1 | //===-- safestack.cpp -----------------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the runtime support for the safe stack protection |
10 | // mechanism. The runtime manages allocation/deallocation of the unsafe stack |
11 | // for the main thread, as well as all pthreads that are created/destroyed |
12 | // during program execution. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #define SANITIZER_COMMON_NO_REDEFINE_BUILTINS |
17 | |
18 | #include "safestack_platform.h" |
19 | #include "safestack_util.h" |
20 | #include "sanitizer_common/sanitizer_internal_defs.h" |
21 | |
22 | #include <errno.h> |
23 | #include <string.h> |
24 | #include <sys/resource.h> |
25 | |
26 | #include "interception/interception.h" |
27 | |
28 | // interception.h drags in sanitizer_redefine_builtins.h, which in turn |
29 | // creates references to __sanitizer_internal_memcpy etc. The interceptors |
30 | // aren't needed here, so just forward to libc. |
31 | extern "C" { |
32 | SANITIZER_INTERFACE_ATTRIBUTE void *__sanitizer_internal_memcpy(void *dest, |
33 | const void *src, |
34 | size_t n) { |
35 | return memcpy(dest: dest, src: src, n: n); |
36 | } |
37 | |
38 | SANITIZER_INTERFACE_ATTRIBUTE void *__sanitizer_internal_memmove( |
39 | void *dest, const void *src, size_t n) { |
40 | return memmove(dest: dest, src: src, n: n); |
41 | } |
42 | |
43 | SANITIZER_INTERFACE_ATTRIBUTE void *__sanitizer_internal_memset(void *s, int c, |
44 | size_t n) { |
45 | return memset(s: s, c: c, n: n); |
46 | } |
47 | } // extern "C" |
48 | |
49 | using namespace safestack; |
50 | |
51 | // TODO: To make accessing the unsafe stack pointer faster, we plan to |
52 | // eventually store it directly in the thread control block data structure on |
53 | // platforms where this structure is pointed to by %fs or %gs. This is exactly |
54 | // the same mechanism as currently being used by the traditional stack |
55 | // protector pass to store the stack guard (see getStackCookieLocation() |
56 | // function above). Doing so requires changing the tcbhead_t struct in glibc |
57 | // on Linux and tcb struct in libc on FreeBSD. |
58 | // |
59 | // For now, store it in a thread-local variable. |
60 | extern "C" { |
61 | __attribute__((visibility( |
62 | "default" ))) __thread void *__safestack_unsafe_stack_ptr = nullptr; |
63 | } |
64 | |
65 | namespace { |
66 | |
67 | // TODO: The runtime library does not currently protect the safe stack beyond |
68 | // relying on the system-enforced ASLR. The protection of the (safe) stack can |
69 | // be provided by three alternative features: |
70 | // |
71 | // 1) Protection via hardware segmentation on x86-32 and some x86-64 |
72 | // architectures: the (safe) stack segment (implicitly accessed via the %ss |
73 | // segment register) can be separated from the data segment (implicitly |
74 | // accessed via the %ds segment register). Dereferencing a pointer to the safe |
75 | // segment would result in a segmentation fault. |
76 | // |
77 | // 2) Protection via software fault isolation: memory writes that are not meant |
78 | // to access the safe stack can be prevented from doing so through runtime |
79 | // instrumentation. One way to do it is to allocate the safe stack(s) in the |
80 | // upper half of the userspace and bitmask the corresponding upper bit of the |
81 | // memory addresses of memory writes that are not meant to access the safe |
82 | // stack. |
83 | // |
84 | // 3) Protection via information hiding on 64 bit architectures: the location |
85 | // of the safe stack(s) can be randomized through secure mechanisms, and the |
86 | // leakage of the stack pointer can be prevented. Currently, libc can leak the |
87 | // stack pointer in several ways (e.g. in longjmp, signal handling, user-level |
88 | // context switching related functions, etc.). These can be fixed in libc and |
89 | // in other low-level libraries, by either eliminating the escaping/dumping of |
90 | // the stack pointer (i.e., %rsp) when that's possible, or by using |
91 | // encryption/PTR_MANGLE (XOR-ing the dumped stack pointer with another secret |
92 | // we control and protect better, as is already done for setjmp in glibc.) |
93 | // Furthermore, a static machine code level verifier can be ran after code |
94 | // generation to make sure that the stack pointer is never written to memory, |
95 | // or if it is, its written on the safe stack. |
96 | // |
97 | // Finally, while the Unsafe Stack pointer is currently stored in a thread |
98 | // local variable, with libc support it could be stored in the TCB (thread |
99 | // control block) as well, eliminating another level of indirection and making |
100 | // such accesses faster. Alternatively, dedicating a separate register for |
101 | // storing it would also be possible. |
102 | |
103 | /// Minimum stack alignment for the unsafe stack. |
104 | const unsigned kStackAlign = 16; |
105 | |
106 | /// Default size of the unsafe stack. This value is only used if the stack |
107 | /// size rlimit is set to infinity. |
108 | const unsigned kDefaultUnsafeStackSize = 0x2800000; |
109 | |
110 | // Per-thread unsafe stack information. It's not frequently accessed, so there |
111 | // it can be kept out of the tcb in normal thread-local variables. |
112 | __thread void *unsafe_stack_start = nullptr; |
113 | __thread size_t unsafe_stack_size = 0; |
114 | __thread size_t unsafe_stack_guard = 0; |
115 | |
116 | inline void *unsafe_stack_alloc(size_t size, size_t guard) { |
117 | SFS_CHECK(size + guard >= size); |
118 | void *addr = Mmap(addr: nullptr, length: size + guard, PROT_READ | PROT_WRITE, |
119 | MAP_PRIVATE | MAP_ANON, fd: -1, offset: 0); |
120 | SFS_CHECK(MAP_FAILED != addr); |
121 | Mprotect(addr, length: guard, PROT_NONE); |
122 | return (char *)addr + guard; |
123 | } |
124 | |
125 | inline void unsafe_stack_setup(void *start, size_t size, size_t guard) { |
126 | SFS_CHECK((char *)start + size >= (char *)start); |
127 | SFS_CHECK((char *)start + guard >= (char *)start); |
128 | void *stack_ptr = (char *)start + size; |
129 | SFS_CHECK((((size_t)stack_ptr) & (kStackAlign - 1)) == 0); |
130 | |
131 | __safestack_unsafe_stack_ptr = stack_ptr; |
132 | unsafe_stack_start = start; |
133 | unsafe_stack_size = size; |
134 | unsafe_stack_guard = guard; |
135 | } |
136 | |
137 | /// Thread data for the cleanup handler |
138 | pthread_key_t thread_cleanup_key; |
139 | |
140 | /// Safe stack per-thread information passed to the thread_start function |
141 | struct tinfo { |
142 | void *(*start_routine)(void *); |
143 | void *start_routine_arg; |
144 | |
145 | void *unsafe_stack_start; |
146 | size_t unsafe_stack_size; |
147 | size_t unsafe_stack_guard; |
148 | }; |
149 | |
150 | /// Wrap the thread function in order to deallocate the unsafe stack when the |
151 | /// thread terminates by returning from its main function. |
152 | void *thread_start(void *arg) { |
153 | struct tinfo *tinfo = (struct tinfo *)arg; |
154 | |
155 | void *(*start_routine)(void *) = tinfo->start_routine; |
156 | void *start_routine_arg = tinfo->start_routine_arg; |
157 | |
158 | // Setup the unsafe stack; this will destroy tinfo content |
159 | unsafe_stack_setup(start: tinfo->unsafe_stack_start, size: tinfo->unsafe_stack_size, |
160 | guard: tinfo->unsafe_stack_guard); |
161 | |
162 | // Make sure out thread-specific destructor will be called |
163 | pthread_setspecific(key: thread_cleanup_key, pointer: (void *)1); |
164 | |
165 | return start_routine(start_routine_arg); |
166 | } |
167 | |
168 | /// Linked list used to store exiting threads stack/thread information. |
169 | struct thread_stack_ll { |
170 | struct thread_stack_ll *next; |
171 | void *stack_base; |
172 | size_t size; |
173 | pid_t pid; |
174 | ThreadId tid; |
175 | }; |
176 | |
177 | /// Linked list of unsafe stacks for threads that are exiting. We delay |
178 | /// unmapping them until the thread exits. |
179 | thread_stack_ll *thread_stacks = nullptr; |
180 | pthread_mutex_t thread_stacks_mutex = PTHREAD_MUTEX_INITIALIZER; |
181 | |
182 | /// Thread-specific data destructor. We want to free the unsafe stack only after |
183 | /// this thread is terminated. libc can call functions in safestack-instrumented |
184 | /// code (like free) after thread-specific data destructors have run. |
185 | void thread_cleanup_handler(void *_iter) { |
186 | SFS_CHECK(unsafe_stack_start != nullptr); |
187 | pthread_setspecific(key: thread_cleanup_key, NULL); |
188 | |
189 | pthread_mutex_lock(mutex: &thread_stacks_mutex); |
190 | // Temporary list to hold the previous threads stacks so we don't hold the |
191 | // thread_stacks_mutex for long. |
192 | thread_stack_ll *temp_stacks = thread_stacks; |
193 | thread_stacks = nullptr; |
194 | pthread_mutex_unlock(mutex: &thread_stacks_mutex); |
195 | |
196 | pid_t pid = getpid(); |
197 | ThreadId tid = GetTid(); |
198 | |
199 | // Free stacks for dead threads |
200 | thread_stack_ll **stackp = &temp_stacks; |
201 | while (*stackp) { |
202 | thread_stack_ll *stack = *stackp; |
203 | if (stack->pid != pid || |
204 | (-1 == TgKill(pid: stack->pid, tid: stack->tid, sig: 0) && errno == ESRCH)) { |
205 | Munmap(addr: stack->stack_base, length: stack->size); |
206 | *stackp = stack->next; |
207 | free(ptr: stack); |
208 | } else |
209 | stackp = &stack->next; |
210 | } |
211 | |
212 | thread_stack_ll *cur_stack = |
213 | (thread_stack_ll *)malloc(size: sizeof(thread_stack_ll)); |
214 | cur_stack->stack_base = (char *)unsafe_stack_start - unsafe_stack_guard; |
215 | cur_stack->size = unsafe_stack_size + unsafe_stack_guard; |
216 | cur_stack->pid = pid; |
217 | cur_stack->tid = tid; |
218 | |
219 | pthread_mutex_lock(mutex: &thread_stacks_mutex); |
220 | // Merge thread_stacks with the current thread's stack and any remaining |
221 | // temp_stacks |
222 | *stackp = thread_stacks; |
223 | cur_stack->next = temp_stacks; |
224 | thread_stacks = cur_stack; |
225 | pthread_mutex_unlock(mutex: &thread_stacks_mutex); |
226 | |
227 | unsafe_stack_start = nullptr; |
228 | } |
229 | |
230 | void EnsureInterceptorsInitialized(); |
231 | |
232 | /// Intercept thread creation operation to allocate and setup the unsafe stack |
233 | INTERCEPTOR(int, pthread_create, pthread_t *thread, |
234 | const pthread_attr_t *attr, |
235 | void *(*start_routine)(void*), void *arg) { |
236 | EnsureInterceptorsInitialized(); |
237 | size_t size = 0; |
238 | size_t guard = 0; |
239 | |
240 | if (attr) { |
241 | pthread_attr_getstacksize(attr: attr, stacksize: &size); |
242 | pthread_attr_getguardsize(attr: attr, guardsize: &guard); |
243 | } else { |
244 | // get pthread default stack size |
245 | pthread_attr_t tmpattr; |
246 | pthread_attr_init(attr: &tmpattr); |
247 | pthread_attr_getstacksize(attr: &tmpattr, stacksize: &size); |
248 | pthread_attr_getguardsize(attr: &tmpattr, guardsize: &guard); |
249 | pthread_attr_destroy(attr: &tmpattr); |
250 | } |
251 | |
252 | #if SANITIZER_SOLARIS |
253 | // Solaris pthread_attr_init initializes stacksize to 0 (the default), so |
254 | // hardcode the actual values as documented in pthread_create(3C). |
255 | if (size == 0) |
256 | # if defined(_LP64) |
257 | size = 2 * 1024 * 1024; |
258 | # else |
259 | size = 1024 * 1024; |
260 | # endif |
261 | #endif |
262 | |
263 | SFS_CHECK(size); |
264 | size = RoundUpTo(size, boundary: kStackAlign); |
265 | |
266 | void *addr = unsafe_stack_alloc(size, guard); |
267 | // Put tinfo at the end of the buffer. guard may be not page aligned. |
268 | // If that is so then some bytes after addr can be mprotected. |
269 | struct tinfo *tinfo = |
270 | (struct tinfo *)(((char *)addr) + size - sizeof(struct tinfo)); |
271 | tinfo->start_routine = start_routine; |
272 | tinfo->start_routine_arg = arg; |
273 | tinfo->unsafe_stack_start = addr; |
274 | tinfo->unsafe_stack_size = size; |
275 | tinfo->unsafe_stack_guard = guard; |
276 | |
277 | return REAL(pthread_create)(thread, attr, thread_start, tinfo); |
278 | } |
279 | |
280 | pthread_mutex_t interceptor_init_mutex = PTHREAD_MUTEX_INITIALIZER; |
281 | bool interceptors_inited = false; |
282 | |
283 | void EnsureInterceptorsInitialized() { |
284 | MutexLock lock(interceptor_init_mutex); |
285 | if (interceptors_inited) |
286 | return; |
287 | |
288 | // Initialize pthread interceptors for thread allocation |
289 | INTERCEPT_FUNCTION(pthread_create); |
290 | |
291 | interceptors_inited = true; |
292 | } |
293 | |
294 | } // namespace |
295 | |
296 | extern "C" __attribute__((visibility("default" ))) |
297 | #if !SANITIZER_CAN_USE_PREINIT_ARRAY |
298 | // On ELF platforms, the constructor is invoked using .preinit_array (see below) |
299 | __attribute__((constructor(0))) |
300 | #endif |
301 | void __safestack_init() { |
302 | // Determine the stack size for the main thread. |
303 | size_t size = kDefaultUnsafeStackSize; |
304 | size_t guard = 4096; |
305 | |
306 | struct rlimit limit; |
307 | if (getrlimit(RLIMIT_STACK, rlimits: &limit) == 0 && limit.rlim_cur != RLIM_INFINITY) |
308 | size = limit.rlim_cur; |
309 | |
310 | // Allocate unsafe stack for main thread |
311 | void *addr = unsafe_stack_alloc(size, guard); |
312 | unsafe_stack_setup(start: addr, size, guard); |
313 | |
314 | // Setup the cleanup handler |
315 | pthread_key_create(key: &thread_cleanup_key, destr_function: thread_cleanup_handler); |
316 | } |
317 | |
318 | #if SANITIZER_CAN_USE_PREINIT_ARRAY |
319 | // On ELF platforms, run safestack initialization before any other constructors. |
320 | // On other platforms we use the constructor attribute to arrange to run our |
321 | // initialization early. |
322 | extern "C" { |
323 | __attribute__((section(".preinit_array" ), |
324 | used)) void (*__safestack_preinit)(void) = __safestack_init; |
325 | } |
326 | #endif |
327 | |
328 | extern "C" |
329 | __attribute__((visibility("default" ))) void *__get_unsafe_stack_bottom() { |
330 | return unsafe_stack_start; |
331 | } |
332 | |
333 | extern "C" |
334 | __attribute__((visibility("default" ))) void *__get_unsafe_stack_top() { |
335 | return (char*)unsafe_stack_start + unsafe_stack_size; |
336 | } |
337 | |
338 | extern "C" |
339 | __attribute__((visibility("default" ))) void *__get_unsafe_stack_start() { |
340 | return unsafe_stack_start; |
341 | } |
342 | |
343 | extern "C" |
344 | __attribute__((visibility("default" ))) void *__get_unsafe_stack_ptr() { |
345 | return __safestack_unsafe_stack_ptr; |
346 | } |
347 | |