Since we can't find a homogeneous value for the resource/cgroup
management of multiple hypervisors, and we have decoupled the
env vars in the Makefile, we don't need the generic ones.
Signed-off-by: Anastassios Nanos <ananos@nubificus.co.uk>
To avoid overriding env vars when multiple hypervisors are
available, we add per-hypervisor vars for static resource
management and cgroups handling. We reflect that in the
relevant config files as well.
Signed-off-by: Anastassios Nanos <ananos@nubificus.co.uk>
It's better to check the container's status before
try to send signal to it. Since there's no need
to send signal to it when the container's stopped.
Signed-off-by: Fupan Li <fupan.lfp@antgroup.com>
Since stop sandbox would be called in multi path,
thus it's better to set and check the sandbox's state.
Fixes: #10042
Signed-off-by: Fupan Li <fupan.lfp@antgroup.com>
This commit aligns the OCI Spec implementation in runtime-rs
with the OCI Spec definitions and related operations provided
by oci-spec-rs. Key changes as below:
(1) Leveraged oci-spec-rs to align Kata Runtime OCI Spec with
the official OCI Spec.
(2) Introduced runtime-spec to separate OCI Spec definitions
from Kata-specific State data structures.
(3) Preserved the original code logic and implementation as
much as possible.
(4) Made minor code adjustments to adhere to Rust programming
conventions;
Fixes#9766
Signed-off-by: Alex Lyn <alex.lyn@antgroup.com>
Hotunplugging memory is not guaranteed or even likely to work.
Nevertheless I'd really like to have this code in for tests and
observation. It shouldn't hurt, from experience so far.
Signed-off-by: Pavel Mores <pmores@redhat.com>
The bulk of this implementation are simple though tedious sanity checks,
alignment computations and logging.
Note that before any hotplugging, we query qemu directly for the current
size of hotplugged memory. This ensures that any request to resize memory
will be properly compared to the actual already available amount and only
necessary amount will be added.
Note also that we borrow checked_next_multiple_of() from CH implementation.
While this might look uncleanly it's just a rather temporary solution since
an equivalent function will apparently be part of std soon, likely the
upcoming 1.75.
Signed-off-by: Pavel Mores <pmores@redhat.com>
The algorithm is rather simple - we query qemu for existing memory devices
to figure out the index of the one we're about to add. Then we add a
backend object and a corresponding frontend device.
Signed-off-by: Pavel Mores <pmores@redhat.com>
When create container failed, it should cleanup the container
thus there's no device/resource left.
Fixes: #10044
Signed-off-by: Fupan Li <fupan.lfp@antgroup.com>
In dragonball Vfio device passthrough scenarois, the first passthrough
device will be allocated slot 0 which is occupied by root device.
It will cause error, looks like as below:
```
...
6: failed to add VFIO passthrough device: NoResource\n
7: no resource available for VFIO device"): unknown
...
```
To address such problem, we adopt another method with no pre-allocated
guest device id and just let dragonball auto allocate guest device id
and return it to runtime. With this idea, add_device will return value
Result<DeviceType> and apply the change to related code.
Fixes#9813
Signed-off-by: Alex Lyn <alex.lyn@antgroup.com>
Fixes: #9532
Instead of call agent.close_stdin in close_io, we call agent.write_stdin
with 0 len data when the stdin pipe ends.
Signed-off-by: Tim Zhang <tim@hyper.sh>
Add a basic runtime-rs `Hypervisor` trait implementation for
AWS Firecracker
- Add basic hypervisor operations (setup / start / stop / add_device)
- Implement AWS Firecracker API on a separate file `fc_api.rs`
- Add support for running jailed (include all sandbox-related content)
- Add initial device support (limited as hotplug is not supported)
- Add separate config for runtime-rs (FC)
Notes:
- devmapper is the only snapshotter supported
- to account for no sharefs support, we copy files in the sandbox (as
in the GO runtime)
- nerdctl spawn is broken (TODO: #7703)
Fixes: #5268
Signed-off-by: George Pyrros <gpyrros@nubificus.co.uk>
Signed-off-by: Anastassios Nanos <ananos@nubificus.co.uk>
Signed-off-by: Charalampos Mainas <cmainas@nubificus.co.uk>
Signed-off-by: George Ntoutsos <gntouts@nubificus.co.uk>
get_vmm_master_tid() currently returns an error with the message "cannot
get qemu pid (though it seems running)" when it finds a valid
QemuInner::qemu_process instance but fails to extract the PID out of it.
This condition however in fact means that a qemu child process was running
(otherwise QemuInner::qemu_process would be None) but isn't anymore (id()
returns None).
Signed-off-by: Pavel Mores <pmores@redhat.com>
Since Hypervisor::stop_vm() is called from the WaitProcess request handling
which appears to be per-container, it can be called multiple times during
kata pod shutdown. Currently the function errors out on any subsequent
call after the initial one since there's no VM to stop anymore. This
commit makes the function tolerate that condition.
While it seems conceivable that sandbox shouldn't be stopped by WaitProcess
handling, and the right fix would then have to happen elsewhere, this
commit at least makes qemu driver's behaviour consistent with other
hypervisor drivers in runtime-rs.
We also slightly improve the error message in case there's no
QemuInner::qemu_process instance.
Signed-off-by: Pavel Mores <pmores@redhat.com>
Since no objections were raised in the linked issue (#9847) this commit
removes the attempt to derive sandbox bundle path from container bundle
path. As described in more detail in the linked issue, this is container
runtime specific and doesn't seem to serve any purpose.
As for implementation, we hoist the only part of
get_shim_info_from_sandbox() that's still useful (getting the socket
address) directly into the caller and remove the function altogether.
Fixes#9847
Signed-off-by: Pavel Mores <pmores@redhat.com>
- Bump tokio to 1.38.0 to fix the security vulnerability
https://rustsec.org/advisories/RUSTSEC-2024-0019
If possible it would be good to add the many runtime-rs creates into the
runtime-rs workspace and provide a centralised version to avoid the updates
in many places.
Signed-off-by: stevenhorsman <steven@uk.ibm.com>
We take advantage of the Inner pattern to enable QemuInner::resize_vcpu()
take `&mut self` which we need to call non-const functions on Qmp.
This runs on Intel architecture but will need to be verified and ported
(if necessary) to other architectures in the future.
Signed-off-by: Pavel Mores <pmores@redhat.com>
The QMP_SOCKET_FILE constant in cmdline_generator.rs is made public to make
it accessible from QemuInner. This is fine for now however if the constant
needs to be accessed from additional places in the future we could consider
moving it to somewhere more visible.
The Debug impl for Qmp is empty since first, we don't actually want it,
it's only forced by Hypervisor trait bounds, and second, it doesn't have
anything to display anyway. If Qmp gets any members in the future that
can be meaningfully displayed they should be handled by Qmp's Debug::fmt().
Signed-off-by: Pavel Mores <pmores@redhat.com>
The constructor handles QMP connection initialisation, too, so there can
be non-functional Qmp instance.
Signed-off-by: Pavel Mores <pmores@redhat.com>
We set the VERSION variable consistently across Makefiles to
'unknown' if the file is empty or not present.
We also use git commands consistently for calculating the COMMIT,
COMMIT_NO variables, not erroring out when building outside of
a git repository.
In create_summary_file we also account for a missing/empty VERSION
file.
This makes e.g. the UVM build process in an environment where we
build outside of git with a minimal/reduced set of files smoother.
Signed-off-by: Manuel Huber <mahuber@microsoft.com>
It shouldn't call the initial_size_manager's setup_config
in the load_config since it had been called in the sandbox's
try_init function.
Fixes: #9778
Signed-off-by: Fupan Li <fupan.lfp@antgroup.com>
For kata container, the container's pid is meaning less to
containerd/crio since the container's pid is belonged to VM,
and containerd/crio couldn't use it. Thus we just return any
tid of kata shim or hypervisor. But since the hypervisor had
been stopped before deleting the container, and it wouldn't
get the hypervisor's tid for some supported hypervisor, thus
we'd better to return the kata shim's pid instead of hypervisor's
tid.
Fixes: #9777
Signed-off-by: Fupan Li <fupan.lfp@antgroup.com>
It creates this line, as the Golang runtime does:
-object rng-random,id=rng0,filename=/dev/urandom -device virtio-rng-pci,rng=rng0
Signed-off-by: Emanuel Lima <emlima@redhat.com>
This option allows to add a debug monitor socket when
`enable_debug = true` to control QEMU within debugging case.
Fixes: #9603
Signed-off-by: Alex Lyn <alex.lyn@antgroup.com>
Replace tab indentation with spaces for the three lines within the ifneq statements, aligning them with the surrounding code.
Fixes:#9692
Signed-off-by: sidneychang <2190206983@qq.com>
The `dial_timeout` works fine for Runtime-go, but is obsoleted in
Runtime-rs.
When the pod cannot connect to the Agent upon starting, we need to adjust
the `reconnect_timeout_ms` to increase the number of connection attempts to
the Agent.
Fixes: #9688
Signed-off-by: Xuewei Niu <niuxuewei.nxw@antgroup.com>
Implementation of QemuCmdLine has a fairly uniform and repetitive structure
that's guided by a set of conventions. These conventions have however been
mostly implicit so far, leading to a superfluous and annoying
request/force-push churn during qemu-rs PR reviews.
This commit aims to make things explicit so that contributors can take them
into account before an initial PR submission.
Signed-off-by: Pavel Mores <pmores@redhat.com>
All these settings are hardcoded as `false` and result in
no extra options on the QEMU command line, like the go
runtime does. There actually not needed :
- we're never going to ask QEMU to survive a guest shutdown
- we're never going to run QEMU daemonized since it prevents
log collection
- we're never going to ask QEMU to start with the guest stopped
No need to keep this code around then.
Signed-off-by: Greg Kurz <groug@kaod.org>
RTC was being built in a wrong fashion on commit #2bc5e3c6e2ab0145fa9e8be95df0d5086c07a517
RTC was being constructed inside the QemuCmdLine struct,
but it should've been built inside the devices vector.
Signed-off-by: Emanuel Lima <emlima@redhat.com>
We should init and asign the runtime instance to runtime
handler, otherwise, if the pause container failed to start,
which means the runtime instance failed to start, then the
following delete & shutdown request wouldn't be run, thus
the dead shim would be left.
Fixes: #9597
Signed-off-by: Fupan Li <fupan.lfp@antgroup.com>
