1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2014-2017 Linaro Ltd. <ard.biesheuvel@linaro.org>
7 #include <linux/ftrace.h>
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/sort.h>
12 #include <asm/cache.h>
13 #include <asm/opcodes.h>
15 #ifdef CONFIG_THUMB2_KERNEL
16 #define PLT_ENT_LDR __opcode_to_mem_thumb32(0xf8dff000 | \
19 #define PLT_ENT_LDR __opcode_to_mem_arm(0xe59ff000 | \
23 static const u32 fixed_plts[] = {
24 #ifdef CONFIG_DYNAMIC_FTRACE
30 static bool in_init(const struct module *mod, unsigned long loc)
32 return loc - (u32)mod->init_layout.base < mod->init_layout.size;
35 static void prealloc_fixed(struct mod_plt_sec *pltsec, struct plt_entries *plt)
39 if (!ARRAY_SIZE(fixed_plts) || pltsec->plt_count)
41 pltsec->plt_count = ARRAY_SIZE(fixed_plts);
43 for (i = 0; i < ARRAY_SIZE(plt->ldr); ++i)
44 plt->ldr[i] = PLT_ENT_LDR;
46 BUILD_BUG_ON(sizeof(fixed_plts) > sizeof(plt->lit));
47 memcpy(plt->lit, fixed_plts, sizeof(fixed_plts));
50 u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val)
52 struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
54 struct plt_entries *plt;
57 /* cache the address, ELF header is available only during module load */
59 pltsec->plt_ent = (struct plt_entries *)pltsec->plt->sh_addr;
60 plt = pltsec->plt_ent;
62 prealloc_fixed(pltsec, plt);
64 for (idx = 0; idx < ARRAY_SIZE(fixed_plts); ++idx)
65 if (plt->lit[idx] == val)
66 return (u32)&plt->ldr[idx];
70 * Look for an existing entry pointing to 'val'. Given that the
71 * relocations are sorted, this will be the last entry we allocated.
74 if (pltsec->plt_count > 0) {
75 plt += (pltsec->plt_count - 1) / PLT_ENT_COUNT;
76 idx = (pltsec->plt_count - 1) % PLT_ENT_COUNT;
78 if (plt->lit[idx] == val)
79 return (u32)&plt->ldr[idx];
81 idx = (idx + 1) % PLT_ENT_COUNT;
87 BUG_ON(pltsec->plt_count * PLT_ENT_SIZE > pltsec->plt->sh_size);
90 /* Populate a new set of entries */
91 *plt = (struct plt_entries){
92 { [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, },
98 return (u32)&plt->ldr[idx];
101 #define cmp_3way(a,b) ((a) < (b) ? -1 : (a) > (b))
103 static int cmp_rel(const void *a, const void *b)
105 const Elf32_Rel *x = a, *y = b;
108 /* sort by type and symbol index */
109 i = cmp_3way(ELF32_R_TYPE(x->r_info), ELF32_R_TYPE(y->r_info));
111 i = cmp_3way(ELF32_R_SYM(x->r_info), ELF32_R_SYM(y->r_info));
115 static bool is_zero_addend_relocation(Elf32_Addr base, const Elf32_Rel *rel)
117 u32 *tval = (u32 *)(base + rel->r_offset);
120 * Do a bitwise compare on the raw addend rather than fully decoding
121 * the offset and doing an arithmetic comparison.
122 * Note that a zero-addend jump/call relocation is encoded taking the
123 * PC bias into account, i.e., -8 for ARM and -4 for Thumb2.
125 switch (ELF32_R_TYPE(rel->r_info)) {
129 case R_ARM_THM_JUMP24:
130 upper = __mem_to_opcode_thumb16(((u16 *)tval)[0]);
131 lower = __mem_to_opcode_thumb16(((u16 *)tval)[1]);
133 return (upper & 0x7ff) == 0x7ff && (lower & 0x2fff) == 0x2ffe;
138 return (__mem_to_opcode_arm(*tval) & 0xffffff) == 0xfffffe;
143 static bool duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num)
145 const Elf32_Rel *prev;
148 * Entries are sorted by type and symbol index. That means that,
149 * if a duplicate entry exists, it must be in the preceding
155 prev = rel + num - 1;
156 return cmp_rel(rel + num, prev) == 0 &&
157 is_zero_addend_relocation(base, prev);
160 /* Count how many PLT entries we may need */
161 static unsigned int count_plts(const Elf32_Sym *syms, Elf32_Addr base,
162 const Elf32_Rel *rel, int num, Elf32_Word dstidx)
164 unsigned int ret = 0;
168 for (i = 0; i < num; i++) {
169 switch (ELF32_R_TYPE(rel[i].r_info)) {
174 case R_ARM_THM_JUMP24:
176 * We only have to consider branch targets that resolve
177 * to symbols that are defined in a different section.
178 * This is not simply a heuristic, it is a fundamental
179 * limitation, since there is no guaranteed way to emit
180 * PLT entries sufficiently close to the branch if the
181 * section size exceeds the range of a branch
182 * instruction. So ignore relocations against defined
183 * symbols if they live in the same section as the
186 s = syms + ELF32_R_SYM(rel[i].r_info);
187 if (s->st_shndx == dstidx)
191 * Jump relocations with non-zero addends against
192 * undefined symbols are supported by the ELF spec, but
193 * do not occur in practice (e.g., 'jump n bytes past
194 * the entry point of undefined function symbol f').
195 * So we need to support them, but there is no need to
196 * take them into consideration when trying to optimize
197 * this code. So let's only check for duplicates when
198 * the addend is zero. (Note that calls into the core
199 * module via init PLT entries could involve section
200 * relative symbol references with non-zero addends, for
201 * which we may end up emitting duplicates, but the init
202 * PLT is released along with the rest of the .init
203 * region as soon as module loading completes.)
205 if (!is_zero_addend_relocation(base, rel + i) ||
206 !duplicate_rel(base, rel, i))
213 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
214 char *secstrings, struct module *mod)
216 unsigned long core_plts = ARRAY_SIZE(fixed_plts);
217 unsigned long init_plts = ARRAY_SIZE(fixed_plts);
218 Elf32_Shdr *s, *sechdrs_end = sechdrs + ehdr->e_shnum;
219 Elf32_Sym *syms = NULL;
222 * To store the PLTs, we expand the .text section for core module code
223 * and for initialization code.
225 for (s = sechdrs; s < sechdrs_end; ++s) {
226 if (strcmp(".plt", secstrings + s->sh_name) == 0)
227 mod->arch.core.plt = s;
228 else if (strcmp(".init.plt", secstrings + s->sh_name) == 0)
229 mod->arch.init.plt = s;
230 else if (s->sh_type == SHT_SYMTAB)
231 syms = (Elf32_Sym *)s->sh_addr;
234 if (!mod->arch.core.plt || !mod->arch.init.plt) {
235 pr_err("%s: module PLT section(s) missing\n", mod->name);
239 pr_err("%s: module symtab section missing\n", mod->name);
243 for (s = sechdrs + 1; s < sechdrs_end; ++s) {
244 Elf32_Rel *rels = (void *)ehdr + s->sh_offset;
245 int numrels = s->sh_size / sizeof(Elf32_Rel);
246 Elf32_Shdr *dstsec = sechdrs + s->sh_info;
248 if (s->sh_type != SHT_REL)
251 /* ignore relocations that operate on non-exec sections */
252 if (!(dstsec->sh_flags & SHF_EXECINSTR))
255 /* sort by type and symbol index */
256 sort(rels, numrels, sizeof(Elf32_Rel), cmp_rel, NULL);
258 if (strncmp(secstrings + dstsec->sh_name, ".init", 5) != 0)
259 core_plts += count_plts(syms, dstsec->sh_addr, rels,
260 numrels, s->sh_info);
262 init_plts += count_plts(syms, dstsec->sh_addr, rels,
263 numrels, s->sh_info);
266 mod->arch.core.plt->sh_type = SHT_NOBITS;
267 mod->arch.core.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
268 mod->arch.core.plt->sh_addralign = L1_CACHE_BYTES;
269 mod->arch.core.plt->sh_size = round_up(core_plts * PLT_ENT_SIZE,
270 sizeof(struct plt_entries));
271 mod->arch.core.plt_count = 0;
272 mod->arch.core.plt_ent = NULL;
274 mod->arch.init.plt->sh_type = SHT_NOBITS;
275 mod->arch.init.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
276 mod->arch.init.plt->sh_addralign = L1_CACHE_BYTES;
277 mod->arch.init.plt->sh_size = round_up(init_plts * PLT_ENT_SIZE,
278 sizeof(struct plt_entries));
279 mod->arch.init.plt_count = 0;
280 mod->arch.init.plt_ent = NULL;
282 pr_debug("%s: plt=%x, init.plt=%x\n", __func__,
283 mod->arch.core.plt->sh_size, mod->arch.init.plt->sh_size);