[mesos] 04/06: Cgroups isolator: added support for nested cgroups during recovery.

[abudnik]

This is an automated email from the ASF dual-hosted git repository.

abudnik pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/mesos.git

commit 969836bd653437fd131530520e8875ac55125bc2
Author: Andrei Budnik <abud...@mesosphere.com>
AuthorDate: Tue Mar 3 14:57:34 2020 +0100

    Cgroups isolator: added support for nested cgroups during recovery.
    
    This patch enables recovery for nested cgroups and implements
    the detection of orphaned nested cgroups.
    
    Review: https://reviews.apache.org/r/71966/
---
 .../mesos/isolators/cgroups/cgroups.cpp            | 41 +++++++++++++++-------
 1 file changed, 29 insertions(+), 12 deletions(-)

diff --git a/src/slave/containerizer/mesos/isolators/cgroups/cgroups.cpp 
b/src/slave/containerizer/mesos/isolators/cgroups/cgroups.cpp
index 09feaf3..bf2a4d8 100644
--- a/src/slave/containerizer/mesos/isolators/cgroups/cgroups.cpp
+++ b/src/slave/containerizer/mesos/isolators/cgroups/cgroups.cpp
@@ -206,10 +206,17 @@ Future<Nothing> CgroupsIsolatorProcess::recover(
   // Recover active containers first.
   vector<Future<Nothing>> recovers;
   foreach (const ContainerState& state, states) {
-    // If we are a nested container, we do not need to recover
-    // anything since only top-level containers will have cgroups
+    // If we are a nested container with shared cgroups, we do not
+    // need to recover anything since its ancestor will have cgroups
     // created for them.
-    if (state.container_id().has_parent()) {
+    const bool shareCgroups =
+      (state.has_container_info() &&
+       state.container_info().has_linux_info() &&
+       state.container_info().linux_info().has_share_cgroups())
+        ? state.container_info().linux_info().share_cgroups()
+        : true;
+
+    if (state.container_id().has_parent() && shareCgroups) {
       continue;
     }
 
@@ -248,7 +255,6 @@ Future<Nothing> CgroupsIsolatorProcess::_recover(
   hashset<ContainerID> unknownOrphans;
 
   foreach (const string& hierarchy, subsystems.keys()) {
-    // TODO(jieyu): Use non-recursive version of `cgroups::get`.
     Try<vector<string>> cgroups = cgroups::get(
         hierarchy,
         flags.cgroups_root);
@@ -267,18 +273,28 @@ Future<Nothing> CgroupsIsolatorProcess::_recover(
         continue;
       }
 
-      ContainerID containerId;
-      containerId.set_value(Path(cgroup).basename());
+      // Need to parse the cgroup to see if it's one we created (i.e.,
+      // matches our separator structure) or one that someone else
+      // created (e.g., in the future we might have nested containers
+      // that are managed by something else rooted within the cgroup
+      // hierarchy).
+      Option<ContainerID> containerId =
+        containerizer::paths::parseCgroupPath(flags.cgroups_root, cgroup);
+
+      if (containerId.isNone()) {
+        LOG(INFO) << "Not recovering cgroup " << cgroup;
+        continue;
+      }
 
       // Skip containerId which already have been recovered.
-      if (infos.contains(containerId)) {
+      if (infos.contains(containerId.get())) {
         continue;
       }
 
-      if (orphans.contains(containerId)) {
-        knownOrphans.insert(containerId);
+      if (orphans.contains(containerId.get())) {
+        knownOrphans.insert(containerId.get());
       } else {
-        unknownOrphans.insert(containerId);
+        unknownOrphans.insert(containerId.get());
       }
     }
   }
@@ -335,7 +351,8 @@ Future<Nothing> CgroupsIsolatorProcess::__recover(
 Future<Nothing> CgroupsIsolatorProcess::___recover(
     const ContainerID& containerId)
 {
-  const string cgroup = path::join(flags.cgroups_root, containerId.value());
+  const string cgroup =
+    containerizer::paths::getCgroupPath(flags.cgroups_root, containerId);
 
   vector<Future<Nothing>> recovers;
   hashset<string> recoveredSubsystems;
@@ -397,7 +414,7 @@ Future<Nothing> CgroupsIsolatorProcess::____recover(
 
   infos[containerId] = Owned<Info>(new Info(
       containerId,
-      path::join(flags.cgroups_root, containerId.value())));
+      containerizer::paths::getCgroupPath(flags.cgroups_root, containerId)));
 
   infos[containerId]->subsystems = recoveredSubsystems;