Re: [Powertop] patch: tell user if the system is running out of FDs

[Jaroslav Skarvada <jskarvad at redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2753 bytes --]

> Hello,
> 
> > I used Youquan's patch, but we have noticed that it can only overcome soft
> > limits,
> > but not hard limits. E.g. on Fedora (and probably others) there is a soft
> > limit
> > of 1024 and a hard limit of 4096, so if powertop opens at least 15 file
> > descriptors
> > per CPU, then this will fail on a system with >273 CPUs.
> 
> that's a pretty big system
> 
> > In case you think we can also temporally increase the hard limit (as we
> > already have
> > the privilege), the attached patch will do it. In case you think the hard
> > limits set by
> > the system administrator are untouchable (e.g. violation of system
> > resources hard limits
> > could result in a bad effects like explosion in a nuclear power plant :),
> > the
> > previously sent patch that fixes the error message needs to be also used
> 
> hm, need to think about this for a moment. both your patches make sense.
> I'd probably go with `temporally increase the hard limit' solution, to make
> someones life easier. what do you, guys, think?
> 
> 
> 	-ss
> 
> > regards
> > 
> > Jaroslav
> 
> > diff --git a/src/main.cpp b/src/main.cpp
> > index 0883424..e0e8a71 100644
> > --- a/src/main.cpp
> > +++ b/src/main.cpp
> > @@ -36,6 +36,8 @@
> >  #include <getopt.h>
> >  #include <unistd.h>
> >  #include <locale.h>
> > +#include <sys/resource.h>
> > +#include <linux/limits.h>
> >  
> >  #include "cpu/cpu.h"
> >  #include "process/process.h"
> > @@ -283,11 +285,16 @@ static void powertop_init(void)
> >  	static char initialized = 0;
> >  	int ret;
> >  	struct statfs st_fs;
> > +	struct rlimit rlmt;
> >  
> >  	if (initialized)
> >  		return;
> >  
> >  	checkroot();
> > +
> > +	rlmt.rlim_cur = rlmt.rlim_max = NR_OPEN;

This seems to be wrong. On my system NR_OPEN is 1024, but
/proc/sys/fs/nr_open gives me 1048576. It seems the linux kernel
hardcodes 1024 * 1024 as the default kernel limit. This default
limit can be raised to:

sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
  -BITS_PER_LONG;

which is 2 Gi - 63 on my 64 bit system. I think we shouldn't
touch/increase this. It seems there is no header definition for these
limits.

I am proposing new patch, which reads the current kernel limit from the
/proc. I am not sure about the portability, but at least it shouldn't
break anything.

Finally I think, the first patch could be applied together with this patch
(probably with the changed message) to notify user that he/she reached
the FDs limit, because there is still some limit. This could also
occur if some process would decrease the /proc/sys/fs/nr_open after
it is read by powertop and before the ulimit is set

regards

Jaroslav

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: powertop-2.4-unlimit-fds.patch --]
[-- Type: text/x-patch, Size: 1335 bytes --]

diff --git a/src/main.cpp b/src/main.cpp
index 0883424..16b1613 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -28,6 +28,7 @@
  *	Arjan van de Ven <arjan@linux.intel.com>
  */
 #include <iostream>
+#include <fstream>
 #include <stdlib.h>
 #include <stdio.h>
 #include <time.h>
@@ -36,6 +37,7 @@
 #include <getopt.h>
 #include <unistd.h>
 #include <locale.h>
+#include <sys/resource.h>
 
 #include "cpu/cpu.h"
 #include "process/process.h"
@@ -60,6 +62,8 @@
 
 #define DEBUGFS_MAGIC          0x64626720
 
+#define NR_OPEN_DEF 1024 * 1024
+
 int debug_learning = 0;
 unsigned time_out = 20;
 int leave_powertop = 0;
@@ -278,16 +282,35 @@ static void checkroot() {
 
 }
 
+static int get_nr_open(void) {
+	int nr_open = NR_OPEN_DEF;
+	ifstream file;
+
+	file.open("/proc/sys/fs/nr_open", ios::in);
+	if (file) {
+		file >> nr_open;
+		if (!file)
+			nr_open = NR_OPEN_DEF;
+		file.close();
+	}
+	return nr_open;
+}
+
 static void powertop_init(void)
 {
 	static char initialized = 0;
 	int ret;
 	struct statfs st_fs;
+	struct rlimit rlmt;
 
 	if (initialized)
 		return;
 
 	checkroot();
+
+	rlmt.rlim_cur = rlmt.rlim_max = get_nr_open();
+	setrlimit (RLIMIT_NOFILE, &rlmt);
+
 	ret = system("/sbin/modprobe cpufreq_stats > /dev/null 2>&1");
 	ret = system("/sbin/modprobe msr > /dev/null 2>&1");
 	statfs("/sys/kernel/debug", &st_fs);