High cpu usage on directory change in mc with fish subshell

[olfway]

I'm using fish and midnight commander from homebrew on mac

fish, version 3.1.0
TERM = xterm-256color

I removed ~/.config/fish to test with empty config

If I run mc with fish subshell and press enter multiple times on some directory fish subshell start using ~130% cpu for some time

I tried to build fish with debug symbols and get backtrace while it was using cpu:

(lldb) bt

* thread #1, queue = 'com.apple.main-thread', stop reason = signal SIGSTOP
  * frame #0: 0x00007fff73a225be libsystem_kernel.dylib`__select + 10
    frame #1: 0x0000000105e65cf1 fish`input_event_queue_t::readb(this=0x00007fca45000a50) at input_common.cpp:78:15
    frame #2: 0x0000000105e6676a fish`input_event_queue_t::readch(this=0x00007fca45000a50) at input_common.cpp:175:20
    frame #3: 0x0000000105e55ab7 fish`inputter_t::readch(this=0x00007fca45000a50, allow_commands=true) at input.cpp:495:33
    frame #4: 0x0000000105f18e8e fish`reader_data_t::read_normal_chars(this=0x00007fca45000618, rls=0x00007ffee9fa0070) at reader.cpp:2393:61
    frame #5: 0x0000000105f1ce70 fish`reader_data_t::readline(this=0x00007fca45000618, nchars_or_0=0) at reader.cpp:3246:38
    frame #6: 0x0000000105f1da3a fish`reader_readline(nchars=0) at reader.cpp:3388:72
    frame #7: 0x0000000105f1e277 fish`read_i(parser=0x00007fca44403cf0) at reader.cpp:2281:33
    frame #8: 0x0000000105f1df26 fish`reader_read(parser=0x00007fca44403cf0, fd=0, io=0x00007ffee9fa05d8) at reader.cpp:3584:37
    frame #9: 0x0000000105c619b8 fish`main(argc=1, argv=0x00007ffee9fa07c8) at fish.cpp:494:19
    frame #10: 0x00007fff738d97fd libdyld.dylib`start + 1
    frame #11: 0x00007fff738d97fd libdyld.dylib`start + 1

[ridiculousfish]

Thanks for filing, I'll try to reproduce!

[ridiculousfish]

What command do you run? I tried:

echo (mc)

and this resulted in "Cannot get terminal settings: Inappropriate ioctl for device (25)" but no high CPU usage.

[olfway]

I run it like

env SHELL=/opt/fish/bin/fish /opt/mc/bin/mc

and then move cursor on any folder and press enter 10-20 times
(so mc enters that folder and then goes back)

[krobelus]

I can reproduce on 3.1.0. Seems to be fixed on current master, probably by e334bec.

With latest master I get around 50% CPU utilization when pressing Enter repeatedly, and around half of the time seems to be spent on UTF8 -> internal UTF32 conversion if I measured correctly. These are directories with 100-200 files. Anyway that doesn't seem harmful.

[olfway]

I found .local/share/fish/fish_history file which is 3.8Mb, ~100k lines

No more high cpu usage after I renamed it

[krobelus]

@olfway great point! This was caused by autosuggestions. Apparently, mc writes commands like
cd (printf '%b' '\0057home\0057johannes\0057git\0057fish\0055shell\0057build') to the fish process, and immediately clears the command line. Note the leading space, which is why it won't be added to the history. Because it doesn't prefix-match any item in the history, fish has to search through the entire history to compute autosuggestions.

Autosuggestions are generally very useful but we can definitely disable history autosuggestion if the command line starts with a space, so this should be fixed now by 390647a.

[zanchey]

Shouldn't the search ignore the leading space?

[krobelus]

I see, it could be convenient to type a leading space and still have history autosuggestions available. Currently you'd have to type the command first with autosuggestions and then prepend the space.

Maybe there's another workaround for mc? I think we could do some minimal caching for failing autosuggestions: after a history search for autosuggestions fails, as long as subsequent edits to the command line only append to it, don't run history autosuggestion again. It'd give not that much improvement, since e334bec already kind of does that, but probably still reduces the load.