drivers/pci/intel-iommu.c

   1 /*
   2  * Copyright (c) 2006, Intel Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Copyright (C) 2006-2008 Intel Corporation
  18  * Author: Ashok Raj <ashok.raj@intel.com>
  19  * Author: Shaohua Li <shaohua.li@intel.com>
  20  * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
  21  * Author: Fenghua Yu <fenghua.yu@intel.com>
  22  */
  23
  24 #include <linux/init.h>
  25 #include <linux/bitmap.h>
  26 #include <linux/debugfs.h>
  27 #include <linux/slab.h>
  28 #include <linux/irq.h>
  29 #include <linux/interrupt.h>
  30 #include <linux/spinlock.h>
  31 #include <linux/pci.h>
  32 #include <linux/dmar.h>
  33 #include <linux/dma-mapping.h>
  34 #include <linux/mempool.h>
  35 #include <linux/timer.h>
  36 #include <linux/iova.h>
  37 #include <linux/iommu.h>
  38 #include <linux/intel-iommu.h>
  39 #include <linux/sysdev.h>
  40 #include <asm/cacheflush.h>
  41 #include <asm/iommu.h>
  42 #include "pci.h"
  43
  44 #define ROOT_SIZE               VTD_PAGE_SIZE
  45 #define CONTEXT_SIZE            VTD_PAGE_SIZE
  46
  47 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
  48 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
  49
  50 #define IOAPIC_RANGE_START      (0xfee00000)
  51 #define IOAPIC_RANGE_END        (0xfeefffff)
  52 #define IOVA_START_ADDR         (0x1000)
  53
  54 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
  55
  56 #define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
  57
  58 #define IOVA_PFN(addr)          ((addr) >> PAGE_SHIFT)
  59 #define DMA_32BIT_PFN           IOVA_PFN(DMA_BIT_MASK(32))
  60 #define DMA_64BIT_PFN           IOVA_PFN(DMA_BIT_MASK(64))
  61
  62 /* global iommu list, set NULL for ignored DMAR units */
  63 static struct intel_iommu **g_iommus;
  64
  65 static int rwbf_quirk;
  66
  67 /*
  68  * 0: Present
  69  * 1-11: Reserved
  70  * 12-63: Context Ptr (12 - (haw-1))
  71  * 64-127: Reserved
  72  */
  73 struct root_entry {
  74         u64     val;
  75         u64     rsvd1;
  76 };
  77 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
  78 static inline bool root_present(struct root_entry *root)
  79 {
  80         return (root->val & 1);
  81 }
  82 static inline void set_root_present(struct root_entry *root)
  83 {
  84         root->val |= 1;
  85 }
  86 static inline void set_root_value(struct root_entry *root, unsigned long value)
  87 {
  88         root->val |= value & VTD_PAGE_MASK;
  89 }
  90
  91 static inline struct context_entry *
  92 get_context_addr_from_root(struct root_entry *root)
  93 {
  94         return (struct context_entry *)
  95                 (root_present(root)?phys_to_virt(
  96                 root->val & VTD_PAGE_MASK) :
  97                 NULL);
  98 }
  99
 100 /*
 101  * low 64 bits:
 102  * 0: present
 103  * 1: fault processing disable
 104  * 2-3: translation type
 105  * 12-63: address space root
 106  * high 64 bits:
 107  * 0-2: address width
 108  * 3-6: aval
 109  * 8-23: domain id
 110  */
 111 struct context_entry {
 112         u64 lo;
 113         u64 hi;
 114 };
 115
 116 static inline bool context_present(struct context_entry *context)
 117 {
 118         return (context->lo & 1);
 119 }
 120 static inline void context_set_present(struct context_entry *context)
 121 {
 122         context->lo |= 1;
 123 }
 124
 125 static inline void context_set_fault_enable(struct context_entry *context)
 126 {
 127         context->lo &= (((u64)-1) << 2) | 1;
 128 }
 129
 130 #define CONTEXT_TT_MULTI_LEVEL 0
 131
 132 static inline void context_set_translation_type(struct context_entry *context,
 133                                                 unsigned long value)
 134 {
 135         context->lo &= (((u64)-1) << 4) | 3;
 136         context->lo |= (value & 3) << 2;
 137 }
 138
 139 static inline void context_set_address_root(struct context_entry *context,
 140                                             unsigned long value)
 141 {
 142         context->lo |= value & VTD_PAGE_MASK;
 143 }
 144
 145 static inline void context_set_address_width(struct context_entry *context,
 146                                              unsigned long value)
 147 {
 148         context->hi |= value & 7;
 149 }
 150
 151 static inline void context_set_domain_id(struct context_entry *context,
 152                                          unsigned long value)
 153 {
 154         context->hi |= (value & ((1 << 16) - 1)) << 8;
 155 }
 156
 157 static inline void context_clear_entry(struct context_entry *context)
 158 {
 159         context->lo = 0;
 160         context->hi = 0;
 161 }
 162
 163 /*
 164  * 0: readable
 165  * 1: writable
 166  * 2-6: reserved
 167  * 7: super page
 168  * 8-10: available
 169  * 11: snoop behavior
 170  * 12-63: Host physcial address
 171  */
 172 struct dma_pte {
 173         u64 val;
 174 };
 175
 176 static inline void dma_clear_pte(struct dma_pte *pte)
 177 {
 178         pte->val = 0;
 179 }
 180
 181 static inline void dma_set_pte_readable(struct dma_pte *pte)
 182 {
 183         pte->val |= DMA_PTE_READ;
 184 }
 185
 186 static inline void dma_set_pte_writable(struct dma_pte *pte)
 187 {
 188         pte->val |= DMA_PTE_WRITE;
 189 }
 190
 191 static inline void dma_set_pte_snp(struct dma_pte *pte)
 192 {
 193         pte->val |= DMA_PTE_SNP;
 194 }
 195
 196 static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)
 197 {
 198         pte->val = (pte->val & ~3) | (prot & 3);
 199 }
 200
 201 static inline u64 dma_pte_addr(struct dma_pte *pte)
 202 {
 203         return (pte->val & VTD_PAGE_MASK);
 204 }
 205
 206 static inline void dma_set_pte_addr(struct dma_pte *pte, u64 addr)
 207 {
 208         pte->val |= (addr & VTD_PAGE_MASK);
 209 }
 210
 211 static inline bool dma_pte_present(struct dma_pte *pte)
 212 {
 213         return (pte->val & 3) != 0;
 214 }
 215
 216 /* devices under the same p2p bridge are owned in one domain */
 217 #define DOMAIN_FLAG_P2P_MULTIPLE_DEVICES (1 << 0)
 218
 219 /* domain represents a virtual machine, more than one devices
 220  * across iommus may be owned in one domain, e.g. kvm guest.
 221  */
 222 #define DOMAIN_FLAG_VIRTUAL_MACHINE     (1 << 1)
 223
 224 struct dmar_domain {
 225         int     id;                     /* domain id */
 226         unsigned long iommu_bmp;        /* bitmap of iommus this domain uses*/
 227
 228         struct list_head devices;       /* all devices' list */
 229         struct iova_domain iovad;       /* iova's that belong to this domain */
 230
 231         struct dma_pte  *pgd;           /* virtual address */
 232         spinlock_t      mapping_lock;   /* page table lock */
 233         int             gaw;            /* max guest address width */
 234
 235         /* adjusted guest address width, 0 is level 2 30-bit */
 236         int             agaw;
 237
 238         int             flags;          /* flags to find out type of domain */
 239
 240         int             iommu_coherency;/* indicate coherency of iommu access */
 241         int             iommu_snooping; /* indicate snooping control feature*/
 242         int             iommu_count;    /* reference count of iommu */
 243         spinlock_t      iommu_lock;     /* protect iommu set in domain */
 244         u64             max_addr;       /* maximum mapped address */
 245 };
 246
 247 /* PCI domain-device relationship */
 248 struct device_domain_info {
 249         struct list_head link;  /* link to domain siblings */
 250         struct list_head global; /* link to global list */
 251         int segment;            /* PCI domain */
 252         u8 bus;                 /* PCI bus number */
 253         u8 devfn;               /* PCI devfn number */
 254         struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
 255         struct dmar_domain *domain; /* pointer to domain */
 256 };
 257
 258 static void flush_unmaps_timeout(unsigned long data);
 259
 260 DEFINE_TIMER(unmap_timer,  flush_unmaps_timeout, 0, 0);
 261
 262 #define HIGH_WATER_MARK 250
 263 struct deferred_flush_tables {
 264         int next;
 265         struct iova *iova[HIGH_WATER_MARK];
 266         struct dmar_domain *domain[HIGH_WATER_MARK];
 267 };
 268
 269 static struct deferred_flush_tables *deferred_flush;
 270
 271 /* bitmap for indexing intel_iommus */
 272 static int g_num_of_iommus;
 273
 274 static DEFINE_SPINLOCK(async_umap_flush_lock);
 275 static LIST_HEAD(unmaps_to_do);
 276
 277 static int timer_on;
 278 static long list_size;
 279
 280 static void domain_remove_dev_info(struct dmar_domain *domain);
 281
 282 #ifdef CONFIG_DMAR_DEFAULT_ON
 283 int dmar_disabled = 0;
 284 #else
 285 int dmar_disabled = 1;
 286 #endif /*CONFIG_DMAR_DEFAULT_ON*/
 287
 288 static int __initdata dmar_map_gfx = 1;
 289 static int dmar_forcedac;
 290 static int intel_iommu_strict;
 291
 292 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
 293 static DEFINE_SPINLOCK(device_domain_lock);
 294 static LIST_HEAD(device_domain_list);
 295
 296 static struct iommu_ops intel_iommu_ops;
 297
 298 static int __init intel_iommu_setup(char *str)
 299 {
 300         if (!str)
 301                 return -EINVAL;
 302         while (*str) {
 303                 if (!strncmp(str, "on", 2)) {
 304                         dmar_disabled = 0;
 305                         printk(KERN_INFO "Intel-IOMMU: enabled\n");
 306                 } else if (!strncmp(str, "off", 3)) {
 307                         dmar_disabled = 1;
 308                         printk(KERN_INFO "Intel-IOMMU: disabled\n");
 309                 } else if (!strncmp(str, "igfx_off", 8)) {
 310                         dmar_map_gfx = 0;
 311                         printk(KERN_INFO
 312                                 "Intel-IOMMU: disable GFX device mapping\n");
 313                 } else if (!strncmp(str, "forcedac", 8)) {
 314                         printk(KERN_INFO
 315                                 "Intel-IOMMU: Forcing DAC for PCI devices\n");
 316                         dmar_forcedac = 1;
 317                 } else if (!strncmp(str, "strict", 6)) {
 318                         printk(KERN_INFO
 319                                 "Intel-IOMMU: disable batched IOTLB flush\n");
 320                         intel_iommu_strict = 1;
 321                 }
 322
 323                 str += strcspn(str, ",");
 324                 while (*str == ',')
 325                         str++;
 326         }
 327         return 0;
 328 }
 329 __setup("intel_iommu=", intel_iommu_setup);
 330
 331 static struct kmem_cache *iommu_domain_cache;
 332 static struct kmem_cache *iommu_devinfo_cache;
 333 static struct kmem_cache *iommu_iova_cache;
 334
 335 static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
 336 {
 337         unsigned int flags;
 338         void *vaddr;
 339
 340         /* trying to avoid low memory issues */
 341         flags = current->flags & PF_MEMALLOC;
 342         current->flags |= PF_MEMALLOC;
 343         vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
 344         current->flags &= (~PF_MEMALLOC | flags);
 345         return vaddr;
 346 }
 347
 348
 349 static inline void *alloc_pgtable_page(void)
 350 {
 351         unsigned int flags;
 352         void *vaddr;
 353
 354         /* trying to avoid low memory issues */
 355         flags = current->flags & PF_MEMALLOC;
 356         current->flags |= PF_MEMALLOC;
 357         vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
 358         current->flags &= (~PF_MEMALLOC | flags);
 359         return vaddr;
 360 }
 361
 362 static inline void free_pgtable_page(void *vaddr)
 363 {
 364         free_page((unsigned long)vaddr);
 365 }
 366
 367 static inline void *alloc_domain_mem(void)
 368 {
 369         return iommu_kmem_cache_alloc(iommu_domain_cache);
 370 }
 371
 372 static void free_domain_mem(void *vaddr)
 373 {
 374         kmem_cache_free(iommu_domain_cache, vaddr);
 375 }
 376
 377 static inline void * alloc_devinfo_mem(void)
 378 {
 379         return iommu_kmem_cache_alloc(iommu_devinfo_cache);
 380 }
 381
 382 static inline void free_devinfo_mem(void *vaddr)
 383 {
 384         kmem_cache_free(iommu_devinfo_cache, vaddr);
 385 }
 386
 387 struct iova *alloc_iova_mem(void)
 388 {
 389         return iommu_kmem_cache_alloc(iommu_iova_cache);
 390 }
 391
 392 void free_iova_mem(struct iova *iova)
 393 {
 394         kmem_cache_free(iommu_iova_cache, iova);
 395 }
 396
 397
 398 static inline int width_to_agaw(int width);
 399
 400 /* calculate agaw for each iommu.
 401  * "SAGAW" may be different across iommus, use a default agaw, and
 402  * get a supported less agaw for iommus that don't support the default agaw.
 403  */
 404 int iommu_calculate_agaw(struct intel_iommu *iommu)
 405 {
 406         unsigned long sagaw;
 407         int agaw = -1;
 408
 409         sagaw = cap_sagaw(iommu->cap);
 410         for (agaw = width_to_agaw(DEFAULT_DOMAIN_ADDRESS_WIDTH);
 411              agaw >= 0; agaw--) {
 412                 if (test_bit(agaw, &sagaw))
 413                         break;
 414         }
 415
 416         return agaw;
 417 }
 418
 419 /* in native case, each domain is related to only one iommu */
 420 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
 421 {
 422         int iommu_id;
 423
 424         BUG_ON(domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE);
 425
 426         iommu_id = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
 427         if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
 428                 return NULL;
 429
 430         return g_iommus[iommu_id];
 431 }
 432
 433 static void domain_update_iommu_coherency(struct dmar_domain *domain)
 434 {
 435         int i;
 436
 437         domain->iommu_coherency = 1;
 438
 439         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
 440         for (; i < g_num_of_iommus; ) {
 441                 if (!ecap_coherent(g_iommus[i]->ecap)) {
 442                         domain->iommu_coherency = 0;
 443                         break;
 444                 }
 445                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
 446         }
 447 }
 448
 449 static void domain_update_iommu_snooping(struct dmar_domain *domain)
 450 {
 451         int i;
 452
 453         domain->iommu_snooping = 1;
 454
 455         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
 456         for (; i < g_num_of_iommus; ) {
 457                 if (!ecap_sc_support(g_iommus[i]->ecap)) {
 458                         domain->iommu_snooping = 0;
 459                         break;
 460                 }
 461                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
 462         }
 463 }
 464
 465 /* Some capabilities may be different across iommus */
 466 static void domain_update_iommu_cap(struct dmar_domain *domain)
 467 {
 468         domain_update_iommu_coherency(domain);
 469         domain_update_iommu_snooping(domain);
 470 }
 471
 472 static struct intel_iommu *device_to_iommu(int segment, u8 bus, u8 devfn)
 473 {
 474         struct dmar_drhd_unit *drhd = NULL;
 475         int i;
 476
 477         for_each_drhd_unit(drhd) {
 478                 if (drhd->ignored)
 479                         continue;
 480                 if (segment != drhd->segment)
 481                         continue;
 482
 483                 for (i = 0; i < drhd->devices_cnt; i++) {
 484                         if (drhd->devices[i] &&
 485                             drhd->devices[i]->bus->number == bus &&
 486                             drhd->devices[i]->devfn == devfn)
 487                                 return drhd->iommu;
 488                         if (drhd->devices[i] &&
 489                             drhd->devices[i]->subordinate &&
 490                             drhd->devices[i]->subordinate->number <= bus &&
 491                             drhd->devices[i]->subordinate->subordinate >= bus)
 492                                 return drhd->iommu;
 493                 }
 494
 495                 if (drhd->include_all)
 496                         return drhd->iommu;
 497         }
 498
 499         return NULL;
 500 }
 501
 502 static void domain_flush_cache(struct dmar_domain *domain,
 503                                void *addr, int size)
 504 {
 505         if (!domain->iommu_coherency)
 506                 clflush_cache_range(addr, size);
 507 }
 508
 509 /* Gets context entry for a given bus and devfn */
 510 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
 511                 u8 bus, u8 devfn)
 512 {
 513         struct root_entry *root;
 514         struct context_entry *context;
 515         unsigned long phy_addr;
 516         unsigned long flags;
 517
 518         spin_lock_irqsave(&iommu->lock, flags);
 519         root = &iommu->root_entry[bus];
 520         context = get_context_addr_from_root(root);
 521         if (!context) {
 522                 context = (struct context_entry *)alloc_pgtable_page();
 523                 if (!context) {
 524                         spin_unlock_irqrestore(&iommu->lock, flags);
 525                         return NULL;
 526                 }
 527                 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
 528                 phy_addr = virt_to_phys((void *)context);
 529                 set_root_value(root, phy_addr);
 530                 set_root_present(root);
 531                 __iommu_flush_cache(iommu, root, sizeof(*root));
 532         }
 533         spin_unlock_irqrestore(&iommu->lock, flags);
 534         return &context[devfn];
 535 }
 536
 537 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
 538 {
 539         struct root_entry *root;
 540         struct context_entry *context;
 541         int ret;
 542         unsigned long flags;
 543
 544         spin_lock_irqsave(&iommu->lock, flags);
 545         root = &iommu->root_entry[bus];
 546         context = get_context_addr_from_root(root);
 547         if (!context) {
 548                 ret = 0;
 549                 goto out;
 550         }
 551         ret = context_present(&context[devfn]);
 552 out:
 553         spin_unlock_irqrestore(&iommu->lock, flags);
 554         return ret;
 555 }
 556
 557 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
 558 {
 559         struct root_entry *root;
 560         struct context_entry *context;
 561         unsigned long flags;
 562
 563         spin_lock_irqsave(&iommu->lock, flags);
 564         root = &iommu->root_entry[bus];
 565         context = get_context_addr_from_root(root);
 566         if (context) {
 567                 context_clear_entry(&context[devfn]);
 568                 __iommu_flush_cache(iommu, &context[devfn], \
 569                         sizeof(*context));
 570         }
 571         spin_unlock_irqrestore(&iommu->lock, flags);
 572 }
 573
 574 static void free_context_table(struct intel_iommu *iommu)
 575 {
 576         struct root_entry *root;
 577         int i;
 578         unsigned long flags;
 579         struct context_entry *context;
 580
 581         spin_lock_irqsave(&iommu->lock, flags);
 582         if (!iommu->root_entry) {
 583                 goto out;
 584         }
 585         for (i = 0; i < ROOT_ENTRY_NR; i++) {
 586                 root = &iommu->root_entry[i];
 587                 context = get_context_addr_from_root(root);
 588                 if (context)
 589                         free_pgtable_page(context);
 590         }
 591         free_pgtable_page(iommu->root_entry);
 592         iommu->root_entry = NULL;
 593 out:
 594         spin_unlock_irqrestore(&iommu->lock, flags);
 595 }
 596
 597 /* page table handling */
 598 #define LEVEL_STRIDE            (9)
 599 #define LEVEL_MASK              (((u64)1 << LEVEL_STRIDE) - 1)
 600
 601 static inline int agaw_to_level(int agaw)
 602 {
 603         return agaw + 2;
 604 }
 605
 606 static inline int agaw_to_width(int agaw)
 607 {
 608         return 30 + agaw * LEVEL_STRIDE;
 609
 610 }
 611
 612 static inline int width_to_agaw(int width)
 613 {
 614         return (width - 30) / LEVEL_STRIDE;
 615 }
 616
 617 static inline unsigned int level_to_offset_bits(int level)
 618 {
 619         return (12 + (level - 1) * LEVEL_STRIDE);
 620 }
 621
 622 static inline int address_level_offset(u64 addr, int level)
 623 {
 624         return ((addr >> level_to_offset_bits(level)) & LEVEL_MASK);
 625 }
 626
 627 static inline u64 level_mask(int level)
 628 {
 629         return ((u64)-1 << level_to_offset_bits(level));
 630 }
 631
 632 static inline u64 level_size(int level)
 633 {
 634         return ((u64)1 << level_to_offset_bits(level));
 635 }
 636
 637 static inline u64 align_to_level(u64 addr, int level)
 638 {
 639         return ((addr + level_size(level) - 1) & level_mask(level));
 640 }
 641
 642 static struct dma_pte * addr_to_dma_pte(struct dmar_domain *domain, u64 addr)
 643 {
 644         int addr_width = agaw_to_width(domain->agaw);
 645         struct dma_pte *parent, *pte = NULL;
 646         int level = agaw_to_level(domain->agaw);
 647         int offset;
 648         unsigned long flags;
 649
 650         BUG_ON(!domain->pgd);
 651
 652         addr &= (((u64)1) << addr_width) - 1;
 653         parent = domain->pgd;
 654
 655         spin_lock_irqsave(&domain->mapping_lock, flags);
 656         while (level > 0) {
 657                 void *tmp_page;
 658
 659                 offset = address_level_offset(addr, level);
 660                 pte = &parent[offset];
 661                 if (level == 1)
 662                         break;
 663
 664                 if (!dma_pte_present(pte)) {
 665                         tmp_page = alloc_pgtable_page();
 666
 667                         if (!tmp_page) {
 668                                 spin_unlock_irqrestore(&domain->mapping_lock,
 669                                         flags);
 670                                 return NULL;
 671                         }
 672                         domain_flush_cache(domain, tmp_page, PAGE_SIZE);
 673                         dma_set_pte_addr(pte, virt_to_phys(tmp_page));
 674                         /*
 675                          * high level table always sets r/w, last level page
 676                          * table control read/write
 677                          */
 678                         dma_set_pte_readable(pte);
 679                         dma_set_pte_writable(pte);
 680                         domain_flush_cache(domain, pte, sizeof(*pte));
 681                 }
 682                 parent = phys_to_virt(dma_pte_addr(pte));
 683                 level--;
 684         }
 685
 686         spin_unlock_irqrestore(&domain->mapping_lock, flags);
 687         return pte;
 688 }
 689
 690 /* return address's pte at specific level */
 691 static struct dma_pte *dma_addr_level_pte(struct dmar_domain *domain, u64 addr,
 692                 int level)
 693 {
 694         struct dma_pte *parent, *pte = NULL;
 695         int total = agaw_to_level(domain->agaw);
 696         int offset;
 697
 698         parent = domain->pgd;
 699         while (level <= total) {
 700                 offset = address_level_offset(addr, total);
 701                 pte = &parent[offset];
 702                 if (level == total)
 703                         return pte;
 704
 705                 if (!dma_pte_present(pte))
 706                         break;
 707                 parent = phys_to_virt(dma_pte_addr(pte));
 708                 total--;
 709         }
 710         return NULL;
 711 }
 712
 713 /* clear one page's page table */
 714 static void dma_pte_clear_one(struct dmar_domain *domain, u64 addr)
 715 {
 716         struct dma_pte *pte = NULL;
 717
 718         /* get last level pte */
 719         pte = dma_addr_level_pte(domain, addr, 1);
 720
 721         if (pte) {
 722                 dma_clear_pte(pte);
 723                 domain_flush_cache(domain, pte, sizeof(*pte));
 724         }
 725 }
 726
 727 /* clear last level pte, a tlb flush should be followed */
 728 static void dma_pte_clear_range(struct dmar_domain *domain, u64 start, u64 end)
 729 {
 730         int addr_width = agaw_to_width(domain->agaw);
 731         int npages;
 732
 733         start &= (((u64)1) << addr_width) - 1;
 734         end &= (((u64)1) << addr_width) - 1;
 735         /* in case it's partial page */
 736         start &= PAGE_MASK;
 737         end = PAGE_ALIGN(end);
 738         npages = (end - start) / VTD_PAGE_SIZE;
 739
 740         /* we don't need lock here, nobody else touches the iova range */
 741         while (npages--) {
 742                 dma_pte_clear_one(domain, start);
 743                 start += VTD_PAGE_SIZE;
 744         }
 745 }
 746
 747 /* free page table pages. last level pte should already be cleared */
 748 static void dma_pte_free_pagetable(struct dmar_domain *domain,
 749         u64 start, u64 end)
 750 {
 751         int addr_width = agaw_to_width(domain->agaw);
 752         struct dma_pte *pte;
 753         int total = agaw_to_level(domain->agaw);
 754         int level;
 755         u64 tmp;
 756
 757         start &= (((u64)1) << addr_width) - 1;
 758         end &= (((u64)1) << addr_width) - 1;
 759
 760         /* we don't need lock here, nobody else touches the iova range */
 761         level = 2;
 762         while (level <= total) {
 763                 tmp = align_to_level(start, level);
 764                 if (tmp >= end || (tmp + level_size(level) > end))
 765                         return;
 766
 767                 while (tmp < end) {
 768                         pte = dma_addr_level_pte(domain, tmp, level);
 769                         if (pte) {
 770                                 free_pgtable_page(
 771                                         phys_to_virt(dma_pte_addr(pte)));
 772                                 dma_clear_pte(pte);
 773                                 domain_flush_cache(domain, pte, sizeof(*pte));
 774                         }
 775                         tmp += level_size(level);
 776                 }
 777                 level++;
 778         }
 779         /* free pgd */
 780         if (start == 0 && end >= ((((u64)1) << addr_width) - 1)) {
 781                 free_pgtable_page(domain->pgd);
 782                 domain->pgd = NULL;
 783         }
 784 }
 785
 786 /* iommu handling */
 787 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
 788 {
 789         struct root_entry *root;
 790         unsigned long flags;
 791
 792         root = (struct root_entry *)alloc_pgtable_page();
 793         if (!root)
 794                 return -ENOMEM;
 795
 796         __iommu_flush_cache(iommu, root, ROOT_SIZE);
 797
 798         spin_lock_irqsave(&iommu->lock, flags);
 799         iommu->root_entry = root;
 800         spin_unlock_irqrestore(&iommu->lock, flags);
 801
 802         return 0;
 803 }
 804
 805 static void iommu_set_root_entry(struct intel_iommu *iommu)
 806 {
 807         void *addr;
 808         u32 cmd, sts;
 809         unsigned long flag;
 810
 811         addr = iommu->root_entry;
 812
 813         spin_lock_irqsave(&iommu->register_lock, flag);
 814         dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
 815
 816         cmd = iommu->gcmd | DMA_GCMD_SRTP;
 817         writel(cmd, iommu->reg + DMAR_GCMD_REG);
 818
 819         /* Make sure hardware complete it */
 820         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
 821                 readl, (sts & DMA_GSTS_RTPS), sts);
 822
 823         spin_unlock_irqrestore(&iommu->register_lock, flag);
 824 }
 825
 826 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
 827 {
 828         u32 val;
 829         unsigned long flag;
 830
 831         if (!rwbf_quirk && !cap_rwbf(iommu->cap))
 832                 return;
 833         val = iommu->gcmd | DMA_GCMD_WBF;
 834
 835         spin_lock_irqsave(&iommu->register_lock, flag);
 836         writel(val, iommu->reg + DMAR_GCMD_REG);
 837
 838         /* Make sure hardware complete it */
 839         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
 840                         readl, (!(val & DMA_GSTS_WBFS)), val);
 841
 842         spin_unlock_irqrestore(&iommu->register_lock, flag);
 843 }
 844
 845 /* return value determine if we need a write buffer flush */
 846 static int __iommu_flush_context(struct intel_iommu *iommu,
 847         u16 did, u16 source_id, u8 function_mask, u64 type,
 848         int non_present_entry_flush)
 849 {
 850         u64 val = 0;
 851         unsigned long flag;
 852
 853         /*
 854          * In the non-present entry flush case, if hardware doesn't cache
 855          * non-present entry we do nothing and if hardware cache non-present
 856          * entry, we flush entries of domain 0 (the domain id is used to cache
 857          * any non-present entries)
 858          */
 859         if (non_present_entry_flush) {
 860                 if (!cap_caching_mode(iommu->cap))
 861                         return 1;
 862                 else
 863                         did = 0;
 864         }
 865
 866         switch (type) {
 867         case DMA_CCMD_GLOBAL_INVL:
 868                 val = DMA_CCMD_GLOBAL_INVL;
 869                 break;
 870         case DMA_CCMD_DOMAIN_INVL:
 871                 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
 872                 break;
 873         case DMA_CCMD_DEVICE_INVL:
 874                 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
 875                         | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
 876                 break;
 877         default:
 878                 BUG();
 879         }
 880         val |= DMA_CCMD_ICC;
 881
 882         spin_lock_irqsave(&iommu->register_lock, flag);
 883         dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
 884
 885         /* Make sure hardware complete it */
 886         IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
 887                 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
 888
 889         spin_unlock_irqrestore(&iommu->register_lock, flag);
 890
 891         /* flush context entry will implicitly flush write buffer */
 892         return 0;
 893 }
 894
 895 /* return value determine if we need a write buffer flush */
 896 static int __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
 897         u64 addr, unsigned int size_order, u64 type,
 898         int non_present_entry_flush)
 899 {
 900         int tlb_offset = ecap_iotlb_offset(iommu->ecap);
 901         u64 val = 0, val_iva = 0;
 902         unsigned long flag;
 903
 904         /*
 905          * In the non-present entry flush case, if hardware doesn't cache
 906          * non-present entry we do nothing and if hardware cache non-present
 907          * entry, we flush entries of domain 0 (the domain id is used to cache
 908          * any non-present entries)
 909          */
 910         if (non_present_entry_flush) {
 911                 if (!cap_caching_mode(iommu->cap))
 912                         return 1;
 913                 else
 914                         did = 0;
 915         }
 916
 917         switch (type) {
 918         case DMA_TLB_GLOBAL_FLUSH:
 919                 /* global flush doesn't need set IVA_REG */
 920                 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
 921                 break;
 922         case DMA_TLB_DSI_FLUSH:
 923                 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
 924                 break;
 925         case DMA_TLB_PSI_FLUSH:
 926                 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
 927                 /* Note: always flush non-leaf currently */
 928                 val_iva = size_order | addr;
 929                 break;
 930         default:
 931                 BUG();
 932         }
 933         /* Note: set drain read/write */
 934 #if 0
 935         /*
 936          * This is probably to be super secure.. Looks like we can
 937          * ignore it without any impact.
 938          */
 939         if (cap_read_drain(iommu->cap))
 940                 val |= DMA_TLB_READ_DRAIN;
 941 #endif
 942         if (cap_write_drain(iommu->cap))
 943                 val |= DMA_TLB_WRITE_DRAIN;
 944
 945         spin_lock_irqsave(&iommu->register_lock, flag);
 946         /* Note: Only uses first TLB reg currently */
 947         if (val_iva)
 948                 dmar_writeq(iommu->reg + tlb_offset, val_iva);
 949         dmar_writeq(iommu->reg + tlb_offset + 8, val);
 950
 951         /* Make sure hardware complete it */
 952         IOMMU_WAIT_OP(iommu, tlb_offset + 8,
 953                 dmar_readq, (!(val & DMA_TLB_IVT)), val);
 954
 955         spin_unlock_irqrestore(&iommu->register_lock, flag);
 956
 957         /* check IOTLB invalidation granularity */
 958         if (DMA_TLB_IAIG(val) == 0)
 959                 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
 960         if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
 961                 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
 962                         (unsigned long long)DMA_TLB_IIRG(type),
 963                         (unsigned long long)DMA_TLB_IAIG(val));
 964         /* flush iotlb entry will implicitly flush write buffer */
 965         return 0;
 966 }
 967
 968 static int iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
 969         u64 addr, unsigned int pages, int non_present_entry_flush)
 970 {
 971         unsigned int mask;
 972
 973         BUG_ON(addr & (~VTD_PAGE_MASK));
 974         BUG_ON(pages == 0);
 975
 976         /* Fallback to domain selective flush if no PSI support */
 977         if (!cap_pgsel_inv(iommu->cap))
 978                 return iommu->flush.flush_iotlb(iommu, did, 0, 0,
 979                                                 DMA_TLB_DSI_FLUSH,
 980                                                 non_present_entry_flush);
 981
 982         /*
 983          * PSI requires page size to be 2 ^ x, and the base address is naturally
 984          * aligned to the size
 985          */
 986         mask = ilog2(__roundup_pow_of_two(pages));
 987         /* Fallback to domain selective flush if size is too big */
 988         if (mask > cap_max_amask_val(iommu->cap))
 989                 return iommu->flush.flush_iotlb(iommu, did, 0, 0,
 990                         DMA_TLB_DSI_FLUSH, non_present_entry_flush);
 991
 992         return iommu->flush.flush_iotlb(iommu, did, addr, mask,
 993                                         DMA_TLB_PSI_FLUSH,
 994                                         non_present_entry_flush);
 995 }
 996
 997 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
 998 {
 999         u32 pmen;
1000         unsigned long flags;
1001
1002         spin_lock_irqsave(&iommu->register_lock, flags);
1003         pmen = readl(iommu->reg + DMAR_PMEN_REG);
1004         pmen &= ~DMA_PMEN_EPM;
1005         writel(pmen, iommu->reg + DMAR_PMEN_REG);
1006
1007         /* wait for the protected region status bit to clear */
1008         IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1009                 readl, !(pmen & DMA_PMEN_PRS), pmen);
1010
1011         spin_unlock_irqrestore(&iommu->register_lock, flags);
1012 }
1013
1014 static int iommu_enable_translation(struct intel_iommu *iommu)
1015 {
1016         u32 sts;
1017         unsigned long flags;
1018
1019         spin_lock_irqsave(&iommu->register_lock, flags);
1020         writel(iommu->gcmd|DMA_GCMD_TE, iommu->reg + DMAR_GCMD_REG);
1021
1022         /* Make sure hardware complete it */
1023         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1024                 readl, (sts & DMA_GSTS_TES), sts);
1025
1026         iommu->gcmd |= DMA_GCMD_TE;
1027         spin_unlock_irqrestore(&iommu->register_lock, flags);
1028         return 0;
1029 }
1030
1031 static int iommu_disable_translation(struct intel_iommu *iommu)
1032 {
1033         u32 sts;
1034         unsigned long flag;
1035
1036         spin_lock_irqsave(&iommu->register_lock, flag);
1037         iommu->gcmd &= ~DMA_GCMD_TE;
1038         writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1039
1040         /* Make sure hardware complete it */
1041         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1042                 readl, (!(sts & DMA_GSTS_TES)), sts);
1043
1044         spin_unlock_irqrestore(&iommu->register_lock, flag);
1045         return 0;
1046 }
1047
1048
1049 static int iommu_init_domains(struct intel_iommu *iommu)
1050 {
1051         unsigned long ndomains;
1052         unsigned long nlongs;
1053
1054         ndomains = cap_ndoms(iommu->cap);
1055         pr_debug("Number of Domains supportd <%ld>\n", ndomains);
1056         nlongs = BITS_TO_LONGS(ndomains);
1057
1058         /* TBD: there might be 64K domains,
1059          * consider other allocation for future chip
1060          */
1061         iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1062         if (!iommu->domain_ids) {
1063                 printk(KERN_ERR "Allocating domain id array failed\n");
1064                 return -ENOMEM;
1065         }
1066         iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
1067                         GFP_KERNEL);
1068         if (!iommu->domains) {
1069                 printk(KERN_ERR "Allocating domain array failed\n");
1070                 kfree(iommu->domain_ids);
1071                 return -ENOMEM;
1072         }
1073
1074         spin_lock_init(&iommu->lock);
1075
1076         /*
1077          * if Caching mode is set, then invalid translations are tagged
1078          * with domainid 0. Hence we need to pre-allocate it.
1079          */
1080         if (cap_caching_mode(iommu->cap))
1081                 set_bit(0, iommu->domain_ids);
1082         return 0;
1083 }
1084
1085
1086 static void domain_exit(struct dmar_domain *domain);
1087 static void vm_domain_exit(struct dmar_domain *domain);
1088
1089 void free_dmar_iommu(struct intel_iommu *iommu)
1090 {
1091         struct dmar_domain *domain;
1092         int i;
1093         unsigned long flags;
1094
1095         i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1096         for (; i < cap_ndoms(iommu->cap); ) {
1097                 domain = iommu->domains[i];
1098                 clear_bit(i, iommu->domain_ids);
1099
1100                 spin_lock_irqsave(&domain->iommu_lock, flags);
1101                 if (--domain->iommu_count == 0) {
1102                         if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
1103                                 vm_domain_exit(domain);
1104                         else
1105                                 domain_exit(domain);
1106                 }
1107                 spin_unlock_irqrestore(&domain->iommu_lock, flags);
1108
1109                 i = find_next_bit(iommu->domain_ids,
1110                         cap_ndoms(iommu->cap), i+1);
1111         }
1112
1113         if (iommu->gcmd & DMA_GCMD_TE)
1114                 iommu_disable_translation(iommu);
1115
1116         if (iommu->irq) {
1117                 set_irq_data(iommu->irq, NULL);
1118                 /* This will mask the irq */
1119                 free_irq(iommu->irq, iommu);
1120                 destroy_irq(iommu->irq);
1121         }
1122
1123         kfree(iommu->domains);
1124         kfree(iommu->domain_ids);
1125
1126         g_iommus[iommu->seq_id] = NULL;
1127
1128         /* if all iommus are freed, free g_iommus */
1129         for (i = 0; i < g_num_of_iommus; i++) {
1130                 if (g_iommus[i])
1131                         break;
1132         }
1133
1134         if (i == g_num_of_iommus)
1135                 kfree(g_iommus);
1136
1137         /* free context mapping */
1138         free_context_table(iommu);
1139 }
1140
1141 static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
1142 {
1143         unsigned long num;
1144         unsigned long ndomains;
1145         struct dmar_domain *domain;
1146         unsigned long flags;
1147
1148         domain = alloc_domain_mem();
1149         if (!domain)
1150                 return NULL;
1151
1152         ndomains = cap_ndoms(iommu->cap);
1153
1154         spin_lock_irqsave(&iommu->lock, flags);
1155         num = find_first_zero_bit(iommu->domain_ids, ndomains);
1156         if (num >= ndomains) {
1157                 spin_unlock_irqrestore(&iommu->lock, flags);
1158                 free_domain_mem(domain);
1159                 printk(KERN_ERR "IOMMU: no free domain ids\n");
1160                 return NULL;
1161         }
1162
1163         set_bit(num, iommu->domain_ids);
1164         domain->id = num;
1165         memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
1166         set_bit(iommu->seq_id, &domain->iommu_bmp);
1167         domain->flags = 0;
1168         iommu->domains[num] = domain;
1169         spin_unlock_irqrestore(&iommu->lock, flags);
1170
1171         return domain;
1172 }
1173
1174 static void iommu_free_domain(struct dmar_domain *domain)
1175 {
1176         unsigned long flags;
1177         struct intel_iommu *iommu;
1178
1179         iommu = domain_get_iommu(domain);
1180
1181         spin_lock_irqsave(&iommu->lock, flags);
1182         clear_bit(domain->id, iommu->domain_ids);
1183         spin_unlock_irqrestore(&iommu->lock, flags);
1184 }
1185
1186 static struct iova_domain reserved_iova_list;
1187 static struct lock_class_key reserved_alloc_key;
1188 static struct lock_class_key reserved_rbtree_key;
1189
1190 static void dmar_init_reserved_ranges(void)
1191 {
1192         struct pci_dev *pdev = NULL;
1193         struct iova *iova;
1194         int i;
1195         u64 addr, size;
1196
1197         init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1198
1199         lockdep_set_class(&reserved_iova_list.iova_alloc_lock,
1200                 &reserved_alloc_key);
1201         lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1202                 &reserved_rbtree_key);
1203
1204         /* IOAPIC ranges shouldn't be accessed by DMA */
1205         iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1206                 IOVA_PFN(IOAPIC_RANGE_END));
1207         if (!iova)
1208                 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1209
1210         /* Reserve all PCI MMIO to avoid peer-to-peer access */
1211         for_each_pci_dev(pdev) {
1212                 struct resource *r;
1213
1214                 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1215                         r = &pdev->resource[i];
1216                         if (!r->flags || !(r->flags & IORESOURCE_MEM))
1217                                 continue;
1218                         addr = r->start;
1219                         addr &= PAGE_MASK;
1220                         size = r->end - addr;
1221                         size = PAGE_ALIGN(size);
1222                         iova = reserve_iova(&reserved_iova_list, IOVA_PFN(addr),
1223                                 IOVA_PFN(size + addr) - 1);
1224                         if (!iova)
1225                                 printk(KERN_ERR "Reserve iova failed\n");
1226                 }
1227         }
1228
1229 }
1230
1231 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1232 {
1233         copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1234 }
1235
1236 static inline int guestwidth_to_adjustwidth(int gaw)
1237 {
1238         int agaw;
1239         int r = (gaw - 12) % 9;
1240
1241         if (r == 0)
1242                 agaw = gaw;
1243         else
1244                 agaw = gaw + 9 - r;
1245         if (agaw > 64)
1246                 agaw = 64;
1247         return agaw;
1248 }
1249
1250 static int domain_init(struct dmar_domain *domain, int guest_width)
1251 {
1252         struct intel_iommu *iommu;
1253         int adjust_width, agaw;
1254         unsigned long sagaw;
1255
1256         init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1257         spin_lock_init(&domain->mapping_lock);
1258         spin_lock_init(&domain->iommu_lock);
1259
1260         domain_reserve_special_ranges(domain);
1261
1262         /* calculate AGAW */
1263         iommu = domain_get_iommu(domain);
1264         if (guest_width > cap_mgaw(iommu->cap))
1265                 guest_width = cap_mgaw(iommu->cap);
1266         domain->gaw = guest_width;
1267         adjust_width = guestwidth_to_adjustwidth(guest_width);
1268         agaw = width_to_agaw(adjust_width);
1269         sagaw = cap_sagaw(iommu->cap);
1270         if (!test_bit(agaw, &sagaw)) {
1271                 /* hardware doesn't support it, choose a bigger one */
1272                 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1273                 agaw = find_next_bit(&sagaw, 5, agaw);
1274                 if (agaw >= 5)
1275                         return -ENODEV;
1276         }
1277         domain->agaw = agaw;
1278         INIT_LIST_HEAD(&domain->devices);
1279
1280         if (ecap_coherent(iommu->ecap))
1281                 domain->iommu_coherency = 1;
1282         else
1283                 domain->iommu_coherency = 0;
1284
1285         if (ecap_sc_support(iommu->ecap))
1286                 domain->iommu_snooping = 1;
1287         else
1288                 domain->iommu_snooping = 0;
1289
1290         domain->iommu_count = 1;
1291
1292         /* always allocate the top pgd */
1293         domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1294         if (!domain->pgd)
1295                 return -ENOMEM;
1296         __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1297         return 0;
1298 }
1299
1300 static void domain_exit(struct dmar_domain *domain)
1301 {
1302         u64 end;
1303
1304         /* Domain 0 is reserved, so dont process it */
1305         if (!domain)
1306                 return;
1307
1308         domain_remove_dev_info(domain);
1309         /* destroy iovas */
1310         put_iova_domain(&domain->iovad);
1311         end = DOMAIN_MAX_ADDR(domain->gaw);
1312         end = end & (~PAGE_MASK);
1313
1314         /* clear ptes */
1315         dma_pte_clear_range(domain, 0, end);
1316
1317         /* free page tables */
1318         dma_pte_free_pagetable(domain, 0, end);
1319
1320         iommu_free_domain(domain);
1321         free_domain_mem(domain);
1322 }
1323
1324 static int domain_context_mapping_one(struct dmar_domain *domain,
1325                                       int segment, u8 bus, u8 devfn)
1326 {
1327         struct context_entry *context;
1328         unsigned long flags;
1329         struct intel_iommu *iommu;
1330         struct dma_pte *pgd;
1331         unsigned long num;
1332         unsigned long ndomains;
1333         int id;
1334         int agaw;
1335
1336         pr_debug("Set context mapping for %02x:%02x.%d\n",
1337                 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1338         BUG_ON(!domain->pgd);
1339
1340         iommu = device_to_iommu(segment, bus, devfn);
1341         if (!iommu)
1342                 return -ENODEV;
1343
1344         context = device_to_context_entry(iommu, bus, devfn);
1345         if (!context)
1346                 return -ENOMEM;
1347         spin_lock_irqsave(&iommu->lock, flags);
1348         if (context_present(context)) {
1349                 spin_unlock_irqrestore(&iommu->lock, flags);
1350                 return 0;
1351         }
1352
1353         id = domain->id;
1354         pgd = domain->pgd;
1355
1356         if (domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE) {
1357                 int found = 0;
1358
1359                 /* find an available domain id for this device in iommu */
1360                 ndomains = cap_ndoms(iommu->cap);
1361                 num = find_first_bit(iommu->domain_ids, ndomains);
1362                 for (; num < ndomains; ) {
1363                         if (iommu->domains[num] == domain) {
1364                                 id = num;
1365                                 found = 1;
1366                                 break;
1367                         }
1368                         num = find_next_bit(iommu->domain_ids,
1369                                             cap_ndoms(iommu->cap), num+1);
1370                 }
1371
1372                 if (found == 0) {
1373                         num = find_first_zero_bit(iommu->domain_ids, ndomains);
1374                         if (num >= ndomains) {
1375                                 spin_unlock_irqrestore(&iommu->lock, flags);
1376                                 printk(KERN_ERR "IOMMU: no free domain ids\n");
1377                                 return -EFAULT;
1378                         }
1379
1380                         set_bit(num, iommu->domain_ids);
1381                         iommu->domains[num] = domain;
1382                         id = num;
1383                 }
1384
1385                 /* Skip top levels of page tables for
1386                  * iommu which has less agaw than default.
1387                  */
1388                 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1389                         pgd = phys_to_virt(dma_pte_addr(pgd));
1390                         if (!dma_pte_present(pgd)) {
1391                                 spin_unlock_irqrestore(&iommu->lock, flags);
1392                                 return -ENOMEM;
1393                         }
1394                 }
1395         }
1396
1397         context_set_domain_id(context, id);
1398         context_set_address_width(context, iommu->agaw);
1399         context_set_address_root(context, virt_to_phys(pgd));
1400         context_set_translation_type(context, CONTEXT_TT_MULTI_LEVEL);
1401         context_set_fault_enable(context);
1402         context_set_present(context);
1403         domain_flush_cache(domain, context, sizeof(*context));
1404
1405         /* it's a non-present to present mapping */
1406         if (iommu->flush.flush_context(iommu, domain->id,
1407                 (((u16)bus) << 8) | devfn, DMA_CCMD_MASK_NOBIT,
1408                 DMA_CCMD_DEVICE_INVL, 1))
1409                 iommu_flush_write_buffer(iommu);
1410         else
1411                 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH, 0);
1412
1413         spin_unlock_irqrestore(&iommu->lock, flags);
1414
1415         spin_lock_irqsave(&domain->iommu_lock, flags);
1416         if (!test_and_set_bit(iommu->seq_id, &domain->iommu_bmp)) {
1417                 domain->iommu_count++;
1418                 domain_update_iommu_cap(domain);
1419         }
1420         spin_unlock_irqrestore(&domain->iommu_lock, flags);
1421         return 0;
1422 }
1423
1424 static int
1425 domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev)
1426 {
1427         int ret;
1428         struct pci_dev *tmp, *parent;
1429
1430         ret = domain_context_mapping_one(domain, pci_domain_nr(pdev->bus),
1431                                          pdev->bus->number, pdev->devfn);
1432         if (ret)
1433                 return ret;
1434
1435         /* dependent device mapping */
1436         tmp = pci_find_upstream_pcie_bridge(pdev);
1437         if (!tmp)
1438                 return 0;
1439         /* Secondary interface's bus number and devfn 0 */
1440         parent = pdev->bus->self;
1441         while (parent != tmp) {
1442                 ret = domain_context_mapping_one(domain,
1443                                                  pci_domain_nr(parent->bus),
1444                                                  parent->bus->number,
1445                                                  parent->devfn);
1446                 if (ret)
1447                         return ret;
1448                 parent = parent->bus->self;
1449         }
1450         if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1451                 return domain_context_mapping_one(domain,
1452                                         pci_domain_nr(tmp->subordinate),
1453                                         tmp->subordinate->number, 0);
1454         else /* this is a legacy PCI bridge */
1455                 return domain_context_mapping_one(domain,
1456                                                   pci_domain_nr(tmp->bus),
1457                                                   tmp->bus->number,
1458                                                   tmp->devfn);
1459 }
1460
1461 static int domain_context_mapped(struct pci_dev *pdev)
1462 {
1463         int ret;
1464         struct pci_dev *tmp, *parent;
1465         struct intel_iommu *iommu;
1466
1467         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
1468                                 pdev->devfn);
1469         if (!iommu)
1470                 return -ENODEV;
1471
1472         ret = device_context_mapped(iommu, pdev->bus->number, pdev->devfn);
1473         if (!ret)
1474                 return ret;
1475         /* dependent device mapping */
1476         tmp = pci_find_upstream_pcie_bridge(pdev);
1477         if (!tmp)
1478                 return ret;
1479         /* Secondary interface's bus number and devfn 0 */
1480         parent = pdev->bus->self;
1481         while (parent != tmp) {
1482                 ret = device_context_mapped(iommu, parent->bus->number,
1483                                             parent->devfn);
1484                 if (!ret)
1485                         return ret;
1486                 parent = parent->bus->self;
1487         }
1488         if (tmp->is_pcie)
1489                 return device_context_mapped(iommu, tmp->subordinate->number,
1490                                              0);
1491         else
1492                 return device_context_mapped(iommu, tmp->bus->number,
1493                                              tmp->devfn);
1494 }
1495
1496 static int
1497 domain_page_mapping(struct dmar_domain *domain, dma_addr_t iova,
1498                         u64 hpa, size_t size, int prot)
1499 {
1500         u64 start_pfn, end_pfn;
1501         struct dma_pte *pte;
1502         int index;
1503         int addr_width = agaw_to_width(domain->agaw);
1504
1505         hpa &= (((u64)1) << addr_width) - 1;
1506
1507         if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1508                 return -EINVAL;
1509         iova &= PAGE_MASK;
1510         start_pfn = ((u64)hpa) >> VTD_PAGE_SHIFT;
1511         end_pfn = (VTD_PAGE_ALIGN(((u64)hpa) + size)) >> VTD_PAGE_SHIFT;
1512         index = 0;
1513         while (start_pfn < end_pfn) {
1514                 pte = addr_to_dma_pte(domain, iova + VTD_PAGE_SIZE * index);
1515                 if (!pte)
1516                         return -ENOMEM;
1517                 /* We don't need lock here, nobody else
1518                  * touches the iova range
1519                  */
1520                 BUG_ON(dma_pte_addr(pte));
1521                 dma_set_pte_addr(pte, start_pfn << VTD_PAGE_SHIFT);
1522                 dma_set_pte_prot(pte, prot);
1523                 if (prot & DMA_PTE_SNP)
1524                         dma_set_pte_snp(pte);
1525                 domain_flush_cache(domain, pte, sizeof(*pte));
1526                 start_pfn++;
1527                 index++;
1528         }
1529         return 0;
1530 }
1531
1532 static void iommu_detach_dev(struct intel_iommu *iommu, u8 bus, u8 devfn)
1533 {
1534         if (!iommu)
1535                 return;
1536
1537         clear_context_table(iommu, bus, devfn);
1538         iommu->flush.flush_context(iommu, 0, 0, 0,
1539                                            DMA_CCMD_GLOBAL_INVL, 0);
1540         iommu->flush.flush_iotlb(iommu, 0, 0, 0,
1541                                          DMA_TLB_GLOBAL_FLUSH, 0);
1542 }
1543
1544 static void domain_remove_dev_info(struct dmar_domain *domain)
1545 {
1546         struct device_domain_info *info;
1547         unsigned long flags;
1548         struct intel_iommu *iommu;
1549
1550         spin_lock_irqsave(&device_domain_lock, flags);
1551         while (!list_empty(&domain->devices)) {
1552                 info = list_entry(domain->devices.next,
1553                         struct device_domain_info, link);
1554                 list_del(&info->link);
1555                 list_del(&info->global);
1556                 if (info->dev)
1557                         info->dev->dev.archdata.iommu = NULL;
1558                 spin_unlock_irqrestore(&device_domain_lock, flags);
1559
1560                 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
1561                 iommu_detach_dev(iommu, info->bus, info->devfn);
1562                 free_devinfo_mem(info);
1563
1564                 spin_lock_irqsave(&device_domain_lock, flags);
1565         }
1566         spin_unlock_irqrestore(&device_domain_lock, flags);
1567 }
1568
1569 /*
1570  * find_domain
1571  * Note: we use struct pci_dev->dev.archdata.iommu stores the info
1572  */
1573 static struct dmar_domain *
1574 find_domain(struct pci_dev *pdev)
1575 {
1576         struct device_domain_info *info;
1577
1578         /* No lock here, assumes no domain exit in normal case */
1579         info = pdev->dev.archdata.iommu;
1580         if (info)
1581                 return info->domain;
1582         return NULL;
1583 }
1584
1585 /* domain is initialized */
1586 static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1587 {
1588         struct dmar_domain *domain, *found = NULL;
1589         struct intel_iommu *iommu;
1590         struct dmar_drhd_unit *drhd;
1591         struct device_domain_info *info, *tmp;
1592         struct pci_dev *dev_tmp;
1593         unsigned long flags;
1594         int bus = 0, devfn = 0;
1595         int segment;
1596
1597         domain = find_domain(pdev);
1598         if (domain)
1599                 return domain;
1600
1601         segment = pci_domain_nr(pdev->bus);
1602
1603         dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1604         if (dev_tmp) {
1605                 if (dev_tmp->is_pcie) {
1606                         bus = dev_tmp->subordinate->number;
1607                         devfn = 0;
1608                 } else {
1609                         bus = dev_tmp->bus->number;
1610                         devfn = dev_tmp->devfn;
1611                 }
1612                 spin_lock_irqsave(&device_domain_lock, flags);
1613                 list_for_each_entry(info, &device_domain_list, global) {
1614                         if (info->segment == segment &&
1615                             info->bus == bus && info->devfn == devfn) {
1616                                 found = info->domain;
1617                                 break;
1618                         }
1619                 }
1620                 spin_unlock_irqrestore(&device_domain_lock, flags);
1621                 /* pcie-pci bridge already has a domain, uses it */
1622                 if (found) {
1623                         domain = found;
1624                         goto found_domain;
1625                 }
1626         }
1627
1628         /* Allocate new domain for the device */
1629         drhd = dmar_find_matched_drhd_unit(pdev);
1630         if (!drhd) {
1631                 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1632                         pci_name(pdev));
1633                 return NULL;
1634         }
1635         iommu = drhd->iommu;
1636
1637         domain = iommu_alloc_domain(iommu);
1638         if (!domain)
1639                 goto error;
1640
1641         if (domain_init(domain, gaw)) {
1642                 domain_exit(domain);
1643                 goto error;
1644         }
1645
1646         /* register pcie-to-pci device */
1647         if (dev_tmp) {
1648                 info = alloc_devinfo_mem();
1649                 if (!info) {
1650                         domain_exit(domain);
1651                         goto error;
1652                 }
1653                 info->segment = segment;
1654                 info->bus = bus;
1655                 info->devfn = devfn;
1656                 info->dev = NULL;
1657                 info->domain = domain;
1658                 /* This domain is shared by devices under p2p bridge */
1659                 domain->flags |= DOMAIN_FLAG_P2P_MULTIPLE_DEVICES;
1660
1661                 /* pcie-to-pci bridge already has a domain, uses it */
1662                 found = NULL;
1663                 spin_lock_irqsave(&device_domain_lock, flags);
1664                 list_for_each_entry(tmp, &device_domain_list, global) {
1665                         if (tmp->segment == segment &&
1666                             tmp->bus == bus && tmp->devfn == devfn) {
1667                                 found = tmp->domain;
1668                                 break;
1669                         }
1670                 }
1671                 if (found) {
1672                         free_devinfo_mem(info);
1673                         domain_exit(domain);
1674                         domain = found;
1675                 } else {
1676                         list_add(&info->link, &domain->devices);
1677                         list_add(&info->global, &device_domain_list);
1678                 }
1679                 spin_unlock_irqrestore(&device_domain_lock, flags);
1680         }
1681
1682 found_domain:
1683         info = alloc_devinfo_mem();
1684         if (!info)
1685                 goto error;
1686         info->segment = segment;
1687         info->bus = pdev->bus->number;
1688         info->devfn = pdev->devfn;
1689         info->dev = pdev;
1690         info->domain = domain;
1691         spin_lock_irqsave(&device_domain_lock, flags);
1692         /* somebody is fast */
1693         found = find_domain(pdev);
1694         if (found != NULL) {
1695                 spin_unlock_irqrestore(&device_domain_lock, flags);
1696                 if (found != domain) {
1697                         domain_exit(domain);
1698                         domain = found;
1699                 }
1700                 free_devinfo_mem(info);
1701                 return domain;
1702         }
1703         list_add(&info->link, &domain->devices);
1704         list_add(&info->global, &device_domain_list);
1705         pdev->dev.archdata.iommu = info;
1706         spin_unlock_irqrestore(&device_domain_lock, flags);
1707         return domain;
1708 error:
1709         /* recheck it here, maybe others set it */
1710         return find_domain(pdev);
1711 }
1712
1713 static int iommu_prepare_identity_map(struct pci_dev *pdev,
1714                                       unsigned long long start,
1715                                       unsigned long long end)
1716 {
1717         struct dmar_domain *domain;
1718         unsigned long size;
1719         unsigned long long base;
1720         int ret;
1721
1722         printk(KERN_INFO
1723                 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1724                 pci_name(pdev), start, end);
1725         /* page table init */
1726         domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1727         if (!domain)
1728                 return -ENOMEM;
1729
1730         /* The address might not be aligned */
1731         base = start & PAGE_MASK;
1732         size = end - base;
1733         size = PAGE_ALIGN(size);
1734         if (!reserve_iova(&domain->iovad, IOVA_PFN(base),
1735                         IOVA_PFN(base + size) - 1)) {
1736                 printk(KERN_ERR "IOMMU: reserve iova failed\n");
1737                 ret = -ENOMEM;
1738                 goto error;
1739         }
1740
1741         pr_debug("Mapping reserved region %lx@%llx for %s\n",
1742                 size, base, pci_name(pdev));
1743         /*
1744          * RMRR range might have overlap with physical memory range,
1745          * clear it first
1746          */
1747         dma_pte_clear_range(domain, base, base + size);
1748
1749         ret = domain_page_mapping(domain, base, base, size,
1750                 DMA_PTE_READ|DMA_PTE_WRITE);
1751         if (ret)
1752                 goto error;
1753
1754         /* context entry init */
1755         ret = domain_context_mapping(domain, pdev);
1756         if (!ret)
1757                 return 0;
1758 error:
1759         domain_exit(domain);
1760         return ret;
1761
1762 }
1763
1764 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1765         struct pci_dev *pdev)
1766 {
1767         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1768                 return 0;
1769         return iommu_prepare_identity_map(pdev, rmrr->base_address,
1770                 rmrr->end_address + 1);
1771 }
1772
1773 #ifdef CONFIG_DMAR_GFX_WA
1774 struct iommu_prepare_data {
1775         struct pci_dev *pdev;
1776         int ret;
1777 };
1778
1779 static int __init iommu_prepare_work_fn(unsigned long start_pfn,
1780                                          unsigned long end_pfn, void *datax)
1781 {
1782         struct iommu_prepare_data *data;
1783
1784         data = (struct iommu_prepare_data *)datax;
1785
1786         data->ret = iommu_prepare_identity_map(data->pdev,
1787                                 start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
1788         return data->ret;
1789
1790 }
1791
1792 static int __init iommu_prepare_with_active_regions(struct pci_dev *pdev)
1793 {
1794         int nid;
1795         struct iommu_prepare_data data;
1796
1797         data.pdev = pdev;
1798         data.ret = 0;
1799
1800         for_each_online_node(nid) {
1801                 work_with_active_regions(nid, iommu_prepare_work_fn, &data);
1802                 if (data.ret)
1803                         return data.ret;
1804         }
1805         return data.ret;
1806 }
1807
1808 static void __init iommu_prepare_gfx_mapping(void)
1809 {
1810         struct pci_dev *pdev = NULL;
1811         int ret;
1812
1813         for_each_pci_dev(pdev) {
1814                 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO ||
1815                                 !IS_GFX_DEVICE(pdev))
1816                         continue;
1817                 printk(KERN_INFO "IOMMU: gfx device %s 1-1 mapping\n",
1818                         pci_name(pdev));
1819                 ret = iommu_prepare_with_active_regions(pdev);
1820                 if (ret)
1821                         printk(KERN_ERR "IOMMU: mapping reserved region failed\n");
1822         }
1823 }
1824 #else /* !CONFIG_DMAR_GFX_WA */
1825 static inline void iommu_prepare_gfx_mapping(void)
1826 {
1827         return;
1828 }
1829 #endif
1830
1831 #ifdef CONFIG_DMAR_FLOPPY_WA
1832 static inline void iommu_prepare_isa(void)
1833 {
1834         struct pci_dev *pdev;
1835         int ret;
1836
1837         pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1838         if (!pdev)
1839                 return;
1840
1841         printk(KERN_INFO "IOMMU: Prepare 0-16M unity mapping for LPC\n");
1842         ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1843
1844         if (ret)
1845                 printk(KERN_ERR "IOMMU: Failed to create 0-64M identity map, "
1846                         "floppy might not work\n");
1847
1848 }
1849 #else
1850 static inline void iommu_prepare_isa(void)
1851 {
1852         return;
1853 }
1854 #endif /* !CONFIG_DMAR_FLPY_WA */
1855
1856 static int __init init_dmars(void)
1857 {
1858         struct dmar_drhd_unit *drhd;
1859         struct dmar_rmrr_unit *rmrr;
1860         struct pci_dev *pdev;
1861         struct intel_iommu *iommu;
1862         int i, ret;
1863
1864         /*
1865          * for each drhd
1866          *    allocate root
1867          *    initialize and program root entry to not present
1868          * endfor
1869          */
1870         for_each_drhd_unit(drhd) {
1871                 g_num_of_iommus++;
1872                 /*
1873                  * lock not needed as this is only incremented in the single
1874                  * threaded kernel __init code path all other access are read
1875                  * only
1876                  */
1877         }
1878
1879         g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
1880                         GFP_KERNEL);
1881         if (!g_iommus) {
1882                 printk(KERN_ERR "Allocating global iommu array failed\n");
1883                 ret = -ENOMEM;
1884                 goto error;
1885         }
1886
1887         deferred_flush = kzalloc(g_num_of_iommus *
1888                 sizeof(struct deferred_flush_tables), GFP_KERNEL);
1889         if (!deferred_flush) {
1890                 kfree(g_iommus);
1891                 ret = -ENOMEM;
1892                 goto error;
1893         }
1894
1895         for_each_drhd_unit(drhd) {
1896                 if (drhd->ignored)
1897                         continue;
1898
1899                 iommu = drhd->iommu;
1900                 g_iommus[iommu->seq_id] = iommu;
1901
1902                 ret = iommu_init_domains(iommu);
1903                 if (ret)
1904                         goto error;
1905
1906                 /*
1907                  * TBD:
1908                  * we could share the same root & context tables
1909                  * amoung all IOMMU's. Need to Split it later.
1910                  */
1911                 ret = iommu_alloc_root_entry(iommu);
1912                 if (ret) {
1913                         printk(KERN_ERR "IOMMU: allocate root entry failed\n");
1914                         goto error;
1915                 }
1916         }
1917
1918         /*
1919          * Start from the sane iommu hardware state.
1920          */
1921         for_each_drhd_unit(drhd) {
1922                 if (drhd->ignored)
1923                         continue;
1924
1925                 iommu = drhd->iommu;
1926
1927                 /*
1928                  * If the queued invalidation is already initialized by us
1929                  * (for example, while enabling interrupt-remapping) then
1930                  * we got the things already rolling from a sane state.
1931                  */
1932                 if (iommu->qi)
1933                         continue;
1934
1935                 /*
1936                  * Clear any previous faults.
1937                  */
1938                 dmar_fault(-1, iommu);
1939                 /*
1940                  * Disable queued invalidation if supported and already enabled
1941                  * before OS handover.
1942                  */
1943                 dmar_disable_qi(iommu);
1944         }
1945
1946         for_each_drhd_unit(drhd) {
1947                 if (drhd->ignored)
1948                         continue;
1949
1950                 iommu = drhd->iommu;
1951
1952                 if (dmar_enable_qi(iommu)) {
1953                         /*
1954                          * Queued Invalidate not enabled, use Register Based
1955                          * Invalidate
1956                          */
1957                         iommu->flush.flush_context = __iommu_flush_context;
1958                         iommu->flush.flush_iotlb = __iommu_flush_iotlb;
1959                         printk(KERN_INFO "IOMMU 0x%Lx: using Register based "
1960                                "invalidation\n",
1961                                (unsigned long long)drhd->reg_base_addr);
1962                 } else {
1963                         iommu->flush.flush_context = qi_flush_context;
1964                         iommu->flush.flush_iotlb = qi_flush_iotlb;
1965                         printk(KERN_INFO "IOMMU 0x%Lx: using Queued "
1966                                "invalidation\n",
1967                                (unsigned long long)drhd->reg_base_addr);
1968                 }
1969         }
1970
1971 #ifdef CONFIG_INTR_REMAP
1972         if (!intr_remapping_enabled) {
1973                 ret = enable_intr_remapping(0);
1974                 if (ret)
1975                         printk(KERN_ERR
1976                                "IOMMU: enable interrupt remapping failed\n");
1977         }
1978 #endif
1979
1980         /*
1981          * For each rmrr
1982          *   for each dev attached to rmrr
1983          *   do
1984          *     locate drhd for dev, alloc domain for dev
1985          *     allocate free domain
1986          *     allocate page table entries for rmrr
1987          *     if context not allocated for bus
1988          *           allocate and init context
1989          *           set present in root table for this bus
1990          *     init context with domain, translation etc
1991          *    endfor
1992          * endfor
1993          */
1994         for_each_rmrr_units(rmrr) {
1995                 for (i = 0; i < rmrr->devices_cnt; i++) {
1996                         pdev = rmrr->devices[i];
1997                         /* some BIOS lists non-exist devices in DMAR table */
1998                         if (!pdev)
1999                                 continue;
2000                         ret = iommu_prepare_rmrr_dev(rmrr, pdev);
2001                         if (ret)
2002                                 printk(KERN_ERR
2003                                  "IOMMU: mapping reserved region failed\n");
2004                 }
2005         }
2006
2007         iommu_prepare_gfx_mapping();
2008
2009         iommu_prepare_isa();
2010
2011         /*
2012          * for each drhd
2013          *   enable fault log
2014          *   global invalidate context cache
2015          *   global invalidate iotlb
2016          *   enable translation
2017          */
2018         for_each_drhd_unit(drhd) {
2019                 if (drhd->ignored)
2020                         continue;
2021                 iommu = drhd->iommu;
2022
2023                 iommu_flush_write_buffer(iommu);
2024
2025                 ret = dmar_set_interrupt(iommu);
2026                 if (ret)
2027                         goto error;
2028
2029                 iommu_set_root_entry(iommu);
2030
2031                 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL,
2032                                            0);
2033                 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH,
2034                                          0);
2035                 iommu_disable_protect_mem_regions(iommu);
2036
2037                 ret = iommu_enable_translation(iommu);
2038                 if (ret)
2039                         goto error;
2040         }
2041
2042         return 0;
2043 error:
2044         for_each_drhd_unit(drhd) {
2045                 if (drhd->ignored)
2046                         continue;
2047                 iommu = drhd->iommu;
2048                 free_iommu(iommu);
2049         }
2050         kfree(g_iommus);
2051         return ret;
2052 }
2053
2054 static inline u64 aligned_size(u64 host_addr, size_t size)
2055 {
2056         u64 addr;
2057         addr = (host_addr & (~PAGE_MASK)) + size;
2058         return PAGE_ALIGN(addr);
2059 }
2060
2061 struct iova *
2062 iommu_alloc_iova(struct dmar_domain *domain, size_t size, u64 end)
2063 {
2064         struct iova *piova;
2065
2066         /* Make sure it's in range */
2067         end = min_t(u64, DOMAIN_MAX_ADDR(domain->gaw), end);
2068         if (!size || (IOVA_START_ADDR + size > end))
2069                 return NULL;
2070
2071         piova = alloc_iova(&domain->iovad,
2072                         size >> PAGE_SHIFT, IOVA_PFN(end), 1);
2073         return piova;
2074 }
2075
2076 static struct iova *
2077 __intel_alloc_iova(struct device *dev, struct dmar_domain *domain,
2078                    size_t size, u64 dma_mask)
2079 {
2080         struct pci_dev *pdev = to_pci_dev(dev);
2081         struct iova *iova = NULL;
2082
2083         if (dma_mask <= DMA_BIT_MASK(32) || dmar_forcedac)
2084                 iova = iommu_alloc_iova(domain, size, dma_mask);
2085         else {
2086                 /*
2087                  * First try to allocate an io virtual address in
2088                  * DMA_BIT_MASK(32) and if that fails then try allocating
2089                  * from higher range
2090                  */
2091                 iova = iommu_alloc_iova(domain, size, DMA_BIT_MASK(32));
2092                 if (!iova)
2093                         iova = iommu_alloc_iova(domain, size, dma_mask);
2094         }
2095
2096         if (!iova) {
2097                 printk(KERN_ERR"Allocating iova for %s failed", pci_name(pdev));
2098                 return NULL;
2099         }
2100
2101         return iova;
2102 }
2103
2104 static struct dmar_domain *
2105 get_valid_domain_for_dev(struct pci_dev *pdev)
2106 {
2107         struct dmar_domain *domain;
2108         int ret;
2109
2110         domain = get_domain_for_dev(pdev,
2111                         DEFAULT_DOMAIN_ADDRESS_WIDTH);
2112         if (!domain) {
2113                 printk(KERN_ERR
2114                         "Allocating domain for %s failed", pci_name(pdev));
2115                 return NULL;
2116         }
2117
2118         /* make sure context mapping is ok */
2119         if (unlikely(!domain_context_mapped(pdev))) {
2120                 ret = domain_context_mapping(domain, pdev);
2121                 if (ret) {
2122                         printk(KERN_ERR
2123                                 "Domain context map for %s failed",
2124                                 pci_name(pdev));
2125                         return NULL;
2126                 }
2127         }
2128
2129         return domain;
2130 }
2131
2132 static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
2133                                      size_t size, int dir, u64 dma_mask)
2134 {
2135         struct pci_dev *pdev = to_pci_dev(hwdev);
2136         struct dmar_domain *domain;
2137         phys_addr_t start_paddr;
2138         struct iova *iova;
2139         int prot = 0;
2140         int ret;
2141         struct intel_iommu *iommu;
2142
2143         BUG_ON(dir == DMA_NONE);
2144         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2145                 return paddr;
2146
2147         domain = get_valid_domain_for_dev(pdev);
2148         if (!domain)
2149                 return 0;
2150
2151         iommu = domain_get_iommu(domain);
2152         size = aligned_size((u64)paddr, size);
2153
2154         iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2155         if (!iova)
2156                 goto error;
2157
2158         start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
2159
2160         /*
2161          * Check if DMAR supports zero-length reads on write only
2162          * mappings..
2163          */
2164         if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2165                         !cap_zlr(iommu->cap))
2166                 prot |= DMA_PTE_READ;
2167         if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2168                 prot |= DMA_PTE_WRITE;
2169         /*
2170          * paddr - (paddr + size) might be partial page, we should map the whole
2171          * page.  Note: if two part of one page are separately mapped, we
2172          * might have two guest_addr mapping to the same host paddr, but this
2173          * is not a big problem
2174          */
2175         ret = domain_page_mapping(domain, start_paddr,
2176                 ((u64)paddr) & PAGE_MASK, size, prot);
2177         if (ret)
2178                 goto error;
2179
2180         /* it's a non-present to present mapping */
2181         ret = iommu_flush_iotlb_psi(iommu, domain->id,
2182                         start_paddr, size >> VTD_PAGE_SHIFT, 1);
2183         if (ret)
2184                 iommu_flush_write_buffer(iommu);
2185
2186         return start_paddr + ((u64)paddr & (~PAGE_MASK));
2187
2188 error:
2189         if (iova)
2190                 __free_iova(&domain->iovad, iova);
2191         printk(KERN_ERR"Device %s request: %zx@%llx dir %d --- failed\n",
2192                 pci_name(pdev), size, (unsigned long long)paddr, dir);
2193         return 0;
2194 }
2195
2196 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
2197                                  unsigned long offset, size_t size,
2198                                  enum dma_data_direction dir,
2199                                  struct dma_attrs *attrs)
2200 {
2201         return __intel_map_single(dev, page_to_phys(page) + offset, size,
2202                                   dir, to_pci_dev(dev)->dma_mask);
2203 }
2204
2205 static void flush_unmaps(void)
2206 {
2207         int i, j;
2208
2209         timer_on = 0;
2210
2211         /* just flush them all */
2212         for (i = 0; i < g_num_of_iommus; i++) {
2213                 struct intel_iommu *iommu = g_iommus[i];
2214                 if (!iommu)
2215                         continue;
2216
2217                 if (deferred_flush[i].next) {
2218                         iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2219                                                  DMA_TLB_GLOBAL_FLUSH, 0);
2220                         for (j = 0; j < deferred_flush[i].next; j++) {
2221                                 __free_iova(&deferred_flush[i].domain[j]->iovad,
2222                                                 deferred_flush[i].iova[j]);
2223                         }
2224                         deferred_flush[i].next = 0;
2225                 }
2226         }
2227
2228         list_size = 0;
2229 }
2230
2231 static void flush_unmaps_timeout(unsigned long data)
2232 {
2233         unsigned long flags;
2234
2235         spin_lock_irqsave(&async_umap_flush_lock, flags);
2236         flush_unmaps();
2237         spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2238 }
2239
2240 static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2241 {
2242         unsigned long flags;
2243         int next, iommu_id;
2244         struct intel_iommu *iommu;
2245
2246         spin_lock_irqsave(&async_umap_flush_lock, flags);
2247         if (list_size == HIGH_WATER_MARK)
2248                 flush_unmaps();
2249
2250         iommu = domain_get_iommu(dom);
2251         iommu_id = iommu->seq_id;
2252
2253         next = deferred_flush[iommu_id].next;
2254         deferred_flush[iommu_id].domain[next] = dom;
2255         deferred_flush[iommu_id].iova[next] = iova;
2256         deferred_flush[iommu_id].next++;
2257
2258         if (!timer_on) {
2259                 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2260                 timer_on = 1;
2261         }
2262         list_size++;
2263         spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2264 }
2265
2266 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
2267                              size_t size, enum dma_data_direction dir,
2268                              struct dma_attrs *attrs)
2269 {
2270         struct pci_dev *pdev = to_pci_dev(dev);
2271         struct dmar_domain *domain;
2272         unsigned long start_addr;
2273         struct iova *iova;
2274         struct intel_iommu *iommu;
2275
2276         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2277                 return;
2278         domain = find_domain(pdev);
2279         BUG_ON(!domain);
2280
2281         iommu = domain_get_iommu(domain);
2282
2283         iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
2284         if (!iova)
2285                 return;
2286
2287         start_addr = iova->pfn_lo << PAGE_SHIFT;
2288         size = aligned_size((u64)dev_addr, size);
2289
2290         pr_debug("Device %s unmapping: %zx@%llx\n",
2291                 pci_name(pdev), size, (unsigned long long)start_addr);
2292
2293         /*  clear the whole page */
2294         dma_pte_clear_range(domain, start_addr, start_addr + size);
2295         /* free page tables */
2296         dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2297         if (intel_iommu_strict) {
2298                 if (iommu_flush_iotlb_psi(iommu,
2299                         domain->id, start_addr, size >> VTD_PAGE_SHIFT, 0))
2300                         iommu_flush_write_buffer(iommu);
2301                 /* free iova */
2302                 __free_iova(&domain->iovad, iova);
2303         } else {
2304                 add_unmap(domain, iova);
2305                 /*
2306                  * queue up the release of the unmap to save the 1/6th of the
2307                  * cpu used up by the iotlb flush operation...
2308                  */
2309         }
2310 }
2311
2312 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr, size_t size,
2313                                int dir)
2314 {
2315         intel_unmap_page(dev, dev_addr, size, dir, NULL);
2316 }
2317
2318 static void *intel_alloc_coherent(struct device *hwdev, size_t size,
2319                                   dma_addr_t *dma_handle, gfp_t flags)
2320 {
2321         void *vaddr;
2322         int order;
2323
2324         size = PAGE_ALIGN(size);
2325         order = get_order(size);
2326         flags &= ~(GFP_DMA | GFP_DMA32);
2327
2328         vaddr = (void *)__get_free_pages(flags, order);
2329         if (!vaddr)
2330                 return NULL;
2331         memset(vaddr, 0, size);
2332
2333         *dma_handle = __intel_map_single(hwdev, virt_to_bus(vaddr), size,
2334                                          DMA_BIDIRECTIONAL,
2335                                          hwdev->coherent_dma_mask);
2336         if (*dma_handle)
2337                 return vaddr;
2338         free_pages((unsigned long)vaddr, order);
2339         return NULL;
2340 }
2341
2342 static void intel_free_coherent(struct device *hwdev, size_t size, void *vaddr,
2343                                 dma_addr_t dma_handle)
2344 {
2345         int order;
2346
2347         size = PAGE_ALIGN(size);
2348         order = get_order(size);
2349
2350         intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2351         free_pages((unsigned long)vaddr, order);
2352 }
2353
2354 static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2355                            int nelems, enum dma_data_direction dir,
2356                            struct dma_attrs *attrs)
2357 {
2358         int i;
2359         struct pci_dev *pdev = to_pci_dev(hwdev);
2360         struct dmar_domain *domain;
2361         unsigned long start_addr;
2362         struct iova *iova;
2363         size_t size = 0;
2364         phys_addr_t addr;
2365         struct scatterlist *sg;
2366         struct intel_iommu *iommu;
2367
2368         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2369                 return;
2370
2371         domain = find_domain(pdev);
2372         BUG_ON(!domain);
2373
2374         iommu = domain_get_iommu(domain);
2375
2376         iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
2377         if (!iova)
2378                 return;
2379         for_each_sg(sglist, sg, nelems, i) {
2380                 addr = page_to_phys(sg_page(sg)) + sg->offset;
2381                 size += aligned_size((u64)addr, sg->length);
2382         }
2383
2384         start_addr = iova->pfn_lo << PAGE_SHIFT;
2385
2386         /*  clear the whole page */
2387         dma_pte_clear_range(domain, start_addr, start_addr + size);
2388         /* free page tables */
2389         dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2390
2391         if (iommu_flush_iotlb_psi(iommu, domain->id, start_addr,
2392                         size >> VTD_PAGE_SHIFT, 0))
2393                 iommu_flush_write_buffer(iommu);
2394
2395         /* free iova */
2396         __free_iova(&domain->iovad, iova);
2397 }
2398
2399 static int intel_nontranslate_map_sg(struct device *hddev,
2400         struct scatterlist *sglist, int nelems, int dir)
2401 {
2402         int i;
2403         struct scatterlist *sg;
2404
2405         for_each_sg(sglist, sg, nelems, i) {
2406                 BUG_ON(!sg_page(sg));
2407                 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
2408                 sg->dma_length = sg->length;
2409         }
2410         return nelems;
2411 }
2412
2413 static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist, int nelems,
2414                         enum dma_data_direction dir, struct dma_attrs *attrs)
2415 {
2416         phys_addr_t addr;
2417         int i;
2418         struct pci_dev *pdev = to_pci_dev(hwdev);
2419         struct dmar_domain *domain;
2420         size_t size = 0;
2421         int prot = 0;
2422         size_t offset = 0;
2423         struct iova *iova = NULL;
2424         int ret;
2425         struct scatterlist *sg;
2426         unsigned long start_addr;
2427         struct intel_iommu *iommu;
2428
2429         BUG_ON(dir == DMA_NONE);
2430         if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2431                 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
2432
2433         domain = get_valid_domain_for_dev(pdev);
2434         if (!domain)
2435                 return 0;
2436
2437         iommu = domain_get_iommu(domain);
2438
2439         for_each_sg(sglist, sg, nelems, i) {
2440                 addr = page_to_phys(sg_page(sg)) + sg->offset;
2441                 size += aligned_size((u64)addr, sg->length);
2442         }
2443
2444         iova = __intel_alloc_iova(hwdev, domain, size, pdev->dma_mask);
2445         if (!iova) {
2446                 sglist->dma_length = 0;
2447                 return 0;
2448         }
2449
2450         /*
2451          * Check if DMAR supports zero-length reads on write only
2452          * mappings..
2453          */
2454         if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2455                         !cap_zlr(iommu->cap))
2456                 prot |= DMA_PTE_READ;
2457         if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2458                 prot |= DMA_PTE_WRITE;
2459
2460         start_addr = iova->pfn_lo << PAGE_SHIFT;
2461         offset = 0;
2462         for_each_sg(sglist, sg, nelems, i) {
2463                 addr = page_to_phys(sg_page(sg)) + sg->offset;
2464                 size = aligned_size((u64)addr, sg->length);
2465                 ret = domain_page_mapping(domain, start_addr + offset,
2466                         ((u64)addr) & PAGE_MASK,
2467                         size, prot);
2468                 if (ret) {
2469                         /*  clear the page */
2470                         dma_pte_clear_range(domain, start_addr,
2471                                   start_addr + offset);
2472                         /* free page tables */
2473                         dma_pte_free_pagetable(domain, start_addr,
2474                                   start_addr + offset);
2475                         /* free iova */
2476                         __free_iova(&domain->iovad, iova);
2477                         return 0;
2478                 }
2479                 sg->dma_address = start_addr + offset +
2480                                 ((u64)addr & (~PAGE_MASK));
2481                 sg->dma_length = sg->length;
2482                 offset += size;
2483         }
2484
2485         /* it's a non-present to present mapping */
2486         if (iommu_flush_iotlb_psi(iommu, domain->id,
2487                         start_addr, offset >> VTD_PAGE_SHIFT, 1))
2488                 iommu_flush_write_buffer(iommu);
2489         return nelems;
2490 }
2491
2492 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
2493 {
2494         return !dma_addr;
2495 }
2496
2497 struct dma_map_ops intel_dma_ops = {
2498         .alloc_coherent = intel_alloc_coherent,
2499         .free_coherent = intel_free_coherent,
2500         .map_sg = intel_map_sg,
2501         .unmap_sg = intel_unmap_sg,
2502         .map_page = intel_map_page,
2503         .unmap_page = intel_unmap_page,
2504         .mapping_error = intel_mapping_error,
2505 };
2506
2507 static inline int iommu_domain_cache_init(void)
2508 {
2509         int ret = 0;
2510
2511         iommu_domain_cache = kmem_cache_create("iommu_domain",
2512                                          sizeof(struct dmar_domain),
2513                                          0,
2514                                          SLAB_HWCACHE_ALIGN,
2515
2516                                          NULL);
2517         if (!iommu_domain_cache) {
2518                 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2519                 ret = -ENOMEM;
2520         }
2521
2522         return ret;
2523 }
2524
2525 static inline int iommu_devinfo_cache_init(void)
2526 {
2527         int ret = 0;
2528
2529         iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2530                                          sizeof(struct device_domain_info),
2531                                          0,
2532                                          SLAB_HWCACHE_ALIGN,
2533                                          NULL);
2534         if (!iommu_devinfo_cache) {
2535                 printk(KERN_ERR "Couldn't create devinfo cache\n");
2536                 ret = -ENOMEM;
2537         }
2538
2539         return ret;
2540 }
2541
2542 static inline int iommu_iova_cache_init(void)
2543 {
2544         int ret = 0;
2545
2546         iommu_iova_cache = kmem_cache_create("iommu_iova",
2547                                          sizeof(struct iova),
2548                                          0,
2549                                          SLAB_HWCACHE_ALIGN,
2550                                          NULL);
2551         if (!iommu_iova_cache) {
2552                 printk(KERN_ERR "Couldn't create iova cache\n");
2553                 ret = -ENOMEM;
2554         }
2555
2556         return ret;
2557 }
2558
2559 static int __init iommu_init_mempool(void)
2560 {
2561         int ret;
2562         ret = iommu_iova_cache_init();
2563         if (ret)
2564                 return ret;
2565
2566         ret = iommu_domain_cache_init();
2567         if (ret)
2568                 goto domain_error;
2569
2570         ret = iommu_devinfo_cache_init();
2571         if (!ret)
2572                 return ret;
2573
2574         kmem_cache_destroy(iommu_domain_cache);
2575 domain_error:
2576         kmem_cache_destroy(iommu_iova_cache);
2577
2578         return -ENOMEM;
2579 }
2580
2581 static void __init iommu_exit_mempool(void)
2582 {
2583         kmem_cache_destroy(iommu_devinfo_cache);
2584         kmem_cache_destroy(iommu_domain_cache);
2585         kmem_cache_destroy(iommu_iova_cache);
2586
2587 }
2588
2589 static void __init init_no_remapping_devices(void)
2590 {
2591         struct dmar_drhd_unit *drhd;
2592
2593         for_each_drhd_unit(drhd) {
2594                 if (!drhd->include_all) {
2595                         int i;
2596                         for (i = 0; i < drhd->devices_cnt; i++)
2597                                 if (drhd->devices[i] != NULL)
2598                                         break;
2599                         /* ignore DMAR unit if no pci devices exist */
2600                         if (i == drhd->devices_cnt)
2601                                 drhd->ignored = 1;
2602                 }
2603         }
2604
2605         if (dmar_map_gfx)
2606                 return;
2607
2608         for_each_drhd_unit(drhd) {
2609                 int i;
2610                 if (drhd->ignored || drhd->include_all)
2611                         continue;
2612
2613                 for (i = 0; i < drhd->devices_cnt; i++)
2614                         if (drhd->devices[i] &&
2615                                 !IS_GFX_DEVICE(drhd->devices[i]))
2616                                 break;
2617
2618                 if (i < drhd->devices_cnt)
2619                         continue;
2620
2621                 /* bypass IOMMU if it is just for gfx devices */
2622                 drhd->ignored = 1;
2623                 for (i = 0; i < drhd->devices_cnt; i++) {
2624                         if (!drhd->devices[i])
2625                                 continue;
2626                         drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
2627                 }
2628         }
2629 }
2630
2631 #ifdef CONFIG_SUSPEND
2632 static int init_iommu_hw(void)
2633 {
2634         struct dmar_drhd_unit *drhd;
2635         struct intel_iommu *iommu = NULL;
2636
2637         for_each_active_iommu(iommu, drhd)
2638                 if (iommu->qi)
2639                         dmar_reenable_qi(iommu);
2640
2641         for_each_active_iommu(iommu, drhd) {
2642                 iommu_flush_write_buffer(iommu);
2643
2644                 iommu_set_root_entry(iommu);
2645
2646                 iommu->flush.flush_context(iommu, 0, 0, 0,
2647                                                 DMA_CCMD_GLOBAL_INVL, 0);
2648                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2649                                                 DMA_TLB_GLOBAL_FLUSH, 0);
2650                 iommu_disable_protect_mem_regions(iommu);
2651                 iommu_enable_translation(iommu);
2652         }
2653
2654         return 0;
2655 }
2656
2657 static void iommu_flush_all(void)
2658 {
2659         struct dmar_drhd_unit *drhd;
2660         struct intel_iommu *iommu;
2661
2662         for_each_active_iommu(iommu, drhd) {
2663                 iommu->flush.flush_context(iommu, 0, 0, 0,
2664                                                 DMA_CCMD_GLOBAL_INVL, 0);
2665                 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
2666                                                 DMA_TLB_GLOBAL_FLUSH, 0);
2667         }
2668 }
2669
2670 static int iommu_suspend(struct sys_device *dev, pm_message_t state)
2671 {
2672         struct dmar_drhd_unit *drhd;
2673         struct intel_iommu *iommu = NULL;
2674         unsigned long flag;
2675
2676         for_each_active_iommu(iommu, drhd) {
2677                 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
2678                                                  GFP_ATOMIC);
2679                 if (!iommu->iommu_state)
2680                         goto nomem;
2681         }
2682
2683         iommu_flush_all();
2684
2685         for_each_active_iommu(iommu, drhd) {
2686                 iommu_disable_translation(iommu);
2687
2688                 spin_lock_irqsave(&iommu->register_lock, flag);
2689
2690                 iommu->iommu_state[SR_DMAR_FECTL_REG] =
2691                         readl(iommu->reg + DMAR_FECTL_REG);
2692                 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
2693                         readl(iommu->reg + DMAR_FEDATA_REG);
2694                 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
2695                         readl(iommu->reg + DMAR_FEADDR_REG);
2696                 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
2697                         readl(iommu->reg + DMAR_FEUADDR_REG);
2698
2699                 spin_unlock_irqrestore(&iommu->register_lock, flag);
2700         }
2701         return 0;
2702
2703 nomem:
2704         for_each_active_iommu(iommu, drhd)
2705                 kfree(iommu->iommu_state);
2706
2707         return -ENOMEM;
2708 }
2709
2710 static int iommu_resume(struct sys_device *dev)
2711 {
2712         struct dmar_drhd_unit *drhd;
2713         struct intel_iommu *iommu = NULL;
2714         unsigned long flag;
2715
2716         if (init_iommu_hw()) {
2717                 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
2718                 return -EIO;
2719         }
2720
2721         for_each_active_iommu(iommu, drhd) {
2722
2723                 spin_lock_irqsave(&iommu->register_lock, flag);
2724
2725                 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
2726                         iommu->reg + DMAR_FECTL_REG);
2727                 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
2728                         iommu->reg + DMAR_FEDATA_REG);
2729                 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
2730                         iommu->reg + DMAR_FEADDR_REG);
2731                 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
2732                         iommu->reg + DMAR_FEUADDR_REG);
2733
2734                 spin_unlock_irqrestore(&iommu->register_lock, flag);
2735         }
2736
2737         for_each_active_iommu(iommu, drhd)
2738                 kfree(iommu->iommu_state);
2739
2740         return 0;
2741 }
2742
2743 static struct sysdev_class iommu_sysclass = {
2744         .name           = "iommu",
2745         .resume         = iommu_resume,
2746         .suspend        = iommu_suspend,
2747 };
2748
2749 static struct sys_device device_iommu = {
2750         .cls    = &iommu_sysclass,
2751 };
2752
2753 static int __init init_iommu_sysfs(void)
2754 {
2755         int error;
2756
2757         error = sysdev_class_register(&iommu_sysclass);
2758         if (error)
2759                 return error;
2760
2761         error = sysdev_register(&device_iommu);
2762         if (error)
2763                 sysdev_class_unregister(&iommu_sysclass);
2764
2765         return error;
2766 }
2767
2768 #else
2769 static int __init init_iommu_sysfs(void)
2770 {
2771         return 0;
2772 }
2773 #endif  /* CONFIG_PM */
2774
2775 int __init intel_iommu_init(void)
2776 {
2777         int ret = 0;
2778
2779         if (dmar_table_init())
2780                 return  -ENODEV;
2781
2782         if (dmar_dev_scope_init())
2783                 return  -ENODEV;
2784
2785         /*
2786          * Check the need for DMA-remapping initialization now.
2787          * Above initialization will also be used by Interrupt-remapping.
2788          */
2789         if (no_iommu || swiotlb || dmar_disabled)
2790                 return -ENODEV;
2791
2792         iommu_init_mempool();
2793         dmar_init_reserved_ranges();
2794
2795         init_no_remapping_devices();
2796
2797         ret = init_dmars();
2798         if (ret) {
2799                 printk(KERN_ERR "IOMMU: dmar init failed\n");
2800                 put_iova_domain(&reserved_iova_list);
2801                 iommu_exit_mempool();
2802                 return ret;
2803         }
2804         printk(KERN_INFO
2805         "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
2806
2807         init_timer(&unmap_timer);
2808         force_iommu = 1;
2809         dma_ops = &intel_dma_ops;
2810         init_iommu_sysfs();
2811
2812         register_iommu(&intel_iommu_ops);
2813
2814         return 0;
2815 }
2816
2817 static int vm_domain_add_dev_info(struct dmar_domain *domain,
2818                                   struct pci_dev *pdev)
2819 {
2820         struct device_domain_info *info;
2821         unsigned long flags;
2822
2823         info = alloc_devinfo_mem();
2824         if (!info)
2825                 return -ENOMEM;
2826
2827         info->segment = pci_domain_nr(pdev->bus);
2828         info->bus = pdev->bus->number;
2829         info->devfn = pdev->devfn;
2830         info->dev = pdev;
2831         info->domain = domain;
2832
2833         spin_lock_irqsave(&device_domain_lock, flags);
2834         list_add(&info->link, &domain->devices);
2835         list_add(&info->global, &device_domain_list);
2836         pdev->dev.archdata.iommu = info;
2837         spin_unlock_irqrestore(&device_domain_lock, flags);
2838
2839         return 0;
2840 }
2841
2842 static void iommu_detach_dependent_devices(struct intel_iommu *iommu,
2843                                            struct pci_dev *pdev)
2844 {
2845         struct pci_dev *tmp, *parent;
2846
2847         if (!iommu || !pdev)
2848                 return;
2849
2850         /* dependent device detach */
2851         tmp = pci_find_upstream_pcie_bridge(pdev);
2852         /* Secondary interface's bus number and devfn 0 */
2853         if (tmp) {
2854                 parent = pdev->bus->self;
2855                 while (parent != tmp) {
2856                         iommu_detach_dev(iommu, parent->bus->number,
2857                                          parent->devfn);
2858                         parent = parent->bus->self;
2859                 }
2860                 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
2861                         iommu_detach_dev(iommu,
2862                                 tmp->subordinate->number, 0);
2863                 else /* this is a legacy PCI bridge */
2864                         iommu_detach_dev(iommu, tmp->bus->number,
2865                                          tmp->devfn);
2866         }
2867 }
2868
2869 static void vm_domain_remove_one_dev_info(struct dmar_domain *domain,
2870                                           struct pci_dev *pdev)
2871 {
2872         struct device_domain_info *info;
2873         struct intel_iommu *iommu;
2874         unsigned long flags;
2875         int found = 0;
2876         struct list_head *entry, *tmp;
2877
2878         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
2879                                 pdev->devfn);
2880         if (!iommu)
2881                 return;
2882
2883         spin_lock_irqsave(&device_domain_lock, flags);
2884         list_for_each_safe(entry, tmp, &domain->devices) {
2885                 info = list_entry(entry, struct device_domain_info, link);
2886                 /* No need to compare PCI domain; it has to be the same */
2887                 if (info->bus == pdev->bus->number &&
2888                     info->devfn == pdev->devfn) {
2889                         list_del(&info->link);
2890                         list_del(&info->global);
2891                         if (info->dev)
2892                                 info->dev->dev.archdata.iommu = NULL;
2893                         spin_unlock_irqrestore(&device_domain_lock, flags);
2894
2895                         iommu_detach_dev(iommu, info->bus, info->devfn);
2896                         iommu_detach_dependent_devices(iommu, pdev);
2897                         free_devinfo_mem(info);
2898
2899                         spin_lock_irqsave(&device_domain_lock, flags);
2900
2901                         if (found)
2902                                 break;
2903                         else
2904                                 continue;
2905                 }
2906
2907                 /* if there is no other devices under the same iommu
2908                  * owned by this domain, clear this iommu in iommu_bmp
2909                  * update iommu count and coherency
2910                  */
2911                 if (iommu == device_to_iommu(info->segment, info->bus,
2912                                             info->devfn))
2913                         found = 1;
2914         }
2915
2916         if (found == 0) {
2917                 unsigned long tmp_flags;
2918                 spin_lock_irqsave(&domain->iommu_lock, tmp_flags);
2919                 clear_bit(iommu->seq_id, &domain->iommu_bmp);
2920                 domain->iommu_count--;
2921                 domain_update_iommu_cap(domain);
2922                 spin_unlock_irqrestore(&domain->iommu_lock, tmp_flags);
2923         }
2924
2925         spin_unlock_irqrestore(&device_domain_lock, flags);
2926 }
2927
2928 static void vm_domain_remove_all_dev_info(struct dmar_domain *domain)
2929 {
2930         struct device_domain_info *info;
2931         struct intel_iommu *iommu;
2932         unsigned long flags1, flags2;
2933
2934         spin_lock_irqsave(&device_domain_lock, flags1);
2935         while (!list_empty(&domain->devices)) {
2936                 info = list_entry(domain->devices.next,
2937                         struct device_domain_info, link);
2938                 list_del(&info->link);
2939                 list_del(&info->global);
2940                 if (info->dev)
2941                         info->dev->dev.archdata.iommu = NULL;
2942
2943                 spin_unlock_irqrestore(&device_domain_lock, flags1);
2944
2945                 iommu = device_to_iommu(info->segment, info->bus, info->devfn);
2946                 iommu_detach_dev(iommu, info->bus, info->devfn);
2947                 iommu_detach_dependent_devices(iommu, info->dev);
2948
2949                 /* clear this iommu in iommu_bmp, update iommu count
2950                  * and capabilities
2951                  */
2952                 spin_lock_irqsave(&domain->iommu_lock, flags2);
2953                 if (test_and_clear_bit(iommu->seq_id,
2954                                        &domain->iommu_bmp)) {
2955                         domain->iommu_count--;
2956                         domain_update_iommu_cap(domain);
2957                 }
2958                 spin_unlock_irqrestore(&domain->iommu_lock, flags2);
2959
2960                 free_devinfo_mem(info);
2961                 spin_lock_irqsave(&device_domain_lock, flags1);
2962         }
2963         spin_unlock_irqrestore(&device_domain_lock, flags1);
2964 }
2965
2966 /* domain id for virtual machine, it won't be set in context */
2967 static unsigned long vm_domid;
2968
2969 static int vm_domain_min_agaw(struct dmar_domain *domain)
2970 {
2971         int i;
2972         int min_agaw = domain->agaw;
2973
2974         i = find_first_bit(&domain->iommu_bmp, g_num_of_iommus);
2975         for (; i < g_num_of_iommus; ) {
2976                 if (min_agaw > g_iommus[i]->agaw)
2977                         min_agaw = g_iommus[i]->agaw;
2978
2979                 i = find_next_bit(&domain->iommu_bmp, g_num_of_iommus, i+1);
2980         }
2981
2982         return min_agaw;
2983 }
2984
2985 static struct dmar_domain *iommu_alloc_vm_domain(void)
2986 {
2987         struct dmar_domain *domain;
2988
2989         domain = alloc_domain_mem();
2990         if (!domain)
2991                 return NULL;
2992
2993         domain->id = vm_domid++;
2994         memset(&domain->iommu_bmp, 0, sizeof(unsigned long));
2995         domain->flags = DOMAIN_FLAG_VIRTUAL_MACHINE;
2996
2997         return domain;
2998 }
2999
3000 static int vm_domain_init(struct dmar_domain *domain, int guest_width)
3001 {
3002         int adjust_width;
3003
3004         init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
3005         spin_lock_init(&domain->mapping_lock);
3006         spin_lock_init(&domain->iommu_lock);
3007
3008         domain_reserve_special_ranges(domain);
3009
3010         /* calculate AGAW */
3011         domain->gaw = guest_width;
3012         adjust_width = guestwidth_to_adjustwidth(guest_width);
3013         domain->agaw = width_to_agaw(adjust_width);
3014
3015         INIT_LIST_HEAD(&domain->devices);
3016
3017         domain->iommu_count = 0;
3018         domain->iommu_coherency = 0;
3019         domain->max_addr = 0;
3020
3021         /* always allocate the top pgd */
3022         domain->pgd = (struct dma_pte *)alloc_pgtable_page();
3023         if (!domain->pgd)
3024                 return -ENOMEM;
3025         domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
3026         return 0;
3027 }
3028
3029 static void iommu_free_vm_domain(struct dmar_domain *domain)
3030 {
3031         unsigned long flags;
3032         struct dmar_drhd_unit *drhd;
3033         struct intel_iommu *iommu;
3034         unsigned long i;
3035         unsigned long ndomains;
3036
3037         for_each_drhd_unit(drhd) {
3038                 if (drhd->ignored)
3039                         continue;
3040                 iommu = drhd->iommu;
3041
3042                 ndomains = cap_ndoms(iommu->cap);
3043                 i = find_first_bit(iommu->domain_ids, ndomains);
3044                 for (; i < ndomains; ) {
3045                         if (iommu->domains[i] == domain) {
3046                                 spin_lock_irqsave(&iommu->lock, flags);
3047                                 clear_bit(i, iommu->domain_ids);
3048                                 iommu->domains[i] = NULL;
3049                                 spin_unlock_irqrestore(&iommu->lock, flags);
3050                                 break;
3051                         }
3052                         i = find_next_bit(iommu->domain_ids, ndomains, i+1);
3053                 }
3054         }
3055 }
3056
3057 static void vm_domain_exit(struct dmar_domain *domain)
3058 {
3059         u64 end;
3060
3061         /* Domain 0 is reserved, so dont process it */
3062         if (!domain)
3063                 return;
3064
3065         vm_domain_remove_all_dev_info(domain);
3066         /* destroy iovas */
3067         put_iova_domain(&domain->iovad);
3068         end = DOMAIN_MAX_ADDR(domain->gaw);
3069         end = end & (~VTD_PAGE_MASK);
3070
3071         /* clear ptes */
3072         dma_pte_clear_range(domain, 0, end);
3073
3074         /* free page tables */
3075         dma_pte_free_pagetable(domain, 0, end);
3076
3077         iommu_free_vm_domain(domain);
3078         free_domain_mem(domain);
3079 }
3080
3081 static int intel_iommu_domain_init(struct iommu_domain *domain)
3082 {
3083         struct dmar_domain *dmar_domain;
3084
3085         dmar_domain = iommu_alloc_vm_domain();
3086         if (!dmar_domain) {
3087                 printk(KERN_ERR
3088                         "intel_iommu_domain_init: dmar_domain == NULL\n");
3089                 return -ENOMEM;
3090         }
3091         if (vm_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
3092                 printk(KERN_ERR
3093                         "intel_iommu_domain_init() failed\n");
3094                 vm_domain_exit(dmar_domain);
3095                 return -ENOMEM;
3096         }
3097         domain->priv = dmar_domain;
3098
3099         return 0;
3100 }
3101
3102 static void intel_iommu_domain_destroy(struct iommu_domain *domain)
3103 {
3104         struct dmar_domain *dmar_domain = domain->priv;
3105
3106         domain->priv = NULL;
3107         vm_domain_exit(dmar_domain);
3108 }
3109
3110 static int intel_iommu_attach_device(struct iommu_domain *domain,
3111                                      struct device *dev)
3112 {
3113         struct dmar_domain *dmar_domain = domain->priv;
3114         struct pci_dev *pdev = to_pci_dev(dev);
3115         struct intel_iommu *iommu;
3116         int addr_width;
3117         u64 end;
3118         int ret;
3119
3120         /* normally pdev is not mapped */
3121         if (unlikely(domain_context_mapped(pdev))) {
3122                 struct dmar_domain *old_domain;
3123
3124                 old_domain = find_domain(pdev);
3125                 if (old_domain) {
3126                         if (dmar_domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE)
3127                                 vm_domain_remove_one_dev_info(old_domain, pdev);
3128                         else
3129                                 domain_remove_dev_info(old_domain);
3130                 }
3131         }
3132
3133         iommu = device_to_iommu(pci_domain_nr(pdev->bus), pdev->bus->number,
3134                                 pdev->devfn);
3135         if (!iommu)
3136                 return -ENODEV;
3137
3138         /* check if this iommu agaw is sufficient for max mapped address */
3139         addr_width = agaw_to_width(iommu->agaw);
3140         end = DOMAIN_MAX_ADDR(addr_width);
3141         end = end & VTD_PAGE_MASK;
3142         if (end < dmar_domain->max_addr) {
3143                 printk(KERN_ERR "%s: iommu agaw (%d) is not "
3144                        "sufficient for the mapped address (%llx)\n",
3145                        __func__, iommu->agaw, dmar_domain->max_addr);
3146                 return -EFAULT;
3147         }
3148
3149         ret = domain_context_mapping(dmar_domain, pdev);
3150         if (ret)
3151                 return ret;
3152
3153         ret = vm_domain_add_dev_info(dmar_domain, pdev);
3154         return ret;
3155 }
3156
3157 static void intel_iommu_detach_device(struct iommu_domain *domain,
3158                                       struct device *dev)
3159 {
3160         struct dmar_domain *dmar_domain = domain->priv;
3161         struct pci_dev *pdev = to_pci_dev(dev);
3162
3163         vm_domain_remove_one_dev_info(dmar_domain, pdev);
3164 }
3165
3166 static int intel_iommu_map_range(struct iommu_domain *domain,
3167                                  unsigned long iova, phys_addr_t hpa,
3168                                  size_t size, int iommu_prot)
3169 {
3170         struct dmar_domain *dmar_domain = domain->priv;
3171         u64 max_addr;
3172         int addr_width;
3173         int prot = 0;
3174         int ret;
3175
3176         if (iommu_prot & IOMMU_READ)
3177                 prot |= DMA_PTE_READ;
3178         if (iommu_prot & IOMMU_WRITE)
3179                 prot |= DMA_PTE_WRITE;
3180         if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
3181                 prot |= DMA_PTE_SNP;
3182
3183         max_addr = (iova & VTD_PAGE_MASK) + VTD_PAGE_ALIGN(size);
3184         if (dmar_domain->max_addr < max_addr) {
3185                 int min_agaw;
3186                 u64 end;
3187
3188                 /* check if minimum agaw is sufficient for mapped address */
3189                 min_agaw = vm_domain_min_agaw(dmar_domain);
3190                 addr_width = agaw_to_width(min_agaw);
3191                 end = DOMAIN_MAX_ADDR(addr_width);
3192                 end = end & VTD_PAGE_MASK;
3193                 if (end < max_addr) {
3194                         printk(KERN_ERR "%s: iommu agaw (%d) is not "
3195                                "sufficient for the mapped address (%llx)\n",
3196                                __func__, min_agaw, max_addr);
3197                         return -EFAULT;
3198                 }
3199                 dmar_domain->max_addr = max_addr;
3200         }
3201
3202         ret = domain_page_mapping(dmar_domain, iova, hpa, size, prot);
3203         return ret;
3204 }
3205
3206 static void intel_iommu_unmap_range(struct iommu_domain *domain,
3207                                     unsigned long iova, size_t size)
3208 {
3209         struct dmar_domain *dmar_domain = domain->priv;
3210         dma_addr_t base;
3211
3212         /* The address might not be aligned */
3213         base = iova & VTD_PAGE_MASK;
3214         size = VTD_PAGE_ALIGN(size);
3215         dma_pte_clear_range(dmar_domain, base, base + size);
3216
3217         if (dmar_domain->max_addr == base + size)
3218                 dmar_domain->max_addr = base;
3219 }
3220
3221 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
3222                                             unsigned long iova)
3223 {
3224         struct dmar_domain *dmar_domain = domain->priv;
3225         struct dma_pte *pte;
3226         u64 phys = 0;
3227
3228         pte = addr_to_dma_pte(dmar_domain, iova);
3229         if (pte)
3230                 phys = dma_pte_addr(pte);
3231
3232         return phys;
3233 }
3234
3235 static int intel_iommu_domain_has_cap(struct iommu_domain *domain,
3236                                       unsigned long cap)
3237 {
3238         struct dmar_domain *dmar_domain = domain->priv;
3239
3240         if (cap == IOMMU_CAP_CACHE_COHERENCY)
3241                 return dmar_domain->iommu_snooping;
3242
3243         return 0;
3244 }
3245
3246 static struct iommu_ops intel_iommu_ops = {
3247         .domain_init    = intel_iommu_domain_init,
3248         .domain_destroy = intel_iommu_domain_destroy,
3249         .attach_dev     = intel_iommu_attach_device,
3250         .detach_dev     = intel_iommu_detach_device,
3251         .map            = intel_iommu_map_range,
3252         .unmap          = intel_iommu_unmap_range,
3253         .iova_to_phys   = intel_iommu_iova_to_phys,
3254         .domain_has_cap = intel_iommu_domain_has_cap,
3255 };
3256
3257 static void __devinit quirk_iommu_rwbf(struct pci_dev *dev)
3258 {
3259         /*
3260          * Mobile 4 Series Chipset neglects to set RWBF capability,
3261          * but needs it:
3262          */
3263         printk(KERN_INFO "DMAR: Forcing write-buffer flush capability\n");
3264         rwbf_quirk = 1;
3265 }
3266
3267 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);