From: Lennart Poettering <lennart@poettering.net>
Date: Thu, 22 Jun 2023 13:09:50 +0000 (+0200)
Subject: async: add explanatory comment
X-Git-Tag: v254-rc1~133^2~4
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e4687bb8a6a2986642e38b5272694ce21e927679;p=thirdparty%2Fsystemd.git

async: add explanatory comment
---

diff --git a/src/shared/async.h b/src/shared/async.h
index 71d044dd618..96148f9006e 100644
--- a/src/shared/async.h
+++ b/src/shared/async.h
@@ -6,6 +6,19 @@
 #include "macro.h"
 #include "rm-rf.h"
 
+/* These functions implement various potentially slow operations that are executed asynchronously. They are
+ * carefully written to not use pthreads, but use fork() or clone() (without CLONE_VM) so that the child does
+ * not share any memory with the parent process, and thus cannot possibly interfere with the malloc()
+ * synchronization locks.
+ *
+ * Background: glibc only synchronizes malloc() locks when doing fork(), but not when doing clone()
+ * (regardless if through glibc's own wrapper or ours). This means if another thread in the parent has the
+ * malloc() lock taken while a thread is cloning, the mutex will remain locked in the child (but the other
+ * thread won't exist there), with no chance to ever be unlocked again. This will result in deadlocks. Hence
+ * one has to make the choice: either never use threads in the parent, or never do memory allocation in the
+ * child, or never use clone()/clone3() and stick to fork() only. Because we need clone()/clone3() we opted
+ * for avoiding threads. */
+
 int asynchronous_sync(pid_t *ret_pid);
 int asynchronous_close(int fd);
 int asynchronous_rm_rf(const char *p, RemoveFlags flags);