Source code
Revision control
Copy as Markdown
Other Tools
// Copyright 2017 The Abseil Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// A low-level allocator that can be used by other low-level
// modules without introducing dependency cycles.
// This allocator is slow and wasteful of memory;
// it should not be used when performance is key.
#include "absl/base/internal/low_level_alloc.h"
#include <type_traits>
#include "absl/base/call_once.h"
#include "absl/base/config.h"
#include "absl/base/internal/direct_mmap.h"
#include "absl/base/internal/scheduling_mode.h"
#include "absl/base/macros.h"
#include "absl/base/thread_annotations.h"
// LowLevelAlloc requires that the platform support low-level
// allocation of virtual memory. Platforms lacking this cannot use
// LowLevelAlloc.
#ifndef ABSL_LOW_LEVEL_ALLOC_MISSING
#ifndef _WIN32
#include <pthread.h>
#include <signal.h>
#include <sys/mman.h>
#include <unistd.h>
#else
#include <windows.h>
#endif
#ifdef __linux__
#include <sys/prctl.h>
#endif
#include <string.h>
#include <algorithm>
#include <atomic>
#include <cerrno>
#include <cstddef>
#include <new> // for placement-new
#include "absl/base/dynamic_annotations.h"
#include "absl/base/internal/raw_logging.h"
#include "absl/base/internal/spinlock.h"
#if defined(MAP_ANON) && !defined(MAP_ANONYMOUS)
#define MAP_ANONYMOUS MAP_ANON
#endif
namespace absl {
ABSL_NAMESPACE_BEGIN
namespace base_internal {
// A first-fit allocator with amortized logarithmic free() time.
// ---------------------------------------------------------------------------
static const int kMaxLevel = 30;
namespace {
// This struct describes one allocated block, or one free block.
struct AllocList {
struct Header {
// Size of entire region, including this field. Must be
// first. Valid in both allocated and unallocated blocks.
uintptr_t size;
// kMagicAllocated or kMagicUnallocated xor this.
uintptr_t magic;
// Pointer to parent arena.
LowLevelAlloc::Arena *arena;
// Aligns regions to 0 mod 2*sizeof(void*).
void *dummy_for_alignment;
} header;
// Next two fields: in unallocated blocks: freelist skiplist data
// in allocated blocks: overlaps with client data
// Levels in skiplist used.
int levels;
// Actually has levels elements. The AllocList node may not have room
// for all kMaxLevel entries. See max_fit in LLA_SkiplistLevels().
AllocList *next[kMaxLevel];
};
} // namespace
// ---------------------------------------------------------------------------
// A trivial skiplist implementation. This is used to keep the freelist
// in address order while taking only logarithmic time per insert and delete.
// An integer approximation of log2(size/base)
// Requires size >= base.
static int IntLog2(size_t size, size_t base) {
int result = 0;
for (size_t i = size; i > base; i >>= 1) { // i == floor(size/2**result)
result++;
}
// floor(size / 2**result) <= base < floor(size / 2**(result-1))
// => log2(size/(base+1)) <= result < 1+log2(size/base)
// => result ~= log2(size/base)
return result;
}
// Return a random integer n: p(n)=1/(2**n) if 1 <= n; p(n)=0 if n < 1.
static int Random(uint32_t *state) {
uint32_t r = *state;
int result = 1;
while ((((r = r * 1103515245 + 12345) >> 30) & 1) == 0) {
result++;
}
*state = r;
return result;
}
// Return a number of skiplist levels for a node of size bytes, where
// base is the minimum node size. Compute level=log2(size / base)+n
// where n is 1 if random is false and otherwise a random number generated with
// the standard distribution for a skiplist: See Random() above.
// Bigger nodes tend to have more skiplist levels due to the log2(size / base)
// term, so first-fit searches touch fewer nodes. "level" is clipped so
// level<kMaxLevel and next[level-1] will fit in the node.
// 0 < LLA_SkiplistLevels(x,y,false) <= LLA_SkiplistLevels(x,y,true) < kMaxLevel
static int LLA_SkiplistLevels(size_t size, size_t base, uint32_t *random) {
// max_fit is the maximum number of levels that will fit in a node for the
// given size. We can't return more than max_fit, no matter what the
// random number generator says.
size_t max_fit = (size - offsetof(AllocList, next)) / sizeof(AllocList *);
int level = IntLog2(size, base) + (random != nullptr ? Random(random) : 1);
if (static_cast<size_t>(level) > max_fit) level = static_cast<int>(max_fit);
if (level > kMaxLevel - 1) level = kMaxLevel - 1;
ABSL_RAW_CHECK(level >= 1, "block not big enough for even one level");
return level;
}
// Return "atleast", the first element of AllocList *head s.t. *atleast >= *e.
// For 0 <= i < head->levels, set prev[i] to "no_greater", where no_greater
// points to the last element at level i in the AllocList less than *e, or is
// head if no such element exists.
static AllocList *LLA_SkiplistSearch(AllocList *head, AllocList *e,
AllocList **prev) {
AllocList *p = head;
for (int level = head->levels - 1; level >= 0; level--) {
for (AllocList *n; (n = p->next[level]) != nullptr && n < e; p = n) {
}
prev[level] = p;
}
return (head->levels == 0) ? nullptr : prev[0]->next[0];
}
// Insert element *e into AllocList *head. Set prev[] as LLA_SkiplistSearch.
// Requires that e->levels be previously set by the caller (using
// LLA_SkiplistLevels())
static void LLA_SkiplistInsert(AllocList *head, AllocList *e,
AllocList **prev) {
LLA_SkiplistSearch(head, e, prev);
for (; head->levels < e->levels; head->levels++) { // extend prev pointers
prev[head->levels] = head; // to all *e's levels
}
for (int i = 0; i != e->levels; i++) { // add element to list
e->next[i] = prev[i]->next[i];
prev[i]->next[i] = e;
}
}
// Remove element *e from AllocList *head. Set prev[] as LLA_SkiplistSearch().
// Requires that e->levels be previous set by the caller (using
// LLA_SkiplistLevels())
static void LLA_SkiplistDelete(AllocList *head, AllocList *e,
AllocList **prev) {
AllocList *found = LLA_SkiplistSearch(head, e, prev);
ABSL_RAW_CHECK(e == found, "element not in freelist");
for (int i = 0; i != e->levels && prev[i]->next[i] == e; i++) {
prev[i]->next[i] = e->next[i];
}
while (head->levels > 0 && head->next[head->levels - 1] == nullptr) {
head->levels--; // reduce head->levels if level unused
}
}
// ---------------------------------------------------------------------------
// Arena implementation
// Metadata for an LowLevelAlloc arena instance.
struct LowLevelAlloc::Arena {
// Constructs an arena with the given LowLevelAlloc flags.
explicit Arena(uint32_t flags_value);
base_internal::SpinLock mu;
// Head of free list, sorted by address
AllocList freelist ABSL_GUARDED_BY(mu);
// Count of allocated blocks
int32_t allocation_count ABSL_GUARDED_BY(mu);
// flags passed to NewArena
const uint32_t flags;
// Result of sysconf(_SC_PAGESIZE)
const size_t pagesize;
// Lowest power of two >= max(16, sizeof(AllocList))
const size_t round_up;
// Smallest allocation block size
const size_t min_size;
// PRNG state
uint32_t random ABSL_GUARDED_BY(mu);
};
namespace {
// Static storage space for the lazily-constructed, default global arena
// instances. We require this space because the whole point of LowLevelAlloc
// is to avoid relying on malloc/new.
alignas(LowLevelAlloc::Arena) unsigned char default_arena_storage[sizeof(
LowLevelAlloc::Arena)];
alignas(LowLevelAlloc::Arena) unsigned char unhooked_arena_storage[sizeof(
LowLevelAlloc::Arena)];
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
alignas(
LowLevelAlloc::Arena) unsigned char unhooked_async_sig_safe_arena_storage
[sizeof(LowLevelAlloc::Arena)];
#endif
// We must use LowLevelCallOnce here to construct the global arenas, rather than
// using function-level statics, to avoid recursively invoking the scheduler.
absl::once_flag create_globals_once;
void CreateGlobalArenas() {
new (&default_arena_storage)
LowLevelAlloc::Arena(LowLevelAlloc::kCallMallocHook);
new (&unhooked_arena_storage) LowLevelAlloc::Arena(0);
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
new (&unhooked_async_sig_safe_arena_storage)
LowLevelAlloc::Arena(LowLevelAlloc::kAsyncSignalSafe);
#endif
}
// Returns a global arena that does not call into hooks. Used by NewArena()
// when kCallMallocHook is not set.
LowLevelAlloc::Arena *UnhookedArena() {
base_internal::LowLevelCallOnce(&create_globals_once, CreateGlobalArenas);
return reinterpret_cast<LowLevelAlloc::Arena *>(&unhooked_arena_storage);
}
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
// Returns a global arena that is async-signal safe. Used by NewArena() when
// kAsyncSignalSafe is set.
LowLevelAlloc::Arena *UnhookedAsyncSigSafeArena() {
base_internal::LowLevelCallOnce(&create_globals_once, CreateGlobalArenas);
return reinterpret_cast<LowLevelAlloc::Arena *>(
&unhooked_async_sig_safe_arena_storage);
}
#endif
} // namespace
// Returns the default arena, as used by LowLevelAlloc::Alloc() and friends.
LowLevelAlloc::Arena *LowLevelAlloc::DefaultArena() {
base_internal::LowLevelCallOnce(&create_globals_once, CreateGlobalArenas);
return reinterpret_cast<LowLevelAlloc::Arena *>(&default_arena_storage);
}
// magic numbers to identify allocated and unallocated blocks
static const uintptr_t kMagicAllocated = 0x4c833e95U;
static const uintptr_t kMagicUnallocated = ~kMagicAllocated;
namespace {
class ABSL_SCOPED_LOCKABLE ArenaLock {
public:
explicit ArenaLock(LowLevelAlloc::Arena *arena)
ABSL_EXCLUSIVE_LOCK_FUNCTION(arena->mu)
: arena_(arena) {
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
if ((arena->flags & LowLevelAlloc::kAsyncSignalSafe) != 0) {
sigset_t all;
sigfillset(&all);
mask_valid_ = pthread_sigmask(SIG_BLOCK, &all, &mask_) == 0;
}
#endif
arena_->mu.Lock();
}
~ArenaLock() { ABSL_RAW_CHECK(left_, "haven't left Arena region"); }
void Leave() ABSL_UNLOCK_FUNCTION() {
arena_->mu.Unlock();
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
if (mask_valid_) {
const int err = pthread_sigmask(SIG_SETMASK, &mask_, nullptr);
if (err != 0) {
ABSL_RAW_LOG(FATAL, "pthread_sigmask failed: %d", err);
}
}
#endif
left_ = true;
}
private:
bool left_ = false; // whether left region
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
bool mask_valid_ = false;
sigset_t mask_; // old mask of blocked signals
#endif
LowLevelAlloc::Arena *arena_;
ArenaLock(const ArenaLock &) = delete;
ArenaLock &operator=(const ArenaLock &) = delete;
};
} // namespace
// create an appropriate magic number for an object at "ptr"
// "magic" should be kMagicAllocated or kMagicUnallocated
inline static uintptr_t Magic(uintptr_t magic, AllocList::Header *ptr) {
return magic ^ reinterpret_cast<uintptr_t>(ptr);
}
namespace {
size_t GetPageSize() {
#ifdef _WIN32
SYSTEM_INFO system_info;
GetSystemInfo(&system_info);
return std::max(system_info.dwPageSize, system_info.dwAllocationGranularity);
#elif defined(__wasm__) || defined(__asmjs__)
return getpagesize();
#else
return static_cast<size_t>(sysconf(_SC_PAGESIZE));
#endif
}
size_t RoundedUpBlockSize() {
// Round up block sizes to a power of two close to the header size.
size_t round_up = 16;
while (round_up < sizeof(AllocList::Header)) {
round_up += round_up;
}
return round_up;
}
} // namespace
LowLevelAlloc::Arena::Arena(uint32_t flags_value)
: mu(base_internal::SCHEDULE_KERNEL_ONLY),
allocation_count(0),
flags(flags_value),
pagesize(GetPageSize()),
round_up(RoundedUpBlockSize()),
min_size(2 * round_up),
random(0) {
freelist.header.size = 0;
freelist.header.magic = Magic(kMagicUnallocated, &freelist.header);
freelist.header.arena = this;
freelist.levels = 0;
memset(freelist.next, 0, sizeof(freelist.next));
}
// L < meta_data_arena->mu
LowLevelAlloc::Arena *LowLevelAlloc::NewArena(uint32_t flags) {
Arena *meta_data_arena = DefaultArena();
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
if ((flags & LowLevelAlloc::kAsyncSignalSafe) != 0) {
meta_data_arena = UnhookedAsyncSigSafeArena();
} else // NOLINT(readability/braces)
#endif
if ((flags & LowLevelAlloc::kCallMallocHook) == 0) {
meta_data_arena = UnhookedArena();
}
Arena *result =
new (AllocWithArena(sizeof(*result), meta_data_arena)) Arena(flags);
return result;
}
// L < arena->mu, L < arena->arena->mu
bool LowLevelAlloc::DeleteArena(Arena *arena) {
ABSL_RAW_CHECK(
arena != nullptr && arena != DefaultArena() && arena != UnhookedArena(),
"may not delete default arena");
ArenaLock section(arena);
if (arena->allocation_count != 0) {
section.Leave();
return false;
}
while (arena->freelist.next[0] != nullptr) {
AllocList *region = arena->freelist.next[0];
size_t size = region->header.size;
arena->freelist.next[0] = region->next[0];
ABSL_RAW_CHECK(
region->header.magic == Magic(kMagicUnallocated, ®ion->header),
"bad magic number in DeleteArena()");
ABSL_RAW_CHECK(region->header.arena == arena,
"bad arena pointer in DeleteArena()");
ABSL_RAW_CHECK(size % arena->pagesize == 0,
"empty arena has non-page-aligned block size");
ABSL_RAW_CHECK(reinterpret_cast<uintptr_t>(region) % arena->pagesize == 0,
"empty arena has non-page-aligned block");
int munmap_result;
#ifdef _WIN32
munmap_result = VirtualFree(region, 0, MEM_RELEASE);
ABSL_RAW_CHECK(munmap_result != 0,
"LowLevelAlloc::DeleteArena: VitualFree failed");
#else
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
if ((arena->flags & LowLevelAlloc::kAsyncSignalSafe) == 0) {
munmap_result = munmap(region, size);
} else {
munmap_result = base_internal::DirectMunmap(region, size);
}
#else
munmap_result = munmap(region, size);
#endif // ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
if (munmap_result != 0) {
ABSL_RAW_LOG(FATAL, "LowLevelAlloc::DeleteArena: munmap failed: %d",
errno);
}
#endif // _WIN32
}
section.Leave();
arena->~Arena();
Free(arena);
return true;
}
// ---------------------------------------------------------------------------
// Addition, checking for overflow. The intent is to die if an external client
// manages to push through a request that would cause arithmetic to fail.
static inline uintptr_t CheckedAdd(uintptr_t a, uintptr_t b) {
uintptr_t sum = a + b;
ABSL_RAW_CHECK(sum >= a, "LowLevelAlloc arithmetic overflow");
return sum;
}
// Return value rounded up to next multiple of align.
// align must be a power of two.
static inline uintptr_t RoundUp(uintptr_t addr, uintptr_t align) {
return CheckedAdd(addr, align - 1) & ~(align - 1);
}
// Equivalent to "return prev->next[i]" but with sanity checking
// that the freelist is in the correct order, that it
// consists of regions marked "unallocated", and that no two regions
// are adjacent in memory (they should have been coalesced).
// L >= arena->mu
static AllocList *Next(int i, AllocList *prev, LowLevelAlloc::Arena *arena) {
ABSL_RAW_CHECK(i < prev->levels, "too few levels in Next()");
AllocList *next = prev->next[i];
if (next != nullptr) {
ABSL_RAW_CHECK(
next->header.magic == Magic(kMagicUnallocated, &next->header),
"bad magic number in Next()");
ABSL_RAW_CHECK(next->header.arena == arena, "bad arena pointer in Next()");
if (prev != &arena->freelist) {
ABSL_RAW_CHECK(prev < next, "unordered freelist");
ABSL_RAW_CHECK(reinterpret_cast<char *>(prev) + prev->header.size <
reinterpret_cast<char *>(next),
"malformed freelist");
}
}
return next;
}
// Coalesce list item "a" with its successor if they are adjacent.
static void Coalesce(AllocList *a) {
AllocList *n = a->next[0];
if (n != nullptr && reinterpret_cast<char *>(a) + a->header.size ==
reinterpret_cast<char *>(n)) {
LowLevelAlloc::Arena *arena = a->header.arena;
a->header.size += n->header.size;
n->header.magic = 0;
n->header.arena = nullptr;
AllocList *prev[kMaxLevel];
LLA_SkiplistDelete(&arena->freelist, n, prev);
LLA_SkiplistDelete(&arena->freelist, a, prev);
a->levels =
LLA_SkiplistLevels(a->header.size, arena->min_size, &arena->random);
LLA_SkiplistInsert(&arena->freelist, a, prev);
}
}
// Adds block at location "v" to the free list
// L >= arena->mu
static void AddToFreelist(void *v, LowLevelAlloc::Arena *arena) {
AllocList *f = reinterpret_cast<AllocList *>(reinterpret_cast<char *>(v) -
sizeof(f->header));
ABSL_RAW_CHECK(f->header.magic == Magic(kMagicAllocated, &f->header),
"bad magic number in AddToFreelist()");
ABSL_RAW_CHECK(f->header.arena == arena,
"bad arena pointer in AddToFreelist()");
f->levels =
LLA_SkiplistLevels(f->header.size, arena->min_size, &arena->random);
AllocList *prev[kMaxLevel];
LLA_SkiplistInsert(&arena->freelist, f, prev);
f->header.magic = Magic(kMagicUnallocated, &f->header);
Coalesce(f); // maybe coalesce with successor
Coalesce(prev[0]); // maybe coalesce with predecessor
}
// Frees storage allocated by LowLevelAlloc::Alloc().
// L < arena->mu
void LowLevelAlloc::Free(void *v) {
if (v != nullptr) {
AllocList *f = reinterpret_cast<AllocList *>(reinterpret_cast<char *>(v) -
sizeof(f->header));
LowLevelAlloc::Arena *arena = f->header.arena;
ArenaLock section(arena);
AddToFreelist(v, arena);
ABSL_RAW_CHECK(arena->allocation_count > 0, "nothing in arena to free");
arena->allocation_count--;
section.Leave();
}
}
// allocates and returns a block of size bytes, to be freed with Free()
// L < arena->mu
static void *DoAllocWithArena(size_t request, LowLevelAlloc::Arena *arena) {
void *result = nullptr;
if (request != 0) {
AllocList *s; // will point to region that satisfies request
ArenaLock section(arena);
// round up with header
size_t req_rnd =
RoundUp(CheckedAdd(request, sizeof(s->header)), arena->round_up);
for (;;) { // loop until we find a suitable region
// find the minimum levels that a block of this size must have
int i = LLA_SkiplistLevels(req_rnd, arena->min_size, nullptr) - 1;
if (i < arena->freelist.levels) { // potential blocks exist
AllocList *before = &arena->freelist; // predecessor of s
while ((s = Next(i, before, arena)) != nullptr &&
s->header.size < req_rnd) {
before = s;
}
if (s != nullptr) { // we found a region
break;
}
}
// we unlock before mmap() both because mmap() may call a callback hook,
// and because it may be slow.
arena->mu.Unlock();
// mmap generous 64K chunks to decrease
// the chances/impact of fragmentation:
size_t new_pages_size = RoundUp(req_rnd, arena->pagesize * 16);
void *new_pages;
#ifdef _WIN32
new_pages = VirtualAlloc(nullptr, new_pages_size,
MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
ABSL_RAW_CHECK(new_pages != nullptr, "VirtualAlloc failed");
#else
#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
if ((arena->flags & LowLevelAlloc::kAsyncSignalSafe) != 0) {
new_pages = base_internal::DirectMmap(nullptr, new_pages_size,
PROT_WRITE|PROT_READ, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
} else {
new_pages = mmap(nullptr, new_pages_size, PROT_WRITE | PROT_READ,
MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
}
#else
new_pages = mmap(nullptr, new_pages_size, PROT_WRITE | PROT_READ,
MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
#endif // ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
if (new_pages == MAP_FAILED) {
ABSL_RAW_LOG(FATAL, "mmap error: %d", errno);
}
#ifdef __linux__
#if defined(PR_SET_VMA) && defined(PR_SET_VMA_ANON_NAME)
// Attempt to name the allocated address range in /proc/$PID/smaps on
// Linux.
//
// This invocation of prctl() may fail if the Linux kernel was not
// configured with the CONFIG_ANON_VMA_NAME option. This is OK since
// the naming of arenas is primarily a debugging aid.
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, new_pages, new_pages_size,
"absl");
#endif
#endif // __linux__
#endif // _WIN32
arena->mu.Lock();
s = reinterpret_cast<AllocList *>(new_pages);
s->header.size = new_pages_size;
// Pretend the block is allocated; call AddToFreelist() to free it.
s->header.magic = Magic(kMagicAllocated, &s->header);
s->header.arena = arena;
AddToFreelist(&s->levels, arena); // insert new region into free list
}
AllocList *prev[kMaxLevel];
LLA_SkiplistDelete(&arena->freelist, s, prev); // remove from free list
// s points to the first free region that's big enough
if (CheckedAdd(req_rnd, arena->min_size) <= s->header.size) {
// big enough to split
AllocList *n =
reinterpret_cast<AllocList *>(req_rnd + reinterpret_cast<char *>(s));
n->header.size = s->header.size - req_rnd;
n->header.magic = Magic(kMagicAllocated, &n->header);
n->header.arena = arena;
s->header.size = req_rnd;
AddToFreelist(&n->levels, arena);
}
s->header.magic = Magic(kMagicAllocated, &s->header);
ABSL_RAW_CHECK(s->header.arena == arena, "");
arena->allocation_count++;
section.Leave();
result = &s->levels;
}
ABSL_ANNOTATE_MEMORY_IS_UNINITIALIZED(result, request);
return result;
}
void *LowLevelAlloc::Alloc(size_t request) {
void *result = DoAllocWithArena(request, DefaultArena());
return result;
}
void *LowLevelAlloc::AllocWithArena(size_t request, Arena *arena) {
ABSL_RAW_CHECK(arena != nullptr, "must pass a valid arena");
void *result = DoAllocWithArena(request, arena);
return result;
}
} // namespace base_internal
ABSL_NAMESPACE_END
} // namespace absl
#endif // ABSL_LOW_LEVEL_ALLOC_MISSING