On Wed, May 15, 2019 at 04:54:04PM +0200, Markus Armbruster wrote:
> Daniel P. Berrangé <berra...@redhat.com> writes:
>
> > On Wed, May 15, 2019 at 08:34:17AM +0200, Markus Armbruster wrote:
> >> Jason Wang <jasow...@redhat.com> writes:
> >>
> >> > On 2019/5/14 下午8:18, Markus Armbruster wrote:
> >> >> Peter Maydell <peter.mayd...@linaro.org> writes:
> >> >>
> >> >>> On Mon, 13 May 2019 at 14:21, Markus Armbruster <arm...@redhat.com>
> >> >>> wrote:
> >> >>>> Perhaps I should do it just for this file while I touch it anyway.
> >> >>>> The
> >> >>>> question to ask: should parse_acl_file() obey the locale for
> >> >>>> whitespace
> >> >>>> recognition?
> >> >>> I vote for "no".
> >> >>>
> >> >>> Q: do we document the format of the ACL file anywhere ?
> >> >> Support for it was added in commit bdef79a2994, v1.1. Just code, no
> >> >> documentation.
> >> >>
> >> >> Grepping for qemu-bridge-helper finds just qemu-options.hx. Contains
> >> >> -help output and some .texi that goes into qemu-doc and the manual page.
> >> >> None of it mentions how qemu-bridge-helper is run, or the ACL file
> >> >> feature, let alone what its format might be.
> >> >>
> >> >> I'm afraid all we have is the commit message. Which doesn't really
> >> >> define the file format, it merely gives a bunch of examples.
> >> >>
> >> >> As far as I can tell, qemu-bridge-helper is for use with -netdev tap and
> >> >> -netdev bridge.
> >> >>
> >> >> Both variations of -netdev call net_bridge_run_helper() to run the
> >> >> helper. First argument is -netdev parameter "helper", default usually
> >> >> "$prefix/libexec/qemu-bridge-helper". Second argument is parameter
> >> >> "br", default "br0".
> >> >>
> >> >> If @helper contains space or tab, net_bridge_run_helper() guesses its a
> >> >> full command, else it guesses its the name of the executable. Bad
> >> >> magic.
> >> >>
> >> >> If it guesses name of executable, it execv()s this executable with
> >> >> arguments "--use-vnet", "--fd=FD", "--br=@bridge".
> >> >>
> >> >> If it guesses full command, it appends "--use-vnet --fd=FD", where FD is
> >> >> the helper's half of the socketpair used to connect QEMU and the helper.
> >> >> It further appends "--br=@bridge", unless @helper contains "--br=".
> >> >> More bad magic.
> >> >>
> >> >> It executes the resulting string with sh -c. Magic cherry on top.
> >> >>
> >> >> When the helper fails, netdev creation fails.
> >> >>
> >> >> The helper we ship with QEMU unconditionally tries to read
> >> >> "$prefix/etc/bridge.conf". Fatal error if this file doesn't exist.
> >> >> Errors in this file are fatal. Errors in files it includes are not
> >> >> fatal; instead, the remainder of the erroneous file is ignored.
> >> >> *Boggle*
> >> >>
> >> >> As far as I can tell, libvirt runs qemu-bridge-helper itself (Paolo's
> >> >> commit 2d80fbb14df). Makes sense, because running QEMU with the
> >> >> necessary privileges would be unwise, and so would be letting it execute
> >> >> setuid helpers. Also bypasses the bad magic in QEMU's
> >> >> net_bridge_run_helper().
> >> >
> >> >
> >> > I don't notice this before. Is this only for the convenience of
> >> > development? I guess libvirt should have native support like adding
> >> > port to bridge/OVS without the help any external command or script.
> >>
> >> Commit 2d80fbb14df hints at the reason:
> >>
> >> <source type='bridge'> uses a helper application to do the necessary
> >> TUN/TAP setup to use an existing network bridge, thus letting
> >> unprivileged users use TUN/TAP interfaces.
> >> ~~~~~~~~~~~~~~~~~~
> >>
> >> The code confirms:
> >>
> >> /* qemuInterfaceBridgeConnect:
> >> * @def: the definition of the VM
> >> * @driver: qemu driver data
> >> * @net: pointer to the VM's interface description
> >> * @tapfd: array of file descriptor return value for the new device
> >> * @tapfdsize: number of file descriptors in @tapfd
> >> *
> >> ---> * Called *only* called if actualType is VIR_DOMAIN_NET_TYPE_NETWORK or
> >> ---> * VIR_DOMAIN_NET_TYPE_BRIDGE (i.e. if the connection is made with a
> >> tap
> >> * device connecting to a bridge device)
> >> */
> >> int
> >> qemuInterfaceBridgeConnect(virDomainDefPtr def,
> >> virQEMUDriverPtr driver,
> >> virDomainNetDefPtr net,
> >> int *tapfd,
> >> size_t *tapfdSize)
> >> {
> >> [...]
> >> ---> if (virQEMUDriverIsPrivileged(driver)) {
> >> [...]
> >> } else {
> >> if (qemuCreateInBridgePortWithHelper(cfg, brname,
> >> &net->ifname,
> >> tapfd, tap_create_flags)
> >> < 0) {
> >> virDomainAuditNetDevice(def, net, tunpath, false);
> >> goto cleanup;
> >> }
> >> [...]
> >> }
> >> [...]
> >> }
> >>
> >> >> qemu-bridge-helper should have a manual page, and its handling of errors
> >> >> in ACL include files needs work. There's probably more; I just glanced
> >> >> at it. I'm not volunteering, though. It lacks a maintainer. Should we
> >> >> add it to Jason's "Network device backends"?
> >> >
> >> >
> >> > Yes.
> >> >
> >> >> -netdev's helper parameter is seriously underdocumented. Document or
> >> >> deprecate?
> >> >
> >> >
> >> > I believe management should only use fd parameter of TAP. If we have
> >> > other, it should be a duplication. So I suggest to deprecate the
> >> > bridge helper and -netdev bridge.
> >>
> >> Objections, anyone?
> >
> > Libvirt runs the qemu bridge helper command directly, and we have
> > applications using this functionality.
>
> Specifically, when libvirt lacks the privileges to set up a TAP fd, it
> farms out the job to setuid qemu-bridge-helper. Correct?
Yes, this is for when using libvirt as an unpriv user - typically the
desktop virt use case.
> > I'd like libvirt to be able to avoid use of the QEMU bridge helper and
> > instead have unprivileged libvirt talk to privileged libvirtd to open a
> > TAP fd on its behalf. If we ever get that done, then libvirt would not
> > need the qemu bridge helper command anymore.
>
> We don't want to deprecate qemu-bridge-helper while libvirt has a
> sensible use for it.
>
> We can still deprecate -netdev tap parameter "helper" and -netdev bridge
> entirely.
No objection to that.
> Once they're gone, qemu-bridge-helper wull have no user within QEMU. We
> could discuss moving it to libvirt then, but I doubt it'll be worth the
> trouble.
I'm fine with that either way.
Regards,
Daniel
--
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org -o- https://fstop138.berrange.com :|
|: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
