Commit 7f64fb47 made the main package, and runc/libcontainer's CriuOpts
depend on criu/rpc. This is not good; among the other things, it makes
it complicated to make c/r optional.
Let's switch CriuOpts.ManageCgroupsMode to a string (yes, it's an APIt
breaking change) and move the cgroup mode string parsing to
libcontainer.
While at it, let's better document ManageCgroupsMode.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
If we get an unexpected error here, it is probably because of a library
or kernel change that could cause our detection logic to be invalid. As
a result, these warnings should be louder so users have a chance to tell
us about them sooner (or so we might notice them before doing a release,
as happened with the 1.2.0 regression).
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
It is possible for LinkAttachProgram to return ErrNotSupported if
program attachment is not supported at all (which doesn't matter in this
case), but it seems possible that upstream will start returning
ErrNotSupported for BPF_F_REPLACE at some point so it's best to make
sure we don't cause additional regressions here.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
In v0.13.0, cilium/ebpf stopped supporting setting BPF_F_REPLACE as an
explicit flag and instead requires us to use link.Anchor to specify
where the program should be attached.
Commit 216175a9ca ("Upgrade Cilium's eBPF library version to 0.16")
did update this correctly for the actual attaching logic, but when
checking for kernel support we still passed BPF_F_REPLACE. This would
result in a generic error being returned, which our feature-support
checking logic would treat as being an error the indicates that
BPF_F_REPLACE *is* supported, resulting in a regression on pre-5.6
kernels.
It turns out that our debug logging saying that this unexpected error
was happening was being output as a result of this change, but nobody
noticed...
Fixes: 216175a9ca ("Upgrade Cilium's eBPF library version to 0.16")
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
A map which is created in func init is only used by capSlice, which is
only used by New, which is only used by runc init. Switch to lazy init
to slightly save on startup time.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
If we are running with privileges and are asked to join an externally
created user namespaces as well as some other namespace that was *not*
created underneath said user namespace, the approach we added in commit
2cd9c31b99 ("nsenter: guarantee correct user namespace ordering")
doesn't work.
While in theory you would want all externally created namespaces to be
sane, it seems that some tools really do create unrelated namespaces and
ask us to join them. Luckily we can just loosely copy what nsenter(1)
appears to do -- we first try to join any namespaces we can (with host
root privileges), then we join any user namespaces, and then we join any
remaining namespaces (now with the user namespace's privileges).
Note that we *do not* have to try to join namespaces after we create our
own user namespace. Namespace permissions are based purely on the owning
user namespace (not the rootuid) so we will not have access to any extra
namespaces once we unshare(CLONE_NEWUSER) (in fact we will not be able
to setns(2) to anything!).
Fixes: 2cd9c31b99 ("nsenter: guarantee correct user namespace ordering")
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
This is basically a no-op change because runc already disallows this,
but it will be needed in future patches when we have to track what
namespaces have already been joined.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
If the sub-cgroup RemovePath has failed for any reason, return the
error right away. This way, we don't have to check for err != nil
before retrying rmdir.
This is a cosmetic change and should not change any functionality.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
An issue with runc 1.2.0 was reported to buildkit, in which
runc delete returns with an error, with the log saying:
> unable to destroy container: unable to remove container's cgroup: open /sys/fs/cgroup/snschvixiy3s74w74fjantrdg: no such file or directory
Apparently, what happens is runc is running with no cgroup access
(because /sys/fs/cgroup is mounted read-only). In this case error to
create a cgroup path (in runc create/run) is ignored, but cgroup removal
(in runc delete) is not.
This is caused by commit d3d7f7d, which changes the cgroup removal
logic in RemovePath. In the current code, if the initial rmdir has
failed (in this case with EROFS), but the subsequent os.ReadDir returns
ENOENT, it is returned (instead of being ignored -- as the path does not
exist and so there is nothing to remove).
Here is the minimal fix for the issue.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
If /run/runc and /usr/bin are on different filesystems, overlayfs may
enable the xino feature which results in the following log message:
kernel: overlayfs: "xino" feature enabled using 3 upper inode bits.
Each time we have to protect /proc/self/exe. So disable xino to remove
the log message (we don't care about the inode numbers of the files
anyway).
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
There is a typo in the comment (ClonedBinary should be CloneBinary), and
the code has changed a bit since then, and it makes more sense to refer
to CloneSelfExe now.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Because we have the overlay solution, we can drop runc-dmz binary
solution since it has too many limitations.
Signed-off-by: lifubang <lifubang@acmcoder.com>
Containerd pre-creates userns and netns before calling runc, which
results in the current code not working when SELinux is enabled,
resulting in the following error:
> runc create failed: unable to start container process: error during
container init: error mounting "mqueue" to rootfs at "/dev/mqueue":
setxattr /path/to/rootfs/dev/mqueue: operation not permitted
The solution is to become root in the user namespace right after
we join it.
Fixes#4466.
Co-authored-by: Wei Fu <fuweid89@gmail.com>
Co-authored-by: Kir Kolyshkin <kolyshkin@gmail.com>
Co-authored-by: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: lifubang <lifubang@acmcoder.com>
This allows to do
runc update $ID --memory=-1 --memory-swap=$VAL
for cgroup v2, i.e. set memory to unlimited and swap to a specific
value.
This was not possible because ConvertMemorySwapToCgroupV2Value rejected
memory=-1 ("unlimited"). In a hindsight, it was a mistake, because if
memory limit is unlimited, we should treat memory+swap limit as just swap
limit.
Revise the unit test; add description to each case.
Fixes: c86be8a2 ("cgroupv2: fix setting MemorySwap")
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Improve readability of ConvertMemorySwapToCgroupV2Value by switching
from a bunch of if statements to a switch, and adding a comment
describing each case.
No functional change.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Since Go 1.19, the same functionality is there in os/exec package.
As we require go 1.22 now, there's no need to have this.
This basically reverts commit 9258eac0 ("libct/start: use execabs for
newuidmap lookup").
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
There is a race situation when we are opening a file, if there is a
small fd was closed at that time, maybe it will be reused by safeExe.
Because of Go stdlib fds shuffling bug, if the fd of safeExe is too
small, go stdlib will dup3 it to another fd, or dup3 a other fd to this
fd, then it will cause the fd type cmd.Path refers to a random path,
and it can lead to an error "permission denied" when starting the process.
Please see #4294 and <https://github.com/golang/go/issues/61751>.
So we should not use the original fd of safeExe, but use the fd after
shuffled by Go stdlib. Because Go stdlib will guarantee this fd refers to
the correct file.
Signed-off-by: lfbzhm <lifubang@acmcoder.com>
Commit b999376fb2 ("nsenter: cloned_binary: remove bindfd logic
entirely") removed the read-only bind-mount logic from our cloned binary
code because it wasn't really safe because a container with
CAP_SYS_ADMIN could remove the MS_RDONLY bit and get write access to
/proc/self/exe (even with user namespaces this could've been an issue
because it's not clear if the flags are locked).
However, copying a binary does seem to have a minor performance impact.
The only way to have no performance impact would be for the kernel to
block these write attempts, but barring that we could try to reduce the
overhead by coming up with a mount that cannot have it's read-only bits
cleared.
The "simplest" solution is to create a temporary overlayfs using
fsopen(2) which uses the directory where runc exists as a lowerdir,
ensuring that the container cannot access the underlying file -- and we
don't have to do any copies.
While fsopen(2) is not free because mount namespace cloning is usually
expensive (and so it seems like the difference would be marginal), some
basic performance testing seems to indicate there is a ~60% improvement
doing it this way and that it has effectively no overhead even when
compared to just using /proc/self/exe directly:
% hyperfine --warmup 50 \
> "./runc-noclone run -b bundle ctr" \
> "./runc-overlayfs run -b bundle ctr" \
> "./runc-memfd run -b bundle ctr"
Benchmark 1: ./runc-noclone run -b bundle ctr
Time (mean ± σ): 13.7 ms ± 0.9 ms [User: 6.0 ms, System: 10.9 ms]
Range (min … max): 11.3 ms … 16.1 ms 184 runs
Benchmark 2: ./runc-overlayfs run -b bundle ctr
Time (mean ± σ): 13.9 ms ± 0.9 ms [User: 6.2 ms, System: 10.8 ms]
Range (min … max): 11.8 ms … 16.0 ms 180 runs
Benchmark 3: ./runc-memfd run -b bundle ctr
Time (mean ± σ): 22.6 ms ± 1.3 ms [User: 5.7 ms, System: 20.7 ms]
Range (min … max): 19.9 ms … 26.5 ms 114 runs
Summary
./runc-noclone run -b bundle ctr ran
1.01 ± 0.09 times faster than ./runc-overlayfs run -b bundle ctr
1.65 ± 0.15 times faster than ./runc-memfd run -b bundle ctr
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
This initWaiter logic was introduced by commit 4ecff8d9, but since the logic of
/proc/self/exe was moved out of runc init in commit 0e9a335, this
seems unnecessary to have initWaiter.
Remove it.
This essentially reverts commit 4ecff8d9.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
While we did set +x when "sealing" regular temporary files, the "is
executable" checks were done before then and would thus fail, causing
the fallback to not work properly.
So just set +x after we create the file. We already have a O_RDWR handle
open when we do the chmod so we won't get permission issues when writing
to the file.
Fixes: e089db3b4a ("dmz: add fallbacks to handle noexec for O_TMPFILE and mktemp()")
Signed-off-by: lifubang <lifubang@acmcoder.com>
We are not really interested in the capabilities of the current process,
so there is no need to load those.
This results in some performance improvement since now the capability
package don't have to parse /proc/self/status.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
The userns package was moved to the moby/sys/userns module
at commit 3778ae603c.
This patch deprecates the old location, and adds it as an alias
for the moby/sys/userns package.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
When preparing to mount container root, we need to make its parent mount
private (i.e. disable propagation), otherwise the new in-container
mounts are leaked to the host.
To find a parent mount, we use to read mountinfo and find the longest
entry which can be a parent of the container root directory.
Unfortunately, due to kernel bug in all Linux kernels older than v5.8
(see [1], [2]), sometimes mountinfo can't be read in its entirety. In
this case, getParentMount may occasionally return a wrong parent mount.
As a result, we do not change the mount propagation to private, and
container mounts are leaked.
Alas, we can not fix the kernel, and reading mountinfo a few times to
ensure its consistency (like it's done in, say, Kubernetes) does not
look like a good solution for performance reasons.
Fortunately, we don't need mountinfo. Let's just traverse the directory
tree, trying to remount it private until we find a mount point (any
error other than EINVAL means we just found it).
Fixes issue 2404.
[1]: https://github.com/kolyshkin/procfs-test
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=9f6c61f96f2d97cbb5f
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Initially, this was a commit to switch from strings.Fields to
strings.SplitN in getCpuUsageBreakdown, since strings.Fields
was probably slower than strings.SplitN in some old Go versions.
Afterwards, strings.Cut was also considered for potential
speed improvements.
After writing a benchmark test, we learned that:
- strings.Fields performance is now adequate;
- strings.SplitN is slower than strings.Fields;
- strings.Cut had <5% performance gain from strings.Fields;
So, remove the TODO and keep the benchmark test.
Signed-off-by: Stavros Panakakis <stavrospanakakis@gmail.com>
The example is too long since it lists too many capabilities.
Simplify it, leaving only two capabilities.
Also, remove ambient capabilities from the set. Inheritable capabilities
were removed earlier by commit 98fe566c, but ambient capabilities can't
be raised without inheritable ones.
Fixes: 98fe566c
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Commit 98fe566c removed inheritable capabilities from the example spec
(used by runc spec) and from the libcontainer/integration test config,
but neglected to also remove ambient capabilities.
An ambient capability could only be set if the same inheritable
capability is set, so as a result of the above change ambient
capabilities were not set (but due to a bug in gocapability package,
those errors are never reported).
Once we start using a library with the fix [1], that bug will become
apparent (both bats-based and libct/int tests will fail).
[1]: https://github.com/kolyshkin/capability/pull/3
Fixes: 98fe566c ("runc: do not set inheritable capabilities")
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
For some rootless container, runc has no access to cgroup,
But the container is still running. So we should return the
`ErrNotRunning` and `ErrCgroupNotExist` error seperatlly.
Signed-off-by: lifubang <lifubang@acmcoder.com>
Shared pid namespace means `runc kill` (or `runc delete -f`) have to
kill all container processes, not just init. To do so, it needs a cgroup
to read the PIDs from.
If there is no cgroup, processes will be leaked, and so such
configuration is bad and should not be allowed. To keep backward
compatibility, though, let's merely warn about this for now.
Alas, the only way to know if cgroup access is available is by returning
an error from Manager.Apply. Amend fs cgroup managers to do so (systemd
doesn't need it, since v1 can't work with rootless, and cgroup v2 does
not have a special rootless case).
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
In these cases, this is exactly what we want to find out.
Slightly improves performance and readability.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
It turns out that the suid and sgid mode bits are silently ignored by
Linux (though the sticky bit is honoured), and some users are requesting
mode bits that are ignored. While returning an error (as securejoin
does) makes some sense, this is a regression.
Ref: https://github.com/cyphar/filepath-securejoin/issues/23
Fixes: dd827f7b71 ("utils: switch to securejoin.MkdirAllHandle")
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
Kir Kolyshkin