[1.1] nsenter: cloned_binary: remove bindfd logic entirely #4392
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
(This is a cherry-pick of b999376.) While the ro-bind-mount trick did eliminate the memory overhead of copying the runc binary for each "runc init" invocation, on machines with very significant container churn, creating a temporary mount namespace on every container invocation can trigger severe lock contention on namespace_sem that makes containers fail to spawn. The only reason we added bindfd in commit 16612d7 ("nsenter: cloned_binary: try to ro-bind /proc/self/exe before copying") was due to a Kubernetes e2e test failure where they had a ridiculously small memory limit. It seems incredibly unlikely that real workloads are running without 10MB to spare for the very short time that runc is interacting with the container. In addition, since the original cloned_binary implementation, cgroupv2 is now almost universally used on modern systems. Unlike cgroupv1, the cgroupv2 memcg implementation does not migrate memory usage when processes change cgroups (even cgroupv1 only did this if you had memory.move_charge_at_immigrate enabled). In addition, because we do the /proc/self/exe clone before synchronising the bootstrap data read, we are guaranteed to do the clone before "runc init" is moved into the container cgroup -- meaning that the memory used by the /proc/self/exe clone is charged against the root cgroup, and thus container workloads should not be affected at all with memfd cloning. The long-term fix for this problem is to block the /proc/self/exe re-opening attack entirely in-kernel, which is something I'm working on[1]. Though it should also be noted that because the memfd is completely separate to the host binary, even attacks like Dirty COW against the runc binary can be defended against with the memfd approach. Of course, once we have in-kernel protection against the /proc/self/exe re-opening attack, we won't have that protection anymore... [1]: https://lwn.net/Articles/934460/ Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
lifubang
approved these changes
Sep 3, 2024
Merged
This was referenced Oct 7, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Backport of #3931.
While the ro-bind-mount trick did eliminate the memory overhead of copying the runc binary for each "runc init" invocation, on machines with very significant container churn, creating a temporary mount namespace on every container invocation can trigger severe lock contention on namespace_sem that makes containers fail to spawn.
The only reason we added bindfd in commit 16612d7 ("nsenter: cloned_binary: try to ro-bind /proc/self/exe before copying") was due to a Kubernetes e2e test failure where they had a ridiculously small memory limit. It seems incredibly unlikely that real workloads are running without 10MB to spare for the very short time that runc is interacting with the container.
In addition, since the original cloned_binary implementation, cgroupv2 is now almost universally used on modern systems. Unlike cgroupv1, the cgroupv2 memcg implementation does not migrate memory usage when processes change cgroups (even cgroupv1 only did this if you had memory.move_charge_at_immigrate enabled). In addition, because we do the /proc/self/exe clone before synchronising the bootstrap data read, we are guaranteed to do the clone before "runc init" is moved into the container cgroup -- meaning that the memory used by the /proc/self/exe clone is charged against the root cgroup, and thus container workloads should not be affected at all with memfd cloning.
The long-term fix for this problem is to block the /proc/self/exe re-opening attack entirely in-kernel, which is something I'm working on. Though it should also be noted that because the memfd is completely separate to the host binary, even attacks like Dirty COW against the runc binary can be defended against with the memfd approach. Of course, once we have in-kernel protection against the /proc/self/exe re-opening attack, we won't have that protection anymore...