View more posts
2021-07-0224 minute read

Protecting reserved roles with PostgreSQL Hooks

Steve ChavezEngineering & PostgREST maintainer

PostgreSQL manages permissions through roles. To create these roles, a database user needs the CREATEROLE privilege, which not only allows role creation but also modification of any role (except superusers).

At Supabase, we use dedicated roles for each of our customers' backend services. For instance, our Storage API  uses the supabase_storage_admin role for connecting to the database. Giving the CREATEROLE privilege to our customers would allow them to drop this role and take their own Storage API down.

And yet, we want to give our customers the ability to manage roles the same as they do it in on-premises databases, with the usual CREATE ROLEALTER ROLE, and DROP ROLE statements.

So, how do we grant them the CREATEROLE privilege and, at the same time, protect our own roles? In this blog post, we explain how we managed to do this by using PostgreSQL Hooks in our SupaUtils extension.

Reserved Roles

PostgreSQL has a list of predefined roles — all prefixed with "pg_" — that cannot be dropped or altered. Attempting to do so will throw an error mentioning that the role is "reserved".

1alter role pg_monitor createdb;
2ERROR:  role name "pg_monitor" is reserved
3DETAIL:  Cannot alter reserved roles.

This mechanism is an internal implementation detail. Unfortunately Postgres doesn't allow us to define our own reserved roles.

RDS reserved roles

Amazon RDS has a similar defense mechanism, all of its predefined roles — prefixed with "rds" — cannot be modified.

1alter role rdsadmin rename to another;
2ERROR: The "rdsadmin" role cannot be renamed.
3DETAIL: The "rdsadmin" role cannot be renamed because either the source
4        or the target name refer to an Amazon RDS reserved role name.
5LOCATION:  handle_rename, rdsutils.c:1534

Again, the error mentions that the role is "reserved".

Also note the rdsutils.c mention. That's not a stock Postgres source file. This means that the logic comes from an RDS extension. We can confirm this is the case by showing the preloaded libraries.

1show shared_preload_libraries;
3  shared_preload_libraries
5 rdsutils,pg_stat_statements

rdsutils can be seen there. Naturally this lead us into thinking we can achieve the same logic with an extension of our own, and thus the SupaUtils idea was born.

Extending PostgreSQL with Hooks

PostgreSQL hooks allow us to extend internal functionality. Hooks can modify behavior at different places, including when running SQL statements.

For example, if we wanted to enforce our own password restrictions whenever a user changes passwords, we could use the check_password_hook to verify the password. We would write out our own Custom Logic, and raise an error if the password fails the password requirements.

This is an images of the lifecycle of a Postgres Hook.

For SupaUtils, we're particularly interested in the ProcessUtility_hook, which allows us to hook into utility statements: every statement except selectinsertdelete or update. They include alter role and drop role, which are the statements we want to hook on.

Hooks are global function pointers

To use hooks, we can override functions pointers that are global. On the Postgres codebase, the ProcessUtility_hook is basically used1 like this:

1// src/backend/tcop/utility.c
3// ProcessUtility_hook is NULL by default
4ProcessUtility_hook_type ProcessUtility_hook = NULL;
6// This function is used for processing all the utility statements
10// call the ProcessUtility_hook if it's not NULL
11if (ProcessUtility_hook)
12  (*ProcessUtility_hook)(PARAMS_OMITTED_FOR_BREVITY);
13// otherwise call the standard function used to process utility statements
15  standard_ProcessUtility(PARAMS_OMITTED_FOR_BREVITY);

As you can see, ProcessUtility_hook is NULL by default, so our extension should set it for the hook to run. Also, the standard_ProcessUtility function is the one that actually does the job of creating or modifying the roles (among other things) so our hook should also call it.

Loading and running the hook

Each extension set in shared_preload_libraries will get its _PG_init function called. This function will allow us to set our hook onto ProcessUtility_hook.

Since hooks are global function pointers, it might be the case that another extension modifies the hook pointer (on its own _PG_init) before us and sets its own hook. So we need to ensure we also run this previously-set hook, before or after our own hook runs.

It's typically2 done like this:

1// variable to store the previous hook
2static ProcessUtility_hook_type prev_hook = NULL;
4// initialize our extension
8  // ProcessUtility_hook has the global function pointer.
9  // Store its value in case another extension already set its own hook.
10  prev_hook = ProcessUtility_hook;
11  // Now override the ProcessUtility_hook with our hook
12  ProcessUtility_hook = our_hook;
15static void
18  // our hook logic goes here
20  // If there was a previous hook, run it after our hook
21  if (prev_hook)
23  // If there's no previous hook, call the standard function
24  else
25    standard_ProcessUtility(PARAMS_OMITTED_FOR_BREVITY);

Setting up the SupaUtils extension

We can use the concepts above to build our extension.

First we'll need a Makefile in order to compile the extension code and include it into our PostgreSQL installation.

1# Makefile
3# Our shared library
4MODULE_big = supautils
6# Our object files to build for the library
7OBJS = src/supautils.o
9# Tell pg_config to pass us the PostgreSQL extensions makefile(PGXS)
10# and include it into our own Makefile through the standard "include" directive.
11PG_CONFIG = pg_config
12PGXS := $(shell $(PG_CONFIG) --pgxs)
13include $(PGXS)

For the source file, we'll start with variable definitions and functions declarations.

1// src/supautils.c
3// include common declarations
4#include "postgres.h"
6// required macro for extension libraries to work
9// variable for the previous hook
10static ProcessUtility_hook_type prev_hook = NULL;
12// variable for our reserved roles configuration parameter
13static char *reserved_roles = NULL;
15// function declaration for extension initialization
16void _PG_init(void);
18// function declaration for our hook
19static void supautils_hook(
20    PlannedStmt *pstmt,
21    const char *queryString,
22    ProcessUtilityContext context,
23    ParamListInfo params,
24    QueryEnvironment *queryEnv,
25    DestReceiver *dest,
26    QueryCompletion *completionTag
29// function declaration for our pure function that will return a reserved role
30static char* look_for_reserved_role(Node *utility_stmt, List *reserved_role_list);

Up next we'll define each one of these function declarations.

Initializing the extension

Let's now _PG_init our extension. Besides setting the hook here, we want to define our reserved roles as a configuration parameter, that way they can be modified by editing the postgresql.conf file. For this, we can use the DefineCustomStringVariable function, which inserts the parameter into Postgres "Grand Unified Configuration"(GUC) system.

4  // Store the previous hook
5  prev_hook = ProcessUtility_hook;
6  // Set our hook
7  ProcessUtility_hook = supautils_hook;
9  // Define our "supautils.reserved_roles" parameter
10  // some arguments are unused so they are left as NULL
11  DefineCustomStringVariable("supautils.reserved_roles",
12  "Comma-separated list of roles that cannot be altered or dropped",
13  NULL,
14  // It will be assigned to the reserved_roles variable
15  &reserved_roles,
16  NULL,
17  // We should be able to reload this parameter without restarting the server,
18  // e.g. with "select pg_reload_conf()".
20  0,

Running the SupaUtils hook

Now that our hook is set, we'll define what it will do. As specified in the ProcessUtility_hook_type, the hook's first parameter is a PlannedStmt, this represents the planned statement — the output from the Postgres planner. This is a step before the statement is executed.

We'll look for the presence of a reserved role in the planned statement. If there's one present, we'll report an error and abort the statement execution step.

1static void
3      // The planned statement
4      PlannedStmt *pstmt,
5      // These parameters are here for completion, we'll not use any of them
6      const char *queryString,
7      ProcessUtilityContext context,
8      ParamListInfo params,
9      QueryEnvironment *queryEnv,
10      DestReceiver *dest,
11      QueryCompletion *completionTag
14  // Get the utility statement from the planned statement
15  Node     *utility_stmt = pstmt->utilityStmt;
17  // Only do the logic if supautils.reserved_roles is not NULL
18  if(reserved_roles){
19    // The found reserved role, assume none was found by default
20    char *reserved_role = NULL;
21    // Temp var for storing the list of reserved roles
22    List *reserved_role_list;
24    // split the comma-separated string into a List by using a
25    // helper function from varlena.h
26    if (!SplitIdentifierString(pstrdup(reserved_roles), ',', &reserved_role_list))
27      // abort and report an error if the splitting fails
28      ereport(ERROR,
30           errmsg("parameter \"%s\" must be a comma-separated list of "
31             "identifiers", reserved_roles)));
33    // look for the reserved role in an internal function
34    reserved_role = look_for_reserved_role(utility_stmt, reserved_role_list);
36    // we're done with the list so free it from memory
37    list_free(reserved_role_list);
39    // abort and report an error if a reserved role was found
40    if(reserved_role)
41      ereport(ERROR,
43         errmsg("\"%s\" is a reserved role, it cannot be modified", reserved_role)));
44  }
46  // Run the previous hook if defined or call the standard function
47  if (prev_hook)
48    prev_hook(pstmt, queryString,
49                context, params, queryEnv,
50                dest, completionTag);
51  else
52    standard_ProcessUtility(pstmt, queryString,
53                context, params, queryEnv,
54                dest, completionTag);

Looking for the reserved role

Lastly, we'll define how we look for the reserved role.

At this stage, we already have the utility statement and the reserved role list. All that's left to do is to define if the utility statement is an ALTER ROLE or DROP ROLE statement, and whether if the role it affects is a reserved one.

1static char*
2look_for_reserved_role(Node *utility_stmt, List *reserved_role_list)
4  // Check the utility statement type
5  switch (utility_stmt->type)
6  {
7    // Matches statements like:
9    case T_AlterRoleStmt:
10    {
11      // cast the utility statement to an alter role statement
12      AlterRoleStmt *stmt = (AlterRoleStmt *) utility_stmt;
14      //RoleSpec has the role name plus some attributes
15      RoleSpec *role = stmt->role;
17      // postgres defines its own list utilities in pg_list.h
18      // ListCell is an element of the list that we'll use for iteration
19      ListCell *role_cell;
21      // pg_list.h includes the foreach macro for iterating over lists
22      foreach(role_cell, reserved_role_list)
23      {
24        // get the list element
25        char *reserved_role = (char *) lfirst(role_cell);
27        // compare the statement role with the reserved role
28        // get_rolespec_name will get the RoleSpec role name,
29        // even in cases where the role is the special case of
30        // "current_user" or "session_user"
31        if (strcmp(get_rolespec_name(role), reserved_role) == 0)
32          return reserved_role;
33      }
35      break;
36    }
38    // Matches statements like:
39    // DROP ROLE role
40    case T_DropRoleStmt:
41    {
42      // cast the utility statement to a drop role statement
43      DropRoleStmt *stmt = (DropRoleStmt *) utility_stmt;
44      ListCell *item;
46      // the logic is the same as before, iterate over the reserved role list
47      // and find a match
48      foreach(item, stmt->roles)
49      {
50        RoleSpec *role = lfirst(item);
51        ListCell *role_cell;
53        foreach(role_cell, reserved_role_list)
54        {
55          char *reserved_role = (char *) lfirst(role_cell);
57          if (strcmp(get_rolespec_name(role), reserved_role) == 0)
58            return reserved_role;
59        }
60      }
62      break;
63    }
65    default:
66      break;
67  }
68  // Didn't find any reserved role on the statement, so return NULL
69  return NULL;

Testing the extension

Now that the code is finished, we can test the extension. Since we already have a Makefile, the extension can be installed by doing make && make install. Then, in postgresql.conf:

1# set the extension as preloaded, this will require a restart
4# the reserved roles
5supautils.reserved_roles="supabase_storage_admin, supabase_auth_admin"

We'll now try to alter or drop the reserved roles:

1alter role supabase_storage_admin nologin password 'fake';
2ERROR:  "supabase_storage_admin" is a reserved role, it cannot be modified
4drop role supabase_auth_admin;
5ERROR:  "supabase_auth_admin" is a reserved role, it cannot be modified
7-- Success!!

Wrapping up

As you can see, PostgreSQL Hooks allow us to intercept SQL statements. There are many types of hooks, you can see unofficial documentation for these at AmatanHead/psql-hooks.

The full SupaUtils code is in our GitHub repository.

By the way, if you like working on PostgreSQL tooling and extensions: we are hiring PostgreSQL experts!

  1. You can see the ProcessUtility_hook declaration here and its usage here.
  2. This is also done on pg_stat_statements and sepgsql.
Related articles
Protecting reserved roles with PostgreSQL Hooks
Developers stay up to date with
Using Supabase in Replit
Postgres as a CRON Server
Cracking PostgreSQL Interview Questions
View all posts
Last post

Supabase Launch Week II: The SQL

22 July 2021
Next post

Supabase Beta June 2021

2 June 2021

Build in a weekend, scale to millions