xref: /openbmc/linux/arch/arm/kernel/module-plts.c (revision 3ddc8b84)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2014-2017 Linaro Ltd. <ard.biesheuvel@linaro.org>
4  */
5 
6 #include <linux/elf.h>
7 #include <linux/ftrace.h>
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/sort.h>
11 #include <linux/moduleloader.h>
12 
13 #include <asm/cache.h>
14 #include <asm/opcodes.h>
15 
16 #ifdef CONFIG_THUMB2_KERNEL
17 #define PLT_ENT_LDR		__opcode_to_mem_thumb32(0xf8dff000 | \
18 							(PLT_ENT_STRIDE - 4))
19 #else
20 #define PLT_ENT_LDR		__opcode_to_mem_arm(0xe59ff000 | \
21 						    (PLT_ENT_STRIDE - 8))
22 #endif
23 
24 static const u32 fixed_plts[] = {
25 #ifdef CONFIG_DYNAMIC_FTRACE
26 	FTRACE_ADDR,
27 	MCOUNT_ADDR,
28 #endif
29 };
30 
31 static void prealloc_fixed(struct mod_plt_sec *pltsec, struct plt_entries *plt)
32 {
33 	int i;
34 
35 	if (!ARRAY_SIZE(fixed_plts) || pltsec->plt_count)
36 		return;
37 	pltsec->plt_count = ARRAY_SIZE(fixed_plts);
38 
39 	for (i = 0; i < ARRAY_SIZE(plt->ldr); ++i)
40 		plt->ldr[i] = PLT_ENT_LDR;
41 
42 	BUILD_BUG_ON(sizeof(fixed_plts) > sizeof(plt->lit));
43 	memcpy(plt->lit, fixed_plts, sizeof(fixed_plts));
44 }
45 
46 u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val)
47 {
48 	struct mod_plt_sec *pltsec = !within_module_init(loc, mod) ?
49 						&mod->arch.core : &mod->arch.init;
50 	struct plt_entries *plt;
51 	int idx;
52 
53 	/* cache the address, ELF header is available only during module load */
54 	if (!pltsec->plt_ent)
55 		pltsec->plt_ent = (struct plt_entries *)pltsec->plt->sh_addr;
56 	plt = pltsec->plt_ent;
57 
58 	prealloc_fixed(pltsec, plt);
59 
60 	for (idx = 0; idx < ARRAY_SIZE(fixed_plts); ++idx)
61 		if (plt->lit[idx] == val)
62 			return (u32)&plt->ldr[idx];
63 
64 	idx = 0;
65 	/*
66 	 * Look for an existing entry pointing to 'val'. Given that the
67 	 * relocations are sorted, this will be the last entry we allocated.
68 	 * (if one exists).
69 	 */
70 	if (pltsec->plt_count > 0) {
71 		plt += (pltsec->plt_count - 1) / PLT_ENT_COUNT;
72 		idx = (pltsec->plt_count - 1) % PLT_ENT_COUNT;
73 
74 		if (plt->lit[idx] == val)
75 			return (u32)&plt->ldr[idx];
76 
77 		idx = (idx + 1) % PLT_ENT_COUNT;
78 		if (!idx)
79 			plt++;
80 	}
81 
82 	pltsec->plt_count++;
83 	BUG_ON(pltsec->plt_count * PLT_ENT_SIZE > pltsec->plt->sh_size);
84 
85 	if (!idx)
86 		/* Populate a new set of entries */
87 		*plt = (struct plt_entries){
88 			{ [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, },
89 			{ val, }
90 		};
91 	else
92 		plt->lit[idx] = val;
93 
94 	return (u32)&plt->ldr[idx];
95 }
96 
97 #define cmp_3way(a,b)	((a) < (b) ? -1 : (a) > (b))
98 
99 static int cmp_rel(const void *a, const void *b)
100 {
101 	const Elf32_Rel *x = a, *y = b;
102 	int i;
103 
104 	/* sort by type and symbol index */
105 	i = cmp_3way(ELF32_R_TYPE(x->r_info), ELF32_R_TYPE(y->r_info));
106 	if (i == 0)
107 		i = cmp_3way(ELF32_R_SYM(x->r_info), ELF32_R_SYM(y->r_info));
108 	return i;
109 }
110 
111 static bool is_zero_addend_relocation(Elf32_Addr base, const Elf32_Rel *rel)
112 {
113 	u32 *tval = (u32 *)(base + rel->r_offset);
114 
115 	/*
116 	 * Do a bitwise compare on the raw addend rather than fully decoding
117 	 * the offset and doing an arithmetic comparison.
118 	 * Note that a zero-addend jump/call relocation is encoded taking the
119 	 * PC bias into account, i.e., -8 for ARM and -4 for Thumb2.
120 	 */
121 	switch (ELF32_R_TYPE(rel->r_info)) {
122 		u16 upper, lower;
123 
124 	case R_ARM_THM_CALL:
125 	case R_ARM_THM_JUMP24:
126 		upper = __mem_to_opcode_thumb16(((u16 *)tval)[0]);
127 		lower = __mem_to_opcode_thumb16(((u16 *)tval)[1]);
128 
129 		return (upper & 0x7ff) == 0x7ff && (lower & 0x2fff) == 0x2ffe;
130 
131 	case R_ARM_CALL:
132 	case R_ARM_PC24:
133 	case R_ARM_JUMP24:
134 		return (__mem_to_opcode_arm(*tval) & 0xffffff) == 0xfffffe;
135 	}
136 	BUG();
137 }
138 
139 static bool duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num)
140 {
141 	const Elf32_Rel *prev;
142 
143 	/*
144 	 * Entries are sorted by type and symbol index. That means that,
145 	 * if a duplicate entry exists, it must be in the preceding
146 	 * slot.
147 	 */
148 	if (!num)
149 		return false;
150 
151 	prev = rel + num - 1;
152 	return cmp_rel(rel + num, prev) == 0 &&
153 	       is_zero_addend_relocation(base, prev);
154 }
155 
156 /* Count how many PLT entries we may need */
157 static unsigned int count_plts(const Elf32_Sym *syms, Elf32_Addr base,
158 			       const Elf32_Rel *rel, int num, Elf32_Word dstidx)
159 {
160 	unsigned int ret = 0;
161 	const Elf32_Sym *s;
162 	int i;
163 
164 	for (i = 0; i < num; i++) {
165 		switch (ELF32_R_TYPE(rel[i].r_info)) {
166 		case R_ARM_CALL:
167 		case R_ARM_PC24:
168 		case R_ARM_JUMP24:
169 		case R_ARM_THM_CALL:
170 		case R_ARM_THM_JUMP24:
171 			/*
172 			 * We only have to consider branch targets that resolve
173 			 * to symbols that are defined in a different section.
174 			 * This is not simply a heuristic, it is a fundamental
175 			 * limitation, since there is no guaranteed way to emit
176 			 * PLT entries sufficiently close to the branch if the
177 			 * section size exceeds the range of a branch
178 			 * instruction. So ignore relocations against defined
179 			 * symbols if they live in the same section as the
180 			 * relocation target.
181 			 */
182 			s = syms + ELF32_R_SYM(rel[i].r_info);
183 			if (s->st_shndx == dstidx)
184 				break;
185 
186 			/*
187 			 * Jump relocations with non-zero addends against
188 			 * undefined symbols are supported by the ELF spec, but
189 			 * do not occur in practice (e.g., 'jump n bytes past
190 			 * the entry point of undefined function symbol f').
191 			 * So we need to support them, but there is no need to
192 			 * take them into consideration when trying to optimize
193 			 * this code. So let's only check for duplicates when
194 			 * the addend is zero. (Note that calls into the core
195 			 * module via init PLT entries could involve section
196 			 * relative symbol references with non-zero addends, for
197 			 * which we may end up emitting duplicates, but the init
198 			 * PLT is released along with the rest of the .init
199 			 * region as soon as module loading completes.)
200 			 */
201 			if (!is_zero_addend_relocation(base, rel + i) ||
202 			    !duplicate_rel(base, rel, i))
203 				ret++;
204 		}
205 	}
206 	return ret;
207 }
208 
209 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
210 			      char *secstrings, struct module *mod)
211 {
212 	unsigned long core_plts = ARRAY_SIZE(fixed_plts);
213 	unsigned long init_plts = ARRAY_SIZE(fixed_plts);
214 	Elf32_Shdr *s, *sechdrs_end = sechdrs + ehdr->e_shnum;
215 	Elf32_Sym *syms = NULL;
216 
217 	/*
218 	 * To store the PLTs, we expand the .text section for core module code
219 	 * and for initialization code.
220 	 */
221 	for (s = sechdrs; s < sechdrs_end; ++s) {
222 		if (strcmp(".plt", secstrings + s->sh_name) == 0)
223 			mod->arch.core.plt = s;
224 		else if (strcmp(".init.plt", secstrings + s->sh_name) == 0)
225 			mod->arch.init.plt = s;
226 		else if (s->sh_type == SHT_SYMTAB)
227 			syms = (Elf32_Sym *)s->sh_addr;
228 	}
229 
230 	if (!mod->arch.core.plt || !mod->arch.init.plt) {
231 		pr_err("%s: module PLT section(s) missing\n", mod->name);
232 		return -ENOEXEC;
233 	}
234 	if (!syms) {
235 		pr_err("%s: module symtab section missing\n", mod->name);
236 		return -ENOEXEC;
237 	}
238 
239 	for (s = sechdrs + 1; s < sechdrs_end; ++s) {
240 		Elf32_Rel *rels = (void *)ehdr + s->sh_offset;
241 		int numrels = s->sh_size / sizeof(Elf32_Rel);
242 		Elf32_Shdr *dstsec = sechdrs + s->sh_info;
243 
244 		if (s->sh_type != SHT_REL)
245 			continue;
246 
247 		/* ignore relocations that operate on non-exec sections */
248 		if (!(dstsec->sh_flags & SHF_EXECINSTR))
249 			continue;
250 
251 		/* sort by type and symbol index */
252 		sort(rels, numrels, sizeof(Elf32_Rel), cmp_rel, NULL);
253 
254 		if (!module_init_layout_section(secstrings + dstsec->sh_name))
255 			core_plts += count_plts(syms, dstsec->sh_addr, rels,
256 						numrels, s->sh_info);
257 		else
258 			init_plts += count_plts(syms, dstsec->sh_addr, rels,
259 						numrels, s->sh_info);
260 	}
261 
262 	mod->arch.core.plt->sh_type = SHT_NOBITS;
263 	mod->arch.core.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
264 	mod->arch.core.plt->sh_addralign = L1_CACHE_BYTES;
265 	mod->arch.core.plt->sh_size = round_up(core_plts * PLT_ENT_SIZE,
266 					       sizeof(struct plt_entries));
267 	mod->arch.core.plt_count = 0;
268 	mod->arch.core.plt_ent = NULL;
269 
270 	mod->arch.init.plt->sh_type = SHT_NOBITS;
271 	mod->arch.init.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
272 	mod->arch.init.plt->sh_addralign = L1_CACHE_BYTES;
273 	mod->arch.init.plt->sh_size = round_up(init_plts * PLT_ENT_SIZE,
274 					       sizeof(struct plt_entries));
275 	mod->arch.init.plt_count = 0;
276 	mod->arch.init.plt_ent = NULL;
277 
278 	pr_debug("%s: plt=%x, init.plt=%x\n", __func__,
279 		 mod->arch.core.plt->sh_size, mod->arch.init.plt->sh_size);
280 	return 0;
281 }
282 
283 bool in_module_plt(unsigned long loc)
284 {
285 	struct module *mod;
286 	bool ret;
287 
288 	preempt_disable();
289 	mod = __module_text_address(loc);
290 	ret = mod && (loc - (u32)mod->arch.core.plt_ent < mod->arch.core.plt_count * PLT_ENT_SIZE ||
291 		      loc - (u32)mod->arch.init.plt_ent < mod->arch.init.plt_count * PLT_ENT_SIZE);
292 	preempt_enable();
293 
294 	return ret;
295 }
296