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