**** BEGIN LOGGING AT Sun Feb 13 02:59:56 2011 Feb 13 03:11:17 build #65 of xburst is complete: Failure [failed compile_6] Build details are at http://tksite.gotdns.org:8010/builders/xburst/builds/65 Feb 13 03:27:26 nbd * r25511 /trunk/ (package/base-files/Makefile toolchain/uClibc/Makefile): remove the shared object stubs of libnsl and libresolv - let the compiler use the static variants instead Feb 13 03:27:30 nbd * r25512 /trunk/package/busybox/config/procps/Config.in: busybox: disable killall5 - nothing uses it Feb 13 03:54:52 nbd: any thoughts on the libnl version bump? Feb 13 03:55:14 have that in my tree already Feb 13 03:55:17 will be committed in a minute Feb 13 03:57:20 nbd: can you fix libubox, relayd and ubus to use PKG_REV to hold their git hashes, like the majority of the other packages? Feb 13 03:57:27 nbd * r25513 /trunk/package/libnl/ (7 files in 2 dirs): libnl: update to version 2.0 (patch by Philip Prindeville) Feb 13 03:58:24 swalker: maybe tomorrow - what's the reason for that though? Feb 13 03:59:41 nbd: consistency mostly and so I don't have to keep resolving the conflicts for my upstream checker Feb 13 03:59:47 ok Feb 13 04:00:26 tomorrow i want to see if i can find some more major bloat sources Feb 13 04:00:44 i think today i freed up at least 100k uncompressed Feb 13 04:07:03 nbd: add librpc to that PKG_REV list as well Feb 13 04:11:00 nbd: thanks. Feb 13 04:15:02 nbd: #8847 (Consider disabling debug output by default in hostapd package in order to decrease application size)? Feb 13 04:24:20 swalker: i frequently request debug logs from people, so i want to keep this in by default **** BEGIN LOGGING AT Sun Feb 13 15:40:14 2011 Feb 13 15:46:40 nbd * r25520 /packages/net/nfs-kernel-server/Makefile: nfs-kernel-server: disable tcp-wrappers for the host build Feb 13 15:49:38 ` Feb 13 15:54:35 build #65 of au1000 is complete: Failure [failed shell compile_3] Build details are at http://tksite.gotdns.org:8010/builders/au1000/builds/65 Feb 13 16:01:12 blogic * r25521 /trunk/target/linux/lantiq/patches/700-dwc_otg.patch: Feb 13 16:01:13 [lantiq] Feb 13 16:01:13 * fixes dwc_otg init when power gpio is set to -1 Feb 13 16:08:41 nbd * r25522 /trunk/target/linux/ar71xx/generic/profiles/00-default.mk: ar71xx: remove vsc73xx ucodes from the default profile - these are only used on some atheros reference designs which are unlikely to have many users Feb 13 16:10:45 build #64 of etrax is complete: Failure [failed compile_3] Build details are at http://tksite.gotdns.org:8010/builders/etrax/builds/64 Feb 13 16:23:18 nbd * r25523 /trunk/package/ppp/Makefile: ppp: move multilink support into a separate build variant - saves around 30k for the default build Feb 13 16:25:35 nbd * r25524 /trunk/package/ppp/Makefile: ppp: add a more descriptive title for ppp-multilink Feb 13 16:46:30 nbd * r25525 /trunk/package/libnl-tiny/src/ (4 files in 2 dirs): libnl-tiny: decrease the size a bit more by removing some unnecessary functions Feb 13 16:54:46 nbd * r25526 /trunk/package/dropbear/Makefile: dropbear: compile out libtomcrypt assertion messages, saves some more space Feb 13 16:58:41 build #64 of iop32x is complete: Failure [failed compile_3] Build details are at http://tksite.gotdns.org:8010/builders/iop32x/builds/64 Feb 13 17:05:37 nbd * r25527 /trunk/package/libnl-tiny/src/ (cache_mngt.c genl_mngt.c include/netlink-local.h): libnl-tiny: remove some more functions to reduce binary size Feb 13 17:25:12 build #86 of at91 is complete: Success [build successful] Build details are at http://tksite.gotdns.org:8010/builders/at91/builds/86 Feb 13 18:03:32 build #74 of gemini is complete: Failure [failed shell shell_4 compile_3] Build details are at http://tksite.gotdns.org:8010/builders/gemini/builds/74 Feb 13 19:05:15 nbd: https://dev.openwrt.org/browser/trunk/package/base-files/files/bin/ipcalc.sh#L3 still references common.awk, #8857 Feb 13 19:05:29 oh, forgot to delete a line Feb 13 19:05:30 thanks Feb 13 19:06:24 nbd * r25528 /trunk/package/base-files/files/bin/ipcalc.sh: base-files: remove a stale reference to /usr/lib/common.awk (thx, swalker) Feb 13 19:08:11 build #63 of adm5120 is complete: Failure [failed compile_3] Build details are at http://tksite.gotdns.org:8010/builders/adm5120/builds/63 Feb 13 19:09:54 nbd * r25529 /trunk/package/ (4 files in 4 dirs): set PKG_RELEASE to PKG_SOURCE_VERSION on a few packages using git Feb 13 19:15:29 * swalker goes to add some more non-consistent versioning hacks Feb 13 19:17:48 at some point it would be nice to move your script to trunk and maintain version check info in the package makefiles Feb 13 19:18:19 might make things easier Feb 13 19:23:46 it's rather ugly shell code, especially the section fixups Feb 13 19:59:11 nbd * r25530 /trunk/tools/squashfs4/patches/ (130-dynamic_lzma_params.patch 130-lzma_props.patch): squashfs4: make the lzma parameters configurable, and change the defaults for slightly improved compression on mips (mostly unchanged on other architectures) Feb 13 20:06:38 tcpdump is missing a librpc dependency it seems Feb 13 20:19:23 hm Feb 13 20:22:41 nbd * r25531 /packages/net/tcpdump/Makefile: tcpdump: do not use rpc support - fixes unnecessary dependency on librpc Feb 13 20:22:45 Wipster: fixed ;) Feb 13 20:23:06 :) thanks Feb 13 21:17:43 nbd: can you commit the vpi.vci patch for the solos driver? thanks. Feb 13 21:18:17 (it's a one liner... and only affects a pr_info()...) Feb 13 22:14:45 nbd: hello, do you see something wrong with this patch for ath9k? Feb 13 22:14:49 https://github.com/dtaht/Cruft/raw/master/bloat/558-ath9k_bufferbloat.patch Feb 13 22:15:29 yes, lots of it Feb 13 22:15:36 btw I tested bufferbloat countermeasures with very good results Feb 13 22:15:40 first of all, it would cripple aggregation Feb 13 22:15:54 then it messes with a hw specific value that has nothing to do with software buffering Feb 13 22:16:10 do you mean hw buffers are that large? Feb 13 22:16:22 no Feb 13 22:16:30 the fifo holds 8 entries on ar9300 Feb 13 22:16:56 seems librpc doesn't compile here: http://paste.openwrt.org/m163c9b43 Feb 13 22:17:21 why are you using uclibc 0.9.31? Feb 13 22:18:37 no specific reason, just to see if it works (which it did for months) Feb 13 22:18:55 i think we're going to remove 0.9.31 soon Feb 13 22:19:03 0.9.32 is better anyway Feb 13 22:19:36 so most things should work now with 0.9.32 ? Feb 13 22:20:04 nbd: i take the patch from here: http://www.bufferbloat.net/projects/bloat/wiki/Bloated_Driver_List Feb 13 22:20:19 the idea is to reduce ring buffers Feb 13 22:20:26 in the drivers Feb 13 22:20:30 the idea is good, but the implementation sucks Feb 13 22:20:43 i'm considering doing something similar to the patch that was posted for iwlagn on linux-wireless Feb 13 22:21:01 i.e. adaptively tune tx queue size based on latency Feb 13 22:21:57 cool! for the entire stack or just for ath9k? Feb 13 22:22:25 i don't know if this would work properly done in the stack Feb 13 22:22:37 maybe once i've moved a-mpdu to mac80211 Feb 13 22:23:48 ok, and when you say the patch cripples aggregation is that just for 802.11N or BG is affected too? Feb 13 22:24:03 11n Feb 13 22:25:21 ok, thank you, I'll stay away from that patch for now :) Feb 13 22:25:54 11n needs quite a bit of buffering to deliver decent performance Feb 13 22:26:33 i don't want to introduce artificial software queueing latency to improve aggregation length Feb 13 22:26:41 so any cutback on buffering needs to be adaptive Feb 13 22:28:10 yea, that's the problem with balancing buffers, too much end with horrible latency but performance is nice Feb 13 22:28:45 with 802.11n it's kind of different compared to wired connections Feb 13 22:29:00 because with 802.11n airtime utilization goes down with more buffering Feb 13 22:30:45 but the bufferbloat theory says that with more buffering it kills TCP congestion mechanisms Feb 13 22:30:54 so the entire flows end transmitting slowly Feb 13 22:31:00 which means more air time Feb 13 22:31:24 I don't think TCP affects air time Feb 13 22:31:51 solca: that theory is for mediums where the maximum capacity doesn't increase with more utilization Feb 13 22:31:55 air time is affected by the 802.11 transmission rate (and modulation) Feb 13 22:33:30 yea I see, it's more complex Feb 13 22:34:34 wireless is a shared medium. That means the constraints are quite different from P2P ethernet Feb 13 22:34:59 And the rate control is way more complex Feb 13 22:35:50 my problem with the bufferbloat crowd is that now that somebody pointed out a real issue, lots of people now try to do quick fixes to anything that looks related, often without careful consideration and tradeoffs Feb 13 22:36:45 = witch hunt :) Feb 13 22:37:20 nbd: yea, but it's something very complex to have an adaptive ring buffer? Feb 13 22:37:38 indeed Feb 13 22:37:46 its the only real solution imo Feb 13 22:37:53 but it's better to keep the status quo than to apply hackish patches like the one you linked to Feb 13 22:38:42 I must admit that reducing txqueuelen on my routers improve my network situation Feb 13 22:38:51 'for every problem there is a solution that is simple, elegant, and wrong' ;) Feb 13 22:38:53 but as you said I was looking for something related Feb 13 22:39:16 and ath9k was a very good target in all those APs ;) Feb 13 22:39:48 with ath9k i was actually considering adding more buffering under some very specific conditions Feb 13 22:40:09 because one of my changes a long time ago already cut the per-queue buffer size down to a quarter Feb 13 22:40:18 and in some situations this does slightly decrease performance Feb 13 22:41:21 wow 4x current size it's very large IMO Feb 13 22:41:36 if current is 512 per queue Feb 13 22:41:39 no Feb 13 22:41:51 it's 128 per queue Feb 13 22:42:00 how much queues? Feb 13 22:42:00 it was always 512 in total Feb 13 22:42:05 4 queues Feb 13 22:42:18 but before my changes, one queue could make use of all buffers and thus starve out the others Feb 13 22:42:35 ohh I see Feb 13 22:42:46 b43 has 128 frames per TX queue, too. And that is the lowest practical value, IMO. I carefully tuned it with actual measurements. Feb 13 22:43:01 that's why ath9k uses the mq qdisc? to handle 4 hw queues? Feb 13 22:43:11 A bigger value can improve performance in some situations Feb 13 22:43:25 especially on N, which we don't really have on b43, yet Feb 13 22:43:27 solca: any proper 802.11 driver needs to use mq with 4 queues Feb 13 22:44:02 because of WMM Feb 13 22:45:20 mb__: interesting, does the proprietary driver do the same thing (various queues)? Feb 13 22:45:38 nbd: so ath9k doesn't have 4 fifos? just 1? Feb 13 22:45:48 solca: I just went back and looked at unpatched 2.6.37. I stand corrected... off-the-shelf linux uses rt_tos2priority() which is seriously obsolete. Feb 13 22:46:15 we patch this function and sent the patch upstream some time ago to netdev. Feb 13 22:46:29 solca: huh? Feb 13 22:46:40 philipp64|laptop: pfifo_fast uses the 4 bit TOS field :) Feb 13 22:46:56 yeah, that's $%^& braindead. Feb 13 22:46:59 nbd: if ath9k have 4 phisical fifos? Feb 13 22:47:15 4 hw buffers? Feb 13 22:47:29 the hardware has 4 queues with separate DMA rings Feb 13 22:47:53 ohh I see, I though you were saying that you use the mq qdisc because of WMM Feb 13 22:48:05 not because of 4 phisical queues Feb 13 22:48:29 what's worse is that it's hard-coded into ip_forward.c and ip_sockglue.c... Feb 13 22:48:54 solca: ath9k does not have anything to do with the scheduler integration Feb 13 22:49:07 but it has to use multiqueue Feb 13 22:49:38 hrm, too bad i can't edit that bufferbloat wiki Feb 13 22:49:46 solca: we use Torsten's patch for CONFIG_IP_DIFFSERV_COMPLIANT. Feb 13 22:50:22 nbd: just for curiosity what would you put there? :) Feb 13 22:50:50 philipp64|laptop: both TOS and DiffServ sucks, SFB with ECN is the way to go IMHO Feb 13 22:51:23 DiffServ rocks if you're using a private network with MPLS. Feb 13 22:51:31 solca: that such quick hacks completely cripple aggregation and thus lead to more unnecessary airtime utilization Feb 13 22:52:07 solca: broadcom hardware has 6 queues, but we only utilize 4 (the required minimum) Feb 13 22:52:44 maybe i should join the bufferbloat channel Feb 13 22:52:55 philipp64|laptop: ohh I see, anyway SFB with ECN FTW! Feb 13 22:52:56 and yell at some people (in a nice way) Feb 13 22:52:56 ;) Feb 13 22:53:43 nbd: yea you should because many ppl will patch your driver with bad results :( Feb 13 22:54:13 mb__: and the propietary driver uses all 6 queues? Feb 13 22:54:22 I don't think so Feb 13 22:54:38 that's too much to fill... Feb 13 22:54:43 There's not much point in using more than 4 Feb 13 22:55:51 the nice thing about the mq qdisc is that theorically one could route packets for each queue as desired but I have never tested that Feb 13 22:59:30 having a discussion with the patch author now Feb 13 22:59:58 nbd: whereto? Feb 13 23:00:19 in #bufferbloat Feb 13 23:00:34 * solca is joining... Feb 13 23:55:02 Is it possible to define static serial ports in hotplug, or must udev be used? Feb 13 23:55:48 (multiple usb serial ports, want them in static order, regardless of order detected) Feb 13 23:59:43 mirko * r25532 /packages/Xorg/lib/qt4/patches/400-do-not-pack-structs.patch: do not pack structs - this causes segmentation faults on ARM architectures - especially when freetype is involved... Feb 14 00:18:49 I hope I didn't ask the wrong question ... Feb 14 00:49:50 KanjiMonster: which was? Feb 14 00:50:13 philipp64|laptop: regarding the sources of the gmac driver for bcm47xx Feb 14 00:59:43 ah. Feb 14 01:06:11 Hi, recently I meet this error when compile in my notebook: http://pastebin.com/8J8ti9p7 **** ENDING LOGGING AT Mon Feb 14 02:59:57 2011