Skip to contentSkip to conversation

PATCH: ausearch doesn't correcly act as a filter when no EOF on input

PATCH Applied Tony Jones 2 participants 3 messages Nov 18, 2008 – Nov 19, 2008

lore
Share & tools
lore

Tony Jones

author
#0 lore raw

This was filed as a bug in our bugzilla.

works : cat /var/log/audit/audit.log | ausearch -i -if /dev/stdin | cat
doesnt: tail -f /var/log/audit/audit.log | ausearch -i -if /dev/stdin | cat

Obviously it's a contrived example, they have more interesting processes each
side of the filter. Issue is that tail -f never indicates EOF and if ausearch
stdout is a pipe (versus a file), the output can remain queued in the pipebuf.

Following patch fixes it, or a simpler patch could unconditionally flush stdout.

I've not looked for similar issues elsewhere.

Tony

ausearch.c renamed from ausearch.c.old +6 -3
@@ -58,11 +58,11 @@
 extern int match(llist *l);
 extern void output_record(llist *l);
 
-static int input_is_pipe(void)
+static int is_pipe(int fd)
 {
 	struct stat st;
 
-	if (fstat(0, &st) == 0) {
+	if (fstat(fd, &st) == 0) {
 		if (S_ISFIFO(st.st_mode)) 
 			pipe_mode = 1;
 	}
@@ -92,7 +92,7 @@
 		rc = process_file(user_file);
 	else if (force_logs)
 		rc = process_logs();
-	else if (input_is_pipe())
+	else if (is_pipe(0))
 		rc = process_stdin();
 	else
 		rc = process_logs();
@@ -175,6 +175,7 @@
 {
 	llist entries; // entries in a record
 	int ret;
+	int flush = is_pipe(1);
 
 	/* For each record in file */
 	list_create(&entries);
@@ -185,6 +186,8 @@
 		}
 		if (match(&entries)) {
 			output_record(&entries);
+			if (flush) 
+				fflush(stdout);
 			found = 1;
 			if (just_one) {
 				list_clear(&entries);
#1 lore raw
Quoted 2 lines · On Monday 17 November 2008 19:17:54 Tony Jones wrote:

Issue is that tail -f never indicates EOF and if ausearch stdout is a pipe (versus a file), the output can remain queued in the pipebuf.

The general problem is a little bigger. ausearch can't tell when an event is complete. It assumes that the kernel serializes things - which it doesn't. With files, it flushes everything when it hits EOF. But pipes don't have that. The solution I was going to implement was to start a timer when the input is a pipe and timeout the currently accumulating event. Flushing the output if using a pipe looks like a good measure, too.

Quoted 2 lines

Following patch fixes it, or a simpler patch could unconditionally flush stdout.

I'd rather not flush all the time otherwise the performance improvement of buffering is lost. I currently have ausearch completely tore apart to rework its event collection logic so that interlaced events are assembled correctly and completely. Maybe in a few days I can revisit this problem.

Quoted 1 line

I've not looked for similar issues elsewhere.

aureport shares much of the same IO design and the auparse library also does not use clock time to finish off events if the input is a pipe.

-Steve

Applied
#2 lore raw
In reply to Tony Jones
Quoted 3 lines · On Monday 17 November 2008 19:17:54 Tony Jones wrote:

Issue is that tail -f never indicates EOF and if ausearch stdout is a pipe (versus a file), the output can remain queued in the pipebuf.

Applied in commit #190.

-Steve