arch/x86/mm/pat.c

   1 /*
   2  * Handle caching attributes in page tables (PAT)
   3  *
   4  * Authors: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
   5  *          Suresh B Siddha <suresh.b.siddha@intel.com>
   6  *
   7  * Loosely based on earlier PAT patchset from Eric Biederman and Andi Kleen.
   8  */
   9
  10 #include <linux/mm.h>
  11 #include <linux/kernel.h>
  12 #include <linux/gfp.h>
  13 #include <linux/fs.h>
  14 #include <linux/bootmem.h>
  15
  16 #include <asm/msr.h>
  17 #include <asm/tlbflush.h>
  18 #include <asm/processor.h>
  19 #include <asm/page.h>
  20 #include <asm/pgtable.h>
  21 #include <asm/pat.h>
  22 #include <asm/e820.h>
  23 #include <asm/cacheflush.h>
  24 #include <asm/fcntl.h>
  25 #include <asm/mtrr.h>
  26 #include <asm/io.h>
  27
  28 int pat_wc_enabled = 1;
  29
  30 static u64 __read_mostly boot_pat_state;
  31
  32 static int nopat(char *str)
  33 {
  34         pat_wc_enabled = 0;
  35         printk(KERN_INFO "x86: PAT support disabled.\n");
  36
  37         return 0;
  38 }
  39 early_param("nopat", nopat);
  40
  41 static int pat_known_cpu(void)
  42 {
  43         if (!pat_wc_enabled)
  44                 return 0;
  45
  46         if (cpu_has_pat)
  47                 return 1;
  48
  49         pat_wc_enabled = 0;
  50         printk(KERN_INFO "CPU and/or kernel does not support PAT.\n");
  51         return 0;
  52 }
  53
  54 enum {
  55         PAT_UC = 0,             /* uncached */
  56         PAT_WC = 1,             /* Write combining */
  57         PAT_WT = 4,             /* Write Through */
  58         PAT_WP = 5,             /* Write Protected */
  59         PAT_WB = 6,             /* Write Back (default) */
  60         PAT_UC_MINUS = 7,       /* UC, but can be overriden by MTRR */
  61 };
  62
  63 #define PAT(x,y)        ((u64)PAT_ ## y << ((x)*8))
  64
  65 void pat_init(void)
  66 {
  67         u64 pat;
  68
  69 #ifndef CONFIG_X86_PAT
  70         nopat(NULL);
  71 #endif
  72
  73         /* Boot CPU enables PAT based on CPU feature */
  74         if (!smp_processor_id() && !pat_known_cpu())
  75                 return;
  76
  77         /* APs enable PAT iff boot CPU has enabled it before */
  78         if (smp_processor_id() && !pat_wc_enabled)
  79                 return;
  80
  81         /* Set PWT to Write-Combining. All other bits stay the same */
  82         /*
  83          * PTE encoding used in Linux:
  84          *      PAT
  85          *      |PCD
  86          *      ||PWT
  87          *      |||
  88          *      000 WB          _PAGE_CACHE_WB
  89          *      001 WC          _PAGE_CACHE_WC
  90          *      010 UC-         _PAGE_CACHE_UC_MINUS
  91          *      011 UC          _PAGE_CACHE_UC
  92          * PAT bit unused
  93          */
  94         pat = PAT(0,WB) | PAT(1,WC) | PAT(2,UC_MINUS) | PAT(3,UC) |
  95               PAT(4,WB) | PAT(5,WC) | PAT(6,UC_MINUS) | PAT(7,UC);
  96
  97         /* Boot CPU check */
  98         if (!smp_processor_id()) {
  99                 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
 100         }
 101
 102         wrmsrl(MSR_IA32_CR_PAT, pat);
 103         printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
 104                smp_processor_id(), boot_pat_state, pat);
 105 }
 106
 107 #undef PAT
 108
 109 static char *cattr_name(unsigned long flags)
 110 {
 111         switch (flags & _PAGE_CACHE_MASK) {
 112                 case _PAGE_CACHE_UC:            return "uncached";
 113                 case _PAGE_CACHE_UC_MINUS:      return "uncached-minus";
 114                 case _PAGE_CACHE_WB:            return "write-back";
 115                 case _PAGE_CACHE_WC:            return "write-combining";
 116                 default:                        return "broken";
 117         }
 118 }
 119
 120 /*
 121  * The global memtype list keeps track of memory type for specific
 122  * physical memory areas. Conflicting memory types in different
 123  * mappings can cause CPU cache corruption. To avoid this we keep track.
 124  *
 125  * The list is sorted based on starting address and can contain multiple
 126  * entries for each address (this allows reference counting for overlapping
 127  * areas). All the aliases have the same cache attributes of course.
 128  * Zero attributes are represented as holes.
 129  *
 130  * Currently the data structure is a list because the number of mappings
 131  * are expected to be relatively small. If this should be a problem
 132  * it could be changed to a rbtree or similar.
 133  *
 134  * memtype_lock protects the whole list.
 135  */
 136
 137 struct memtype {
 138         u64 start;
 139         u64 end;
 140         unsigned long type;
 141         struct list_head nd;
 142 };
 143
 144 static LIST_HEAD(memtype_list);
 145 static DEFINE_SPINLOCK(memtype_lock);   /* protects memtype list */
 146
 147 /*
 148  * Does intersection of PAT memory type and MTRR memory type and returns
 149  * the resulting memory type as PAT understands it.
 150  * (Type in pat and mtrr will not have same value)
 151  * The intersection is based on "Effective Memory Type" tables in IA-32
 152  * SDM vol 3a
 153  */
 154 static int pat_x_mtrr_type(u64 start, u64 end, unsigned long prot,
 155                                 unsigned long *ret_prot)
 156 {
 157         unsigned long pat_type;
 158         u8 mtrr_type;
 159
 160         mtrr_type = mtrr_type_lookup(start, end);
 161         if (mtrr_type == 0xFF) {                /* MTRR not enabled */
 162                 *ret_prot = prot;
 163                 return 0;
 164         }
 165         if (mtrr_type == 0xFE) {                /* MTRR match error */
 166                 *ret_prot = _PAGE_CACHE_UC;
 167                 return -1;
 168         }
 169         if (mtrr_type != MTRR_TYPE_UNCACHABLE &&
 170             mtrr_type != MTRR_TYPE_WRBACK &&
 171             mtrr_type != MTRR_TYPE_WRCOMB) {    /* MTRR type unhandled */
 172                 *ret_prot = _PAGE_CACHE_UC;
 173                 return -1;
 174         }
 175
 176         pat_type = prot & _PAGE_CACHE_MASK;
 177         prot &= (~_PAGE_CACHE_MASK);
 178
 179         /* Currently doing intersection by hand. Optimize it later. */
 180         if (pat_type == _PAGE_CACHE_WC) {
 181                 *ret_prot = prot | _PAGE_CACHE_WC;
 182         } else if (pat_type == _PAGE_CACHE_UC_MINUS) {
 183                 *ret_prot = prot | _PAGE_CACHE_UC_MINUS;
 184         } else if (pat_type == _PAGE_CACHE_UC ||
 185                    mtrr_type == MTRR_TYPE_UNCACHABLE) {
 186                 *ret_prot = prot | _PAGE_CACHE_UC;
 187         } else if (mtrr_type == MTRR_TYPE_WRCOMB) {
 188                 *ret_prot = prot | _PAGE_CACHE_WC;
 189         } else {
 190                 *ret_prot = prot | _PAGE_CACHE_WB;
 191         }
 192
 193         return 0;
 194 }
 195
 196 /*
 197  * req_type typically has one of the:
 198  * - _PAGE_CACHE_WB
 199  * - _PAGE_CACHE_WC
 200  * - _PAGE_CACHE_UC_MINUS
 201  * - _PAGE_CACHE_UC
 202  *
 203  * req_type will have a special case value '-1', when requester want to inherit
 204  * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
 205  *
 206  * If ret_type is NULL, function will return an error if it cannot reserve the
 207  * region with req_type. If ret_type is non-null, function will return
 208  * available type in ret_type in case of no error. In case of any error
 209  * it will return a negative return value.
 210  */
 211 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
 212                         unsigned long *ret_type)
 213 {
 214         struct memtype *new_entry = NULL;
 215         struct memtype *parse;
 216         unsigned long actual_type;
 217         int err = 0;
 218
 219         /* Only track when pat_wc_enabled */
 220         if (!pat_wc_enabled) {
 221                 /* This is identical to page table setting without PAT */
 222                 if (ret_type) {
 223                         if (req_type == -1) {
 224                                 *ret_type = _PAGE_CACHE_WB;
 225                         } else {
 226                                 *ret_type = req_type;
 227                         }
 228                 }
 229                 return 0;
 230         }
 231
 232         /* Low ISA region is always mapped WB in page table. No need to track */
 233         if (start >= ISA_START_ADDRESS && (end - 1) <= ISA_END_ADDRESS) {
 234                 if (ret_type)
 235                         *ret_type = _PAGE_CACHE_WB;
 236
 237                 return 0;
 238         }
 239
 240         if (req_type == -1) {
 241                 /*
 242                  * Special case where caller wants to inherit from mtrr or
 243                  * existing pat mapping, defaulting to UC_MINUS in case of
 244                  * no match.
 245                  */
 246                 u8 mtrr_type = mtrr_type_lookup(start, end);
 247                 if (mtrr_type == 0xFE) { /* MTRR match error */
 248                         err = -1;
 249                 }
 250
 251                 if (mtrr_type == MTRR_TYPE_WRBACK) {
 252                         req_type = _PAGE_CACHE_WB;
 253                         actual_type = _PAGE_CACHE_WB;
 254                 } else {
 255                         req_type = _PAGE_CACHE_UC_MINUS;
 256                         actual_type = _PAGE_CACHE_UC_MINUS;
 257                 }
 258         } else {
 259                 req_type &= _PAGE_CACHE_MASK;
 260                 err = pat_x_mtrr_type(start, end, req_type, &actual_type);
 261         }
 262
 263         if (err) {
 264                 if (ret_type)
 265                         *ret_type = actual_type;
 266
 267                 return -EINVAL;
 268         }
 269
 270         new_entry  = kmalloc(sizeof(struct memtype), GFP_KERNEL);
 271         if (!new_entry)
 272                 return -ENOMEM;
 273
 274         new_entry->start = start;
 275         new_entry->end = end;
 276         new_entry->type = actual_type;
 277
 278         if (ret_type)
 279                 *ret_type = actual_type;
 280
 281         spin_lock(&memtype_lock);
 282
 283         /* Search for existing mapping that overlaps the current range */
 284         list_for_each_entry(parse, &memtype_list, nd) {
 285                 struct memtype *saved_ptr;
 286
 287                 if (parse->start >= end) {
 288                         pr_debug("New Entry\n");
 289                         list_add(&new_entry->nd, parse->nd.prev);
 290                         new_entry = NULL;
 291                         break;
 292                 }
 293
 294                 if (start <= parse->start && end >= parse->start) {
 295                         if (actual_type != parse->type && ret_type) {
 296                                 actual_type = parse->type;
 297                                 *ret_type = actual_type;
 298                                 new_entry->type = actual_type;
 299                         }
 300
 301                         if (actual_type != parse->type) {
 302                                 printk(
 303                 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
 304                                         current->comm, current->pid,
 305                                         start, end,
 306                                         cattr_name(actual_type),
 307                                         cattr_name(parse->type));
 308                                 err = -EBUSY;
 309                                 break;
 310                         }
 311
 312                         saved_ptr = parse;
 313                         /*
 314                          * Check to see whether the request overlaps more
 315                          * than one entry in the list
 316                          */
 317                         list_for_each_entry_continue(parse, &memtype_list, nd) {
 318                                 if (end <= parse->start) {
 319                                         break;
 320                                 }
 321
 322                                 if (actual_type != parse->type) {
 323                                         printk(
 324                 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
 325                                                 current->comm, current->pid,
 326                                                 start, end,
 327                                                 cattr_name(actual_type),
 328                                                 cattr_name(parse->type));
 329                                         err = -EBUSY;
 330                                         break;
 331                                 }
 332                         }
 333
 334                         if (err) {
 335                                 break;
 336                         }
 337
 338                         pr_debug("Overlap at 0x%Lx-0x%Lx\n",
 339                                saved_ptr->start, saved_ptr->end);
 340                         /* No conflict. Go ahead and add this new entry */
 341                         list_add(&new_entry->nd, saved_ptr->nd.prev);
 342                         new_entry = NULL;
 343                         break;
 344                 }
 345
 346                 if (start < parse->end) {
 347                         if (actual_type != parse->type && ret_type) {
 348                                 actual_type = parse->type;
 349                                 *ret_type = actual_type;
 350                                 new_entry->type = actual_type;
 351                         }
 352
 353                         if (actual_type != parse->type) {
 354                                 printk(
 355                 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
 356                                         current->comm, current->pid,
 357                                         start, end,
 358                                         cattr_name(actual_type),
 359                                         cattr_name(parse->type));
 360                                 err = -EBUSY;
 361                                 break;
 362                         }
 363
 364                         saved_ptr = parse;
 365                         /*
 366                          * Check to see whether the request overlaps more
 367                          * than one entry in the list
 368                          */
 369                         list_for_each_entry_continue(parse, &memtype_list, nd) {
 370                                 if (end <= parse->start) {
 371                                         break;
 372                                 }
 373
 374                                 if (actual_type != parse->type) {
 375                                         printk(
 376                 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
 377                                                 current->comm, current->pid,
 378                                                 start, end,
 379                                                 cattr_name(actual_type),
 380                                                 cattr_name(parse->type));
 381                                         err = -EBUSY;
 382                                         break;
 383                                 }
 384                         }
 385
 386                         if (err) {
 387                                 break;
 388                         }
 389
 390                         printk(KERN_INFO "Overlap at 0x%Lx-0x%Lx\n",
 391                                saved_ptr->start, saved_ptr->end);
 392                         /* No conflict. Go ahead and add this new entry */
 393                         list_add(&new_entry->nd, &saved_ptr->nd);
 394                         new_entry = NULL;
 395                         break;
 396                 }
 397         }
 398
 399         if (err) {
 400                 printk(KERN_INFO
 401         "reserve_memtype failed 0x%Lx-0x%Lx, track %s, req %s\n",
 402                         start, end, cattr_name(new_entry->type),
 403                         cattr_name(req_type));
 404                 kfree(new_entry);
 405                 spin_unlock(&memtype_lock);
 406                 return err;
 407         }
 408
 409         if (new_entry) {
 410                 /* No conflict. Not yet added to the list. Add to the tail */
 411                 list_add_tail(&new_entry->nd, &memtype_list);
 412                 pr_debug("New Entry\n");
 413         }
 414
 415         if (ret_type) {
 416                 pr_debug(
 417         "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
 418                         start, end, cattr_name(actual_type),
 419                         cattr_name(req_type), cattr_name(*ret_type));
 420         } else {
 421                 pr_debug(
 422         "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s\n",
 423                         start, end, cattr_name(actual_type),
 424                         cattr_name(req_type));
 425         }
 426
 427         spin_unlock(&memtype_lock);
 428         return err;
 429 }
 430
 431 int free_memtype(u64 start, u64 end)
 432 {
 433         struct memtype *ml;
 434         int err = -EINVAL;
 435
 436         /* Only track when pat_wc_enabled */
 437         if (!pat_wc_enabled) {
 438                 return 0;
 439         }
 440
 441         /* Low ISA region is always mapped WB. No need to track */
 442         if (start >= ISA_START_ADDRESS && end <= ISA_END_ADDRESS) {
 443                 return 0;
 444         }
 445
 446         spin_lock(&memtype_lock);
 447         list_for_each_entry(ml, &memtype_list, nd) {
 448                 if (ml->start == start && ml->end == end) {
 449                         list_del(&ml->nd);
 450                         kfree(ml);
 451                         err = 0;
 452                         break;
 453                 }
 454         }
 455         spin_unlock(&memtype_lock);
 456
 457         if (err) {
 458                 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
 459                         current->comm, current->pid, start, end);
 460         }
 461
 462         pr_debug("free_memtype request 0x%Lx-0x%Lx\n", start, end);
 463         return err;
 464 }
 465
 466
 467 /*
 468  * /dev/mem mmap interface. The memtype used for mapping varies:
 469  * - Use UC for mappings with O_SYNC flag
 470  * - Without O_SYNC flag, if there is any conflict in reserve_memtype,
 471  *   inherit the memtype from existing mapping.
 472  * - Else use UC_MINUS memtype (for backward compatibility with existing
 473  *   X drivers.
 474  */
 475 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
 476                                 unsigned long size, pgprot_t vma_prot)
 477 {
 478         return vma_prot;
 479 }
 480
 481 #ifdef CONFIG_NONPROMISC_DEVMEM
 482 /* This check is done in drivers/char/mem.c in case of NONPROMISC_DEVMEM*/
 483 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
 484 {
 485         return 1;
 486 }
 487 #else
 488 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
 489 {
 490         u64 from = ((u64)pfn) << PAGE_SHIFT;
 491         u64 to = from + size;
 492         u64 cursor = from;
 493
 494         while (cursor < to) {
 495                 if (!devmem_is_allowed(pfn)) {
 496                         printk(KERN_INFO
 497                 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
 498                                 current->comm, from, to);
 499                         return 0;
 500                 }
 501                 cursor += PAGE_SIZE;
 502                 pfn++;
 503         }
 504         return 1;
 505 }
 506 #endif /* CONFIG_NONPROMISC_DEVMEM */
 507
 508 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
 509                                 unsigned long size, pgprot_t *vma_prot)
 510 {
 511         u64 offset = ((u64) pfn) << PAGE_SHIFT;
 512         unsigned long flags = _PAGE_CACHE_UC_MINUS;
 513         unsigned long ret_flags;
 514         int retval;
 515
 516         if (!range_is_allowed(pfn, size))
 517                 return 0;
 518
 519         if (file->f_flags & O_SYNC) {
 520                 flags = _PAGE_CACHE_UC;
 521         }
 522
 523 #ifdef CONFIG_X86_32
 524         /*
 525          * On the PPro and successors, the MTRRs are used to set
 526          * memory types for physical addresses outside main memory,
 527          * so blindly setting UC or PWT on those pages is wrong.
 528          * For Pentiums and earlier, the surround logic should disable
 529          * caching for the high addresses through the KEN pin, but
 530          * we maintain the tradition of paranoia in this code.
 531          */
 532         if (!pat_wc_enabled &&
 533             ! ( test_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability) ||
 534                 test_bit(X86_FEATURE_K6_MTRR, boot_cpu_data.x86_capability) ||
 535                 test_bit(X86_FEATURE_CYRIX_ARR, boot_cpu_data.x86_capability) ||
 536                 test_bit(X86_FEATURE_CENTAUR_MCR, boot_cpu_data.x86_capability)) &&
 537            (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
 538                 flags = _PAGE_CACHE_UC;
 539         }
 540 #endif
 541
 542         /*
 543          * With O_SYNC, we can only take UC mapping. Fail if we cannot.
 544          * Without O_SYNC, we want to get
 545          * - WB for WB-able memory and no other conflicting mappings
 546          * - UC_MINUS for non-WB-able memory with no other conflicting mappings
 547          * - Inherit from confliting mappings otherwise
 548          */
 549         if (flags != _PAGE_CACHE_UC_MINUS) {
 550                 retval = reserve_memtype(offset, offset + size, flags, NULL);
 551         } else {
 552                 retval = reserve_memtype(offset, offset + size, -1, &ret_flags);
 553         }
 554
 555         if (retval < 0)
 556                 return 0;
 557
 558         flags = ret_flags;
 559
 560         if (pfn <= max_pfn_mapped &&
 561             ioremap_change_attr((unsigned long)__va(offset), size, flags) < 0) {
 562                 free_memtype(offset, offset + size);
 563                 printk(KERN_INFO
 564                 "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
 565                         current->comm, current->pid,
 566                         cattr_name(flags),
 567                         offset, offset + size);
 568                 return 0;
 569         }
 570
 571         *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
 572                              flags);
 573         return 1;
 574 }
 575
 576 void map_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
 577 {
 578         u64 addr = (u64)pfn << PAGE_SHIFT;
 579         unsigned long flags;
 580         unsigned long want_flags = (pgprot_val(vma_prot) & _PAGE_CACHE_MASK);
 581
 582         reserve_memtype(addr, addr + size, want_flags, &flags);
 583         if (flags != want_flags) {
 584                 printk(KERN_INFO
 585                 "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
 586                         current->comm, current->pid,
 587                         cattr_name(want_flags),
 588                         addr, addr + size,
 589                         cattr_name(flags));
 590         }
 591 }
 592
 593 void unmap_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
 594 {
 595         u64 addr = (u64)pfn << PAGE_SHIFT;
 596
 597         free_memtype(addr, addr + size);
 598 }
 599