Merge pull request #109342 from Mic92/wrappers

This commit is contained in:
Jörg Thalheim 2021-01-27 14:32:38 +00:00 committed by GitHub
commit 0998756db2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 181 additions and 174 deletions

View file

@ -10,16 +10,8 @@ let
(n: v: (if v ? program then v else v // {program=n;}))
wrappers);
securityWrapper = pkgs.stdenv.mkDerivation {
name = "security-wrapper";
phases = [ "installPhase" "fixupPhase" ];
buildInputs = [ pkgs.libcap pkgs.libcap_ng pkgs.linuxHeaders ];
hardeningEnable = [ "pie" ];
installPhase = ''
mkdir -p $out/bin
$CC -Wall -O2 -DWRAPPER_DIR=\"${parentWrapperDir}\" \
-lcap-ng -lcap ${./wrapper.c} -o $out/bin/security-wrapper
'';
securityWrapper = pkgs.callPackage ./wrapper.nix {
inherit parentWrapperDir;
};
###### Activation script for the setcap wrappers

View file

@ -4,15 +4,17 @@
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/xattr.h>
#include <fcntl.h>
#include <dirent.h>
#include <assert.h>
#include <errno.h>
#include <linux/capability.h>
#include <sys/capability.h>
#include <sys/prctl.h>
#include <limits.h>
#include <cap-ng.h>
#include <stdint.h>
#include <syscall.h>
#include <byteswap.h>
// Make sure assertions are not compiled out, we use them to codify
// invariants about this program and we want it to fail fast and
@ -23,182 +25,172 @@ extern char **environ;
// The WRAPPER_DIR macro is supplied at compile time so that it cannot
// be changed at runtime
static char * wrapperDir = WRAPPER_DIR;
static char *wrapper_dir = WRAPPER_DIR;
// Wrapper debug variable name
static char * wrapperDebug = "WRAPPER_DEBUG";
static char *wrapper_debug = "WRAPPER_DEBUG";
// Update the capabilities of the running process to include the given
// capability in the Ambient set.
static void set_ambient_cap(cap_value_t cap)
{
capng_get_caps_process();
#define CAP_SETPCAP 8
if (capng_update(CAPNG_ADD, CAPNG_INHERITABLE, (unsigned long) cap))
{
perror("cannot raise the capability into the Inheritable set\n");
exit(1);
#if __BYTE_ORDER == __BIG_ENDIAN
#define LE32_TO_H(x) bswap_32(x)
#else
#define LE32_TO_H(x) (x)
#endif
int get_last_cap(unsigned *last_cap) {
FILE* file = fopen("/proc/sys/kernel/cap_last_cap", "r");
if (file == NULL) {
int saved_errno = errno;
fprintf(stderr, "failed to open /proc/sys/kernel/cap_last_cap: %s\n", strerror(errno));
return -saved_errno;
}
capng_apply(CAPNG_SELECT_CAPS);
if (prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_RAISE, (unsigned long) cap, 0, 0))
{
perror("cannot raise the capability into the Ambient set\n");
exit(1);
int res = fscanf(file, "%u", last_cap);
if (res == EOF) {
int saved_errno = errno;
fprintf(stderr, "could not read number from /proc/sys/kernel/cap_last_cap: %s\n", strerror(errno));
return -saved_errno;
}
fclose(file);
return 0;
}
// Given the path to this program, fetch its configured capability set
// (as set by `setcap ... /path/to/file`) and raise those capabilities
// into the Ambient set.
static int make_caps_ambient(const char *selfPath)
{
cap_t caps = cap_get_file(selfPath);
if(!caps)
{
if(getenv(wrapperDebug))
fprintf(stderr, "no caps set or could not retrieve the caps for this file, not doing anything...");
static int make_caps_ambient(const char *self_path) {
struct vfs_ns_cap_data data = {};
int r = getxattr(self_path, "security.capability", &data, sizeof(data));
if (r < 0) {
if (errno == ENODATA) {
// no capabilities set
return 0;
}
fprintf(stderr, "cannot get capabilities for %s: %s", self_path, strerror(errno));
return 1;
}
// We use `cap_to_text` and iteration over the tokenized result
// string because, as of libcap's current release, there is no
// facility for retrieving an array of `cap_value_t`'s that can be
// given to `prctl` in order to lift that capability into the
// Ambient set.
//
// Some discussion was had around shot-gunning all of the
// capabilities we know about into the Ambient set but that has a
// security smell and I deemed the risk of the current
// implementation crashing the program to be lower than the risk
// of a privilege escalation security hole being introduced by
// raising all capabilities, even ones we didn't intend for the
// program, into the Ambient set.
//
// `cap_t` which is returned by `cap_get_*` is an opaque type and
// even if we could retrieve the bitmasks (which, as far as I can
// tell we cannot) in order to get the `cap_value_t`
// representation for each capability we would have to take the
// total number of capabilities supported and iterate over the
// sequence of integers up-to that maximum total, testing each one
// against the bitmask ((bitmask >> n) & 1) to see if it's set and
// aggregating each "capability integer n" that is set in the
// bitmask.
//
// That, combined with the fact that we can't easily get the
// bitmask anyway seemed much more brittle than fetching the
// `cap_t`, transforming it into a textual representation,
// tokenizing the string, and using `cap_from_name` on the token
// to get the `cap_value_t` that we need for `prctl`. There is
// indeed risk involved if the output string format of
// `cap_to_text` ever changes but at this time the combination of
// factors involving the below list have led me to the conclusion
// that the best implementation at this time is reading then
// parsing with *lots of documentation* about why we're doing it
// this way.
//
// 1. No explicit API for fetching an array of `cap_value_t`'s or
// for transforming a `cap_t` into such a representation
// 2. The risk of a crash is lower than lifting all capabilities
// into the Ambient set
// 3. libcap is depended on heavily in the Linux ecosystem so
// there is a high chance that the output representation of
// `cap_to_text` will not change which reduces our risk that
// this parsing step will cause a crash
//
// The preferred method, should it ever be available in the
// future, would be to use libcap API's to transform the result
// from a `cap_get_*` into an array of `cap_value_t`'s that can
// then be given to prctl.
//
// - Parnell
ssize_t capLen;
char* capstr = cap_to_text(caps, &capLen);
cap_free(caps);
// TODO: For now, we assume that cap_to_text always starts its
// result string with " =" and that the first capability is listed
// immediately after that. We should verify this.
assert(capLen >= 2);
capstr += 2;
char* saveptr = NULL;
for(char* tok = strtok_r(capstr, ",", &saveptr); tok; tok = strtok_r(NULL, ",", &saveptr))
{
cap_value_t capnum;
if (cap_from_name(tok, &capnum))
{
if(getenv(wrapperDebug))
fprintf(stderr, "cap_from_name failed, skipping: %s", tok);
}
else if (capnum == CAP_SETPCAP)
{
// Check for the cap_setpcap capability, we set this on the
// wrapper so it can elevate the capabilities to the Ambient
// set but we do not want to propagate it down into the
// wrapped program.
//
// TODO: what happens if that's the behavior you want
// though???? I'm preferring a strict vs. loose policy here.
if(getenv(wrapperDebug))
fprintf(stderr, "cap_setpcap in set, skipping it\n");
}
else
{
set_ambient_cap(capnum);
if(getenv(wrapperDebug))
fprintf(stderr, "raised %s into the Ambient capability set\n", tok);
}
size_t size;
uint32_t version = LE32_TO_H(data.magic_etc) & VFS_CAP_REVISION_MASK;
switch (version) {
case VFS_CAP_REVISION_1:
size = VFS_CAP_U32_1;
break;
case VFS_CAP_REVISION_2:
case VFS_CAP_REVISION_3:
size = VFS_CAP_U32_3;
break;
default:
fprintf(stderr, "BUG! Unsupported capability version 0x%x on %s. Report to NixOS bugtracker\n", version, self_path);
return 1;
}
const struct __user_cap_header_struct header = {
.version = _LINUX_CAPABILITY_VERSION_3,
.pid = getpid(),
};
struct __user_cap_data_struct user_data[2] = {};
for (size_t i = 0; i < size; i++) {
// merge inheritable & permitted into one
user_data[i].permitted = user_data[i].inheritable =
LE32_TO_H(data.data[i].inheritable) | LE32_TO_H(data.data[i].permitted);
}
if (syscall(SYS_capset, &header, &user_data) < 0) {
fprintf(stderr, "failed to inherit capabilities: %s", strerror(errno));
return 1;
}
unsigned last_cap;
r = get_last_cap(&last_cap);
if (r < 0) {
return 1;
}
uint64_t set = user_data[0].permitted | (uint64_t)user_data[1].permitted << 32;
for (unsigned cap = 0; cap < last_cap; cap++) {
if (!(set & (1ULL << cap))) {
continue;
}
// Check for the cap_setpcap capability, we set this on the
// wrapper so it can elevate the capabilities to the Ambient
// set but we do not want to propagate it down into the
// wrapped program.
//
// TODO: what happens if that's the behavior you want
// though???? I'm preferring a strict vs. loose policy here.
if (cap == CAP_SETPCAP) {
if(getenv(wrapper_debug)) {
fprintf(stderr, "cap_setpcap in set, skipping it\n");
}
continue;
}
if (prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_RAISE, (unsigned long) cap, 0, 0)) {
fprintf(stderr, "cannot raise the capability %d into the ambient set: %s\n", cap, strerror(errno));
return 1;
}
if (getenv(wrapper_debug)) {
fprintf(stderr, "raised %d into the ambient capability set\n", cap);
}
}
cap_free(capstr);
return 0;
}
int main(int argc, char * * argv)
{
// I *think* it's safe to assume that a path from a symbolic link
// should safely fit within the PATH_MAX system limit. Though I'm
// not positive it's safe...
char selfPath[PATH_MAX];
int selfPathSize = readlink("/proc/self/exe", selfPath, sizeof(selfPath));
int readlink_malloc(const char *p, char **ret) {
size_t l = FILENAME_MAX+1;
int r;
assert(selfPathSize > 0);
for (;;) {
char *c = calloc(l, sizeof(char));
if (!c) {
return -ENOMEM;
}
// Assert we have room for the zero byte, this ensures the path
// isn't being truncated because it's too big for the buffer.
//
// A better way to handle this might be to use something like the
// whereami library (https://github.com/gpakosz/whereami) or a
// loop that resizes the buffer and re-reads the link if the
// contents are being truncated.
assert(selfPathSize < sizeof(selfPath));
ssize_t n = readlink(p, c, l-1);
if (n < 0) {
r = -errno;
free(c);
return r;
}
// Set the zero byte since readlink doesn't do that for us.
selfPath[selfPathSize] = '\0';
if ((size_t) n < l-1) {
c[n] = 0;
*ret = c;
return 0;
}
free(c);
l *= 2;
}
}
int main(int argc, char **argv) {
char *self_path = NULL;
int self_path_size = readlink_malloc("/proc/self/exe", &self_path);
if (self_path_size < 0) {
fprintf(stderr, "cannot readlink /proc/self/exe: %s", strerror(-self_path_size));
}
// Make sure that we are being executed from the right location,
// i.e., `safeWrapperDir'. This is to prevent someone from creating
// i.e., `safe_wrapper_dir'. This is to prevent someone from creating
// hard link `X' from some other location, along with a false
// `X.real' file, to allow arbitrary programs from being executed
// with elevated capabilities.
int len = strlen(wrapperDir);
if (len > 0 && '/' == wrapperDir[len - 1])
int len = strlen(wrapper_dir);
if (len > 0 && '/' == wrapper_dir[len - 1])
--len;
assert(!strncmp(selfPath, wrapperDir, len));
assert('/' == wrapperDir[0]);
assert('/' == selfPath[len]);
assert(!strncmp(self_path, wrapper_dir, len));
assert('/' == wrapper_dir[0]);
assert('/' == self_path[len]);
// Make *really* *really* sure that we were executed as
// `selfPath', and not, say, as some other setuid program. That
// `self_path', and not, say, as some other setuid program. That
// is, our effective uid/gid should match the uid/gid of
// `selfPath'.
// `self_path'.
struct stat st;
assert(lstat(selfPath, &st) != -1);
assert(lstat(self_path, &st) != -1);
assert(!(st.st_mode & S_ISUID) || (st.st_uid == geteuid()));
assert(!(st.st_mode & S_ISGID) || (st.st_gid == getegid()));
@ -207,33 +199,35 @@ int main(int argc, char * * argv)
assert(!(st.st_mode & (S_IWGRP | S_IWOTH)));
// Read the path of the real (wrapped) program from <self>.real.
char realFN[PATH_MAX + 10];
int realFNSize = snprintf (realFN, sizeof(realFN), "%s.real", selfPath);
assert (realFNSize < sizeof(realFN));
char real_fn[PATH_MAX + 10];
int real_fn_size = snprintf(real_fn, sizeof(real_fn), "%s.real", self_path);
assert(real_fn_size < sizeof(real_fn));
int fdSelf = open(realFN, O_RDONLY);
assert (fdSelf != -1);
int fd_self = open(real_fn, O_RDONLY);
assert(fd_self != -1);
char sourceProg[PATH_MAX];
len = read(fdSelf, sourceProg, PATH_MAX);
assert (len != -1);
assert (len < sizeof(sourceProg));
assert (len > 0);
sourceProg[len] = 0;
char source_prog[PATH_MAX];
len = read(fd_self, source_prog, PATH_MAX);
assert(len != -1);
assert(len < sizeof(source_prog));
assert(len > 0);
source_prog[len] = 0;
close(fdSelf);
close(fd_self);
// Read the capabilities set on the wrapper and raise them in to
// the Ambient set so the program we're wrapping receives the
// the ambient set so the program we're wrapping receives the
// capabilities too!
make_caps_ambient(selfPath);
if (make_caps_ambient(self_path) != 0) {
free(self_path);
return 1;
}
free(self_path);
execve(sourceProg, argv, environ);
execve(source_prog, argv, environ);
fprintf(stderr, "%s: cannot run `%s': %s\n",
argv[0], sourceProg, strerror(errno));
argv[0], source_prog, strerror(errno));
exit(1);
return 1;
}

View file

@ -0,0 +1,21 @@
{ stdenv, linuxHeaders, parentWrapperDir, debug ? false }:
# For testing:
# $ nix-build -E 'with import <nixpkgs> {}; pkgs.callPackage ./wrapper.nix { parentWrapperDir = "/run/wrappers"; debug = true; }'
stdenv.mkDerivation {
name = "security-wrapper";
buildInputs = [ linuxHeaders ];
dontUnpack = true;
hardeningEnable = [ "pie" ];
CFLAGS = [
''-DWRAPPER_DIR="${parentWrapperDir}"''
] ++ (if debug then [
"-Werror" "-Og" "-g"
] else [
"-Wall" "-O2"
]);
dontStrip = debug;
installPhase = ''
mkdir -p $out/bin
$CC $CFLAGS ${./wrapper.c} -o $out/bin/security-wrapper
'';
}