intel-iommu: Create new iommu_domain_identity_map() function
[linux-2.6/verdex.git] / drivers / pci / intel-iommu.c
blob3a4f347e2f882dd4d70be001edf2a655d86cdcc6
1 /*
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
11 * more details.
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>
42 #include "pci.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 #ifndef PHYSICAL_PAGE_MASK
65 #define PHYSICAL_PAGE_MASK PAGE_MASK
66 #endif
68 /* global iommu list, set NULL for ignored DMAR units */
69 static struct intel_iommu **g_iommus;
71 static int rwbf_quirk;
74 * 0: Present
75 * 1-11: Reserved
76 * 12-63: Context Ptr (12 - (haw-1))
77 * 64-127: Reserved
79 struct root_entry {
80 u64 val;
81 u64 rsvd1;
83 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
84 static inline bool root_present(struct root_entry *root)
86 return (root->val & 1);
88 static inline void set_root_present(struct root_entry *root)
90 root->val |= 1;
92 static inline void set_root_value(struct root_entry *root, unsigned long value)
94 root->val |= value & VTD_PAGE_MASK;
97 static inline struct context_entry *
98 get_context_addr_from_root(struct root_entry *root)
100 return (struct context_entry *)
101 (root_present(root)?phys_to_virt(
102 root->val & VTD_PAGE_MASK) :
103 NULL);
107 * low 64 bits:
108 * 0: present
109 * 1: fault processing disable
110 * 2-3: translation type
111 * 12-63: address space root
112 * high 64 bits:
113 * 0-2: address width
114 * 3-6: aval
115 * 8-23: domain id
117 struct context_entry {
118 u64 lo;
119 u64 hi;
122 static inline bool context_present(struct context_entry *context)
124 return (context->lo & 1);
126 static inline void context_set_present(struct context_entry *context)
128 context->lo |= 1;
131 static inline void context_set_fault_enable(struct context_entry *context)
133 context->lo &= (((u64)-1) << 2) | 1;
136 static inline void context_set_translation_type(struct context_entry *context,
137 unsigned long value)
139 context->lo &= (((u64)-1) << 4) | 3;
140 context->lo |= (value & 3) << 2;
143 static inline void context_set_address_root(struct context_entry *context,
144 unsigned long value)
146 context->lo |= value & VTD_PAGE_MASK;
149 static inline void context_set_address_width(struct context_entry *context,
150 unsigned long value)
152 context->hi |= value & 7;
155 static inline void context_set_domain_id(struct context_entry *context,
156 unsigned long value)
158 context->hi |= (value & ((1 << 16) - 1)) << 8;
161 static inline void context_clear_entry(struct context_entry *context)
163 context->lo = 0;
164 context->hi = 0;
168 * 0: readable
169 * 1: writable
170 * 2-6: reserved
171 * 7: super page
172 * 8-10: available
173 * 11: snoop behavior
174 * 12-63: Host physcial address
176 struct dma_pte {
177 u64 val;
180 static inline void dma_clear_pte(struct dma_pte *pte)
182 pte->val = 0;
185 static inline void dma_set_pte_readable(struct dma_pte *pte)
187 pte->val |= DMA_PTE_READ;
190 static inline void dma_set_pte_writable(struct dma_pte *pte)
192 pte->val |= DMA_PTE_WRITE;
195 static inline void dma_set_pte_snp(struct dma_pte *pte)
197 pte->val |= DMA_PTE_SNP;
200 static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)
202 pte->val = (pte->val & ~3) | (prot & 3);
205 static inline u64 dma_pte_addr(struct dma_pte *pte)
207 return (pte->val & VTD_PAGE_MASK);
210 static inline void dma_set_pte_addr(struct dma_pte *pte, u64 addr)
212 pte->val |= (addr & VTD_PAGE_MASK);
215 static inline bool dma_pte_present(struct dma_pte *pte)
217 return (pte->val & 3) != 0;
221 * This domain is a statically identity mapping domain.
222 * 1. This domain creats a static 1:1 mapping to all usable memory.
223 * 2. It maps to each iommu if successful.
224 * 3. Each iommu mapps to this domain if successful.
226 struct dmar_domain *si_domain;
228 /* devices under the same p2p bridge are owned in one domain */
229 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
231 /* domain represents a virtual machine, more than one devices
232 * across iommus may be owned in one domain, e.g. kvm guest.
234 #define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 1)
236 /* si_domain contains mulitple devices */
237 #define DOMAIN_FLAG_STATIC_IDENTITY (1 << 2)
239 struct dmar_domain {
240 int id; /* domain id */
241 unsigned long iommu_bmp; /* bitmap of iommus this domain uses*/
243 struct list_head devices; /* all devices' list */
244 struct iova_domain iovad; /* iova's that belong to this domain */
246 struct dma_pte *pgd; /* virtual address */
247 spinlock_t mapping_lock; /* page table lock */
248 int gaw; /* max guest address width */
250 /* adjusted guest address width, 0 is level 2 30-bit */
251 int agaw;
253 int flags; /* flags to find out type of domain */
255 int iommu_coherency;/* indicate coherency of iommu access */
256 int iommu_snooping; /* indicate snooping control feature*/
257 int iommu_count; /* reference count of iommu */
258 spinlock_t iommu_lock; /* protect iommu set in domain */
259 u64 max_addr; /* maximum mapped address */
262 /* PCI domain-device relationship */
263 struct device_domain_info {
264 struct list_head link; /* link to domain siblings */
265 struct list_head global; /* link to global list */
266 int segment; /* PCI domain */
267 u8 bus; /* PCI bus number */
268 u8 devfn; /* PCI devfn number */
269 struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
270 struct intel_iommu *iommu; /* IOMMU used by this device */
271 struct dmar_domain *domain; /* pointer to domain */
274 static void flush_unmaps_timeout(unsigned long data);
276 DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
278 #define HIGH_WATER_MARK 250
279 struct deferred_flush_tables {
280 int next;
281 struct iova *iova[HIGH_WATER_MARK];
282 struct dmar_domain *domain[HIGH_WATER_MARK];
285 static struct deferred_flush_tables *deferred_flush;
287 /* bitmap for indexing intel_iommus */
288 static int g_num_of_iommus;
290 static DEFINE_SPINLOCK(async_umap_flush_lock);
291 static LIST_HEAD(unmaps_to_do);
293 static int timer_on;
294 static long list_size;
296 static void domain_remove_dev_info(struct dmar_domain *domain);
298 #ifdef CONFIG_DMAR_DEFAULT_ON
299 int dmar_disabled = 0;
300 #else
301 int dmar_disabled = 1;
302 #endif /*CONFIG_DMAR_DEFAULT_ON*/
304 static int __initdata dmar_map_gfx = 1;
305 static int dmar_forcedac;
306 static int intel_iommu_strict;
308 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
309 static DEFINE_SPINLOCK(device_domain_lock);
310 static LIST_HEAD(device_domain_list);
312 static struct iommu_ops intel_iommu_ops;
314 static int __init intel_iommu_setup(char *str)
316 if (!str)
317 return -EINVAL;
318 while (*str) {
319 if (!strncmp(str, "on", 2)) {
320 dmar_disabled = 0;
321 printk(KERN_INFO "Intel-IOMMU: enabled\n");
322 } else if (!strncmp(str, "off", 3)) {
323 dmar_disabled = 1;
324 printk(KERN_INFO "Intel-IOMMU: disabled\n");
325 } else if (!strncmp(str, "igfx_off", 8)) {
326 dmar_map_gfx = 0;
327 printk(KERN_INFO
328 "Intel-IOMMU: disable GFX device mapping\n");
329 } else if (!strncmp(str, "forcedac", 8)) {
330 printk(KERN_INFO
331 "Intel-IOMMU: Forcing DAC for PCI devices\n");
332 dmar_forcedac = 1;
333 } else if (!strncmp(str, "strict", 6)) {
334 printk(KERN_INFO
335 "Intel-IOMMU: disable batched IOTLB flush\n");
336 intel_iommu_strict = 1;
339 str += strcspn(str, ",");
340 while (*str == ',')
341 str++;
343 return 0;
345 __setup("intel_iommu=", intel_iommu_setup);
347 static struct kmem_cache *iommu_domain_cache;
348 static struct kmem_cache *iommu_devinfo_cache;
349 static struct kmem_cache *iommu_iova_cache;
351 static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
353 unsigned int flags;
354 void *vaddr;
356 /* trying to avoid low memory issues */
357 flags = current->flags & PF_MEMALLOC;
358 current->flags |= PF_MEMALLOC;
359 vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
360 current->flags &= (~PF_MEMALLOC | flags);
361 return vaddr;
365 static inline void *alloc_pgtable_page(void)
367 unsigned int flags;
368 void *vaddr;
370 /* trying to avoid low memory issues */
371 flags = current->flags & PF_MEMALLOC;
372 current->flags |= PF_MEMALLOC;
373 vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
374 current->flags &= (~PF_MEMALLOC | flags);
375 return vaddr;
378 static inline void free_pgtable_page(void *vaddr)
380 free_page((unsigned long)vaddr);
383 static inline void *alloc_domain_mem(void)
385 return iommu_kmem_cache_alloc(iommu_domain_cache);
388 static void free_domain_mem(void *vaddr)
390 kmem_cache_free(iommu_domain_cache, vaddr);
393 static inline void * alloc_devinfo_mem(void)
395 return iommu_kmem_cache_alloc(iommu_devinfo_cache);
398 static inline void free_devinfo_mem(void *vaddr)
400 kmem_cache_free(iommu_devinfo_cache, vaddr);
403 struct iova *alloc_iova_mem(void)
405 return iommu_kmem_cache_alloc(iommu_iova_cache);
408 void free_iova_mem(struct iova *iova)
410 kmem_cache_free(iommu_iova_cache, iova);
414 static inline int width_to_agaw(int width);
416 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
418 unsigned long sagaw;
419 int agaw = -1;
421 sagaw = cap_sagaw(iommu->cap);
422 for (agaw = width_to_agaw(max_gaw);
423 agaw >= 0; agaw--) {
424 if (test_bit(agaw, &sagaw))
425 break;
428 return agaw;
432 * Calculate max SAGAW for each iommu.
434 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
436 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
440 * calculate agaw for each iommu.
441 * "SAGAW" may be different across iommus, use a default agaw, and
442 * get a supported less agaw for iommus that don't support the default agaw.
444 int iommu_calculate_agaw(struct intel_iommu *iommu)
446 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
449 /* This functionin only returns single iommu in a domain */
450 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
452 int iommu_id;
454 /* si_domain and vm domain should not get here. */
455 BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
456 BUG_ON(domain->flags & DOMAIN_FLAG_STATIC_IDENTITY);
458 iommu_id = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
459 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
460 return NULL;
462 return g_iommus[iommu_id];
465 static void domain_update_iommu_coherency(struct dmar_domain *domain)
467 int i;
469 domain->iommu_coherency = 1;
471 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
472 for (; i < g_num_of_iommus; ) {
473 if (!ecap_coherent(g_iommus[i]->ecap)) {
474 domain->iommu_coherency = 0;
475 break;
477 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
481 static void domain_update_iommu_snooping(struct dmar_domain *domain)
483 int i;
485 domain->iommu_snooping = 1;
487 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
488 for (; i < g_num_of_iommus; ) {
489 if (!ecap_sc_support(g_iommus[i]->ecap)) {
490 domain->iommu_snooping = 0;
491 break;
493 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
497 /* Some capabilities may be different across iommus */
498 static void domain_update_iommu_cap(struct dmar_domain *domain)
500 domain_update_iommu_coherency(domain);
501 domain_update_iommu_snooping(domain);
504 static struct intel_iommu *device_to_iommu(int segment, u8 bus, u8 devfn)
506 struct dmar_drhd_unit *drhd = NULL;
507 int i;
509 for_each_drhd_unit(drhd) {
510 if (drhd->ignored)
511 continue;
512 if (segment != drhd->segment)
513 continue;
515 for (i = 0; i < drhd->devices_cnt; i++) {
516 if (drhd->devices[i] &&
517 drhd->devices[i]->bus->number == bus &&
518 drhd->devices[i]->devfn == devfn)
519 return drhd->iommu;
520 if (drhd->devices[i] &&
521 drhd->devices[i]->subordinate &&
522 drhd->devices[i]->subordinate->number <= bus &&
523 drhd->devices[i]->subordinate->subordinate >= bus)
524 return drhd->iommu;
527 if (drhd->include_all)
528 return drhd->iommu;
531 return NULL;
534 static void domain_flush_cache(struct dmar_domain *domain,
535 void *addr, int size)
537 if (!domain->iommu_coherency)
538 clflush_cache_range(addr, size);
541 /* Gets context entry for a given bus and devfn */
542 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
543 u8 bus, u8 devfn)
545 struct root_entry *root;
546 struct context_entry *context;
547 unsigned long phy_addr;
548 unsigned long flags;
550 spin_lock_irqsave(&iommu->lock, flags);
551 root = &iommu->root_entry[bus];
552 context = get_context_addr_from_root(root);
553 if (!context) {
554 context = (struct context_entry *)alloc_pgtable_page();
555 if (!context) {
556 spin_unlock_irqrestore(&iommu->lock, flags);
557 return NULL;
559 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
560 phy_addr = virt_to_phys((void *)context);
561 set_root_value(root, phy_addr);
562 set_root_present(root);
563 __iommu_flush_cache(iommu, root, sizeof(*root));
565 spin_unlock_irqrestore(&iommu->lock, flags);
566 return &context[devfn];
569 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
571 struct root_entry *root;
572 struct context_entry *context;
573 int ret;
574 unsigned long flags;
576 spin_lock_irqsave(&iommu->lock, flags);
577 root = &iommu->root_entry[bus];
578 context = get_context_addr_from_root(root);
579 if (!context) {
580 ret = 0;
581 goto out;
583 ret = context_present(&context[devfn]);
584 out:
585 spin_unlock_irqrestore(&iommu->lock, flags);
586 return ret;
589 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
591 struct root_entry *root;
592 struct context_entry *context;
593 unsigned long flags;
595 spin_lock_irqsave(&iommu->lock, flags);
596 root = &iommu->root_entry[bus];
597 context = get_context_addr_from_root(root);
598 if (context) {
599 context_clear_entry(&context[devfn]);
600 __iommu_flush_cache(iommu, &context[devfn], \
601 sizeof(*context));
603 spin_unlock_irqrestore(&iommu->lock, flags);
606 static void free_context_table(struct intel_iommu *iommu)
608 struct root_entry *root;
609 int i;
610 unsigned long flags;
611 struct context_entry *context;
613 spin_lock_irqsave(&iommu->lock, flags);
614 if (!iommu->root_entry) {
615 goto out;
617 for (i = 0; i < ROOT_ENTRY_NR; i++) {
618 root = &iommu->root_entry[i];
619 context = get_context_addr_from_root(root);
620 if (context)
621 free_pgtable_page(context);
623 free_pgtable_page(iommu->root_entry);
624 iommu->root_entry = NULL;
625 out:
626 spin_unlock_irqrestore(&iommu->lock, flags);
629 /* page table handling */
630 #define LEVEL_STRIDE (9)
631 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
633 static inline int agaw_to_level(int agaw)
635 return agaw + 2;
638 static inline int agaw_to_width(int agaw)
640 return 30 + agaw * LEVEL_STRIDE;
644 static inline int width_to_agaw(int width)
646 return (width - 30) / LEVEL_STRIDE;
649 static inline unsigned int level_to_offset_bits(int level)
651 return (12 + (level - 1) * LEVEL_STRIDE);
654 static inline int address_level_offset(u64 addr, int level)
656 return ((addr >> level_to_offset_bits(level)) & LEVEL_MASK);
659 static inline u64 level_mask(int level)
661 return ((u64)-1 << level_to_offset_bits(level));
664 static inline u64 level_size(int level)
666 return ((u64)1 << level_to_offset_bits(level));
669 static inline u64 align_to_level(u64 addr, int level)
671 return ((addr + level_size(level) - 1) & level_mask(level));
674 static struct dma_pte * addr_to_dma_pte(struct dmar_domain *domain, u64 addr)
676 int addr_width = agaw_to_width(domain->agaw);
677 struct dma_pte *parent, *pte = NULL;
678 int level = agaw_to_level(domain->agaw);
679 int offset;
680 unsigned long flags;
682 BUG_ON(!domain->pgd);
684 addr &= (((u64)1) << addr_width) - 1;
685 parent = domain->pgd;
687 spin_lock_irqsave(&domain->mapping_lock, flags);
688 while (level > 0) {
689 void *tmp_page;
691 offset = address_level_offset(addr, level);
692 pte = &parent[offset];
693 if (level == 1)
694 break;
696 if (!dma_pte_present(pte)) {
697 tmp_page = alloc_pgtable_page();
699 if (!tmp_page) {
700 spin_unlock_irqrestore(&domain->mapping_lock,
701 flags);
702 return NULL;
704 domain_flush_cache(domain, tmp_page, PAGE_SIZE);
705 dma_set_pte_addr(pte, virt_to_phys(tmp_page));
707 * high level table always sets r/w, last level page
708 * table control read/write
710 dma_set_pte_readable(pte);
711 dma_set_pte_writable(pte);
712 domain_flush_cache(domain, pte, sizeof(*pte));
714 parent = phys_to_virt(dma_pte_addr(pte));
715 level--;
718 spin_unlock_irqrestore(&domain->mapping_lock, flags);
719 return pte;
722 /* return address's pte at specific level */
723 static struct dma_pte *dma_addr_level_pte(struct dmar_domain *domain, u64 addr,
724 int level)
726 struct dma_pte *parent, *pte = NULL;
727 int total = agaw_to_level(domain->agaw);
728 int offset;
730 parent = domain->pgd;
731 while (level <= total) {
732 offset = address_level_offset(addr, total);
733 pte = &parent[offset];
734 if (level == total)
735 return pte;
737 if (!dma_pte_present(pte))
738 break;
739 parent = phys_to_virt(dma_pte_addr(pte));
740 total--;
742 return NULL;
745 /* clear one page's page table */
746 static void dma_pte_clear_one(struct dmar_domain *domain, u64 addr)
748 struct dma_pte *pte = NULL;
750 /* get last level pte */
751 pte = dma_addr_level_pte(domain, addr, 1);
753 if (pte) {
754 dma_clear_pte(pte);
755 domain_flush_cache(domain, pte, sizeof(*pte));
759 /* clear last level pte, a tlb flush should be followed */
760 static void dma_pte_clear_range(struct dmar_domain *domain, u64 start, u64 end)
762 int addr_width = agaw_to_width(domain->agaw);
763 int npages;
765 start &= (((u64)1) << addr_width) - 1;
766 end &= (((u64)1) << addr_width) - 1;
767 /* in case it's partial page */
768 start &= PAGE_MASK;
769 end = PAGE_ALIGN(end);
770 npages = (end - start) / VTD_PAGE_SIZE;
772 /* we don't need lock here, nobody else touches the iova range */
773 while (npages--) {
774 dma_pte_clear_one(domain, start);
775 start += VTD_PAGE_SIZE;
779 /* free page table pages. last level pte should already be cleared */
780 static void dma_pte_free_pagetable(struct dmar_domain *domain,
781 u64 start, u64 end)
783 int addr_width = agaw_to_width(domain->agaw);
784 struct dma_pte *pte;
785 int total = agaw_to_level(domain->agaw);
786 int level;
787 u64 tmp;
789 start &= (((u64)1) << addr_width) - 1;
790 end &= (((u64)1) << addr_width) - 1;
792 /* we don't need lock here, nobody else touches the iova range */
793 level = 2;
794 while (level <= total) {
795 tmp = align_to_level(start, level);
796 if (tmp >= end || (tmp + level_size(level) > end))
797 return;
799 while (tmp < end) {
800 pte = dma_addr_level_pte(domain, tmp, level);
801 if (pte) {
802 free_pgtable_page(
803 phys_to_virt(dma_pte_addr(pte)));
804 dma_clear_pte(pte);
805 domain_flush_cache(domain, pte, sizeof(*pte));
807 tmp += level_size(level);
809 level++;
811 /* free pgd */
812 if (start == 0 && end >= ((((u64)1) << addr_width) - 1)) {
813 free_pgtable_page(domain->pgd);
814 domain->pgd = NULL;
818 /* iommu handling */
819 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
821 struct root_entry *root;
822 unsigned long flags;
824 root = (struct root_entry *)alloc_pgtable_page();
825 if (!root)
826 return -ENOMEM;
828 __iommu_flush_cache(iommu, root, ROOT_SIZE);
830 spin_lock_irqsave(&iommu->lock, flags);
831 iommu->root_entry = root;
832 spin_unlock_irqrestore(&iommu->lock, flags);
834 return 0;
837 static void iommu_set_root_entry(struct intel_iommu *iommu)
839 void *addr;
840 u32 sts;
841 unsigned long flag;
843 addr = iommu->root_entry;
845 spin_lock_irqsave(&iommu->register_lock, flag);
846 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
848 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
850 /* Make sure hardware complete it */
851 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
852 readl, (sts & DMA_GSTS_RTPS), sts);
854 spin_unlock_irqrestore(&iommu->register_lock, flag);
857 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
859 u32 val;
860 unsigned long flag;
862 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
863 return;
865 spin_lock_irqsave(&iommu->register_lock, flag);
866 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
868 /* Make sure hardware complete it */
869 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
870 readl, (!(val & DMA_GSTS_WBFS)), val);
872 spin_unlock_irqrestore(&iommu->register_lock, flag);
875 /* return value determine if we need a write buffer flush */
876 static void __iommu_flush_context(struct intel_iommu *iommu,
877 u16 did, u16 source_id, u8 function_mask,
878 u64 type)
880 u64 val = 0;
881 unsigned long flag;
883 switch (type) {
884 case DMA_CCMD_GLOBAL_INVL:
885 val = DMA_CCMD_GLOBAL_INVL;
886 break;
887 case DMA_CCMD_DOMAIN_INVL:
888 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
889 break;
890 case DMA_CCMD_DEVICE_INVL:
891 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
892 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
893 break;
894 default:
895 BUG();
897 val |= DMA_CCMD_ICC;
899 spin_lock_irqsave(&iommu->register_lock, flag);
900 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
902 /* Make sure hardware complete it */
903 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
904 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
906 spin_unlock_irqrestore(&iommu->register_lock, flag);
909 /* return value determine if we need a write buffer flush */
910 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
911 u64 addr, unsigned int size_order, u64 type)
913 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
914 u64 val = 0, val_iva = 0;
915 unsigned long flag;
917 switch (type) {
918 case DMA_TLB_GLOBAL_FLUSH:
919 /* global flush doesn't need set IVA_REG */
920 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
921 break;
922 case DMA_TLB_DSI_FLUSH:
923 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
924 break;
925 case DMA_TLB_PSI_FLUSH:
926 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
927 /* Note: always flush non-leaf currently */
928 val_iva = size_order | addr;
929 break;
930 default:
931 BUG();
933 /* Note: set drain read/write */
934 #if 0
936 * This is probably to be super secure.. Looks like we can
937 * ignore it without any impact.
939 if (cap_read_drain(iommu->cap))
940 val |= DMA_TLB_READ_DRAIN;
941 #endif
942 if (cap_write_drain(iommu->cap))
943 val |= DMA_TLB_WRITE_DRAIN;
945 spin_lock_irqsave(&iommu->register_lock, flag);
946 /* Note: Only uses first TLB reg currently */
947 if (val_iva)
948 dmar_writeq(iommu->reg + tlb_offset, val_iva);
949 dmar_writeq(iommu->reg + tlb_offset + 8, val);
951 /* Make sure hardware complete it */
952 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
953 dmar_readq, (!(val & DMA_TLB_IVT)), val);
955 spin_unlock_irqrestore(&iommu->register_lock, flag);
957 /* check IOTLB invalidation granularity */
958 if (DMA_TLB_IAIG(val) == 0)
959 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
960 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
961 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
962 (unsigned long long)DMA_TLB_IIRG(type),
963 (unsigned long long)DMA_TLB_IAIG(val));
966 static struct device_domain_info *iommu_support_dev_iotlb(
967 struct dmar_domain *domain, int segment, u8 bus, u8 devfn)
969 int found = 0;
970 unsigned long flags;
971 struct device_domain_info *info;
972 struct intel_iommu *iommu = device_to_iommu(segment, bus, devfn);
974 if (!ecap_dev_iotlb_support(iommu->ecap))
975 return NULL;
977 if (!iommu->qi)
978 return NULL;
980 spin_lock_irqsave(&device_domain_lock, flags);
981 list_for_each_entry(info, &domain->devices, link)
982 if (info->bus == bus && info->devfn == devfn) {
983 found = 1;
984 break;
986 spin_unlock_irqrestore(&device_domain_lock, flags);
988 if (!found || !info->dev)
989 return NULL;
991 if (!pci_find_ext_capability(info->dev, PCI_EXT_CAP_ID_ATS))
992 return NULL;
994 if (!dmar_find_matched_atsr_unit(info->dev))
995 return NULL;
997 info->iommu = iommu;
999 return info;
1002 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
1004 if (!info)
1005 return;
1007 pci_enable_ats(info->dev, VTD_PAGE_SHIFT);
1010 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1012 if (!info->dev || !pci_ats_enabled(info->dev))
1013 return;
1015 pci_disable_ats(info->dev);
1018 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1019 u64 addr, unsigned mask)
1021 u16 sid, qdep;
1022 unsigned long flags;
1023 struct device_domain_info *info;
1025 spin_lock_irqsave(&device_domain_lock, flags);
1026 list_for_each_entry(info, &domain->devices, link) {
1027 if (!info->dev || !pci_ats_enabled(info->dev))
1028 continue;
1030 sid = info->bus << 8 | info->devfn;
1031 qdep = pci_ats_queue_depth(info->dev);
1032 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1034 spin_unlock_irqrestore(&device_domain_lock, flags);
1037 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
1038 u64 addr, unsigned int pages)
1040 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1042 BUG_ON(addr & (~VTD_PAGE_MASK));
1043 BUG_ON(pages == 0);
1046 * Fallback to domain selective flush if no PSI support or the size is
1047 * too big.
1048 * PSI requires page size to be 2 ^ x, and the base address is naturally
1049 * aligned to the size
1051 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1052 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1053 DMA_TLB_DSI_FLUSH);
1054 else
1055 iommu->flush.flush_iotlb(iommu, did, addr, mask,
1056 DMA_TLB_PSI_FLUSH);
1059 * In caching mode, domain ID 0 is reserved for non-present to present
1060 * mapping flush. Device IOTLB doesn't need to be flushed in this case.
1062 if (!cap_caching_mode(iommu->cap) || did)
1063 iommu_flush_dev_iotlb(iommu->domains[did], addr, mask);
1066 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1068 u32 pmen;
1069 unsigned long flags;
1071 spin_lock_irqsave(&iommu->register_lock, flags);
1072 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1073 pmen &= ~DMA_PMEN_EPM;
1074 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1076 /* wait for the protected region status bit to clear */
1077 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1078 readl, !(pmen & DMA_PMEN_PRS), pmen);
1080 spin_unlock_irqrestore(&iommu->register_lock, flags);
1083 static int iommu_enable_translation(struct intel_iommu *iommu)
1085 u32 sts;
1086 unsigned long flags;
1088 spin_lock_irqsave(&iommu->register_lock, flags);
1089 iommu->gcmd |= DMA_GCMD_TE;
1090 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1092 /* Make sure hardware complete it */
1093 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1094 readl, (sts & DMA_GSTS_TES), sts);
1096 spin_unlock_irqrestore(&iommu->register_lock, flags);
1097 return 0;
1100 static int iommu_disable_translation(struct intel_iommu *iommu)
1102 u32 sts;
1103 unsigned long flag;
1105 spin_lock_irqsave(&iommu->register_lock, flag);
1106 iommu->gcmd &= ~DMA_GCMD_TE;
1107 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1109 /* Make sure hardware complete it */
1110 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1111 readl, (!(sts & DMA_GSTS_TES)), sts);
1113 spin_unlock_irqrestore(&iommu->register_lock, flag);
1114 return 0;
1118 static int iommu_init_domains(struct intel_iommu *iommu)
1120 unsigned long ndomains;
1121 unsigned long nlongs;
1123 ndomains = cap_ndoms(iommu->cap);
1124 pr_debug("Number of Domains supportd <%ld>\n", ndomains);
1125 nlongs = BITS_TO_LONGS(ndomains);
1127 /* TBD: there might be 64K domains,
1128 * consider other allocation for future chip
1130 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1131 if (!iommu->domain_ids) {
1132 printk(KERN_ERR "Allocating domain id array failed\n");
1133 return -ENOMEM;
1135 iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1136 GFP_KERNEL);
1137 if (!iommu->domains) {
1138 printk(KERN_ERR "Allocating domain array failed\n");
1139 kfree(iommu->domain_ids);
1140 return -ENOMEM;
1143 spin_lock_init(&iommu->lock);
1146 * if Caching mode is set, then invalid translations are tagged
1147 * with domainid 0. Hence we need to pre-allocate it.
1149 if (cap_caching_mode(iommu->cap))
1150 set_bit(0, iommu->domain_ids);
1151 return 0;
1155 static void domain_exit(struct dmar_domain *domain);
1156 static void vm_domain_exit(struct dmar_domain *domain);
1158 void free_dmar_iommu(struct intel_iommu *iommu)
1160 struct dmar_domain *domain;
1161 int i;
1162 unsigned long flags;
1164 i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1165 for (; i < cap_ndoms(iommu->cap); ) {
1166 domain = iommu->domains[i];
1167 clear_bit(i, iommu->domain_ids);
1169 spin_lock_irqsave(&domain->iommu_lock, flags);
1170 if (--domain->iommu_count == 0) {
1171 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
1172 vm_domain_exit(domain);
1173 else
1174 domain_exit(domain);
1176 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1178 i = find_next_bit(iommu->domain_ids,
1179 cap_ndoms(iommu->cap), i+1);
1182 if (iommu->gcmd & DMA_GCMD_TE)
1183 iommu_disable_translation(iommu);
1185 if (iommu->irq) {
1186 set_irq_data(iommu->irq, NULL);
1187 /* This will mask the irq */
1188 free_irq(iommu->irq, iommu);
1189 destroy_irq(iommu->irq);
1192 kfree(iommu->domains);
1193 kfree(iommu->domain_ids);
1195 g_iommus[iommu->seq_id] = NULL;
1197 /* if all iommus are freed, free g_iommus */
1198 for (i = 0; i < g_num_of_iommus; i++) {
1199 if (g_iommus[i])
1200 break;
1203 if (i == g_num_of_iommus)
1204 kfree(g_iommus);
1206 /* free context mapping */
1207 free_context_table(iommu);
1210 static struct dmar_domain *alloc_domain(void)
1212 struct dmar_domain *domain;
1214 domain = alloc_domain_mem();
1215 if (!domain)
1216 return NULL;
1218 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
1219 domain->flags = 0;
1221 return domain;
1224 static int iommu_attach_domain(struct dmar_domain *domain,
1225 struct intel_iommu *iommu)
1227 int num;
1228 unsigned long ndomains;
1229 unsigned long flags;
1231 ndomains = cap_ndoms(iommu->cap);
1233 spin_lock_irqsave(&iommu->lock, flags);
1235 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1236 if (num >= ndomains) {
1237 spin_unlock_irqrestore(&iommu->lock, flags);
1238 printk(KERN_ERR "IOMMU: no free domain ids\n");
1239 return -ENOMEM;
1242 domain->id = num;
1243 set_bit(num, iommu->domain_ids);
1244 set_bit(iommu->seq_id, &domain->iommu_bmp);
1245 iommu->domains[num] = domain;
1246 spin_unlock_irqrestore(&iommu->lock, flags);
1248 return 0;
1251 static void iommu_detach_domain(struct dmar_domain *domain,
1252 struct intel_iommu *iommu)
1254 unsigned long flags;
1255 int num, ndomains;
1256 int found = 0;
1258 spin_lock_irqsave(&iommu->lock, flags);
1259 ndomains = cap_ndoms(iommu->cap);
1260 num = find_first_bit(iommu->domain_ids, ndomains);
1261 for (; num < ndomains; ) {
1262 if (iommu->domains[num] == domain) {
1263 found = 1;
1264 break;
1266 num = find_next_bit(iommu->domain_ids,
1267 cap_ndoms(iommu->cap), num+1);
1270 if (found) {
1271 clear_bit(num, iommu->domain_ids);
1272 clear_bit(iommu->seq_id, &domain->iommu_bmp);
1273 iommu->domains[num] = NULL;
1275 spin_unlock_irqrestore(&iommu->lock, flags);
1278 static struct iova_domain reserved_iova_list;
1279 static struct lock_class_key reserved_alloc_key;
1280 static struct lock_class_key reserved_rbtree_key;
1282 static void dmar_init_reserved_ranges(void)
1284 struct pci_dev *pdev = NULL;
1285 struct iova *iova;
1286 int i;
1287 u64 addr, size;
1289 init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1291 lockdep_set_class(&reserved_iova_list.iova_alloc_lock,
1292 &reserved_alloc_key);
1293 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1294 &reserved_rbtree_key);
1296 /* IOAPIC ranges shouldn't be accessed by DMA */
1297 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1298 IOVA_PFN(IOAPIC_RANGE_END));
1299 if (!iova)
1300 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1302 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1303 for_each_pci_dev(pdev) {
1304 struct resource *r;
1306 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1307 r = &pdev->resource[i];
1308 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1309 continue;
1310 addr = r->start;
1311 addr &= PHYSICAL_PAGE_MASK;
1312 size = r->end - addr;
1313 size = PAGE_ALIGN(size);
1314 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(addr),
1315 IOVA_PFN(size + addr) - 1);
1316 if (!iova)
1317 printk(KERN_ERR "Reserve iova failed\n");
1323 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1325 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1328 static inline int guestwidth_to_adjustwidth(int gaw)
1330 int agaw;
1331 int r = (gaw - 12) % 9;
1333 if (r == 0)
1334 agaw = gaw;
1335 else
1336 agaw = gaw + 9 - r;
1337 if (agaw > 64)
1338 agaw = 64;
1339 return agaw;
1342 static int domain_init(struct dmar_domain *domain, int guest_width)
1344 struct intel_iommu *iommu;
1345 int adjust_width, agaw;
1346 unsigned long sagaw;
1348 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1349 spin_lock_init(&domain->mapping_lock);
1350 spin_lock_init(&domain->iommu_lock);
1352 domain_reserve_special_ranges(domain);
1354 /* calculate AGAW */
1355 iommu = domain_get_iommu(domain);
1356 if (guest_width > cap_mgaw(iommu->cap))
1357 guest_width = cap_mgaw(iommu->cap);
1358 domain->gaw = guest_width;
1359 adjust_width = guestwidth_to_adjustwidth(guest_width);
1360 agaw = width_to_agaw(adjust_width);
1361 sagaw = cap_sagaw(iommu->cap);
1362 if (!test_bit(agaw, &sagaw)) {
1363 /* hardware doesn't support it, choose a bigger one */
1364 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1365 agaw = find_next_bit(&sagaw, 5, agaw);
1366 if (agaw >= 5)
1367 return -ENODEV;
1369 domain->agaw = agaw;
1370 INIT_LIST_HEAD(&domain->devices);
1372 if (ecap_coherent(iommu->ecap))
1373 domain->iommu_coherency = 1;
1374 else
1375 domain->iommu_coherency = 0;
1377 if (ecap_sc_support(iommu->ecap))
1378 domain->iommu_snooping = 1;
1379 else
1380 domain->iommu_snooping = 0;
1382 domain->iommu_count = 1;
1384 /* always allocate the top pgd */
1385 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1386 if (!domain->pgd)
1387 return -ENOMEM;
1388 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1389 return 0;
1392 static void domain_exit(struct dmar_domain *domain)
1394 struct dmar_drhd_unit *drhd;
1395 struct intel_iommu *iommu;
1396 u64 end;
1398 /* Domain 0 is reserved, so dont process it */
1399 if (!domain)
1400 return;
1402 domain_remove_dev_info(domain);
1403 /* destroy iovas */
1404 put_iova_domain(&domain->iovad);
1405 end = DOMAIN_MAX_ADDR(domain->gaw);
1406 end = end & (~PAGE_MASK);
1408 /* clear ptes */
1409 dma_pte_clear_range(domain, 0, end);
1411 /* free page tables */
1412 dma_pte_free_pagetable(domain, 0, end);
1414 for_each_active_iommu(iommu, drhd)
1415 if (test_bit(iommu->seq_id, &domain->iommu_bmp))
1416 iommu_detach_domain(domain, iommu);
1418 free_domain_mem(domain);
1421 static int domain_context_mapping_one(struct dmar_domain *domain, int segment,
1422 u8 bus, u8 devfn, int translation)
1424 struct context_entry *context;
1425 unsigned long flags;
1426 struct intel_iommu *iommu;
1427 struct dma_pte *pgd;
1428 unsigned long num;
1429 unsigned long ndomains;
1430 int id;
1431 int agaw;
1432 struct device_domain_info *info = NULL;
1434 pr_debug("Set context mapping for %02x:%02x.%d\n",
1435 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1437 BUG_ON(!domain->pgd);
1438 BUG_ON(translation != CONTEXT_TT_PASS_THROUGH &&
1439 translation != CONTEXT_TT_MULTI_LEVEL);
1441 iommu = device_to_iommu(segment, bus, devfn);
1442 if (!iommu)
1443 return -ENODEV;
1445 context = device_to_context_entry(iommu, bus, devfn);
1446 if (!context)
1447 return -ENOMEM;
1448 spin_lock_irqsave(&iommu->lock, flags);
1449 if (context_present(context)) {
1450 spin_unlock_irqrestore(&iommu->lock, flags);
1451 return 0;
1454 id = domain->id;
1455 pgd = domain->pgd;
1457 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
1458 domain->flags & DOMAIN_FLAG_STATIC_IDENTITY) {
1459 int found = 0;
1461 /* find an available domain id for this device in iommu */
1462 ndomains = cap_ndoms(iommu->cap);
1463 num = find_first_bit(iommu->domain_ids, ndomains);
1464 for (; num < ndomains; ) {
1465 if (iommu->domains[num] == domain) {
1466 id = num;
1467 found = 1;
1468 break;
1470 num = find_next_bit(iommu->domain_ids,
1471 cap_ndoms(iommu->cap), num+1);
1474 if (found == 0) {
1475 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1476 if (num >= ndomains) {
1477 spin_unlock_irqrestore(&iommu->lock, flags);
1478 printk(KERN_ERR "IOMMU: no free domain ids\n");
1479 return -EFAULT;
1482 set_bit(num, iommu->domain_ids);
1483 set_bit(iommu->seq_id, &domain->iommu_bmp);
1484 iommu->domains[num] = domain;
1485 id = num;
1488 /* Skip top levels of page tables for
1489 * iommu which has less agaw than default.
1491 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1492 pgd = phys_to_virt(dma_pte_addr(pgd));
1493 if (!dma_pte_present(pgd)) {
1494 spin_unlock_irqrestore(&iommu->lock, flags);
1495 return -ENOMEM;
1500 context_set_domain_id(context, id);
1502 if (translation != CONTEXT_TT_PASS_THROUGH) {
1503 info = iommu_support_dev_iotlb(domain, segment, bus, devfn);
1504 translation = info ? CONTEXT_TT_DEV_IOTLB :
1505 CONTEXT_TT_MULTI_LEVEL;
1508 * In pass through mode, AW must be programmed to indicate the largest
1509 * AGAW value supported by hardware. And ASR is ignored by hardware.
1511 if (unlikely(translation == CONTEXT_TT_PASS_THROUGH))
1512 context_set_address_width(context, iommu->msagaw);
1513 else {
1514 context_set_address_root(context, virt_to_phys(pgd));
1515 context_set_address_width(context, iommu->agaw);
1518 context_set_translation_type(context, translation);
1519 context_set_fault_enable(context);
1520 context_set_present(context);
1521 domain_flush_cache(domain, context, sizeof(*context));
1524 * It's a non-present to present mapping. If hardware doesn't cache
1525 * non-present entry we only need to flush the write-buffer. If the
1526 * _does_ cache non-present entries, then it does so in the special
1527 * domain #0, which we have to flush:
1529 if (cap_caching_mode(iommu->cap)) {
1530 iommu->flush.flush_context(iommu, 0,
1531 (((u16)bus) << 8) | devfn,
1532 DMA_CCMD_MASK_NOBIT,
1533 DMA_CCMD_DEVICE_INVL);
1534 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
1535 } else {
1536 iommu_flush_write_buffer(iommu);
1538 iommu_enable_dev_iotlb(info);
1539 spin_unlock_irqrestore(&iommu->lock, flags);
1541 spin_lock_irqsave(&domain->iommu_lock, flags);
1542 if (!test_and_set_bit(iommu->seq_id, &domain->iommu_bmp)) {
1543 domain->iommu_count++;
1544 domain_update_iommu_cap(domain);
1546 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1547 return 0;
1550 static int
1551 domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev,
1552 int translation)
1554 int ret;
1555 struct pci_dev *tmp, *parent;
1557 ret = domain_context_mapping_one(domain, pci_domain_nr(pdev->bus),
1558 pdev->bus->number, pdev->devfn,
1559 translation);
1560 if (ret)
1561 return ret;
1563 /* dependent device mapping */
1564 tmp = pci_find_upstream_pcie_bridge(pdev);
1565 if (!tmp)
1566 return 0;
1567 /* Secondary interface's bus number and devfn 0 */
1568 parent = pdev->bus->self;
1569 while (parent != tmp) {
1570 ret = domain_context_mapping_one(domain,
1571 pci_domain_nr(parent->bus),
1572 parent->bus->number,
1573 parent->devfn, translation);
1574 if (ret)
1575 return ret;
1576 parent = parent->bus->self;
1578 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1579 return domain_context_mapping_one(domain,
1580 pci_domain_nr(tmp->subordinate),
1581 tmp->subordinate->number, 0,
1582 translation);
1583 else /* this is a legacy PCI bridge */
1584 return domain_context_mapping_one(domain,
1585 pci_domain_nr(tmp->bus),
1586 tmp->bus->number,
1587 tmp->devfn,
1588 translation);
1591 static int domain_context_mapped(struct pci_dev *pdev)
1593 int ret;
1594 struct pci_dev *tmp, *parent;
1595 struct intel_iommu *iommu;
1597 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
1598 pdev->devfn);
1599 if (!iommu)
1600 return -ENODEV;
1602 ret = device_context_mapped(iommu, pdev->bus->number, pdev->devfn);
1603 if (!ret)
1604 return ret;
1605 /* dependent device mapping */
1606 tmp = pci_find_upstream_pcie_bridge(pdev);
1607 if (!tmp)
1608 return ret;
1609 /* Secondary interface's bus number and devfn 0 */
1610 parent = pdev->bus->self;
1611 while (parent != tmp) {
1612 ret = device_context_mapped(iommu, parent->bus->number,
1613 parent->devfn);
1614 if (!ret)
1615 return ret;
1616 parent = parent->bus->self;
1618 if (tmp->is_pcie)
1619 return device_context_mapped(iommu, tmp->subordinate->number,
1621 else
1622 return device_context_mapped(iommu, tmp->bus->number,
1623 tmp->devfn);
1626 static int
1627 domain_page_mapping(struct dmar_domain *domain, dma_addr_t iova,
1628 u64 hpa, size_t size, int prot)
1630 u64 start_pfn, end_pfn;
1631 struct dma_pte *pte;
1632 int index;
1633 int addr_width = agaw_to_width(domain->agaw);
1635 hpa &= (((u64)1) << addr_width) - 1;
1637 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1638 return -EINVAL;
1639 iova &= PAGE_MASK;
1640 start_pfn = ((u64)hpa) >> VTD_PAGE_SHIFT;
1641 end_pfn = (VTD_PAGE_ALIGN(((u64)hpa) + size)) >> VTD_PAGE_SHIFT;
1642 index = 0;
1643 while (start_pfn < end_pfn) {
1644 pte = addr_to_dma_pte(domain, iova + VTD_PAGE_SIZE * index);
1645 if (!pte)
1646 return -ENOMEM;
1647 /* We don't need lock here, nobody else
1648 * touches the iova range
1650 BUG_ON(dma_pte_addr(pte));
1651 dma_set_pte_addr(pte, start_pfn << VTD_PAGE_SHIFT);
1652 dma_set_pte_prot(pte, prot);
1653 if (prot & DMA_PTE_SNP)
1654 dma_set_pte_snp(pte);
1655 domain_flush_cache(domain, pte, sizeof(*pte));
1656 start_pfn++;
1657 index++;
1659 return 0;
1662 static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
1664 if (!iommu)
1665 return;
1667 clear_context_table(iommu, bus, devfn);
1668 iommu->flush.flush_context(iommu, 0, 0, 0,
1669 DMA_CCMD_GLOBAL_INVL);
1670 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
1673 static void domain_remove_dev_info(struct dmar_domain *domain)
1675 struct device_domain_info *info;
1676 unsigned long flags;
1677 struct intel_iommu *iommu;
1679 spin_lock_irqsave(&device_domain_lock, flags);
1680 while (!list_empty(&domain->devices)) {
1681 info = list_entry(domain->devices.next,
1682 struct device_domain_info, link);
1683 list_del(&info->link);
1684 list_del(&info->global);
1685 if (info->dev)
1686 info->dev->dev.archdata.iommu = NULL;
1687 spin_unlock_irqrestore(&device_domain_lock, flags);
1689 iommu_disable_dev_iotlb(info);
1690 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
1691 iommu_detach_dev(iommu, info->bus, info->devfn);
1692 free_devinfo_mem(info);
1694 spin_lock_irqsave(&device_domain_lock, flags);
1696 spin_unlock_irqrestore(&device_domain_lock, flags);
1700 * find_domain
1701 * Note: we use struct pci_dev->dev.archdata.iommu stores the info
1703 static struct dmar_domain *
1704 find_domain(struct pci_dev *pdev)
1706 struct device_domain_info *info;
1708 /* No lock here, assumes no domain exit in normal case */
1709 info = pdev->dev.archdata.iommu;
1710 if (info)
1711 return info->domain;
1712 return NULL;
1715 /* domain is initialized */
1716 static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1718 struct dmar_domain *domain, *found = NULL;
1719 struct intel_iommu *iommu;
1720 struct dmar_drhd_unit *drhd;
1721 struct device_domain_info *info, *tmp;
1722 struct pci_dev *dev_tmp;
1723 unsigned long flags;
1724 int bus = 0, devfn = 0;
1725 int segment;
1726 int ret;
1728 domain = find_domain(pdev);
1729 if (domain)
1730 return domain;
1732 segment = pci_domain_nr(pdev->bus);
1734 dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1735 if (dev_tmp) {
1736 if (dev_tmp->is_pcie) {
1737 bus = dev_tmp->subordinate->number;
1738 devfn = 0;
1739 } else {
1740 bus = dev_tmp->bus->number;
1741 devfn = dev_tmp->devfn;
1743 spin_lock_irqsave(&device_domain_lock, flags);
1744 list_for_each_entry(info, &device_domain_list, global) {
1745 if (info->segment == segment &&
1746 info->bus == bus && info->devfn == devfn) {
1747 found = info->domain;
1748 break;
1751 spin_unlock_irqrestore(&device_domain_lock, flags);
1752 /* pcie-pci bridge already has a domain, uses it */
1753 if (found) {
1754 domain = found;
1755 goto found_domain;
1759 domain = alloc_domain();
1760 if (!domain)
1761 goto error;
1763 /* Allocate new domain for the device */
1764 drhd = dmar_find_matched_drhd_unit(pdev);
1765 if (!drhd) {
1766 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1767 pci_name(pdev));
1768 return NULL;
1770 iommu = drhd->iommu;
1772 ret = iommu_attach_domain(domain, iommu);
1773 if (ret) {
1774 domain_exit(domain);
1775 goto error;
1778 if (domain_init(domain, gaw)) {
1779 domain_exit(domain);
1780 goto error;
1783 /* register pcie-to-pci device */
1784 if (dev_tmp) {
1785 info = alloc_devinfo_mem();
1786 if (!info) {
1787 domain_exit(domain);
1788 goto error;
1790 info->segment = segment;
1791 info->bus = bus;
1792 info->devfn = devfn;
1793 info->dev = NULL;
1794 info->domain = domain;
1795 /* This domain is shared by devices under p2p bridge */
1796 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
1798 /* pcie-to-pci bridge already has a domain, uses it */
1799 found = NULL;
1800 spin_lock_irqsave(&device_domain_lock, flags);
1801 list_for_each_entry(tmp, &device_domain_list, global) {
1802 if (tmp->segment == segment &&
1803 tmp->bus == bus && tmp->devfn == devfn) {
1804 found = tmp->domain;
1805 break;
1808 if (found) {
1809 free_devinfo_mem(info);
1810 domain_exit(domain);
1811 domain = found;
1812 } else {
1813 list_add(&info->link, &domain->devices);
1814 list_add(&info->global, &device_domain_list);
1816 spin_unlock_irqrestore(&device_domain_lock, flags);
1819 found_domain:
1820 info = alloc_devinfo_mem();
1821 if (!info)
1822 goto error;
1823 info->segment = segment;
1824 info->bus = pdev->bus->number;
1825 info->devfn = pdev->devfn;
1826 info->dev = pdev;
1827 info->domain = domain;
1828 spin_lock_irqsave(&device_domain_lock, flags);
1829 /* somebody is fast */
1830 found = find_domain(pdev);
1831 if (found != NULL) {
1832 spin_unlock_irqrestore(&device_domain_lock, flags);
1833 if (found != domain) {
1834 domain_exit(domain);
1835 domain = found;
1837 free_devinfo_mem(info);
1838 return domain;
1840 list_add(&info->link, &domain->devices);
1841 list_add(&info->global, &device_domain_list);
1842 pdev->dev.archdata.iommu = info;
1843 spin_unlock_irqrestore(&device_domain_lock, flags);
1844 return domain;
1845 error:
1846 /* recheck it here, maybe others set it */
1847 return find_domain(pdev);
1850 static int iommu_identity_mapping;
1852 static int iommu_domain_identity_map(struct dmar_domain *domain,
1853 unsigned long long start,
1854 unsigned long long end)
1856 unsigned long size;
1857 unsigned long long base;
1859 /* The address might not be aligned */
1860 base = start & PAGE_MASK;
1861 size = end - base;
1862 size = PAGE_ALIGN(size);
1863 if (!reserve_iova(&domain->iovad, IOVA_PFN(base),
1864 IOVA_PFN(base + size) - 1)) {
1865 printk(KERN_ERR "IOMMU: reserve iova failed\n");
1866 return -ENOMEM;
1869 pr_debug("Mapping reserved region %lx@%llx for domain %d\n",
1870 size, base, domain->id);
1872 * RMRR range might have overlap with physical memory range,
1873 * clear it first
1875 dma_pte_clear_range(domain, base, base + size);
1877 return domain_page_mapping(domain, base, base, size,
1878 DMA_PTE_READ|DMA_PTE_WRITE);
1881 static int iommu_prepare_identity_map(struct pci_dev *pdev,
1882 unsigned long long start,
1883 unsigned long long end)
1885 struct dmar_domain *domain;
1886 int ret;
1888 printk(KERN_INFO
1889 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1890 pci_name(pdev), start, end);
1892 if (iommu_identity_mapping)
1893 domain = si_domain;
1894 else
1895 /* page table init */
1896 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1897 if (!domain)
1898 return -ENOMEM;
1900 ret = iommu_domain_identity_map(domain, start, end);
1901 if (ret)
1902 goto error;
1904 /* context entry init */
1905 ret = domain_context_mapping(domain, pdev, CONTEXT_TT_MULTI_LEVEL);
1906 if (ret)
1907 goto error;
1909 return 0;
1911 error:
1912 domain_exit(domain);
1913 return ret;
1916 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1917 struct pci_dev *pdev)
1919 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1920 return 0;
1921 return iommu_prepare_identity_map(pdev, rmrr->base_address,
1922 rmrr->end_address + 1);
1925 struct iommu_prepare_data {
1926 struct pci_dev *pdev;
1927 int ret;
1930 static int __init iommu_prepare_work_fn(unsigned long start_pfn,
1931 unsigned long end_pfn, void *datax)
1933 struct iommu_prepare_data *data;
1935 data = (struct iommu_prepare_data *)datax;
1937 data->ret = iommu_prepare_identity_map(data->pdev,
1938 start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
1939 return data->ret;
1943 static int __init iommu_prepare_with_active_regions(struct pci_dev *pdev)
1945 int nid;
1946 struct iommu_prepare_data data;
1948 data.pdev = pdev;
1949 data.ret = 0;
1951 for_each_online_node(nid) {
1952 work_with_active_regions(nid, iommu_prepare_work_fn, &data);
1953 if (data.ret)
1954 return data.ret;
1956 return data.ret;
1959 #ifdef CONFIG_DMAR_GFX_WA
1960 static void __init iommu_prepare_gfx_mapping(void)
1962 struct pci_dev *pdev = NULL;
1963 int ret;
1965 for_each_pci_dev(pdev) {
1966 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO ||
1967 !IS_GFX_DEVICE(pdev))
1968 continue;
1969 printk(KERN_INFO "IOMMU: gfx device %s 1-1 mapping\n",
1970 pci_name(pdev));
1971 ret = iommu_prepare_with_active_regions(pdev);
1972 if (ret)
1973 printk(KERN_ERR "IOMMU: mapping reserved region failed\n");
1976 #else /* !CONFIG_DMAR_GFX_WA */
1977 static inline void iommu_prepare_gfx_mapping(void)
1979 return;
1981 #endif
1983 #ifdef CONFIG_DMAR_FLOPPY_WA
1984 static inline void iommu_prepare_isa(void)
1986 struct pci_dev *pdev;
1987 int ret;
1989 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1990 if (!pdev)
1991 return;
1993 printk(KERN_INFO "IOMMU: Prepare 0-16M unity mapping for LPC\n");
1994 ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1996 if (ret)
1997 printk(KERN_ERR "IOMMU: Failed to create 0-64M identity map, "
1998 "floppy might not work\n");
2001 #else
2002 static inline void iommu_prepare_isa(void)
2004 return;
2006 #endif /* !CONFIG_DMAR_FLPY_WA */
2008 /* Initialize each context entry as pass through.*/
2009 static int __init init_context_pass_through(void)
2011 struct pci_dev *pdev = NULL;
2012 struct dmar_domain *domain;
2013 int ret;
2015 for_each_pci_dev(pdev) {
2016 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2017 ret = domain_context_mapping(domain, pdev,
2018 CONTEXT_TT_PASS_THROUGH);
2019 if (ret)
2020 return ret;
2022 return 0;
2025 static int md_domain_init(struct dmar_domain *domain, int guest_width);
2026 static int si_domain_init(void)
2028 struct dmar_drhd_unit *drhd;
2029 struct intel_iommu *iommu;
2030 int ret = 0;
2032 si_domain = alloc_domain();
2033 if (!si_domain)
2034 return -EFAULT;
2037 for_each_active_iommu(iommu, drhd) {
2038 ret = iommu_attach_domain(si_domain, iommu);
2039 if (ret) {
2040 domain_exit(si_domain);
2041 return -EFAULT;
2045 if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2046 domain_exit(si_domain);
2047 return -EFAULT;
2050 si_domain->flags = DOMAIN_FLAG_STATIC_IDENTITY;
2052 return 0;
2055 static void domain_remove_one_dev_info(struct dmar_domain *domain,
2056 struct pci_dev *pdev);
2057 static int identity_mapping(struct pci_dev *pdev)
2059 struct device_domain_info *info;
2061 if (likely(!iommu_identity_mapping))
2062 return 0;
2065 list_for_each_entry(info, &si_domain->devices, link)
2066 if (info->dev == pdev)
2067 return 1;
2068 return 0;
2071 static int domain_add_dev_info(struct dmar_domain *domain,
2072 struct pci_dev *pdev)
2074 struct device_domain_info *info;
2075 unsigned long flags;
2077 info = alloc_devinfo_mem();
2078 if (!info)
2079 return -ENOMEM;
2081 info->segment = pci_domain_nr(pdev->bus);
2082 info->bus = pdev->bus->number;
2083 info->devfn = pdev->devfn;
2084 info->dev = pdev;
2085 info->domain = domain;
2087 spin_lock_irqsave(&device_domain_lock, flags);
2088 list_add(&info->link, &domain->devices);
2089 list_add(&info->global, &device_domain_list);
2090 pdev->dev.archdata.iommu = info;
2091 spin_unlock_irqrestore(&device_domain_lock, flags);
2093 return 0;
2096 static int iommu_prepare_static_identity_mapping(void)
2098 struct pci_dev *pdev = NULL;
2099 int ret;
2101 ret = si_domain_init();
2102 if (ret)
2103 return -EFAULT;
2105 printk(KERN_INFO "IOMMU: Setting identity map:\n");
2106 for_each_pci_dev(pdev) {
2107 ret = iommu_prepare_with_active_regions(pdev);
2108 if (ret) {
2109 printk(KERN_INFO "1:1 mapping to one domain failed.\n");
2110 return -EFAULT;
2112 ret = domain_add_dev_info(si_domain, pdev);
2113 if (ret)
2114 return ret;
2117 return 0;
2120 int __init init_dmars(void)
2122 struct dmar_drhd_unit *drhd;
2123 struct dmar_rmrr_unit *rmrr;
2124 struct pci_dev *pdev;
2125 struct intel_iommu *iommu;
2126 int i, ret;
2127 int pass_through = 1;
2130 * In case pass through can not be enabled, iommu tries to use identity
2131 * mapping.
2133 if (iommu_pass_through)
2134 iommu_identity_mapping = 1;
2137 * for each drhd
2138 * allocate root
2139 * initialize and program root entry to not present
2140 * endfor
2142 for_each_drhd_unit(drhd) {
2143 g_num_of_iommus++;
2145 * lock not needed as this is only incremented in the single
2146 * threaded kernel __init code path all other access are read
2147 * only
2151 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
2152 GFP_KERNEL);
2153 if (!g_iommus) {
2154 printk(KERN_ERR "Allocating global iommu array failed\n");
2155 ret = -ENOMEM;
2156 goto error;
2159 deferred_flush = kzalloc(g_num_of_iommus *
2160 sizeof(struct deferred_flush_tables), GFP_KERNEL);
2161 if (!deferred_flush) {
2162 kfree(g_iommus);
2163 ret = -ENOMEM;
2164 goto error;
2167 for_each_drhd_unit(drhd) {
2168 if (drhd->ignored)
2169 continue;
2171 iommu = drhd->iommu;
2172 g_iommus[iommu->seq_id] = iommu;
2174 ret = iommu_init_domains(iommu);
2175 if (ret)
2176 goto error;
2179 * TBD:
2180 * we could share the same root & context tables
2181 * amoung all IOMMU's. Need to Split it later.
2183 ret = iommu_alloc_root_entry(iommu);
2184 if (ret) {
2185 printk(KERN_ERR "IOMMU: allocate root entry failed\n");
2186 goto error;
2188 if (!ecap_pass_through(iommu->ecap))
2189 pass_through = 0;
2191 if (iommu_pass_through)
2192 if (!pass_through) {
2193 printk(KERN_INFO
2194 "Pass Through is not supported by hardware.\n");
2195 iommu_pass_through = 0;
2199 * Start from the sane iommu hardware state.
2201 for_each_drhd_unit(drhd) {
2202 if (drhd->ignored)
2203 continue;
2205 iommu = drhd->iommu;
2208 * If the queued invalidation is already initialized by us
2209 * (for example, while enabling interrupt-remapping) then
2210 * we got the things already rolling from a sane state.
2212 if (iommu->qi)
2213 continue;
2216 * Clear any previous faults.
2218 dmar_fault(-1, iommu);
2220 * Disable queued invalidation if supported and already enabled
2221 * before OS handover.
2223 dmar_disable_qi(iommu);
2226 for_each_drhd_unit(drhd) {
2227 if (drhd->ignored)
2228 continue;
2230 iommu = drhd->iommu;
2232 if (dmar_enable_qi(iommu)) {
2234 * Queued Invalidate not enabled, use Register Based
2235 * Invalidate
2237 iommu->flush.flush_context = __iommu_flush_context;
2238 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2239 printk(KERN_INFO "IOMMU 0x%Lx: using Register based "
2240 "invalidation\n",
2241 (unsigned long long)drhd->reg_base_addr);
2242 } else {
2243 iommu->flush.flush_context = qi_flush_context;
2244 iommu->flush.flush_iotlb = qi_flush_iotlb;
2245 printk(KERN_INFO "IOMMU 0x%Lx: using Queued "
2246 "invalidation\n",
2247 (unsigned long long)drhd->reg_base_addr);
2252 * If pass through is set and enabled, context entries of all pci
2253 * devices are intialized by pass through translation type.
2255 if (iommu_pass_through) {
2256 ret = init_context_pass_through();
2257 if (ret) {
2258 printk(KERN_ERR "IOMMU: Pass through init failed.\n");
2259 iommu_pass_through = 0;
2264 * If pass through is not set or not enabled, setup context entries for
2265 * identity mappings for rmrr, gfx, and isa and may fall back to static
2266 * identity mapping if iommu_identity_mapping is set.
2268 if (!iommu_pass_through) {
2269 if (iommu_identity_mapping)
2270 iommu_prepare_static_identity_mapping();
2272 * For each rmrr
2273 * for each dev attached to rmrr
2274 * do
2275 * locate drhd for dev, alloc domain for dev
2276 * allocate free domain
2277 * allocate page table entries for rmrr
2278 * if context not allocated for bus
2279 * allocate and init context
2280 * set present in root table for this bus
2281 * init context with domain, translation etc
2282 * endfor
2283 * endfor
2285 printk(KERN_INFO "IOMMU: Setting RMRR:\n");
2286 for_each_rmrr_units(rmrr) {
2287 for (i = 0; i < rmrr->devices_cnt; i++) {
2288 pdev = rmrr->devices[i];
2290 * some BIOS lists non-exist devices in DMAR
2291 * table.
2293 if (!pdev)
2294 continue;
2295 ret = iommu_prepare_rmrr_dev(rmrr, pdev);
2296 if (ret)
2297 printk(KERN_ERR
2298 "IOMMU: mapping reserved region failed\n");
2302 iommu_prepare_gfx_mapping();
2304 iommu_prepare_isa();
2308 * for each drhd
2309 * enable fault log
2310 * global invalidate context cache
2311 * global invalidate iotlb
2312 * enable translation
2314 for_each_drhd_unit(drhd) {
2315 if (drhd->ignored)
2316 continue;
2317 iommu = drhd->iommu;
2319 iommu_flush_write_buffer(iommu);
2321 ret = dmar_set_interrupt(iommu);
2322 if (ret)
2323 goto error;
2325 iommu_set_root_entry(iommu);
2327 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
2328 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2329 iommu_disable_protect_mem_regions(iommu);
2331 ret = iommu_enable_translation(iommu);
2332 if (ret)
2333 goto error;
2336 return 0;
2337 error:
2338 for_each_drhd_unit(drhd) {
2339 if (drhd->ignored)
2340 continue;
2341 iommu = drhd->iommu;
2342 free_iommu(iommu);
2344 kfree(g_iommus);
2345 return ret;
2348 static inline u64 aligned_size(u64 host_addr, size_t size)
2350 u64 addr;
2351 addr = (host_addr & (~PAGE_MASK)) + size;
2352 return PAGE_ALIGN(addr);
2355 struct iova *
2356 iommu_alloc_iova(struct dmar_domain *domain, size_t size, u64 end)
2358 struct iova *piova;
2360 /* Make sure it's in range */
2361 end = min_t(u64, DOMAIN_MAX_ADDR(domain->gaw), end);
2362 if (!size || (IOVA_START_ADDR + size > end))
2363 return NULL;
2365 piova = alloc_iova(&domain->iovad,
2366 size >> PAGE_SHIFT, IOVA_PFN(end), 1);
2367 return piova;
2370 static struct iova *
2371 __intel_alloc_iova(struct device *dev, struct dmar_domain *domain,
2372 size_t size, u64 dma_mask)
2374 struct pci_dev *pdev = to_pci_dev(dev);
2375 struct iova *iova = NULL;
2377 if (dma_mask <= DMA_BIT_MASK(32) || dmar_forcedac)
2378 iova = iommu_alloc_iova(domain, size, dma_mask);
2379 else {
2381 * First try to allocate an io virtual address in
2382 * DMA_BIT_MASK(32) and if that fails then try allocating
2383 * from higher range
2385 iova = iommu_alloc_iova(domain, size, DMA_BIT_MASK(32));
2386 if (!iova)
2387 iova = iommu_alloc_iova(domain, size, dma_mask);
2390 if (!iova) {
2391 printk(KERN_ERR"Allocating iova for %s failed", pci_name(pdev));
2392 return NULL;
2395 return iova;
2398 static struct dmar_domain *
2399 get_valid_domain_for_dev(struct pci_dev *pdev)
2401 struct dmar_domain *domain;
2402 int ret;
2404 domain = get_domain_for_dev(pdev,
2405 DEFAULT_DOMAIN_ADDRESS_WIDTH);
2406 if (!domain) {
2407 printk(KERN_ERR
2408 "Allocating domain for %s failed", pci_name(pdev));
2409 return NULL;
2412 /* make sure context mapping is ok */
2413 if (unlikely(!domain_context_mapped(pdev))) {
2414 ret = domain_context_mapping(domain, pdev,
2415 CONTEXT_TT_MULTI_LEVEL);
2416 if (ret) {
2417 printk(KERN_ERR
2418 "Domain context map for %s failed",
2419 pci_name(pdev));
2420 return NULL;
2424 return domain;
2427 static int iommu_dummy(struct pci_dev *pdev)
2429 return pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
2432 /* Check if the pdev needs to go through non-identity map and unmap process.*/
2433 static int iommu_no_mapping(struct pci_dev *pdev)
2435 int found;
2437 if (!iommu_identity_mapping)
2438 return iommu_dummy(pdev);
2440 found = identity_mapping(pdev);
2441 if (found) {
2442 if (pdev->dma_mask > DMA_BIT_MASK(32))
2443 return 1;
2444 else {
2446 * 32 bit DMA is removed from si_domain and fall back
2447 * to non-identity mapping.
2449 domain_remove_one_dev_info(si_domain, pdev);
2450 printk(KERN_INFO "32bit %s uses non-identity mapping\n",
2451 pci_name(pdev));
2452 return 0;
2454 } else {
2456 * In case of a detached 64 bit DMA device from vm, the device
2457 * is put into si_domain for identity mapping.
2459 if (pdev->dma_mask > DMA_BIT_MASK(32)) {
2460 int ret;
2461 ret = domain_add_dev_info(si_domain, pdev);
2462 if (!ret) {
2463 printk(KERN_INFO "64bit %s uses identity mapping\n",
2464 pci_name(pdev));
2465 return 1;
2470 return iommu_dummy(pdev);
2473 static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
2474 size_t size, int dir, u64 dma_mask)
2476 struct pci_dev *pdev = to_pci_dev(hwdev);
2477 struct dmar_domain *domain;
2478 phys_addr_t start_paddr;
2479 struct iova *iova;
2480 int prot = 0;
2481 int ret;
2482 struct intel_iommu *iommu;
2484 BUG_ON(dir == DMA_NONE);
2486 if (iommu_no_mapping(pdev))
2487 return paddr;
2489 domain = get_valid_domain_for_dev(pdev);
2490 if (!domain)
2491 return 0;
2493 iommu = domain_get_iommu(domain);
2494 size = aligned_size((u64)paddr, size);
2496 iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2497 if (!iova)
2498 goto error;
2500 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
2503 * Check if DMAR supports zero-length reads on write only
2504 * mappings..
2506 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2507 !cap_zlr(iommu->cap))
2508 prot |= DMA_PTE_READ;
2509 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2510 prot |= DMA_PTE_WRITE;
2512 * paddr - (paddr + size) might be partial page, we should map the whole
2513 * page. Note: if two part of one page are separately mapped, we
2514 * might have two guest_addr mapping to the same host paddr, but this
2515 * is not a big problem
2517 ret = domain_page_mapping(domain, start_paddr,
2518 ((u64)paddr) & PHYSICAL_PAGE_MASK,
2519 size, prot);
2520 if (ret)
2521 goto error;
2523 /* it's a non-present to present mapping. Only flush if caching mode */
2524 if (cap_caching_mode(iommu->cap))
2525 iommu_flush_iotlb_psi(iommu, 0, start_paddr,
2526 size >> VTD_PAGE_SHIFT);
2527 else
2528 iommu_flush_write_buffer(iommu);
2530 return start_paddr + ((u64)paddr & (~PAGE_MASK));
2532 error:
2533 if (iova)
2534 __free_iova(&domain->iovad, iova);
2535 printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
2536 pci_name(pdev), size, (unsigned long long)paddr, dir);
2537 return 0;
2540 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
2541 unsigned long offset, size_t size,
2542 enum dma_data_direction dir,
2543 struct dma_attrs *attrs)
2545 return __intel_map_single(dev, page_to_phys(page) + offset, size,
2546 dir, to_pci_dev(dev)->dma_mask);
2549 static void flush_unmaps(void)
2551 int i, j;
2553 timer_on = 0;
2555 /* just flush them all */
2556 for (i = 0; i < g_num_of_iommus; i++) {
2557 struct intel_iommu *iommu = g_iommus[i];
2558 if (!iommu)
2559 continue;
2561 if (!deferred_flush[i].next)
2562 continue;
2564 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2565 DMA_TLB_GLOBAL_FLUSH);
2566 for (j = 0; j < deferred_flush[i].next; j++) {
2567 unsigned long mask;
2568 struct iova *iova = deferred_flush[i].iova[j];
2570 mask = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT;
2571 mask = ilog2(mask >> VTD_PAGE_SHIFT);
2572 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
2573 iova->pfn_lo << PAGE_SHIFT, mask);
2574 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
2576 deferred_flush[i].next = 0;
2579 list_size = 0;
2582 static void flush_unmaps_timeout(unsigned long data)
2584 unsigned long flags;
2586 spin_lock_irqsave(&async_umap_flush_lock, flags);
2587 flush_unmaps();
2588 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2591 static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2593 unsigned long flags;
2594 int next, iommu_id;
2595 struct intel_iommu *iommu;
2597 spin_lock_irqsave(&async_umap_flush_lock, flags);
2598 if (list_size == HIGH_WATER_MARK)
2599 flush_unmaps();
2601 iommu = domain_get_iommu(dom);
2602 iommu_id = iommu->seq_id;
2604 next = deferred_flush[iommu_id].next;
2605 deferred_flush[iommu_id].domain[next] = dom;
2606 deferred_flush[iommu_id].iova[next] = iova;
2607 deferred_flush[iommu_id].next++;
2609 if (!timer_on) {
2610 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2611 timer_on = 1;
2613 list_size++;
2614 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2617 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
2618 size_t size, enum dma_data_direction dir,
2619 struct dma_attrs *attrs)
2621 struct pci_dev *pdev = to_pci_dev(dev);
2622 struct dmar_domain *domain;
2623 unsigned long start_addr;
2624 struct iova *iova;
2625 struct intel_iommu *iommu;
2627 if (iommu_no_mapping(pdev))
2628 return;
2630 domain = find_domain(pdev);
2631 BUG_ON(!domain);
2633 iommu = domain_get_iommu(domain);
2635 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
2636 if (!iova)
2637 return;
2639 start_addr = iova->pfn_lo << PAGE_SHIFT;
2640 size = aligned_size((u64)dev_addr, size);
2642 pr_debug("Device %s unmapping: %zx@%llx\n",
2643 pci_name(pdev), size, (unsigned long long)start_addr);
2645 /* clear the whole page */
2646 dma_pte_clear_range(domain, start_addr, start_addr + size);
2647 /* free page tables */
2648 dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2649 if (intel_iommu_strict) {
2650 iommu_flush_iotlb_psi(iommu, domain->id, start_addr,
2651 size >> VTD_PAGE_SHIFT);
2652 /* free iova */
2653 __free_iova(&domain->iovad, iova);
2654 } else {
2655 add_unmap(domain, iova);
2657 * queue up the release of the unmap to save the 1/6th of the
2658 * cpu used up by the iotlb flush operation...
2663 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr, size_t size,
2664 int dir)
2666 intel_unmap_page(dev, dev_addr, size, dir, NULL);
2669 static void *intel_alloc_coherent(struct device *hwdev, size_t size,
2670 dma_addr_t *dma_handle, gfp_t flags)
2672 void *vaddr;
2673 int order;
2675 size = PAGE_ALIGN(size);
2676 order = get_order(size);
2677 flags &= ~(GFP_DMA | GFP_DMA32);
2679 vaddr = (void *)__get_free_pages(flags, order);
2680 if (!vaddr)
2681 return NULL;
2682 memset(vaddr, 0, size);
2684 *dma_handle = __intel_map_single(hwdev, virt_to_bus(vaddr), size,
2685 DMA_BIDIRECTIONAL,
2686 hwdev->coherent_dma_mask);
2687 if (*dma_handle)
2688 return vaddr;
2689 free_pages((unsigned long)vaddr, order);
2690 return NULL;
2693 static void intel_free_coherent(struct device *hwdev, size_t size, void *vaddr,
2694 dma_addr_t dma_handle)
2696 int order;
2698 size = PAGE_ALIGN(size);
2699 order = get_order(size);
2701 intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2702 free_pages((unsigned long)vaddr, order);
2705 static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2706 int nelems, enum dma_data_direction dir,
2707 struct dma_attrs *attrs)
2709 int i;
2710 struct pci_dev *pdev = to_pci_dev(hwdev);
2711 struct dmar_domain *domain;
2712 unsigned long start_addr;
2713 struct iova *iova;
2714 size_t size = 0;
2715 phys_addr_t addr;
2716 struct scatterlist *sg;
2717 struct intel_iommu *iommu;
2719 if (iommu_no_mapping(pdev))
2720 return;
2722 domain = find_domain(pdev);
2723 BUG_ON(!domain);
2725 iommu = domain_get_iommu(domain);
2727 iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
2728 if (!iova)
2729 return;
2730 for_each_sg(sglist, sg, nelems, i) {
2731 addr = page_to_phys(sg_page(sg)) + sg->offset;
2732 size += aligned_size((u64)addr, sg->length);
2735 start_addr = iova->pfn_lo << PAGE_SHIFT;
2737 /* clear the whole page */
2738 dma_pte_clear_range(domain, start_addr, start_addr + size);
2739 /* free page tables */
2740 dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2742 iommu_flush_iotlb_psi(iommu, domain->id, start_addr,
2743 size >> VTD_PAGE_SHIFT);
2745 /* free iova */
2746 __free_iova(&domain->iovad, iova);
2749 static int intel_nontranslate_map_sg(struct device *hddev,
2750 struct scatterlist *sglist, int nelems, int dir)
2752 int i;
2753 struct scatterlist *sg;
2755 for_each_sg(sglist, sg, nelems, i) {
2756 BUG_ON(!sg_page(sg));
2757 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
2758 sg->dma_length = sg->length;
2760 return nelems;
2763 static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist, int nelems,
2764 enum dma_data_direction dir, struct dma_attrs *attrs)
2766 phys_addr_t addr;
2767 int i;
2768 struct pci_dev *pdev = to_pci_dev(hwdev);
2769 struct dmar_domain *domain;
2770 size_t size = 0;
2771 int prot = 0;
2772 size_t offset = 0;
2773 struct iova *iova = NULL;
2774 int ret;
2775 struct scatterlist *sg;
2776 unsigned long start_addr;
2777 struct intel_iommu *iommu;
2779 BUG_ON(dir == DMA_NONE);
2780 if (iommu_no_mapping(pdev))
2781 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
2783 domain = get_valid_domain_for_dev(pdev);
2784 if (!domain)
2785 return 0;
2787 iommu = domain_get_iommu(domain);
2789 for_each_sg(sglist, sg, nelems, i) {
2790 addr = page_to_phys(sg_page(sg)) + sg->offset;
2791 size += aligned_size((u64)addr, sg->length);
2794 iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2795 if (!iova) {
2796 sglist->dma_length = 0;
2797 return 0;
2801 * Check if DMAR supports zero-length reads on write only
2802 * mappings..
2804 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2805 !cap_zlr(iommu->cap))
2806 prot |= DMA_PTE_READ;
2807 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2808 prot |= DMA_PTE_WRITE;
2810 start_addr = iova->pfn_lo << PAGE_SHIFT;
2811 offset = 0;
2812 for_each_sg(sglist, sg, nelems, i) {
2813 addr = page_to_phys(sg_page(sg)) + sg->offset;
2814 size = aligned_size((u64)addr, sg->length);
2815 ret = domain_page_mapping(domain, start_addr + offset,
2816 ((u64)addr) & PHYSICAL_PAGE_MASK,
2817 size, prot);
2818 if (ret) {
2819 /* clear the page */
2820 dma_pte_clear_range(domain, start_addr,
2821 start_addr + offset);
2822 /* free page tables */
2823 dma_pte_free_pagetable(domain, start_addr,
2824 start_addr + offset);
2825 /* free iova */
2826 __free_iova(&domain->iovad, iova);
2827 return 0;
2829 sg->dma_address = start_addr + offset +
2830 ((u64)addr & (~PAGE_MASK));
2831 sg->dma_length = sg->length;
2832 offset += size;
2835 /* it's a non-present to present mapping. Only flush if caching mode */
2836 if (cap_caching_mode(iommu->cap))
2837 iommu_flush_iotlb_psi(iommu, 0, start_addr,
2838 offset >> VTD_PAGE_SHIFT);
2839 else
2840 iommu_flush_write_buffer(iommu);
2842 return nelems;
2845 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
2847 return !dma_addr;
2850 struct dma_map_ops intel_dma_ops = {
2851 .alloc_coherent = intel_alloc_coherent,
2852 .free_coherent = intel_free_coherent,
2853 .map_sg = intel_map_sg,
2854 .unmap_sg = intel_unmap_sg,
2855 .map_page = intel_map_page,
2856 .unmap_page = intel_unmap_page,
2857 .mapping_error = intel_mapping_error,
2860 static inline int iommu_domain_cache_init(void)
2862 int ret = 0;
2864 iommu_domain_cache = kmem_cache_create("iommu_domain",
2865 sizeof(struct dmar_domain),
2867 SLAB_HWCACHE_ALIGN,
2869 NULL);
2870 if (!iommu_domain_cache) {
2871 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2872 ret = -ENOMEM;
2875 return ret;
2878 static inline int iommu_devinfo_cache_init(void)
2880 int ret = 0;
2882 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2883 sizeof(struct device_domain_info),
2885 SLAB_HWCACHE_ALIGN,
2886 NULL);
2887 if (!iommu_devinfo_cache) {
2888 printk(KERN_ERR "Couldn't create devinfo cache\n");
2889 ret = -ENOMEM;
2892 return ret;
2895 static inline int iommu_iova_cache_init(void)
2897 int ret = 0;
2899 iommu_iova_cache = kmem_cache_create("iommu_iova",
2900 sizeof(struct iova),
2902 SLAB_HWCACHE_ALIGN,
2903 NULL);
2904 if (!iommu_iova_cache) {
2905 printk(KERN_ERR "Couldn't create iova cache\n");
2906 ret = -ENOMEM;
2909 return ret;
2912 static int __init iommu_init_mempool(void)
2914 int ret;
2915 ret = iommu_iova_cache_init();
2916 if (ret)
2917 return ret;
2919 ret = iommu_domain_cache_init();
2920 if (ret)
2921 goto domain_error;
2923 ret = iommu_devinfo_cache_init();
2924 if (!ret)
2925 return ret;
2927 kmem_cache_destroy(iommu_domain_cache);
2928 domain_error:
2929 kmem_cache_destroy(iommu_iova_cache);
2931 return -ENOMEM;
2934 static void __init iommu_exit_mempool(void)
2936 kmem_cache_destroy(iommu_devinfo_cache);
2937 kmem_cache_destroy(iommu_domain_cache);
2938 kmem_cache_destroy(iommu_iova_cache);
2942 static void __init init_no_remapping_devices(void)
2944 struct dmar_drhd_unit *drhd;
2946 for_each_drhd_unit(drhd) {
2947 if (!drhd->include_all) {
2948 int i;
2949 for (i = 0; i < drhd->devices_cnt; i++)
2950 if (drhd->devices[i] != NULL)
2951 break;
2952 /* ignore DMAR unit if no pci devices exist */
2953 if (i == drhd->devices_cnt)
2954 drhd->ignored = 1;
2958 if (dmar_map_gfx)
2959 return;
2961 for_each_drhd_unit(drhd) {
2962 int i;
2963 if (drhd->ignored || drhd->include_all)
2964 continue;
2966 for (i = 0; i < drhd->devices_cnt; i++)
2967 if (drhd->devices[i] &&
2968 !IS_GFX_DEVICE(drhd->devices[i]))
2969 break;
2971 if (i < drhd->devices_cnt)
2972 continue;
2974 /* bypass IOMMU if it is just for gfx devices */
2975 drhd->ignored = 1;
2976 for (i = 0; i < drhd->devices_cnt; i++) {
2977 if (!drhd->devices[i])
2978 continue;
2979 drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
2984 #ifdef CONFIG_SUSPEND
2985 static int init_iommu_hw(void)
2987 struct dmar_drhd_unit *drhd;
2988 struct intel_iommu *iommu = NULL;
2990 for_each_active_iommu(iommu, drhd)
2991 if (iommu->qi)
2992 dmar_reenable_qi(iommu);
2994 for_each_active_iommu(iommu, drhd) {
2995 iommu_flush_write_buffer(iommu);
2997 iommu_set_root_entry(iommu);
2999 iommu->flush.flush_context(iommu, 0, 0, 0,
3000 DMA_CCMD_GLOBAL_INVL);
3001 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3002 DMA_TLB_GLOBAL_FLUSH);
3003 iommu_disable_protect_mem_regions(iommu);
3004 iommu_enable_translation(iommu);
3007 return 0;
3010 static void iommu_flush_all(void)
3012 struct dmar_drhd_unit *drhd;
3013 struct intel_iommu *iommu;
3015 for_each_active_iommu(iommu, drhd) {
3016 iommu->flush.flush_context(iommu, 0, 0, 0,
3017 DMA_CCMD_GLOBAL_INVL);
3018 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3019 DMA_TLB_GLOBAL_FLUSH);
3023 static int iommu_suspend(struct sys_device *dev, pm_message_t state)
3025 struct dmar_drhd_unit *drhd;
3026 struct intel_iommu *iommu = NULL;
3027 unsigned long flag;
3029 for_each_active_iommu(iommu, drhd) {
3030 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3031 GFP_ATOMIC);
3032 if (!iommu->iommu_state)
3033 goto nomem;
3036 iommu_flush_all();
3038 for_each_active_iommu(iommu, drhd) {
3039 iommu_disable_translation(iommu);
3041 spin_lock_irqsave(&iommu->register_lock, flag);
3043 iommu->iommu_state[SR_DMAR_FECTL_REG] =
3044 readl(iommu->reg + DMAR_FECTL_REG);
3045 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
3046 readl(iommu->reg + DMAR_FEDATA_REG);
3047 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
3048 readl(iommu->reg + DMAR_FEADDR_REG);
3049 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
3050 readl(iommu->reg + DMAR_FEUADDR_REG);
3052 spin_unlock_irqrestore(&iommu->register_lock, flag);
3054 return 0;
3056 nomem:
3057 for_each_active_iommu(iommu, drhd)
3058 kfree(iommu->iommu_state);
3060 return -ENOMEM;
3063 static int iommu_resume(struct sys_device *dev)
3065 struct dmar_drhd_unit *drhd;
3066 struct intel_iommu *iommu = NULL;
3067 unsigned long flag;
3069 if (init_iommu_hw()) {
3070 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
3071 return -EIO;
3074 for_each_active_iommu(iommu, drhd) {
3076 spin_lock_irqsave(&iommu->register_lock, flag);
3078 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
3079 iommu->reg + DMAR_FECTL_REG);
3080 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
3081 iommu->reg + DMAR_FEDATA_REG);
3082 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
3083 iommu->reg + DMAR_FEADDR_REG);
3084 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
3085 iommu->reg + DMAR_FEUADDR_REG);
3087 spin_unlock_irqrestore(&iommu->register_lock, flag);
3090 for_each_active_iommu(iommu, drhd)
3091 kfree(iommu->iommu_state);
3093 return 0;
3096 static struct sysdev_class iommu_sysclass = {
3097 .name = "iommu",
3098 .resume = iommu_resume,
3099 .suspend = iommu_suspend,
3102 static struct sys_device device_iommu = {
3103 .cls = &iommu_sysclass,
3106 static int __init init_iommu_sysfs(void)
3108 int error;
3110 error = sysdev_class_register(&iommu_sysclass);
3111 if (error)
3112 return error;
3114 error = sysdev_register(&device_iommu);
3115 if (error)
3116 sysdev_class_unregister(&iommu_sysclass);
3118 return error;
3121 #else
3122 static int __init init_iommu_sysfs(void)
3124 return 0;
3126 #endif /* CONFIG_PM */
3128 int __init intel_iommu_init(void)
3130 int ret = 0;
3132 if (dmar_table_init())
3133 return -ENODEV;
3135 if (dmar_dev_scope_init())
3136 return -ENODEV;
3139 * Check the need for DMA-remapping initialization now.
3140 * Above initialization will also be used by Interrupt-remapping.
3142 if (no_iommu || (swiotlb && !iommu_pass_through) || dmar_disabled)
3143 return -ENODEV;
3145 iommu_init_mempool();
3146 dmar_init_reserved_ranges();
3148 init_no_remapping_devices();
3150 ret = init_dmars();
3151 if (ret) {
3152 printk(KERN_ERR "IOMMU: dmar init failed\n");
3153 put_iova_domain(&reserved_iova_list);
3154 iommu_exit_mempool();
3155 return ret;
3157 printk(KERN_INFO
3158 "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
3160 init_timer(&unmap_timer);
3161 force_iommu = 1;
3163 if (!iommu_pass_through) {
3164 printk(KERN_INFO
3165 "Multi-level page-table translation for DMAR.\n");
3166 dma_ops = &intel_dma_ops;
3167 } else
3168 printk(KERN_INFO
3169 "DMAR: Pass through translation for DMAR.\n");
3171 init_iommu_sysfs();
3173 register_iommu(&intel_iommu_ops);
3175 return 0;
3178 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
3179 struct pci_dev *pdev)
3181 struct pci_dev *tmp, *parent;
3183 if (!iommu || !pdev)
3184 return;
3186 /* dependent device detach */
3187 tmp = pci_find_upstream_pcie_bridge(pdev);
3188 /* Secondary interface's bus number and devfn 0 */
3189 if (tmp) {
3190 parent = pdev->bus->self;
3191 while (parent != tmp) {
3192 iommu_detach_dev(iommu, parent->bus->number,
3193 parent->devfn);
3194 parent = parent->bus->self;
3196 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
3197 iommu_detach_dev(iommu,
3198 tmp->subordinate->number, 0);
3199 else /* this is a legacy PCI bridge */
3200 iommu_detach_dev(iommu, tmp->bus->number,
3201 tmp->devfn);
3205 static void domain_remove_one_dev_info(struct dmar_domain *domain,
3206 struct pci_dev *pdev)
3208 struct device_domain_info *info;
3209 struct intel_iommu *iommu;
3210 unsigned long flags;
3211 int found = 0;
3212 struct list_head *entry, *tmp;
3214 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3215 pdev->devfn);
3216 if (!iommu)
3217 return;
3219 spin_lock_irqsave(&device_domain_lock, flags);
3220 list_for_each_safe(entry, tmp, &domain->devices) {
3221 info = list_entry(entry, struct device_domain_info, link);
3222 /* No need to compare PCI domain; it has to be the same */
3223 if (info->bus == pdev->bus->number &&
3224 info->devfn == pdev->devfn) {
3225 list_del(&info->link);
3226 list_del(&info->global);
3227 if (info->dev)
3228 info->dev->dev.archdata.iommu = NULL;
3229 spin_unlock_irqrestore(&device_domain_lock, flags);
3231 iommu_disable_dev_iotlb(info);
3232 iommu_detach_dev(iommu, info->bus, info->devfn);
3233 iommu_detach_dependent_devices(iommu, pdev);
3234 free_devinfo_mem(info);
3236 spin_lock_irqsave(&device_domain_lock, flags);
3238 if (found)
3239 break;
3240 else
3241 continue;
3244 /* if there is no other devices under the same iommu
3245 * owned by this domain, clear this iommu in iommu_bmp
3246 * update iommu count and coherency
3248 if (iommu == device_to_iommu(info->segment, info->bus,
3249 info->devfn))
3250 found = 1;
3253 if (found == 0) {
3254 unsigned long tmp_flags;
3255 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
3256 clear_bit(iommu->seq_id, &domain->iommu_bmp);
3257 domain->iommu_count--;
3258 domain_update_iommu_cap(domain);
3259 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
3262 spin_unlock_irqrestore(&device_domain_lock, flags);
3265 static void vm_domain_remove_all_dev_info(struct dmar_domain *domain)
3267 struct device_domain_info *info;
3268 struct intel_iommu *iommu;
3269 unsigned long flags1, flags2;
3271 spin_lock_irqsave(&device_domain_lock, flags1);
3272 while (!list_empty(&domain->devices)) {
3273 info = list_entry(domain->devices.next,
3274 struct device_domain_info, link);
3275 list_del(&info->link);
3276 list_del(&info->global);
3277 if (info->dev)
3278 info->dev->dev.archdata.iommu = NULL;
3280 spin_unlock_irqrestore(&device_domain_lock, flags1);
3282 iommu_disable_dev_iotlb(info);
3283 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
3284 iommu_detach_dev(iommu, info->bus, info->devfn);
3285 iommu_detach_dependent_devices(iommu, info->dev);
3287 /* clear this iommu in iommu_bmp, update iommu count
3288 * and capabilities
3290 spin_lock_irqsave(&domain->iommu_lock, flags2);
3291 if (test_and_clear_bit(iommu->seq_id,
3292 &domain->iommu_bmp)) {
3293 domain->iommu_count--;
3294 domain_update_iommu_cap(domain);
3296 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
3298 free_devinfo_mem(info);
3299 spin_lock_irqsave(&device_domain_lock, flags1);
3301 spin_unlock_irqrestore(&device_domain_lock, flags1);
3304 /* domain id for virtual machine, it won't be set in context */
3305 static unsigned long vm_domid;
3307 static int vm_domain_min_agaw(struct dmar_domain *domain)
3309 int i;
3310 int min_agaw = domain->agaw;
3312 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
3313 for (; i < g_num_of_iommus; ) {
3314 if (min_agaw > g_iommus[i]->agaw)
3315 min_agaw = g_iommus[i]->agaw;
3317 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
3320 return min_agaw;
3323 static struct dmar_domain *iommu_alloc_vm_domain(void)
3325 struct dmar_domain *domain;
3327 domain = alloc_domain_mem();
3328 if (!domain)
3329 return NULL;
3331 domain->id = vm_domid++;
3332 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
3333 domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
3335 return domain;
3338 static int md_domain_init(struct dmar_domain *domain, int guest_width)
3340 int adjust_width;
3342 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
3343 spin_lock_init(&domain->mapping_lock);
3344 spin_lock_init(&domain->iommu_lock);
3346 domain_reserve_special_ranges(domain);
3348 /* calculate AGAW */
3349 domain->gaw = guest_width;
3350 adjust_width = guestwidth_to_adjustwidth(guest_width);
3351 domain->agaw = width_to_agaw(adjust_width);
3353 INIT_LIST_HEAD(&domain->devices);
3355 domain->iommu_count = 0;
3356 domain->iommu_coherency = 0;
3357 domain->max_addr = 0;
3359 /* always allocate the top pgd */
3360 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
3361 if (!domain->pgd)
3362 return -ENOMEM;
3363 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
3364 return 0;
3367 static void iommu_free_vm_domain(struct dmar_domain *domain)
3369 unsigned long flags;
3370 struct dmar_drhd_unit *drhd;
3371 struct intel_iommu *iommu;
3372 unsigned long i;
3373 unsigned long ndomains;
3375 for_each_drhd_unit(drhd) {
3376 if (drhd->ignored)
3377 continue;
3378 iommu = drhd->iommu;
3380 ndomains = cap_ndoms(iommu->cap);
3381 i = find_first_bit(iommu->domain_ids, ndomains);
3382 for (; i < ndomains; ) {
3383 if (iommu->domains[i] == domain) {
3384 spin_lock_irqsave(&iommu->lock, flags);
3385 clear_bit(i, iommu->domain_ids);
3386 iommu->domains[i] = NULL;
3387 spin_unlock_irqrestore(&iommu->lock, flags);
3388 break;
3390 i = find_next_bit(iommu->domain_ids, ndomains, i+1);
3395 static void vm_domain_exit(struct dmar_domain *domain)
3397 u64 end;
3399 /* Domain 0 is reserved, so dont process it */
3400 if (!domain)
3401 return;
3403 vm_domain_remove_all_dev_info(domain);
3404 /* destroy iovas */
3405 put_iova_domain(&domain->iovad);
3406 end = DOMAIN_MAX_ADDR(domain->gaw);
3407 end = end & (~VTD_PAGE_MASK);
3409 /* clear ptes */
3410 dma_pte_clear_range(domain, 0, end);
3412 /* free page tables */
3413 dma_pte_free_pagetable(domain, 0, end);
3415 iommu_free_vm_domain(domain);
3416 free_domain_mem(domain);
3419 static int intel_iommu_domain_init(struct iommu_domain *domain)
3421 struct dmar_domain *dmar_domain;
3423 dmar_domain = iommu_alloc_vm_domain();
3424 if (!dmar_domain) {
3425 printk(KERN_ERR
3426 "intel_iommu_domain_init: dmar_domain == NULL\n");
3427 return -ENOMEM;
3429 if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
3430 printk(KERN_ERR
3431 "intel_iommu_domain_init() failed\n");
3432 vm_domain_exit(dmar_domain);
3433 return -ENOMEM;
3435 domain->priv = dmar_domain;
3437 return 0;
3440 static void intel_iommu_domain_destroy(struct iommu_domain *domain)
3442 struct dmar_domain *dmar_domain = domain->priv;
3444 domain->priv = NULL;
3445 vm_domain_exit(dmar_domain);
3448 static int intel_iommu_attach_device(struct iommu_domain *domain,
3449 struct device *dev)
3451 struct dmar_domain *dmar_domain = domain->priv;
3452 struct pci_dev *pdev = to_pci_dev(dev);
3453 struct intel_iommu *iommu;
3454 int addr_width;
3455 u64 end;
3456 int ret;
3458 /* normally pdev is not mapped */
3459 if (unlikely(domain_context_mapped(pdev))) {
3460 struct dmar_domain *old_domain;
3462 old_domain = find_domain(pdev);
3463 if (old_domain) {
3464 if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
3465 dmar_domain->flags & DOMAIN_FLAG_STATIC_IDENTITY)
3466 domain_remove_one_dev_info(old_domain, pdev);
3467 else
3468 domain_remove_dev_info(old_domain);
3472 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3473 pdev->devfn);
3474 if (!iommu)
3475 return -ENODEV;
3477 /* check if this iommu agaw is sufficient for max mapped address */
3478 addr_width = agaw_to_width(iommu->agaw);
3479 end = DOMAIN_MAX_ADDR(addr_width);
3480 end = end & VTD_PAGE_MASK;
3481 if (end < dmar_domain->max_addr) {
3482 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3483 "sufficient for the mapped address (%llx)\n",
3484 __func__, iommu->agaw, dmar_domain->max_addr);
3485 return -EFAULT;
3488 ret = domain_add_dev_info(dmar_domain, pdev);
3489 if (ret)
3490 return ret;
3492 ret = domain_context_mapping(dmar_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
3493 return ret;
3496 static void intel_iommu_detach_device(struct iommu_domain *domain,
3497 struct device *dev)
3499 struct dmar_domain *dmar_domain = domain->priv;
3500 struct pci_dev *pdev = to_pci_dev(dev);
3502 domain_remove_one_dev_info(dmar_domain, pdev);
3505 static int intel_iommu_map_range(struct iommu_domain *domain,
3506 unsigned long iova, phys_addr_t hpa,
3507 size_t size, int iommu_prot)
3509 struct dmar_domain *dmar_domain = domain->priv;
3510 u64 max_addr;
3511 int addr_width;
3512 int prot = 0;
3513 int ret;
3515 if (iommu_prot & IOMMU_READ)
3516 prot |= DMA_PTE_READ;
3517 if (iommu_prot & IOMMU_WRITE)
3518 prot |= DMA_PTE_WRITE;
3519 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
3520 prot |= DMA_PTE_SNP;
3522 max_addr = (iova & VTD_PAGE_MASK) + VTD_PAGE_ALIGN(size);
3523 if (dmar_domain->max_addr < max_addr) {
3524 int min_agaw;
3525 u64 end;
3527 /* check if minimum agaw is sufficient for mapped address */
3528 min_agaw = vm_domain_min_agaw(dmar_domain);
3529 addr_width = agaw_to_width(min_agaw);
3530 end = DOMAIN_MAX_ADDR(addr_width);
3531 end = end & VTD_PAGE_MASK;
3532 if (end < max_addr) {
3533 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3534 "sufficient for the mapped address (%llx)\n",
3535 __func__, min_agaw, max_addr);
3536 return -EFAULT;
3538 dmar_domain->max_addr = max_addr;
3541 ret = domain_page_mapping(dmar_domain, iova, hpa, size, prot);
3542 return ret;
3545 static void intel_iommu_unmap_range(struct iommu_domain *domain,
3546 unsigned long iova, size_t size)
3548 struct dmar_domain *dmar_domain = domain->priv;
3549 dma_addr_t base;
3551 /* The address might not be aligned */
3552 base = iova & VTD_PAGE_MASK;
3553 size = VTD_PAGE_ALIGN(size);
3554 dma_pte_clear_range(dmar_domain, base, base + size);
3556 if (dmar_domain->max_addr == base + size)
3557 dmar_domain->max_addr = base;
3560 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
3561 unsigned long iova)
3563 struct dmar_domain *dmar_domain = domain->priv;
3564 struct dma_pte *pte;
3565 u64 phys = 0;
3567 pte = addr_to_dma_pte(dmar_domain, iova);
3568 if (pte)
3569 phys = dma_pte_addr(pte);
3571 return phys;
3574 static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
3575 unsigned long cap)
3577 struct dmar_domain *dmar_domain = domain->priv;
3579 if (cap == IOMMU_CAP_CACHE_COHERENCY)
3580 return dmar_domain->iommu_snooping;
3582 return 0;
3585 static struct iommu_ops intel_iommu_ops = {
3586 .domain_init = intel_iommu_domain_init,
3587 .domain_destroy = intel_iommu_domain_destroy,
3588 .attach_dev = intel_iommu_attach_device,
3589 .detach_dev = intel_iommu_detach_device,
3590 .map = intel_iommu_map_range,
3591 .unmap = intel_iommu_unmap_range,
3592 .iova_to_phys = intel_iommu_iova_to_phys,
3593 .domain_has_cap = intel_iommu_domain_has_cap,
3596 static void __devinit quirk_iommu_rwbf(struct pci_dev *dev)
3599 * Mobile 4 Series Chipset neglects to set RWBF capability,
3600 * but needs it:
3602 printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
3603 rwbf_quirk = 1;
3606 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);