Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * s_lock.h
4 : * Hardware-dependent implementation of spinlocks.
5 : *
6 : * NOTE: none of the macros in this file are intended to be called directly.
7 : * Call them through the hardware-independent 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 : * bool S_LOCK_FREE(slock_t *lock)
25 : * Tests if the lock is free. Returns true if free, false if locked.
26 : * This does *not* change the state of the lock.
27 : *
28 : * void SPIN_DELAY(void)
29 : * Delay operation to occur inside spinlock wait loop.
30 : *
31 : * Note to implementors: there are default implementations for all these
32 : * macros at the bottom of the file. Check if your platform can use
33 : * these or needs to override them.
34 : *
35 : * Usually, S_LOCK() is implemented in terms of even lower-level macros
36 : * TAS() and TAS_SPIN():
37 : *
38 : * int TAS(slock_t *lock)
39 : * Atomic test-and-set instruction. Attempt to acquire the lock,
40 : * but do *not* wait. Returns 0 if successful, nonzero if unable
41 : * to acquire the lock.
42 : *
43 : * int TAS_SPIN(slock_t *lock)
44 : * Like TAS(), but this version is used when waiting for a lock
45 : * previously found to be contended. By default, this is the
46 : * same as TAS(), but on some architectures it's better to poll a
47 : * contended lock using an unlocked instruction and retry the
48 : * atomic test-and-set only when it appears free.
49 : *
50 : * TAS() and TAS_SPIN() are NOT part of the API, and should never be called
51 : * directly.
52 : *
53 : * CAUTION: on some platforms TAS() and/or TAS_SPIN() may sometimes report
54 : * failure to acquire a lock even when the lock is not locked. For example,
55 : * on Alpha TAS() will "fail" if interrupted. Therefore a retry loop must
56 : * always be used, even if you are certain the lock is free.
57 : *
58 : * It is the responsibility of these macros to make sure that the compiler
59 : * does not re-order accesses to shared memory to precede the actual lock
60 : * acquisition, or follow the lock release. Prior to PostgreSQL 9.5, this
61 : * was the caller's responsibility, which meant that callers had to use
62 : * volatile-qualified pointers to refer to both the spinlock itself and the
63 : * shared data being accessed within the spinlocked critical section. This
64 : * was notationally awkward, easy to forget (and thus error-prone), and
65 : * prevented some useful compiler optimizations. For these reasons, we
66 : * now require that the macros themselves prevent compiler re-ordering,
67 : * so that the caller doesn't need to take special precautions.
68 : *
69 : * On platforms with weak memory ordering, the TAS(), TAS_SPIN(), and
70 : * S_UNLOCK() macros must further include hardware-level memory fence
71 : * instructions to prevent similar re-ordering at the hardware level.
72 : * TAS() and TAS_SPIN() must guarantee that loads and stores issued after
73 : * the macro are not executed until the lock has been obtained. Conversely,
74 : * S_UNLOCK() must guarantee that loads and stores issued before the macro
75 : * have been executed before the lock is released.
76 : *
77 : * On most supported platforms, TAS() uses a tas() function written
78 : * in assembly language to execute a hardware atomic-test-and-set
79 : * instruction. Equivalent OS-supplied mutex routines could be used too.
80 : *
81 : * If no system-specific TAS() is available (ie, HAVE_SPINLOCKS is not
82 : * defined), then we fall back on an emulation that uses SysV semaphores
83 : * (see spin.c). This emulation will be MUCH MUCH slower than a proper TAS()
84 : * implementation, because of the cost of a kernel call per lock or unlock.
85 : * An old report is that Postgres spends around 40% of its time in semop(2)
86 : * when using the SysV semaphore code.
87 : *
88 : *
89 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
90 : * Portions Copyright (c) 1994, Regents of the University of California
91 : *
92 : * src/include/storage/s_lock.h
93 : *
94 : *-------------------------------------------------------------------------
95 : */
96 : #ifndef S_LOCK_H
97 : #define S_LOCK_H
98 :
99 : #ifdef FRONTEND
100 : #error "s_lock.h may not be included from frontend code"
101 : #endif
102 :
103 : #ifdef HAVE_SPINLOCKS /* skip spinlocks if requested */
104 :
105 : #if defined(__GNUC__) || defined(__INTEL_COMPILER)
106 : /*************************************************************************
107 : * All the gcc inlines
108 : * Gcc consistently defines the CPU as __cpu__.
109 : * Other compilers use __cpu or __cpu__ so we test for both in those cases.
110 : */
111 :
112 : /*----------
113 : * Standard gcc asm format (assuming "volatile slock_t *lock"):
114 :
115 : __asm__ __volatile__(
116 : " instruction \n"
117 : " instruction \n"
118 : " instruction \n"
119 : : "=r"(_res), "+m"(*lock) // return register, in/out lock value
120 : : "r"(lock) // lock pointer, in input register
121 : : "memory", "cc"); // show clobbered registers here
122 :
123 : * The output-operands list (after first colon) should always include
124 : * "+m"(*lock), whether or not the asm code actually refers to this
125 : * operand directly. This ensures that gcc believes the value in the
126 : * lock variable is used and set by the asm code. Also, the clobbers
127 : * list (after third colon) should always include "memory"; this prevents
128 : * gcc from thinking it can cache the values of shared-memory fields
129 : * across the asm code. Add "cc" if your asm code changes the condition
130 : * code register, and also list any temp registers the code uses.
131 : *----------
132 : */
133 :
134 :
135 : #ifdef __i386__ /* 32-bit i386 */
136 : #define HAS_TEST_AND_SET
137 :
138 : typedef unsigned char slock_t;
139 :
140 : #define TAS(lock) tas(lock)
141 :
142 : static __inline__ int
143 : tas(volatile slock_t *lock)
144 : {
145 : slock_t _res = 1;
146 :
147 : /*
148 : * Use a non-locking test before asserting the bus lock. Note that the
149 : * extra test appears to be a small loss on some x86 platforms and a small
150 : * win on others; it's by no means clear that we should keep it.
151 : *
152 : * When this was last tested, we didn't have separate TAS() and TAS_SPIN()
153 : * macros. Nowadays it probably would be better to do a non-locking test
154 : * in TAS_SPIN() but not in TAS(), like on x86_64, but no-one's done the
155 : * testing to verify that. Without some empirical evidence, better to
156 : * leave it alone.
157 : */
158 : __asm__ __volatile__(
159 : " cmpb $0,%1 \n"
160 : " jne 1f \n"
161 : " lock \n"
162 : " xchgb %0,%1 \n"
163 : "1: \n"
164 : : "+q"(_res), "+m"(*lock)
165 : : /* no inputs */
166 : : "memory", "cc");
167 : return (int) _res;
168 : }
169 :
170 : #define SPIN_DELAY() spin_delay()
171 :
172 : static __inline__ void
173 : spin_delay(void)
174 : {
175 : /*
176 : * This sequence is equivalent to the PAUSE instruction ("rep" is
177 : * ignored by old IA32 processors if the following instruction is
178 : * not a string operation); the IA-32 Architecture Software
179 : * Developer's Manual, Vol. 3, Section 7.7.2 describes why using
180 : * PAUSE in the inner loop of a spin lock is necessary for good
181 : * performance:
182 : *
183 : * The PAUSE instruction improves the performance of IA-32
184 : * processors supporting Hyper-Threading Technology when
185 : * executing spin-wait loops and other routines where one
186 : * thread is accessing a shared lock or semaphore in a tight
187 : * polling loop. When executing a spin-wait loop, the
188 : * processor can suffer a severe performance penalty when
189 : * exiting the loop because it detects a possible memory order
190 : * violation and flushes the core processor's pipeline. The
191 : * PAUSE instruction provides a hint to the processor that the
192 : * code sequence is a spin-wait loop. The processor uses this
193 : * hint to avoid the memory order violation and prevent the
194 : * pipeline flush. In addition, the PAUSE instruction
195 : * de-pipelines the spin-wait loop to prevent it from
196 : * consuming execution resources excessively.
197 : */
198 : __asm__ __volatile__(
199 : " rep; nop \n");
200 : }
201 :
202 : #endif /* __i386__ */
203 :
204 :
205 : #ifdef __x86_64__ /* AMD Opteron, Intel EM64T */
206 : #define HAS_TEST_AND_SET
207 :
208 : typedef unsigned char slock_t;
209 :
210 : #define TAS(lock) tas(lock)
211 :
212 : /*
213 : * On Intel EM64T, it's a win to use a non-locking test before the xchg proper,
214 : * but only when spinning.
215 : *
216 : * See also Implementing Scalable Atomic Locks for Multi-Core Intel(tm) EM64T
217 : * and IA32, by Michael Chynoweth and Mary R. Lee. As of this writing, it is
218 : * available at:
219 : * http://software.intel.com/en-us/articles/implementing-scalable-atomic-locks-for-multi-core-intel-em64t-and-ia32-architectures
220 : */
221 : #define TAS_SPIN(lock) (*(lock) ? 1 : TAS(lock))
222 :
223 : static __inline__ int
224 92429884 : tas(volatile slock_t *lock)
225 : {
226 92429884 : slock_t _res = 1;
227 :
228 92429884 : __asm__ __volatile__(
229 : " lock \n"
230 : " xchgb %0,%1 \n"
231 : : "+q"(_res), "+m"(*lock)
232 : : /* no inputs */
233 : : "memory", "cc");
234 92429884 : return (int) _res;
235 : }
236 :
237 : #define SPIN_DELAY() spin_delay()
238 :
239 : static __inline__ void
240 178234 : spin_delay(void)
241 : {
242 : /*
243 : * Adding a PAUSE in the spin delay loop is demonstrably a no-op on
244 : * Opteron, but it may be of some use on EM64T, so we keep it.
245 : */
246 178234 : __asm__ __volatile__(
247 : " rep; nop \n");
248 178234 : }
249 :
250 : #endif /* __x86_64__ */
251 :
252 :
253 : /*
254 : * On ARM and ARM64, we use __sync_lock_test_and_set(int *, int) if available.
255 : *
256 : * We use the int-width variant of the builtin because it works on more chips
257 : * than other widths.
258 : */
259 : #if defined(__arm__) || defined(__arm) || defined(__aarch64__)
260 : #ifdef HAVE_GCC__SYNC_INT32_TAS
261 : #define HAS_TEST_AND_SET
262 :
263 : #define TAS(lock) tas(lock)
264 :
265 : typedef int slock_t;
266 :
267 : static __inline__ int
268 : tas(volatile slock_t *lock)
269 : {
270 : return __sync_lock_test_and_set(lock, 1);
271 : }
272 :
273 : #define S_UNLOCK(lock) __sync_lock_release(lock)
274 :
275 : /*
276 : * Using an ISB instruction to delay in spinlock loops appears beneficial on
277 : * high-core-count ARM64 processors. It seems mostly a wash for smaller gear,
278 : * and ISB doesn't exist at all on pre-v7 ARM chips.
279 : */
280 : #if defined(__aarch64__)
281 :
282 : #define SPIN_DELAY() spin_delay()
283 :
284 : static __inline__ void
285 : spin_delay(void)
286 : {
287 : __asm__ __volatile__(
288 : " isb; \n");
289 : }
290 :
291 : #endif /* __aarch64__ */
292 : #endif /* HAVE_GCC__SYNC_INT32_TAS */
293 : #endif /* __arm__ || __arm || __aarch64__ */
294 :
295 :
296 : /* S/390 and S/390x Linux (32- and 64-bit zSeries) */
297 : #if defined(__s390__) || defined(__s390x__)
298 : #define HAS_TEST_AND_SET
299 :
300 : typedef unsigned int slock_t;
301 :
302 : #define TAS(lock) tas(lock)
303 :
304 : static __inline__ int
305 : tas(volatile slock_t *lock)
306 : {
307 : int _res = 0;
308 :
309 : __asm__ __volatile__(
310 : " cs %0,%3,0(%2) \n"
311 : : "+d"(_res), "+m"(*lock)
312 : : "a"(lock), "d"(1)
313 : : "memory", "cc");
314 : return _res;
315 : }
316 :
317 : #endif /* __s390__ || __s390x__ */
318 :
319 :
320 : #if defined(__sparc__) /* Sparc */
321 : /*
322 : * Solaris has always run sparc processors in TSO (total store) mode, but
323 : * linux didn't use to and the *BSDs still don't. So, be careful about
324 : * acquire/release semantics. The CPU will treat superfluous members as
325 : * NOPs, so it's just code space.
326 : */
327 : #define HAS_TEST_AND_SET
328 :
329 : typedef unsigned char slock_t;
330 :
331 : #define TAS(lock) tas(lock)
332 :
333 : static __inline__ int
334 : tas(volatile slock_t *lock)
335 : {
336 : slock_t _res;
337 :
338 : /*
339 : * See comment in src/backend/port/tas/sunstudio_sparc.s for why this
340 : * uses "ldstub", and that file uses "cas". gcc currently generates
341 : * sparcv7-targeted binaries, so "cas" use isn't possible.
342 : */
343 : __asm__ __volatile__(
344 : " ldstub [%2], %0 \n"
345 : : "=r"(_res), "+m"(*lock)
346 : : "r"(lock)
347 : : "memory");
348 : #if defined(__sparcv7) || defined(__sparc_v7__)
349 : /*
350 : * No stbar or membar available, luckily no actually produced hardware
351 : * requires a barrier.
352 : */
353 : #elif defined(__sparcv8) || defined(__sparc_v8__)
354 : /* stbar is available (and required for both PSO, RMO), membar isn't */
355 : __asm__ __volatile__ ("stbar \n":::"memory");
356 : #else
357 : /*
358 : * #LoadStore (RMO) | #LoadLoad (RMO) together are the appropriate acquire
359 : * barrier for sparcv8+ upwards.
360 : */
361 : __asm__ __volatile__ ("membar #LoadStore | #LoadLoad \n":::"memory");
362 : #endif
363 : return (int) _res;
364 : }
365 :
366 : #if defined(__sparcv7) || defined(__sparc_v7__)
367 : /*
368 : * No stbar or membar available, luckily no actually produced hardware
369 : * requires a barrier. We fall through to the default gcc definition of
370 : * S_UNLOCK in this case.
371 : */
372 : #elif defined(__sparcv8) || defined(__sparc_v8__)
373 : /* stbar is available (and required for both PSO, RMO), membar isn't */
374 : #define S_UNLOCK(lock) \
375 : do \
376 : { \
377 : __asm__ __volatile__ ("stbar \n":::"memory"); \
378 : *((volatile slock_t *) (lock)) = 0; \
379 : } while (0)
380 : #else
381 : /*
382 : * #LoadStore (RMO) | #StoreStore (RMO, PSO) together are the appropriate
383 : * release barrier for sparcv8+ upwards.
384 : */
385 : #define S_UNLOCK(lock) \
386 : do \
387 : { \
388 : __asm__ __volatile__ ("membar #LoadStore | #StoreStore \n":::"memory"); \
389 : *((volatile slock_t *) (lock)) = 0; \
390 : } while (0)
391 : #endif
392 :
393 : #endif /* __sparc__ */
394 :
395 :
396 : /* PowerPC */
397 : #if defined(__ppc__) || defined(__powerpc__) || defined(__ppc64__) || defined(__powerpc64__)
398 : #define HAS_TEST_AND_SET
399 :
400 : typedef unsigned int slock_t;
401 :
402 : #define TAS(lock) tas(lock)
403 :
404 : /* On PPC, it's a win to use a non-locking test before the lwarx */
405 : #define TAS_SPIN(lock) (*(lock) ? 1 : TAS(lock))
406 :
407 : /*
408 : * The second operand of addi can hold a constant zero or a register number,
409 : * hence constraint "=&b" to avoid allocating r0. "b" stands for "address
410 : * base register"; most operands having this register-or-zero property are
411 : * address bases, e.g. the second operand of lwax.
412 : *
413 : * NOTE: per the Enhanced PowerPC Architecture manual, v1.0 dated 7-May-2002,
414 : * an isync is a sufficient synchronization barrier after a lwarx/stwcx loop.
415 : * But if the spinlock is in ordinary memory, we can use lwsync instead for
416 : * better performance.
417 : */
418 : static __inline__ int
419 : tas(volatile slock_t *lock)
420 : {
421 : slock_t _t;
422 : int _res;
423 :
424 : __asm__ __volatile__(
425 : " lwarx %0,0,%3,1 \n"
426 : " cmpwi %0,0 \n"
427 : " bne 1f \n"
428 : " addi %0,%0,1 \n"
429 : " stwcx. %0,0,%3 \n"
430 : " beq 2f \n"
431 : "1: \n"
432 : " li %1,1 \n"
433 : " b 3f \n"
434 : "2: \n"
435 : " lwsync \n"
436 : " li %1,0 \n"
437 : "3: \n"
438 : : "=&b"(_t), "=r"(_res), "+m"(*lock)
439 : : "r"(lock)
440 : : "memory", "cc");
441 : return _res;
442 : }
443 :
444 : /*
445 : * PowerPC S_UNLOCK is almost standard but requires a "sync" instruction.
446 : * But we can use lwsync instead for better performance.
447 : */
448 : #define S_UNLOCK(lock) \
449 : do \
450 : { \
451 : __asm__ __volatile__ (" lwsync \n" ::: "memory"); \
452 : *((volatile slock_t *) (lock)) = 0; \
453 : } while (0)
454 :
455 : #endif /* powerpc */
456 :
457 :
458 : #if defined(__mips__) && !defined(__sgi) /* non-SGI MIPS */
459 : #define HAS_TEST_AND_SET
460 :
461 : typedef unsigned int slock_t;
462 :
463 : #define TAS(lock) tas(lock)
464 :
465 : /*
466 : * Original MIPS-I processors lacked the LL/SC instructions, but if we are
467 : * so unfortunate as to be running on one of those, we expect that the kernel
468 : * will handle the illegal-instruction traps and emulate them for us. On
469 : * anything newer (and really, MIPS-I is extinct) LL/SC is the only sane
470 : * choice because any other synchronization method must involve a kernel
471 : * call. Unfortunately, many toolchains still default to MIPS-I as the
472 : * codegen target; if the symbol __mips shows that that's the case, we
473 : * have to force the assembler to accept LL/SC.
474 : *
475 : * R10000 and up processors require a separate SYNC, which has the same
476 : * issues as LL/SC.
477 : */
478 : #if __mips < 2
479 : #define MIPS_SET_MIPS2 " .set mips2 \n"
480 : #else
481 : #define MIPS_SET_MIPS2
482 : #endif
483 :
484 : static __inline__ int
485 : tas(volatile slock_t *lock)
486 : {
487 : volatile slock_t *_l = lock;
488 : int _res;
489 : int _tmp;
490 :
491 : __asm__ __volatile__(
492 : " .set push \n"
493 : MIPS_SET_MIPS2
494 : " .set noreorder \n"
495 : " .set nomacro \n"
496 : " ll %0, %2 \n"
497 : " or %1, %0, 1 \n"
498 : " sc %1, %2 \n"
499 : " xori %1, 1 \n"
500 : " or %0, %0, %1 \n"
501 : " sync \n"
502 : " .set pop "
503 : : "=&r" (_res), "=&r" (_tmp), "+R" (*_l)
504 : : /* no inputs */
505 : : "memory");
506 : return _res;
507 : }
508 :
509 : /* MIPS S_UNLOCK is almost standard but requires a "sync" instruction */
510 : #define S_UNLOCK(lock) \
511 : do \
512 : { \
513 : __asm__ __volatile__( \
514 : " .set push \n" \
515 : MIPS_SET_MIPS2 \
516 : " .set noreorder \n" \
517 : " .set nomacro \n" \
518 : " sync \n" \
519 : " .set pop " \
520 : : /* no outputs */ \
521 : : /* no inputs */ \
522 : : "memory"); \
523 : *((volatile slock_t *) (lock)) = 0; \
524 : } while (0)
525 :
526 : #endif /* __mips__ && !__sgi */
527 :
528 :
529 : #if defined(__hppa) || defined(__hppa__) /* HP PA-RISC */
530 : /*
531 : * HP's PA-RISC
532 : *
533 : * Because LDCWX requires a 16-byte-aligned address, we declare slock_t as a
534 : * 16-byte struct. The active word in the struct is whichever has the aligned
535 : * address; the other three words just sit at -1.
536 : */
537 : #define HAS_TEST_AND_SET
538 :
539 : typedef struct
540 : {
541 : int sema[4];
542 : } slock_t;
543 :
544 : #define TAS_ACTIVE_WORD(lock) ((volatile int *) (((uintptr_t) (lock) + 15) & ~15))
545 :
546 : static __inline__ int
547 : tas(volatile slock_t *lock)
548 : {
549 : volatile int *lockword = TAS_ACTIVE_WORD(lock);
550 : int lockval;
551 :
552 : /*
553 : * The LDCWX instruction atomically clears the target word and
554 : * returns the previous value. Hence, if the instruction returns
555 : * 0, someone else has already acquired the lock before we tested
556 : * it (i.e., we have failed).
557 : *
558 : * Notice that this means that we actually clear the word to set
559 : * the lock and set the word to clear the lock. This is the
560 : * opposite behavior from the SPARC LDSTUB instruction. For some
561 : * reason everything that H-P does is rather baroque...
562 : *
563 : * For details about the LDCWX instruction, see the "Precision
564 : * Architecture and Instruction Reference Manual" (09740-90014 of June
565 : * 1987), p. 5-38.
566 : */
567 : __asm__ __volatile__(
568 : " ldcwx 0(0,%2),%0 \n"
569 : : "=r"(lockval), "+m"(*lockword)
570 : : "r"(lockword)
571 : : "memory");
572 : return (lockval == 0);
573 : }
574 :
575 : #define S_UNLOCK(lock) \
576 : do { \
577 : __asm__ __volatile__("" : : : "memory"); \
578 : *TAS_ACTIVE_WORD(lock) = -1; \
579 : } while (0)
580 :
581 : #define S_INIT_LOCK(lock) \
582 : do { \
583 : volatile slock_t *lock_ = (lock); \
584 : lock_->sema[0] = -1; \
585 : lock_->sema[1] = -1; \
586 : lock_->sema[2] = -1; \
587 : lock_->sema[3] = -1; \
588 : } while (0)
589 :
590 : #define S_LOCK_FREE(lock) (*TAS_ACTIVE_WORD(lock) != 0)
591 :
592 : #endif /* __hppa || __hppa__ */
593 :
594 :
595 : /*
596 : * If we have no platform-specific knowledge, but we found that the compiler
597 : * provides __sync_lock_test_and_set(), use that. Prefer the int-width
598 : * version over the char-width version if we have both, on the rather dubious
599 : * grounds that that's known to be more likely to work in the ARM ecosystem.
600 : * (But we dealt with ARM above.)
601 : */
602 : #if !defined(HAS_TEST_AND_SET)
603 :
604 : #if defined(HAVE_GCC__SYNC_INT32_TAS)
605 : #define HAS_TEST_AND_SET
606 :
607 : #define TAS(lock) tas(lock)
608 :
609 : typedef int slock_t;
610 :
611 : static __inline__ int
612 : tas(volatile slock_t *lock)
613 : {
614 : return __sync_lock_test_and_set(lock, 1);
615 : }
616 :
617 : #define S_UNLOCK(lock) __sync_lock_release(lock)
618 :
619 : #elif defined(HAVE_GCC__SYNC_CHAR_TAS)
620 : #define HAS_TEST_AND_SET
621 :
622 : #define TAS(lock) tas(lock)
623 :
624 : typedef char slock_t;
625 :
626 : static __inline__ int
627 : tas(volatile slock_t *lock)
628 : {
629 : return __sync_lock_test_and_set(lock, 1);
630 : }
631 :
632 : #define S_UNLOCK(lock) __sync_lock_release(lock)
633 :
634 : #endif /* HAVE_GCC__SYNC_INT32_TAS */
635 :
636 : #endif /* !defined(HAS_TEST_AND_SET) */
637 :
638 :
639 : /*
640 : * Default implementation of S_UNLOCK() for gcc/icc.
641 : *
642 : * Note that this implementation is unsafe for any platform that can reorder
643 : * a memory access (either load or store) after a following store. That
644 : * happens not to be possible on x86 and most legacy architectures (some are
645 : * single-processor!), but many modern systems have weaker memory ordering.
646 : * Those that do must define their own version of S_UNLOCK() rather than
647 : * relying on this one.
648 : */
649 : #if !defined(S_UNLOCK)
650 : #define S_UNLOCK(lock) \
651 : do { __asm__ __volatile__("" : : : "memory"); *(lock) = 0; } while (0)
652 : #endif
653 :
654 : #endif /* defined(__GNUC__) || defined(__INTEL_COMPILER) */
655 :
656 :
657 : /*
658 : * ---------------------------------------------------------------------
659 : * Platforms that use non-gcc inline assembly:
660 : * ---------------------------------------------------------------------
661 : */
662 :
663 : #if !defined(HAS_TEST_AND_SET) /* We didn't trigger above, let's try here */
664 :
665 : /* These are in sunstudio_(sparc|x86).s */
666 :
667 : #if defined(__SUNPRO_C) && (defined(__i386) || defined(__x86_64__) || defined(__sparc__) || defined(__sparc))
668 : #define HAS_TEST_AND_SET
669 :
670 : #if defined(__i386) || defined(__x86_64__) || defined(__sparcv9) || defined(__sparcv8plus)
671 : typedef unsigned int slock_t;
672 : #else
673 : typedef unsigned char slock_t;
674 : #endif
675 :
676 : extern slock_t pg_atomic_cas(volatile slock_t *lock, slock_t with,
677 : slock_t cmp);
678 :
679 : #define TAS(a) (pg_atomic_cas((a), 1, 0) != 0)
680 : #endif
681 :
682 :
683 : #ifdef _MSC_VER
684 : typedef LONG slock_t;
685 :
686 : #define HAS_TEST_AND_SET
687 : #define TAS(lock) (InterlockedCompareExchange(lock, 1, 0))
688 :
689 : #define SPIN_DELAY() spin_delay()
690 :
691 : /* If using Visual C++ on Win64, inline assembly is unavailable.
692 : * Use a _mm_pause intrinsic instead of rep nop.
693 : */
694 : #if defined(_WIN64)
695 : static __forceinline void
696 : spin_delay(void)
697 : {
698 : _mm_pause();
699 : }
700 : #else
701 : static __forceinline void
702 : spin_delay(void)
703 : {
704 : /* See comment for gcc code. Same code, MASM syntax */
705 : __asm rep nop;
706 : }
707 : #endif
708 :
709 : #include <intrin.h>
710 : #pragma intrinsic(_ReadWriteBarrier)
711 :
712 : #define S_UNLOCK(lock) \
713 : do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)
714 :
715 : #endif
716 :
717 :
718 : #endif /* !defined(HAS_TEST_AND_SET) */
719 :
720 :
721 : /* Blow up if we didn't have any way to do spinlocks */
722 : #ifndef HAS_TEST_AND_SET
723 : #error PostgreSQL does not have native spinlock support on this platform. To continue the compilation, rerun configure using --disable-spinlocks. However, performance will be poor. Please report this to pgsql-bugs@lists.postgresql.org.
724 : #endif
725 :
726 :
727 : #else /* !HAVE_SPINLOCKS */
728 :
729 :
730 : /*
731 : * Fake spinlock implementation using semaphores --- slow and prone
732 : * to fall foul of kernel limits on number of semaphores, so don't use this
733 : * unless you must! The subroutines appear in spin.c.
734 : */
735 : typedef int slock_t;
736 :
737 : extern bool s_lock_free_sema(volatile slock_t *lock);
738 : extern void s_unlock_sema(volatile slock_t *lock);
739 : extern void s_init_lock_sema(volatile slock_t *lock, bool nested);
740 : extern int tas_sema(volatile slock_t *lock);
741 :
742 : #define S_LOCK_FREE(lock) s_lock_free_sema(lock)
743 : #define S_UNLOCK(lock) s_unlock_sema(lock)
744 : #define S_INIT_LOCK(lock) s_init_lock_sema(lock, false)
745 : #define TAS(lock) tas_sema(lock)
746 :
747 :
748 : #endif /* HAVE_SPINLOCKS */
749 :
750 :
751 : /*
752 : * Default Definitions - override these above as needed.
753 : */
754 :
755 : #if !defined(S_LOCK)
756 : #define S_LOCK(lock) \
757 : (TAS(lock) ? s_lock((lock), __FILE__, __LINE__, __func__) : 0)
758 : #endif /* S_LOCK */
759 :
760 : #if !defined(S_LOCK_FREE)
761 : #define S_LOCK_FREE(lock) (*(lock) == 0)
762 : #endif /* S_LOCK_FREE */
763 :
764 : #if !defined(S_UNLOCK)
765 : /*
766 : * Our default implementation of S_UNLOCK is essentially *(lock) = 0. This
767 : * is unsafe if the platform can reorder a memory access (either load or
768 : * store) after a following store; platforms where this is possible must
769 : * define their own S_UNLOCK. But CPU reordering is not the only concern:
770 : * if we simply defined S_UNLOCK() as an inline macro, the compiler might
771 : * reorder instructions from inside the critical section to occur after the
772 : * lock release. Since the compiler probably can't know what the external
773 : * function s_unlock is doing, putting the same logic there should be adequate.
774 : * A sufficiently-smart globally optimizing compiler could break that
775 : * assumption, though, and the cost of a function call for every spinlock
776 : * release may hurt performance significantly, so we use this implementation
777 : * only for platforms where we don't know of a suitable intrinsic. For the
778 : * most part, those are relatively obscure platform/compiler combinations to
779 : * which the PostgreSQL project does not have access.
780 : */
781 : #define USE_DEFAULT_S_UNLOCK
782 : extern void s_unlock(volatile slock_t *lock);
783 : #define S_UNLOCK(lock) s_unlock(lock)
784 : #endif /* S_UNLOCK */
785 :
786 : #if !defined(S_INIT_LOCK)
787 : #define S_INIT_LOCK(lock) S_UNLOCK(lock)
788 : #endif /* S_INIT_LOCK */
789 :
790 : #if !defined(SPIN_DELAY)
791 : #define SPIN_DELAY() ((void) 0)
792 : #endif /* SPIN_DELAY */
793 :
794 : #if !defined(TAS)
795 : extern int tas(volatile slock_t *lock); /* in port/.../tas.s, or
796 : * s_lock.c */
797 :
798 : #define TAS(lock) tas(lock)
799 : #endif /* TAS */
800 :
801 : #if !defined(TAS_SPIN)
802 : #define TAS_SPIN(lock) TAS(lock)
803 : #endif /* TAS_SPIN */
804 :
805 :
806 : /*
807 : * Platform-independent out-of-line support routines
808 : */
809 : extern int s_lock(volatile slock_t *lock, const char *file, int line, const char *func);
810 :
811 : /* Support for dynamic adjustment of spins_per_delay */
812 : #define DEFAULT_SPINS_PER_DELAY 100
813 :
814 : extern void set_spins_per_delay(int shared_spins_per_delay);
815 : extern int update_spins_per_delay(int shared_spins_per_delay);
816 :
817 : /*
818 : * Support for spin delay which is useful in various places where
819 : * spinlock-like procedures take place.
820 : */
821 : typedef struct
822 : {
823 : int spins;
824 : int delays;
825 : int cur_delay;
826 : const char *file;
827 : int line;
828 : const char *func;
829 : } SpinDelayStatus;
830 :
831 : static inline void
832 47423808 : init_spin_delay(SpinDelayStatus *status,
833 : const char *file, int line, const char *func)
834 : {
835 47423808 : status->spins = 0;
836 47423808 : status->delays = 0;
837 47423808 : status->cur_delay = 0;
838 47423808 : status->file = file;
839 47423808 : status->line = line;
840 47423808 : status->func = func;
841 47423808 : }
842 :
843 : #define init_local_spin_delay(status) init_spin_delay(status, __FILE__, __LINE__, __func__)
844 : extern void perform_spin_delay(SpinDelayStatus *status);
845 : extern void finish_spin_delay(SpinDelayStatus *status);
846 :
847 : #endif /* S_LOCK_H */
|
