Merge git://git.infradead.org/~dwmw2/iommu-2.6.31
[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
60 #define IOVA_PFN(addr)          ((addr) >> PAGE_SHIFT)
61 #define DMA_32BIT_PFN           IOVA_PFN(DMA_BIT_MASK(32))
62 #define DMA_64BIT_PFN           IOVA_PFN(DMA_BIT_MASK(64))
63
64 #ifndef PHYSICAL_PAGE_MASK
65 #define PHYSICAL_PAGE_MASK PAGE_MASK
66 #endif
67
68 /* global iommu list, set NULL for ignored DMAR units */
69 static struct intel_iommu **g_iommus;
70
71 static int rwbf_quirk;
72
73 /*
74  * 0: Present
75  * 1-11: Reserved
76  * 12-63: Context Ptr (12 - (haw-1))
77  * 64-127: Reserved
78  */
79 struct root_entry {
80         u64     val;
81         u64     rsvd1;
82 };
83 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
84 static inline bool root_present(struct root_entry *root)
85 {
86         return (root->val & 1);
87 }
88 static inline void set_root_present(struct root_entry *root)
89 {
90         root->val |= 1;
91 }
92 static inline void set_root_value(struct root_entry *root, unsigned long value)
93 {
94         root->val |= value & VTD_PAGE_MASK;
95 }
96
97 static inline struct context_entry *
98 get_context_addr_from_root(struct root_entry *root)
99 {
100         return (struct context_entry *)
101                 (root_present(root)?phys_to_virt(
102                 root->val & VTD_PAGE_MASK) :
103                 NULL);
104 }
105
106 /*
107  * low 64 bits:
108  * 0: present
109  * 1: fault processing disable
110  * 2-3: translation type
111  * 12-63: address space root
112  * high 64 bits:
113  * 0-2: address width
114  * 3-6: aval
115  * 8-23: domain id
116  */
117 struct context_entry {
118         u64 lo;
119         u64 hi;
120 };
121
122 static inline bool context_present(struct context_entry *context)
123 {
124         return (context->lo & 1);
125 }
126 static inline void context_set_present(struct context_entry *context)
127 {
128         context->lo |= 1;
129 }
130
131 static inline void context_set_fault_enable(struct context_entry *context)
132 {
133         context->lo &= (((u64)-1) << 2) | 1;
134 }
135
136 static inline void context_set_translation_type(struct context_entry *context,
137                                                 unsigned long value)
138 {
139         context->lo &= (((u64)-1) << 4) | 3;
140         context->lo |= (value & 3) << 2;
141 }
142
143 static inline void context_set_address_root(struct context_entry *context,
144                                             unsigned long value)
145 {
146         context->lo |= value & VTD_PAGE_MASK;
147 }
148
149 static inline void context_set_address_width(struct context_entry *context,
150                                              unsigned long value)
151 {
152         context->hi |= value & 7;
153 }
154
155 static inline void context_set_domain_id(struct context_entry *context,
156                                          unsigned long value)
157 {
158         context->hi |= (value & ((1 << 16) - 1)) << 8;
159 }
160
161 static inline void context_clear_entry(struct context_entry *context)
162 {
163         context->lo = 0;
164         context->hi = 0;
165 }
166
167 /*
168  * 0: readable
169  * 1: writable
170  * 2-6: reserved
171  * 7: super page
172  * 8-10: available
173  * 11: snoop behavior
174  * 12-63: Host physcial address
175  */
176 struct dma_pte {
177         u64 val;
178 };
179
180 static inline void dma_clear_pte(struct dma_pte *pte)
181 {
182         pte->val = 0;
183 }
184
185 static inline void dma_set_pte_readable(struct dma_pte *pte)
186 {
187         pte->val |= DMA_PTE_READ;
188 }
189
190 static inline void dma_set_pte_writable(struct dma_pte *pte)
191 {
192         pte->val |= DMA_PTE_WRITE;
193 }
194
195 static inline void dma_set_pte_snp(struct dma_pte *pte)
196 {
197         pte->val |= DMA_PTE_SNP;
198 }
199
200 static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)
201 {
202         pte->val = (pte->val & ~3) | (prot & 3);
203 }
204
205 static inline u64 dma_pte_addr(struct dma_pte *pte)
206 {
207         return (pte->val & VTD_PAGE_MASK);
208 }
209
210 static inline void dma_set_pte_addr(struct dma_pte *pte, u64 addr)
211 {
212         pte->val |= (addr & VTD_PAGE_MASK);
213 }
214
215 static inline bool dma_pte_present(struct dma_pte *pte)
216 {
217         return (pte->val & 3) != 0;
218 }
219
220 /* devices under the same p2p bridge are owned in one domain */
221 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
222
223 /* domain represents a virtual machine, more than one devices
224  * across iommus may be owned in one domain, e.g. kvm guest.
225  */
226 #define DOMAIN_FLAG_VIRTUAL_MACHINE     (1 << 1)
227
228 struct dmar_domain {
229         int     id;                     /* domain id */
230         unsigned long iommu_bmp;        /* bitmap of iommus this domain uses*/
231
232         struct list_head devices;       /* all devices' list */
233         struct iova_domain iovad;       /* iova's that belong to this domain */
234
235         struct dma_pte  *pgd;           /* virtual address */
236         spinlock_t      mapping_lock;   /* page table lock */
237         int             gaw;            /* max guest address width */
238
239         /* adjusted guest address width, 0 is level 2 30-bit */
240         int             agaw;
241
242         int             flags;          /* flags to find out type of domain */
243
244         int             iommu_coherency;/* indicate coherency of iommu access */
245         int             iommu_snooping; /* indicate snooping control feature*/
246         int             iommu_count;    /* reference count of iommu */
247         spinlock_t      iommu_lock;     /* protect iommu set in domain */
248         u64             max_addr;       /* maximum mapped address */
249 };
250
251 /* PCI domain-device relationship */
252 struct device_domain_info {
253         struct list_head link;  /* link to domain siblings */
254         struct list_head global; /* link to global list */
255         int segment;            /* PCI domain */
256         u8 bus;                 /* PCI bus number */
257         u8 devfn;               /* PCI devfn number */
258         struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
259         struct intel_iommu *iommu; /* IOMMU used by this device */
260         struct dmar_domain *domain; /* pointer to domain */
261 };
262
263 static void flush_unmaps_timeout(unsigned long data);
264
265 DEFINE_TIMER(unmap_timer,  flush_unmaps_timeout, 0, 0);
266
267 #define HIGH_WATER_MARK 250
268 struct deferred_flush_tables {
269         int next;
270         struct iova *iova[HIGH_WATER_MARK];
271         struct dmar_domain *domain[HIGH_WATER_MARK];
272 };
273
274 static struct deferred_flush_tables *deferred_flush;
275
276 /* bitmap for indexing intel_iommus */
277 static int g_num_of_iommus;
278
279 static DEFINE_SPINLOCK(async_umap_flush_lock);
280 static LIST_HEAD(unmaps_to_do);
281
282 static int timer_on;
283 static long list_size;
284
285 static void domain_remove_dev_info(struct dmar_domain *domain);
286
287 #ifdef CONFIG_DMAR_DEFAULT_ON
288 int dmar_disabled = 0;
289 #else
290 int dmar_disabled = 1;
291 #endif /*CONFIG_DMAR_DEFAULT_ON*/
292
293 static int __initdata dmar_map_gfx = 1;
294 static int dmar_forcedac;
295 static int intel_iommu_strict;
296
297 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
298 static DEFINE_SPINLOCK(device_domain_lock);
299 static LIST_HEAD(device_domain_list);
300
301 static struct iommu_ops intel_iommu_ops;
302
303 static int __init intel_iommu_setup(char *str)
304 {
305         if (!str)
306                 return -EINVAL;
307         while (*str) {
308                 if (!strncmp(str, "on", 2)) {
309                         dmar_disabled = 0;
310                         printk(KERN_INFO "Intel-IOMMU: enabled\n");
311                 } else if (!strncmp(str, "off", 3)) {
312                         dmar_disabled = 1;
313                         printk(KERN_INFO "Intel-IOMMU: disabled\n");
314                 } else if (!strncmp(str, "igfx_off", 8)) {
315                         dmar_map_gfx = 0;
316                         printk(KERN_INFO
317                                 "Intel-IOMMU: disable GFX device mapping\n");
318                 } else if (!strncmp(str, "forcedac", 8)) {
319                         printk(KERN_INFO
320                                 "Intel-IOMMU: Forcing DAC for PCI devices\n");
321                         dmar_forcedac = 1;
322                 } else if (!strncmp(str, "strict", 6)) {
323                         printk(KERN_INFO
324                                 "Intel-IOMMU: disable batched IOTLB flush\n");
325                         intel_iommu_strict = 1;
326                 }
327
328                 str += strcspn(str, ",");
329                 while (*str == ',')
330                         str++;
331         }
332         return 0;
333 }
334 __setup("intel_iommu=", intel_iommu_setup);
335
336 static struct kmem_cache *iommu_domain_cache;
337 static struct kmem_cache *iommu_devinfo_cache;
338 static struct kmem_cache *iommu_iova_cache;
339
340 static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
341 {
342         unsigned int flags;
343         void *vaddr;
344
345         /* trying to avoid low memory issues */
346         flags = current->flags & PF_MEMALLOC;
347         current->flags |= PF_MEMALLOC;
348         vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
349         current->flags &= (~PF_MEMALLOC | flags);
350         return vaddr;
351 }
352
353
354 static inline void *alloc_pgtable_page(void)
355 {
356         unsigned int flags;
357         void *vaddr;
358
359         /* trying to avoid low memory issues */
360         flags = current->flags & PF_MEMALLOC;
361         current->flags |= PF_MEMALLOC;
362         vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
363         current->flags &= (~PF_MEMALLOC | flags);
364         return vaddr;
365 }
366
367 static inline void free_pgtable_page(void *vaddr)
368 {
369         free_page((unsigned long)vaddr);
370 }
371
372 static inline void *alloc_domain_mem(void)
373 {
374         return iommu_kmem_cache_alloc(iommu_domain_cache);
375 }
376
377 static void free_domain_mem(void *vaddr)
378 {
379         kmem_cache_free(iommu_domain_cache, vaddr);
380 }
381
382 static inline void * alloc_devinfo_mem(void)
383 {
384         return iommu_kmem_cache_alloc(iommu_devinfo_cache);
385 }
386
387 static inline void free_devinfo_mem(void *vaddr)
388 {
389         kmem_cache_free(iommu_devinfo_cache, vaddr);
390 }
391
392 struct iova *alloc_iova_mem(void)
393 {
394         return iommu_kmem_cache_alloc(iommu_iova_cache);
395 }
396
397 void free_iova_mem(struct iova *iova)
398 {
399         kmem_cache_free(iommu_iova_cache, iova);
400 }
401
402
403 static inline int width_to_agaw(int width);
404
405 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
406 {
407         unsigned long sagaw;
408         int agaw = -1;
409
410         sagaw = cap_sagaw(iommu->cap);
411         for (agaw = width_to_agaw(max_gaw);
412              agaw >= 0; agaw--) {
413                 if (test_bit(agaw, &sagaw))
414                         break;
415         }
416
417         return agaw;
418 }
419
420 /*
421  * Calculate max SAGAW for each iommu.
422  */
423 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
424 {
425         return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
426 }
427
428 /*
429  * calculate agaw for each iommu.
430  * "SAGAW" may be different across iommus, use a default agaw, and
431  * get a supported less agaw for iommus that don't support the default agaw.
432  */
433 int iommu_calculate_agaw(struct intel_iommu *iommu)
434 {
435         return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
436 }
437
438 /* in native case, each domain is related to only one iommu */
439 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
440 {
441         int iommu_id;
442
443         BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
444
445         iommu_id = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
446         if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
447                 return NULL;
448
449         return g_iommus[iommu_id];
450 }
451
452 static void domain_update_iommu_coherency(struct dmar_domain *domain)
453 {
454         int i;
455
456         domain->iommu_coherency = 1;
457
458         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
459         for (; i < g_num_of_iommus; ) {
460                 if (!ecap_coherent(g_iommus[i]->ecap)) {
461                         domain->iommu_coherency = 0;
462                         break;
463                 }
464                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
465         }
466 }
467
468 static void domain_update_iommu_snooping(struct dmar_domain *domain)
469 {
470         int i;
471
472         domain->iommu_snooping = 1;
473
474         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
475         for (; i < g_num_of_iommus; ) {
476                 if (!ecap_sc_support(g_iommus[i]->ecap)) {
477                         domain->iommu_snooping = 0;
478                         break;
479                 }
480                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
481         }
482 }
483
484 /* Some capabilities may be different across iommus */
485 static void domain_update_iommu_cap(struct dmar_domain *domain)
486 {
487         domain_update_iommu_coherency(domain);
488         domain_update_iommu_snooping(domain);
489 }
490
491 static struct intel_iommu *device_to_iommu(int segment, u8 bus, u8 devfn)
492 {
493         struct dmar_drhd_unit *drhd = NULL;
494         int i;
495
496         for_each_drhd_unit(drhd) {
497                 if (drhd->ignored)
498                         continue;
499                 if (segment != drhd->segment)
500                         continue;
501
502                 for (i = 0; i < drhd->devices_cnt; i++) {
503                         if (drhd->devices[i] &&
504                             drhd->devices[i]->bus->number == bus &&
505                             drhd->devices[i]->devfn == devfn)
506                                 return drhd->iommu;
507                         if (drhd->devices[i] &&
508                             drhd->devices[i]->subordinate &&
509                             drhd->devices[i]->subordinate->number <= bus &&
510                             drhd->devices[i]->subordinate->subordinate >= bus)
511                                 return drhd->iommu;
512                 }
513
514                 if (drhd->include_all)
515                         return drhd->iommu;
516         }
517
518         return NULL;
519 }
520
521 static void domain_flush_cache(struct dmar_domain *domain,
522                                void *addr, int size)
523 {
524         if (!domain->iommu_coherency)
525                 clflush_cache_range(addr, size);
526 }
527
528 /* Gets context entry for a given bus and devfn */
529 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
530                 u8 bus, u8 devfn)
531 {
532         struct root_entry *root;
533         struct context_entry *context;
534         unsigned long phy_addr;
535         unsigned long flags;
536
537         spin_lock_irqsave(&iommu->lock, flags);
538         root = &iommu->root_entry[bus];
539         context = get_context_addr_from_root(root);
540         if (!context) {
541                 context = (struct context_entry *)alloc_pgtable_page();
542                 if (!context) {
543                         spin_unlock_irqrestore(&iommu->lock, flags);
544                         return NULL;
545                 }
546                 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
547                 phy_addr = virt_to_phys((void *)context);
548                 set_root_value(root, phy_addr);
549                 set_root_present(root);
550                 __iommu_flush_cache(iommu, root, sizeof(*root));
551         }
552         spin_unlock_irqrestore(&iommu->lock, flags);
553         return &context[devfn];
554 }
555
556 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
557 {
558         struct root_entry *root;
559         struct context_entry *context;
560         int ret;
561         unsigned long flags;
562
563         spin_lock_irqsave(&iommu->lock, flags);
564         root = &iommu->root_entry[bus];
565         context = get_context_addr_from_root(root);
566         if (!context) {
567                 ret = 0;
568                 goto out;
569         }
570         ret = context_present(&context[devfn]);
571 out:
572         spin_unlock_irqrestore(&iommu->lock, flags);
573         return ret;
574 }
575
576 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
577 {
578         struct root_entry *root;
579         struct context_entry *context;
580         unsigned long flags;
581
582         spin_lock_irqsave(&iommu->lock, flags);
583         root = &iommu->root_entry[bus];
584         context = get_context_addr_from_root(root);
585         if (context) {
586                 context_clear_entry(&context[devfn]);
587                 __iommu_flush_cache(iommu, &context[devfn], \
588                         sizeof(*context));
589         }
590         spin_unlock_irqrestore(&iommu->lock, flags);
591 }
592
593 static void free_context_table(struct intel_iommu *iommu)
594 {
595         struct root_entry *root;
596         int i;
597         unsigned long flags;
598         struct context_entry *context;
599
600         spin_lock_irqsave(&iommu->lock, flags);
601         if (!iommu->root_entry) {
602                 goto out;
603         }
604         for (i = 0; i < ROOT_ENTRY_NR; i++) {
605                 root = &iommu->root_entry[i];
606                 context = get_context_addr_from_root(root);
607                 if (context)
608                         free_pgtable_page(context);
609         }
610         free_pgtable_page(iommu->root_entry);
611         iommu->root_entry = NULL;
612 out:
613         spin_unlock_irqrestore(&iommu->lock, flags);
614 }
615
616 /* page table handling */
617 #define LEVEL_STRIDE            (9)
618 #define LEVEL_MASK              (((u64)1 << LEVEL_STRIDE) - 1)
619
620 static inline int agaw_to_level(int agaw)
621 {
622         return agaw + 2;
623 }
624
625 static inline int agaw_to_width(int agaw)
626 {
627         return 30 + agaw * LEVEL_STRIDE;
628
629 }
630
631 static inline int width_to_agaw(int width)
632 {
633         return (width - 30) / LEVEL_STRIDE;
634 }
635
636 static inline unsigned int level_to_offset_bits(int level)
637 {
638         return (12 + (level - 1) * LEVEL_STRIDE);
639 }
640
641 static inline int address_level_offset(u64 addr, int level)
642 {
643         return ((addr >> level_to_offset_bits(level)) & LEVEL_MASK);
644 }
645
646 static inline u64 level_mask(int level)
647 {
648         return ((u64)-1 << level_to_offset_bits(level));
649 }
650
651 static inline u64 level_size(int level)
652 {
653         return ((u64)1 << level_to_offset_bits(level));
654 }
655
656 static inline u64 align_to_level(u64 addr, int level)
657 {
658         return ((addr + level_size(level) - 1) & level_mask(level));
659 }
660
661 static struct dma_pte * addr_to_dma_pte(struct dmar_domain *domain, u64 addr)
662 {
663         int addr_width = agaw_to_width(domain->agaw);
664         struct dma_pte *parent, *pte = NULL;
665         int level = agaw_to_level(domain->agaw);
666         int offset;
667         unsigned long flags;
668
669         BUG_ON(!domain->pgd);
670
671         addr &= (((u64)1) << addr_width) - 1;
672         parent = domain->pgd;
673
674         spin_lock_irqsave(&domain->mapping_lock, flags);
675         while (level > 0) {
676                 void *tmp_page;
677
678                 offset = address_level_offset(addr, level);
679                 pte = &parent[offset];
680                 if (level == 1)
681                         break;
682
683                 if (!dma_pte_present(pte)) {
684                         tmp_page = alloc_pgtable_page();
685
686                         if (!tmp_page) {
687                                 spin_unlock_irqrestore(&domain->mapping_lock,
688                                         flags);
689                                 return NULL;
690                         }
691                         domain_flush_cache(domain, tmp_page, PAGE_SIZE);
692                         dma_set_pte_addr(pte, virt_to_phys(tmp_page));
693                         /*
694                          * high level table always sets r/w, last level page
695                          * table control read/write
696                          */
697                         dma_set_pte_readable(pte);
698                         dma_set_pte_writable(pte);
699                         domain_flush_cache(domain, pte, sizeof(*pte));
700                 }
701                 parent = phys_to_virt(dma_pte_addr(pte));
702                 level--;
703         }
704
705         spin_unlock_irqrestore(&domain->mapping_lock, flags);
706         return pte;
707 }
708
709 /* return address's pte at specific level */
710 static struct dma_pte *dma_addr_level_pte(struct dmar_domain *domain, u64 addr,
711                 int level)
712 {
713         struct dma_pte *parent, *pte = NULL;
714         int total = agaw_to_level(domain->agaw);
715         int offset;
716
717         parent = domain->pgd;
718         while (level <= total) {
719                 offset = address_level_offset(addr, total);
720                 pte = &parent[offset];
721                 if (level == total)
722                         return pte;
723
724                 if (!dma_pte_present(pte))
725                         break;
726                 parent = phys_to_virt(dma_pte_addr(pte));
727                 total--;
728         }
729         return NULL;
730 }
731
732 /* clear one page's page table */
733 static void dma_pte_clear_one(struct dmar_domain *domain, u64 addr)
734 {
735         struct dma_pte *pte = NULL;
736
737         /* get last level pte */
738         pte = dma_addr_level_pte(domain, addr, 1);
739
740         if (pte) {
741                 dma_clear_pte(pte);
742                 domain_flush_cache(domain, pte, sizeof(*pte));
743         }
744 }
745
746 /* clear last level pte, a tlb flush should be followed */
747 static void dma_pte_clear_range(struct dmar_domain *domain, u64 start, u64 end)
748 {
749         int addr_width = agaw_to_width(domain->agaw);
750         int npages;
751
752         start &= (((u64)1) << addr_width) - 1;
753         end &= (((u64)1) << addr_width) - 1;
754         /* in case it's partial page */
755         start &= PAGE_MASK;
756         end = PAGE_ALIGN(end);
757         npages = (end - start) / VTD_PAGE_SIZE;
758
759         /* we don't need lock here, nobody else touches the iova range */
760         while (npages--) {
761                 dma_pte_clear_one(domain, start);
762                 start += VTD_PAGE_SIZE;
763         }
764 }
765
766 /* free page table pages. last level pte should already be cleared */
767 static void dma_pte_free_pagetable(struct dmar_domain *domain,
768         u64 start, u64 end)
769 {
770         int addr_width = agaw_to_width(domain->agaw);
771         struct dma_pte *pte;
772         int total = agaw_to_level(domain->agaw);
773         int level;
774         u64 tmp;
775
776         start &= (((u64)1) << addr_width) - 1;
777         end &= (((u64)1) << addr_width) - 1;
778
779         /* we don't need lock here, nobody else touches the iova range */
780         level = 2;
781         while (level <= total) {
782                 tmp = align_to_level(start, level);
783                 if (tmp >= end || (tmp + level_size(level) > end))
784                         return;
785
786                 while (tmp < end) {
787                         pte = dma_addr_level_pte(domain, tmp, level);
788                         if (pte) {
789                                 free_pgtable_page(
790                                         phys_to_virt(dma_pte_addr(pte)));
791                                 dma_clear_pte(pte);
792                                 domain_flush_cache(domain, pte, sizeof(*pte));
793                         }
794                         tmp += level_size(level);
795                 }
796                 level++;
797         }
798         /* free pgd */
799         if (start == 0 && end >= ((((u64)1) << addr_width) - 1)) {
800                 free_pgtable_page(domain->pgd);
801                 domain->pgd = NULL;
802         }
803 }
804
805 /* iommu handling */
806 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
807 {
808         struct root_entry *root;
809         unsigned long flags;
810
811         root = (struct root_entry *)alloc_pgtable_page();
812         if (!root)
813                 return -ENOMEM;
814
815         __iommu_flush_cache(iommu, root, ROOT_SIZE);
816
817         spin_lock_irqsave(&iommu->lock, flags);
818         iommu->root_entry = root;
819         spin_unlock_irqrestore(&iommu->lock, flags);
820
821         return 0;
822 }
823
824 static void iommu_set_root_entry(struct intel_iommu *iommu)
825 {
826         void *addr;
827         u32 sts;
828         unsigned long flag;
829
830         addr = iommu->root_entry;
831
832         spin_lock_irqsave(&iommu->register_lock, flag);
833         dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
834
835         writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
836
837         /* Make sure hardware complete it */
838         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
839                       readl, (sts & DMA_GSTS_RTPS), sts);
840
841         spin_unlock_irqrestore(&iommu->register_lock, flag);
842 }
843
844 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
845 {
846         u32 val;
847         unsigned long flag;
848
849         if (!rwbf_quirk && !cap_rwbf(iommu->cap))
850                 return;
851
852         spin_lock_irqsave(&iommu->register_lock, flag);
853         writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
854
855         /* Make sure hardware complete it */
856         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
857                       readl, (!(val & DMA_GSTS_WBFS)), val);
858
859         spin_unlock_irqrestore(&iommu->register_lock, flag);
860 }
861
862 /* return value determine if we need a write buffer flush */
863 static void __iommu_flush_context(struct intel_iommu *iommu,
864                                   u16 did, u16 source_id, u8 function_mask,
865                                   u64 type)
866 {
867         u64 val = 0;
868         unsigned long flag;
869
870         switch (type) {
871         case DMA_CCMD_GLOBAL_INVL:
872                 val = DMA_CCMD_GLOBAL_INVL;
873                 break;
874         case DMA_CCMD_DOMAIN_INVL:
875                 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
876                 break;
877         case DMA_CCMD_DEVICE_INVL:
878                 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
879                         | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
880                 break;
881         default:
882                 BUG();
883         }
884         val |= DMA_CCMD_ICC;
885
886         spin_lock_irqsave(&iommu->register_lock, flag);
887         dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
888
889         /* Make sure hardware complete it */
890         IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
891                 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
892
893         spin_unlock_irqrestore(&iommu->register_lock, flag);
894 }
895
896 /* return value determine if we need a write buffer flush */
897 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
898                                 u64 addr, unsigned int size_order, u64 type)
899 {
900         int tlb_offset = ecap_iotlb_offset(iommu->ecap);
901         u64 val = 0, val_iva = 0;
902         unsigned long flag;
903
904         switch (type) {
905         case DMA_TLB_GLOBAL_FLUSH:
906                 /* global flush doesn't need set IVA_REG */
907                 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
908                 break;
909         case DMA_TLB_DSI_FLUSH:
910                 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
911                 break;
912         case DMA_TLB_PSI_FLUSH:
913                 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
914                 /* Note: always flush non-leaf currently */
915                 val_iva = size_order | addr;
916                 break;
917         default:
918                 BUG();
919         }
920         /* Note: set drain read/write */
921 #if 0
922         /*
923          * This is probably to be super secure.. Looks like we can
924          * ignore it without any impact.
925          */
926         if (cap_read_drain(iommu->cap))
927                 val |= DMA_TLB_READ_DRAIN;
928 #endif
929         if (cap_write_drain(iommu->cap))
930                 val |= DMA_TLB_WRITE_DRAIN;
931
932         spin_lock_irqsave(&iommu->register_lock, flag);
933         /* Note: Only uses first TLB reg currently */
934         if (val_iva)
935                 dmar_writeq(iommu->reg + tlb_offset, val_iva);
936         dmar_writeq(iommu->reg + tlb_offset + 8, val);
937
938         /* Make sure hardware complete it */
939         IOMMU_WAIT_OP(iommu, tlb_offset + 8,
940                 dmar_readq, (!(val & DMA_TLB_IVT)), val);
941
942         spin_unlock_irqrestore(&iommu->register_lock, flag);
943
944         /* check IOTLB invalidation granularity */
945         if (DMA_TLB_IAIG(val) == 0)
946                 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
947         if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
948                 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
949                         (unsigned long long)DMA_TLB_IIRG(type),
950                         (unsigned long long)DMA_TLB_IAIG(val));
951 }
952
953 static struct device_domain_info *iommu_support_dev_iotlb(
954         struct dmar_domain *domain, int segment, u8 bus, u8 devfn)
955 {
956         int found = 0;
957         unsigned long flags;
958         struct device_domain_info *info;
959         struct intel_iommu *iommu = device_to_iommu(segment, bus, devfn);
960
961         if (!ecap_dev_iotlb_support(iommu->ecap))
962                 return NULL;
963
964         if (!iommu->qi)
965                 return NULL;
966
967         spin_lock_irqsave(&device_domain_lock, flags);
968         list_for_each_entry(info, &domain->devices, link)
969                 if (info->bus == bus && info->devfn == devfn) {
970                         found = 1;
971                         break;
972                 }
973         spin_unlock_irqrestore(&device_domain_lock, flags);
974
975         if (!found || !info->dev)
976                 return NULL;
977
978         if (!pci_find_ext_capability(info->dev, PCI_EXT_CAP_ID_ATS))
979                 return NULL;
980
981         if (!dmar_find_matched_atsr_unit(info->dev))
982                 return NULL;
983
984         info->iommu = iommu;
985
986         return info;
987 }
988
989 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
990 {
991         if (!info)
992                 return;
993
994         pci_enable_ats(info->dev, VTD_PAGE_SHIFT);
995 }
996
997 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
998 {
999         if (!info->dev || !pci_ats_enabled(info->dev))
1000                 return;
1001
1002         pci_disable_ats(info->dev);
1003 }
1004
1005 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1006                                   u64 addr, unsigned mask)
1007 {
1008         u16 sid, qdep;
1009         unsigned long flags;
1010         struct device_domain_info *info;
1011
1012         spin_lock_irqsave(&device_domain_lock, flags);
1013         list_for_each_entry(info, &domain->devices, link) {
1014                 if (!info->dev || !pci_ats_enabled(info->dev))
1015                         continue;
1016
1017                 sid = info->bus << 8 | info->devfn;
1018                 qdep = pci_ats_queue_depth(info->dev);
1019                 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1020         }
1021         spin_unlock_irqrestore(&device_domain_lock, flags);
1022 }
1023
1024 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
1025                                   u64 addr, unsigned int pages)
1026 {
1027         unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1028
1029         BUG_ON(addr & (~VTD_PAGE_MASK));
1030         BUG_ON(pages == 0);
1031
1032         /*
1033          * Fallback to domain selective flush if no PSI support or the size is
1034          * too big.
1035          * PSI requires page size to be 2 ^ x, and the base address is naturally
1036          * aligned to the size
1037          */
1038         if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1039                 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1040                                                 DMA_TLB_DSI_FLUSH);
1041         else
1042                 iommu->flush.flush_iotlb(iommu, did, addr, mask,
1043                                                 DMA_TLB_PSI_FLUSH);
1044         if (did)
1045                 iommu_flush_dev_iotlb(iommu->domains[did], addr, mask);
1046 }
1047
1048 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1049 {
1050         u32 pmen;
1051         unsigned long flags;
1052
1053         spin_lock_irqsave(&iommu->register_lock, flags);
1054         pmen = readl(iommu->reg + DMAR_PMEN_REG);
1055         pmen &= ~DMA_PMEN_EPM;
1056         writel(pmen, iommu->reg + DMAR_PMEN_REG);
1057
1058         /* wait for the protected region status bit to clear */
1059         IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1060                 readl, !(pmen & DMA_PMEN_PRS), pmen);
1061
1062         spin_unlock_irqrestore(&iommu->register_lock, flags);
1063 }
1064
1065 static int iommu_enable_translation(struct intel_iommu *iommu)
1066 {
1067         u32 sts;
1068         unsigned long flags;
1069
1070         spin_lock_irqsave(&iommu->register_lock, flags);
1071         iommu->gcmd |= DMA_GCMD_TE;
1072         writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1073
1074         /* Make sure hardware complete it */
1075         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1076                       readl, (sts & DMA_GSTS_TES), sts);
1077
1078         spin_unlock_irqrestore(&iommu->register_lock, flags);
1079         return 0;
1080 }
1081
1082 static int iommu_disable_translation(struct intel_iommu *iommu)
1083 {
1084         u32 sts;
1085         unsigned long flag;
1086
1087         spin_lock_irqsave(&iommu->register_lock, flag);
1088         iommu->gcmd &= ~DMA_GCMD_TE;
1089         writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1090
1091         /* Make sure hardware complete it */
1092         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1093                       readl, (!(sts & DMA_GSTS_TES)), sts);
1094
1095         spin_unlock_irqrestore(&iommu->register_lock, flag);
1096         return 0;
1097 }
1098
1099
1100 static int iommu_init_domains(struct intel_iommu *iommu)
1101 {
1102         unsigned long ndomains;
1103         unsigned long nlongs;
1104
1105         ndomains = cap_ndoms(iommu->cap);
1106         pr_debug("Number of Domains supportd <%ld>\n", ndomains);
1107         nlongs = BITS_TO_LONGS(ndomains);
1108
1109         /* TBD: there might be 64K domains,
1110          * consider other allocation for future chip
1111          */
1112         iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1113         if (!iommu->domain_ids) {
1114                 printk(KERN_ERR "Allocating domain id array failed\n");
1115                 return -ENOMEM;
1116         }
1117         iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1118                         GFP_KERNEL);
1119         if (!iommu->domains) {
1120                 printk(KERN_ERR "Allocating domain array failed\n");
1121                 kfree(iommu->domain_ids);
1122                 return -ENOMEM;
1123         }
1124
1125         spin_lock_init(&iommu->lock);
1126
1127         /*
1128          * if Caching mode is set, then invalid translations are tagged
1129          * with domainid 0. Hence we need to pre-allocate it.
1130          */
1131         if (cap_caching_mode(iommu->cap))
1132                 set_bit(0, iommu->domain_ids);
1133         return 0;
1134 }
1135
1136
1137 static void domain_exit(struct dmar_domain *domain);
1138 static void vm_domain_exit(struct dmar_domain *domain);
1139
1140 void free_dmar_iommu(struct intel_iommu *iommu)
1141 {
1142         struct dmar_domain *domain;
1143         int i;
1144         unsigned long flags;
1145
1146         i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1147         for (; i < cap_ndoms(iommu->cap); ) {
1148                 domain = iommu->domains[i];
1149                 clear_bit(i, iommu->domain_ids);
1150
1151                 spin_lock_irqsave(&domain->iommu_lock, flags);
1152                 if (--domain->iommu_count == 0) {
1153                         if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
1154                                 vm_domain_exit(domain);
1155                         else
1156                                 domain_exit(domain);
1157                 }
1158                 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1159
1160                 i = find_next_bit(iommu->domain_ids,
1161                         cap_ndoms(iommu->cap), i+1);
1162         }
1163
1164         if (iommu->gcmd & DMA_GCMD_TE)
1165                 iommu_disable_translation(iommu);
1166
1167         if (iommu->irq) {
1168                 set_irq_data(iommu->irq, NULL);
1169                 /* This will mask the irq */
1170                 free_irq(iommu->irq, iommu);
1171                 destroy_irq(iommu->irq);
1172         }
1173
1174         kfree(iommu->domains);
1175         kfree(iommu->domain_ids);
1176
1177         g_iommus[iommu->seq_id] = NULL;
1178
1179         /* if all iommus are freed, free g_iommus */
1180         for (i = 0; i < g_num_of_iommus; i++) {
1181                 if (g_iommus[i])
1182                         break;
1183         }
1184
1185         if (i == g_num_of_iommus)
1186                 kfree(g_iommus);
1187
1188         /* free context mapping */
1189         free_context_table(iommu);
1190 }
1191
1192 static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
1193 {
1194         unsigned long num;
1195         unsigned long ndomains;
1196         struct dmar_domain *domain;
1197         unsigned long flags;
1198
1199         domain = alloc_domain_mem();
1200         if (!domain)
1201                 return NULL;
1202
1203         ndomains = cap_ndoms(iommu->cap);
1204
1205         spin_lock_irqsave(&iommu->lock, flags);
1206         num = find_first_zero_bit(iommu->domain_ids, ndomains);
1207         if (num >= ndomains) {
1208                 spin_unlock_irqrestore(&iommu->lock, flags);
1209                 free_domain_mem(domain);
1210                 printk(KERN_ERR "IOMMU: no free domain ids\n");
1211                 return NULL;
1212         }
1213
1214         set_bit(num, iommu->domain_ids);
1215         domain->id = num;
1216         memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
1217         set_bit(iommu->seq_id, &domain->iommu_bmp);
1218         domain->flags = 0;
1219         iommu->domains[num] = domain;
1220         spin_unlock_irqrestore(&iommu->lock, flags);
1221
1222         return domain;
1223 }
1224
1225 static void iommu_free_domain(struct dmar_domain *domain)
1226 {
1227         unsigned long flags;
1228         struct intel_iommu *iommu;
1229
1230         iommu = domain_get_iommu(domain);
1231
1232         spin_lock_irqsave(&iommu->lock, flags);
1233         clear_bit(domain->id, iommu->domain_ids);
1234         spin_unlock_irqrestore(&iommu->lock, flags);
1235 }
1236
1237 static struct iova_domain reserved_iova_list;
1238 static struct lock_class_key reserved_alloc_key;
1239 static struct lock_class_key reserved_rbtree_key;
1240
1241 static void dmar_init_reserved_ranges(void)
1242 {
1243         struct pci_dev *pdev = NULL;
1244         struct iova *iova;
1245         int i;
1246         u64 addr, size;
1247
1248         init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1249
1250         lockdep_set_class(&reserved_iova_list.iova_alloc_lock,
1251                 &reserved_alloc_key);
1252         lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1253                 &reserved_rbtree_key);
1254
1255         /* IOAPIC ranges shouldn't be accessed by DMA */
1256         iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1257                 IOVA_PFN(IOAPIC_RANGE_END));
1258         if (!iova)
1259                 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1260
1261         /* Reserve all PCI MMIO to avoid peer-to-peer access */
1262         for_each_pci_dev(pdev) {
1263                 struct resource *r;
1264
1265                 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1266                         r = &pdev->resource[i];
1267                         if (!r->flags || !(r->flags & IORESOURCE_MEM))
1268                                 continue;
1269                         addr = r->start;
1270                         addr &= PHYSICAL_PAGE_MASK;
1271                         size = r->end - addr;
1272                         size = PAGE_ALIGN(size);
1273                         iova = reserve_iova(&reserved_iova_list, IOVA_PFN(addr),
1274                                 IOVA_PFN(size + addr) - 1);
1275                         if (!iova)
1276                                 printk(KERN_ERR "Reserve iova failed\n");
1277                 }
1278         }
1279
1280 }
1281
1282 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1283 {
1284         copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1285 }
1286
1287 static inline int guestwidth_to_adjustwidth(int gaw)
1288 {
1289         int agaw;
1290         int r = (gaw - 12) % 9;
1291
1292         if (r == 0)
1293                 agaw = gaw;
1294         else
1295                 agaw = gaw + 9 - r;
1296         if (agaw > 64)
1297                 agaw = 64;
1298         return agaw;
1299 }
1300
1301 static int domain_init(struct dmar_domain *domain, int guest_width)
1302 {
1303         struct intel_iommu *iommu;
1304         int adjust_width, agaw;
1305         unsigned long sagaw;
1306
1307         init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1308         spin_lock_init(&domain->mapping_lock);
1309         spin_lock_init(&domain->iommu_lock);
1310
1311         domain_reserve_special_ranges(domain);
1312
1313         /* calculate AGAW */
1314         iommu = domain_get_iommu(domain);
1315         if (guest_width > cap_mgaw(iommu->cap))
1316                 guest_width = cap_mgaw(iommu->cap);
1317         domain->gaw = guest_width;
1318         adjust_width = guestwidth_to_adjustwidth(guest_width);
1319         agaw = width_to_agaw(adjust_width);
1320         sagaw = cap_sagaw(iommu->cap);
1321         if (!test_bit(agaw, &sagaw)) {
1322                 /* hardware doesn't support it, choose a bigger one */
1323                 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1324                 agaw = find_next_bit(&sagaw, 5, agaw);
1325                 if (agaw >= 5)
1326                         return -ENODEV;
1327         }
1328         domain->agaw = agaw;
1329         INIT_LIST_HEAD(&domain->devices);
1330
1331         if (ecap_coherent(iommu->ecap))
1332                 domain->iommu_coherency = 1;
1333         else
1334                 domain->iommu_coherency = 0;
1335
1336         if (ecap_sc_support(iommu->ecap))
1337                 domain->iommu_snooping = 1;
1338         else
1339                 domain->iommu_snooping = 0;
1340
1341         domain->iommu_count = 1;
1342
1343         /* always allocate the top pgd */
1344         domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1345         if (!domain->pgd)
1346                 return -ENOMEM;
1347         __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1348         return 0;
1349 }
1350
1351 static void domain_exit(struct dmar_domain *domain)
1352 {
1353         u64 end;
1354
1355         /* Domain 0 is reserved, so dont process it */
1356         if (!domain)
1357                 return;
1358
1359         domain_remove_dev_info(domain);
1360         /* destroy iovas */
1361         put_iova_domain(&domain->iovad);
1362         end = DOMAIN_MAX_ADDR(domain->gaw);
1363         end = end & (~PAGE_MASK);
1364
1365         /* clear ptes */
1366         dma_pte_clear_range(domain, 0, end);
1367
1368         /* free page tables */
1369         dma_pte_free_pagetable(domain, 0, end);
1370
1371         iommu_free_domain(domain);
1372         free_domain_mem(domain);
1373 }
1374
1375 static int domain_context_mapping_one(struct dmar_domain *domain, int segment,
1376                                  u8 bus, u8 devfn, int translation)
1377 {
1378         struct context_entry *context;
1379         unsigned long flags;
1380         struct intel_iommu *iommu;
1381         struct dma_pte *pgd;
1382         unsigned long num;
1383         unsigned long ndomains;
1384         int id;
1385         int agaw;
1386         struct device_domain_info *info = NULL;
1387
1388         pr_debug("Set context mapping for %02x:%02x.%d\n",
1389                 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1390
1391         BUG_ON(!domain->pgd);
1392         BUG_ON(translation != CONTEXT_TT_PASS_THROUGH &&
1393                translation != CONTEXT_TT_MULTI_LEVEL);
1394
1395         iommu = device_to_iommu(segment, bus, devfn);
1396         if (!iommu)
1397                 return -ENODEV;
1398
1399         context = device_to_context_entry(iommu, bus, devfn);
1400         if (!context)
1401                 return -ENOMEM;
1402         spin_lock_irqsave(&iommu->lock, flags);
1403         if (context_present(context)) {
1404                 spin_unlock_irqrestore(&iommu->lock, flags);
1405                 return 0;
1406         }
1407
1408         id = domain->id;
1409         pgd = domain->pgd;
1410
1411         if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE) {
1412                 int found = 0;
1413
1414                 /* find an available domain id for this device in iommu */
1415                 ndomains = cap_ndoms(iommu->cap);
1416                 num = find_first_bit(iommu->domain_ids, ndomains);
1417                 for (; num < ndomains; ) {
1418                         if (iommu->domains[num] == domain) {
1419                                 id = num;
1420                                 found = 1;
1421                                 break;
1422                         }
1423                         num = find_next_bit(iommu->domain_ids,
1424                                             cap_ndoms(iommu->cap), num+1);
1425                 }
1426
1427                 if (found == 0) {
1428                         num = find_first_zero_bit(iommu->domain_ids, ndomains);
1429                         if (num >= ndomains) {
1430                                 spin_unlock_irqrestore(&iommu->lock, flags);
1431                                 printk(KERN_ERR "IOMMU: no free domain ids\n");
1432                                 return -EFAULT;
1433                         }
1434
1435                         set_bit(num, iommu->domain_ids);
1436                         iommu->domains[num] = domain;
1437                         id = num;
1438                 }
1439
1440                 /* Skip top levels of page tables for
1441                  * iommu which has less agaw than default.
1442                  */
1443                 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1444                         pgd = phys_to_virt(dma_pte_addr(pgd));
1445                         if (!dma_pte_present(pgd)) {
1446                                 spin_unlock_irqrestore(&iommu->lock, flags);
1447                                 return -ENOMEM;
1448                         }
1449                 }
1450         }
1451
1452         context_set_domain_id(context, id);
1453
1454         if (translation != CONTEXT_TT_PASS_THROUGH) {
1455                 info = iommu_support_dev_iotlb(domain, segment, bus, devfn);
1456                 translation = info ? CONTEXT_TT_DEV_IOTLB :
1457                                      CONTEXT_TT_MULTI_LEVEL;
1458         }
1459         /*
1460          * In pass through mode, AW must be programmed to indicate the largest
1461          * AGAW value supported by hardware. And ASR is ignored by hardware.
1462          */
1463         if (unlikely(translation == CONTEXT_TT_PASS_THROUGH))
1464                 context_set_address_width(context, iommu->msagaw);
1465         else {
1466                 context_set_address_root(context, virt_to_phys(pgd));
1467                 context_set_address_width(context, iommu->agaw);
1468         }
1469
1470         context_set_translation_type(context, translation);
1471         context_set_fault_enable(context);
1472         context_set_present(context);
1473         domain_flush_cache(domain, context, sizeof(*context));
1474
1475         /*
1476          * It's a non-present to present mapping. If hardware doesn't cache
1477          * non-present entry we only need to flush the write-buffer. If the
1478          * _does_ cache non-present entries, then it does so in the special
1479          * domain #0, which we have to flush:
1480          */
1481         if (cap_caching_mode(iommu->cap)) {
1482                 iommu->flush.flush_context(iommu, 0,
1483                                            (((u16)bus) << 8) | devfn,
1484                                            DMA_CCMD_MASK_NOBIT,
1485                                            DMA_CCMD_DEVICE_INVL);
1486                 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
1487         } else {
1488                 iommu_flush_write_buffer(iommu);
1489         }
1490         iommu_enable_dev_iotlb(info);
1491         spin_unlock_irqrestore(&iommu->lock, flags);
1492
1493         spin_lock_irqsave(&domain->iommu_lock, flags);
1494         if (!test_and_set_bit(iommu->seq_id, &domain->iommu_bmp)) {
1495                 domain->iommu_count++;
1496                 domain_update_iommu_cap(domain);
1497         }
1498         spin_unlock_irqrestore(&domain->iommu_lock, flags);
1499         return 0;
1500 }
1501
1502 static int
1503 domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev,
1504                         int translation)
1505 {
1506         int ret;
1507         struct pci_dev *tmp, *parent;
1508
1509         ret = domain_context_mapping_one(domain, pci_domain_nr(pdev->bus),
1510                                          pdev->bus->number, pdev->devfn,
1511                                          translation);
1512         if (ret)
1513                 return ret;
1514
1515         /* dependent device mapping */
1516         tmp = pci_find_upstream_pcie_bridge(pdev);
1517         if (!tmp)
1518                 return 0;
1519         /* Secondary interface's bus number and devfn 0 */
1520         parent = pdev->bus->self;
1521         while (parent != tmp) {
1522                 ret = domain_context_mapping_one(domain,
1523                                                  pci_domain_nr(parent->bus),
1524                                                  parent->bus->number,
1525                                                  parent->devfn, translation);
1526                 if (ret)
1527                         return ret;
1528                 parent = parent->bus->self;
1529         }
1530         if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1531                 return domain_context_mapping_one(domain,
1532                                         pci_domain_nr(tmp->subordinate),
1533                                         tmp->subordinate->number, 0,
1534                                         translation);
1535         else /* this is a legacy PCI bridge */
1536                 return domain_context_mapping_one(domain,
1537                                                   pci_domain_nr(tmp->bus),
1538                                                   tmp->bus->number,
1539                                                   tmp->devfn,
1540                                                   translation);
1541 }
1542
1543 static int domain_context_mapped(struct pci_dev *pdev)
1544 {
1545         int ret;
1546         struct pci_dev *tmp, *parent;
1547         struct intel_iommu *iommu;
1548
1549         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
1550                                 pdev->devfn);
1551         if (!iommu)
1552                 return -ENODEV;
1553
1554         ret = device_context_mapped(iommu, pdev->bus->number, pdev->devfn);
1555         if (!ret)
1556                 return ret;
1557         /* dependent device mapping */
1558         tmp = pci_find_upstream_pcie_bridge(pdev);
1559         if (!tmp)
1560                 return ret;
1561         /* Secondary interface's bus number and devfn 0 */
1562         parent = pdev->bus->self;
1563         while (parent != tmp) {
1564                 ret = device_context_mapped(iommu, parent->bus->number,
1565                                             parent->devfn);
1566                 if (!ret)
1567                         return ret;
1568                 parent = parent->bus->self;
1569         }
1570         if (tmp->is_pcie)
1571                 return device_context_mapped(iommu, tmp->subordinate->number,
1572                                              0);
1573         else
1574                 return device_context_mapped(iommu, tmp->bus->number,
1575                                              tmp->devfn);
1576 }
1577
1578 static int
1579 domain_page_mapping(struct dmar_domain *domain, dma_addr_t iova,
1580                         u64 hpa, size_t size, int prot)
1581 {
1582         u64 start_pfn, end_pfn;
1583         struct dma_pte *pte;
1584         int index;
1585         int addr_width = agaw_to_width(domain->agaw);
1586
1587         hpa &= (((u64)1) << addr_width) - 1;
1588
1589         if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1590                 return -EINVAL;
1591         iova &= PAGE_MASK;
1592         start_pfn = ((u64)hpa) >> VTD_PAGE_SHIFT;
1593         end_pfn = (VTD_PAGE_ALIGN(((u64)hpa) + size)) >> VTD_PAGE_SHIFT;
1594         index = 0;
1595         while (start_pfn < end_pfn) {
1596                 pte = addr_to_dma_pte(domain, iova + VTD_PAGE_SIZE * index);
1597                 if (!pte)
1598                         return -ENOMEM;
1599                 /* We don't need lock here, nobody else
1600                  * touches the iova range
1601                  */
1602                 BUG_ON(dma_pte_addr(pte));
1603                 dma_set_pte_addr(pte, start_pfn << VTD_PAGE_SHIFT);
1604                 dma_set_pte_prot(pte, prot);
1605                 if (prot & DMA_PTE_SNP)
1606                         dma_set_pte_snp(pte);
1607                 domain_flush_cache(domain, pte, sizeof(*pte));
1608                 start_pfn++;
1609                 index++;
1610         }
1611         return 0;
1612 }
1613
1614 static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
1615 {
1616         if (!iommu)
1617                 return;
1618
1619         clear_context_table(iommu, bus, devfn);
1620         iommu->flush.flush_context(iommu, 0, 0, 0,
1621                                            DMA_CCMD_GLOBAL_INVL);
1622         iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
1623 }
1624
1625 static void domain_remove_dev_info(struct dmar_domain *domain)
1626 {
1627         struct device_domain_info *info;
1628         unsigned long flags;
1629         struct intel_iommu *iommu;
1630
1631         spin_lock_irqsave(&device_domain_lock, flags);
1632         while (!list_empty(&domain->devices)) {
1633                 info = list_entry(domain->devices.next,
1634                         struct device_domain_info, link);
1635                 list_del(&info->link);
1636                 list_del(&info->global);
1637                 if (info->dev)
1638                         info->dev->dev.archdata.iommu = NULL;
1639                 spin_unlock_irqrestore(&device_domain_lock, flags);
1640
1641                 iommu_disable_dev_iotlb(info);
1642                 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
1643                 iommu_detach_dev(iommu, info->bus, info->devfn);
1644                 free_devinfo_mem(info);
1645
1646                 spin_lock_irqsave(&device_domain_lock, flags);
1647         }
1648         spin_unlock_irqrestore(&device_domain_lock, flags);
1649 }
1650
1651 /*
1652  * find_domain
1653  * Note: we use struct pci_dev->dev.archdata.iommu stores the info
1654  */
1655 static struct dmar_domain *
1656 find_domain(struct pci_dev *pdev)
1657 {
1658         struct device_domain_info *info;
1659
1660         /* No lock here, assumes no domain exit in normal case */
1661         info = pdev->dev.archdata.iommu;
1662         if (info)
1663                 return info->domain;
1664         return NULL;
1665 }
1666
1667 /* domain is initialized */
1668 static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1669 {
1670         struct dmar_domain *domain, *found = NULL;
1671         struct intel_iommu *iommu;
1672         struct dmar_drhd_unit *drhd;
1673         struct device_domain_info *info, *tmp;
1674         struct pci_dev *dev_tmp;
1675         unsigned long flags;
1676         int bus = 0, devfn = 0;
1677         int segment;
1678
1679         domain = find_domain(pdev);
1680         if (domain)
1681                 return domain;
1682
1683         segment = pci_domain_nr(pdev->bus);
1684
1685         dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1686         if (dev_tmp) {
1687                 if (dev_tmp->is_pcie) {
1688                         bus = dev_tmp->subordinate->number;
1689                         devfn = 0;
1690                 } else {
1691                         bus = dev_tmp->bus->number;
1692                         devfn = dev_tmp->devfn;
1693                 }
1694                 spin_lock_irqsave(&device_domain_lock, flags);
1695                 list_for_each_entry(info, &device_domain_list, global) {
1696                         if (info->segment == segment &&
1697                             info->bus == bus && info->devfn == devfn) {
1698                                 found = info->domain;
1699                                 break;
1700                         }
1701                 }
1702                 spin_unlock_irqrestore(&device_domain_lock, flags);
1703                 /* pcie-pci bridge already has a domain, uses it */
1704                 if (found) {
1705                         domain = found;
1706                         goto found_domain;
1707                 }
1708         }
1709
1710         /* Allocate new domain for the device */
1711         drhd = dmar_find_matched_drhd_unit(pdev);
1712         if (!drhd) {
1713                 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1714                         pci_name(pdev));
1715                 return NULL;
1716         }
1717         iommu = drhd->iommu;
1718
1719         domain = iommu_alloc_domain(iommu);
1720         if (!domain)
1721                 goto error;
1722
1723         if (domain_init(domain, gaw)) {
1724                 domain_exit(domain);
1725                 goto error;
1726         }
1727
1728         /* register pcie-to-pci device */
1729         if (dev_tmp) {
1730                 info = alloc_devinfo_mem();
1731                 if (!info) {
1732                         domain_exit(domain);
1733                         goto error;
1734                 }
1735                 info->segment = segment;
1736                 info->bus = bus;
1737                 info->devfn = devfn;
1738                 info->dev = NULL;
1739                 info->domain = domain;
1740                 /* This domain is shared by devices under p2p bridge */
1741                 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
1742
1743                 /* pcie-to-pci bridge already has a domain, uses it */
1744                 found = NULL;
1745                 spin_lock_irqsave(&device_domain_lock, flags);
1746                 list_for_each_entry(tmp, &device_domain_list, global) {
1747                         if (tmp->segment == segment &&
1748                             tmp->bus == bus && tmp->devfn == devfn) {
1749                                 found = tmp->domain;
1750                                 break;
1751                         }
1752                 }
1753                 if (found) {
1754                         free_devinfo_mem(info);
1755                         domain_exit(domain);
1756                         domain = found;
1757                 } else {
1758                         list_add(&info->link, &domain->devices);
1759                         list_add(&info->global, &device_domain_list);
1760                 }
1761                 spin_unlock_irqrestore(&device_domain_lock, flags);
1762         }
1763
1764 found_domain:
1765         info = alloc_devinfo_mem();
1766         if (!info)
1767                 goto error;
1768         info->segment = segment;
1769         info->bus = pdev->bus->number;
1770         info->devfn = pdev->devfn;
1771         info->dev = pdev;
1772         info->domain = domain;
1773         spin_lock_irqsave(&device_domain_lock, flags);
1774         /* somebody is fast */
1775         found = find_domain(pdev);
1776         if (found != NULL) {
1777                 spin_unlock_irqrestore(&device_domain_lock, flags);
1778                 if (found != domain) {
1779                         domain_exit(domain);
1780                         domain = found;
1781                 }
1782                 free_devinfo_mem(info);
1783                 return domain;
1784         }
1785         list_add(&info->link, &domain->devices);
1786         list_add(&info->global, &device_domain_list);
1787         pdev->dev.archdata.iommu = info;
1788         spin_unlock_irqrestore(&device_domain_lock, flags);
1789         return domain;
1790 error:
1791         /* recheck it here, maybe others set it */
1792         return find_domain(pdev);
1793 }
1794
1795 static int iommu_prepare_identity_map(struct pci_dev *pdev,
1796                                       unsigned long long start,
1797                                       unsigned long long end)
1798 {
1799         struct dmar_domain *domain;
1800         unsigned long size;
1801         unsigned long long base;
1802         int ret;
1803
1804         printk(KERN_INFO
1805                 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1806                 pci_name(pdev), start, end);
1807         /* page table init */
1808         domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1809         if (!domain)
1810                 return -ENOMEM;
1811
1812         /* The address might not be aligned */
1813         base = start & PAGE_MASK;
1814         size = end - base;
1815         size = PAGE_ALIGN(size);
1816         if (!reserve_iova(&domain->iovad, IOVA_PFN(base),
1817                         IOVA_PFN(base + size) - 1)) {
1818                 printk(KERN_ERR "IOMMU: reserve iova failed\n");
1819                 ret = -ENOMEM;
1820                 goto error;
1821         }
1822
1823         pr_debug("Mapping reserved region %lx@%llx for %s\n",
1824                 size, base, pci_name(pdev));
1825         /*
1826          * RMRR range might have overlap with physical memory range,
1827          * clear it first
1828          */
1829         dma_pte_clear_range(domain, base, base + size);
1830
1831         ret = domain_page_mapping(domain, base, base, size,
1832                 DMA_PTE_READ|DMA_PTE_WRITE);
1833         if (ret)
1834                 goto error;
1835
1836         /* context entry init */
1837         ret = domain_context_mapping(domain, pdev, CONTEXT_TT_MULTI_LEVEL);
1838         if (!ret)
1839                 return 0;
1840 error:
1841         domain_exit(domain);
1842         return ret;
1843
1844 }
1845
1846 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1847         struct pci_dev *pdev)
1848 {
1849         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1850                 return 0;
1851         return iommu_prepare_identity_map(pdev, rmrr->base_address,
1852                 rmrr->end_address + 1);
1853 }
1854
1855 #ifdef CONFIG_DMAR_GFX_WA
1856 struct iommu_prepare_data {
1857         struct pci_dev *pdev;
1858         int ret;
1859 };
1860
1861 static int __init iommu_prepare_work_fn(unsigned long start_pfn,
1862                                          unsigned long end_pfn, void *datax)
1863 {
1864         struct iommu_prepare_data *data;
1865
1866         data = (struct iommu_prepare_data *)datax;
1867
1868         data->ret = iommu_prepare_identity_map(data->pdev,
1869                                 start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
1870         return data->ret;
1871
1872 }
1873
1874 static int __init iommu_prepare_with_active_regions(struct pci_dev *pdev)
1875 {
1876         int nid;
1877         struct iommu_prepare_data data;
1878
1879         data.pdev = pdev;
1880         data.ret = 0;
1881
1882         for_each_online_node(nid) {
1883                 work_with_active_regions(nid, iommu_prepare_work_fn, &data);
1884                 if (data.ret)
1885                         return data.ret;
1886         }
1887         return data.ret;
1888 }
1889
1890 static void __init iommu_prepare_gfx_mapping(void)
1891 {
1892         struct pci_dev *pdev = NULL;
1893         int ret;
1894
1895         for_each_pci_dev(pdev) {
1896                 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO ||
1897                                 !IS_GFX_DEVICE(pdev))
1898                         continue;
1899                 printk(KERN_INFO "IOMMU: gfx device %s 1-1 mapping\n",
1900                         pci_name(pdev));
1901                 ret = iommu_prepare_with_active_regions(pdev);
1902                 if (ret)
1903                         printk(KERN_ERR "IOMMU: mapping reserved region failed\n");
1904         }
1905 }
1906 #else /* !CONFIG_DMAR_GFX_WA */
1907 static inline void iommu_prepare_gfx_mapping(void)
1908 {
1909         return;
1910 }
1911 #endif
1912
1913 #ifdef CONFIG_DMAR_FLOPPY_WA
1914 static inline void iommu_prepare_isa(void)
1915 {
1916         struct pci_dev *pdev;
1917         int ret;
1918
1919         pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1920         if (!pdev)
1921                 return;
1922
1923         printk(KERN_INFO "IOMMU: Prepare 0-16M unity mapping for LPC\n");
1924         ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1925
1926         if (ret)
1927                 printk(KERN_ERR "IOMMU: Failed to create 0-64M identity map, "
1928                         "floppy might not work\n");
1929
1930 }
1931 #else
1932 static inline void iommu_prepare_isa(void)
1933 {
1934         return;
1935 }
1936 #endif /* !CONFIG_DMAR_FLPY_WA */
1937
1938 /* Initialize each context entry as pass through.*/
1939 static int __init init_context_pass_through(void)
1940 {
1941         struct pci_dev *pdev = NULL;
1942         struct dmar_domain *domain;
1943         int ret;
1944
1945         for_each_pci_dev(pdev) {
1946                 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1947                 ret = domain_context_mapping(domain, pdev,
1948                                              CONTEXT_TT_PASS_THROUGH);
1949                 if (ret)
1950                         return ret;
1951         }
1952         return 0;
1953 }
1954
1955 static int __init init_dmars(void)
1956 {
1957         struct dmar_drhd_unit *drhd;
1958         struct dmar_rmrr_unit *rmrr;
1959         struct pci_dev *pdev;
1960         struct intel_iommu *iommu;
1961         int i, ret;
1962         int pass_through = 1;
1963
1964         /*
1965          * for each drhd
1966          *    allocate root
1967          *    initialize and program root entry to not present
1968          * endfor
1969          */
1970         for_each_drhd_unit(drhd) {
1971                 g_num_of_iommus++;
1972                 /*
1973                  * lock not needed as this is only incremented in the single
1974                  * threaded kernel __init code path all other access are read
1975                  * only
1976                  */
1977         }
1978
1979         g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
1980                         GFP_KERNEL);
1981         if (!g_iommus) {
1982                 printk(KERN_ERR "Allocating global iommu array failed\n");
1983                 ret = -ENOMEM;
1984                 goto error;
1985         }
1986
1987         deferred_flush = kzalloc(g_num_of_iommus *
1988                 sizeof(struct deferred_flush_tables), GFP_KERNEL);
1989         if (!deferred_flush) {
1990                 kfree(g_iommus);
1991                 ret = -ENOMEM;
1992                 goto error;
1993         }
1994
1995         for_each_drhd_unit(drhd) {
1996                 if (drhd->ignored)
1997                         continue;
1998
1999                 iommu = drhd->iommu;
2000                 g_iommus[iommu->seq_id] = iommu;
2001
2002                 ret = iommu_init_domains(iommu);
2003                 if (ret)
2004                         goto error;
2005
2006                 /*
2007                  * TBD:
2008                  * we could share the same root & context tables
2009                  * amoung all IOMMU's. Need to Split it later.
2010                  */
2011                 ret = iommu_alloc_root_entry(iommu);
2012                 if (ret) {
2013                         printk(KERN_ERR "IOMMU: allocate root entry failed\n");
2014                         goto error;
2015                 }
2016                 if (!ecap_pass_through(iommu->ecap))
2017                         pass_through = 0;
2018         }
2019         if (iommu_pass_through)
2020                 if (!pass_through) {
2021                         printk(KERN_INFO
2022                                "Pass Through is not supported by hardware.\n");
2023                         iommu_pass_through = 0;
2024                 }
2025
2026         /*
2027          * Start from the sane iommu hardware state.
2028          */
2029         for_each_drhd_unit(drhd) {
2030                 if (drhd->ignored)
2031                         continue;
2032
2033                 iommu = drhd->iommu;
2034
2035                 /*
2036                  * If the queued invalidation is already initialized by us
2037                  * (for example, while enabling interrupt-remapping) then
2038                  * we got the things already rolling from a sane state.
2039                  */
2040                 if (iommu->qi)
2041                         continue;
2042
2043                 /*
2044                  * Clear any previous faults.
2045                  */
2046                 dmar_fault(-1, iommu);
2047                 /*
2048                  * Disable queued invalidation if supported and already enabled
2049                  * before OS handover.
2050                  */
2051                 dmar_disable_qi(iommu);
2052         }
2053
2054         for_each_drhd_unit(drhd) {
2055                 if (drhd->ignored)
2056                         continue;
2057
2058                 iommu = drhd->iommu;
2059
2060                 if (dmar_enable_qi(iommu)) {
2061                         /*
2062                          * Queued Invalidate not enabled, use Register Based
2063                          * Invalidate
2064                          */
2065                         iommu->flush.flush_context = __iommu_flush_context;
2066                         iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2067                         printk(KERN_INFO "IOMMU 0x%Lx: using Register based "
2068                                "invalidation\n",
2069                                (unsigned long long)drhd->reg_base_addr);
2070                 } else {
2071                         iommu->flush.flush_context = qi_flush_context;
2072                         iommu->flush.flush_iotlb = qi_flush_iotlb;
2073                         printk(KERN_INFO "IOMMU 0x%Lx: using Queued "
2074                                "invalidation\n",
2075                                (unsigned long long)drhd->reg_base_addr);
2076                 }
2077         }
2078
2079         /*
2080          * If pass through is set and enabled, context entries of all pci
2081          * devices are intialized by pass through translation type.
2082          */
2083         if (iommu_pass_through) {
2084                 ret = init_context_pass_through();
2085                 if (ret) {
2086                         printk(KERN_ERR "IOMMU: Pass through init failed.\n");
2087                         iommu_pass_through = 0;
2088                 }
2089         }
2090
2091         /*
2092          * If pass through is not set or not enabled, setup context entries for
2093          * identity mappings for rmrr, gfx, and isa.
2094          */
2095         if (!iommu_pass_through) {
2096                 /*
2097                  * For each rmrr
2098                  *   for each dev attached to rmrr
2099                  *   do
2100                  *     locate drhd for dev, alloc domain for dev
2101                  *     allocate free domain
2102                  *     allocate page table entries for rmrr
2103                  *     if context not allocated for bus
2104                  *           allocate and init context
2105                  *           set present in root table for this bus
2106                  *     init context with domain, translation etc
2107                  *    endfor
2108                  * endfor
2109                  */
2110                 for_each_rmrr_units(rmrr) {
2111                         for (i = 0; i < rmrr->devices_cnt; i++) {
2112                                 pdev = rmrr->devices[i];
2113                                 /*
2114                                  * some BIOS lists non-exist devices in DMAR
2115                                  * table.
2116                                  */
2117                                 if (!pdev)
2118                                         continue;
2119                                 ret = iommu_prepare_rmrr_dev(rmrr, pdev);
2120                                 if (ret)
2121                                         printk(KERN_ERR
2122                                  "IOMMU: mapping reserved region failed\n");
2123                         }
2124                 }
2125
2126                 iommu_prepare_gfx_mapping();
2127
2128                 iommu_prepare_isa();
2129         }
2130
2131         /*
2132          * for each drhd
2133          *   enable fault log
2134          *   global invalidate context cache
2135          *   global invalidate iotlb
2136          *   enable translation
2137          */
2138         for_each_drhd_unit(drhd) {
2139                 if (drhd->ignored)
2140                         continue;
2141                 iommu = drhd->iommu;
2142
2143                 iommu_flush_write_buffer(iommu);
2144
2145                 ret = dmar_set_interrupt(iommu);
2146                 if (ret)
2147                         goto error;
2148
2149                 iommu_set_root_entry(iommu);
2150
2151                 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
2152                 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2153                 iommu_disable_protect_mem_regions(iommu);
2154
2155                 ret = iommu_enable_translation(iommu);
2156                 if (ret)
2157                         goto error;
2158         }
2159
2160         return 0;
2161 error:
2162         for_each_drhd_unit(drhd) {
2163                 if (drhd->ignored)
2164                         continue;
2165                 iommu = drhd->iommu;
2166                 free_iommu(iommu);
2167         }
2168         kfree(g_iommus);
2169         return ret;
2170 }
2171
2172 static inline u64 aligned_size(u64 host_addr, size_t size)
2173 {
2174         u64 addr;
2175         addr = (host_addr & (~PAGE_MASK)) + size;
2176         return PAGE_ALIGN(addr);
2177 }
2178
2179 struct iova *
2180 iommu_alloc_iova(struct dmar_domain *domain, size_t size, u64 end)
2181 {
2182         struct iova *piova;
2183
2184         /* Make sure it's in range */
2185         end = min_t(u64, DOMAIN_MAX_ADDR(domain->gaw), end);
2186         if (!size || (IOVA_START_ADDR + size > end))
2187                 return NULL;
2188
2189         piova = alloc_iova(&domain->iovad,
2190                         size >> PAGE_SHIFT, IOVA_PFN(end), 1);
2191         return piova;
2192 }
2193
2194 static struct iova *
2195 __intel_alloc_iova(struct device *dev, struct dmar_domain *domain,
2196                    size_t size, u64 dma_mask)
2197 {
2198         struct pci_dev *pdev = to_pci_dev(dev);
2199         struct iova *iova = NULL;
2200
2201         if (dma_mask <= DMA_BIT_MASK(32) || dmar_forcedac)
2202                 iova = iommu_alloc_iova(domain, size, dma_mask);
2203         else {
2204                 /*
2205                  * First try to allocate an io virtual address in
2206                  * DMA_BIT_MASK(32) and if that fails then try allocating
2207                  * from higher range
2208                  */
2209                 iova = iommu_alloc_iova(domain, size, DMA_BIT_MASK(32));
2210                 if (!iova)
2211                         iova = iommu_alloc_iova(domain, size, dma_mask);
2212         }
2213
2214         if (!iova) {
2215                 printk(KERN_ERR"Allocating iova for %s failed", pci_name(pdev));
2216                 return NULL;
2217         }
2218
2219         return iova;
2220 }
2221
2222 static struct dmar_domain *
2223 get_valid_domain_for_dev(struct pci_dev *pdev)
2224 {
2225         struct dmar_domain *domain;
2226         int ret;
2227
2228         domain = get_domain_for_dev(pdev,
2229                         DEFAULT_DOMAIN_ADDRESS_WIDTH);
2230         if (!domain) {
2231                 printk(KERN_ERR
2232                         "Allocating domain for %s failed", pci_name(pdev));
2233                 return NULL;
2234         }
2235
2236         /* make sure context mapping is ok */
2237         if (unlikely(!domain_context_mapped(pdev))) {
2238                 ret = domain_context_mapping(domain, pdev,
2239                                              CONTEXT_TT_MULTI_LEVEL);
2240                 if (ret) {
2241                         printk(KERN_ERR
2242                                 "Domain context map for %s failed",
2243                                 pci_name(pdev));
2244                         return NULL;
2245                 }
2246         }
2247
2248         return domain;
2249 }
2250
2251 static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
2252                                      size_t size, int dir, u64 dma_mask)
2253 {
2254         struct pci_dev *pdev = to_pci_dev(hwdev);
2255         struct dmar_domain *domain;
2256         phys_addr_t start_paddr;
2257         struct iova *iova;
2258         int prot = 0;
2259         int ret;
2260         struct intel_iommu *iommu;
2261
2262         BUG_ON(dir == DMA_NONE);
2263         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2264                 return paddr;
2265
2266         domain = get_valid_domain_for_dev(pdev);
2267         if (!domain)
2268                 return 0;
2269
2270         iommu = domain_get_iommu(domain);
2271         size = aligned_size((u64)paddr, size);
2272
2273         iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2274         if (!iova)
2275                 goto error;
2276
2277         start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
2278
2279         /*
2280          * Check if DMAR supports zero-length reads on write only
2281          * mappings..
2282          */
2283         if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2284                         !cap_zlr(iommu->cap))
2285                 prot |= DMA_PTE_READ;
2286         if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2287                 prot |= DMA_PTE_WRITE;
2288         /*
2289          * paddr - (paddr + size) might be partial page, we should map the whole
2290          * page.  Note: if two part of one page are separately mapped, we
2291          * might have two guest_addr mapping to the same host paddr, but this
2292          * is not a big problem
2293          */
2294         ret = domain_page_mapping(domain, start_paddr,
2295                                   ((u64)paddr) & PHYSICAL_PAGE_MASK,
2296                                   size, prot);
2297         if (ret)
2298                 goto error;
2299
2300         /* it's a non-present to present mapping. Only flush if caching mode */
2301         if (cap_caching_mode(iommu->cap))
2302                 iommu_flush_iotlb_psi(iommu, 0, start_paddr,
2303                                       size >> VTD_PAGE_SHIFT);
2304         else
2305                 iommu_flush_write_buffer(iommu);
2306
2307         return start_paddr + ((u64)paddr & (~PAGE_MASK));
2308
2309 error:
2310         if (iova)
2311                 __free_iova(&domain->iovad, iova);
2312         printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
2313                 pci_name(pdev), size, (unsigned long long)paddr, dir);
2314         return 0;
2315 }
2316
2317 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
2318                                  unsigned long offset, size_t size,
2319                                  enum dma_data_direction dir,
2320                                  struct dma_attrs *attrs)
2321 {
2322         return __intel_map_single(dev, page_to_phys(page) + offset, size,
2323                                   dir, to_pci_dev(dev)->dma_mask);
2324 }
2325
2326 static void flush_unmaps(void)
2327 {
2328         int i, j;
2329
2330         timer_on = 0;
2331
2332         /* just flush them all */
2333         for (i = 0; i < g_num_of_iommus; i++) {
2334                 struct intel_iommu *iommu = g_iommus[i];
2335                 if (!iommu)
2336                         continue;
2337
2338                 if (!deferred_flush[i].next)
2339                         continue;
2340
2341                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2342                                          DMA_TLB_GLOBAL_FLUSH);
2343                 for (j = 0; j < deferred_flush[i].next; j++) {
2344                         unsigned long mask;
2345                         struct iova *iova = deferred_flush[i].iova[j];
2346
2347                         mask = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT;
2348                         mask = ilog2(mask >> VTD_PAGE_SHIFT);
2349                         iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
2350                                         iova->pfn_lo << PAGE_SHIFT, mask);
2351                         __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
2352                 }
2353                 deferred_flush[i].next = 0;
2354         }
2355
2356         list_size = 0;
2357 }
2358
2359 static void flush_unmaps_timeout(unsigned long data)
2360 {
2361         unsigned long flags;
2362
2363         spin_lock_irqsave(&async_umap_flush_lock, flags);
2364         flush_unmaps();
2365         spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2366 }
2367
2368 static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2369 {
2370         unsigned long flags;
2371         int next, iommu_id;
2372         struct intel_iommu *iommu;
2373
2374         spin_lock_irqsave(&async_umap_flush_lock, flags);
2375         if (list_size == HIGH_WATER_MARK)
2376                 flush_unmaps();
2377
2378         iommu = domain_get_iommu(dom);
2379         iommu_id = iommu->seq_id;
2380
2381         next = deferred_flush[iommu_id].next;
2382         deferred_flush[iommu_id].domain[next] = dom;
2383         deferred_flush[iommu_id].iova[next] = iova;
2384         deferred_flush[iommu_id].next++;
2385
2386         if (!timer_on) {
2387                 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2388                 timer_on = 1;
2389         }
2390         list_size++;
2391         spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2392 }
2393
2394 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
2395                              size_t size, enum dma_data_direction dir,
2396                              struct dma_attrs *attrs)
2397 {
2398         struct pci_dev *pdev = to_pci_dev(dev);
2399         struct dmar_domain *domain;
2400         unsigned long start_addr;
2401         struct iova *iova;
2402         struct intel_iommu *iommu;
2403
2404         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2405                 return;
2406         domain = find_domain(pdev);
2407         BUG_ON(!domain);
2408
2409         iommu = domain_get_iommu(domain);
2410
2411         iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
2412         if (!iova)
2413                 return;
2414
2415         start_addr = iova->pfn_lo << PAGE_SHIFT;
2416         size = aligned_size((u64)dev_addr, size);
2417
2418         pr_debug("Device %s unmapping: %zx@%llx\n",
2419                 pci_name(pdev), size, (unsigned long long)start_addr);
2420
2421         /*  clear the whole page */
2422         dma_pte_clear_range(domain, start_addr, start_addr + size);
2423         /* free page tables */
2424         dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2425         if (intel_iommu_strict) {
2426                 iommu_flush_iotlb_psi(iommu, domain->id, start_addr,
2427                                       size >> VTD_PAGE_SHIFT);
2428                 /* free iova */
2429                 __free_iova(&domain->iovad, iova);
2430         } else {
2431                 add_unmap(domain, iova);
2432                 /*
2433                  * queue up the release of the unmap to save the 1/6th of the
2434                  * cpu used up by the iotlb flush operation...
2435                  */
2436         }
2437 }
2438
2439 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr, size_t size,
2440                                int dir)
2441 {
2442         intel_unmap_page(dev, dev_addr, size, dir, NULL);
2443 }
2444
2445 static void *intel_alloc_coherent(struct device *hwdev, size_t size,
2446                                   dma_addr_t *dma_handle, gfp_t flags)
2447 {
2448         void *vaddr;
2449         int order;
2450
2451         size = PAGE_ALIGN(size);
2452         order = get_order(size);
2453         flags &= ~(GFP_DMA | GFP_DMA32);
2454
2455         vaddr = (void *)__get_free_pages(flags, order);
2456         if (!vaddr)
2457                 return NULL;
2458         memset(vaddr, 0, size);
2459
2460         *dma_handle = __intel_map_single(hwdev, virt_to_bus(vaddr), size,
2461                                          DMA_BIDIRECTIONAL,
2462                                          hwdev->coherent_dma_mask);
2463         if (*dma_handle)
2464                 return vaddr;
2465         free_pages((unsigned long)vaddr, order);
2466         return NULL;
2467 }
2468
2469 static void intel_free_coherent(struct device *hwdev, size_t size, void *vaddr,
2470                                 dma_addr_t dma_handle)
2471 {
2472         int order;
2473
2474         size = PAGE_ALIGN(size);
2475         order = get_order(size);
2476
2477         intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2478         free_pages((unsigned long)vaddr, order);
2479 }
2480
2481 static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2482                            int nelems, enum dma_data_direction dir,
2483                            struct dma_attrs *attrs)
2484 {
2485         int i;
2486         struct pci_dev *pdev = to_pci_dev(hwdev);
2487         struct dmar_domain *domain;
2488         unsigned long start_addr;
2489         struct iova *iova;
2490         size_t size = 0;
2491         phys_addr_t addr;
2492         struct scatterlist *sg;
2493         struct intel_iommu *iommu;
2494
2495         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2496                 return;
2497
2498         domain = find_domain(pdev);
2499         BUG_ON(!domain);
2500
2501         iommu = domain_get_iommu(domain);
2502
2503         iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
2504         if (!iova)
2505                 return;
2506         for_each_sg(sglist, sg, nelems, i) {
2507                 addr = page_to_phys(sg_page(sg)) + sg->offset;
2508                 size += aligned_size((u64)addr, sg->length);
2509         }
2510
2511         start_addr = iova->pfn_lo << PAGE_SHIFT;
2512
2513         /*  clear the whole page */
2514         dma_pte_clear_range(domain, start_addr, start_addr + size);
2515         /* free page tables */
2516         dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2517
2518         iommu_flush_iotlb_psi(iommu, domain->id, start_addr,
2519                               size >> VTD_PAGE_SHIFT);
2520
2521         /* free iova */
2522         __free_iova(&domain->iovad, iova);
2523 }
2524
2525 static int intel_nontranslate_map_sg(struct device *hddev,
2526         struct scatterlist *sglist, int nelems, int dir)
2527 {
2528         int i;
2529         struct scatterlist *sg;
2530
2531         for_each_sg(sglist, sg, nelems, i) {
2532                 BUG_ON(!sg_page(sg));
2533                 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
2534                 sg->dma_length = sg->length;
2535         }
2536         return nelems;
2537 }
2538
2539 static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist, int nelems,
2540                         enum dma_data_direction dir, struct dma_attrs *attrs)
2541 {
2542         phys_addr_t addr;
2543         int i;
2544         struct pci_dev *pdev = to_pci_dev(hwdev);
2545         struct dmar_domain *domain;
2546         size_t size = 0;
2547         int prot = 0;
2548         size_t offset = 0;
2549         struct iova *iova = NULL;
2550         int ret;
2551         struct scatterlist *sg;
2552         unsigned long start_addr;
2553         struct intel_iommu *iommu;
2554
2555         BUG_ON(dir == DMA_NONE);
2556         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2557                 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
2558
2559         domain = get_valid_domain_for_dev(pdev);
2560         if (!domain)
2561                 return 0;
2562
2563         iommu = domain_get_iommu(domain);
2564
2565         for_each_sg(sglist, sg, nelems, i) {
2566                 addr = page_to_phys(sg_page(sg)) + sg->offset;
2567                 size += aligned_size((u64)addr, sg->length);
2568         }
2569
2570         iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2571         if (!iova) {
2572                 sglist->dma_length = 0;
2573                 return 0;
2574         }
2575
2576         /*
2577          * Check if DMAR supports zero-length reads on write only
2578          * mappings..
2579          */
2580         if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2581                         !cap_zlr(iommu->cap))
2582                 prot |= DMA_PTE_READ;
2583         if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2584                 prot |= DMA_PTE_WRITE;
2585
2586         start_addr = iova->pfn_lo << PAGE_SHIFT;
2587         offset = 0;
2588         for_each_sg(sglist, sg, nelems, i) {
2589                 addr = page_to_phys(sg_page(sg)) + sg->offset;
2590                 size = aligned_size((u64)addr, sg->length);
2591                 ret = domain_page_mapping(domain, start_addr + offset,
2592                                           ((u64)addr) & PHYSICAL_PAGE_MASK,
2593                                           size, prot);
2594                 if (ret) {
2595                         /*  clear the page */
2596                         dma_pte_clear_range(domain, start_addr,
2597                                   start_addr + offset);
2598                         /* free page tables */
2599                         dma_pte_free_pagetable(domain, start_addr,
2600                                   start_addr + offset);
2601                         /* free iova */
2602                         __free_iova(&domain->iovad, iova);
2603                         return 0;
2604                 }
2605                 sg->dma_address = start_addr + offset +
2606                                 ((u64)addr & (~PAGE_MASK));
2607                 sg->dma_length = sg->length;
2608                 offset += size;
2609         }
2610
2611         /* it's a non-present to present mapping. Only flush if caching mode */
2612         if (cap_caching_mode(iommu->cap))
2613                 iommu_flush_iotlb_psi(iommu, 0, start_addr,
2614                                       offset >> VTD_PAGE_SHIFT);
2615         else
2616                 iommu_flush_write_buffer(iommu);
2617
2618         return nelems;
2619 }
2620
2621 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
2622 {
2623         return !dma_addr;
2624 }
2625
2626 struct dma_map_ops intel_dma_ops = {
2627         .alloc_coherent = intel_alloc_coherent,
2628         .free_coherent = intel_free_coherent,
2629         .map_sg = intel_map_sg,
2630         .unmap_sg = intel_unmap_sg,
2631         .map_page = intel_map_page,
2632         .unmap_page = intel_unmap_page,
2633         .mapping_error = intel_mapping_error,
2634 };
2635
2636 static inline int iommu_domain_cache_init(void)
2637 {
2638         int ret = 0;
2639
2640         iommu_domain_cache = kmem_cache_create("iommu_domain",
2641                                          sizeof(struct dmar_domain),
2642                                          0,
2643                                          SLAB_HWCACHE_ALIGN,
2644
2645                                          NULL);
2646         if (!iommu_domain_cache) {
2647                 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2648                 ret = -ENOMEM;
2649         }
2650
2651         return ret;
2652 }
2653
2654 static inline int iommu_devinfo_cache_init(void)
2655 {
2656         int ret = 0;
2657
2658         iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2659                                          sizeof(struct device_domain_info),
2660                                          0,
2661                                          SLAB_HWCACHE_ALIGN,
2662                                          NULL);
2663         if (!iommu_devinfo_cache) {
2664                 printk(KERN_ERR "Couldn't create devinfo cache\n");
2665                 ret = -ENOMEM;
2666         }
2667
2668         return ret;
2669 }
2670
2671 static inline int iommu_iova_cache_init(void)
2672 {
2673         int ret = 0;
2674
2675         iommu_iova_cache = kmem_cache_create("iommu_iova",
2676                                          sizeof(struct iova),
2677                                          0,
2678                                          SLAB_HWCACHE_ALIGN,
2679                                          NULL);
2680         if (!iommu_iova_cache) {
2681                 printk(KERN_ERR "Couldn't create iova cache\n");
2682                 ret = -ENOMEM;
2683         }
2684
2685         return ret;
2686 }
2687
2688 static int __init iommu_init_mempool(void)
2689 {
2690         int ret;
2691         ret = iommu_iova_cache_init();
2692         if (ret)
2693                 return ret;
2694
2695         ret = iommu_domain_cache_init();
2696         if (ret)
2697                 goto domain_error;
2698
2699         ret = iommu_devinfo_cache_init();
2700         if (!ret)
2701                 return ret;
2702
2703         kmem_cache_destroy(iommu_domain_cache);
2704 domain_error:
2705         kmem_cache_destroy(iommu_iova_cache);
2706
2707         return -ENOMEM;
2708 }
2709
2710 static void __init iommu_exit_mempool(void)
2711 {
2712         kmem_cache_destroy(iommu_devinfo_cache);
2713         kmem_cache_destroy(iommu_domain_cache);
2714         kmem_cache_destroy(iommu_iova_cache);
2715
2716 }
2717
2718 static void __init init_no_remapping_devices(void)
2719 {
2720         struct dmar_drhd_unit *drhd;
2721
2722         for_each_drhd_unit(drhd) {
2723                 if (!drhd->include_all) {
2724                         int i;
2725                         for (i = 0; i < drhd->devices_cnt; i++)
2726                                 if (drhd->devices[i] != NULL)
2727                                         break;
2728                         /* ignore DMAR unit if no pci devices exist */
2729                         if (i == drhd->devices_cnt)
2730                                 drhd->ignored = 1;
2731                 }
2732         }
2733
2734         if (dmar_map_gfx)
2735                 return;
2736
2737         for_each_drhd_unit(drhd) {
2738                 int i;
2739                 if (drhd->ignored || drhd->include_all)
2740                         continue;
2741
2742                 for (i = 0; i < drhd->devices_cnt; i++)
2743                         if (drhd->devices[i] &&
2744                                 !IS_GFX_DEVICE(drhd->devices[i]))
2745                                 break;
2746
2747                 if (i < drhd->devices_cnt)
2748                         continue;
2749
2750                 /* bypass IOMMU if it is just for gfx devices */
2751                 drhd->ignored = 1;
2752                 for (i = 0; i < drhd->devices_cnt; i++) {
2753                         if (!drhd->devices[i])
2754                                 continue;
2755                         drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
2756                 }
2757         }
2758 }
2759
2760 #ifdef CONFIG_SUSPEND
2761 static int init_iommu_hw(void)
2762 {
2763         struct dmar_drhd_unit *drhd;
2764         struct intel_iommu *iommu = NULL;
2765
2766         for_each_active_iommu(iommu, drhd)
2767                 if (iommu->qi)
2768                         dmar_reenable_qi(iommu);
2769
2770         for_each_active_iommu(iommu, drhd) {
2771                 iommu_flush_write_buffer(iommu);
2772
2773                 iommu_set_root_entry(iommu);
2774
2775                 iommu->flush.flush_context(iommu, 0, 0, 0,
2776                                            DMA_CCMD_GLOBAL_INVL);
2777                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2778                                          DMA_TLB_GLOBAL_FLUSH);
2779                 iommu_disable_protect_mem_regions(iommu);
2780                 iommu_enable_translation(iommu);
2781         }
2782
2783         return 0;
2784 }
2785
2786 static void iommu_flush_all(void)
2787 {
2788         struct dmar_drhd_unit *drhd;
2789         struct intel_iommu *iommu;
2790
2791         for_each_active_iommu(iommu, drhd) {
2792                 iommu->flush.flush_context(iommu, 0, 0, 0,
2793                                            DMA_CCMD_GLOBAL_INVL);
2794                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2795                                          DMA_TLB_GLOBAL_FLUSH);
2796         }
2797 }
2798
2799 static int iommu_suspend(struct sys_device *dev, pm_message_t state)
2800 {
2801         struct dmar_drhd_unit *drhd;
2802         struct intel_iommu *iommu = NULL;
2803         unsigned long flag;
2804
2805         for_each_active_iommu(iommu, drhd) {
2806                 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
2807                                                  GFP_ATOMIC);
2808                 if (!iommu->iommu_state)
2809                         goto nomem;
2810         }
2811
2812         iommu_flush_all();
2813
2814         for_each_active_iommu(iommu, drhd) {
2815                 iommu_disable_translation(iommu);
2816
2817                 spin_lock_irqsave(&iommu->register_lock, flag);
2818
2819                 iommu->iommu_state[SR_DMAR_FECTL_REG] =
2820                         readl(iommu->reg + DMAR_FECTL_REG);
2821                 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
2822                         readl(iommu->reg + DMAR_FEDATA_REG);
2823                 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
2824                         readl(iommu->reg + DMAR_FEADDR_REG);
2825                 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
2826                         readl(iommu->reg + DMAR_FEUADDR_REG);
2827
2828                 spin_unlock_irqrestore(&iommu->register_lock, flag);
2829         }
2830         return 0;
2831
2832 nomem:
2833         for_each_active_iommu(iommu, drhd)
2834                 kfree(iommu->iommu_state);
2835
2836         return -ENOMEM;
2837 }
2838
2839 static int iommu_resume(struct sys_device *dev)
2840 {
2841         struct dmar_drhd_unit *drhd;
2842         struct intel_iommu *iommu = NULL;
2843         unsigned long flag;
2844
2845         if (init_iommu_hw()) {
2846                 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
2847                 return -EIO;
2848         }
2849
2850         for_each_active_iommu(iommu, drhd) {
2851
2852                 spin_lock_irqsave(&iommu->register_lock, flag);
2853
2854                 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
2855                         iommu->reg + DMAR_FECTL_REG);
2856                 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
2857                         iommu->reg + DMAR_FEDATA_REG);
2858                 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
2859                         iommu->reg + DMAR_FEADDR_REG);
2860                 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
2861                         iommu->reg + DMAR_FEUADDR_REG);
2862
2863                 spin_unlock_irqrestore(&iommu->register_lock, flag);
2864         }
2865
2866         for_each_active_iommu(iommu, drhd)
2867                 kfree(iommu->iommu_state);
2868
2869         return 0;
2870 }
2871
2872 static struct sysdev_class iommu_sysclass = {
2873         .name           = "iommu",
2874         .resume         = iommu_resume,
2875         .suspend        = iommu_suspend,
2876 };
2877
2878 static struct sys_device device_iommu = {
2879         .cls    = &iommu_sysclass,
2880 };
2881
2882 static int __init init_iommu_sysfs(void)
2883 {
2884         int error;
2885
2886         error = sysdev_class_register(&iommu_sysclass);
2887         if (error)
2888                 return error;
2889
2890         error = sysdev_register(&device_iommu);
2891         if (error)
2892                 sysdev_class_unregister(&iommu_sysclass);
2893
2894         return error;
2895 }
2896
2897 #else
2898 static int __init init_iommu_sysfs(void)
2899 {
2900         return 0;
2901 }
2902 #endif  /* CONFIG_PM */
2903
2904 int __init intel_iommu_init(void)
2905 {
2906         int ret = 0;
2907
2908         if (dmar_table_init())
2909                 return  -ENODEV;
2910
2911         if (dmar_dev_scope_init())
2912                 return  -ENODEV;
2913
2914         /*
2915          * Check the need for DMA-remapping initialization now.
2916          * Above initialization will also be used by Interrupt-remapping.
2917          */
2918         if (no_iommu || (swiotlb && !iommu_pass_through) || dmar_disabled)
2919                 return -ENODEV;
2920
2921         iommu_init_mempool();
2922         dmar_init_reserved_ranges();
2923
2924         init_no_remapping_devices();
2925
2926         ret = init_dmars();
2927         if (ret) {
2928                 printk(KERN_ERR "IOMMU: dmar init failed\n");
2929                 put_iova_domain(&reserved_iova_list);
2930                 iommu_exit_mempool();
2931                 return ret;
2932         }
2933         printk(KERN_INFO
2934         "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
2935
2936         init_timer(&unmap_timer);
2937         force_iommu = 1;
2938
2939         if (!iommu_pass_through) {
2940                 printk(KERN_INFO
2941                        "Multi-level page-table translation for DMAR.\n");
2942                 dma_ops = &intel_dma_ops;
2943         } else
2944                 printk(KERN_INFO
2945                        "DMAR: Pass through translation for DMAR.\n");
2946
2947         init_iommu_sysfs();
2948
2949         register_iommu(&intel_iommu_ops);
2950
2951         return 0;
2952 }
2953
2954 static int vm_domain_add_dev_info(struct dmar_domain *domain,
2955                                   struct pci_dev *pdev)
2956 {
2957         struct device_domain_info *info;
2958         unsigned long flags;
2959
2960         info = alloc_devinfo_mem();
2961         if (!info)
2962                 return -ENOMEM;
2963
2964         info->segment = pci_domain_nr(pdev->bus);
2965         info->bus = pdev->bus->number;
2966         info->devfn = pdev->devfn;
2967         info->dev = pdev;
2968         info->domain = domain;
2969
2970         spin_lock_irqsave(&device_domain_lock, flags);
2971         list_add(&info->link, &domain->devices);
2972         list_add(&info->global, &device_domain_list);
2973         pdev->dev.archdata.iommu = info;
2974         spin_unlock_irqrestore(&device_domain_lock, flags);
2975
2976         return 0;
2977 }
2978
2979 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
2980                                            struct pci_dev *pdev)
2981 {
2982         struct pci_dev *tmp, *parent;
2983
2984         if (!iommu || !pdev)
2985                 return;
2986
2987         /* dependent device detach */
2988         tmp = pci_find_upstream_pcie_bridge(pdev);
2989         /* Secondary interface's bus number and devfn 0 */
2990         if (tmp) {
2991                 parent = pdev->bus->self;
2992                 while (parent != tmp) {
2993                         iommu_detach_dev(iommu, parent->bus->number,
2994                                          parent->devfn);
2995                         parent = parent->bus->self;
2996                 }
2997                 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
2998                         iommu_detach_dev(iommu,
2999                                 tmp->subordinate->number, 0);
3000                 else /* this is a legacy PCI bridge */
3001                         iommu_detach_dev(iommu, tmp->bus->number,
3002                                          tmp->devfn);
3003         }
3004 }
3005
3006 static void vm_domain_remove_one_dev_info(struct dmar_domain *domain,
3007                                           struct pci_dev *pdev)
3008 {
3009         struct device_domain_info *info;
3010         struct intel_iommu *iommu;
3011         unsigned long flags;
3012         int found = 0;
3013         struct list_head *entry, *tmp;
3014
3015         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3016                                 pdev->devfn);
3017         if (!iommu)
3018                 return;
3019
3020         spin_lock_irqsave(&device_domain_lock, flags);
3021         list_for_each_safe(entry, tmp, &domain->devices) {
3022                 info = list_entry(entry, struct device_domain_info, link);
3023                 /* No need to compare PCI domain; it has to be the same */
3024                 if (info->bus == pdev->bus->number &&
3025                     info->devfn == pdev->devfn) {
3026                         list_del(&info->link);
3027                         list_del(&info->global);
3028                         if (info->dev)
3029                                 info->dev->dev.archdata.iommu = NULL;
3030                         spin_unlock_irqrestore(&device_domain_lock, flags);
3031
3032                         iommu_disable_dev_iotlb(info);
3033                         iommu_detach_dev(iommu, info->bus, info->devfn);
3034                         iommu_detach_dependent_devices(iommu, pdev);
3035                         free_devinfo_mem(info);
3036
3037                         spin_lock_irqsave(&device_domain_lock, flags);
3038
3039                         if (found)
3040                                 break;
3041                         else
3042                                 continue;
3043                 }
3044
3045                 /* if there is no other devices under the same iommu
3046                  * owned by this domain, clear this iommu in iommu_bmp
3047                  * update iommu count and coherency
3048                  */
3049                 if (iommu == device_to_iommu(info->segment, info->bus,
3050                                             info->devfn))
3051                         found = 1;
3052         }
3053
3054         if (found == 0) {
3055                 unsigned long tmp_flags;
3056                 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
3057                 clear_bit(iommu->seq_id, &domain->iommu_bmp);
3058                 domain->iommu_count--;
3059                 domain_update_iommu_cap(domain);
3060                 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
3061         }
3062
3063         spin_unlock_irqrestore(&device_domain_lock, flags);
3064 }
3065
3066 static void vm_domain_remove_all_dev_info(struct dmar_domain *domain)
3067 {
3068         struct device_domain_info *info;
3069         struct intel_iommu *iommu;
3070         unsigned long flags1, flags2;
3071
3072         spin_lock_irqsave(&device_domain_lock, flags1);
3073         while (!list_empty(&domain->devices)) {
3074                 info = list_entry(domain->devices.next,
3075                         struct device_domain_info, link);
3076                 list_del(&info->link);
3077                 list_del(&info->global);
3078                 if (info->dev)
3079                         info->dev->dev.archdata.iommu = NULL;
3080
3081                 spin_unlock_irqrestore(&device_domain_lock, flags1);
3082
3083                 iommu_disable_dev_iotlb(info);
3084                 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
3085                 iommu_detach_dev(iommu, info->bus, info->devfn);
3086                 iommu_detach_dependent_devices(iommu, info->dev);
3087
3088                 /* clear this iommu in iommu_bmp, update iommu count
3089                  * and capabilities
3090                  */
3091                 spin_lock_irqsave(&domain->iommu_lock, flags2);
3092                 if (test_and_clear_bit(iommu->seq_id,
3093                                        &domain->iommu_bmp)) {
3094                         domain->iommu_count--;
3095                         domain_update_iommu_cap(domain);
3096                 }
3097                 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
3098
3099                 free_devinfo_mem(info);
3100                 spin_lock_irqsave(&device_domain_lock, flags1);
3101         }
3102         spin_unlock_irqrestore(&device_domain_lock, flags1);
3103 }
3104
3105 /* domain id for virtual machine, it won't be set in context */
3106 static unsigned long vm_domid;
3107
3108 static int vm_domain_min_agaw(struct dmar_domain *domain)
3109 {
3110         int i;
3111         int min_agaw = domain->agaw;
3112
3113         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
3114         for (; i < g_num_of_iommus; ) {
3115                 if (min_agaw > g_iommus[i]->agaw)
3116                         min_agaw = g_iommus[i]->agaw;
3117
3118                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
3119         }
3120
3121         return min_agaw;
3122 }
3123
3124 static struct dmar_domain *iommu_alloc_vm_domain(void)
3125 {
3126         struct dmar_domain *domain;
3127
3128         domain = alloc_domain_mem();
3129         if (!domain)
3130                 return NULL;
3131
3132         domain->id = vm_domid++;
3133         memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
3134         domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
3135
3136         return domain;
3137 }
3138
3139 static int vm_domain_init(struct dmar_domain *domain, int guest_width)
3140 {
3141         int adjust_width;
3142
3143         init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
3144         spin_lock_init(&domain->mapping_lock);
3145         spin_lock_init(&domain->iommu_lock);
3146
3147         domain_reserve_special_ranges(domain);
3148
3149         /* calculate AGAW */
3150         domain->gaw = guest_width;
3151         adjust_width = guestwidth_to_adjustwidth(guest_width);
3152         domain->agaw = width_to_agaw(adjust_width);
3153
3154         INIT_LIST_HEAD(&domain->devices);
3155
3156         domain->iommu_count = 0;
3157         domain->iommu_coherency = 0;
3158         domain->max_addr = 0;
3159
3160         /* always allocate the top pgd */
3161         domain->pgd = (struct dma_pte *)alloc_pgtable_page();
3162         if (!domain->pgd)
3163                 return -ENOMEM;
3164         domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
3165         return 0;
3166 }
3167
3168 static void iommu_free_vm_domain(struct dmar_domain *domain)
3169 {
3170         unsigned long flags;
3171         struct dmar_drhd_unit *drhd;
3172         struct intel_iommu *iommu;
3173         unsigned long i;
3174         unsigned long ndomains;
3175
3176         for_each_drhd_unit(drhd) {
3177                 if (drhd->ignored)
3178                         continue;
3179                 iommu = drhd->iommu;
3180
3181                 ndomains = cap_ndoms(iommu->cap);
3182                 i = find_first_bit(iommu->domain_ids, ndomains);
3183                 for (; i < ndomains; ) {
3184                         if (iommu->domains[i] == domain) {
3185                                 spin_lock_irqsave(&iommu->lock, flags);
3186                                 clear_bit(i, iommu->domain_ids);
3187                                 iommu->domains[i] = NULL;
3188                                 spin_unlock_irqrestore(&iommu->lock, flags);
3189                                 break;
3190                         }
3191                         i = find_next_bit(iommu->domain_ids, ndomains, i+1);
3192                 }
3193         }
3194 }
3195
3196 static void vm_domain_exit(struct dmar_domain *domain)
3197 {
3198         u64 end;
3199
3200         /* Domain 0 is reserved, so dont process it */
3201         if (!domain)
3202                 return;
3203
3204         vm_domain_remove_all_dev_info(domain);
3205         /* destroy iovas */
3206         put_iova_domain(&domain->iovad);
3207         end = DOMAIN_MAX_ADDR(domain->gaw);
3208         end = end & (~VTD_PAGE_MASK);
3209
3210         /* clear ptes */
3211         dma_pte_clear_range(domain, 0, end);
3212
3213         /* free page tables */
3214         dma_pte_free_pagetable(domain, 0, end);
3215
3216         iommu_free_vm_domain(domain);
3217         free_domain_mem(domain);
3218 }
3219
3220 static int intel_iommu_domain_init(struct iommu_domain *domain)
3221 {
3222         struct dmar_domain *dmar_domain;
3223
3224         dmar_domain = iommu_alloc_vm_domain();
3225         if (!dmar_domain) {
3226                 printk(KERN_ERR
3227                         "intel_iommu_domain_init: dmar_domain == NULL\n");
3228                 return -ENOMEM;
3229         }
3230         if (vm_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
3231                 printk(KERN_ERR
3232                         "intel_iommu_domain_init() failed\n");
3233                 vm_domain_exit(dmar_domain);
3234                 return -ENOMEM;
3235         }
3236         domain->priv = dmar_domain;
3237
3238         return 0;
3239 }
3240
3241 static void intel_iommu_domain_destroy(struct iommu_domain *domain)
3242 {
3243         struct dmar_domain *dmar_domain = domain->priv;
3244
3245         domain->priv = NULL;
3246         vm_domain_exit(dmar_domain);
3247 }
3248
3249 static int intel_iommu_attach_device(struct iommu_domain *domain,
3250                                      struct device *dev)
3251 {
3252         struct dmar_domain *dmar_domain = domain->priv;
3253         struct pci_dev *pdev = to_pci_dev(dev);
3254         struct intel_iommu *iommu;
3255         int addr_width;
3256         u64 end;
3257         int ret;
3258
3259         /* normally pdev is not mapped */
3260         if (unlikely(domain_context_mapped(pdev))) {
3261                 struct dmar_domain *old_domain;
3262
3263                 old_domain = find_domain(pdev);
3264                 if (old_domain) {
3265                         if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
3266                                 vm_domain_remove_one_dev_info(old_domain, pdev);
3267                         else
3268                                 domain_remove_dev_info(old_domain);
3269                 }
3270         }
3271
3272         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3273                                 pdev->devfn);
3274         if (!iommu)
3275                 return -ENODEV;
3276
3277         /* check if this iommu agaw is sufficient for max mapped address */
3278         addr_width = agaw_to_width(iommu->agaw);
3279         end = DOMAIN_MAX_ADDR(addr_width);
3280         end = end & VTD_PAGE_MASK;
3281         if (end < dmar_domain->max_addr) {
3282                 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3283                        "sufficient for the mapped address (%llx)\n",
3284                        __func__, iommu->agaw, dmar_domain->max_addr);
3285                 return -EFAULT;
3286         }
3287
3288         ret = vm_domain_add_dev_info(dmar_domain, pdev);
3289         if (ret)
3290                 return ret;
3291
3292         ret = domain_context_mapping(dmar_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
3293         return ret;
3294 }
3295
3296 static void intel_iommu_detach_device(struct iommu_domain *domain,
3297                                       struct device *dev)
3298 {
3299         struct dmar_domain *dmar_domain = domain->priv;
3300         struct pci_dev *pdev = to_pci_dev(dev);
3301
3302         vm_domain_remove_one_dev_info(dmar_domain, pdev);
3303 }
3304
3305 static int intel_iommu_map_range(struct iommu_domain *domain,
3306                                  unsigned long iova, phys_addr_t hpa,
3307                                  size_t size, int iommu_prot)
3308 {
3309         struct dmar_domain *dmar_domain = domain->priv;
3310         u64 max_addr;
3311         int addr_width;
3312         int prot = 0;
3313         int ret;
3314
3315         if (iommu_prot & IOMMU_READ)
3316                 prot |= DMA_PTE_READ;
3317         if (iommu_prot & IOMMU_WRITE)
3318                 prot |= DMA_PTE_WRITE;
3319         if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
3320                 prot |= DMA_PTE_SNP;
3321
3322         max_addr = (iova & VTD_PAGE_MASK) + VTD_PAGE_ALIGN(size);
3323         if (dmar_domain->max_addr < max_addr) {
3324                 int min_agaw;
3325                 u64 end;
3326
3327                 /* check if minimum agaw is sufficient for mapped address */
3328                 min_agaw = vm_domain_min_agaw(dmar_domain);
3329                 addr_width = agaw_to_width(min_agaw);
3330                 end = DOMAIN_MAX_ADDR(addr_width);
3331                 end = end & VTD_PAGE_MASK;
3332                 if (end < max_addr) {
3333                         printk(KERN_ERR "%s: iommu agaw (%d) is not "
3334                                "sufficient for the mapped address (%llx)\n",
3335                                __func__, min_agaw, max_addr);
3336                         return -EFAULT;
3337                 }
3338                 dmar_domain->max_addr = max_addr;
3339         }
3340
3341         ret = domain_page_mapping(dmar_domain, iova, hpa, size, prot);
3342         return ret;
3343 }
3344
3345 static void intel_iommu_unmap_range(struct iommu_domain *domain,
3346                                     unsigned long iova, size_t size)
3347 {
3348         struct dmar_domain *dmar_domain = domain->priv;
3349         dma_addr_t base;
3350
3351         /* The address might not be aligned */
3352         base = iova & VTD_PAGE_MASK;
3353         size = VTD_PAGE_ALIGN(size);
3354         dma_pte_clear_range(dmar_domain, base, base + size);
3355
3356         if (dmar_domain->max_addr == base + size)
3357                 dmar_domain->max_addr = base;
3358 }
3359
3360 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
3361                                             unsigned long iova)
3362 {
3363         struct dmar_domain *dmar_domain = domain->priv;
3364         struct dma_pte *pte;
3365         u64 phys = 0;
3366
3367         pte = addr_to_dma_pte(dmar_domain, iova);
3368         if (pte)
3369                 phys = dma_pte_addr(pte);
3370
3371         return phys;
3372 }
3373
3374 static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
3375                                       unsigned long cap)
3376 {
3377         struct dmar_domain *dmar_domain = domain->priv;
3378
3379         if (cap == IOMMU_CAP_CACHE_COHERENCY)
3380                 return dmar_domain->iommu_snooping;
3381
3382         return 0;
3383 }
3384
3385 static struct iommu_ops intel_iommu_ops = {
3386         .domain_init    = intel_iommu_domain_init,
3387         .domain_destroy = intel_iommu_domain_destroy,
3388         .attach_dev     = intel_iommu_attach_device,
3389         .detach_dev     = intel_iommu_detach_device,
3390         .map            = intel_iommu_map_range,
3391         .unmap          = intel_iommu_unmap_range,
3392         .iova_to_phys   = intel_iommu_iova_to_phys,
3393         .domain_has_cap = intel_iommu_domain_has_cap,
3394 };
3395
3396 static void __devinit quirk_iommu_rwbf(struct pci_dev *dev)
3397 {
3398         /*
3399          * Mobile 4 Series Chipset neglects to set RWBF capability,
3400          * but needs it:
3401          */
3402         printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
3403         rwbf_quirk = 1;
3404 }
3405
3406 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);