Commit | Line | Data |
---|---|---|
fd045f6c AB |
1 | /* |
2 | * Copyright (C) 2014-2016 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 | u64 module_emit_plt_entry(struct module *mod, const Elf64_Rela *rela, | |
30 | Elf64_Sym *sym) | |
31 | { | |
32 | struct plt_entry *plt = (struct plt_entry *)mod->arch.plt->sh_addr; | |
33 | int i = mod->arch.plt_num_entries; | |
34 | u64 val = sym->st_value + rela->r_addend; | |
35 | ||
36 | /* | |
37 | * We only emit PLT entries against undefined (SHN_UNDEF) symbols, | |
38 | * which are listed in the ELF symtab section, but without a type | |
39 | * or a size. | |
40 | * So, similar to how the module loader uses the Elf64_Sym::st_value | |
41 | * field to store the resolved addresses of undefined symbols, let's | |
42 | * borrow the Elf64_Sym::st_size field (whose value is never used by | |
43 | * the module loader, even for symbols that are defined) to record | |
44 | * the address of a symbol's associated PLT entry as we emit it for a | |
45 | * zero addend relocation (which is the only kind we have to deal with | |
46 | * in practice). This allows us to find duplicates without having to | |
47 | * go through the table every time. | |
48 | */ | |
49 | if (rela->r_addend == 0 && sym->st_size != 0) { | |
50 | BUG_ON(sym->st_size < (u64)plt || sym->st_size >= (u64)&plt[i]); | |
51 | return sym->st_size; | |
52 | } | |
53 | ||
54 | mod->arch.plt_num_entries++; | |
55 | BUG_ON(mod->arch.plt_num_entries > mod->arch.plt_max_entries); | |
56 | ||
57 | /* | |
58 | * MOVK/MOVN/MOVZ opcode: | |
59 | * +--------+------------+--------+-----------+-------------+---------+ | |
60 | * | sf[31] | opc[30:29] | 100101 | hw[22:21] | imm16[20:5] | Rd[4:0] | | |
61 | * +--------+------------+--------+-----------+-------------+---------+ | |
62 | * | |
63 | * Rd := 0x10 (x16) | |
64 | * hw := 0b00 (no shift), 0b01 (lsl #16), 0b10 (lsl #32) | |
65 | * opc := 0b11 (MOVK), 0b00 (MOVN), 0b10 (MOVZ) | |
66 | * sf := 1 (64-bit variant) | |
67 | */ | |
68 | plt[i] = (struct plt_entry){ | |
69 | cpu_to_le32(0x92800010 | (((~val ) & 0xffff)) << 5), | |
70 | cpu_to_le32(0xf2a00010 | ((( val >> 16) & 0xffff)) << 5), | |
71 | cpu_to_le32(0xf2c00010 | ((( val >> 32) & 0xffff)) << 5), | |
72 | cpu_to_le32(0xd61f0200) | |
73 | }; | |
74 | ||
75 | if (rela->r_addend == 0) | |
76 | sym->st_size = (u64)&plt[i]; | |
77 | ||
78 | return (u64)&plt[i]; | |
79 | } | |
80 | ||
81 | #define cmp_3way(a,b) ((a) < (b) ? -1 : (a) > (b)) | |
82 | ||
83 | static int cmp_rela(const void *a, const void *b) | |
84 | { | |
85 | const Elf64_Rela *x = a, *y = b; | |
86 | int i; | |
87 | ||
88 | /* sort by type, symbol index and addend */ | |
89 | i = cmp_3way(ELF64_R_TYPE(x->r_info), ELF64_R_TYPE(y->r_info)); | |
90 | if (i == 0) | |
91 | i = cmp_3way(ELF64_R_SYM(x->r_info), ELF64_R_SYM(y->r_info)); | |
92 | if (i == 0) | |
93 | i = cmp_3way(x->r_addend, y->r_addend); | |
94 | return i; | |
95 | } | |
96 | ||
97 | static bool duplicate_rel(const Elf64_Rela *rela, int num) | |
98 | { | |
99 | /* | |
100 | * Entries are sorted by type, symbol index and addend. That means | |
101 | * that, if a duplicate entry exists, it must be in the preceding | |
102 | * slot. | |
103 | */ | |
104 | return num > 0 && cmp_rela(rela + num, rela + num - 1) == 0; | |
105 | } | |
106 | ||
107 | static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num) | |
108 | { | |
109 | unsigned int ret = 0; | |
110 | Elf64_Sym *s; | |
111 | int i; | |
112 | ||
113 | for (i = 0; i < num; i++) { | |
114 | switch (ELF64_R_TYPE(rela[i].r_info)) { | |
115 | case R_AARCH64_JUMP26: | |
116 | case R_AARCH64_CALL26: | |
117 | /* | |
118 | * We only have to consider branch targets that resolve | |
119 | * to undefined symbols. This is not simply a heuristic, | |
120 | * it is a fundamental limitation, since the PLT itself | |
121 | * is part of the module, and needs to be within 128 MB | |
122 | * as well, so modules can never grow beyond that limit. | |
123 | */ | |
124 | s = syms + ELF64_R_SYM(rela[i].r_info); | |
125 | if (s->st_shndx != SHN_UNDEF) | |
126 | break; | |
127 | ||
128 | /* | |
129 | * Jump relocations with non-zero addends against | |
130 | * undefined symbols are supported by the ELF spec, but | |
131 | * do not occur in practice (e.g., 'jump n bytes past | |
132 | * the entry point of undefined function symbol f'). | |
133 | * So we need to support them, but there is no need to | |
134 | * take them into consideration when trying to optimize | |
135 | * this code. So let's only check for duplicates when | |
136 | * the addend is zero: this allows us to record the PLT | |
137 | * entry address in the symbol table itself, rather than | |
138 | * having to search the list for duplicates each time we | |
139 | * emit one. | |
140 | */ | |
141 | if (rela[i].r_addend != 0 || !duplicate_rel(rela, i)) | |
142 | ret++; | |
143 | break; | |
144 | } | |
145 | } | |
146 | return ret; | |
147 | } | |
148 | ||
149 | int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs, | |
150 | char *secstrings, struct module *mod) | |
151 | { | |
152 | unsigned long plt_max_entries = 0; | |
153 | Elf64_Sym *syms = NULL; | |
154 | int i; | |
155 | ||
156 | /* | |
157 | * Find the empty .plt section so we can expand it to store the PLT | |
158 | * entries. Record the symtab address as well. | |
159 | */ | |
160 | for (i = 0; i < ehdr->e_shnum; i++) { | |
161 | if (strcmp(".plt", secstrings + sechdrs[i].sh_name) == 0) | |
162 | mod->arch.plt = sechdrs + i; | |
163 | else if (sechdrs[i].sh_type == SHT_SYMTAB) | |
164 | syms = (Elf64_Sym *)sechdrs[i].sh_addr; | |
165 | } | |
166 | ||
167 | if (!mod->arch.plt) { | |
168 | pr_err("%s: module PLT section missing\n", mod->name); | |
169 | return -ENOEXEC; | |
170 | } | |
171 | if (!syms) { | |
172 | pr_err("%s: module symtab section missing\n", mod->name); | |
173 | return -ENOEXEC; | |
174 | } | |
175 | ||
176 | for (i = 0; i < ehdr->e_shnum; i++) { | |
177 | Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset; | |
178 | int numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela); | |
179 | Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info; | |
180 | ||
181 | if (sechdrs[i].sh_type != SHT_RELA) | |
182 | continue; | |
183 | ||
184 | /* ignore relocations that operate on non-exec sections */ | |
185 | if (!(dstsec->sh_flags & SHF_EXECINSTR)) | |
186 | continue; | |
187 | ||
188 | /* sort by type, symbol index and addend */ | |
189 | sort(rels, numrels, sizeof(Elf64_Rela), cmp_rela, NULL); | |
190 | ||
191 | plt_max_entries += count_plts(syms, rels, numrels); | |
192 | } | |
193 | ||
194 | mod->arch.plt->sh_type = SHT_NOBITS; | |
195 | mod->arch.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC; | |
196 | mod->arch.plt->sh_addralign = L1_CACHE_BYTES; | |
197 | mod->arch.plt->sh_size = plt_max_entries * sizeof(struct plt_entry); | |
198 | mod->arch.plt_num_entries = 0; | |
199 | mod->arch.plt_max_entries = plt_max_entries; | |
200 | return 0; | |
201 | } |