From c1618e86feb3c4fbeee13dcaa3995b081bc123fa Mon Sep 17 00:00:00 2001
From: Leif Hedstrom <zwoop@apache.org>
Date: Wed, 18 Mar 2026 14:39:06 -0600
Subject: [PATCH 1/2] Parallelize dir-sync on graceful shutdowns

This will make normal ATS shutdowns come down quicker, specially on
boxes with a lot of drives. This is similar to the new parallel sync
options under normal operations, except this will always Parallelize
as much as needed.
---
 src/iocore/cache/CacheDir.cc | 35 +++++++++++++++++++++++++++++------
 1 file changed, 29 insertions(+), 6 deletions(-)

diff --git a/src/iocore/cache/CacheDir.cc b/src/iocore/cache/CacheDir.cc
index 413d11c6219..fec7b6d48cf 100644
--- a/src/iocore/cache/CacheDir.cc
+++ b/src/iocore/cache/CacheDir.cc
@@ -33,6 +33,7 @@
 #include "ts/ats_probe.h"
 #include "iocore/eventsystem/Tasks.h"
 
+#include <thread>
 #include <unordered_map>
 
 #ifdef LOOP_CHECK_MODE
@@ -958,20 +959,42 @@ Directory::entries_used()
 }
 
 /*
- * this function flushes the cache meta data to disk when
+ * This function flushes the cache meta data to disk when
  * the cache is shutdown. Must *NOT* be used during regular
- * operation.
+ * operation. Stripes are synced in parallel, one thread per
+ * physical disk.
  */
 
 void
 sync_cache_dir_on_shutdown()
 {
-  Dbg(dbg_ctl_cache_dir_sync, "sync started");
-  EThread *t = reinterpret_cast<EThread *>(0xdeadbeef);
+  Dbg(dbg_ctl_cache_dir_sync, "shutdown sync started");
+
+  std::unordered_map<CacheDisk *, std::vector<int>> drive_stripe_map;
+
   for (int i = 0; i < gnstripes; i++) {
-    gstripes[i]->shutdown(t);
+    drive_stripe_map[gstripes[i]->disk].push_back(i);
   }
-  Dbg(dbg_ctl_cache_dir_sync, "sync done");
+
+  std::vector<std::thread> threads;
+
+  threads.reserve(drive_stripe_map.size());
+  for (auto &[disk, indices] : drive_stripe_map) {
+    Dbg(dbg_ctl_cache_dir_sync, "Disk %s: syncing %zu stripe(s)", disk->path, indices.size());
+    threads.emplace_back([&indices]() {
+      EThread *t = reinterpret_cast<EThread *>(0xdeadbeef);
+
+      for (int idx : indices) {
+        gstripes[idx]->shutdown(t);
+      }
+    });
+  }
+
+  for (auto &thr : threads) {
+    thr.join();
+  }
+
+  Dbg(dbg_ctl_cache_dir_sync, "shutdown sync done");
 }
 
 int

From 3cfb438b971bdd27630761faa10d7a226352f78e Mon Sep 17 00:00:00 2001
From: Leif Hedstrom <zwoop@apache.org>
Date: Wed, 18 Mar 2026 15:32:26 -0600
Subject: [PATCH 2/2] Address Copilot's review comments

Fix lambda reference capture bug where all threads processed the last disk's
stripes instead of their own; fix shared 0xdeadbeef EThread* sentinel by using
a thread_local variable to give each OS thread a unique identity for
MUTEX_TAKE_LOCK.
---
 src/iocore/cache/CacheDir.cc | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/src/iocore/cache/CacheDir.cc b/src/iocore/cache/CacheDir.cc
index fec7b6d48cf..efdd155cfd6 100644
--- a/src/iocore/cache/CacheDir.cc
+++ b/src/iocore/cache/CacheDir.cc
@@ -981,10 +981,13 @@ sync_cache_dir_on_shutdown()
   threads.reserve(drive_stripe_map.size());
   for (auto &[disk, indices] : drive_stripe_map) {
     Dbg(dbg_ctl_cache_dir_sync, "Disk %s: syncing %zu stripe(s)", disk->path, indices.size());
-    threads.emplace_back([&indices]() {
-      EThread *t = reinterpret_cast<EThread *>(0xdeadbeef);
+    auto stripe_indices = indices;
+    threads.emplace_back([stripe_indices]() {
+      // Use a thread_local variable to give each OS thread a unique EThread* sentinel instead of 0xdeadbeef.
+      thread_local char thread_sentinel;
+      EThread          *t = reinterpret_cast<EThread *>(&thread_sentinel);
 
-      for (int idx : indices) {
+      for (int idx : stripe_indices) {
         gstripes[idx]->shutdown(t);
       }
     });