virtio-devices: Attach and detach endpoints from domain properly

Properly detach a device from a domain if that device is already
attached to another domain on an attach request (following section
5.13.6.3.2 of the virtio-iommu spec). Resolves nested virtualization
reboot.

Signed-off-by: Andrew Carp <acarp@crusoeenergy.com>

pci/src/lib.rs

@@ -26,7 +26,7 @@ pub use self::device::{
 };
 pub use self::msi::{msi_num_enabled_vectors, MsiCap, MsiConfig};
 pub use self::msix::{MsixCap, MsixConfig, MsixTableEntry, MSIX_CONFIG_ID, MSIX_TABLE_ENTRY_SIZE};
-pub use self::vfio::{VfioPciDevice, VfioPciError};
+pub use self::vfio::{MmioRegion, VfioDmaMapping, VfioPciDevice, VfioPciError};
 pub use self::vfio_user::{VfioUserDmaMapping, VfioUserPciDevice, VfioUserPciDeviceError};
 use serde::de::Visitor;
 use std::fmt::{self, Display};

pci/src/vfio.rs

@@ -32,11 +32,12 @@ use vm_allocator::page_size::{
     align_page_size_down, align_page_size_up, is_4k_aligned, is_4k_multiple, is_page_size_aligned,
 };
 use vm_allocator::{AddressAllocator, SystemAllocator};
+use vm_device::dma_mapping::ExternalDmaMapping;
 use vm_device::interrupt::{
     InterruptIndex, InterruptManager, InterruptSourceGroup, MsiIrqGroupConfig,
 };
 use vm_device::{BusDevice, Resource};
-use vm_memory::{Address, GuestAddress, GuestUsize};
+use vm_memory::{Address, GuestAddress, GuestAddressSpace, GuestMemory, GuestUsize};
 use vm_migration::{
     Migratable, MigratableError, Pausable, Snapshot, Snapshottable, Transportable, VersionMapped,
 };
@@ -274,6 +275,48 @@ pub struct MmioRegion {
     pub(crate) index: u32,
     pub(crate) user_memory_regions: Vec<UserMemoryRegion>,
 }
+
+trait MmioRegionRange {
+    fn check_range(&self, guest_addr: u64, size: u64) -> bool;
+    fn find_user_address(&self, guest_addr: u64) -> Result<u64, io::Error>;
+}
+
+impl MmioRegionRange for Vec<MmioRegion> {
+    // Check if a guest address is within the range of mmio regions
+    fn check_range(&self, guest_addr: u64, size: u64) -> bool {
+        for region in self.iter() {
+            let Some(guest_addr_end) = guest_addr.checked_add(size) else {
+                return false;
+            };
+            let Some(region_end) = region.start.raw_value().checked_add(region.length) else {
+                return false;
+            };
+            if guest_addr >= region.start.raw_value() && guest_addr_end <= region_end {
+                return true;
+            }
+        }
+        false
+    }
+
+    // Locate the user region address for a guest address within all mmio regions
+    fn find_user_address(&self, guest_addr: u64) -> Result<u64, io::Error> {
+        for region in self.iter() {
+            for user_region in region.user_memory_regions.iter() {
+                if guest_addr >= user_region.start
+                    && guest_addr < user_region.start + user_region.size
+                {
+                    return Ok(user_region.host_addr + (guest_addr - user_region.start));
+                }
+            }
+        }
+
+        Err(io::Error::new(
+            io::ErrorKind::Other,
+            format!("unable to find user address: 0x{guest_addr:x}"),
+        ))
+    }
+}
+
 #[derive(Debug, Error)]
 pub enum VfioError {
     #[error("Kernel VFIO error: {0}")]
@@ -1885,3 +1928,80 @@ impl Snapshottable for VfioPciDevice {
 }
 impl Transportable for VfioPciDevice {}
 impl Migratable for VfioPciDevice {}
+
+/// This structure implements the ExternalDmaMapping trait. It is meant to
+/// be used when the caller tries to provide a way to update the mappings
+/// associated with a specific VFIO container.
+pub struct VfioDmaMapping<M: GuestAddressSpace> {
+    container: Arc<VfioContainer>,
+    memory: Arc<M>,
+    mmio_regions: Arc<Mutex<Vec<MmioRegion>>>,
+}
+
+impl<M: GuestAddressSpace> VfioDmaMapping<M> {
+    /// Create a DmaMapping object.
+    /// # Parameters
+    /// * `container`: VFIO container object.
+    /// * `memory`: guest memory to mmap.
+    /// * `mmio_regions`: mmio_regions to mmap.
+    pub fn new(
+        container: Arc<VfioContainer>,
+        memory: Arc<M>,
+        mmio_regions: Arc<Mutex<Vec<MmioRegion>>>,
+    ) -> Self {
+        VfioDmaMapping {
+            container,
+            memory,
+            mmio_regions,
+        }
+    }
+}
+
+impl<M: GuestAddressSpace + Sync + Send> ExternalDmaMapping for VfioDmaMapping<M> {
+    fn map(&self, iova: u64, gpa: u64, size: u64) -> std::result::Result<(), io::Error> {
+        let mem = self.memory.memory();
+        let guest_addr = GuestAddress(gpa);
+        let user_addr = if mem.check_range(guest_addr, size as usize) {
+            match mem.get_host_address(guest_addr) {
+                Ok(t) => t as u64,
+                Err(e) => {
+                    return Err(io::Error::new(
+                        io::ErrorKind::Other,
+                        format!("unable to retrieve user address for gpa 0x{gpa:x} from guest memory region: {e}")
+                    ));
+                }
+            }
+        } else if self.mmio_regions.lock().unwrap().check_range(gpa, size) {
+            self.mmio_regions.lock().unwrap().find_user_address(gpa)?
+        } else {
+            return Err(io::Error::new(
+                io::ErrorKind::Other,
+                format!("failed to locate guest address 0x{gpa:x} in guest memory"),
+            ));
+        };
+
+        self.container
+            .vfio_dma_map(iova, size, user_addr)
+            .map_err(|e| {
+                io::Error::new(
+                    io::ErrorKind::Other,
+                    format!(
+                        "failed to map memory for VFIO container, \
+                         iova 0x{iova:x}, gpa 0x{gpa:x}, size 0x{size:x}: {e:?}"
+                    ),
+                )
+            })
+    }
+
+    fn unmap(&self, iova: u64, size: u64) -> std::result::Result<(), io::Error> {
+        self.container.vfio_dma_unmap(iova, size).map_err(|e| {
+            io::Error::new(
+                io::ErrorKind::Other,
+                format!(
+                    "failed to unmap memory for VFIO container, \
+                     iova 0x{iova:x}, size 0x{size:x}: {e:?}"
+                ),
+            )
+        })
+    }
+}

virtio-devices/src/iommu.rs

@@ -407,6 +407,15 @@ impl Request {
                     let bypass =
                         (req.flags & VIRTIO_IOMMU_ATTACH_F_BYPASS) == VIRTIO_IOMMU_ATTACH_F_BYPASS;
 
+                    let mut old_domain_id = domain_id;
+                    if let Some(&id) = mapping.endpoints.read().unwrap().get(&endpoint) {
+                        old_domain_id = id;
+                    }
+
+                    if old_domain_id != domain_id {
+                        detach_endpoint_from_domain(endpoint, old_domain_id, mapping, ext_mapping)?;
+                    }
+
                     // Add endpoint associated with specific domain
                     mapping
                         .endpoints
@@ -414,6 +423,19 @@ impl Request {
                         .unwrap()
                         .insert(endpoint, domain_id);
 
+                    // If any other mappings exist in the domain for other containers,
+                    // make sure to issue these mappings for the new endpoint/container
+                    if let Some(domain_mappings) = &mapping.domains.read().unwrap().get(&domain_id)
+                    {
+                        if let Some(ext_map) = ext_mapping.get(&endpoint) {
+                            for (virt_start, addr_map) in &domain_mappings.mappings {
+                                ext_map
+                                    .map(*virt_start, addr_map.gpa, addr_map.size)
+                                    .map_err(Error::ExternalUnmapping)?;
+                            }
+                        }
+                    }
+
                     // Add new domain with no mapping if the entry didn't exist yet
                     let mut domains = mapping.domains.write().unwrap();
                     let domain = Domain {
@@ -439,22 +461,7 @@ impl Request {
                     let endpoint = req.endpoint;
 
                     // Remove endpoint associated with specific domain
-                    mapping.endpoints.write().unwrap().remove(&endpoint);
+                    detach_endpoint_from_domain(endpoint, domain_id, mapping, ext_mapping)?;
-
-                    // After all endpoints have been successfully detached from a
-                    // domain, the domain can be removed. This means we must remove
-                    // the mappings associated with this domain.
-                    if mapping
-                        .endpoints
-                        .write()
-                        .unwrap()
-                        .iter()
-                        .filter(|(_, &d)| d == domain_id)
-                        .count()
-                        == 0
-                    {
-                        mapping.domains.write().unwrap().remove(&domain_id);
-                    }
                 }
                 VIRTIO_IOMMU_T_MAP => {
                     if desc_size_left != size_of::<VirtioIommuReqMap>() {
@@ -491,7 +498,9 @@ impl Request {
                         .map(|(&e, _)| e)
                         .collect();
 
-                    // Trigger external mapping if necessary.
+                    // For viommu all endpoints receive their own VFIO container, as a result
+                    // Each endpoint within the domain needs to be separately mapped, as the
+                    // mapping is done on a per-container level, not a per-domain level
                     for endpoint in endpoints {
                         if let Some(ext_map) = ext_mapping.get(&endpoint) {
                             let size = req.virt_end - req.virt_start + 1;
@@ -640,6 +649,41 @@ impl Request {
     }
 }
 
+fn detach_endpoint_from_domain(
+    endpoint: u32,
+    domain_id: u32,
+    mapping: &Arc<IommuMapping>,
+    ext_mapping: &BTreeMap<u32, Arc<dyn ExternalDmaMapping>>,
+) -> result::Result<(), Error> {
+    // Remove endpoint associated with specific domain
+    mapping.endpoints.write().unwrap().remove(&endpoint);
+
+    // Trigger external unmapping for the endpoint if necessary.
+    if let Some(domain_mappings) = &mapping.domains.read().unwrap().get(&domain_id) {
+        if let Some(ext_map) = ext_mapping.get(&endpoint) {
+            for (virt_start, addr_map) in &domain_mappings.mappings {
+                ext_map
+                    .unmap(*virt_start, addr_map.size)
+                    .map_err(Error::ExternalUnmapping)?;
+            }
+        }
+    }
+
+    if mapping
+        .endpoints
+        .write()
+        .unwrap()
+        .iter()
+        .filter(|(_, &d)| d == domain_id)
+        .count()
+        == 0
+    {
+        mapping.domains.write().unwrap().remove(&domain_id);
+    }
+
+    Ok(())
+}
+
 struct IommuEpollHandler {
     mem: GuestMemoryAtomic<GuestMemoryMmap>,
     request_queue: Queue,

vm-device/src/dma_mapping/mod.rs

@@ -2,8 +2,6 @@
 //
 // SPDX-License-Identifier: Apache-2.0 OR BSD-3-Clause
 
-pub mod vfio;
-
 /// Trait meant for triggering the DMA mapping update related to an external
 /// device not managed fully through virtio. It is dedicated to virtio-iommu
 /// in order to trigger the map update anytime the mapping is updated from the

vm-device/src/dma_mapping/vfio.rs

@@ -1,70 +0,0 @@
-// Copyright © 2021 Intel Corporation
-//
-// SPDX-License-Identifier: Apache-2.0 OR BSD-3-Clause
-
-use crate::dma_mapping::ExternalDmaMapping;
-use std::io;
-use std::sync::Arc;
-use vfio_ioctls::VfioContainer;
-use vm_memory::{GuestAddress, GuestAddressSpace, GuestMemory};
-
-/// This structure implements the ExternalDmaMapping trait. It is meant to
-/// be used when the caller tries to provide a way to update the mappings
-/// associated with a specific VFIO container.
-pub struct VfioDmaMapping<M: GuestAddressSpace> {
-    container: Arc<VfioContainer>,
-    memory: Arc<M>,
-}
-
-impl<M: GuestAddressSpace> VfioDmaMapping<M> {
-    /// Create a DmaMapping object.
-    ///
-    /// # Parameters
-    /// * `container`: VFIO container object.
-    /// * `memory·: guest memory to mmap.
-    pub fn new(container: Arc<VfioContainer>, memory: Arc<M>) -> Self {
-        VfioDmaMapping { container, memory }
-    }
-}
-
-impl<M: GuestAddressSpace + Sync + Send> ExternalDmaMapping for VfioDmaMapping<M> {
-    fn map(&self, iova: u64, gpa: u64, size: u64) -> std::result::Result<(), io::Error> {
-        let mem = self.memory.memory();
-        let guest_addr = GuestAddress(gpa);
-        let user_addr = if mem.check_range(guest_addr, size as usize) {
-            mem.get_host_address(guest_addr).unwrap() as u64
-        } else {
-            return Err(io::Error::new(
-                io::ErrorKind::Other,
-                format!(
-                    "failed to convert guest address 0x{gpa:x} into \
-                     host user virtual address"
-                ),
-            ));
-        };
-
-        self.container
-            .vfio_dma_map(iova, size, user_addr)
-            .map_err(|e| {
-                io::Error::new(
-                    io::ErrorKind::Other,
-                    format!(
-                        "failed to map memory for VFIO container, \
-                         iova 0x{iova:x}, gpa 0x{gpa:x}, size 0x{size:x}: {e:?}"
-                    ),
-                )
-            })
-    }
-
-    fn unmap(&self, iova: u64, size: u64) -> std::result::Result<(), io::Error> {
-        self.container.vfio_dma_unmap(iova, size).map_err(|e| {
-            io::Error::new(
-                io::ErrorKind::Other,
-                format!(
-                    "failed to unmap memory for VFIO container, \
-                     iova 0x{iova:x}, size 0x{size:x}: {e:?}"
-                ),
-            )
-        })
-    }
-}

vmm/src/device_manager.rs

@@ -58,8 +58,8 @@ use libc::{
     O_TMPFILE, PROT_READ, PROT_WRITE, TCSANOW,
 };
 use pci::{
-    DeviceRelocation, PciBarRegionType, PciBdf, PciDevice, VfioPciDevice, VfioUserDmaMapping,
+    DeviceRelocation, MmioRegion, PciBarRegionType, PciBdf, PciDevice, VfioDmaMapping,
-    VfioUserPciDevice, VfioUserPciDeviceError,
+    VfioPciDevice, VfioUserDmaMapping, VfioUserPciDevice, VfioUserPciDeviceError,
 };
 use rate_limiter::group::RateLimiterGroup;
 use seccompiler::SeccompAction;
@@ -85,7 +85,6 @@ use virtio_devices::{
 };
 use virtio_devices::{Endpoint, IommuMapping};
 use vm_allocator::{AddressAllocator, SystemAllocator};
-use vm_device::dma_mapping::vfio::VfioDmaMapping;
 use vm_device::dma_mapping::ExternalDmaMapping;
 use vm_device::interrupt::{
     InterruptIndex, InterruptManager, LegacyIrqGroupConfig, MsiIrqGroupConfig,
@@ -966,6 +965,8 @@ pub struct DeviceManager {
     snapshot: Option<Snapshot>,
 
     rate_limit_groups: HashMap<String, Arc<RateLimiterGroup>>,
+
+    mmio_regions: Arc<Mutex<Vec<MmioRegion>>>,
 }
 
 impl DeviceManager {
@@ -1196,6 +1197,7 @@ impl DeviceManager {
             acpi_platform_addresses: AcpiPlatformAddresses::default(),
             snapshot,
             rate_limit_groups,
+            mmio_regions: Arc::new(Mutex::new(Vec::new())),
         };
 
         let device_manager = Arc::new(Mutex::new(device_manager));
@@ -3424,6 +3426,7 @@ impl DeviceManager {
             let vfio_mapping = Arc::new(VfioDmaMapping::new(
                 Arc::clone(&vfio_container),
                 Arc::new(self.memory_manager.lock().unwrap().guest_memory()),
+                Arc::clone(&self.mmio_regions),
             ));
 
             if let Some(iommu) = &self.iommu_device {
@@ -3468,6 +3471,7 @@ impl DeviceManager {
             let vfio_mapping = Arc::new(VfioDmaMapping::new(
                 Arc::clone(&vfio_container),
                 Arc::new(self.memory_manager.lock().unwrap().guest_memory()),
+                Arc::clone(&self.mmio_regions),
             ));
 
             for virtio_mem_device in self.virtio_mem_devices.iter() {
@@ -3530,6 +3534,10 @@ impl DeviceManager {
             .map_mmio_regions()
             .map_err(DeviceManagerError::VfioMapRegion)?;
 
+        for mmio_region in vfio_pci_device.lock().unwrap().mmio_regions() {
+            self.mmio_regions.lock().unwrap().push(mmio_region);
+        }
+
         let mut node = device_node!(vfio_name, vfio_pci_device);
 
         // Update the device tree with correct resource information.
@@ -4201,12 +4209,21 @@ impl DeviceManager {
 
         let (pci_device, bus_device, virtio_device, remove_dma_handler) = match pci_device_handle {
             // No need to remove any virtio-mem mapping here as the container outlives all devices
-            PciDeviceHandle::Vfio(vfio_pci_device) => (
+            PciDeviceHandle::Vfio(vfio_pci_device) => {
+                for mmio_region in vfio_pci_device.lock().unwrap().mmio_regions() {
+                    self.mmio_regions
+                        .lock()
+                        .unwrap()
+                        .retain(|x| x.start != mmio_region.start)
+                }
+
+                (
                     Arc::clone(&vfio_pci_device) as Arc<Mutex<dyn PciDevice>>,
                     Arc::clone(&vfio_pci_device) as Arc<Mutex<dyn BusDevice>>,
                     None as Option<Arc<Mutex<dyn virtio_devices::VirtioDevice>>>,
                     false,
-            ),
+                )
+            }
             PciDeviceHandle::Virtio(virtio_pci_device) => {
                 let dev = virtio_pci_device.lock().unwrap();
                 let bar_addr = dev.config_bar_addr();