arch/s390/mm/dump_pagetables.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/set_memory.h>
   3 #include <linux/ptdump.h>
   4 #include <linux/seq_file.h>
   5 #include <linux/debugfs.h>
   6 #include <linux/mm.h>
   7 #include <linux/kfence.h>
   8 #include <linux/kasan.h>
   9 #include <asm/ptdump.h>
  10 #include <asm/kasan.h>
  11 #include <asm/nospec-branch.h>
  12 #include <asm/sections.h>
  13
  14 static unsigned long max_addr;
  15
  16 struct addr_marker {
  17         unsigned long start_address;
  18         const char *name;
  19 };
  20
  21 enum address_markers_idx {
  22         IDENTITY_BEFORE_NR = 0,
  23         IDENTITY_BEFORE_END_NR,
  24         KERNEL_START_NR,
  25         KERNEL_END_NR,
  26 #ifdef CONFIG_KFENCE
  27         KFENCE_START_NR,
  28         KFENCE_END_NR,
  29 #endif
  30         IDENTITY_AFTER_NR,
  31         IDENTITY_AFTER_END_NR,
  32 #ifdef CONFIG_KASAN
  33         KASAN_SHADOW_START_NR,
  34         KASAN_SHADOW_END_NR,
  35 #endif
  36         VMEMMAP_NR,
  37         VMEMMAP_END_NR,
  38         VMALLOC_NR,
  39         VMALLOC_END_NR,
  40         MODULES_NR,
  41         MODULES_END_NR,
  42 };
  43
  44 static struct addr_marker address_markers[] = {
  45         [IDENTITY_BEFORE_NR]    = {0, "Identity Mapping Start"},
  46         [IDENTITY_BEFORE_END_NR] = {(unsigned long)_stext, "Identity Mapping End"},
  47         [KERNEL_START_NR]       = {(unsigned long)_stext, "Kernel Image Start"},
  48         [KERNEL_END_NR]         = {(unsigned long)_end, "Kernel Image End"},
  49 #ifdef CONFIG_KFENCE
  50         [KFENCE_START_NR]       = {0, "KFence Pool Start"},
  51         [KFENCE_END_NR]         = {0, "KFence Pool End"},
  52 #endif
  53         [IDENTITY_AFTER_NR]     = {(unsigned long)_end, "Identity Mapping Start"},
  54         [IDENTITY_AFTER_END_NR] = {0, "Identity Mapping End"},
  55 #ifdef CONFIG_KASAN
  56         [KASAN_SHADOW_START_NR] = {KASAN_SHADOW_START, "Kasan Shadow Start"},
  57         [KASAN_SHADOW_END_NR]   = {KASAN_SHADOW_END, "Kasan Shadow End"},
  58 #endif
  59         [VMEMMAP_NR]            = {0, "vmemmap Area Start"},
  60         [VMEMMAP_END_NR]        = {0, "vmemmap Area End"},
  61         [VMALLOC_NR]            = {0, "vmalloc Area Start"},
  62         [VMALLOC_END_NR]        = {0, "vmalloc Area End"},
  63         [MODULES_NR]            = {0, "Modules Area Start"},
  64         [MODULES_END_NR]        = {0, "Modules Area End"},
  65         { -1, NULL }
  66 };
  67
  68 struct pg_state {
  69         struct ptdump_state ptdump;
  70         struct seq_file *seq;
  71         int level;
  72         unsigned int current_prot;
  73         bool check_wx;
  74         unsigned long wx_pages;
  75         unsigned long start_address;
  76         const struct addr_marker *marker;
  77 };
  78
  79 #define pt_dump_seq_printf(m, fmt, args...)     \
  80 ({                                              \
  81         struct seq_file *__m = (m);             \
  82                                                 \
  83         if (__m)                                \
  84                 seq_printf(__m, fmt, ##args);   \
  85 })
  86
  87 #define pt_dump_seq_puts(m, fmt)                \
  88 ({                                              \
  89         struct seq_file *__m = (m);             \
  90                                                 \
  91         if (__m)                                \
  92                 seq_printf(__m, fmt);           \
  93 })
  94
  95 static void print_prot(struct seq_file *m, unsigned int pr, int level)
  96 {
  97         static const char * const level_name[] =
  98                 { "ASCE", "PGD", "PUD", "PMD", "PTE" };
  99
 100         pt_dump_seq_printf(m, "%s ", level_name[level]);
 101         if (pr & _PAGE_INVALID) {
 102                 pt_dump_seq_printf(m, "I\n");
 103                 return;
 104         }
 105         pt_dump_seq_puts(m, (pr & _PAGE_PROTECT) ? "RO " : "RW ");
 106         pt_dump_seq_puts(m, (pr & _PAGE_NOEXEC) ? "NX\n" : "X\n");
 107 }
 108
 109 static void note_prot_wx(struct pg_state *st, unsigned long addr)
 110 {
 111 #ifdef CONFIG_DEBUG_WX
 112         if (!st->check_wx)
 113                 return;
 114         if (st->current_prot & _PAGE_INVALID)
 115                 return;
 116         if (st->current_prot & _PAGE_PROTECT)
 117                 return;
 118         if (st->current_prot & _PAGE_NOEXEC)
 119                 return;
 120         /*
 121          * The first lowcore page is W+X if spectre mitigations are using
 122          * trampolines or the BEAR enhancements facility is not installed,
 123          * in which case we have two lpswe instructions in lowcore that need
 124          * to be executable.
 125          */
 126         if (addr == PAGE_SIZE && (nospec_uses_trampoline() || !static_key_enabled(&cpu_has_bear)))
 127                 return;
 128         WARN_ONCE(1, "s390/mm: Found insecure W+X mapping at address %pS\n",
 129                   (void *)st->start_address);
 130         st->wx_pages += (addr - st->start_address) / PAGE_SIZE;
 131 #endif /* CONFIG_DEBUG_WX */
 132 }
 133
 134 static void note_page(struct ptdump_state *pt_st, unsigned long addr, int level, u64 val)
 135 {
 136         int width = sizeof(unsigned long) * 2;
 137         static const char units[] = "KMGTPE";
 138         const char *unit = units;
 139         unsigned long delta;
 140         struct pg_state *st;
 141         struct seq_file *m;
 142         unsigned int prot;
 143
 144         st = container_of(pt_st, struct pg_state, ptdump);
 145         m = st->seq;
 146         prot = val & (_PAGE_PROTECT | _PAGE_NOEXEC);
 147         if (level == 4 && (val & _PAGE_INVALID))
 148                 prot = _PAGE_INVALID;
 149         /* For pmd_none() & friends val gets passed as zero. */
 150         if (level != 4 && !val)
 151                 prot = _PAGE_INVALID;
 152         /* Final flush from generic code. */
 153         if (level == -1)
 154                 addr = max_addr;
 155         if (st->level == -1) {
 156                 pt_dump_seq_printf(m, "---[ %s ]---\n", st->marker->name);
 157                 st->start_address = addr;
 158                 st->current_prot = prot;
 159                 st->level = level;
 160         } else if (prot != st->current_prot || level != st->level ||
 161                    addr >= st->marker[1].start_address) {
 162                 note_prot_wx(st, addr);
 163                 pt_dump_seq_printf(m, "0x%0*lx-0x%0*lx ",
 164                                    width, st->start_address,
 165                                    width, addr);
 166                 delta = (addr - st->start_address) >> 10;
 167                 while (!(delta & 0x3ff) && unit[1]) {
 168                         delta >>= 10;
 169                         unit++;
 170                 }
 171                 pt_dump_seq_printf(m, "%9lu%c ", delta, *unit);
 172                 print_prot(m, st->current_prot, st->level);
 173                 while (addr >= st->marker[1].start_address) {
 174                         st->marker++;
 175                         pt_dump_seq_printf(m, "---[ %s ]---\n", st->marker->name);
 176                 }
 177                 st->start_address = addr;
 178                 st->current_prot = prot;
 179                 st->level = level;
 180         }
 181 }
 182
 183 #ifdef CONFIG_DEBUG_WX
 184 void ptdump_check_wx(void)
 185 {
 186         struct pg_state st = {
 187                 .ptdump = {
 188                         .note_page = note_page,
 189                         .range = (struct ptdump_range[]) {
 190                                 {.start = 0, .end = max_addr},
 191                                 {.start = 0, .end = 0},
 192                         }
 193                 },
 194                 .seq = NULL,
 195                 .level = -1,
 196                 .current_prot = 0,
 197                 .check_wx = true,
 198                 .wx_pages = 0,
 199                 .start_address = 0,
 200                 .marker = (struct addr_marker[]) {
 201                         { .start_address =  0, .name = NULL},
 202                         { .start_address = -1, .name = NULL},
 203                 },
 204         };
 205
 206         if (!MACHINE_HAS_NX)
 207                 return;
 208         ptdump_walk_pgd(&st.ptdump, &init_mm, NULL);
 209         if (st.wx_pages)
 210                 pr_warn("Checked W+X mappings: FAILED, %lu W+X pages found\n", st.wx_pages);
 211         else
 212                 pr_info("Checked W+X mappings: passed, no %sW+X pages found\n",
 213                         (nospec_uses_trampoline() || !static_key_enabled(&cpu_has_bear)) ?
 214                         "unexpected " : "");
 215 }
 216 #endif /* CONFIG_DEBUG_WX */
 217
 218 #ifdef CONFIG_PTDUMP_DEBUGFS
 219 static int ptdump_show(struct seq_file *m, void *v)
 220 {
 221         struct pg_state st = {
 222                 .ptdump = {
 223                         .note_page = note_page,
 224                         .range = (struct ptdump_range[]) {
 225                                 {.start = 0, .end = max_addr},
 226                                 {.start = 0, .end = 0},
 227                         }
 228                 },
 229                 .seq = m,
 230                 .level = -1,
 231                 .current_prot = 0,
 232                 .check_wx = false,
 233                 .wx_pages = 0,
 234                 .start_address = 0,
 235                 .marker = address_markers,
 236         };
 237
 238         get_online_mems();
 239         mutex_lock(&cpa_mutex);
 240         ptdump_walk_pgd(&st.ptdump, &init_mm, NULL);
 241         mutex_unlock(&cpa_mutex);
 242         put_online_mems();
 243         return 0;
 244 }
 245 DEFINE_SHOW_ATTRIBUTE(ptdump);
 246 #endif /* CONFIG_PTDUMP_DEBUGFS */
 247
 248 /*
 249  * Heapsort from lib/sort.c is not a stable sorting algorithm, do a simple
 250  * insertion sort to preserve the original order of markers with the same
 251  * start address.
 252  */
 253 static void sort_address_markers(void)
 254 {
 255         struct addr_marker tmp;
 256         int i, j;
 257
 258         for (i = 1; i < ARRAY_SIZE(address_markers) - 1; i++) {
 259                 tmp = address_markers[i];
 260                 for (j = i - 1; j >= 0 && address_markers[j].start_address > tmp.start_address; j--)
 261                         address_markers[j + 1] = address_markers[j];
 262                 address_markers[j + 1] = tmp;
 263         }
 264 }
 265
 266 static int pt_dump_init(void)
 267 {
 268 #ifdef CONFIG_KFENCE
 269         unsigned long kfence_start = (unsigned long)__kfence_pool;
 270 #endif
 271         /*
 272          * Figure out the maximum virtual address being accessible with the
 273          * kernel ASCE. We need this to keep the page table walker functions
 274          * from accessing non-existent entries.
 275          */
 276         max_addr = (S390_lowcore.kernel_asce & _REGION_ENTRY_TYPE_MASK) >> 2;
 277         max_addr = 1UL << (max_addr * 11 + 31);
 278         address_markers[IDENTITY_AFTER_END_NR].start_address = ident_map_size;
 279         address_markers[MODULES_NR].start_address = MODULES_VADDR;
 280         address_markers[MODULES_END_NR].start_address = MODULES_END;
 281         address_markers[VMEMMAP_NR].start_address = (unsigned long) vmemmap;
 282         address_markers[VMEMMAP_END_NR].start_address = (unsigned long)vmemmap + vmemmap_size;
 283         address_markers[VMALLOC_NR].start_address = VMALLOC_START;
 284         address_markers[VMALLOC_END_NR].start_address = VMALLOC_END;
 285 #ifdef CONFIG_KFENCE
 286         address_markers[KFENCE_START_NR].start_address = kfence_start;
 287         address_markers[KFENCE_END_NR].start_address = kfence_start + KFENCE_POOL_SIZE;
 288 #endif
 289         sort_address_markers();
 290 #ifdef CONFIG_PTDUMP_DEBUGFS
 291         debugfs_create_file("kernel_page_tables", 0400, NULL, NULL, &ptdump_fops);
 292 #endif /* CONFIG_PTDUMP_DEBUGFS */
 293         return 0;
 294 }
 295 device_initcall(pt_dump_init);