Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 10 Sep 2020 18:46:10 +0200
From: Mickaël Salaün <>
Cc: Mickaël Salaün <>,
	Aleksa Sarai <>,
	Alexei Starovoitov <>,
	Al Viro <>,
	Andrew Morton <>,
	Andy Lutomirski <>,
	Arnd Bergmann <>,
	Casey Schaufler <>,
	Christian Brauner <>,
	Christian Heimes <>,
	Daniel Borkmann <>,
	Deven Bowers <>,
	Dmitry Vyukov <>,
	Eric Biggers <>,
	Eric Chiang <>,
	Florian Weimer <>,
	James Morris <>,
	Jan Kara <>,
	Jann Horn <>,
	Jonathan Corbet <>,
	Kees Cook <>,
	Lakshmi Ramasubramanian <>,
	Matthew Garrett <>,
	Matthew Wilcox <>,
	Michael Kerrisk <>,
	Miklos Szeredi <>,
	Mimi Zohar <>,
	Philippe Trébuchet <>,
	Scott Shell <>,
	Sean Christopherson <>,
	Shuah Khan <>,
	Steve Dower <>,
	Steve Grubb <>,
	Tetsuo Handa <>,
	Thibaut Sautereau <>,
	Vincent Strubel <>,,,,,,
	Thibaut Sautereau <>,
	Mickaël Salaün <>
Subject: [RFC PATCH v9 1/3] fs: Add introspect_access(2) syscall implementation and related sysctl

From: Mickaël Salaün <>

The introspect_access() syscall enables user space tasks to check that
files are allowed to be executed or interpreted by user space.  This may
allow script interpreters to check execution permission before reading
commands from a file, or dynamic linkers to allow shared object loading.
This may be seen as a way for a trusted task (e.g. interpreter) to check
the trustworthiness of files (e.g. scripts) before extending its control
flow graph with new ones originating from these files.

The security policy is consistently managed by the kernel through a
sysctl or implemented by an LSM thanks to the inode_permission hook and
a new kernel flag: MAY_INTROSPECTION_EXEC .

The new sysctl fs.introspection_policy enables system administrators to
enforce two complementary security policies according to the installed
system: enforce the noexec mount option, and enforce executable file
permission.  Indeed, because of compatibility with installed systems,
only system administrators are able to check that this new enforcement
is in line with the system mount points and file permissions.

The underlying idea is to be able to restrict scripts interpretation
according to a policy defined by the system administrator.  For this to
be possible, script interpreters must use introspect_access(2) with the
X_OK mode.  To be fully effective, these interpreters also need to
handle the other ways to execute code: command line parameters (e.g.,
option -e for Perl), module loading (e.g., option -m for Python), stdin,
file sourcing, environment variables, configuration files, etc.
According to the threat model, it may be acceptable to allow some script
interpreters (e.g. Bash) to interpret commands from stdin, may it be a
TTY or a pipe, because it may not be enough to (directly) perform

Even without enforced security policy, user space interpreters can use
this syscall to try as much as possible to enforce the system policy at
their level, knowing that it will not break anything on running systems
which do not care about this feature.  However, on systems which want
this feature enforced, there will be knowledgeable people (i.e.  system
administrator who configured fs.introspection_policy deliberately) to
manage it.

Because introspect_access(2) with X_OK mode is a mean to enforce a
system-wide security policy (but not application-centric policies), it
does not make sense for user space to check the sysctl value.  Indeed,
this new flag only enables to extend the system ability to enforce a
policy thanks to (some trusted) user space collaboration.  Moreover,
additional security policies could be managed by LSMs.  This is a
best-effort approach from the application developer point of view:

introspect_access(2) with X_OK should not be confused with the O_EXEC
flag (for open) which is intended for execute-only, which obviously
doesn't work for scripts.  However, a similar behavior could be
implemented in user space with O_PATH:

Being able to restrict execution also enables to protect the kernel by
restricting arbitrary syscalls that an attacker could perform with a
crafted binary or certain script languages.  It also improves multilevel
isolation by reducing the ability of an attacker to use side channels
with specific code.  These restrictions can natively be enforced for ELF
binaries (with the noexec mount option) but require this kernel
extension to properly handle scripts (e.g. Python, Perl).  To get a
consistent execution policy, additional memory restrictions should also
be enforced (e.g. thanks to SELinux).

This is a new implementation of a patch initially written by
Vincent Strubel for CLIP OS 4:
This patch has been used for more than 12 years with customized script
interpreters.  Some examples (with the original O_MAYEXEC) can be found

Co-developed-by: Thibaut Sautereau <>
Signed-off-by: Thibaut Sautereau <>
Signed-off-by: Mickaël Salaün <>
Cc: Al Viro <>
Cc: Andrew Morton <>
Cc: Arnd Bergmann <>
Cc: Jonathan Corbet <>
Cc: Kees Cook <>
Cc: Vincent Strubel <>

Changes since v8:
* Add a dedicated syscall introspect_access() (requested by Al Viro).
* Rename the sysctl fs.interpreted_access to fs.introspection_policy .
* Update documentation.

Changes since v7:
* Replaces openat2/O_MAYEXEC with faccessat2/X_OK/AT_INTERPRETED .
  Switching to an FD-based syscall was suggested by Al Viro and Jann
* Handle special file descriptors.
* Add a compatibility mode for execute/read check.
* Move the sysctl policy from fs/namei.c to fs/open.c for the new
* Rename the sysctl from fs.open_mayexec_enforce to
  fs.interpreted_access .
* Update documentation accordingly.

Changes since v6:
* Allow opening pipes, block devices and character devices with
  O_MAYEXEC when there is no enforced policy, but forbid any non-regular
  file opened with O_MAYEXEC otherwise (i.e. for any enforced policy).
* Add a paragraph about the non-regular files policy.
* Move path_noexec() calls out of the fast-path (suggested by Kees
* Do not set __FMODE_EXEC for now because of inconsistent behavior:
* Returns EISDIR when opening a directory with O_MAYEXEC.
* Removed Deven Bowers and Kees Cook Reviewed-by tags because of the
  current update.

Changes since v5:
* Remove the static enforcement configuration through Kconfig because it
  makes the code more simple like this, and because the current sysctl
  configuration can only be set with CAP_SYS_ADMIN, the same way mount
  options (i.e. noexec) can be set.  If an harden distro wants to
  enforce a configuration, it should restrict capabilities or sysctl
  configuration.  Furthermore, an LSM can easily leverage O_MAYEXEC to
  fit its need.
* Move checks from inode_permission() to may_open() and make the error
  codes more consistent according to file types (in line with a previous
  commit): opening a directory with O_MAYEXEC returns EISDIR and other
  non-regular file types may return EACCES.
* In may_open(), when OMAYEXEC_ENFORCE_FILE is set, replace explicit
  call to generic_permission() with an artificial MAY_EXEC to avoid
  double calls.  This makes sense especially when an LSM policy forbids
  execution of a file.
* Replace the custom proc_omayexec() with
  proc_dointvec_minmax_sysadmin(), and then replace the CAP_MAC_ADMIN
  check with a CAP_SYS_ADMIN one (suggested by Kees Cook and Stephen
* Use BIT() (suggested by Kees Cook).
* Rename variables (suggested by Kees Cook).
* Reword the kconfig help.
* Import the documentation patch (suggested by Kees Cook):
* Update documentation and add article.

Changes since v4:
* Add kernel configuration options to enforce O_MAYEXEC at build time,
  and disable the sysctl in such case (requested by James Morris).
* Reword commit message.

Changes since v3:
* Switch back to O_MAYEXEC, but only handle it with openat2(2) which
  checks unknown flags (suggested by Aleksa Sarai). Cf.

Changes since v2:
* Replace O_MAYEXEC with RESOLVE_MAYEXEC from openat2(2).  This change
  enables to not break existing application using bogus O_* flags that
  may be ignored by current kernels by using a new dedicated flag, only
  usable through openat2(2) (suggested by Jeff Layton).  Using this flag
  will results in an error if the running kernel does not support it.
  User space needs to manage this case, as with other RESOLVE_* flags.
  The best effort approach to security (for most common distros) will
  simply consists of ignoring such an error and retry without
  RESOLVE_MAYEXEC.  However, a fully controlled system may which to
  error out if such an inconsistency is detected.
* Cosmetic changes.

Changes since v1:
* Set __FMODE_EXEC when using O_MAYEXEC to make this information
  available through the new fanotify/FAN_OPEN_EXEC event (suggested by
  Jan Kara and Matthew Bobrowski):
* Move code from Yama to the FS subsystem (suggested by Kees Cook).
* Make omayexec_inode_permission() static (suggested by Jann Horn).
* Use mode 0600 for the sysctl.
* Only match regular files (not directories nor other types), which
  follows the same semantic as commit 73601ea5b7b1 ("fs/open.c: allow
  opening only regular files during execve()").
 Documentation/admin-guide/sysctl/fs.rst | 50 ++++++++++++++++
 fs/open.c                               | 79 +++++++++++++++++++++++++
 include/linux/fs.h                      |  3 +
 include/linux/syscalls.h                |  1 +
 kernel/sysctl.c                         | 12 +++-
 5 files changed, 143 insertions(+), 2 deletions(-)

diff --git a/Documentation/admin-guide/sysctl/fs.rst b/Documentation/admin-guide/sysctl/fs.rst
index f48277a0a850..2f244e968a1d 100644
--- a/Documentation/admin-guide/sysctl/fs.rst
+++ b/Documentation/admin-guide/sysctl/fs.rst
@@ -36,6 +36,7 @@ Currently, these files are in /proc/sys/fs:
 - inode-max
 - inode-nr
 - inode-state
+- introspection_policy
 - nr_open
 - overflowuid
 - overflowgid
@@ -165,6 +166,55 @@ system needs to prune the inode list instead of allocating
+An interpreter can call :manpage:`introspect_access(2)` with an ``X_OK`` mode
+to check that opened regular files are expected to be executable.  If the file
+is not identified as executable, then the syscall returns -EACCES.  This may
+allow a script interpreter to check executable permission before reading
+commands from a file, or a dynamic linker to only load executable shared
+objects.  One interesting use case is to enforce a "write xor execute" policy
+through interpreters.
+The ability to restrict code execution must be thought as a system-wide policy,
+which first starts by restricting mount points with the ``noexec`` option.
+This option is also automatically applied to special filesystems such as /proc .
+This prevents files on such mount points to be directly executed by the kernel
+or mapped as executable memory (e.g. libraries).  With script interpreters
+using :manpage:`introspect_access(2)`, the executable permission can then be
+checked before reading commands from files.  This makes it possible to enforce
+the ``noexec`` at the interpreter level, and thus propagates this security
+policy to scripts.  To be fully effective, these interpreters also need to
+handle the other ways to execute code: command line parameters (e.g., option
+``-e`` for Perl), module loading (e.g., option ``-m`` for Python), stdin, file
+sourcing, environment variables, configuration files, etc.  According to the
+threat model, it may be acceptable to allow some script interpreters (e.g.
+Bash) to interpret commands from stdin, may it be a TTY or a pipe, because it
+may not be enough to (directly) perform syscalls.
+There are two complementary security policies: enforce the ``noexec`` mount
+option, and enforce executable file permission.  These policies are handled by
+the ``fs.introspection_policy`` sysctl (writable only with ``CAP_SYS_ADMIN``)
+as a bitmask:
+1 - Mount restriction: checks that the mount options for the underlying VFS
+    mount do not prevent execution.
+2 - File permission restriction: checks that the file is marked as
+    executable for the current process (e.g., POSIX permissions, ACLs).
+Note that as long as a policy is enforced, checking any non-regular file with
+:manpage:`introspect_access(2)` returns -EACCES (e.g. TTYs, pipe), even when
+such a file is marked as executable or is on an executable mount point.
+Code samples can be found in
+tools/testing/selftests/interpreter/introspection_policy_test.c and interpreter
+patches (for the original O_MAYEXEC) are available at
+ .
+See also an overview article: .
 overflowgid & overflowuid
diff --git a/fs/open.c b/fs/open.c
index 9af548fb841b..390cef411236 100644
--- a/fs/open.c
+++ b/fs/open.c
@@ -32,6 +32,7 @@
 #include <linux/ima.h>
 #include <linux/dnotify.h>
 #include <linux/compat.h>
+#include <linux/sysctl.h>
 #include "internal.h"
@@ -482,6 +483,84 @@ SYSCALL_DEFINE2(access, const char __user *, filename, int, mode)
 	return do_faccessat(AT_FDCWD, filename, mode, 0);
+int sysctl_introspection_policy __read_mostly;
+SYSCALL_DEFINE3(introspect_access, const int, fd, const int, mode, const int, flags)
+	int mask, err = -EACCES;
+	struct fd f;
+	struct inode *inode;
+	if (flags)
+		return -EINVAL;
+	/* Only allows X_OK for now. */
+	if (mode != S_IXOTH)
+		return -EINVAL;
+	mask = MAY_EXEC;
+	f = fdget(fd);
+	if (!f.file)
+		return -EBADF;
+	inode = d_backing_inode(f.file->f_path.dentry);
+	/*
+	 * For compatibility reasons, without a defined security policy (via
+	 * sysctl or LSM), we must map the execute permission to the read
+	 * permission.  Indeed, from user space point of view, being able to
+	 * execute data (e.g. scripts) implies to be able to read this data.
+	 *
+	 * The MAY_INTROSPECTION_EXEC bit is set to enable LSMs to add custom
+	 * checks, while being compatible with current policies.
+	 */
+	if ((mask & MAY_EXEC)) {
+		/*
+		 * If there is a system-wide execute policy enforced, then
+		 * forbids access to non-regular files and special superblocks.
+		 */
+		if ((sysctl_introspection_policy & (INTROSPECTION_EXEC_MOUNT |
+			if (!S_ISREG(inode->i_mode))
+				goto out_fd;
+			/*
+			 * Denies access to pseudo filesystems that will never
+			 * be mountable (e.g. sockfs, pipefs) but can still be
+			 * reachable through /proc/self/fd, or memfd-like file
+			 * descriptors, or nsfs-like files.
+			 *
+			 * According to the tests, SB_NOEXEC seems to be only
+			 * used by proc and nsfs filesystems.  Is it correct?
+			 */
+			if ((f.file->f_path.dentry->d_sb->s_flags &
+				goto out_fd;
+		}
+		if ((sysctl_introspection_policy & INTROSPECTION_EXEC_MOUNT) &&
+				path_noexec(&f.file->f_path))
+			goto out_fd;
+		/*
+		 * For compatibility reasons, if the system-wide policy doesn't
+		 * enforce file permission checks, then replaces the execute
+		 * permission request with a read permission request.
+		 */
+		if (!(sysctl_introspection_policy & INTROSPECTION_EXEC_FILE))
+			mask &= ~MAY_EXEC;
+		/* To be executed *by* user space, files must be readable. */
+		mask |= MAY_READ;
+	}
+	err = inode_permission(inode, mask | MAY_ACCESS);
+	fdput(f);
+	return err;
 SYSCALL_DEFINE1(chdir, const char __user *, filename)
 	struct path path;
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 7519ae003a08..3f9c4fe199ce 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -83,6 +83,7 @@ extern int sysctl_protected_symlinks;
 extern int sysctl_protected_hardlinks;
 extern int sysctl_protected_fifos;
 extern int sysctl_protected_regular;
+extern int sysctl_introspection_policy;
 typedef __kernel_rwf_t rwf_t;
@@ -101,6 +102,8 @@ typedef int (dio_iodone_t)(struct kiocb *iocb, loff_t offset,
 #define MAY_CHDIR		0x00000040
 /* called from RCU mode, don't block */
 #define MAY_NOT_BLOCK		0x00000080
+/* introspection accesses, cf. introspect_access(2) */
+#define MAY_INTROSPECTION_EXEC	0x00000100
  * flags in file.f_mode.  Note that FMODE_READ and FMODE_WRITE must correspond
diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
index 75ac7f8ae93c..e8cb6846dea2 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -429,6 +429,7 @@ asmlinkage long sys_fallocate(int fd, int mode, loff_t offset, loff_t len);
 asmlinkage long sys_faccessat(int dfd, const char __user *filename, int mode);
 asmlinkage long sys_faccessat2(int dfd, const char __user *filename, int mode,
 			       int flags);
+asmlinkage long sys_introspect_access(int fd, int mode, int flags);
 asmlinkage long sys_chdir(const char __user *filename);
 asmlinkage long sys_fchdir(unsigned int fd);
 asmlinkage long sys_chroot(const char __user *filename);
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index 09e70ee2332e..d9c2aca9a0c0 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -113,6 +113,7 @@ static int sixty = 60;
 static int __maybe_unused neg_one = -1;
 static int __maybe_unused two = 2;
+static int __maybe_unused three = 3;
 static int __maybe_unused four = 4;
 static unsigned long zero_ul;
 static unsigned long one_ul = 1;
@@ -887,7 +888,6 @@ static int proc_taint(struct ctl_table *table, int write,
 	return err;
 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 				void *buffer, size_t *lenp, loff_t *ppos)
@@ -896,7 +896,6 @@ static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 	return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
  * struct do_proc_dointvec_minmax_conv_param - proc_dointvec_minmax() range checking structure
@@ -3293,6 +3292,15 @@ static struct ctl_table fs_table[] = {
 		.extra1		= SYSCTL_ZERO,
 		.extra2		= &two,
+	{
+		.procname       = "introspection_policy",
+		.data           = &sysctl_introspection_policy,
+		.maxlen         = sizeof(int),
+		.mode           = 0600,
+		.proc_handler	= proc_dointvec_minmax_sysadmin,
+		.extra1		= SYSCTL_ZERO,
+		.extra2		= &three,
+	},
 		.procname	= "binfmt_misc",

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.