* Rename parsePoolObjsetFile to parseLinuxPoolObjsetFile to better reflect
it's scope
* Create a new parseFreeBSDPoolObjsetStats function, to generate a list
of per pool metrics to be queried via sysctl
---------
Signed-off-by: Conall O'Brien <conall@conall.net>
* Optionally fetch ARP stats via rtnetlink instead of procfs
Implement collection of ARP stats via rtnetlink to work around
shortcomings in the output of /proc/net/arp, which truncates InfiniBand
link-layer addresses.
Fixes: #2776
---------
Signed-off-by: Daniel Swarbrick <daniel.swarbrick@gmail.com>
Co-authored-by: Ben Kochie <superq@gmail.com>
Despite being quite hard to provoke (< 10% in my testing), the btrfs
collector would occasionally leave stale FDs relating to btrfs
mountpoints, making the filesystems unable to be unmounted.
Fixes: #2772.
Signed-off-by: Daniel Swarbrick <daniel.swarbrick@gmail.com>
Revert changes to node_cpu_info and add new node_cpu_frequency_hertz
metric for measuring CPU frequency from /proc/cpuinfo
Signed-off-by: John Kordich <jkordich@gmail.com>
For CPUs which don't have an available (or insertable) cpufreq driver,
the /proc/cpuinfo file can sometimes have accurate CPU core frequency
measurements. This change replaces the constant value of "1" for the
"node_cpu_info" metric with the parsed CPU MHz value from
/proc/cpuinfo for each core.
Signed-off-by: John Kordich <jkordich@gmail.com>
Ensure that unwanted tests are correctly excluded when various build
tags are specified, i.e. when the code that they test would be excluded
from compilation.
Signed-off-by: Daniel Swarbrick <daniel.swarbrick@gmail.com>
Drop redundant GOOS build tags at start of file if the constraint is
already specified by the filename, e.g. foo_GOOS.go or
foo_GOOS_GOARCH.go, avoiding potential confusion in future.
cf. https://pkg.go.dev/cmd/go#hdr-Build_constraints
Signed-off-by: Daniel Swarbrick <daniel.swarbrick@gmail.com>
On some platforms, `msg.Attributes.Stats64` is `nil` because the kernel doesn't
expose 64-bit stats. In that case, return `msg.Attributes.Stats` instead, which
are the 32-bit equivalent.
Note that `RXOtherhostDropped` isn't available in that case, so we hardcode it
to zero.
Fixes#2756.
Signed-off-by: Benoît Knecht <bknecht@protonmail.ch>
Add a new pattern to allow moving the flag handling out of the collector
package and into a separate package. This allows collector package users
to create their own config or flag handling.
Signed-off-by: Ben Kochie <superq@gmail.com>
Use the correct include value to the device filter function.
* Add new bogus hwmon fixture.
* Update end-to-end test to use hwmon chip include flag.
Signed-off-by: Ben Kochie <superq@gmail.com>
prefix.
Leave an annotation about using regexps instead of device_filter.go, so
@SuperQ doesn't need to remember everything.
Signed-off-by: Conall O'Brien <conall@conall.net>
* Add include and exclude flags chip name flags to hwmon collector, following example in systemd collector
---------
Signed-off-by: Conall O'Brien <conall@conall.net>
Co-authored-by: Ben Kochie <superq@gmail.com>
This change adds the ability to process multiple stat calls in parallel.
Processing is rate-limited based on the new flag
`collector.filesystem.stat-workers` (default 4).
Caveat: filesystem stats information is no longer in the same order as
returned by `/proc/1/mounts`. This should not be an issue.
Caveat: This change currently uses unbuffered channels to prove
correctness without reliance on buffers. Buffered channels will yield
superior performance.
Signed-off-by: Erica Mays <erica@emays.dev>
Read missing dev_id, name_assign_type, and addr_assign_type
from sysfs, since they only take a device-specific lock and
not the whole RTNL lock. This means reading them is much less
impactful on other system processes than many of the other
attributes in sysfs that do take the RTNL lock.
Signed-off-by: Dan Williams <dcbw@redhat.com>
