libsanitizer/sanitizer_common/sanitizer_procmaps_mac.cc

   1 //===-- sanitizer_procmaps_mac.cc -----------------------------------------===//
   2 //
   3 // This file is distributed under the University of Illinois Open Source
   4 // License. See LICENSE.TXT for details.
   5 //
   6 //===----------------------------------------------------------------------===//
   7 //
   8 // Information about the process mappings (Mac-specific parts).
   9 //===----------------------------------------------------------------------===//
  10
  11 #include "sanitizer_platform.h"
  12 #if SANITIZER_MAC
  13 #include "sanitizer_common.h"
  14 #include "sanitizer_placement_new.h"
  15 #include "sanitizer_procmaps.h"
  16
  17 #include <mach-o/dyld.h>
  18 #include <mach-o/loader.h>
  19 #include <mach/mach.h>
  20
  21 // These are not available in older macOS SDKs.
  22 #ifndef CPU_SUBTYPE_X86_64_H
  23 #define CPU_SUBTYPE_X86_64_H  ((cpu_subtype_t)8)   /* Haswell */
  24 #endif
  25 #ifndef CPU_SUBTYPE_ARM_V7S
  26 #define CPU_SUBTYPE_ARM_V7S   ((cpu_subtype_t)11)  /* Swift */
  27 #endif
  28 #ifndef CPU_SUBTYPE_ARM_V7K
  29 #define CPU_SUBTYPE_ARM_V7K   ((cpu_subtype_t)12)
  30 #endif
  31 #ifndef CPU_TYPE_ARM64
  32 #define CPU_TYPE_ARM64        (CPU_TYPE_ARM | CPU_ARCH_ABI64)
  33 #endif
  34
  35 namespace __sanitizer {
  36
  37 // Contains information used to iterate through sections.
  38 struct MemoryMappedSegmentData {
  39   char name[kMaxSegName];
  40   uptr nsects;
  41   char *current_load_cmd_addr;
  42   u32 lc_type;
  43   uptr base_virt_addr;
  44   uptr addr_mask;
  45 };
  46
  47 template <typename Section>
  48 static void NextSectionLoad(LoadedModule *module, MemoryMappedSegmentData *data,
  49                             bool isWritable) {
  50   const Section *sc = (const Section *)data->current_load_cmd_addr;
  51   data->current_load_cmd_addr += sizeof(Section);
  52
  53   uptr sec_start = (sc->addr & data->addr_mask) + data->base_virt_addr;
  54   uptr sec_end = sec_start + sc->size;
  55   module->addAddressRange(sec_start, sec_end, /*executable=*/false, isWritable,
  56                           sc->sectname);
  57 }
  58
  59 void MemoryMappedSegment::AddAddressRanges(LoadedModule *module) {
  60   // Don't iterate over sections when the caller hasn't set up the
  61   // data pointer, when there are no sections, or when the segment
  62   // is executable. Avoid iterating over executable sections because
  63   // it will confuse libignore, and because the extra granularity
  64   // of information is not needed by any sanitizers.
  65   if (!data_ || !data_->nsects || IsExecutable()) {
  66     module->addAddressRange(start, end, IsExecutable(), IsWritable(),
  67                             data_ ? data_->name : nullptr);
  68     return;
  69   }
  70
  71   do {
  72     if (data_->lc_type == LC_SEGMENT) {
  73       NextSectionLoad<struct section>(module, data_, IsWritable());
  74 #ifdef MH_MAGIC_64
  75     } else if (data_->lc_type == LC_SEGMENT_64) {
  76       NextSectionLoad<struct section_64>(module, data_, IsWritable());
  77 #endif
  78     }
  79   } while (--data_->nsects);
  80 }
  81
  82 MemoryMappingLayout::MemoryMappingLayout(bool cache_enabled) {
  83   Reset();
  84 }
  85
  86 MemoryMappingLayout::~MemoryMappingLayout() {
  87 }
  88
  89 // More information about Mach-O headers can be found in mach-o/loader.h
  90 // Each Mach-O image has a header (mach_header or mach_header_64) starting with
  91 // a magic number, and a list of linker load commands directly following the
  92 // header.
  93 // A load command is at least two 32-bit words: the command type and the
  94 // command size in bytes. We're interested only in segment load commands
  95 // (LC_SEGMENT and LC_SEGMENT_64), which tell that a part of the file is mapped
  96 // into the task's address space.
  97 // The |vmaddr|, |vmsize| and |fileoff| fields of segment_command or
  98 // segment_command_64 correspond to the memory address, memory size and the
  99 // file offset of the current memory segment.
 100 // Because these fields are taken from the images as is, one needs to add
 101 // _dyld_get_image_vmaddr_slide() to get the actual addresses at runtime.
 102
 103 void MemoryMappingLayout::Reset() {
 104   // Count down from the top.
 105   // TODO(glider): as per man 3 dyld, iterating over the headers with
 106   // _dyld_image_count is thread-unsafe. We need to register callbacks for
 107   // adding and removing images which will invalidate the MemoryMappingLayout
 108   // state.
 109   data_.current_image = _dyld_image_count();
 110   data_.current_load_cmd_count = -1;
 111   data_.current_load_cmd_addr = 0;
 112   data_.current_magic = 0;
 113   data_.current_filetype = 0;
 114   data_.current_arch = kModuleArchUnknown;
 115   internal_memset(data_.current_uuid, 0, kModuleUUIDSize);
 116 }
 117
 118 // The dyld load address should be unchanged throughout process execution,
 119 // and it is expensive to compute once many libraries have been loaded,
 120 // so cache it here and do not reset.
 121 static mach_header *dyld_hdr = 0;
 122 static const char kDyldPath[] = "/usr/lib/dyld";
 123 static const int kDyldImageIdx = -1;
 124
 125 // static
 126 void MemoryMappingLayout::CacheMemoryMappings() {
 127   // No-op on Mac for now.
 128 }
 129
 130 void MemoryMappingLayout::LoadFromCache() {
 131   // No-op on Mac for now.
 132 }
 133
 134 // _dyld_get_image_header() and related APIs don't report dyld itself.
 135 // We work around this by manually recursing through the memory map
 136 // until we hit a Mach header matching dyld instead. These recurse
 137 // calls are expensive, but the first memory map generation occurs
 138 // early in the process, when dyld is one of the only images loaded,
 139 // so it will be hit after only a few iterations.
 140 static mach_header *get_dyld_image_header() {
 141   mach_port_name_t port;
 142   if (task_for_pid(mach_task_self(), internal_getpid(), &port) !=
 143       KERN_SUCCESS) {
 144     return nullptr;
 145   }
 146
 147   unsigned depth = 1;
 148   vm_size_t size = 0;
 149   vm_address_t address = 0;
 150   kern_return_t err = KERN_SUCCESS;
 151   mach_msg_type_number_t count = VM_REGION_SUBMAP_INFO_COUNT_64;
 152
 153   while (true) {
 154     struct vm_region_submap_info_64 info;
 155     err = vm_region_recurse_64(port, &address, &size, &depth,
 156                                (vm_region_info_t)&info, &count);
 157     if (err != KERN_SUCCESS) return nullptr;
 158
 159     if (size >= sizeof(mach_header) && info.protection & kProtectionRead) {
 160       mach_header *hdr = (mach_header *)address;
 161       if ((hdr->magic == MH_MAGIC || hdr->magic == MH_MAGIC_64) &&
 162           hdr->filetype == MH_DYLINKER) {
 163         return hdr;
 164       }
 165     }
 166     address += size;
 167   }
 168 }
 169
 170 const mach_header *get_dyld_hdr() {
 171   if (!dyld_hdr) dyld_hdr = get_dyld_image_header();
 172
 173   return dyld_hdr;
 174 }
 175
 176 // Next and NextSegmentLoad were inspired by base/sysinfo.cc in
 177 // Google Perftools, https://github.com/gperftools/gperftools.
 178
 179 // NextSegmentLoad scans the current image for the next segment load command
 180 // and returns the start and end addresses and file offset of the corresponding
 181 // segment.
 182 // Note that the segment addresses are not necessarily sorted.
 183 template <u32 kLCSegment, typename SegmentCommand>
 184 static bool NextSegmentLoad(MemoryMappedSegment *segment,
 185 MemoryMappedSegmentData *seg_data, MemoryMappingLayoutData &layout_data) {
 186   const char *lc = layout_data.current_load_cmd_addr;
 187   layout_data.current_load_cmd_addr += ((const load_command *)lc)->cmdsize;
 188   if (((const load_command *)lc)->cmd == kLCSegment) {
 189     const SegmentCommand* sc = (const SegmentCommand *)lc;
 190     uptr base_virt_addr, addr_mask;
 191     if (layout_data.current_image == kDyldImageIdx) {
 192       base_virt_addr = (uptr)get_dyld_hdr();
 193       // vmaddr is masked with 0xfffff because on macOS versions < 10.12,
 194       // it contains an absolute address rather than an offset for dyld.
 195       // To make matters even more complicated, this absolute address
 196       // isn't actually the absolute segment address, but the offset portion
 197       // of the address is accurate when combined with the dyld base address,
 198       // and the mask will give just this offset.
 199       addr_mask = 0xfffff;
 200     } else {
 201       base_virt_addr =
 202           (uptr)_dyld_get_image_vmaddr_slide(layout_data.current_image);
 203       addr_mask = ~0;
 204     }
 205
 206     segment->start = (sc->vmaddr & addr_mask) + base_virt_addr;
 207     segment->end = segment->start + sc->vmsize;
 208     // Most callers don't need section information, so only fill this struct
 209     // when required.
 210     if (seg_data) {
 211       seg_data->nsects = sc->nsects;
 212       seg_data->current_load_cmd_addr =
 213           (char *)lc + sizeof(SegmentCommand);
 214       seg_data->lc_type = kLCSegment;
 215       seg_data->base_virt_addr = base_virt_addr;
 216       seg_data->addr_mask = addr_mask;
 217       internal_strncpy(seg_data->name, sc->segname,
 218                        ARRAY_SIZE(seg_data->name));
 219     }
 220
 221     // Return the initial protection.
 222     segment->protection = sc->initprot;
 223     segment->offset = (layout_data.current_filetype ==
 224                        /*MH_EXECUTE*/ 0x2)
 225                           ? sc->vmaddr
 226                           : sc->fileoff;
 227     if (segment->filename) {
 228       const char *src = (layout_data.current_image == kDyldImageIdx)
 229                             ? kDyldPath
 230                             : _dyld_get_image_name(layout_data.current_image);
 231       internal_strncpy(segment->filename, src, segment->filename_size);
 232     }
 233     segment->arch = layout_data.current_arch;
 234     internal_memcpy(segment->uuid, layout_data.current_uuid, kModuleUUIDSize);
 235     return true;
 236   }
 237   return false;
 238 }
 239
 240 ModuleArch ModuleArchFromCpuType(cpu_type_t cputype, cpu_subtype_t cpusubtype) {
 241   cpusubtype = cpusubtype & ~CPU_SUBTYPE_MASK;
 242   switch (cputype) {
 243     case CPU_TYPE_I386:
 244       return kModuleArchI386;
 245     case CPU_TYPE_X86_64:
 246       if (cpusubtype == CPU_SUBTYPE_X86_64_ALL) return kModuleArchX86_64;
 247       if (cpusubtype == CPU_SUBTYPE_X86_64_H) return kModuleArchX86_64H;
 248       CHECK(0 && "Invalid subtype of x86_64");
 249       return kModuleArchUnknown;
 250     case CPU_TYPE_ARM:
 251       if (cpusubtype == CPU_SUBTYPE_ARM_V6) return kModuleArchARMV6;
 252       if (cpusubtype == CPU_SUBTYPE_ARM_V7) return kModuleArchARMV7;
 253       if (cpusubtype == CPU_SUBTYPE_ARM_V7S) return kModuleArchARMV7S;
 254       if (cpusubtype == CPU_SUBTYPE_ARM_V7K) return kModuleArchARMV7K;
 255       CHECK(0 && "Invalid subtype of ARM");
 256       return kModuleArchUnknown;
 257     case CPU_TYPE_ARM64:
 258       return kModuleArchARM64;
 259     default:
 260       CHECK(0 && "Invalid CPU type");
 261       return kModuleArchUnknown;
 262   }
 263 }
 264
 265 static const load_command *NextCommand(const load_command *lc) {
 266   return (const load_command *)((char *)lc + lc->cmdsize);
 267 }
 268
 269 static void FindUUID(const load_command *first_lc, u8 *uuid_output) {
 270   for (const load_command *lc = first_lc; lc->cmd != 0; lc = NextCommand(lc)) {
 271     if (lc->cmd != LC_UUID) continue;
 272
 273     const uuid_command *uuid_lc = (const uuid_command *)lc;
 274     const uint8_t *uuid = &uuid_lc->uuid[0];
 275     internal_memcpy(uuid_output, uuid, kModuleUUIDSize);
 276     return;
 277   }
 278 }
 279
 280 static bool IsModuleInstrumented(const load_command *first_lc) {
 281   for (const load_command *lc = first_lc; lc->cmd != 0; lc = NextCommand(lc)) {
 282     if (lc->cmd != LC_LOAD_DYLIB) continue;
 283
 284     const dylib_command *dylib_lc = (const dylib_command *)lc;
 285     uint32_t dylib_name_offset = dylib_lc->dylib.name.offset;
 286     const char *dylib_name = ((const char *)dylib_lc) + dylib_name_offset;
 287     dylib_name = StripModuleName(dylib_name);
 288     if (dylib_name != 0 && (internal_strstr(dylib_name, "libclang_rt."))) {
 289       return true;
 290     }
 291   }
 292   return false;
 293 }
 294
 295 bool MemoryMappingLayout::Next(MemoryMappedSegment *segment) {
 296   for (; data_.current_image >= kDyldImageIdx; data_.current_image--) {
 297     const mach_header *hdr = (data_.current_image == kDyldImageIdx)
 298                                  ? get_dyld_hdr()
 299                                  : _dyld_get_image_header(data_.current_image);
 300     if (!hdr) continue;
 301     if (data_.current_load_cmd_count < 0) {
 302       // Set up for this image;
 303       data_.current_load_cmd_count = hdr->ncmds;
 304       data_.current_magic = hdr->magic;
 305       data_.current_filetype = hdr->filetype;
 306       data_.current_arch = ModuleArchFromCpuType(hdr->cputype, hdr->cpusubtype);
 307       switch (data_.current_magic) {
 308 #ifdef MH_MAGIC_64
 309         case MH_MAGIC_64: {
 310           data_.current_load_cmd_addr = (char *)hdr + sizeof(mach_header_64);
 311           break;
 312         }
 313 #endif
 314         case MH_MAGIC: {
 315           data_.current_load_cmd_addr = (char *)hdr + sizeof(mach_header);
 316           break;
 317         }
 318         default: {
 319           continue;
 320         }
 321       }
 322       FindUUID((const load_command *)data_.current_load_cmd_addr,
 323                data_.current_uuid);
 324       data_.current_instrumented = IsModuleInstrumented(
 325           (const load_command *)data_.current_load_cmd_addr);
 326     }
 327
 328     for (; data_.current_load_cmd_count >= 0; data_.current_load_cmd_count--) {
 329       switch (data_.current_magic) {
 330         // data_.current_magic may be only one of MH_MAGIC, MH_MAGIC_64.
 331 #ifdef MH_MAGIC_64
 332         case MH_MAGIC_64: {
 333           if (NextSegmentLoad<LC_SEGMENT_64, struct segment_command_64>(
 334           segment, segment->data_, data_))
 335             return true;
 336           break;
 337         }
 338 #endif
 339         case MH_MAGIC: {
 340           if (NextSegmentLoad<LC_SEGMENT, struct segment_command>(
 341           segment, segment->data_, data_))
 342             return true;
 343           break;
 344         }
 345       }
 346     }
 347     // If we get here, no more load_cmd's in this image talk about
 348     // segments.  Go on to the next image.
 349   }
 350   return false;
 351 }
 352
 353 void MemoryMappingLayout::DumpListOfModules(
 354     InternalMmapVectorNoCtor<LoadedModule> *modules) {
 355   Reset();
 356   InternalScopedString module_name(kMaxPathLength);
 357   MemoryMappedSegment segment(module_name.data(), kMaxPathLength);
 358   MemoryMappedSegmentData data;
 359   segment.data_ = &data;
 360   while (Next(&segment)) {
 361     if (segment.filename[0] == '\0') continue;
 362     LoadedModule *cur_module = nullptr;
 363     if (!modules->empty() &&
 364         0 == internal_strcmp(segment.filename, modules->back().full_name())) {
 365       cur_module = &modules->back();
 366     } else {
 367       modules->push_back(LoadedModule());
 368       cur_module = &modules->back();
 369       cur_module->set(segment.filename, segment.start, segment.arch,
 370                       segment.uuid, data_.current_instrumented);
 371     }
 372     segment.AddAddressRanges(cur_module);
 373   }
 374 }
 375
 376 }  // namespace __sanitizer
 377
 378 #endif  // SANITIZER_MAC