On Wed, Aug 2, 2023 at 2:39 PM Aleksa Sarai <cyphar@xxxxxxxxxx> wrote: > > On 2023-08-02, Jeff Xu <jeffxu@xxxxxxxxxxxx> wrote: > > On Tue, Aug 1, 2023 at 6:05 PM Aleksa Sarai <cyphar@xxxxxxxxxx> wrote: > > > > > This thread is getting longer with different topics, I will try to > > respond with trimmed interleaved replies [1] > > There are 3 topics (logging/'migration/ratcheting), this response will > > be regarding ratcheting. > > The migration and ratcheting topics are interconnected because the > migration issue makes ratcheting an even more severe issue. But I'll > respond to each thread separately. > > > [1] https://www.kernel.org/doc/html/latest/process/submitting-patches.html?highlight=signed%20off#use-trimmed-interleaved-replies-in-email-discussions > > > > > > > > > > > > * The ratcheting mechanism for vm.memfd_noexec doesn't make sense as a > > > > > > > security mechanism because a CAP_SYS_ADMIN capable user can create > > > > > > > executable binaries in a hidden tmpfs very easily, not to mention the > > > > > > > many other things they can do. > > > > > > > > > > > > > By further limiting CAP_SYS_ADMIN, an attacker can't modify this > > > > > > sysctl even after compromising some system service with high > > > > > > privilege, YAMA has the same approach for ptrace_scope=3 > > > > > > > > > > Personally, I also think this behaviour from YAMA is a little goofy too, > > > > > but given that it only locks the most extreme setting and there is no > > > > > way to get around the most extreme setting, I guess it makes some sense > > > > > (not to mention it's an LSM and so there is an argument that it should > > > > > be possible to lock out privileged users from modifying it). > > > > > There are many other security sysctls, and very few have this behaviour > > > > > because it doesn't make much sense in most cases. > > > > > > > > > > > In addition, this sysctl is pid_name spaced, this means child > > > > > > pid_namespace will alway have the same or stricter security setting > > > > > > than its parent, this allows admin to maintain a tree like view. If we > > > > > > allow the child pid namespace to elevate its setting, then the > > > > > > system-wide setting is no longer meaningful. > > > > > > > > > > "no longer meaningful" is too strong of a statement imho. It is still > > > > > useful for constraining non-root processes and presumably ChromeOS > > > > > disallows random processes to do CLONE_NEWUSER (otherwise the protection > > > > > of this sysctl is pointless) so in practice for ChromeOS there is no > > > > > change in the attack surface. > > > > > > > > > > (FWIW, I think tying this to the user namespace would've made more sense > > > > > since this is about privilege restrictions, but that ship has sailed.) > > > > > > > > > The reason that this sysctl is a PID namespace is that I hope a > > > > container and host can have different sysctl values, e.g. host will > > > > allow runc's use of X mfd, while a container doesn't want X mfd. . > > > > To clarify what you meant, do you mean this: when a container is in > > > > its own pid_namespace, and has "=2", the programs inside the container > > > > can still use CLONE_NEWUSER to break out "=2" ? > > > > > > With the current implementation, this is not possible. My point was that > > > even if it were possible to lower the sysctl, ChromeOS presumably > > > already blocks the operations that a user would be able to use to create > > > a memfd (an unprivileged user cannot CLONE_NEWPID to modify the sysctl > > > without CLONE_NEWUSER, which is presumably blocked on ChromeOS due to > > > the other security concerns). > > > > > > > > > > > > The code sample shared in this patch set indicates that the attacker > > > > > > already has the ability of creating tmpfs and executing complex steps, > > > > > > at that point, it doesn't matter if the code execution is from memfd > > > > > > or not. For a safe by default system such as ChromeOS, attackers won't > > > > > > easily run arbitrary code, memfd is one of the open doors for that, so > > > > > > we are disabling executable memfd in ChromeOS. In other words: if an > > > > > > attacker can already execute the arbitrary code as sample given in > > > > > > ChromeOS, without using executable memfd, then memfd is no longer the > > > > > > thing we need to worry about, the arbitrary code execution is already > > > > > > achieved by the attacker. Even though I use ChromeOS as an example, I > > > > > > think the same type of threat model applies to any system that wants > > > > > > to disable executable memfd entirely. > > > > > > > > > > I understand the threat model this sysctl is blocking, my point is that > > > > > blocking CAP_SYS_ADMIN from modifying the setting doesn't make sense > > > > > from that threat model. An attacker that manages to trick some process > > > > > into creating a memfd with an executable payload is not going to be able > > > > > to change the sysctl setting (unless there's a confused deputy with > > > > > CAP_SYS_ADMIN, in which case you have much bigger issues). > > > > > > > > > It is the reverse. An attacker that manages to trick some > > > > CAP_SYSADMIN processes into changing this sysctl value (i.e. lower the > > > > setting to 0 if no ratcheting), will be able to continue to use mfd as > > > > part of the attack chain. > > > > In chromeOS, an attacker that can change sysctl might not necessarily > > > > gain full arbitrary code execution already. As I mentioned previously, > > > > the main threat model here is to prevent arbitrary code execution > > > > through mfd. If an attacker already gains arbitrary code execution, > > > > at that point, we no longer worry about mfd. > > > > > > If an attacker can trick a privileged process into writing to arbitrary > > > sysctls, the system has much bigger issues than arbitrary (presumably > > > unprivileged) code execution. On the other hand, requiring you to reboot > > > a server due to a misconfigured sysctl *is* broken. > > > > > > Again, at the very least, not even allowing capable(CAP_SYS_ADMIN) to > > > change the setting is actually broken. > > > > > > > > If a CAP_SYS_ADMIN-capable user wants to change the sysctl, blocking it > > > > > doesn't add any security because that process could create a memfd-like > > > > > fd to execute without issues. > > > > >What practical attack does this ratcheting > > > > > mechanism protect against? (This is a question you can answer with the > > > > > YAMA sysctl, but not this one AFAICS.) > > > > > > > > > > But even if you feel that allowing this in child user namespaces is > > > > > unsafe or undesirable, it's absolutely necessary that > > > > > capable(CAP_SYS_ADMIN) should be able to un-brick the running system by > > > > > changing the sysctl. The alternative is that you need to reboot your > > > > > server in order to un-set a sysctl that broke some application you run. > > > > > > > > > > > > > > Also, by the same token, this ratcheting mechanism doesn't make sense > > > > > with =1 *at all* because it could break programs in a way that would > > > > > require a reboot but it's not a "security setting" (and the YAMA sysctl > > > > > mentioned only locks the sysctl at the highest setting). > > > > > > > > > I think a system should use "=0" when it is unsure about its program's > > > > need or not need executable memfd. Technically, it is not that this > > > > sysctl breaks the user, but the admin made the mistake to set the > > > > wrong sysctl value, and an admin should know what they are doing for a > > > > sysctl. Yes. rebooting increases the steps to undo the mistake, but > > > > that could be an incentive for the admin to fully test its programs > > > > before turning on this sysctl - and avoid unexpected runtime errors. > > > > > > I don't think this stance is really acceptable -- if an admin that has > > > privileges to load kernel modules is not able to disable a sysctl that > > > can break working programs without rebooting there is > > > > > > When this sysctl was first proposed a few years ago (when kernel folks > > > found out that runc was using executable memfds), my understanding is > > > that the long-term goal was to switch programs to have > > > non-executable-memfds by default on most distributions. Making it > > > impossible for an admin to lower the sysctl value flies in the face of > > > this goal. > > > > > > At the very least, being unable to lower the sysctl from =1 to =0 is > > > just broken (even if you use the yama example -- yama only locks the > > > sysctl at highest possible setting, not on lower settings). But in my > > > view, having this sysctl ratchet at all doesn't make sense. > > > > > To reiterate/summarize the current mechanism for vm.memfd_noexec > > > > 1> It is a pid namespace sysctl, init ns and child pid ns can have > > different setting values. > > 2> child pid ns inherits parent's pid ns's sysctl at the time of fork. > > 3> There are 3 values for the sysctl, each higher value is more > > restrictive than the lower one. Once set, doesn't allow downgrading. > > > > It can be used as following: > > 1> > > init ns: vm.memfd_noexec = 2 (at boot time) > > Not allow executable memfd for the entire system, including its containers. > > > > 2> > > init ns: vm.memfd_noexec = 0 or 1 > > container (child init namespace) vm.memfd_noexec = 2. > > The host allows runc's usage of executable memfd during container > > creation. Inside the container, executable memfd is not allowed. > > > > The inherence + not allow downgrading is to reason with how > > vm.memfd_noexec is applied in the process tree. > > Without it, essentially we are losing the hierarchy view across the > > process tree and a process can evaluate its capability by modifying > > the setting. I think that is a less secure approach I would not > > prefer. > > If you really want the hierarchical aspect, we can implement it so that > it's _actually_ hierarchical like so: > > * By default, your setting is the same as your parent (this is checked > by going up the pidns tree -- a-la cgroups). This is less efficient > but you want a hierarchy, so we can do it this way instead. > * If you set a specific setting, that takes precedence but only if it's > a greater or equal setting to your parent. > * Trying to set a lower setting than your parent fails regardless of > privileges. > > This will allow *privileged users* to lower the setting, but only if > the parent pidns also has a lower setting. This allows a system admin to > enforce the setting. It seems to me that this fulfils all the > requirements you have. > > Most importantly, this would allow for a hierarchical view without > having a sysctl that will break systems and nobody will use. I need to > re-iterate this point -- nobody is going to use this sysctl as it > currently works because it ratchets in a way that admins cannot undo. In > practice this would mean you would need to reboot your whole datacenter > if you didn't catch that an update to one of you dependencies didn't > pass a required *noop* flag to memfd_create(). > Yes. I agree this is another way to implement a hierarchical view, which is a little more costly, because it goes up the process tree. I respectfully disagree that nobody will use the current sysctl though, I can still see that a container might want this, e.g. a small container that doesn't require a lot of refactoring to add NX, and restarting container usually isn't a problem, and container might like the fact that downgrade is denied at run time. Thanks Best regards, -Jeff