feat: Support reclaiming page cache for agent

agent/src/config/config.rs

@@ -1987,6 +1987,7 @@ pub struct Tunning {
     pub cpu_affinity: Vec<usize>,
     pub process_scheduling_priority: usize,
     pub idle_memory_trimming: bool,
+    pub page_cache_reclaim_percentage: u8,
     #[serde(with = "humantime_serde")]
     pub resource_monitoring_interval: Duration,
 }
@@ -1997,6 +1998,7 @@ impl Default for Tunning {
             cpu_affinity: vec![],
             process_scheduling_priority: 0,
             idle_memory_trimming: false,
+            page_cache_reclaim_percentage: 100,
             resource_monitoring_interval: Duration::from_secs(10),
         }
     }

agent/src/config/handler.rs

@@ -207,6 +207,7 @@ pub struct EnvironmentConfig {
     pub system_load_circuit_breaker_threshold: f32,
     pub system_load_circuit_breaker_recover: f32,
     pub system_load_circuit_breaker_metric: agent::SystemLoadMetric,
+    pub page_cache_reclaim_percentage: u8,
 }
 
 #[derive(Clone, PartialEq, Eq, Debug)]
@@ -1602,6 +1603,7 @@ impl TryFrom<(Config, UserConfig)> for ModuleConfig {
                     .circuit_breakers
                     .relative_sys_load
                     .metric,
+                page_cache_reclaim_percentage: conf.global.tunning.page_cache_reclaim_percentage,
             },
             synchronizer: SynchronizerConfig {
                 sync_interval: conf.global.communication.proactive_request_interval,
@@ -3941,6 +3943,13 @@ impl ConfigHandler {
             tunning.resource_monitoring_interval = new_tunning.resource_monitoring_interval;
             restart_agent = !first_run;
         }
+        if tunning.page_cache_reclaim_percentage != new_tunning.page_cache_reclaim_percentage {
+            info!(
+                "Update global.tunning.page_cache_reclaim_percentage from {:?} to {:?}.",
+                tunning.page_cache_reclaim_percentage, new_tunning.page_cache_reclaim_percentage
+            );
+            tunning.page_cache_reclaim_percentage = new_tunning.page_cache_reclaim_percentage;
+        }
 
         // dev
         let dev = &mut config.dev;

agent/src/monitor.rs

@@ -35,6 +35,7 @@ use crate::{
     config::handler::EnvironmentAccess,
     error::{Error, Result},
     utils::{
+        cgroups,
         process::{get_current_sys_memory_percentage, get_file_and_size_sum},
         stats::{
             self, Collector, Countable, Counter, CounterType, CounterValue, RefCountable,
@@ -318,6 +319,16 @@ impl RefCountable for SysStatusBroker {
                 Err(_) => CounterValue::Unsigned(0),
             },
         ));
+        #[cfg(target_os = "linux")]
+        metrics.push((
+            "page_cache",
+            CounterType::Gauged,
+            if let Some(m_stat) = cgroups::memory_info() {
+                CounterValue::Unsigned(m_stat.stat.cache)
+            } else {
+                CounterValue::Unsigned(0)
+            },
+        ));
         metrics
     }
 }

agent/src/utils/cgroups/linux.rs

@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 
+use std::path::{Path, PathBuf};
 use std::sync::{Arc, Condvar, Mutex};
 use std::thread::JoinHandle;
 use std::time::Duration;
@@ -24,9 +25,14 @@ use crate::config::handler::EnvironmentAccess;
 use crate::utils::environment::is_kernel_available;
 
 use arc_swap::access::Access;
-use cgroups_rs::cgroup_builder::*;
-use cgroups_rs::*;
-use log::{info, warn};
+use cgroups_rs::{
+    cgroup_builder::CgroupBuilder,
+    cpu::CpuController,
+    hierarchies,
+    memory::{MemController, Memory},
+    Cgroup, CgroupPid, Controller, CpuResources, MemoryResources, Resources,
+};
+use log::{debug, info, trace, warn};
 use public::consts::{DEFAULT_CPU_CFS_PERIOD_US, PROCESS_NAME};
 
 pub struct Cgroups {
@@ -40,40 +46,34 @@ pub struct Cgroups {
 
 const CHECK_INTERVAL: Duration = Duration::from_secs(1);
 
+fn cgroups_supported() -> bool {
+    let Ok(fs) = fs::read_to_string("/proc/filesystems") else {
+        return false;
+    };
+    fs.lines()
+        .any(|line| line.to_lowercase().contains("cgroup"))
+}
+
 impl Cgroups {
     /// 创建cgroup hierarchy
     pub fn new(pid: u64, config: EnvironmentAccess) -> Result<Self, Error> {
-        let contents = match fs::read_to_string("/proc/filesystems") {
-            Ok(file_contents) => file_contents,
-            Err(e) => {
-                return Err(Error::CgroupsNotSupported(e.to_string()));
-            }
-        };
-        let mut cgroup_supported = false;
-        for line in contents.lines() {
-            // 检查系统是否支持cgroup
-            if line.to_lowercase().contains("cgroup") {
-                cgroup_supported = true;
-                break;
-            }
-        }
-        if !cgroup_supported {
-            return Err(Error::CgroupsNotSupported(format!(
-                "cgroups v1 or v2 is not found."
-            )));
+        if !cgroups_supported() {
+            return Err(Error::CgroupsNotSupported(
+                "read /proc/filesystems failed or cgroups/cgroups2 not found.".to_string(),
+            ));
         }
         let hier = hierarchies::auto();
         let is_v2 = hier.v2();
         let cg: Cgroup = CgroupBuilder::new(PROCESS_NAME).build(hier);
-        let cpus: &cpu::CpuController = match cg.controller_of() {
+        let cpus: &CpuController = match cg.controller_of() {
             Some(controller) => controller,
             None => {
                 return Err(Error::CpuControllerSetFailed(format!(
                     "maybe cgroups is not installed"
                 )));
             }
         };
-        let mem: &memory::MemController = match cg.controller_of() {
+        let mem: &MemController = match cg.controller_of() {
             Some(controller) => controller,
             None => {
                 return Err(Error::MemControllerSetFailed(format!(
@@ -229,3 +229,135 @@ pub fn is_cgroup_procs_writable() -> bool {
     const MIN_KERNEL_VERSION_CGROUP_PROCS: &str = "3";
     is_kernel_available(MIN_KERNEL_VERSION_CGROUP_PROCS)
 }
+
+const PID1_ROOT: &str = "/proc/1/root";
+
+/*
+ * The path of container memory cgroup from its own namespace is `/sys/fs/cgroup/memory`.
+ * However, the path is mounted read-only, making it impossible to reclaim memory cache with:
+ *
+ *     `echo 0 > /sys/fs/cgroup/memory/memory.force_empty`
+ *
+ * The approach here is to take a little detour, visiting the actual memory cgroup path from global cgroup namespace.
+ * The path would be:
+ *
+ *     /proc/1/root/{memory_cgroup_mount_point}/{cgroup_path}/memory.force_empty
+ */
+fn get_memory_cgroup_path() -> procfs::ProcResult<Option<PathBuf>> {
+    let mut path = PathBuf::from(PID1_ROOT);
+    let proc = procfs::process::Process::myself()?;
+
+    let Some(mount_info) = proc
+        .mountinfo()?
+        .into_iter()
+        .find(|m| m.fs_type == "cgroup" && m.super_options.contains_key("memory"))
+    else {
+        debug!("memory cgroup not found");
+        return Ok(None);
+    };
+    trace!(
+        "memory cgroup mount point: {}",
+        mount_info.mount_point.display()
+    );
+    let mut mount_point = mount_info.mount_point.components();
+    mount_point.next(); // skip "/"
+    path.extend(mount_point);
+
+    let Some(cg_info) = proc
+        .cgroups()?
+        .into_iter()
+        .find(|cg| cg.controllers.iter().any(|c| c == "memory"))
+    else {
+        return Ok(None);
+    };
+    trace!("memory cgroup path: {}", cg_info.pathname);
+    if cg_info.pathname == "/" {
+        debug!("memory cgroup is mounted on root");
+        return Ok(None);
+    }
+    let mut cg_path = Path::new(&cg_info.pathname).components();
+    cg_path.next(); // skip "/"
+    path.extend(cg_path);
+
+    trace!("memory cgroup path: {}", path.display());
+    Ok(Some(path))
+}
+
+fn cgroups_v1_check() -> bool {
+    if !cgroups_supported() {
+        debug!("cgroups not supported for this system");
+        return false;
+    }
+    if hierarchies::is_cgroup2_unified_mode() {
+        debug!("cgroups v2 is not supported");
+        return false;
+    }
+
+    true
+}
+
+pub(crate) fn memory_info() -> Option<Memory> {
+    if !cgroups_v1_check() {
+        return None;
+    }
+
+    let mem_mount = match get_memory_cgroup_path() {
+        Ok(Some(path)) => path,
+        Ok(None) => {
+            debug!("cgroups memory mount point not found or is invalid");
+            return None;
+        }
+        Err(e) => {
+            warn!("get memory path failed: {e}");
+            return None;
+        }
+    };
+
+    Some(MemController::new(mem_mount.clone(), false).memory_stat())
+}
+
+pub(crate) fn page_cache_reclaim_check(threshold: u8) -> bool {
+    if threshold >= 100 {
+        return false;
+    }
+    if !cgroups_v1_check() {
+        return false;
+    }
+
+    let mem_mount = match get_memory_cgroup_path() {
+        Ok(Some(path)) => path,
+        Ok(None) => {
+            debug!("cgroups memory mount point not found or is invalid");
+            return false;
+        }
+        Err(e) => {
+            warn!("get memory path failed: {e}");
+            return false;
+        }
+    };
+    let mc = MemController::new(mem_mount.clone(), false);
+    let mut reclaim_path = mem_mount;
+    reclaim_path.set_file_name("memory.force_empty");
+
+    let m_stat = mc.memory_stat();
+    let percentage = m_stat.stat.cache * 100 / m_stat.limit_in_bytes as u64;
+    if percentage < threshold as u64 {
+        debug!("cache / limit = {percentage}% < {threshold}%");
+        return false;
+    }
+
+    debug!("cache before reclaim: {}", m_stat.stat.cache);
+    if let Err(e) = fs::write(&reclaim_path, b"0") {
+        warn!(
+            "reclaim memory cache write to {} failed: {e}",
+            reclaim_path.display()
+        );
+        return false;
+    }
+    info!(
+        "page cache reclaimed {} -> {}",
+        m_stat.stat.cache,
+        mc.memory_stat().stat.cache
+    );
+    true
+}

agent/src/utils/guard.rs

@@ -355,6 +355,8 @@ impl Guard {
         let in_container = running_in_container();
         let cgroups_disabled = self.cgroups_disabled;
         let mut last_exceeded = get_timestamp(0);
+        #[cfg(target_os = "linux")]
+        let mut last_page_reclaim = Instant::now();
 
         let thread = thread::Builder::new().name("guard".to_owned()).spawn(move || {
             let mut system_load = SystemLoadGuard::new(system.clone(), exception_handler.clone());
@@ -436,6 +438,12 @@ impl Guard {
                     unsafe { let _ = malloc_trim(0); }
                 }
 
+                #[cfg(target_os = "linux")]
+                if last_page_reclaim.elapsed() >= Duration::from_secs(60) {
+                    last_page_reclaim = Instant::now();
+                    let _ = crate::utils::cgroups::page_cache_reclaim_check(config.page_cache_reclaim_percentage);
+                }
+
                 // Periodic memory checks are necessary:
                 // Cgroups does not count the memory of RssFile, and AF_PACKET Block occupies RssFile.
                 // Therefore, using Cgroups to limit the memory usage may not be accurate in some scenarios.

server/agent_config/README-CH.md

@@ -652,6 +652,42 @@ global:
 
 开启闲置内存修剪特性，将降低 agent 内存使用量，但可能会损失 agent 处理性能。
 
+### Page Cache 回收间隔 {#global.tunning.page_cache_reclaim_interval}
+
+**标签**:
+
+`hot_update`
+
+**FQCN**:
+
+`global.tunning.page_cache_reclaim_interval`
+
+Upgrade from old version: `static_config.page-cache-reclaim-interval`
+
+**默认值**:
+```yaml
+global:
+  tunning:
+    page_cache_reclaim_interval: 0
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | duration |
+| Range | [0, '1d'] |
+
+**详细描述**:
+
+Cgroup 的内存使用量包括匿名内存和文件页缓存。在某些情况下，仅仅是文件页缓存就可能导致
+cgroup 因为内存不足杀死 agent 进程。为了避免这种情况，agent 将定期强制清空文件页缓存，
+且由于 agent 的文件 I/O 量不大，这不太可能对 agent 的性能造成影响，但同一 cgroup 下的其他
+进程可能会受到影响。不建议设置很小的值。
+设置该值为 0 时，该特性不生效。
+注意：
+- 该特性仅支持 cgroups v1。
+- 如果 agent 的 memory cgroup 路径是 “/”，该特性不生效。
+
 ### 资源监控间隔 {#global.tunning.resource_monitoring_interval}
 
 **标签**:
@@ -7048,7 +7084,7 @@ processors:
 | Key  | Value                        |
 | ---- | ---------------------------- |
 | Type | duration |
-| Range | ['1s', '10s'] |
+| Range | ['1s', '20s'] |
 
 **详细描述**:
 
@@ -7079,7 +7115,7 @@ processors:
 | Key  | Value                        |
 | ---- | ---------------------------- |
 | Type | duration |
-| Range | ['0s', '10s'] |
+| Range | ['0s', '20s'] |
 
 **详细描述**: