marinone94's picture
Training in progress, epoch 0
1ce325b
raw
history blame
13.5 kB
/* Memory mapping wrappers.
* ARM and MinGW ports contributed by Hideo Okuma and Tomoyuki Yoshimura at
* NICT.
*/
#include "mmap.hh"
#include "exception.hh"
#include "file.hh"
#include "scoped.hh"
#include <iostream>
#include <cassert>
#include <fcntl.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <cstdlib>
#if defined(_WIN32) || defined(_WIN64)
#include <windows.h>
#include <io.h>
#else
#include <sys/mman.h>
#include <unistd.h>
#endif
namespace util {
std::size_t SizePage() {
#if defined(_WIN32) || defined(_WIN64)
SYSTEM_INFO si;
GetSystemInfo(&si);
return si.dwAllocationGranularity;
#else
return sysconf(_SC_PAGE_SIZE);
#endif
}
scoped_mmap::~scoped_mmap() {
if (data_ != (void*)-1) {
try {
// Thanks Denis Filimonov for pointing out NFS likes msync first.
SyncOrThrow(data_, size_);
UnmapOrThrow(data_, size_);
} catch (const util::ErrnoException &e) {
std::cerr << e.what();
abort();
}
}
}
namespace {
template <class T> T RoundUpPow2(T value, T mult) {
return ((value - 1) & ~(mult - 1)) + mult;
}
std::size_t RoundUpSize(const scoped_memory &mem) {
switch(mem.source()) {
case scoped_memory::MMAP_ROUND_1G_ALLOCATED:
return RoundUpPow2<std::size_t>(mem.size(), 1ULL << 30);
case scoped_memory::MMAP_ROUND_2M_ALLOCATED:
return RoundUpPow2<std::size_t>(mem.size(), 1ULL << 21);
case scoped_memory::MMAP_ROUND_PAGE_ALLOCATED:
return RoundUpPow2<std::size_t>(mem.size(), static_cast<std::size_t>(SizePage()));
default:
return mem.size();
}
}
} // namespace
scoped_memory::scoped_memory(std::size_t size, bool zeroed) : data_(NULL), size_(0), source_(NONE_ALLOCATED) {
HugeMalloc(size, zeroed, *this);
}
void scoped_memory::reset(void *data, std::size_t size, Alloc source) {
switch(source_) {
case MMAP_ROUND_1G_ALLOCATED:
case MMAP_ROUND_2M_ALLOCATED:
case MMAP_ROUND_PAGE_ALLOCATED:
case MMAP_ALLOCATED:
scoped_mmap(data_, RoundUpSize(*this));
break;
case MALLOC_ALLOCATED:
free(data_);
break;
case NONE_ALLOCATED:
break;
}
data_ = data;
size_ = size;
source_ = source;
}
const int kFileFlags =
#if defined(_WIN32) || defined(_WIN64)
0 // MapOrThrow ignores flags on windows
#elif defined(MAP_FILE)
MAP_FILE | MAP_SHARED
#else
MAP_SHARED
#endif
;
void *MapOrThrow(std::size_t size, bool for_write, int flags, bool prefault, int fd, uint64_t offset) {
#ifdef MAP_POPULATE // Linux specific
if (prefault) {
flags |= MAP_POPULATE;
}
#endif
#if defined(_WIN32) || defined(_WIN64)
int protectC = for_write ? PAGE_READWRITE : PAGE_READONLY;
int protectM = for_write ? FILE_MAP_WRITE : FILE_MAP_READ;
uint64_t total_size = size + offset;
HANDLE hMapping = CreateFileMapping((HANDLE)_get_osfhandle(fd), NULL, protectC, total_size >> 32, static_cast<DWORD>(total_size), NULL);
UTIL_THROW_IF(!hMapping, ErrnoException, "CreateFileMapping failed");
LPVOID ret = MapViewOfFile(hMapping, protectM, offset >> 32, offset, size);
CloseHandle(hMapping);
UTIL_THROW_IF(!ret, ErrnoException, "MapViewOfFile failed");
#else
int protect = for_write ? (PROT_READ | PROT_WRITE) : PROT_READ;
void *ret;
UTIL_THROW_IF((ret = mmap(NULL, size, protect, flags, fd, offset)) == MAP_FAILED, ErrnoException, "mmap failed for size " << size << " at offset " << offset);
# ifdef MADV_HUGEPAGE
/* We like huge pages but it's fine if we can't have them. Note that huge
* pages are not supported for file-backed mmap on linux.
*/
madvise(ret, size, MADV_HUGEPAGE);
# endif
#endif
return ret;
}
void SyncOrThrow(void *start, size_t length) {
#if defined(_WIN32) || defined(_WIN64)
UTIL_THROW_IF(!::FlushViewOfFile(start, length), ErrnoException, "Failed to sync mmap");
#else
UTIL_THROW_IF(length && msync(start, length, MS_SYNC), ErrnoException, "Failed to sync mmap");
#endif
}
void UnmapOrThrow(void *start, size_t length) {
#if defined(_WIN32) || defined(_WIN64)
UTIL_THROW_IF(!::UnmapViewOfFile(start), ErrnoException, "Failed to unmap a file");
#else
UTIL_THROW_IF(munmap(start, length), ErrnoException, "munmap failed with " << start << " for length " << length);
#endif
}
// Linux huge pages.
#ifdef __linux__
namespace {
bool TryHuge(std::size_t size, bool populate, uint8_t alignment_bits, scoped_memory::Alloc huge_scheme, scoped_memory &to) {
// Don't bother with these cases.
if (size < (1ULL << alignment_bits) || (1ULL << alignment_bits) < SizePage())
return false;
// First try: Linux >= 3.8 with manually configured hugetlb pages available.
int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB | (alignment_bits << 26 /* This is MAP_HUGE_SHIFT but some headers are too old. */);
if (populate) flags |= MAP_POPULATE;
void *ret = mmap(NULL, size, PROT_READ | PROT_WRITE, flags, -1, 0);
if (ret != MAP_FAILED) {
to.reset(ret, size, huge_scheme);
return true;
}
// There weren't pages in a sysadmin-created pool. Let's get aligned memory
// and hope transparent huge pages kicks in. Align to a multiple of the huge
// page size by overallocating. I feel bad about doing this, but it's also how
// posix_memalign is implemented. And the memory is virtual.
// Round up requested size to multiple of page size. This will allow the pages after to be munmapped.
std::size_t size_up = RoundUpPow2(size, SizePage());
std::size_t ask = size_up + (1 << alignment_bits) - SizePage();
// Don't populate because this is asking for more than we will use.
scoped_mmap larger(mmap(NULL, ask, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0), ask);
if (larger.get() == MAP_FAILED) return false;
// Throw out pages before the alignment point.
uintptr_t base = reinterpret_cast<uintptr_t>(larger.get());
// Round up to next multiple of alignment.
uintptr_t rounded_up = RoundUpPow2(base, static_cast<uintptr_t>(1) << alignment_bits);
if (base != rounded_up) {
// If this throws an exception (which it shouldn't) then we want to unmap the whole thing by keeping it in larger.
UnmapOrThrow(larger.get(), rounded_up - base);
larger.steal();
larger.reset(reinterpret_cast<void*>(rounded_up), ask - (rounded_up - base));
}
// Throw out pages after the requested size.
assert(larger.size() >= size_up);
if (larger.size() > size_up) {
// This is where we assume size_up is a multiple of page size.
UnmapOrThrow(static_cast<uint8_t*>(larger.get()) + size_up, larger.size() - size_up);
larger.reset(larger.steal(), size_up);
}
#ifdef MADV_HUGEPAGE
madvise(larger.get(), size_up, MADV_HUGEPAGE);
#endif
to.reset(larger.steal(), size, scoped_memory::MMAP_ROUND_PAGE_ALLOCATED);
return true;
}
} // namespace
#endif
void HugeMalloc(std::size_t size, bool zeroed, scoped_memory &to) {
to.reset();
#ifdef __linux__
// TODO: architectures/page sizes other than 2^21 and 2^30.
// Attempt 1 GB pages.
// If the user asked for zeroed memory, assume they want it populated.
if (size >= (1ULL << 30) && TryHuge(size, zeroed, 30, scoped_memory::MMAP_ROUND_1G_ALLOCATED, to))
return;
// Attempt 2 MB pages.
if (size >= (1ULL << 21) && TryHuge(size, zeroed, 21, scoped_memory::MMAP_ROUND_2M_ALLOCATED, to))
return;
#endif // __linux__
// Non-linux will always do this, as will small allocations on Linux.
to.reset(zeroed ? calloc(1, size) : malloc(size), size, scoped_memory::MALLOC_ALLOCATED);
UTIL_THROW_IF(!to.get(), ErrnoException, "Failed to allocate " << size << " bytes");
}
namespace {
#ifdef __linux__
const std::size_t kTransitionHuge = std::max<std::size_t>(1ULL << 21, SizePage());
#endif // __linux__
void ReplaceAndCopy(std::size_t to, bool zero_new, scoped_memory &mem) {
scoped_memory replacement;
HugeMalloc(to, zero_new, replacement);
memcpy(replacement.get(), mem.get(), mem.size());
// This can't throw.
mem.reset(replacement.get(), replacement.size(), replacement.source());
replacement.steal();
}
} // namespace
void HugeRealloc(std::size_t to, bool zero_new, scoped_memory &mem) {
if (!to) {
mem.reset();
return;
}
switch (mem.source()) {
case scoped_memory::NONE_ALLOCATED:
HugeMalloc(to, zero_new, mem);
return;
#ifdef __linux__
// TODO really need to collapse these cases with a number.
case scoped_memory::MMAP_ROUND_1G_ALLOCATED:
case scoped_memory::MMAP_ROUND_2M_ALLOCATED:
case scoped_memory::MMAP_ROUND_PAGE_ALLOCATED:
case scoped_memory::MMAP_ALLOCATED:
// Downsizing below barrier?
if (to <= SizePage()) {
scoped_malloc replacement(malloc(to));
memcpy(replacement.get(), mem.get(), std::min(to, mem.size()));
if (zero_new && to > mem.size())
memset(static_cast<uint8_t*>(replacement.get()) + mem.size(), 0, to - mem.size());
mem.reset(replacement.release(), to, scoped_memory::MALLOC_ALLOCATED);
} else {
// main path: try to mremap.
void *new_addr = mremap(mem.get(), RoundUpSize(mem), to, MREMAP_MAYMOVE);
if (new_addr != MAP_FAILED) {
scoped_memory::Alloc source(mem.source()); // steal resets mem.source()
mem.steal(); // let go otherwise reset() will free it first
mem.reset(new_addr, to, source);
} else {
// Reallocating huge pages can fail with EINVAL.
// https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/mm/mremap.c?id=refs/tags/v3.19#n346
ReplaceAndCopy(to, zero_new, mem);
}
}
return;
#endif // __linux__
case scoped_memory::MALLOC_ALLOCATED:
#ifdef __linux__
// Transition larger allocations to huge pages, but don't keep trying if we're still malloc allocated.
if (to >= kTransitionHuge && mem.size() < kTransitionHuge) {
ReplaceAndCopy(to, zero_new, mem);
return;
}
#endif // __linux__
{
void *new_addr = std::realloc(mem.get(), to);
UTIL_THROW_IF(!new_addr, ErrnoException, "realloc to " << to << " bytes failed.");
if (zero_new && to > mem.size())
memset(static_cast<uint8_t*>(new_addr) + mem.size(), 0, to - mem.size());
mem.steal();
mem.reset(new_addr, to, scoped_memory::MALLOC_ALLOCATED);
}
return;
default:
UTIL_THROW(Exception, "HugeRealloc called with type " << mem.source());
}
}
void MapRead(LoadMethod method, int fd, uint64_t offset, std::size_t size, scoped_memory &out) {
switch (method) {
case LAZY:
out.reset(MapOrThrow(size, false, kFileFlags, false, fd, offset), size, scoped_memory::MMAP_ALLOCATED);
break;
case POPULATE_OR_LAZY:
#ifdef MAP_POPULATE
case POPULATE_OR_READ:
#endif
out.reset(MapOrThrow(size, false, kFileFlags, true, fd, offset), size, scoped_memory::MMAP_ALLOCATED);
break;
#ifndef MAP_POPULATE
case POPULATE_OR_READ:
#endif
case READ:
HugeMalloc(size, false, out);
SeekOrThrow(fd, offset);
ReadOrThrow(fd, out.get(), size);
break;
case PARALLEL_READ:
UTIL_THROW(Exception, "Parallel read was removed from this repo.");
break;
}
}
void *MapZeroedWrite(int fd, std::size_t size) {
ResizeOrThrow(fd, 0);
ResizeOrThrow(fd, size);
return MapOrThrow(size, true, kFileFlags, false, fd, 0);
}
void *MapZeroedWrite(const char *name, std::size_t size, scoped_fd &file) {
file.reset(CreateOrThrow(name));
try {
return MapZeroedWrite(file.get(), size);
} catch (ErrnoException &e) {
e << " in file " << name;
throw;
}
}
Rolling::Rolling(const Rolling &copy_from, uint64_t increase) {
*this = copy_from;
IncreaseBase(increase);
}
Rolling &Rolling::operator=(const Rolling &copy_from) {
fd_ = copy_from.fd_;
file_begin_ = copy_from.file_begin_;
file_end_ = copy_from.file_end_;
for_write_ = copy_from.for_write_;
block_ = copy_from.block_;
read_bound_ = copy_from.read_bound_;
current_begin_ = 0;
if (copy_from.IsPassthrough()) {
current_end_ = copy_from.current_end_;
ptr_ = copy_from.ptr_;
} else {
// Force call on next mmap.
current_end_ = 0;
ptr_ = NULL;
}
return *this;
}
Rolling::Rolling(int fd, bool for_write, std::size_t block, std::size_t read_bound, uint64_t offset, uint64_t amount) {
current_begin_ = 0;
current_end_ = 0;
fd_ = fd;
file_begin_ = offset;
file_end_ = offset + amount;
for_write_ = for_write;
block_ = block;
read_bound_ = read_bound;
}
void *Rolling::ExtractNonRolling(scoped_memory &out, uint64_t index, std::size_t size) {
out.reset();
if (IsPassthrough()) return static_cast<uint8_t*>(get()) + index;
uint64_t offset = index + file_begin_;
// Round down to multiple of page size.
uint64_t cruft = offset % static_cast<uint64_t>(SizePage());
std::size_t map_size = static_cast<std::size_t>(size + cruft);
out.reset(MapOrThrow(map_size, for_write_, kFileFlags, true, fd_, offset - cruft), map_size, scoped_memory::MMAP_ALLOCATED);
return static_cast<uint8_t*>(out.get()) + static_cast<std::size_t>(cruft);
}
void Rolling::Roll(uint64_t index) {
assert(!IsPassthrough());
std::size_t amount;
if (file_end_ - (index + file_begin_) > static_cast<uint64_t>(block_)) {
amount = block_;
current_end_ = index + amount - read_bound_;
} else {
amount = file_end_ - (index + file_begin_);
current_end_ = index + amount;
}
ptr_ = static_cast<uint8_t*>(ExtractNonRolling(mem_, index, amount)) - index;
current_begin_ = index;
}
} // namespace util