arch/x86/mm/dump_pagetables.c

   1 /*
   2  * Debug helper to dump the current kernel pagetables of the system
   3  * so that we can see what the various memory ranges are set to.
   4  *
   5  * (C) Copyright 2008 Intel Corporation
   6  *
   7  * Author: Arjan van de Ven <arjan@linux.intel.com>
   8  *
   9  * This program is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU General Public License
  11  * as published by the Free Software Foundation; version 2
  12  * of the License.
  13  */
  14
  15 #include <linux/debugfs.h>
  16 #include <linux/mm.h>
  17 #include <linux/init.h>
  18 #include <linux/sched.h>
  19 #include <linux/seq_file.h>
  20
  21 #include <asm/pgtable.h>
  22
  23 /*
  24  * The dumper groups pagetable entries of the same type into one, and for
  25  * that it needs to keep some state when walking, and flush this state
  26  * when a "break" in the continuity is found.
  27  */
  28 struct pg_state {
  29         int level;
  30         pgprot_t current_prot;
  31         unsigned long start_address;
  32         unsigned long current_address;
  33         const struct addr_marker *marker;
  34         unsigned long lines;
  35         bool to_dmesg;
  36         bool check_wx;
  37         unsigned long wx_pages;
  38 };
  39
  40 struct addr_marker {
  41         unsigned long start_address;
  42         const char *name;
  43         unsigned long max_lines;
  44 };
  45
  46 /* indices for address_markers; keep sync'd w/ address_markers below */
  47 enum address_markers_idx {
  48         USER_SPACE_NR = 0,
  49 #ifdef CONFIG_X86_64
  50         KERNEL_SPACE_NR,
  51         LOW_KERNEL_NR,
  52         VMALLOC_START_NR,
  53         VMEMMAP_START_NR,
  54 # ifdef CONFIG_X86_ESPFIX64
  55         ESPFIX_START_NR,
  56 # endif
  57         HIGH_KERNEL_NR,
  58         MODULES_VADDR_NR,
  59         MODULES_END_NR,
  60 #else
  61         KERNEL_SPACE_NR,
  62         VMALLOC_START_NR,
  63         VMALLOC_END_NR,
  64 # ifdef CONFIG_HIGHMEM
  65         PKMAP_BASE_NR,
  66 # endif
  67         FIXADDR_START_NR,
  68 #endif
  69 };
  70
  71 /* Address space markers hints */
  72 static struct addr_marker address_markers[] = {
  73         { 0, "User Space" },
  74 #ifdef CONFIG_X86_64
  75         { 0x8000000000000000UL, "Kernel Space" },
  76         { 0/* PAGE_OFFSET */,   "Low Kernel Mapping" },
  77         { 0/* VMALLOC_START */, "vmalloc() Area" },
  78         { 0/* VMEMMAP_START */, "Vmemmap" },
  79 # ifdef CONFIG_X86_ESPFIX64
  80         { ESPFIX_BASE_ADDR,     "ESPfix Area", 16 },
  81 # endif
  82 # ifdef CONFIG_EFI
  83         { EFI_VA_END,           "EFI Runtime Services" },
  84 # endif
  85         { __START_KERNEL_map,   "High Kernel Mapping" },
  86         { MODULES_VADDR,        "Modules" },
  87         { MODULES_END,          "End Modules" },
  88 #else
  89         { PAGE_OFFSET,          "Kernel Mapping" },
  90         { 0/* VMALLOC_START */, "vmalloc() Area" },
  91         { 0/*VMALLOC_END*/,     "vmalloc() End" },
  92 # ifdef CONFIG_HIGHMEM
  93         { 0/*PKMAP_BASE*/,      "Persistent kmap() Area" },
  94 # endif
  95         { 0/*FIXADDR_START*/,   "Fixmap Area" },
  96 #endif
  97         { -1, NULL }            /* End of list */
  98 };
  99
 100 /* Multipliers for offsets within the PTEs */
 101 #define PTE_LEVEL_MULT (PAGE_SIZE)
 102 #define PMD_LEVEL_MULT (PTRS_PER_PTE * PTE_LEVEL_MULT)
 103 #define PUD_LEVEL_MULT (PTRS_PER_PMD * PMD_LEVEL_MULT)
 104 #define PGD_LEVEL_MULT (PTRS_PER_PUD * PUD_LEVEL_MULT)
 105
 106 #define pt_dump_seq_printf(m, to_dmesg, fmt, args...)           \
 107 ({                                                              \
 108         if (to_dmesg)                                   \
 109                 printk(KERN_INFO fmt, ##args);                  \
 110         else                                                    \
 111                 if (m)                                          \
 112                         seq_printf(m, fmt, ##args);             \
 113 })
 114
 115 #define pt_dump_cont_printf(m, to_dmesg, fmt, args...)          \
 116 ({                                                              \
 117         if (to_dmesg)                                   \
 118                 printk(KERN_CONT fmt, ##args);                  \
 119         else                                                    \
 120                 if (m)                                          \
 121                         seq_printf(m, fmt, ##args);             \
 122 })
 123
 124 /*
 125  * Print a readable form of a pgprot_t to the seq_file
 126  */
 127 static void printk_prot(struct seq_file *m, pgprot_t prot, int level, bool dmsg)
 128 {
 129         pgprotval_t pr = pgprot_val(prot);
 130         static const char * const level_name[] =
 131                 { "cr3", "pgd", "pud", "pmd", "pte" };
 132
 133         if (!pgprot_val(prot)) {
 134                 /* Not present */
 135                 pt_dump_cont_printf(m, dmsg, "                              ");
 136         } else {
 137                 if (pr & _PAGE_USER)
 138                         pt_dump_cont_printf(m, dmsg, "USR ");
 139                 else
 140                         pt_dump_cont_printf(m, dmsg, "    ");
 141                 if (pr & _PAGE_RW)
 142                         pt_dump_cont_printf(m, dmsg, "RW ");
 143                 else
 144                         pt_dump_cont_printf(m, dmsg, "ro ");
 145                 if (pr & _PAGE_PWT)
 146                         pt_dump_cont_printf(m, dmsg, "PWT ");
 147                 else
 148                         pt_dump_cont_printf(m, dmsg, "    ");
 149                 if (pr & _PAGE_PCD)
 150                         pt_dump_cont_printf(m, dmsg, "PCD ");
 151                 else
 152                         pt_dump_cont_printf(m, dmsg, "    ");
 153
 154                 /* Bit 7 has a different meaning on level 3 vs 4 */
 155                 if (level <= 3 && pr & _PAGE_PSE)
 156                         pt_dump_cont_printf(m, dmsg, "PSE ");
 157                 else
 158                         pt_dump_cont_printf(m, dmsg, "    ");
 159                 if ((level == 4 && pr & _PAGE_PAT) ||
 160                     ((level == 3 || level == 2) && pr & _PAGE_PAT_LARGE))
 161                         pt_dump_cont_printf(m, dmsg, "PAT ");
 162                 else
 163                         pt_dump_cont_printf(m, dmsg, "    ");
 164                 if (pr & _PAGE_GLOBAL)
 165                         pt_dump_cont_printf(m, dmsg, "GLB ");
 166                 else
 167                         pt_dump_cont_printf(m, dmsg, "    ");
 168                 if (pr & _PAGE_NX)
 169                         pt_dump_cont_printf(m, dmsg, "NX ");
 170                 else
 171                         pt_dump_cont_printf(m, dmsg, "x  ");
 172         }
 173         pt_dump_cont_printf(m, dmsg, "%s\n", level_name[level]);
 174 }
 175
 176 /*
 177  * On 64 bits, sign-extend the 48 bit address to 64 bit
 178  */
 179 static unsigned long normalize_addr(unsigned long u)
 180 {
 181 #ifdef CONFIG_X86_64
 182         return (signed long)(u << 16) >> 16;
 183 #else
 184         return u;
 185 #endif
 186 }
 187
 188 /*
 189  * This function gets called on a break in a continuous series
 190  * of PTE entries; the next one is different so we need to
 191  * print what we collected so far.
 192  */
 193 static void note_page(struct seq_file *m, struct pg_state *st,
 194                       pgprot_t new_prot, int level)
 195 {
 196         pgprotval_t prot, cur;
 197         static const char units[] = "BKMGTPE";
 198
 199         /*
 200          * If we have a "break" in the series, we need to flush the state that
 201          * we have now. "break" is either changing perms, levels or
 202          * address space marker.
 203          */
 204         prot = pgprot_val(new_prot);
 205         cur = pgprot_val(st->current_prot);
 206
 207         if (!st->level) {
 208                 /* First entry */
 209                 st->current_prot = new_prot;
 210                 st->level = level;
 211                 st->marker = address_markers;
 212                 st->lines = 0;
 213                 pt_dump_seq_printf(m, st->to_dmesg, "---[ %s ]---\n",
 214                                    st->marker->name);
 215         } else if (prot != cur || level != st->level ||
 216                    st->current_address >= st->marker[1].start_address) {
 217                 const char *unit = units;
 218                 unsigned long delta;
 219                 int width = sizeof(unsigned long) * 2;
 220                 pgprotval_t pr = pgprot_val(st->current_prot);
 221
 222                 if (st->check_wx && (pr & _PAGE_RW) && !(pr & _PAGE_NX)) {
 223                         WARN_ONCE(1,
 224                                   "x86/mm: Found insecure W+X mapping at address %p/%pS\n",
 225                                   (void *)st->start_address,
 226                                   (void *)st->start_address);
 227                         st->wx_pages += (st->current_address -
 228                                          st->start_address) / PAGE_SIZE;
 229                 }
 230
 231                 /*
 232                  * Now print the actual finished series
 233                  */
 234                 if (!st->marker->max_lines ||
 235                     st->lines < st->marker->max_lines) {
 236                         pt_dump_seq_printf(m, st->to_dmesg,
 237                                            "0x%0*lx-0x%0*lx   ",
 238                                            width, st->start_address,
 239                                            width, st->current_address);
 240
 241                         delta = st->current_address - st->start_address;
 242                         while (!(delta & 1023) && unit[1]) {
 243                                 delta >>= 10;
 244                                 unit++;
 245                         }
 246                         pt_dump_cont_printf(m, st->to_dmesg, "%9lu%c ",
 247                                             delta, *unit);
 248                         printk_prot(m, st->current_prot, st->level,
 249                                     st->to_dmesg);
 250                 }
 251                 st->lines++;
 252
 253                 /*
 254                  * We print markers for special areas of address space,
 255                  * such as the start of vmalloc space etc.
 256                  * This helps in the interpretation.
 257                  */
 258                 if (st->current_address >= st->marker[1].start_address) {
 259                         if (st->marker->max_lines &&
 260                             st->lines > st->marker->max_lines) {
 261                                 unsigned long nskip =
 262                                         st->lines - st->marker->max_lines;
 263                                 pt_dump_seq_printf(m, st->to_dmesg,
 264                                                    "... %lu entr%s skipped ... \n",
 265                                                    nskip,
 266                                                    nskip == 1 ? "y" : "ies");
 267                         }
 268                         st->marker++;
 269                         st->lines = 0;
 270                         pt_dump_seq_printf(m, st->to_dmesg, "---[ %s ]---\n",
 271                                            st->marker->name);
 272                 }
 273
 274                 st->start_address = st->current_address;
 275                 st->current_prot = new_prot;
 276                 st->level = level;
 277         }
 278 }
 279
 280 static void walk_pte_level(struct seq_file *m, struct pg_state *st, pmd_t addr,
 281                                                         unsigned long P)
 282 {
 283         int i;
 284         pte_t *start;
 285         pgprotval_t prot;
 286
 287         start = (pte_t *) pmd_page_vaddr(addr);
 288         for (i = 0; i < PTRS_PER_PTE; i++) {
 289                 prot = pte_flags(*start);
 290                 st->current_address = normalize_addr(P + i * PTE_LEVEL_MULT);
 291                 note_page(m, st, __pgprot(prot), 4);
 292                 start++;
 293         }
 294 }
 295
 296 #if PTRS_PER_PMD > 1
 297
 298 static void walk_pmd_level(struct seq_file *m, struct pg_state *st, pud_t addr,
 299                                                         unsigned long P)
 300 {
 301         int i;
 302         pmd_t *start;
 303         pgprotval_t prot;
 304
 305         start = (pmd_t *) pud_page_vaddr(addr);
 306         for (i = 0; i < PTRS_PER_PMD; i++) {
 307                 st->current_address = normalize_addr(P + i * PMD_LEVEL_MULT);
 308                 if (!pmd_none(*start)) {
 309                         if (pmd_large(*start) || !pmd_present(*start)) {
 310                                 prot = pmd_flags(*start);
 311                                 note_page(m, st, __pgprot(prot), 3);
 312                         } else {
 313                                 walk_pte_level(m, st, *start,
 314                                                P + i * PMD_LEVEL_MULT);
 315                         }
 316                 } else
 317                         note_page(m, st, __pgprot(0), 3);
 318                 start++;
 319         }
 320 }
 321
 322 #else
 323 #define walk_pmd_level(m,s,a,p) walk_pte_level(m,s,__pmd(pud_val(a)),p)
 324 #define pud_large(a) pmd_large(__pmd(pud_val(a)))
 325 #define pud_none(a)  pmd_none(__pmd(pud_val(a)))
 326 #endif
 327
 328 #if PTRS_PER_PUD > 1
 329
 330 /*
 331  * This is an optimization for CONFIG_DEBUG_WX=y + CONFIG_KASAN=y
 332  * KASAN fills page tables with the same values. Since there is no
 333  * point in checking page table more than once we just skip repeated
 334  * entries. This saves us dozens of seconds during boot.
 335  */
 336 static bool pud_already_checked(pud_t *prev_pud, pud_t *pud, bool checkwx)
 337 {
 338         return checkwx && prev_pud && (pud_val(*prev_pud) == pud_val(*pud));
 339 }
 340
 341 static void walk_pud_level(struct seq_file *m, struct pg_state *st, pgd_t addr,
 342                                                         unsigned long P)
 343 {
 344         int i;
 345         pud_t *start;
 346         pgprotval_t prot;
 347         pud_t *prev_pud = NULL;
 348
 349         start = (pud_t *) pgd_page_vaddr(addr);
 350
 351         for (i = 0; i < PTRS_PER_PUD; i++) {
 352                 st->current_address = normalize_addr(P + i * PUD_LEVEL_MULT);
 353                 if (!pud_none(*start) &&
 354                     !pud_already_checked(prev_pud, start, st->check_wx)) {
 355                         if (pud_large(*start) || !pud_present(*start)) {
 356                                 prot = pud_flags(*start);
 357                                 note_page(m, st, __pgprot(prot), 2);
 358                         } else {
 359                                 walk_pmd_level(m, st, *start,
 360                                                P + i * PUD_LEVEL_MULT);
 361                         }
 362                 } else
 363                         note_page(m, st, __pgprot(0), 2);
 364
 365                 prev_pud = start;
 366                 start++;
 367         }
 368 }
 369
 370 #else
 371 #define walk_pud_level(m,s,a,p) walk_pmd_level(m,s,__pud(pgd_val(a)),p)
 372 #define pgd_large(a) pud_large(__pud(pgd_val(a)))
 373 #define pgd_none(a)  pud_none(__pud(pgd_val(a)))
 374 #endif
 375
 376 static inline bool is_hypervisor_range(int idx)
 377 {
 378 #ifdef CONFIG_X86_64
 379         /*
 380          * ffff800000000000 - ffff87ffffffffff is reserved for
 381          * the hypervisor.
 382          */
 383         return  (idx >= pgd_index(__PAGE_OFFSET) - 16) &&
 384                 (idx <  pgd_index(__PAGE_OFFSET));
 385 #else
 386         return false;
 387 #endif
 388 }
 389
 390 static void ptdump_walk_pgd_level_core(struct seq_file *m, pgd_t *pgd,
 391                                        bool checkwx)
 392 {
 393 #ifdef CONFIG_X86_64
 394         pgd_t *start = (pgd_t *) &init_level4_pgt;
 395 #else
 396         pgd_t *start = swapper_pg_dir;
 397 #endif
 398         pgprotval_t prot;
 399         int i;
 400         struct pg_state st = {};
 401
 402         if (pgd) {
 403                 start = pgd;
 404                 st.to_dmesg = true;
 405         }
 406
 407         st.check_wx = checkwx;
 408         if (checkwx)
 409                 st.wx_pages = 0;
 410
 411         for (i = 0; i < PTRS_PER_PGD; i++) {
 412                 st.current_address = normalize_addr(i * PGD_LEVEL_MULT);
 413                 if (!pgd_none(*start) && !is_hypervisor_range(i)) {
 414                         if (pgd_large(*start) || !pgd_present(*start)) {
 415                                 prot = pgd_flags(*start);
 416                                 note_page(m, &st, __pgprot(prot), 1);
 417                         } else {
 418                                 walk_pud_level(m, &st, *start,
 419                                                i * PGD_LEVEL_MULT);
 420                         }
 421                 } else
 422                         note_page(m, &st, __pgprot(0), 1);
 423
 424                 cond_resched();
 425                 start++;
 426         }
 427
 428         /* Flush out the last page */
 429         st.current_address = normalize_addr(PTRS_PER_PGD*PGD_LEVEL_MULT);
 430         note_page(m, &st, __pgprot(0), 0);
 431         if (!checkwx)
 432                 return;
 433         if (st.wx_pages)
 434                 pr_info("x86/mm: Checked W+X mappings: FAILED, %lu W+X pages found.\n",
 435                         st.wx_pages);
 436         else
 437                 pr_info("x86/mm: Checked W+X mappings: passed, no W+X pages found.\n");
 438 }
 439
 440 void ptdump_walk_pgd_level(struct seq_file *m, pgd_t *pgd)
 441 {
 442         ptdump_walk_pgd_level_core(m, pgd, false);
 443 }
 444 EXPORT_SYMBOL_GPL(ptdump_walk_pgd_level);
 445
 446 void ptdump_walk_pgd_level_checkwx(void)
 447 {
 448         ptdump_walk_pgd_level_core(NULL, NULL, true);
 449 }
 450
 451 static int __init pt_dump_init(void)
 452 {
 453         /*
 454          * Various markers are not compile-time constants, so assign them
 455          * here.
 456          */
 457 #ifdef CONFIG_X86_64
 458         address_markers[LOW_KERNEL_NR].start_address = PAGE_OFFSET;
 459         address_markers[VMALLOC_START_NR].start_address = VMALLOC_START;
 460         address_markers[VMEMMAP_START_NR].start_address = VMEMMAP_START;
 461 #endif
 462 #ifdef CONFIG_X86_32
 463         address_markers[VMALLOC_START_NR].start_address = VMALLOC_START;
 464         address_markers[VMALLOC_END_NR].start_address = VMALLOC_END;
 465 # ifdef CONFIG_HIGHMEM
 466         address_markers[PKMAP_BASE_NR].start_address = PKMAP_BASE;
 467 # endif
 468         address_markers[FIXADDR_START_NR].start_address = FIXADDR_START;
 469 #endif
 470
 471         return 0;
 472 }
 473 __initcall(pt_dump_init);