elf_builder.h revision 6065402316da2b51eed5fc34cffbd991766bd408
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, 90 const std::string& name, 91 Elf_Word type, 92 Elf_Word flags, 93 const Section* link, 94 Elf_Word info, 95 Elf_Word align, 96 Elf_Word entsize) 97 : OutputStream(name), 98 owner_(owner), 99 header_(), 100 section_index_(0), 101 name_(name), 102 link_(link), 103 started_(false), 104 finished_(false), 105 phdr_flags_(PF_R), 106 phdr_type_(0) { 107 DCHECK_GE(align, 1u); 108 header_.sh_type = type; 109 header_.sh_flags = flags; 110 header_.sh_info = info; 111 header_.sh_addralign = align; 112 header_.sh_entsize = entsize; 113 } 114 115 // Start writing of this section. 116 void Start() { 117 CHECK(!started_); 118 CHECK(!finished_); 119 started_ = true; 120 auto& sections = owner_->sections_; 121 // Check that the previous section is complete. 122 CHECK(sections.empty() || sections.back()->finished_); 123 // The first ELF section index is 1. Index 0 is reserved for NULL. 124 section_index_ = sections.size() + 1; 125 // Page-align if we switch between allocated and non-allocated sections, 126 // or if we change the type of allocation (e.g. executable vs non-executable). 127 if (!sections.empty()) { 128 if (header_.sh_flags != sections.back()->header_.sh_flags) { 129 header_.sh_addralign = kPageSize; 130 } 131 } 132 // Align file position. 133 if (header_.sh_type != SHT_NOBITS) { 134 header_.sh_offset = owner_->AlignFileOffset(header_.sh_addralign); 135 } else { 136 header_.sh_offset = 0; 137 } 138 // Align virtual memory address. 139 if ((header_.sh_flags & SHF_ALLOC) != 0) { 140 header_.sh_addr = owner_->AlignVirtualAddress(header_.sh_addralign); 141 } else { 142 header_.sh_addr = 0; 143 } 144 // Push this section on the list of written sections. 145 sections.push_back(this); 146 } 147 148 // Finish writing of this section. 149 void End() { 150 CHECK(started_); 151 CHECK(!finished_); 152 finished_ = true; 153 if (header_.sh_type == SHT_NOBITS) { 154 CHECK_GT(header_.sh_size, 0u); 155 } else { 156 // Use the current file position to determine section size. 157 off_t file_offset = owner_->stream_.Seek(0, kSeekCurrent); 158 CHECK_GE(file_offset, (off_t)header_.sh_offset); 159 header_.sh_size = file_offset - header_.sh_offset; 160 } 161 if ((header_.sh_flags & SHF_ALLOC) != 0) { 162 owner_->virtual_address_ += header_.sh_size; 163 } 164 } 165 166 // Returns true if the section was written to disk. 167 // (Used to check whether we have .text when writing JIT debug info) 168 bool Exists() const { 169 return finished_; 170 } 171 172 // Get the location of this section in virtual memory. 173 Elf_Addr GetAddress() const { 174 CHECK(started_); 175 return header_.sh_addr; 176 } 177 178 // Returns the size of the content of this section. 179 Elf_Word GetSize() const { 180 if (finished_) { 181 return header_.sh_size; 182 } else { 183 CHECK(started_); 184 CHECK_NE(header_.sh_type, (Elf_Word)SHT_NOBITS); 185 return owner_->stream_.Seek(0, kSeekCurrent) - header_.sh_offset; 186 } 187 } 188 189 // Write this section as "NOBITS" section. (used for the .bss section) 190 // This means that the ELF file does not contain the initial data for this section 191 // and it will be zero-initialized when the ELF file is loaded in the running program. 192 void WriteNoBitsSection(Elf_Word size) { 193 DCHECK_NE(header_.sh_flags & SHF_ALLOC, 0u); 194 header_.sh_type = SHT_NOBITS; 195 Start(); 196 header_.sh_size = size; 197 End(); 198 } 199 200 // This function always succeeds to simplify code. 201 // Use builder's Good() to check the actual status. 202 bool WriteFully(const void* buffer, size_t byte_count) OVERRIDE { 203 CHECK(started_); 204 CHECK(!finished_); 205 return owner_->stream_.WriteFully(buffer, byte_count); 206 } 207 208 // This function always succeeds to simplify code. 209 // Use builder's Good() to check the actual status. 210 off_t Seek(off_t offset, Whence whence) OVERRIDE { 211 // Forward the seek as-is and trust the caller to use it reasonably. 212 return owner_->stream_.Seek(offset, whence); 213 } 214 215 // This function flushes the output and returns whether it succeeded. 216 // If there was a previous failure, this does nothing and returns false, i.e. failed. 217 bool Flush() OVERRIDE { 218 return owner_->stream_.Flush(); 219 } 220 221 Elf_Word GetSectionIndex() const { 222 DCHECK(started_); 223 DCHECK_NE(section_index_, 0u); 224 return section_index_; 225 } 226 227 private: 228 ElfBuilder<ElfTypes>* owner_; 229 Elf_Shdr header_; 230 Elf_Word section_index_; 231 const std::string name_; 232 const Section* const link_; 233 bool started_; 234 bool finished_; 235 Elf_Word phdr_flags_; 236 Elf_Word phdr_type_; 237 238 friend class ElfBuilder; 239 240 DISALLOW_COPY_AND_ASSIGN(Section); 241 }; 242 243 class CachedSection : public Section { 244 public: 245 CachedSection(ElfBuilder<ElfTypes>* owner, 246 const std::string& name, 247 Elf_Word type, 248 Elf_Word flags, 249 const Section* link, 250 Elf_Word info, 251 Elf_Word align, 252 Elf_Word entsize) 253 : Section(owner, name, type, flags, link, info, align, entsize), cache_() { } 254 255 Elf_Word Add(const void* data, size_t length) { 256 Elf_Word offset = cache_.size(); 257 const uint8_t* d = reinterpret_cast<const uint8_t*>(data); 258 cache_.insert(cache_.end(), d, d + length); 259 return offset; 260 } 261 262 Elf_Word GetCacheSize() { 263 return cache_.size(); 264 } 265 266 void Write() { 267 this->WriteFully(cache_.data(), cache_.size()); 268 cache_.clear(); 269 cache_.shrink_to_fit(); 270 } 271 272 void WriteCachedSection() { 273 this->Start(); 274 Write(); 275 this->End(); 276 } 277 278 private: 279 std::vector<uint8_t> cache_; 280 }; 281 282 // Writer of .dynstr section. 283 class CachedStringSection FINAL : public CachedSection { 284 public: 285 CachedStringSection(ElfBuilder<ElfTypes>* owner, 286 const std::string& name, 287 Elf_Word flags, 288 Elf_Word align) 289 : CachedSection(owner, 290 name, 291 SHT_STRTAB, 292 flags, 293 /* link */ nullptr, 294 /* info */ 0, 295 align, 296 /* entsize */ 0) { } 297 298 Elf_Word Add(const std::string& name) { 299 if (CachedSection::GetCacheSize() == 0u) { 300 DCHECK(name.empty()); 301 } 302 return CachedSection::Add(name.c_str(), name.length() + 1); 303 } 304 }; 305 306 // Writer of .strtab and .shstrtab sections. 307 class StringSection FINAL : public Section { 308 public: 309 StringSection(ElfBuilder<ElfTypes>* owner, 310 const std::string& name, 311 Elf_Word flags, 312 Elf_Word align) 313 : Section(owner, 314 name, 315 SHT_STRTAB, 316 flags, 317 /* link */ nullptr, 318 /* info */ 0, 319 align, 320 /* entsize */ 0), 321 current_offset_(0) { 322 } 323 324 Elf_Word Write(const std::string& name) { 325 if (current_offset_ == 0) { 326 DCHECK(name.empty()); 327 } 328 Elf_Word offset = current_offset_; 329 this->WriteFully(name.c_str(), name.length() + 1); 330 current_offset_ += name.length() + 1; 331 return offset; 332 } 333 334 private: 335 Elf_Word current_offset_; 336 }; 337 338 // Writer of .dynsym and .symtab sections. 339 class SymbolSection FINAL : public CachedSection { 340 public: 341 SymbolSection(ElfBuilder<ElfTypes>* owner, 342 const std::string& name, 343 Elf_Word type, 344 Elf_Word flags, 345 Section* strtab) 346 : CachedSection(owner, 347 name, 348 type, 349 flags, 350 strtab, 351 /* info */ 0, 352 sizeof(Elf_Off), 353 sizeof(Elf_Sym)) { 354 // The symbol table always has to start with NULL symbol. 355 Elf_Sym null_symbol = Elf_Sym(); 356 CachedSection::Add(&null_symbol, sizeof(null_symbol)); 357 } 358 359 // Buffer symbol for this section. It will be written later. 360 // If the symbol's section is null, it will be considered absolute (SHN_ABS). 361 // (we use this in JIT to reference code which is stored outside the debug ELF file) 362 void Add(Elf_Word name, 363 const Section* section, 364 Elf_Addr addr, 365 bool is_relative, 366 Elf_Word size, 367 uint8_t binding, 368 uint8_t type, 369 uint8_t other = 0) { 370 DCHECK(section != nullptr || !is_relative); 371 Elf_Addr abs_addr = addr + (is_relative ? section->GetAddress() : 0); 372 Elf_Word section_index = 373 (section != nullptr) ? section->GetSectionIndex() : static_cast<Elf_Word>(SHN_ABS); 374 Add(name, section_index, abs_addr, size, binding, type, other); 375 } 376 377 void Add(Elf_Word name, 378 Elf_Word section_index, 379 Elf_Addr addr, 380 Elf_Word size, 381 uint8_t binding, 382 uint8_t type, 383 uint8_t other = 0) { 384 Elf_Sym sym = Elf_Sym(); 385 sym.st_name = name; 386 sym.st_value = addr; 387 sym.st_size = size; 388 sym.st_other = other; 389 sym.st_shndx = section_index; 390 sym.st_info = (binding << 4) + (type & 0xf); 391 CachedSection::Add(&sym, sizeof(sym)); 392 } 393 }; 394 395 ElfBuilder(InstructionSet isa, OutputStream* output) 396 : isa_(isa), 397 stream_(output), 398 rodata_(this, ".rodata", SHT_PROGBITS, SHF_ALLOC, nullptr, 0, kPageSize, 0), 399 text_(this, ".text", SHT_PROGBITS, SHF_ALLOC | SHF_EXECINSTR, nullptr, 0, kPageSize, 0), 400 bss_(this, ".bss", SHT_NOBITS, SHF_ALLOC, nullptr, 0, kPageSize, 0), 401 dynstr_(this, ".dynstr", SHF_ALLOC, kPageSize), 402 dynsym_(this, ".dynsym", SHT_DYNSYM, SHF_ALLOC, &dynstr_), 403 hash_(this, ".hash", SHT_HASH, SHF_ALLOC, &dynsym_, 0, sizeof(Elf_Word), sizeof(Elf_Word)), 404 dynamic_(this, ".dynamic", SHT_DYNAMIC, SHF_ALLOC, &dynstr_, 0, kPageSize, sizeof(Elf_Dyn)), 405 eh_frame_(this, ".eh_frame", SHT_PROGBITS, SHF_ALLOC, nullptr, 0, kPageSize, 0), 406 eh_frame_hdr_(this, ".eh_frame_hdr", SHT_PROGBITS, SHF_ALLOC, nullptr, 0, 4, 0), 407 strtab_(this, ".strtab", 0, 1), 408 symtab_(this, ".symtab", SHT_SYMTAB, 0, &strtab_), 409 debug_frame_(this, ".debug_frame", SHT_PROGBITS, 0, nullptr, 0, sizeof(Elf_Addr), 0), 410 debug_info_(this, ".debug_info", SHT_PROGBITS, 0, nullptr, 0, 1, 0), 411 debug_line_(this, ".debug_line", SHT_PROGBITS, 0, nullptr, 0, 1, 0), 412 shstrtab_(this, ".shstrtab", 0, 1), 413 started_(false), 414 write_program_headers_(false), 415 loaded_size_(0u), 416 virtual_address_(0) { 417 text_.phdr_flags_ = PF_R | PF_X; 418 bss_.phdr_flags_ = PF_R | PF_W; 419 dynamic_.phdr_flags_ = PF_R | PF_W; 420 dynamic_.phdr_type_ = PT_DYNAMIC; 421 eh_frame_hdr_.phdr_type_ = PT_GNU_EH_FRAME; 422 } 423 ~ElfBuilder() {} 424 425 InstructionSet GetIsa() { return isa_; } 426 Section* GetRoData() { return &rodata_; } 427 Section* GetText() { return &text_; } 428 Section* GetBss() { return &bss_; } 429 StringSection* GetStrTab() { return &strtab_; } 430 SymbolSection* GetSymTab() { return &symtab_; } 431 Section* GetEhFrame() { return &eh_frame_; } 432 Section* GetEhFrameHdr() { return &eh_frame_hdr_; } 433 Section* GetDebugFrame() { return &debug_frame_; } 434 Section* GetDebugInfo() { return &debug_info_; } 435 Section* GetDebugLine() { return &debug_line_; } 436 437 // Encode patch locations as LEB128 list of deltas between consecutive addresses. 438 // (exposed publicly for tests) 439 static void EncodeOatPatches(const ArrayRef<const uintptr_t>& locations, 440 std::vector<uint8_t>* buffer) { 441 buffer->reserve(buffer->size() + locations.size() * 2); // guess 2 bytes per ULEB128. 442 uintptr_t address = 0; // relative to start of section. 443 for (uintptr_t location : locations) { 444 DCHECK_GE(location, address) << "Patch locations are not in sorted order"; 445 EncodeUnsignedLeb128(buffer, dchecked_integral_cast<uint32_t>(location - address)); 446 address = location; 447 } 448 } 449 450 void WritePatches(const char* name, const ArrayRef<const uintptr_t>& patch_locations) { 451 std::vector<uint8_t> buffer; 452 EncodeOatPatches(patch_locations, &buffer); 453 std::unique_ptr<Section> s(new Section(this, name, SHT_OAT_PATCH, 0, nullptr, 0, 1, 0)); 454 s->Start(); 455 s->WriteFully(buffer.data(), buffer.size()); 456 s->End(); 457 other_sections_.push_back(std::move(s)); 458 } 459 460 void WriteSection(const char* name, const std::vector<uint8_t>* buffer) { 461 std::unique_ptr<Section> s(new Section(this, name, SHT_PROGBITS, 0, nullptr, 0, 1, 0)); 462 s->Start(); 463 s->WriteFully(buffer->data(), buffer->size()); 464 s->End(); 465 other_sections_.push_back(std::move(s)); 466 } 467 468 // Reserve space for ELF header and program headers. 469 // We do not know the number of headers until later, so 470 // it is easiest to just reserve a fixed amount of space. 471 // Program headers are required for loading by the linker. 472 // It is possible to omit them for ELF files used for debugging. 473 void Start(bool write_program_headers = true) { 474 int size = sizeof(Elf_Ehdr); 475 if (write_program_headers) { 476 size += sizeof(Elf_Phdr) * kMaxProgramHeaders; 477 } 478 stream_.Seek(size, kSeekSet); 479 started_ = true; 480 virtual_address_ += size; 481 write_program_headers_ = write_program_headers; 482 } 483 484 void End() { 485 DCHECK(started_); 486 487 // Note: loaded_size_ == 0 for tests that don't write .rodata, .text, .bss, 488 // .dynstr, dynsym, .hash and .dynamic. These tests should not read loaded_size_. 489 // TODO: Either refactor the .eh_frame creation so that it counts towards loaded_size_, 490 // or remove all support for .eh_frame. (The currently unused .eh_frame counts towards 491 // the virtual_address_ but we don't consider it for loaded_size_.) 492 CHECK(loaded_size_ == 0 || loaded_size_ == RoundUp(virtual_address_, kPageSize)) 493 << loaded_size_ << " " << virtual_address_; 494 495 // Write section names and finish the section headers. 496 shstrtab_.Start(); 497 shstrtab_.Write(""); 498 for (auto* section : sections_) { 499 section->header_.sh_name = shstrtab_.Write(section->name_); 500 if (section->link_ != nullptr) { 501 section->header_.sh_link = section->link_->GetSectionIndex(); 502 } 503 } 504 shstrtab_.End(); 505 506 // Write section headers at the end of the ELF file. 507 std::vector<Elf_Shdr> shdrs; 508 shdrs.reserve(1u + sections_.size()); 509 shdrs.push_back(Elf_Shdr()); // NULL at index 0. 510 for (auto* section : sections_) { 511 shdrs.push_back(section->header_); 512 } 513 Elf_Off section_headers_offset; 514 section_headers_offset = AlignFileOffset(sizeof(Elf_Off)); 515 stream_.WriteFully(shdrs.data(), shdrs.size() * sizeof(shdrs[0])); 516 517 // Flush everything else before writing the program headers. This should prevent 518 // the OS from reordering writes, so that we don't end up with valid headers 519 // and partially written data if we suddenly lose power, for example. 520 stream_.Flush(); 521 522 // The main ELF header. 523 Elf_Ehdr elf_header = MakeElfHeader(isa_); 524 elf_header.e_shoff = section_headers_offset; 525 elf_header.e_shnum = shdrs.size(); 526 elf_header.e_shstrndx = shstrtab_.GetSectionIndex(); 527 528 // Program headers (i.e. mmap instructions). 529 std::vector<Elf_Phdr> phdrs; 530 if (write_program_headers_) { 531 phdrs = MakeProgramHeaders(); 532 CHECK_LE(phdrs.size(), kMaxProgramHeaders); 533 elf_header.e_phoff = sizeof(Elf_Ehdr); 534 elf_header.e_phnum = phdrs.size(); 535 } 536 537 stream_.Seek(0, kSeekSet); 538 stream_.WriteFully(&elf_header, sizeof(elf_header)); 539 stream_.WriteFully(phdrs.data(), phdrs.size() * sizeof(phdrs[0])); 540 stream_.Flush(); 541 } 542 543 // The running program does not have access to section headers 544 // and the loader is not supposed to use them either. 545 // The dynamic sections therefore replicates some of the layout 546 // information like the address and size of .rodata and .text. 547 // It also contains other metadata like the SONAME. 548 // The .dynamic section is found using the PT_DYNAMIC program header. 549 void PrepareDynamicSection(const std::string& elf_file_path, 550 Elf_Word rodata_size, 551 Elf_Word text_size, 552 Elf_Word bss_size) { 553 std::string soname(elf_file_path); 554 size_t directory_separator_pos = soname.rfind('/'); 555 if (directory_separator_pos != std::string::npos) { 556 soname = soname.substr(directory_separator_pos + 1); 557 } 558 559 // Calculate addresses of .text, .bss and .dynstr. 560 DCHECK_EQ(rodata_.header_.sh_addralign, static_cast<Elf_Word>(kPageSize)); 561 DCHECK_EQ(text_.header_.sh_addralign, static_cast<Elf_Word>(kPageSize)); 562 DCHECK_EQ(bss_.header_.sh_addralign, static_cast<Elf_Word>(kPageSize)); 563 DCHECK_EQ(dynstr_.header_.sh_addralign, static_cast<Elf_Word>(kPageSize)); 564 Elf_Word rodata_address = rodata_.GetAddress(); 565 Elf_Word text_address = RoundUp(rodata_address + rodata_size, kPageSize); 566 Elf_Word bss_address = RoundUp(text_address + text_size, kPageSize); 567 Elf_Word dynstr_address = RoundUp(bss_address + bss_size, kPageSize); 568 569 // Cache .dynstr, .dynsym and .hash data. 570 dynstr_.Add(""); // dynstr should start with empty string. 571 Elf_Word rodata_index = rodata_.GetSectionIndex(); 572 Elf_Word oatdata = dynstr_.Add("oatdata"); 573 dynsym_.Add(oatdata, rodata_index, rodata_address, rodata_size, STB_GLOBAL, STT_OBJECT); 574 if (text_size != 0u) { 575 Elf_Word text_index = rodata_index + 1u; 576 Elf_Word oatexec = dynstr_.Add("oatexec"); 577 dynsym_.Add(oatexec, text_index, text_address, text_size, STB_GLOBAL, STT_OBJECT); 578 Elf_Word oatlastword = dynstr_.Add("oatlastword"); 579 Elf_Word oatlastword_address = text_address + text_size - 4; 580 dynsym_.Add(oatlastword, text_index, oatlastword_address, 4, STB_GLOBAL, STT_OBJECT); 581 } else if (rodata_size != 0) { 582 // rodata_ can be size 0 for dwarf_test. 583 Elf_Word oatlastword = dynstr_.Add("oatlastword"); 584 Elf_Word oatlastword_address = rodata_address + rodata_size - 4; 585 dynsym_.Add(oatlastword, rodata_index, oatlastword_address, 4, STB_GLOBAL, STT_OBJECT); 586 } 587 if (bss_size != 0u) { 588 Elf_Word bss_index = rodata_index + 1u + (text_size != 0 ? 1u : 0u); 589 Elf_Word oatbss = dynstr_.Add("oatbss"); 590 dynsym_.Add(oatbss, bss_index, bss_address, bss_size, STB_GLOBAL, STT_OBJECT); 591 Elf_Word oatbsslastword = dynstr_.Add("oatbsslastword"); 592 Elf_Word bsslastword_address = bss_address + bss_size - 4; 593 dynsym_.Add(oatbsslastword, bss_index, bsslastword_address, 4, STB_GLOBAL, STT_OBJECT); 594 } 595 Elf_Word soname_offset = dynstr_.Add(soname); 596 597 // We do not really need a hash-table since there is so few entries. 598 // However, the hash-table is the only way the linker can actually 599 // determine the number of symbols in .dynsym so it is required. 600 int count = dynsym_.GetCacheSize() / sizeof(Elf_Sym); // Includes NULL. 601 std::vector<Elf_Word> hash; 602 hash.push_back(1); // Number of buckets. 603 hash.push_back(count); // Number of chains. 604 // Buckets. Having just one makes it linear search. 605 hash.push_back(1); // Point to first non-NULL symbol. 606 // Chains. This creates linked list of symbols. 607 hash.push_back(0); // Dummy entry for the NULL symbol. 608 for (int i = 1; i < count - 1; i++) { 609 hash.push_back(i + 1); // Each symbol points to the next one. 610 } 611 hash.push_back(0); // Last symbol terminates the chain. 612 hash_.Add(hash.data(), hash.size() * sizeof(hash[0])); 613 614 // Calculate addresses of .dynsym, .hash and .dynamic. 615 DCHECK_EQ(dynstr_.header_.sh_flags, dynsym_.header_.sh_flags); 616 DCHECK_EQ(dynsym_.header_.sh_flags, hash_.header_.sh_flags); 617 Elf_Word dynsym_address = 618 RoundUp(dynstr_address + dynstr_.GetCacheSize(), dynsym_.header_.sh_addralign); 619 Elf_Word hash_address = 620 RoundUp(dynsym_address + dynsym_.GetCacheSize(), hash_.header_.sh_addralign); 621 DCHECK_EQ(dynamic_.header_.sh_addralign, static_cast<Elf_Word>(kPageSize)); 622 Elf_Word dynamic_address = RoundUp(hash_address + dynsym_.GetCacheSize(), kPageSize); 623 624 Elf_Dyn dyns[] = { 625 { DT_HASH, { hash_address } }, 626 { DT_STRTAB, { dynstr_address } }, 627 { DT_SYMTAB, { dynsym_address } }, 628 { DT_SYMENT, { sizeof(Elf_Sym) } }, 629 { DT_STRSZ, { dynstr_.GetCacheSize() } }, 630 { DT_SONAME, { soname_offset } }, 631 { DT_NULL, { 0 } }, 632 }; 633 dynamic_.Add(&dyns, sizeof(dyns)); 634 635 loaded_size_ = RoundUp(dynamic_address + dynamic_.GetCacheSize(), kPageSize); 636 } 637 638 void WriteDynamicSection() { 639 dynstr_.WriteCachedSection(); 640 dynsym_.WriteCachedSection(); 641 hash_.WriteCachedSection(); 642 dynamic_.WriteCachedSection(); 643 644 CHECK_EQ(loaded_size_, RoundUp(dynamic_.GetAddress() + dynamic_.GetSize(), kPageSize)); 645 } 646 647 Elf_Word GetLoadedSize() { 648 CHECK_NE(loaded_size_, 0u); 649 return loaded_size_; 650 } 651 652 // Returns true if all writes and seeks on the output stream succeeded. 653 bool Good() { 654 return stream_.Good(); 655 } 656 657 // Returns the builder's internal stream. 658 OutputStream* GetStream() { 659 return &stream_; 660 } 661 662 off_t AlignFileOffset(size_t alignment) { 663 return stream_.Seek(RoundUp(stream_.Seek(0, kSeekCurrent), alignment), kSeekSet); 664 } 665 666 Elf_Addr AlignVirtualAddress(size_t alignment) { 667 return virtual_address_ = RoundUp(virtual_address_, alignment); 668 } 669 670 private: 671 static Elf_Ehdr MakeElfHeader(InstructionSet isa) { 672 Elf_Ehdr elf_header = Elf_Ehdr(); 673 switch (isa) { 674 case kArm: 675 // Fall through. 676 case kThumb2: { 677 elf_header.e_machine = EM_ARM; 678 elf_header.e_flags = EF_ARM_EABI_VER5; 679 break; 680 } 681 case kArm64: { 682 elf_header.e_machine = EM_AARCH64; 683 elf_header.e_flags = 0; 684 break; 685 } 686 case kX86: { 687 elf_header.e_machine = EM_386; 688 elf_header.e_flags = 0; 689 break; 690 } 691 case kX86_64: { 692 elf_header.e_machine = EM_X86_64; 693 elf_header.e_flags = 0; 694 break; 695 } 696 case kMips: { 697 elf_header.e_machine = EM_MIPS; 698 elf_header.e_flags = (EF_MIPS_NOREORDER | 699 EF_MIPS_PIC | 700 EF_MIPS_CPIC | 701 EF_MIPS_ABI_O32 | 702 EF_MIPS_ARCH_32R2); 703 break; 704 } 705 case kMips64: { 706 elf_header.e_machine = EM_MIPS; 707 elf_header.e_flags = (EF_MIPS_NOREORDER | 708 EF_MIPS_PIC | 709 EF_MIPS_CPIC | 710 EF_MIPS_ARCH_64R6); 711 break; 712 } 713 case kNone: { 714 LOG(FATAL) << "No instruction set"; 715 break; 716 } 717 default: { 718 LOG(FATAL) << "Unknown instruction set " << isa; 719 } 720 } 721 722 elf_header.e_ident[EI_MAG0] = ELFMAG0; 723 elf_header.e_ident[EI_MAG1] = ELFMAG1; 724 elf_header.e_ident[EI_MAG2] = ELFMAG2; 725 elf_header.e_ident[EI_MAG3] = ELFMAG3; 726 elf_header.e_ident[EI_CLASS] = (sizeof(Elf_Addr) == sizeof(Elf32_Addr)) 727 ? ELFCLASS32 : ELFCLASS64;; 728 elf_header.e_ident[EI_DATA] = ELFDATA2LSB; 729 elf_header.e_ident[EI_VERSION] = EV_CURRENT; 730 elf_header.e_ident[EI_OSABI] = ELFOSABI_LINUX; 731 elf_header.e_ident[EI_ABIVERSION] = 0; 732 elf_header.e_type = ET_DYN; 733 elf_header.e_version = 1; 734 elf_header.e_entry = 0; 735 elf_header.e_ehsize = sizeof(Elf_Ehdr); 736 elf_header.e_phentsize = sizeof(Elf_Phdr); 737 elf_header.e_shentsize = sizeof(Elf_Shdr); 738 elf_header.e_phoff = sizeof(Elf_Ehdr); 739 return elf_header; 740 } 741 742 // Create program headers based on written sections. 743 std::vector<Elf_Phdr> MakeProgramHeaders() { 744 CHECK(!sections_.empty()); 745 std::vector<Elf_Phdr> phdrs; 746 { 747 // The program headers must start with PT_PHDR which is used in 748 // loaded process to determine the number of program headers. 749 Elf_Phdr phdr = Elf_Phdr(); 750 phdr.p_type = PT_PHDR; 751 phdr.p_flags = PF_R; 752 phdr.p_offset = phdr.p_vaddr = phdr.p_paddr = sizeof(Elf_Ehdr); 753 phdr.p_filesz = phdr.p_memsz = 0; // We need to fill this later. 754 phdr.p_align = sizeof(Elf_Off); 755 phdrs.push_back(phdr); 756 // Tell the linker to mmap the start of file to memory. 757 Elf_Phdr load = Elf_Phdr(); 758 load.p_type = PT_LOAD; 759 load.p_flags = PF_R; 760 load.p_offset = load.p_vaddr = load.p_paddr = 0; 761 load.p_filesz = load.p_memsz = sections_[0]->header_.sh_offset; 762 load.p_align = kPageSize; 763 phdrs.push_back(load); 764 } 765 // Create program headers for sections. 766 for (auto* section : sections_) { 767 const Elf_Shdr& shdr = section->header_; 768 if ((shdr.sh_flags & SHF_ALLOC) != 0 && shdr.sh_size != 0) { 769 // PT_LOAD tells the linker to mmap part of the file. 770 // The linker can only mmap page-aligned sections. 771 // Single PT_LOAD may contain several ELF sections. 772 Elf_Phdr& prev = phdrs.back(); 773 Elf_Phdr load = Elf_Phdr(); 774 load.p_type = PT_LOAD; 775 load.p_flags = section->phdr_flags_; 776 load.p_offset = shdr.sh_offset; 777 load.p_vaddr = load.p_paddr = shdr.sh_addr; 778 load.p_filesz = (shdr.sh_type != SHT_NOBITS ? shdr.sh_size : 0u); 779 load.p_memsz = shdr.sh_size; 780 load.p_align = shdr.sh_addralign; 781 if (prev.p_type == load.p_type && 782 prev.p_flags == load.p_flags && 783 prev.p_filesz == prev.p_memsz && // Do not merge .bss 784 load.p_filesz == load.p_memsz) { // Do not merge .bss 785 // Merge this PT_LOAD with the previous one. 786 Elf_Word size = shdr.sh_offset + shdr.sh_size - prev.p_offset; 787 prev.p_filesz = size; 788 prev.p_memsz = size; 789 } else { 790 // If we are adding new load, it must be aligned. 791 CHECK_EQ(shdr.sh_addralign, (Elf_Word)kPageSize); 792 phdrs.push_back(load); 793 } 794 } 795 } 796 for (auto* section : sections_) { 797 const Elf_Shdr& shdr = section->header_; 798 if ((shdr.sh_flags & SHF_ALLOC) != 0 && shdr.sh_size != 0) { 799 // Other PT_* types allow the program to locate interesting 800 // parts of memory at runtime. They must overlap with PT_LOAD. 801 if (section->phdr_type_ != 0) { 802 Elf_Phdr phdr = Elf_Phdr(); 803 phdr.p_type = section->phdr_type_; 804 phdr.p_flags = section->phdr_flags_; 805 phdr.p_offset = shdr.sh_offset; 806 phdr.p_vaddr = phdr.p_paddr = shdr.sh_addr; 807 phdr.p_filesz = phdr.p_memsz = shdr.sh_size; 808 phdr.p_align = shdr.sh_addralign; 809 phdrs.push_back(phdr); 810 } 811 } 812 } 813 // Set the size of the initial PT_PHDR. 814 CHECK_EQ(phdrs[0].p_type, (Elf_Word)PT_PHDR); 815 phdrs[0].p_filesz = phdrs[0].p_memsz = phdrs.size() * sizeof(Elf_Phdr); 816 817 return phdrs; 818 } 819 820 InstructionSet isa_; 821 822 ErrorDelayingOutputStream stream_; 823 824 Section rodata_; 825 Section text_; 826 Section bss_; 827 CachedStringSection dynstr_; 828 SymbolSection dynsym_; 829 CachedSection hash_; 830 CachedSection dynamic_; 831 Section eh_frame_; 832 Section eh_frame_hdr_; 833 StringSection strtab_; 834 SymbolSection symtab_; 835 Section debug_frame_; 836 Section debug_info_; 837 Section debug_line_; 838 StringSection shstrtab_; 839 std::vector<std::unique_ptr<Section>> other_sections_; 840 841 // List of used section in the order in which they were written. 842 std::vector<Section*> sections_; 843 844 bool started_; 845 bool write_program_headers_; 846 847 // The size of the memory taken by the ELF file when loaded. 848 size_t loaded_size_; 849 850 // Used for allocation of virtual address space. 851 Elf_Addr virtual_address_; 852 853 DISALLOW_COPY_AND_ASSIGN(ElfBuilder); 854}; 855 856} // namespace art 857 858#endif // ART_COMPILER_ELF_BUILDER_H_ 859