mm: fix exec activate_mm vs TLB shootdown and lazy tlb switching race
[linux/fpc-iii.git] / arch / arm / kernel / module-plts.c
blob3d0c2e4dda1d23ae92ca40d4b40441d45bd02616
1 /*
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.
7 */
9 #include <linux/elf.h>
10 #include <linux/kernel.h>
11 #include <linux/module.h>
12 #include <linux/sort.h>
14 #include <asm/cache.h>
15 #include <asm/opcodes.h>
17 #define PLT_ENT_STRIDE L1_CACHE_BYTES
18 #define PLT_ENT_COUNT (PLT_ENT_STRIDE / sizeof(u32))
19 #define PLT_ENT_SIZE (sizeof(struct plt_entries) / PLT_ENT_COUNT)
21 #ifdef CONFIG_THUMB2_KERNEL
22 #define PLT_ENT_LDR __opcode_to_mem_thumb32(0xf8dff000 | \
23 (PLT_ENT_STRIDE - 4))
24 #else
25 #define PLT_ENT_LDR __opcode_to_mem_arm(0xe59ff000 | \
26 (PLT_ENT_STRIDE - 8))
27 #endif
29 struct plt_entries {
30 u32 ldr[PLT_ENT_COUNT];
31 u32 lit[PLT_ENT_COUNT];
34 static bool in_init(const struct module *mod, unsigned long loc)
36 return loc - (u32)mod->init_layout.base < mod->init_layout.size;
39 u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val)
41 struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
42 &mod->arch.init;
44 struct plt_entries *plt = (struct plt_entries *)pltsec->plt->sh_addr;
45 int idx = 0;
48 * Look for an existing entry pointing to 'val'. Given that the
49 * relocations are sorted, this will be the last entry we allocated.
50 * (if one exists).
52 if (pltsec->plt_count > 0) {
53 plt += (pltsec->plt_count - 1) / PLT_ENT_COUNT;
54 idx = (pltsec->plt_count - 1) % PLT_ENT_COUNT;
56 if (plt->lit[idx] == val)
57 return (u32)&plt->ldr[idx];
59 idx = (idx + 1) % PLT_ENT_COUNT;
60 if (!idx)
61 plt++;
64 pltsec->plt_count++;
65 BUG_ON(pltsec->plt_count * PLT_ENT_SIZE > pltsec->plt->sh_size);
67 if (!idx)
68 /* Populate a new set of entries */
69 *plt = (struct plt_entries){
70 { [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, },
71 { val, }
73 else
74 plt->lit[idx] = val;
76 return (u32)&plt->ldr[idx];
79 #define cmp_3way(a,b) ((a) < (b) ? -1 : (a) > (b))
81 static int cmp_rel(const void *a, const void *b)
83 const Elf32_Rel *x = a, *y = b;
84 int i;
86 /* sort by type and symbol index */
87 i = cmp_3way(ELF32_R_TYPE(x->r_info), ELF32_R_TYPE(y->r_info));
88 if (i == 0)
89 i = cmp_3way(ELF32_R_SYM(x->r_info), ELF32_R_SYM(y->r_info));
90 return i;
93 static bool is_zero_addend_relocation(Elf32_Addr base, const Elf32_Rel *rel)
95 u32 *tval = (u32 *)(base + rel->r_offset);
98 * Do a bitwise compare on the raw addend rather than fully decoding
99 * the offset and doing an arithmetic comparison.
100 * Note that a zero-addend jump/call relocation is encoded taking the
101 * PC bias into account, i.e., -8 for ARM and -4 for Thumb2.
103 switch (ELF32_R_TYPE(rel->r_info)) {
104 u16 upper, lower;
106 case R_ARM_THM_CALL:
107 case R_ARM_THM_JUMP24:
108 upper = __mem_to_opcode_thumb16(((u16 *)tval)[0]);
109 lower = __mem_to_opcode_thumb16(((u16 *)tval)[1]);
111 return (upper & 0x7ff) == 0x7ff && (lower & 0x2fff) == 0x2ffe;
113 case R_ARM_CALL:
114 case R_ARM_PC24:
115 case R_ARM_JUMP24:
116 return (__mem_to_opcode_arm(*tval) & 0xffffff) == 0xfffffe;
118 BUG();
121 static bool duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num)
123 const Elf32_Rel *prev;
126 * Entries are sorted by type and symbol index. That means that,
127 * if a duplicate entry exists, it must be in the preceding
128 * slot.
130 if (!num)
131 return false;
133 prev = rel + num - 1;
134 return cmp_rel(rel + num, prev) == 0 &&
135 is_zero_addend_relocation(base, prev);
138 /* Count how many PLT entries we may need */
139 static unsigned int count_plts(const Elf32_Sym *syms, Elf32_Addr base,
140 const Elf32_Rel *rel, int num, Elf32_Word dstidx)
142 unsigned int ret = 0;
143 const Elf32_Sym *s;
144 int i;
146 for (i = 0; i < num; i++) {
147 switch (ELF32_R_TYPE(rel[i].r_info)) {
148 case R_ARM_CALL:
149 case R_ARM_PC24:
150 case R_ARM_JUMP24:
151 case R_ARM_THM_CALL:
152 case R_ARM_THM_JUMP24:
154 * We only have to consider branch targets that resolve
155 * to symbols that are defined in a different section.
156 * This is not simply a heuristic, it is a fundamental
157 * limitation, since there is no guaranteed way to emit
158 * PLT entries sufficiently close to the branch if the
159 * section size exceeds the range of a branch
160 * instruction. So ignore relocations against defined
161 * symbols if they live in the same section as the
162 * relocation target.
164 s = syms + ELF32_R_SYM(rel[i].r_info);
165 if (s->st_shndx == dstidx)
166 break;
169 * Jump relocations with non-zero addends against
170 * undefined symbols are supported by the ELF spec, but
171 * do not occur in practice (e.g., 'jump n bytes past
172 * the entry point of undefined function symbol f').
173 * So we need to support them, but there is no need to
174 * take them into consideration when trying to optimize
175 * this code. So let's only check for duplicates when
176 * the addend is zero. (Note that calls into the core
177 * module via init PLT entries could involve section
178 * relative symbol references with non-zero addends, for
179 * which we may end up emitting duplicates, but the init
180 * PLT is released along with the rest of the .init
181 * region as soon as module loading completes.)
183 if (!is_zero_addend_relocation(base, rel + i) ||
184 !duplicate_rel(base, rel, i))
185 ret++;
188 return ret;
191 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
192 char *secstrings, struct module *mod)
194 unsigned long core_plts = 0;
195 unsigned long init_plts = 0;
196 Elf32_Shdr *s, *sechdrs_end = sechdrs + ehdr->e_shnum;
197 Elf32_Sym *syms = NULL;
200 * To store the PLTs, we expand the .text section for core module code
201 * and for initialization code.
203 for (s = sechdrs; s < sechdrs_end; ++s) {
204 if (strcmp(".plt", secstrings + s->sh_name) == 0)
205 mod->arch.core.plt = s;
206 else if (strcmp(".init.plt", secstrings + s->sh_name) == 0)
207 mod->arch.init.plt = s;
208 else if (s->sh_type == SHT_SYMTAB)
209 syms = (Elf32_Sym *)s->sh_addr;
212 if (!mod->arch.core.plt || !mod->arch.init.plt) {
213 pr_err("%s: module PLT section(s) missing\n", mod->name);
214 return -ENOEXEC;
216 if (!syms) {
217 pr_err("%s: module symtab section missing\n", mod->name);
218 return -ENOEXEC;
221 for (s = sechdrs + 1; s < sechdrs_end; ++s) {
222 Elf32_Rel *rels = (void *)ehdr + s->sh_offset;
223 int numrels = s->sh_size / sizeof(Elf32_Rel);
224 Elf32_Shdr *dstsec = sechdrs + s->sh_info;
226 if (s->sh_type != SHT_REL)
227 continue;
229 /* ignore relocations that operate on non-exec sections */
230 if (!(dstsec->sh_flags & SHF_EXECINSTR))
231 continue;
233 /* sort by type and symbol index */
234 sort(rels, numrels, sizeof(Elf32_Rel), cmp_rel, NULL);
236 if (strncmp(secstrings + dstsec->sh_name, ".init", 5) != 0)
237 core_plts += count_plts(syms, dstsec->sh_addr, rels,
238 numrels, s->sh_info);
239 else
240 init_plts += count_plts(syms, dstsec->sh_addr, rels,
241 numrels, s->sh_info);
244 mod->arch.core.plt->sh_type = SHT_NOBITS;
245 mod->arch.core.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
246 mod->arch.core.plt->sh_addralign = L1_CACHE_BYTES;
247 mod->arch.core.plt->sh_size = round_up(core_plts * PLT_ENT_SIZE,
248 sizeof(struct plt_entries));
249 mod->arch.core.plt_count = 0;
251 mod->arch.init.plt->sh_type = SHT_NOBITS;
252 mod->arch.init.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
253 mod->arch.init.plt->sh_addralign = L1_CACHE_BYTES;
254 mod->arch.init.plt->sh_size = round_up(init_plts * PLT_ENT_SIZE,
255 sizeof(struct plt_entries));
256 mod->arch.init.plt_count = 0;
258 pr_debug("%s: plt=%x, init.plt=%x\n", __func__,
259 mod->arch.core.plt->sh_size, mod->arch.init.plt->sh_size);
260 return 0;