Linux: LD_PRELOAD + -z,initfirst

Question

我正在编写一个共享对象，它应该被 LD_PRELOAD 编辑到进程中。
在那个共享对象中，我有一些初始化，比如

__attribute__((constructor)) void initFunc();

我希望在流程中的任何其他代码之前被调用。
对于只是可执行文件的进程，这可以正常工作，但如果进程本身依赖于其他一些共享对象，则这些对象会在我的 LD_PRELOAD 共享对象之前初始化。

我尝试为链接器提供选项 -Wl,-z,initfirst 但这似乎根本没有任何效果。当我使用 LD_DEBUG=files 运行进程时，我仍然看到应用程序在我之前如此启动。

我正在运行 CentOS 5.5

Answer 1

问题是加载器只支持一个带-z initfirst 的共享库，而 libpthread.so(几乎所有东西都使用)已经有这个集合。即使您使用 LD_PRELOAD 加载库，也会首先调用 libpthread 的构造函数。

您可以通过修补加载程序以使用 -z initfirst 支持多个共享库来解决这个问题。这是 ld.so 2.21 版的补丁，它保留了二进制 ABI，但从 initfirst 库中创建了一个链接列表，并首先使用 LD_PRELOAD 构造函数调用它们。

diff --git a/elf/dl-load.c b/elf/dl-load.c
index 6dd8550..ac3b079 100644
--- a/elf/dl-load.c
+++ b/elf/dl-load.c
@@ -1387,7 +1387,27 @@ cannot enable executable stack as shared object requires");

   /* Remember whether this object must be initialized first.  */
   if (l->l_flags_1 & DF_1_INITFIRST)
-    GL(dl_initfirst) = l;
+    {
+#if 0
+      struct initfirst_list *first = malloc(sizeof(*first));
+      first->which = l;
+      first->next = GL(dl_initfirst);
+      GL(dl_initfirst) = first;
+#else
+      struct initfirst_list *node = malloc(sizeof(*node));
+      node->which = l;
+      node->next = NULL;
+      struct initfirst_list *it = GL(dl_initfirst);
+      if (!it)
+        GL(dl_initfirst) = node;
+      else
+        {
+          while (it->next)
+            it = it->next;
+          it->next = node;
+        }
+#endif
+    }

   /* Finally the file information.  */
   l->l_dev = st.st_dev;
diff --git a/elf/dl-map-segments.h b/elf/dl-map-segments.h
index baaa813..bca961c 100644
--- a/elf/dl-map-segments.h
+++ b/elf/dl-map-segments.h
@@ -55,7 +55,11 @@ _dl_map_segments (struct link_map *l, int fd,
       /* Remember which part of the address space this object uses.  */
       l->l_map_start = (ElfW(Addr)) __mmap ((void *) mappref, maplength,
                                             c->prot,
+#if 0
                                             MAP_COPY|MAP_FILE,
+#else
+                                            MAP_COPY|MAP_FILE|MAP_32BIT,
+#endif
                                             fd, c->mapoff);
       if (__glibc_unlikely ((void *) l->l_map_start == MAP_FAILED))
         return DL_MAP_SEGMENTS_ERROR_MAP_SEGMENT;
diff --git a/elf/dl-support.c b/elf/dl-support.c
index 835dcb3..9ea0c05 100644
--- a/elf/dl-support.c
+++ b/elf/dl-support.c
@@ -148,7 +148,7 @@ struct r_search_path_elem *_dl_all_dirs;
 struct r_search_path_elem *_dl_init_all_dirs;

 /* The object to be initialized first.  */
-struct link_map *_dl_initfirst;
+struct initfirst_list *_dl_initfirst;

 /* Descriptor to write debug messages to.  */
 int _dl_debug_fd = STDERR_FILENO;
diff --git a/sysdeps/generic/ldsodefs.h b/sysdeps/generic/ldsodefs.h
index b421931..7bb7a69 100644
--- a/sysdeps/generic/ldsodefs.h
+++ b/sysdeps/generic/ldsodefs.h
@@ -318,7 +318,11 @@ struct rtld_global
   EXTERN unsigned long long _dl_load_adds;

   /* The object to be initialized first.  */
-  EXTERN struct link_map *_dl_initfirst;
+  /*EXTERN struct link_map *_dl_initfirst;*/
+  EXTERN struct initfirst_list {
+    struct link_map *which;
+    struct initfirst_list *next;
+  } *_dl_initfirst;

 #if HP_SMALL_TIMING_AVAIL || defined HP_TIMING_PAD
   /* Start time on CPU clock.  */

我想您可以尝试破解 libpthread 以不使用 -z initfirst，但这似乎是最简单的选择。我已经成功地使用它来获得一个先于其他调用的构造函数。您只需确保您的 LD_PRELOADed 库不使用 libc，因为 libc 的构造函数将首先被调用，而在多线程程序中，libc 依赖于 libpthread，因此 pthread 的构造函数将在此之前被调用。

这是一个例子。我用 -pthread 编译了一个 hello, world 程序(否则没有问题)。我写了一个小库，它是 LD_PRELOADed 并且不依赖于 libc。使用默认加载程序，您无法首先调用您的 init 函数:

$ cat hello.c 
#include <stdio.h>

int main() {
    puts("Hello, world!");
    return 0;
}
$ gcc hello.c -o hello -pthread
$ cat superearly.c
#include <unistd.h> 
#include <sys/syscall.h>

long write(int fd, const void *buffer, size_t len) {
    unsigned long ret;
    __asm__ __volatile__("syscall" : "=a"(ret) : "a"(__NR_write), "D"(fd),
        "S"(buffer), "d"(len));
    return ret;
}

void hello(void) __attribute__((constructor));
void hello(void) {
    write(STDOUT_FILENO, "Got in first!\n", 14);
}
$ gcc superearly.c -fPIC -shared -nostdlib -o libsuperearly.so -Wl,-z,initfirst
$ LD_DEBUG=libs LD_PRELOAD=./libsuperearly.so ./hello
    19997:     find library=libpthread.so.0 [0]; searching
    19997:      search cache=/etc/ld.so.cache
    19997:       trying file=/lib/x86_64-linux-gnu/libpthread.so.0
    19997:
    19997:     find library=libc.so.6 [0]; searching
    19997:      search cache=/etc/ld.so.cache
    19997:       trying file=/lib/x86_64-linux-gnu/libc.so.6
    19997:
    19997:
    19997:     calling init: /lib/x86_64-linux-gnu/libpthread.so.0
    19997:
    19997:
    19997:     calling init: /lib/x86_64-linux-gnu/libc.so.6
    19997:
    19997:
    19997:     calling init: ./libsuperearly.so
    19997:
Got in first!
    19997:
    19997:     initialize program: ./hello
    19997:
    19997:
    19997:     transferring control: ./hello
    19997:
Hello, world!
    19997:
    19997:     calling fini: ./hello [0]
    19997:
    19997:
    19997:     calling fini: /lib/x86_64-linux-gnu/libpthread.so.0 [0]
    19997:
$

但是，使用我上面的补丁修补过的 ld.so:

$ LD_DEBUG=libs LD_PRELOAD=./libsuperearly.so ~/libc/lib/ld-2.21.so ./hello
    19986:     find library=libpthread.so.0 [0]; searching
    19986:      search cache=/home/user/libc/etc/ld.so.cache
    19986:       trying file=/home/user/libc/lib/libpthread.so.0
    19986:
    19986:     find library=libc.so.6 [0]; searching
    19986:      search cache=/home/user/libc/etc/ld.so.cache
    19986:       trying file=/home/user/libc/lib/libc.so.6
    19986:
    19986:
    19986:     calling init: ./libsuperearly.so
    19986:
Got in first!
    19986:
    19986:     calling init: /home/user/libc/lib/libpthread.so.0
    19986:
    19986:
    19986:     calling init: /home/user/libc/lib/libc.so.6
    19986:
    19986:
    19986:     initialize program: ./hello
    19986:
Hello, world!
    19986:
    19986:     calling fini: ./hello [0]
    19986:
    19986:
    19986:     calling fini: /home/user/libc/lib/libpthread.so.0 [0]
    19986:
$