Implements support for arbitrary Performance Co-Pilot
metrics with per-process instance domains to form new
htop columns. The column-to-metric mappings are setup
using configuration files which will be documented via
man pages as part of a follow-up commit.
We provide an initial set of column configurations so
as to provide new capabilities to pcp-htop: including
configs for containers, open fd counts, scheduler run
queue time, tcp/udp bytes/calls sent/recv, delay acct,
virtual machine guests, detailed virtual memory, swap.
Note there is a change to the configuration file path
resolution algorithm introduced for 'dynamic meters'.
First, look in any custom PCP_HTOP_DIR location. Then
iterate, in priority order, users home directory, then
local sysadmins files in /etc/pcp/htop, then readonly
configuration files below /usr/share/pcp/htop. This
final location becomes the preferred place for our own
shipped meter and column files.
The Settings file (htoprc) writing code is updated to
not using the numeric identifier for dynamic columns.
The same strategy used for dynamic meters is used here
where we write Dynamic(name) so the name can be setup
once more at start. Regular (static) columns writing
to htoprc - i.e. numerically indexed - is unchanged.
Currently htop does not support offline CPUs and hot-swapping, e.g. via
echo 0 > /sys/devices/system/cpu/cpu2/online
Split the current single cpuCount variable into activeCPUs and
existingCPUs.
Supersedes: #650
Related: #580
This commit is based on exploratory work by Sohaib Mohamed.
The end goal is two-fold - to support addition of Meters we
build via configuration files for both the PCP platform and
for scripts ( https://github.com/htop-dev/htop/issues/526 )
Here, we focus on generic code and the PCP support. A new
class DynamicMeter is introduced - it uses the special case
'param' field handling that previously was used only by the
CPUMeter, such that every runtime-configured Meter is given
a unique identifier. Unlike with the CPUMeter this is used
internally only. When reading/writing to htoprc instead of
CPU(N) - where N is an integer param (CPU number) - we use
the string name for each meter. For example, if we have a
configuration for a DynamicMeter for some Redis metrics, we
might read and write "Dynamic(redis)". This identifier is
subsequently matched (back) up to the configuration file so
we're able to re-create arbitrary user configurations.
The PCP platform configuration file format is fairly simple.
We expand configs from several directories, including the
users homedir alongside htoprc (below htop/meters/) and also
/etc/pcp/htop/meters. The format will be described via a
new pcp-htop(5) man page, but its basically ini-style and
each Meter has one or more metric expressions associated, as
well as specifications for labels, color and so on via a dot
separated notation for individual metrics within the Meter.
A few initial sample configuration files are provided below
./pcp/meters that give the general idea. The PCP "derived"
metric specification - see pmRegisterDerived(3) - is used
as the syntax for specifying metrics in PCP DynamicMeters.
Refactor the sample time code to make one call to gettimeofday
(aka the realtime clock in clock_gettime, when available) and
one to the monotonic clock. Stores each in more appropriately
named ProcessList fields for ready access when needed. Every
platform gets the opportunity to provide their own clock code,
and the existing Mac OS X specific code is moved below darwin
instead of in Compat.
A couple of leftover time(2) calls are converted to use these
ProcessList fields as well, instead of yet again sampling the
system clock.
Related to https://github.com/htop-dev/htop/pull/574
The end goal is to consolidate all the points in htop that can only work in
live-only mode today, so that will be able to inject PCP archive mode and have
a chance at it working.
The biggest problem we've got at this moment is all the places that are
independently asking the kernel to 'give me the time right now'.
Each of those needs to be audited and ultimately changed to allow platforms to
manage their own idea of time.
So, all the calls to gettimeofday(2) and time(2) are potential problems.
Ultimately I want to get these down to just one or two.
Related to https://github.com/htop-dev/htop/pull/574
* Set process data for:
- minflt
- majflt
- processor
- nlwp
* Drop unimplemented nlwp column
* Scan userland threads
* Mark a 'Thread is currently on a CPU.' with 'R', and processes
'Currently runnable' with 'P', do confine with man:ps(1) and Linux.
See https://man.openbsd.org/ps.1
* Show CPU frequency
Use similar calculation than procps.
Show AvailableMemory in text mode.
Use total minus available memory instead of manually computed used-
memory as fraction part in bar mode (if available).
According to the Linux kernel documentation, "SwapCached" tracks "memory
that once was swapped out, is swapped back in but still also is
in the swapfile (if memory is needed it doesn't need to be swapped out
AGAIN because it is already in the swapfile. This saves I/O)."
The global ProcessList structure contains a couple of unused
fields. 'sharedMem' has never been used by any Meter, since
its not been anything other than zero in Linux /proc/meminfo
for many, many years. The freeMem field is only used in the
usedMem calculation, so it can reside on the stack like some
other memory variables used within-calculations-only and not
exposed to the user via a Meter.
Generic data, as CPU and memory usage, are used by Meters.
In paused mode they would stop receiving updates and especially Graph
Meters would stop showing continuous data.
Improves: #214Closes: #253
Reasoning:
- implementation was unsound -- broke down when I added a fairly
basic macro definition expanding to a struct initializer in a *.c
file.
- made it way too easy (e.g. via otherwise totally innocuous git
commands) to end up with timestamps such that it always ran
MakeHeader.py but never used its output, leading to overbuild noise
when running what should be a null 'make'.
- but mostly: it's just an awkward way of dealing with C code.
Daniel Lange