UbuntuIRC / 2008 /04 /11 /#ubuntu-kernel.txt
niansa
Initial commit
4aa5fce
=== inuka_desk_ is now known as inuka_desk
=== jayc__ is now known as jayc_
[05:05] <osmosis> i dont understand what this means, http://www.news.com/8301-13580_3-9867657-39.html
[05:17] <osmosis> zul: ping
=== jayc__ is now known as jayc_
[05:56] <tjaalton> hmh, resume from hibernate failed, ie. it booted normally and not from the file
=== asac_ is now known as asac
=== macd_ is now known as macd
=== Whoopie_ is now known as Whoopie
[10:45] <cking> xivulon: Hi, I've been looking at your notes for the Wubi issues
[10:46] <xivulon> just added a new one!
[10:46] <xivulon> was waiting for you in fact...
[10:47] <xivulon> let me know if I need to clarify anything
[10:48] <cking> xivulon: I managed to screw up my /home today playing around with this - I was offline for a while sorting my machine out :-(
[10:49] <xivulon> sorry to hear that :(
[10:49] <cking> xivulon: It's all part of the fun :-)
[10:49] <xivulon> You are invited to explain the fun part to my wife...
[10:50] <cking> xivulon: I looked at the notes - the problem is umounting / and then /home - can one actually do this in practice?
[10:50] <cking> xivulon: I was thinking of using pivot_root but I haven't much experience with this, so I am not sure if this a possible way of doing things
[10:50] <xivulon> unmounting is not a problem except that it will be the last thing you can do (without a ramdisk)
[10:51] <xivulon> see my last comment on that
[10:51] <cking> ..but with a ramdisk, one could umount /?
[10:51] <xivulon> I would assume so
[10:51] <xivulon> but I have never tried that
[10:52] <cking> ..me neither. I think we may need some expertise from someone like Colin Watson.
[10:53] <xivulon> eheh...
[10:53] <cking> But I think it will lead to a sane shutdown that will ensure / and /home are flushed and umounted properly
[10:53] <cking> ..otherwise one will always see this kind of corruption of /
[10:54] <cking> if one can do this then the vm dirty hacks are no longer required
[10:54] <xivulon> I'd think that if we could remount /host ro (safely) it would be good enough
[10:55] <xivulon> see my umountroot attachment for that, unfortunately ntfs-3g complains about remounting...
[10:55] <xivulon> cking the vm dirty hacks are still required in case someone hard-reboot
[10:56] <cking> xivulon: yep - true about the hard-reboot - I overlooked that detail.
[10:56] <xivulon> In fact based on user feedback all the problems of fs corruption were due to hard reboots
[10:57] <xivulon> I haven't seen any complaint about fs corruption because of normal reboot
[10:57] <xivulon> I'd guess that remounting root ro + the sysctl hacks is enough in most cases (and maybe some fsck help...)
[10:57] <cking> xivulon: I managed it :-) I did started removing a huge tree of source and then rebooted - and I got a bit of a mess
[10:58] <xivulon> that is with the sysctl hacks?
[10:59] <xivulon> cking was disconnected
[11:00] <cking> xivulon: what happened there?
[11:00] <cjwatson> there was a netsplit, too
[11:00] <cking> ..a netsplit?
[11:00] <cjwatson> IRC networks consist of many servers, in general, which are connected in a graph; if a pair of servers lose their connection, that's a netsplit
[11:00] <cking> ah.. I am enlightened
[11:00] <cjwatson> which results in the users on the network being effectively partitioned in terms of their ability to talk to each other
[11:01] <xivulon> cking missed replies after my last comment
[11:01] <xivulon> didn't miss cjwatson explanation though :)
[11:01] <cking> xivulon: yep, I kind of missed the last few lines
[11:02] <cjwatson> 10:56 <cjwatson> cking: I think you misunderstand the order of events slightly, though
[11:02] <cjwatson> 10:56 <cjwatson> cking: the NTFS filesystem in question is actually /host, not /home
[11:02] <cjwatson> 10:56 <cjwatson> cking: and Wubi is a little odd here - the way it works is that / is actually a loop-mounted filesystem *on top of* an NTFS filesystem that is then moved to /host
[11:02] <cjwatson> 10:57 <cjwatson> cking: so the filesystems are the other way round from what you'd expect
[11:02] <cjwatson> in case that was missed
[11:02] <cking> thanks
[11:03] <cking> .. I keep on writing /home when I mean /host - it's a constant brain type of mine
[11:03] <cking> ^type^typo. I need coffee.
[11:03] <xivulon> the crux though is that /host cannot be remounted r/o IMHO
[11:03] <cjwatson> I suspect that you cannot actually unmount / and then expect to be able to unmount /host
[11:03] <cjwatson> because you won't have a reference to /host any more
[11:03] <xivulon> cking did you experience fs corruption when rebooting (rc6.d) and with the sysctl hacks in place?
[11:04] <cjwatson> you'd have to move all the mount points around, and frankly, I do not think that is a viable approach at this point
[11:04] <xivulon> was /host to be corrupted or / and did corruption involve the journal?
[11:04] <cking> xivulon: yep
[11:04] <cjwatson> we need something pretty simple and foolproof, and playing Tetris with your mount points isn't that :)
[11:05] <cking> cjwatson: I'm not sure if there is a solution that is simple and foolproof.
[11:05] <xivulon> not sure if we can try fixing /host (ntfs-3g) remount
[11:06] <cking> xivulon: one could do that - but wasn't there a big risk that NTFS could get corrupted?
[11:06] <xivulon> that was my understanding from last szaka comment
[11:07] <xivulon> did ask him some more info by mail but he didn't reply
[11:07] <cjwatson> if the NTFS filesystem is never remounted read-only or unmounted (as is the case at the moment, I understand), why does that affect the kernel's ability to writeback changes to the ext3 /?
[11:07] <cjwatson> or rather, what is breaking the kernel's ability to writeback changes?
[11:08] <cking> cjwatson: from the way I see it, we can either force disk writebacks using the vm flush hacks - but this does not necessarily guarantee a fulling sync'd fs.. or..
[11:08] <cjwatson> IIRC the kernel always syncs just before reboot
[11:09] <cking> .. we can fool around with the umounting. The former may help when systems power down in an uncontrolled way, the latter makes sure the filesystems are coherently umounted at shutdown
[11:10] <cking> cjwatson: true.
[11:10] <xivulon> ...and yet you ended up with a corrupted fs
[11:10] <cking> ..my concern is that the /host is not being properly written back to.
[11:12] <cking> I believe /host is not being umounted which worries me.
[11:12] <xivulon> it certainly is not
[11:12] <xivulon> or you would lose /
[11:12] <xivulon> and hence /sbin/reboot
[11:13] <cjwatson> remounting it read-only ought to be as good as unmounting from the perspective of avoiding corruption, but that's been problematic due to ntfs-3g bugs
[11:14] <xivulon> cking, on a side note, you also suggested using sync mount option on the loopfile, I assume you meant the /host fs
[11:16] <cking> xivulon: yes, if it's possible on ntfs. The sync is not required on the loop file since it's mounted ro at the shutdown. However, the sync,dirsync and commit=1 could help on power outages I suppose.
[11:17] <xivulon> but isn't host that has to sync automatically? the fact that the loopfile syncs to host does not help if host does not sync to disc, correct?
[11:18] <cking> xivulon: true - yep, it's clear I haven't thought that one through.
[11:20] <cking> xivulon: I suppose the outstanding questions are: can ntfs-3g be fixed in time to enable a ro remount, and if so, can one avoid NTFS corruption?
[11:21] <cking> xivulon: ..also, do the vm dirty tweaks do anything useful in reality?
[11:21] <xivulon> yes, that would be the best avenue, I had that bug open for quite some time...
[11:21] <xivulon> but did not bother too much because had zero bug reports about fs corruption... so far...
[11:22] <cking> xivulon: ..thirdly.. does the ntfs-3g support sync and commit=1 or are these only applicable to ext3?
[11:22] <xivulon> cking I know that in wubi 7.04 we did not have them we had more reports about people having fs corruption after hard reboot
[11:22] <xivulon> that is based on a completely subjective measure
[11:22] <cking> xivulon: ..yes but it's comforting to hear this.
[11:22] <xivulon> might be due to better user education also (added a few faqs)
[11:23] <xivulon> I would not know how to test that in an objective way
[11:23] <xivulon> for ntfs-3g we would need to ask szaka
[11:23] <xivulon> I know nothing about its internals
[11:25] <xivulon> but if you see #186117 that didn't go too far
[11:25] <cking> xivulon: that's the best path to take if he is available.
[11:26] <cking> xivulon: Meanwhile I will google around to see if there are any other ways to make sure the filesystem is flushed other than using the vm dirty hacks,
[11:27] <cking> ..one never knows if there is something better.
[11:30] <xivulon> cjwatson as plan b couldn't we still try to go for /sbin/reboot on ramdisk (but leaving current behaviour whenver that is not strcilt needed)?
[11:35] <cjwatson> it's very risky, you need to copy over any libraries it needs as well; it needs /proc and /dev; and who knows what else you might have to play catch-up with
[11:35] <cjwatson> I have a suspicion changing ntfs-3g would be safer
[11:35] <xivulon> cking still other route might be to ensure that ntfs-3g does actually empy his cache whenever the sync command is issued. I think.
[11:36] <xivulon> if / is ro and synced properly, rebooting should not be too drastic even if /host is rw
[11:36] <cking> xivulon: an extra sync can only help
[11:37] <cking> xivulon: I recall that twiddle with (one of) the vm dirty options may also forces a sync.
[11:37] <xivulon> cking you might want to try to reproduce fs corruption with new initramfs + sysctl + umountroot, as per thread
[11:38] <xivulon> as per bug attachments...
[11:38] <cking> yep. Will do.
[11:38] * cking need to reboot to try this.. rebooting ..
[11:38] <xivulon> FS corruption of the host is probably far more important than the guest corruption. And for the latter I'd focus on journal loss.
[11:41] <cking> xivulon: NTFS + ntfs-3g's behaviour is not really well know in this scenario. That's my concern. I'd hate too see Wubi users hacked off because their NTFS is screwed up
[11:42] <cking> ..so sync sync sync and then a fixed ntfs-3g mount -o remount may be the best way forward.
[11:43] <cking> ..one can never sync enough.
[11:43] <cking> :-)
[11:43] <xivulon> only confort I can give you is that with almost 1 million downloads all the ntfs corruption claims were due to hard reboots. And I have no evidence that ntfs-3g makes things any worse
[11:43] <cking> 1 million(!!)
[11:43] <xivulon> since you would still risk fs corruption when hard rebooting
[11:46] <xivulon> I discussed that with szaka more than once, and he always asserted that the chances of fs corruption are not made any worse by the use of ntfs-3g
[11:47] <xivulon> yep we are close to a million
[11:48] <xivulon> but of course now that wubi is official, host corruption would make very bad press...
[11:55] <cking> cking: rebooting ..
=== Lure_ is now known as Lure
[12:36] <xivulon> lamont: I did do some basic tests yesterday for #206113 and commented on the bug let me know if you need anything else
[13:30] <lamont> my most recent dist-upgrade (-15.27 et al) killed sound...
[13:35] <amitk> lamont: does it also lead to a long boot time?
[13:35] <lamont> dunno
[13:35] <lamont> I wasn't watching...
[13:35] <lamont> and actually, I think it's related to session management, not to the kernel upgrade...
[13:36] <amitk> lamont: hmm.. how so?
[13:36] <lamont> I think I saw this before, now I just need to remember how I got told to fix it...
[13:40] <lamont> [ 54.214217] iTCO_wdt: Found a ICH5 or ICH5R TCO device (Version=1, TCOBASE=0xf860)
[13:41] <lamont> not seeing where I figured it out before, istr it was simply muted hard somewhere
[13:41] * lamont afk for about 15 min
[14:12] <lamont> amitk: so I take back what I said... let's blame the kernel for now... debugging ideas?
[14:17] <amitk> lamont: sound-applet reports no mixers?
[14:18] <xivulon> lamont, in case you missed it, did the tests you asked, see comment in 206113
[14:18] <lamont> xivulon: yeah - thanks
[14:18] <xivulon> let me know if you need anything else
[14:18] <lamont> sure
[14:19] <lamont> amitk: bringing the system fully current to now and rebooting (new lrm), and then I think I'll need something more basic in the way of instructions...
[14:19] * lamont iz sound neub
[14:28] <lamont> amitk: anything I should tweak before I reboot
[14:28] <lamont> ?
[14:28] <lamont> Open Volume Control --> No volume control GStreamer plugins and/or devices found
[14:31] <amitk> lamont: just note if you reboot takes longer than usual
[14:32] <pwnguin> is -16 known to not boot? i cant find a report on it
[14:33] <rtg> pwnguin: on what? -16 is still not widely distributed as meta was just uploaded a few minutes ago.
[14:33] <pwnguin> ah
[14:34] <pwnguin> on my laptop, toshiba tecra m7. -generic
[14:34] <rtg> pwnguin: is this a regression?
[14:34] <pwnguin> yea
[14:34] <pwnguin> on the other hand, i dont have -meta
[14:34] <rtg> from -15 to -16?
[14:34] <pwnguin> yep
[14:34] <pwnguin> im on -15 right now
[14:34] <rtg> pwnguin: meta won't affect the boot.
[14:35] <pwnguin> rtg: i'd feel better about not missing part of the package with meta though
[14:35] <rtg> pwnguin: at what point in the boot process does it stop?
[14:35] <pwnguin> after turning off the graphical boot, hardware detection is the last message
[14:36] <pwnguin> i booted into recovery mode
[14:36] <pwnguin> and
[14:36] <pwnguin> the last two modules giving messages were iwl and cs
[14:36] <amitk> pwnguin, wait for a few minutes
[14:36] <amitk> it should continue booting
[14:36] <lamont> unsurprisingly, that didn't fix either of my issues
[14:36] <amitk> rtg: I see this too, I was woring on Classmate when I hit this
[14:37] <pwnguin> a few minutes?
[14:37] <pwnguin> thats a hell of a timeout ;)
[14:37] <amitk> pwnguin: I am trying to determine if you are seeing the same problem as me ;)
[14:37] <pwnguin> well
[14:37] <pwnguin> lemme reboot it i guess
[14:37] <rtg> amitk: everything since -15 has been custom binary, except for the mmc timeout.
[14:37] <pwnguin> heh
[14:38] <pwnguin> the mmc timeout
[14:38] <pwnguin> interesting; i upgraded because i wanted to test that
[14:38] <amitk> rtg: yeah.. I saw the changelog, and even the mmc timeout was 10ms, not 2 minutes
[14:38] <rtg> better go through the git log and make sure.
[14:38] <smb_tp> It is if it was 2ms before and not 2min
[14:39] <rtg> brb, rebooting after upgrade.
[14:39] <smb_tp> It was just a one liner changing mmc_delay from 2 to 10
[14:39] <pwnguin> its supposed to be ms
[14:39] <pwnguin> it could be seconds
[14:39] <pwnguin> i never found that particular function
[14:40] <smb_tp> It is an inline function in one of the headers
[14:40] <amitk> core.h:static inline void mmc_delay(unsigned int ms)
[14:41] <amitk> pwnguin: can you confirm that the boot contineus after a few minutes? albeit w/o sound
[14:41] <pwnguin> on a related note, the sdhci developer's asked me to retry with MMC_DEBUG
[14:41] <pwnguin> amitk: not yet
[14:41] <pwnguin> give it a minute or so
[14:42] <amitk> pwnguin: just wait a few more please, that'll tell me that it is reproducible behaviour
[14:43] <amitk> pwnguin: What does Alt+F1 show?
[14:43] <pwnguin> doh
[14:43] <pwnguin> nothing much
[14:43] <amitk> F2?
[14:43] <amitk> I forget which one it is
[14:43] <pwnguin> kinit: no resume, doing normal boot
[14:43] <lamont> 8
[14:43] <pwnguin> loading manual drivers
[14:43] <pwnguin> i forgot to set verbose
[14:44] <amitk> pwnguin: care to disable "quiet splash" in the kernel cmdline on the next boot?
[14:44] <pwnguin> indeed
[14:44] <amitk> thanks
[14:45] <amitk> hmm. now to see what changed in lum
[14:45] <lamont> amitk: my boot seems to proceed at normal-ish pace
[14:46] <pwnguin> lets see. theres hda_codec at 30 saying it cant find my audio. iwl at 31 saying there's 11 tunable channels for bg, 13 for a, and several I/O port probes for module cs
[14:46] <amitk> rtg: 607ab6f78fa5d51b4dab72d218455a858c499c6f in lum
[14:47] <pwnguin> if you want logs, im afraid i dont have a serial port =(
[14:47] <amitk> rtg: smb_tp: that commit looks like it will affect every sound driver
[14:47] <rtg> amitk: is that the culprit?
[14:48] <pwnguin> huzah
[14:48] <pwnguin> fail
[14:48] <amitk> rtg: I am going to revert and try on the classmate
[14:49] <rtg> amitk: just rename snd_core to something, then try.
[14:49] <pwnguin> amitk: after 2 minutes of no activity, it wrote some new text
[14:50] <pwnguin> it seems to still be stuck though
[14:50] <amitk> pwnguin: I'll bet it'll boot eventually :)
[14:52] <smb_tp> Hm, might be bad if register all will eventually call register...
[14:52] <pwnguin> heh
[14:52] <pwnguin> well, it doesn't seem to be spinning
[14:52] <amitk> rtg: crap.. just realised soundcore.ko comes from kernel, not ubuntu directory in /lib/modules :)
[14:53] <rtg> amitk: I was just looking to see where that file gets compiled.
[14:54] <amitk> rtg: that seems to improve things a lot
[14:54] <amitk> I'm going to revert the commit if you have no objections
[14:55] <rtg> amitk: can tyou rebuild lum with that patch reverted?
[14:55] <amitk> rtg: yeah.. doing that now
[14:55] <xivulon> cking in your tests did you try mounting /host with sync & co?
[14:55] <rtg> amitk: removing soundcore indicates it is the sound sub-system, but isn't proof positive.
[14:56] <xivulon> would it help to have 1 sec sleep before rebooting?
[14:58] <cking> xivulon: I couldn't see to create any filesystem corruptions whatsoever.
[14:58] <rtg> amitk: I started a new release in LUM
[14:58] <cking> xivulon: what is interesting is looking at /proc/vmstat | grep dirty
[14:59] <amitk> rtg: I've been seeing this since morning, but thought it was due to my usb patch for classmate suspend
[14:59] <cking> ..even with the agressive vm dirty tweaks the data is not getting written back that quickly
[15:00] <cking> ..one can see the writes by doing a bit of an ugly hack: sudo echo 1 > /proc/sys/vm/block_dump and looking at dmesg
[15:00] <amitk> pwnguin: you are running -15?
[15:00] <pwnguin> no
[15:01] <pwnguin> -16
[15:01] <amitk> you compiled yourself?
[15:01] <pwnguin> nope =(
[15:01] <cking> xivulon: ..even with the most aggressive vm dirty settings one can observe writes being flushed a few seconds after the initial write...
[15:02] <pwnguin> i noticed -16 in the repos
[15:02] <amitk> rtg: ^^
[15:02] <cking> xivulon: ..which could explain why some users are seeing corruptions even when we are forcing writebacks to occur as soon as possible.
[15:02] <rtg> amitk: I saw that.
[15:02] <cking> xivulon: ..because ASAP is not really that ASAP :-(
[15:02] <xivulon> hmm
[15:02] <rtg> pwnguin: do you also have -16 LUM installed?
[15:03] <xivulon> is that an ntfs-3g / fuse implementation issue?
[15:03] <amitk> I only see -16 LRM, no LUM or kernel
[15:03] <xivulon> cking what is your suggestion based on that?
[15:03] <pwnguin> well my mirror wins i guess
[15:03] <amitk> ohh and I see -16 linux-libc-dev
[15:03] <cking> xivulon: no. I initially thought it was all the layering in the ntfs-3g, fuse, loopback etc.. but one can observe this on a normally mounted filesystem.
[15:04] <smb_tp> amitk: Got that installed already
[15:04] <xivulon> when you say "couldn't see to create any filesystem corruptions whatsoever" do you mean you couldn't create corruptions or that you were simply not testing that
[15:04] <xivulon> so it's a kernel issue...
[15:04] <cking> xivulon: so it appears what the kernel documentation says and what is actually happening differ.
[15:05] <xivulon> not to pleased to hear that
[15:05] <cking> xivulon: as for the corruptions: I hammered the system with loads of creat's and unlinks and forced a reboot and everything was OK...
[15:05] <amitk> smb_tp: no sideeffects?
[15:06] <cking> xivulon: and also I repeated this several times and pulled the plug and got the normal fsck'ing fixes but I did not get any major failures
[15:06] <xivulon> cking that is in line with lack of reports on the matter. was this with the new initramfs / umountroot or in a plain vanilla case?
[15:06] <cking> xivulon: it was using the new scripts from the bug report
[15:07] <smb_tp> amitk: None I noted, but I guess this only affects programs comiled on that host. and I didn't compile any user-space
[15:07] <xivulon> well that is good news
[15:07] <xivulon> for once...
[15:07] <amitk> rebooting now....
[15:07] <cking> xivulon: I left some notes on the bug report so one can see what I mean about the write backs being a little bit too lazy for my liking
[15:08] <xivulon> could it be that the mount options (sync & co) interfere with sysctl?
[15:09] <cking> xivulon: no. when I saw the less than perfect performance I tried it on a vanilla configuration to do a sanity check - i.e. no insane sync,dirsync,commit=1 flags.
[15:09] <xivulon> cking: would you suggest we keep mount options and/or sysctl settings?
[15:10] <cking> ..and I saw the same "problem". We are probably seeing the pdflush pushing out all it can before it gets rescheduled or something. I need to shove a whole load of debug in to see why it is a little less aggressive than expecyed
[15:11] <xivulon> so I guess you'll need some more time for the tests then we can reassess
[15:11] <xivulon> I would not think though that in the light of what you say the new initrd/umountroot should make much difference
[15:11] <cking> xivulon: well, closer inspection shows that the mount options are ignored by ntfs - so we just fall back to the sysctl's.
[15:11] <xivulon> since /host is still not remounted ro (at least in my tests)
[15:12] <cking> xivulon: agreed
[15:13] <cking> xivulon: I basically need to figure out why pdflush is not being totally true to the vm dirty knobs
[15:13] <amitk> rtg: revert works and sound is back
[15:13] <xivulon> sure, keep in mind that we'll need to do a feature freeze exception for each change, if something is not strictly necesseray (proven to be so) we should skip it
[15:13] <amitk> rtg: I've pushed the revert to lum
[15:13] <xivulon> that applies to initrd, sysctl and umountroot
[15:15] <cking> xivulon: yep. Well I think the real issue is that we want to flush blocks out asap and I need to figure out why the pdflush does not quite match the behaviour as described on the lid
[15:15] <cking> ^on the lid^in the docs^
[15:15] <cking> ..and then I'm a little worried about tinkering with the pdflush daemon this late in a code freeze :-(
[15:16] <cking> ..but it could be something obvious once I've got my head around it.
[15:17] <cking> xivulon: any ideas on anything else we could do?
[15:17] <xivulon> not really
[15:21] <cking> xivulon: OK I investigate pdflush and the vm dirty knobs
[15:22] <cking> s/I/I will/
[15:36] <xivulon> cking, maybe doublecheck whether initramfs/umountroot make any difference
[15:38] <cking> xivulon: I think most productive thing is to see if we can get pdflush to really do aggressive writebacks to limit loopbacke'd fs corruption
[15:39] <cking> xivulon: I hope it is something obvious
[15:40] <xivulon> agree
[15:41] <xivulon> by the way, wouldn't something like that be relevant also for vm?
[15:43] <xivulon> that too is a nested fs and hence subject to the same vulnerabilities
[15:43] <xivulon> I am surprised I am the only one running into this
[15:51] <xivulon> at least for the power-loss part (normal reboot is very different of course)
[16:20] <rtg> amitk: I verified the cx88 revert. it definitely makes a difference.
[16:21] <amitk> rtg: half the patch was mutexes, it had to ;)
[16:21] <rtg> amitk: I looked through it when I applied it. everything looked balanced, _and_ it came from upstream.
[16:22] <amitk> rtg: I guess it needs another part from upstream
[16:23] <rtg> amitk: could be. now I'm f\gonna have to figure out what the real cx88 crash was .
=== jayc__ is now known as jayc_
[17:33] <nxvl> i have just updated to -16 kernel and it hangs
[17:34] <rtg> nxvl: LUM is currently percolating with a fix. wait for linux-ubuntu-modules-2.6.24 16.22
[17:34] <nxvl> ok
[17:34] <nxvl> so it was a know issue?
[17:34] <pwnguin> heh
[17:34] <pwnguin> only recently
[17:35] <rtg> nxvl: it became a known issue fairly quickly.
[17:35] <nxvl> yes of course it has no more than one day uploaded
[17:35] <nxvl> but if you already have what you need to fix it
[17:35] <nxvl> i'm ok
[17:35] <pwnguin> 16.22?
[17:35] <rtg> https://launchpad.net/ubuntu/+source/linux-ubuntu-modules-2.6.24/2.6.24-16.22
[17:36] <pwnguin> i have linux-headers at 16.30
[17:36] <pwnguin> is that . not kept in sync?
[17:37] <rtg> pwnguin: The LUM package version is not related to the headers package version (except for the ABI number)
[17:38] <pwnguin> ok
[17:39] <nxvl> so, fix is released and we only need to wait until it reach the mirrors?
[17:39] <rtg> yep
[17:39] <pwnguin> not booting is a fairly high priority bug ;)
[17:40] <pwnguin> espcially when someone like amitk and reproduce it
[17:41] <nxvl> ok thanks!
[17:58] <nxvl> btw
[17:58] <nxvl> i forgot to report something else
[17:58] <nxvl> i hav a lenovo T61
[17:58] <nxvl> with hotplug DVD-rom
[17:59] <nxvl> but when iunplug it the system hangs
[18:40] <mario_limonciell> rtg, i was thinking this morning, could you see to it that LBM is shipped as available on the DVD image to install?
[18:40] <mario_limonciell> in it's pool directory or similar
[18:41] <rtg> mario_limonciell: thats a question for cjwatson
[18:42] <mario_limonciell> okay i'll ask him
[18:52] <crimsun> rtg: are there are traces from the bootup problems (WRT #212960 in l-u-m -16)?
[18:52] <crimsun> amitk: ^
[18:52] <crimsun> are there, sheesh. can't type.
[18:53] <rtg> crimsun: there don't seem to be. I'm looking at it now. Frank's original patch works, but the one Leann attached does not (it locks up)
[18:53] <rtg> crimsun: I'm trying to decide if it is really the root of the problem.
[18:54] <crimsun> rtg: I don't think it is, after closer inspection.
[18:54] <rtg> crimsun: can there really be that much list activity going on? Or is the use of mutexes just jostling the thread timing.
[18:55] <rtg> crimsun: as far as I can remember, PCI device discovery is single threaded, right?
=== inuka_desk_ is now known as inuka_desk
[19:09] <crimsun> rtg: there shouldn't be much activity, and it is single-threaded IIRC.
[19:15] <rtg> crimsun: that's what I thought. I'm not sure Frank's analysis is correct (about the lists getting corrupted).
[19:17] <pepie34> I can't catch an AP with my AR5008 madwifi card since the two last -14 and -15 kernel
[19:17] <pepie34> still no problem with the -12 kernel and the same userland as the others new
[19:22] <mario_limonciell> hm what is creating the symlink between cpufreq directories on two core systems initially?
[19:22] <mario_limonciell> i had thought it'd be powernowd doing it, but I don't see evidence of that
[19:23] <pepie34> any idea ? for the regression on madwifi
[19:23] <smb_tp> rtg: Hm, I might be stupid but in the report snd_card_register_all is called and later snd_device_new. If thats both in core, how should that work?
[19:24] <pepie34> I should be more precise as it is a AR5XX8 chipset i had to use the madwifi svn and not the module in the restricted package
[19:24] <pepie34> it works fine on feisty/gutsy and hardy since the 2.6.24-12 kernel
[19:29] <rtg> smb_tp: it clearly must work because its not a problem for everyone. I'm still looking at the cx88 driver. However, right now its lunch time.
[19:30] <rtg> pepie34: you are on your own 'cause I haven't pulled the madwifi branch that supports the AR5008. I've been waiting until they make it official.
[19:30] <smb_tp> rtg: Enjoy. I was just thinking of the same codepath with mutexes
[19:38] <pepie34> rtg i just want to warn then that since the -14 the madwifi-svn does not work and you will get all the mac user really angry
[19:50] <lamont> amitk: so will this new kernel make my soundz werk>?
[20:24] * pwnguin does a zomg sd works again dance
[20:24] <pwnguin> rtg: thanks
[20:49] <infinity> lamont: You don't need sound.
[20:49] <lamont> infinity: do too
[20:50] <infinity> lamont: The kernel team disagrees.
[20:54] <rtg> lamont: we've gone out of our way to wreck only _your_ audio support.
[20:55] <lamont> rtg: I feel so much more productive without the noise-canceling music
[20:56] <rtg> lamont: I thought you had audio once upon a time?
[21:23] <sistpoty> hi, can someone please look at the FFe in bug #185634 and give a comment? thanks!
[21:23] <ubotu> Launchpad bug 185634 in ubuntu-meta "uvcvideo: iSight firmware loading does not work" [Medium,Confirmed] https://launchpad.net/bugs/185634
[23:56] <tjaalton> is there something in -16 which broke nvidia from loading? bug 215778
[23:56] <ubotu> Launchpad bug 215778 in linux-restricted-modules-2.6.24 "2.6.24-16.30 kernel update - nvidia: module license 'NVIDIA' taints kernel" [Undecided,Confirmed] https://launchpad.net/bugs/215778
[23:57] <tjaalton> lrm debdiff looked good, so it shouldn't be due to the dh_strip fix