about summary refs log tree commit diff
diff options
context:
space:
mode:
authorFlorian Weimer <fweimer@redhat.com>2017-02-28 15:28:45 +0100
committerFlorian Weimer <fweimer@redhat.com>2017-02-28 15:36:17 +0100
commitcf0bd2f73bd65beab613865bba567d7787836888 (patch)
tree8cde5a2a9547382fac6546c9397aebbb914cb485
parent37fb019cb02656d0ce0b8d40d56fe8c42f0d1658 (diff)
downloadglibc-cf0bd2f73bd65beab613865bba567d7787836888.tar.gz
glibc-cf0bd2f73bd65beab613865bba567d7787836888.tar.xz
glibc-cf0bd2f73bd65beab613865bba567d7787836888.zip
sunrpc: Improvements for UDP client timeout handling [BZ #20257]
This commit fixes various aspects in the UDP client timeout handling.
Timeouts are now applied in a more consistent fashion.  Discarded UDP
packets no longer prevent the timeout from happening at all.
-rw-r--r--ChangeLog23
-rw-r--r--inet/Makefile8
-rw-r--r--inet/deadline.c122
-rw-r--r--inet/net-internal.h89
-rw-r--r--inet/tst-deadline.c188
-rw-r--r--sunrpc/Makefile10
-rw-r--r--sunrpc/clnt_udp.c127
-rw-r--r--sunrpc/tst-udp-garbage.c104
-rw-r--r--sunrpc/tst-udp-nonblocking.c333
-rw-r--r--sunrpc/tst-udp-timeout.c402
10 files changed, 1347 insertions, 59 deletions
diff --git a/ChangeLog b/ChangeLog
index edccd36002..b3d7adae20 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,28 @@
 2017-02-28  Florian Weimer  <fweimer@redhat.com>
 
+	[BZ #20257]
+	* inet/Makefile (routines): Add deadline.
+	(tests-static): Add tst-deadline.
+	* inet/net-internal.h (struct deadline_current_time)
+	(__deadline_current_time, struct deadline, __deadline_is_infinite)
+	(__deadline_elapsed, __deadline_first, __deadline_from_timeval)
+	(__deadline_to_ms, __is_timeval_valid_timeout): Declare.
+	* inet/deadline.c: New file.
+	* inet/tst-deadline.c: Likewise.
+	* sunrpc/Makefile (tests): Add tst-udp-nonblocking,
+	tst-udp-timeout, tst-udp-garbage.
+	(tst-udp-nonblocking, tst-udp-timeout): Link against libc.so
+	explicitly.
+	(tst-udp-garbage): Likewise.  Also link against thread library.
+	* sunrpc/clnt_udp.c (struct cu_data): Mention in comment that the
+	struct layout is part of the ABI.
+	(clntudp_call): Rework timeout handling.
+	* sunrpc/tst-udp-garbage.c: New file.
+	* sunrpc/tst-udp-nonblocking.c: Likewise.
+	* sunrpc/tst-udp-timeout.c: Likewise.
+
+2017-02-28  Florian Weimer  <fweimer@redhat.com>
+
 	[BZ #5010]
 	* sunrpc/svc.c (svc_is_mapped): Remove.
 	(svc_unregister): Obtain mapped status while the service is still
diff --git a/inet/Makefile b/inet/Makefile
index 010792af8f..6a7d3e0664 100644
--- a/inet/Makefile
+++ b/inet/Makefile
@@ -45,14 +45,18 @@ routines := htonl htons		\
 	    in6_addr getnameinfo if_index ifaddrs inet6_option \
 	    getipv4sourcefilter setipv4sourcefilter \
 	    getsourcefilter setsourcefilter inet6_opt inet6_rth \
-	    inet6_scopeid_pton
+	    inet6_scopeid_pton deadline
 
 aux := check_pf check_native ifreq
 
 tests := htontest test_ifindex tst-ntoa tst-ether_aton tst-network \
 	 tst-gethnm test-ifaddrs bug-if1 test-inet6_opt tst-ether_line \
 	 tst-getni1 tst-getni2 tst-inet6_rth tst-checks tst-checks-posix \
-	 tst-sockaddr tst-inet6_scopeid_pton test-hnto-types
+	 tst-sockaddr tst-inet6_scopeid_pton test-hnto-types tst-deadline
+
+# tst-deadline must be linked statically so that we can access
+# internal functions.
+tests-static += tst-deadline
 
 include ../Rules
 
diff --git a/inet/deadline.c b/inet/deadline.c
new file mode 100644
index 0000000000..c1fa415a39
--- /dev/null
+++ b/inet/deadline.c
@@ -0,0 +1,122 @@
+/* Computing deadlines for timeouts.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <net-internal.h>
+
+#include <assert.h>
+#include <limits.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <time.h>
+
+struct deadline_current_time internal_function
+__deadline_current_time (void)
+{
+  struct deadline_current_time result;
+  if (__clock_gettime (CLOCK_MONOTONIC, &result.current) != 0)
+    {
+      struct timeval current_tv;
+      if (__gettimeofday (&current_tv, NULL) == 0)
+        __libc_fatal ("Fatal error: gettimeofday system call failed\n");
+      result.current.tv_sec = current_tv.tv_sec;
+      result.current.tv_nsec = current_tv.tv_usec * 1000;
+    }
+  assert (result.current.tv_sec >= 0);
+  return result;
+}
+
+/* A special deadline value for which __deadline_is_infinite is
+   true.  */
+static inline struct deadline
+infinite_deadline (void)
+{
+  return (struct deadline) { { -1, -1 } };
+}
+
+struct deadline internal_function
+__deadline_from_timeval (struct deadline_current_time current,
+                         struct timeval tv)
+{
+  assert (__is_timeval_valid_timeout (tv));
+
+  /* Compute second-based deadline.  Perform the addition in
+     uintmax_t, which is unsigned, to simply overflow detection.  */
+  uintmax_t sec = current.current.tv_sec;
+  sec += tv.tv_sec;
+  if (sec < (uintmax_t) tv.tv_sec)
+    return infinite_deadline ();
+
+  /* Compute nanosecond deadline.  */
+  int nsec = current.current.tv_nsec + tv.tv_usec * 1000;
+  if (nsec >= 1000 * 1000 * 1000)
+    {
+      /* Carry nanosecond overflow to seconds.  */
+      nsec -= 1000 * 1000 * 1000;
+      if (sec + 1 < sec)
+        return infinite_deadline ();
+      ++sec;
+    }
+  /* This uses a GCC extension, otherwise these casts for detecting
+     overflow would not be defined.  */
+  if ((time_t) sec < 0 || sec != (uintmax_t) (time_t) sec)
+    return infinite_deadline ();
+
+  return (struct deadline) { { sec, nsec } };
+}
+
+int internal_function
+__deadline_to_ms (struct deadline_current_time current,
+                  struct deadline deadline)
+{
+  if (__deadline_is_infinite (deadline))
+    return INT_MAX;
+
+  if (current.current.tv_sec > deadline.absolute.tv_sec
+      || (current.current.tv_sec == deadline.absolute.tv_sec
+          && current.current.tv_nsec >= deadline.absolute.tv_nsec))
+    return 0;
+  time_t sec = deadline.absolute.tv_sec - current.current.tv_sec;
+  if (sec >= INT_MAX)
+    /* This value will overflow below.  */
+    return INT_MAX;
+  int nsec = deadline.absolute.tv_nsec - current.current.tv_nsec;
+  if (nsec < 0)
+    {
+      /* Borrow from the seconds field.  */
+      assert (sec > 0);
+      --sec;
+      nsec += 1000 * 1000 * 1000;
+    }
+
+  /* Prepare for rounding up to milliseconds.  */
+  nsec += 999999;
+  if (nsec > 1000 * 1000 * 1000)
+    {
+      assert (sec < INT_MAX);
+      ++sec;
+      nsec -= 1000 * 1000 * 1000;
+    }
+
+  unsigned int msec = nsec / (1000 * 1000);
+  if (sec > INT_MAX / 1000)
+    return INT_MAX;
+  msec += sec * 1000;
+  if (msec > INT_MAX)
+    return INT_MAX;
+  return msec;
+}
diff --git a/inet/net-internal.h b/inet/net-internal.h
index 087597ed99..2b2632c7ba 100644
--- a/inet/net-internal.h
+++ b/inet/net-internal.h
@@ -20,11 +20,100 @@
 #define _NET_INTERNAL_H 1
 
 #include <arpa/inet.h>
+#include <stdbool.h>
 #include <stdint.h>
+#include <sys/time.h>
 
 int __inet6_scopeid_pton (const struct in6_addr *address,
                           const char *scope, uint32_t *result)
   internal_function attribute_hidden;
 libc_hidden_proto (__inet6_scopeid_pton)
 
+
+/* Deadline handling for enforcing timeouts.
+
+   Code should call __deadline_current_time to obtain the current time
+   and cache it locally.  The cache needs updating after every
+   long-running or potentially blocking operation.  Deadlines relative
+   to the current time can be computed using __deadline_from_timeval.
+   The deadlines may have to be recomputed in response to certain
+   events (such as an incoming packet), but they are absolute (not
+   relative to the current time).  A timeout suitable for use with the
+   poll function can be computed from such a deadline using
+   __deadline_to_ms.
+
+   The fields in the structs defined belowed should only be used
+   within the implementation.  */
+
+/* Cache of the current time.  Used to compute deadlines from relative
+   timeouts and vice versa.  */
+struct deadline_current_time
+{
+  struct timespec current;
+};
+
+/* Return the current time.  Terminates the process if the current
+   time is not available.  */
+struct deadline_current_time __deadline_current_time (void)
+  internal_function attribute_hidden;
+
+/* Computed absolute deadline.  */
+struct deadline
+{
+  struct timespec absolute;
+};
+
+
+/* For internal use only.  */
+static inline bool
+__deadline_is_infinite (struct deadline deadline)
+{
+  return deadline.absolute.tv_nsec < 0;
+}
+
+/* Return true if the current time is at the deadline or past it.  */
+static inline bool
+__deadline_elapsed (struct deadline_current_time current,
+                    struct deadline deadline)
+{
+  return !__deadline_is_infinite (deadline)
+    && (current.current.tv_sec > deadline.absolute.tv_sec
+        || (current.current.tv_sec == deadline.absolute.tv_sec
+            && current.current.tv_nsec >= deadline.absolute.tv_nsec));
+}
+
+/* Return the deadline which occurs first.  */
+static inline struct deadline
+__deadline_first (struct deadline left, struct deadline right)
+{
+  if (__deadline_is_infinite (right)
+      || left.absolute.tv_sec < right.absolute.tv_sec
+      || (left.absolute.tv_sec == right.absolute.tv_sec
+          && left.absolute.tv_nsec < right.absolute.tv_nsec))
+    return left;
+  else
+    return right;
+}
+
+/* Add TV to the current time and return it.  Returns a special
+   infinite absolute deadline on overflow.  */
+struct deadline __deadline_from_timeval (struct deadline_current_time,
+                                         struct timeval tv)
+  internal_function attribute_hidden;
+
+/* Compute the number of milliseconds until the specified deadline,
+   from the current time in the argument.  The result is mainly for
+   use with poll.  If the deadline has already passed, return 0.  If
+   the result would overflow an int, return INT_MAX.  */
+int __deadline_to_ms (struct deadline_current_time, struct deadline)
+  internal_function attribute_hidden;
+
+/* Return true if TV.tv_sec is non-negative and TV.tv_usec is in the
+   interval [0, 999999].  */
+static inline bool
+__is_timeval_valid_timeout (struct timeval tv)
+{
+  return tv.tv_sec >= 0 && tv.tv_usec >= 0 && tv.tv_usec < 1000 * 1000;
+}
+
 #endif /* _NET_INTERNAL_H */
diff --git a/inet/tst-deadline.c b/inet/tst-deadline.c
new file mode 100644
index 0000000000..ed04345c35
--- /dev/null
+++ b/inet/tst-deadline.c
@@ -0,0 +1,188 @@
+/* Tests for computing deadlines for timeouts.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <inet/net-internal.h>
+#include <limits.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <support/check.h>
+
+/* Find the maximum value which can be represented in a time_t.  */
+static time_t
+time_t_max (void)
+{
+  _Static_assert (0 > (time_t) -1, "time_t is signed");
+  uintmax_t current = 1;
+  while (true)
+    {
+      uintmax_t next = current * 2;
+      /* This cannot happen because time_t is signed.  */
+      TEST_VERIFY_EXIT (next > current);
+      ++next;
+      if ((time_t) next < 0 || next != (uintmax_t) (time_t) next)
+        /* Value cannot be represented in time_t.  Return the previous
+           value. */
+        return current;
+      current = next;
+    }
+}
+
+static int
+do_test (void)
+{
+  {
+    struct deadline_current_time current_time = __deadline_current_time ();
+    TEST_VERIFY (current_time.current.tv_sec >= 0);
+    current_time = __deadline_current_time ();
+    /* Due to CLOCK_MONOTONIC, either seconds or nanoseconds are
+       greater than zero.  This is also true for the gettimeofday
+       fallback.  */
+    TEST_VERIFY (current_time.current.tv_sec >= 0);
+    TEST_VERIFY (current_time.current.tv_sec > 0
+                 || current_time.current.tv_nsec > 0);
+  }
+
+  /* Check basic computations of deadlines.  */
+  struct deadline_current_time current_time = { { 1, 123456789 } };
+  struct deadline deadline = __deadline_from_timeval
+    (current_time, (struct timeval) { 0, 1 });
+  TEST_VERIFY (deadline.absolute.tv_sec == 1);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 123457789);
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 1);
+
+  deadline = __deadline_from_timeval
+    (current_time, ((struct timeval) { 0, 2 }));
+  TEST_VERIFY (deadline.absolute.tv_sec == 1);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 123458789);
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 1);
+
+  deadline = __deadline_from_timeval
+    (current_time, ((struct timeval) { 1, 0 }));
+  TEST_VERIFY (deadline.absolute.tv_sec == 2);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 123456789);
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 1000);
+
+  /* Check if timeouts are correctly rounded up to the next
+     millisecond.  */
+  for (int i = 0; i < 999999; ++i)
+    {
+      ++current_time.current.tv_nsec;
+      TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 1000);
+    }
+
+  /* A full millisecond has elapsed, so the time to the deadline is
+     now less than 1000.  */
+  ++current_time.current.tv_nsec;
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 999);
+
+  /* Check __deadline_to_ms carry-over.  */
+  current_time = (struct deadline_current_time) { { 9, 123456789 } };
+  deadline = (struct deadline) { { 10, 122456789 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 999);
+  deadline = (struct deadline) { { 10, 122456790 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 1000);
+  deadline = (struct deadline) { { 10, 123456788 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 1000);
+  deadline = (struct deadline) { { 10, 123456789 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 1000);
+
+  /* Check __deadline_to_ms overflow.  */
+  deadline = (struct deadline) { { INT_MAX - 1, 1 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == INT_MAX);
+
+  /* Check __deadline_to_ms for elapsed deadlines.  */
+  current_time = (struct deadline_current_time) { { 9, 123456789 } };
+  deadline.absolute = current_time.current;
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 0);
+  current_time = (struct deadline_current_time) { { 9, 123456790 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 0);
+  current_time = (struct deadline_current_time) { { 10, 0 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 0);
+  current_time = (struct deadline_current_time) { { 10, 123456788 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 0);
+  current_time = (struct deadline_current_time) { { 10, 123456789 } };
+  TEST_VERIFY (__deadline_to_ms (current_time, deadline) == 0);
+
+  /* Check carry-over in __deadline_from_timeval.  */
+  current_time = (struct deadline_current_time) { { 9, 998000001 } };
+  for (int i = 0; i < 2000; ++i)
+    {
+      deadline = __deadline_from_timeval
+        (current_time, (struct timeval) { 1, i });
+      TEST_VERIFY (deadline.absolute.tv_sec == 10);
+      TEST_VERIFY (deadline.absolute.tv_nsec == 998000001 + i * 1000);
+    }
+  for (int i = 2000; i < 3000; ++i)
+    {
+      deadline = __deadline_from_timeval
+        (current_time, (struct timeval) { 2, i });
+      TEST_VERIFY (deadline.absolute.tv_sec == 12);
+      TEST_VERIFY (deadline.absolute.tv_nsec == 1 + (i - 2000) * 1000);
+    }
+
+  /* Check infinite deadlines.  */
+  deadline = __deadline_from_timeval
+    ((struct deadline_current_time) { { 0, 1000 * 1000 * 1000 - 1000 } },
+     (struct timeval) { time_t_max (), 1 });
+  TEST_VERIFY (__deadline_is_infinite (deadline));
+  deadline = __deadline_from_timeval
+    ((struct deadline_current_time) { { 0, 1000 * 1000 * 1000 - 1001 } },
+     (struct timeval) { time_t_max (), 1 });
+  TEST_VERIFY (!__deadline_is_infinite (deadline));
+  deadline = __deadline_from_timeval
+    ((struct deadline_current_time)
+       { { time_t_max (), 1000 * 1000 * 1000 - 1000 } },
+     (struct timeval) { 0, 1 });
+  TEST_VERIFY (__deadline_is_infinite (deadline));
+  deadline = __deadline_from_timeval
+    ((struct deadline_current_time)
+       { { time_t_max () / 2 + 1, 0 } },
+     (struct timeval) { time_t_max () / 2 + 1, 0 });
+  TEST_VERIFY (__deadline_is_infinite (deadline));
+
+  /* Check __deadline_first behavior.  */
+  deadline = __deadline_first
+    ((struct deadline) { { 1, 2 } },
+     (struct deadline) { { 1, 3 } });
+  TEST_VERIFY (deadline.absolute.tv_sec == 1);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 2);
+  deadline = __deadline_first
+    ((struct deadline) { { 1, 3 } },
+     (struct deadline) { { 1, 2 } });
+  TEST_VERIFY (deadline.absolute.tv_sec == 1);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 2);
+  deadline = __deadline_first
+    ((struct deadline) { { 1, 2 } },
+     (struct deadline) { { 2, 1 } });
+  TEST_VERIFY (deadline.absolute.tv_sec == 1);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 2);
+  deadline = __deadline_first
+    ((struct deadline) { { 1, 2 } },
+     (struct deadline) { { 2, 4 } });
+  TEST_VERIFY (deadline.absolute.tv_sec == 1);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 2);
+  deadline = __deadline_first
+    ((struct deadline) { { 2, 4 } },
+     (struct deadline) { { 1, 2 } });
+  TEST_VERIFY (deadline.absolute.tv_sec == 1);
+  TEST_VERIFY (deadline.absolute.tv_nsec == 2);
+
+  return 0;
+}
+
+#include <support/test-driver.c>
diff --git a/sunrpc/Makefile b/sunrpc/Makefile
index 0249e10545..063a76161c 100644
--- a/sunrpc/Makefile
+++ b/sunrpc/Makefile
@@ -93,12 +93,13 @@ rpcgen-objs = rpc_main.o rpc_hout.o rpc_cout.o rpc_parse.o \
 extra-objs = $(rpcgen-objs) $(addprefix cross-,$(rpcgen-objs))
 others += rpcgen
 
-tests = tst-xdrmem tst-xdrmem2 test-rpcent tst-udp-error
+tests = tst-xdrmem tst-xdrmem2 test-rpcent tst-udp-error tst-udp-timeout \
+  tst-udp-nonblocking
 xtests := tst-getmyaddr
 
 ifeq ($(have-thread-library),yes)
 xtests += thrsvc
-tests += tst-svc_register
+tests += tst-svc_register tst-udp-garbage
 endif
 
 ifeq ($(run-built-tests),yes)
@@ -238,3 +239,8 @@ $(rpcgen-tests): $(objpfx)%.out: %.x $(objpfx)rpcgen
 	$(built-program-cmd) -c $< -o $@; \
 	$(evaluate-test)
 endif
+
+$(objpfx)tst-udp-timeout: $(common-objpfx)linkobj/libc.so
+$(objpfx)tst-udp-nonblocking: $(common-objpfx)linkobj/libc.so
+$(objpfx)tst-udp-garbage: \
+  $(common-objpfx)linkobj/libc.so $(shared-thread-library)
diff --git a/sunrpc/clnt_udp.c b/sunrpc/clnt_udp.c
index 1de25cb771..6ce16eb298 100644
--- a/sunrpc/clnt_udp.c
+++ b/sunrpc/clnt_udp.c
@@ -55,6 +55,7 @@
 #endif
 
 #include <kernel-features.h>
+#include <inet/net-internal.h>
 
 extern u_long _create_xid (void);
 
@@ -80,7 +81,9 @@ static const struct clnt_ops udp_ops =
 };
 
 /*
- * Private data kept per client handle
+ * Private data kept per client handle.  This private struct is
+ * unfortunately part of the ABI; ypbind contains a copy of it and
+ * accesses it through CLIENT::cl_private field.
  */
 struct cu_data
   {
@@ -278,28 +281,38 @@ clntudp_call (/* client handle */
   int inlen;
   socklen_t fromlen;
   struct pollfd fd;
-  int milliseconds = (cu->cu_wait.tv_sec * 1000) +
-    (cu->cu_wait.tv_usec / 1000);
   struct sockaddr_in from;
   struct rpc_msg reply_msg;
   XDR reply_xdrs;
-  struct timeval time_waited;
   bool_t ok;
   int nrefreshes = 2;		/* number of times to refresh cred */
-  struct timeval timeout;
   int anyup;			/* any network interface up */
 
-  if (cu->cu_total.tv_usec == -1)
-    {
-      timeout = utimeout;	/* use supplied timeout */
-    }
-  else
+  struct deadline_current_time current_time = __deadline_current_time ();
+  struct deadline total_deadline; /* Determined once by overall timeout.  */
+  struct deadline response_deadline; /* Determined anew for each query.  */
+
+  /* Choose the timeout value.  For non-sending usage (xargs == NULL),
+     the total deadline does not matter, only cu->cu_wait is used
+     below.  */
+  if (xargs != NULL)
     {
-      timeout = cu->cu_total;	/* use default timeout */
+      struct timeval tv;
+      if (cu->cu_total.tv_usec == -1)
+	/* Use supplied timeout.  */
+	tv = utimeout;
+      else
+	/* Use default timeout.  */
+	tv = cu->cu_total;
+      if (!__is_timeval_valid_timeout (tv))
+	return (cu->cu_error.re_status = RPC_TIMEDOUT);
+      total_deadline = __deadline_from_timeval (current_time, tv);
     }
 
-  time_waited.tv_sec = 0;
-  time_waited.tv_usec = 0;
+  /* Guard against bad timeout specification.  */
+  if (!__is_timeval_valid_timeout (cu->cu_wait))
+    return (cu->cu_error.re_status = RPC_TIMEDOUT);
+
 call_again:
   xdrs = &(cu->cu_outxdrs);
   if (xargs == NULL)
@@ -325,27 +338,46 @@ send_again:
       return (cu->cu_error.re_status = RPC_CANTSEND);
     }
 
-  /*
-   * Hack to provide rpc-based message passing
-   */
-  if (timeout.tv_sec == 0 && timeout.tv_usec == 0)
-    {
-      return (cu->cu_error.re_status = RPC_TIMEDOUT);
-    }
+  /* sendto may have blocked, so recompute the current time.  */
+  current_time = __deadline_current_time ();
  get_reply:
-  /*
-   * sub-optimal code appears here because we have
-   * some clock time to spare while the packets are in flight.
-   * (We assume that this is actually only executed once.)
-   */
+  response_deadline = __deadline_from_timeval (current_time, cu->cu_wait);
+
   reply_msg.acpted_rply.ar_verf = _null_auth;
   reply_msg.acpted_rply.ar_results.where = resultsp;
   reply_msg.acpted_rply.ar_results.proc = xresults;
   fd.fd = cu->cu_sock;
   fd.events = POLLIN;
   anyup = 0;
+
+  /* Per-response retry loop.  current_time must be up-to-date at the
+     top of the loop.  */
   for (;;)
     {
+      int milliseconds;
+      if (xargs != NULL)
+	{
+	  if (__deadline_elapsed (current_time, total_deadline))
+	    /* Overall timeout expired.  */
+	    return (cu->cu_error.re_status = RPC_TIMEDOUT);
+	  milliseconds = __deadline_to_ms
+	    (current_time, __deadline_first (total_deadline,
+					     response_deadline));
+	  if (milliseconds == 0)
+	    /* Per-query timeout expired.  */
+	    goto send_again;
+	}
+      else
+	{
+	  /* xatgs == NULL.  Collect a response without sending a
+	     query.  In this mode, we need to ignore the total
+	     deadline.  */
+	  milliseconds = __deadline_to_ms (current_time, response_deadline);
+	  if (milliseconds == 0)
+	    /* Cannot send again, so bail out.  */
+	    return (cu->cu_error.re_status = RPC_CANTSEND);
+	}
+
       switch (__poll (&fd, 1, milliseconds))
 	{
 
@@ -356,27 +388,10 @@ send_again:
 	      if (!anyup)
 		return (cu->cu_error.re_status = RPC_CANTRECV);
 	    }
-
-	  time_waited.tv_sec += cu->cu_wait.tv_sec;
-	  time_waited.tv_usec += cu->cu_wait.tv_usec;
-	  while (time_waited.tv_usec >= 1000000)
-	    {
-	      time_waited.tv_sec++;
-	      time_waited.tv_usec -= 1000000;
-	    }
-	  if ((time_waited.tv_sec < timeout.tv_sec) ||
-	      ((time_waited.tv_sec == timeout.tv_sec) &&
-	       (time_waited.tv_usec < timeout.tv_usec)))
-	    goto send_again;
-	  return (cu->cu_error.re_status = RPC_TIMEDOUT);
-
-	  /*
-	   * buggy in other cases because time_waited is not being
-	   * updated.
-	   */
+	  goto next_response;
 	case -1:
 	  if (errno == EINTR)
-	    continue;
+	    goto next_response;
 	  cu->cu_error.re_errno = errno;
 	  return (cu->cu_error.re_status = RPC_CANTRECV);
 	}
@@ -440,20 +455,22 @@ send_again:
       if (inlen < 0)
 	{
 	  if (errno == EWOULDBLOCK)
-	    continue;
+	    goto next_response;
 	  cu->cu_error.re_errno = errno;
 	  return (cu->cu_error.re_status = RPC_CANTRECV);
 	}
-      if (inlen < 4)
-	continue;
-
-      /* see if reply transaction id matches sent id.
-	Don't do this if we only wait for a replay */
-      if (xargs != NULL
-	  && memcmp (cu->cu_inbuf, cu->cu_outbuf, sizeof (u_int32_t)) != 0)
-	continue;
-      /* we now assume we have the proper reply */
-      break;
+      /* Accept the response if the packet is sufficiently long and
+	 the transaction ID matches the query (if available).  */
+      if (inlen >= 4
+	  && (xargs == NULL
+	      || memcmp (cu->cu_inbuf, cu->cu_outbuf,
+			 sizeof (u_int32_t)) == 0))
+	break;
+
+    next_response:
+      /* Update the current time because poll and recvmsg waited for
+	 an unknown time.  */
+      current_time = __deadline_current_time ();
     }
 
   /*
diff --git a/sunrpc/tst-udp-garbage.c b/sunrpc/tst-udp-garbage.c
new file mode 100644
index 0000000000..4abda93f08
--- /dev/null
+++ b/sunrpc/tst-udp-garbage.c
@@ -0,0 +1,104 @@
+/* Test that garbage packets do not affect timeout handling.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <netinet/in.h>
+#include <rpc/clnt.h>
+#include <rpc/svc.h>
+#include <stdbool.h>
+#include <support/check.h>
+#include <support/namespace.h>
+#include <support/xsocket.h>
+#include <support/xthread.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+/* Descriptor for the server UDP socket.  */
+static int server_fd;
+
+static void *
+garbage_sender_thread (void *unused)
+{
+  while (true)
+    {
+      struct sockaddr_storage sa;
+      socklen_t salen = sizeof (sa);
+      char buf[1];
+      if (recvfrom (server_fd, buf, sizeof (buf), 0,
+                    (struct sockaddr *) &sa, &salen) < 0)
+        FAIL_EXIT1 ("recvfrom: %m");
+
+      /* Send garbage packets indefinitely.  */
+      buf[0] = 0;
+      while (true)
+        {
+          /* sendto can fail if the client closed the socket.  */
+          if (sendto (server_fd, buf, sizeof (buf), 0,
+                      (struct sockaddr *) &sa, salen) < 0)
+            break;
+
+          /* Wait a bit, to avoid burning too many CPU cycles in a
+             tight loop.  The wait period must be much shorter than
+             the client timeouts configured below.  */
+          usleep (50 * 1000);
+        }
+    }
+}
+
+static int
+do_test (void)
+{
+  support_become_root ();
+  support_enter_network_namespace ();
+
+  server_fd = xsocket (AF_INET, SOCK_DGRAM | SOCK_CLOEXEC, IPPROTO_UDP);
+  struct sockaddr_in server_address =
+    {
+      .sin_family = AF_INET,
+      .sin_addr.s_addr = htonl (INADDR_LOOPBACK),
+    };
+  xbind (server_fd,
+         (struct sockaddr *) &server_address, sizeof (server_address));
+  {
+    socklen_t sinlen = sizeof (server_address);
+    xgetsockname (server_fd, (struct sockaddr *) &server_address, &sinlen);
+    TEST_VERIFY (sizeof (server_address) == sinlen);
+  }
+
+  /* Garbage packet source.  */
+  xpthread_detach (xpthread_create (NULL, garbage_sender_thread, NULL));
+
+  /* Test client.  Use an arbitrary timeout of one second, which is
+     much longer than the garbage packet interval, but still
+     reasonably short, so that the test completes quickly.  */
+  int client_fd = RPC_ANYSOCK;
+  CLIENT *clnt = clntudp_create (&server_address,
+                                 1, 2, /* Arbitrary RPC endpoint numbers.  */
+                                 (struct timeval) { 1, 0 },
+                                 &client_fd);
+  if (clnt == NULL)
+    FAIL_EXIT1 ("clntudp_create: %m");
+
+  TEST_VERIFY (clnt_call (clnt, 3, /* Arbitrary RPC procedure number.  */
+                          (xdrproc_t) xdr_void, NULL,
+                          (xdrproc_t) xdr_void, NULL,
+                          ((struct timeval) { 1, 0 })));
+
+  return 0;
+}
+
+#include <support/test-driver.c>
diff --git a/sunrpc/tst-udp-nonblocking.c b/sunrpc/tst-udp-nonblocking.c
new file mode 100644
index 0000000000..1d6a7f4b56
--- /dev/null
+++ b/sunrpc/tst-udp-nonblocking.c
@@ -0,0 +1,333 @@
+/* Test non-blocking use of the UDP client.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <netinet/in.h>
+#include <rpc/clnt.h>
+#include <rpc/svc.h>
+#include <stdbool.h>
+#include <string.h>
+#include <support/check.h>
+#include <support/namespace.h>
+#include <support/test-driver.h>
+#include <support/xsocket.h>
+#include <support/xunistd.h>
+#include <sys/socket.h>
+#include <time.h>
+#include <unistd.h>
+
+/* Test data serialization and deserialization.   */
+
+struct test_query
+{
+  uint32_t a;
+  uint32_t b;
+  uint32_t timeout_ms;
+};
+
+static bool_t
+xdr_test_query (XDR *xdrs, void *data, ...)
+{
+  struct test_query *p = data;
+  return xdr_uint32_t (xdrs, &p->a)
+    && xdr_uint32_t (xdrs, &p->b)
+    && xdr_uint32_t (xdrs, &p->timeout_ms);
+}
+
+struct test_response
+{
+  uint32_t server_id;
+  uint32_t seq;
+  uint32_t sum;
+};
+
+static bool_t
+xdr_test_response (XDR *xdrs, void *data, ...)
+{
+  struct test_response *p = data;
+  return xdr_uint32_t (xdrs, &p->server_id)
+    && xdr_uint32_t (xdrs, &p->seq)
+    && xdr_uint32_t (xdrs, &p->sum);
+}
+
+/* Implementation of the test server.  */
+
+enum
+  {
+    /* Number of test servers to run. */
+    SERVER_COUNT = 3,
+
+    /* RPC parameters, chosen at random.  */
+    PROGNUM = 8242,
+    VERSNUM = 19654,
+
+    /* Main RPC operation.  */
+    PROC_ADD = 1,
+
+    /* Request process termination.  */
+    PROC_EXIT,
+
+    /* Special exit status to mark successful processing.  */
+    EXIT_MARKER = 55,
+  };
+
+/* Set by the parent process to tell test servers apart.  */
+static int server_id;
+
+/* Implementation of the test server.  */
+static void
+server_dispatch (struct svc_req *request, SVCXPRT *transport)
+{
+  /* Query sequence number.  */
+  static uint32_t seq = 0;
+  ++seq;
+  static bool proc_add_seen;
+
+  if (test_verbose)
+    printf ("info: server_dispatch server_id=%d seq=%u rq_proc=%lu\n",
+            server_id, seq, request->rq_proc);
+
+  switch (request->rq_proc)
+    {
+    case PROC_ADD:
+      {
+        struct test_query query;
+        memset (&query, 0xc0, sizeof (query));
+        TEST_VERIFY_EXIT
+          (svc_getargs (transport, xdr_test_query,
+                        (void *) &query));
+
+        if (test_verbose)
+          printf ("  a=%u b=%u timeout_ms=%u\n",
+                  query.a, query.b, query.timeout_ms);
+
+        usleep (query.timeout_ms * 1000);
+
+        struct test_response response =
+          {
+            .server_id = server_id,
+            .seq = seq,
+            .sum = query.a + query.b,
+          };
+        TEST_VERIFY (svc_sendreply (transport, xdr_test_response,
+                                    (void *) &response));
+        if (test_verbose)
+          printf ("  server id %d response seq=%u sent\n", server_id, seq);
+        proc_add_seen = true;
+      }
+      break;
+
+    case PROC_EXIT:
+      TEST_VERIFY (proc_add_seen);
+      TEST_VERIFY (svc_sendreply (transport, (xdrproc_t) xdr_void, NULL));
+      _exit (EXIT_MARKER);
+      break;
+
+    default:
+      FAIL_EXIT1 ("invalid rq_proc value: %lu", request->rq_proc);
+      break;
+    }
+}
+
+/* Return the number seconds since an arbitrary point in time.  */
+static double
+get_ticks (void)
+{
+  {
+    struct timespec ts;
+    if (clock_gettime (CLOCK_MONOTONIC, &ts) == 0)
+      return ts.tv_sec + ts.tv_nsec * 1e-9;
+  }
+  {
+    struct timeval tv;
+    TEST_VERIFY_EXIT (gettimeofday (&tv, NULL) == 0);
+    return tv.tv_sec + tv.tv_usec * 1e-6;
+  }
+}
+
+static int
+do_test (void)
+{
+  support_become_root ();
+  support_enter_network_namespace ();
+
+  /* Information about the test servers.  */
+  struct
+  {
+    SVCXPRT *transport;
+    struct sockaddr_in address;
+    pid_t pid;
+    uint32_t xid;
+  } servers[SERVER_COUNT];
+
+  /* Spawn the test servers.  */
+  for (int i = 0; i < SERVER_COUNT; ++i)
+    {
+      servers[i].transport = svcudp_create (RPC_ANYSOCK);
+      TEST_VERIFY_EXIT (servers[i].transport != NULL);
+      servers[i].address = (struct sockaddr_in)
+        {
+          .sin_family = AF_INET,
+          .sin_addr.s_addr = htonl (INADDR_LOOPBACK),
+          .sin_port = htons (servers[i].transport->xp_port),
+        };
+      servers[i].xid = 0xabcd0101 + i;
+      if (test_verbose)
+        printf ("info: setting up server %d xid=%x on port %d\n",
+                i, servers[i].xid, servers[i].transport->xp_port);
+
+      server_id = i;
+      servers[i].pid = xfork ();
+      if (servers[i].pid == 0)
+        {
+          TEST_VERIFY (svc_register (servers[i].transport,
+                                     PROGNUM, VERSNUM, server_dispatch, 0));
+          svc_run ();
+          FAIL_EXIT1 ("supposed to be unreachable");
+        }
+      /* We need to close the socket so that we do not accidentally
+         consume the request.  */
+      TEST_VERIFY (close (servers[i].transport->xp_sock) == 0);
+    }
+
+
+  /* The following code mirrors what ypbind does.  */
+
+  /* Copied from clnt_udp.c (like ypbind).  */
+  struct cu_data
+  {
+    int cu_sock;
+    bool_t cu_closeit;
+    struct sockaddr_in cu_raddr;
+    int cu_rlen;
+    struct timeval cu_wait;
+    struct timeval cu_total;
+    struct rpc_err cu_error;
+    XDR cu_outxdrs;
+    u_int cu_xdrpos;
+    u_int cu_sendsz;
+    char *cu_outbuf;
+    u_int cu_recvsz;
+    char cu_inbuf[1];
+  };
+
+  int client_socket = xsocket (AF_INET, SOCK_DGRAM | SOCK_NONBLOCK, 0);
+  CLIENT *clnt = clntudp_create (&servers[0].address, PROGNUM, VERSNUM,
+                                 /* 5 seconds per-response timeout.  */
+                                 ((struct timeval) { 5, 0 }),
+                                 &client_socket);
+  TEST_VERIFY (clnt != NULL);
+  clnt->cl_auth = authunix_create_default ();
+  {
+    struct timeval zero = { 0, 0 };
+    TEST_VERIFY (clnt_control (clnt, CLSET_TIMEOUT, (void *) &zero));
+  }
+
+  /* Poke at internal data structures (like ypbind).  */
+  struct cu_data *cu = (struct cu_data *) clnt->cl_private;
+
+  /* Send a ping to each server.  */
+  double before_pings = get_ticks ();
+  for (int i = 0; i < SERVER_COUNT; ++i)
+    {
+      if (test_verbose)
+        printf ("info: sending server %d ping\n", i);
+      /* Reset the xid because it is changed by each invocation of
+         clnt_call.  Subtract one to compensate for the xid update
+         during the call.  */
+      *((u_int32_t *) (cu->cu_outbuf)) = servers[i].xid - 1;
+      cu->cu_raddr = servers[i].address;
+
+      struct test_query query = { .a = 100, .b = i + 1 };
+      if (i == 1)
+        /* Shorter timeout to prefer this server.  These timeouts must
+           be much shorter than the 5-second per-response timeout
+           configured with clntudp_create.  */
+        query.timeout_ms = 700;
+      else
+        query.timeout_ms = 1400;
+      struct test_response response = { 0 };
+      /* NB: Do not check the return value.  The server reply will
+         prove that the call worked.  */
+      double before_one_ping = get_ticks ();
+      clnt_call (clnt, PROC_ADD,
+                 xdr_test_query, (void *) &query,
+                 xdr_test_response, (void *) &response,
+                 ((struct timeval) { 0, 0 }));
+      double after_one_ping = get_ticks ();
+      if (test_verbose)
+        printf ("info: non-blocking send took %f seconds\n",
+                after_one_ping - before_one_ping);
+      /* clnt_call should return immediately.  Accept some delay in
+         case the process is descheduled.  */
+      TEST_VERIFY (after_one_ping - before_one_ping < 0.3);
+    }
+
+  /* Collect the non-blocking response.  */
+  if (test_verbose)
+    printf ("info: collecting response\n");
+  struct test_response response = { 0 };
+  TEST_VERIFY
+    (clnt_call (clnt, PROC_ADD, NULL, NULL,
+                xdr_test_response, (void *) &response,
+                ((struct timeval) { 0, 0 })) == RPC_SUCCESS);
+  double after_pings = get_ticks ();
+  if (test_verbose)
+    printf ("info: send/receive took %f seconds\n",
+            after_pings - before_pings);
+  /* Expected timeout is 0.7 seconds.  */
+  TEST_VERIFY (0.7 <= after_pings - before_pings);
+  TEST_VERIFY (after_pings - before_pings < 1.2);
+
+  uint32_t xid;
+  memcpy (&xid, &cu->cu_inbuf, sizeof (xid));
+  if (test_verbose)
+    printf ("info: non-blocking response: xid=%x server_id=%u seq=%u sum=%u\n",
+            xid, response.server_id, response.seq, response.sum);
+  /* Check that the reply from the preferred server was used.  */
+  TEST_VERIFY (servers[1].xid == xid);
+  TEST_VERIFY (response.server_id == 1);
+  TEST_VERIFY (response.seq == 1);
+  TEST_VERIFY (response.sum == 102);
+
+  auth_destroy (clnt->cl_auth);
+  clnt_destroy (clnt);
+
+  for (int i = 0; i < SERVER_COUNT; ++i)
+    {
+      if (test_verbose)
+        printf ("info: requesting server %d termination\n", i);
+      client_socket = RPC_ANYSOCK;
+      clnt = clntudp_create (&servers[i].address, PROGNUM, VERSNUM,
+                             ((struct timeval) { 5, 0 }),
+                             &client_socket);
+      TEST_VERIFY_EXIT (clnt != NULL);
+      TEST_VERIFY (clnt_call (clnt, PROC_EXIT,
+                              (xdrproc_t) xdr_void, NULL,
+                              (xdrproc_t) xdr_void, NULL,
+                              ((struct timeval) { 3, 0 })) == RPC_SUCCESS);
+      clnt_destroy (clnt);
+
+      int status;
+      xwaitpid (servers[i].pid, &status, 0);
+      TEST_VERIFY (WIFEXITED (status) && WEXITSTATUS (status) == EXIT_MARKER);
+    }
+
+  return 0;
+}
+
+#include <support/test-driver.c>
diff --git a/sunrpc/tst-udp-timeout.c b/sunrpc/tst-udp-timeout.c
new file mode 100644
index 0000000000..db9943a03e
--- /dev/null
+++ b/sunrpc/tst-udp-timeout.c
@@ -0,0 +1,402 @@
+/* Test timeout handling in the UDP client.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <netinet/in.h>
+#include <rpc/clnt.h>
+#include <rpc/svc.h>
+#include <stdbool.h>
+#include <string.h>
+#include <support/check.h>
+#include <support/namespace.h>
+#include <support/test-driver.h>
+#include <support/xsocket.h>
+#include <support/xunistd.h>
+#include <sys/socket.h>
+#include <time.h>
+#include <unistd.h>
+
+/* Test data serialization and deserialization.   */
+
+struct test_query
+{
+  uint32_t a;
+  uint32_t b;
+  uint32_t timeout_ms;
+  uint32_t wait_for_seq;
+  uint32_t garbage_packets;
+};
+
+static bool_t
+xdr_test_query (XDR *xdrs, void *data, ...)
+{
+  struct test_query *p = data;
+  return xdr_uint32_t (xdrs, &p->a)
+    && xdr_uint32_t (xdrs, &p->b)
+    && xdr_uint32_t (xdrs, &p->timeout_ms)
+    && xdr_uint32_t (xdrs, &p->wait_for_seq)
+    && xdr_uint32_t (xdrs, &p->garbage_packets);
+}
+
+struct test_response
+{
+  uint32_t seq;
+  uint32_t sum;
+};
+
+static bool_t
+xdr_test_response (XDR *xdrs, void *data, ...)
+{
+  struct test_response *p = data;
+  return xdr_uint32_t (xdrs, &p->seq)
+    && xdr_uint32_t (xdrs, &p->sum);
+}
+
+/* Implementation of the test server.  */
+
+enum
+  {
+    /* RPC parameters, chosen at random.  */
+    PROGNUM = 15717,
+    VERSNUM = 13689,
+
+    /* Main RPC operation.  */
+    PROC_ADD = 1,
+
+    /* Reset the sequence number.  */
+    PROC_RESET_SEQ,
+
+    /* Request process termination.  */
+    PROC_EXIT,
+
+    /* Special exit status to mark successful processing.  */
+    EXIT_MARKER = 55,
+  };
+
+static void
+server_dispatch (struct svc_req *request, SVCXPRT *transport)
+{
+  /* Query sequence number.  */
+  static uint32_t seq = 0;
+  ++seq;
+
+  if (test_verbose)
+    printf ("info: server_dispatch seq=%u rq_proc=%lu\n",
+            seq, request->rq_proc);
+
+  switch (request->rq_proc)
+    {
+    case PROC_ADD:
+      {
+        struct test_query query;
+        memset (&query, 0xc0, sizeof (query));
+        TEST_VERIFY_EXIT
+          (svc_getargs (transport, xdr_test_query,
+                        (void *) &query));
+
+        if (test_verbose)
+          printf ("  a=%u b=%u timeout_ms=%u wait_for_seq=%u"
+                  " garbage_packets=%u\n",
+                  query.a, query.b, query.timeout_ms, query.wait_for_seq,
+                  query.garbage_packets);
+
+        if (seq < query.wait_for_seq)
+          {
+            /* No response at this point.  */
+            if (test_verbose)
+              printf ("  skipped response\n");
+            break;
+          }
+
+        if (query.garbage_packets > 0)
+          {
+            int per_packet_timeout;
+            if (query.timeout_ms > 0)
+              per_packet_timeout
+                = query.timeout_ms * 1000 / query.garbage_packets;
+            else
+              per_packet_timeout = 0;
+
+            char buf[20];
+            memset (&buf, 0xc0, sizeof (buf));
+            for (int i = 0; i < query.garbage_packets; ++i)
+              {
+                /* 13 is relatively prime to 20 = sizeof (buf) + 1, so
+                   the len variable will cover the entire interval
+                   [0, 20] if query.garbage_packets is sufficiently
+                   large.  */
+                size_t len = (i * 13 + 1) % (sizeof (buf) + 1);
+                TEST_VERIFY (sendto (transport->xp_sock,
+                                     buf, len, MSG_NOSIGNAL,
+                                     (struct sockaddr *) &transport->xp_raddr,
+                                     transport->xp_addrlen) == len);
+                if (per_packet_timeout > 0)
+                  usleep (per_packet_timeout);
+              }
+          }
+        else if (query.timeout_ms > 0)
+          usleep (query.timeout_ms * 1000);
+
+        struct test_response response =
+          {
+            .seq = seq,
+            .sum = query.a + query.b,
+          };
+        TEST_VERIFY (svc_sendreply (transport, xdr_test_response,
+                                    (void *) &response));
+      }
+      break;
+
+    case PROC_RESET_SEQ:
+      seq = 0;
+      TEST_VERIFY (svc_sendreply (transport, (xdrproc_t) xdr_void, NULL));
+      break;
+
+    case PROC_EXIT:
+      TEST_VERIFY (svc_sendreply (transport, (xdrproc_t) xdr_void, NULL));
+      _exit (EXIT_MARKER);
+      break;
+
+    default:
+      FAIL_EXIT1 ("invalid rq_proc value: %lu", request->rq_proc);
+      break;
+    }
+}
+
+/* Implementation of the test client.  */
+
+static struct test_response
+test_call (CLIENT *clnt, int proc, struct test_query query,
+           struct timeval timeout)
+{
+  if (test_verbose)
+    printf ("info: test_call proc=%d timeout=%lu.%06lu\n",
+            proc, (unsigned long) timeout.tv_sec,
+            (unsigned long) timeout.tv_usec);
+  struct test_response response;
+  TEST_VERIFY_EXIT (clnt_call (clnt, proc,
+                               xdr_test_query, (void *) &query,
+                               xdr_test_response, (void *) &response,
+                               timeout)
+                    == RPC_SUCCESS);
+  return response;
+}
+
+static void
+test_call_timeout (CLIENT *clnt, int proc, struct test_query query,
+                   struct timeval timeout)
+{
+  struct test_response response;
+  TEST_VERIFY (clnt_call (clnt, proc,
+                          xdr_test_query, (void *) &query,
+                          xdr_test_response, (void *) &response,
+                          timeout)
+               == RPC_TIMEDOUT);
+}
+
+/* Complete one regular RPC call to drain the server socket
+   buffer.  Resets the sequence number.  */
+static void
+test_call_flush (CLIENT *clnt)
+{
+  /* This needs a longer timeout to flush out all pending requests.
+     The choice of 5 seconds is larger than the per-response timeouts
+     requested via the timeout_ms field.  */
+  if (test_verbose)
+    printf ("info: flushing pending queries\n");
+  TEST_VERIFY_EXIT (clnt_call (clnt, PROC_RESET_SEQ,
+                               (xdrproc_t) xdr_void, NULL,
+                               (xdrproc_t) xdr_void, NULL,
+                               ((struct timeval) { 5, 0 }))
+                    == RPC_SUCCESS);
+}
+
+/* Return the number seconds since an arbitrary point in time.  */
+static double
+get_ticks (void)
+{
+  {
+    struct timespec ts;
+    if (clock_gettime (CLOCK_MONOTONIC, &ts) == 0)
+      return ts.tv_sec + ts.tv_nsec * 1e-9;
+  }
+  {
+    struct timeval tv;
+    TEST_VERIFY_EXIT (gettimeofday (&tv, NULL) == 0);
+    return tv.tv_sec + tv.tv_usec * 1e-6;
+  }
+}
+
+static void
+test_udp_server (int port)
+{
+  struct sockaddr_in sin =
+    {
+      .sin_family = AF_INET,
+      .sin_addr.s_addr = htonl (INADDR_LOOPBACK),
+      .sin_port = htons (port)
+    };
+  int sock = RPC_ANYSOCK;
+
+  /* The client uses a 1.5 second timeout for retries.  The timeouts
+     are arbitrary, but chosen so that there is a substantial gap
+     between them, but the total time spent waiting is not too
+     large.  */
+  CLIENT *clnt = clntudp_create (&sin, PROGNUM, VERSNUM,
+                                 (struct timeval) { 1, 500 * 1000 },
+                                 &sock);
+  TEST_VERIFY_EXIT (clnt != NULL);
+
+  /* Basic call/response test.  */
+  struct test_response response = test_call
+    (clnt, PROC_ADD,
+     (struct test_query) { .a = 17, .b = 4 },
+     (struct timeval) { 3, 0 });
+  TEST_VERIFY (response.sum == 21);
+  TEST_VERIFY (response.seq == 1);
+
+  /* Check that garbage packets do not interfere with timeout
+     processing.  */
+  double before = get_ticks ();
+  response = test_call
+    (clnt, PROC_ADD,
+     (struct test_query) {
+       .a = 19, .b = 4, .timeout_ms = 500, .garbage_packets = 21,
+     },
+     (struct timeval) { 3, 0 });
+  TEST_VERIFY (response.sum == 23);
+  TEST_VERIFY (response.seq == 2);
+  double after = get_ticks ();
+  if (test_verbose)
+    printf ("info: 21 garbage packets took %f seconds\n", after - before);
+  /* Expected timeout is 0.5 seconds.  Add some slack in case process
+     scheduling delays processing the query or response, but do not
+     accept a retry (which would happen at 1.5 seconds).  */
+  TEST_VERIFY (0.5 <= after - before);
+  TEST_VERIFY (after - before < 1.2);
+  test_call_flush (clnt);
+
+  /* Check that missing a response introduces a 1.5 second timeout, as
+     requested when calling clntudp_create.  */
+  before = get_ticks ();
+  response = test_call
+    (clnt, PROC_ADD,
+     (struct test_query) { .a = 170, .b = 40, .wait_for_seq = 2 },
+     (struct timeval) { 3, 0 });
+  TEST_VERIFY (response.sum == 210);
+  TEST_VERIFY (response.seq == 2);
+  after = get_ticks ();
+  if (test_verbose)
+    printf ("info: skipping one response took %f seconds\n",
+            after - before);
+  /* Expected timeout is 1.5 seconds.  Do not accept a second retry
+     (which would happen at 3 seconds).  */
+  TEST_VERIFY (1.5 <= after - before);
+  TEST_VERIFY (after - before < 2.9);
+  test_call_flush (clnt);
+
+  /* Check that the overall timeout wins against the per-query
+     timeout.  */
+  before = get_ticks ();
+  test_call_timeout
+    (clnt, PROC_ADD,
+     (struct test_query) { .a = 170, .b = 41, .wait_for_seq = 2 },
+     (struct timeval) { 0, 750 * 1000 });
+  after = get_ticks ();
+  if (test_verbose)
+    printf ("info: 0.75 second timeout took %f seconds\n",
+            after - before);
+  TEST_VERIFY (0.75 <= after - before);
+  TEST_VERIFY (after - before < 1.4);
+  test_call_flush (clnt);
+
+  for (int with_garbage = 0; with_garbage < 2; ++with_garbage)
+    {
+      /* Check that no response at all causes the client to bail out.  */
+      before = get_ticks ();
+      test_call_timeout
+        (clnt, PROC_ADD,
+         (struct test_query) {
+           .a = 170, .b = 40, .timeout_ms = 1200,
+           .garbage_packets = with_garbage * 21
+         },
+         (struct timeval) { 0, 750 * 1000 });
+      after = get_ticks ();
+      if (test_verbose)
+        printf ("info: test_udp_server: 0.75 second timeout took %f seconds"
+                " (garbage %d)\n",
+                after - before, with_garbage);
+      TEST_VERIFY (0.75 <= after - before);
+      TEST_VERIFY (after - before < 1.4);
+      test_call_flush (clnt);
+
+      /* As above, but check the total timeout.  */
+      before = get_ticks ();
+      test_call_timeout
+        (clnt, PROC_ADD,
+         (struct test_query) {
+           .a = 170, .b = 40, .timeout_ms = 3000,
+           .garbage_packets = with_garbage * 30
+         },
+         (struct timeval) { 2, 300 * 1000 });
+      after = get_ticks ();
+      if (test_verbose)
+        printf ("info: test_udp_server: 2.3 second timeout took %f seconds"
+                " (garbage %d)\n",
+                after - before, with_garbage);
+      TEST_VERIFY (2.3 <= after - before);
+      TEST_VERIFY (after - before < 3.0);
+      test_call_flush (clnt);
+    }
+
+  TEST_VERIFY_EXIT (clnt_call (clnt, PROC_EXIT,
+                               (xdrproc_t) xdr_void, NULL,
+                               (xdrproc_t) xdr_void, NULL,
+                               ((struct timeval) { 5, 0 }))
+                    == RPC_SUCCESS);
+  clnt_destroy (clnt);
+}
+
+static int
+do_test (void)
+{
+  support_become_root ();
+  support_enter_network_namespace ();
+
+  SVCXPRT *transport = svcudp_create (RPC_ANYSOCK);
+  TEST_VERIFY_EXIT (transport != NULL);
+  TEST_VERIFY (svc_register (transport, PROGNUM, VERSNUM, server_dispatch, 0));
+
+  pid_t pid = xfork ();
+  if (pid == 0)
+    {
+      svc_run ();
+      FAIL_EXIT1 ("supposed to be unreachable");
+    }
+  test_udp_server (transport->xp_port);
+
+  int status;
+  xwaitpid (pid, &status, 0);
+  TEST_VERIFY (WIFEXITED (status) && WEXITSTATUS (status) == EXIT_MARKER);
+
+  SVC_DESTROY (transport);
+  return 0;
+}
+
+/* The minimum run time is around 17 seconds.  */
+#define TIMEOUT 25
+#include <support/test-driver.c>