This will permit us to extend the internals of systemd.Manager to include
further information about the system, such as whether cgroupv1, cgroupv2 or
both are in effect.
Furthermore, it allows a future refactor of moving more of UseSystemd() code
into the factory initialization function.
Signed-off-by: Filipe Brandenburger <filbranden@gmail.com>
Minor refactoring to use the filePair struct for both init sock and log pipe
Co-authored-by: Julia Nedialkova <julianedialkova@hotmail.com>
Signed-off-by: Georgi Sabev <georgethebeatle@gmail.com>
In the exception handling of initProcess.start(), we need to add the
missing IntelRdtManager.Destroy() handler in defer func.
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Bump logrus so that we can use logrus.StandardLogger().Logf instead
Co-authored-by: Julia Nedialkova <julianedialkova@hotmail.com>
Signed-off-by: Georgi Sabev <georgethebeatle@gmail.com>
We discovered in umoci that setting a dummy type of "none" would result
in file-based bind-mounts no longer working properly, which is caused by
a restriction for when specconv will change the device type to "bind" to
work around rootfs_linux.go's ... issues.
However, bind-mounts don't have a type (and Linux will ignore any type
specifier you give it) because the type is copied from the source of the
bind-mount. So we should always overwrite it to avoid user confusion.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Refactor configuring logging into a reusable component
so that it can be nicely used in both main() and init process init()
Co-authored-by: Georgi Sabev <georgethebeatle@gmail.com>
Co-authored-by: Giuseppe Capizzi <gcapizzi@pivotal.io>
Co-authored-by: Claudia Beresford <cberesford@pivotal.io>
Signed-off-by: Danail Branekov <danailster@gmail.com>
Add support for children processes logging (including nsexec).
A pipe is used to send logs from children to parent in JSON.
The JSON format used is the same used by logrus JSON formatted,
i.e. children process can use standard logrus APIs.
Signed-off-by: Marco Vedovati <mvedovati@suse.com>
Whenever processes are spawned using nsexec, a zombie runc:[1:CHILD]
process will always be created and will need to be reaped by the parent
Signed-off-by: Alex Fang <littlelightlittlefire@gmail.com>
secure_getenv is a Glibc extension and so this code does not compile
on Musl libc any more after this patch.
secure_getenv is only intended to be used in setuid binaries, in
order that they should not trust their environment. It simply returns
NULL if the binary is running setuid. If runc was installed setuid,
the user can already do anything as root, so it is game over, so this
check is not needed.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Work is ongoing in the kernel to support different kernel
keyrings per user namespace. We want to allow SELinux to manage
kernel keyrings inside of the container.
Currently when runc creates the kernel keyring it gets the label which runc is
running with ususally `container_runtime_t`, with this change the kernel keyring
will be labeled with the container process label container_t:s0:C1,c2.
Container running as container_t:s0:c1,c2 can manage keyrings with the same label.
This change required a revendoring or the SELinux go bindings.
github.com/opencontainers/selinux.
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
There are some circumstances where sendfile(2) can fail (one example is
that AppArmor appears to block writing to deleted files with sendfile(2)
under some circumstances) and so we need to have a userspace fallback.
It's fairly trivial (and handles short-writes).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
The usage of memfd_create(2) and other copying techniques is quite
wasteful, despite attempts to minimise it with _LIBCONTAINER_STATEDIR.
memfd_create(2) added ~10M of memory usage to the cgroup associated with
the container, which can result in some setups getting OOM'd (or just
hogging the hosts' memory when you have lots of created-but-not-started
containers sticking around).
The easiest way of solving this is by creating a read-only bind-mount of
the binary, opening that read-only bindmount, and then umounting it to
ensure that the host won't accidentally be re-mounted read-write. This
avoids all copying and cleans up naturally like the other techniques
used. Unfortunately, like the O_TMPFILE fallback, this requires being
able to create a file inside _LIBCONTAINER_STATEDIR (since bind-mounting
over the most obvious path -- /proc/self/exe -- is a *very bad idea*).
Unfortunately detecting this isn't fool-proof -- on a system with a
read-only root filesystem (that might become read-write during "runc
init" execution), we cannot tell whether we have already done an ro
remount. As a partial mitigation, we store a _LIBCONTAINER_CLONED_BINARY
environment variable which is checked *alongside* the protection being
present.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Writing a file to tmpfs actually incurs a memcg penalty, and thus the
benefit of being able to disable memfd_create(2) with
_LIBCONTAINER_DISABLE_MEMFD_CLONE is fairly minimal -- though it should
be noted that quite a few distributions don't use tmpfs for /tmp (and
instead have it as a regular directory or subvolume of the host
filesystem).
Since runc must have write access to the state directory anyway (and the
state directory is usually not on a tmpfs) we can use that instead of
/tmp -- avoiding potential memcg costs with no real downside.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
In order to get around the memfd_create(2) requirement, 0a8e4117e7
("nsenter: clone /proc/self/exe to avoid exposing host binary to
container") added an O_TMPFILE fallback. However, this fallback was
flawed in two ways:
* It required O_TMPFILE which is relatively new (having been added to
Linux 3.11).
* The fallback choice was made at compile-time, not runtime. This
results in several complications when it comes to running binaries
on different machines to the ones they were built on.
The easiest way to resolve these things is to have fallbacks work in a
more procedural way (though it does make the code unfortunately more
complicated) and to add a new fallback that uses mkotemp(3).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
For a variety of reasons, sendfile(2) can end up doing a short-copy so
we need to just loop until we hit the binary size. Since /proc/self/exe
is tautologically our own binary, there's no chance someone is going to
modify it underneath us (or changing the size).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
This makes use of the vendored in Go bindings and removes the copy of
the CRIU RPC interface definition. runc now relies on go-criu for RPC
definition and hopefully more CRIU functions can be used in the future
from the CRIU Go bindings.
Signed-off-by: Adrian Reber <areber@redhat.com>
My first attempt to simplify this and make it less costly focussed on
the way constructors are called. I was under the impression that the ELF
specification mandated that arg, argv, and actually even envp need to be
passed to functions located in the .init_arry section (aka
"constructors"). Actually, the specifications is (cf. [2]):
SHT_INIT_ARRAY
This section contains an array of pointers to initialization functions,
as described in ``Initialization and Termination Functions'' in Chapter
5. Each pointer in the array is taken as a parameterless procedure with
a void return.
which means that this becomes a libc specific decision. Glibc passes
down those args, musl doesn't. So this approach can't work. However, we
can at least remove the environment parsing part based on POSIX since
[1] mandates that there should be an environ variable defined in
unistd.h which provides access to the environment. See also the relevant
Open Group specification [1].
[1]: http://pubs.opengroup.org/onlinepubs/9699919799/
[2]: http://www.sco.com/developers/gabi/latest/ch4.sheader.html#init_array
Fixes: CVE-2019-5736
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
The detection for scope properties (whether scope units support
DefaultDependencies= or Delegate=) has always been broken, since systemd
refuses to create scopes unless at least one PID is attached to it (and
this has been so since scope units were introduced in systemd v205.)
This can be seen in journal logs whenever a container is started with
libpod:
Feb 11 15:08:07 myhost systemd[1]: libcontainer-12345-systemd-test-default-dependencies.scope: Scope has no PIDs. Refusing.
Feb 11 15:08:07 myhost systemd[1]: libcontainer-12345-systemd-test-default-dependencies.scope: Scope has no PIDs. Refusing.
Since this logic never worked, just assume both attributes are supported
(which is what the code does when detection fails for this reason, since
it's looking for an "unknown attribute" or "read-only attribute" to mark
them as false) and skip the detection altogether.
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
runc creates all missing mountpoints when it starts a container, this
commit also creates those mountpoints during restore. Now it is possible
to restore a container using the same, but newly created rootfs just as
during container start.
Signed-off-by: Adrian Reber <areber@redhat.com>
During rootfs setup all mountpoints (directory and files) are created
before bind mounting the bind mounts. This does not happen during
container restore via CRIU. If restoring in an identical but newly created
rootfs, the restore fails right now. This just factors out the code to
create the bind mount mountpoints so that it also can be used during
restore.
Signed-off-by: Adrian Reber <areber@redhat.com>
There are quite a few circumstances where /proc/self/exe pointing to a
pretty important container binary is a _bad_ thing, so to avoid this we
have to make a copy (preferably doing self-clean-up and not being
writeable).
We require memfd_create(2) -- though there is an O_TMPFILE fallback --
but we can always extend this to use a scratch MNT_DETACH overlayfs or
tmpfs. The main downside to this approach is no page-cache sharing for
the runc binary (which overlayfs would give us) but this is far less
complicated.
This is only done during nsenter so that it happens transparently to the
Go code, and any libcontainer users benefit from it. This also makes
ExtraFiles and --preserve-fds handling trivial (because we don't need to
worry about it).
Fixes: CVE-2019-5736
Co-developed-by: Christian Brauner <christian.brauner@ubuntu.com>
Signed-off-by: Aleksa Sarai <asarai@suse.de>
For some reason, libcontainer/integration has a whole bunch of incorrect
usages of libcontainer.Factory -- causing test failures with a set of
security patches that will be published soon. Fixing ths is fairly
trivial (switch to creating a new libcontainer.Factory once in each
process, rather than creating one in TestMain globally).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
When creating a new user namespace, the kernel doesn't allow to mount
a new procfs or sysfs file system if there is not already one instance
fully visible in the current mount namespace.
When using --no-pivot we were effectively inhibiting this protection
from the kernel, as /proc and /sys from the host are still present in
the container mount namespace.
A container without full access to /proc could then create a new user
namespace, and from there able to mount a fully visible /proc, bypassing
the limitations in the container.
A simple reproducer for this issue is:
unshare -mrfp sh -c "mount -t proc none /proc && echo c > /proc/sysrq-trigger"
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
since commit df3fa115f9 it is not
possible to set a kernel memory limit when using the systemd cgroups
backend as we use cgroup.Apply twice.
Skip enabling kernel memory if there are already tasks in the cgroup.
Without this patch, runc fails with:
container_linux.go:344: starting container process caused
"process_linux.go:311: applying cgroup configuration for process
caused \"failed to set memory.kmem.limit_in_bytes, because either
tasks have already joined this cgroup or it has children\""
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
This patch fixes a corner case when destroy a container:
If we start a container without 'intelRdt' config set, and then we run
“runc update --l3-cache-schema/--mem-bw-schema” to add 'intelRdt' config
implicitly.
Now if we enter "exit" from the container inside, we will pass through
linuxContainer.Destroy() -> state.destroy() -> intelRdtManager.Destroy().
But in IntelRdtManager.Destroy(), IntelRdtManager.Path is still null
string, it hasn’t been initialized yet. As a result, the created rdt
group directory during "runc update" will not be removed as expected.
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
CRIU 3.11 introduces configuration files:
https://criu.org/Configuration_fileshttps://lisas.de/~adrian/posts/2018-Nov-08-criu-configuration-files.html
This enables the user to influence CRIU's behaviour without code changes
if using new CRIU features or if the user wants to enable certain CRIU
behaviour without always specifying certain options.
With this it is possible to write 'tcp-established' to the configuration
file:
$ echo tcp-established > /etc/criu/runc.conf
and from now on all checkpoints will preserve the state of established
TCP connections. This removes the need to always use
$ runc checkpoint --tcp-stablished
If the goal is to always checkpoint with '--tcp-established'
It also adds the possibility for unexpected CRIU behaviour if the user
created a configuration file at some point in time and forgets about it.
As a result of the discussion in https://github.com/opencontainers/runc/pull/1933
it is now also possible to define a CRIU configuration file for each
container with the annotation 'org.criu.config'.
If 'org.criu.config' does not exist, runc will tell CRIU to use
'/etc/criu/runc.conf' if it exists.
If 'org.criu.config' is set to an empty string (''), runc will tell CRIU
to not use any runc specific configuration file at all.
If 'org.criu.config' is set to a non-empty string, runc will use that
value as an additional configuration file for CRIU.
With the annotation the user can decide to use the default configuration
file ('/etc/criu/runc.conf'), none or a container specific configuration
file.
Signed-off-by: Adrian Reber <areber@redhat.com>
The kernel will sometimes return EINVAL when writing a pid to a
cgroup.procs file. It does so when the task being added still has the
state TASK_NEW.
See: https://elixir.bootlin.com/linux/v4.8/source/kernel/sched/core.c#L8286
Co-authored-by: Danail Branekov <danailster@gmail.com>
Signed-off-by: Tom Godkin <tgodkin@pivotal.io>
Signed-off-by: Danail Branekov <danailster@gmail.com>
when restore container from a checkpoint directory, we should get
pid from criu notify, since c.initProcess has not been created.
Signed-off-by: Ace-Tang <aceapril@126.com>
When built with nokmem we explicitly are disabling support for kmemcg,
but it is a strict specification requirement that if we cannot fulfil an
aspect of the container configuration that we error out.
Completely ignoring explicitly-requested kmemcg limits with nokmem would
undoubtably lead to problems.
Fixes: 6a2c155968 ("libcontainer: ability to compile without kmem")
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Finish off the work started in a344b2d6 (sync up `HookState` with OCI
spec `State`, 2016-12-19, #1201).
And drop HookState, since there's no need for a local alias for
specs.State.
Also set c.initProcess in newInitProcess to support OCIState calls
from within initProcess.start(). I think the cyclic references
between linuxContainer and initProcess are unfortunate, but didn't
want to address that here.
I've also left the timing of the Prestart hooks alone, although the
spec calls for them to happen before start (not as part of creation)
[1,2]. Once the timing gets fixed we can drop the
initProcessStartTime hacks which initProcess.start currently needs.
I'm not sure why we trigger the prestart hooks in response to both
procReady and procHooks. But we've had two prestart rounds in
initProcess.start since 2f276498 (Move pre-start hooks after container
mounts, 2016-02-17, #568). I've left that alone too.
I really think we should have len() guards to avoid computing the
state when .Hooks is non-nil but the particular phase we're looking at
is empty. Aleksa, however, is adamantly against them [3] citing a
risk of sloppy copy/pastes causing the hook slice being len-guarded to
diverge from the hook slice being iterated over within the guard. I
think that ort of thing is very lo-risk, because:
* We shouldn't be copy/pasting this, right? DRY for the win :).
* There's only ever a few lines between the guard and the guarded
loop. That makes broken copy/pastes easy to catch in review.
* We should have test coverage for these. Guarding with the wrong
slice is certainly not the only thing you can break with a sloppy
copy/paste.
But I'm not a maintainer ;).
[1]: https://github.com/opencontainers/runtime-spec/blob/v1.0.0/config.md#prestart
[2]: https://github.com/opencontainers/runc/issues/1710
[3]: https://github.com/opencontainers/runc/pull/1741#discussion_r233331570
Signed-off-by: W. Trevor King <wking@tremily.us>
MBA Software Controller feature is introduced in Linux kernel v4.18.
It is a software enhancement to mitigate some limitations in MBA which
describes in kernel documentation. It also makes the interface more user
friendly - we could specify memory bandwidth in "MBps" (Mega Bytes per
second) as well as in "percentages".
The kernel underneath would use a software feedback mechanism or a
"Software Controller" which reads the actual bandwidth using MBM
counters and adjust the memory bandwidth percentages to ensure:
"actual memory bandwidth < user specified memory bandwidth".
We could enable this feature through mount option "-o mba_MBps":
mount -t resctrl resctrl -o mba_MBps /sys/fs/resctrl
In runc, we handle both memory bandwidth schemata in unified format:
"MB:<cache_id0>=bandwidth0;<cache_id1>=bandwidth1;..."
The unit of memory bandwidth is specified in "percentages" by default,
and in "MBps" if MBA Software Controller is enabled.
For more information about Intel RDT and MBA Software Controller:
https://www.kernel.org/doc/Documentation/x86/intel_rdt_ui.txt
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Commit fe898e7862 (PR #1350) enables kernel memory accounting
for all cgroups created by libcontainer -- even if kmem limit is
not configured.
Kernel memory accounting is known to be broken in some kernels,
specifically the ones from RHEL7 (including RHEL 7.5). Those
kernels do not support kernel memory reclaim, and are prone to
oopses. Unconditionally enabling kmem acct on such kernels lead
to bugs, such as
* https://github.com/opencontainers/runc/issues/1725
* https://github.com/kubernetes/kubernetes/issues/61937
* https://github.com/moby/moby/issues/29638
This commit gives a way to compile runc without kernel memory setting
support. To do so, use something like
make BUILDTAGS="seccomp nokmem"
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Cgroup namespace can be configured in `config.json` as other
namespaces. Here is an example:
```
"namespaces": [
{
"type": "pid"
},
{
"type": "network"
},
{
"type": "ipc"
},
{
"type": "uts"
},
{
"type": "mount"
},
{
"type": "cgroup"
}
],
```
Note that if you want to run a container which has shared cgroup ns with
another container, then it's strongly recommended that you set
proper `CgroupsPath` of both containers(the second container's cgroup
path must be the subdirectory of the first one). Or there might be
some unexpected results.
Signed-off-by: Yuanhong Peng <pengyuanhong@huawei.com>
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
This is a regression from 06f789cf26
when the user namespace was configured without a privileged helper.
To allow a single mapping in an user namespace, it is necessary to set
/proc/self/setgroups to "deny".
For a simple reproducer, the user namespace can be created with
"unshare -r".
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
Currently runc applies PidsLimit restriction by writing directly to
cgroup's pids.max, without notifying systemd. As a consequence, when the
later updates the context of the corresponding scope, pids.max is reset
to the value of systemd's TasksMax property.
This can be easily reproduced this way (I'm using "postfix" here just an
example, any unrelated but existing service will do):
# CTR=`docker run --pids-limit 111 --detach --rm busybox /bin/sleep 8h`
# cat /sys/fs/cgroup/pids/system.slice/docker-${CTR}.scope/pids.max
111
# systemctl disable --now postfix
# systemctl enable --now postfix
# cat /sys/fs/cgroup/pids/system.slice/docker-${CTR}.scope/pids.max
max
This patch adds TasksAccounting=true and TasksMax=PidsLimit to the
properties sent to systemd.
Signed-off-by: Sergio Lopez <slp@redhat.com>
This is a very simple implementation because it doesn't require any
configuration unlike the other namespaces, and in its current state it
only masks paths.
This feature is available in Linux 4.6+ and is enabled by default for
kernels compiled with CONFIG_CGROUP=y.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
Linux kernel v4.15 introduces better diagnostics for Intel RDT operation
errors. If any error returns when making new directories or writing to
any of the control file in resctrl filesystem, reading file
/sys/fs/resctrl/info/last_cmd_status could provide more information that
can be conveyed in the error returns from file operations.
Some examples:
echo "L3:0=f3;1=ff" > /sys/fs/resctrl/container_id/schemata
-bash: echo: write error: Invalid argument
cat /sys/fs/resctrl/info/last_cmd_status
mask f3 has non-consecutive 1-bits
echo "MB:0=0;1=110" > /sys/fs/resctrl/container_id/schemata
-bash: echo: write error: Invalid argument
cat /sys/fs/resctrl/info/last_cmd_status
MB value 0 out of range [10,100]
cd /sys/fs/resctrl
mkdir 1 2 3 4 5 6 7 8
mkdir: cannot create directory '8': No space left on device
cat /sys/fs/resctrl/info/last_cmd_status
out of CLOSIDs
See 'last_cmd_status' for more details in kernel documentation:
https://www.kernel.org/doc/Documentation/x86/intel_rdt_ui.txt
In runc, we could append the diagnostics information to the error
message of Intel RDT operation errors to provide more user-friendly
information.
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Double check if Intel RDT sub-features are available in "resource
control" filesystem. Intel RDT sub-features can be selectively disabled
or enabled by kernel command line (e.g., rdt=!l3cat,mba) in 4.14 and
newer kernel.
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Memory Bandwidth Allocation (MBA) is a resource allocation sub-feature
of Intel Resource Director Technology (RDT) which is supported on some
Intel Xeon platforms. Intel RDT/MBA provides indirect and approximate
throttle over memory bandwidth for the software. A user controls the
resource by indicating the percentage of maximum memory bandwidth.
Hardware details of Intel RDT/MBA can be found in section 17.18 of
Intel Software Developer Manual:
https://software.intel.com/en-us/articles/intel-sdm
In Linux 4.12 kernel and newer, Intel RDT/MBA is enabled by kernel
config CONFIG_INTEL_RDT. If hardware support, CPU flags `rdt_a` and
`mba` will be set in /proc/cpuinfo.
Intel RDT "resource control" filesystem hierarchy:
mount -t resctrl resctrl /sys/fs/resctrl
tree /sys/fs/resctrl
/sys/fs/resctrl/
|-- info
| |-- L3
| | |-- cbm_mask
| | |-- min_cbm_bits
| | |-- num_closids
| |-- MB
| |-- bandwidth_gran
| |-- delay_linear
| |-- min_bandwidth
| |-- num_closids
|-- ...
|-- schemata
|-- tasks
|-- <container_id>
|-- ...
|-- schemata
|-- tasks
For MBA support for `runc`, we will reuse the infrastructure and code
base of Intel RDT/CAT which implemented in #1279. We could also make
use of `tasks` and `schemata` configuration for memory bandwidth
resource constraints.
The file `tasks` has a list of tasks that belongs to this group (e.g.,
<container_id>" group). Tasks can be added to a group by writing the
task ID to the "tasks" file (which will automatically remove them from
the previous group to which they belonged). New tasks created by
fork(2) and clone(2) are added to the same group as their parent.
The file `schemata` has a list of all the resources available to this
group. Each resource (L3 cache, memory bandwidth) has its own line and
format.
Memory bandwidth schema:
It has allocation values for memory bandwidth on each socket, which
contains L3 cache id and memory bandwidth percentage.
Format: "MB:<cache_id0>=bandwidth0;<cache_id1>=bandwidth1;..."
The minimum bandwidth percentage value for each CPU model is predefined
and can be looked up through "info/MB/min_bandwidth". The bandwidth
granularity that is allocated is also dependent on the CPU model and
can be looked up at "info/MB/bandwidth_gran". The available bandwidth
control steps are: min_bw + N * bw_gran. Intermediate values are
rounded to the next control step available on the hardware.
For more information about Intel RDT kernel interface:
https://www.kernel.org/doc/Documentation/x86/intel_rdt_ui.txt
An example for runc:
Consider a two-socket machine with two L3 caches where the minimum
memory bandwidth of 10% with a memory bandwidth granularity of 10%.
Tasks inside the container may use a maximum memory bandwidth of 20%
on socket 0 and 70% on socket 1.
"linux": {
"intelRdt": {
"memBwSchema": "MB:0=20;1=70"
}
}
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
Respect the container's cgroup path when finding the container's
cgroup mount point, which is useful in multi-tenant environments, where
containers have their own unique cgroup mounts
Signed-off-by: Danail Branekov <danailster@gmail.com>
Signed-off-by: Oliver Stenbom <ostenbom@pivotal.io>
Signed-off-by: Giuseppe Capizzi <gcapizzi@pivotal.io>
Fix duplicate entries and missing entries in getCgroupMountsHelper
Add test for testing cgroup mounts on bedrock linux
Stop relying on number of subsystems for cgroups
LGTMs: @crosbymichael @cyphar
Closes#1817
While all modern kernels (and I do mean _all_ of them -- this syscall
was added in 2.6.10 before git had begun development!) have support for
this syscall, LXC has a default seccomp profile that returns ENOSYS for
this syscall. For most syscalls this would be a deal-breaker, and our
use of session keyrings is security-based there are a few mitigating
factors that make this change not-completely-insane:
* We already have a flag that disables the use of session keyrings
(for older kernels that had system-wide keyring limits and so
on). So disabling it is not a new idea.
* While the primary justification of using session keys *is*
security-based, it's more of a security-by-obscurity protection.
The main defense keyrings have is VFS credentials -- which is
something that users already have better security tools for
(setuid(2) and user namespaces).
* Given the security justification you might argue that we
shouldn't silently ignore this. However, the only way for the
kernel to return -ENOSYS is either being ridiculously old (at
which point we wouldn't work anyway) or that there is a seccomp
profile in place blocking it.
Given that the seccomp profile (if malicious) could very easily
just return 0 or a silly return code (or something even more
clever with seccomp-bpf) and trick us without this patch, there
isn't much of a significant change in how much seccomp can trick
us with or without this patch.
Given all of that over-analysis, I'm pretty convinced there isn't a
security problem in this very specific case and it will help out the
ChromeOS folks by allowing Docker to run inside their LXC container
setup. I'd be happy to be proven wrong.
Ref: https://bugs.chromium.org/p/chromium/issues/detail?id=860565
Signed-off-by: Aleksa Sarai <asarai@suse.de>
This PR decomposes `libcontainer/configs.Config.Rootless bool` into `RootlessEUID bool` and
`RootlessCgroups bool`, so as to make "runc-in-userns" to be more compatible with "rootful" runc.
`RootlessEUID` denotes that runc is being executed as a non-root user (euid != 0) in
the current user namespace. `RootlessEUID` is almost identical to the former `Rootless`
except cgroups stuff.
`RootlessCgroups` denotes that runc is unlikely to have the full access to cgroups.
`RootlessCgroups` is set to false if runc is executed as the root (euid == 0) in the initial namespace.
Otherwise `RootlessCgroups` is set to true.
(Hint: if `RootlessEUID` is true, `RootlessCgroups` becomes true as well)
When runc is executed as the root (euid == 0) in an user namespace (e.g. by Docker-in-LXD, Podman, Usernetes),
`RootlessEUID` is set to false but `RootlessCgroups` is set to true.
So, "runc-in-userns" behaves almost same as "rootful" runc except that cgroups errors are ignored.
This PR does not have any impact on CLI flags and `state.json`.
Note about CLI:
* Now `runc --rootless=(auto|true|false)` CLI flag is only used for setting `RootlessCgroups`.
* Now `runc spec --rootless` is only required when `RootlessEUID` is set to true.
For runc-in-userns, `runc spec` without `--rootless` should work, when sufficient numbers of
UID/GID are mapped.
Note about `$XDG_RUNTIME_DIR` (e.g. `/run/user/1000`):
* `$XDG_RUNTIME_DIR` is ignored if runc is being executed as the root (euid == 0) in the initial namespace, for backward compatibility.
(`/run/runc` is used)
* If runc is executed as the root (euid == 0) in an user namespace, `$XDG_RUNTIME_DIR` is honored if `$USER != "" && $USER != "root"`.
This allows unprivileged users to allow execute runc as the root in userns, without mounting writable `/run/runc`.
Note about `state.json`:
* `rootless` is set to true when `RootlessEUID == true && RootlessCgroups == true`.
Signed-off-by: Akihiro Suda <suda.akihiro@lab.ntt.co.jp>
it is now allowed to bind mount /proc. This is useful for rootless
containers when the PID namespace is shared with the host.
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
subgid is defined per user, not group (see subgid(5))
This commit also adds support for specifying subuid owner with a numeric UID.
Signed-off-by: Akihiro Suda <suda.akihiro@lab.ntt.co.jp>
Using CRIU to checkpoint and restore a container into an existing
network namespace is not possible.
If the network namespace is defined like
{
"type": "network",
"path": "/run/netns/test"
}
there is the expectation that the restored container is again running in
the network namespace specified with 'path'.
This adds the new CRIU 'external namespace' feature to runc, where
during checkpointing that specific namespace is referenced and during
restore CRIU tries to restore the container in exactly that
namespace.
This breaks/fixes current runc behavior. If, without this patch, runc
restores a container with such a network namespace definition, it is
ignored and CRIU recreates a network namespace without a name.
With this patch runc uses the network namespace path (if available) to
checkpoint and restore the container in just that network namespace.
Restore will now fail if a container was checkpointed with a network
namespace path set and if that network namespace path does not exist
during restore.
runc still falls back to the old behavior if CRIU older than 3.11 is
installed.
Fixes#1786
Related to https://github.com/projectatomic/libpod/pull/469
Thanks to Andrei Vagin for all the help in getting the interface between
CRIU and runc right!
Signed-off-by: Adrian Reber <areber@redhat.com>
MOVE_MOUNT will fail under certain situations.
You are not allowed to MS_MOVE if the parent directory is shared.
man mount
...
The move operation
Move a mounted tree to another place (atomically). The call is:
mount --move olddir newdir
This will cause the contents which previously appeared under olddir to
now be accessible under newdir. The physical location of the files is
not changed. Note that olddir has to be a mountpoint.
Note also that moving a mount residing under a shared mount is invalid
and unsupported. Use findmnt -o TARGET,PROPAGATION to see the current
propagation flags.
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
This will help runc's init to not spawn many threads on large systems when
launched with max procs by the caller.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
Add a mountinfo from a bedrock linux system with 4 strata, and include
it for tests
Signed-off-by: Jay Kamat <jaygkamat@gmail.com>
Signed-off-by: Daniel Dao <dqminh89@gmail.com>
When there are complicated mount setups, there can be multiple mount
points which have the subsystem we are looking for. Instead of
counting the mountpoints, tick off subsystems until we have found them
all.
Without the 'all' flag, ignore duplicate subsystems after the first.
Signed-off-by: Daniel Dao <dqminh89@gmail.com>
These sysctls are namespaced by CLONE_NEWUTS, and we need to use
"kernel.domainname" if we want users to be able to set an NIS domainname
on Linux. However we disallow "kernel.hostname" because it would
conflict with the "hostname" field and cause confusion (but we include a
helpful message to make it clearer to the user).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
It turns out that MIPS uses uint32 in the device number returned by
stat(2), so explicitly wrap everything to make the compiler happy. I
really wish that Go had C-like numeric type promotion.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
This fixes the following compilation error on 32bit ARM:
```
$ GOARCH=arm GOARCH=6 go build ./libcontainer/system/
libcontainer/system/linux.go:119:89: constant 4294967295 overflows int
```
Signed-off-by: Tibor Vass <tibor@docker.com>
When running in a new unserNS as root, don't require a mapping to be
present in the configuration file. We are already skipping the test
for a new userns to be present.
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
We need to lock the threads for the SetProcessLabel to work,
should also call SetProcessLabel("") after the container starts
to go back to the default SELinux behaviour.
Once you call SetProcessLabel, then any process executed by runc
will run with this label, even if the process is for setup rather
then the container.
It is always safest to call the SELinux calls just before the exec of the
container, so that other processes do not get started with the incorrect label.
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
There is a race in runc exec when the init process stops just before
the check for the container status. It is then wrongly assumed that
we are trying to start an init process instead of an exec process.
This commit add an Init field to libcontainer Process to distinguish
between init and exec processes to prevent this race.
Signed-off-by: Mrunal Patel <mrunalp@gmail.com>
Include a rootless argument for isIgnorableError to avoid people
accidentally using isIgnorableError when they shouldn't (we don't ignore
any errors when running as root as that really isn't safe).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
So that, if a timeout happens and we decide to stop blocking on the
operation, the writer will not block when they try to report the result
of the operation.
This should address Issue #1780 and it's a follow up for PR #1683,
PR #1754 and PR #1772.
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
When joining an existing namespace, don't default to configuring a
loopback interface in that namespace.
Its creator should have done that, and we don't want to fail to create
the container when we don't have sufficient privileges to configure the
network namespace.
Signed-off-by: Nalin Dahyabhai <nalin@redhat.com>
Starting with systemd 237, in preparation for cgroup v2, delegation is
only now available for scopes, not slices.
Update libcontainer code to detect whether delegation is available on
both and use that information when creating new slices.
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
The channel was introduced in #1683 to work around a race condition.
However, the check for error in StartTransientUnit ignores the error for
an already existing unit, and in that case there will be no notification
from DBus (so waiting on the channel will make it hang.)
Later PR #1754 added a timeout, which worked around the issue, but we
can fix this correctly by only waiting on the channel when there is no
error. Fix the code to do so.
The timeout handling was kept, since there might be other cases where
this situation occurs (https://bugzilla.redhat.com/show_bug.cgi?id=1548358
mentions calling this code from inside a container, it's unclear whether
an existing container was in use or not, so not sure whether this would
have fixed that bug as well.)
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
There is no reason to set the container state to "running" as a
temporary value when exec'ing a process on a container in "created"
state. The problem doing this is that consumers of the libcontainer
library might use it by keeping pointers in memory. In this case,
the container state will indicate that the container is running, which
is wrong, and this will end up with a failure on the next action
because the check for the container state transition will complain.
Fixes#1767
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Previously if oomScoreAdj was not set in config.json we would implicitly
set oom_score_adj to 0. This is not allowed according to the spec:
> If oomScoreAdj is not set, the runtime MUST NOT change the value of
> oom_score_adj.
Change this so that we do not modify oom_score_adj if oomScoreAdj is not
present in the configuration. While this modifies our internal
configuration types, the on-disk format is still compatible.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
In some cases, /sys/fs/cgroups is mounted read-only. In rootless
containers we can consider this effectively identical to having cgroups
that we don't have write permission to -- because the user isn't
responsible for the read-only setup and cannot modify it. The rules are
identical to when /sys/fs/cgroups is not writable by the unprivileged
user.
An example of this is the default configuration of Docker, where cgroups
are mounted as read-only as a preventative security measure.
Reported-by: Vladimir Rutsky <rutsky@google.com>
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Currently if a confined container process tries to list these directories
AVC's are generated because they are labeled with external labels. Adding
the mountlabel will remove these AVC's.
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
Currently Manager accepts nil cgroups when calling Apply, but it will panic then trying to call Destroy with the same config.
Signed-off-by: Denys Smirnov <denys@sourced.tech>
The function is called even if the usernamespace is not set.
This results having wrong uid/gid set on devices.
This fix add a test to check if usernamespace is set befor calling
setupUserNamespace.
Fixes#1742
Signed-off-by: Julien Lavesque <julien.lavesque@gmail.com>
These were added in 244c9fc4 (*: console rewrite, 2016-06-04, #1018)
alongside procConsole and the associated handling. procConsole and
that handling were removed in 00a0ecf5 (Add separate console socket,
2017-03-02, #1356), but 00a0ecf5 missed this comment.
Signed-off-by: W. Trevor King <wking@tremily.us>
gocapability has supported 0 as "the current PID" since
syndtr/gocapability@5e7cce49 (Allow to use the zero value for pid to
operate with the current task, 2015-01-15, syndtr/gocapability#2).
libcontainer was ported to that approach in 444cc298 (namespaces:
allow to use pid namespace without mount namespace, 2015-01-27,
docker/libcontainer#358), but the change was clobbered by 22df5551
(Merge branch 'master' into api, 2015-02-19, docker/libcontainer#388)
which landed via 5b73860e (Merge pull request #388 from docker/api,
2015-02-19, docker/libcontainer#388). This commit restores the
changes from 444cc298.
Signed-off-by: W. Trevor King <wking@tremily.us>
The helper DRYs up the transition tests and makes it easy to get
complete coverage for invalid transitions.
I'm also using t.Run() for subtests. Run() is new in Go 1.7 [1], but
runc dropped support for 1.6 back in e773f96b (update go version at
travis-ci, 2017-02-20, #1335).
[1]: https://blog.golang.org/subtests
Signed-off-by: W. Trevor King <wking@tremily.us>
Technically, this change should not be necessary, as the kernel
documentation claims that if you call clone(flags|CLONE_NEWUSER), the
new user namespace will be the owner of all other namespaces created in
@flags. Unfortunately this isn't always the case, due to various
additional semantics and kernel bugs.
One particular instance is SELinux, which acts very strangely towards
the IPC namespace and mqueue. If you unshare the IPC namespace *before*
you map a user in the user namespace, the IPC namespace's internal
kern-mount for mqueue will be labelled incorrectly and the container
won't be able to access it. The only way of solving this is to unshare
IPC *after* the user has been mapped and we have changed to that user.
I've also heard of this happening to the NET namespace while talking to
some LXC folks, though I haven't personally seen that issue.
This change matches our handling of user namespaces to be the same as
how LXC handles these problems.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
When starting a container with `runc start` or `runc run`, the stub
process (runc[2:INIT]) opens a fifo for writing. Its parent runc process
will open the same fifo for reading. In this way, they synchronize.
If the stub process exits at the wrong time, the parent runc process
will block forever.
This can happen when racing 2 runc operations against each other: `runc
run/start`, and `runc delete`. It could also happen for other reasons,
e.g. the kernel's OOM killer may select the stub process.
This commit resolves this race by racing the opening of the exec fifo
from the runc parent process against the stub process exiting. If the
stub process exits before we open the fifo, we return an error.
Another solution is to wait on the stub process. However, it seems it
would require more refactoring to avoid calling wait multiple times on
the same process, which is an error.
Signed-off-by: Craig Furman <cfurman@pivotal.io>
Annotations weren't passed to hooks. This patch fixes that by passing
annotations to stdin for hooks.
Signed-off-by: Antonio Murdaca <runcom@redhat.com>
libapparmor is integrated in libcontainer using cgo but is only used to
call a single function: aa_change_onexec. It turns out this function is
simple enough (writing a string to a file in /proc/<n>/attr/...) to be
re-implemented locally in libcontainer in plain Go.
This allows to drop the dependency on libapparmor and the corresponding
cgo integration.
Fixes#1674
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
When a subreaper is enabled, it might expect to reap a process and
retrieve its exit code. That's the reason why this patch is giving
the possibility to define the usage of a subreaper as a consumer of
libcontainer. Relying on this information, libcontainer will not
wait for signalled processes in case a subreaper has been set.
Fixes#1677
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
TestGetContainerStateAfterUpdate creates its state.json file on the current
directory which turns out to be the host runc directory. Thus whenever
the test completes it leaves the state.json file behind thus
a) poluting the local git repository
b) changing the host file system violating the principle of doing
everything in an isolated container environment
This change would create a new temporary (in-container) directory and use it as
linuxContainer.root
Signed-off-by: Tom Godkin <tgodkin@pivotal.io>
runc currently only support Linux platform, and since we dont intend to expose
the support to other platform, removing all other platforms placeholder code.
`libcontainer/configs` still being used in
https://github.com/moby/moby/blob/master/daemon/daemon_windows.go so
keeping it for now.
After this, we probably should also rename files to drop linux suffices
if possible.
Signed-off-by: Daniel Dao <dqminh89@gmail.com>
runc is not supported on FreeBSD, so remove all FreeBSD specific bits.
As suggested by @crosbymichael in #1653
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Selinux related code has been moved to the selinux package
(https://github.com/opencontainers/selinux) and therefore xattr related
code can be deleted from libcontainer
Signed-off-by: Danail Branekov <danailster@gmail.com>
When running runc tests with temp directory with size 500M copying
busybox without preserving hardlinks causes the folder to inflate to
roughly 330M. Copying busybox twice in certain tests causes the /tmp
directory to overfill. Using `-a` preserves links which busybox uses to
implement its choice of binary to run.
Signed-off-by: Tom Godkin <tgodkin@pivotal.io>
runc shouldn't depend on docker and be more self-contained.
Removing github.com/pkg/symlink dep is the first step to not depend on docker anymore
Signed-off-by: Vincent Demeester <vincent@sbr.pm>
Previously we would handle the "unmapped stdio" case by just doing a
simple check, however this didn't handle cases where the overflow_uid
was actually mapped in the user namespace. Instead of doing some
userspace checks, just try to do the fchown(2) and ignore EINVAL
(unmapped) or EPERM (lacking privilege over inode) errors.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Multiple conditions were previously allowed to be placed upon the
same syscall argument. Restore this behavior.
Signed-off-by: Matthew Heon <mheon@redhat.com>
This fixes a GetStats() issue introduced in #1590:
If Intel RDT is enabled by hardware and kernel, but intelRdt is not
specified in original config, GetStats() will return error unexpectedly
because we haven't called Apply() to create intelrdt group or attach
tasks for this container. As a result, runc events command will have no
output.
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
The Major and Minor functions were added for Linux in golang/sys@85d1495
which is already vendored in. Use these functions instead of the local
re-implementation.
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
This allows runc to be used as a target for docker's reexec module that
depends on a correct argv0 to select which process entrypoint to invoke.
Without this patch, when runc re-execs argv0 is set to "/proc/self/exe"
and the reexec module doesn't know what to do with it.
Signed-off-by: Petros Angelatos <petrosagg@gmail.com>
There are essentially two possible implementations for Setuid/Setgid on
Linux, either using SYS_SETUID32/SYS_SETGID32 or SYS_SETUID/SYS_SETGID,
depending on the architecture (see golang/go#1435 for why Setuid/Setgid
aren currently implemented for Linux neither in syscall nor in
golang.org/x/sys/unix).
Reduce duplication by merging the currently implemented variants and
adjusting the build tags accordingly.
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
This commit ensures we write the expected freezer cgroup state after
every state check, in case the state check does not give the expected
result. This can happen when a new task is created and prevents the
whole cgroup to be FROZEN, leaving the state into FREEZING instead.
This patch prevents the case of an infinite loop to happen.
Fixes https://github.com/opencontainers/runc/issues/1609
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Both Process.Kill() and Process.Wait() can return errors that don't impact the correct behaviour of terminate.
Instead of letting these get returned and logged, which causes confusion, silently ignore them.
Currently the test needs to be a string test as the errors are private to the runtime packages, so its our only option.
This can be seen if init fails during the setns.
Signed-off-by: Steven Hartland <steven.hartland@multiplay.co.uk>
Due to the semantics of chroot(2) when it comes to mount namespaces, it
is not generally safe to use MS_PRIVATE as a mount propgation when using
chroot(2). The reason for this is that this effectively results in a set
of mount references being held by the chroot'd namespace which the
namespace cannot free. pivot_root(2) does not have this issue because
the @old_root can be unmounted by the process.
Ultimately, --no-pivot is not really necessary anymore as a commonly
used option since f8e6b5af5e ("rootfs: make pivot_root not use a
temporary directory") resolved the read-only issue. But if someone
really needs to use it, MS_PRIVATE is never a good idea.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Signed-off-by: Ed King <eking@pivotal.io>
Signed-off-by: Gabriel Rosenhouse <grosenhouse@pivotal.io>
Signed-off-by: Konstantinos Karampogias <konstantinos.karampogias@swisscom.com>
The benefit for doing this within runc is that it works well with
userns.
Actually, runc already does the same thing for mount points.
Signed-off-by: Akihiro Suda <suda.akihiro@lab.ntt.co.jp>
Signed-off-by: Will Martin <wmartin@pivotal.io>
Signed-off-by: Petar Petrov <pppepito86@gmail.com>
Signed-off-by: Ed King <eking@pivotal.io>
Signed-off-by: Roberto Jimenez Sanchez <jszroberto@gmail.com>
Signed-off-by: Thomas Godkin <tgodkin@pivotal.io>
The code in prepareRoot (e385f67a0e/libcontainer/rootfs_linux.go (L599-L605))
attempts to default the rootfs mount to `rslave`. However, since the spec
conversion has already defaulted it to `rprivate`, that code doesn't
actually ever do anything.
This changes the spec conversion code to accept "" and treat it as 0.
Implicitly, this makes rootfs propagation default to `rslave`, which is
a part of fixing the moby bug https://github.com/moby/moby/issues/34672
Alternate implementatoins include changing this defaulting to be
`rslave` and removing the defaulting code in prepareRoot, or skipping
the mapping entirely for "", but I think this change is the cleanest of
those options.
Signed-off-by: Euan Kemp <euan.kemp@coreos.com>
In current implementation:
Either Intel RDT is not enabled by hardware and kernel, or intelRdt is
not specified in original config, we don't init IntelRdtManager in the
container to handle intelrdt constraint. It is a tradeoff that Intel RDT
has hardware limitation to support only limited number of groups.
This patch makes a minor change to support update command:
Whether or not intelRdt is specified in config, we always init
IntelRdtManager in the container if Intel RDT is enabled. If intelRdt is
not specified in original config, we just don't Apply() to create
intelrdt group or attach tasks for this container.
In update command, we could re-enable through IntelRdtManager.Apply()
and then update intelrdt constraint.
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
This fix tries to address the warnings caused by static build
with go 1.9. As systemd needs dlopen/dlclose, the following warnings
will be generated for static build in go 1.9:
```
root@f4b077232050:/go/src/github.com/opencontainers/runc# make static
CGO_ENABLED=1 go build -tags "seccomp cgo static_build" -ldflags "-w -extldflags -static -X main.gitCommit="1c81e2a794c6e26a4c650142ae8893c47f619764" -X main.version=1.0.0-rc4+dev " -o runc .
/tmp/go-link-113476657/000007.o: In function `_cgo_a5acef59ed3f_Cfunc_dlopen':
/tmp/go-build/github.com/opencontainers/runc/vendor/github.com/coreos/pkg/dlopen/_obj/cgo-gcc-prolog:76: warning: Using 'dlopen' in statically linked applications requires at runtime the shared libraries from the glibc version used for linking
```
This fix disables systemd when `static_build` flag is on (apply_nosystemd.go
is used instead).
This fix also fixes a small bug in `apply_nosystemd.go` for return value.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
Now that rootless containers have support for multiple uid and gid
mappings, allow --user to work as expected. If the user is not mapped,
an error occurs (as usual).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
With the addition of our new{uid,gid}map support, we used to call
execvp(3) from inside nsexec. This would mean that the path resolution
for the binaries would happen in nsexec. Move the resolution to the
initial setup code, and pass the absolute path to nsexec.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
After quite a bit of debugging, I found that previous versions of this
patchset did not include newgidmap in a rootless setting. Fix this by
passing it whenever group mappings are applied, and also providing some
better checking for try_mapping_tool. This commit also includes some
stylistic improvements.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Take advantage of the newuidmap/newgidmap tools to allow multiple
users/groups to be mapped into the new user namespace in the rootless
case.
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
[ rebased to handle intelrdt changes. ]
Signed-off-by: Aleksa Sarai <asarai@suse.de>
This is the follow-up PR of #1279 to fix remaining issues:
Use init() to avoid race condition in IsIntelRdtEnabled().
Add also rename some variables and functions.
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
This applies cgroups earlier for container creation before the init
process starts running and forking off any additional processes.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
With the help of userfaultfd CRIU supports lazy migration. Lazy
migration means that memory pages are only transferred from the
migration source to the migration destination on page fault.
This enables to reduce the downtime during process or container
migration to a minimum as the memory does not need to be transferred
during migration.
Lazy migration currently depends on userfaultfd being available on the
current Linux kernel and if the used CRIU version supports lazy
migration. Both dependencies can be checked by querying CRIU via RPC if
the lazy migration feature is available. Using feature checking instead
of version comparison enables runC to use CRIU features from the
criu-dev branch. This way the user can decide if lazy migration should
be available by choosing the right kernel and CRIU branch.
To use lazy migration the CRIU process during dump needs to dump
everything besides the memory pages and then it opens a network port
waiting for remote page fault requests:
# runc checkpoint httpd --lazy-pages --page-server 0.0.0.0:27 \
--status-fd /tmp/postcopy-pipe
In this example CRIU will hang/wait once it has opened the network port
and wait for network connection. As runC waits for CRIU to finish it
will also hang until the lazy migration has finished. To know when the
restore on the destination side can start the '--status-fd' parameter is
used:
#️ runc checkpoint --help | grep status
--status-fd value criu writes \0 to this FD once lazy-pages is ready
The parameter '--status-fd' is directly from CRIU and this way the
process outside of runC which controls the migration knows exactly when
to transfer the checkpoint (without memory pages) to the destination and
that the restore can be started.
On the destination side it is necessary to start CRIU in 'lazy-pages'
mode like this:
# criu lazy-pages --page-server --address 192.168.122.3 --port 27 \
-D checkpoint
and tell runC to do a lazy restore:
# runc restore -d --image-path checkpoint --work-path checkpoint \
--lazy-pages httpd
If both processes on the restore side have the same working directory
'criu lazy-pages' creates a unix domain socket where it waits for
requests from the actual restore. runC starts CRIU restore in lazy
restore mode and talks to 'criu lazy-pages' that it wants to restore
memory pages on demand. CRIU continues to restore the process and once
the process is running and accesses the first non-existing memory page
the 'criu lazy-pages' server will request the page from the source
system. Thus all pages from the source system will be transferred to the
destination system. Once all pages have been transferred runC on the
source system will end and the container will have finished migration.
This can also be combined with CRIU's pre-copy support. The combination
of pre-copy and post-copy (lazy migration) provides the possibility to
migrate containers with minimal downtimes.
Some additional background about post-copy migration can be found in
these articles:
https://lisas.de/~adrian/?p=1253https://lisas.de/~adrian/?p=1183
Signed-off-by: Adrian Reber <areber@redhat.com>
Before adding the actual lazy migration support, this adds the feature
check for lazy-pages. Right now lazy migration, which is based on
userfaultd is only available in the criu-dev branch and not yet in a
release. As the check does not dependent on a certain version but on
a CRIU feature which can be queried it can be part of runC without a new
version check depending on a feature from criu-dev.
Signed-off-by: Adrian Reber <areber@redhat.com>
About Intel RDT/CAT feature:
Intel platforms with new Xeon CPU support Intel Resource Director Technology
(RDT). Cache Allocation Technology (CAT) is a sub-feature of RDT, which
currently supports L3 cache resource allocation.
This feature provides a way for the software to restrict cache allocation to a
defined 'subset' of L3 cache which may be overlapping with other 'subsets'.
The different subsets are identified by class of service (CLOS) and each CLOS
has a capacity bitmask (CBM).
For more information about Intel RDT/CAT can be found in the section 17.17
of Intel Software Developer Manual.
About Intel RDT/CAT kernel interface:
In Linux 4.10 kernel or newer, the interface is defined and exposed via
"resource control" filesystem, which is a "cgroup-like" interface.
Comparing with cgroups, it has similar process management lifecycle and
interfaces in a container. But unlike cgroups' hierarchy, it has single level
filesystem layout.
Intel RDT "resource control" filesystem hierarchy:
mount -t resctrl resctrl /sys/fs/resctrl
tree /sys/fs/resctrl
/sys/fs/resctrl/
|-- info
| |-- L3
| |-- cbm_mask
| |-- min_cbm_bits
| |-- num_closids
|-- cpus
|-- schemata
|-- tasks
|-- <container_id>
|-- cpus
|-- schemata
|-- tasks
For runc, we can make use of `tasks` and `schemata` configuration for L3 cache
resource constraints.
The file `tasks` has a list of tasks that belongs to this group (e.g.,
<container_id>" group). Tasks can be added to a group by writing the task ID
to the "tasks" file (which will automatically remove them from the previous
group to which they belonged). New tasks created by fork(2) and clone(2) are
added to the same group as their parent. If a pid is not in any sub group, it
Is in root group.
The file `schemata` has allocation bitmasks/values for L3 cache on each socket,
which contains L3 cache id and capacity bitmask (CBM).
Format: "L3:<cache_id0>=<cbm0>;<cache_id1>=<cbm1>;..."
For example, on a two-socket machine, L3's schema line could be `L3:0=ff;1=c0`
which means L3 cache id 0's CBM is 0xff, and L3 cache id 1's CBM is 0xc0.
The valid L3 cache CBM is a *contiguous bits set* and number of bits that can
be set is less than the max bit. The max bits in the CBM is varied among
supported Intel Xeon platforms. In Intel RDT "resource control" filesystem
layout, the CBM in a group should be a subset of the CBM in root. Kernel will
check if it is valid when writing. e.g., 0xfffff in root indicates the max bits
of CBM is 20 bits, which mapping to entire L3 cache capacity. Some valid CBM
values to set in a group: 0xf, 0xf0, 0x3ff, 0x1f00 and etc.
For more information about Intel RDT/CAT kernel interface:
https://www.kernel.org/doc/Documentation/x86/intel_rdt_ui.txt
An example for runc:
Consider a two-socket machine with two L3 caches where the default CBM is
0xfffff and the max CBM length is 20 bits. With this configuration, tasks
inside the container only have access to the "upper" 80% of L3 cache id 0 and
the "lower" 50% L3 cache id 1:
"linux": {
"intelRdt": {
"l3CacheSchema": "L3:0=ffff0;1=3ff"
}
}
Signed-off-by: Xiaochen Shen <xiaochen.shen@intel.com>
This mirrors the standard_init_linux.go seccomp code, which only applies
seccomp early if NoNewPrivileges is enabled. Otherwise it's done
immediately before execve to reduce the amount of syscalls necessary for
users to enable in their seccomp profiles.
Signed-off-by: Aleksa Sarai <asarai@suse.de>
Fixes: #1557
I'm not quite sure about the root cause, looks like
systemd still want them to be uint64.
Signed-off-by: Qiang Huang <h.huangqiang@huawei.com>
While we have significant protections in place against CVE-2016-9962, we
still were holding onto a file descriptor that referenced the host
filesystem. This meant that in certain scenarios it was still possible
for a semi-privileged container to gain access to the host filesystem
(if they had CAP_SYS_PTRACE).
Instead, open the FIFO itself using a O_PATH. This allows us to
reference the FIFO directly without providing the ability for
directory-level access. When opening the FIFO inside the init process,
open it through procfs to re-open the actual FIFO (this is currently the
only supported way to open such a file descriptor).
Signed-off-by: Aleksa Sarai <asarai@suse.de>
The documentation here:
https://docs.docker.com/engine/security/userns-remap/#user-namespace-known-limitations
says that readonly containers can't be used with user namespaces do to some
kernel restriction. In fact, there is a special case in the kernel to be
able to do stuff like this, so let's use it.
This takes us from:
ubuntu@docker:~$ docker run -it --read-only ubuntu
docker: Error response from daemon: oci runtime error: container_linux.go:262: starting container process caused "process_linux.go:339: container init caused \"rootfs_linux.go:125: remounting \\\"/dev\\\" as readonly caused \\\"operation not permitted\\\"\"".
to:
ubuntu@docker:~$ docker-runc --version
runc version 1.0.0-rc4+dev
commit: ae2948042b08ad3d6d13cd09f40a50ffff4fc688-dirty
spec: 1.0.0
ubuntu@docker:~$ docker run -it --read-only ubuntu
root@181e2acb909a:/# touch foo
touch: cannot touch 'foo': Read-only file system
Signed-off-by: Tycho Andersen <tycho@docker.com>
congfig.Sysctl setting is duplicated.
when contianer is rootless and Linux is nil, runc will panic.
Signed-off-by: Ma Shimiao <mashimiao.fnst@cn.fujitsu.com>
state.json should be a reflection of the container's
realtime state, including resource configurations,
so we should update state.json after updating container
resources.
Signed-off-by: Qiang Huang <h.huangqiang@huawei.com>
Fixes: #1228
It can be reproduced by applying this patch:
```diff
@@ -45,6 +46,7 @@ func registerMemoryEvent(cgDir string, evName string, arg string) (<-chan struct
go func() {
defer func() {
close(ch)
+ <-time.After(1 * time.Second)
eventfd.Close()
evFile.Close()
}()
```
We can close channel after fds were closed.
Signed-off-by: Qiang Huang <h.huangqiang@huawei.com>
Linux is not always not nil.
If Linux is nil, panic will occur.
Signed-off-by: Ma Shimiao <mashimiao.fnst@cn.fujitsu.com>
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
Before this change, some file type would be treated as char devices
(e.g. symlinks).
Signed-off-by: Kenfe-Mickael Laventure <mickael.laventure@gmail.com>
This allows the libcontainer to automatically clean up runc:[1:CHILD]
processes created as part of nsenter.
Signed-off-by: Alex Fang <littlelightlittlefire@gmail.com>
With this runC also uses RPC to ask CRIU for its version. CRIU supports
a VERSION RPC since CRIU 3.0 and using the RPC interface does not
require parsing the console output of CRIU (which could change anytime).
For older CRIU versions which do not yet have the VERSION RPC runC falls
back to its old CRIU output parsing mode.
Once CRIU 3.0 is the minimum version required for runC the old code can
be removed.
v2:
* adapt to changes in the previous patches based on the review
Signed-off-by: Adrian Reber <areber@redhat.com>
Update criurpc.proto for the upcoming VERSION RPC.
This includes lazy_pages for the upcoming lazy migration support.
Signed-off-by: Adrian Reber <areber@redhat.com>
To use the CRIU VERSION RPC the criuSwrk function is adapted to work
with CriuOpts set to 'nil' as CriuOpts is not required for the VERSION
RPC.
Also do not print c.criuVersion if it is '0' as the first RPC call will
always be the VERSION call and only after that the version will be
known.
Signed-off-by: Adrian Reber <areber@redhat.com>
If the version of criu has already been determined there is no need to
ask criu for the version again. Use the value from c.criuVersion.
v2:
* reduce unnecessary code movement in the patch series
* factor out the criu version parsing into a separate function
Signed-off-by: Adrian Reber <areber@redhat.com>
The checkCriuVersion function used a string to specify the minimum
version required. This is more comfortable for an external interface
but for an internal function this added unnecessary complexity. This
changes to version string like '1.5.2' to an integer like 10502. This is
already the format used internally in the function.
Signed-off-by: Adrian Reber <areber@redhat.com>
go's switch statement doesn't need an explicit break. Remove it where
that is the case and add a comment to indicate the purpose where the
removal would lead to an empty case.
Found with honnef.co/go/tools/cmd/staticcheck
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Handle err return value of fmt.Scanf, os.Pipe and unix.ParseUnixRights.
Found with honnef.co/go/tools/cmd/staticcheck
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
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>
Kenta Tada
Mrunal Patel
Filipe Brandenburger
Georgi Sabev
Xiaochen Shen
Aleksa Sarai
Danail Branekov
Marco Vedovati
Michael Crosby
Alex Fang
Filipe Brandenburger
Justin Cormack
Daniel J Walsh
Adrian Reber
Christian Brauner
Giuseppe Scrivano
JoeWrightss
Tom Godkin
Ace-Tang
Lifubang
W. Trevor King
Kir Kolyshkin
Yuanhong Peng
Chris Aniszczyk
Sergio Lopez
Dominik Süß
Jonathan Marler
Mike Brown
Akihiro Suda
Yan Zhu
ChangFeng
Qiang Huang
Mrunal Patel
Jay Kamat
Alban Crequy
Daniel Dao
Tibor Vass
Derek Carr
dlorenc
Bin Chen
Antonio Murdaca
Nalin Dahyabhai
Sebastien Boeuf
Tamal Saha
vikaschoudhary16
Denys Smirnov
Vincent Batts
ynirk
Allen Sun
ravisantoshgudimetla
Ed King
Will Martin
Tobias Klauser
Allen Sun
Ma Shimiao
Seth Jennings
Vincent Demeester
Danail Branekov
Lorenzo Fontana
Matthew Heon
Petros Angelatos
Steven Hartland
Konstantinos Karampogias
Euan Kemp
s7v7nislands
Yong Tang
Tycho Andersen
Nikolas Sepos
Kenfe-Mickael Laventure
yangshukui