Annotation of OpenXM_contrib2/asir2000/gc/include/private/gc_locks.h, Revision 1.3
1.1 noro 1: /*
2: * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
3: * Copyright (c) 1991-1994 by Xerox Corporation. All rights reserved.
4: * Copyright (c) 1996-1999 by Silicon Graphics. All rights reserved.
5: * Copyright (c) 1999 by Hewlett-Packard Company. All rights reserved.
6: *
7: *
8: * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
9: * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
10: *
11: * Permission is hereby granted to use or copy this program
12: * for any purpose, provided the above notices are retained on all copies.
13: * Permission to modify the code and to distribute modified code is granted,
14: * provided the above notices are retained, and a notice that the code was
15: * modified is included with the above copyright notice.
16: */
17:
18: #ifndef GC_LOCKS_H
19: #define GC_LOCKS_H
20:
21: /*
22: * Mutual exclusion between allocator/collector routines.
23: * Needed if there is more than one allocator thread.
24: * FASTLOCK() is assumed to try to acquire the lock in a cheap and
25: * dirty way that is acceptable for a few instructions, e.g. by
26: * inhibiting preemption. This is assumed to have succeeded only
27: * if a subsequent call to FASTLOCK_SUCCEEDED() returns TRUE.
28: * FASTUNLOCK() is called whether or not FASTLOCK_SUCCEEDED().
29: * If signals cannot be tolerated with the FASTLOCK held, then
30: * FASTLOCK should disable signals. The code executed under
31: * FASTLOCK is otherwise immune to interruption, provided it is
32: * not restarted.
33: * DCL_LOCK_STATE declares any local variables needed by LOCK and UNLOCK
34: * and/or DISABLE_SIGNALS and ENABLE_SIGNALS and/or FASTLOCK.
35: * (There is currently no equivalent for FASTLOCK.)
36: *
37: * In the PARALLEL_MARK case, we also need to define a number of
38: * other inline finctions here:
39: * GC_bool GC_compare_and_exchange( volatile GC_word *addr,
40: * GC_word old, GC_word new )
41: * GC_word GC_atomic_add( volatile GC_word *addr, GC_word how_much )
42: * void GC_memory_barrier( )
43: *
44: */
45: # ifdef THREADS
1.3 ! noro 46: void GC_noop1 GC_PROTO((word));
1.1 noro 47: # ifdef PCR_OBSOLETE /* Faster, but broken with multiple lwp's */
48: # include "th/PCR_Th.h"
49: # include "th/PCR_ThCrSec.h"
50: extern struct PCR_Th_MLRep GC_allocate_ml;
51: # define DCL_LOCK_STATE PCR_sigset_t GC_old_sig_mask
52: # define LOCK() PCR_Th_ML_Acquire(&GC_allocate_ml)
53: # define UNLOCK() PCR_Th_ML_Release(&GC_allocate_ml)
54: # define UNLOCK() PCR_Th_ML_Release(&GC_allocate_ml)
55: # define FASTLOCK() PCR_ThCrSec_EnterSys()
56: /* Here we cheat (a lot): */
57: # define FASTLOCK_SUCCEEDED() (*(int *)(&GC_allocate_ml) == 0)
58: /* TRUE if nobody currently holds the lock */
59: # define FASTUNLOCK() PCR_ThCrSec_ExitSys()
60: # endif
61: # ifdef PCR
62: # include <base/PCR_Base.h>
63: # include <th/PCR_Th.h>
64: extern PCR_Th_ML GC_allocate_ml;
65: # define DCL_LOCK_STATE \
66: PCR_ERes GC_fastLockRes; PCR_sigset_t GC_old_sig_mask
67: # define LOCK() PCR_Th_ML_Acquire(&GC_allocate_ml)
68: # define UNLOCK() PCR_Th_ML_Release(&GC_allocate_ml)
69: # define FASTLOCK() (GC_fastLockRes = PCR_Th_ML_Try(&GC_allocate_ml))
70: # define FASTLOCK_SUCCEEDED() (GC_fastLockRes == PCR_ERes_okay)
71: # define FASTUNLOCK() {\
72: if( FASTLOCK_SUCCEEDED() ) PCR_Th_ML_Release(&GC_allocate_ml); }
73: # endif
74: # ifdef SRC_M3
75: extern GC_word RT0u__inCritical;
76: # define LOCK() RT0u__inCritical++
77: # define UNLOCK() RT0u__inCritical--
78: # endif
1.3 ! noro 79: # ifdef GC_SOLARIS_THREADS
1.1 noro 80: # include <thread.h>
81: # include <signal.h>
82: extern mutex_t GC_allocate_ml;
83: # define LOCK() mutex_lock(&GC_allocate_ml);
84: # define UNLOCK() mutex_unlock(&GC_allocate_ml);
85: # endif
86:
87: /* Try to define GC_TEST_AND_SET and a matching GC_CLEAR for spin lock */
88: /* acquisition and release. We need this for correct operation of the */
89: /* incremental GC. */
90: # ifdef __GNUC__
91: # if defined(I386)
92: inline static int GC_test_and_set(volatile unsigned int *addr) {
93: int oldval;
94: /* Note: the "xchg" instruction does not need a "lock" prefix */
95: __asm__ __volatile__("xchgl %0, %1"
96: : "=r"(oldval), "=m"(*(addr))
97: : "0"(1), "m"(*(addr)) : "memory");
98: return oldval;
99: }
100: # define GC_TEST_AND_SET_DEFINED
101: # endif
102: # if defined(IA64)
103: inline static int GC_test_and_set(volatile unsigned int *addr) {
104: long oldval, n = 1;
105: __asm__ __volatile__("xchg4 %0=%1,%2"
106: : "=r"(oldval), "=m"(*addr)
107: : "r"(n), "1"(*addr) : "memory");
108: return oldval;
109: }
110: # define GC_TEST_AND_SET_DEFINED
111: /* Should this handle post-increment addressing?? */
112: inline static void GC_clear(volatile unsigned int *addr) {
113: __asm__ __volatile__("st4.rel %0=r0" : "=m" (*addr) : : "memory");
114: }
115: # define GC_CLEAR_DEFINED
116: # endif
117: # ifdef SPARC
118: inline static int GC_test_and_set(volatile unsigned int *addr) {
119: int oldval;
120:
121: __asm__ __volatile__("ldstub %1,%0"
122: : "=r"(oldval), "=m"(*addr)
123: : "m"(*addr) : "memory");
124: return oldval;
125: }
126: # define GC_TEST_AND_SET_DEFINED
127: # endif
128: # ifdef M68K
129: /* Contributed by Tony Mantler. I'm not sure how well it was */
130: /* tested. */
131: inline static int GC_test_and_set(volatile unsigned int *addr) {
132: char oldval; /* this must be no longer than 8 bits */
133:
134: /* The return value is semi-phony. */
135: /* 'tas' sets bit 7 while the return */
136: /* value pretends bit 0 was set */
137: __asm__ __volatile__(
138: "tas %1@; sne %0; negb %0"
139: : "=d" (oldval)
140: : "a" (addr) : "memory");
141: return oldval;
142: }
143: # define GC_TEST_AND_SET_DEFINED
144: # endif
145: # if defined(POWERPC)
146: inline static int GC_test_and_set(volatile unsigned int *addr) {
147: int oldval;
1.3 ! noro 148: int temp = 1; /* locked value */
1.1 noro 149:
150: __asm__ __volatile__(
1.3 ! noro 151: "1:\tlwarx %0,0,%3\n" /* load and reserve */
! 152: "\tcmpwi %0, 0\n" /* if load is */
! 153: "\tbne 2f\n" /* non-zero, return already set */
! 154: "\tstwcx. %2,0,%1\n" /* else store conditional */
! 155: "\tbne- 1b\n" /* retry if lost reservation */
! 156: "2:\t\n" /* oldval is zero if we set */
1.1 noro 157: : "=&r"(oldval), "=p"(addr)
158: : "r"(temp), "1"(addr)
159: : "memory");
160: return (int)oldval;
161: }
162: # define GC_TEST_AND_SET_DEFINED
163: inline static void GC_clear(volatile unsigned int *addr) {
1.3 ! noro 164: __asm__ __volatile__("eieio" : : : "memory");
1.1 noro 165: *(addr) = 0;
166: }
167: # define GC_CLEAR_DEFINED
168: # endif
169: # if defined(ALPHA)
170: inline static int GC_test_and_set(volatile unsigned int * addr)
171: {
172: unsigned long oldvalue;
173: unsigned long temp;
174:
175: __asm__ __volatile__(
176: "1: ldl_l %0,%1\n"
177: " and %0,%3,%2\n"
178: " bne %2,2f\n"
179: " xor %0,%3,%0\n"
180: " stl_c %0,%1\n"
181: " beq %0,3f\n"
182: " mb\n"
183: "2:\n"
184: ".section .text2,\"ax\"\n"
185: "3: br 1b\n"
186: ".previous"
187: :"=&r" (temp), "=m" (*addr), "=&r" (oldvalue)
188: :"Ir" (1), "m" (*addr)
189: :"memory");
190:
191: return oldvalue;
192: }
193: # define GC_TEST_AND_SET_DEFINED
194: /* Should probably also define GC_clear, since it needs */
195: /* a memory barrier ?? */
196: # endif /* ALPHA */
197: # ifdef ARM32
198: inline static int GC_test_and_set(volatile unsigned int *addr) {
199: int oldval;
200: /* SWP on ARM is very similar to XCHG on x86. Doesn't lock the
201: * bus because there are no SMP ARM machines. If/when there are,
202: * this code will likely need to be updated. */
203: /* See linuxthreads/sysdeps/arm/pt-machine.h in glibc-2.1 */
204: __asm__ __volatile__("swp %0, %1, [%2]"
205: : "=r"(oldval)
206: : "r"(1), "r"(addr)
207: : "memory");
208: return oldval;
209: }
210: # define GC_TEST_AND_SET_DEFINED
211: # endif /* ARM32 */
212: # endif /* __GNUC__ */
213: # if (defined(ALPHA) && !defined(__GNUC__))
214: # define GC_test_and_set(addr) __cxx_test_and_set_atomic(addr, 1)
215: # define GC_TEST_AND_SET_DEFINED
216: # endif
217: # if defined(MSWIN32)
218: # define GC_test_and_set(addr) InterlockedExchange((LPLONG)addr,1)
219: # define GC_TEST_AND_SET_DEFINED
220: # endif
221: # ifdef MIPS
1.3 ! noro 222: # ifdef LINUX
! 223: # include <sys/tas.h>
! 224: # define GC_test_and_set(addr) _test_and_set((int *) addr,1)
! 225: # define GC_TEST_AND_SET_DEFINED
! 226: # elif __mips < 3 || !(defined (_ABIN32) || defined(_ABI64)) \
1.1 noro 227: || !defined(_COMPILER_VERSION) || _COMPILER_VERSION < 700
1.3 ! noro 228: # ifdef __GNUC__
! 229: # define GC_test_and_set(addr) _test_and_set(addr,1)
! 230: # else
! 231: # define GC_test_and_set(addr) test_and_set(addr,1)
! 232: # endif
1.1 noro 233: # else
1.3 ! noro 234: # define GC_test_and_set(addr) __test_and_set(addr,1)
1.1 noro 235: # define GC_clear(addr) __lock_release(addr);
236: # define GC_CLEAR_DEFINED
237: # endif
238: # define GC_TEST_AND_SET_DEFINED
239: # endif /* MIPS */
240: # if 0 /* defined(HP_PA) */
241: /* The official recommendation seems to be to not use ldcw from */
242: /* user mode. Since multithreaded incremental collection doesn't */
243: /* work anyway on HP_PA, this shouldn't be a major loss. */
244:
245: /* "set" means 0 and "clear" means 1 here. */
246: # define GC_test_and_set(addr) !GC_test_and_clear(addr);
247: # define GC_TEST_AND_SET_DEFINED
1.3 ! noro 248: # define GC_clear(addr) GC_noop1((word)(addr)); *(volatile unsigned int *)addr = 1;
1.1 noro 249: /* The above needs a memory barrier! */
250: # define GC_CLEAR_DEFINED
251: # endif
252: # if defined(GC_TEST_AND_SET_DEFINED) && !defined(GC_CLEAR_DEFINED)
253: # ifdef __GNUC__
254: inline static void GC_clear(volatile unsigned int *addr) {
255: /* Try to discourage gcc from moving anything past this. */
256: __asm__ __volatile__(" " : : : "memory");
257: *(addr) = 0;
258: }
259: # else
260: /* The function call in the following should prevent the */
261: /* compiler from moving assignments to below the UNLOCK. */
262: # define GC_clear(addr) GC_noop1((word)(addr)); \
263: *((volatile unsigned int *)(addr)) = 0;
264: # endif
265: # define GC_CLEAR_DEFINED
266: # endif /* !GC_CLEAR_DEFINED */
267:
268: # if !defined(GC_TEST_AND_SET_DEFINED)
269: # define USE_PTHREAD_LOCKS
270: # endif
271:
1.3 ! noro 272: # if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
! 273: && !defined(GC_IRIX_THREADS) && !defined(GC_WIN32_THREADS)
1.1 noro 274: # define NO_THREAD (pthread_t)(-1)
275: # include <pthread.h>
276: # if defined(PARALLEL_MARK)
277: /* We need compare-and-swap to update mark bits, where it's */
278: /* performance critical. If USE_MARK_BYTES is defined, it is */
279: /* no longer needed for this purpose. However we use it in */
280: /* either case to implement atomic fetch-and-add, though that's */
281: /* less performance critical, and could perhaps be done with */
282: /* a lock. */
283: # if defined(GENERIC_COMPARE_AND_SWAP)
284: /* Probably not useful, except for debugging. */
285: /* We do use GENERIC_COMPARE_AND_SWAP on PA_RISC, but we */
286: /* minimize its use. */
287: extern pthread_mutex_t GC_compare_and_swap_lock;
288:
289: /* Note that if GC_word updates are not atomic, a concurrent */
290: /* reader should acquire GC_compare_and_swap_lock. On */
291: /* currently supported platforms, such updates are atomic. */
292: extern GC_bool GC_compare_and_exchange(volatile GC_word *addr,
293: GC_word old, GC_word new_val);
294: # endif /* GENERIC_COMPARE_AND_SWAP */
295: # if defined(I386)
296: # if !defined(GENERIC_COMPARE_AND_SWAP)
297: /* Returns TRUE if the comparison succeeded. */
298: inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
299: GC_word old,
300: GC_word new_val)
301: {
302: char result;
303: __asm__ __volatile__("lock; cmpxchgl %2, %0; setz %1"
304: : "=m"(*(addr)), "=r"(result)
305: : "r" (new_val), "0"(*(addr)), "a"(old) : "memory");
306: return (GC_bool) result;
307: }
308: # endif /* !GENERIC_COMPARE_AND_SWAP */
309: inline static void GC_memory_write_barrier()
310: {
311: /* We believe the processor ensures at least processor */
312: /* consistent ordering. Thus a compiler barrier */
313: /* should suffice. */
314: __asm__ __volatile__("" : : : "memory");
315: }
316: # endif /* I386 */
317: # if defined(IA64)
318: # if !defined(GENERIC_COMPARE_AND_SWAP)
319: inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
320: GC_word old, GC_word new_val)
321: {
322: unsigned long oldval;
323: __asm__ __volatile__("mov ar.ccv=%4 ;; cmpxchg8.rel %0=%1,%2,ar.ccv"
324: : "=r"(oldval), "=m"(*addr)
325: : "r"(new_val), "1"(*addr), "r"(old) : "memory");
326: return (oldval == old);
327: }
328: # endif /* !GENERIC_COMPARE_AND_SWAP */
329: # if 0
330: /* Shouldn't be needed; we use volatile stores instead. */
331: inline static void GC_memory_write_barrier()
332: {
333: __asm__ __volatile__("mf" : : : "memory");
334: }
335: # endif /* 0 */
336: # endif /* IA64 */
337: # if !defined(GENERIC_COMPARE_AND_SWAP)
338: /* Returns the original value of *addr. */
339: inline static GC_word GC_atomic_add(volatile GC_word *addr,
340: GC_word how_much)
341: {
342: GC_word old;
343: do {
344: old = *addr;
345: } while (!GC_compare_and_exchange(addr, old, old+how_much));
346: return old;
347: }
348: # else /* GENERIC_COMPARE_AND_SWAP */
349: /* So long as a GC_word can be atomically updated, it should */
350: /* be OK to read *addr without a lock. */
351: extern GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much);
352: # endif /* GENERIC_COMPARE_AND_SWAP */
353:
354: # endif /* PARALLEL_MARK */
355:
356: # if !defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_LOCKS)
357: /* In the THREAD_LOCAL_ALLOC case, the allocation lock tends to */
358: /* be held for long periods, if it is held at all. Thus spinning */
359: /* and sleeping for fixed periods are likely to result in */
360: /* significant wasted time. We thus rely mostly on queued locks. */
361: # define USE_SPIN_LOCK
362: extern volatile unsigned int GC_allocate_lock;
363: extern void GC_lock(void);
364: /* Allocation lock holder. Only set if acquired by client through */
365: /* GC_call_with_alloc_lock. */
366: # ifdef GC_ASSERTIONS
367: # define LOCK() \
368: { if (GC_test_and_set(&GC_allocate_lock)) GC_lock(); \
369: SET_LOCK_HOLDER(); }
370: # define UNLOCK() \
371: { GC_ASSERT(I_HOLD_LOCK()); UNSET_LOCK_HOLDER(); \
372: GC_clear(&GC_allocate_lock); }
373: # else
374: # define LOCK() \
375: { if (GC_test_and_set(&GC_allocate_lock)) GC_lock(); }
376: # define UNLOCK() \
377: GC_clear(&GC_allocate_lock)
378: # endif /* !GC_ASSERTIONS */
379: # if 0
380: /* Another alternative for OSF1 might be: */
381: # include <sys/mman.h>
382: extern msemaphore GC_allocate_semaphore;
383: # define LOCK() { if (msem_lock(&GC_allocate_semaphore, MSEM_IF_NOWAIT) \
384: != 0) GC_lock(); else GC_allocate_lock = 1; }
385: /* The following is INCORRECT, since the memory model is too weak. */
386: /* Is this true? Presumably msem_unlock has the right semantics? */
387: /* - HB */
388: # define UNLOCK() { GC_allocate_lock = 0; \
389: msem_unlock(&GC_allocate_semaphore, 0); }
390: # endif /* 0 */
391: # else /* THREAD_LOCAL_ALLOC || USE_PTHREAD_LOCKS */
392: # ifndef USE_PTHREAD_LOCKS
393: # define USE_PTHREAD_LOCKS
394: # endif
395: # endif /* THREAD_LOCAL_ALLOC */
396: # ifdef USE_PTHREAD_LOCKS
397: # include <pthread.h>
398: extern pthread_mutex_t GC_allocate_ml;
399: # ifdef GC_ASSERTIONS
400: # define LOCK() \
401: { GC_lock(); \
402: SET_LOCK_HOLDER(); }
403: # define UNLOCK() \
404: { GC_ASSERT(I_HOLD_LOCK()); UNSET_LOCK_HOLDER(); \
405: pthread_mutex_unlock(&GC_allocate_ml); }
406: # else /* !GC_ASSERTIONS */
407: # define LOCK() \
408: { if (0 != pthread_mutex_trylock(&GC_allocate_ml)) GC_lock(); }
409: # define UNLOCK() pthread_mutex_unlock(&GC_allocate_ml)
410: # endif /* !GC_ASSERTIONS */
411: # endif /* USE_PTHREAD_LOCKS */
412: # define SET_LOCK_HOLDER() GC_lock_holder = pthread_self()
413: # define UNSET_LOCK_HOLDER() GC_lock_holder = NO_THREAD
414: # define I_HOLD_LOCK() (pthread_equal(GC_lock_holder, pthread_self()))
415: extern VOLATILE GC_bool GC_collecting;
416: # define ENTER_GC() GC_collecting = 1;
417: # define EXIT_GC() GC_collecting = 0;
418: extern void GC_lock(void);
419: extern pthread_t GC_lock_holder;
420: # ifdef GC_ASSERTIONS
421: extern pthread_t GC_mark_lock_holder;
422: # endif
1.3 ! noro 423: # endif /* GC_PTHREADS with linux_threads.c implementation */
! 424: # if defined(GC_IRIX_THREADS)
1.1 noro 425: # include <pthread.h>
426: /* This probably should never be included, but I can't test */
427: /* on Irix anymore. */
428: # include <mutex.h>
429:
430: extern unsigned long GC_allocate_lock;
431: /* This is not a mutex because mutexes that obey the (optional) */
432: /* POSIX scheduling rules are subject to convoys in high contention */
433: /* applications. This is basically a spin lock. */
434: extern pthread_t GC_lock_holder;
435: extern void GC_lock(void);
436: /* Allocation lock holder. Only set if acquired by client through */
437: /* GC_call_with_alloc_lock. */
438: # define SET_LOCK_HOLDER() GC_lock_holder = pthread_self()
439: # define NO_THREAD (pthread_t)(-1)
440: # define UNSET_LOCK_HOLDER() GC_lock_holder = NO_THREAD
441: # define I_HOLD_LOCK() (pthread_equal(GC_lock_holder, pthread_self()))
1.3 ! noro 442: # define LOCK() { if (GC_test_and_set(&GC_allocate_lock)) GC_lock(); }
1.1 noro 443: # define UNLOCK() GC_clear(&GC_allocate_lock);
444: extern VOLATILE GC_bool GC_collecting;
445: # define ENTER_GC() \
446: { \
447: GC_collecting = 1; \
448: }
449: # define EXIT_GC() GC_collecting = 0;
1.3 ! noro 450: # endif /* GC_IRIX_THREADS */
! 451: # if defined(GC_WIN32_THREADS)
! 452: # if defined(GC_PTHREADS)
! 453: # include <pthread.h>
! 454: extern pthread_mutex_t GC_allocate_ml;
! 455: # define LOCK() pthread_mutex_lock(&GC_allocate_ml)
! 456: # define UNLOCK() pthread_mutex_unlock(&GC_allocate_ml)
! 457: # else
! 458: # include <windows.h>
! 459: GC_API CRITICAL_SECTION GC_allocate_ml;
! 460: # define LOCK() EnterCriticalSection(&GC_allocate_ml);
! 461: # define UNLOCK() LeaveCriticalSection(&GC_allocate_ml);
! 462: # endif
1.1 noro 463: # endif
464: # ifndef SET_LOCK_HOLDER
465: # define SET_LOCK_HOLDER()
466: # define UNSET_LOCK_HOLDER()
467: # define I_HOLD_LOCK() FALSE
468: /* Used on platforms were locks can be reacquired, */
469: /* so it doesn't matter if we lie. */
470: # endif
471: # else /* !THREADS */
472: # define LOCK()
473: # define UNLOCK()
474: # endif /* !THREADS */
475: # ifndef SET_LOCK_HOLDER
476: # define SET_LOCK_HOLDER()
477: # define UNSET_LOCK_HOLDER()
478: # define I_HOLD_LOCK() FALSE
479: /* Used on platforms were locks can be reacquired, */
480: /* so it doesn't matter if we lie. */
481: # endif
482: # ifndef ENTER_GC
483: # define ENTER_GC()
484: # define EXIT_GC()
485: # endif
486:
487: # ifndef DCL_LOCK_STATE
488: # define DCL_LOCK_STATE
489: # endif
490: # ifndef FASTLOCK
491: # define FASTLOCK() LOCK()
492: # define FASTLOCK_SUCCEEDED() TRUE
493: # define FASTUNLOCK() UNLOCK()
494: # endif
495:
496: #endif /* GC_LOCKS_H */
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>