add id mount

Signed-off-by: lengrongfu <lenronfu@gmail.com>

crates/libcgroups/Cargo.toml

@@ -22,7 +22,8 @@ cgroupsv2_devices = ["rbpf", "libbpf-sys", "errno", "libc", "nix/dir"]
 [dependencies]
 nix = { version = "0.28.0", features = ["signal", "user", "fs"] }
 procfs = "0.16.0"
-oci-spec = { version = "~0.6.4", features = ["runtime"] }
+#oci-spec = { version = "~0.6.4", features = ["runtime"] }
+oci-spec={git="https://github.com/lengrongfu/oci-spec-rs",branch = "feat/add_idmapped_mount",features = ["runtime"] }
 fixedbitset = "0.5.7"
 serde = { version = "1.0", features = ["derive"] }
 rbpf = { version = "0.2.0", optional = true }
@@ -34,7 +35,8 @@ tracing = { version = "0.1.40", features = ["attributes"] }
 
 [dev-dependencies]
 anyhow = "1.0"
-oci-spec = { version = "~0.6.4", features = ["proptests", "runtime"] }
+#oci-spec = { version = "~0.6.4", features = ["proptests", "runtime"] }
+oci-spec={git="https://github.com/lengrongfu/oci-spec-rs",branch = "feat/add_idmapped_mount",features = ["proptests", "runtime"]}
 quickcheck = "1"
 mockall = { version = "0.12.1", features = [] }
 clap = "4.1.6"

crates/libcontainer/Cargo.toml

@@ -39,7 +39,8 @@ nix = { version = "0.28.0", features = [
     "term",
     "hostname",
 ] }
-oci-spec = { version = "~0.6.4", features = ["runtime"] }
+#oci-spec = { version = "~0.6.4", features = ["runtime"] }
+oci-spec={git="https://github.com/lengrongfu/oci-spec-rs",branch = "feat/add_idmapped_mount",features = ["runtime"] }
 once_cell = "1.19.0"
 procfs = "0.16.0"
 prctl = "1.0.0"
@@ -56,7 +57,8 @@ safe-path = "0.1.0"
 nc = "0.8.20"
 
 [dev-dependencies]
-oci-spec = { version = "~0.6.4", features = ["proptests", "runtime"] }
+#oci-spec = { version = "~0.6.4", features = ["proptests", "runtime"] }
+oci-spec={git="https://github.com/lengrongfu/oci-spec-rs",branch = "feat/add_idmapped_mount",features = ["proptests", "runtime"]}
 quickcheck = "1"
 serial_test = "3.1.1"
 tempfile = "3"

crates/libcontainer/src/process/channel.rs

@@ -4,6 +4,7 @@ use nix::unistd::Pid;
 
 use crate::channel::{channel, Receiver, Sender};
 use crate::process::message::Message;
+use crate::rootfs::mount::{IdMountParam, IdMountSource};
 
 #[derive(Debug, thiserror::Error)]
 pub enum ChannelError {
@@ -57,6 +58,13 @@ impl MainSender {
         Ok(())
     }
 
+    // process mount place
+    pub fn process_mount_place(&mut self, m: IdMountParam) -> Result<(), ChannelError> {
+        tracing::debug!("send process mapping place request");
+        self.sender.send(Message::SendConfigureMount(m))?;
+        Ok(())
+    }
+
     pub fn seccomp_notify_request(&mut self, fd: RawFd) -> Result<(), ChannelError> {
         self.sender
             .send_fds(Message::SeccompNotify, &[fd.as_raw_fd()])?;
@@ -186,6 +194,23 @@ impl MainReceiver {
         }
     }
 
+    pub fn wait_process_mount_place(&mut self) -> Result<IdMountParam,ChannelError> {
+        let msg = self
+            .receiver
+            .recv()
+            .map_err(|err| ChannelError::ReceiveError {
+                msg: "waiting for process mount place".to_string(),
+                source: err,
+            })?;
+        match msg {
+            Message::SendConfigureMount(m) => Ok(m),
+            msg => Err(ChannelError::UnexpectedMessage {
+                expected: Message::SendConfigureMount(Default::default()),
+                received: msg,
+            })
+        }
+    }
+
     pub fn close(&self) -> Result<(), ChannelError> {
         self.receiver.close()?;
 
@@ -270,6 +295,12 @@ impl InitSender {
         Ok(())
     }
 
+    pub fn send_mount_source(&mut self, ms: &IdMountSource) -> Result<(),ChannelError> {
+        tracing::debug!("send mount source request");
+        self.sender.send(Message::ReceiveMountFd(ms.clone()))?;
+        Ok(())
+    }
+
     pub fn close(&self) -> Result<(), ChannelError> {
         self.sender.close()?;
 
@@ -301,6 +332,22 @@ impl InitReceiver {
         }
     }
 
+    pub fn wait_for_mount_source(&mut self) -> Result<IdMountSource,ChannelError> {
+        let msg = self.receiver.recv().map_err(|err| ChannelError::ReceiveError {
+            msg: "waiting for get mount source".to_string(),
+            source: err,
+        })?;
+        match msg {
+            Message::ReceiveMountFd(ms) => Ok(ms),
+            msg => Err(
+                ChannelError::UnexpectedMessage {
+                    expected: Message::ReceiveMountFd(IdMountSource { file: 0 }),
+                    received: msg,
+                }
+            )
+        }
+    }
+
     pub fn close(&self) -> Result<(), ChannelError> {
         self.receiver.close()?;
 

crates/libcontainer/src/process/container_init_process.rs

@@ -323,13 +323,20 @@ pub fn container_init_process(
 
         let in_user_ns = utils::is_in_new_userns().map_err(InitProcessError::Io)?;
         let bind_service = namespaces.get(LinuxNamespaceType::User)?.is_some() || in_user_ns;
+        let mut ns_path: Option<PathBuf> = None;
+        if let Some(user_namespace) = namespaces.get(LinuxNamespaceType::User)? {
+            ns_path = Some(user_namespace.path().clone().unwrap());
+        }
         let rootfs = RootFS::new();
         rootfs
             .prepare_rootfs(
                 spec,
                 rootfs_path,
                 bind_service,
                 namespaces.get(LinuxNamespaceType::Cgroup)?.is_some(),
+                ns_path,
+                main_sender,
+                init_receiver,
             )
             .map_err(|err| {
                 tracing::error!(?err, "failed to prepare rootfs");

crates/libcontainer/src/process/container_main_process.rs

@@ -1,11 +1,14 @@
+use std::mem;
+use std::fs::File;
+use std::os::fd::{AsRawFd, RawFd};
 use nix::sys::wait::{waitpid, WaitStatus};
 use nix::unistd::Pid;
 
 use crate::process::args::ContainerArgs;
 use crate::process::fork::{self, CloneCb};
 use crate::process::intel_rdt::setup_intel_rdt;
 use crate::process::{channel, container_intermediate_process};
-use crate::syscall::SyscallError;
+use crate::syscall::{linux, SyscallError};
 use crate::user_ns::UserNamespaceConfig;
 
 #[derive(Debug, thiserror::Error)]
@@ -115,6 +118,44 @@ pub fn container_main_process(container_args: &ContainerArgs) -> Result<(Pid, bo
     // The intermediate process will send the init pid once it forks the init
     // process.  The intermediate process should exit after this point.
     let init_pid = main_receiver.wait_for_intermediate_ready()?;
+
+    loop {
+        let id_map = main_receiver.wait_process_mount_place().unwrap();
+        if id_map.end {
+            break;
+        }
+        let mut flags = libc::OPEN_TREE_CLONE | libc::OPEN_TREE_CLOEXEC;
+        let id_map_flags = id_map.flags;
+        if id_map_flags&libc::MS_REC == libc::MS_REC {
+            flags |= libc::AT_RECURSIVE as std::os::raw::c_uint;
+        }
+        let mount_file = unsafe {
+            libc::syscall(libc::SYS_open_tree,libc::AT_FDCWD,id_map.source,flags)
+        };
+        let mut userns_file: RawFd = -1;
+        if id_map.user_ns_path != "" {
+            let file = File::open(id_map.user_ns_path).unwrap();
+            userns_file = file.as_raw_fd();
+        } else {
+            // TODO we need use uid and git go get fd
+        }
+
+        let mat = &linux::MountAttr{
+            attr_set: linux::MOUNT_ATTR_IDMAP,
+            attr_clr: 0,
+            propagation: 0,
+            userns_fd: userns_file as u64,
+        };
+        let mut set_attr_flags = libc::AT_EMPTY_PATH;
+        if id_map.recursive {
+            set_attr_flags |= libc::AT_RECURSIVE;
+        }
+        syscall.mount_setattr(mount_file as i32, "".as_ref(), set_attr_flags as u32, mat,mem::size_of::<linux::MountAttr>(),).map_err(|err| {
+            tracing::error!(?err, "failed to mount_setattr");
+            ProcessError::SyscallOther(err)
+        }).unwrap();
+    }
+
     let mut need_to_clean_up_intel_rdt_subdirectory = false;
 
     if let Some(linux) = container_args.spec.linux() {

crates/libcontainer/src/process/message.rs

@@ -1,6 +1,7 @@
 use core::fmt;
 
 use serde::{Deserialize, Serialize};
+use crate::rootfs::mount::{IdMountParam, IdMountSource};
 
 /// Used as a wrapper for messages to be sent between child and parent processes
 #[derive(Debug, Serialize, Deserialize, Clone)]
@@ -12,6 +13,8 @@ pub enum Message {
     SeccompNotify,
     SeccompNotifyDone,
     ExecFailed(String),
+    SendConfigureMount(IdMountParam),
+    ReceiveMountFd(IdMountSource),
 }
 
 impl fmt::Display for Message {
@@ -24,6 +27,8 @@ impl fmt::Display for Message {
             Message::SeccompNotify => write!(f, "SeccompNotify"),
             Message::SeccompNotifyDone => write!(f, "SeccompNotifyDone"),
             Message::ExecFailed(s) => write!(f, "ExecFailed({})", s),
+            Message::SendConfigureMount(_) => write!(f,"SendConfigureMount"),
+            Message::ReceiveMountFd(_) => write!(f,"ReceiveMountFd"),
         }
     }
 }