security_inode_init_security() inode field requirements

security_inode_init_security() inode field requirements

[Steven Whitehouse]

Hi,

I'm wondering whether there is a list somewhere of fields which
security_inode_init_security() requires are set in an inode when it is
called? In particular, does it matter if the inode number itself is
unset when security_inode_init_security() is called?

The problem that I'm looking at is the use of multiple transactions
during inode creation when some combination of ACLs/LSMs/IMA are turned
on. What we have currently (in GFS2, there are other fs which follow
broadly the same solution though) is this:

1. Create inode in core
2. Create inode on disk
3. Add xattrs one at a time for ACLs/LSMs/IMA
4. Link inode into directory

Steps 2 through 4 require separate transactions at the moment. What I'd
like to do is to be able to get the details of the xattrs ahead of time
such that the allocation of the inode can be one and the same allocation
as that for the xattr blocks. That allows merging of the transactions
into one and a greatly simplified error path too. This would look
something like:

1. Create in-core inode and init required fields
2. Get details of all xattrs for the inode
3. Alloc on disk inode and blocks for xattrs as needed
4. Link into directory

In this case, steps 2 through 4 are within a single transaction. We
don't actually need to have the content of the xattrs ahead of
allocating the inode, just the length (or even just a max length, if it
is not too large). However the easiest solution would just be to move
the call to security_inode_init_security() to the point before we've
allocated the inode (and thus we don't know its number yet) but after
we've filled out all the remaining fields if that makes sense?

So I just wanted to check whether that would break anything,

Steve.

Re: security_inode_init_security() inode field requirements

[Mimi Zohar]

On Fri, 2013-03-01 at 10:12 +0000, Steven Whitehouse wrote:
> Hi,
> 
> I'm wondering whether there is a list somewhere of fields which
> security_inode_init_security() requires are set in an inode when it is
> called? In particular, does it matter if the inode number itself is
> unset when security_inode_init_security() is called?
> 
> The problem that I'm looking at is the use of multiple transactions
> during inode creation when some combination of ACLs/LSMs/IMA are turned
> on. What we have currently (in GFS2, there are other fs which follow
> broadly the same solution though) is this:
> 
> 1. Create inode in core
> 2. Create inode on disk
> 3. Add xattrs one at a time for ACLs/LSMs/IMA
> 4. Link inode into directory
> 
> Steps 2 through 4 require separate transactions at the moment. What I'd
> like to do is to be able to get the details of the xattrs ahead of time
> such that the allocation of the inode can be one and the same allocation
> as that for the xattr blocks. That allows merging of the transactions
> into one and a greatly simplified error path too. This would look
> something like:
> 
> 1. Create in-core inode and init required fields
> 2. Get details of all xattrs for the inode
> 3. Alloc on disk inode and blocks for xattrs as needed
> 4. Link into directory
> 
> In this case, steps 2 through 4 are within a single transaction. We
> don't actually need to have the content of the xattrs ahead of
> allocating the inode, just the length (or even just a max length, if it
> is not too large). However the easiest solution would just be to move
> the call to security_inode_init_security() to the point before we've
> allocated the inode (and thus we don't know its number yet) but after
> we've filled out all the remaining fields if that makes sense?
> 
> So I just wanted to check whether that would break anything,

Hi Steve,

Included in security_inode_init_security() is the call to
evm_inode_init_security() to write the 'security.evm' extended
attribute.  'security.evm' is an HMAC of the security extended
attributes and other file metadata, including the inode.  For an exact
list of other metadata included in the HMAC calculation refer to
hmac_add_misc().  (The UUID is being added to the HMAC calculation in
this open window.)

thanks,

Mimi

Re: security_inode_init_security() inode field requirements

[Steven Whitehouse]

Hi,

On Fri, 2013-03-01 at 07:27 -0500, Mimi Zohar wrote:
> On Fri, 2013-03-01 at 10:12 +0000, Steven Whitehouse wrote:
> > Hi,
> > 
> > I'm wondering whether there is a list somewhere of fields which
> > security_inode_init_security() requires are set in an inode when it is
> > called? In particular, does it matter if the inode number itself is
> > unset when security_inode_init_security() is called?
> > 
> > The problem that I'm looking at is the use of multiple transactions
> > during inode creation when some combination of ACLs/LSMs/IMA are turned
> > on. What we have currently (in GFS2, there are other fs which follow
> > broadly the same solution though) is this:
> > 
> > 1. Create inode in core
> > 2. Create inode on disk
> > 3. Add xattrs one at a time for ACLs/LSMs/IMA
> > 4. Link inode into directory
> > 
> > Steps 2 through 4 require separate transactions at the moment. What I'd
> > like to do is to be able to get the details of the xattrs ahead of time
> > such that the allocation of the inode can be one and the same allocation
> > as that for the xattr blocks. That allows merging of the transactions
> > into one and a greatly simplified error path too. This would look
> > something like:
> > 
> > 1. Create in-core inode and init required fields
> > 2. Get details of all xattrs for the inode
> > 3. Alloc on disk inode and blocks for xattrs as needed
> > 4. Link into directory
> > 
> > In this case, steps 2 through 4 are within a single transaction. We
> > don't actually need to have the content of the xattrs ahead of
> > allocating the inode, just the length (or even just a max length, if it
> > is not too large). However the easiest solution would just be to move
> > the call to security_inode_init_security() to the point before we've
> > allocated the inode (and thus we don't know its number yet) but after
> > we've filled out all the remaining fields if that makes sense?
> > 
> > So I just wanted to check whether that would break anything,
> 
> Hi Steve,
> 
> Included in security_inode_init_security() is the call to
> evm_inode_init_security() to write the 'security.evm' extended
> attribute.  'security.evm' is an HMAC of the security extended
> attributes and other file metadata, including the inode.  For an exact
> list of other metadata included in the HMAC calculation refer to
> hmac_add_misc().  (The UUID is being added to the HMAC calculation in
> this open window.)
> 
> thanks,
> 
> Mimi
> 

Ok... but it is using inode->i_ino in that case, as well as the
generation number too. So that presumably can only be done after the
inode has been allocated, since we need to know its location in order to
know its inode number. Also the generation number is assigned at
allocation time. One potential issue though... the inode->i_ino is an
unsigned long, so that on 32 bit archs, that will be (on GFS2) a
truncated version of the full 64 bit inode number. I'm not sure if that
matters or not for EVM.

So since that appears to rule out doing the security init bits ahead of
the allocation of the inode, is it possible to get a maximum value for
the size of the xattr which EVM will add? Likewise for the other LSMs
too?

Steve.

Re: security_inode_init_security() inode field requirements

[Mimi Zohar]

On Fri, 2013-03-01 at 13:11 +0000, Steven Whitehouse wrote:
> Hi,
> 
> On Fri, 2013-03-01 at 07:27 -0500, Mimi Zohar wrote:
> > On Fri, 2013-03-01 at 10:12 +0000, Steven Whitehouse wrote:
> > > Hi,
> > > 
> > > I'm wondering whether there is a list somewhere of fields which
> > > security_inode_init_security() requires are set in an inode when it is
> > > called? In particular, does it matter if the inode number itself is
> > > unset when security_inode_init_security() is called?
> > > 
> > > The problem that I'm looking at is the use of multiple transactions
> > > during inode creation when some combination of ACLs/LSMs/IMA are turned
> > > on. What we have currently (in GFS2, there are other fs which follow
> > > broadly the same solution though) is this:
> > > 
> > > 1. Create inode in core
> > > 2. Create inode on disk
> > > 3. Add xattrs one at a time for ACLs/LSMs/IMA
> > > 4. Link inode into directory
> > > 
> > > Steps 2 through 4 require separate transactions at the moment. What I'd
> > > like to do is to be able to get the details of the xattrs ahead of time
> > > such that the allocation of the inode can be one and the same allocation
> > > as that for the xattr blocks. That allows merging of the transactions
> > > into one and a greatly simplified error path too. This would look
> > > something like:
> > > 
> > > 1. Create in-core inode and init required fields
> > > 2. Get details of all xattrs for the inode
> > > 3. Alloc on disk inode and blocks for xattrs as needed
> > > 4. Link into directory
> > > 
> > > In this case, steps 2 through 4 are within a single transaction. We
> > > don't actually need to have the content of the xattrs ahead of
> > > allocating the inode, just the length (or even just a max length, if it
> > > is not too large). However the easiest solution would just be to move
> > > the call to security_inode_init_security() to the point before we've
> > > allocated the inode (and thus we don't know its number yet) but after
> > > we've filled out all the remaining fields if that makes sense?
> > > 
> > > So I just wanted to check whether that would break anything,
> > 
> > Hi Steve,
> > 
> > Included in security_inode_init_security() is the call to
> > evm_inode_init_security() to write the 'security.evm' extended
> > attribute.  'security.evm' is an HMAC of the security extended
> > attributes and other file metadata, including the inode.  For an exact
> > list of other metadata included in the HMAC calculation refer to
> > hmac_add_misc().  (The UUID is being added to the HMAC calculation in
> > this open window.)
> > 
> > thanks,
> > 
> > Mimi
> > 
> 
> Ok... but it is using inode->i_ino in that case, as well as the
> generation number too. So that presumably can only be done after the
> inode has been allocated, since we need to know its location in order to
> know its inode number. Also the generation number is assigned at
> allocation time. One potential issue though... the inode->i_ino is an
> unsigned long, so that on 32 bit archs, that will be (on GFS2) a
> truncated version of the full 64 bit inode number. I'm not sure if that
> matters or not for EVM.

The reason for adding additional information to the HMAC calculation is
to prevent a cut & paste attack, taking a valid 'security.evm' from one
file and using it for another file.

> So since that appears to rule out doing the security init bits ahead of
> the allocation of the inode, is it possible to get a maximum value for
> the size of the xattr which EVM will add? Likewise for the other LSMs
> too?

Although 'security.evm' can contain either an HMAC or a digital
signature, new inodes are created with an HMAC, which is hardcoded as
hmac(sha1).  On a running system, the existing 'security.evm' digital
signatures are replaced with an HMAC.

SELinux defines INITCONTEXTLEN as 255, which seems to be the maximum
xattr length, but I'm not sure.  From the smack header:

/*
 * Smack labels were limited to 23 characters for a long time.
 */
#define SMK_LABELLEN    24
#define SMK_LONGLABEL   256

thanks,

Mimi

Re: security_inode_init_security() inode field requirements

[Eric Paris]

SELinux has no maximum   :-(

Realistically there are a couple of interfaces that limit things to
4k, but labels on files on disk could be even larger than that!

255 will fit most every label, but not necessarily all of them.


I know ext4 on Fedora allocates inodes which left about 255 bytes for
selinux.selinux, but will place the xattr in another block if it
happens to be larger than 255.  This is rare, but certainly
possible....

We use the inode->i_mode.

In debug/error path we use:
  inode->i_sb inode->i_no

We could use the parent dir sb instead of the new inode->i_sb.  We
don't have to print the i_no when we hit a failure, but it is just
about the only information that can help for debugging/figuring out
which file had a failure..

-Eric

On Fri, Mar 1, 2013 at 9:07 AM, Mimi Zohar <zohar@linux.vnet.ibm.com> wrote:
> On Fri, 2013-03-01 at 13:11 +0000, Steven Whitehouse wrote:
>> Hi,
>>
>> On Fri, 2013-03-01 at 07:27 -0500, Mimi Zohar wrote:
>> > On Fri, 2013-03-01 at 10:12 +0000, Steven Whitehouse wrote:
>> > > Hi,
>> > >
>> > > I'm wondering whether there is a list somewhere of fields which
>> > > security_inode_init_security() requires are set in an inode when it is
>> > > called? In particular, does it matter if the inode number itself is
>> > > unset when security_inode_init_security() is called?
>> > >
>> > > The problem that I'm looking at is the use of multiple transactions
>> > > during inode creation when some combination of ACLs/LSMs/IMA are turned
>> > > on. What we have currently (in GFS2, there are other fs which follow
>> > > broadly the same solution though) is this:
>> > >
>> > > 1. Create inode in core
>> > > 2. Create inode on disk
>> > > 3. Add xattrs one at a time for ACLs/LSMs/IMA
>> > > 4. Link inode into directory
>> > >
>> > > Steps 2 through 4 require separate transactions at the moment. What I'd
>> > > like to do is to be able to get the details of the xattrs ahead of time
>> > > such that the allocation of the inode can be one and the same allocation
>> > > as that for the xattr blocks. That allows merging of the transactions
>> > > into one and a greatly simplified error path too. This would look
>> > > something like:
>> > >
>> > > 1. Create in-core inode and init required fields
>> > > 2. Get details of all xattrs for the inode
>> > > 3. Alloc on disk inode and blocks for xattrs as needed
>> > > 4. Link into directory
>> > >
>> > > In this case, steps 2 through 4 are within a single transaction. We
>> > > don't actually need to have the content of the xattrs ahead of
>> > > allocating the inode, just the length (or even just a max length, if it
>> > > is not too large). However the easiest solution would just be to move
>> > > the call to security_inode_init_security() to the point before we've
>> > > allocated the inode (and thus we don't know its number yet) but after
>> > > we've filled out all the remaining fields if that makes sense?
>> > >
>> > > So I just wanted to check whether that would break anything,
>> >
>> > Hi Steve,
>> >
>> > Included in security_inode_init_security() is the call to
>> > evm_inode_init_security() to write the 'security.evm' extended
>> > attribute.  'security.evm' is an HMAC of the security extended
>> > attributes and other file metadata, including the inode.  For an exact
>> > list of other metadata included in the HMAC calculation refer to
>> > hmac_add_misc().  (The UUID is being added to the HMAC calculation in
>> > this open window.)
>> >
>> > thanks,
>> >
>> > Mimi
>> >
>>
>> Ok... but it is using inode->i_ino in that case, as well as the
>> generation number too. So that presumably can only be done after the
>> inode has been allocated, since we need to know its location in order to
>> know its inode number. Also the generation number is assigned at
>> allocation time. One potential issue though... the inode->i_ino is an
>> unsigned long, so that on 32 bit archs, that will be (on GFS2) a
>> truncated version of the full 64 bit inode number. I'm not sure if that
>> matters or not for EVM.
>
> The reason for adding additional information to the HMAC calculation is
> to prevent a cut & paste attack, taking a valid 'security.evm' from one
> file and using it for another file.
>
>> So since that appears to rule out doing the security init bits ahead of
>> the allocation of the inode, is it possible to get a maximum value for
>> the size of the xattr which EVM will add? Likewise for the other LSMs
>> too?
>
> Although 'security.evm' can contain either an HMAC or a digital
> signature, new inodes are created with an HMAC, which is hardcoded as
> hmac(sha1).  On a running system, the existing 'security.evm' digital
> signatures are replaced with an HMAC.
>
> SELinux defines INITCONTEXTLEN as 255, which seems to be the maximum
> xattr length, but I'm not sure.  From the smack header:
>
> /*
>  * Smack labels were limited to 23 characters for a long time.
>  */
> #define SMK_LABELLEN    24
> #define SMK_LONGLABEL   256
>
> thanks,
>
> Mimi
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: security_inode_init_security() inode field requirements

[Steven Whitehouse]

Hi,

On Fri, 2013-03-01 at 10:13 -0500, Eric Paris wrote:
> SELinux has no maximum   :-(
> 
> Realistically there are a couple of interfaces that limit things to
> 4k, but labels on files on disk could be even larger than that!
> 
> 255 will fit most every label, but not necessarily all of them.
> 
> 
> I know ext4 on Fedora allocates inodes which left about 255 bytes for
> selinux.selinux, but will place the xattr in another block if it
> happens to be larger than 255.  This is rare, but certainly
> possible....
> 
> We use the inode->i_mode.
> 
> In debug/error path we use:
>   inode->i_sb inode->i_no
> 
> We could use the parent dir sb instead of the new inode->i_sb.  We
> don't have to print the i_no when we hit a failure, but it is just
> about the only information that can help for debugging/figuring out
> which file had a failure..
> 
> -Eric
> 
So it sounds like setting the selinux label before the allocation of the
inode wouldn't be too much of a problem. That would give us the size
ahead of time. Maybe EVM is the only thing which needs to be an
exception in terms of being done after the inode number is set, and if
that has a fairly small maximum size, then that could still work I
think.

Having said that, this is turning out to be a fair bit more complicated
than I'd hoped :(

Steve.

Re: security_inode_init_security() inode field requirements

[Casey Schaufler]

On 3/1/2013 2:12 AM, Steven Whitehouse wrote:
> Hi,
>
> I'm wondering whether there is a list somewhere of fields which
> security_inode_init_security() requires are set in an inode when it is
> called? In particular, does it matter if the inode number itself is
> unset when security_inode_init_security() is called?
>
> The problem that I'm looking at is the use of multiple transactions
> during inode creation when some combination of ACLs/LSMs/IMA are turned
> on. What we have currently (in GFS2, there are other fs which follow
> broadly the same solution though) is this:
>
> 1. Create inode in core
> 2. Create inode on disk
> 3. Add xattrs one at a time for ACLs/LSMs/IMA
> 4. Link inode into directory
>
> Steps 2 through 4 require separate transactions at the moment. What I'd
> like to do is to be able to get the details of the xattrs ahead of time
> such that the allocation of the inode can be one and the same allocation
> as that for the xattr blocks. That allows merging of the transactions
> into one and a greatly simplified error path too. This would look
> something like:
>
> 1. Create in-core inode and init required fields
> 2. Get details of all xattrs for the inode

Although none of the existing LSMs do so, you have to allow for
the case where an LSM uses multiple labels or lists of labels,
or makes decisions based on the order in which inodes are processed.
As Mimi points out, the details of the attributes may depend on
the state of initialization.

> 3. Alloc on disk inode and blocks for xattrs as needed
> 4. Link into directory
>
> In this case, steps 2 through 4 are within a single transaction. We
> don't actually need to have the content of the xattrs ahead of
> allocating the inode, just the length (or even just a max length, if it
> is not too large). However the easiest solution would just be to move
> the call to security_inode_init_security() to the point before we've
> allocated the inode (and thus we don't know its number yet) but after
> we've filled out all the remaining fields if that makes sense?
>
> So I just wanted to check whether that would break anything,
>
> Steve.
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>