Add missing WCE (write-cache enable) property support. This not only
an enhancement, but also a fix for a bug.
Right now, when vhost_user_blk uses a qcow2 image, it doesn't write
the QCOW2 metadata until the guest explicitly requests a flush. In
practice, this is equivalent to the write back semantic.
Without WCE, the guest assumes write through for the virtio_blk
device, and doesn't send those flush requests. Adding support for WCE,
and enabling it by default, we ensure the guest does send said
requests.
Supporting "WCE = false" would require updating our qcow2
implementation to ensure that, when required, it honors the write
through semantics by not deferring the updates to QCOW2 metadata.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Add support for opening the disk images with O_DIRECT. This allows
bypassing the host's file system cache, which is useful to avoid
polluting its cache and for better data integrity.
This mode of operation can be enabled by adding the "direct=<bool>"
parameter to the "backend" argument:
./target/debug/vhost_user_blk --backend image=test.raw,sock=/tmp/vhostblk,direct=true
The "direct" parameter defaults to "false", to preserve the original
behavior.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Use RawFile as backend instead of File. This allows us to abstract
the access to the actual image with a specialized layer, so we have a
place where we can deal with the low-level peculiarities.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Doing I/O on an image opened with O_DIRECT requires to adhere to
certain restrictions, requiring the following elements to be aligned:
- Address of the source/destination memory buffer.
- File offset.
- Length of the data to be read/written.
The actual alignment value depends on various elements, and according
to open(2) "(...) there is currently no filesystem-independent
interface for an application to discover these restrictions (...)".
To discover such value, we iterate through a list of alignments
(currently, 512 and 4096) calling pread() with each one and checking
if the operation succeeded.
We also extend RawFile so it can be used as a backend for QcowFile,
so the later can be easily adapted to support O_DIRECT too.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Update queue number with 4 to verify if vhost-user-net device
and backend could work well with multiple queue.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
There are two new options num_queues and queue_size defined for
virtio-net, add them in test_valid_vm_config_net which is used
to validate that both the CLI and the OpenAPI will generate the
same configuration.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
Update the common part in net_util.rs under vm-virtio to add mq
support, meanwhile enable mq for virtio-net device, vhost-user-net
device and vhost-user-net backend. Multiple threads will be created,
one thread will be responsible to handle one queue pair separately.
To gain the better performance, it requires to have the same amount
of vcpus as queue pair numbers defined for the net device, due to
the cpu affinity.
Multiple thread support is not added for vhost-user-net backend
currently, it will be added in future.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
Add num_queues and queue_size for virtio-net device to make them configurable,
while add the associated options in command line.
Update cloud-hypervisor.yaml with the new options for NetConfig.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
Since the common parts are put into net_util.rs under vm-virtio,
refactoring code for virtio-net device, vhost-user-net device
and backend to shrink the code size and improve readability
meanwhile.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
This specifies how much address space should be reserved for hotplugging
of RAM. This space is reserved by adding move the start of the device
area by the desired amount.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
With the latest version of virtiofsd, some changes have been made
to the socket path parameter. This needs to be updated in the cloud
hypervisor codebase, so that our CI keeps running correctly.
Fixes#611
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The cache= option from the virtiofsd backend is meant to be used
depending on how the user wants to use the guest page cache. For most
use cases, we want to bypass the guest page cache to reduce the guest
memory footprint, which is why "cache=none" should be the default.
"cache=always" is the other option when the user wants to do something
specific by using the guest page cache, but should not be the default
anymore.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
With latest kernel, virtio-fs mount command has been simplified, and
this needs to be applied everywhere in our tests and documentation.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Because we don't always reach the expected footprint improvements with
KSM, let's review the numbers. By reducing the expectations and
increasing the amount of pages to scan, this should stabilize the CI.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This patch is to check if block device is readonly
when backend set readonly=true.
The lsblk command can show the RO value in the guest.
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
The current backend only support rw, and we also need
add readonly support.
The new command:
vhost_user_blk \
--backend "image=/home/test.img, \
sock=/home/path/vhost.socket, \
readonly=true"
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
The goal here is to ensure that CLI and OpenAPI both behave as closely
as possible, and also that they behave as expected.
Leveraging the reorganization of the code, we can now compare two
VmConfig structures generated from one CLI entry on one side, and from
an OpenAPI entry (JSON payload) on the other side.
Fixes#535
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This brings more modularity to the code, which will be helpful when we
will later test the CLI and OpenAPI generate the same VmConfig output.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This brings more modularity to the code, which will be helpful when we
will later test the CLI and OpenAPI generate the same VmConfig output.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This brings more modularity to the code, which will be helpful when we
will later test the CLI and OpenAPI generate the same VmConfig output.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
On our CI the /tmp filesystem is mounted as tmpfs and this is the
location where the test disk images are located. When the CI worker
nodes have less memory and fewer CPUs the tmpfs fills up as the tests
run in parallel.
Introduce a mechanism to reduce the parallelism of the tests based on
starvation of the tmpfs disk availability.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
By default, and in order to avoid falling into the legacy CLI usage, the
CPU argument should at least include "boot=" to define the number of
CPUs.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
In order to validate that multiple devices can be passed through and
they are still fully functional, this patch extends the existing VFIO
test to pass a second virtio-net device, and verifies that both
interfaces are functional by ssh'ing into each network interface.
Fixes#503
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The easiest way to detect if the kernel is willing to accept hotplug
vCPUs is to check the dmesg output.
Switch the test to bionic as the Clear Cloud image lacks "dmesg."
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
In order to validate the new virtio-fs daemon written in Rust is
behaving correctly, a new integration test has been added. Important to
note that for now, only a test with cache=none and dax=off can be added
since the daemon does not support shared memory region yet.
The long term goal being to replace virtiofsd with vhost_user_daemon
once it will reach parity regarding the supported features.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Because the vhost_user_backend crate needs some changes to support
moving the process to a different mount namespace and perform a pivot
root, it is not possible to change '/' to the given shared directory.
This commit, as a temporary measure, let the code point at the given
shared directory.
The long term solution is to perform the mount namespace change and the
pivot root as this will provide greater security.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This patch implements a vhost-user-fs daemon based on Rust. It only
supports communicating through the virtqueues. The support for the
shared memory region associated with DAX will be added later.
It relies on all the code copied over from the crosvm repository, based
on the commit 961461350c0b6824e5f20655031bf6c6bf6b7c30.
It also relies on the vhost_user_backend crate, limiting the amount of
code needed to get this daemon up and running.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Instead of using bash and awk, using Rust allows us to retrieve
information about a VM process with the right permissions as we are not
forced to spawn a new child process.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The test validates that when the mergeable option is enabled, the
resulting PSS for two instances of cloud-hypervisor is lower than two
instances not using the mergeable flag.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
In order to let the user indicate if the persistent memory pages should
be marked as mergeable or not, a new option is being introduced.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
In order to let the user indicate if the guest RAM pages should be
marked as mergeable or not, a new option is being introduced.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
When vmm.ping give a response, we expect get the version from
the VMM not the vmm create
Signed-off-by: Jose Carlos Venegas Munoz <jose.carlos.venegas.munoz@intel.com>
vmm.ping will help to check if http API server is up and
running.
This also removes the vmm.info endpoint.
Signed-off-by: Jose Carlos Venegas Munoz <jose.carlos.venegas.munoz@intel.com>
Use the new vhost-user-blk backend for the integration tests,
eliminating the need for building vubd using the implementation in
QEMU.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Create a vhost-user-blk backend using vhost-user-backend and following
the conventions established by the existing vhost-user-net
implementation.
This backend is based on https://github.com/slp/vhost-user-backend,
but a bit simplified, making it closer to the original implementation
in Firecracker. The main features missing are EVENT_IDX, support for
asynchronous I/O and multiqueue, but it's still fully functional and
provides a good starting point for evolving it into a more complete
implementation.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Extend VhostUserBackend trait with protocol_features(), so device
backend implementations can freely define which protocol features they
want to support.
Signed-off-by: Sergio Lopez <slp@redhat.com>
A new ClearLinux image has been uploaded to the Azure storage account.
It is based off of the ClearLinux cloudguest image 31310 version, with
two extra bundles added to it.
First bundle is sysadmin-basic to include utility like netcat, and the
second bundle is iperf, adding the iperf binary to the image.
The image is 2G in size.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
When VFIO devices are created and if the device is attached to the
virtual IOMMU, the ExternalDmaMapping trait implementation is created
and associated with the device. The idea is to build a hash map of
device IDs with their associated trait implementation.
This hash map is provided to the virtual IOMMU device so that it knows
how to properly trigger external mappings associated with VFIO devices.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Having the virtual IOMMU created with --iommu is one thing, but we also
need a way to decide if a VFIO device should be attached to the virtual
IOMMU or not. That's why we introduce an extra option "iommu" with the
value "on" or "off". By default, the device is not attached, which means
"iommu=off".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Clean up the error handling and ensure that where possible errors are
propagated. Make use of std::convert::From in order to translate error
types.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Simplify the check for the unusual situation where the memory is not
configured by using .ok_or() on the option to convert it to a result.
This cleans up a bunch of extra indentation.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Remove messages that are left over from the development of the project
that represent normal operation for the backend. This cleans up the
console output and improves performance.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
We pause a VM from the API, then SSH'ing into it should fail.
After resuming, SSH'ing should work again.
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
Because the L2 VM running in the VFIO integration test is actually
running as L3 (since the CI runs in a VM), it can take quite some
time for this VM to boot.
The way to solve this issue is to extend the sleep time before to try
communicating with the L2 VM, but also to speed up the boot time by
using virtio-console instead of serial. We suspect the use of serial,
implying PIO VM exits for each character on the serial port is quite
expensive compared to the paravirtualized console.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Azure virtual machines can have private IPs in the 172.16.x.x range,
causing some issues with the VFIO test. By using 172.17.x.x for this
test, we avoid IP conflicts.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Now that our custom kernel includes all the patches for the full support
of virtio-iommu, we can go one step further by attaching the virtio-net
device to the virtual IOMMU and use it to SSH some commands validating
both disks and the network card are isolated into their own IOMMU group.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Now that cloud-hypervisor can expose a virtual IOMMU to its guest VM,
the integration test validating the VFIO support with virtio-net can be
updated to use cloud-hypervisor exclusively.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Having the virtual IOMMU created with --iommu is one thing, but we also
need a way to decide if a virtio-vsock device should be attached to this
virtual IOMMU or not. That's why we introduce an extra option "iommu"
with the value "on" or "off". By default, the device is not attached,
which means "iommu=off".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Having the virtual IOMMU created with --iommu is one thing, but we also
need a way to decide if a virtio-console device should be attached to
this virtual IOMMU or not. That's why we introduce an extra option
"iommu" with the value "on" or "off". By default, the device is not
attached, which means "iommu=off".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Having the virtual IOMMU created with --iommu is one thing, but we also
need a way to decide if a virtio-pmem device should be attached to this
virtual IOMMU or not. That's why we introduce an extra option "iommu"
with the value "on" or "off". By default, the device is not attached,
which means "iommu=off".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Having the virtual IOMMU created with --iommu is one thing, but we also
need a way to decide if a virtio-rng device should be attached to this
virtual IOMMU or not. That's why we introduce an extra option "iommu"
with the value "on" or "off". By default, the device is not attached,
which means "iommu=off".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Having the virtual IOMMU created with --iommu is one thing, but we also
need a way to decide if a virtio-net device should be attached to this
virtual IOMMU or not. That's why we introduce an extra option "iommu"
with the value "on" or "off". By default, the device is not attached,
which means "iommu=off".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Having the virtual IOMMU created with --iommu is one thing, but we also
need a way to decide if a virtio-blk device should be attached to this
virtual IOMMU or not. That's why we introduce an extra option "iommu"
with the value "on" or "off". By default, the device is not attached,
which means "iommu=off".
One side effect of this new option is that we had to introduce a new
option for the disk path, simply called "path=".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
We use the serde crate to serialize and deserialize the VmVConfig
structure. This structure will be passed from the HTTP API caller as a
JSON payload and we need to deserialize it into a VmConfig.
For a convenient use of the HTTP API, we also provide Default traits
implementations for some of the VmConfig fields (vCPUs, memory, etc...).
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
The kernel path was the only mandatory command line option.
With the addition of the --api-socket option, we can run without a
kernel path and get it later through the API.
Since we can end up with VM configurations that are no longer valid by
default, we need to provide a validation check for it. For now, if the
kernel path is not defined, the VM configuration is invalid.
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
The API server will unconditionally run through a UNIX domain socket
which default path is /run/user/<uid>/cloud-hypervisor.<pid>.
The --api-socket command line option allows to override that default
value with some custom socket path.
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
With the API server socket option, we will be able to support a model
where the user can start cloud-hypervisor with no options or an
alternative API server socket path. In this case, we don't want to try
to start a new guest VM, and for that we need to know if the user has
set any VM configuration at all. Grouping all VM configuration specific
options together is one way to be able to know about it.
If the user has not set any VM configuration, we only start the API
server. If it has set anything, we will verify that the overall
configuration is valid and will implicitly convert that configuration
into a request to the API server.
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
An integration test relying on the new vhost-user-net backend now
replaces the previous test using the QEMU test backend. This allows
us to avoid building the QEMU backend, and we now really exercise the
vhost-user-net implementation as it is used for the ssh communication
in this test.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Create vhost-user-net backend with Tap interface, to offload network
transaction from cloud-hypervisor. The goal is to provide flexibility
about the backend being in use, but also more security as it will allow
users to isolate the backend with different security profiles since it
will run as a dedicated process on the host.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
We now start the main VMM thread, which will be listening for VM and IPC
related events.
In order to start the configured VM, we no longer directly call the VM
API but we use the IPC instead, to first create and then start a VM.
Fixes: #303
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
Unlike the Vmm structure we removed with commit bdfd1a3f, this new one
is really meant to represent the VM monitoring/management object.
For that, we implement a control loop that will replace the one that's
currently embedded within the Vm structure itself.
This will allow us to decouple the VM lifecycle management from the VM
object itself, by having a constantly running VMM control loop.
Besides the VM specific events (exit, reset, stdin for now), the VMM
control loop also handles all the Cloud Hypervisor IPC requests.
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
Once passed to the VM creation routine, a VmConfig structure is
immutable. We can simply carry a Arc of it instead of a reference.
This also allows us to remove any lifetime bound from our VM.
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
The Vmm structure is just a placeholder for the KVM instance. We can
create it directly from the VM creation routine instead.
Signed-off-by: Samuel Ortiz <sameo@linux.intel.com>
Now that vhost-user has been fixed regarding size of the virtqueues,
booting a VM with the firmware from a vhost-user-blk backend actually
works. That's why this commit updates the previously introduced
integration test to make it use the firmware instead of direct kernel
boot.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This commit extends the existing integration test related to
vhost-user-blk by validating the block image contains one file
"foo" containing "bar".
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Currently we need use backend device from Qemu to test vhost-user-blk
device. Once the rust backend is ready, we will replace it.
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Now that we have expanded our address space we should add automated
testing to try VMs that use a large amount of RAM.
As the hypervisor does an anonymous mmap() for the backing of memory and
during a typical test boot the guest will not touch it all it should be
possible to test large RAM VMs even if that exceeds the RAM of the host
machine.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Label tests that definitely can't function with virtio-mmio (because
they use the firmware) and within those that can be used mark individual
assertions that will no longer hold (around PCI.)
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
This patch extends the current set of integration tests to correctly
validate that virtio-vsock is functional. It establishes a communication
between host and guest relying on the newly integrated vsock device.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Rely on the newly generated Clear Linux image for the integration
testing of cloud-hypervisor. The image has been generated using the
Clear Linux clr-installer tooling, which means it is in compliance with
the Clear Linux licensing.
This new image contains one more bundle that was not part of the default
cloudguest image. This bundle is basic-sysadmin, and contains both nc
and socat utilities.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The goal here is to decouple the Guest instance from the ssh connection
to send some commands to the guest. The reason being to allow ssh
commands to be issued from a different thread, which can be useful to
wait for the end of a command with a thread.join().
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
As part of the reboot test check that the binary cleanly terminated
after the subsequent shutdown.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The new flag vsock is meant to be used in order to create a VM with a
virtio-vsock device attached to it. Two parameters are needed with this
device, "cid" representing the guest context ID, and "sock" representing
the UNIX socket path which can be accessed from the host.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The default number of MSI-X vector allocated was 2, which is the minimum
defined by the virtio specification. The reason for this minimum is that
virtio needs at least one interrupt to signal that configuration changed
and at least one to specify something happened regarding the virtqueues.
But this current implementation is not optimal because our VMM supports
as many MSI-X vectors as allowed by the MSI-X specification (2048 max).
For that reason, the current patch relies on the number of virtqueues
needed by the virtio device to determine the right amount of MSI-X
vectors needed. It's important not to forget the dedicated vector for
any configuration change too.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Being able to reboot requires us to identify all the resources we are
leaking and cleaning those up before we can enable reboot. For now if
the user requests a reboot then shutdown instead.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>