Adding upstream version 115.7.0esr.upstream/115.7.0esrupstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

1 files changed, 1050 insertions, 0 deletions

diff --git a/js/src/gc/Memory.cpp b/js/src/gc/Memory.cpp
new file mode 100644
index 0000000000..c5bdaa14bf
--- /dev/null
+++ b/js/src/gc/Memory.cpp
@@ -0,0 +1,1050 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: set ts=8 sts=2 et sw=2 tw=80:
+ * This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "gc/Memory.h"
+
+#include "mozilla/Atomics.h"
+#include "mozilla/MathAlgorithms.h"
+#include "mozilla/RandomNum.h"
+#include "mozilla/TaggedAnonymousMemory.h"
+
+#include "jit/JitOptions.h"
+#include "js/HeapAPI.h"
+#include "js/Utility.h"
+#include "util/Memory.h"
+
+#ifdef XP_WIN
+
+#  include "util/WindowsWrapper.h"
+#  include <psapi.h>
+
+#else
+
+#  include <algorithm>
+#  include <errno.h>
+#  include <unistd.h>
+
+#  if !defined(__wasi__)
+#    include <sys/mman.h>
+#    include <sys/resource.h>
+#    include <sys/stat.h>
+#    include <sys/types.h>
+#  endif  // !defined(__wasi__)
+
+#endif  // !XP_WIN
+
+namespace js {
+namespace gc {
+
+/*
+ * System allocation functions generally require the allocation size
+ * to be an integer multiple of the page size of the running process.
+ */
+static size_t pageSize = 0;
+
+/* The OS allocation granularity may not match the page size. */
+static size_t allocGranularity = 0;
+
+/* The number of bits used by addresses on this platform. */
+static size_t numAddressBits = 0;
+
+/* An estimate of the number of bytes available for virtual memory. */
+static size_t virtualMemoryLimit = size_t(-1);
+
+/*
+ * System allocation functions may hand out regions of memory in increasing or
+ * decreasing order. This ordering is used as a hint during chunk alignment to
+ * reduce the number of system calls. On systems with 48-bit addresses, our
+ * workarounds to obtain 47-bit pointers cause addresses to be handed out in
+ * increasing order.
+ *
+ * We do not use the growth direction on Windows, as constraints on VirtualAlloc
+ * would make its application failure prone and complex. Tests indicate that
+ * VirtualAlloc always hands out regions of memory in increasing order.
+ */
+#if defined(XP_DARWIN)
+static mozilla::Atomic<int, mozilla::Relaxed> growthDirection(1);
+#elif defined(XP_UNIX)
+static mozilla::Atomic<int, mozilla::Relaxed> growthDirection(0);
+#endif
+
+/*
+ * Data from OOM crashes shows there may be up to 24 chunk-sized but unusable
+ * chunks available in low memory situations. These chunks may all need to be
+ * used up before we gain access to remaining *alignable* chunk-sized regions,
+ * so we use a generous limit of 32 unusable chunks to ensure we reach them.
+ */
+static const int MaxLastDitchAttempts = 32;
+
+#ifdef JS_64BIT
+/*
+ * On some 64-bit platforms we can use a random, scattershot allocator that
+ * tries addresses from the available range at random. If the address range
+ * is large enough this will have a high chance of success and additionally
+ * makes the memory layout of our process less predictable.
+ *
+ * However, not all 64-bit platforms have a very large address range. For
+ * example, AArch64 on Linux defaults to using 39-bit addresses to limit the
+ * number of translation tables used. On such configurations the scattershot
+ * approach to allocation creates a conflict with our desire to reserve large
+ * regions of memory for applications like WebAssembly: Small allocations may
+ * inadvertently block off all available 4-6GiB regions, and conversely
+ * reserving such regions may lower the success rate for smaller allocations to
+ * unacceptable levels.
+ *
+ * So we make a compromise: Instead of using the scattershot on all 64-bit
+ * platforms, we only use it on platforms that meet a minimum requirement for
+ * the available address range. In addition we split the address range,
+ * reserving the upper half for huge allocations and the lower half for smaller
+ * allocations. We use a limit of 43 bits so that at least 42 bits are available
+ * for huge allocations - this matches the 8TiB per process address space limit
+ * that we're already subject to on Windows.
+ */
+static const size_t MinAddressBitsForRandomAlloc = 43;
+
+/* The lower limit for huge allocations. This is fairly arbitrary. */
+static const size_t HugeAllocationSize = 1024 * 1024 * 1024;
+
+/* The minimum and maximum valid addresses that can be allocated into. */
+static size_t minValidAddress = 0;
+static size_t maxValidAddress = 0;
+
+/* The upper limit for smaller allocations and the lower limit for huge ones. */
+static size_t hugeSplit = 0;
+#endif
+
+size_t SystemPageSize() { return pageSize; }
+
+size_t SystemAddressBits() { return numAddressBits; }
+
+size_t VirtualMemoryLimit() { return virtualMemoryLimit; }
+
+bool UsingScattershotAllocator() {
+#ifdef JS_64BIT
+  return numAddressBits >= MinAddressBitsForRandomAlloc;
+#else
+  return false;
+#endif
+}
+
+enum class Commit : bool {
+  No = false,
+  Yes = true,
+};
+
+#ifdef XP_WIN
+enum class PageAccess : DWORD {
+  None = PAGE_NOACCESS,
+  Read = PAGE_READONLY,
+  ReadWrite = PAGE_READWRITE,
+  Execute = PAGE_EXECUTE,
+  ReadExecute = PAGE_EXECUTE_READ,
+  ReadWriteExecute = PAGE_EXECUTE_READWRITE,
+};
+#elif defined(__wasi__)
+enum class PageAccess : int {
+  None = 0,
+  Read = 0,
+  ReadWrite = 0,
+  Execute = 0,
+  ReadExecute = 0,
+  ReadWriteExecute = 0,
+};
+#else
+enum class PageAccess : int {
+  None = PROT_NONE,
+  Read = PROT_READ,
+  ReadWrite = PROT_READ | PROT_WRITE,
+  Execute = PROT_EXEC,
+  ReadExecute = PROT_READ | PROT_EXEC,
+  ReadWriteExecute = PROT_READ | PROT_WRITE | PROT_EXEC,
+};
+#endif
+
+template <bool AlwaysGetNew = true>
+static bool TryToAlignChunk(void** aRegion, void** aRetainedRegion,
+                            size_t length, size_t alignment);
+
+#ifndef __wasi__
+static void* MapAlignedPagesSlow(size_t length, size_t alignment);
+#endif  // wasi
+static void* MapAlignedPagesLastDitch(size_t length, size_t alignment);
+
+#ifdef JS_64BIT
+static void* MapAlignedPagesRandom(size_t length, size_t alignment);
+#endif
+
+void* TestMapAlignedPagesLastDitch(size_t length, size_t alignment) {
+  return MapAlignedPagesLastDitch(length, alignment);
+}
+
+bool DecommitEnabled() { return SystemPageSize() == PageSize; }
+
+/* Returns the offset from the nearest aligned address at or below |region|. */
+static inline size_t OffsetFromAligned(void* region, size_t alignment) {
+  return uintptr_t(region) % alignment;
+}
+
+template <Commit commit, PageAccess prot>
+static inline void* MapInternal(void* desired, size_t length) {
+  void* region = nullptr;
+#ifdef XP_WIN
+  DWORD flags =
+      (commit == Commit::Yes ? MEM_RESERVE | MEM_COMMIT : MEM_RESERVE);
+  region = VirtualAlloc(desired, length, flags, DWORD(prot));
+#elif defined(__wasi__)
+  if (int err = posix_memalign(&region, gc::SystemPageSize(), length)) {
+    MOZ_RELEASE_ASSERT(err == ENOMEM);
+    return nullptr;
+  }
+  if (region) {
+    memset(region, 0, length);
+  }
+#else
+  int flags = MAP_PRIVATE | MAP_ANON;
+  region = MozTaggedAnonymousMmap(desired, length, int(prot), flags, -1, 0,
+                                  "js-gc-heap");
+  if (region == MAP_FAILED) {
+    return nullptr;
+  }
+#endif
+  return region;
+}
+
+static inline void UnmapInternal(void* region, size_t length) {
+  MOZ_ASSERT(region && OffsetFromAligned(region, allocGranularity) == 0);
+  MOZ_ASSERT(length > 0 && length % pageSize == 0);
+
+#ifdef XP_WIN
+  MOZ_RELEASE_ASSERT(VirtualFree(region, 0, MEM_RELEASE) != 0);
+#elif defined(__wasi__)
+  free(region);
+#else
+  if (munmap(region, length)) {
+    MOZ_RELEASE_ASSERT(errno == ENOMEM);
+  }
+#endif
+}
+
+template <Commit commit = Commit::Yes, PageAccess prot = PageAccess::ReadWrite>
+static inline void* MapMemory(size_t length) {
+  MOZ_ASSERT(length > 0);
+
+  return MapInternal<commit, prot>(nullptr, length);
+}
+
+/*
+ * Attempts to map memory at the given address, but allows the system
+ * to return a different address that may still be suitable.
+ */
+template <Commit commit = Commit::Yes, PageAccess prot = PageAccess::ReadWrite>
+static inline void* MapMemoryAtFuzzy(void* desired, size_t length) {
+  MOZ_ASSERT(desired && OffsetFromAligned(desired, allocGranularity) == 0);
+  MOZ_ASSERT(length > 0);
+
+  // Note that some platforms treat the requested address as a hint, so the
+  // returned address might not match the requested address.
+  return MapInternal<commit, prot>(desired, length);
+}
+
+/*
+ * Attempts to map memory at the given address, returning nullptr if
+ * the system returns any address other than the requested one.
+ */
+template <Commit commit = Commit::Yes, PageAccess prot = PageAccess::ReadWrite>
+static inline void* MapMemoryAt(void* desired, size_t length) {
+  MOZ_ASSERT(desired && OffsetFromAligned(desired, allocGranularity) == 0);
+  MOZ_ASSERT(length > 0);
+
+  void* region = MapInternal<commit, prot>(desired, length);
+  if (!region) {
+    return nullptr;
+  }
+
+  // On some platforms mmap treats the desired address as a hint, so
+  // check that the address we got is the address we requested.
+  if (region != desired) {
+    UnmapInternal(region, length);
+    return nullptr;
+  }
+  return region;
+}
+
+#ifdef JS_64BIT
+
+/* Returns a random number in the given range. */
+static inline uint64_t GetNumberInRange(uint64_t minNum, uint64_t maxNum) {
+  const uint64_t MaxRand = UINT64_C(0xffffffffffffffff);
+  maxNum -= minNum;
+  uint64_t binSize = 1 + (MaxRand - maxNum) / (maxNum + 1);
+
+  uint64_t rndNum;
+  do {
+    mozilla::Maybe<uint64_t> result;
+    do {
+      result = mozilla::RandomUint64();
+    } while (!result);
+    rndNum = result.value() / binSize;
+  } while (rndNum > maxNum);
+
+  return minNum + rndNum;
+}
+
+#  ifndef XP_WIN
+static inline uint64_t FindAddressLimitInner(size_t highBit, size_t tries);
+
+/*
+ * The address range available to applications depends on both hardware and
+ * kernel configuration. For example, AArch64 on Linux uses addresses with
+ * 39 significant bits by default, but can be configured to use addresses with
+ * 48 significant bits by enabling a 4th translation table. Unfortunately,
+ * there appears to be no standard way to query the limit at runtime
+ * (Windows exposes this via GetSystemInfo()).
+ *
+ * This function tries to find the address limit by performing a binary search
+ * on the index of the most significant set bit in the addresses it attempts to
+ * allocate. As the requested address is often treated as a hint by the
+ * operating system, we use the actual returned addresses to narrow the range.
+ * We return the number of bits of an address that may be set.
+ */
+static size_t FindAddressLimit() {
+  // Use 32 bits as a lower bound in case we keep getting nullptr.
+  uint64_t low = 31;
+  uint64_t highestSeen = (UINT64_C(1) << 32) - allocGranularity - 1;
+
+  // Exclude 48-bit and 47-bit addresses first.
+  uint64_t high = 47;
+  for (; high >= std::max(low, UINT64_C(46)); --high) {
+    highestSeen = std::max(FindAddressLimitInner(high, 4), highestSeen);
+    low = mozilla::FloorLog2(highestSeen);
+  }
+  // If those didn't work, perform a modified binary search.
+  while (high - 1 > low) {
+    uint64_t middle = low + (high - low) / 2;
+    highestSeen = std::max(FindAddressLimitInner(middle, 4), highestSeen);
+    low = mozilla::FloorLog2(highestSeen);
+    if (highestSeen < (UINT64_C(1) << middle)) {
+      high = middle;
+    }
+  }
+  // We can be sure of the lower bound, but check the upper bound again.
+  do {
+    high = low + 1;
+    highestSeen = std::max(FindAddressLimitInner(high, 8), highestSeen);
+    low = mozilla::FloorLog2(highestSeen);
+  } while (low >= high);
+
+  // `low` is the highest set bit, so `low + 1` is the number of bits.
+  return low + 1;
+}
+
+static inline uint64_t FindAddressLimitInner(size_t highBit, size_t tries) {
+  const size_t length = allocGranularity;  // Used as both length and alignment.
+
+  uint64_t highestSeen = 0;
+  uint64_t startRaw = UINT64_C(1) << highBit;
+  uint64_t endRaw = 2 * startRaw - length - 1;
+  uint64_t start = (startRaw + length - 1) / length;
+  uint64_t end = (endRaw - (length - 1)) / length;
+  for (size_t i = 0; i < tries; ++i) {
+    uint64_t desired = length * GetNumberInRange(start, end);
+    void* address = MapMemoryAtFuzzy(reinterpret_cast<void*>(desired), length);
+    uint64_t actual = uint64_t(address);
+    if (address) {
+      UnmapInternal(address, length);
+    }
+    if (actual > highestSeen) {
+      highestSeen = actual;
+      if (actual >= startRaw) {
+        break;
+      }
+    }
+  }
+  return highestSeen;
+}
+#  endif  // !defined(XP_WIN)
+
+#endif  // defined(JS_64BIT)
+
+void InitMemorySubsystem() {
+  if (pageSize == 0) {
+#ifdef XP_WIN
+    SYSTEM_INFO sysinfo;
+    GetSystemInfo(&sysinfo);
+    pageSize = sysinfo.dwPageSize;
+    allocGranularity = sysinfo.dwAllocationGranularity;
+#else
+    pageSize = size_t(sysconf(_SC_PAGESIZE));
+    allocGranularity = pageSize;
+#endif
+#ifdef JS_64BIT
+#  ifdef XP_WIN
+    minValidAddress = size_t(sysinfo.lpMinimumApplicationAddress);
+    maxValidAddress = size_t(sysinfo.lpMaximumApplicationAddress);
+    numAddressBits = mozilla::FloorLog2(maxValidAddress) + 1;
+#  else
+    // No standard way to determine these, so fall back to FindAddressLimit().
+    numAddressBits = FindAddressLimit();
+    minValidAddress = allocGranularity;
+    maxValidAddress = (UINT64_C(1) << numAddressBits) - 1 - allocGranularity;
+#  endif
+    // Sanity check the address to ensure we don't use more than 47 bits.
+    uint64_t maxJSAddress = UINT64_C(0x00007fffffffffff) - allocGranularity;
+    if (maxValidAddress > maxJSAddress) {
+      maxValidAddress = maxJSAddress;
+      hugeSplit = UINT64_C(0x00003fffffffffff) - allocGranularity;
+    } else {
+      hugeSplit = (UINT64_C(1) << (numAddressBits - 1)) - 1 - allocGranularity;
+    }
+#else  // !defined(JS_64BIT)
+    numAddressBits = 32;
+#endif
+#ifdef RLIMIT_AS
+    if (jit::HasJitBackend()) {
+      rlimit as_limit;
+      if (getrlimit(RLIMIT_AS, &as_limit) == 0 &&
+          as_limit.rlim_max != RLIM_INFINITY) {
+        virtualMemoryLimit = as_limit.rlim_max;
+      }
+    }
+#endif
+  }
+}
+
+#ifdef JS_64BIT
+/* The JS engine uses 47-bit pointers; all higher bits must be clear. */
+static inline bool IsInvalidRegion(void* region, size_t length) {
+  const uint64_t invalidPointerMask = UINT64_C(0xffff800000000000);
+  return (uintptr_t(region) + length - 1) & invalidPointerMask;
+}
+#endif
+
+void* MapAlignedPages(size_t length, size_t alignment) {
+  MOZ_RELEASE_ASSERT(length > 0 && alignment > 0);
+  MOZ_RELEASE_ASSERT(length % pageSize == 0);
+  MOZ_RELEASE_ASSERT(std::max(alignment, allocGranularity) %
+                         std::min(alignment, allocGranularity) ==
+                     0);
+
+  // Smaller alignments aren't supported by the allocation functions.
+  if (alignment < allocGranularity) {
+    alignment = allocGranularity;
+  }
+
+#ifdef __wasi__
+  void* region = nullptr;
+  if (int err = posix_memalign(&region, alignment, length)) {
+    MOZ_ASSERT(err == ENOMEM);
+    return nullptr;
+  }
+  MOZ_ASSERT(region != nullptr);
+  memset(region, 0, length);
+  return region;
+#else
+
+#  ifdef JS_64BIT
+  // Use the scattershot allocator if the address range is large enough.
+  if (UsingScattershotAllocator()) {
+    void* region = MapAlignedPagesRandom(length, alignment);
+
+    MOZ_RELEASE_ASSERT(!IsInvalidRegion(region, length));
+    MOZ_ASSERT(OffsetFromAligned(region, alignment) == 0);
+
+    return region;
+  }
+#  endif
+
+  // Try to allocate the region. If the returned address is aligned,
+  // either we OOMed (region is nullptr) or we're done.
+  void* region = MapMemory(length);
+  if (OffsetFromAligned(region, alignment) == 0) {
+    return region;
+  }
+
+  // Try to align the region. On success, TryToAlignChunk() returns
+  // true and we can return the aligned region immediately.
+  void* retainedRegion;
+  if (TryToAlignChunk(&region, &retainedRegion, length, alignment)) {
+    MOZ_ASSERT(region && OffsetFromAligned(region, alignment) == 0);
+    MOZ_ASSERT(!retainedRegion);
+    return region;
+  }
+
+  // On failure, the unaligned region is retained unless we OOMed. We don't
+  // use the retained region on this path (see the last ditch allocator).
+  if (retainedRegion) {
+    UnmapInternal(retainedRegion, length);
+  }
+
+  // If it fails to align the given region, TryToAlignChunk() returns the
+  // next valid region that we might be able to align (unless we OOMed).
+  if (region) {
+    MOZ_ASSERT(OffsetFromAligned(region, alignment) != 0);
+    UnmapInternal(region, length);
+  }
+
+  // Since we couldn't align the first region, fall back to allocating a
+  // region large enough that we can definitely align it.
+  region = MapAlignedPagesSlow(length, alignment);
+  if (!region) {
+    // If there wasn't enough contiguous address space left for that,
+    // try to find an alignable region using the last ditch allocator.
+    region = MapAlignedPagesLastDitch(length, alignment);
+  }
+
+  // At this point we should either have an aligned region or nullptr.
+  MOZ_ASSERT(OffsetFromAligned(region, alignment) == 0);
+  return region;
+#endif  // !__wasi__
+}
+
+#ifdef JS_64BIT
+
+/*
+ * This allocator takes advantage of the large address range on some 64-bit
+ * platforms to allocate in a scattershot manner, choosing addresses at random
+ * from the range. By controlling the range we can avoid returning addresses
+ * that have more than 47 significant bits (as required by SpiderMonkey).
+ * This approach also has some other advantages over the methods employed by
+ * the other allocation functions in this file:
+ * 1) Allocations are extremely likely to succeed on the first try.
+ * 2) The randomness makes our memory layout becomes harder to predict.
+ * 3) The low probability of reusing regions guards against use-after-free.
+ *
+ * The main downside is that detecting physical OOM situations becomes more
+ * difficult; to guard against this, we occasionally try a regular allocation.
+ * In addition, sprinkling small allocations throughout the full address range
+ * might get in the way of large address space reservations such as those
+ * employed by WebAssembly. To avoid this (or the opposite problem of such
+ * reservations reducing the chance of success for smaller allocations) we
+ * split the address range in half, with one half reserved for huge allocations
+ * and the other for regular (usually chunk sized) allocations.
+ */
+static void* MapAlignedPagesRandom(size_t length, size_t alignment) {
+  uint64_t minNum, maxNum;
+  if (length < HugeAllocationSize) {
+    // Use the lower half of the range.
+    minNum = (minValidAddress + alignment - 1) / alignment;
+    maxNum = (hugeSplit - (length - 1)) / alignment;
+  } else {
+    // Use the upper half of the range.
+    minNum = (hugeSplit + 1 + alignment - 1) / alignment;
+    maxNum = (maxValidAddress - (length - 1)) / alignment;
+  }
+
+  // Try to allocate in random aligned locations.
+  void* region = nullptr;
+  for (size_t i = 1; i <= 1024; ++i) {
+    if (i & 0xf) {
+      uint64_t desired = alignment * GetNumberInRange(minNum, maxNum);
+      region = MapMemoryAtFuzzy(reinterpret_cast<void*>(desired), length);
+      if (!region) {
+        continue;
+      }
+    } else {
+      // Check for OOM.
+      region = MapMemory(length);
+      if (!region) {
+        return nullptr;
+      }
+    }
+    if (IsInvalidRegion(region, length)) {
+      UnmapInternal(region, length);
+      continue;
+    }
+    if (OffsetFromAligned(region, alignment) == 0) {
+      return region;
+    }
+    void* retainedRegion = nullptr;
+    if (TryToAlignChunk<false>(&region, &retainedRegion, length, alignment)) {
+      MOZ_ASSERT(region && OffsetFromAligned(region, alignment) == 0);
+      MOZ_ASSERT(!retainedRegion);
+      return region;
+    }
+    MOZ_ASSERT(region && !retainedRegion);
+    UnmapInternal(region, length);
+  }
+
+  if (numAddressBits < 48) {
+    // Try the reliable fallback of overallocating.
+    // Note: This will not respect the address space split.
+    region = MapAlignedPagesSlow(length, alignment);
+    if (region) {
+      return region;
+    }
+  }
+  if (length < HugeAllocationSize) {
+    MOZ_CRASH("Couldn't allocate even after 1000 tries!");
+  }
+
+  return nullptr;
+}
+
+#endif  // defined(JS_64BIT)
+
+#ifndef __wasi__
+static void* MapAlignedPagesSlow(size_t length, size_t alignment) {
+  void* alignedRegion = nullptr;
+  do {
+    size_t reserveLength = length + alignment - pageSize;
+#  ifdef XP_WIN
+    // Don't commit the requested pages as we won't use the region directly.
+    void* region = MapMemory<Commit::No>(reserveLength);
+#  else
+    void* region = MapMemory(reserveLength);
+#  endif
+    if (!region) {
+      return nullptr;
+    }
+    alignedRegion =
+        reinterpret_cast<void*>(AlignBytes(uintptr_t(region), alignment));
+#  ifdef XP_WIN
+    // Windows requires that map and unmap calls be matched, so deallocate
+    // and immediately reallocate at the desired (aligned) address.
+    UnmapInternal(region, reserveLength);
+    alignedRegion = MapMemoryAt(alignedRegion, length);
+#  else
+    // munmap allows us to simply unmap the pages that don't interest us.
+    if (alignedRegion != region) {
+      UnmapInternal(region, uintptr_t(alignedRegion) - uintptr_t(region));
+    }
+    void* regionEnd =
+        reinterpret_cast<void*>(uintptr_t(region) + reserveLength);
+    void* alignedEnd =
+        reinterpret_cast<void*>(uintptr_t(alignedRegion) + length);
+    if (alignedEnd != regionEnd) {
+      UnmapInternal(alignedEnd, uintptr_t(regionEnd) - uintptr_t(alignedEnd));
+    }
+#  endif
+    // On Windows we may have raced with another thread; if so, try again.
+  } while (!alignedRegion);
+
+  return alignedRegion;
+}
+#endif  // wasi
+
+/*
+ * In a low memory or high fragmentation situation, alignable chunks of the
+ * desired length may still be available, even if there are no more contiguous
+ * free chunks that meet the |length + alignment - pageSize| requirement of
+ * MapAlignedPagesSlow. In this case, try harder to find an alignable chunk
+ * by temporarily holding onto the unaligned parts of each chunk until the
+ * allocator gives us a chunk that either is, or can be aligned.
+ */
+static void* MapAlignedPagesLastDitch(size_t length, size_t alignment) {
+  void* tempMaps[MaxLastDitchAttempts];
+  int attempt = 0;
+  void* region = MapMemory(length);
+  if (OffsetFromAligned(region, alignment) == 0) {
+    return region;
+  }
+  for (; attempt < MaxLastDitchAttempts; ++attempt) {
+    if (TryToAlignChunk(&region, tempMaps + attempt, length, alignment)) {
+      MOZ_ASSERT(region && OffsetFromAligned(region, alignment) == 0);
+      MOZ_ASSERT(!tempMaps[attempt]);
+      break;  // Success!
+    }
+    if (!region || !tempMaps[attempt]) {
+      break;  // We ran out of memory, so give up.
+    }
+  }
+  if (OffsetFromAligned(region, alignment)) {
+    UnmapInternal(region, length);
+    region = nullptr;
+  }
+  while (--attempt >= 0) {
+    UnmapInternal(tempMaps[attempt], length);
+  }
+  return region;
+}
+
+#ifdef XP_WIN
+
+/*
+ * On Windows, map and unmap calls must be matched, so we deallocate the
+ * unaligned chunk, then reallocate the unaligned part to block off the
+ * old address and force the allocator to give us a new one.
+ */
+template <bool>
+static bool TryToAlignChunk(void** aRegion, void** aRetainedRegion,
+                            size_t length, size_t alignment) {
+  void* region = *aRegion;
+  MOZ_ASSERT(region && OffsetFromAligned(region, alignment) != 0);
+
+  size_t retainedLength = 0;
+  void* retainedRegion = nullptr;
+  do {
+    size_t offset = OffsetFromAligned(region, alignment);
+    if (offset == 0) {
+      // If the address is aligned, either we hit OOM or we're done.
+      break;
+    }
+    UnmapInternal(region, length);
+    retainedLength = alignment - offset;
+    retainedRegion = MapMemoryAt<Commit::No>(region, retainedLength);
+    region = MapMemory(length);
+
+    // If retainedRegion is null here, we raced with another thread.
+  } while (!retainedRegion);
+
+  bool result = OffsetFromAligned(region, alignment) == 0;
+  if (result && retainedRegion) {
+    UnmapInternal(retainedRegion, retainedLength);
+    retainedRegion = nullptr;
+  }
+
+  *aRegion = region;
+  *aRetainedRegion = retainedRegion;
+  return region && result;
+}
+
+#else  // !defined(XP_WIN)
+
+/*
+ * mmap calls don't have to be matched with calls to munmap, so we can unmap
+ * just the pages we don't need. However, as we don't know a priori if addresses
+ * are handed out in increasing or decreasing order, we have to try both
+ * directions (depending on the environment, one will always fail).
+ */
+template <bool AlwaysGetNew>
+static bool TryToAlignChunk(void** aRegion, void** aRetainedRegion,
+                            size_t length, size_t alignment) {
+  void* regionStart = *aRegion;
+  MOZ_ASSERT(regionStart && OffsetFromAligned(regionStart, alignment) != 0);
+
+  bool addressesGrowUpward = growthDirection > 0;
+  bool directionUncertain = -8 < growthDirection && growthDirection <= 8;
+  size_t offsetLower = OffsetFromAligned(regionStart, alignment);
+  size_t offsetUpper = alignment - offsetLower;
+  for (size_t i = 0; i < 2; ++i) {
+    if (addressesGrowUpward) {
+      void* upperStart =
+          reinterpret_cast<void*>(uintptr_t(regionStart) + offsetUpper);
+      void* regionEnd =
+          reinterpret_cast<void*>(uintptr_t(regionStart) + length);
+      if (MapMemoryAt(regionEnd, offsetUpper)) {
+        UnmapInternal(regionStart, offsetUpper);
+        if (directionUncertain) {
+          ++growthDirection;
+        }
+        regionStart = upperStart;
+        break;
+      }
+    } else {
+      auto* lowerStart =
+          reinterpret_cast<void*>(uintptr_t(regionStart) - offsetLower);
+      auto* lowerEnd = reinterpret_cast<void*>(uintptr_t(lowerStart) + length);
+      if (MapMemoryAt(lowerStart, offsetLower)) {
+        UnmapInternal(lowerEnd, offsetLower);
+        if (directionUncertain) {
+          --growthDirection;
+        }
+        regionStart = lowerStart;
+        break;
+      }
+    }
+    // If we're confident in the growth direction, don't try the other.
+    if (!directionUncertain) {
+      break;
+    }
+    addressesGrowUpward = !addressesGrowUpward;
+  }
+
+  void* retainedRegion = nullptr;
+  bool result = OffsetFromAligned(regionStart, alignment) == 0;
+  if (AlwaysGetNew && !result) {
+    // If our current chunk cannot be aligned, just get a new one.
+    retainedRegion = regionStart;
+    regionStart = MapMemory(length);
+    // Our new region might happen to already be aligned.
+    result = OffsetFromAligned(regionStart, alignment) == 0;
+    if (result) {
+      UnmapInternal(retainedRegion, length);
+      retainedRegion = nullptr;
+    }
+  }
+
+  *aRegion = regionStart;
+  *aRetainedRegion = retainedRegion;
+  return regionStart && result;
+}
+
+#endif
+
+void UnmapPages(void* region, size_t length) {
+  MOZ_RELEASE_ASSERT(region &&
+                     OffsetFromAligned(region, allocGranularity) == 0);
+  MOZ_RELEASE_ASSERT(length > 0 && length % pageSize == 0);
+
+  // ASan does not automatically unpoison memory, so we have to do this here.
+  MOZ_MAKE_MEM_UNDEFINED(region, length);
+
+  UnmapInternal(region, length);
+}
+
+static void CheckDecommit(void* region, size_t length) {
+  MOZ_RELEASE_ASSERT(region);
+  MOZ_RELEASE_ASSERT(length > 0);
+
+  // pageSize == ArenaSize doesn't necessarily hold, but this function is
+  // used by the GC to decommit unused Arenas, so we don't want to assert
+  // if pageSize > ArenaSize.
+  MOZ_ASSERT(OffsetFromAligned(region, ArenaSize) == 0);
+  MOZ_ASSERT(length % ArenaSize == 0);
+
+  MOZ_RELEASE_ASSERT(OffsetFromAligned(region, pageSize) == 0);
+  MOZ_RELEASE_ASSERT(length % pageSize == 0);
+}
+
+bool MarkPagesUnusedSoft(void* region, size_t length) {
+  MOZ_ASSERT(DecommitEnabled());
+  CheckDecommit(region, length);
+
+  MOZ_MAKE_MEM_NOACCESS(region, length);
+
+#if defined(XP_WIN)
+  return VirtualAlloc(region, length, MEM_RESET,
+                      DWORD(PageAccess::ReadWrite)) == region;
+#elif defined(__wasi__)
+  return 0;
+#else
+  int status;
+  do {
+#  if defined(XP_DARWIN)
+    status = madvise(region, length, MADV_FREE_REUSABLE);
+#  elif defined(XP_SOLARIS)
+    status = posix_madvise(region, length, POSIX_MADV_DONTNEED);
+#  else
+    status = madvise(region, length, MADV_DONTNEED);
+#  endif
+  } while (status == -1 && errno == EAGAIN);
+  return status == 0;
+#endif
+}
+
+bool MarkPagesUnusedHard(void* region, size_t length) {
+  CheckDecommit(region, length);
+
+  MOZ_MAKE_MEM_NOACCESS(region, length);
+
+  if (!DecommitEnabled()) {
+    return true;
+  }
+
+#if defined(XP_WIN)
+  return VirtualFree(region, length, MEM_DECOMMIT);
+#else
+  return MarkPagesUnusedSoft(region, length);
+#endif
+}
+
+void MarkPagesInUseSoft(void* region, size_t length) {
+  MOZ_ASSERT(DecommitEnabled());
+  CheckDecommit(region, length);
+
+#if defined(XP_DARWIN)
+  while (madvise(region, length, MADV_FREE_REUSE) == -1 && errno == EAGAIN) {
+  }
+#endif
+
+  MOZ_MAKE_MEM_UNDEFINED(region, length);
+}
+
+bool MarkPagesInUseHard(void* region, size_t length) {
+  if (js::oom::ShouldFailWithOOM()) {
+    return false;
+  }
+
+  CheckDecommit(region, length);
+
+  MOZ_MAKE_MEM_UNDEFINED(region, length);
+
+  if (!DecommitEnabled()) {
+    return true;
+  }
+
+#if defined(XP_WIN)
+  return VirtualAlloc(region, length, MEM_COMMIT,
+                      DWORD(PageAccess::ReadWrite)) == region;
+#else
+  return true;
+#endif
+}
+
+size_t GetPageFaultCount() {
+#ifdef XP_WIN
+  PROCESS_MEMORY_COUNTERS pmc;
+  if (GetProcessMemoryInfo(GetCurrentProcess(), &pmc, sizeof(pmc)) == 0) {
+    return 0;
+  }
+  return pmc.PageFaultCount;
+#elif defined(__wasi__)
+  return 0;
+#else
+  struct rusage usage;
+  int err = getrusage(RUSAGE_SELF, &usage);
+  if (err) {
+    return 0;
+  }
+  return usage.ru_majflt;
+#endif
+}
+
+void* AllocateMappedContent(int fd, size_t offset, size_t length,
+                            size_t alignment) {
+#ifdef __wasi__
+  MOZ_CRASH("Not yet supported for WASI");
+#else
+  if (length == 0 || alignment == 0 || offset % alignment != 0 ||
+      std::max(alignment, allocGranularity) %
+              std::min(alignment, allocGranularity) !=
+          0) {
+    return nullptr;
+  }
+
+  size_t alignedOffset = offset - (offset % allocGranularity);
+  size_t alignedLength = length + (offset % allocGranularity);
+
+  // We preallocate the mapping using MapAlignedPages, which expects
+  // the length parameter to be an integer multiple of the page size.
+  size_t mappedLength = alignedLength;
+  if (alignedLength % pageSize != 0) {
+    mappedLength += pageSize - alignedLength % pageSize;
+  }
+
+#  ifdef XP_WIN
+  HANDLE hFile = reinterpret_cast<HANDLE>(intptr_t(fd));
+
+  // This call will fail if the file does not exist.
+  HANDLE hMap =
+      CreateFileMappingW(hFile, nullptr, PAGE_READONLY, 0, 0, nullptr);
+  if (!hMap) {
+    return nullptr;
+  }
+
+  DWORD offsetH = uint32_t(uint64_t(alignedOffset) >> 32);
+  DWORD offsetL = uint32_t(alignedOffset);
+
+  uint8_t* map = nullptr;
+  for (;;) {
+    // The value of a pointer is technically only defined while the region
+    // it points to is allocated, so explicitly treat this one as a number.
+    uintptr_t region = uintptr_t(MapAlignedPages(mappedLength, alignment));
+    if (region == 0) {
+      break;
+    }
+    UnmapInternal(reinterpret_cast<void*>(region), mappedLength);
+    // If the offset or length are out of bounds, this call will fail.
+    map = static_cast<uint8_t*>(
+        MapViewOfFileEx(hMap, FILE_MAP_COPY, offsetH, offsetL, alignedLength,
+                        reinterpret_cast<void*>(region)));
+
+    // Retry if another thread mapped the address we were trying to use.
+    if (map || GetLastError() != ERROR_INVALID_ADDRESS) {
+      break;
+    }
+  }
+
+  // This just decreases the file mapping object's internal reference count;
+  // it won't actually be destroyed until we unmap the associated view.
+  CloseHandle(hMap);
+
+  if (!map) {
+    return nullptr;
+  }
+#  else  // !defined(XP_WIN)
+  // Sanity check the offset and length, as mmap does not do this for us.
+  struct stat st;
+  if (fstat(fd, &st) || offset >= uint64_t(st.st_size) ||
+      length > uint64_t(st.st_size) - offset) {
+    return nullptr;
+  }
+
+  void* region = MapAlignedPages(mappedLength, alignment);
+  if (!region) {
+    return nullptr;
+  }
+
+  // Calling mmap with MAP_FIXED will replace the previous mapping, allowing
+  // us to reuse the region we obtained without racing with other threads.
+  uint8_t* map =
+      static_cast<uint8_t*>(mmap(region, alignedLength, PROT_READ | PROT_WRITE,
+                                 MAP_PRIVATE | MAP_FIXED, fd, alignedOffset));
+  if (map == MAP_FAILED) {
+    UnmapInternal(region, mappedLength);
+    return nullptr;
+  }
+#  endif
+
+#  ifdef DEBUG
+  // Zero out data before and after the desired mapping to catch errors early.
+  if (offset != alignedOffset) {
+    memset(map, 0, offset - alignedOffset);
+  }
+  if (alignedLength % pageSize) {
+    memset(map + alignedLength, 0, pageSize - (alignedLength % pageSize));
+  }
+#  endif
+
+  return map + (offset - alignedOffset);
+#endif  // __wasi__
+}
+
+void DeallocateMappedContent(void* region, size_t length) {
+#ifdef __wasi__
+  MOZ_CRASH("Not yet supported for WASI");
+#else
+  if (!region) {
+    return;
+  }
+
+  // Due to bug 1502562, the following assertion does not currently hold.
+  // MOZ_RELEASE_ASSERT(length > 0);
+
+  // Calculate the address originally returned by the system call.
+  // This is needed because AllocateMappedContent returns a pointer
+  // that might be offset from the mapping, as the beginning of a
+  // mapping must be aligned with the allocation granularity.
+  uintptr_t map = uintptr_t(region) - (uintptr_t(region) % allocGranularity);
+#  ifdef XP_WIN
+  MOZ_RELEASE_ASSERT(UnmapViewOfFile(reinterpret_cast<void*>(map)) != 0);
+#  else
+  size_t alignedLength = length + (uintptr_t(region) % allocGranularity);
+  if (munmap(reinterpret_cast<void*>(map), alignedLength)) {
+    MOZ_RELEASE_ASSERT(errno == ENOMEM);
+  }
+#  endif
+#endif  // __wasi__
+}
+
+static inline void ProtectMemory(void* region, size_t length, PageAccess prot) {
+  MOZ_RELEASE_ASSERT(region && OffsetFromAligned(region, pageSize) == 0);
+  MOZ_RELEASE_ASSERT(length > 0 && length % pageSize == 0);
+#ifdef XP_WIN
+  DWORD oldProtect;
+  MOZ_RELEASE_ASSERT(VirtualProtect(region, length, DWORD(prot), &oldProtect) !=
+                     0);
+#elif defined(__wasi__)
+  /* nothing */
+#else
+  MOZ_RELEASE_ASSERT(mprotect(region, length, int(prot)) == 0);
+#endif
+}
+
+void ProtectPages(void* region, size_t length) {
+  ProtectMemory(region, length, PageAccess::None);
+}
+
+void MakePagesReadOnly(void* region, size_t length) {
+  ProtectMemory(region, length, PageAccess::Read);
+}
+
+void UnprotectPages(void* region, size_t length) {
+  ProtectMemory(region, length, PageAccess::ReadWrite);
+}
+
+}  // namespace gc
+}  // namespace js