1/* Copyright (C) 2002-2020 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3 Contributed by Ulrich Drepper <drepper@redhat.com>, 2002.
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#include <assert.h>
20#include <errno.h>
21#include <signal.h>
22#include <stdint.h>
23#include <string.h>
24#include <unistd.h>
25#include <sys/mman.h>
26#include <sys/param.h>
27#include <dl-sysdep.h>
28#include <dl-tls.h>
29#include <tls.h>
30#include <list.h>
31#include <lowlevellock.h>
32#include <futex-internal.h>
33#include <kernel-features.h>
34#include <stack-aliasing.h>
35
36
37#ifndef NEED_SEPARATE_REGISTER_STACK
38
39/* Most architectures have exactly one stack pointer. Some have more. */
40# define STACK_VARIABLES void *stackaddr = NULL
41
42/* How to pass the values to the 'create_thread' function. */
43# define STACK_VARIABLES_ARGS stackaddr
44
45/* How to declare function which gets there parameters. */
46# define STACK_VARIABLES_PARMS void *stackaddr
47
48/* How to declare allocate_stack. */
49# define ALLOCATE_STACK_PARMS void **stack
50
51/* This is how the function is called. We do it this way to allow
52 other variants of the function to have more parameters. */
53# define ALLOCATE_STACK(attr, pd) allocate_stack (attr, pd, &stackaddr)
54
55#else
56
57/* We need two stacks. The kernel will place them but we have to tell
58 the kernel about the size of the reserved address space. */
59# define STACK_VARIABLES void *stackaddr = NULL; size_t stacksize = 0
60
61/* How to pass the values to the 'create_thread' function. */
62# define STACK_VARIABLES_ARGS stackaddr, stacksize
63
64/* How to declare function which gets there parameters. */
65# define STACK_VARIABLES_PARMS void *stackaddr, size_t stacksize
66
67/* How to declare allocate_stack. */
68# define ALLOCATE_STACK_PARMS void **stack, size_t *stacksize
69
70/* This is how the function is called. We do it this way to allow
71 other variants of the function to have more parameters. */
72# define ALLOCATE_STACK(attr, pd) \
73 allocate_stack (attr, pd, &stackaddr, &stacksize)
74
75#endif
76
77
78/* Default alignment of stack. */
79#ifndef STACK_ALIGN
80# define STACK_ALIGN __alignof__ (long double)
81#endif
82
83/* Default value for minimal stack size after allocating thread
84 descriptor and guard. */
85#ifndef MINIMAL_REST_STACK
86# define MINIMAL_REST_STACK 4096
87#endif
88
89
90/* Newer kernels have the MAP_STACK flag to indicate a mapping is used for
91 a stack. Use it when possible. */
92#ifndef MAP_STACK
93# define MAP_STACK 0
94#endif
95
96/* This yields the pointer that TLS support code calls the thread pointer. */
97#if TLS_TCB_AT_TP
98# define TLS_TPADJ(pd) (pd)
99#elif TLS_DTV_AT_TP
100# define TLS_TPADJ(pd) ((struct pthread *)((char *) (pd) + TLS_PRE_TCB_SIZE))
101#endif
102
103/* Cache handling for not-yet free stacks. */
104
105/* Maximum size in kB of cache. */
106static size_t stack_cache_maxsize = 40 * 1024 * 1024; /* 40MiBi by default. */
107static size_t stack_cache_actsize;
108
109/* Mutex protecting this variable. */
110static int stack_cache_lock = LLL_LOCK_INITIALIZER;
111
112/* List of queued stack frames. */
113static LIST_HEAD (stack_cache);
114
115/* List of the stacks in use. */
116static LIST_HEAD (stack_used);
117
118/* We need to record what list operations we are going to do so that,
119 in case of an asynchronous interruption due to a fork() call, we
120 can correct for the work. */
121static uintptr_t in_flight_stack;
122
123/* List of the threads with user provided stacks in use. No need to
124 initialize this, since it's done in __pthread_initialize_minimal. */
125list_t __stack_user __attribute__ ((nocommon));
126hidden_data_def (__stack_user)
127
128
129/* Check whether the stack is still used or not. */
130#define FREE_P(descr) ((descr)->tid <= 0)
131
132
133static void
134stack_list_del (list_t *elem)
135{
136 in_flight_stack = (uintptr_t) elem;
137
138 atomic_write_barrier ();
139
140 list_del (elem);
141
142 atomic_write_barrier ();
143
144 in_flight_stack = 0;
145}
146
147
148static void
149stack_list_add (list_t *elem, list_t *list)
150{
151 in_flight_stack = (uintptr_t) elem | 1;
152
153 atomic_write_barrier ();
154
155 list_add (elem, list);
156
157 atomic_write_barrier ();
158
159 in_flight_stack = 0;
160}
161
162
163/* We create a double linked list of all cache entries. Double linked
164 because this allows removing entries from the end. */
165
166
167/* Get a stack frame from the cache. We have to match by size since
168 some blocks might be too small or far too large. */
169static struct pthread *
170get_cached_stack (size_t *sizep, void **memp)
171{
172 size_t size = *sizep;
173 struct pthread *result = NULL;
174 list_t *entry;
175
176 lll_lock (stack_cache_lock, LLL_PRIVATE);
177
178 /* Search the cache for a matching entry. We search for the
179 smallest stack which has at least the required size. Note that
180 in normal situations the size of all allocated stacks is the
181 same. As the very least there are only a few different sizes.
182 Therefore this loop will exit early most of the time with an
183 exact match. */
184 list_for_each (entry, &stack_cache)
185 {
186 struct pthread *curr;
187
188 curr = list_entry (entry, struct pthread, list);
189 if (FREE_P (curr) && curr->stackblock_size >= size)
190 {
191 if (curr->stackblock_size == size)
192 {
193 result = curr;
194 break;
195 }
196
197 if (result == NULL
198 || result->stackblock_size > curr->stackblock_size)
199 result = curr;
200 }
201 }
202
203 if (__builtin_expect (result == NULL, 0)
204 /* Make sure the size difference is not too excessive. In that
205 case we do not use the block. */
206 || __builtin_expect (result->stackblock_size > 4 * size, 0))
207 {
208 /* Release the lock. */
209 lll_unlock (stack_cache_lock, LLL_PRIVATE);
210
211 return NULL;
212 }
213
214 /* Don't allow setxid until cloned. */
215 result->setxid_futex = -1;
216
217 /* Dequeue the entry. */
218 stack_list_del (&result->list);
219
220 /* And add to the list of stacks in use. */
221 stack_list_add (&result->list, &stack_used);
222
223 /* And decrease the cache size. */
224 stack_cache_actsize -= result->stackblock_size;
225
226 /* Release the lock early. */
227 lll_unlock (stack_cache_lock, LLL_PRIVATE);
228
229 /* Report size and location of the stack to the caller. */
230 *sizep = result->stackblock_size;
231 *memp = result->stackblock;
232
233 /* Cancellation handling is back to the default. */
234 result->cancelhandling = 0;
235 result->cleanup = NULL;
236
237 /* No pending event. */
238 result->nextevent = NULL;
239
240 /* Clear the DTV. */
241 dtv_t *dtv = GET_DTV (TLS_TPADJ (result));
242 for (size_t cnt = 0; cnt < dtv[-1].counter; ++cnt)
243 free (dtv[1 + cnt].pointer.to_free);
244 memset (dtv, '\0', (dtv[-1].counter + 1) * sizeof (dtv_t));
245
246 /* Re-initialize the TLS. */
247 _dl_allocate_tls_init (TLS_TPADJ (result));
248
249 return result;
250}
251
252
253/* Free stacks until cache size is lower than LIMIT. */
254static void
255free_stacks (size_t limit)
256{
257 /* We reduce the size of the cache. Remove the last entries until
258 the size is below the limit. */
259 list_t *entry;
260 list_t *prev;
261
262 /* Search from the end of the list. */
263 list_for_each_prev_safe (entry, prev, &stack_cache)
264 {
265 struct pthread *curr;
266
267 curr = list_entry (entry, struct pthread, list);
268 if (FREE_P (curr))
269 {
270 /* Unlink the block. */
271 stack_list_del (entry);
272
273 /* Account for the freed memory. */
274 stack_cache_actsize -= curr->stackblock_size;
275
276 /* Free the memory associated with the ELF TLS. */
277 _dl_deallocate_tls (TLS_TPADJ (curr), false);
278
279 /* Remove this block. This should never fail. If it does
280 something is really wrong. */
281 if (__munmap (curr->stackblock, curr->stackblock_size) != 0)
282 abort ();
283
284 /* Maybe we have freed enough. */
285 if (stack_cache_actsize <= limit)
286 break;
287 }
288 }
289}
290
291/* Free all the stacks on cleanup. */
292void
293__nptl_stacks_freeres (void)
294{
295 free_stacks (0);
296}
297
298/* Add a stack frame which is not used anymore to the stack. Must be
299 called with the cache lock held. */
300static inline void
301__attribute ((always_inline))
302queue_stack (struct pthread *stack)
303{
304 /* We unconditionally add the stack to the list. The memory may
305 still be in use but it will not be reused until the kernel marks
306 the stack as not used anymore. */
307 stack_list_add (&stack->list, &stack_cache);
308
309 stack_cache_actsize += stack->stackblock_size;
310 if (__glibc_unlikely (stack_cache_actsize > stack_cache_maxsize))
311 free_stacks (stack_cache_maxsize);
312}
313
314
315static int
316change_stack_perm (struct pthread *pd
317#ifdef NEED_SEPARATE_REGISTER_STACK
318 , size_t pagemask
319#endif
320 )
321{
322#ifdef NEED_SEPARATE_REGISTER_STACK
323 void *stack = (pd->stackblock
324 + (((((pd->stackblock_size - pd->guardsize) / 2)
325 & pagemask) + pd->guardsize) & pagemask));
326 size_t len = pd->stackblock + pd->stackblock_size - stack;
327#elif _STACK_GROWS_DOWN
328 void *stack = pd->stackblock + pd->guardsize;
329 size_t len = pd->stackblock_size - pd->guardsize;
330#elif _STACK_GROWS_UP
331 void *stack = pd->stackblock;
332 size_t len = (uintptr_t) pd - pd->guardsize - (uintptr_t) pd->stackblock;
333#else
334# error "Define either _STACK_GROWS_DOWN or _STACK_GROWS_UP"
335#endif
336 if (__mprotect (stack, len, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
337 return errno;
338
339 return 0;
340}
341
342/* Return the guard page position on allocated stack. */
343static inline char *
344__attribute ((always_inline))
345guard_position (void *mem, size_t size, size_t guardsize, struct pthread *pd,
346 size_t pagesize_m1)
347{
348#ifdef NEED_SEPARATE_REGISTER_STACK
349 return mem + (((size - guardsize) / 2) & ~pagesize_m1);
350#elif _STACK_GROWS_DOWN
351 return mem;
352#elif _STACK_GROWS_UP
353 return (char *) (((uintptr_t) pd - guardsize) & ~pagesize_m1);
354#endif
355}
356
357/* Based on stack allocated with PROT_NONE, setup the required portions with
358 'prot' flags based on the guard page position. */
359static inline int
360setup_stack_prot (char *mem, size_t size, char *guard, size_t guardsize,
361 const int prot)
362{
363 char *guardend = guard + guardsize;
364#if _STACK_GROWS_DOWN && !defined(NEED_SEPARATE_REGISTER_STACK)
365 /* As defined at guard_position, for architectures with downward stack
366 the guard page is always at start of the allocated area. */
367 if (__mprotect (guardend, size - guardsize, prot) != 0)
368 return errno;
369#else
370 size_t mprots1 = (uintptr_t) guard - (uintptr_t) mem;
371 if (__mprotect (mem, mprots1, prot) != 0)
372 return errno;
373 size_t mprots2 = ((uintptr_t) mem + size) - (uintptr_t) guardend;
374 if (__mprotect (guardend, mprots2, prot) != 0)
375 return errno;
376#endif
377 return 0;
378}
379
380/* Mark the memory of the stack as usable to the kernel. It frees everything
381 except for the space used for the TCB itself. */
382static __always_inline void
383advise_stack_range (void *mem, size_t size, uintptr_t pd, size_t guardsize)
384{
385 uintptr_t sp = (uintptr_t) CURRENT_STACK_FRAME;
386 size_t pagesize_m1 = __getpagesize () - 1;
387#if _STACK_GROWS_DOWN && !defined(NEED_SEPARATE_REGISTER_STACK)
388 size_t freesize = (sp - (uintptr_t) mem) & ~pagesize_m1;
389 assert (freesize < size);
390 if (freesize > PTHREAD_STACK_MIN)
391 __madvise (mem, freesize - PTHREAD_STACK_MIN, MADV_DONTNEED);
392#else
393 /* Page aligned start of memory to free (higher than or equal
394 to current sp plus the minimum stack size). */
395 uintptr_t freeblock = (sp + PTHREAD_STACK_MIN + pagesize_m1) & ~pagesize_m1;
396 uintptr_t free_end = (pd - guardsize) & ~pagesize_m1;
397 if (free_end > freeblock)
398 {
399 size_t freesize = free_end - freeblock;
400 assert (freesize < size);
401 __madvise ((void*) freeblock, freesize, MADV_DONTNEED);
402 }
403#endif
404}
405
406/* Returns a usable stack for a new thread either by allocating a
407 new stack or reusing a cached stack of sufficient size.
408 ATTR must be non-NULL and point to a valid pthread_attr.
409 PDP must be non-NULL. */
410static int
411allocate_stack (const struct pthread_attr *attr, struct pthread **pdp,
412 ALLOCATE_STACK_PARMS)
413{
414 struct pthread *pd;
415 size_t size;
416 size_t pagesize_m1 = __getpagesize () - 1;
417
418 assert (powerof2 (pagesize_m1 + 1));
419 assert (TCB_ALIGNMENT >= STACK_ALIGN);
420
421 /* Get the stack size from the attribute if it is set. Otherwise we
422 use the default we determined at start time. */
423 if (attr->stacksize != 0)
424 size = attr->stacksize;
425 else
426 {
427 lll_lock (__default_pthread_attr_lock, LLL_PRIVATE);
428 size = __default_pthread_attr.stacksize;
429 lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE);
430 }
431
432 /* Get memory for the stack. */
433 if (__glibc_unlikely (attr->flags & ATTR_FLAG_STACKADDR))
434 {
435 uintptr_t adj;
436 char *stackaddr = (char *) attr->stackaddr;
437
438 /* Assume the same layout as the _STACK_GROWS_DOWN case, with struct
439 pthread at the top of the stack block. Later we adjust the guard
440 location and stack address to match the _STACK_GROWS_UP case. */
441 if (_STACK_GROWS_UP)
442 stackaddr += attr->stacksize;
443
444 /* If the user also specified the size of the stack make sure it
445 is large enough. */
446 if (attr->stacksize != 0
447 && attr->stacksize < (__static_tls_size + MINIMAL_REST_STACK))
448 return EINVAL;
449
450 /* Adjust stack size for alignment of the TLS block. */
451#if TLS_TCB_AT_TP
452 adj = ((uintptr_t) stackaddr - TLS_TCB_SIZE)
453 & __static_tls_align_m1;
454 assert (size > adj + TLS_TCB_SIZE);
455#elif TLS_DTV_AT_TP
456 adj = ((uintptr_t) stackaddr - __static_tls_size)
457 & __static_tls_align_m1;
458 assert (size > adj);
459#endif
460
461 /* The user provided some memory. Let's hope it matches the
462 size... We do not allocate guard pages if the user provided
463 the stack. It is the user's responsibility to do this if it
464 is wanted. */
465#if TLS_TCB_AT_TP
466 pd = (struct pthread *) ((uintptr_t) stackaddr
467 - TLS_TCB_SIZE - adj);
468#elif TLS_DTV_AT_TP
469 pd = (struct pthread *) (((uintptr_t) stackaddr
470 - __static_tls_size - adj)
471 - TLS_PRE_TCB_SIZE);
472#endif
473
474 /* The user provided stack memory needs to be cleared. */
475 memset (pd, '\0', sizeof (struct pthread));
476
477 /* The first TSD block is included in the TCB. */
478 pd->specific[0] = pd->specific_1stblock;
479
480 /* Remember the stack-related values. */
481 pd->stackblock = (char *) stackaddr - size;
482 pd->stackblock_size = size;
483
484 /* This is a user-provided stack. It will not be queued in the
485 stack cache nor will the memory (except the TLS memory) be freed. */
486 pd->user_stack = true;
487
488 /* This is at least the second thread. */
489 pd->header.multiple_threads = 1;
490#ifndef TLS_MULTIPLE_THREADS_IN_TCB
491 __pthread_multiple_threads = *__libc_multiple_threads_ptr = 1;
492#endif
493
494#ifdef NEED_DL_SYSINFO
495 SETUP_THREAD_SYSINFO (pd);
496#endif
497
498 /* Don't allow setxid until cloned. */
499 pd->setxid_futex = -1;
500
501 /* Allocate the DTV for this thread. */
502 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
503 {
504 /* Something went wrong. */
505 assert (errno == ENOMEM);
506 return errno;
507 }
508
509
510 /* Prepare to modify global data. */
511 lll_lock (stack_cache_lock, LLL_PRIVATE);
512
513 /* And add to the list of stacks in use. */
514 list_add (&pd->list, &__stack_user);
515
516 lll_unlock (stack_cache_lock, LLL_PRIVATE);
517 }
518 else
519 {
520 /* Allocate some anonymous memory. If possible use the cache. */
521 size_t guardsize;
522 size_t reqsize;
523 void *mem;
524 const int prot = (PROT_READ | PROT_WRITE
525 | ((GL(dl_stack_flags) & PF_X) ? PROT_EXEC : 0));
526
527 /* Adjust the stack size for alignment. */
528 size &= ~__static_tls_align_m1;
529 assert (size != 0);
530
531 /* Make sure the size of the stack is enough for the guard and
532 eventually the thread descriptor. */
533 guardsize = (attr->guardsize + pagesize_m1) & ~pagesize_m1;
534 if (guardsize < attr->guardsize || size + guardsize < guardsize)
535 /* Arithmetic overflow. */
536 return EINVAL;
537 size += guardsize;
538 if (__builtin_expect (size < ((guardsize + __static_tls_size
539 + MINIMAL_REST_STACK + pagesize_m1)
540 & ~pagesize_m1),
541 0))
542 /* The stack is too small (or the guard too large). */
543 return EINVAL;
544
545 /* Try to get a stack from the cache. */
546 reqsize = size;
547 pd = get_cached_stack (&size, &mem);
548 if (pd == NULL)
549 {
550 /* To avoid aliasing effects on a larger scale than pages we
551 adjust the allocated stack size if necessary. This way
552 allocations directly following each other will not have
553 aliasing problems. */
554#if MULTI_PAGE_ALIASING != 0
555 if ((size % MULTI_PAGE_ALIASING) == 0)
556 size += pagesize_m1 + 1;
557#endif
558
559 /* If a guard page is required, avoid committing memory by first
560 allocate with PROT_NONE and then reserve with required permission
561 excluding the guard page. */
562 mem = __mmap (NULL, size, (guardsize == 0) ? prot : PROT_NONE,
563 MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
564
565 if (__glibc_unlikely (mem == MAP_FAILED))
566 return errno;
567
568 /* SIZE is guaranteed to be greater than zero.
569 So we can never get a null pointer back from mmap. */
570 assert (mem != NULL);
571
572 /* Place the thread descriptor at the end of the stack. */
573#if TLS_TCB_AT_TP
574 pd = (struct pthread *) ((((uintptr_t) mem + size)
575 - TLS_TCB_SIZE)
576 & ~__static_tls_align_m1);
577#elif TLS_DTV_AT_TP
578 pd = (struct pthread *) ((((uintptr_t) mem + size
579 - __static_tls_size)
580 & ~__static_tls_align_m1)
581 - TLS_PRE_TCB_SIZE);
582#endif
583
584 /* Now mprotect the required region excluding the guard area. */
585 if (__glibc_likely (guardsize > 0))
586 {
587 char *guard = guard_position (mem, size, guardsize, pd,
588 pagesize_m1);
589 if (setup_stack_prot (mem, size, guard, guardsize, prot) != 0)
590 {
591 __munmap (mem, size);
592 return errno;
593 }
594 }
595
596 /* Remember the stack-related values. */
597 pd->stackblock = mem;
598 pd->stackblock_size = size;
599 /* Update guardsize for newly allocated guardsize to avoid
600 an mprotect in guard resize below. */
601 pd->guardsize = guardsize;
602
603 /* We allocated the first block thread-specific data array.
604 This address will not change for the lifetime of this
605 descriptor. */
606 pd->specific[0] = pd->specific_1stblock;
607
608 /* This is at least the second thread. */
609 pd->header.multiple_threads = 1;
610#ifndef TLS_MULTIPLE_THREADS_IN_TCB
611 __pthread_multiple_threads = *__libc_multiple_threads_ptr = 1;
612#endif
613
614#ifdef NEED_DL_SYSINFO
615 SETUP_THREAD_SYSINFO (pd);
616#endif
617
618 /* Don't allow setxid until cloned. */
619 pd->setxid_futex = -1;
620
621 /* Allocate the DTV for this thread. */
622 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
623 {
624 /* Something went wrong. */
625 assert (errno == ENOMEM);
626
627 /* Free the stack memory we just allocated. */
628 (void) __munmap (mem, size);
629
630 return errno;
631 }
632
633
634 /* Prepare to modify global data. */
635 lll_lock (stack_cache_lock, LLL_PRIVATE);
636
637 /* And add to the list of stacks in use. */
638 stack_list_add (&pd->list, &stack_used);
639
640 lll_unlock (stack_cache_lock, LLL_PRIVATE);
641
642
643 /* There might have been a race. Another thread might have
644 caused the stacks to get exec permission while this new
645 stack was prepared. Detect if this was possible and
646 change the permission if necessary. */
647 if (__builtin_expect ((GL(dl_stack_flags) & PF_X) != 0
648 && (prot & PROT_EXEC) == 0, 0))
649 {
650 int err = change_stack_perm (pd
651#ifdef NEED_SEPARATE_REGISTER_STACK
652 , ~pagesize_m1
653#endif
654 );
655 if (err != 0)
656 {
657 /* Free the stack memory we just allocated. */
658 (void) __munmap (mem, size);
659
660 return err;
661 }
662 }
663
664
665 /* Note that all of the stack and the thread descriptor is
666 zeroed. This means we do not have to initialize fields
667 with initial value zero. This is specifically true for
668 the 'tid' field which is always set back to zero once the
669 stack is not used anymore and for the 'guardsize' field
670 which will be read next. */
671 }
672
673 /* Create or resize the guard area if necessary. */
674 if (__glibc_unlikely (guardsize > pd->guardsize))
675 {
676 char *guard = guard_position (mem, size, guardsize, pd,
677 pagesize_m1);
678 if (__mprotect (guard, guardsize, PROT_NONE) != 0)
679 {
680 mprot_error:
681 lll_lock (stack_cache_lock, LLL_PRIVATE);
682
683 /* Remove the thread from the list. */
684 stack_list_del (&pd->list);
685
686 lll_unlock (stack_cache_lock, LLL_PRIVATE);
687
688 /* Get rid of the TLS block we allocated. */
689 _dl_deallocate_tls (TLS_TPADJ (pd), false);
690
691 /* Free the stack memory regardless of whether the size
692 of the cache is over the limit or not. If this piece
693 of memory caused problems we better do not use it
694 anymore. Uh, and we ignore possible errors. There
695 is nothing we could do. */
696 (void) __munmap (mem, size);
697
698 return errno;
699 }
700
701 pd->guardsize = guardsize;
702 }
703 else if (__builtin_expect (pd->guardsize - guardsize > size - reqsize,
704 0))
705 {
706 /* The old guard area is too large. */
707
708#ifdef NEED_SEPARATE_REGISTER_STACK
709 char *guard = mem + (((size - guardsize) / 2) & ~pagesize_m1);
710 char *oldguard = mem + (((size - pd->guardsize) / 2) & ~pagesize_m1);
711
712 if (oldguard < guard
713 && __mprotect (oldguard, guard - oldguard, prot) != 0)
714 goto mprot_error;
715
716 if (__mprotect (guard + guardsize,
717 oldguard + pd->guardsize - guard - guardsize,
718 prot) != 0)
719 goto mprot_error;
720#elif _STACK_GROWS_DOWN
721 if (__mprotect ((char *) mem + guardsize, pd->guardsize - guardsize,
722 prot) != 0)
723 goto mprot_error;
724#elif _STACK_GROWS_UP
725 char *new_guard = (char *)(((uintptr_t) pd - guardsize)
726 & ~pagesize_m1);
727 char *old_guard = (char *)(((uintptr_t) pd - pd->guardsize)
728 & ~pagesize_m1);
729 /* The guard size difference might be > 0, but once rounded
730 to the nearest page the size difference might be zero. */
731 if (new_guard > old_guard
732 && __mprotect (old_guard, new_guard - old_guard, prot) != 0)
733 goto mprot_error;
734#endif
735
736 pd->guardsize = guardsize;
737 }
738 /* The pthread_getattr_np() calls need to get passed the size
739 requested in the attribute, regardless of how large the
740 actually used guardsize is. */
741 pd->reported_guardsize = guardsize;
742 }
743
744 /* Initialize the lock. We have to do this unconditionally since the
745 stillborn thread could be canceled while the lock is taken. */
746 pd->lock = LLL_LOCK_INITIALIZER;
747
748 /* The robust mutex lists also need to be initialized
749 unconditionally because the cleanup for the previous stack owner
750 might have happened in the kernel. */
751 pd->robust_head.futex_offset = (offsetof (pthread_mutex_t, __data.__lock)
752 - offsetof (pthread_mutex_t,
753 __data.__list.__next));
754 pd->robust_head.list_op_pending = NULL;
755#if __PTHREAD_MUTEX_HAVE_PREV
756 pd->robust_prev = &pd->robust_head;
757#endif
758 pd->robust_head.list = &pd->robust_head;
759
760 /* We place the thread descriptor at the end of the stack. */
761 *pdp = pd;
762
763#if _STACK_GROWS_DOWN
764 void *stacktop;
765
766# if TLS_TCB_AT_TP
767 /* The stack begins before the TCB and the static TLS block. */
768 stacktop = ((char *) (pd + 1) - __static_tls_size);
769# elif TLS_DTV_AT_TP
770 stacktop = (char *) (pd - 1);
771# endif
772
773# ifdef NEED_SEPARATE_REGISTER_STACK
774 *stack = pd->stackblock;
775 *stacksize = stacktop - *stack;
776# else
777 *stack = stacktop;
778# endif
779#else
780 *stack = pd->stackblock;
781#endif
782
783 return 0;
784}
785
786
787void
788__deallocate_stack (struct pthread *pd)
789{
790 lll_lock (stack_cache_lock, LLL_PRIVATE);
791
792 /* Remove the thread from the list of threads with user defined
793 stacks. */
794 stack_list_del (&pd->list);
795
796 /* Not much to do. Just free the mmap()ed memory. Note that we do
797 not reset the 'used' flag in the 'tid' field. This is done by
798 the kernel. If no thread has been created yet this field is
799 still zero. */
800 if (__glibc_likely (! pd->user_stack))
801 (void) queue_stack (pd);
802 else
803 /* Free the memory associated with the ELF TLS. */
804 _dl_deallocate_tls (TLS_TPADJ (pd), false);
805
806 lll_unlock (stack_cache_lock, LLL_PRIVATE);
807}
808
809
810int
811__make_stacks_executable (void **stack_endp)
812{
813 /* First the main thread's stack. */
814 int err = _dl_make_stack_executable (stack_endp);
815 if (err != 0)
816 return err;
817
818#ifdef NEED_SEPARATE_REGISTER_STACK
819 const size_t pagemask = ~(__getpagesize () - 1);
820#endif
821
822 lll_lock (stack_cache_lock, LLL_PRIVATE);
823
824 list_t *runp;
825 list_for_each (runp, &stack_used)
826 {
827 err = change_stack_perm (list_entry (runp, struct pthread, list)
828#ifdef NEED_SEPARATE_REGISTER_STACK
829 , pagemask
830#endif
831 );
832 if (err != 0)
833 break;
834 }
835
836 /* Also change the permission for the currently unused stacks. This
837 might be wasted time but better spend it here than adding a check
838 in the fast path. */
839 if (err == 0)
840 list_for_each (runp, &stack_cache)
841 {
842 err = change_stack_perm (list_entry (runp, struct pthread, list)
843#ifdef NEED_SEPARATE_REGISTER_STACK
844 , pagemask
845#endif
846 );
847 if (err != 0)
848 break;
849 }
850
851 lll_unlock (stack_cache_lock, LLL_PRIVATE);
852
853 return err;
854}
855
856
857/* In case of a fork() call the memory allocation in the child will be
858 the same but only one thread is running. All stacks except that of
859 the one running thread are not used anymore. We have to recycle
860 them. */
861void
862__reclaim_stacks (void)
863{
864 struct pthread *self = (struct pthread *) THREAD_SELF;
865
866 /* No locking necessary. The caller is the only stack in use. But
867 we have to be aware that we might have interrupted a list
868 operation. */
869
870 if (in_flight_stack != 0)
871 {
872 bool add_p = in_flight_stack & 1;
873 list_t *elem = (list_t *) (in_flight_stack & ~(uintptr_t) 1);
874
875 if (add_p)
876 {
877 /* We always add at the beginning of the list. So in this case we
878 only need to check the beginning of these lists to see if the
879 pointers at the head of the list are inconsistent. */
880 list_t *l = NULL;
881
882 if (stack_used.next->prev != &stack_used)
883 l = &stack_used;
884 else if (stack_cache.next->prev != &stack_cache)
885 l = &stack_cache;
886
887 if (l != NULL)
888 {
889 assert (l->next->prev == elem);
890 elem->next = l->next;
891 elem->prev = l;
892 l->next = elem;
893 }
894 }
895 else
896 {
897 /* We can simply always replay the delete operation. */
898 elem->next->prev = elem->prev;
899 elem->prev->next = elem->next;
900 }
901 }
902
903 /* Mark all stacks except the still running one as free. */
904 list_t *runp;
905 list_for_each (runp, &stack_used)
906 {
907 struct pthread *curp = list_entry (runp, struct pthread, list);
908 if (curp != self)
909 {
910 /* This marks the stack as free. */
911 curp->tid = 0;
912
913 /* Account for the size of the stack. */
914 stack_cache_actsize += curp->stackblock_size;
915
916 if (curp->specific_used)
917 {
918 /* Clear the thread-specific data. */
919 memset (curp->specific_1stblock, '\0',
920 sizeof (curp->specific_1stblock));
921
922 curp->specific_used = false;
923
924 for (size_t cnt = 1; cnt < PTHREAD_KEY_1STLEVEL_SIZE; ++cnt)
925 if (curp->specific[cnt] != NULL)
926 {
927 memset (curp->specific[cnt], '\0',
928 sizeof (curp->specific_1stblock));
929
930 /* We have allocated the block which we do not
931 free here so re-set the bit. */
932 curp->specific_used = true;
933 }
934 }
935 }
936 }
937
938 /* Add the stack of all running threads to the cache. */
939 list_splice (&stack_used, &stack_cache);
940
941 /* Remove the entry for the current thread to from the cache list
942 and add it to the list of running threads. Which of the two
943 lists is decided by the user_stack flag. */
944 stack_list_del (&self->list);
945
946 /* Re-initialize the lists for all the threads. */
947 INIT_LIST_HEAD (&stack_used);
948 INIT_LIST_HEAD (&__stack_user);
949
950 if (__glibc_unlikely (THREAD_GETMEM (self, user_stack)))
951 list_add (&self->list, &__stack_user);
952 else
953 list_add (&self->list, &stack_used);
954
955 /* There is one thread running. */
956 __nptl_nthreads = 1;
957
958 in_flight_stack = 0;
959
960 /* Initialize locks. */
961 stack_cache_lock = LLL_LOCK_INITIALIZER;
962 __default_pthread_attr_lock = LLL_LOCK_INITIALIZER;
963}
964
965
966static void
967setxid_mark_thread (struct xid_command *cmdp, struct pthread *t)
968{
969 int ch;
970
971 /* Wait until this thread is cloned. */
972 if (t->setxid_futex == -1
973 && ! atomic_compare_and_exchange_bool_acq (&t->setxid_futex, -2, -1))
974 do
975 futex_wait_simple (&t->setxid_futex, -2, FUTEX_PRIVATE);
976 while (t->setxid_futex == -2);
977
978 /* Don't let the thread exit before the setxid handler runs. */
979 t->setxid_futex = 0;
980
981 do
982 {
983 ch = t->cancelhandling;
984
985 /* If the thread is exiting right now, ignore it. */
986 if ((ch & EXITING_BITMASK) != 0)
987 {
988 /* Release the futex if there is no other setxid in
989 progress. */
990 if ((ch & SETXID_BITMASK) == 0)
991 {
992 t->setxid_futex = 1;
993 futex_wake (&t->setxid_futex, 1, FUTEX_PRIVATE);
994 }
995 return;
996 }
997 }
998 while (atomic_compare_and_exchange_bool_acq (&t->cancelhandling,
999 ch | SETXID_BITMASK, ch));
1000}
1001
1002
1003static void
1004setxid_unmark_thread (struct xid_command *cmdp, struct pthread *t)
1005{
1006 int ch;
1007
1008 do
1009 {
1010 ch = t->cancelhandling;
1011 if ((ch & SETXID_BITMASK) == 0)
1012 return;
1013 }
1014 while (atomic_compare_and_exchange_bool_acq (&t->cancelhandling,
1015 ch & ~SETXID_BITMASK, ch));
1016
1017 /* Release the futex just in case. */
1018 t->setxid_futex = 1;
1019 futex_wake (&t->setxid_futex, 1, FUTEX_PRIVATE);
1020}
1021
1022
1023static int
1024setxid_signal_thread (struct xid_command *cmdp, struct pthread *t)
1025{
1026 if ((t->cancelhandling & SETXID_BITMASK) == 0)
1027 return 0;
1028
1029 int val;
1030 pid_t pid = __getpid ();
1031 INTERNAL_SYSCALL_DECL (err);
1032 val = INTERNAL_SYSCALL_CALL (tgkill, err, pid, t->tid, SIGSETXID);
1033
1034 /* If this failed, it must have had not started yet or else exited. */
1035 if (!INTERNAL_SYSCALL_ERROR_P (val, err))
1036 {
1037 atomic_increment (&cmdp->cntr);
1038 return 1;
1039 }
1040 else
1041 return 0;
1042}
1043
1044/* Check for consistency across set*id system call results. The abort
1045 should not happen as long as all privileges changes happen through
1046 the glibc wrappers. ERROR must be 0 (no error) or an errno
1047 code. */
1048void
1049attribute_hidden
1050__nptl_setxid_error (struct xid_command *cmdp, int error)
1051{
1052 do
1053 {
1054 int olderror = cmdp->error;
1055 if (olderror == error)
1056 break;
1057 if (olderror != -1)
1058 {
1059 /* Mismatch between current and previous results. Save the
1060 error value to memory so that is not clobbered by the
1061 abort function and preserved in coredumps. */
1062 volatile int xid_err __attribute__((unused)) = error;
1063 abort ();
1064 }
1065 }
1066 while (atomic_compare_and_exchange_bool_acq (&cmdp->error, error, -1));
1067}
1068
1069int
1070attribute_hidden
1071__nptl_setxid (struct xid_command *cmdp)
1072{
1073 int signalled;
1074 int result;
1075 lll_lock (stack_cache_lock, LLL_PRIVATE);
1076
1077 __xidcmd = cmdp;
1078 cmdp->cntr = 0;
1079 cmdp->error = -1;
1080
1081 struct pthread *self = THREAD_SELF;
1082
1083 /* Iterate over the list with system-allocated threads first. */
1084 list_t *runp;
1085 list_for_each (runp, &stack_used)
1086 {
1087 struct pthread *t = list_entry (runp, struct pthread, list);
1088 if (t == self)
1089 continue;
1090
1091 setxid_mark_thread (cmdp, t);
1092 }
1093
1094 /* Now the list with threads using user-allocated stacks. */
1095 list_for_each (runp, &__stack_user)
1096 {
1097 struct pthread *t = list_entry (runp, struct pthread, list);
1098 if (t == self)
1099 continue;
1100
1101 setxid_mark_thread (cmdp, t);
1102 }
1103
1104 /* Iterate until we don't succeed in signalling anyone. That means
1105 we have gotten all running threads, and their children will be
1106 automatically correct once started. */
1107 do
1108 {
1109 signalled = 0;
1110
1111 list_for_each (runp, &stack_used)
1112 {
1113 struct pthread *t = list_entry (runp, struct pthread, list);
1114 if (t == self)
1115 continue;
1116
1117 signalled += setxid_signal_thread (cmdp, t);
1118 }
1119
1120 list_for_each (runp, &__stack_user)
1121 {
1122 struct pthread *t = list_entry (runp, struct pthread, list);
1123 if (t == self)
1124 continue;
1125
1126 signalled += setxid_signal_thread (cmdp, t);
1127 }
1128
1129 int cur = cmdp->cntr;
1130 while (cur != 0)
1131 {
1132 futex_wait_simple ((unsigned int *) &cmdp->cntr, cur,
1133 FUTEX_PRIVATE);
1134 cur = cmdp->cntr;
1135 }
1136 }
1137 while (signalled != 0);
1138
1139 /* Clean up flags, so that no thread blocks during exit waiting
1140 for a signal which will never come. */
1141 list_for_each (runp, &stack_used)
1142 {
1143 struct pthread *t = list_entry (runp, struct pthread, list);
1144 if (t == self)
1145 continue;
1146
1147 setxid_unmark_thread (cmdp, t);
1148 }
1149
1150 list_for_each (runp, &__stack_user)
1151 {
1152 struct pthread *t = list_entry (runp, struct pthread, list);
1153 if (t == self)
1154 continue;
1155
1156 setxid_unmark_thread (cmdp, t);
1157 }
1158
1159 /* This must be last, otherwise the current thread might not have
1160 permissions to send SIGSETXID syscall to the other threads. */
1161 INTERNAL_SYSCALL_DECL (err);
1162 result = INTERNAL_SYSCALL_NCS (cmdp->syscall_no, err, 3,
1163 cmdp->id[0], cmdp->id[1], cmdp->id[2]);
1164 int error = 0;
1165 if (__glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (result, err)))
1166 {
1167 error = INTERNAL_SYSCALL_ERRNO (result, err);
1168 __set_errno (error);
1169 result = -1;
1170 }
1171 __nptl_setxid_error (cmdp, error);
1172
1173 lll_unlock (stack_cache_lock, LLL_PRIVATE);
1174 return result;
1175}
1176
1177static inline void __attribute__((always_inline))
1178init_one_static_tls (struct pthread *curp, struct link_map *map)
1179{
1180# if TLS_TCB_AT_TP
1181 void *dest = (char *) curp - map->l_tls_offset;
1182# elif TLS_DTV_AT_TP
1183 void *dest = (char *) curp + map->l_tls_offset + TLS_PRE_TCB_SIZE;
1184# else
1185# error "Either TLS_TCB_AT_TP or TLS_DTV_AT_TP must be defined"
1186# endif
1187
1188 /* Initialize the memory. */
1189 memset (__mempcpy (dest, map->l_tls_initimage, map->l_tls_initimage_size),
1190 '\0', map->l_tls_blocksize - map->l_tls_initimage_size);
1191}
1192
1193void
1194attribute_hidden
1195__pthread_init_static_tls (struct link_map *map)
1196{
1197 lll_lock (stack_cache_lock, LLL_PRIVATE);
1198
1199 /* Iterate over the list with system-allocated threads first. */
1200 list_t *runp;
1201 list_for_each (runp, &stack_used)
1202 init_one_static_tls (list_entry (runp, struct pthread, list), map);
1203
1204 /* Now the list with threads using user-allocated stacks. */
1205 list_for_each (runp, &__stack_user)
1206 init_one_static_tls (list_entry (runp, struct pthread, list), map);
1207
1208 lll_unlock (stack_cache_lock, LLL_PRIVATE);
1209}
1210
1211
1212void
1213attribute_hidden
1214__wait_lookup_done (void)
1215{
1216 lll_lock (stack_cache_lock, LLL_PRIVATE);
1217
1218 struct pthread *self = THREAD_SELF;
1219
1220 /* Iterate over the list with system-allocated threads first. */
1221 list_t *runp;
1222 list_for_each (runp, &stack_used)
1223 {
1224 struct pthread *t = list_entry (runp, struct pthread, list);
1225 if (t == self || t->header.gscope_flag == THREAD_GSCOPE_FLAG_UNUSED)
1226 continue;
1227
1228 int *const gscope_flagp = &t->header.gscope_flag;
1229
1230 /* We have to wait until this thread is done with the global
1231 scope. First tell the thread that we are waiting and
1232 possibly have to be woken. */
1233 if (atomic_compare_and_exchange_bool_acq (gscope_flagp,
1234 THREAD_GSCOPE_FLAG_WAIT,
1235 THREAD_GSCOPE_FLAG_USED))
1236 continue;
1237
1238 do
1239 futex_wait_simple ((unsigned int *) gscope_flagp,
1240 THREAD_GSCOPE_FLAG_WAIT, FUTEX_PRIVATE);
1241 while (*gscope_flagp == THREAD_GSCOPE_FLAG_WAIT);
1242 }
1243
1244 /* Now the list with threads using user-allocated stacks. */
1245 list_for_each (runp, &__stack_user)
1246 {
1247 struct pthread *t = list_entry (runp, struct pthread, list);
1248 if (t == self || t->header.gscope_flag == THREAD_GSCOPE_FLAG_UNUSED)
1249 continue;
1250
1251 int *const gscope_flagp = &t->header.gscope_flag;
1252
1253 /* We have to wait until this thread is done with the global
1254 scope. First tell the thread that we are waiting and
1255 possibly have to be woken. */
1256 if (atomic_compare_and_exchange_bool_acq (gscope_flagp,
1257 THREAD_GSCOPE_FLAG_WAIT,
1258 THREAD_GSCOPE_FLAG_USED))
1259 continue;
1260
1261 do
1262 futex_wait_simple ((unsigned int *) gscope_flagp,
1263 THREAD_GSCOPE_FLAG_WAIT, FUTEX_PRIVATE);
1264 while (*gscope_flagp == THREAD_GSCOPE_FLAG_WAIT);
1265 }
1266
1267 lll_unlock (stack_cache_lock, LLL_PRIVATE);
1268}
1269