link; /* link to domain siblings */
419 struct list_head global; /* link to global list */
420 u8 bus; /* PCI bus number */
421 u8 devfn; /* PCI devfn number */
422 struct device *dev; /* it's NULL for PCIe-to-PCI bridge */
423 struct intel_iommu *iommu; /* IOMMU used by this device */
424 struct dmar_domain *domain; /* pointer to domain */
427 struct dmar_rmrr_unit {
428 struct list_head list; /* list of rmrr units */
429 struct acpi_dmar_header *hdr; /* ACPI header */
430 u64 base_address; /* reserved base address*/
431 u64 end_address; /* reserved end address */
432 struct dmar_dev_scope *devices; /* target devices */
433 int devices_cnt; /* target device count */
436 struct dmar_atsr_unit {
437 struct list_head list; /* list of ATSR units */
438 struct acpi_dmar_header *hdr; /* ACPI header */
439 struct dmar_dev_scope *devices; /* target devices */
440 int devices_cnt; /* target device count */
441 u8 include_all:1; /* include all ports */
444 static LIST_HEAD(dmar_atsr_units);
445 static LIST_HEAD(dmar_rmrr_units);
447 #define for_each_rmrr_units(rmrr) \
448 list_for_each_entry(rmrr, &dmar_rmrr_units, list)
450 static void flush_unmaps_timeout(unsigned long data);
452 static DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
454 #define HIGH_WATER_MARK 250
455 struct deferred_flush_tables {
457 struct iova *iova[HIGH_WATER_MARK];
458 struct dmar_domain *domain[HIGH_WATER_MARK];
459 struct page *freelist[HIGH_WATER_MARK];
462 static struct deferred_flush_tables *deferred_flush;
464 /* bitmap for indexing intel_iommus */
465 static int g_num_of_iommus;
467 static DEFINE_SPINLOCK(async_umap_flush_lock);
468 static LIST_HEAD(unmaps_to_do);
471 static long list_size;
473 static void domain_exit(struct dmar_domain *domain);
474 static void domain_remove_dev_info(struct dmar_domain *domain);
475 static void dmar_remove_one_dev_info(struct dmar_domain *domain,
477 static void domain_context_clear(struct intel_iommu *iommu,
479 static int domain_detach_iommu(struct dmar_domain *domain,
480 struct intel_iommu *iommu);
482 #ifdef CONFIG_INTEL_IOMMU_DEFAULT_ON
483 int dmar_disabled = 0;
485 int dmar_disabled = 1;
486 #endif /*CONFIG_INTEL_IOMMU_DEFAULT_ON*/
488 int intel_iommu_enabled = 0;
489 EXPORT_SYMBOL_GPL(intel_iommu_enabled);
491 static int dmar_map_gfx = 1;
492 static int dmar_forcedac;
493 static int intel_iommu_strict;
494 static int intel_iommu_superpage = 1;
495 static int intel_iommu_ecs = 1;
497 /* We only actually use ECS when PASID support (on the new bit 40)
498 * is also advertised. Some early implementations — the ones with
499 * PASID support on bit 28 — have issues even when we *only* use
500 * extended root/context tables. */
501 #define ecs_enabled(iommu) (intel_iommu_ecs && ecap_ecs(iommu->ecap) && \
502 ecap_pasid(iommu->ecap))
504 int intel_iommu_gfx_mapped;
505 EXPORT_SYMBOL_GPL(intel_iommu_gfx_mapped);
507 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
508 static DEFINE_SPINLOCK(device_domain_lock);
509 static LIST_HEAD(device_domain_list);
511 static const struct iommu_ops intel_iommu_ops;
513 static bool translation_pre_enabled(struct intel_iommu *iommu)
515 return (iommu->flags & VTD_FLAG_TRANS_PRE_ENABLED);
518 static void clear_translation_pre_enabled(struct intel_iommu *iommu)
520 iommu->flags &= ~VTD_FLAG_TRANS_PRE_ENABLED;
523 static void init_translation_status(struct intel_iommu *iommu)
527 gsts = readl(iommu->reg + DMAR_GSTS_REG);
528 if (gsts & DMA_GSTS_TES)
529 iommu->flags |= VTD_FLAG_TRANS_PRE_ENABLED;
532 /* Convert generic 'struct iommu_domain to private struct dmar_domain */
533 static struct dmar_domain *to_dmar_domain(struct iommu_domain *dom)
535 return container_of(dom, struct dmar_domain, domain);
538 static int __init intel_iommu_setup(char *str)
543 if (!strncmp(str, "on", 2)) {
545 pr_info("IOMMU enabled\n");
546 } else if (!strncmp(str, "off", 3)) {
548 pr_info("IOMMU disabled\n");
549 } else if (!strncmp(str, "igfx_off", 8)) {
551 pr_info("Disable GFX device mapping\n");
552 } else if (!strncmp(str, "forcedac", 8)) {
553 pr_info("Forcing DAC for PCI devices\n");
555 } else if (!strncmp(str, "strict", 6)) {
556 pr_info("Disable batched IOTLB flush\n");
557 intel_iommu_strict = 1;
558 } else if (!strncmp(str, "sp_off", 6)) {
559 pr_info("Disable supported super page\n");
560 intel_iommu_superpage = 0;
561 } else if (!strncmp(str, "ecs_off", 7)) {
563 "Intel-IOMMU: disable extended context table support\n");
567 str += strcspn(str, ",");
573 __setup("intel_iommu=", intel_iommu_setup);
575 static struct kmem_cache *iommu_domain_cache;
576 static struct kmem_cache *iommu_devinfo_cache;
578 static struct dmar_domain* get_iommu_domain(struct intel_iommu *iommu, u16 did)
580 struct dmar_domain **domains;
583 domains = iommu->domains[idx];
587 return domains[did & 0xff];
590 static void set_iommu_domain(struct intel_iommu *iommu, u16 did,
591 struct dmar_domain *domain)
593 struct dmar_domain **domains;
596 if (!iommu->domains[idx]) {
597 size_t size = 256 * sizeof(struct dmar_domain *);
598 iommu->domains[idx] = kzalloc(size, GFP_ATOMIC);
601 domains = iommu->domains[idx];
602 if (WARN_ON(!domains))
605 domains[did & 0xff] = domain;
608 static inline void *alloc_pgtable_page(int node)
613 page = alloc_pages_node(node, GFP_ATOMIC | __GFP_ZERO, 0);
615 vaddr = page_address(page);
619 static inline void free_pgtable_page(void *vaddr)
621 free_page((unsigned long)vaddr);
624 static inline void *alloc_domain_mem(void)
626 return kmem_cache_alloc(iommu_domain_cache, GFP_ATOMIC);
629 static void free_domain_mem(void *vaddr)
631 kmem_cache_free(iommu_domain_cache, vaddr);
634 static inline void * alloc_devinfo_mem(void)
636 return kmem_cache_alloc(iommu_devinfo_cache, GFP_ATOMIC);
639 static inline void free_devinfo_mem(void *vaddr)
641 kmem_cache_free(iommu_devinfo_cache, vaddr);
644 static inline int domain_type_is_vm(struct dmar_domain *domain)
646 return domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE;
649 static inline int domain_type_is_si(struct dmar_domain *domain)
651 return domain->flags & DOMAIN_FLAG_STATIC_IDENTITY;
654 static inline int domain_type_is_vm_or_si(struct dmar_domain *domain)
656 return domain->flags & (DOMAIN_FLAG_VIRTUAL_MACHINE |
657 DOMAIN_FLAG_STATIC_IDENTITY);
660 static inline int domain_pfn_supported(struct dmar_domain *domain,
663 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
665 return !(addr_width < BITS_PER_LONG && pfn >> addr_width);
668 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
673 sagaw = cap_sagaw(iommu->cap);
674 for (agaw = width_to_agaw(max_gaw);
676 if (test_bit(agaw, &sagaw))
684 * Calculate max SAGAW for each iommu.
686 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
688 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
692 * calculate agaw for each iommu.
693 * "SAGAW" may be different across iommus, use a default agaw, and
694 * get a supported less agaw for iommus that don't support the default agaw.
696 int iommu_calculate_agaw(struct intel_iommu *iommu)
698 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
701 /* This functionin only returns single iommu in a domain */
702 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
706 /* si_domain and vm domain should not get here. */
707 BUG_ON(domain_type_is_vm_or_si(domain));
708 for_each_domain_iommu(iommu_id, domain)
711 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
714 return g_iommus[iommu_id];
717 static void domain_update_iommu_coherency(struct dmar_domain *domain)
719 struct dmar_drhd_unit *drhd;
720 struct intel_iommu *iommu;
724 domain->iommu_coherency = 1;
726 for_each_domain_iommu(i, domain) {
728 if (!ecap_coherent(g_iommus[i]->ecap)) {
729 domain->iommu_coherency = 0;
736 /* No hardware attached; use lowest common denominator */
738 for_each_active_iommu(iommu, drhd) {
739 if (!ecap_coherent(iommu->ecap)) {
740 domain->iommu_coherency = 0;
747 static int domain_update_iommu_snooping(struct intel_iommu *skip)
749 struct dmar_drhd_unit *drhd;
750 struct intel_iommu *iommu;
754 for_each_active_iommu(iommu, drhd) {
756 if (!ecap_sc_support(iommu->ecap)) {
767 static int domain_update_iommu_superpage(struct intel_iommu *skip)
769 struct dmar_drhd_unit *drhd;
770 struct intel_iommu *iommu;
773 if (!intel_iommu_superpage) {
777 /* set iommu_superpage to the smallest common denominator */
779 for_each_active_iommu(iommu, drhd) {
781 mask &= cap_super_page_val(iommu->cap);
791 /* Some capabilities may be different across iommus */
792 static void domain_update_iommu_cap(struct dmar_domain *domain)
794 domain_update_iommu_coherency(domain);
795 domain->iommu_snooping = domain_update_iommu_snooping(NULL);
796 domain->iommu_superpage = domain_update_iommu_superpage(NULL);
799 static inline struct context_entry *iommu_context_addr(struct intel_iommu *iommu,
800 u8 bus, u8 devfn, int alloc)
802 struct root_entry *root = &iommu->root_entry[bus];
803 struct context_entry *context;
806 if (ecs_enabled(iommu)) {
815 context = phys_to_virt(*entry & VTD_PAGE_MASK);
817 unsigned long phy_addr;
821 context = alloc_pgtable_page(iommu->node);
825 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
826 phy_addr = virt_to_phys((void *)context);
827 *entry = phy_addr | 1;
828 __iommu_flush_cache(iommu, entry, sizeof(*entry));
830 return &context[devfn];
833 static int iommu_dummy(struct device *dev)
835 return dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
838 static struct intel_iommu *device_to_iommu(struct device *dev, u8 *bus, u8 *devfn)
840 struct dmar_drhd_unit *drhd = NULL;
841 struct intel_iommu *iommu;
843 struct pci_dev *ptmp, *pdev = NULL;
847 if (iommu_dummy(dev))
850 if (dev_is_pci(dev)) {
851 pdev = to_pci_dev(dev);
852 segment = pci_domain_nr(pdev->bus);
853 } else if (has_acpi_companion(dev))
854 dev = &ACPI_COMPANION(dev)->dev;
857 for_each_active_iommu(iommu, drhd) {
858 if (pdev && segment != drhd->segment)
861 for_each_active_dev_scope(drhd->devices,
862 drhd->devices_cnt, i, tmp) {
864 *bus = drhd->devices[i].bus;
865 *devfn = drhd->devices[i].devfn;
869 if (!pdev || !dev_is_pci(tmp))
872 ptmp = to_pci_dev(tmp);
873 if (ptmp->subordinate &&
874 ptmp->subordinate->number <= pdev->bus->number &&
875 ptmp->subordinate->busn_res.end >= pdev->bus->number)
879 if (pdev && drhd->include_all) {
881 *bus = pdev->bus->number;
882 *devfn = pdev->devfn;
893 static void domain_flush_cache(struct dmar_domain *domain,
894 void *addr, int size)
896 if (!domain->iommu_coherency)
897 clflush_cache_range(addr, size);
900 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
902 struct context_entry *context;
906 spin_lock_irqsave(&iommu->lock, flags);
907 context = iommu_context_addr(iommu, bus, devfn, 0);
909 ret = context_present(context);
910 spin_unlock_irqrestore(&iommu->lock, flags);
914 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
916 struct context_entry *context;
919 spin_lock_irqsave(&iommu->lock, flags);
920 context = iommu_context_addr(iommu, bus, devfn, 0);
922 context_clear_entry(context);
923 __iommu_flush_cache(iommu, context, sizeof(*context));
925 spin_unlock_irqrestore(&iommu->lock, flags);
928 static void free_context_table(struct intel_iommu *iommu)
932 struct context_entry *context;
934 spin_lock_irqsave(&iommu->lock, flags);
935 if (!iommu->root_entry) {
938 for (i = 0; i < ROOT_ENTRY_NR; i++) {
939 context = iommu_context_addr(iommu, i, 0, 0);
941 free_pgtable_page(context);
943 if (!ecs_enabled(iommu))
946 context = iommu_context_addr(iommu, i, 0x80, 0);
948 free_pgtable_page(context);
951 free_pgtable_page(iommu->root_entry);
952 iommu->root_entry = NULL;
954 spin_unlock_irqrestore(&iommu->lock, flags);
957 static struct dma_pte *pfn_to_dma_pte(struct dmar_domain *domain,
958 unsigned long pfn, int *target_level)
960 struct dma_pte *parent, *pte = NULL;
961 int level = agaw_to_level(domain->agaw);
964 BUG_ON(!domain->pgd);
966 if (!domain_pfn_supported(domain, pfn))
967 /* Address beyond IOMMU's addressing capabilities. */
970 parent = domain->pgd;
975 offset = pfn_level_offset(pfn, level);
976 pte = &parent[offset];
977 if (!*target_level && (dma_pte_superpage(pte) || !dma_pte_present(pte)))
979 if (level == *target_level)
982 if (!dma_pte_present(pte)) {
985 tmp_page = alloc_pgtable_page(domain->nid);
990 domain_flush_cache(domain, tmp_page, VTD_PAGE_SIZE);
991 pteval = ((uint64_t)virt_to_dma_pfn(tmp_page) << VTD_PAGE_SHIFT) | DMA_PTE_READ | DMA_PTE_WRITE;
992 if (cmpxchg64(&pte->val, 0ULL, pteval))
993 /* Someone else set it while we were thinking; use theirs. */
994 free_pgtable_page(tmp_page);
996 domain_flush_cache(domain, pte, sizeof(*pte));
1001 parent = phys_to_virt(dma_pte_addr(pte));
1006 *target_level = level;
1012 /* return address's pte at specific level */
1013 static struct dma_pte *dma_pfn_level_pte(struct dmar_domain *domain,
1015 int level, int *large_page)
1017 struct dma_pte *parent, *pte = NULL;
1018 int total = agaw_to_level(domain->agaw);
1021 parent = domain->pgd;
1022 while (level <= total) {
1023 offset = pfn_level_offset(pfn, total);
1024 pte = &parent[offset];
1028 if (!dma_pte_present(pte)) {
1029 *large_page = total;
1033 if (dma_pte_superpage(pte)) {
1034 *large_page = total;
1038 parent = phys_to_virt(dma_pte_addr(pte));
1044 /* clear last level pte, a tlb flush should be followed */
1045 static void dma_pte_clear_range(struct dmar_domain *domain,
1046 unsigned long start_pfn,
1047 unsigned long last_pfn)
1049 unsigned int large_page = 1;
1050 struct dma_pte *first_pte, *pte;
1052 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1053 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1054 BUG_ON(start_pfn > last_pfn);
1056 /* we don't need lock here; nobody else touches the iova range */
1059 first_pte = pte = dma_pfn_level_pte(domain, start_pfn, 1, &large_page);
1061 start_pfn = align_to_level(start_pfn + 1, large_page + 1);
1066 start_pfn += lvl_to_nr_pages(large_page);
1068 } while (start_pfn <= last_pfn && !first_pte_in_page(pte));
1070 domain_flush_cache(domain, first_pte,
1071 (void *)pte - (void *)first_pte);
1073 } while (start_pfn && start_pfn <= last_pfn);
1076 static void dma_pte_free_level(struct dmar_domain *domain, int level,
1077 struct dma_pte *pte, unsigned long pfn,
1078 unsigned long start_pfn, unsigned long last_pfn)
1080 pfn = max(start_pfn, pfn);
1081 pte = &pte[pfn_level_offset(pfn, level)];
1084 unsigned long level_pfn;
1085 struct dma_pte *level_pte;
1087 if (!dma_pte_present(pte) || dma_pte_superpage(pte))
1090 level_pfn = pfn & level_mask(level - 1);
1091 level_pte = phys_to_virt(dma_pte_addr(pte));
1094 dma_pte_free_level(domain, level - 1, level_pte,
1095 level_pfn, start_pfn, last_pfn);
1097 /* If range covers entire pagetable, free it */
1098 if (!(start_pfn > level_pfn ||
1099 last_pfn < level_pfn + level_size(level) - 1)) {
1101 domain_flush_cache(domain, pte, sizeof(*pte));
1102 free_pgtable_page(level_pte);
1105 pfn += level_size(level);
1106 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
1109 /* free page table pages. last level pte should already be cleared */
1110 static void dma_pte_free_pagetable(struct dmar_domain *domain,
1111 unsigned long start_pfn,
1112 unsigned long last_pfn)
1114 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1115 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1116 BUG_ON(start_pfn > last_pfn);
1118 dma_pte_clear_range(domain, start_pfn, last_pfn);
1120 /* We don't need lock here; nobody else touches the iova range */
1121 dma_pte_free_level(domain, agaw_to_level(domain->agaw),
1122 domain->pgd, 0, start_pfn, last_pfn);
1125 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
1126 free_pgtable_page(domain->pgd);
1131 /* When a page at a given level is being unlinked from its parent, we don't
1132 need to *modify* it at all. All we need to do is make a list of all the
1133 pages which can be freed just as soon as we've flushed the IOTLB and we
1134 know the hardware page-walk will no longer touch them.
1135 The 'pte' argument is the *parent* PTE, pointing to the page that is to
1137 static struct page *dma_pte_list_pagetables(struct dmar_domain *domain,
1138 int level, struct dma_pte *pte,
1139 struct page *freelist)
1143 pg = pfn_to_page(dma_pte_addr(pte) >> PAGE_SHIFT);
1144 pg->freelist = freelist;
1150 pte = page_address(pg);
1152 if (dma_pte_present(pte) && !dma_pte_superpage(pte))
1153 freelist = dma_pte_list_pagetables(domain, level - 1,
1156 } while (!first_pte_in_page(pte));
1161 static struct page *dma_pte_clear_level(struct dmar_domain *domain, int level,
1162 struct dma_pte *pte, unsigned long pfn,
1163 unsigned long start_pfn,
1164 unsigned long last_pfn,
1165 struct page *freelist)
1167 struct dma_pte *first_pte = NULL, *last_pte = NULL;
1169 pfn = max(start_pfn, pfn);
1170 pte = &pte[pfn_level_offset(pfn, level)];
1173 unsigned long level_pfn;
1175 if (!dma_pte_present(pte))
1178 level_pfn = pfn & level_mask(level);
1180 /* If range covers entire pagetable, free it */
1181 if (start_pfn <= level_pfn &&
1182 last_pfn >= level_pfn + level_size(level) - 1) {
1183 /* These suborbinate page tables are going away entirely. Don't
1184 bother to clear them; we're just going to *free* them. */
1185 if (level > 1 && !dma_pte_superpage(pte))
1186 freelist = dma_pte_list_pagetables(domain, level - 1, pte, freelist);
1192 } else if (level > 1) {
1193 /* Recurse down into a level that isn't *entirely* obsolete */
1194 freelist = dma_pte_clear_level(domain, level - 1,
1195 phys_to_virt(dma_pte_addr(pte)),
1196 level_pfn, start_pfn, last_pfn,
1200 pfn += level_size(level);
1201 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
1204 domain_flush_cache(domain, first_pte,
1205 (void *)++last_pte - (void *)first_pte);
1210 /* We can't just free the pages because the IOMMU may still be walking
1211 the page tables, and may have cached the intermediate levels. The
1212 pages can only be freed after the IOTLB flush has been done. */
1213 struct page *domain_unmap(struct dmar_domain *domain,
1214 unsigned long start_pfn,
1215 unsigned long last_pfn)
1217 struct page *freelist = NULL;
1219 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1220 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1221 BUG_ON(start_pfn > last_pfn);
1223 /* we don't need lock here; nobody else touches the iova range */
1224 freelist = dma_pte_clear_level(domain, agaw_to_level(domain->agaw),
1225 domain->pgd, 0, start_pfn, last_pfn, NULL);
1228 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
1229 struct page *pgd_page = virt_to_page(domain->pgd);
1230 pgd_page->freelist = freelist;
1231 freelist = pgd_page;
1239 void dma_free_pagelist(struct page *freelist)
1243 while ((pg = freelist)) {
1244 freelist = pg->freelist;
1245 free_pgtable_page(page_address(pg));
1249 /* iommu handling */
1250 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
1252 struct root_entry *root;
1253 unsigned long flags;
1255 root = (struct root_entry *)alloc_pgtable_page(iommu->node);
1257 pr_err("Allocating root entry for %s failed\n",
1262 __iommu_flush_cache(iommu, root, ROOT_SIZE);
1264 spin_lock_irqsave(&iommu->lock, flags);
1265 iommu->root_entry = root;
1266 spin_unlock_irqrestore(&iommu->lock, flags);
1271 static void iommu_set_root_entry(struct intel_iommu *iommu)
1277 addr = virt_to_phys(iommu->root_entry);
1278 if (ecs_enabled(iommu))
1279 addr |= DMA_RTADDR_RTT;
1281 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1282 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, addr);
1284 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
1286 /* Make sure hardware complete it */
1287 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1288 readl, (sts & DMA_GSTS_RTPS), sts);
1290 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1293 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
1298 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
1301 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1302 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
1304 /* Make sure hardware complete it */
1305 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1306 readl, (!(val & DMA_GSTS_WBFS)), val);
1308 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1311 /* return value determine if we need a write buffer flush */
1312 static void __iommu_flush_context(struct intel_iommu *iommu,
1313 u16 did, u16 source_id, u8 function_mask,
1320 case DMA_CCMD_GLOBAL_INVL:
1321 val = DMA_CCMD_GLOBAL_INVL;
1323 case DMA_CCMD_DOMAIN_INVL:
1324 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
1326 case DMA_CCMD_DEVICE_INVL:
1327 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
1328 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
1333 val |= DMA_CCMD_ICC;
1335 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1336 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
1338 /* Make sure hardware complete it */
1339 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
1340 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
1342 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1345 /* return value determine if we need a write buffer flush */
1346 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
1347 u64 addr, unsigned int size_order, u64 type)
1349 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
1350 u64 val = 0, val_iva = 0;
1354 case DMA_TLB_GLOBAL_FLUSH:
1355 /* global flush doesn't need set IVA_REG */
1356 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
1358 case DMA_TLB_DSI_FLUSH:
1359 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1361 case DMA_TLB_PSI_FLUSH:
1362 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1363 /* IH bit is passed in as part of address */
1364 val_iva = size_order | addr;
1369 /* Note: set drain read/write */
1372 * This is probably to be super secure.. Looks like we can
1373 * ignore it without any impact.
1375 if (cap_read_drain(iommu->cap))
1376 val |= DMA_TLB_READ_DRAIN;
1378 if (cap_write_drain(iommu->cap))
1379 val |= DMA_TLB_WRITE_DRAIN;
1381 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1382 /* Note: Only uses first TLB reg currently */
1384 dmar_writeq(iommu->reg + tlb_offset, val_iva);
1385 dmar_writeq(iommu->reg + tlb_offset + 8, val);
1387 /* Make sure hardware complete it */
1388 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
1389 dmar_readq, (!(val & DMA_TLB_IVT)), val);
1391 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1393 /* check IOTLB invalidation granularity */
1394 if (DMA_TLB_IAIG(val) == 0)
1395 pr_err("Flush IOTLB failed\n");
1396 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
1397 pr_debug("TLB flush request %Lx, actual %Lx\n",
1398 (unsigned long long)DMA_TLB_IIRG(type),
1399 (unsigned long long)DMA_TLB_IAIG(val));
1402 static struct device_domain_info *
1403 iommu_support_dev_iotlb (struct dmar_domain *domain, struct intel_iommu *iommu,
1407 unsigned long flags;
1408 struct device_domain_info *info;
1409 struct pci_dev *pdev;
1411 if (!ecap_dev_iotlb_support(iommu->ecap))
1417 spin_lock_irqsave(&device_domain_lock, flags);
1418 list_for_each_entry(info, &domain->devices, link)
1419 if (info->iommu == iommu && info->bus == bus &&
1420 info->devfn == devfn) {
1424 spin_unlock_irqrestore(&device_domain_lock, flags);
1426 if (!found || !info->dev || !dev_is_pci(info->dev))
1429 pdev = to_pci_dev(info->dev);
1431 if (!pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS))
1434 if (!dmar_find_matched_atsr_unit(pdev))
1440 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
1442 if (!info || !dev_is_pci(info->dev))
1445 pci_enable_ats(to_pci_dev(info->dev), VTD_PAGE_SHIFT);
1448 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1450 if (!info->dev || !dev_is_pci(info->dev) ||
1451 !pci_ats_enabled(to_pci_dev(info->dev)))
1454 pci_disable_ats(to_pci_dev(info->dev));
1457 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1458 u64 addr, unsigned mask)
1461 unsigned long flags;
1462 struct device_domain_info *info;
1464 spin_lock_irqsave(&device_domain_lock, flags);
1465 list_for_each_entry(info, &domain->devices, link) {
1466 struct pci_dev *pdev;
1467 if (!info->dev || !dev_is_pci(info->dev))
1470 pdev = to_pci_dev(info->dev);
1471 if (!pci_ats_enabled(pdev))
1474 sid = info->bus << 8 | info->devfn;
1475 qdep = pci_ats_queue_depth(pdev);
1476 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1478 spin_unlock_irqrestore(&device_domain_lock, flags);
1481 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu,
1482 struct dmar_domain *domain,
1483 unsigned long pfn, unsigned int pages,
1486 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1487 uint64_t addr = (uint64_t)pfn << VTD_PAGE_SHIFT;
1488 u16 did = domain->iommu_did[iommu->seq_id];
1495 * Fallback to domain selective flush if no PSI support or the size is
1497 * PSI requires page size to be 2 ^ x, and the base address is naturally
1498 * aligned to the size
1500 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1501 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1504 iommu->flush.flush_iotlb(iommu, did, addr | ih, mask,
1508 * In caching mode, changes of pages from non-present to present require
1509 * flush. However, device IOTLB doesn't need to be flushed in this case.
1511 if (!cap_caching_mode(iommu->cap) || !map)
1512 iommu_flush_dev_iotlb(get_iommu_domain(iommu, did),
1516 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1519 unsigned long flags;
1521 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1522 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1523 pmen &= ~DMA_PMEN_EPM;
1524 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1526 /* wait for the protected region status bit to clear */
1527 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1528 readl, !(pmen & DMA_PMEN_PRS), pmen);
1530 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1533 static void iommu_enable_translation(struct intel_iommu *iommu)
1536 unsigned long flags;
1538 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1539 iommu->gcmd |= DMA_GCMD_TE;
1540 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1542 /* Make sure hardware complete it */
1543 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1544 readl, (sts & DMA_GSTS_TES), sts);
1546 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1549 static void iommu_disable_translation(struct intel_iommu *iommu)
1554 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1555 iommu->gcmd &= ~DMA_GCMD_TE;
1556 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1558 /* Make sure hardware complete it */
1559 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1560 readl, (!(sts & DMA_GSTS_TES)), sts);
1562 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1566 static int iommu_init_domains(struct intel_iommu *iommu)
1568 u32 ndomains, nlongs;
1571 ndomains = cap_ndoms(iommu->cap);
1572 pr_debug("%s: Number of Domains supported <%d>\n",
1573 iommu->name, ndomains);
1574 nlongs = BITS_TO_LONGS(ndomains);
1576 spin_lock_init(&iommu->lock);
1578 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1579 if (!iommu->domain_ids) {
1580 pr_err("%s: Allocating domain id array failed\n",
1585 size = ((ndomains >> 8) + 1) * sizeof(struct dmar_domain **);
1586 iommu->domains = kzalloc(size, GFP_KERNEL);
1588 if (iommu->domains) {
1589 size = 256 * sizeof(struct dmar_domain *);
1590 iommu->domains[0] = kzalloc(size, GFP_KERNEL);
1593 if (!iommu->domains || !iommu->domains[0]) {
1594 pr_err("%s: Allocating domain array failed\n",
1596 kfree(iommu->domain_ids);
1597 kfree(iommu->domains);
1598 iommu->domain_ids = NULL;
1599 iommu->domains = NULL;
1606 * If Caching mode is set, then invalid translations are tagged
1607 * with domain-id 0, hence we need to pre-allocate it. We also
1608 * use domain-id 0 as a marker for non-allocated domain-id, so
1609 * make sure it is not used for a real domain.
1611 set_bit(0, iommu->domain_ids);
1616 static void disable_dmar_iommu(struct intel_iommu *iommu)
1618 struct device_domain_info *info, *tmp;
1620 if (!iommu->domains || !iommu->domain_ids)
1623 list_for_each_entry_safe(info, tmp, &device_domain_list, global) {
1624 struct dmar_domain *domain;
1626 if (info->iommu != iommu)
1629 if (!info->dev || !info->domain)
1632 domain = info->domain;
1634 dmar_remove_one_dev_info(domain, info->dev);
1636 if (!domain_type_is_vm_or_si(domain))
1637 domain_exit(domain);
1640 if (iommu->gcmd & DMA_GCMD_TE)
1641 iommu_disable_translation(iommu);
1644 static void free_dmar_iommu(struct intel_iommu *iommu)
1646 if ((iommu->domains) && (iommu->domain_ids)) {
1647 int elems = (cap_ndoms(iommu->cap) >> 8) + 1;
1650 for (i = 0; i < elems; i++)
1651 kfree(iommu->domains[i]);
1652 kfree(iommu->domains);
1653 kfree(iommu->domain_ids);
1654 iommu->domains = NULL;
1655 iommu->domain_ids = NULL;
1658 g_iommus[iommu->seq_id] = NULL;
1660 /* free context mapping */
1661 free_context_table(iommu);
1664 static struct dmar_domain *alloc_domain(int flags)
1666 struct dmar_domain *domain;
1668 domain = alloc_domain_mem();
1672 memset(domain, 0, sizeof(*domain));
1674 domain->flags = flags;
1675 spin_lock_init(&domain->iommu_lock);
1676 INIT_LIST_HEAD(&domain->devices);
1681 static int __iommu_attach_domain(struct dmar_domain *domain,
1682 struct intel_iommu *iommu)
1685 unsigned long ndomains;
1687 num = domain->iommu_did[iommu->seq_id];
1691 ndomains = cap_ndoms(iommu->cap);
1692 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1694 if (num < ndomains) {
1695 set_bit(num, iommu->domain_ids);
1696 set_iommu_domain(iommu, num, domain);
1697 domain->iommu_did[iommu->seq_id] = num;
1703 pr_err("%s: No free domain ids\n", iommu->name);
1708 static int iommu_attach_domain(struct dmar_domain *domain,
1709 struct intel_iommu *iommu)
1712 unsigned long flags;
1714 spin_lock_irqsave(&iommu->lock, flags);
1715 num = __iommu_attach_domain(domain, iommu);
1716 spin_unlock_irqrestore(&iommu->lock, flags);
1721 static void iommu_detach_domain(struct dmar_domain *domain,
1722 struct intel_iommu *iommu)
1724 unsigned long flags;
1727 spin_lock_irqsave(&iommu->lock, flags);
1729 num = domain->iommu_did[iommu->seq_id];
1734 clear_bit(num, iommu->domain_ids);
1735 set_iommu_domain(iommu, num, NULL);
1737 spin_unlock_irqrestore(&iommu->lock, flags);
1740 static void domain_attach_iommu(struct dmar_domain *domain,
1741 struct intel_iommu *iommu)
1743 unsigned long flags;
1745 spin_lock_irqsave(&domain->iommu_lock, flags);
1746 domain->iommu_refcnt[iommu->seq_id] += 1;
1747 domain->iommu_count += 1;
1748 if (domain->iommu_refcnt[iommu->seq_id] == 1) {
1749 domain->nid = iommu->node;
1750 domain_update_iommu_cap(domain);
1752 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1755 static int domain_detach_iommu(struct dmar_domain *domain,
1756 struct intel_iommu *iommu)
1758 unsigned long flags;
1759 int count = INT_MAX;
1761 spin_lock_irqsave(&domain->iommu_lock, flags);
1762 domain->iommu_refcnt[iommu->seq_id] -= 1;
1763 count = --domain->iommu_count;
1764 if (domain->iommu_refcnt[iommu->seq_id] == 0) {
1765 domain_update_iommu_cap(domain);
1766 domain->iommu_did[iommu->seq_id] = 0;
1768 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1773 static struct iova_domain reserved_iova_list;
1774 static struct lock_class_key reserved_rbtree_key;
1776 static int dmar_init_reserved_ranges(void)
1778 struct pci_dev *pdev = NULL;
1782 init_iova_domain(&reserved_iova_list, VTD_PAGE_SIZE, IOVA_START_PFN,
1785 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1786 &reserved_rbtree_key);
1788 /* IOAPIC ranges shouldn't be accessed by DMA */
1789 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1790 IOVA_PFN(IOAPIC_RANGE_END));
1792 pr_err("Reserve IOAPIC range failed\n");
1796 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1797 for_each_pci_dev(pdev) {
1800 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1801 r = &pdev->resource[i];
1802 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1804 iova = reserve_iova(&reserved_iova_list,
1808 pr_err("Reserve iova failed\n");
1816 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1818 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1821 static inline int guestwidth_to_adjustwidth(int gaw)
1824 int r = (gaw - 12) % 9;
1835 static int domain_init(struct dmar_domain *domain, struct intel_iommu *iommu,
1838 int adjust_width, agaw;
1839 unsigned long sagaw;
1841 init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
1843 domain_reserve_special_ranges(domain);
1845 /* calculate AGAW */
1846 if (guest_width > cap_mgaw(iommu->cap))
1847 guest_width = cap_mgaw(iommu->cap);
1848 domain->gaw = guest_width;
1849 adjust_width = guestwidth_to_adjustwidth(guest_width);
1850 agaw = width_to_agaw(adjust_width);
1851 sagaw = cap_sagaw(iommu->cap);
1852 if (!test_bit(agaw, &sagaw)) {
1853 /* hardware doesn't support it, choose a bigger one */
1854 pr_debug("Hardware doesn't support agaw %d\n", agaw);
1855 agaw = find_next_bit(&sagaw, 5, agaw);
1859 domain->agaw = agaw;
1861 if (ecap_coherent(iommu->ecap))
1862 domain->iommu_coherency = 1;
1864 domain->iommu_coherency = 0;
1866 if (ecap_sc_support(iommu->ecap))
1867 domain->iommu_snooping = 1;
1869 domain->iommu_snooping = 0;
1871 if (intel_iommu_superpage)
1872 domain->iommu_superpage = fls(cap_super_page_val(iommu->cap));
1874 domain->iommu_superpage = 0;
1876 domain->nid = iommu->node;
1878 /* always allocate the top pgd */
1879 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
1882 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1886 static void domain_exit(struct dmar_domain *domain)
1888 struct page *freelist = NULL;
1891 /* Domain 0 is reserved, so dont process it */
1895 /* Flush any lazy unmaps that may reference this domain */
1896 if (!intel_iommu_strict)
1897 flush_unmaps_timeout(0);
1899 /* remove associated devices */
1900 domain_remove_dev_info(domain);
1903 put_iova_domain(&domain->iovad);
1905 freelist = domain_unmap(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1907 /* clear attached or cached domains */
1909 for_each_domain_iommu(i, domain)
1910 iommu_detach_domain(domain, g_iommus[i]);
1913 dma_free_pagelist(freelist);
1915 free_domain_mem(domain);
1918 static int domain_context_mapping_one(struct dmar_domain *domain,
1919 struct intel_iommu *iommu,
1922 int translation = CONTEXT_TT_MULTI_LEVEL;
1923 struct device_domain_info *info = NULL;
1924 struct context_entry *context;
1925 unsigned long flags;
1926 struct dma_pte *pgd;
1930 if (hw_pass_through && domain_type_is_si(domain))
1931 translation = CONTEXT_TT_PASS_THROUGH;
1933 pr_debug("Set context mapping for %02x:%02x.%d\n",
1934 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1936 BUG_ON(!domain->pgd);
1938 spin_lock_irqsave(&iommu->lock, flags);
1939 context = iommu_context_addr(iommu, bus, devfn, 1);
1940 spin_unlock_irqrestore(&iommu->lock, flags);
1943 spin_lock_irqsave(&iommu->lock, flags);
1944 if (context_present(context)) {
1945 spin_unlock_irqrestore(&iommu->lock, flags);
1951 id = __iommu_attach_domain(domain, iommu);
1953 spin_unlock_irqrestore(&iommu->lock, flags);
1954 pr_err("%s: No free domain ids\n", iommu->name);
1958 context_clear_entry(context);
1959 context_set_domain_id(context, id);
1962 * Skip top levels of page tables for iommu which has less agaw
1963 * than default. Unnecessary for PT mode.
1965 if (translation != CONTEXT_TT_PASS_THROUGH) {
1966 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1967 pgd = phys_to_virt(dma_pte_addr(pgd));
1968 if (!dma_pte_present(pgd)) {
1969 spin_unlock_irqrestore(&iommu->lock, flags);
1974 info = iommu_support_dev_iotlb(domain, iommu, bus, devfn);
1975 translation = info ? CONTEXT_TT_DEV_IOTLB :
1976 CONTEXT_TT_MULTI_LEVEL;
1978 context_set_address_root(context, virt_to_phys(pgd));
1979 context_set_address_width(context, iommu->agaw);
1982 * In pass through mode, AW must be programmed to
1983 * indicate the largest AGAW value supported by
1984 * hardware. And ASR is ignored by hardware.
1986 context_set_address_width(context, iommu->msagaw);
1989 context_set_translation_type(context, translation);
1990 context_set_fault_enable(context);
1991 context_set_present(context);
1992 domain_flush_cache(domain, context, sizeof(*context));
1995 * It's a non-present to present mapping. If hardware doesn't cache
1996 * non-present entry we only need to flush the write-buffer. If the
1997 * _does_ cache non-present entries, then it does so in the special
1998 * domain #0, which we have to flush:
2000 if (cap_caching_mode(iommu->cap)) {
2001 iommu->flush.flush_context(iommu, 0,
2002 (((u16)bus) << 8) | devfn,
2003 DMA_CCMD_MASK_NOBIT,
2004 DMA_CCMD_DEVICE_INVL);
2005 iommu->flush.flush_iotlb(iommu, id, 0, 0, DMA_TLB_DSI_FLUSH);
2007 iommu_flush_write_buffer(iommu);
2009 iommu_enable_dev_iotlb(info);
2010 spin_unlock_irqrestore(&iommu->lock, flags);
2012 domain_attach_iommu(domain, iommu);
2017 struct domain_context_mapping_data {
2018 struct dmar_domain *domain;
2019 struct intel_iommu *iommu;
2022 static int domain_context_mapping_cb(struct pci_dev *pdev,
2023 u16 alias, void *opaque)
2025 struct domain_context_mapping_data *data = opaque;
2027 return domain_context_mapping_one(data->domain, data->iommu,
2028 PCI_BUS_NUM(alias), alias & 0xff);
2032 domain_context_mapping(struct dmar_domain *domain, struct device *dev)
2034 struct intel_iommu *iommu;
2036 struct domain_context_mapping_data data;
2038 iommu = device_to_iommu(dev, &bus, &devfn);
2042 if (!dev_is_pci(dev))
2043 return domain_context_mapping_one(domain, iommu, bus, devfn);
2045 data.domain = domain;
2048 return pci_for_each_dma_alias(to_pci_dev(dev),
2049 &domain_context_mapping_cb, &data);
2052 static int domain_context_mapped_cb(struct pci_dev *pdev,
2053 u16 alias, void *opaque)
2055 struct intel_iommu *iommu = opaque;
2057 return !device_context_mapped(iommu, PCI_BUS_NUM(alias), alias & 0xff);
2060 static int domain_context_mapped(struct device *dev)
2062 struct intel_iommu *iommu;
2065 iommu = device_to_iommu(dev, &bus, &devfn);
2069 if (!dev_is_pci(dev))
2070 return device_context_mapped(iommu, bus, devfn);
2072 return !pci_for_each_dma_alias(to_pci_dev(dev),
2073 domain_context_mapped_cb, iommu);
2076 /* Returns a number of VTD pages, but aligned to MM page size */
2077 static inline unsigned long aligned_nrpages(unsigned long host_addr,
2080 host_addr &= ~PAGE_MASK;
2081 return PAGE_ALIGN(host_addr + size) >> VTD_PAGE_SHIFT;
2084 /* Return largest possible superpage level for a given mapping */
2085 static inline int hardware_largepage_caps(struct dmar_domain *domain,
2086 unsigned long iov_pfn,
2087 unsigned long phy_pfn,
2088 unsigned long pages)
2090 int support, level = 1;
2091 unsigned long pfnmerge;
2093 support = domain->iommu_superpage;
2095 /* To use a large page, the virtual *and* physical addresses
2096 must be aligned to 2MiB/1GiB/etc. Lower bits set in either
2097 of them will mean we have to use smaller pages. So just
2098 merge them and check both at once. */
2099 pfnmerge = iov_pfn | phy_pfn;
2101 while (support && !(pfnmerge & ~VTD_STRIDE_MASK)) {
2102 pages >>= VTD_STRIDE_SHIFT;
2105 pfnmerge >>= VTD_STRIDE_SHIFT;
2112 static int __domain_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2113 struct scatterlist *sg, unsigned long phys_pfn,
2114 unsigned long nr_pages, int prot)
2116 struct dma_pte *first_pte = NULL, *pte = NULL;
2117 phys_addr_t uninitialized_var(pteval);
2118 unsigned long sg_res = 0;
2119 unsigned int largepage_lvl = 0;
2120 unsigned long lvl_pages = 0;
2122 BUG_ON(!domain_pfn_supported(domain, iov_pfn + nr_pages - 1));
2124 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
2127 prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
2131 pteval = ((phys_addr_t)phys_pfn << VTD_PAGE_SHIFT) | prot;
2134 while (nr_pages > 0) {
2138 sg_res = aligned_nrpages(sg->offset, sg->length);
2139 sg->dma_address = ((dma_addr_t)iov_pfn << VTD_PAGE_SHIFT) + sg->offset;
2140 sg->dma_length = sg->length;
2141 pteval = page_to_phys(sg_page(sg)) | prot;
2142 phys_pfn = pteval >> VTD_PAGE_SHIFT;
2146 largepage_lvl = hardware_largepage_caps(domain, iov_pfn, phys_pfn, sg_res);
2148 first_pte = pte = pfn_to_dma_pte(domain, iov_pfn, &largepage_lvl);
2151 /* It is large page*/
2152 if (largepage_lvl > 1) {
2153 pteval |= DMA_PTE_LARGE_PAGE;
2154 lvl_pages = lvl_to_nr_pages(largepage_lvl);
2156 * Ensure that old small page tables are
2157 * removed to make room for superpage,
2160 dma_pte_free_pagetable(domain, iov_pfn,
2161 iov_pfn + lvl_pages - 1);
2163 pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;
2167 /* We don't need lock here, nobody else
2168 * touches the iova range
2170 tmp = cmpxchg64_local(&pte->val, 0ULL, pteval);
2172 static int dumps = 5;
2173 pr_crit("ERROR: DMA PTE for vPFN 0x%lx already set (to %llx not %llx)\n",
2174 iov_pfn, tmp, (unsigned long long)pteval);
2177 debug_dma_dump_mappings(NULL);
2182 lvl_pages = lvl_to_nr_pages(largepage_lvl);
2184 BUG_ON(nr_pages < lvl_pages);
2185 BUG_ON(sg_res < lvl_pages);
2187 nr_pages -= lvl_pages;
2188 iov_pfn += lvl_pages;
2189 phys_pfn += lvl_pages;
2190 pteval += lvl_pages * VTD_PAGE_SIZE;
2191 sg_res -= lvl_pages;
2193 /* If the next PTE would be the first in a new page, then we
2194 need to flush the cache on the entries we've just written.
2195 And then we'll need to recalculate 'pte', so clear it and
2196 let it get set again in the if (!pte) block above.
2198 If we're done (!nr_pages) we need to flush the cache too.
2200 Also if we've been setting superpages, we may need to
2201 recalculate 'pte' and switch back to smaller pages for the
2202 end of the mapping, if the trailing size is not enough to
2203 use another superpage (i.e. sg_res < lvl_pages). */
2205 if (!nr_pages || first_pte_in_page(pte) ||
2206 (largepage_lvl > 1 && sg_res < lvl_pages)) {
2207 domain_flush_cache(domain, first_pte,
2208 (void *)pte - (void *)first_pte);
2212 if (!sg_res && nr_pages)
2218 static inline int domain_sg_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2219 struct scatterlist *sg, unsigned long nr_pages,
2222 return __domain_mapping(domain, iov_pfn, sg, 0, nr_pages, prot);
2225 static inline int domain_pfn_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2226 unsigned long phys_pfn, unsigned long nr_pages,
2229 return __domain_mapping(domain, iov_pfn, NULL, phys_pfn, nr_pages, prot);
2232 static void domain_context_clear_one(struct intel_iommu *iommu, u8 bus, u8 devfn)
2237 clear_context_table(iommu, bus, devfn);
2238 iommu->flush.flush_context(iommu, 0, 0, 0,
2239 DMA_CCMD_GLOBAL_INVL);
2240 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2243 static inline void unlink_domain_info(struct device_domain_info *info)
2245 assert_spin_locked(&device_domain_lock);
2246 list_del(&info->link);
2247 list_del(&info->global);
2249 info->dev->archdata.iommu = NULL;
2252 static void domain_remove_dev_info(struct dmar_domain *domain)
2254 struct device_domain_info *info, *tmp;
2256 list_for_each_entry_safe(info, tmp, &domain->devices, link)
2257 dmar_remove_one_dev_info(domain, info->dev);
2262 * Note: we use struct device->archdata.iommu stores the info
2264 static struct dmar_domain *find_domain(struct device *dev)
2266 struct device_domain_info *info;
2268 /* No lock here, assumes no domain exit in normal case */
2269 info = dev->archdata.iommu;
2271 return info->domain;
2275 static inline struct device_domain_info *
2276 dmar_search_domain_by_dev_info(int segment, int bus, int devfn)
2278 struct device_domain_info *info;
2280 list_for_each_entry(info, &device_domain_list, global)
2281 if (info->iommu->segment == segment && info->bus == bus &&
2282 info->devfn == devfn)
2288 static struct dmar_domain *dmar_insert_one_dev_info(struct intel_iommu *iommu,
2291 struct dmar_domain *domain)
2293 struct dmar_domain *found = NULL;
2294 struct device_domain_info *info;
2295 unsigned long flags;
2297 info = alloc_devinfo_mem();
2302 info->devfn = devfn;
2304 info->domain = domain;
2305 info->iommu = iommu;
2307 spin_lock_irqsave(&device_domain_lock, flags);
2309 found = find_domain(dev);
2311 struct device_domain_info *info2;
2312 info2 = dmar_search_domain_by_dev_info(iommu->segment, bus, devfn);
2314 found = info2->domain;
2317 spin_unlock_irqrestore(&device_domain_lock, flags);
2318 free_devinfo_mem(info);
2319 /* Caller must free the original domain */
2323 list_add(&info->link, &domain->devices);
2324 list_add(&info->global, &device_domain_list);
2326 dev->archdata.iommu = info;
2327 spin_unlock_irqrestore(&device_domain_lock, flags);
2329 if (dev && domain_context_mapping(domain, dev)) {
2330 pr_err("Domain context map for %s failed\n", dev_name(dev));
2331 dmar_remove_one_dev_info(domain, dev);
2338 static int get_last_alias(struct pci_dev *pdev, u16 alias, void *opaque)
2340 *(u16 *)opaque = alias;
2344 /* domain is initialized */
2345 static struct dmar_domain *get_domain_for_dev(struct device *dev, int gaw)
2347 struct device_domain_info *info = NULL;
2348 struct dmar_domain *domain, *tmp;
2349 struct intel_iommu *iommu;
2350 unsigned long flags;
2354 domain = find_domain(dev);
2358 iommu = device_to_iommu(dev, &bus, &devfn);
2362 if (dev_is_pci(dev)) {
2363 struct pci_dev *pdev = to_pci_dev(dev);
2365 pci_for_each_dma_alias(pdev, get_last_alias, &dma_alias);
2367 spin_lock_irqsave(&device_domain_lock, flags);
2368 info = dmar_search_domain_by_dev_info(pci_domain_nr(pdev->bus),
2369 PCI_BUS_NUM(dma_alias),
2372 iommu = info->iommu;
2373 domain = info->domain;
2375 spin_unlock_irqrestore(&device_domain_lock, flags);
2377 /* DMA alias already has a domain, uses it */
2382 /* Allocate and initialize new domain for the device */
2383 domain = alloc_domain(0);
2386 if (iommu_attach_domain(domain, iommu) < 0) {
2387 free_domain_mem(domain);
2390 domain_attach_iommu(domain, iommu);
2391 if (domain_init(domain, iommu, gaw)) {
2392 domain_exit(domain);
2396 /* register PCI DMA alias device */
2397 if (dev_is_pci(dev)) {
2398 tmp = dmar_insert_one_dev_info(iommu, PCI_BUS_NUM(dma_alias),
2399 dma_alias & 0xff, NULL, domain);
2401 if (!tmp || tmp != domain) {
2402 domain_exit(domain);
2411 tmp = dmar_insert_one_dev_info(iommu, bus, devfn, dev, domain);
2413 if (!tmp || tmp != domain) {
2414 domain_exit(domain);
2421 static int iommu_identity_mapping;
2422 #define IDENTMAP_ALL 1
2423 #define IDENTMAP_GFX 2
2424 #define IDENTMAP_AZALIA 4
2426 static int iommu_domain_identity_map(struct dmar_domain *domain,
2427 unsigned long long start,
2428 unsigned long long end)
2430 unsigned long first_vpfn = start >> VTD_PAGE_SHIFT;
2431 unsigned long last_vpfn = end >> VTD_PAGE_SHIFT;
2433 if (!reserve_iova(&domain->iovad, dma_to_mm_pfn(first_vpfn),
2434 dma_to_mm_pfn(last_vpfn))) {
2435 pr_err("Reserving iova failed\n");
2439 pr_debug("Mapping reserved region %llx-%llx\n", start, end);
2441 * RMRR range might have overlap with physical memory range,
2444 dma_pte_clear_range(domain, first_vpfn, last_vpfn);
2446 return domain_pfn_mapping(domain, first_vpfn, first_vpfn,
2447 last_vpfn - first_vpfn + 1,
2448 DMA_PTE_READ|DMA_PTE_WRITE);
2451 static int iommu_prepare_identity_map(struct device *dev,
2452 unsigned long long start,
2453 unsigned long long end)
2455 struct dmar_domain *domain;
2458 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2462 /* For _hardware_ passthrough, don't bother. But for software
2463 passthrough, we do it anyway -- it may indicate a memory
2464 range which is reserved in E820, so which didn't get set
2465 up to start with in si_domain */
2466 if (domain == si_domain && hw_pass_through) {
2467 pr_warn("Ignoring identity map for HW passthrough device %s [0x%Lx - 0x%Lx]\n",
2468 dev_name(dev), start, end);
2472 pr_info("Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
2473 dev_name(dev), start, end);
2476 WARN(1, "Your BIOS is broken; RMRR ends before it starts!\n"
2477 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2478 dmi_get_system_info(DMI_BIOS_VENDOR),
2479 dmi_get_system_info(DMI_BIOS_VERSION),
2480 dmi_get_system_info(DMI_PRODUCT_VERSION));
2485 if (end >> agaw_to_width(domain->agaw)) {
2486 WARN(1, "Your BIOS is broken; RMRR exceeds permitted address width (%d bits)\n"
2487 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2488 agaw_to_width(domain->agaw),
2489 dmi_get_system_info(DMI_BIOS_VENDOR),
2490 dmi_get_system_info(DMI_BIOS_VERSION),
2491 dmi_get_system_info(DMI_PRODUCT_VERSION));
2496 ret = iommu_domain_identity_map(domain, start, end);
2503 domain_exit(domain);
2507 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
2510 if (dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2512 return iommu_prepare_identity_map(dev, rmrr->base_address,
2516 #ifdef CONFIG_INTEL_IOMMU_FLOPPY_WA
2517 static inline void iommu_prepare_isa(void)
2519 struct pci_dev *pdev;
2522 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
2526 pr_info("Prepare 0-16MiB unity mapping for LPC\n");
2527 ret = iommu_prepare_identity_map(&pdev->dev, 0, 16*1024*1024 - 1);
2530 pr_err("Failed to create 0-16MiB identity map - floppy might not work\n");
2535 static inline void iommu_prepare_isa(void)
2539 #endif /* !CONFIG_INTEL_IOMMU_FLPY_WA */
2541 static int md_domain_init(struct dmar_domain *domain, int guest_width);
2543 static int __init si_domain_init(int hw)
2547 si_domain = alloc_domain(DOMAIN_FLAG_STATIC_IDENTITY);
2551 if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2552 domain_exit(si_domain);
2556 pr_debug("Identity mapping domain allocated\n");
2561 for_each_online_node(nid) {
2562 unsigned long start_pfn, end_pfn;
2565 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
2566 ret = iommu_domain_identity_map(si_domain,
2567 PFN_PHYS(start_pfn), PFN_PHYS(end_pfn));
2576 static int identity_mapping(struct device *dev)
2578 struct device_domain_info *info;
2580 if (likely(!iommu_identity_mapping))
2583 info = dev->archdata.iommu;
2584 if (info && info != DUMMY_DEVICE_DOMAIN_INFO)
2585 return (info->domain == si_domain);
2590 static int domain_add_dev_info(struct dmar_domain *domain, struct device *dev)
2592 struct dmar_domain *ndomain;
2593 struct intel_iommu *iommu;
2596 iommu = device_to_iommu(dev, &bus, &devfn);
2600 ndomain = dmar_insert_one_dev_info(iommu, bus, devfn, dev, domain);
2601 if (ndomain != domain)
2607 static bool device_has_rmrr(struct device *dev)
2609 struct dmar_rmrr_unit *rmrr;
2614 for_each_rmrr_units(rmrr) {
2616 * Return TRUE if this RMRR contains the device that
2619 for_each_active_dev_scope(rmrr->devices,
2620 rmrr->devices_cnt, i, tmp)
2631 * There are a couple cases where we need to restrict the functionality of
2632 * devices associated with RMRRs. The first is when evaluating a device for
2633 * identity mapping because problems exist when devices are moved in and out
2634 * of domains and their respective RMRR information is lost. This means that
2635 * a device with associated RMRRs will never be in a "passthrough" domain.
2636 * The second is use of the device through the IOMMU API. This interface
2637 * expects to have full control of the IOVA space for the device. We cannot
2638 * satisfy both the requirement that RMRR access is maintained and have an
2639 * unencumbered IOVA space. We also have no ability to quiesce the device's
2640 * use of the RMRR space or even inform the IOMMU API user of the restriction.
2641 * We therefore prevent devices associated with an RMRR from participating in
2642 * the IOMMU API, which eliminates them from device assignment.
2644 * In both cases we assume that PCI USB devices with RMRRs have them largely
2645 * for historical reasons and that the RMRR space is not actively used post
2646 * boot. This exclusion may change if vendors begin to abuse it.
2648 * The same exception is made for graphics devices, with the requirement that
2649 * any use of the RMRR regions will be torn down before assigning the device
2652 static bool device_is_rmrr_locked(struct device *dev)
2654 if (!device_has_rmrr(dev))
2657 if (dev_is_pci(dev)) {
2658 struct pci_dev *pdev = to_pci_dev(dev);
2660 if (IS_USB_DEVICE(pdev) || IS_GFX_DEVICE(pdev))
2667 static int iommu_should_identity_map(struct device *dev, int startup)
2670 if (dev_is_pci(dev)) {
2671 struct pci_dev *pdev = to_pci_dev(dev);
2673 if (device_is_rmrr_locked(dev))
2676 if ((iommu_identity_mapping & IDENTMAP_AZALIA) && IS_AZALIA(pdev))
2679 if ((iommu_identity_mapping & IDENTMAP_GFX) && IS_GFX_DEVICE(pdev))
2682 if (!(iommu_identity_mapping & IDENTMAP_ALL))
2686 * We want to start off with all devices in the 1:1 domain, and
2687 * take them out later if we find they can't access all of memory.
2689 * However, we can't do this for PCI devices behind bridges,
2690 * because all PCI devices behind the same bridge will end up
2691 * with the same source-id on their transactions.
2693 * Practically speaking, we can't change things around for these
2694 * devices at run-time, because we can't be sure there'll be no
2695 * DMA transactions in flight for any of their siblings.
2697 * So PCI devices (unless they're on the root bus) as well as
2698 * their parent PCI-PCI or PCIe-PCI bridges must be left _out_ of
2699 * the 1:1 domain, just in _case_ one of their siblings turns out
2700 * not to be able to map all of memory.
2702 if (!pci_is_pcie(pdev)) {
2703 if (!pci_is_root_bus(pdev->bus))
2705 if (pdev->class >> 8 == PCI_CLASS_BRIDGE_PCI)
2707 } else if (pci_pcie_type(pdev) == PCI_EXP_TYPE_PCI_BRIDGE)
2710 if (device_has_rmrr(dev))
2715 * At boot time, we don't yet know if devices will be 64-bit capable.
2716 * Assume that they will — if they turn out not to be, then we can
2717 * take them out of the 1:1 domain later.
2721 * If the device's dma_mask is less than the system's memory
2722 * size then this is not a candidate for identity mapping.
2724 u64 dma_mask = *dev->dma_mask;
2726 if (dev->coherent_dma_mask &&
2727 dev->coherent_dma_mask < dma_mask)
2728 dma_mask = dev->coherent_dma_mask;
2730 return dma_mask >= dma_get_required_mask(dev);
2736 static int __init dev_prepare_static_identity_mapping(struct device *dev, int hw)
2740 if (!iommu_should_identity_map(dev, 1))
2743 ret = domain_add_dev_info(si_domain, dev);
2745 pr_info("%s identity mapping for device %s\n",
2746 hw ? "Hardware" : "Software", dev_name(dev));
2747 else if (ret == -ENODEV)
2748 /* device not associated with an iommu */
2755 static int __init iommu_prepare_static_identity_mapping(int hw)
2757 struct pci_dev *pdev = NULL;
2758 struct dmar_drhd_unit *drhd;
2759 struct intel_iommu *iommu;
2764 for_each_pci_dev(pdev) {
2765 ret = dev_prepare_static_identity_mapping(&pdev->dev, hw);
2770 for_each_active_iommu(iommu, drhd)
2771 for_each_active_dev_scope(drhd->devices, drhd->devices_cnt, i, dev) {
2772 struct acpi_device_physical_node *pn;
2773 struct acpi_device *adev;
2775 if (dev->bus != &acpi_bus_type)
2778 adev= to_acpi_device(dev);
2779 mutex_lock(&adev->physical_node_lock);
2780 list_for_each_entry(pn, &adev->physical_node_list, node) {
2781 ret = dev_prepare_static_identity_mapping(pn->dev, hw);
2785 mutex_unlock(&adev->physical_node_lock);
2793 static void intel_iommu_init_qi(struct intel_iommu *iommu)
2796 * Start from the sane iommu hardware state.
2797 * If the queued invalidation is already initialized by us
2798 * (for example, while enabling interrupt-remapping) then
2799 * we got the things already rolling from a sane state.
2803 * Clear any previous faults.
2805 dmar_fault(-1, iommu);
2807 * Disable queued invalidation if supported and already enabled
2808 * before OS handover.
2810 dmar_disable_qi(iommu);
2813 if (dmar_enable_qi(iommu)) {
2815 * Queued Invalidate not enabled, use Register Based Invalidate
2817 iommu->flush.flush_context = __iommu_flush_context;
2818 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2819 pr_info("%s: Using Register based invalidation\n",
2822 iommu->flush.flush_context = qi_flush_context;
2823 iommu->flush.flush_iotlb = qi_flush_iotlb;
2824 pr_info("%s: Using Queued invalidation\n", iommu->name);
2828 static int copy_context_table(struct intel_iommu *iommu,
2829 struct root_entry *old_re,
2830 struct context_entry **tbl,
2833 struct context_entry *old_ce = NULL, *new_ce = NULL, ce;
2834 int tbl_idx, pos = 0, idx, devfn, ret = 0, did;
2835 phys_addr_t old_ce_phys;
2837 tbl_idx = ext ? bus * 2 : bus;
2839 for (devfn = 0; devfn < 256; devfn++) {
2840 /* First calculate the correct index */
2841 idx = (ext ? devfn * 2 : devfn) % 256;
2844 /* First save what we may have and clean up */
2846 tbl[tbl_idx] = new_ce;
2847 __iommu_flush_cache(iommu, new_ce,
2857 old_ce_phys = root_entry_lctp(old_re);
2859 old_ce_phys = root_entry_uctp(old_re);
2862 if (ext && devfn == 0) {
2863 /* No LCTP, try UCTP */
2872 old_ce = ioremap_cache(old_ce_phys, PAGE_SIZE);
2876 new_ce = alloc_pgtable_page(iommu->node);
2883 /* Now copy the context entry */
2886 if (!__context_present(&ce))
2889 did = context_domain_id(&ce);
2890 if (did >= 0 && did < cap_ndoms(iommu->cap))
2891 set_bit(did, iommu->domain_ids);
2894 * We need a marker for copied context entries. This
2895 * marker needs to work for the old format as well as
2896 * for extended context entries.
2898 * Bit 67 of the context entry is used. In the old
2899 * format this bit is available to software, in the
2900 * extended format it is the PGE bit, but PGE is ignored
2901 * by HW if PASIDs are disabled (and thus still
2904 * So disable PASIDs first and then mark the entry
2905 * copied. This means that we don't copy PASID
2906 * translations from the old kernel, but this is fine as
2907 * faults there are not fatal.
2909 context_clear_pasid_enable(&ce);
2910 context_set_copied(&ce);
2915 tbl[tbl_idx + pos] = new_ce;
2917 __iommu_flush_cache(iommu, new_ce, VTD_PAGE_SIZE);
2926 static int copy_translation_tables(struct intel_iommu *iommu)
2928 struct context_entry **ctxt_tbls;
2929 struct root_entry *old_rt;
2930 phys_addr_t old_rt_phys;
2931 int ctxt_table_entries;
2932 unsigned long flags;
2937 rtaddr_reg = dmar_readq(iommu->reg + DMAR_RTADDR_REG);
2938 ext = !!(rtaddr_reg & DMA_RTADDR_RTT);
2939 new_ext = !!ecap_ecs(iommu->ecap);
2942 * The RTT bit can only be changed when translation is disabled,
2943 * but disabling translation means to open a window for data
2944 * corruption. So bail out and don't copy anything if we would
2945 * have to change the bit.
2950 old_rt_phys = rtaddr_reg & VTD_PAGE_MASK;
2954 old_rt = ioremap_cache(old_rt_phys, PAGE_SIZE);
2958 /* This is too big for the stack - allocate it from slab */
2959 ctxt_table_entries = ext ? 512 : 256;
2961 ctxt_tbls = kzalloc(ctxt_table_entries * sizeof(void *), GFP_KERNEL);
2965 for (bus = 0; bus < 256; bus++) {
2966 ret = copy_context_table(iommu, &old_rt[bus],
2967 ctxt_tbls, bus, ext);
2969 pr_err("%s: Failed to copy context table for bus %d\n",
2975 spin_lock_irqsave(&iommu->lock, flags);
2977 /* Context tables are copied, now write them to the root_entry table */
2978 for (bus = 0; bus < 256; bus++) {
2979 int idx = ext ? bus * 2 : bus;
2982 if (ctxt_tbls[idx]) {
2983 val = virt_to_phys(ctxt_tbls[idx]) | 1;
2984 iommu->root_entry[bus].lo = val;
2987 if (!ext || !ctxt_tbls[idx + 1])
2990 val = virt_to_phys(ctxt_tbls[idx + 1]) | 1;
2991 iommu->root_entry[bus].hi = val;
2994 spin_unlock_irqrestore(&iommu->lock, flags);
2998 __iommu_flush_cache(iommu, iommu->root_entry, PAGE_SIZE);
3008 static int __init init_dmars(void)
3010 struct dmar_drhd_unit *drhd;
3011 struct dmar_rmrr_unit *rmrr;
3012 bool copied_tables = false;
3014 struct intel_iommu *iommu;
3020 * initialize and program root entry to not present
3023 for_each_drhd_unit(drhd) {
3025 * lock not needed as this is only incremented in the single
3026 * threaded kernel __init code path all other access are read
3029 if (g_num_of_iommus < DMAR_UNITS_SUPPORTED) {
3033 pr_err_once("Exceeded %d IOMMUs\n", DMAR_UNITS_SUPPORTED);
3036 /* Preallocate enough resources for IOMMU hot-addition */
3037 if (g_num_of_iommus < DMAR_UNITS_SUPPORTED)
3038 g_num_of_iommus = DMAR_UNITS_SUPPORTED;
3040 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
3043 pr_err("Allocating global iommu array failed\n");
3048 deferred_flush = kzalloc(g_num_of_iommus *
3049 sizeof(struct deferred_flush_tables), GFP_KERNEL);
3050 if (!deferred_flush) {
3055 for_each_active_iommu(iommu, drhd) {
3056 g_iommus[iommu->seq_id] = iommu;
3058 intel_iommu_init_qi(iommu);
3060 ret = iommu_init_domains(iommu);
3064 init_translation_status(iommu);
3066 if (translation_pre_enabled(iommu) && !is_kdump_kernel()) {
3067 iommu_disable_translation(iommu);
3068 clear_translation_pre_enabled(iommu);
3069 pr_warn("Translation was enabled for %s but we are not in kdump mode\n",
3075 * we could share the same root & context tables
3076 * among all IOMMU's. Need to Split it later.
3078 ret = iommu_alloc_root_entry(iommu);
3082 if (translation_pre_enabled(iommu)) {
3083 pr_info("Translation already enabled - trying to copy translation structures\n");
3085 ret = copy_translation_tables(iommu);
3088 * We found the IOMMU with translation
3089 * enabled - but failed to copy over the
3090 * old root-entry table. Try to proceed
3091 * by disabling translation now and
3092 * allocating a clean root-entry table.
3093 * This might cause DMAR faults, but
3094 * probably the dump will still succeed.
3096 pr_err("Failed to copy translation tables from previous kernel for %s\n",
3098 iommu_disable_translation(iommu);
3099 clear_translation_pre_enabled(iommu);
3101 pr_info("Copied translation tables from previous kernel for %s\n",
3103 copied_tables = true;
3107 iommu_flush_write_buffer(iommu);
3108 iommu_set_root_entry(iommu);
3109 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
3110 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
3112 if (!ecap_pass_through(iommu->ecap))
3113 hw_pass_through = 0;
3116 if (iommu_pass_through)
3117 iommu_identity_mapping |= IDENTMAP_ALL;
3119 #ifdef CONFIG_INTEL_IOMMU_BROKEN_GFX_WA
3120 iommu_identity_mapping |= IDENTMAP_GFX;
3123 if (iommu_identity_mapping) {
3124 ret = si_domain_init(hw_pass_through);
3129 check_tylersburg_isoch();
3132 * If we copied translations from a previous kernel in the kdump
3133 * case, we can not assign the devices to domains now, as that
3134 * would eliminate the old mappings. So skip this part and defer
3135 * the assignment to device driver initialization time.
3141 * If pass through is not set or not enabled, setup context entries for
3142 * identity mappings for rmrr, gfx, and isa and may fall back to static
3143 * identity mapping if iommu_identity_mapping is set.
3145 if (iommu_identity_mapping) {
3146 ret = iommu_prepare_static_identity_mapping(hw_pass_through);
3148 pr_crit("Failed to setup IOMMU pass-through\n");
3154 * for each dev attached to rmrr
3156 * locate drhd for dev, alloc domain for dev
3157 * allocate free domain
3158 * allocate page table entries for rmrr
3159 * if context not allocated for bus
3160 * allocate and init context
3161 * set present in root table for this bus
3162 * init context with domain, translation etc
3166 pr_info("Setting RMRR:\n");
3167 for_each_rmrr_units(rmrr) {
3168 /* some BIOS lists non-exist devices in DMAR table. */
3169 for_each_active_dev_scope(rmrr->devices, rmrr->devices_cnt,
3171 ret = iommu_prepare_rmrr_dev(rmrr, dev);
3173 pr_err("Mapping reserved region failed\n");
3177 iommu_prepare_isa();
3184 * global invalidate context cache
3185 * global invalidate iotlb
3186 * enable translation
3188 for_each_iommu(iommu, drhd) {
3189 if (drhd->ignored) {
3191 * we always have to disable PMRs or DMA may fail on
3195 iommu_disable_protect_mem_regions(iommu);
3199 iommu_flush_write_buffer(iommu);
3201 ret = dmar_set_interrupt(iommu);
3205 if (!translation_pre_enabled(iommu))
3206 iommu_enable_translation(iommu);
3208 iommu_disable_protect_mem_regions(iommu);
3214 for_each_active_iommu(iommu, drhd) {
3215 disable_dmar_iommu(iommu);
3216 free_dmar_iommu(iommu);
3218 kfree(deferred_flush);
3225 /* This takes a number of _MM_ pages, not VTD pages */
3226 static struct iova *intel_alloc_iova(struct device *dev,
3227 struct dmar_domain *domain,
3228 unsigned long nrpages, uint64_t dma_mask)
3230 struct iova *iova = NULL;
3232 /* Restrict dma_mask to the width that the iommu can handle */
3233 dma_mask = min_t(uint64_t, DOMAIN_MAX_ADDR(domain->gaw), dma_mask);
3235 if (!dmar_forcedac && dma_mask > DMA_BIT_MASK(32)) {
3237 * First try to allocate an io virtual address in
3238 * DMA_BIT_MASK(32) and if that fails then try allocating
3241 iova = alloc_iova(&domain->iovad, nrpages,
3242 IOVA_PFN(DMA_BIT_MASK(32)), 1);
3246 iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
3247 if (unlikely(!iova)) {
3248 pr_err("Allocating %ld-page iova for %s failed",
3249 nrpages, dev_name(dev));
3256 static struct dmar_domain *__get_valid_domain_for_dev(struct device *dev)
3258 struct dmar_domain *domain;
3260 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
3262 pr_err("Allocating domain for %s failed\n",
3270 static inline struct dmar_domain *get_valid_domain_for_dev(struct device *dev)
3272 struct device_domain_info *info;
3274 /* No lock here, assumes no domain exit in normal case */
3275 info = dev->archdata.iommu;
3277 return info->domain;
3279 return __get_valid_domain_for_dev(dev);
3282 /* Check if the dev needs to go through non-identity map and unmap process.*/
3283 static int iommu_no_mapping(struct device *dev)
3287 if (iommu_dummy(dev))
3290 if (!iommu_identity_mapping)
3293 found = identity_mapping(dev);
3295 if (iommu_should_identity_map(dev, 0))
3299 * 32 bit DMA is removed from si_domain and fall back
3300 * to non-identity mapping.
3302 dmar_remove_one_dev_info(si_domain, dev);
3303 pr_info("32bit %s uses non-identity mapping\n",
3309 * In case of a detached 64 bit DMA device from vm, the device
3310 * is put into si_domain for identity mapping.
3312 if (iommu_should_identity_map(dev, 0)) {
3314 ret = domain_add_dev_info(si_domain, dev);
3316 pr_info("64bit %s uses identity mapping\n",
3326 static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
3327 size_t size, int dir, u64 dma_mask)
3329 struct dmar_domain *domain;
3330 phys_addr_t start_paddr;
3334 struct intel_iommu *iommu;
3335 unsigned long paddr_pfn = paddr >> PAGE_SHIFT;
3337 BUG_ON(dir == DMA_NONE);
3339 if (iommu_no_mapping(dev))
3342 domain = get_valid_domain_for_dev(dev);
3346 iommu = domain_get_iommu(domain);
3347 size = aligned_nrpages(paddr, size);
3349 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size), dma_mask);
3354 * Check if DMAR supports zero-length reads on write only
3357 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3358 !cap_zlr(iommu->cap))
3359 prot |= DMA_PTE_READ;
3360 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3361 prot |= DMA_PTE_WRITE;
3363 * paddr - (paddr + size) might be partial page, we should map the whole
3364 * page. Note: if two part of one page are separately mapped, we
3365 * might have two guest_addr mapping to the same host paddr, but this
3366 * is not a big problem
3368 ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
3369 mm_to_dma_pfn(paddr_pfn), size, prot);
3373 /* it's a non-present to present mapping. Only flush if caching mode */
3374 if (cap_caching_mode(iommu->cap))
3375 iommu_flush_iotlb_psi(iommu, domain,
3376 mm_to_dma_pfn(iova->pfn_lo),
3379 iommu_flush_write_buffer(iommu);
3381 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
3382 start_paddr += paddr & ~PAGE_MASK;
3387 __free_iova(&domain->iovad, iova);
3388 pr_err("Device %s request: %zx@%llx dir %d --- failed\n",
3389 dev_name(dev), size, (unsigned long long)paddr, dir);
3393 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
3394 unsigned long offset, size_t size,
3395 enum dma_data_direction dir,
3396 struct dma_attrs *attrs)
3398 return __intel_map_single(dev, page_to_phys(page) + offset, size,
3399 dir, *dev->dma_mask);
3402 static void flush_unmaps(void)
3408 /* just flush them all */
3409 for (i = 0; i < g_num_of_iommus; i++) {
3410 struct intel_iommu *iommu = g_iommus[i];
3414 if (!deferred_flush[i].next)
3417 /* In caching mode, global flushes turn emulation expensive */
3418 if (!cap_caching_mode(iommu->cap))
3419 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3420 DMA_TLB_GLOBAL_FLUSH);
3421 for (j = 0; j < deferred_flush[i].next; j++) {
3423 struct iova *iova = deferred_flush[i].iova[j];
3424 struct dmar_domain *domain = deferred_flush[i].domain[j];
3426 /* On real hardware multiple invalidations are expensive */
3427 if (cap_caching_mode(iommu->cap))
3428 iommu_flush_iotlb_psi(iommu, domain,
3429 iova->pfn_lo, iova_size(iova),
3430 !deferred_flush[i].freelist[j], 0);
3432 mask = ilog2(mm_to_dma_pfn(iova_size(iova)));
3433 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
3434 (uint64_t)iova->pfn_lo << PAGE_SHIFT, mask);
3436 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
3437 if (deferred_flush[i].freelist[j])
3438 dma_free_pagelist(deferred_flush[i].freelist[j]);
3440 deferred_flush[i].next = 0;
3446 static void flush_unmaps_timeout(unsigned long data)
3448 unsigned long flags;
3450 spin_lock_irqsave(&async_umap_flush_lock, flags);
3452 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3455 static void add_unmap(struct dmar_domain *dom, struct iova *iova, struct page *freelist)
3457 unsigned long flags;
3459 struct intel_iommu *iommu;
3461 spin_lock_irqsave(&async_umap_flush_lock, flags);
3462 if (list_size == HIGH_WATER_MARK)
3465 iommu = domain_get_iommu(dom);
3466 iommu_id = iommu->seq_id;
3468 next = deferred_flush[iommu_id].next;
3469 deferred_flush[iommu_id].domain[next] = dom;
3470 deferred_flush[iommu_id].iova[next] = iova;
3471 deferred_flush[iommu_id].freelist[next] = freelist;
3472 deferred_flush[iommu_id].next++;
3475 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
3479 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3482 static void intel_unmap(struct device *dev, dma_addr_t dev_addr)
3484 struct dmar_domain *domain;
3485 unsigned long start_pfn, last_pfn;
3487 struct intel_iommu *iommu;
3488 struct page *freelist;
3490 if (iommu_no_mapping(dev))
3493 domain = find_domain(dev);
3496 iommu = domain_get_iommu(domain);
3498 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
3499 if (WARN_ONCE(!iova, "Driver unmaps unmatched page at PFN %llx\n",
3500 (unsigned long long)dev_addr))
3503 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
3504 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
3506 pr_debug("Device %s unmapping: pfn %lx-%lx\n",
3507 dev_name(dev), start_pfn, last_pfn);
3509 freelist = domain_unmap(domain, start_pfn, last_pfn);
3511 if (intel_iommu_strict) {
3512 iommu_flush_iotlb_psi(iommu, domain, start_pfn,
3513 last_pfn - start_pfn + 1, !freelist, 0);
3515 __free_iova(&domain->iovad, iova);
3516 dma_free_pagelist(freelist);
3518 add_unmap(domain, iova, freelist);
3520 * queue up the release of the unmap to save the 1/6th of the
3521 * cpu used up by the iotlb flush operation...
3526 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
3527 size_t size, enum dma_data_direction dir,
3528 struct dma_attrs *attrs)
3530 intel_unmap(dev, dev_addr);
3533 static void *intel_alloc_coherent(struct device *dev, size_t size,
3534 dma_addr_t *dma_handle, gfp_t flags,
3535 struct dma_attrs *attrs)
3537 struct page *page = NULL;
3540 size = PAGE_ALIGN(size);
3541 order = get_order(size);
3543 if (!iommu_no_mapping(dev))
3544 flags &= ~(GFP_DMA | GFP_DMA32);
3545 else if (dev->coherent_dma_mask < dma_get_required_mask(dev)) {
3546 if (dev->coherent_dma_mask < DMA_BIT_MASK(32))
3552 if (flags & __GFP_WAIT) {
3553 unsigned int count = size >> PAGE_SHIFT;
3555 page = dma_alloc_from_contiguous(dev, count, order);
3556 if (page && iommu_no_mapping(dev) &&
3557 page_to_phys(page) + size > dev->coherent_dma_mask) {
3558 dma_release_from_contiguous(dev, page, count);
3564 page = alloc_pages(flags, order);
3567 memset(page_address(page), 0, size);
3569 *dma_handle = __intel_map_single(dev, page_to_phys(page), size,
3571 dev->coherent_dma_mask);
3573 return page_address(page);
3574 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3575 __free_pages(page, order);
3580 static void intel_free_coherent(struct device *dev, size_t size, void *vaddr,
3581 dma_addr_t dma_handle, struct dma_attrs *attrs)
3584 struct page *page = virt_to_page(vaddr);
3586 size = PAGE_ALIGN(size);
3587 order = get_order(size);
3589 intel_unmap(dev, dma_handle);
3590 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3591 __free_pages(page, order);
3594 static void intel_unmap_sg(struct device *dev, struct scatterlist *sglist,
3595 int nelems, enum dma_data_direction dir,
3596 struct dma_attrs *attrs)
3598 intel_unmap(dev, sglist[0].dma_address);
3601 static int intel_nontranslate_map_sg(struct device *hddev,
3602 struct scatterlist *sglist, int nelems, int dir)
3605 struct scatterlist *sg;
3607 for_each_sg(sglist, sg, nelems, i) {
3608 BUG_ON(!sg_page(sg));
3609 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
3610 sg->dma_length = sg->length;
3615 static int intel_map_sg(struct device *dev, struct scatterlist *sglist, int nelems,
3616 enum dma_data_direction dir, struct dma_attrs *attrs)
3619 struct dmar_domain *domain;
3622 struct iova *iova = NULL;
3624 struct scatterlist *sg;
3625 unsigned long start_vpfn;
3626 struct intel_iommu *iommu;
3628 BUG_ON(dir == DMA_NONE);
3629 if (iommu_no_mapping(dev))
3630 return intel_nontranslate_map_sg(dev, sglist, nelems, dir);
3632 domain = get_valid_domain_for_dev(dev);
3636 iommu = domain_get_iommu(domain);
3638 for_each_sg(sglist, sg, nelems, i)
3639 size += aligned_nrpages(sg->offset, sg->length);
3641 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size),
3644 sglist->dma_length = 0;
3649 * Check if DMAR supports zero-length reads on write only
3652 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3653 !cap_zlr(iommu->cap))
3654 prot |= DMA_PTE_READ;
3655 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3656 prot |= DMA_PTE_WRITE;
3658 start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
3660 ret = domain_sg_mapping(domain, start_vpfn, sglist, size, prot);
3661 if (unlikely(ret)) {
3662 dma_pte_free_pagetable(domain, start_vpfn,
3663 start_vpfn + size - 1);
3664 __free_iova(&domain->iovad, iova);
3668 /* it's a non-present to present mapping. Only flush if caching mode */
3669 if (cap_caching_mode(iommu->cap))
3670 iommu_flush_iotlb_psi(iommu, domain, start_vpfn, size, 0, 1);
3672 iommu_flush_write_buffer(iommu);
3677 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
3682 struct dma_map_ops intel_dma_ops = {
3683 .alloc = intel_alloc_coherent,
3684 .free = intel_free_coherent,
3685 .map_sg = intel_map_sg,
3686 .unmap_sg = intel_unmap_sg,
3687 .map_page = intel_map_page,
3688 .unmap_page = intel_unmap_page,
3689 .mapping_error = intel_mapping_error,
3692 static inline int iommu_domain_cache_init(void)
3696 iommu_domain_cache = kmem_cache_create("iommu_domain",
3697 sizeof(struct dmar_domain),
3702 if (!iommu_domain_cache) {
3703 pr_err("Couldn't create iommu_domain cache\n");
3710 static inline int iommu_devinfo_cache_init(void)
3714 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
3715 sizeof(struct device_domain_info),
3719 if (!iommu_devinfo_cache) {
3720 pr_err("Couldn't create devinfo cache\n");
3727 static int __init iommu_init_mempool(void)
3730 ret = iommu_iova_cache_init();
3734 ret = iommu_domain_cache_init();
3738 ret = iommu_devinfo_cache_init();
3742 kmem_cache_destroy(iommu_domain_cache);
3744 iommu_iova_cache_destroy();
3749 static void __init iommu_exit_mempool(void)
3751 kmem_cache_destroy(iommu_devinfo_cache);
3752 kmem_cache_destroy(iommu_domain_cache);
3753 iommu_iova_cache_destroy();
3756 static void quirk_ioat_snb_local_iommu(struct pci_dev *pdev)
3758 struct dmar_drhd_unit *drhd;
3762 /* We know that this device on this chipset has its own IOMMU.
3763 * If we find it under a different IOMMU, then the BIOS is lying
3764 * to us. Hope that the IOMMU for this device is actually
3765 * disabled, and it needs no translation...
3767 rc = pci_bus_read_config_dword(pdev->bus, PCI_DEVFN(0, 0), 0xb0, &vtbar);
3769 /* "can't" happen */
3770 dev_info(&pdev->dev, "failed to run vt-d quirk\n");
3773 vtbar &= 0xffff0000;
3775 /* we know that the this iommu should be at offset 0xa000 from vtbar */
3776 drhd = dmar_find_matched_drhd_unit(pdev);
3777 if (WARN_TAINT_ONCE(!drhd || drhd->reg_base_addr - vtbar != 0xa000,
3778 TAINT_FIRMWARE_WORKAROUND,
3779 "BIOS assigned incorrect VT-d unit for Intel(R) QuickData Technology device\n"))
3780 pdev->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3782 DECLARE_PCI_FIXUP_ENABLE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB, quirk_ioat_snb_local_iommu);
3784 static void __init init_no_remapping_devices(void)
3786 struct dmar_drhd_unit *drhd;
3790 for_each_drhd_unit(drhd) {
3791 if (!drhd->include_all) {
3792 for_each_active_dev_scope(drhd->devices,
3793 drhd->devices_cnt, i, dev)
3795 /* ignore DMAR unit if no devices exist */
3796 if (i == drhd->devices_cnt)
3801 for_each_active_drhd_unit(drhd) {
3802 if (drhd->include_all)
3805 for_each_active_dev_scope(drhd->devices,
3806 drhd->devices_cnt, i, dev)
3807 if (!dev_is_pci(dev) || !IS_GFX_DEVICE(to_pci_dev(dev)))
3809 if (i < drhd->devices_cnt)
3812 /* This IOMMU has *only* gfx devices. Either bypass it or
3813 set the gfx_mapped flag, as appropriate */
3815 intel_iommu_gfx_mapped = 1;
3818 for_each_active_dev_scope(drhd->devices,
3819 drhd->devices_cnt, i, dev)
3820 dev->archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3825 #ifdef CONFIG_SUSPEND
3826 static int init_iommu_hw(void)
3828 struct dmar_drhd_unit *drhd;
3829 struct intel_iommu *iommu = NULL;
3831 for_each_active_iommu(iommu, drhd)
3833 dmar_reenable_qi(iommu);
3835 for_each_iommu(iommu, drhd) {
3836 if (drhd->ignored) {
3838 * we always have to disable PMRs or DMA may fail on
3842 iommu_disable_protect_mem_regions(iommu);
3846 iommu_flush_write_buffer(iommu);
3848 iommu_set_root_entry(iommu);
3850 iommu->flush.flush_context(iommu, 0, 0, 0,
3851 DMA_CCMD_GLOBAL_INVL);
3852 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
3853 iommu_enable_translation(iommu);
3854 iommu_disable_protect_mem_regions(iommu);
3860 static void iommu_flush_all(void)
3862 struct dmar_drhd_unit *drhd;
3863 struct intel_iommu *iommu;
3865 for_each_active_iommu(iommu, drhd) {
3866 iommu->flush.flush_context(iommu, 0, 0, 0,
3867 DMA_CCMD_GLOBAL_INVL);
3868 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3869 DMA_TLB_GLOBAL_FLUSH);
3873 static int iommu_suspend(void)
3875 struct dmar_drhd_unit *drhd;
3876 struct intel_iommu *iommu = NULL;
3879 for_each_active_iommu(iommu, drhd) {
3880 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3882 if (!iommu->iommu_state)
3888 for_each_active_iommu(iommu, drhd) {
3889 iommu_disable_translation(iommu);
3891 raw_spin_lock_irqsave(&iommu->register_lock, flag);
3893 iommu->iommu_state[SR_DMAR_FECTL_REG] =
3894 readl(iommu->reg + DMAR_FECTL_REG);
3895 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
3896 readl(iommu->reg + DMAR_FEDATA_REG);
3897 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
3898 readl(iommu->reg + DMAR_FEADDR_REG);
3899 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
3900 readl(iommu->reg + DMAR_FEUADDR_REG);
3902 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
3907 for_each_active_iommu(iommu, drhd)
3908 kfree(iommu->iommu_state);
3913 static void iommu_resume(void)
3915 struct dmar_drhd_unit *drhd;
3916 struct intel_iommu *iommu = NULL;
3919 if (init_iommu_hw()) {
3921 panic("tboot: IOMMU setup failed, DMAR can not resume!\n");
3923 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
3927 for_each_active_iommu(iommu, drhd) {
3929 raw_spin_lock_irqsave(&iommu->register_lock, flag);
3931 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
3932 iommu->reg + DMAR_FECTL_REG);
3933 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
3934 iommu->reg + DMAR_FEDATA_REG);
3935 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
3936 iommu->reg + DMAR_FEADDR_REG);
3937 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
3938 iommu->reg + DMAR_FEUADDR_REG);
3940 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
3943 for_each_active_iommu(iommu, drhd)
3944 kfree(iommu->iommu_state);
3947 static struct syscore_ops iommu_syscore_ops = {
3948 .resume = iommu_resume,
3949 .suspend = iommu_suspend,
3952 static void __init init_iommu_pm_ops(void)
3954 register_syscore_ops(&iommu_syscore_ops);
3958 static inline void init_iommu_pm_ops(void) {}
3959 #endif /* CONFIG_PM */
3962 int __init dmar_parse_one_rmrr(struct acpi_dmar_header *header, void *arg)
3964 struct acpi_dmar_reserved_memory *rmrr;
3965 struct dmar_rmrr_unit *rmrru;
3967 rmrru = kzalloc(sizeof(*rmrru), GFP_KERNEL);
3971 rmrru->hdr = header;
3972 rmrr = (struct acpi_dmar_reserved_memory *)header;
3973 rmrru->base_address = rmrr->base_address;
3974 rmrru->end_address = rmrr->end_address;
3975 rmrru->devices = dmar_alloc_dev_scope((void *)(rmrr + 1),
3976 ((void *)rmrr) + rmrr->header.length,
3977 &rmrru->devices_cnt);
3978 if (rmrru->devices_cnt && rmrru->devices == NULL) {
3983 list_add(&rmrru->list, &dmar_rmrr_units);
3988 static struct dmar_atsr_unit *dmar_find_atsr(struct acpi_dmar_atsr *atsr)
3990 struct dmar_atsr_unit *atsru;
3991 struct acpi_dmar_atsr *tmp;
3993 list_for_each_entry_rcu(atsru, &dmar_atsr_units, list) {
3994 tmp = (struct acpi_dmar_atsr *)atsru->hdr;
3995 if (atsr->segment != tmp->segment)
3997 if (atsr->header.length != tmp->header.length)
3999 if (memcmp(atsr, tmp, atsr->header.length) == 0)
4006 int dmar_parse_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4008 struct acpi_dmar_atsr *atsr;
4009 struct dmar_atsr_unit *atsru;
4011 if (system_state != SYSTEM_BOOTING && !intel_iommu_enabled)
4014 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4015 atsru = dmar_find_atsr(atsr);
4019 atsru = kzalloc(sizeof(*atsru) + hdr->length, GFP_KERNEL);
4024 * If memory is allocated from slab by ACPI _DSM method, we need to
4025 * copy the memory content because the memory buffer will be freed
4028 atsru->hdr = (void *)(atsru + 1);
4029 memcpy(atsru->hdr, hdr, hdr->length);
4030 atsru->include_all = atsr->flags & 0x1;
4031 if (!atsru->include_all) {
4032 atsru->devices = dmar_alloc_dev_scope((void *)(atsr + 1),
4033 (void *)atsr + atsr->header.length,
4034 &atsru->devices_cnt);
4035 if (atsru->devices_cnt && atsru->devices == NULL) {
4041 list_add_rcu(&atsru->list, &dmar_atsr_units);
4046 static void intel_iommu_free_atsr(struct dmar_atsr_unit *atsru)
4048 dmar_free_dev_scope(&atsru->devices, &atsru->devices_cnt);
4052 int dmar_release_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4054 struct acpi_dmar_atsr *atsr;
4055 struct dmar_atsr_unit *atsru;
4057 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4058 atsru = dmar_find_atsr(atsr);
4060 list_del_rcu(&atsru->list);
4062 intel_iommu_free_atsr(atsru);
4068 int dmar_check_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4072 struct acpi_dmar_atsr *atsr;
4073 struct dmar_atsr_unit *atsru;
4075 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4076 atsru = dmar_find_atsr(atsr);
4080 if (!atsru->include_all && atsru->devices && atsru->devices_cnt)
4081 for_each_active_dev_scope(atsru->devices, atsru->devices_cnt,
4088 static int intel_iommu_add(struct dmar_drhd_unit *dmaru)
4091 struct intel_iommu *iommu = dmaru->iommu;
4093 if (g_iommus[iommu->seq_id])
4096 if (hw_pass_through && !ecap_pass_through(iommu->ecap)) {
4097 pr_warn("%s: Doesn't support hardware pass through.\n",
4101 if (!ecap_sc_support(iommu->ecap) &&
4102 domain_update_iommu_snooping(iommu)) {
4103 pr_warn("%s: Doesn't support snooping.\n",
4107 sp = domain_update_iommu_superpage(iommu) - 1;
4108 if (sp >= 0 && !(cap_super_page_val(iommu->cap) & (1 << sp))) {
4109 pr_warn("%s: Doesn't support large page.\n",
4115 * Disable translation if already enabled prior to OS handover.
4117 if (iommu->gcmd & DMA_GCMD_TE)
4118 iommu_disable_translation(iommu);
4120 g_iommus[iommu->seq_id] = iommu;
4121 ret = iommu_init_domains(iommu);
4123 ret = iommu_alloc_root_entry(iommu);
4127 if (dmaru->ignored) {
4129 * we always have to disable PMRs or DMA may fail on this device
4132 iommu_disable_protect_mem_regions(iommu);
4136 intel_iommu_init_qi(iommu);
4137 iommu_flush_write_buffer(iommu);
4138 ret = dmar_set_interrupt(iommu);
4142 iommu_set_root_entry(iommu);
4143 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
4144 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
4145 iommu_enable_translation(iommu);
4147 iommu_disable_protect_mem_regions(iommu);
4151 disable_dmar_iommu(iommu);
4153 free_dmar_iommu(iommu);
4157 int dmar_iommu_hotplug(struct dmar_drhd_unit *dmaru, bool insert)
4160 struct intel_iommu *iommu = dmaru->iommu;
4162 if (!intel_iommu_enabled)
4168 ret = intel_iommu_add(dmaru);
4170 disable_dmar_iommu(iommu);
4171 free_dmar_iommu(iommu);
4177 static void intel_iommu_free_dmars(void)
4179 struct dmar_rmrr_unit *rmrru, *rmrr_n;
4180 struct dmar_atsr_unit *atsru, *atsr_n;
4182 list_for_each_entry_safe(rmrru, rmrr_n, &dmar_rmrr_units, list) {
4183 list_del(&rmrru->list);
4184 dmar_free_dev_scope(&rmrru->devices, &rmrru->devices_cnt);
4188 list_for_each_entry_safe(atsru, atsr_n, &dmar_atsr_units, list) {
4189 list_del(&atsru->list);
4190 intel_iommu_free_atsr(atsru);
4194 int dmar_find_matched_atsr_unit(struct pci_dev *dev)
4197 struct pci_bus *bus;
4198 struct pci_dev *bridge = NULL;
4200 struct acpi_dmar_atsr *atsr;
4201 struct dmar_atsr_unit *atsru;
4203 dev = pci_physfn(dev);
4204 for (bus = dev->bus; bus; bus = bus->parent) {
4206 if (!bridge || !pci_is_pcie(bridge) ||
4207 pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE)
4209 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_ROOT_PORT)
4216 list_for_each_entry_rcu(atsru, &dmar_atsr_units, list) {
4217 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
4218 if (atsr->segment != pci_domain_nr(dev->bus))
4221 for_each_dev_scope(atsru->devices, atsru->devices_cnt, i, tmp)
4222 if (tmp == &bridge->dev)
4225 if (atsru->include_all)
4235 int dmar_iommu_notify_scope_dev(struct dmar_pci_notify_info *info)
4238 struct dmar_rmrr_unit *rmrru;
4239 struct dmar_atsr_unit *atsru;
4240 struct acpi_dmar_atsr *atsr;
4241 struct acpi_dmar_reserved_memory *rmrr;
4243 if (!intel_iommu_enabled && system_state != SYSTEM_BOOTING)
4246 list_for_each_entry(rmrru, &dmar_rmrr_units, list) {
4247 rmrr = container_of(rmrru->hdr,
4248 struct acpi_dmar_reserved_memory, header);
4249 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
4250 ret = dmar_insert_dev_scope(info, (void *)(rmrr + 1),
4251 ((void *)rmrr) + rmrr->header.length,
4252 rmrr->segment, rmrru->devices,
4253 rmrru->devices_cnt);
4256 } else if (info->event == BUS_NOTIFY_DEL_DEVICE) {
4257 dmar_remove_dev_scope(info, rmrr->segment,
4258 rmrru->devices, rmrru->devices_cnt);
4262 list_for_each_entry(atsru, &dmar_atsr_units, list) {
4263 if (atsru->include_all)
4266 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
4267 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
4268 ret = dmar_insert_dev_scope(info, (void *)(atsr + 1),
4269 (void *)atsr + atsr->header.length,
4270 atsr->segment, atsru->devices,
4271 atsru->devices_cnt);
4276 } else if (info->event == BUS_NOTIFY_DEL_DEVICE) {
4277 if (dmar_remove_dev_scope(info, atsr->segment,
4278 atsru->devices, atsru->devices_cnt))
4287 * Here we only respond to action of unbound device from driver.
4289 * Added device is not attached to its DMAR domain here yet. That will happen
4290 * when mapping the device to iova.
4292 static int device_notifier(struct notifier_block *nb,
4293 unsigned long action, void *data)
4295 struct device *dev = data;
4296 struct dmar_domain *domain;
4298 if (iommu_dummy(dev))
4301 if (action != BUS_NOTIFY_REMOVED_DEVICE)
4304 domain = find_domain(dev);
4308 down_read(&dmar_global_lock);
4309 dmar_remove_one_dev_info(domain, dev);
4310 if (!domain_type_is_vm_or_si(domain) && list_empty(&domain->devices))
4311 domain_exit(domain);
4312 up_read(&dmar_global_lock);
4317 static struct notifier_block device_nb = {
4318 .notifier_call = device_notifier,
4321 static int intel_iommu_memory_notifier(struct notifier_block *nb,
4322 unsigned long val, void *v)
4324 struct memory_notify *mhp = v;
4325 unsigned long long start, end;
4326 unsigned long start_vpfn, last_vpfn;
4329 case MEM_GOING_ONLINE:
4330 start = mhp->start_pfn << PAGE_SHIFT;
4331 end = ((mhp->start_pfn + mhp->nr_pages) << PAGE_SHIFT) - 1;
4332 if (iommu_domain_identity_map(si_domain, start, end)) {
4333 pr_warn("Failed to build identity map for [%llx-%llx]\n",
4340 case MEM_CANCEL_ONLINE:
4341 start_vpfn = mm_to_dma_pfn(mhp->start_pfn);
4342 last_vpfn = mm_to_dma_pfn(mhp->start_pfn + mhp->nr_pages - 1);
4343 while (start_vpfn <= last_vpfn) {
4345 struct dmar_drhd_unit *drhd;
4346 struct intel_iommu *iommu;
4347 struct page *freelist;
4349 iova = find_iova(&si_domain->iovad, start_vpfn);
4351 pr_debug("Failed get IOVA for PFN %lx\n",
4356 iova = split_and_remove_iova(&si_domain->iovad, iova,
4357 start_vpfn, last_vpfn);
4359 pr_warn("Failed to split IOVA PFN [%lx-%lx]\n",
4360 start_vpfn, last_vpfn);
4364 freelist = domain_unmap(si_domain, iova->pfn_lo,
4368 for_each_active_iommu(iommu, drhd)
4369 iommu_flush_iotlb_psi(iommu, si_domain,
4370 iova->pfn_lo, iova_size(iova),
4373 dma_free_pagelist(freelist);
4375 start_vpfn = iova->pfn_hi + 1;
4376 free_iova_mem(iova);
4384 static struct notifier_block intel_iommu_memory_nb = {
4385 .notifier_call = intel_iommu_memory_notifier,
4390 static ssize_t intel_iommu_show_version(struct device *dev,
4391 struct device_attribute *attr,
4394 struct intel_iommu *iommu = dev_get_drvdata(dev);
4395 u32 ver = readl(iommu->reg + DMAR_VER_REG);
4396 return sprintf(buf, "%d:%d\n",
4397 DMAR_VER_MAJOR(ver), DMAR_VER_MINOR(ver));
4399 static DEVICE_ATTR(version, S_IRUGO, intel_iommu_show_version, NULL);
4401 static ssize_t intel_iommu_show_address(struct device *dev,
4402 struct device_attribute *attr,
4405 struct intel_iommu *iommu = dev_get_drvdata(dev);
4406 return sprintf(buf, "%llx\n", iommu->reg_phys);
4408 static DEVICE_ATTR(address, S_IRUGO, intel_iommu_show_address, NULL);
4410 static ssize_t intel_iommu_show_cap(struct device *dev,
4411 struct device_attribute *attr,
4414 struct intel_iommu *iommu = dev_get_drvdata(dev);
4415 return sprintf(buf, "%llx\n", iommu->cap);
4417 static DEVICE_ATTR(cap, S_IRUGO, intel_iommu_show_cap, NULL);
4419 static ssize_t intel_iommu_show_ecap(struct device *dev,
4420 struct device_attribute *attr,
4423 struct intel_iommu *iommu = dev_get_drvdata(dev);
4424 return sprintf(buf, "%llx\n", iommu->ecap);
4426 static DEVICE_ATTR(ecap, S_IRUGO, intel_iommu_show_ecap, NULL);
4428 static ssize_t intel_iommu_show_ndoms(struct device *dev,
4429 struct device_attribute *attr,
4432 struct intel_iommu *iommu = dev_get_drvdata(dev);
4433 return sprintf(buf, "%ld\n", cap_ndoms(iommu->cap));
4435 static DEVICE_ATTR(domains_supported, S_IRUGO, intel_iommu_show_ndoms, NULL);
4437 static ssize_t intel_iommu_show_ndoms_used(struct device *dev,
4438 struct device_attribute *attr,
4441 struct intel_iommu *iommu = dev_get_drvdata(dev);
4442 return sprintf(buf, "%d\n", bitmap_weight(iommu->domain_ids,
4443 cap_ndoms(iommu->cap)));
4445 static DEVICE_ATTR(domains_used, S_IRUGO, intel_iommu_show_ndoms_used, NULL);
4447 static struct attribute *intel_iommu_attrs[] = {
4448 &dev_attr_version.attr,
4449 &dev_attr_address.attr,
4451 &dev_attr_ecap.attr,
4452 &dev_attr_domains_supported.attr,
4453 &dev_attr_domains_used.attr,
4457 static struct attribute_group intel_iommu_group = {
4458 .name = "intel-iommu",
4459 .attrs = intel_iommu_attrs,
4462 const struct attribute_group *intel_iommu_groups[] = {
4467 int __init intel_iommu_init(void)
4470 struct dmar_drhd_unit *drhd;
4471 struct intel_iommu *iommu;
4473 /* VT-d is required for a TXT/tboot launch, so enforce that */
4474 force_on = tboot_force_iommu();
4476 if (iommu_init_mempool()) {
4478 panic("tboot: Failed to initialize iommu memory\n");
4482 down_write(&dmar_global_lock);
4483 if (dmar_table_init()) {
4485 panic("tboot: Failed to initialize DMAR table\n");
4489 if (dmar_dev_scope_init() < 0) {
4491 panic("tboot: Failed to initialize DMAR device scope\n");
4495 if (no_iommu || dmar_disabled)
4498 if (list_empty(&dmar_rmrr_units))
4499 pr_info("No RMRR found\n");
4501 if (list_empty(&dmar_atsr_units))
4502 pr_info("No ATSR found\n");
4504 if (dmar_init_reserved_ranges()) {
4506 panic("tboot: Failed to reserve iommu ranges\n");
4507 goto out_free_reserved_range;
4510 init_no_remapping_devices();
4515 panic("tboot: Failed to initialize DMARs\n");
4516 pr_err("Initialization failed\n");
4517 goto out_free_reserved_range;
4519 up_write(&dmar_global_lock);
4520 pr_info("Intel(R) Virtualization Technology for Directed I/O\n");
4522 init_timer(&unmap_timer);
4523 #ifdef CONFIG_SWIOTLB
4526 dma_ops = &intel_dma_ops;
4528 init_iommu_pm_ops();
4530 for_each_active_iommu(iommu, drhd)
4531 iommu->iommu_dev = iommu_device_create(NULL, iommu,
4535 bus_set_iommu(&pci_bus_type, &intel_iommu_ops);
4536 bus_register_notifier(&pci_bus_type, &device_nb);
4537 if (si_domain && !hw_pass_through)
4538 register_memory_notifier(&intel_iommu_memory_nb);
4540 intel_iommu_enabled = 1;
4544 out_free_reserved_range:
4545 put_iova_domain(&reserved_iova_list);
4547 intel_iommu_free_dmars();
4548 up_write(&dmar_global_lock);
4549 iommu_exit_mempool();
4553 static int domain_context_clear_one_cb(struct pci_dev *pdev, u16 alias, void *opaque)
4555 struct intel_iommu *iommu = opaque;
4557 domain_context_clear_one(iommu, PCI_BUS_NUM(alias), alias & 0xff);
4562 * NB - intel-iommu lacks any sort of reference counting for the users of
4563 * dependent devices. If multiple endpoints have intersecting dependent
4564 * devices, unbinding the driver from any one of them will possibly leave
4565 * the others unable to operate.
4567 static void domain_context_clear(struct intel_iommu *iommu, struct device *dev)
4569 if (!iommu || !dev || !dev_is_pci(dev))
4572 pci_for_each_dma_alias(to_pci_dev(dev), &domain_context_clear_one_cb, iommu);
4575 static void dmar_remove_one_dev_info(struct dmar_domain *domain,
4578 struct device_domain_info *info;
4579 struct intel_iommu *iommu;
4580 unsigned long flags;
4583 iommu = device_to_iommu(dev, &bus, &devfn);
4587 info = dev->archdata.iommu;
4592 spin_lock_irqsave(&device_domain_lock, flags);
4593 unlink_domain_info(info);
4594 spin_unlock_irqrestore(&device_domain_lock, flags);
4596 iommu_disable_dev_iotlb(info);
4597 domain_context_clear(iommu, dev);
4598 free_devinfo_mem(info);
4599 domain_detach_iommu(domain, iommu);
4601 spin_lock_irqsave(&domain->iommu_lock, flags);
4602 if (!domain->iommu_refcnt[iommu->seq_id])
4603 iommu_detach_domain(domain, iommu);
4604 spin_unlock_irqrestore(&domain->iommu_lock, flags);
4607 static int md_domain_init(struct dmar_domain *domain, int guest_width)
4611 init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
4613 domain_reserve_special_ranges(domain);
4615 /* calculate AGAW */
4616 domain->gaw = guest_width;
4617 adjust_width = guestwidth_to_adjustwidth(guest_width);
4618 domain->agaw = width_to_agaw(adjust_width);
4620 domain->iommu_coherency = 0;
4621 domain->iommu_snooping = 0;
4622 domain->iommu_superpage = 0;
4623 domain->max_addr = 0;
4625 /* always allocate the top pgd */
4626 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
4629 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
4633 static struct iommu_domain *intel_iommu_domain_alloc(unsigned type)
4635 struct dmar_domain *dmar_domain;
4636 struct iommu_domain *domain;
4638 if (type != IOMMU_DOMAIN_UNMANAGED)
4641 dmar_domain = alloc_domain(DOMAIN_FLAG_VIRTUAL_MACHINE);
4643 pr_err("Can't allocate dmar_domain\n");
4646 if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
4647 pr_err("Domain initialization failed\n");
4648 domain_exit(dmar_domain);
4651 domain_update_iommu_cap(dmar_domain);
4653 domain = &dmar_domain->domain;
4654 domain->geometry.aperture_start = 0;
4655 domain->geometry.aperture_end = __DOMAIN_MAX_ADDR(dmar_domain->gaw);
4656 domain->geometry.force_aperture = true;
4661 static void intel_iommu_domain_free(struct iommu_domain *domain)
4663 domain_exit(to_dmar_domain(domain));
4666 static int intel_iommu_attach_device(struct iommu_domain *domain,
4669 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4670 struct intel_iommu *iommu;
4674 if (device_is_rmrr_locked(dev)) {
4675 dev_warn(dev, "Device is ineligible for IOMMU domain attach due to platform RMRR requirement. Contact your platform vendor.\n");
4679 /* normally dev is not mapped */
4680 if (unlikely(domain_context_mapped(dev))) {
4681 struct dmar_domain *old_domain;
4683 old_domain = find_domain(dev);
4685 if (domain_type_is_vm_or_si(dmar_domain))
4686 dmar_remove_one_dev_info(old_domain, dev);
4688 domain_remove_dev_info(old_domain);
4690 if (!domain_type_is_vm_or_si(old_domain) &&
4691 list_empty(&old_domain->devices))
4692 domain_exit(old_domain);
4696 iommu = device_to_iommu(dev, &bus, &devfn);
4700 /* check if this iommu agaw is sufficient for max mapped address */
4701 addr_width = agaw_to_width(iommu->agaw);
4702 if (addr_width > cap_mgaw(iommu->cap))
4703 addr_width = cap_mgaw(iommu->cap);
4705 if (dmar_domain->max_addr > (1LL << addr_width)) {
4706 pr_err("%s: iommu width (%d) is not "
4707 "sufficient for the mapped address (%llx)\n",
4708 __func__, addr_width, dmar_domain->max_addr);
4711 dmar_domain->gaw = addr_width;
4714 * Knock out extra levels of page tables if necessary
4716 while (iommu->agaw < dmar_domain->agaw) {
4717 struct dma_pte *pte;
4719 pte = dmar_domain->pgd;
4720 if (dma_pte_present(pte)) {
4721 dmar_domain->pgd = (struct dma_pte *)
4722 phys_to_virt(dma_pte_addr(pte));
4723 free_pgtable_page(pte);
4725 dmar_domain->agaw--;
4728 return domain_add_dev_info(dmar_domain, dev);
4731 static void intel_iommu_detach_device(struct iommu_domain *domain,
4734 dmar_remove_one_dev_info(to_dmar_domain(domain), dev);
4737 static int intel_iommu_map(struct iommu_domain *domain,
4738 unsigned long iova, phys_addr_t hpa,
4739 size_t size, int iommu_prot)
4741 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4746 if (iommu_prot & IOMMU_READ)
4747 prot |= DMA_PTE_READ;
4748 if (iommu_prot & IOMMU_WRITE)
4749 prot |= DMA_PTE_WRITE;
4750 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
4751 prot |= DMA_PTE_SNP;
4753 max_addr = iova + size;
4754 if (dmar_domain->max_addr < max_addr) {
4757 /* check if minimum agaw is sufficient for mapped address */
4758 end = __DOMAIN_MAX_ADDR(dmar_domain->gaw) + 1;
4759 if (end < max_addr) {
4760 pr_err("%s: iommu width (%d) is not "
4761 "sufficient for the mapped address (%llx)\n",
4762 __func__, dmar_domain->gaw, max_addr);
4765 dmar_domain->max_addr = max_addr;
4767 /* Round up size to next multiple of PAGE_SIZE, if it and
4768 the low bits of hpa would take us onto the next page */
4769 size = aligned_nrpages(hpa, size);
4770 ret = domain_pfn_mapping(dmar_domain, iova >> VTD_PAGE_SHIFT,
4771 hpa >> VTD_PAGE_SHIFT, size, prot);
4775 static size_t intel_iommu_unmap(struct iommu_domain *domain,
4776 unsigned long iova, size_t size)
4778 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4779 struct page *freelist = NULL;
4780 struct intel_iommu *iommu;
4781 unsigned long start_pfn, last_pfn;
4782 unsigned int npages;
4783 int iommu_id, level = 0;
4785 /* Cope with horrid API which requires us to unmap more than the
4786 size argument if it happens to be a large-page mapping. */
4787 if (!pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level))
4790 if (size < VTD_PAGE_SIZE << level_to_offset_bits(level))
4791 size = VTD_PAGE_SIZE << level_to_offset_bits(level);
4793 start_pfn = iova >> VTD_PAGE_SHIFT;
4794 last_pfn = (iova + size - 1) >> VTD_PAGE_SHIFT;
4796 freelist = domain_unmap(dmar_domain, start_pfn, last_pfn);
4798 npages = last_pfn - start_pfn + 1;
4800 for_each_domain_iommu(iommu_id, dmar_domain) {
4801 iommu = g_iommus[iommu_id];
4803 iommu_flush_iotlb_psi(g_iommus[iommu_id], dmar_domain,
4804 start_pfn, npages, !freelist, 0);
4807 dma_free_pagelist(freelist);
4809 if (dmar_domain->max_addr == iova + size)
4810 dmar_domain->max_addr = iova;
4815 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
4818 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4819 struct dma_pte *pte;
4823 pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level);
4825 phys = dma_pte_addr(pte);
4830 static bool intel_iommu_capable(enum iommu_cap cap)
4832 if (cap == IOMMU_CAP_CACHE_COHERENCY)
4833 return domain_update_iommu_snooping(NULL) == 1;
4834 if (cap == IOMMU_CAP_INTR_REMAP)
4835 return irq_remapping_enabled == 1;
4840 static int intel_iommu_add_device(struct device *dev)
4842 struct intel_iommu *iommu;
4843 struct iommu_group *group;
4846 iommu = device_to_iommu(dev, &bus, &devfn);
4850 iommu_device_link(iommu->iommu_dev, dev);
4852 group = iommu_group_get_for_dev(dev);
4855 return PTR_ERR(group);
4857 iommu_group_put(group);
4861 static void intel_iommu_remove_device(struct device *dev)
4863 struct intel_iommu *iommu;
4866 iommu = device_to_iommu(dev, &bus, &devfn);
4870 iommu_group_remove_device(dev);
4872 iommu_device_unlink(iommu->iommu_dev, dev);
4875 static const struct iommu_ops intel_iommu_ops = {
4876 .capable = intel_iommu_capable,
4877 .domain_alloc = intel_iommu_domain_alloc,
4878 .domain_free = intel_iommu_domain_free,
4879 .attach_dev = intel_iommu_attach_device,
4880 .detach_dev = intel_iommu_detach_device,
4881 .map = intel_iommu_map,
4882 .unmap = intel_iommu_unmap,
4883 .map_sg = default_iommu_map_sg,
4884 .iova_to_phys = intel_iommu_iova_to_phys,
4885 .add_device = intel_iommu_add_device,
4886 .remove_device = intel_iommu_remove_device,
4887 .pgsize_bitmap = INTEL_IOMMU_PGSIZES,
4890 static void quirk_iommu_g4x_gfx(struct pci_dev *dev)
4892 /* G4x/GM45 integrated gfx dmar support is totally busted. */
4893 pr_info("Disabling IOMMU for graphics on this chipset\n");
4897 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_g4x_gfx);
4898 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_g4x_gfx);
4899 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_g4x_gfx);
4900 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_g4x_gfx);
4901 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_g4x_gfx);
4902 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_g4x_gfx);
4903 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_g4x_gfx);
4905 static void quirk_iommu_rwbf(struct pci_dev *dev)
4908 * Mobile 4 Series Chipset neglects to set RWBF capability,
4909 * but needs it. Same seems to hold for the desktop versions.
4911 pr_info("Forcing write-buffer flush capability\n");
4915 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);
4916 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_rwbf);
4917 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_rwbf);
4918 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_rwbf);
4919 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_rwbf);
4920 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_rwbf);
4921 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_rwbf);
4924 #define GGC_MEMORY_SIZE_MASK (0xf << 8)
4925 #define GGC_MEMORY_SIZE_NONE (0x0 << 8)
4926 #define GGC_MEMORY_SIZE_1M (0x1 << 8)
4927 #define GGC_MEMORY_SIZE_2M (0x3 << 8)
4928 #define GGC_MEMORY_VT_ENABLED (0x8 << 8)
4929 #define GGC_MEMORY_SIZE_2M_VT (0x9 << 8)
4930 #define GGC_MEMORY_SIZE_3M_VT (0xa << 8)
4931 #define GGC_MEMORY_SIZE_4M_VT (0xb << 8)
4933 static void quirk_calpella_no_shadow_gtt(struct pci_dev *dev)
4937 if (pci_read_config_word(dev, GGC, &ggc))
4940 if (!(ggc & GGC_MEMORY_VT_ENABLED)) {
4941 pr_info("BIOS has allocated no shadow GTT; disabling IOMMU for graphics\n");
4943 } else if (dmar_map_gfx) {
4944 /* we have to ensure the gfx device is idle before we flush */
4945 pr_info("Disabling batched IOTLB flush on Ironlake\n");
4946 intel_iommu_strict = 1;
4949 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0040, quirk_calpella_no_shadow_gtt);
4950 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0044, quirk_calpella_no_shadow_gtt);
4951 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0062, quirk_calpella_no_shadow_gtt);
4952 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x006a, quirk_calpella_no_shadow_gtt);
4954 /* On Tylersburg chipsets, some BIOSes have been known to enable the
4955 ISOCH DMAR unit for the Azalia sound device, but not give it any
4956 TLB entries, which causes it to deadlock. Check for that. We do
4957 this in a function called from init_dmars(), instead of in a PCI
4958 quirk, because we don't want to print the obnoxious "BIOS broken"
4959 message if VT-d is actually disabled.
4961 static void __init check_tylersburg_isoch(void)
4963 struct pci_dev *pdev;
4964 uint32_t vtisochctrl;
4966 /* If there's no Azalia in the system anyway, forget it. */
4967 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x3a3e, NULL);
4972 /* System Management Registers. Might be hidden, in which case
4973 we can't do the sanity check. But that's OK, because the
4974 known-broken BIOSes _don't_ actually hide it, so far. */
4975 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x342e, NULL);
4979 if (pci_read_config_dword(pdev, 0x188, &vtisochctrl)) {
4986 /* If Azalia DMA is routed to the non-isoch DMAR unit, fine. */
4987 if (vtisochctrl & 1)
4990 /* Drop all bits other than the number of TLB entries */
4991 vtisochctrl &= 0x1c;
4993 /* If we have the recommended number of TLB entries (16), fine. */
4994 if (vtisochctrl == 0x10)
4997 /* Zero TLB entries? You get to ride the short bus to school. */
4999 WARN(1, "Your BIOS is broken; DMA routed to ISOCH DMAR unit but no TLB space.\n"
5000 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
5001 dmi_get_system_info(DMI_BIOS_VENDOR),
5002 dmi_get_system_info(DMI_BIOS_VERSION),
5003 dmi_get_system_info(DMI_PRODUCT_VERSION));
5004 iommu_identity_mapping |= IDENTMAP_AZALIA;
5008 pr_warn("Recommended TLB entries for ISOCH unit is 16; your BIOS set %d\n",