[RFC PATCH 1/2] ima: don't access a file's integrity status before an IMA policy is loaded

From: Test <redacted>

Only after an IMA policy is loaded, check, save, or update the cached
file's integrity status.

Signed-off-by: Mimi Zohar <zohar@linux.ibm.com>
---
 security/integrity/ima/ima_main.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/security/integrity/ima/ima_main.c b/security/integrity/ima/ima_main.c
index 9ef748ea829f..9d1196f712e1 100644
--- a/security/integrity/ima/ima_main.c
+++ b/security/integrity/ima/ima_main.c

@@ -606,6 +606,9 @@ void ima_post_create_tmpfile(struct user_namespace *mnt_userns,
 	struct integrity_iint_cache *iint;
 	int must_appraise;
 
+	if (!ima_policy_flag || !S_ISREG(inode->i_mode))
+		return;
+
 	must_appraise = ima_must_appraise(mnt_userns, inode, MAY_ACCESS,
 					  FILE_CHECK);
 	if (!must_appraise)

@@ -636,6 +639,9 @@ void ima_post_path_mknod(struct user_namespace *mnt_userns,
 	struct inode *inode = dentry->d_inode;
 	int must_appraise;
 
+	if (!ima_policy_flag || !S_ISREG(inode->i_mode))
+		return;
+
 	must_appraise = ima_must_appraise(mnt_userns, inode, MAY_ACCESS,
 					  FILE_CHECK);
 	if (!must_appraise)

-- 
2.27.0

From: Test <redacted>

The integrity's "iint_cache" is initialized at security_init().  Only
after an IMA policy is loaded, which is initialized at late_initcall,
is a file's integrity status stored in the "iint_cache".

All integrity_inode_get() callers first verify that the IMA policy has
been loaded, before calling it.  Yet for some reason, it is still being
called, causing a NULL pointer dereference.

As reported by Dmitry Vyukov:
in qemu:
qemu-system-x86_64       -enable-kvm     -machine q35,nvdimm -cpu
max,migratable=off -smp 4       -m 4G,slots=4,maxmem=16G        -hda
wheezy.img      -kernel arch/x86/boot/bzImage   -nographic -vga std
 -soundhw all     -usb -usbdevice tablet  -bt hci -bt device:keyboard
   -net user,host=10.0.2.10,hostfwd=tcp::10022-:22 -net
nic,model=virtio-net-pci   -object
memory-backend-file,id=pmem1,share=off,mem-path=/dev/zero,size=64M
  -device nvdimm,id=nvdimm1,memdev=pmem1  -append "console=ttyS0
root=/dev/sda earlyprintk=serial rodata=n oops=panic panic_on_warn=1
panic=86400 lsm=smack numa=fake=2 nopcid dummy_hcd.num=8"   -pidfile
vm_pid -m 2G -cpu host

But it crashes on NULL deref in integrity_inode_get during boot:

Run /sbin/init as init process
BUG: kernel NULL pointer dereference, address: 000000000000001c
PGD 0 P4D 0
Oops: 0000 [#1] PREEMPT SMP KASAN
CPU: 3 PID: 1 Comm: swapper/0 Not tainted 5.12.0-rc2+ #97
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS
rel-1.13.0-44-g88ab0c15525c-prebuilt.qemu.org 04/01/2014
RIP: 0010:kmem_cache_alloc+0x2b/0x370 mm/slub.c:2920
Code: 57 41 56 41 55 41 54 41 89 f4 55 48 89 fd 53 48 83 ec 10 44 8b
3d d9 1f 90 0b 65 48 8b 04 25 28 00 00 00 48 89 44 24 08 31 c0 <8b> 5f
1c 4cf
RSP: 0000:ffffc9000032f9d8 EFLAGS: 00010246
RAX: 0000000000000000 RBX: ffff888017fc4f00 RCX: 0000000000000000
RDX: ffff888040220000 RSI: 0000000000000c40 RDI: 0000000000000000
RBP: 0000000000000000 R08: 0000000000000000 R09: ffff888019263627
R10: ffffffff83937cd1 R11: 0000000000000000 R12: 0000000000000c40
R13: ffff888019263538 R14: 0000000000000000 R15: 0000000000ffffff
FS:  0000000000000000(0000) GS:ffff88802d180000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 000000000000001c CR3: 000000000b48e000 CR4: 0000000000750ee0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
PKRU: 55555554
Call Trace:
 integrity_inode_get+0x47/0x260 security/integrity/iint.c:105
 process_measurement+0x33d/0x17e0 security/integrity/ima/ima_main.c:237
 ima_bprm_check+0xde/0x210 security/integrity/ima/ima_main.c:474
 security_bprm_check+0x7d/0xa0 security/security.c:845
 search_binary_handler fs/exec.c:1708 [inline]
 exec_binprm fs/exec.c:1761 [inline]
 bprm_execve fs/exec.c:1830 [inline]
 bprm_execve+0x764/0x19a0 fs/exec.c:1792
 kernel_execve+0x370/0x460 fs/exec.c:1973
 try_to_run_init_process+0x14/0x4e init/main.c:1366
 kernel_init+0x11d/0x1b8 init/main.c:1477
 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294
Modules linked in:
CR2: 000000000000001c
---[ end trace 22d601a500de7d79 ]---

Before calling kmem_cache_alloc(), check that the iint_cache has
been initialized.

Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Mimi Zohar <zohar@linux.ibm.com>
---
 security/integrity/iint.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/security/integrity/iint.c b/security/integrity/iint.c
index 1d20003243c3..80b5ae7bb712 100644
--- a/security/integrity/iint.c
+++ b/security/integrity/iint.c

@@ -97,6 +97,15 @@ struct integrity_iint_cache *integrity_inode_get(struct inode *inode)
 	struct rb_node **p;
 	struct rb_node *node, *parent = NULL;
 	struct integrity_iint_cache *iint, *test_iint;
+	static int once = 0;
+
+	if (!iint_cache) { /* shouldn't get here */
+		if (!once) {
+			dump_stack();
+			once = 1;
+		}
+		return NULL;
+	}
 
 	iint = integrity_iint_find(inode);
 	if (iint)

-- 
2.27.0

On 2021/03/20 5:03, Mimi Zohar wrote:

The integrity's "iint_cache" is initialized at security_init().  Only
after an IMA policy is loaded, which is initialized at late_initcall,
is a file's integrity status stored in the "iint_cache".

All integrity_inode_get() callers first verify that the IMA policy has
been loaded, before calling it.  Yet for some reason, it is still being
called, causing a NULL pointer dereference.

As reported by Dmitry Vyukov:
in qemu:
qemu-system-x86_64       -enable-kvm     -machine q35,nvdimm -cpu
max,migratable=off -smp 4       -m 4G,slots=4,maxmem=16G        -hda
wheezy.img      -kernel arch/x86/boot/bzImage   -nographic -vga std
 -soundhw all     -usb -usbdevice tablet  -bt hci -bt device:keyboard
   -net user,host=10.0.2.10,hostfwd=tcp::10022-:22 -net
nic,model=virtio-net-pci   -object
memory-backend-file,id=pmem1,share=off,mem-path=/dev/zero,size=64M
  -device nvdimm,id=nvdimm1,memdev=pmem1  -append "console=ttyS0
root=/dev/sda earlyprintk=serial rodata=n oops=panic panic_on_warn=1
panic=86400 lsm=smack numa=fake=2 nopcid dummy_hcd.num=8"   -pidfile
vm_pid -m 2G -cpu host

I tried similar command line (without "-enable-kvm" and without "-cpu host"
as I'm running from VMware, without "-soundhw all", without "-machine q35,nvdimm"
and "-device nvdimm,id=nvdimm1,memdev=pmem1" etc.) on 5.12-rc4. While I was finally
able to hit similar crash when I used "-smp 1" instead of "-smp 4", I suspect
this is not a integrity module's problem but a memory initialization/corruption
problem, for I got various different crashes (INT3) at memory allocation when
I was trimming command line options trying to reproduce the same crash.

Dmitry, do you get different crashes by changing command line arguments?

On Mon, Mar 22, 2021 at 8:11 AM Tetsuo Handa
[off-list ref] wrote:

On 2021/03/20 5:03, Mimi Zohar wrote:

quoted

The integrity's "iint_cache" is initialized at security_init().  Only
after an IMA policy is loaded, which is initialized at late_initcall,
is a file's integrity status stored in the "iint_cache".

All integrity_inode_get() callers first verify that the IMA policy has
been loaded, before calling it.  Yet for some reason, it is still being
called, causing a NULL pointer dereference.

As reported by Dmitry Vyukov:
in qemu:
qemu-system-x86_64       -enable-kvm     -machine q35,nvdimm -cpu
max,migratable=off -smp 4       -m 4G,slots=4,maxmem=16G        -hda
wheezy.img      -kernel arch/x86/boot/bzImage   -nographic -vga std
 -soundhw all     -usb -usbdevice tablet  -bt hci -bt device:keyboard
   -net user,host=10.0.2.10,hostfwd=tcp::10022-:22 -net
nic,model=virtio-net-pci   -object
memory-backend-file,id=pmem1,share=off,mem-path=/dev/zero,size=64M
  -device nvdimm,id=nvdimm1,memdev=pmem1  -append "console=ttyS0
root=/dev/sda earlyprintk=serial rodata=n oops=panic panic_on_warn=1
panic=86400 lsm=smack numa=fake=2 nopcid dummy_hcd.num=8"   -pidfile
vm_pid -m 2G -cpu host

I tried similar command line (without "-enable-kvm" and without "-cpu host"
as I'm running from VMware, without "-soundhw all", without "-machine q35,nvdimm"
and "-device nvdimm,id=nvdimm1,memdev=pmem1" etc.) on 5.12-rc4. While I was finally
able to hit similar crash when I used "-smp 1" instead of "-smp 4", I suspect
this is not a integrity module's problem but a memory initialization/corruption
problem, for I got various different crashes (INT3) at memory allocation when
I was trimming command line options trying to reproduce the same crash.

Dmitry, do you get different crashes by changing command line arguments?

No, I have not seen any other crashes, one the reported one.

On 2021/03/20 5:03, Mimi Zohar wrote:

The integrity's "iint_cache" is initialized at security_init().  Only
after an IMA policy is loaded, which is initialized at late_initcall,
is a file's integrity status stored in the "iint_cache".

All integrity_inode_get() callers first verify that the IMA policy has
been loaded, before calling it.  Yet for some reason, it is still being
called, causing a NULL pointer dereference.

qemu-system-x86_64 (...snipped...) lsm=smack (...snipped...)

Hmm, why are you using lsm=smack instead of security=smack ?
Since use of lsm= overrides CONFIG_LSM="lockdown,yama,safesetid,integrity,tomoyo,smack,bpf" settings,
only smack is activated, which means that integrity_iintcache_init() will not be called by

  DEFINE_LSM(integrity) = {
  	.name = "integrity",
  	.init = integrity_iintcache_init,
  };

declaration. That's the reason iint_cache == NULL when integrity_inode_get() is called.

On Tue, 2021-03-23 at 10:46 +0900, Tetsuo Handa wrote:

On 2021/03/20 5:03, Mimi Zohar wrote:

quoted

The integrity's "iint_cache" is initialized at security_init().  Only
after an IMA policy is loaded, which is initialized at late_initcall,
is a file's integrity status stored in the "iint_cache".

All integrity_inode_get() callers first verify that the IMA policy has
been loaded, before calling it.  Yet for some reason, it is still being
called, causing a NULL pointer dereference.

qemu-system-x86_64 (...snipped...) lsm=smack (...snipped...)

Hmm, why are you using lsm=smack instead of security=smack ?
Since use of lsm= overrides CONFIG_LSM="lockdown,yama,safesetid,integrity,tomoyo,smack,bpf" settings,
only smack is activated, which means that integrity_iintcache_init() will not be called by

  DEFINE_LSM(integrity) = {
  	.name = "integrity",
  	.init = integrity_iintcache_init,
  };

declaration. That's the reason iint_cache == NULL when integrity_inode_get() is called.

That's exactly the problem, but since we don't control how the system
is configured or which parameters are supplied on the boot command
line, the kernel needs to at least provide some explanation instead of
dereferencing a NULL pointer.

FYI, "security=" is being deprecated.   From Documentation/admin-
guide/kernel-parameters.txt:

       security=  [SECURITY] Choose a legacy "major" security module to
                        enable at boot. This has been deprecated by the
                        "lsm=" parameter.

Please take a look at the newer version of this patch.   Do you want to
add any tags?

thanks,

Mimi

On 2021/03/23 21:09, Mimi Zohar wrote:

Please take a look at the newer version of this patch.   Do you want to
add any tags?

Oh, I didn't know that you already posted the newer version.

This looks strange. If "lsm=" parameter must include "integrity",
it implies that nobody is allowed to disable "integrity" at boot.
Then, why not unconditionally call integrity_iintcache_init() by
not counting on DEFINE_LSM(integrity) declaration?

 	iint = integrity_iint_find(inode);
 	if (iint)
 		return iint;

On 2021/03/23 22:37, Tetsuo Handa wrote:

On 2021/03/23 21:09, Mimi Zohar wrote:

quoted

Please take a look at the newer version of this patch.   Do you want to
add any tags?

Oh, I didn't know that you already posted the newer version.

quoted

diff --git a/security/integrity/iint.c b/security/integrity/iint.c
index 1d20003243c3..0ba01847e836 100644
--- a/security/integrity/iint.c
+++ b/security/integrity/iint.c

@@ -98,6 +98,14 @@ struct integrity_iint_cache *integrity_inode_get(struct inode *inode)
 	struct rb_node *node, *parent = NULL;
 	struct integrity_iint_cache *iint, *test_iint;
 
+	/*
+	 * The integrity's "iint_cache" is initialized at security_init(),
+	 * unless it is not included in the ordered list of LSMs enabled
+	 * on the boot command line.
+	 */
+	if (!iint_cache)
+		panic("%s: lsm=integrity required.\n", __func__);
+

This looks strange. If "lsm=" parameter must include "integrity",
it implies that nobody is allowed to disable "integrity" at boot.
Then, why not unconditionally call integrity_iintcache_init() by
not counting on DEFINE_LSM(integrity) declaration?

Or, I think below one is also possible.

diff --git a/security/integrity/iint.c b/security/integrity/iint.c
index 1d20003243c3..37afc5168891 100644
--- a/security/integrity/iint.c
+++ b/security/integrity/iint.c

@@ -19,6 +19,7 @@
 #include <linux/uaccess.h>
 #include <linux/security.h>
 #include <linux/lsm_hooks.h>
+#include <linux/sched/mm.h>
 #include "integrity.h"
 
 static struct rb_root integrity_iint_tree = RB_ROOT;

@@ -85,6 +86,20 @@ static void iint_free(struct integrity_iint_cache *iint)
 	kmem_cache_free(iint_cache, iint);
 }
 
+static void init_once(void *foo)
+{
+	struct integrity_iint_cache *iint = foo;
+
+	memset(iint, 0, sizeof(*iint));
+	iint->ima_file_status = INTEGRITY_UNKNOWN;
+	iint->ima_mmap_status = INTEGRITY_UNKNOWN;
+	iint->ima_bprm_status = INTEGRITY_UNKNOWN;
+	iint->ima_read_status = INTEGRITY_UNKNOWN;
+	iint->ima_creds_status = INTEGRITY_UNKNOWN;
+	iint->evm_status = INTEGRITY_UNKNOWN;
+	mutex_init(&iint->mutex);
+}
+
 /**
  * integrity_inode_get - find or allocate an iint associated with an inode
  * @inode: pointer to the inode

@@ -102,6 +117,18 @@ struct integrity_iint_cache *integrity_inode_get(struct inode *inode)
 	if (iint)
 		return iint;
 
+	if (!iint_cache) {
+		static DEFINE_MUTEX(lock);
+		unsigned int flags = memalloc_nofs_save();
+
+		mutex_lock(&lock);
+		if (!iint_cache)
+			iint_cache = kmem_cache_create("iint_cache",
+						       sizeof(struct integrity_iint_cache),
+						       0, SLAB_PANIC, init_once);
+		mutex_unlock(&lock);
+		memalloc_nofs_restore(flags);
+	}
 	iint = kmem_cache_alloc(iint_cache, GFP_NOFS);
 	if (!iint)
 		return NULL;

@@ -150,25 +177,8 @@ void integrity_inode_free(struct inode *inode)
 	iint_free(iint);
 }
 
-static void init_once(void *foo)
-{
-	struct integrity_iint_cache *iint = foo;
-
-	memset(iint, 0, sizeof(*iint));
-	iint->ima_file_status = INTEGRITY_UNKNOWN;
-	iint->ima_mmap_status = INTEGRITY_UNKNOWN;
-	iint->ima_bprm_status = INTEGRITY_UNKNOWN;
-	iint->ima_read_status = INTEGRITY_UNKNOWN;
-	iint->ima_creds_status = INTEGRITY_UNKNOWN;
-	iint->evm_status = INTEGRITY_UNKNOWN;
-	mutex_init(&iint->mutex);
-}
-
 static int __init integrity_iintcache_init(void)
 {
-	iint_cache =
-	    kmem_cache_create("iint_cache", sizeof(struct integrity_iint_cache),
-			      0, SLAB_PANIC, init_once);
 	return 0;
 }
 DEFINE_LSM(integrity) = {

On Tue, 2021-03-23 at 23:01 +0900, Tetsuo Handa wrote:

On 2021/03/23 22:37, Tetsuo Handa wrote:

quoted

On 2021/03/23 21:09, Mimi Zohar wrote:

quoted

Please take a look at the newer version of this patch.   Do you want to
add any tags?

Oh, I didn't know that you already posted the newer version.

quoted

diff --git a/security/integrity/iint.c b/security/integrity/iint.c
index 1d20003243c3..0ba01847e836 100644
--- a/security/integrity/iint.c
+++ b/security/integrity/iint.c

@@ -98,6 +98,14 @@ struct integrity_iint_cache *integrity_inode_get(struct inode *inode)
 	struct rb_node *node, *parent = NULL;
 	struct integrity_iint_cache *iint, *test_iint;
 
+	/*
+	 * The integrity's "iint_cache" is initialized at security_init(),
+	 * unless it is not included in the ordered list of LSMs enabled
+	 * on the boot command line.
+	 */
+	if (!iint_cache)
+		panic("%s: lsm=integrity required.\n", __func__);
+

This looks strange. If "lsm=" parameter must include "integrity",
it implies that nobody is allowed to disable "integrity" at boot.

Integrity isn't always required.  Only when something tries to use it,
does it need to be enabled.  Since both integrity and the integrity
caller are runtime dependent, it is up to the user/admin to specify
"integrity" as an "lsm=" option.

quoted

Then, why not unconditionally call integrity_iintcache_init() by
not counting on DEFINE_LSM(integrity) declaration?

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Mimi

On 2021/03/23 23:47, Mimi Zohar wrote:

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

Mimi

On 2021/03/24 1:13, Mimi Zohar wrote:

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

On Wed, 2021-03-24 at 19:10 +0900, Tetsuo Handa wrote:

On 2021/03/24 1:13, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

User in this case, is anyone rebooting the system and is intentionally
changing the default values, dropping the "integrity" option on the
boot command line.

Mimi

On 2021/03/24 20:10, Mimi Zohar wrote:

On Wed, 2021-03-24 at 19:10 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/24 1:13, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

User in this case, is anyone rebooting the system and is intentionally
changing the default values, dropping the "integrity" option on the
boot command line.

OK. Then, I expect that the system boots instead of calling panic().
That user is explicitly asking for disabling "integrity" for some reason.

On Wed, Mar 24, 2021 at 12:21 PM Tetsuo Handa
[off-list ref] wrote:

On 2021/03/24 20:10, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 19:10 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/24 1:13, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

User in this case, is anyone rebooting the system and is intentionally
changing the default values, dropping the "integrity" option on the
boot command line.

OK. Then, I expect that the system boots instead of calling panic().
That user is explicitly asking for disabling "integrity" for some reason.

That was actually my intention. The prebuilt kernel that I use for
things has all LSMs enabled, but then I needed to try some workload
with only 1 specific LSM, so I gave a different lsm= argument.

On Wed, 2021-03-24 at 12:37 +0100, Dmitry Vyukov wrote:

On Wed, Mar 24, 2021 at 12:21 PM Tetsuo Handa
[off-list ref] wrote:

quoted

On 2021/03/24 20:10, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 19:10 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/24 1:13, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

User in this case, is anyone rebooting the system and is intentionally
changing the default values, dropping the "integrity" option on the
boot command line.

OK. Then, I expect that the system boots instead of calling panic().
That user is explicitly asking for disabling "integrity" for some reason.

That was actually my intention. The prebuilt kernel that I use for
things has all LSMs enabled, but then I needed to try some workload
with only 1 specific LSM, so I gave a different lsm= argument.

IMA/EVM is dependent on "integrity".  Was your intention to also
disable IMA and EVM?  If so, when disabling "integrity", don't load an
IMA policy.

Mimi

On Wed, Mar 24, 2021 at 12:49 PM Mimi Zohar [off-list ref] wrote:

On Wed, 2021-03-24 at 12:37 +0100, Dmitry Vyukov wrote:

quoted

On Wed, Mar 24, 2021 at 12:21 PM Tetsuo Handa
[off-list ref] wrote:

quoted

On 2021/03/24 20:10, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 19:10 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/24 1:13, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

User in this case, is anyone rebooting the system and is intentionally
changing the default values, dropping the "integrity" option on the
boot command line.

OK. Then, I expect that the system boots instead of calling panic().
That user is explicitly asking for disabling "integrity" for some reason.

That was actually my intention. The prebuilt kernel that I use for
things has all LSMs enabled, but then I needed to try some workload
with only 1 specific LSM, so I gave a different lsm= argument.

IMA/EVM is dependent on "integrity".  Was your intention to also
disable IMA and EVM?

I think, yes... or not sure. I was trying to test a bug that requires
a different major LSM and all minor LSMs are presumably irrelevant. I
dropped existing lsm= arg and added something like lsm=apparmor.

If so, when disabling "integrity", don't load an
IMA policy.

I don't really know what this means. I guess it simply comes from the
image? If so, there was no easy way to avoid loading.

On Wed, 2021-03-24 at 12:58 +0100, Dmitry Vyukov wrote:

On Wed, Mar 24, 2021 at 12:49 PM Mimi Zohar [off-list ref] wrote:

quoted

On Wed, 2021-03-24 at 12:37 +0100, Dmitry Vyukov wrote:

quoted

On Wed, Mar 24, 2021 at 12:21 PM Tetsuo Handa
[off-list ref] wrote:

quoted

On 2021/03/24 20:10, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 19:10 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/24 1:13, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

User in this case, is anyone rebooting the system and is intentionally
changing the default values, dropping the "integrity" option on the
boot command line.

OK. Then, I expect that the system boots instead of calling panic().
That user is explicitly asking for disabling "integrity" for some reason.

That was actually my intention. The prebuilt kernel that I use for
things has all LSMs enabled, but then I needed to try some workload
with only 1 specific LSM, so I gave a different lsm= argument.

IMA/EVM is dependent on "integrity".  Was your intention to also
disable IMA and EVM?

I think, yes... or not sure. I was trying to test a bug that requires
a different major LSM and all minor LSMs are presumably irrelevant. I
dropped existing lsm= arg and added something like lsm=apparmor.

quoted

If so, when disabling "integrity", don't load an
IMA policy.

I don't really know what this means. I guess it simply comes from the
image? If so, there was no easy way to avoid loading.

There are a couple of builtin IMA policies, which may be loaded on boot
by specifying on the boot command line "ima_policy=".   Unless the boot
command line "ima_policy=" option is specified, no policy will loaded.

A custom IMA policy may subsequently be loaded, normally in the
initramfs, by echo'ing the file pathname to
/sys/kernel/security/ima/policy.

Mimi

On 3/24/2021 4:58 AM, Dmitry Vyukov wrote:

On Wed, Mar 24, 2021 at 12:49 PM Mimi Zohar [off-list ref] wrote:

quoted

On Wed, 2021-03-24 at 12:37 +0100, Dmitry Vyukov wrote:

quoted

On Wed, Mar 24, 2021 at 12:21 PM Tetsuo Handa
[off-list ref] wrote:

quoted

On 2021/03/24 20:10, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 19:10 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/24 1:13, Mimi Zohar wrote:

quoted

On Wed, 2021-03-24 at 00:14 +0900, Tetsuo Handa wrote:

quoted

On 2021/03/23 23:47, Mimi Zohar wrote:

quoted

Initially I also questioned making "integrity" an LSM.  Perhaps it's
time to reconsider.   For now, it makes sense to just fix the NULL
pointer dereferencing.

Do we think calling panic() as "fix the NULL pointer dereferencing" ?

Not supplying "integrity" as an "lsm=" option is a user error.  There
are only two options - allow or deny the caller to proceed.   If the
user is expecting the integrity subsystem to be properly working,
returning a NULL and allowing the system to boot (RFC patch version)
does not make sense.   Better to fail early.

What does the "user" mean? Those who load the vmlinux?
Only the "root" user (so called administrators)?
Any users including other than "root" user?

If the user means those who load the vmlinux, that user is explicitly asking
for disabling "integrity" for some reason. In that case, it is a bug if
booting with "integrity" disabled is impossible.

If the user means something other than those who load the vmlinux,
is there a possibility that that user (especially non "root" users) is
allowed to try to use "integrity" ? If processes other than global init
process can try to use "integrity", wouldn't it be a DoS attack vector?
Please explain in the descripotion why calling panic() does not cause
DoS attack vector.

User in this case, is anyone rebooting the system and is intentionally
changing the default values, dropping the "integrity" option on the
boot command line.

OK. Then, I expect that the system boots instead of calling panic().
That user is explicitly asking for disabling "integrity" for some reason.

That was actually my intention. The prebuilt kernel that I use for
things has all LSMs enabled, but then I needed to try some workload
with only 1 specific LSM, so I gave a different lsm= argument.

IMA/EVM is dependent on "integrity".  Was your intention to also
disable IMA and EVM?

I think, yes... or not sure. I was trying to test a bug that requires
a different major LSM and all minor LSMs are presumably irrelevant. I
dropped existing lsm= arg and added something like lsm=apparmor.

This is the legacy case that security= supports. If you specify
security=apparmor you will get all the "minor" LSMs you have compiled
in and the "major" LSM you've specified, AppArmor in this case. This
is exactly the behavior you used to get before lsm= was introduced.

quoted

If so, when disabling "integrity", don't load an
IMA policy.

I don't really know what this means. I guess it simply comes from the
image? If so, there was no easy way to avoid loading.