Обсуждение: Re: [PATCHES] Users/Groups -> Roles
* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> Stephen Frost <sfrost@snowman.net> writes:
> > Attached please find files and patches associated with moving from the
> > User/Group system currently in place to Roles, as discussed
> > previously.
>
> I have cleaned this up a bit and committed it. I normally wouldn't
> commit an incomplete patch, but this change is blocking Alvaro's work
> on dependencies for shared objects, so I felt it was best to get the
> catalog changes in now. That will let Alvaro work on dependencies
> while I sort out the unfinished bits of roles, which I intend to do
> over the next day or so.
Great, glad to hear it. I hope you got a chance to look over the open
items in the 'milestones' file. I'd really like to see the grammar be
fixed to match SQL spec for GRANT ROLE/REVOKE ROLE. I think an approach
to take there might be to try and get GrantRoleStmt and GrantStmt to use
the same productions at the end of the line if possible or something
along those lines.
Also, I've been looking through the diff between my tree and what you
committed to CVS and had a couple comments (just my 2c: I think it would
have been alot easier using SVN to see exaclty what was different from
my patch vs. other changes since my last CVS up):
First, sorry about the gratuitous name changes, it helped me find every place I needed to look at the code and think
aboutif it needed to be changed in some way (ie: Int32GetDatum -> ObjectIdGetDatum, etc). I had planned on changing
someof them back to minimize the patch but kind of ran out of time.
Second, looks like I missed fixing an owner check in pg_proc.c Current CVS has, line 269: if (GetUserId() !=
oldproc->proowner&& !superuser()) Which is not a sufficient owner check. This should by fixed by doing a proper
pg_proc_ownercheck,ie: if (!pg_proc_ownercheck(HeapTupleGetOid(oldtup), GetUserId()))
Third, I feel it's incorrect to only allow superuser() to change ownership of objects under a role-based system.
Usersmust be able to create objects owned by a role they're in (as opposed to owned only by themselves). Without this
thereis no way for a given role to allow other roles to perform owner-level actions on objects which they create. The
pointof adding roles was to allow owner-level actions on objects to more than a single user or the superuser.
Requiringthe superuser to get involved with every table creation defeats much of the point. This should really be
possibleeither by explicitly changing the ownership of an object using ALTER ... OWNER, or by a SET ROLE followed by
CREATETABLE, etc. SET ROLE is defined by the SQL specification, though we don't support it specifically yet
(shouldn'tbe too difficult to add now though). Certainly if we accept that SET ROLE should be supported and that
objectsthen created should be owned by the role set in SET ROLE we should be willing to support non-superusers doing
ALTER... OWNER given that they could effectively do the same thing via SET ROLE (though with much more difficulty,
whichhas no appreciable gain).
Fourth, not that I use it, but, it looks like my changes to src/interfaces/ecpg/preproc/preproc.y were lost. Not sure
ifthat was intentional or not (I wouldn't think so... I do wish ecpg could just be the differences necessary for ecpg
andbe based off the main parser somehow, but that'd be a rather large change). Oh, and in that same boat,
src/tools/pgindent/pgindentalso appears to not have gotten the changes that I made.
> Many thanks for your work on this!
Happy to have helped though frustrated that you seem to have removed the part that I was originally looking for. I
don'tfeel that's justification for having it (I feel I've addressed that above) but it certainly would have been nice
tobe aware of that earlier and perhaps to have discussed the issues around it a bit more before being so close to the
featurefreeze (I know, alot my fault, but there it is).
Thanks,
Stephen
Stephen Frost <sfrost@snowman.net> writes:
> Also, I've been looking through the diff between my tree and what you
> committed to CVS and had a couple comments
> First, sorry about the gratuitous name changes, it helped me find
> every place I needed to look at the code and think about if it needed
> to be changed in some way (ie: Int32GetDatum -> ObjectIdGetDatum,
> etc). I had planned on changing some of them back to minimize the
> patch but kind of ran out of time.
No problem, I figured that was why you'd done it, but changing them back
helped me to understand the patch also ;-)
> Second, looks like I missed fixing an owner check in pg_proc.c
Got it. I was wondering if there were more --- might be worth checking
all the superuser() calls.
> Third, I feel it's incorrect to only allow superuser() to change
> ownership of objects under a role-based system.
I took that out because it struck me as a likely security hole; we don't
allow non-superuser users to give away objects now, and we shouldn't
allow non-superuser roles to do so either. Moreover the tests you had
were inconsistent (not same test everyplace).
> Users must be able to
> create objects owned by a role they're in (as opposed to owned only
> by themselves).
This is what SET SESSION AUTHORIZATION/SET ROLE is for, no? You set the
auth to a role you are allowed to be in, then create the object. I do
notice that we don't have this yet, but it's surely a required piece of
the puzzle.
> Fourth, not that I use it, but, it looks like my changes to
> src/interfaces/ecpg/preproc/preproc.y were lost. Not sure if that was
> intentional or not
Yeah, it was. I leave it to Michael Meskes to sync ecpg with the main
parser; on the occasions where I've tried to do it for him, things
didn't work out well.
> I do wish ecpg could just
> be the differences necessary for ecpg and be based off the main parser
> somehow,
Me too, but I haven't seen a way yet.
> src/tools/pgindent/pgindent also appears to not have gotten the
> changes that I made.
That's an automatically generated list; there's no need to edit it.
regards, tom lane
Stephen Frost <sfrost@snowman.net> writes:
> The code I had for this was:
> if (!pg_class_ownercheck(tuple,GetUserId()) ||
> !is_role_member(newowner,GetUserId()))
> That needs a check for superuser though because while the test will pass
> on the 'pg_class_ownercheck' side, it won't on the 'is_role_member' side
Um, right, that was another problem I had with it --- at one point the
regression tests were failing because the superuser wasn't allowed to
reassign object ownership ...
I'm still fairly concerned about the security implications of letting
ordinary users reassign object ownership. The fact that SET ROLE would
let you *create* an object with ownership X is a long way away from
saying that you should be allowed to change an *existing* object to have
ownership X. This is particularly so if you are a member of a couple of
different roles with different memberships: you will be able to cause
objects to become effectively owned by certain other people, or make
them stop being effectively owned by those people. I don't have a clear
trouble case in mind at the moment, but this sure sounds like the stuff
of routine security-hole reports. (Altering the ownership of a SECURITY
DEFINER function, in particular, sounds like a great path for a cracker
to pursue.)
> One place I recall seeing one and not being sure if it should be a new
> *_ownercheck() function or not was in the 2PC patch- twophase.c, line
> 380:
This one I think we can leave...
regards, tom lane
* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> Stephen Frost <sfrost@snowman.net> writes:
> > Second, looks like I missed fixing an owner check in pg_proc.c
>
> Got it. I was wondering if there were more --- might be worth checking
> all the superuser() calls.
Yeah, let's come up with a decision about what exactly we should allow
and then perhaps I can go through all of the superuser() calls and see
what needs fixing up.
> > Third, I feel it's incorrect to only allow superuser() to change
> > ownership of objects under a role-based system.
>
> I took that out because it struck me as a likely security hole; we don't
> allow non-superuser users to give away objects now, and we shouldn't
> allow non-superuser roles to do so either. Moreover the tests you had
> were inconsistent (not same test everyplace).
Sorry about them being inconsistent, I didn't intend for them to be.
I went through a couple of iterations of them trying to do the check the
'right' way. Thinking back on it, even the checks I ended up with were
wrong (in the superuser case), though I think they were closer.
Basically my thought was to allow the same thing you could do w/ SET
ROLE, etc:
If you are the owner of the object to be changed (following the normal
owner checking rules) AND would still be considered the owner of the
object *after* the change, then you can change the ownership.
The code I had for this was:
if (!pg_class_ownercheck(tuple,GetUserId()) || !is_role_member(newowner,GetUserId()))
That needs a check for superuser though because while the test will pass
on the 'pg_class_ownercheck' side, it won't on the 'is_role_member' side
(currently anyway, I suppose a superuser could be considered to be in
any role, so we could change is_role_member to always return true for
superusers, that'd probably make pg_group look ugly though, either way):
if (!superuser() && !(pg_class_ownercheck(tupe,GetUserId()) &&
is_role_member(newowner,GetUserId())))
I think that's the correct check and can be done the same way for pretty
much all of the objects. Were there other security concerns you had?
I'd be happy to look through the superuser() checks in commands/ and
develop a patch following what I described above, as well as looking for
other cases where we should be using the *_ownercheck() functions.
One place I recall seeing one and not being sure if it should be a new
*_ownercheck() function or not was in the 2PC patch- twophase.c, line
380:
if (user != gxact->owner && !superuser_arg(user))
Wasn't sure if that made sense to have *_ownercheck, or, even if we
added one for it, if it made sense to check is_member_of_role() for
prepared transactions. I don't think SQL has anything to say about it,
anyone know what other DBs do here?
> > Users must be able to
> > create objects owned by a role they're in (as opposed to owned only
> > by themselves).
>
> This is what SET SESSION AUTHORIZATION/SET ROLE is for, no? You set the
> auth to a role you are allowed to be in, then create the object. I do
> notice that we don't have this yet, but it's surely a required piece of
> the puzzle.
(Technically I think SET SESSION AUTHORIZATION is different from SET
ROLE, but anyway)
Right, that's another way to do it (as I mentioned), and that lets you
do ownership changes, but they're much more painful:
CONNECT AS joe;
CREATE TABLE abc as SELECT name,avg(a),sum(b) FROM reallybigtable;
-- Whoops, I meant for abc to be owned by role C so sally can add her
-- column to it later, or vacuum/analyze it, whatever
GRANT SELECT ON abc TO C; -- Might not be necessary
ALTER TABLE abc RENAME TO abc_temp;
SET ROLE C;
CREATE TABLE abc AS SELECT * FROM abc_temp; -- Could be big :(
SET ROLE NONE; -- Might be just 'SET ROLE;'? Gotta check the spec
DROP TABLE abc_temp;
I don't really see the point in making users go through all of these
hoops to do an ownership change. In the end, it's the same result near
as I can tell...
> Yeah, it was. I leave it to Michael Meskes to sync ecpg with the main
> parser; on the occasions where I've tried to do it for him, things
> didn't work out well.
Ah, ok.
> > src/tools/pgindent/pgindent also appears to not have gotten the
> > changes that I made.
>
> That's an automatically generated list; there's no need to edit it.
Hah, silly me.
Thanks,
Stephen
* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> Stephen Frost <sfrost@snowman.net> writes:
> > That needs a check for superuser though because while the test will pass
> > on the 'pg_class_ownercheck' side, it won't on the 'is_role_member' side
>
> Um, right, that was another problem I had with it --- at one point the
> regression tests were failing because the superuser wasn't allowed to
> reassign object ownership ...
Yeah, sorry about that.
> I'm still fairly concerned about the security implications of letting
> ordinary users reassign object ownership. The fact that SET ROLE would
> let you *create* an object with ownership X is a long way away from
> saying that you should be allowed to change an *existing* object to have
> ownership X. This is particularly so if you are a member of a couple of
> different roles with different memberships: you will be able to cause
> objects to become effectively owned by certain other people, or make
> them stop being effectively owned by those people. I don't have a clear
> trouble case in mind at the moment, but this sure sounds like the stuff
> of routine security-hole reports. (Altering the ownership of a SECURITY
> DEFINER function, in particular, sounds like a great path for a cracker
> to pursue.)
SET ROLE also lets you *drop* an object owned by that role. Or alter
it. Or CREATE OR REPLACE FUNCTION ...
I can understand your concern. The specific use case I'm thinking about
is where a user creates an object, does some work on it, and then wants
to change its ownership to be owned by a role which that user is in. I
find myself doing that a fair bit (as superuser atm). One thing I don't
like about limiting it to that is that you then can't go back without
the whole drop/create business or getting an admin.
This also isn't stuff that couldn't be done through other means, even in
the SECURITY DEFINER function case, you just need to drop, set role,
create. Having a role with members be able to own objects isn't meant
to replace the privileges system and I don't expect people to try to use
it to.
I can perhaps see a special case for SECURITY DEFINER functions but if
we're going to special case them I'd think we'd need to make them only
be creatable/modifiable at all by superusers or add another flag to the
role to allow that.
Thanks,
Stephen
On Tue, Jun 28, 2005 at 14:45:06 -0400, Stephen Frost <sfrost@snowman.net> wrote: > > If you are the owner of the object to be changed (following the normal > owner checking rules) AND would still be considered the owner of the > object *after* the change, then you can change the ownership. That still isn't a good idea, because the new owner may not have had access to create the object you just gave to them. Or you may not have had access to drop the object you just gave away. That is going to be a security hole.
On Tue, Jun 28, 2005 at 14:52:07 -0500, Bruno Wolff III <bruno@wolff.to> wrote: > On Tue, Jun 28, 2005 at 14:45:06 -0400, > Stephen Frost <sfrost@snowman.net> wrote: > > > > If you are the owner of the object to be changed (following the normal > > owner checking rules) AND would still be considered the owner of the > > object *after* the change, then you can change the ownership. > > That still isn't a good idea, because the new owner may not have had > access to create the object you just gave to them. Or you may not have > had access to drop the object you just gave away. That is going to > be a security hole. Thinking about it some more, drops wouldn't be an issue since the owner can always drop objects. Creating objects in particular schemas or databases is not something that all roles may be able to do.
* Bruno Wolff III (bruno@wolff.to) wrote:
> On Tue, Jun 28, 2005 at 14:45:06 -0400,
> Stephen Frost <sfrost@snowman.net> wrote:
> >
> > If you are the owner of the object to be changed (following the normal
> > owner checking rules) AND would still be considered the owner of the
> > object *after* the change, then you can change the ownership.
>
> That still isn't a good idea, because the new owner may not have had
> access to create the object you just gave to them. Or you may not have
> had access to drop the object you just gave away. That is going to
> be a security hole.
If you're considered the owner of an object then you have access to drop
it already. You have to be a member of the role to which you're
changing the ownership. That role not having permission to create the
object in place is an interesting question. That's an issue for SET
ROLE too, to some extent I think, do you still have your role's
permissions after you've SET ROLE to another role? If not then you'd
have to grant CREATE on the schema to the role in order to create
objects owned by that role, and I don't think that's necessairly
something you'd want to do.
Thanks,
Stephen
Stephen Frost wrote: > I can perhaps see a special case for SECURITY DEFINER functions but if > we're going to special case them I'd think we'd need to make them only > be creatable/modifiable at all by superusers or add another flag to the > role to allow that. I agree that owner changes of SECURITY DEFINER functions seem dangerous. I would follow Stephen's idea that SECURITY DEFINER functions should only be creatable/modifiable by superusers. This would be similar to unix, where setting the suid/sgid bits is usually only allowed to root. Best Regards, Michael Paesold
Stephen Frost wrote: > If you're considered the owner of an object then you have access to drop > it already. You have to be a member of the role to which you're > changing the ownership. That role not having permission to create the > object in place is an interesting question. That's an issue for SET > ROLE too, to some extent I think, do you still have your role's > permissions after you've SET ROLE to another role? For me this would be the "natural" way how SET ROLE would behave. This is unix'ism again, but using setuid to become another user, you loose the privileges of the old user context. Therefore SET ROLE should not inherit privileges from the other role. This seems to be the safes approach. Nevertheless, what does the standard say? > If not then you'd > have to grant CREATE on the schema to the role in order to create > objects owned by that role, and I don't think that's necessairly > something you'd want to do. Right, that's an issue. But since the new role will be the *owner* of the object, it *should* really have create-privileges in that schema. So the above way seems to be correct anyway. Best Regards, Michael Paesold
* Bruno Wolff III (bruno@wolff.to) wrote:
> Thinking about it some more, drops wouldn't be an issue since the owner
> can always drop objects.
Right.
> Creating objects in particular schemas or databases is not something that
> all roles may be able to do.
Yeah, I'm not entirely sure what I think about this issue. If you're
not allowed to change ownership of objects and SET ROLE drops your
regular ROLE's privileges then the role which owns the object originally
(and which you're required to be in) must have had create access to that
schema at some point.
I can see requiring the role that's changing the ownership to have
create access to the schema in which the object that's being changed is
in.
Thanks,
Stephen
* Michael Paesold (mpaesold@gmx.at) wrote:
> Stephen Frost wrote:
> >If you're considered the owner of an object then you have access to drop
> >it already. You have to be a member of the role to which you're
> >changing the ownership. That role not having permission to create the
> >object in place is an interesting question. That's an issue for SET
> >ROLE too, to some extent I think, do you still have your role's
> >permissions after you've SET ROLE to another role?
>
> For me this would be the "natural" way how SET ROLE would behave. This is
> unix'ism again, but using setuid to become another user, you loose the
> privileges of the old user context.
> Therefore SET ROLE should not inherit privileges from the other role. This
> seems to be the safes approach.
>
> Nevertheless, what does the standard say?
Hmm, it says there's a stack and that the thing on top is what's
currently used, so it sounds like it would drop the privs too, but imv
it's not entirely clear.
> >If not then you'd
> >have to grant CREATE on the schema to the role in order to create
> >objects owned by that role, and I don't think that's necessairly
> >something you'd want to do.
>
> Right, that's an issue. But since the new role will be the *owner* of the
> object, it *should* really have create-privileges in that schema. So the
> above way seems to be correct anyway.
I'm not entirely sure that you'd necessairly want the role to have
create privileges on the schema even when it owns things in the schema
but the more I think about it that doesn't seem all that unreasonable
either. I don't think it'd be very difficult to add such a check to the
ALTER OWNER code too though.
In general, and perhaps as a unix'ism to some extent, I don't
particularly like having to su to people. To get all the other
permissions which the role has you don't have to 'su' currently, and
personally I like that and think that's correct for a role-based
environment (unlike unix where you have users and groups).
Thanks,
Stephen
Stephen Frost <sfrost@snowman.net> writes:
> * Bruno Wolff III (bruno@wolff.to) wrote:
>> Creating objects in particular schemas or databases is not something that
>> all roles may be able to do.
> Yeah, I'm not entirely sure what I think about this issue.
We have a precedent, which is that RENAME checks for create rights.
If you want to lean on the argument that this is just a shortcut for
dropping the object and then recreating it somewhere else, then you
need (a) the right to drop the object --- which is inherent in being
the old owner, and (b) the right to create the new object, which means
that (b1) you can become the role you wish to have owning the object,
and (b2) *as that role* you would have the rights needed to create the
object.
Stephen's original analysis covers (a) and (b1) but not (b2). With (b2)
I'd agree that it's just a useful shortcut.
I don't see a need to treat SECURITY DEFINER functions as
superuser-only. We've had that facility since 7.3 or so and no one
has complained that it's too dangerous.
regards, tom lane
* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> Stephen Frost <sfrost@snowman.net> writes:
> > * Bruno Wolff III (bruno@wolff.to) wrote:
> >> Creating objects in particular schemas or databases is not something that
> >> all roles may be able to do.
>
> > Yeah, I'm not entirely sure what I think about this issue.
>
> We have a precedent, which is that RENAME checks for create rights.
Ah, ok. Precedent is good.
> If you want to lean on the argument that this is just a shortcut for
> dropping the object and then recreating it somewhere else, then you
> need (a) the right to drop the object --- which is inherent in being
> the old owner, and (b) the right to create the new object, which means
> that (b1) you can become the role you wish to have owning the object,
> and (b2) *as that role* you would have the rights needed to create the
> object.
>
> Stephen's original analysis covers (a) and (b1) but not (b2). With (b2)
> I'd agree that it's just a useful shortcut.
Right. Ok, I'll develop a patch which covers (a), (b1) and (b2). I'll
also go through all of the superuser() calls in src/backend/commands/
and check for other places we may need *_ownercheck calls.
I expect to have the patch done either tonight or tommorow.
Thanks,
Stephen
I notice that AddRoleMems/DelRoleMems assume that ADMIN OPTION is not
inherited indirectly; that is it must be granted directly to you.
This seems wrong; SQL99 has under <privileges>
19) B has the WITH ADMIN OPTION on a role if a role authorization descriptor identifies the role as
grantedto B WITH ADMIN OPTION or a role authorization descriptor identifies it as granted WITH
ADMINOPTION to another applicable role for B.
and in the Access Rules for <grant role statement>
1) Every role identified by <role granted> shall be contained in the applicable roles for A and the
correspondingrole authorization descriptors shall specify WITH ADMIN OPTION.
I can't see any support in the spec for the idea that WITH ADMIN OPTION
doesn't flow through role memberships in the same way as ordinary
membership; can you quote someplace that implies this?
regards, tom lane
* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> I notice that AddRoleMems/DelRoleMems assume that ADMIN OPTION is not
> inherited indirectly; that is it must be granted directly to you.
> This seems wrong; SQL99 has under <privileges>
>
> 19) B has the WITH ADMIN OPTION on a role if a role authorization
> descriptor identifies the role as granted to B WITH ADMIN OPTION
> or a role authorization descriptor identifies it as granted WITH
> ADMIN OPTION to another applicable role for B.
>
> and in the Access Rules for <grant role statement>
>
> 1) Every role identified by <role granted> shall be contained
> in the applicable roles for A and the corresponding role
> authorization descriptors shall specify WITH ADMIN OPTION.
>
> I can't see any support in the spec for the idea that WITH ADMIN OPTION
> doesn't flow through role memberships in the same way as ordinary
> membership; can you quote someplace that implies this?
Hrm, no, sorry, I just interpreted the 'Access Rules' line for <grant
role statement> differently. That is to say:
1) Every role identified by <role granted> shall be contained (Alright, all the roles which you're granting, right)
in the applicable roles for A and the corresponding role (A must be in all the roles which are being granted)
authorization descriptors shall specify WITH ADMIN OPTION. (the grants to A for those rules specify ADMIN
OPTION)
This came across to me as meaning "there must exist an authorization
descriptor such that the granted-role equals <role granted>, the grantee
is A and WITH ADMIN OPTION is set". That could only be true if the
grant was done explicitly. Reading from 19 above (which I don't recall
seeing before, or at least not reading very carefully) I think you're
right. Either way is fine with me.
Thanks,
Stephen