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 struct plt_entry {
  15         /*
  16          * A program that conforms to the AArch64 Procedure Call Standard
  17          * (AAPCS64) must assume that a veneer that alters IP0 (x16) and/or
  18          * IP1 (x17) may be inserted at any branch instruction that is
  19          * exposed to a relocation that supports long branches. Since that
  20          * is exactly what we are dealing with here, we are free to use x16
  21          * as a scratch register in the PLT veneers.
  22          */
  23         __le32  mov0;   /* movn x16, #0x....                    */
  24         __le32  mov1;   /* movk x16, #0x...., lsl #16           */
  25         __le32  mov2;   /* movk x16, #0x...., lsl #32           */
  26         __le32  br;     /* br   x16                             */
  27 };
  28
  29 static bool in_init(const struct module *mod, void *loc)
  30 {
  31         return (u64)loc - (u64)mod->init_layout.base < mod->init_layout.size;
  32 }
  33
  34 u64 module_emit_plt_entry(struct module *mod, void *loc, const Elf64_Rela *rela,
  35                           Elf64_Sym *sym)
  36 {
  37         struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
  38                                                           &mod->arch.init;
  39         struct plt_entry *plt = (struct plt_entry *)pltsec->plt->sh_addr;
  40         int i = pltsec->plt_num_entries;
  41         u64 val = sym->st_value + rela->r_addend;
  42
  43         /*
  44          * MOVK/MOVN/MOVZ opcode:
  45          * +--------+------------+--------+-----------+-------------+---------+
  46          * | sf[31] | opc[30:29] | 100101 | hw[22:21] | imm16[20:5] | Rd[4:0] |
  47          * +--------+------------+--------+-----------+-------------+---------+
  48          *
  49          * Rd     := 0x10 (x16)
  50          * hw     := 0b00 (no shift), 0b01 (lsl #16), 0b10 (lsl #32)
  51          * opc    := 0b11 (MOVK), 0b00 (MOVN), 0b10 (MOVZ)
  52          * sf     := 1 (64-bit variant)
  53          */
  54         plt[i] = (struct plt_entry){
  55                 cpu_to_le32(0x92800010 | (((~val      ) & 0xffff)) << 5),
  56                 cpu_to_le32(0xf2a00010 | ((( val >> 16) & 0xffff)) << 5),
  57                 cpu_to_le32(0xf2c00010 | ((( val >> 32) & 0xffff)) << 5),
  58                 cpu_to_le32(0xd61f0200)
  59         };
  60
  61         /*
  62          * Check if the entry we just created is a duplicate. Given that the
  63          * relocations are sorted, this will be the last entry we allocated.
  64          * (if one exists).
  65          */
  66         if (i > 0 &&
  67             plt[i].mov0 == plt[i - 1].mov0 &&
  68             plt[i].mov1 == plt[i - 1].mov1 &&
  69             plt[i].mov2 == plt[i - 1].mov2)
  70                 return (u64)&plt[i - 1];
  71
  72         pltsec->plt_num_entries++;
  73         BUG_ON(pltsec->plt_num_entries > pltsec->plt_max_entries);
  74
  75         return (u64)&plt[i];
  76 }
  77
  78 #define cmp_3way(a,b)   ((a) < (b) ? -1 : (a) > (b))
  79
  80 static int cmp_rela(const void *a, const void *b)
  81 {
  82         const Elf64_Rela *x = a, *y = b;
  83         int i;
  84
  85         /* sort by type, symbol index and addend */
  86         i = cmp_3way(ELF64_R_TYPE(x->r_info), ELF64_R_TYPE(y->r_info));
  87         if (i == 0)
  88                 i = cmp_3way(ELF64_R_SYM(x->r_info), ELF64_R_SYM(y->r_info));
  89         if (i == 0)
  90                 i = cmp_3way(x->r_addend, y->r_addend);
  91         return i;
  92 }
  93
  94 static bool duplicate_rel(const Elf64_Rela *rela, int num)
  95 {
  96         /*
  97          * Entries are sorted by type, symbol index and addend. That means
  98          * that, if a duplicate entry exists, it must be in the preceding
  99          * slot.
 100          */
 101         return num > 0 && cmp_rela(rela + num, rela + num - 1) == 0;
 102 }
 103
 104 static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
 105                                Elf64_Word dstidx)
 106 {
 107         unsigned int ret = 0;
 108         Elf64_Sym *s;
 109         int i;
 110
 111         for (i = 0; i < num; i++) {
 112                 switch (ELF64_R_TYPE(rela[i].r_info)) {
 113                 case R_AARCH64_JUMP26:
 114                 case R_AARCH64_CALL26:
 115                         /*
 116                          * We only have to consider branch targets that resolve
 117                          * to symbols that are defined in a different section.
 118                          * This is not simply a heuristic, it is a fundamental
 119                          * limitation, since there is no guaranteed way to emit
 120                          * PLT entries sufficiently close to the branch if the
 121                          * section size exceeds the range of a branch
 122                          * instruction. So ignore relocations against defined
 123                          * symbols if they live in the same section as the
 124                          * relocation target.
 125                          */
 126                         s = syms + ELF64_R_SYM(rela[i].r_info);
 127                         if (s->st_shndx == dstidx)
 128                                 break;
 129
 130                         /*
 131                          * Jump relocations with non-zero addends against
 132                          * undefined symbols are supported by the ELF spec, but
 133                          * do not occur in practice (e.g., 'jump n bytes past
 134                          * the entry point of undefined function symbol f').
 135                          * So we need to support them, but there is no need to
 136                          * take them into consideration when trying to optimize
 137                          * this code. So let's only check for duplicates when
 138                          * the addend is zero: this allows us to record the PLT
 139                          * entry address in the symbol table itself, rather than
 140                          * having to search the list for duplicates each time we
 141                          * emit one.
 142                          */
 143                         if (rela[i].r_addend != 0 || !duplicate_rel(rela, i))
 144                                 ret++;
 145                         break;
 146                 }
 147         }
 148         return ret;
 149 }
 150
 151 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
 152                               char *secstrings, struct module *mod)
 153 {
 154         unsigned long core_plts = 0;
 155         unsigned long init_plts = 0;
 156         Elf64_Sym *syms = NULL;
 157         int i;
 158
 159         /*
 160          * Find the empty .plt section so we can expand it to store the PLT
 161          * entries. Record the symtab address as well.
 162          */
 163         for (i = 0; i < ehdr->e_shnum; i++) {
 164                 if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
 165                         mod->arch.core.plt = sechdrs + i;
 166                 else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt"))
 167                         mod->arch.init.plt = sechdrs + i;
 168                 else if (sechdrs[i].sh_type == SHT_SYMTAB)
 169                         syms = (Elf64_Sym *)sechdrs[i].sh_addr;
 170         }
 171
 172         if (!mod->arch.core.plt || !mod->arch.init.plt) {
 173                 pr_err("%s: module PLT section(s) missing\n", mod->name);
 174                 return -ENOEXEC;
 175         }
 176         if (!syms) {
 177                 pr_err("%s: module symtab section missing\n", mod->name);
 178                 return -ENOEXEC;
 179         }
 180
 181         for (i = 0; i < ehdr->e_shnum; i++) {
 182                 Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset;
 183                 int numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
 184                 Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info;
 185
 186                 if (sechdrs[i].sh_type != SHT_RELA)
 187                         continue;
 188
 189                 /* ignore relocations that operate on non-exec sections */
 190                 if (!(dstsec->sh_flags & SHF_EXECINSTR))
 191                         continue;
 192
 193                 /* sort by type, symbol index and addend */
 194                 sort(rels, numrels, sizeof(Elf64_Rela), cmp_rela, NULL);
 195
 196                 if (strncmp(secstrings + dstsec->sh_name, ".init", 5) != 0)
 197                         core_plts += count_plts(syms, rels, numrels,
 198                                                 sechdrs[i].sh_info);
 199                 else
 200                         init_plts += count_plts(syms, rels, numrels,
 201                                                 sechdrs[i].sh_info);
 202         }
 203
 204         mod->arch.core.plt->sh_type = SHT_NOBITS;
 205         mod->arch.core.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
 206         mod->arch.core.plt->sh_addralign = L1_CACHE_BYTES;
 207         mod->arch.core.plt->sh_size = (core_plts  + 1) * sizeof(struct plt_entry);
 208         mod->arch.core.plt_num_entries = 0;
 209         mod->arch.core.plt_max_entries = core_plts;
 210
 211         mod->arch.init.plt->sh_type = SHT_NOBITS;
 212         mod->arch.init.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
 213         mod->arch.init.plt->sh_addralign = L1_CACHE_BYTES;
 214         mod->arch.init.plt->sh_size = (init_plts + 1) * sizeof(struct plt_entry);
 215         mod->arch.init.plt_num_entries = 0;
 216         mod->arch.init.plt_max_entries = init_plts;
 217
 218         return 0;
 219 }