首页 > 漏洞分析 > CVE-2023-36036 Windows Cloud Files Mini Filter Driver 权限提升漏洞分析

CVE-2023-36036 Windows Cloud Files Mini Filter Driver 权限提升漏洞分析

基本信息

Windows Cloud Files Mini Filter 驱动中存在越界写入漏洞，在解析Reparse point数据时，由于memcpy函数的长度参数用户可控，源内存可控，导致攻击者可以构造恶意结构并传递给Windows Cloud Files Mini Filter 驱动，造成越界写入，并在内核执行任意代码。

影响版本

略

环境搭建

Windows 10 23年10月补丁

技术分析&调试

cldflt.sys驱动中实现了云文件的各项功能，diff该驱动，修改函数如下：

在HsmpRpiDecompressBuffer函数中有如下修改，对*(_WORD *)(a1 + 10)添加了一个判断，是否>0x4000，

如果大于则抛出错误 0xC000CF02对应 STATUS_CLOUD_FILE_METADATA_CORRUPT

https://endodermis59.rssing.com/chan-36587470/all_p6.html

AI输出

HsmpRpiDecompressBuffer函数的作用是解压压缩后的Reparse Point数据。

主要功能包括:

1. 校验传入数据的完整性和魔数是否正确
2. 如果数据被压缩,则根据原长度分配解压缓冲区
3. 调用RtlDecompressBuffer进行实际解压
4. 检查解压后数据长度是否匹配
5. 如果解压成功,返回解压后的数据
6. 否则返回错误码

所以它是一个典型的压缩数据解压函数,接收原始压缩数据,校验->分配缓冲区->解压->返回解压后数据的过程。

通过解压让后续代码可以处理未压缩的Reparse Point数据,一般在需要提交/更新数据时会解压。

主要作用就是将压缩后的Reparse Point还原为可读的未压缩数据。

HsmpRpiDecompressBuffer由 HsmpRpReadBuffer调用

__int64 __fastcall HsmpRpReadBuffer(PFLT_INSTANCE Instance, PFILE_OBJECT FileObject, unsigned __int16 **a3)
{
...

  *a3 = 0i64;
  v6 = 1024;
  OutputBuffer = (unsigned __int16 *)ExAllocatePoolWithTag(PagedPool, 0x400ui64, 0x70527348u);
  v8 = OutputBuffer;
 ......
  }
  LODWORD(v9) = HsmpRpiDecompressBuffer((__int64)v8, v6, a3);
  HsmDbgBreakOnStatus((unsigned int)v9);
  if ( (int)v9 < 0 )
  {
    v16 = WPP_GLOBAL_Control;
    if ( WPP_GLOBAL_Control == (PDEVICE_OBJECT)&WPP_GLOBAL_Control
      || (HIDWORD(WPP_GLOBAL_Control->Timer) & 1) == 0
      || BYTE1(WPP_GLOBAL_Control->Timer) < 2u )
    {
      goto LABEL_8;
    }
    v17 = 20i64;
    goto LABEL_30;
  }
  if ( *a3 != v8 )
LABEL_8:
    ExFreePoolWithTag(v8, 0x70527348u);
  return (unsigned int)v9;
}

AI解释

这段代码实现了解析和解压Reparse Point数据的功能:

1. 分配OutputBuffer内存用于读取Reparse数据。
2. 调用FltFsControlFile读取Reparse数据到OutputBuffer。
3. 如果读取失败,并且错误码表示buffer太小,则重新分配更大的buffer并重试读取。
4. 调用HsmpRpiDecompressBuffer对读取到的数据进行解压。
5. 如果解压成功,返回解压后的数据,否则返回错误码。
6. 释放原始数据内存。
7. 其中包含了详细的错误和日志记录逻辑。

所以总的来说,这段代码用于从文件系统读取Reparse Point数据,然后对数据进行解压和解析,最后返回解析后的数据。

主要功能聚焦在读取压缩数据并解压这块。

HsmpRpReadBuffer由 HsmpRpCommitNoLock和 HsmpSetupContexts调用。

HsmpRpCommitNoLock解释如下

这个C函数实现了HsmpRpCommitNoLock,它似乎是用于在没有锁的情况下提交页面重播(Page Replay)数据。

主要的逻辑如下:

1. 校验和准备输入数据
2. 分配内存池来存储提交的数据
3. 构建数据结构,填充各种元数据
4. 尝试压缩数据
5. 将数据写入文件
6. 清理临时数据结构和内存

具体来说,这个函数做了以下工作:

1. 验证输入参数的有效性
2. 为输出缓冲区分配内存
3. 构建输出缓冲区的数据结构
4. 填充输出缓冲区的头部
5. 将输入缓冲区的数据复制到输出缓冲区
6. 计算校验和
7. 尝试压缩输出缓冲区
8. 标记文件属性
9. 将输出缓冲区的数据写入文件
10. 重置文件属性
11. 释放临时缓冲区和内存

所以总的来说,这个函数的主要目的是准备并提交页面重播数据,同时处理必要的校验、压缩和清理工作。

在 HsmpRpCommitNoLock中有如下代码，可以看到在前面diff中出现的0x4000和0x3FFC，可以猜测漏洞产生于该函数中

LABEL_156:
  PoolWithTag = (unsigned int *)ExAllocatePoolWithTag(PagedPool, 0x4000ui64, 0x70527348u);
  v142 = PoolWithTag;
  v11 = (char *)PoolWithTag;
  if ( PoolWithTag )
  {
    memset(PoolWithTag, 0, 0x4000ui64);
    v57 = InputBuffer;
    v58 = v11 + 4;
    if ( v8 && *((_WORD *)v8 + 7) > 0xAu )
      v57 = *((_WORD *)v8 + 7);
    v59 = (unsigned int *)(v58 + 8);
    *((_WORD *)v58 + 6) = 0;
    v9 = (unsigned __int64)(v58 + 16);
    *((_WORD *)v58 + 7) = v57;
    *((_DWORD *)v58 + 2) = 8 * v57 + 16;
    *(_DWORD *)v58 = 'pReF';
    memset(v58 + 16, 0, 8i64 * v57);
    if ( *((_WORD *)v58 + 7) )
    {
      v60 = *v59;
      if ( ((v60 + 3) & 0xFFFFFFFFFFFFFFFCui64) + 1 <= 0x3FFC )// 12 偏移
      {
        *v59 = (v60 + 3) & 0xFFFFFFFC;
        if ( *(_WORD *)v9 )
          *((_WORD *)v58 + 6) |= 1u;
        *(_WORD *)v9 = 7;
        LODWORD(v9) = 0;
        *((_WORD *)v58 + 9) = 1;
        v61 = *v59;
        *((_DWORD *)v58 + 5) = v61;
        v58[v61] = 1;

继续审查代码，发现在HsmpRpCommitNoLock中有如下代码，在do while循环中调用memmove函数时，传入的src来源于 HsmpRpReadBuffer解压后的element[10]数据，dst为ExAllocatePoolWithTag分配的大小为0x4000的内存。长度参数来源于ElementInfos[10].Length，不难看出由此可以造成越界写入，且用户可控。

  v32 = 0i64;
  if ( (v9 & 0x80000000) == 0i64 )
  {
    v8 = (char *)P + 12;
...
                {
                  if ( (_DWORD)v54 && (_WORD)v55 )
                    v167 = &v8[v54];
                  else
                    v167 = v32;
    .....
  PoolWithTag = (unsigned int *)ExAllocatePoolWithTag(PagedPool, 0x4000ui64, 0x70527348u);
  v142 = PoolWithTag;
  v11 = (char *)PoolWithTag;
  if ( PoolWithTag )
  {
    memset(PoolWithTag, 0, 0x4000ui64);
    v57 = InputBuffer;
    v58 = v11 + 4;
    if ( v8 && *((_WORD *)v8 + 7) > 0xAu )
      v57 = *((_WORD *)v8 + 7);
    v59 = (unsigned int *)(v58 + 8);
    *((_WORD *)v58 + 6) = 0;
    v9 = (unsigned __int64)(v58 + 16);
    *((_WORD *)v58 + 7) = v57;
    *((_DWORD *)v58 + 2) = 8 * v57 + 16;
    *(_DWORD *)v58 = 'pReF';
    memset(v58 + 16, 0, 8i64 * v57);
   .....
            }
            *v59 += v109;
   .....
            if ( *((_WORD *)v58 + 28) )
              *((_WORD *)v58 + 6) |= 1u;
            *v59 = (v113 + 3) & 0xFFFFFFFC;
          ....
            *v59 += v114;
       .....
       v117 = (char *)v167;
       v107 = (char *)Src;
            *v59 = (v118 + 3) & 0xFFFFFFFC;
            *((_WORD *)v58 + 32) = 17;
            *((_WORD *)v58 + 33) = v119;
            v121 = *v59;
            *((_DWORD *)v58 + 17) = v121;
            if ( &v58[v121] != v117 )
            {
              memmove(&v58[v121], v117, v120);
......
		v125 = 10;
           do
        {
          v126 = v125;
          *(HSM_ELEMENT_INFO *)&v58[8 * v125 + 16] = v124->ElementInfos[v125];
          memmove(&v58[*v59], (char *)v124 + v124->ElementInfos[v125].Offset, v124->ElementInfos[v125].Length);
          ++v125;
          *(_DWORD *)&v58[8 * v126 + 20] = *v59;
          *v59 += *(unsigned __int16 *)&v58[8 * v126 + 18];
        }
        while ( v125 < v124->NumberOfElements );
 ...
  if ( v14 )
    ExFreePoolWithTag(v14, 0x70527348u);
  if ( v11 )
    ExFreePoolWithTag(v11, 0x70527348u);
  return (unsigned int)v9;
}

搜索Reparse point RtlCompressBuffer，找到文章，根据文章 _REPARSE_DATA_BUFFER定义如下，可以知道传入 HsmpRpiDecompressBuffer的是 REPARSE_DATA_BUFFER，其中 ReparseTag为IO_REPARSE_TAG_CLOUD_3 值 0x9000301A 并且在结构体 HsmReparseBufferRaw的RawData成员中存储了由 (RtlCompressBuffer压缩的数据 HsmReparseBufferRaw

    // Handled by cldflt.sys!HsmpRpReadBuffer
    struct {
      USHORT Flags;    // Flags (0x8000 = not compressed)
      USHORT Length;   // Length of the data (uncompressed)
      BYTE RawData[1]; // To be RtlDecompressBuffer-ed
    } HsmReparseBufferRaw;

_REPARSE_DATA_BUFFER定义

typedef struct _REPARSE_DATA_BUFFER {
  ULONG ReparseTag;         // Reparse tag type
  USHORT ReparseDataLength; // Length of the reparse data
  USHORT Reserved;          // Used internally by NTFS to store remaining length
  union {
    // Structure for IO_REPARSE_TAG_SYMLINK
    // Handled by nt!IoCompleteRequest
    struct {
      USHORT SubstituteNameOffset;
      USHORT SubstituteNameLength;
      USHORT PrintNameOffset;
      USHORT PrintNameLength;
      ULONG Flags;
      WCHAR PathBuffer[1];
      /* Example of distinction between substitute and print names:
      // mklink /d ldrive c:\
      // SubstituteName: c:\\??\
      // PrintName: c:\
      */
    } SymbolicLinkReparseBuffer;

    // Structure for IO_REPARSE_TAG_MOUNT_POINT
    // Handled by nt!IoCompleteRequest
    struct {
      USHORT SubstituteNameOffset;
      USHORT SubstituteNameLength;
      USHORT PrintNameOffset;
      USHORT PrintNameLength;
      WCHAR PathBuffer[1];
    } MountPointReparseBuffer;

    // Structure for IO_REPARSE_TAG_WIM
    // Handled by wimmount!FPOpenReparseTarget->wimserv.dll
    // (wimsrv!ImageExtract)
    struct {
      GUID ImageGuid;           // GUID of the mounted VIM image
      BYTE ImagePathHash[0x14]; // Hash of the path to the file within the
                                // image
    } WimImageReparseBuffer;

    // Structure for IO_REPARSE_TAG_WOF
    // Handled by FSCTL_GET_EXTERNAL_BACKING, FSCTL_SET_EXTERNAL_BACKING in
    // NTFS (Windows 10+)
    struct {
      //-- WOF_EXTERNAL_INFO --------------------
      ULONG Wof_Version;  // Should be 1 (WOF_CURRENT_VERSION)
      ULONG Wof_Provider; // Should be 2 (WOF_PROVIDER_FILE)

      //-- FILE_PROVIDER_EXTERNAL_INFO_V1 --------------------
      ULONG FileInfo_Version; // Should be 1 (FILE_PROVIDER_CURRENT_VERSION)
      ULONG
      FileInfo_Algorithm; // Usually 0 (FILE_PROVIDER_COMPRESSION_XPRESS4K)
    } WofReparseBuffer;

    // Structure for IO_REPARSE_TAG_APPEXECLINK
    struct {
      ULONG StringCount;   // Number of the strings in the StringList, separated
                           // by '\0'
      WCHAR StringList[1]; // Multistring (strings separated by '\0',
                           // terminated by '\0\0')
    } AppExecLinkReparseBuffer;

    // Structure for IO_REPARSE_TAG_WCI (0x80000018)
    struct {
      ULONG Version; // Expected to be 1 by wcifs.sys
      ULONG Reserved;
      GUID LookupGuid;      // GUID used for lookup in wcifs!WcLookupLayer
      USHORT WciNameLength; // Length of the WCI subname, in bytes
      WCHAR WciName[1];     // The WCI subname (not zero terminated)
    } WcifsReparseBuffer;

    // Handled by cldflt.sys!HsmpRpReadBuffer
    struct {
      USHORT Flags;    // Flags (0x8000 = not compressed)
      USHORT Length;   // Length of the data (uncompressed)
      BYTE RawData[1]; // To be RtlDecompressBuffer-ed
    } HsmReparseBufferRaw;

    // Dummy structure
    struct {
      UCHAR DataBuffer[1];
    } GenericReparseBuffer;
  } DUMMYUNIONNAME;
} REPARSE_DATA_BUFFER, *PREPARSE_DATA_BUFFER;

在这个Github仓库中实现了对Reparse point的解析，其中定义了HSM_REPARSE_DATA

typedef struct _HSM_ELEMENT_INFO
{
    USHORT Type;                        // Type of the element (?). One of HSM_ELEMENT_TYPE_XXX
    USHORT Length;                      // Length of the element data in bytes
    ULONG  Offset;                      // Offset of the element data, relative to begin of HSM_DATA. Aligned to 4 bytes
} HSM_ELEMENT_INFO, *PHSM_ELEMENT_INFO;

typedef struct _HSM_DATA
{
    ULONG  Magic;                       // 0x70527442 ('pRtB') for bitmap data, 0x70526546 ('FeRp') for file data
    ULONG  Crc32;                       // CRC32 of the following data (calculated by RtlComputeCrc32)
    ULONG  Length;                      // Length of the entire HSM_DATA in bytes
    USHORT Flags;                       // HSM_DATA_XXXX
    USHORT NumberOfElements;            // Number of elements
    HSM_ELEMENT_INFO ElementInfos[1];   // Array of element infos. There is fixed maximal items for bitmap and reparse data
} HSM_DATA, *PHSM_DATA;

typedef struct _HSM_REPARSE_DATA
{
    USHORT Flags;                       // Lower 8 bits is revision (must be 1 as of Windows 10 16299)
                                        // Flags: 0x8000 = Data needs to be decompressed by RtlCompressBuffer
    USHORT Length;                      // Length of the HSM_REPARSE_DATA structure (including "Flags" and "Length")

    HSM_DATA FileData;                  // HSM data
} HSM_REPARSE_DATA, *PHSM_REPARSE_DATA;

对应在 REPARSE_DATA_BUFFER的偏移如下

0:000> dt pa
Local var @ 0xa8444fec08 Type _REPARSE_DATA_BUFFER*
0x000001e0`ef867690
   +0x000 ReparseTag       : 0x9000301a
   +0x004 ReparseDataLength : 0x4008
   +0x006 Reserved         : 0
   +0x008 SymbolicLinkReparseBuffer : _REPARSE_DATA_BUFFER::<unnamed-tag>::<unnamed-type-SymbolicLinkReparseBuffer>
   +0x008 MountPointReparseBuffer : _REPARSE_DATA_BUFFER::<unnamed-tag>::<unnamed-type-MountPointReparseBuffer>
   +0x008 WimImageReparseBuffer : _REPARSE_DATA_BUFFER::<unnamed-tag>::<unnamed-type-WimImageReparseBuffer>
   +0x008 WofReparseBuffer : _REPARSE_DATA_BUFFER::<unnamed-tag>::<unnamed-type-WofReparseBuffer>
   +0x008 AppExecLinkReparseBuffer : _REPARSE_DATA_BUFFER::<unnamed-tag>::<unnamed-type-AppExecLinkReparseBuffer>
   +0x008 WcifsReparseBuffer : _REPARSE_DATA_BUFFER::<unnamed-tag>::<unnamed-type-WcifsReparseBuffer>
   +0x008 hsm_reparse_data : _HSM_REPARSE_DATA
   +0x008 GenericReparseBuffer : _REPARSE_DATA_BUFFER::<unnamed-tag>::<unnamed-type-GenericReparseBuffer>
0:000> dx -r1 (*((poc3!_HSM_REPARSE_DATA *)0x1e0ef867698))
(*((poc3!_HSM_REPARSE_DATA *)0x1e0ef867698))                 [Type: _HSM_REPARSE_DATA]
    [+0x000] Flags            : 0x8001 [Type: unsigned short]   // 8
    [+0x002] Length           : 0x4008 [Type: unsigned short]   // 10
    [+0x004] FileData         [Type: _HSM_DATA]                 // 12
0:000> dx -r1 (*((poc3!_HSM_DATA *)0x1e0ef86769c))
(*((poc3!_HSM_DATA *)0x1e0ef86769c))                 [Type: _HSM_DATA]
    [+0x000] Magic            : 0x70526546 [Type: unsigned long] // 12
    [+0x004] Crc32            : 0x31e13b17 [Type: unsigned long] // 16
    [+0x008] Length           : 0x4004 [Type: unsigned long]     // 20
    [+0x00c] Flags            : 0x2 [Type: unsigned short]       // 24
    [+0x00e] NumberOfElements : 0xb [Type: unsigned short]       // 26
    [+0x010] ElementInfos     [Type: _HSM_ELEMENT_INFO [10)     // 28

PoC构造

将结构体导入到ida中，在HsmpRpCommitNoLock中首先对ReparseTag进行验证，而后将hsm_reparse_data和对应的长度导入到 HsmpRpValidateBuffer函数中验证。

  if ( (reparse_data_buffer->ReparseTag & 0xFFFF0FFF) != dword_1C00235D0 )
  {
    LODWORD(v9) = -1073688821;
    HsmDbgBreakOnStatus(3221278475i64);
    if ( WPP_GLOBAL_Control != (PDEVICE_OBJECT)&WPP_GLOBAL_Control
      && (HIDWORD(WPP_GLOBAL_Control->Timer) & 1) != 0
      && BYTE1(WPP_GLOBAL_Control->Timer) >= 2u )
    {
      WPP_SF_qiqDDd(
        WPP_GLOBAL_Control->AttachedDevice,
        2i64,
        v30,
        a2,
        *(_QWORD *)(v5 + 32),
        v29,
        dword_1C00235D0,
        reparse_data_buffer->ReparseTag);
    }
    goto LABEL_8;
  }
  ReparseDataLength = reparse_data_buffer->ReparseDataLength;
  v9 = (unsigned int)HsmpRpValidateBuffer(&reparse_data_buffer->DUMMYUNIONNAME.hsm_reparse_data, ReparseDataLength);

在 HsmpRpValidateBuffer函数中对HSM_DATA结构体的一些字段做了如下校验。

reparse_data_buffer->ReparseDataLength > 4
reparse_data_buffer->hsm_reparse_data.Flags=1
reparse_data_buffer->hsm_reparse_data.FileData.Magic = ‘pReF’
reparse_data_buffer->hsm_reparse_data.FileData.Flags = 2, 并且reparse_data_buffer->hsm_reparse_data.FileData.Crc32 == RtlComputeCrc32(0, (PUCHAR)&a1->FileData.Length, v2 - 8
NumberOfElements 不为0，且最大为10，最后一个以NONE结尾

特别的，从如下代码中可以看到对ElementInfos[0]和ElementInfos[1]进行了校验，容易得出如下条件：

NumberOfElements > 1
FileData.Length >= 0x20
`FileData.ElementInfos[1].Type == 0xA
FileData.ElementInfos[1].Offset >= 8 * NumberOfElements + 16 && FileData.ElementInfos[1].Offset < FileData.Length
FileData.ElementInfos[1].Length == 4
FileData.ElementInfos[1].Length + FileData.ElementInfos[1].Offset < 65535

  if ( (unsigned __int16)NumberOfElements > 1u
    && (unsigned int)Length >= 0x20
    && (v22 = a1->FileData.ElementInfos[1].Type, v22 < 0x12u)
    && ((v23 = a1->FileData.ElementInfos[1].Offset, !(_DWORD)v23) || v23 >= hsm_data_length)
    && (unsigned int)v23 <= (unsigned int)Length
    && (v24 = a1->FileData.ElementInfos[1].Length, v24 <= (unsigned int)Length)
    && v24 + (unsigned int)v23 >= (unsigned int)v23
    && v24 + (unsigned int)v23 <= (unsigned int)Length
    && v22 == 10
    && v24 == 4 )
  {
    v5 = *(ULONG *)((char *)&p_FileData->Magic + v23);
    IsReparseBufferSupported = 0;
  }
  else
  {
    IsReparseBufferSupported = 0xC0000225;
  }

如下代码对ElementInfos[2]进行了校验，有如下：

FileData.ElementInfos[2].Offset < FileData.Length
FileData.ElementInfos[2].Length < FileData.Length
FileData.ElementInfos[2].Type == 6

  if ( (element_1_Data & 0x10) != 0 )
    return IsReparseBufferSupported;
  v27 = a1->FileData.Length;
  if ( v27 < 0x18
    || (v28 = a1->FileData.NumberOfElements, (unsigned __int16)v28 <= 2u)
    || v27 < 0x28
    || (v29 = a1->FileData.ElementInfos[2].Type, v29 >= 0x12u)
    || (v30 = a1->FileData.ElementInfos[2].Offset, (_DWORD)v30) && v30 < 8 * v28 + 16
    || (unsigned int)v30 > v27
    || (v31 = a1->FileData.ElementInfos[2].Length, v31 > v27)
    || v31 + (unsigned int)v30 < (unsigned int)v30
    || v31 + (unsigned int)v30 > v27
    || v29 != 6
    || (IsReparseBufferSupported = 0, v31 != 8) )
  {
    IsReparseBufferSupported = 0xC0000225;
  }

后面还有一堆校验逻辑就不贴了。

在 HsmpRpCommitNoLock中对 HsmpRpValidateBuffer返回值做了校验，如果IsReparseBufferSupported不为0则会进入报错逻辑，而在 HsmpRpValidateBuffer

  IsReparseBufferSupported = (unsigned int)HsmpRpValidateBuffer(
                                             &reparse_data_buffer->DUMMYUNIONNAME.hsm_reparse_data,
                                             ReparseDataLength);
  HsmDbgBreakOnStatus(IsReparseBufferSupported);
  v32 = 0i64;
  if ( (IsReparseBufferSupported & 0x80000000) == 0i64 )
  {
...
  }
  else
  {
    HsmDbgBreakOnCorruption();
    if ( a4 == (_BYTE)v32 )
    {
      if ( WPP_GLOBAL_Control != (PDEVICE_OBJECT)&WPP_GLOBAL_Control
        && (HIDWORD(WPP_GLOBAL_Control->Timer) & 1) != 0

在HsmpRpValidateBuffer中可以看到当通过第一次校验后，如果ElementInfos[1]的Data & 0x10 则会直接返回，此时IsReparseBufferSupported=0能通过校验。

  if ( (unsigned __int16)NumberOfElements > 1u
    && (unsigned int)Length >= 0x20
    && (v22 = a1->FileData.ElementInfos[1].Type, v22 < 0x12u)
    && ((v23 = a1->FileData.ElementInfos[1].Offset, !(_DWORD)v23) || v23 >= hsm_data_length)
    && (unsigned int)v23 <= (unsigned int)Length
    && (v24 = a1->FileData.ElementInfos[1].Length, v24 <= (unsigned int)Length)
    && v24 + (unsigned int)v23 >= (unsigned int)v23
    && v24 + (unsigned int)v23 <= (unsigned int)Length
    && v22 == 10
    && v24 == 4 )
  {
    element_1_Data = *(ULONG *)((char *)&p_FileData->Magic + v23);
    IsReparseBufferSupported = 0;
  }
  else
  {
    IsReparseBufferSupported = 0xC0000225;
  }
  HsmDbgBreakOnStatus(IsReparseBufferSupported);
  if ( (IsReparseBufferSupported & 0x80000000) != 0 )
  {
    v25 = WPP_GLOBAL_Control;
    if ( WPP_GLOBAL_Control == (PDEVICE_OBJECT)&WPP_GLOBAL_Control
      || (HIDWORD(WPP_GLOBAL_Control->Timer) & 1) == 0
      || BYTE1(WPP_GLOBAL_Control->Timer) < 2u )
    {
      return IsReparseBufferSupported;
    }
    v26 = 24i64;
    goto LABEL_163;
  }
  if ( (element_1_Data & 0x10) != 0 )
    return IsReparseBufferSupported;

通过构造ElementInfos[0]和ElementInfos[1]可以通过HsmpRpValidateBuffer校验，而后漏洞触发点会读取ElementInfos[10]的数据和Length通过memcpy进行拷贝，所以还需要构造ElementInfos[10]的数据，并且ElementInfos[10]的Length需要超过目标缓冲区，特别的在计算CRC32后，需要通过RtlCompressBuffer压缩目标数据，并放入到FileData处。

构造多大的缓冲区？根据前面补丁分析，在补丁中限制了ReparseDataLength < 0x4000，所以超过四千的部分会造成溢出，如果想溢出8个字节则需要构造0x4008 + 8 = 0x4010，依此类推，在构造缓冲区时。

如何将构造好的数据传递给驱动并在目标位置触发呢？在网上查到有类似漏洞分析文章 Windows云文件迷你过滤器驱动程序中的提权漏洞(CVE-2021-31969)，不难看出CVE-2021-31969修复和本次分析的漏洞CVE-2023-36036修复位置类似，都对ReparseDataLength进行了判断，所以本次PoC编写也可以借鉴。

在CVE-2021-31969分析文章中贴出了部分PoC，结合这部分PoC和前面的结构体，写出PoC也就不难了。

动态调试

在如下两个位置下断点

bp cldflt!HsmpRpCommitNoLock
bp cldflt!HsmpRpCommitNoLock+0x13de

运行poc，可以看到已经进入HsmpRpCommitNoLock函数

1: kd> g
Breakpoint 0 hit
cldflt!HsmpRpCommitNoLock:
fffff804`6f6a1e88 48895c2420      mov     qword ptr [rsp+20h],rbx

继续运行，触发第二个断点

0: kd> g
Breakpoint 1 hit
cldflt!HsmpRpCommitNoLock+0x13de:
fffff804`6f6a3266 e81571faff      call    cldflt!memcpy (fffff804`6f64a380)

此时memmove已经被优化为memcpy，而要拷贝的长度为0x3f94，dst所在的堆大小为0x4000，dst指向偏移0x74处，最多有0x3f8c大小，所以memcpy拷贝时会越界写入8个字节，造成堆溢出。

1: kd> rr8
r8=0000000000003f94
1: kd> !pool rcx
Pool page ffffd980717f7074 region is Paged pool
*ffffd980717f7000 : large page allocation, tag is HsRp, size is 0x4000 bytes
		Owning component : Unknown (update pooltag.txt)

继续运行，则在memcpy内部触发异常，因为尝试往未分配的内存里面写入00

0: kd> u
cldflt!memcpy+0x165:
fffff800`8186a4e5 0f2941f0        movaps  xmmword ptr [rcx-10h],xmm0
0: kd> !pool rcx - 0x10
Pool page ffffe5028e4fa000 region is Paged pool
ffffe5028e4fa000 is not a valid large pool allocation, checking large session pool...
ffffe5028e4fa000 is not valid pool. Checking for freed (or corrupt) pool
Address ffffe5028e4fa000 could not be read. It may be a freed, invalid or paged out page
0: kd> rxmm0
mm0=0000000000000000

对应代码为

        if ( v25 )
          *(_OWORD *)(v15 + v25 - 16) = *(_OWORD *)(v15 + v25 - 16 + v13);
        *(__m128 *)(v15 - 0x10) = v14;

以下为调用栈

1: kd> k
 # Child-SP          RetAddr               Call Site
00 fffffb8a`8a45e4f8 fffff804`63717f82     nt!DbgBreakPointWithStatus
01 fffffb8a`8a45e500 fffff804`63717566     nt!KiBugCheckDebugBreak+0x12
02 fffffb8a`8a45e560 fffff804`635fd747     nt!KeBugCheck2+0x946
03 fffffb8a`8a45ec70 fffff804`63638f6f     nt!KeBugCheckEx+0x107
04 fffffb8a`8a45ecb0 fffff804`63430730     nt!MiSystemFault+0x1de5ff
05 fffffb8a`8a45edb0 fffff804`6360d1d8     nt!MmAccessFault+0x400
06 fffffb8a`8a45ef50 fffff804`6f64a4e1     nt!KiPageFault+0x358
07 fffffb8a`8a45f0e8 fffff804`6f6a326b     cldflt!memcpy+0x161
08 fffffb8a`8a45f0f0 fffff804`6f6a983b     cldflt!HsmpRpCommitNoLock+0x13e3
09 fffffb8a`8a45f230 fffff804`6f66f0d7     cldflt!HsmiOpUpdatePlaceholderDirectory+0x57f
0a fffffb8a`8a45f320 fffff804`6f674b65     cldflt!HsmFltProcessUpdatePlaceholder+0x443
0b fffffb8a`8a45f3d0 fffff804`6f6a4504     cldflt!HsmFltProcessHSMControl+0x3d5
0c fffffb8a`8a45f500 fffff804`647264cc     cldflt!HsmFltPreFILE_SYSTEM_CONTROL+0x6a4
0d fffffb8a`8a45f5a0 fffff804`64725f7a     FLTMGR!FltpPerformPreCallbacksWorker+0x36c
0e fffffb8a`8a45f6c0 fffff804`64725021     FLTMGR!FltpPassThroughInternal+0xca
0f fffffb8a`8a45f710 fffff804`6475ae2f     FLTMGR!FltpPassThrough+0x541
10 fffffb8a`8a45f7a0 fffff804`63410665     FLTMGR!FltpFsControl+0xbf
11 fffffb8a`8a45f800 fffff804`6380142c     nt!IofCallDriver+0x55
12 fffffb8a`8a45f840 fffff804`63801081     nt!IopSynchronousServiceTail+0x34c
13 fffffb8a`8a45f8e0 fffff804`638d9ed6     nt!IopXxxControlFile+0xc71
14 fffffb8a`8a45fa20 fffff804`63610ef5     nt!NtFsControlFile+0x56
15 fffffb8a`8a45fa90 00007ff9`c648d704     nt!KiSystemServiceCopyEnd+0x25
16 00000056`01aff5b8 00007ff6`5e59167f     ntdll!NtFsControlFile+0x14
17 00000056`01aff5c0 00000000`000001bc     0x00007ff6`5e59167f
18 00000056`01aff5c8 00000000`00000000     0x1bc

PoC会在过几天上传到GitHub

https://github.com/Chestnuts4/POC

小结

本次漏洞分析离不开业内前辈逆向得出的_HSM_REPARSE_DATA结构体信息，这个结构体微软没有公开的文档，相关资料也很少。可以看到早在2018年，就已经逆向出了HSM相关数据结构信息。目前只有这一个仓库有相关信息，向前辈致敬。

/*****************************************************************************/
/* ReparseDataHsm.h                       Copyright (c) Ladislav Zezula 2018 */
/*---------------------------------------------------------------------------*/
/* Interface of the HSM reparse data structures                              */
/*---------------------------------------------------------------------------*/
/*   Date    Ver   Who  Comment                                              */
/* --------  ----  ---  -------                                              */
/* 06.09.18  1.00  Lad  The first version of ReparseDataHsm.h                */
/*****************************************************************************/

这里引用一下前辈的主页。

https://www.zezula.net/en/tools/main.html

整体来看，这个漏洞原理和触发方式较为简单，在使用memcpy之前没有校验长度，而修复也简单，再解压之前验证长度是否超过0x4000，超过则认为数据有错，进入到错误逻辑，从而在源头阻止了触发漏洞逻辑。

在漏洞修复处在修复上个整数下溢的漏洞时，开发人员只修复当时的整数下溢漏洞，没有去考虑长度会不会过长，某些程度来说这也是开发的粗心大意导致了这个漏洞留到现在。

在编写PoC参考了其他安全研究员已有的分析。

至于Exploit部分还得再研究一下。

参考链接

https://msrc.microsoft.com/update-guide/vulnerability/CVE-2023-36036

https://zhuanlan.zhihu.com/p/392194464

https://github.com/microsoft/Windows-classic-samples/tree/main/Samples/CloudMirror

https://learn.microsoft.com/en-us/windows/win32/cfapi/cloud-filter-reference

https://learn.microsoft.com/zh-cn/windows/win32/cfapi/cloud-files-functions

https://learn.microsoft.com/en-us/windows/win32/api/_cloudapi/

Created at 2023-11-24T15:49:32+08:00

创建于:Friday, November 24,2023

最后修改于: Monday, November 27,2023