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         bool to_dmesg;
  35 };
  36
  37 struct addr_marker {
  38         unsigned long start_address;
  39         const char *name;
  40         unsigned long max_lines;
  41 };
  42
  43 /* indices for address_markers; keep sync'd w/ address_markers below */
  44 enum address_markers_idx {
  45         USER_SPACE_NR = 0,
  46 #ifdef CONFIG_X86_64
  47         KERNEL_SPACE_NR,
  48         LOW_KERNEL_NR,
  49         VMALLOC_START_NR,
  50         VMEMMAP_START_NR,
  51 # ifdef CONFIG_X86_ESPFIX64
  52         ESPFIX_START_NR,
  53 # endif
  54         HIGH_KERNEL_NR,
  55         MODULES_VADDR_NR,
  56         MODULES_END_NR,
  57 #else
  58         KERNEL_SPACE_NR,
  59         VMALLOC_START_NR,
  60         VMALLOC_END_NR,
  61 # ifdef CONFIG_HIGHMEM
  62         PKMAP_BASE_NR,
  63 # endif
  64         FIXADDR_START_NR,
  65 #endif
  66 };
  67
  68 /* Address space markers hints */
  69 static struct addr_marker address_markers[] = {
  70         { 0, "User Space" },
  71 #ifdef CONFIG_X86_64
  72         { 0x8000000000000000UL, "Kernel Space" },
  73         { PAGE_OFFSET,          "Low Kernel Mapping" },
  74         { VMALLOC_START,        "vmalloc() Area" },
  75         { VMEMMAP_START,        "Vmemmap" },
  76 # ifdef CONFIG_X86_ESPFIX64
  77         { ESPFIX_BASE_ADDR,     "ESPfix Area", 16 },
  78 # endif
  79         { __START_KERNEL_map,   "High Kernel Mapping" },
  80         { MODULES_VADDR,        "Modules" },
  81         { MODULES_END,          "End Modules" },
  82 #else
  83         { PAGE_OFFSET,          "Kernel Mapping" },
  84         { 0/* VMALLOC_START */, "vmalloc() Area" },
  85         { 0/*VMALLOC_END*/,     "vmalloc() End" },
  86 # ifdef CONFIG_HIGHMEM
  87         { 0/*PKMAP_BASE*/,      "Persisent kmap() Area" },
  88 # endif
  89         { 0/*FIXADDR_START*/,   "Fixmap Area" },
  90 #endif
  91         { -1, NULL }            /* End of list */
  92 };
  93
  94 /* Multipliers for offsets within the PTEs */
  95 #define PTE_LEVEL_MULT (PAGE_SIZE)
  96 #define PMD_LEVEL_MULT (PTRS_PER_PTE * PTE_LEVEL_MULT)
  97 #define PUD_LEVEL_MULT (PTRS_PER_PMD * PMD_LEVEL_MULT)
  98 #define PGD_LEVEL_MULT (PTRS_PER_PUD * PUD_LEVEL_MULT)
  99
 100 #define pt_dump_seq_printf(m, to_dmesg, fmt, args...)           \
 101 ({                                                              \
 102         if (to_dmesg)                                   \
 103                 printk(KERN_INFO fmt, ##args);                  \
 104         else                                                    \
 105                 if (m)                                          \
 106                         seq_printf(m, fmt, ##args);             \
 107 })
 108
 109 #define pt_dump_cont_printf(m, to_dmesg, fmt, args...)          \
 110 ({                                                              \
 111         if (to_dmesg)                                   \
 112                 printk(KERN_CONT fmt, ##args);                  \
 113         else                                                    \
 114                 if (m)                                          \
 115                         seq_printf(m, fmt, ##args);             \
 116 })
 117
 118 /*
 119  * Print a readable form of a pgprot_t to the seq_file
 120  */
 121 static void printk_prot(struct seq_file *m, pgprot_t prot, int level, bool dmsg)
 122 {
 123         pgprotval_t pr = pgprot_val(prot);
 124         static const char * const level_name[] =
 125                 { "cr3", "pgd", "pud", "pmd", "pte" };
 126
 127         if (!pgprot_val(prot)) {
 128                 /* Not present */
 129                 pt_dump_cont_printf(m, dmsg, "                              ");
 130         } else {
 131                 if (pr & _PAGE_USER)
 132                         pt_dump_cont_printf(m, dmsg, "USR ");
 133                 else
 134                         pt_dump_cont_printf(m, dmsg, "    ");
 135                 if (pr & _PAGE_RW)
 136                         pt_dump_cont_printf(m, dmsg, "RW ");
 137                 else
 138                         pt_dump_cont_printf(m, dmsg, "ro ");
 139                 if (pr & _PAGE_PWT)
 140                         pt_dump_cont_printf(m, dmsg, "PWT ");
 141                 else
 142                         pt_dump_cont_printf(m, dmsg, "    ");
 143                 if (pr & _PAGE_PCD)
 144                         pt_dump_cont_printf(m, dmsg, "PCD ");
 145                 else
 146                         pt_dump_cont_printf(m, dmsg, "    ");
 147
 148                 /* Bit 7 has a different meaning on level 3 vs 4 */
 149                 if (level <= 3 && pr & _PAGE_PSE)
 150                         pt_dump_cont_printf(m, dmsg, "PSE ");
 151                 else
 152                         pt_dump_cont_printf(m, dmsg, "    ");
 153                 if ((level == 4 && pr & _PAGE_PAT) ||
 154                     ((level == 3 || level == 2) && pr & _PAGE_PAT_LARGE))
 155                         pt_dump_cont_printf(m, dmsg, "pat ");
 156                 else
 157                         pt_dump_cont_printf(m, dmsg, "    ");
 158                 if (pr & _PAGE_GLOBAL)
 159                         pt_dump_cont_printf(m, dmsg, "GLB ");
 160                 else
 161                         pt_dump_cont_printf(m, dmsg, "    ");
 162                 if (pr & _PAGE_NX)
 163                         pt_dump_cont_printf(m, dmsg, "NX ");
 164                 else
 165                         pt_dump_cont_printf(m, dmsg, "x  ");
 166         }
 167         pt_dump_cont_printf(m, dmsg, "%s\n", level_name[level]);
 168 }
 169
 170 /*
 171  * On 64 bits, sign-extend the 48 bit address to 64 bit
 172  */
 173 static unsigned long normalize_addr(unsigned long u)
 174 {
 175 #ifdef CONFIG_X86_64
 176         return (signed long)(u << 16) >> 16;
 177 #else
 178         return u;
 179 #endif
 180 }
 181
 182 /*
 183  * This function gets called on a break in a continuous series
 184  * of PTE entries; the next one is different so we need to
 185  * print what we collected so far.
 186  */
 187 static void note_page(struct seq_file *m, struct pg_state *st,
 188                       pgprot_t new_prot, int level)
 189 {
 190         pgprotval_t prot, cur;
 191         static const char units[] = "BKMGTPE";
 192
 193         /*
 194          * If we have a "break" in the series, we need to flush the state that
 195          * we have now. "break" is either changing perms, levels or
 196          * address space marker.
 197          */
 198         prot = pgprot_val(new_prot) & PTE_FLAGS_MASK;
 199         cur = pgprot_val(st->current_prot) & PTE_FLAGS_MASK;
 200
 201         if (!st->level) {
 202                 /* First entry */
 203                 st->current_prot = new_prot;
 204                 st->level = level;
 205                 st->marker = address_markers;
 206                 st->lines = 0;
 207                 pt_dump_seq_printf(m, st->to_dmesg, "---[ %s ]---\n",
 208                                    st->marker->name);
 209         } else if (prot != cur || level != st->level ||
 210                    st->current_address >= st->marker[1].start_address) {
 211                 const char *unit = units;
 212                 unsigned long delta;
 213                 int width = sizeof(unsigned long) * 2;
 214
 215                 /*
 216                  * Now print the actual finished series
 217                  */
 218                 if (!st->marker->max_lines ||
 219                     st->lines < st->marker->max_lines) {
 220                         pt_dump_seq_printf(m, st->to_dmesg,
 221                                            "0x%0*lx-0x%0*lx   ",
 222                                            width, st->start_address,
 223                                            width, st->current_address);
 224
 225                         delta = st->current_address - st->start_address;
 226                         while (!(delta & 1023) && unit[1]) {
 227                                 delta >>= 10;
 228                                 unit++;
 229                         }
 230                         pt_dump_cont_printf(m, st->to_dmesg, "%9lu%c ",
 231                                             delta, *unit);
 232                         printk_prot(m, st->current_prot, st->level,
 233                                     st->to_dmesg);
 234                 }
 235                 st->lines++;
 236
 237                 /*
 238                  * We print markers for special areas of address space,
 239                  * such as the start of vmalloc space etc.
 240                  * This helps in the interpretation.
 241                  */
 242                 if (st->current_address >= st->marker[1].start_address) {
 243                         if (st->marker->max_lines &&
 244                             st->lines > st->marker->max_lines) {
 245                                 unsigned long nskip =
 246                                         st->lines - st->marker->max_lines;
 247                                 pt_dump_seq_printf(m, st->to_dmesg,
 248                                                    "... %lu entr%s skipped ... \n",
 249                                                    nskip,
 250                                                    nskip == 1 ? "y" : "ies");
 251                         }
 252                         st->marker++;
 253                         st->lines = 0;
 254                         pt_dump_seq_printf(m, st->to_dmesg, "---[ %s ]---\n",
 255                                            st->marker->name);
 256                 }
 257
 258                 st->start_address = st->current_address;
 259                 st->current_prot = new_prot;
 260                 st->level = level;
 261         }
 262 }
 263
 264 static void walk_pte_level(struct seq_file *m, struct pg_state *st, pmd_t addr,
 265                                                         unsigned long P)
 266 {
 267         int i;
 268         pte_t *start;
 269
 270         start = (pte_t *) pmd_page_vaddr(addr);
 271         for (i = 0; i < PTRS_PER_PTE; i++) {
 272                 pgprot_t prot = pte_pgprot(*start);
 273
 274                 st->current_address = normalize_addr(P + i * PTE_LEVEL_MULT);
 275                 note_page(m, st, prot, 4);
 276                 start++;
 277         }
 278 }
 279
 280 #if PTRS_PER_PMD > 1
 281
 282 static void walk_pmd_level(struct seq_file *m, struct pg_state *st, pud_t addr,
 283                                                         unsigned long P)
 284 {
 285         int i;
 286         pmd_t *start;
 287
 288         start = (pmd_t *) pud_page_vaddr(addr);
 289         for (i = 0; i < PTRS_PER_PMD; i++) {
 290                 st->current_address = normalize_addr(P + i * PMD_LEVEL_MULT);
 291                 if (!pmd_none(*start)) {
 292                         pgprotval_t prot = pmd_val(*start) & PTE_FLAGS_MASK;
 293
 294                         if (pmd_large(*start) || !pmd_present(*start))
 295                                 note_page(m, st, __pgprot(prot), 3);
 296                         else
 297                                 walk_pte_level(m, st, *start,
 298                                                P + i * PMD_LEVEL_MULT);
 299                 } else
 300                         note_page(m, st, __pgprot(0), 3);
 301                 start++;
 302         }
 303 }
 304
 305 #else
 306 #define walk_pmd_level(m,s,a,p) walk_pte_level(m,s,__pmd(pud_val(a)),p)
 307 #define pud_large(a) pmd_large(__pmd(pud_val(a)))
 308 #define pud_none(a)  pmd_none(__pmd(pud_val(a)))
 309 #endif
 310
 311 #if PTRS_PER_PUD > 1
 312
 313 static void walk_pud_level(struct seq_file *m, struct pg_state *st, pgd_t addr,
 314                                                         unsigned long P)
 315 {
 316         int i;
 317         pud_t *start;
 318
 319         start = (pud_t *) pgd_page_vaddr(addr);
 320
 321         for (i = 0; i < PTRS_PER_PUD; i++) {
 322                 st->current_address = normalize_addr(P + i * PUD_LEVEL_MULT);
 323                 if (!pud_none(*start)) {
 324                         pgprotval_t prot = pud_val(*start) & PTE_FLAGS_MASK;
 325
 326                         if (pud_large(*start) || !pud_present(*start))
 327                                 note_page(m, st, __pgprot(prot), 2);
 328                         else
 329                                 walk_pmd_level(m, st, *start,
 330                                                P + i * PUD_LEVEL_MULT);
 331                 } else
 332                         note_page(m, st, __pgprot(0), 2);
 333
 334                 start++;
 335         }
 336 }
 337
 338 #else
 339 #define walk_pud_level(m,s,a,p) walk_pmd_level(m,s,__pud(pgd_val(a)),p)
 340 #define pgd_large(a) pud_large(__pud(pgd_val(a)))
 341 #define pgd_none(a)  pud_none(__pud(pgd_val(a)))
 342 #endif
 343
 344 void ptdump_walk_pgd_level(struct seq_file *m, pgd_t *pgd)
 345 {
 346 #ifdef CONFIG_X86_64
 347         pgd_t *start = (pgd_t *) &init_level4_pgt;
 348 #else
 349         pgd_t *start = swapper_pg_dir;
 350 #endif
 351         int i;
 352         struct pg_state st = {};
 353
 354         if (pgd) {
 355                 start = pgd;
 356                 st.to_dmesg = true;
 357         }
 358
 359         for (i = 0; i < PTRS_PER_PGD; i++) {
 360                 st.current_address = normalize_addr(i * PGD_LEVEL_MULT);
 361                 if (!pgd_none(*start)) {
 362                         pgprotval_t prot = pgd_val(*start) & PTE_FLAGS_MASK;
 363
 364                         if (pgd_large(*start) || !pgd_present(*start))
 365                                 note_page(m, &st, __pgprot(prot), 1);
 366                         else
 367                                 walk_pud_level(m, &st, *start,
 368                                                i * PGD_LEVEL_MULT);
 369                 } else
 370                         note_page(m, &st, __pgprot(0), 1);
 371
 372                 start++;
 373         }
 374
 375         /* Flush out the last page */
 376         st.current_address = normalize_addr(PTRS_PER_PGD*PGD_LEVEL_MULT);
 377         note_page(m, &st, __pgprot(0), 0);
 378 }
 379
 380 static int ptdump_show(struct seq_file *m, void *v)
 381 {
 382         ptdump_walk_pgd_level(m, NULL);
 383         return 0;
 384 }
 385
 386 static int ptdump_open(struct inode *inode, struct file *filp)
 387 {
 388         return single_open(filp, ptdump_show, NULL);
 389 }
 390
 391 static const struct file_operations ptdump_fops = {
 392         .open           = ptdump_open,
 393         .read           = seq_read,
 394         .llseek         = seq_lseek,
 395         .release        = single_release,
 396 };
 397
 398 static int pt_dump_init(void)
 399 {
 400         struct dentry *pe;
 401
 402 #ifdef CONFIG_X86_32
 403         /* Not a compile-time constant on x86-32 */
 404         address_markers[VMALLOC_START_NR].start_address = VMALLOC_START;
 405         address_markers[VMALLOC_END_NR].start_address = VMALLOC_END;
 406 # ifdef CONFIG_HIGHMEM
 407         address_markers[PKMAP_BASE_NR].start_address = PKMAP_BASE;
 408 # endif
 409         address_markers[FIXADDR_START_NR].start_address = FIXADDR_START;
 410 #endif
 411
 412         pe = debugfs_create_file("kernel_page_tables", 0600, NULL, NULL,
 413                                  &ptdump_fops);
 414         if (!pe)
 415                 return -ENOMEM;
 416
 417         return 0;
 418 }
 419
 420 __initcall(pt_dump_init);
 421 MODULE_LICENSE("GPL");
 422 MODULE_AUTHOR("Arjan van de Ven <arjan@linux.intel.com>");
 423 MODULE_DESCRIPTION("Kernel debugging helper that dumps pagetables");