add schedule entity

Signed-off-by: lengrongfu <[email protected]>
Signed-off-by: lengrongfu <[email protected]>

crates/libcontainer/Cargo.toml

@@ -53,6 +53,7 @@ regex = "1.10.2"
 thiserror = "1.0.50"
 tracing = { version = "0.1.40", features = ["attributes"] }
 safe-path = "0.1.0"
+nc = "0.8.18"
 
 [dev-dependencies]
 oci-spec = { version = "~0.6.4", features = ["proptests", "runtime"] }

crates/libcontainer/src/container/tenant_builder.rs

@@ -4,7 +4,7 @@ use nix::unistd::{self, close, pipe2, read, Pid};
 use oci_spec::runtime::{
     Capabilities as SpecCapabilities, Capability as SpecCapability, LinuxBuilder,
     LinuxCapabilities, LinuxCapabilitiesBuilder, LinuxNamespace, LinuxNamespaceBuilder,
-    LinuxNamespaceType, Process, ProcessBuilder, Spec,
+    LinuxNamespaceType, LinuxSchedulerPolicy, Process, ProcessBuilder, Spec,
 };
 use procfs::process::Namespace;
 
@@ -222,6 +222,59 @@ impl TenantContainerBuilder {
                     }
                 }
             }
+
+            if let Some(sc) = process.scheduler() {
+                let policy = sc.policy();
+                if let Some(nice) = sc.nice() {
+                    // https://man7.org/linux/man-pages/man2/sched_setattr.2.html#top_of_page
+                    if (*policy != LinuxSchedulerPolicy::SchedBatch
+                        || *policy != LinuxSchedulerPolicy::SchedOther)
+                        && *nice < -20
+                        && *nice > 19
+                    {
+                        tracing::error!(?nice, "invalid scheduler.nice: '{}'", nice);
+                        Err(ErrInvalidSpec::Scheduler)?;
+                    }
+                }
+                if let Some(priority) = sc.priority() {
+                    if *priority != 0
+                        && (*policy != LinuxSchedulerPolicy::SchedFifo
+                            && *policy != LinuxSchedulerPolicy::SchedRr)
+                    {
+                        tracing::error!(?policy,"scheduler.priority can only be specified for SchedFIFO or SchedRR policy");
+                        Err(ErrInvalidSpec::Scheduler)?;
+                    }
+                }
+                if *policy != LinuxSchedulerPolicy::SchedDeadline {
+                    if let Some(runtime) = sc.runtime() {
+                        if *runtime != 0 {
+                            tracing::error!(
+                                ?runtime,
+                                "scheduler runtime can only be specified for SchedDeadline policy"
+                            );
+                            Err(ErrInvalidSpec::Scheduler)?;
+                        }
+                    }
+                    if let Some(deadline) = sc.deadline() {
+                        if *deadline != 0 {
+                            tracing::error!(
+                                ?deadline,
+                                "scheduler deadline can only be specified for SchedDeadline policy"
+                            );
+                            Err(ErrInvalidSpec::Scheduler)?;
+                        }
+                    }
+                    if let Some(period) = sc.period() {
+                        if *period != 0 {
+                            tracing::error!(
+                                ?period,
+                                "scheduler period can only be specified for SchedDeadline policy"
+                            );
+                            Err(ErrInvalidSpec::Scheduler)?;
+                        }
+                    }
+                }
+            }
         }
 
         utils::validate_spec_for_new_user_ns(spec)?;

crates/libcontainer/src/error.rs

@@ -92,4 +92,6 @@ pub enum ErrInvalidSpec {
     AppArmorNotEnabled,
     #[error("invalid io priority or class.")]
     IoPriority,
+    #[error("invalid scheduler config for process")]
+    Scheduler,
 }

crates/libcontainer/src/process/container_init_process.rs

@@ -7,13 +7,18 @@ use crate::{
     capabilities, hooks, namespaces::Namespaces, process::channel, rootfs::RootFS, tty,
     user_ns::UserNamespaceConfig, utils,
 };
+use nc;
 use nix::mount::MsFlags;
 use nix::sched::CloneFlags;
 use nix::sys::stat::Mode;
 use nix::unistd::setsid;
 use nix::unistd::{self, Gid, Uid};
-use oci_spec::runtime::{IOPriorityClass, LinuxIOPriority, LinuxNamespaceType, Spec, User};
+use oci_spec::runtime::{
+    IOPriorityClass, LinuxIOPriority, LinuxNamespaceType, LinuxSchedulerFlag, LinuxSchedulerPolicy,
+    Scheduler, Spec, User,
+};
 use std::collections::HashMap;
+use std::mem;
 use std::os::unix::io::AsRawFd;
 use std::{
     env, fs,
@@ -74,6 +79,8 @@ pub enum InitProcessError {
     WorkloadValidation(#[from] workload::ExecutorValidationError),
     #[error("invalid io priority class: {0}")]
     IoPriorityClass(String),
+    #[error("call exec sched_setattr error: {0}")]
+    SchedSetattr(String),
 }
 
 type Result<T> = std::result::Result<T, InitProcessError>;
@@ -288,6 +295,8 @@ pub fn container_init_process(
 
     set_io_priority(syscall.as_ref(), proc.io_priority())?;
 
+    setup_scheduler(proc.scheduler())?;
+
     // set up tty if specified
     if let Some(csocketfd) = args.console_socket {
         tty::setup_console(&csocketfd).map_err(|err| {
@@ -741,6 +750,58 @@ fn set_io_priority(syscall: &dyn Syscall, io_priority_op: &Option<LinuxIOPriorit
     Ok(())
 }
 
+/// Set the RT priority of a thread
+fn setup_scheduler(sc_op: &Option<Scheduler>) -> Result<()> {
+    if let Some(sc) = sc_op {
+        let policy: u32 = match sc.policy() {
+            LinuxSchedulerPolicy::SchedOther => 0,
+            LinuxSchedulerPolicy::SchedFifo => 1,
+            LinuxSchedulerPolicy::SchedRr => 2,
+            LinuxSchedulerPolicy::SchedBatch => 3,
+            LinuxSchedulerPolicy::SchedIso => 4,
+            LinuxSchedulerPolicy::SchedIdle => 5,
+            LinuxSchedulerPolicy::SchedDeadline => 6,
+        };
+        let mut flags_value: u64 = 0;
+        if let Some(flags) = sc.flags() {
+            for flag in flags {
+                match flag {
+                    LinuxSchedulerFlag::SchedResetOnFork => flags_value |= 0x01,
+                    LinuxSchedulerFlag::SchedFlagReclaim => flags_value |= 0x02,
+                    LinuxSchedulerFlag::SchedFlagDLOverrun => flags_value |= 0x04,
+                    LinuxSchedulerFlag::SchedFlagKeepPolicy => flags_value |= 0x08,
+                    LinuxSchedulerFlag::SchedFlagKeepParams => flags_value |= 0x10,
+                    LinuxSchedulerFlag::SchedFlagUtilClampMin => flags_value |= 0x20,
+                    LinuxSchedulerFlag::SchedFlagUtilClampMax => flags_value |= 0x40,
+                }
+            }
+        }
+        let mut a = nc::sched_attr_t {
+            size: mem::size_of::<nc::sched_attr_t>().try_into().unwrap(),
+            sched_policy: policy,
+            sched_flags: flags_value,
+            sched_nice: sc.nice().unwrap_or(0),
+            sched_priority: sc.priority().unwrap_or(0) as u32,
+            sched_runtime: sc.runtime().unwrap_or(0),
+            sched_deadline: sc.deadline().unwrap_or(0),
+            sched_period: sc.period().unwrap_or(0),
+            sched_util_min: 0,
+            sched_util_max: 0,
+        };
+        unsafe {
+            let result = nc::sched_setattr(0, &mut a, 0);
+            match result {
+                Ok(_) => {}
+                Err(err) => {
+                    tracing::error!(?err, "error setting scheduler");
+                    Err(InitProcessError::SchedSetattr(err.to_string()))?;
+                }
+            }
+        };
+    }
+    Ok(())
+}
+
 #[cfg(feature = "libseccomp")]
 fn sync_seccomp(
     fd: Option<i32>,

tests/integration_test/src/main.rs

@@ -9,6 +9,7 @@ use crate::tests::lifecycle::{ContainerCreate, ContainerLifecycle};
 use crate::tests::linux_ns_itype::get_ns_itype_tests;
 use crate::tests::mounts_recursive::get_mounts_recursive_test;
 use crate::tests::pidfile::get_pidfile_test;
+use crate::tests::process::get_scheduler_test;
 use crate::tests::readonly_paths::get_ro_paths_test;
 use crate::tests::seccomp::get_seccomp_test;
 use crate::tests::seccomp_notify::get_seccomp_notify_test;
@@ -103,6 +104,7 @@ fn main() -> Result<()> {
     let mounts_recursive = get_mounts_recursive_test();
     let intel_rdt = get_intel_rdt_test();
     let sysctl = get_sysctl_test();
+    let scheduler = get_scheduler_test();
 
     tm.add_test_group(Box::new(cl));
     tm.add_test_group(Box::new(cc));
@@ -123,6 +125,7 @@ fn main() -> Result<()> {
     tm.add_test_group(Box::new(mounts_recursive));
     tm.add_test_group(Box::new(intel_rdt));
     tm.add_test_group(Box::new(sysctl));
+    tm.add_test_group(Box::new(scheduler));
 
     tm.add_cleanup(Box::new(cgroups::cleanup_v1));
     tm.add_cleanup(Box::new(cgroups::cleanup_v2));

tests/integration_test/src/tests/mod.rs

@@ -7,6 +7,7 @@ pub mod lifecycle;
 pub mod linux_ns_itype;
 pub mod mounts_recursive;
 pub mod pidfile;
+pub mod process;
 pub mod readonly_paths;
 pub mod seccomp;
 pub mod seccomp_notify;

tests/integration_test/src/tests/process/mod.rs

@@ -0,0 +1,3 @@
+mod scheduler;
+
+pub use scheduler::get_scheduler_test;

tests/integration_test/src/tests/process/scheduler.rs

@@ -0,0 +1,51 @@
+use anyhow::{Context, Result};
+use oci_spec::runtime::{
+    LinuxSchedulerPolicy, ProcessBuilder, SchedulerBuilder, Spec, SpecBuilder,
+};
+use test_framework::{test_result, Test, TestGroup, TestResult};
+
+use crate::utils::test_inside_container;
+
+////////// ANCHOR: create_spec policy: u32
+fn create_spec(policy: LinuxSchedulerPolicy) -> Result<Spec> {
+    let sc = SchedulerBuilder::default()
+        .policy(policy)
+        .nice(0i32)
+        .build()
+        .unwrap();
+    SpecBuilder::default()
+        .process(
+            ProcessBuilder::default()
+                .args(
+                    ["runtimetest", "hello_world"]
+                        .iter()
+                        .map(|s| s.to_string())
+                        .collect::<Vec<String>>(),
+                )
+                .scheduler(sc)
+                .build()?,
+        )
+        .build()
+        .context("failed to create spec")
+}
+
+fn scheduler_policy_other_test() -> TestResult {
+    let spec = test_result!(create_spec(LinuxSchedulerPolicy::SchedOther));
+    test_inside_container(spec, &|_| Ok(()))
+}
+
+fn scheduler_policy_batch_test() -> TestResult {
+    let spec = test_result!(create_spec(LinuxSchedulerPolicy::SchedBatch));
+    test_inside_container(spec, &|_| Ok(()))
+}
+
+pub fn get_scheduler_test() -> TestGroup {
+    let mut scheduler_policy_group = TestGroup::new("set_scheduler_policy");
+    let policy_fifo_test = Test::new("policy_fifo", Box::new(scheduler_policy_other_test));
+    let policy_rr_test = Test::new("policy_rr", Box::new(scheduler_policy_batch_test));
+
+    scheduler_policy_group.add(vec![Box::new(policy_fifo_test)]);
+    scheduler_policy_group.add(vec![Box::new(policy_rr_test)]);
+
+    scheduler_policy_group
+}