matthiaspetermann / netbsd_exporter Goto Github PK

How to reproduce:

Simply build and run on a 32bit machine. I tested my usual candidates
for this: port-i386, port-sparc, port-vax. I assume this code was written
and tested on 64bit machines only.

Example output (snipped irrelevant parts):

netbsd_fs_size_bytes{device="wd0a",mountpoint="/"} -1958563840
netbsd_fs_used_bytes{device="wd0a",mountpoint="/"} -1555032064
netbsd_fs_free_bytes{device="wd0a",mountpoint="/"} -1808840704

df says:

/dev/wd0a 26G 2.6G 22G 10% /

How to fix:

This is essentially the old C language trap that identically named integer types are
of different size depending on platform word size. In this case, long is
8 bytes (64bit) on a 64bit machine (which is large enough here) and
4 bytes (32bit) on a 32bit machine (which will cheerfully overflow
here).

There are reasons int64_t, uint64_t, int32_t and friends are defined
these days. Of course with that, gcc rightfully complains about
the formatting args, because uint64_t is long long int on 32bit and
long int on 64bit machines. So I ended up using long long which is
64bit on both.

With that, it works correctly on both 32bit and 64bit machines. Tested
platforms: port-i386, port-amd64, port-vax, port-sparc, port-sparc64.

And for those who no longer have 32bit machines, qemu works very well
(that's how I run my port-sparc machine on an arm64 host).

Patch follows

--- netbsd_exporter.h.orig      2024-05-09 18:58:29.961248161 +0200
+++ netbsd_exporter.h   2024-05-09 19:49:00.021634195 +0200
@@ -31,7 +31,7 @@
 int option_http_header = 1;
 int option_syslog = 1;

-void print_filesystem_metric(const char*, const char*, const char*, long );
+void print_filesystem_metric(const char*, const char*, const char*, long long unsigned int);

 void print_disk_io_metric(const char*, long long unsigned int, long long unsigned int );

--- netbsd_exporter.c.orig      2024-05-07 11:40:44.013487932 +0200
+++ netbsd_exporter.c   2024-05-09 19:49:15.832451252 +0200
@@ -53,8 +53,8 @@
 #include "netbsd_exporter.h"
 #include "version.h"

-void print_filesystem_metric(const char* metric, const char* device, const char* mountpoint, long value) {
-    printf("netbsd_fs_%s_bytes{device=\"%s\",mountpoint=\"%s\"} %ld\n", metric, device, mountpoint, value);
+void print_filesystem_metric(const char* metric, const char* device, const char* mountpoint, long long unsigned int value) {
+    printf("netbsd_fs_%s_bytes{device=\"%s\",mountpoint=\"%s\"} %llu\n", metric, device, mountpoint, value);
 }

 void print_disk_io_metric(const char* device, long long unsigned int rbytes, long long unsigned int wbytes) {

The utilization of the individual CPUs would be interesting (or all of them together) - is a different metric than the load.

(reported by Thomas Klausner)

Temperatures (via envsys(4)) could also be displayed.

(Reported by Thomas Klausner)

The snmp_exporter against my Fritzbox returns the load as a
value with labels, so:
laLoadInt{laNames="Load-1"} 10
laLoadInt{laNames="Load-15"} 17
laLoadInt{laNames="Load-5"} 10

The netbsd_exporter has 3 metrics for this:
netbsd_load1 0.059082
netbsd_load5 0.083496
netbsd_load15 0.141113

I find this a bit more convenient with labels, because I can
graph construction I only have to include one metric.

Not a big deal and if you disagree, that's ok too, I just
just wanted to mention it.

(Reported by Thomas Klausner)

"Beim Basteln des neuen Dashboards ist mir aufgefallen, dass
prometheus/grafana die Namen ein paar der Metriken nicht mögen, weil
sie nicht auf "_total" o.ä. enden. Ich hänge dir die Screenshots an.
(es funktioniert, aber es gibt so ein Warnungsdreieck bei den Graphen)"

(reported by Thomas Klausner)