• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/moduleloader.h>
11 #include <linux/sort.h>
12 
13 #include <asm/kvm_pkvm_module.h>
14 
__get_adrp_add_pair(u64 dst,u64 pc,enum aarch64_insn_register reg)15 static struct plt_entry __get_adrp_add_pair(u64 dst, u64 pc,
16 					    enum aarch64_insn_register reg)
17 {
18 	u32 adrp, add;
19 
20 	adrp = aarch64_insn_gen_adr(pc, dst, reg, AARCH64_INSN_ADR_TYPE_ADRP);
21 	add = aarch64_insn_gen_add_sub_imm(reg, reg, dst % SZ_4K,
22 					   AARCH64_INSN_VARIANT_64BIT,
23 					   AARCH64_INSN_ADSB_ADD);
24 
25 	return (struct plt_entry){ cpu_to_le32(adrp), cpu_to_le32(add) };
26 }
27 
get_plt_entry(u64 dst,void * pc)28 struct plt_entry get_plt_entry(u64 dst, void *pc)
29 {
30 	struct plt_entry plt;
31 	static u32 br;
32 
33 	if (!br)
34 		br = aarch64_insn_gen_branch_reg(AARCH64_INSN_REG_16,
35 						 AARCH64_INSN_BRANCH_NOLINK);
36 
37 	plt = __get_adrp_add_pair(dst, (u64)pc, AARCH64_INSN_REG_16);
38 	plt.br = cpu_to_le32(br);
39 
40 	return plt;
41 }
42 
plt_entries_equal(const struct plt_entry * a,const struct plt_entry * b)43 static bool plt_entries_equal(const struct plt_entry *a,
44 			      const struct plt_entry *b)
45 {
46 	u64 p, q;
47 
48 	/*
49 	 * Check whether both entries refer to the same target:
50 	 * do the cheapest checks first.
51 	 * If the 'add' or 'br' opcodes are different, then the target
52 	 * cannot be the same.
53 	 */
54 	if (a->add != b->add || a->br != b->br)
55 		return false;
56 
57 	p = ALIGN_DOWN((u64)a, SZ_4K);
58 	q = ALIGN_DOWN((u64)b, SZ_4K);
59 
60 	/*
61 	 * If the 'adrp' opcodes are the same then we just need to check
62 	 * that they refer to the same 4k region.
63 	 */
64 	if (a->adrp == b->adrp && p == q)
65 		return true;
66 
67 	return (p + aarch64_insn_adrp_get_offset(le32_to_cpu(a->adrp))) ==
68 	       (q + aarch64_insn_adrp_get_offset(le32_to_cpu(b->adrp)));
69 }
70 
module_emit_plt_entry(struct module * mod,Elf64_Shdr * sechdrs,void * loc,const Elf64_Rela * rela,Elf64_Sym * sym)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 = !within_module_init((unsigned long)loc, mod) ?
76 						&mod->arch.core : &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
module_emit_veneer_for_adrp(struct module * mod,Elf64_Shdr * sechdrs,void * loc,u64 val)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 = !within_module_init((unsigned long)loc, mod) ?
107 						&mod->arch.core : &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 
cmp_rela(const void * a,const void * b)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 
duplicate_rel(const Elf64_Rela * rela,int num)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 
count_plts(Elf64_Sym * syms,Elf64_Rela * rela,int num,Elf64_Word dstidx,Elf_Shdr * dstsec)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 			/*
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 (!cpus_have_final_cap(ARM64_WORKAROUND_843419))
206 				break;
207 
208 			/*
209 			 * Determine the minimal safe alignment for this ADRP
210 			 * instruction: the section alignment at which it is
211 			 * guaranteed not to appear at a vulnerable offset.
212 			 *
213 			 * This comes down to finding the least significant zero
214 			 * bit in bits [11:3] of the section offset, and
215 			 * increasing the section's alignment so that the
216 			 * resulting address of this instruction is guaranteed
217 			 * to equal the offset in that particular bit (as well
218 			 * as all less significant bits). This ensures that the
219 			 * address modulo 4 KB != 0xfff8 or 0xfffc (which would
220 			 * have all ones in bits [11:3])
221 			 */
222 			min_align = 2ULL << ffz(rela[i].r_offset | 0x7);
223 
224 			/*
225 			 * Allocate veneer space for each ADRP that may appear
226 			 * at a vulnerable offset nonetheless. At relocation
227 			 * time, some of these will remain unused since some
228 			 * ADRP instructions can be patched to ADR instructions
229 			 * instead.
230 			 */
231 			if (min_align > SZ_4K)
232 				ret++;
233 			else
234 				dstsec->sh_addralign = max(dstsec->sh_addralign,
235 							   min_align);
236 			break;
237 		}
238 	}
239 
240 	if (cpus_have_final_cap(ARM64_WORKAROUND_843419)) {
241 		/*
242 		 * Add some slack so we can skip PLT slots that may trigger
243 		 * the erratum due to the placement of the ADRP instruction.
244 		 */
245 		ret += DIV_ROUND_UP(ret, (SZ_4K / sizeof(struct plt_entry)));
246 	}
247 
248 	return ret;
249 }
250 
branch_rela_needs_plt(Elf64_Sym * syms,Elf64_Rela * rela,Elf64_Word dstidx)251 static bool branch_rela_needs_plt(Elf64_Sym *syms, Elf64_Rela *rela,
252 				  Elf64_Word dstidx)
253 {
254 
255 	Elf64_Sym *s = syms + ELF64_R_SYM(rela->r_info);
256 
257 	if (s->st_shndx == dstidx)
258 		return false;
259 
260 	return ELF64_R_TYPE(rela->r_info) == R_AARCH64_JUMP26 ||
261 	       ELF64_R_TYPE(rela->r_info) == R_AARCH64_CALL26;
262 }
263 
264 /* Group branch PLT relas at the front end of the array. */
partition_branch_plt_relas(Elf64_Sym * syms,Elf64_Rela * rela,int numrels,Elf64_Word dstidx)265 static int partition_branch_plt_relas(Elf64_Sym *syms, Elf64_Rela *rela,
266 				      int numrels, Elf64_Word dstidx)
267 {
268 	int i = 0, j = numrels - 1;
269 
270 	while (i < j) {
271 		if (branch_rela_needs_plt(syms, &rela[i], dstidx))
272 			i++;
273 		else if (branch_rela_needs_plt(syms, &rela[j], dstidx))
274 			swap(rela[i], rela[j]);
275 		else
276 			j--;
277 	}
278 
279 	return i;
280 }
281 
module_frob_arch_sections(Elf_Ehdr * ehdr,Elf_Shdr * sechdrs,char * secstrings,struct module * mod)282 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
283 			      char *secstrings, struct module *mod)
284 {
285 	unsigned long core_plts = 0;
286 	unsigned long init_plts = 0;
287 	Elf64_Sym *syms = NULL;
288 	Elf_Shdr *pltsec, *tramp = NULL;
289 	int i;
290 
291 	/*
292 	 * Find the empty .plt section so we can expand it to store the PLT
293 	 * entries. Record the symtab address as well.
294 	 */
295 	for (i = 0; i < ehdr->e_shnum; i++) {
296 		if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
297 			mod->arch.core.plt_shndx = i;
298 		else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt"))
299 			mod->arch.init.plt_shndx = i;
300 		else if (!strcmp(secstrings + sechdrs[i].sh_name,
301 				 ".text.ftrace_trampoline"))
302 			tramp = sechdrs + i;
303 		else if (sechdrs[i].sh_type == SHT_SYMTAB)
304 			syms = (Elf64_Sym *)sechdrs[i].sh_addr;
305 	}
306 
307 	if (!mod->arch.core.plt_shndx || !mod->arch.init.plt_shndx) {
308 		pr_err("%s: module PLT section(s) missing\n", mod->name);
309 		return -ENOEXEC;
310 	}
311 	if (!syms) {
312 		pr_err("%s: module symtab section missing\n", mod->name);
313 		return -ENOEXEC;
314 	}
315 
316 	for (i = 0; i < ehdr->e_shnum; i++) {
317 		Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset;
318 		int nents, numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
319 		Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info;
320 
321 		if (sechdrs[i].sh_type != SHT_RELA)
322 			continue;
323 
324 		/* ignore relocations that operate on non-exec sections */
325 		if (!(dstsec->sh_flags & SHF_EXECINSTR))
326 			continue;
327 
328 		/*
329 		 * sort branch relocations requiring a PLT by type, symbol index
330 		 * and addend
331 		 */
332 		nents = partition_branch_plt_relas(syms, rels, numrels,
333 						   sechdrs[i].sh_info);
334 		if (nents)
335 			sort(rels, nents, sizeof(Elf64_Rela), cmp_rela, NULL);
336 
337 		if (!module_init_layout_section(secstrings + dstsec->sh_name))
338 			core_plts += count_plts(syms, rels, numrels,
339 						sechdrs[i].sh_info, dstsec);
340 		else
341 			init_plts += count_plts(syms, rels, numrels,
342 						sechdrs[i].sh_info, dstsec);
343 	}
344 
345 	pltsec = sechdrs + mod->arch.core.plt_shndx;
346 	pltsec->sh_type = SHT_NOBITS;
347 	pltsec->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
348 	pltsec->sh_addralign = L1_CACHE_BYTES;
349 	pltsec->sh_size = (core_plts  + 1) * sizeof(struct plt_entry);
350 	mod->arch.core.plt_num_entries = 0;
351 	mod->arch.core.plt_max_entries = core_plts;
352 
353 	pltsec = sechdrs + mod->arch.init.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 = (init_plts + 1) * sizeof(struct plt_entry);
358 	mod->arch.init.plt_num_entries = 0;
359 	mod->arch.init.plt_max_entries = init_plts;
360 
361 	if (tramp) {
362 		tramp->sh_type = SHT_NOBITS;
363 		tramp->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
364 		tramp->sh_addralign = __alignof__(struct plt_entry);
365 		tramp->sh_size = NR_FTRACE_PLTS * sizeof(struct plt_entry);
366 	}
367 
368 #if IS_ENABLED(CONFIG_KVM)
369 	pkvm_el2_mod_frob_sections(ehdr, sechdrs, secstrings);
370 #endif
371 
372 	return 0;
373 }
374