| 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/seq_file.h> |
| 11 | #include <linux/bootmem.h> |
| 12 | #include <linux/debugfs.h> |
| 13 | #include <linux/kernel.h> |
| 14 | #include <linux/module.h> |
| 15 | #include <linux/gfp.h> |
| 16 | #include <linux/mm.h> |
| 17 | #include <linux/fs.h> |
| 18 | #include <linux/rbtree.h> |
| 19 | |
| 20 | #include <asm/cacheflush.h> |
| 21 | #include <asm/processor.h> |
| 22 | #include <asm/tlbflush.h> |
| 23 | #include <asm/pgtable.h> |
| 24 | #include <asm/fcntl.h> |
| 25 | #include <asm/e820.h> |
| 26 | #include <asm/mtrr.h> |
| 27 | #include <asm/page.h> |
| 28 | #include <asm/msr.h> |
| 29 | #include <asm/pat.h> |
| 30 | #include <asm/io.h> |
| 31 | |
| 32 | #ifdef CONFIG_X86_PAT |
| 33 | int __read_mostly pat_enabled = 1; |
| 34 | |
| 35 | static inline void pat_disable(const char *reason) |
| 36 | { |
| 37 | pat_enabled = 0; |
| 38 | printk(KERN_INFO "%s\n", reason); |
| 39 | } |
| 40 | |
| 41 | static int __init nopat(char *str) |
| 42 | { |
| 43 | pat_disable("PAT support disabled."); |
| 44 | return 0; |
| 45 | } |
| 46 | early_param("nopat", nopat); |
| 47 | #else |
| 48 | static inline void pat_disable(const char *reason) |
| 49 | { |
| 50 | (void)reason; |
| 51 | } |
| 52 | #endif |
| 53 | |
| 54 | |
| 55 | static int debug_enable; |
| 56 | |
| 57 | static int __init pat_debug_setup(char *str) |
| 58 | { |
| 59 | debug_enable = 1; |
| 60 | return 0; |
| 61 | } |
| 62 | __setup("debugpat", pat_debug_setup); |
| 63 | |
| 64 | #define dprintk(fmt, arg...) \ |
| 65 | do { if (debug_enable) printk(KERN_INFO fmt, ##arg); } while (0) |
| 66 | |
| 67 | |
| 68 | static u64 __read_mostly boot_pat_state; |
| 69 | |
| 70 | enum { |
| 71 | PAT_UC = 0, /* uncached */ |
| 72 | PAT_WC = 1, /* Write combining */ |
| 73 | PAT_WT = 4, /* Write Through */ |
| 74 | PAT_WP = 5, /* Write Protected */ |
| 75 | PAT_WB = 6, /* Write Back (default) */ |
| 76 | PAT_UC_MINUS = 7, /* UC, but can be overriden by MTRR */ |
| 77 | }; |
| 78 | |
| 79 | #define PAT(x, y) ((u64)PAT_ ## y << ((x)*8)) |
| 80 | |
| 81 | void pat_init(void) |
| 82 | { |
| 83 | u64 pat; |
| 84 | bool boot_cpu = !boot_pat_state; |
| 85 | |
| 86 | if (!pat_enabled) |
| 87 | return; |
| 88 | |
| 89 | if (!cpu_has_pat) { |
| 90 | if (!boot_pat_state) { |
| 91 | pat_disable("PAT not supported by CPU."); |
| 92 | return; |
| 93 | } else { |
| 94 | /* |
| 95 | * If this happens we are on a secondary CPU, but |
| 96 | * switched to PAT on the boot CPU. We have no way to |
| 97 | * undo PAT. |
| 98 | */ |
| 99 | printk(KERN_ERR "PAT enabled, " |
| 100 | "but not supported by secondary CPU\n"); |
| 101 | BUG(); |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | /* Set PWT to Write-Combining. All other bits stay the same */ |
| 106 | /* |
| 107 | * PTE encoding used in Linux: |
| 108 | * PAT |
| 109 | * |PCD |
| 110 | * ||PWT |
| 111 | * ||| |
| 112 | * 000 WB _PAGE_CACHE_WB |
| 113 | * 001 WC _PAGE_CACHE_WC |
| 114 | * 010 UC- _PAGE_CACHE_UC_MINUS |
| 115 | * 011 UC _PAGE_CACHE_UC |
| 116 | * PAT bit unused |
| 117 | */ |
| 118 | pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) | |
| 119 | PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC); |
| 120 | |
| 121 | /* Boot CPU check */ |
| 122 | if (!boot_pat_state) |
| 123 | rdmsrl(MSR_IA32_CR_PAT, boot_pat_state); |
| 124 | |
| 125 | wrmsrl(MSR_IA32_CR_PAT, pat); |
| 126 | |
| 127 | if (boot_cpu) |
| 128 | printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n", |
| 129 | smp_processor_id(), boot_pat_state, pat); |
| 130 | } |
| 131 | |
| 132 | #undef PAT |
| 133 | |
| 134 | static char *cattr_name(unsigned long flags) |
| 135 | { |
| 136 | switch (flags & _PAGE_CACHE_MASK) { |
| 137 | case _PAGE_CACHE_UC: return "uncached"; |
| 138 | case _PAGE_CACHE_UC_MINUS: return "uncached-minus"; |
| 139 | case _PAGE_CACHE_WB: return "write-back"; |
| 140 | case _PAGE_CACHE_WC: return "write-combining"; |
| 141 | default: return "broken"; |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | /* |
| 146 | * The global memtype list keeps track of memory type for specific |
| 147 | * physical memory areas. Conflicting memory types in different |
| 148 | * mappings can cause CPU cache corruption. To avoid this we keep track. |
| 149 | * |
| 150 | * The list is sorted based on starting address and can contain multiple |
| 151 | * entries for each address (this allows reference counting for overlapping |
| 152 | * areas). All the aliases have the same cache attributes of course. |
| 153 | * Zero attributes are represented as holes. |
| 154 | * |
| 155 | * The data structure is a list that is also organized as an rbtree |
| 156 | * sorted on the start address of memtype range. |
| 157 | * |
| 158 | * memtype_lock protects both the linear list and rbtree. |
| 159 | */ |
| 160 | |
| 161 | struct memtype { |
| 162 | u64 start; |
| 163 | u64 end; |
| 164 | unsigned long type; |
| 165 | struct list_head nd; |
| 166 | struct rb_node rb; |
| 167 | }; |
| 168 | |
| 169 | static struct rb_root memtype_rbroot = RB_ROOT; |
| 170 | static LIST_HEAD(memtype_list); |
| 171 | static DEFINE_SPINLOCK(memtype_lock); /* protects memtype list */ |
| 172 | |
| 173 | static struct memtype *memtype_rb_search(struct rb_root *root, u64 start) |
| 174 | { |
| 175 | struct rb_node *node = root->rb_node; |
| 176 | struct memtype *last_lower = NULL; |
| 177 | |
| 178 | while (node) { |
| 179 | struct memtype *data = container_of(node, struct memtype, rb); |
| 180 | |
| 181 | if (data->start < start) { |
| 182 | last_lower = data; |
| 183 | node = node->rb_right; |
| 184 | } else if (data->start > start) { |
| 185 | node = node->rb_left; |
| 186 | } else |
| 187 | return data; |
| 188 | } |
| 189 | |
| 190 | /* Will return NULL if there is no entry with its start <= start */ |
| 191 | return last_lower; |
| 192 | } |
| 193 | |
| 194 | static void memtype_rb_insert(struct rb_root *root, struct memtype *data) |
| 195 | { |
| 196 | struct rb_node **new = &(root->rb_node); |
| 197 | struct rb_node *parent = NULL; |
| 198 | |
| 199 | while (*new) { |
| 200 | struct memtype *this = container_of(*new, struct memtype, rb); |
| 201 | |
| 202 | parent = *new; |
| 203 | if (data->start <= this->start) |
| 204 | new = &((*new)->rb_left); |
| 205 | else if (data->start > this->start) |
| 206 | new = &((*new)->rb_right); |
| 207 | } |
| 208 | |
| 209 | rb_link_node(&data->rb, parent, new); |
| 210 | rb_insert_color(&data->rb, root); |
| 211 | } |
| 212 | |
| 213 | /* |
| 214 | * Does intersection of PAT memory type and MTRR memory type and returns |
| 215 | * the resulting memory type as PAT understands it. |
| 216 | * (Type in pat and mtrr will not have same value) |
| 217 | * The intersection is based on "Effective Memory Type" tables in IA-32 |
| 218 | * SDM vol 3a |
| 219 | */ |
| 220 | static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type) |
| 221 | { |
| 222 | /* |
| 223 | * Look for MTRR hint to get the effective type in case where PAT |
| 224 | * request is for WB. |
| 225 | */ |
| 226 | if (req_type == _PAGE_CACHE_WB) { |
| 227 | u8 mtrr_type; |
| 228 | |
| 229 | mtrr_type = mtrr_type_lookup(start, end); |
| 230 | if (mtrr_type != MTRR_TYPE_WRBACK) |
| 231 | return _PAGE_CACHE_UC_MINUS; |
| 232 | |
| 233 | return _PAGE_CACHE_WB; |
| 234 | } |
| 235 | |
| 236 | return req_type; |
| 237 | } |
| 238 | |
| 239 | static int |
| 240 | chk_conflict(struct memtype *new, struct memtype *entry, unsigned long *type) |
| 241 | { |
| 242 | if (new->type != entry->type) { |
| 243 | if (type) { |
| 244 | new->type = entry->type; |
| 245 | *type = entry->type; |
| 246 | } else |
| 247 | goto conflict; |
| 248 | } |
| 249 | |
| 250 | /* check overlaps with more than one entry in the list */ |
| 251 | list_for_each_entry_continue(entry, &memtype_list, nd) { |
| 252 | if (new->end <= entry->start) |
| 253 | break; |
| 254 | else if (new->type != entry->type) |
| 255 | goto conflict; |
| 256 | } |
| 257 | return 0; |
| 258 | |
| 259 | conflict: |
| 260 | printk(KERN_INFO "%s:%d conflicting memory types " |
| 261 | "%Lx-%Lx %s<->%s\n", current->comm, current->pid, new->start, |
| 262 | new->end, cattr_name(new->type), cattr_name(entry->type)); |
| 263 | return -EBUSY; |
| 264 | } |
| 265 | |
| 266 | static int pat_pagerange_is_ram(unsigned long start, unsigned long end) |
| 267 | { |
| 268 | int ram_page = 0, not_rampage = 0; |
| 269 | unsigned long page_nr; |
| 270 | |
| 271 | for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT); |
| 272 | ++page_nr) { |
| 273 | /* |
| 274 | * For legacy reasons, physical address range in the legacy ISA |
| 275 | * region is tracked as non-RAM. This will allow users of |
| 276 | * /dev/mem to map portions of legacy ISA region, even when |
| 277 | * some of those portions are listed(or not even listed) with |
| 278 | * different e820 types(RAM/reserved/..) |
| 279 | */ |
| 280 | if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) && |
| 281 | page_is_ram(page_nr)) |
| 282 | ram_page = 1; |
| 283 | else |
| 284 | not_rampage = 1; |
| 285 | |
| 286 | if (ram_page == not_rampage) |
| 287 | return -1; |
| 288 | } |
| 289 | |
| 290 | return ram_page; |
| 291 | } |
| 292 | |
| 293 | /* |
| 294 | * For RAM pages, we use page flags to mark the pages with appropriate type. |
| 295 | * Here we do two pass: |
| 296 | * - Find the memtype of all the pages in the range, look for any conflicts |
| 297 | * - In case of no conflicts, set the new memtype for pages in the range |
| 298 | * |
| 299 | * Caller must hold memtype_lock for atomicity. |
| 300 | */ |
| 301 | static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type, |
| 302 | unsigned long *new_type) |
| 303 | { |
| 304 | struct page *page; |
| 305 | u64 pfn; |
| 306 | |
| 307 | if (req_type == _PAGE_CACHE_UC) { |
| 308 | /* We do not support strong UC */ |
| 309 | WARN_ON_ONCE(1); |
| 310 | req_type = _PAGE_CACHE_UC_MINUS; |
| 311 | } |
| 312 | |
| 313 | for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) { |
| 314 | unsigned long type; |
| 315 | |
| 316 | page = pfn_to_page(pfn); |
| 317 | type = get_page_memtype(page); |
| 318 | if (type != -1) { |
| 319 | printk(KERN_INFO "reserve_ram_pages_type failed " |
| 320 | "0x%Lx-0x%Lx, track 0x%lx, req 0x%lx\n", |
| 321 | start, end, type, req_type); |
| 322 | if (new_type) |
| 323 | *new_type = type; |
| 324 | |
| 325 | return -EBUSY; |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | if (new_type) |
| 330 | *new_type = req_type; |
| 331 | |
| 332 | for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) { |
| 333 | page = pfn_to_page(pfn); |
| 334 | set_page_memtype(page, req_type); |
| 335 | } |
| 336 | return 0; |
| 337 | } |
| 338 | |
| 339 | static int free_ram_pages_type(u64 start, u64 end) |
| 340 | { |
| 341 | struct page *page; |
| 342 | u64 pfn; |
| 343 | |
| 344 | for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) { |
| 345 | page = pfn_to_page(pfn); |
| 346 | set_page_memtype(page, -1); |
| 347 | } |
| 348 | return 0; |
| 349 | } |
| 350 | |
| 351 | /* |
| 352 | * req_type typically has one of the: |
| 353 | * - _PAGE_CACHE_WB |
| 354 | * - _PAGE_CACHE_WC |
| 355 | * - _PAGE_CACHE_UC_MINUS |
| 356 | * - _PAGE_CACHE_UC |
| 357 | * |
| 358 | * req_type will have a special case value '-1', when requester want to inherit |
| 359 | * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS. |
| 360 | * |
| 361 | * If new_type is NULL, function will return an error if it cannot reserve the |
| 362 | * region with req_type. If new_type is non-NULL, function will return |
| 363 | * available type in new_type in case of no error. In case of any error |
| 364 | * it will return a negative return value. |
| 365 | */ |
| 366 | int reserve_memtype(u64 start, u64 end, unsigned long req_type, |
| 367 | unsigned long *new_type) |
| 368 | { |
| 369 | struct memtype *new, *entry; |
| 370 | unsigned long actual_type; |
| 371 | struct list_head *where; |
| 372 | int is_range_ram; |
| 373 | int err = 0; |
| 374 | |
| 375 | BUG_ON(start >= end); /* end is exclusive */ |
| 376 | |
| 377 | if (!pat_enabled) { |
| 378 | /* This is identical to page table setting without PAT */ |
| 379 | if (new_type) { |
| 380 | if (req_type == -1) |
| 381 | *new_type = _PAGE_CACHE_WB; |
| 382 | else if (req_type == _PAGE_CACHE_WC) |
| 383 | *new_type = _PAGE_CACHE_UC_MINUS; |
| 384 | else |
| 385 | *new_type = req_type & _PAGE_CACHE_MASK; |
| 386 | } |
| 387 | return 0; |
| 388 | } |
| 389 | |
| 390 | /* Low ISA region is always mapped WB in page table. No need to track */ |
| 391 | if (is_ISA_range(start, end - 1)) { |
| 392 | if (new_type) |
| 393 | *new_type = _PAGE_CACHE_WB; |
| 394 | return 0; |
| 395 | } |
| 396 | |
| 397 | /* |
| 398 | * Call mtrr_lookup to get the type hint. This is an |
| 399 | * optimization for /dev/mem mmap'ers into WB memory (BIOS |
| 400 | * tools and ACPI tools). Use WB request for WB memory and use |
| 401 | * UC_MINUS otherwise. |
| 402 | */ |
| 403 | actual_type = pat_x_mtrr_type(start, end, req_type & _PAGE_CACHE_MASK); |
| 404 | |
| 405 | if (new_type) |
| 406 | *new_type = actual_type; |
| 407 | |
| 408 | is_range_ram = pat_pagerange_is_ram(start, end); |
| 409 | if (is_range_ram == 1) { |
| 410 | |
| 411 | spin_lock(&memtype_lock); |
| 412 | err = reserve_ram_pages_type(start, end, req_type, new_type); |
| 413 | spin_unlock(&memtype_lock); |
| 414 | |
| 415 | return err; |
| 416 | } else if (is_range_ram < 0) { |
| 417 | return -EINVAL; |
| 418 | } |
| 419 | |
| 420 | new = kmalloc(sizeof(struct memtype), GFP_KERNEL); |
| 421 | if (!new) |
| 422 | return -ENOMEM; |
| 423 | |
| 424 | new->start = start; |
| 425 | new->end = end; |
| 426 | new->type = actual_type; |
| 427 | |
| 428 | spin_lock(&memtype_lock); |
| 429 | |
| 430 | /* Search for existing mapping that overlaps the current range */ |
| 431 | where = NULL; |
| 432 | list_for_each_entry(entry, &memtype_list, nd) { |
| 433 | if (end <= entry->start) { |
| 434 | where = entry->nd.prev; |
| 435 | break; |
| 436 | } else if (start <= entry->start) { /* end > entry->start */ |
| 437 | err = chk_conflict(new, entry, new_type); |
| 438 | if (!err) { |
| 439 | dprintk("Overlap at 0x%Lx-0x%Lx\n", |
| 440 | entry->start, entry->end); |
| 441 | where = entry->nd.prev; |
| 442 | } |
| 443 | break; |
| 444 | } else if (start < entry->end) { /* start > entry->start */ |
| 445 | err = chk_conflict(new, entry, new_type); |
| 446 | if (!err) { |
| 447 | dprintk("Overlap at 0x%Lx-0x%Lx\n", |
| 448 | entry->start, entry->end); |
| 449 | |
| 450 | /* |
| 451 | * Move to right position in the linked |
| 452 | * list to add this new entry |
| 453 | */ |
| 454 | list_for_each_entry_continue(entry, |
| 455 | &memtype_list, nd) { |
| 456 | if (start <= entry->start) { |
| 457 | where = entry->nd.prev; |
| 458 | break; |
| 459 | } |
| 460 | } |
| 461 | } |
| 462 | break; |
| 463 | } |
| 464 | } |
| 465 | |
| 466 | if (err) { |
| 467 | printk(KERN_INFO "reserve_memtype failed 0x%Lx-0x%Lx, " |
| 468 | "track %s, req %s\n", |
| 469 | start, end, cattr_name(new->type), cattr_name(req_type)); |
| 470 | kfree(new); |
| 471 | spin_unlock(&memtype_lock); |
| 472 | |
| 473 | return err; |
| 474 | } |
| 475 | |
| 476 | if (where) |
| 477 | list_add(&new->nd, where); |
| 478 | else |
| 479 | list_add_tail(&new->nd, &memtype_list); |
| 480 | |
| 481 | memtype_rb_insert(&memtype_rbroot, new); |
| 482 | |
| 483 | spin_unlock(&memtype_lock); |
| 484 | |
| 485 | dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n", |
| 486 | start, end, cattr_name(new->type), cattr_name(req_type), |
| 487 | new_type ? cattr_name(*new_type) : "-"); |
| 488 | |
| 489 | return err; |
| 490 | } |
| 491 | |
| 492 | int free_memtype(u64 start, u64 end) |
| 493 | { |
| 494 | struct memtype *entry, *saved_entry; |
| 495 | int err = -EINVAL; |
| 496 | int is_range_ram; |
| 497 | |
| 498 | if (!pat_enabled) |
| 499 | return 0; |
| 500 | |
| 501 | /* Low ISA region is always mapped WB. No need to track */ |
| 502 | if (is_ISA_range(start, end - 1)) |
| 503 | return 0; |
| 504 | |
| 505 | is_range_ram = pat_pagerange_is_ram(start, end); |
| 506 | if (is_range_ram == 1) { |
| 507 | |
| 508 | spin_lock(&memtype_lock); |
| 509 | err = free_ram_pages_type(start, end); |
| 510 | spin_unlock(&memtype_lock); |
| 511 | |
| 512 | return err; |
| 513 | } else if (is_range_ram < 0) { |
| 514 | return -EINVAL; |
| 515 | } |
| 516 | |
| 517 | spin_lock(&memtype_lock); |
| 518 | |
| 519 | entry = memtype_rb_search(&memtype_rbroot, start); |
| 520 | if (unlikely(entry == NULL)) |
| 521 | goto unlock_ret; |
| 522 | |
| 523 | /* |
| 524 | * Saved entry points to an entry with start same or less than what |
| 525 | * we searched for. Now go through the list in both directions to look |
| 526 | * for the entry that matches with both start and end, with list stored |
| 527 | * in sorted start address |
| 528 | */ |
| 529 | saved_entry = entry; |
| 530 | list_for_each_entry_from(entry, &memtype_list, nd) { |
| 531 | if (entry->start == start && entry->end == end) { |
| 532 | rb_erase(&entry->rb, &memtype_rbroot); |
| 533 | list_del(&entry->nd); |
| 534 | kfree(entry); |
| 535 | err = 0; |
| 536 | break; |
| 537 | } else if (entry->start > start) { |
| 538 | break; |
| 539 | } |
| 540 | } |
| 541 | |
| 542 | if (!err) |
| 543 | goto unlock_ret; |
| 544 | |
| 545 | entry = saved_entry; |
| 546 | list_for_each_entry_reverse(entry, &memtype_list, nd) { |
| 547 | if (entry->start == start && entry->end == end) { |
| 548 | rb_erase(&entry->rb, &memtype_rbroot); |
| 549 | list_del(&entry->nd); |
| 550 | kfree(entry); |
| 551 | err = 0; |
| 552 | break; |
| 553 | } else if (entry->start < start) { |
| 554 | break; |
| 555 | } |
| 556 | } |
| 557 | unlock_ret: |
| 558 | spin_unlock(&memtype_lock); |
| 559 | |
| 560 | if (err) { |
| 561 | printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n", |
| 562 | current->comm, current->pid, start, end); |
| 563 | } |
| 564 | |
| 565 | dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end); |
| 566 | |
| 567 | return err; |
| 568 | } |
| 569 | |
| 570 | |
| 571 | /** |
| 572 | * lookup_memtype - Looksup the memory type for a physical address |
| 573 | * @paddr: physical address of which memory type needs to be looked up |
| 574 | * |
| 575 | * Only to be called when PAT is enabled |
| 576 | * |
| 577 | * Returns _PAGE_CACHE_WB, _PAGE_CACHE_WC, _PAGE_CACHE_UC_MINUS or |
| 578 | * _PAGE_CACHE_UC |
| 579 | */ |
| 580 | static unsigned long lookup_memtype(u64 paddr) |
| 581 | { |
| 582 | int rettype = _PAGE_CACHE_WB; |
| 583 | struct memtype *entry; |
| 584 | |
| 585 | if (is_ISA_range(paddr, paddr + PAGE_SIZE - 1)) |
| 586 | return rettype; |
| 587 | |
| 588 | if (pat_pagerange_is_ram(paddr, paddr + PAGE_SIZE)) { |
| 589 | struct page *page; |
| 590 | spin_lock(&memtype_lock); |
| 591 | page = pfn_to_page(paddr >> PAGE_SHIFT); |
| 592 | rettype = get_page_memtype(page); |
| 593 | spin_unlock(&memtype_lock); |
| 594 | /* |
| 595 | * -1 from get_page_memtype() implies RAM page is in its |
| 596 | * default state and not reserved, and hence of type WB |
| 597 | */ |
| 598 | if (rettype == -1) |
| 599 | rettype = _PAGE_CACHE_WB; |
| 600 | |
| 601 | return rettype; |
| 602 | } |
| 603 | |
| 604 | spin_lock(&memtype_lock); |
| 605 | |
| 606 | entry = memtype_rb_search(&memtype_rbroot, paddr); |
| 607 | if (entry != NULL) |
| 608 | rettype = entry->type; |
| 609 | else |
| 610 | rettype = _PAGE_CACHE_UC_MINUS; |
| 611 | |
| 612 | spin_unlock(&memtype_lock); |
| 613 | return rettype; |
| 614 | } |
| 615 | |
| 616 | /** |
| 617 | * io_reserve_memtype - Request a memory type mapping for a region of memory |
| 618 | * @start: start (physical address) of the region |
| 619 | * @end: end (physical address) of the region |
| 620 | * @type: A pointer to memtype, with requested type. On success, requested |
| 621 | * or any other compatible type that was available for the region is returned |
| 622 | * |
| 623 | * On success, returns 0 |
| 624 | * On failure, returns non-zero |
| 625 | */ |
| 626 | int io_reserve_memtype(resource_size_t start, resource_size_t end, |
| 627 | unsigned long *type) |
| 628 | { |
| 629 | resource_size_t size = end - start; |
| 630 | unsigned long req_type = *type; |
| 631 | unsigned long new_type; |
| 632 | int ret; |
| 633 | |
| 634 | WARN_ON_ONCE(iomem_map_sanity_check(start, size)); |
| 635 | |
| 636 | ret = reserve_memtype(start, end, req_type, &new_type); |
| 637 | if (ret) |
| 638 | goto out_err; |
| 639 | |
| 640 | if (!is_new_memtype_allowed(start, size, req_type, new_type)) |
| 641 | goto out_free; |
| 642 | |
| 643 | if (kernel_map_sync_memtype(start, size, new_type) < 0) |
| 644 | goto out_free; |
| 645 | |
| 646 | *type = new_type; |
| 647 | return 0; |
| 648 | |
| 649 | out_free: |
| 650 | free_memtype(start, end); |
| 651 | ret = -EBUSY; |
| 652 | out_err: |
| 653 | return ret; |
| 654 | } |
| 655 | |
| 656 | /** |
| 657 | * io_free_memtype - Release a memory type mapping for a region of memory |
| 658 | * @start: start (physical address) of the region |
| 659 | * @end: end (physical address) of the region |
| 660 | */ |
| 661 | void io_free_memtype(resource_size_t start, resource_size_t end) |
| 662 | { |
| 663 | free_memtype(start, end); |
| 664 | } |
| 665 | |
| 666 | pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn, |
| 667 | unsigned long size, pgprot_t vma_prot) |
| 668 | { |
| 669 | return vma_prot; |
| 670 | } |
| 671 | |
| 672 | #ifdef CONFIG_STRICT_DEVMEM |
| 673 | /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/ |
| 674 | static inline int range_is_allowed(unsigned long pfn, unsigned long size) |
| 675 | { |
| 676 | return 1; |
| 677 | } |
| 678 | #else |
| 679 | /* This check is needed to avoid cache aliasing when PAT is enabled */ |
| 680 | static inline int range_is_allowed(unsigned long pfn, unsigned long size) |
| 681 | { |
| 682 | u64 from = ((u64)pfn) << PAGE_SHIFT; |
| 683 | u64 to = from + size; |
| 684 | u64 cursor = from; |
| 685 | |
| 686 | if (!pat_enabled) |
| 687 | return 1; |
| 688 | |
| 689 | while (cursor < to) { |
| 690 | if (!devmem_is_allowed(pfn)) { |
| 691 | printk(KERN_INFO |
| 692 | "Program %s tried to access /dev/mem between %Lx->%Lx.\n", |
| 693 | current->comm, from, to); |
| 694 | return 0; |
| 695 | } |
| 696 | cursor += PAGE_SIZE; |
| 697 | pfn++; |
| 698 | } |
| 699 | return 1; |
| 700 | } |
| 701 | #endif /* CONFIG_STRICT_DEVMEM */ |
| 702 | |
| 703 | int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn, |
| 704 | unsigned long size, pgprot_t *vma_prot) |
| 705 | { |
| 706 | unsigned long flags = _PAGE_CACHE_WB; |
| 707 | |
| 708 | if (!range_is_allowed(pfn, size)) |
| 709 | return 0; |
| 710 | |
| 711 | if (file->f_flags & O_SYNC) { |
| 712 | flags = _PAGE_CACHE_UC_MINUS; |
| 713 | } |
| 714 | |
| 715 | #ifdef CONFIG_X86_32 |
| 716 | /* |
| 717 | * On the PPro and successors, the MTRRs are used to set |
| 718 | * memory types for physical addresses outside main memory, |
| 719 | * so blindly setting UC or PWT on those pages is wrong. |
| 720 | * For Pentiums and earlier, the surround logic should disable |
| 721 | * caching for the high addresses through the KEN pin, but |
| 722 | * we maintain the tradition of paranoia in this code. |
| 723 | */ |
| 724 | if (!pat_enabled && |
| 725 | !(boot_cpu_has(X86_FEATURE_MTRR) || |
| 726 | boot_cpu_has(X86_FEATURE_K6_MTRR) || |
| 727 | boot_cpu_has(X86_FEATURE_CYRIX_ARR) || |
| 728 | boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) && |
| 729 | (pfn << PAGE_SHIFT) >= __pa(high_memory)) { |
| 730 | flags = _PAGE_CACHE_UC; |
| 731 | } |
| 732 | #endif |
| 733 | |
| 734 | *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) | |
| 735 | flags); |
| 736 | return 1; |
| 737 | } |
| 738 | |
| 739 | /* |
| 740 | * Change the memory type for the physial address range in kernel identity |
| 741 | * mapping space if that range is a part of identity map. |
| 742 | */ |
| 743 | int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags) |
| 744 | { |
| 745 | unsigned long id_sz; |
| 746 | |
| 747 | if (base >= __pa(high_memory)) |
| 748 | return 0; |
| 749 | |
| 750 | id_sz = (__pa(high_memory) < base + size) ? |
| 751 | __pa(high_memory) - base : |
| 752 | size; |
| 753 | |
| 754 | if (ioremap_change_attr((unsigned long)__va(base), id_sz, flags) < 0) { |
| 755 | printk(KERN_INFO |
| 756 | "%s:%d ioremap_change_attr failed %s " |
| 757 | "for %Lx-%Lx\n", |
| 758 | current->comm, current->pid, |
| 759 | cattr_name(flags), |
| 760 | base, (unsigned long long)(base + size)); |
| 761 | return -EINVAL; |
| 762 | } |
| 763 | return 0; |
| 764 | } |
| 765 | |
| 766 | /* |
| 767 | * Internal interface to reserve a range of physical memory with prot. |
| 768 | * Reserved non RAM regions only and after successful reserve_memtype, |
| 769 | * this func also keeps identity mapping (if any) in sync with this new prot. |
| 770 | */ |
| 771 | static int reserve_pfn_range(u64 paddr, unsigned long size, pgprot_t *vma_prot, |
| 772 | int strict_prot) |
| 773 | { |
| 774 | int is_ram = 0; |
| 775 | int ret; |
| 776 | unsigned long want_flags = (pgprot_val(*vma_prot) & _PAGE_CACHE_MASK); |
| 777 | unsigned long flags = want_flags; |
| 778 | |
| 779 | is_ram = pat_pagerange_is_ram(paddr, paddr + size); |
| 780 | |
| 781 | /* |
| 782 | * reserve_pfn_range() for RAM pages. We do not refcount to keep |
| 783 | * track of number of mappings of RAM pages. We can assert that |
| 784 | * the type requested matches the type of first page in the range. |
| 785 | */ |
| 786 | if (is_ram) { |
| 787 | if (!pat_enabled) |
| 788 | return 0; |
| 789 | |
| 790 | flags = lookup_memtype(paddr); |
| 791 | if (want_flags != flags) { |
| 792 | printk(KERN_WARNING |
| 793 | "%s:%d map pfn RAM range req %s for %Lx-%Lx, got %s\n", |
| 794 | current->comm, current->pid, |
| 795 | cattr_name(want_flags), |
| 796 | (unsigned long long)paddr, |
| 797 | (unsigned long long)(paddr + size), |
| 798 | cattr_name(flags)); |
| 799 | *vma_prot = __pgprot((pgprot_val(*vma_prot) & |
| 800 | (~_PAGE_CACHE_MASK)) | |
| 801 | flags); |
| 802 | } |
| 803 | return 0; |
| 804 | } |
| 805 | |
| 806 | ret = reserve_memtype(paddr, paddr + size, want_flags, &flags); |
| 807 | if (ret) |
| 808 | return ret; |
| 809 | |
| 810 | if (flags != want_flags) { |
| 811 | if (strict_prot || |
| 812 | !is_new_memtype_allowed(paddr, size, want_flags, flags)) { |
| 813 | free_memtype(paddr, paddr + size); |
| 814 | printk(KERN_ERR "%s:%d map pfn expected mapping type %s" |
| 815 | " for %Lx-%Lx, got %s\n", |
| 816 | current->comm, current->pid, |
| 817 | cattr_name(want_flags), |
| 818 | (unsigned long long)paddr, |
| 819 | (unsigned long long)(paddr + size), |
| 820 | cattr_name(flags)); |
| 821 | return -EINVAL; |
| 822 | } |
| 823 | /* |
| 824 | * We allow returning different type than the one requested in |
| 825 | * non strict case. |
| 826 | */ |
| 827 | *vma_prot = __pgprot((pgprot_val(*vma_prot) & |
| 828 | (~_PAGE_CACHE_MASK)) | |
| 829 | flags); |
| 830 | } |
| 831 | |
| 832 | if (kernel_map_sync_memtype(paddr, size, flags) < 0) { |
| 833 | free_memtype(paddr, paddr + size); |
| 834 | return -EINVAL; |
| 835 | } |
| 836 | return 0; |
| 837 | } |
| 838 | |
| 839 | /* |
| 840 | * Internal interface to free a range of physical memory. |
| 841 | * Frees non RAM regions only. |
| 842 | */ |
| 843 | static void free_pfn_range(u64 paddr, unsigned long size) |
| 844 | { |
| 845 | int is_ram; |
| 846 | |
| 847 | is_ram = pat_pagerange_is_ram(paddr, paddr + size); |
| 848 | if (is_ram == 0) |
| 849 | free_memtype(paddr, paddr + size); |
| 850 | } |
| 851 | |
| 852 | /* |
| 853 | * track_pfn_vma_copy is called when vma that is covering the pfnmap gets |
| 854 | * copied through copy_page_range(). |
| 855 | * |
| 856 | * If the vma has a linear pfn mapping for the entire range, we get the prot |
| 857 | * from pte and reserve the entire vma range with single reserve_pfn_range call. |
| 858 | */ |
| 859 | int track_pfn_vma_copy(struct vm_area_struct *vma) |
| 860 | { |
| 861 | resource_size_t paddr; |
| 862 | unsigned long prot; |
| 863 | unsigned long vma_size = vma->vm_end - vma->vm_start; |
| 864 | pgprot_t pgprot; |
| 865 | |
| 866 | if (is_linear_pfn_mapping(vma)) { |
| 867 | /* |
| 868 | * reserve the whole chunk covered by vma. We need the |
| 869 | * starting address and protection from pte. |
| 870 | */ |
| 871 | if (follow_phys(vma, vma->vm_start, 0, &prot, &paddr)) { |
| 872 | WARN_ON_ONCE(1); |
| 873 | return -EINVAL; |
| 874 | } |
| 875 | pgprot = __pgprot(prot); |
| 876 | return reserve_pfn_range(paddr, vma_size, &pgprot, 1); |
| 877 | } |
| 878 | |
| 879 | return 0; |
| 880 | } |
| 881 | |
| 882 | /* |
| 883 | * track_pfn_vma_new is called when a _new_ pfn mapping is being established |
| 884 | * for physical range indicated by pfn and size. |
| 885 | * |
| 886 | * prot is passed in as a parameter for the new mapping. If the vma has a |
| 887 | * linear pfn mapping for the entire range reserve the entire vma range with |
| 888 | * single reserve_pfn_range call. |
| 889 | */ |
| 890 | int track_pfn_vma_new(struct vm_area_struct *vma, pgprot_t *prot, |
| 891 | unsigned long pfn, unsigned long size) |
| 892 | { |
| 893 | unsigned long flags; |
| 894 | resource_size_t paddr; |
| 895 | unsigned long vma_size = vma->vm_end - vma->vm_start; |
| 896 | |
| 897 | if (is_linear_pfn_mapping(vma)) { |
| 898 | /* reserve the whole chunk starting from vm_pgoff */ |
| 899 | paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT; |
| 900 | return reserve_pfn_range(paddr, vma_size, prot, 0); |
| 901 | } |
| 902 | |
| 903 | if (!pat_enabled) |
| 904 | return 0; |
| 905 | |
| 906 | /* for vm_insert_pfn and friends, we set prot based on lookup */ |
| 907 | flags = lookup_memtype(pfn << PAGE_SHIFT); |
| 908 | *prot = __pgprot((pgprot_val(vma->vm_page_prot) & (~_PAGE_CACHE_MASK)) | |
| 909 | flags); |
| 910 | |
| 911 | return 0; |
| 912 | } |
| 913 | |
| 914 | /* |
| 915 | * untrack_pfn_vma is called while unmapping a pfnmap for a region. |
| 916 | * untrack can be called for a specific region indicated by pfn and size or |
| 917 | * can be for the entire vma (in which case size can be zero). |
| 918 | */ |
| 919 | void untrack_pfn_vma(struct vm_area_struct *vma, unsigned long pfn, |
| 920 | unsigned long size) |
| 921 | { |
| 922 | resource_size_t paddr; |
| 923 | unsigned long vma_size = vma->vm_end - vma->vm_start; |
| 924 | |
| 925 | if (is_linear_pfn_mapping(vma)) { |
| 926 | /* free the whole chunk starting from vm_pgoff */ |
| 927 | paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT; |
| 928 | free_pfn_range(paddr, vma_size); |
| 929 | return; |
| 930 | } |
| 931 | } |
| 932 | |
| 933 | pgprot_t pgprot_writecombine(pgprot_t prot) |
| 934 | { |
| 935 | if (pat_enabled) |
| 936 | return __pgprot(pgprot_val(prot) | _PAGE_CACHE_WC); |
| 937 | else |
| 938 | return pgprot_noncached(prot); |
| 939 | } |
| 940 | EXPORT_SYMBOL_GPL(pgprot_writecombine); |
| 941 | |
| 942 | #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT) |
| 943 | |
| 944 | /* get Nth element of the linked list */ |
| 945 | static struct memtype *memtype_get_idx(loff_t pos) |
| 946 | { |
| 947 | struct memtype *list_node, *print_entry; |
| 948 | int i = 1; |
| 949 | |
| 950 | print_entry = kmalloc(sizeof(struct memtype), GFP_KERNEL); |
| 951 | if (!print_entry) |
| 952 | return NULL; |
| 953 | |
| 954 | spin_lock(&memtype_lock); |
| 955 | list_for_each_entry(list_node, &memtype_list, nd) { |
| 956 | if (pos == i) { |
| 957 | *print_entry = *list_node; |
| 958 | spin_unlock(&memtype_lock); |
| 959 | return print_entry; |
| 960 | } |
| 961 | ++i; |
| 962 | } |
| 963 | spin_unlock(&memtype_lock); |
| 964 | kfree(print_entry); |
| 965 | |
| 966 | return NULL; |
| 967 | } |
| 968 | |
| 969 | static void *memtype_seq_start(struct seq_file *seq, loff_t *pos) |
| 970 | { |
| 971 | if (*pos == 0) { |
| 972 | ++*pos; |
| 973 | seq_printf(seq, "PAT memtype list:\n"); |
| 974 | } |
| 975 | |
| 976 | return memtype_get_idx(*pos); |
| 977 | } |
| 978 | |
| 979 | static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos) |
| 980 | { |
| 981 | ++*pos; |
| 982 | return memtype_get_idx(*pos); |
| 983 | } |
| 984 | |
| 985 | static void memtype_seq_stop(struct seq_file *seq, void *v) |
| 986 | { |
| 987 | } |
| 988 | |
| 989 | static int memtype_seq_show(struct seq_file *seq, void *v) |
| 990 | { |
| 991 | struct memtype *print_entry = (struct memtype *)v; |
| 992 | |
| 993 | seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type), |
| 994 | print_entry->start, print_entry->end); |
| 995 | kfree(print_entry); |
| 996 | |
| 997 | return 0; |
| 998 | } |
| 999 | |
| 1000 | static const struct seq_operations memtype_seq_ops = { |
| 1001 | .start = memtype_seq_start, |
| 1002 | .next = memtype_seq_next, |
| 1003 | .stop = memtype_seq_stop, |
| 1004 | .show = memtype_seq_show, |
| 1005 | }; |
| 1006 | |
| 1007 | static int memtype_seq_open(struct inode *inode, struct file *file) |
| 1008 | { |
| 1009 | return seq_open(file, &memtype_seq_ops); |
| 1010 | } |
| 1011 | |
| 1012 | static const struct file_operations memtype_fops = { |
| 1013 | .open = memtype_seq_open, |
| 1014 | .read = seq_read, |
| 1015 | .llseek = seq_lseek, |
| 1016 | .release = seq_release, |
| 1017 | }; |
| 1018 | |
| 1019 | static int __init pat_memtype_list_init(void) |
| 1020 | { |
| 1021 | debugfs_create_file("pat_memtype_list", S_IRUSR, arch_debugfs_dir, |
| 1022 | NULL, &memtype_fops); |
| 1023 | return 0; |
| 1024 | } |
| 1025 | |
| 1026 | late_initcall(pat_memtype_list_init); |
| 1027 | |
| 1028 | #endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */ |