xref: /openbmc/linux/arch/arm64/kernel/module-plts.c (revision b737eecd)
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 bool plt_entries_equal(const struct plt_entry *a, const struct plt_entry *b)
41 {
42 	u64 p, q;
43 
44 	/*
45 	 * Check whether both entries refer to the same target:
46 	 * do the cheapest checks first.
47 	 * If the 'add' or 'br' opcodes are different, then the target
48 	 * cannot be the same.
49 	 */
50 	if (a->add != b->add || a->br != b->br)
51 		return false;
52 
53 	p = ALIGN_DOWN((u64)a, SZ_4K);
54 	q = ALIGN_DOWN((u64)b, SZ_4K);
55 
56 	/*
57 	 * If the 'adrp' opcodes are the same then we just need to check
58 	 * that they refer to the same 4k region.
59 	 */
60 	if (a->adrp == b->adrp && p == q)
61 		return true;
62 
63 	return (p + aarch64_insn_adrp_get_offset(le32_to_cpu(a->adrp))) ==
64 	       (q + aarch64_insn_adrp_get_offset(le32_to_cpu(b->adrp)));
65 }
66 
67 static bool in_init(const struct module *mod, void *loc)
68 {
69 	return (u64)loc - (u64)mod->init_layout.base < mod->init_layout.size;
70 }
71 
72 u64 module_emit_plt_entry(struct module *mod, Elf64_Shdr *sechdrs,
73 			  void *loc, const Elf64_Rela *rela,
74 			  Elf64_Sym *sym)
75 {
76 	struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
77 							  &mod->arch.init;
78 	struct plt_entry *plt = (struct plt_entry *)sechdrs[pltsec->plt_shndx].sh_addr;
79 	int i = pltsec->plt_num_entries;
80 	int j = i - 1;
81 	u64 val = sym->st_value + rela->r_addend;
82 
83 	if (is_forbidden_offset_for_adrp(&plt[i].adrp))
84 		i++;
85 
86 	plt[i] = get_plt_entry(val, &plt[i]);
87 
88 	/*
89 	 * Check if the entry we just created is a duplicate. Given that the
90 	 * relocations are sorted, this will be the last entry we allocated.
91 	 * (if one exists).
92 	 */
93 	if (j >= 0 && plt_entries_equal(plt + i, plt + j))
94 		return (u64)&plt[j];
95 
96 	pltsec->plt_num_entries += i - j;
97 	if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
98 		return 0;
99 
100 	return (u64)&plt[i];
101 }
102 
103 #ifdef CONFIG_ARM64_ERRATUM_843419
104 u64 module_emit_veneer_for_adrp(struct module *mod, Elf64_Shdr *sechdrs,
105 				void *loc, u64 val)
106 {
107 	struct mod_plt_sec *pltsec = !in_init(mod, loc) ? &mod->arch.core :
108 							  &mod->arch.init;
109 	struct plt_entry *plt = (struct plt_entry *)sechdrs[pltsec->plt_shndx].sh_addr;
110 	int i = pltsec->plt_num_entries++;
111 	u32 br;
112 	int rd;
113 
114 	if (WARN_ON(pltsec->plt_num_entries > pltsec->plt_max_entries))
115 		return 0;
116 
117 	if (is_forbidden_offset_for_adrp(&plt[i].adrp))
118 		i = pltsec->plt_num_entries++;
119 
120 	/* get the destination register of the ADRP instruction */
121 	rd = aarch64_insn_decode_register(AARCH64_INSN_REGTYPE_RD,
122 					  le32_to_cpup((__le32 *)loc));
123 
124 	br = aarch64_insn_gen_branch_imm((u64)&plt[i].br, (u64)loc + 4,
125 					 AARCH64_INSN_BRANCH_NOLINK);
126 
127 	plt[i] = __get_adrp_add_pair(val, (u64)&plt[i], rd);
128 	plt[i].br = cpu_to_le32(br);
129 
130 	return (u64)&plt[i];
131 }
132 #endif
133 
134 #define cmp_3way(a, b)	((a) < (b) ? -1 : (a) > (b))
135 
136 static int cmp_rela(const void *a, const void *b)
137 {
138 	const Elf64_Rela *x = a, *y = b;
139 	int i;
140 
141 	/* sort by type, symbol index and addend */
142 	i = cmp_3way(ELF64_R_TYPE(x->r_info), ELF64_R_TYPE(y->r_info));
143 	if (i == 0)
144 		i = cmp_3way(ELF64_R_SYM(x->r_info), ELF64_R_SYM(y->r_info));
145 	if (i == 0)
146 		i = cmp_3way(x->r_addend, y->r_addend);
147 	return i;
148 }
149 
150 static bool duplicate_rel(const Elf64_Rela *rela, int num)
151 {
152 	/*
153 	 * Entries are sorted by type, symbol index and addend. That means
154 	 * that, if a duplicate entry exists, it must be in the preceding
155 	 * slot.
156 	 */
157 	return num > 0 && cmp_rela(rela + num, rela + num - 1) == 0;
158 }
159 
160 static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
161 			       Elf64_Word dstidx, Elf_Shdr *dstsec)
162 {
163 	unsigned int ret = 0;
164 	Elf64_Sym *s;
165 	int i;
166 
167 	for (i = 0; i < num; i++) {
168 		u64 min_align;
169 
170 		switch (ELF64_R_TYPE(rela[i].r_info)) {
171 		case R_AARCH64_JUMP26:
172 		case R_AARCH64_CALL26:
173 			if (!IS_ENABLED(CONFIG_RANDOMIZE_BASE))
174 				break;
175 
176 			/*
177 			 * We only have to consider branch targets that resolve
178 			 * to symbols that are defined in a different section.
179 			 * This is not simply a heuristic, it is a fundamental
180 			 * limitation, since there is no guaranteed way to emit
181 			 * PLT entries sufficiently close to the branch if the
182 			 * section size exceeds the range of a branch
183 			 * instruction. So ignore relocations against defined
184 			 * symbols if they live in the same section as the
185 			 * relocation target.
186 			 */
187 			s = syms + ELF64_R_SYM(rela[i].r_info);
188 			if (s->st_shndx == dstidx)
189 				break;
190 
191 			/*
192 			 * Jump relocations with non-zero addends against
193 			 * undefined symbols are supported by the ELF spec, but
194 			 * do not occur in practice (e.g., 'jump n bytes past
195 			 * the entry point of undefined function symbol f').
196 			 * So we need to support them, but there is no need to
197 			 * take them into consideration when trying to optimize
198 			 * this code. So let's only check for duplicates when
199 			 * the addend is zero: this allows us to record the PLT
200 			 * entry address in the symbol table itself, rather than
201 			 * having to search the list for duplicates each time we
202 			 * emit one.
203 			 */
204 			if (rela[i].r_addend != 0 || !duplicate_rel(rela, i))
205 				ret++;
206 			break;
207 		case R_AARCH64_ADR_PREL_PG_HI21_NC:
208 		case R_AARCH64_ADR_PREL_PG_HI21:
209 			if (!IS_ENABLED(CONFIG_ARM64_ERRATUM_843419) ||
210 			    !cpus_have_const_cap(ARM64_WORKAROUND_843419))
211 				break;
212 
213 			/*
214 			 * Determine the minimal safe alignment for this ADRP
215 			 * instruction: the section alignment at which it is
216 			 * guaranteed not to appear at a vulnerable offset.
217 			 *
218 			 * This comes down to finding the least significant zero
219 			 * bit in bits [11:3] of the section offset, and
220 			 * increasing the section's alignment so that the
221 			 * resulting address of this instruction is guaranteed
222 			 * to equal the offset in that particular bit (as well
223 			 * as all less signficant bits). This ensures that the
224 			 * address modulo 4 KB != 0xfff8 or 0xfffc (which would
225 			 * have all ones in bits [11:3])
226 			 */
227 			min_align = 2ULL << ffz(rela[i].r_offset | 0x7);
228 
229 			/*
230 			 * Allocate veneer space for each ADRP that may appear
231 			 * at a vulnerable offset nonetheless. At relocation
232 			 * time, some of these will remain unused since some
233 			 * ADRP instructions can be patched to ADR instructions
234 			 * instead.
235 			 */
236 			if (min_align > SZ_4K)
237 				ret++;
238 			else
239 				dstsec->sh_addralign = max(dstsec->sh_addralign,
240 							   min_align);
241 			break;
242 		}
243 	}
244 
245 	if (IS_ENABLED(CONFIG_ARM64_ERRATUM_843419) &&
246 	    cpus_have_const_cap(ARM64_WORKAROUND_843419))
247 		/*
248 		 * Add some slack so we can skip PLT slots that may trigger
249 		 * the erratum due to the placement of the ADRP instruction.
250 		 */
251 		ret += DIV_ROUND_UP(ret, (SZ_4K / sizeof(struct plt_entry)));
252 
253 	return ret;
254 }
255 
256 static bool branch_rela_needs_plt(Elf64_Sym *syms, Elf64_Rela *rela,
257 				  Elf64_Word dstidx)
258 {
259 
260 	Elf64_Sym *s = syms + ELF64_R_SYM(rela->r_info);
261 
262 	if (s->st_shndx == dstidx)
263 		return false;
264 
265 	return ELF64_R_TYPE(rela->r_info) == R_AARCH64_JUMP26 ||
266 	       ELF64_R_TYPE(rela->r_info) == R_AARCH64_CALL26;
267 }
268 
269 /* Group branch PLT relas at the front end of the array. */
270 static int partition_branch_plt_relas(Elf64_Sym *syms, Elf64_Rela *rela,
271 				      int numrels, Elf64_Word dstidx)
272 {
273 	int i = 0, j = numrels - 1;
274 
275 	if (!IS_ENABLED(CONFIG_RANDOMIZE_BASE))
276 		return 0;
277 
278 	while (i < j) {
279 		if (branch_rela_needs_plt(syms, &rela[i], dstidx))
280 			i++;
281 		else if (branch_rela_needs_plt(syms, &rela[j], dstidx))
282 			swap(rela[i], rela[j]);
283 		else
284 			j--;
285 	}
286 
287 	return i;
288 }
289 
290 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
291 			      char *secstrings, struct module *mod)
292 {
293 	unsigned long core_plts = 0;
294 	unsigned long init_plts = 0;
295 	Elf64_Sym *syms = NULL;
296 	Elf_Shdr *pltsec, *tramp = NULL;
297 	int i;
298 
299 	/*
300 	 * Find the empty .plt section so we can expand it to store the PLT
301 	 * entries. Record the symtab address as well.
302 	 */
303 	for (i = 0; i < ehdr->e_shnum; i++) {
304 		if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
305 			mod->arch.core.plt_shndx = i;
306 		else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt"))
307 			mod->arch.init.plt_shndx = i;
308 		else if (!strcmp(secstrings + sechdrs[i].sh_name,
309 				 ".text.ftrace_trampoline"))
310 			tramp = sechdrs + i;
311 		else if (sechdrs[i].sh_type == SHT_SYMTAB)
312 			syms = (Elf64_Sym *)sechdrs[i].sh_addr;
313 	}
314 
315 	if (!mod->arch.core.plt_shndx || !mod->arch.init.plt_shndx) {
316 		pr_err("%s: module PLT section(s) missing\n", mod->name);
317 		return -ENOEXEC;
318 	}
319 	if (!syms) {
320 		pr_err("%s: module symtab section missing\n", mod->name);
321 		return -ENOEXEC;
322 	}
323 
324 	for (i = 0; i < ehdr->e_shnum; i++) {
325 		Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset;
326 		int nents, numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
327 		Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info;
328 
329 		if (sechdrs[i].sh_type != SHT_RELA)
330 			continue;
331 
332 		/* ignore relocations that operate on non-exec sections */
333 		if (!(dstsec->sh_flags & SHF_EXECINSTR))
334 			continue;
335 
336 		/*
337 		 * sort branch relocations requiring a PLT by type, symbol index
338 		 * and addend
339 		 */
340 		nents = partition_branch_plt_relas(syms, rels, numrels,
341 						   sechdrs[i].sh_info);
342 		if (nents)
343 			sort(rels, nents, sizeof(Elf64_Rela), cmp_rela, NULL);
344 
345 		if (!str_has_prefix(secstrings + dstsec->sh_name, ".init"))
346 			core_plts += count_plts(syms, rels, numrels,
347 						sechdrs[i].sh_info, dstsec);
348 		else
349 			init_plts += count_plts(syms, rels, numrels,
350 						sechdrs[i].sh_info, dstsec);
351 	}
352 
353 	pltsec = sechdrs + mod->arch.core.plt_shndx;
354 	pltsec->sh_type = SHT_NOBITS;
355 	pltsec->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
356 	pltsec->sh_addralign = L1_CACHE_BYTES;
357 	pltsec->sh_size = (core_plts  + 1) * sizeof(struct plt_entry);
358 	mod->arch.core.plt_num_entries = 0;
359 	mod->arch.core.plt_max_entries = core_plts;
360 
361 	pltsec = sechdrs + mod->arch.init.plt_shndx;
362 	pltsec->sh_type = SHT_NOBITS;
363 	pltsec->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
364 	pltsec->sh_addralign = L1_CACHE_BYTES;
365 	pltsec->sh_size = (init_plts + 1) * sizeof(struct plt_entry);
366 	mod->arch.init.plt_num_entries = 0;
367 	mod->arch.init.plt_max_entries = init_plts;
368 
369 	if (tramp) {
370 		tramp->sh_type = SHT_NOBITS;
371 		tramp->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
372 		tramp->sh_addralign = __alignof__(struct plt_entry);
373 		tramp->sh_size = NR_FTRACE_PLTS * sizeof(struct plt_entry);
374 	}
375 
376 	return 0;
377 }
378