Merge branch 'datatable-roles-redesign' into datatable-roles-redesign-part-2

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BRoYE5ZeAVvrDYdfhDAYXb

backend/ee-repo-ref.txt

@@ -1 +1 @@
-6f26308c67acf9fcc45773b373aa30a2593b665c
+85e15fc93b238159fc90acfc8427d29e219b88ca

backend/windmill-api-integration-tests/tests/datatable_roles.rs

@@ -1009,6 +1009,68 @@ async fn an_entry_without_roles_cannot_newly_reach_a_database_under_roles(
     Ok(())
 }
 
+#[sqlx::test(migrations = "../migrations", fixtures("base", "datatable_roles"))]
+async fn an_alias_saved_elsewhere_waits_for_roles_going_on_for_its_database(
+    db: Pool<Postgres>,
+) -> anyhow::Result<()> {
+    initialize_tracing().await;
+    sqlx::query(
+        r#"UPDATE workspace_settings SET datatable = jsonb_set(datatable, '{datatables,other}',
+             '{"database": {"resource_type": "instance", "resource_path": "dt_other"}}')
+           WHERE workspace_id = 'test-workspace'"#,
+    )
+    .execute(&db)
+    .await?;
+
+    // Roles going on for `dt_other`, not committed yet: it holds only its own workspace's settings
+    // row, so an alias saved from another workspace that looked for roles now would miss them.
+    let enabling = {
+        let mut tx = db.begin().await?;
+        windmill_common::datatable_roles::lock_instance_databases_governance(
+            &mut *tx,
+            ["dt_other"],
+        )
+        .await?;
+        sqlx::query(
+            r#"UPDATE workspace_settings SET datatable = jsonb_set(datatable,
+                 '{datatables,other,permissions}',
+                 '{"default_role": "admin", "roles": {"admin": {"tenants": ["*"]}}}')
+               WHERE workspace_id = 'test-workspace'"#,
+        )
+        .execute(&mut *tx)
+        .await?;
+        tx
+    };
+
+    let server = ApiServer::start(db.clone()).await?;
+    let url = format!(
+        "http://localhost:{}/api/w/wm-fork-dt/workspaces/edit_datatable_config",
+        server.addr.port()
+    );
+    let save = tokio::spawn(
+        authed(client().post(&url), "SECRET_TOKEN")
+            .json(&json!({ "settings": { "datatables": {
+                "direct": { "database": { "resource_type": "instance", "resource_path": "dt_other" } }
+            } } }))
+            .send(),
+    );
+    tokio::time::sleep(std::time::Duration::from_millis(500)).await;
+    assert!(
+        !save.is_finished(),
+        "an alias was saved while roles were going on for its database"
+    );
+    enabling.commit().await?;
+
+    let resp = save.await??;
+    let status = resp.status();
+    let body = resp.text().await?;
+    assert!(
+        status == 400 && body.contains("which a data table under roles uses"),
+        "the alias reached the database whose roles went on while it waited ({status}): {body}"
+    );
+    Ok(())
+}
+
 #[cfg(not(all(feature = "private", feature = "enterprise")))]
 const ENTERPRISE_REFUSAL: &str = "Data table roles are a Windmill Enterprise Edition feature";
 
@@ -1098,22 +1160,25 @@ async fn without_the_enterprise_edition_a_data_table_under_roles_is_refused_a_co
         assert!(err.to_string().contains(ENTERPRISE_REFUSAL), "{err}");
     }
 
-    // Not under roles, it resolves as it always has; naming a role on it is refused.
+    // Not under roles, it resolves as it always has, including when `admin` is named — which every
+    // migration does; naming any other role on it is refused.
     sqlx::query(
         "UPDATE workspace_settings SET datatable = datatable #- '{datatables,main,permissions}'
          WHERE workspace_id = 'test-workspace'",
     )
     .execute(&db)
     .await?;
-    let resolved = get_datatable_resource_from_db(
-        &db,
-        "test-workspace",
-        "main",
-        None,
-        DatatableAccess::NoIdentity,
-    )
-    .await?;
-    assert_eq!(resolved["dbname"], "dt_main", "{resolved}");
+    for role in [None, Some("admin")] {
+        let resolved = get_datatable_resource_from_db(
+            &db,
+            "test-workspace",
+            "main",
+            role,
+            DatatableAccess::NoIdentity,
+        )
+        .await?;
+        assert_eq!(resolved["dbname"], "dt_main", "{resolved}");
+    }
     let err = get_datatable_resource_from_db(
         &db,
         "test-workspace",

backend/windmill-api-workspaces/src/datatable_migrations.rs

@@ -174,6 +174,10 @@ async fn datatable_database_arg(
     // default role — which is what `ensure_migration_role_allowed` gated it as, and which is the
     // only role a DDL statement can be expected to succeed under. A migration that does declare a
     // role overrides this: the annotation wins over the reference.
+    //
+    // A legacy name containing `?` cannot be migrated through this reference: the appended query
+    // makes it neither an exact name nor a parseable one. Accepted on purpose, since such names can
+    // no longer be created and none are expected to carry migrations.
     Ok(to_raw_value(&format!(
         "datatable://{datatable_name}?role={ADMIN_DATATABLE_ROLE}"
     )))

backend/windmill-api-workspaces/src/workspaces.rs

@@ -3892,50 +3892,64 @@ async fn edit_datatable_config(
     // entry through a declared rename alone, and a settings sync never declares one, so an entry
     // without roles that newly points at such a database — a name added, or an existing one
     // repointed — would answer everyone there as `admin`. That holds whichever workspace governs it.
-    let governed_elsewhere: Vec<String> = sqlx::query_scalar(
-        "SELECT DISTINCT dt.value->'database'->>'resource_path' FROM workspace_settings ws
-         CROSS JOIN LATERAL jsonb_each(COALESCE(ws.datatable->'datatables', '{}'::jsonb)) dt
-         WHERE ws.workspace_id <> $1 AND dt.value ? 'permissions'
-           AND dt.value->'database'->>'resource_type' = 'instance'",
+    let newly_pointed: Vec<(&String, &str)> = new_config
+        .settings
+        .datatables
+        .iter()
+        .filter(|(_, dt)| dt.permissions.is_none())
+        .filter_map(|(name, dt)| {
+            let db = dt
+                .database
+                .as_ref()
+                .filter(|d| d.resource_type == DataTableCatalogResourceType::Instance)?;
+            let lookup = rename_src
+                .get(name.as_str())
+                .copied()
+                .unwrap_or(name.as_str());
+            let repointed = old_datatables
+                .get(lookup)
+                .and_then(|old| old.database.as_ref())
+                .is_none_or(|old_db| {
+                    old_db.resource_type != db.resource_type
+                        || old_db.resource_path != db.resource_path
+                });
+            repointed.then_some((name, db.resource_path.as_str()))
+        })
+        .collect();
+    // Another workspace turning roles on for the same database holds only its own settings row, so
+    // without this the scan below could read past its uncommitted write.
+    windmill_common::datatable_roles::lock_instance_databases_governance(
+        &mut *tx,
+        newly_pointed.iter().map(|(_, dbname)| *dbname),
     )
-    .bind(&w_id)
-    .fetch_all(&mut *tx)
     .await?;
-    for (name, dt) in new_config.settings.datatables.iter() {
-        if dt.permissions.is_some() {
-            continue;
-        }
-        let Some(db) = dt
-            .database
-            .as_ref()
-            .filter(|d| d.resource_type == DataTableCatalogResourceType::Instance)
-        else {
-            continue;
-        };
-        let lookup = rename_src
-            .get(name.as_str())
-            .copied()
-            .unwrap_or(name.as_str());
-        let repointed = old_datatables
-            .get(lookup)
-            .and_then(|old| old.database.as_ref())
-            .is_none_or(|old_db| {
-                old_db.resource_type != db.resource_type || old_db.resource_path != db.resource_path
-            });
+    let governed_elsewhere: Vec<String> = if newly_pointed.is_empty() {
+        vec![]
+    } else {
+        sqlx::query_scalar(
+            "SELECT DISTINCT dt.value->'database'->>'resource_path' FROM workspace_settings ws
+             CROSS JOIN LATERAL jsonb_each(COALESCE(ws.datatable->'datatables', '{}'::jsonb)) dt
+             WHERE ws.workspace_id <> $1 AND dt.value ? 'permissions'
+               AND dt.value->'database'->>'resource_type' = 'instance'",
+        )
+        .bind(&w_id)
+        .fetch_all(&mut *tx)
+        .await?
+    };
+    for (name, dbname) in newly_pointed {
         let governed_here = old_datatables.values().any(|old| {
             old.permissions.is_some()
                 && old.database.as_ref().is_some_and(|d| {
                     d.resource_type == DataTableCatalogResourceType::Instance
-                        && d.resource_path == db.resource_path
+                        && d.resource_path == dbname
                 })
         });
-        if repointed && (governed_here || governed_elsewhere.contains(&db.resource_path)) {
+        if governed_here || governed_elsewhere.iter().any(|g| g == dbname) {
             return Err(Error::BadRequest(format!(
-                "Data table '{name}' would point at database '{}', which a data table under roles \
-                 uses, without carrying those roles: everyone reaching '{name}' would connect there \
-                 as `admin`. Rename the data table under roles from the data table settings, which \
-                 carries its roles, or turn its roles off first.",
-                db.resource_path
+                "Data table '{name}' would point at database '{dbname}', which a data table under \
+                 roles uses, without carrying those roles: everyone reaching '{name}' would connect \
+                 there as `admin`. Rename the data table under roles from the data table settings, \
+                 which carries its roles, or turn its roles off first."
             )));
         }
     }

backend/windmill-common/src/datatable_roles.rs

@@ -140,6 +140,25 @@ pub async fn lock_datatable_streams(conn: &mut sqlx::PgConnection, exclusive: bo
     Ok(())
 }
 
+/// Whether an instance database is reached only through entries under roles is decided by two
+/// writes that lock different workspaces' settings rows: turning roles on for one entry, and a
+/// settings save pointing an entry without roles at the database. Each holds this for every
+/// database it decides on, so neither reads past the other's uncommitted write. Held for the
+/// transaction; the names are locked in sorted order so two holders cannot deadlock.
+pub async fn lock_instance_databases_governance<'a>(
+    conn: &mut sqlx::PgConnection,
+    dbnames: impl IntoIterator<Item = &'a str>,
+) -> Result<()> {
+    let dbnames: std::collections::BTreeSet<&str> = dbnames.into_iter().collect();
+    for dbname in dbnames {
+        sqlx::query("SELECT pg_advisory_xact_lock(hashtext('datatable_instance_database:' || $1))")
+            .bind(dbname)
+            .execute(&mut *conn)
+            .await?;
+    }
+    Ok(())
+}
+
 /// Disclosure: returns every role's stored Postgres password in plaintext. Any server path that
 /// has to resolve or name a role may call it — including handlers open to a workspace member, who
 /// need the names — but callers MUST NOT let `pwd` reach a response, a log line, an audit record

backend/windmill-common/src/datatable_roles_oss.rs

@@ -163,8 +163,8 @@ mod ce {
         Err(unavailable())
     }
 
-    /// A data table not under roles, asked for no role, is not a role decision and passes, as it
-    /// did before roles existed. Anything else is refused.
+    /// A data table not under roles, asked for no role or for `admin`, is not a role decision and
+    /// passes, as it did before roles existed. Anything else is refused.
     pub(crate) async fn ensure_can_use_datatable_role(
         db: &DB,
         w_id: &str,
@@ -174,7 +174,9 @@ mod ce {
         _context: &str,
     ) -> Result<()> {
         let governing = resolve_governing_datatable(db, w_id, name).await?;
-        if governing.datatable.permissions.is_none() && role.is_none() {
+        if governing.datatable.permissions.is_none()
+            && role.is_none_or(|r| r == crate::datatable_roles::ADMIN_DATATABLE_ROLE)
+        {
             Ok(())
         } else {
             Err(unavailable())

backend/windmill-common/src/workspaces.rs

@@ -1664,9 +1664,10 @@ pub async fn get_datatable_resource_from_db(
 ) -> Result<serde_json::Value> {
     let governing = resolve_governing_datatable(db, w_id, name).await?;
     let db_resource = resolve_datatable_connection_unchecked(db, &governing, false).await?;
-    // Not under roles and asked for none: the `admin` connection, as before roles existed, in
-    // every edition. Anything else is a role decision.
-    if governing.datatable.permissions.is_none() && role.is_none() {
+    // Not under roles and asked for none, or for `admin` by name: the `admin` connection, as before
+    // roles existed, in every edition. Anything else is a role decision. Every migration names
+    // `admin` explicitly, so an edition without roles must not treat that as one.
+    if governing.datatable.permissions.is_none() && role.is_none_or(|r| r == ADMIN_DATATABLE_ROLE) {
         return Ok(db_resource);
     }
     crate::datatable_roles_oss::resolve_datatable_role_connection(