Add support for virtio-pci transport #5240
Open
+4,949
−1,450
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## feature/pcie #5240 +/- ##
================================================
+ Coverage 79.23% 80.26% +1.02%
================================================
Files 262 265 +3
Lines 29347 30967 +1620
================================================
+ Hits 23254 24856 +1602
- Misses 6093 6111 +18
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
bchalios
force-pushed
the
virtio-pci
branch
7 times, most recently
from
a80faa9 to
15ed3eb
Compare
bchalios
force-pushed
the
virtio-pci
branch
6 times, most recently
from
61c6be1 to
9d18f11
Compare
bchalios
force-pushed
the
virtio-pci
branch
15 times, most recently
from
c0ada38 to
507cbce
Compare
bchalios
force-pushed
the
virtio-pci
branch
2 times, most recently
from
8436358 to
a82f14c
Compare
We need the new KvmIrqRouting FamStruct wrapper from kvm-bindings, which though forces us to update vmm-sys-util to 0.14.0 and also bump all downstream dependencies of vmm-sys-util to use that version. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Define thiserror::Error and displaydoc::Display for various error types in the vended PCI crate. This way we can embed them in our error types downstream. Also export a few types and struct fields that were private and we will be needing them. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Instead of returning an `EventFd` type, which will actually force us to clone the file descriptor in the Firecracker side. Signed-off-by: Babis Chalios <bchalios@amazon.es>
This is code we are not going to use in Firecracker. Remove it, so we can keep the crates we vend as minimal as possible, including only things we are actually using. Signed-off-by: Babis Chalios <bchalios@amazon.es>
We'd like to be able to store Vm within an atomic reference so we can pass it around and share it with other components. The main issue with doing this change is that we need Vm to be `mut` during initialization and the builder.rs code was creating Vmm with Vm embedded in it. To solve this, we break down the initialization of the Vmm object. We first create its individual parts (Vm, Kvm and DeviceManager), perform any necessary initialization logic on Vm and once this done add it within an Arc. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Add logic to track the device interrupts used by the microVM. This is not strictly needed right now, but we will need it when adding support for MSI-X interrupts. MSI-X interrupts are configured at runtime and we need to interact with KVM to set the interruput routes. To do it, we need to keep track all of the interrupts the VM is using. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Enable Vm to vend and manage MSI/MSI-X interrupts. This adds the logic to create a set of MSI vectors and then handle their lifetime. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Vm object is now maintaining information about the interrupts (both traditional IRQs and MSI-X vectors) that are being used by microVM devices. Derive Serialize/Deserialize add logic for recreating objects for relevant types. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Commit d8c2714 (refactor: use VirtioInterrupt in VirtIO devices) which refactored devices to use new VirtioInterrupt type introduced a bug with the index used to trigger a queue interrupt. Instead of using the actual queue index, we were using the index of the used descriptor. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Apparently, PCI needs Queue::size to be initialized to the maximum possible size supported by the device, otherwise initialization fails. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Remove the flags in FADT that were declaring we do not support MSI and PCI ASPM. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Merge the device-related errors that DeviceManager might return. This way, we can avoid adding yet another error type for PCI devices and reduce some the variants of StartMicrovmError. Suggested-by: Egor Lazarchuk <yegorlz@amazon.co.uk> Signed-off-by: Babis Chalios <bchalios@amazon.es>
Add a VirtIO PCI transport implementation. When a Firecracker microVM is launched with --enable-pci, we will create all VirtIO devices using the PCI transport layer. Snapshotting of VirtIO PCI devices is not supported and we will add this functionality in later commit. Add a couple of tests that ensure that PCI configuration space is what expected. We read common fields and make sure the BAR we allocate for the VirtIO device is what expected. Signed-off-by: Babis Chalios <bchalios@amazon.es>
We are now calling KVM_CHECK_EXTENSION for checking the KVM_CAP_MSI_DEVID capability. We are also calling KVM_SET_GSI_ROUTING to set the interrupts routes and KVM_IRQFD to set/unset interrupt lines. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Add some unit tests to PciSegment. We now test that the next_device_bdf() method and the initialization logic work as expected. We also check that the configuration space of the PCI segment is correctly registered with the MMIO and, on x86, PIO bus. Signed-off-by: Babis Chalios <bchalios@amazon.es>
vm-allocator now allows us to (De)serialize IdAllocator and AddressAllocator types. Add ResourceAllocator in DeviceManager snapshot state and restore it when loading a snapshot. Like this we can avoid doing the ExactMatch allocations during snapshot resumes for reserving the exact same MMIO ranges. Moreover, change DeviceManager and PciDevices to provide save/restore functionality via the Persist trait. Like that we can avoid first creating the objects and then restoring their state, overwriting their fields. Signed-off-by: Babis Chalios <bchalios@amazon.es>
VirtIO MMIO restore logic activates the device the moment we restore the device state, if the device was activated when snapshotted. Move the activation responsibility to the logic the restores the MMIO transport. The reason for this change is that that's how it will be done for the PCI transport. Unifying this will allow us reusing the same types for restoring the non-transport state of devices. Note that we needed to change the way Net devices are saved/restored. RxBuffer type of Net devices holds RX descriptors that we have parsed from the Queue ahead of time. The way we restored this info was manipulating the queue to re-parse the RX descriptors during the restore phase. However, we need the device to be activated to do so, which now isn't. So, instead of storing this info inside the snapshot make sure we have flushed everything before taking the snapshot. Also, simplify a bit the types that we use for serializing/deserializing the state of a device. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Support serializing the device-specific and transport state of a VirtIO device that uses the PCI transport. Signed-off-by: Babis Chalios <bchalios@amazon.es>
ResourceAllocator object was part of DeviceManager since it is (mainly) devices that use it. ResourceAllocator is as well the object that implements (in a dummy way, for the moment) the DeviceRelocation trait which PciDevices use to move the address space of a PciDevice when triggered from the guest. Problem with DeviceRelocation is that it also needs the Vm file descriptor to perform the relocation, because we need to move register the new IO event fd for VirtIO devices. To make things simpler, move ResourceAllocator inside the Vm object. In subsequent commit we will remove the DeviceRelocation from ResourceAllocator and move it to Vm instead. This has the nice secondary effect that we were able to simplify the signature of many device-related methods that received Vm and ResourceAllocator arguments. Signed-off-by: Babis Chalios <bchalios@amazon.es>
We had previously added MMIO and Port IO buses inside ResourceAllocator so that we could implement DeviceRelocation for the type. Now, we will delegate device relocation responsibilities to ArchVm instead. That is because device relocation requires access to the Vm file descriptor as well. As a result, we can move buses to the Vm object itself. Add MMIO bus to VmCommon as both architectures use it. Add PortIO bus for x86 architecture only. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Implement PCI device relocation logic for Vm type. Up until now, we were only implementing it in a dummy way. Make sure we correctly deallocate previously allocate ranges and unregister them from the corresponding bus. Try to allocate new addresses and register them with the bus. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Add support for ITS device which provides support for MSI interrupts on ARM architecture. This is currently supported only on systems with GICv3 interrupt controller. In order to make saving/restore of ITS state work properly, we need to change the order in which we restore redistributor register GICR_CTLR. We need to make sure that this register is restored last. Otherwise, restoring GICR_PROPBASER doesn't have any effect and ITS depends on it in order to save/restore ITS tables to/from guest memory. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Refactor the test code that inserts VirtIO devices in a Vmm object and then add a test which creates a Vmm with PCI devices and then serializes and deserializes the device manager and ensures that everything is as restored as expected. Signed-off-by: Babis Chalios <bchalios@amazon.es>
Use pci_enabled fixture for boot time, block, and network tests to create PCI microVM variants as well. Signed-off-by: Babis Chalios <bchalios@amazon.es>
We only pass pci=off if PCI is disabled in Firecracker. Adapt tests and comments to reflect that. Signed-off-by: Babis Chalios <bchalios@amazon.es>
So that we don't have to downcast VirtioDevice trait objects to the actual device type before calling the logic to process events for each device. Signed-off-by: Babis Chalios <bchalios@amazon.es>
roypat
reviewed
Jun 30, 2025
| impl MsiVector { | ||
| /// Enable vector | ||
| fn enable(&self, vmfd: &VmFd) -> Result<(), errno::Error> { | ||
| if !self.enabled.load(Ordering::Acquire) { |
There was a problem hiding this comment.
should this be an compare_exchange (with a rollback if register_irqfd fails, although I guess we dont really expect it to fail) instead to prevent the function from racing with itself? Can it even race with itself?
|
|
||
| impl InterruptSourceGroup for MsiVectorGroup { | ||
| fn enable(&self) -> vm_device::interrupt::Result<()> { | ||
| for (_, route) in self.irq_routes.iter() { |
There was a problem hiding this comment.
nit:
Suggested change
| for (_, route) in self.irq_routes.iter() { | |
| for route in self.irq_routes.values() { |
(there's a few more cases of this)
| &self, | ||
| index: vm_device::interrupt::InterruptIndex, | ||
| ) -> vm_device::interrupt::Result<()> { | ||
| if let Some(route) = self.irq_routes.get(&index) { |
There was a problem hiding this comment.
nit: reuse the notifier function from below here maybe?
|
|
||
| if self.common.fd.check_extension(kvm_ioctls::Cap::MsiDevid) { | ||
| // On AArch64, there is limitation on the range of the 'devid', | ||
| // it cannot be greater than 65536 (the max of u16). |
| for (i, gsi) in resource_allocator | ||
| .allocate_gsi(count as u32)? | ||
| .iter() | ||
| .enumerate() |
There was a problem hiding this comment.
.zip(0u32..) here instead of enumerate() lets you avoid the cast below
| } | ||
|
|
||
| Err(std::io::Error::other(format!( | ||
| "MSI-X update: invalid vector index {index}" |
There was a problem hiding this comment.
as an aside (irrelevant to this PR because its vendored code): I really dislike this error handling using std::io::Error::other
Comment on lines
433
to
434
| let (device_activated, interrupt_status, cap_pci_cfg_info) = | ||
| (false, 0, VirtioPciCfgCapInfo::default()); |
There was a problem hiding this comment.
nit: I dont quite get why we keep these 3 in local variables instead of just inlining them into the struct initializer below
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Changes
This PR adds support for using a PCI transport for VirtIO devices. The PR adds PCI support for x86 and Aarch64 systems both for booted and snapshot-restored microVMs. The VirtIO-PCI transport is used whenever Firecracker is launched with PCI enabled. Otherwise, we fall back to the MMIO transport.
This PR also adds support for MSI-X interrupts for VirtIO devices. When using the PCI transport VirtIO devices always use MSI-X type of interrupts. MMIO devices still use the IRQ based ones.
Since we needed to change various things in order to support PCI along with MMIO transport the PR brings a few quality of life changes in the code base:
ResourceAllocator, so that we free old memory and allocate new one, the MMIO and Port IO buses, so that we register the device to its new allocated address range and the KVM Vm file descriptor so that we correctly set the new IO event fds. As a result, we move buses, and resource allocator inside theVmobject and make the latter responsible for the relocation when needed. This moves simplify the signatures of a lot of functions that were using a combination ofResourceAllocator,VmandBusobjects. In the future, we might want to even moveDeviceManagerinside theVmobject. This should simplify things even further.serdesupport in the new version ofvm-allocatorcrate, so that we don't have to perform allocations with fixed addresses upon snapshot restore.Reason
Support VirtIO devices with PCI transport
License Acceptance
By submitting this pull request, I confirm that my contribution is made under
the terms of the Apache 2.0 license. For more information on following Developer
Certificate of Origin and signing off your commits, please check
CONTRIBUTING.md.PR Checklist
tools/devtool checkstyleto verify that the PR passes theautomated style checks.
how they are solving the problem in a clear and encompassing way.
in the PR.
CHANGELOG.md.Runbook for Firecracker API changes.
integration tests.
TODO.rust-vmm.