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 #ifdef CONFIG_X86_PAT
  29 int __read_mostly pat_enabled = 1;
  30
  31 void __cpuinit pat_disable(char *reason)
  32 {
  33         pat_enabled = 0;
  34         printk(KERN_INFO "%s\n", reason);
  35 }
  36
  37 static int __init nopat(char *str)
  38 {
  39         pat_disable("PAT support disabled.");
  40         return 0;
  41 }
  42 early_param("nopat", nopat);
  43 #endif
  44
  45
  46 static int debug_enable;
  47 static int __init pat_debug_setup(char *str)
  48 {
  49         debug_enable = 1;
  50         return 0;
  51 }
  52 __setup("debugpat", pat_debug_setup);
  53
  54 #define dprintk(fmt, arg...) \
  55         do { if (debug_enable) printk(KERN_INFO fmt, ##arg); } while (0)
  56
  57
  58 static u64 __read_mostly boot_pat_state;
  59
  60 enum {
  61         PAT_UC = 0,             /* uncached */
  62         PAT_WC = 1,             /* Write combining */
  63         PAT_WT = 4,             /* Write Through */
  64         PAT_WP = 5,             /* Write Protected */
  65         PAT_WB = 6,             /* Write Back (default) */
  66         PAT_UC_MINUS = 7,       /* UC, but can be overriden by MTRR */
  67 };
  68
  69 #define PAT(x, y)       ((u64)PAT_ ## y << ((x)*8))
  70
  71 void pat_init(void)
  72 {
  73         u64 pat;
  74
  75         if (!pat_enabled)
  76                 return;
  77
  78         /* Paranoia check. */
  79         if (!cpu_has_pat && boot_pat_state) {
  80                 /*
  81                  * If this happens we are on a secondary CPU, but
  82                  * switched to PAT on the boot CPU. We have no way to
  83                  * undo PAT.
  84                  */
  85                 printk(KERN_ERR "PAT enabled, "
  86                        "but not supported by secondary CPU\n");
  87                 BUG();
  88         }
  89
  90         /* Set PWT to Write-Combining. All other bits stay the same */
  91         /*
  92          * PTE encoding used in Linux:
  93          *      PAT
  94          *      |PCD
  95          *      ||PWT
  96          *      |||
  97          *      000 WB          _PAGE_CACHE_WB
  98          *      001 WC          _PAGE_CACHE_WC
  99          *      010 UC-         _PAGE_CACHE_UC_MINUS
 100          *      011 UC          _PAGE_CACHE_UC
 101          * PAT bit unused
 102          */
 103         pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) |
 104               PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC);
 105
 106         /* Boot CPU check */
 107         if (!boot_pat_state)
 108                 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
 109
 110         wrmsrl(MSR_IA32_CR_PAT, pat);
 111         printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
 112                smp_processor_id(), boot_pat_state, pat);
 113 }
 114
 115 #undef PAT
 116
 117 static char *cattr_name(unsigned long flags)
 118 {
 119         switch (flags & _PAGE_CACHE_MASK) {
 120         case _PAGE_CACHE_UC:            return "uncached";
 121         case _PAGE_CACHE_UC_MINUS:      return "uncached-minus";
 122         case _PAGE_CACHE_WB:            return "write-back";
 123         case _PAGE_CACHE_WC:            return "write-combining";
 124         default:                        return "broken";
 125         }
 126 }
 127
 128 /*
 129  * The global memtype list keeps track of memory type for specific
 130  * physical memory areas. Conflicting memory types in different
 131  * mappings can cause CPU cache corruption. To avoid this we keep track.
 132  *
 133  * The list is sorted based on starting address and can contain multiple
 134  * entries for each address (this allows reference counting for overlapping
 135  * areas). All the aliases have the same cache attributes of course.
 136  * Zero attributes are represented as holes.
 137  *
 138  * Currently the data structure is a list because the number of mappings
 139  * are expected to be relatively small. If this should be a problem
 140  * it could be changed to a rbtree or similar.
 141  *
 142  * memtype_lock protects the whole list.
 143  */
 144
 145 struct memtype {
 146         u64 start;
 147         u64 end;
 148         unsigned long type;
 149         struct list_head nd;
 150 };
 151
 152 static LIST_HEAD(memtype_list);
 153 static DEFINE_SPINLOCK(memtype_lock);   /* protects memtype list */
 154
 155 /*
 156  * Does intersection of PAT memory type and MTRR memory type and returns
 157  * the resulting memory type as PAT understands it.
 158  * (Type in pat and mtrr will not have same value)
 159  * The intersection is based on "Effective Memory Type" tables in IA-32
 160  * SDM vol 3a
 161  */
 162 static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
 163 {
 164         /*
 165          * Look for MTRR hint to get the effective type in case where PAT
 166          * request is for WB.
 167          */
 168         if (req_type == _PAGE_CACHE_WB) {
 169                 u8 mtrr_type;
 170
 171                 mtrr_type = mtrr_type_lookup(start, end);
 172                 if (mtrr_type == MTRR_TYPE_UNCACHABLE)
 173                         return _PAGE_CACHE_UC;
 174                 if (mtrr_type == MTRR_TYPE_WRCOMB)
 175                         return _PAGE_CACHE_WC;
 176         }
 177
 178         return req_type;
 179 }
 180
 181 static int chk_conflict(struct memtype *new, struct memtype *entry,
 182                         unsigned long *type)
 183 {
 184         if (new->type != entry->type) {
 185                 if (type) {
 186                         new->type = entry->type;
 187                         *type = entry->type;
 188                 } else
 189                         goto conflict;
 190         }
 191
 192          /* check overlaps with more than one entry in the list */
 193         list_for_each_entry_continue(entry, &memtype_list, nd) {
 194                 if (new->end <= entry->start)
 195                         break;
 196                 else if (new->type != entry->type)
 197                         goto conflict;
 198         }
 199         return 0;
 200
 201  conflict:
 202         printk(KERN_INFO "%s:%d conflicting memory types "
 203                "%Lx-%Lx %s<->%s\n", current->comm, current->pid, new->start,
 204                new->end, cattr_name(new->type), cattr_name(entry->type));
 205         return -EBUSY;
 206 }
 207
 208 /*
 209  * req_type typically has one of the:
 210  * - _PAGE_CACHE_WB
 211  * - _PAGE_CACHE_WC
 212  * - _PAGE_CACHE_UC_MINUS
 213  * - _PAGE_CACHE_UC
 214  *
 215  * req_type will have a special case value '-1', when requester want to inherit
 216  * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
 217  *
 218  * If new_type is NULL, function will return an error if it cannot reserve the
 219  * region with req_type. If new_type is non-NULL, function will return
 220  * available type in new_type in case of no error. In case of any error
 221  * it will return a negative return value.
 222  */
 223 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
 224                         unsigned long *new_type)
 225 {
 226         struct memtype *new, *entry;
 227         unsigned long actual_type;
 228         struct list_head *where;
 229         int err = 0;
 230
 231         BUG_ON(start >= end); /* end is exclusive */
 232
 233         if (!pat_enabled) {
 234                 /* This is identical to page table setting without PAT */
 235                 if (new_type) {
 236                         if (req_type == -1)
 237                                 *new_type = _PAGE_CACHE_WB;
 238                         else
 239                                 *new_type = req_type & _PAGE_CACHE_MASK;
 240                 }
 241                 return 0;
 242         }
 243
 244         /* Low ISA region is always mapped WB in page table. No need to track */
 245         if (is_ISA_range(start, end - 1)) {
 246                 if (new_type)
 247                         *new_type = _PAGE_CACHE_WB;
 248                 return 0;
 249         }
 250
 251         if (req_type == -1) {
 252                 /*
 253                  * Call mtrr_lookup to get the type hint. This is an
 254                  * optimization for /dev/mem mmap'ers into WB memory (BIOS
 255                  * tools and ACPI tools). Use WB request for WB memory and use
 256                  * UC_MINUS otherwise.
 257                  */
 258                 u8 mtrr_type = mtrr_type_lookup(start, end);
 259
 260                 if (mtrr_type == MTRR_TYPE_WRBACK)
 261                         actual_type = _PAGE_CACHE_WB;
 262                 else
 263                         actual_type = _PAGE_CACHE_UC_MINUS;
 264         } else
 265                 actual_type = pat_x_mtrr_type(start, end,
 266                                               req_type & _PAGE_CACHE_MASK);
 267
 268         new  = kmalloc(sizeof(struct memtype), GFP_KERNEL);
 269         if (!new)
 270                 return -ENOMEM;
 271
 272         new->start = start;
 273         new->end = end;
 274         new->type = actual_type;
 275
 276         if (new_type)
 277                 *new_type = actual_type;
 278
 279         spin_lock(&memtype_lock);
 280
 281         /* Search for existing mapping that overlaps the current range */
 282         where = NULL;
 283         list_for_each_entry(entry, &memtype_list, nd) {
 284                 if (end <= entry->start) {
 285                         where = entry->nd.prev;
 286                         break;
 287                 } else if (start <= entry->start) { /* end > entry->start */
 288                         err = chk_conflict(new, entry, new_type);
 289                         if (!err) {
 290                                 dprintk("Overlap at 0x%Lx-0x%Lx\n",
 291                                         entry->start, entry->end);
 292                                 where = entry->nd.prev;
 293                         }
 294                         break;
 295                 } else if (start < entry->end) { /* start > entry->start */
 296                         err = chk_conflict(new, entry, new_type);
 297                         if (!err) {
 298                                 dprintk("Overlap at 0x%Lx-0x%Lx\n",
 299                                         entry->start, entry->end);
 300                                 where = &entry->nd;
 301                         }
 302                         break;
 303                 }
 304         }
 305
 306         if (err) {
 307                 printk(KERN_INFO "reserve_memtype failed 0x%Lx-0x%Lx, "
 308                        "track %s, req %s\n",
 309                        start, end, cattr_name(new->type), cattr_name(req_type));
 310                 kfree(new);
 311                 spin_unlock(&memtype_lock);
 312                 return err;
 313         }
 314
 315         if (where)
 316                 list_add(&new->nd, where);
 317         else
 318                 list_add_tail(&new->nd, &memtype_list);
 319
 320         spin_unlock(&memtype_lock);
 321
 322         dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
 323                 start, end, cattr_name(new->type), cattr_name(req_type),
 324                 new_type ? cattr_name(*new_type) : "-");
 325
 326         return err;
 327 }
 328
 329 int free_memtype(u64 start, u64 end)
 330 {
 331         struct memtype *entry;
 332         int err = -EINVAL;
 333
 334         if (!pat_enabled)
 335                 return 0;
 336
 337         /* Low ISA region is always mapped WB. No need to track */
 338         if (is_ISA_range(start, end - 1))
 339                 return 0;
 340
 341         spin_lock(&memtype_lock);
 342         list_for_each_entry(entry, &memtype_list, nd) {
 343                 if (entry->start == start && entry->end == end) {
 344                         list_del(&entry->nd);
 345                         kfree(entry);
 346                         err = 0;
 347                         break;
 348                 }
 349         }
 350         spin_unlock(&memtype_lock);
 351
 352         if (err) {
 353                 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
 354                         current->comm, current->pid, start, end);
 355         }
 356
 357         dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
 358         return err;
 359 }
 360
 361
 362 /*
 363  * /dev/mem mmap interface. The memtype used for mapping varies:
 364  * - Use UC for mappings with O_SYNC flag
 365  * - Without O_SYNC flag, if there is any conflict in reserve_memtype,
 366  *   inherit the memtype from existing mapping.
 367  * - Else use UC_MINUS memtype (for backward compatibility with existing
 368  *   X drivers.
 369  */
 370 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
 371                                 unsigned long size, pgprot_t vma_prot)
 372 {
 373         return vma_prot;
 374 }
 375
 376 #ifdef CONFIG_NONPROMISC_DEVMEM
 377 /* This check is done in drivers/char/mem.c in case of NONPROMISC_DEVMEM*/
 378 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
 379 {
 380         return 1;
 381 }
 382 #else
 383 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
 384 {
 385         u64 from = ((u64)pfn) << PAGE_SHIFT;
 386         u64 to = from + size;
 387         u64 cursor = from;
 388
 389         while (cursor < to) {
 390                 if (!devmem_is_allowed(pfn)) {
 391                         printk(KERN_INFO
 392                 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
 393                                 current->comm, from, to);
 394                         return 0;
 395                 }
 396                 cursor += PAGE_SIZE;
 397                 pfn++;
 398         }
 399         return 1;
 400 }
 401 #endif /* CONFIG_NONPROMISC_DEVMEM */
 402
 403 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
 404                                 unsigned long size, pgprot_t *vma_prot)
 405 {
 406         u64 offset = ((u64) pfn) << PAGE_SHIFT;
 407         unsigned long flags = _PAGE_CACHE_UC_MINUS;
 408         int retval;
 409
 410         if (!range_is_allowed(pfn, size))
 411                 return 0;
 412
 413         if (file->f_flags & O_SYNC) {
 414                 flags = _PAGE_CACHE_UC;
 415         }
 416
 417 #ifdef CONFIG_X86_32
 418         /*
 419          * On the PPro and successors, the MTRRs are used to set
 420          * memory types for physical addresses outside main memory,
 421          * so blindly setting UC or PWT on those pages is wrong.
 422          * For Pentiums and earlier, the surround logic should disable
 423          * caching for the high addresses through the KEN pin, but
 424          * we maintain the tradition of paranoia in this code.
 425          */
 426         if (!pat_enabled &&
 427             !(boot_cpu_has(X86_FEATURE_MTRR) ||
 428               boot_cpu_has(X86_FEATURE_K6_MTRR) ||
 429               boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
 430               boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
 431             (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
 432                 flags = _PAGE_CACHE_UC;
 433         }
 434 #endif
 435
 436         /*
 437          * With O_SYNC, we can only take UC mapping. Fail if we cannot.
 438          * Without O_SYNC, we want to get
 439          * - WB for WB-able memory and no other conflicting mappings
 440          * - UC_MINUS for non-WB-able memory with no other conflicting mappings
 441          * - Inherit from confliting mappings otherwise
 442          */
 443         if (flags != _PAGE_CACHE_UC_MINUS) {
 444                 retval = reserve_memtype(offset, offset + size, flags, NULL);
 445         } else {
 446                 retval = reserve_memtype(offset, offset + size, -1, &flags);
 447         }
 448
 449         if (retval < 0)
 450                 return 0;
 451
 452         if (((pfn < max_low_pfn_mapped) ||
 453              (pfn >= (1UL<<(32 - PAGE_SHIFT)) && pfn < max_pfn_mapped)) &&
 454             ioremap_change_attr((unsigned long)__va(offset), size, flags) < 0) {
 455                 free_memtype(offset, offset + size);
 456                 printk(KERN_INFO
 457                 "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
 458                         current->comm, current->pid,
 459                         cattr_name(flags),
 460                         offset, (unsigned long long)(offset + size));
 461                 return 0;
 462         }
 463
 464         *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
 465                              flags);
 466         return 1;
 467 }
 468
 469 void map_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
 470 {
 471         u64 addr = (u64)pfn << PAGE_SHIFT;
 472         unsigned long flags;
 473         unsigned long want_flags = (pgprot_val(vma_prot) & _PAGE_CACHE_MASK);
 474
 475         reserve_memtype(addr, addr + size, want_flags, &flags);
 476         if (flags != want_flags) {
 477                 printk(KERN_INFO
 478                 "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
 479                         current->comm, current->pid,
 480                         cattr_name(want_flags),
 481                         addr, (unsigned long long)(addr + size),
 482                         cattr_name(flags));
 483         }
 484 }
 485
 486 void unmap_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
 487 {
 488         u64 addr = (u64)pfn << PAGE_SHIFT;
 489
 490         free_memtype(addr, addr + size);
 491 }