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         AMODE31_START_NR,
  25         AMODE31_END_NR,
  26         KERNEL_START_NR,
  27         KERNEL_END_NR,
  28 #ifdef CONFIG_KFENCE
  29         KFENCE_START_NR,
  30         KFENCE_END_NR,
  31 #endif
  32         IDENTITY_AFTER_NR,
  33         IDENTITY_AFTER_END_NR,
  34 #ifdef CONFIG_KASAN
  35         KASAN_SHADOW_START_NR,
  36         KASAN_SHADOW_END_NR,
  37 #endif
  38         VMEMMAP_NR,
  39         VMEMMAP_END_NR,
  40         VMALLOC_NR,
  41         VMALLOC_END_NR,
  42         MODULES_NR,
  43         MODULES_END_NR,
  44 };
  45
  46 static struct addr_marker address_markers[] = {
  47         [IDENTITY_BEFORE_NR]    = {0, "Identity Mapping Start"},
  48         [IDENTITY_BEFORE_END_NR] = {(unsigned long)_stext, "Identity Mapping End"},
  49         [AMODE31_START_NR]      = {0, "Amode31 Area Start"},
  50         [AMODE31_END_NR]        = {0, "Amode31 Area End"},
  51         [KERNEL_START_NR]       = {(unsigned long)_stext, "Kernel Image Start"},
  52         [KERNEL_END_NR]         = {(unsigned long)_end, "Kernel Image End"},
  53 #ifdef CONFIG_KFENCE
  54         [KFENCE_START_NR]       = {0, "KFence Pool Start"},
  55         [KFENCE_END_NR]         = {0, "KFence Pool End"},
  56 #endif
  57         [IDENTITY_AFTER_NR]     = {(unsigned long)_end, "Identity Mapping Start"},
  58         [IDENTITY_AFTER_END_NR] = {0, "Identity Mapping End"},
  59 #ifdef CONFIG_KASAN
  60         [KASAN_SHADOW_START_NR] = {KASAN_SHADOW_START, "Kasan Shadow Start"},
  61         [KASAN_SHADOW_END_NR]   = {KASAN_SHADOW_END, "Kasan Shadow End"},
  62 #endif
  63         [VMEMMAP_NR]            = {0, "vmemmap Area Start"},
  64         [VMEMMAP_END_NR]        = {0, "vmemmap Area End"},
  65         [VMALLOC_NR]            = {0, "vmalloc Area Start"},
  66         [VMALLOC_END_NR]        = {0, "vmalloc Area End"},
  67         [MODULES_NR]            = {0, "Modules Area Start"},
  68         [MODULES_END_NR]        = {0, "Modules Area End"},
  69         { -1, NULL }
  70 };
  71
  72 struct pg_state {
  73         struct ptdump_state ptdump;
  74         struct seq_file *seq;
  75         int level;
  76         unsigned int current_prot;
  77         bool check_wx;
  78         unsigned long wx_pages;
  79         unsigned long start_address;
  80         const struct addr_marker *marker;
  81 };
  82
  83 #define pt_dump_seq_printf(m, fmt, args...)     \
  84 ({                                              \
  85         struct seq_file *__m = (m);             \
  86                                                 \
  87         if (__m)                                \
  88                 seq_printf(__m, fmt, ##args);   \
  89 })
  90
  91 #define pt_dump_seq_puts(m, fmt)                \
  92 ({                                              \
  93         struct seq_file *__m = (m);             \
  94                                                 \
  95         if (__m)                                \
  96                 seq_printf(__m, fmt);           \
  97 })
  98
  99 static void print_prot(struct seq_file *m, unsigned int pr, int level)
 100 {
 101         static const char * const level_name[] =
 102                 { "ASCE", "PGD", "PUD", "PMD", "PTE" };
 103
 104         pt_dump_seq_printf(m, "%s ", level_name[level]);
 105         if (pr & _PAGE_INVALID) {
 106                 pt_dump_seq_printf(m, "I\n");
 107                 return;
 108         }
 109         pt_dump_seq_puts(m, (pr & _PAGE_PROTECT) ? "RO " : "RW ");
 110         pt_dump_seq_puts(m, (pr & _PAGE_NOEXEC) ? "NX\n" : "X\n");
 111 }
 112
 113 static void note_prot_wx(struct pg_state *st, unsigned long addr)
 114 {
 115 #ifdef CONFIG_DEBUG_WX
 116         if (!st->check_wx)
 117                 return;
 118         if (st->current_prot & _PAGE_INVALID)
 119                 return;
 120         if (st->current_prot & _PAGE_PROTECT)
 121                 return;
 122         if (st->current_prot & _PAGE_NOEXEC)
 123                 return;
 124         /*
 125          * The first lowcore page is W+X if spectre mitigations are using
 126          * trampolines or the BEAR enhancements facility is not installed,
 127          * in which case we have two lpswe instructions in lowcore that need
 128          * to be executable.
 129          */
 130         if (addr == PAGE_SIZE && (nospec_uses_trampoline() || !static_key_enabled(&cpu_has_bear)))
 131                 return;
 132         WARN_ONCE(1, "s390/mm: Found insecure W+X mapping at address %pS\n",
 133                   (void *)st->start_address);
 134         st->wx_pages += (addr - st->start_address) / PAGE_SIZE;
 135 #endif /* CONFIG_DEBUG_WX */
 136 }
 137
 138 static void note_page(struct ptdump_state *pt_st, unsigned long addr, int level, u64 val)
 139 {
 140         int width = sizeof(unsigned long) * 2;
 141         static const char units[] = "KMGTPE";
 142         const char *unit = units;
 143         unsigned long delta;
 144         struct pg_state *st;
 145         struct seq_file *m;
 146         unsigned int prot;
 147
 148         st = container_of(pt_st, struct pg_state, ptdump);
 149         m = st->seq;
 150         prot = val & (_PAGE_PROTECT | _PAGE_NOEXEC);
 151         if (level == 4 && (val & _PAGE_INVALID))
 152                 prot = _PAGE_INVALID;
 153         /* For pmd_none() & friends val gets passed as zero. */
 154         if (level != 4 && !val)
 155                 prot = _PAGE_INVALID;
 156         /* Final flush from generic code. */
 157         if (level == -1)
 158                 addr = max_addr;
 159         if (st->level == -1) {
 160                 pt_dump_seq_printf(m, "---[ %s ]---\n", st->marker->name);
 161                 st->start_address = addr;
 162                 st->current_prot = prot;
 163                 st->level = level;
 164         } else if (prot != st->current_prot || level != st->level ||
 165                    addr >= st->marker[1].start_address) {
 166                 note_prot_wx(st, addr);
 167                 pt_dump_seq_printf(m, "0x%0*lx-0x%0*lx ",
 168                                    width, st->start_address,
 169                                    width, addr);
 170                 delta = (addr - st->start_address) >> 10;
 171                 while (!(delta & 0x3ff) && unit[1]) {
 172                         delta >>= 10;
 173                         unit++;
 174                 }
 175                 pt_dump_seq_printf(m, "%9lu%c ", delta, *unit);
 176                 print_prot(m, st->current_prot, st->level);
 177                 while (addr >= st->marker[1].start_address) {
 178                         st->marker++;
 179                         pt_dump_seq_printf(m, "---[ %s ]---\n", st->marker->name);
 180                 }
 181                 st->start_address = addr;
 182                 st->current_prot = prot;
 183                 st->level = level;
 184         }
 185 }
 186
 187 #ifdef CONFIG_DEBUG_WX
 188 void ptdump_check_wx(void)
 189 {
 190         struct pg_state st = {
 191                 .ptdump = {
 192                         .note_page = note_page,
 193                         .range = (struct ptdump_range[]) {
 194                                 {.start = 0, .end = max_addr},
 195                                 {.start = 0, .end = 0},
 196                         }
 197                 },
 198                 .seq = NULL,
 199                 .level = -1,
 200                 .current_prot = 0,
 201                 .check_wx = true,
 202                 .wx_pages = 0,
 203                 .start_address = 0,
 204                 .marker = (struct addr_marker[]) {
 205                         { .start_address =  0, .name = NULL},
 206                         { .start_address = -1, .name = NULL},
 207                 },
 208         };
 209
 210         if (!MACHINE_HAS_NX)
 211                 return;
 212         ptdump_walk_pgd(&st.ptdump, &init_mm, NULL);
 213         if (st.wx_pages)
 214                 pr_warn("Checked W+X mappings: FAILED, %lu W+X pages found\n", st.wx_pages);
 215         else
 216                 pr_info("Checked W+X mappings: passed, no %sW+X pages found\n",
 217                         (nospec_uses_trampoline() || !static_key_enabled(&cpu_has_bear)) ?
 218                         "unexpected " : "");
 219 }
 220 #endif /* CONFIG_DEBUG_WX */
 221
 222 #ifdef CONFIG_PTDUMP_DEBUGFS
 223 static int ptdump_show(struct seq_file *m, void *v)
 224 {
 225         struct pg_state st = {
 226                 .ptdump = {
 227                         .note_page = note_page,
 228                         .range = (struct ptdump_range[]) {
 229                                 {.start = 0, .end = max_addr},
 230                                 {.start = 0, .end = 0},
 231                         }
 232                 },
 233                 .seq = m,
 234                 .level = -1,
 235                 .current_prot = 0,
 236                 .check_wx = false,
 237                 .wx_pages = 0,
 238                 .start_address = 0,
 239                 .marker = address_markers,
 240         };
 241
 242         get_online_mems();
 243         mutex_lock(&cpa_mutex);
 244         ptdump_walk_pgd(&st.ptdump, &init_mm, NULL);
 245         mutex_unlock(&cpa_mutex);
 246         put_online_mems();
 247         return 0;
 248 }
 249 DEFINE_SHOW_ATTRIBUTE(ptdump);
 250 #endif /* CONFIG_PTDUMP_DEBUGFS */
 251
 252 /*
 253  * Heapsort from lib/sort.c is not a stable sorting algorithm, do a simple
 254  * insertion sort to preserve the original order of markers with the same
 255  * start address.
 256  */
 257 static void sort_address_markers(void)
 258 {
 259         struct addr_marker tmp;
 260         int i, j;
 261
 262         for (i = 1; i < ARRAY_SIZE(address_markers) - 1; i++) {
 263                 tmp = address_markers[i];
 264                 for (j = i - 1; j >= 0 && address_markers[j].start_address > tmp.start_address; j--)
 265                         address_markers[j + 1] = address_markers[j];
 266                 address_markers[j + 1] = tmp;
 267         }
 268 }
 269
 270 static int pt_dump_init(void)
 271 {
 272 #ifdef CONFIG_KFENCE
 273         unsigned long kfence_start = (unsigned long)__kfence_pool;
 274 #endif
 275         /*
 276          * Figure out the maximum virtual address being accessible with the
 277          * kernel ASCE. We need this to keep the page table walker functions
 278          * from accessing non-existent entries.
 279          */
 280         max_addr = (S390_lowcore.kernel_asce & _REGION_ENTRY_TYPE_MASK) >> 2;
 281         max_addr = 1UL << (max_addr * 11 + 31);
 282         address_markers[IDENTITY_AFTER_END_NR].start_address = ident_map_size;
 283         address_markers[AMODE31_START_NR].start_address = __samode31;
 284         address_markers[AMODE31_END_NR].start_address = __eamode31;
 285         address_markers[MODULES_NR].start_address = MODULES_VADDR;
 286         address_markers[MODULES_END_NR].start_address = MODULES_END;
 287         address_markers[VMEMMAP_NR].start_address = (unsigned long) vmemmap;
 288         address_markers[VMEMMAP_END_NR].start_address = (unsigned long)vmemmap + vmemmap_size;
 289         address_markers[VMALLOC_NR].start_address = VMALLOC_START;
 290         address_markers[VMALLOC_END_NR].start_address = VMALLOC_END;
 291 #ifdef CONFIG_KFENCE
 292         address_markers[KFENCE_START_NR].start_address = kfence_start;
 293         address_markers[KFENCE_END_NR].start_address = kfence_start + KFENCE_POOL_SIZE;
 294 #endif
 295         sort_address_markers();
 296 #ifdef CONFIG_PTDUMP_DEBUGFS
 297         debugfs_create_file("kernel_page_tables", 0400, NULL, NULL, &ptdump_fops);
 298 #endif /* CONFIG_PTDUMP_DEBUGFS */
 299         return 0;
 300 }
 301 device_initcall(pt_dump_init);