1 /*
2 * Copyright (C) 2016 The Android Open Source Project
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * * Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * * Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in
12 * the documentation and/or other materials provided with the
13 * distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
16 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
17 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
18 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
19 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
21 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
22 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
23 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
25 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 */
28
29 #include "linker_main.h"
30
31 #include <link.h>
32 #include <stdlib.h>
33 #include <sys/auxv.h>
34 #include <sys/prctl.h>
35
36 #include "linker.h"
37 #include "linker_auxv.h"
38 #include "linker_cfi.h"
39 #include "linker_debug.h"
40 #include "linker_debuggerd.h"
41 #include "linker_gdb_support.h"
42 #include "linker_globals.h"
43 #include "linker_phdr.h"
44 #include "linker_relocate.h"
45 #include "linker_relocs.h"
46 #include "linker_tls.h"
47 #include "linker_utils.h"
48
49 #include "private/KernelArgumentBlock.h"
50 #include "private/bionic_call_ifunc_resolver.h"
51 #include "private/bionic_globals.h"
52 #include "private/bionic_tls.h"
53
54 #include "android-base/unique_fd.h"
55 #include "android-base/strings.h"
56 #include "android-base/stringprintf.h"
57
58 #include <async_safe/log.h>
59 #include <bionic/libc_init_common.h>
60 #include <bionic/pthread_internal.h>
61
62 #include <vector>
63
64 __LIBC_HIDDEN__ extern "C" void _start();
65
66 static ElfW(Addr) get_elf_exec_load_bias(const ElfW(Ehdr)* elf);
67
68 static void get_elf_base_from_phdr(const ElfW(Phdr)* phdr_table, size_t phdr_count,
69 ElfW(Addr)* base, ElfW(Addr)* load_bias);
70
71 static void set_bss_vma_name(soinfo* si);
72
73 void __libc_init_mte(const memtag_dynamic_entries_t* memtag_dynamic_entries, const void* phdr_start,
74 size_t phdr_count, uintptr_t load_bias, void* stack_top);
75
76 // These should be preserved static to avoid emitting
77 // RELATIVE relocations for the part of the code running
78 // before linker links itself.
79
80 // TODO (dimtiry): remove somain, rename solist to solist_head
81 static soinfo* solist;
82 static soinfo* sonext;
83 static soinfo* somain; // main process, always the one after libdl_info
84 static soinfo* solinker;
85 static soinfo* vdso; // vdso if present
86
solist_add_soinfo(soinfo * si)87 void solist_add_soinfo(soinfo* si) {
88 sonext->next = si;
89 sonext = si;
90 }
91
solist_remove_soinfo(soinfo * si)92 bool solist_remove_soinfo(soinfo* si) {
93 soinfo *prev = nullptr, *trav;
94 for (trav = solist; trav != nullptr; trav = trav->next) {
95 if (trav == si) {
96 break;
97 }
98 prev = trav;
99 }
100
101 if (trav == nullptr) {
102 // si was not in solist
103 PRINT("name \"%s\"@%p is not in solist!", si->get_realpath(), si);
104 return false;
105 }
106
107 // prev will never be null, because the first entry in solist is
108 // always the static libdl_info.
109 CHECK(prev != nullptr);
110 prev->next = si->next;
111 if (si == sonext) {
112 sonext = prev;
113 }
114
115 return true;
116 }
117
solist_get_head()118 soinfo* solist_get_head() {
119 return solist;
120 }
121
solist_get_somain()122 soinfo* solist_get_somain() {
123 return somain;
124 }
125
solist_get_vdso()126 soinfo* solist_get_vdso() {
127 return vdso;
128 }
129
130 bool g_is_ldd;
131 int g_ld_debug_verbosity;
132
133 static std::vector<std::string> g_ld_preload_names;
134
135 static std::vector<soinfo*> g_ld_preloads;
136
parse_path(const char * path,const char * delimiters,std::vector<std::string> * resolved_paths)137 static void parse_path(const char* path, const char* delimiters,
138 std::vector<std::string>* resolved_paths) {
139 std::vector<std::string> paths;
140 split_path(path, delimiters, &paths);
141 resolve_paths(paths, resolved_paths);
142 }
143
parse_LD_LIBRARY_PATH(const char * path)144 static void parse_LD_LIBRARY_PATH(const char* path) {
145 std::vector<std::string> ld_libary_paths;
146 parse_path(path, ":", &ld_libary_paths);
147 g_default_namespace.set_ld_library_paths(std::move(ld_libary_paths));
148 }
149
parse_LD_PRELOAD(const char * path)150 static void parse_LD_PRELOAD(const char* path) {
151 g_ld_preload_names.clear();
152 if (path != nullptr) {
153 // We have historically supported ':' as well as ' ' in LD_PRELOAD.
154 g_ld_preload_names = android::base::Split(path, " :");
155 g_ld_preload_names.erase(std::remove_if(g_ld_preload_names.begin(), g_ld_preload_names.end(),
156 [](const std::string& s) { return s.empty(); }),
157 g_ld_preload_names.end());
158 }
159 }
160
161 // An empty list of soinfos
162 static soinfo_list_t g_empty_list;
163
add_vdso()164 static void add_vdso() {
165 ElfW(Ehdr)* ehdr_vdso = reinterpret_cast<ElfW(Ehdr)*>(getauxval(AT_SYSINFO_EHDR));
166 if (ehdr_vdso == nullptr) {
167 return;
168 }
169
170 soinfo* si = soinfo_alloc(&g_default_namespace, "[vdso]", nullptr, 0, 0);
171
172 si->phdr = reinterpret_cast<ElfW(Phdr)*>(reinterpret_cast<char*>(ehdr_vdso) + ehdr_vdso->e_phoff);
173 si->phnum = ehdr_vdso->e_phnum;
174 si->base = reinterpret_cast<ElfW(Addr)>(ehdr_vdso);
175 si->size = phdr_table_get_load_size(si->phdr, si->phnum);
176 si->load_bias = get_elf_exec_load_bias(ehdr_vdso);
177
178 si->prelink_image();
179 si->link_image(SymbolLookupList(si), si, nullptr, nullptr);
180 // prevents accidental unloads...
181 si->set_dt_flags_1(si->get_dt_flags_1() | DF_1_NODELETE);
182 si->set_linked();
183 si->call_constructors();
184
185 vdso = si;
186 }
187
188 // Initializes an soinfo's link_map_head field using other fields from the
189 // soinfo (phdr, phnum, load_bias). The soinfo's realpath must not change after
190 // this function is called.
init_link_map_head(soinfo & info)191 static void init_link_map_head(soinfo& info) {
192 auto& map = info.link_map_head;
193 map.l_addr = info.load_bias;
194 map.l_name = const_cast<char*>(info.get_realpath());
195 phdr_table_get_dynamic_section(info.phdr, info.phnum, info.load_bias, &map.l_ld, nullptr);
196 }
197
198 extern "C" int __system_properties_init(void);
199
200 struct ExecutableInfo {
201 std::string path;
202 struct stat file_stat;
203 const ElfW(Phdr)* phdr;
204 size_t phdr_count;
205 ElfW(Addr) entry_point;
206 bool should_pad_segments;
207 };
208
get_executable_info(const char * arg_path)209 static ExecutableInfo get_executable_info(const char* arg_path) {
210 ExecutableInfo result = {};
211 char const* exe_path = "/proc/self/exe";
212
213 // Stat "/proc/self/exe" instead of executable_path because
214 // the executable could be unlinked by this point and it should
215 // not cause a crash (see http://b/31084669)
216 if (TEMP_FAILURE_RETRY(stat(exe_path, &result.file_stat) == -1)) {
217 // Fallback to argv[0] for the case where /proc isn't available
218 if (TEMP_FAILURE_RETRY(stat(arg_path, &result.file_stat) == -1)) {
219 async_safe_fatal("unable to stat either \"/proc/self/exe\" or \"%s\": %s",
220 arg_path, strerror(errno));
221 }
222 exe_path = arg_path;
223 }
224
225 // Path might be a symlink; we need the target so that we get the right
226 // linker configuration later.
227 char sym_path[PATH_MAX];
228 result.path = std::string(realpath(exe_path, sym_path) != nullptr ? sym_path : exe_path);
229
230 result.phdr = reinterpret_cast<const ElfW(Phdr)*>(getauxval(AT_PHDR));
231 result.phdr_count = getauxval(AT_PHNUM);
232 result.entry_point = getauxval(AT_ENTRY);
233 return result;
234 }
235
236 #if defined(__LP64__)
237 static char kFallbackLinkerPath[] = "/system/bin/linker64";
238 #else
239 static char kFallbackLinkerPath[] = "/system/bin/linker";
240 #endif
241
242 __printflike(1, 2)
__linker_error(const char * fmt,...)243 static void __linker_error(const char* fmt, ...) {
244 va_list ap;
245
246 va_start(ap, fmt);
247 async_safe_format_fd_va_list(STDERR_FILENO, fmt, ap);
248 va_end(ap);
249
250 va_start(ap, fmt);
251 async_safe_format_log_va_list(ANDROID_LOG_FATAL, "linker", fmt, ap);
252 va_end(ap);
253
254 _exit(EXIT_FAILURE);
255 }
256
__linker_cannot_link(const char * argv0)257 static void __linker_cannot_link(const char* argv0) {
258 __linker_error("CANNOT LINK EXECUTABLE \"%s\": %s\n",
259 argv0,
260 linker_get_error_buffer());
261 }
262
263 // Load an executable. Normally the kernel has already loaded the executable when the linker
264 // starts. The linker can be invoked directly on an executable, though, and then the linker must
265 // load it. This function doesn't load dependencies or resolve relocations.
load_executable(const char * orig_path)266 static ExecutableInfo load_executable(const char* orig_path) {
267 ExecutableInfo result = {};
268
269 if (orig_path[0] != '/') {
270 __linker_error("error: expected absolute path: \"%s\"\n", orig_path);
271 }
272
273 off64_t file_offset;
274 android::base::unique_fd fd(open_executable(orig_path, &file_offset, &result.path));
275 if (fd.get() == -1) {
276 __linker_error("error: unable to open file \"%s\"\n", orig_path);
277 }
278
279 if (TEMP_FAILURE_RETRY(fstat(fd.get(), &result.file_stat)) == -1) {
280 __linker_error("error: unable to stat \"%s\": %s\n", result.path.c_str(), strerror(errno));
281 }
282
283 ElfReader elf_reader;
284 if (!elf_reader.Read(result.path.c_str(), fd.get(), file_offset, result.file_stat.st_size)) {
285 __linker_error("error: %s\n", linker_get_error_buffer());
286 }
287 address_space_params address_space;
288 if (!elf_reader.Load(&address_space)) {
289 __linker_error("error: %s\n", linker_get_error_buffer());
290 }
291
292 result.phdr = elf_reader.loaded_phdr();
293 result.phdr_count = elf_reader.phdr_count();
294 result.entry_point = elf_reader.entry_point();
295 result.should_pad_segments = elf_reader.should_pad_segments();
296 return result;
297 }
298
platform_properties_init()299 static void platform_properties_init() {
300 #if defined(__aarch64__)
301 const unsigned long hwcap2 = getauxval(AT_HWCAP2);
302 g_platform_properties.bti_supported = (hwcap2 & HWCAP2_BTI) != 0;
303 #endif
304 }
305
linker_main(KernelArgumentBlock & args,const char * exe_to_load)306 static ElfW(Addr) linker_main(KernelArgumentBlock& args, const char* exe_to_load) {
307 ProtectedDataGuard guard;
308
309 #if TIMING
310 struct timeval t0, t1;
311 gettimeofday(&t0, 0);
312 #endif
313
314 // Sanitize the environment.
315 __libc_init_AT_SECURE(args.envp);
316
317 // Initialize system properties
318 __system_properties_init(); // may use 'environ'
319
320 // Initialize platform properties.
321 platform_properties_init();
322
323 // Register the debuggerd signal handler.
324 linker_debuggerd_init();
325
326 g_linker_logger.ResetState();
327
328 // Enable debugging logs?
329 const char* LD_DEBUG = getenv("LD_DEBUG");
330 if (LD_DEBUG != nullptr) {
331 g_ld_debug_verbosity = atoi(LD_DEBUG);
332 }
333
334 if (getenv("LD_SHOW_AUXV") != nullptr) ld_show_auxv(args.auxv);
335
336 #if defined(__LP64__)
337 INFO("[ Android dynamic linker (64-bit) ]");
338 #else
339 INFO("[ Android dynamic linker (32-bit) ]");
340 #endif
341
342 // These should have been sanitized by __libc_init_AT_SECURE, but the test
343 // doesn't cost us anything.
344 const char* ldpath_env = nullptr;
345 const char* ldpreload_env = nullptr;
346 if (!getauxval(AT_SECURE)) {
347 ldpath_env = getenv("LD_LIBRARY_PATH");
348 if (ldpath_env != nullptr) {
349 INFO("[ LD_LIBRARY_PATH set to \"%s\" ]", ldpath_env);
350 }
351 ldpreload_env = getenv("LD_PRELOAD");
352 if (ldpreload_env != nullptr) {
353 INFO("[ LD_PRELOAD set to \"%s\" ]", ldpreload_env);
354 }
355 }
356
357 const ExecutableInfo exe_info = exe_to_load ? load_executable(exe_to_load) :
358 get_executable_info(args.argv[0]);
359
360 INFO("[ Linking executable \"%s\" ]", exe_info.path.c_str());
361
362 // Initialize the main exe's soinfo.
363 soinfo* si = soinfo_alloc(&g_default_namespace,
364 exe_info.path.c_str(), &exe_info.file_stat,
365 0, RTLD_GLOBAL);
366 somain = si;
367 si->phdr = exe_info.phdr;
368 si->phnum = exe_info.phdr_count;
369 si->set_should_pad_segments(exe_info.should_pad_segments);
370 get_elf_base_from_phdr(si->phdr, si->phnum, &si->base, &si->load_bias);
371 si->size = phdr_table_get_load_size(si->phdr, si->phnum);
372 si->dynamic = nullptr;
373 si->set_main_executable();
374 init_link_map_head(*si);
375
376 set_bss_vma_name(si);
377
378 // Use the executable's PT_INTERP string as the solinker filename in the
379 // dynamic linker's module list. gdb reads both PT_INTERP and the module list,
380 // and if the paths for the linker are different, gdb will report that the
381 // PT_INTERP linker path was unloaded once the module list is initialized.
382 // There are three situations to handle:
383 // - the APEX linker (/system/bin/linker[64] -> /apex/.../linker[64])
384 // - the ASAN linker (/system/bin/linker_asan[64] -> /apex/.../linker[64])
385 // - the bootstrap linker (/system/bin/bootstrap/linker[64])
386 const char *interp = phdr_table_get_interpreter_name(somain->phdr, somain->phnum,
387 somain->load_bias);
388 if (interp == nullptr) {
389 // This case can happen if the linker attempts to execute itself
390 // (e.g. "linker64 /system/bin/linker64").
391 interp = kFallbackLinkerPath;
392 }
393 solinker->set_realpath(interp);
394 init_link_map_head(*solinker);
395
396 #if defined(__aarch64__)
397 if (exe_to_load == nullptr) {
398 // Kernel does not add PROT_BTI to executable pages of the loaded ELF.
399 // Apply appropriate protections here if it is needed.
400 auto note_gnu_property = GnuPropertySection(somain);
401 if (note_gnu_property.IsBTICompatible() &&
402 (phdr_table_protect_segments(somain->phdr, somain->phnum, somain->load_bias,
403 somain->should_pad_segments(), ¬e_gnu_property) < 0)) {
404 __linker_error("error: can't protect segments for \"%s\": %s", exe_info.path.c_str(),
405 strerror(errno));
406 }
407 }
408 #endif
409
410 // Register the main executable and the linker upfront to have
411 // gdb aware of them before loading the rest of the dependency
412 // tree.
413 //
414 // gdb expects the linker to be in the debug shared object list.
415 // Without this, gdb has trouble locating the linker's ".text"
416 // and ".plt" sections. Gdb could also potentially use this to
417 // relocate the offset of our exported 'rtld_db_dlactivity' symbol.
418 //
419 insert_link_map_into_debug_map(&si->link_map_head);
420 insert_link_map_into_debug_map(&solinker->link_map_head);
421
422 add_vdso();
423
424 ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(si->base);
425
426 // For security reasons we dropped non-PIE support in API level 21,
427 // and the NDK no longer supports earlier API levels.
428 if (elf_hdr->e_type != ET_DYN) {
429 __linker_error("error: %s: Android only supports position-independent "
430 "executables (-fPIE)\n", exe_info.path.c_str());
431 }
432
433 // Use LD_LIBRARY_PATH and LD_PRELOAD (but only if we aren't setuid/setgid).
434 parse_LD_LIBRARY_PATH(ldpath_env);
435 parse_LD_PRELOAD(ldpreload_env);
436
437 std::vector<android_namespace_t*> namespaces = init_default_namespaces(exe_info.path.c_str());
438
439 if (!si->prelink_image()) __linker_cannot_link(g_argv[0]);
440
441 // add somain to global group
442 si->set_dt_flags_1(si->get_dt_flags_1() | DF_1_GLOBAL);
443 // ... and add it to all other linked namespaces
444 for (auto linked_ns : namespaces) {
445 if (linked_ns != &g_default_namespace) {
446 linked_ns->add_soinfo(somain);
447 somain->add_secondary_namespace(linked_ns);
448 }
449 }
450
451 linker_setup_exe_static_tls(g_argv[0]);
452
453 // Load ld_preloads and dependencies.
454 std::vector<const char*> needed_library_name_list;
455 size_t ld_preloads_count = 0;
456
457 for (const auto& ld_preload_name : g_ld_preload_names) {
458 needed_library_name_list.push_back(ld_preload_name.c_str());
459 ++ld_preloads_count;
460 }
461
462 for_each_dt_needed(si, [&](const char* name) {
463 needed_library_name_list.push_back(name);
464 });
465
466 const char** needed_library_names = &needed_library_name_list[0];
467 size_t needed_libraries_count = needed_library_name_list.size();
468
469 if (needed_libraries_count > 0 &&
470 !find_libraries(&g_default_namespace,
471 si,
472 needed_library_names,
473 needed_libraries_count,
474 nullptr,
475 &g_ld_preloads,
476 ld_preloads_count,
477 RTLD_GLOBAL,
478 nullptr,
479 true /* add_as_children */,
480 &namespaces)) {
481 __linker_cannot_link(g_argv[0]);
482 } else if (needed_libraries_count == 0) {
483 if (!si->link_image(SymbolLookupList(si), si, nullptr, nullptr)) {
484 __linker_cannot_link(g_argv[0]);
485 }
486 si->increment_ref_count();
487 }
488
489 // Exit early for ldd. We don't want to run the code that was loaded, so skip
490 // the constructor calls. Skip CFI setup because it would call __cfi_init in
491 // libdl.so.
492 if (g_is_ldd) _exit(EXIT_SUCCESS);
493
494 #if defined(__aarch64__)
495 // This has to happen after the find_libraries, which will have collected any possible
496 // libraries that request memtag_stack in the dynamic section.
497 __libc_init_mte(somain->memtag_dynamic_entries(), somain->phdr, somain->phnum, somain->load_bias,
498 args.argv);
499 #endif
500
501 linker_finalize_static_tls();
502 __libc_init_main_thread_final();
503
504 if (!get_cfi_shadow()->InitialLinkDone(solist)) __linker_cannot_link(g_argv[0]);
505
506 si->call_pre_init_constructors();
507 si->call_constructors();
508
509 #if TIMING
510 gettimeofday(&t1, nullptr);
511 PRINT("LINKER TIME: %s: %d microseconds", g_argv[0],
512 static_cast<int>(((static_cast<long long>(t1.tv_sec) * 1000000LL) +
513 static_cast<long long>(t1.tv_usec)) -
514 ((static_cast<long long>(t0.tv_sec) * 1000000LL) +
515 static_cast<long long>(t0.tv_usec))));
516 #endif
517 #if STATS
518 print_linker_stats();
519 #endif
520 #if TIMING || STATS
521 fflush(stdout);
522 #endif
523
524 // We are about to hand control over to the executable loaded. We don't want
525 // to leave dirty pages behind unnecessarily.
526 purge_unused_memory();
527
528 ElfW(Addr) entry = exe_info.entry_point;
529 TRACE("[ Ready to execute \"%s\" @ %p ]", si->get_realpath(), reinterpret_cast<void*>(entry));
530 return entry;
531 }
532
533 /* Compute the load-bias of an existing executable. This shall only
534 * be used to compute the load bias of an executable or shared library
535 * that was loaded by the kernel itself.
536 *
537 * Input:
538 * elf -> address of ELF header, assumed to be at the start of the file.
539 * Return:
540 * load bias, i.e. add the value of any p_vaddr in the file to get
541 * the corresponding address in memory.
542 */
get_elf_exec_load_bias(const ElfW (Ehdr)* elf)543 static ElfW(Addr) get_elf_exec_load_bias(const ElfW(Ehdr)* elf) {
544 ElfW(Addr) offset = elf->e_phoff;
545 const ElfW(Phdr)* phdr_table =
546 reinterpret_cast<const ElfW(Phdr)*>(reinterpret_cast<uintptr_t>(elf) + offset);
547 const ElfW(Phdr)* phdr_end = phdr_table + elf->e_phnum;
548
549 for (const ElfW(Phdr)* phdr = phdr_table; phdr < phdr_end; phdr++) {
550 if (phdr->p_type == PT_LOAD) {
551 return reinterpret_cast<ElfW(Addr)>(elf) + phdr->p_offset - phdr->p_vaddr;
552 }
553 }
554 return 0;
555 }
556
557 /* Find the load bias and base address of an executable or shared object loaded
558 * by the kernel. The ELF file's PHDR table must have a PT_PHDR entry.
559 *
560 * A VDSO doesn't have a PT_PHDR entry in its PHDR table.
561 */
get_elf_base_from_phdr(const ElfW (Phdr)* phdr_table,size_t phdr_count,ElfW (Addr)* base,ElfW (Addr)* load_bias)562 static void get_elf_base_from_phdr(const ElfW(Phdr)* phdr_table, size_t phdr_count,
563 ElfW(Addr)* base, ElfW(Addr)* load_bias) {
564 for (size_t i = 0; i < phdr_count; ++i) {
565 if (phdr_table[i].p_type == PT_PHDR) {
566 *load_bias = reinterpret_cast<ElfW(Addr)>(phdr_table) - phdr_table[i].p_vaddr;
567 *base = reinterpret_cast<ElfW(Addr)>(phdr_table) - phdr_table[i].p_offset;
568 return;
569 }
570 }
571 async_safe_fatal("Could not find a PHDR: broken executable?");
572 }
573
574 /*
575 * Set anonymous VMA name for .bss section. For DSOs loaded by the linker, this
576 * is done by ElfReader. This function is here for DSOs loaded by the kernel,
577 * namely the linker itself and the main executable.
578 */
set_bss_vma_name(soinfo * si)579 static void set_bss_vma_name(soinfo* si) {
580 for (size_t i = 0; i < si->phnum; ++i) {
581 auto phdr = &si->phdr[i];
582
583 if (phdr->p_type != PT_LOAD) {
584 continue;
585 }
586
587 ElfW(Addr) seg_start = phdr->p_vaddr + si->load_bias;
588 ElfW(Addr) seg_page_end = page_end(seg_start + phdr->p_memsz);
589 ElfW(Addr) seg_file_end = page_end(seg_start + phdr->p_filesz);
590
591 if (seg_page_end > seg_file_end) {
592 prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME,
593 reinterpret_cast<void*>(seg_file_end), seg_page_end - seg_file_end,
594 ".bss");
595 }
596 }
597 }
598
599 #if defined(USE_RELA)
600 using RelType = ElfW(Rela);
601 const unsigned kRelTag = DT_RELA;
602 const unsigned kRelSzTag = DT_RELASZ;
603 #else
604 using RelType = ElfW(Rel);
605 const unsigned kRelTag = DT_REL;
606 const unsigned kRelSzTag = DT_RELSZ;
607 #endif
608
609 extern __LIBC_HIDDEN__ ElfW(Ehdr) __ehdr_start;
610
call_ifunc_resolvers_for_section(RelType * begin,RelType * end)611 static void call_ifunc_resolvers_for_section(RelType* begin, RelType* end) {
612 auto ehdr = reinterpret_cast<ElfW(Addr)>(&__ehdr_start);
613 for (RelType *r = begin; r != end; ++r) {
614 if (ELFW(R_TYPE)(r->r_info) != R_GENERIC_IRELATIVE) {
615 continue;
616 }
617 ElfW(Addr)* offset = reinterpret_cast<ElfW(Addr)*>(ehdr + r->r_offset);
618 #if defined(USE_RELA)
619 ElfW(Addr) resolver = ehdr + r->r_addend;
620 #else
621 ElfW(Addr) resolver = ehdr + *offset;
622 #endif
623 *offset = __bionic_call_ifunc_resolver(resolver);
624 }
625 }
626
relocate_linker()627 static void relocate_linker() {
628 // The linker should only have relative relocations (in RELR) and IRELATIVE
629 // relocations. Find the IRELATIVE relocations using the DT_JMPREL and
630 // DT_PLTRELSZ, or DT_RELA/DT_RELASZ (DT_REL/DT_RELSZ on ILP32).
631 auto ehdr = reinterpret_cast<ElfW(Addr)>(&__ehdr_start);
632 auto* phdr = reinterpret_cast<ElfW(Phdr)*>(ehdr + __ehdr_start.e_phoff);
633 for (size_t i = 0; i != __ehdr_start.e_phnum; ++i) {
634 if (phdr[i].p_type != PT_DYNAMIC) {
635 continue;
636 }
637 auto *dyn = reinterpret_cast<ElfW(Dyn)*>(ehdr + phdr[i].p_vaddr);
638 ElfW(Addr) relr = 0, relrsz = 0, pltrel = 0, pltrelsz = 0, rel = 0, relsz = 0;
639 for (size_t j = 0, size = phdr[i].p_filesz / sizeof(ElfW(Dyn)); j != size; ++j) {
640 const auto tag = dyn[j].d_tag;
641 const auto val = dyn[j].d_un.d_ptr;
642 // We don't currently handle IRELATIVE relocations in DT_ANDROID_REL[A].
643 // We disabled DT_ANDROID_REL[A] at build time; verify that it was actually disabled.
644 CHECK(tag != DT_ANDROID_REL && tag != DT_ANDROID_RELA);
645 if (tag == DT_RELR || tag == DT_ANDROID_RELR) {
646 relr = val;
647 } else if (tag == DT_RELRSZ || tag == DT_ANDROID_RELRSZ) {
648 relrsz = val;
649 } else if (tag == DT_JMPREL) {
650 pltrel = val;
651 } else if (tag == DT_PLTRELSZ) {
652 pltrelsz = val;
653 } else if (tag == kRelTag) {
654 rel = val;
655 } else if (tag == kRelSzTag) {
656 relsz = val;
657 }
658 }
659 // Apply RELR relocations first so that the GOT is initialized for ifunc
660 // resolvers.
661 if (relr && relrsz) {
662 relocate_relr(reinterpret_cast<ElfW(Relr*)>(ehdr + relr),
663 reinterpret_cast<ElfW(Relr*)>(ehdr + relr + relrsz), ehdr);
664 }
665 if (pltrel && pltrelsz) {
666 call_ifunc_resolvers_for_section(reinterpret_cast<RelType*>(ehdr + pltrel),
667 reinterpret_cast<RelType*>(ehdr + pltrel + pltrelsz));
668 }
669 if (rel && relsz) {
670 call_ifunc_resolvers_for_section(reinterpret_cast<RelType*>(ehdr + rel),
671 reinterpret_cast<RelType*>(ehdr + rel + relsz));
672 }
673 }
674 }
675
676 // Usable before ifunc resolvers have been called. This function is compiled with -ffreestanding.
linker_memclr(void * dst,size_t cnt)677 static void linker_memclr(void* dst, size_t cnt) {
678 for (size_t i = 0; i < cnt; ++i) {
679 reinterpret_cast<char*>(dst)[i] = '\0';
680 }
681 }
682
683 // Detect an attempt to run the linker on itself. e.g.:
684 // /system/bin/linker64 /system/bin/linker64
685 // Use priority-1 to run this constructor before other constructors.
detect_self_exec()686 __attribute__((constructor(1))) static void detect_self_exec() {
687 // Normally, the linker initializes the auxv global before calling its
688 // constructors. If the linker loads itself, though, the first loader calls
689 // the second loader's constructors before calling __linker_init.
690 if (__libc_shared_globals()->auxv != nullptr) {
691 return;
692 }
693 #if defined(__i386__)
694 // We don't have access to the auxv struct from here, so use the int 0x80
695 // fallback.
696 __libc_sysinfo = reinterpret_cast<void*>(__libc_int0x80);
697 #endif
698 __linker_error("error: linker cannot load itself\n");
699 }
700
701 static ElfW(Addr) __attribute__((noinline))
702 __linker_init_post_relocation(KernelArgumentBlock& args, soinfo& linker_so);
703
704 /*
705 * This is the entry point for the linker, called from begin.S. This
706 * method is responsible for fixing the linker's own relocations, and
707 * then calling __linker_init_post_relocation().
708 *
709 * Because this method is called before the linker has fixed it's own
710 * relocations, any attempt to reference an extern variable, extern
711 * function, or other GOT reference will generate a segfault.
712 */
__linker_init(void * raw_args)713 extern "C" ElfW(Addr) __linker_init(void* raw_args) {
714 // Unlock the loader mutex immediately before transferring to the executable's
715 // entry point. This must happen after destructors are called in this function
716 // (e.g. ~soinfo), so declare this variable very early.
717 struct DlMutexUnlocker {
718 ~DlMutexUnlocker() { pthread_mutex_unlock(&g_dl_mutex); }
719 } unlocker;
720
721 // Initialize TLS early so system calls and errno work.
722 KernelArgumentBlock args(raw_args);
723 bionic_tcb temp_tcb __attribute__((uninitialized));
724 linker_memclr(&temp_tcb, sizeof(temp_tcb));
725 __libc_init_main_thread_early(args, &temp_tcb);
726
727 // When the linker is run by itself (rather than as an interpreter for
728 // another program), AT_BASE is 0.
729 ElfW(Addr) linker_addr = getauxval(AT_BASE);
730 if (linker_addr == 0) {
731 // The AT_PHDR and AT_PHNUM aux values describe this linker instance, so use
732 // the phdr to find the linker's base address.
733 ElfW(Addr) load_bias;
734 get_elf_base_from_phdr(
735 reinterpret_cast<ElfW(Phdr)*>(getauxval(AT_PHDR)), getauxval(AT_PHNUM),
736 &linker_addr, &load_bias);
737 }
738
739 ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(linker_addr);
740 ElfW(Phdr)* phdr = reinterpret_cast<ElfW(Phdr)*>(linker_addr + elf_hdr->e_phoff);
741
742 // Relocate the linker. This step will initialize the GOT, which is needed for
743 // accessing non-hidden global variables. (On some targets, the stack
744 // protector uses GOT accesses rather than TLS.) Relocating the linker will
745 // also call the linker's ifunc resolvers so that string.h functions can be
746 // used.
747 relocate_linker();
748
749 soinfo tmp_linker_so(nullptr, nullptr, nullptr, 0, 0);
750
751 tmp_linker_so.base = linker_addr;
752 tmp_linker_so.size = phdr_table_get_load_size(phdr, elf_hdr->e_phnum);
753 tmp_linker_so.load_bias = get_elf_exec_load_bias(elf_hdr);
754 tmp_linker_so.dynamic = nullptr;
755 tmp_linker_so.phdr = phdr;
756 tmp_linker_so.phnum = elf_hdr->e_phnum;
757 tmp_linker_so.set_linker_flag();
758
759 if (!tmp_linker_so.prelink_image()) __linker_cannot_link(args.argv[0]);
760 if (!tmp_linker_so.link_image(SymbolLookupList(&tmp_linker_so), &tmp_linker_so, nullptr, nullptr)) __linker_cannot_link(args.argv[0]);
761
762 return __linker_init_post_relocation(args, tmp_linker_so);
763 }
764
765 /*
766 * This code is called after the linker has linked itself and fixed its own
767 * GOT. It is safe to make references to externs and other non-local data at
768 * this point. The compiler sometimes moves GOT references earlier in a
769 * function, so avoid inlining this function (http://b/80503879).
770 */
771 static ElfW(Addr) __attribute__((noinline))
__linker_init_post_relocation(KernelArgumentBlock & args,soinfo & tmp_linker_so)772 __linker_init_post_relocation(KernelArgumentBlock& args, soinfo& tmp_linker_so) {
773 // Finish initializing the main thread.
774 __libc_init_main_thread_late();
775
776 // We didn't protect the linker's RELRO pages in link_image because we
777 // couldn't make system calls on x86 at that point, but we can now...
778 if (!tmp_linker_so.protect_relro()) __linker_cannot_link(args.argv[0]);
779
780 // And we can set VMA name for the bss section now
781 set_bss_vma_name(&tmp_linker_so);
782
783 // Initialize the linker's static libc's globals
784 __libc_init_globals();
785
786 // A constructor could spawn a thread that calls into the loader, so as soon
787 // as we've called a constructor, we need to hold the lock until transferring
788 // to the entry point.
789 pthread_mutex_lock(&g_dl_mutex);
790
791 // Initialize the linker's own global variables
792 tmp_linker_so.call_constructors();
793
794 // Setting the linker soinfo's soname can allocate heap memory, so delay it until here.
795 for (const ElfW(Dyn)* d = tmp_linker_so.dynamic; d->d_tag != DT_NULL; ++d) {
796 if (d->d_tag == DT_SONAME) {
797 tmp_linker_so.set_soname(tmp_linker_so.get_string(d->d_un.d_val));
798 }
799 }
800
801 // When the linker is run directly rather than acting as PT_INTERP, parse
802 // arguments and determine the executable to load. When it's instead acting
803 // as PT_INTERP, AT_ENTRY will refer to the loaded executable rather than the
804 // linker's _start.
805 const char* exe_to_load = nullptr;
806 if (getauxval(AT_ENTRY) == reinterpret_cast<uintptr_t>(&_start)) {
807 if (args.argc == 3 && !strcmp(args.argv[1], "--list")) {
808 // We're being asked to behave like ldd(1).
809 g_is_ldd = true;
810 exe_to_load = args.argv[2];
811 } else if (args.argc <= 1 || !strcmp(args.argv[1], "--help")) {
812 async_safe_format_fd(STDOUT_FILENO,
813 "Usage: %s [--list] PROGRAM [ARGS-FOR-PROGRAM...]\n"
814 " %s [--list] path.zip!/PROGRAM [ARGS-FOR-PROGRAM...]\n"
815 "\n"
816 "A helper program for linking dynamic executables. Typically, the kernel loads\n"
817 "this program because it's the PT_INTERP of a dynamic executable.\n"
818 "\n"
819 "This program can also be run directly to load and run a dynamic executable. The\n"
820 "executable can be inside a zip file if it's stored uncompressed and at a\n"
821 "page-aligned offset.\n"
822 "\n"
823 "The --list option gives behavior equivalent to ldd(1) on other systems.\n",
824 args.argv[0], args.argv[0]);
825 _exit(EXIT_SUCCESS);
826 } else {
827 exe_to_load = args.argv[1];
828 __libc_shared_globals()->initial_linker_arg_count = 1;
829 }
830 }
831
832 // store argc/argv/envp to use them for calling constructors
833 g_argc = args.argc - __libc_shared_globals()->initial_linker_arg_count;
834 g_argv = args.argv + __libc_shared_globals()->initial_linker_arg_count;
835 g_envp = args.envp;
836 __libc_shared_globals()->init_progname = g_argv[0];
837
838 // Initialize static variables. Note that in order to
839 // get correct libdl_info we need to call constructors
840 // before get_libdl_info().
841 sonext = solist = solinker = get_libdl_info(tmp_linker_so);
842 g_default_namespace.add_soinfo(solinker);
843
844 ElfW(Addr) start_address = linker_main(args, exe_to_load);
845
846 INFO("[ Jumping to _start (%p)... ]", reinterpret_cast<void*>(start_address));
847
848 // Return the address that the calling assembly stub should jump to.
849 return start_address;
850 }
851