1 /* Register destructors for C++ TLS variables declared with thread_local.
2 Copyright (C) 2013-2023 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
9
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
14
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <https://www.gnu.org/licenses/>. */
18
19 /* CONCURRENCY NOTES:
20
21 This documents concurrency for the non-POD TLS destructor registration,
22 calling and destruction. The functions __cxa_thread_atexit_impl,
23 _dl_close_worker and __call_tls_dtors are the three main routines that may
24 run concurrently and access shared data. The shared data in all possible
25 combinations of all three functions are the link map list, a link map for a
26 DSO and the link map member l_tls_dtor_count.
27
28 __cxa_thread_atexit_impl acquires the dl_load_lock before accessing any
29 shared state and hence multiple of its instances can safely execute
30 concurrently.
31
32 _dl_close_worker acquires the dl_load_lock before accessing any shared state
33 as well and hence can concurrently execute multiple of its own instances as
34 well as those of __cxa_thread_atexit_impl safely. Not all accesses to
35 l_tls_dtor_count are protected by the dl_load_lock, so we need to
36 synchronize using atomics.
37
38 __call_tls_dtors accesses the l_tls_dtor_count without taking the lock; it
39 decrements the value by one. It does not need the big lock because it does
40 not access any other shared state except for the current DSO link map and
41 its member l_tls_dtor_count.
42
43 Correspondingly, _dl_close_worker loads l_tls_dtor_count and if it is zero,
44 unloads the DSO, thus deallocating the current link map. This is the goal
45 of maintaining l_tls_dtor_count - to unload the DSO and free resources if
46 there are no pending destructors to be called.
47
48 We want to eliminate the inconsistent state where the DSO is unloaded in
49 _dl_close_worker before it is used in __call_tls_dtors. This could happen
50 if __call_tls_dtors uses the link map after it sets l_tls_dtor_count to 0,
51 since _dl_close_worker will conclude from the 0 l_tls_dtor_count value that
52 it is safe to unload the DSO. Hence, to ensure that this does not happen,
53 the following conditions must be met:
54
55 1. In _dl_close_worker, the l_tls_dtor_count load happens before the DSO is
56 unloaded and its link map is freed
57 2. The link map dereference in __call_tls_dtors happens before the
58 l_tls_dtor_count dereference.
59
60 To ensure this, the l_tls_dtor_count decrement in __call_tls_dtors should
61 have release semantics and the load in _dl_close_worker should have acquire
62 semantics.
63
64 Concurrent executions of __call_tls_dtors should only ensure that the value
65 is accessed atomically; no reordering constraints need to be considered.
66 Likewise for the increment of l_tls_dtor_count in __cxa_thread_atexit_impl.
67
68 There is still a possibility on concurrent execution of _dl_close_worker and
69 __call_tls_dtors where _dl_close_worker reads the value of l_tls_dtor_count
70 as 1, __call_tls_dtors decrements the value of l_tls_dtor_count but
71 _dl_close_worker does not unload the DSO, having read the old value. This
72 is not very different from a case where __call_tls_dtors is called after
73 _dl_close_worker on the DSO and hence is an accepted execution. */
74
75 #include <stdio.h>
76 #include <stdlib.h>
77 #include <ldsodefs.h>
78 #include <pointer_guard.h>
79
80 typedef void (*dtor_func) (void *);
81
82 struct dtor_list
83 {
84 dtor_func func;
85 void *obj;
86 struct link_map *map;
87 struct dtor_list *next;
88 };
89
90 static __thread struct dtor_list *tls_dtor_list;
91 static __thread void *dso_symbol_cache;
92 static __thread struct link_map *lm_cache;
93
94 /* Register a destructor for TLS variables declared with the 'thread_local'
95 keyword. This function is only called from code generated by the C++
96 compiler. FUNC is the destructor function and OBJ is the object to be
97 passed to the destructor. DSO_SYMBOL is the __dso_handle symbol that each
98 DSO has at a unique address in its map, added from crtbegin.o during the
99 linking phase. */
100 int
101 __cxa_thread_atexit_impl (dtor_func func, void *obj, void *dso_symbol)
102 {
103 PTR_MANGLE (func);
104
105 /* Prepend. */
106 struct dtor_list *new = calloc (1, sizeof (struct dtor_list));
107 if (__glibc_unlikely (new == NULL))
108 __libc_fatal ("Fatal glibc error: failed to register TLS destructor: "
109 "out of memory\n");
110 new->func = func;
111 new->obj = obj;
112 new->next = tls_dtor_list;
113 tls_dtor_list = new;
114
115 /* We have to acquire the big lock to prevent a racing dlclose from pulling
116 our DSO from underneath us while we're setting up our destructor. */
117 __rtld_lock_lock_recursive (GL(dl_load_lock));
118
119 /* See if we already encountered the DSO. */
120 if (__glibc_unlikely (dso_symbol_cache != dso_symbol))
121 {
122 ElfW(Addr) caller = (ElfW(Addr)) dso_symbol;
123
124 struct link_map *l = _dl_find_dso_for_object (caller);
125
126 /* If the address is not recognized the call comes from the main
127 program (we hope). */
128 lm_cache = l ? l : GL(dl_ns)[LM_ID_BASE]._ns_loaded;
129 }
130
131 /* This increment may only be concurrently observed either by the decrement
132 in __call_tls_dtors since the other l_tls_dtor_count access in
133 _dl_close_worker is protected by the dl_load_lock. The execution in
134 __call_tls_dtors does not really depend on this value beyond the fact that
135 it should be atomic, so Relaxed MO should be sufficient. */
136 atomic_fetch_add_relaxed (&lm_cache->l_tls_dtor_count, 1);
137 __rtld_lock_unlock_recursive (GL(dl_load_lock));
138
139 new->map = lm_cache;
140
141 return 0;
142 }
143
144 /* Call the destructors. This is called either when a thread returns from the
145 initial function or when the process exits via the exit function. */
146 void
147 __call_tls_dtors (void)
148 {
149 while (tls_dtor_list)
150 {
151 struct dtor_list *cur = tls_dtor_list;
152 dtor_func func = cur->func;
153 PTR_DEMANGLE (func);
154
155 tls_dtor_list = tls_dtor_list->next;
156 func (cur->obj);
157
158 /* Ensure that the MAP dereference happens before
159 l_tls_dtor_count decrement. That way, we protect this access from a
160 potential DSO unload in _dl_close_worker, which happens when
161 l_tls_dtor_count is 0. See CONCURRENCY NOTES for more detail. */
162 atomic_fetch_add_release (&cur->map->l_tls_dtor_count, -1);
163 free (cur);
164 }
165 }
166 libc_hidden_def (__call_tls_dtors)