Re: [PATCH] SE-PgSQL/tiny rev.2193 - Mailing list pgsql-hackers

Robert Haas wrote:
> 2009/7/16 KaiGai Kohei <kaigai@ak.jp.nec.com>:
>> Yes, the tiny version will not give any advantages in security without
>> future enhancements.
>> It is not difficult to add object classes and permissions.
>> If necessary, I'll add checks them with corresponding permissions.
>>
>> One anxiety is PostgreSQL specific object class, such as LANGUAGE.
>> It's not clear for me whether the maintainer of the SELinux security
>> policy accept these kind of object classes, or not.
>> I would like to implement them except for PostgreSQL specific object
>> class in this phase.
> 
> I'm starting to think that there's just no hope of this matching up
> well enough with the way PostgreSQL already works to have a chance of
> being accepted.

I believe different security mechanisms can have different viewpoints,
security models, criteria to make its access control decision and so on.
Some of SELinux part can match with PostgreSQL's permission with one-to-one,
such as db_database:{connect} and ACL_CONNECT.
But, we don't need to mind different security mechanism has different stuffs.

For example, Linux applies its DAC permission checks at inode_permission()
which also calls selinux_inode_permission(). It is the one-to-one mapping
part. But SELinux (LSM in actually) also put its security hooks in other
points, such as selinux_inode_rename() to check file renaming permissions,
selinux_socket_accept() to check accept connections permissions although
vanilla OS does not have corresponding permissions.

What I would like to say is that it is quite natural different security
mechanism has not-identical security models and so on.

>> Here is a few differences in access control model between PostgreSQL and
>> SELinux, so I could not map all the SELinux permissions on the pg_xxx_aclcheck()
>> mechanism.
>>
>> For example, ExecCheckRTEPerms() checks permissions on the tables and
>> columns appeared in the user given query. When the user have SELECT
>> permission on the required table, it bypasses to check permissions on
>> the columns.
>> SELinux's security model needs to check permissions on all the required
>> objects. For example, "SELECT A,B FROM T" requires the client to have
>> db_table:{select} on T and db_column:{select} on A and B.
> 
> Isn't this a purely arbitrary decision on your part to implement
> incompatible semantics?  I don't see why it can't check for
> db_table:{select} and if that fails then check for db_column:[select}
> on each column?  Maybe that's not legit, I don't understand SE-Linux
> well enough to know.  But I think we need to get someone from the
> SE-Linux community involved to help review and consider these kinds of
> issues, because it is obvious that we don't have the expertise in the
> PostgreSQL community.

Indeed, operating system does not have well analogy, such as a relationship
between tables and columns. However, SELinux's rule is simple.

Please imagine a situation when we write a security policy module.
When we would like to control accesses a certain object (such as column),
all we need to focus on is users' privileges on the column only.
If checks on columns are bypassed when he is allowed to access to the
table, we need to focus on both of the table and the column.
It seems to me it goes against to the principle in SELinux.
(Note that I never say database ACL model is bad. Both models have
their philosophies, and I believe both are worth each other.)

If you need any more comments from other persons in SELinux community,
I try to call Joshua Brindle who had been invoked in pgsql-hackers
several months ago.

>> For other example, some of pg_xxx_aclcheck() is bypassed when the client
>> has superuser privilege. In this case, SELinux requires the client to
>> have both of db_database:{superuser} and a certain permission.
> 
> Surely you can't just transform (A OR B) into (A AND B) and pretend
> that's the same thing...

It needs to change code path but it is not reasonable when user
disables SE-PgSQL. So, I thought separated security hooks are
reasonable, because it is replaced by empty macro if disabled.

>> Sorry, I could not read it from the previous suggestions.
>> If you have been suggesting it repeatedly, I'm sorry so much.
> 
> I think the language barrier is part of what is making this a very
> difficult process.  Your English is surely better than my Japanese,
> but we are definitely going around in circles.
> 
> ...Robert
> 

Thanks,
-- 
OSS Platform Development Division, NEC
KaiGai Kohei <kaigai@ak.jp.nec.com>