mjhjmtu

I need to upload the output of the current script, so I added a trap and set -ex, e.g

#!/bin/bash

exec &> /tmp/error.log
trap 'cat /tmp/error.log; curl http://127.0.0.1/error.php?hostname=$(hostname) -F file=@/tmp/error.log' EXIT

set -ex
wtfwtf

When I execute it, I'm always getting this error, and the PHP script did not receive the whole file

%> cat /tmp/error.log
1.sh: line 6: wtfwtf: command not found
cat: /tmp/error.log: input file is output file

So far, the only solution is to copy the error.log to a new file and upload it, e.g

#!/bin/bash

exec &> /tmp/error.log
trap 'cp /tmp/error.log 123; curl http://127.0.0.1/error.php?hostname=$(hostname) -F file=@123' EXIT

set -ex
wtfwtf

Is there any better way to do this?

With the exec, you are redirecting all output of the script to a specific log file.

In your trap, you want to display the contents of the log file using cat. Since all output is also redirected to that file, GNU cat notices that its input file and standard output stream (which is inherited from the shell) are the same thing, and refuses to perform its task.

BSD cat does not do the same check as GNU cat does, which, if the script is not interrupted, results in an infinitely large log file with the same few lines repeated over and over.

A workaround is to save the original standard output file descriptor, do the redirection as before, and then reinstate it in the trap.

#!/bin/bash

exec 3>&1 # make fd 3 copy of original fd 1
exec >/tmp/error.log 2>&1

# in the trap, make fd 1 copy of fd 3 and close fd 3 (i.e. move fd 3 to fd 1)
trap 'exec 1>&3-; cat /tmp/error.log; curl "http://127.0.0.1/error.php?hostname=$(hostname)" -F file=@/tmp/error.log' EXIT

set -ex
wtfwtf

This makes a copy of file descriptor 1 (as fd 3) before redirecting it to the log file. In the trap, we move this copy back to fd 1 and do the output.

Note that the standard error stream, in the trap, in this example, is still connected to the log file. Therefore, if the curl generates a diagnostic message, this will be saved in the log file rather than being displayed on the terminal (or wherever the original standard error stream was connected to).

Taking the comment from Stéphane Chazelas into account:

#!/bin/sh

exit_handler () 
 # 1. Make standard output be the original standard error
 # (by using fd 3, which is a copy of original fd 2)
 # 2. Do the same with standard error
 # 3. Close fd 3.
 exec >&3 2>&3 3>&-
 cat "$logfile"
 curl "some URL" -F "file=@$logfile"


logfile='/var/log/myscript.log'

# Truncate the logfile.
: >"$logfile"

# 1. Make fd 3 a copy of standard error (fd 2)
# 2. Redirect original standard output to the logfile (appending)
# 3. Redirect original standard error to the logfile (will also append)
exec 3>&2 >>"$logfile" 2>&1

# Use shell function for exit trap (for neatness)
trap exit_handler EXIT

set -ex
wtfwtf

His point is that the logfile is only for diagnostic messages anyway, so it makes more sense to output the logfile to the original standard error stream.

He also points out that it's dangerous to use a fixed filename in a world-writable directory such as /tmp. This is because no check is put in place in the script to make sure that this file does not already exist (someone or some malware could have created a /tmp/error.log symlink to /etc/passwd or your ~/.bashrc for instance). His solution to this is to use a dedicated persistent log file for the script under /var/log instead (the file is persistent, but the contents will be cleared when running the script).

A variation of this would be to use mktemp to create a unique filename under $TMPDIR (and then remove that file in the EXIT trap, unless curl failed in which case the rm would not be executed since set -e is in effect):

#!/bin/sh

exit_handler () 
 # 1. Make standard output be the original standard error
 # (by using fd 3, which is a copy of original fd 2)
 # 2. Do the same with standard error
 # 3. Close fd 3.
 exec >&3 2>&3 3>&-
 cat "$logfile"
 curl "some URL" -F "file=@$logfile"
 rm -f "$logfile"


logfile=$( mktemp )

# 1. Make fd 3 a copy of standard error (fd 2)
# 2. Redirect original standard output to the logfile (appending)
# 3. Redirect original standard error to the logfile (will also append)
exec 3>&2 >>"$logfile" 2>&1

# Use shell function for exit trap (for neatness)
trap exit_handler EXIT

set -ex
wtfwtf

Your second example works, but only because you're not using cat on the log file, not because of copying it.

Minor nitpick: URLs on the command line should probably always be at least double-quoted as they tend to contain characters that the shell may interpret as special (for example ?).