dragonball: add seccomp support for dragonball

This commit modifies seccomp framework to
support different restrictions for different threads.

Signed-off-by: wangxinge <wangxinge@bupt.edu.cn>

src/dragonball/Cargo.lock

@@ -344,20 +344,26 @@ name = "dbs-pci"
 version = "0.1.0"
 dependencies = [
  "byteorder",
+ "dbs-address-space",
  "dbs-allocator",
  "dbs-arch",
  "dbs-boot",
  "dbs-device",
  "dbs-interrupt",
+ "dbs-utils",
+ "dbs-virtio-devices",
  "downcast-rs",
  "kvm-bindings",
  "kvm-ioctls",
  "libc",
  "log",
+ "serde",
  "thiserror 1.0.48",
  "vfio-bindings",
  "vfio-ioctls",
+ "virtio-queue",
  "vm-memory",
+ "vmm-sys-util",
 ]
 
 [[package]]
@@ -1810,9 +1816,9 @@ checksum = "94143f37725109f92c262ed2cf5e59bce7498c01bcc1502d7b9afe439a4e9f49"
 
 [[package]]
 name = "seccompiler"
-version = "0.2.0"
+version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e01d1292a1131b22ccea49f30bd106f1238b5ddeec1a98d39268dcc31d540e68"
+checksum = "a4ae55de56877481d112a559bbc12667635fdaf5e005712fd4e2b2fa50ffc884"
 dependencies = [
  "libc",
 ]

src/dragonball/Cargo.toml

@@ -33,7 +33,7 @@ event-manager = "0.2.1"
 kvm-bindings = "0.6.0"
 kvm-ioctls = "0.12.0"
 linux-loader = "0.8.0"
-seccompiler = "0.2.0"
+seccompiler = "0.5.0"
 vfio-bindings = "0.3.0"
 vfio-ioctls = "0.1.0"
 virtio-bindings = "0.1.0"

src/dragonball/src/api/v1/vmm_action.rs

@@ -479,14 +479,13 @@ impl VmmService {
         use self::StartMicroVmError::MicroVMAlreadyRunning;
         use self::VmmActionError::StartMicroVm;
 
-        let vmm_seccomp_filter = vmm.vmm_seccomp_filter();
-        let vcpu_seccomp_filter = vmm.vcpu_seccomp_filter();
+        let seccomp_filters = vmm.seccomp_filters();
         let vm = vmm.get_vm_mut().ok_or(VmmActionError::InvalidVMID)?;
         if vm.is_vm_initialized() {
             return Err(StartMicroVm(MicroVMAlreadyRunning));
         }
 
-        vm.start_microvm(event_mgr, vmm_seccomp_filter, vcpu_seccomp_filter)
+        vm.start_microvm(event_mgr, seccomp_filters)
             .map(|_| VmmData::Empty)
             .map_err(StartMicroVm)
     }

src/dragonball/src/error.rs

@@ -219,6 +219,10 @@ pub enum StartMicroVmError {
     /// Failed to register DMA memory address range.
     #[error("failure while registering DMA address range: {0:?}")]
     RegisterDMAAddress(#[source] VfioDeviceError),
+
+    /// Cannot build seccomp filters.
+    #[error("failure while configuring seccomp filters: {0}")]
+    SeccompFilters(#[source] seccompiler::Error),
 }
 
 /// Errors associated with starting the instance.

src/dragonball/src/lib.rs

@@ -48,7 +48,7 @@ mod vmm;
 
 pub use self::error::StartMicroVmError;
 pub use self::io_manager::IoManagerCached;
-pub use self::vmm::Vmm;
+pub use self::vmm::{Vmm, ALL_THREADS, VCPU_THREAD, VMM_THREAD};
 
 /// Success exit code.
 pub const EXIT_CODE_OK: u8 = 0;

src/dragonball/src/vm/mod.rs

@@ -1,6 +1,7 @@
 // Copyright (C) 2021 Alibaba Cloud. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
+use std::collections::HashMap;
 use std::io::{self, Read, Seek, SeekFrom};
 use std::ops::Deref;
 use std::os::unix::io::RawFd;
@@ -18,6 +19,7 @@ use dbs_utils::time::TimestampUs;
 use kvm_ioctls::VmFd;
 use linux_loader::loader::{KernelLoader, KernelLoaderResult};
 use seccompiler::BpfProgram;
+use seccompiler::{apply_filter_all_threads, Error as SecError};
 use serde_derive::{Deserialize, Serialize};
 use slog::{error, info};
 use vm_memory::{Bytes, GuestAddress, GuestAddressSpace};
@@ -42,6 +44,7 @@ use crate::resource_manager::ResourceManager;
 use crate::vcpu::{VcpuManager, VcpuManagerError};
 #[cfg(feature = "hotplug")]
 use crate::vcpu::{VcpuResizeError, VcpuResizeInfo};
+use crate::{ALL_THREADS, VCPU_THREAD, VMM_THREAD};
 #[cfg(target_arch = "aarch64")]
 use dbs_arch::gic::Error as GICError;
 
@@ -709,10 +712,27 @@ impl Vm {
     pub fn start_microvm(
         &mut self,
         event_mgr: &mut EventManager,
-        vmm_seccomp_filter: BpfProgram,
-        vcpu_seccomp_filter: BpfProgram,
+        seccomp_filters: HashMap<String, BpfProgram>,
     ) -> std::result::Result<(), StartMicroVmError> {
         info!(self.logger, "VM: received instance start command");
+
+        if let Some(process_seccomp_filter) = seccomp_filters.get(ALL_THREADS) {
+            // Load seccomp filters for the whole process.
+            // Execution panics if filters cannot be loaded, use --seccomp-level=0 if skipping filters
+            // altogether is the desired behaviour.
+            if let Err(e) = apply_filter_all_threads(process_seccomp_filter) {
+                if !matches!(e, SecError::EmptyFilter) {
+                    error!(
+                        self.logger,
+                        "VM: failed to apply process-wide seccomp filters: {}", e
+                    );
+                    return Err(StartMicroVmError::SeccompFilters(e));
+                }
+            } else {
+                info!(self.logger, "VM: process-wide seccomp filters applied");
+            }
+        }
+
         if self.is_vm_initialized() {
             return Err(StartMicroVmError::MicroVMAlreadyRunning);
         }
@@ -738,8 +758,14 @@ impl Vm {
                 AddressManagerError::GuestMemoryNotInitialized,
             ))?;
 
-        self.init_vcpu_manager(vm_as.clone(), vcpu_seccomp_filter)
-            .map_err(StartMicroVmError::Vcpu)?;
+        self.init_vcpu_manager(
+            vm_as.clone(),
+            seccomp_filters
+                .get(VCPU_THREAD)
+                .cloned()
+                .unwrap_or_default(),
+        )
+        .map_err(StartMicroVmError::Vcpu)?;
         self.init_microvm(event_mgr.epoll_manager(), vm_as.clone(), request_ts)?;
         self.init_configure_system(&vm_as)?;
         #[cfg(feature = "dbs-upcall")]
@@ -751,7 +777,7 @@ impl Vm {
         info!(self.logger, "VM: start vcpus");
         self.vcpu_manager()
             .map_err(StartMicroVmError::Vcpu)?
-            .start_boot_vcpus(vmm_seccomp_filter)
+            .start_boot_vcpus(seccomp_filters.get(VMM_THREAD).cloned().unwrap_or_default())
             .map_err(StartMicroVmError::Vcpu)?;
 
         // Use expect() to crash if the other thread poisoned this lock.

src/dragonball/src/vmm.rs

@@ -6,6 +6,7 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the THIRD-PARTY file.
 
+use std::collections::HashMap;
 use std::fmt::Formatter;
 use std::os::unix::io::RawFd;
 use std::sync::{Arc, Mutex, RwLock};
@@ -22,6 +23,19 @@ use crate::event_manager::{EventContext, EventManager};
 use crate::vm::Vm;
 use crate::{EXIT_CODE_GENERIC_ERROR, EXIT_CODE_OK};
 
+// Thread types to which the seccomp restrictions apply
+// Currently the restrictions are applied to the whole process
+// More thread types can be added later to accommodate more granular constraints.
+
+/// Union of restrictions for all threads;
+/// The corresponding restrictions are applied to the whole process
+/// This type should be used with any other thread type which has specific restrictions
+pub const ALL_THREADS: &str = "all";
+/// VMM thread
+pub const VMM_THREAD: &str = "vmm";
+/// Vcpu thread
+pub const VCPU_THREAD: &str = "vcpu";
+
 /// Global coordinator to manage API servers, virtual machines, upgrade etc.
 ///
 /// Originally firecracker assumes an VMM only manages an VM, and doesn't distinguish VMM and VM.
@@ -41,8 +55,7 @@ pub struct Vmm {
     // Will change to a HashMap when enabling 1 VMM with multiple VMs.
     vm: Vm,
 
-    vcpu_seccomp_filter: BpfProgram,
-    vmm_seccomp_filter: BpfProgram,
+    seccomp_filters: HashMap<String, BpfProgram>,
 }
 
 impl Vmm {
@@ -50,8 +63,7 @@ impl Vmm {
     pub fn new(
         api_shared_info: Arc<RwLock<InstanceInfo>>,
         api_event_fd: EventFd,
-        vmm_seccomp_filter: BpfProgram,
-        vcpu_seccomp_filter: BpfProgram,
+        seccomp_filters: HashMap<String, BpfProgram>,
         kvm_fd: Option<RawFd>,
     ) -> Result<Self> {
         let epoll_manager = EpollManager::default();
@@ -59,8 +71,7 @@ impl Vmm {
             api_shared_info,
             api_event_fd,
             epoll_manager,
-            vmm_seccomp_filter,
-            vcpu_seccomp_filter,
+            seccomp_filters,
             kvm_fd,
         )
     }
@@ -70,8 +81,7 @@ impl Vmm {
         api_shared_info: Arc<RwLock<InstanceInfo>>,
         api_event_fd: EventFd,
         epoll_manager: EpollManager,
-        vmm_seccomp_filter: BpfProgram,
-        vcpu_seccomp_filter: BpfProgram,
+        seccomp_filters: HashMap<String, BpfProgram>,
         kvm_fd: Option<RawFd>,
     ) -> Result<Self> {
         let vm = Vm::new(kvm_fd, api_shared_info, epoll_manager.clone())?;
@@ -81,8 +91,7 @@ impl Vmm {
             event_ctx,
             epoll_manager,
             vm,
-            vcpu_seccomp_filter,
-            vmm_seccomp_filter,
+            seccomp_filters,
         })
     }
 
@@ -96,14 +105,9 @@ impl Vmm {
         Some(&mut self.vm)
     }
 
-    /// Get the seccomp rules for vCPU threads.
-    pub fn vcpu_seccomp_filter(&self) -> BpfProgram {
-        self.vcpu_seccomp_filter.clone()
-    }
-
-    /// Get the seccomp rules for VMM threads.
-    pub fn vmm_seccomp_filter(&self) -> BpfProgram {
-        self.vmm_seccomp_filter.clone()
+    /// Get all seccomp rules.
+    pub fn seccomp_filters(&self) -> HashMap<String, BpfProgram> {
+        self.seccomp_filters.clone()
     }
 
     /// Run the event loop to service API requests.
@@ -206,14 +210,15 @@ impl std::fmt::Debug for Vmm {
         f.debug_struct("Vmm")
             .field("event_ctx", &self.event_ctx)
             .field("vm", &self.vm.shared_info())
-            .field("vcpu_seccomp_filter", &self.vcpu_seccomp_filter)
-            .field("vmm_seccomp_filter", &self.vmm_seccomp_filter)
+            .field("seccomp_filters", &self.seccomp_filters)
             .finish()
     }
 }
 
 #[cfg(test)]
 pub(crate) mod tests {
+    use std::collections::HashMap;
+
     use test_utils::skip_if_not_root;
 
     use super::*;
@@ -221,17 +226,8 @@ pub(crate) mod tests {
     pub fn create_vmm_instance(epoll_manager: EpollManager) -> Vmm {
         let info = Arc::new(RwLock::new(InstanceInfo::default()));
         let event_fd = EventFd::new(libc::EFD_NONBLOCK).unwrap();
-        let seccomp_filter: BpfProgram = Vec::new();
 
-        Vmm::new_with_epoll_manager(
-            info,
-            event_fd,
-            epoll_manager,
-            seccomp_filter.clone(),
-            seccomp_filter,
-            None,
-        )
-        .unwrap()
+        Vmm::new_with_epoll_manager(info, event_fd, epoll_manager, HashMap::new(), None).unwrap()
     }
 
     #[test]