Wednesday, February 17, 2016

psutil 4.0.0 and how to get "real" process memory and environ in Python

New psutil 4.0.0 is out, with some interesting news about process memory metrics. I'll just get straight to the point and describe what's new.

"Real" process memory info

Determining how much memory a process really uses is not an easy matter (see this and this). RSS (Resident Set Size), which is what most people usually rely on, is misleading because it includes both the memory which is unique to the process and the memory shared with other processes. What would be more interesting in terms of profiling is the memory which would be freed if the process was terminated right now. In the Linux world this is called USS (Unique Set Size), and this is the major feature which was introduced in psutil 4.0.0 (not only for Linux but also for Windows and OSX).

USS memory

The USS (Unique Set Size) is the memory which is unique to a process and which would be freed if the process was terminated right now. On Linux this can be determined by parsing all the "private" blocks in /proc/pid/smaps. The Firefox team pushed this further and managed to do the same also on OSX and Windows, which is great. New version of psutil is now able to do the same:
>>> psutil.Process().memory_full_info()
pfullmem(rss=101990, vms=521888, shared=38804, text=28200, lib=0, data=59672, dirty=0, 
         uss=81623, pss=91788, swap=0)

PSS and swap

On Linux there are two additional metrics which can also be determined via /proc/pid/smaps: PSS and swap. PSS, aka "Proportional Set Size", represents the amount of memory shared with other processes, accounted in a way that the amount is divided evenly between the processes that share it. I.e. if a process has 10 MBs all to itself (USS) and 10 MBs shared with another process, its PSS will be 15 MBs. "swap" is simply the amount of memory that has been swapped out to disk. With memory_full_info() it is possible to implement a tool like this, similar to smem on Linux, which provides a list of processes sorted by "USS". It is interesting to notice how RSS differs from USS:
~/svn/psutil$ ./scripts/
PID     User    Cmdline                            USS     PSS    Swap     RSS
3986    giampao /usr/bin/python3 /usr/bin/indi   15.3M   16.6M      0B   25.6M
3906    giampao /usr/lib/ibus/ibus-ui-gtk3       17.6M   18.1M      0B   26.7M
3991    giampao python /usr/bin/hp-systray -x    19.0M   23.3M      0B   40.7M
3830    giampao /usr/bin/ibus-daemon --daemoni   19.0M   19.0M      0B   21.4M
20529   giampao /opt/sublime_text/plugin_host    19.9M   20.1M      0B   22.0M
3990    giampao nautilus -n                      20.6M   29.9M      0B   50.2M
3898    giampao /usr/lib/unity/unity-panel-ser   27.1M   27.9M      0B   37.7M
4176    giampao /usr/lib/evolution/evolution-c   35.7M   36.2M      0B   41.5M
20712   giampao /usr/bin/python -B /home/giamp   45.6M   45.9M      0B   49.4M
3880    giampao /usr/lib/x86_64-linux-gnu/hud/   51.6M   52.7M      0B   61.3M
20513   giampao /opt/sublime_text/sublime_text   65.8M   73.0M      0B   87.9M
3976    giampao compiz                          115.0M  117.0M      0B  130.9M
32486   giampao skype                           145.1M  147.5M      0B  149.6M


In order to get these values (USS, PSS and swap) we need to pass through the whole process address space. This usually requires higher user privileges and is considerably slower than getting the "usual" memory metrics via Process.memory_info(), which is probably the reason why tools like ps and top show RSS/VMS instead of USS. A big thanks goes to the Mozilla team which figured out all this stuff on Windows and OSX, and to Eric Rahm who put the PRs for psutil together (see #744, #745 and #746). For those of you who don't use Python and would like to port the code on other languages here's the interesting parts:

Memory type percent

After reorganizing process memory APIs I decided to add a new memtype parameter to Process.memory_percent(). With this it is now possible to compare a specific memory type (not only RSS) against the total physical memory. E.g.
>>> psutil.Process().memory_percent(memtype='pss')

Process environ

Second biggest improvement in psutil 4.0.0 is the ability to determine the process environment variables. This opens interesting possibilities about process recognition and monitoring techniques. For instance, one might start a process by passing a certain custom environment variable, then iterate over all processes to find the one of interest (and figure out whether it's running or whatever):
import psutil
for p in psutil.process_iter():
        env = p.environ()
    except psutil.Error:
        if 'MYAPP' in env:
Process environ was a long standing issue (year 2009) who I gave up to implement because the Windows implementation worked for the current process only. Frank Benkstein solved that and the process environ can now be determined on Linux, Windows and OSX for all processes (of course you may still bump into AccessDenied for processes owned by another user):
>>> import psutil
>>> from pprint import pprint as pp
>>> pp(psutil.Process().environ())
 'COLORTERM': 'gnome-terminal',
 'COMPIZ_BIN_PATH': '/usr/bin/',
 'HOME': '/home/giampaolo',
 'PWD': '/home/giampaolo/svn/psutil',
It must be noted that the resulting dict usually does not reflect changes made after the process started (e.g. os.environ['MYAPP'] = '1'). Again, for whoever is interested in doing this in other languages, here's the interesting parts:

Extended disk IO stats

psutil.disk_io_counters() has been extended to report additional metrics on Linux and FreeBSD:
  • busy_time, which is the time spent doing actual I/Os (in milliseconds).
  • read_merged_count and write_merged_count (Linux only), which is number of merged reads and writes (see iostats doc)
With these new metrics it is possible to have a better representation of actual disk utilization, similarly to iostat command on Linux.

OS constants

Given the increasing number of platform-specific metrics I added a new set of constants to quickly differentiate what platform you're on: psutil.LINUX, psutil.WINDOWS, etc.

Main bug fixes

  • #734: on Python 3 invalid UTF-8 data was not correctly handled for proces name(), cwd(), exe(), cmdline() and open_files() methods, resulting in UnicodeDecodeError. This was affecting all platforms. Now surrogateescape error handler is used as a workaround for replacing the corrupted data.
  • #761: [Windows] psutil.boot_time() no longer wraps to 0 after 49 days.
  • #767: [Linux] disk_io_counters() may raise ValueError on 2.6 kernels and it's  broken on 2.4 kernels.
  • #764: psutil can now be compiled on NetBSD-6.X.
  • #704: psutil can now be compiled on Solaris sparc.
Complete list of bug fixes is available here.

Porting code

Being 4.0.0 a major version, I took the chance to (lightly) change / break some APIs.
  • Process.memory_info() no longer returns just an (rss, vms) namedtuple. Instead it returns a namedtuple of variable length, changing depending on the platform (rss and vms are always present though, also on Windows). Basically it returns the same result of old process_memory_info_ex(). This shouldn't break your existent code, unless you were doing "rss, vms = p.memory_info()".
  • At the same time process_memory_info_ex() is now deprecated. The method is still there as an alias for memory_info(), issuing a deprecation warning.
  • psutil.disk_io_counters() returns 2 additional fields on Linux and 1 additional field on FreeBSD. 
  • psutil.disk_io_counters() on NetBSD and OpenBSD no longer return write_count and read_count metrics because the kernel do not provide them (we were returning the busy time instead). I also don't expect this to be a big issue because NetBSD and OpenBSD support is very recent.

Final notes and looking for a job

OK, this is it. I would like to spend a couple more words to announce the world that I'm currently unemployed and looking for a remote gig again! =) I want remote because my plan for this year is to remain in Prague (Czech Republic) and possibly spend 2-3 months in Asia. If you know about any company who's looking for a Python backend dev who can work from afar feel free to share my Linkedin profile or mail me at g.rodola [at] gmail [dot] com.

External links

Friday, February 12, 2016

How to always execute exit functions in Python

...or why atexit.register() and signal.signal() are evil

UPDATE (2016-02-13): this recipe no longer handles SIGINT, SIGQUIT and SIGABRT as aliases for "application exit" because it was a bad idea. It only handles SIGTERM. Also it no longer support Windows because signal.signal() implementation is too different than POSIX.

Many people erroneously think that any function registered via atexit module is guaranteed to always be executed when the program terminates. You may have noticed this is not the case when, for example, you daemonize your app in production then try to stop it or restart it: the cleanup functions will not be executed. This is because functions registered wth atexit module are not called when the program is killed by a signal:
import atexit, os, signal

def cleanup():
    print("on exit")  # XXX this never gets printed

os.kill(os.getpid(), signal.SIGTERM)

It must be noted that the same thing would happen if instead of atexit.register() we would use a "finally" clause. It turns out the correct way to make sure the exit function is always called in case a signal is received is to register it via signal.signal(). That has a drawback though: in case a third-party module has already registered a function for that signal (SIGTERM or whatever), your new function will overwrite the old one:

import os, signal

def old(*args):
    print("old")  # XXX this never gets printed

def new(*args):

signal.signal(signal.SIGTERM, old)
signal.signal(signal.SIGTERM, new)
os.kill(os.getpid(), signal.SIGTERM)

Also, we would still have to use atexit.register() so that the function is called also on "clean" interpreter exit and take into account other signals other than SIGTERM which would cause the process to terminate. This recipe attempts to address all these issues so that:
  •  the exit function is always executed for all exit signals (SIGTERM, SIGINT, SIGQUIT, SIGABRT) on SIGTERM and on "clean" interpreter exit.
  • any exit function(s) previously registered via atexit.register() or signal.signal() will be executed as well (after the new one). 
  • It must be noted that the exit function will never be executed in case of SIGKILL, SIGSTOP or os._exit().

The code

Function / decorator which tries very hard to register a function to
be executed at importerer exit.

Author: Giampaolo Rodola'
License: MIT

from __future__ import print_function
import atexit
import os
import functools
import signal
import sys

_registered_exit_funs = set()
_executed_exit_funs = set()

def register_exit_fun(fun=None, signals=[signal.SIGTERM],
                      logfun=lambda s: print(s, file=sys.stderr)):
    """Register a function which will be executed on "normal"
    interpreter exit or in case one of the `signals` is received
    by this process (differently from atexit.register()).
    Also, it makes sure to execute any other function which was
    previously registered via signal.signal(). If any, it will be
    executed after our own `fun`.

    Functions which were already registered or executed via this
    function will be ignored.

    Note: there's no way to escape SIGKILL, SIGSTOP or os._exit(0)
    so don't bother trying.

    You can use this either as a function or as a decorator:

        def cleanup():

        # ...or


    Note about Windows: I tested this some time ago and didn't work
    exactly the same as on UNIX, then I didn't care about it
    anymore and didn't test since then so may not work on Windows.


    - fun: a callable
    - signals: a list of signals for which this function will be
      executed (default SIGTERM)
    - logfun: a logging function which is called when a signal is
      received. Default: print to standard error. May be set to
      None if no logging is desired.
    def stringify_sig(signum):
        if sys.version_info < (3, 5):
            smap = dict([(getattr(signal, x), x) for x in dir(signal)
                         if x.startswith('SIG')])
            return smap.get(signum, signum)
            return signum

    def fun_wrapper():
        if fun not in _executed_exit_funs:

    def signal_wrapper(signum=None, frame=None):
        if signum is not None:
            if logfun is not None:
                logfun("signal {} received by process with PID {}".format(
                    stringify_sig(signum), os.getpid()))
        # Only return the original signal this process was hit with
        # in case fun returns with no errors, otherwise process will
        # return with sig 1.
        if signum is not None:
            if signum == signal.SIGINT:
                raise KeyboardInterrupt
            # XXX - should we do the same for SIGTERM / SystemExit?

    def register_fun(fun, signals):
        if not callable(fun):
            raise TypeError("{!r} is not callable".format(fun))
        set([fun])  # raise exc if obj is not hash-able

        signals = set(signals)
        for sig in signals:
            # Register function for this signal and pop() the previously
            # registered one (if any). This can either be a callable,
            # SIG_IGN (ignore signal) or SIG_DFL (perform default action
            # for signal).
            old_handler = signal.signal(sig, signal_wrapper)
            if old_handler not in (signal.SIG_DFL, signal.SIG_IGN):
                # ...just for extra safety.
                if not callable(old_handler):
                # This is needed otherwise we'll get a KeyboardInterrupt
                # strace on interpreter exit, even if the process exited
                # with sig 0.
                if (sig == signal.SIGINT and
                        old_handler is signal.default_int_handler):
                # There was a function which was already registered for this
                # signal. Register it again so it will get executed (after our
                # new fun).
                if old_handler not in _registered_exit_funs:

        # This further registration will be executed in case of clean
        # interpreter exit (no signals received).
        if fun not in _registered_exit_funs or not signals:

    # This piece of machinery handles 3 usage cases. register_exit_fun()
    # used as:
    # - a function
    # - a decorator without parentheses
    # - a decorator with parentheses
    if fun is None:
        def outer(fun):
            return register_fun(fun, signals)
        return outer
        register_fun(fun, signals)
        return fun


As a function:
def cleanup():

As a decorator:
def cleanup():

Unit tests

This recipe is hosted on ActiveState and has a full set of unittests. It works with Python 2 and 3.

Notes about Windows

On Windows signals are only partially supported meaning a function which was previously registered via signal.signal() will be executed only on interpreter exit, but not if the process receives a signal. Apparently this is a limitation either of Windows or the signal module (most likely Windows).

Because of how different signal.signal() behaves on Windows, this code is UNIX only:

Proposal for stdlib inclusion

The fact that atexit module does not handle signals and that signal.signal() overwrites previously registered handlers is unfortunate. It is also confusing because it is not immediately clear which one you are supposed to use (and it turns out you're supposed to use both). Most of the times you have no idea (or don't care) that you're overwriting another exit function. As a user, I would just want to execute an exit function, no matter what, possibly without messing with whatever a module I've previously imported has done with signal.signal(). To me this suggests there could be space for something like "atexit.register_w_signals".

External discussions