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

<braunr> now that there is a pthread_hurd_cond_timedwait_np function
  available, we could replace the ulgy timers in pfinet

IRC, freenode, #hurd, 2013-04-02

<braunr> youpi: also, i think i know why fakeroot is slow on the hurd
<braunr> well, pfinet actually
<youpi> tcp flow control?
<braunr> i think it's because of our timer resolution
<youpi> ah
<braunr> and i think i spotted a small mistake in the timer emulation code
<braunr> it's so obvious i even doubt it's actually true :)
<braunr> see timer_emul.c:timer_function
<braunr> the code checks for timers->expires < jiffies
<braunr> shouldn't it be "<="
<braunr> ?
<youpi> well it'd be equivalent
<youpi> if ==, then wait becomes 0 in the else
<braunr> see the second scheck
<braunr> the one performed right before running the callback
<youpi> ah, the while?
<braunr> yes
<youpi> I'd say <= could do it yes
<braunr> ok
<braunr> i have hurd packages ready to be tested
<youpi> although I'm unsure it'd make a difference
<youpi> do you notice some?
<braunr> just waiting for my current glibc to finish building
<braunr> and i'll test right after
<braunr> i think it would actually
<youpi> in which case?
<braunr> the timer resolution is 10ms
<braunr> it's huge
<braunr> well, i hope it fixes fakeroot slowness :)
<youpi> so timers below that would trigger immediately?
<braunr> or equal
<braunr> taht's the problem
<braunr> for 10ms, timers that have expired must wait for the next tick to
<youpi> I include "equal" in below
<youpi> actually :)
<braunr> then yes :)
<braunr> soryy i never know when equal is implicit
<youpi> because they boil down to expires = jiffies
<youpi> in french it's implicit
<youpi> but anyway here equal is not really important
<braunr> right
<braunr> why ?
<youpi> it's the fact that 1ms would be rounded up to 10ms, not down to 0ms
<braunr> well, doing that seems reasonable
<youpi> or rather, rounded down to 0ms, but waited for 10ms because of the
<braunr> we do want timers not to fire before the time event
<youpi> I'm however wondering which part of the code would have timer below
<braunr> well, a select with low timeout from a client perhaps ?
<youpi> but then we have to round up
<youpi> as you say we don't want to fire before some time
<braunr> well yes
<braunr> that's fine
<youpi> all that being said, linux has been having 10ms clock for a long
<braunr> but when the timer fires, i.e. when expires == jiffie, we do want
  it to fire
<youpi> highres timers are only relativeley recent
<braunr> not at jiffie + 1
<youpi> braunr: ah, right, so we don't wait for 20ms instead of 10ms
<braunr> yes
<youpi> ok, so it's not a miracle fix, but could bring 2times fix
<braunr> well, pfinet eats 40% of my processor when this problem occurs
<braunr> which i'm seeing right now
<youpi> yes, I've seen that too
<braunr> so it may be a visible win
<braunr> build finished :) let's see
<youpi> Mmm, thinking again
<youpi> indeed
<youpi> when expires become jiffies
<youpi> we mach_msg (0)
<youpi> but don't do anything
<youpi> and then restart
<youpi> so just eaten cpu for nothing
<braunr> is there a small package that builds fast and uses fakeroot a lot
<youpi> something like a package which has a lot of data to install at make
<braunr> or i can rebuild glibc with -nc
<youpi> ok, I've checked in linux
<pinotree> libarchive's testsuite used to intermittently failing under
<youpi> it's definitely <= which is meant
<braunr> it looks better
<braunr> but still not 100% cpu
<youpi> at any rate, as long as it doesn't seem to break anything, please
  push the fix
<youpi> it definitely makes sense
<youpi> (and I don't see why it could break anything)
<braunr> ok
<braunr> i'll look into this timer problem a bit more, there may be other
  code involved
<braunr> yes, schedule_timeout could need a review
<braunr> actually, fakeroot rm -rf * is a good test
<braunr> and it's still damn slow

IRC, freenode, #hurd, 2013-11-04

<braunr> i think i know why fakeroot is slow no
<braunr> now
<braunr> schedule_timeout as implemented in pfinet can only be awaken by a
<braunr> even when the expected even comes in earlier
<braunr> so yes, the proper solution is to rewrite the timers using
  interruptible_sleep_on_timeout (and in turn
<braunr> hm no, it's still not that straightforward :(

IRC, freenode, #hurd, 2013-11-05

<braunr> youpi: i found the bug slowing down fakeroot-tcp
<braunr> it's actually a bug that slows down anything using the loopback
<braunr> (although there still is a problem with fakeroot chown)
<youpi> oh!
<braunr> basically
<braunr> the loopback device calls netif_rx from its xmit function
<braunr> which is perfectly fine
<braunr> except the glue code makes mark_bh (used to raise bottom halves)
  broadcast a condition
<braunr> and since netif_rx is called from within xmit, which is called
  from the net_bh worker thread
<braunr> the thread itself is never waiting for the condition when it is
<braunr> it's very simple to fix, i'll send a patch later
<braunr> netcat to netcat now consumes 100% cpu
<braunr> as does fakeroot ls -Rl
<braunr> but for some reason fakeroot chown is still extremely slow
<braunr> and i've seen deadlocks in glibc (e.g. setlocale() getting the
  locale lock, which is locked again in case libfakeroot fails and calls
<braunr> so still a bit of debugging work needed

IRC, freenode, #hurd, 2013-11-06

<braunr> chown being slow with fakeroot-tcp can also be seen on linux

<teythoon> did your recent patch improve the performance of fakeroot-tcp ?
<braunr> yes
<teythoon> very nice :)
<braunr> but fakeroot chown is still slow
<braunr> although it's also slow on linux
<braunr> so i'm not looking into that any more for the time being
<braunr> as long as it's not used recursively on huge directories, it's

IRC, freenode, #hurd, 2013-11-09

<teythoon> braunr: fakeroot-tcp is indeed much faster now :)