projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
intel-iommu: Don't set identity mapping for bypassed graphics devices
[deliverable/linux.git] / drivers / pci / intel-iommu.c
CommitLineData
ba395927
KA		 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 *
98bcef56 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>
5b6985ce 21 * Author: Fenghua Yu <fenghua.yu@intel.com>
ba395927
KA		 22 */
23
24#include <linux/init.h>
25#include <linux/bitmap.h>
5e0d2a6f 26#include <linux/debugfs.h>
ba395927
KA		 27#include <linux/slab.h>
28#include <linux/irq.h>
29#include <linux/interrupt.h>
ba395927
KA		 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>
5e0d2a6f 35#include <linux/timer.h>
38717946 36#include <linux/iova.h>
5d450806 37#include <linux/iommu.h>
38717946 38#include <linux/intel-iommu.h>
f59c7b69 39#include <linux/sysdev.h>
ba395927 40#include <asm/cacheflush.h>
46a7fa27 41#include <asm/iommu.h>
ba395927
KA		 42#include "pci.h"
43
5b6985ce
FY		 44#define ROOT_SIZE VTD_PAGE_SIZE
45#define CONTEXT_SIZE VTD_PAGE_SIZE
46
ba395927
KA		 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
4ed0d3e6
FY		 56#define MAX_AGAW_WIDTH 64
57
ba395927 58#define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
595badf5 59#define DOMAIN_MAX_PFN(gaw) ((((u64)1) << (gaw-VTD_PAGE_SHIFT)) - 1)
ba395927 60
f27be03b 61#define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
284901a9 62#define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
6a35528a 63#define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
5e0d2a6f 64
fd18de50 65
dd4e8319
DW		 66/* VT-d pages must always be _smaller_ than MM pages. Otherwise things
67 are never going to work. */
68static inline unsigned long dma_to_mm_pfn(unsigned long dma_pfn)
69{
70 return dma_pfn >> (PAGE_SHIFT - VTD_PAGE_SHIFT);
71}
72
73static inline unsigned long mm_to_dma_pfn(unsigned long mm_pfn)
74{
75 return mm_pfn << (PAGE_SHIFT - VTD_PAGE_SHIFT);
76}
77static inline unsigned long page_to_dma_pfn(struct page *pg)
78{
79 return mm_to_dma_pfn(page_to_pfn(pg));
80}
81static inline unsigned long virt_to_dma_pfn(void *p)
82{
83 return page_to_dma_pfn(virt_to_page(p));
84}
85
d9630fe9
WH		 86/* global iommu list, set NULL for ignored DMAR units */
87static struct intel_iommu **g_iommus;
88
9af88143
DW		 89static int rwbf_quirk;
90
46b08e1a
MM		 91/*
92 * 0: Present
93 * 1-11: Reserved
94 * 12-63: Context Ptr (12 - (haw-1))
95 * 64-127: Reserved
96 */
97struct root_entry {
98 u64 val;
99 u64 rsvd1;
100};
101#define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
102static inline bool root_present(struct root_entry *root)
103{
104 return (root->val & 1);
105}
106static inline void set_root_present(struct root_entry *root)
107{
108 root->val |= 1;
109}
110static inline void set_root_value(struct root_entry *root, unsigned long value)
111{
112 root->val |= value & VTD_PAGE_MASK;
113}
114
115static inline struct context_entry *
116get_context_addr_from_root(struct root_entry *root)
117{
118 return (struct context_entry *)
119 (root_present(root)?phys_to_virt(
120 root->val & VTD_PAGE_MASK) :
121 NULL);
122}
123
7a8fc25e
MM		 124/*
125 * low 64 bits:
126 * 0: present
127 * 1: fault processing disable
128 * 2-3: translation type
129 * 12-63: address space root
130 * high 64 bits:
131 * 0-2: address width
132 * 3-6: aval
133 * 8-23: domain id
134 */
135struct context_entry {
136 u64 lo;
137 u64 hi;
138};
c07e7d21
MM		 139
140static inline bool context_present(struct context_entry *context)
141{
142 return (context->lo & 1);
143}
144static inline void context_set_present(struct context_entry *context)
145{
146 context->lo |= 1;
147}
148
149static inline void context_set_fault_enable(struct context_entry *context)
150{
151 context->lo &= (((u64)-1) << 2) | 1;
152}
153
c07e7d21
MM		 154static inline void context_set_translation_type(struct context_entry *context,
155 unsigned long value)
156{
157 context->lo &= (((u64)-1) << 4) | 3;
158 context->lo |= (value & 3) << 2;
159}
160
161static inline void context_set_address_root(struct context_entry *context,
162 unsigned long value)
163{
164 context->lo |= value & VTD_PAGE_MASK;
165}
166
167static inline void context_set_address_width(struct context_entry *context,
168 unsigned long value)
169{
170 context->hi |= value & 7;
171}
172
173static inline void context_set_domain_id(struct context_entry *context,
174 unsigned long value)
175{
176 context->hi |= (value & ((1 << 16) - 1)) << 8;
177}
178
179static inline void context_clear_entry(struct context_entry *context)
180{
181 context->lo = 0;
182 context->hi = 0;
183}
7a8fc25e 184
622ba12a
MM		 185/*
186 * 0: readable
187 * 1: writable
188 * 2-6: reserved
189 * 7: super page
9cf06697
SY		 190 * 8-10: available
191 * 11: snoop behavior
622ba12a
MM		 192 * 12-63: Host physcial address
193 */
194struct dma_pte {
195 u64 val;
196};
622ba12a 197
19c239ce
MM		 198static inline void dma_clear_pte(struct dma_pte *pte)
199{
200 pte->val = 0;
201}
202
203static inline void dma_set_pte_readable(struct dma_pte *pte)
204{
205 pte->val |= DMA_PTE_READ;
206}
207
208static inline void dma_set_pte_writable(struct dma_pte *pte)
209{
210 pte->val |= DMA_PTE_WRITE;
211}
212
9cf06697
SY		 213static inline void dma_set_pte_snp(struct dma_pte *pte)
214{
215 pte->val |= DMA_PTE_SNP;
216}
217
19c239ce
MM		 218static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)
219{
220 pte->val = (pte->val & ~3) | (prot & 3);
221}
222
223static inline u64 dma_pte_addr(struct dma_pte *pte)
224{
c85994e4
DW		 225#ifdef CONFIG_64BIT
226 return pte->val & VTD_PAGE_MASK;
227#else
228 /* Must have a full atomic 64-bit read */
229 return __cmpxchg64(pte, 0ULL, 0ULL) & VTD_PAGE_MASK;
230#endif
19c239ce
MM		 231}
232
dd4e8319 233static inline void dma_set_pte_pfn(struct dma_pte *pte, unsigned long pfn)
19c239ce 234{
dd4e8319 235 pte->val |= (uint64_t)pfn << VTD_PAGE_SHIFT;
19c239ce
MM		 236}
237
238static inline bool dma_pte_present(struct dma_pte *pte)
239{
240 return (pte->val & 3) != 0;
241}
622ba12a 242
75e6bf96
DW		 243static inline int first_pte_in_page(struct dma_pte *pte)
244{
245 return !((unsigned long)pte & ~VTD_PAGE_MASK);
246}
247
2c2e2c38
FY		 248/*
249 * This domain is a statically identity mapping domain.
250 * 1. This domain creats a static 1:1 mapping to all usable memory.
251 * 2. It maps to each iommu if successful.
252 * 3. Each iommu mapps to this domain if successful.
253 */
254struct dmar_domain *si_domain;
255
3b5410e7 256/* devices under the same p2p bridge are owned in one domain */
cdc7b837 257#define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
3b5410e7 258
1ce28feb
WH		 259/* domain represents a virtual machine, more than one devices
260 * across iommus may be owned in one domain, e.g. kvm guest.
261 */
262#define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 1)
263
2c2e2c38
FY		 264/* si_domain contains mulitple devices */
265#define DOMAIN_FLAG_STATIC_IDENTITY (1 << 2)
266
99126f7c
MM		 267struct dmar_domain {
268 int id; /* domain id */
8c11e798 269 unsigned long iommu_bmp; /* bitmap of iommus this domain uses*/
99126f7c
MM		 270
271 struct list_head devices; /* all devices' list */
272 struct iova_domain iovad; /* iova's that belong to this domain */
273
274 struct dma_pte *pgd; /* virtual address */
99126f7c
MM		 275 int gaw; /* max guest address width */
276
277 /* adjusted guest address width, 0 is level 2 30-bit */
278 int agaw;
279
3b5410e7 280 int flags; /* flags to find out type of domain */
8e604097
WH		 281
282 int iommu_coherency;/* indicate coherency of iommu access */
58c610bd 283 int iommu_snooping; /* indicate snooping control feature*/
c7151a8d
WH		 284 int iommu_count; /* reference count of iommu */
285 spinlock_t iommu_lock; /* protect iommu set in domain */
fe40f1e0 286 u64 max_addr; /* maximum mapped address */
99126f7c
MM		 287};
288
a647dacb
MM		 289/* PCI domain-device relationship */
290struct device_domain_info {
291 struct list_head link; /* link to domain siblings */
292 struct list_head global; /* link to global list */
276dbf99
DW		 293 int segment; /* PCI domain */
294 u8 bus; /* PCI bus number */
a647dacb
MM		 295 u8 devfn; /* PCI devfn number */
296 struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
93a23a72 297 struct intel_iommu *iommu; /* IOMMU used by this device */
a647dacb
MM		 298 struct dmar_domain *domain; /* pointer to domain */
299};
300
5e0d2a6f 301static void flush_unmaps_timeout(unsigned long data);
302
303DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
304
80b20dd8 305#define HIGH_WATER_MARK 250
306struct deferred_flush_tables {
307 int next;
308 struct iova *iova[HIGH_WATER_MARK];
309 struct dmar_domain *domain[HIGH_WATER_MARK];
310};
311
312static struct deferred_flush_tables *deferred_flush;
313
5e0d2a6f 314/* bitmap for indexing intel_iommus */
5e0d2a6f 315static int g_num_of_iommus;
316
317static DEFINE_SPINLOCK(async_umap_flush_lock);
318static LIST_HEAD(unmaps_to_do);
319
320static int timer_on;
321static long list_size;
5e0d2a6f 322
ba395927
KA		 323static void domain_remove_dev_info(struct dmar_domain *domain);
324
0cd5c3c8
KM		 325#ifdef CONFIG_DMAR_DEFAULT_ON
326int dmar_disabled = 0;
327#else
328int dmar_disabled = 1;
329#endif /*CONFIG_DMAR_DEFAULT_ON*/
330
ba395927 331static int __initdata dmar_map_gfx = 1;
7d3b03ce 332static int dmar_forcedac;
5e0d2a6f 333static int intel_iommu_strict;
ba395927
KA		 334
335#define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
336static DEFINE_SPINLOCK(device_domain_lock);
337static LIST_HEAD(device_domain_list);
338
a8bcbb0d
JR		 339static struct iommu_ops intel_iommu_ops;
340
ba395927
KA		 341static int __init intel_iommu_setup(char *str)
342{
343 if (!str)
344 return -EINVAL;
345 while (*str) {
0cd5c3c8
KM		 346 if (!strncmp(str, "on", 2)) {
347 dmar_disabled = 0;
348 printk(KERN_INFO "Intel-IOMMU: enabled\n");
349 } else if (!strncmp(str, "off", 3)) {
ba395927 350 dmar_disabled = 1;
0cd5c3c8 351 printk(KERN_INFO "Intel-IOMMU: disabled\n");
ba395927
KA		 352 } else if (!strncmp(str, "igfx_off", 8)) {
353 dmar_map_gfx = 0;
354 printk(KERN_INFO
355 "Intel-IOMMU: disable GFX device mapping\n");
7d3b03ce 356 } else if (!strncmp(str, "forcedac", 8)) {
5e0d2a6f 357 printk(KERN_INFO
7d3b03ce
KA		 358 "Intel-IOMMU: Forcing DAC for PCI devices\n");
359 dmar_forcedac = 1;
5e0d2a6f 360 } else if (!strncmp(str, "strict", 6)) {
361 printk(KERN_INFO
362 "Intel-IOMMU: disable batched IOTLB flush\n");
363 intel_iommu_strict = 1;
ba395927
KA		 364 }
365
366 str += strcspn(str, ",");
367 while (*str == ',')
368 str++;
369 }
370 return 0;
371}
372__setup("intel_iommu=", intel_iommu_setup);
373
374static struct kmem_cache *iommu_domain_cache;
375static struct kmem_cache *iommu_devinfo_cache;
376static struct kmem_cache *iommu_iova_cache;
377
eb3fa7cb
KA		 378static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
379{
380 unsigned int flags;
381 void *vaddr;
382
383 /* trying to avoid low memory issues */
384 flags = current->flags & PF_MEMALLOC;
385 current->flags |= PF_MEMALLOC;
386 vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
387 current->flags &= (~PF_MEMALLOC | flags);
388 return vaddr;
389}
390
391
ba395927
KA		 392static inline void *alloc_pgtable_page(void)
393{
eb3fa7cb
KA		 394 unsigned int flags;
395 void *vaddr;
396
397 /* trying to avoid low memory issues */
398 flags = current->flags & PF_MEMALLOC;
399 current->flags |= PF_MEMALLOC;
400 vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
401 current->flags &= (~PF_MEMALLOC | flags);
402 return vaddr;
ba395927
KA		 403}
404
405static inline void free_pgtable_page(void *vaddr)
406{
407 free_page((unsigned long)vaddr);
408}
409
410static inline void *alloc_domain_mem(void)
411{
eb3fa7cb 412 return iommu_kmem_cache_alloc(iommu_domain_cache);
ba395927
KA		 413}
414
38717946 415static void free_domain_mem(void *vaddr)
ba395927
KA		 416{
417 kmem_cache_free(iommu_domain_cache, vaddr);
418}
419
420static inline void * alloc_devinfo_mem(void)
421{
eb3fa7cb 422 return iommu_kmem_cache_alloc(iommu_devinfo_cache);
ba395927
KA		 423}
424
425static inline void free_devinfo_mem(void *vaddr)
426{
427 kmem_cache_free(iommu_devinfo_cache, vaddr);
428}
429
430struct iova *alloc_iova_mem(void)
431{
eb3fa7cb 432 return iommu_kmem_cache_alloc(iommu_iova_cache);
ba395927
KA		 433}
434
435void free_iova_mem(struct iova *iova)
436{
437 kmem_cache_free(iommu_iova_cache, iova);
438}
439
1b573683
WH		 440
441static inline int width_to_agaw(int width);
442
4ed0d3e6 443static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
1b573683
WH		 444{
445 unsigned long sagaw;
446 int agaw = -1;
447
448 sagaw = cap_sagaw(iommu->cap);
4ed0d3e6 449 for (agaw = width_to_agaw(max_gaw);
1b573683
WH		 450 agaw >= 0; agaw--) {
451 if (test_bit(agaw, &sagaw))
452 break;
453 }
454
455 return agaw;
456}
457
4ed0d3e6
FY		 458/*
459 * Calculate max SAGAW for each iommu.
460 */
461int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
462{
463 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
464}
465
466/*
467 * calculate agaw for each iommu.
468 * "SAGAW" may be different across iommus, use a default agaw, and
469 * get a supported less agaw for iommus that don't support the default agaw.
470 */
471int iommu_calculate_agaw(struct intel_iommu *iommu)
472{
473 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
474}
475
2c2e2c38 476/* This functionin only returns single iommu in a domain */
8c11e798
WH		 477static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
478{
479 int iommu_id;
480
2c2e2c38 481 /* si_domain and vm domain should not get here. */
1ce28feb 482 BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
2c2e2c38 483 BUG_ON(domain->flags & DOMAIN_FLAG_STATIC_IDENTITY);
1ce28feb 484
8c11e798
WH		 485 iommu_id = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
486 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
487 return NULL;
488
489 return g_iommus[iommu_id];
490}
491
8e604097
WH		 492static void domain_update_iommu_coherency(struct dmar_domain *domain)
493{
494 int i;
495
496 domain->iommu_coherency = 1;
497
498 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
499 for (; i < g_num_of_iommus; ) {
500 if (!ecap_coherent(g_iommus[i]->ecap)) {
501 domain->iommu_coherency = 0;
502 break;
503 }
504 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
505 }
506}
507
58c610bd
SY		 508static void domain_update_iommu_snooping(struct dmar_domain *domain)
509{
510 int i;
511
512 domain->iommu_snooping = 1;
513
514 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
515 for (; i < g_num_of_iommus; ) {
516 if (!ecap_sc_support(g_iommus[i]->ecap)) {
517 domain->iommu_snooping = 0;
518 break;
519 }
520 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
521 }
522}
523
524/* Some capabilities may be different across iommus */
525static void domain_update_iommu_cap(struct dmar_domain *domain)
526{
527 domain_update_iommu_coherency(domain);
528 domain_update_iommu_snooping(domain);
529}
530
276dbf99 531static struct intel_iommu *device_to_iommu(int segment, u8 bus, u8 devfn)
c7151a8d
WH		 532{
533 struct dmar_drhd_unit *drhd = NULL;
534 int i;
535
536 for_each_drhd_unit(drhd) {
537 if (drhd->ignored)
538 continue;
276dbf99
DW		 539 if (segment != drhd->segment)
540 continue;
c7151a8d 541
924b6231 542 for (i = 0; i < drhd->devices_cnt; i++) {
288e4877
DH		 543 if (drhd->devices[i] &&
544 drhd->devices[i]->bus->number == bus &&
c7151a8d
WH		 545 drhd->devices[i]->devfn == devfn)
546 return drhd->iommu;
4958c5dc
DW		 547 if (drhd->devices[i] &&
548 drhd->devices[i]->subordinate &&
924b6231
DW		 549 drhd->devices[i]->subordinate->number <= bus &&
550 drhd->devices[i]->subordinate->subordinate >= bus)
551 return drhd->iommu;
552 }
c7151a8d
WH		 553
554 if (drhd->include_all)
555 return drhd->iommu;
556 }
557
558 return NULL;
559}
560
5331fe6f
WH		 561static void domain_flush_cache(struct dmar_domain *domain,
562 void *addr, int size)
563{
564 if (!domain->iommu_coherency)
565 clflush_cache_range(addr, size);
566}
567
ba395927
KA		 568/* Gets context entry for a given bus and devfn */
569static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
570 u8 bus, u8 devfn)
571{
572 struct root_entry *root;
573 struct context_entry *context;
574 unsigned long phy_addr;
575 unsigned long flags;
576
577 spin_lock_irqsave(&iommu->lock, flags);
578 root = &iommu->root_entry[bus];
579 context = get_context_addr_from_root(root);
580 if (!context) {
581 context = (struct context_entry *)alloc_pgtable_page();
582 if (!context) {
583 spin_unlock_irqrestore(&iommu->lock, flags);
584 return NULL;
585 }
5b6985ce 586 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
ba395927
KA		 587 phy_addr = virt_to_phys((void *)context);
588 set_root_value(root, phy_addr);
589 set_root_present(root);
590 __iommu_flush_cache(iommu, root, sizeof(*root));
591 }
592 spin_unlock_irqrestore(&iommu->lock, flags);
593 return &context[devfn];
594}
595
596static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
597{
598 struct root_entry *root;
599 struct context_entry *context;
600 int ret;
601 unsigned long flags;
602
603 spin_lock_irqsave(&iommu->lock, flags);
604 root = &iommu->root_entry[bus];
605 context = get_context_addr_from_root(root);
606 if (!context) {
607 ret = 0;
608 goto out;
609 }
c07e7d21 610 ret = context_present(&context[devfn]);
ba395927
KA		 611out:
612 spin_unlock_irqrestore(&iommu->lock, flags);
613 return ret;
614}
615
616static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
617{
618 struct root_entry *root;
619 struct context_entry *context;
620 unsigned long flags;
621
622 spin_lock_irqsave(&iommu->lock, flags);
623 root = &iommu->root_entry[bus];
624 context = get_context_addr_from_root(root);
625 if (context) {
c07e7d21 626 context_clear_entry(&context[devfn]);
ba395927
KA		 627 __iommu_flush_cache(iommu, &context[devfn], \
628 sizeof(*context));
629 }
630 spin_unlock_irqrestore(&iommu->lock, flags);
631}
632
633static void free_context_table(struct intel_iommu *iommu)
634{
635 struct root_entry *root;
636 int i;
637 unsigned long flags;
638 struct context_entry *context;
639
640 spin_lock_irqsave(&iommu->lock, flags);
641 if (!iommu->root_entry) {
642 goto out;
643 }
644 for (i = 0; i < ROOT_ENTRY_NR; i++) {
645 root = &iommu->root_entry[i];
646 context = get_context_addr_from_root(root);
647 if (context)
648 free_pgtable_page(context);
649 }
650 free_pgtable_page(iommu->root_entry);
651 iommu->root_entry = NULL;
652out:
653 spin_unlock_irqrestore(&iommu->lock, flags);
654}
655
656/* page table handling */
657#define LEVEL_STRIDE (9)
658#define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
659
660static inline int agaw_to_level(int agaw)
661{
662 return agaw + 2;
663}
664
665static inline int agaw_to_width(int agaw)
666{
667 return 30 + agaw * LEVEL_STRIDE;
668
669}
670
671static inline int width_to_agaw(int width)
672{
673 return (width - 30) / LEVEL_STRIDE;
674}
675
676static inline unsigned int level_to_offset_bits(int level)
677{
6660c63a 678 return (level - 1) * LEVEL_STRIDE;
ba395927
KA		 679}
680
77dfa56c 681static inline int pfn_level_offset(unsigned long pfn, int level)
ba395927 682{
6660c63a 683 return (pfn >> level_to_offset_bits(level)) & LEVEL_MASK;
ba395927
KA		 684}
685
6660c63a 686static inline unsigned long level_mask(int level)
ba395927 687{
6660c63a 688 return -1UL << level_to_offset_bits(level);
ba395927
KA		 689}
690
6660c63a 691static inline unsigned long level_size(int level)
ba395927 692{
6660c63a 693 return 1UL << level_to_offset_bits(level);
ba395927
KA		 694}
695
6660c63a 696static inline unsigned long align_to_level(unsigned long pfn, int level)
ba395927 697{
6660c63a 698 return (pfn + level_size(level) - 1) & level_mask(level);
ba395927
KA		 699}
700
b026fd28
DW		 701static struct dma_pte *pfn_to_dma_pte(struct dmar_domain *domain,
702 unsigned long pfn)
ba395927 703{
b026fd28 704 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
ba395927
KA		 705 struct dma_pte *parent, *pte = NULL;
706 int level = agaw_to_level(domain->agaw);
707 int offset;
ba395927
KA		 708
709 BUG_ON(!domain->pgd);
b026fd28 710 BUG_ON(addr_width < BITS_PER_LONG && pfn >> addr_width);
ba395927
KA		 711 parent = domain->pgd;
712
ba395927
KA		 713 while (level > 0) {
714 void *tmp_page;
715
b026fd28 716 offset = pfn_level_offset(pfn, level);
ba395927
KA		 717 pte = &parent[offset];
718 if (level == 1)
719 break;
720
19c239ce 721 if (!dma_pte_present(pte)) {
c85994e4
DW		 722 uint64_t pteval;
723
ba395927
KA		 724 tmp_page = alloc_pgtable_page();
725
206a73c1 726 if (!tmp_page)
ba395927 727 return NULL;
206a73c1 728
c85994e4
DW		 729 domain_flush_cache(domain, tmp_page, VTD_PAGE_SIZE);
730 pteval = (virt_to_dma_pfn(tmp_page) << VTD_PAGE_SHIFT) | DMA_PTE_READ | DMA_PTE_WRITE;
731 if (cmpxchg64(&pte->val, 0ULL, pteval)) {
732 /* Someone else set it while we were thinking; use theirs. */
733 free_pgtable_page(tmp_page);
734 } else {
735 dma_pte_addr(pte);
736 domain_flush_cache(domain, pte, sizeof(*pte));
737 }
ba395927 738 }
19c239ce 739 parent = phys_to_virt(dma_pte_addr(pte));
ba395927
KA		 740 level--;
741 }
742
ba395927
KA		 743 return pte;
744}
745
746/* return address's pte at specific level */
90dcfb5e
DW		 747static struct dma_pte *dma_pfn_level_pte(struct dmar_domain *domain,
748 unsigned long pfn,
749 int level)
ba395927
KA		 750{
751 struct dma_pte *parent, *pte = NULL;
752 int total = agaw_to_level(domain->agaw);
753 int offset;
754
755 parent = domain->pgd;
756 while (level <= total) {
90dcfb5e 757 offset = pfn_level_offset(pfn, total);
ba395927
KA		 758 pte = &parent[offset];
759 if (level == total)
760 return pte;
761
19c239ce 762 if (!dma_pte_present(pte))
ba395927 763 break;
19c239ce 764 parent = phys_to_virt(dma_pte_addr(pte));
ba395927
KA		 765 total--;
766 }
767 return NULL;
768}
769
ba395927 770/* clear last level pte, a tlb flush should be followed */
595badf5
DW		 771static void dma_pte_clear_range(struct dmar_domain *domain,
772 unsigned long start_pfn,
773 unsigned long last_pfn)
ba395927 774{
04b18e65 775 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
310a5ab9 776 struct dma_pte *first_pte, *pte;
66eae846 777
04b18e65 778 BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
595badf5 779 BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
ba395927 780
04b18e65 781 /* we don't need lock here; nobody else touches the iova range */
595badf5 782 while (start_pfn <= last_pfn) {
310a5ab9
DW		 783 first_pte = pte = dma_pfn_level_pte(domain, start_pfn, 1);
784 if (!pte) {
785 start_pfn = align_to_level(start_pfn + 1, 2);
786 continue;
787 }
75e6bf96 788 do {
310a5ab9
DW		 789 dma_clear_pte(pte);
790 start_pfn++;
791 pte++;
75e6bf96
DW		 792 } while (start_pfn <= last_pfn && !first_pte_in_page(pte));
793
310a5ab9
DW		 794 domain_flush_cache(domain, first_pte,
795 (void *)pte - (void *)first_pte);
ba395927
KA		 796 }
797}
798
799/* free page table pages. last level pte should already be cleared */
800static void dma_pte_free_pagetable(struct dmar_domain *domain,
d794dc9b
DW		 801 unsigned long start_pfn,
802 unsigned long last_pfn)
ba395927 803{
6660c63a 804 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
f3a0a52f 805 struct dma_pte *first_pte, *pte;
ba395927
KA		 806 int total = agaw_to_level(domain->agaw);
807 int level;
6660c63a 808 unsigned long tmp;
ba395927 809
6660c63a
DW		 810 BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
811 BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
ba395927 812
f3a0a52f 813 /* We don't need lock here; nobody else touches the iova range */
ba395927
KA		 814 level = 2;
815 while (level <= total) {
6660c63a
DW		 816 tmp = align_to_level(start_pfn, level);
817
f3a0a52f 818 /* If we can't even clear one PTE at this level, we're done */
6660c63a 819 if (tmp + level_size(level) - 1 > last_pfn)
ba395927
KA		 820 return;
821
3d7b0e41 822 while (tmp + level_size(level) - 1 <= last_pfn) {
f3a0a52f
DW		 823 first_pte = pte = dma_pfn_level_pte(domain, tmp, level);
824 if (!pte) {
825 tmp = align_to_level(tmp + 1, level + 1);
826 continue;
827 }
75e6bf96 828 do {
6a43e574
DW		 829 if (dma_pte_present(pte)) {
830 free_pgtable_page(phys_to_virt(dma_pte_addr(pte)));
831 dma_clear_pte(pte);
832 }
f3a0a52f
DW		 833 pte++;
834 tmp += level_size(level);
75e6bf96
DW		 835 } while (!first_pte_in_page(pte) &&
836 tmp + level_size(level) - 1 <= last_pfn);
837
f3a0a52f
DW		 838 domain_flush_cache(domain, first_pte,
839 (void *)pte - (void *)first_pte);
840
ba395927
KA		 841 }
842 level++;
843 }
844 /* free pgd */
d794dc9b 845 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
ba395927
KA		 846 free_pgtable_page(domain->pgd);
847 domain->pgd = NULL;
848 }
849}
850
851/* iommu handling */
852static int iommu_alloc_root_entry(struct intel_iommu *iommu)
853{
854 struct root_entry *root;
855 unsigned long flags;
856
857 root = (struct root_entry *)alloc_pgtable_page();
858 if (!root)
859 return -ENOMEM;
860
5b6985ce 861 __iommu_flush_cache(iommu, root, ROOT_SIZE);
ba395927
KA		 862
863 spin_lock_irqsave(&iommu->lock, flags);
864 iommu->root_entry = root;
865 spin_unlock_irqrestore(&iommu->lock, flags);
866
867 return 0;
868}
869
ba395927
KA		 870static void iommu_set_root_entry(struct intel_iommu *iommu)
871{
872 void *addr;
c416daa9 873 u32 sts;
ba395927
KA		 874 unsigned long flag;
875
876 addr = iommu->root_entry;
877
878 spin_lock_irqsave(&iommu->register_lock, flag);
879 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
880
c416daa9 881 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
ba395927
KA		 882
883 /* Make sure hardware complete it */
884 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
c416daa9 885 readl, (sts & DMA_GSTS_RTPS), sts);
ba395927
KA		 886
887 spin_unlock_irqrestore(&iommu->register_lock, flag);
888}
889
890static void iommu_flush_write_buffer(struct intel_iommu *iommu)
891{
892 u32 val;
893 unsigned long flag;
894
9af88143 895 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
ba395927 896 return;
ba395927
KA		 897
898 spin_lock_irqsave(&iommu->register_lock, flag);
462b60f6 899 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
ba395927
KA		 900
901 /* Make sure hardware complete it */
902 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
c416daa9 903 readl, (!(val & DMA_GSTS_WBFS)), val);
ba395927
KA		 904
905 spin_unlock_irqrestore(&iommu->register_lock, flag);
906}
907
908/* return value determine if we need a write buffer flush */
4c25a2c1
DW		 909static void __iommu_flush_context(struct intel_iommu *iommu,
910 u16 did, u16 source_id, u8 function_mask,
911 u64 type)
ba395927
KA		 912{
913 u64 val = 0;
914 unsigned long flag;
915
ba395927
KA		 916 switch (type) {
917 case DMA_CCMD_GLOBAL_INVL:
918 val = DMA_CCMD_GLOBAL_INVL;
919 break;
920 case DMA_CCMD_DOMAIN_INVL:
921 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
922 break;
923 case DMA_CCMD_DEVICE_INVL:
924 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
925 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
926 break;
927 default:
928 BUG();
929 }
930 val |= DMA_CCMD_ICC;
931
932 spin_lock_irqsave(&iommu->register_lock, flag);
933 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
934
935 /* Make sure hardware complete it */
936 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
937 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
938
939 spin_unlock_irqrestore(&iommu->register_lock, flag);
ba395927
KA		 940}
941
ba395927 942/* return value determine if we need a write buffer flush */
1f0ef2aa
DW		 943static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
944 u64 addr, unsigned int size_order, u64 type)
ba395927
KA		 945{
946 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
947 u64 val = 0, val_iva = 0;
948 unsigned long flag;
949
ba395927
KA		 950 switch (type) {
951 case DMA_TLB_GLOBAL_FLUSH:
952 /* global flush doesn't need set IVA_REG */
953 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
954 break;
955 case DMA_TLB_DSI_FLUSH:
956 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
957 break;
958 case DMA_TLB_PSI_FLUSH:
959 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
960 /* Note: always flush non-leaf currently */
961 val_iva = size_order | addr;
962 break;
963 default:
964 BUG();
965 }
966 /* Note: set drain read/write */
967#if 0
968 /*
969 * This is probably to be super secure.. Looks like we can
970 * ignore it without any impact.
971 */
972 if (cap_read_drain(iommu->cap))
973 val |= DMA_TLB_READ_DRAIN;
974#endif
975 if (cap_write_drain(iommu->cap))
976 val |= DMA_TLB_WRITE_DRAIN;
977
978 spin_lock_irqsave(&iommu->register_lock, flag);
979 /* Note: Only uses first TLB reg currently */
980 if (val_iva)
981 dmar_writeq(iommu->reg + tlb_offset, val_iva);
982 dmar_writeq(iommu->reg + tlb_offset + 8, val);
983
984 /* Make sure hardware complete it */
985 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
986 dmar_readq, (!(val & DMA_TLB_IVT)), val);
987
988 spin_unlock_irqrestore(&iommu->register_lock, flag);
989
990 /* check IOTLB invalidation granularity */
991 if (DMA_TLB_IAIG(val) == 0)
992 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
993 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
994 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
5b6985ce
FY		 995 (unsigned long long)DMA_TLB_IIRG(type),
996 (unsigned long long)DMA_TLB_IAIG(val));
ba395927
KA		 997}
998
93a23a72
YZ		 999static struct device_domain_info *iommu_support_dev_iotlb(
1000 struct dmar_domain *domain, int segment, u8 bus, u8 devfn)
1001{
1002 int found = 0;
1003 unsigned long flags;
1004 struct device_domain_info *info;
1005 struct intel_iommu *iommu = device_to_iommu(segment, bus, devfn);
1006
1007 if (!ecap_dev_iotlb_support(iommu->ecap))
1008 return NULL;
1009
1010 if (!iommu->qi)
1011 return NULL;
1012
1013 spin_lock_irqsave(&device_domain_lock, flags);
1014 list_for_each_entry(info, &domain->devices, link)
1015 if (info->bus == bus && info->devfn == devfn) {
1016 found = 1;
1017 break;
1018 }
1019 spin_unlock_irqrestore(&device_domain_lock, flags);
1020
1021 if (!found || !info->dev)
1022 return NULL;
1023
1024 if (!pci_find_ext_capability(info->dev, PCI_EXT_CAP_ID_ATS))
1025 return NULL;
1026
1027 if (!dmar_find_matched_atsr_unit(info->dev))
1028 return NULL;
1029
1030 info->iommu = iommu;
1031
1032 return info;
1033}
1034
1035static void iommu_enable_dev_iotlb(struct device_domain_info *info)
ba395927 1036{
93a23a72
YZ		 1037 if (!info)
1038 return;
1039
1040 pci_enable_ats(info->dev, VTD_PAGE_SHIFT);
1041}
1042
1043static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1044{
1045 if (!info->dev || !pci_ats_enabled(info->dev))
1046 return;
1047
1048 pci_disable_ats(info->dev);
1049}
1050
1051static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1052 u64 addr, unsigned mask)
1053{
1054 u16 sid, qdep;
1055 unsigned long flags;
1056 struct device_domain_info *info;
1057
1058 spin_lock_irqsave(&device_domain_lock, flags);
1059 list_for_each_entry(info, &domain->devices, link) {
1060 if (!info->dev || !pci_ats_enabled(info->dev))
1061 continue;
1062
1063 sid = info->bus << 8 | info->devfn;
1064 qdep = pci_ats_queue_depth(info->dev);
1065 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1066 }
1067 spin_unlock_irqrestore(&device_domain_lock, flags);
1068}
1069
1f0ef2aa 1070static void iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
03d6a246 1071 unsigned long pfn, unsigned int pages)
ba395927 1072{
9dd2fe89 1073 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
03d6a246 1074 uint64_t addr = (uint64_t)pfn << VTD_PAGE_SHIFT;
ba395927 1075
ba395927
KA		 1076 BUG_ON(pages == 0);
1077
ba395927 1078 /*
9dd2fe89
YZ		 1079 * Fallback to domain selective flush if no PSI support or the size is
1080 * too big.
ba395927
KA		 1081 * PSI requires page size to be 2 ^ x, and the base address is naturally
1082 * aligned to the size
1083 */
9dd2fe89
YZ		 1084 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1085 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1f0ef2aa 1086 DMA_TLB_DSI_FLUSH);
9dd2fe89
YZ		 1087 else
1088 iommu->flush.flush_iotlb(iommu, did, addr, mask,
1089 DMA_TLB_PSI_FLUSH);
bf92df30
YZ		 1090
1091 /*
1092 * In caching mode, domain ID 0 is reserved for non-present to present
1093 * mapping flush. Device IOTLB doesn't need to be flushed in this case.
1094 */
1095 if (!cap_caching_mode(iommu->cap) || did)
93a23a72 1096 iommu_flush_dev_iotlb(iommu->domains[did], addr, mask);
ba395927
KA		 1097}
1098
f8bab735 1099static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1100{
1101 u32 pmen;
1102 unsigned long flags;
1103
1104 spin_lock_irqsave(&iommu->register_lock, flags);
1105 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1106 pmen &= ~DMA_PMEN_EPM;
1107 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1108
1109 /* wait for the protected region status bit to clear */
1110 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1111 readl, !(pmen & DMA_PMEN_PRS), pmen);
1112
1113 spin_unlock_irqrestore(&iommu->register_lock, flags);
1114}
1115
ba395927
KA		 1116static int iommu_enable_translation(struct intel_iommu *iommu)
1117{
1118 u32 sts;
1119 unsigned long flags;
1120
1121 spin_lock_irqsave(&iommu->register_lock, flags);
c416daa9
DW		 1122 iommu->gcmd |= DMA_GCMD_TE;
1123 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
ba395927
KA		 1124
1125 /* Make sure hardware complete it */
1126 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
c416daa9 1127 readl, (sts & DMA_GSTS_TES), sts);
ba395927 1128
ba395927
KA		 1129 spin_unlock_irqrestore(&iommu->register_lock, flags);
1130 return 0;
1131}
1132
1133static int iommu_disable_translation(struct intel_iommu *iommu)
1134{
1135 u32 sts;
1136 unsigned long flag;
1137
1138 spin_lock_irqsave(&iommu->register_lock, flag);
1139 iommu->gcmd &= ~DMA_GCMD_TE;
1140 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1141
1142 /* Make sure hardware complete it */
1143 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
c416daa9 1144 readl, (!(sts & DMA_GSTS_TES)), sts);
ba395927
KA		 1145
1146 spin_unlock_irqrestore(&iommu->register_lock, flag);
1147 return 0;
1148}
1149
3460a6d9 1150
ba395927
KA		 1151static int iommu_init_domains(struct intel_iommu *iommu)
1152{
1153 unsigned long ndomains;
1154 unsigned long nlongs;
1155
1156 ndomains = cap_ndoms(iommu->cap);
1157 pr_debug("Number of Domains supportd <%ld>\n", ndomains);
1158 nlongs = BITS_TO_LONGS(ndomains);
1159
1160 /* TBD: there might be 64K domains,
1161 * consider other allocation for future chip
1162 */
1163 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1164 if (!iommu->domain_ids) {
1165 printk(KERN_ERR "Allocating domain id array failed\n");
1166 return -ENOMEM;
1167 }
1168 iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1169 GFP_KERNEL);
1170 if (!iommu->domains) {
1171 printk(KERN_ERR "Allocating domain array failed\n");
1172 kfree(iommu->domain_ids);
1173 return -ENOMEM;
1174 }
1175
e61d98d8
SS		 1176 spin_lock_init(&iommu->lock);
1177
ba395927
KA		 1178 /*
1179 * if Caching mode is set, then invalid translations are tagged
1180 * with domainid 0. Hence we need to pre-allocate it.
1181 */
1182 if (cap_caching_mode(iommu->cap))
1183 set_bit(0, iommu->domain_ids);
1184 return 0;
1185}
ba395927 1186
ba395927
KA		 1187
1188static void domain_exit(struct dmar_domain *domain);
5e98c4b1 1189static void vm_domain_exit(struct dmar_domain *domain);
e61d98d8
SS		 1190
1191void free_dmar_iommu(struct intel_iommu *iommu)
ba395927
KA		 1192{
1193 struct dmar_domain *domain;
1194 int i;
c7151a8d 1195 unsigned long flags;
ba395927 1196
ba395927
KA		 1197 i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1198 for (; i < cap_ndoms(iommu->cap); ) {
1199 domain = iommu->domains[i];
1200 clear_bit(i, iommu->domain_ids);
c7151a8d
WH		 1201
1202 spin_lock_irqsave(&domain->iommu_lock, flags);
5e98c4b1
WH		 1203 if (--domain->iommu_count == 0) {
1204 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
1205 vm_domain_exit(domain);
1206 else
1207 domain_exit(domain);
1208 }
c7151a8d
WH		 1209 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1210
ba395927
KA		 1211 i = find_next_bit(iommu->domain_ids,
1212 cap_ndoms(iommu->cap), i+1);
1213 }
1214
1215 if (iommu->gcmd & DMA_GCMD_TE)
1216 iommu_disable_translation(iommu);
1217
1218 if (iommu->irq) {
1219 set_irq_data(iommu->irq, NULL);
1220 /* This will mask the irq */
1221 free_irq(iommu->irq, iommu);
1222 destroy_irq(iommu->irq);
1223 }
1224
1225 kfree(iommu->domains);
1226 kfree(iommu->domain_ids);
1227
d9630fe9
WH		 1228 g_iommus[iommu->seq_id] = NULL;
1229
1230 /* if all iommus are freed, free g_iommus */
1231 for (i = 0; i < g_num_of_iommus; i++) {
1232 if (g_iommus[i])
1233 break;
1234 }
1235
1236 if (i == g_num_of_iommus)
1237 kfree(g_iommus);
1238
ba395927
KA		 1239 /* free context mapping */
1240 free_context_table(iommu);
ba395927
KA		 1241}
1242
2c2e2c38 1243static struct dmar_domain *alloc_domain(void)
ba395927 1244{
ba395927 1245 struct dmar_domain *domain;
ba395927
KA		 1246
1247 domain = alloc_domain_mem();
1248 if (!domain)
1249 return NULL;
1250
2c2e2c38
FY		 1251 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
1252 domain->flags = 0;
1253
1254 return domain;
1255}
1256
1257static int iommu_attach_domain(struct dmar_domain *domain,
1258 struct intel_iommu *iommu)
1259{
1260 int num;
1261 unsigned long ndomains;
1262 unsigned long flags;
1263
ba395927
KA		 1264 ndomains = cap_ndoms(iommu->cap);
1265
1266 spin_lock_irqsave(&iommu->lock, flags);
2c2e2c38 1267
ba395927
KA		 1268 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1269 if (num >= ndomains) {
1270 spin_unlock_irqrestore(&iommu->lock, flags);
ba395927 1271 printk(KERN_ERR "IOMMU: no free domain ids\n");
2c2e2c38 1272 return -ENOMEM;
ba395927
KA		 1273 }
1274
ba395927 1275 domain->id = num;
2c2e2c38 1276 set_bit(num, iommu->domain_ids);
8c11e798 1277 set_bit(iommu->seq_id, &domain->iommu_bmp);
ba395927
KA		 1278 iommu->domains[num] = domain;
1279 spin_unlock_irqrestore(&iommu->lock, flags);
1280
2c2e2c38 1281 return 0;
ba395927
KA		 1282}
1283
2c2e2c38
FY		 1284static void iommu_detach_domain(struct dmar_domain *domain,
1285 struct intel_iommu *iommu)
ba395927
KA		 1286{
1287 unsigned long flags;
2c2e2c38
FY		 1288 int num, ndomains;
1289 int found = 0;
ba395927 1290
8c11e798 1291 spin_lock_irqsave(&iommu->lock, flags);
2c2e2c38
FY		 1292 ndomains = cap_ndoms(iommu->cap);
1293 num = find_first_bit(iommu->domain_ids, ndomains);
1294 for (; num < ndomains; ) {
1295 if (iommu->domains[num] == domain) {
1296 found = 1;
1297 break;
1298 }
1299 num = find_next_bit(iommu->domain_ids,
1300 cap_ndoms(iommu->cap), num+1);
1301 }
1302
1303 if (found) {
1304 clear_bit(num, iommu->domain_ids);
1305 clear_bit(iommu->seq_id, &domain->iommu_bmp);
1306 iommu->domains[num] = NULL;
1307 }
8c11e798 1308 spin_unlock_irqrestore(&iommu->lock, flags);
ba395927
KA		 1309}
1310
1311static struct iova_domain reserved_iova_list;
8a443df4
MG		 1312static struct lock_class_key reserved_alloc_key;
1313static struct lock_class_key reserved_rbtree_key;
ba395927
KA		 1314
1315static void dmar_init_reserved_ranges(void)
1316{
1317 struct pci_dev *pdev = NULL;
1318 struct iova *iova;
1319 int i;
ba395927 1320
f661197e 1321 init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
ba395927 1322
8a443df4
MG		 1323 lockdep_set_class(&reserved_iova_list.iova_alloc_lock,
1324 &reserved_alloc_key);
1325 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1326 &reserved_rbtree_key);
1327
ba395927
KA		 1328 /* IOAPIC ranges shouldn't be accessed by DMA */
1329 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1330 IOVA_PFN(IOAPIC_RANGE_END));
1331 if (!iova)
1332 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1333
1334 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1335 for_each_pci_dev(pdev) {
1336 struct resource *r;
1337
1338 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1339 r = &pdev->resource[i];
1340 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1341 continue;
1a4a4551
DW		 1342 iova = reserve_iova(&reserved_iova_list,
1343 IOVA_PFN(r->start),
1344 IOVA_PFN(r->end));
ba395927
KA		 1345 if (!iova)
1346 printk(KERN_ERR "Reserve iova failed\n");
1347 }
1348 }
1349
1350}
1351
1352static void domain_reserve_special_ranges(struct dmar_domain *domain)
1353{
1354 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1355}
1356
1357static inline int guestwidth_to_adjustwidth(int gaw)
1358{
1359 int agaw;
1360 int r = (gaw - 12) % 9;
1361
1362 if (r == 0)
1363 agaw = gaw;
1364 else
1365 agaw = gaw + 9 - r;
1366 if (agaw > 64)
1367 agaw = 64;
1368 return agaw;
1369}
1370
1371static int domain_init(struct dmar_domain *domain, int guest_width)
1372{
1373 struct intel_iommu *iommu;
1374 int adjust_width, agaw;
1375 unsigned long sagaw;
1376
f661197e 1377 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
c7151a8d 1378 spin_lock_init(&domain->iommu_lock);
ba395927
KA		 1379
1380 domain_reserve_special_ranges(domain);
1381
1382 /* calculate AGAW */
8c11e798 1383 iommu = domain_get_iommu(domain);
ba395927
KA		 1384 if (guest_width > cap_mgaw(iommu->cap))
1385 guest_width = cap_mgaw(iommu->cap);
1386 domain->gaw = guest_width;
1387 adjust_width = guestwidth_to_adjustwidth(guest_width);
1388 agaw = width_to_agaw(adjust_width);
1389 sagaw = cap_sagaw(iommu->cap);
1390 if (!test_bit(agaw, &sagaw)) {
1391 /* hardware doesn't support it, choose a bigger one */
1392 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1393 agaw = find_next_bit(&sagaw, 5, agaw);
1394 if (agaw >= 5)
1395 return -ENODEV;
1396 }
1397 domain->agaw = agaw;
1398 INIT_LIST_HEAD(&domain->devices);
1399
8e604097
WH		 1400 if (ecap_coherent(iommu->ecap))
1401 domain->iommu_coherency = 1;
1402 else
1403 domain->iommu_coherency = 0;
1404
58c610bd
SY		 1405 if (ecap_sc_support(iommu->ecap))
1406 domain->iommu_snooping = 1;
1407 else
1408 domain->iommu_snooping = 0;
1409
c7151a8d
WH		 1410 domain->iommu_count = 1;
1411
ba395927
KA		 1412 /* always allocate the top pgd */
1413 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1414 if (!domain->pgd)
1415 return -ENOMEM;
5b6985ce 1416 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
ba395927
KA		 1417 return 0;
1418}
1419
1420static void domain_exit(struct dmar_domain *domain)
1421{
2c2e2c38
FY		 1422 struct dmar_drhd_unit *drhd;
1423 struct intel_iommu *iommu;
ba395927
KA		 1424
1425 /* Domain 0 is reserved, so dont process it */
1426 if (!domain)
1427 return;
1428
1429 domain_remove_dev_info(domain);
1430 /* destroy iovas */
1431 put_iova_domain(&domain->iovad);
ba395927
KA		 1432
1433 /* clear ptes */
595badf5 1434 dma_pte_clear_range(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
ba395927
KA		 1435
1436 /* free page tables */
d794dc9b 1437 dma_pte_free_pagetable(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
ba395927 1438
2c2e2c38
FY		 1439 for_each_active_iommu(iommu, drhd)
1440 if (test_bit(iommu->seq_id, &domain->iommu_bmp))
1441 iommu_detach_domain(domain, iommu);
1442
ba395927
KA		 1443 free_domain_mem(domain);
1444}
1445
4ed0d3e6
FY		 1446static int domain_context_mapping_one(struct dmar_domain *domain, int segment,
1447 u8 bus, u8 devfn, int translation)
ba395927
KA		 1448{
1449 struct context_entry *context;
ba395927 1450 unsigned long flags;
5331fe6f 1451 struct intel_iommu *iommu;
ea6606b0
WH		 1452 struct dma_pte *pgd;
1453 unsigned long num;
1454 unsigned long ndomains;
1455 int id;
1456 int agaw;
93a23a72 1457 struct device_domain_info *info = NULL;
ba395927
KA		 1458
1459 pr_debug("Set context mapping for %02x:%02x.%d\n",
1460 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
4ed0d3e6 1461
ba395927 1462 BUG_ON(!domain->pgd);
4ed0d3e6
FY		 1463 BUG_ON(translation != CONTEXT_TT_PASS_THROUGH &&
1464 translation != CONTEXT_TT_MULTI_LEVEL);
5331fe6f 1465
276dbf99 1466 iommu = device_to_iommu(segment, bus, devfn);
5331fe6f
WH		 1467 if (!iommu)
1468 return -ENODEV;
1469
ba395927
KA		 1470 context = device_to_context_entry(iommu, bus, devfn);
1471 if (!context)
1472 return -ENOMEM;
1473 spin_lock_irqsave(&iommu->lock, flags);
c07e7d21 1474 if (context_present(context)) {
ba395927
KA		 1475 spin_unlock_irqrestore(&iommu->lock, flags);
1476 return 0;
1477 }
1478
ea6606b0
WH		 1479 id = domain->id;
1480 pgd = domain->pgd;
1481
2c2e2c38
FY		 1482 if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
1483 domain->flags & DOMAIN_FLAG_STATIC_IDENTITY) {
ea6606b0
WH		 1484 int found = 0;
1485
1486 /* find an available domain id for this device in iommu */
1487 ndomains = cap_ndoms(iommu->cap);
1488 num = find_first_bit(iommu->domain_ids, ndomains);
1489 for (; num < ndomains; ) {
1490 if (iommu->domains[num] == domain) {
1491 id = num;
1492 found = 1;
1493 break;
1494 }
1495 num = find_next_bit(iommu->domain_ids,
1496 cap_ndoms(iommu->cap), num+1);
1497 }
1498
1499 if (found == 0) {
1500 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1501 if (num >= ndomains) {
1502 spin_unlock_irqrestore(&iommu->lock, flags);
1503 printk(KERN_ERR "IOMMU: no free domain ids\n");
1504 return -EFAULT;
1505 }
1506
1507 set_bit(num, iommu->domain_ids);
2c2e2c38 1508 set_bit(iommu->seq_id, &domain->iommu_bmp);
ea6606b0
WH		 1509 iommu->domains[num] = domain;
1510 id = num;
1511 }
1512
1513 /* Skip top levels of page tables for
1514 * iommu which has less agaw than default.
1515 */
1516 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1517 pgd = phys_to_virt(dma_pte_addr(pgd));
1518 if (!dma_pte_present(pgd)) {
1519 spin_unlock_irqrestore(&iommu->lock, flags);
1520 return -ENOMEM;
1521 }
1522 }
1523 }
1524
1525 context_set_domain_id(context, id);
4ed0d3e6 1526
93a23a72
YZ		 1527 if (translation != CONTEXT_TT_PASS_THROUGH) {
1528 info = iommu_support_dev_iotlb(domain, segment, bus, devfn);
1529 translation = info ? CONTEXT_TT_DEV_IOTLB :
1530 CONTEXT_TT_MULTI_LEVEL;
1531 }
4ed0d3e6
FY		 1532 /*
1533 * In pass through mode, AW must be programmed to indicate the largest
1534 * AGAW value supported by hardware. And ASR is ignored by hardware.
1535 */
93a23a72 1536 if (unlikely(translation == CONTEXT_TT_PASS_THROUGH))
4ed0d3e6 1537 context_set_address_width(context, iommu->msagaw);
93a23a72
YZ		 1538 else {
1539 context_set_address_root(context, virt_to_phys(pgd));
1540 context_set_address_width(context, iommu->agaw);
1541 }
4ed0d3e6
FY		 1542
1543 context_set_translation_type(context, translation);
c07e7d21
MM		 1544 context_set_fault_enable(context);
1545 context_set_present(context);
5331fe6f 1546 domain_flush_cache(domain, context, sizeof(*context));
ba395927 1547
4c25a2c1
DW		 1548 /*
1549 * It's a non-present to present mapping. If hardware doesn't cache
1550 * non-present entry we only need to flush the write-buffer. If the
1551 * _does_ cache non-present entries, then it does so in the special
1552 * domain #0, which we have to flush:
1553 */
1554 if (cap_caching_mode(iommu->cap)) {
1555 iommu->flush.flush_context(iommu, 0,
1556 (((u16)bus) << 8) | devfn,
1557 DMA_CCMD_MASK_NOBIT,
1558 DMA_CCMD_DEVICE_INVL);
1f0ef2aa 1559 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
4c25a2c1 1560 } else {
ba395927 1561 iommu_flush_write_buffer(iommu);
4c25a2c1 1562 }
93a23a72 1563 iommu_enable_dev_iotlb(info);
ba395927 1564 spin_unlock_irqrestore(&iommu->lock, flags);
c7151a8d
WH		 1565
1566 spin_lock_irqsave(&domain->iommu_lock, flags);
1567 if (!test_and_set_bit(iommu->seq_id, &domain->iommu_bmp)) {
1568 domain->iommu_count++;
58c610bd 1569 domain_update_iommu_cap(domain);
c7151a8d
WH		 1570 }
1571 spin_unlock_irqrestore(&domain->iommu_lock, flags);
ba395927
KA		 1572 return 0;
1573}
1574
1575static int
4ed0d3e6
FY		 1576domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev,
1577 int translation)
ba395927
KA		 1578{
1579 int ret;
1580 struct pci_dev *tmp, *parent;
1581
276dbf99 1582 ret = domain_context_mapping_one(domain, pci_domain_nr(pdev->bus),
4ed0d3e6
FY		 1583 pdev->bus->number, pdev->devfn,
1584 translation);
ba395927
KA		 1585 if (ret)
1586 return ret;
1587
1588 /* dependent device mapping */
1589 tmp = pci_find_upstream_pcie_bridge(pdev);
1590 if (!tmp)
1591 return 0;
1592 /* Secondary interface's bus number and devfn 0 */
1593 parent = pdev->bus->self;
1594 while (parent != tmp) {
276dbf99
DW		 1595 ret = domain_context_mapping_one(domain,
1596 pci_domain_nr(parent->bus),
1597 parent->bus->number,
4ed0d3e6 1598 parent->devfn, translation);
ba395927
KA		 1599 if (ret)
1600 return ret;
1601 parent = parent->bus->self;
1602 }
1603 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1604 return domain_context_mapping_one(domain,
276dbf99 1605 pci_domain_nr(tmp->subordinate),
4ed0d3e6
FY		 1606 tmp->subordinate->number, 0,
1607 translation);
ba395927
KA		 1608 else /* this is a legacy PCI bridge */
1609 return domain_context_mapping_one(domain,
276dbf99
DW		 1610 pci_domain_nr(tmp->bus),
1611 tmp->bus->number,
4ed0d3e6
FY		 1612 tmp->devfn,
1613 translation);
ba395927
KA		 1614}
1615
5331fe6f 1616static int domain_context_mapped(struct pci_dev *pdev)
ba395927
KA		 1617{
1618 int ret;
1619 struct pci_dev *tmp, *parent;
5331fe6f
WH		 1620 struct intel_iommu *iommu;
1621
276dbf99
DW		 1622 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
1623 pdev->devfn);
5331fe6f
WH		 1624 if (!iommu)
1625 return -ENODEV;
ba395927 1626
276dbf99 1627 ret = device_context_mapped(iommu, pdev->bus->number, pdev->devfn);
ba395927
KA		 1628 if (!ret)
1629 return ret;
1630 /* dependent device mapping */
1631 tmp = pci_find_upstream_pcie_bridge(pdev);
1632 if (!tmp)
1633 return ret;
1634 /* Secondary interface's bus number and devfn 0 */
1635 parent = pdev->bus->self;
1636 while (parent != tmp) {
8c11e798 1637 ret = device_context_mapped(iommu, parent->bus->number,
276dbf99 1638 parent->devfn);
ba395927
KA		 1639 if (!ret)
1640 return ret;
1641 parent = parent->bus->self;
1642 }
1643 if (tmp->is_pcie)
276dbf99
DW		 1644 return device_context_mapped(iommu, tmp->subordinate->number,
1645 0);
ba395927 1646 else
276dbf99
DW		 1647 return device_context_mapped(iommu, tmp->bus->number,
1648 tmp->devfn);
ba395927
KA		 1649}
1650
9051aa02
DW		 1651static int __domain_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1652 struct scatterlist *sg, unsigned long phys_pfn,
1653 unsigned long nr_pages, int prot)
e1605495
DW		 1654{
1655 struct dma_pte *first_pte = NULL, *pte = NULL;
9051aa02 1656 phys_addr_t uninitialized_var(pteval);
e1605495 1657 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
9051aa02 1658 unsigned long sg_res;
e1605495
DW		 1659
1660 BUG_ON(addr_width < BITS_PER_LONG && (iov_pfn + nr_pages - 1) >> addr_width);
1661
1662 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1663 return -EINVAL;
1664
1665 prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
1666
9051aa02
DW		 1667 if (sg)
1668 sg_res = 0;
1669 else {
1670 sg_res = nr_pages + 1;
1671 pteval = ((phys_addr_t)phys_pfn << VTD_PAGE_SHIFT) | prot;
1672 }
1673
e1605495 1674 while (nr_pages--) {
c85994e4
DW		 1675 uint64_t tmp;
1676
e1605495
DW		 1677 if (!sg_res) {
1678 sg_res = (sg->offset + sg->length + VTD_PAGE_SIZE - 1) >> VTD_PAGE_SHIFT;
1679 sg->dma_address = ((dma_addr_t)iov_pfn << VTD_PAGE_SHIFT) + sg->offset;
1680 sg->dma_length = sg->length;
1681 pteval = page_to_phys(sg_page(sg)) | prot;
1682 }
1683 if (!pte) {
1684 first_pte = pte = pfn_to_dma_pte(domain, iov_pfn);
1685 if (!pte)
1686 return -ENOMEM;
1687 }
1688 /* We don't need lock here, nobody else
1689 * touches the iova range
1690 */
7766a3fb 1691 tmp = cmpxchg64_local(&pte->val, 0ULL, pteval);
c85994e4 1692 if (tmp) {
1bf20f0d 1693 static int dumps = 5;
c85994e4
DW		 1694 printk(KERN_CRIT "ERROR: DMA PTE for vPFN 0x%lx already set (to %llx not %llx)\n",
1695 iov_pfn, tmp, (unsigned long long)pteval);
1bf20f0d
DW		 1696 if (dumps) {
1697 dumps--;
1698 debug_dma_dump_mappings(NULL);
1699 }
1700 WARN_ON(1);
1701 }
e1605495 1702 pte++;
75e6bf96 1703 if (!nr_pages || first_pte_in_page(pte)) {
e1605495
DW		 1704 domain_flush_cache(domain, first_pte,
1705 (void *)pte - (void *)first_pte);
1706 pte = NULL;
1707 }
1708 iov_pfn++;
1709 pteval += VTD_PAGE_SIZE;
1710 sg_res--;
1711 if (!sg_res)
1712 sg = sg_next(sg);
1713 }
1714 return 0;
1715}
1716
9051aa02
DW		 1717static inline int domain_sg_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1718 struct scatterlist *sg, unsigned long nr_pages,
1719 int prot)
ba395927 1720{
9051aa02
DW		 1721 return __domain_mapping(domain, iov_pfn, sg, 0, nr_pages, prot);
1722}
6f6a00e4 1723
9051aa02
DW		 1724static inline int domain_pfn_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
1725 unsigned long phys_pfn, unsigned long nr_pages,
1726 int prot)
1727{
1728 return __domain_mapping(domain, iov_pfn, NULL, phys_pfn, nr_pages, prot);
ba395927
KA		 1729}
1730
c7151a8d 1731static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
ba395927 1732{
c7151a8d
WH		 1733 if (!iommu)
1734 return;
8c11e798
WH		 1735
1736 clear_context_table(iommu, bus, devfn);
1737 iommu->flush.flush_context(iommu, 0, 0, 0,
4c25a2c1 1738 DMA_CCMD_GLOBAL_INVL);
1f0ef2aa 1739 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
ba395927
KA		 1740}
1741
1742static void domain_remove_dev_info(struct dmar_domain *domain)
1743{
1744 struct device_domain_info *info;
1745 unsigned long flags;
c7151a8d 1746 struct intel_iommu *iommu;
ba395927
KA		 1747
1748 spin_lock_irqsave(&device_domain_lock, flags);
1749 while (!list_empty(&domain->devices)) {
1750 info = list_entry(domain->devices.next,
1751 struct device_domain_info, link);
1752 list_del(&info->link);
1753 list_del(&info->global);
1754 if (info->dev)
358dd8ac 1755 info->dev->dev.archdata.iommu = NULL;
ba395927
KA		 1756 spin_unlock_irqrestore(&device_domain_lock, flags);
1757
93a23a72 1758 iommu_disable_dev_iotlb(info);
276dbf99 1759 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
c7151a8d 1760 iommu_detach_dev(iommu, info->bus, info->devfn);
ba395927
KA		 1761 free_devinfo_mem(info);
1762
1763 spin_lock_irqsave(&device_domain_lock, flags);
1764 }
1765 spin_unlock_irqrestore(&device_domain_lock, flags);
1766}
1767
1768/*
1769 * find_domain
358dd8ac 1770 * Note: we use struct pci_dev->dev.archdata.iommu stores the info
ba395927 1771 */
38717946 1772static struct dmar_domain *
ba395927
KA		 1773find_domain(struct pci_dev *pdev)
1774{
1775 struct device_domain_info *info;
1776
1777 /* No lock here, assumes no domain exit in normal case */
358dd8ac 1778 info = pdev->dev.archdata.iommu;
ba395927
KA		 1779 if (info)
1780 return info->domain;
1781 return NULL;
1782}
1783
ba395927
KA		 1784/* domain is initialized */
1785static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1786{
1787 struct dmar_domain *domain, *found = NULL;
1788 struct intel_iommu *iommu;
1789 struct dmar_drhd_unit *drhd;
1790 struct device_domain_info *info, *tmp;
1791 struct pci_dev *dev_tmp;
1792 unsigned long flags;
1793 int bus = 0, devfn = 0;
276dbf99 1794 int segment;
2c2e2c38 1795 int ret;
ba395927
KA		 1796
1797 domain = find_domain(pdev);
1798 if (domain)
1799 return domain;
1800
276dbf99
DW		 1801 segment = pci_domain_nr(pdev->bus);
1802
ba395927
KA		 1803 dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1804 if (dev_tmp) {
1805 if (dev_tmp->is_pcie) {
1806 bus = dev_tmp->subordinate->number;
1807 devfn = 0;
1808 } else {
1809 bus = dev_tmp->bus->number;
1810 devfn = dev_tmp->devfn;
1811 }
1812 spin_lock_irqsave(&device_domain_lock, flags);
1813 list_for_each_entry(info, &device_domain_list, global) {
276dbf99
DW		 1814 if (info->segment == segment &&
1815 info->bus == bus && info->devfn == devfn) {
ba395927
KA		 1816 found = info->domain;
1817 break;
1818 }
1819 }
1820 spin_unlock_irqrestore(&device_domain_lock, flags);
1821 /* pcie-pci bridge already has a domain, uses it */
1822 if (found) {
1823 domain = found;
1824 goto found_domain;
1825 }
1826 }
1827
2c2e2c38
FY		 1828 domain = alloc_domain();
1829 if (!domain)
1830 goto error;
1831
ba395927
KA		 1832 /* Allocate new domain for the device */
1833 drhd = dmar_find_matched_drhd_unit(pdev);
1834 if (!drhd) {
1835 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1836 pci_name(pdev));
1837 return NULL;
1838 }
1839 iommu = drhd->iommu;
1840
2c2e2c38
FY		 1841 ret = iommu_attach_domain(domain, iommu);
1842 if (ret) {
1843 domain_exit(domain);
ba395927 1844 goto error;
2c2e2c38 1845 }
ba395927
KA		 1846
1847 if (domain_init(domain, gaw)) {
1848 domain_exit(domain);
1849 goto error;
1850 }
1851
1852 /* register pcie-to-pci device */
1853 if (dev_tmp) {
1854 info = alloc_devinfo_mem();
1855 if (!info) {
1856 domain_exit(domain);
1857 goto error;
1858 }
276dbf99 1859 info->segment = segment;
ba395927
KA		 1860 info->bus = bus;
1861 info->devfn = devfn;
1862 info->dev = NULL;
1863 info->domain = domain;
1864 /* This domain is shared by devices under p2p bridge */
3b5410e7 1865 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
ba395927
KA		 1866
1867 /* pcie-to-pci bridge already has a domain, uses it */
1868 found = NULL;
1869 spin_lock_irqsave(&device_domain_lock, flags);
1870 list_for_each_entry(tmp, &device_domain_list, global) {
276dbf99
DW		 1871 if (tmp->segment == segment &&
1872 tmp->bus == bus && tmp->devfn == devfn) {
ba395927
KA		 1873 found = tmp->domain;
1874 break;
1875 }
1876 }
1877 if (found) {
1878 free_devinfo_mem(info);
1879 domain_exit(domain);
1880 domain = found;
1881 } else {
1882 list_add(&info->link, &domain->devices);
1883 list_add(&info->global, &device_domain_list);
1884 }
1885 spin_unlock_irqrestore(&device_domain_lock, flags);
1886 }
1887
1888found_domain:
1889 info = alloc_devinfo_mem();
1890 if (!info)
1891 goto error;
276dbf99 1892 info->segment = segment;
ba395927
KA		 1893 info->bus = pdev->bus->number;
1894 info->devfn = pdev->devfn;
1895 info->dev = pdev;
1896 info->domain = domain;
1897 spin_lock_irqsave(&device_domain_lock, flags);
1898 /* somebody is fast */
1899 found = find_domain(pdev);
1900 if (found != NULL) {
1901 spin_unlock_irqrestore(&device_domain_lock, flags);
1902 if (found != domain) {
1903 domain_exit(domain);
1904 domain = found;
1905 }
1906 free_devinfo_mem(info);
1907 return domain;
1908 }
1909 list_add(&info->link, &domain->devices);
1910 list_add(&info->global, &device_domain_list);
358dd8ac 1911 pdev->dev.archdata.iommu = info;
ba395927
KA		 1912 spin_unlock_irqrestore(&device_domain_lock, flags);
1913 return domain;
1914error:
1915 /* recheck it here, maybe others set it */
1916 return find_domain(pdev);
1917}
1918
2c2e2c38
FY		 1919static int iommu_identity_mapping;
1920
b213203e
DW		 1921static int iommu_domain_identity_map(struct dmar_domain *domain,
1922 unsigned long long start,
1923 unsigned long long end)
ba395927 1924{
c5395d5c
DW		 1925 unsigned long first_vpfn = start >> VTD_PAGE_SHIFT;
1926 unsigned long last_vpfn = end >> VTD_PAGE_SHIFT;
1927
1928 if (!reserve_iova(&domain->iovad, dma_to_mm_pfn(first_vpfn),
1929 dma_to_mm_pfn(last_vpfn))) {
ba395927 1930 printk(KERN_ERR "IOMMU: reserve iova failed\n");
b213203e 1931 return -ENOMEM;
ba395927
KA		 1932 }
1933
c5395d5c
DW		 1934 pr_debug("Mapping reserved region %llx-%llx for domain %d\n",
1935 start, end, domain->id);
ba395927
KA		 1936 /*
1937 * RMRR range might have overlap with physical memory range,
1938 * clear it first
1939 */
c5395d5c 1940 dma_pte_clear_range(domain, first_vpfn, last_vpfn);
ba395927 1941
c5395d5c
DW		 1942 return domain_pfn_mapping(domain, first_vpfn, first_vpfn,
1943 last_vpfn - first_vpfn + 1,
61df7443 1944 DMA_PTE_READ|DMA_PTE_WRITE);
b213203e
DW		 1945}
1946
1947static int iommu_prepare_identity_map(struct pci_dev *pdev,
1948 unsigned long long start,
1949 unsigned long long end)
1950{
1951 struct dmar_domain *domain;
1952 int ret;
1953
1954 printk(KERN_INFO
1955 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1956 pci_name(pdev), start, end);
1957
c7ab48d2 1958 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
b213203e
DW		 1959 if (!domain)
1960 return -ENOMEM;
1961
1962 ret = iommu_domain_identity_map(domain, start, end);
ba395927
KA		 1963 if (ret)
1964 goto error;
1965
1966 /* context entry init */
4ed0d3e6 1967 ret = domain_context_mapping(domain, pdev, CONTEXT_TT_MULTI_LEVEL);
b213203e
DW		 1968 if (ret)
1969 goto error;
1970
1971 return 0;
1972
1973 error:
ba395927
KA		 1974 domain_exit(domain);
1975 return ret;
ba395927
KA		 1976}
1977
1978static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1979 struct pci_dev *pdev)
1980{
358dd8ac 1981 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
ba395927
KA		 1982 return 0;
1983 return iommu_prepare_identity_map(pdev, rmrr->base_address,
1984 rmrr->end_address + 1);
1985}
1986
49a0429e
KA		 1987#ifdef CONFIG_DMAR_FLOPPY_WA
1988static inline void iommu_prepare_isa(void)
1989{
1990 struct pci_dev *pdev;
1991 int ret;
1992
1993 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1994 if (!pdev)
1995 return;
1996
c7ab48d2 1997 printk(KERN_INFO "IOMMU: Prepare 0-16MiB unity mapping for LPC\n");
49a0429e
KA		 1998 ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1999
2000 if (ret)
c7ab48d2
DW		 2001 printk(KERN_ERR "IOMMU: Failed to create 0-16MiB identity map; "
2002 "floppy might not work\n");
49a0429e
KA		 2003
2004}
2005#else
2006static inline void iommu_prepare_isa(void)
2007{
2008 return;
2009}
2010#endif /* !CONFIG_DMAR_FLPY_WA */
2011
4ed0d3e6
FY		 2012/* Initialize each context entry as pass through.*/
2013static int __init init_context_pass_through(void)
2014{
2015 struct pci_dev *pdev = NULL;
2016 struct dmar_domain *domain;
2017 int ret;
2018
2019 for_each_pci_dev(pdev) {
2020 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2021 ret = domain_context_mapping(domain, pdev,
2022 CONTEXT_TT_PASS_THROUGH);
2023 if (ret)
2024 return ret;
2025 }
2026 return 0;
2027}
2028
2c2e2c38 2029static int md_domain_init(struct dmar_domain *domain, int guest_width);
c7ab48d2
DW		 2030
2031static int __init si_domain_work_fn(unsigned long start_pfn,
2032 unsigned long end_pfn, void *datax)
2033{
2034 int *ret = datax;
2035
2036 *ret = iommu_domain_identity_map(si_domain,
2037 (uint64_t)start_pfn << PAGE_SHIFT,
2038 (uint64_t)end_pfn << PAGE_SHIFT);
2039 return *ret;
2040
2041}
2042
2c2e2c38
FY		 2043static int si_domain_init(void)
2044{
2045 struct dmar_drhd_unit *drhd;
2046 struct intel_iommu *iommu;
c7ab48d2 2047 int nid, ret = 0;
2c2e2c38
FY		 2048
2049 si_domain = alloc_domain();
2050 if (!si_domain)
2051 return -EFAULT;
2052
c7ab48d2 2053 pr_debug("Identity mapping domain is domain %d\n", si_domain->id);
2c2e2c38
FY		 2054
2055 for_each_active_iommu(iommu, drhd) {
2056 ret = iommu_attach_domain(si_domain, iommu);
2057 if (ret) {
2058 domain_exit(si_domain);
2059 return -EFAULT;
2060 }
2061 }
2062
2063 if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2064 domain_exit(si_domain);
2065 return -EFAULT;
2066 }
2067
2068 si_domain->flags = DOMAIN_FLAG_STATIC_IDENTITY;
2069
c7ab48d2
DW		 2070 for_each_online_node(nid) {
2071 work_with_active_regions(nid, si_domain_work_fn, &ret);
2072 if (ret)
2073 return ret;
2074 }
2075
2c2e2c38
FY		 2076 return 0;
2077}
2078
2079static void domain_remove_one_dev_info(struct dmar_domain *domain,
2080 struct pci_dev *pdev);
2081static int identity_mapping(struct pci_dev *pdev)
2082{
2083 struct device_domain_info *info;
2084
2085 if (likely(!iommu_identity_mapping))
2086 return 0;
2087
2088
2089 list_for_each_entry(info, &si_domain->devices, link)
2090 if (info->dev == pdev)
2091 return 1;
2092 return 0;
2093}
2094
2095static int domain_add_dev_info(struct dmar_domain *domain,
2096 struct pci_dev *pdev)
2097{
2098 struct device_domain_info *info;
2099 unsigned long flags;
2100
2101 info = alloc_devinfo_mem();
2102 if (!info)
2103 return -ENOMEM;
2104
2105 info->segment = pci_domain_nr(pdev->bus);
2106 info->bus = pdev->bus->number;
2107 info->devfn = pdev->devfn;
2108 info->dev = pdev;
2109 info->domain = domain;
2110
2111 spin_lock_irqsave(&device_domain_lock, flags);
2112 list_add(&info->link, &domain->devices);
2113 list_add(&info->global, &device_domain_list);
2114 pdev->dev.archdata.iommu = info;
2115 spin_unlock_irqrestore(&device_domain_lock, flags);
2116
2117 return 0;
2118}
2119
2120static int iommu_prepare_static_identity_mapping(void)
2121{
2c2e2c38
FY		 2122 struct pci_dev *pdev = NULL;
2123 int ret;
2124
2125 ret = si_domain_init();
2126 if (ret)
2127 return -EFAULT;
2128
2c2e2c38 2129 for_each_pci_dev(pdev) {
c7ab48d2
DW		 2130 printk(KERN_INFO "IOMMU: identity mapping for device %s\n",
2131 pci_name(pdev));
2132
2133 ret = domain_context_mapping(si_domain, pdev,
2134 CONTEXT_TT_MULTI_LEVEL);
2135 if (ret)
2136 return ret;
2c2e2c38
FY		 2137 ret = domain_add_dev_info(si_domain, pdev);
2138 if (ret)
2139 return ret;
2140 }
2141
2142 return 0;
2143}
2144
2145int __init init_dmars(void)
ba395927
KA		 2146{
2147 struct dmar_drhd_unit *drhd;
2148 struct dmar_rmrr_unit *rmrr;
2149 struct pci_dev *pdev;
2150 struct intel_iommu *iommu;
9d783ba0 2151 int i, ret;
4ed0d3e6 2152 int pass_through = 1;
ba395927 2153
2c2e2c38
FY		 2154 /*
2155 * In case pass through can not be enabled, iommu tries to use identity
2156 * mapping.
2157 */
2158 if (iommu_pass_through)
2159 iommu_identity_mapping = 1;
2160
ba395927
KA		 2161 /*
2162 * for each drhd
2163 * allocate root
2164 * initialize and program root entry to not present
2165 * endfor
2166 */
2167 for_each_drhd_unit(drhd) {
5e0d2a6f 2168 g_num_of_iommus++;
2169 /*
2170 * lock not needed as this is only incremented in the single
2171 * threaded kernel __init code path all other access are read
2172 * only
2173 */
2174 }
2175
d9630fe9
WH		 2176 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
2177 GFP_KERNEL);
2178 if (!g_iommus) {
2179 printk(KERN_ERR "Allocating global iommu array failed\n");
2180 ret = -ENOMEM;
2181 goto error;
2182 }
2183
80b20dd8 2184 deferred_flush = kzalloc(g_num_of_iommus *
2185 sizeof(struct deferred_flush_tables), GFP_KERNEL);
2186 if (!deferred_flush) {
d9630fe9 2187 kfree(g_iommus);
5e0d2a6f 2188 ret = -ENOMEM;
2189 goto error;
2190 }
2191
5e0d2a6f 2192 for_each_drhd_unit(drhd) {
2193 if (drhd->ignored)
2194 continue;
1886e8a9
SS		 2195
2196 iommu = drhd->iommu;
d9630fe9 2197 g_iommus[iommu->seq_id] = iommu;
ba395927 2198
e61d98d8
SS		 2199 ret = iommu_init_domains(iommu);
2200 if (ret)
2201 goto error;
2202
ba395927
KA		 2203 /*
2204 * TBD:
2205 * we could share the same root & context tables
2206 * amoung all IOMMU's. Need to Split it later.
2207 */
2208 ret = iommu_alloc_root_entry(iommu);
2209 if (ret) {
2210 printk(KERN_ERR "IOMMU: allocate root entry failed\n");
2211 goto error;
2212 }
4ed0d3e6
FY		 2213 if (!ecap_pass_through(iommu->ecap))
2214 pass_through = 0;
ba395927 2215 }
4ed0d3e6
FY		 2216 if (iommu_pass_through)
2217 if (!pass_through) {
2218 printk(KERN_INFO
2219 "Pass Through is not supported by hardware.\n");
2220 iommu_pass_through = 0;
2221 }
ba395927 2222
1531a6a6
SS		 2223 /*
2224 * Start from the sane iommu hardware state.
2225 */
a77b67d4
YS		 2226 for_each_drhd_unit(drhd) {
2227 if (drhd->ignored)
2228 continue;
2229
2230 iommu = drhd->iommu;
1531a6a6
SS		 2231
2232 /*
2233 * If the queued invalidation is already initialized by us
2234 * (for example, while enabling interrupt-remapping) then
2235 * we got the things already rolling from a sane state.
2236 */
2237 if (iommu->qi)
2238 continue;
2239
2240 /*
2241 * Clear any previous faults.
2242 */
2243 dmar_fault(-1, iommu);
2244 /*
2245 * Disable queued invalidation if supported and already enabled
2246 * before OS handover.
2247 */
2248 dmar_disable_qi(iommu);
2249 }
2250
2251 for_each_drhd_unit(drhd) {
2252 if (drhd->ignored)
2253 continue;
2254
2255 iommu = drhd->iommu;
2256
a77b67d4
YS		 2257 if (dmar_enable_qi(iommu)) {
2258 /*
2259 * Queued Invalidate not enabled, use Register Based
2260 * Invalidate
2261 */
2262 iommu->flush.flush_context = __iommu_flush_context;
2263 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2264 printk(KERN_INFO "IOMMU 0x%Lx: using Register based "
b4e0f9eb
FT		 2265 "invalidation\n",
2266 (unsigned long long)drhd->reg_base_addr);
a77b67d4
YS		 2267 } else {
2268 iommu->flush.flush_context = qi_flush_context;
2269 iommu->flush.flush_iotlb = qi_flush_iotlb;
2270 printk(KERN_INFO "IOMMU 0x%Lx: using Queued "
b4e0f9eb
FT		 2271 "invalidation\n",
2272 (unsigned long long)drhd->reg_base_addr);
a77b67d4
YS		 2273 }
2274 }
2275
ba395927 2276 /*
4ed0d3e6
FY		 2277 * If pass through is set and enabled, context entries of all pci
2278 * devices are intialized by pass through translation type.
ba395927 2279 */
4ed0d3e6
FY		 2280 if (iommu_pass_through) {
2281 ret = init_context_pass_through();
2282 if (ret) {
2283 printk(KERN_ERR "IOMMU: Pass through init failed.\n");
2284 iommu_pass_through = 0;
ba395927
KA		 2285 }
2286 }
2287
ba395927 2288 /*
4ed0d3e6 2289 * If pass through is not set or not enabled, setup context entries for
2c2e2c38
FY		 2290 * identity mappings for rmrr, gfx, and isa and may fall back to static
2291 * identity mapping if iommu_identity_mapping is set.
ba395927 2292 */
4ed0d3e6 2293 if (!iommu_pass_through) {
2c2e2c38
FY		 2294 if (iommu_identity_mapping)
2295 iommu_prepare_static_identity_mapping();
4ed0d3e6
FY		 2296 /*
2297 * For each rmrr
2298 * for each dev attached to rmrr
2299 * do
2300 * locate drhd for dev, alloc domain for dev
2301 * allocate free domain
2302 * allocate page table entries for rmrr
2303 * if context not allocated for bus
2304 * allocate and init context
2305 * set present in root table for this bus
2306 * init context with domain, translation etc
2307 * endfor
2308 * endfor
2309 */
2c2e2c38 2310 printk(KERN_INFO "IOMMU: Setting RMRR:\n");
4ed0d3e6
FY		 2311 for_each_rmrr_units(rmrr) {
2312 for (i = 0; i < rmrr->devices_cnt; i++) {
2313 pdev = rmrr->devices[i];
2314 /*
2315 * some BIOS lists non-exist devices in DMAR
2316 * table.
2317 */
2318 if (!pdev)
2319 continue;
2320 ret = iommu_prepare_rmrr_dev(rmrr, pdev);
2321 if (ret)
2322 printk(KERN_ERR
ba395927 2323 "IOMMU: mapping reserved region failed\n");
4ed0d3e6 2324 }
ba395927 2325 }
ba395927 2326
4ed0d3e6
FY		 2327 iommu_prepare_isa();
2328 }
49a0429e 2329
ba395927
KA		 2330 /*
2331 * for each drhd
2332 * enable fault log
2333 * global invalidate context cache
2334 * global invalidate iotlb
2335 * enable translation
2336 */
2337 for_each_drhd_unit(drhd) {
2338 if (drhd->ignored)
2339 continue;
2340 iommu = drhd->iommu;
ba395927
KA		 2341
2342 iommu_flush_write_buffer(iommu);
2343
3460a6d9
KA		 2344 ret = dmar_set_interrupt(iommu);
2345 if (ret)
2346 goto error;
2347
ba395927
KA		 2348 iommu_set_root_entry(iommu);
2349
4c25a2c1 2350 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
1f0ef2aa 2351 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
f8bab735 2352 iommu_disable_protect_mem_regions(iommu);
2353
ba395927
KA		 2354 ret = iommu_enable_translation(iommu);
2355 if (ret)
2356 goto error;
2357 }
2358
2359 return 0;
2360error:
2361 for_each_drhd_unit(drhd) {
2362 if (drhd->ignored)
2363 continue;
2364 iommu = drhd->iommu;
2365 free_iommu(iommu);
2366 }
d9630fe9 2367 kfree(g_iommus);
ba395927
KA		 2368 return ret;
2369}
2370
5a5e02a6 2371/* Returns a number of VTD pages, but aligned to MM page size */
88cb6a74
DW		 2372static inline unsigned long aligned_nrpages(unsigned long host_addr,
2373 size_t size)
ba395927 2374{
88cb6a74 2375 host_addr &= ~PAGE_MASK;
5a5e02a6 2376 return PAGE_ALIGN(host_addr + size) >> VTD_PAGE_SHIFT;
ba395927
KA		 2377}
2378
5a5e02a6 2379/* This takes a number of _MM_ pages, not VTD pages */
875764de
DW		 2380static struct iova *intel_alloc_iova(struct device *dev,
2381 struct dmar_domain *domain,
2382 unsigned long nrpages, uint64_t dma_mask)
ba395927 2383{
ba395927 2384 struct pci_dev *pdev = to_pci_dev(dev);
ba395927 2385 struct iova *iova = NULL;
ba395927 2386
875764de
DW		 2387 /* Restrict dma_mask to the width that the iommu can handle */
2388 dma_mask = min_t(uint64_t, DOMAIN_MAX_ADDR(domain->gaw), dma_mask);
2389
2390 if (!dmar_forcedac && dma_mask > DMA_BIT_MASK(32)) {
ba395927
KA		 2391 /*
2392 * First try to allocate an io virtual address in
284901a9 2393 * DMA_BIT_MASK(32) and if that fails then try allocating
3609801e 2394 * from higher range
ba395927 2395 */
875764de
DW		 2396 iova = alloc_iova(&domain->iovad, nrpages,
2397 IOVA_PFN(DMA_BIT_MASK(32)), 1);
2398 if (iova)
2399 return iova;
2400 }
2401 iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
2402 if (unlikely(!iova)) {
2403 printk(KERN_ERR "Allocating %ld-page iova for %s failed",
2404 nrpages, pci_name(pdev));
f76aec76
KA		 2405 return NULL;
2406 }
2407
2408 return iova;
2409}
2410
2411static struct dmar_domain *
2412get_valid_domain_for_dev(struct pci_dev *pdev)
2413{
2414 struct dmar_domain *domain;
2415 int ret;
2416
2417 domain = get_domain_for_dev(pdev,
2418 DEFAULT_DOMAIN_ADDRESS_WIDTH);
2419 if (!domain) {
2420 printk(KERN_ERR
2421 "Allocating domain for %s failed", pci_name(pdev));
4fe05bbc 2422 return NULL;
ba395927
KA		 2423 }
2424
2425 /* make sure context mapping is ok */
5331fe6f 2426 if (unlikely(!domain_context_mapped(pdev))) {
4ed0d3e6
FY		 2427 ret = domain_context_mapping(domain, pdev,
2428 CONTEXT_TT_MULTI_LEVEL);
f76aec76
KA		 2429 if (ret) {
2430 printk(KERN_ERR
2431 "Domain context map for %s failed",
2432 pci_name(pdev));
4fe05bbc 2433 return NULL;
f76aec76 2434 }
ba395927
KA		 2435 }
2436
f76aec76
KA		 2437 return domain;
2438}
2439
2c2e2c38
FY		 2440static int iommu_dummy(struct pci_dev *pdev)
2441{
2442 return pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
2443}
2444
2445/* Check if the pdev needs to go through non-identity map and unmap process.*/
2446static int iommu_no_mapping(struct pci_dev *pdev)
2447{
2448 int found;
2449
1e4c64c4
DW		 2450 if (iommu_dummy(pdev))
2451 return 1;
2452
2c2e2c38 2453 if (!iommu_identity_mapping)
1e4c64c4 2454 return 0;
2c2e2c38
FY		 2455
2456 found = identity_mapping(pdev);
2457 if (found) {
2458 if (pdev->dma_mask > DMA_BIT_MASK(32))
2459 return 1;
2460 else {
2461 /*
2462 * 32 bit DMA is removed from si_domain and fall back
2463 * to non-identity mapping.
2464 */
2465 domain_remove_one_dev_info(si_domain, pdev);
2466 printk(KERN_INFO "32bit %s uses non-identity mapping\n",
2467 pci_name(pdev));
2468 return 0;
2469 }
2470 } else {
2471 /*
2472 * In case of a detached 64 bit DMA device from vm, the device
2473 * is put into si_domain for identity mapping.
2474 */
2475 if (pdev->dma_mask > DMA_BIT_MASK(32)) {
2476 int ret;
2477 ret = domain_add_dev_info(si_domain, pdev);
2478 if (!ret) {
2479 printk(KERN_INFO "64bit %s uses identity mapping\n",
2480 pci_name(pdev));
2481 return 1;
2482 }
2483 }
2484 }
2485
1e4c64c4 2486 return 0;
2c2e2c38
FY		 2487}
2488
bb9e6d65
FT		 2489static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
2490 size_t size, int dir, u64 dma_mask)
f76aec76
KA		 2491{
2492 struct pci_dev *pdev = to_pci_dev(hwdev);
f76aec76 2493 struct dmar_domain *domain;
5b6985ce 2494 phys_addr_t start_paddr;
f76aec76
KA		 2495 struct iova *iova;
2496 int prot = 0;
6865f0d1 2497 int ret;
8c11e798 2498 struct intel_iommu *iommu;
f76aec76
KA		 2499
2500 BUG_ON(dir == DMA_NONE);
2c2e2c38
FY		 2501
2502 if (iommu_no_mapping(pdev))
6865f0d1 2503 return paddr;
f76aec76
KA		 2504
2505 domain = get_valid_domain_for_dev(pdev);
2506 if (!domain)
2507 return 0;
2508
8c11e798 2509 iommu = domain_get_iommu(domain);
88cb6a74 2510 size = aligned_nrpages(paddr, size);
f76aec76 2511
5a5e02a6
DW		 2512 iova = intel_alloc_iova(hwdev, domain, dma_to_mm_pfn(size),
2513 pdev->dma_mask);
f76aec76
KA		 2514 if (!iova)
2515 goto error;
2516
ba395927
KA		 2517 /*
2518 * Check if DMAR supports zero-length reads on write only
2519 * mappings..
2520 */
2521 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
8c11e798 2522 !cap_zlr(iommu->cap))
ba395927
KA		 2523 prot |= DMA_PTE_READ;
2524 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2525 prot |= DMA_PTE_WRITE;
2526 /*
6865f0d1 2527 * paddr - (paddr + size) might be partial page, we should map the whole
ba395927 2528 * page. Note: if two part of one page are separately mapped, we
6865f0d1 2529 * might have two guest_addr mapping to the same host paddr, but this
ba395927
KA		 2530 * is not a big problem
2531 */
0ab36de2
DW		 2532 ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
2533 paddr >> VTD_PAGE_SHIFT, size, prot);
ba395927
KA		 2534 if (ret)
2535 goto error;
2536
1f0ef2aa
DW		 2537 /* it's a non-present to present mapping. Only flush if caching mode */
2538 if (cap_caching_mode(iommu->cap))
03d6a246 2539 iommu_flush_iotlb_psi(iommu, 0, mm_to_dma_pfn(iova->pfn_lo), size);
1f0ef2aa 2540 else
8c11e798 2541 iommu_flush_write_buffer(iommu);
f76aec76 2542
03d6a246
DW		 2543 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
2544 start_paddr += paddr & ~PAGE_MASK;
2545 return start_paddr;
ba395927 2546
ba395927 2547error:
f76aec76
KA		 2548 if (iova)
2549 __free_iova(&domain->iovad, iova);
4cf2e75d 2550 printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
5b6985ce 2551 pci_name(pdev), size, (unsigned long long)paddr, dir);
ba395927
KA		 2552 return 0;
2553}
2554
ffbbef5c
FT		 2555static dma_addr_t intel_map_page(struct device *dev, struct page *page,
2556 unsigned long offset, size_t size,
2557 enum dma_data_direction dir,
2558 struct dma_attrs *attrs)
bb9e6d65 2559{
ffbbef5c
FT		 2560 return __intel_map_single(dev, page_to_phys(page) + offset, size,
2561 dir, to_pci_dev(dev)->dma_mask);
bb9e6d65
FT		 2562}
2563
5e0d2a6f 2564static void flush_unmaps(void)
2565{
80b20dd8 2566 int i, j;
5e0d2a6f 2567
5e0d2a6f 2568 timer_on = 0;
2569
2570 /* just flush them all */
2571 for (i = 0; i < g_num_of_iommus; i++) {
a2bb8459
WH		 2572 struct intel_iommu *iommu = g_iommus[i];
2573 if (!iommu)
2574 continue;
c42d9f32 2575
9dd2fe89
YZ		 2576 if (!deferred_flush[i].next)
2577 continue;
2578
2579 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
93a23a72 2580 DMA_TLB_GLOBAL_FLUSH);
9dd2fe89 2581 for (j = 0; j < deferred_flush[i].next; j++) {
93a23a72
YZ		 2582 unsigned long mask;
2583 struct iova *iova = deferred_flush[i].iova[j];
2584
2585 mask = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT;
2586 mask = ilog2(mask >> VTD_PAGE_SHIFT);
2587 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
2588 iova->pfn_lo << PAGE_SHIFT, mask);
2589 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
80b20dd8 2590 }
9dd2fe89 2591 deferred_flush[i].next = 0;
5e0d2a6f 2592 }
2593
5e0d2a6f 2594 list_size = 0;
5e0d2a6f 2595}
2596
2597static void flush_unmaps_timeout(unsigned long data)
2598{
80b20dd8 2599 unsigned long flags;
2600
2601 spin_lock_irqsave(&async_umap_flush_lock, flags);
5e0d2a6f 2602 flush_unmaps();
80b20dd8 2603 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
5e0d2a6f 2604}
2605
2606static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2607{
2608 unsigned long flags;
80b20dd8 2609 int next, iommu_id;
8c11e798 2610 struct intel_iommu *iommu;
5e0d2a6f 2611
2612 spin_lock_irqsave(&async_umap_flush_lock, flags);
80b20dd8 2613 if (list_size == HIGH_WATER_MARK)
2614 flush_unmaps();
2615
8c11e798
WH		 2616 iommu = domain_get_iommu(dom);
2617 iommu_id = iommu->seq_id;
c42d9f32 2618
80b20dd8 2619 next = deferred_flush[iommu_id].next;
2620 deferred_flush[iommu_id].domain[next] = dom;
2621 deferred_flush[iommu_id].iova[next] = iova;
2622 deferred_flush[iommu_id].next++;
5e0d2a6f 2623
2624 if (!timer_on) {
2625 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2626 timer_on = 1;
2627 }
2628 list_size++;
2629 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2630}
2631
ffbbef5c
FT		 2632static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
2633 size_t size, enum dma_data_direction dir,
2634 struct dma_attrs *attrs)
ba395927 2635{
ba395927 2636 struct pci_dev *pdev = to_pci_dev(dev);
f76aec76 2637 struct dmar_domain *domain;
d794dc9b 2638 unsigned long start_pfn, last_pfn;
ba395927 2639 struct iova *iova;
8c11e798 2640 struct intel_iommu *iommu;
ba395927 2641
2c2e2c38 2642 if (iommu_no_mapping(pdev))
f76aec76 2643 return;
2c2e2c38 2644
ba395927
KA		 2645 domain = find_domain(pdev);
2646 BUG_ON(!domain);
2647
8c11e798
WH		 2648 iommu = domain_get_iommu(domain);
2649
ba395927 2650 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
85b98276
DW		 2651 if (WARN_ONCE(!iova, "Driver unmaps unmatched page at PFN %llx\n",
2652 (unsigned long long)dev_addr))
ba395927 2653 return;
ba395927 2654
d794dc9b
DW		 2655 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
2656 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
ba395927 2657
d794dc9b
DW		 2658 pr_debug("Device %s unmapping: pfn %lx-%lx\n",
2659 pci_name(pdev), start_pfn, last_pfn);
ba395927 2660
f76aec76 2661 /* clear the whole page */
d794dc9b
DW		 2662 dma_pte_clear_range(domain, start_pfn, last_pfn);
2663
f76aec76 2664 /* free page tables */
d794dc9b
DW		 2665 dma_pte_free_pagetable(domain, start_pfn, last_pfn);
2666
5e0d2a6f 2667 if (intel_iommu_strict) {
03d6a246 2668 iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
d794dc9b 2669 last_pfn - start_pfn + 1);
5e0d2a6f 2670 /* free iova */
2671 __free_iova(&domain->iovad, iova);
2672 } else {
2673 add_unmap(domain, iova);
2674 /*
2675 * queue up the release of the unmap to save the 1/6th of the
2676 * cpu used up by the iotlb flush operation...
2677 */
5e0d2a6f 2678 }
ba395927
KA		 2679}
2680
d7ab5c46
FT		 2681static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr, size_t size,
2682 int dir)
ffbbef5c
FT		 2683{
2684 intel_unmap_page(dev, dev_addr, size, dir, NULL);
2685}
2686
d7ab5c46
FT		 2687static void *intel_alloc_coherent(struct device *hwdev, size_t size,
2688 dma_addr_t *dma_handle, gfp_t flags)
ba395927
KA		 2689{
2690 void *vaddr;
2691 int order;
2692
5b6985ce 2693 size = PAGE_ALIGN(size);
ba395927
KA		 2694 order = get_order(size);
2695 flags &= ~(GFP_DMA | GFP_DMA32);
2696
2697 vaddr = (void *)__get_free_pages(flags, order);
2698 if (!vaddr)
2699 return NULL;
2700 memset(vaddr, 0, size);
2701
bb9e6d65
FT		 2702 *dma_handle = __intel_map_single(hwdev, virt_to_bus(vaddr), size,
2703 DMA_BIDIRECTIONAL,
2704 hwdev->coherent_dma_mask);
ba395927
KA		 2705 if (*dma_handle)
2706 return vaddr;
2707 free_pages((unsigned long)vaddr, order);
2708 return NULL;
2709}
2710
d7ab5c46
FT		 2711static void intel_free_coherent(struct device *hwdev, size_t size, void *vaddr,
2712 dma_addr_t dma_handle)
ba395927
KA		 2713{
2714 int order;
2715
5b6985ce 2716 size = PAGE_ALIGN(size);
ba395927
KA		 2717 order = get_order(size);
2718
2719 intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2720 free_pages((unsigned long)vaddr, order);
2721}
2722
d7ab5c46
FT		 2723static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2724 int nelems, enum dma_data_direction dir,
2725 struct dma_attrs *attrs)
ba395927 2726{
ba395927
KA		 2727 struct pci_dev *pdev = to_pci_dev(hwdev);
2728 struct dmar_domain *domain;
d794dc9b 2729 unsigned long start_pfn, last_pfn;
f76aec76 2730 struct iova *iova;
8c11e798 2731 struct intel_iommu *iommu;
ba395927 2732
2c2e2c38 2733 if (iommu_no_mapping(pdev))
ba395927
KA		 2734 return;
2735
2736 domain = find_domain(pdev);
8c11e798
WH		 2737 BUG_ON(!domain);
2738
2739 iommu = domain_get_iommu(domain);
ba395927 2740
c03ab37c 2741 iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
85b98276
DW		 2742 if (WARN_ONCE(!iova, "Driver unmaps unmatched sglist at PFN %llx\n",
2743 (unsigned long long)sglist[0].dma_address))
f76aec76 2744 return;
f76aec76 2745
d794dc9b
DW		 2746 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
2747 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
f76aec76
KA		 2748
2749 /* clear the whole page */
d794dc9b
DW		 2750 dma_pte_clear_range(domain, start_pfn, last_pfn);
2751
f76aec76 2752 /* free page tables */
d794dc9b 2753 dma_pte_free_pagetable(domain, start_pfn, last_pfn);
f76aec76 2754
03d6a246 2755 iommu_flush_iotlb_psi(iommu, domain->id, start_pfn,
d794dc9b 2756 (last_pfn - start_pfn + 1));
f76aec76
KA		 2757
2758 /* free iova */
2759 __free_iova(&domain->iovad, iova);
ba395927
KA		 2760}
2761
ba395927 2762static int intel_nontranslate_map_sg(struct device *hddev,
c03ab37c 2763 struct scatterlist *sglist, int nelems, int dir)
ba395927
KA		 2764{
2765 int i;
c03ab37c 2766 struct scatterlist *sg;
ba395927 2767
c03ab37c 2768 for_each_sg(sglist, sg, nelems, i) {
12d4d40e 2769 BUG_ON(!sg_page(sg));
4cf2e75d 2770 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
c03ab37c 2771 sg->dma_length = sg->length;
ba395927
KA		 2772 }
2773 return nelems;
2774}
2775
d7ab5c46
FT		 2776static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist, int nelems,
2777 enum dma_data_direction dir, struct dma_attrs *attrs)
ba395927 2778{
ba395927 2779 int i;
ba395927
KA		 2780 struct pci_dev *pdev = to_pci_dev(hwdev);
2781 struct dmar_domain *domain;
f76aec76
KA		 2782 size_t size = 0;
2783 int prot = 0;
b536d24d 2784 size_t offset_pfn = 0;
f76aec76
KA		 2785 struct iova *iova = NULL;
2786 int ret;
c03ab37c 2787 struct scatterlist *sg;
b536d24d 2788 unsigned long start_vpfn;
8c11e798 2789 struct intel_iommu *iommu;
ba395927
KA		 2790
2791 BUG_ON(dir == DMA_NONE);
2c2e2c38 2792 if (iommu_no_mapping(pdev))
c03ab37c 2793 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
ba395927 2794
f76aec76
KA		 2795 domain = get_valid_domain_for_dev(pdev);
2796 if (!domain)
2797 return 0;
2798
8c11e798
WH		 2799 iommu = domain_get_iommu(domain);
2800
b536d24d 2801 for_each_sg(sglist, sg, nelems, i)
88cb6a74 2802 size += aligned_nrpages(sg->offset, sg->length);
f76aec76 2803
5a5e02a6
DW		 2804 iova = intel_alloc_iova(hwdev, domain, dma_to_mm_pfn(size),
2805 pdev->dma_mask);
f76aec76 2806 if (!iova) {
c03ab37c 2807 sglist->dma_length = 0;
f76aec76
KA		 2808 return 0;
2809 }
2810
2811 /*
2812 * Check if DMAR supports zero-length reads on write only
2813 * mappings..
2814 */
2815 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
8c11e798 2816 !cap_zlr(iommu->cap))
f76aec76
KA		 2817 prot |= DMA_PTE_READ;
2818 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2819 prot |= DMA_PTE_WRITE;
2820
b536d24d 2821 start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
e1605495
DW		 2822
2823 ret = domain_sg_mapping(domain, start_vpfn, sglist, mm_to_dma_pfn(size), prot);
2824 if (unlikely(ret)) {
2825 /* clear the page */
2826 dma_pte_clear_range(domain, start_vpfn,
2827 start_vpfn + size - 1);
2828 /* free page tables */
2829 dma_pte_free_pagetable(domain, start_vpfn,
2830 start_vpfn + size - 1);
2831 /* free iova */
2832 __free_iova(&domain->iovad, iova);
2833 return 0;
ba395927
KA		 2834 }
2835
1f0ef2aa
DW		 2836 /* it's a non-present to present mapping. Only flush if caching mode */
2837 if (cap_caching_mode(iommu->cap))
03d6a246 2838 iommu_flush_iotlb_psi(iommu, 0, start_vpfn, offset_pfn);
1f0ef2aa 2839 else
8c11e798 2840 iommu_flush_write_buffer(iommu);
1f0ef2aa 2841
ba395927
KA		 2842 return nelems;
2843}
2844
dfb805e8
FT		 2845static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
2846{
2847 return !dma_addr;
2848}
2849
160c1d8e 2850struct dma_map_ops intel_dma_ops = {
ba395927
KA		 2851 .alloc_coherent = intel_alloc_coherent,
2852 .free_coherent = intel_free_coherent,
ba395927
KA		 2853 .map_sg = intel_map_sg,
2854 .unmap_sg = intel_unmap_sg,
ffbbef5c
FT		 2855 .map_page = intel_map_page,
2856 .unmap_page = intel_unmap_page,
dfb805e8 2857 .mapping_error = intel_mapping_error,
ba395927
KA		 2858};
2859
2860static inline int iommu_domain_cache_init(void)
2861{
2862 int ret = 0;
2863
2864 iommu_domain_cache = kmem_cache_create("iommu_domain",
2865 sizeof(struct dmar_domain),
2866 0,
2867 SLAB_HWCACHE_ALIGN,
2868
2869 NULL);
2870 if (!iommu_domain_cache) {
2871 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2872 ret = -ENOMEM;
2873 }
2874
2875 return ret;
2876}
2877
2878static inline int iommu_devinfo_cache_init(void)
2879{
2880 int ret = 0;
2881
2882 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2883 sizeof(struct device_domain_info),
2884 0,
2885 SLAB_HWCACHE_ALIGN,
ba395927
KA		 2886 NULL);
2887 if (!iommu_devinfo_cache) {
2888 printk(KERN_ERR "Couldn't create devinfo cache\n");
2889 ret = -ENOMEM;
2890 }
2891
2892 return ret;
2893}
2894
2895static inline int iommu_iova_cache_init(void)
2896{
2897 int ret = 0;
2898
2899 iommu_iova_cache = kmem_cache_create("iommu_iova",
2900 sizeof(struct iova),
2901 0,
2902 SLAB_HWCACHE_ALIGN,
ba395927
KA		 2903 NULL);
2904 if (!iommu_iova_cache) {
2905 printk(KERN_ERR "Couldn't create iova cache\n");
2906 ret = -ENOMEM;
2907 }
2908
2909 return ret;
2910}
2911
2912static int __init iommu_init_mempool(void)
2913{
2914 int ret;
2915 ret = iommu_iova_cache_init();
2916 if (ret)
2917 return ret;
2918
2919 ret = iommu_domain_cache_init();
2920 if (ret)
2921 goto domain_error;
2922
2923 ret = iommu_devinfo_cache_init();
2924 if (!ret)
2925 return ret;
2926
2927 kmem_cache_destroy(iommu_domain_cache);
2928domain_error:
2929 kmem_cache_destroy(iommu_iova_cache);
2930
2931 return -ENOMEM;
2932}
2933
2934static void __init iommu_exit_mempool(void)
2935{
2936 kmem_cache_destroy(iommu_devinfo_cache);
2937 kmem_cache_destroy(iommu_domain_cache);
2938 kmem_cache_destroy(iommu_iova_cache);
2939
2940}
2941
ba395927
KA		 2942static void __init init_no_remapping_devices(void)
2943{
2944 struct dmar_drhd_unit *drhd;
2945
2946 for_each_drhd_unit(drhd) {
2947 if (!drhd->include_all) {
2948 int i;
2949 for (i = 0; i < drhd->devices_cnt; i++)
2950 if (drhd->devices[i] != NULL)
2951 break;
2952 /* ignore DMAR unit if no pci devices exist */
2953 if (i == drhd->devices_cnt)
2954 drhd->ignored = 1;
2955 }
2956 }
2957
2958 if (dmar_map_gfx)
2959 return;
2960
2961 for_each_drhd_unit(drhd) {
2962 int i;
2963 if (drhd->ignored || drhd->include_all)
2964 continue;
2965
2966 for (i = 0; i < drhd->devices_cnt; i++)
2967 if (drhd->devices[i] &&
2968 !IS_GFX_DEVICE(drhd->devices[i]))
2969 break;
2970
2971 if (i < drhd->devices_cnt)
2972 continue;
2973
2974 /* bypass IOMMU if it is just for gfx devices */
2975 drhd->ignored = 1;
2976 for (i = 0; i < drhd->devices_cnt; i++) {
2977 if (!drhd->devices[i])
2978 continue;
358dd8ac 2979 drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
ba395927
KA		 2980 }
2981 }
2982}
2983
f59c7b69
FY		 2984#ifdef CONFIG_SUSPEND
2985static int init_iommu_hw(void)
2986{
2987 struct dmar_drhd_unit *drhd;
2988 struct intel_iommu *iommu = NULL;
2989
2990 for_each_active_iommu(iommu, drhd)
2991 if (iommu->qi)
2992 dmar_reenable_qi(iommu);
2993
2994 for_each_active_iommu(iommu, drhd) {
2995 iommu_flush_write_buffer(iommu);
2996
2997 iommu_set_root_entry(iommu);
2998
2999 iommu->flush.flush_context(iommu, 0, 0, 0,
1f0ef2aa 3000 DMA_CCMD_GLOBAL_INVL);
f59c7b69 3001 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
1f0ef2aa 3002 DMA_TLB_GLOBAL_FLUSH);
f59c7b69
FY		 3003 iommu_disable_protect_mem_regions(iommu);
3004 iommu_enable_translation(iommu);
3005 }
3006
3007 return 0;
3008}
3009
3010static void iommu_flush_all(void)
3011{
3012 struct dmar_drhd_unit *drhd;
3013 struct intel_iommu *iommu;
3014
3015 for_each_active_iommu(iommu, drhd) {
3016 iommu->flush.flush_context(iommu, 0, 0, 0,
1f0ef2aa 3017 DMA_CCMD_GLOBAL_INVL);
f59c7b69 3018 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
1f0ef2aa 3019 DMA_TLB_GLOBAL_FLUSH);
f59c7b69
FY		 3020 }
3021}
3022
3023static int iommu_suspend(struct sys_device *dev, pm_message_t state)
3024{
3025 struct dmar_drhd_unit *drhd;
3026 struct intel_iommu *iommu = NULL;
3027 unsigned long flag;
3028
3029 for_each_active_iommu(iommu, drhd) {
3030 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3031 GFP_ATOMIC);
3032 if (!iommu->iommu_state)
3033 goto nomem;
3034 }
3035
3036 iommu_flush_all();
3037
3038 for_each_active_iommu(iommu, drhd) {
3039 iommu_disable_translation(iommu);
3040
3041 spin_lock_irqsave(&iommu->register_lock, flag);
3042
3043 iommu->iommu_state[SR_DMAR_FECTL_REG] =
3044 readl(iommu->reg + DMAR_FECTL_REG);
3045 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
3046 readl(iommu->reg + DMAR_FEDATA_REG);
3047 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
3048 readl(iommu->reg + DMAR_FEADDR_REG);
3049 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
3050 readl(iommu->reg + DMAR_FEUADDR_REG);
3051
3052 spin_unlock_irqrestore(&iommu->register_lock, flag);
3053 }
3054 return 0;
3055
3056nomem:
3057 for_each_active_iommu(iommu, drhd)
3058 kfree(iommu->iommu_state);
3059
3060 return -ENOMEM;
3061}
3062
3063static int iommu_resume(struct sys_device *dev)
3064{
3065 struct dmar_drhd_unit *drhd;
3066 struct intel_iommu *iommu = NULL;
3067 unsigned long flag;
3068
3069 if (init_iommu_hw()) {
3070 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
3071 return -EIO;
3072 }
3073
3074 for_each_active_iommu(iommu, drhd) {
3075
3076 spin_lock_irqsave(&iommu->register_lock, flag);
3077
3078 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
3079 iommu->reg + DMAR_FECTL_REG);
3080 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
3081 iommu->reg + DMAR_FEDATA_REG);
3082 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
3083 iommu->reg + DMAR_FEADDR_REG);
3084 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
3085 iommu->reg + DMAR_FEUADDR_REG);
3086
3087 spin_unlock_irqrestore(&iommu->register_lock, flag);
3088 }
3089
3090 for_each_active_iommu(iommu, drhd)
3091 kfree(iommu->iommu_state);
3092
3093 return 0;
3094}
3095
3096static struct sysdev_class iommu_sysclass = {
3097 .name = "iommu",
3098 .resume = iommu_resume,
3099 .suspend = iommu_suspend,
3100};
3101
3102static struct sys_device device_iommu = {
3103 .cls = &iommu_sysclass,
3104};
3105
3106static int __init init_iommu_sysfs(void)
3107{
3108 int error;
3109
3110 error = sysdev_class_register(&iommu_sysclass);
3111 if (error)
3112 return error;
3113
3114 error = sysdev_register(&device_iommu);
3115 if (error)
3116 sysdev_class_unregister(&iommu_sysclass);
3117
3118 return error;
3119}
3120
3121#else
3122static int __init init_iommu_sysfs(void)
3123{
3124 return 0;
3125}
3126#endif /* CONFIG_PM */
3127
ba395927
KA		 3128int __init intel_iommu_init(void)
3129{
3130 int ret = 0;
3131
ba395927
KA		 3132 if (dmar_table_init())
3133 return -ENODEV;
3134
1886e8a9
SS		 3135 if (dmar_dev_scope_init())
3136 return -ENODEV;
3137
2ae21010
SS		 3138 /*
3139 * Check the need for DMA-remapping initialization now.
3140 * Above initialization will also be used by Interrupt-remapping.
3141 */
4ed0d3e6 3142 if (no_iommu || (swiotlb && !iommu_pass_through) || dmar_disabled)
2ae21010
SS		 3143 return -ENODEV;
3144
ba395927
KA		 3145 iommu_init_mempool();
3146 dmar_init_reserved_ranges();
3147
3148 init_no_remapping_devices();
3149
3150 ret = init_dmars();
3151 if (ret) {
3152 printk(KERN_ERR "IOMMU: dmar init failed\n");
3153 put_iova_domain(&reserved_iova_list);
3154 iommu_exit_mempool();
3155 return ret;
3156 }
3157 printk(KERN_INFO
3158 "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
3159
5e0d2a6f 3160 init_timer(&unmap_timer);
ba395927 3161 force_iommu = 1;
4ed0d3e6
FY		 3162
3163 if (!iommu_pass_through) {
3164 printk(KERN_INFO
3165 "Multi-level page-table translation for DMAR.\n");
3166 dma_ops = &intel_dma_ops;
3167 } else
3168 printk(KERN_INFO
3169 "DMAR: Pass through translation for DMAR.\n");
3170
f59c7b69 3171 init_iommu_sysfs();
a8bcbb0d
JR		 3172
3173 register_iommu(&intel_iommu_ops);
3174
ba395927
KA		 3175 return 0;
3176}
e820482c 3177
3199aa6b
HW		 3178static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
3179 struct pci_dev *pdev)
3180{
3181 struct pci_dev *tmp, *parent;
3182
3183 if (!iommu || !pdev)
3184 return;
3185
3186 /* dependent device detach */
3187 tmp = pci_find_upstream_pcie_bridge(pdev);
3188 /* Secondary interface's bus number and devfn 0 */
3189 if (tmp) {
3190 parent = pdev->bus->self;
3191 while (parent != tmp) {
3192 iommu_detach_dev(iommu, parent->bus->number,
276dbf99 3193 parent->devfn);
3199aa6b
HW		 3194 parent = parent->bus->self;
3195 }
3196 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
3197 iommu_detach_dev(iommu,
3198 tmp->subordinate->number, 0);
3199 else /* this is a legacy PCI bridge */
276dbf99
DW		 3200 iommu_detach_dev(iommu, tmp->bus->number,
3201 tmp->devfn);
3199aa6b
HW		 3202 }
3203}
3204
2c2e2c38 3205static void domain_remove_one_dev_info(struct dmar_domain *domain,
c7151a8d
WH		 3206 struct pci_dev *pdev)
3207{
3208 struct device_domain_info *info;
3209 struct intel_iommu *iommu;
3210 unsigned long flags;
3211 int found = 0;
3212 struct list_head *entry, *tmp;
3213
276dbf99
DW		 3214 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3215 pdev->devfn);
c7151a8d
WH		 3216 if (!iommu)
3217 return;
3218
3219 spin_lock_irqsave(&device_domain_lock, flags);
3220 list_for_each_safe(entry, tmp, &domain->devices) {
3221 info = list_entry(entry, struct device_domain_info, link);
276dbf99 3222 /* No need to compare PCI domain; it has to be the same */
c7151a8d
WH		 3223 if (info->bus == pdev->bus->number &&
3224 info->devfn == pdev->devfn) {
3225 list_del(&info->link);
3226 list_del(&info->global);
3227 if (info->dev)
3228 info->dev->dev.archdata.iommu = NULL;
3229 spin_unlock_irqrestore(&device_domain_lock, flags);
3230
93a23a72 3231 iommu_disable_dev_iotlb(info);
c7151a8d 3232 iommu_detach_dev(iommu, info->bus, info->devfn);
3199aa6b 3233 iommu_detach_dependent_devices(iommu, pdev);
c7151a8d
WH		 3234 free_devinfo_mem(info);
3235
3236 spin_lock_irqsave(&device_domain_lock, flags);
3237
3238 if (found)
3239 break;
3240 else
3241 continue;
3242 }
3243
3244 /* if there is no other devices under the same iommu
3245 * owned by this domain, clear this iommu in iommu_bmp
3246 * update iommu count and coherency
3247 */
276dbf99
DW		 3248 if (iommu == device_to_iommu(info->segment, info->bus,
3249 info->devfn))
c7151a8d
WH		 3250 found = 1;
3251 }
3252
3253 if (found == 0) {
3254 unsigned long tmp_flags;
3255 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
3256 clear_bit(iommu->seq_id, &domain->iommu_bmp);
3257 domain->iommu_count--;
58c610bd 3258 domain_update_iommu_cap(domain);
c7151a8d
WH		 3259 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
3260 }
3261
3262 spin_unlock_irqrestore(&device_domain_lock, flags);
3263}
3264
3265static void vm_domain_remove_all_dev_info(struct dmar_domain *domain)
3266{
3267 struct device_domain_info *info;
3268 struct intel_iommu *iommu;
3269 unsigned long flags1, flags2;
3270
3271 spin_lock_irqsave(&device_domain_lock, flags1);
3272 while (!list_empty(&domain->devices)) {
3273 info = list_entry(domain->devices.next,
3274 struct device_domain_info, link);
3275 list_del(&info->link);
3276 list_del(&info->global);
3277 if (info->dev)
3278 info->dev->dev.archdata.iommu = NULL;
3279
3280 spin_unlock_irqrestore(&device_domain_lock, flags1);
3281
93a23a72 3282 iommu_disable_dev_iotlb(info);
276dbf99 3283 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
c7151a8d 3284 iommu_detach_dev(iommu, info->bus, info->devfn);
3199aa6b 3285 iommu_detach_dependent_devices(iommu, info->dev);
c7151a8d
WH		 3286
3287 /* clear this iommu in iommu_bmp, update iommu count
58c610bd 3288 * and capabilities
c7151a8d
WH		 3289 */
3290 spin_lock_irqsave(&domain->iommu_lock, flags2);
3291 if (test_and_clear_bit(iommu->seq_id,
3292 &domain->iommu_bmp)) {
3293 domain->iommu_count--;
58c610bd 3294 domain_update_iommu_cap(domain);
c7151a8d
WH		 3295 }
3296 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
3297
3298 free_devinfo_mem(info);
3299 spin_lock_irqsave(&device_domain_lock, flags1);
3300 }
3301 spin_unlock_irqrestore(&device_domain_lock, flags1);
3302}
3303
5e98c4b1
WH		 3304/* domain id for virtual machine, it won't be set in context */
3305static unsigned long vm_domid;
3306
fe40f1e0
WH		 3307static int vm_domain_min_agaw(struct dmar_domain *domain)
3308{
3309 int i;
3310 int min_agaw = domain->agaw;
3311
3312 i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
3313 for (; i < g_num_of_iommus; ) {
3314 if (min_agaw > g_iommus[i]->agaw)
3315 min_agaw = g_iommus[i]->agaw;
3316
3317 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
3318 }
3319
3320 return min_agaw;
3321}
3322
5e98c4b1
WH		 3323static struct dmar_domain *iommu_alloc_vm_domain(void)
3324{
3325 struct dmar_domain *domain;
3326
3327 domain = alloc_domain_mem();
3328 if (!domain)
3329 return NULL;
3330
3331 domain->id = vm_domid++;
3332 memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
3333 domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
3334
3335 return domain;
3336}
3337
2c2e2c38 3338static int md_domain_init(struct dmar_domain *domain, int guest_width)
5e98c4b1
WH		 3339{
3340 int adjust_width;
3341
3342 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
5e98c4b1
WH		 3343 spin_lock_init(&domain->iommu_lock);
3344
3345 domain_reserve_special_ranges(domain);
3346
3347 /* calculate AGAW */
3348 domain->gaw = guest_width;
3349 adjust_width = guestwidth_to_adjustwidth(guest_width);
3350 domain->agaw = width_to_agaw(adjust_width);
3351
3352 INIT_LIST_HEAD(&domain->devices);
3353
3354 domain->iommu_count = 0;
3355 domain->iommu_coherency = 0;
fe40f1e0 3356 domain->max_addr = 0;
5e98c4b1
WH		 3357
3358 /* always allocate the top pgd */
3359 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
3360 if (!domain->pgd)
3361 return -ENOMEM;
3362 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
3363 return 0;
3364}
3365
3366static void iommu_free_vm_domain(struct dmar_domain *domain)
3367{
3368 unsigned long flags;
3369 struct dmar_drhd_unit *drhd;
3370 struct intel_iommu *iommu;
3371 unsigned long i;
3372 unsigned long ndomains;
3373
3374 for_each_drhd_unit(drhd) {
3375 if (drhd->ignored)
3376 continue;
3377 iommu = drhd->iommu;
3378
3379 ndomains = cap_ndoms(iommu->cap);
3380 i = find_first_bit(iommu->domain_ids, ndomains);
3381 for (; i < ndomains; ) {
3382 if (iommu->domains[i] == domain) {
3383 spin_lock_irqsave(&iommu->lock, flags);
3384 clear_bit(i, iommu->domain_ids);
3385 iommu->domains[i] = NULL;
3386 spin_unlock_irqrestore(&iommu->lock, flags);
3387 break;
3388 }
3389 i = find_next_bit(iommu->domain_ids, ndomains, i+1);
3390 }
3391 }
3392}
3393
3394static void vm_domain_exit(struct dmar_domain *domain)
3395{
5e98c4b1
WH		 3396 /* Domain 0 is reserved, so dont process it */
3397 if (!domain)
3398 return;
3399
3400 vm_domain_remove_all_dev_info(domain);
3401 /* destroy iovas */
3402 put_iova_domain(&domain->iovad);
5e98c4b1
WH		 3403
3404 /* clear ptes */
595badf5 3405 dma_pte_clear_range(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
5e98c4b1
WH		 3406
3407 /* free page tables */
d794dc9b 3408 dma_pte_free_pagetable(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
5e98c4b1
WH		 3409
3410 iommu_free_vm_domain(domain);
3411 free_domain_mem(domain);
3412}
3413
5d450806 3414static int intel_iommu_domain_init(struct iommu_domain *domain)
38717946 3415{
5d450806 3416 struct dmar_domain *dmar_domain;
38717946 3417
5d450806
JR		 3418 dmar_domain = iommu_alloc_vm_domain();
3419 if (!dmar_domain) {
38717946 3420 printk(KERN_ERR
5d450806
JR		 3421 "intel_iommu_domain_init: dmar_domain == NULL\n");
3422 return -ENOMEM;
38717946 3423 }
2c2e2c38 3424 if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
38717946 3425 printk(KERN_ERR
5d450806
JR		 3426 "intel_iommu_domain_init() failed\n");
3427 vm_domain_exit(dmar_domain);
3428 return -ENOMEM;
38717946 3429 }
5d450806 3430 domain->priv = dmar_domain;
faa3d6f5 3431
5d450806 3432 return 0;
38717946 3433}
38717946 3434
5d450806 3435static void intel_iommu_domain_destroy(struct iommu_domain *domain)
38717946 3436{
5d450806
JR		 3437 struct dmar_domain *dmar_domain = domain->priv;
3438
3439 domain->priv = NULL;
3440 vm_domain_exit(dmar_domain);
38717946 3441}
38717946 3442
4c5478c9
JR		 3443static int intel_iommu_attach_device(struct iommu_domain *domain,
3444 struct device *dev)
38717946 3445{
4c5478c9
JR		 3446 struct dmar_domain *dmar_domain = domain->priv;
3447 struct pci_dev *pdev = to_pci_dev(dev);
fe40f1e0
WH		 3448 struct intel_iommu *iommu;
3449 int addr_width;
3450 u64 end;
faa3d6f5
WH		 3451 int ret;
3452
3453 /* normally pdev is not mapped */
3454 if (unlikely(domain_context_mapped(pdev))) {
3455 struct dmar_domain *old_domain;
3456
3457 old_domain = find_domain(pdev);
3458 if (old_domain) {
2c2e2c38
FY		 3459 if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE ||
3460 dmar_domain->flags & DOMAIN_FLAG_STATIC_IDENTITY)
3461 domain_remove_one_dev_info(old_domain, pdev);
faa3d6f5
WH		 3462 else
3463 domain_remove_dev_info(old_domain);
3464 }
3465 }
3466
276dbf99
DW		 3467 iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3468 pdev->devfn);
fe40f1e0
WH		 3469 if (!iommu)
3470 return -ENODEV;
3471
3472 /* check if this iommu agaw is sufficient for max mapped address */
3473 addr_width = agaw_to_width(iommu->agaw);
3474 end = DOMAIN_MAX_ADDR(addr_width);
3475 end = end & VTD_PAGE_MASK;
4c5478c9 3476 if (end < dmar_domain->max_addr) {
fe40f1e0
WH		 3477 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3478 "sufficient for the mapped address (%llx)\n",
4c5478c9 3479 __func__, iommu->agaw, dmar_domain->max_addr);
fe40f1e0
WH		 3480 return -EFAULT;
3481 }
3482
2c2e2c38 3483 ret = domain_add_dev_info(dmar_domain, pdev);
faa3d6f5
WH		 3484 if (ret)
3485 return ret;
3486
93a23a72 3487 ret = domain_context_mapping(dmar_domain, pdev, CONTEXT_TT_MULTI_LEVEL);
faa3d6f5 3488 return ret;
38717946 3489}
38717946 3490
4c5478c9
JR		 3491static void intel_iommu_detach_device(struct iommu_domain *domain,
3492 struct device *dev)
38717946 3493{
4c5478c9
JR		 3494 struct dmar_domain *dmar_domain = domain->priv;
3495 struct pci_dev *pdev = to_pci_dev(dev);
3496
2c2e2c38 3497 domain_remove_one_dev_info(dmar_domain, pdev);
faa3d6f5 3498}
c7151a8d 3499
dde57a21
JR		 3500static int intel_iommu_map_range(struct iommu_domain *domain,
3501 unsigned long iova, phys_addr_t hpa,
3502 size_t size, int iommu_prot)
faa3d6f5 3503{
dde57a21 3504 struct dmar_domain *dmar_domain = domain->priv;
fe40f1e0
WH		 3505 u64 max_addr;
3506 int addr_width;
dde57a21 3507 int prot = 0;
faa3d6f5 3508 int ret;
fe40f1e0 3509
dde57a21
JR		 3510 if (iommu_prot & IOMMU_READ)
3511 prot |= DMA_PTE_READ;
3512 if (iommu_prot & IOMMU_WRITE)
3513 prot |= DMA_PTE_WRITE;
9cf06697
SY		 3514 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
3515 prot |= DMA_PTE_SNP;
dde57a21 3516
163cc52c 3517 max_addr = iova + size;
dde57a21 3518 if (dmar_domain->max_addr < max_addr) {
fe40f1e0
WH		 3519 int min_agaw;
3520 u64 end;
3521
3522 /* check if minimum agaw is sufficient for mapped address */
dde57a21 3523 min_agaw = vm_domain_min_agaw(dmar_domain);
fe40f1e0
WH		 3524 addr_width = agaw_to_width(min_agaw);
3525 end = DOMAIN_MAX_ADDR(addr_width);
3526 end = end & VTD_PAGE_MASK;
3527 if (end < max_addr) {
3528 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3529 "sufficient for the mapped address (%llx)\n",
3530 __func__, min_agaw, max_addr);
3531 return -EFAULT;
3532 }
dde57a21 3533 dmar_domain->max_addr = max_addr;
fe40f1e0 3534 }
ad051221
DW		 3535 /* Round up size to next multiple of PAGE_SIZE, if it and
3536 the low bits of hpa would take us onto the next page */
88cb6a74 3537 size = aligned_nrpages(hpa, size);
ad051221
DW		 3538 ret = domain_pfn_mapping(dmar_domain, iova >> VTD_PAGE_SHIFT,
3539 hpa >> VTD_PAGE_SHIFT, size, prot);
faa3d6f5 3540 return ret;
38717946 3541}
38717946 3542
dde57a21
JR		 3543static void intel_iommu_unmap_range(struct iommu_domain *domain,
3544 unsigned long iova, size_t size)
38717946 3545{
dde57a21 3546 struct dmar_domain *dmar_domain = domain->priv;
faa3d6f5 3547
163cc52c
DW		 3548 dma_pte_clear_range(dmar_domain, iova >> VTD_PAGE_SHIFT,
3549 (iova + size - 1) >> VTD_PAGE_SHIFT);
fe40f1e0 3550
163cc52c
DW		 3551 if (dmar_domain->max_addr == iova + size)
3552 dmar_domain->max_addr = iova;
38717946 3553}
38717946 3554
d14d6577
JR		 3555static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
3556 unsigned long iova)
38717946 3557{
d14d6577 3558 struct dmar_domain *dmar_domain = domain->priv;
38717946 3559 struct dma_pte *pte;
faa3d6f5 3560 u64 phys = 0;
38717946 3561
b026fd28 3562 pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT);
38717946 3563 if (pte)
faa3d6f5 3564 phys = dma_pte_addr(pte);
38717946 3565
faa3d6f5 3566 return phys;
38717946 3567}
a8bcbb0d 3568
dbb9fd86
SY		 3569static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
3570 unsigned long cap)
3571{
3572 struct dmar_domain *dmar_domain = domain->priv;
3573
3574 if (cap == IOMMU_CAP_CACHE_COHERENCY)
3575 return dmar_domain->iommu_snooping;
3576
3577 return 0;
3578}
3579
a8bcbb0d
JR		 3580static struct iommu_ops intel_iommu_ops = {
3581 .domain_init = intel_iommu_domain_init,
3582 .domain_destroy = intel_iommu_domain_destroy,
3583 .attach_dev = intel_iommu_attach_device,
3584 .detach_dev = intel_iommu_detach_device,
3585 .map = intel_iommu_map_range,
3586 .unmap = intel_iommu_unmap_range,
3587 .iova_to_phys = intel_iommu_iova_to_phys,
dbb9fd86 3588 .domain_has_cap = intel_iommu_domain_has_cap,
a8bcbb0d 3589};
9af88143
DW		 3590
3591static void __devinit quirk_iommu_rwbf(struct pci_dev *dev)
3592{
3593 /*
3594 * Mobile 4 Series Chipset neglects to set RWBF capability,
3595 * but needs it:
3596 */
3597 printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
3598 rwbf_quirk = 1;
3599}
3600
3601DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);
Linux kernel - Deliverables
This page took 0.428366 seconds and 5 git commands to generate.