Several fixes for concurrent index builds:

schmiddy · schmiddy · commit 42357353a7e4 · 2012-12-13T19:12:05.000-07:00
* Use poll() if it is available, or select() otherwise, to
   efficiently wait on index builds in worker queries to finish.
 * fix off-by-one error when initially assigning workers
 * move PQsetnonblocking() calls to setup_workers()
diff --git a/bin/pg_repack.c b/bin/pg_repack.c
@@ -24,17 +24,32 @@ const char *PROGRAM_VERSION = "unknown";
 
 #include "pgut/pgut-fe.h"
 
+#include <errno.h>
 #include <string.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <time.h>
 
+
+#ifdef HAVE_POLL_H
+#include <poll.h>
+#endif
+#ifdef HAVE_SYS_POLL_H
+#include <sys/poll.h>
+#endif
+#ifdef HAVE_SYS_SELECT_H
+#include <sys/select.h>
+#endif
+
+
 /*
  * APPLY_COUNT: Number of applied logs per transaction. Larger values
  * could be faster, but will be long transactions in the REDO phase.
  */
 #define APPLY_COUNT		1000
 
+/* poll() or select() timeout, in seconds */
+#define POLL_TIMEOUT    3
 
 /* Compile an array of existing transactions which are active during
  * pg_repack's setup. Some transactions we can safely ignore:
@@ -633,7 +648,7 @@ rebuild_indexes(const repack_table *table)
 			 */
 			index_jobs[i].status = FINISHED;
 		}
-		else if (i <= workers.num_workers) {
+		else if (i < workers.num_workers) {
 			/* Assign available worker to build an index. */
 			index_jobs[i].status = INPROGRESS;
 			index_jobs[i].worker_idx = i;
@@ -656,23 +671,67 @@ rebuild_indexes(const repack_table *table)
 	}
 	PQclear(res);
 
+	/* How many workers we kicked off earlier. */
+	num_active_workers = num_indexes > workers.num_workers ? workers.num_workers : num_indexes;
+
 	if (workers.num_workers > 1)
 	{
-		/* How many workers we kicked off earlier. */
-		num_active_workers = num_indexes > workers.num_workers ? workers.num_workers : num_indexes;
+		int freed_worker = -1;
+		int ret;
+
+/* Prefer poll() over select(), following PostgreSQL custom. */
+#ifdef HAVE_POLL
+		struct pollfd *input_fds;
+
+		input_fds = pgut_malloc(sizeof(struct pollfd) * num_active_workers);
+		for (i = 0; i < num_active_workers; i++)
+		{
+			input_fds[i].fd = PQsocket(workers.conns[i]);
+			input_fds[i].events = POLLIN | POLLERR;
+			input_fds[i].revents = 0;
+		}
+#else
+		fd_set input_mask;
+		struct timeval timeout;
+		/* select() needs the highest-numbered socket descriptor */
+		int max_fd = 0;
+
+		FD_ZERO(&input_mask);
+		for (i = 0; i < num_active_workers; i++)
+		{
+			FD_SET(PQsocket(workers.conns[i]), &input_mask);
+			if (PQsocket(workers.conns[i]) > max_fd)
+				max_fd = PQsocket(workers.conns[i]);
+		}
+#endif
 
 		/* Now go through our index builds, and look for any which is
 		 * reported complete. Reassign that worker to the next index to
 		 * be built, if any.
 		 */
-		while (num_active_workers)
+		while (num_active_workers > 0)
 		{
-			int freed_worker = -1;
+			elog(DEBUG2, "polling %d active workers", num_active_workers);
+
+#ifdef HAVE_POLL
+			ret = poll(input_fds, num_active_workers, POLL_TIMEOUT * 1000);
+#else
+			/* re-initialize timeout before each invocation of select()
+			 * just in case select() modifies timeout to indicate remaining
+			 * time.
+			 */
+			timeout.tv_sec = POLL_TIMEOUT;
+			timeout.tv_usec = 0;
+			ret = select(max_fd + 1, &input_mask, NULL, NULL, &timeout);
+#endif
+			if (ret < 0 && errno != EINTR)
+				elog(ERROR, "poll() failed: %d, %d", ret, errno);
 
 			for (i = 0; i < num_indexes; i++)
 			{
 				if (index_jobs[i].status == INPROGRESS)
 				{
+					Assert(index_jobs[i].worker_idx >= 0);
 					/* Must call PQconsumeInput before we can check PQisBusy */
 					if (PQconsumeInput(workers.conns[index_jobs[i].worker_idx]) != 1)
 					{
@@ -699,7 +758,13 @@ rebuild_indexes(const repack_table *table)
 							}
 							PQclear(res);
 						}
-
+						
+						/* We are only going to re-queue one worker, even
+						 * though more than one index build might be finished.
+						 * Any other jobs which may be finished will
+						 * just have to wait for the next pass through the
+						 * poll()/select() loop.
+						 */
 						freed_worker = index_jobs[i].worker_idx;
 						index_jobs[i].status = FINISHED;
 						num_active_workers--;
@@ -733,7 +798,6 @@ rebuild_indexes(const repack_table *table)
 				}
 				freed_worker = -1;
 			}
-			sleep(1);
 		}
 
 	}
@@ -980,7 +1044,6 @@ repack_one_table(const repack_table *table, const char *orderby)
 		goto cleanup;
 	}
 
-
 	/*
 	 * 4. Apply log to temp table until no tuples are left in the log
 	 * and all of the old transactions are finished.
diff --git a/bin/pgut/pgut-fe.c b/bin/pgut/pgut-fe.c
@@ -82,7 +82,11 @@ setup_workers(int num_workers)
  			 */
  			elog(DEBUG2, "Setting up worker conn %d", i);
 
- 			/* Don't confuse pgut_connections by using pgut_connect() */
+ 			/* Don't confuse pgut_connections by using pgut_connect()
+			 *
+			 * XXX: could use PQconnectStart() and PQconnectPoll() to
+			 * open these connections in non-blocking manner.
+			 */
  			conn = PQconnectdb(buf.data);
  			if (PQstatus(conn) == CONNECTION_OK)
  			{
@@ -94,6 +98,15 @@ setup_workers(int num_workers)
 					 PQerrorMessage(conn));
 				break;
 			}
+
+            /* Make sure each worker connection can work in non-blocking
+             * mode.
+             */
+            if (PQsetnonblocking(workers.conns[i], 1))
+			{
+				elog(ERROR, "Unable to set worker connection %d "
+					 "non-blocking.", i);
+			}
  		}
 		/* In case we bailed out of setting up all workers, record
 		 * how many successful worker conns we actually have.
