From af756d78ac042a2eed2417c5250d4b675d43bf93 Mon Sep 17 00:00:00 2001
From: Adrian Kummerlaender
Date: Wed, 17 Feb 2016 15:02:48 +0100
Subject: Implement static allocator for initialization

The previous interposition logic based on plain usage of `dlsym` analogously to various online examples led to a deadlock during _neovim_ startup. This deadlock was caused by _neovim_'s custom memory allocation library _jemalloc_ because it calls `mmap` during its initialization phase. The problem with calling `mmap` during initialization is that this already leads to executing `libChangeLog`'s `mmap` version whoes static `actual_mmap` function pointer is not initialized at this point in time. This is detected and leads to a call to `dlsym` to remedy this situation. Sadly `dlsym` in turn requires memory allocation using `calloc` which leads us back to initializing _jemalloc_ and as such to a deadlock.

I first saw this as a bug in _jemalloc_ which seemed to be confirmed by a short search in my search engine of choice. This prompted me to create an appropriate [bug report](https://github.com/jemalloc/jemalloc/issues/329) which was dismissed as a problem in the way `mmap` was interposed and not as a bug in the library. Thus it seems to be accepted practice that it is not the responsibility of a custom memory allocator to cater to the initialization needs of other libraries relying on function interposition. This is of course a valid position as the whole issue is a kind of _chicken and egg_ problem where both sides can be argued.

To cut to the chase I was left with the only option of working around this deadlock by adapting `libChangeLog` to call `dlsym` without relying on the wrapped application's memory allocator of choice. The most straight forward way to do this is to provide another custom memory allocator alongside the _payload_ function interpositions of `mmap` and friends.

`init/alloc.cc` implements such a selectively transparent memory allocator that offers a small static buffer for usage in the context of executing `dlsym`.The choice between forwarding memory allocation requests to the wrapped application's allocator and using the static buffer is governed by `init::dlsymContext`. This tiny helper class maintains an `dlsym_level` counter by posing as a scope guard.

The end result of this extension to `libChangeLog` is that it now also works with applications using _jemalloc_ such as _neovim_ and should overall be much more robust during its initialization phase.
---
 src/change_log.cc | 45 +++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 43 insertions(+), 2 deletions(-)

(limited to 'src/change_log.cc')

diff --git a/src/change_log.cc b/src/change_log.cc
index e0dc179..5f3b4f8 100644
--- a/src/change_log.cc
+++ b/src/change_log.cc
@@ -1,5 +1,6 @@
 #include "actual.h"
 
+#include "init/alloc.h"
 #include "utility/io.h"
 #include "utility/logger.h"
 #include "tracking/path_matcher.h"
@@ -14,8 +15,8 @@ static std::unique_ptr<utility::Logger>              logger;
 static std::unique_ptr<tracking::PathMatcher>        matcher;
 static std::unique_ptr<tracking::ChangeTracker>      tracker;
 
-void init() __attribute__ ((constructor));
-void init() {
+void initialize() __attribute__ ((constructor));
+void initialize() {
 	if ( getenv("CHANGE_LOG_TARGET") != NULL ) {
 		fd_guard = std::make_unique<utility::FileDescriptorGuard>(
 			getenv("CHANGE_LOG_TARGET")
@@ -81,6 +82,46 @@ inline void track_remove(const std::string& path) {
 	}
 }
 
+void free(void* ptr) {
+	static actual::ptr<void, void*> actual_free{};
+
+	if ( !actual_free ) {
+		actual_free = actual::get_ptr<decltype(actual_free)>("free");
+	}
+
+	if ( !init::from_static_buffer(ptr) ) {
+		actual_free(ptr);
+	}
+}
+
+void* malloc(size_t size) {
+	static actual::ptr<void*, size_t> actual_malloc{};
+
+	if ( init::dlsymContext::is_active() ) {
+		return init::static_malloc(size);
+	} else {
+		if ( !actual_malloc ) {
+			actual_malloc = actual::get_ptr<decltype(actual_malloc)>("malloc");
+		}
+
+		return actual_malloc(size);
+	}
+}
+
+void* calloc(size_t block, size_t size) {
+	static actual::ptr<void*, size_t, size_t> actual_calloc{};
+
+	if ( init::dlsymContext::is_active() ) {
+		return init::static_calloc(block, size);
+	} else {
+		if ( !actual_calloc ) {
+			actual_calloc = actual::get_ptr<decltype(actual_calloc)>("calloc");
+		}
+
+		return actual_calloc(block, size);
+	}
+}
+
 ssize_t write(int fd, const void* buffer, size_t count) {
 	static actual::ptr<ssize_t, int, const void*, size_t> actual_write{};
 
-- 
cgit v1.2.3