lockedpool.cpp raw

   1  // Copyright (c) 2016-present The Bitcoin Core developers
   2  // Distributed under the MIT software license, see the accompanying
   3  // file COPYING or http://www.opensource.org/licenses/mit-license.php.
   4  
   5  #include <support/lockedpool.h>
   6  #include <support/cleanse.h>
   7  
   8  #ifdef WIN32
   9  #include <windows.h>
  10  #else
  11  #include <sys/mman.h>
  12  #include <sys/resource.h>
  13  #include <unistd.h>
  14  #endif
  15  
  16  #include <algorithm>
  17  #include <limits>
  18  #include <stdexcept>
  19  #include <utility>
  20  #ifdef ARENA_DEBUG
  21  #include <iomanip>
  22  #include <iostream>
  23  #endif
  24  
  25  LockedPoolManager* LockedPoolManager::_instance = nullptr;
  26  
  27  /*******************************************************************************/
  28  // Utilities
  29  //
  30  /** Align up to power of 2 */
  31  static inline size_t align_up(size_t x, size_t align)
  32  {
  33      return (x + align - 1) & ~(align - 1);
  34  }
  35  
  36  /*******************************************************************************/
  37  // Implementation: Arena
  38  
  39  Arena::Arena(void *base_in, size_t size_in, size_t alignment_in):
  40      base(base_in), end(static_cast<char*>(base_in) + size_in), alignment(alignment_in)
  41  {
  42      // Start with one free chunk that covers the entire arena
  43      auto it = size_to_free_chunk.emplace(size_in, base);
  44      chunks_free.emplace(base, it);
  45      chunks_free_end.emplace(static_cast<char*>(base) + size_in, it);
  46  }
  47  
  48  Arena::~Arena() = default;
  49  
  50  void* Arena::alloc(size_t size)
  51  {
  52      // Round to next multiple of alignment
  53      size = align_up(size, alignment);
  54  
  55      // Don't handle zero-sized chunks
  56      if (size == 0)
  57          return nullptr;
  58  
  59      // Pick a large enough free-chunk. Returns an iterator pointing to the first element that is not less than key.
  60      // This allocation strategy is best-fit. According to "Dynamic Storage Allocation: A Survey and Critical Review",
  61      // Wilson et. al. 1995, https://www.scs.stanford.edu/14wi-cs140/sched/readings/wilson.pdf, best-fit and first-fit
  62      // policies seem to work well in practice.
  63      auto size_ptr_it = size_to_free_chunk.lower_bound(size);
  64      if (size_ptr_it == size_to_free_chunk.end())
  65          return nullptr;
  66  
  67      // Create the used-chunk, taking its space from the end of the free-chunk
  68      const size_t size_remaining = size_ptr_it->first - size;
  69      char* const free_chunk = static_cast<char*>(size_ptr_it->second);
  70      auto allocated = chunks_used.emplace(free_chunk + size_remaining, size).first;
  71      chunks_free_end.erase(free_chunk + size_ptr_it->first);
  72      if (size_ptr_it->first == size) {
  73          // whole chunk is used up
  74          chunks_free.erase(size_ptr_it->second);
  75      } else {
  76          // still some memory left in the chunk
  77          auto it_remaining = size_to_free_chunk.emplace(size_remaining, size_ptr_it->second);
  78          chunks_free[size_ptr_it->second] = it_remaining;
  79          chunks_free_end.emplace(free_chunk + size_remaining, it_remaining);
  80      }
  81      size_to_free_chunk.erase(size_ptr_it);
  82  
  83      return allocated->first;
  84  }
  85  
  86  void Arena::free(void *ptr)
  87  {
  88      // Freeing the nullptr pointer is OK.
  89      if (ptr == nullptr) {
  90          return;
  91      }
  92  
  93      // Remove chunk from used map
  94      auto i = chunks_used.find(ptr);
  95      if (i == chunks_used.end()) {
  96          throw std::runtime_error("Arena: invalid or double free");
  97      }
  98      auto freed = std::make_pair(static_cast<char*>(i->first), i->second);
  99      chunks_used.erase(i);
 100  
 101      // coalesce freed with previous chunk
 102      auto prev = chunks_free_end.find(freed.first);
 103      if (prev != chunks_free_end.end()) {
 104          freed.first -= prev->second->first;
 105          freed.second += prev->second->first;
 106          size_to_free_chunk.erase(prev->second);
 107          chunks_free_end.erase(prev);
 108      }
 109  
 110      // coalesce freed with chunk after freed
 111      auto next = chunks_free.find(freed.first + freed.second);
 112      if (next != chunks_free.end()) {
 113          freed.second += next->second->first;
 114          size_to_free_chunk.erase(next->second);
 115          chunks_free.erase(next);
 116      }
 117  
 118      // Add/set space with coalesced free chunk
 119      auto it = size_to_free_chunk.emplace(freed.second, freed.first);
 120      chunks_free[freed.first] = it;
 121      chunks_free_end[freed.first + freed.second] = it;
 122  }
 123  
 124  Arena::Stats Arena::stats() const
 125  {
 126      Arena::Stats r{ 0, 0, 0, chunks_used.size(), chunks_free.size() };
 127      for (const auto& chunk: chunks_used)
 128          r.used += chunk.second;
 129      for (const auto& chunk: chunks_free)
 130          r.free += chunk.second->first;
 131      r.total = r.used + r.free;
 132      return r;
 133  }
 134  
 135  #ifdef ARENA_DEBUG
 136  static void printchunk(void* base, size_t sz, bool used) {
 137      std::cout <<
 138          "0x" << std::hex << std::setw(16) << std::setfill('0') << base <<
 139          " 0x" << std::hex << std::setw(16) << std::setfill('0') << sz <<
 140          " 0x" << used << std::endl;
 141  }
 142  void Arena::walk() const
 143  {
 144      for (const auto& chunk: chunks_used)
 145          printchunk(chunk.first, chunk.second, true);
 146      std::cout << std::endl;
 147      for (const auto& chunk: chunks_free)
 148          printchunk(chunk.first, chunk.second->first, false);
 149      std::cout << std::endl;
 150  }
 151  #endif
 152  
 153  /*******************************************************************************/
 154  // Implementation: Win32LockedPageAllocator
 155  
 156  #ifdef WIN32
 157  /** LockedPageAllocator specialized for Windows.
 158   */
 159  class Win32LockedPageAllocator: public LockedPageAllocator
 160  {
 161  public:
 162      Win32LockedPageAllocator();
 163      void* AllocateLocked(size_t len, bool *lockingSuccess) override;
 164      void FreeLocked(void* addr, size_t len) override;
 165      size_t GetLimit() override;
 166  private:
 167      size_t page_size;
 168  };
 169  
 170  Win32LockedPageAllocator::Win32LockedPageAllocator()
 171  {
 172      // Determine system page size in bytes
 173      SYSTEM_INFO sSysInfo;
 174      GetSystemInfo(&sSysInfo);
 175      page_size = sSysInfo.dwPageSize;
 176  }
 177  void *Win32LockedPageAllocator::AllocateLocked(size_t len, bool *lockingSuccess)
 178  {
 179      len = align_up(len, page_size);
 180      void *addr = VirtualAlloc(nullptr, len, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE);
 181      if (addr) {
 182          // VirtualLock is used to attempt to keep keying material out of swap. Note
 183          // that it does not provide this as a guarantee, but, in practice, memory
 184          // that has been VirtualLock'd almost never gets written to the pagefile
 185          // except in rare circumstances where memory is extremely low.
 186          *lockingSuccess = VirtualLock(const_cast<void*>(addr), len) != 0;
 187      }
 188      return addr;
 189  }
 190  void Win32LockedPageAllocator::FreeLocked(void* addr, size_t len)
 191  {
 192      len = align_up(len, page_size);
 193      memory_cleanse(addr, len);
 194      VirtualUnlock(const_cast<void*>(addr), len);
 195  }
 196  
 197  size_t Win32LockedPageAllocator::GetLimit()
 198  {
 199      size_t min, max;
 200      if(GetProcessWorkingSetSize(GetCurrentProcess(), &min, &max) != 0) {
 201          return min;
 202      }
 203      return std::numeric_limits<size_t>::max();
 204  }
 205  #endif
 206  
 207  /*******************************************************************************/
 208  // Implementation: PosixLockedPageAllocator
 209  
 210  #ifndef WIN32
 211  /** LockedPageAllocator specialized for OSes that don't try to be
 212   * special snowflakes.
 213   */
 214  class PosixLockedPageAllocator: public LockedPageAllocator
 215  {
 216  public:
 217      PosixLockedPageAllocator();
 218      void* AllocateLocked(size_t len, bool *lockingSuccess) override;
 219      void FreeLocked(void* addr, size_t len) override;
 220      size_t GetLimit() override;
 221  private:
 222      size_t page_size;
 223  };
 224  
 225  PosixLockedPageAllocator::PosixLockedPageAllocator()
 226  {
 227      // Determine system page size in bytes
 228  #if defined(PAGESIZE) // defined in limits.h
 229      page_size = PAGESIZE;
 230  #else                   // assume some POSIX OS
 231      page_size = sysconf(_SC_PAGESIZE);
 232  #endif
 233  }
 234  
 235  void *PosixLockedPageAllocator::AllocateLocked(size_t len, bool *lockingSuccess)
 236  {
 237      void *addr;
 238      len = align_up(len, page_size);
 239      addr = mmap(nullptr, len, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
 240      if (addr == MAP_FAILED) {
 241          return nullptr;
 242      }
 243      if (addr) {
 244          *lockingSuccess = mlock(addr, len) == 0;
 245  #if defined(MADV_DONTDUMP) // Linux
 246          madvise(addr, len, MADV_DONTDUMP);
 247  #elif defined(MADV_NOCORE) // FreeBSD
 248          madvise(addr, len, MADV_NOCORE);
 249  #endif
 250      }
 251      return addr;
 252  }
 253  void PosixLockedPageAllocator::FreeLocked(void* addr, size_t len)
 254  {
 255      len = align_up(len, page_size);
 256      memory_cleanse(addr, len);
 257      munlock(addr, len);
 258      munmap(addr, len);
 259  }
 260  size_t PosixLockedPageAllocator::GetLimit()
 261  {
 262  #ifdef RLIMIT_MEMLOCK
 263      struct rlimit rlim;
 264      if (getrlimit(RLIMIT_MEMLOCK, &rlim) == 0) {
 265          if (rlim.rlim_cur != RLIM_INFINITY &&
 266              std::cmp_less_equal(rlim.rlim_cur, static_cast<rlim_t>(std::numeric_limits<size_t>::max()))) {
 267              return rlim.rlim_cur;
 268          }
 269      }
 270  #endif
 271      return std::numeric_limits<size_t>::max();
 272  }
 273  #endif
 274  
 275  /*******************************************************************************/
 276  // Implementation: LockedPool
 277  
 278  LockedPool::LockedPool(std::unique_ptr<LockedPageAllocator> allocator_in, LockingFailed_Callback lf_cb_in)
 279      : allocator(std::move(allocator_in)), lf_cb(lf_cb_in)
 280  {
 281  }
 282  
 283  LockedPool::~LockedPool() = default;
 284  
 285  void* LockedPool::alloc(size_t size)
 286  {
 287      std::lock_guard<std::mutex> lock(mutex);
 288  
 289      // Don't handle impossible sizes
 290      if (size == 0 || size > ARENA_SIZE)
 291          return nullptr;
 292  
 293      // Try allocating from each current arena
 294      for (auto &arena: arenas) {
 295          void *addr = arena.alloc(size);
 296          if (addr) {
 297              return addr;
 298          }
 299      }
 300      // If that fails, create a new one
 301      if (new_arena(ARENA_SIZE, ARENA_ALIGN)) {
 302          return arenas.back().alloc(size);
 303      }
 304      return nullptr;
 305  }
 306  
 307  void LockedPool::free(void *ptr)
 308  {
 309      std::lock_guard<std::mutex> lock(mutex);
 310      // TODO we can do better than this linear search by keeping a map of arena
 311      // extents to arena, and looking up the address.
 312      for (auto &arena: arenas) {
 313          if (arena.addressInArena(ptr)) {
 314              arena.free(ptr);
 315              return;
 316          }
 317      }
 318      throw std::runtime_error("LockedPool: invalid address not pointing to any arena");
 319  }
 320  
 321  LockedPool::Stats LockedPool::stats() const
 322  {
 323      std::lock_guard<std::mutex> lock(mutex);
 324      LockedPool::Stats r{0, 0, 0, cumulative_bytes_locked, 0, 0};
 325      for (const auto &arena: arenas) {
 326          Arena::Stats i = arena.stats();
 327          r.used += i.used;
 328          r.free += i.free;
 329          r.total += i.total;
 330          r.chunks_used += i.chunks_used;
 331          r.chunks_free += i.chunks_free;
 332      }
 333      return r;
 334  }
 335  
 336  bool LockedPool::new_arena(size_t size, size_t align)
 337  {
 338      bool locked;
 339      // If this is the first arena, handle this specially: Cap the upper size
 340      // by the process limit. This makes sure that the first arena will at least
 341      // be locked. An exception to this is if the process limit is 0:
 342      // in this case no memory can be locked at all so we'll skip past this logic.
 343      if (arenas.empty()) {
 344          size_t limit = allocator->GetLimit();
 345          if (limit > 0) {
 346              size = std::min(size, limit);
 347          }
 348      }
 349      void *addr = allocator->AllocateLocked(size, &locked);
 350      if (!addr) {
 351          return false;
 352      }
 353      if (locked) {
 354          cumulative_bytes_locked += size;
 355      } else if (lf_cb) { // Call the locking-failed callback if locking failed
 356          if (!lf_cb()) { // If the callback returns false, free the memory and fail, otherwise consider the user warned and proceed.
 357              allocator->FreeLocked(addr, size);
 358              return false;
 359          }
 360      }
 361      arenas.emplace_back(allocator.get(), addr, size, align);
 362      return true;
 363  }
 364  
 365  LockedPool::LockedPageArena::LockedPageArena(LockedPageAllocator *allocator_in, void *base_in, size_t size_in, size_t align_in):
 366      Arena(base_in, size_in, align_in), base(base_in), size(size_in), allocator(allocator_in)
 367  {
 368  }
 369  LockedPool::LockedPageArena::~LockedPageArena()
 370  {
 371      allocator->FreeLocked(base, size);
 372  }
 373  
 374  /*******************************************************************************/
 375  // Implementation: LockedPoolManager
 376  //
 377  LockedPoolManager::LockedPoolManager(std::unique_ptr<LockedPageAllocator> allocator_in):
 378      LockedPool(std::move(allocator_in), &LockedPoolManager::LockingFailed)
 379  {
 380  }
 381  
 382  bool LockedPoolManager::LockingFailed()
 383  {
 384      // TODO: log something but how? without including util.h
 385      return true;
 386  }
 387  
 388  void LockedPoolManager::CreateInstance()
 389  {
 390      // Using a local static instance guarantees that the object is initialized
 391      // when it's first needed and also deinitialized after all objects that use
 392      // it are done with it.  I can think of one unlikely scenario where we may
 393      // have a static deinitialization order/problem, but the check in
 394      // LockedPoolManagerBase's destructor helps us detect if that ever happens.
 395  #ifdef WIN32
 396      std::unique_ptr<LockedPageAllocator> allocator(new Win32LockedPageAllocator());
 397  #else
 398      std::unique_ptr<LockedPageAllocator> allocator(new PosixLockedPageAllocator());
 399  #endif
 400      static LockedPoolManager instance(std::move(allocator));
 401      LockedPoolManager::_instance = &instance;
 402  }
 403  
 404  LockedPoolManager& LockedPoolManager::Instance()
 405  {
 406      static std::once_flag init_flag;
 407      std::call_once(init_flag, LockedPoolManager::CreateInstance);
 408      return *LockedPoolManager::_instance;
 409  }
 410