On Fri, Jun 01, 2018 at 02:08:38PM +0200, Kashyap Chamarthy wrote: > Signed-off-by: Kashyap Chamarthy <kchamart(a)redhat.com&gt; > --- > src/qemu/qemu.conf | 5 +++-- > 1 file changed, 3 insertions(+), 2 deletions(-) > Why?

On Fri, Jun 01, 2018 at 02:46:40PM +0200, Peter Krempa wrote: > On Fri, Jun 01, 2018 at 13:32:20 +0100, Daniel Berrange wrote:

> > The reason the config file documents 'root' is because that is what > > configure defaults to. If you pass --with-qemu-user to configure, > > we don't update the config file example though, and perhaps we should.

> > Alternatively should we make configure defualt to 'qemu' instead of > > 'root', since it is generally considered insane to run QEMU as root. > > But user 'qemu' is not by default present on all systems. Even the > libvirt.spec file creates the account. Yes, that's the reason configure defaults to 'root', but I really hate the fact that we default to a config that no one should ever run in practice. We could check for existance of 'qemu' in configure and complain if it is missing, but that's painful in itself as it is valid to build on a host without the user, as long as it exists at runtime. I tend to think we should just blindly use qemu/qemu by default and document that creating these accounts is a requirement. Users will quickly see if they're missing when they try to start a guest.

> As a second thought, we generally use commented-out bits that are the > non-default configuration. So this fits the pattern in the extent that > any sane distro specified it's own user/group using the configure > options and if for any reason the user wants to run this as root it's > done just by uncommenting it. Most commented out bits are not a security flaw if uncommented though. The fact that we show 'user=root' in the config file though puts across the misleading idea that it is a reasonable thing todo, when in fact it is a horribly insecure thing todo.

On Fri, Jun 01, 2018 at 05:05:30PM +0200, Kashyap Chamarthy wrote: > On Fri, Jun 01, 2018 at 02:11:12PM +0100, Daniel P. Berrangé wrote: > > On Fri, Jun 01, 2018 at 02:46:40PM +0200, Peter Krempa wrote: > > > On Fri, Jun 01, 2018 at 13:32:20 +0100, Daniel Berrange wrote: > > [...] > > > > > The reason the config file documents 'root' is because that is what > > > > configure defaults to. If you pass --with-qemu-user to configure, > > > > we don't update the config file example though, and perhaps we should. > > Thanks for that 'configure' context. > > > > > Alternatively should we make configure defualt to 'qemu' instead of > > > > 'root', since it is generally considered insane to run QEMU as root. > > > > > > But user 'qemu' is not by default present on all systems. Even the > > > libvirt.spec file creates the account. > > > > Yes, that's the reason configure defaults to 'root', but I really hate > > the fact that we default to a config that no one should ever run in > > practice. > > Nobody should be using ./configure with the defaults in practice anyway. For real world usage, people should go through their distribution's packages which should have specified --with-qemu-user. > > We could check for existance of 'qemu' in configure and complain if > > it is missing, but that's painful in itself as it is valid to build > > on a host without the user, as long as it exists at runtime. > > We probe for many things that are only relevant at runtime (paths to various binaries come to mind). For many things we try to be smart and try to guess what the user wants and turn off configure options when dependencies for drivers and features are missing, instead of building everything by default and letting the user disable stuff manually if they don't want the dependencies. Which brings me to the question: who are the ./configure defaults for? Distributions usually specify all possible configure options when building the packages and end users IMO should not be building from the source. Should we try to build everything and error out when it's not present? Or try to build as much as we can with the libaries and dependencies currently present in the system?

> > I tend to think we should just blindly use qemu/qemu by default and > > document that creating these accounts is a requirement. Users will > > quickly see if they're missing when they try to start a guest. > > I'll try to audit what user all the different distributions (that > matter) use to launch QEMU. If they are all are using 'qemu' anyway, > then probably we can just go with 'qemu:qemu', and document the > requirement as such. Just by defaulting to qemu:qemu if they are present in the system you can eliminate many setups that would default to root.

But if we start requiring things we consider sensible at configure time, I'd happily clean up more of our automagic and let the users frugal on system libraries supply --without- arguments for stuff they don't want.

> > > As a second thought, we generally use commented-out bits that are the > > > non-default configuration. So this fits the pattern in the extent that > > > any sane distro specified it's own user/group using the configure > > > options and if for any reason the user wants to run this as root it's > > > done just by uncommenting it. > > > > Most commented out bits are not a security flaw if uncommented though. > > The fact that we show 'user=root' in the config file though puts across > > the misleading idea that it is a reasonable thing todo, when in fact it > > is a horribly insecure thing todo. > In that case, #allow_disk_format_probing = 1

should also probably go. Maybe {vnc,tls}_listen = "0.0.0.0" as well.