Columns for delaystats

[brendangregg]

Please add column for delaystats to htop. They've been in Linux for a long time, and show delays for different types of resources. There's not many tools that use them, so far.

The Linux kernel source ships with an example: https://github.com/torvalds/linux/blob/master/tools/accounting/getdelays.c

tools/accounting> ./getdelays -d -p $$
print delayacct stats ON
PID	4961


CPU             count     real total  virtual total    delay total  delay average
                  367      160000000      150819518        1942403          0.005ms
IO              count    delay total  delay average
                   40      306740944              7ms
SWAP            count    delay total  delay average
                    0              0              0ms
RECLAIM         count    delay total  delay average
                    0              0              0ms

It's the delay totals that are interesting. They are measured in nanoseconds, but presenting them in htop I think it would make more sense as percentages. Eg, to show 50% CPU delays for a process (thread) would mean it's spending half it's time waiting for a turn on CPU.

[andrestc]

Is this a good starting issue for someone who is new to the project(does not know the codebase)? If so, I would like to try and tackle this.

[brendangregg]

@andrestc it's a moderate change. A trivial addition would be a field that comes from a /proc file that htop is already reading. A bit harder would be adding a new /proc file that wasn't read before.

In this case, delaystats (sorry, I should be using the real name, which is "delay accounting") is read via the netlink interface, which is completely different, and more difficult, than a /proc reader. htop currently doesn't have a netlink reader (AFAIK). Fortunately, getdelays.c in the Linux source shows what needs to be done, but yes, it means adding a whole new data source to htop, which is why I'd call it moderate.

Is it good for a beginner? I'm not sure -- it would at least need a determined beginner who understood it's not, say, a 5 line trivial change. It'd probably end up a few hundred lines of code. Most of which may come from getdelays.c.

[andrestc]

@brendangregg thanks for the detailed answer.

I will give this a shot. But, as I might take slightly longer than someone with more experience on htop and its implementation details and I don't want to slow down anything, anyone can feel free to do it and beat me to it

[hishamhm]

@andrestc Yes, that would be an excellent way to get started contributing to htop! @brendangregg's assessment is fully on point. It's not a quick patch, but it should be pretty self-contained, consisting basically of additions rather than modifications to the code base.

I won't be able to work on it myself over the next few weeks, so thank you for the initiative! Feel free to tag me via Github mentions if you have any questions about your work-in-progress.

[andrestc]

Great, @hishamhm. I will probably be able to get some code done on the next weekend.

In the meantime, getdelays.c and others, e.g, the man for libnetlink, recommend the use of higher level libraries like libmnl or libnl. Should I aim to use one of these libraries and add such dependency to htop or just approach it like getdelays.c and just handle the parsing and etc myself?

[hishamhm]

You can add the dependency, as that means less code to be added to htop itself. You can look at how hwloc was added as an optional dependency and follow that (adding the check/flag to configure.ac and then guarding the code with appropriate #defines).

[hishamhm]

Linux delay stats are now merged! Major kudos to @andrestc for the implementation and thanks @brendangregg for the suggestion. A new release of htop with the feature is coming soon! :)

[pmucci]

HI folks, it's worth noting in the documentation that delay stats are only available to root users. Even if you own the process, it seems delay accounting is only available to the root user or those with CAP_SYS_ADMIN capability.