path: root/open_issues/clock_gettime.mdwn

[[!meta copyright="Copyright © 2011, 2013, 2014 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]]."]]"""]]

[[!meta title="clock_gettime"]]

[[!tag open_issue_glibc open_issue_gnumach]]

Missing `clock_gettime(CLOCK_MONOTONIC)` (e.g. for iceweasel)

It could be a mere matter of extending the
[[mapped-time_interface|microkernel/mach/gnumach/interface/device/time]]:
add it to
`mapped_time_value_t` in gnumach, handle it in `gnumach/kern/mach_clock.c`, and
make `clock_gettime` use it.

BTW, also make `gettimeofday()` use it, since it's way more efficient and some
applications assume that it is.

What about adding a nanosecond-precision clock, too?  --[[tschwinge]]


# IRC, freenode, #hurd, 2011-08-26

    < pinotree> youpi: thing is: apparently i found a simple way to have a
      monotonic clock as mmap-able device inside gnumach
    < pinotree> currently, in kern/mach_clock.c there's a variable 'time',
      which gets increased on clock interrupt, and optionally modified by
      host_set_time
    < pinotree> ()
    < pinotree> if i add a new variable next to it, only increasing it on
      interrupt but not modifying it at all otherwise, would that give me a
      monotonic clock?
    < pinotree> at least on sme basic tests i did, it seems it could work that
      way
    < youpi> yes, it should work
    < braunr> sure
    < youpi> and that's the way I was considering implementing it


# IRC, freenode, #hurd, 2011-09-06

    <pinotree> yeah, i had a draft of improved idea for also handling
      nanoseconds
    <tschwinge> pinotree: Ah, nice, I thought about nanoseconds as well.
    <tschwinge> pinotree, youpi: This memory page is all-zero by default,
      right?
    <tschwinge> Can't we then say that its last int is a version code, and if
      it is 0 (as it is now), we only have the normal mapped time field, if it
      is 1, we also have the monotonic cliock and ns precision on address 8 and
      16 (or whatever)?
    <tschwinge> In case that isn't your plan anyway.
    <youpi> it's all-zero, yes
    <tschwinge> Or, we say if a field is != 0 it is valid.
    <youpi> making the last int a version code limits the size to one page
    <youpi> I was thinking a field != 0  being valid is simpler
    <youpi> but it's probably a problem too
    <youpi> in that glibc usually caches whether interfaces are supported
    <tschwinge> Wrap-around?
    <youpi> for some clocks, it may be valid that the value is 0
    <youpi> wrap-around is another issue too
    <tschwinge> Well, then we can do the version-field thing, but put it right
      after the current time field (address 8, I think)?
    <youpi> yes
    <youpi> it's a bit ugly, but it's hidden behind the structure
    <tschwinge> It's not too bad, I think.
    <youpi> yes
    <tschwinge> And it will forever be a witness of the evolving of this
      map_time interface.  :-)


# IRC, freenode, #hurd, 2013-02-11

In context of [[select]].

    <pinotree> braunr: would you send for review (and inclusion) your
      time_data_t addition?
    <pinotree> this way we could add nanosecs-based utime rpc (and then their
      implementation in libc)
    <braunr> pinotree: it's part of the hurd branch
    <braunr> do you want it sent separately ?
    <pinotree> yeah
    <braunr> ok
    <braunr> let me get it right first :)
    <pinotree> sure :)


## IRC, freenode, #hurd, 2013-02-12

    <braunr> pinotree:
      http://git.savannah.gnu.org/cgit/hurd/hurd.git/commit/?h=rbraun/select_timeout_pthread_v2&id=6ec50e62d9792c803d00cbff1cab2c0b3675690a
    <pinotree> uh nice
    <pinotree> will need two small inline functions to convert time_data_t <->
      timespec, but that's it
    <braunr> hm right
    <braunr> i could have thought about it
    <braunr> but i'll leave it for another patch :p
    <pinotree> oh sure, no hurry


## IRC, freenode, #hurd, 2013-02-19

    <youpi> braunr: about time_data_t, I get it's needed that it be an array
    <youpi> so it can be passed by reference, not by value?
    <braunr> by address, yes
    <braunr> that's the difference between array and struct


## IRC, freenode, #hurd, 2013-02-25

    <youpi> braunr: why did you want to see time_data passed as pointer, not as
      struct?
    <braunr> to microoptimize
    <braunr> the struct is 2 64-bit integers
    <youpi> well, we already pass structs along in a few cases,
      e.g. io_statbuf_t, rusage_t, etc.
    <youpi> be it written t[0].sec or t->sec, it seems odd
    <youpi> copying 2 64bit integers is not much compared to the potential for
      bugs here
    <braunr> bugs ?
    <youpi> yes, as in trying to access t[1], passing a wrong pointer, etc.
    <youpi> or the reader frowning on "why is this case different than the
      others?"
    <braunr> well, i'm already usually frowning when i see what mig does ..
    <youpi> right
    <youpi> on the plus side, it's only the client side, i.e. mostly glibc,
      which sees the t[0]
    <braunr> and the practice established by my patch is to convert to struct
      timespec as soon as possible
    <braunr> the direct use of this type is therefore limited
    <youpi> could we define time_data_t as a struct time_data * instead of
      struct time_data[1] ?
    <youpi> (in the.h)
    <youpi> that would make more sense to define a struct time_data, and pass a
      pointer to it
    <braunr> i'm not sure
    <braunr> the mach server writing guide was very clear about array implying
      a C array too
    <braunr> and i remember having compilation problems before doing that
    <braunr> but i don't remember their nature exactly 
    <youpi> I'm not sure to understand what you said about converting to struct
      timespec
    <youpi> what makes it not possible now?
    <youpi> and what is the relation with being an array or a pointer?
    <braunr> concerning struct timespec, what i mean is that the functions
      called by the mig stub code directly convert time_data_t to a struct
      timespec (which is the real type used throughout the hurd code)
    <braunr> about the rest, i'm not sure, i'd have to try again
    <braunr> mig just assumes it's an array
    <youpi> and why not just using struct timespec?
    <youpi> (for the mig type too)
    <braunr> my brain can't correctly compute variable sized types in mig
      definition files
    <braunr> i wanted something that would remain correct for the 64-bit port

[[64-bit_port]], [[mig_portable_rpc_declarations]].

    <youpi> ah, you mean because tv_nsec is a long, which will not be the same
      type?
    <braunr> and tv_sec being a time_t (thus a long too)
    <youpi> but we have the same issue e.g. for the rusage structure, don't we?
    <braunr> yes
    <youpi> so we'll have to fix things for that too anyway
    <braunr> sure
    <youpi> making a special case will not necessarily help
    <braunr> but it doesn't mean new interfaces have to be buggy too
    <youpi> well, using the proper type in the server itself is nicer
    <youpi> instead of having to convert
    <braunr> yes
    <braunr> i'm not exactly sure where to declare struct timespec then
    <braunr> should it be declared in hurd_types.h, and simply reused by the
      libc headers ?
    <youpi> ? AIUI, it's the converse, hurd_types.h uses the struct timespec
      from libc headers, and defines timespec_t
    <braunr> ok
    <youpi> timespec_t being the internal type whose definition gets done right
      for mig to do the right thing
    <braunr> yes
    <braunr> i see
    <braunr> so, you'd like a struct of integer_t instead of an array of
      signed64
    <youpi> for our current 32bit userland yes
    <braunr> do you want to make the changes yourself or should i add a new
      branch ?
    <youpi> and we'll make that a 64bit struct when we have a64bit userland


# IRC, freenode, #hurd, 2013-04-06

    <tschwinge> pinotree: You had once been working on adding nsec-procision
      timestamps to GNU Mach's maptime interface (or what the name is).  Is
      that blocked on something or just waiting to be continued?
    <pinotree> blocked on me needing to learn more the proper way to do
      "atomic" update of the struct with time :)


# IRC, freenode, #hurd, 2013-09-04

    <teythoon> do we have CLOCK_MONOTONIC ?
    <braunr> teythoon: i think we do but it's actually a simple offset from
      CLOCK_REALTIME .. :)
    <teythoon> ah never mind, I do hate this posix time interface anyways
    <braunr> really ?
    <braunr> i think librt is decent


# Candidate for [[vDSO]] code?


# IRC, freenode, #hurd, 2014-02-23

    <desrt> GLib (gthread-posix.c): Unexpected error from C library during
      'pthread_condattr_setclock': Invalid argument.  Aborting.
    <desrt> uh oh...
    <desrt> time to go digging in glibc i guess...
    <braunr> what are you trying to run ?
    <desrt> glib
    <braunr> with what ?
    <desrt> just running glib's test suite under jhbuild
    <desrt> i maintain glib and i made some changes recently -- i wanted to
      make sure they didn't break the hurd
    <desrt> and it seems they have ;/
    <braunr> well
    <braunr> the hurd doesn't completely comply with posix 2008
    <desrt> long story short: we've keyed our timed waits on condition
      variables to the monotonic clock for a long time now, but we never tested
      that it actually worked
    <desrt> so i just added an assert -- and indeed it fails on hurd
    <braunr> our glibc lies about supporting timers
    <braunr> good thinking
    <braunr> we don't support the monotonic clock
    <desrt> clock_gettime(CLOCK_MONOTONIC) seems to work
    <braunr> and you should know that, even if clock selection and timers are
      available (which posix 2008 requires), it's still optional
    <braunr> no, glibc lies
    <desrt> !!
    <braunr> our "support" is a mere hack shifting CLOCK_REALTIME
    <desrt> it should at least lie consistently :)
    <braunr> we need to implement CLOCK_MONOTONIC properly
    <desrt> ya... that would be very nice indeed
    <braunr> not that hard either
    <desrt> i agree!
    <braunr> we just have to do it right
    <desrt> fwiw, i plan to keep this assert in glib
    <braunr> yes, it's good
    <desrt> is there anywhere i can file a bug to give you guys some advance
      warning?
    <braunr> i don't think it's needed
    <braunr> we know the problem
    <desrt> k -- consider yourself warned, then :)
    <braunr> and it's been a bigger concern recently
    <desrt> awesome.  glad i don't have to do anything :)
    <braunr> if it's not already done, i suggest you check for the
      CLOCK_MONOTONIC option
    <desrt> fwiw, i'm trying to get a regular debian/gnu/hurd build of
      glib/gtk/etc setup
    <braunr> regular ?
    <desrt> ya... out of git master on a daily basis
    <braunr> from sources ?
    <braunr> oh nice
    <desrt> we recently set this up for freebsd as well
    <braunr> few maintainers take the pain :)
    <desrt> our non-linux 'problem discovery' is a bit crap before now :/
    <braunr> i guess that's pretty normal
    <braunr> i don't consider it the responsibility of the maintainers to test
      every possible platform
    <desrt> glib is a bit unique -- portability is our business
    <braunr> taking our patches into consideration is what we ask most
    <braunr> right
    <desrt> and the "please take the patches" thing is something we want to
      stop doing
    <braunr> why ?
    <desrt> mostly because we often look at a patch that someone sent a few
      years ago and say "do we even still need this?"
    <desrt> and have no way to know
    <braunr> uh
    <desrt> you would not believe how many patches like this we've
      accumulated...
    <braunr> but if we send it now ? :)
    <desrt> braunr: new policy is roughly this:
      https://wiki.gnome.org/Projects/GLib/SupportedPlatforms
    <desrt> ie: fixes for issues that are general portability improvements and
      POSIX compliance are welcome...
    <desrt> patches that introduce platform-specific #ifdef sections are
      rejected unless we have a regular builder to test that code
    <braunr> i see
    <braunr> again, regarding portability, don't consider CLOCK_MONOTONIC to be
      readily available, check for it
    <braunr> an #error would be enough but it has to be checked
    <desrt> it basically comes down to: we don't want to have code in our
      version control that we have no possible way of testing
    <braunr> yes
    <desrt> braunr: we do check for it
    <braunr> ok
    <desrt> we assert() if clock_gettime(CLOCK_MONOTONIC) fails
    <braunr> no i mean
    <desrt> as POSIX said it should if CLOCK_MONOTONIC is not supported
    <desrt> if you lie to us.... well, not much we can do
    <braunr> POSIX_MONOTONIC_CLOCK
    <braunr> _POSIX_MONOTONIC_CLOCK
    <desrt> this is actually defined to 0 on most platforms...
    <desrt> which does not mean that it's unsupported -- it means that the
      runtime must be ready to deal with it not actually existing at runtime
    <braunr> really ?
    <desrt> yes
    <desrt> we used to rely on this and got a bug that we were doing it wrong
      :)
    <desrt> and indeed, even on linux, both with glibc and uclibc:
    <desrt> /usr/include/bits/posix_opt.h:#define _POSIX_MONOTONIC_CLOCK
      0
    <desrt> /usr/include/uClibc/bits/posix_opt.h:#define _POSIX_MONOTONIC_CLOCK
      0
    <braunr> ok it's described in 2.1.6 Options
    <braunr> so your check is appropriate
    <desrt> so does clock_gettime(MONOTONIC) on debian/hurd get me realtime?
    <braunr> either that, or a value shifted from it
    <desrt> if so, i'll just hack out the condattr_setclock() check and proceed
      trying to build past glib...
    * desrt checks
    <desrt> as it is, even the build of glib fails since we use some tools
      linked against ourselves during the build process...
    <desrt> 1393124084790000 1393124084790000
    <desrt> those look the same....
    <braunr> heh
    <desrt> i also notice that your clocks are not very high precision :)
    <braunr> that's right
    <desrt> HZ = 100, i guess
    <braunr> yes
    <desrt> fair enough
    <desrt> our mainloop doesn't support better-than-millisecond accuracy yet
      anyway :)
    <desrt> (although it will soon...)
    <braunr> nice