storage: Add storvsc_manager and disk_storvsc

Author: eric135

Cargo.lock

@@ -1373,6 +1373,22 @@ dependencies = [
  "vm_resource",
 ]
 
+[[package]]
+name = "disk_storvsc"
+version = "0.0.0"
+dependencies = [
+ "disk_backend",
+ "futures",
+ "inspect",
+ "scsi_buffers",
+ "scsi_defs",
+ "storvsc_driver",
+ "storvsp_protocol",
+ "tracing",
+ "vmbus_user_channel",
+ "zerocopy 0.8.24",
+]
+
 [[package]]
 name = "disk_striped"
 version = "0.0.0"
@@ -6926,10 +6942,12 @@ dependencies = [
 name = "storvsc_driver"
 version = "0.0.0"
 dependencies = [
+ "anyhow",
  "futures",
  "futures-concurrency",
  "guestmem",
  "inspect",
+ "mesh",
  "mesh_channel",
  "pal_async",
  "scsi_buffers",
@@ -6941,6 +6959,7 @@ dependencies = [
  "thiserror 2.0.12",
  "tracing",
  "tracing_helpers",
+ "user_driver",
  "vmbus_async",
  "vmbus_channel",
  "vmbus_ring",
@@ -6996,6 +7015,7 @@ version = "0.0.0"
 dependencies = [
  "arbitrary",
  "guid",
+ "inspect",
  "open_enum",
  "scsi_defs",
  "zerocopy 0.8.24",
@@ -7777,6 +7797,7 @@ dependencies = [
  "disk_blockdevice",
  "disk_get_vmgs",
  "disk_nvme",
+ "disk_storvsc",
  "firmware_pcat",
  "firmware_uefi",
  "firmware_uefi_custom_vars",
@@ -7843,7 +7864,9 @@ dependencies = [
  "sparse_mmap",
  "state_unit",
  "storage_string",
+ "storvsc_driver",
  "storvsp",
+ "storvsp_protocol",
  "storvsp_resources",
  "tee_call",
  "test_with_tracing",

Cargo.toml

@@ -268,6 +268,7 @@ disk_layered = { path = "vm/devices/storage/disk_layered" }
 disk_nvme = { path = "vm/devices/storage/disk_nvme" }
 disk_delay = { path = "vm/devices/storage/disk_delay" }
 disk_prwrap = { path = "vm/devices/storage/disk_prwrap" }
+disk_storvsc = { path = "vm/devices/storage/disk_storvsc" }
 disk_striped = { path = "vm/devices/storage/disk_striped" }
 disk_vhd1 = { path = "vm/devices/storage/disk_vhd1" }
 disk_vhdmp = { path = "vm/devices/storage/disk_vhdmp" }

openhcl/openhcl_boot/src/main.rs

@@ -158,7 +158,9 @@ fn build_kernel_command_line(
         "rdinit=/underhill-init",
         // Default to user-mode NVMe driver.
         "OPENHCL_NVME_VFIO=1",
-        // The next three items reduce the memory overhead of the storvsc driver.
+        // Default to user-mode storvsc driver.
+        "OPENHCL_STORVSC_USERMODE=1",
+        // The next three items reduce the memory overhead of the kernel storvsc driver.
         // Since it is only used for DVD, performance is not critical.
         "hv_storvsc.storvsc_vcpus_per_sub_channel=2048",
         // Fix number of hardware queues at 2.

openhcl/rootfs.config

@@ -35,7 +35,7 @@ file /lib/modules/001/pci-hyperv.ko         ${OPENHCL_KERNEL_PATH}/build/native/
 
 # Storvsc is last because it sometimes takes a long time to load and should not
 # block other device startup.
-file /lib/modules/999/hv_storvsc.ko         ${OPENHCL_KERNEL_PATH}/build/native/bin/${OPENHCL_KERNEL_ARCH}/modules/kernel/drivers/scsi/hv_storvsc.ko      0644 0 0
+#file /lib/modules/999/hv_storvsc.ko         ${OPENHCL_KERNEL_PATH}/build/native/bin/${OPENHCL_KERNEL_ARCH}/modules/kernel/drivers/scsi/hv_storvsc.ko      0644 0 0
 
 # These nodes are needed for early logging before devfs is mounted.
 nod /dev/null      0666 0 0 c 1  3

openhcl/underhill_core/Cargo.toml

@@ -48,6 +48,7 @@ disk_backend_resources.workspace = true
 disk_blockdevice.workspace = true
 disk_get_vmgs.workspace = true
 disk_nvme.workspace = true
+disk_storvsc.workspace = true
 firmware_uefi.workspace = true
 firmware_uefi_custom_vars.workspace = true
 hyperv_ic_guest.workspace = true
@@ -79,7 +80,9 @@ scsidisk.workspace = true
 scsidisk_resources.workspace = true
 serial_16550_resources.workspace = true
 storage_string.workspace = true
+storvsc_driver.workspace = true
 storvsp.workspace = true
+storvsp_protocol.workspace = true
 storvsp_resources.workspace = true
 tpm_resources.workspace = true
 tpm = { workspace = true, features = ["tpm"] }

openhcl/underhill_core/src/dispatch/mod.rs

@@ -16,6 +16,8 @@ use crate::reference_time::ReferenceTime;
 use crate::servicing;
 use crate::servicing::NvmeSavedState;
 use crate::servicing::ServicingState;
+use crate::servicing::StorvscSavedState;
+use crate::storvsc_manager::StorvscManager;
 use crate::vmbus_relay_unit::VmbusRelayHandle;
 use crate::worker::FirmwareType;
 use crate::worker::NetworkSettingsError;
@@ -144,6 +146,7 @@ pub(crate) struct LoadedVm {
     pub uevent_listener: Arc<UeventListener>,
     pub resolver: ResourceResolver,
     pub nvme_manager: Option<NvmeManager>,
+    pub storvsc_manager: Option<StorvscManager>,
     pub emuplat_servicing: EmuplatServicing,
     pub device_interfaces: Option<DeviceInterfaces>,
     pub vmbus_client: Option<vmbus_client::VmbusClient>,
@@ -323,6 +326,7 @@ impl LoadedVm {
                         resp.field("vmgs", self.vmgs.as_ref().map(|x| &x.0));
                         resp.field("network", &self.network_settings);
                         resp.field("nvme", &self.nvme_manager);
+                        resp.field("storvsc", &self.storvsc_manager);
                         resp.field("resolver", &self.resolver);
                         resp.field(
                             "vtl0_memory_map",
@@ -575,6 +579,13 @@ impl LoadedVm {
                 }
             };
 
+            // Reset all user-mode storvsc devices.
+            let shutdown_storvsc = async {
+                if let Some(storvsc_manager) = self.storvsc_manager.take() {
+                    storvsc_manager.shutdown().instrument(tracing::info_span!("shutdown_storvsc", CVM_ALLOWED, %correlation_id)).await;
+                }
+            };
+
             // Unbind drivers from the PCI devices to prepare for a kernel
             // restart.
             let shutdown_pci = async {
@@ -583,7 +594,7 @@ impl LoadedVm {
                     .await
             };
 
-            let (r, (), ()) = (shutdown_pci, shutdown_mana, shutdown_nvme).join().await;
+            let (r, (), (), ()) = (shutdown_pci, shutdown_mana, shutdown_nvme, shutdown_storvsc).join().await;
             r?;
 
             Ok(state)
@@ -731,6 +742,17 @@ impl LoadedVm {
             None
         };
 
+        let storvsc_state = if let Some(s) = &self.storvsc_manager {
+            s.save()
+                .instrument(tracing::info_span!("storvsc_manager_save", CVM_ALLOWED))
+                .await
+                .map(|state| StorvscSavedState {
+                    storvsc_state: state,
+                })
+        } else {
+            None
+        };
+
         let units = self.save_units().await.context("state unit save failed")?;
         let vmgs = if let Some((vmgs_thin_client, vmgs_disk_metadata, _)) = self.vmgs.as_ref() {
             Some((
@@ -770,6 +792,7 @@ impl LoadedVm {
                 nvme_state,
                 dma_manager_state,
                 vmbus_client,
+                storvsc_state,
             },
             units,
         };

openhcl/underhill_core/src/dispatch/vtl2_settings_worker.rs

@@ -5,6 +5,7 @@
 
 use super::LoadedVm;
 use crate::nvme_manager::manager::NvmeDiskConfig;
+use crate::storvsc_manager::StorvscDiskConfig;
 use crate::worker::NicConfig;
 use anyhow::Context;
 use cvm_tracing::CVM_ALLOWED;
@@ -243,6 +244,7 @@ pub struct DeviceInterfaces {
     scsi_dvds: HashMap<StorageDevicePath, mesh::Sender<SimpleScsiDvdRequest>>,
     scsi_request: HashMap<Guid, mesh::Sender<ScsiControllerRequest>>,
     use_nvme_vfio: bool,
+    use_storvsc_usermode: bool,
 }
 
 impl Vtl2SettingsWorker {
@@ -376,6 +378,7 @@ impl Vtl2SettingsWorker {
                         &StorageContext {
                             uevent_listener,
                             use_nvme_vfio: self.interfaces.use_nvme_vfio,
+                            use_storvsc_usermode: self.interfaces.use_storvsc_usermode,
                         },
                         &disk,
                         false,
@@ -394,6 +397,7 @@ impl Vtl2SettingsWorker {
                         &StorageContext {
                             uevent_listener,
                             use_nvme_vfio: self.interfaces.use_nvme_vfio,
+                            use_storvsc_usermode: self.interfaces.use_storvsc_usermode,
                         },
                         &disk,
                         false,
@@ -957,6 +961,7 @@ async fn make_disk_type_from_physical_devices(
 struct StorageContext<'a> {
     uevent_listener: &'a UeventListener,
     use_nvme_vfio: bool,
+    use_storvsc_usermode: bool,
 }
 
 #[instrument(skip_all, fields(CVM_ALLOWED))]
@@ -1000,6 +1005,36 @@ async fn make_disk_type_from_physical_device(
         }));
     }
 
+    // Special case for VScsi when using usermode storvsc.
+    if storage_context.use_storvsc_usermode
+        && matches!(
+            single_device.device_type,
+            underhill_config::DeviceType::VScsi
+        )
+    {
+        // Wait for the SCSI controller to arrive.
+        let devpath = uio_path(&controller_instance_id);
+        async {
+            ctx.until_cancelled(storage_context.uevent_listener.wait_for_devpath(&devpath))
+                .await??;
+            Ok(())
+        }
+        .await
+        .map_err(|err| Error::StorageCannotFindVtl2Device {
+            device_type: single_device.device_type,
+            instance_id: controller_instance_id,
+            sub_device_path,
+            source: err,
+        })?;
+
+        // Cannot validate device actually exists yet. Check this later
+        // TODO: Is this true?
+        return Ok(Resource::new(StorvscDiskConfig {
+            instance_guid: controller_instance_id,
+            lun: sub_device_path as u8,
+        }));
+    }
+
     // Wait for the device to arrive.
     let devname = async {
         let devname = ctx
@@ -1428,6 +1463,7 @@ pub async fn create_storage_controllers_from_vtl2_settings(
     ctx: &mut CancelContext,
     uevent_listener: &UeventListener,
     use_nvme_vfio: bool,
+    use_storvsc_usermode: bool,
     settings: &Vtl2SettingsDynamic,
     sub_channels: u16,
     is_restoring: bool,
@@ -1443,6 +1479,7 @@ pub async fn create_storage_controllers_from_vtl2_settings(
     let storage_context = StorageContext {
         uevent_listener,
         use_nvme_vfio,
+        use_storvsc_usermode,
     };
     let ide_controller =
         make_ide_controller_config(ctx, &storage_context, settings, is_restoring).await?;
@@ -1643,6 +1680,11 @@ fn vpci_path(instance_id: &Guid) -> (String, PathBuf) {
     (pci_id, devpath)
 }
 
+fn uio_path(instance_id: &Guid) -> PathBuf {
+    // TODO: Is this enough?
+    PathBuf::from(format!("/sys/bus/vmbus/devices/{instance_id}/uio"))
+}
+
 /// Waits for the PCI path to get populated. The PCI path is just a symlink to the actual
 /// device path. This should be called once the device path is available.
 pub async fn wait_for_pci_path(pci_id: &String) {
@@ -1782,6 +1824,7 @@ impl InitialControllers {
         uevent_listener: &UeventListener,
         dps: &DevicePlatformSettings,
         use_nvme_vfio: bool,
+        use_storvsc_usermode: bool,
         is_restoring: bool,
         default_io_queue_depth: u32,
     ) -> anyhow::Result<Self> {
@@ -1801,6 +1844,7 @@ impl InitialControllers {
                 &mut context,
                 uevent_listener,
                 use_nvme_vfio,
+                use_storvsc_usermode,
                 dynamic,
                 fixed.scsi_sub_channels,
                 is_restoring,
@@ -1851,6 +1895,7 @@ impl InitialControllers {
                 scsi_dvds,
                 scsi_request,
                 use_nvme_vfio,
+                use_storvsc_usermode,
             },
         };
 

openhcl/underhill_core/src/lib.rs

@@ -20,6 +20,7 @@ mod nvme_manager;
 mod options;
 mod reference_time;
 mod servicing;
+mod storvsc_manager;
 mod threadpool_vm_task_backend;
 mod vmbus_relay_unit;
 mod vmgs_logger;
@@ -330,6 +331,7 @@ async fn launch_workers(
         disable_uefi_frontpage: opt.disable_uefi_frontpage,
         guest_state_encryption_policy: opt.guest_state_encryption_policy,
         attempt_ak_cert_callback: opt.attempt_ak_cert_callback,
+        storvsc_usermode: opt.storvsc_usermode,
     };
 
     let (mut remote_console_cfg, framebuffer_access) =

openhcl/underhill_core/src/options.rs

@@ -187,6 +187,10 @@ pub struct Options {
     /// (HCL_ATTEMPT_AK_CERT_CALLBACK=1) Attempt to renew the AK cert.
     /// If not specified, use the configuration in DPSv2 ManagementVtlFeatures.
     pub attempt_ak_cert_callback: Option<bool>,
+
+    /// (OPENHCL_STORVSC_USERMODE=1)
+    /// Use the user-mode storvsc driver instead of the Linux driver.
+    pub storvsc_usermode: bool,
 }
 
 impl Options {
@@ -322,6 +326,7 @@ impl Options {
             .map_err(|e| tracing::warn!("failed to parse HCL_ATTEMPT_AK_CERT_CALLBACK: {:#}", e))
             .ok()
             .flatten();
+        let storvsc_usermode = parse_env_bool("OPENHCL_STORVSC_USERMODE");
 
         let mut args = std::env::args().chain(extra_args);
         // Skip our own filename.
@@ -381,6 +386,7 @@ impl Options {
             disable_uefi_frontpage,
             guest_state_encryption_policy,
             attempt_ak_cert_callback,
+            storvsc_usermode,
         })
     }