Age Owner TLA Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * s_lock.h
4 : * Implementation of spinlocks.
5 : *
6 : * NOTE: none of the macros in this file are intended to be called directly.
7 : * Call them through the macros in spin.h.
8 : *
9 : * The following hardware-dependent macros must be provided for each
10 : * supported platform:
11 : *
12 : * void S_INIT_LOCK(slock_t *lock)
13 : * Initialize a spinlock (to the unlocked state).
14 : *
15 : * int S_LOCK(slock_t *lock)
16 : * Acquire a spinlock, waiting if necessary.
17 : * Time out and abort() if unable to acquire the lock in a
18 : * "reasonable" amount of time --- typically ~ 1 minute.
19 : * Should return number of "delays"; see s_lock.c
20 : *
21 : * void S_UNLOCK(slock_t *lock)
22 : * Unlock a previously acquired lock.
23 : *
24 : * void SPIN_DELAY(void)
25 : * Delay operation to occur inside spinlock wait loop.
26 : *
27 : * Note to implementors: there are default implementations for all these
28 : * macros at the bottom of the file. Check if your platform can use
29 : * these or needs to override them.
30 : *
31 : * Usually, S_LOCK() is implemented in terms of even lower-level macros
32 : * TAS() and TAS_SPIN():
33 : *
34 : * int TAS(slock_t *lock)
35 : * Atomic test-and-set instruction. Attempt to acquire the lock,
36 : * but do *not* wait. Returns 0 if successful, nonzero if unable
37 : * to acquire the lock.
38 : *
39 : * int TAS_SPIN(slock_t *lock)
40 : * Like TAS(), but this version is used when waiting for a lock
41 : * previously found to be contended. By default, this is the
42 : * same as TAS(), but on some architectures it's better to poll a
43 : * contended lock using an unlocked instruction and retry the
44 : * atomic test-and-set only when it appears free.
45 : *
46 : * TAS() and TAS_SPIN() are NOT part of the API, and should never be called
47 : * directly. A platform must provide either S_LOCK() or a TAS() for the
48 : * default S_LOCK() to be built on. Currently, all supported platforms do
49 : * the latter, so that is probably the best place to start if adding a new
50 : * one.
51 : *
52 : * It is the responsibility of these macros to make sure that the compiler
53 : * does not re-order accesses to shared memory to precede the actual lock
54 : * acquisition, or follow the lock release.
55 : *
56 : * On platforms with weak memory ordering, the TAS(), TAS_SPIN(), and
57 : * S_UNLOCK() macros must further include hardware-level memory fence
58 : * instructions to prevent similar re-ordering at the hardware level.
59 : * TAS() and TAS_SPIN() must guarantee that loads and stores issued after
60 : * the macro are not executed until the lock has been obtained. Conversely,
61 : * S_UNLOCK() must guarantee that loads and stores issued before the macro
62 : * have been executed before the lock is released.
63 : *
64 : * On most supported platforms, TAS() uses a tas() function written
65 : * in assembly language to execute a hardware atomic-test-and-set
66 : * instruction. Equivalent compiler intrinsics are another popular option.
67 : *
68 : *
69 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
70 : * Portions Copyright (c) 1994, Regents of the University of California
71 : *
72 : * src/include/storage/s_lock.h
73 : *
74 : *-------------------------------------------------------------------------
75 : */
76 : #ifndef S_LOCK_H
77 : #define S_LOCK_H
78 :
79 : #ifdef FRONTEND
80 : #error "s_lock.h may not be included from frontend code"
81 : #endif
82 :
83 : #if defined(__GNUC__) || defined(__INTEL_COMPILER)
84 : /*************************************************************************
85 : * All the gcc inlines
86 : * Gcc consistently defines the CPU as __cpu__.
87 : * Other compilers use __cpu or __cpu__ so we test for both in those cases.
88 : */
89 :
90 : /*----------
91 : * Standard gcc asm format (assuming "volatile slock_t *lock"):
92 :
93 : __asm__ __volatile__(
94 : " instruction \n"
95 : " instruction \n"
96 : " instruction \n"
97 : : "=r"(_res), "+m"(*lock) // return register, in/out lock value
98 : : "r"(lock) // lock pointer, in input register
99 : : "memory", "cc"); // show clobbered registers here
100 :
101 : * The output-operands list (after first colon) should always include
102 : * "+m"(*lock), whether or not the asm code actually refers to this
103 : * operand directly. This ensures that gcc believes the value in the
104 : * lock variable is used and set by the asm code. Also, the clobbers
105 : * list (after third colon) should always include "memory"; this prevents
106 : * gcc from thinking it can cache the values of shared-memory fields
107 : * across the asm code. Add "cc" if your asm code changes the condition
108 : * code register, and also list any temp registers the code uses.
109 : *
110 : * If you need branch target labels within the asm block, include "%="
111 : * in the label names to make them distinct across multiple asm blocks
112 : * within a source file.
113 : *----------
114 : */
115 :
116 :
117 : #ifdef __i386__ /* 32-bit i386 */
118 :
119 : typedef unsigned char slock_t;
120 :
121 : #define TAS(lock) tas(lock)
122 :
123 : static inline int
124 : tas(volatile slock_t *lock)
125 : {
126 : slock_t _res = 1;
127 :
128 : /*
129 : * Use a non-locking test before asserting the bus lock. Note that the
130 : * extra test appears to be a small loss on some x86 platforms and a small
131 : * win on others; it's by no means clear that we should keep it.
132 : *
133 : * When this was last tested, we didn't have separate TAS() and TAS_SPIN()
134 : * macros. Nowadays it probably would be better to do a non-locking test
135 : * in TAS_SPIN() but not in TAS(), like on x86_64, but no-one's done the
136 : * testing to verify that. Without some empirical evidence, better to
137 : * leave it alone.
138 : */
139 : __asm__ __volatile__(
140 : " cmpb $0,%1 \n"
141 : " jne TAS%=_out \n"
142 : " lock \n"
143 : " xchgb %0,%1 \n"
144 : "TAS%=_out: \n"
145 : : "+q"(_res), "+m"(*lock)
146 : : /* no inputs */
147 : : "memory", "cc");
148 : return (int) _res;
149 : }
150 :
151 : #define SPIN_DELAY() spin_delay()
152 :
153 : static inline void
154 : spin_delay(void)
155 : {
156 : /*
157 : * This sequence is equivalent to the PAUSE instruction ("rep" is
158 : * ignored by old IA32 processors if the following instruction is
159 : * not a string operation); the IA-32 Architecture Software
160 : * Developer's Manual, Vol. 3, Section 7.7.2 describes why using
161 : * PAUSE in the inner loop of a spin lock is necessary for good
162 : * performance:
163 : *
164 : * The PAUSE instruction improves the performance of IA-32
165 : * processors supporting Hyper-Threading Technology when
166 : * executing spin-wait loops and other routines where one
167 : * thread is accessing a shared lock or semaphore in a tight
168 : * polling loop. When executing a spin-wait loop, the
169 : * processor can suffer a severe performance penalty when
170 : * exiting the loop because it detects a possible memory order
171 : * violation and flushes the core processor's pipeline. The
172 : * PAUSE instruction provides a hint to the processor that the
173 : * code sequence is a spin-wait loop. The processor uses this
174 : * hint to avoid the memory order violation and prevent the
175 : * pipeline flush. In addition, the PAUSE instruction
176 : * de-pipelines the spin-wait loop to prevent it from
177 : * consuming execution resources excessively.
178 : */
179 : __asm__ __volatile__(
180 : " rep; nop \n");
181 : }
182 :
183 : #endif /* __i386__ */
184 :
185 :
186 : #ifdef __x86_64__ /* AMD Opteron, Intel EM64T */
187 :
188 : typedef unsigned char slock_t;
189 :
190 : #define TAS(lock) tas(lock)
191 :
192 : /*
193 : * On Intel EM64T, it's a win to use a non-locking test before the xchg proper,
194 : * but only when spinning.
195 : *
196 : * See also Implementing Scalable Atomic Locks for Multi-Core Intel(tm) EM64T
197 : * and IA32, by Michael Chynoweth and Mary R. Lee. As of this writing, it is
198 : * available at:
199 : * http://software.intel.com/en-us/articles/implementing-scalable-atomic-locks-for-multi-core-intel-em64t-and-ia32-architectures
200 : */
201 : #define TAS_SPIN(lock) (*(lock) ? 1 : TAS(lock))
202 :
203 : static inline int
7671 tgl@sss.pgh.pa.us 204 CBC 63254000 : tas(volatile slock_t *lock)
205 : {
1479 andres@anarazel.de 206 63254000 : slock_t _res = 1;
207 :
7671 tgl@sss.pgh.pa.us 208 63254000 : __asm__ __volatile__(
209 : " lock \n"
210 : " xchgb %0,%1 \n"
211 : : "+q"(_res), "+m"(*lock)
212 : : /* no inputs */
213 : : "memory", "cc");
214 63254000 : return (int) _res;
215 : }
216 :
217 : #define SPIN_DELAY() spin_delay()
218 :
219 : static inline void
220 761744 : spin_delay(void)
221 : {
222 : /*
223 : * Adding a PAUSE in the spin delay loop is demonstrably a no-op on
224 : * Opteron, but it may be of some use on EM64T, so we keep it.
225 : */
226 761744 : __asm__ __volatile__(
227 : " rep; nop \n");
228 761744 : }
229 :
230 : #endif /* __x86_64__ */
231 :
232 :
233 : /*
234 : * On ARM and ARM64, we use __sync_lock_test_and_set(int *, int) if available.
235 : *
236 : * We use the int-width variant of the builtin because it works on more chips
237 : * than other widths.
238 : */
239 : #if defined(__arm__) || defined(__aarch64__)
240 : #ifdef HAVE_GCC__SYNC_INT32_TAS
241 :
242 : #define TAS(lock) tas(lock)
243 :
244 : typedef int slock_t;
245 :
246 : static inline int
247 : tas(volatile slock_t *lock)
248 : {
249 : return __sync_lock_test_and_set(lock, 1);
250 : }
251 :
252 : #define S_UNLOCK(lock) __sync_lock_release(lock)
253 :
254 : #if defined(__aarch64__)
255 :
256 : /*
257 : * On ARM64, it's a win to use a non-locking test before the TAS proper. It
258 : * may be a win on 32-bit ARM, too, but nobody's tested it yet.
259 : */
260 : #define TAS_SPIN(lock) (*(lock) ? 1 : TAS(lock))
261 :
262 : #define SPIN_DELAY() spin_delay()
263 :
264 : static inline void
265 : spin_delay(void)
266 : {
267 : /*
268 : * Using an ISB instruction to delay in spinlock loops appears beneficial
269 : * on high-core-count ARM64 processors. It seems mostly a wash for smaller
270 : * gear, and ISB doesn't exist at all on pre-v7 ARM chips.
271 : */
272 : __asm__ __volatile__(
273 : " isb; \n");
274 : }
275 :
276 : #endif /* __aarch64__ */
277 : #endif /* HAVE_GCC__SYNC_INT32_TAS */
278 : #endif /* __arm__ || __aarch64__ */
279 :
280 :
281 : /* S/390 and S/390x Linux (32- and 64-bit zSeries) */
282 : #if defined(__s390__) || defined(__s390x__)
283 :
284 : typedef unsigned int slock_t;
285 :
286 : #define TAS(lock) tas(lock)
287 :
288 : static inline int
289 : tas(volatile slock_t *lock)
290 : {
291 : int _res = 0;
292 :
293 : __asm__ __volatile__(
294 : " cs %0,%3,0(%2) \n"
295 : : "+d"(_res), "+m"(*lock)
296 : : "a"(lock), "d"(1)
297 : : "memory", "cc");
298 : return _res;
299 : }
300 :
301 : #endif /* __s390__ || __s390x__ */
302 :
303 :
304 : #if defined(__sparc__) /* Sparc */
305 : /*
306 : * Solaris has always run sparc processors in TSO (total store) mode, but
307 : * linux didn't use to and the *BSDs still don't. So, be careful about
308 : * acquire/release semantics. The CPU will treat superfluous members as
309 : * NOPs, so it's just code space.
310 : */
311 :
312 : typedef unsigned char slock_t;
313 :
314 : #define TAS(lock) tas(lock)
315 :
316 : static inline int
317 : tas(volatile slock_t *lock)
318 : {
319 : slock_t _res;
320 :
321 : /*
322 : * "cas" would be better than "ldstub", but it is only present on
323 : * sparcv8plus and later, while some platforms still support sparcv7 or
324 : * sparcv8. Also, "cas" requires that the system be running in TSO mode.
325 : */
326 : __asm__ __volatile__(
327 : " ldstub [%2], %0 \n"
328 : : "=r"(_res), "+m"(*lock)
329 : : "r"(lock)
330 : : "memory");
331 : #if defined(__sparcv7) || defined(__sparc_v7__)
332 : /*
333 : * No stbar or membar available, luckily no actually produced hardware
334 : * requires a barrier.
335 : */
336 : #elif defined(__sparcv8) || defined(__sparc_v8__)
337 : /* stbar is available (and required for both PSO, RMO), membar isn't */
338 : __asm__ __volatile__ ("stbar \n":::"memory");
339 : #else
340 : /*
341 : * #LoadStore (RMO) | #LoadLoad (RMO) together are the appropriate acquire
342 : * barrier for sparcv8+ upwards.
343 : */
344 : __asm__ __volatile__ ("membar #LoadStore | #LoadLoad \n":::"memory");
345 : #endif
346 : return (int) _res;
347 : }
348 :
349 : #if defined(__sparcv7) || defined(__sparc_v7__)
350 : /*
351 : * No stbar or membar available, luckily no actually produced hardware
352 : * requires a barrier. We fall through to the default gcc definition of
353 : * S_UNLOCK in this case.
354 : */
355 : #elif defined(__sparcv8) || defined(__sparc_v8__)
356 : /* stbar is available (and required for both PSO, RMO), membar isn't */
357 : #define S_UNLOCK(lock) \
358 : do \
359 : { \
360 : __asm__ __volatile__ ("stbar \n":::"memory"); \
361 : *((volatile slock_t *) (lock)) = 0; \
362 : } while (0)
363 : #else
364 : /*
365 : * #LoadStore (RMO) | #StoreStore (RMO, PSO) together are the appropriate
366 : * release barrier for sparcv8+ upwards.
367 : */
368 : #define S_UNLOCK(lock) \
369 : do \
370 : { \
371 : __asm__ __volatile__ ("membar #LoadStore | #StoreStore \n":::"memory"); \
372 : *((volatile slock_t *) (lock)) = 0; \
373 : } while (0)
374 : #endif
375 :
376 : #endif /* __sparc__ */
377 :
378 :
379 : /* PowerPC */
380 : #if defined(__powerpc__) || defined(__powerpc64__)
381 :
382 : typedef unsigned int slock_t;
383 :
384 : #define TAS(lock) tas(lock)
385 :
386 : /* On PPC, it's a win to use a non-locking test before the lwarx */
387 : #define TAS_SPIN(lock) (*(lock) ? 1 : TAS(lock))
388 :
389 : /*
390 : * The second operand of addi can hold a constant zero or a register number,
391 : * hence constraint "=&b" to avoid allocating r0. "b" stands for "address
392 : * base register"; most operands having this register-or-zero property are
393 : * address bases, e.g. the second operand of lwax.
394 : *
395 : * NOTE: per the Enhanced PowerPC Architecture manual, v1.0 dated 7-May-2002,
396 : * an isync is a sufficient synchronization barrier after a lwarx/stwcx loop.
397 : * But if the spinlock is in ordinary memory, we can use lwsync instead for
398 : * better performance.
399 : */
400 : static inline int
401 : tas(volatile slock_t *lock)
402 : {
403 : slock_t _t;
404 : int _res;
405 :
406 : __asm__ __volatile__(
407 : " lwarx %0,0,%3,1 \n"
408 : " cmpwi %0,0 \n"
409 : " bne TAS%=_fail \n"
410 : " addi %0,%0,1 \n"
411 : " stwcx. %0,0,%3 \n"
412 : " beq TAS%=_ok \n"
413 : "TAS%=_fail: \n"
414 : " li %1,1 \n"
415 : " b TAS%=_out \n"
416 : "TAS%=_ok: \n"
417 : " lwsync \n"
418 : " li %1,0 \n"
419 : "TAS%=_out: \n"
420 : : "=&b"(_t), "=r"(_res), "+m"(*lock)
421 : : "r"(lock)
422 : : "memory", "cc");
423 : return _res;
424 : }
425 :
426 : /*
427 : * PowerPC S_UNLOCK is almost standard but requires a "sync" instruction.
428 : * But we can use lwsync instead for better performance.
429 : */
430 : #define S_UNLOCK(lock) \
431 : do \
432 : { \
433 : __asm__ __volatile__ (" lwsync \n" ::: "memory"); \
434 : *((volatile slock_t *) (lock)) = 0; \
435 : } while (0)
436 :
437 : #endif /* powerpc */
438 :
439 :
440 : #if defined(__mips__) && !defined(__sgi) /* non-SGI MIPS */
441 :
442 : typedef unsigned int slock_t;
443 :
444 : #define TAS(lock) tas(lock)
445 :
446 : /*
447 : * Original MIPS-I processors lacked the LL/SC instructions, but if we are
448 : * so unfortunate as to be running on one of those, we expect that the kernel
449 : * will handle the illegal-instruction traps and emulate them for us. On
450 : * anything newer (and really, MIPS-I is extinct) LL/SC is the only sane
451 : * choice because any other synchronization method must involve a kernel
452 : * call. Unfortunately, many toolchains still default to MIPS-I as the
453 : * codegen target; if the symbol __mips shows that that's the case, we
454 : * have to force the assembler to accept LL/SC.
455 : *
456 : * R10000 and up processors require a separate SYNC, which has the same
457 : * issues as LL/SC.
458 : */
459 : #if __mips < 2
460 : #define MIPS_SET_MIPS2 " .set mips2 \n"
461 : #else
462 : #define MIPS_SET_MIPS2
463 : #endif
464 :
465 : static inline int
466 : tas(volatile slock_t *lock)
467 : {
468 : volatile slock_t *_l = lock;
469 : int _res;
470 : int _tmp;
471 :
472 : __asm__ __volatile__(
473 : " .set push \n"
474 : MIPS_SET_MIPS2
475 : " .set noreorder \n"
476 : " .set nomacro \n"
477 : " ll %0, %2 \n"
478 : " or %1, %0, 1 \n"
479 : " sc %1, %2 \n"
480 : " xori %1, 1 \n"
481 : " or %0, %0, %1 \n"
482 : " sync \n"
483 : " .set pop "
484 : : "=&r" (_res), "=&r" (_tmp), "+R" (*_l)
485 : : /* no inputs */
486 : : "memory");
487 : return _res;
488 : }
489 :
490 : /* MIPS S_UNLOCK is almost standard but requires a "sync" instruction */
491 : #define S_UNLOCK(lock) \
492 : do \
493 : { \
494 : __asm__ __volatile__( \
495 : " .set push \n" \
496 : MIPS_SET_MIPS2 \
497 : " .set noreorder \n" \
498 : " .set nomacro \n" \
499 : " sync \n" \
500 : " .set pop " \
501 : : /* no outputs */ \
502 : : /* no inputs */ \
503 : : "memory"); \
504 : *((volatile slock_t *) (lock)) = 0; \
505 : } while (0)
506 :
507 : #endif /* __mips__ && !__sgi */
508 :
509 :
510 :
511 : /*
512 : * If we have no platform-specific knowledge, but we found that the compiler
513 : * provides __sync_lock_test_and_set(), use that. Prefer the int-width
514 : * version over the char-width version if we have both, on the rather dubious
515 : * grounds that that's known to be more likely to work in the ARM ecosystem.
516 : * (But we dealt with ARM above.)
517 : */
518 : #if !defined(TAS)
519 :
520 : #if defined(HAVE_GCC__SYNC_INT32_TAS)
521 :
522 : #define TAS(lock) tas(lock)
523 :
524 : typedef int slock_t;
525 :
526 : static inline int
527 : tas(volatile slock_t *lock)
528 : {
529 : return __sync_lock_test_and_set(lock, 1);
530 : }
531 :
532 : #define S_UNLOCK(lock) __sync_lock_release(lock)
533 :
534 : #elif defined(HAVE_GCC__SYNC_CHAR_TAS)
535 :
536 : #define TAS(lock) tas(lock)
537 :
538 : typedef char slock_t;
539 :
540 : static inline int
541 : tas(volatile slock_t *lock)
542 : {
543 : return __sync_lock_test_and_set(lock, 1);
544 : }
545 :
546 : #define S_UNLOCK(lock) __sync_lock_release(lock)
547 :
548 : #endif /* HAVE_GCC__SYNC_INT32_TAS */
549 :
550 : #endif /* !defined(TAS) */
551 :
552 :
553 : /*
554 : * Default implementation of S_UNLOCK() for gcc/icc.
555 : *
556 : * Note that this implementation is unsafe for any platform that can reorder
557 : * a memory access (either load or store) after a following store. That
558 : * happens not to be possible on x86 and most legacy architectures (some are
559 : * single-processor!), but many modern systems have weaker memory ordering.
560 : * Those that do must define their own version of S_UNLOCK() rather than
561 : * relying on this one.
562 : */
563 : #if !defined(S_UNLOCK)
564 : #define S_UNLOCK(lock) \
565 : do { __asm__ __volatile__("" : : : "memory"); *(lock) = 0; } while (0)
566 : #endif
567 :
568 : #endif /* defined(__GNUC__) || defined(__INTEL_COMPILER) */
569 :
570 :
571 : /*
572 : * ---------------------------------------------------------------------
573 : * Platforms that use non-gcc inline assembly:
574 : * ---------------------------------------------------------------------
575 : */
576 :
577 : #if !defined(TAS) /* We didn't trigger above, let's try here */
578 :
579 : #ifdef _MSC_VER
580 : typedef LONG slock_t;
581 :
582 : #define TAS(lock) (InterlockedCompareExchange(lock, 1, 0))
583 :
584 : #define SPIN_DELAY() spin_delay()
585 :
586 : #ifdef __aarch64__
587 : static __forceinline void
588 : spin_delay(void)
589 : {
590 : /*
591 : * Research indicates ISB is better than __yield() on AArch64. See
592 : * https://postgr.es/m/1c2a29b8-5b1e-44f7-a871-71ec5fefc120%40app.fastmail.com.
593 : */
594 : __isb(_ARM64_BARRIER_SY);
595 : }
596 : #elif defined(_WIN64)
597 : static __forceinline void
598 : spin_delay(void)
599 : {
600 : /*
601 : * If using Visual C++ on Win64, inline assembly is unavailable.
602 : * Use a _mm_pause intrinsic instead of rep nop.
603 : */
604 : _mm_pause();
605 : }
606 : #else
607 : static __forceinline void
608 : spin_delay(void)
609 : {
610 : /* See comment for gcc code. Same code, MASM syntax */
611 : __asm rep nop;
612 : }
613 : #endif
614 :
615 : #include <intrin.h>
616 :
617 : #ifdef __aarch64__
618 :
619 : /* _ReadWriteBarrier() is insufficient on non-TSO architectures. */
620 : #pragma intrinsic(_InterlockedExchange)
621 : #define S_UNLOCK(lock) _InterlockedExchange(lock, 0)
622 :
623 : #else
624 :
625 : #pragma intrinsic(_ReadWriteBarrier)
626 : #define S_UNLOCK(lock) \
627 : do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)
628 :
629 : #endif
630 : #endif
631 :
632 :
633 : #endif /* !defined(TAS) */
634 :
635 :
636 : /*
637 : * Default Definitions - override these above as needed.
638 : */
639 :
640 : /*
641 : * Make sure S_LOCK is defined, either explicitly for the platform or via a TAS
642 : * macro for the platform.
643 : */
644 : #if !defined(S_LOCK)
645 : #ifdef TAS
646 : #define USE_DEFAULT_S_LOCK
647 : extern int s_lock(volatile slock_t *lock, const char *file, int line, const char *func);
648 : #define S_LOCK(lock) \
649 : (TAS(lock) ? s_lock((lock), __FILE__, __LINE__, __func__) : 0)
650 : #else
651 : #error PostgreSQL does not have spinlock support on this platform. Please report this to pgsql-bugs@lists.postgresql.org.
652 : #endif /* TAS */
653 : #endif /* S_LOCK */
654 :
655 : #if !defined(S_UNLOCK)
656 : /*
657 : * Our default implementation of S_UNLOCK is essentially *(lock) = 0. This
658 : * is unsafe if the platform can reorder a memory access (either load or
659 : * store) after a following store; platforms where this is possible must
660 : * define their own S_UNLOCK. But CPU reordering is not the only concern:
661 : * if we simply defined S_UNLOCK() as an inline macro, the compiler might
662 : * reorder instructions from inside the critical section to occur after the
663 : * lock release. Since the compiler probably can't know what the external
664 : * function s_unlock is doing, putting the same logic there should be adequate.
665 : * A sufficiently-smart globally optimizing compiler could break that
666 : * assumption, though, and the cost of a function call for every spinlock
667 : * release may hurt performance significantly, so we use this implementation
668 : * only for platforms where we don't know of a suitable intrinsic. For the
669 : * most part, those are relatively obscure platform/compiler combinations to
670 : * which the PostgreSQL project does not have access.
671 : */
672 : #define USE_DEFAULT_S_UNLOCK
673 : extern void s_unlock(volatile slock_t *lock);
674 : #define S_UNLOCK(lock) s_unlock(lock)
675 : #endif /* S_UNLOCK */
676 :
677 : #if !defined(S_INIT_LOCK)
678 : #define S_INIT_LOCK(lock) S_UNLOCK(lock)
679 : #endif /* S_INIT_LOCK */
680 :
681 : #if !defined(SPIN_DELAY)
682 : #define SPIN_DELAY() ((void) 0)
683 : #endif /* SPIN_DELAY */
684 :
685 : /*
686 : * TAS_SPIN is only needed by the default S_LOCK's helper function (s_lock()),
687 : * so we only provide a default when there is a TAS to base it on.
688 : */
689 : #if !defined(TAS_SPIN) && defined(TAS)
690 : #define TAS_SPIN(lock) TAS(lock)
691 : #endif /* ! TAS_SPIN && TAS */
692 :
693 :
694 : /*
695 : * Platform-independent out-of-line support routines
696 : */
697 :
698 : /* Support for dynamic adjustment of spins_per_delay */
699 : #define DEFAULT_SPINS_PER_DELAY 100
700 :
701 : extern void set_spins_per_delay(int shared_spins_per_delay);
702 : extern int update_spins_per_delay(int shared_spins_per_delay);
703 :
704 : /*
705 : * Support for spin delay which is useful in various places where
706 : * spinlock-like procedures take place.
707 : */
708 : typedef struct
709 : {
710 : int spins;
711 : int delays;
712 : int cur_delay;
713 : const char *file;
714 : int line;
715 : const char *func;
716 : } SpinDelayStatus;
717 :
718 : static inline void
3833 andres@anarazel.de 719 30860 : init_spin_delay(SpinDelayStatus *status,
720 : const char *file, int line, const char *func)
721 : {
722 30860 : status->spins = 0;
723 30860 : status->delays = 0;
724 30860 : status->cur_delay = 0;
725 30860 : status->file = file;
726 30860 : status->line = line;
727 30860 : status->func = func;
728 30860 : }
729 :
730 : #define init_local_spin_delay(status) init_spin_delay(status, __FILE__, __LINE__, __func__)
731 : extern void perform_spin_delay(SpinDelayStatus *status);
732 : extern void finish_spin_delay(SpinDelayStatus *status);
733 :
734 : #endif /* S_LOCK_H */
|
