This package is to provide unix.* wrappers to ensure that:
- they retry on EINTR;
- a "rich" error is returned on failure.
A first such wrapper, Sendmsg, is introduced.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
The errrolint linter can finally ignore errors from Close,
and it also ignores direct comparisons of errors from x/sys/unix.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This simplifies the code flow and basically removes the last
filepath.Clean, which is not necessary in either case:
- for absolute path, single filepath.Clean is enough (as it is
guaranteed to remove all dot and dot-dot elements);
- for relative path, filepath.Rel calls Clean at the end
(which is even documented).
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Using strings.CutPrefix (available since Go 1.20) instead of
strings.HasPrefix and/or strings.TrimPrefix makes the code
a tad more straightforward.
No functional change.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Using strings.HasPrefix with strings.TrimPrefix results in doing the
same thing (checking if prefix exists) twice. In this case, using
strings.TrimPrefix right away is sufficient.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This release includes a minor breaking API change that requires us to
rework the types of our wrappers, but there is no practical behaviour
change.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.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>
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>
It is available since Go 1.21 and is defined during compile time
(i.e. based on GOARCH during build).
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
filepath-securejoin has a bunch of extra hardening features and is very
well-tested, so we should use it instead of our own homebrew solution.
A lot of rootfs_linux.go callers pass a SecureJoin'd path, which means
we need to keep the wrapper helpers in utils, but at least the core
logic is no longer in runc. In future we will want to remove this dodgy
logic and just use file handles for everything (using libpathrs,
ideally).
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
While we use SecureJoin to try to make all of our target paths inside
the container safe, SecureJoin is not safe against an attacker than can
change the path after we "resolve" it.
os.MkdirAll can inadvertently follow symlinks and thus an attacker could
end up tricking runc into creating empty directories on the host (note
that the container doesn't get access to these directories, and the host
just sees empty directories). However, this could potentially cause DoS
issues by (for instance) creating a directory in a conf.d directory for
a daemon that doesn't handle subdirectories properly.
In addition, the handling for creating file bind-mounts did a plain
open(O_CREAT) on the SecureJoin'd path, which is even more obviously
unsafe (luckily we didn't use O_TRUNC, or this bug could've allowed an
attacker to cause data loss...). Regardless of the symlink issue,
opening an untrusted file could result in a DoS if the file is a hung
tty or some other "nasty" file. We can use mknodat to safely create a
regular file without opening anything anyway (O_CREAT|O_EXCL would also
work but it makes the logic a bit more complicated, and we don't want to
open the file for any particular reason anyway).
libpathrs[1] is the long-term solution for these kinds of problems, but
for now we can patch this particular issue by creating a more restricted
MkdirAll that refuses to resolve symlinks and does the creation using
file descriptors. This is loosely based on a more secure version that
filepath-securejoin now has[2] and will be added to libpathrs soon[3].
[1]: https://github.com/openSUSE/libpathrs
[2]: https://github.com/cyphar/filepath-securejoin/releases/tag/v0.3.0
[3]: https://github.com/openSUSE/libpathrs/issues/10
Fixes: CVE-2024-45310
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
The logic for how we create mountpoints is spread over each mountpoint
preparation function, when in reality the behaviour is pretty uniform
with only a handful of exceptions. So just move it all to one function
that is easier to understand.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
We close the logfd before execve so there's no need to special case it.
In addition, it turns out that (*os.File).Fd() doesn't handle the case
where the file was closed and so it seems suspect to use that kind of
check.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
If we leak a file descriptor referencing the host filesystem, an
attacker could use a /proc/self/fd magic-link as the source for execve
to execute a host binary in the container. This would allow the binary
itself (or a process inside the container in the 'runc exec' case) to
write to a host binary, leading to a container escape.
The simple solution is to make sure we close all file descriptors
immediately before the execve(2) step. Doing this earlier can lead to very
serious issues in Go (as file descriptors can be reused, any (*os.File)
reference could start silently operating on a different file) so we have
to do it as late as possible.
Unfortunately, there are some Go runtime file descriptors that we must
not close (otherwise the Go scheduler panics randomly). The only way of
being sure which file descriptors cannot be closed is to sneakily
go:linkname the runtime internal "internal/poll.IsPollDescriptor"
function. This is almost certainly not recommended but there isn't any
other way to be absolutely sure, while also closing any other possible
files.
In addition, we can keep the logrus forwarding logfd open because you
cannot execve a pipe and the contents of the pipe are so restricted
(JSON-encoded in a format we pick) that it seems unlikely you could even
construct shellcode. Closing the logfd causes issues if there is an
error returned from execve.
In mainline runc, runc-dmz protects us against this attack because the
intermediate execve(2) closes all of the O_CLOEXEC internal runc file
descriptors and thus runc-dmz cannot access them to attack the host.
Fixes: GHSA-xr7r-f8xq-vfvv CVE-2024-21626
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
While it doesn't make much of a practical difference, it seems far more
reasonable to use os.NewFile to wrap all of our passed file descriptors
to make sure they're tracked by the Go runtime and that we don't
double-close them.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
With the idmap work, we will have a tainted Go thread in our
thread-group that has a different mount namespace to the other threads.
It seems that (due to some bad luck) the Go scheduler tends to make this
thread the thread-group leader in our tests, which results in very
baffling failures where /proc/self/mountinfo produces gibberish results.
In order to avoid this, switch to using /proc/thread-self for everything
that is thread-local. This primarily includes switching all file
descriptor paths (CLONE_FS), all of the places that check the current
cgroup (technically we never will run a single runc thread in a separate
cgroup, but better to be safe than sorry), and the aforementioned
mountinfo code. We don't need to do anything for the following because
the results we need aren't thread-local:
* Checks that certain namespaces are supported by stat(2)ing
/proc/self/ns/...
* /proc/self/exe and /proc/self/cmdline are not thread-local.
* While threads can be in different cgroups, we do not do this for the
runc binary (or libcontainer) and thus we do not need to switch to
the thread-local version of /proc/self/cgroups.
* All of the CLONE_NEWUSER files are not thread-local because you
cannot set the usernamespace of a single thread (setns(CLONE_NEWUSER)
is blocked for multi-threaded programs).
Note that we have to use runtime.LockOSThread when we have an open
handle to a tid-specific procfs file that we are operating on multiple
times. Go can reschedule us such that we are running on a different
thread and then kill the original thread (causing -ENOENT or similarly
confusing errors). This is not strictly necessary for most usages of
/proc/thread-self (such as using /proc/thread-self/fd/$n directly) since
only operating on the actual inodes associated with the tid requires
this locking, but because of the pre-3.17 fallback for CentOS, we have
to do this in most cases.
In addition, CentOS's kernel is too old for /proc/thread-self, which
requires us to emulate it -- however in rootfs_linux.go, we are in the
container pid namespace but /proc is the host's procfs. This leads to
the incredibly frustrating situation where there is no way (on pre-4.1
Linux) to figure out which /proc/self/task/... entry refers to the
current tid. We can just use /proc/self in this case.
Yes this is all pretty ugly. I also wish it wasn't necessary.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
We have different requirements for the initial configuration and
initWaiter pipe (just send netlink and JSON blobs with no complicated
handling needed for message coalescing) and the packet-based
synchronisation pipe.
Tests with switching everything to SOCK_SEQPACKET lead to endless issues
with runc hanging on start-up because random things would try to do
short reads (which SOCK_SEQPACKET will not allow and the Go stdlib
explicitly treats as a streaming source), so splitting it was the only
reasonable solution. Even doing somewhat dodgy tricks such as adding a
Read() wrapper which actually calls ReadPacket() and makes it seem like
a stream source doesn't work -- and is a bit too magical.
One upside is that doing it this way makes the difference between the
modes clearer -- INITPIPE is still used for initWaiter syncrhonisation
but aside from that all other synchronisation is done by SYNCPIPE.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
This includes quite a few cleanups and improvements to the way we do
synchronisation. The core behaviour is unchanged, but switching to
embedding json.RawMessage into the synchronisation structure will allow
us to do more complicated synchronisation operations in future patches.
The file descriptor passing through the synchronisation system feature
will be used as part of the idmapped-mount and bind-mount-source
features when switching that code to use the new mount API outside of
nsexec.c.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
close_range(2) is far more efficient than a readdir of /proc/self/fd and
then doing a syscall for each file descriptor.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
Using strings.Split generates temporary strings for GC to collect.
Rewrite the function to not do that.
Also, add a second return value, so that the caller can distinguish
between an empty value found and no key found cases.
Fix the test accordingly.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Since commit 8850636eb3 (February 2015) this function is no longer
used (replaced by (*ConfigValidator).rootfs), so let's remove it,
together with its unit tests (which were added by commit 917c1f6d6 in
April 2016).
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
SendFds is a helper function for sending a set of file descriptors and a message
over a unix domain socket.
Signed-off-by: Mauricio Vásquez <mauricio@kinvolk.io>
Go 1.17 introduce this new (and better) way to specify build tags.
For more info, see https://golang.org/design/draft-gobuild.
As a way to seamlessly switch from old to new build tags, gofmt (and
gopls) from go 1.17 adds the new tags along with the old ones.
Later, when go < 1.17 is no longer supported, the old build tags
can be removed.
Now, as I started to use latest gopls (v0.7.1), it adds these tags
while I edit. Rather than to randomly add new build tags, I guess
it is better to do it once for all files.
Mind that previous commits removed some tags that were useless,
so this one only touches packages that can at least be built
on non-linux.
Brought to you by
go1.17 fmt ./...
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Only some libcontainer packages can be built on non-linux platforms
(not that it make sense, but at least go build succeeds). Let's call
these "good" packages.
For all other packages (i.e. ones that fail to build with GOOS other
than linux), it does not make sense to have linux build tag (as they
are broken already, and thus are not and can not be used on anything
other than Linux).
Remove linux build tag for all non-"good" packages.
This was mostly done by the following script, with just a few manual
fixes on top.
function list_good_pkgs() {
for pkg in $(find . -type d -print); do
GOOS=freebsd go build $pkg 2>/dev/null \
&& GOOS=solaris go build $pkg 2>/dev/null \
&& echo $pkg
done | sed -e 's|^./||' | tr '\n' '|' | sed -e 's/|$//'
}
function remove_tag() {
sed -i -e '\|^// +build linux$|d' $1
go fmt $1
}
SKIP="^("$(list_good_pkgs)")"
for f in $(git ls-files . | grep .go$); do
if echo $f | grep -qE "$SKIP"; then
echo skip $f
continue
fi
echo proc $f
remove_tag $f
done
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This should result in no change when the error is printed, but make the
errors returned unwrappable, meaning errors.As and errors.Is will work.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
gofumpt (mvdan.cc/gofumpt) is a fork of gofmt with stricter rules.
Brought to you by
git ls-files \*.go | grep -v ^vendor/ | xargs gofumpt -s -w
Looking at the diff, all these changes make sense.
Also, replace gofmt with gofumpt in golangci.yml.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Because the target of a mount is inside a container (which may be a
volume that is shared with another container), there exists a race
condition where the target of the mount may change to a path containing
a symlink after we have sanitised the path -- resulting in us
inadvertently mounting the path outside of the container.
This is not immediately useful because we are in a mount namespace with
MS_SLAVE mount propagation applied to "/", so we cannot mount on top of
host paths in the host namespace. However, if any subsequent mountpoints
in the configuration use a subdirectory of that host path as a source,
those subsequent mounts will use an attacker-controlled source path
(resolved within the host rootfs) -- allowing the bind-mounting of "/"
into the container.
While arguably configuration issues like this are not entirely within
runc's threat model, within the context of Kubernetes (and possibly
other container managers that provide semi-arbitrary container creation
privileges to untrusted users) this is a legitimate issue. Since we
cannot block mounting from the host into the container, we need to block
the first stage of this attack (mounting onto a path outside the
container).
The long-term plan to solve this would be to migrate to libpathrs, but
as a stop-gap we implement libpathrs-like path verification through
readlink(/proc/self/fd/$n) and then do mount operations through the
procfd once it's been verified to be inside the container. The target
could move after we've checked it, but if it is inside the container
then we can assume that it is safe for the same reason that libpathrs
operations would be safe.
A slight wrinkle is the "copyup" functionality we provide for tmpfs,
which is the only case where we want to do a mount on the host
filesystem. To facilitate this, I split out the copy-up functionality
entirely so that the logic isn't interspersed with the regular tmpfs
logic. In addition, all dependencies on m.Destination being overwritten
have been removed since that pattern was just begging to be a source of
more mount-target bugs (we do still have to modify m.Destination for
tmpfs-copyup but we only do it temporarily).
Fixes: CVE-2021-30465
Reported-by: Etienne Champetier <champetier.etienne@gmail.com>
Co-authored-by: Noah Meyerhans <nmeyerha@amazon.com>
Reviewed-by: Samuel Karp <skarp@amazon.com>
Reviewed-by: Kir Kolyshkin <kolyshkin@gmail.com> (@kolyshkin)
Reviewed-by: Akihiro Suda <akihiro.suda.cz@hco.ntt.co.jp>
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
Having -EPERM is the default was a fairly significant mistake from a
future-proofing standpoint in that it makes any new syscall return a
non-ignorable error (from glibc's point of view). We need to correct
this now because faccessat2(2) is something glibc critically needs to
have support for, but they're blocked on container runtimes because we
return -EPERM unconditionally (leading to confusion in glibc). This is
also a problem we're probably going to keep running into in the future.
Unfortunately there are several issues which stop us from having a clean
solution to this problem:
1. libseccomp has several limitations which require us to emulate
behaviour we want:
a. We cannot do logic based on syscall number, meaning we cannot
specify a "largest known syscall number";
b. libseccomp doesn't know in which kernel version a syscall was
added, and has no API for "minimum kernel version" so we cannot
simply ask libseccomp to generate sane -ENOSYS rules for us.
c. Additional seccomp rules for the same syscall are not treated as
distinct rules -- if rules overlap, seccomp will merge them. This
means we cannot add per-syscall -EPERM fallbacks;
d. There is no inverse operation for SCMP_CMP_MASKED_EQ;
e. libseccomp does not allow you to specify multiple rules for a
single argument, making it impossible to invert OR rules for
arguments.
2. The runtime-spec does not have any way of specifying:
a. The errno for the default action;
b. The minimum kernel version or "newest syscall at time of profile
creation"; nor
c. Which syscalls were intentionally excluded from the allow list
(weird syscalls that are no longer used were excluded entirely,
but Docker et al expect those syscalls to get EPERM not ENOSYS).
3. Certain syscalls should not return -ENOSYS (especially only for
certain argument combinations) because this could also trigger glibc
confusion. This means we have to return -EPERM for certain syscalls
but not as a global default.
4. There is not an obvious (and reasonable) upper limit to syscall
numbers, so we cannot create a set of rules for each syscall above
the largest syscall number in libseccomp. This means we must handle
inverse rules as described below.
5. Any syscall can be specified multiple times, which can make
generation of hotfix rules much harder.
As a result, we have to work around all of these things by coming up
with a heuristic to stop the bleeding. In the future we could hopefully
improve the situation in the runtime-spec and libseccomp.
The solution applied here is to prepend a "stub" filter which returns
-ENOSYS if the requested syscall has a larger syscall number than any
syscall mentioned in the filter. The reason for this specific rule is
that syscall numbers are (roughly) allocated sequentially and thus newer
syscalls will (usually) have a larger syscall number -- thus causing our
filters to produce -ENOSYS if the filter was written before the syscall
existed.
Sadly this is not a perfect solution because syscalls can be added
out-of-order and the syscall table can contain holes for several
releases. Unfortuntely we do not have a nicer solution at the moment
because there is no library which provides information about which Linux
version a syscall was introduced in. Until that exists, this workaround
will have to be good enough.
The above behaviour only happens if the default action is a blocking
action (in other words it is not SCMP_ACT_LOG or SCMP_ACT_ALLOW). If the
default action is permissive then we don't do any patching.
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
This is an additional mitigation for CVE-2019-16884. The primary problem
is that Docker can be coerced into bind-mounting a file system on top of
/proc (resulting in label-related writes to /proc no longer happening).
While we are working on mitigations against permitting the mounts, this
helps avoid our code from being tricked into writing to non-procfs
files. This is not a perfect solution (after all, there might be a
bind-mount of a different procfs file over the target) but in order to
exploit that you would need to be able to tweak a config.json pretty
specifically (which thankfully Docker doesn't allow).
Specifically this stops AppArmor from not labeling a process silently
due to /proc/self/attr/... being incorrectly set, and stops any
accidental fd leaks because /proc/self/fd/... is not real.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
This moves all console code to use github.com/containerd/console library to
handle console I/O. Also move to use EpollConsole by default when user requests
a terminal so we can still cope when the other side temporarily goes away.
Signed-off-by: Daniel Dao <dqminh89@gmail.com>
Since syscall is outdated and broken for some architectures,
use x/sys/unix instead.
There are still some dependencies on the syscall package that will
remain in syscall for the forseeable future:
Errno
Signal
SysProcAttr
Additionally:
- os still uses syscall, so it needs to be kept for anything
returning *os.ProcessState, such as process.Wait.
Signed-off-by: Christy Perez <christy@linux.vnet.ibm.com>
Kir Kolyshkin