fix non-responsive Ctrl-C during disk-usage sorting

[danfarino]

Ctrl-C wasn't being noticed during a disk-usage sorting, effectively meaning there was no way to cancel the operation except to wait it out.

Tested on Ubuntu 20.04 and an old MacOS 10.11.6 (El Capitan).

Thanks for a great program!
-Dan

[jarun]

We want to skip a check at every file. Can we do this when we encounter a directory instead?

Please share the performance data of current and new in cache-dropped state as well.

[jarun]

Unrelated to this PR, currently 1 thread processes a sub-directory tree under the directory where you do du. Say, at /, 1 thread processes each thread. The problem is - /bin pr /lib takes muck longer. See if we can do 1 dir (not the full tree) per thread. I guess if we try that, we will have to store the subtree's top directory info with the directories as the size should reflect for that dir visually.

[jarun]

We can't add a full uint here. Why is the bit not working? It is written in global scope and the threads only read it.

[jarun]

I don't think the threads need to read this atomically as none of them changes it.

[danfarino]

Thanks for taking a look. I'm pretty new to nnn, and don't have the bandwidth to really dig in and learn the code in depth. That said, I saw that g_state.interrupt was being written to in the signal handler but those writes weren't being noticed in other parts of the program. Adding atomic operations fixed that issue on my machines. I had to make interrupt a uint_t instead of a bit field so that it would work with the atomic operations. (It's been years since I've done any serious C programming, so if I missed an obvious way to make it work without changing that, even better.)

I added the atomic operations around access to active_threads since I saw while (active_threads == NUM_DU_THREADS); and spinlocks definitely need atomic operations to work properly.

Please feel free to incorporate whatever ideas from this PR you'd like, and change the rest.

Thanks!

[jarun]

That said, I saw that g_state.interrupt was being written to in the signal handler

That's global. Also we clear in global scope. Currently dirwalk doesn't add a new dir when the interrupt is set.

I'm pretty new to nnn, and don't have the bandwidth to really dig in and learn the code in depth.

Please feel free to incorporate whatever ideas from this PR you'd like, and change the rest.

I will refer to this PR and incorporate the changes when I get the time.

[jarun]

Commit 4c705a9 should work for you.

[danfarino]

Yes! Works great so far. Much simpler than what I had. Thanks!

[jarun]

Thanks for the confirmation.