Bug 315641

I have noticed when a system has over 3,000 open TCP connections, while the underlying Linux OS can handle it and the app managing the 3,000+ open TCP connections can handle it, ksysguardd ramps up to 100% CPU utilization at this point.  I tried with up to 12,000 open TCP sockets and while everything else was fine, ksysguardd slowed to a crawl while running at 100% CPU utilization.  I found the source of the problem in /ksysguard/ksysguardd/netstat.c.  Specifically the get_num_sockets() function on line 166 loops through the file /proc/net/tcp and counts up the lines, which can become quite large when many TCP sockets are open.  What I did to fix this was to tap into the file /proc/net/sockstat and just grab the one line I need.  I noticed that the /proc/net/tcp results and /proc/net/sockstat listing when combining the inuse and tw stats don't always match up, but combined (inuse and tw) they usually match up to manual counting and is a whole heck of a lot faster.

Here is the code I added / changed in order to facilitate this dramatic performance boost:

On line 176 after the get_num_sockets() fuction:
int get_num_sockets_sockstat(FILE *netstat, const char *stat_type)
{
	char line[1024];
	int line_count= 0;
	int inuse, orphan, tw, alloc, mem;
	inuse= orphan= tw= alloc= mem= 0;
	char *read_result= NULL;

	read_result= fgets(line, 1024, netstat); /* Read and throw away sockets total */
	read_result= fgets(line, 1024, netstat); /* Read TCP stats */
	if (strcmp("tcp", stat_type) == 0) { /* Get TCP stats */
		if (read_result != NULL) {
			sscanf (line, "TCP: inuse %i orphan %i tw %i alloc %i mem %i",
				&inuse, &orphan, &tw, &alloc, &mem);
			line_count= inuse + tw;
		}
	} else if (strcmp("udp", stat_type) == 0) { /* Get UDP stat */
		read_result= fgets(line, 1024, netstat); /* Read UDP stat */
		if (read_result != NULL) {
			sscanf (line, "UDP: inuse %i mem %i",
				&inuse, &mem);
			line_count= inuse + tw;
		}
	} else if (strcmp("raw", stat_type) == 0) { /* Get RAW stat */
		read_result= fgets(line, 1024, netstat); /* Read and throw away UDP stat */
		read_result= fgets(line, 1024, netstat); /* Read and throw away UDP Lite stat */
		read_result= fgets(line, 1024, netstat); /* Read RAW stat */
		if (read_result != NULL) {
			sscanf (line, "RAW: inuse %i",
				&inuse);
			line_count= inuse;
		}
	}

	return line_count;
} /* End get_num_sockets_sockstat */


And for what is now my line 274 function updateNetStat() I changed the file opened for TCP, UDP, and RAW to open the /proc/net/sockstat file instead.

Reproducible: Always

Steps to Reproduce:
1. Open ksysguard and monitor open TCP sockets.
2. Run a program that can potentially have lots of open TCP sockets such as Apache.
3. Run script to open and hold open TCP connections as this is the lazy way to get a whole lot of connections to be open at once.
Actual Results:  
ksysguardd reached maximum CPU utilization around 3,000 open TCP connections and slowed to a crawl as the connection count went up to 12,000 open TCP connections.

The above patch I made seems to have fixed this problem.  Of course with anything open source you are free to poke at it and let me know what you think.

Expected Results:  
No extra CPU load as more TCP connections are open as ksysguardd is supposed to monitor activity, not grind through intensive analysis every time it is polled.

This was run on Ubuntu 12.04 LTS 64-bit.  The source was downloaded from the included package manager.