]> git.ipfire.org Git - thirdparty/postgresql.git/commitdiff
libpq-oauth: Avoid overflow for very large intervals
authorJacob Champion <jchampion@postgresql.org>
Fri, 31 Jul 2026 18:13:41 +0000 (11:13 -0700)
committerJacob Champion <jchampion@postgresql.org>
Fri, 31 Jul 2026 18:13:41 +0000 (11:13 -0700)
The slow_down interval parsing code checks explicitly for overflow, but
since it does that after the signed overflow has already occurred, we
end up inviting undefined behavior from the compiler anyway.

Use checked arithmetic instead. set_timer() takes a long int in order to
interface nicely with libcurl, so use an int32 as the interval counter
and clamp to LONG_MAX during conversion to milliseconds.

Backpatch to 18, where libpq-oauth was introduced.

Reported-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Discussion: https://postgr.es/m/qtclihmrkq67ach3xjxyi4qcksstin5qxwsnkqefkmotxwh4g6%40ae2bj6jvcmry
Backpatch-through: 18

src/interfaces/libpq-oauth/oauth-curl.c

index d4dcc4cd7a53af2e7a3db047c3513e0431c509eb..bb29f67933e43a5aee11283d065096c49c741527 100644 (file)
@@ -28,6 +28,7 @@
 #error libpq-oauth is not supported on this platform
 #endif
 
+#include "common/int.h"
 #include "common/jsonapi.h"
 #include "mb/pg_wchar.h"
 #include "oauth-curl.h"
@@ -136,7 +137,7 @@ struct device_authz
 
        /* Fields below are parsed from the corresponding string above. */
        int                     expires_in;
-       int                     interval;
+       int32           interval;
 };
 
 static void
@@ -1020,7 +1021,7 @@ parse_json_number(const char *s)
  * expensive network polling loop.) Tests may remove the lower bound with
  * PGOAUTHDEBUG, for improved performance.
  */
-static int
+static int32
 parse_interval(struct async_ctx *actx, const char *interval_str)
 {
        double          parsed;
@@ -1031,8 +1032,8 @@ parse_interval(struct async_ctx *actx, const char *interval_str)
        if (parsed < 1)
                return (actx->debug_flags & OAUTHDEBUG_UNSAFE_DOS_ENDPOINT) ? 0 : 1;
 
-       else if (parsed >= INT_MAX)
-               return INT_MAX;
+       else if (parsed >= INT32_MAX)
+               return INT32_MAX;
 
        return parsed;
 }
@@ -2620,10 +2621,7 @@ handle_token_response(struct async_ctx *actx, char **token)
         */
        if (strcmp(err->error, "slow_down") == 0)
        {
-               int                     prev_interval = actx->authz.interval;
-
-               actx->authz.interval += 5;
-               if (actx->authz.interval < prev_interval)
+               if (pg_add_s32_overflow(actx->authz.interval, 5, &actx->authz.interval))
                {
                        actx_error(actx, "slow_down interval overflow");
                        goto token_cleanup;
@@ -2964,8 +2962,25 @@ pg_fe_run_oauth_flow_impl(PGconn *conn, PGoauthBearerRequestV2 *request,
                                 * Wait for the required interval before issuing the next
                                 * request.
                                 */
-                               if (!set_timer(actx, actx->authz.interval * 1000))
-                                       goto error_return;
+                               {
+                                       /*
+                                        * Avoid overflow of long int. (By the time we reach
+                                        * LONG_MAX milliseconds -- 24 days on 32-bit platforms --
+                                        * continuing to honor slow_down requests seems pretty
+                                        * pointless anyway.)
+                                        */
+                                       int64           interval_ms;
+
+                                       if (pg_mul_s64_overflow(actx->authz.interval, 1000,
+                                                                                       &interval_ms)
+                                               || (interval_ms > LONG_MAX))
+                                       {
+                                               interval_ms = LONG_MAX;
+                                       }
+
+                                       if (!set_timer(actx, (long) interval_ms))
+                                               goto error_return;
+                               }
 
                                /*
                                 * No Curl requests are running, so we can simplify by having