Fix assorted race conditions in the new timeout infrastructure.

tglsfdc · tglsfdc · commit 16e1b7a1b7f7 · 2013-11-29T16:41:00.000-05:00
Prevent handle_sig_alarm from losing control partway through due to a query
cancel (either an asynchronous SIGINT, or a cancel triggered by one of the
timeout handler functions).  That would at least result in failure to
schedule any required future interrupt, and might result in actual
corruption of timeout.c's data structures, if the interrupt happened while
we were updating those.

We could still lose control if an asynchronous SIGINT arrives just as the
function is entered.  This wouldn't break any data structures, but it would
have the same effect as if the SIGALRM interrupt had been silently lost:
we'd not fire any currently-due handlers, nor schedule any new interrupt.
To forestall that scenario, forcibly reschedule any pending timer interrupt
during AbortTransaction and AbortSubTransaction.  We can avoid any extra
kernel call in most cases by not doing that until we've allowed
LockErrorCleanup to kill the DEADLOCK_TIMEOUT and LOCK_TIMEOUT events.

Another hazard is that some platforms (at least Linux and *BSD) block a
signal before calling its handler and then unblock it on return.  When we
longjmp out of the handler, the unblock doesn't happen, and the signal is
left blocked indefinitely.  Again, we can fix that by forcibly unblocking
signals during AbortTransaction and AbortSubTransaction.

These latter two problems do not manifest when the longjmp reaches
postgres.c, because the error recovery code there kills all pending timeout
events anyway, and it uses sigsetjmp(..., 1) so that the appropriate signal
mask is restored.  So errors thrown outside any transaction should be OK
already, and cleaning up in AbortTransaction and AbortSubTransaction should
be enough to fix these issues.  (We're assuming that any code that catches
a query cancel error and doesn't re-throw it will do at least a
subtransaction abort to clean up; but that was pretty much required already
by other subsystems.)

Lastly, ProcSleep should not clear the LOCK_TIMEOUT indicator flag when
disabling that event: if a lock timeout interrupt happened after the lock
was granted, the ensuing query cancel is still going to happen at the next
CHECK_FOR_INTERRUPTS, and we want to report it as a lock timeout not a user
cancel.

Per reports from Dan Wood.

Back-patch to 9.3 where the new timeout handling infrastructure was
introduced.  We may at some point decide to back-patch the signal
unblocking changes further, but I'll desist from that until we hear
actual field complaints about it.
diff --git a/src/backend/access/transam/xact.c b/src/backend/access/transam/xact.c
@@ -34,6 +34,7 @@
 #include "commands/trigger.h"
 #include "executor/spi.h"
 #include "libpq/be-fsstubs.h"
+#include "libpq/pqsignal.h"
 #include "miscadmin.h"
 #include "pgstat.h"
 #include "replication/walsender.h"
@@ -52,6 +53,7 @@
 #include "utils/memutils.h"
 #include "utils/relmapper.h"
 #include "utils/snapmgr.h"
+#include "utils/timeout.h"
 #include "utils/timestamp.h"
 #include "pg_trace.h"
 
@@ -2296,6 +2298,22 @@ AbortTransaction(void)
 	 */
 	LockErrorCleanup();
 
+	/*
+	 * If any timeout events are still active, make sure the timeout interrupt
+	 * is scheduled.  This covers possible loss of a timeout interrupt due to
+	 * longjmp'ing out of the SIGINT handler (see notes in handle_sig_alarm).
+	 * We delay this till after LockErrorCleanup so that we don't uselessly
+	 * reschedule lock or deadlock check timeouts.
+	 */
+	reschedule_timeouts();
+
+	/*
+	 * Re-enable signals, in case we got here by longjmp'ing out of a signal
+	 * handler.  We do this fairly early in the sequence so that the timeout
+	 * infrastructure will be functional if needed while aborting.
+	 */
+	PG_SETMASK(&UnBlockSig);
+
 	/*
 	 * check the current transaction state
 	 */
@@ -4222,8 +4240,28 @@ AbortSubTransaction(void)
 	AbortBufferIO();
 	UnlockBuffers();
 
+	/*
+	 * Also clean up any open wait for lock, since the lock manager will choke
+	 * if we try to wait for another lock before doing this.
+	 */
 	LockErrorCleanup();
 
+	/*
+	 * If any timeout events are still active, make sure the timeout interrupt
+	 * is scheduled.  This covers possible loss of a timeout interrupt due to
+	 * longjmp'ing out of the SIGINT handler (see notes in handle_sig_alarm).
+	 * We delay this till after LockErrorCleanup so that we don't uselessly
+	 * reschedule lock or deadlock check timeouts.
+	 */
+	reschedule_timeouts();
+
+	/*
+	 * Re-enable signals, in case we got here by longjmp'ing out of a signal
+	 * handler.  We do this fairly early in the sequence so that the timeout
+	 * infrastructure will be functional if needed while aborting.
+	 */
+	PG_SETMASK(&UnBlockSig);
+
 	/*
 	 * check the current transaction state
 	 */
diff --git a/src/backend/postmaster/autovacuum.c b/src/backend/postmaster/autovacuum.c
@@ -493,9 +493,8 @@ AutoVacLauncherMain(int argc, char *argv[])
 		HOLD_INTERRUPTS();
 
 		/* Forget any pending QueryCancel or timeout request */
-		QueryCancelPending = false;
 		disable_all_timeouts(false);
-		QueryCancelPending = false;		/* again in case timeout occurred */
+		QueryCancelPending = false;		/* second to avoid race condition */
 
 		/* Report the error to the server log */
 		EmitErrorReport();
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
@@ -1266,7 +1266,10 @@ ProcSleep(LOCALLOCK *locallock, LockMethod lockMethodTable)
 	} while (myWaitStatus == STATUS_WAITING);
 
 	/*
-	 * Disable the timers, if they are still running
+	 * Disable the timers, if they are still running.  As in LockErrorCleanup,
+	 * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
+	 * already caused QueryCancelPending to become set, we want the cancel to
+	 * be reported as a lock timeout, not a user cancel.
 	 */
 	if (LockTimeout > 0)
 	{
@@ -1275,7 +1278,7 @@ ProcSleep(LOCALLOCK *locallock, LockMethod lockMethodTable)
 		timeouts[0].id = DEADLOCK_TIMEOUT;
 		timeouts[0].keep_indicator = false;
 		timeouts[1].id = LOCK_TIMEOUT;
-		timeouts[1].keep_indicator = false;
+		timeouts[1].keep_indicator = true;
 		disable_timeouts(timeouts, 2);
 	}
 	else
diff --git a/src/backend/tcop/postgres.c b/src/backend/tcop/postgres.c
@@ -3803,6 +3803,13 @@ PostgresMain(int argc, char *argv[],
 	 * always active, we have at least some chance of recovering from an error
 	 * during error recovery.  (If we get into an infinite loop thereby, it
 	 * will soon be stopped by overflow of elog.c's internal state stack.)
+	 *
+	 * Note that we use sigsetjmp(..., 1), so that this function's signal mask
+	 * (to wit, UnBlockSig) will be restored when longjmp'ing to here.  This
+	 * is essential in case we longjmp'd out of a signal handler on a platform
+	 * where that leaves the signal blocked.  It's not redundant with the
+	 * unblock in AbortTransaction() because the latter is only called if we
+	 * were inside a transaction.
 	 */
 
 	if (sigsetjmp(local_sigjmp_buf, 1) != 0)
@@ -3823,11 +3830,17 @@ PostgresMain(int argc, char *argv[],
 
 		/*
 		 * Forget any pending QueryCancel request, since we're returning to
-		 * the idle loop anyway, and cancel any active timeout requests.
+		 * the idle loop anyway, and cancel any active timeout requests.  (In
+		 * future we might want to allow some timeout requests to survive, but
+		 * at minimum it'd be necessary to do reschedule_timeouts(), in case
+		 * we got here because of a query cancel interrupting the SIGALRM
+		 * interrupt handler.)	Note in particular that we must clear the
+		 * statement and lock timeout indicators, to prevent any future plain
+		 * query cancels from being misreported as timeouts in case we're
+		 * forgetting a timeout cancel.
 		 */
-		QueryCancelPending = false;
 		disable_all_timeouts(false);
-		QueryCancelPending = false;		/* again in case timeout occurred */
+		QueryCancelPending = false;		/* second to avoid race condition */
 
 		/*
 		 * Turn off these interrupts too.  This is only needed here and not in
diff --git a/src/backend/utils/misc/timeout.c b/src/backend/utils/misc/timeout.c
@@ -16,6 +16,7 @@
 
 #include <sys/time.h>
 
+#include "miscadmin.h"
 #include "storage/proc.h"
 #include "utils/timeout.h"
 #include "utils/timestamp.h"
@@ -259,6 +260,23 @@ handle_sig_alarm(SIGNAL_ARGS)
 {
 	int			save_errno = errno;
 
+	/*
+	 * We may be executing while ImmediateInterruptOK is true (e.g., when
+	 * mainline is waiting for a lock).  If SIGINT or similar arrives while
+	 * this code is running, we'd lose control and perhaps leave our data
+	 * structures in an inconsistent state.  Hold off interrupts to prevent
+	 * that.
+	 *
+	 * Note: it's possible for a SIGINT to interrupt handle_sig_alarm even
+	 * before we reach HOLD_INTERRUPTS(); the net effect would be as if the
+	 * SIGALRM event had been silently lost.  Therefore error recovery must
+	 * include some action that will allow any lost interrupt to be
+	 * rescheduled.  Disabling some or all timeouts is sufficient, or if
+	 * that's not appropriate, reschedule_timeouts() can be called.  Also, the
+	 * signal blocking hazard described below applies here too.
+	 */
+	HOLD_INTERRUPTS();
+
 	/*
 	 * SIGALRM is always cause for waking anything waiting on the process
 	 * latch.  Cope with MyProc not being there, as the startup process also
@@ -311,6 +329,20 @@ handle_sig_alarm(SIGNAL_ARGS)
 		}
 	}
 
+	/*
+	 * Re-allow query cancel, and then service any cancel request that arrived
+	 * meanwhile (this might in particular include a cancel request fired by
+	 * one of the timeout handlers).
+	 *
+	 * Note: a longjmp from here is safe so far as our own data structures are
+	 * concerned; but on platforms that block a signal before calling the
+	 * handler and then un-block it on return, longjmping out of the signal
+	 * handler leaves SIGALRM still blocked.  Error cleanup is responsible for
+	 * unblocking any blocked signals.
+	 */
+	RESUME_INTERRUPTS();
+	CHECK_FOR_INTERRUPTS();
+
 	errno = save_errno;
 }
 
@@ -387,6 +419,30 @@ RegisterTimeout(TimeoutId id, timeout_handler_proc handler)
 	return id;
 }
 
+/*
+ * Reschedule any pending SIGALRM interrupt.
+ *
+ * This can be used during error recovery in case query cancel resulted in loss
+ * of a SIGALRM event (due to longjmp'ing out of handle_sig_alarm before it
+ * could do anything).	But note it's not necessary if any of the public
+ * enable_ or disable_timeout functions are called in the same area, since
+ * those all do schedule_alarm() internally if needed.
+ */
+void
+reschedule_timeouts(void)
+{
+	/* For flexibility, allow this to be called before we're initialized. */
+	if (!all_timeouts_initialized)
+		return;
+
+	/* Disable timeout interrupts for safety. */
+	disable_alarm();
+
+	/* Reschedule the interrupt, if any timeouts remain active. */
+	if (num_active_timeouts > 0)
+		schedule_alarm(GetCurrentTimestamp());
+}
+
 /*
  * Enable the specified timeout to fire after the specified delay.
  *
diff --git a/src/include/utils/timeout.h b/src/include/utils/timeout.h
@@ -67,6 +67,7 @@ typedef struct
 /* timeout setup */
 extern void InitializeTimeouts(void);
 extern TimeoutId RegisterTimeout(TimeoutId id, timeout_handler_proc handler);
+extern void reschedule_timeouts(void);
 
 /* timeout operation */
 extern void enable_timeout_after(TimeoutId id, int delay_ms);
