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 <sys/auxv.h>
33 
34 #include "linker.h"
35 #include "linker_cfi.h"
36 #include "linker_debug.h"
37 #include "linker_debuggerd.h"
38 #include "linker_gdb_support.h"
39 #include "linker_globals.h"
40 #include "linker_phdr.h"
41 #include "linker_relocate.h"
42 #include "linker_tls.h"
43 #include "linker_utils.h"
44 
45 #include "private/bionic_auxv.h"
46 #include "private/bionic_call_ifunc_resolver.h"
47 #include "private/bionic_globals.h"
48 #include "private/bionic_tls.h"
49 #include "private/KernelArgumentBlock.h"
50 
51 #include "android-base/unique_fd.h"
52 #include "android-base/strings.h"
53 #include "android-base/stringprintf.h"
54 
55 #include <async_safe/log.h>
56 #include <bionic/libc_init_common.h>
57 #include <bionic/pthread_internal.h>
58 
59 #include <vector>
60 
61 __LIBC_HIDDEN__ extern "C" void _start();
62 
63 static ElfW(Addr) get_elf_exec_load_bias(const ElfW(Ehdr)* elf);
64 
65 static void get_elf_base_from_phdr(const ElfW(Phdr)* phdr_table, size_t phdr_count,
66                                    ElfW(Addr)* base, ElfW(Addr)* load_bias);
67 
68 static void set_bss_vma_name(soinfo* si);
69 
70 void __libc_init_mte(const void* phdr_start, size_t phdr_count, uintptr_t load_bias);
71 
72 // These should be preserved static to avoid emitting
73 // RELATIVE relocations for the part of the code running
74 // before linker links itself.
75 
76 // TODO (dimtiry): remove somain, rename solist to solist_head
77 static soinfo* solist;
78 static soinfo* sonext;
79 static soinfo* somain; // main process, always the one after libdl_info
80 static soinfo* solinker;
81 static soinfo* vdso; // vdso if present
82 
solist_add_soinfo(soinfo * si)83 void solist_add_soinfo(soinfo* si) {
84   sonext->next = si;
85   sonext = si;
86 }
87 
solist_remove_soinfo(soinfo * si)88 bool solist_remove_soinfo(soinfo* si) {
89   soinfo *prev = nullptr, *trav;
90   for (trav = solist; trav != nullptr; trav = trav->next) {
91     if (trav == si) {
92       break;
93     }
94     prev = trav;
95   }
96 
97   if (trav == nullptr) {
98     // si was not in solist
99     PRINT("name \"%s\"@%p is not in solist!", si->get_realpath(), si);
100     return false;
101   }
102 
103   // prev will never be null, because the first entry in solist is
104   // always the static libdl_info.
105   CHECK(prev != nullptr);
106   prev->next = si->next;
107   if (si == sonext) {
108     sonext = prev;
109   }
110 
111   return true;
112 }
113 
solist_get_head()114 soinfo* solist_get_head() {
115   return solist;
116 }
117 
solist_get_somain()118 soinfo* solist_get_somain() {
119   return somain;
120 }
121 
solist_get_vdso()122 soinfo* solist_get_vdso() {
123   return vdso;
124 }
125 
126 bool g_is_ldd;
127 int g_ld_debug_verbosity;
128 
129 static std::vector<std::string> g_ld_preload_names;
130 
131 static std::vector<soinfo*> g_ld_preloads;
132 
parse_path(const char * path,const char * delimiters,std::vector<std::string> * resolved_paths)133 static void parse_path(const char* path, const char* delimiters,
134                        std::vector<std::string>* resolved_paths) {
135   std::vector<std::string> paths;
136   split_path(path, delimiters, &paths);
137   resolve_paths(paths, resolved_paths);
138 }
139 
parse_LD_LIBRARY_PATH(const char * path)140 static void parse_LD_LIBRARY_PATH(const char* path) {
141   std::vector<std::string> ld_libary_paths;
142   parse_path(path, ":", &ld_libary_paths);
143   g_default_namespace.set_ld_library_paths(std::move(ld_libary_paths));
144 }
145 
parse_LD_PRELOAD(const char * path)146 static void parse_LD_PRELOAD(const char* path) {
147   g_ld_preload_names.clear();
148   if (path != nullptr) {
149     // We have historically supported ':' as well as ' ' in LD_PRELOAD.
150     g_ld_preload_names = android::base::Split(path, " :");
151     g_ld_preload_names.erase(std::remove_if(g_ld_preload_names.begin(), g_ld_preload_names.end(),
152                                             [](const std::string& s) { return s.empty(); }),
153                              g_ld_preload_names.end());
154   }
155 }
156 
157 // An empty list of soinfos
158 static soinfo_list_t g_empty_list;
159 
add_vdso()160 static void add_vdso() {
161   ElfW(Ehdr)* ehdr_vdso = reinterpret_cast<ElfW(Ehdr)*>(getauxval(AT_SYSINFO_EHDR));
162   if (ehdr_vdso == nullptr) {
163     return;
164   }
165 
166   soinfo* si = soinfo_alloc(&g_default_namespace, "[vdso]", nullptr, 0, 0);
167 
168   si->phdr = reinterpret_cast<ElfW(Phdr)*>(reinterpret_cast<char*>(ehdr_vdso) + ehdr_vdso->e_phoff);
169   si->phnum = ehdr_vdso->e_phnum;
170   si->base = reinterpret_cast<ElfW(Addr)>(ehdr_vdso);
171   si->size = phdr_table_get_load_size(si->phdr, si->phnum);
172   si->load_bias = get_elf_exec_load_bias(ehdr_vdso);
173 
174   si->prelink_image();
175   si->link_image(SymbolLookupList(si), si, nullptr, nullptr);
176   // prevents accidental unloads...
177   si->set_dt_flags_1(si->get_dt_flags_1() | DF_1_NODELETE);
178   si->set_linked();
179   si->call_constructors();
180 
181   vdso = si;
182 }
183 
184 // Initializes an soinfo's link_map_head field using other fields from the
185 // soinfo (phdr, phnum, load_bias). The soinfo's realpath must not change after
186 // this function is called.
init_link_map_head(soinfo & info)187 static void init_link_map_head(soinfo& info) {
188   auto& map = info.link_map_head;
189   map.l_addr = info.load_bias;
190   map.l_name = const_cast<char*>(info.get_realpath());
191   phdr_table_get_dynamic_section(info.phdr, info.phnum, info.load_bias, &map.l_ld, nullptr);
192 }
193 
194 extern "C" int __system_properties_init(void);
195 
196 struct ExecutableInfo {
197   std::string path;
198   struct stat file_stat;
199   const ElfW(Phdr)* phdr;
200   size_t phdr_count;
201   ElfW(Addr) entry_point;
202 };
203 
get_executable_info()204 static ExecutableInfo get_executable_info() {
205   ExecutableInfo result = {};
206 
207   if (is_first_stage_init()) {
208     // /proc fs is not mounted when first stage init starts. Therefore we can't
209     // use /proc/self/exe for init.
210     stat("/init", &result.file_stat);
211 
212     // /init may be a symlink, so try to read it as such.
213     char path[PATH_MAX];
214     ssize_t path_len = readlink("/init", path, sizeof(path));
215     if (path_len == -1 || path_len >= static_cast<ssize_t>(sizeof(path))) {
216       result.path = "/init";
217     } else {
218       result.path = std::string(path, path_len);
219     }
220   } else {
221     // Stat "/proc/self/exe" instead of executable_path because
222     // the executable could be unlinked by this point and it should
223     // not cause a crash (see http://b/31084669)
224     if (TEMP_FAILURE_RETRY(stat("/proc/self/exe", &result.file_stat)) != 0) {
225       async_safe_fatal("unable to stat \"/proc/self/exe\": %s", strerror(errno));
226     }
227     char path[PATH_MAX];
228     ssize_t path_len = readlink("/proc/self/exe", path, sizeof(path));
229     if (path_len == -1 || path_len >= static_cast<ssize_t>(sizeof(path))) {
230       async_safe_fatal("readlink('/proc/self/exe') failed: %s", strerror(errno));
231     }
232     result.path = std::string(path, path_len);
233   }
234 
235   result.phdr = reinterpret_cast<const ElfW(Phdr)*>(getauxval(AT_PHDR));
236   result.phdr_count = getauxval(AT_PHNUM);
237   result.entry_point = getauxval(AT_ENTRY);
238   return result;
239 }
240 
241 #if defined(__LP64__)
242 static char kFallbackLinkerPath[] = "/system/bin/linker64";
243 #else
244 static char kFallbackLinkerPath[] = "/system/bin/linker";
245 #endif
246 
247 __printflike(1, 2)
__linker_error(const char * fmt,...)248 static void __linker_error(const char* fmt, ...) {
249   va_list ap;
250 
251   va_start(ap, fmt);
252   async_safe_format_fd_va_list(STDERR_FILENO, fmt, ap);
253   va_end(ap);
254 
255   va_start(ap, fmt);
256   async_safe_format_log_va_list(ANDROID_LOG_FATAL, "linker", fmt, ap);
257   va_end(ap);
258 
259   _exit(EXIT_FAILURE);
260 }
261 
__linker_cannot_link(const char * argv0)262 static void __linker_cannot_link(const char* argv0) {
263   __linker_error("CANNOT LINK EXECUTABLE \"%s\": %s\n",
264                  argv0,
265                  linker_get_error_buffer());
266 }
267 
268 // Load an executable. Normally the kernel has already loaded the executable when the linker
269 // starts. The linker can be invoked directly on an executable, though, and then the linker must
270 // load it. This function doesn't load dependencies or resolve relocations.
load_executable(const char * orig_path)271 static ExecutableInfo load_executable(const char* orig_path) {
272   ExecutableInfo result = {};
273 
274   if (orig_path[0] != '/') {
275     __linker_error("error: expected absolute path: \"%s\"\n", orig_path);
276   }
277 
278   off64_t file_offset;
279   android::base::unique_fd fd(open_executable(orig_path, &file_offset, &result.path));
280   if (fd.get() == -1) {
281     __linker_error("error: unable to open file \"%s\"\n", orig_path);
282   }
283 
284   if (TEMP_FAILURE_RETRY(fstat(fd.get(), &result.file_stat)) == -1) {
285     __linker_error("error: unable to stat \"%s\": %s\n", result.path.c_str(), strerror(errno));
286   }
287 
288   ElfReader elf_reader;
289   if (!elf_reader.Read(result.path.c_str(), fd.get(), file_offset, result.file_stat.st_size)) {
290     __linker_error("error: %s\n", linker_get_error_buffer());
291   }
292   address_space_params address_space;
293   if (!elf_reader.Load(&address_space)) {
294     __linker_error("error: %s\n", linker_get_error_buffer());
295   }
296 
297   result.phdr = elf_reader.loaded_phdr();
298   result.phdr_count = elf_reader.phdr_count();
299   result.entry_point = elf_reader.entry_point();
300   return result;
301 }
302 
platform_properties_init()303 static void platform_properties_init() {
304 #if defined(__aarch64__)
305   const unsigned long hwcap2 = getauxval(AT_HWCAP2);
306   g_platform_properties.bti_supported = (hwcap2 & HWCAP2_BTI) != 0;
307 #endif
308 }
309 
linker_main(KernelArgumentBlock & args,const char * exe_to_load)310 static ElfW(Addr) linker_main(KernelArgumentBlock& args, const char* exe_to_load) {
311   ProtectedDataGuard guard;
312 
313 #if TIMING
314   struct timeval t0, t1;
315   gettimeofday(&t0, 0);
316 #endif
317 
318   // Sanitize the environment.
319   __libc_init_AT_SECURE(args.envp);
320 
321   // Initialize system properties
322   __system_properties_init(); // may use 'environ'
323 
324   // Initialize platform properties.
325   platform_properties_init();
326 
327   // Register the debuggerd signal handler.
328   linker_debuggerd_init();
329 
330   g_linker_logger.ResetState();
331 
332   // Get a few environment variables.
333   const char* LD_DEBUG = getenv("LD_DEBUG");
334   if (LD_DEBUG != nullptr) {
335     g_ld_debug_verbosity = atoi(LD_DEBUG);
336   }
337 
338 #if defined(__LP64__)
339   INFO("[ Android dynamic linker (64-bit) ]");
340 #else
341   INFO("[ Android dynamic linker (32-bit) ]");
342 #endif
343 
344   // These should have been sanitized by __libc_init_AT_SECURE, but the test
345   // doesn't cost us anything.
346   const char* ldpath_env = nullptr;
347   const char* ldpreload_env = nullptr;
348   if (!getauxval(AT_SECURE)) {
349     ldpath_env = getenv("LD_LIBRARY_PATH");
350     if (ldpath_env != nullptr) {
351       INFO("[ LD_LIBRARY_PATH set to \"%s\" ]", ldpath_env);
352     }
353     ldpreload_env = getenv("LD_PRELOAD");
354     if (ldpreload_env != nullptr) {
355       INFO("[ LD_PRELOAD set to \"%s\" ]", ldpreload_env);
356     }
357   }
358 
359   const ExecutableInfo exe_info = exe_to_load ? load_executable(exe_to_load) :
360                                                 get_executable_info();
361 
362   INFO("[ Linking executable \"%s\" ]", exe_info.path.c_str());
363 
364   // Initialize the main exe's soinfo.
365   soinfo* si = soinfo_alloc(&g_default_namespace,
366                             exe_info.path.c_str(), &exe_info.file_stat,
367                             0, RTLD_GLOBAL);
368   somain = si;
369   si->phdr = exe_info.phdr;
370   si->phnum = exe_info.phdr_count;
371   get_elf_base_from_phdr(si->phdr, si->phnum, &si->base, &si->load_bias);
372   si->size = phdr_table_get_load_size(si->phdr, si->phnum);
373   si->dynamic = nullptr;
374   si->set_main_executable();
375   init_link_map_head(*si);
376 
377   set_bss_vma_name(si);
378 
379   // Use the executable's PT_INTERP string as the solinker filename in the
380   // dynamic linker's module list. gdb reads both PT_INTERP and the module list,
381   // and if the paths for the linker are different, gdb will report that the
382   // PT_INTERP linker path was unloaded once the module list is initialized.
383   // There are three situations to handle:
384   //  - the APEX linker (/system/bin/linker[64] -> /apex/.../linker[64])
385   //  - the ASAN linker (/system/bin/linker_asan[64] -> /apex/.../linker[64])
386   //  - the bootstrap linker (/system/bin/bootstrap/linker[64])
387   const char *interp = phdr_table_get_interpreter_name(somain->phdr, somain->phnum,
388                                                        somain->load_bias);
389   if (interp == nullptr) {
390     // This case can happen if the linker attempts to execute itself
391     // (e.g. "linker64 /system/bin/linker64").
392     interp = kFallbackLinkerPath;
393   }
394   solinker->set_realpath(interp);
395   init_link_map_head(*solinker);
396 
397 #if defined(__aarch64__)
398   if (exe_to_load == nullptr) {
399     // Kernel does not add PROT_BTI to executable pages of the loaded ELF.
400     // Apply appropriate protections here if it is needed.
401     auto note_gnu_property = GnuPropertySection(somain);
402     if (note_gnu_property.IsBTICompatible() &&
403         (phdr_table_protect_segments(somain->phdr, somain->phnum, somain->load_bias,
404                                      &note_gnu_property) < 0)) {
405       __linker_error("error: can't protect segments for \"%s\": %s", exe_info.path.c_str(),
406                      strerror(errno));
407     }
408   }
409 
410   __libc_init_mte(somain->phdr, somain->phnum, somain->load_bias);
411 #endif
412 
413   // Register the main executable and the linker upfront to have
414   // gdb aware of them before loading the rest of the dependency
415   // tree.
416   //
417   // gdb expects the linker to be in the debug shared object list.
418   // Without this, gdb has trouble locating the linker's ".text"
419   // and ".plt" sections. Gdb could also potentially use this to
420   // relocate the offset of our exported 'rtld_db_dlactivity' symbol.
421   //
422   insert_link_map_into_debug_map(&si->link_map_head);
423   insert_link_map_into_debug_map(&solinker->link_map_head);
424 
425   add_vdso();
426 
427   ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(si->base);
428 
429   // We haven't supported non-PIE since Lollipop for security reasons.
430   if (elf_hdr->e_type != ET_DYN) {
431     // We don't use async_safe_fatal here because we don't want a tombstone:
432     // even after several years we still find ourselves on app compatibility
433     // investigations because some app's trying to launch an executable that
434     // hasn't worked in at least three years, and we've "helpfully" dropped a
435     // tombstone for them. The tombstone never provided any detail relevant to
436     // fixing the problem anyway, and the utility of drawing extra attention
437     // to the problem is non-existent at this late date.
438     async_safe_format_fd(STDERR_FILENO,
439                          "\"%s\": error: Android 5.0 and later only support "
440                          "position-independent executables (-fPIE).\n",
441                          g_argv[0]);
442     _exit(EXIT_FAILURE);
443   }
444 
445   // Use LD_LIBRARY_PATH and LD_PRELOAD (but only if we aren't setuid/setgid).
446   parse_LD_LIBRARY_PATH(ldpath_env);
447   parse_LD_PRELOAD(ldpreload_env);
448 
449   std::vector<android_namespace_t*> namespaces = init_default_namespaces(exe_info.path.c_str());
450 
451   if (!si->prelink_image()) __linker_cannot_link(g_argv[0]);
452 
453   // add somain to global group
454   si->set_dt_flags_1(si->get_dt_flags_1() | DF_1_GLOBAL);
455   // ... and add it to all other linked namespaces
456   for (auto linked_ns : namespaces) {
457     if (linked_ns != &g_default_namespace) {
458       linked_ns->add_soinfo(somain);
459       somain->add_secondary_namespace(linked_ns);
460     }
461   }
462 
463   linker_setup_exe_static_tls(g_argv[0]);
464 
465   // Load ld_preloads and dependencies.
466   std::vector<const char*> needed_library_name_list;
467   size_t ld_preloads_count = 0;
468 
469   for (const auto& ld_preload_name : g_ld_preload_names) {
470     needed_library_name_list.push_back(ld_preload_name.c_str());
471     ++ld_preloads_count;
472   }
473 
474   for_each_dt_needed(si, [&](const char* name) {
475     needed_library_name_list.push_back(name);
476   });
477 
478   const char** needed_library_names = &needed_library_name_list[0];
479   size_t needed_libraries_count = needed_library_name_list.size();
480 
481   if (needed_libraries_count > 0 &&
482       !find_libraries(&g_default_namespace,
483                       si,
484                       needed_library_names,
485                       needed_libraries_count,
486                       nullptr,
487                       &g_ld_preloads,
488                       ld_preloads_count,
489                       RTLD_GLOBAL,
490                       nullptr,
491                       true /* add_as_children */,
492                       &namespaces)) {
493     __linker_cannot_link(g_argv[0]);
494   } else if (needed_libraries_count == 0) {
495     if (!si->link_image(SymbolLookupList(si), si, nullptr, nullptr)) {
496       __linker_cannot_link(g_argv[0]);
497     }
498     si->increment_ref_count();
499   }
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 // TODO: There is a similar ifunc resolver calling loop in libc_init_static.cpp, but that version
600 // uses weak symbols, which don't work in the linker prior to its relocation. This version also
601 // supports a load bias. When we stop supporting the gold linker in the NDK, then maybe we can use
602 // non-weak definitions and merge the two loops.
603 #if defined(USE_RELA)
604 extern __LIBC_HIDDEN__ ElfW(Rela) __rela_iplt_start[], __rela_iplt_end[];
605 
call_ifunc_resolvers(ElfW (Addr)load_bias)606 static void call_ifunc_resolvers(ElfW(Addr) load_bias) {
607   for (ElfW(Rela) *r = __rela_iplt_start; r != __rela_iplt_end; ++r) {
608     ElfW(Addr)* offset = reinterpret_cast<ElfW(Addr)*>(r->r_offset + load_bias);
609     ElfW(Addr) resolver = r->r_addend + load_bias;
610     *offset = __bionic_call_ifunc_resolver(resolver);
611   }
612 }
613 #else
614 extern __LIBC_HIDDEN__ ElfW(Rel) __rel_iplt_start[], __rel_iplt_end[];
615 
call_ifunc_resolvers(ElfW (Addr)load_bias)616 static void call_ifunc_resolvers(ElfW(Addr) load_bias) {
617   for (ElfW(Rel) *r = __rel_iplt_start; r != __rel_iplt_end; ++r) {
618     ElfW(Addr)* offset = reinterpret_cast<ElfW(Addr)*>(r->r_offset + load_bias);
619     ElfW(Addr) resolver = *offset + load_bias;
620     *offset = __bionic_call_ifunc_resolver(resolver);
621   }
622 }
623 #endif
624 
625 // Usable before ifunc resolvers have been called. This function is compiled with -ffreestanding.
linker_memclr(void * dst,size_t cnt)626 static void linker_memclr(void* dst, size_t cnt) {
627   for (size_t i = 0; i < cnt; ++i) {
628     reinterpret_cast<char*>(dst)[i] = '\0';
629   }
630 }
631 
632 // Detect an attempt to run the linker on itself. e.g.:
633 //   /system/bin/linker64 /system/bin/linker64
634 // Use priority-1 to run this constructor before other constructors.
detect_self_exec()635 __attribute__((constructor(1))) static void detect_self_exec() {
636   // Normally, the linker initializes the auxv global before calling its
637   // constructors. If the linker loads itself, though, the first loader calls
638   // the second loader's constructors before calling __linker_init.
639   if (__libc_shared_globals()->auxv != nullptr) {
640     return;
641   }
642 #if defined(__i386__)
643   // We don't have access to the auxv struct from here, so use the int 0x80
644   // fallback.
645   __libc_sysinfo = reinterpret_cast<void*>(__libc_int0x80);
646 #endif
647   __linker_error("error: linker cannot load itself\n");
648 }
649 
650 static ElfW(Addr) __attribute__((noinline))
651 __linker_init_post_relocation(KernelArgumentBlock& args, soinfo& linker_so);
652 
653 /*
654  * This is the entry point for the linker, called from begin.S. This
655  * method is responsible for fixing the linker's own relocations, and
656  * then calling __linker_init_post_relocation().
657  *
658  * Because this method is called before the linker has fixed it's own
659  * relocations, any attempt to reference an extern variable, extern
660  * function, or other GOT reference will generate a segfault.
661  */
__linker_init(void * raw_args)662 extern "C" ElfW(Addr) __linker_init(void* raw_args) {
663   // Initialize TLS early so system calls and errno work.
664   KernelArgumentBlock args(raw_args);
665   bionic_tcb temp_tcb __attribute__((uninitialized));
666   linker_memclr(&temp_tcb, sizeof(temp_tcb));
667   __libc_init_main_thread_early(args, &temp_tcb);
668 
669   // When the linker is run by itself (rather than as an interpreter for
670   // another program), AT_BASE is 0.
671   ElfW(Addr) linker_addr = getauxval(AT_BASE);
672   if (linker_addr == 0) {
673     // The AT_PHDR and AT_PHNUM aux values describe this linker instance, so use
674     // the phdr to find the linker's base address.
675     ElfW(Addr) load_bias;
676     get_elf_base_from_phdr(
677       reinterpret_cast<ElfW(Phdr)*>(getauxval(AT_PHDR)), getauxval(AT_PHNUM),
678       &linker_addr, &load_bias);
679   }
680 
681   ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(linker_addr);
682   ElfW(Phdr)* phdr = reinterpret_cast<ElfW(Phdr)*>(linker_addr + elf_hdr->e_phoff);
683 
684   // string.h functions must not be used prior to calling the linker's ifunc resolvers.
685   const ElfW(Addr) load_bias = get_elf_exec_load_bias(elf_hdr);
686   call_ifunc_resolvers(load_bias);
687 
688   soinfo tmp_linker_so(nullptr, nullptr, nullptr, 0, 0);
689 
690   tmp_linker_so.base = linker_addr;
691   tmp_linker_so.size = phdr_table_get_load_size(phdr, elf_hdr->e_phnum);
692   tmp_linker_so.load_bias = load_bias;
693   tmp_linker_so.dynamic = nullptr;
694   tmp_linker_so.phdr = phdr;
695   tmp_linker_so.phnum = elf_hdr->e_phnum;
696   tmp_linker_so.set_linker_flag();
697 
698   // Prelink the linker so we can access linker globals.
699   if (!tmp_linker_so.prelink_image()) __linker_cannot_link(args.argv[0]);
700   if (!tmp_linker_so.link_image(SymbolLookupList(&tmp_linker_so), &tmp_linker_so, nullptr, nullptr)) __linker_cannot_link(args.argv[0]);
701 
702   return __linker_init_post_relocation(args, tmp_linker_so);
703 }
704 
705 /*
706  * This code is called after the linker has linked itself and fixed its own
707  * GOT. It is safe to make references to externs and other non-local data at
708  * this point. The compiler sometimes moves GOT references earlier in a
709  * function, so avoid inlining this function (http://b/80503879).
710  */
711 static ElfW(Addr) __attribute__((noinline))
__linker_init_post_relocation(KernelArgumentBlock & args,soinfo & tmp_linker_so)712 __linker_init_post_relocation(KernelArgumentBlock& args, soinfo& tmp_linker_so) {
713   // Finish initializing the main thread.
714   __libc_init_main_thread_late();
715 
716   // We didn't protect the linker's RELRO pages in link_image because we
717   // couldn't make system calls on x86 at that point, but we can now...
718   if (!tmp_linker_so.protect_relro()) __linker_cannot_link(args.argv[0]);
719 
720   // And we can set VMA name for the bss section now
721   set_bss_vma_name(&tmp_linker_so);
722 
723   // Initialize the linker's static libc's globals
724   __libc_init_globals();
725 
726   // Initialize the linker's own global variables
727   tmp_linker_so.call_constructors();
728 
729   // Setting the linker soinfo's soname can allocate heap memory, so delay it until here.
730   for (const ElfW(Dyn)* d = tmp_linker_so.dynamic; d->d_tag != DT_NULL; ++d) {
731     if (d->d_tag == DT_SONAME) {
732       tmp_linker_so.set_soname(tmp_linker_so.get_string(d->d_un.d_val));
733     }
734   }
735 
736   // When the linker is run directly rather than acting as PT_INTERP, parse
737   // arguments and determine the executable to load. When it's instead acting
738   // as PT_INTERP, AT_ENTRY will refer to the loaded executable rather than the
739   // linker's _start.
740   const char* exe_to_load = nullptr;
741   if (getauxval(AT_ENTRY) == reinterpret_cast<uintptr_t>(&_start)) {
742     if (args.argc == 3 && !strcmp(args.argv[1], "--list")) {
743       // We're being asked to behave like ldd(1).
744       g_is_ldd = true;
745       exe_to_load = args.argv[2];
746     } else if (args.argc <= 1 || !strcmp(args.argv[1], "--help")) {
747       async_safe_format_fd(STDOUT_FILENO,
748          "Usage: %s [--list] PROGRAM [ARGS-FOR-PROGRAM...]\n"
749          "       %s [--list] path.zip!/PROGRAM [ARGS-FOR-PROGRAM...]\n"
750          "\n"
751          "A helper program for linking dynamic executables. Typically, the kernel loads\n"
752          "this program because it's the PT_INTERP of a dynamic executable.\n"
753          "\n"
754          "This program can also be run directly to load and run a dynamic executable. The\n"
755          "executable can be inside a zip file if it's stored uncompressed and at a\n"
756          "page-aligned offset.\n"
757          "\n"
758          "The --list option gives behavior equivalent to ldd(1) on other systems.\n",
759          args.argv[0], args.argv[0]);
760       _exit(EXIT_SUCCESS);
761     } else {
762       exe_to_load = args.argv[1];
763       __libc_shared_globals()->initial_linker_arg_count = 1;
764     }
765   }
766 
767   // store argc/argv/envp to use them for calling constructors
768   g_argc = args.argc - __libc_shared_globals()->initial_linker_arg_count;
769   g_argv = args.argv + __libc_shared_globals()->initial_linker_arg_count;
770   g_envp = args.envp;
771   __libc_shared_globals()->init_progname = g_argv[0];
772 
773   // Initialize static variables. Note that in order to
774   // get correct libdl_info we need to call constructors
775   // before get_libdl_info().
776   sonext = solist = solinker = get_libdl_info(tmp_linker_so);
777   g_default_namespace.add_soinfo(solinker);
778 
779   ElfW(Addr) start_address = linker_main(args, exe_to_load);
780 
781   if (g_is_ldd) _exit(EXIT_SUCCESS);
782 
783   INFO("[ Jumping to _start (%p)... ]", reinterpret_cast<void*>(start_address));
784 
785   // Return the address that the calling assembly stub should jump to.
786   return start_address;
787 }
788