<MichaelRaskin>
I wonder what is the good way to manage the export list of an FS VM
<MichaelRaskin>
Maybe just dynamically launching a filtering process that adds a prefix to all requests…
<MichaelRaskin>
I think VM machine migration often means things that cannot be done as «migrated as source code»
<MichaelRaskin>
(I also think that a VM-per-application-instance implies that you have a very limited amount of VM images that are then given different connections to store)
<MichaelRaskin>
I think it is cheap to have multiple storage VMs, so you can still divide data into sensitivity domains, but without an obligation to align this perfectly with application VM allocation
<MichaelRaskin>
BTW, how long does it take to spin up and shutdown an empty CrosVM in your test runs?
<MichaelRaskin>
Hmm. If we need to mount different subsets of a single tree, and probably want to do it with a single server and filtering proxies, we might just as well ask the proxies to make sure the paths are well-formed and have no .. and are UTF-8-clean etc.
<MichaelRaskin>
Good question what happens to symlinks in all the mess
<qyliss>
I think the server doesn't need to care much about symlinks
<MichaelRaskin>
The server doesn't, but the system design doe
<MichaelRaskin>
does
<MichaelRaskin>
Do we imply that we mount just different subsets of the same tree (split into multiple parts by sensitivity etc.), or is layout customisable?
<qyliss>
I don't see why it shouldn't be customisable
<MichaelRaskin>
Neither do I, but then the user will need some reminder of how symlinks work
<qyliss>
I think if you're using symlinks you're responsible for understanding how they work.
<qyliss>
if this work has any impact, a very small proportion of users know what a symlink is
<MichaelRaskin>
Maybe. We are going to break a nontrivial amount of assumptions, though
<qyliss>
how so?
<qyliss>
mount points will work the same as any other unix systetm
<MichaelRaskin>
Kind of
<MichaelRaskin>
A typical unix system doesn't have the same subtree mounted at different paths all the time
<MichaelRaskin>
And to get to thepoint where nobody knows what a symlink is, you need to get through the phase where people who do know use and like the system
<MichaelRaskin>
Hm. I wonder at what level do we want to enforce read-only subtrees
<qyliss>
that would be done by the fs server
<MichaelRaskin>
Not sure what is the exact plan.
<MichaelRaskin>
The FS server will have a socket per subtree and per its read-only status?
<qyliss>
something like that
<MichaelRaskin>
If we can get cheap enough VMs, there is a benefit in having a single FS server socket for the actual FS server, and then multiple filtering VMs that enforce RO and subtrees
<MichaelRaskin>
Advantages: no dynamic reconfiguration of VMs and no dynamic reconfiguration of servers
<MichaelRaskin>
(Also, if we have dead-simple filtering proxies, we can skip «reinvent the filesystem»)
<MichaelRaskin>
(And looking at BtrFS history, reinventing the FS is not what you want to do in a single year)
<qyliss>
MichaelRaskin: oh, btw, takes 1.5 seconds or so to get to the point of a terminal appearing
<MichaelRaskin>
I guess you have a better CPU than I do.
<qyliss>
actually, that time it took more like 3s
<qyliss>
not slow, anyway
<qyliss>
this is on 2012 hardware
<MichaelRaskin>
Not sure what is the CPU year for Thinkpad W530
<qyliss>
this is a 220
<qyliss>
x220
<qyliss>
so probably 1 generation later
<qyliss>
how long does it take for you?
<MichaelRaskin>
Well, I do have an extra layer of container around
pie_ has quit [Ping timeout: 260 seconds]
<qyliss>
Yeah we should be comparing like for like
<MichaelRaskin>
But purely the dmesg part takes strictly more than 2 seconds, every time
<qyliss>
Yeah, at the moment I'm getting 2s of that, and then 1s later the window appears
<MichaelRaskin>
But this is just to 9p-mount and exit
<qyliss>
I've tagged mktuntap 1.0, btw. Seems like the thing to do since it has another user.
<qyliss>
If somebody else wants to submit to Nixpkgs they're most welcome to, but I won't self-submit.
<MichaelRaskin>
In a sense, as long as all the ChromiumOS stuff from SpectrumOS repo is not submitted upstream, mktuntap is a small thing (as it seems to be most useful for CrosVM: Qemu can select interface name on its own)
<qyliss>
I'm sure there are other use-cases
<qyliss>
execline-style tap/tun device configuration feels generally useful to me
<MichaelRaskin>
I wonder what is safer by now, wlroots or libvncclient
pie_ has joined #spectrum
pie_ has quit [Ping timeout: 258 seconds]
pie_ has joined #spectrum
tilpner_ has joined #spectrum
tilpner has quit [Ping timeout: 268 seconds]
<MichaelRaskin>
BTW: re: 9P over IPv6: in principle, if for non-containerised VMs you want IPv6, one could also do a connect() then mount via an FD
<MichaelRaskin>
I think authenticated 9P mounts work like that anyway