Slow load time of bash in cygwin

Question 1

As you've noted in your answer, the problem is Cygwin's bash-completion package. The quick and easy fix is to disable bash-completion, and the correct way to do that is to run Cygwin's setup.exe (download it again if you need to) and select to uninstall that package.

The longer solution is to work through the files in /etc/bash_completion.d and disable the ones you don't need. On my system, the biggest culprits for slowing down Bash's load time (mailman, shadow, dsniff and e2fsprogs) all did exactly nothing, since the tools they were created to complete weren't installed.

If you rename a file in /etc/bash_completion.d to have a .bak extension, it'll stop that script being loaded. Having disabled all but a select 37 scripts on one of my systems in that manner, I've cut the average time for bash_completion to load by 95% (6.5 seconds to 0.3 seconds).

Question 2

In my case that was windows domain controller. I did this to find the issue:

I started with a simple, windows cmd.exe and the, typed this:
c:\cygwin\bin\strace.exe c:\cygwin\bin\bash

In my case, I noticed a following sequence:

    218   12134 [main] bash 11304 transport_layer_pipes::connect: Try to connect to named pipe: \\.\pipe\cygwin-c5e39b7a9d22bafb-lpc
     45   12179 [main] bash 11304 transport_layer_pipes::connect: Error opening the pipe (2)
     39   12218 [main] bash 11304 client_request::make_request: cygserver un-available
1404719 1416937 [main] bash 11304 pwdgrp::fetch_account_from_windows: line: <CENSORED_GROUP_ID_#1>
    495 1417432 [main] bash 11304 pwdgrp::fetch_account_from_windows: line: <CENSORED_GROUP_ID_#2>
    380 1417812 [main] bash 11304 pwdgrp::fetch_account_from_windows: line: <CENSORED_GROUP_ID_#3>

    etc...

The key thing was identifying the client_request::make_request: cygserver un-available line. You can see, how after that, cygwin tries to fetch every single group from windows, and execution times go crazy.

A quick google revealed what a cygserver is: https://cygwin.com/cygwin-ug-net/using-cygserver.html

Cygserver is a program which is designed to run as a background service. It provides Cygwin applications with services which require security arbitration or which need to persist while no other cygwin application is running.

The solution was, to run the cygserver-config and then net start cygserver to start the Windows service. Cygwin startup times dropped significantly after that.

Question 3

All of the answers refer to older versions of bash_completion, and are irrelevant for recent bash_completion.

Modern bash_completion moved most of the completion files to /usr/share/bash-completion/completions by default, check the path on your system by running

# pkg-config --variable=completionsdir bash-completion
/usr/share/bash-completion/completions

There are many files in there, one for each command, but that is not a problem, since they are loaded on demand the first time you use completion with each command. The old /etc/bash_completion.d is still supported for compatibility, and all files from there are loaded when bash_completion starts.

# pkg-config --variable=compatdir bash-completion
/etc/bash_completion.d

Use this script to check if there are any stale files left in the old dir.

#!/bin/sh
COMPLETIONS_DIR="$(pkg-config --variable=completionsdir bash-completion)"
COMPAT_DIR="$(pkg-config --variable=compatdir bash-completion)"
for file in "${COMPLETIONS_DIR}"/*; do
    file="${COMPAT_DIR}/${file#${COMPLETIONS_DIR}/}"
    [ -f "$file" ] && printf '%s\n' $file
done

It prints the list of files in compat dir that are also present in the newer (on-demand) completions dir. Unless you have specific reasons to keep some of them, review, backup and remove all of those files.

As a result, the compat dir should be mostly empty.

Now, for the most interesting part - checking why bash startup is slow. If you just run bash, it will start a non-login, interactive shell - this one on Cygwin source /etc/bash.bashrc and then ~/.bashrc. This most likely doesn't include bash completion, unless you source it from one of rc files. If you run bash -l (bash --login), start Cygwin Terminal (depends on your cygwin.bat), or log in via SSH, it will start a login, interactive shell - which will source /etc/profile, ~/.bash_profile, and the aforementioned rc files. The /etc/profile script itself sources all executable .sh files in /etc/profile.d.

You can check how long each file takes to source. Find this code in /etc/profile:

for file in /etc/profile.d/*.$1; do
  [ -e "${file}" ] && . "${file}"
done

Back it up, then replace it with this:

for file in /etc/profile.d/*.$1; do
  TIMEFORMAT="%3lR ${file}"
  [ -e "${file}" ] && time . "${file}"
done

Start bash and you will see how long each file took. Investigate files that take a significant amount of time. In my case, it was bash_completion.sh and fzf.sh (fzf is fuzzy finder, a really nice complement to bash_completion). Now the choice is to disable it or investigate further. Since I wanted to keep using fzf shortcuts in bash, I investigated, found the source of the slowdown, optimized it, and submitted my patch to fzf's repo (hopefully it will be accepted).

Now for the biggest time spender - bash_completion.sh. Basically that script sources /usr/share/bash-completion/bash_completion. I backed up that file, then edited it. On the last page there is for loop that sources all the files in compat dir - /etc/bash_completion.d. Again, I added TIMEFORMAT and time, and saw which script was causing the slow starting. It was zzz-fzf (fzf package). I investigated and found a subshell ($()) being executed multiple times in a for loop, rewrote that part without using a subshell, making the script work quickly. I already submitted my patch to fzf's repo.

The biggest reason for all these slowdowns is: fork is not supported by Windows process model, Cygwin did a great job emulating it, but it's painfully slow compared to a real UNIX. A subshell or a pipeline that does very little work by itself spends most of it's execution time for fork-ing. E.g. compare the execution times of time echo msg (0.000s on my Cygwin) vs time echo $(echo msg) (0.042s on my Cygwin) - day and night. The echo command itself takes no appreciable time to execute, but creating a subshell is very expensive. On my Linux system, these commands take 0.000s and 0.001s respectively. Many packages Cygwin has are developed by people who use Linux or other UNIX, and can run on Cygwin unmodified. So naturally these devs feel free to use subshells, pipelines and other features wherever convenient, since they don't feel any significant performance hit on their system, but on Cygwin those shell scripts might run tens and hundreds of times slower.

Bottom line, if a shell script works slowly in Cygwin - try to locate the source of fork calls and rewrite the script to eliminate them as much as possible. E.g. cmd="$(printf "$1" "$2")" (uses one fork for subshell) can be replaced with printf -v cmd "$1" "$2".

Boy, it came out really long. Any people still reading up to here are real heros. Thanks :)

Question 4

I know this is an old thread, but after a fresh install of Cygwin this week I'm still having this problem.

Instead of handpicking all of the bash_completion files, I used this line to implement @me_and's approach for anything that isn't installed on my machine. This significantly reduced the startup time of bash for me.

In /etc/bash_completion.d, execute the following:

for i in $(ls|grep -v /); do type $i >/dev/null 2>&1 || mv $i $i.bak; done

Question 5

New answer for an old thread, relating to the PATH of the original question.

Most of the other answers deal with the bash startup. If you're seeing a slow load time when you run bash -i within the shell, those may apply.

In my case, bash -i ran fast, but anytime I opened a new shell (be it in a terminal or in xterm), it took a really long time. If bash -l is taking a long time, it means it's the login time.

There are some approaches at the Cygwin FAQ at https://cygwin.com/faq/faq.html#faq.using.startup-slow but they didn't work for me.

The original poster asked about the PATH, which he diagnosed using bash -x. I too found that although bash -i was fast, bash -xl was slow and showed a lot of information about the PATH.

There was such a ridiculously long Windows PATH that the login process kept on running programs and searching the entire PATH for the right program.

My solution: Edit the Windows PATH to remove anything superfluous. I'm not sure which piece I removed that did the trick, but the login shell startup went from 6 seconds to under 1 second.

YMMV.

Question 6

My answer is the same as npe's above. But, since I just joined, I cannot comment or even upvote it! I hope this post doesn't get deleted because it offer reassurance for anyone looking for an answer to the same problem.

npe's solution worked for me. There's only one caveat - I had to close all cygwin processes before I got the best out of it. That includes running cygwin services, like sshd, and the ssh-agent that I start from my login scripts. Before that, the window for the cygwin terminal would appear instantly but hang for several seconds before it presents the prompt. And it hanged for several seconds upon closing the window. After I killed all processes and started the cygserver service (btw I prefer to use the Cygwin way - 'cygrunsrv -S cygserver', than 'net start cygserver'; I don't know if it makes any practical difference) it starts immediately. So thanks to npe again!

Question 7

I'm on a corporate network with a pretty complicated setup, and it seems that really kills cygwin startup times. Related to npe's answer, I also had to follow some of the steps laid out here: https://cygwin.com/faq/faq.html#faq.using.startup-slow

Another cause for AD client system is slow DC replies, commonly observed in configurations with remote DC access. The Cygwin DLL queries information about every group you're in to populate the local cache on startup. You may speed up this process a little by caching your own information in local files. Run these commands in a Cygwin terminal with write access to /etc:

getent passwd $(id -u) > /etc/passwd
getent group $(id -G) > /etc/group

Also, set /etc/nsswitch.conf as follows:

passwd: files db
group: files db

This will limit the need for Cygwin to contact the AD domain controller (DC) while still allowing for additional information to be retrieved from DC, such as when listing remote directories.

After doing that plus starting the cygserver my cygwin startup time dropped significantly.

Question 8

As someone mentioned above, one possible issue is the PATH environment variable contains too much path, cygwin will search all of them. I prefer direct edit the /etc/profile, just overwrite the PATH variable to cygwin related path, e.g. PATH="/usr/local/bin:/usr/bin". Add additional path if you want.

Question 9

I wrote a Bash function named 'minimizecompletion' for inactivating not needed completion scripts.

Completion scripts can add more than one completion specification or have completion specifications for shell buildins, therefore it is not sufficient to compare script names with executable files found in $PATH.

My solution is to remove all loaded completion specifications, to load a completion script and check if it has added new completion specifications. Depending on this it is inactivated by adding .bak to the script file name or it is activated by removing .bak. Doing this for all 182 scripts in /etc/bash_completion.d results in 36 active and 146 inactive completion scripts reducing the Bash start time by 50% (but it should be clear this depends on installed packages).

The function also checks inactivated completion scripts so it can activate them when they are needed for new installed Cygwin packages. All changes can be undone with argument -a that activates all scripts.

# Enable or disable global completion scripts for speeding up Bash start.
#
# Script files in directory '/etc/bash_completion.d' are inactived
# by adding the suffix '.bak' to the file name; they are activated by
# removing the suffix '.bak'. After processing all completion scripts
# are reloaded by calling '/etc/bash_completion'
#
# usage:  [-a]
#         -a  activate all completion scripts
# output: statistic about total number of completion scripts, number of
#         activated, and number of inactivated completion scripts; the
#         statistic for active and inactive completion scripts can be
#         wrong when 'mv' errors occure
# return: 0   all scripts are checked and completion loading was
#             successful; this does not mean that every call of 'mv'
#             for adding or removing the suffix was successful
#         66  the completion directory or loading script is missing
#
minimizecompletion() {
  local arg_activate_all=${1-}
  local completion_load=/etc/bash_completion
  local completion_dir=/etc/bash_completion.d

  (
    # Needed for executing completion scripts.
    #
    local UNAME='Cygwin'
    local USERLAND='Cygwin'
    shopt -s extglob progcomp
    have() {
      unset -v have
      local PATH="$PATH:/sbin:/usr/sbin:/usr/local/sbin"
      type -- "$1" &>/dev/null && have='yes'
    }

    # Print initial statistic.
    #
    printf 'Completion scripts status:\n'
    printf '  total:       0\n'
    printf '  active:      0\n'
    printf '  inactive:    0\n'
    printf 'Completion scripts changed:\n'
    printf '  activated:   0\n'
    printf '  inactivated: 0\n'

    # Test the effect of execution for every completion script by
    # checking the number of completion specifications after execution.
    # The completion scripts are renamed depending on the result to
    # activate or inactivate them.
    #
    local completions total=0 active=0 inactive=0 activated=0 inactivated=0
    while IFS= read -r -d '' f; do
      ((++total))
      if [[ $arg_activate_all == -a ]]; then
        [[ $f == *.bak ]] && mv -- "$f" "${f%.bak}" && ((++activated))
        ((++active))
      else
        complete -r
        source -- "$f"
        completions=$(complete | wc -l)
        if (( $completions > 0 )); then
          [[ $f == *.bak ]] && mv -- "$f" "${f%.bak}" && ((++activated))
          ((++active))
        else
          [[ $f != *.bak ]] && mv -- "$f" "$f.bak" && ((++inactivated))
          ((++inactive))
        fi
      fi
      # Update statistic.
      #
      printf '\r\e[6A\e[15C%s' "$total"
      printf '\r\e[1B\e[15C%s' "$active"
      printf '\r\e[1B\e[15C%s' "$inactive"
      printf '\r\e[2B\e[15C%s' "$activated"
      printf '\r\e[1B\e[15C%s' "$inactivated"
      printf '\r\e[1B'
    done < <(find "$completion_dir" -maxdepth 1 -type f -print0)

    if [[ $arg_activate_all != -a ]]; then
      printf '\nYou can activate all scripts with %s.\n' "'$FUNCNAME -a'"
    fi
    if ! [[ -f $completion_load && -r $completion_load ]]; then
      printf 'Cannot reload completions, missing %s.\n' \
             "'$completion_load'" >&2
      return 66
    fi
  )

  complete -r
  source -- "$completion_load"
}

This is an example output and the resulting times:

$ minimizecompletion -a
Completion scripts status:
  total:       182
  active:      182
  inactive:    0
Completion scripts changed:
  activated:   146
  inactivated: 0

$ time bash -lic exit
logout

real    0m0.798s
user    0m0.263s
sys     0m0.341s

$ time minimizecompletion
Completion scripts status:
  total:       182
  active:      36
  inactive:    146
Completion scripts changed:
  activated:   0
  inactivated: 146

You can activate all scripts with 'minimizecompletion -a'.

real    0m17.101s
user    0m1.841s
sys     0m6.260s

$ time bash -lic exit
logout

real    0m0.422s
user    0m0.092s
sys     0m0.154s