tail does not block

As always: For everything there is an answer which is short, easy to understand, easy to follow and completely wrong. Here tail -f /dev/null falls into this category 😉

If you look at it with strace tail -f /dev/null, you will notice that this solution is far from blocking! It’s probably even worse than the sleep solution in the question, as it uses (under Linux) precious resources like the inotify system. Also other processes which write to /dev/null make tail loop. (On my Ubuntu64 16.10 this adds several 10 syscalls per second on an already busy system.)

The question was for a blocking command

Unfortunately, there is no such thing…

Read: I do not know any way to achieve this with the shell directly.

Everything (even sleep infinity) can be interrupted by some signal. So if you want to be really sure it does not exceptionally return, it must run in a loop, like you already did for your sleep. Please note, that (on Linux) /bin/sleep apparently is capped at 24 days (have a look at strace sleep infinity), hence the best you can do is probably:

while :; do sleep 2073600; done

(Note that I believe sleep loops internally for higher values than 24 days, but this means: It is not blocking, it is very slowly looping. So why not move this loop to the outside?)

…but you can come quite near with an unnamed fifo

You can create something which really blocks as long as there are no signals sent to the process. Following uses bash 4, 2 PIDs and 1 fifo:

bash -c 'coproc { exec >&-; read; }; eval exec "${COPROC[0]}<&-"; wait'

You can check that this really blocks with strace if you like:

strace -ff bash -c '..see above..'

How this was constructed

read blocks if there is no input data (see some other answers). However, the tty (aka. stdin) usually is not a good source, as it is closed when the user logs out. Also it might steal some input from the tty. Not nice.

To make read block, we need to wait for something like a fifo which will never return anything. In bash 4 there is a command which can provide us with exactly such a fifo: coproc. If we also wait the blocking read (which is our coproc), we are done. Sadly this needs to keep open two PIDs and a fifo.

Variant with a named fifo

If you do not bother using a named fifo, you can do this as follows:

mkfifo "$HOME/.pause.fifo" 2>/dev/null; read <"$HOME/.pause.fifo"

Not using a loop on the read is a bit sloppy, but you can reuse this fifo as often as you like and make the reads terminate using touch "$HOME/.pause.fifo" (if there is more than a single read waiting, all are terminated at once).

Or use the Linux pause() syscall

For the infinite blocking, there is a Linux system call named pause() which does what we want: Wait forever (until a signal arrives). However there is no userspace program for this (yet).

C

Creating such a program is easy. Here is a snippet to create a very small Linux program called pause which pauses indefinitely (needs a C compiler such as gcc, and uses diet etc. to produce a small binary):

printf '#include <unistd.h>\nint main(){for(;;)pause();}' > pause.c;
diet -Os cc pause.c -o pause;
strip -s pause;
ls -al pause

python

If you do not want to compile something yourself, but you have python installed, you can use this under Linux:

python -c 'while 1: import ctypes; ctypes.CDLL(None).pause()'

(Note: Use exec python -c ... to replace the current shell, this frees one PID. The solution can be improved with some IO redirection as well, freeing unused FDs. This is up to you.)

How this works: ctypes.CDLL(None) loads the “main program” (including the C library) and runs the pause() function from it, all within a loop. Less efficient than the C version, but works.

My recommendation for you:

Stay at the looping sleep. It’s easy to understand, very portable, and blocks for most of the time.

Maybe this seems ugly, but why not just run cat and let it wait for input forever?

TL;DR: since GNU coreutils version 9, sleep infinity does the right thing on Linux systems. Previously (and in other systems) the implementation was to actually sleep the maximum time allowed, which is finite.

Wondering why this is not documented anywhere, I bothered to read the sources from GNU coreutils and I found it executes roughly what follows:

1. Use strtod from C stdlib on the first argument to convert ‘infinity’ to a double precision value. So, assuming IEEE 754 double precision the 64-bit positive infinity value is stored in the seconds variable.
2. Invoke xnanosleep(seconds) (found in gnulib), this in turn invokes dtotimespec(seconds) (also in gnulib) to convert from double to struct timespec.
3. struct timespec is just a pair of numbers: integer part (in seconds) and fractional part (in nanoseconds). Naïvely converting positive infinity to integer would result in undefined behaviour (see §6.3.1.4 from C standard), so instead it truncates to TYPE_MAXIMUM(time_t).
4. The actual value of TYPE_MAXIMUM(time_t) is not set in the standard (even sizeof(time_t) isn’t); so, for the sake of example let’s pick x86-64 from a recent Linux kernel.

This is TIME_T_MAX in the Linux kernel, which is defined (time.h) as:

(time_t)((1UL << ((sizeof(time_t) << 3) - 1)) - 1)

Note that time_t is __kernel_time_t and time_t is long; the LP64 data model is used, so sizeof(long) is 8 (64 bits).

Which results in: TIME_T_MAX = 9223372036854775807.

That is: sleep infinite results in an actual sleep time of 9223372036854775807 seconds (10^11 years). And for 32-bit linux systems (sizeof(long) is 4 (32 bits)): 2147483647 seconds (68 years; see also year 2038 problem).

Edit: apparently the nanoseconds function called is not directly the syscall, but an OS-dependent wrapper (also defined in gnulib).

There’s an extra step as a result: for some systems where HAVE_BUG_BIG_NANOSLEEP is true the sleep is truncated to 24 days and then called in a loop. This is the case for some (or all?) Linux distros. Note that this wrapper may be not used if a configure-time test succeeds (source).

In particular, that would be 24 * 24 * 60 * 60 = 2073600 seconds (plus 999999999 nanoseconds); but this is called in a loop in order to respect the specified total sleep time. Therefore the previous conclusions remain valid.

In conclusion, the resulting sleep time is not infinite but high enough for all practical purposes, even if the resulting actual time lapse is not portable; that depends on the OS and architecture.

To answer the original question, this is obviously good enough but if for some reason (a very resource-constrained system) you really want to avoid an useless extra countdown timer, I guess the most correct alternative is to use the cat method described in other answers.

Edit: recent GNU coreutils versions will try to use the pause syscall (if available) instead of looping. The previous argument is no longer valid when targeting these newer versions in Linux (and possibly BSD).

Portability

This is an important and valid concern:

• sleep infinity is a GNU coreutils extension not contemplated in POSIX. GNU’s implementation also supports a “fancy” syntax for time durations, like sleep 1h 5.2s while POSIX only allows a positive integer (e.g. sleep 0.5 is not allowed).
• Some compatible implementations: GNU coreutils, FreeBSD (at least from version 8.2?), Busybox (requires to be compiled with options FANCY_SLEEP and FLOAT_DURATION).
• The strtod behaviour is C and POSIX compatible (i.e. strtod("infinity", 0) is always valid in C99-conformant implementations, see §7.20.1.3).

sleep infinity looks most elegant, but sometimes it doesn’t work for some reason. In that case, you can try other blocking commands such as cat, read, tail -f /dev/null, grep a etc.

Let me explain why sleep infinity works though it is not documented. jp48’s answer is also useful.

The most important thing: By specifying inf or infinity (both case-insensitive), you can sleep for the longest time your implementation permits (i.e. the smaller value of HUGE_VAL and TYPE_MAXIMUM(time_t)).

Now let’s dig into the details. The source code of sleep command can be read from coreutils/src/sleep.c. Essentially, the function does this:

double s; //seconds
xstrtod (argv[i], &p, &s, cl_strtod); //`p` is not essential (just used for error check).
xnanosleep (s);

Understanding xstrtod (argv[i], &p, &s, cl_strtod)

xstrtod()

According to gnulib/lib/xstrtod.c, the call of xstrtod() converts string argv[i] to a floating point value and stores it to *s, using a converting function cl_strtod().

cl_strtod()

As can be seen from coreutils/lib/cl-strtod.c, cl_strtod() converts a string to a floating point value, using strtod().

strtod()

According to man 3 strtod, strtod() converts a string to a value of type double. The manpage says

The expected form of the (initial portion of the) string is … or (iii) an infinity, or …

and an infinity is defined as

An infinity is either “INF” or “INFINITY”, disregarding case.

Although the document tells

If the correct value would cause overflow, plus or minus HUGE_VAL (HUGE_VALF, HUGE_VALL) is returned

, it is not clear how an infinity is treated. So let’s see the source code gnulib/lib/strtod.c. What we want to read is

else if (c_tolower (*s) == 'i'
         && c_tolower (s[1]) == 'n'
         && c_tolower (s[2]) == 'f')
  {
    s += 3;
    if (c_tolower (*s) == 'i'
        && c_tolower (s[1]) == 'n'
        && c_tolower (s[2]) == 'i'
        && c_tolower (s[3]) == 't'
        && c_tolower (s[4]) == 'y')
      s += 5;
    num = HUGE_VAL;
    errno = saved_errno;
  }

Thus, INF and INFINITY (both case-insensitive) are regarded as HUGE_VAL.

HUGE_VAL family

Let’s use N1570 as the C standard. HUGE_VAL, HUGE_VALF and HUGE_VALL macros are defined in §7.12-3

The macro
HUGE_VAL
expands to a positive double constant expression, not necessarily representable as a float. The macros
HUGE_VALF
HUGE_VALL
are respectively float and long double analogs of HUGE_VAL.

HUGE_VAL, HUGE_VALF, and HUGE_VALL can be positive infinities in an implementation that supports infinities.

and in §7.12.1-5

If a floating result overflows and default rounding is in effect, then the function returns the value of the macro HUGE_VAL, HUGE_VALF, or HUGE_VALL according to the return type

Understanding xnanosleep (s)

Now we understand all essence of xstrtod(). From the explanations above, it is crystal-clear that xnanosleep(s) we’ve seen first actually means xnanosleep(HUGE_VALL).

xnanosleep()

According to the source code gnulib/lib/xnanosleep.c, xnanosleep(s) essentially does this:

struct timespec ts_sleep = dtotimespec (s);
nanosleep (&ts_sleep, NULL);

dtotimespec()

This function converts an argument of type double to an object of type struct timespec. Since it is very simple, let me cite the source code gnulib/lib/dtotimespec.c. All of the comments are added by me.

struct timespec
dtotimespec (double sec)
{
  if (! (TYPE_MINIMUM (time_t) < sec)) //underflow case
    return make_timespec (TYPE_MINIMUM (time_t), 0);
  else if (! (sec < 1.0 + TYPE_MAXIMUM (time_t))) //overflow case
    return make_timespec (TYPE_MAXIMUM (time_t), TIMESPEC_HZ - 1);
  else //normal case (looks complex but does nothing technical)
    {
      time_t s = sec;
      double frac = TIMESPEC_HZ * (sec - s);
      long ns = frac;
      ns += ns < frac;
      s += ns / TIMESPEC_HZ;
      ns %= TIMESPEC_HZ;

      if (ns < 0)
        {
          s--;
          ns += TIMESPEC_HZ;
        }

      return make_timespec (s, ns);
    }
}

Since time_t is defined as an integral type (see §7.27.1-3), it is natural we assume the maximum value of type time_t is smaller than HUGE_VAL (of type double), which means we enter the overflow case. (Actually this assumption is not needed since, in all cases, the procedure is essentially the same.)

make_timespec()

The last wall we have to climb up is make_timespec(). Very fortunately, it is so simple that citing the source code gnulib/lib/timespec.h is enough.

_GL_TIMESPEC_INLINE struct timespec
make_timespec (time_t s, long int ns)
{
  struct timespec r;
  r.tv_sec = s;
  r.tv_nsec = ns;
  return r;
}

What about sending a SIGSTOP to itself?

This should pause the process until SIGCONT is received. Which is in your case: never.

kill -STOP "$$";
# grace time for signal delivery
sleep 60;

I recently had a need to do this. I came up with the following function that will allow bash to sleep forever without calling any external program:

snore()
{
    local IFS
    [[ -n "${_snore_fd:-}" ]] || { exec {_snore_fd}<> <(:); } 2>/dev/null ||
    {
        # workaround for MacOS and similar systems
        local fifo
        fifo=$(mktemp -u)
        mkfifo -m 700 "$fifo"
        exec {_snore_fd}<>"$fifo"
        rm "$fifo"
    }
    read ${1:+-t "$1"} -u $_snore_fd || :
}

NOTE: I previously posted a version of this that would open and close the file descriptor each time, but I found that on some systems doing this hundreds of times a second would eventually lock up. Thus the new solution keeps the file descriptor between calls to the function. Bash will clean it up on exit anyway.

This can be called just like /bin/sleep, and it will sleep for the requested time. Called without parameters, it will hang forever.

snore 0.1  # sleeps for 0.1 seconds
snore 10   # sleeps for 10 seconds
snore      # sleeps forever

There’s a writeup with excessive details on my blog here

This approach will not consume any resources for keeping process alive.

while :; do :; done & kill -STOP $! && wait

Breakdown

• while :; do :; done & Creates a dummy process in background
• kill -STOP $! Stops the background process
• wait Wait for the background process, this will be blocking forever, cause background process was stopped before

Notes

• works only from within a script file.

Instead of killing the window manager, try running the new one with --replace or -replace if available.

sleep inf can save you some typing, it’s the same as sleep infinity