Revert "Revert "monitor the disk I/O of the server""

This reverts commit fa587455496a1aa45cc9ad7ff1280a927e74493b.

18 files changed, 362 insertions, 5 deletions

diff --git a/Porting b/Porting
index a1bf0e9..2570b3b 100644
--- a/Porting
+++ b/Porting
@@ -74,6 +74,15 @@ struct system_info
     int    *memory;      /* memory information */
 };
 
+The io_info struct is filled in by get_io_info.
+struct io_info {
+    int64_t reads;
+    int64_t readsectors;
+    int64_t writes;
+    int64_t writesectors;
+};
+
+
 The last three pointers each point to an array of integers.  The
 length of the array is determined by the length of the corresponding
 _names array in the statics structure.  Furthermore, if an entry in a
@@ -116,6 +125,8 @@ int proc_compare(caddr_t, caddr_t)
 
 	qsort comparison function
 
+void get_io_info(struct io_info *)
+
 get_process_info is called immediately after get_system_info.  In
 fact, the two functions could be rolled in to one.  The reason they
 are not is mostly historical.
diff --git a/display.c b/display.c
index 5091587..2d785c2 100644
--- a/display.c
+++ b/display.c
@@ -72,6 +72,10 @@ static int	y_mem = Y_MEM;
 static int	x_swap = -1;
 static int	y_swap = -1;
 static int	y_message = Y_MESSAGE;
+static int	x_db = X_DB;
+static int	y_db = Y_DB;
+static int	x_io = X_IO;
+static int	y_io = Y_IO;
 static int	x_header = X_HEADER;
 static int	y_header = Y_HEADER;
 static int	x_idlecursor = X_IDLECURSOR;
@@ -1196,6 +1200,57 @@ u_swap(long *stats)
 }
 
 /*
+ *     *_db(db_info) - print "DB activity: " followed by database activity
+ */
+void
+i_db(struct db_info *db_info)
+{
+	u_db(db_info);
+}
+
+void
+u_db(struct db_info *db_info)
+{
+	char buf[128];
+
+	display_write(x_db, y_db, 0, 0, "DB activity: ");
+	snprintf(buf, sizeof(buf),
+			"%3ld tps, %2ld rollbs/s, %3ld buffer r/s, %2ld hit%%, %6ld row r/s, %4ld row w/s ",
+			 db_info->numXact, db_info->numRollback,
+			 db_info->numBlockRead,
+			 db_info->numBlockRead + db_info->numBlockHit > 0 ?
+					((int64_t)(db_info->numBlockHit * 100 /
+					(db_info->numBlockRead + db_info->numBlockHit))) : 0,
+			 db_info->numTupleFetched,
+			 db_info->numTupleAltered);
+	display_write(-1, -1, 0, 0, buf);
+}
+
+/*
+ *     *_io(io_info) - print "DB I/O: " followed by IO summary
+ */
+void
+i_io(struct io_info *io_info)
+{
+	char buf[128];
+	display_write(x_io, y_io, 0, 0, "DB I/O: ");
+
+	snprintf(buf, sizeof(buf),
+			"%5ld reads/s, %5ld KB/s, %5ld writes/s, %5ld KB/s  ",
+			io_info->reads,
+			io_info->readsectors / 2,
+			io_info->writes,
+			io_info->writesectors / 2);
+	display_write(-1, -1, 0, 0, buf);
+}
+
+void
+u_io(struct io_info *io_info)
+{
+	i_io(io_info);
+}
+
+/*
  *	*_message() - print the next pending message line, or erase the one
  *				  that is there.
  *
diff --git a/display.h b/display.h
index f5b8c2a..4cc15d0 100644
--- a/display.h
+++ b/display.h
@@ -28,6 +28,10 @@ void		i_memory(long *stats);
 void		u_memory(long *stats);
 void		i_swap(long *stats);
 void		u_swap(long *stats);
+void		i_db(struct db_info *db_info);
+void		u_db(struct db_info *db_info);
+void		i_io(struct io_info *io_info);
+void		u_io(struct io_info *io_info);
 void		i_message();
 void		u_message();
 void		i_header(char *text);
diff --git a/layout.h b/layout.h
index c64f2d0..cda9db4 100644
--- a/layout.h
+++ b/layout.h
@@ -27,13 +27,17 @@
 #define  Y_CPUSTATES	2
 #define  X_MEM		8
 #define  Y_MEM		3
+#define  X_DB		0
+#define  Y_DB		4
+#define  X_IO		0
+#define  Y_IO		5
 #define  X_SWAP		6
-#define  Y_SWAP		4
-#define  Y_MESSAGE	4
+#define  Y_SWAP		6
+#define  Y_MESSAGE	6
 #define  X_HEADER	0
-#define  Y_HEADER	5
+#define  Y_HEADER	7
 #define  X_IDLECURSOR	0
-#define  Y_IDLECURSOR	4
-#define  Y_PROCS	6
+#define  Y_IDLECURSOR	6
+#define  Y_PROCS	8
 
 #endif							/* _LAYOUT_H_ */
diff --git a/machine.h b/machine.h
index 4604a5a..dbee31d 100644
--- a/machine.h
+++ b/machine.h
@@ -97,6 +97,31 @@ struct system_info
  */
 
 /*
+ * Database activity information
+ */
+struct db_info {
+	int numDb;
+	int64_t numXact;
+	int64_t numRollback;
+	int64_t numBlockRead;
+	int64_t numBlockHit;
+	int64_t numTupleFetched;
+	int64_t numTupleAltered;
+	int64_t numConflict;
+};
+
+/*
+ * Info on reads/writes happening on disk.
+ * On Linux, this can be obtained from /proc/diskstats.
+ */
+struct io_info {
+	int64_t reads;
+	int64_t readsectors;
+	int64_t writes;
+	int64_t writesectors;
+};
+
+/*
  * the process_select struct tells get_process_info what processes we
  * are interested in seeing
  */
@@ -114,6 +139,15 @@ int			machine_init(struct statics *);
 void		get_system_info(struct system_info *);
 caddr_t		get_process_info(struct system_info *, struct process_select *, int,
 							 struct pg_conninfo_ctx *, int);
+#ifdef __linux__
+caddr_t get_process_info(struct system_info *, struct process_select *, int,
+				 struct pg_conninfo_ctx *, int);
+#else
+caddr_t get_process_info(struct system_info *, struct process_select *, int,
+				 char *);
+#endif /* __linux__ */
+void		get_database_info(struct db_info *, struct pg_conninfo_ctx *);
+void		get_io_info(struct io_info *);
 char	   *format_header(char *);
 #if defined(__linux__) || defined (__FreeBSD__)
 char	   *format_next_io(caddr_t);
diff --git a/machine/m_aix43.c b/machine/m_aix43.c
index 52cf39a..72a0f43 100644
--- a/machine/m_aix43.c
+++ b/machine/m_aix43.c
@@ -741,3 +741,10 @@ proc_owner(int pid)
 
 	return (-1);
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_aix5.c b/machine/m_aix5.c
index 0656b57..c07ba26 100644
--- a/machine/m_aix5.c
+++ b/machine/m_aix5.c
@@ -828,3 +828,10 @@ int			pid;
 
 	return (-1);
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_common.c b/machine/m_common.c
index fb9c6a6..fb7833a 100644
--- a/machine/m_common.c
+++ b/machine/m_common.c
@@ -13,6 +13,13 @@
 
 #include "machine.h"
 
+/* Query to fetch information about database activity */
+#define QUERY_STAT_DB \
+		"SELECT datid, datname, numbackends, xact_commit, xact_rollback, \n" \
+		"       blks_read, blks_hit, tup_returned, tup_fetched, \n" \
+		"       tup_inserted, tup_updated, tup_deleted, conflicts \n" \
+		"FROM pg_stat_database;"
+
 char	   *backendstatenames[] =
 {
 	"", "idle", "active", "idltxn", "fast", "abort", "disabl", NULL
@@ -26,6 +33,84 @@ char	   *procstatenames[] =
 
 char		fmt_header_replication[] =
 "    PID USERNAME APPLICATION          CLIENT STATE     PRIMARY    SENT       WRITE      FLUSH      REPLAY      SLAG  WLAG  FLAG  RLAG";
+/*
+ * Get database info via the above QUERY_STAT_DB info.
+ * Returns rate info on the various statistics by comparing current
+ * values with previous values.
+ */
+void
+get_database_info(struct db_info *db_info, struct pg_conninfo_ctx *conninfo)
+{
+	struct timeval thistime;
+	double		timediff;
+	int			i;
+	int			rows;
+	PGresult   *pgresult = NULL;
+	struct db_info cur_info;
+	static struct timeval lasttime;
+	static struct db_info last_db_info;
+
+	/* calculate the time difference since our last check */
+	gettimeofday(&thistime, 0);
+	if (lasttime.tv_sec)
+		timediff = ((thistime.tv_sec - lasttime.tv_sec) +
+					(thistime.tv_usec - lasttime.tv_usec) * 1e-6);
+	else
+		timediff = 0;
+
+	lasttime = thistime;
+
+	rows = 0;
+	connect_to_db(conninfo);
+	if (conninfo->connection != NULL)
+	{
+		pgresult = PQexec(conninfo->connection, QUERY_STAT_DB);
+		if (PQresultStatus(pgresult) == PGRES_TUPLES_OK)
+			rows = PQntuples(pgresult);
+
+	}
+	if (rows == 0)
+	{
+		/* Database probably stopped, clear current and last */
+		memset(&last_db_info, 0, sizeof(last_db_info));
+	}
+	memset(&cur_info, 0, sizeof(cur_info));
+	for (i = 0; i < rows; i++)
+	{
+		PQgetvalue(pgresult, i, 2);
+		/* Count all databases, even with no active backends */
+		cur_info.numDb++;
+		cur_info.numXact += atoi(PQgetvalue(pgresult, i, 3));
+		cur_info.numRollback += atoi(PQgetvalue(pgresult, i, 4));
+		cur_info.numBlockRead += atoi(PQgetvalue(pgresult, i, 5));
+		cur_info.numBlockHit += atoi(PQgetvalue(pgresult, i, 6));
+		cur_info.numTupleFetched += atoi(PQgetvalue(pgresult, i, 8));
+		cur_info.numTupleAltered += atoi(PQgetvalue(pgresult, i, 9)) +
+			atoi(PQgetvalue(pgresult, i, 10)) +
+			atoi(PQgetvalue(pgresult, i, 11));
+		cur_info.numConflict += atoi(PQgetvalue(pgresult, i, 12));
+	}
+	if (pgresult != NULL)
+		PQclear(pgresult);
+	disconnect_from_db(conninfo);
+	if (timediff <= 0)
+	{
+		last_db_info = cur_info;
+		memset(db_info, 0, sizeof(*db_info));
+		return;
+	}
+
+	/* Compute the rate information */
+	db_info->numDb = cur_info.numDb;
+	db_info->numXact = (double)(cur_info.numXact - last_db_info.numXact) / timediff;
+	db_info->numRollback = (double)(cur_info.numRollback - last_db_info.numRollback) / timediff;
+	db_info->numBlockRead = (double)(cur_info.numBlockRead - last_db_info.numBlockRead) / timediff;
+	db_info->numBlockHit = (double)(cur_info.numBlockHit - last_db_info.numBlockHit) / timediff;
+	db_info->numTupleFetched = (double)(cur_info.numTupleFetched - last_db_info.numTupleFetched) / timediff;
+	db_info->numTupleAltered = (double)(cur_info.numTupleAltered - last_db_info.numTupleAltered) / timediff;
+	db_info->numConflict = (double)(cur_info.numConflict - last_db_info.numConflict) / timediff;
+	last_db_info = cur_info;
+}
 
 void
 update_state(int *pgstate, char *state)
diff --git a/machine/m_freebsd.c b/machine/m_freebsd.c
index 4f78043..8e20b6a 100644
--- a/machine/m_freebsd.c
+++ b/machine/m_freebsd.c
@@ -1208,3 +1208,10 @@ topproccmp(struct pg_proc *e1, struct pg_proc *e2)
 {
 	return (e1->pid < e2->pid ? -1 : e1->pid > e2->pid);
 }
+void
+
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_linux.c b/machine/m_linux.c
index 50a09cb..79b2825 100644
--- a/machine/m_linux.c
+++ b/machine/m_linux.c
@@ -15,6 +15,8 @@
  * AUTHOR: Richard Henderson <rth@tamu.edu>
  * Order support added by Alexey Klimkin <kad@klon.tme.mcst.ru>
  * Ported to 2.4 by William LeFebvre
+ *
+ * Portions Copyright (c) 2013 VMware, Inc. All Rights Reserved.
  */
 
 #define _GNU_SOURCE
@@ -1370,3 +1372,76 @@ compare_writes(const void *v1, const void *v2)
 
 	return (result);
 }
+
+/*
+ * Get IO information for the SCSI devices in the system.  Returns
+ * read/write IOs per second and bandwidth by comparing current values
+ * with previous values.
+ */
+void
+get_io_info(struct io_info *io_info)
+{
+	struct timeval thistime;
+	double timediff;
+	static struct timeval lasttime;
+	struct io_info cur_info;
+	static struct io_info last_io_info;
+	FILE *fp;
+	char line[256];
+	int major, minor;
+	char dev_name[32];
+	int64_t reads, readsectors, skip, writes, writesectors;
+	int i;
+
+	/* calculate the time difference since our last check */
+	gettimeofday(&thistime, 0);
+	if (lasttime.tv_sec)
+		timediff = ((thistime.tv_sec - lasttime.tv_sec) +
+					(thistime.tv_usec - lasttime.tv_usec) * 1e-6);
+	else
+		timediff = 0;
+
+	lasttime = thistime;
+
+	fp = fopen("/proc/diskstats", "r");
+	if (fp == NULL)
+	{
+		memset(io_info, 0, sizeof(*io_info));
+		return;
+	}
+
+	memset(&cur_info, 0, sizeof(cur_info));
+	while (fgets(line, 256, fp) != NULL)
+	{
+		i = sscanf(line, "%d %d %31s %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld",
+				   &major, &minor, dev_name,
+				   &reads, &skip, &readsectors, &skip,
+				   &writes, &skip, &writesectors, &skip,
+				   &skip, &skip, &skip);
+		if (i != 14)
+			continue;
+
+		/* Total up full scsi devices (not partitions) */
+		if (major == 8 && (minor % 16) == 0)
+		{
+			cur_info.reads += reads;
+			cur_info.readsectors += readsectors;
+			cur_info.writes += writes;
+			cur_info.writesectors += writesectors;
+		}
+    }
+	fclose(fp);
+	if (timediff == 0)
+	{
+		last_io_info = cur_info;
+		memset(io_info, 0, sizeof(*io_info));
+		return;
+	}
+
+	/* Compute the rate information */
+	io_info->reads = (double)(cur_info.reads - last_io_info.reads) / timediff;
+	io_info->readsectors = (double)(cur_info.readsectors - last_io_info.readsectors) / timediff;
+	io_info->writes = (double)(cur_info.writes - last_io_info.writes) / timediff;
+	io_info->writesectors = (double)(cur_info.writesectors - last_io_info.writesectors) / timediff;
+	last_io_info = cur_info;
+}
diff --git a/machine/m_netbsd.c b/machine/m_netbsd.c
index 6f9f566..6265817 100644
--- a/machine/m_netbsd.c
+++ b/machine/m_netbsd.c
@@ -903,3 +903,10 @@ u_int64_t  *diffs;
 	for (i = 0; i < cnt; i++)
 		*out++ = (int) ((*diffs++ * 1000 + half_total) / total_change);
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+    /* Not supported yet */
+    memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_openbsd.c b/machine/m_openbsd.c
index 6508106..27461e3 100644
--- a/machine/m_openbsd.c
+++ b/machine/m_openbsd.c
@@ -866,3 +866,10 @@ topproccmp(struct pg_proc *e1, struct pg_proc *e2)
 {
 	return (e1->pid < e2->pid ? -1 : e1->pid > e2->pid);
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_sco5.c b/machine/m_sco5.c
index 6a4b2e3..6c62c27 100644
--- a/machine/m_sco5.c
+++ b/machine/m_sco5.c
@@ -864,3 +864,10 @@ compare_time(struct proc **pp1, struct proc **pp2)
 
 	return (result);
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_sunos5.c b/machine/m_sunos5.c
index 02a0a58..22cb4d4 100644
--- a/machine/m_sunos5.c
+++ b/machine/m_sunos5.c
@@ -1799,3 +1799,10 @@ reallocproc(int n)
 			op->oldpid = -1;
 	}
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_svr4.c b/machine/m_svr4.c
index fbdf7da..f4f9ed3 100644
--- a/machine/m_svr4.c
+++ b/machine/m_svr4.c
@@ -749,3 +749,10 @@ struct sysinfo *percpu_si;
 		percpu_p++;
 	}
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/machine/m_svr5.c b/machine/m_svr5.c
index 7014013..2b9cbc5 100644
--- a/machine/m_svr5.c
+++ b/machine/m_svr5.c
@@ -1437,3 +1437,10 @@ kmet_verify(
 	}
 	return;
 }
+
+void
+get_io_info(struct io_info *io_info)
+{
+	/* Not supported yet */
+	memset(io_info, 0, sizeof(*io_info));
+}
diff --git a/pg_top.c b/pg_top.c
index 2bb78b2..fee2c18 100644
--- a/pg_top.c
+++ b/pg_top.c
@@ -124,6 +124,8 @@ void		(*d_procstates) (int, int *) = i_procstates;
 void		(*d_cpustates) (int64_t *) = i_cpustates;
 void		(*d_memory) (long *) = i_memory;
 void		(*d_swap) (long *) = i_swap;
+void		(*d_db) (struct db_info *) = i_db;
+void		(*d_io) (struct io_info *) = i_io;
 void		(*d_message) () = i_message;
 void		(*d_process) (int, char *) = i_process;
 
@@ -196,6 +198,12 @@ do_display(struct pg_top_context *pgtctx)
 									   pgtctx->order_index, &pgtctx->conninfo, pgtctx->mode);
 	}
 
+	/* Get database activity information */
+	get_database_info(&pgtctx->db_info, &pgtctx->conninfo);
+
+	/* Get database I/O information */
+	get_io_info(&pgtctx->io_info);
+
 	/* display the load averages */
 	(*d_loadave) (pgtctx->system_info.last_pid, pgtctx->system_info.load_avg);
 
@@ -236,6 +244,12 @@ do_display(struct pg_top_context *pgtctx)
 	/* display memory stats */
 	(*d_memory) (pgtctx->system_info.memory);
 
+	/* display database activity */
+	(*d_db) (&pgtctx->db_info);
+
+	/* display database I/O */
+	(*d_io) (&pgtctx->io_info);
+
 	/* display swap stats */
 	(*d_swap) (pgtctx->system_info.swap);
 
@@ -332,6 +346,8 @@ do_display(struct pg_top_context *pgtctx)
 				d_procstates = u_procstates;
 				d_cpustates = u_cpustates;
 				d_memory = u_memory;
+				d_db = u_db;
+				d_io = u_io;
 				d_swap = u_swap;
 				d_message = u_message;
 				pgtctx->d_header = u_header;
@@ -538,6 +554,8 @@ reset_display(struct pg_top_context *pgtctx)
 	d_cpustates = i_cpustates;
 	d_memory = i_memory;
 	d_swap = i_swap;
+	d_db = i_db;
+	d_io = i_io;
 	d_message = i_message;
 	pgtctx->d_header = i_header;
 	d_process = i_process;
@@ -1001,6 +1019,12 @@ main(int argc, char *argv[])
 									  &pgtctx.conninfo, -1);
 		}
 
+		/* Get database activity information */
+		get_database_info(&pgtctx.db_info, &pgtctx.conninfo);
+
+		/* Get database I/O information */
+		get_io_info(&pgtctx.io_info);
+
 		pgtctx.timeout.tv_sec = 1;
 		pgtctx.timeout.tv_usec = 0;
 		select(0, NULL, NULL, NULL, &pgtctx.timeout);
diff --git a/pg_top.h b/pg_top.h
index b3af7ec..4156754 100644
--- a/pg_top.h
+++ b/pg_top.h
@@ -101,6 +101,8 @@ struct pg_top_context
 								 * system. */
 	int			order_index;
 	char	   *order_name;
+	struct db_info db_info;
+	struct io_info io_info;
 	struct process_select ps;
 	char		show_tags;
 	struct statics statics;