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/module.h>
  18 #include <linux/seq_file.h>
  19
  20 #include <asm/pgtable.h>
  21
  22 /*
  23  * The dumper groups pagetable entries of the same type into one, and for
  24  * that it needs to keep some state when walking, and flush this state
  25  * when a "break" in the continuity is found.
  26  */
  27 struct pg_state {
  28         int level;
  29         pgprot_t current_prot;
  30         unsigned long start_address;
  31         unsigned long current_address;
  32         const struct addr_marker *marker;
  33         unsigned long lines;
  34 };
  35
  36 struct addr_marker {
  37         unsigned long start_address;
  38         const char *name;
  39         unsigned long max_lines;
  40 };
  41
  42 /* indices for address_markers; keep sync'd w/ address_markers below */
  43 enum address_markers_idx {
  44         USER_SPACE_NR = 0,
  45 #ifdef CONFIG_X86_64
  46         KERNEL_SPACE_NR,
  47         LOW_KERNEL_NR,
  48         VMALLOC_START_NR,
  49         VMEMMAP_START_NR,
  50         ESPFIX_START_NR,
  51         HIGH_KERNEL_NR,
  52         MODULES_VADDR_NR,
  53         MODULES_END_NR,
  54 #else
  55         KERNEL_SPACE_NR,
  56         VMALLOC_START_NR,
  57         VMALLOC_END_NR,
  58 # ifdef CONFIG_HIGHMEM
  59         PKMAP_BASE_NR,
  60 # endif
  61         FIXADDR_START_NR,
  62 #endif
  63 };
  64
  65 /* Address space markers hints */
  66 static struct addr_marker address_markers[] = {
  67         { 0, "User Space" },
  68 #ifdef CONFIG_X86_64
  69         { 0x8000000000000000UL, "Kernel Space" },
  70         { PAGE_OFFSET,          "Low Kernel Mapping" },
  71         { VMALLOC_START,        "vmalloc() Area" },
  72         { VMEMMAP_START,        "Vmemmap" },
  73         { ESPFIX_BASE_ADDR,     "ESPfix Area", 16 },
  74         { __START_KERNEL_map,   "High Kernel Mapping" },
  75         { MODULES_VADDR,        "Modules" },
  76         { MODULES_END,          "End Modules" },
  77 #else
  78         { PAGE_OFFSET,          "Kernel Mapping" },
  79         { 0/* VMALLOC_START */, "vmalloc() Area" },
  80         { 0/*VMALLOC_END*/,     "vmalloc() End" },
  81 # ifdef CONFIG_HIGHMEM
  82         { 0/*PKMAP_BASE*/,      "Persisent kmap() Area" },
  83 # endif
  84         { 0/*FIXADDR_START*/,   "Fixmap Area" },
  85 #endif
  86         { -1, NULL }            /* End of list */
  87 };
  88
  89 /* Multipliers for offsets within the PTEs */
  90 #define PTE_LEVEL_MULT (PAGE_SIZE)
  91 #define PMD_LEVEL_MULT (PTRS_PER_PTE * PTE_LEVEL_MULT)
  92 #define PUD_LEVEL_MULT (PTRS_PER_PMD * PMD_LEVEL_MULT)
  93 #define PGD_LEVEL_MULT (PTRS_PER_PUD * PUD_LEVEL_MULT)
  94
  95 /*
  96  * Print a readable form of a pgprot_t to the seq_file
  97  */
  98 static void printk_prot(struct seq_file *m, pgprot_t prot, int level)
  99 {
 100         pgprotval_t pr = pgprot_val(prot);
 101         static const char * const level_name[] =
 102                 { "cr3", "pgd", "pud", "pmd", "pte" };
 103
 104         if (!pgprot_val(prot)) {
 105                 /* Not present */
 106                 seq_printf(m, "                          ");
 107         } else {
 108                 if (pr & _PAGE_USER)
 109                         seq_printf(m, "USR ");
 110                 else
 111                         seq_printf(m, "    ");
 112                 if (pr & _PAGE_RW)
 113                         seq_printf(m, "RW ");
 114                 else
 115                         seq_printf(m, "ro ");
 116                 if (pr & _PAGE_PWT)
 117                         seq_printf(m, "PWT ");
 118                 else
 119                         seq_printf(m, "    ");
 120                 if (pr & _PAGE_PCD)
 121                         seq_printf(m, "PCD ");
 122                 else
 123                         seq_printf(m, "    ");
 124
 125                 /* Bit 9 has a different meaning on level 3 vs 4 */
 126                 if (level <= 3) {
 127                         if (pr & _PAGE_PSE)
 128                                 seq_printf(m, "PSE ");
 129                         else
 130                                 seq_printf(m, "    ");
 131                 } else {
 132                         if (pr & _PAGE_PAT)
 133                                 seq_printf(m, "pat ");
 134                         else
 135                                 seq_printf(m, "    ");
 136                 }
 137                 if (pr & _PAGE_GLOBAL)
 138                         seq_printf(m, "GLB ");
 139                 else
 140                         seq_printf(m, "    ");
 141                 if (pr & _PAGE_NX)
 142                         seq_printf(m, "NX ");
 143                 else
 144                         seq_printf(m, "x  ");
 145         }
 146         seq_printf(m, "%s\n", level_name[level]);
 147 }
 148
 149 /*
 150  * On 64 bits, sign-extend the 48 bit address to 64 bit
 151  */
 152 static unsigned long normalize_addr(unsigned long u)
 153 {
 154 #ifdef CONFIG_X86_64
 155         return (signed long)(u << 16) >> 16;
 156 #else
 157         return u;
 158 #endif
 159 }
 160
 161 /*
 162  * This function gets called on a break in a continuous series
 163  * of PTE entries; the next one is different so we need to
 164  * print what we collected so far.
 165  */
 166 static void note_page(struct seq_file *m, struct pg_state *st,
 167                       pgprot_t new_prot, int level)
 168 {
 169         pgprotval_t prot, cur;
 170         static const char units[] = "BKMGTPE";
 171
 172         /*
 173          * If we have a "break" in the series, we need to flush the state that
 174          * we have now. "break" is either changing perms, levels or
 175          * address space marker.
 176          */
 177         prot = pgprot_val(new_prot) & PTE_FLAGS_MASK;
 178         cur = pgprot_val(st->current_prot) & PTE_FLAGS_MASK;
 179
 180         if (!st->level) {
 181                 /* First entry */
 182                 st->current_prot = new_prot;
 183                 st->level = level;
 184                 st->marker = address_markers;
 185                 st->lines = 0;
 186                 seq_printf(m, "---[ %s ]---\n", st->marker->name);
 187         } else if (prot != cur || level != st->level ||
 188                    st->current_address >= st->marker[1].start_address) {
 189                 const char *unit = units;
 190                 unsigned long delta;
 191                 int width = sizeof(unsigned long) * 2;
 192
 193                 /*
 194                  * Now print the actual finished series
 195                  */
 196                 if (!st->marker->max_lines ||
 197                     st->lines < st->marker->max_lines) {
 198                         seq_printf(m, "0x%0*lx-0x%0*lx   ",
 199                                    width, st->start_address,
 200                                    width, st->current_address);
 201
 202                         delta = (st->current_address - st->start_address);
 203                         while (!(delta & 1023) && unit[1]) {
 204                                 delta >>= 10;
 205                                 unit++;
 206                         }
 207                         seq_printf(m, "%9lu%c ", delta, *unit);
 208                         printk_prot(m, st->current_prot, st->level);
 209                 }
 210                 st->lines++;
 211
 212                 /*
 213                  * We print markers for special areas of address space,
 214                  * such as the start of vmalloc space etc.
 215                  * This helps in the interpretation.
 216                  */
 217                 if (st->current_address >= st->marker[1].start_address) {
 218                         if (st->marker->max_lines &&
 219                             st->lines > st->marker->max_lines) {
 220                                 unsigned long nskip =
 221                                         st->lines - st->marker->max_lines;
 222                                 seq_printf(m, "... %lu entr%s skipped ... \n",
 223                                            nskip, nskip == 1 ? "y" : "ies");
 224                         }
 225                         st->marker++;
 226                         st->lines = 0;
 227                         seq_printf(m, "---[ %s ]---\n", st->marker->name);
 228                 }
 229
 230                 st->start_address = st->current_address;
 231                 st->current_prot = new_prot;
 232                 st->level = level;
 233         }
 234 }
 235
 236 static void walk_pte_level(struct seq_file *m, struct pg_state *st, pmd_t addr,
 237                                                         unsigned long P)
 238 {
 239         int i;
 240         pte_t *start;
 241
 242         start = (pte_t *) pmd_page_vaddr(addr);
 243         for (i = 0; i < PTRS_PER_PTE; i++) {
 244                 pgprot_t prot = pte_pgprot(*start);
 245
 246                 st->current_address = normalize_addr(P + i * PTE_LEVEL_MULT);
 247                 note_page(m, st, prot, 4);
 248                 start++;
 249         }
 250 }
 251
 252 #if PTRS_PER_PMD > 1
 253
 254 static void walk_pmd_level(struct seq_file *m, struct pg_state *st, pud_t addr,
 255                                                         unsigned long P)
 256 {
 257         int i;
 258         pmd_t *start;
 259
 260         start = (pmd_t *) pud_page_vaddr(addr);
 261         for (i = 0; i < PTRS_PER_PMD; i++) {
 262                 st->current_address = normalize_addr(P + i * PMD_LEVEL_MULT);
 263                 if (!pmd_none(*start)) {
 264                         pgprotval_t prot = pmd_val(*start) & PTE_FLAGS_MASK;
 265
 266                         if (pmd_large(*start) || !pmd_present(*start))
 267                                 note_page(m, st, __pgprot(prot), 3);
 268                         else
 269                                 walk_pte_level(m, st, *start,
 270                                                P + i * PMD_LEVEL_MULT);
 271                 } else
 272                         note_page(m, st, __pgprot(0), 3);
 273                 start++;
 274         }
 275 }
 276
 277 #else
 278 #define walk_pmd_level(m,s,a,p) walk_pte_level(m,s,__pmd(pud_val(a)),p)
 279 #define pud_large(a) pmd_large(__pmd(pud_val(a)))
 280 #define pud_none(a)  pmd_none(__pmd(pud_val(a)))
 281 #endif
 282
 283 #if PTRS_PER_PUD > 1
 284
 285 static void walk_pud_level(struct seq_file *m, struct pg_state *st, pgd_t addr,
 286                                                         unsigned long P)
 287 {
 288         int i;
 289         pud_t *start;
 290
 291         start = (pud_t *) pgd_page_vaddr(addr);
 292
 293         for (i = 0; i < PTRS_PER_PUD; i++) {
 294                 st->current_address = normalize_addr(P + i * PUD_LEVEL_MULT);
 295                 if (!pud_none(*start)) {
 296                         pgprotval_t prot = pud_val(*start) & PTE_FLAGS_MASK;
 297
 298                         if (pud_large(*start) || !pud_present(*start))
 299                                 note_page(m, st, __pgprot(prot), 2);
 300                         else
 301                                 walk_pmd_level(m, st, *start,
 302                                                P + i * PUD_LEVEL_MULT);
 303                 } else
 304                         note_page(m, st, __pgprot(0), 2);
 305
 306                 start++;
 307         }
 308 }
 309
 310 #else
 311 #define walk_pud_level(m,s,a,p) walk_pmd_level(m,s,__pud(pgd_val(a)),p)
 312 #define pgd_large(a) pud_large(__pud(pgd_val(a)))
 313 #define pgd_none(a)  pud_none(__pud(pgd_val(a)))
 314 #endif
 315
 316 static void walk_pgd_level(struct seq_file *m)
 317 {
 318 #ifdef CONFIG_X86_64
 319         pgd_t *start = (pgd_t *) &init_level4_pgt;
 320 #else
 321         pgd_t *start = swapper_pg_dir;
 322 #endif
 323         int i;
 324         struct pg_state st;
 325
 326         memset(&st, 0, sizeof(st));
 327
 328         for (i = 0; i < PTRS_PER_PGD; i++) {
 329                 st.current_address = normalize_addr(i * PGD_LEVEL_MULT);
 330                 if (!pgd_none(*start)) {
 331                         pgprotval_t prot = pgd_val(*start) & PTE_FLAGS_MASK;
 332
 333                         if (pgd_large(*start) || !pgd_present(*start))
 334                                 note_page(m, &st, __pgprot(prot), 1);
 335                         else
 336                                 walk_pud_level(m, &st, *start,
 337                                                i * PGD_LEVEL_MULT);
 338                 } else
 339                         note_page(m, &st, __pgprot(0), 1);
 340
 341                 start++;
 342         }
 343
 344         /* Flush out the last page */
 345         st.current_address = normalize_addr(PTRS_PER_PGD*PGD_LEVEL_MULT);
 346         note_page(m, &st, __pgprot(0), 0);
 347 }
 348
 349 static int ptdump_show(struct seq_file *m, void *v)
 350 {
 351         walk_pgd_level(m);
 352         return 0;
 353 }
 354
 355 static int ptdump_open(struct inode *inode, struct file *filp)
 356 {
 357         return single_open(filp, ptdump_show, NULL);
 358 }
 359
 360 static const struct file_operations ptdump_fops = {
 361         .open           = ptdump_open,
 362         .read           = seq_read,
 363         .llseek         = seq_lseek,
 364         .release        = single_release,
 365 };
 366
 367 static int pt_dump_init(void)
 368 {
 369         struct dentry *pe;
 370
 371 #ifdef CONFIG_X86_32
 372         /* Not a compile-time constant on x86-32 */
 373         address_markers[VMALLOC_START_NR].start_address = VMALLOC_START;
 374         address_markers[VMALLOC_END_NR].start_address = VMALLOC_END;
 375 # ifdef CONFIG_HIGHMEM
 376         address_markers[PKMAP_BASE_NR].start_address = PKMAP_BASE;
 377 # endif
 378         address_markers[FIXADDR_START_NR].start_address = FIXADDR_START;
 379 #endif
 380
 381         pe = debugfs_create_file("kernel_page_tables", 0600, NULL, NULL,
 382                                  &ptdump_fops);
 383         if (!pe)
 384                 return -ENOMEM;
 385
 386         return 0;
 387 }
 388
 389 __initcall(pt_dump_init);
 390 MODULE_LICENSE("GPL");
 391 MODULE_AUTHOR("Arjan van de Ven <arjan@linux.intel.com>");
 392 MODULE_DESCRIPTION("Kernel debugging helper that dumps pagetables");