2 * Copyright (c) 2006, Intel Corporation.
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * You should have received a copy of the GNU General Public License along with
14 * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
15 * Place - Suite 330, Boston, MA 02111-1307 USA.
17 * Copyright (C) 2006-2008 Intel Corporation
18 * Author: Ashok Raj <ashok.raj@intel.com>
19 * Author: Shaohua Li <shaohua.li@intel.com>
20 * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
21 * Author: Fenghua Yu <fenghua.yu@intel.com>
24 #include <linux/init.h>
25 #include <linux/bitmap.h>
26 #include <linux/debugfs.h>
27 #include <linux/slab.h>
28 #include <linux/irq.h>
29 #include <linux/interrupt.h>
30 #include <linux/spinlock.h>
31 #include <linux/pci.h>
32 #include <linux/dmar.h>
33 #include <linux/dma-mapping.h>
34 #include <linux/mempool.h>
35 #include <linux/timer.h>
36 #include <linux/iova.h>
37 #include <linux/iommu.h>
38 #include <linux/intel-iommu.h>
39 #include <linux/sysdev.h>
40 #include <asm/cacheflush.h>
41 #include <asm/iommu.h>
44 #define ROOT_SIZE VTD_PAGE_SIZE
45 #define CONTEXT_SIZE VTD_PAGE_SIZE
47 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
48 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
50 #define IOAPIC_RANGE_START (0xfee00000)
51 #define IOAPIC_RANGE_END (0xfeefffff)
52 #define IOVA_START_ADDR (0x1000)
54 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
56 #define MAX_AGAW_WIDTH 64
58 #define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
60 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
61 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
62 #define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
64 /* global iommu list, set NULL for ignored DMAR units */
65 static struct intel_iommu **g_iommus;
67 static int rwbf_quirk;
72 * 12-63: Context Ptr (12 - (haw-1))
79 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
80 static inline bool root_present(struct root_entry *root)
82 return (root->val & 1);
84 static inline void set_root_present(struct root_entry *root)
88 static inline void set_root_value(struct root_entry *root, unsigned long value)
90 root->val |= value & VTD_PAGE_MASK;
93 static inline struct context_entry *
94 get_context_addr_from_root(struct root_entry *root)
96 return (struct context_entry *)
97 (root_present(root)?phys_to_virt(
98 root->val & VTD_PAGE_MASK) :
105 * 1: fault processing disable
106 * 2-3: translation type
107 * 12-63: address space root
113 struct context_entry {
118 static inline bool context_present(struct context_entry *context)
120 return (context->lo & 1);
122 static inline void context_set_present(struct context_entry *context)
127 static inline void context_set_fault_enable(struct context_entry *context)
129 context->lo &= (((u64)-1) << 2) | 1;
132 static inline void context_set_translation_type(struct context_entry *context,
135 context->lo &= (((u64)-1) << 4) | 3;
136 context->lo |= (value & 3) << 2;
139 static inline void context_set_address_root(struct context_entry *context,
142 context->lo |= value & VTD_PAGE_MASK;
145 static inline void context_set_address_width(struct context_entry *context,
148 context->hi |= value & 7;
151 static inline void context_set_domain_id(struct context_entry *context,
154 context->hi |= (value & ((1 << 16) - 1)) << 8;
157 static inline void context_clear_entry(struct context_entry *context)
170 * 12-63: Host physcial address
176 static inline void dma_clear_pte(struct dma_pte *pte)
181 static inline void dma_set_pte_readable(struct dma_pte *pte)
183 pte->val |= DMA_PTE_READ;
186 static inline void dma_set_pte_writable(struct dma_pte *pte)
188 pte->val |= DMA_PTE_WRITE;
191 static inline void dma_set_pte_snp(struct dma_pte *pte)
193 pte->val |= DMA_PTE_SNP;
196 static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)
198 pte->val = (pte->val & ~3) | (prot & 3);
201 static inline u64 dma_pte_addr(struct dma_pte *pte)
203 return (pte->val & VTD_PAGE_MASK);
206 static inline void dma_set_pte_addr(struct dma_pte *pte, u64 addr)
208 pte->val |= (addr & VTD_PAGE_MASK);
211 static inline bool dma_pte_present(struct dma_pte *pte)
213 return (pte->val & 3) != 0;
216 /* devices under the same p2p bridge are owned in one domain */
217 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
219 /* domain represents a virtual machine, more than one devices
220 * across iommus may be owned in one domain, e.g. kvm guest.
222 #define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 1)
225 int id; /* domain id */
226 unsigned long iommu_bmp; /* bitmap of iommus this domain uses*/
228 struct list_head devices; /* all devices' list */
229 struct iova_domain iovad; /* iova's that belong to this domain */
231 struct dma_pte *pgd; /* virtual address */
232 spinlock_t mapping_lock; /* page table lock */
233 int gaw; /* max guest address width */
235 /* adjusted guest address width, 0 is level 2 30-bit */
238 int flags; /* flags to find out type of domain */
240 int iommu_coherency;/* indicate coherency of iommu access */
241 int iommu_snooping; /* indicate snooping control feature*/
242 int iommu_count; /* reference count of iommu */
243 spinlock_t iommu_lock; /* protect iommu set in domain */
244 u64 max_addr; /* maximum mapped address */
247 /* PCI domain-device relationship */
248 struct device_domain_info {
249 struct list_head link; /* link to domain siblings */
250 struct list_head global; /* link to global list */
251 int segment; /* PCI domain */
252 u8 bus; /* PCI bus number */
253 u8 devfn; /* PCI devfn number */
254 struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
255 struct intel_iommu *iommu; /* IOMMU used by this device */
256 struct dmar_domain *domain; /* pointer to domain */
259 static void flush_unmaps_timeout(unsigned long data);
261 DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
263 #define HIGH_WATER_MARK 250
264 struct deferred_flush_tables {
266 struct iova *iova[HIGH_WATER_MARK];
267 struct dmar_domain *domain[HIGH_WATER_MARK];
270 static struct deferred_flush_tables *deferred_flush;
272 /* bitmap for indexing intel_iommus */
273 static int g_num_of_iommus;
275 static DEFINE_SPINLOCK(async_umap_flush_lock);
276 static LIST_HEAD(unmaps_to_do);
279 static long list_size;
281 static void domain_remove_dev_info(struct dmar_domain *domain);
283 #ifdef CONFIG_DMAR_DEFAULT_ON
284 int dmar_disabled = 0;
286 int dmar_disabled = 1;
287 #endif /*CONFIG_DMAR_DEFAULT_ON*/
289 static int __initdata dmar_map_gfx = 1;
290 static int dmar_forcedac;
291 static int intel_iommu_strict;
293 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
294 static DEFINE_SPINLOCK(device_domain_lock);
295 static LIST_HEAD(device_domain_list);
297 static struct iommu_ops intel_iommu_ops;
299 static int __init intel_iommu_setup(char *str)
304 if (!strncmp(str, "on", 2)) {
306 printk(KERN_INFO "Intel-IOMMU: enabled\n");
307 } else if (!strncmp(str, "off", 3)) {
309 printk(KERN_INFO "Intel-IOMMU: disabled\n");
310 } else if (!strncmp(str, "igfx_off", 8)) {
313 "Intel-IOMMU: disable GFX device mapping\n");
314 } else if (!strncmp(str, "forcedac", 8)) {
316 "Intel-IOMMU: Forcing DAC for PCI devices\n");
318 } else if (!strncmp(str, "strict", 6)) {
320 "Intel-IOMMU: disable batched IOTLB flush\n");
321 intel_iommu_strict = 1;
324 str += strcspn(str, ",");
330 __setup("intel_iommu=", intel_iommu_setup);
332 static struct kmem_cache *iommu_domain_cache;
333 static struct kmem_cache *iommu_devinfo_cache;
334 static struct kmem_cache *iommu_iova_cache;
336 static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
341 /* trying to avoid low memory issues */
342 flags = current->flags & PF_MEMALLOC;
343 current->flags |= PF_MEMALLOC;
344 vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
345 current->flags &= (~PF_MEMALLOC | flags);
350 static inline void *alloc_pgtable_page(void)
355 /* trying to avoid low memory issues */
356 flags = current->flags & PF_MEMALLOC;
357 current->flags |= PF_MEMALLOC;
358 vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
359 current->flags &= (~PF_MEMALLOC | flags);
363 static inline void free_pgtable_page(void *vaddr)
365 free_page((unsigned long)vaddr);
368 static inline void *alloc_domain_mem(void)
370 return iommu_kmem_cache_alloc(iommu_domain_cache);
373 static void free_domain_mem(void *vaddr)
375 kmem_cache_free(iommu_domain_cache, vaddr);
378 static inline void * alloc_devinfo_mem(void)
380 return iommu_kmem_cache_alloc(iommu_devinfo_cache);
383 static inline void free_devinfo_mem(void *vaddr)
385 kmem_cache_free(iommu_devinfo_cache, vaddr);
388 struct iova *alloc_iova_mem(void)
390 return iommu_kmem_cache_alloc(iommu_iova_cache);
393 void free_iova_mem(struct iova *iova)
395 kmem_cache_free(iommu_iova_cache, iova);
399 static inline int width_to_agaw(int width);
401 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
406 sagaw = cap_sagaw(iommu->cap);
407 for (agaw = width_to_agaw(max_gaw);
409 if (test_bit(agaw, &sagaw))
417 * Calculate max SAGAW for each iommu.
419 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
421 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
425 * calculate agaw for each iommu.
426 * "SAGAW" may be different across iommus, use a default agaw, and
427 * get a supported less agaw for iommus that don't support the default agaw.
429 int iommu_calculate_agaw(struct intel_iommu *iommu)
431 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
434 /* in native case, each domain is related to only one iommu */
435 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
439 BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
441 iommu_id = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
442 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
445 return g_iommus[iommu_id];
448 static void domain_update_iommu_coherency(struct dmar_domain *domain)
452 domain->iommu_coherency = 1;
454 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
455 for (; i < g_num_of_iommus; ) {
456 if (!ecap_coherent(g_iommus[i]->ecap)) {
457 domain->iommu_coherency = 0;
460 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
464 static void domain_update_iommu_snooping(struct dmar_domain *domain)
468 domain->iommu_snooping = 1;
470 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
471 for (; i < g_num_of_iommus; ) {
472 if (!ecap_sc_support(g_iommus[i]->ecap)) {
473 domain->iommu_snooping = 0;
476 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
480 /* Some capabilities may be different across iommus */
481 static void domain_update_iommu_cap(struct dmar_domain *domain)
483 domain_update_iommu_coherency(domain);
484 domain_update_iommu_snooping(domain);
487 static struct intel_iommu *device_to_iommu(int segment, u8 bus, u8 devfn)
489 struct dmar_drhd_unit *drhd = NULL;
492 for_each_drhd_unit(drhd) {
495 if (segment != drhd->segment)
498 for (i = 0; i < drhd->devices_cnt; i++) {
499 if (drhd->devices[i] &&
500 drhd->devices[i]->bus->number == bus &&
501 drhd->devices[i]->devfn == devfn)
503 if (drhd->devices[i] &&
504 drhd->devices[i]->subordinate &&
505 drhd->devices[i]->subordinate->number <= bus &&
506 drhd->devices[i]->subordinate->subordinate >= bus)
510 if (drhd->include_all)
517 static void domain_flush_cache(struct dmar_domain *domain,
518 void *addr, int size)
520 if (!domain->iommu_coherency)
521 clflush_cache_range(addr, size);
524 /* Gets context entry for a given bus and devfn */
525 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
528 struct root_entry *root;
529 struct context_entry *context;
530 unsigned long phy_addr;
533 spin_lock_irqsave(&iommu->lock, flags);
534 root = &iommu->root_entry[bus];
535 context = get_context_addr_from_root(root);
537 context = (struct context_entry *)alloc_pgtable_page();
539 spin_unlock_irqrestore(&iommu->lock, flags);
542 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
543 phy_addr = virt_to_phys((void *)context);
544 set_root_value(root, phy_addr);
545 set_root_present(root);
546 __iommu_flush_cache(iommu, root, sizeof(*root));
548 spin_unlock_irqrestore(&iommu->lock, flags);
549 return &context[devfn];
552 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
554 struct root_entry *root;
555 struct context_entry *context;
559 spin_lock_irqsave(&iommu->lock, flags);
560 root = &iommu->root_entry[bus];
561 context = get_context_addr_from_root(root);
566 ret = context_present(&context[devfn]);
568 spin_unlock_irqrestore(&iommu->lock, flags);
572 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
574 struct root_entry *root;
575 struct context_entry *context;
578 spin_lock_irqsave(&iommu->lock, flags);
579 root = &iommu->root_entry[bus];
580 context = get_context_addr_from_root(root);
582 context_clear_entry(&context[devfn]);
583 __iommu_flush_cache(iommu, &context[devfn], \
586 spin_unlock_irqrestore(&iommu->lock, flags);
589 static void free_context_table(struct intel_iommu *iommu)
591 struct root_entry *root;
594 struct context_entry *context;
596 spin_lock_irqsave(&iommu->lock, flags);
597 if (!iommu->root_entry) {
600 for (i = 0; i < ROOT_ENTRY_NR; i++) {
601 root = &iommu->root_entry[i];
602 context = get_context_addr_from_root(root);
604 free_pgtable_page(context);
606 free_pgtable_page(iommu->root_entry);
607 iommu->root_entry = NULL;
609 spin_unlock_irqrestore(&iommu->lock, flags);
612 /* page table handling */
613 #define LEVEL_STRIDE (9)
614 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
616 static inline int agaw_to_level(int agaw)
621 static inline int agaw_to_width(int agaw)
623 return 30 + agaw * LEVEL_STRIDE;
627 static inline int width_to_agaw(int width)
629 return (width - 30) / LEVEL_STRIDE;
632 static inline unsigned int level_to_offset_bits(int level)
634 return (12 + (level - 1) * LEVEL_STRIDE);
637 static inline int address_level_offset(u64 addr, int level)
639 return ((addr >> level_to_offset_bits(level)) & LEVEL_MASK);
642 static inline u64 level_mask(int level)
644 return ((u64)-1 << level_to_offset_bits(level));
647 static inline u64 level_size(int level)
649 return ((u64)1 << level_to_offset_bits(level));
652 static inline u64 align_to_level(u64 addr, int level)
654 return ((addr + level_size(level) - 1) & level_mask(level));
657 static struct dma_pte * addr_to_dma_pte(struct dmar_domain *domain, u64 addr)
659 int addr_width = agaw_to_width(domain->agaw);
660 struct dma_pte *parent, *pte = NULL;
661 int level = agaw_to_level(domain->agaw);
665 BUG_ON(!domain->pgd);
667 addr &= (((u64)1) << addr_width) - 1;
668 parent = domain->pgd;
670 spin_lock_irqsave(&domain->mapping_lock, flags);
674 offset = address_level_offset(addr, level);
675 pte = &parent[offset];
679 if (!dma_pte_present(pte)) {
680 tmp_page = alloc_pgtable_page();
683 spin_unlock_irqrestore(&domain->mapping_lock,
687 domain_flush_cache(domain, tmp_page, PAGE_SIZE);
688 dma_set_pte_addr(pte, virt_to_phys(tmp_page));
690 * high level table always sets r/w, last level page
691 * table control read/write
693 dma_set_pte_readable(pte);
694 dma_set_pte_writable(pte);
695 domain_flush_cache(domain, pte, sizeof(*pte));
697 parent = phys_to_virt(dma_pte_addr(pte));
701 spin_unlock_irqrestore(&domain->mapping_lock, flags);
705 /* return address's pte at specific level */
706 static struct dma_pte *dma_addr_level_pte(struct dmar_domain *domain, u64 addr,
709 struct dma_pte *parent, *pte = NULL;
710 int total = agaw_to_level(domain->agaw);
713 parent = domain->pgd;
714 while (level <= total) {
715 offset = address_level_offset(addr, total);
716 pte = &parent[offset];
720 if (!dma_pte_present(pte))
722 parent = phys_to_virt(dma_pte_addr(pte));
728 /* clear one page's page table */
729 static void dma_pte_clear_one(struct dmar_domain *domain, u64 addr)
731 struct dma_pte *pte = NULL;
733 /* get last level pte */
734 pte = dma_addr_level_pte(domain, addr, 1);
738 domain_flush_cache(domain, pte, sizeof(*pte));
742 /* clear last level pte, a tlb flush should be followed */
743 static void dma_pte_clear_range(struct dmar_domain *domain, u64 start, u64 end)
745 int addr_width = agaw_to_width(domain->agaw);
748 start &= (((u64)1) << addr_width) - 1;
749 end &= (((u64)1) << addr_width) - 1;
750 /* in case it's partial page */
752 end = PAGE_ALIGN(end);
753 npages = (end - start) / VTD_PAGE_SIZE;
755 /* we don't need lock here, nobody else touches the iova range */
757 dma_pte_clear_one(domain, start);
758 start += VTD_PAGE_SIZE;
762 /* free page table pages. last level pte should already be cleared */
763 static void dma_pte_free_pagetable(struct dmar_domain *domain,
766 int addr_width = agaw_to_width(domain->agaw);
768 int total = agaw_to_level(domain->agaw);
772 start &= (((u64)1) << addr_width) - 1;
773 end &= (((u64)1) << addr_width) - 1;
775 /* we don't need lock here, nobody else touches the iova range */
777 while (level <= total) {
778 tmp = align_to_level(start, level);
779 if (tmp >= end || (tmp + level_size(level) > end))
783 pte = dma_addr_level_pte(domain, tmp, level);
786 phys_to_virt(dma_pte_addr(pte)));
788 domain_flush_cache(domain, pte, sizeof(*pte));
790 tmp += level_size(level);
795 if (start == 0 && end >= ((((u64)1) << addr_width) - 1)) {
796 free_pgtable_page(domain->pgd);
802 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
804 struct root_entry *root;
807 root = (struct root_entry *)alloc_pgtable_page();
811 __iommu_flush_cache(iommu, root, ROOT_SIZE);
813 spin_lock_irqsave(&iommu->lock, flags);
814 iommu->root_entry = root;
815 spin_unlock_irqrestore(&iommu->lock, flags);
820 static void iommu_set_root_entry(struct intel_iommu *iommu)
826 addr = iommu->root_entry;
828 spin_lock_irqsave(&iommu->register_lock, flag);
829 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
831 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
833 /* Make sure hardware complete it */
834 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
835 readl, (sts & DMA_GSTS_RTPS), sts);
837 spin_unlock_irqrestore(&iommu->register_lock, flag);
840 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
845 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
848 spin_lock_irqsave(&iommu->register_lock, flag);
849 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
851 /* Make sure hardware complete it */
852 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
853 readl, (!(val & DMA_GSTS_WBFS)), val);
855 spin_unlock_irqrestore(&iommu->register_lock, flag);
858 /* return value determine if we need a write buffer flush */
859 static void __iommu_flush_context(struct intel_iommu *iommu,
860 u16 did, u16 source_id, u8 function_mask,
867 case DMA_CCMD_GLOBAL_INVL:
868 val = DMA_CCMD_GLOBAL_INVL;
870 case DMA_CCMD_DOMAIN_INVL:
871 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
873 case DMA_CCMD_DEVICE_INVL:
874 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
875 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
882 spin_lock_irqsave(&iommu->register_lock, flag);
883 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
885 /* Make sure hardware complete it */
886 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
887 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
889 spin_unlock_irqrestore(&iommu->register_lock, flag);
892 /* return value determine if we need a write buffer flush */
893 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
894 u64 addr, unsigned int size_order, u64 type)
896 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
897 u64 val = 0, val_iva = 0;
901 case DMA_TLB_GLOBAL_FLUSH:
902 /* global flush doesn't need set IVA_REG */
903 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
905 case DMA_TLB_DSI_FLUSH:
906 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
908 case DMA_TLB_PSI_FLUSH:
909 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
910 /* Note: always flush non-leaf currently */
911 val_iva = size_order | addr;
916 /* Note: set drain read/write */
919 * This is probably to be super secure.. Looks like we can
920 * ignore it without any impact.
922 if (cap_read_drain(iommu->cap))
923 val |= DMA_TLB_READ_DRAIN;
925 if (cap_write_drain(iommu->cap))
926 val |= DMA_TLB_WRITE_DRAIN;
928 spin_lock_irqsave(&iommu->register_lock, flag);
929 /* Note: Only uses first TLB reg currently */
931 dmar_writeq(iommu->reg + tlb_offset, val_iva);
932 dmar_writeq(iommu->reg + tlb_offset + 8, val);
934 /* Make sure hardware complete it */
935 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
936 dmar_readq, (!(val & DMA_TLB_IVT)), val);
938 spin_unlock_irqrestore(&iommu->register_lock, flag);
940 /* check IOTLB invalidation granularity */
941 if (DMA_TLB_IAIG(val) == 0)
942 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
943 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
944 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
945 (unsigned long long)DMA_TLB_IIRG(type),
946 (unsigned long long)DMA_TLB_IAIG(val));
949 static struct device_domain_info *iommu_support_dev_iotlb(
950 struct dmar_domain *domain, int segment, u8 bus, u8 devfn)
954 struct device_domain_info *info;
955 struct intel_iommu *iommu = device_to_iommu(segment, bus, devfn);
957 if (!ecap_dev_iotlb_support(iommu->ecap))
963 spin_lock_irqsave(&device_domain_lock, flags);
964 list_for_each_entry(info, &domain->devices, link)
965 if (info->bus == bus && info->devfn == devfn) {
969 spin_unlock_irqrestore(&device_domain_lock, flags);
971 if (!found || !info->dev)
974 if (!pci_find_ext_capability(info->dev, PCI_EXT_CAP_ID_ATS))
977 if (!dmar_find_matched_atsr_unit(info->dev))
985 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
990 pci_enable_ats(info->dev, VTD_PAGE_SHIFT);
993 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
995 if (!info->dev || !pci_ats_enabled(info->dev))
998 pci_disable_ats(info->dev);
1001 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1002 u64 addr, unsigned mask)
1005 unsigned long flags;
1006 struct device_domain_info *info;
1008 spin_lock_irqsave(&device_domain_lock, flags);
1009 list_for_each_entry(info, &domain->devices, link) {
1010 if (!info->dev || !pci_ats_enabled(info->dev))
1013 sid = info->bus << 8 | info->devfn;
1014 qdep = pci_ats_queue_depth(info->dev);
1015 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1017 spin_unlock_irqrestore(&device_domain_lock, flags);
1020 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
1021 u64 addr, unsigned int pages)
1023 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1025 BUG_ON(addr & (~VTD_PAGE_MASK));
1029 * Fallback to domain selective flush if no PSI support or the size is
1031 * PSI requires page size to be 2 ^ x, and the base address is naturally
1032 * aligned to the size
1034 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1035 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1038 iommu->flush.flush_iotlb(iommu, did, addr, mask,
1041 iommu_flush_dev_iotlb(iommu->domains[did], addr, mask);
1044 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1047 unsigned long flags;
1049 spin_lock_irqsave(&iommu->register_lock, flags);
1050 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1051 pmen &= ~DMA_PMEN_EPM;
1052 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1054 /* wait for the protected region status bit to clear */
1055 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1056 readl, !(pmen & DMA_PMEN_PRS), pmen);
1058 spin_unlock_irqrestore(&iommu->register_lock, flags);
1061 static int iommu_enable_translation(struct intel_iommu *iommu)
1064 unsigned long flags;
1066 spin_lock_irqsave(&iommu->register_lock, flags);
1067 iommu->gcmd |= DMA_GCMD_TE;
1068 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1070 /* Make sure hardware complete it */
1071 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1072 readl, (sts & DMA_GSTS_TES), sts);
1074 spin_unlock_irqrestore(&iommu->register_lock, flags);
1078 static int iommu_disable_translation(struct intel_iommu *iommu)
1083 spin_lock_irqsave(&iommu->register_lock, flag);
1084 iommu->gcmd &= ~DMA_GCMD_TE;
1085 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1087 /* Make sure hardware complete it */
1088 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1089 readl, (!(sts & DMA_GSTS_TES)), sts);
1091 spin_unlock_irqrestore(&iommu->register_lock, flag);
1096 static int iommu_init_domains(struct intel_iommu *iommu)
1098 unsigned long ndomains;
1099 unsigned long nlongs;
1101 ndomains = cap_ndoms(iommu->cap);
1102 pr_debug("Number of Domains supportd <%ld>\n", ndomains);
1103 nlongs = BITS_TO_LONGS(ndomains);
1105 /* TBD: there might be 64K domains,
1106 * consider other allocation for future chip
1108 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1109 if (!iommu->domain_ids) {
1110 printk(KERN_ERR "Allocating domain id array failed\n");
1113 iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1115 if (!iommu->domains) {
1116 printk(KERN_ERR "Allocating domain array failed\n");
1117 kfree(iommu->domain_ids);
1121 spin_lock_init(&iommu->lock);
1124 * if Caching mode is set, then invalid translations are tagged
1125 * with domainid 0. Hence we need to pre-allocate it.
1127 if (cap_caching_mode(iommu->cap))
1128 set_bit(0, iommu->domain_ids);
1133 static void domain_exit(struct dmar_domain *domain);
1134 static void vm_domain_exit(struct dmar_domain *domain);
1136 void free_dmar_iommu(struct intel_iommu *iommu)
1138 struct dmar_domain *domain;
1140 unsigned long flags;
1142 i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1143 for (; i < cap_ndoms(iommu->cap); ) {
1144 domain = iommu->domains[i];
1145 clear_bit(i, iommu->domain_ids);
1147 spin_lock_irqsave(&domain->iommu_lock, flags);
1148 if (--domain->iommu_count == 0) {
1149 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
1150 vm_domain_exit(domain);
1152 domain_exit(domain);
1154 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1156 i = find_next_bit(iommu->domain_ids,
1157 cap_ndoms(iommu->cap), i+1);
1160 if (iommu->gcmd & DMA_GCMD_TE)
1161 iommu_disable_translation(iommu);
1164 set_irq_data(iommu->irq, NULL);
1165 /* This will mask the irq */
1166 free_irq(iommu->irq, iommu);
1167 destroy_irq(iommu->irq);
1170 kfree(iommu->domains);
1171 kfree(iommu->domain_ids);
1173 g_iommus[iommu->seq_id] = NULL;
1175 /* if all iommus are freed, free g_iommus */
1176 for (i = 0; i < g_num_of_iommus; i++) {
1181 if (i == g_num_of_iommus)
1184 /* free context mapping */
1185 free_context_table(iommu);
1188 static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
1191 unsigned long ndomains;
1192 struct dmar_domain *domain;
1193 unsigned long flags;
1195 domain = alloc_domain_mem();
1199 ndomains = cap_ndoms(iommu->cap);
1201 spin_lock_irqsave(&iommu->lock, flags);
1202 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1203 if (num >= ndomains) {
1204 spin_unlock_irqrestore(&iommu->lock, flags);
1205 free_domain_mem(domain);
1206 printk(KERN_ERR "IOMMU: no free domain ids\n");
1210 set_bit(num, iommu->domain_ids);
1212 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
1213 set_bit(iommu->seq_id, &domain->iommu_bmp);
1215 iommu->domains[num] = domain;
1216 spin_unlock_irqrestore(&iommu->lock, flags);
1221 static void iommu_free_domain(struct dmar_domain *domain)
1223 unsigned long flags;
1224 struct intel_iommu *iommu;
1226 iommu = domain_get_iommu(domain);
1228 spin_lock_irqsave(&iommu->lock, flags);
1229 clear_bit(domain->id, iommu->domain_ids);
1230 spin_unlock_irqrestore(&iommu->lock, flags);
1233 static struct iova_domain reserved_iova_list;
1234 static struct lock_class_key reserved_alloc_key;
1235 static struct lock_class_key reserved_rbtree_key;
1237 static void dmar_init_reserved_ranges(void)
1239 struct pci_dev *pdev = NULL;
1244 init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1246 lockdep_set_class(&reserved_iova_list.iova_alloc_lock,
1247 &reserved_alloc_key);
1248 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1249 &reserved_rbtree_key);
1251 /* IOAPIC ranges shouldn't be accessed by DMA */
1252 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1253 IOVA_PFN(IOAPIC_RANGE_END));
1255 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1257 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1258 for_each_pci_dev(pdev) {
1261 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1262 r = &pdev->resource[i];
1263 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1267 size = r->end - addr;
1268 size = PAGE_ALIGN(size);
1269 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(addr),
1270 IOVA_PFN(size + addr) - 1);
1272 printk(KERN_ERR "Reserve iova failed\n");
1278 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1280 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1283 static inline int guestwidth_to_adjustwidth(int gaw)
1286 int r = (gaw - 12) % 9;
1297 static int domain_init(struct dmar_domain *domain, int guest_width)
1299 struct intel_iommu *iommu;
1300 int adjust_width, agaw;
1301 unsigned long sagaw;
1303 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1304 spin_lock_init(&domain->mapping_lock);
1305 spin_lock_init(&domain->iommu_lock);
1307 domain_reserve_special_ranges(domain);
1309 /* calculate AGAW */
1310 iommu = domain_get_iommu(domain);
1311 if (guest_width > cap_mgaw(iommu->cap))
1312 guest_width = cap_mgaw(iommu->cap);
1313 domain->gaw = guest_width;
1314 adjust_width = guestwidth_to_adjustwidth(guest_width);
1315 agaw = width_to_agaw(adjust_width);
1316 sagaw = cap_sagaw(iommu->cap);
1317 if (!test_bit(agaw, &sagaw)) {
1318 /* hardware doesn't support it, choose a bigger one */
1319 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1320 agaw = find_next_bit(&sagaw, 5, agaw);
1324 domain->agaw = agaw;
1325 INIT_LIST_HEAD(&domain->devices);
1327 if (ecap_coherent(iommu->ecap))
1328 domain->iommu_coherency = 1;
1330 domain->iommu_coherency = 0;
1332 if (ecap_sc_support(iommu->ecap))
1333 domain->iommu_snooping = 1;
1335 domain->iommu_snooping = 0;
1337 domain->iommu_count = 1;
1339 /* always allocate the top pgd */
1340 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1343 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1347 static void domain_exit(struct dmar_domain *domain)
1351 /* Domain 0 is reserved, so dont process it */
1355 domain_remove_dev_info(domain);
1357 put_iova_domain(&domain->iovad);
1358 end = DOMAIN_MAX_ADDR(domain->gaw);
1359 end = end & (~PAGE_MASK);
1362 dma_pte_clear_range(domain, 0, end);
1364 /* free page tables */
1365 dma_pte_free_pagetable(domain, 0, end);
1367 iommu_free_domain(domain);
1368 free_domain_mem(domain);
1371 static int domain_context_mapping_one(struct dmar_domain *domain, int segment,
1372 u8 bus, u8 devfn, int translation)
1374 struct context_entry *context;
1375 unsigned long flags;
1376 struct intel_iommu *iommu;
1377 struct dma_pte *pgd;
1379 unsigned long ndomains;
1382 struct device_domain_info *info = NULL;
1384 pr_debug("Set context mapping for %02x:%02x.%d\n",
1385 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1387 BUG_ON(!domain->pgd);
1388 BUG_ON(translation != CONTEXT_TT_PASS_THROUGH &&
1389 translation != CONTEXT_TT_MULTI_LEVEL);
1391 iommu = device_to_iommu(segment, bus, devfn);
1395 context = device_to_context_entry(iommu, bus, devfn);
1398 spin_lock_irqsave(&iommu->lock, flags);
1399 if (context_present(context)) {
1400 spin_unlock_irqrestore(&iommu->lock, flags);
1407 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE) {
1410 /* find an available domain id for this device in iommu */
1411 ndomains = cap_ndoms(iommu->cap);
1412 num = find_first_bit(iommu->domain_ids, ndomains);
1413 for (; num < ndomains; ) {
1414 if (iommu->domains[num] == domain) {
1419 num = find_next_bit(iommu->domain_ids,
1420 cap_ndoms(iommu->cap), num+1);
1424 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1425 if (num >= ndomains) {
1426 spin_unlock_irqrestore(&iommu->lock, flags);
1427 printk(KERN_ERR "IOMMU: no free domain ids\n");
1431 set_bit(num, iommu->domain_ids);
1432 iommu->domains[num] = domain;
1436 /* Skip top levels of page tables for
1437 * iommu which has less agaw than default.
1439 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1440 pgd = phys_to_virt(dma_pte_addr(pgd));
1441 if (!dma_pte_present(pgd)) {
1442 spin_unlock_irqrestore(&iommu->lock, flags);
1448 context_set_domain_id(context, id);
1450 if (translation != CONTEXT_TT_PASS_THROUGH) {
1451 info = iommu_support_dev_iotlb(domain, segment, bus, devfn);
1452 translation = info ? CONTEXT_TT_DEV_IOTLB :
1453 CONTEXT_TT_MULTI_LEVEL;
1456 * In pass through mode, AW must be programmed to indicate the largest
1457 * AGAW value supported by hardware. And ASR is ignored by hardware.
1459 if (unlikely(translation == CONTEXT_TT_PASS_THROUGH))
1460 context_set_address_width(context, iommu->msagaw);
1462 context_set_address_root(context, virt_to_phys(pgd));
1463 context_set_address_width(context, iommu->agaw);
1466 context_set_translation_type(context, translation);
1467 context_set_fault_enable(context);
1468 context_set_present(context);
1469 domain_flush_cache(domain, context, sizeof(*context));
1472 * It's a non-present to present mapping. If hardware doesn't cache
1473 * non-present entry we only need to flush the write-buffer. If the
1474 * _does_ cache non-present entries, then it does so in the special
1475 * domain #0, which we have to flush:
1477 if (cap_caching_mode(iommu->cap)) {
1478 iommu->flush.flush_context(iommu, 0,
1479 (((u16)bus) << 8) | devfn,
1480 DMA_CCMD_MASK_NOBIT,
1481 DMA_CCMD_DEVICE_INVL);
1482 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
1484 iommu_flush_write_buffer(iommu);
1486 iommu_enable_dev_iotlb(info);
1487 spin_unlock_irqrestore(&iommu->lock, flags);
1489 spin_lock_irqsave(&domain->iommu_lock, flags);
1490 if (!test_and_set_bit(iommu->seq_id, &domain->iommu_bmp)) {
1491 domain->iommu_count++;
1492 domain_update_iommu_cap(domain);
1494 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1499 domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev,
1503 struct pci_dev *tmp, *parent;
1505 ret = domain_context_mapping_one(domain, pci_domain_nr(pdev->bus),
1506 pdev->bus->number, pdev->devfn,
1511 /* dependent device mapping */
1512 tmp = pci_find_upstream_pcie_bridge(pdev);
1515 /* Secondary interface's bus number and devfn 0 */
1516 parent = pdev->bus->self;
1517 while (parent != tmp) {
1518 ret = domain_context_mapping_one(domain,
1519 pci_domain_nr(parent->bus),
1520 parent->bus->number,
1521 parent->devfn, translation);
1524 parent = parent->bus->self;
1526 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1527 return domain_context_mapping_one(domain,
1528 pci_domain_nr(tmp->subordinate),
1529 tmp->subordinate->number, 0,
1531 else /* this is a legacy PCI bridge */
1532 return domain_context_mapping_one(domain,
1533 pci_domain_nr(tmp->bus),
1539 static int domain_context_mapped(struct pci_dev *pdev)
1542 struct pci_dev *tmp, *parent;
1543 struct intel_iommu *iommu;
1545 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
1550 ret = device_context_mapped(iommu, pdev->bus->number, pdev->devfn);
1553 /* dependent device mapping */
1554 tmp = pci_find_upstream_pcie_bridge(pdev);
1557 /* Secondary interface's bus number and devfn 0 */
1558 parent = pdev->bus->self;
1559 while (parent != tmp) {
1560 ret = device_context_mapped(iommu, parent->bus->number,
1564 parent = parent->bus->self;
1567 return device_context_mapped(iommu, tmp->subordinate->number,
1570 return device_context_mapped(iommu, tmp->bus->number,
1575 domain_page_mapping(struct dmar_domain *domain, dma_addr_t iova,
1576 u64 hpa, size_t size, int prot)
1578 u64 start_pfn, end_pfn;
1579 struct dma_pte *pte;
1581 int addr_width = agaw_to_width(domain->agaw);
1583 hpa &= (((u64)1) << addr_width) - 1;
1585 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1588 start_pfn = ((u64)hpa) >> VTD_PAGE_SHIFT;
1589 end_pfn = (VTD_PAGE_ALIGN(((u64)hpa) + size)) >> VTD_PAGE_SHIFT;
1591 while (start_pfn < end_pfn) {
1592 pte = addr_to_dma_pte(domain, iova + VTD_PAGE_SIZE * index);
1595 /* We don't need lock here, nobody else
1596 * touches the iova range
1598 BUG_ON(dma_pte_addr(pte));
1599 dma_set_pte_addr(pte, start_pfn << VTD_PAGE_SHIFT);
1600 dma_set_pte_prot(pte, prot);
1601 if (prot & DMA_PTE_SNP)
1602 dma_set_pte_snp(pte);
1603 domain_flush_cache(domain, pte, sizeof(*pte));
1610 static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
1615 clear_context_table(iommu, bus, devfn);
1616 iommu->flush.flush_context(iommu, 0, 0, 0,
1617 DMA_CCMD_GLOBAL_INVL);
1618 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
1621 static void domain_remove_dev_info(struct dmar_domain *domain)
1623 struct device_domain_info *info;
1624 unsigned long flags;
1625 struct intel_iommu *iommu;
1627 spin_lock_irqsave(&device_domain_lock, flags);
1628 while (!list_empty(&domain->devices)) {
1629 info = list_entry(domain->devices.next,
1630 struct device_domain_info, link);
1631 list_del(&info->link);
1632 list_del(&info->global);
1634 info->dev->dev.archdata.iommu = NULL;
1635 spin_unlock_irqrestore(&device_domain_lock, flags);
1637 iommu_disable_dev_iotlb(info);
1638 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
1639 iommu_detach_dev(iommu, info->bus, info->devfn);
1640 free_devinfo_mem(info);
1642 spin_lock_irqsave(&device_domain_lock, flags);
1644 spin_unlock_irqrestore(&device_domain_lock, flags);
1649 * Note: we use struct pci_dev->dev.archdata.iommu stores the info
1651 static struct dmar_domain *
1652 find_domain(struct pci_dev *pdev)
1654 struct device_domain_info *info;
1656 /* No lock here, assumes no domain exit in normal case */
1657 info = pdev->dev.archdata.iommu;
1659 return info->domain;
1663 /* domain is initialized */
1664 static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1666 struct dmar_domain *domain, *found = NULL;
1667 struct intel_iommu *iommu;
1668 struct dmar_drhd_unit *drhd;
1669 struct device_domain_info *info, *tmp;
1670 struct pci_dev *dev_tmp;
1671 unsigned long flags;
1672 int bus = 0, devfn = 0;
1675 domain = find_domain(pdev);
1679 segment = pci_domain_nr(pdev->bus);
1681 dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1683 if (dev_tmp->is_pcie) {
1684 bus = dev_tmp->subordinate->number;
1687 bus = dev_tmp->bus->number;
1688 devfn = dev_tmp->devfn;
1690 spin_lock_irqsave(&device_domain_lock, flags);
1691 list_for_each_entry(info, &device_domain_list, global) {
1692 if (info->segment == segment &&
1693 info->bus == bus && info->devfn == devfn) {
1694 found = info->domain;
1698 spin_unlock_irqrestore(&device_domain_lock, flags);
1699 /* pcie-pci bridge already has a domain, uses it */
1706 /* Allocate new domain for the device */
1707 drhd = dmar_find_matched_drhd_unit(pdev);
1709 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1713 iommu = drhd->iommu;
1715 domain = iommu_alloc_domain(iommu);
1719 if (domain_init(domain, gaw)) {
1720 domain_exit(domain);
1724 /* register pcie-to-pci device */
1726 info = alloc_devinfo_mem();
1728 domain_exit(domain);
1731 info->segment = segment;
1733 info->devfn = devfn;
1735 info->domain = domain;
1736 /* This domain is shared by devices under p2p bridge */
1737 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
1739 /* pcie-to-pci bridge already has a domain, uses it */
1741 spin_lock_irqsave(&device_domain_lock, flags);
1742 list_for_each_entry(tmp, &device_domain_list, global) {
1743 if (tmp->segment == segment &&
1744 tmp->bus == bus && tmp->devfn == devfn) {
1745 found = tmp->domain;
1750 free_devinfo_mem(info);
1751 domain_exit(domain);
1754 list_add(&info->link, &domain->devices);
1755 list_add(&info->global, &device_domain_list);
1757 spin_unlock_irqrestore(&device_domain_lock, flags);
1761 info = alloc_devinfo_mem();
1764 info->segment = segment;
1765 info->bus = pdev->bus->number;
1766 info->devfn = pdev->devfn;
1768 info->domain = domain;
1769 spin_lock_irqsave(&device_domain_lock, flags);
1770 /* somebody is fast */
1771 found = find_domain(pdev);
1772 if (found != NULL) {
1773 spin_unlock_irqrestore(&device_domain_lock, flags);
1774 if (found != domain) {
1775 domain_exit(domain);
1778 free_devinfo_mem(info);
1781 list_add(&info->link, &domain->devices);
1782 list_add(&info->global, &device_domain_list);
1783 pdev->dev.archdata.iommu = info;
1784 spin_unlock_irqrestore(&device_domain_lock, flags);
1787 /* recheck it here, maybe others set it */
1788 return find_domain(pdev);
1791 static int iommu_prepare_identity_map(struct pci_dev *pdev,
1792 unsigned long long start,
1793 unsigned long long end)
1795 struct dmar_domain *domain;
1797 unsigned long long base;
1801 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1802 pci_name(pdev), start, end);
1803 /* page table init */
1804 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1808 /* The address might not be aligned */
1809 base = start & PAGE_MASK;
1811 size = PAGE_ALIGN(size);
1812 if (!reserve_iova(&domain->iovad, IOVA_PFN(base),
1813 IOVA_PFN(base + size) - 1)) {
1814 printk(KERN_ERR "IOMMU: reserve iova failed\n");
1819 pr_debug("Mapping reserved region %lx@%llx for %s\n",
1820 size, base, pci_name(pdev));
1822 * RMRR range might have overlap with physical memory range,
1825 dma_pte_clear_range(domain, base, base + size);
1827 ret = domain_page_mapping(domain, base, base, size,
1828 DMA_PTE_READ|DMA_PTE_WRITE);
1832 /* context entry init */
1833 ret = domain_context_mapping(domain, pdev, CONTEXT_TT_MULTI_LEVEL);
1837 domain_exit(domain);
1842 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1843 struct pci_dev *pdev)
1845 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1847 return iommu_prepare_identity_map(pdev, rmrr->base_address,
1848 rmrr->end_address + 1);
1851 #ifdef CONFIG_DMAR_GFX_WA
1852 struct iommu_prepare_data {
1853 struct pci_dev *pdev;
1857 static int __init iommu_prepare_work_fn(unsigned long start_pfn,
1858 unsigned long end_pfn, void *datax)
1860 struct iommu_prepare_data *data;
1862 data = (struct iommu_prepare_data *)datax;
1864 data->ret = iommu_prepare_identity_map(data->pdev,
1865 start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
1870 static int __init iommu_prepare_with_active_regions(struct pci_dev *pdev)
1873 struct iommu_prepare_data data;
1878 for_each_online_node(nid) {
1879 work_with_active_regions(nid, iommu_prepare_work_fn, &data);
1886 static void __init iommu_prepare_gfx_mapping(void)
1888 struct pci_dev *pdev = NULL;
1891 for_each_pci_dev(pdev) {
1892 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO ||
1893 !IS_GFX_DEVICE(pdev))
1895 printk(KERN_INFO "IOMMU: gfx device %s 1-1 mapping\n",
1897 ret = iommu_prepare_with_active_regions(pdev);
1899 printk(KERN_ERR "IOMMU: mapping reserved region failed\n");
1902 #else /* !CONFIG_DMAR_GFX_WA */
1903 static inline void iommu_prepare_gfx_mapping(void)
1909 #ifdef CONFIG_DMAR_FLOPPY_WA
1910 static inline void iommu_prepare_isa(void)
1912 struct pci_dev *pdev;
1915 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1919 printk(KERN_INFO "IOMMU: Prepare 0-16M unity mapping for LPC\n");
1920 ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1923 printk(KERN_ERR "IOMMU: Failed to create 0-64M identity map, "
1924 "floppy might not work\n");
1928 static inline void iommu_prepare_isa(void)
1932 #endif /* !CONFIG_DMAR_FLPY_WA */
1934 /* Initialize each context entry as pass through.*/
1935 static int __init init_context_pass_through(void)
1937 struct pci_dev *pdev = NULL;
1938 struct dmar_domain *domain;
1941 for_each_pci_dev(pdev) {
1942 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1943 ret = domain_context_mapping(domain, pdev,
1944 CONTEXT_TT_PASS_THROUGH);
1951 static int __init init_dmars(void)
1953 struct dmar_drhd_unit *drhd;
1954 struct dmar_rmrr_unit *rmrr;
1955 struct pci_dev *pdev;
1956 struct intel_iommu *iommu;
1958 int pass_through = 1;
1963 * initialize and program root entry to not present
1966 for_each_drhd_unit(drhd) {
1969 * lock not needed as this is only incremented in the single
1970 * threaded kernel __init code path all other access are read
1975 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
1978 printk(KERN_ERR "Allocating global iommu array failed\n");
1983 deferred_flush = kzalloc(g_num_of_iommus *
1984 sizeof(struct deferred_flush_tables), GFP_KERNEL);
1985 if (!deferred_flush) {
1991 for_each_drhd_unit(drhd) {
1995 iommu = drhd->iommu;
1996 g_iommus[iommu->seq_id] = iommu;
1998 ret = iommu_init_domains(iommu);
2004 * we could share the same root & context tables
2005 * amoung all IOMMU's. Need to Split it later.
2007 ret = iommu_alloc_root_entry(iommu);
2009 printk(KERN_ERR "IOMMU: allocate root entry failed\n");
2012 if (!ecap_pass_through(iommu->ecap))
2015 if (iommu_pass_through)
2016 if (!pass_through) {
2018 "Pass Through is not supported by hardware.\n");
2019 iommu_pass_through = 0;
2023 * Start from the sane iommu hardware state.
2025 for_each_drhd_unit(drhd) {
2029 iommu = drhd->iommu;
2032 * If the queued invalidation is already initialized by us
2033 * (for example, while enabling interrupt-remapping) then
2034 * we got the things already rolling from a sane state.
2040 * Clear any previous faults.
2042 dmar_fault(-1, iommu);
2044 * Disable queued invalidation if supported and already enabled
2045 * before OS handover.
2047 dmar_disable_qi(iommu);
2050 for_each_drhd_unit(drhd) {
2054 iommu = drhd->iommu;
2056 if (dmar_enable_qi(iommu)) {
2058 * Queued Invalidate not enabled, use Register Based
2061 iommu->flush.flush_context = __iommu_flush_context;
2062 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2063 printk(KERN_INFO "IOMMU 0x%Lx: using Register based "
2065 (unsigned long long)drhd->reg_base_addr);
2067 iommu->flush.flush_context = qi_flush_context;
2068 iommu->flush.flush_iotlb = qi_flush_iotlb;
2069 printk(KERN_INFO "IOMMU 0x%Lx: using Queued "
2071 (unsigned long long)drhd->reg_base_addr);
2075 #ifdef CONFIG_INTR_REMAP
2076 if (!intr_remapping_enabled) {
2077 ret = enable_intr_remapping(0);
2080 "IOMMU: enable interrupt remapping failed\n");
2084 * If pass through is set and enabled, context entries of all pci
2085 * devices are intialized by pass through translation type.
2087 if (iommu_pass_through) {
2088 ret = init_context_pass_through();
2090 printk(KERN_ERR "IOMMU: Pass through init failed.\n");
2091 iommu_pass_through = 0;
2096 * If pass through is not set or not enabled, setup context entries for
2097 * identity mappings for rmrr, gfx, and isa.
2099 if (!iommu_pass_through) {
2102 * for each dev attached to rmrr
2104 * locate drhd for dev, alloc domain for dev
2105 * allocate free domain
2106 * allocate page table entries for rmrr
2107 * if context not allocated for bus
2108 * allocate and init context
2109 * set present in root table for this bus
2110 * init context with domain, translation etc
2114 for_each_rmrr_units(rmrr) {
2115 for (i = 0; i < rmrr->devices_cnt; i++) {
2116 pdev = rmrr->devices[i];
2118 * some BIOS lists non-exist devices in DMAR
2123 ret = iommu_prepare_rmrr_dev(rmrr, pdev);
2126 "IOMMU: mapping reserved region failed\n");
2130 iommu_prepare_gfx_mapping();
2132 iommu_prepare_isa();
2138 * global invalidate context cache
2139 * global invalidate iotlb
2140 * enable translation
2142 for_each_drhd_unit(drhd) {
2145 iommu = drhd->iommu;
2147 iommu_flush_write_buffer(iommu);
2149 ret = dmar_set_interrupt(iommu);
2153 iommu_set_root_entry(iommu);
2155 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
2156 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2157 iommu_disable_protect_mem_regions(iommu);
2159 ret = iommu_enable_translation(iommu);
2166 for_each_drhd_unit(drhd) {
2169 iommu = drhd->iommu;
2176 static inline u64 aligned_size(u64 host_addr, size_t size)
2179 addr = (host_addr & (~PAGE_MASK)) + size;
2180 return PAGE_ALIGN(addr);
2184 iommu_alloc_iova(struct dmar_domain *domain, size_t size, u64 end)
2188 /* Make sure it's in range */
2189 end = min_t(u64, DOMAIN_MAX_ADDR(domain->gaw), end);
2190 if (!size || (IOVA_START_ADDR + size > end))
2193 piova = alloc_iova(&domain->iovad,
2194 size >> PAGE_SHIFT, IOVA_PFN(end), 1);
2198 static struct iova *
2199 __intel_alloc_iova(struct device *dev, struct dmar_domain *domain,
2200 size_t size, u64 dma_mask)
2202 struct pci_dev *pdev = to_pci_dev(dev);
2203 struct iova *iova = NULL;
2205 if (dma_mask <= DMA_BIT_MASK(32) || dmar_forcedac)
2206 iova = iommu_alloc_iova(domain, size, dma_mask);
2209 * First try to allocate an io virtual address in
2210 * DMA_BIT_MASK(32) and if that fails then try allocating
2213 iova = iommu_alloc_iova(domain, size, DMA_BIT_MASK(32));
2215 iova = iommu_alloc_iova(domain, size, dma_mask);
2219 printk(KERN_ERR"Allocating iova for %s failed", pci_name(pdev));
2226 static struct dmar_domain *
2227 get_valid_domain_for_dev(struct pci_dev *pdev)
2229 struct dmar_domain *domain;
2232 domain = get_domain_for_dev(pdev,
2233 DEFAULT_DOMAIN_ADDRESS_WIDTH);
2236 "Allocating domain for %s failed", pci_name(pdev));
2240 /* make sure context mapping is ok */
2241 if (unlikely(!domain_context_mapped(pdev))) {
2242 ret = domain_context_mapping(domain, pdev,
2243 CONTEXT_TT_MULTI_LEVEL);
2246 "Domain context map for %s failed",
2255 static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
2256 size_t size, int dir, u64 dma_mask)
2258 struct pci_dev *pdev = to_pci_dev(hwdev);
2259 struct dmar_domain *domain;
2260 phys_addr_t start_paddr;
2264 struct intel_iommu *iommu;
2266 BUG_ON(dir == DMA_NONE);
2267 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2270 domain = get_valid_domain_for_dev(pdev);
2274 iommu = domain_get_iommu(domain);
2275 size = aligned_size((u64)paddr, size);
2277 iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2281 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
2284 * Check if DMAR supports zero-length reads on write only
2287 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2288 !cap_zlr(iommu->cap))
2289 prot |= DMA_PTE_READ;
2290 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2291 prot |= DMA_PTE_WRITE;
2293 * paddr - (paddr + size) might be partial page, we should map the whole
2294 * page. Note: if two part of one page are separately mapped, we
2295 * might have two guest_addr mapping to the same host paddr, but this
2296 * is not a big problem
2298 ret = domain_page_mapping(domain, start_paddr,
2299 ((u64)paddr) & PAGE_MASK, size, prot);
2303 /* it's a non-present to present mapping. Only flush if caching mode */
2304 if (cap_caching_mode(iommu->cap))
2305 iommu_flush_iotlb_psi(iommu, 0, start_paddr,
2306 size >> VTD_PAGE_SHIFT);
2308 iommu_flush_write_buffer(iommu);
2310 return start_paddr + ((u64)paddr & (~PAGE_MASK));
2314 __free_iova(&domain->iovad, iova);
2315 printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
2316 pci_name(pdev), size, (unsigned long long)paddr, dir);
2320 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
2321 unsigned long offset, size_t size,
2322 enum dma_data_direction dir,
2323 struct dma_attrs *attrs)
2325 return __intel_map_single(dev, page_to_phys(page) + offset, size,
2326 dir, to_pci_dev(dev)->dma_mask);
2329 static void flush_unmaps(void)
2335 /* just flush them all */
2336 for (i = 0; i < g_num_of_iommus; i++) {
2337 struct intel_iommu *iommu = g_iommus[i];
2341 if (!deferred_flush[i].next)
2344 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2345 DMA_TLB_GLOBAL_FLUSH);
2346 for (j = 0; j < deferred_flush[i].next; j++) {
2348 struct iova *iova = deferred_flush[i].iova[j];
2350 mask = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT;
2351 mask = ilog2(mask >> VTD_PAGE_SHIFT);
2352 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
2353 iova->pfn_lo << PAGE_SHIFT, mask);
2354 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
2356 deferred_flush[i].next = 0;
2362 static void flush_unmaps_timeout(unsigned long data)
2364 unsigned long flags;
2366 spin_lock_irqsave(&async_umap_flush_lock, flags);
2368 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2371 static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2373 unsigned long flags;
2375 struct intel_iommu *iommu;
2377 spin_lock_irqsave(&async_umap_flush_lock, flags);
2378 if (list_size == HIGH_WATER_MARK)
2381 iommu = domain_get_iommu(dom);
2382 iommu_id = iommu->seq_id;
2384 next = deferred_flush[iommu_id].next;
2385 deferred_flush[iommu_id].domain[next] = dom;
2386 deferred_flush[iommu_id].iova[next] = iova;
2387 deferred_flush[iommu_id].next++;
2390 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2394 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2397 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
2398 size_t size, enum dma_data_direction dir,
2399 struct dma_attrs *attrs)
2401 struct pci_dev *pdev = to_pci_dev(dev);
2402 struct dmar_domain *domain;
2403 unsigned long start_addr;
2405 struct intel_iommu *iommu;
2407 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2409 domain = find_domain(pdev);
2412 iommu = domain_get_iommu(domain);
2414 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
2418 start_addr = iova->pfn_lo << PAGE_SHIFT;
2419 size = aligned_size((u64)dev_addr, size);
2421 pr_debug("Device %s unmapping: %zx@%llx\n",
2422 pci_name(pdev), size, (unsigned long long)start_addr);
2424 /* clear the whole page */
2425 dma_pte_clear_range(domain, start_addr, start_addr + size);
2426 /* free page tables */
2427 dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2428 if (intel_iommu_strict) {
2429 iommu_flush_iotlb_psi(iommu, domain->id, start_addr,
2430 size >> VTD_PAGE_SHIFT);
2432 __free_iova(&domain->iovad, iova);
2434 add_unmap(domain, iova);
2436 * queue up the release of the unmap to save the 1/6th of the
2437 * cpu used up by the iotlb flush operation...
2442 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr, size_t size,
2445 intel_unmap_page(dev, dev_addr, size, dir, NULL);
2448 static void *intel_alloc_coherent(struct device *hwdev, size_t size,
2449 dma_addr_t *dma_handle, gfp_t flags)
2454 size = PAGE_ALIGN(size);
2455 order = get_order(size);
2456 flags &= ~(GFP_DMA | GFP_DMA32);
2458 vaddr = (void *)__get_free_pages(flags, order);
2461 memset(vaddr, 0, size);
2463 *dma_handle = __intel_map_single(hwdev, virt_to_bus(vaddr), size,
2465 hwdev->coherent_dma_mask);
2468 free_pages((unsigned long)vaddr, order);
2472 static void intel_free_coherent(struct device *hwdev, size_t size, void *vaddr,
2473 dma_addr_t dma_handle)
2477 size = PAGE_ALIGN(size);
2478 order = get_order(size);
2480 intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2481 free_pages((unsigned long)vaddr, order);
2484 static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2485 int nelems, enum dma_data_direction dir,
2486 struct dma_attrs *attrs)
2489 struct pci_dev *pdev = to_pci_dev(hwdev);
2490 struct dmar_domain *domain;
2491 unsigned long start_addr;
2495 struct scatterlist *sg;
2496 struct intel_iommu *iommu;
2498 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2501 domain = find_domain(pdev);
2504 iommu = domain_get_iommu(domain);
2506 iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
2509 for_each_sg(sglist, sg, nelems, i) {
2510 addr = page_to_phys(sg_page(sg)) + sg->offset;
2511 size += aligned_size((u64)addr, sg->length);
2514 start_addr = iova->pfn_lo << PAGE_SHIFT;
2516 /* clear the whole page */
2517 dma_pte_clear_range(domain, start_addr, start_addr + size);
2518 /* free page tables */
2519 dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2521 iommu_flush_iotlb_psi(iommu, domain->id, start_addr,
2522 size >> VTD_PAGE_SHIFT);
2525 __free_iova(&domain->iovad, iova);
2528 static int intel_nontranslate_map_sg(struct device *hddev,
2529 struct scatterlist *sglist, int nelems, int dir)
2532 struct scatterlist *sg;
2534 for_each_sg(sglist, sg, nelems, i) {
2535 BUG_ON(!sg_page(sg));
2536 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
2537 sg->dma_length = sg->length;
2542 static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist, int nelems,
2543 enum dma_data_direction dir, struct dma_attrs *attrs)
2547 struct pci_dev *pdev = to_pci_dev(hwdev);
2548 struct dmar_domain *domain;
2552 struct iova *iova = NULL;
2554 struct scatterlist *sg;
2555 unsigned long start_addr;
2556 struct intel_iommu *iommu;
2558 BUG_ON(dir == DMA_NONE);
2559 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2560 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
2562 domain = get_valid_domain_for_dev(pdev);
2566 iommu = domain_get_iommu(domain);
2568 for_each_sg(sglist, sg, nelems, i) {
2569 addr = page_to_phys(sg_page(sg)) + sg->offset;
2570 size += aligned_size((u64)addr, sg->length);
2573 iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2575 sglist->dma_length = 0;
2580 * Check if DMAR supports zero-length reads on write only
2583 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2584 !cap_zlr(iommu->cap))
2585 prot |= DMA_PTE_READ;
2586 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2587 prot |= DMA_PTE_WRITE;
2589 start_addr = iova->pfn_lo << PAGE_SHIFT;
2591 for_each_sg(sglist, sg, nelems, i) {
2592 addr = page_to_phys(sg_page(sg)) + sg->offset;
2593 size = aligned_size((u64)addr, sg->length);
2594 ret = domain_page_mapping(domain, start_addr + offset,
2595 ((u64)addr) & PAGE_MASK,
2598 /* clear the page */
2599 dma_pte_clear_range(domain, start_addr,
2600 start_addr + offset);
2601 /* free page tables */
2602 dma_pte_free_pagetable(domain, start_addr,
2603 start_addr + offset);
2605 __free_iova(&domain->iovad, iova);
2608 sg->dma_address = start_addr + offset +
2609 ((u64)addr & (~PAGE_MASK));
2610 sg->dma_length = sg->length;
2614 /* it's a non-present to present mapping. Only flush if caching mode */
2615 if (cap_caching_mode(iommu->cap))
2616 iommu_flush_iotlb_psi(iommu, 0, start_addr,
2617 offset >> VTD_PAGE_SHIFT);
2619 iommu_flush_write_buffer(iommu);
2624 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
2629 struct dma_map_ops intel_dma_ops = {
2630 .alloc_coherent = intel_alloc_coherent,
2631 .free_coherent = intel_free_coherent,
2632 .map_sg = intel_map_sg,
2633 .unmap_sg = intel_unmap_sg,
2634 .map_page = intel_map_page,
2635 .unmap_page = intel_unmap_page,
2636 .mapping_error = intel_mapping_error,
2639 static inline int iommu_domain_cache_init(void)
2643 iommu_domain_cache = kmem_cache_create("iommu_domain",
2644 sizeof(struct dmar_domain),
2649 if (!iommu_domain_cache) {
2650 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2657 static inline int iommu_devinfo_cache_init(void)
2661 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2662 sizeof(struct device_domain_info),
2666 if (!iommu_devinfo_cache) {
2667 printk(KERN_ERR "Couldn't create devinfo cache\n");
2674 static inline int iommu_iova_cache_init(void)
2678 iommu_iova_cache = kmem_cache_create("iommu_iova",
2679 sizeof(struct iova),
2683 if (!iommu_iova_cache) {
2684 printk(KERN_ERR "Couldn't create iova cache\n");
2691 static int __init iommu_init_mempool(void)
2694 ret = iommu_iova_cache_init();
2698 ret = iommu_domain_cache_init();
2702 ret = iommu_devinfo_cache_init();
2706 kmem_cache_destroy(iommu_domain_cache);
2708 kmem_cache_destroy(iommu_iova_cache);
2713 static void __init iommu_exit_mempool(void)
2715 kmem_cache_destroy(iommu_devinfo_cache);
2716 kmem_cache_destroy(iommu_domain_cache);
2717 kmem_cache_destroy(iommu_iova_cache);
2721 static void __init init_no_remapping_devices(void)
2723 struct dmar_drhd_unit *drhd;
2725 for_each_drhd_unit(drhd) {
2726 if (!drhd->include_all) {
2728 for (i = 0; i < drhd->devices_cnt; i++)
2729 if (drhd->devices[i] != NULL)
2731 /* ignore DMAR unit if no pci devices exist */
2732 if (i == drhd->devices_cnt)
2740 for_each_drhd_unit(drhd) {
2742 if (drhd->ignored || drhd->include_all)
2745 for (i = 0; i < drhd->devices_cnt; i++)
2746 if (drhd->devices[i] &&
2747 !IS_GFX_DEVICE(drhd->devices[i]))
2750 if (i < drhd->devices_cnt)
2753 /* bypass IOMMU if it is just for gfx devices */
2755 for (i = 0; i < drhd->devices_cnt; i++) {
2756 if (!drhd->devices[i])
2758 drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
2763 #ifdef CONFIG_SUSPEND
2764 static int init_iommu_hw(void)
2766 struct dmar_drhd_unit *drhd;
2767 struct intel_iommu *iommu = NULL;
2769 for_each_active_iommu(iommu, drhd)
2771 dmar_reenable_qi(iommu);
2773 for_each_active_iommu(iommu, drhd) {
2774 iommu_flush_write_buffer(iommu);
2776 iommu_set_root_entry(iommu);
2778 iommu->flush.flush_context(iommu, 0, 0, 0,
2779 DMA_CCMD_GLOBAL_INVL);
2780 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2781 DMA_TLB_GLOBAL_FLUSH);
2782 iommu_disable_protect_mem_regions(iommu);
2783 iommu_enable_translation(iommu);
2789 static void iommu_flush_all(void)
2791 struct dmar_drhd_unit *drhd;
2792 struct intel_iommu *iommu;
2794 for_each_active_iommu(iommu, drhd) {
2795 iommu->flush.flush_context(iommu, 0, 0, 0,
2796 DMA_CCMD_GLOBAL_INVL);
2797 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2798 DMA_TLB_GLOBAL_FLUSH);
2802 static int iommu_suspend(struct sys_device *dev, pm_message_t state)
2804 struct dmar_drhd_unit *drhd;
2805 struct intel_iommu *iommu = NULL;
2808 for_each_active_iommu(iommu, drhd) {
2809 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
2811 if (!iommu->iommu_state)
2817 for_each_active_iommu(iommu, drhd) {
2818 iommu_disable_translation(iommu);
2820 spin_lock_irqsave(&iommu->register_lock, flag);
2822 iommu->iommu_state[SR_DMAR_FECTL_REG] =
2823 readl(iommu->reg + DMAR_FECTL_REG);
2824 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
2825 readl(iommu->reg + DMAR_FEDATA_REG);
2826 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
2827 readl(iommu->reg + DMAR_FEADDR_REG);
2828 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
2829 readl(iommu->reg + DMAR_FEUADDR_REG);
2831 spin_unlock_irqrestore(&iommu->register_lock, flag);
2836 for_each_active_iommu(iommu, drhd)
2837 kfree(iommu->iommu_state);
2842 static int iommu_resume(struct sys_device *dev)
2844 struct dmar_drhd_unit *drhd;
2845 struct intel_iommu *iommu = NULL;
2848 if (init_iommu_hw()) {
2849 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
2853 for_each_active_iommu(iommu, drhd) {
2855 spin_lock_irqsave(&iommu->register_lock, flag);
2857 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
2858 iommu->reg + DMAR_FECTL_REG);
2859 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
2860 iommu->reg + DMAR_FEDATA_REG);
2861 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
2862 iommu->reg + DMAR_FEADDR_REG);
2863 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
2864 iommu->reg + DMAR_FEUADDR_REG);
2866 spin_unlock_irqrestore(&iommu->register_lock, flag);
2869 for_each_active_iommu(iommu, drhd)
2870 kfree(iommu->iommu_state);
2875 static struct sysdev_class iommu_sysclass = {
2877 .resume = iommu_resume,
2878 .suspend = iommu_suspend,
2881 static struct sys_device device_iommu = {
2882 .cls = &iommu_sysclass,
2885 static int __init init_iommu_sysfs(void)
2889 error = sysdev_class_register(&iommu_sysclass);
2893 error = sysdev_register(&device_iommu);
2895 sysdev_class_unregister(&iommu_sysclass);
2901 static int __init init_iommu_sysfs(void)
2905 #endif /* CONFIG_PM */
2907 int __init intel_iommu_init(void)
2911 if (dmar_table_init())
2914 if (dmar_dev_scope_init())
2918 * Check the need for DMA-remapping initialization now.
2919 * Above initialization will also be used by Interrupt-remapping.
2921 if (no_iommu || (swiotlb && !iommu_pass_through) || dmar_disabled)
2924 iommu_init_mempool();
2925 dmar_init_reserved_ranges();
2927 init_no_remapping_devices();
2931 printk(KERN_ERR "IOMMU: dmar init failed\n");
2932 put_iova_domain(&reserved_iova_list);
2933 iommu_exit_mempool();
2937 "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
2939 init_timer(&unmap_timer);
2942 if (!iommu_pass_through) {
2944 "Multi-level page-table translation for DMAR.\n");
2945 dma_ops = &intel_dma_ops;
2948 "DMAR: Pass through translation for DMAR.\n");
2952 register_iommu(&intel_iommu_ops);
2957 static int vm_domain_add_dev_info(struct dmar_domain *domain,
2958 struct pci_dev *pdev)
2960 struct device_domain_info *info;
2961 unsigned long flags;
2963 info = alloc_devinfo_mem();
2967 info->segment = pci_domain_nr(pdev->bus);
2968 info->bus = pdev->bus->number;
2969 info->devfn = pdev->devfn;
2971 info->domain = domain;
2973 spin_lock_irqsave(&device_domain_lock, flags);
2974 list_add(&info->link, &domain->devices);
2975 list_add(&info->global, &device_domain_list);
2976 pdev->dev.archdata.iommu = info;
2977 spin_unlock_irqrestore(&device_domain_lock, flags);
2982 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
2983 struct pci_dev *pdev)
2985 struct pci_dev *tmp, *parent;
2987 if (!iommu || !pdev)
2990 /* dependent device detach */
2991 tmp = pci_find_upstream_pcie_bridge(pdev);
2992 /* Secondary interface's bus number and devfn 0 */
2994 parent = pdev->bus->self;
2995 while (parent != tmp) {
2996 iommu_detach_dev(iommu, parent->bus->number,
2998 parent = parent->bus->self;
3000 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
3001 iommu_detach_dev(iommu,
3002 tmp->subordinate->number, 0);
3003 else /* this is a legacy PCI bridge */
3004 iommu_detach_dev(iommu, tmp->bus->number,
3009 static void vm_domain_remove_one_dev_info(struct dmar_domain *domain,
3010 struct pci_dev *pdev)
3012 struct device_domain_info *info;
3013 struct intel_iommu *iommu;
3014 unsigned long flags;
3016 struct list_head *entry, *tmp;
3018 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3023 spin_lock_irqsave(&device_domain_lock, flags);
3024 list_for_each_safe(entry, tmp, &domain->devices) {
3025 info = list_entry(entry, struct device_domain_info, link);
3026 /* No need to compare PCI domain; it has to be the same */
3027 if (info->bus == pdev->bus->number &&
3028 info->devfn == pdev->devfn) {
3029 list_del(&info->link);
3030 list_del(&info->global);
3032 info->dev->dev.archdata.iommu = NULL;
3033 spin_unlock_irqrestore(&device_domain_lock, flags);
3035 iommu_disable_dev_iotlb(info);
3036 iommu_detach_dev(iommu, info->bus, info->devfn);
3037 iommu_detach_dependent_devices(iommu, pdev);
3038 free_devinfo_mem(info);
3040 spin_lock_irqsave(&device_domain_lock, flags);
3048 /* if there is no other devices under the same iommu
3049 * owned by this domain, clear this iommu in iommu_bmp
3050 * update iommu count and coherency
3052 if (iommu == device_to_iommu(info->segment, info->bus,
3058 unsigned long tmp_flags;
3059 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
3060 clear_bit(iommu->seq_id, &domain->iommu_bmp);
3061 domain->iommu_count--;
3062 domain_update_iommu_cap(domain);
3063 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
3066 spin_unlock_irqrestore(&device_domain_lock, flags);
3069 static void vm_domain_remove_all_dev_info(struct dmar_domain *domain)
3071 struct device_domain_info *info;
3072 struct intel_iommu *iommu;
3073 unsigned long flags1, flags2;
3075 spin_lock_irqsave(&device_domain_lock, flags1);
3076 while (!list_empty(&domain->devices)) {
3077 info = list_entry(domain->devices.next,
3078 struct device_domain_info, link);
3079 list_del(&info->link);
3080 list_del(&info->global);
3082 info->dev->dev.archdata.iommu = NULL;
3084 spin_unlock_irqrestore(&device_domain_lock, flags1);
3086 iommu_disable_dev_iotlb(info);
3087 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
3088 iommu_detach_dev(iommu, info->bus, info->devfn);
3089 iommu_detach_dependent_devices(iommu, info->dev);
3091 /* clear this iommu in iommu_bmp, update iommu count
3094 spin_lock_irqsave(&domain->iommu_lock, flags2);
3095 if (test_and_clear_bit(iommu->seq_id,
3096 &domain->iommu_bmp)) {
3097 domain->iommu_count--;
3098 domain_update_iommu_cap(domain);
3100 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
3102 free_devinfo_mem(info);
3103 spin_lock_irqsave(&device_domain_lock, flags1);
3105 spin_unlock_irqrestore(&device_domain_lock, flags1);
3108 /* domain id for virtual machine, it won't be set in context */
3109 static unsigned long vm_domid;
3111 static int vm_domain_min_agaw(struct dmar_domain *domain)
3114 int min_agaw = domain->agaw;
3116 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
3117 for (; i < g_num_of_iommus; ) {
3118 if (min_agaw > g_iommus[i]->agaw)
3119 min_agaw = g_iommus[i]->agaw;
3121 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
3127 static struct dmar_domain *iommu_alloc_vm_domain(void)
3129 struct dmar_domain *domain;
3131 domain = alloc_domain_mem();
3135 domain->id = vm_domid++;
3136 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
3137 domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
3142 static int vm_domain_init(struct dmar_domain *domain, int guest_width)
3146 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
3147 spin_lock_init(&domain->mapping_lock);
3148 spin_lock_init(&domain->iommu_lock);
3150 domain_reserve_special_ranges(domain);
3152 /* calculate AGAW */
3153 domain->gaw = guest_width;
3154 adjust_width = guestwidth_to_adjustwidth(guest_width);
3155 domain->agaw = width_to_agaw(adjust_width);
3157 INIT_LIST_HEAD(&domain->devices);
3159 domain->iommu_count = 0;
3160 domain->iommu_coherency = 0;
3161 domain->max_addr = 0;
3163 /* always allocate the top pgd */
3164 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
3167 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
3171 static void iommu_free_vm_domain(struct dmar_domain *domain)
3173 unsigned long flags;
3174 struct dmar_drhd_unit *drhd;
3175 struct intel_iommu *iommu;
3177 unsigned long ndomains;
3179 for_each_drhd_unit(drhd) {
3182 iommu = drhd->iommu;
3184 ndomains = cap_ndoms(iommu->cap);
3185 i = find_first_bit(iommu->domain_ids, ndomains);
3186 for (; i < ndomains; ) {
3187 if (iommu->domains[i] == domain) {
3188 spin_lock_irqsave(&iommu->lock, flags);
3189 clear_bit(i, iommu->domain_ids);
3190 iommu->domains[i] = NULL;
3191 spin_unlock_irqrestore(&iommu->lock, flags);
3194 i = find_next_bit(iommu->domain_ids, ndomains, i+1);
3199 static void vm_domain_exit(struct dmar_domain *domain)
3203 /* Domain 0 is reserved, so dont process it */
3207 vm_domain_remove_all_dev_info(domain);
3209 put_iova_domain(&domain->iovad);
3210 end = DOMAIN_MAX_ADDR(domain->gaw);
3211 end = end & (~VTD_PAGE_MASK);
3214 dma_pte_clear_range(domain, 0, end);
3216 /* free page tables */
3217 dma_pte_free_pagetable(domain, 0, end);
3219 iommu_free_vm_domain(domain);
3220 free_domain_mem(domain);
3223 static int intel_iommu_domain_init(struct iommu_domain *domain)
3225 struct dmar_domain *dmar_domain;
3227 dmar_domain = iommu_alloc_vm_domain();
3230 "intel_iommu_domain_init: dmar_domain == NULL\n");
3233 if (vm_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
3235 "intel_iommu_domain_init() failed\n");
3236 vm_domain_exit(dmar_domain);
3239 domain->priv = dmar_domain;
3244 static void intel_iommu_domain_destroy(struct iommu_domain *domain)
3246 struct dmar_domain *dmar_domain = domain->priv;
3248 domain->priv = NULL;
3249 vm_domain_exit(dmar_domain);
3252 static int intel_iommu_attach_device(struct iommu_domain *domain,
3255 struct dmar_domain *dmar_domain = domain->priv;
3256 struct pci_dev *pdev = to_pci_dev(dev);
3257 struct intel_iommu *iommu;
3262 /* normally pdev is not mapped */
3263 if (unlikely(domain_context_mapped(pdev))) {
3264 struct dmar_domain *old_domain;
3266 old_domain = find_domain(pdev);
3268 if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
3269 vm_domain_remove_one_dev_info(old_domain, pdev);
3271 domain_remove_dev_info(old_domain);
3275 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3280 /* check if this iommu agaw is sufficient for max mapped address */
3281 addr_width = agaw_to_width(iommu->agaw);
3282 end = DOMAIN_MAX_ADDR(addr_width);
3283 end = end & VTD_PAGE_MASK;
3284 if (end < dmar_domain->max_addr) {
3285 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3286 "sufficient for the mapped address (%llx)\n",
3287 __func__, iommu->agaw, dmar_domain->max_addr);
3291 ret = vm_domain_add_dev_info(dmar_domain, pdev);
3295 ret = domain_context_mapping(dmar_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
3299 static void intel_iommu_detach_device(struct iommu_domain *domain,
3302 struct dmar_domain *dmar_domain = domain->priv;
3303 struct pci_dev *pdev = to_pci_dev(dev);
3305 vm_domain_remove_one_dev_info(dmar_domain, pdev);
3308 static int intel_iommu_map_range(struct iommu_domain *domain,
3309 unsigned long iova, phys_addr_t hpa,
3310 size_t size, int iommu_prot)
3312 struct dmar_domain *dmar_domain = domain->priv;
3318 if (iommu_prot & IOMMU_READ)
3319 prot |= DMA_PTE_READ;
3320 if (iommu_prot & IOMMU_WRITE)
3321 prot |= DMA_PTE_WRITE;
3322 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
3323 prot |= DMA_PTE_SNP;
3325 max_addr = (iova & VTD_PAGE_MASK) + VTD_PAGE_ALIGN(size);
3326 if (dmar_domain->max_addr < max_addr) {
3330 /* check if minimum agaw is sufficient for mapped address */
3331 min_agaw = vm_domain_min_agaw(dmar_domain);
3332 addr_width = agaw_to_width(min_agaw);
3333 end = DOMAIN_MAX_ADDR(addr_width);
3334 end = end & VTD_PAGE_MASK;
3335 if (end < max_addr) {
3336 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3337 "sufficient for the mapped address (%llx)\n",
3338 __func__, min_agaw, max_addr);
3341 dmar_domain->max_addr = max_addr;
3344 ret = domain_page_mapping(dmar_domain, iova, hpa, size, prot);
3348 static void intel_iommu_unmap_range(struct iommu_domain *domain,
3349 unsigned long iova, size_t size)
3351 struct dmar_domain *dmar_domain = domain->priv;
3354 /* The address might not be aligned */
3355 base = iova & VTD_PAGE_MASK;
3356 size = VTD_PAGE_ALIGN(size);
3357 dma_pte_clear_range(dmar_domain, base, base + size);
3359 if (dmar_domain->max_addr == base + size)
3360 dmar_domain->max_addr = base;
3363 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
3366 struct dmar_domain *dmar_domain = domain->priv;
3367 struct dma_pte *pte;
3370 pte = addr_to_dma_pte(dmar_domain, iova);
3372 phys = dma_pte_addr(pte);
3377 static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
3380 struct dmar_domain *dmar_domain = domain->priv;
3382 if (cap == IOMMU_CAP_CACHE_COHERENCY)
3383 return dmar_domain->iommu_snooping;
3388 static struct iommu_ops intel_iommu_ops = {
3389 .domain_init = intel_iommu_domain_init,
3390 .domain_destroy = intel_iommu_domain_destroy,
3391 .attach_dev = intel_iommu_attach_device,
3392 .detach_dev = intel_iommu_detach_device,
3393 .map = intel_iommu_map_range,
3394 .unmap = intel_iommu_unmap_range,
3395 .iova_to_phys = intel_iommu_iova_to_phys,
3396 .domain_has_cap = intel_iommu_domain_has_cap,
3399 static void __devinit quirk_iommu_rwbf(struct pci_dev *dev)
3402 * Mobile 4 Series Chipset neglects to set RWBF capability,
3405 printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
3409 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);