Processes sometimes don't show up?

[Atemu]

Sometimes when I start iotop-c -o when a process is already running, that process will not be shown. It does show a Total DISK WRITE including that process' IO and a few btrfs threads but their small IO doesn't add up to the total.

This does not happen with the python iotop.

The issue is very spurious and hard to reproduce.

[bbonev]

In curses mode, -o depends on several things:

• if the graph is hidden, it will use IO or R+W (depending on the current kernel.task_delayacct)
• if the graph is visible, it will use the visible values from the current graph type

There is a 'feature' with newer kernels - changing kernel.task_delayacct will not affect already running processes.

And there is one more moving part - it is possible for a process to have R/W with IO being zero (or more precisely said, rounded off to zero).

One way to see if that is the case is to change the graph type to R+W when this happens.

[Atemu]

kernel.task_delayacct was enabled in my case and the exact command was sudo iotop-c --hide-prio --hide-graph -o -e.

[bbonev]

With the combination of --hide-graph and kernel.task_delayacct=1, -o to will only show processes with non-zero IO wait. Unless kernel.task_delayacct is activated as kernel boot parameter, there may be processes that will perform IO and have IO wait with zero value - current kernels do not start tracking that for existing processes after activating task_delayacct.

I understand and agree that this is confusing, but I see no possible fix for it. If I am missing some way to improve this, please share it.

[Atemu]

With the combination of --hide-graph and kernel.task_delayacct=1, -o to will only show processes with non-zero IO wait.

Is it not possible to filter using another metric?
Or make --hide-graph keep the same behaviour as with the graph for filtering purposes but simply don't draw it?

Python iotop -o has been showing all processes just fine for me without delayacct (just with a broken swapin column which I don't care about).

[bbonev]

Maybe it will be saner to change the -o logic to do an OR between IO, R and W. What do you think?

[Atemu]

That sounds sane to me; any process that causes any sort of disk IO should show up in -o.

Is the IO metric only for disk IO or also IO to tmpfs and the like?

[bbonev]

@pabs3 I'd appreciate your thoughts on this. IMO, it is better to make it more user friendly and closer to the user's expectations than to blindly stick to exactly what iotop-py is doing

[pabs3]

Yes, definitely better to be user friendly than copying iotop-py.

-- bye, pabs

https://bonedaddy.net/pabs3/

[bbonev]

I believe that this got fixed in b787f78c578563f3c2ca158e02619695167f2b03