Two shebang papercuts

2019 Nov 08

You’ve written a Ruby script on your Mac. The script demands great performance, so you add --enable=jit to its shebang:

#!/usr/bin/env ruby --enable=jit

print "Hello, world!"

You test on your Mac and the program greets everybody in record time. Ship it!

Some time later, a user reports a problem. When they run the script, it seems rather slow; in fact, they’ve waited for hours and no one has been greeted.

The script still works (and it’s fast!) on your machine, so you ask the user about their environment. One difference stands out: they’re using Linux. So you spin up a Linux box and try the script. Sure enough, it seems stuck. You reach for strace, and see the same spew looping ad nauseam:

you@yourbox:~$ sudo strace ./hello
execve("./hello", ["./hello"], 0x7ffe721b38b0 /* 52 vars */) = 0
brk(NULL)                               = 0x55b53fddf000
arch_prctl(0x3001 /* ARCH_??? */, 0x7fff6a434040) = -1 EINVAL (Invalid argument)
access("/etc/ld.so.preload", R_OK)      = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=211342, ...}) = 0
mmap(NULL, 211342, PROT_READ, MAP_PRIVATE, 3, 0) = 0x7f4088423000
close(3)                                = 0
openat(AT_FDCWD, "/usr/lib/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
read(3, "\177ELF\2\1\1\3\0\0\0\0\0\0\0\0\3\0>\0\1\0\0\0`r\2\0\0\0\0\0"..., 832) = 832
lseek(3, 64, SEEK_SET)                  = 64
read(3, "\6\0\0\0\4\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0"..., 784) = 784
lseek(3, 848, SEEK_SET)                 = 848
read(3, "\4\0\0\0\20\0\0\0\5\0\0\0GNU\0\2\0\0\300\4\0\0\0\3\0\0\0\0\0\0\0", 32) = 32
lseek(3, 880, SEEK_SET)                 = 880
read(3, "\4\0\0\0\24\0\0\0\3\0\0\0GNU\0003\321\363P\3617(e\35t\335*V\272\321\344"..., 68) = 68
fstat(3, {st_mode=S_IFREG|0755, st_size=2149496, ...}) = 0
mmap(NULL, 8192, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f4088421000
lseek(3, 64, SEEK_SET)                  = 64
read(3, "\6\0\0\0\4\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0"..., 784) = 784
lseek(3, 848, SEEK_SET)                 = 848
read(3, "\4\0\0\0\20\0\0\0\5\0\0\0GNU\0\2\0\0\300\4\0\0\0\3\0\0\0\0\0\0\0", 32) = 32
lseek(3, 880, SEEK_SET)                 = 880
read(3, "\4\0\0\0\24\0\0\0\3\0\0\0GNU\0003\321\363P\3617(e\35t\335*V\272\321\344"..., 68) = 68
mmap(NULL, 1860536, PROT_READ, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x7f408825a000
mprotect(0x7f408827f000, 1671168, PROT_NONE) = 0
mmap(0x7f408827f000, 1363968, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x25000) = 0x7f408827f000
mmap(0x7f40883cc000, 303104, PROT_READ, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x172000) = 0x7f40883cc000
mmap(0x7f4088417000, 24576, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x1bc000) = 0x7f4088417000
mmap(0x7f408841d000, 13240, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x7f408841d000
close(3)                                = 0
arch_prctl(ARCH_SET_FS, 0x7f4088422580) = 0
mprotect(0x7f4088417000, 12288, PROT_READ) = 0
mprotect(0x55b53f654000, 4096, PROT_READ) = 0
mprotect(0x7f4088481000, 4096, PROT_READ) = 0
munmap(0x7f4088423000, 211342)          = 0
brk(NULL)                               = 0x55b53fddf000
brk(0x55b53fe00000)                     = 0x55b53fe00000
openat(AT_FDCWD, "/usr/lib/locale/locale-archive", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=3035216, ...}) = 0
mmap(NULL, 3035216, PROT_READ, MAP_PRIVATE, 3, 0) = 0x7f4087f74000
close(3)                                = 0
execve("./hello", ["./hello"], 0x7ffe721b38b0 /* 52 vars */) = 0
brk(NULL)                               = 0x55b53fddf000
arch_prctl(0x3001 /* ARCH_??? */, 0x7fff6a434040) = -1 EINVAL (Invalid argument)
...

What’s going on? The script is just exec-ing itself in a seemingly infinite loop, and not really doing anything else. Why?

It turns out this comes from combining two papercuts, one which applies to shebangs in general, and another which applies to any which invoke /usr/bin/env.

the multiple argument shebang papercut

You can think of the shebang #!/usr/bin/env ruby --enable=jit as having two parts:

Among other differences in shebangs, Linux and macOS treat the remainder differently. Linux parses the remainder into exactly one argument. In our ./hello example, Linux parses the shebang’s remainder to "ruby --enable=jit", combines it with the user’s arguments {"./hello"}, and executes "/usr/bin/env" with arguments {"/usr/bin/env", "ruby --enable=jit", "./hello"}.

macOS splits the remainder into several arguments if there are spaces. For ./hello, macOS parses the remainder to {"ruby", "--enable=jit"} and executes "/usr/bin/env" with arguments {"env", "ruby", "--enable=jit", "./hello"}.

It seems like the script would fail on Linux with this shebang, because it’s not asking env to find and run ruby. But it’s not saying something like, program not found: ruby --enable=jit, it’s hanging in that loop. Something else is afoot.

the #!/usr/bin/env shebang papercut

The first papercut doesn’t fully explain this loopy behavior. Another paper-sharp shebang, which also tries to pass two arguments to /usr/bin/env, fails differently on Linux:

#!/usr/bin/env ruby --verbose

echo "Goodbye, world"
you@yourbox:~$ ./goodbye
/usr/bin/env: ‘ruby --verbose’: No such file or directory

The difference is in the = character. Besides finding the first interpreter on a user’s PATH, env has another use: to “run a program in a modified environment”. One way it does this is to take arguments of the form VARIABLE=VALUE before the program and its arguments, and set those environment variables in the executed program’s process. For instance:

you@yourbox:~$ env DISPLAY=:0 xeyes -fg dodgerblue

executes {"xeyes", "-fg", "dodgerblue"} with the shell’s environment, plus the environment variable "DISPLAY" set to ":0".

Combining these papercuts, when Linux comes to execute ./hello:

  1. An exec-family function reads "./hello"’s contents and parses the shebang into {"/usr/bin/env", "ruby --enable=jit"}.
  2. The parsed shebang is combined with the argv passed to exec*() to ultimately execute {"/usr/bin/env", "ruby --enable=jit", "./hello"}.
  3. env parses its arguments, sets the environment variable "ruby --enable" to the value "jit" and executes the program "./hello".
  4. GOTO 1

./goodbye fails differently because env interprets the argument ruby --verbose as the program to run, since there’s no = character in it.

a solution: -S

Fortunately, modern versions of env have a solution. macOS and FreeBSD support the -S option, which does, quote the macOS manual:

   -S string
           Split apart the given string into multiple strings, and process each of the
           resulting strings as separate arguments to the env utility.  The -S option
           recognizes some special character escape sequences and also supports environment-
           variable substitution, as described below.

On Linux, GNU’s coreutils added the same flag to env in version 8.30, released July 2018.

So you can write a fancy, optionful shebang like:

#!/usr/bin/env -S ruby --enable=jit --verbose

puts "Hello, world!"

And ./hello will execute what you wanted.

references

Andries E. Brouwer posted a good writeup of shebang behavior on many unix-ish platforms.