intel-iommu: Introduce domain_sg_mapping() to speed up intel_map_sg()
[linux-2.6.git] / drivers / pci / intel-iommu.c
1 /*
2  * Copyright (c) 2006, Intel Corporation.
3  *
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.
7  *
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.
12  *
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.
16  *
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>
22  */
23
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"
43
44 #define ROOT_SIZE               VTD_PAGE_SIZE
45 #define CONTEXT_SIZE            VTD_PAGE_SIZE
46
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)
49
50 #define IOAPIC_RANGE_START      (0xfee00000)
51 #define IOAPIC_RANGE_END        (0xfeefffff)
52 #define IOVA_START_ADDR         (0x1000)
53
54 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
55
56 #define MAX_AGAW_WIDTH 64
57
58 #define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
59 #define DOMAIN_MAX_PFN(gaw)  ((((u64)1) << (gaw-VTD_PAGE_SHIFT)) - 1)
60
61 #define IOVA_PFN(addr)          ((addr) >> PAGE_SHIFT)
62 #define DMA_32BIT_PFN           IOVA_PFN(DMA_BIT_MASK(32))
63 #define DMA_64BIT_PFN           IOVA_PFN(DMA_BIT_MASK(64))
64
65
66 /* VT-d pages must always be _smaller_ than MM pages. Otherwise things
67    are never going to work. */
68 static inline unsigned long dma_to_mm_pfn(unsigned long dma_pfn)
69 {
70         return dma_pfn >> (PAGE_SHIFT - VTD_PAGE_SHIFT);
71 }
72
73 static inline unsigned long mm_to_dma_pfn(unsigned long mm_pfn)
74 {
75         return mm_pfn << (PAGE_SHIFT - VTD_PAGE_SHIFT);
76 }
77 static inline unsigned long page_to_dma_pfn(struct page *pg)
78 {
79         return mm_to_dma_pfn(page_to_pfn(pg));
80 }
81 static inline unsigned long virt_to_dma_pfn(void *p)
82 {
83         return page_to_dma_pfn(virt_to_page(p));
84 }
85
86 /* global iommu list, set NULL for ignored DMAR units */
87 static struct intel_iommu **g_iommus;
88
89 static int rwbf_quirk;
90
91 /*
92  * 0: Present
93  * 1-11: Reserved
94  * 12-63: Context Ptr (12 - (haw-1))
95  * 64-127: Reserved
96  */
97 struct root_entry {
98         u64     val;
99         u64     rsvd1;
100 };
101 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
102 static inline bool root_present(struct root_entry *root)
103 {
104         return (root->val & 1);
105 }
106 static inline void set_root_present(struct root_entry *root)
107 {
108         root->val |= 1;
109 }
110 static inline void set_root_value(struct root_entry *root, unsigned long value)
111 {
112         root->val |= value & VTD_PAGE_MASK;
113 }
114
115 static inline struct context_entry *
116 get_context_addr_from_root(struct root_entry *root)
117 {
118         return (struct context_entry *)
119                 (root_present(root)?phys_to_virt(
120                 root->val & VTD_PAGE_MASK) :
121                 NULL);
122 }
123
124 /*
125  * low 64 bits:
126  * 0: present
127  * 1: fault processing disable
128  * 2-3: translation type
129  * 12-63: address space root
130  * high 64 bits:
131  * 0-2: address width
132  * 3-6: aval
133  * 8-23: domain id
134  */
135 struct context_entry {
136         u64 lo;
137         u64 hi;
138 };
139
140 static inline bool context_present(struct context_entry *context)
141 {
142         return (context->lo & 1);
143 }
144 static inline void context_set_present(struct context_entry *context)
145 {
146         context->lo |= 1;
147 }
148
149 static inline void context_set_fault_enable(struct context_entry *context)
150 {
151         context->lo &= (((u64)-1) << 2) | 1;
152 }
153
154 static inline void context_set_translation_type(struct context_entry *context,
155                                                 unsigned long value)
156 {
157         context->lo &= (((u64)-1) << 4) | 3;
158         context->lo |= (value & 3) << 2;
159 }
160
161 static inline void context_set_address_root(struct context_entry *context,
162                                             unsigned long value)
163 {
164         context->lo |= value & VTD_PAGE_MASK;
165 }
166
167 static inline void context_set_address_width(struct context_entry *context,
168                                              unsigned long value)
169 {
170         context->hi |= value & 7;
171 }
172
173 static inline void context_set_domain_id(struct context_entry *context,
174                                          unsigned long value)
175 {
176         context->hi |= (value & ((1 << 16) - 1)) << 8;
177 }
178
179 static inline void context_clear_entry(struct context_entry *context)
180 {
181         context->lo = 0;
182         context->hi = 0;
183 }
184
185 /*
186  * 0: readable
187  * 1: writable
188  * 2-6: reserved
189  * 7: super page
190  * 8-10: available
191  * 11: snoop behavior
192  * 12-63: Host physcial address
193  */
194 struct dma_pte {
195         u64 val;
196 };
197
198 static inline void dma_clear_pte(struct dma_pte *pte)
199 {
200         pte->val = 0;
201 }
202
203 static inline void dma_set_pte_readable(struct dma_pte *pte)
204 {
205         pte->val |= DMA_PTE_READ;
206 }
207
208 static inline void dma_set_pte_writable(struct dma_pte *pte)
209 {
210         pte->val |= DMA_PTE_WRITE;
211 }
212
213 static inline void dma_set_pte_snp(struct dma_pte *pte)
214 {
215         pte->val |= DMA_PTE_SNP;
216 }
217
218 static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)
219 {
220         pte->val = (pte->val & ~3) | (prot & 3);
221 }
222
223 static inline u64 dma_pte_addr(struct dma_pte *pte)
224 {
225         return (pte->val & VTD_PAGE_MASK);
226 }
227
228 static inline void dma_set_pte_pfn(struct dma_pte *pte, unsigned long pfn)
229 {
230         pte->val |= (uint64_t)pfn << VTD_PAGE_SHIFT;
231 }
232
233 static inline bool dma_pte_present(struct dma_pte *pte)
234 {
235         return (pte->val & 3) != 0;
236 }
237
238 /*
239  * This domain is a statically identity mapping domain.
240  *      1. This domain creats a static 1:1 mapping to all usable memory.
241  *      2. It maps to each iommu if successful.
242  *      3. Each iommu mapps to this domain if successful.
243  */
244 struct dmar_domain *si_domain;
245
246 /* devices under the same p2p bridge are owned in one domain */
247 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
248
249 /* domain represents a virtual machine, more than one devices
250  * across iommus may be owned in one domain, e.g. kvm guest.
251  */
252 #define DOMAIN_FLAG_VIRTUAL_MACHINE     (1 << 1)
253
254 /* si_domain contains mulitple devices */
255 #define DOMAIN_FLAG_STATIC_IDENTITY     (1 << 2)
256
257 struct dmar_domain {
258         int     id;                     /* domain id */
259         unsigned long iommu_bmp;        /* bitmap of iommus this domain uses*/
260
261         struct list_head devices;       /* all devices' list */
262         struct iova_domain iovad;       /* iova's that belong to this domain */
263
264         struct dma_pte  *pgd;           /* virtual address */
265         spinlock_t      mapping_lock;   /* page table lock */
266         int             gaw;            /* max guest address width */
267
268         /* adjusted guest address width, 0 is level 2 30-bit */
269         int             agaw;
270
271         int             flags;          /* flags to find out type of domain */
272
273         int             iommu_coherency;/* indicate coherency of iommu access */
274         int             iommu_snooping; /* indicate snooping control feature*/
275         int             iommu_count;    /* reference count of iommu */
276         spinlock_t      iommu_lock;     /* protect iommu set in domain */
277         u64             max_addr;       /* maximum mapped address */
278 };
279
280 /* PCI domain-device relationship */
281 struct device_domain_info {
282         struct list_head link;  /* link to domain siblings */
283         struct list_head global; /* link to global list */
284         int segment;            /* PCI domain */
285         u8 bus;                 /* PCI bus number */
286         u8 devfn;               /* PCI devfn number */
287         struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
288         struct intel_iommu *iommu; /* IOMMU used by this device */
289         struct dmar_domain *domain; /* pointer to domain */
290 };
291
292 static void flush_unmaps_timeout(unsigned long data);
293
294 DEFINE_TIMER(unmap_timer,  flush_unmaps_timeout, 0, 0);
295
296 #define HIGH_WATER_MARK 250
297 struct deferred_flush_tables {
298         int next;
299         struct iova *iova[HIGH_WATER_MARK];
300         struct dmar_domain *domain[HIGH_WATER_MARK];
301 };
302
303 static struct deferred_flush_tables *deferred_flush;
304
305 /* bitmap for indexing intel_iommus */
306 static int g_num_of_iommus;
307
308 static DEFINE_SPINLOCK(async_umap_flush_lock);
309 static LIST_HEAD(unmaps_to_do);
310
311 static int timer_on;
312 static long list_size;
313
314 static void domain_remove_dev_info(struct dmar_domain *domain);
315
316 #ifdef CONFIG_DMAR_DEFAULT_ON
317 int dmar_disabled = 0;
318 #else
319 int dmar_disabled = 1;
320 #endif /*CONFIG_DMAR_DEFAULT_ON*/
321
322 static int __initdata dmar_map_gfx = 1;
323 static int dmar_forcedac;
324 static int intel_iommu_strict;
325
326 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
327 static DEFINE_SPINLOCK(device_domain_lock);
328 static LIST_HEAD(device_domain_list);
329
330 static struct iommu_ops intel_iommu_ops;
331
332 static int __init intel_iommu_setup(char *str)
333 {
334         if (!str)
335                 return -EINVAL;
336         while (*str) {
337                 if (!strncmp(str, "on", 2)) {
338                         dmar_disabled = 0;
339                         printk(KERN_INFO "Intel-IOMMU: enabled\n");
340                 } else if (!strncmp(str, "off", 3)) {
341                         dmar_disabled = 1;
342                         printk(KERN_INFO "Intel-IOMMU: disabled\n");
343                 } else if (!strncmp(str, "igfx_off", 8)) {
344                         dmar_map_gfx = 0;
345                         printk(KERN_INFO
346                                 "Intel-IOMMU: disable GFX device mapping\n");
347                 } else if (!strncmp(str, "forcedac", 8)) {
348                         printk(KERN_INFO
349                                 "Intel-IOMMU: Forcing DAC for PCI devices\n");
350                         dmar_forcedac = 1;
351                 } else if (!strncmp(str, "strict", 6)) {
352                         printk(KERN_INFO
353                                 "Intel-IOMMU: disable batched IOTLB flush\n");
354                         intel_iommu_strict = 1;
355                 }
356
357                 str += strcspn(str, ",");
358                 while (*str == ',')
359                         str++;
360         }
361         return 0;
362 }
363 __setup("intel_iommu=", intel_iommu_setup);
364
365 static struct kmem_cache *iommu_domain_cache;
366 static struct kmem_cache *iommu_devinfo_cache;
367 static struct kmem_cache *iommu_iova_cache;
368
369 static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
370 {
371         unsigned int flags;
372         void *vaddr;
373
374         /* trying to avoid low memory issues */
375         flags = current->flags & PF_MEMALLOC;
376         current->flags |= PF_MEMALLOC;
377         vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
378         current->flags &= (~PF_MEMALLOC | flags);
379         return vaddr;
380 }
381
382
383 static inline void *alloc_pgtable_page(void)
384 {
385         unsigned int flags;
386         void *vaddr;
387
388         /* trying to avoid low memory issues */
389         flags = current->flags & PF_MEMALLOC;
390         current->flags |= PF_MEMALLOC;
391         vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
392         current->flags &= (~PF_MEMALLOC | flags);
393         return vaddr;
394 }
395
396 static inline void free_pgtable_page(void *vaddr)
397 {
398         free_page((unsigned long)vaddr);
399 }
400
401 static inline void *alloc_domain_mem(void)
402 {
403         return iommu_kmem_cache_alloc(iommu_domain_cache);
404 }
405
406 static void free_domain_mem(void *vaddr)
407 {
408         kmem_cache_free(iommu_domain_cache, vaddr);
409 }
410
411 static inline void * alloc_devinfo_mem(void)
412 {
413         return iommu_kmem_cache_alloc(iommu_devinfo_cache);
414 }
415
416 static inline void free_devinfo_mem(void *vaddr)
417 {
418         kmem_cache_free(iommu_devinfo_cache, vaddr);
419 }
420
421 struct iova *alloc_iova_mem(void)
422 {
423         return iommu_kmem_cache_alloc(iommu_iova_cache);
424 }
425
426 void free_iova_mem(struct iova *iova)
427 {
428         kmem_cache_free(iommu_iova_cache, iova);
429 }
430
431
432 static inline int width_to_agaw(int width);
433
434 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
435 {
436         unsigned long sagaw;
437         int agaw = -1;
438
439         sagaw = cap_sagaw(iommu->cap);
440         for (agaw = width_to_agaw(max_gaw);
441              agaw >= 0; agaw--) {
442                 if (test_bit(agaw, &sagaw))
443                         break;
444         }
445
446         return agaw;
447 }
448
449 /*
450  * Calculate max SAGAW for each iommu.
451  */
452 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
453 {
454         return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
455 }
456
457 /*
458  * calculate agaw for each iommu.
459  * "SAGAW" may be different across iommus, use a default agaw, and
460  * get a supported less agaw for iommus that don't support the default agaw.
461  */
462 int iommu_calculate_agaw(struct intel_iommu *iommu)
463 {
464         return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
465 }
466
467 /* This functionin only returns single iommu in a domain */
468 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
469 {
470         int iommu_id;
471
472         /* si_domain and vm domain should not get here. */
473         BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
474         BUG_ON(domain->flags & DOMAIN_FLAG_STATIC_IDENTITY);
475
476         iommu_id = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
477         if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
478                 return NULL;
479
480         return g_iommus[iommu_id];
481 }
482
483 static void domain_update_iommu_coherency(struct dmar_domain *domain)
484 {
485         int i;
486
487         domain->iommu_coherency = 1;
488
489         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
490         for (; i < g_num_of_iommus; ) {
491                 if (!ecap_coherent(g_iommus[i]->ecap)) {
492                         domain->iommu_coherency = 0;
493                         break;
494                 }
495                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
496         }
497 }
498
499 static void domain_update_iommu_snooping(struct dmar_domain *domain)
500 {
501         int i;
502
503         domain->iommu_snooping = 1;
504
505         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
506         for (; i < g_num_of_iommus; ) {
507                 if (!ecap_sc_support(g_iommus[i]->ecap)) {
508                         domain->iommu_snooping = 0;
509                         break;
510                 }
511                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
512         }
513 }
514
515 /* Some capabilities may be different across iommus */
516 static void domain_update_iommu_cap(struct dmar_domain *domain)
517 {
518         domain_update_iommu_coherency(domain);
519         domain_update_iommu_snooping(domain);
520 }
521
522 static struct intel_iommu *device_to_iommu(int segment, u8 bus, u8 devfn)
523 {
524         struct dmar_drhd_unit *drhd = NULL;
525         int i;
526
527         for_each_drhd_unit(drhd) {
528                 if (drhd->ignored)
529                         continue;
530                 if (segment != drhd->segment)
531                         continue;
532
533                 for (i = 0; i < drhd->devices_cnt; i++) {
534                         if (drhd->devices[i] &&
535                             drhd->devices[i]->bus->number == bus &&
536                             drhd->devices[i]->devfn == devfn)
537                                 return drhd->iommu;
538                         if (drhd->devices[i] &&
539                             drhd->devices[i]->subordinate &&
540                             drhd->devices[i]->subordinate->number <= bus &&
541                             drhd->devices[i]->subordinate->subordinate >= bus)
542                                 return drhd->iommu;
543                 }
544
545                 if (drhd->include_all)
546                         return drhd->iommu;
547         }
548
549         return NULL;
550 }
551
552 static void domain_flush_cache(struct dmar_domain *domain,
553                                void *addr, int size)
554 {
555         if (!domain->iommu_coherency)
556                 clflush_cache_range(addr, size);
557 }
558
559 /* Gets context entry for a given bus and devfn */
560 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
561                 u8 bus, u8 devfn)
562 {
563         struct root_entry *root;
564         struct context_entry *context;
565         unsigned long phy_addr;
566         unsigned long flags;
567
568         spin_lock_irqsave(&iommu->lock, flags);
569         root = &iommu->root_entry[bus];
570         context = get_context_addr_from_root(root);
571         if (!context) {
572                 context = (struct context_entry *)alloc_pgtable_page();
573                 if (!context) {
574                         spin_unlock_irqrestore(&iommu->lock, flags);
575                         return NULL;
576                 }
577                 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
578                 phy_addr = virt_to_phys((void *)context);
579                 set_root_value(root, phy_addr);
580                 set_root_present(root);
581                 __iommu_flush_cache(iommu, root, sizeof(*root));
582         }
583         spin_unlock_irqrestore(&iommu->lock, flags);
584         return &context[devfn];
585 }
586
587 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
588 {
589         struct root_entry *root;
590         struct context_entry *context;
591         int ret;
592         unsigned long flags;
593
594         spin_lock_irqsave(&iommu->lock, flags);
595         root = &iommu->root_entry[bus];
596         context = get_context_addr_from_root(root);
597         if (!context) {
598                 ret = 0;
599                 goto out;
600         }
601         ret = context_present(&context[devfn]);
602 out:
603         spin_unlock_irqrestore(&iommu->lock, flags);
604         return ret;
605 }
606
607 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
608 {
609         struct root_entry *root;
610         struct context_entry *context;
611         unsigned long flags;
612
613         spin_lock_irqsave(&iommu->lock, flags);
614         root = &iommu->root_entry[bus];
615         context = get_context_addr_from_root(root);
616         if (context) {
617                 context_clear_entry(&context[devfn]);
618                 __iommu_flush_cache(iommu, &context[devfn], \
619                         sizeof(*context));
620         }
621         spin_unlock_irqrestore(&iommu->lock, flags);
622 }
623
624 static void free_context_table(struct intel_iommu *iommu)
625 {
626         struct root_entry *root;
627         int i;
628         unsigned long flags;
629         struct context_entry *context;
630
631         spin_lock_irqsave(&iommu->lock, flags);
632         if (!iommu->root_entry) {
633                 goto out;
634         }
635         for (i = 0; i < ROOT_ENTRY_NR; i++) {
636                 root = &iommu->root_entry[i];
637                 context = get_context_addr_from_root(root);
638                 if (context)
639                         free_pgtable_page(context);
640         }
641         free_pgtable_page(iommu->root_entry);
642         iommu->root_entry = NULL;
643 out:
644         spin_unlock_irqrestore(&iommu->lock, flags);
645 }
646
647 /* page table handling */
648 #define LEVEL_STRIDE            (9)
649 #define LEVEL_MASK              (((u64)1 << LEVEL_STRIDE) - 1)
650
651 static inline int agaw_to_level(int agaw)
652 {
653         return agaw + 2;
654 }
655
656 static inline int agaw_to_width(int agaw)
657 {
658         return 30 + agaw * LEVEL_STRIDE;
659
660 }
661
662 static inline int width_to_agaw(int width)
663 {
664         return (width - 30) / LEVEL_STRIDE;
665 }
666
667 static inline unsigned int level_to_offset_bits(int level)
668 {
669         return (level - 1) * LEVEL_STRIDE;
670 }
671
672 static inline int pfn_level_offset(unsigned long pfn, int level)
673 {
674         return (pfn >> level_to_offset_bits(level)) & LEVEL_MASK;
675 }
676
677 static inline unsigned long level_mask(int level)
678 {
679         return -1UL << level_to_offset_bits(level);
680 }
681
682 static inline unsigned long level_size(int level)
683 {
684         return 1UL << level_to_offset_bits(level);
685 }
686
687 static inline unsigned long align_to_level(unsigned long pfn, int level)
688 {
689         return (pfn + level_size(level) - 1) & level_mask(level);
690 }
691
692 static struct dma_pte *pfn_to_dma_pte(struct dmar_domain *domain,
693                                       unsigned long pfn)
694 {
695         int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
696         struct dma_pte *parent, *pte = NULL;
697         int level = agaw_to_level(domain->agaw);
698         int offset;
699         unsigned long flags;
700
701         BUG_ON(!domain->pgd);
702         BUG_ON(addr_width < BITS_PER_LONG && pfn >> addr_width);
703         parent = domain->pgd;
704
705         spin_lock_irqsave(&domain->mapping_lock, flags);
706         while (level > 0) {
707                 void *tmp_page;
708
709                 offset = pfn_level_offset(pfn, level);
710                 pte = &parent[offset];
711                 if (level == 1)
712                         break;
713
714                 if (!dma_pte_present(pte)) {
715                         tmp_page = alloc_pgtable_page();
716
717                         if (!tmp_page) {
718                                 spin_unlock_irqrestore(&domain->mapping_lock,
719                                         flags);
720                                 return NULL;
721                         }
722                         domain_flush_cache(domain, tmp_page, PAGE_SIZE);
723                         dma_set_pte_pfn(pte, virt_to_dma_pfn(tmp_page));
724                         /*
725                          * high level table always sets r/w, last level page
726                          * table control read/write
727                          */
728                         dma_set_pte_readable(pte);
729                         dma_set_pte_writable(pte);
730                         domain_flush_cache(domain, pte, sizeof(*pte));
731                 }
732                 parent = phys_to_virt(dma_pte_addr(pte));
733                 level--;
734         }
735
736         spin_unlock_irqrestore(&domain->mapping_lock, flags);
737         return pte;
738 }
739
740 /* return address's pte at specific level */
741 static struct dma_pte *dma_pfn_level_pte(struct dmar_domain *domain,
742                                          unsigned long pfn,
743                                          int level)
744 {
745         struct dma_pte *parent, *pte = NULL;
746         int total = agaw_to_level(domain->agaw);
747         int offset;
748
749         parent = domain->pgd;
750         while (level <= total) {
751                 offset = pfn_level_offset(pfn, total);
752                 pte = &parent[offset];
753                 if (level == total)
754                         return pte;
755
756                 if (!dma_pte_present(pte))
757                         break;
758                 parent = phys_to_virt(dma_pte_addr(pte));
759                 total--;
760         }
761         return NULL;
762 }
763
764 /* clear last level pte, a tlb flush should be followed */
765 static void dma_pte_clear_range(struct dmar_domain *domain,
766                                 unsigned long start_pfn,
767                                 unsigned long last_pfn)
768 {
769         int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
770         struct dma_pte *first_pte, *pte;
771
772         BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
773         BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
774
775         /* we don't need lock here; nobody else touches the iova range */
776         while (start_pfn <= last_pfn) {
777                 first_pte = pte = dma_pfn_level_pte(domain, start_pfn, 1);
778                 if (!pte) {
779                         start_pfn = align_to_level(start_pfn + 1, 2);
780                         continue;
781                 }
782                 while (start_pfn <= last_pfn &&
783                        (unsigned long)pte >> VTD_PAGE_SHIFT ==
784                        (unsigned long)first_pte >> VTD_PAGE_SHIFT) {
785                         dma_clear_pte(pte);
786                         start_pfn++;
787                         pte++;
788                 }
789                 domain_flush_cache(domain, first_pte,
790                                    (void *)pte - (void *)first_pte);
791         }
792 }
793
794 /* free page table pages. last level pte should already be cleared */
795 static void dma_pte_free_pagetable(struct dmar_domain *domain,
796                                    unsigned long start_pfn,
797                                    unsigned long last_pfn)
798 {
799         int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
800         struct dma_pte *pte;
801         int total = agaw_to_level(domain->agaw);
802         int level;
803         unsigned long tmp;
804
805         BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
806         BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
807
808         /* we don't need lock here, nobody else touches the iova range */
809         level = 2;
810         while (level <= total) {
811                 tmp = align_to_level(start_pfn, level);
812
813                 /* Only clear this pte/pmd if we're asked to clear its
814                    _whole_ range */
815                 if (tmp + level_size(level) - 1 > last_pfn)
816                         return;
817
818                 while (tmp <= last_pfn) {
819                         pte = dma_pfn_level_pte(domain, tmp, level);
820                         if (pte) {
821                                 free_pgtable_page(
822                                         phys_to_virt(dma_pte_addr(pte)));
823                                 dma_clear_pte(pte);
824                                 domain_flush_cache(domain, pte, sizeof(*pte));
825                         }
826                         tmp += level_size(level);
827                 }
828                 level++;
829         }
830         /* free pgd */
831         if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
832                 free_pgtable_page(domain->pgd);
833                 domain->pgd = NULL;
834         }
835 }
836
837 /* iommu handling */
838 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
839 {
840         struct root_entry *root;
841         unsigned long flags;
842
843         root = (struct root_entry *)alloc_pgtable_page();
844         if (!root)
845                 return -ENOMEM;
846
847         __iommu_flush_cache(iommu, root, ROOT_SIZE);
848
849         spin_lock_irqsave(&iommu->lock, flags);
850         iommu->root_entry = root;
851         spin_unlock_irqrestore(&iommu->lock, flags);
852
853         return 0;
854 }
855
856 static void iommu_set_root_entry(struct intel_iommu *iommu)
857 {
858         void *addr;
859         u32 sts;
860         unsigned long flag;
861
862         addr = iommu->root_entry;
863
864         spin_lock_irqsave(&iommu->register_lock, flag);
865         dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
866
867         writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
868
869         /* Make sure hardware complete it */
870         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
871                       readl, (sts & DMA_GSTS_RTPS), sts);
872
873         spin_unlock_irqrestore(&iommu->register_lock, flag);
874 }
875
876 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
877 {
878         u32 val;
879         unsigned long flag;
880
881         if (!rwbf_quirk && !cap_rwbf(iommu->cap))
882                 return;
883
884         spin_lock_irqsave(&iommu->register_lock, flag);
885         writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
886
887         /* Make sure hardware complete it */
888         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
889                       readl, (!(val & DMA_GSTS_WBFS)), val);
890
891         spin_unlock_irqrestore(&iommu->register_lock, flag);
892 }
893
894 /* return value determine if we need a write buffer flush */
895 static void __iommu_flush_context(struct intel_iommu *iommu,
896                                   u16 did, u16 source_id, u8 function_mask,
897                                   u64 type)
898 {
899         u64 val = 0;
900         unsigned long flag;
901
902         switch (type) {
903         case DMA_CCMD_GLOBAL_INVL:
904                 val = DMA_CCMD_GLOBAL_INVL;
905                 break;
906         case DMA_CCMD_DOMAIN_INVL:
907                 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
908                 break;
909         case DMA_CCMD_DEVICE_INVL:
910                 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
911                         | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
912                 break;
913         default:
914                 BUG();
915         }
916         val |= DMA_CCMD_ICC;
917
918         spin_lock_irqsave(&iommu->register_lock, flag);
919         dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
920
921         /* Make sure hardware complete it */
922         IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
923                 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
924
925         spin_unlock_irqrestore(&iommu->register_lock, flag);
926 }
927
928 /* return value determine if we need a write buffer flush */
929 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
930                                 u64 addr, unsigned int size_order, u64 type)
931 {
932         int tlb_offset = ecap_iotlb_offset(iommu->ecap);
933         u64 val = 0, val_iva = 0;
934         unsigned long flag;
935
936         switch (type) {
937         case DMA_TLB_GLOBAL_FLUSH:
938                 /* global flush doesn't need set IVA_REG */
939                 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
940                 break;
941         case DMA_TLB_DSI_FLUSH:
942                 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
943                 break;
944         case DMA_TLB_PSI_FLUSH:
945                 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
946                 /* Note: always flush non-leaf currently */
947                 val_iva = size_order | addr;
948                 break;
949         default:
950                 BUG();
951         }
952         /* Note: set drain read/write */
953 #if 0
954         /*
955          * This is probably to be super secure.. Looks like we can
956          * ignore it without any impact.
957          */
958         if (cap_read_drain(iommu->cap))
959                 val |= DMA_TLB_READ_DRAIN;
960 #endif
961         if (cap_write_drain(iommu->cap))
962                 val |= DMA_TLB_WRITE_DRAIN;
963
964         spin_lock_irqsave(&iommu->register_lock, flag);
965         /* Note: Only uses first TLB reg currently */
966         if (val_iva)
967                 dmar_writeq(iommu->reg + tlb_offset, val_iva);
968         dmar_writeq(iommu->reg + tlb_offset + 8, val);
969
970         /* Make sure hardware complete it */
971         IOMMU_WAIT_OP(iommu, tlb_offset + 8,
972                 dmar_readq, (!(val & DMA_TLB_IVT)), val);
973
974         spin_unlock_irqrestore(&iommu->register_lock, flag);
975
976         /* check IOTLB invalidation granularity */
977         if (DMA_TLB_IAIG(val) == 0)
978                 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
979         if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
980                 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
981                         (unsigned long long)DMA_TLB_IIRG(type),
982                         (unsigned long long)DMA_TLB_IAIG(val));
983 }
984
985 static struct device_domain_info *iommu_support_dev_iotlb(
986         struct dmar_domain *domain, int segment, u8 bus, u8 devfn)
987 {
988         int found = 0;
989         unsigned long flags;
990         struct device_domain_info *info;
991         struct intel_iommu *iommu = device_to_iommu(segment, bus, devfn);
992
993         if (!ecap_dev_iotlb_support(iommu->ecap))
994                 return NULL;
995
996         if (!iommu->qi)
997                 return NULL;
998
999         spin_lock_irqsave(&device_domain_lock, flags);
1000         list_for_each_entry(info, &domain->devices, link)
1001                 if (info->bus == bus && info->devfn == devfn) {
1002                         found = 1;
1003                         break;
1004                 }
1005         spin_unlock_irqrestore(&device_domain_lock, flags);
1006
1007         if (!found || !info->dev)
1008                 return NULL;
1009
1010         if (!pci_find_ext_capability(info->dev, PCI_EXT_CAP_ID_ATS))
1011                 return NULL;
1012
1013         if (!dmar_find_matched_atsr_unit(info->dev))
1014                 return NULL;
1015
1016         info->iommu = iommu;
1017
1018         return info;
1019 }
1020
1021 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
1022 {
1023         if (!info)
1024                 return;
1025
1026         pci_enable_ats(info->dev, VTD_PAGE_SHIFT);
1027 }
1028
1029 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1030 {
1031         if (!info->dev || !pci_ats_enabled(info->dev))
1032                 return;
1033
1034         pci_disable_ats(info->dev);
1035 }
1036
1037 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1038                                   u64 addr, unsigned mask)
1039 {
1040         u16 sid, qdep;
1041         unsigned long flags;
1042         struct device_domain_info *info;
1043
1044         spin_lock_irqsave(&device_domain_lock, flags);
1045         list_for_each_entry(info, &domain->devices, link) {
1046                 if (!info->dev || !pci_ats_enabled(info->dev))
1047                         continue;
1048
1049                 sid = info->bus << 8 | info->devfn;
1050                 qdep = pci_ats_queue_depth(info->dev);
1051                 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1052         }
1053         spin_unlock_irqrestore(&device_domain_lock, flags);
1054 }
1055
1056 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
1057                                   unsigned long pfn, unsigned int pages)
1058 {
1059         unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1060         uint64_t addr = (uint64_t)pfn << VTD_PAGE_SHIFT;
1061
1062         BUG_ON(pages == 0);
1063
1064         /*
1065          * Fallback to domain selective flush if no PSI support or the size is
1066          * too big.
1067          * PSI requires page size to be 2 ^ x, and the base address is naturally
1068          * aligned to the size
1069          */
1070         if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1071                 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1072                                                 DMA_TLB_DSI_FLUSH);
1073         else
1074                 iommu->flush.flush_iotlb(iommu, did, addr, mask,
1075                                                 DMA_TLB_PSI_FLUSH);
1076
1077         /*
1078          * In caching mode, domain ID 0 is reserved for non-present to present
1079          * mapping flush. Device IOTLB doesn't need to be flushed in this case.
1080          */
1081         if (!cap_caching_mode(iommu->cap) || did)
1082                 iommu_flush_dev_iotlb(iommu->domains[did], addr, mask);
1083 }
1084
1085 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1086 {
1087         u32 pmen;
1088         unsigned long flags;
1089
1090         spin_lock_irqsave(&iommu->register_lock, flags);
1091         pmen = readl(iommu->reg + DMAR_PMEN_REG);
1092         pmen &= ~DMA_PMEN_EPM;
1093         writel(pmen, iommu->reg + DMAR_PMEN_REG);
1094
1095         /* wait for the protected region status bit to clear */
1096         IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1097                 readl, !(pmen & DMA_PMEN_PRS), pmen);
1098
1099         spin_unlock_irqrestore(&iommu->register_lock, flags);
1100 }
1101
1102 static int iommu_enable_translation(struct intel_iommu *iommu)
1103 {
1104         u32 sts;
1105         unsigned long flags;
1106
1107         spin_lock_irqsave(&iommu->register_lock, flags);
1108         iommu->gcmd |= DMA_GCMD_TE;
1109         writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1110
1111         /* Make sure hardware complete it */
1112         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1113                       readl, (sts & DMA_GSTS_TES), sts);
1114
1115         spin_unlock_irqrestore(&iommu->register_lock, flags);
1116         return 0;
1117 }
1118
1119 static int iommu_disable_translation(struct intel_iommu *iommu)
1120 {
1121         u32 sts;
1122         unsigned long flag;
1123
1124         spin_lock_irqsave(&iommu->register_lock, flag);
1125         iommu->gcmd &= ~DMA_GCMD_TE;
1126         writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1127
1128         /* Make sure hardware complete it */
1129         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1130                       readl, (!(sts & DMA_GSTS_TES)), sts);
1131
1132         spin_unlock_irqrestore(&iommu->register_lock, flag);
1133         return 0;
1134 }
1135
1136
1137 static int iommu_init_domains(struct intel_iommu *iommu)
1138 {
1139         unsigned long ndomains;
1140         unsigned long nlongs;
1141
1142         ndomains = cap_ndoms(iommu->cap);
1143         pr_debug("Number of Domains supportd <%ld>\n", ndomains);
1144         nlongs = BITS_TO_LONGS(ndomains);
1145
1146         /* TBD: there might be 64K domains,
1147          * consider other allocation for future chip
1148          */
1149         iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1150         if (!iommu->domain_ids) {
1151                 printk(KERN_ERR "Allocating domain id array failed\n");
1152                 return -ENOMEM;
1153         }
1154         iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1155                         GFP_KERNEL);
1156         if (!iommu->domains) {
1157                 printk(KERN_ERR "Allocating domain array failed\n");
1158                 kfree(iommu->domain_ids);
1159                 return -ENOMEM;
1160         }
1161
1162         spin_lock_init(&iommu->lock);
1163
1164         /*
1165          * if Caching mode is set, then invalid translations are tagged
1166          * with domainid 0. Hence we need to pre-allocate it.
1167          */
1168         if (cap_caching_mode(iommu->cap))
1169                 set_bit(0, iommu->domain_ids);
1170         return 0;
1171 }
1172
1173
1174 static void domain_exit(struct dmar_domain *domain);
1175 static void vm_domain_exit(struct dmar_domain *domain);
1176
1177 void free_dmar_iommu(struct intel_iommu *iommu)
1178 {
1179         struct dmar_domain *domain;
1180         int i;
1181         unsigned long flags;
1182
1183         i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1184         for (; i < cap_ndoms(iommu->cap); ) {
1185                 domain = iommu->domains[i];
1186                 clear_bit(i, iommu->domain_ids);
1187
1188                 spin_lock_irqsave(&domain->iommu_lock, flags);
1189                 if (--domain->iommu_count == 0) {
1190                         if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
1191                                 vm_domain_exit(domain);
1192                         else
1193                                 domain_exit(domain);
1194                 }
1195                 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1196
1197                 i = find_next_bit(iommu->domain_ids,
1198                         cap_ndoms(iommu->cap), i+1);
1199         }
1200
1201         if (iommu->gcmd & DMA_GCMD_TE)
1202                 iommu_disable_translation(iommu);
1203
1204         if (iommu->irq) {
1205                 set_irq_data(iommu->irq, NULL);
1206                 /* This will mask the irq */
1207                 free_irq(iommu->irq, iommu);
1208                 destroy_irq(iommu->irq);
1209         }
1210
1211         kfree(iommu->domains);
1212         kfree(iommu->domain_ids);
1213
1214         g_iommus[iommu->seq_id] = NULL;
1215
1216         /* if all iommus are freed, free g_iommus */
1217         for (i = 0; i < g_num_of_iommus; i++) {
1218                 if (g_iommus[i])
1219                         break;
1220         }
1221
1222         if (i == g_num_of_iommus)
1223                 kfree(g_iommus);
1224
1225         /* free context mapping */
1226         free_context_table(iommu);
1227 }
1228
1229 static struct dmar_domain *alloc_domain(void)
1230 {
1231         struct dmar_domain *domain;
1232
1233         domain = alloc_domain_mem();
1234         if (!domain)
1235                 return NULL;
1236
1237         memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
1238         domain->flags = 0;
1239
1240         return domain;
1241 }
1242
1243 static int iommu_attach_domain(struct dmar_domain *domain,
1244                                struct intel_iommu *iommu)
1245 {
1246         int num;
1247         unsigned long ndomains;
1248         unsigned long flags;
1249
1250         ndomains = cap_ndoms(iommu->cap);
1251
1252         spin_lock_irqsave(&iommu->lock, flags);
1253
1254         num = find_first_zero_bit(iommu->domain_ids, ndomains);
1255         if (num >= ndomains) {
1256                 spin_unlock_irqrestore(&iommu->lock, flags);
1257                 printk(KERN_ERR "IOMMU: no free domain ids\n");
1258                 return -ENOMEM;
1259         }
1260
1261         domain->id = num;
1262         set_bit(num, iommu->domain_ids);
1263         set_bit(iommu->seq_id, &domain->iommu_bmp);
1264         iommu->domains[num] = domain;
1265         spin_unlock_irqrestore(&iommu->lock, flags);
1266
1267         return 0;
1268 }
1269
1270 static void iommu_detach_domain(struct dmar_domain *domain,
1271                                 struct intel_iommu *iommu)
1272 {
1273         unsigned long flags;
1274         int num, ndomains;
1275         int found = 0;
1276
1277         spin_lock_irqsave(&iommu->lock, flags);
1278         ndomains = cap_ndoms(iommu->cap);
1279         num = find_first_bit(iommu->domain_ids, ndomains);
1280         for (; num < ndomains; ) {
1281                 if (iommu->domains[num] == domain) {
1282                         found = 1;
1283                         break;
1284                 }
1285                 num = find_next_bit(iommu->domain_ids,
1286                                     cap_ndoms(iommu->cap), num+1);
1287         }
1288
1289         if (found) {
1290                 clear_bit(num, iommu->domain_ids);
1291                 clear_bit(iommu->seq_id, &domain->iommu_bmp);
1292                 iommu->domains[num] = NULL;
1293         }
1294         spin_unlock_irqrestore(&iommu->lock, flags);
1295 }
1296
1297 static struct iova_domain reserved_iova_list;
1298 static struct lock_class_key reserved_alloc_key;
1299 static struct lock_class_key reserved_rbtree_key;
1300
1301 static void dmar_init_reserved_ranges(void)
1302 {
1303         struct pci_dev *pdev = NULL;
1304         struct iova *iova;
1305         int i;
1306
1307         init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1308
1309         lockdep_set_class(&reserved_iova_list.iova_alloc_lock,
1310                 &reserved_alloc_key);
1311         lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1312                 &reserved_rbtree_key);
1313
1314         /* IOAPIC ranges shouldn't be accessed by DMA */
1315         iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1316                 IOVA_PFN(IOAPIC_RANGE_END));
1317         if (!iova)
1318                 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1319
1320         /* Reserve all PCI MMIO to avoid peer-to-peer access */
1321         for_each_pci_dev(pdev) {
1322                 struct resource *r;
1323
1324                 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1325                         r = &pdev->resource[i];
1326                         if (!r->flags || !(r->flags & IORESOURCE_MEM))
1327                                 continue;
1328                         iova = reserve_iova(&reserved_iova_list,
1329                                             IOVA_PFN(r->start),
1330                                             IOVA_PFN(r->end));
1331                         if (!iova)
1332                                 printk(KERN_ERR "Reserve iova failed\n");
1333                 }
1334         }
1335
1336 }
1337
1338 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1339 {
1340         copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1341 }
1342
1343 static inline int guestwidth_to_adjustwidth(int gaw)
1344 {
1345         int agaw;
1346         int r = (gaw - 12) % 9;
1347
1348         if (r == 0)
1349                 agaw = gaw;
1350         else
1351                 agaw = gaw + 9 - r;
1352         if (agaw > 64)
1353                 agaw = 64;
1354         return agaw;
1355 }
1356
1357 static int domain_init(struct dmar_domain *domain, int guest_width)
1358 {
1359         struct intel_iommu *iommu;
1360         int adjust_width, agaw;
1361         unsigned long sagaw;
1362
1363         init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1364         spin_lock_init(&domain->mapping_lock);
1365         spin_lock_init(&domain->iommu_lock);
1366
1367         domain_reserve_special_ranges(domain);
1368
1369         /* calculate AGAW */
1370         iommu = domain_get_iommu(domain);
1371         if (guest_width > cap_mgaw(iommu->cap))
1372                 guest_width = cap_mgaw(iommu->cap);
1373         domain->gaw = guest_width;
1374         adjust_width = guestwidth_to_adjustwidth(guest_width);
1375         agaw = width_to_agaw(adjust_width);
1376         sagaw = cap_sagaw(iommu->cap);
1377         if (!test_bit(agaw, &sagaw)) {
1378                 /* hardware doesn't support it, choose a bigger one */
1379                 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1380                 agaw = find_next_bit(&sagaw, 5, agaw);
1381                 if (agaw >= 5)
1382                         return -ENODEV;
1383         }
1384         domain->agaw = agaw;
1385         INIT_LIST_HEAD(&domain->devices);
1386
1387         if (ecap_coherent(iommu->ecap))
1388                 domain->iommu_coherency = 1;
1389         else
1390                 domain->iommu_coherency = 0;
1391
1392         if (ecap_sc_support(iommu->ecap))
1393                 domain->iommu_snooping = 1;
1394         else
1395                 domain->iommu_snooping = 0;
1396
1397         domain->iommu_count = 1;
1398
1399         /* always allocate the top pgd */
1400         domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1401         if (!domain->pgd)
1402                 return -ENOMEM;
1403         __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1404         return 0;
1405 }
1406
1407 static void domain_exit(struct dmar_domain *domain)
1408 {
1409         struct dmar_drhd_unit *drhd;
1410         struct intel_iommu *iommu;
1411
1412         /* Domain 0 is reserved, so dont process it */
1413         if (!domain)
1414                 return;
1415
1416         domain_remove_dev_info(domain);
1417         /* destroy iovas */
1418         put_iova_domain(&domain->iovad);
1419
1420         /* clear ptes */
1421         dma_pte_clear_range(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1422
1423         /* free page tables */
1424         dma_pte_free_pagetable(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1425
1426         for_each_active_iommu(iommu, drhd)
1427                 if (test_bit(iommu->seq_id, &domain->iommu_bmp))
1428                         iommu_detach_domain(domain, iommu);
1429
1430         free_domain_mem(domain);
1431 }
1432
1433 static int domain_context_mapping_one(struct dmar_domain *domain, int segment,
1434                                  u8 bus, u8 devfn, int translation)
1435 {
1436         struct context_entry *context;
1437         unsigned long flags;
1438         struct intel_iommu *iommu;
1439         struct dma_pte *pgd;
1440         unsigned long num;
1441         unsigned long ndomains;
1442         int id;
1443         int agaw;
1444         struct device_domain_info *info = NULL;
1445
1446         pr_debug("Set context mapping for %02x:%02x.%d\n",
1447                 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1448
1449         BUG_ON(!domain->pgd);
1450         BUG_ON(translation != CONTEXT_TT_PASS_THROUGH &&
1451                translation != CONTEXT_TT_MULTI_LEVEL);
1452
1453         iommu = device_to_iommu(segment, bus, devfn);
1454         if (!iommu)
1455                 return -ENODEV;
1456
1457         context = device_to_context_entry(iommu, bus, devfn);
1458         if (!context)
1459                 return -ENOMEM;
1460         spin_lock_irqsave(&iommu->lock, flags);
1461         if (context_present(context)) {
1462                 spin_unlock_irqrestore(&iommu->lock, flags);
1463                 return 0;
1464         }
1465
1466         id = domain->id;
1467         pgd = domain->pgd;
1468
1469         if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
1470             domain->flags & DOMAIN_FLAG_STATIC_IDENTITY) {
1471                 int found = 0;
1472
1473                 /* find an available domain id for this device in iommu */
1474                 ndomains = cap_ndoms(iommu->cap);
1475                 num = find_first_bit(iommu->domain_ids, ndomains);
1476                 for (; num < ndomains; ) {
1477                         if (iommu->domains[num] == domain) {
1478                                 id = num;
1479                                 found = 1;
1480                                 break;
1481                         }
1482                         num = find_next_bit(iommu->domain_ids,
1483                                             cap_ndoms(iommu->cap), num+1);
1484                 }
1485
1486                 if (found == 0) {
1487                         num = find_first_zero_bit(iommu->domain_ids, ndomains);
1488                         if (num >= ndomains) {
1489                                 spin_unlock_irqrestore(&iommu->lock, flags);
1490                                 printk(KERN_ERR "IOMMU: no free domain ids\n");
1491                                 return -EFAULT;
1492                         }
1493
1494                         set_bit(num, iommu->domain_ids);
1495                         set_bit(iommu->seq_id, &domain->iommu_bmp);
1496                         iommu->domains[num] = domain;
1497                         id = num;
1498                 }
1499
1500                 /* Skip top levels of page tables for
1501                  * iommu which has less agaw than default.
1502                  */
1503                 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1504                         pgd = phys_to_virt(dma_pte_addr(pgd));
1505                         if (!dma_pte_present(pgd)) {
1506                                 spin_unlock_irqrestore(&iommu->lock, flags);
1507                                 return -ENOMEM;
1508                         }
1509                 }
1510         }
1511
1512         context_set_domain_id(context, id);
1513
1514         if (translation != CONTEXT_TT_PASS_THROUGH) {
1515                 info = iommu_support_dev_iotlb(domain, segment, bus, devfn);
1516                 translation = info ? CONTEXT_TT_DEV_IOTLB :
1517                                      CONTEXT_TT_MULTI_LEVEL;
1518         }
1519         /*
1520          * In pass through mode, AW must be programmed to indicate the largest
1521          * AGAW value supported by hardware. And ASR is ignored by hardware.
1522          */
1523         if (unlikely(translation == CONTEXT_TT_PASS_THROUGH))
1524                 context_set_address_width(context, iommu->msagaw);
1525         else {
1526                 context_set_address_root(context, virt_to_phys(pgd));
1527                 context_set_address_width(context, iommu->agaw);
1528         }
1529
1530         context_set_translation_type(context, translation);
1531         context_set_fault_enable(context);
1532         context_set_present(context);
1533         domain_flush_cache(domain, context, sizeof(*context));
1534
1535         /*
1536          * It's a non-present to present mapping. If hardware doesn't cache
1537          * non-present entry we only need to flush the write-buffer. If the
1538          * _does_ cache non-present entries, then it does so in the special
1539          * domain #0, which we have to flush:
1540          */
1541         if (cap_caching_mode(iommu->cap)) {
1542                 iommu->flush.flush_context(iommu, 0,
1543                                            (((u16)bus) << 8) | devfn,
1544                                            DMA_CCMD_MASK_NOBIT,
1545                                            DMA_CCMD_DEVICE_INVL);
1546                 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
1547         } else {
1548                 iommu_flush_write_buffer(iommu);
1549         }
1550         iommu_enable_dev_iotlb(info);
1551         spin_unlock_irqrestore(&iommu->lock, flags);
1552
1553         spin_lock_irqsave(&domain->iommu_lock, flags);
1554         if (!test_and_set_bit(iommu->seq_id, &domain->iommu_bmp)) {
1555                 domain->iommu_count++;
1556                 domain_update_iommu_cap(domain);
1557         }
1558         spin_unlock_irqrestore(&domain->iommu_lock, flags);
1559         return 0;
1560 }
1561
1562 static int
1563 domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev,
1564                         int translation)
1565 {
1566         int ret;
1567         struct pci_dev *tmp, *parent;
1568
1569         ret = domain_context_mapping_one(domain, pci_domain_nr(pdev->bus),
1570                                          pdev->bus->number, pdev->devfn,
1571                                          translation);
1572         if (ret)
1573                 return ret;
1574
1575         /* dependent device mapping */
1576         tmp = pci_find_upstream_pcie_bridge(pdev);
1577         if (!tmp)
1578                 return 0;
1579         /* Secondary interface's bus number and devfn 0 */
1580         parent = pdev->bus->self;
1581         while (parent != tmp) {
1582                 ret = domain_context_mapping_one(domain,
1583                                                  pci_domain_nr(parent->bus),
1584                                                  parent->bus->number,
1585                                                  parent->devfn, translation);
1586                 if (ret)
1587                         return ret;
1588                 parent = parent->bus->self;
1589         }
1590         if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1591                 return domain_context_mapping_one(domain,
1592                                         pci_domain_nr(tmp->subordinate),
1593                                         tmp->subordinate->number, 0,
1594                                         translation);
1595         else /* this is a legacy PCI bridge */
1596                 return domain_context_mapping_one(domain,
1597                                                   pci_domain_nr(tmp->bus),
1598                                                   tmp->bus->number,
1599                                                   tmp->devfn,
1600                                                   translation);
1601 }
1602
1603 static int domain_context_mapped(struct pci_dev *pdev)
1604 {
1605         int ret;
1606         struct pci_dev *tmp, *parent;
1607         struct intel_iommu *iommu;
1608
1609         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
1610                                 pdev->devfn);
1611         if (!iommu)
1612                 return -ENODEV;
1613
1614         ret = device_context_mapped(iommu, pdev->bus->number, pdev->devfn);
1615         if (!ret)
1616                 return ret;
1617         /* dependent device mapping */
1618         tmp = pci_find_upstream_pcie_bridge(pdev);
1619         if (!tmp)
1620                 return ret;
1621         /* Secondary interface's bus number and devfn 0 */
1622         parent = pdev->bus->self;
1623         while (parent != tmp) {
1624                 ret = device_context_mapped(iommu, parent->bus->number,
1625                                             parent->devfn);
1626                 if (!ret)
1627                         return ret;
1628                 parent = parent->bus->self;
1629         }
1630         if (tmp->is_pcie)
1631                 return device_context_mapped(iommu, tmp->subordinate->number,
1632                                              0);
1633         else
1634                 return device_context_mapped(iommu, tmp->bus->number,
1635                                              tmp->devfn);
1636 }
1637
1638 static int domain_sg_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1639                              struct scatterlist *sg, unsigned long nr_pages,
1640                              int prot)
1641 {
1642         struct dma_pte *first_pte = NULL, *pte = NULL;
1643         uint64_t pteval;
1644         int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
1645         unsigned long sg_res = 0;
1646
1647         BUG_ON(addr_width < BITS_PER_LONG && (iov_pfn + nr_pages - 1) >> addr_width);
1648
1649         if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1650                 return -EINVAL;
1651
1652         prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
1653
1654         while (nr_pages--) {
1655                 if (!sg_res) {
1656                         sg_res = (sg->offset + sg->length + VTD_PAGE_SIZE - 1) >> VTD_PAGE_SHIFT;
1657                         sg->dma_address = ((dma_addr_t)iov_pfn << VTD_PAGE_SHIFT) + sg->offset;
1658                         sg->dma_length = sg->length;
1659                         pteval = page_to_phys(sg_page(sg)) | prot;
1660                 }
1661                 if (!pte) {
1662                         first_pte = pte = pfn_to_dma_pte(domain, iov_pfn);
1663                         if (!pte)
1664                                 return -ENOMEM;
1665                 }
1666                 /* We don't need lock here, nobody else
1667                  * touches the iova range
1668                  */
1669                 BUG_ON(dma_pte_addr(pte));
1670                 pte->val = pteval;
1671                 pte++;
1672                 if (!nr_pages ||
1673                     (unsigned long)pte >> VTD_PAGE_SHIFT !=
1674                     (unsigned long)first_pte >> VTD_PAGE_SHIFT) {
1675                         domain_flush_cache(domain, first_pte,
1676                                            (void *)pte - (void *)first_pte);
1677                         pte = NULL;
1678                 }
1679                 iov_pfn++;
1680                 pteval += VTD_PAGE_SIZE;
1681                 sg_res--;
1682                 if (!sg_res)
1683                         sg = sg_next(sg);
1684         }
1685         return 0;
1686 }
1687
1688 static int domain_pfn_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1689                               unsigned long phys_pfn, unsigned long nr_pages,
1690                               int prot)
1691 {
1692         struct dma_pte *first_pte = NULL, *pte = NULL;
1693         int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
1694
1695         BUG_ON(addr_width < BITS_PER_LONG && (iov_pfn + nr_pages - 1) >> addr_width);
1696
1697         if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1698                 return -EINVAL;
1699
1700         prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
1701
1702         while (nr_pages--) {
1703                 if (!pte) {
1704                         first_pte = pte = pfn_to_dma_pte(domain, iov_pfn);
1705                         if (!pte)
1706                                 return -ENOMEM;
1707                 }
1708                 /* We don't need lock here, nobody else
1709                  * touches the iova range
1710                  */
1711                 BUG_ON(dma_pte_addr(pte));
1712                 pte->val = (phys_pfn << VTD_PAGE_SHIFT) | prot;
1713                 pte++;
1714                 if (!nr_pages ||
1715                     (unsigned long)pte >> VTD_PAGE_SHIFT !=
1716                     (unsigned long)first_pte >> VTD_PAGE_SHIFT) {
1717                         domain_flush_cache(domain, first_pte,
1718                                            (void *)pte - (void *)first_pte);
1719                         pte = NULL;
1720                 }
1721                 iov_pfn++;
1722                 phys_pfn++;
1723         }
1724         return 0;
1725 }
1726
1727 static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
1728 {
1729         if (!iommu)
1730                 return;
1731
1732         clear_context_table(iommu, bus, devfn);
1733         iommu->flush.flush_context(iommu, 0, 0, 0,
1734                                            DMA_CCMD_GLOBAL_INVL);
1735         iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
1736 }
1737
1738 static void domain_remove_dev_info(struct dmar_domain *domain)
1739 {
1740         struct device_domain_info *info;
1741         unsigned long flags;
1742         struct intel_iommu *iommu;
1743
1744         spin_lock_irqsave(&device_domain_lock, flags);
1745         while (!list_empty(&domain->devices)) {
1746                 info = list_entry(domain->devices.next,
1747                         struct device_domain_info, link);
1748                 list_del(&info->link);
1749                 list_del(&info->global);
1750                 if (info->dev)
1751                         info->dev->dev.archdata.iommu = NULL;
1752                 spin_unlock_irqrestore(&device_domain_lock, flags);
1753
1754                 iommu_disable_dev_iotlb(info);
1755                 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
1756                 iommu_detach_dev(iommu, info->bus, info->devfn);
1757                 free_devinfo_mem(info);
1758
1759                 spin_lock_irqsave(&device_domain_lock, flags);
1760         }
1761         spin_unlock_irqrestore(&device_domain_lock, flags);
1762 }
1763
1764 /*
1765  * find_domain
1766  * Note: we use struct pci_dev->dev.archdata.iommu stores the info
1767  */
1768 static struct dmar_domain *
1769 find_domain(struct pci_dev *pdev)
1770 {
1771         struct device_domain_info *info;
1772
1773         /* No lock here, assumes no domain exit in normal case */
1774         info = pdev->dev.archdata.iommu;
1775         if (info)
1776                 return info->domain;
1777         return NULL;
1778 }
1779
1780 /* domain is initialized */
1781 static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1782 {
1783         struct dmar_domain *domain, *found = NULL;
1784         struct intel_iommu *iommu;
1785         struct dmar_drhd_unit *drhd;
1786         struct device_domain_info *info, *tmp;
1787         struct pci_dev *dev_tmp;
1788         unsigned long flags;
1789         int bus = 0, devfn = 0;
1790         int segment;
1791         int ret;
1792
1793         domain = find_domain(pdev);
1794         if (domain)
1795                 return domain;
1796
1797         segment = pci_domain_nr(pdev->bus);
1798
1799         dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1800         if (dev_tmp) {
1801                 if (dev_tmp->is_pcie) {
1802                         bus = dev_tmp->subordinate->number;
1803                         devfn = 0;
1804                 } else {
1805                         bus = dev_tmp->bus->number;
1806                         devfn = dev_tmp->devfn;
1807                 }
1808                 spin_lock_irqsave(&device_domain_lock, flags);
1809                 list_for_each_entry(info, &device_domain_list, global) {
1810                         if (info->segment == segment &&
1811                             info->bus == bus && info->devfn == devfn) {
1812                                 found = info->domain;
1813                                 break;
1814                         }
1815                 }
1816                 spin_unlock_irqrestore(&device_domain_lock, flags);
1817                 /* pcie-pci bridge already has a domain, uses it */
1818                 if (found) {
1819                         domain = found;
1820                         goto found_domain;
1821                 }
1822         }
1823
1824         domain = alloc_domain();
1825         if (!domain)
1826                 goto error;
1827
1828         /* Allocate new domain for the device */
1829         drhd = dmar_find_matched_drhd_unit(pdev);
1830         if (!drhd) {
1831                 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1832                         pci_name(pdev));
1833                 return NULL;
1834         }
1835         iommu = drhd->iommu;
1836
1837         ret = iommu_attach_domain(domain, iommu);
1838         if (ret) {
1839                 domain_exit(domain);
1840                 goto error;
1841         }
1842
1843         if (domain_init(domain, gaw)) {
1844                 domain_exit(domain);
1845                 goto error;
1846         }
1847
1848         /* register pcie-to-pci device */
1849         if (dev_tmp) {
1850                 info = alloc_devinfo_mem();
1851                 if (!info) {
1852                         domain_exit(domain);
1853                         goto error;
1854                 }
1855                 info->segment = segment;
1856                 info->bus = bus;
1857                 info->devfn = devfn;
1858                 info->dev = NULL;
1859                 info->domain = domain;
1860                 /* This domain is shared by devices under p2p bridge */
1861                 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
1862
1863                 /* pcie-to-pci bridge already has a domain, uses it */
1864                 found = NULL;
1865                 spin_lock_irqsave(&device_domain_lock, flags);
1866                 list_for_each_entry(tmp, &device_domain_list, global) {
1867                         if (tmp->segment == segment &&
1868                             tmp->bus == bus && tmp->devfn == devfn) {
1869                                 found = tmp->domain;
1870                                 break;
1871                         }
1872                 }
1873                 if (found) {
1874                         free_devinfo_mem(info);
1875                         domain_exit(domain);
1876                         domain = found;
1877                 } else {
1878                         list_add(&info->link, &domain->devices);
1879                         list_add(&info->global, &device_domain_list);
1880                 }
1881                 spin_unlock_irqrestore(&device_domain_lock, flags);
1882         }
1883
1884 found_domain:
1885         info = alloc_devinfo_mem();
1886         if (!info)
1887                 goto error;
1888         info->segment = segment;
1889         info->bus = pdev->bus->number;
1890         info->devfn = pdev->devfn;
1891         info->dev = pdev;
1892         info->domain = domain;
1893         spin_lock_irqsave(&device_domain_lock, flags);
1894         /* somebody is fast */
1895         found = find_domain(pdev);
1896         if (found != NULL) {
1897                 spin_unlock_irqrestore(&device_domain_lock, flags);
1898                 if (found != domain) {
1899                         domain_exit(domain);
1900                         domain = found;
1901                 }
1902                 free_devinfo_mem(info);
1903                 return domain;
1904         }
1905         list_add(&info->link, &domain->devices);
1906         list_add(&info->global, &device_domain_list);
1907         pdev->dev.archdata.iommu = info;
1908         spin_unlock_irqrestore(&device_domain_lock, flags);
1909         return domain;
1910 error:
1911         /* recheck it here, maybe others set it */
1912         return find_domain(pdev);
1913 }
1914
1915 static int iommu_identity_mapping;
1916
1917 static int iommu_domain_identity_map(struct dmar_domain *domain,
1918                                      unsigned long long start,
1919                                      unsigned long long end)
1920 {
1921         unsigned long first_vpfn = start >> VTD_PAGE_SHIFT;
1922         unsigned long last_vpfn = end >> VTD_PAGE_SHIFT;
1923
1924         if (!reserve_iova(&domain->iovad, dma_to_mm_pfn(first_vpfn),
1925                           dma_to_mm_pfn(last_vpfn))) {
1926                 printk(KERN_ERR "IOMMU: reserve iova failed\n");
1927                 return -ENOMEM;
1928         }
1929
1930         pr_debug("Mapping reserved region %llx-%llx for domain %d\n",
1931                  start, end, domain->id);
1932         /*
1933          * RMRR range might have overlap with physical memory range,
1934          * clear it first
1935          */
1936         dma_pte_clear_range(domain, first_vpfn, last_vpfn);
1937
1938         return domain_pfn_mapping(domain, first_vpfn, first_vpfn,
1939                                   last_vpfn - first_vpfn + 1,
1940                                   DMA_PTE_READ|DMA_PTE_WRITE);
1941 }
1942
1943 static int iommu_prepare_identity_map(struct pci_dev *pdev,
1944                                       unsigned long long start,
1945                                       unsigned long long end)
1946 {
1947         struct dmar_domain *domain;
1948         int ret;
1949
1950         printk(KERN_INFO
1951                "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1952                pci_name(pdev), start, end);
1953
1954         domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1955         if (!domain)
1956                 return -ENOMEM;
1957
1958         ret = iommu_domain_identity_map(domain, start, end);
1959         if (ret)
1960                 goto error;
1961
1962         /* context entry init */
1963         ret = domain_context_mapping(domain, pdev, CONTEXT_TT_MULTI_LEVEL);
1964         if (ret)
1965                 goto error;
1966
1967         return 0;
1968
1969  error:
1970         domain_exit(domain);
1971         return ret;
1972 }
1973
1974 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1975         struct pci_dev *pdev)
1976 {
1977         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1978                 return 0;
1979         return iommu_prepare_identity_map(pdev, rmrr->base_address,
1980                 rmrr->end_address + 1);
1981 }
1982
1983 #ifdef CONFIG_DMAR_FLOPPY_WA
1984 static inline void iommu_prepare_isa(void)
1985 {
1986         struct pci_dev *pdev;
1987         int ret;
1988
1989         pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1990         if (!pdev)
1991                 return;
1992
1993         printk(KERN_INFO "IOMMU: Prepare 0-16MiB unity mapping for LPC\n");
1994         ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1995
1996         if (ret)
1997                 printk(KERN_ERR "IOMMU: Failed to create 0-16MiB identity map; "
1998                        "floppy might not work\n");
1999
2000 }
2001 #else
2002 static inline void iommu_prepare_isa(void)
2003 {
2004         return;
2005 }
2006 #endif /* !CONFIG_DMAR_FLPY_WA */
2007
2008 /* Initialize each context entry as pass through.*/
2009 static int __init init_context_pass_through(void)
2010 {
2011         struct pci_dev *pdev = NULL;
2012         struct dmar_domain *domain;
2013         int ret;
2014
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;
2021         }
2022         return 0;
2023 }
2024
2025 static int md_domain_init(struct dmar_domain *domain, int guest_width);
2026
2027 static int __init si_domain_work_fn(unsigned long start_pfn,
2028                                     unsigned long end_pfn, void *datax)
2029 {
2030         int *ret = datax;
2031
2032         *ret = iommu_domain_identity_map(si_domain,
2033                                          (uint64_t)start_pfn << PAGE_SHIFT,
2034                                          (uint64_t)end_pfn << PAGE_SHIFT);
2035         return *ret;
2036
2037 }
2038
2039 static int si_domain_init(void)
2040 {
2041         struct dmar_drhd_unit *drhd;
2042         struct intel_iommu *iommu;
2043         int nid, ret = 0;
2044
2045         si_domain = alloc_domain();
2046         if (!si_domain)
2047                 return -EFAULT;
2048
2049         pr_debug("Identity mapping domain is domain %d\n", si_domain->id);
2050
2051         for_each_active_iommu(iommu, drhd) {
2052                 ret = iommu_attach_domain(si_domain, iommu);
2053                 if (ret) {
2054                         domain_exit(si_domain);
2055                         return -EFAULT;
2056                 }
2057         }
2058
2059         if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2060                 domain_exit(si_domain);
2061                 return -EFAULT;
2062         }
2063
2064         si_domain->flags = DOMAIN_FLAG_STATIC_IDENTITY;
2065
2066         for_each_online_node(nid) {
2067                 work_with_active_regions(nid, si_domain_work_fn, &ret);
2068                 if (ret)
2069                         return ret;
2070         }
2071
2072         return 0;
2073 }
2074
2075 static void domain_remove_one_dev_info(struct dmar_domain *domain,
2076                                           struct pci_dev *pdev);
2077 static int identity_mapping(struct pci_dev *pdev)
2078 {
2079         struct device_domain_info *info;
2080
2081         if (likely(!iommu_identity_mapping))
2082                 return 0;
2083
2084
2085         list_for_each_entry(info, &si_domain->devices, link)
2086                 if (info->dev == pdev)
2087                         return 1;
2088         return 0;
2089 }
2090
2091 static int domain_add_dev_info(struct dmar_domain *domain,
2092                                   struct pci_dev *pdev)
2093 {
2094         struct device_domain_info *info;
2095         unsigned long flags;
2096
2097         info = alloc_devinfo_mem();
2098         if (!info)
2099                 return -ENOMEM;
2100
2101         info->segment = pci_domain_nr(pdev->bus);
2102         info->bus = pdev->bus->number;
2103         info->devfn = pdev->devfn;
2104         info->dev = pdev;
2105         info->domain = domain;
2106
2107         spin_lock_irqsave(&device_domain_lock, flags);
2108         list_add(&info->link, &domain->devices);
2109         list_add(&info->global, &device_domain_list);
2110         pdev->dev.archdata.iommu = info;
2111         spin_unlock_irqrestore(&device_domain_lock, flags);
2112
2113         return 0;
2114 }
2115
2116 static int iommu_prepare_static_identity_mapping(void)
2117 {
2118         struct pci_dev *pdev = NULL;
2119         int ret;
2120
2121         ret = si_domain_init();
2122         if (ret)
2123                 return -EFAULT;
2124
2125         for_each_pci_dev(pdev) {
2126                 printk(KERN_INFO "IOMMU: identity mapping for device %s\n",
2127                        pci_name(pdev));
2128
2129                 ret = domain_context_mapping(si_domain, pdev,
2130                                              CONTEXT_TT_MULTI_LEVEL);
2131                 if (ret)
2132                         return ret;
2133                 ret = domain_add_dev_info(si_domain, pdev);
2134                 if (ret)
2135                         return ret;
2136         }
2137
2138         return 0;
2139 }
2140
2141 int __init init_dmars(void)
2142 {
2143         struct dmar_drhd_unit *drhd;
2144         struct dmar_rmrr_unit *rmrr;
2145         struct pci_dev *pdev;
2146         struct intel_iommu *iommu;
2147         int i, ret;
2148         int pass_through = 1;
2149
2150         /*
2151          * In case pass through can not be enabled, iommu tries to use identity
2152          * mapping.
2153          */
2154         if (iommu_pass_through)
2155                 iommu_identity_mapping = 1;
2156
2157         /*
2158          * for each drhd
2159          *    allocate root
2160          *    initialize and program root entry to not present
2161          * endfor
2162          */
2163         for_each_drhd_unit(drhd) {
2164                 g_num_of_iommus++;
2165                 /*
2166                  * lock not needed as this is only incremented in the single
2167                  * threaded kernel __init code path all other access are read
2168                  * only
2169                  */
2170         }
2171
2172         g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
2173                         GFP_KERNEL);
2174         if (!g_iommus) {
2175                 printk(KERN_ERR "Allocating global iommu array failed\n");
2176                 ret = -ENOMEM;
2177                 goto error;
2178         }
2179
2180         deferred_flush = kzalloc(g_num_of_iommus *
2181                 sizeof(struct deferred_flush_tables), GFP_KERNEL);
2182         if (!deferred_flush) {
2183                 kfree(g_iommus);
2184                 ret = -ENOMEM;
2185                 goto error;
2186         }
2187
2188         for_each_drhd_unit(drhd) {
2189                 if (drhd->ignored)
2190                         continue;
2191
2192                 iommu = drhd->iommu;
2193                 g_iommus[iommu->seq_id] = iommu;
2194
2195                 ret = iommu_init_domains(iommu);
2196                 if (ret)
2197                         goto error;
2198
2199                 /*
2200                  * TBD:
2201                  * we could share the same root & context tables
2202                  * amoung all IOMMU's. Need to Split it later.
2203                  */
2204                 ret = iommu_alloc_root_entry(iommu);
2205                 if (ret) {
2206                         printk(KERN_ERR "IOMMU: allocate root entry failed\n");
2207                         goto error;
2208                 }
2209                 if (!ecap_pass_through(iommu->ecap))
2210                         pass_through = 0;
2211         }
2212         if (iommu_pass_through)
2213                 if (!pass_through) {
2214                         printk(KERN_INFO
2215                                "Pass Through is not supported by hardware.\n");
2216                         iommu_pass_through = 0;
2217                 }
2218
2219         /*
2220          * Start from the sane iommu hardware state.
2221          */
2222         for_each_drhd_unit(drhd) {
2223                 if (drhd->ignored)
2224                         continue;
2225
2226                 iommu = drhd->iommu;
2227
2228                 /*
2229                  * If the queued invalidation is already initialized by us
2230                  * (for example, while enabling interrupt-remapping) then
2231                  * we got the things already rolling from a sane state.
2232                  */
2233                 if (iommu->qi)
2234                         continue;
2235
2236                 /*
2237                  * Clear any previous faults.
2238                  */
2239                 dmar_fault(-1, iommu);
2240                 /*
2241                  * Disable queued invalidation if supported and already enabled
2242                  * before OS handover.
2243                  */
2244                 dmar_disable_qi(iommu);
2245         }
2246
2247         for_each_drhd_unit(drhd) {
2248                 if (drhd->ignored)
2249                         continue;
2250
2251                 iommu = drhd->iommu;
2252
2253                 if (dmar_enable_qi(iommu)) {
2254                         /*
2255                          * Queued Invalidate not enabled, use Register Based
2256                          * Invalidate
2257                          */
2258                         iommu->flush.flush_context = __iommu_flush_context;
2259                         iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2260                         printk(KERN_INFO "IOMMU 0x%Lx: using Register based "
2261                                "invalidation\n",
2262                                (unsigned long long)drhd->reg_base_addr);
2263                 } else {
2264                         iommu->flush.flush_context = qi_flush_context;
2265                         iommu->flush.flush_iotlb = qi_flush_iotlb;
2266                         printk(KERN_INFO "IOMMU 0x%Lx: using Queued "
2267                                "invalidation\n",
2268                                (unsigned long long)drhd->reg_base_addr);
2269                 }
2270         }
2271
2272         /*
2273          * If pass through is set and enabled, context entries of all pci
2274          * devices are intialized by pass through translation type.
2275          */
2276         if (iommu_pass_through) {
2277                 ret = init_context_pass_through();
2278                 if (ret) {
2279                         printk(KERN_ERR "IOMMU: Pass through init failed.\n");
2280                         iommu_pass_through = 0;
2281                 }
2282         }
2283
2284         /*
2285          * If pass through is not set or not enabled, setup context entries for
2286          * identity mappings for rmrr, gfx, and isa and may fall back to static
2287          * identity mapping if iommu_identity_mapping is set.
2288          */
2289         if (!iommu_pass_through) {
2290                 if (iommu_identity_mapping)
2291                         iommu_prepare_static_identity_mapping();
2292                 /*
2293                  * For each rmrr
2294                  *   for each dev attached to rmrr
2295                  *   do
2296                  *     locate drhd for dev, alloc domain for dev
2297                  *     allocate free domain
2298                  *     allocate page table entries for rmrr
2299                  *     if context not allocated for bus
2300                  *           allocate and init context
2301                  *           set present in root table for this bus
2302                  *     init context with domain, translation etc
2303                  *    endfor
2304                  * endfor
2305                  */
2306                 printk(KERN_INFO "IOMMU: Setting RMRR:\n");
2307                 for_each_rmrr_units(rmrr) {
2308                         for (i = 0; i < rmrr->devices_cnt; i++) {
2309                                 pdev = rmrr->devices[i];
2310                                 /*
2311                                  * some BIOS lists non-exist devices in DMAR
2312                                  * table.
2313                                  */
2314                                 if (!pdev)
2315                                         continue;
2316                                 ret = iommu_prepare_rmrr_dev(rmrr, pdev);
2317                                 if (ret)
2318                                         printk(KERN_ERR
2319                                  "IOMMU: mapping reserved region failed\n");
2320                         }
2321                 }
2322
2323                 iommu_prepare_isa();
2324         }
2325
2326         /*
2327          * for each drhd
2328          *   enable fault log
2329          *   global invalidate context cache
2330          *   global invalidate iotlb
2331          *   enable translation
2332          */
2333         for_each_drhd_unit(drhd) {
2334                 if (drhd->ignored)
2335                         continue;
2336                 iommu = drhd->iommu;
2337
2338                 iommu_flush_write_buffer(iommu);
2339
2340                 ret = dmar_set_interrupt(iommu);
2341                 if (ret)
2342                         goto error;
2343
2344                 iommu_set_root_entry(iommu);
2345
2346                 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
2347                 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2348                 iommu_disable_protect_mem_regions(iommu);
2349
2350                 ret = iommu_enable_translation(iommu);
2351                 if (ret)
2352                         goto error;
2353         }
2354
2355         return 0;
2356 error:
2357         for_each_drhd_unit(drhd) {
2358                 if (drhd->ignored)
2359                         continue;
2360                 iommu = drhd->iommu;
2361                 free_iommu(iommu);
2362         }
2363         kfree(g_iommus);
2364         return ret;
2365 }
2366
2367 static inline unsigned long aligned_nrpages(unsigned long host_addr,
2368                                             size_t size)
2369 {
2370         host_addr &= ~PAGE_MASK;
2371         host_addr += size + PAGE_SIZE - 1;
2372
2373         return host_addr >> VTD_PAGE_SHIFT;
2374 }
2375
2376 static struct iova *intel_alloc_iova(struct device *dev,
2377                                      struct dmar_domain *domain,
2378                                      unsigned long nrpages, uint64_t dma_mask)
2379 {
2380         struct pci_dev *pdev = to_pci_dev(dev);
2381         struct iova *iova = NULL;
2382
2383         /* Restrict dma_mask to the width that the iommu can handle */
2384         dma_mask = min_t(uint64_t, DOMAIN_MAX_ADDR(domain->gaw), dma_mask);
2385
2386         if (!dmar_forcedac && dma_mask > DMA_BIT_MASK(32)) {
2387                 /*
2388                  * First try to allocate an io virtual address in
2389                  * DMA_BIT_MASK(32) and if that fails then try allocating
2390                  * from higher range
2391                  */
2392                 iova = alloc_iova(&domain->iovad, nrpages,
2393                                   IOVA_PFN(DMA_BIT_MASK(32)), 1);
2394                 if (iova)
2395                         return iova;
2396         }
2397         iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
2398         if (unlikely(!iova)) {
2399                 printk(KERN_ERR "Allocating %ld-page iova for %s failed",
2400                        nrpages, pci_name(pdev));
2401                 return NULL;
2402         }
2403
2404         return iova;
2405 }
2406
2407 static struct dmar_domain *
2408 get_valid_domain_for_dev(struct pci_dev *pdev)
2409 {
2410         struct dmar_domain *domain;
2411         int ret;
2412
2413         domain = get_domain_for_dev(pdev,
2414                         DEFAULT_DOMAIN_ADDRESS_WIDTH);
2415         if (!domain) {
2416                 printk(KERN_ERR
2417                         "Allocating domain for %s failed", pci_name(pdev));
2418                 return NULL;
2419         }
2420
2421         /* make sure context mapping is ok */
2422         if (unlikely(!domain_context_mapped(pdev))) {
2423                 ret = domain_context_mapping(domain, pdev,
2424                                              CONTEXT_TT_MULTI_LEVEL);
2425                 if (ret) {
2426                         printk(KERN_ERR
2427                                 "Domain context map for %s failed",
2428                                 pci_name(pdev));
2429                         return NULL;
2430                 }
2431         }
2432
2433         return domain;
2434 }
2435
2436 static int iommu_dummy(struct pci_dev *pdev)
2437 {
2438         return pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
2439 }
2440
2441 /* Check if the pdev needs to go through non-identity map and unmap process.*/
2442 static int iommu_no_mapping(struct pci_dev *pdev)
2443 {
2444         int found;
2445
2446         if (!iommu_identity_mapping)
2447                 return iommu_dummy(pdev);
2448
2449         found = identity_mapping(pdev);
2450         if (found) {
2451                 if (pdev->dma_mask > DMA_BIT_MASK(32))
2452                         return 1;
2453                 else {
2454                         /*
2455                          * 32 bit DMA is removed from si_domain and fall back
2456                          * to non-identity mapping.
2457                          */
2458                         domain_remove_one_dev_info(si_domain, pdev);
2459                         printk(KERN_INFO "32bit %s uses non-identity mapping\n",
2460                                pci_name(pdev));
2461                         return 0;
2462                 }
2463         } else {
2464                 /*
2465                  * In case of a detached 64 bit DMA device from vm, the device
2466                  * is put into si_domain for identity mapping.
2467                  */
2468                 if (pdev->dma_mask > DMA_BIT_MASK(32)) {
2469                         int ret;
2470                         ret = domain_add_dev_info(si_domain, pdev);
2471                         if (!ret) {
2472                                 printk(KERN_INFO "64bit %s uses identity mapping\n",
2473                                        pci_name(pdev));
2474                                 return 1;
2475                         }
2476                 }
2477         }
2478
2479         return iommu_dummy(pdev);
2480 }
2481
2482 static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
2483                                      size_t size, int dir, u64 dma_mask)
2484 {
2485         struct pci_dev *pdev = to_pci_dev(hwdev);
2486         struct dmar_domain *domain;
2487         phys_addr_t start_paddr;
2488         struct iova *iova;
2489         int prot = 0;
2490         int ret;
2491         struct intel_iommu *iommu;
2492
2493         BUG_ON(dir == DMA_NONE);
2494
2495         if (iommu_no_mapping(pdev))
2496                 return paddr;
2497
2498         domain = get_valid_domain_for_dev(pdev);
2499         if (!domain)
2500                 return 0;
2501
2502         iommu = domain_get_iommu(domain);
2503         size = aligned_nrpages(paddr, size);
2504
2505         iova = intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2506         if (!iova)
2507                 goto error;
2508
2509         /*
2510          * Check if DMAR supports zero-length reads on write only
2511          * mappings..
2512          */
2513         if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2514                         !cap_zlr(iommu->cap))
2515                 prot |= DMA_PTE_READ;
2516         if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2517                 prot |= DMA_PTE_WRITE;
2518         /*
2519          * paddr - (paddr + size) might be partial page, we should map the whole
2520          * page.  Note: if two part of one page are separately mapped, we
2521          * might have two guest_addr mapping to the same host paddr, but this
2522          * is not a big problem
2523          */
2524         ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
2525                                  paddr >> VTD_PAGE_SHIFT, size, prot);
2526         if (ret)
2527                 goto error;
2528
2529         /* it's a non-present to present mapping. Only flush if caching mode */
2530         if (cap_caching_mode(iommu->cap))
2531                 iommu_flush_iotlb_psi(iommu, 0, mm_to_dma_pfn(iova->pfn_lo), size);
2532         else
2533                 iommu_flush_write_buffer(iommu);
2534
2535         start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
2536         start_paddr += paddr & ~PAGE_MASK;
2537         return start_paddr;
2538
2539 error:
2540         if (iova)
2541                 __free_iova(&domain->iovad, iova);
2542         printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
2543                 pci_name(pdev), size, (unsigned long long)paddr, dir);
2544         return 0;
2545 }
2546
2547 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
2548                                  unsigned long offset, size_t size,
2549                                  enum dma_data_direction dir,
2550                                  struct dma_attrs *attrs)
2551 {
2552         return __intel_map_single(dev, page_to_phys(page) + offset, size,
2553                                   dir, to_pci_dev(dev)->dma_mask);
2554 }
2555
2556 static void flush_unmaps(void)
2557 {
2558         int i, j;
2559
2560         timer_on = 0;
2561
2562         /* just flush them all */
2563         for (i = 0; i < g_num_of_iommus; i++) {
2564                 struct intel_iommu *iommu = g_iommus[i];
2565                 if (!iommu)
2566                         continue;
2567
2568                 if (!deferred_flush[i].next)
2569                         continue;
2570
2571                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2572                                          DMA_TLB_GLOBAL_FLUSH);
2573                 for (j = 0; j < deferred_flush[i].next; j++) {
2574                         unsigned long mask;
2575                         struct iova *iova = deferred_flush[i].iova[j];
2576
2577                         mask = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT;
2578                         mask = ilog2(mask >> VTD_PAGE_SHIFT);
2579                         iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
2580                                         iova->pfn_lo << PAGE_SHIFT, mask);
2581                         __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
2582                 }
2583                 deferred_flush[i].next = 0;
2584         }
2585
2586         list_size = 0;
2587 }
2588
2589 static void flush_unmaps_timeout(unsigned long data)
2590 {
2591         unsigned long flags;
2592
2593         spin_lock_irqsave(&async_umap_flush_lock, flags);
2594         flush_unmaps();
2595         spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2596 }
2597
2598 static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2599 {
2600         unsigned long flags;
2601         int next, iommu_id;
2602         struct intel_iommu *iommu;
2603
2604         spin_lock_irqsave(&async_umap_flush_lock, flags);
2605         if (list_size == HIGH_WATER_MARK)
2606                 flush_unmaps();
2607
2608         iommu = domain_get_iommu(dom);
2609         iommu_id = iommu->seq_id;
2610
2611         next = deferred_flush[iommu_id].next;
2612         deferred_flush[iommu_id].domain[next] = dom;
2613         deferred_flush[iommu_id].iova[next] = iova;
2614         deferred_flush[iommu_id].next++;
2615
2616         if (!timer_on) {
2617                 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2618                 timer_on = 1;
2619         }
2620         list_size++;
2621         spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2622 }
2623
2624 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
2625                              size_t size, enum dma_data_direction dir,
2626                              struct dma_attrs *attrs)
2627 {
2628         struct pci_dev *pdev = to_pci_dev(dev);
2629         struct dmar_domain *domain;
2630         unsigned long start_pfn, last_pfn;
2631         struct iova *iova;
2632         struct intel_iommu *iommu;
2633
2634         if (iommu_no_mapping(pdev))
2635                 return;
2636
2637         domain = find_domain(pdev);
2638         BUG_ON(!domain);
2639
2640         iommu = domain_get_iommu(domain);
2641
2642         iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
2643         if (!iova)
2644                 return;
2645
2646         start_pfn = mm_to_dma_pfn(iova->pfn_lo);
2647         last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
2648
2649         pr_debug("Device %s unmapping: pfn %lx-%lx\n",
2650                  pci_name(pdev), start_pfn, last_pfn);
2651
2652         /*  clear the whole page */
2653         dma_pte_clear_range(domain, start_pfn, last_pfn);
2654
2655         /* free page tables */
2656         dma_pte_free_pagetable(domain, start_pfn, last_pfn);
2657
2658         if (intel_iommu_strict) {
2659                 iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
2660                                       last_pfn - start_pfn + 1);
2661                 /* free iova */
2662                 __free_iova(&domain->iovad, iova);
2663         } else {
2664                 add_unmap(domain, iova);
2665                 /*
2666                  * queue up the release of the unmap to save the 1/6th of the
2667                  * cpu used up by the iotlb flush operation...
2668                  */
2669         }
2670 }
2671
2672 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr, size_t size,
2673                                int dir)
2674 {
2675         intel_unmap_page(dev, dev_addr, size, dir, NULL);
2676 }
2677
2678 static void *intel_alloc_coherent(struct device *hwdev, size_t size,
2679                                   dma_addr_t *dma_handle, gfp_t flags)
2680 {
2681         void *vaddr;
2682         int order;
2683
2684         size = PAGE_ALIGN(size);
2685         order = get_order(size);
2686         flags &= ~(GFP_DMA | GFP_DMA32);
2687
2688         vaddr = (void *)__get_free_pages(flags, order);
2689         if (!vaddr)
2690                 return NULL;
2691         memset(vaddr, 0, size);
2692
2693         *dma_handle = __intel_map_single(hwdev, virt_to_bus(vaddr), size,
2694                                          DMA_BIDIRECTIONAL,
2695                                          hwdev->coherent_dma_mask);
2696         if (*dma_handle)
2697                 return vaddr;
2698         free_pages((unsigned long)vaddr, order);
2699         return NULL;
2700 }
2701
2702 static void intel_free_coherent(struct device *hwdev, size_t size, void *vaddr,
2703                                 dma_addr_t dma_handle)
2704 {
2705         int order;
2706
2707         size = PAGE_ALIGN(size);
2708         order = get_order(size);
2709
2710         intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2711         free_pages((unsigned long)vaddr, order);
2712 }
2713
2714 static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2715                            int nelems, enum dma_data_direction dir,
2716                            struct dma_attrs *attrs)
2717 {
2718         struct pci_dev *pdev = to_pci_dev(hwdev);
2719         struct dmar_domain *domain;
2720         unsigned long start_pfn, last_pfn;
2721         struct iova *iova;
2722         struct intel_iommu *iommu;
2723
2724         if (iommu_no_mapping(pdev))
2725                 return;
2726
2727         domain = find_domain(pdev);
2728         BUG_ON(!domain);
2729
2730         iommu = domain_get_iommu(domain);
2731
2732         iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
2733         if (!iova)
2734                 return;
2735
2736         start_pfn = mm_to_dma_pfn(iova->pfn_lo);
2737         last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
2738
2739         /*  clear the whole page */
2740         dma_pte_clear_range(domain, start_pfn, last_pfn);
2741
2742         /* free page tables */
2743         dma_pte_free_pagetable(domain, start_pfn, last_pfn);
2744
2745         iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
2746                               (last_pfn - start_pfn + 1));
2747
2748         /* free iova */
2749         __free_iova(&domain->iovad, iova);
2750 }
2751
2752 static int intel_nontranslate_map_sg(struct device *hddev,
2753         struct scatterlist *sglist, int nelems, int dir)
2754 {
2755         int i;
2756         struct scatterlist *sg;
2757
2758         for_each_sg(sglist, sg, nelems, i) {
2759                 BUG_ON(!sg_page(sg));
2760                 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
2761                 sg->dma_length = sg->length;
2762         }
2763         return nelems;
2764 }
2765
2766 static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist, int nelems,
2767                         enum dma_data_direction dir, struct dma_attrs *attrs)
2768 {
2769         int i;
2770         struct pci_dev *pdev = to_pci_dev(hwdev);
2771         struct dmar_domain *domain;
2772         size_t size = 0;
2773         int prot = 0;
2774         size_t offset_pfn = 0;
2775         struct iova *iova = NULL;
2776         int ret;
2777         struct scatterlist *sg;
2778         unsigned long start_vpfn;
2779         struct intel_iommu *iommu;
2780
2781         BUG_ON(dir == DMA_NONE);
2782         if (iommu_no_mapping(pdev))
2783                 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
2784
2785         domain = get_valid_domain_for_dev(pdev);
2786         if (!domain)
2787                 return 0;
2788
2789         iommu = domain_get_iommu(domain);
2790
2791         for_each_sg(sglist, sg, nelems, i)
2792                 size += aligned_nrpages(sg->offset, sg->length);
2793
2794         iova = intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2795         if (!iova) {
2796                 sglist->dma_length = 0;
2797                 return 0;
2798         }
2799
2800         /*
2801          * Check if DMAR supports zero-length reads on write only
2802          * mappings..
2803          */
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;
2809
2810         start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
2811
2812         ret = domain_sg_mapping(domain, start_vpfn, sglist, mm_to_dma_pfn(size), prot);
2813         if (unlikely(ret)) {
2814                 /*  clear the page */
2815                 dma_pte_clear_range(domain, start_vpfn,
2816                                     start_vpfn + size - 1);
2817                 /* free page tables */
2818                 dma_pte_free_pagetable(domain, start_vpfn,
2819                                        start_vpfn + size - 1);
2820                 /* free iova */
2821                 __free_iova(&domain->iovad, iova);
2822                 return 0;
2823         }
2824
2825         /* it's a non-present to present mapping. Only flush if caching mode */
2826         if (cap_caching_mode(iommu->cap))
2827                 iommu_flush_iotlb_psi(iommu, 0, start_vpfn, offset_pfn);
2828         else
2829                 iommu_flush_write_buffer(iommu);
2830
2831         return nelems;
2832 }
2833
2834 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
2835 {
2836         return !dma_addr;
2837 }
2838
2839 struct dma_map_ops intel_dma_ops = {
2840         .alloc_coherent = intel_alloc_coherent,
2841         .free_coherent = intel_free_coherent,
2842         .map_sg = intel_map_sg,
2843         .unmap_sg = intel_unmap_sg,
2844         .map_page = intel_map_page,
2845         .unmap_page = intel_unmap_page,
2846         .mapping_error = intel_mapping_error,
2847 };
2848
2849 static inline int iommu_domain_cache_init(void)
2850 {
2851         int ret = 0;
2852
2853         iommu_domain_cache = kmem_cache_create("iommu_domain",
2854                                          sizeof(struct dmar_domain),
2855                                          0,
2856                                          SLAB_HWCACHE_ALIGN,
2857
2858                                          NULL);
2859         if (!iommu_domain_cache) {
2860                 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2861                 ret = -ENOMEM;
2862         }
2863
2864         return ret;
2865 }
2866
2867 static inline int iommu_devinfo_cache_init(void)
2868 {
2869         int ret = 0;
2870
2871         iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2872                                          sizeof(struct device_domain_info),
2873                                          0,
2874                                          SLAB_HWCACHE_ALIGN,
2875                                          NULL);
2876         if (!iommu_devinfo_cache) {
2877                 printk(KERN_ERR "Couldn't create devinfo cache\n");
2878                 ret = -ENOMEM;
2879         }
2880
2881         return ret;
2882 }
2883
2884 static inline int iommu_iova_cache_init(void)
2885 {
2886         int ret = 0;
2887
2888         iommu_iova_cache = kmem_cache_create("iommu_iova",
2889                                          sizeof(struct iova),
2890                                          0,
2891                                          SLAB_HWCACHE_ALIGN,
2892                                          NULL);
2893         if (!iommu_iova_cache) {
2894                 printk(KERN_ERR "Couldn't create iova cache\n");
2895                 ret = -ENOMEM;
2896         }
2897
2898         return ret;
2899 }
2900
2901 static int __init iommu_init_mempool(void)
2902 {
2903         int ret;
2904         ret = iommu_iova_cache_init();
2905         if (ret)
2906                 return ret;
2907
2908         ret = iommu_domain_cache_init();
2909         if (ret)
2910                 goto domain_error;
2911
2912         ret = iommu_devinfo_cache_init();
2913         if (!ret)
2914                 return ret;
2915
2916         kmem_cache_destroy(iommu_domain_cache);
2917 domain_error:
2918         kmem_cache_destroy(iommu_iova_cache);
2919
2920         return -ENOMEM;
2921 }
2922
2923 static void __init iommu_exit_mempool(void)
2924 {
2925         kmem_cache_destroy(iommu_devinfo_cache);
2926         kmem_cache_destroy(iommu_domain_cache);
2927         kmem_cache_destroy(iommu_iova_cache);
2928
2929 }
2930
2931 static void __init init_no_remapping_devices(void)
2932 {
2933         struct dmar_drhd_unit *drhd;
2934
2935         for_each_drhd_unit(drhd) {
2936                 if (!drhd->include_all) {
2937                         int i;
2938                         for (i = 0; i < drhd->devices_cnt; i++)
2939                                 if (drhd->devices[i] != NULL)
2940                                         break;
2941                         /* ignore DMAR unit if no pci devices exist */
2942                         if (i == drhd->devices_cnt)
2943                                 drhd->ignored = 1;
2944                 }
2945         }
2946
2947         if (dmar_map_gfx)
2948                 return;
2949
2950         for_each_drhd_unit(drhd) {
2951                 int i;
2952                 if (drhd->ignored || drhd->include_all)
2953                         continue;
2954
2955                 for (i = 0; i < drhd->devices_cnt; i++)
2956                         if (drhd->devices[i] &&
2957                                 !IS_GFX_DEVICE(drhd->devices[i]))
2958                                 break;
2959
2960                 if (i < drhd->devices_cnt)
2961                         continue;
2962
2963                 /* bypass IOMMU if it is just for gfx devices */
2964                 drhd->ignored = 1;
2965                 for (i = 0; i < drhd->devices_cnt; i++) {
2966                         if (!drhd->devices[i])
2967                                 continue;
2968                         drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
2969                 }
2970         }
2971 }
2972
2973 #ifdef CONFIG_SUSPEND
2974 static int init_iommu_hw(void)
2975 {
2976         struct dmar_drhd_unit *drhd;
2977         struct intel_iommu *iommu = NULL;
2978
2979         for_each_active_iommu(iommu, drhd)
2980                 if (iommu->qi)
2981                         dmar_reenable_qi(iommu);
2982
2983         for_each_active_iommu(iommu, drhd) {
2984                 iommu_flush_write_buffer(iommu);
2985
2986                 iommu_set_root_entry(iommu);
2987
2988                 iommu->flush.flush_context(iommu, 0, 0, 0,
2989                                            DMA_CCMD_GLOBAL_INVL);
2990                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2991                                          DMA_TLB_GLOBAL_FLUSH);
2992                 iommu_disable_protect_mem_regions(iommu);
2993                 iommu_enable_translation(iommu);
2994         }
2995
2996         return 0;
2997 }
2998
2999 static void iommu_flush_all(void)
3000 {
3001         struct dmar_drhd_unit *drhd;
3002         struct intel_iommu *iommu;
3003
3004         for_each_active_iommu(iommu, drhd) {
3005                 iommu->flush.flush_context(iommu, 0, 0, 0,
3006                                            DMA_CCMD_GLOBAL_INVL);
3007                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3008                                          DMA_TLB_GLOBAL_FLUSH);
3009         }
3010 }
3011
3012 static int iommu_suspend(struct sys_device *dev, pm_message_t state)
3013 {
3014         struct dmar_drhd_unit *drhd;
3015         struct intel_iommu *iommu = NULL;
3016         unsigned long flag;
3017
3018         for_each_active_iommu(iommu, drhd) {
3019                 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3020                                                  GFP_ATOMIC);
3021                 if (!iommu->iommu_state)
3022                         goto nomem;
3023         }
3024
3025         iommu_flush_all();
3026
3027         for_each_active_iommu(iommu, drhd) {
3028                 iommu_disable_translation(iommu);
3029
3030                 spin_lock_irqsave(&iommu->register_lock, flag);
3031
3032                 iommu->iommu_state[SR_DMAR_FECTL_REG] =
3033                         readl(iommu->reg + DMAR_FECTL_REG);
3034                 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
3035                         readl(iommu->reg + DMAR_FEDATA_REG);
3036                 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
3037                         readl(iommu->reg + DMAR_FEADDR_REG);
3038                 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
3039                         readl(iommu->reg + DMAR_FEUADDR_REG);
3040
3041                 spin_unlock_irqrestore(&iommu->register_lock, flag);
3042         }
3043         return 0;
3044
3045 nomem:
3046         for_each_active_iommu(iommu, drhd)
3047                 kfree(iommu->iommu_state);
3048
3049         return -ENOMEM;
3050 }
3051
3052 static int iommu_resume(struct sys_device *dev)
3053 {
3054         struct dmar_drhd_unit *drhd;
3055         struct intel_iommu *iommu = NULL;
3056         unsigned long flag;
3057
3058         if (init_iommu_hw()) {
3059                 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
3060                 return -EIO;
3061         }
3062
3063         for_each_active_iommu(iommu, drhd) {
3064
3065                 spin_lock_irqsave(&iommu->register_lock, flag);
3066
3067                 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
3068                         iommu->reg + DMAR_FECTL_REG);
3069                 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
3070                         iommu->reg + DMAR_FEDATA_REG);
3071                 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
3072                         iommu->reg + DMAR_FEADDR_REG);
3073                 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
3074                         iommu->reg + DMAR_FEUADDR_REG);
3075
3076                 spin_unlock_irqrestore(&iommu->register_lock, flag);
3077         }
3078
3079         for_each_active_iommu(iommu, drhd)
3080                 kfree(iommu->iommu_state);
3081
3082         return 0;
3083 }
3084
3085 static struct sysdev_class iommu_sysclass = {
3086         .name           = "iommu",
3087         .resume         = iommu_resume,
3088         .suspend        = iommu_suspend,
3089 };
3090
3091 static struct sys_device device_iommu = {
3092         .cls    = &iommu_sysclass,
3093 };
3094
3095 static int __init init_iommu_sysfs(void)
3096 {
3097         int error;
3098
3099         error = sysdev_class_register(&iommu_sysclass);
3100         if (error)
3101                 return error;
3102
3103         error = sysdev_register(&device_iommu);
3104         if (error)
3105                 sysdev_class_unregister(&iommu_sysclass);
3106
3107         return error;
3108 }
3109
3110 #else
3111 static int __init init_iommu_sysfs(void)
3112 {
3113         return 0;
3114 }
3115 #endif  /* CONFIG_PM */
3116
3117 int __init intel_iommu_init(void)
3118 {
3119         int ret = 0;
3120
3121         if (dmar_table_init())
3122                 return  -ENODEV;
3123
3124         if (dmar_dev_scope_init())
3125                 return  -ENODEV;
3126
3127         /*
3128          * Check the need for DMA-remapping initialization now.
3129          * Above initialization will also be used by Interrupt-remapping.
3130          */
3131         if (no_iommu || (swiotlb && !iommu_pass_through) || dmar_disabled)
3132                 return -ENODEV;
3133
3134         iommu_init_mempool();
3135         dmar_init_reserved_ranges();
3136
3137         init_no_remapping_devices();
3138
3139         ret = init_dmars();
3140         if (ret) {
3141                 printk(KERN_ERR "IOMMU: dmar init failed\n");
3142                 put_iova_domain(&reserved_iova_list);
3143                 iommu_exit_mempool();
3144                 return ret;
3145         }
3146         printk(KERN_INFO
3147         "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
3148
3149         init_timer(&unmap_timer);
3150         force_iommu = 1;
3151
3152         if (!iommu_pass_through) {
3153                 printk(KERN_INFO
3154                        "Multi-level page-table translation for DMAR.\n");
3155                 dma_ops = &intel_dma_ops;
3156         } else
3157                 printk(KERN_INFO
3158                        "DMAR: Pass through translation for DMAR.\n");
3159
3160         init_iommu_sysfs();
3161
3162         register_iommu(&intel_iommu_ops);
3163
3164         return 0;
3165 }
3166
3167 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
3168                                            struct pci_dev *pdev)
3169 {
3170         struct pci_dev *tmp, *parent;
3171
3172         if (!iommu || !pdev)
3173                 return;
3174
3175         /* dependent device detach */
3176         tmp = pci_find_upstream_pcie_bridge(pdev);
3177         /* Secondary interface's bus number and devfn 0 */
3178         if (tmp) {
3179                 parent = pdev->bus->self;
3180                 while (parent != tmp) {
3181                         iommu_detach_dev(iommu, parent->bus->number,
3182                                          parent->devfn);
3183                         parent = parent->bus->self;
3184                 }
3185                 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
3186                         iommu_detach_dev(iommu,
3187                                 tmp->subordinate->number, 0);
3188                 else /* this is a legacy PCI bridge */
3189                         iommu_detach_dev(iommu, tmp->bus->number,
3190                                          tmp->devfn);
3191         }
3192 }
3193
3194 static void domain_remove_one_dev_info(struct dmar_domain *domain,
3195                                           struct pci_dev *pdev)
3196 {
3197         struct device_domain_info *info;
3198         struct intel_iommu *iommu;
3199         unsigned long flags;
3200         int found = 0;
3201         struct list_head *entry, *tmp;
3202
3203         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3204                                 pdev->devfn);
3205         if (!iommu)
3206                 return;
3207
3208         spin_lock_irqsave(&device_domain_lock, flags);
3209         list_for_each_safe(entry, tmp, &domain->devices) {
3210                 info = list_entry(entry, struct device_domain_info, link);
3211                 /* No need to compare PCI domain; it has to be the same */
3212                 if (info->bus == pdev->bus->number &&
3213                     info->devfn == pdev->devfn) {
3214                         list_del(&info->link);
3215                         list_del(&info->global);
3216                         if (info->dev)
3217                                 info->dev->dev.archdata.iommu = NULL;
3218                         spin_unlock_irqrestore(&device_domain_lock, flags);
3219
3220                         iommu_disable_dev_iotlb(info);
3221                         iommu_detach_dev(iommu, info->bus, info->devfn);
3222                         iommu_detach_dependent_devices(iommu, pdev);
3223                         free_devinfo_mem(info);
3224
3225                         spin_lock_irqsave(&device_domain_lock, flags);
3226
3227                         if (found)
3228                                 break;
3229                         else
3230                                 continue;
3231                 }
3232
3233                 /* if there is no other devices under the same iommu
3234                  * owned by this domain, clear this iommu in iommu_bmp
3235                  * update iommu count and coherency
3236                  */
3237                 if (iommu == device_to_iommu(info->segment, info->bus,
3238                                             info->devfn))
3239                         found = 1;
3240         }
3241
3242         if (found == 0) {
3243                 unsigned long tmp_flags;
3244                 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
3245                 clear_bit(iommu->seq_id, &domain->iommu_bmp);
3246                 domain->iommu_count--;
3247                 domain_update_iommu_cap(domain);
3248                 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
3249         }
3250
3251         spin_unlock_irqrestore(&device_domain_lock, flags);
3252 }
3253
3254 static void vm_domain_remove_all_dev_info(struct dmar_domain *domain)
3255 {
3256         struct device_domain_info *info;
3257         struct intel_iommu *iommu;
3258         unsigned long flags1, flags2;
3259
3260         spin_lock_irqsave(&device_domain_lock, flags1);
3261         while (!list_empty(&domain->devices)) {
3262                 info = list_entry(domain->devices.next,
3263                         struct device_domain_info, link);
3264                 list_del(&info->link);
3265                 list_del(&info->global);
3266                 if (info->dev)
3267                         info->dev->dev.archdata.iommu = NULL;
3268
3269                 spin_unlock_irqrestore(&device_domain_lock, flags1);
3270
3271                 iommu_disable_dev_iotlb(info);
3272                 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
3273                 iommu_detach_dev(iommu, info->bus, info->devfn);
3274                 iommu_detach_dependent_devices(iommu, info->dev);
3275
3276                 /* clear this iommu in iommu_bmp, update iommu count
3277                  * and capabilities
3278                  */
3279                 spin_lock_irqsave(&domain->iommu_lock, flags2);
3280                 if (test_and_clear_bit(iommu->seq_id,
3281                                        &domain->iommu_bmp)) {
3282                         domain->iommu_count--;
3283                         domain_update_iommu_cap(domain);
3284                 }
3285                 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
3286
3287                 free_devinfo_mem(info);
3288                 spin_lock_irqsave(&device_domain_lock, flags1);
3289         }
3290         spin_unlock_irqrestore(&device_domain_lock, flags1);
3291 }
3292
3293 /* domain id for virtual machine, it won't be set in context */
3294 static unsigned long vm_domid;
3295
3296 static int vm_domain_min_agaw(struct dmar_domain *domain)
3297 {
3298         int i;
3299         int min_agaw = domain->agaw;
3300
3301         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
3302         for (; i < g_num_of_iommus; ) {
3303                 if (min_agaw > g_iommus[i]->agaw)
3304                         min_agaw = g_iommus[i]->agaw;
3305
3306                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
3307         }
3308
3309         return min_agaw;
3310 }
3311
3312 static struct dmar_domain *iommu_alloc_vm_domain(void)
3313 {
3314         struct dmar_domain *domain;
3315
3316         domain = alloc_domain_mem();
3317         if (!domain)
3318                 return NULL;
3319
3320         domain->id = vm_domid++;
3321         memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
3322         domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
3323
3324         return domain;
3325 }
3326
3327 static int md_domain_init(struct dmar_domain *domain, int guest_width)
3328 {
3329         int adjust_width;
3330
3331         init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
3332         spin_lock_init(&domain->mapping_lock);
3333         spin_lock_init(&domain->iommu_lock);
3334
3335         domain_reserve_special_ranges(domain);
3336
3337         /* calculate AGAW */
3338         domain->gaw = guest_width;
3339         adjust_width = guestwidth_to_adjustwidth(guest_width);
3340         domain->agaw = width_to_agaw(adjust_width);
3341
3342         INIT_LIST_HEAD(&domain->devices);
3343
3344         domain->iommu_count = 0;
3345         domain->iommu_coherency = 0;
3346         domain->max_addr = 0;
3347
3348         /* always allocate the top pgd */
3349         domain->pgd = (struct dma_pte *)alloc_pgtable_page();
3350         if (!domain->pgd)
3351                 return -ENOMEM;
3352         domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
3353         return 0;
3354 }
3355
3356 static void iommu_free_vm_domain(struct dmar_domain *domain)
3357 {
3358         unsigned long flags;
3359         struct dmar_drhd_unit *drhd;
3360         struct intel_iommu *iommu;
3361         unsigned long i;
3362         unsigned long ndomains;
3363
3364         for_each_drhd_unit(drhd) {
3365                 if (drhd->ignored)
3366                         continue;
3367                 iommu = drhd->iommu;
3368
3369                 ndomains = cap_ndoms(iommu->cap);
3370                 i = find_first_bit(iommu->domain_ids, ndomains);
3371                 for (; i < ndomains; ) {
3372                         if (iommu->domains[i] == domain) {
3373                                 spin_lock_irqsave(&iommu->lock, flags);
3374                                 clear_bit(i, iommu->domain_ids);
3375                                 iommu->domains[i] = NULL;
3376                                 spin_unlock_irqrestore(&iommu->lock, flags);
3377                                 break;
3378                         }
3379                         i = find_next_bit(iommu->domain_ids, ndomains, i+1);
3380                 }
3381         }
3382 }
3383
3384 static void vm_domain_exit(struct dmar_domain *domain)
3385 {
3386         /* Domain 0 is reserved, so dont process it */
3387         if (!domain)
3388                 return;
3389
3390         vm_domain_remove_all_dev_info(domain);
3391         /* destroy iovas */
3392         put_iova_domain(&domain->iovad);
3393
3394         /* clear ptes */
3395         dma_pte_clear_range(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
3396
3397         /* free page tables */
3398         dma_pte_free_pagetable(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
3399
3400         iommu_free_vm_domain(domain);
3401         free_domain_mem(domain);
3402 }
3403
3404 static int intel_iommu_domain_init(struct iommu_domain *domain)
3405 {
3406         struct dmar_domain *dmar_domain;
3407
3408         dmar_domain = iommu_alloc_vm_domain();
3409         if (!dmar_domain) {
3410                 printk(KERN_ERR
3411                         "intel_iommu_domain_init: dmar_domain == NULL\n");
3412                 return -ENOMEM;
3413         }
3414         if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
3415                 printk(KERN_ERR
3416                         "intel_iommu_domain_init() failed\n");
3417                 vm_domain_exit(dmar_domain);
3418                 return -ENOMEM;
3419         }
3420         domain->priv = dmar_domain;
3421
3422         return 0;
3423 }
3424
3425 static void intel_iommu_domain_destroy(struct iommu_domain *domain)
3426 {
3427         struct dmar_domain *dmar_domain = domain->priv;
3428
3429         domain->priv = NULL;
3430         vm_domain_exit(dmar_domain);
3431 }
3432
3433 static int intel_iommu_attach_device(struct iommu_domain *domain,
3434                                      struct device *dev)
3435 {
3436         struct dmar_domain *dmar_domain = domain->priv;
3437         struct pci_dev *pdev = to_pci_dev(dev);
3438         struct intel_iommu *iommu;
3439         int addr_width;
3440         u64 end;
3441         int ret;
3442
3443         /* normally pdev is not mapped */
3444         if (unlikely(domain_context_mapped(pdev))) {
3445                 struct dmar_domain *old_domain;
3446
3447                 old_domain = find_domain(pdev);
3448                 if (old_domain) {
3449                         if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
3450                             dmar_domain->flags & DOMAIN_FLAG_STATIC_IDENTITY)
3451                                 domain_remove_one_dev_info(old_domain, pdev);
3452                         else
3453                                 domain_remove_dev_info(old_domain);
3454                 }
3455         }
3456
3457         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3458                                 pdev->devfn);
3459         if (!iommu)
3460                 return -ENODEV;
3461
3462         /* check if this iommu agaw is sufficient for max mapped address */
3463         addr_width = agaw_to_width(iommu->agaw);
3464         end = DOMAIN_MAX_ADDR(addr_width);
3465         end = end & VTD_PAGE_MASK;
3466         if (end < dmar_domain->max_addr) {
3467                 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3468                        "sufficient for the mapped address (%llx)\n",
3469                        __func__, iommu->agaw, dmar_domain->max_addr);
3470                 return -EFAULT;
3471         }
3472
3473         ret = domain_add_dev_info(dmar_domain, pdev);
3474         if (ret)
3475                 return ret;
3476
3477         ret = domain_context_mapping(dmar_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
3478         return ret;
3479 }
3480
3481 static void intel_iommu_detach_device(struct iommu_domain *domain,
3482                                       struct device *dev)
3483 {
3484         struct dmar_domain *dmar_domain = domain->priv;
3485         struct pci_dev *pdev = to_pci_dev(dev);
3486
3487         domain_remove_one_dev_info(dmar_domain, pdev);
3488 }
3489
3490 static int intel_iommu_map_range(struct iommu_domain *domain,
3491                                  unsigned long iova, phys_addr_t hpa,
3492                                  size_t size, int iommu_prot)
3493 {
3494         struct dmar_domain *dmar_domain = domain->priv;
3495         u64 max_addr;
3496         int addr_width;
3497         int prot = 0;
3498         int ret;
3499
3500         if (iommu_prot & IOMMU_READ)
3501                 prot |= DMA_PTE_READ;
3502         if (iommu_prot & IOMMU_WRITE)
3503                 prot |= DMA_PTE_WRITE;
3504         if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
3505                 prot |= DMA_PTE_SNP;
3506
3507         max_addr = iova + size;
3508         if (dmar_domain->max_addr < max_addr) {
3509                 int min_agaw;
3510                 u64 end;
3511
3512                 /* check if minimum agaw is sufficient for mapped address */
3513                 min_agaw = vm_domain_min_agaw(dmar_domain);
3514                 addr_width = agaw_to_width(min_agaw);
3515                 end = DOMAIN_MAX_ADDR(addr_width);
3516                 end = end & VTD_PAGE_MASK;
3517                 if (end < max_addr) {
3518                         printk(KERN_ERR "%s: iommu agaw (%d) is not "
3519                                "sufficient for the mapped address (%llx)\n",
3520                                __func__, min_agaw, max_addr);
3521                         return -EFAULT;
3522                 }
3523                 dmar_domain->max_addr = max_addr;
3524         }
3525         /* Round up size to next multiple of PAGE_SIZE, if it and
3526            the low bits of hpa would take us onto the next page */
3527         size = aligned_nrpages(hpa, size);
3528         ret = domain_pfn_mapping(dmar_domain, iova >> VTD_PAGE_SHIFT,
3529                                  hpa >> VTD_PAGE_SHIFT, size, prot);
3530         return ret;
3531 }
3532
3533 static void intel_iommu_unmap_range(struct iommu_domain *domain,
3534                                     unsigned long iova, size_t size)
3535 {
3536         struct dmar_domain *dmar_domain = domain->priv;
3537
3538         dma_pte_clear_range(dmar_domain, iova >> VTD_PAGE_SHIFT,
3539                             (iova + size - 1) >> VTD_PAGE_SHIFT);
3540
3541         if (dmar_domain->max_addr == iova + size)
3542                 dmar_domain->max_addr = iova;
3543 }
3544
3545 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
3546                                             unsigned long iova)
3547 {
3548         struct dmar_domain *dmar_domain = domain->priv;
3549         struct dma_pte *pte;
3550         u64 phys = 0;
3551
3552         pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT);
3553         if (pte)
3554                 phys = dma_pte_addr(pte);
3555
3556         return phys;
3557 }
3558
3559 static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
3560                                       unsigned long cap)
3561 {
3562         struct dmar_domain *dmar_domain = domain->priv;
3563
3564         if (cap == IOMMU_CAP_CACHE_COHERENCY)
3565                 return dmar_domain->iommu_snooping;
3566
3567         return 0;
3568 }
3569
3570 static struct iommu_ops intel_iommu_ops = {
3571         .domain_init    = intel_iommu_domain_init,
3572         .domain_destroy = intel_iommu_domain_destroy,
3573         .attach_dev     = intel_iommu_attach_device,
3574         .detach_dev     = intel_iommu_detach_device,
3575         .map            = intel_iommu_map_range,
3576         .unmap          = intel_iommu_unmap_range,
3577         .iova_to_phys   = intel_iommu_iova_to_phys,
3578         .domain_has_cap = intel_iommu_domain_has_cap,
3579 };
3580
3581 static void __devinit quirk_iommu_rwbf(struct pci_dev *dev)
3582 {
3583         /*
3584          * Mobile 4 Series Chipset neglects to set RWBF capability,
3585          * but needs it:
3586          */
3587         printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
3588         rwbf_quirk = 1;
3589 }
3590
3591 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);