[v4,2/5] security: Rewrite security_old_inode_init_security()

Message ID 20221110094639.3086409-3-roberto.sassu@huaweicloud.com
State New
Headers
Series evm: Prepare for moving to the LSM infrastructure |

Commit Message

Roberto Sassu Nov. 10, 2022, 9:46 a.m. UTC
  From: Roberto Sassu <roberto.sassu@huawei.com>

Rewrite security_old_inode_init_security() to call
security_inode_init_security() before making changes to support multiple
LSMs providing xattrs. Do it so that the required changes are done only in
one place.

Define the security_initxattrs() callback and pass it to
security_inode_init_security() as argument, to obtain the first xattr
provided by LSMs.

This behavior is a bit different from the current one. Before this patch
calling call_int_hook() could cause multiple LSMs to provide an xattr,
since call_int_hook() does not stop when an LSM returns zero. The caller of
security_old_inode_init_security() receives the last xattr set. The pointer
of the xattr value of previous LSMs is lost, causing memory leaks.

However, in practice, this scenario does not happen as the only in-tree
LSMs providing an xattr at inode creation time are SELinux and Smack, which
are mutually exclusive.

Signed-off-by: Roberto Sassu <roberto.sassu@huawei.com>
---
 security/security.c | 58 +++++++++++++++++++++++++++++++++++++--------
 1 file changed, 48 insertions(+), 10 deletions(-)
  

Comments

Mimi Zohar Nov. 17, 2022, 1:03 p.m. UTC | #1
Hi Roberto,

On Thu, 2022-11-10 at 10:46 +0100, Roberto Sassu wrote:
> From: Roberto Sassu <roberto.sassu@huawei.com>
> 
> Rewrite security_old_inode_init_security() to call
> security_inode_init_security() before making changes to support multiple
> LSMs providing xattrs. Do it so that the required changes are done only in
> one place.

Only security_inode_init_security() has support for EVM.   Making
security_old_inode_init_security() a wrapper for
security_inode_init_security() could result in security.evm extended
attributes being created that previously weren't created.

In fact ocfs2 defines ocfs2_init_security_get() as a wrapper for both
the old and new inode_init_security calls based on the caller's
preference.   Only mknod and symlink seem to use the old function. 
Wondering why do they differentiate between callers?  (Cc'ing the ocfs2
mailing list as they're affected by this change.)

"[PATCH v4 1/5] reiserfs: Add missing calls to
reiserfs_security_free()"  fixed a memory leak.  I couldn't tell if
there was a similar memory leak in ocfs2, the only other user of
security_old_inode_init_security().

As ocfs2 already defines initxattrs, that leaves only reiserfs missing
initxattrs().  A better, cleaner solution would be to define one.

thanks,

Mimi

> 
> Define the security_initxattrs() callback and pass it to
> security_inode_init_security() as argument, to obtain the first xattr
> provided by LSMs.
> 
> This behavior is a bit different from the current one. Before this patch
> calling call_int_hook() could cause multiple LSMs to provide an xattr,
> since call_int_hook() does not stop when an LSM returns zero. The caller of
> security_old_inode_init_security() receives the last xattr set. The pointer
> of the xattr value of previous LSMs is lost, causing memory leaks.
> 
> However, in practice, this scenario does not happen as the only in-tree
> LSMs providing an xattr at inode creation time are SELinux and Smack, which
> are mutually exclusive.
> 
> Signed-off-by: Roberto Sassu <roberto.sassu@huawei.com>b
  
Roberto Sassu Nov. 18, 2022, 9:04 a.m. UTC | #2
On 11/17/2022 2:03 PM, Mimi Zohar wrote:
> Hi Roberto,
> 
> On Thu, 2022-11-10 at 10:46 +0100, Roberto Sassu wrote:
>> From: Roberto Sassu <roberto.sassu@huawei.com>
>>
>> Rewrite security_old_inode_init_security() to call
>> security_inode_init_security() before making changes to support multiple
>> LSMs providing xattrs. Do it so that the required changes are done only in
>> one place.
> 
> Only security_inode_init_security() has support for EVM.   Making
> security_old_inode_init_security() a wrapper for
> security_inode_init_security() could result in security.evm extended
> attributes being created that previously weren't created.

Hi Mimi

yes, I thought about this problem. In fact, it should not matter too 
much. Since security_old_inode_init_security() supports setting only one 
xattr: if there is an LSM xattr, that one will be set, and the EVM one 
will be discarded; if there is no LSM xattr, EVM would not add one.

> In fact ocfs2 defines ocfs2_init_security_get() as a wrapper for both
> the old and new inode_init_security calls based on the caller's
> preference.   Only mknod and symlink seem to use the old function.
> Wondering why do they differentiate between callers?  (Cc'ing the ocfs2
> mailing list as they're affected by this change.)
> 
> "[PATCH v4 1/5] reiserfs: Add missing calls to
> reiserfs_security_free()"  fixed a memory leak.  I couldn't tell if
> there was a similar memory leak in ocfs2, the only other user of
> security_old_inode_init_security().

Will look into it.

> As ocfs2 already defines initxattrs, that leaves only reiserfs missing
> initxattrs().  A better, cleaner solution would be to define one.

Yes, great idea!

Thanks

Roberto

> thanks,
> 
> Mimi
> 
>>
>> Define the security_initxattrs() callback and pass it to
>> security_inode_init_security() as argument, to obtain the first xattr
>> provided by LSMs.
>>
>> This behavior is a bit different from the current one. Before this patch
>> calling call_int_hook() could cause multiple LSMs to provide an xattr,
>> since call_int_hook() does not stop when an LSM returns zero. The caller of
>> security_old_inode_init_security() receives the last xattr set. The pointer
>> of the xattr value of previous LSMs is lost, causing memory leaks.
>>
>> However, in practice, this scenario does not happen as the only in-tree
>> LSMs providing an xattr at inode creation time are SELinux and Smack, which
>> are mutually exclusive.
>>
>> Signed-off-by: Roberto Sassu <roberto.sassu@huawei.com>b
  
Roberto Sassu Nov. 21, 2022, 9:45 a.m. UTC | #3
On Thu, 2022-11-17 at 08:03 -0500, Mimi Zohar wrote:
> Hi Roberto,
> 
> On Thu, 2022-11-10 at 10:46 +0100, Roberto Sassu wrote:
> > From: Roberto Sassu <roberto.sassu@huawei.com>
> > 
> > Rewrite security_old_inode_init_security() to call
> > security_inode_init_security() before making changes to support multiple
> > LSMs providing xattrs. Do it so that the required changes are done only in
> > one place.
> 
> Only security_inode_init_security() has support for EVM.   Making
> security_old_inode_init_security() a wrapper for
> security_inode_init_security() could result in security.evm extended
> attributes being created that previously weren't created.
> 
> In fact ocfs2 defines ocfs2_init_security_get() as a wrapper for both
> the old and new inode_init_security calls based on the caller's
> preference.   Only mknod and symlink seem to use the old function. 
> Wondering why do they differentiate between callers?  (Cc'ing the ocfs2
> mailing list as they're affected by this change.)
> 
> "[PATCH v4 1/5] reiserfs: Add missing calls to
> reiserfs_security_free()"  fixed a memory leak.  I couldn't tell if
> there was a similar memory leak in ocfs2, the only other user of
> security_old_inode_init_security().

From what I see, there is no memory leak there.

> As ocfs2 already defines initxattrs, that leaves only reiserfs missing
> initxattrs().  A better, cleaner solution would be to define one.

If I understood why security_old_inode_init_security() is called
instead of security_inode_init_security(), the reason seems that the
filesystem code uses the length of the obtained xattr to make some
calculations (e.g. reserve space). The xattr is written at a later
time.

Since for reiserfs there is a plan to deprecate it, it probably
wouldn't be worth to support the creation of multiple xattrs. I would
define a callback to take the first xattr and make a copy, so that
calling security_inode_init_security() + reiserfs_initxattrs() is
equivalent to calling security_old_inode_init_security().

But then, this is what anyway I was doing with the
security_initxattrs() callback, for all callers of security_old_inode_i
nit_security().

Also, security_old_inode_init_security() is exported to kernel modules.
Maybe, it is used somewhere. So, unless we plan to remove it
completely, it should be probably be fixed to avoid multiple LSMs
successfully setting an xattr, and losing the memory of all except the
last (which this patch fixes by calling security_inode_init_security())
.

If there is still the preference, I will implement the reiserfs
callback and make a fix for security_old_inode_init_security().

Thanks

Roberto

> thanks,
> 
> Mimi
> 
> > Define the security_initxattrs() callback and pass it to
> > security_inode_init_security() as argument, to obtain the first xattr
> > provided by LSMs.
> > 
> > This behavior is a bit different from the current one. Before this patch
> > calling call_int_hook() could cause multiple LSMs to provide an xattr,
> > since call_int_hook() does not stop when an LSM returns zero. The caller of
> > security_old_inode_init_security() receives the last xattr set. The pointer
> > of the xattr value of previous LSMs is lost, causing memory leaks.
> > 
> > However, in practice, this scenario does not happen as the only in-tree
> > LSMs providing an xattr at inode creation time are SELinux and Smack, which
> > are mutually exclusive.
> > 
> > Signed-off-by: Roberto Sassu <roberto.sassu@huawei.com>b
  
Mimi Zohar Nov. 21, 2022, 8:54 p.m. UTC | #4
On Mon, 2022-11-21 at 10:45 +0100, Roberto Sassu wrote:
> > As ocfs2 already defines initxattrs, that leaves only reiserfs missing
> > initxattrs().  A better, cleaner solution would be to define one.
> 
> If I understood why security_old_inode_init_security() is called
> instead of security_inode_init_security(), the reason seems that the
> filesystem code uses the length of the obtained xattr to make some
> calculations (e.g. reserve space). The xattr is written at a later
> time.
> 
> Since for reiserfs there is a plan to deprecate it, it probably
> wouldn't be worth to support the creation of multiple xattrs. I would
> define a callback to take the first xattr and make a copy, so that
> calling security_inode_init_security() + reiserfs_initxattrs() is
> equivalent to calling security_old_inode_init_security().
> 
> But then, this is what anyway I was doing with the
> security_initxattrs() callback, for all callers of security_old_inode_i
> nit_security().
> 
> Also, security_old_inode_init_security() is exported to kernel modules.
> Maybe, it is used somewhere. So, unless we plan to remove it
> completely, it should be probably be fixed to avoid multiple LSMs
> successfully setting an xattr, and losing the memory of all except the
> last (which this patch fixes by calling security_inode_init_security())
> .
> 
> If there is still the preference, I will implement the reiserfs
> callback and make a fix for security_old_inode_init_security().

There's no sense in doing both, as the purpose of defining a reiserfs
initxattrs function was to clean up this code making it more readable.

Mimi
  
Paul Moore Nov. 21, 2022, 11:55 p.m. UTC | #5
On Mon, Nov 21, 2022 at 3:54 PM Mimi Zohar <zohar@linux.ibm.com> wrote:
>
> On Mon, 2022-11-21 at 10:45 +0100, Roberto Sassu wrote:
> > > As ocfs2 already defines initxattrs, that leaves only reiserfs missing
> > > initxattrs().  A better, cleaner solution would be to define one.
> >
> > If I understood why security_old_inode_init_security() is called
> > instead of security_inode_init_security(), the reason seems that the
> > filesystem code uses the length of the obtained xattr to make some
> > calculations (e.g. reserve space). The xattr is written at a later
> > time.
> >
> > Since for reiserfs there is a plan to deprecate it, it probably
> > wouldn't be worth to support the creation of multiple xattrs. I would
> > define a callback to take the first xattr and make a copy, so that
> > calling security_inode_init_security() + reiserfs_initxattrs() is
> > equivalent to calling security_old_inode_init_security().

FWIW, reiserfs isn't going to be removed until 2025, I'm hopeful we
can remove the IMA/EVM special cases before then :)

> > But then, this is what anyway I was doing with the
> > security_initxattrs() callback, for all callers of security_old_inode_i
> > nit_security().
> >
> > Also, security_old_inode_init_security() is exported to kernel modules.
> > Maybe, it is used somewhere. So, unless we plan to remove it
> > completely, it should be probably be fixed to avoid multiple LSMs
> > successfully setting an xattr, and losing the memory of all except the
> > last (which this patch fixes by calling security_inode_init_security()).

I would much rather remove security_old_inode_init_security() then
worry about what out-of-tree modules might be using it.  Hopefully we
can resolve the ocfs2 usage and get ocfs2 exclusively on the new hook
without too much trouble, which means all we have left is reiserfs ...
how difficult would you expect the conversion to be for reiserfs?

> > If there is still the preference, I will implement the reiserfs
> > callback and make a fix for security_old_inode_init_security().
>
> There's no sense in doing both, as the purpose of defining a reiserfs
> initxattrs function was to clean up this code making it more readable.
>
> Mimi
  
Roberto Sassu Nov. 22, 2022, 8:29 a.m. UTC | #6
On Mon, 2022-11-21 at 18:55 -0500, Paul Moore wrote:
> On Mon, Nov 21, 2022 at 3:54 PM Mimi Zohar <zohar@linux.ibm.com> wrote:
> > On Mon, 2022-11-21 at 10:45 +0100, Roberto Sassu wrote:
> > > > As ocfs2 already defines initxattrs, that leaves only reiserfs missing
> > > > initxattrs().  A better, cleaner solution would be to define one.
> > > 
> > > If I understood why security_old_inode_init_security() is called
> > > instead of security_inode_init_security(), the reason seems that the
> > > filesystem code uses the length of the obtained xattr to make some
> > > calculations (e.g. reserve space). The xattr is written at a later
> > > time.
> > > 
> > > Since for reiserfs there is a plan to deprecate it, it probably
> > > wouldn't be worth to support the creation of multiple xattrs. I would
> > > define a callback to take the first xattr and make a copy, so that
> > > calling security_inode_init_security() + reiserfs_initxattrs() is
> > > equivalent to calling security_old_inode_init_security().
> 
> FWIW, reiserfs isn't going to be removed until 2025, I'm hopeful we
> can remove the IMA/EVM special cases before then :)

Well, we are not that far...

> > > But then, this is what anyway I was doing with the
> > > security_initxattrs() callback, for all callers of security_old_inode_i
> > > nit_security().
> > > 
> > > Also, security_old_inode_init_security() is exported to kernel modules.
> > > Maybe, it is used somewhere. So, unless we plan to remove it
> > > completely, it should be probably be fixed to avoid multiple LSMs
> > > successfully setting an xattr, and losing the memory of all except the
> > > last (which this patch fixes by calling security_inode_init_security()).
> 
> I would much rather remove security_old_inode_init_security() then
> worry about what out-of-tree modules might be using it.  Hopefully we
> can resolve the ocfs2 usage and get ocfs2 exclusively on the new hook
> without too much trouble, which means all we have left is reiserfs ...
> how difficult would you expect the conversion to be for reiserfs?

Ok for removing security_old_inode_init_security().

For reiserfs, I guess maintaining the current behavior of setting only
one xattr should be easy. For multiple xattrs, I need to understand
exactly how many blocks need to be reserved.

> > > If there is still the preference, I will implement the reiserfs
> > > callback and make a fix for security_old_inode_init_security().
> > 
> > There's no sense in doing both, as the purpose of defining a reiserfs
> > initxattrs function was to clean up this code making it more readable.

The fix for security_old_inode_init_security(), stopping at the first
LSM returning zero, was to avoid the memory leak. Will not be needed if
the function is removed.

Roberto
  

Patch

diff --git a/security/security.c b/security/security.c
index 79d82cb6e469..a0e9b4ce2341 100644
--- a/security/security.c
+++ b/security/security.c
@@ -1089,20 +1089,34 @@  int security_dentry_create_files_as(struct dentry *dentry, int mode,
 }
 EXPORT_SYMBOL(security_dentry_create_files_as);
 
+static int security_initxattrs(struct inode *inode, const struct xattr *xattrs,
+			       void *fs_info)
+{
+	struct xattr *dest = (struct xattr *)fs_info;
+
+	dest->name = xattrs->name;
+	dest->value = xattrs->value;
+	dest->value_len = xattrs->value_len;
+	return 0;
+}
+
 int security_inode_init_security(struct inode *inode, struct inode *dir,
 				 const struct qstr *qstr,
 				 const initxattrs initxattrs, void *fs_data)
 {
 	struct xattr new_xattrs[MAX_LSM_EVM_XATTR + 1];
 	struct xattr *lsm_xattr, *evm_xattr, *xattr;
-	int ret;
+	int ret = -EOPNOTSUPP;
 
 	if (unlikely(IS_PRIVATE(inode)))
-		return 0;
+		goto out_exit;
 
-	if (!initxattrs)
-		return call_int_hook(inode_init_security, -EOPNOTSUPP, inode,
-				     dir, qstr, NULL, NULL, NULL);
+	if (!initxattrs ||
+	    (initxattrs == &security_initxattrs && !fs_data)) {
+		ret = call_int_hook(inode_init_security, -EOPNOTSUPP, inode,
+				    dir, qstr, NULL, NULL, NULL);
+		goto out_exit;
+	}
 	memset(new_xattrs, 0, sizeof(new_xattrs));
 	lsm_xattr = new_xattrs;
 	ret = call_int_hook(inode_init_security, -EOPNOTSUPP, inode, dir, qstr,
@@ -1118,8 +1132,19 @@  int security_inode_init_security(struct inode *inode, struct inode *dir,
 		goto out;
 	ret = initxattrs(inode, new_xattrs, fs_data);
 out:
-	for (xattr = new_xattrs; xattr->value != NULL; xattr++)
+	for (xattr = new_xattrs; xattr->value != NULL; xattr++) {
+		/*
+		 * Xattr value freed by the caller of
+		 * security_old_inode_init_security().
+		 */
+		if (xattr == new_xattrs && initxattrs == &security_initxattrs &&
+		    !ret)
+			continue;
 		kfree(xattr->value);
+	}
+out_exit:
+	if (initxattrs == &security_initxattrs)
+		return ret;
 	return (ret == -EOPNOTSUPP) ? 0 : ret;
 }
 EXPORT_SYMBOL(security_inode_init_security);
@@ -1136,10 +1161,23 @@  int security_old_inode_init_security(struct inode *inode, struct inode *dir,
 				     const struct qstr *qstr, const char **name,
 				     void **value, size_t *len)
 {
-	if (unlikely(IS_PRIVATE(inode)))
-		return -EOPNOTSUPP;
-	return call_int_hook(inode_init_security, -EOPNOTSUPP, inode, dir,
-			     qstr, name, value, len);
+	struct xattr xattr = {};
+	struct xattr *lsm_xattr = (value) ? &xattr : NULL;
+	int ret;
+
+	ret = security_inode_init_security(inode, dir, qstr,
+					   security_initxattrs, lsm_xattr);
+	if (ret)
+		return ret;
+
+	if (name)
+		*name = lsm_xattr->name;
+	if (value)
+		*value = lsm_xattr->value;
+	if (len)
+		*len = lsm_xattr->value_len;
+
+	return 0;
 }
 EXPORT_SYMBOL(security_old_inode_init_security);