Vasily Averin <vvs@xxxxxxxxxxxxx> writes: > On 05/19/2014 11:30 PM, Bart De Schuymer wrote: >> As pointed out by Maciej, always >> starting from init_net isn't really an option in case of nested >> namespaces (start from the parent's namespace instead). > > Dear Bart, Serge, Maciej > thank you very much for your feedback! I am missing the context which makes raises this issue. > I've analyzed possibility to inherit settings from parent net-namespace, > discovered problems described below and finally decided to follow > Maciej's way (a) "use some kernel defaults", with adding an ability > to change pre-compiled kernel defaults. > > Below you can found more detailed description of discovered problems. > > 1) there are no (easy) ways to find parent of given network namespace. > > Network namespaces in kernel are not hierarchical but flat, > "struct net" have no reference to parent netns, and my collegians expect > that Eric Biederman will likely object to adding a parent netns pointer. > > Without this reference I do not see any good ways to copy parents > settings. Copying settings can easily happen at netowkr namespace creation time. Copying at any other time is too weird to even think about. So no you don't need a parent network namespace pointer to enable copying. > 2) settings inheriting does not work if subsystem module is loaded after > creation of network namespace. > > In this case all namespaces get pre-compiled defaults settings, and seems > there are no ways to apply "adjusted" setting to all already existing > netns. setns. > Moreover there is curious situation: to apply required sysctl settings > during module loading, Red Hat recommends to force "sysctl -p" execution > via install command in modprobe.conf > https://bugzilla.redhat.com/show_bug.cgi?id=634735#c7 > > However if module is loaded from inside one of network namespaces > it does not work! Why not? The appropriate events should fire globally. And note in most use cases participants in a network namespace won't have permissions to load modules. > In this case sysctl is executed inside netns. > If assigned sysctl key is not virtualized -- sysctl command can fail > if key is virtualized -- setting in current netns will be adjusted, > but not -- in init_net, that looks unexpected for me. >From what little you have said. This sounds like a don't do that then situation. Certainly if a module is the kernel triggers request module the module will be loaded in the initial set of namespaces. > I believe initial subsystem settings of newly created namespace should > not differ from initial settings of newly created subsystem in already > existing namespace. In case in-kernel setting inheriting this behavior > cannot be reached, additional subsystem tuning is required anyway. You are arguing that creation of a network namespace should use the kernel's default values for sysctls? That is a fairly reasonable position to take. > Therefore Maceiej's variant (a) "use some kernel defaults" looks like > right choice for me. If parent wants to assign some adjusted settings > in child environments -- it can only force loading of required modules > and apply required settings directly. > > At the same time I would like to have an ability to change pre-compiled > defaults somehow. In my patch I'm going to add new module options, that > allows node owner to specify wished "safe" settings before module loading, > and change them via sysfs after this. Why sysfs and not sysctl? It is not clear to me what is going on, from the limited details I see in this message it sounds like there may be a bit of overdesign and tackling problems that do not matter in the real world going on. For any kernel settings that apply to a network namespace we have two very basic choices. - Set them to default values when a namespace is initialized. - Copy them from somewhere when the namespace is created. Last I looked at that code we were copying sysctl values from the initial network namespace instead of the creators network namespace. Which has always seemed a bit silly to me. In general most people don't care and this does not cause an issue for most folks, or we could not have gone 5+ years without addressing it. For most things any practical program at this point is going to have to set the sysctls it cares about because it is going to have to run on existing kernels. Beyond that I don't have a strong opinion but we could either set values to well expected defaults, or copy them from the creators previous network namespace. Both would give deterministic results without any significant chance of breaking userspace today. Is their a compelling use case in this conversation that could weigh the decision of which semantics make the most sense? Adding sysfs entries or module parameters to change the action of sysctls sounds like there is something broken somewhere. Unfortunately it is not clear to me where that somewhere is. Eric -- To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
