Обсуждение: Re: Clarification on Role Access Rights to Table Indexes
> As it stands, a superuser can prewarm an index (because she bypasses all
> privilege checks including this one), but nobody else can.
> privilege checks including this one), but nobody else can.
That's not fully true. Any role can prewarm an index if the role has the correct privileges.
postgres=# GRANT CREATE ON SCHEMA PUBLIC TO alpha;
GRANTpostgres=# SET ROLE alpha;
SET
postgres=> CREATE TABLE tab(id INT);
CREATE TABLE
postgres=> CREATE INDEX tab_idx ON tab(id);
CREATE INDEX
postgres=> SELECT pg_prewarm('tab_idx');
pg_prewarm
------------
1
(1 row)
Don't know what stopped it from prewarming the catalog table index.
Maybe it's checking who is the owner of the table. Although in code
it's just checking the ACL_SELECT [1]. I will be debugging more here
and maybe create a patch for the same.
it's just checking the ACL_SELECT [1]. I will be debugging more here
and maybe create a patch for the same.
postgres=# RESET ROLE;
RESET
postgres=# CREATE TABLE superuser_tab(id INT);
CREATE TABLE
postgres=# CREATE INDEX idx_superuser_tab ON superuser_tab(id);
CREATE INDEX
postgres=# GRANT SELECT ON superuser_tab TO alpha;
GRANT
postgres=# SET ROLE alpha;
SET
postgres=> SELECT pg_prewarm('superuser_tab');
pg_prewarm
------------
0
(1 row)
RESET
postgres=# CREATE TABLE superuser_tab(id INT);
CREATE TABLE
postgres=# CREATE INDEX idx_superuser_tab ON superuser_tab(id);
CREATE INDEX
postgres=# GRANT SELECT ON superuser_tab TO alpha;
GRANT
postgres=# SET ROLE alpha;
SET
postgres=> SELECT pg_prewarm('superuser_tab');
pg_prewarm
------------
0
(1 row)
postgres=> SELECT pg_prewarm('idx_superuser_tab');
ERROR: permission denied for index idx_superuser_tab
ERROR: permission denied for index idx_superuser_tab
postgres=> RESET ROLE;
RESET
postgres=# ALTER TABLE superuser_tab OWNER TO alpha;
ALTER TABLE
postgres=# SET ROLE alpha;
SET
postgres=> SELECT pg_prewarm('idx_superuser_tab');
pg_prewarm
------------
1
(1 row)
RESET
postgres=# ALTER TABLE superuser_tab OWNER TO alpha;
ALTER TABLE
postgres=# SET ROLE alpha;
SET
postgres=> SELECT pg_prewarm('idx_superuser_tab');
pg_prewarm
------------
1
(1 row)
But I agree we should just check the table privileges even in the case of
indexes to decide whether to prewarm or not, as indexes don't have any privileges
of their own.
> I think -hackers would be the appropriate location for that.
I am shifting this to -hackers mailing list instead of general.
[1] https://github.com/postgres/postgres/blob/master/contrib/pg_prewarm/pg_prewarm.c#L108-L110
Regards,
Ayush Vatsa
SDE AWS
indexes to decide whether to prewarm or not, as indexes don't have any privileges
of their own.
> I think -hackers would be the appropriate location for that.
I am shifting this to -hackers mailing list instead of general.
[1] https://github.com/postgres/postgres/blob/master/contrib/pg_prewarm/pg_prewarm.c#L108-L110
Regards,
Ayush Vatsa
SDE AWS
Ayush Vatsa <ayushvatsa1810@gmail.com> writes:
>> As it stands, a superuser can prewarm an index (because she bypasses all
>> privilege checks including this one), but nobody else can.
> That's not fully true. Any role can prewarm an index if the role has the
> correct privileges.
Ah, right. An index will have null pg_class.relacl, which'll be
interpreted as "owner has all rights", so it will work for the
table owner too. Likely this explains the lack of prior complaints.
It's still a poor design IMO.
regards, tom lane
On Mon, Feb 17, 2025 at 2:57 PM Tom Lane <tgl@sss.pgh.pa.us> wrote: > Ayush Vatsa <ayushvatsa1810@gmail.com> writes: > >> As it stands, a superuser can prewarm an index (because she bypasses all > >> privilege checks including this one), but nobody else can. > > > That's not fully true. Any role can prewarm an index if the role has the > > correct privileges. > > Ah, right. An index will have null pg_class.relacl, which'll be > interpreted as "owner has all rights", so it will work for the > table owner too. Likely this explains the lack of prior complaints. > It's still a poor design IMO. I'm not sure if I'd call that a "design". Sounds like I just made a mistake here. -- Robert Haas EDB: http://www.enterprisedb.com
> I'm not sure if I'd call that a "design". Sounds like I just made a
> mistake here.
> mistake here.
Thanks Robert for confirming, let me submit a patch to fix the same.
Regards
Ayush Vatsa
ADE AWS
Ayush Vatsa <ayushvatsa1810@gmail.com> writes:
> Thanks Robert for confirming, let me submit a patch to fix the same.
Well, the first thing you need is consensus on what the behavior
should be instead.
I have a very vague recollection that we concluded that SELECT
privilege was a reasonable check because if you have that you
could manually prewarm by reading the table. That would lead
to the conclusion that the minimal fix is to look at the owning
table's privileges instead of the index's own privileges.
Or we could switch to using ownership, which'd keep the code
simple but some users might complain it's too restrictive.
While I mentioned built-in roles earlier, I now think those mostly
carry more privilege than should be required here, given the analogy
to SELECT.
regards, tom lane
On Mon, Feb 17, 2025 at 3:02 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Ayush Vatsa <ayushvatsa1810@gmail.com> writes:
> Thanks Robert for confirming, let me submit a patch to fix the same.
Well, the first thing you need is consensus on what the behavior
should be instead.
I have a very vague recollection that we concluded that SELECT
privilege was a reasonable check because if you have that you
could manually prewarm by reading the table. That would lead
to the conclusion that the minimal fix is to look at the owning
table's privileges instead of the index's own privileges.
I feel like if you can blow up the cache by loading an entire table into memory with just select privilege on the table we should be ok with allowing the same person to name an index on the same table and load it into the cache too.
David J.
On Mon, Feb 17, 2025 at 5:18 PM David G. Johnston <david.g.johnston@gmail.com> wrote: >> I have a very vague recollection that we concluded that SELECT >> privilege was a reasonable check because if you have that you >> could manually prewarm by reading the table. That would lead >> to the conclusion that the minimal fix is to look at the owning >> table's privileges instead of the index's own privileges. > > I feel like if you can blow up the cache by loading an entire table into memory with just select privilege on the tablewe should be ok with allowing the same person to name an index on the same table and load it into the cache too. +1. -- Robert Haas EDB: http://www.enterprisedb.com
Robert Haas <robertmhaas@gmail.com> writes:
> On Mon, Feb 17, 2025 at 5:18 PM David G. Johnston
> <david.g.johnston@gmail.com> wrote:
>>> I have a very vague recollection that we concluded that SELECT
>>> privilege was a reasonable check because if you have that you
>>> could manually prewarm by reading the table. That would lead
>>> to the conclusion that the minimal fix is to look at the owning
>>> table's privileges instead of the index's own privileges.
>> I feel like if you can blow up the cache by loading an entire table into memory with just select privilege on the
tablewe should be ok with allowing the same person to name an index on the same table and load it into the cache too.
> +1.
Is that a +1 for the specific design of "check SELECT on the index's
table", or just a +1 for changing something here?
regards, tom lane
On Tue, Feb 18, 2025 at 11:02 AM Tom Lane <tgl@sss.pgh.pa.us> wrote: > Is that a +1 for the specific design of "check SELECT on the index's > table", or just a +1 for changing something here? That is a +1 for the specific design of "check SELECT on the index's table". I don't want to be closed-minded: if you have some strong reason for believing that's the wrong thing to do, I'm all ears. However, I'm presently of the view that it is exactly the right thing to do, to the point where I don't currently understand why there's anything to think about here. -- Robert Haas EDB: http://www.enterprisedb.com
Robert Haas <robertmhaas@gmail.com> writes:
> That is a +1 for the specific design of "check SELECT on the index's
> table". I don't want to be closed-minded: if you have some strong
> reason for believing that's the wrong thing to do, I'm all ears.
> However, I'm presently of the view that it is exactly the right thing
> to do, to the point where I don't currently understand why there's
> anything to think about here.
I have no objection to it, but I wasn't as entirely convinced
as you are that it's the only plausible answer.
One specific thing I'm slightly worried about is that a naive
implementation would probably cause this function to lock the
table after the index, risking deadlock against queries that
take the locks in the more conventional order. I don't recall
what if anything we've done about that in other places
(-ENOCAFFEINE).
regards, tom lane
On Tue, Feb 18, 2025 at 11:30 AM Tom Lane <tgl@sss.pgh.pa.us> wrote: > I have no objection to it, but I wasn't as entirely convinced > as you are that it's the only plausible answer. Hmm, OK. > One specific thing I'm slightly worried about is that a naive > implementation would probably cause this function to lock the > table after the index, risking deadlock against queries that > take the locks in the more conventional order. I don't recall > what if anything we've done about that in other places > (-ENOCAFFEINE). Yeah, that seems like a good thing to worry about from an implementation point of view but it doesn't seem like a reason to question the basic design choice. In general, if you can use a table, you also get to use its indexes, so that interpretation seems natural to me here, also. Now, if somebody finds a problem with requiring only SELECT permission, I could see changing the requirements for both tables and indexes, but I find it harder to imagine that we'd want those things to work differently from each other. Of course I'm willing to be convinced that there's a good reason for them to be different; I just can't currently imagine what it might be. -- Robert Haas EDB: http://www.enterprisedb.com
Hello Everyone,
It seems there's a general consensus that we should maintain a
original design to support pg_prewarm, with a minor adjustment:
when querying indexes, we should verify the privileges of the parent table.
I’ve attached a patch for this, which includes some test cases as well.
Let me know if it needs any changes.
Regards,
Ayush Vatsa
SDE AWS
Вложения
Added the CF entry for the same - https://commitfest.postgresql.org/patch/5583/
Regards,
Ayush Vatsa
Ayush Vatsa
SDE AWS
On Wed, Feb 19, 2025 at 03:53:48PM +0530, Ayush Vatsa wrote: > It seems there's a general consensus that we should maintain a > original design to support pg_prewarm, with a minor adjustment: > when querying indexes, we should verify the privileges of the parent table. > > I´ve attached a patch for this, which includes some test cases as well. > Let me know if it needs any changes. + tableOid = IndexGetRelation(relOid, false); + aclresult = pg_class_aclcheck(tableOid, GetUserId(), ACL_SELECT); I'm wondering whether setting missing_ok to true is correct here. IIUC we should have an AccessShareLock on the index, but I don't know if that's enough protection. The only other similar coding pattern I'm aware of is RangeVarCallbackForReindexIndex(), which sets missing_ok to false and attempts to gracefully handle a missing table. Of course, maybe that's wrong, too. Perhaps it's all close enough in practice. If we get it wrong, you might get a slightly less helpful error message when the table is concurrently dropped, which isn't so bad. -- nathan
> I'm wondering whether setting missing_ok to true is correct here. IIUC we
> should have an AccessShareLock on the index, but I don't know if that's
> enough protection.
> should have an AccessShareLock on the index, but I don't know if that's
> enough protection.
Since we are already opening the relation with rel = relation_open(relOid, AccessShareLock);,
if relOid does not exist, it will throw an error. If it does exist, we acquire an AccessShareLock,
preventing it from being dropped.
By the time we reach IndexGetRelation(), we can be confident that relOid exists and is
protected by the lock. Given this, it makes sense to keep missing_ok = false here.
Let me know if you agree or if you see any scenario where missing_ok = true would be preferable—I can update the condition accordingly.
Thanks!
Ayush Vatsa
SDE AWS
On Sat, Mar 08, 2025 at 08:34:40PM +0530, Ayush Vatsa wrote: >> I'm wondering whether setting missing_ok to true is correct here. IIUC we >> should have an AccessShareLock on the index, but I don't know if that's >> enough protection. > > Since we are already opening the relation with rel = relation_open(relOid, > AccessShareLock);, > if relOid does not exist, it will throw an error. If it does exist, we > acquire an AccessShareLock, > preventing it from being dropped. > > By the time we reach IndexGetRelation(), we can be confident that relOid > exists and is > protected by the lock. Given this, it makes sense to keep missing_ok = false > here. > > Let me know if you agree or if you see any scenario where > missing_ok = true would be preferable-I can update the condition > accordingly. Right, we will have a lock on the index, but my concern is that we won't have a lock on its table. I was specifically concerned that a concurrent DROP TABLE could cause IndexGetRelation() to fail, i.e., emit a gross "cache lookup failed" error. From a quick test and skim of the relevant code, I think your patch is fine, though. IndexGetRelation() retrieves the table OID from pg_index, so the OID should definitely be valid. And IIUC DROP TABLE first acquires a lock on the table and its dependent objects (e.g., indexes) before any actual deletions, so AFAICT there's no problem with using it in pg_class_aclcheck() and get_rel_name(), either. -- nathan
> From a quick test and skim of the relevant
> code, I think your patch is fine, though
Thanks for reviewing.
> (e.g., indexes) before any actual deletions, so AFAICT there's no problem
> with using it in pg_class_aclcheck() and get_rel_name(), either.
> code, I think your patch is fine, though
Thanks for reviewing.
> And IIUC
> DROP TABLE first acquires a lock on the table and its dependent objects> (e.g., indexes) before any actual deletions, so AFAICT there's no problem
> with using it in pg_class_aclcheck() and get_rel_name(), either.
True, I have also verified that from [1], hence I think we are safe here.
Maybe we can move ahead with the patch if we can see no other concerns.
Maybe we can move ahead with the patch if we can see no other concerns.
Thanks,
Ayush Vatsa
SDE AWS
Ayush Vatsa
SDE AWS
On Sun, Mar 09, 2025 at 03:01:41AM +0530, Ayush Vatsa wrote: > Maybe we can move ahead with the patch if we can see no other concerns. I think we should allow some time in case others want to review the patch. I do see a concern upthread about increased deadlock risk [0], but your patch doesn't lock the table, but unless I'm wrong [1] (which is always possible), it doesn't need to lock it. Anyway, here is a tidied up patch. [0] https://postgr.es/m/1246906.1739896202%40sss.pgh.pa.us [1] https://postgr.es/m/Z8yxsm9ZWVkHlPbV%40nathan -- nathan
Вложения
Nathan Bossart <nathandbossart@gmail.com> writes:
> I do see a concern upthread about increased deadlock risk [0], but your
> patch doesn't lock the table, but unless I'm wrong [1] (which is always
> possible), it doesn't need to lock it.
It bothers me a bit that this proposes to do something as complicated
as pg_class_aclcheck on a table we have no lock on. As you say, the
lock we hold on the index would prevent DROP TABLE, but that doesn't
mean we won't have any issues with other DDL on the table. Still,
taking a lock would be bad because of the deadlock hazard, and I
think the potential for conflicts with concurrent DDL is nonzero in
a lot of other places. So I don't have any concrete reason to object.
ReindexIndex() faces this same problem and solves it with some
very complex code that manages to get the table's lock first.
But I see that it's also doing pg_class_aclcheck on a table
it hasn't locked yet, so I don't think that adopting its approach
would do anything useful for us here.
regards, tom lane
On Sat, Mar 08, 2025 at 05:17:40PM -0500, Tom Lane wrote:
> It bothers me a bit that this proposes to do something as complicated
> as pg_class_aclcheck on a table we have no lock on. As you say, the
> lock we hold on the index would prevent DROP TABLE, but that doesn't
> mean we won't have any issues with other DDL on the table. Still,
> taking a lock would be bad because of the deadlock hazard, and I
> think the potential for conflicts with concurrent DDL is nonzero in
> a lot of other places. So I don't have any concrete reason to object.
>
> ReindexIndex() faces this same problem and solves it with some
> very complex code that manages to get the table's lock first.
> But I see that it's also doing pg_class_aclcheck on a table
> it hasn't locked yet, so I don't think that adopting its approach
> would do anything useful for us here.
I noticed that amcheck's bt_index_check_internal() handles this problem,
and I think that approach could be adapted here:
relkind = get_rel_relkind(relOid);
if (relkind == RELKIND_INDEX || relkind == RELKIND_PARTITIONED_INDEX)
{
permOid = IndexGetRelation(relOid, true);
if (OidIsValid(permOid))
LockRelationOid(permOid, AccessShareLock);
else
fail = true;
}
else
permOid = relOid;
rel = relation_open(relOid, AccessShareLock);
if (fail ||
(permOid != relOid && permOid != IndexGetRelation(relOid, false)))
ereport(ERROR,
(errcode(ERRCODE_UNDEFINED_TABLE),
errmsg("could not find parent table of index \"%s\"",
RelationGetRelationName(rel))));
aclresult = pg_class_aclcheck(permOid, GetUserId(), ACL_SELECT);
if (aclresult != ACLCHECK_OK)
aclcheck_error(aclresult, get_relkind_objtype(rel->rd_rel->relkind), get_rel_name(relOid));
if (permOid != relOid)
UnlockRelationOid(permOid, AccessShareLock);
stats_lock_check_privileges() does something similar, but it's not as
cautious about the "heapid != IndexGetRelation(indrelid, false)" race
condition. Maybe RangeVarCallbackForReindexIndex() should be smarter about
this, too. That being said, this is a fair amount of complexity to handle
something that is in theory extremely rare...
--
nathan
Nathan Bossart <nathandbossart@gmail.com> writes:
> On Sat, Mar 08, 2025 at 05:17:40PM -0500, Tom Lane wrote:
>> ReindexIndex() faces this same problem and solves it with some
>> very complex code that manages to get the table's lock first.
> I noticed that amcheck's bt_index_check_internal() handles this problem,
> ...
> stats_lock_check_privileges() does something similar, but it's not as
> cautious about the "heapid != IndexGetRelation(indrelid, false)" race
> condition.
Egad, we've already got three inconsistent implementations of this
functionality? I think the first step must be to unify them into
a common implementation, if at all possible.
regards, tom lane
On Sun, Mar 09, 2025 at 11:48:05AM -0400, Tom Lane wrote: > Nathan Bossart <nathandbossart@gmail.com> writes: >> On Sat, Mar 08, 2025 at 05:17:40PM -0500, Tom Lane wrote: >>> ReindexIndex() faces this same problem and solves it with some >>> very complex code that manages to get the table's lock first. > >> I noticed that amcheck's bt_index_check_internal() handles this problem, >> ... >> stats_lock_check_privileges() does something similar, but it's not as >> cautious about the "heapid != IndexGetRelation(indrelid, false)" race >> condition. > > Egad, we've already got three inconsistent implementations of this > functionality? I think the first step must be to unify them into > a common implementation, if at all possible. Agreed. I worry that trying to unify each bespoke implementation into a single function might result in an unwieldy mess, but I'll give it a shot... -- nathan
On Sun, 9 Mar 2025 at 03:27, Nathan Bossart <nathandbossart@gmail.com> wrote: > > On Sun, Mar 09, 2025 at 03:01:41AM +0530, Ayush Vatsa wrote: > > Maybe we can move ahead with the patch if we can see no other concerns. > > I think we should allow some time in case others want to review the patch. > I do see a concern upthread about increased deadlock risk [0], but your > patch doesn't lock the table, but unless I'm wrong [1] (which is always > possible), it doesn't need to lock it. > > Anyway, here is a tidied up patch. I noticed that Tom Lane's comment from [1] is not addressed. I'm changing the commitfest entry status to Waiting on Author, Please address them and update the status to Needs Review. [1] - https://www.postgresql.org/message-id/279947.1741535285%40sss.pgh.pa.us Regards, Vignesh
On Mon, Mar 10, 2025 at 10:15:19AM -0500, Nathan Bossart wrote: > On Sun, Mar 09, 2025 at 11:48:05AM -0400, Tom Lane wrote: >> Nathan Bossart <nathandbossart@gmail.com> writes: >>> On Sat, Mar 08, 2025 at 05:17:40PM -0500, Tom Lane wrote: >>>> ReindexIndex() faces this same problem and solves it with some >>>> very complex code that manages to get the table's lock first. >> >>> I noticed that amcheck's bt_index_check_internal() handles this problem, >>> ... >>> stats_lock_check_privileges() does something similar, but it's not as >>> cautious about the "heapid != IndexGetRelation(indrelid, false)" race >>> condition. >> >> Egad, we've already got three inconsistent implementations of this >> functionality? I think the first step must be to unify them into >> a common implementation, if at all possible. > > Agreed. I worry that trying to unify each bespoke implementation into a > single function might result in an unwieldy mess, but I'll give it a > shot... I tried to unify these, but each one seems to be just different enough to make it not worth the trouble. Instead, I took a look at each implementation: * amcheck's amcheck_lock_relation_and_check() seems correct to me. * stats_lock_check_privileges() appears to be missing the second IndexGetRelation() check after locking the table and index, so I added that in 0001. Since this code is new to v18, I proposed to back-patch 0001 there. * RangeVarCallbackForReindexIndex() was checking privileges on the table before locking it, so I reversed it in 0002. Interestingly, this caused test errors because LockRelationOid() checks for invalidation messages, so the pg_class_aclcheck() call started failing with unhelpful errors due to concurrently dropped relations. To deal with that, I switched to pg_class_aclcheck_ext() so that we can handle missing relations. Furthermore, I noticed that this callback seems to assume that as long as the index does not change between calls, its table won't, either. That's probably always true in practice, but even if it's completely true, I see no reason to rely on it. So, I simplified the code to unconditionally unlock any previously-locked table and to lock whatever IndexGetRelation() returns. This could probably be back-patched, but in the absence of any reports or any reproducible bugs, I don't think we should. * 0003 fixes pg_prewarm's privilege checks by following a similar pattern. This probably ought to get back-patched to all supported versions. -- nathan
Вложения
Nathan Bossart <nathandbossart@gmail.com> writes:
> * RangeVarCallbackForReindexIndex() was checking privileges on the table
> before locking it, so I reversed it in 0002.
Don't we do that intentionally, to make sure someone can't cause DOS
on a table they have no privileges on?
regards, tom lane
On Wed, Sep 24, 2025 at 12:13:34PM -0400, Tom Lane wrote: > Nathan Bossart <nathandbossart@gmail.com> writes: >> * RangeVarCallbackForReindexIndex() was checking privileges on the table >> before locking it, so I reversed it in 0002. > > Don't we do that intentionally, to make sure someone can't cause DOS > on a table they have no privileges on? Ah, right. I switched it back in v4. -- nathan
Вложения
On Wed, Sep 24, 2025 at 11:52:09AM -0500, Nathan Bossart wrote: > Ah, right. I switched it back in v4. Unless more feedback materializes, I'm planning to commit these sometime next week. -- nathan
On Wed, 2025-09-24 at 11:52 -0500, Nathan Bossart wrote:
> On Wed, Sep 24, 2025 at 12:13:34PM -0400, Tom Lane wrote:
> > Nathan Bossart <nathandbossart@gmail.com> writes:
> > > * RangeVarCallbackForReindexIndex() was checking privileges on
> > > the table
> > > before locking it, so I reversed it in 0002.
> >
> > Don't we do that intentionally, to make sure someone can't cause
> > DOS
> > on a table they have no privileges on?
>
> Ah, right. I switched it back in v4.
v4-0001 looks good to me.
Just to make sure I understand: the actual problem would only happen
with OID wraparound, right?
Regards,
Jeff Davis
On Wed, 2025-10-08 at 20:06 -0700, Jeff Davis wrote:
> On Wed, 2025-09-24 at 11:52 -0500, Nathan Bossart wrote:
> > On Wed, Sep 24, 2025 at 12:13:34PM -0400, Tom Lane wrote:
> > > Nathan Bossart <nathandbossart@gmail.com> writes:
> > > > * RangeVarCallbackForReindexIndex() was checking privileges on
> > > > the table
> > > > before locking it, so I reversed it in 0002.
> > >
> > > Don't we do that intentionally, to make sure someone can't cause
> > > DOS
> > > on a table they have no privileges on?
> >
> > Ah, right. I switched it back in v4.
>
> v4-0001 looks good to me.
Actually, now I'm unsure. v4-0001 is taking a lock on the table before
checking privileges, whereas v4-0002 is going to some effort to avoid
that. Is that because the latter is taking a ShareLock?
Regards,
Jeff Davis
On Wed, Oct 08, 2025 at 08:28:01PM -0700, Jeff Davis wrote: > Actually, now I'm unsure. v4-0001 is taking a lock on the table before > checking privileges, whereas v4-0002 is going to some effort to avoid > that. Is that because the latter is taking a ShareLock? I was confused by this, too. We seem to go to great lengths to avoid taking a lock before checking permissions in RangeVarGetRelidExtended(), but in pg_prewarm() and this stats code, we are taking the lock first. pg_prewarm() can't use RangeVarGetRelid because you give it the OID, but I'm not seeing why stat_utils.c can't use it. We should probably fix this. I wouldn't be surprised if there are other examples. -- nathan
On Thu, Oct 09, 2025 at 10:39:32AM -0500, Nathan Bossart wrote: > On Wed, Oct 08, 2025 at 08:28:01PM -0700, Jeff Davis wrote: >> Actually, now I'm unsure. v4-0001 is taking a lock on the table before >> checking privileges, whereas v4-0002 is going to some effort to avoid >> that. Is that because the latter is taking a ShareLock? > > I was confused by this, too. We seem to go to great lengths to avoid > taking a lock before checking permissions in RangeVarGetRelidExtended(), > but in pg_prewarm() and this stats code, we are taking the lock first. > pg_prewarm() can't use RangeVarGetRelid because you give it the OID, but > I'm not seeing why stat_utils.c can't use it. We should probably fix this. > I wouldn't be surprised if there are other examples. I spent some time trying to change pg_prewarm() to check permissions before locking and came up with the attached. There are certainly issues with the patch, but this at least demonstrates the complexity required. I'm tempted to say that this is more trouble than it's worth, but it does feel a little weird to leave it as-is. There's a similar pattern in get_rel_from_relname() in dblink.c, which also seems to only be used with an AccessShareLock (like pg_prewarm). My best guess from reading lots of code, commit messages, and old e-mails in the archives is that the original check-privileges-before-locking work was never completed. I'm currently leaning towards continuing with v4 of the patch set. 0001 and 0003 are a little weird in that a concurrent change could lead to a "could not find parent table" ERROR, but IIUC that is an extremely remote possibility. -- nathan
Вложения
On Thu, Oct 09, 2025 at 04:18:03PM -0500, Nathan Bossart wrote: > There's a similar pattern in get_rel_from_relname() in dblink.c, which also > seems to only be used with an AccessShareLock (like pg_prewarm). My best > guess from reading lots of code, commit messages, and old e-mails in the > archives is that the original check-privileges-before-locking work was > never completed. I added an 0004 that changes dblink to use RangeVarGetRelidExtended(). > I'm currently leaning towards continuing with v4 of the patch set. 0001 > and 0003 are a little weird in that a concurrent change could lead to a > "could not find parent table" ERROR, but IIUC that is an extremely remote > possibility. After sleeping on it, I still think this is the right call. In any case, I've spent way too much time on this stuff, so I plan to commit the attached soon. -- nathan
Вложения
On Fri, 2025-10-10 at 11:26 -0500, Nathan Bossart wrote:
> On Thu, Oct 09, 2025 at 04:18:03PM -0500, Nathan Bossart wrote:
> > There's a similar pattern in get_rel_from_relname() in dblink.c,
> > which also
> > seems to only be used with an AccessShareLock (like pg_prewarm).
> > My best
> > guess from reading lots of code, commit messages, and old e-mails
> > in the
> > archives is that the original check-privileges-before-locking work
> > was
> > never completed.
Interesting, thank you for the analysis.
> > I'm currently leaning towards continuing with v4 of the patch set.
> > 0001
> > and 0003 are a little weird in that a concurrent change could lead
> > to a
> > "could not find parent table" ERROR, but IIUC that is an extremely
> > remote
> > possibility.
>
> After sleeping on it, I still think this is the right call. In any
> case,
> I've spent way too much time on this stuff, so I plan to commit the
> attached soon.
I'm OK with that. v5-0001 is an improvement over the current situation.
Regards,
Jeff Davis
On Wed, 2025-09-24 at 12:13 -0400, Tom Lane wrote:
> Don't we do that intentionally, to make sure someone can't cause DOS
> on a table they have no privileges on?
Is this only a problem for strong locks (ShareLock or greater)?
Strong locks are a problem when you have a pattern like a long running
query that holds an AccessShareLock, and then an unprivileged user
requests an AccessExclusiveLock, forcing other queries to queue up
behind it, and the queue doesn't clear until the long running query
finishes.
But weaker locks don't seem to have that problem, right?
Regards,
Jeff Davis
On Fri, Oct 10, 2025 at 11:31:03AM -0700, Jeff Davis wrote: > On Fri, 2025-10-10 at 11:26 -0500, Nathan Bossart wrote: >> After sleeping on it, I still think this is the right call. In any >> case, I've spent way too much time on this stuff, so I plan to commit >> the attached soon. > > I'm OK with that. v5-0001 is an improvement over the current situation. Okay, I lied. I spent even more time on these patches and came up with the attached. Here's a summary of what's going on: * 0001 moves the code for stats clearing/setting to use RangeVarGetRelidExtended(). The existing code looks up the relation, locks it, and then checks permissions. There's no guarantee that the relation you looked up didn't concurrently change before locking, and locking before privilege checks is troublesome from a DOS perspective. One downside of using RangeVarGetRelidExtended() is that we can't use AccessShareLock for regular indexes, but I'm not sure that's really a problem since we're already using ShareUpdateExclusiveLock for everything else. The RangeVarGetRelidExtended() callback is similar to the one modified by 0002. This should be back-patched to v18. * 0002 fixes the RangeVarGetRelidExtended() callback for REINDEX INDEX to handle unlikely scenarios involving OID reuse (e.g., lookup returns the same index OID for a different table). I did confirm there was a bug here by concurrently re-creating an index with the same OID for a heap with a different OID (via the pg_upgrade support functions). In previous versions of this patch, I tried to fix this by unconditionally unlocking the heap at the beginning of the callback, but upon further inspection, I noticed that creates deadlock hazards because we might've already locked the index. (We need to lock the heap first.) In v6, I've just added an ERROR for these extremely unlikely scenarios. I've also replaced all early returns in this function with ERRORs (except for the invalid relId case). AFAICT the extra checks are unecessary, and even if they were necessary, I think they break some of the code related to heap locking in subtle ways. Some callbacks do these extra checks, and others do not, and AFAIK there haven't been any reported problems either way. 0002 should be back-patched to v13, but it will look a little different on v16 and newer, i.e., before MAINTAIN was added. * 0003 fixes the privilege checks in pg_prewarm by using a similar approach to amcheck_lock_relation_and_check(). This seems correct to me, but it does add more locking. This should be back-patched to v13. * 0004 is a small patch to teach dblink to use RangeVarGetRelidExtended(). I believe this code predates that function. I don't intend to back-patch this one. -- nathan
Вложения
Jeff Davis <pgsql@j-davis.com> writes:
> On Wed, 2025-09-24 at 12:13 -0400, Tom Lane wrote:
>> Don't we do that intentionally, to make sure someone can't cause DOS
>> on a table they have no privileges on?
> Is this only a problem for strong locks (ShareLock or greater)?
> Strong locks are a problem when you have a pattern like a long running
> query that holds an AccessShareLock, and then an unprivileged user
> requests an AccessExclusiveLock, forcing other queries to queue up
> behind it, and the queue doesn't clear until the long running query
> finishes.
> But weaker locks don't seem to have that problem, right?
I don't think so. Even AccessShareLock is enough to block another
session trying to acquire AccessExclusiveLock, and then not only
have you DoS'd that session, but everything else trying to access
the table will queue up behind the AccessExclusiveLock request.
So it's only not-a-problem if nothing anywhere in the system wants
non-sharable locks.
regards, tom lane
On Mon, 2025-10-13 at 17:21 -0400, Tom Lane wrote:
> I don't think so. Even AccessShareLock is enough to block another
> session trying to acquire AccessExclusiveLock, and then not only
> have you DoS'd that session, but everything else trying to access
> the table will queue up behind the AccessExclusiveLock request.
> So it's only not-a-problem if nothing anywhere in the system wants
> non-sharable locks.
I tried imagining how that could be a problem, but couldn't come up
with anything. If the privilege check is right after the lock, then
either:
(a) The malicious AccessShareLock is granted, then is quickly released
when the privilege check fails and the transaction aborts; or
(b) The malicious AccessShareLock is queued behind a legitimate
AccessExclusiveLock, in which case every other lock would be queued up
as well. As soon as the AccessExclusiveLock is released, the
AccessShareLock would be granted, but quickly released when the
privilege check fails.
For it to be a problem, the malicious lock needs to be strong enough to
conflict with a lock level weaker than itself, i.e. ShareLock or
stronger.
I'm not sure we save anything by being lazier for weaker lock levels,
so perhaps the point is irrelevant. But if I'm missing something,
please let me know.
Regards,
Jeff Davis
On Mon, 2025-10-13 at 14:30 -0500, Nathan Bossart wrote:
> * 0001 moves the code for stats clearing/setting to use
> RangeVarGetRelidExtended(). The existing code looks up the relation,
> locks
> it, and then checks permissions. There's no guarantee that the
> relation
> you looked up didn't concurrently change before locking, and locking
> before
> privilege checks is troublesome from a DOS perspective. One downside
> of
> using RangeVarGetRelidExtended() is that we can't use AccessShareLock
> for
> regular indexes, but I'm not sure that's really a problem since we're
> already using ShareUpdateExclusiveLock for everything else. The
> RangeVarGetRelidExtended() callback is similar to the one modified by
> 0002.
> This should be back-patched to v18.
We tried to match the locking behavior for analyze. Originally, that's
because we were using in-place updates, which required those specific
kinds of locks. Now that the in-place code is gone, then I think it's
OK to use ShareUpdateExclusiveLock for indexes, too, but it is a
notable difference in behavior.
Including Corey in case he has comments.
As for the patch itself, it looks good to me. Stylistically I might
have kept the "index_oid" variable, which makes some of the tests a bit
clearer, but I don't have a strong opinion.
The unlikely scenarios are a bit confusing. I'd probably error for
either case. Also, the error message on the second scenario is wrong if
the previous lookup was a table, I think.
> * 0002 fixes the RangeVarGetRelidExtended() callback for REINDEX
> INDEX to
> handle unlikely scenarios involving OID reuse (e.g., lookup returns
> the
> same index OID for a different table). I did confirm there was a bug
> here
> by concurrently re-creating an index with the same OID for a heap
> with a
> different OID (via the pg_upgrade support functions). In previous
> versions
> of this patch, I tried to fix this by unconditionally unlocking the
> heap at
> the beginning of the callback, but upon further inspection, I noticed
> that
> creates deadlock hazards because we might've already locked the
> index. (We
> need to lock the heap first.) In v6, I've just added an ERROR for
> these
> extremely unlikely scenarios. I've also replaced all early returns
> in this
> function with ERRORs (except for the invalid relId case).
+1 for throwing errors when we have race conditions combined with name
reuse. Looks fine to me.
>
> * 0003 fixes the privilege checks in pg_prewarm by using a similar
> approach
> to amcheck_lock_relation_and_check(). This seems correct to me, but
> it
> does add more locking. This should be back-patched to v13.
IIUC this is locking before the privilege check. Is there a reason why
we think this is OK here (and in amcheck_lock_relation_and_check()) but
not for the stats?
> * 0004 is a small patch to teach dblink to use
> RangeVarGetRelidExtended().
> I believe this code predates that function. I don't intend to back-
> patch
> this one.
Looks good.
Regards,
Jeff Davis
Thanks for reviewing. On Mon, Oct 13, 2025 at 07:23:36PM -0700, Jeff Davis wrote: > The unlikely scenarios are a bit confusing. I'd probably error for > either case. Also, the error message on the second scenario is wrong if > the previous lookup was a table, I think. Yeah, I think that's a better idea. > IIUC this is locking before the privilege check. Is there a reason why > we think this is OK here (and in amcheck_lock_relation_and_check()) but > not for the stats? For amcheck, AFAICT there aren't actually any ACL checks within the code because the function is restricted to superuser by default. For pg_prewarm, I don't know. You do have to install the extension before using it, but once installed, it's available to everyone by default. My guess is that it just hasn't been a problem in the field. Regardless, fixing the lock-before-privilege-checks behavior doesn't strike me as a bug, so I think we ought to proceed with something like 0003 for back-patching purposes and then to rework it further for v19. Does that sound okay to you? >> * 0004 is a small patch to teach dblink to use >> RangeVarGetRelidExtended(). I believe this code predates that >> function. I don't intend to back-patch this one. > > Looks good. I'm going to go commit this one now to get it out of the way. -- nathan
Jeff Davis <pgsql@j-davis.com> writes:
> On Mon, 2025-10-13 at 17:21 -0400, Tom Lane wrote:
>> I don't think so. Even AccessShareLock is enough to block another
>> session trying to acquire AccessExclusiveLock, and then not only
>> have you DoS'd that session, but everything else trying to access
>> the table will queue up behind the AccessExclusiveLock request.
>> So it's only not-a-problem if nothing anywhere in the system wants
>> non-sharable locks.
> I tried imagining how that could be a problem, but couldn't come up
> with anything. If the privilege check is right after the lock, then
> either:
> (a) The malicious AccessShareLock is granted, then is quickly released
> when the privilege check fails and the transaction aborts; or
> (b) The malicious AccessShareLock is queued behind a legitimate
> AccessExclusiveLock, in which case every other lock would be queued up
> as well. As soon as the AccessExclusiveLock is released, the
> AccessShareLock would be granted, but quickly released when the
> privilege check fails.
> For it to be a problem, the malicious lock needs to be strong enough to
> conflict with a lock level weaker than itself, i.e. ShareLock or
> stronger.
Robert might remember better, but I think the assumption behind
the current design of RangeVarGetRelidExtended is that it's not
okay to take a lock in the first place if you don't have the
privilege to do so. Your analysis here supposes that it's okay
to take a lock without privileges so long as you can't block someone
else for very long, where "very long" is not tightly defined but
hopefully isn't controllable by the malicious user. So that's
moving the goalposts somewhat, but you might get people to sign
onto it with more careful analysis of the worst-case delay.
(The thing I'd worry about is whether it's possible to block
execution of the privilege check, or even just make it slow.)
Given that definition, I think you're right that it's possible to
identify cases where lock-then-check can't cause meaningful DoS.
RangeVarGetRelidExtended has to cope with the general case, so
that's not an argument for simplifying it, but we might not need
equivalent complexity everywhere.
regards, tom lane
On Tue, 2025-10-14 at 11:05 -0500, Nathan Bossart wrote:
> For
> pg_prewarm, I don't know. You do have to install the extension
> before
> using it, but once installed, it's available to everyone by default.
> My
> guess is that it just hasn't been a problem in the field.
If we start with an OID, what's the right way to do these kinds of
checks? Could we do an ACL check, then lock it, then do an ACL check
again to catch OID wraparound?
Last-minute suggestions on 0003:
* Add a comment around the privOid check to explain that, if the
object is an index, we must check the privileges on the table instead.
* Clarify in the comment that the race against index drop/recreation
involves OID wraparound.
+1 to the patch and backpatch.
As a separate thought, I'm wondering if we should do more to enforce
the idea that we check the privileges and owner of an index's table,
and never the index itself. That's for another discussion, though.
> Regardless, fixing the lock-before-privilege-checks behavior doesn't
> strike
> me as a bug, so I think we ought to proceed with something like 0003
> for
> back-patching purposes and then to rework it further for v19. Does
> that
> sound okay to you?
According to the current rules[1], it does seem to technically be a
bug, but as far as I can tell, not one of much consequence.
Regards,
Jeff Davis
[1]
https://www.postgresql.org/message-id/976405.1760459426@sss.pgh.pa.us
I've committed 0004. Here is an updated patch set. -- nathan
Вложения
On Tue, Oct 14, 2025 at 10:01:37AM -0700, Jeff Davis wrote: > If we start with an OID, what's the right way to do these kinds of > checks? Could we do an ACL check, then lock it, then do an ACL check > again to catch OID wraparound? I tried something like this upthread [0]. My feeling was that this was a lot of complexity for not a lot of gain. Perhaps it's still worth doing, though. [0] https://postgr.es/m/aOgmi6avE6qMw_6t%40nathan -- nathan
I just pushed 0001, and longfin and sifaka have very quickly reminded me that we don't require C11 on v18. Will fix shortly... -- nathan
I've committed everything except for 0003, which I probably won't get to until Friday. Note that I decided against back-patching 0002 because of the presumed rarity of and lack of reports for the bug. -- nathan
On Wed, Oct 15, 2025 at 04:38:05PM -0500, Nathan Bossart wrote: > I've committed everything except for 0003, which I probably won't get to > until Friday. Note that I decided against back-patching 0002 because of > the presumed rarity of and lack of reports for the bug. Everything tracked in this thread is now committed. -- nathan