[[!meta copyright="Copyright © 2009, 2011, 2013 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_porting]]
[[!toc]]
# `time`
Neither the `time` executable from the GNU time package work completely
correctly, nor does the GNU Bash built-in one.
tschwinge@flubber:~ $ \time sleep 2
0.00user 0.00system 9:38:00elapsed 0%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+0minor)pagefaults 0swaps
tschwinge@flubber:~ $ \time sleep 4
0.00user 0.00system 18:50:25elapsed 0%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+0minor)pagefaults 0swaps
tschwinge@flubber:~ $ \time sleep 6
0.00user 0.00system 28:00:53elapsed 0%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+0minor)pagefaults 0swaps
tschwinge@flubber:~ $ time sleep 2
real 0m2.093s
user 0m0.000s
sys 0m0.011s
tschwinge@flubber:~ $ time sleep 4
real 0m4.083s
user 0m0.000s
sys 0m0.010s
tschwinge@flubber:~ $ time sleep 6
real 0m6.164s
user 0m0.000s
sys 0m0.010s
GNU time's *elapsed* value is off by some factor.
$ \time factor 1111111111111111111
1111111111111111111: 1111111111111111111
0.00user 0.00system 52:39:24elapsed 0%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+0minor)pagefaults 0swaps
$ time factor 1111111111111111111
1111111111111111111: 1111111111111111111
real 0m11.424s
user 0m0.000s
sys 0m0.010s
As above; also here all the running time should be attributed to *user* time.
This is probably a [[!taglink open_issue_gnumach]].
## 2011-09-02
Might want to revisit this, and take Xen [[!tag open_issue_xen]] into account
-- I believe flubber has already been Xenified at that time.
### IRC, freenode, #hurd, 2011-09-02
While testing some [[performance/IPC_virtual_copy]] performance issues:
<tschwinge> And I can confirm that with dd if=/dev/zero of=/dev/null bs=4k
running, a parallel sleep 10 takes about 20 s (on strauss).
## 2013-03-30/31
Investigating time's `configure`, a difference of the output between Linux and
Hurd shows:
-checking for wait3 that fills in rusage... yes
+checking for wait3 that fills in rusage... no
This causes a different code path in `resuse.c` to be used; such code path does
not get a define for `HZ`, which is then defined with a fallback value of 60.
[[!debbug 704283]] has been filed with a fix for this no-wait3 case.
# `times`
## guile
### IRC, freenode, #hurd, 2013-08-21
<nalaginrut> does guile2 on hurd fixed? times issue
<teythoon> nalaginrut: does not look good
<teythoon> scheme@(guile-user)> (times)
<teythoon> $1 = #(0 0 0 0 0)
<nalaginrut> well, seems not a fixed version, if there's fixed version
<nalaginrut> since it's not Guile's bug, I can do nothing for it
<teythoon> ah
<nalaginrut> in spite of this, Guile2 works I think
<nalaginrut> all tests passed but 2 fail
<nalaginrut> one of the failure is version shows "UNKNOWN" which is
trivials
<teythoon> well, did you try to fix the times issue in Hurd?
<nalaginrut> I didn't , I have to get more familiar with hurd first
<nalaginrut> I'm playing hurd these days
<teythoon> :)
<nalaginrut> anyway, I think times issue is beyond my ability at present
<nalaginrut> ;-P
<teythoon> times is implemented in the glibc, in sysdeps/mach/hurd/times.c
<teythoon> don't say that before you had a look
<nalaginrut> yes, you're right
<nalaginrut> but I think times has something to do with the kernel time
mechanism, dunno if it's related to the issue
<nalaginrut> how did you get the times.c under hurd?
<nalaginrut> apt-get source glibc?
<teythoon> well, I'd clone git://sourceware.org/git/glibc.git
<teythoon> and yes, the kernel is involved
<teythoon> task_info is used to obtain the actual values
<teythoon>
http://www.gnu.org/software/hurd/gnumach-doc/Task-Information.html
<teythoon> I'd guess that something fails, but the times(2) interface is
not able to communicate the exact failure
<nalaginrut> maybe it's not proper to get src from upstream git? since it's
OK under Linux which uses it too
<nalaginrut> but apt-get source glibc has nothing
<teythoon> so I would copy the times(2) implementation from the libc so
that you can modify it and run it as a standalone program
<teythoon> well, the libc has system dependent stuff, times(2) on Linux is
different from the Hurd version
<teythoon> it has to be
<nalaginrut> alright, I got what you mean ;-)
<teythoon> and the debian libc is built from the eglibc sources, so the
source package is called eglibc iirc
<nalaginrut> ah~I'll try
<teythoon> have you tried to rpctrace your times test program? the small c
snippet you posted the other day?
<nalaginrut> I haven't build all the tools & debug environment on my hurd
;-(
<teythoon> what tools?
<nalaginrut> well, I don't even have git on it, and I'm installing but
speed is slow, I'm looking for a new mirror
<teythoon> ah well, no need to do all this on the Hurd directly
<teythoon> building the libc takes like ages anyway
<nalaginrut> oops ;-)
<nalaginrut> I'll take your advice to concentrate on times.c only
<teythoon> oh well, it might be difficult after all, not sure though
<teythoon> times sends two task_info messages, once with TASK_BASIC_INFO,
once with TASK_THREAD_TIMES_INFO
<teythoon> here is the relevant rpctrace of your test program:
<teythoon> task131(pid14726)->task_info (1 10) = 0 {0 25 153427968 643072 0
0 0 0 1377065590 570000}
<teythoon> task131(pid14726)->task_info (3 4) = 0 {0 0 0 10000}
<teythoon> ok, I don't know enough about that to be honest, but
TASK_THREAD_TIMES_INFO behaves funny
<teythoon> I put a sleep(1) into your test program, and if I rpctrace it,
it behaves differently o_O
* nalaginrut is reading task-information page to get what it could be
<nalaginrut> maybe I have to do the same steps under Linux to find some
clue
<teythoon> no, this is Mach specific, there is no such thing on Linux
<teythoon> on Linux, times(2) is a system call
<teythoon> on Hurd, times is a function implemented in the libc that
behaves roughly the same way
<nalaginrut> OK~so different
<teythoon> look at struct task_basic_info and struct task_thread_times_info
in the task-information page for the meaning of the values in the
rpctrace
<teythoon> yes, very
<braunr> nalaginrut: you may want to try a patch i did but which is still
waiting to be merged in glibc
<nalaginrut> braunr: ah~thanks for did it ;-)
<nalaginrut> can I have the link?
<braunr> i'm getting it
<braunr> teythoon: funny things happen with rpctrace, that's expected
<braunr> keep in mind rpctrace doesn't behave like ptrace at all
<braunr> it acts as a proxy
<braunr> nalaginrut:
http://git.savannah.gnu.org/cgit/hurd/glibc.git/commit/?h=rbraun/getclktck_100_hz&id=90404d6d1aa01f6ce1557841f5a675bb6a30f508
<braunr> nalaginrut: you need to add it to the debian eglibc patch list,
rebuild the packages, and install the resulting .debs
<braunr> if you have trouble doing it, i'll make packages when i have time
<nalaginrut> braunr: I think your test result is expected? ;-)
<braunr> what test result ?
<nalaginrut> times test under that patch
<braunr> yes
<braunr> but i have no idea if it will work
<braunr> my patch fixes a mismatch between glibc and the procfs server
<braunr> nothing more
<braunr> it may help, it may not, that's what i'd like to know
<nalaginrut> hah~thanks for that
<nalaginrut> I get source from apt-get, then manually modified the files,
no much code ;-)
<nalaginrut> compiling
<nalaginrut> there is no cpuinfo in /proc?
<teythoon> no
<nalaginrut> a feature need to be done? or there's another way for that?
<teythoon> well, it hasn't been implemented
<teythoon> do you need that? what for?
<nalaginrut> compiling error, I realized I should use gcc-4.7
<pinotree> how are you building?
<nalaginrut> I just happened to play proc while compiling, and found
there's no
<nalaginrut> cxa_finalize.c:48:1: error: tcbhead_t has no member
named multiple_threads
<nalaginrut> I changed to gcc-4.7
<pinotree> just edit the sources, and then dpkg-buildpackage -nc -us -uc
<pinotree> that will rebuild the debian package as it would be in a debian
build, making sure all the build dependencies are there, etc
<pinotree> doing it different than that is just wrong™
<nalaginrut> ok, doing
<pinotree> were you really doing ./configure etc yourself?
<nalaginrut> well, I can't wait till it's done, I'll let it compile and
check it out tomorrow
<nalaginrut> I used configure, yes ;-P
<pinotree> not good
<nalaginrut> I have to go, thanks for help guys
### IRC, freenode, #hurd, 2013-08-22
< nalaginrut> eglibc was done by dpkg-buildpackage, then how to install it?
(sorry I'm a brand new debian users)
< nalaginrut> oh~I found it
< nalaginrut> yes, (times) returns reasonable result ;-)
* nalaginrut is trying 'make check'
< nalaginrut> unfortunately, it can't pass the test though, I'm researching
it, anyway, we made first step
< nalaginrut> for Hurd internal-time-units-per-second will be 1000
< nalaginrut> , but the elapsed time is far larger than (* 2
internal-time-units-per-second)
< nalaginrut> I think the different of two returned clocks after 1 second
should be the TIME_UNITS_PER_SECOND, in principle
< nalaginrut> but I'm not sure if it's elibc or Guile bug
< nalaginrut> dunno, maybe clock tick should be 1000?
< nalaginrut> well, I'll try clock per second as 1000
< braunr> nalaginrut: clock tick (or actually, the obsolete notion of a
clock tick in userspace) should be 100
< braunr> nalaginrut: how did you come with 1000 ?
< nalaginrut> braunr: Guile set TIME_UNITS_PER_SECOND to 1000 when there's
no 8bytes size and doesn't define HAVE_CLOCK_GETTIME
< nalaginrut> #if SCM_SIZEOF_LONG >= 8 && defined HAVE_CLOCK_GETTIME
< nalaginrut> #define TIME_UNITS_PER_SECOND 1000000000
< nalaginrut> #else
< nalaginrut> #define TIME_UNITS_PER_SECOND 1000
< nalaginrut> #endif
< nalaginrut> and the test for 'times' used time-units-per-second
< pinotree> what has sizeof(long) have to do with time units per second?
< nalaginrut> dunno, maybe the representation of time?
< nalaginrut> the test failed since the difference between two clocks after
1sec is too large
< nalaginrut> and for the test context, it should small than 2 times of
units-per-second
< nalaginrut> should be smaller
< nalaginrut> sorry for bad English
< pinotree> aren't you basically looking for clock_getres?
< nalaginrut> pinotree: I don't understand what you mean
< pinotree>
http://pubs.opengroup.org/onlinepubs/9699919799/functions/clock_getres.html
< nalaginrut> I wonder if there's a standard CLK_PER_SEC for Hurd
< nalaginrut> or it can be modified as wish
< pinotree> why do you need it?
< nalaginrut> the difference is 10,000,000, which can never less than
2*clock_per_second
< nalaginrut> pinotree: I don't need it, but I want to know if there's a
standard value
< braunr> nalaginrut: ok so, this is entirely a guile thing
< braunr> nalaginrut: did you test with my patch ?
< nalaginrut> braunr: yes, 'times' works fine
< braunr> but even with that, a tets fails ?
< braunr> test*
< nalaginrut> well, I can't say works fine, the proper description is "now
it has reasonable result"
< braunr> youpi: could you bring
http://darnassus.sceen.net/gitweb/savannah_mirror/glibc.git/commit/90404d6d1aa01f6ce1557841f5a675bb6a30f508
into debian glibc btw ?
< nalaginrut> braunr: it failed the test since the clock run too fast, but
it should be smaller than 2*clk-per-sec
< braunr> i don't get that
< braunr> can you show the code that checks the condition ?
< nalaginrut> braunr: http://pastebin.com/sG3QxnPt
< braunr> * 0.5 internal-time-units-per-second ?
< nalaginrut> for C users, it's just like
a=times(...);sleep(1);b=times(...); then time-units-per-sec/2 <= (b-a) <=
time-units-per-sec*2
< braunr> ah ok
< nalaginrut> the test passes when it's true
< braunr> so basically, it says sleep(1) sleeps for more than 2 seconds
< braunr> can you check the actual value ?
< braunr> b-a
< nalaginrut> hold on for minutes
< nalaginrut> it's 10,000,000
< nalaginrut> for clk-per-sec=1000,000,000, it's OK
< nalaginrut> but for 100 or 1000, it's too small
< braunr> let's forget 100
< braunr> guile uses 1000
< nalaginrut> OK
< braunr> but i still don't get why
< nalaginrut> so I asked if there's standard value, or it can be ajustified
< nalaginrut> adjusted
< braunr> ok so, times are expressed in clock ticks
< braunr> are you sure you're using a patched glibc ?
< nalaginrut> yes I used your patch, and the 'times' get reasonable result
< braunr> then
< braunr> 11:28 < nalaginrut> it's 10,000,000
< braunr> doesn't make sense
< nalaginrut> hmm
< braunr> anhd i don't understand the test
< braunr> what's tms:clock new ?
< nalaginrut> it's actually the return value of 'times'
< nalaginrut> Guile wrap the clock_t and tms to a vector, then we can get
all the thing in a row
< nalaginrut> 'new' is a variable which was gotten after 1 sec
< braunr> let's see what this does exactly
< nalaginrut> equal to "new = times(...)"
< nalaginrut> 'tms' equal to (clock_t (struct tms))
< nalaginrut> we have to pass in the struct pointer to get the struct
values filled, but for Guile we don't use pointer, times actually returns
two things: clock_t and struct tms
< nalaginrut> and Guile returns them as a vector in a row, that's it
< braunr> nalaginrut: test this please:
http://darnassus.sceen.net/~rbraun/test.c
< braunr> i don't have a patched libc here
< braunr> i'll build one right now
< nalaginrut> clock ticks: 1000000
< braunr> and this seems reasonable to you ?
< braunr> anyway, i think the guile test is bugged
< nalaginrut> no, the reasonable is not for this
< braunr> does it ever get the clock tick value from sysconf() ?
< nalaginrut> I say reasonable since it's always 0 both for clock and tms,
before apply your patch
< braunr> uh no
< braunr> i have the same value, without my patch
< nalaginrut> so I said "I can't say it works fine"
< braunr> either the test is wrong because it doesn't use sysconf()
< nalaginrut> anyway, I don't think times should return "all zero"
< braunr> or the clock values have already been ocnverted
< braunr> but it doesn't
< braunr> you did something wrong
< nalaginrut> with your patch it doesn't
< braunr> without neither
< braunr> 11:43 < braunr> i have the same value, without my patch
< nalaginrut> well, it's too strange
< braunr> check how the test actually gets the clock values
< braunr> also, are your running in vbox ?
< braunr> you*
< nalaginrut> no ,it's physical machine
< braunr> oh
< braunr> nice
< braunr> note that vbox has timing issues
< nalaginrut> I thought I should give you some info of CPU, but there's no
/proc/cpuinfo
< braunr> shouldn't be needed
< nalaginrut> OK
< braunr> run my test again with an unpatched glibc
< braunr> just to make sure it produces the same result
< braunr> and
< nalaginrut> so the clock-per-sec is machine independent for Hurd I think
< braunr> 11:46 < braunr> check how the test actually gets the clock values
< nalaginrut> since it's implemented in userland
< braunr> clock-per-sec is always system dependent
< braunr> All times reported are in clock ticks.
< braunr> The number of clock ticks per second can be obtained
using:
< braunr> sysconf(_SC_CLK_TCK);
< braunr> 11:46 < braunr> check how the test actually gets the clock values
< braunr> to see if they're converted before reaching the test code or not
* nalaginrut is building eglibc
< braunr> building ?
< braunr> what for ?
< nalaginrut> I modified it to 1000, now it's useless
< braunr> we want it to 100 either way
< nalaginrut> and how to reinstall eglibc under debian?
< braunr> it's obsolete, procfs already uses 100, and 100 is low enough to
avoid overflows in practically all cases
< braunr> aptitude install libc0.3=<version>
< nalaginrut> OK
< braunr> aptitude show -v libc0.3
< braunr> for the list of available versions
< nalaginrut> out of topic, what's the meaning of the code in
quantize_timeval ?
< nalaginrut> tv->tv_usec = ((tv->tv_usec + (quantum - 1)) / quantum) *
quantum;
< nalaginrut> I can't understand this line
< braunr> scaling and rounding i guess
< nalaginrut> hmm...but quantum seems always set to 1?
< nalaginrut> 100/__getclktck()
< braunr> ah right
< braunr> old crap from the past
< nalaginrut> and clk-tck is 100
< braunr> the author probably anticipated clk_ticks could vary
< braunr> in practice it doesn't, and that's why it's been made obsolete
< nalaginrut> I wonder if it could be vary
< braunr> no
< nalaginrut> alright
< nalaginrut> why not just assign it to 1?
< braunr> 11:55 < braunr> old crap from the past
< braunr> the hurd is 20 years old
< braunr> like linux
< nalaginrut> oh~
< braunr> but with a lot less maintenance
< nalaginrut> braunr: well, I tried the original eglibc, your test was
clock ticks: 1000000
< nalaginrut> but in Guile, (times) ==> (0 0 0 0 0)
< nalaginrut> the reasonable result maybe: #(4491527510000000 80000000 0 0
0)
< braunr> 11:46 < braunr> check how the test actually gets the clock values
< braunr> ah, he left
### IRC, freenode, #hurd, 2013-08-23
< braunr> nalaginrut: times() doesn't seem to be affected by my patch at
all
< nalaginrut> braunr: but it did in my machine
< nalaginrut> well, I think you mean it doesn't affect your C test code
< braunr> i'm almost sure something was wrong in your test
< braunr> keep using the official debian glibc package
< nalaginrut> I don't think it's test issue, since every time (times)
return zero, the test can never get correct result
< braunr> times doesn't return 0
< braunr> for sleep(1), i always have the right result, except in
microseconds
< nalaginrut> times in Guile always return #(0 0 0 0 0)
< braunr> (microseconds is the native mach time unit)
< braunr> well, guile does something wrong
< nalaginrut> after sleep 1, it's 0 again, so it's none sense
< braunr> 11:46 < braunr> check how the test actually gets the clock values
< braunr> not on my system
< nalaginrut> but (times) returns reasonable result after applied your
patch
< braunr> that's not normal, since times isn't affected by my patch
< nalaginrut> oops
< braunr> you need to look for what happens in guile between the times()
call and the #(0 0 0 0 0) values
< nalaginrut> well, I tried many times between patch or non-patch, I think
there's no mistake
< nalaginrut> I read the 'times' code in Guile, there's nothing strange,
just call 'times' and put all the result to a vector
< braunr> which means there is no conversion
< braunr> in which case the test is plain wrong since there MUST also be a
call to sysconf()
< braunr> to obtain the right clock ticks value
< braunr> is your box reachable with ssh ?
< nalaginrut> oh~wait, seems there's a quotient operation, I'm checking
< nalaginrut> factor = scm_quotient (scm_from_long (TIME_UNITS_PER_SECOND),
< nalaginrut> scm_from_long (ticks_per_second));
< braunr> iirc, TIME_UNITS_PER_SECOND is hardcoded
< nalaginrut> unless factor is zero
< nalaginrut> yes, it's hardcoded
< braunr> that's completely non portable and wrong
< nalaginrut> you suggest to call sysconf?
< braunr> yes
< braunr> but i don't have the code in mind
< braunr> what is ticks_per_second ?
< nalaginrut> OK, that's one issue, we have to find why times return 0
< braunr> 14:14 < braunr> is your box reachable with ssh ?
< braunr> i'd like to make sure times returns 0 at your side
< braunr> because it doesn't at mine
< nalaginrut> no
< braunr> until i can reproduce, i can't consider there is a problem
< nalaginrut> I think it's unreachable for outer space
< nalaginrut> well, if you want to reproduce, just get guile src of debian
< braunr> guile 2.0 ?
< nalaginrut> yes, apt-get source guile-2.0
< nalaginrut> I'm checking ticks_per_second
< braunr> got the source, how do i test
< braunr> ?
< nalaginrut> you have to build it, and run ./meta/guile, then you don't
have to install it
< nalaginrut> and try (times)
< braunr> aw libgc
< nalaginrut> the reasonable result should be #(4313401920000000 110000000
20000000 0 0) or something alike
< nalaginrut> but #(0 0 0 0 0) in each time is not reasonable apparently
< nalaginrut> maybe you need apt-get build-dep guile-2.0?
< braunr> already done
< nalaginrut> building Guile2 may take very long time
< nalaginrut> about 30 minutes in my old machine
< braunr> then it should take just a few minutes on mine
< nalaginrut> alright it's not very long, I've spent 8 hours for gcc in LFS
< braunr> 8 hours ?
< braunr> takes 5-10 minutes on a common machine ..
< nalaginrut> but it's Celeron566 at that time...
< braunr> ah, that old
< nalaginrut> include bootstrap, so very long
< braunr> nalaginrut: i got the test failure from the build procedure, how
do i run it manually ?
< nalaginrut> braunr: ./meta/guile -L test-suite
test-suite/tests/time.test
< nalaginrut> braunr: or make check for all
< braunr> put a print after the schedule() and before the return nil; in
runtime_mstart, since that's the body of new threads
< nlightnfotis> unfortunately, I can't confirm this with goroutines
running; the assertion failure aborts before I can get anything useful
< braunr> you can
< braunr> make sure there is a \n in the message, since stdout is line
buffered by default
< braunr> if you don't reach that code, it means threads don't exit
< braunr> at least goroutine threads
< braunr> btw, where is the main thread running ?
< nlightnfotis> I just checked there is a \n at the end.
< nlightnfotis> "<braunr> btw, where is the main thread running " could you
elaborate a little bit on this?
< braunr> what does main() after initializing the runtime ?
< braunr> +do
< nlightnfotis> the runtime main or the process's main?
< braunr> the process
< braunr> nlightnfotis: what we're interested in is knowing whether main()
exits or not
< nlightnfotis> braunr: I can see there are about 4 functions of interest:
runtime_main (the main goroutine, and I can imagine 1st thread)
< nlightnfotis> main_init (I don't know what it does, will check this out
now)
< nlightnfotis> main_main (not sure about this one either)
< nlightnfotis> and runtime_exit (0)
< braunr> i can see that too
< braunr> i'm asking about main()
< nlightnfotis> which seems to be the function that terminates the main
thread
< nlightnfotis> <braunr> nlightnfotis: what we're interested in is knowing
whether main() exits or not --> my theory is runtime_exit (0) exits the
process' main. Seeing as at various times go programs echo $? == 0.
< nlightnfotis> let me research that a little bit
< nlightnfotis> braunr: that will require a bit more studying. main_main()
and main_init() are both expanded to assembly tags if I understand it
correctly.
< nlightnfotis> main.main and __go_init_main respectively.
< braunr> why are you looking from there instead of looking from main() ?
< nlightnfotis> are we not looking out if main exits?
< braunr> we are
< braunr> so why look at main_main ?
< braunr> or anything else than main ?
< nlightnfotis> these are called inside runtime_main and I figured out they
might have a clue
< braunr> runtime_main != main
< braunr> (except if there is aliasing)
< nlightnfotis> there is still the possibility that runtime_main is the
main function and that runtime_exit(0) exits it.
< braunr> there is no doubt that main is main
< braunr> (almost)
< nlightnfotis> and I just found out that there is no main in assembly
produced from go. Only main.main
< braunr> check the elf headers for the entry point then
< nlightnfotis> braunr: I went through the headers, and found the process'
main. You can find it in <gcc_root>/libgo/runtime/go-main.c
< nlightnfotis> it seems very strange though: It creates a new thread, then
aborts?
< braunr> nlightnfotis: see :)
< braunr> nlightnfotis: add traces there
< nlightnfotis> braunr: can you look into that piece of code to check out
something I don't understand?
< nlightnfotis> braunr: I can not seem able to find __go_go 's definition
< nlightnfotis> only a declaration in runtime.h
< braunr>
https://github.com/NlightNFotis/gcc/blob/master/libgo/runtime/proc.c,
line 1552
< nlightnfotis> gee thanx. For a strange kind of fashion, I was looking for
it in runtime.c
< braunr> use git grep
< braunr> or tags/cscope
< nlightnfotis> braunr: yep! runtime_exit does seem to terminate a go
process that was not otherwise abnormally terminated.
< braunr> ?
< braunr> is it called or not ?
< braunr> runtime_exit is a macro on exit()
< braunr> so we already know what it does
< nlightnfotis> it is called
< braunr> ok
< braunr> that's not normal :)
< nlightnfotis> for a simple program
< braunr> uh ?
< nlightnfotis> for one that has a go routine
< braunr> but
< nlightnfotis> it doesn't
< nlightnfotis> it's expected
< braunr> ok
< braunr> that makes sense
< braunr> well, trace
< braunr> keep tracing
< braunr> for example in main()
< braunr> is runtime_mstart() actually reached ?
< nlightnfotis> yeah main and runtime_main were my next two targets
< braunr> good
< nlightnfotis> and now I followed your advice and it does compiler much
faster
< braunr> so, it looks like the main thread just becomes a mere kernel
thread
< braunr> running runtime_mstart() and fetching goroutines as needed
< braunr> after your traces, i'd suggest running a small go test program,
with one simple goroutine (doesn't crash right ?)
< braunr> and trace context switching
< braunr> but after the traces
< braunr> one important trace is to understand why runtime_exit gets called
< nlightnfotis> it does crash even with 1 goroutine
< braunr> oh
< braunr> when doesn't it crash ?
< nlightnfotis> when it has 0 goroutines
< nlightnfotis> it works as expected
< nlightnfotis> but anything involving goroutines crashes
< nlightnfotis> and goroutines are very important; everything in the
standard library involves goroutines
< braunr> ok
< braunr> doesn't change what i suggested, good
< braunr> 1/ find out why runtime_exit gets called
< braunr> 2/ trace context switching with 1 goroutine
< nlightnfotis> on it.
< braunr> in all cases, make all your goroutines (including the main one)
*not* return
< braunr> so that you don't deal with goroutine destruction yet
< nlightnfotis> runtime_mstart in main doesn't to be run at all. So the
path is __go_go and then return from it.
< nlightnfotis> *doesn't seem
### IRC, freenode, #hurd, 2013-08-26
< braunr> youpi: my glibc clock patch looks incomplete btw
< youpi> which one?
< youpi> ah, the ticks one?
< braunr> yes
< braunr> it doesn't change the values returned by times
< braunr> as a side effect, the load average bumps to 2+ on an idle machine
### IRC, freenode, #hurd, 2013-08-27
< nalaginrut> braunr: have you tried Guile2 on your machine? ;-)
< braunr> nalaginrut: no
< braunr> nalaginrut: but i saw the code actually does use sysconf()
< nalaginrut> braunr: yes, for ticks_per_second
< braunr> i had to look myself to find it out, you didn't say it, despite
me asking multiple times
< braunr> it won't make debugging easier ;p
< braunr> nalaginrut: also, the return value of times is actually *never*
used
< braunr> i don't know why you've been talking about it so much
< nalaginrut> braunr: I'm sorry, it's first time to look stime.c for me
< braunr> the interesting function is get_internal_run_time_times()
< nalaginrut> what do you mean about "the return value of times is actually
*never* used"? in which context?
< braunr> see get_internal_run_time_times
< braunr> struct tms time_buffer;
< braunr> times(&time_buffer);
< braunr> return ...
< braunr> and yes, the user and system time reported in struct tms are 0
< braunr> let's see what posix has to say about it
< pinotree> it says it will return (clock_t)-1 for errors, but no standard
errors are defined yet
< nalaginrut> but I don't think get_internal_run_time_times has something
to do with scm_times
< braunr> well, i don't see any other call to times()
< braunr> i've asked you repeatedly to look for how guile fetches the data
< braunr> i think it's done in get_internal_run_time_times
< braunr> what makes you think otherwise ?
< braunr> our times() seems to behave fine, other than the units of the
return value
< nalaginrut> I don't understand what do you mean?
get_internal_run_time_times is unrelated to scm_times which is actually
"times" in Scheme code
< braunr> ok
< nalaginrut> I think we're talking about "times" activity, right?
< braunr> ok so result is a vector
< braunr> with the return value and the four values in struct tms
< nalaginrut> yes
< braunr> and what looks interesting is
< braunr> factor = scm_quotient (scm_from_long (TIME_UNITS_PER_SECOND),
scm_from_long (ticks_per_second));
< braunr> SCM_SIMPLE_VECTOR_SET (result, 0, scm_product (scm_from_long
(rv), factor));
< braunr> TIME_UNITS_PER_SECOND is 1000
< nalaginrut> yes, it means (clock_t *
(TIME_UNITS_PER_SECOND/ticks_per_second)), though I've no idea why it
does this
< braunr> normalizing values i guess
< nalaginrut> I wonder if the factor should be 1, just guessing
< braunr> let's see what our clock tick really is
< braunr> 1000000 on an unmodified libc
< braunr> 100 with my patch
< nalaginrut> so what's the problem?
< nalaginrut> all the values were multiplied by ticks, it's fair for the
subtraction
< nalaginrut> I think the problem is clock is too large for the difference
between utime and utime(sleep 1)
< nalaginrut> oops, is too small
< nalaginrut> sorry, I confused,
< nalaginrut> the problem is the difference of clock is too large for
2*internal-time-units-per-second
< nalaginrut> and actually, internal-time-units-per-second is
SCM_TIME_UNITS_PER_SECOND
< nalaginrut> but without your patch, 'times' would return zeros all the
time, which is never meet the condition: SCM_TIME_UNITS_PER_SECOND/2 <=
(clock2 - clock1)
< nalaginrut> well, maybe your point is
TIME_UNITS_PER_SECOND/ticks_per_second is too small without your patch,
which causes the scm_to_long cast give a 0 value
< nalaginrut> s/cast/casting
< nalaginrut> when ticks_per_second is 100, the factor would be 10, which
seems to be reasonable
< nalaginrut> s/scm_to_long/scm_from_long
< nalaginrut> well, I have to checkout this
< nalaginrut> OK, let me reconstruct the point: ticks_per_second so too
large that makes the factor becomes zero
< nalaginrut> but decrease ticks_per_second to 100 causes the clock become
too large than TIME_UNITS_PER_SECOND
< braunr> 10:59 < nalaginrut> but without your patch, 'times' would return
zeros all the time, which is never meet the condition:
SCM_TIME_UNITS_PER_SECOND/2 <= (clock2 - clock1)
< braunr> until you prove me otherwise, this is plain wrong
< braunr> times() never returned me 0
< braunr> so let's see, this gives us a factor of 1000 / 1000000
< braunr> so the problem is factor being 0
< braunr> that's why *guile* times returns 0
< braunr> with my patch it should return 10
< nalaginrut> braunr: I'm sorry I mean "stime" in Scheme returns zeros
< nalaginrut> yes, I think the problem is factor
< nalaginrut> the factor
< braunr> now why doesn't my patch fix it all ?
< braunr> ah yes, rv is still in microseconds
< braunr> that's what i've been telling youpi recently, my patch is
incomplete
< braunr> i'll cook a quick fix, give me a few minutes please
< nalaginrut> but it fixed something ;-)
< braunr> well, guile makes a stupid assumption here
< braunr> so it's not really a fix
< nalaginrut> braunr: should I ask some info about TIME_UNITS_PER_SECOND
from Guile community?
< nalaginrut> or it doesn't help
< braunr> what do you want to ask them ?
< nalaginrut> since I don't know how this value was chosen
< nalaginrut> dunno, I'll ask if you need it
< nalaginrut> I just think maybe you need this info
< braunr> well
< braunr> my plan is to align the hurd on what other archs do
< braunr> i.e. set clk_tck to 100
< braunr> in which case this won't be a problem any more
< braunr> now you could warn them about the protability issue
< braunr> i'm not sure if they would care though
< nalaginrut> the warning is useful for the future
< nalaginrut> and it's not hard to make a change I think, for a constant,
but it depends on the maintainers
< braunr> it's not that simple
< braunr> time related things can easily overflow in the future
< nalaginrut> alright
< braunr> refer to the 2038 end-of-the-world bug
< nalaginrut> so how can I describe the warning/suggestion to them?
< braunr> i'm not sure
< braunr> tell them the TIME_UNITS_PER_SECOND isn't appropriate for larger
values of clk_tck
< braunr> dammit, microseconds are hardcoded everywhere in
sysdeps/mach/hurd ... >(
< braunr> nalaginrut: my new patch seems to fix the problem
< braunr> nalaginrut: i've built debian packages with which you can
directly test
< braunr> nalaginrut: deb http://ftp.sceen.net/debian-hurd-i386
experimental/
< braunr> Totals for this test run:
< braunr> passes: 38605
< braunr> failures: 0
< braunr> unexpected passes: 0
< braunr> expected failures: 7
< braunr> unresolved test cases: 578
< braunr> untested test cases: 1
< braunr> unsupported test cases: 10
< braunr> errors: 0
< braunr> PASS: check-guile
< braunr> =============
< braunr> 1 test passed
< braunr> =============
< braunr> :)
< braunr> youpi: the branch i added to glibc contains a working patch for
clock_t in centiseconds
< youpi> k
### IRC, freenode, #hurd, 2013-08-28
<nalaginrut> braunr: well, looks great! I'll try it soon~
<nalaginrut> braunr: BTW, where is the patch/
<mark_weaver> braunr: what was needed to get guile working on the hurd?
<mark_weaver> well, if the fix wasn't to guile, I don't need the details.
<braunr> 04:53 < nalaginrut> braunr: BTW, where is the patch/
<braunr> there is hardly anyone here at 5am
<braunr> nalaginrut:
http://git.savannah.gnu.org/cgit/hurd/glibc.git/log/?h=rbraun/clock_t_centiseconds
<nalaginrut> braunr: thanks for that, but why not use a constant for 100?
<braunr> nalaginrut: i don't know where to define it
<braunr> it's glibc, you don't define new stuff mindlessly
<youpi> braunr: about your centiseconds patch, did you run the libc
testsuite with it?
<mark_weaver> it does seem a shame to reduce the resolution of the timers
from microseconds to centiseconds. I wonder if that could be avoided.
<youpi> by fixing all applications which assume centiseconds
<mark_weaver> *nod* well, if there's such a problem in Guile, I'd be glad
to fix that.
<braunr> youpi: no
<mark_weaver> I see that there's a macro CLOCKS_PER_SEC that programs
should consult.
<youpi> braunr: ok, I'll do then
<braunr> mark_weaver: why is it a shame ?
<braunr> it's not clock or timer resolution
<youpi> it's clock_t resolution
<braunr> it's an obsolete api to measure average cpu usage
<braunr> having such a big value on the other hand reduces the cpu usage
durations
<mark_weaver> braunr: good point :) I confess to being mostly ignorant of
these APIs.
<mark_weaver> Though Guile should still consult CLOCKS_PER_SEC instead of
assuming centiseconds. If it's making an improper assumption, I'd like
to know so I can fix it.
<braunr> the improper assumption is that there are less than 1000 clock
ticks per second
<mark_weaver> do you know off-hand of some code in Guile that is making
improper assumptions?
<braunr> yes
<braunr> let me find it
<mark_weaver> thanks
<braunr> factor = scm_quotient (scm_from_long (TIME_UNITS_PER_SECOND),
<braunr> scm_from_long (ticks_per_second));
<braunr> it seems guile attempts to normalize all times values to
TIME_UNITS_PER_SECOND
<braunr> while i think it would be better off using ticks_per_second (clock
ticks as provided by sysconf())
<braunr> attempting to normalize here causes factor to become 0 if
TIME_UNITS_PER_SECOND < ticks_per_second
<mark_weaver> ah, I see.
<mark_weaver> I'll take care of it. thanks for the pointer!
<youpi> braunr: I've commited the centisecond patch to debian's glibc
### IRC, freenode, #hurd, 2013-08-29
<nalaginrut> braunr: Guile2 works smoothly now, let me try something cool
with it
<braunr> nalaginrut: nice
### IRC, OFTC, #debian-hurd, 2013-09-29
<pinotree> youpi: is the latest glibc carrying the changes related to
timing? what about gb guile-2.0 with it?
<youpi> it does
<youpi> so that was the only issue with guile?
<youpi> well at least we'll see
<pinotree> iirc yes
<pinotree> according to nalaginrut and the latest build log, it'd seem so
<youpi> started
<youpi> yay, guile-2.0 :)
<pinotree> yay