arch/arm64/kernel/module-plts.c

   1 /*
   2  * Copyright (C) 2014-2017 Linaro Ltd. <ard.biesheuvel@linaro.org>
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 as
   6  * published by the Free Software Foundation.
   7  */
   8
   9 #include <linux/elf.h>
  10 #include <linux/kernel.h>
  11 #include <linux/module.h>
  12 #include <linux/sort.h>
  13
  14 static bool in_init(const struct module *mod, void *loc)
  15 {
  16         return (u64)loc - (u64)mod->init_layout.base < mod->init_layout.size;
  17 }
  18
  19 u64 module_emit_plt_entry(struct module *mod, void *loc, const Elf64_Rela *rela,
  20                           Elf64_Sym *sym)
  21 {
  22         struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
  23                                                           &mod->arch.init;
  24         struct plt_entry *plt = (struct plt_entry *)pltsec->plt->sh_addr;
  25         int i = pltsec->plt_num_entries;
  26         u64 val = sym->st_value + rela->r_addend;
  27
  28         plt[i] = get_plt_entry(val);
  29
  30         /*
  31          * Check if the entry we just created is a duplicate. Given that the
  32          * relocations are sorted, this will be the last entry we allocated.
  33          * (if one exists).
  34          */
  35         if (i > 0 && plt_entries_equal(plt + i, plt + i - 1))
  36                 return (u64)&plt[i - 1];
  37
  38         pltsec->plt_num_entries++;
  39         if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
  40                 return 0;
  41
  42         return (u64)&plt[i];
  43 }
  44
  45 #ifdef CONFIG_ARM64_ERRATUM_843419
  46 u64 module_emit_veneer_for_adrp(struct module *mod, void *loc, u64 val)
  47 {
  48         struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
  49                                                           &mod->arch.init;
  50         struct plt_entry *plt = (struct plt_entry *)pltsec->plt->sh_addr;
  51         int i = pltsec->plt_num_entries++;
  52         u32 mov0, mov1, mov2, br;
  53         int rd;
  54
  55         if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
  56                 return 0;
  57
  58         /* get the destination register of the ADRP instruction */
  59         rd = aarch64_insn_decode_register(AARCH64_INSN_REGTYPE_RD,
  60                                           le32_to_cpup((__le32 *)loc));
  61
  62         /* generate the veneer instructions */
  63         mov0 = aarch64_insn_gen_movewide(rd, (u16)~val, 0,
  64                                          AARCH64_INSN_VARIANT_64BIT,
  65                                          AARCH64_INSN_MOVEWIDE_INVERSE);
  66         mov1 = aarch64_insn_gen_movewide(rd, (u16)(val >> 16), 16,
  67                                          AARCH64_INSN_VARIANT_64BIT,
  68                                          AARCH64_INSN_MOVEWIDE_KEEP);
  69         mov2 = aarch64_insn_gen_movewide(rd, (u16)(val >> 32), 32,
  70                                          AARCH64_INSN_VARIANT_64BIT,
  71                                          AARCH64_INSN_MOVEWIDE_KEEP);
  72         br = aarch64_insn_gen_branch_imm((u64)&plt[i].br, (u64)loc + 4,
  73                                          AARCH64_INSN_BRANCH_NOLINK);
  74
  75         plt[i] = (struct plt_entry){
  76                         cpu_to_le32(mov0),
  77                         cpu_to_le32(mov1),
  78                         cpu_to_le32(mov2),
  79                         cpu_to_le32(br)
  80                 };
  81
  82         return (u64)&plt[i];
  83 }
  84 #endif
  85
  86 #define cmp_3way(a,b)   ((a) < (b) ? -1 : (a) > (b))
  87
  88 static int cmp_rela(const void *a, const void *b)
  89 {
  90         const Elf64_Rela *x = a, *y = b;
  91         int i;
  92
  93         /* sort by type, symbol index and addend */
  94         i = cmp_3way(ELF64_R_TYPE(x->r_info), ELF64_R_TYPE(y->r_info));
  95         if (i == 0)
  96                 i = cmp_3way(ELF64_R_SYM(x->r_info), ELF64_R_SYM(y->r_info));
  97         if (i == 0)
  98                 i = cmp_3way(x->r_addend, y->r_addend);
  99         return i;
 100 }
 101
 102 static bool duplicate_rel(const Elf64_Rela *rela, int num)
 103 {
 104         /*
 105          * Entries are sorted by type, symbol index and addend. That means
 106          * that, if a duplicate entry exists, it must be in the preceding
 107          * slot.
 108          */
 109         return num > 0 && cmp_rela(rela + num, rela + num - 1) == 0;
 110 }
 111
 112 static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
 113                                Elf64_Word dstidx, Elf_Shdr *dstsec)
 114 {
 115         unsigned int ret = 0;
 116         Elf64_Sym *s;
 117         int i;
 118
 119         for (i = 0; i < num; i++) {
 120                 u64 min_align;
 121
 122                 switch (ELF64_R_TYPE(rela[i].r_info)) {
 123                 case R_AARCH64_JUMP26:
 124                 case R_AARCH64_CALL26:
 125                         if (!IS_ENABLED(CONFIG_RANDOMIZE_BASE))
 126                                 break;
 127
 128                         /*
 129                          * We only have to consider branch targets that resolve
 130                          * to symbols that are defined in a different section.
 131                          * This is not simply a heuristic, it is a fundamental
 132                          * limitation, since there is no guaranteed way to emit
 133                          * PLT entries sufficiently close to the branch if the
 134                          * section size exceeds the range of a branch
 135                          * instruction. So ignore relocations against defined
 136                          * symbols if they live in the same section as the
 137                          * relocation target.
 138                          */
 139                         s = syms + ELF64_R_SYM(rela[i].r_info);
 140                         if (s->st_shndx == dstidx)
 141                                 break;
 142
 143                         /*
 144                          * Jump relocations with non-zero addends against
 145                          * undefined symbols are supported by the ELF spec, but
 146                          * do not occur in practice (e.g., 'jump n bytes past
 147                          * the entry point of undefined function symbol f').
 148                          * So we need to support them, but there is no need to
 149                          * take them into consideration when trying to optimize
 150                          * this code. So let's only check for duplicates when
 151                          * the addend is zero: this allows us to record the PLT
 152                          * entry address in the symbol table itself, rather than
 153                          * having to search the list for duplicates each time we
 154                          * emit one.
 155                          */
 156                         if (rela[i].r_addend != 0 || !duplicate_rel(rela, i))
 157                                 ret++;
 158                         break;
 159                 case R_AARCH64_ADR_PREL_PG_HI21_NC:
 160                 case R_AARCH64_ADR_PREL_PG_HI21:
 161                         if (!IS_ENABLED(CONFIG_ARM64_ERRATUM_843419) ||
 162                             !cpus_have_const_cap(ARM64_WORKAROUND_843419))
 163                                 break;
 164
 165                         /*
 166                          * Determine the minimal safe alignment for this ADRP
 167                          * instruction: the section alignment at which it is
 168                          * guaranteed not to appear at a vulnerable offset.
 169                          *
 170                          * This comes down to finding the least significant zero
 171                          * bit in bits [11:3] of the section offset, and
 172                          * increasing the section's alignment so that the
 173                          * resulting address of this instruction is guaranteed
 174                          * to equal the offset in that particular bit (as well
 175                          * as all less signficant bits). This ensures that the
 176                          * address modulo 4 KB != 0xfff8 or 0xfffc (which would
 177                          * have all ones in bits [11:3])
 178                          */
 179                         min_align = 2ULL << ffz(rela[i].r_offset | 0x7);
 180
 181                         /*
 182                          * Allocate veneer space for each ADRP that may appear
 183                          * at a vulnerable offset nonetheless. At relocation
 184                          * time, some of these will remain unused since some
 185                          * ADRP instructions can be patched to ADR instructions
 186                          * instead.
 187                          */
 188                         if (min_align > SZ_4K)
 189                                 ret++;
 190                         else
 191                                 dstsec->sh_addralign = max(dstsec->sh_addralign,
 192                                                            min_align);
 193                         break;
 194                 }
 195         }
 196         return ret;
 197 }
 198
 199 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
 200                               char *secstrings, struct module *mod)
 201 {
 202         unsigned long core_plts = 0;
 203         unsigned long init_plts = 0;
 204         Elf64_Sym *syms = NULL;
 205         Elf_Shdr *tramp = NULL;
 206         int i;
 207
 208         /*
 209          * Find the empty .plt section so we can expand it to store the PLT
 210          * entries. Record the symtab address as well.
 211          */
 212         for (i = 0; i < ehdr->e_shnum; i++) {
 213                 if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
 214                         mod->arch.core.plt = sechdrs + i;
 215                 else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt"))
 216                         mod->arch.init.plt = sechdrs + i;
 217                 else if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE) &&
 218                          !strcmp(secstrings + sechdrs[i].sh_name,
 219                                  ".text.ftrace_trampoline"))
 220                         tramp = sechdrs + i;
 221                 else if (sechdrs[i].sh_type == SHT_SYMTAB)
 222                         syms = (Elf64_Sym *)sechdrs[i].sh_addr;
 223         }
 224
 225         if (!mod->arch.core.plt || !mod->arch.init.plt) {
 226                 pr_err("%s: module PLT section(s) missing\n", mod->name);
 227                 return -ENOEXEC;
 228         }
 229         if (!syms) {
 230                 pr_err("%s: module symtab section missing\n", mod->name);
 231                 return -ENOEXEC;
 232         }
 233
 234         for (i = 0; i < ehdr->e_shnum; i++) {
 235                 Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset;
 236                 int numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
 237                 Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info;
 238
 239                 if (sechdrs[i].sh_type != SHT_RELA)
 240                         continue;
 241
 242                 /* ignore relocations that operate on non-exec sections */
 243                 if (!(dstsec->sh_flags & SHF_EXECINSTR))
 244                         continue;
 245
 246                 /* sort by type, symbol index and addend */
 247                 sort(rels, numrels, sizeof(Elf64_Rela), cmp_rela, NULL);
 248
 249                 if (strncmp(secstrings + dstsec->sh_name, ".init", 5) != 0)
 250                         core_plts += count_plts(syms, rels, numrels,
 251                                                 sechdrs[i].sh_info, dstsec);
 252                 else
 253                         init_plts += count_plts(syms, rels, numrels,
 254                                                 sechdrs[i].sh_info, dstsec);
 255         }
 256
 257         mod->arch.core.plt->sh_type = SHT_NOBITS;
 258         mod->arch.core.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
 259         mod->arch.core.plt->sh_addralign = L1_CACHE_BYTES;
 260         mod->arch.core.plt->sh_size = (core_plts  + 1) * sizeof(struct plt_entry);
 261         mod->arch.core.plt_num_entries = 0;
 262         mod->arch.core.plt_max_entries = core_plts;
 263
 264         mod->arch.init.plt->sh_type = SHT_NOBITS;
 265         mod->arch.init.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
 266         mod->arch.init.plt->sh_addralign = L1_CACHE_BYTES;
 267         mod->arch.init.plt->sh_size = (init_plts + 1) * sizeof(struct plt_entry);
 268         mod->arch.init.plt_num_entries = 0;
 269         mod->arch.init.plt_max_entries = init_plts;
 270
 271         if (tramp) {
 272                 tramp->sh_type = SHT_NOBITS;
 273                 tramp->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
 274                 tramp->sh_addralign = __alignof__(struct plt_entry);
 275                 tramp->sh_size = sizeof(struct plt_entry);
 276         }
 277
 278         return 0;
 279 }