Documentation improvement and minor code cleanups for the latch facility.

Improve the documentation around weak-memory-ordering risks, and do a pass
of general editorialization on the comments in the latch code.  Make the
Windows latch code more like the Unix latch code where feasible; in
particular provide the same Assert checks in both implementations.
Fix poorly-placed WaitLatch call in syncrep.c.

This patch resolves, for the moment, concerns around weak-memory-ordering
bugs in latch-related code: we have documented the restrictions and checked
that existing calls meet them.  In 9.2 I hope that we will install suitable
memory barrier instructions in SetLatch/ResetLatch, so that their callers
don't need to be quite so careful.

src/backend/port/unix_latch.c

@@ -3,60 +3,6 @@
  * unix_latch.c
  *	  Routines for inter-process latches
  *
- * A latch is a boolean variable, with operations that let you to sleep
- * until it is set. A latch can be set from another process, or a signal
- * handler within the same process.
- *
- * The latch interface is a reliable replacement for the common pattern of
- * using pg_usleep() or select() to wait until a signal arrives, where the
- * signal handler sets a global variable. Because on some platforms, an
- * incoming signal doesn't interrupt sleep, and even on platforms where it
- * does there is a race condition if the signal arrives just before
- * entering the sleep, the common pattern must periodically wake up and
- * poll the global variable. pselect() system call was invented to solve
- * the problem, but it is not portable enough. Latches are designed to
- * overcome these limitations, allowing you to sleep without polling and
- * ensuring a quick response to signals from other processes.
- *
- * There are two kinds of latches: local and shared. A local latch is
- * initialized by InitLatch, and can only be set from the same process.
- * A local latch can be used to wait for a signal to arrive, by calling
- * SetLatch in the signal handler. A shared latch resides in shared memory,
- * and must be initialized at postmaster startup by InitSharedLatch. Before
- * a shared latch can be waited on, it must be associated with a process
- * with OwnLatch. Only the process owning the latch can wait on it, but any
- * process can set it.
- *
- * There are three basic operations on a latch:
- *
- * SetLatch		- Sets the latch
- * ResetLatch	- Clears the latch, allowing it to be set again
- * WaitLatch	- Waits for the latch to become set
- *
- * The correct pattern to wait for an event is:
- *
- * for (;;)
- * {
- *	   ResetLatch();
- *	   if (work to do)
- *		   Do Stuff();
- *
- *	   WaitLatch();
- * }
- *
- * It's important to reset the latch *before* checking if there's work to
- * do. Otherwise, if someone sets the latch between the check and the
- * ResetLatch call, you will miss it and Wait will block.
- *
- * To wake up the waiter, you must first set a global flag or something
- * else that the main loop tests in the "if (work to do)" part, and call
- * SetLatch *after* that. SetLatch is designed to return quickly if the
- * latch is already set.
- *
- *
- * Implementation
- * --------------
- *
  * The Unix implementation uses the so-called self-pipe trick to overcome
  * the race condition involved with select() and setting a global flag
  * in the signal handler. When a latch is set and the current process
@@ -65,8 +11,8 @@
  * interrupt select() on all platforms, and even on platforms where it
  * does, a signal that arrives just before the select() call does not
  * prevent the select() from entering sleep. An incoming byte on a pipe
- * however reliably interrupts the sleep, and makes select() to return
- * immediately if the signal arrives just before select() begins.
+ * however reliably interrupts the sleep, and causes select() to return
+ * immediately even if the signal arrives before select() begins.
  *
  * When SetLatch is called from the same process that owns the latch,
  * SetLatch writes the byte directly to the pipe. If it's owned by another
@@ -99,7 +45,7 @@
 /* Are we currently in WaitLatch? The signal handler would like to know. */
 static volatile sig_atomic_t waiting = false;
 
-/* Read and write end of the self-pipe */
+/* Read and write ends of the self-pipe */
 static int	selfpipe_readfd = -1;
 static int	selfpipe_writefd = -1;
 
@@ -115,7 +61,7 @@ static void sendSelfPipeByte(void);
 void
 InitLatch(volatile Latch *latch)
 {
-	/* Initialize the self pipe if this is our first latch in the process */
+	/* Initialize the self-pipe if this is our first latch in the process */
 	if (selfpipe_readfd == -1)
 		initSelfPipe();
 
@@ -126,13 +72,14 @@ InitLatch(volatile Latch *latch)
 
 /*
  * Initialize a shared latch that can be set from other processes. The latch
- * is initially owned by no-one, use OwnLatch to associate it with the
+ * is initially owned by no-one; use OwnLatch to associate it with the
  * current process.
  *
  * InitSharedLatch needs to be called in postmaster before forking child
  * processes, usually right after allocating the shared memory block
- * containing the latch with ShmemInitStruct. The Unix implementation
- * doesn't actually require that, but the Windows one does.
+ * containing the latch with ShmemInitStruct. (The Unix implementation
+ * doesn't actually require that, but the Windows one does.) Because of
+ * this restriction, we have no concurrency issues to worry about here.
  */
 void
 InitSharedLatch(volatile Latch *latch)
@@ -144,23 +91,30 @@ InitSharedLatch(volatile Latch *latch)
 
 /*
  * Associate a shared latch with the current process, allowing it to
- * wait on it.
+ * wait on the latch.
  *
- * Make sure that latch_sigusr1_handler() is called from the SIGUSR1 signal
- * handler, as shared latches use SIGUSR1 to for inter-process communication.
+ * Although there is a sanity check for latch-already-owned, we don't do
+ * any sort of locking here, meaning that we could fail to detect the error
+ * if two processes try to own the same latch at about the same time.  If
+ * there is any risk of that, caller must provide an interlock to prevent it.
+ *
+ * In any process that calls OwnLatch(), make sure that
+ * latch_sigusr1_handler() is called from the SIGUSR1 signal handler,
+ * as shared latches use SIGUSR1 for inter-process communication.
  */
 void
 OwnLatch(volatile Latch *latch)
 {
 	Assert(latch->is_shared);
 
-	/* Initialize the self pipe if this is our first latch in the process */
+	/* Initialize the self-pipe if this is our first latch in this process */
 	if (selfpipe_readfd == -1)
 		initSelfPipe();
 
 	/* sanity check */
 	if (latch->owner_pid != 0)
 		elog(ERROR, "latch already owned");
+
 	latch->owner_pid = MyProcPid;
 }
 
@@ -172,6 +126,7 @@ DisownLatch(volatile Latch *latch)
 {
 	Assert(latch->is_shared);
 	Assert(latch->owner_pid == MyProcPid);
+
 	latch->owner_pid = 0;
 }
 
@@ -229,21 +184,31 @@ WaitLatchOrSocket(volatile Latch *latch, pgsocket sock, bool forRead,
 		int			hifd;
 
 		/*
-		 * Clear the pipe, and check if the latch is set already. If someone
+		 * Clear the pipe, then check if the latch is set already. If someone
 		 * sets the latch between this and the select() below, the setter will
 		 * write a byte to the pipe (or signal us and the signal handler will
 		 * do that), and the select() will return immediately.
+		 *
+		 * Note: we assume that the kernel calls involved in drainSelfPipe()
+		 * and SetLatch() will provide adequate synchronization on machines
+		 * with weak memory ordering, so that we cannot miss seeing is_set
+		 * if the signal byte is already in the pipe when we drain it.
 		 */
 		drainSelfPipe();
+
 		if (latch->is_set)
 		{
 			result = 1;
 			break;
 		}
 
+		/* Must wait ... set up the event masks for select() */
 		FD_ZERO(&input_mask);
+		FD_ZERO(&output_mask);
+
 		FD_SET(selfpipe_readfd, &input_mask);
 		hifd = selfpipe_readfd;
+
 		if (sock != PGINVALID_SOCKET && forRead)
 		{
 			FD_SET(sock, &input_mask);
@@ -251,7 +216,6 @@ WaitLatchOrSocket(volatile Latch *latch, pgsocket sock, bool forRead,
 				hifd = sock;
 		}
 
-		FD_ZERO(&output_mask);
 		if (sock != PGINVALID_SOCKET && forWrite)
 		{
 			FD_SET(sock, &output_mask);
@@ -288,14 +252,23 @@ WaitLatchOrSocket(volatile Latch *latch, pgsocket sock, bool forRead,
 }
 
 /*
- * Sets a latch and wakes up anyone waiting on it. Returns quickly if the
- * latch is already set.
+ * Sets a latch and wakes up anyone waiting on it.
+ *
+ * This is cheap if the latch is already set, otherwise not so much.
  */
 void
 SetLatch(volatile Latch *latch)
 {
 	pid_t		owner_pid;
 
+	/*
+	 * XXX there really ought to be a memory barrier operation right here,
+	 * to ensure that any flag variables we might have changed get flushed
+	 * to main memory before we check/set is_set.  Without that, we have to
+	 * require that callers provide their own synchronization for machines
+	 * with weak memory ordering (see latch.h).
+	 */
+
 	/* Quick exit if already set */
 	if (latch->is_set)
 		return;
@@ -307,13 +280,21 @@ SetLatch(volatile Latch *latch)
 	 * we're in a signal handler. We use the self-pipe to wake up the select()
 	 * in that case. If it's another process, send a signal.
 	 *
-	 * Fetch owner_pid only once, in case the owner simultaneously disowns the
-	 * latch and clears owner_pid. XXX: This assumes that pid_t is atomic,
-	 * which isn't guaranteed to be true! In practice, the effective range of
-	 * pid_t fits in a 32 bit integer, and so should be atomic. In the worst
-	 * case, we might end up signaling wrong process if the right one disowns
-	 * the latch just as we fetch owner_pid. Even then, you're very unlucky if
-	 * a process with that bogus pid exists.
+	 * Fetch owner_pid only once, in case the latch is concurrently getting
+	 * owned or disowned. XXX: This assumes that pid_t is atomic, which isn't
+	 * guaranteed to be true! In practice, the effective range of pid_t fits
+	 * in a 32 bit integer, and so should be atomic. In the worst case, we
+	 * might end up signaling the wrong process. Even then, you're very
+	 * unlucky if a process with that bogus pid exists and belongs to
+	 * Postgres; and PG database processes should handle excess SIGUSR1
+	 * interrupts without a problem anyhow.
+	 *
+	 * Another sort of race condition that's possible here is for a new process
+	 * to own the latch immediately after we look, so we don't signal it.
+	 * This is okay so long as all callers of ResetLatch/WaitLatch follow the
+	 * standard coding convention of waiting at the bottom of their loops,
+	 * not the top, so that they'll correctly process latch-setting events that
+	 * happen before they enter the loop.
 	 */
 	owner_pid = latch->owner_pid;
 	if (owner_pid == 0)
@@ -335,11 +316,23 @@ ResetLatch(volatile Latch *latch)
 	Assert(latch->owner_pid == MyProcPid);
 
 	latch->is_set = false;
+
+	/*
+	 * XXX there really ought to be a memory barrier operation right here, to
+	 * ensure that the write to is_set gets flushed to main memory before we
+	 * examine any flag variables.  Otherwise a concurrent SetLatch might
+	 * falsely conclude that it needn't signal us, even though we have missed
+	 * seeing some flag updates that SetLatch was supposed to inform us of.
+	 * For the moment, callers must supply their own synchronization of flag
+	 * variables (see latch.h).
+	 */
 }
 
 /*
- * SetLatch uses SIGUSR1 to wake up the process waiting on the latch. Wake
- * up WaitLatch.
+ * SetLatch uses SIGUSR1 to wake up the process waiting on the latch.
+ *
+ * Wake up WaitLatch, if we're waiting.  (We might not be, since SIGUSR1 is
+ * overloaded for multiple purposes.)
  */
 void
 latch_sigusr1_handler(void)

src/backend/port/win32_latch.c

@@ -1,9 +1,10 @@
 /*-------------------------------------------------------------------------
  *
  * win32_latch.c
- *	  Windows implementation of latches.
+ *	  Routines for inter-process latches
  *
- * See unix_latch.c for information on usage.
+ * See unix_latch.c for header comments for the exported functions;
+ * the API presented here is supposed to be the same as there.
  *
  * The Windows implementation uses Windows events that are inherited by
  * all postmaster child processes.
@@ -23,7 +24,6 @@
 #include <unistd.h>
 
 #include "miscadmin.h"
-#include "replication/walsender.h"
 #include "storage/latch.h"
 #include "storage/shmem.h"
 
@@ -88,7 +88,7 @@ WaitLatch(volatile Latch *latch, long timeout)
 }
 
 int
-WaitLatchOrSocket(volatile Latch *latch, SOCKET sock, bool forRead,
+WaitLatchOrSocket(volatile Latch *latch, pgsocket sock, bool forRead,
 				  bool forWrite, long timeout)
 {
 	DWORD		rc;
@@ -98,6 +98,9 @@ WaitLatchOrSocket(volatile Latch *latch, SOCKET sock, bool forRead,
 	int			numevents;
 	int			result = 0;
 
+	if (latch->owner_pid != MyProcPid)
+		elog(ERROR, "cannot wait on a latch owned by another process");
+
 	latchevent = latch->event;
 
 	events[0] = latchevent;
@@ -187,15 +190,10 @@ SetLatch(volatile Latch *latch)
 
 	/*
 	 * See if anyone's waiting for the latch. It can be the current process if
-	 * we're in a signal handler. Use a local variable here in case the latch
-	 * is just disowned between the test and the SetEvent call, and event
-	 * field set to NULL.
+	 * we're in a signal handler.
 	 *
-	 * Fetch handle field only once, in case the owner simultaneously disowns
-	 * the latch and clears handle. This assumes that HANDLE is atomic, which
-	 * isn't guaranteed to be true! In practice, it should be, and in the
-	 * worst case we end up calling SetEvent with a bogus handle, and SetEvent
-	 * will return an error with no harm done.
+	 * Use a local variable here just in case somebody changes the event field
+	 * concurrently (which really should not happen).
 	 */
 	handle = latch->event;
 	if (handle)
@@ -212,5 +210,8 @@ SetLatch(volatile Latch *latch)
 void
 ResetLatch(volatile Latch *latch)
 {
+	/* Only the owner should reset the latch */
+	Assert(latch->owner_pid == MyProcPid);
+
 	latch->is_set = false;
 }

src/backend/replication/syncrep.c

@@ -166,13 +166,6 @@ SyncRepWaitForLSN(XLogRecPtr XactCommitLSN)
 	{
 		int			syncRepState;
 
-		/*
-		 * Wait on latch for up to 60 seconds. This allows us to check for
-		 * postmaster death regularly while waiting. Note that timeout here
-		 * does not necessarily release from loop.
-		 */
-		WaitLatch(&MyProc->waitLatch, 60000000L);
-
 		/* Must reset the latch before testing state. */
 		ResetLatch(&MyProc->waitLatch);
 
@@ -184,6 +177,12 @@ SyncRepWaitForLSN(XLogRecPtr XactCommitLSN)
 		 * walsender changes the state to SYNC_REP_WAIT_COMPLETE, it will
 		 * never update it again, so we can't be seeing a stale value in that
 		 * case.
+		 *
+		 * Note: on machines with weak memory ordering, the acquisition of
+		 * the lock is essential to avoid race conditions: we cannot be sure
+		 * the sender's state update has reached main memory until we acquire
+		 * the lock.  We could get rid of this dance if SetLatch/ResetLatch
+		 * contained memory barriers.
 		 */
 		syncRepState = MyProc->syncRepState;
 		if (syncRepState == SYNC_REP_WAITING)
@@ -246,6 +245,13 @@ SyncRepWaitForLSN(XLogRecPtr XactCommitLSN)
 			SyncRepCancelWait();
 			break;
 		}
+
+		/*
+		 * Wait on latch for up to 60 seconds. This allows us to check for
+		 * cancel/die signal or postmaster death regularly while waiting. Note
+		 * that timeout here does not necessarily release from loop.
+		 */
+		WaitLatch(&MyProc->waitLatch, 60000000L);
 	}
 
 	/*

src/backend/storage/lmgr/proc.c

@@ -338,7 +338,7 @@ InitProcess(void)
 	MyProc->waitLSN.xrecoff = 0;
 	MyProc->syncRepState = SYNC_REP_NOT_WAITING;
 	SHMQueueElemInit(&(MyProc->syncRepLinks));
-	OwnLatch((Latch *) &MyProc->waitLatch);
+	OwnLatch(&MyProc->waitLatch);
 
 	/*
 	 * We might be reusing a semaphore that belonged to a failed process. So

src/include/storage/latch.h

@@ -3,6 +3,66 @@
  * latch.h
  *	  Routines for interprocess latches
  *
+ * A latch is a boolean variable, with operations that let processes sleep
+ * until it is set. A latch can be set from another process, or a signal
+ * handler within the same process.
+ *
+ * The latch interface is a reliable replacement for the common pattern of
+ * using pg_usleep() or select() to wait until a signal arrives, where the
+ * signal handler sets a flag variable. Because on some platforms an
+ * incoming signal doesn't interrupt sleep, and even on platforms where it
+ * does there is a race condition if the signal arrives just before
+ * entering the sleep, the common pattern must periodically wake up and
+ * poll the flag variable. The pselect() system call was invented to solve
+ * this problem, but it is not portable enough. Latches are designed to
+ * overcome these limitations, allowing you to sleep without polling and
+ * ensuring quick response to signals from other processes.
+ *
+ * There are two kinds of latches: local and shared. A local latch is
+ * initialized by InitLatch, and can only be set from the same process.
+ * A local latch can be used to wait for a signal to arrive, by calling
+ * SetLatch in the signal handler. A shared latch resides in shared memory,
+ * and must be initialized at postmaster startup by InitSharedLatch. Before
+ * a shared latch can be waited on, it must be associated with a process
+ * with OwnLatch. Only the process owning the latch can wait on it, but any
+ * process can set it.
+ *
+ * There are three basic operations on a latch:
+ *
+ * SetLatch		- Sets the latch
+ * ResetLatch	- Clears the latch, allowing it to be set again
+ * WaitLatch	- Waits for the latch to become set
+ *
+ * WaitLatch includes a provision for timeouts (which should hopefully not
+ * be necessary once the code is fully latch-ified).
+ * See unix_latch.c for detailed specifications for the exported functions.
+ *
+ * The correct pattern to wait for event(s) is:
+ *
+ * for (;;)
+ * {
+ *	   ResetLatch();
+ *	   if (work to do)
+ *		   Do Stuff();
+ *	   WaitLatch();
+ * }
+ *
+ * It's important to reset the latch *before* checking if there's work to
+ * do. Otherwise, if someone sets the latch between the check and the
+ * ResetLatch call, you will miss it and Wait will incorrectly block.
+ *
+ * To wake up the waiter, you must first set a global flag or something
+ * else that the wait loop tests in the "if (work to do)" part, and call
+ * SetLatch *after* that. SetLatch is designed to return quickly if the
+ * latch is already set.
+ *
+ * Presently, when using a shared latch for interprocess signalling, the
+ * flag variable(s) set by senders and inspected by the wait loop must
+ * be protected by spinlocks or LWLocks, else it is possible to miss events
+ * on machines with weak memory ordering (such as PPC).  This restriction
+ * will be lifted in future by inserting suitable memory barriers into
+ * SetLatch and ResetLatch.
+ *
  *
  * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
@@ -44,16 +104,17 @@ extern int WaitLatchOrSocket(volatile Latch *latch, pgsocket sock,
 extern void SetLatch(volatile Latch *latch);
 extern void ResetLatch(volatile Latch *latch);
 
-#define TestLatch(latch) (((volatile Latch *) latch)->is_set)
+/* beware of memory ordering issues if you use this macro! */
+#define TestLatch(latch) (((volatile Latch *) (latch))->is_set)
 
 /*
- * Unix implementation uses SIGUSR1 for inter-process signaling, Win32 doesn't
- * need this.
+ * Unix implementation uses SIGUSR1 for inter-process signaling.
+ * Win32 doesn't need this.
  */
 #ifndef WIN32
 extern void latch_sigusr1_handler(void);
 #else
-#define latch_sigusr1_handler()
+#define latch_sigusr1_handler()  ((void) 0)
 #endif
 
 #endif   /* LATCH_H */