Skip to end of metadata
Go to start of metadata

TOCTOU (time-of-check, time-of-use) race condition is possible when two or more concurrent processes are operating on a shared file system [Seacord 2013b]. Typically, the first access is a check to verify some attribute of the file, followed by a call to use the file. An attacker can alter the file between the two accesses, or replace the file with a symbolic or hard link to a different file. These TOCTOU conditions can be exploited when a program performs two or more file operations on the same file name or path name.

A program that performs two or more file operations on a single file name or path name creates a race window between the two file operations. This race window comes from the assumption that the file name or path name refers to the same resource both times. If an attacker can modify the file, remove it, or replace it with a different file, then this assumption will not hold.

Noncompliant Code Example

If an existing file is opened for writing with the w mode argument, the file's previous contents (if any) are destroyed. This noncompliant code example tries to prevent an existing file from being overwritten by first opening it for reading before opening it for writing. An attacker can exploit the race window between the two calls to fopen() to overwrite an existing file.

#include <stdio.h>

void open_some_file(const char *file) {
  FILE *f = fopen(file, "r");
  if (NULL != f) {
    /* File exists, handle error */
  } else {
    if (fclose(f) == EOF) {
      /* Handle error */
    }
    f = fopen(file, "w");
    if (NULL == f) {
      /* Handle error */
    }
 
    /* Write to file */
    if (fclose(f) == EOF) {
      /* Handle error */
    }
  }
}

Compliant Solution

This compliant solution invokes fopen() at a single location and uses the x mode of fopen(), which was added in C11. This mode causes fopen() to fail if the file exists. This check and subsequent open is performed without creating a race window. The x mode provides exclusive access to the file only if the host environment provides this support.

#include <stdio.h>

void open_some_file(const char *file) {
  FILE *f = fopen(file, "wx");
  if (NULL == f) {
    /* Handle error */
  }
  /* Write to file */
  if (fclose(f) == EOF) {
    /* Handle error */
  }
}

Compliant Solution (POSIX)

This compliant solution uses the O_CREAT and O_EXCL flags of POSIX's open() function. These flags cause open() to fail if the file exists.

#include <stdio.h>
#include <unistd.h>
#include <fcntl.h>

void open_some_file(const char *file) {
  int fd = open(file, O_CREAT | O_EXCL | O_WRONLY);
  if (-1 != fd) {
    FILE *f = fdopen(fd, "w");
    if (NULL != f) {
      /* Write to file */

      if (fclose(f) == EOF) {
        /* Handle error */
      }
    } else {
      if (close(fd) == -1) {
        /* Handle error */
      }
    }
  }
}

Exceptions

FIO45-C-EX1: TOCTOU race conditions require that the vulnerable process is more privileged than the attacker; otherwise there is nothing to be gained from a successful attack. An unprivileged process is not subject to this rule.

FIO45-C-EX2: Accessing a file name or path name multiple times is permitted if the file referenced resides in a secure directory. (For more information, see FIO15-C. Ensure that file operations are performed in a secure directory.)

FIO45-C-EX3: Accessing a file name or path name multiple times is permitted if the program can verify that every operation operates on the same file.

This POSIX code example verifies that each subsequent file access operates on the same file. In POSIX, every file can be uniquely identified by using its device and i-node attributes. This code example checks that a file name refers to a regular file (and not a directory, symbolic link, or other special file) by invoking lstat(). This call also retrieves its device and i-node. The file is subsequently opened. Finally, the program verifies that the file that was opened is the same one (matching device and i-nodes) as the file that was confirmed as a regular file.


#include <sys/stat.h>
#include <fcntl.h>

int open_regular_file(char *filename, int flags) {
  struct stat lstat_info;
  struct stat fstat_info;
  int f;
 
  if (lstat(filename, &lstat_info) == -1) {
    /* File does not exist, handle error */
  }
 
  if (!S_ISREG(lstat_info.st_mode)) {
    /* File is not a regular file, handle error */
  }
 
  if ((f = open(filename, flags)) == -1) {
    /* File has disappeared, handle error */
  }
 
  if (fstat(f, &fstat_info) == -1) {
    /* Handle error */
  }
 
  if (lstat_info.st_ino != fstat_info.st_ino  ||
      lstat_info.st_dev != fstat_info.st_dev) {
    /* Open file is not the expected regular file, handle error */
  }
 
  /* f is the expected regular open file */
  return f;
}

Risk Assessment

TOCTOU race conditions can result in unexpected behavior, including privilege escalation.

Rule

Severity

Likelihood

Remediation Cost

Priority

Level

FIO45-C

High

Probable

High

P6

L2

Automated Detection

Tool

Version

Checker

Description

CodeSonar
5.1p0
IO.RACEFile system race condition
Coverity
2017.07

TOCTOU

Implemented
Klocwork
2018
SV.TOCTOU.FILE_ACCESS
LDRA tool suite
9.7.1
75 DPartially implemented
Parasoft C/C++test

10.4.2

CERT_C-FIO45-a

Avoid race conditions while accessing files

Polyspace Bug Finder

R2018a

File access between time of check and use (TOCTOU)

File or folder might change state due to access race

Related Vulnerabilities

Search for vulnerabilities resulting from the violation of this rule on the CERT website.

Bibliography

[Seacord 2013b]Chapter 7, "Files"



  

2 Comments

  1. The example in FIO45-EX3 is flawed. Consider this test:

    $ echo abc >foo && stat foo && echo xyz >bar && rm foo && ln -s bar foo && stat foo
    File: ‘foo’
    Size: 4 Blocks: 8 IO Block: 4096 regular file
    Device: fd02h/64770d Inode: 17706253 Links: 1
    Access: (0664/-rw-rw-r--) Uid: ( 1000/ msebor) Gid: ( 1000/ msebor)
    Context: unconfined_u:object_r:user_home_t:s0
    Access: 2014-04-04 14:09:50.000780213 -0600
    Modify: 2014-04-04 14:09:50.000780213 -0600
    Change: 2014-04-04 14:09:50.000780213 -0600
    Birth: -
    File: ‘foo’ -> ‘bar’
    Size: 3 Blocks: 0 IO Block: 4096 symbolic link
    Device: fd02h/64770d Inode: 17706253 Links: 1
    Access: (0777/lrwxrwxrwx) Uid: ( 1000/ msebor) Gid: ( 1000/ msebor)
    Context: unconfined_u:object_r:user_home_t:s0
    Access: 2014-04-04 14:09:50.003780171 -0600
    Modify: 2014-04-04 14:09:50.003780171 -0600
    Change: 2014-04-04 14:09:50.003780171 -0600
    Birth: -

    It shows that there is a window between the call to access and open when it's possible to remove the original file operated on by access and replace it with a link to a different file without the open_regular_file function detecting the change, thus making it susceptible to the TOCTOU race condition discussed above.

    1. Fascinating. AFAICT this race exploit works because Martin's script relies on the fact that the malicious foo is created immediately after the benign foo is deleted. Which means the OS is presumably reusing the same inode for both files. IOW the exploit script has its own race window :) It seems to fail if you create a second file between the rm and ln commands, and it fails if you merely move foo rather than delete it.

      So the exploit is useless if the attacker is unable to delete the file in question. I've amended the exception to specify this.