From e3caf98a8f5446aeeee033e29e44991bdc881a9a Mon Sep 17 00:00:00 2001
From: eldritch horrors <pennae@lix.systems>
Date: Wed, 25 Jun 2025 15:54:19 +0200
Subject: [PATCH] libstore: wait for cgroups to die on cleanup

killing a cgroup via `cgroup.kill` is not synchronous, we need to give
the processes in the group some time to wake up and exit. due to a few
historical accidents in the codebase we cannot do this asycnhronously,
e.g. with a kj promise without creating yet more problems. we will, at
some point in the future, have to move cgroup management into the main
daemon rather than doing it with RAII wrappers within every subdaemon.

Change-Id: I03bf9060144b5737729f2b05c25771c674fd154c
---
 lix/libutil/cgroup.cc | 52 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)

diff --git a/lix/libutil/cgroup.cc b/lix/libutil/cgroup.cc
index e772da7ab..a1fe6d8bf 100644
--- a/lix/libutil/cgroup.cc
+++ b/lix/libutil/cgroup.cc
@@ -1,4 +1,8 @@
+#include "error.hh"
+#include "file-descriptor.hh"
 #include "logging.hh"
+#include <fcntl.h>
+#include <sys/poll.h>
 #if __linux__
 
 #include "regex.hh"
@@ -116,6 +120,54 @@ destroyCgroup(const std::string & name, const std::filesystem::path & aliveCgrou
 
     killCgroup(name, aliveCgroup);
 
+    // FIXME this should be done asynchronously, but we can't do that without
+    // a proper cgroup management entity at a higher level (eg in Worker). we
+    // can either synchronously kill the cgroup and have simple RAII classes,
+    // or we can have a management entity that can asynchronously destroy our
+    // cgroups since asynchronous destructors unfortunately don't exist here.
+    // FIXME we should also check that no new processes are added, but we can
+    // only do this without *even more* code duplication after checkInterrupt
+    // has been excised. until then a timeout on group death will have to do.
+    {
+        auto eventsFile = aliveCgroup / "cgroup.events";
+        AutoCloseFD events(open(eventsFile.c_str(), O_RDONLY));
+        if (!events) {
+            throw SysError("failed to open %s", eventsFile);
+        }
+
+        constexpr int WAIT_MS = 1000;
+        constexpr std::chrono::seconds TIMEOUT{120};
+
+        const auto started = std::chrono::steady_clock::now();
+
+        bool populated = true;
+        while (populated && (std::chrono::steady_clock::now() - started) < TIMEOUT) {
+            // there's only two keys today, this should be fine for a while
+            std::array<char, 1024> buf = {};
+            const auto got = pread(events.get(), buf.data(), buf.size(), 0);
+            if (got < 0) {
+                throw SysError("reading %s", eventsFile);
+            }
+            auto lines = tokenizeString<Strings>(std::string_view{buf.data(), size_t(got)}, "\n");
+            for (const auto & line : lines) {
+                auto tokens = tokenizeString<Strings>(line);
+                if (tokens.front() == "populated" && tokens.back() == "0") {
+                    populated = false;
+                    break;
+                }
+            }
+
+            debug("cgroup %s isn't empty yet, waiting for a while", aliveCgroup);
+            pollfd pfd{.fd = events.get(), .events = POLLPRI | POLLERR};
+            if (poll(&pfd, 1, WAIT_MS) < 0) {
+                throw SysError("polling %s", eventsFile);
+            }
+        }
+
+        // if the cgroup is still populated after waiting we just continue.
+        // cleanup will fail, but we can't do any better than this for now.
+    }
+
     Result<CgroupStats> stats = readStatistics(aliveCgroup);
 
     if (rmdir(aliveCgroup.c_str()) == -1) {