elf_builder.h revision ae5d2738a2b941b543c3fd478af910d4cd16f2ba
1/*
2 * Copyright (C) 2015 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#ifndef ART_COMPILER_ELF_BUILDER_H_
18#define ART_COMPILER_ELF_BUILDER_H_
19
20#include <vector>
21
22#include "arch/instruction_set.h"
23#include "base/bit_utils.h"
24#include "base/casts.h"
25#include "base/unix_file/fd_file.h"
26#include "elf_utils.h"
27#include "leb128.h"
28#include "linker/error_delaying_output_stream.h"
29#include "utils/array_ref.h"
30
31namespace art {
32
33// Writes ELF file.
34//
35// The basic layout of the elf file:
36//   Elf_Ehdr                    - The ELF header.
37//   Elf_Phdr[]                  - Program headers for the linker.
38//   .rodata                     - DEX files and oat metadata.
39//   .text                       - Compiled code.
40//   .bss                        - Zero-initialized writeable section.
41//   .dynstr                     - Names for .dynsym.
42//   .dynsym                     - A few oat-specific dynamic symbols.
43//   .hash                       - Hash-table for .dynsym.
44//   .dynamic                    - Tags which let the linker locate .dynsym.
45//   .strtab                     - Names for .symtab.
46//   .symtab                     - Debug symbols.
47//   .eh_frame                   - Unwind information (CFI).
48//   .eh_frame_hdr               - Index of .eh_frame.
49//   .debug_frame                - Unwind information (CFI).
50//   .debug_frame.oat_patches    - Addresses for relocation.
51//   .debug_info                 - Debug information.
52//   .debug_info.oat_patches     - Addresses for relocation.
53//   .debug_abbrev               - Decoding information for .debug_info.
54//   .debug_str                  - Strings for .debug_info.
55//   .debug_line                 - Line number tables.
56//   .debug_line.oat_patches     - Addresses for relocation.
57//   .text.oat_patches           - Addresses for relocation.
58//   .shstrtab                   - Names of ELF sections.
59//   Elf_Shdr[]                  - Section headers.
60//
61// Some section are optional (the debug sections in particular).
62//
63// We try write the section data directly into the file without much
64// in-memory buffering.  This means we generally write sections based on the
65// dependency order (e.g. .dynamic points to .dynsym which points to .text).
66//
67// In the cases where we need to buffer, we write the larger section first
68// and buffer the smaller one (e.g. .strtab is bigger than .symtab).
69//
70// The debug sections are written last for easier stripping.
71//
72template <typename ElfTypes>
73class ElfBuilder FINAL {
74 public:
75  static constexpr size_t kMaxProgramHeaders = 16;
76  using Elf_Addr = typename ElfTypes::Addr;
77  using Elf_Off = typename ElfTypes::Off;
78  using Elf_Word = typename ElfTypes::Word;
79  using Elf_Sword = typename ElfTypes::Sword;
80  using Elf_Ehdr = typename ElfTypes::Ehdr;
81  using Elf_Shdr = typename ElfTypes::Shdr;
82  using Elf_Sym = typename ElfTypes::Sym;
83  using Elf_Phdr = typename ElfTypes::Phdr;
84  using Elf_Dyn = typename ElfTypes::Dyn;
85
86  // Base class of all sections.
87  class Section : public OutputStream {
88   public:
89    Section(ElfBuilder<ElfTypes>* owner, const std::string& name,
90            Elf_Word type, Elf_Word flags, const Section* link,
91            Elf_Word info, Elf_Word align, Elf_Word entsize)
92        : OutputStream(name), owner_(owner), header_(),
93          section_index_(0), name_(name), link_(link),
94          started_(false), finished_(false), phdr_flags_(PF_R), phdr_type_(0) {
95      DCHECK_GE(align, 1u);
96      header_.sh_type = type;
97      header_.sh_flags = flags;
98      header_.sh_info = info;
99      header_.sh_addralign = align;
100      header_.sh_entsize = entsize;
101    }
102
103    // Start writing of this section.
104    void Start() {
105      CHECK(!started_);
106      CHECK(!finished_);
107      started_ = true;
108      auto& sections = owner_->sections_;
109      // Check that the previous section is complete.
110      CHECK(sections.empty() || sections.back()->finished_);
111      // The first ELF section index is 1. Index 0 is reserved for NULL.
112      section_index_ = sections.size() + 1;
113      // Page-align if we switch between allocated and non-allocated sections,
114      // or if we change the type of allocation (e.g. executable vs non-executable).
115      if (!sections.empty()) {
116        if (header_.sh_flags != sections.back()->header_.sh_flags) {
117          header_.sh_addralign = kPageSize;
118        }
119      }
120      // Align file position.
121      if (header_.sh_type != SHT_NOBITS) {
122        header_.sh_offset = owner_->AlignFileOffset(header_.sh_addralign);
123      } else {
124        header_.sh_offset = 0;
125      }
126      // Align virtual memory address.
127      if ((header_.sh_flags & SHF_ALLOC) != 0) {
128        header_.sh_addr = owner_->AlignVirtualAddress(header_.sh_addralign);
129      } else {
130        header_.sh_addr = 0;
131      }
132      // Push this section on the list of written sections.
133      sections.push_back(this);
134    }
135
136    // Finish writing of this section.
137    void End() {
138      CHECK(started_);
139      CHECK(!finished_);
140      finished_ = true;
141      if (header_.sh_type == SHT_NOBITS) {
142        CHECK_GT(header_.sh_size, 0u);
143      } else {
144        // Use the current file position to determine section size.
145        off_t file_offset = owner_->stream_.Seek(0, kSeekCurrent);
146        CHECK_GE(file_offset, (off_t)header_.sh_offset);
147        header_.sh_size = file_offset - header_.sh_offset;
148      }
149      if ((header_.sh_flags & SHF_ALLOC) != 0) {
150        owner_->virtual_address_ += header_.sh_size;
151      }
152    }
153
154    // Returns true if the section was written to disk.
155    // (Used to check whether we have .text when writing JIT debug info)
156    bool Exists() const {
157      return finished_;
158    }
159
160    // Get the location of this section in virtual memory.
161    Elf_Addr GetAddress() const {
162      CHECK(started_);
163      return header_.sh_addr;
164    }
165
166    // Returns the size of the content of this section.
167    Elf_Word GetSize() const {
168      if (finished_) {
169        return header_.sh_size;
170      } else {
171        CHECK(started_);
172        CHECK_NE(header_.sh_type, (Elf_Word)SHT_NOBITS);
173        return owner_->stream_.Seek(0, kSeekCurrent) - header_.sh_offset;
174      }
175    }
176
177    // Write this section as "NOBITS" section. (used for the .bss section)
178    // This means that the ELF file does not contain the initial data for this section
179    // and it will be zero-initialized when the ELF file is loaded in the running program.
180    void WriteNoBitsSection(Elf_Word size) {
181      DCHECK_NE(header_.sh_flags & SHF_ALLOC, 0u);
182      header_.sh_type = SHT_NOBITS;
183      Start();
184      header_.sh_size = size;
185      End();
186    }
187
188    // This function always succeeds to simplify code.
189    // Use builder's Good() to check the actual status.
190    bool WriteFully(const void* buffer, size_t byte_count) OVERRIDE {
191      CHECK(started_);
192      CHECK(!finished_);
193      return owner_->stream_.WriteFully(buffer, byte_count);
194    }
195
196    // This function always succeeds to simplify code.
197    // Use builder's Good() to check the actual status.
198    off_t Seek(off_t offset, Whence whence) OVERRIDE {
199      // Forward the seek as-is and trust the caller to use it reasonably.
200      return owner_->stream_.Seek(offset, whence);
201    }
202
203    // This function flushes the output and returns whether it succeeded.
204    // If there was a previous failure, this does nothing and returns false, i.e. failed.
205    bool Flush() OVERRIDE {
206      return owner_->stream_.Flush();
207    }
208
209    Elf_Word GetSectionIndex() const {
210      DCHECK(started_);
211      DCHECK_NE(section_index_, 0u);
212      return section_index_;
213    }
214
215   private:
216    ElfBuilder<ElfTypes>* owner_;
217    Elf_Shdr header_;
218    Elf_Word section_index_;
219    const std::string name_;
220    const Section* const link_;
221    bool started_;
222    bool finished_;
223    Elf_Word phdr_flags_;
224    Elf_Word phdr_type_;
225
226    friend class ElfBuilder;
227
228    DISALLOW_COPY_AND_ASSIGN(Section);
229  };
230
231  // Writer of .dynstr .strtab and .shstrtab sections.
232  class StringSection FINAL : public Section {
233   public:
234    StringSection(ElfBuilder<ElfTypes>* owner, const std::string& name,
235                  Elf_Word flags, Elf_Word align)
236        : Section(owner, name, SHT_STRTAB, flags, nullptr, 0, align, 0),
237          current_offset_(0) {
238    }
239
240    Elf_Word Write(const std::string& name) {
241      if (current_offset_ == 0) {
242        DCHECK(name.empty());
243      }
244      Elf_Word offset = current_offset_;
245      this->WriteFully(name.c_str(), name.length() + 1);
246      current_offset_ += name.length() + 1;
247      return offset;
248    }
249
250   private:
251    Elf_Word current_offset_;
252  };
253
254  // Writer of .dynsym and .symtab sections.
255  class SymbolSection FINAL : public Section {
256   public:
257    SymbolSection(ElfBuilder<ElfTypes>* owner, const std::string& name,
258                  Elf_Word type, Elf_Word flags, StringSection* strtab)
259        : Section(owner, name, type, flags, strtab, 0,
260                  sizeof(Elf_Off), sizeof(Elf_Sym)) {
261    }
262
263    // Buffer symbol for this section.  It will be written later.
264    // If the symbol's section is null, it will be considered absolute (SHN_ABS).
265    // (we use this in JIT to reference code which is stored outside the debug ELF file)
266    void Add(Elf_Word name, const Section* section,
267             Elf_Addr addr, bool is_relative, Elf_Word size,
268             uint8_t binding, uint8_t type, uint8_t other = 0) {
269      Elf_Sym sym = Elf_Sym();
270      sym.st_name = name;
271      sym.st_value = addr + (is_relative ? section->GetAddress() : 0);
272      sym.st_size = size;
273      sym.st_other = other;
274      sym.st_shndx = (section != nullptr ? section->GetSectionIndex()
275                                         : static_cast<Elf_Word>(SHN_ABS));
276      sym.st_info = (binding << 4) + (type & 0xf);
277      symbols_.push_back(sym);
278    }
279
280    void Write() {
281      // The symbol table always has to start with NULL symbol.
282      Elf_Sym null_symbol = Elf_Sym();
283      this->WriteFully(&null_symbol, sizeof(null_symbol));
284      this->WriteFully(symbols_.data(), symbols_.size() * sizeof(symbols_[0]));
285      symbols_.clear();
286      symbols_.shrink_to_fit();
287    }
288
289   private:
290    std::vector<Elf_Sym> symbols_;
291  };
292
293  ElfBuilder(InstructionSet isa, OutputStream* output)
294      : isa_(isa),
295        stream_(output),
296        rodata_(this, ".rodata", SHT_PROGBITS, SHF_ALLOC, nullptr, 0, kPageSize, 0),
297        text_(this, ".text", SHT_PROGBITS, SHF_ALLOC | SHF_EXECINSTR, nullptr, 0, kPageSize, 0),
298        bss_(this, ".bss", SHT_NOBITS, SHF_ALLOC, nullptr, 0, kPageSize, 0),
299        dynstr_(this, ".dynstr", SHF_ALLOC, kPageSize),
300        dynsym_(this, ".dynsym", SHT_DYNSYM, SHF_ALLOC, &dynstr_),
301        hash_(this, ".hash", SHT_HASH, SHF_ALLOC, &dynsym_, 0, sizeof(Elf_Word), sizeof(Elf_Word)),
302        dynamic_(this, ".dynamic", SHT_DYNAMIC, SHF_ALLOC, &dynstr_, 0, kPageSize, sizeof(Elf_Dyn)),
303        eh_frame_(this, ".eh_frame", SHT_PROGBITS, SHF_ALLOC, nullptr, 0, kPageSize, 0),
304        eh_frame_hdr_(this, ".eh_frame_hdr", SHT_PROGBITS, SHF_ALLOC, nullptr, 0, 4, 0),
305        strtab_(this, ".strtab", 0, 1),
306        symtab_(this, ".symtab", SHT_SYMTAB, 0, &strtab_),
307        debug_frame_(this, ".debug_frame", SHT_PROGBITS, 0, nullptr, 0, sizeof(Elf_Addr), 0),
308        debug_info_(this, ".debug_info", SHT_PROGBITS, 0, nullptr, 0, 1, 0),
309        debug_line_(this, ".debug_line", SHT_PROGBITS, 0, nullptr, 0, 1, 0),
310        shstrtab_(this, ".shstrtab", 0, 1),
311        started_(false),
312        virtual_address_(0) {
313    text_.phdr_flags_ = PF_R | PF_X;
314    bss_.phdr_flags_ = PF_R | PF_W;
315    dynamic_.phdr_flags_ = PF_R | PF_W;
316    dynamic_.phdr_type_ = PT_DYNAMIC;
317    eh_frame_hdr_.phdr_type_ = PT_GNU_EH_FRAME;
318  }
319  ~ElfBuilder() {}
320
321  InstructionSet GetIsa() { return isa_; }
322  Section* GetRoData() { return &rodata_; }
323  Section* GetText() { return &text_; }
324  Section* GetBss() { return &bss_; }
325  StringSection* GetStrTab() { return &strtab_; }
326  SymbolSection* GetSymTab() { return &symtab_; }
327  Section* GetEhFrame() { return &eh_frame_; }
328  Section* GetEhFrameHdr() { return &eh_frame_hdr_; }
329  Section* GetDebugFrame() { return &debug_frame_; }
330  Section* GetDebugInfo() { return &debug_info_; }
331  Section* GetDebugLine() { return &debug_line_; }
332
333  // Encode patch locations as LEB128 list of deltas between consecutive addresses.
334  // (exposed publicly for tests)
335  static void EncodeOatPatches(const ArrayRef<const uintptr_t>& locations,
336                               std::vector<uint8_t>* buffer) {
337    buffer->reserve(buffer->size() + locations.size() * 2);  // guess 2 bytes per ULEB128.
338    uintptr_t address = 0;  // relative to start of section.
339    for (uintptr_t location : locations) {
340      DCHECK_GE(location, address) << "Patch locations are not in sorted order";
341      EncodeUnsignedLeb128(buffer, dchecked_integral_cast<uint32_t>(location - address));
342      address = location;
343    }
344  }
345
346  void WritePatches(const char* name, const ArrayRef<const uintptr_t>& patch_locations) {
347    std::vector<uint8_t> buffer;
348    EncodeOatPatches(patch_locations, &buffer);
349    std::unique_ptr<Section> s(new Section(this, name, SHT_OAT_PATCH, 0, nullptr, 0, 1, 0));
350    s->Start();
351    s->WriteFully(buffer.data(), buffer.size());
352    s->End();
353    other_sections_.push_back(std::move(s));
354  }
355
356  void WriteSection(const char* name, const std::vector<uint8_t>* buffer) {
357    std::unique_ptr<Section> s(new Section(this, name, SHT_PROGBITS, 0, nullptr, 0, 1, 0));
358    s->Start();
359    s->WriteFully(buffer->data(), buffer->size());
360    s->End();
361    other_sections_.push_back(std::move(s));
362  }
363
364  // Reserve space for ELF header and program headers.
365  // We do not know the number of headers until later, so
366  // it is easiest to just reserve a fixed amount of space.
367  // Program headers are required for loading by the linker.
368  // It is possible to omit them for ELF files used for debugging.
369  void Start(bool write_program_headers = true) {
370    int size = sizeof(Elf_Ehdr);
371    if (write_program_headers) {
372      size += sizeof(Elf_Phdr) * kMaxProgramHeaders;
373    }
374    stream_.Seek(size, kSeekSet);
375    started_ = true;
376    virtual_address_ += size;
377    write_program_headers_ = write_program_headers;
378  }
379
380  void End() {
381    DCHECK(started_);
382
383    // Write section names and finish the section headers.
384    shstrtab_.Start();
385    shstrtab_.Write("");
386    for (auto* section : sections_) {
387      section->header_.sh_name = shstrtab_.Write(section->name_);
388      if (section->link_ != nullptr) {
389        section->header_.sh_link = section->link_->GetSectionIndex();
390      }
391    }
392    shstrtab_.End();
393
394    // Write section headers at the end of the ELF file.
395    std::vector<Elf_Shdr> shdrs;
396    shdrs.reserve(1u + sections_.size());
397    shdrs.push_back(Elf_Shdr());  // NULL at index 0.
398    for (auto* section : sections_) {
399      shdrs.push_back(section->header_);
400    }
401    Elf_Off section_headers_offset;
402    section_headers_offset = AlignFileOffset(sizeof(Elf_Off));
403    stream_.WriteFully(shdrs.data(), shdrs.size() * sizeof(shdrs[0]));
404
405    // Flush everything else before writing the program headers. This should prevent
406    // the OS from reordering writes, so that we don't end up with valid headers
407    // and partially written data if we suddenly lose power, for example.
408    stream_.Flush();
409
410    // The main ELF header.
411    Elf_Ehdr elf_header = MakeElfHeader(isa_);
412    elf_header.e_shoff = section_headers_offset;
413    elf_header.e_shnum = shdrs.size();
414    elf_header.e_shstrndx = shstrtab_.GetSectionIndex();
415
416    // Program headers (i.e. mmap instructions).
417    std::vector<Elf_Phdr> phdrs;
418    if (write_program_headers_) {
419      phdrs = MakeProgramHeaders();
420      CHECK_LE(phdrs.size(), kMaxProgramHeaders);
421      elf_header.e_phoff = sizeof(Elf_Ehdr);
422      elf_header.e_phnum = phdrs.size();
423    }
424
425    stream_.Seek(0, kSeekSet);
426    stream_.WriteFully(&elf_header, sizeof(elf_header));
427    stream_.WriteFully(phdrs.data(), phdrs.size() * sizeof(phdrs[0]));
428    stream_.Flush();
429  }
430
431  // The running program does not have access to section headers
432  // and the loader is not supposed to use them either.
433  // The dynamic sections therefore replicates some of the layout
434  // information like the address and size of .rodata and .text.
435  // It also contains other metadata like the SONAME.
436  // The .dynamic section is found using the PT_DYNAMIC program header.
437  void WriteDynamicSection(const std::string& elf_file_path) {
438    std::string soname(elf_file_path);
439    size_t directory_separator_pos = soname.rfind('/');
440    if (directory_separator_pos != std::string::npos) {
441      soname = soname.substr(directory_separator_pos + 1);
442    }
443
444    dynstr_.Start();
445    dynstr_.Write("");  // dynstr should start with empty string.
446    dynsym_.Add(dynstr_.Write("oatdata"), &rodata_, 0, true,
447                rodata_.GetSize(), STB_GLOBAL, STT_OBJECT);
448    if (text_.GetSize() != 0u) {
449      dynsym_.Add(dynstr_.Write("oatexec"), &text_, 0, true,
450                  text_.GetSize(), STB_GLOBAL, STT_OBJECT);
451      dynsym_.Add(dynstr_.Write("oatlastword"), &text_, text_.GetSize() - 4,
452                  true, 4, STB_GLOBAL, STT_OBJECT);
453    } else if (rodata_.GetSize() != 0) {
454      // rodata_ can be size 0 for dwarf_test.
455      dynsym_.Add(dynstr_.Write("oatlastword"), &rodata_, rodata_.GetSize() - 4,
456                  true, 4, STB_GLOBAL, STT_OBJECT);
457    }
458    if (bss_.finished_) {
459      dynsym_.Add(dynstr_.Write("oatbss"), &bss_,
460                  0, true, bss_.GetSize(), STB_GLOBAL, STT_OBJECT);
461      dynsym_.Add(dynstr_.Write("oatbsslastword"), &bss_,
462                  bss_.GetSize() - 4, true, 4, STB_GLOBAL, STT_OBJECT);
463    }
464    Elf_Word soname_offset = dynstr_.Write(soname);
465    dynstr_.End();
466
467    dynsym_.Start();
468    dynsym_.Write();
469    dynsym_.End();
470
471    // We do not really need a hash-table since there is so few entries.
472    // However, the hash-table is the only way the linker can actually
473    // determine the number of symbols in .dynsym so it is required.
474    hash_.Start();
475    int count = dynsym_.GetSize() / sizeof(Elf_Sym);  // Includes NULL.
476    std::vector<Elf_Word> hash;
477    hash.push_back(1);  // Number of buckets.
478    hash.push_back(count);  // Number of chains.
479    // Buckets.  Having just one makes it linear search.
480    hash.push_back(1);  // Point to first non-NULL symbol.
481    // Chains.  This creates linked list of symbols.
482    hash.push_back(0);  // Dummy entry for the NULL symbol.
483    for (int i = 1; i < count - 1; i++) {
484      hash.push_back(i + 1);  // Each symbol points to the next one.
485    }
486    hash.push_back(0);  // Last symbol terminates the chain.
487    hash_.WriteFully(hash.data(), hash.size() * sizeof(hash[0]));
488    hash_.End();
489
490    dynamic_.Start();
491    Elf_Dyn dyns[] = {
492      { DT_HASH, { hash_.GetAddress() } },
493      { DT_STRTAB, { dynstr_.GetAddress() } },
494      { DT_SYMTAB, { dynsym_.GetAddress() } },
495      { DT_SYMENT, { sizeof(Elf_Sym) } },
496      { DT_STRSZ, { dynstr_.GetSize() } },
497      { DT_SONAME, { soname_offset } },
498      { DT_NULL, { 0 } },
499    };
500    dynamic_.WriteFully(&dyns, sizeof(dyns));
501    dynamic_.End();
502  }
503
504  // Returns true if all writes and seeks on the output stream succeeded.
505  bool Good() {
506    return stream_.Good();
507  }
508
509  // Returns the builder's internal stream.
510  OutputStream* GetStream() {
511    return &stream_;
512  }
513
514  off_t AlignFileOffset(size_t alignment) {
515     return stream_.Seek(RoundUp(stream_.Seek(0, kSeekCurrent), alignment), kSeekSet);
516  }
517
518  Elf_Addr AlignVirtualAddress(size_t alignment) {
519     return virtual_address_ = RoundUp(virtual_address_, alignment);
520  }
521
522 private:
523  static Elf_Ehdr MakeElfHeader(InstructionSet isa) {
524    Elf_Ehdr elf_header = Elf_Ehdr();
525    switch (isa) {
526      case kArm:
527        // Fall through.
528      case kThumb2: {
529        elf_header.e_machine = EM_ARM;
530        elf_header.e_flags = EF_ARM_EABI_VER5;
531        break;
532      }
533      case kArm64: {
534        elf_header.e_machine = EM_AARCH64;
535        elf_header.e_flags = 0;
536        break;
537      }
538      case kX86: {
539        elf_header.e_machine = EM_386;
540        elf_header.e_flags = 0;
541        break;
542      }
543      case kX86_64: {
544        elf_header.e_machine = EM_X86_64;
545        elf_header.e_flags = 0;
546        break;
547      }
548      case kMips: {
549        elf_header.e_machine = EM_MIPS;
550        elf_header.e_flags = (EF_MIPS_NOREORDER |
551                               EF_MIPS_PIC       |
552                               EF_MIPS_CPIC      |
553                               EF_MIPS_ABI_O32   |
554                               EF_MIPS_ARCH_32R2);
555        break;
556      }
557      case kMips64: {
558        elf_header.e_machine = EM_MIPS;
559        elf_header.e_flags = (EF_MIPS_NOREORDER |
560                               EF_MIPS_PIC       |
561                               EF_MIPS_CPIC      |
562                               EF_MIPS_ARCH_64R6);
563        break;
564      }
565      case kNone: {
566        LOG(FATAL) << "No instruction set";
567        break;
568      }
569      default: {
570        LOG(FATAL) << "Unknown instruction set " << isa;
571      }
572    }
573
574    elf_header.e_ident[EI_MAG0]       = ELFMAG0;
575    elf_header.e_ident[EI_MAG1]       = ELFMAG1;
576    elf_header.e_ident[EI_MAG2]       = ELFMAG2;
577    elf_header.e_ident[EI_MAG3]       = ELFMAG3;
578    elf_header.e_ident[EI_CLASS]      = (sizeof(Elf_Addr) == sizeof(Elf32_Addr))
579                                         ? ELFCLASS32 : ELFCLASS64;;
580    elf_header.e_ident[EI_DATA]       = ELFDATA2LSB;
581    elf_header.e_ident[EI_VERSION]    = EV_CURRENT;
582    elf_header.e_ident[EI_OSABI]      = ELFOSABI_LINUX;
583    elf_header.e_ident[EI_ABIVERSION] = 0;
584    elf_header.e_type = ET_DYN;
585    elf_header.e_version = 1;
586    elf_header.e_entry = 0;
587    elf_header.e_ehsize = sizeof(Elf_Ehdr);
588    elf_header.e_phentsize = sizeof(Elf_Phdr);
589    elf_header.e_shentsize = sizeof(Elf_Shdr);
590    elf_header.e_phoff = sizeof(Elf_Ehdr);
591    return elf_header;
592  }
593
594  // Create program headers based on written sections.
595  std::vector<Elf_Phdr> MakeProgramHeaders() {
596    CHECK(!sections_.empty());
597    std::vector<Elf_Phdr> phdrs;
598    {
599      // The program headers must start with PT_PHDR which is used in
600      // loaded process to determine the number of program headers.
601      Elf_Phdr phdr = Elf_Phdr();
602      phdr.p_type    = PT_PHDR;
603      phdr.p_flags   = PF_R;
604      phdr.p_offset  = phdr.p_vaddr = phdr.p_paddr = sizeof(Elf_Ehdr);
605      phdr.p_filesz  = phdr.p_memsz = 0;  // We need to fill this later.
606      phdr.p_align   = sizeof(Elf_Off);
607      phdrs.push_back(phdr);
608      // Tell the linker to mmap the start of file to memory.
609      Elf_Phdr load = Elf_Phdr();
610      load.p_type    = PT_LOAD;
611      load.p_flags   = PF_R;
612      load.p_offset  = load.p_vaddr = load.p_paddr = 0;
613      load.p_filesz  = load.p_memsz = sections_[0]->header_.sh_offset;
614      load.p_align   = kPageSize;
615      phdrs.push_back(load);
616    }
617    // Create program headers for sections.
618    for (auto* section : sections_) {
619      const Elf_Shdr& shdr = section->header_;
620      if ((shdr.sh_flags & SHF_ALLOC) != 0 && shdr.sh_size != 0) {
621        // PT_LOAD tells the linker to mmap part of the file.
622        // The linker can only mmap page-aligned sections.
623        // Single PT_LOAD may contain several ELF sections.
624        Elf_Phdr& prev = phdrs.back();
625        Elf_Phdr load = Elf_Phdr();
626        load.p_type   = PT_LOAD;
627        load.p_flags  = section->phdr_flags_;
628        load.p_offset = shdr.sh_offset;
629        load.p_vaddr  = load.p_paddr = shdr.sh_addr;
630        load.p_filesz = (shdr.sh_type != SHT_NOBITS ? shdr.sh_size : 0u);
631        load.p_memsz  = shdr.sh_size;
632        load.p_align  = shdr.sh_addralign;
633        if (prev.p_type == load.p_type &&
634            prev.p_flags == load.p_flags &&
635            prev.p_filesz == prev.p_memsz &&  // Do not merge .bss
636            load.p_filesz == load.p_memsz) {  // Do not merge .bss
637          // Merge this PT_LOAD with the previous one.
638          Elf_Word size = shdr.sh_offset + shdr.sh_size - prev.p_offset;
639          prev.p_filesz = size;
640          prev.p_memsz  = size;
641        } else {
642          // If we are adding new load, it must be aligned.
643          CHECK_EQ(shdr.sh_addralign, (Elf_Word)kPageSize);
644          phdrs.push_back(load);
645        }
646      }
647    }
648    for (auto* section : sections_) {
649      const Elf_Shdr& shdr = section->header_;
650      if ((shdr.sh_flags & SHF_ALLOC) != 0 && shdr.sh_size != 0) {
651        // Other PT_* types allow the program to locate interesting
652        // parts of memory at runtime. They must overlap with PT_LOAD.
653        if (section->phdr_type_ != 0) {
654          Elf_Phdr phdr = Elf_Phdr();
655          phdr.p_type   = section->phdr_type_;
656          phdr.p_flags  = section->phdr_flags_;
657          phdr.p_offset = shdr.sh_offset;
658          phdr.p_vaddr  = phdr.p_paddr = shdr.sh_addr;
659          phdr.p_filesz = phdr.p_memsz = shdr.sh_size;
660          phdr.p_align  = shdr.sh_addralign;
661          phdrs.push_back(phdr);
662        }
663      }
664    }
665    // Set the size of the initial PT_PHDR.
666    CHECK_EQ(phdrs[0].p_type, (Elf_Word)PT_PHDR);
667    phdrs[0].p_filesz = phdrs[0].p_memsz = phdrs.size() * sizeof(Elf_Phdr);
668
669    return phdrs;
670  }
671
672  InstructionSet isa_;
673
674  ErrorDelayingOutputStream stream_;
675
676  Section rodata_;
677  Section text_;
678  Section bss_;
679  StringSection dynstr_;
680  SymbolSection dynsym_;
681  Section hash_;
682  Section dynamic_;
683  Section eh_frame_;
684  Section eh_frame_hdr_;
685  StringSection strtab_;
686  SymbolSection symtab_;
687  Section debug_frame_;
688  Section debug_info_;
689  Section debug_line_;
690  StringSection shstrtab_;
691  std::vector<std::unique_ptr<Section>> other_sections_;
692
693  // List of used section in the order in which they were written.
694  std::vector<Section*> sections_;
695
696  bool started_;
697
698  // Used for allocation of virtual address space.
699  Elf_Addr virtual_address_;
700
701  size_t write_program_headers_;
702
703  DISALLOW_COPY_AND_ASSIGN(ElfBuilder);
704};
705
706}  // namespace art
707
708#endif  // ART_COMPILER_ELF_BUILDER_H_
709