path: root/hurd/translator/procfs/jkoenig/discussion.mdwn

[[!meta copyright="Copyright © 2010, 2011, 2012 Free Software Foundation,
Inc."]]

[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
id="license" text="Permission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant
Sections, no Front-Cover Texts, and no Back-Cover Texts.  A copy of the license
is included in the section entitled [[GNU Free Documentation
License|/fdl]]."]]"""]]

[[!tag open_issue_hurd]]

[[!toc]]


# Miscellaneous

IRC, #hurd, around September 2010

    <youpi> jkoenig: is it not possible to provide a /proc/self which points at
      the client's pid?
    <pinotree> (also, shouldn't /proc/version say something else than "Linux"?)
    <youpi> to make linux tools work, no :/
    <youpi> kfreebsd does that too
    <pinotree> really?
    <youpi> yes
    <youpi> (kfreebsd, not freebsd)
    <pinotree> does kbsd's one print just "Linux version x.y.z" too, or
      something more eg in a second line?
    <pinotree> (as curiosity)
    <youpi> % cat /proc/version
    <youpi> Linux version 2.6.16 (des@freebsd.org) (gcc version 4.3.5) #4 Sun
      Dec 18 04:30:00 CET 1977
    <pinotree> k
    <giselher> I had some problems with killall5 to read the pid from /proc, Is
      this now more reliable?
    <youpi> I haven't tested with jkoenig's implementation
    [...]
    <pinotree> looks like he did 'self' too, see rootdir_entries[] in rootdir.c
    <youpi> but it doesn't point at self
    <antrik> youpi: there is no way to provide /proc/self, because the server
      doesn't know the identity of the client
    <youpi> :/
    <antrik> youpi: using the existing mechanisms, we would need another magic
      lookup type
    <antrik> an alternative idea I discussed with cfhammer once would be for
      the client to voluntarily provide it's identity to the server... but that
      would be a rather fundamental change that requires careful consideration
    <antrik> also, object migration could be used, so the implementation would
      be provided by the server, but the execution would happen in the
      client... but that's even more involved :-)
    <youpi> but we've seen how much that'd help with a lot of other stuff
    <antrik> I'm not sure whether we discussed this on the ML at some point, or
      only on IRC
    <youpi> it "just" needs to be commited :)
    <antrik> in either case, it can't hurt to bring this up again :-)


# root group

IRC, #hurd, around October 2010

    <pinotree> the only glitch is that files/dirs have the right user as
      owner, but always with root group


# `/proc/[PID]/stat` being 400 and not 444, and some more

IRC, freenode, #hurd, 2011-03-27

    <pochu> is there a reason for /proc/$pid/stat to be 400 and not 444 like on
      Linux?
    <pochu> there is an option to procfs to make it 444 like Linux
    <pochu> jkoenig: ^
    <jkoenig> pochu, hi
    <jkoenig> /proc/$pid/stat reveals information which is not usually
      available on Hurd
    <jkoenig> so I made it 400 by default to avoid leaking anything
    <pochu> is there a security risk in providing that info?
    <jkoenig> probably not so much, but it seemed like it's not really a
      descision procfs should make
    <jkoenig> I'm not sure which information we're speaking about, though, I
      just remember the abstract reason.
    <pochu> things like the pid, the memory, the priority, the state...
    <pochu> sounds safe to expose
    <jkoenig> also it's 0444 by default in "compatible" mode
    <jkoenig> (which is necessary for the linux tools to work well)
    <pochu> yeah I saw that :)
    <pochu> my question is, should we change it to 0444 by default? if there
      are no security risks and this improves compatibility, sounds like a good
      thing to me
    <pochu> we're already 'leaking' part of that info through e.g. ps
    <jkoenig> I think /proc should be translated by /hurd/procfs --compatible
      by default (I'm not sure whether it's already the case)
    <jkoenig> also I'm not sure why hurd-ps is setuid root, rather than the
      proc server being less paranoid, but maybe I'm missing something.
    <pochu> jkoenig: it's not, at least not on Debian
    <pochu> youpi: hi, what do you think about starting procfs with
      --compatible by default?
    <pochu> youpi: or changing /proc/$pid/stat to 0444 like on Linux
      (--compatible does that among a few other things)
    <youpi> I guess you need it for something?
    <pochu> I'm porting libgtop :)
    <youpi> k
    <pochu> though I still think we should do this in procfs itself
    <youpi> ymmv
    <jkoenig> pochu, youpi, --compatible is also needed because mach's high
      reported sysconf(_SC_CLK_TCK) makes some integers overflow (IIRC)
    <youpi> agreed
    <jkoenig> luckily, tools which use procfs usually try to detect the value
      /proc uses rather than rely on CLK_TCK
    <jkoenig> (so we can choose whatever reasonable value we want)

IRC, freenode, #hurd, 2011-03-28

    <antrik> jkoenig: does procfs expose any information that is not available
      to everyone through the proc server?...
    <antrik> also, why is --compatible not the default; or rather, why is there
      even another mode? the whole point of procfs is compatibility...
    <jkoenig> antrik, yes, through the <pid>/environ and (as mentionned above)
      <pid>/stat files, but I've been careful to make these files readable only
      to the process owner
    <jkoenig> --compatible is not the default because it relaxes this paranoia
      wrt. the stat file, and does not conform to the specification with regard
      to clock tick counters
    <antrik> what specification?
    <jkoenig> the linux proc(5) manpage
    <jkoenig> which says clock tick counters are in units of
      1/sysconf(_SC_CLK_TCK)
    <antrik> so you are saying that there is some information that the Hurd
      proc server doesn't expose to unprivileged processes, but linux /proc
      does?
    <jkoenig> yes
    <antrik> that's odd. I wonder what the reasoning behind that could be
    <antrik> but this information is available through Hurd ps?
    <antrik> BTW, what exactly is _SC_CLK_TCK supposed to be?
    <pinotree> jkoenig: hm, just tried with two random processes on linux
      (2.6.32), and enrivon is 400
    <pinotree> (which makes sense, as you could have sensible informations eg
      in http_proxy or other envvars)
    <jkoenig> antrik, CLK_TCK is similar to HZ (maybe clock resolution instead
      of time slices ?)
    <jkoenig> sysconf(3) says "The number of clock ticks per second."
    <jkoenig> antrik, I don't remember precisely what information this was, but
      ps-hurd is setuid root.
    <jkoenig> anyway, if you run procfs --compatible as a user and try to read
      foo/1/stat, the result is an I/O error, which is the result of the proc
      server denying access.
    <antrik> but Linux /proc acutally uses HZ as the unit IIRC? or is
      _SC_CLK_TCK=HZ on Linux?...
    <jkoenig> I expect they're equal.
    <jkoenig> in practice procps uses heuristics to guess what value /proc uses
      (for compatibility purposes with older kernels)
    <jkoenig> I don't think HZ is POSIX, while _SC_CLK_TCK is specifies as the
      unit for (at least) the values returned by times()
    <jkoenig> s/specifies/specified/
    <jkoenig> antrik, some the information is fetched directly from mach by
      libps, and understandably, the proc server does not give the task port to
      anyone who asks.
    <antrik> well, as long as the information is exposed through ps, there is
      no point in hiding it in procfs...
    <antrik> and I'm aware of the crazy guessing in libproc... I was actually
      mentoring the previous procfs implementation
    <antrik> (though I never got around to look at his buggy code...)
    <jkoenig> ok

IRC, freenode, #hurd, 2011-07-22

    <pinotree> hm, why /proc/$pid/stat is 600 instead of 644 of linux?
    <jkoenig> pinotree, it reveals information which, while not that sensitive,
      would not be available to users through the normal proc interface.
    <jkoenig> (it's available through the ps command which is setuid root)
    <jkoenig> we discussed at some point making it 644, IIRC.
    <pinotree> hm, then why is it not a problem on eg linux?
    <jkoenig> (btw you can change it with the -s option.)
    <jkoenig> pinotree, it's not a problem because the information is not that
      sensitive, but when rewriting procfs I preferred to play it self and
      consider it's not procfs' job to decide what is sensitive or not.
    <jkoenig> IIRC it's not sensitive but you need the task port to query it.
    <jkoenig> like, thread times or something.
    <pinotree> status is 644 though
    <jkoenig> but status contains information which anyone can ask to the proc
      server anyway, I think.


# `/proc/mounts`, `/proc/[PID]/mounts`

IRC, freenode, #hurd, 2011-07-25

    < pinotree> jkoenig: btw, what do you think about providing empty
      /proc/mounts and /proc/$pid/mounts files?
    < jkoenig> pinotree, I guess one would have to evaluate the consequences
      wrt. existing use cases (in other words, "I have absolutely no clue
      whatsoever about whether that would be desirable" :-)
    < jkoenig> pinotree, the thing is, an error message like "/proc/mounts: No
      such file or directory" is rather explicit, whereas errors which would be
      caused by missing data in /proc/mounts would maybe be harder to track
    < braunr> this seems reasonable though
    < braunr> there already are many servers with e.g. grsecurity or chrooted
      environments where mounts is empty
    < pinotree> well, currently we also have an empty mtab
    < braunr> pinotree: but what do you need that for ?
    < braunr> pinotree: the init system ?
    < pinotree> and the mnt C api already returns no entries (or it bails out,
      i don't remember)
    < pinotree> not a strict need


# `/proc/[PID]/auxv`, `/proc/[PID]/exe`, `/proc/[PID]/mem`

Needed by glibc's `pldd` tool (commit
11988f8f9656042c3dfd9002ac85dff33173b9bd).


# `/proc/self/exe`

[[!message-id "alpine.LFD.2.02.1110111111260.2016@akari"]].  Needed by glibc's
`stdlib/tst-secure-getenv.c`.

Is it generally possible to use something like the following instead?
Disadvantage is that every program using this needs to be patched.

    #include <dlfcn.h>
    [...]
    Dl_info DLInfo;
    int err = dladdr(&main, &DLInfo);
    if (err == 0)
      [...]
    /* Pathname of shared object that contains address: DLInfo.dli_fname.  */
    /* Filter it through realpath.  */

This is used in `[LLVM]/lib/Support/Unix/Path.inc`.


# `/proc/[PID]/fd/`

## IRC, freenode, #hurd, 2012-04-24

    <antrik> braunr: /proc/*/fd can be implemented in several ways. none of
      them would require undue centralisation
    <antrik> braunr: the easiest would be adding one more type of magic lookup
      to the existing magic lookup mechanism
    <antrik> wait, I mean /proc/self... for /proc/*/fd it's even more
      straighforward -- we might even have a magic lookup for that already
    <pinotree> i guess the ideal thing would be implement that fd logic in
      libps
    <antrik> pinotree: nope. it doesn't need to ask proc (or any other server)
      at all. it's local information. that's what we have the magic lookups for
    <antrik> one option we were considering at some point would be using the
      object migration mechanism, so the actual handling would still happen
      client-side, but the server could supply the code doing it. this would
      allow servers to add arbitrary magic lookup methods without any global
      modifications... but has other downsides :-)
    <gnu_srs> youpi: How much info for /proc/*/fd is possible to get from
      libps? Re: d-h@
    <youpi> see my mail
    <youpi> I don't think there is an interface for that
    <youpi> processes handle fds themselves
    <youpi> so libps would have to peek in there
    <youpi> and I don't remember having seen any code like that
    <gnu_srs> 10:17:17< antrik> wait, I mean /proc/self... for /proc/*/fd it's
      even more straighforward -- we might even have a magic lookup for that
      already
    <gnu_srs> pinotree: For me that does not ring a bell on RPCs. Don't know
      what magic means,,
    <youpi> for /proc/self/fd we have a magic lookup
    <youpi> for /proc/pid/fd, I don't think we have
    <gnu_srs> magic lookup*
    <gnu_srs> magic lookup == RPC?
    <youpi> magic lookup is a kind of answer to the lookup RPC
    <youpi> that basically says "it's somewhere else, see there"
    <youpi> the magic FD lookup tells the process "it's your FD number x"
    <youpi> which works for /proc/self/fd, but not /proc/pid/fd
    <civodul> youpi, gnu_srs: regarding FDs, there the msg_get_fd RPC that
      could be used
    <civodul> `msgport' should have --get-fd, actually
    <youpi> civodul: I assumed that the reason why msgport doesn't have it is
      that it didn't exist
    <youpi> so we can get a port on the fd
    <youpi> but then how to know what it is?
    <civodul> youpi: ah, you mean for the /proc/X/fd symlinks?
    <civodul> good question
    <civodul> it's not designed to be mapped back to names, indeed :-)
    <antrik> youpi: yeah, I realized myself that only /proc/self/fd is trivial
    <antrik> BTW, in Linux it's nor real symlinks. it's magic, with some very
      strange (but useful in certain situations) semantics
    <antrik> not real symlinks
    <antrik> it's very weird for example for fd connected to files that have
      been unlinked. it looks like a broken symlink, but when dereferencing
      (e.g. with cp), you get the actual file contents...


# `/proc/[PID]/maps`

[[!GNU_Savannah_bug 32770]]


## IRC, OFTC, #debian-hurd, 2012-06-20

    <pinotree> bdefreese: the two elfutils tests fail because there are no
      /proc/$pid/maps files
    <pinotree> that code is quite relying on linux features, like locating the
      linux kernel executables and their modules, etc
    <pinotree> (see eg libdwfl/linux-kernel-modules.c)
    <pinotree> refactor elfutils to have the linux parts executed only on linux
      :D
    <bdefreese> Oh yeah, the maintainer already seems really thrilled about
      Hurd..  Did you see
      http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=662041 ?
    <pinotree> kurt is generally helpful with us (= hurd)
    <pinotree> most probably there he is complaining that we let elfutils build
      with nocheck (ie skipping the test suite run) instead of investigate and
      report why the test suite failed


# `/proc/[PID]/cwd`

## IRC, freenode, #hurd, 2012-06-30

    * pinotree has a local work to add the /proc/$pid/cwd symlink, but relying
        on "internal" (but exported) glibc functions


# "Unusual" PIDs

Not actually related to procfs, but here seems to be a convenient place for
filing these:


## IRC, freenode, #hurd, 2012-08-10

    <braunr> too bad the proc server has pid 0
    <braunr> top & co won't show it


## IRC, OFTC, #debian-hurd, 2012-09-18

    <pinotree> youpi: did you see
      https://enc.com.au/2012/09/careful-with-pids/'
    <pinotree> ?
    <youpi> nope


# CPU Usage

## IRC, freenode, #hurd, 2013-01-30

    <gnu_srs> Hi, htop seems to report CPU usage correct, but not top, is that
      a known issue? 
    <youpi> does your /proc have the -c flag?
    <gnu_srs> /hurd/procfs -c
    <youpi> I don't remember which way it works, but iirc depending on whether
      -c is there or not, it will work or not
    <youpi> problem being that nothing says linux' clock is 100Hz, but a lot of
      programs assume it
    <gnu_srs> seems like htop gets it right though
    <youpi> possibly just by luc
    <youpi> k


### IRC, freenode, #hurd, 2013-01-31

    <braunr> both htop and top seem to have problems report the cpu time
    <braunr> so i expect the problem to be in procfs