Files
llama.cpp/src/llama-mmap.cpp
T
LumpiastyandClaude Opus 5.5 cce0ac1f09 llama: drop the model file's page cache after a copying load
With -lm none the GGUF is read() into our own buffers, but the kernel keeps
the file in page cache afterwards, charged to the pod's cgroup: about 20 GB
after a 35B load. On 2026-10-24 the Talos OOM controller killed the pod right
after a load under exactly that pressure. Those pages are pure waste once the
data is copied.

llama_file::drop_cache() wraps posix_fadvise(POSIX_FADV_DONTNEED) (Linux only,
errors logged at debug). load_all_data() drops each copied tensor's range right
after reading it, so the cache never grows past about one tensor, and drops the
whole file at the final cleanup to catch the metadata and partial pages at
tensor edges. Both only when the loader holds no mapping of the files
(mappings.empty(): not mmap, not mmap+mlock, no lazy tensors), so nothing that
is mapped is ever advised away. Split GGUFs are covered (every file in files).
clip.cpp always copies the mmproj via ifstream, so it drops the whole file
after loading it.

Measured on the pod, gemma-4-E2B UD-Q4_K_XL (2.5 GiB) + mmproj BF16 (0.9 GiB),
CPU only, -lm none, cold cache, same binary with the drops disabled as the
control (4 rounds, rotated order):
- file pages resident after load (mincore): 2431-2444 / 940 MiB -> 0 / 0
- peak model-file residency during load: 2.5 GiB -> 1.2-1.5 GiB (largest tensor)
- pod cgroup file after load: +3.37 GiB -> +0; host MemFree after load
  0.6 GB -> 2.5-2.8 GB (MemAvailable counts clean cache, so it cannot show this)
- load time: 5.87 s -> 6.82 s median (+0.95 s); a post-load-only drop is
  6.56 s (+0.68 s). The cost is synchronous page freeing that reclaim would
  otherwise do later; -lm dio (direct IO) costs the same (+0.9 s) and does not
  cover metadata or the mmproj.
- greedy output byte-identical; -lm mmap keeps the model file fully resident
  and mapped (2498 MiB, no refaults); -lzm on skips the drop (file mapped).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-25 13:13:00 +01:00

930 lines
30 KiB
C++

#include "llama-mmap.h"
#include "llama-impl.h"
#include "ggml.h"
#include <cstring>
#include <climits>
#include <cstdlib>
#include <stdexcept>
#include <cerrno>
#include <algorithm>
#ifdef __has_include
#if __has_include(<unistd.h>)
#include <unistd.h>
#include <fcntl.h>
#include <sys/stat.h>
#if defined(_POSIX_MAPPED_FILES)
#include <sys/mman.h>
#endif
#if defined(_POSIX_MEMLOCK_RANGE)
#include <sys/resource.h>
#endif
#endif
#endif
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#ifndef PATH_MAX
#define PATH_MAX MAX_PATH
#endif
#include <io.h>
#endif
#if defined(__APPLE__)
#include <TargetConditionals.h>
#endif
#ifdef _WIN32
# define llama_mmap_ftell _ftelli64
# define llama_mmap_fseek _fseeki64
#else
# define llama_mmap_ftell ftello
# define llama_mmap_fseek fseeko
#endif
// TODO: consider moving to llama-impl.h if needed in more places
#if defined(_WIN32)
static std::string llama_format_win_err(DWORD err) {
LPSTR buf;
size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
if (!size) {
return "FormatMessageA failed";
}
std::string ret(buf, size);
LocalFree(buf);
return ret;
}
#endif
// llama_file
struct llama_file::impl {
#if defined(_WIN32)
HANDLE fp_win32;
std::string GetErrorMessageWin32(DWORD error_code) const {
std::string ret;
LPSTR lpMsgBuf = NULL;
DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
if (!bufLen) {
ret = format("Win32 error code: %lx", error_code);
} else {
ret = lpMsgBuf;
LocalFree(lpMsgBuf);
}
return ret;
}
impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) {
fp = ggml_fopen(fname, mode);
if (fp == NULL) {
throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
}
fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp));
seek(0, SEEK_END);
size = tell();
seek(0, SEEK_SET);
}
impl(FILE * file) : owns_fp(false) {
fp = file;
fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp));
seek(0, SEEK_END);
size = tell();
seek(0, SEEK_SET);
}
size_t tell() const {
LARGE_INTEGER li;
li.QuadPart = 0;
BOOL ret = SetFilePointerEx(fp_win32, li, &li, FILE_CURRENT);
if (!ret) {
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
return li.QuadPart;
}
void seek(size_t offset, int whence) const {
static_assert(SEEK_SET == FILE_BEGIN, "SEEK_SET != FILE_BEGIN");
static_assert(SEEK_CUR == FILE_CURRENT, "SEEK_CUR != FILE_CURRENT");
static_assert(SEEK_END == FILE_END, "SEEK_END != FILE_END");
LARGE_INTEGER li;
li.QuadPart = offset;
BOOL ret = SetFilePointerEx(fp_win32, li, NULL, whence);
if (!ret) {
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
}
void read_raw(void * ptr, size_t len) {
size_t bytes_read = 0;
while (bytes_read < len) {
size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
DWORD chunk_read = 0;
BOOL result = ReadFile(fp_win32, reinterpret_cast<char*>(ptr) + bytes_read, chunk_size, &chunk_read, NULL);
if (!result) {
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
if (chunk_read < chunk_size || chunk_read == 0) {
throw std::runtime_error("unexpectedly reached end of file");
}
bytes_read += chunk_read;
}
}
uint32_t read_u32() {
uint32_t val;
read_raw(&val, sizeof(val));
return val;
}
void write_raw(const void * ptr, size_t len) const {
size_t bytes_written = 0;
while (bytes_written < len) {
size_t chunk_size = std::min<size_t>(len - bytes_written, 64*1024*1024);
DWORD chunk_written = 0;
BOOL result = WriteFile(fp_win32, reinterpret_cast<char const*>(ptr) + bytes_written, chunk_size, &chunk_written, NULL);
if (!result) {
throw std::runtime_error(format("write error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
if (chunk_written < chunk_size || chunk_written == 0) {
throw std::runtime_error("unexpectedly failed to write bytes");
}
bytes_written += chunk_written;
}
}
void write_u32(uint32_t val) const {
write_raw(&val, sizeof(val));
}
bool has_direct_io() const {
return true;
}
~impl() {
if (fp && owns_fp) {
std::fclose(fp);
}
}
#else
impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) : fname(fname) {
#ifdef __linux__
// Try unbuffered I/O for read only
if (use_direct_io && std::strcmp(mode, "rb") == 0) {
if (init_fd()) {
return;
}
LLAMA_LOG_WARN("Failed to open file '%s' with error: %s. Falling back to buffered I/O",
fname, strerror(errno));
}
#endif
init_fp(mode);
}
#ifdef __linux__
bool init_fd() {
fd = open(fname.c_str(), O_RDONLY | O_DIRECT);
if (fd != -1) {
struct stat file_stats{};
fstat(fd, &file_stats);
size = file_stats.st_size;
alignment = file_stats.st_blksize;
off_t ret = lseek(fd, 0, SEEK_SET);
if (ret == -1) {
throw std::runtime_error(format("seek error: %s", strerror(errno)));
}
return true;
}
return false;
}
#endif
void init_fp(const char * mode) {
fp = ggml_fopen(fname.c_str(), mode);
if (fp == NULL) {
throw std::runtime_error(format("failed to open %s: %s", fname.c_str(), strerror(errno)));
}
seek(0, SEEK_END);
size = tell();
seek(0, SEEK_SET);
}
impl(FILE * file) : fname("(file*)"), owns_fp(false) {
fp = file;
seek(0, SEEK_END);
size = tell();
seek(0, SEEK_SET);
}
size_t tell() const {
if (fd == -1) {
off_t ret = llama_mmap_ftell(fp);
if (ret == -1) {
throw std::runtime_error(format("ftell error: %s", strerror(errno)));
}
return (size_t) ret;
}
off_t pos = lseek(fd, 0, SEEK_CUR);
if (pos == -1) {
throw std::runtime_error(format("lseek error: %s", strerror(errno)));
}
return (size_t) pos;
}
void seek(size_t offset, int whence) const {
off_t ret = 0;
if (fd == -1) {
ret = llama_mmap_fseek(fp, offset, whence);
} else {
ret = lseek(fd, offset, whence);
}
if (ret == -1) {
throw std::runtime_error(format("seek error: %s", strerror(errno)));
}
}
void read_raw_unsafe(void * ptr, size_t len) {
if (len == 0) {
return;
}
errno = 0;
if (fd == -1) {
const size_t curr_off = tell();
const size_t to_read = std::min(len, size - curr_off);
std::size_t ret = std::fread(ptr, to_read, 1, fp);
if (ferror(fp)) {
throw std::runtime_error(format("read error: %s", strerror(errno)));
}
if (to_read > 0 && ret != 1) {
throw std::runtime_error("unexpectedly reached end of file");
}
} else {
size_t bytes_read = 0;
while (bytes_read < len) {
const size_t to_read = len - bytes_read;
ssize_t ret = ::read(fd, reinterpret_cast<char *>(ptr) + bytes_read, to_read);
if (ret == -1) {
if (errno == EINTR) {
continue; // Interrupted by signal, retry
}
// Fallback to std::fread in case the DMA controller cannot access the buffer
if (errno == EFAULT || errno == EINVAL) {
LLAMA_LOG_WARN("%s: Falling back to buffered IO due to %s\n", __func__, strerror(errno));
auto curr_off = tell();
close(fd);
fd = -1;
alignment = 1;
init_fp("rb");
seek(curr_off, SEEK_SET);
read_raw_unsafe(ptr, len);
return;
}
throw std::runtime_error(format("read error: %s", strerror(errno)));
}
if (ret == 0) {
// EOF: allow if this read was only pulling alignment padding past file end
off_t pos = lseek(fd, 0, SEEK_CUR);
if (pos != -1 && (size_t) pos == size) {
std::memset(reinterpret_cast<char *>(ptr) + bytes_read, 0, len - bytes_read);
return;
}
throw std::runtime_error("unexpectedly reached end of file");
}
bytes_read += (size_t) ret;
}
}
}
void read_aligned_chunk(void * dest, size_t size) {
size_t offset = tell();
off_t aligned_offset = offset & ~(alignment - 1);
off_t offset_from_alignment = offset - aligned_offset;
size_t bytes_to_read = (offset_from_alignment + size + alignment - 1) & ~(alignment - 1);
// stage through a bounded buffer, so that a large tensor is not held in memory twice while it loads
const size_t buffer_size = std::min<size_t>(bytes_to_read, LLAMA_DIRECT_IO_BUFFER_SIZE);
void * raw_buffer = nullptr;
int ret = posix_memalign(&raw_buffer, alignment, buffer_size);
if (ret != 0) {
throw std::runtime_error(format("posix_memalign failed with error %d", ret));
}
struct aligned_buffer_deleter {
void operator()(void * p) const { free(p); }
};
std::unique_ptr<void, aligned_buffer_deleter> buffer(raw_buffer);
seek(aligned_offset, SEEK_SET);
size_t skip = offset_from_alignment;
size_t copied = 0;
for (size_t done = 0; done < bytes_to_read; ) {
const size_t n = std::min(buffer_size, bytes_to_read - done);
read_raw_unsafe(buffer.get(), n);
const size_t count = std::min(n - skip, size - copied);
memcpy(reinterpret_cast<char *>(dest) + copied, reinterpret_cast<char *>(buffer.get()) + skip, count);
copied += count;
skip = 0;
done += n;
}
}
void read_raw(void * ptr, size_t len) {
if (has_direct_io()) {
read_aligned_chunk(ptr, len);
} else {
read_raw_unsafe(ptr, len);
}
}
uint32_t read_u32() {
uint32_t ret;
read_raw(&ret, sizeof(ret));
return ret;
}
void write_raw(const void * ptr, size_t len) const {
if (len == 0) {
return;
}
errno = 0;
size_t ret = std::fwrite(ptr, len, 1, fp);
if (ret != 1) {
throw std::runtime_error(format("write error: %s", strerror(errno)));
}
}
void write_u32(uint32_t val) const {
write_raw(&val, sizeof(val));
}
bool has_direct_io() const {
return fd != -1 && alignment > 1;
}
~impl() {
if (fd != -1) {
close(fd);
} else if (owns_fp) {
std::fclose(fp);
}
}
int fd = -1;
std::string fname;
#endif
size_t read_alignment() const {
return alignment;
}
size_t alignment = 1;
FILE * fp{};
size_t size{};
bool owns_fp = true;
};
llama_file::llama_file(const char * fname, const char * mode, const bool use_direct_io) :
pimpl(std::make_unique<impl>(fname, mode, use_direct_io)) {}
llama_file::llama_file(FILE * file) : pimpl(std::make_unique<impl>(file)) {}
llama_file::~llama_file() = default;
size_t llama_file::tell() const { return pimpl->tell(); }
size_t llama_file::size() const { return pimpl->size; }
size_t llama_file::read_alignment() const { return pimpl->read_alignment(); }
bool llama_file::has_direct_io() const { return pimpl->has_direct_io(); }
void llama_file::drop_cache(size_t offset, size_t len) const {
#ifdef __linux__
// only a hint: the pages are clean, a failure just leaves them cached
const int err = posix_fadvise(file_id(), (off_t) offset, (off_t) len, POSIX_FADV_DONTNEED);
if (err != 0) {
LLAMA_LOG_DEBUG("%s: posix_fadvise(.., POSIX_FADV_DONTNEED) failed: %s\n", __func__, strerror(err));
}
#else
GGML_UNUSED(offset);
GGML_UNUSED(len);
#endif
}
int llama_file::file_id() const {
#ifdef _WIN32
return _fileno(pimpl->fp);
#else
if (pimpl->fd != -1) {
return pimpl->fd;
}
#if defined(fileno)
return fileno(pimpl->fp);
#else
return ::fileno(pimpl->fp);
#endif
#endif
}
void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
void llama_file::read_raw(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
#ifdef _WIN32
void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
#else
void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw_unsafe(ptr, len); }
#endif
uint32_t llama_file::read_u32() { return pimpl->read_u32(); }
void llama_file::write_raw(const void * ptr, size_t len) const { pimpl->write_raw(ptr, len); }
void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }
// llama_mmap
#if defined(_POSIX_MAPPED_FILES) || defined(_WIN32)
// merge `ranges` and return their complement within [0, limit)
static llama_mmap::ranges ranges_complement(llama_mmap::ranges ranges, size_t limit) {
llama_mmap::ranges res;
std::sort(ranges.begin(), ranges.end());
size_t pos = 0;
for (const auto & range : ranges) {
const size_t beg = std::min(range.first, limit);
const size_t end = std::min(range.second, limit);
if (beg > pos) {
res.emplace_back(pos, beg);
}
pos = std::max(pos, end);
}
if (pos < limit) {
res.emplace_back(pos, limit);
}
return res;
}
#endif
struct llama_mmap::impl {
#ifdef _POSIX_MAPPED_FILES
std::vector<std::pair<size_t, size_t>> mapped_fragments;
impl(struct llama_file * file, size_t prefetch, bool numa, const llama_mmap::ranges & lazy_ranges) {
size = file->size();
int fd = file->file_id();
int flags = MAP_SHARED;
if (numa) { prefetch = 0; }
#ifdef __linux__
if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) {
LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n",
strerror(errno));
}
// MAP_POPULATE would fault in the lazy ranges too
if (prefetch && lazy_ranges.empty()) { flags |= MAP_POPULATE; }
#endif
addr = mmap(NULL, file->size(), PROT_READ, flags, fd, 0);
if (addr == MAP_FAILED) {
throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
}
// page-aligned madvise over [beg, end), clamped to the file
auto advise = [&](size_t beg, size_t end, int advice, const char * name) {
const size_t page_size = sysconf(_SC_PAGESIZE);
beg = beg & ~(page_size - 1);
end = std::min((end + page_size - 1) & ~(page_size - 1), file->size());
if (beg >= end) {
return;
}
if (posix_madvise((char *) addr + beg, end - beg, advice)) {
LLAMA_LOG_WARN("warning: posix_madvise(.., %s) failed: %s\n", name, strerror(errno));
}
};
if (prefetch > 0) {
for (const auto & range : ranges_complement(lazy_ranges, std::min(file->size(), prefetch))) {
advise(range.first, range.second, POSIX_MADV_WILLNEED, "POSIX_MADV_WILLNEED");
}
}
for (const auto & range : lazy_ranges) {
advise(range.first, range.second, POSIX_MADV_RANDOM, "POSIX_MADV_RANDOM");
}
if (numa) {
if (posix_madvise(addr, file->size(), POSIX_MADV_RANDOM)) {
LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
strerror(errno));
}
}
mapped_fragments.emplace_back(0, file->size());
}
static void align_range(size_t * first, size_t * last, size_t page_size) {
size_t offset_in_page = *first & (page_size - 1);
size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
*first += offset_to_page;
*last = *last & ~(page_size - 1);
if (*last <= *first) {
*last = *first;
}
}
void unmap_fragment(size_t first, size_t last) {
int page_size = sysconf(_SC_PAGESIZE);
align_range(&first, &last, page_size);
size_t len = last - first;
if (len == 0) {
return;
}
GGML_ASSERT(first % page_size == 0);
GGML_ASSERT(last % page_size == 0);
GGML_ASSERT(last > first);
void * next_page_start = (uint8_t *) addr + first;
if (munmap(next_page_start, len)) {
LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
}
std::vector<std::pair<size_t, size_t>> new_mapped_fragments;
for (const auto & frag : mapped_fragments) {
if (frag.first < first && frag.second > last) {
new_mapped_fragments.emplace_back(frag.first, first);
new_mapped_fragments.emplace_back(last, frag.second);
} else if (frag.first < first && frag.second > first) {
new_mapped_fragments.emplace_back(frag.first, first);
} else if (frag.first < last && frag.second > last) {
new_mapped_fragments.emplace_back(last, frag.second);
} else if (frag.first >= first && frag.second <= last) {
} else {
new_mapped_fragments.push_back(frag);
}
}
mapped_fragments = std::move(new_mapped_fragments);
}
~impl() {
for (const auto & frag : mapped_fragments) {
if (munmap((char *) addr + frag.first, frag.second - frag.first)) {
LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
}
}
}
#elif defined(_WIN32)
HANDLE hMapping = nullptr;
impl(struct llama_file * file, size_t prefetch, bool numa, const llama_mmap::ranges & lazy_ranges) {
GGML_UNUSED(numa);
size = file->size();
HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id());
hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
if (hMapping == NULL) {
DWORD error = GetLastError();
throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
}
addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
DWORD error = GetLastError();
if (addr == NULL) {
CloseHandle(hMapping);
throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
}
if (prefetch > 0) {
#if _WIN32_WINNT >= 0x602
BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
if (pPrefetchVirtualMemory) {
std::vector<WIN32_MEMORY_RANGE_ENTRY> entries;
for (const auto & range : ranges_complement(lazy_ranges, std::min(size, prefetch))) {
WIN32_MEMORY_RANGE_ENTRY entry;
entry.VirtualAddress = (char *) addr + range.first;
entry.NumberOfBytes = (SIZE_T) (range.second - range.first);
entries.push_back(entry);
}
if (!entries.empty() &&
!pPrefetchVirtualMemory(GetCurrentProcess(), (ULONG_PTR) entries.size(), entries.data(), 0)) {
LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
#else
LLAMA_LOG_DEBUG("skipping PrefetchVirtualMemory because _WIN32_WINNT < 0x602\n");
#endif
}
}
void unmap_fragment(size_t first, size_t last) {
GGML_UNUSED(first);
GGML_UNUSED(last);
}
~impl() {
if (hMapping) {
if (addr) {
if (!UnmapViewOfFile(addr)) {
LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
if (!CloseHandle(hMapping)) {
LLAMA_LOG_WARN("warning: CloseHandle failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
}
#else
impl(struct llama_file * file, size_t prefetch, bool numa, const llama_mmap::ranges & lazy_ranges) {
GGML_UNUSED(file);
GGML_UNUSED(prefetch);
GGML_UNUSED(numa);
GGML_UNUSED(lazy_ranges);
throw std::runtime_error("mmap not supported");
}
void unmap_fragment(size_t first, size_t last) {
GGML_UNUSED(first);
GGML_UNUSED(last);
throw std::runtime_error("mmap not supported");
}
#endif
void * addr;
size_t size;
};
llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa,
const ranges & lazy_ranges) : pimpl(std::make_unique<impl>(file, prefetch, numa, lazy_ranges)) {}
llama_mmap::~llama_mmap() = default;
size_t llama_mmap::size() const { return pimpl->size; }
void * llama_mmap::addr() const { return pimpl->addr; }
void llama_mmap::unmap_fragment(size_t first, size_t last) { pimpl->unmap_fragment(first, last); }
#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
const bool llama_mmap::SUPPORTED = true;
#else
const bool llama_mmap::SUPPORTED = false;
#endif
// llama_mlock
struct llama_mlock::impl {
#ifdef _POSIX_MEMLOCK_RANGE
static size_t lock_granularity() {
return (size_t) sysconf(_SC_PAGESIZE);
}
bool raw_lock(const void * addr, size_t size) const {
if (!mlock(addr, size)) {
return true;
}
#ifdef __APPLE__
#define MLOCK_SUGGESTION \
"Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
"decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n"
#else
#define MLOCK_SUGGESTION \
"Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n"
#endif
char* errmsg = std::strerror(errno);
bool suggest = (errno == ENOMEM);
#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX) || defined(__HAIKU__)
// visionOS/tvOS/Haiku don't support RLIMIT_MEMLOCK
// Skip resource limit checks on these platforms
suggest = false;
#else
struct rlimit lock_limit;
if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
suggest = false;
}
if (suggest && ((uint64_t)lock_limit.rlim_max > (uint64_t)lock_limit.rlim_cur + size)) {
suggest = false;
}
#endif
LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
return false;
}
static void raw_unlock(void * addr, size_t size) {
if (munlock(addr, size)) {
LLAMA_LOG_WARN("warning: failed to munlock buffer: %s\n", std::strerror(errno));
}
}
#elif defined(_WIN32)
static size_t lock_granularity() {
SYSTEM_INFO si;
GetSystemInfo(&si);
return (size_t) si.dwPageSize;
}
bool raw_lock(void * ptr, size_t len) const {
for (int tries = 1; ; tries++) {
if (VirtualLock(ptr, len)) {
return true;
}
if (tries == 2) {
LLAMA_LOG_WARN("warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
len, size, llama_format_win_err(GetLastError()).c_str());
return false;
}
SIZE_T min_ws_size, max_ws_size;
if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
LLAMA_LOG_WARN("warning: GetProcessWorkingSetSize failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
return false;
}
size_t increment = len + 1048576;
min_ws_size += increment;
max_ws_size += increment;
if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
LLAMA_LOG_WARN("warning: SetProcessWorkingSetSize failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
return false;
}
}
}
static void raw_unlock(void * ptr, size_t len) {
if (!VirtualUnlock(ptr, len)) {
LLAMA_LOG_WARN("warning: failed to VirtualUnlock buffer: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
#else
static size_t lock_granularity() {
return (size_t) 65536;
}
bool raw_lock(const void * addr, size_t len) const {
LLAMA_LOG_WARN("warning: mlock not supported on this system\n");
return false;
}
static void raw_unlock(const void * addr, size_t len) {}
#endif
impl() : addr(NULL), size(0), failed_already(false) {}
void init(void * ptr) {
GGML_ASSERT(addr == NULL && size == 0);
addr = ptr;
}
void grow_to(size_t target_size) {
GGML_ASSERT(addr);
if (failed_already) {
return;
}
size_t granularity = lock_granularity();
target_size = (target_size + granularity - 1) & ~(granularity - 1);
if (target_size > size) {
if (raw_lock((uint8_t *) addr + size, target_size - size)) {
size = target_size;
} else {
failed_already = true;
}
}
}
void * addr;
size_t size;
bool failed_already;
};
llama_mlock::llama_mlock() : pimpl(std::make_unique<impl>()) {}
llama_mlock::~llama_mlock() = default;
void llama_mlock::init(void * ptr) { pimpl->init(ptr); }
void llama_mlock::grow_to(size_t target_size) { pimpl->grow_to(target_size); }
#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
const bool llama_mlock::SUPPORTED = true;
#else
const bool llama_mlock::SUPPORTED = false;
#endif
void llama_prefetch(llama_memory_ranges mr) {
#if defined(__linux__) || defined(__APPLE__) || (defined(_WIN32) && _WIN32_WINNT >= 0x602)
if (mr.empty()) {
return;
}
#if defined(_WIN32)
using prefetch_virtual_memory_t = BOOL (WINAPI *)(HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
static const auto pPrefetchVirtualMemory = (prefetch_virtual_memory_t) (void *) GetProcAddress(GetModuleHandleW(L"kernel32.dll"), "PrefetchVirtualMemory");
if (!pPrefetchVirtualMemory) {
return;
}
static const long page_size = [] {
SYSTEM_INFO info;
GetSystemInfo(&info);
return (long) info.dwPageSize;
}();
#else
static const long page_size = sysconf(_SC_PAGESIZE);
#endif
if (page_size <= 0) {
return;
}
const size_t page = (size_t) page_size;
std::sort(mr.begin(), mr.end(), [](const llama_memory_range & a, const llama_memory_range & b) {
return (uintptr_t) a.addr < (uintptr_t) b.addr;
});
uintptr_t begin = 0, end = 0;
size_t n_ranges = 0;
for (const auto & range : mr) {
if (!range.addr || range.size == 0) {
continue;
}
const uintptr_t pointer = (uintptr_t) range.addr;
const uintptr_t first = pointer / page * page;
const uintptr_t last = (pointer + range.size + page - 1) / page * page;
if (end && first > end) {
mr[n_ranges++] = { (const void *) begin, end - begin };
end = 0;
}
if (!end) {
begin = first;
}
end = std::max(end, last);
}
if (end) {
mr[n_ranges++] = { (const void *) begin, end - begin };
}
mr.resize(n_ranges);
#if defined(_WIN32)
// prefetch the merged ranges in one call, so the reads can be issued concurrently
std::vector<WIN32_MEMORY_RANGE_ENTRY> entries;
entries.reserve(mr.size());
for (const auto & range : mr) {
entries.push_back({ const_cast<PVOID>(range.addr), (SIZE_T) range.size });
}
if (!entries.empty() && !pPrefetchVirtualMemory(GetCurrentProcess(), (ULONG_PTR) entries.size(), entries.data(), 0)) {
LLAMA_LOG_WARN("llama_prefetch: PrefetchVirtualMemory failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
#else
for (const auto & range : mr) {
// TODO: in a follow-up, use posix_madvise(.., POSIX_MADV_WILLNEED) to match the rest of this file
if (madvise(const_cast<void *>(range.addr), range.size, MADV_WILLNEED) != 0) {
LLAMA_LOG_WARN("llama_prefetch: madvise(MADV_WILLNEED) failed: %s\n", strerror(errno));
return;
}
}
#endif
#else
GGML_UNUSED(mr);
#endif
}
size_t llama_path_max() {
return PATH_MAX;
}