1 // Copyright 2014 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4 
5 // Implementation notes:
6 //
7 // We need to remove a piece from the ELF shared library.  However, we also
8 // want to avoid fixing DWARF cfi data and relative relocation addresses.
9 // So after packing we shift offets and starting address of the RX segment
10 // while preserving code/data vaddrs location.
11 // This requires some fixups for symtab/hash/gnu_hash dynamic section addresses.
12 
13 #include "elf_file.h"
14 
15 #include <stdlib.h>
16 #include <sys/types.h>
17 #include <unistd.h>
18 #include <algorithm>
19 #include <string>
20 #include <vector>
21 
22 #include "debug.h"
23 #include "elf_traits.h"
24 #include "libelf.h"
25 #include "packer.h"
26 
27 namespace relocation_packer {
28 
29 // Out-of-band dynamic tags used to indicate the offset and size of the
30 // android packed relocations section.
31 static constexpr int32_t DT_ANDROID_REL = DT_LOOS + 2;
32 static constexpr int32_t DT_ANDROID_RELSZ = DT_LOOS + 3;
33 
34 static constexpr int32_t DT_ANDROID_RELA = DT_LOOS + 4;
35 static constexpr int32_t DT_ANDROID_RELASZ = DT_LOOS + 5;
36 
37 static constexpr uint32_t SHT_ANDROID_REL = SHT_LOOS + 1;
38 static constexpr uint32_t SHT_ANDROID_RELA = SHT_LOOS + 2;
39 
40 static const size_t kPageSize = 4096;
41 
42 // Alignment to preserve, in bytes.  This must be at least as large as the
43 // largest d_align and sh_addralign values found in the loaded file.
44 // Out of caution for RELRO page alignment, we preserve to a complete target
45 // page.  See http://www.airs.com/blog/archives/189.
46 static const size_t kPreserveAlignment = kPageSize;
47 
48 // Get section data.  Checks that the section has exactly one data entry,
49 // so that the section size and the data size are the same.  True in
50 // practice for all sections we resize when packing or unpacking.  Done
51 // by ensuring that a call to elf_getdata(section, data) returns NULL as
52 // the next data entry.
GetSectionData(Elf_Scn * section)53 static Elf_Data* GetSectionData(Elf_Scn* section) {
54   Elf_Data* data = elf_getdata(section, NULL);
55   CHECK(data && elf_getdata(section, data) == NULL);
56   return data;
57 }
58 
59 // Rewrite section data.  Allocates new data and makes it the data element's
60 // buffer.  Relies on program exit to free allocated data.
RewriteSectionData(Elf_Scn * section,const void * section_data,size_t size)61 static void RewriteSectionData(Elf_Scn* section,
62                                const void* section_data,
63                                size_t size) {
64   Elf_Data* data = GetSectionData(section);
65   CHECK(size == data->d_size);
66   uint8_t* area = new uint8_t[size];
67   memcpy(area, section_data, size);
68   data->d_buf = area;
69 }
70 
71 // Verbose ELF header logging.
72 template <typename Ehdr>
VerboseLogElfHeader(const Ehdr * elf_header)73 static void VerboseLogElfHeader(const Ehdr* elf_header) {
74   VLOG(1) << "e_phoff = " << elf_header->e_phoff;
75   VLOG(1) << "e_shoff = " << elf_header->e_shoff;
76   VLOG(1) << "e_ehsize = " << elf_header->e_ehsize;
77   VLOG(1) << "e_phentsize = " << elf_header->e_phentsize;
78   VLOG(1) << "e_phnum = " << elf_header->e_phnum;
79   VLOG(1) << "e_shnum = " << elf_header->e_shnum;
80   VLOG(1) << "e_shstrndx = " << elf_header->e_shstrndx;
81 }
82 
83 // Verbose ELF program header logging.
84 template <typename Phdr>
VerboseLogProgramHeader(size_t program_header_index,const Phdr * program_header)85 static void VerboseLogProgramHeader(size_t program_header_index,
86                              const Phdr* program_header) {
87   std::string type;
88   switch (program_header->p_type) {
89     case PT_NULL: type = "NULL"; break;
90     case PT_LOAD: type = "LOAD"; break;
91     case PT_DYNAMIC: type = "DYNAMIC"; break;
92     case PT_INTERP: type = "INTERP"; break;
93     case PT_PHDR: type = "PHDR"; break;
94     case PT_GNU_RELRO: type = "GNU_RELRO"; break;
95     case PT_GNU_STACK: type = "GNU_STACK"; break;
96     case PT_ARM_EXIDX: type = "EXIDX"; break;
97     default: type = "(OTHER)"; break;
98   }
99   VLOG(1) << "phdr[" << program_header_index << "] : " << type;
100   VLOG(1) << "  p_offset = " << program_header->p_offset;
101   VLOG(1) << "  p_vaddr = " << program_header->p_vaddr;
102   VLOG(1) << "  p_paddr = " << program_header->p_paddr;
103   VLOG(1) << "  p_filesz = " << program_header->p_filesz;
104   VLOG(1) << "  p_memsz = " << program_header->p_memsz;
105   VLOG(1) << "  p_flags = " << program_header->p_flags;
106   VLOG(1) << "  p_align = " << program_header->p_align;
107 }
108 
109 // Verbose ELF section header logging.
110 template <typename Shdr>
VerboseLogSectionHeader(const std::string & section_name,const Shdr * section_header)111 static void VerboseLogSectionHeader(const std::string& section_name,
112                              const Shdr* section_header) {
113   VLOG(1) << "section " << section_name;
114   VLOG(1) << "  sh_addr = " << section_header->sh_addr;
115   VLOG(1) << "  sh_offset = " << section_header->sh_offset;
116   VLOG(1) << "  sh_size = " << section_header->sh_size;
117   VLOG(1) << "  sh_entsize = " << section_header->sh_entsize;
118   VLOG(1) << "  sh_addralign = " << section_header->sh_addralign;
119 }
120 
121 // Verbose ELF section data logging.
VerboseLogSectionData(const Elf_Data * data)122 static void VerboseLogSectionData(const Elf_Data* data) {
123   VLOG(1) << "  data";
124   VLOG(1) << "    d_buf = " << data->d_buf;
125   VLOG(1) << "    d_off = " << data->d_off;
126   VLOG(1) << "    d_size = " << data->d_size;
127   VLOG(1) << "    d_align = " << data->d_align;
128 }
129 
130 // Load the complete ELF file into a memory image in libelf, and identify
131 // the .rel.dyn or .rela.dyn, .dynamic, and .android.rel.dyn or
132 // .android.rela.dyn sections.  No-op if the ELF file has already been loaded.
133 template <typename ELF>
Load()134 bool ElfFile<ELF>::Load() {
135   if (elf_)
136     return true;
137 
138   Elf* elf = elf_begin(fd_, ELF_C_RDWR, NULL);
139   CHECK(elf);
140 
141   if (elf_kind(elf) != ELF_K_ELF) {
142     LOG(ERROR) << "File not in ELF format";
143     return false;
144   }
145 
146   auto elf_header = ELF::getehdr(elf);
147   if (!elf_header) {
148     LOG(ERROR) << "Failed to load ELF header: " << elf_errmsg(elf_errno());
149     return false;
150   }
151 
152   if (elf_header->e_type != ET_DYN) {
153     LOG(ERROR) << "ELF file is not a shared object";
154     return false;
155   }
156 
157   // Require that our endianness matches that of the target, and that both
158   // are little-endian.  Safe for all current build/target combinations.
159   const int endian = elf_header->e_ident[EI_DATA];
160   CHECK(endian == ELFDATA2LSB);
161   CHECK(__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__);
162 
163   const int file_class = elf_header->e_ident[EI_CLASS];
164   VLOG(1) << "endian = " << endian << ", file class = " << file_class;
165   VerboseLogElfHeader(elf_header);
166 
167   auto elf_program_header = ELF::getphdr(elf);
168   CHECK(elf_program_header != nullptr);
169 
170   const typename ELF::Phdr* dynamic_program_header = NULL;
171   for (size_t i = 0; i < elf_header->e_phnum; ++i) {
172     auto program_header = &elf_program_header[i];
173     VerboseLogProgramHeader(i, program_header);
174 
175     if (program_header->p_type == PT_DYNAMIC) {
176       CHECK(dynamic_program_header == NULL);
177       dynamic_program_header = program_header;
178     }
179   }
180   CHECK(dynamic_program_header != nullptr);
181 
182   size_t string_index;
183   elf_getshdrstrndx(elf, &string_index);
184 
185   // Notes of the dynamic relocations, packed relocations, and .dynamic
186   // sections.  Found while iterating sections, and later stored in class
187   // attributes.
188   Elf_Scn* found_relocations_section = nullptr;
189   Elf_Scn* found_dynamic_section = nullptr;
190 
191   // Notes of relocation section types seen.  We require one or the other of
192   // these; both is unsupported.
193   bool has_rel_relocations = false;
194   bool has_rela_relocations = false;
195   bool has_android_relocations = false;
196 
197   Elf_Scn* section = NULL;
198   while ((section = elf_nextscn(elf, section)) != nullptr) {
199     auto section_header = ELF::getshdr(section);
200     std::string name = elf_strptr(elf, string_index, section_header->sh_name);
201     VerboseLogSectionHeader(name, section_header);
202 
203     // Note relocation section types.
204     if (section_header->sh_type == SHT_REL || section_header->sh_type == SHT_ANDROID_REL) {
205       has_rel_relocations = true;
206     }
207     if (section_header->sh_type == SHT_RELA || section_header->sh_type == SHT_ANDROID_RELA) {
208       has_rela_relocations = true;
209     }
210 
211     // Note special sections as we encounter them.
212     if ((name == ".rel.dyn" || name == ".rela.dyn") &&
213         section_header->sh_size > 0) {
214       found_relocations_section = section;
215 
216       // Note if relocation section is already packed
217       has_android_relocations =
218           section_header->sh_type == SHT_ANDROID_REL ||
219           section_header->sh_type == SHT_ANDROID_RELA;
220     }
221 
222     if (section_header->sh_offset == dynamic_program_header->p_offset) {
223       found_dynamic_section = section;
224     }
225 
226     // Ensure we preserve alignment, repeated later for the data block(s).
227     CHECK(section_header->sh_addralign <= kPreserveAlignment);
228 
229     Elf_Data* data = NULL;
230     while ((data = elf_getdata(section, data)) != NULL) {
231       CHECK(data->d_align <= kPreserveAlignment);
232       VerboseLogSectionData(data);
233     }
234   }
235 
236   // Loading failed if we did not find the required special sections.
237   if (!found_relocations_section) {
238     LOG(ERROR) << "Missing or empty .rel.dyn or .rela.dyn section";
239     return false;
240   }
241   if (!found_dynamic_section) {
242     LOG(ERROR) << "Missing .dynamic section";
243     return false;
244   }
245 
246   // Loading failed if we could not identify the relocations type.
247   if (!has_rel_relocations && !has_rela_relocations) {
248     LOG(ERROR) << "No relocations sections found";
249     return false;
250   }
251   if (has_rel_relocations && has_rela_relocations) {
252     LOG(ERROR) << "Multiple relocations sections with different types found, "
253                << "not currently supported";
254     return false;
255   }
256 
257   elf_ = elf;
258   relocations_section_ = found_relocations_section;
259   dynamic_section_ = found_dynamic_section;
260   relocations_type_ = has_rel_relocations ? REL : RELA;
261   has_android_relocations_ = has_android_relocations;
262   return true;
263 }
264 
265 // Helper for ResizeSection().  Adjust the main ELF header for the hole.
266 template <typename ELF>
AdjustElfHeaderForHole(typename ELF::Ehdr * elf_header,typename ELF::Off hole_start,ssize_t hole_size)267 static void AdjustElfHeaderForHole(typename ELF::Ehdr* elf_header,
268                                    typename ELF::Off hole_start,
269                                    ssize_t hole_size) {
270   if (elf_header->e_phoff > hole_start) {
271     elf_header->e_phoff += hole_size;
272     VLOG(1) << "e_phoff adjusted to " << elf_header->e_phoff;
273   }
274   if (elf_header->e_shoff > hole_start) {
275     elf_header->e_shoff += hole_size;
276     VLOG(1) << "e_shoff adjusted to " << elf_header->e_shoff;
277   }
278 }
279 
280 // Helper for ResizeSection().  Adjust all section headers for the hole.
281 template <typename ELF>
AdjustSectionHeadersForHole(Elf * elf,typename ELF::Off hole_start,ssize_t hole_size)282 static void AdjustSectionHeadersForHole(Elf* elf,
283                                         typename ELF::Off hole_start,
284                                         ssize_t hole_size) {
285   size_t string_index;
286   elf_getshdrstrndx(elf, &string_index);
287 
288   Elf_Scn* section = NULL;
289   while ((section = elf_nextscn(elf, section)) != NULL) {
290     auto section_header = ELF::getshdr(section);
291     std::string name = elf_strptr(elf, string_index, section_header->sh_name);
292 
293     if (section_header->sh_offset > hole_start) {
294       section_header->sh_offset += hole_size;
295       VLOG(1) << "section " << name
296               << " sh_offset adjusted to " << section_header->sh_offset;
297     } else {
298       section_header->sh_addr -= hole_size;
299       VLOG(1) << "section " << name
300               << " sh_addr adjusted to " << section_header->sh_addr;
301     }
302   }
303 }
304 
305 // Helpers for ResizeSection().  On packing, reduce p_align for LOAD segments
306 // to 4kb if larger.  On unpacking, restore p_align for LOAD segments if
307 // packing reduced it to 4kb.  Return true if p_align was changed.
308 template <typename ELF>
ClampLoadSegmentAlignment(typename ELF::Phdr * program_header)309 static bool ClampLoadSegmentAlignment(typename ELF::Phdr* program_header) {
310   CHECK(program_header->p_type == PT_LOAD);
311 
312   // If large, reduce p_align for a LOAD segment to page size on packing.
313   if (program_header->p_align > kPageSize) {
314     program_header->p_align = kPageSize;
315     return true;
316   }
317   return false;
318 }
319 
320 template <typename ELF>
RestoreLoadSegmentAlignment(typename ELF::Phdr * program_headers,size_t count,typename ELF::Phdr * program_header)321 static bool RestoreLoadSegmentAlignment(typename ELF::Phdr* program_headers,
322                                         size_t count,
323                                         typename ELF::Phdr* program_header) {
324   CHECK(program_header->p_type == PT_LOAD);
325 
326   // If p_align was reduced on packing, restore it to its previous value
327   // on unpacking.  We do this by searching for a different LOAD segment
328   // and setting p_align to that of the other LOAD segment found.
329   //
330   // Relies on the following observations:
331   //   - a packable ELF executable has more than one LOAD segment;
332   //   - before packing all LOAD segments have the same p_align;
333   //   - on packing we reduce only one LOAD segment's p_align.
334   if (program_header->p_align == kPageSize) {
335     for (size_t i = 0; i < count; ++i) {
336       typename ELF::Phdr* other_header = &program_headers[i];
337       if (other_header->p_type == PT_LOAD && other_header != program_header) {
338         program_header->p_align = other_header->p_align;
339         return true;
340       }
341     }
342     LOG(WARNING) << "Cannot find a LOAD segment from which to restore p_align";
343   }
344   return false;
345 }
346 
347 template <typename ELF>
AdjustLoadSegmentAlignment(typename ELF::Phdr * program_headers,size_t count,typename ELF::Phdr * program_header,ssize_t hole_size)348 static bool AdjustLoadSegmentAlignment(typename ELF::Phdr* program_headers,
349                                        size_t count,
350                                        typename ELF::Phdr* program_header,
351                                        ssize_t hole_size) {
352   CHECK(program_header->p_type == PT_LOAD);
353 
354   bool status = false;
355   if (hole_size < 0) {
356     status = ClampLoadSegmentAlignment<ELF>(program_header);
357   } else if (hole_size > 0) {
358     status = RestoreLoadSegmentAlignment<ELF>(program_headers,
359                                               count,
360                                               program_header);
361   }
362   return status;
363 }
364 
365 // Helper for ResizeSection().  Adjust the offsets of any program headers
366 // that have offsets currently beyond the hole start, and adjust the
367 // virtual and physical addrs (and perhaps alignment) of the others.
368 template <typename ELF>
AdjustProgramHeaderFields(typename ELF::Phdr * program_headers,size_t count,typename ELF::Off hole_start,ssize_t hole_size)369 static void AdjustProgramHeaderFields(typename ELF::Phdr* program_headers,
370                                       size_t count,
371                                       typename ELF::Off hole_start,
372                                       ssize_t hole_size) {
373   int alignment_changes = 0;
374   for (size_t i = 0; i < count; ++i) {
375     typename ELF::Phdr* program_header = &program_headers[i];
376 
377     // Do not adjust PT_GNU_STACK - it confuses gdb and results
378     // in incorrect unwinding if the executable is stripped after
379     // packing.
380     if (program_header->p_type == PT_GNU_STACK) {
381       continue;
382     }
383 
384     if (program_header->p_offset > hole_start) {
385       // The hole start is past this segment, so adjust offset.
386       program_header->p_offset += hole_size;
387       VLOG(1) << "phdr[" << i
388               << "] p_offset adjusted to "<< program_header->p_offset;
389     } else {
390       program_header->p_vaddr -= hole_size;
391       program_header->p_paddr -= hole_size;
392 
393       // If packing, clamp LOAD segment alignment to 4kb to prevent strip
394       // from adjusting it unnecessarily if run on a packed file.  If
395       // unpacking, attempt to restore a reduced alignment to its previous
396       // value.  Ensure that we do this on at most one LOAD segment.
397       if (program_header->p_type == PT_LOAD) {
398         alignment_changes += AdjustLoadSegmentAlignment<ELF>(program_headers,
399                                                              count,
400                                                              program_header,
401                                                              hole_size);
402         LOG_IF(FATAL, alignment_changes > 1)
403             << "Changed p_align on more than one LOAD segment";
404       }
405 
406       VLOG(1) << "phdr[" << i
407               << "] p_vaddr adjusted to "<< program_header->p_vaddr
408               << "; p_paddr adjusted to "<< program_header->p_paddr
409               << "; p_align adjusted to "<< program_header->p_align;
410     }
411   }
412 }
413 
414 // Helper for ResizeSection().  Find the first loadable segment in the
415 // file.  We expect it to map from file offset zero.
416 template <typename ELF>
FindLoadSegmentForHole(typename ELF::Phdr * program_headers,size_t count,typename ELF::Off hole_start)417 static typename ELF::Phdr* FindLoadSegmentForHole(typename ELF::Phdr* program_headers,
418                                                   size_t count,
419                                                   typename ELF::Off hole_start) {
420   for (size_t i = 0; i < count; ++i) {
421     typename ELF::Phdr* program_header = &program_headers[i];
422 
423     if (program_header->p_type == PT_LOAD &&
424         program_header->p_offset <= hole_start &&
425         (program_header->p_offset + program_header->p_filesz) >= hole_start ) {
426       return program_header;
427     }
428   }
429   LOG(FATAL) << "Cannot locate a LOAD segment with hole_start=0x" << std::hex << hole_start;
430   NOTREACHED();
431 
432   return nullptr;
433 }
434 
435 // Helper for ResizeSection().  Rewrite program headers.
436 template <typename ELF>
RewriteProgramHeadersForHole(Elf * elf,typename ELF::Off hole_start,ssize_t hole_size)437 static void RewriteProgramHeadersForHole(Elf* elf,
438                                          typename ELF::Off hole_start,
439                                          ssize_t hole_size) {
440   const typename ELF::Ehdr* elf_header = ELF::getehdr(elf);
441   CHECK(elf_header);
442 
443   typename ELF::Phdr* elf_program_header = ELF::getphdr(elf);
444   CHECK(elf_program_header);
445 
446   const size_t program_header_count = elf_header->e_phnum;
447 
448   // Locate the segment that we can overwrite to form the new LOAD entry,
449   // and the segment that we are going to split into two parts.
450   typename ELF::Phdr* target_load_header =
451       FindLoadSegmentForHole<ELF>(elf_program_header, program_header_count, hole_start);
452 
453   VLOG(1) << "phdr[" << target_load_header - elf_program_header << "] adjust";
454   // Adjust PT_LOAD program header memsz and filesz
455   target_load_header->p_filesz += hole_size;
456   target_load_header->p_memsz += hole_size;
457 
458   // Adjust the offsets and p_vaddrs
459   AdjustProgramHeaderFields<ELF>(elf_program_header,
460                                  program_header_count,
461                                  hole_start,
462                                  hole_size);
463 }
464 
465 // Helper for ResizeSection().  Locate and return the dynamic section.
466 template <typename ELF>
GetDynamicSection(Elf * elf)467 static Elf_Scn* GetDynamicSection(Elf* elf) {
468   const typename ELF::Ehdr* elf_header = ELF::getehdr(elf);
469   CHECK(elf_header);
470 
471   const typename ELF::Phdr* elf_program_header = ELF::getphdr(elf);
472   CHECK(elf_program_header);
473 
474   // Find the program header that describes the dynamic section.
475   const typename ELF::Phdr* dynamic_program_header = NULL;
476   for (size_t i = 0; i < elf_header->e_phnum; ++i) {
477     const typename ELF::Phdr* program_header = &elf_program_header[i];
478 
479     if (program_header->p_type == PT_DYNAMIC) {
480       dynamic_program_header = program_header;
481     }
482   }
483   CHECK(dynamic_program_header);
484 
485   // Now find the section with the same offset as this program header.
486   Elf_Scn* dynamic_section = NULL;
487   Elf_Scn* section = NULL;
488   while ((section = elf_nextscn(elf, section)) != NULL) {
489     typename ELF::Shdr* section_header = ELF::getshdr(section);
490 
491     if (section_header->sh_offset == dynamic_program_header->p_offset) {
492       dynamic_section = section;
493     }
494   }
495   CHECK(dynamic_section != NULL);
496 
497   return dynamic_section;
498 }
499 
500 // Helper for ResizeSection().  Adjust the .dynamic section for the hole.
501 template <typename ELF>
AdjustDynamicSectionForHole(Elf_Scn * dynamic_section,typename ELF::Off hole_start,ssize_t hole_size,relocations_type_t relocations_type)502 void ElfFile<ELF>::AdjustDynamicSectionForHole(Elf_Scn* dynamic_section,
503                                                typename ELF::Off hole_start,
504                                                ssize_t hole_size,
505                                                relocations_type_t relocations_type) {
506   CHECK(relocations_type != NONE);
507   Elf_Data* data = GetSectionData(dynamic_section);
508 
509   auto dynamic_base = reinterpret_cast<typename ELF::Dyn*>(data->d_buf);
510   std::vector<typename ELF::Dyn> dynamics(
511       dynamic_base,
512       dynamic_base + data->d_size / sizeof(dynamics[0]));
513 
514   if (hole_size > 0) { // expanding
515     hole_start += hole_size;
516   }
517 
518   for (size_t i = 0; i < dynamics.size(); ++i) {
519     typename ELF::Dyn* dynamic = &dynamics[i];
520     const typename ELF::Sword tag = dynamic->d_tag;
521 
522     // Any tags that hold offsets are adjustment candidates.
523     const bool is_adjustable = (tag == DT_PLTGOT ||
524                                 tag == DT_HASH ||
525                                 tag == DT_GNU_HASH ||
526                                 tag == DT_STRTAB ||
527                                 tag == DT_SYMTAB ||
528                                 tag == DT_RELA ||
529                                 tag == DT_INIT ||
530                                 tag == DT_FINI ||
531                                 tag == DT_REL ||
532                                 tag == DT_JMPREL ||
533                                 tag == DT_INIT_ARRAY ||
534                                 tag == DT_FINI_ARRAY ||
535                                 tag == DT_VERSYM ||
536                                 tag == DT_VERNEED ||
537                                 tag == DT_VERDEF ||
538                                 tag == DT_ANDROID_REL||
539                                 tag == DT_ANDROID_RELA);
540 
541     if (is_adjustable && dynamic->d_un.d_ptr <= hole_start) {
542       dynamic->d_un.d_ptr -= hole_size;
543       VLOG(1) << "dynamic[" << i << "] " << dynamic->d_tag
544               << " d_ptr adjusted to " << dynamic->d_un.d_ptr;
545     }
546 
547     // DT_RELSZ or DT_RELASZ indicate the overall size of relocations.
548     // Only one will be present.  Adjust by hole size.
549     if (tag == DT_RELSZ || tag == DT_RELASZ || tag == DT_ANDROID_RELSZ || tag == DT_ANDROID_RELASZ) {
550       dynamic->d_un.d_val += hole_size;
551       VLOG(1) << "dynamic[" << i << "] " << dynamic->d_tag
552               << " d_val adjusted to " << dynamic->d_un.d_val;
553     }
554 
555     // Special case: DT_MIPS_RLD_MAP2 stores the difference between dynamic
556     // entry address and the address of the _r_debug (used by GDB)
557     // since the dynamic section and target address are on the
558     // different sides of the hole it needs to be adjusted accordingly
559     if (tag == DT_MIPS_RLD_MAP2) {
560       dynamic->d_un.d_val += hole_size;
561       VLOG(1) << "dynamic[" << i << "] " << dynamic->d_tag
562               << " d_val adjusted to " << dynamic->d_un.d_val;
563     }
564 
565     // Ignore DT_RELCOUNT and DT_RELACOUNT: (1) nobody uses them and
566     // technically (2) the relative relocation count is not changed.
567 
568     // DT_RELENT and DT_RELAENT don't change, ignore them as well.
569   }
570 
571   void* section_data = &dynamics[0];
572   size_t bytes = dynamics.size() * sizeof(dynamics[0]);
573   RewriteSectionData(dynamic_section, section_data, bytes);
574 }
575 
576 // Resize a section.  If the new size is larger than the current size, open
577 // up a hole by increasing file offsets that come after the hole.  If smaller
578 // than the current size, remove the hole by decreasing those offsets.
579 template <typename ELF>
ResizeSection(Elf * elf,Elf_Scn * section,size_t new_size,typename ELF::Word new_sh_type,relocations_type_t relocations_type)580 void ElfFile<ELF>::ResizeSection(Elf* elf, Elf_Scn* section, size_t new_size,
581                                  typename ELF::Word new_sh_type,
582                                  relocations_type_t relocations_type) {
583 
584   size_t string_index;
585   elf_getshdrstrndx(elf, &string_index);
586   auto section_header = ELF::getshdr(section);
587   std::string name = elf_strptr(elf, string_index, section_header->sh_name);
588 
589   if (section_header->sh_size == new_size) {
590     return;
591   }
592 
593   // Require that the section size and the data size are the same.  True
594   // in practice for all sections we resize when packing or unpacking.
595   Elf_Data* data = GetSectionData(section);
596   CHECK(data->d_off == 0 && data->d_size == section_header->sh_size);
597 
598   // Require that the section is not zero-length (that is, has allocated
599   // data that we can validly expand).
600   CHECK(data->d_size && data->d_buf);
601 
602   const auto hole_start = section_header->sh_offset;
603   const ssize_t hole_size = new_size - data->d_size;
604 
605   VLOG_IF(1, (hole_size > 0)) << "expand section (" << name << ") size: " <<
606       data->d_size << " -> " << (data->d_size + hole_size);
607   VLOG_IF(1, (hole_size < 0)) << "shrink section (" << name << ") size: " <<
608       data->d_size << " -> " << (data->d_size + hole_size);
609 
610   // libelf overrides sh_entsize for known sh_types, so it does not matter what we set
611   // for SHT_REL/SHT_RELA.
612   typename ELF::Xword new_entsize =
613       (new_sh_type == SHT_ANDROID_REL || new_sh_type == SHT_ANDROID_RELA) ? 1 : 0;
614 
615   VLOG(1) << "Update section (" << name << ") entry size: " <<
616       section_header->sh_entsize << " -> " << new_entsize;
617 
618   // Resize the data and the section header.
619   data->d_size += hole_size;
620   section_header->sh_size += hole_size;
621   section_header->sh_entsize = new_entsize;
622   section_header->sh_type = new_sh_type;
623 
624   // Add the hole size to all offsets in the ELF file that are after the
625   // start of the hole.  If the hole size is positive we are expanding the
626   // section to create a new hole; if negative, we are closing up a hole.
627 
628   // Start with the main ELF header.
629   typename ELF::Ehdr* elf_header = ELF::getehdr(elf);
630   AdjustElfHeaderForHole<ELF>(elf_header, hole_start, hole_size);
631 
632   // Adjust all section headers.
633   AdjustSectionHeadersForHole<ELF>(elf, hole_start, hole_size);
634 
635   // Rewrite the program headers to either split or coalesce segments,
636   // and adjust dynamic entries to match.
637   RewriteProgramHeadersForHole<ELF>(elf, hole_start, hole_size);
638 
639   Elf_Scn* dynamic_section = GetDynamicSection<ELF>(elf);
640   AdjustDynamicSectionForHole(dynamic_section, hole_start, hole_size, relocations_type);
641 }
642 
643 // Find the first slot in a dynamics array with the given tag.  The array
644 // always ends with a free (unused) element, and which we exclude from the
645 // search.  Returns dynamics->size() if not found.
646 template <typename ELF>
FindDynamicEntry(typename ELF::Sword tag,std::vector<typename ELF::Dyn> * dynamics)647 static size_t FindDynamicEntry(typename ELF::Sword tag,
648                                std::vector<typename ELF::Dyn>* dynamics) {
649   // Loop until the penultimate entry.  We exclude the end sentinel.
650   for (size_t i = 0; i < dynamics->size() - 1; ++i) {
651     if (dynamics->at(i).d_tag == tag) {
652       return i;
653     }
654   }
655 
656   // The tag was not found.
657   return dynamics->size();
658 }
659 
660 // Replace dynamic entry.
661 template <typename ELF>
ReplaceDynamicEntry(typename ELF::Sword tag,const typename ELF::Dyn & dyn,std::vector<typename ELF::Dyn> * dynamics)662 static void ReplaceDynamicEntry(typename ELF::Sword tag,
663                                 const typename ELF::Dyn& dyn,
664                                 std::vector<typename ELF::Dyn>* dynamics) {
665   const size_t slot = FindDynamicEntry<ELF>(tag, dynamics);
666   if (slot == dynamics->size()) {
667     LOG(FATAL) << "Dynamic slot is not found for tag=" << tag;
668   }
669 
670   // Replace this entry with the one supplied.
671   dynamics->at(slot) = dyn;
672   VLOG(1) << "dynamic[" << slot << "] overwritten with " << dyn.d_tag;
673 }
674 
675 // Remove relative entries from dynamic relocations and write as packed
676 // data into android packed relocations.
677 template <typename ELF>
PackRelocations()678 bool ElfFile<ELF>::PackRelocations() {
679   // Load the ELF file into libelf.
680   if (!Load()) {
681     LOG(ERROR) << "Failed to load as ELF";
682     return false;
683   }
684 
685   // Retrieve the current dynamic relocations section data.
686   Elf_Data* data = GetSectionData(relocations_section_);
687   // we always pack rela, because packed format is pretty much the same
688   std::vector<typename ELF::Rela> relocations;
689 
690   if (relocations_type_ == REL) {
691     // Convert data to a vector of relocations.
692     const typename ELF::Rel* relocations_base = reinterpret_cast<typename ELF::Rel*>(data->d_buf);
693     ConvertRelArrayToRelaVector(relocations_base,
694         data->d_size / sizeof(typename ELF::Rel), &relocations);
695     VLOG(1) << "Relocations   : REL";
696   } else if (relocations_type_ == RELA) {
697     // Convert data to a vector of relocations with addends.
698     const typename ELF::Rela* relocations_base = reinterpret_cast<typename ELF::Rela*>(data->d_buf);
699     relocations = std::vector<typename ELF::Rela>(
700         relocations_base,
701         relocations_base + data->d_size / sizeof(relocations[0]));
702 
703     VLOG(1) << "Relocations   : RELA";
704   } else {
705     NOTREACHED();
706   }
707 
708   return PackTypedRelocations(&relocations);
709 }
710 
711 // Helper for PackRelocations().  Rel type is one of ELF::Rel or ELF::Rela.
712 template <typename ELF>
PackTypedRelocations(std::vector<typename ELF::Rela> * relocations)713 bool ElfFile<ELF>::PackTypedRelocations(std::vector<typename ELF::Rela>* relocations) {
714   typedef typename ELF::Rela Rela;
715 
716   if (has_android_relocations_) {
717     LOG(INFO) << "Relocation table is already packed";
718     return true;
719   }
720 
721   // If no relocations then we have nothing packable.  Perhaps
722   // the shared object has already been packed?
723   if (relocations->empty()) {
724     LOG(ERROR) << "No relocations found";
725     return false;
726   }
727 
728   const size_t rel_size =
729       relocations_type_ == RELA ? sizeof(typename ELF::Rela) : sizeof(typename ELF::Rel);
730   const size_t initial_bytes = relocations->size() * rel_size;
731 
732   VLOG(1) << "Unpacked                   : " << initial_bytes << " bytes";
733   std::vector<uint8_t> packed;
734   RelocationPacker<ELF> packer;
735 
736   // Pack relocations: dry run to estimate memory savings.
737   packer.PackRelocations(*relocations, &packed);
738   const size_t packed_bytes_estimate = packed.size() * sizeof(packed[0]);
739   VLOG(1) << "Packed         (no padding): " << packed_bytes_estimate << " bytes";
740 
741   if (packed.empty()) {
742     VLOG(1) << "Too few relocations to pack";
743     return true;
744   }
745 
746   // Pre-calculate the size of the hole we will close up when we rewrite
747   // dynamic relocations.  We have to adjust relocation addresses to
748   // account for this.
749   typename ELF::Shdr* section_header = ELF::getshdr(relocations_section_);
750   ssize_t hole_size = initial_bytes - packed_bytes_estimate;
751 
752   // hole_size needs to be page_aligned.
753   hole_size -= hole_size % kPreserveAlignment;
754 
755   VLOG(1) << "Compaction                 : " << hole_size << " bytes";
756 
757   // Adjusting for alignment may have removed any packing benefit.
758   if (hole_size == 0) {
759     VLOG(1) << "Too few relocations to pack after alignment";
760     return true;
761   }
762 
763   if (hole_size <= 0) {
764     VLOG(1) << "Packing relocations saves no space";
765     return true;
766   }
767 
768   size_t data_padding_bytes = is_padding_relocations_ ?
769       initial_bytes - packed_bytes_estimate :
770       initial_bytes - hole_size - packed_bytes_estimate;
771 
772   // pad data
773   std::vector<uint8_t> padding(data_padding_bytes, 0);
774   packed.insert(packed.end(), padding.begin(), padding.end());
775 
776   const void* packed_data = &packed[0];
777 
778   // Run a loopback self-test as a check that packing is lossless.
779   std::vector<Rela> unpacked;
780   packer.UnpackRelocations(packed, &unpacked);
781   CHECK(unpacked.size() == relocations->size());
782   CHECK(!memcmp(&unpacked[0],
783                 &relocations->at(0),
784                 unpacked.size() * sizeof(unpacked[0])));
785 
786   // Rewrite the current dynamic relocations section with packed one then shrink it to size.
787   const size_t bytes = packed.size() * sizeof(packed[0]);
788   ResizeSection(elf_, relocations_section_, bytes,
789       relocations_type_ == REL ? SHT_ANDROID_REL : SHT_ANDROID_RELA, relocations_type_);
790   RewriteSectionData(relocations_section_, packed_data, bytes);
791 
792   // TODO (dimitry): fix string table and replace .rel.dyn/plt with .android.rel.dyn/plt
793 
794   // Rewrite .dynamic and rename relocation tags describing the packed android
795   // relocations.
796   Elf_Data* data = GetSectionData(dynamic_section_);
797   const typename ELF::Dyn* dynamic_base = reinterpret_cast<typename ELF::Dyn*>(data->d_buf);
798   std::vector<typename ELF::Dyn> dynamics(
799       dynamic_base,
800       dynamic_base + data->d_size / sizeof(dynamics[0]));
801   section_header = ELF::getshdr(relocations_section_);
802   {
803     typename ELF::Dyn dyn;
804     dyn.d_tag = relocations_type_ == REL ? DT_ANDROID_REL : DT_ANDROID_RELA;
805     dyn.d_un.d_ptr = section_header->sh_addr;
806     ReplaceDynamicEntry<ELF>(relocations_type_ == REL ? DT_REL : DT_RELA, dyn, &dynamics);
807   }
808   {
809     typename ELF::Dyn dyn;
810     dyn.d_tag = relocations_type_ == REL ? DT_ANDROID_RELSZ : DT_ANDROID_RELASZ;
811     dyn.d_un.d_val = section_header->sh_size;
812     ReplaceDynamicEntry<ELF>(relocations_type_ == REL ? DT_RELSZ : DT_RELASZ, dyn, &dynamics);
813   }
814 
815   const void* dynamics_data = &dynamics[0];
816   const size_t dynamics_bytes = dynamics.size() * sizeof(dynamics[0]);
817   RewriteSectionData(dynamic_section_, dynamics_data, dynamics_bytes);
818 
819   Flush();
820   return true;
821 }
822 
823 // Find packed relative relocations in the packed android relocations
824 // section, unpack them, and rewrite the dynamic relocations section to
825 // contain unpacked data.
826 template <typename ELF>
UnpackRelocations()827 bool ElfFile<ELF>::UnpackRelocations() {
828   // Load the ELF file into libelf.
829   if (!Load()) {
830     LOG(ERROR) << "Failed to load as ELF";
831     return false;
832   }
833 
834   typename ELF::Shdr* section_header = ELF::getshdr(relocations_section_);
835   // Retrieve the current packed android relocations section data.
836   Elf_Data* data = GetSectionData(relocations_section_);
837 
838   // Convert data to a vector of bytes.
839   const uint8_t* packed_base = reinterpret_cast<uint8_t*>(data->d_buf);
840   std::vector<uint8_t> packed(
841       packed_base,
842       packed_base + data->d_size / sizeof(packed[0]));
843 
844   if ((section_header->sh_type == SHT_ANDROID_RELA || section_header->sh_type == SHT_ANDROID_REL) &&
845       packed.size() > 3 &&
846       packed[0] == 'A' &&
847       packed[1] == 'P' &&
848       packed[2] == 'S' &&
849       packed[3] == '2') {
850     LOG(INFO) << "Relocations   : " << (relocations_type_ == REL ? "REL" : "RELA");
851   } else {
852     LOG(ERROR) << "Packed relocations not found (not packed?)";
853     return false;
854   }
855 
856   return UnpackTypedRelocations(packed);
857 }
858 
859 // Helper for UnpackRelocations().  Rel type is one of ELF::Rel or ELF::Rela.
860 template <typename ELF>
UnpackTypedRelocations(const std::vector<uint8_t> & packed)861 bool ElfFile<ELF>::UnpackTypedRelocations(const std::vector<uint8_t>& packed) {
862   // Unpack the data to re-materialize the relative relocations.
863   const size_t packed_bytes = packed.size() * sizeof(packed[0]);
864   LOG(INFO) << "Packed           : " << packed_bytes << " bytes";
865   std::vector<typename ELF::Rela> unpacked_relocations;
866   RelocationPacker<ELF> packer;
867   packer.UnpackRelocations(packed, &unpacked_relocations);
868 
869   const size_t relocation_entry_size =
870       relocations_type_ == REL ? sizeof(typename ELF::Rel) : sizeof(typename ELF::Rela);
871   const size_t unpacked_bytes = unpacked_relocations.size() * relocation_entry_size;
872   LOG(INFO) << "Unpacked         : " << unpacked_bytes << " bytes";
873 
874   // Retrieve the current dynamic relocations section data.
875   Elf_Data* data = GetSectionData(relocations_section_);
876 
877   LOG(INFO) << "Relocations      : " << unpacked_relocations.size() << " entries";
878 
879   // If we found the same number of null relocation entries in the dynamic
880   // relocations section as we hold as unpacked relative relocations, then
881   // this is a padded file.
882 
883   const bool is_padded = packed_bytes == unpacked_bytes;
884 
885   // Unless padded, pre-apply relative relocations to account for the
886   // hole, and pre-adjust all relocation offsets accordingly.
887   typename ELF::Shdr* section_header = ELF::getshdr(relocations_section_);
888 
889   if (!is_padded) {
890     LOG(INFO) << "Expansion     : " << unpacked_bytes - packed_bytes << " bytes";
891   }
892 
893   // Rewrite the current dynamic relocations section with unpacked version of
894   // relocations.
895   const void* section_data = nullptr;
896   std::vector<typename ELF::Rel> unpacked_rel_relocations;
897   if (relocations_type_ == RELA) {
898     section_data = &unpacked_relocations[0];
899   } else if (relocations_type_ == REL) {
900     ConvertRelaVectorToRelVector(unpacked_relocations, &unpacked_rel_relocations);
901     section_data = &unpacked_rel_relocations[0];
902   } else {
903     NOTREACHED();
904   }
905 
906   ResizeSection(elf_, relocations_section_, unpacked_bytes,
907       relocations_type_ == REL ? SHT_REL : SHT_RELA, relocations_type_);
908   RewriteSectionData(relocations_section_, section_data, unpacked_bytes);
909 
910   // Rewrite .dynamic to remove two tags describing packed android relocations.
911   data = GetSectionData(dynamic_section_);
912   const typename ELF::Dyn* dynamic_base = reinterpret_cast<typename ELF::Dyn*>(data->d_buf);
913   std::vector<typename ELF::Dyn> dynamics(
914       dynamic_base,
915       dynamic_base + data->d_size / sizeof(dynamics[0]));
916   {
917     typename ELF::Dyn dyn;
918     dyn.d_tag = relocations_type_ == REL ? DT_REL : DT_RELA;
919     dyn.d_un.d_ptr = section_header->sh_addr;
920     ReplaceDynamicEntry<ELF>(relocations_type_ == REL ? DT_ANDROID_REL : DT_ANDROID_RELA,
921         dyn, &dynamics);
922   }
923 
924   {
925     typename ELF::Dyn dyn;
926     dyn.d_tag = relocations_type_ == REL ? DT_RELSZ : DT_RELASZ;
927     dyn.d_un.d_val = section_header->sh_size;
928     ReplaceDynamicEntry<ELF>(relocations_type_ == REL ? DT_ANDROID_RELSZ : DT_ANDROID_RELASZ,
929         dyn, &dynamics);
930   }
931 
932   const void* dynamics_data = &dynamics[0];
933   const size_t dynamics_bytes = dynamics.size() * sizeof(dynamics[0]);
934   RewriteSectionData(dynamic_section_, dynamics_data, dynamics_bytes);
935 
936   Flush();
937   return true;
938 }
939 
940 // Flush rewritten shared object file data.
941 template <typename ELF>
Flush()942 void ElfFile<ELF>::Flush() {
943   // Flag all ELF data held in memory as needing to be written back to the
944   // file, and tell libelf that we have controlled the file layout.
945   elf_flagelf(elf_, ELF_C_SET, ELF_F_DIRTY);
946   elf_flagelf(elf_, ELF_C_SET, ELF_F_LAYOUT);
947 
948   // Write ELF data back to disk.
949   const off_t file_bytes = elf_update(elf_, ELF_C_WRITE);
950   if (file_bytes == -1) {
951     LOG(ERROR) << "elf_update failed: " << elf_errmsg(elf_errno());
952   }
953 
954   CHECK(file_bytes > 0);
955   VLOG(1) << "elf_update returned: " << file_bytes;
956 
957   // Clean up libelf, and truncate the output file to the number of bytes
958   // written by elf_update().
959   elf_end(elf_);
960   elf_ = NULL;
961   const int truncate = ftruncate(fd_, file_bytes);
962   CHECK(truncate == 0);
963 }
964 
965 template <typename ELF>
ConvertRelArrayToRelaVector(const typename ELF::Rel * rel_array,size_t rel_array_size,std::vector<typename ELF::Rela> * rela_vector)966 void ElfFile<ELF>::ConvertRelArrayToRelaVector(const typename ELF::Rel* rel_array,
967                                                size_t rel_array_size,
968                                                std::vector<typename ELF::Rela>* rela_vector) {
969   for (size_t i = 0; i<rel_array_size; ++i) {
970     typename ELF::Rela rela;
971     rela.r_offset = rel_array[i].r_offset;
972     rela.r_info = rel_array[i].r_info;
973     rela.r_addend = 0;
974     rela_vector->push_back(rela);
975   }
976 }
977 
978 template <typename ELF>
ConvertRelaVectorToRelVector(const std::vector<typename ELF::Rela> & rela_vector,std::vector<typename ELF::Rel> * rel_vector)979 void ElfFile<ELF>::ConvertRelaVectorToRelVector(const std::vector<typename ELF::Rela>& rela_vector,
980                                                 std::vector<typename ELF::Rel>* rel_vector) {
981   for (auto rela : rela_vector) {
982     typename ELF::Rel rel;
983     rel.r_offset = rela.r_offset;
984     rel.r_info = rela.r_info;
985     CHECK(rela.r_addend == 0);
986     rel_vector->push_back(rel);
987   }
988 }
989 
990 template class ElfFile<ELF32_traits>;
991 template class ElfFile<ELF64_traits>;
992 
993 }  // namespace relocation_packer
994