1 /*
2  * parse_vdso.c: Linux reference vDSO parser
3  * Written by Andrew Lutomirski, 2011-2014.
4  *
5  * This code is meant to be linked in to various programs that run on Linux.
6  * As such, it is available with as few restrictions as possible.  This file
7  * is licensed under the Creative Commons Zero License, version 1.0,
8  * available at http://creativecommons.org/publicdomain/zero/1.0/legalcode
9  *
10  * The vDSO is a regular ELF DSO that the kernel maps into user space when
11  * it starts a program.  It works equally well in statically and dynamically
12  * linked binaries.
13  *
14  * This code is tested on x86.  In principle it should work on any
15  * architecture that has a vDSO.
16  */
17 
18 #include <stdbool.h>
19 #include <stdint.h>
20 #include <string.h>
21 #include <limits.h>
22 #include <elf.h>
23 
24 #include "parse_vdso.h"
25 
26 /* And here's the code. */
27 #ifndef ELF_BITS
28 # if ULONG_MAX > 0xffffffffUL
29 #  define ELF_BITS 64
30 # else
31 #  define ELF_BITS 32
32 # endif
33 #endif
34 
35 #define ELF_BITS_XFORM2(bits, x) Elf##bits##_##x
36 #define ELF_BITS_XFORM(bits, x) ELF_BITS_XFORM2(bits, x)
37 #define ELF(x) ELF_BITS_XFORM(ELF_BITS, x)
38 
39 #ifdef __s390x__
40 #define ELF_HASH_ENTRY ELF(Xword)
41 #else
42 #define ELF_HASH_ENTRY ELF(Word)
43 #endif
44 
45 static struct vdso_info
46 {
47 	bool valid;
48 
49 	/* Load information */
50 	uintptr_t load_addr;
51 	uintptr_t load_offset;  /* load_addr - recorded vaddr */
52 
53 	/* Symbol table */
54 	ELF(Sym) *symtab;
55 	const char *symstrings;
56 	ELF_HASH_ENTRY *bucket, *chain;
57 	ELF_HASH_ENTRY nbucket, nchain;
58 
59 	/* Version table */
60 	ELF(Versym) *versym;
61 	ELF(Verdef) *verdef;
62 } vdso_info;
63 
64 /*
65  * Straight from the ELF specification...and then tweaked slightly, in order to
66  * avoid a few clang warnings.
67  */
elf_hash(const char * name)68 static unsigned long elf_hash(const char *name)
69 {
70 	unsigned long h = 0, g;
71 	const unsigned char *uch_name = (const unsigned char *)name;
72 
73 	while (*uch_name)
74 	{
75 		h = (h << 4) + *uch_name++;
76 		g = h & 0xf0000000;
77 		if (g)
78 			h ^= g >> 24;
79 		h &= ~g;
80 	}
81 	return h;
82 }
83 
vdso_init_from_sysinfo_ehdr(uintptr_t base)84 void vdso_init_from_sysinfo_ehdr(uintptr_t base)
85 {
86 	size_t i;
87 	bool found_vaddr = false;
88 
89 	vdso_info.valid = false;
90 
91 	vdso_info.load_addr = base;
92 
93 	ELF(Ehdr) *hdr = (ELF(Ehdr)*)base;
94 	if (hdr->e_ident[EI_CLASS] !=
95 	    (ELF_BITS == 32 ? ELFCLASS32 : ELFCLASS64)) {
96 		return;  /* Wrong ELF class -- check ELF_BITS */
97 	}
98 
99 	ELF(Phdr) *pt = (ELF(Phdr)*)(vdso_info.load_addr + hdr->e_phoff);
100 	ELF(Dyn) *dyn = 0;
101 
102 	/*
103 	 * We need two things from the segment table: the load offset
104 	 * and the dynamic table.
105 	 */
106 	for (i = 0; i < hdr->e_phnum; i++)
107 	{
108 		if (pt[i].p_type == PT_LOAD && !found_vaddr) {
109 			found_vaddr = true;
110 			vdso_info.load_offset =	base
111 				+ (uintptr_t)pt[i].p_offset
112 				- (uintptr_t)pt[i].p_vaddr;
113 		} else if (pt[i].p_type == PT_DYNAMIC) {
114 			dyn = (ELF(Dyn)*)(base + pt[i].p_offset);
115 		}
116 	}
117 
118 	if (!found_vaddr || !dyn)
119 		return;  /* Failed */
120 
121 	/*
122 	 * Fish out the useful bits of the dynamic table.
123 	 */
124 	ELF_HASH_ENTRY *hash = 0;
125 	vdso_info.symstrings = 0;
126 	vdso_info.symtab = 0;
127 	vdso_info.versym = 0;
128 	vdso_info.verdef = 0;
129 	for (i = 0; dyn[i].d_tag != DT_NULL; i++) {
130 		switch (dyn[i].d_tag) {
131 		case DT_STRTAB:
132 			vdso_info.symstrings = (const char *)
133 				((uintptr_t)dyn[i].d_un.d_ptr
134 				 + vdso_info.load_offset);
135 			break;
136 		case DT_SYMTAB:
137 			vdso_info.symtab = (ELF(Sym) *)
138 				((uintptr_t)dyn[i].d_un.d_ptr
139 				 + vdso_info.load_offset);
140 			break;
141 		case DT_HASH:
142 			hash = (ELF_HASH_ENTRY *)
143 				((uintptr_t)dyn[i].d_un.d_ptr
144 				 + vdso_info.load_offset);
145 			break;
146 		case DT_VERSYM:
147 			vdso_info.versym = (ELF(Versym) *)
148 				((uintptr_t)dyn[i].d_un.d_ptr
149 				 + vdso_info.load_offset);
150 			break;
151 		case DT_VERDEF:
152 			vdso_info.verdef = (ELF(Verdef) *)
153 				((uintptr_t)dyn[i].d_un.d_ptr
154 				 + vdso_info.load_offset);
155 			break;
156 		}
157 	}
158 	if (!vdso_info.symstrings || !vdso_info.symtab || !hash)
159 		return;  /* Failed */
160 
161 	if (!vdso_info.verdef)
162 		vdso_info.versym = 0;
163 
164 	/* Parse the hash table header. */
165 	vdso_info.nbucket = hash[0];
166 	vdso_info.nchain = hash[1];
167 	vdso_info.bucket = &hash[2];
168 	vdso_info.chain = &hash[vdso_info.nbucket + 2];
169 
170 	/* That's all we need. */
171 	vdso_info.valid = true;
172 }
173 
vdso_match_version(ELF (Versym)ver,const char * name,ELF (Word)hash)174 static bool vdso_match_version(ELF(Versym) ver,
175 			       const char *name, ELF(Word) hash)
176 {
177 	/*
178 	 * This is a helper function to check if the version indexed by
179 	 * ver matches name (which hashes to hash).
180 	 *
181 	 * The version definition table is a mess, and I don't know how
182 	 * to do this in better than linear time without allocating memory
183 	 * to build an index.  I also don't know why the table has
184 	 * variable size entries in the first place.
185 	 *
186 	 * For added fun, I can't find a comprehensible specification of how
187 	 * to parse all the weird flags in the table.
188 	 *
189 	 * So I just parse the whole table every time.
190 	 */
191 
192 	/* First step: find the version definition */
193 	ver &= 0x7fff;  /* Apparently bit 15 means "hidden" */
194 	ELF(Verdef) *def = vdso_info.verdef;
195 	while(true) {
196 		if ((def->vd_flags & VER_FLG_BASE) == 0
197 		    && (def->vd_ndx & 0x7fff) == ver)
198 			break;
199 
200 		if (def->vd_next == 0)
201 			return false;  /* No definition. */
202 
203 		def = (ELF(Verdef) *)((char *)def + def->vd_next);
204 	}
205 
206 	/* Now figure out whether it matches. */
207 	ELF(Verdaux) *aux = (ELF(Verdaux)*)((char *)def + def->vd_aux);
208 	return def->vd_hash == hash
209 		&& !strcmp(name, vdso_info.symstrings + aux->vda_name);
210 }
211 
vdso_sym(const char * version,const char * name)212 void *vdso_sym(const char *version, const char *name)
213 {
214 	unsigned long ver_hash;
215 	if (!vdso_info.valid)
216 		return 0;
217 
218 	ver_hash = elf_hash(version);
219 	ELF(Word) chain = vdso_info.bucket[elf_hash(name) % vdso_info.nbucket];
220 
221 	for (; chain != STN_UNDEF; chain = vdso_info.chain[chain]) {
222 		ELF(Sym) *sym = &vdso_info.symtab[chain];
223 
224 		/* Check for a defined global or weak function w/ right name. */
225 		if (ELF64_ST_TYPE(sym->st_info) != STT_FUNC &&
226 		    ELF64_ST_TYPE(sym->st_info) != STT_NOTYPE)
227 			continue;
228 		if (ELF64_ST_BIND(sym->st_info) != STB_GLOBAL &&
229 		    ELF64_ST_BIND(sym->st_info) != STB_WEAK)
230 			continue;
231 		if (sym->st_shndx == SHN_UNDEF)
232 			continue;
233 		if (strcmp(name, vdso_info.symstrings + sym->st_name))
234 			continue;
235 
236 		/* Check symbol version. */
237 		if (vdso_info.versym
238 		    && !vdso_match_version(vdso_info.versym[chain],
239 					   version, ver_hash))
240 			continue;
241 
242 		return (void *)(vdso_info.load_offset + sym->st_value);
243 	}
244 
245 	return 0;
246 }
247 
vdso_init_from_auxv(void * auxv)248 void vdso_init_from_auxv(void *auxv)
249 {
250 	ELF(auxv_t) *elf_auxv = auxv;
251 	for (int i = 0; elf_auxv[i].a_type != AT_NULL; i++)
252 	{
253 		if (elf_auxv[i].a_type == AT_SYSINFO_EHDR) {
254 			vdso_init_from_sysinfo_ehdr(elf_auxv[i].a_un.a_val);
255 			return;
256 		}
257 	}
258 
259 	vdso_info.valid = false;
260 }
261