/*
* Copyright (c) 1994 by Xerox Corporation. All rights reserved.
* Copyright (c) 1996 by Silicon Graphics. All rights reserved.
* Copyright (c) 1998 by Fergus Henderson. All rights reserved.
* Copyright (c) 2000-2008 by Hewlett-Packard Development Company.
* All rights reserved.
* Copyright (c) 2008-2021 Ivan Maidanski
*
* THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
* OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
*
* Permission is hereby granted to use or copy this program
* for any purpose, provided the above notices are retained on all copies.
* Permission to modify the code and to distribute modified code is granted,
* provided the above notices are retained, and a notice that the code was
* modified is included with the above copyright notice.
*/
#include "private/gc_priv.h"
#if defined(GC_WIN32_THREADS)
#ifdef THREAD_LOCAL_ALLOC
# include "private/thread_local_alloc.h"
#endif /* THREAD_LOCAL_ALLOC */
/* Allocation lock declarations. */
#if !defined(USE_PTHREAD_LOCKS)
GC_INNER CRITICAL_SECTION GC_allocate_ml;
# ifdef GC_ASSERTIONS
GC_INNER DWORD GC_lock_holder = NO_THREAD;
/* Thread id for current holder of allocation lock */
# endif
#else
GC_INNER pthread_mutex_t GC_allocate_ml = PTHREAD_MUTEX_INITIALIZER;
# ifdef GC_ASSERTIONS
GC_INNER unsigned long GC_lock_holder = NO_THREAD;
# endif
#endif
# include <pthread.h>
# ifdef CAN_CALL_ATFORK
# include <unistd.h>
# endif
# ifdef EMULATE_PTHREAD_SEMAPHORE
# include "private/darwin_semaphore.h"
# else
# include <semaphore.h>
# endif
#elif !defined(MSWINCE)
# include <process.h> /* For _beginthreadex, _endthreadex */
# include <errno.h> /* for errno, EAGAIN */
#endif /* !GC_PTHREADS && !MSWINCE */
/* PUSHED_REGS_COUNT is the number of copied registers in copy_ptr_regs. */
static ptr_t copy_ptr_regs(word *regs, const CONTEXT *pcontext);
#if defined(I386)
# ifdef WOW64_THREAD_CONTEXT_WORKAROUND
# define PUSHED_REGS_COUNT 9
# else
# define PUSHED_REGS_COUNT 7
# endif
#elif defined(X86_64)
# ifdef XMM_CANT_STORE_PTRS
/* If pointers can't be located in Xmm registers. */
# define PUSHED_REGS_COUNT 15
# else
/* gcc-13 may store pointers into SIMD registers when */
/* certain compiler optimizations are enabled. */
# define PUSHED_REGS_COUNT (15+32)
# endif
#elif defined(SHx)
# define PUSHED_REGS_COUNT 15
#elif defined(ARM32)
# define PUSHED_REGS_COUNT 13
#elif defined(AARCH64)
# define PUSHED_REGS_COUNT 30
#elif defined(MIPS) || defined(ALPHA)
# define PUSHED_REGS_COUNT 28
#elif defined(PPC)
# define PUSHED_REGS_COUNT 29
#endif
/* DllMain-based thread registration is currently incompatible */
/* with thread-local allocation, pthreads and WinCE. */
#if (defined(GC_DLL) || defined(GC_INSIDE_DLL)) && !defined(NO_CRT) \
&& !defined(GC_NO_THREADS_DISCOVERY) && !defined(MSWINCE) \
&& !defined(THREAD_LOCAL_ALLOC) && !defined(GC_PTHREADS)
/* This code operates in two distinct modes, depending on */
/* the setting of GC_win32_dll_threads. */
/* If GC_win32_dll_threads is set, all threads in the process */
/* are implicitly registered with the GC by DllMain. */
/* No explicit registration is required, and attempts at */
/* explicit registration are ignored. This mode is */
/* very different from the Posix operation of the collector. */
/* In this mode access to the thread table is lock-free. */
/* Hence there is a static limit on the number of threads. */
# ifdef GC_DISCOVER_TASK_THREADS
/* GC_DISCOVER_TASK_THREADS should be used if DllMain-based */
/* thread registration is required but it is impossible to */
/* call GC_use_threads_discovery before other GC routines. */
# define GC_win32_dll_threads TRUE
# else
STATIC GC_bool GC_win32_dll_threads = FALSE;
/* GC_win32_dll_threads must be set (if needed) at the */
/* application initialization time, i.e. before any */
/* collector or thread calls. We make it a "dynamic" */
/* option only to avoid multiple library versions. */
# endif
#else
/* If GC_win32_dll_threads is FALSE (or the collector is */
/* built without GC_DLL defined), things operate in a way */
/* that is very similar to Posix platforms, and new threads */
/* must be registered with the collector, e.g. by using */
/* preprocessor-based interception of the thread primitives. */
/* In this case, we use a real data structure for the thread */
/* table. Note that there is no equivalent of linker-based */
/* call interception, since we don't have ELF-like */
/* facilities. The Windows analog appears to be "API */
/* hooking", which really seems to be a standard way to */
/* do minor binary rewriting (?). I'd prefer not to have */
/* the basic collector rely on such facilities, but an */
/* optional package that intercepts thread calls this way */
/* would probably be nice. */
# ifndef GC_NO_THREADS_DISCOVERY
# define GC_NO_THREADS_DISCOVERY
# endif
# define GC_win32_dll_threads FALSE
# undef MAX_THREADS
# define MAX_THREADS 1 /* dll_thread_table[] is always empty. */
#endif /* GC_NO_THREADS_DISCOVERY */
/* We have two versions of the thread table. Which one */
/* we use depends on whether GC_win32_dll_threads */
/* is set. Note that before initialization, we don't */
/* add any entries to either table, even if DllMain is */
/* called. The main thread will be added on */
/* initialization. */
/* The type of the first argument to InterlockedExchange. */
/* Documented to be LONG volatile *, but at least gcc likes */
/* this better. */
typedef LONG * IE_t;
/* GC_use_threads_discovery() is currently incompatible with pthreads */
/* and WinCE. It might be possible to get DllMain-based thread */
/* registration to work with Cygwin, but if you try it then you are on */
/* your own. */
GC_API void GC_CALL GC_use_threads_discovery(void)
{
# ifdef GC_NO_THREADS_DISCOVERY
ABORT("GC DllMain-based thread registration unsupported");
# else
/* Turn on GC_win32_dll_threads. */
GC_ASSERT(!parallel_initialized);
/* Note that GC_use_threads_discovery is expected to be called by */
/* the client application (not from DllMain) at start-up. */
# ifndef GC_DISCOVER_TASK_THREADS
GC_win32_dll_threads = TRUE;
# endif
GC_init_parallel();
# endif
}
#define ADDR_LIMIT ((ptr_t)GC_WORD_MAX)
struct GC_Thread_Rep {
union {
# ifndef GC_NO_THREADS_DISCOVERY
volatile AO_t in_use;
/* Updated without lock. */
/* We assert that unused */
/* entries have invalid ids of */
/* zero and zero stack fields. */
/* Used only with GC_win32_dll_threads. */
LONG long_in_use; /* The same but of the type that */
/* matches the first argument of */
/* InterlockedExchange(); volatile is */
/* omitted because the ancient version */
/* of the prototype lacks the qualifier.*/
# endif
struct GC_Thread_Rep * next;
/* Hash table link without */
/* GC_win32_dll_threads. */
/* More recently allocated threads */
/* with a given pthread id come */
/* first. (All but the first are */
/* guaranteed to be dead, but we may */
/* not yet have registered the join.) */
} tm; /* table_management */
DWORD id;
# ifdef MSWINCE
/* According to MSDN specs for WinCE targets: */
/* - DuplicateHandle() is not applicable to thread handles; and */
/* - the value returned by GetCurrentThreadId() could be used as */
/* a "real" thread handle (for SuspendThread(), ResumeThread() and */
/* GetThreadContext()). */
# define THREAD_HANDLE(t) (HANDLE)(word)(t)->id
# else
HANDLE handle;
# define THREAD_HANDLE(t) (t)->handle
# endif
ptr_t stack_base; /* The cold end of the stack. */
/* 0 ==> entry not valid. */
/* !in_use ==> stack_base == 0 */
ptr_t last_stack_min; /* Last known minimum (hottest) address */
/* in stack or ADDR_LIMIT if unset */
# ifdef IA64
ptr_t backing_store_end;
ptr_t backing_store_ptr;
# elif defined(I386)
ptr_t initial_stack_base;
/* The cold end of the stack saved by */
/* GC_record_stack_base (never modified */
/* by GC_set_stackbottom). */
# endif
ptr_t thread_blocked_sp; /* Protected by GC lock. */
/* NULL value means thread unblocked. */
/* If set to non-NULL, thread will */
/* acquire GC lock before doing any */
/* pointer manipulations. Thus it does */
/* not need to stop this thread. */
struct GC_traced_stack_sect_s *traced_stack_sect;
/* Points to the "stack section" data */
/* held in stack by the innermost */
/* GC_call_with_gc_active() of this */
/* thread. May be NULL. */
# ifndef GC_NO_FINALIZATION
unsigned short finalizer_skipped;
unsigned char finalizer_nested;
/* Used by GC_check_finalizer_nested() */
/* to minimize the level of recursion */
/* when a client finalizer allocates */
/* memory (initially both are 0). */
# endif
unsigned char suspended; /* really of GC_bool type */
# ifdef GC_PTHREADS
unsigned char flags; /* Protected by GC lock. */
# define FINISHED 1 /* Thread has exited. */
# define DETACHED 2 /* Thread is intended to be detached. */
# define KNOWN_FINISHED(t) (((t) -> flags) & FINISHED)
pthread_t pthread_id;
void *status; /* hold exit value until join in case it's a pointer */
# else
# define KNOWN_FINISHED(t) 0
# endif
# if defined(WOW64_THREAD_CONTEXT_WORKAROUND) && defined(MSWINRT_FLAVOR)
PNT_TIB tib;
# endif
# ifdef RETRY_GET_THREAD_CONTEXT
ptr_t context_sp;
word context_regs[PUSHED_REGS_COUNT];
/* Populated as part of GC_suspend() as */
/* resume/suspend loop may be needed for the */
/* call to GetThreadContext() to succeed. */
# endif
};
#ifndef GC_NO_THREADS_DISCOVERY
/* We track thread attachments while the world is supposed to be */
/* stopped. Unfortunately, we cannot stop them from starting, since */
/* blocking in DllMain seems to cause the world to deadlock. Thus, */
/* we have to recover if we notice this in the middle of marking. */
STATIC volatile AO_t GC_attached_thread = FALSE;
/* We assumed that volatile ==> memory ordering, at least among */
/* volatiles. This code should consistently use atomic_ops. */
STATIC volatile GC_bool GC_please_stop = FALSE;
#elif defined(GC_ASSERTIONS)
STATIC GC_bool GC_please_stop = FALSE;
#endif /* GC_NO_THREADS_DISCOVERY && GC_ASSERTIONS */
#if defined(WRAP_MARK_SOME) && !defined(GC_PTHREADS)
/* Return TRUE if an thread was attached since we last asked or */
/* since GC_attached_thread was explicitly reset. */
GC_INNER GC_bool GC_started_thread_while_stopped(void)
{
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
# ifdef AO_HAVE_compare_and_swap_release
if (AO_compare_and_swap_release(&GC_attached_thread, TRUE,
FALSE /* stored */))
return TRUE;
# else
AO_nop_full(); /* Prior heap reads need to complete earlier. */
if (AO_load(&GC_attached_thread)) {
AO_store(&GC_attached_thread, FALSE);
return TRUE;
}
# endif
}
# endif
return FALSE;
}
#endif /* WRAP_MARK_SOME */
/* Thread table used if GC_win32_dll_threads is set. */
/* This is a fixed size array. */
/* Since we use runtime conditionals, both versions */
/* are always defined. */
# ifndef MAX_THREADS
# define MAX_THREADS 512
# endif
/* Things may get quite slow for large numbers of threads, */
/* since we look them up with sequential search. */
volatile struct GC_Thread_Rep dll_thread_table[MAX_THREADS];
STATIC volatile LONG GC_max_thread_index = 0;
/* Largest index in dll_thread_table */
/* that was ever used. */
/* And now the version used if GC_win32_dll_threads is not set. */
/* This is a chained hash table, with much of the code borrowed */
/* from the Posix implementation. */
#ifndef THREAD_TABLE_SZ
# define THREAD_TABLE_SZ 256 /* Power of 2 (for speed). */
#endif
#define THREAD_TABLE_INDEX(id) /* id is of DWORD type */ \
(int)((((id) >> 8) ^ (id)) % THREAD_TABLE_SZ)
STATIC GC_thread GC_threads[THREAD_TABLE_SZ];
/* It may not be safe to allocate when we register the first thread. */
/* Thus we allocated one statically. It does not contain any pointer */
/* field we need to push ("next" and "status" fields are unused). */
static struct GC_Thread_Rep first_thread;
static GC_bool first_thread_used = FALSE;
/* Add a thread to GC_threads. We assume it wasn't already there. */
/* Caller holds allocation lock. */
/* Unlike the pthreads version, the id field is set by the caller. */
STATIC GC_thread GC_new_thread(DWORD id)
{
int hv = THREAD_TABLE_INDEX(id);
GC_thread result;
# ifdef DEBUG_THREADS
GC_log_printf("Creating thread 0x%lx\n", (long)id);
if (GC_threads[hv] != NULL)
GC_log_printf("Hash collision at GC_threads[%d]\n", hv);
# endif
GC_ASSERT(I_HOLD_LOCK());
if (!EXPECT(first_thread_used, TRUE)) {
result = &first_thread;
first_thread_used = TRUE;
GC_ASSERT(NULL == GC_threads[hv]);
} else {
GC_ASSERT(!GC_win32_dll_threads);
result = (struct GC_Thread_Rep *)
GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
if (result == 0) return(0);
}
/* result -> id = id; Done by caller. */
result -> tm.next = GC_threads[hv];
GC_threads[hv] = result;
# ifdef GC_PTHREADS
GC_ASSERT(result -> flags == 0);
# endif
GC_ASSERT(result -> thread_blocked_sp == NULL);
if (EXPECT(result != &first_thread, TRUE))
GC_dirty(result);
return(result);
}
GC_INLINE void GC_record_stack_base(GC_vthread me,
const struct GC_stack_base *sb)
{
me -> stack_base = (ptr_t)sb->mem_base;
# ifdef IA64
me -> backing_store_end = (ptr_t)sb->reg_base;
# elif defined(I386)
me -> initial_stack_base = (ptr_t)sb->mem_base;
# endif
if (me -> stack_base == NULL)
ABORT("Bad stack base in GC_register_my_thread");
}
/* This may be called from DllMain, and hence operates under unusual */
/* constraints. In particular, it must be lock-free if */
/* GC_win32_dll_threads is set. Always called from the thread being */
/* added. If GC_win32_dll_threads is not set, we already hold the */
/* allocation lock except possibly during single-threaded startup code. */
/* Does not initialize thread local free lists. */
STATIC GC_thread GC_register_my_thread_inner(const struct GC_stack_base *sb,
DWORD thread_id)
{
GC_vthread me;
/* The following should be a no-op according to the Win32 */
/* documentation. There is empirical evidence that it */
/* isn't. - HB */
# if defined(MPROTECT_VDB) && !defined(CYGWIN32)
if (GC_auto_incremental
# ifdef GWW_VDB
&& !GC_gww_dirty_init()
# endif
)
GC_set_write_fault_handler();
# endif
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
int i;
/* It appears to be unsafe to acquire a lock here, since this */
/* code is apparently not preemptible on some systems. */
/* (This is based on complaints, not on Microsoft's official */
/* documentation, which says this should perform "only simple */
/* initialization tasks".) */
/* Hence we make do with nonblocking synchronization. */
/* It has been claimed that DllMain is really only executed with */
/* a particular system lock held, and thus careful use of locking */
/* around code that doesn't call back into the system libraries */
/* might be OK. But this has not been tested across all Win32 */
/* variants. */
for (i = 0;
InterlockedExchange(&dll_thread_table[i].tm.long_in_use, 1) != 0;
i++) {
/* Compare-and-swap would make this cleaner, but that's not */
/* supported before Windows 98 and NT 4.0. In Windows 2000, */
/* InterlockedExchange is supposed to be replaced by */
/* InterlockedExchangePointer, but that's not really what I */
/* want here. */
/* FIXME: We should eventually declare Windows 95 dead and use */
/* AO_ primitives here. */
if (i == MAX_THREADS - 1)
ABORT("Too many threads");
}
/* Update GC_max_thread_index if necessary. The following is */
/* safe, and unlike CompareExchange-based solutions seems to work */
/* on all Windows95 and later platforms. */
/* Unfortunately, GC_max_thread_index may be temporarily out of */
/* bounds, so readers have to compensate. */
while (i > GC_max_thread_index) {
InterlockedIncrement((IE_t)&GC_max_thread_index);
}
if (GC_max_thread_index >= MAX_THREADS) {
/* We overshot due to simultaneous increments. */
/* Setting it to MAX_THREADS-1 is always safe. */
GC_max_thread_index = MAX_THREADS - 1;
}
me = dll_thread_table + i;
} else
# endif
/* else */ /* Not using DllMain */ {
GC_ASSERT(I_HOLD_LOCK());
GC_in_thread_creation = TRUE; /* OK to collect from unknown thread. */
me = GC_new_thread(thread_id);
GC_in_thread_creation = FALSE;
if (me == 0)
ABORT("Failed to allocate memory for thread registering");
}
# ifdef GC_PTHREADS
/* me can be NULL -> segfault */
me -> pthread_id = pthread_self();
# endif
# ifndef MSWINCE
/* GetCurrentThread() returns a pseudohandle (a const value). */
if (!DuplicateHandle(GetCurrentProcess(), GetCurrentThread(),
GetCurrentProcess(),
(HANDLE*)&(me -> handle),
0 /* dwDesiredAccess */, FALSE /* bInheritHandle */,
DUPLICATE_SAME_ACCESS)) {
ABORT_ARG1("DuplicateHandle failed",
": errcode= 0x%X", (unsigned)GetLastError());
}
# endif
# if defined(WOW64_THREAD_CONTEXT_WORKAROUND) && defined(MSWINRT_FLAVOR)
/* Lookup TIB value via a call to NtCurrentTeb() on thread */
/* registration rather than calling GetThreadSelectorEntry() which */
/* is not available on UWP. */
me -> tib = (PNT_TIB)NtCurrentTeb();
# endif
me -> last_stack_min = ADDR_LIMIT;
GC_record_stack_base(me, sb);
/* Up until this point, GC_push_all_stacks considers this thread */
/* invalid. */
/* Up until this point, this entry is viewed as reserved but invalid */
/* by GC_delete_thread. */
me -> id = thread_id;
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
if (GC_please_stop) {
AO_store(&GC_attached_thread, TRUE);
AO_nop_full(); /* Later updates must become visible after this. */
}
/* We'd like to wait here, but can't, since waiting in DllMain */
/* provokes deadlocks. */
/* Thus we force marking to be restarted instead. */
} else
# endif
/* else */ {
GC_ASSERT(!GC_please_stop);
/* Otherwise both we and the thread stopping code would be */
/* holding the allocation lock. */
}
return (GC_thread)(me);
}
/*
* GC_max_thread_index may temporarily be larger than MAX_THREADS.
* To avoid subscript errors, we check on access.
*/
GC_INLINE LONG GC_get_max_thread_index(void)
{
LONG my_max = GC_max_thread_index;
if (my_max >= MAX_THREADS) return MAX_THREADS - 1;
return my_max;
}
/* Return the GC_thread corresponding to a thread id. May be called */
/* without a lock, but should be called in contexts in which the */
/* requested thread cannot be asynchronously deleted, e.g. from the */
/* thread itself. */
/* This version assumes that either GC_win32_dll_threads is set, or */
/* we hold the allocator lock. */
/* Also used (for assertion checking only) from thread_local_alloc.c. */
STATIC GC_thread GC_lookup_thread_inner(DWORD thread_id)
{
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
int i;
LONG my_max = GC_get_max_thread_index();
for (i = 0; i <= my_max &&
(!AO_load_acquire(&dll_thread_table[i].tm.in_use)
|| dll_thread_table[i].id != thread_id);
/* Must still be in_use, since nobody else can store our */
/* thread_id. */
i++) {
/* empty */
}
return i <= my_max ? (GC_thread)(dll_thread_table + i) : NULL;
} else
# endif
/* else */ {
GC_thread p = GC_threads[THREAD_TABLE_INDEX(thread_id)];
GC_ASSERT(I_HOLD_LOCK());
while (p != 0 && p -> id != thread_id) p = p -> tm.next;
return(p);
}
}
#ifndef GC_NO_FINALIZATION
/* Called by GC_finalize() (in case of an allocation failure observed). */
/* GC_reset_finalizer_nested() is the same as in pthread_support.c. */
GC_INNER void GC_reset_finalizer_nested(void)
{
GC_thread me = GC_lookup_thread_inner(GetCurrentThreadId());
/* GC_check_finalizer_nested() is the same as in pthread_support.c. */
GC_INNER unsigned char *GC_check_finalizer_nested(void)
{
GC_thread me = GC_lookup_thread_inner(GetCurrentThreadId());
unsigned nesting_level;
CHECK_LOOKUP_MY_THREAD(me);
nesting_level = me->finalizer_nested;
if (nesting_level) {
/* We are inside another GC_invoke_finalizers(). */
/* Skip some implicitly-called GC_invoke_finalizers() */
/* depending on the nesting (recursion) level. */
if (++me->finalizer_skipped < (1U << nesting_level)) return NULL;
me->finalizer_skipped = 0;
}
me->finalizer_nested = (unsigned char)(nesting_level + 1);
return &me->finalizer_nested;
}
#endif /* !GC_NO_FINALIZATION */
#if defined(GC_ASSERTIONS) && defined(THREAD_LOCAL_ALLOC)
/* This is called from thread-local GC_malloc(). */
GC_bool GC_is_thread_tsd_valid(void *tsd)
{
GC_thread me;
DCL_LOCK_STATE;
/* Make sure thread descriptor t is not protected by the VDB */
/* implementation. */
/* Used to prevent write faults when the world is (partially) stopped, */
/* since it may have been stopped with a system lock held, and that */
/* lock may be required for fault handling. */
#if defined(MPROTECT_VDB)
# define UNPROTECT_THREAD(t) \
if (!GC_win32_dll_threads && GC_auto_incremental \
&& t != &first_thread) { \
GC_ASSERT(SMALL_OBJ(GC_size(t))); \
GC_remove_protection(HBLKPTR(t), 1, FALSE); \
} else (void)0
#else
# define UNPROTECT_THREAD(t) (void)0
#endif
#ifdef CYGWIN32
# define GC_PTHREAD_PTRVAL(pthread_id) pthread_id
#elif defined(GC_WIN32_PTHREADS) || defined(GC_PTHREADS_PARAMARK)
# include <pthread.h> /* to check for winpthreads */
# if defined(__WINPTHREADS_VERSION_MAJOR)
# define GC_PTHREAD_PTRVAL(pthread_id) pthread_id
# else
# define GC_PTHREAD_PTRVAL(pthread_id) pthread_id.p
# endif
#endif
/* If a thread has been joined, but we have not yet */
/* been notified, then there may be more than one thread */
/* in the table with the same Win32 thread id. */
/* This is OK, but we need a way to delete a specific one. */
/* Assumes we hold the allocation lock unless */
/* GC_win32_dll_threads is set. Does not actually free */
/* GC_thread entry (only unlinks it). */
/* If GC_win32_dll_threads is set it should be called from the */
/* thread being deleted. */
STATIC void GC_delete_gc_thread_no_free(GC_vthread t)
{
# ifndef MSWINCE
CloseHandle(t->handle);
# endif
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
/* This is intended to be lock-free. */
/* It is either called synchronously from the thread being */
/* deleted, or by the joining thread. */
/* In this branch asynchronous changes to (*t) are possible. */
/* It's not allowed to call GC_printf (and the friends) here, */
/* see GC_stop_world() for the information. */
t -> stack_base = 0;
t -> id = 0;
t -> suspended = FALSE;
# ifdef RETRY_GET_THREAD_CONTEXT
t -> context_sp = NULL;
# endif
AO_store_release(&t->tm.in_use, FALSE);
} else
# endif
/* else */ {
DWORD id = ((GC_thread)t) -> id;
/* Cast away volatile qualifier, since we have lock. */
int hv = THREAD_TABLE_INDEX(id);
GC_thread p = GC_threads[hv];
GC_thread prev = NULL;
GC_ASSERT(I_HOLD_LOCK());
while (p != (GC_thread)t) {
prev = p;
p = p -> tm.next;
}
if (prev == 0) {
GC_threads[hv] = p -> tm.next;
} else {
GC_ASSERT(prev != &first_thread);
prev -> tm.next = p -> tm.next;
GC_dirty(prev);
}
}
}
/* Delete a thread from GC_threads. We assume it is there. */
/* (The code intentionally traps if it wasn't.) Assumes we */
/* hold the allocation lock unless GC_win32_dll_threads is set. */
/* If GC_win32_dll_threads is set then it should be called from */
/* the thread being deleted. It is also safe to delete the */
/* main thread (unless GC_win32_dll_threads). */
STATIC void GC_delete_thread(DWORD id)
{
if (GC_win32_dll_threads) {
GC_vthread t = GC_lookup_thread_inner(id);
/* Check GC is initialized and the current thread is registered. */
LOCK();
GC_ASSERT(GC_lookup_thread_inner(GetCurrentThreadId()) != 0);
UNLOCK();
# endif
# if !defined(GC_ALWAYS_MULTITHREADED) && !defined(PARALLEL_MARK) \
&& !defined(GC_NO_THREADS_DISCOVERY)
/* GC_init() does not call GC_init_parallel() in this case. */
parallel_initialized = TRUE;
# endif
GC_start_mark_threads();
set_need_to_lock();
}
if (GC_need_to_lock == FALSE)
ABORT("Threads explicit registering is not previously enabled");
/* We lock here, since we want to wait for an ongoing GC. */
LOCK();
me = GC_lookup_thread_inner(thread_id);
if (me == 0) {
me = GC_register_my_thread_inner(sb, thread_id);
# ifdef GC_PTHREADS
# if defined(CPPCHECK)
GC_noop1(me->flags);
# endif
me -> flags |= DETACHED;
/* Treat as detached, since we do not need to worry about */
/* pointer results. */
# else
(void)me;
# endif
} else
# ifdef GC_PTHREADS
/* else */ if ((me -> flags & FINISHED) != 0) {
GC_record_stack_base(me, sb);
me -> flags &= ~FINISHED; /* but not DETACHED */
} else
# endif
/* else */ {
UNLOCK();
return GC_DUPLICATE;
}
#ifdef GC_DISABLE_INCREMENTAL
# define GC_wait_for_gc_completion(wait_for_all) (void)(wait_for_all)
#else
/* Similar to that in pthread_support.c. */
STATIC void GC_wait_for_gc_completion(GC_bool wait_for_all)
{
GC_ASSERT(I_HOLD_LOCK());
if (GC_incremental && GC_collection_in_progress()) {
word old_gc_no = GC_gc_no;
/* Make sure that no part of our stack is still on the mark stack, */
/* since it's about to be unmapped. */
do {
ENTER_GC();
GC_in_thread_creation = TRUE;
GC_collect_a_little_inner(1);
GC_in_thread_creation = FALSE;
EXIT_GC();
if (GC_win32_dll_threads) {
# if defined(THREAD_LOCAL_ALLOC)
/* Can't happen: see GC_use_threads_discovery(). */
GC_ASSERT(FALSE);
# else
/* FIXME: Should we just ignore this? */
GC_delete_thread(GetCurrentThreadId());
# endif
} else {
# if defined(THREAD_LOCAL_ALLOC) || defined(GC_PTHREADS)
GC_thread me;
# endif
DWORD thread_id = GetCurrentThreadId();
LOCK();
GC_wait_for_gc_completion(FALSE);
# if defined(THREAD_LOCAL_ALLOC) || defined(GC_PTHREADS)
me = GC_lookup_thread_inner(thread_id);
CHECK_LOOKUP_MY_THREAD(me);
GC_ASSERT(!KNOWN_FINISHED(me));
# endif
# if defined(THREAD_LOCAL_ALLOC)
GC_ASSERT(GC_getspecific(GC_thread_key) == &me->tlfs);
GC_destroy_thread_local(&(me->tlfs));
# endif
# ifdef GC_PTHREADS
if ((me -> flags & DETACHED) == 0) {
me -> flags |= FINISHED;
} else
# endif
/* else */ {
GC_delete_thread(thread_id);
}
# if defined(THREAD_LOCAL_ALLOC)
/* It is required to call remove_specific defined in specific.c. */
GC_remove_specific(GC_thread_key);
# endif
UNLOCK();
}
return GC_SUCCESS;
}
/* Wrapper for functions that are likely to block for an appreciable */
/* length of time. */
/* GC_do_blocking_inner() is nearly the same as in pthread_support.c */
GC_INNER void GC_do_blocking_inner(ptr_t data, void * context GC_ATTR_UNUSED)
{
struct blocking_data * d = (struct blocking_data *) data;
DWORD thread_id = GetCurrentThreadId();
GC_thread me;
# ifdef IA64
ptr_t stack_ptr = GC_save_regs_in_stack();
# endif
DCL_LOCK_STATE;
LOCK();
me = GC_lookup_thread_inner(thread_id);
CHECK_LOOKUP_MY_THREAD(me);
GC_ASSERT(me -> thread_blocked_sp == NULL);
# ifdef IA64
me -> backing_store_ptr = stack_ptr;
# endif
me -> thread_blocked_sp = (ptr_t) &d; /* save approx. sp */
/* Save context here if we want to support precise stack marking */
UNLOCK();
d -> client_data = (d -> fn)(d -> client_data);
LOCK(); /* This will block if the world is stopped. */
# if defined(CPPCHECK)
GC_noop1((word)me->thread_blocked_sp);
# endif
me -> thread_blocked_sp = NULL;
UNLOCK();
}
/* GC_call_with_gc_active() has the opposite to GC_do_blocking() */
/* functionality. It might be called from a user function invoked by */
/* GC_do_blocking() to temporarily back allow calling any GC function */
/* and/or manipulating pointers to the garbage collected heap. */
GC_API void * GC_CALL GC_call_with_gc_active(GC_fn_type fn,
void * client_data)
{
struct GC_traced_stack_sect_s stacksect;
DWORD thread_id = GetCurrentThreadId();
GC_thread me;
DCL_LOCK_STATE;
LOCK(); /* This will block if the world is stopped. */
me = GC_lookup_thread_inner(thread_id);
CHECK_LOOKUP_MY_THREAD(me);
/* Adjust our stack bottom pointer (this could happen unless */
/* GC_get_stack_base() was used which returned GC_SUCCESS). */
GC_ASSERT(me -> stack_base != NULL);
if ((word)me->stack_base < (word)(&stacksect)) {
me -> stack_base = (ptr_t)(&stacksect);
# if defined(I386)
me -> initial_stack_base = me -> stack_base;
# endif
}
if (me -> thread_blocked_sp == NULL) {
/* We are not inside GC_do_blocking() - do nothing more. */
UNLOCK();
client_data = fn(client_data);
/* Prevent treating the above as a tail call. */
GC_noop1(COVERT_DATAFLOW(&stacksect));
return client_data; /* result */
}
/* Setup new "stack section". */
stacksect.saved_stack_ptr = me -> thread_blocked_sp;
# ifdef IA64
/* This is the same as in GC_call_with_stack_base(). */
stacksect.backing_store_end = GC_save_regs_in_stack();
/* Unnecessarily flushes register stack, */
/* but that probably doesn't hurt. */
stacksect.saved_backing_store_ptr = me -> backing_store_ptr;
# endif
stacksect.prev = me -> traced_stack_sect;
me -> thread_blocked_sp = NULL;
me -> traced_stack_sect = &stacksect;
GC_ASSERT(I_HOLD_LOCK());
if (NULL == t) { /* current thread? */
t = GC_lookup_thread_inner(GetCurrentThreadId());
CHECK_LOOKUP_MY_THREAD(t);
}
GC_ASSERT(!KNOWN_FINISHED(t));
GC_ASSERT(NULL == t -> thread_blocked_sp
&& NULL == t -> traced_stack_sect); /* for now */
t -> stack_base = (ptr_t)sb->mem_base;
t -> last_stack_min = ADDR_LIMIT; /* reset the known minimum */
# ifdef IA64
t -> backing_store_end = (ptr_t)sb->reg_base;
# endif
}
LOCK();
me = GC_lookup_thread_inner(thread_id);
CHECK_LOOKUP_MY_THREAD(me); /* the thread is assumed to be registered */
sb -> mem_base = me -> stack_base;
# ifdef IA64
sb -> reg_base = me -> backing_store_end;
# endif
UNLOCK();
return (void *)me; /* gc_thread_handle */
}
#ifdef GC_PTHREADS
/* A quick-and-dirty cache of the mapping between pthread_t */
/* and Win32 thread id. */
# define PTHREAD_MAP_SIZE 512
DWORD GC_pthread_map_cache[PTHREAD_MAP_SIZE] = {0};
# define PTHREAD_MAP_INDEX(pthread_id) \
((NUMERIC_THREAD_ID(pthread_id) >> 5) % PTHREAD_MAP_SIZE)
/* It appears pthread_t is really a pointer type ... */
# define SET_PTHREAD_MAP_CACHE(pthread_id, win32_id) \
(void)(GC_pthread_map_cache[PTHREAD_MAP_INDEX(pthread_id)] = (win32_id))
# define GET_PTHREAD_MAP_CACHE(pthread_id) \
GC_pthread_map_cache[PTHREAD_MAP_INDEX(pthread_id)]
/* Return a GC_thread corresponding to a given pthread_t. */
/* Returns 0 if it's not there. */
/* We assume that this is only called for pthread ids that */
/* have not yet terminated or are still joinable, and */
/* cannot be concurrently terminated. */
/* Assumes we do NOT hold the allocation lock. */
STATIC GC_thread GC_lookup_pthread(pthread_t id)
{
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
int i;
LONG my_max = GC_get_max_thread_index();
for (i = 0; i <= my_max &&
(!AO_load_acquire(&dll_thread_table[i].tm.in_use)
|| !THREAD_EQUAL(dll_thread_table[i].pthread_id, id));
/* Must still be in_use, since nobody else can */
/* store our thread_id. */
i++) {
/* empty */
}
return i <= my_max ? (GC_thread)(dll_thread_table + i) : NULL;
} else
# endif
/* else */ {
/* We first try the cache. If that fails, we use a very slow */
/* approach. */
DWORD win32_id = GET_PTHREAD_MAP_CACHE(id);
int hv_guess = THREAD_TABLE_INDEX(win32_id);
int hv;
GC_thread p;
DCL_LOCK_STATE;
LOCK();
for (p = GC_threads[hv_guess]; 0 != p; p = p -> tm.next) {
if (THREAD_EQUAL(p -> pthread_id, id))
goto foundit;
}
for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
for (p = GC_threads[hv]; 0 != p; p = p -> tm.next) {
if (THREAD_EQUAL(p -> pthread_id, id))
goto foundit;
}
}
p = 0;
foundit:
UNLOCK();
return p;
}
}
#endif /* GC_PTHREADS */
#ifdef CAN_HANDLE_FORK
/* Similar to that in pthread_support.c but also rehashes the table */
/* since hash map key (thread_id) differs from that in the parent. */
STATIC void GC_remove_all_threads_but_me(void)
{
int hv;
GC_thread me = NULL;
DWORD thread_id;
pthread_t pthread_id = pthread_self(); /* same as in parent */
for (p = GC_threads[hv]; 0 != p; p = next) {
next = p -> tm.next;
if (THREAD_EQUAL(p -> pthread_id, pthread_id)
&& me == NULL) { /* ignore dead threads with the same id */
me = p;
p -> tm.next = 0;
} else {
# ifdef THREAD_LOCAL_ALLOC
if ((p -> flags & FINISHED) == 0) {
/* Cannot call GC_destroy_thread_local here (see the */
/* corresponding comment in pthread_support.c). */
GC_remove_specific_after_fork(GC_thread_key, p -> pthread_id);
}
# endif
if (&first_thread != p)
GC_INTERNAL_FREE(p);
}
}
GC_threads[hv] = NULL;
}
/* Put "me" back to GC_threads. */
GC_ASSERT(me != NULL);
thread_id = GetCurrentThreadId(); /* differs from that in parent */
GC_threads[THREAD_TABLE_INDEX(thread_id)] = me;
/* Update Win32 thread Id and handle. */
me -> id = thread_id;
# ifndef MSWINCE
if (!DuplicateHandle(GetCurrentProcess(), GetCurrentThread(),
GetCurrentProcess(), (HANDLE *)&me->handle,
0 /* dwDesiredAccess */, FALSE /* bInheritHandle */,
DUPLICATE_SAME_ACCESS))
ABORT("DuplicateHandle failed");
# endif
# if defined(THREAD_LOCAL_ALLOC) && !defined(USE_CUSTOM_SPECIFIC)
/* For Cygwin, we need to re-assign thread-local pointer to */
/* 'tlfs' (it is OK to call GC_destroy_thread_local and */
/* GC_free_inner before this action). */
if (GC_setspecific(GC_thread_key, &me->tlfs) != 0)
ABORT("GC_setspecific failed (in child)");
# endif
}
static void fork_child_proc(void)
{
# ifdef PARALLEL_MARK
if (GC_parallel) {
GC_release_mark_lock();
GC_parallel = FALSE; /* or GC_markers_m1 = 0 */
/* Turn off parallel marking in the child, since we are */
/* probably just going to exec, and we would have to */
/* restart mark threads. */
}
# endif
GC_remove_all_threads_but_me();
UNLOCK();
}
/* Routines for fork handling by client (no-op if pthread_atfork works). */
GC_API void GC_CALL GC_atfork_prepare(void)
{
if (!EXPECT(GC_is_initialized, TRUE)) GC_init();
if (GC_handle_fork <= 0)
fork_prepare_proc();
}
void GC_push_thread_structures(void)
{
GC_ASSERT(I_HOLD_LOCK());
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
/* Unlike the other threads implementations, the thread table */
/* here contains no pointers to the collectible heap (note also */
/* that GC_PTHREADS is incompatible with DllMain-based thread */
/* registration). Thus we have no private structures we need */
/* to preserve. */
} else
# endif
/* else */ {
GC_push_all(&GC_threads, (ptr_t)(&GC_threads) + sizeof(GC_threads));
}
# if defined(THREAD_LOCAL_ALLOC) && defined(USE_CUSTOM_SPECIFIC)
GC_PUSH_ALL_SYM(GC_thread_key);
/* Just in case we ever use our own TLS implementation. */
# endif
}
# ifdef MSWINCE
/* SuspendThread() will fail if thread is running kernel code. */
while (SuspendThread(THREAD_HANDLE(t)) == (DWORD)-1) {
GC_release_dirty_lock();
Sleep(10); /* in millis */
GC_acquire_dirty_lock();
}
# elif defined(RETRY_GET_THREAD_CONTEXT)
for (retry_cnt = 0;;) {
/* Apparently the Windows 95 GetOpenFileName call creates */
/* a thread that does not properly get cleaned up, and */
/* SuspendThread on its descriptor may provoke a crash. */
/* This reduces the probability of that event, though it still */
/* appears there is a race here. */
if (GetExitCodeThread(t -> handle, &exitCode)
&& exitCode != STILL_ACTIVE) {
GC_release_dirty_lock();
# ifdef GC_PTHREADS
t -> stack_base = 0; /* prevent stack from being pushed */
# else
/* This breaks pthread_join on Cygwin, which is guaranteed to */
/* only see user threads. */
GC_ASSERT(GC_win32_dll_threads);
GC_delete_gc_thread_no_free(t);
# endif
return;
}
if (SuspendThread(t->handle) != (DWORD)-1) {
CONTEXT context;
context.ContextFlags = GET_THREAD_CONTEXT_FLAGS;
if (GetThreadContext(t->handle, &context)) {
/* TODO: WoW64 extra workaround: if CONTEXT_EXCEPTION_ACTIVE */
/* then Sleep(1) and retry. */
t->context_sp = copy_ptr_regs(t->context_regs, &context);
break; /* success; the context pointer registers are saved */
}
/* Resume the thread, try to suspend it in a better location. */
if (ResumeThread(t->handle) == (DWORD)-1)
ABORT("ResumeThread failed in suspend loop");
}
if (retry_cnt > 1) {
GC_release_dirty_lock();
Sleep(0); /* yield */
GC_acquire_dirty_lock();
}
if (++retry_cnt >= MAX_SUSPEND_THREAD_RETRIES)
ABORT("SuspendThread loop failed"); /* something must be wrong */
}
# else
if (GetExitCodeThread(t -> handle, &exitCode)
&& exitCode != STILL_ACTIVE) {
GC_release_dirty_lock();
# ifdef GC_PTHREADS
t -> stack_base = 0; /* prevent stack from being pushed */
# else
GC_ASSERT(GC_win32_dll_threads);
GC_delete_gc_thread_no_free(t);
# endif
return;
}
if (SuspendThread(t -> handle) == (DWORD)-1)
ABORT("SuspendThread failed");
# endif
t -> suspended = (unsigned char)TRUE;
GC_release_dirty_lock();
if (GC_on_thread_event)
GC_on_thread_event(GC_EVENT_THREAD_SUSPENDED, THREAD_HANDLE(t));
}
if (!GC_thr_initialized)
ABORT("GC_stop_world() called before GC_thr_init()");
GC_ASSERT(I_HOLD_LOCK());
/* This code is the same as in pthread_stop_world.c */
# ifdef PARALLEL_MARK
if (GC_parallel) {
GC_acquire_mark_lock();
GC_ASSERT(GC_fl_builder_count == 0);
/* We should have previously waited for it to become zero. */
}
# endif /* PARALLEL_MARK */
# if !defined(GC_NO_THREADS_DISCOVERY) || defined(GC_ASSERTIONS)
GC_please_stop = TRUE;
# endif
# if (defined(MSWIN32) && !defined(CONSOLE_LOG)) || defined(MSWINCE)
GC_ASSERT(!GC_write_disabled);
EnterCriticalSection(&GC_write_cs);
/* It's not allowed to call GC_printf() (and friends) here down to */
/* LeaveCriticalSection (same applies recursively to GC_suspend, */
/* GC_delete_gc_thread_no_free, GC_get_max_thread_index, GC_size */
/* and GC_remove_protection). */
# ifdef GC_ASSERTIONS
GC_write_disabled = TRUE;
# endif
# endif
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
int i;
int my_max;
/* Any threads being created during this loop will end up setting */
/* GC_attached_thread when they start. This will force marking */
/* to restart. This is not ideal, but hopefully correct. */
AO_store(&GC_attached_thread, FALSE);
my_max = (int)GC_get_max_thread_index();
for (i = 0; i <= my_max; i++) {
GC_vthread t = dll_thread_table + i;
if (t -> stack_base != 0 && t -> thread_blocked_sp == NULL
&& t -> id != thread_id) {
GC_suspend((GC_thread)t);
}
}
} else
# endif
/* else */ {
GC_thread t;
int i;
for (i = 0; i < THREAD_TABLE_SZ; i++) {
for (t = GC_threads[i]; t != 0; t = t -> tm.next) {
if (t -> stack_base != 0 && t -> thread_blocked_sp == NULL
&& !KNOWN_FINISHED(t) && t -> id != thread_id) {
GC_suspend(t);
}
}
}
}
# if (defined(MSWIN32) && !defined(CONSOLE_LOG)) || defined(MSWINCE)
# ifdef GC_ASSERTIONS
GC_write_disabled = FALSE;
# endif
LeaveCriticalSection(&GC_write_cs);
# endif
# ifdef PARALLEL_MARK
if (GC_parallel)
GC_release_mark_lock();
# endif
}
GC_ASSERT(I_HOLD_LOCK());
if (GC_win32_dll_threads) {
LONG my_max = GC_get_max_thread_index();
int i;
for (i = 0; i <= my_max; i++) {
GC_thread t = (GC_thread)(dll_thread_table + i);
if (t -> suspended) {
# ifdef DEBUG_THREADS
GC_log_printf("Resuming 0x%x\n", (int)t->id);
# endif
GC_ASSERT(t -> stack_base != 0 && t -> id != thread_id);
if (ResumeThread(THREAD_HANDLE(t)) == (DWORD)-1)
ABORT("ResumeThread failed");
t -> suspended = FALSE;
if (GC_on_thread_event)
GC_on_thread_event(GC_EVENT_THREAD_UNSUSPENDED, THREAD_HANDLE(t));
}
/* Else thread is unregistered or not suspended. */
}
} else {
GC_thread t;
int i;
for (i = 0; i < THREAD_TABLE_SZ; i++) {
for (t = GC_threads[i]; t != 0; t = t -> tm.next) {
if (t -> suspended) {
# ifdef DEBUG_THREADS
GC_log_printf("Resuming 0x%x\n", (int)t->id);
# endif
GC_ASSERT(t -> stack_base != 0 && t -> id != thread_id);
if (ResumeThread(THREAD_HANDLE(t)) == (DWORD)-1)
ABORT("ResumeThread failed");
UNPROTECT_THREAD(t);
t -> suspended = FALSE;
if (GC_on_thread_event)
GC_on_thread_event(GC_EVENT_THREAD_UNSUSPENDED, THREAD_HANDLE(t));
} else {
# ifdef DEBUG_THREADS
GC_log_printf("Not resuming thread 0x%x as it is not suspended\n",
(int)t->id);
# endif
}
}
}
}
# if !defined(GC_NO_THREADS_DISCOVERY) || defined(GC_ASSERTIONS)
GC_please_stop = FALSE;
# endif
}
#ifdef MSWINCE
/* The VirtualQuery calls below won't work properly on some old WinCE */
/* versions, but since each stack is restricted to an aligned 64 KiB */
/* region of virtual memory we can just take the next lowest multiple */
/* of 64 KiB. The result of this macro must not be used as its */
/* argument later and must not be used as the lower bound for sp */
/* check (since the stack may be bigger than 64 KiB). */
# define GC_wince_evaluate_stack_min(s) \
(ptr_t)(((word)(s) - 1) & ~(word)0xFFFF)
#elif defined(GC_ASSERTIONS)
# define GC_dont_query_stack_min FALSE
#endif
/* A cache holding the results of the recent VirtualQuery call. */
/* Protected by the allocation lock. */
static ptr_t last_address = 0;
static MEMORY_BASIC_INFORMATION last_info;
/* Probe stack memory region (starting at "s") to find out its */
/* lowest address (i.e. stack top). */
/* S must be a mapped address inside the region, NOT the first */
/* unmapped address. */
STATIC ptr_t GC_get_stack_min(ptr_t s)
{
ptr_t bottom;
GC_ASSERT(I_HOLD_LOCK());
if (s != last_address) {
VirtualQuery(s, &last_info, sizeof(last_info));
last_address = s;
}
do {
bottom = (ptr_t)last_info.BaseAddress;
VirtualQuery(bottom - 1, &last_info, sizeof(last_info));
last_address = bottom - 1;
} while ((last_info.Protect & PAGE_READWRITE)
&& !(last_info.Protect & PAGE_GUARD));
return(bottom);
}
/* Return true if the page at s has protections appropriate */
/* for a stack page. */
static GC_bool may_be_in_stack(ptr_t s)
{
GC_ASSERT(I_HOLD_LOCK());
if (s != last_address) {
VirtualQuery(s, &last_info, sizeof(last_info));
last_address = s;
}
return (last_info.Protect & PAGE_READWRITE)
&& !(last_info.Protect & PAGE_GUARD);
}
/* Copy all registers that might point into the heap. Frame */
/* pointer registers are included in case client code was */
/* compiled with the 'omit frame pointer' optimization. */
/* The context register values are stored to regs argument */
/* which is expected to be of PUSHED_REGS_COUNT length exactly. */
/* The functions returns the context stack pointer value. */
static ptr_t copy_ptr_regs(word *regs, const CONTEXT *pcontext) {
ptr_t sp;
int cnt = 0;
# define context (*pcontext)
# define PUSH1(reg) (regs[cnt++] = (word)pcontext->reg)
# define PUSH2(r1,r2) (PUSH1(r1), PUSH1(r2))
# define PUSH4(r1,r2,r3,r4) (PUSH2(r1,r2), PUSH2(r3,r4))
# define PUSH8_LH(r1,r2,r3,r4) (PUSH4(r1.Low,r1.High,r2.Low,r2.High), \
PUSH4(r3.Low,r3.High,r4.Low,r4.High))
# if defined(I386)
# ifdef WOW64_THREAD_CONTEXT_WORKAROUND
PUSH2(ContextFlags, SegFs); /* cannot contain pointers */
# endif
PUSH4(Edi,Esi,Ebx,Edx), PUSH2(Ecx,Eax), PUSH1(Ebp);
sp = (ptr_t)context.Esp;
# elif defined(X86_64)
PUSH4(Rax,Rcx,Rdx,Rbx); PUSH2(Rbp, Rsi); PUSH1(Rdi);
PUSH4(R8, R9, R10, R11); PUSH4(R12, R13, R14, R15);
# ifndef XMM_CANT_STORE_PTRS
PUSH8_LH(Xmm0, Xmm1, Xmm2, Xmm3);
PUSH8_LH(Xmm4, Xmm5, Xmm6, Xmm7);
PUSH8_LH(Xmm8, Xmm9, Xmm10, Xmm11);
PUSH8_LH(Xmm12, Xmm13, Xmm14, Xmm15);
# endif
sp = (ptr_t)context.Rsp;
# elif defined(ARM32)
PUSH4(R0,R1,R2,R3),PUSH4(R4,R5,R6,R7),PUSH4(R8,R9,R10,R11);
PUSH1(R12);
sp = (ptr_t)context.Sp;
# elif defined(AARCH64)
PUSH4(X0,X1,X2,X3),PUSH4(X4,X5,X6,X7),PUSH4(X8,X9,X10,X11);
PUSH4(X12,X13,X14,X15),PUSH4(X16,X17,X18,X19),PUSH4(X20,X21,X22,X23);
PUSH4(X24,X25,X26,X27),PUSH1(X28);
PUSH1(Lr);
sp = (ptr_t)context.Sp;
# elif defined(SHx)
PUSH4(R0,R1,R2,R3), PUSH4(R4,R5,R6,R7), PUSH4(R8,R9,R10,R11);
PUSH2(R12,R13), PUSH1(R14);
sp = (ptr_t)context.R15;
# elif defined(MIPS)
PUSH4(IntAt,IntV0,IntV1,IntA0), PUSH4(IntA1,IntA2,IntA3,IntT0);
PUSH4(IntT1,IntT2,IntT3,IntT4), PUSH4(IntT5,IntT6,IntT7,IntS0);
PUSH4(IntS1,IntS2,IntS3,IntS4), PUSH4(IntS5,IntS6,IntS7,IntT8);
PUSH4(IntT9,IntK0,IntK1,IntS8);
sp = (ptr_t)context.IntSp;
# elif defined(PPC)
PUSH4(Gpr0, Gpr3, Gpr4, Gpr5), PUSH4(Gpr6, Gpr7, Gpr8, Gpr9);
PUSH4(Gpr10,Gpr11,Gpr12,Gpr14), PUSH4(Gpr15,Gpr16,Gpr17,Gpr18);
PUSH4(Gpr19,Gpr20,Gpr21,Gpr22), PUSH4(Gpr23,Gpr24,Gpr25,Gpr26);
PUSH4(Gpr27,Gpr28,Gpr29,Gpr30), PUSH1(Gpr31);
sp = (ptr_t)context.Gpr1;
# elif defined(ALPHA)
PUSH4(IntV0,IntT0,IntT1,IntT2), PUSH4(IntT3,IntT4,IntT5,IntT6);
PUSH4(IntT7,IntS0,IntS1,IntS2), PUSH4(IntS3,IntS4,IntS5,IntFp);
PUSH4(IntA0,IntA1,IntA2,IntA3), PUSH4(IntA4,IntA5,IntT8,IntT9);
PUSH4(IntT10,IntT11,IntT12,IntAt);
sp = (ptr_t)context.IntSp;
# elif defined(CPPCHECK)
sp = (ptr_t)(word)cnt; /* to workaround "cnt not used" false positive */
# else
# error Architecture is not supported
# endif
# undef context
GC_ASSERT(cnt == PUSHED_REGS_COUNT);
return sp;
}
STATIC word GC_push_stack_for(GC_thread thread, DWORD me)
{
ptr_t sp, stack_min;
struct GC_traced_stack_sect_s *traced_stack_sect =
thread -> traced_stack_sect;
if (thread -> id == me) {
GC_ASSERT(thread -> thread_blocked_sp == NULL);
sp = GC_approx_sp();
} else if ((sp = thread -> thread_blocked_sp) == NULL) {
/* Use saved sp value for blocked threads. */
# ifdef RETRY_GET_THREAD_CONTEXT
/* We cache context when suspending the thread since it may */
/* require looping. */
word *regs = thread->context_regs;
if (thread->suspended) {
sp = thread->context_sp;
} else
# else
word regs[PUSHED_REGS_COUNT];
# endif
/* else */ {
CONTEXT context;
/* For unblocked threads call GetThreadContext(). */
context.ContextFlags = GET_THREAD_CONTEXT_FLAGS;
if (GetThreadContext(THREAD_HANDLE(thread), &context)) {
sp = copy_ptr_regs(regs, &context);
} else {
# ifdef RETRY_GET_THREAD_CONTEXT
/* At least, try to use the stale context if saved. */
sp = thread->context_sp;
if (NULL == sp) {
/* Skip the current thread, anyway its stack will */
/* be pushed when the world is stopped. */
return 0;
}
# else
ABORT("GetThreadContext failed");
# endif
}
}
# ifdef THREAD_LOCAL_ALLOC
GC_ASSERT(thread->suspended || !GC_world_stopped);
# endif
if (!GetThreadSelectorEntry(THREAD_HANDLE(thread), SegFs,
&selector))
ABORT("GetThreadSelectorEntry failed");
tib = (PNT_TIB)(selector.BaseLow
| (selector.HighWord.Bits.BaseMid << 16)
| (selector.HighWord.Bits.BaseHi << 24));
# endif
# ifdef DEBUG_THREADS
GC_log_printf("TIB stack limit/base: %p .. %p\n",
(void *)tib->StackLimit, (void *)tib->StackBase);
# endif
GC_ASSERT(!((word)thread->stack_base
COOLER_THAN (word)tib->StackBase));
if (thread->stack_base != thread->initial_stack_base
/* We are in a coroutine. */
&& ((word)thread->stack_base <= (word)tib->StackLimit
|| (word)tib->StackBase < (word)thread->stack_base)) {
/* The coroutine stack is not within TIB stack. */
WARN("GetThreadContext might return stale register values"
" including ESP= %p\n", sp);
/* TODO: Because of WoW64 bug, there is no guarantee that */
/* sp really points to the stack top but, for now, we do */
/* our best as the TIB stack limit/base cannot be used */
/* while we are inside a coroutine. */
} else {
/* GetThreadContext() might return stale register values, */
/* so we scan the entire stack region (down to the stack */
/* limit). There is no 100% guarantee that all the */
/* registers are pushed but we do our best (the proper */
/* solution would be to fix it inside Windows). */
sp = (ptr_t)tib->StackLimit;
}
} /* else */
# ifdef DEBUG_THREADS
else {
static GC_bool logged;
if (!logged
&& (ContextFlags & CONTEXT_EXCEPTION_REPORTING) == 0) {
GC_log_printf("CONTEXT_EXCEPTION_REQUEST not supported\n");
logged = TRUE;
}
}
# endif
}
# endif /* WOW64_THREAD_CONTEXT_WORKAROUND */
} /* ! current thread */
# ifdef STACKPTR_CORRECTOR_AVAILABLE
if (GC_sp_corrector != 0)
GC_sp_corrector((void **)&sp, (void *)(thread -> pthread_id));
# endif
/* Set stack_min to the lowest address in the thread stack, */
/* or to an address in the thread stack no larger than sp, */
/* taking advantage of the old value to avoid slow traversals */
/* of large stacks. */
if (thread -> last_stack_min == ADDR_LIMIT) {
# ifdef MSWINCE
if (GC_dont_query_stack_min) {
stack_min = GC_wince_evaluate_stack_min(traced_stack_sect != NULL ?
(ptr_t)traced_stack_sect : thread -> stack_base);
/* Keep last_stack_min value unmodified. */
} else
# endif
/* else */ {
stack_min = GC_get_stack_min(traced_stack_sect != NULL ?
(ptr_t)traced_stack_sect : thread -> stack_base);
UNPROTECT_THREAD(thread);
thread -> last_stack_min = stack_min;
}
} else {
/* First, adjust the latest known minimum stack address if we */
/* are inside GC_call_with_gc_active(). */
if (traced_stack_sect != NULL &&
(word)thread->last_stack_min > (word)traced_stack_sect) {
UNPROTECT_THREAD(thread);
thread -> last_stack_min = (ptr_t)traced_stack_sect;
}
if ((word)sp < (word)thread->stack_base
&& (word)sp >= (word)thread->last_stack_min) {
stack_min = sp;
} else {
/* In the current thread it is always safe to use sp value. */
if (may_be_in_stack(thread -> id == me &&
(word)sp < (word)thread->last_stack_min ?
sp : thread -> last_stack_min)) {
stack_min = (ptr_t)last_info.BaseAddress;
/* Do not probe rest of the stack if sp is correct. */
if ((word)sp < (word)stack_min
|| (word)sp >= (word)thread->stack_base)
stack_min = GC_get_stack_min(thread -> last_stack_min);
} else {
/* Stack shrunk? Is this possible? */
stack_min = GC_get_stack_min(thread -> stack_base);
}
UNPROTECT_THREAD(thread);
thread -> last_stack_min = stack_min;
}
}
if ((word)sp >= (word)stack_min && (word)sp < (word)thread->stack_base) {
# ifdef DEBUG_THREADS
GC_log_printf("Pushing stack for 0x%x from sp %p to %p from 0x%x\n",
(int)thread->id, (void *)sp, (void *)thread->stack_base,
(int)me);
# endif
GC_push_all_stack_sections(sp, thread->stack_base, traced_stack_sect);
} else {
/* If not current thread then it is possible for sp to point to */
/* the guarded (untouched yet) page just below the current */
/* stack_min of the thread. */
if (thread -> id == me || (word)sp >= (word)thread->stack_base
|| (word)(sp + GC_page_size) < (word)stack_min)
WARN("Thread stack pointer %p out of range, pushing everything\n",
sp);
# ifdef DEBUG_THREADS
GC_log_printf("Pushing stack for 0x%x from (min) %p to %p from 0x%x\n",
(int)thread->id, (void *)stack_min,
(void *)thread->stack_base, (int)me);
# endif
/* Push everything - ignore "traced stack section" data. */
GC_push_all_stack(stack_min, thread->stack_base);
}
return thread->stack_base - sp; /* stack grows down */
}
/* We hold allocation lock. Should do exactly the right thing if the */
/* world is stopped. Should not fail if it isn't. */
GC_INNER void GC_push_all_stacks(void)
{
DWORD thread_id = GetCurrentThreadId();
GC_bool found_me = FALSE;
# ifndef SMALL_CONFIG
unsigned nthreads = 0;
# endif
word total_size = 0;
# ifndef GC_NO_THREADS_DISCOVERY
if (GC_win32_dll_threads) {
int i;
LONG my_max = GC_get_max_thread_index();
for (i = 0; i <= my_max; i++) {
GC_thread t = (GC_thread)(dll_thread_table + i);
if (t -> tm.in_use && t -> stack_base) {
# ifndef SMALL_CONFIG
++nthreads;
# endif
total_size += GC_push_stack_for(t, thread_id);
if (t -> id == thread_id) found_me = TRUE;
}
}
} else
# endif
/* else */ {
int i;
for (i = 0; i < THREAD_TABLE_SZ; i++) {
GC_thread t;
for (t = GC_threads[i]; t != 0; t = t -> tm.next) {
if (!KNOWN_FINISHED(t) && t -> stack_base) {
# ifndef SMALL_CONFIG
++nthreads;
# endif
total_size += GC_push_stack_for(t, thread_id);
if (t -> id == thread_id) found_me = TRUE;
}
}
}
}
# ifndef SMALL_CONFIG
GC_VERBOSE_LOG_PRINTF("Pushed %d thread stacks%s\n", nthreads,
GC_win32_dll_threads ?
" based on DllMain thread tracking" : "");
# endif
if (!found_me && !GC_in_thread_creation)
ABORT("Collecting from unknown thread");
GC_total_stacksize = total_size;
}
static ptr_t marker_sp[MAX_MARKERS - 1]; /* The cold end of the stack */
/* for markers. */
# ifdef IA64
static ptr_t marker_bsp[MAX_MARKERS - 1];
# endif
static ptr_t marker_last_stack_min[MAX_MARKERS - 1];
/* Last known minimum (hottest) address */
/* in stack (or ADDR_LIMIT if unset) */
/* for markers. */
#endif /* PARALLEL_MARK */
/* Find stack with the lowest address which overlaps the */
/* interval [start, limit). */
/* Return stack bounds in *lo and *hi. If no such stack */
/* is found, both *hi and *lo will be set to an address */
/* higher than limit. */
GC_INNER void GC_get_next_stack(char *start, char *limit,
char **lo, char **hi)
{
int i;
char * current_min = ADDR_LIMIT; /* Least in-range stack base */
ptr_t *plast_stack_min = NULL; /* Address of last_stack_min */
/* field for thread corresponding */
/* to current_min. */
GC_thread thread = NULL; /* Either NULL or points to the */
/* thread's hash table entry */
/* containing *plast_stack_min. */
/* First set current_min, ignoring limit. */
if (GC_win32_dll_threads) {
LONG my_max = GC_get_max_thread_index();
for (i = 0; i <= my_max; i++) {
ptr_t s = (ptr_t)(dll_thread_table[i].stack_base);
if ((word)s > (word)start && (word)s < (word)current_min) {
/* Update address of last_stack_min. */
plast_stack_min = (ptr_t * /* no volatile */)
&dll_thread_table[i].last_stack_min;
current_min = s;
# if defined(CPPCHECK)
/* To avoid a warning that thread is always null. */
thread = (GC_thread)&dll_thread_table[i];
# endif
}
}
} else {
for (i = 0; i < THREAD_TABLE_SZ; i++) {
GC_thread t;
for (t = GC_threads[i]; t != 0; t = t -> tm.next) {
ptr_t s = t -> stack_base;
if ((word)s > (word)start && (word)s < (word)current_min) {
/* Update address of last_stack_min. */
plast_stack_min = &t -> last_stack_min;
thread = t; /* Remember current thread to unprotect. */
current_min = s;
}
}
}
# ifdef PARALLEL_MARK
for (i = 0; i < GC_markers_m1; ++i) {
ptr_t s = marker_sp[i];
# ifdef IA64
/* FIXME: not implemented */
# endif
if ((word)s > (word)start && (word)s < (word)current_min) {
GC_ASSERT(marker_last_stack_min[i] != NULL);
plast_stack_min = &marker_last_stack_min[i];
current_min = s;
thread = NULL; /* Not a thread's hash table entry. */
}
}
# endif
}
if ((word)current_min > (word)limit && !may_be_in_stack(limit)) {
/* Skip the rest since the memory region at limit address is */
/* not a stack (so the lowest address of the found stack would */
/* be above the limit value anyway). */
*lo = ADDR_LIMIT;
return;
}
/* Get the minimum address of the found stack by probing its memory */
/* region starting from the recent known minimum (if set). */
if (*plast_stack_min == ADDR_LIMIT
|| !may_be_in_stack(*plast_stack_min)) {
/* Unsafe to start from last_stack_min value. */
*lo = GC_get_stack_min(current_min);
} else {
/* Use the recent value to optimize search for min address. */
*lo = GC_get_stack_min(*plast_stack_min);
}
/* Remember current stack_min value. */
if (thread != NULL) {
UNPROTECT_THREAD(thread);
}
*plast_stack_min = *lo;
}
#ifdef PARALLEL_MARK
# if !defined(GC_PTHREADS_PARAMARK)
STATIC HANDLE GC_marker_cv[MAX_MARKERS - 1] = {0};
/* Events with manual reset (one for each */
/* mark helper). */
STATIC DWORD GC_marker_Id[MAX_MARKERS - 1] = {0};
/* This table is used for mapping helper */
/* threads ID to mark helper index (linear */
/* search is used since the mapping contains */
/* only a few entries). */
# endif
# if defined(GC_PTHREADS) && defined(HAVE_PTHREAD_SETNAME_NP_WITH_TID)
static void set_marker_thread_name(unsigned id)
{
/* This code is the same as in pthread_support.c. */
char name_buf[16]; /* pthread_setname_np may fail for longer names */
int len = sizeof("GC-marker-") - 1;
/* Compose the name manually as snprintf may be unavailable or */
/* "%u directive output may be truncated" warning may occur. */
BCOPY("GC-marker-", name_buf, len);
if (id >= 10)
name_buf[len++] = (char)('0' + (id / 10) % 10);
name_buf[len] = (char)('0' + id % 10);
name_buf[len + 1] = '\0';
if (pthread_setname_np(pthread_self(), name_buf) != 0)
WARN("pthread_setname_np failed\n", 0);
}
# elif !defined(MSWINCE)
/* A pointer to SetThreadDescription() which is available since */
/* Windows 10. The function prototype is in processthreadsapi.h. */
static FARPROC setThreadDescription_fn;
static void set_marker_thread_name(unsigned id)
{
WCHAR name_buf[16];
int len = sizeof(L"GC-marker-") / sizeof(WCHAR) - 1;
HRESULT hr;
if (!setThreadDescription_fn) return; /* missing SetThreadDescription */
/* Compose the name manually as swprintf may be unavailable. */
BCOPY(L"GC-marker-", name_buf, len * sizeof(WCHAR));
if (id >= 10)
name_buf[len++] = (WCHAR)('0' + (id / 10) % 10);
name_buf[len] = (WCHAR)('0' + id % 10);
name_buf[len + 1] = 0;
/* Invoke SetThreadDescription(). Cast the function pointer to word */
/* first to avoid "incompatible function types" compiler warning. */
hr = (*(HRESULT (WINAPI *)(HANDLE, const WCHAR *))
(word)setThreadDescription_fn)(GetCurrentThread(), name_buf);
if (FAILED(hr))
WARN("SetThreadDescription failed\n", 0);
}
# else
# define set_marker_thread_name(id) (void)(id)
# endif
/* GC_mark_thread() is the same as in pthread_support.c */
# ifdef GC_PTHREADS_PARAMARK
STATIC void * GC_mark_thread(void * id)
# elif defined(MSWINCE)
STATIC DWORD WINAPI GC_mark_thread(LPVOID id)
# else
STATIC unsigned __stdcall GC_mark_thread(void * id)
# endif
{
word my_mark_no = 0;
if ((word)id == GC_WORD_MAX) return 0; /* to prevent a compiler warning */
set_marker_thread_name((unsigned)(word)id);
marker_sp[(word)id] = GC_approx_sp();
# ifdef IA64
marker_bsp[(word)id] = GC_save_regs_in_stack();
# endif
# if !defined(GC_PTHREADS_PARAMARK)
GC_marker_Id[(word)id] = GetCurrentThreadId();
# endif
/* Inform GC_start_mark_threads about completion of marker data init. */
GC_acquire_mark_lock();
if (0 == --GC_fl_builder_count) /* count may have a negative value */
GC_notify_all_builder();
for (;; ++my_mark_no) {
if (my_mark_no - GC_mark_no > (word)2) {
/* resynchronize if we get far off, e.g. because GC_mark_no */
/* wrapped. */
my_mark_no = GC_mark_no;
}
# ifdef DEBUG_THREADS
GC_log_printf("Starting mark helper for mark number %lu\n",
(unsigned long)my_mark_no);
# endif
GC_help_marker(my_mark_no);
}
}
/* GC_mark_threads[] is unused here unlike that in pthread_support.c */
static int available_markers_m1 = 0;
# ifdef GC_PTHREADS_PARAMARK
# include <pthread.h>
# if defined(GC_ASSERTIONS) && !defined(NUMERIC_THREAD_ID)
# define NUMERIC_THREAD_ID(id) (unsigned long)(word)GC_PTHREAD_PTRVAL(id)
/* Id not guaranteed to be unique. */
# endif
GC_ASSERT(I_HOLD_LOCK());
ASSERT_CANCEL_DISABLED();
if (available_markers_m1 <= 0 || GC_parallel) return;
/* Skip if parallel markers disabled or already started. */
GC_wait_for_gc_completion(TRUE);
# ifdef CAN_HANDLE_FORK
/* Reset mark_cv state after forking (as in pthread_support.c). */
{
pthread_cond_t mark_cv_local = PTHREAD_COND_INITIALIZER;
BCOPY(&mark_cv_local, &mark_cv, sizeof(mark_cv));
}
# endif
GC_ASSERT(GC_fl_builder_count == 0);
if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
ABORT("pthread_attr_setdetachstate failed");
# ifndef NO_MARKER_SPECIAL_SIGMASK
/* Apply special signal mask to GC marker threads, and don't drop */
/* user defined signals by GC marker threads. */
if (sigfillset(&set) != 0)
ABORT("sigfillset failed");
if (pthread_sigmask(SIG_BLOCK, &set, &oldset) < 0) {
WARN("pthread_sigmask set failed, no markers started\n", 0);
GC_markers_m1 = 0;
(void)pthread_attr_destroy(&attr);
return;
}
# endif /* !NO_MARKER_SPECIAL_SIGMASK */
/* To have proper GC_parallel value in GC_help_marker. */
GC_markers_m1 = available_markers_m1;
for (i = 0; i < available_markers_m1; ++i) {
marker_last_stack_min[i] = ADDR_LIMIT;
if (0 != pthread_create(&new_thread, &attr,
GC_mark_thread, (void *)(word)i)) {
WARN("Marker thread creation failed\n", 0);
/* Don't try to create other marker threads. */
GC_markers_m1 = i;
break;
}
}
/* GC_acquire/release_mark_lock(), GC_wait_builder/marker(), */
/* GC_wait_for_reclaim(), GC_notify_all_builder/marker() are the same */
/* as in pthread_support.c except that GC_generic_lock() is not used. */
/* Collector must wait for a freelist builders for 2 reasons: */
/* 1) Mark bits may still be getting examined without lock. */
/* 2) Partial free lists referenced only by locals may not be */
/* scanned correctly, e.g. if they contain "pointer-free" objects, */
/* since the free-list link may be ignored. */
STATIC void GC_wait_builder(void)
{
UNSET_MARK_LOCK_HOLDER;
if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
ABORT("pthread_cond_wait failed");
}
GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
SET_MARK_LOCK_HOLDER;
}
GC_INNER void GC_release_mark_lock(void)
{
UNSET_MARK_LOCK_HOLDER;
if (InterlockedExchange(&GC_mark_mutex_state, 0 /* unlocked */) < 0) {
/* wake a waiter */
if (SetEvent(mark_mutex_event) == FALSE)
ABORT("SetEvent failed");
}
}
/* In GC_wait_for_reclaim/GC_notify_all_builder() we emulate POSIX */
/* cond_wait/cond_broadcast() primitives with WinAPI Event object */
/* (working in "manual reset" mode). This works here because */
/* GC_notify_all_builder() is always called holding lock on */
/* mark_mutex and the checked condition (GC_fl_builder_count == 0) */
/* is the only one for which broadcasting on builder_cv is performed. */
GC_INNER void GC_wait_for_reclaim(void)
{
GC_ASSERT(builder_cv != 0);
for (;;) {
GC_acquire_mark_lock();
if (GC_fl_builder_count == 0)
break;
if (ResetEvent(builder_cv) == FALSE)
ABORT("ResetEvent failed");
GC_release_mark_lock();
if (WaitForSingleObject(builder_cv, INFINITE) == WAIT_FAILED)
ABORT("WaitForSingleObject failed");
}
GC_release_mark_lock();
}
/* Clear the thread entry even if we exit with an exception. */
/* This is probably pointless, since an uncaught exception is */
/* supposed to result in the process being killed. */
# if !defined(__GNUC__) && !defined(NO_CRT)
ret = NULL; /* to suppress "might be uninitialized" compiler warning */
__try
# endif
{
ret = (void *)(word)(*start)(param);
}
# if !defined(__GNUC__) && !defined(NO_CRT)
__finally
# endif
{
GC_unregister_my_thread();
}
# ifdef DEBUG_THREADS
GC_log_printf("thread 0x%lx returned from start routine\n",
(long)GetCurrentThreadId());
# endif
return ret;
}
# if !defined(CYGWIN32) && !defined(MSWINCE) && !defined(MSWIN_XBOX1) \
&& !defined(NO_CRT)
GC_API GC_uintptr_t GC_CALL GC_beginthreadex(
void *security, unsigned stack_size,
unsigned (__stdcall *start_address)(void *),
void *arglist, unsigned initflag,
unsigned *thrdaddr)
{
if (!EXPECT(parallel_initialized, TRUE))
GC_init_parallel();
/* make sure GC is initialized (i.e. main thread is */
/* attached, tls initialized). */
# ifdef DEBUG_THREADS
GC_log_printf("About to create a thread from 0x%lx\n",
(long)GetCurrentThreadId());
# endif
if (GC_win32_dll_threads) {
return _beginthreadex(security, stack_size, start_address,
arglist, initflag, thrdaddr);
} else {
GC_uintptr_t thread_h;
thread_args *args =
(thread_args *)GC_malloc_uncollectable(sizeof(thread_args));
/* Handed off to and deallocated by child thread. */
if (NULL == args) {
/* MSDN docs say _beginthreadex() returns 0 on error and sets */
/* errno to either EAGAIN (too many threads) or EINVAL (the */
/* argument is invalid or the stack size is incorrect), so we */
/* set errno to EAGAIN on "not enough memory". */
errno = EAGAIN;
return 0;
}
/* set up thread arguments */
args -> start = (LPTHREAD_START_ROUTINE)start_address;
args -> param = arglist;
GC_dirty(args);
REACHABLE_AFTER_DIRTY(arglist);
GC_API void GC_CALL GC_set_markers_count(unsigned markers GC_ATTR_UNUSED)
{
/* The same implementation as in pthread_support.c. */
# ifdef PARALLEL_MARK
required_markers_cnt = markers < MAX_MARKERS ? markers : MAX_MARKERS;
# endif
}
#ifdef WOW64_THREAD_CONTEXT_WORKAROUND
# ifdef MSWINRT_FLAVOR
/* Available on WinRT but we have to declare it manually. */
__declspec(dllimport) HMODULE WINAPI GetModuleHandleW(LPCWSTR);
# endif
static GC_bool is_wow64_process(HMODULE hK32 GC_ATTR_UNUSED)
{
BOOL is_wow64;
# ifdef MSWINRT_FLAVOR
/* Try to use IsWow64Process2 as it handles different WoW64 cases. */
HMODULE hWow64 = GetModuleHandleW(L"api-ms-win-core-wow64-l1-1-1.dll");
# ifdef CAN_HANDLE_FORK
/* Prepare for forks if requested. */
if (GC_handle_fork) {
# ifdef CAN_CALL_ATFORK
if (pthread_atfork(fork_prepare_proc, fork_parent_proc,
fork_child_proc) == 0) {
/* Handlers successfully registered. */
GC_handle_fork = 1;
} else
# endif
/* else */ if (GC_handle_fork != -1)
ABORT("pthread_atfork failed");
}
# endif
# ifdef WOW64_THREAD_CONTEXT_WORKAROUND
/* Set isWow64 flag. */
isWow64 = is_wow64_process(hK32);
# endif
/* Add the initial thread, so we can stop it. */
sb.mem_base = GC_stackbottom;
GC_ASSERT(sb.mem_base != NULL);
# ifdef IA64
sb.reg_base = GC_register_stackbottom;
# endif
# if defined(PARALLEL_MARK)
{
char * markers_string = GETENV("GC_MARKERS");
int markers = required_markers_cnt;
if (markers_string != NULL) {
markers = atoi(markers_string);
if (markers <= 0 || markers > MAX_MARKERS) {
WARN("Too big or invalid number of mark threads: %" WARN_PRIdPTR
"; using maximum threads\n", (signed_word)markers);
markers = MAX_MARKERS;
}
} else if (0 == markers) {
/* Unless the client sets the desired number of */
/* parallel markers, it is determined based on the */
/* number of CPU cores. */
# ifdef MSWINCE
/* There is no GetProcessAffinityMask() in WinCE. */
/* GC_sysinfo is already initialized. */
markers = (int)GC_sysinfo.dwNumberOfProcessors;
# else
# ifdef _WIN64
DWORD_PTR procMask = 0;
DWORD_PTR sysMask;
# else
DWORD procMask = 0;
DWORD sysMask;
# endif
int ncpu = 0;
if (
# ifdef __cplusplus
GetProcessAffinityMask(GetCurrentProcess(), &procMask, &sysMask)
# else
/* Cast args to void* for compatibility with some old SDKs. */
GetProcessAffinityMask(GetCurrentProcess(),
(void *)&procMask, (void *)&sysMask)
# endif
&& procMask) {
do {
ncpu++;
} while ((procMask &= procMask - 1) != 0);
}
markers = ncpu;
# endif
# if defined(GC_MIN_MARKERS) && !defined(CPPCHECK)
/* This is primarily for testing on systems without getenv(). */
if (markers < GC_MIN_MARKERS)
markers = GC_MIN_MARKERS;
# endif
if (markers > MAX_MARKERS)
markers = MAX_MARKERS; /* silently limit the value */
}
available_markers_m1 = markers - 1;
}
/* After the join, thread id may have been recycled. */
t = GC_lookup_pthread(pthread_id);
result = pthread_join(pthread_id, retval);
if (0 == result) {
LOCK();
if ((t -> flags & FINISHED) != 0) {
GC_delete_gc_thread_no_free(t);
GC_INTERNAL_FREE(t);
}
UNLOCK();
}
/* Cygwin-pthreads calls CreateThread internally, but it's not easily */
/* interceptable by us..., so intercept pthread_create instead. */
GC_API int GC_pthread_create(pthread_t *new_thread,
GC_PTHREAD_CREATE_CONST pthread_attr_t *attr,
void *(*start_routine)(void *), void *arg)
{
int result;
struct start_info si;
if (!EXPECT(parallel_initialized, TRUE))
GC_init_parallel();
/* make sure GC is initialized (i.e. main thread is attached) */
GC_ASSERT(!GC_win32_dll_threads);
if (sem_init(&si.registered, GC_SEM_INIT_PSHARED, 0) != 0)
ABORT("sem_init failed");
si.start_routine = start_routine;
si.arg = arg;
si.detached = 0;
if (attr != NULL
&& pthread_attr_getdetachstate(attr, &si.detached) != 0)
ABORT("pthread_attr_getdetachstate failed");
# ifdef DEBUG_THREADS
GC_log_printf("About to create a thread from %p(0x%lx)\n",
(void *)GC_PTHREAD_PTRVAL(pthread_self()),
(long)GetCurrentThreadId());
# endif
START_MARK_THREADS();
set_need_to_lock();
result = pthread_create(new_thread, attr, GC_pthread_start, &si);
/* Wait until child has been added to the thread table. */
/* This also ensures that we hold onto the stack-allocated si */
/* until the child is done with it. */
if (0 == result) {
while (0 != sem_wait(&si.registered)) {
if (EINTR != errno) ABORT("sem_wait failed");
}
}
sem_destroy(&si.registered);
return(result);
}
GC_ASSERT(!GC_win32_dll_threads);
/* If a GC occurs before the thread is registered, that GC will */
/* ignore this thread. That's fine, since it will block trying to */
/* acquire the allocation lock, and won't yet hold interesting */
/* pointers. */
LOCK();
/* We register the thread here instead of in the parent, so that */
/* we don't need to hold the allocation lock during pthread_create. */
me = GC_register_my_thread_inner(sb, thread_id);
SET_PTHREAD_MAP_CACHE(pthread_id, thread_id);
GC_ASSERT(me != &first_thread);
me -> pthread_id = pthread_id;
if (si->detached) me -> flags |= DETACHED;
# ifdef THREAD_LOCAL_ALLOC
GC_init_thread_local(&me->tlfs);
# endif
UNLOCK();
start = si -> start_routine;
start_arg = si -> arg;
sem_post(&(si -> registered)); /* Last action on si. */
/* OK to deallocate. */
pthread_cleanup_push(GC_thread_exit_proc, (void *)me);
result = (*start)(start_arg);
me -> status = result;
GC_dirty(me);
pthread_cleanup_pop(1);
# ifdef DEBUG_THREADS
GC_log_printf("thread %p(0x%x) returned from start routine\n",
(void *)GC_PTHREAD_PTRVAL(pthread_id), (int)thread_id);
# endif
return(result);
}
LOCK();
GC_wait_for_gc_completion(FALSE);
# if defined(THREAD_LOCAL_ALLOC)
GC_ASSERT(GC_getspecific(GC_thread_key) == &me->tlfs);
GC_destroy_thread_local(&(me->tlfs));
# endif
if (me -> flags & DETACHED) {
GC_delete_thread(GetCurrentThreadId());
} else {
/* deallocate it as part of join */
me -> flags |= FINISHED;
}
# if defined(THREAD_LOCAL_ALLOC)
/* It is required to call remove_specific defined in specific.c. */
GC_remove_specific(GC_thread_key);
# endif
UNLOCK();
}
# ifndef GC_NO_PTHREAD_SIGMASK
/* pthreads-win32 does not support sigmask. */
/* So, nothing required here... */
GC_API int GC_pthread_sigmask(int how, const sigset_t *set,
sigset_t *oset)
{
return pthread_sigmask(how, set, oset);
}
# endif /* !GC_NO_PTHREAD_SIGMASK */
GC_API int GC_pthread_detach(pthread_t thread)
{
int result;
GC_thread t;
DCL_LOCK_STATE;
GC_ASSERT(!GC_win32_dll_threads);
t = GC_lookup_pthread(thread);
result = pthread_detach(thread);
if (result == 0) {
LOCK();
t -> flags |= DETACHED;
/* Here the pthread thread id may have been recycled. */
if ((t -> flags & FINISHED) != 0) {
GC_delete_gc_thread_no_free(t);
GC_INTERNAL_FREE(t);
}
UNLOCK();
}
return result;
}
#elif !defined(GC_NO_THREADS_DISCOVERY)
/* We avoid acquiring locks here, since this doesn't seem to be */
/* preemptible. This may run with an uninitialized collector, in */
/* which case we don't do much. This implies that no threads other */
/* than the main one should be created with an uninitialized */
/* collector. (The alternative of initializing the collector here */
/* seems dangerous, since DllMain is limited in what it can do.) */
/* Note that GC_use_threads_discovery should be called by the */
/* client application at start-up to activate automatic thread */
/* registration (it is the default GC behavior); */
/* to always have automatic thread registration turned on, the GC */
/* should be compiled with -D GC_DISCOVER_TASK_THREADS. */
if (!GC_win32_dll_threads && parallel_initialized) return TRUE;
switch (reason) {
case DLL_THREAD_ATTACH:
# ifdef PARALLEL_MARK
/* Don't register marker threads. */
if (GC_parallel) {
/* We could reach here only if parallel_initialized == FALSE. */
break;
}
# endif
/* FALLTHRU */
case DLL_PROCESS_ATTACH:
/* This may run with the collector uninitialized. */
thread_id = GetCurrentThreadId();
if (parallel_initialized && GC_main_thread != thread_id) {
struct GC_stack_base sb;
/* Don't lock here. */
# ifdef GC_ASSERTIONS
int sb_result =
# endif
GC_get_stack_base(&sb);
GC_ASSERT(sb_result == GC_SUCCESS);
GC_register_my_thread_inner(&sb, thread_id);
} /* o.w. we already did it during GC_thr_init, called by GC_init */
break;
case DLL_THREAD_DETACH:
/* We are hopefully running in the context of the exiting thread. */
if (GC_win32_dll_threads) {
GC_delete_thread(GetCurrentThreadId());
}
break;
case DLL_PROCESS_DETACH:
if (GC_win32_dll_threads) {
int i;
int my_max = (int)GC_get_max_thread_index();
for (i = 0; i <= my_max; ++i) {
if (AO_load(&(dll_thread_table[i].tm.in_use)))
GC_delete_gc_thread_no_free(&dll_thread_table[i]);
}
GC_deinit();
}
break;
}
return TRUE;
}
#endif /* !GC_NO_THREADS_DISCOVERY && !GC_PTHREADS */
/* Perform all initializations, including those that */
/* may require allocation. */
/* Called without allocation lock. */
/* Must be called before a second thread is created. */
GC_INNER void GC_init_parallel(void)
{
# if defined(THREAD_LOCAL_ALLOC)
GC_thread me;
DCL_LOCK_STATE;
# endif
if (parallel_initialized) return;
parallel_initialized = TRUE;
/* GC_init() calls us back, so set flag first. */
if (!GC_is_initialized) GC_init();
# if defined(CPPCHECK) && !defined(GC_NO_THREADS_DISCOVERY)
GC_noop1((word)&GC_DllMain);
# endif
if (GC_win32_dll_threads) {
set_need_to_lock();
/* Cannot intercept thread creation. Hence we don't know if */
/* other threads exist. However, client is not allowed to */
/* create other threads before collector initialization. */
/* Thus it's OK not to lock before this. */
}
/* Initialize thread local free lists if used. */
# if defined(THREAD_LOCAL_ALLOC)
LOCK();
me = GC_lookup_thread_inner(GetCurrentThreadId());
CHECK_LOOKUP_MY_THREAD(me);
GC_init_thread_local(&me->tlfs);
UNLOCK();
# endif
}
#if defined(USE_PTHREAD_LOCKS)
/* Support for pthread locking code. */
/* pthread_mutex_trylock may not win here, */
/* due to builtin support for spinning first? */
/* We must explicitly mark ptrfree and gcj free lists, since the free */
/* list links wouldn't otherwise be found. We also set them in the */
/* normal free lists, since that involves touching less memory than if */
/* we scanned them normally. */
GC_INNER void GC_mark_thread_local_free_lists(void)
{
int i;
GC_thread p;
for (i = 0; i < THREAD_TABLE_SZ; ++i) {
for (p = GC_threads[i]; 0 != p; p = p -> tm.next) {
if (!KNOWN_FINISHED(p)) {
# ifdef DEBUG_THREADS
GC_log_printf("Marking thread locals for 0x%x\n", (int)p -> id);
# endif
GC_mark_thread_local_fls_for(&(p->tlfs));
}
}
}
}
# if defined(GC_ASSERTIONS)
/* Check that all thread-local free-lists are completely marked. */
/* also check that thread-specific-data structures are marked. */
void GC_check_tls(void)
{
int i;
GC_thread p;
for (i = 0; i < THREAD_TABLE_SZ; ++i) {
for (p = GC_threads[i]; 0 != p; p = p -> tm.next) {
if (!KNOWN_FINISHED(p))
GC_check_tls_for(&(p->tlfs));
}
}
# if defined(USE_CUSTOM_SPECIFIC)
if (GC_thread_key != 0)
GC_check_tsd_marks(GC_thread_key);
# endif
}
# endif /* GC_ASSERTIONS */