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>
38 #include <linux/iova.h>
39 #include <linux/iommu.h>
40 #include <linux/intel-iommu.h>
41 #include <linux/syscore_ops.h>
42 #include <linux/tboot.h>
43 #include <linux/dmi.h>
44 #include <linux/pci-ats.h>
45 #include <linux/memblock.h>
46 #include <linux/dma-contiguous.h>
47 #include <linux/crash_dump.h>
48 #include <asm/irq_remapping.h>
49 #include <asm/cacheflush.h>
50 #include <asm/iommu.h>
52 #include "irq_remapping.h"
54 #define ROOT_SIZE VTD_PAGE_SIZE
55 #define CONTEXT_SIZE VTD_PAGE_SIZE
57 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
58 #define IS_USB_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_SERIAL_USB)
59 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
60 #define IS_AZALIA(pdev) ((pdev)->vendor == 0x8086 && (pdev)->device == 0x3a3e)
62 #define IOAPIC_RANGE_START (0xfee00000)
63 #define IOAPIC_RANGE_END (0xfeefffff)
64 #define IOVA_START_ADDR (0x1000)
66 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
68 #define MAX_AGAW_WIDTH 64
69 #define MAX_AGAW_PFN_WIDTH (MAX_AGAW_WIDTH - VTD_PAGE_SHIFT)
71 #define __DOMAIN_MAX_PFN(gaw) ((((uint64_t)1) << (gaw-VTD_PAGE_SHIFT)) - 1)
72 #define __DOMAIN_MAX_ADDR(gaw) ((((uint64_t)1) << gaw) - 1)
74 /* We limit DOMAIN_MAX_PFN to fit in an unsigned long, and DOMAIN_MAX_ADDR
75 to match. That way, we can use 'unsigned long' for PFNs with impunity. */
76 #define DOMAIN_MAX_PFN(gaw) ((unsigned long) min_t(uint64_t, \
77 __DOMAIN_MAX_PFN(gaw), (unsigned long)-1))
78 #define DOMAIN_MAX_ADDR(gaw) (((uint64_t)__DOMAIN_MAX_PFN(gaw)) << VTD_PAGE_SHIFT)
80 /* IO virtual address start page frame number */
81 #define IOVA_START_PFN (1)
83 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
84 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
85 #define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
87 /* page table handling */
88 #define LEVEL_STRIDE (9)
89 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
92 * This bitmap is used to advertise the page sizes our hardware support
93 * to the IOMMU core, which will then use this information to split
94 * physically contiguous memory regions it is mapping into page sizes
97 * Traditionally the IOMMU core just handed us the mappings directly,
98 * after making sure the size is an order of a 4KiB page and that the
99 * mapping has natural alignment.
101 * To retain this behavior, we currently advertise that we support
102 * all page sizes that are an order of 4KiB.
104 * If at some point we'd like to utilize the IOMMU core's new behavior,
105 * we could change this to advertise the real page sizes we support.
107 #define INTEL_IOMMU_PGSIZES (~0xFFFUL)
109 static inline int agaw_to_level(int agaw)
114 static inline int agaw_to_width(int agaw)
116 return min_t(int, 30 + agaw * LEVEL_STRIDE, MAX_AGAW_WIDTH);
119 static inline int width_to_agaw(int width)
121 return DIV_ROUND_UP(width - 30, LEVEL_STRIDE);
124 static inline unsigned int level_to_offset_bits(int level)
126 return (level - 1) * LEVEL_STRIDE;
129 static inline int pfn_level_offset(unsigned long pfn, int level)
131 return (pfn >> level_to_offset_bits(level)) & LEVEL_MASK;
134 static inline unsigned long level_mask(int level)
136 return -1UL << level_to_offset_bits(level);
139 static inline unsigned long level_size(int level)
141 return 1UL << level_to_offset_bits(level);
144 static inline unsigned long align_to_level(unsigned long pfn, int level)
146 return (pfn + level_size(level) - 1) & level_mask(level);
149 static inline unsigned long lvl_to_nr_pages(unsigned int lvl)
151 return 1 << min_t(int, (lvl - 1) * LEVEL_STRIDE, MAX_AGAW_PFN_WIDTH);
154 /* VT-d pages must always be _smaller_ than MM pages. Otherwise things
155 are never going to work. */
156 static inline unsigned long dma_to_mm_pfn(unsigned long dma_pfn)
158 return dma_pfn >> (PAGE_SHIFT - VTD_PAGE_SHIFT);
161 static inline unsigned long mm_to_dma_pfn(unsigned long mm_pfn)
163 return mm_pfn << (PAGE_SHIFT - VTD_PAGE_SHIFT);
165 static inline unsigned long page_to_dma_pfn(struct page *pg)
167 return mm_to_dma_pfn(page_to_pfn(pg));
169 static inline unsigned long virt_to_dma_pfn(void *p)
171 return page_to_dma_pfn(virt_to_page(p));
174 /* global iommu list, set NULL for ignored DMAR units */
175 static struct intel_iommu **g_iommus;
177 static void __init check_tylersburg_isoch(void);
178 static int rwbf_quirk;
181 * set to 1 to panic kernel if can't successfully enable VT-d
182 * (used when kernel is launched w/ TXT)
184 static int force_on = 0;
189 * 12-63: Context Ptr (12 - (haw-1))
196 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
199 * Take a root_entry and return the Lower Context Table Pointer (LCTP)
202 static phys_addr_t root_entry_lctp(struct root_entry *re)
207 return re->lo & VTD_PAGE_MASK;
211 * Take a root_entry and return the Upper Context Table Pointer (UCTP)
214 static phys_addr_t root_entry_uctp(struct root_entry *re)
219 return re->hi & VTD_PAGE_MASK;
224 * 1: fault processing disable
225 * 2-3: translation type
226 * 12-63: address space root
232 struct context_entry {
237 static inline void context_clear_pasid_enable(struct context_entry *context)
239 context->lo &= ~(1ULL << 11);
242 static inline bool context_pasid_enabled(struct context_entry *context)
244 return !!(context->lo & (1ULL << 11));
247 static inline void context_set_copied(struct context_entry *context)
249 context->hi |= (1ull << 3);
252 static inline bool context_copied(struct context_entry *context)
254 return !!(context->hi & (1ULL << 3));
257 static inline bool __context_present(struct context_entry *context)
259 return (context->lo & 1);
262 static inline bool context_present(struct context_entry *context)
264 return context_pasid_enabled(context) ?
265 __context_present(context) :
266 __context_present(context) && !context_copied(context);
269 static inline void context_set_present(struct context_entry *context)
274 static inline void context_set_fault_enable(struct context_entry *context)
276 context->lo &= (((u64)-1) << 2) | 1;
279 static inline void context_set_translation_type(struct context_entry *context,
282 context->lo &= (((u64)-1) << 4) | 3;
283 context->lo |= (value & 3) << 2;
286 static inline void context_set_address_root(struct context_entry *context,
289 context->lo &= ~VTD_PAGE_MASK;
290 context->lo |= value & VTD_PAGE_MASK;
293 static inline void context_set_address_width(struct context_entry *context,
296 context->hi |= value & 7;
299 static inline void context_set_domain_id(struct context_entry *context,
302 context->hi |= (value & ((1 << 16) - 1)) << 8;
305 static inline int context_domain_id(struct context_entry *c)
307 return((c->hi >> 8) & 0xffff);
310 static inline void context_clear_entry(struct context_entry *context)
323 * 12-63: Host physcial address
329 static inline void dma_clear_pte(struct dma_pte *pte)
334 static inline u64 dma_pte_addr(struct dma_pte *pte)
337 return pte->val & VTD_PAGE_MASK;
339 /* Must have a full atomic 64-bit read */
340 return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
344 static inline bool dma_pte_present(struct dma_pte *pte)
346 return (pte->val & 3) != 0;
349 static inline bool dma_pte_superpage(struct dma_pte *pte)
351 return (pte->val & DMA_PTE_LARGE_PAGE);
354 static inline int first_pte_in_page(struct dma_pte *pte)
356 return !((unsigned long)pte & ~VTD_PAGE_MASK);
360 * This domain is a statically identity mapping domain.
361 * 1. This domain creats a static 1:1 mapping to all usable memory.
362 * 2. It maps to each iommu if successful.
363 * 3. Each iommu mapps to this domain if successful.
365 static struct dmar_domain *si_domain;
366 static int hw_pass_through = 1;
369 * Domain represents a virtual machine, more than one devices
370 * across iommus may be owned in one domain, e.g. kvm guest.
372 #define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 0)
374 /* si_domain contains mulitple devices */
375 #define DOMAIN_FLAG_STATIC_IDENTITY (1 << 1)
377 #define for_each_domain_iommu(idx, domain) \
378 for (idx = 0; idx < g_num_of_iommus; idx++) \
379 if (domain->iommu_refcnt[idx])
382 int nid; /* node id */
384 unsigned iommu_refcnt[DMAR_UNITS_SUPPORTED];
385 /* Refcount of devices per iommu */
388 u16 iommu_did[DMAR_UNITS_SUPPORTED];
389 /* Domain ids per IOMMU. Use u16 since
390 * domain ids are 16 bit wide according
391 * to VT-d spec, section 9.3 */
393 struct list_head devices; /* all devices' list */
394 struct iova_domain iovad; /* iova's that belong to this domain */
396 struct dma_pte *pgd; /* virtual address */
397 int gaw; /* max guest address width */
399 /* adjusted guest address width, 0 is level 2 30-bit */
402 int flags; /* flags to find out type of domain */
404 int iommu_coherency;/* indicate coherency of iommu access */
405 int iommu_snooping; /* indicate snooping control feature*/
406 int iommu_count; /* reference count of iommu */
407 int iommu_superpage;/* Level of superpages supported:
408 0 == 4KiB (no superpages), 1 == 2MiB,
409 2 == 1GiB, 3 == 512GiB, 4 == 1TiB */
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 u8 pasid_supported:3;
429 struct device *dev; /* it's NULL for PCIe-to-PCI bridge */
430 struct intel_iommu *iommu; /* IOMMU used by this device */
431 struct dmar_domain *domain; /* pointer to domain */
434 struct dmar_rmrr_unit {
435 struct list_head list; /* list of rmrr units */
436 struct acpi_dmar_header *hdr; /* ACPI header */
437 u64 base_address; /* reserved base address*/
438 u64 end_address; /* reserved end address */
439 struct dmar_dev_scope *devices; /* target devices */
440 int devices_cnt; /* target device count */
443 struct dmar_atsr_unit {
444 struct list_head list; /* list of ATSR units */
445 struct acpi_dmar_header *hdr; /* ACPI header */
446 struct dmar_dev_scope *devices; /* target devices */
447 int devices_cnt; /* target device count */
448 u8 include_all:1; /* include all ports */
451 static LIST_HEAD(dmar_atsr_units);
452 static LIST_HEAD(dmar_rmrr_units);
454 #define for_each_rmrr_units(rmrr) \
455 list_for_each_entry(rmrr, &dmar_rmrr_units, list)
457 static void flush_unmaps_timeout(unsigned long data);
459 static DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
461 #define HIGH_WATER_MARK 250
462 struct deferred_flush_tables {
464 struct iova *iova[HIGH_WATER_MARK];
465 struct dmar_domain *domain[HIGH_WATER_MARK];
466 struct page *freelist[HIGH_WATER_MARK];
469 static struct deferred_flush_tables *deferred_flush;
471 /* bitmap for indexing intel_iommus */
472 static int g_num_of_iommus;
474 static DEFINE_SPINLOCK(async_umap_flush_lock);
475 static LIST_HEAD(unmaps_to_do);
478 static long list_size;
480 static void domain_exit(struct dmar_domain *domain);
481 static void domain_remove_dev_info(struct dmar_domain *domain);
482 static void dmar_remove_one_dev_info(struct dmar_domain *domain,
484 static void __dmar_remove_one_dev_info(struct device_domain_info *info);
485 static void domain_context_clear(struct intel_iommu *iommu,
487 static int domain_detach_iommu(struct dmar_domain *domain,
488 struct intel_iommu *iommu);
490 #ifdef CONFIG_INTEL_IOMMU_DEFAULT_ON
491 int dmar_disabled = 0;
493 int dmar_disabled = 1;
494 #endif /*CONFIG_INTEL_IOMMU_DEFAULT_ON*/
496 int intel_iommu_enabled = 0;
497 EXPORT_SYMBOL_GPL(intel_iommu_enabled);
499 static int dmar_map_gfx = 1;
500 static int dmar_forcedac;
501 static int intel_iommu_strict;
502 static int intel_iommu_superpage = 1;
503 static int intel_iommu_ecs = 1;
504 static int intel_iommu_pasid28;
505 static int iommu_identity_mapping;
507 #define IDENTMAP_ALL 1
508 #define IDENTMAP_GFX 2
509 #define IDENTMAP_AZALIA 4
511 /* Broadwell and Skylake have broken ECS support — normal so-called "second
512 * level" translation of DMA requests-without-PASID doesn't actually happen
513 * unless you also set the NESTE bit in an extended context-entry. Which of
514 * course means that SVM doesn't work because it's trying to do nested
515 * translation of the physical addresses it finds in the process page tables,
516 * through the IOVA->phys mapping found in the "second level" page tables.
518 * The VT-d specification was retroactively changed to change the definition
519 * of the capability bits and pretend that Broadwell/Skylake never happened...
520 * but unfortunately the wrong bit was changed. It's ECS which is broken, but
521 * for some reason it was the PASID capability bit which was redefined (from
522 * bit 28 on BDW/SKL to bit 40 in future).
524 * So our test for ECS needs to eschew those implementations which set the old
525 * PASID capabiity bit 28, since those are the ones on which ECS is broken.
526 * Unless we are working around the 'pasid28' limitations, that is, by putting
527 * the device into passthrough mode for normal DMA and thus masking the bug.
529 #define ecs_enabled(iommu) (intel_iommu_ecs && ecap_ecs(iommu->ecap) && \
530 (intel_iommu_pasid28 || !ecap_broken_pasid(iommu->ecap)))
531 /* PASID support is thus enabled if ECS is enabled and *either* of the old
532 * or new capability bits are set. */
533 #define pasid_enabled(iommu) (ecs_enabled(iommu) && \
534 (ecap_pasid(iommu->ecap) || ecap_broken_pasid(iommu->ecap)))
536 int intel_iommu_gfx_mapped;
537 EXPORT_SYMBOL_GPL(intel_iommu_gfx_mapped);
539 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
540 static DEFINE_SPINLOCK(device_domain_lock);
541 static LIST_HEAD(device_domain_list);
543 static const struct iommu_ops intel_iommu_ops;
545 static bool translation_pre_enabled(struct intel_iommu *iommu)
547 return (iommu->flags & VTD_FLAG_TRANS_PRE_ENABLED);
550 static void clear_translation_pre_enabled(struct intel_iommu *iommu)
552 iommu->flags &= ~VTD_FLAG_TRANS_PRE_ENABLED;
555 static void init_translation_status(struct intel_iommu *iommu)
559 gsts = readl(iommu->reg + DMAR_GSTS_REG);
560 if (gsts & DMA_GSTS_TES)
561 iommu->flags |= VTD_FLAG_TRANS_PRE_ENABLED;
564 /* Convert generic 'struct iommu_domain to private struct dmar_domain */
565 static struct dmar_domain *to_dmar_domain(struct iommu_domain *dom)
567 return container_of(dom, struct dmar_domain, domain);
570 static int __init intel_iommu_setup(char *str)
575 if (!strncmp(str, "on", 2)) {
577 pr_info("IOMMU enabled\n");
578 } else if (!strncmp(str, "off", 3)) {
580 pr_info("IOMMU disabled\n");
581 } else if (!strncmp(str, "igfx_off", 8)) {
583 pr_info("Disable GFX device mapping\n");
584 } else if (!strncmp(str, "forcedac", 8)) {
585 pr_info("Forcing DAC for PCI devices\n");
587 } else if (!strncmp(str, "strict", 6)) {
588 pr_info("Disable batched IOTLB flush\n");
589 intel_iommu_strict = 1;
590 } else if (!strncmp(str, "sp_off", 6)) {
591 pr_info("Disable supported super page\n");
592 intel_iommu_superpage = 0;
593 } else if (!strncmp(str, "ecs_off", 7)) {
595 "Intel-IOMMU: disable extended context table support\n");
597 } else if (!strncmp(str, "pasid28", 7)) {
599 "Intel-IOMMU: enable pre-production PASID support\n");
600 intel_iommu_pasid28 = 1;
601 iommu_identity_mapping |= IDENTMAP_GFX;
604 str += strcspn(str, ",");
610 __setup("intel_iommu=", intel_iommu_setup);
612 static struct kmem_cache *iommu_domain_cache;
613 static struct kmem_cache *iommu_devinfo_cache;
615 static struct dmar_domain* get_iommu_domain(struct intel_iommu *iommu, u16 did)
617 struct dmar_domain **domains;
620 domains = iommu->domains[idx];
624 return domains[did & 0xff];
627 static void set_iommu_domain(struct intel_iommu *iommu, u16 did,
628 struct dmar_domain *domain)
630 struct dmar_domain **domains;
633 if (!iommu->domains[idx]) {
634 size_t size = 256 * sizeof(struct dmar_domain *);
635 iommu->domains[idx] = kzalloc(size, GFP_ATOMIC);
638 domains = iommu->domains[idx];
639 if (WARN_ON(!domains))
642 domains[did & 0xff] = domain;
645 static inline void *alloc_pgtable_page(int node)
650 page = alloc_pages_node(node, GFP_ATOMIC | __GFP_ZERO, 0);
652 vaddr = page_address(page);
656 static inline void free_pgtable_page(void *vaddr)
658 free_page((unsigned long)vaddr);
661 static inline void *alloc_domain_mem(void)
663 return kmem_cache_alloc(iommu_domain_cache, GFP_ATOMIC);
666 static void free_domain_mem(void *vaddr)
668 kmem_cache_free(iommu_domain_cache, vaddr);
671 static inline void * alloc_devinfo_mem(void)
673 return kmem_cache_alloc(iommu_devinfo_cache, GFP_ATOMIC);
676 static inline void free_devinfo_mem(void *vaddr)
678 kmem_cache_free(iommu_devinfo_cache, vaddr);
681 static inline int domain_type_is_vm(struct dmar_domain *domain)
683 return domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE;
686 static inline int domain_type_is_si(struct dmar_domain *domain)
688 return domain->flags & DOMAIN_FLAG_STATIC_IDENTITY;
691 static inline int domain_type_is_vm_or_si(struct dmar_domain *domain)
693 return domain->flags & (DOMAIN_FLAG_VIRTUAL_MACHINE |
694 DOMAIN_FLAG_STATIC_IDENTITY);
697 static inline int domain_pfn_supported(struct dmar_domain *domain,
700 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
702 return !(addr_width < BITS_PER_LONG && pfn >> addr_width);
705 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
710 sagaw = cap_sagaw(iommu->cap);
711 for (agaw = width_to_agaw(max_gaw);
713 if (test_bit(agaw, &sagaw))
721 * Calculate max SAGAW for each iommu.
723 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
725 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
729 * calculate agaw for each iommu.
730 * "SAGAW" may be different across iommus, use a default agaw, and
731 * get a supported less agaw for iommus that don't support the default agaw.
733 int iommu_calculate_agaw(struct intel_iommu *iommu)
735 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
738 /* This functionin only returns single iommu in a domain */
739 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
743 /* si_domain and vm domain should not get here. */
744 BUG_ON(domain_type_is_vm_or_si(domain));
745 for_each_domain_iommu(iommu_id, domain)
748 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
751 return g_iommus[iommu_id];
754 static void domain_update_iommu_coherency(struct dmar_domain *domain)
756 struct dmar_drhd_unit *drhd;
757 struct intel_iommu *iommu;
761 domain->iommu_coherency = 1;
763 for_each_domain_iommu(i, domain) {
765 if (!ecap_coherent(g_iommus[i]->ecap)) {
766 domain->iommu_coherency = 0;
773 /* No hardware attached; use lowest common denominator */
775 for_each_active_iommu(iommu, drhd) {
776 if (!ecap_coherent(iommu->ecap)) {
777 domain->iommu_coherency = 0;
784 static int domain_update_iommu_snooping(struct intel_iommu *skip)
786 struct dmar_drhd_unit *drhd;
787 struct intel_iommu *iommu;
791 for_each_active_iommu(iommu, drhd) {
793 if (!ecap_sc_support(iommu->ecap)) {
804 static int domain_update_iommu_superpage(struct intel_iommu *skip)
806 struct dmar_drhd_unit *drhd;
807 struct intel_iommu *iommu;
810 if (!intel_iommu_superpage) {
814 /* set iommu_superpage to the smallest common denominator */
816 for_each_active_iommu(iommu, drhd) {
818 mask &= cap_super_page_val(iommu->cap);
828 /* Some capabilities may be different across iommus */
829 static void domain_update_iommu_cap(struct dmar_domain *domain)
831 domain_update_iommu_coherency(domain);
832 domain->iommu_snooping = domain_update_iommu_snooping(NULL);
833 domain->iommu_superpage = domain_update_iommu_superpage(NULL);
836 static inline struct context_entry *iommu_context_addr(struct intel_iommu *iommu,
837 u8 bus, u8 devfn, int alloc)
839 struct root_entry *root = &iommu->root_entry[bus];
840 struct context_entry *context;
844 if (ecs_enabled(iommu)) {
852 context = phys_to_virt(*entry & VTD_PAGE_MASK);
854 unsigned long phy_addr;
858 context = alloc_pgtable_page(iommu->node);
862 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
863 phy_addr = virt_to_phys((void *)context);
864 *entry = phy_addr | 1;
865 __iommu_flush_cache(iommu, entry, sizeof(*entry));
867 return &context[devfn];
870 static int iommu_dummy(struct device *dev)
872 return dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
875 static struct intel_iommu *device_to_iommu(struct device *dev, u8 *bus, u8 *devfn)
877 struct dmar_drhd_unit *drhd = NULL;
878 struct intel_iommu *iommu;
880 struct pci_dev *ptmp, *pdev = NULL;
884 if (iommu_dummy(dev))
887 if (dev_is_pci(dev)) {
888 pdev = to_pci_dev(dev);
889 segment = pci_domain_nr(pdev->bus);
890 } else if (has_acpi_companion(dev))
891 dev = &ACPI_COMPANION(dev)->dev;
894 for_each_active_iommu(iommu, drhd) {
895 if (pdev && segment != drhd->segment)
898 for_each_active_dev_scope(drhd->devices,
899 drhd->devices_cnt, i, tmp) {
901 *bus = drhd->devices[i].bus;
902 *devfn = drhd->devices[i].devfn;
906 if (!pdev || !dev_is_pci(tmp))
909 ptmp = to_pci_dev(tmp);
910 if (ptmp->subordinate &&
911 ptmp->subordinate->number <= pdev->bus->number &&
912 ptmp->subordinate->busn_res.end >= pdev->bus->number)
916 if (pdev && drhd->include_all) {
918 *bus = pdev->bus->number;
919 *devfn = pdev->devfn;
930 static void domain_flush_cache(struct dmar_domain *domain,
931 void *addr, int size)
933 if (!domain->iommu_coherency)
934 clflush_cache_range(addr, size);
937 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
939 struct context_entry *context;
943 spin_lock_irqsave(&iommu->lock, flags);
944 context = iommu_context_addr(iommu, bus, devfn, 0);
946 ret = context_present(context);
947 spin_unlock_irqrestore(&iommu->lock, flags);
951 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
953 struct context_entry *context;
956 spin_lock_irqsave(&iommu->lock, flags);
957 context = iommu_context_addr(iommu, bus, devfn, 0);
959 context_clear_entry(context);
960 __iommu_flush_cache(iommu, context, sizeof(*context));
962 spin_unlock_irqrestore(&iommu->lock, flags);
965 static void free_context_table(struct intel_iommu *iommu)
969 struct context_entry *context;
971 spin_lock_irqsave(&iommu->lock, flags);
972 if (!iommu->root_entry) {
975 for (i = 0; i < ROOT_ENTRY_NR; i++) {
976 context = iommu_context_addr(iommu, i, 0, 0);
978 free_pgtable_page(context);
980 if (!ecs_enabled(iommu))
983 context = iommu_context_addr(iommu, i, 0x80, 0);
985 free_pgtable_page(context);
988 free_pgtable_page(iommu->root_entry);
989 iommu->root_entry = NULL;
991 spin_unlock_irqrestore(&iommu->lock, flags);
994 static struct dma_pte *pfn_to_dma_pte(struct dmar_domain *domain,
995 unsigned long pfn, int *target_level)
997 struct dma_pte *parent, *pte = NULL;
998 int level = agaw_to_level(domain->agaw);
1001 BUG_ON(!domain->pgd);
1003 if (!domain_pfn_supported(domain, pfn))
1004 /* Address beyond IOMMU's addressing capabilities. */
1007 parent = domain->pgd;
1012 offset = pfn_level_offset(pfn, level);
1013 pte = &parent[offset];
1014 if (!*target_level && (dma_pte_superpage(pte) || !dma_pte_present(pte)))
1016 if (level == *target_level)
1019 if (!dma_pte_present(pte)) {
1022 tmp_page = alloc_pgtable_page(domain->nid);
1027 domain_flush_cache(domain, tmp_page, VTD_PAGE_SIZE);
1028 pteval = ((uint64_t)virt_to_dma_pfn(tmp_page) << VTD_PAGE_SHIFT) | DMA_PTE_READ | DMA_PTE_WRITE;
1029 if (cmpxchg64(&pte->val, 0ULL, pteval))
1030 /* Someone else set it while we were thinking; use theirs. */
1031 free_pgtable_page(tmp_page);
1033 domain_flush_cache(domain, pte, sizeof(*pte));
1038 parent = phys_to_virt(dma_pte_addr(pte));
1043 *target_level = level;
1049 /* return address's pte at specific level */
1050 static struct dma_pte *dma_pfn_level_pte(struct dmar_domain *domain,
1052 int level, int *large_page)
1054 struct dma_pte *parent, *pte = NULL;
1055 int total = agaw_to_level(domain->agaw);
1058 parent = domain->pgd;
1059 while (level <= total) {
1060 offset = pfn_level_offset(pfn, total);
1061 pte = &parent[offset];
1065 if (!dma_pte_present(pte)) {
1066 *large_page = total;
1070 if (dma_pte_superpage(pte)) {
1071 *large_page = total;
1075 parent = phys_to_virt(dma_pte_addr(pte));
1081 /* clear last level pte, a tlb flush should be followed */
1082 static void dma_pte_clear_range(struct dmar_domain *domain,
1083 unsigned long start_pfn,
1084 unsigned long last_pfn)
1086 unsigned int large_page = 1;
1087 struct dma_pte *first_pte, *pte;
1089 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1090 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1091 BUG_ON(start_pfn > last_pfn);
1093 /* we don't need lock here; nobody else touches the iova range */
1096 first_pte = pte = dma_pfn_level_pte(domain, start_pfn, 1, &large_page);
1098 start_pfn = align_to_level(start_pfn + 1, large_page + 1);
1103 start_pfn += lvl_to_nr_pages(large_page);
1105 } while (start_pfn <= last_pfn && !first_pte_in_page(pte));
1107 domain_flush_cache(domain, first_pte,
1108 (void *)pte - (void *)first_pte);
1110 } while (start_pfn && start_pfn <= last_pfn);
1113 static void dma_pte_free_level(struct dmar_domain *domain, int level,
1114 struct dma_pte *pte, unsigned long pfn,
1115 unsigned long start_pfn, unsigned long last_pfn)
1117 pfn = max(start_pfn, pfn);
1118 pte = &pte[pfn_level_offset(pfn, level)];
1121 unsigned long level_pfn;
1122 struct dma_pte *level_pte;
1124 if (!dma_pte_present(pte) || dma_pte_superpage(pte))
1127 level_pfn = pfn & level_mask(level - 1);
1128 level_pte = phys_to_virt(dma_pte_addr(pte));
1131 dma_pte_free_level(domain, level - 1, level_pte,
1132 level_pfn, start_pfn, last_pfn);
1134 /* If range covers entire pagetable, free it */
1135 if (!(start_pfn > level_pfn ||
1136 last_pfn < level_pfn + level_size(level) - 1)) {
1138 domain_flush_cache(domain, pte, sizeof(*pte));
1139 free_pgtable_page(level_pte);
1142 pfn += level_size(level);
1143 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
1146 /* free page table pages. last level pte should already be cleared */
1147 static void dma_pte_free_pagetable(struct dmar_domain *domain,
1148 unsigned long start_pfn,
1149 unsigned long last_pfn)
1151 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1152 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1153 BUG_ON(start_pfn > last_pfn);
1155 dma_pte_clear_range(domain, start_pfn, last_pfn);
1157 /* We don't need lock here; nobody else touches the iova range */
1158 dma_pte_free_level(domain, agaw_to_level(domain->agaw),
1159 domain->pgd, 0, start_pfn, last_pfn);
1162 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
1163 free_pgtable_page(domain->pgd);
1168 /* When a page at a given level is being unlinked from its parent, we don't
1169 need to *modify* it at all. All we need to do is make a list of all the
1170 pages which can be freed just as soon as we've flushed the IOTLB and we
1171 know the hardware page-walk will no longer touch them.
1172 The 'pte' argument is the *parent* PTE, pointing to the page that is to
1174 static struct page *dma_pte_list_pagetables(struct dmar_domain *domain,
1175 int level, struct dma_pte *pte,
1176 struct page *freelist)
1180 pg = pfn_to_page(dma_pte_addr(pte) >> PAGE_SHIFT);
1181 pg->freelist = freelist;
1187 pte = page_address(pg);
1189 if (dma_pte_present(pte) && !dma_pte_superpage(pte))
1190 freelist = dma_pte_list_pagetables(domain, level - 1,
1193 } while (!first_pte_in_page(pte));
1198 static struct page *dma_pte_clear_level(struct dmar_domain *domain, int level,
1199 struct dma_pte *pte, unsigned long pfn,
1200 unsigned long start_pfn,
1201 unsigned long last_pfn,
1202 struct page *freelist)
1204 struct dma_pte *first_pte = NULL, *last_pte = NULL;
1206 pfn = max(start_pfn, pfn);
1207 pte = &pte[pfn_level_offset(pfn, level)];
1210 unsigned long level_pfn;
1212 if (!dma_pte_present(pte))
1215 level_pfn = pfn & level_mask(level);
1217 /* If range covers entire pagetable, free it */
1218 if (start_pfn <= level_pfn &&
1219 last_pfn >= level_pfn + level_size(level) - 1) {
1220 /* These suborbinate page tables are going away entirely. Don't
1221 bother to clear them; we're just going to *free* them. */
1222 if (level > 1 && !dma_pte_superpage(pte))
1223 freelist = dma_pte_list_pagetables(domain, level - 1, pte, freelist);
1229 } else if (level > 1) {
1230 /* Recurse down into a level that isn't *entirely* obsolete */
1231 freelist = dma_pte_clear_level(domain, level - 1,
1232 phys_to_virt(dma_pte_addr(pte)),
1233 level_pfn, start_pfn, last_pfn,
1237 pfn += level_size(level);
1238 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
1241 domain_flush_cache(domain, first_pte,
1242 (void *)++last_pte - (void *)first_pte);
1247 /* We can't just free the pages because the IOMMU may still be walking
1248 the page tables, and may have cached the intermediate levels. The
1249 pages can only be freed after the IOTLB flush has been done. */
1250 static struct page *domain_unmap(struct dmar_domain *domain,
1251 unsigned long start_pfn,
1252 unsigned long last_pfn)
1254 struct page *freelist = NULL;
1256 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1257 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1258 BUG_ON(start_pfn > last_pfn);
1260 /* we don't need lock here; nobody else touches the iova range */
1261 freelist = dma_pte_clear_level(domain, agaw_to_level(domain->agaw),
1262 domain->pgd, 0, start_pfn, last_pfn, NULL);
1265 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
1266 struct page *pgd_page = virt_to_page(domain->pgd);
1267 pgd_page->freelist = freelist;
1268 freelist = pgd_page;
1276 static void dma_free_pagelist(struct page *freelist)
1280 while ((pg = freelist)) {
1281 freelist = pg->freelist;
1282 free_pgtable_page(page_address(pg));
1286 /* iommu handling */
1287 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
1289 struct root_entry *root;
1290 unsigned long flags;
1292 root = (struct root_entry *)alloc_pgtable_page(iommu->node);
1294 pr_err("Allocating root entry for %s failed\n",
1299 __iommu_flush_cache(iommu, root, ROOT_SIZE);
1301 spin_lock_irqsave(&iommu->lock, flags);
1302 iommu->root_entry = root;
1303 spin_unlock_irqrestore(&iommu->lock, flags);
1308 static void iommu_set_root_entry(struct intel_iommu *iommu)
1314 addr = virt_to_phys(iommu->root_entry);
1315 if (ecs_enabled(iommu))
1316 addr |= DMA_RTADDR_RTT;
1318 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1319 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, addr);
1321 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
1323 /* Make sure hardware complete it */
1324 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1325 readl, (sts & DMA_GSTS_RTPS), sts);
1327 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1330 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
1335 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
1338 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1339 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
1341 /* Make sure hardware complete it */
1342 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1343 readl, (!(val & DMA_GSTS_WBFS)), val);
1345 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1348 /* return value determine if we need a write buffer flush */
1349 static void __iommu_flush_context(struct intel_iommu *iommu,
1350 u16 did, u16 source_id, u8 function_mask,
1357 case DMA_CCMD_GLOBAL_INVL:
1358 val = DMA_CCMD_GLOBAL_INVL;
1360 case DMA_CCMD_DOMAIN_INVL:
1361 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
1363 case DMA_CCMD_DEVICE_INVL:
1364 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
1365 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
1370 val |= DMA_CCMD_ICC;
1372 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1373 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
1375 /* Make sure hardware complete it */
1376 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
1377 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
1379 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1382 /* return value determine if we need a write buffer flush */
1383 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
1384 u64 addr, unsigned int size_order, u64 type)
1386 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
1387 u64 val = 0, val_iva = 0;
1391 case DMA_TLB_GLOBAL_FLUSH:
1392 /* global flush doesn't need set IVA_REG */
1393 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
1395 case DMA_TLB_DSI_FLUSH:
1396 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1398 case DMA_TLB_PSI_FLUSH:
1399 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1400 /* IH bit is passed in as part of address */
1401 val_iva = size_order | addr;
1406 /* Note: set drain read/write */
1409 * This is probably to be super secure.. Looks like we can
1410 * ignore it without any impact.
1412 if (cap_read_drain(iommu->cap))
1413 val |= DMA_TLB_READ_DRAIN;
1415 if (cap_write_drain(iommu->cap))
1416 val |= DMA_TLB_WRITE_DRAIN;
1418 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1419 /* Note: Only uses first TLB reg currently */
1421 dmar_writeq(iommu->reg + tlb_offset, val_iva);
1422 dmar_writeq(iommu->reg + tlb_offset + 8, val);
1424 /* Make sure hardware complete it */
1425 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
1426 dmar_readq, (!(val & DMA_TLB_IVT)), val);
1428 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1430 /* check IOTLB invalidation granularity */
1431 if (DMA_TLB_IAIG(val) == 0)
1432 pr_err("Flush IOTLB failed\n");
1433 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
1434 pr_debug("TLB flush request %Lx, actual %Lx\n",
1435 (unsigned long long)DMA_TLB_IIRG(type),
1436 (unsigned long long)DMA_TLB_IAIG(val));
1439 static struct device_domain_info *
1440 iommu_support_dev_iotlb (struct dmar_domain *domain, struct intel_iommu *iommu,
1443 struct device_domain_info *info;
1445 assert_spin_locked(&device_domain_lock);
1450 list_for_each_entry(info, &domain->devices, link)
1451 if (info->iommu == iommu && info->bus == bus &&
1452 info->devfn == devfn) {
1453 if (info->ats_supported && info->dev)
1461 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
1463 struct pci_dev *pdev;
1465 if (!info || !dev_is_pci(info->dev))
1468 pdev = to_pci_dev(info->dev);
1470 #ifdef CONFIG_INTEL_IOMMU_SVM
1471 /* The PCIe spec, in its wisdom, declares that the behaviour of
1472 the device if you enable PASID support after ATS support is
1473 undefined. So always enable PASID support on devices which
1474 have it, even if we can't yet know if we're ever going to
1476 if (info->pasid_supported && !pci_enable_pasid(pdev, info->pasid_supported & ~1))
1477 info->pasid_enabled = 1;
1479 if (info->pri_supported && !pci_reset_pri(pdev) && !pci_enable_pri(pdev, 32))
1480 info->pri_enabled = 1;
1482 if (info->ats_supported && !pci_enable_ats(pdev, VTD_PAGE_SHIFT)) {
1483 info->ats_enabled = 1;
1484 info->ats_qdep = pci_ats_queue_depth(pdev);
1488 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1490 struct pci_dev *pdev;
1492 if (!dev_is_pci(info->dev))
1495 pdev = to_pci_dev(info->dev);
1497 if (info->ats_enabled) {
1498 pci_disable_ats(pdev);
1499 info->ats_enabled = 0;
1501 #ifdef CONFIG_INTEL_IOMMU_SVM
1502 if (info->pri_enabled) {
1503 pci_disable_pri(pdev);
1504 info->pri_enabled = 0;
1506 if (info->pasid_enabled) {
1507 pci_disable_pasid(pdev);
1508 info->pasid_enabled = 0;
1513 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1514 u64 addr, unsigned mask)
1517 unsigned long flags;
1518 struct device_domain_info *info;
1520 spin_lock_irqsave(&device_domain_lock, flags);
1521 list_for_each_entry(info, &domain->devices, link) {
1522 if (!info->ats_enabled)
1525 sid = info->bus << 8 | info->devfn;
1526 qdep = info->ats_qdep;
1527 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1529 spin_unlock_irqrestore(&device_domain_lock, flags);
1532 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu,
1533 struct dmar_domain *domain,
1534 unsigned long pfn, unsigned int pages,
1537 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1538 uint64_t addr = (uint64_t)pfn << VTD_PAGE_SHIFT;
1539 u16 did = domain->iommu_did[iommu->seq_id];
1546 * Fallback to domain selective flush if no PSI support or the size is
1548 * PSI requires page size to be 2 ^ x, and the base address is naturally
1549 * aligned to the size
1551 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1552 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1555 iommu->flush.flush_iotlb(iommu, did, addr | ih, mask,
1559 * In caching mode, changes of pages from non-present to present require
1560 * flush. However, device IOTLB doesn't need to be flushed in this case.
1562 if (!cap_caching_mode(iommu->cap) || !map)
1563 iommu_flush_dev_iotlb(get_iommu_domain(iommu, did),
1567 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1570 unsigned long flags;
1572 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1573 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1574 pmen &= ~DMA_PMEN_EPM;
1575 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1577 /* wait for the protected region status bit to clear */
1578 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1579 readl, !(pmen & DMA_PMEN_PRS), pmen);
1581 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1584 static void iommu_enable_translation(struct intel_iommu *iommu)
1587 unsigned long flags;
1589 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1590 iommu->gcmd |= DMA_GCMD_TE;
1591 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1593 /* Make sure hardware complete it */
1594 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1595 readl, (sts & DMA_GSTS_TES), sts);
1597 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1600 static void iommu_disable_translation(struct intel_iommu *iommu)
1605 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1606 iommu->gcmd &= ~DMA_GCMD_TE;
1607 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1609 /* Make sure hardware complete it */
1610 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1611 readl, (!(sts & DMA_GSTS_TES)), sts);
1613 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1617 static int iommu_init_domains(struct intel_iommu *iommu)
1619 u32 ndomains, nlongs;
1622 ndomains = cap_ndoms(iommu->cap);
1623 pr_debug("%s: Number of Domains supported <%d>\n",
1624 iommu->name, ndomains);
1625 nlongs = BITS_TO_LONGS(ndomains);
1627 spin_lock_init(&iommu->lock);
1629 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1630 if (!iommu->domain_ids) {
1631 pr_err("%s: Allocating domain id array failed\n",
1636 size = ((ndomains >> 8) + 1) * sizeof(struct dmar_domain **);
1637 iommu->domains = kzalloc(size, GFP_KERNEL);
1639 if (iommu->domains) {
1640 size = 256 * sizeof(struct dmar_domain *);
1641 iommu->domains[0] = kzalloc(size, GFP_KERNEL);
1644 if (!iommu->domains || !iommu->domains[0]) {
1645 pr_err("%s: Allocating domain array failed\n",
1647 kfree(iommu->domain_ids);
1648 kfree(iommu->domains);
1649 iommu->domain_ids = NULL;
1650 iommu->domains = NULL;
1657 * If Caching mode is set, then invalid translations are tagged
1658 * with domain-id 0, hence we need to pre-allocate it. We also
1659 * use domain-id 0 as a marker for non-allocated domain-id, so
1660 * make sure it is not used for a real domain.
1662 set_bit(0, iommu->domain_ids);
1667 static void disable_dmar_iommu(struct intel_iommu *iommu)
1669 struct device_domain_info *info, *tmp;
1670 unsigned long flags;
1672 if (!iommu->domains || !iommu->domain_ids)
1676 spin_lock_irqsave(&device_domain_lock, flags);
1677 list_for_each_entry_safe(info, tmp, &device_domain_list, global) {
1678 struct dmar_domain *domain;
1680 if (info->iommu != iommu)
1683 if (!info->dev || !info->domain)
1686 domain = info->domain;
1688 __dmar_remove_one_dev_info(info);
1690 if (!domain_type_is_vm_or_si(domain)) {
1692 * The domain_exit() function can't be called under
1693 * device_domain_lock, as it takes this lock itself.
1694 * So release the lock here and re-run the loop
1697 spin_unlock_irqrestore(&device_domain_lock, flags);
1698 domain_exit(domain);
1702 spin_unlock_irqrestore(&device_domain_lock, flags);
1704 if (iommu->gcmd & DMA_GCMD_TE)
1705 iommu_disable_translation(iommu);
1708 static void free_dmar_iommu(struct intel_iommu *iommu)
1710 if ((iommu->domains) && (iommu->domain_ids)) {
1711 int elems = (cap_ndoms(iommu->cap) >> 8) + 1;
1714 for (i = 0; i < elems; i++)
1715 kfree(iommu->domains[i]);
1716 kfree(iommu->domains);
1717 kfree(iommu->domain_ids);
1718 iommu->domains = NULL;
1719 iommu->domain_ids = NULL;
1722 g_iommus[iommu->seq_id] = NULL;
1724 /* free context mapping */
1725 free_context_table(iommu);
1727 #ifdef CONFIG_INTEL_IOMMU_SVM
1728 if (pasid_enabled(iommu)) {
1729 if (ecap_prs(iommu->ecap))
1730 intel_svm_finish_prq(iommu);
1731 intel_svm_free_pasid_tables(iommu);
1736 static struct dmar_domain *alloc_domain(int flags)
1738 struct dmar_domain *domain;
1740 domain = alloc_domain_mem();
1744 memset(domain, 0, sizeof(*domain));
1746 domain->flags = flags;
1747 INIT_LIST_HEAD(&domain->devices);
1752 /* Must be called with iommu->lock */
1753 static int domain_attach_iommu(struct dmar_domain *domain,
1754 struct intel_iommu *iommu)
1756 unsigned long ndomains;
1759 assert_spin_locked(&device_domain_lock);
1760 assert_spin_locked(&iommu->lock);
1762 domain->iommu_refcnt[iommu->seq_id] += 1;
1763 domain->iommu_count += 1;
1764 if (domain->iommu_refcnt[iommu->seq_id] == 1) {
1765 ndomains = cap_ndoms(iommu->cap);
1766 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1768 if (num >= ndomains) {
1769 pr_err("%s: No free domain ids\n", iommu->name);
1770 domain->iommu_refcnt[iommu->seq_id] -= 1;
1771 domain->iommu_count -= 1;
1775 set_bit(num, iommu->domain_ids);
1776 set_iommu_domain(iommu, num, domain);
1778 domain->iommu_did[iommu->seq_id] = num;
1779 domain->nid = iommu->node;
1781 domain_update_iommu_cap(domain);
1787 static int domain_detach_iommu(struct dmar_domain *domain,
1788 struct intel_iommu *iommu)
1790 int num, count = INT_MAX;
1792 assert_spin_locked(&device_domain_lock);
1793 assert_spin_locked(&iommu->lock);
1795 domain->iommu_refcnt[iommu->seq_id] -= 1;
1796 count = --domain->iommu_count;
1797 if (domain->iommu_refcnt[iommu->seq_id] == 0) {
1798 num = domain->iommu_did[iommu->seq_id];
1799 clear_bit(num, iommu->domain_ids);
1800 set_iommu_domain(iommu, num, NULL);
1802 domain_update_iommu_cap(domain);
1803 domain->iommu_did[iommu->seq_id] = 0;
1809 static struct iova_domain reserved_iova_list;
1810 static struct lock_class_key reserved_rbtree_key;
1812 static int dmar_init_reserved_ranges(void)
1814 struct pci_dev *pdev = NULL;
1818 init_iova_domain(&reserved_iova_list, VTD_PAGE_SIZE, IOVA_START_PFN,
1821 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1822 &reserved_rbtree_key);
1824 /* IOAPIC ranges shouldn't be accessed by DMA */
1825 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1826 IOVA_PFN(IOAPIC_RANGE_END));
1828 pr_err("Reserve IOAPIC range failed\n");
1832 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1833 for_each_pci_dev(pdev) {
1836 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1837 r = &pdev->resource[i];
1838 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1840 iova = reserve_iova(&reserved_iova_list,
1844 pr_err("Reserve iova failed\n");
1852 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1854 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1857 static inline int guestwidth_to_adjustwidth(int gaw)
1860 int r = (gaw - 12) % 9;
1871 static int domain_init(struct dmar_domain *domain, struct intel_iommu *iommu,
1874 int adjust_width, agaw;
1875 unsigned long sagaw;
1877 init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
1879 domain_reserve_special_ranges(domain);
1881 /* calculate AGAW */
1882 if (guest_width > cap_mgaw(iommu->cap))
1883 guest_width = cap_mgaw(iommu->cap);
1884 domain->gaw = guest_width;
1885 adjust_width = guestwidth_to_adjustwidth(guest_width);
1886 agaw = width_to_agaw(adjust_width);
1887 sagaw = cap_sagaw(iommu->cap);
1888 if (!test_bit(agaw, &sagaw)) {
1889 /* hardware doesn't support it, choose a bigger one */
1890 pr_debug("Hardware doesn't support agaw %d\n", agaw);
1891 agaw = find_next_bit(&sagaw, 5, agaw);
1895 domain->agaw = agaw;
1897 if (ecap_coherent(iommu->ecap))
1898 domain->iommu_coherency = 1;
1900 domain->iommu_coherency = 0;
1902 if (ecap_sc_support(iommu->ecap))
1903 domain->iommu_snooping = 1;
1905 domain->iommu_snooping = 0;
1907 if (intel_iommu_superpage)
1908 domain->iommu_superpage = fls(cap_super_page_val(iommu->cap));
1910 domain->iommu_superpage = 0;
1912 domain->nid = iommu->node;
1914 /* always allocate the top pgd */
1915 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
1918 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1922 static void domain_exit(struct dmar_domain *domain)
1924 struct page *freelist = NULL;
1926 /* Domain 0 is reserved, so dont process it */
1930 /* Flush any lazy unmaps that may reference this domain */
1931 if (!intel_iommu_strict)
1932 flush_unmaps_timeout(0);
1934 /* Remove associated devices and clear attached or cached domains */
1936 domain_remove_dev_info(domain);
1940 put_iova_domain(&domain->iovad);
1942 freelist = domain_unmap(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1944 dma_free_pagelist(freelist);
1946 free_domain_mem(domain);
1949 static int domain_context_mapping_one(struct dmar_domain *domain,
1950 struct intel_iommu *iommu,
1953 u16 did = domain->iommu_did[iommu->seq_id];
1954 int translation = CONTEXT_TT_MULTI_LEVEL;
1955 struct device_domain_info *info = NULL;
1956 struct context_entry *context;
1957 unsigned long flags;
1958 struct dma_pte *pgd;
1963 if (hw_pass_through && domain_type_is_si(domain))
1964 translation = CONTEXT_TT_PASS_THROUGH;
1966 pr_debug("Set context mapping for %02x:%02x.%d\n",
1967 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1969 BUG_ON(!domain->pgd);
1971 spin_lock_irqsave(&device_domain_lock, flags);
1972 spin_lock(&iommu->lock);
1975 context = iommu_context_addr(iommu, bus, devfn, 1);
1980 if (context_present(context))
1985 context_clear_entry(context);
1986 context_set_domain_id(context, did);
1989 * Skip top levels of page tables for iommu which has less agaw
1990 * than default. Unnecessary for PT mode.
1992 if (translation != CONTEXT_TT_PASS_THROUGH) {
1993 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1995 pgd = phys_to_virt(dma_pte_addr(pgd));
1996 if (!dma_pte_present(pgd))
2000 info = iommu_support_dev_iotlb(domain, iommu, bus, devfn);
2001 if (info && info->ats_supported)
2002 translation = CONTEXT_TT_DEV_IOTLB;
2004 translation = CONTEXT_TT_MULTI_LEVEL;
2006 context_set_address_root(context, virt_to_phys(pgd));
2007 context_set_address_width(context, iommu->agaw);
2010 * In pass through mode, AW must be programmed to
2011 * indicate the largest AGAW value supported by
2012 * hardware. And ASR is ignored by hardware.
2014 context_set_address_width(context, iommu->msagaw);
2017 context_set_translation_type(context, translation);
2018 context_set_fault_enable(context);
2019 context_set_present(context);
2020 domain_flush_cache(domain, context, sizeof(*context));
2023 * It's a non-present to present mapping. If hardware doesn't cache
2024 * non-present entry we only need to flush the write-buffer. If the
2025 * _does_ cache non-present entries, then it does so in the special
2026 * domain #0, which we have to flush:
2028 if (cap_caching_mode(iommu->cap)) {
2029 iommu->flush.flush_context(iommu, 0,
2030 (((u16)bus) << 8) | devfn,
2031 DMA_CCMD_MASK_NOBIT,
2032 DMA_CCMD_DEVICE_INVL);
2033 iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
2035 iommu_flush_write_buffer(iommu);
2037 iommu_enable_dev_iotlb(info);
2042 spin_unlock(&iommu->lock);
2043 spin_unlock_irqrestore(&device_domain_lock, flags);
2048 struct domain_context_mapping_data {
2049 struct dmar_domain *domain;
2050 struct intel_iommu *iommu;
2053 static int domain_context_mapping_cb(struct pci_dev *pdev,
2054 u16 alias, void *opaque)
2056 struct domain_context_mapping_data *data = opaque;
2058 return domain_context_mapping_one(data->domain, data->iommu,
2059 PCI_BUS_NUM(alias), alias & 0xff);
2063 domain_context_mapping(struct dmar_domain *domain, struct device *dev)
2065 struct intel_iommu *iommu;
2067 struct domain_context_mapping_data data;
2069 iommu = device_to_iommu(dev, &bus, &devfn);
2073 if (!dev_is_pci(dev))
2074 return domain_context_mapping_one(domain, iommu, bus, devfn);
2076 data.domain = domain;
2079 return pci_for_each_dma_alias(to_pci_dev(dev),
2080 &domain_context_mapping_cb, &data);
2083 static int domain_context_mapped_cb(struct pci_dev *pdev,
2084 u16 alias, void *opaque)
2086 struct intel_iommu *iommu = opaque;
2088 return !device_context_mapped(iommu, PCI_BUS_NUM(alias), alias & 0xff);
2091 static int domain_context_mapped(struct device *dev)
2093 struct intel_iommu *iommu;
2096 iommu = device_to_iommu(dev, &bus, &devfn);
2100 if (!dev_is_pci(dev))
2101 return device_context_mapped(iommu, bus, devfn);
2103 return !pci_for_each_dma_alias(to_pci_dev(dev),
2104 domain_context_mapped_cb, iommu);
2107 /* Returns a number of VTD pages, but aligned to MM page size */
2108 static inline unsigned long aligned_nrpages(unsigned long host_addr,
2111 host_addr &= ~PAGE_MASK;
2112 return PAGE_ALIGN(host_addr + size) >> VTD_PAGE_SHIFT;
2115 /* Return largest possible superpage level for a given mapping */
2116 static inline int hardware_largepage_caps(struct dmar_domain *domain,
2117 unsigned long iov_pfn,
2118 unsigned long phy_pfn,
2119 unsigned long pages)
2121 int support, level = 1;
2122 unsigned long pfnmerge;
2124 support = domain->iommu_superpage;
2126 /* To use a large page, the virtual *and* physical addresses
2127 must be aligned to 2MiB/1GiB/etc. Lower bits set in either
2128 of them will mean we have to use smaller pages. So just
2129 merge them and check both at once. */
2130 pfnmerge = iov_pfn | phy_pfn;
2132 while (support && !(pfnmerge & ~VTD_STRIDE_MASK)) {
2133 pages >>= VTD_STRIDE_SHIFT;
2136 pfnmerge >>= VTD_STRIDE_SHIFT;
2143 static int __domain_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2144 struct scatterlist *sg, unsigned long phys_pfn,
2145 unsigned long nr_pages, int prot)
2147 struct dma_pte *first_pte = NULL, *pte = NULL;
2148 phys_addr_t uninitialized_var(pteval);
2149 unsigned long sg_res = 0;
2150 unsigned int largepage_lvl = 0;
2151 unsigned long lvl_pages = 0;
2153 BUG_ON(!domain_pfn_supported(domain, iov_pfn + nr_pages - 1));
2155 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
2158 prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
2162 pteval = ((phys_addr_t)phys_pfn << VTD_PAGE_SHIFT) | prot;
2165 while (nr_pages > 0) {
2169 sg_res = aligned_nrpages(sg->offset, sg->length);
2170 sg->dma_address = ((dma_addr_t)iov_pfn << VTD_PAGE_SHIFT) + sg->offset;
2171 sg->dma_length = sg->length;
2172 pteval = page_to_phys(sg_page(sg)) | prot;
2173 phys_pfn = pteval >> VTD_PAGE_SHIFT;
2177 largepage_lvl = hardware_largepage_caps(domain, iov_pfn, phys_pfn, sg_res);
2179 first_pte = pte = pfn_to_dma_pte(domain, iov_pfn, &largepage_lvl);
2182 /* It is large page*/
2183 if (largepage_lvl > 1) {
2184 unsigned long nr_superpages, end_pfn;
2186 pteval |= DMA_PTE_LARGE_PAGE;
2187 lvl_pages = lvl_to_nr_pages(largepage_lvl);
2189 nr_superpages = sg_res / lvl_pages;
2190 end_pfn = iov_pfn + nr_superpages * lvl_pages - 1;
2193 * Ensure that old small page tables are
2194 * removed to make room for superpage(s).
2196 dma_pte_free_pagetable(domain, iov_pfn, end_pfn);
2198 pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;
2202 /* We don't need lock here, nobody else
2203 * touches the iova range
2205 tmp = cmpxchg64_local(&pte->val, 0ULL, pteval);
2207 static int dumps = 5;
2208 pr_crit("ERROR: DMA PTE for vPFN 0x%lx already set (to %llx not %llx)\n",
2209 iov_pfn, tmp, (unsigned long long)pteval);
2212 debug_dma_dump_mappings(NULL);
2217 lvl_pages = lvl_to_nr_pages(largepage_lvl);
2219 BUG_ON(nr_pages < lvl_pages);
2220 BUG_ON(sg_res < lvl_pages);
2222 nr_pages -= lvl_pages;
2223 iov_pfn += lvl_pages;
2224 phys_pfn += lvl_pages;
2225 pteval += lvl_pages * VTD_PAGE_SIZE;
2226 sg_res -= lvl_pages;
2228 /* If the next PTE would be the first in a new page, then we
2229 need to flush the cache on the entries we've just written.
2230 And then we'll need to recalculate 'pte', so clear it and
2231 let it get set again in the if (!pte) block above.
2233 If we're done (!nr_pages) we need to flush the cache too.
2235 Also if we've been setting superpages, we may need to
2236 recalculate 'pte' and switch back to smaller pages for the
2237 end of the mapping, if the trailing size is not enough to
2238 use another superpage (i.e. sg_res < lvl_pages). */
2240 if (!nr_pages || first_pte_in_page(pte) ||
2241 (largepage_lvl > 1 && sg_res < lvl_pages)) {
2242 domain_flush_cache(domain, first_pte,
2243 (void *)pte - (void *)first_pte);
2247 if (!sg_res && nr_pages)
2253 static inline int domain_sg_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2254 struct scatterlist *sg, unsigned long nr_pages,
2257 return __domain_mapping(domain, iov_pfn, sg, 0, nr_pages, prot);
2260 static inline int domain_pfn_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2261 unsigned long phys_pfn, unsigned long nr_pages,
2264 return __domain_mapping(domain, iov_pfn, NULL, phys_pfn, nr_pages, prot);
2267 static void domain_context_clear_one(struct intel_iommu *iommu, u8 bus, u8 devfn)
2272 clear_context_table(iommu, bus, devfn);
2273 iommu->flush.flush_context(iommu, 0, 0, 0,
2274 DMA_CCMD_GLOBAL_INVL);
2275 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2278 static inline void unlink_domain_info(struct device_domain_info *info)
2280 assert_spin_locked(&device_domain_lock);
2281 list_del(&info->link);
2282 list_del(&info->global);
2284 info->dev->archdata.iommu = NULL;
2287 static void domain_remove_dev_info(struct dmar_domain *domain)
2289 struct device_domain_info *info, *tmp;
2290 unsigned long flags;
2292 spin_lock_irqsave(&device_domain_lock, flags);
2293 list_for_each_entry_safe(info, tmp, &domain->devices, link)
2294 __dmar_remove_one_dev_info(info);
2295 spin_unlock_irqrestore(&device_domain_lock, flags);
2300 * Note: we use struct device->archdata.iommu stores the info
2302 static struct dmar_domain *find_domain(struct device *dev)
2304 struct device_domain_info *info;
2306 /* No lock here, assumes no domain exit in normal case */
2307 info = dev->archdata.iommu;
2309 return info->domain;
2313 static inline struct device_domain_info *
2314 dmar_search_domain_by_dev_info(int segment, int bus, int devfn)
2316 struct device_domain_info *info;
2318 list_for_each_entry(info, &device_domain_list, global)
2319 if (info->iommu->segment == segment && info->bus == bus &&
2320 info->devfn == devfn)
2326 static struct dmar_domain *dmar_insert_one_dev_info(struct intel_iommu *iommu,
2329 struct dmar_domain *domain)
2331 struct dmar_domain *found = NULL;
2332 struct device_domain_info *info;
2333 unsigned long flags;
2336 info = alloc_devinfo_mem();
2341 info->devfn = devfn;
2342 info->ats_supported = info->pasid_supported = info->pri_supported = 0;
2343 info->ats_enabled = info->pasid_enabled = info->pri_enabled = 0;
2346 info->domain = domain;
2347 info->iommu = iommu;
2349 if (dev && dev_is_pci(dev)) {
2350 struct pci_dev *pdev = to_pci_dev(info->dev);
2352 if (ecap_dev_iotlb_support(iommu->ecap) &&
2353 pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS) &&
2354 dmar_find_matched_atsr_unit(pdev))
2355 info->ats_supported = 1;
2357 if (ecs_enabled(iommu)) {
2358 if (pasid_enabled(iommu)) {
2359 int features = pci_pasid_features(pdev);
2361 info->pasid_supported = features | 1;
2364 if (info->ats_supported && ecap_prs(iommu->ecap) &&
2365 pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI))
2366 info->pri_supported = 1;
2370 spin_lock_irqsave(&device_domain_lock, flags);
2372 found = find_domain(dev);
2375 struct device_domain_info *info2;
2376 info2 = dmar_search_domain_by_dev_info(iommu->segment, bus, devfn);
2378 found = info2->domain;
2384 spin_unlock_irqrestore(&device_domain_lock, flags);
2385 free_devinfo_mem(info);
2386 /* Caller must free the original domain */
2390 spin_lock(&iommu->lock);
2391 ret = domain_attach_iommu(domain, iommu);
2392 spin_unlock(&iommu->lock);
2395 spin_unlock_irqrestore(&device_domain_lock, flags);
2396 free_devinfo_mem(info);
2400 list_add(&info->link, &domain->devices);
2401 list_add(&info->global, &device_domain_list);
2403 dev->archdata.iommu = info;
2404 spin_unlock_irqrestore(&device_domain_lock, flags);
2406 if (dev && domain_context_mapping(domain, dev)) {
2407 pr_err("Domain context map for %s failed\n", dev_name(dev));
2408 dmar_remove_one_dev_info(domain, dev);
2415 static int get_last_alias(struct pci_dev *pdev, u16 alias, void *opaque)
2417 *(u16 *)opaque = alias;
2421 /* domain is initialized */
2422 static struct dmar_domain *get_domain_for_dev(struct device *dev, int gaw)
2424 struct device_domain_info *info = NULL;
2425 struct dmar_domain *domain, *tmp;
2426 struct intel_iommu *iommu;
2427 u16 req_id, dma_alias;
2428 unsigned long flags;
2431 domain = find_domain(dev);
2435 iommu = device_to_iommu(dev, &bus, &devfn);
2439 req_id = ((u16)bus << 8) | devfn;
2441 if (dev_is_pci(dev)) {
2442 struct pci_dev *pdev = to_pci_dev(dev);
2444 pci_for_each_dma_alias(pdev, get_last_alias, &dma_alias);
2446 spin_lock_irqsave(&device_domain_lock, flags);
2447 info = dmar_search_domain_by_dev_info(pci_domain_nr(pdev->bus),
2448 PCI_BUS_NUM(dma_alias),
2451 iommu = info->iommu;
2452 domain = info->domain;
2454 spin_unlock_irqrestore(&device_domain_lock, flags);
2456 /* DMA alias already has a domain, uses it */
2461 /* Allocate and initialize new domain for the device */
2462 domain = alloc_domain(0);
2465 if (domain_init(domain, iommu, gaw)) {
2466 domain_exit(domain);
2470 /* register PCI DMA alias device */
2471 if (req_id != dma_alias && dev_is_pci(dev)) {
2472 tmp = dmar_insert_one_dev_info(iommu, PCI_BUS_NUM(dma_alias),
2473 dma_alias & 0xff, NULL, domain);
2475 if (!tmp || tmp != domain) {
2476 domain_exit(domain);
2485 tmp = dmar_insert_one_dev_info(iommu, bus, devfn, dev, domain);
2487 if (!tmp || tmp != domain) {
2488 domain_exit(domain);
2495 static int iommu_domain_identity_map(struct dmar_domain *domain,
2496 unsigned long long start,
2497 unsigned long long end)
2499 unsigned long first_vpfn = start >> VTD_PAGE_SHIFT;
2500 unsigned long last_vpfn = end >> VTD_PAGE_SHIFT;
2502 if (!reserve_iova(&domain->iovad, dma_to_mm_pfn(first_vpfn),
2503 dma_to_mm_pfn(last_vpfn))) {
2504 pr_err("Reserving iova failed\n");
2508 pr_debug("Mapping reserved region %llx-%llx\n", start, end);
2510 * RMRR range might have overlap with physical memory range,
2513 dma_pte_clear_range(domain, first_vpfn, last_vpfn);
2515 return domain_pfn_mapping(domain, first_vpfn, first_vpfn,
2516 last_vpfn - first_vpfn + 1,
2517 DMA_PTE_READ|DMA_PTE_WRITE);
2520 static int domain_prepare_identity_map(struct device *dev,
2521 struct dmar_domain *domain,
2522 unsigned long long start,
2523 unsigned long long end)
2525 /* For _hardware_ passthrough, don't bother. But for software
2526 passthrough, we do it anyway -- it may indicate a memory
2527 range which is reserved in E820, so which didn't get set
2528 up to start with in si_domain */
2529 if (domain == si_domain && hw_pass_through) {
2530 pr_warn("Ignoring identity map for HW passthrough device %s [0x%Lx - 0x%Lx]\n",
2531 dev_name(dev), start, end);
2535 pr_info("Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
2536 dev_name(dev), start, end);
2539 WARN(1, "Your BIOS is broken; RMRR ends before it starts!\n"
2540 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2541 dmi_get_system_info(DMI_BIOS_VENDOR),
2542 dmi_get_system_info(DMI_BIOS_VERSION),
2543 dmi_get_system_info(DMI_PRODUCT_VERSION));
2547 if (end >> agaw_to_width(domain->agaw)) {
2548 WARN(1, "Your BIOS is broken; RMRR exceeds permitted address width (%d bits)\n"
2549 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2550 agaw_to_width(domain->agaw),
2551 dmi_get_system_info(DMI_BIOS_VENDOR),
2552 dmi_get_system_info(DMI_BIOS_VERSION),
2553 dmi_get_system_info(DMI_PRODUCT_VERSION));
2557 return iommu_domain_identity_map(domain, start, end);
2560 static int iommu_prepare_identity_map(struct device *dev,
2561 unsigned long long start,
2562 unsigned long long end)
2564 struct dmar_domain *domain;
2567 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2571 ret = domain_prepare_identity_map(dev, domain, start, end);
2573 domain_exit(domain);
2578 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
2581 if (dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2583 return iommu_prepare_identity_map(dev, rmrr->base_address,
2587 #ifdef CONFIG_INTEL_IOMMU_FLOPPY_WA
2588 static inline void iommu_prepare_isa(void)
2590 struct pci_dev *pdev;
2593 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
2597 pr_info("Prepare 0-16MiB unity mapping for LPC\n");
2598 ret = iommu_prepare_identity_map(&pdev->dev, 0, 16*1024*1024 - 1);
2601 pr_err("Failed to create 0-16MiB identity map - floppy might not work\n");
2606 static inline void iommu_prepare_isa(void)
2610 #endif /* !CONFIG_INTEL_IOMMU_FLPY_WA */
2612 static int md_domain_init(struct dmar_domain *domain, int guest_width);
2614 static int __init si_domain_init(int hw)
2618 si_domain = alloc_domain(DOMAIN_FLAG_STATIC_IDENTITY);
2622 if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2623 domain_exit(si_domain);
2627 pr_debug("Identity mapping domain allocated\n");
2632 for_each_online_node(nid) {
2633 unsigned long start_pfn, end_pfn;
2636 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
2637 ret = iommu_domain_identity_map(si_domain,
2638 PFN_PHYS(start_pfn), PFN_PHYS(end_pfn));
2647 static int identity_mapping(struct device *dev)
2649 struct device_domain_info *info;
2651 if (likely(!iommu_identity_mapping))
2654 info = dev->archdata.iommu;
2655 if (info && info != DUMMY_DEVICE_DOMAIN_INFO)
2656 return (info->domain == si_domain);
2661 static int domain_add_dev_info(struct dmar_domain *domain, struct device *dev)
2663 struct dmar_domain *ndomain;
2664 struct intel_iommu *iommu;
2667 iommu = device_to_iommu(dev, &bus, &devfn);
2671 ndomain = dmar_insert_one_dev_info(iommu, bus, devfn, dev, domain);
2672 if (ndomain != domain)
2678 static bool device_has_rmrr(struct device *dev)
2680 struct dmar_rmrr_unit *rmrr;
2685 for_each_rmrr_units(rmrr) {
2687 * Return TRUE if this RMRR contains the device that
2690 for_each_active_dev_scope(rmrr->devices,
2691 rmrr->devices_cnt, i, tmp)
2702 * There are a couple cases where we need to restrict the functionality of
2703 * devices associated with RMRRs. The first is when evaluating a device for
2704 * identity mapping because problems exist when devices are moved in and out
2705 * of domains and their respective RMRR information is lost. This means that
2706 * a device with associated RMRRs will never be in a "passthrough" domain.
2707 * The second is use of the device through the IOMMU API. This interface
2708 * expects to have full control of the IOVA space for the device. We cannot
2709 * satisfy both the requirement that RMRR access is maintained and have an
2710 * unencumbered IOVA space. We also have no ability to quiesce the device's
2711 * use of the RMRR space or even inform the IOMMU API user of the restriction.
2712 * We therefore prevent devices associated with an RMRR from participating in
2713 * the IOMMU API, which eliminates them from device assignment.
2715 * In both cases we assume that PCI USB devices with RMRRs have them largely
2716 * for historical reasons and that the RMRR space is not actively used post
2717 * boot. This exclusion may change if vendors begin to abuse it.
2719 * The same exception is made for graphics devices, with the requirement that
2720 * any use of the RMRR regions will be torn down before assigning the device
2723 static bool device_is_rmrr_locked(struct device *dev)
2725 if (!device_has_rmrr(dev))
2728 if (dev_is_pci(dev)) {
2729 struct pci_dev *pdev = to_pci_dev(dev);
2731 if (IS_USB_DEVICE(pdev) || IS_GFX_DEVICE(pdev))
2738 static int iommu_should_identity_map(struct device *dev, int startup)
2741 if (dev_is_pci(dev)) {
2742 struct pci_dev *pdev = to_pci_dev(dev);
2744 if (device_is_rmrr_locked(dev))
2747 if ((iommu_identity_mapping & IDENTMAP_AZALIA) && IS_AZALIA(pdev))
2750 if ((iommu_identity_mapping & IDENTMAP_GFX) && IS_GFX_DEVICE(pdev))
2753 if (!(iommu_identity_mapping & IDENTMAP_ALL))
2757 * We want to start off with all devices in the 1:1 domain, and
2758 * take them out later if we find they can't access all of memory.
2760 * However, we can't do this for PCI devices behind bridges,
2761 * because all PCI devices behind the same bridge will end up
2762 * with the same source-id on their transactions.
2764 * Practically speaking, we can't change things around for these
2765 * devices at run-time, because we can't be sure there'll be no
2766 * DMA transactions in flight for any of their siblings.
2768 * So PCI devices (unless they're on the root bus) as well as
2769 * their parent PCI-PCI or PCIe-PCI bridges must be left _out_ of
2770 * the 1:1 domain, just in _case_ one of their siblings turns out
2771 * not to be able to map all of memory.
2773 if (!pci_is_pcie(pdev)) {
2774 if (!pci_is_root_bus(pdev->bus))
2776 if (pdev->class >> 8 == PCI_CLASS_BRIDGE_PCI)
2778 } else if (pci_pcie_type(pdev) == PCI_EXP_TYPE_PCI_BRIDGE)
2781 if (device_has_rmrr(dev))
2786 * At boot time, we don't yet know if devices will be 64-bit capable.
2787 * Assume that they will — if they turn out not to be, then we can
2788 * take them out of the 1:1 domain later.
2792 * If the device's dma_mask is less than the system's memory
2793 * size then this is not a candidate for identity mapping.
2795 u64 dma_mask = *dev->dma_mask;
2797 if (dev->coherent_dma_mask &&
2798 dev->coherent_dma_mask < dma_mask)
2799 dma_mask = dev->coherent_dma_mask;
2801 return dma_mask >= dma_get_required_mask(dev);
2807 static int __init dev_prepare_static_identity_mapping(struct device *dev, int hw)
2811 if (!iommu_should_identity_map(dev, 1))
2814 ret = domain_add_dev_info(si_domain, dev);
2816 pr_info("%s identity mapping for device %s\n",
2817 hw ? "Hardware" : "Software", dev_name(dev));
2818 else if (ret == -ENODEV)
2819 /* device not associated with an iommu */
2826 static int __init iommu_prepare_static_identity_mapping(int hw)
2828 struct pci_dev *pdev = NULL;
2829 struct dmar_drhd_unit *drhd;
2830 struct intel_iommu *iommu;
2835 for_each_pci_dev(pdev) {
2836 ret = dev_prepare_static_identity_mapping(&pdev->dev, hw);
2841 for_each_active_iommu(iommu, drhd)
2842 for_each_active_dev_scope(drhd->devices, drhd->devices_cnt, i, dev) {
2843 struct acpi_device_physical_node *pn;
2844 struct acpi_device *adev;
2846 if (dev->bus != &acpi_bus_type)
2849 adev= to_acpi_device(dev);
2850 mutex_lock(&adev->physical_node_lock);
2851 list_for_each_entry(pn, &adev->physical_node_list, node) {
2852 ret = dev_prepare_static_identity_mapping(pn->dev, hw);
2856 mutex_unlock(&adev->physical_node_lock);
2864 static void intel_iommu_init_qi(struct intel_iommu *iommu)
2867 * Start from the sane iommu hardware state.
2868 * If the queued invalidation is already initialized by us
2869 * (for example, while enabling interrupt-remapping) then
2870 * we got the things already rolling from a sane state.
2874 * Clear any previous faults.
2876 dmar_fault(-1, iommu);
2878 * Disable queued invalidation if supported and already enabled
2879 * before OS handover.
2881 dmar_disable_qi(iommu);
2884 if (dmar_enable_qi(iommu)) {
2886 * Queued Invalidate not enabled, use Register Based Invalidate
2888 iommu->flush.flush_context = __iommu_flush_context;
2889 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2890 pr_info("%s: Using Register based invalidation\n",
2893 iommu->flush.flush_context = qi_flush_context;
2894 iommu->flush.flush_iotlb = qi_flush_iotlb;
2895 pr_info("%s: Using Queued invalidation\n", iommu->name);
2899 static int copy_context_table(struct intel_iommu *iommu,
2900 struct root_entry *old_re,
2901 struct context_entry **tbl,
2904 int tbl_idx, pos = 0, idx, devfn, ret = 0, did;
2905 struct context_entry *new_ce = NULL, ce;
2906 struct context_entry *old_ce = NULL;
2907 struct root_entry re;
2908 phys_addr_t old_ce_phys;
2910 tbl_idx = ext ? bus * 2 : bus;
2911 memcpy(&re, old_re, sizeof(re));
2913 for (devfn = 0; devfn < 256; devfn++) {
2914 /* First calculate the correct index */
2915 idx = (ext ? devfn * 2 : devfn) % 256;
2918 /* First save what we may have and clean up */
2920 tbl[tbl_idx] = new_ce;
2921 __iommu_flush_cache(iommu, new_ce,
2931 old_ce_phys = root_entry_lctp(&re);
2933 old_ce_phys = root_entry_uctp(&re);
2936 if (ext && devfn == 0) {
2937 /* No LCTP, try UCTP */
2946 old_ce = memremap(old_ce_phys, PAGE_SIZE,
2951 new_ce = alloc_pgtable_page(iommu->node);
2958 /* Now copy the context entry */
2959 memcpy(&ce, old_ce + idx, sizeof(ce));
2961 if (!__context_present(&ce))
2964 did = context_domain_id(&ce);
2965 if (did >= 0 && did < cap_ndoms(iommu->cap))
2966 set_bit(did, iommu->domain_ids);
2969 * We need a marker for copied context entries. This
2970 * marker needs to work for the old format as well as
2971 * for extended context entries.
2973 * Bit 67 of the context entry is used. In the old
2974 * format this bit is available to software, in the
2975 * extended format it is the PGE bit, but PGE is ignored
2976 * by HW if PASIDs are disabled (and thus still
2979 * So disable PASIDs first and then mark the entry
2980 * copied. This means that we don't copy PASID
2981 * translations from the old kernel, but this is fine as
2982 * faults there are not fatal.
2984 context_clear_pasid_enable(&ce);
2985 context_set_copied(&ce);
2990 tbl[tbl_idx + pos] = new_ce;
2992 __iommu_flush_cache(iommu, new_ce, VTD_PAGE_SIZE);
3001 static int copy_translation_tables(struct intel_iommu *iommu)
3003 struct context_entry **ctxt_tbls;
3004 struct root_entry *old_rt;
3005 phys_addr_t old_rt_phys;
3006 int ctxt_table_entries;
3007 unsigned long flags;
3012 rtaddr_reg = dmar_readq(iommu->reg + DMAR_RTADDR_REG);
3013 ext = !!(rtaddr_reg & DMA_RTADDR_RTT);
3014 new_ext = !!ecap_ecs(iommu->ecap);
3017 * The RTT bit can only be changed when translation is disabled,
3018 * but disabling translation means to open a window for data
3019 * corruption. So bail out and don't copy anything if we would
3020 * have to change the bit.
3025 old_rt_phys = rtaddr_reg & VTD_PAGE_MASK;
3029 old_rt = memremap(old_rt_phys, PAGE_SIZE, MEMREMAP_WB);
3033 /* This is too big for the stack - allocate it from slab */
3034 ctxt_table_entries = ext ? 512 : 256;
3036 ctxt_tbls = kzalloc(ctxt_table_entries * sizeof(void *), GFP_KERNEL);
3040 for (bus = 0; bus < 256; bus++) {
3041 ret = copy_context_table(iommu, &old_rt[bus],
3042 ctxt_tbls, bus, ext);
3044 pr_err("%s: Failed to copy context table for bus %d\n",
3050 spin_lock_irqsave(&iommu->lock, flags);
3052 /* Context tables are copied, now write them to the root_entry table */
3053 for (bus = 0; bus < 256; bus++) {
3054 int idx = ext ? bus * 2 : bus;
3057 if (ctxt_tbls[idx]) {
3058 val = virt_to_phys(ctxt_tbls[idx]) | 1;
3059 iommu->root_entry[bus].lo = val;
3062 if (!ext || !ctxt_tbls[idx + 1])
3065 val = virt_to_phys(ctxt_tbls[idx + 1]) | 1;
3066 iommu->root_entry[bus].hi = val;
3069 spin_unlock_irqrestore(&iommu->lock, flags);
3073 __iommu_flush_cache(iommu, iommu->root_entry, PAGE_SIZE);
3083 static int __init init_dmars(void)
3085 struct dmar_drhd_unit *drhd;
3086 struct dmar_rmrr_unit *rmrr;
3087 bool copied_tables = false;
3089 struct intel_iommu *iommu;
3095 * initialize and program root entry to not present
3098 for_each_drhd_unit(drhd) {
3100 * lock not needed as this is only incremented in the single
3101 * threaded kernel __init code path all other access are read
3104 if (g_num_of_iommus < DMAR_UNITS_SUPPORTED) {
3108 pr_err_once("Exceeded %d IOMMUs\n", DMAR_UNITS_SUPPORTED);
3111 /* Preallocate enough resources for IOMMU hot-addition */
3112 if (g_num_of_iommus < DMAR_UNITS_SUPPORTED)
3113 g_num_of_iommus = DMAR_UNITS_SUPPORTED;
3115 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
3118 pr_err("Allocating global iommu array failed\n");
3123 deferred_flush = kzalloc(g_num_of_iommus *
3124 sizeof(struct deferred_flush_tables), GFP_KERNEL);
3125 if (!deferred_flush) {
3130 for_each_active_iommu(iommu, drhd) {
3131 g_iommus[iommu->seq_id] = iommu;
3133 intel_iommu_init_qi(iommu);
3135 ret = iommu_init_domains(iommu);
3139 init_translation_status(iommu);
3141 if (translation_pre_enabled(iommu) && !is_kdump_kernel()) {
3142 iommu_disable_translation(iommu);
3143 clear_translation_pre_enabled(iommu);
3144 pr_warn("Translation was enabled for %s but we are not in kdump mode\n",
3150 * we could share the same root & context tables
3151 * among all IOMMU's. Need to Split it later.
3153 ret = iommu_alloc_root_entry(iommu);
3157 if (translation_pre_enabled(iommu)) {
3158 pr_info("Translation already enabled - trying to copy translation structures\n");
3160 ret = copy_translation_tables(iommu);
3163 * We found the IOMMU with translation
3164 * enabled - but failed to copy over the
3165 * old root-entry table. Try to proceed
3166 * by disabling translation now and
3167 * allocating a clean root-entry table.
3168 * This might cause DMAR faults, but
3169 * probably the dump will still succeed.
3171 pr_err("Failed to copy translation tables from previous kernel for %s\n",
3173 iommu_disable_translation(iommu);
3174 clear_translation_pre_enabled(iommu);
3176 pr_info("Copied translation tables from previous kernel for %s\n",
3178 copied_tables = true;
3182 if (!ecap_pass_through(iommu->ecap))
3183 hw_pass_through = 0;
3184 #ifdef CONFIG_INTEL_IOMMU_SVM
3185 if (pasid_enabled(iommu))
3186 intel_svm_alloc_pasid_tables(iommu);
3191 * Now that qi is enabled on all iommus, set the root entry and flush
3192 * caches. This is required on some Intel X58 chipsets, otherwise the
3193 * flush_context function will loop forever and the boot hangs.
3195 for_each_active_iommu(iommu, drhd) {
3196 iommu_flush_write_buffer(iommu);
3197 iommu_set_root_entry(iommu);
3198 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
3199 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
3202 if (iommu_pass_through)
3203 iommu_identity_mapping |= IDENTMAP_ALL;
3205 #ifdef CONFIG_INTEL_IOMMU_BROKEN_GFX_WA
3206 iommu_identity_mapping |= IDENTMAP_GFX;
3209 if (iommu_identity_mapping) {
3210 ret = si_domain_init(hw_pass_through);
3215 check_tylersburg_isoch();
3218 * If we copied translations from a previous kernel in the kdump
3219 * case, we can not assign the devices to domains now, as that
3220 * would eliminate the old mappings. So skip this part and defer
3221 * the assignment to device driver initialization time.
3227 * If pass through is not set or not enabled, setup context entries for
3228 * identity mappings for rmrr, gfx, and isa and may fall back to static
3229 * identity mapping if iommu_identity_mapping is set.
3231 if (iommu_identity_mapping) {
3232 ret = iommu_prepare_static_identity_mapping(hw_pass_through);
3234 pr_crit("Failed to setup IOMMU pass-through\n");
3240 * for each dev attached to rmrr
3242 * locate drhd for dev, alloc domain for dev
3243 * allocate free domain
3244 * allocate page table entries for rmrr
3245 * if context not allocated for bus
3246 * allocate and init context
3247 * set present in root table for this bus
3248 * init context with domain, translation etc
3252 pr_info("Setting RMRR:\n");
3253 for_each_rmrr_units(rmrr) {
3254 /* some BIOS lists non-exist devices in DMAR table. */
3255 for_each_active_dev_scope(rmrr->devices, rmrr->devices_cnt,
3257 ret = iommu_prepare_rmrr_dev(rmrr, dev);
3259 pr_err("Mapping reserved region failed\n");
3263 iommu_prepare_isa();
3270 * global invalidate context cache
3271 * global invalidate iotlb
3272 * enable translation
3274 for_each_iommu(iommu, drhd) {
3275 if (drhd->ignored) {
3277 * we always have to disable PMRs or DMA may fail on
3281 iommu_disable_protect_mem_regions(iommu);
3285 iommu_flush_write_buffer(iommu);
3287 #ifdef CONFIG_INTEL_IOMMU_SVM
3288 if (pasid_enabled(iommu) && ecap_prs(iommu->ecap)) {
3289 ret = intel_svm_enable_prq(iommu);
3294 ret = dmar_set_interrupt(iommu);
3298 if (!translation_pre_enabled(iommu))
3299 iommu_enable_translation(iommu);
3301 iommu_disable_protect_mem_regions(iommu);
3307 for_each_active_iommu(iommu, drhd) {
3308 disable_dmar_iommu(iommu);
3309 free_dmar_iommu(iommu);
3311 kfree(deferred_flush);
3318 /* This takes a number of _MM_ pages, not VTD pages */
3319 static struct iova *intel_alloc_iova(struct device *dev,
3320 struct dmar_domain *domain,
3321 unsigned long nrpages, uint64_t dma_mask)
3323 struct iova *iova = NULL;
3325 /* Restrict dma_mask to the width that the iommu can handle */
3326 dma_mask = min_t(uint64_t, DOMAIN_MAX_ADDR(domain->gaw), dma_mask);
3327 /* Ensure we reserve the whole size-aligned region */
3328 nrpages = __roundup_pow_of_two(nrpages);
3330 if (!dmar_forcedac && dma_mask > DMA_BIT_MASK(32)) {
3332 * First try to allocate an io virtual address in
3333 * DMA_BIT_MASK(32) and if that fails then try allocating
3336 iova = alloc_iova(&domain->iovad, nrpages,
3337 IOVA_PFN(DMA_BIT_MASK(32)), 1);
3341 iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
3342 if (unlikely(!iova)) {
3343 pr_err("Allocating %ld-page iova for %s failed",
3344 nrpages, dev_name(dev));
3351 static struct dmar_domain *__get_valid_domain_for_dev(struct device *dev)
3353 struct dmar_rmrr_unit *rmrr;
3354 struct dmar_domain *domain;
3355 struct device *i_dev;
3358 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
3360 pr_err("Allocating domain for %s failed\n",
3365 /* We have a new domain - setup possible RMRRs for the device */
3367 for_each_rmrr_units(rmrr) {
3368 for_each_active_dev_scope(rmrr->devices, rmrr->devices_cnt,
3373 ret = domain_prepare_identity_map(dev, domain,
3377 dev_err(dev, "Mapping reserved region failed\n");
3385 static inline struct dmar_domain *get_valid_domain_for_dev(struct device *dev)
3387 struct device_domain_info *info;
3389 /* No lock here, assumes no domain exit in normal case */
3390 info = dev->archdata.iommu;
3392 return info->domain;
3394 return __get_valid_domain_for_dev(dev);
3397 /* Check if the dev needs to go through non-identity map and unmap process.*/
3398 static int iommu_no_mapping(struct device *dev)
3402 if (iommu_dummy(dev))
3405 if (!iommu_identity_mapping)
3408 found = identity_mapping(dev);
3410 if (iommu_should_identity_map(dev, 0))
3414 * 32 bit DMA is removed from si_domain and fall back
3415 * to non-identity mapping.
3417 dmar_remove_one_dev_info(si_domain, dev);
3418 pr_info("32bit %s uses non-identity mapping\n",
3424 * In case of a detached 64 bit DMA device from vm, the device
3425 * is put into si_domain for identity mapping.
3427 if (iommu_should_identity_map(dev, 0)) {
3429 ret = domain_add_dev_info(si_domain, dev);
3431 pr_info("64bit %s uses identity mapping\n",
3441 static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
3442 size_t size, int dir, u64 dma_mask)
3444 struct dmar_domain *domain;
3445 phys_addr_t start_paddr;
3449 struct intel_iommu *iommu;
3450 unsigned long paddr_pfn = paddr >> PAGE_SHIFT;
3452 BUG_ON(dir == DMA_NONE);
3454 if (iommu_no_mapping(dev))
3457 domain = get_valid_domain_for_dev(dev);
3461 iommu = domain_get_iommu(domain);
3462 size = aligned_nrpages(paddr, size);
3464 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size), dma_mask);
3469 * Check if DMAR supports zero-length reads on write only
3472 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3473 !cap_zlr(iommu->cap))
3474 prot |= DMA_PTE_READ;
3475 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3476 prot |= DMA_PTE_WRITE;
3478 * paddr - (paddr + size) might be partial page, we should map the whole
3479 * page. Note: if two part of one page are separately mapped, we
3480 * might have two guest_addr mapping to the same host paddr, but this
3481 * is not a big problem
3483 ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
3484 mm_to_dma_pfn(paddr_pfn), size, prot);
3488 /* it's a non-present to present mapping. Only flush if caching mode */
3489 if (cap_caching_mode(iommu->cap))
3490 iommu_flush_iotlb_psi(iommu, domain,
3491 mm_to_dma_pfn(iova->pfn_lo),
3494 iommu_flush_write_buffer(iommu);
3496 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
3497 start_paddr += paddr & ~PAGE_MASK;
3502 __free_iova(&domain->iovad, iova);
3503 pr_err("Device %s request: %zx@%llx dir %d --- failed\n",
3504 dev_name(dev), size, (unsigned long long)paddr, dir);
3508 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
3509 unsigned long offset, size_t size,
3510 enum dma_data_direction dir,
3511 struct dma_attrs *attrs)
3513 return __intel_map_single(dev, page_to_phys(page) + offset, size,
3514 dir, *dev->dma_mask);
3517 static void flush_unmaps(void)
3523 /* just flush them all */
3524 for (i = 0; i < g_num_of_iommus; i++) {
3525 struct intel_iommu *iommu = g_iommus[i];
3529 if (!deferred_flush[i].next)
3532 /* In caching mode, global flushes turn emulation expensive */
3533 if (!cap_caching_mode(iommu->cap))
3534 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3535 DMA_TLB_GLOBAL_FLUSH);
3536 for (j = 0; j < deferred_flush[i].next; j++) {
3538 struct iova *iova = deferred_flush[i].iova[j];
3539 struct dmar_domain *domain = deferred_flush[i].domain[j];
3541 /* On real hardware multiple invalidations are expensive */
3542 if (cap_caching_mode(iommu->cap))
3543 iommu_flush_iotlb_psi(iommu, domain,
3544 iova->pfn_lo, iova_size(iova),
3545 !deferred_flush[i].freelist[j], 0);
3547 mask = ilog2(mm_to_dma_pfn(iova_size(iova)));
3548 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
3549 (uint64_t)iova->pfn_lo << PAGE_SHIFT, mask);
3551 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
3552 if (deferred_flush[i].freelist[j])
3553 dma_free_pagelist(deferred_flush[i].freelist[j]);
3555 deferred_flush[i].next = 0;
3561 static void flush_unmaps_timeout(unsigned long data)
3563 unsigned long flags;
3565 spin_lock_irqsave(&async_umap_flush_lock, flags);
3567 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3570 static void add_unmap(struct dmar_domain *dom, struct iova *iova, struct page *freelist)
3572 unsigned long flags;
3574 struct intel_iommu *iommu;
3576 spin_lock_irqsave(&async_umap_flush_lock, flags);
3577 if (list_size == HIGH_WATER_MARK)
3580 iommu = domain_get_iommu(dom);
3581 iommu_id = iommu->seq_id;
3583 next = deferred_flush[iommu_id].next;
3584 deferred_flush[iommu_id].domain[next] = dom;
3585 deferred_flush[iommu_id].iova[next] = iova;
3586 deferred_flush[iommu_id].freelist[next] = freelist;
3587 deferred_flush[iommu_id].next++;
3590 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
3594 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3597 static void intel_unmap(struct device *dev, dma_addr_t dev_addr)
3599 struct dmar_domain *domain;
3600 unsigned long start_pfn, last_pfn;
3602 struct intel_iommu *iommu;
3603 struct page *freelist;
3605 if (iommu_no_mapping(dev))
3608 domain = find_domain(dev);
3611 iommu = domain_get_iommu(domain);
3613 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
3614 if (WARN_ONCE(!iova, "Driver unmaps unmatched page at PFN %llx\n",
3615 (unsigned long long)dev_addr))
3618 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
3619 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
3621 pr_debug("Device %s unmapping: pfn %lx-%lx\n",
3622 dev_name(dev), start_pfn, last_pfn);
3624 freelist = domain_unmap(domain, start_pfn, last_pfn);
3626 if (intel_iommu_strict) {
3627 iommu_flush_iotlb_psi(iommu, domain, start_pfn,
3628 last_pfn - start_pfn + 1, !freelist, 0);
3630 __free_iova(&domain->iovad, iova);
3631 dma_free_pagelist(freelist);
3633 add_unmap(domain, iova, freelist);
3635 * queue up the release of the unmap to save the 1/6th of the
3636 * cpu used up by the iotlb flush operation...
3641 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
3642 size_t size, enum dma_data_direction dir,
3643 struct dma_attrs *attrs)
3645 intel_unmap(dev, dev_addr);
3648 static void *intel_alloc_coherent(struct device *dev, size_t size,
3649 dma_addr_t *dma_handle, gfp_t flags,
3650 struct dma_attrs *attrs)
3652 struct page *page = NULL;
3655 size = PAGE_ALIGN(size);
3656 order = get_order(size);
3658 if (!iommu_no_mapping(dev))
3659 flags &= ~(GFP_DMA | GFP_DMA32);
3660 else if (dev->coherent_dma_mask < dma_get_required_mask(dev)) {
3661 if (dev->coherent_dma_mask < DMA_BIT_MASK(32))
3667 if (gfpflags_allow_blocking(flags)) {
3668 unsigned int count = size >> PAGE_SHIFT;
3670 page = dma_alloc_from_contiguous(dev, count, order);
3671 if (page && iommu_no_mapping(dev) &&
3672 page_to_phys(page) + size > dev->coherent_dma_mask) {
3673 dma_release_from_contiguous(dev, page, count);
3679 page = alloc_pages(flags, order);
3682 memset(page_address(page), 0, size);
3684 *dma_handle = __intel_map_single(dev, page_to_phys(page), size,
3686 dev->coherent_dma_mask);
3688 return page_address(page);
3689 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3690 __free_pages(page, order);
3695 static void intel_free_coherent(struct device *dev, size_t size, void *vaddr,
3696 dma_addr_t dma_handle, struct dma_attrs *attrs)
3699 struct page *page = virt_to_page(vaddr);
3701 size = PAGE_ALIGN(size);
3702 order = get_order(size);
3704 intel_unmap(dev, dma_handle);
3705 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3706 __free_pages(page, order);
3709 static void intel_unmap_sg(struct device *dev, struct scatterlist *sglist,
3710 int nelems, enum dma_data_direction dir,
3711 struct dma_attrs *attrs)
3713 intel_unmap(dev, sglist[0].dma_address);
3716 static int intel_nontranslate_map_sg(struct device *hddev,
3717 struct scatterlist *sglist, int nelems, int dir)
3720 struct scatterlist *sg;
3722 for_each_sg(sglist, sg, nelems, i) {
3723 BUG_ON(!sg_page(sg));
3724 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
3725 sg->dma_length = sg->length;
3730 static int intel_map_sg(struct device *dev, struct scatterlist *sglist, int nelems,
3731 enum dma_data_direction dir, struct dma_attrs *attrs)
3734 struct dmar_domain *domain;
3737 struct iova *iova = NULL;
3739 struct scatterlist *sg;
3740 unsigned long start_vpfn;
3741 struct intel_iommu *iommu;
3743 BUG_ON(dir == DMA_NONE);
3744 if (iommu_no_mapping(dev))
3745 return intel_nontranslate_map_sg(dev, sglist, nelems, dir);
3747 domain = get_valid_domain_for_dev(dev);
3751 iommu = domain_get_iommu(domain);
3753 for_each_sg(sglist, sg, nelems, i)
3754 size += aligned_nrpages(sg->offset, sg->length);
3756 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size),
3759 sglist->dma_length = 0;
3764 * Check if DMAR supports zero-length reads on write only
3767 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3768 !cap_zlr(iommu->cap))
3769 prot |= DMA_PTE_READ;
3770 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3771 prot |= DMA_PTE_WRITE;
3773 start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
3775 ret = domain_sg_mapping(domain, start_vpfn, sglist, size, prot);
3776 if (unlikely(ret)) {
3777 dma_pte_free_pagetable(domain, start_vpfn,
3778 start_vpfn + size - 1);
3779 __free_iova(&domain->iovad, iova);
3783 /* it's a non-present to present mapping. Only flush if caching mode */
3784 if (cap_caching_mode(iommu->cap))
3785 iommu_flush_iotlb_psi(iommu, domain, start_vpfn, size, 0, 1);
3787 iommu_flush_write_buffer(iommu);
3792 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
3797 struct dma_map_ops intel_dma_ops = {
3798 .alloc = intel_alloc_coherent,
3799 .free = intel_free_coherent,
3800 .map_sg = intel_map_sg,
3801 .unmap_sg = intel_unmap_sg,
3802 .map_page = intel_map_page,
3803 .unmap_page = intel_unmap_page,
3804 .mapping_error = intel_mapping_error,
3807 static inline int iommu_domain_cache_init(void)
3811 iommu_domain_cache = kmem_cache_create("iommu_domain",
3812 sizeof(struct dmar_domain),
3817 if (!iommu_domain_cache) {
3818 pr_err("Couldn't create iommu_domain cache\n");
3825 static inline int iommu_devinfo_cache_init(void)
3829 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
3830 sizeof(struct device_domain_info),
3834 if (!iommu_devinfo_cache) {
3835 pr_err("Couldn't create devinfo cache\n");
3842 static int __init iommu_init_mempool(void)
3845 ret = iova_cache_get();
3849 ret = iommu_domain_cache_init();
3853 ret = iommu_devinfo_cache_init();
3857 kmem_cache_destroy(iommu_domain_cache);
3864 static void __init iommu_exit_mempool(void)
3866 kmem_cache_destroy(iommu_devinfo_cache);
3867 kmem_cache_destroy(iommu_domain_cache);
3871 static void quirk_ioat_snb_local_iommu(struct pci_dev *pdev)
3873 struct dmar_drhd_unit *drhd;
3877 /* We know that this device on this chipset has its own IOMMU.
3878 * If we find it under a different IOMMU, then the BIOS is lying
3879 * to us. Hope that the IOMMU for this device is actually
3880 * disabled, and it needs no translation...
3882 rc = pci_bus_read_config_dword(pdev->bus, PCI_DEVFN(0, 0), 0xb0, &vtbar);
3884 /* "can't" happen */
3885 dev_info(&pdev->dev, "failed to run vt-d quirk\n");
3888 vtbar &= 0xffff0000;
3890 /* we know that the this iommu should be at offset 0xa000 from vtbar */
3891 drhd = dmar_find_matched_drhd_unit(pdev);
3892 if (WARN_TAINT_ONCE(!drhd || drhd->reg_base_addr - vtbar != 0xa000,
3893 TAINT_FIRMWARE_WORKAROUND,
3894 "BIOS assigned incorrect VT-d unit for Intel(R) QuickData Technology device\n"))
3895 pdev->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3897 DECLARE_PCI_FIXUP_ENABLE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB, quirk_ioat_snb_local_iommu);
3899 static void __init init_no_remapping_devices(void)
3901 struct dmar_drhd_unit *drhd;
3905 for_each_drhd_unit(drhd) {
3906 if (!drhd->include_all) {
3907 for_each_active_dev_scope(drhd->devices,
3908 drhd->devices_cnt, i, dev)
3910 /* ignore DMAR unit if no devices exist */
3911 if (i == drhd->devices_cnt)
3916 for_each_active_drhd_unit(drhd) {
3917 if (drhd->include_all)
3920 for_each_active_dev_scope(drhd->devices,
3921 drhd->devices_cnt, i, dev)
3922 if (!dev_is_pci(dev) || !IS_GFX_DEVICE(to_pci_dev(dev)))
3924 if (i < drhd->devices_cnt)
3927 /* This IOMMU has *only* gfx devices. Either bypass it or
3928 set the gfx_mapped flag, as appropriate */
3930 intel_iommu_gfx_mapped = 1;
3933 for_each_active_dev_scope(drhd->devices,
3934 drhd->devices_cnt, i, dev)
3935 dev->archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3940 #ifdef CONFIG_SUSPEND
3941 static int init_iommu_hw(void)
3943 struct dmar_drhd_unit *drhd;
3944 struct intel_iommu *iommu = NULL;
3946 for_each_active_iommu(iommu, drhd)
3948 dmar_reenable_qi(iommu);
3950 for_each_iommu(iommu, drhd) {
3951 if (drhd->ignored) {
3953 * we always have to disable PMRs or DMA may fail on
3957 iommu_disable_protect_mem_regions(iommu);
3961 iommu_flush_write_buffer(iommu);
3963 iommu_set_root_entry(iommu);
3965 iommu->flush.flush_context(iommu, 0, 0, 0,
3966 DMA_CCMD_GLOBAL_INVL);
3967 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
3968 iommu_enable_translation(iommu);
3969 iommu_disable_protect_mem_regions(iommu);
3975 static void iommu_flush_all(void)
3977 struct dmar_drhd_unit *drhd;
3978 struct intel_iommu *iommu;
3980 for_each_active_iommu(iommu, drhd) {
3981 iommu->flush.flush_context(iommu, 0, 0, 0,
3982 DMA_CCMD_GLOBAL_INVL);
3983 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3984 DMA_TLB_GLOBAL_FLUSH);
3988 static int iommu_suspend(void)
3990 struct dmar_drhd_unit *drhd;
3991 struct intel_iommu *iommu = NULL;
3994 for_each_active_iommu(iommu, drhd) {
3995 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3997 if (!iommu->iommu_state)
4003 for_each_active_iommu(iommu, drhd) {
4004 iommu_disable_translation(iommu);
4006 raw_spin_lock_irqsave(&iommu->register_lock, flag);
4008 iommu->iommu_state[SR_DMAR_FECTL_REG] =
4009 readl(iommu->reg + DMAR_FECTL_REG);
4010 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
4011 readl(iommu->reg + DMAR_FEDATA_REG);
4012 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
4013 readl(iommu->reg + DMAR_FEADDR_REG);
4014 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
4015 readl(iommu->reg + DMAR_FEUADDR_REG);
4017 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
4022 for_each_active_iommu(iommu, drhd)
4023 kfree(iommu->iommu_state);
4028 static void iommu_resume(void)
4030 struct dmar_drhd_unit *drhd;
4031 struct intel_iommu *iommu = NULL;
4034 if (init_iommu_hw()) {
4036 panic("tboot: IOMMU setup failed, DMAR can not resume!\n");
4038 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
4042 for_each_active_iommu(iommu, drhd) {
4044 raw_spin_lock_irqsave(&iommu->register_lock, flag);
4046 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
4047 iommu->reg + DMAR_FECTL_REG);
4048 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
4049 iommu->reg + DMAR_FEDATA_REG);
4050 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
4051 iommu->reg + DMAR_FEADDR_REG);
4052 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
4053 iommu->reg + DMAR_FEUADDR_REG);
4055 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
4058 for_each_active_iommu(iommu, drhd)
4059 kfree(iommu->iommu_state);
4062 static struct syscore_ops iommu_syscore_ops = {
4063 .resume = iommu_resume,
4064 .suspend = iommu_suspend,
4067 static void __init init_iommu_pm_ops(void)
4069 register_syscore_ops(&iommu_syscore_ops);
4073 static inline void init_iommu_pm_ops(void) {}
4074 #endif /* CONFIG_PM */
4077 int __init dmar_parse_one_rmrr(struct acpi_dmar_header *header, void *arg)
4079 struct acpi_dmar_reserved_memory *rmrr;
4080 struct dmar_rmrr_unit *rmrru;
4082 rmrru = kzalloc(sizeof(*rmrru), GFP_KERNEL);
4086 rmrru->hdr = header;
4087 rmrr = (struct acpi_dmar_reserved_memory *)header;
4088 rmrru->base_address = rmrr->base_address;
4089 rmrru->end_address = rmrr->end_address;
4090 rmrru->devices = dmar_alloc_dev_scope((void *)(rmrr + 1),
4091 ((void *)rmrr) + rmrr->header.length,
4092 &rmrru->devices_cnt);
4093 if (rmrru->devices_cnt && rmrru->devices == NULL) {
4098 list_add(&rmrru->list, &dmar_rmrr_units);
4103 static struct dmar_atsr_unit *dmar_find_atsr(struct acpi_dmar_atsr *atsr)
4105 struct dmar_atsr_unit *atsru;
4106 struct acpi_dmar_atsr *tmp;
4108 list_for_each_entry_rcu(atsru, &dmar_atsr_units, list) {
4109 tmp = (struct acpi_dmar_atsr *)atsru->hdr;
4110 if (atsr->segment != tmp->segment)
4112 if (atsr->header.length != tmp->header.length)
4114 if (memcmp(atsr, tmp, atsr->header.length) == 0)
4121 int dmar_parse_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4123 struct acpi_dmar_atsr *atsr;
4124 struct dmar_atsr_unit *atsru;
4126 if (system_state != SYSTEM_BOOTING && !intel_iommu_enabled)
4129 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4130 atsru = dmar_find_atsr(atsr);
4134 atsru = kzalloc(sizeof(*atsru) + hdr->length, GFP_KERNEL);
4139 * If memory is allocated from slab by ACPI _DSM method, we need to
4140 * copy the memory content because the memory buffer will be freed
4143 atsru->hdr = (void *)(atsru + 1);
4144 memcpy(atsru->hdr, hdr, hdr->length);
4145 atsru->include_all = atsr->flags & 0x1;
4146 if (!atsru->include_all) {
4147 atsru->devices = dmar_alloc_dev_scope((void *)(atsr + 1),
4148 (void *)atsr + atsr->header.length,
4149 &atsru->devices_cnt);
4150 if (atsru->devices_cnt && atsru->devices == NULL) {
4156 list_add_rcu(&atsru->list, &dmar_atsr_units);
4161 static void intel_iommu_free_atsr(struct dmar_atsr_unit *atsru)
4163 dmar_free_dev_scope(&atsru->devices, &atsru->devices_cnt);
4167 int dmar_release_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4169 struct acpi_dmar_atsr *atsr;
4170 struct dmar_atsr_unit *atsru;
4172 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4173 atsru = dmar_find_atsr(atsr);
4175 list_del_rcu(&atsru->list);
4177 intel_iommu_free_atsr(atsru);
4183 int dmar_check_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4187 struct acpi_dmar_atsr *atsr;
4188 struct dmar_atsr_unit *atsru;
4190 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4191 atsru = dmar_find_atsr(atsr);
4195 if (!atsru->include_all && atsru->devices && atsru->devices_cnt) {
4196 for_each_active_dev_scope(atsru->devices, atsru->devices_cnt,
4204 static int intel_iommu_add(struct dmar_drhd_unit *dmaru)
4207 struct intel_iommu *iommu = dmaru->iommu;
4209 if (g_iommus[iommu->seq_id])
4212 if (hw_pass_through && !ecap_pass_through(iommu->ecap)) {
4213 pr_warn("%s: Doesn't support hardware pass through.\n",
4217 if (!ecap_sc_support(iommu->ecap) &&
4218 domain_update_iommu_snooping(iommu)) {
4219 pr_warn("%s: Doesn't support snooping.\n",
4223 sp = domain_update_iommu_superpage(iommu) - 1;
4224 if (sp >= 0 && !(cap_super_page_val(iommu->cap) & (1 << sp))) {
4225 pr_warn("%s: Doesn't support large page.\n",
4231 * Disable translation if already enabled prior to OS handover.
4233 if (iommu->gcmd & DMA_GCMD_TE)
4234 iommu_disable_translation(iommu);
4236 g_iommus[iommu->seq_id] = iommu;
4237 ret = iommu_init_domains(iommu);
4239 ret = iommu_alloc_root_entry(iommu);
4243 #ifdef CONFIG_INTEL_IOMMU_SVM
4244 if (pasid_enabled(iommu))
4245 intel_svm_alloc_pasid_tables(iommu);
4248 if (dmaru->ignored) {
4250 * we always have to disable PMRs or DMA may fail on this device
4253 iommu_disable_protect_mem_regions(iommu);
4257 intel_iommu_init_qi(iommu);
4258 iommu_flush_write_buffer(iommu);
4260 #ifdef CONFIG_INTEL_IOMMU_SVM
4261 if (pasid_enabled(iommu) && ecap_prs(iommu->ecap)) {
4262 ret = intel_svm_enable_prq(iommu);
4267 ret = dmar_set_interrupt(iommu);
4271 iommu_set_root_entry(iommu);
4272 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
4273 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
4274 iommu_enable_translation(iommu);
4276 iommu_disable_protect_mem_regions(iommu);
4280 disable_dmar_iommu(iommu);
4282 free_dmar_iommu(iommu);
4286 int dmar_iommu_hotplug(struct dmar_drhd_unit *dmaru, bool insert)
4289 struct intel_iommu *iommu = dmaru->iommu;
4291 if (!intel_iommu_enabled)
4297 ret = intel_iommu_add(dmaru);
4299 disable_dmar_iommu(iommu);
4300 free_dmar_iommu(iommu);
4306 static void intel_iommu_free_dmars(void)
4308 struct dmar_rmrr_unit *rmrru, *rmrr_n;
4309 struct dmar_atsr_unit *atsru, *atsr_n;
4311 list_for_each_entry_safe(rmrru, rmrr_n, &dmar_rmrr_units, list) {
4312 list_del(&rmrru->list);
4313 dmar_free_dev_scope(&rmrru->devices, &rmrru->devices_cnt);
4317 list_for_each_entry_safe(atsru, atsr_n, &dmar_atsr_units, list) {
4318 list_del(&atsru->list);
4319 intel_iommu_free_atsr(atsru);
4323 int dmar_find_matched_atsr_unit(struct pci_dev *dev)
4326 struct pci_bus *bus;
4327 struct pci_dev *bridge = NULL;
4329 struct acpi_dmar_atsr *atsr;
4330 struct dmar_atsr_unit *atsru;
4332 dev = pci_physfn(dev);
4333 for (bus = dev->bus; bus; bus = bus->parent) {
4335 /* If it's an integrated device, allow ATS */
4338 /* Connected via non-PCIe: no ATS */
4339 if (!pci_is_pcie(bridge) ||
4340 pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE)
4342 /* If we found the root port, look it up in the ATSR */
4343 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_ROOT_PORT)
4348 list_for_each_entry_rcu(atsru, &dmar_atsr_units, list) {
4349 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
4350 if (atsr->segment != pci_domain_nr(dev->bus))
4353 for_each_dev_scope(atsru->devices, atsru->devices_cnt, i, tmp)
4354 if (tmp == &bridge->dev)
4357 if (atsru->include_all)
4367 int dmar_iommu_notify_scope_dev(struct dmar_pci_notify_info *info)
4370 struct dmar_rmrr_unit *rmrru;
4371 struct dmar_atsr_unit *atsru;
4372 struct acpi_dmar_atsr *atsr;
4373 struct acpi_dmar_reserved_memory *rmrr;
4375 if (!intel_iommu_enabled && system_state != SYSTEM_BOOTING)
4378 list_for_each_entry(rmrru, &dmar_rmrr_units, list) {
4379 rmrr = container_of(rmrru->hdr,
4380 struct acpi_dmar_reserved_memory, header);
4381 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
4382 ret = dmar_insert_dev_scope(info, (void *)(rmrr + 1),
4383 ((void *)rmrr) + rmrr->header.length,
4384 rmrr->segment, rmrru->devices,
4385 rmrru->devices_cnt);
4388 } else if (info->event == BUS_NOTIFY_REMOVED_DEVICE) {
4389 dmar_remove_dev_scope(info, rmrr->segment,
4390 rmrru->devices, rmrru->devices_cnt);
4394 list_for_each_entry(atsru, &dmar_atsr_units, list) {
4395 if (atsru->include_all)
4398 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
4399 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
4400 ret = dmar_insert_dev_scope(info, (void *)(atsr + 1),
4401 (void *)atsr + atsr->header.length,
4402 atsr->segment, atsru->devices,
4403 atsru->devices_cnt);
4408 } else if (info->event == BUS_NOTIFY_REMOVED_DEVICE) {
4409 if (dmar_remove_dev_scope(info, atsr->segment,
4410 atsru->devices, atsru->devices_cnt))
4419 * Here we only respond to action of unbound device from driver.
4421 * Added device is not attached to its DMAR domain here yet. That will happen
4422 * when mapping the device to iova.
4424 static int device_notifier(struct notifier_block *nb,
4425 unsigned long action, void *data)
4427 struct device *dev = data;
4428 struct dmar_domain *domain;
4430 if (iommu_dummy(dev))
4433 if (action != BUS_NOTIFY_REMOVED_DEVICE)
4436 domain = find_domain(dev);
4440 dmar_remove_one_dev_info(domain, dev);
4441 if (!domain_type_is_vm_or_si(domain) && list_empty(&domain->devices))
4442 domain_exit(domain);
4447 static struct notifier_block device_nb = {
4448 .notifier_call = device_notifier,
4451 static int intel_iommu_memory_notifier(struct notifier_block *nb,
4452 unsigned long val, void *v)
4454 struct memory_notify *mhp = v;
4455 unsigned long long start, end;
4456 unsigned long start_vpfn, last_vpfn;
4459 case MEM_GOING_ONLINE:
4460 start = mhp->start_pfn << PAGE_SHIFT;
4461 end = ((mhp->start_pfn + mhp->nr_pages) << PAGE_SHIFT) - 1;
4462 if (iommu_domain_identity_map(si_domain, start, end)) {
4463 pr_warn("Failed to build identity map for [%llx-%llx]\n",
4470 case MEM_CANCEL_ONLINE:
4471 start_vpfn = mm_to_dma_pfn(mhp->start_pfn);
4472 last_vpfn = mm_to_dma_pfn(mhp->start_pfn + mhp->nr_pages - 1);
4473 while (start_vpfn <= last_vpfn) {
4475 struct dmar_drhd_unit *drhd;
4476 struct intel_iommu *iommu;
4477 struct page *freelist;
4479 iova = find_iova(&si_domain->iovad, start_vpfn);
4481 pr_debug("Failed get IOVA for PFN %lx\n",
4486 iova = split_and_remove_iova(&si_domain->iovad, iova,
4487 start_vpfn, last_vpfn);
4489 pr_warn("Failed to split IOVA PFN [%lx-%lx]\n",
4490 start_vpfn, last_vpfn);
4494 freelist = domain_unmap(si_domain, iova->pfn_lo,
4498 for_each_active_iommu(iommu, drhd)
4499 iommu_flush_iotlb_psi(iommu, si_domain,
4500 iova->pfn_lo, iova_size(iova),
4503 dma_free_pagelist(freelist);
4505 start_vpfn = iova->pfn_hi + 1;
4506 free_iova_mem(iova);
4514 static struct notifier_block intel_iommu_memory_nb = {
4515 .notifier_call = intel_iommu_memory_notifier,
4520 static ssize_t intel_iommu_show_version(struct device *dev,
4521 struct device_attribute *attr,
4524 struct intel_iommu *iommu = dev_get_drvdata(dev);
4525 u32 ver = readl(iommu->reg + DMAR_VER_REG);
4526 return sprintf(buf, "%d:%d\n",
4527 DMAR_VER_MAJOR(ver), DMAR_VER_MINOR(ver));
4529 static DEVICE_ATTR(version, S_IRUGO, intel_iommu_show_version, NULL);
4531 static ssize_t intel_iommu_show_address(struct device *dev,
4532 struct device_attribute *attr,
4535 struct intel_iommu *iommu = dev_get_drvdata(dev);
4536 return sprintf(buf, "%llx\n", iommu->reg_phys);
4538 static DEVICE_ATTR(address, S_IRUGO, intel_iommu_show_address, NULL);
4540 static ssize_t intel_iommu_show_cap(struct device *dev,
4541 struct device_attribute *attr,
4544 struct intel_iommu *iommu = dev_get_drvdata(dev);
4545 return sprintf(buf, "%llx\n", iommu->cap);
4547 static DEVICE_ATTR(cap, S_IRUGO, intel_iommu_show_cap, NULL);
4549 static ssize_t intel_iommu_show_ecap(struct device *dev,
4550 struct device_attribute *attr,
4553 struct intel_iommu *iommu = dev_get_drvdata(dev);
4554 return sprintf(buf, "%llx\n", iommu->ecap);
4556 static DEVICE_ATTR(ecap, S_IRUGO, intel_iommu_show_ecap, NULL);
4558 static ssize_t intel_iommu_show_ndoms(struct device *dev,
4559 struct device_attribute *attr,
4562 struct intel_iommu *iommu = dev_get_drvdata(dev);
4563 return sprintf(buf, "%ld\n", cap_ndoms(iommu->cap));
4565 static DEVICE_ATTR(domains_supported, S_IRUGO, intel_iommu_show_ndoms, NULL);
4567 static ssize_t intel_iommu_show_ndoms_used(struct device *dev,
4568 struct device_attribute *attr,
4571 struct intel_iommu *iommu = dev_get_drvdata(dev);
4572 return sprintf(buf, "%d\n", bitmap_weight(iommu->domain_ids,
4573 cap_ndoms(iommu->cap)));
4575 static DEVICE_ATTR(domains_used, S_IRUGO, intel_iommu_show_ndoms_used, NULL);
4577 static struct attribute *intel_iommu_attrs[] = {
4578 &dev_attr_version.attr,
4579 &dev_attr_address.attr,
4581 &dev_attr_ecap.attr,
4582 &dev_attr_domains_supported.attr,
4583 &dev_attr_domains_used.attr,
4587 static struct attribute_group intel_iommu_group = {
4588 .name = "intel-iommu",
4589 .attrs = intel_iommu_attrs,
4592 const struct attribute_group *intel_iommu_groups[] = {
4597 int __init intel_iommu_init(void)
4600 struct dmar_drhd_unit *drhd;
4601 struct intel_iommu *iommu;
4603 /* VT-d is required for a TXT/tboot launch, so enforce that */
4604 force_on = tboot_force_iommu();
4606 if (iommu_init_mempool()) {
4608 panic("tboot: Failed to initialize iommu memory\n");
4612 down_write(&dmar_global_lock);
4613 if (dmar_table_init()) {
4615 panic("tboot: Failed to initialize DMAR table\n");
4619 if (dmar_dev_scope_init() < 0) {
4621 panic("tboot: Failed to initialize DMAR device scope\n");
4625 if (no_iommu || dmar_disabled)
4628 if (list_empty(&dmar_rmrr_units))
4629 pr_info("No RMRR found\n");
4631 if (list_empty(&dmar_atsr_units))
4632 pr_info("No ATSR found\n");
4634 if (dmar_init_reserved_ranges()) {
4636 panic("tboot: Failed to reserve iommu ranges\n");
4637 goto out_free_reserved_range;
4640 init_no_remapping_devices();
4645 panic("tboot: Failed to initialize DMARs\n");
4646 pr_err("Initialization failed\n");
4647 goto out_free_reserved_range;
4649 up_write(&dmar_global_lock);
4650 pr_info("Intel(R) Virtualization Technology for Directed I/O\n");
4652 init_timer(&unmap_timer);
4653 #ifdef CONFIG_SWIOTLB
4656 dma_ops = &intel_dma_ops;
4658 init_iommu_pm_ops();
4660 for_each_active_iommu(iommu, drhd)
4661 iommu->iommu_dev = iommu_device_create(NULL, iommu,
4665 bus_set_iommu(&pci_bus_type, &intel_iommu_ops);
4666 bus_register_notifier(&pci_bus_type, &device_nb);
4667 if (si_domain && !hw_pass_through)
4668 register_memory_notifier(&intel_iommu_memory_nb);
4670 intel_iommu_enabled = 1;
4674 out_free_reserved_range:
4675 put_iova_domain(&reserved_iova_list);
4677 intel_iommu_free_dmars();
4678 up_write(&dmar_global_lock);
4679 iommu_exit_mempool();
4683 static int domain_context_clear_one_cb(struct pci_dev *pdev, u16 alias, void *opaque)
4685 struct intel_iommu *iommu = opaque;
4687 domain_context_clear_one(iommu, PCI_BUS_NUM(alias), alias & 0xff);
4692 * NB - intel-iommu lacks any sort of reference counting for the users of
4693 * dependent devices. If multiple endpoints have intersecting dependent
4694 * devices, unbinding the driver from any one of them will possibly leave
4695 * the others unable to operate.
4697 static void domain_context_clear(struct intel_iommu *iommu, struct device *dev)
4699 if (!iommu || !dev || !dev_is_pci(dev))
4702 pci_for_each_dma_alias(to_pci_dev(dev), &domain_context_clear_one_cb, iommu);
4705 static void __dmar_remove_one_dev_info(struct device_domain_info *info)
4707 struct intel_iommu *iommu;
4708 unsigned long flags;
4710 assert_spin_locked(&device_domain_lock);
4715 iommu = info->iommu;
4718 iommu_disable_dev_iotlb(info);
4719 domain_context_clear(iommu, info->dev);
4722 unlink_domain_info(info);
4724 spin_lock_irqsave(&iommu->lock, flags);
4725 domain_detach_iommu(info->domain, iommu);
4726 spin_unlock_irqrestore(&iommu->lock, flags);
4728 free_devinfo_mem(info);
4731 static void dmar_remove_one_dev_info(struct dmar_domain *domain,
4734 struct device_domain_info *info;
4735 unsigned long flags;
4737 spin_lock_irqsave(&device_domain_lock, flags);
4738 info = dev->archdata.iommu;
4739 __dmar_remove_one_dev_info(info);
4740 spin_unlock_irqrestore(&device_domain_lock, flags);
4743 static int md_domain_init(struct dmar_domain *domain, int guest_width)
4747 init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
4749 domain_reserve_special_ranges(domain);
4751 /* calculate AGAW */
4752 domain->gaw = guest_width;
4753 adjust_width = guestwidth_to_adjustwidth(guest_width);
4754 domain->agaw = width_to_agaw(adjust_width);
4756 domain->iommu_coherency = 0;
4757 domain->iommu_snooping = 0;
4758 domain->iommu_superpage = 0;
4759 domain->max_addr = 0;
4761 /* always allocate the top pgd */
4762 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
4765 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
4769 static struct iommu_domain *intel_iommu_domain_alloc(unsigned type)
4771 struct dmar_domain *dmar_domain;
4772 struct iommu_domain *domain;
4774 if (type != IOMMU_DOMAIN_UNMANAGED)
4777 dmar_domain = alloc_domain(DOMAIN_FLAG_VIRTUAL_MACHINE);
4779 pr_err("Can't allocate dmar_domain\n");
4782 if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
4783 pr_err("Domain initialization failed\n");
4784 domain_exit(dmar_domain);
4787 domain_update_iommu_cap(dmar_domain);
4789 domain = &dmar_domain->domain;
4790 domain->geometry.aperture_start = 0;
4791 domain->geometry.aperture_end = __DOMAIN_MAX_ADDR(dmar_domain->gaw);
4792 domain->geometry.force_aperture = true;
4797 static void intel_iommu_domain_free(struct iommu_domain *domain)
4799 domain_exit(to_dmar_domain(domain));
4802 static int intel_iommu_attach_device(struct iommu_domain *domain,
4805 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4806 struct intel_iommu *iommu;
4810 if (device_is_rmrr_locked(dev)) {
4811 dev_warn(dev, "Device is ineligible for IOMMU domain attach due to platform RMRR requirement. Contact your platform vendor.\n");
4815 /* normally dev is not mapped */
4816 if (unlikely(domain_context_mapped(dev))) {
4817 struct dmar_domain *old_domain;
4819 old_domain = find_domain(dev);
4822 dmar_remove_one_dev_info(old_domain, dev);
4825 if (!domain_type_is_vm_or_si(old_domain) &&
4826 list_empty(&old_domain->devices))
4827 domain_exit(old_domain);
4831 iommu = device_to_iommu(dev, &bus, &devfn);
4835 /* check if this iommu agaw is sufficient for max mapped address */
4836 addr_width = agaw_to_width(iommu->agaw);
4837 if (addr_width > cap_mgaw(iommu->cap))
4838 addr_width = cap_mgaw(iommu->cap);
4840 if (dmar_domain->max_addr > (1LL << addr_width)) {
4841 pr_err("%s: iommu width (%d) is not "
4842 "sufficient for the mapped address (%llx)\n",
4843 __func__, addr_width, dmar_domain->max_addr);
4846 dmar_domain->gaw = addr_width;
4849 * Knock out extra levels of page tables if necessary
4851 while (iommu->agaw < dmar_domain->agaw) {
4852 struct dma_pte *pte;
4854 pte = dmar_domain->pgd;
4855 if (dma_pte_present(pte)) {
4856 dmar_domain->pgd = (struct dma_pte *)
4857 phys_to_virt(dma_pte_addr(pte));
4858 free_pgtable_page(pte);
4860 dmar_domain->agaw--;
4863 return domain_add_dev_info(dmar_domain, dev);
4866 static void intel_iommu_detach_device(struct iommu_domain *domain,
4869 dmar_remove_one_dev_info(to_dmar_domain(domain), dev);
4872 static int intel_iommu_map(struct iommu_domain *domain,
4873 unsigned long iova, phys_addr_t hpa,
4874 size_t size, int iommu_prot)
4876 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4881 if (iommu_prot & IOMMU_READ)
4882 prot |= DMA_PTE_READ;
4883 if (iommu_prot & IOMMU_WRITE)
4884 prot |= DMA_PTE_WRITE;
4885 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
4886 prot |= DMA_PTE_SNP;
4888 max_addr = iova + size;
4889 if (dmar_domain->max_addr < max_addr) {
4892 /* check if minimum agaw is sufficient for mapped address */
4893 end = __DOMAIN_MAX_ADDR(dmar_domain->gaw) + 1;
4894 if (end < max_addr) {
4895 pr_err("%s: iommu width (%d) is not "
4896 "sufficient for the mapped address (%llx)\n",
4897 __func__, dmar_domain->gaw, max_addr);
4900 dmar_domain->max_addr = max_addr;
4902 /* Round up size to next multiple of PAGE_SIZE, if it and
4903 the low bits of hpa would take us onto the next page */
4904 size = aligned_nrpages(hpa, size);
4905 ret = domain_pfn_mapping(dmar_domain, iova >> VTD_PAGE_SHIFT,
4906 hpa >> VTD_PAGE_SHIFT, size, prot);
4910 static size_t intel_iommu_unmap(struct iommu_domain *domain,
4911 unsigned long iova, size_t size)
4913 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4914 struct page *freelist = NULL;
4915 struct intel_iommu *iommu;
4916 unsigned long start_pfn, last_pfn;
4917 unsigned int npages;
4918 int iommu_id, level = 0;
4920 /* Cope with horrid API which requires us to unmap more than the
4921 size argument if it happens to be a large-page mapping. */
4922 BUG_ON(!pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level));
4924 if (size < VTD_PAGE_SIZE << level_to_offset_bits(level))
4925 size = VTD_PAGE_SIZE << level_to_offset_bits(level);
4927 start_pfn = iova >> VTD_PAGE_SHIFT;
4928 last_pfn = (iova + size - 1) >> VTD_PAGE_SHIFT;
4930 freelist = domain_unmap(dmar_domain, start_pfn, last_pfn);
4932 npages = last_pfn - start_pfn + 1;
4934 for_each_domain_iommu(iommu_id, dmar_domain) {
4935 iommu = g_iommus[iommu_id];
4937 iommu_flush_iotlb_psi(g_iommus[iommu_id], dmar_domain,
4938 start_pfn, npages, !freelist, 0);
4941 dma_free_pagelist(freelist);
4943 if (dmar_domain->max_addr == iova + size)
4944 dmar_domain->max_addr = iova;
4949 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
4952 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4953 struct dma_pte *pte;
4957 pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level);
4959 phys = dma_pte_addr(pte);
4964 static bool intel_iommu_capable(enum iommu_cap cap)
4966 if (cap == IOMMU_CAP_CACHE_COHERENCY)
4967 return domain_update_iommu_snooping(NULL) == 1;
4968 if (cap == IOMMU_CAP_INTR_REMAP)
4969 return irq_remapping_enabled == 1;
4974 static int intel_iommu_add_device(struct device *dev)
4976 struct intel_iommu *iommu;
4977 struct iommu_group *group;
4980 iommu = device_to_iommu(dev, &bus, &devfn);
4984 iommu_device_link(iommu->iommu_dev, dev);
4986 group = iommu_group_get_for_dev(dev);
4989 return PTR_ERR(group);
4991 iommu_group_put(group);
4995 static void intel_iommu_remove_device(struct device *dev)
4997 struct intel_iommu *iommu;
5000 iommu = device_to_iommu(dev, &bus, &devfn);
5004 iommu_group_remove_device(dev);
5006 iommu_device_unlink(iommu->iommu_dev, dev);
5009 #ifdef CONFIG_INTEL_IOMMU_SVM
5010 int intel_iommu_enable_pasid(struct intel_iommu *iommu, struct intel_svm_dev *sdev)
5012 struct device_domain_info *info;
5013 struct context_entry *context;
5014 struct dmar_domain *domain;
5015 unsigned long flags;
5019 domain = get_valid_domain_for_dev(sdev->dev);
5023 spin_lock_irqsave(&device_domain_lock, flags);
5024 spin_lock(&iommu->lock);
5027 info = sdev->dev->archdata.iommu;
5028 if (!info || !info->pasid_supported)
5031 context = iommu_context_addr(iommu, info->bus, info->devfn, 0);
5032 if (WARN_ON(!context))
5035 ctx_lo = context[0].lo;
5037 sdev->did = domain->iommu_did[iommu->seq_id];
5038 sdev->sid = PCI_DEVID(info->bus, info->devfn);
5040 if (!(ctx_lo & CONTEXT_PASIDE)) {
5041 context[1].hi = (u64)virt_to_phys(iommu->pasid_state_table);
5042 context[1].lo = (u64)virt_to_phys(iommu->pasid_table) | ecap_pss(iommu->ecap);
5044 /* CONTEXT_TT_MULTI_LEVEL and CONTEXT_TT_DEV_IOTLB are both
5045 * extended to permit requests-with-PASID if the PASIDE bit
5046 * is set. which makes sense. For CONTEXT_TT_PASS_THROUGH,
5047 * however, the PASIDE bit is ignored and requests-with-PASID
5048 * are unconditionally blocked. Which makes less sense.
5049 * So convert from CONTEXT_TT_PASS_THROUGH to one of the new
5050 * "guest mode" translation types depending on whether ATS
5051 * is available or not. Annoyingly, we can't use the new
5052 * modes *unless* PASIDE is set. */
5053 if ((ctx_lo & CONTEXT_TT_MASK) == (CONTEXT_TT_PASS_THROUGH << 2)) {
5054 ctx_lo &= ~CONTEXT_TT_MASK;
5055 if (info->ats_supported)
5056 ctx_lo |= CONTEXT_TT_PT_PASID_DEV_IOTLB << 2;
5058 ctx_lo |= CONTEXT_TT_PT_PASID << 2;
5060 ctx_lo |= CONTEXT_PASIDE;
5061 if (iommu->pasid_state_table)
5062 ctx_lo |= CONTEXT_DINVE;
5063 if (info->pri_supported)
5064 ctx_lo |= CONTEXT_PRS;
5065 context[0].lo = ctx_lo;
5067 iommu->flush.flush_context(iommu, sdev->did, sdev->sid,
5068 DMA_CCMD_MASK_NOBIT,
5069 DMA_CCMD_DEVICE_INVL);
5072 /* Enable PASID support in the device, if it wasn't already */
5073 if (!info->pasid_enabled)
5074 iommu_enable_dev_iotlb(info);
5076 if (info->ats_enabled) {
5077 sdev->dev_iotlb = 1;
5078 sdev->qdep = info->ats_qdep;
5079 if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
5085 spin_unlock(&iommu->lock);
5086 spin_unlock_irqrestore(&device_domain_lock, flags);
5091 struct intel_iommu *intel_svm_device_to_iommu(struct device *dev)
5093 struct intel_iommu *iommu;
5096 if (iommu_dummy(dev)) {
5098 "No IOMMU translation for device; cannot enable SVM\n");
5102 iommu = device_to_iommu(dev, &bus, &devfn);
5104 dev_err(dev, "No IOMMU for device; cannot enable SVM\n");
5108 if (!iommu->pasid_table) {
5109 dev_err(dev, "PASID not enabled on IOMMU; cannot enable SVM\n");
5115 #endif /* CONFIG_INTEL_IOMMU_SVM */
5117 static const struct iommu_ops intel_iommu_ops = {
5118 .capable = intel_iommu_capable,
5119 .domain_alloc = intel_iommu_domain_alloc,
5120 .domain_free = intel_iommu_domain_free,
5121 .attach_dev = intel_iommu_attach_device,
5122 .detach_dev = intel_iommu_detach_device,
5123 .map = intel_iommu_map,
5124 .unmap = intel_iommu_unmap,
5125 .map_sg = default_iommu_map_sg,
5126 .iova_to_phys = intel_iommu_iova_to_phys,
5127 .add_device = intel_iommu_add_device,
5128 .remove_device = intel_iommu_remove_device,
5129 .device_group = pci_device_group,
5130 .pgsize_bitmap = INTEL_IOMMU_PGSIZES,
5133 static void quirk_iommu_g4x_gfx(struct pci_dev *dev)
5135 /* G4x/GM45 integrated gfx dmar support is totally busted. */
5136 pr_info("Disabling IOMMU for graphics on this chipset\n");
5140 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_g4x_gfx);
5141 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_g4x_gfx);
5142 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_g4x_gfx);
5143 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_g4x_gfx);
5144 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_g4x_gfx);
5145 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_g4x_gfx);
5146 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_g4x_gfx);
5148 static void quirk_iommu_rwbf(struct pci_dev *dev)
5151 * Mobile 4 Series Chipset neglects to set RWBF capability,
5152 * but needs it. Same seems to hold for the desktop versions.
5154 pr_info("Forcing write-buffer flush capability\n");
5158 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);
5159 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_rwbf);
5160 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_rwbf);
5161 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_rwbf);
5162 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_rwbf);
5163 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_rwbf);
5164 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_rwbf);
5167 #define GGC_MEMORY_SIZE_MASK (0xf << 8)
5168 #define GGC_MEMORY_SIZE_NONE (0x0 << 8)
5169 #define GGC_MEMORY_SIZE_1M (0x1 << 8)
5170 #define GGC_MEMORY_SIZE_2M (0x3 << 8)
5171 #define GGC_MEMORY_VT_ENABLED (0x8 << 8)
5172 #define GGC_MEMORY_SIZE_2M_VT (0x9 << 8)
5173 #define GGC_MEMORY_SIZE_3M_VT (0xa << 8)
5174 #define GGC_MEMORY_SIZE_4M_VT (0xb << 8)
5176 static void quirk_calpella_no_shadow_gtt(struct pci_dev *dev)
5180 if (pci_read_config_word(dev, GGC, &ggc))
5183 if (!(ggc & GGC_MEMORY_VT_ENABLED)) {
5184 pr_info("BIOS has allocated no shadow GTT; disabling IOMMU for graphics\n");
5186 } else if (dmar_map_gfx) {
5187 /* we have to ensure the gfx device is idle before we flush */
5188 pr_info("Disabling batched IOTLB flush on Ironlake\n");
5189 intel_iommu_strict = 1;
5192 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0040, quirk_calpella_no_shadow_gtt);
5193 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0044, quirk_calpella_no_shadow_gtt);
5194 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0062, quirk_calpella_no_shadow_gtt);
5195 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x006a, quirk_calpella_no_shadow_gtt);
5197 /* On Tylersburg chipsets, some BIOSes have been known to enable the
5198 ISOCH DMAR unit for the Azalia sound device, but not give it any
5199 TLB entries, which causes it to deadlock. Check for that. We do
5200 this in a function called from init_dmars(), instead of in a PCI
5201 quirk, because we don't want to print the obnoxious "BIOS broken"
5202 message if VT-d is actually disabled.
5204 static void __init check_tylersburg_isoch(void)
5206 struct pci_dev *pdev;
5207 uint32_t vtisochctrl;
5209 /* If there's no Azalia in the system anyway, forget it. */
5210 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x3a3e, NULL);
5215 /* System Management Registers. Might be hidden, in which case
5216 we can't do the sanity check. But that's OK, because the
5217 known-broken BIOSes _don't_ actually hide it, so far. */
5218 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x342e, NULL);
5222 if (pci_read_config_dword(pdev, 0x188, &vtisochctrl)) {
5229 /* If Azalia DMA is routed to the non-isoch DMAR unit, fine. */
5230 if (vtisochctrl & 1)
5233 /* Drop all bits other than the number of TLB entries */
5234 vtisochctrl &= 0x1c;
5236 /* If we have the recommended number of TLB entries (16), fine. */
5237 if (vtisochctrl == 0x10)
5240 /* Zero TLB entries? You get to ride the short bus to school. */
5242 WARN(1, "Your BIOS is broken; DMA routed to ISOCH DMAR unit but no TLB space.\n"
5243 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
5244 dmi_get_system_info(DMI_BIOS_VENDOR),
5245 dmi_get_system_info(DMI_BIOS_VERSION),
5246 dmi_get_system_info(DMI_PRODUCT_VERSION));
5247 iommu_identity_mapping |= IDENTMAP_AZALIA;
5251 pr_warn("Recommended TLB entries for ISOCH unit is 16; your BIOS set %d\n",