I'm trying to write a Linux shell script (preferably bash),
supposedly named detach.sh, to safely detach programs from a terminal,
such that:
Invocation: ./detach.sh prog [arg1 arg2 ...].
Is exec-able, eg. by running this in your shell:
exec ./detach.sh prog [arg1 arg2 ...]
With proper quoting (mainly handling of arguments containing whitespaces).
Discards the outputs (since they are unneeded).
Does not use screen, tmux, etc.
(same reason with 4, plus no need for an extra babysitting process).
Uses (reasonably) portable commands and programs,
and no things like start-stop-daemon which is quite distro-specific.
I have thought of several ways (shebang lines #!/bin/bash neglected
for the sake of briefness):
nohup:
nohup "$#" >& /dev/null &
disown:
"$#" >& /dev/null &
disown
setsid:
setsid "$#" >& /dev/null &
Using a subshell:
("$#" >& /dev/null &)
nohup/setsid combined with subshell:
# Or alternatively:
# (nohup "$#" >& /dev/null &)
(setsid "$#" >& /dev/null &)
When using gedit as the test program (substituting the "$#" part),
condition 1 can be satisfied with all the above methods,
but condition 2 can be satisfied with none.
However, if an arbitrary program (but not a shell builtin) is appended to script 5,
all the conditions seem to be satisfied (at least for me in the gedit case).
For example:
(setsid "$#" >& /dev/null &)
# Not just `true' because it is also a shell builtin.
/bin/true
Anyone with an idea about an explanation of the above phenomenons
and how to correctly implement the requirements?
EDIT:
With condition 2, I mean the program should be detached from the terminal but runs as usual otherwise. For example, with the gedit case, the condition fails if gedit just exits immediately right after the process of the script has ended.
Upon closer investigation, these previously unnoticed facts were revealed:
Both scripts 3 and 5 (the setsid variant only) will
satisfy all the conditions if a /bin/true is appended to the script.
These scripts, as modified in fact 1, will work as well if
/bin/true is replaced with for i in {0..9999}; do :; done.
Therefore we can conclude that:
(From fact 1)
Multiple levels of detaching (as in script 5) is unnecessary,
and the key is to use the right utility (setsid).
(From fact 2)
A suitable delay before bash exit is necessary for the success of the script.
(Calling external program /bin/true consumes some time,
just like the pure-bash time consumer for i in {0..9999}; do :; done.)
I have not looked at the source code, but I guess a possible explanation
is that bash may exit before setsid finishes configuring the execution
environment of the program to run, if an appropriate delay is not applied.
And finally, an optimal solution should be
#!/bin/bash
setsid "$#" >& /dev/null &
sleep 0.01
EDIT 1:
The necessity of a delay has been explained here. Many thanks to #wilx!
EDIT 2:
(Thanks to #MateiDavid) we seem to have forgotten to redirect the standard input, and a better way would be:
#!/bin/bash
setsid "$#" >& /dev/null < /dev/null &
I think you need to do setsid "$#" >& /dev/null & wait so that the controlling terminal does not disappear before setsid manages to fork the child.
UPDATE
It seems to me that this works both on command line and as argument of -c:
(setsid tail -F /var/log/messages >& /dev/null) & disown
You are trying to create a UNIX daemon process (i.e., a process that has no controlling terminal and that is its own session leader). The setsid command should do this for you, but you are responsible for closing all file descriptors that are open on the terminal you are abandoning. This can be done by redirecting them to /dev/null or using the shell's syntax for closing file descriptors (e.g., 2>&- and 0<&- in Bash).
Related
I have a python file that I run using a .sh file providing some args. When I run this file on the terminal using bash myscript.sh it runs fine, prints progress on the console. When I use nohup like nohup bash myscript.sh > log.log nothing gets saved to log.log file but the processes are running (it's a 4 GPU process and I can see the GPU usage in top and nvidia-smi). As soon as I kill the process using either ctrl+c or kill command, all the output gets printed to the file at once along with keyboard interrupt or process killed message.
I have tried nohup myscript.sh &> log.log nohup myscript.sh &>> log.log but the issue remains the same. What's the reason for such a behaviour?
myscript.sh runs a python file somewhat like
python main.py --arg1 val1 --arg2 val2
I tried
python -u main.py
but it doesn't help. I know the script is working fine as it occupies exactly same amount of memory as it should.
Use stdbuf:
nohup stdbuf -oL bash myscript.sh > log.log
Your problem is related to output buffering. In general, non-interactive output to files tend to be block buffered. The -oL changes output buffering to line mode. There is also the less efficient -o0 to change it to unbuffered.
I know how to redirect stdout to a file:
exec > foo.log
echo test
this will put the 'test' into the foo.log file.
Now I want to redirect the output into the log file AND keep it on stdout
i.e. it can be done trivially from outside the script:
script | tee foo.log
but I want to do declare it within the script itself
I tried
exec | tee foo.log
but it didn't work.
#!/usr/bin/env bash
# Redirect stdout ( > ) into a named pipe ( >() ) running "tee"
exec > >(tee -i logfile.txt)
# Without this, only stdout would be captured - i.e. your
# log file would not contain any error messages.
# SEE (and upvote) the answer by Adam Spiers, which keeps STDERR
# as a separate stream - I did not want to steal from him by simply
# adding his answer to mine.
exec 2>&1
echo "foo"
echo "bar" >&2
Note that this is bash, not sh. If you invoke the script with sh myscript.sh, you will get an error along the lines of syntax error near unexpected token '>'.
If you are working with signal traps, you might want to use the tee -i option to avoid disruption of the output if a signal occurs. (Thanks to JamesThomasMoon1979 for the comment.)
Tools that change their output depending on whether they write to a pipe or a terminal (ls using colors and columnized output, for example) will detect the above construct as meaning that they output to a pipe.
There are options to enforce the colorizing / columnizing (e.g. ls -C --color=always). Note that this will result in the color codes being written to the logfile as well, making it less readable.
The accepted answer does not preserve STDERR as a separate file descriptor. That means
./script.sh >/dev/null
will not output bar to the terminal, only to the logfile, and
./script.sh 2>/dev/null
will output both foo and bar to the terminal. Clearly that's not
the behaviour a normal user is likely to expect. This can be
fixed by using two separate tee processes both appending to the same
log file:
#!/bin/bash
# See (and upvote) the comment by JamesThomasMoon1979
# explaining the use of the -i option to tee.
exec > >(tee -ia foo.log)
exec 2> >(tee -ia foo.log >&2)
echo "foo"
echo "bar" >&2
(Note that the above does not initially truncate the log file - if you want that behaviour you should add
>foo.log
to the top of the script.)
The POSIX.1-2008 specification of tee(1) requires that output is unbuffered, i.e. not even line-buffered, so in this case it is possible that STDOUT and STDERR could end up on the same line of foo.log; however that could also happen on the terminal, so the log file will be a faithful reflection of what could be seen on the terminal, if not an exact mirror of it. If you want the STDOUT lines cleanly separated from the STDERR lines, consider using two log files, possibly with date stamp prefixes on each line to allow chronological reassembly later on.
Solution for busybox, macOS bash, and non-bash shells
The accepted answer is certainly the best choice for bash. I'm working in a Busybox environment without access to bash, and it does not understand the exec > >(tee log.txt) syntax. It also does not do exec >$PIPE properly, trying to create an ordinary file with the same name as the named pipe, which fails and hangs.
Hopefully this would be useful to someone else who doesn't have bash.
Also, for anyone using a named pipe, it is safe to rm $PIPE, because that unlinks the pipe from the VFS, but the processes that use it still maintain a reference count on it until they are finished.
Note the use of $* is not necessarily safe.
#!/bin/sh
if [ "$SELF_LOGGING" != "1" ]
then
# The parent process will enter this branch and set up logging
# Create a named piped for logging the child's output
PIPE=tmp.fifo
mkfifo $PIPE
# Launch the child process with stdout redirected to the named pipe
SELF_LOGGING=1 sh $0 $* >$PIPE &
# Save PID of child process
PID=$!
# Launch tee in a separate process
tee logfile <$PIPE &
# Unlink $PIPE because the parent process no longer needs it
rm $PIPE
# Wait for child process, which is running the rest of this script
wait $PID
# Return the error code from the child process
exit $?
fi
# The rest of the script goes here
Inside your script file, put all of the commands within parentheses, like this:
(
echo start
ls -l
echo end
) | tee foo.log
Easy way to make a bash script log to syslog. The script output is available both through /var/log/syslog and through stderr. syslog will add useful metadata, including timestamps.
Add this line at the top:
exec &> >(logger -t myscript -s)
Alternatively, send the log to a separate file:
exec &> >(ts |tee -a /tmp/myscript.output >&2 )
This requires moreutils (for the ts command, which adds timestamps).
Using the accepted answer my script kept returning exceptionally early (right after 'exec > >(tee ...)') leaving the rest of my script running in the background. As I couldn't get that solution to work my way I found another solution/work around to the problem:
# Logging setup
logfile=mylogfile
mkfifo ${logfile}.pipe
tee < ${logfile}.pipe $logfile &
exec &> ${logfile}.pipe
rm ${logfile}.pipe
# Rest of my script
This makes output from script go from the process, through the pipe into the sub background process of 'tee' that logs everything to disc and to original stdout of the script.
Note that 'exec &>' redirects both stdout and stderr, we could redirect them separately if we like, or change to 'exec >' if we just want stdout.
Even thou the pipe is removed from the file system in the beginning of the script it will continue to function until the processes finishes. We just can't reference it using the file name after the rm-line.
Bash 4 has a coproc command which establishes a named pipe to a command and allows you to communicate through it.
Can't say I'm comfortable with any of the solutions based on exec. I prefer to use tee directly, so I make the script call itself with tee when requested:
# my script:
check_tee_output()
{
# copy (append) stdout and stderr to log file if TEE is unset or true
if [[ -z $TEE || "$TEE" == true ]]; then
echo '-------------------------------------------' >> log.txt
echo '***' $(date) $0 $# >> log.txt
TEE=false $0 $# 2>&1 | tee --append log.txt
exit $?
fi
}
check_tee_output $#
rest of my script
This allows you to do this:
your_script.sh args # tee
TEE=true your_script.sh args # tee
TEE=false your_script.sh args # don't tee
export TEE=false
your_script.sh args # tee
You can customize this, e.g. make tee=false the default instead, make TEE hold the log file instead, etc. I guess this solution is similar to jbarlow's, but simpler, maybe mine has limitations that I have not come across yet.
Neither of these is a perfect solution, but here are a couple things you could try:
exec >foo.log
tail -f foo.log &
# rest of your script
or
PIPE=tmp.fifo
mkfifo $PIPE
exec >$PIPE
tee foo.log <$PIPE &
# rest of your script
rm $PIPE
The second one would leave a pipe file sitting around if something goes wrong with your script, which may or may not be a problem (i.e. maybe you could rm it in the parent shell afterwards).
function run () {
nohup python $1 > nohup.out &
}
On the command line I call this as "run scriptname.py" and bash executes the following command:
python scriptname.py > nohup.out &
Can you help me translate this to fish.
I have this so far..
function run
bash -c "nohup python $1 > nohup.out &"
end
When I call source on ~/.config/fish/config.fish
This exists simply saying
Error when reading file: ~/.config/fish/config.fish
without providing any helpful hints as to what the error is.
There's really no need to execute bash here, fish can also execute nohup, the redirections also work and such.
There's a minor difference in that, instead of $1 and $2 and so on, arguments to fish functions are stored in the $argv array.
function run
nohup python $argv > nohup.out &
end
This will expand $argv to all elements of that as one element each, so run script.py banana would run nohup python script.py banana > nohup.out &. If you truly want just one argument to be passed, you'd need $argv[1].
I actually have no idea why your definition should cause an error when sourcing config.fish - which fish version are you using?
This is a perfectly valid (and more correct) replacement for your function in fish:
function run
bash -c 'nohup python "$#" > nohup.out &' _ $argv
end
This is an equivalent to the native-bash function:
run() {
nohup python "$#" </dev/null >nohup.out 2>&1 &
}
...which, personally, I would suggest rewriting to use disown rather than nohup.
With respect to the error seen from fish, I'd suggest paying attention to any other (not syntax-related) which may have impacted whether your file could be read -- permissions, etc.
As part of a bash script, I want to run a program repeatedly, and redirect the output to less. The program has an interactive element, so the goal is that when you exit the program via the window's X button, it is restarted via the script. This part works great, but when I use a pipe to less, the program does not automatically restart until I go to the console and press q. The relevant part of the script:
while :
do
program | less
done
I want to make less quit itself when the pipe closes, so that the program restarts without any user intervention. (That way it behaves just as if the pipe was not there, except while the program is running you can consult the console to view the output of the current run.)
Alternative solutions to this problem are also welcome.
Instead of exiting less, could you simply aggregate the output of each run of program?
while :
do
program
done | less
Having less exit when program would be at odds with one useful feature of less, which is that it can buffer the output of a program that exits before you finish reading its output.
UPDATE: Here's an attempt at using a background process to kill less when it is time. It assumes that the only program reading the output file is the less to kill.
while :
do
( program > /tmp/$$-program-output; kill $(lsof -Fp | cut -c2-) ) &
less /tmp/$$-program-output
done
program writes its output to a file. Once it exits, the kill command uses lsof to
find out what process is reading the file, then kills it. Note that there is a race condition; less needs to start before program exists. If that's a problem, it can
probably be worked around, but I'll avoid cluttering the answer otherwise.
You may try to kill the process group program and less belong to instead of using kill and lsof.
#!/bin/bash
trap 'kill 0' EXIT
while :
do
# script command gives sh -c own process group id (only sh -c cmd gets killed, not entire script!)
# FreeBSD script command
script -q /dev/null sh -c '(trap "kill -HUP -- -$$" EXIT; echo hello; sleep 5; echo world) | less -E -c'
# GNU script command
#script -q -c 'sh -c "(trap \"kill -HUP -- -$$\" EXIT; echo hello; sleep 5; echo world) | less -E -c"' /dev/null
printf '\n%s\n\n' "you now may ctrl-c the program: $0" 1>&2
sleep 3
done
While I agree with chepner's suggestion, if you really want individual less instances, I think this item for the man page will help you:
-e or --quit-at-eof
Causes less to automatically exit the second time it reaches end-of-file. By default,
the only way to exit less is via the "q" command.
-E or --QUIT-AT-EOF
Causes less to automatically exit the first time it reaches end-of-file.
you would make this option visible to less in the LESS envir variable
export LESS="-E"
while : ; do
program | less
done
IHTH
In a C program I can write argv[0] and the new name shows up in a ps listing.
How can I do this in bash?
You can do it when running a new program via exec -a <newname>.
Just for the record, even though it does not exactly answer the original poster's question, this is something trivial to do with zsh:
ARGV0=emacs nethack
I've had a chance to go through the source for bash and it does not look like there is any support for writing to argv[0].
I'm assuming you've got a shell script that you wish to execute such that the script process itself has a new argv[0]. For example (I've only tested this in bash, so i'm using that, but this may work elsewhere).
#!/bin/bash
echo "process $$ here, first arg was $1"
ps -p $$
The output will be something like this:
$ ./script arg1
process 70637 here, first arg was arg1
PID TTY TIME CMD
70637 ttys003 0:00.00 /bin/bash ./script arg1
So ps shows the shell, /bin/bash in this case. Now try your interactive shell's exec -a, but in a subshell so you don't blow away the interactive shell:
$ (exec -a MyScript ./script arg1)
process 70936 here, first arg was arg1
PID TTY TIME CMD
70936 ttys008 0:00.00 /bin/bash /path/to/script arg1
Woops, still showing /bin/bash. what happened? The exec -a probably did set argv[0], but then a new instance of bash started because the operating system read #!/bin/bash at the top of your script. Ok, what if we perform the exec'ing inside the script somehow? First, we need some way of detecting whether this is the "first" execution of the script, or the second, execed instance, otherwise the second instance will exec again, and on and on in an infinite loop. Next, we need the executable to not be a file with a #!/bin/bash line at the top, to prevent the OS from changing our desired argv[0]. Here's my attempt:
$ cat ./script
#!/bin/bash
__second_instance="__second_instance_$$"
[[ -z ${!__second_instance} ]] && {
declare -x "__second_instance_$$=true"
exec -a MyScript "$SHELL" "$0" "$#"
}
echo "process $$ here, first arg was $1"
ps -p $$
Thanks to this answer, I first test for the environment variable __second_instance_$$, based on the PID (which does not change through exec) so that it won't collide with other scripts using this technique. If it's empty, I assume this is the first instance, and I export that environment variable, then exec. But, importantly, I do not exec this script, but I exec the shell binary directly, with this script ($0) as an argument, passing along all the other arguments as well ($#). The environment variable is a bit of a hack.
Now the output is this:
$ ./script arg1
process 71143 here, first arg was arg1
PID TTY TIME CMD
71143 ttys008 0:00.01 MyScript ./script arg1
That's almost there. The argv[0] is MyScript like I want, but there's that extra arg ./script in there which is a consequence of executing the shell directly (rather than via the OS's #! processing). Unfortunately, I don't know how to get any better than this.
Update for Bash 5.0
Looks like Bash 5.0 adds support for writing to special variable BASH_ARGV0, so this should become far simpler to accomplish.
(see release announcement)
( exec -a foo bash -c 'echo $0' )
ps and others inspect two things, none of which is argv0: /proc/PID/comm (for the "process name") and /proc/PID/cmdline (for the command-line). Assigning to argv0 will not change what ps shows in the CMD column, but it will change what the process usually sees as its own name (in output messages, for example).
To change the CMD column, write to /proc/PID/comm:
echo -n mynewname >/proc/$$/comm; ps
You cannot write to or modify /proc/PID/cmdline in any way.
Process can set their own "title" by writing to the memory area in which argv & envp are located (note that this is different than setting BASH_ARGV0). This has the side effect of changing /proc/PID/cmdline as well, which is what some daemons do in order to prettify (hide?) their command lines. libbsd's setproctitle() does exactly that, but you cannot do that in Bash without support of external tools.
I will just add that this must be possible at runtime, at least in some environments. Assigning $0 in perl on linux does change what shows up in ps. I do not know how that is implemented, however. If I can find out, i'll update this.
edit:
Based on how perl does it, it is non-trivial. I doubt there is any bask built in way at runtime but don't know for sure. You can see how perl does sets the process name at runtime.
Copy the bash executable to a different name.
You can do this in the script itself...
cp /bin/bash ./new-name
PATH=$PATH:.
exec new-name $0
If you are trying to pretend you are not a shell script you can rename the script itself to something cool or even " " (a single space) so
exec new-name " "
Will execute bash your script and appears in the ps list as just new-name.
OK so calling a script " " is a very bad idea :)
Basically, to change the name
bash script
rename bash and rename the script.
If you are worried, as Mr McDoom. apparently is, about copying a binary to a new name (which is entirely safe) you could also create a symlink
ln -s /bin/bash ./MyFunkyName
./MyFunkyName
This way, the symlink is what appears in the ps list. (again use PATH=$PATH:. if you dont want the ./)