interspersed within the article

thinking I might try out mesa-devel to see if the situation is any different as long as libva actually picks up on the drivers

suspecting this is some kind of regression since the hardware support should be here, Mesa just isn't picking up on it

might be worth trying out the amdgpu DDX driver instead of modesetting, since it does seem like VAAPI does have to interact with the X server

grasping at straws though

I don't want to have to test on Linux, but... I might have to

welp, mesa-devel's VAAPI driver reports VP9 support

Firefox picks it up as well

now running into another issue, youtube playback under Firefox drops tons of frames, like roughly half. immediately suspected picom, but disabling picom didn't change anything

guess I'll have to try the amdgpu DDX after all, see if the situation is any different there

exact same issue on the amdgpu DDX :/

seeing some reports of framerate issues on AMDGPU with kern.sched.steal_thresh at the default value of 2, apparently setting it to 1 fixes it for some people. I can give it a try, but are there going to be any side effects of messing with that setting?

guessing the steal_thresh parameter will affect the latency/throughput tradeoff regarding preemption?

kern.sched.steal_thresh seemingly has no impact, and I'm also finding some youtube videos that don't drop frames except for 1 or 2 at the very start of playback

also weird is that it reports dropping basically 50% of frames, but the video looks more or less like 60fps with a few hitches

i have 5 directories with rar files in them how do I extract them all once

uh, does anyone know why latest does not have packages for chromium or ungoogled-chromium?

i am looking at epair (man.freebsd.org/cgi/man.cgi?query=e….0-CURRENT&arch=default&format=html) and tap (man.freebsd.org/cgi/man.cgi?query=t….0-CURRENT&arch=default&format=html) and why would one use epair for jails and tap for bhyve.. is there any reason.. other then convention, that i may be missing?

Title: epair(4)

voy4g3r2: epair is designed for two instances of the kernel IP stack to communicate with each other (for example, the host, and a VNET jail). tap is designed for a userland process to communicate with the kernel IP stack, which is what bhyve does

a tap is an epair where one end of the connection is a userland process instead of a network interface

hwhich is why they mentione VMnet devices in tap not epair

i am trying to move the tap/epair stuff to the intel nics and i did not comment too well..

ah and i setup a jail with vnet.. hrm

lw: thank you, i was about to drop the epair stuff as i am like.. i can do this with tap.. so i thought

so when i have ifconfig_tap0="link0 up" the link0 is that userland / kernel bridge?

link0 is a flag that tells the kernel not to set the interface down when the process using it (bhyve) exits

('linkX' are a set of generic interface flags that are used for interface-specific functionality like this when there's no generic flag for it)

oh yeah, i have those notes.. we should NOT bring down these taps with the bhyve as we may want to 'reuse'

which would lead me to believe.. since i have 3 nic , i really should have a link0 link1 link2 each assocaited with a nic

because if i would "reuse" link0 between different bridges that would just be a mess

no... link0 doesn't mean "link number 0", it means "the interface-specific configuration flag number 0", which for tap, means "do not down the interface when the process exits". all your tap interfaces should have link0 set

any way I can install chromium from quarterly without completely switching to quarterly? seems like asking for dependency issues, tbh

so this mess: dpaste.org/Gar51

Title: dpaste/Gar51 (Plain Text)

i made sure to add some type of comment

i am not a fan of the one cloned_interfaces entry but it will work

well lets see what happens.. because re0 needs to go

voy4g3r2: having three interfaces on the same subnet is a bit unusual, you probably just want one bridge with all the tap/epair interfaces and igb0 in it?

and then don't use re0/igb1 at all

if i ever decide to execute the mikrotik stuff i want to use igb1 as an isolated network all together

i am decomissioning the re0 now :) just baby steps

1.) get the igb0,1 working 2.) get a bridge setup on igb0 3.) shutdown re0

well it looks like it worked ifconfig bridge0 down

and still have music going and bhyve is still up

bah.. bridge1 not getting an ip.. deep diving again

you can only have one bridge

you can make more, but you can't actually use them

I just encountered this today

that's not true, i have a system right here using two bridges just fine

huh, then I don't know why my second bridge didn't work

I don't have the url handy that said to not make multiple

there are quite a few ways to configure bridges wrongly which are not well documented, and also some related driver bugs like the ixgbe(4) vlan issue

(which don't prevent having multiple bridges, but can cause otherwise reasonable configurations to not work for no apparent reason)

i am pretty good at configuring things wrongly

but i figured the bridge on different nic would be okay as the if_bridge manual pages state, as long as you cloned the interfaces, i should be good

make me wonder how the heck my -CURRENT bhyve is working as the config points to bridge1 device

if anyone here has run into this issue with Firefox dropping frames in videos, despite no apparent hardware bottleneck, and has managed to find a way to work around it, I'd love to hear. it seems specific to Firefox, because I tried out Falkon, and despite that being restricted to software decoding, videos play back smoothly

on Firefox, WebRender is hardware accelerated and I've got VAAPI enabled to decode both VP9 and H.264 but at least VP9 drops about 50% of frames on most 60fps videos, the issue apparently doesn't show up on 30fps videos

voy4g3r2: For what it’s worth: gist.github.com/Crest/3c89deed4e9d8e9f5160ccdbee4c4745

Title: mk-epair · GitHub

so I haven't tested this solution extensively, so an obligatory "knock on wood", but setting Firefox's layout.frame_rate setting from -1 (unlimited?) to 60 (the monitor's refresh rate) seems to have almost completely fixed youtube's frame drops

down to about a 1% drop rate from nearly 50%

michaeldexter: is there an advantage to epair over tun (for bhyve)

rtprio: bhyve cannot use epair at all, as far as i know

jails only, eh

right. tap is for when the traffic originates from userland (as in bhyve), epair is for when the traffic originates from the kernel, which is mostly jails, although you could perhaps use it for other things

I hope the reason the gzdoom port is still on a release from 2021 is just because nobody is actively maintaining it, and not any technical reasons

i've wondered about, if you have a bunch of bhyve VMs with tap interfaces in a bridge, putting one end of an epair interface into the bridge as well to make the host's network config / firewalling easier... but i've never actually tried that. (i would be concerned about performance, but if_bridge performance is already quite bad)

gzdoom's port is on 4.7.1 whereas the latest stable release is 4.11.3 from last october

guess I will have to try compiling it to see

is it that bad?

regarding GZDoom? 4.7.1 is old enough to be lacking compatibility with newer map/mod releases, afaik

guess I need to hold cmake's hand here. the ninja build file it generated doesn't put /usr/local/include in the include path, evidently

I would think that'd be a given on FreeBSD

especially weird since it has found libraries I've got in /usr/local

tm512: regarding your question on figuring out why something doesn't have packages, try portsfallout.com you can reach the poudriere web interface of the builders from there too to find skipped ports

Title: Dashboard - FreeBSD pkg-fallout

thanks for the pointer

there doesn't appear to be any entry for chromium in the 140amd64-default build environment, but the package is just nowhere to be found in the latest repo still

lw: what do you mean if_bridge's performance is bad?

sfox: i don't have a link off hand, but received knowledge if that network performance via if_bridge is noticably lower than via a normal network interface

hm, beefy12.nyi.freebsd.org/build.html?…140amd64-default&build=cd2e1424dd41 has build/timeout for *chromium

trying to compile the latest GZDoom source from git. eventually, it gets to a file which includes sys/stat.h, which leads to:

/usr/include/sys/time.h:326:5: error: member access into incomplete type 'struct timeval'

reminding me of one of the annoyances that helped move me away from daily driving FreeBSD, and not really something that's FreeBSD's fault, just the fact that many developers don't really ever keep it in mind

nimaje: that gives "unable to connect" for me

regarding these compilation issues, I guess investigating further is gonna have to wait until tomorrow, but if anyone knows why sys/stat.h would include sys/time.h in a way that would ever break compilation, I'd like to know

errors like this make me think another header is not being included somewhere, or a necessary define is missing

no ipv6? they are mirrored for that reason pkg-status.freebsd.org/beefy12/buil…140amd64-default&build=cd2e1424dd41

Title: Poudriere bulk results

that link works

hm, but struct timeval gets unconditionally defined by sys/_timeval.h which unconditionally gets included by sys/time.h, so it seems strange that it would be incomplete

this is where sys/stat.h is being included: github.com/ZDoom/gzdoom/blob/master…lesystem/source/fs_findfile.cpp#L70

Title: gzdoom/src/common/filesystem/source/fs_findfile.cpp at master · ZDoom/gzdoom · GitHub

I don't think I saw any patches to this file in the gzdoom port, so either this file didn't exist in 4.7.1, or they've done something since then to break it

seems the file did not exist in 4.7.1, they were using something else for filesystem stuff

can you confirm, that on your system too sys/time.h unconditionally includes sys/_timeval.h and that unconditionally defines struct timeval ? (well, except for the include guards) I really can't see how struct timeval could be incomplete at that point

voy4g3r2: tap and epair interfaces are similar in that they connect two parts of software inside the same (freebsd) system

but epair connects two kernel network stacks (e.g. the host and a vnet enabled jail)

while a tap interface connects a userspace process (e.g. bhyve, openvpn) with the kernel network stack

so the tap interface is both a network interface and a character device

if you write to the tap device the kernel sees the data as ethernet frame

and if you send on the tap device the userspace sees it a data (to be interpreted as ethernet frame) on the device file

epairs a different from all other clonable network interfaces because they always exist pairs

if you ask the kernel to create epair0 it will (attempt to) create epair0a and epair0b as a pair

you can't have one without the other

if you destroy epair0b it also destroys epair0a (works the other way around too)

nimaje: yes, on line 37

regarding the _timeval.h include anyway

and that does indeed define struct timeval

the member access that triggers the "incomplete type" error is line 326 in sys/time.h

@sc@s

oops, sorry

nimaje: I may look at the preprocessor output when I get the opportunity to continue investigating this

crest: yes i did learn about that neat epair thing.. it makes an a and b

voy4g3r2: did my script work for you?

(the one michaeldexter linked to)

oh i did not see that .. scrolling up buffer

and if not is there something missing that would help?

voy4g3r2: gist.github.com/Crest/3c89deed4e9d8e9f5160ccdbee4c4745

Title: mk-epair · GitHub

it's a shell script to setup epairs

intersting i never "created" epairs just added them to /etc/rc.conf and they just appear...

but i do have to say my use of jail is limited.. i use bhyve more

but have thought about setting up dnsmasq to help with local dns.. as bind seemed to be too much

e.g. `mk-epair -n host-$name -N jail-$name -m bridge0 -J $name` could be used as exec.created inside a jail.conf

to let the kernel pick the next available epair unit number

rename both interfaces

add one end to a bridge on the host

and move the other to the vnet enabled jail

if any step fails the script cleans up (e.g. destroys the epair under its current name)

it looks quite clean and easy to follow

the script handles the tedious steps creating the pair, infering the names of both elements of the pair, renaming them as desired, setting MAC addresses, moving them to jails, adding them to bridges, adding them to groups

and can also run ifconfig inside the appropriate jail/host context

the latter feature depends on the ifconfig -j flag that only got added in freebsd 14

i don't know if it was backported to 13.3

yeah i am on 14

but good tidbit to know

its an annoying thing to automate well, because you have to recover to a sane operational state from partial errors

i do not fully understand why just doing this.. would NOT suffice dpaste.org/hkKK0

Title: dpaste/hkKK0 (Plain Text)

but i take it that requires a reboot and rebooting a big server is not the best

or just bringing down network services and restarting.. which would impact EVERYTHIGN on that host

it doesn't require a reboot if you know the right commands to the netif rc.d script

you can use something like service netif start epair10

to only start a single interface

but yes running service netif restart on a remote system is a bad idea

you have to restart at least the routing service too

because otherwise you're left without a default route

yeah, i have edone that a few times and hading to move the server to an area with a monitor

this time, i kept the re0 bridge around just in case something went boom

you can do it protected by screen/tmux/nohup if you like play fast and loose with a lab system

you config should work correctly

but it's all static and doesn't rename the interface to meaninful names

yeah.. it should but bridge1 is NOT getting an ip.. but

i like to embed the jail/guest name in interface names

oh agree

thankfully i had igb0 and re0 operational .. so i could get into the box

you can rename interfaces in rc.conf

you can automate it all with sysrc

but i prefer to keep my jail config in one place (the per jail jail.conf included from the global jail.conf)

and create the interface on demand

can you show the output of ifconfig bridge1?

the rc.conf snippet you shared assigns 192.168.0.30/24 to bridge1

why do you configure two bridges (bridge0 and bridge1) with the same IPv4 prefix 192.168.0./24

*192.168.0.0/24

(leaving out equal-cost multipath routing) there can be only one most specific route for each destination

and both 192.168.0.25/24 and 192.168.0.30/24 imply a network route for 192.168.0.0/24 directly attached to conflicting interfaces (bridge0 vs bridge1)

why did you create two bridges?

i would recommend that you create a lagg(4) interface containing igb0 and igb1 and add the lagg interface as first member interface to the remaining single bridge

if supported and enabled on your switch you can use LACP to share the load among all working members of the lagg

but each flow uses only the bandwidth of the link it's hashed to

so if you want to exceed the bandwidth of a single link you have to use multiple connections

i am not sure.. i am cat herding now.. and will need to think a little

did not want to give impression i just went away

btw why did you set the tap interfaces to "link0 up"?

wouldn't it be easier to use vmnet${unit} instead of tap${unit}?

vmnet is just tap with link0 as default, so there's no particular benefit from doing that

lw: less lines in rc.conf

and it's atomic on interface creation

crest: i rather be verbose, as i am learning

crest: is it? don't you have to set vmnetX up anyway?

you can't fail with a tap interface created, but the link0 flag not applied

lw: not for bhyve. it will bring the tap interface up when it starts

crest: the routing stuff, i did not think there would be an impact by using /24 on both.. my objective is to REMOVE the re0 bridge eventually

and you shouldn't have any IP addresses configured directly on bridge member interfaces

because having the addresses on the bridge members instead of only the bridge would cause really nasty IP link scope violations

which instantly break IPv6 and over time will break IPv4 too

could explain why i have "lags" with the freaking bhyve -CURRENT

i login and it just "waits" a few seconds

but now i am jut fixating on that issue.

how do you login?

ssh

iirc UseDNS defaults to yes

so it could just be waiting for a slow reverse DNS lookup

try to set UseDNS no in /etc/ssh/sshd_config, run service sshd restart and try again

or make sure even missing reverse DNS entry fail (near) instantly

well i have /etc/hosts messes on all the machiens which is probably making the problem worse

the dnsmasq setup is after this

voy4g3r2: ssh wants to look up other record types too

e.g. SSHFP

try ssh -vv, press enter a quickly a few times while it hangs

logout and look at the line just above the block of empty liens

*lines

if the delay is long enough to do it like that

of course there are better ways to do it, but this is the easiest if your terminal has enough scrollback

and no fancy shell config (e.g. a clock with second resolution that updates in the prompt) because that would cause the terminal to jump down agian

*again

but regarding the link0 flag on tap/vmnet: you want the bridge to not even attempt to deliver frames to the tap interfaces if the interface is down

because copying each multicast/broadcast/unknown unicast frame to the tap interfaces only to discard them is wasteful

the flag is important if you want to reuse tap interfaces in a routed network configuration

e.g. a /31 or /127 prefix per tap interface and the host redistributing all direct routes to tap*/vmnet* into a dynamic routing protocol

so then would that just mean it would be "easier" to isolate bhyve (tap) related traffic on one bridge and have another bridge with jails (epair)

because with the link0 the addresses and the routes that use them as nexthop would be deleted everytime the tap device is closed

voy4g3r2: no

there is no downside to mixing vnet enabled jails and bhyve guests on a shared bridge

only bad routing

:) that i introduced

just use epairs for the jails

err cat herding again (code word for meeting)

and tap interfaces for bhyve

yeah that is what i am using them for but it looks like my routing on the bridge itself is no bueno

Does anyone have experience with USB (SIM card) modems? I'm wondering which one has the best compatibility with FreeBSD.

Ideally something on a budget --- I don't need fast speed or anything of the sort.

crest: i am going to walk through and figure this all out, would it be cool to share iwth you my notes? i am thinking i am making a post about this.

voy4g3r2: sure. just send me a DM if you don't want it to get lost in the noise

DarkUranium: anything that can aggregate bands. Personally I'd avoid "USB stick" format and choose between smartphone tethering or a dedicated device like ZTE MF286D (which is ancient, underpowered, and cheap) over WiFi

I have my own Jail scheduler (jailer) and I'm going to be doing some massive changes in the codebase but keeping the front-end the same. I was wondering, what's the best way to write tests on FreeBSD? Is it still kyua? cheers

antranigv: yes, kyua is still the preferred way

dstolfa thank you.

not the best, but the standard

That's okay, I just need to run some code (jailer create jail0, jailer subcommand somearg) and then make sure that it does what it's supposed to do

e.g. returns 0, actually creates the jail files, etc

Also, if there's a getting started guide for Kyua, that would be even better

so I can first write the tests and then write the code, like a #SoftwareEngineer

perhaps see pf tests, I seem to recall them using a lot of jail infrastructure

antranigv: github.com/freebsd/kyua/wiki/Quickstart

Title: Quickstart · freebsd/kyua Wiki · GitHub

you probably care about tests written in shell, so: github.com/freebsd/kyua/wiki/Quickstart-atf-sh-test

Title: Quickstart atf sh test · freebsd/kyua Wiki · GitHub

dstolfa thank you <3

Is there a way to set sysctl values in loader.conf? I need to set "debug.debugger_on_panic=0", I tried setting this in sysctl.conf but it appears that that happens too late

My crash is at least earlier than those settings get applied

TMM_ I think there is. one sec lemme check

TMM_ I assume you also cannot get into single-user mode?

No, this crashes during PCI bus enumeration, and before my keyboard works. So to do anything I'll have to get the kernel to dump immediately so I can recover a minidump

I now get a ddb prompt I can't interact with :)

TMM_ I see. I just noticed that I was able to do that with OmniOS (which also use FreeBSD's bootloader), but lemme check if kernel supports that

I love how I keep forgetting to forward the ports into my IRC bouncer

TMM_ I still can't find it :D

I actually have time to debug these gzdoom compilation errors now

this is the full error from this invocation of clang++ termbin.com/mtt3

why are you compiling gzdoom? sounds exciting

it's funny you mention it but that uses cmake right?

i guess when porting to freebsd you have to uncmake it?

we use cmake in the gzdoom port

yes, but no. cmake works fine on FreeBSD

the only reason I haven't updated the gzdoom port is because I'm not ready for people to bitch at me about having to disable i386

I have a vacation coming up, so maybe it's the right time

well upstream gzdoom has broken at least some stuff, that's what I'm trying to sort out

johnjaye: just trying to build it because I'd like to play some Doom

that error's going to be because of needing sys/time.h included

most likely

but sys/stat.h is including sys/time.h

well, for __BSD_VISIBLE

there could be conflicting _*_SOURCE definitions being defined that would mean __BSD_VISIBLE doesn't get defined

where would I include sys/time.h, including it before sys/stat.h gives the same incomplete type errors

hmm, odd

oh, that's gross

I bet one of the includes outside of the FileSys namespace there ends up transitively including sys/time.h

so then include guards mean that we don't get _timeval.h *inside* the namespace like it wants

gack

hooray for C++

antranigv: no worries, I can't find it either. Maybe I just need to rebuild the kernel with some other option

kevans: if you figure something out I would be happy to test it

tp512: not going to have time for a while, but I'll think on it a bit

what they're doing is dirty and fragile imo

lovely, so 14-STABLE keeps panicking on me

openzfs zap_get_leaf_byblk calls _sx_slock_int and _sx_slock_hard which page faults

tp512: (re: gzdoom) do you actually need newer features past 4.7.1, or would games/gzdoom be able to fill the urge for now?

kevans: honestly, just trying to go with git HEAD since that's what I've done for the last 6 years, heh. 4.7.1 would probably suffice

*nod*

though the fact that this system might panic at any moment is kinda discouraging

I'm on the feb 29th snapshot, might be worth seeing if this is already something that's been fixed, though I'm torn on whether to use pkgbase

also hoping this is just a software issue and there's not an issue with this SSD causing the page faults

I wonder if these panics are related to the merging of openzfs 2.2

seems it's firefox ultimately doing something that leads to the crash and coincidentally I'm running firefox on an encrypted zfs dataset so I wonder if the panics are related to encryption somehow

has anyone else on 14-STABLE run into this? termbin.com/4vep

going to test w/o running the browser on this encrypted dataset. the idea was so that stuff like login cookies and would be safe if my laptop got stolen, but not something I have to worry about atm

kevans, how's the ecc stuff coming along? :)