1/* Copyright (C) 2002-2023 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3
4 The GNU C Library is free software; you can redistribute it and/or
5 modify it under the terms of the GNU Lesser General Public
6 License as published by the Free Software Foundation; either
7 version 2.1 of the License, or (at your option) any later version.
8
9 The GNU C Library is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 Lesser General Public License for more details.
13
14 You should have received a copy of the GNU Lesser General Public
15 License along with the GNU C Library; if not, see
16 <https://www.gnu.org/licenses/>. */
17
18#include <assert.h>
19#include <errno.h>
20#include <signal.h>
21#include <stdint.h>
22#include <string.h>
23#include <unistd.h>
24#include <sys/mman.h>
25#include <sys/param.h>
26#include <dl-sysdep.h>
27#include <dl-tls.h>
28#include <tls.h>
29#include <list.h>
30#include <lowlevellock.h>
31#include <futex-internal.h>
32#include <kernel-features.h>
33#include <nptl-stack.h>
34#include <libc-lock.h>
35#include <tls-internal.h>
36
37/* Default alignment of stack. */
38#ifndef STACK_ALIGN
39# define STACK_ALIGN __alignof__ (long double)
40#endif
41
42/* Default value for minimal stack size after allocating thread
43 descriptor and guard. */
44#ifndef MINIMAL_REST_STACK
45# define MINIMAL_REST_STACK 4096
46#endif
47
48
49/* Newer kernels have the MAP_STACK flag to indicate a mapping is used for
50 a stack. Use it when possible. */
51#ifndef MAP_STACK
52# define MAP_STACK 0
53#endif
54
55/* Get a stack frame from the cache. We have to match by size since
56 some blocks might be too small or far too large. */
57static struct pthread *
58get_cached_stack (size_t *sizep, void **memp)
59{
60 size_t size = *sizep;
61 struct pthread *result = NULL;
62 list_t *entry;
63
64 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
65
66 /* Search the cache for a matching entry. We search for the
67 smallest stack which has at least the required size. Note that
68 in normal situations the size of all allocated stacks is the
69 same. As the very least there are only a few different sizes.
70 Therefore this loop will exit early most of the time with an
71 exact match. */
72 list_for_each (entry, &GL (dl_stack_cache))
73 {
74 struct pthread *curr;
75
76 curr = list_entry (entry, struct pthread, list);
77 if (__nptl_stack_in_use (curr) && curr->stackblock_size >= size)
78 {
79 if (curr->stackblock_size == size)
80 {
81 result = curr;
82 break;
83 }
84
85 if (result == NULL
86 || result->stackblock_size > curr->stackblock_size)
87 result = curr;
88 }
89 }
90
91 if (__builtin_expect (result == NULL, 0)
92 /* Make sure the size difference is not too excessive. In that
93 case we do not use the block. */
94 || __builtin_expect (result->stackblock_size > 4 * size, 0))
95 {
96 /* Release the lock. */
97 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
98
99 return NULL;
100 }
101
102 /* Don't allow setxid until cloned. */
103 result->setxid_futex = -1;
104
105 /* Dequeue the entry. */
106 __nptl_stack_list_del (&result->list);
107
108 /* And add to the list of stacks in use. */
109 __nptl_stack_list_add (&result->list, &GL (dl_stack_used));
110
111 /* And decrease the cache size. */
112 GL (dl_stack_cache_actsize) -= result->stackblock_size;
113
114 /* Release the lock early. */
115 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
116
117 /* Report size and location of the stack to the caller. */
118 *sizep = result->stackblock_size;
119 *memp = result->stackblock;
120
121 /* Cancellation handling is back to the default. */
122 result->cancelhandling = 0;
123 result->cleanup = NULL;
124 result->setup_failed = 0;
125
126 /* No pending event. */
127 result->nextevent = NULL;
128
129 result->exiting = false;
130 __libc_lock_init (result->exit_lock);
131 memset (&result->tls_state, 0, sizeof result->tls_state);
132
133 /* Clear the DTV. */
134 dtv_t *dtv = GET_DTV (TLS_TPADJ (result));
135 for (size_t cnt = 0; cnt < dtv[-1].counter; ++cnt)
136 free (dtv[1 + cnt].pointer.to_free);
137 memset (dtv, '\0', (dtv[-1].counter + 1) * sizeof (dtv_t));
138
139 /* Re-initialize the TLS. */
140 _dl_allocate_tls_init (TLS_TPADJ (result), true);
141
142 return result;
143}
144
145/* Return the guard page position on allocated stack. */
146static inline char *
147__attribute ((always_inline))
148guard_position (void *mem, size_t size, size_t guardsize, struct pthread *pd,
149 size_t pagesize_m1)
150{
151#ifdef NEED_SEPARATE_REGISTER_STACK
152 return mem + (((size - guardsize) / 2) & ~pagesize_m1);
153#elif _STACK_GROWS_DOWN
154 return mem;
155#elif _STACK_GROWS_UP
156 return (char *) (((uintptr_t) pd - guardsize) & ~pagesize_m1);
157#endif
158}
159
160/* Based on stack allocated with PROT_NONE, setup the required portions with
161 'prot' flags based on the guard page position. */
162static inline int
163setup_stack_prot (char *mem, size_t size, char *guard, size_t guardsize,
164 const int prot)
165{
166 char *guardend = guard + guardsize;
167#if _STACK_GROWS_DOWN && !defined(NEED_SEPARATE_REGISTER_STACK)
168 /* As defined at guard_position, for architectures with downward stack
169 the guard page is always at start of the allocated area. */
170 if (__mprotect (guardend, size - guardsize, prot) != 0)
171 return errno;
172#else
173 size_t mprots1 = (uintptr_t) guard - (uintptr_t) mem;
174 if (__mprotect (mem, mprots1, prot) != 0)
175 return errno;
176 size_t mprots2 = ((uintptr_t) mem + size) - (uintptr_t) guardend;
177 if (__mprotect (guardend, mprots2, prot) != 0)
178 return errno;
179#endif
180 return 0;
181}
182
183/* Mark the memory of the stack as usable to the kernel. It frees everything
184 except for the space used for the TCB itself. */
185static __always_inline void
186advise_stack_range (void *mem, size_t size, uintptr_t pd, size_t guardsize)
187{
188 uintptr_t sp = (uintptr_t) CURRENT_STACK_FRAME;
189 size_t pagesize_m1 = __getpagesize () - 1;
190#if _STACK_GROWS_DOWN && !defined(NEED_SEPARATE_REGISTER_STACK)
191 size_t freesize = (sp - (uintptr_t) mem) & ~pagesize_m1;
192 assert (freesize < size);
193 if (freesize > PTHREAD_STACK_MIN)
194 __madvise (mem, freesize - PTHREAD_STACK_MIN, MADV_DONTNEED);
195#else
196 /* Page aligned start of memory to free (higher than or equal
197 to current sp plus the minimum stack size). */
198 uintptr_t freeblock = (sp + PTHREAD_STACK_MIN + pagesize_m1) & ~pagesize_m1;
199 uintptr_t free_end = (pd - guardsize) & ~pagesize_m1;
200 if (free_end > freeblock)
201 {
202 size_t freesize = free_end - freeblock;
203 assert (freesize < size);
204 __madvise ((void*) freeblock, freesize, MADV_DONTNEED);
205 }
206#endif
207}
208
209/* Returns a usable stack for a new thread either by allocating a
210 new stack or reusing a cached stack of sufficient size.
211 ATTR must be non-NULL and point to a valid pthread_attr.
212 PDP must be non-NULL. */
213static int
214allocate_stack (const struct pthread_attr *attr, struct pthread **pdp,
215 void **stack, size_t *stacksize)
216{
217 struct pthread *pd;
218 size_t size;
219 size_t pagesize_m1 = __getpagesize () - 1;
220 size_t tls_static_size_for_stack = __nptl_tls_static_size_for_stack ();
221 size_t tls_static_align_m1 = GLRO (dl_tls_static_align) - 1;
222
223 assert (powerof2 (pagesize_m1 + 1));
224 assert (TCB_ALIGNMENT >= STACK_ALIGN);
225
226 /* Get the stack size from the attribute if it is set. Otherwise we
227 use the default we determined at start time. */
228 if (attr->stacksize != 0)
229 size = attr->stacksize;
230 else
231 {
232 lll_lock (__default_pthread_attr_lock, LLL_PRIVATE);
233 size = __default_pthread_attr.internal.stacksize;
234 lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE);
235 }
236
237 /* Get memory for the stack. */
238 if (__glibc_unlikely (attr->flags & ATTR_FLAG_STACKADDR))
239 {
240 uintptr_t adj;
241 char *stackaddr = (char *) attr->stackaddr;
242
243 /* Assume the same layout as the _STACK_GROWS_DOWN case, with struct
244 pthread at the top of the stack block. Later we adjust the guard
245 location and stack address to match the _STACK_GROWS_UP case. */
246 if (_STACK_GROWS_UP)
247 stackaddr += attr->stacksize;
248
249 /* If the user also specified the size of the stack make sure it
250 is large enough. */
251 if (attr->stacksize != 0
252 && attr->stacksize < (tls_static_size_for_stack
253 + MINIMAL_REST_STACK))
254 return EINVAL;
255
256 /* Adjust stack size for alignment of the TLS block. */
257#if TLS_TCB_AT_TP
258 adj = ((uintptr_t) stackaddr - TLS_TCB_SIZE)
259 & tls_static_align_m1;
260 assert (size > adj + TLS_TCB_SIZE);
261#elif TLS_DTV_AT_TP
262 adj = ((uintptr_t) stackaddr - tls_static_size_for_stack)
263 & tls_static_align_m1;
264 assert (size > adj);
265#endif
266
267 /* The user provided some memory. Let's hope it matches the
268 size... We do not allocate guard pages if the user provided
269 the stack. It is the user's responsibility to do this if it
270 is wanted. */
271#if TLS_TCB_AT_TP
272 pd = (struct pthread *) ((uintptr_t) stackaddr
273 - TLS_TCB_SIZE - adj);
274#elif TLS_DTV_AT_TP
275 pd = (struct pthread *) (((uintptr_t) stackaddr
276 - tls_static_size_for_stack - adj)
277 - TLS_PRE_TCB_SIZE);
278#endif
279
280 /* The user provided stack memory needs to be cleared. */
281 memset (pd, '\0', sizeof (struct pthread));
282
283 /* The first TSD block is included in the TCB. */
284 pd->specific[0] = pd->specific_1stblock;
285
286 /* Remember the stack-related values. */
287 pd->stackblock = (char *) stackaddr - size;
288 pd->stackblock_size = size;
289
290 /* This is a user-provided stack. It will not be queued in the
291 stack cache nor will the memory (except the TLS memory) be freed. */
292 pd->user_stack = true;
293
294 /* This is at least the second thread. */
295 pd->header.multiple_threads = 1;
296
297#ifdef NEED_DL_SYSINFO
298 SETUP_THREAD_SYSINFO (pd);
299#endif
300
301 /* Don't allow setxid until cloned. */
302 pd->setxid_futex = -1;
303
304 /* Allocate the DTV for this thread. */
305 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
306 {
307 /* Something went wrong. */
308 assert (errno == ENOMEM);
309 return errno;
310 }
311
312
313 /* Prepare to modify global data. */
314 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
315
316 /* And add to the list of stacks in use. */
317 list_add (&pd->list, &GL (dl_stack_user));
318
319 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
320 }
321 else
322 {
323 /* Allocate some anonymous memory. If possible use the cache. */
324 size_t guardsize;
325 size_t reported_guardsize;
326 size_t reqsize;
327 void *mem;
328 const int prot = (PROT_READ | PROT_WRITE
329 | ((GL(dl_stack_flags) & PF_X) ? PROT_EXEC : 0));
330
331 /* Adjust the stack size for alignment. */
332 size &= ~tls_static_align_m1;
333 assert (size != 0);
334
335 /* Make sure the size of the stack is enough for the guard and
336 eventually the thread descriptor. On some targets there is
337 a minimum guard size requirement, ARCH_MIN_GUARD_SIZE, so
338 internally enforce it (unless the guard was disabled), but
339 report the original guard size for backward compatibility:
340 before POSIX 2008 the guardsize was specified to be one page
341 by default which is observable via pthread_attr_getguardsize
342 and pthread_getattr_np. */
343 guardsize = (attr->guardsize + pagesize_m1) & ~pagesize_m1;
344 reported_guardsize = guardsize;
345 if (guardsize > 0 && guardsize < ARCH_MIN_GUARD_SIZE)
346 guardsize = ARCH_MIN_GUARD_SIZE;
347 if (guardsize < attr->guardsize || size + guardsize < guardsize)
348 /* Arithmetic overflow. */
349 return EINVAL;
350 size += guardsize;
351 if (__builtin_expect (size < ((guardsize + tls_static_size_for_stack
352 + MINIMAL_REST_STACK + pagesize_m1)
353 & ~pagesize_m1),
354 0))
355 /* The stack is too small (or the guard too large). */
356 return EINVAL;
357
358 /* Try to get a stack from the cache. */
359 reqsize = size;
360 pd = get_cached_stack (&size, &mem);
361 if (pd == NULL)
362 {
363 /* If a guard page is required, avoid committing memory by first
364 allocate with PROT_NONE and then reserve with required permission
365 excluding the guard page. */
366 mem = __mmap (NULL, size, (guardsize == 0) ? prot : PROT_NONE,
367 MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
368
369 if (__glibc_unlikely (mem == MAP_FAILED))
370 return errno;
371
372 /* Do madvise in case the tunable glibc.pthread.stack_hugetlb is
373 set to 0, disabling hugetlb. */
374 if (__glibc_unlikely (__nptl_stack_hugetlb == 0)
375 && __madvise (mem, size, MADV_NOHUGEPAGE) != 0)
376 return errno;
377
378 /* SIZE is guaranteed to be greater than zero.
379 So we can never get a null pointer back from mmap. */
380 assert (mem != NULL);
381
382 /* Place the thread descriptor at the end of the stack. */
383#if TLS_TCB_AT_TP
384 pd = (struct pthread *) ((((uintptr_t) mem + size)
385 - TLS_TCB_SIZE)
386 & ~tls_static_align_m1);
387#elif TLS_DTV_AT_TP
388 pd = (struct pthread *) ((((uintptr_t) mem + size
389 - tls_static_size_for_stack)
390 & ~tls_static_align_m1)
391 - TLS_PRE_TCB_SIZE);
392#endif
393
394 /* Now mprotect the required region excluding the guard area. */
395 if (__glibc_likely (guardsize > 0))
396 {
397 char *guard = guard_position (mem, size, guardsize, pd,
398 pagesize_m1);
399 if (setup_stack_prot (mem, size, guard, guardsize, prot) != 0)
400 {
401 __munmap (mem, size);
402 return errno;
403 }
404 }
405
406 /* Remember the stack-related values. */
407 pd->stackblock = mem;
408 pd->stackblock_size = size;
409 /* Update guardsize for newly allocated guardsize to avoid
410 an mprotect in guard resize below. */
411 pd->guardsize = guardsize;
412
413 /* We allocated the first block thread-specific data array.
414 This address will not change for the lifetime of this
415 descriptor. */
416 pd->specific[0] = pd->specific_1stblock;
417
418 /* This is at least the second thread. */
419 pd->header.multiple_threads = 1;
420
421#ifdef NEED_DL_SYSINFO
422 SETUP_THREAD_SYSINFO (pd);
423#endif
424
425 /* Don't allow setxid until cloned. */
426 pd->setxid_futex = -1;
427
428 /* Allocate the DTV for this thread. */
429 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
430 {
431 /* Something went wrong. */
432 assert (errno == ENOMEM);
433
434 /* Free the stack memory we just allocated. */
435 (void) __munmap (mem, size);
436
437 return errno;
438 }
439
440
441 /* Prepare to modify global data. */
442 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
443
444 /* And add to the list of stacks in use. */
445 __nptl_stack_list_add (&pd->list, &GL (dl_stack_used));
446
447 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
448
449
450 /* There might have been a race. Another thread might have
451 caused the stacks to get exec permission while this new
452 stack was prepared. Detect if this was possible and
453 change the permission if necessary. */
454 if (__builtin_expect ((GL(dl_stack_flags) & PF_X) != 0
455 && (prot & PROT_EXEC) == 0, 0))
456 {
457 int err = __nptl_change_stack_perm (pd);
458 if (err != 0)
459 {
460 /* Free the stack memory we just allocated. */
461 (void) __munmap (mem, size);
462
463 return err;
464 }
465 }
466
467
468 /* Note that all of the stack and the thread descriptor is
469 zeroed. This means we do not have to initialize fields
470 with initial value zero. This is specifically true for
471 the 'tid' field which is always set back to zero once the
472 stack is not used anymore and for the 'guardsize' field
473 which will be read next. */
474 }
475
476 /* Create or resize the guard area if necessary. */
477 if (__glibc_unlikely (guardsize > pd->guardsize))
478 {
479 char *guard = guard_position (mem, size, guardsize, pd,
480 pagesize_m1);
481 if (__mprotect (guard, guardsize, PROT_NONE) != 0)
482 {
483 mprot_error:
484 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
485
486 /* Remove the thread from the list. */
487 __nptl_stack_list_del (&pd->list);
488
489 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
490
491 /* Get rid of the TLS block we allocated. */
492 _dl_deallocate_tls (TLS_TPADJ (pd), false);
493
494 /* Free the stack memory regardless of whether the size
495 of the cache is over the limit or not. If this piece
496 of memory caused problems we better do not use it
497 anymore. Uh, and we ignore possible errors. There
498 is nothing we could do. */
499 (void) __munmap (mem, size);
500
501 return errno;
502 }
503
504 pd->guardsize = guardsize;
505 }
506 else if (__builtin_expect (pd->guardsize - guardsize > size - reqsize,
507 0))
508 {
509 /* The old guard area is too large. */
510
511#ifdef NEED_SEPARATE_REGISTER_STACK
512 char *guard = mem + (((size - guardsize) / 2) & ~pagesize_m1);
513 char *oldguard = mem + (((size - pd->guardsize) / 2) & ~pagesize_m1);
514
515 if (oldguard < guard
516 && __mprotect (oldguard, guard - oldguard, prot) != 0)
517 goto mprot_error;
518
519 if (__mprotect (guard + guardsize,
520 oldguard + pd->guardsize - guard - guardsize,
521 prot) != 0)
522 goto mprot_error;
523#elif _STACK_GROWS_DOWN
524 if (__mprotect ((char *) mem + guardsize, pd->guardsize - guardsize,
525 prot) != 0)
526 goto mprot_error;
527#elif _STACK_GROWS_UP
528 char *new_guard = (char *)(((uintptr_t) pd - guardsize)
529 & ~pagesize_m1);
530 char *old_guard = (char *)(((uintptr_t) pd - pd->guardsize)
531 & ~pagesize_m1);
532 /* The guard size difference might be > 0, but once rounded
533 to the nearest page the size difference might be zero. */
534 if (new_guard > old_guard
535 && __mprotect (old_guard, new_guard - old_guard, prot) != 0)
536 goto mprot_error;
537#endif
538
539 pd->guardsize = guardsize;
540 }
541 /* The pthread_getattr_np() calls need to get passed the size
542 requested in the attribute, regardless of how large the
543 actually used guardsize is. */
544 pd->reported_guardsize = reported_guardsize;
545 }
546
547 /* Initialize the lock. We have to do this unconditionally since the
548 stillborn thread could be canceled while the lock is taken. */
549 pd->lock = LLL_LOCK_INITIALIZER;
550
551 /* The robust mutex lists also need to be initialized
552 unconditionally because the cleanup for the previous stack owner
553 might have happened in the kernel. */
554 pd->robust_head.futex_offset = (offsetof (pthread_mutex_t, __data.__lock)
555 - offsetof (pthread_mutex_t,
556 __data.__list.__next));
557 pd->robust_head.list_op_pending = NULL;
558#if __PTHREAD_MUTEX_HAVE_PREV
559 pd->robust_prev = &pd->robust_head;
560#endif
561 pd->robust_head.list = &pd->robust_head;
562
563 /* We place the thread descriptor at the end of the stack. */
564 *pdp = pd;
565
566 void *stacktop;
567
568#if TLS_TCB_AT_TP
569 /* The stack begins before the TCB and the static TLS block. */
570 stacktop = ((char *) (pd + 1) - tls_static_size_for_stack);
571#elif TLS_DTV_AT_TP
572 stacktop = (char *) (pd - 1);
573#endif
574
575 *stacksize = stacktop - pd->stackblock;
576 *stack = pd->stackblock;
577
578 return 0;
579}
580