xref: /openbmc/linux/arch/arm64/kernel/module-plts.c (revision a44e4f3a)
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/kernel.h>
8 #include <linux/module.h>
9 #include <linux/sort.h>
10 
11 static struct plt_entry __get_adrp_add_pair(u64 dst, u64 pc,
12 					    enum aarch64_insn_register reg)
13 {
14 	u32 adrp, add;
15 
16 	adrp = aarch64_insn_gen_adr(pc, dst, reg, AARCH64_INSN_ADR_TYPE_ADRP);
17 	add = aarch64_insn_gen_add_sub_imm(reg, reg, dst % SZ_4K,
18 					   AARCH64_INSN_VARIANT_64BIT,
19 					   AARCH64_INSN_ADSB_ADD);
20 
21 	return (struct plt_entry){ cpu_to_le32(adrp), cpu_to_le32(add) };
22 }
23 
24 struct plt_entry get_plt_entry(u64 dst, void *pc)
25 {
26 	struct plt_entry plt;
27 	static u32 br;
28 
29 	if (!br)
30 		br = aarch64_insn_gen_branch_reg(AARCH64_INSN_REG_16,
31 						 AARCH64_INSN_BRANCH_NOLINK);
32 
33 	plt = __get_adrp_add_pair(dst, (u64)pc, AARCH64_INSN_REG_16);
34 	plt.br = cpu_to_le32(br);
35 
36 	return plt;
37 }
38 
39 bool plt_entries_equal(const struct plt_entry *a, const struct plt_entry *b)
40 {
41 	u64 p, q;
42 
43 	/*
44 	 * Check whether both entries refer to the same target:
45 	 * do the cheapest checks first.
46 	 * If the 'add' or 'br' opcodes are different, then the target
47 	 * cannot be the same.
48 	 */
49 	if (a->add != b->add || a->br != b->br)
50 		return false;
51 
52 	p = ALIGN_DOWN((u64)a, SZ_4K);
53 	q = ALIGN_DOWN((u64)b, SZ_4K);
54 
55 	/*
56 	 * If the 'adrp' opcodes are the same then we just need to check
57 	 * that they refer to the same 4k region.
58 	 */
59 	if (a->adrp == b->adrp && p == q)
60 		return true;
61 
62 	return (p + aarch64_insn_adrp_get_offset(le32_to_cpu(a->adrp))) ==
63 	       (q + aarch64_insn_adrp_get_offset(le32_to_cpu(b->adrp)));
64 }
65 
66 static bool in_init(const struct module *mod, void *loc)
67 {
68 	return (u64)loc - (u64)mod->init_layout.base < mod->init_layout.size;
69 }
70 
71 u64 module_emit_plt_entry(struct module *mod, Elf64_Shdr *sechdrs,
72 			  void *loc, const Elf64_Rela *rela,
73 			  Elf64_Sym *sym)
74 {
75 	struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
76 							  &mod->arch.init;
77 	struct plt_entry *plt = (struct plt_entry *)sechdrs[pltsec->plt_shndx].sh_addr;
78 	int i = pltsec->plt_num_entries;
79 	int j = i - 1;
80 	u64 val = sym->st_value + rela->r_addend;
81 
82 	if (is_forbidden_offset_for_adrp(&plt[i].adrp))
83 		i++;
84 
85 	plt[i] = get_plt_entry(val, &plt[i]);
86 
87 	/*
88 	 * Check if the entry we just created is a duplicate. Given that the
89 	 * relocations are sorted, this will be the last entry we allocated.
90 	 * (if one exists).
91 	 */
92 	if (j >= 0 && plt_entries_equal(plt + i, plt + j))
93 		return (u64)&plt[j];
94 
95 	pltsec->plt_num_entries += i - j;
96 	if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
97 		return 0;
98 
99 	return (u64)&plt[i];
100 }
101 
102 #ifdef CONFIG_ARM64_ERRATUM_843419
103 u64 module_emit_veneer_for_adrp(struct module *mod, Elf64_Shdr *sechdrs,
104 				void *loc, u64 val)
105 {
106 	struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
107 							  &mod->arch.init;
108 	struct plt_entry *plt = (struct plt_entry *)sechdrs[pltsec->plt_shndx].sh_addr;
109 	int i = pltsec->plt_num_entries++;
110 	u32 br;
111 	int rd;
112 
113 	if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
114 		return 0;
115 
116 	if (is_forbidden_offset_for_adrp(&plt[i].adrp))
117 		i = pltsec->plt_num_entries++;
118 
119 	/* get the destination register of the ADRP instruction */
120 	rd = aarch64_insn_decode_register(AARCH64_INSN_REGTYPE_RD,
121 					  le32_to_cpup((__le32 *)loc));
122 
123 	br = aarch64_insn_gen_branch_imm((u64)&plt[i].br, (u64)loc + 4,
124 					 AARCH64_INSN_BRANCH_NOLINK);
125 
126 	plt[i] = __get_adrp_add_pair(val, (u64)&plt[i], rd);
127 	plt[i].br = cpu_to_le32(br);
128 
129 	return (u64)&plt[i];
130 }
131 #endif
132 
133 #define cmp_3way(a,b)	((a) < (b) ? -1 : (a) > (b))
134 
135 static int cmp_rela(const void *a, const void *b)
136 {
137 	const Elf64_Rela *x = a, *y = b;
138 	int i;
139 
140 	/* sort by type, symbol index and addend */
141 	i = cmp_3way(ELF64_R_TYPE(x->r_info), ELF64_R_TYPE(y->r_info));
142 	if (i == 0)
143 		i = cmp_3way(ELF64_R_SYM(x->r_info), ELF64_R_SYM(y->r_info));
144 	if (i == 0)
145 		i = cmp_3way(x->r_addend, y->r_addend);
146 	return i;
147 }
148 
149 static bool duplicate_rel(const Elf64_Rela *rela, int num)
150 {
151 	/*
152 	 * Entries are sorted by type, symbol index and addend. That means
153 	 * that, if a duplicate entry exists, it must be in the preceding
154 	 * slot.
155 	 */
156 	return num > 0 && cmp_rela(rela + num, rela + num - 1) == 0;
157 }
158 
159 static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
160 			       Elf64_Word dstidx, Elf_Shdr *dstsec)
161 {
162 	unsigned int ret = 0;
163 	Elf64_Sym *s;
164 	int i;
165 
166 	for (i = 0; i < num; i++) {
167 		u64 min_align;
168 
169 		switch (ELF64_R_TYPE(rela[i].r_info)) {
170 		case R_AARCH64_JUMP26:
171 		case R_AARCH64_CALL26:
172 			if (!IS_ENABLED(CONFIG_RANDOMIZE_BASE))
173 				break;
174 
175 			/*
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
184 			 * relocation target.
185 			 */
186 			s = syms + ELF64_R_SYM(rela[i].r_info);
187 			if (s->st_shndx == dstidx)
188 				break;
189 
190 			/*
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: this allows us to record the PLT
199 			 * entry address in the symbol table itself, rather than
200 			 * having to search the list for duplicates each time we
201 			 * emit one.
202 			 */
203 			if (rela[i].r_addend != 0 || !duplicate_rel(rela, i))
204 				ret++;
205 			break;
206 		case R_AARCH64_ADR_PREL_PG_HI21_NC:
207 		case R_AARCH64_ADR_PREL_PG_HI21:
208 			if (!IS_ENABLED(CONFIG_ARM64_ERRATUM_843419) ||
209 			    !cpus_have_const_cap(ARM64_WORKAROUND_843419))
210 				break;
211 
212 			/*
213 			 * Determine the minimal safe alignment for this ADRP
214 			 * instruction: the section alignment at which it is
215 			 * guaranteed not to appear at a vulnerable offset.
216 			 *
217 			 * This comes down to finding the least significant zero
218 			 * bit in bits [11:3] of the section offset, and
219 			 * increasing the section's alignment so that the
220 			 * resulting address of this instruction is guaranteed
221 			 * to equal the offset in that particular bit (as well
222 			 * as all less signficant bits). This ensures that the
223 			 * address modulo 4 KB != 0xfff8 or 0xfffc (which would
224 			 * have all ones in bits [11:3])
225 			 */
226 			min_align = 2ULL << ffz(rela[i].r_offset | 0x7);
227 
228 			/*
229 			 * Allocate veneer space for each ADRP that may appear
230 			 * at a vulnerable offset nonetheless. At relocation
231 			 * time, some of these will remain unused since some
232 			 * ADRP instructions can be patched to ADR instructions
233 			 * instead.
234 			 */
235 			if (min_align > SZ_4K)
236 				ret++;
237 			else
238 				dstsec->sh_addralign = max(dstsec->sh_addralign,
239 							   min_align);
240 			break;
241 		}
242 	}
243 
244 	if (IS_ENABLED(CONFIG_ARM64_ERRATUM_843419) &&
245 	    cpus_have_const_cap(ARM64_WORKAROUND_843419))
246 		/*
247 		 * Add some slack so we can skip PLT slots that may trigger
248 		 * the erratum due to the placement of the ADRP instruction.
249 		 */
250 		ret += DIV_ROUND_UP(ret, (SZ_4K / sizeof(struct plt_entry)));
251 
252 	return ret;
253 }
254 
255 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
256 			      char *secstrings, struct module *mod)
257 {
258 	unsigned long core_plts = 0;
259 	unsigned long init_plts = 0;
260 	Elf64_Sym *syms = NULL;
261 	Elf_Shdr *pltsec, *tramp = NULL;
262 	int i;
263 
264 	/*
265 	 * Find the empty .plt section so we can expand it to store the PLT
266 	 * entries. Record the symtab address as well.
267 	 */
268 	for (i = 0; i < ehdr->e_shnum; i++) {
269 		if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
270 			mod->arch.core.plt_shndx = i;
271 		else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt"))
272 			mod->arch.init.plt_shndx = i;
273 		else if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE) &&
274 			 !strcmp(secstrings + sechdrs[i].sh_name,
275 				 ".text.ftrace_trampoline"))
276 			tramp = sechdrs + i;
277 		else if (sechdrs[i].sh_type == SHT_SYMTAB)
278 			syms = (Elf64_Sym *)sechdrs[i].sh_addr;
279 	}
280 
281 	if (!mod->arch.core.plt_shndx || !mod->arch.init.plt_shndx) {
282 		pr_err("%s: module PLT section(s) missing\n", mod->name);
283 		return -ENOEXEC;
284 	}
285 	if (!syms) {
286 		pr_err("%s: module symtab section missing\n", mod->name);
287 		return -ENOEXEC;
288 	}
289 
290 	for (i = 0; i < ehdr->e_shnum; i++) {
291 		Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset;
292 		int numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
293 		Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info;
294 
295 		if (sechdrs[i].sh_type != SHT_RELA)
296 			continue;
297 
298 		/* ignore relocations that operate on non-exec sections */
299 		if (!(dstsec->sh_flags & SHF_EXECINSTR))
300 			continue;
301 
302 		/* sort by type, symbol index and addend */
303 		sort(rels, numrels, sizeof(Elf64_Rela), cmp_rela, NULL);
304 
305 		if (!str_has_prefix(secstrings + dstsec->sh_name, ".init"))
306 			core_plts += count_plts(syms, rels, numrels,
307 						sechdrs[i].sh_info, dstsec);
308 		else
309 			init_plts += count_plts(syms, rels, numrels,
310 						sechdrs[i].sh_info, dstsec);
311 	}
312 
313 	pltsec = sechdrs + mod->arch.core.plt_shndx;
314 	pltsec->sh_type = SHT_NOBITS;
315 	pltsec->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
316 	pltsec->sh_addralign = L1_CACHE_BYTES;
317 	pltsec->sh_size = (core_plts  + 1) * sizeof(struct plt_entry);
318 	mod->arch.core.plt_num_entries = 0;
319 	mod->arch.core.plt_max_entries = core_plts;
320 
321 	pltsec = sechdrs + mod->arch.init.plt_shndx;
322 	pltsec->sh_type = SHT_NOBITS;
323 	pltsec->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
324 	pltsec->sh_addralign = L1_CACHE_BYTES;
325 	pltsec->sh_size = (init_plts + 1) * sizeof(struct plt_entry);
326 	mod->arch.init.plt_num_entries = 0;
327 	mod->arch.init.plt_max_entries = init_plts;
328 
329 	if (tramp) {
330 		tramp->sh_type = SHT_NOBITS;
331 		tramp->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
332 		tramp->sh_addralign = __alignof__(struct plt_entry);
333 		tramp->sh_size = sizeof(struct plt_entry);
334 	}
335 
336 	return 0;
337 }
338