blob: 19a9a4342e5fa2ac84df6aa0d86e3f33135144ad [file] [log] [blame]
// Copyright 2015 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
// Author: Ken Chen <kenchen@google.com>
//
// hugepage text library to remap process executable segment with hugepages.
#include "chromeos/hugepage_text/hugepage_text.h"
#include <link.h>
#include <sys/mman.h>
#include <sys/resource.h>
#include <sys/time.h>
#include "base/bit_cast.h"
#include "base/logging.h"
namespace chromeos {
#ifndef MAP_HUGETLB
#define MAP_HUGETLB 0x40000
#endif
#ifndef TMPFS_MAGIC
#define TMPFS_MAGIC 0x01021994
#endif
#ifndef MADV_HUGEPAGE
#define MADV_HUGEPAGE 14
#endif
const int kHpageShift = 21;
const int kHpageSize = (1 << kHpageShift);
const int kHpageMask = (~(kHpageSize - 1));
const int kProtection = (PROT_READ | PROT_WRITE);
const int kMremapFlags = (MREMAP_MAYMOVE | MREMAP_FIXED);
// The number of hugepages we want to use to map chrome text section
// to hugepages. With the help of AutoFDO, the hot functions are grouped
// in to a small area of the binary.
const int kNumHugePages = 15;
// Get an anonymous mapping backed by explicit transparent hugepage
// Return NULL if such mapping can not be established.
static void* GetTransparentHugepageMapping(const size_t hsize) {
// setup explicit transparent hugepage segment
char* addr = static_cast<char*>(mmap(NULL, hsize + kHpageSize, kProtection,
MAP_ANONYMOUS | MAP_PRIVATE, -1, 0));
if (addr == MAP_FAILED) {
PLOG(INFO) << "unable to mmap anon pages, fall back to small page";
return NULL;
}
// remove unaligned head and tail regions
size_t head_gap = kHpageSize - (size_t)addr % kHpageSize;
size_t tail_gap = kHpageSize - head_gap;
munmap(addr, head_gap);
munmap(addr + head_gap + hsize, tail_gap);
void* haddr = addr + head_gap;
if (madvise(haddr, hsize, MADV_HUGEPAGE)) {
PLOG(INFO) << "no transparent hugepage support, fall back to small page";
munmap(haddr, hsize);
return NULL;
}
if (mlock(haddr, hsize)) {
PLOG(INFO) << "Mlocking text pages failed";
}
return haddr;
}
// memcpy for word-aligned data which is not instrumented by AddressSanitizer.
ATTRIBUTE_NO_SANITIZE_ADDRESS
static void NoAsanAlignedMemcpy(void* dst, void* src, size_t size) {
DCHECK_EQ(0U, size % sizeof(uintptr_t)); // size is a multiple of word size.
DCHECK_EQ(0U, reinterpret_cast<uintptr_t>(dst) % sizeof(uintptr_t));
DCHECK_EQ(0U, reinterpret_cast<uintptr_t>(src) % sizeof(uintptr_t));
uintptr_t* d = reinterpret_cast<uintptr_t*>(dst);
uintptr_t* s = reinterpret_cast<uintptr_t*>(src);
for (size_t i = 0; i < size / sizeof(uintptr_t); i++)
d[i] = s[i];
}
// Remaps text segment at address "vaddr" to hugepage backed mapping via mremap
// syscall. The virtual address does not change. When this call returns, the
// backing physical memory will be changed from small page to hugetlb page.
//
// Inputs: vaddr, the starting virtual address to remap to hugepage
// hsize, size of the memory segment to remap in bytes
// Return: none
// Effect: physical backing page changed from small page to hugepage. If there
// are error condition, the remapping operation is aborted.
static void MremapHugetlbText(void* vaddr, const size_t hsize) {
DCHECK_EQ(0ul, reinterpret_cast<uintptr_t>(vaddr) & ~kHpageMask);
void* haddr = GetTransparentHugepageMapping(hsize);
if (haddr == NULL)
return;
// Copy text segment to hugepage mapping. We are using a non-asan memcpy,
// otherwise it would be flagged as a bunch of out of bounds reads.
NoAsanAlignedMemcpy(haddr, vaddr, hsize);
// change mapping protection to read only now that it has done the copy
if (mprotect(haddr, hsize, PROT_READ | PROT_EXEC)) {
PLOG(INFO) << "can not change protection to r-x, fall back to small page";
munmap(haddr, hsize);
return;
}
// remap hugepage text on top of existing small page mapping
if (mremap(haddr, hsize, hsize, kMremapFlags, vaddr) == MAP_FAILED) {
PLOG(INFO) << "unable to mremap hugepage mapping, fall back to small page";
munmap(haddr, hsize);
return;
}
}
// Top level text remapping function.
//
// Inputs: vaddr, the starting virtual address to remap to hugepage
// segsize, size of the memory segment to remap in bytes
// Return: none
// Effect: physical backing page changed from small page to hugepage. If there
// are error condition, the remaping operation is aborted.
static void RemapHugetlbText(void* vaddr, const size_t segsize) {
// remove unaligned head regions
uintptr_t head_gap =
(kHpageSize - reinterpret_cast<uintptr_t>(vaddr) % kHpageSize) %
kHpageSize;
uintptr_t addr = reinterpret_cast<uintptr_t>(vaddr) + head_gap;
if (segsize < head_gap)
return;
size_t hsize = segsize - head_gap;
hsize = hsize & kHpageMask;
if (hsize > kHpageSize * kNumHugePages)
hsize = kHpageSize * kNumHugePages;
if (hsize == 0)
return;
MremapHugetlbText(reinterpret_cast<void*>(addr), hsize);
}
// For a given ELF program header descriptor, iterates over all segments within
// it and find the first segment that has PT_LOAD and is executable, call
// RemapHugetlbText().
//
// Additionally, since these pages are important, we attempt to lock them into
// memory.
//
// Inputs: info: pointer to a struct dl_phdr_info that describes the DSO.
// size: size of the above structure (not used in this function).
// data: user param (not used in this function).
// Return: always return true. The value is propagated by dl_iterate_phdr().
static int FilterElfHeader(struct dl_phdr_info* info, size_t size, void* data) {
void* vaddr;
int segsize;
// From dl_iterate_phdr's man page: "The first object visited by callback is
// the main program. For the main program, the dlpi_name field will be an
// empty string." Hence, no "is this the Chrome we're looking for?" checks are
// necessary.
for (int i = 0; i < info->dlpi_phnum; i++) {
if (info->dlpi_phdr[i].p_type == PT_LOAD &&
info->dlpi_phdr[i].p_flags == (PF_R | PF_X)) {
vaddr = bit_cast<void*>(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
segsize = info->dlpi_phdr[i].p_filesz;
RemapHugetlbText(vaddr, segsize);
// Only re-map the first text segment.
return 1;
}
}
return 1;
}
// Main function. This function will iterate all ELF segments, attempt to remap
// parts of the text segment from small page to hugepage, and mlock in all of
// the hugepages. Any errors will cause the failing piece of this to be rolled
// back, so nothing world-ending can come from this function (hopefully ;) ).
void InitHugepagesAndMlockSelf(void) {
dl_iterate_phdr(FilterElfHeader, 0);
}
} // namespace chromeos