Обсуждение: BUG #17549: wrong index scan plan with RLS

The following bug has been logged on the website:

Bug reference:      17549
Logged by:          Zhao Rui
Email address:      875941708@qq.com
PostgreSQL version: 14.4
Operating system:   All
Description:

You can reproduce in this way:

create table abc (a integer, b text);
insert into abc select (random()*(10^4))::integer, (random()*(10^4))::text
from generate_series(1,100000);
create index on abc(a, lower(b));

ALTER TABLE abc enable ROW LEVEL SECURITY;
ALTER TABLE abc FORCE  ROW LEVEL SECURITY;
CREATE POLICY abc_id_iso_ply on abc to CURRENT_USER USING (a =
(current_setting('app.a'::text))::int);

# for bypass user, index scan works fine
explain analyse select * from abc where a=1 and lower(b)='1234';
     Index Scan using abc_a_lower_idx on abc
         Index Cond: ((a = 1) AND (lower(b) = '1234'::text))

# for RLS user, index scan can only use column a, and filter by lower(b)
set app.a=1;
explain analyse select * from abc where a=1 and lower(b)='1234';
     Index Scan using abc_a_lower_idx on abc
         Index Cond: (a = 1)
         Filter: (lower(b) = '1234'::text)

This only occurs when using non-leak-proof functional index. Everything
works fine in following way:
create index on abc(a, b);
explain analyse select * from abc where a=1 and b='1234';

I think crucial function is restriction_is_securely_promotable. Maybe it is
too strict to reject normal clause match.
Could you please recheck RLS with functional index?

PG Bug reporting form <noreply@postgresql.org> writes:
> # for RLS user, index scan can only use column a, and filter by lower(b)
> set app.a=1;
> explain analyse select * from abc where a=1 and lower(b)='1234';
>      Index Scan using abc_a_lower_idx on abc
>          Index Cond: (a = 1)
>          Filter: (lower(b) = '1234'::text)

AFAICS this is operating as designed.  It's unsafe to apply the
non-leakproof condition until we've verified that the row has a = 1.

In the particular case shown here, it might be all right to do it,
but cases such as bitmap indexscans or lossy index opclasses could
result in live re-evaluations of the indexqual conditions at some
rows.  So we can't safely allow lower(b) to become part of the
indexquals.

            regards, tom lane