2 * Copyright (C) 2014-2017 Linaro Ltd. <ard.biesheuvel@linaro.org>
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.
10 #include <linux/ftrace.h>
11 #include <linux/kernel.h>
12 #include <linux/module.h>
13 #include <linux/sort.h>
15 #include <asm/cache.h>
16 #include <asm/opcodes.h>
18 #ifdef CONFIG_THUMB2_KERNEL
19 #define PLT_ENT_LDR __opcode_to_mem_thumb32(0xf8dff000 | \
22 #define PLT_ENT_LDR __opcode_to_mem_arm(0xe59ff000 | \
26 static const u32 fixed_plts[] = {
27 #ifdef CONFIG_DYNAMIC_FTRACE
33 static bool in_init(const struct module *mod, unsigned long loc)
35 return loc - (u32)mod->init_layout.base < mod->init_layout.size;
38 static void prealloc_fixed(struct mod_plt_sec *pltsec, struct plt_entries *plt)
42 if (!ARRAY_SIZE(fixed_plts) || pltsec->plt_count)
44 pltsec->plt_count = ARRAY_SIZE(fixed_plts);
46 for (i = 0; i < ARRAY_SIZE(plt->ldr); ++i)
47 plt->ldr[i] = PLT_ENT_LDR;
49 BUILD_BUG_ON(sizeof(fixed_plts) > sizeof(plt->lit));
50 memcpy(plt->lit, fixed_plts, sizeof(fixed_plts));
53 u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val)
55 struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
57 struct plt_entries *plt;
60 /* cache the address, ELF header is available only during module load */
62 pltsec->plt_ent = (struct plt_entries *)pltsec->plt->sh_addr;
63 plt = pltsec->plt_ent;
65 prealloc_fixed(pltsec, plt);
67 for (idx = 0; idx < ARRAY_SIZE(fixed_plts); ++idx)
68 if (plt->lit[idx] == val)
69 return (u32)&plt->ldr[idx];
73 * Look for an existing entry pointing to 'val'. Given that the
74 * relocations are sorted, this will be the last entry we allocated.
77 if (pltsec->plt_count > 0) {
78 plt += (pltsec->plt_count - 1) / PLT_ENT_COUNT;
79 idx = (pltsec->plt_count - 1) % PLT_ENT_COUNT;
81 if (plt->lit[idx] == val)
82 return (u32)&plt->ldr[idx];
84 idx = (idx + 1) % PLT_ENT_COUNT;
90 BUG_ON(pltsec->plt_count * PLT_ENT_SIZE > pltsec->plt->sh_size);
93 /* Populate a new set of entries */
94 *plt = (struct plt_entries){
95 { [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, },
101 return (u32)&plt->ldr[idx];
104 #define cmp_3way(a,b) ((a) < (b) ? -1 : (a) > (b))
106 static int cmp_rel(const void *a, const void *b)
108 const Elf32_Rel *x = a, *y = b;
111 /* sort by type and symbol index */
112 i = cmp_3way(ELF32_R_TYPE(x->r_info), ELF32_R_TYPE(y->r_info));
114 i = cmp_3way(ELF32_R_SYM(x->r_info), ELF32_R_SYM(y->r_info));
118 static bool is_zero_addend_relocation(Elf32_Addr base, const Elf32_Rel *rel)
120 u32 *tval = (u32 *)(base + rel->r_offset);
123 * Do a bitwise compare on the raw addend rather than fully decoding
124 * the offset and doing an arithmetic comparison.
125 * Note that a zero-addend jump/call relocation is encoded taking the
126 * PC bias into account, i.e., -8 for ARM and -4 for Thumb2.
128 switch (ELF32_R_TYPE(rel->r_info)) {
132 case R_ARM_THM_JUMP24:
133 upper = __mem_to_opcode_thumb16(((u16 *)tval)[0]);
134 lower = __mem_to_opcode_thumb16(((u16 *)tval)[1]);
136 return (upper & 0x7ff) == 0x7ff && (lower & 0x2fff) == 0x2ffe;
141 return (__mem_to_opcode_arm(*tval) & 0xffffff) == 0xfffffe;
146 static bool duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num)
148 const Elf32_Rel *prev;
151 * Entries are sorted by type and symbol index. That means that,
152 * if a duplicate entry exists, it must be in the preceding
158 prev = rel + num - 1;
159 return cmp_rel(rel + num, prev) == 0 &&
160 is_zero_addend_relocation(base, prev);
163 /* Count how many PLT entries we may need */
164 static unsigned int count_plts(const Elf32_Sym *syms, Elf32_Addr base,
165 const Elf32_Rel *rel, int num, Elf32_Word dstidx)
167 unsigned int ret = 0;
171 for (i = 0; i < num; i++) {
172 switch (ELF32_R_TYPE(rel[i].r_info)) {
177 case R_ARM_THM_JUMP24:
179 * We only have to consider branch targets that resolve
180 * to symbols that are defined in a different section.
181 * This is not simply a heuristic, it is a fundamental
182 * limitation, since there is no guaranteed way to emit
183 * PLT entries sufficiently close to the branch if the
184 * section size exceeds the range of a branch
185 * instruction. So ignore relocations against defined
186 * symbols if they live in the same section as the
189 s = syms + ELF32_R_SYM(rel[i].r_info);
190 if (s->st_shndx == dstidx)
194 * Jump relocations with non-zero addends against
195 * undefined symbols are supported by the ELF spec, but
196 * do not occur in practice (e.g., 'jump n bytes past
197 * the entry point of undefined function symbol f').
198 * So we need to support them, but there is no need to
199 * take them into consideration when trying to optimize
200 * this code. So let's only check for duplicates when
201 * the addend is zero. (Note that calls into the core
202 * module via init PLT entries could involve section
203 * relative symbol references with non-zero addends, for
204 * which we may end up emitting duplicates, but the init
205 * PLT is released along with the rest of the .init
206 * region as soon as module loading completes.)
208 if (!is_zero_addend_relocation(base, rel + i) ||
209 !duplicate_rel(base, rel, i))
216 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
217 char *secstrings, struct module *mod)
219 unsigned long core_plts = ARRAY_SIZE(fixed_plts);
220 unsigned long init_plts = ARRAY_SIZE(fixed_plts);
221 Elf32_Shdr *s, *sechdrs_end = sechdrs + ehdr->e_shnum;
222 Elf32_Sym *syms = NULL;
225 * To store the PLTs, we expand the .text section for core module code
226 * and for initialization code.
228 for (s = sechdrs; s < sechdrs_end; ++s) {
229 if (strcmp(".plt", secstrings + s->sh_name) == 0)
230 mod->arch.core.plt = s;
231 else if (strcmp(".init.plt", secstrings + s->sh_name) == 0)
232 mod->arch.init.plt = s;
233 else if (s->sh_type == SHT_SYMTAB)
234 syms = (Elf32_Sym *)s->sh_addr;
237 if (!mod->arch.core.plt || !mod->arch.init.plt) {
238 pr_err("%s: module PLT section(s) missing\n", mod->name);
242 pr_err("%s: module symtab section missing\n", mod->name);
246 for (s = sechdrs + 1; s < sechdrs_end; ++s) {
247 Elf32_Rel *rels = (void *)ehdr + s->sh_offset;
248 int numrels = s->sh_size / sizeof(Elf32_Rel);
249 Elf32_Shdr *dstsec = sechdrs + s->sh_info;
251 if (s->sh_type != SHT_REL)
254 /* ignore relocations that operate on non-exec sections */
255 if (!(dstsec->sh_flags & SHF_EXECINSTR))
258 /* sort by type and symbol index */
259 sort(rels, numrels, sizeof(Elf32_Rel), cmp_rel, NULL);
261 if (strncmp(secstrings + dstsec->sh_name, ".init", 5) != 0)
262 core_plts += count_plts(syms, dstsec->sh_addr, rels,
263 numrels, s->sh_info);
265 init_plts += count_plts(syms, dstsec->sh_addr, rels,
266 numrels, s->sh_info);
269 mod->arch.core.plt->sh_type = SHT_NOBITS;
270 mod->arch.core.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
271 mod->arch.core.plt->sh_addralign = L1_CACHE_BYTES;
272 mod->arch.core.plt->sh_size = round_up(core_plts * PLT_ENT_SIZE,
273 sizeof(struct plt_entries));
274 mod->arch.core.plt_count = 0;
275 mod->arch.core.plt_ent = NULL;
277 mod->arch.init.plt->sh_type = SHT_NOBITS;
278 mod->arch.init.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
279 mod->arch.init.plt->sh_addralign = L1_CACHE_BYTES;
280 mod->arch.init.plt->sh_size = round_up(init_plts * PLT_ENT_SIZE,
281 sizeof(struct plt_entries));
282 mod->arch.init.plt_count = 0;
283 mod->arch.init.plt_ent = NULL;
285 pr_debug("%s: plt=%x, init.plt=%x\n", __func__,
286 mod->arch.core.plt->sh_size, mod->arch.init.plt->sh_size);