Before this patch, both the RdtEnabled and BlockIOEnabled are provided
by services/tasks pkg. Since the services/tasks can be pkg plugin which
can be initialized multiple times or concurrently. It will fire data-race
issue as there is no mutex to protect `enable`.
This patch is aimed to provide wrapper pkgs to use intel/{blockio,rdt}
safely.
Signed-off-by: Wei Fu <fuweid89@gmail.com>
/etc/cni has to be readable for non-root users (0755), because /etc/cni/tuning/allowlist.conf is used for rootless mode too.
This file was introduced in CNI plugins 1.2.0 (containernetworking/plugins PR 693), and its path is hard-coded.
Signed-off-by: Akihiro Suda <akihiro.suda.cz@hco.ntt.co.jp>
Some of this code was originally added in b7b1200dd3,
which likely meant to initialize the slice with a length to reduce allocations,
however, instead of initializing with a zero-length and a capacity, it
initialized the slice with a fixed length, which was corrected in commit
0c63c42f81.
This patch initializes the slice with a zero-length and expected capacity.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Comments in initPlatform for Windows states that the options were
Linux specific. Additionally properly wrap an error after trying
to setup CDI on Linux.
Signed-off-by: Danny Canter <danny@dcantah.dev>
The sandbox and container both have the userns config. Lets make sure
they are the same, therefore consistent.
Signed-off-by: Rodrigo Campos <rodrigoca@microsoft.com>
Currently we require that c.containerSpec() does not return an error
if test.err is not set.
However, if the require fails (i.e. it indeed returned an error) the
rest of the code is executed anyways. The rest of the code assumes it
did not return an error (so code assumes spec is not nil). This fails
miserably if it indeed returned an error, as spec is nil and go crashes
while running the unit tests.
Let's require it is not an error, so code does not continue to execute
if that fails and go doesn't crash.
In the test.err case is not harmful the bug of using assert, but let's
switch it to require too as that is what we really want.
Signed-off-by: Rodrigo Campos <rodrigoca@microsoft.com>
A domainname field was recently added to the OCI spec. Prior to this
folks would need to set this with a sysctl, but now runtimes should be
able to setdomainname(2). There's an open change to runc at the moment
to add support for this so I've just left testing as a couple spec
validations in CRI until that's in and usable.
Signed-off-by: Danny Canter <danny@dcantah.dev>
This patch requests the OCI runtime to create a userns when the CRI
message includes such request.
Signed-off-by: Rodrigo Campos <rodrigoca@microsoft.com>
This allows user namespace support to progress, either by allowing
snapshotters to deal with ownership, or falling back to containerd doing
a recursive chown.
In the future, when snapshotters implement idmap mounts, they should
report the "remap-ids" capability.
Co-authored-by: Rodrigo Campos <rodrigoca@microsoft.com>
Signed-off-by: Rodrigo Campos <rodrigoca@microsoft.com>
Signed-off-by: David Leadbeater <dgl@dgl.cx>
As explained in the comments, this patch lets the OCI runtime create the
netns when userns are in use. This is needed because the netns needs to
be owned by the userns (otherwise can't modify the IP, etc.).
Before this patch, we are creating the netns and then starting the pod
sandbox asking to join this netns. This can't never work with userns, as
the userns needs to be created first for the netns ownership to be
correct.
One option would be to also create the userns in containerd, then create
the netns. But this is painful (needs tricks with the go runtime,
special care to write the mapping, etc.).
So, we just let the OCI runtime create the userns and netns, that
creates them with the proper ownership.
As requested by Mike Brown, the current code when userns is not used is
left unchanged. We can unify the cases (with and without userns) in a
future release.
Signed-off-by: Rodrigo Campos <rodrigoca@microsoft.com>
Due to when we were updating the pod sandboxes underlying container
object, the pointer to the sandbox would have the right info, but
the on-disk representation of the data was behind. This would cause
the data returned from loading any sandboxes after a restart to have
no CNI result or IP information for the pod.
This change does an additional update to the on-disk container info
right after we invoke the CNI plugin so the metadata for the CNI result
and other networking information is properly flushed to disk.
Signed-off-by: Danny Canter <danny@dcantah.dev>
While we need to support CRI v1alpha2, the implementation doesn't have
to be tied to gogo/protobuf.
Signed-off-by: Kazuyoshi Kato <katokazu@amazon.com>
We do a ton of host networking checks around the CRI plugin, all mainly
doing the same thing of checking the different quirks on various platforms
(for windows are we a HostProcess pod, for linux is namespace mode the
right thing, darwin doesn't have CNI support etc.) which could all be
bundled up into a small helper that can be re-used.
Signed-off-by: Danny Canter <danny@dcantah.dev>
OCI runtime spec defines memory.swap as 'limit of memory+Swap usage'
so setting them to equal should disable the swap. Also, this change
should make containerd behaviour same as other runtimes e.g
'cri-dockerd/dockershim' and won't be impacted when user turn on
'NodeSwap' (https://github.com/kubernetes/enhancements/issues/2400) feature.
Signed-off-by: Qasim Sarfraz <qasimsarfraz@microsoft.com>
Remove direct invocation of old v0.1.0 NRI plugins. They
can be enabled using the revised NRI API and the v0.1.0
adapter plugin.
Signed-off-by: Krisztian Litkey <krisztian.litkey@intel.com>
Implement the adaptation interface required by the NRI
service plugin to handle CRI sandboxes and containers.
Hook the NRI service plugin into CRI request processing.
Signed-off-by: Krisztian Litkey <krisztian.litkey@intel.com>
In golang when copy a slice, if the slice is initialized with a
desired length, then appending to it will cause the size double.
Signed-off-by: bin liu <liubin0329@gmail.com>
Go 1.18 and up now provides a strings.Cut() which is better suited for
splitting key/value pairs (and similar constructs), and performs better:
```go
func BenchmarkSplit(b *testing.B) {
b.ReportAllocs()
data := []string{"12hello=world", "12hello=", "12=hello", "12hello"}
for i := 0; i < b.N; i++ {
for _, s := range data {
_ = strings.SplitN(s, "=", 2)[0]
}
}
}
func BenchmarkCut(b *testing.B) {
b.ReportAllocs()
data := []string{"12hello=world", "12hello=", "12=hello", "12hello"}
for i := 0; i < b.N; i++ {
for _, s := range data {
_, _, _ = strings.Cut(s, "=")
}
}
}
```
BenchmarkSplit
BenchmarkSplit-10 8244206 128.0 ns/op 128 B/op 4 allocs/op
BenchmarkCut
BenchmarkCut-10 54411998 21.80 ns/op 0 B/op 0 allocs/op
While looking at occurrences of `strings.Split()`, I also updated some for alternatives,
or added some constraints; for cases where an specific number of items is expected, I used `strings.SplitN()`
with a suitable limit. This prevents (theoretical) unlimited splits.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Signed-off-by: Swagat Bora <sbora@amazon.com>
Add spans around image unpack operations
Use image.ref to denote image name and image.id for the image config digest
Add top-level spand and record errors in the CRI instrumentation service
Old TODO stating that pkg/cri/opts's `WithWindowsNetworkNamespace`
should be moved to the main containerd pkg was out of date as thats
already been done (well, to the /oci package). This just removes it
and swaps all uses of `WithWindowsNetworkNamespace` to the oci
packages impl.
Signed-off-by: Danny Canter <danny@dcantah.dev>
Remove nolint-comments that weren't hit by linters, and remove the "structcheck"
and "varcheck" linters, as they have been deprecated:
WARN [runner] The linter 'structcheck' is deprecated (since v1.49.0) due to: The owner seems to have abandoned the linter. Replaced by unused.
WARN [runner] The linter 'varcheck' is deprecated (since v1.49.0) due to: The owner seems to have abandoned the linter. Replaced by unused.
WARN [linters context] structcheck is disabled because of generics. You can track the evolution of the generics support by following the https://github.com/golangci/golangci-lint/issues/2649.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
- fix "nolint" comments to be in the correct format (`//nolint:<linters>[,<linter>`
no leading space, required colon (`:`) and linters.
- remove "nolint" comments for errcheck, which is disabled in our config.
- remove "nolint" comments that were no longer needed (nolintlint).
- where known, add a comment describing why a "nolint" was applied.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
This `//nolint` was added in f5c7ac9272
to suppress warnings about the `NameToCertificate` function being deprecated:
// Deprecated: NameToCertificate only allows associating a single certificate
// with a given name. Leave that field nil to let the library select the first
// compatible chain from Certificates.
Looking at that, it was deprecated in Go 1.14 through
eb93c684d4
(https://go-review.googlesource.com/c/go/+/205059), which describes:
crypto/tls: select only compatible chains from Certificates
Now that we have a full implementation of the logic to check certificate
compatibility, we can let applications just list multiple chains in
Certificates (for example, an RSA and an ECDSA one) and choose the most
appropriate automatically.
NameToCertificate only maps each name to one chain, so simply deprecate
it, and while at it simplify its implementation by not stripping
trailing dots from the SNI (which is specified not to have any, see RFC
6066, Section 3) and by not supporting multi-level wildcards, which are
not a thing in the WebPKI (and in crypto/x509).
We should at least have a comment describing why we are ignoring this, but preferably
review whether we should still use it.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Currently CDI registry is reconfigured on every
WithCDI call, which is a relatively heavy operation.
This happens because cdi.GetRegistry(cdi.WithSpecDirs(cdiSpecDirs...))
unconditionally reconfigures the registry (clears fs notify watch,
sets up new watch, rescans directories).
Moving configuration to the criService.initPlatform should result
in performing registry configuration only once on the service start.
Signed-off-by: Ed Bartosh <eduard.bartosh@intel.com>
As WithCDI is CRI-only API it makes sense to move it
out of oci module.
This move can also fix possible issues with this API when
CRI plugin is disabled.
Signed-off-by: Ed Bartosh <eduard.bartosh@intel.com>
pkg/cri/sbserver/cri_fuzzer.go and pkg/cri/server/cri_fuzzer.go were
mostly the same.
This commit merges them together and move the unified fuzzer to
contrib/fuzz again to sort out dependencies. pkg/cri/ shouldn't consume
cmd/.
Signed-off-by: Kazuyoshi Kato <katokazu@amazon.com>
`ioutil` has been deprecated by golang. All the code in `ioutil` just
forwards functionality to code in either the `io` or `os` packages.
See https://github.com/golang/go/pull/51961 for more info.
Signed-off-by: Jeff Widman <jeff@jeffwidman.com>
All of the CRI store related packages all use the standard errdefs
errors now for if a key doesn't or already exists (ErrAlreadyExists,
ErrNotFound), but the comments for the methods still referenced
some unused package specific error definitions. This change just
updates the comments to reflect what errors are actually returned
and adds comments for some previously undocumented exported functions.
Signed-off-by: Daniel Canter <dcanter@microsoft.com>
All containers except the pause container, mount `/dev/shm" with flags
`nosuid,nodev,noexec`. So change mount options for pause container to
keep consistence.
This also helps to solve issues of failing to mount `/dev/shm` when
pod/container level user namespace is enabled.
Fixes: #6911
Signed-off-by: Jiang Liu <gerry@linux.alibaba.com>
Signed-off-by: Lei Wang <wllenyj@linux.alibaba.com>
This's an optimization to get rid of redundant `/dev/shm" mounts for pause container.
In `oci.defaultMounts`, there is a default `/dev/shm` mount which is redundant for
pause container.
Fixes: #6911
Signed-off-by: Jiang Liu <gerry@linux.alibaba.com>
Signed-off-by: Lei Wang <wllenyj@linux.alibaba.com>
The TestPodAnnotationPassthroughContainerSpec test and the
TestContainerAnnotationPassthroughContainerSpec test both depend on a
platform-specific implementation of criService.containerSpec, which is
unimplemented on FreeBSD.
The TestSandboxContainerSpec depends on a platform-specific
implementation oc criService.sandboxContainerSpec, which is
unimplemented on FreeBSD.
Signed-off-by: Samuel Karp <me@samuelkarp.com>
This change does a couple things to remove some cruft/unused functionality
in the Windows snapshotter, as well as add a way to specify the rootfs
size in bytes for a Windows container via a new field added in the CRI api in
k8s 1.24. Setting the rootfs/scratch volume size was assumed to be working
prior to this but turns out not to be the case.
Previously I'd added a change to pass any annotations in the containerd
snapshot form (containerd.io/snapshot/*) as labels for the containers
rootfs snapshot. This was added as a means for a client to be able to provide
containerd.io/snapshot/io.microsoft.container.storage.rootfs.size-gb as an
annotation and have that be translated to a label and ultimately set the
size for the scratch volume in Windows. However, this actually only worked if
interfacing with the CRI api directly (crictl) as Kubernetes itself will
fail to validate annotations that if split by "/" end up with > 2 parts,
which the snapshot labels will (containerd.io / snapshot / foobarbaz).
With this in mind, passing the annotations and filtering to
containerd.io/snapshot/* is moot, so I've removed this code in favor of
a new `snapshotterOpts()` function that will return platform specific
snapshotter options if ones exist. Now on Windows we can just check if
RootfsSizeInBytes is set on the WindowsContainerResources struct and
then return a snapshotter option that sets the right label.
So all in all this change:
- Gets rid of code to pass CRI annotations as labels down to snapshotters.
- Gets rid of the functionality to create a 1GB sized scratch disk if
the client provided a size < 20GB. This code is not used currently and
has a few logical shortcomings as it won't be able to create the disk
if a container is already running and using the same base layer. WCIFS
(driver that handles the unioning of windows container layers together)
holds open handles to some files that we need to delete to create the
1GB scratch disk is the underlying problem.
- Deprecates the containerd.io/snapshot/io.microsoft.container.storage.rootfs.size-gb
label in favor of a new containerd.io/snapshot/windows/rootfs.sizebytes label.
The previous label/annotation wasn't being used by us, and from a cursory
github search wasn't being used by anyone else either. Now that there is a CRI
field to specify the size, this should just be a field that users can set
on their pod specs and don't need to concern themselves with what it eventually
gets translated to, but non-CRI clients can still use the new label/deprecated
label as usual.
- Add test to cri integration suite to validate expanding the rootfs size.
Signed-off-by: Daniel Canter <dcanter@microsoft.com>
