xref: /openbmc/linux/arch/arm64/kernel/module-plts.c (revision 7f8256ae)
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 
12 static struct plt_entry __get_adrp_add_pair(u64 dst, u64 pc,
13 					    enum aarch64_insn_register reg)
14 {
15 	u32 adrp, add;
16 
17 	adrp = aarch64_insn_gen_adr(pc, dst, reg, AARCH64_INSN_ADR_TYPE_ADRP);
18 	add = aarch64_insn_gen_add_sub_imm(reg, reg, dst % SZ_4K,
19 					   AARCH64_INSN_VARIANT_64BIT,
20 					   AARCH64_INSN_ADSB_ADD);
21 
22 	return (struct plt_entry){ cpu_to_le32(adrp), cpu_to_le32(add) };
23 }
24 
25 struct plt_entry get_plt_entry(u64 dst, void *pc)
26 {
27 	struct plt_entry plt;
28 	static u32 br;
29 
30 	if (!br)
31 		br = aarch64_insn_gen_branch_reg(AARCH64_INSN_REG_16,
32 						 AARCH64_INSN_BRANCH_NOLINK);
33 
34 	plt = __get_adrp_add_pair(dst, (u64)pc, AARCH64_INSN_REG_16);
35 	plt.br = cpu_to_le32(br);
36 
37 	return plt;
38 }
39 
40 static bool plt_entries_equal(const struct plt_entry *a,
41 			      const struct plt_entry *b)
42 {
43 	u64 p, q;
44 
45 	/*
46 	 * Check whether both entries refer to the same target:
47 	 * do the cheapest checks first.
48 	 * If the 'add' or 'br' opcodes are different, then the target
49 	 * cannot be the same.
50 	 */
51 	if (a->add != b->add || a->br != b->br)
52 		return false;
53 
54 	p = ALIGN_DOWN((u64)a, SZ_4K);
55 	q = ALIGN_DOWN((u64)b, SZ_4K);
56 
57 	/*
58 	 * If the 'adrp' opcodes are the same then we just need to check
59 	 * that they refer to the same 4k region.
60 	 */
61 	if (a->adrp == b->adrp && p == q)
62 		return true;
63 
64 	return (p + aarch64_insn_adrp_get_offset(le32_to_cpu(a->adrp))) ==
65 	       (q + aarch64_insn_adrp_get_offset(le32_to_cpu(b->adrp)));
66 }
67 
68 u64 module_emit_plt_entry(struct module *mod, Elf64_Shdr *sechdrs,
69 			  void *loc, const Elf64_Rela *rela,
70 			  Elf64_Sym *sym)
71 {
72 	struct mod_plt_sec *pltsec = !within_module_init((unsigned long)loc, mod) ?
73 						&mod->arch.core : &mod->arch.init;
74 	struct plt_entry *plt = (struct plt_entry *)sechdrs[pltsec->plt_shndx].sh_addr;
75 	int i = pltsec->plt_num_entries;
76 	int j = i - 1;
77 	u64 val = sym->st_value + rela->r_addend;
78 
79 	if (is_forbidden_offset_for_adrp(&plt[i].adrp))
80 		i++;
81 
82 	plt[i] = get_plt_entry(val, &plt[i]);
83 
84 	/*
85 	 * Check if the entry we just created is a duplicate. Given that the
86 	 * relocations are sorted, this will be the last entry we allocated.
87 	 * (if one exists).
88 	 */
89 	if (j >= 0 && plt_entries_equal(plt + i, plt + j))
90 		return (u64)&plt[j];
91 
92 	pltsec->plt_num_entries += i - j;
93 	if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
94 		return 0;
95 
96 	return (u64)&plt[i];
97 }
98 
99 #ifdef CONFIG_ARM64_ERRATUM_843419
100 u64 module_emit_veneer_for_adrp(struct module *mod, Elf64_Shdr *sechdrs,
101 				void *loc, u64 val)
102 {
103 	struct mod_plt_sec *pltsec = !within_module_init((unsigned long)loc, mod) ?
104 						&mod->arch.core : &mod->arch.init;
105 	struct plt_entry *plt = (struct plt_entry *)sechdrs[pltsec->plt_shndx].sh_addr;
106 	int i = pltsec->plt_num_entries++;
107 	u32 br;
108 	int rd;
109 
110 	if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
111 		return 0;
112 
113 	if (is_forbidden_offset_for_adrp(&plt[i].adrp))
114 		i = pltsec->plt_num_entries++;
115 
116 	/* get the destination register of the ADRP instruction */
117 	rd = aarch64_insn_decode_register(AARCH64_INSN_REGTYPE_RD,
118 					  le32_to_cpup((__le32 *)loc));
119 
120 	br = aarch64_insn_gen_branch_imm((u64)&plt[i].br, (u64)loc + 4,
121 					 AARCH64_INSN_BRANCH_NOLINK);
122 
123 	plt[i] = __get_adrp_add_pair(val, (u64)&plt[i], rd);
124 	plt[i].br = cpu_to_le32(br);
125 
126 	return (u64)&plt[i];
127 }
128 #endif
129 
130 #define cmp_3way(a, b)	((a) < (b) ? -1 : (a) > (b))
131 
132 static int cmp_rela(const void *a, const void *b)
133 {
134 	const Elf64_Rela *x = a, *y = b;
135 	int i;
136 
137 	/* sort by type, symbol index and addend */
138 	i = cmp_3way(ELF64_R_TYPE(x->r_info), ELF64_R_TYPE(y->r_info));
139 	if (i == 0)
140 		i = cmp_3way(ELF64_R_SYM(x->r_info), ELF64_R_SYM(y->r_info));
141 	if (i == 0)
142 		i = cmp_3way(x->r_addend, y->r_addend);
143 	return i;
144 }
145 
146 static bool duplicate_rel(const Elf64_Rela *rela, int num)
147 {
148 	/*
149 	 * Entries are sorted by type, symbol index and addend. That means
150 	 * that, if a duplicate entry exists, it must be in the preceding
151 	 * slot.
152 	 */
153 	return num > 0 && cmp_rela(rela + num, rela + num - 1) == 0;
154 }
155 
156 static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
157 			       Elf64_Word dstidx, Elf_Shdr *dstsec)
158 {
159 	unsigned int ret = 0;
160 	Elf64_Sym *s;
161 	int i;
162 
163 	for (i = 0; i < num; i++) {
164 		u64 min_align;
165 
166 		switch (ELF64_R_TYPE(rela[i].r_info)) {
167 		case R_AARCH64_JUMP26:
168 		case R_AARCH64_CALL26:
169 			if (!IS_ENABLED(CONFIG_RANDOMIZE_BASE))
170 				break;
171 
172 			/*
173 			 * We only have to consider branch targets that resolve
174 			 * to symbols that are defined in a different section.
175 			 * This is not simply a heuristic, it is a fundamental
176 			 * limitation, since there is no guaranteed way to emit
177 			 * PLT entries sufficiently close to the branch if the
178 			 * section size exceeds the range of a branch
179 			 * instruction. So ignore relocations against defined
180 			 * symbols if they live in the same section as the
181 			 * relocation target.
182 			 */
183 			s = syms + ELF64_R_SYM(rela[i].r_info);
184 			if (s->st_shndx == dstidx)
185 				break;
186 
187 			/*
188 			 * Jump relocations with non-zero addends against
189 			 * undefined symbols are supported by the ELF spec, but
190 			 * do not occur in practice (e.g., 'jump n bytes past
191 			 * the entry point of undefined function symbol f').
192 			 * So we need to support them, but there is no need to
193 			 * take them into consideration when trying to optimize
194 			 * this code. So let's only check for duplicates when
195 			 * the addend is zero: this allows us to record the PLT
196 			 * entry address in the symbol table itself, rather than
197 			 * having to search the list for duplicates each time we
198 			 * emit one.
199 			 */
200 			if (rela[i].r_addend != 0 || !duplicate_rel(rela, i))
201 				ret++;
202 			break;
203 		case R_AARCH64_ADR_PREL_PG_HI21_NC:
204 		case R_AARCH64_ADR_PREL_PG_HI21:
205 			if (!IS_ENABLED(CONFIG_ARM64_ERRATUM_843419) ||
206 			    !cpus_have_const_cap(ARM64_WORKAROUND_843419))
207 				break;
208 
209 			/*
210 			 * Determine the minimal safe alignment for this ADRP
211 			 * instruction: the section alignment at which it is
212 			 * guaranteed not to appear at a vulnerable offset.
213 			 *
214 			 * This comes down to finding the least significant zero
215 			 * bit in bits [11:3] of the section offset, and
216 			 * increasing the section's alignment so that the
217 			 * resulting address of this instruction is guaranteed
218 			 * to equal the offset in that particular bit (as well
219 			 * as all less significant bits). This ensures that the
220 			 * address modulo 4 KB != 0xfff8 or 0xfffc (which would
221 			 * have all ones in bits [11:3])
222 			 */
223 			min_align = 2ULL << ffz(rela[i].r_offset | 0x7);
224 
225 			/*
226 			 * Allocate veneer space for each ADRP that may appear
227 			 * at a vulnerable offset nonetheless. At relocation
228 			 * time, some of these will remain unused since some
229 			 * ADRP instructions can be patched to ADR instructions
230 			 * instead.
231 			 */
232 			if (min_align > SZ_4K)
233 				ret++;
234 			else
235 				dstsec->sh_addralign = max(dstsec->sh_addralign,
236 							   min_align);
237 			break;
238 		}
239 	}
240 
241 	if (IS_ENABLED(CONFIG_ARM64_ERRATUM_843419) &&
242 	    cpus_have_const_cap(ARM64_WORKAROUND_843419))
243 		/*
244 		 * Add some slack so we can skip PLT slots that may trigger
245 		 * the erratum due to the placement of the ADRP instruction.
246 		 */
247 		ret += DIV_ROUND_UP(ret, (SZ_4K / sizeof(struct plt_entry)));
248 
249 	return ret;
250 }
251 
252 static bool branch_rela_needs_plt(Elf64_Sym *syms, Elf64_Rela *rela,
253 				  Elf64_Word dstidx)
254 {
255 
256 	Elf64_Sym *s = syms + ELF64_R_SYM(rela->r_info);
257 
258 	if (s->st_shndx == dstidx)
259 		return false;
260 
261 	return ELF64_R_TYPE(rela->r_info) == R_AARCH64_JUMP26 ||
262 	       ELF64_R_TYPE(rela->r_info) == R_AARCH64_CALL26;
263 }
264 
265 /* Group branch PLT relas at the front end of the array. */
266 static int partition_branch_plt_relas(Elf64_Sym *syms, Elf64_Rela *rela,
267 				      int numrels, Elf64_Word dstidx)
268 {
269 	int i = 0, j = numrels - 1;
270 
271 	if (!IS_ENABLED(CONFIG_RANDOMIZE_BASE))
272 		return 0;
273 
274 	while (i < j) {
275 		if (branch_rela_needs_plt(syms, &rela[i], dstidx))
276 			i++;
277 		else if (branch_rela_needs_plt(syms, &rela[j], dstidx))
278 			swap(rela[i], rela[j]);
279 		else
280 			j--;
281 	}
282 
283 	return i;
284 }
285 
286 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
287 			      char *secstrings, struct module *mod)
288 {
289 	unsigned long core_plts = 0;
290 	unsigned long init_plts = 0;
291 	Elf64_Sym *syms = NULL;
292 	Elf_Shdr *pltsec, *tramp = NULL;
293 	int i;
294 
295 	/*
296 	 * Find the empty .plt section so we can expand it to store the PLT
297 	 * entries. Record the symtab address as well.
298 	 */
299 	for (i = 0; i < ehdr->e_shnum; i++) {
300 		if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
301 			mod->arch.core.plt_shndx = i;
302 		else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt"))
303 			mod->arch.init.plt_shndx = i;
304 		else if (!strcmp(secstrings + sechdrs[i].sh_name,
305 				 ".text.ftrace_trampoline"))
306 			tramp = sechdrs + i;
307 		else if (sechdrs[i].sh_type == SHT_SYMTAB)
308 			syms = (Elf64_Sym *)sechdrs[i].sh_addr;
309 	}
310 
311 	if (!mod->arch.core.plt_shndx || !mod->arch.init.plt_shndx) {
312 		pr_err("%s: module PLT section(s) missing\n", mod->name);
313 		return -ENOEXEC;
314 	}
315 	if (!syms) {
316 		pr_err("%s: module symtab section missing\n", mod->name);
317 		return -ENOEXEC;
318 	}
319 
320 	for (i = 0; i < ehdr->e_shnum; i++) {
321 		Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset;
322 		int nents, numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
323 		Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info;
324 
325 		if (sechdrs[i].sh_type != SHT_RELA)
326 			continue;
327 
328 		/* ignore relocations that operate on non-exec sections */
329 		if (!(dstsec->sh_flags & SHF_EXECINSTR))
330 			continue;
331 
332 		/*
333 		 * sort branch relocations requiring a PLT by type, symbol index
334 		 * and addend
335 		 */
336 		nents = partition_branch_plt_relas(syms, rels, numrels,
337 						   sechdrs[i].sh_info);
338 		if (nents)
339 			sort(rels, nents, sizeof(Elf64_Rela), cmp_rela, NULL);
340 
341 		if (!str_has_prefix(secstrings + dstsec->sh_name, ".init"))
342 			core_plts += count_plts(syms, rels, numrels,
343 						sechdrs[i].sh_info, dstsec);
344 		else
345 			init_plts += count_plts(syms, rels, numrels,
346 						sechdrs[i].sh_info, dstsec);
347 	}
348 
349 	pltsec = sechdrs + mod->arch.core.plt_shndx;
350 	pltsec->sh_type = SHT_NOBITS;
351 	pltsec->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
352 	pltsec->sh_addralign = L1_CACHE_BYTES;
353 	pltsec->sh_size = (core_plts  + 1) * sizeof(struct plt_entry);
354 	mod->arch.core.plt_num_entries = 0;
355 	mod->arch.core.plt_max_entries = core_plts;
356 
357 	pltsec = sechdrs + mod->arch.init.plt_shndx;
358 	pltsec->sh_type = SHT_NOBITS;
359 	pltsec->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
360 	pltsec->sh_addralign = L1_CACHE_BYTES;
361 	pltsec->sh_size = (init_plts + 1) * sizeof(struct plt_entry);
362 	mod->arch.init.plt_num_entries = 0;
363 	mod->arch.init.plt_max_entries = init_plts;
364 
365 	if (tramp) {
366 		tramp->sh_type = SHT_NOBITS;
367 		tramp->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
368 		tramp->sh_addralign = __alignof__(struct plt_entry);
369 		tramp->sh_size = NR_FTRACE_PLTS * sizeof(struct plt_entry);
370 	}
371 
372 	return 0;
373 }
374