Re: [LUG] NVidia

["M. J. Everitt" <m.j.everitt@xxxxxxx>]

On 29/10/18 17:55, mr meowski wrote:
> On 29/10/2018 17:26, M. J. Everitt wrote:
>
>>> I'm long overdue a major system overhaul, but not sure which way I'll go
>>> yet. I have had some issues with the meltdown/spectre patches on one
>>> system, so I've had to revert. I'm now running the ck patchset, which makes
>>> a noticeable difference in performance/memory usage.
>>> Next step is to change C library over to musl .. but that is a longer-term
>>> plan, depending on when this Big Overhaul takes place ..
>>>
>>> That said, I am loosely involved in the Gentoo kernel project, and am
>>> working on the new CI framework for it, so I have access to some clever
>>> people. Alas Greg K-H left Gentoo after the last hiatus (-sigh-) .. so we
>>> can't poke him directly so easily any more.
>>>
>>> MJE
>>>
>>>
>>>
>> Also anecdotally, one of the ZFS devs is a Gentoo dev also .. so if you 
>> can point to the specific issues, I can probably stir the pot a little ..
>> #zfsonlinux on Freenode IRC might be a resource to tap into.
>>
>> MJE
>>
> Wow you're really well connected! I'm just some little random bloke who 
> files bug reports and politely asks on forums for kernel devs to explain 
> what I'm doing wrong.
>
> I don't like IRC much and never have but that's a good idea actually - 
> perhaps a live chat with the people further up the kernel food chain can 
> sort things out. To be fair, this is a (known, recurring) bug rather 
> than me just being dumb - it's some weird old Solaris 64 bit specific 
> ICP assembly code that keeps sneaking back in and triggering fatal 
> compiler warnings in the ZFS module build. I did even dig around in the 
> tools/objtool kernel source a bit and experimented with editing the C 
> return codes around the specific function to pass over the buggy check 
> but rapidly realised I was seriously out of my depth.
>
> As you're already using the -ck patchset I seriously recommend you take 
> a look at pf too:
>
> https://gitlab.com/post-factum/pf-kernel/wikis/README
> https://gpo.zugaina.org/sys-kernel/pf-sources
>
> BFS, BFQ, UKSM, Graysky GCC patches, block multiqueue... I manually 
> patch in ZFS, Wireguard, a lot of old weird crap for dealing with my 
> beloved legacy UNIX systems, proper UFS r/w, AUFS... it's a long, long 
> list. The resulting properly packaged debs are signed, checked into a 
> repo and ready for my fleet of systems to install. I've been doing this 
> for so many years now that when I touch an Ubuntu system running stock 
> kernels I frequently wonder why it's so broken. Oh, and BPF too. That 
> won't work properly with compiling in support.
>
> Cheers
Yes I've briefly seen post-factum in IRC too, and correspond periodically
with 'con' of ck on another server. There was some hiccup with the
scheduler that pf- was using that made me double-take, but since I've also
been shoe-horned into adopting the zen-sources package which is presently
maintainerless, so that we have a better spread of resources in the distro.

I'm connected with folks who like kernel and system-level debugging down to
register level, although most of it does admittedly go right over my head -
some really interesting, clever people. I also know the guy who wrote
s6-init (and associated libs) but he's got a new job now, and is lamenting
having less time for FOSS projects! Oh and the lead dev of musl, because of
my connection with Adelie linux's conception ...

Anyhow, back to ZFS itself, I'll do a bit of digging, see what I can find
out. I met a few avid Gentoo users at FOSDEM '17 who swore by it, so if it
can be readily reproduced, it shouldn't be an issue to move it forward
somewhat .. :D

Best,
Michael.
aka veremitz on IRC :D

Attachment: signature.asc
Description: OpenPGP digital signature

-- 
The Mailing List for the Devon & Cornwall LUG
https://mailman.dclug.org.uk/listinfo/list
FAQ: http://www.dcglug.org.uk/listfaq