aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--arch/arm/include/asm/module.h6
-rw-r--r--arch/arm/kernel/module-plts.c243
-rw-r--r--arch/arm/kernel/module.lds3
3 files changed, 147 insertions, 105 deletions
diff --git a/arch/arm/include/asm/module.h b/arch/arm/include/asm/module.h
index e358b7966c06..464748b9fd7d 100644
--- a/arch/arm/include/asm/module.h
+++ b/arch/arm/include/asm/module.h
@@ -23,10 +23,8 @@ struct mod_arch_specific {
23 struct unwind_table *unwind[ARM_SEC_MAX]; 23 struct unwind_table *unwind[ARM_SEC_MAX];
24#endif 24#endif
25#ifdef CONFIG_ARM_MODULE_PLTS 25#ifdef CONFIG_ARM_MODULE_PLTS
26 struct elf32_shdr *core_plt; 26 struct elf32_shdr *plt;
27 struct elf32_shdr *init_plt; 27 int plt_count;
28 int core_plt_count;
29 int init_plt_count;
30#endif 28#endif
31}; 29};
32 30
diff --git a/arch/arm/kernel/module-plts.c b/arch/arm/kernel/module-plts.c
index 0c7efc3446c0..3a5cba90c971 100644
--- a/arch/arm/kernel/module-plts.c
+++ b/arch/arm/kernel/module-plts.c
@@ -9,6 +9,7 @@
9#include <linux/elf.h> 9#include <linux/elf.h>
10#include <linux/kernel.h> 10#include <linux/kernel.h>
11#include <linux/module.h> 11#include <linux/module.h>
12#include <linux/sort.h>
12 13
13#include <asm/cache.h> 14#include <asm/cache.h>
14#include <asm/opcodes.h> 15#include <asm/opcodes.h>
@@ -30,154 +31,198 @@ struct plt_entries {
30 u32 lit[PLT_ENT_COUNT]; 31 u32 lit[PLT_ENT_COUNT];
31}; 32};
32 33
33static bool in_init(const struct module *mod, u32 addr) 34u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val)
34{ 35{
35 return addr - (u32)mod->init_layout.base < mod->init_layout.size; 36 struct plt_entries *plt = (struct plt_entries *)mod->arch.plt->sh_addr;
37 int idx = 0;
38
39 /*
40 * Look for an existing entry pointing to 'val'. Given that the
41 * relocations are sorted, this will be the last entry we allocated.
42 * (if one exists).
43 */
44 if (mod->arch.plt_count > 0) {
45 plt += (mod->arch.plt_count - 1) / PLT_ENT_COUNT;
46 idx = (mod->arch.plt_count - 1) % PLT_ENT_COUNT;
47
48 if (plt->lit[idx] == val)
49 return (u32)&plt->ldr[idx];
50
51 idx = (idx + 1) % PLT_ENT_COUNT;
52 if (!idx)
53 plt++;
54 }
55
56 mod->arch.plt_count++;
57 BUG_ON(mod->arch.plt_count * PLT_ENT_SIZE > mod->arch.plt->sh_size);
58
59 if (!idx)
60 /* Populate a new set of entries */
61 *plt = (struct plt_entries){
62 { [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, },
63 { val, }
64 };
65 else
66 plt->lit[idx] = val;
67
68 return (u32)&plt->ldr[idx];
36} 69}
37 70
38u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val) 71#define cmp_3way(a,b) ((a) < (b) ? -1 : (a) > (b))
72
73static int cmp_rel(const void *a, const void *b)
39{ 74{
40 struct plt_entries *plt, *plt_end; 75 const Elf32_Rel *x = a, *y = b;
41 int c, *count; 76 int i;
42
43 if (in_init(mod, loc)) {
44 plt = (void *)mod->arch.init_plt->sh_addr;
45 plt_end = (void *)plt + mod->arch.init_plt->sh_size;
46 count = &mod->arch.init_plt_count;
47 } else {
48 plt = (void *)mod->arch.core_plt->sh_addr;
49 plt_end = (void *)plt + mod->arch.core_plt->sh_size;
50 count = &mod->arch.core_plt_count;
51 }
52 77
53 /* Look for an existing entry pointing to 'val' */ 78 /* sort by type and symbol index */
54 for (c = *count; plt < plt_end; c -= PLT_ENT_COUNT, plt++) { 79 i = cmp_3way(ELF32_R_TYPE(x->r_info), ELF32_R_TYPE(y->r_info));
55 int i; 80 if (i == 0)
56 81 i = cmp_3way(ELF32_R_SYM(x->r_info), ELF32_R_SYM(y->r_info));
57 if (!c) { 82 return i;
58 /* Populate a new set of entries */ 83}
59 *plt = (struct plt_entries){ 84
60 { [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, }, 85static bool is_zero_addend_relocation(Elf32_Addr base, const Elf32_Rel *rel)
61 { val, } 86{
62 }; 87 u32 *tval = (u32 *)(base + rel->r_offset);
63 ++*count; 88
64 return (u32)plt->ldr; 89 /*
65 } 90 * Do a bitwise compare on the raw addend rather than fully decoding
66 for (i = 0; i < PLT_ENT_COUNT; i++) { 91 * the offset and doing an arithmetic comparison.
67 if (!plt->lit[i]) { 92 * Note that a zero-addend jump/call relocation is encoded taking the
68 plt->lit[i] = val; 93 * PC bias into account, i.e., -8 for ARM and -4 for Thumb2.
69 ++*count; 94 */
70 } 95 switch (ELF32_R_TYPE(rel->r_info)) {
71 if (plt->lit[i] == val) 96 u16 upper, lower;
72 return (u32)&plt->ldr[i]; 97
73 } 98 case R_ARM_THM_CALL:
99 case R_ARM_THM_JUMP24:
100 upper = __mem_to_opcode_thumb16(((u16 *)tval)[0]);
101 lower = __mem_to_opcode_thumb16(((u16 *)tval)[1]);
102
103 return (upper & 0x7ff) == 0x7ff && (lower & 0x2fff) == 0x2ffe;
104
105 case R_ARM_CALL:
106 case R_ARM_PC24:
107 case R_ARM_JUMP24:
108 return (__mem_to_opcode_arm(*tval) & 0xffffff) == 0xfffffe;
74 } 109 }
75 BUG(); 110 BUG();
76} 111}
77 112
78static int duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num, 113static bool duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num)
79 u32 mask)
80{ 114{
81 u32 *loc1, *loc2; 115 const Elf32_Rel *prev;
82 int i;
83 116
84 for (i = 0; i < num; i++) { 117 /*
85 if (rel[i].r_info != rel[num].r_info) 118 * Entries are sorted by type and symbol index. That means that,
86 continue; 119 * if a duplicate entry exists, it must be in the preceding
120 * slot.
121 */
122 if (!num)
123 return false;
87 124
88 /* 125 prev = rel + num - 1;
89 * Identical relocation types against identical symbols can 126 return cmp_rel(rel + num, prev) == 0 &&
90 * still result in different PLT entries if the addend in the 127 is_zero_addend_relocation(base, prev);
91 * place is different. So resolve the target of the relocation
92 * to compare the values.
93 */
94 loc1 = (u32 *)(base + rel[i].r_offset);
95 loc2 = (u32 *)(base + rel[num].r_offset);
96 if (((*loc1 ^ *loc2) & mask) == 0)
97 return 1;
98 }
99 return 0;
100} 128}
101 129
102/* Count how many PLT entries we may need */ 130/* Count how many PLT entries we may need */
103static unsigned int count_plts(Elf32_Addr base, const Elf32_Rel *rel, int num) 131static unsigned int count_plts(const Elf32_Sym *syms, Elf32_Addr base,
132 const Elf32_Rel *rel, int num)
104{ 133{
105 unsigned int ret = 0; 134 unsigned int ret = 0;
135 const Elf32_Sym *s;
106 int i; 136 int i;
107 137
108 /* 138 for (i = 0; i < num; i++) {
109 * Sure, this is order(n^2), but it's usually short, and not
110 * time critical
111 */
112 for (i = 0; i < num; i++)
113 switch (ELF32_R_TYPE(rel[i].r_info)) { 139 switch (ELF32_R_TYPE(rel[i].r_info)) {
114 case R_ARM_CALL: 140 case R_ARM_CALL:
115 case R_ARM_PC24: 141 case R_ARM_PC24:
116 case R_ARM_JUMP24: 142 case R_ARM_JUMP24:
117 if (!duplicate_rel(base, rel, i,
118 __opcode_to_mem_arm(0x00ffffff)))
119 ret++;
120 break;
121#ifdef CONFIG_THUMB2_KERNEL
122 case R_ARM_THM_CALL: 143 case R_ARM_THM_CALL:
123 case R_ARM_THM_JUMP24: 144 case R_ARM_THM_JUMP24:
124 if (!duplicate_rel(base, rel, i, 145 /*
125 __opcode_to_mem_thumb32(0x07ff2fff))) 146 * We only have to consider branch targets that resolve
147 * to undefined symbols. This is not simply a heuristic,
148 * it is a fundamental limitation, since the PLT itself
149 * is part of the module, and needs to be within range
150 * as well, so modules can never grow beyond that limit.
151 */
152 s = syms + ELF32_R_SYM(rel[i].r_info);
153 if (s->st_shndx != SHN_UNDEF)
154 break;
155
156 /*
157 * Jump relocations with non-zero addends against
158 * undefined symbols are supported by the ELF spec, but
159 * do not occur in practice (e.g., 'jump n bytes past
160 * the entry point of undefined function symbol f').
161 * So we need to support them, but there is no need to
162 * take them into consideration when trying to optimize
163 * this code. So let's only check for duplicates when
164 * the addend is zero.
165 */
166 if (!is_zero_addend_relocation(base, rel + i) ||
167 !duplicate_rel(base, rel, i))
126 ret++; 168 ret++;
127#endif
128 } 169 }
170 }
129 return ret; 171 return ret;
130} 172}
131 173
132int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs, 174int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
133 char *secstrings, struct module *mod) 175 char *secstrings, struct module *mod)
134{ 176{
135 unsigned long core_plts = 0, init_plts = 0; 177 unsigned long plts = 0;
136 Elf32_Shdr *s, *sechdrs_end = sechdrs + ehdr->e_shnum; 178 Elf32_Shdr *s, *sechdrs_end = sechdrs + ehdr->e_shnum;
179 Elf32_Sym *syms = NULL;
137 180
138 /* 181 /*
139 * To store the PLTs, we expand the .text section for core module code 182 * To store the PLTs, we expand the .text section for core module code
140 * and the .init.text section for initialization code. 183 * and for initialization code.
141 */ 184 */
142 for (s = sechdrs; s < sechdrs_end; ++s) 185 for (s = sechdrs; s < sechdrs_end; ++s) {
143 if (strcmp(".core.plt", secstrings + s->sh_name) == 0) 186 if (strcmp(".plt", secstrings + s->sh_name) == 0)
144 mod->arch.core_plt = s; 187 mod->arch.plt = s;
145 else if (strcmp(".init.plt", secstrings + s->sh_name) == 0) 188 else if (s->sh_type == SHT_SYMTAB)
146 mod->arch.init_plt = s; 189 syms = (Elf32_Sym *)s->sh_addr;
147 190 }
148 if (!mod->arch.core_plt || !mod->arch.init_plt) { 191
149 pr_err("%s: sections missing\n", mod->name); 192 if (!mod->arch.plt) {
193 pr_err("%s: module PLT section missing\n", mod->name);
194 return -ENOEXEC;
195 }
196 if (!syms) {
197 pr_err("%s: module symtab section missing\n", mod->name);
150 return -ENOEXEC; 198 return -ENOEXEC;
151 } 199 }
152 200
153 for (s = sechdrs + 1; s < sechdrs_end; ++s) { 201 for (s = sechdrs + 1; s < sechdrs_end; ++s) {
154 const Elf32_Rel *rels = (void *)ehdr + s->sh_offset; 202 Elf32_Rel *rels = (void *)ehdr + s->sh_offset;
155 int numrels = s->sh_size / sizeof(Elf32_Rel); 203 int numrels = s->sh_size / sizeof(Elf32_Rel);
156 Elf32_Shdr *dstsec = sechdrs + s->sh_info; 204 Elf32_Shdr *dstsec = sechdrs + s->sh_info;
157 205
158 if (s->sh_type != SHT_REL) 206 if (s->sh_type != SHT_REL)
159 continue; 207 continue;
160 208
161 if (strstr(secstrings + s->sh_name, ".init")) 209 /* ignore relocations that operate on non-exec sections */
162 init_plts += count_plts(dstsec->sh_addr, rels, numrels); 210 if (!(dstsec->sh_flags & SHF_EXECINSTR))
163 else 211 continue;
164 core_plts += count_plts(dstsec->sh_addr, rels, numrels); 212
213 /* sort by type and symbol index */
214 sort(rels, numrels, sizeof(Elf32_Rel), cmp_rel, NULL);
215
216 plts += count_plts(syms, dstsec->sh_addr, rels, numrels);
165 } 217 }
166 218
167 mod->arch.core_plt->sh_type = SHT_NOBITS; 219 mod->arch.plt->sh_type = SHT_NOBITS;
168 mod->arch.core_plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC; 220 mod->arch.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
169 mod->arch.core_plt->sh_addralign = L1_CACHE_BYTES; 221 mod->arch.plt->sh_addralign = L1_CACHE_BYTES;
170 mod->arch.core_plt->sh_size = round_up(core_plts * PLT_ENT_SIZE, 222 mod->arch.plt->sh_size = round_up(plts * PLT_ENT_SIZE,
171 sizeof(struct plt_entries)); 223 sizeof(struct plt_entries));
172 mod->arch.core_plt_count = 0; 224 mod->arch.plt_count = 0;
173 225
174 mod->arch.init_plt->sh_type = SHT_NOBITS; 226 pr_debug("%s: plt=%x\n", __func__, mod->arch.plt->sh_size);
175 mod->arch.init_plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
176 mod->arch.init_plt->sh_addralign = L1_CACHE_BYTES;
177 mod->arch.init_plt->sh_size = round_up(init_plts * PLT_ENT_SIZE,
178 sizeof(struct plt_entries));
179 mod->arch.init_plt_count = 0;
180 pr_debug("%s: core.plt=%x, init.plt=%x\n", __func__,
181 mod->arch.core_plt->sh_size, mod->arch.init_plt->sh_size);
182 return 0; 227 return 0;
183} 228}
diff --git a/arch/arm/kernel/module.lds b/arch/arm/kernel/module.lds
index 3682fa107918..05881e2b414c 100644
--- a/arch/arm/kernel/module.lds
+++ b/arch/arm/kernel/module.lds
@@ -1,4 +1,3 @@
1SECTIONS { 1SECTIONS {
2 .core.plt : { BYTE(0) } 2 .plt : { BYTE(0) }
3 .init.plt : { BYTE(0) }
4} 3}
generated by cgit v1.2.2 (git 2.25.0) at 2025-06-01 12:19:58 -0400