1 /*
2 * kmp_alloc.cpp -- private/shared dynamic memory allocation and management
3 */
4
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "kmp.h"
14 #include "kmp_io.h"
15 #include "kmp_wrapper_malloc.h"
16
17 // Disable bget when it is not used
18 #if KMP_USE_BGET
19
20 /* Thread private buffer management code */
21
22 typedef int (*bget_compact_t)(size_t, int);
23 typedef void *(*bget_acquire_t)(size_t);
24 typedef void (*bget_release_t)(void *);
25
26 /* NOTE: bufsize must be a signed datatype */
27
28 #if KMP_OS_WINDOWS
29 #if KMP_ARCH_X86 || KMP_ARCH_ARM
30 typedef kmp_int32 bufsize;
31 #else
32 typedef kmp_int64 bufsize;
33 #endif
34 #else
35 typedef ssize_t bufsize;
36 #endif // KMP_OS_WINDOWS
37
38 /* The three modes of operation are, fifo search, lifo search, and best-fit */
39
40 typedef enum bget_mode {
41 bget_mode_fifo = 0,
42 bget_mode_lifo = 1,
43 bget_mode_best = 2
44 } bget_mode_t;
45
46 static void bpool(kmp_info_t *th, void *buffer, bufsize len);
47 static void *bget(kmp_info_t *th, bufsize size);
48 static void *bgetz(kmp_info_t *th, bufsize size);
49 static void *bgetr(kmp_info_t *th, void *buffer, bufsize newsize);
50 static void brel(kmp_info_t *th, void *buf);
51 static void bectl(kmp_info_t *th, bget_compact_t compact,
52 bget_acquire_t acquire, bget_release_t release,
53 bufsize pool_incr);
54
55 /* BGET CONFIGURATION */
56 /* Buffer allocation size quantum: all buffers allocated are a
57 multiple of this size. This MUST be a power of two. */
58
59 /* On IA-32 architecture with Linux* OS, malloc() does not
60 ensure 16 byte alignment */
61
62 #if KMP_ARCH_X86 || !KMP_HAVE_QUAD
63
64 #define SizeQuant 8
65 #define AlignType double
66
67 #else
68
69 #define SizeQuant 16
70 #define AlignType _Quad
71
72 #endif
73
74 // Define this symbol to enable the bstats() function which calculates the
75 // total free space in the buffer pool, the largest available buffer, and the
76 // total space currently allocated.
77 #define BufStats 1
78
79 #ifdef KMP_DEBUG
80
81 // Define this symbol to enable the bpoold() function which dumps the buffers
82 // in a buffer pool.
83 #define BufDump 1
84
85 // Define this symbol to enable the bpoolv() function for validating a buffer
86 // pool.
87 #define BufValid 1
88
89 // Define this symbol to enable the bufdump() function which allows dumping the
90 // contents of an allocated or free buffer.
91 #define DumpData 1
92
93 #ifdef NOT_USED_NOW
94
95 // Wipe free buffers to a guaranteed pattern of garbage to trip up miscreants
96 // who attempt to use pointers into released buffers.
97 #define FreeWipe 1
98
99 // Use a best fit algorithm when searching for space for an allocation request.
100 // This uses memory more efficiently, but allocation will be much slower.
101 #define BestFit 1
102
103 #endif /* NOT_USED_NOW */
104 #endif /* KMP_DEBUG */
105
106 static bufsize bget_bin_size[] = {
107 0,
108 // 1 << 6, /* .5 Cache line */
109 1 << 7, /* 1 Cache line, new */
110 1 << 8, /* 2 Cache lines */
111 1 << 9, /* 4 Cache lines, new */
112 1 << 10, /* 8 Cache lines */
113 1 << 11, /* 16 Cache lines, new */
114 1 << 12, 1 << 13, /* new */
115 1 << 14, 1 << 15, /* new */
116 1 << 16, 1 << 17, 1 << 18, 1 << 19, 1 << 20, /* 1MB */
117 1 << 21, /* 2MB */
118 1 << 22, /* 4MB */
119 1 << 23, /* 8MB */
120 1 << 24, /* 16MB */
121 1 << 25, /* 32MB */
122 };
123
124 #define MAX_BGET_BINS (int)(sizeof(bget_bin_size) / sizeof(bufsize))
125
126 struct bfhead;
127
128 // Declare the interface, including the requested buffer size type, bufsize.
129
130 /* Queue links */
131 typedef struct qlinks {
132 struct bfhead *flink; /* Forward link */
133 struct bfhead *blink; /* Backward link */
134 } qlinks_t;
135
136 /* Header in allocated and free buffers */
137 typedef struct bhead2 {
138 kmp_info_t *bthr; /* The thread which owns the buffer pool */
139 bufsize prevfree; /* Relative link back to previous free buffer in memory or
140 0 if previous buffer is allocated. */
141 bufsize bsize; /* Buffer size: positive if free, negative if allocated. */
142 } bhead2_t;
143
144 /* Make sure the bhead structure is a multiple of SizeQuant in size. */
145 typedef union bhead {
146 KMP_ALIGN(SizeQuant)
147 AlignType b_align;
148 char b_pad[sizeof(bhead2_t) + (SizeQuant - (sizeof(bhead2_t) % SizeQuant))];
149 bhead2_t bb;
150 } bhead_t;
151 #define BH(p) ((bhead_t *)(p))
152
153 /* Header in directly allocated buffers (by acqfcn) */
154 typedef struct bdhead {
155 bufsize tsize; /* Total size, including overhead */
156 bhead_t bh; /* Common header */
157 } bdhead_t;
158 #define BDH(p) ((bdhead_t *)(p))
159
160 /* Header in free buffers */
161 typedef struct bfhead {
162 bhead_t bh; /* Common allocated/free header */
163 qlinks_t ql; /* Links on free list */
164 } bfhead_t;
165 #define BFH(p) ((bfhead_t *)(p))
166
167 typedef struct thr_data {
168 bfhead_t freelist[MAX_BGET_BINS];
169 #if BufStats
170 size_t totalloc; /* Total space currently allocated */
171 long numget, numrel; /* Number of bget() and brel() calls */
172 long numpblk; /* Number of pool blocks */
173 long numpget, numprel; /* Number of block gets and rels */
174 long numdget, numdrel; /* Number of direct gets and rels */
175 #endif /* BufStats */
176
177 /* Automatic expansion block management functions */
178 bget_compact_t compfcn;
179 bget_acquire_t acqfcn;
180 bget_release_t relfcn;
181
182 bget_mode_t mode; /* what allocation mode to use? */
183
184 bufsize exp_incr; /* Expansion block size */
185 bufsize pool_len; /* 0: no bpool calls have been made
186 -1: not all pool blocks are the same size
187 >0: (common) block size for all bpool calls made so far
188 */
189 bfhead_t *last_pool; /* Last pool owned by this thread (delay deallocation) */
190 } thr_data_t;
191
192 /* Minimum allocation quantum: */
193 #define QLSize (sizeof(qlinks_t))
194 #define SizeQ ((SizeQuant > QLSize) ? SizeQuant : QLSize)
195 #define MaxSize \
196 (bufsize)( \
197 ~(((bufsize)(1) << (sizeof(bufsize) * CHAR_BIT - 1)) | (SizeQuant - 1)))
198 // Maximum for the requested size.
199
200 /* End sentinel: value placed in bsize field of dummy block delimiting
201 end of pool block. The most negative number which will fit in a
202 bufsize, defined in a way that the compiler will accept. */
203
204 #define ESent \
205 ((bufsize)(-(((((bufsize)1) << ((int)sizeof(bufsize) * 8 - 2)) - 1) * 2) - 2))
206
207 /* Thread Data management routines */
bget_get_bin(bufsize size)208 static int bget_get_bin(bufsize size) {
209 // binary chop bins
210 int lo = 0, hi = MAX_BGET_BINS - 1;
211
212 KMP_DEBUG_ASSERT(size > 0);
213
214 while ((hi - lo) > 1) {
215 int mid = (lo + hi) >> 1;
216 if (size < bget_bin_size[mid])
217 hi = mid - 1;
218 else
219 lo = mid;
220 }
221
222 KMP_DEBUG_ASSERT((lo >= 0) && (lo < MAX_BGET_BINS));
223
224 return lo;
225 }
226
set_thr_data(kmp_info_t * th)227 static void set_thr_data(kmp_info_t *th) {
228 int i;
229 thr_data_t *data;
230
231 data = (thr_data_t *)((!th->th.th_local.bget_data)
232 ? __kmp_allocate(sizeof(*data))
233 : th->th.th_local.bget_data);
234
235 memset(data, '\0', sizeof(*data));
236
237 for (i = 0; i < MAX_BGET_BINS; ++i) {
238 data->freelist[i].ql.flink = &data->freelist[i];
239 data->freelist[i].ql.blink = &data->freelist[i];
240 }
241
242 th->th.th_local.bget_data = data;
243 th->th.th_local.bget_list = 0;
244 #if !USE_CMP_XCHG_FOR_BGET
245 #ifdef USE_QUEUING_LOCK_FOR_BGET
246 __kmp_init_lock(&th->th.th_local.bget_lock);
247 #else
248 __kmp_init_bootstrap_lock(&th->th.th_local.bget_lock);
249 #endif /* USE_LOCK_FOR_BGET */
250 #endif /* ! USE_CMP_XCHG_FOR_BGET */
251 }
252
get_thr_data(kmp_info_t * th)253 static thr_data_t *get_thr_data(kmp_info_t *th) {
254 thr_data_t *data;
255
256 data = (thr_data_t *)th->th.th_local.bget_data;
257
258 KMP_DEBUG_ASSERT(data != 0);
259
260 return data;
261 }
262
263 /* Walk the free list and release the enqueued buffers */
__kmp_bget_dequeue(kmp_info_t * th)264 static void __kmp_bget_dequeue(kmp_info_t *th) {
265 void *p = TCR_SYNC_PTR(th->th.th_local.bget_list);
266
267 if (p != 0) {
268 #if USE_CMP_XCHG_FOR_BGET
269 {
270 volatile void *old_value = TCR_SYNC_PTR(th->th.th_local.bget_list);
271 while (!KMP_COMPARE_AND_STORE_PTR(&th->th.th_local.bget_list,
272 CCAST(void *, old_value), nullptr)) {
273 KMP_CPU_PAUSE();
274 old_value = TCR_SYNC_PTR(th->th.th_local.bget_list);
275 }
276 p = CCAST(void *, old_value);
277 }
278 #else /* ! USE_CMP_XCHG_FOR_BGET */
279 #ifdef USE_QUEUING_LOCK_FOR_BGET
280 __kmp_acquire_lock(&th->th.th_local.bget_lock, __kmp_gtid_from_thread(th));
281 #else
282 __kmp_acquire_bootstrap_lock(&th->th.th_local.bget_lock);
283 #endif /* USE_QUEUING_LOCK_FOR_BGET */
284
285 p = (void *)th->th.th_local.bget_list;
286 th->th.th_local.bget_list = 0;
287
288 #ifdef USE_QUEUING_LOCK_FOR_BGET
289 __kmp_release_lock(&th->th.th_local.bget_lock, __kmp_gtid_from_thread(th));
290 #else
291 __kmp_release_bootstrap_lock(&th->th.th_local.bget_lock);
292 #endif
293 #endif /* USE_CMP_XCHG_FOR_BGET */
294
295 /* Check again to make sure the list is not empty */
296 while (p != 0) {
297 void *buf = p;
298 bfhead_t *b = BFH(((char *)p) - sizeof(bhead_t));
299
300 KMP_DEBUG_ASSERT(b->bh.bb.bsize != 0);
301 KMP_DEBUG_ASSERT(((kmp_uintptr_t)TCR_PTR(b->bh.bb.bthr) & ~1) ==
302 (kmp_uintptr_t)th); // clear possible mark
303 KMP_DEBUG_ASSERT(b->ql.blink == 0);
304
305 p = (void *)b->ql.flink;
306
307 brel(th, buf);
308 }
309 }
310 }
311
312 /* Chain together the free buffers by using the thread owner field */
__kmp_bget_enqueue(kmp_info_t * th,void * buf,kmp_int32 rel_gtid)313 static void __kmp_bget_enqueue(kmp_info_t *th, void *buf
314 #ifdef USE_QUEUING_LOCK_FOR_BGET
315 ,
316 kmp_int32 rel_gtid
317 #endif
318 ) {
319 bfhead_t *b = BFH(((char *)buf) - sizeof(bhead_t));
320
321 KMP_DEBUG_ASSERT(b->bh.bb.bsize != 0);
322 KMP_DEBUG_ASSERT(((kmp_uintptr_t)TCR_PTR(b->bh.bb.bthr) & ~1) ==
323 (kmp_uintptr_t)th); // clear possible mark
324
325 b->ql.blink = 0;
326
327 KC_TRACE(10, ("__kmp_bget_enqueue: moving buffer to T#%d list\n",
328 __kmp_gtid_from_thread(th)));
329
330 #if USE_CMP_XCHG_FOR_BGET
331 {
332 volatile void *old_value = TCR_PTR(th->th.th_local.bget_list);
333 /* the next pointer must be set before setting bget_list to buf to avoid
334 exposing a broken list to other threads, even for an instant. */
335 b->ql.flink = BFH(CCAST(void *, old_value));
336
337 while (!KMP_COMPARE_AND_STORE_PTR(&th->th.th_local.bget_list,
338 CCAST(void *, old_value), buf)) {
339 KMP_CPU_PAUSE();
340 old_value = TCR_PTR(th->th.th_local.bget_list);
341 /* the next pointer must be set before setting bget_list to buf to avoid
342 exposing a broken list to other threads, even for an instant. */
343 b->ql.flink = BFH(CCAST(void *, old_value));
344 }
345 }
346 #else /* ! USE_CMP_XCHG_FOR_BGET */
347 #ifdef USE_QUEUING_LOCK_FOR_BGET
348 __kmp_acquire_lock(&th->th.th_local.bget_lock, rel_gtid);
349 #else
350 __kmp_acquire_bootstrap_lock(&th->th.th_local.bget_lock);
351 #endif
352
353 b->ql.flink = BFH(th->th.th_local.bget_list);
354 th->th.th_local.bget_list = (void *)buf;
355
356 #ifdef USE_QUEUING_LOCK_FOR_BGET
357 __kmp_release_lock(&th->th.th_local.bget_lock, rel_gtid);
358 #else
359 __kmp_release_bootstrap_lock(&th->th.th_local.bget_lock);
360 #endif
361 #endif /* USE_CMP_XCHG_FOR_BGET */
362 }
363
364 /* insert buffer back onto a new freelist */
__kmp_bget_insert_into_freelist(thr_data_t * thr,bfhead_t * b)365 static void __kmp_bget_insert_into_freelist(thr_data_t *thr, bfhead_t *b) {
366 int bin;
367
368 KMP_DEBUG_ASSERT(((size_t)b) % SizeQuant == 0);
369 KMP_DEBUG_ASSERT(b->bh.bb.bsize % SizeQuant == 0);
370
371 bin = bget_get_bin(b->bh.bb.bsize);
372
373 KMP_DEBUG_ASSERT(thr->freelist[bin].ql.blink->ql.flink ==
374 &thr->freelist[bin]);
375 KMP_DEBUG_ASSERT(thr->freelist[bin].ql.flink->ql.blink ==
376 &thr->freelist[bin]);
377
378 b->ql.flink = &thr->freelist[bin];
379 b->ql.blink = thr->freelist[bin].ql.blink;
380
381 thr->freelist[bin].ql.blink = b;
382 b->ql.blink->ql.flink = b;
383 }
384
385 /* unlink the buffer from the old freelist */
__kmp_bget_remove_from_freelist(bfhead_t * b)386 static void __kmp_bget_remove_from_freelist(bfhead_t *b) {
387 KMP_DEBUG_ASSERT(b->ql.blink->ql.flink == b);
388 KMP_DEBUG_ASSERT(b->ql.flink->ql.blink == b);
389
390 b->ql.blink->ql.flink = b->ql.flink;
391 b->ql.flink->ql.blink = b->ql.blink;
392 }
393
394 /* GET STATS -- check info on free list */
bcheck(kmp_info_t * th,bufsize * max_free,bufsize * total_free)395 static void bcheck(kmp_info_t *th, bufsize *max_free, bufsize *total_free) {
396 thr_data_t *thr = get_thr_data(th);
397 int bin;
398
399 *total_free = *max_free = 0;
400
401 for (bin = 0; bin < MAX_BGET_BINS; ++bin) {
402 bfhead_t *b, *best;
403
404 best = &thr->freelist[bin];
405 b = best->ql.flink;
406
407 while (b != &thr->freelist[bin]) {
408 *total_free += (b->bh.bb.bsize - sizeof(bhead_t));
409 if ((best == &thr->freelist[bin]) || (b->bh.bb.bsize < best->bh.bb.bsize))
410 best = b;
411
412 /* Link to next buffer */
413 b = b->ql.flink;
414 }
415
416 if (*max_free < best->bh.bb.bsize)
417 *max_free = best->bh.bb.bsize;
418 }
419
420 if (*max_free > (bufsize)sizeof(bhead_t))
421 *max_free -= sizeof(bhead_t);
422 }
423
424 /* BGET -- Allocate a buffer. */
bget(kmp_info_t * th,bufsize requested_size)425 static void *bget(kmp_info_t *th, bufsize requested_size) {
426 thr_data_t *thr = get_thr_data(th);
427 bufsize size = requested_size;
428 bfhead_t *b;
429 void *buf;
430 int compactseq = 0;
431 int use_blink = 0;
432 /* For BestFit */
433 bfhead_t *best;
434
435 if (size < 0 || size + sizeof(bhead_t) > MaxSize) {
436 return NULL;
437 }
438
439 __kmp_bget_dequeue(th); /* Release any queued buffers */
440
441 if (size < (bufsize)SizeQ) { // Need at least room for the queue links.
442 size = SizeQ;
443 }
444 #if defined(SizeQuant) && (SizeQuant > 1)
445 size = (size + (SizeQuant - 1)) & (~(SizeQuant - 1));
446 #endif
447
448 size += sizeof(bhead_t); // Add overhead in allocated buffer to size required.
449 KMP_DEBUG_ASSERT(size >= 0);
450 KMP_DEBUG_ASSERT(size % SizeQuant == 0);
451
452 use_blink = (thr->mode == bget_mode_lifo);
453
454 /* If a compact function was provided in the call to bectl(), wrap
455 a loop around the allocation process to allow compaction to
456 intervene in case we don't find a suitable buffer in the chain. */
457
458 for (;;) {
459 int bin;
460
461 for (bin = bget_get_bin(size); bin < MAX_BGET_BINS; ++bin) {
462 /* Link to next buffer */
463 b = (use_blink ? thr->freelist[bin].ql.blink
464 : thr->freelist[bin].ql.flink);
465
466 if (thr->mode == bget_mode_best) {
467 best = &thr->freelist[bin];
468
469 /* Scan the free list searching for the first buffer big enough
470 to hold the requested size buffer. */
471 while (b != &thr->freelist[bin]) {
472 if (b->bh.bb.bsize >= (bufsize)size) {
473 if ((best == &thr->freelist[bin]) ||
474 (b->bh.bb.bsize < best->bh.bb.bsize)) {
475 best = b;
476 }
477 }
478
479 /* Link to next buffer */
480 b = (use_blink ? b->ql.blink : b->ql.flink);
481 }
482 b = best;
483 }
484
485 while (b != &thr->freelist[bin]) {
486 if ((bufsize)b->bh.bb.bsize >= (bufsize)size) {
487
488 // Buffer is big enough to satisfy the request. Allocate it to the
489 // caller. We must decide whether the buffer is large enough to split
490 // into the part given to the caller and a free buffer that remains
491 // on the free list, or whether the entire buffer should be removed
492 // from the free list and given to the caller in its entirety. We
493 // only split the buffer if enough room remains for a header plus the
494 // minimum quantum of allocation.
495 if ((b->bh.bb.bsize - (bufsize)size) >
496 (bufsize)(SizeQ + (sizeof(bhead_t)))) {
497 bhead_t *ba, *bn;
498
499 ba = BH(((char *)b) + (b->bh.bb.bsize - (bufsize)size));
500 bn = BH(((char *)ba) + size);
501
502 KMP_DEBUG_ASSERT(bn->bb.prevfree == b->bh.bb.bsize);
503
504 /* Subtract size from length of free block. */
505 b->bh.bb.bsize -= (bufsize)size;
506
507 /* Link allocated buffer to the previous free buffer. */
508 ba->bb.prevfree = b->bh.bb.bsize;
509
510 /* Plug negative size into user buffer. */
511 ba->bb.bsize = -size;
512
513 /* Mark this buffer as owned by this thread. */
514 TCW_PTR(ba->bb.bthr,
515 th); // not an allocated address (do not mark it)
516 /* Mark buffer after this one not preceded by free block. */
517 bn->bb.prevfree = 0;
518
519 // unlink buffer from old freelist, and reinsert into new freelist
520 __kmp_bget_remove_from_freelist(b);
521 __kmp_bget_insert_into_freelist(thr, b);
522 #if BufStats
523 thr->totalloc += (size_t)size;
524 thr->numget++; /* Increment number of bget() calls */
525 #endif
526 buf = (void *)((((char *)ba) + sizeof(bhead_t)));
527 KMP_DEBUG_ASSERT(((size_t)buf) % SizeQuant == 0);
528 return buf;
529 } else {
530 bhead_t *ba;
531
532 ba = BH(((char *)b) + b->bh.bb.bsize);
533
534 KMP_DEBUG_ASSERT(ba->bb.prevfree == b->bh.bb.bsize);
535
536 /* The buffer isn't big enough to split. Give the whole
537 shebang to the caller and remove it from the free list. */
538
539 __kmp_bget_remove_from_freelist(b);
540 #if BufStats
541 thr->totalloc += (size_t)b->bh.bb.bsize;
542 thr->numget++; /* Increment number of bget() calls */
543 #endif
544 /* Negate size to mark buffer allocated. */
545 b->bh.bb.bsize = -(b->bh.bb.bsize);
546
547 /* Mark this buffer as owned by this thread. */
548 TCW_PTR(ba->bb.bthr, th); // not an allocated address (do not mark)
549 /* Zero the back pointer in the next buffer in memory
550 to indicate that this buffer is allocated. */
551 ba->bb.prevfree = 0;
552
553 /* Give user buffer starting at queue links. */
554 buf = (void *)&(b->ql);
555 KMP_DEBUG_ASSERT(((size_t)buf) % SizeQuant == 0);
556 return buf;
557 }
558 }
559
560 /* Link to next buffer */
561 b = (use_blink ? b->ql.blink : b->ql.flink);
562 }
563 }
564
565 /* We failed to find a buffer. If there's a compact function defined,
566 notify it of the size requested. If it returns TRUE, try the allocation
567 again. */
568
569 if ((thr->compfcn == 0) || (!(*thr->compfcn)(size, ++compactseq))) {
570 break;
571 }
572 }
573
574 /* No buffer available with requested size free. */
575
576 /* Don't give up yet -- look in the reserve supply. */
577 if (thr->acqfcn != 0) {
578 if (size > (bufsize)(thr->exp_incr - sizeof(bhead_t))) {
579 /* Request is too large to fit in a single expansion block.
580 Try to satisfy it by a direct buffer acquisition. */
581 bdhead_t *bdh;
582
583 size += sizeof(bdhead_t) - sizeof(bhead_t);
584
585 KE_TRACE(10, ("%%%%%% MALLOC( %d )\n", (int)size));
586
587 /* richryan */
588 bdh = BDH((*thr->acqfcn)((bufsize)size));
589 if (bdh != NULL) {
590
591 // Mark the buffer special by setting size field of its header to zero.
592 bdh->bh.bb.bsize = 0;
593
594 /* Mark this buffer as owned by this thread. */
595 TCW_PTR(bdh->bh.bb.bthr, th); // don't mark buffer as allocated,
596 // because direct buffer never goes to free list
597 bdh->bh.bb.prevfree = 0;
598 bdh->tsize = size;
599 #if BufStats
600 thr->totalloc += (size_t)size;
601 thr->numget++; /* Increment number of bget() calls */
602 thr->numdget++; /* Direct bget() call count */
603 #endif
604 buf = (void *)(bdh + 1);
605 KMP_DEBUG_ASSERT(((size_t)buf) % SizeQuant == 0);
606 return buf;
607 }
608
609 } else {
610
611 /* Try to obtain a new expansion block */
612 void *newpool;
613
614 KE_TRACE(10, ("%%%%%% MALLOCB( %d )\n", (int)thr->exp_incr));
615
616 /* richryan */
617 newpool = (*thr->acqfcn)((bufsize)thr->exp_incr);
618 KMP_DEBUG_ASSERT(((size_t)newpool) % SizeQuant == 0);
619 if (newpool != NULL) {
620 bpool(th, newpool, thr->exp_incr);
621 buf = bget(
622 th, requested_size); /* This can't, I say, can't get into a loop. */
623 return buf;
624 }
625 }
626 }
627
628 /* Still no buffer available */
629
630 return NULL;
631 }
632
633 /* BGETZ -- Allocate a buffer and clear its contents to zero. We clear
634 the entire contents of the buffer to zero, not just the
635 region requested by the caller. */
636
bgetz(kmp_info_t * th,bufsize size)637 static void *bgetz(kmp_info_t *th, bufsize size) {
638 char *buf = (char *)bget(th, size);
639
640 if (buf != NULL) {
641 bhead_t *b;
642 bufsize rsize;
643
644 b = BH(buf - sizeof(bhead_t));
645 rsize = -(b->bb.bsize);
646 if (rsize == 0) {
647 bdhead_t *bd;
648
649 bd = BDH(buf - sizeof(bdhead_t));
650 rsize = bd->tsize - (bufsize)sizeof(bdhead_t);
651 } else {
652 rsize -= sizeof(bhead_t);
653 }
654
655 KMP_DEBUG_ASSERT(rsize >= size);
656
657 (void)memset(buf, 0, (bufsize)rsize);
658 }
659 return ((void *)buf);
660 }
661
662 /* BGETR -- Reallocate a buffer. This is a minimal implementation,
663 simply in terms of brel() and bget(). It could be
664 enhanced to allow the buffer to grow into adjacent free
665 blocks and to avoid moving data unnecessarily. */
666
bgetr(kmp_info_t * th,void * buf,bufsize size)667 static void *bgetr(kmp_info_t *th, void *buf, bufsize size) {
668 void *nbuf;
669 bufsize osize; /* Old size of buffer */
670 bhead_t *b;
671
672 nbuf = bget(th, size);
673 if (nbuf == NULL) { /* Acquire new buffer */
674 return NULL;
675 }
676 if (buf == NULL) {
677 return nbuf;
678 }
679 b = BH(((char *)buf) - sizeof(bhead_t));
680 osize = -b->bb.bsize;
681 if (osize == 0) {
682 /* Buffer acquired directly through acqfcn. */
683 bdhead_t *bd;
684
685 bd = BDH(((char *)buf) - sizeof(bdhead_t));
686 osize = bd->tsize - (bufsize)sizeof(bdhead_t);
687 } else {
688 osize -= sizeof(bhead_t);
689 }
690
691 KMP_DEBUG_ASSERT(osize > 0);
692
693 (void)KMP_MEMCPY((char *)nbuf, (char *)buf, /* Copy the data */
694 (size_t)((size < osize) ? size : osize));
695 brel(th, buf);
696
697 return nbuf;
698 }
699
700 /* BREL -- Release a buffer. */
brel(kmp_info_t * th,void * buf)701 static void brel(kmp_info_t *th, void *buf) {
702 thr_data_t *thr = get_thr_data(th);
703 bfhead_t *b, *bn;
704 kmp_info_t *bth;
705
706 KMP_DEBUG_ASSERT(buf != NULL);
707 KMP_DEBUG_ASSERT(((size_t)buf) % SizeQuant == 0);
708
709 b = BFH(((char *)buf) - sizeof(bhead_t));
710
711 if (b->bh.bb.bsize == 0) { /* Directly-acquired buffer? */
712 bdhead_t *bdh;
713
714 bdh = BDH(((char *)buf) - sizeof(bdhead_t));
715 KMP_DEBUG_ASSERT(b->bh.bb.prevfree == 0);
716 #if BufStats
717 thr->totalloc -= (size_t)bdh->tsize;
718 thr->numdrel++; /* Number of direct releases */
719 thr->numrel++; /* Increment number of brel() calls */
720 #endif /* BufStats */
721 #ifdef FreeWipe
722 (void)memset((char *)buf, 0x55, (size_t)(bdh->tsize - sizeof(bdhead_t)));
723 #endif /* FreeWipe */
724
725 KE_TRACE(10, ("%%%%%% FREE( %p )\n", (void *)bdh));
726
727 KMP_DEBUG_ASSERT(thr->relfcn != 0);
728 (*thr->relfcn)((void *)bdh); /* Release it directly. */
729 return;
730 }
731
732 bth = (kmp_info_t *)((kmp_uintptr_t)TCR_PTR(b->bh.bb.bthr) &
733 ~1); // clear possible mark before comparison
734 if (bth != th) {
735 /* Add this buffer to be released by the owning thread later */
736 __kmp_bget_enqueue(bth, buf
737 #ifdef USE_QUEUING_LOCK_FOR_BGET
738 ,
739 __kmp_gtid_from_thread(th)
740 #endif
741 );
742 return;
743 }
744
745 /* Buffer size must be negative, indicating that the buffer is allocated. */
746 if (b->bh.bb.bsize >= 0) {
747 bn = NULL;
748 }
749 KMP_DEBUG_ASSERT(b->bh.bb.bsize < 0);
750
751 /* Back pointer in next buffer must be zero, indicating the same thing: */
752
753 KMP_DEBUG_ASSERT(BH((char *)b - b->bh.bb.bsize)->bb.prevfree == 0);
754
755 #if BufStats
756 thr->numrel++; /* Increment number of brel() calls */
757 thr->totalloc += (size_t)b->bh.bb.bsize;
758 #endif
759
760 /* If the back link is nonzero, the previous buffer is free. */
761
762 if (b->bh.bb.prevfree != 0) {
763 /* The previous buffer is free. Consolidate this buffer with it by adding
764 the length of this buffer to the previous free buffer. Note that we
765 subtract the size in the buffer being released, since it's negative to
766 indicate that the buffer is allocated. */
767 bufsize size = b->bh.bb.bsize;
768
769 /* Make the previous buffer the one we're working on. */
770 KMP_DEBUG_ASSERT(BH((char *)b - b->bh.bb.prevfree)->bb.bsize ==
771 b->bh.bb.prevfree);
772 b = BFH(((char *)b) - b->bh.bb.prevfree);
773 b->bh.bb.bsize -= size;
774
775 /* unlink the buffer from the old freelist */
776 __kmp_bget_remove_from_freelist(b);
777 } else {
778 /* The previous buffer isn't allocated. Mark this buffer size as positive
779 (i.e. free) and fall through to place the buffer on the free list as an
780 isolated free block. */
781 b->bh.bb.bsize = -b->bh.bb.bsize;
782 }
783
784 /* insert buffer back onto a new freelist */
785 __kmp_bget_insert_into_freelist(thr, b);
786
787 /* Now we look at the next buffer in memory, located by advancing from
788 the start of this buffer by its size, to see if that buffer is
789 free. If it is, we combine this buffer with the next one in
790 memory, dechaining the second buffer from the free list. */
791 bn = BFH(((char *)b) + b->bh.bb.bsize);
792 if (bn->bh.bb.bsize > 0) {
793
794 /* The buffer is free. Remove it from the free list and add
795 its size to that of our buffer. */
796 KMP_DEBUG_ASSERT(BH((char *)bn + bn->bh.bb.bsize)->bb.prevfree ==
797 bn->bh.bb.bsize);
798
799 __kmp_bget_remove_from_freelist(bn);
800
801 b->bh.bb.bsize += bn->bh.bb.bsize;
802
803 /* unlink the buffer from the old freelist, and reinsert it into the new
804 * freelist */
805 __kmp_bget_remove_from_freelist(b);
806 __kmp_bget_insert_into_freelist(thr, b);
807
808 /* Finally, advance to the buffer that follows the newly
809 consolidated free block. We must set its backpointer to the
810 head of the consolidated free block. We know the next block
811 must be an allocated block because the process of recombination
812 guarantees that two free blocks will never be contiguous in
813 memory. */
814 bn = BFH(((char *)b) + b->bh.bb.bsize);
815 }
816 #ifdef FreeWipe
817 (void)memset(((char *)b) + sizeof(bfhead_t), 0x55,
818 (size_t)(b->bh.bb.bsize - sizeof(bfhead_t)));
819 #endif
820 KMP_DEBUG_ASSERT(bn->bh.bb.bsize < 0);
821
822 /* The next buffer is allocated. Set the backpointer in it to point
823 to this buffer; the previous free buffer in memory. */
824
825 bn->bh.bb.prevfree = b->bh.bb.bsize;
826
827 /* If a block-release function is defined, and this free buffer
828 constitutes the entire block, release it. Note that pool_len
829 is defined in such a way that the test will fail unless all
830 pool blocks are the same size. */
831 if (thr->relfcn != 0 &&
832 b->bh.bb.bsize == (bufsize)(thr->pool_len - sizeof(bhead_t))) {
833 #if BufStats
834 if (thr->numpblk !=
835 1) { /* Do not release the last buffer until finalization time */
836 #endif
837
838 KMP_DEBUG_ASSERT(b->bh.bb.prevfree == 0);
839 KMP_DEBUG_ASSERT(BH((char *)b + b->bh.bb.bsize)->bb.bsize == ESent);
840 KMP_DEBUG_ASSERT(BH((char *)b + b->bh.bb.bsize)->bb.prevfree ==
841 b->bh.bb.bsize);
842
843 /* Unlink the buffer from the free list */
844 __kmp_bget_remove_from_freelist(b);
845
846 KE_TRACE(10, ("%%%%%% FREE( %p )\n", (void *)b));
847
848 (*thr->relfcn)(b);
849 #if BufStats
850 thr->numprel++; /* Nr of expansion block releases */
851 thr->numpblk--; /* Total number of blocks */
852 KMP_DEBUG_ASSERT(thr->numpblk == thr->numpget - thr->numprel);
853
854 // avoid leaving stale last_pool pointer around if it is being dealloced
855 if (thr->last_pool == b)
856 thr->last_pool = 0;
857 } else {
858 thr->last_pool = b;
859 }
860 #endif /* BufStats */
861 }
862 }
863
864 /* BECTL -- Establish automatic pool expansion control */
bectl(kmp_info_t * th,bget_compact_t compact,bget_acquire_t acquire,bget_release_t release,bufsize pool_incr)865 static void bectl(kmp_info_t *th, bget_compact_t compact,
866 bget_acquire_t acquire, bget_release_t release,
867 bufsize pool_incr) {
868 thr_data_t *thr = get_thr_data(th);
869
870 thr->compfcn = compact;
871 thr->acqfcn = acquire;
872 thr->relfcn = release;
873 thr->exp_incr = pool_incr;
874 }
875
876 /* BPOOL -- Add a region of memory to the buffer pool. */
bpool(kmp_info_t * th,void * buf,bufsize len)877 static void bpool(kmp_info_t *th, void *buf, bufsize len) {
878 /* int bin = 0; */
879 thr_data_t *thr = get_thr_data(th);
880 bfhead_t *b = BFH(buf);
881 bhead_t *bn;
882
883 __kmp_bget_dequeue(th); /* Release any queued buffers */
884
885 #ifdef SizeQuant
886 len &= ~(SizeQuant - 1);
887 #endif
888 if (thr->pool_len == 0) {
889 thr->pool_len = len;
890 } else if (len != thr->pool_len) {
891 thr->pool_len = -1;
892 }
893 #if BufStats
894 thr->numpget++; /* Number of block acquisitions */
895 thr->numpblk++; /* Number of blocks total */
896 KMP_DEBUG_ASSERT(thr->numpblk == thr->numpget - thr->numprel);
897 #endif /* BufStats */
898
899 /* Since the block is initially occupied by a single free buffer,
900 it had better not be (much) larger than the largest buffer
901 whose size we can store in bhead.bb.bsize. */
902 KMP_DEBUG_ASSERT(len - sizeof(bhead_t) <= -((bufsize)ESent + 1));
903
904 /* Clear the backpointer at the start of the block to indicate that
905 there is no free block prior to this one. That blocks
906 recombination when the first block in memory is released. */
907 b->bh.bb.prevfree = 0;
908
909 /* Create a dummy allocated buffer at the end of the pool. This dummy
910 buffer is seen when a buffer at the end of the pool is released and
911 blocks recombination of the last buffer with the dummy buffer at
912 the end. The length in the dummy buffer is set to the largest
913 negative number to denote the end of the pool for diagnostic
914 routines (this specific value is not counted on by the actual
915 allocation and release functions). */
916 len -= sizeof(bhead_t);
917 b->bh.bb.bsize = (bufsize)len;
918 /* Set the owner of this buffer */
919 TCW_PTR(b->bh.bb.bthr,
920 (kmp_info_t *)((kmp_uintptr_t)th |
921 1)); // mark the buffer as allocated address
922
923 /* Chain the new block to the free list. */
924 __kmp_bget_insert_into_freelist(thr, b);
925
926 #ifdef FreeWipe
927 (void)memset(((char *)b) + sizeof(bfhead_t), 0x55,
928 (size_t)(len - sizeof(bfhead_t)));
929 #endif
930 bn = BH(((char *)b) + len);
931 bn->bb.prevfree = (bufsize)len;
932 /* Definition of ESent assumes two's complement! */
933 KMP_DEBUG_ASSERT((~0) == -1 && (bn != 0));
934
935 bn->bb.bsize = ESent;
936 }
937
938 /* BFREED -- Dump the free lists for this thread. */
bfreed(kmp_info_t * th)939 static void bfreed(kmp_info_t *th) {
940 int bin = 0, count = 0;
941 int gtid = __kmp_gtid_from_thread(th);
942 thr_data_t *thr = get_thr_data(th);
943
944 #if BufStats
945 __kmp_printf_no_lock("__kmp_printpool: T#%d total=%" KMP_UINT64_SPEC
946 " get=%" KMP_INT64_SPEC " rel=%" KMP_INT64_SPEC
947 " pblk=%" KMP_INT64_SPEC " pget=%" KMP_INT64_SPEC
948 " prel=%" KMP_INT64_SPEC " dget=%" KMP_INT64_SPEC
949 " drel=%" KMP_INT64_SPEC "\n",
950 gtid, (kmp_uint64)thr->totalloc, (kmp_int64)thr->numget,
951 (kmp_int64)thr->numrel, (kmp_int64)thr->numpblk,
952 (kmp_int64)thr->numpget, (kmp_int64)thr->numprel,
953 (kmp_int64)thr->numdget, (kmp_int64)thr->numdrel);
954 #endif
955
956 for (bin = 0; bin < MAX_BGET_BINS; ++bin) {
957 bfhead_t *b;
958
959 for (b = thr->freelist[bin].ql.flink; b != &thr->freelist[bin];
960 b = b->ql.flink) {
961 bufsize bs = b->bh.bb.bsize;
962
963 KMP_DEBUG_ASSERT(b->ql.blink->ql.flink == b);
964 KMP_DEBUG_ASSERT(b->ql.flink->ql.blink == b);
965 KMP_DEBUG_ASSERT(bs > 0);
966
967 count += 1;
968
969 __kmp_printf_no_lock(
970 "__kmp_printpool: T#%d Free block: 0x%p size %6ld bytes.\n", gtid, b,
971 (long)bs);
972 #ifdef FreeWipe
973 {
974 char *lerr = ((char *)b) + sizeof(bfhead_t);
975 if ((bs > sizeof(bfhead_t)) &&
976 ((*lerr != 0x55) ||
977 (memcmp(lerr, lerr + 1, (size_t)(bs - (sizeof(bfhead_t) + 1))) !=
978 0))) {
979 __kmp_printf_no_lock("__kmp_printpool: T#%d (Contents of above "
980 "free block have been overstored.)\n",
981 gtid);
982 }
983 }
984 #endif
985 }
986 }
987
988 if (count == 0)
989 __kmp_printf_no_lock("__kmp_printpool: T#%d No free blocks\n", gtid);
990 }
991
__kmp_initialize_bget(kmp_info_t * th)992 void __kmp_initialize_bget(kmp_info_t *th) {
993 KMP_DEBUG_ASSERT(SizeQuant >= sizeof(void *) && (th != 0));
994
995 set_thr_data(th);
996
997 bectl(th, (bget_compact_t)0, (bget_acquire_t)malloc, (bget_release_t)free,
998 (bufsize)__kmp_malloc_pool_incr);
999 }
1000
__kmp_finalize_bget(kmp_info_t * th)1001 void __kmp_finalize_bget(kmp_info_t *th) {
1002 thr_data_t *thr;
1003 bfhead_t *b;
1004
1005 KMP_DEBUG_ASSERT(th != 0);
1006
1007 #if BufStats
1008 thr = (thr_data_t *)th->th.th_local.bget_data;
1009 KMP_DEBUG_ASSERT(thr != NULL);
1010 b = thr->last_pool;
1011
1012 /* If a block-release function is defined, and this free buffer constitutes
1013 the entire block, release it. Note that pool_len is defined in such a way
1014 that the test will fail unless all pool blocks are the same size. */
1015
1016 // Deallocate the last pool if one exists because we no longer do it in brel()
1017 if (thr->relfcn != 0 && b != 0 && thr->numpblk != 0 &&
1018 b->bh.bb.bsize == (bufsize)(thr->pool_len - sizeof(bhead_t))) {
1019 KMP_DEBUG_ASSERT(b->bh.bb.prevfree == 0);
1020 KMP_DEBUG_ASSERT(BH((char *)b + b->bh.bb.bsize)->bb.bsize == ESent);
1021 KMP_DEBUG_ASSERT(BH((char *)b + b->bh.bb.bsize)->bb.prevfree ==
1022 b->bh.bb.bsize);
1023
1024 /* Unlink the buffer from the free list */
1025 __kmp_bget_remove_from_freelist(b);
1026
1027 KE_TRACE(10, ("%%%%%% FREE( %p )\n", (void *)b));
1028
1029 (*thr->relfcn)(b);
1030 thr->numprel++; /* Nr of expansion block releases */
1031 thr->numpblk--; /* Total number of blocks */
1032 KMP_DEBUG_ASSERT(thr->numpblk == thr->numpget - thr->numprel);
1033 }
1034 #endif /* BufStats */
1035
1036 /* Deallocate bget_data */
1037 if (th->th.th_local.bget_data != NULL) {
1038 __kmp_free(th->th.th_local.bget_data);
1039 th->th.th_local.bget_data = NULL;
1040 }
1041 }
1042
kmpc_set_poolsize(size_t size)1043 void kmpc_set_poolsize(size_t size) {
1044 bectl(__kmp_get_thread(), (bget_compact_t)0, (bget_acquire_t)malloc,
1045 (bget_release_t)free, (bufsize)size);
1046 }
1047
kmpc_get_poolsize(void)1048 size_t kmpc_get_poolsize(void) {
1049 thr_data_t *p;
1050
1051 p = get_thr_data(__kmp_get_thread());
1052
1053 return p->exp_incr;
1054 }
1055
kmpc_set_poolmode(int mode)1056 void kmpc_set_poolmode(int mode) {
1057 thr_data_t *p;
1058
1059 if (mode == bget_mode_fifo || mode == bget_mode_lifo ||
1060 mode == bget_mode_best) {
1061 p = get_thr_data(__kmp_get_thread());
1062 p->mode = (bget_mode_t)mode;
1063 }
1064 }
1065
kmpc_get_poolmode(void)1066 int kmpc_get_poolmode(void) {
1067 thr_data_t *p;
1068
1069 p = get_thr_data(__kmp_get_thread());
1070
1071 return p->mode;
1072 }
1073
kmpc_get_poolstat(size_t * maxmem,size_t * allmem)1074 void kmpc_get_poolstat(size_t *maxmem, size_t *allmem) {
1075 kmp_info_t *th = __kmp_get_thread();
1076 bufsize a, b;
1077
1078 __kmp_bget_dequeue(th); /* Release any queued buffers */
1079
1080 bcheck(th, &a, &b);
1081
1082 *maxmem = a;
1083 *allmem = b;
1084 }
1085
kmpc_poolprint(void)1086 void kmpc_poolprint(void) {
1087 kmp_info_t *th = __kmp_get_thread();
1088
1089 __kmp_bget_dequeue(th); /* Release any queued buffers */
1090
1091 bfreed(th);
1092 }
1093
1094 #endif // #if KMP_USE_BGET
1095
kmpc_malloc(size_t size)1096 void *kmpc_malloc(size_t size) {
1097 void *ptr;
1098 ptr = bget(__kmp_entry_thread(), (bufsize)(size + sizeof(ptr)));
1099 if (ptr != NULL) {
1100 // save allocated pointer just before one returned to user
1101 *(void **)ptr = ptr;
1102 ptr = (void **)ptr + 1;
1103 }
1104 return ptr;
1105 }
1106
1107 #define IS_POWER_OF_TWO(n) (((n) & ((n)-1)) == 0)
1108
kmpc_aligned_malloc(size_t size,size_t alignment)1109 void *kmpc_aligned_malloc(size_t size, size_t alignment) {
1110 void *ptr;
1111 void *ptr_allocated;
1112 KMP_DEBUG_ASSERT(alignment < 32 * 1024); // Alignment should not be too big
1113 if (!IS_POWER_OF_TWO(alignment)) {
1114 // AC: do we need to issue a warning here?
1115 errno = EINVAL;
1116 return NULL;
1117 }
1118 size = size + sizeof(void *) + alignment;
1119 ptr_allocated = bget(__kmp_entry_thread(), (bufsize)size);
1120 if (ptr_allocated != NULL) {
1121 // save allocated pointer just before one returned to user
1122 ptr = (void *)(((kmp_uintptr_t)ptr_allocated + sizeof(void *) + alignment) &
1123 ~(alignment - 1));
1124 *((void **)ptr - 1) = ptr_allocated;
1125 } else {
1126 ptr = NULL;
1127 }
1128 return ptr;
1129 }
1130
kmpc_calloc(size_t nelem,size_t elsize)1131 void *kmpc_calloc(size_t nelem, size_t elsize) {
1132 void *ptr;
1133 ptr = bgetz(__kmp_entry_thread(), (bufsize)(nelem * elsize + sizeof(ptr)));
1134 if (ptr != NULL) {
1135 // save allocated pointer just before one returned to user
1136 *(void **)ptr = ptr;
1137 ptr = (void **)ptr + 1;
1138 }
1139 return ptr;
1140 }
1141
kmpc_realloc(void * ptr,size_t size)1142 void *kmpc_realloc(void *ptr, size_t size) {
1143 void *result = NULL;
1144 if (ptr == NULL) {
1145 // If pointer is NULL, realloc behaves like malloc.
1146 result = bget(__kmp_entry_thread(), (bufsize)(size + sizeof(ptr)));
1147 // save allocated pointer just before one returned to user
1148 if (result != NULL) {
1149 *(void **)result = result;
1150 result = (void **)result + 1;
1151 }
1152 } else if (size == 0) {
1153 // If size is 0, realloc behaves like free.
1154 // The thread must be registered by the call to kmpc_malloc() or
1155 // kmpc_calloc() before.
1156 // So it should be safe to call __kmp_get_thread(), not
1157 // __kmp_entry_thread().
1158 KMP_ASSERT(*((void **)ptr - 1));
1159 brel(__kmp_get_thread(), *((void **)ptr - 1));
1160 } else {
1161 result = bgetr(__kmp_entry_thread(), *((void **)ptr - 1),
1162 (bufsize)(size + sizeof(ptr)));
1163 if (result != NULL) {
1164 *(void **)result = result;
1165 result = (void **)result + 1;
1166 }
1167 }
1168 return result;
1169 }
1170
1171 // NOTE: the library must have already been initialized by a previous allocate
kmpc_free(void * ptr)1172 void kmpc_free(void *ptr) {
1173 if (!__kmp_init_serial) {
1174 return;
1175 }
1176 if (ptr != NULL) {
1177 kmp_info_t *th = __kmp_get_thread();
1178 __kmp_bget_dequeue(th); /* Release any queued buffers */
1179 // extract allocated pointer and free it
1180 KMP_ASSERT(*((void **)ptr - 1));
1181 brel(th, *((void **)ptr - 1));
1182 }
1183 }
1184
___kmp_thread_malloc(kmp_info_t * th,size_t size KMP_SRC_LOC_DECL)1185 void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL) {
1186 void *ptr;
1187 KE_TRACE(30, ("-> __kmp_thread_malloc( %p, %d ) called from %s:%d\n", th,
1188 (int)size KMP_SRC_LOC_PARM));
1189 ptr = bget(th, (bufsize)size);
1190 KE_TRACE(30, ("<- __kmp_thread_malloc() returns %p\n", ptr));
1191 return ptr;
1192 }
1193
___kmp_thread_calloc(kmp_info_t * th,size_t nelem,size_t elsize KMP_SRC_LOC_DECL)1194 void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
1195 size_t elsize KMP_SRC_LOC_DECL) {
1196 void *ptr;
1197 KE_TRACE(30, ("-> __kmp_thread_calloc( %p, %d, %d ) called from %s:%d\n", th,
1198 (int)nelem, (int)elsize KMP_SRC_LOC_PARM));
1199 ptr = bgetz(th, (bufsize)(nelem * elsize));
1200 KE_TRACE(30, ("<- __kmp_thread_calloc() returns %p\n", ptr));
1201 return ptr;
1202 }
1203
___kmp_thread_realloc(kmp_info_t * th,void * ptr,size_t size KMP_SRC_LOC_DECL)1204 void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
1205 size_t size KMP_SRC_LOC_DECL) {
1206 KE_TRACE(30, ("-> __kmp_thread_realloc( %p, %p, %d ) called from %s:%d\n", th,
1207 ptr, (int)size KMP_SRC_LOC_PARM));
1208 ptr = bgetr(th, ptr, (bufsize)size);
1209 KE_TRACE(30, ("<- __kmp_thread_realloc() returns %p\n", ptr));
1210 return ptr;
1211 }
1212
___kmp_thread_free(kmp_info_t * th,void * ptr KMP_SRC_LOC_DECL)1213 void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL) {
1214 KE_TRACE(30, ("-> __kmp_thread_free( %p, %p ) called from %s:%d\n", th,
1215 ptr KMP_SRC_LOC_PARM));
1216 if (ptr != NULL) {
1217 __kmp_bget_dequeue(th); /* Release any queued buffers */
1218 brel(th, ptr);
1219 }
1220 KE_TRACE(30, ("<- __kmp_thread_free()\n"));
1221 }
1222
1223 /* OMP 5.0 Memory Management support */
1224 static const char *kmp_mk_lib_name;
1225 static void *h_memkind;
1226 /* memkind experimental API: */
1227 // memkind_alloc
1228 static void *(*kmp_mk_alloc)(void *k, size_t sz);
1229 // memkind_free
1230 static void (*kmp_mk_free)(void *kind, void *ptr);
1231 // memkind_check_available
1232 static int (*kmp_mk_check)(void *kind);
1233 // kinds we are going to use
1234 static void **mk_default;
1235 static void **mk_interleave;
1236 static void **mk_hbw;
1237 static void **mk_hbw_interleave;
1238 static void **mk_hbw_preferred;
1239 static void **mk_hugetlb;
1240 static void **mk_hbw_hugetlb;
1241 static void **mk_hbw_preferred_hugetlb;
1242 static void **mk_dax_kmem;
1243 static void **mk_dax_kmem_all;
1244 static void **mk_dax_kmem_preferred;
1245
1246 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
chk_kind(void *** pkind)1247 static inline void chk_kind(void ***pkind) {
1248 KMP_DEBUG_ASSERT(pkind);
1249 if (*pkind) // symbol found
1250 if (kmp_mk_check(**pkind)) // kind not available or error
1251 *pkind = NULL;
1252 }
1253 #endif
1254
__kmp_init_memkind()1255 void __kmp_init_memkind() {
1256 // as of 2018-07-31 memkind does not support Windows*, exclude it for now
1257 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
1258 // use of statically linked memkind is problematic, as it depends on libnuma
1259 kmp_mk_lib_name = "libmemkind.so";
1260 h_memkind = dlopen(kmp_mk_lib_name, RTLD_LAZY);
1261 if (h_memkind) {
1262 kmp_mk_check = (int (*)(void *))dlsym(h_memkind, "memkind_check_available");
1263 kmp_mk_alloc =
1264 (void *(*)(void *, size_t))dlsym(h_memkind, "memkind_malloc");
1265 kmp_mk_free = (void (*)(void *, void *))dlsym(h_memkind, "memkind_free");
1266 mk_default = (void **)dlsym(h_memkind, "MEMKIND_DEFAULT");
1267 if (kmp_mk_check && kmp_mk_alloc && kmp_mk_free && mk_default &&
1268 !kmp_mk_check(*mk_default)) {
1269 __kmp_memkind_available = 1;
1270 mk_interleave = (void **)dlsym(h_memkind, "MEMKIND_INTERLEAVE");
1271 chk_kind(&mk_interleave);
1272 mk_hbw = (void **)dlsym(h_memkind, "MEMKIND_HBW");
1273 chk_kind(&mk_hbw);
1274 mk_hbw_interleave = (void **)dlsym(h_memkind, "MEMKIND_HBW_INTERLEAVE");
1275 chk_kind(&mk_hbw_interleave);
1276 mk_hbw_preferred = (void **)dlsym(h_memkind, "MEMKIND_HBW_PREFERRED");
1277 chk_kind(&mk_hbw_preferred);
1278 mk_hugetlb = (void **)dlsym(h_memkind, "MEMKIND_HUGETLB");
1279 chk_kind(&mk_hugetlb);
1280 mk_hbw_hugetlb = (void **)dlsym(h_memkind, "MEMKIND_HBW_HUGETLB");
1281 chk_kind(&mk_hbw_hugetlb);
1282 mk_hbw_preferred_hugetlb =
1283 (void **)dlsym(h_memkind, "MEMKIND_HBW_PREFERRED_HUGETLB");
1284 chk_kind(&mk_hbw_preferred_hugetlb);
1285 mk_dax_kmem = (void **)dlsym(h_memkind, "MEMKIND_DAX_KMEM");
1286 chk_kind(&mk_dax_kmem);
1287 mk_dax_kmem_all = (void **)dlsym(h_memkind, "MEMKIND_DAX_KMEM_ALL");
1288 chk_kind(&mk_dax_kmem_all);
1289 mk_dax_kmem_preferred =
1290 (void **)dlsym(h_memkind, "MEMKIND_DAX_KMEM_PREFERRED");
1291 chk_kind(&mk_dax_kmem_preferred);
1292 KE_TRACE(25, ("__kmp_init_memkind: memkind library initialized\n"));
1293 return; // success
1294 }
1295 dlclose(h_memkind); // failure
1296 }
1297 #else // !(KMP_OS_UNIX && KMP_DYNAMIC_LIB)
1298 kmp_mk_lib_name = "";
1299 #endif // !(KMP_OS_UNIX && KMP_DYNAMIC_LIB)
1300 h_memkind = NULL;
1301 kmp_mk_check = NULL;
1302 kmp_mk_alloc = NULL;
1303 kmp_mk_free = NULL;
1304 mk_default = NULL;
1305 mk_interleave = NULL;
1306 mk_hbw = NULL;
1307 mk_hbw_interleave = NULL;
1308 mk_hbw_preferred = NULL;
1309 mk_hugetlb = NULL;
1310 mk_hbw_hugetlb = NULL;
1311 mk_hbw_preferred_hugetlb = NULL;
1312 mk_dax_kmem = NULL;
1313 mk_dax_kmem_all = NULL;
1314 mk_dax_kmem_preferred = NULL;
1315 }
1316
__kmp_fini_memkind()1317 void __kmp_fini_memkind() {
1318 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
1319 if (__kmp_memkind_available)
1320 KE_TRACE(25, ("__kmp_fini_memkind: finalize memkind library\n"));
1321 if (h_memkind) {
1322 dlclose(h_memkind);
1323 h_memkind = NULL;
1324 }
1325 kmp_mk_check = NULL;
1326 kmp_mk_alloc = NULL;
1327 kmp_mk_free = NULL;
1328 mk_default = NULL;
1329 mk_interleave = NULL;
1330 mk_hbw = NULL;
1331 mk_hbw_interleave = NULL;
1332 mk_hbw_preferred = NULL;
1333 mk_hugetlb = NULL;
1334 mk_hbw_hugetlb = NULL;
1335 mk_hbw_preferred_hugetlb = NULL;
1336 mk_dax_kmem = NULL;
1337 mk_dax_kmem_all = NULL;
1338 mk_dax_kmem_preferred = NULL;
1339 #endif
1340 }
1341
__kmpc_init_allocator(int gtid,omp_memspace_handle_t ms,int ntraits,omp_alloctrait_t traits[])1342 omp_allocator_handle_t __kmpc_init_allocator(int gtid, omp_memspace_handle_t ms,
1343 int ntraits,
1344 omp_alloctrait_t traits[]) {
1345 // OpenMP 5.0 only allows predefined memspaces
1346 KMP_DEBUG_ASSERT(ms == omp_default_mem_space || ms == omp_low_lat_mem_space ||
1347 ms == omp_large_cap_mem_space || ms == omp_const_mem_space ||
1348 ms == omp_high_bw_mem_space);
1349 kmp_allocator_t *al;
1350 int i;
1351 al = (kmp_allocator_t *)__kmp_allocate(sizeof(kmp_allocator_t)); // zeroed
1352 al->memspace = ms; // not used currently
1353 for (i = 0; i < ntraits; ++i) {
1354 switch (traits[i].key) {
1355 case omp_atk_sync_hint:
1356 case omp_atk_access:
1357 case omp_atk_pinned:
1358 break;
1359 case omp_atk_alignment:
1360 __kmp_type_convert(traits[i].value, &(al->alignment));
1361 KMP_ASSERT(IS_POWER_OF_TWO(al->alignment));
1362 break;
1363 case omp_atk_pool_size:
1364 al->pool_size = traits[i].value;
1365 break;
1366 case omp_atk_fallback:
1367 al->fb = (omp_alloctrait_value_t)traits[i].value;
1368 KMP_DEBUG_ASSERT(
1369 al->fb == omp_atv_default_mem_fb || al->fb == omp_atv_null_fb ||
1370 al->fb == omp_atv_abort_fb || al->fb == omp_atv_allocator_fb);
1371 break;
1372 case omp_atk_fb_data:
1373 al->fb_data = RCAST(kmp_allocator_t *, traits[i].value);
1374 break;
1375 case omp_atk_partition:
1376 al->memkind = RCAST(void **, traits[i].value);
1377 break;
1378 default:
1379 KMP_ASSERT2(0, "Unexpected allocator trait");
1380 }
1381 }
1382 if (al->fb == 0) {
1383 // set default allocator
1384 al->fb = omp_atv_default_mem_fb;
1385 al->fb_data = (kmp_allocator_t *)omp_default_mem_alloc;
1386 } else if (al->fb == omp_atv_allocator_fb) {
1387 KMP_ASSERT(al->fb_data != NULL);
1388 } else if (al->fb == omp_atv_default_mem_fb) {
1389 al->fb_data = (kmp_allocator_t *)omp_default_mem_alloc;
1390 }
1391 if (__kmp_memkind_available) {
1392 // Let's use memkind library if available
1393 if (ms == omp_high_bw_mem_space) {
1394 if (al->memkind == (void *)omp_atv_interleaved && mk_hbw_interleave) {
1395 al->memkind = mk_hbw_interleave;
1396 } else if (mk_hbw_preferred) {
1397 // AC: do not try to use MEMKIND_HBW for now, because memkind library
1398 // cannot reliably detect exhaustion of HBW memory.
1399 // It could be possible using hbw_verify_memory_region() but memkind
1400 // manual says: "Using this function in production code may result in
1401 // serious performance penalty".
1402 al->memkind = mk_hbw_preferred;
1403 } else {
1404 // HBW is requested but not available --> return NULL allocator
1405 __kmp_free(al);
1406 return omp_null_allocator;
1407 }
1408 } else if (ms == omp_large_cap_mem_space) {
1409 if (mk_dax_kmem_all) {
1410 // All pmem nodes are visited
1411 al->memkind = mk_dax_kmem_all;
1412 } else if (mk_dax_kmem) {
1413 // Only closest pmem node is visited
1414 al->memkind = mk_dax_kmem;
1415 } else {
1416 __kmp_free(al);
1417 return omp_null_allocator;
1418 }
1419 } else {
1420 if (al->memkind == (void *)omp_atv_interleaved && mk_interleave) {
1421 al->memkind = mk_interleave;
1422 } else {
1423 al->memkind = mk_default;
1424 }
1425 }
1426 } else {
1427 if (ms == omp_high_bw_mem_space) {
1428 // cannot detect HBW memory presence without memkind library
1429 __kmp_free(al);
1430 return omp_null_allocator;
1431 }
1432 }
1433 return (omp_allocator_handle_t)al;
1434 }
1435
__kmpc_destroy_allocator(int gtid,omp_allocator_handle_t allocator)1436 void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t allocator) {
1437 if (allocator > kmp_max_mem_alloc)
1438 __kmp_free(allocator);
1439 }
1440
__kmpc_set_default_allocator(int gtid,omp_allocator_handle_t allocator)1441 void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t allocator) {
1442 if (allocator == omp_null_allocator)
1443 allocator = omp_default_mem_alloc;
1444 __kmp_threads[gtid]->th.th_def_allocator = allocator;
1445 }
1446
__kmpc_get_default_allocator(int gtid)1447 omp_allocator_handle_t __kmpc_get_default_allocator(int gtid) {
1448 return __kmp_threads[gtid]->th.th_def_allocator;
1449 }
1450
1451 typedef struct kmp_mem_desc { // Memory block descriptor
1452 void *ptr_alloc; // Pointer returned by allocator
1453 size_t size_a; // Size of allocated memory block (initial+descriptor+align)
1454 size_t size_orig; // Original size requested
1455 void *ptr_align; // Pointer to aligned memory, returned
1456 kmp_allocator_t *allocator; // allocator
1457 } kmp_mem_desc_t;
1458 static int alignment = sizeof(void *); // let's align to pointer size
1459
__kmpc_alloc(int gtid,size_t size,omp_allocator_handle_t allocator)1460 void *__kmpc_alloc(int gtid, size_t size, omp_allocator_handle_t allocator) {
1461 void *ptr = NULL;
1462 kmp_allocator_t *al;
1463 KMP_DEBUG_ASSERT(__kmp_init_serial);
1464
1465 if (size == 0)
1466 return NULL;
1467
1468 if (allocator == omp_null_allocator)
1469 allocator = __kmp_threads[gtid]->th.th_def_allocator;
1470
1471 KE_TRACE(25, ("__kmpc_alloc: T#%d (%d, %p)\n", gtid, (int)size, allocator));
1472 al = RCAST(kmp_allocator_t *, CCAST(omp_allocator_handle_t, allocator));
1473
1474 int sz_desc = sizeof(kmp_mem_desc_t);
1475 kmp_mem_desc_t desc;
1476 kmp_uintptr_t addr; // address returned by allocator
1477 kmp_uintptr_t addr_align; // address to return to caller
1478 kmp_uintptr_t addr_descr; // address of memory block descriptor
1479 int align = alignment; // default alignment
1480 if (allocator > kmp_max_mem_alloc && al->alignment > 0) {
1481 align = al->alignment; // alignment requested by user
1482 }
1483 desc.size_orig = size;
1484 desc.size_a = size + sz_desc + align;
1485
1486 if (__kmp_memkind_available) {
1487 if (allocator < kmp_max_mem_alloc) {
1488 // pre-defined allocator
1489 if (allocator == omp_high_bw_mem_alloc && mk_hbw_preferred) {
1490 ptr = kmp_mk_alloc(*mk_hbw_preferred, desc.size_a);
1491 } else if (allocator == omp_large_cap_mem_alloc && mk_dax_kmem_all) {
1492 ptr = kmp_mk_alloc(*mk_dax_kmem_all, desc.size_a);
1493 } else {
1494 ptr = kmp_mk_alloc(*mk_default, desc.size_a);
1495 }
1496 } else if (al->pool_size > 0) {
1497 // custom allocator with pool size requested
1498 kmp_uint64 used =
1499 KMP_TEST_THEN_ADD64((kmp_int64 *)&al->pool_used, desc.size_a);
1500 if (used + desc.size_a > al->pool_size) {
1501 // not enough space, need to go fallback path
1502 KMP_TEST_THEN_ADD64((kmp_int64 *)&al->pool_used, -desc.size_a);
1503 if (al->fb == omp_atv_default_mem_fb) {
1504 al = (kmp_allocator_t *)omp_default_mem_alloc;
1505 ptr = kmp_mk_alloc(*mk_default, desc.size_a);
1506 } else if (al->fb == omp_atv_abort_fb) {
1507 KMP_ASSERT(0); // abort fallback requested
1508 } else if (al->fb == omp_atv_allocator_fb) {
1509 KMP_ASSERT(al != al->fb_data);
1510 al = al->fb_data;
1511 return __kmpc_alloc(gtid, size, (omp_allocator_handle_t)al);
1512 } // else ptr == NULL;
1513 } else {
1514 // pool has enough space
1515 ptr = kmp_mk_alloc(*al->memkind, desc.size_a);
1516 if (ptr == NULL) {
1517 if (al->fb == omp_atv_default_mem_fb) {
1518 al = (kmp_allocator_t *)omp_default_mem_alloc;
1519 ptr = kmp_mk_alloc(*mk_default, desc.size_a);
1520 } else if (al->fb == omp_atv_abort_fb) {
1521 KMP_ASSERT(0); // abort fallback requested
1522 } else if (al->fb == omp_atv_allocator_fb) {
1523 KMP_ASSERT(al != al->fb_data);
1524 al = al->fb_data;
1525 return __kmpc_alloc(gtid, size, (omp_allocator_handle_t)al);
1526 }
1527 }
1528 }
1529 } else {
1530 // custom allocator, pool size not requested
1531 ptr = kmp_mk_alloc(*al->memkind, desc.size_a);
1532 if (ptr == NULL) {
1533 if (al->fb == omp_atv_default_mem_fb) {
1534 al = (kmp_allocator_t *)omp_default_mem_alloc;
1535 ptr = kmp_mk_alloc(*mk_default, desc.size_a);
1536 } else if (al->fb == omp_atv_abort_fb) {
1537 KMP_ASSERT(0); // abort fallback requested
1538 } else if (al->fb == omp_atv_allocator_fb) {
1539 KMP_ASSERT(al != al->fb_data);
1540 al = al->fb_data;
1541 return __kmpc_alloc(gtid, size, (omp_allocator_handle_t)al);
1542 }
1543 }
1544 }
1545 } else if (allocator < kmp_max_mem_alloc) {
1546 // pre-defined allocator
1547 if (allocator == omp_high_bw_mem_alloc) {
1548 // ptr = NULL;
1549 } else if (allocator == omp_large_cap_mem_alloc) {
1550 // warnings?
1551 } else {
1552 ptr = __kmp_thread_malloc(__kmp_thread_from_gtid(gtid), desc.size_a);
1553 }
1554 } else if (al->pool_size > 0) {
1555 // custom allocator with pool size requested
1556 kmp_uint64 used =
1557 KMP_TEST_THEN_ADD64((kmp_int64 *)&al->pool_used, desc.size_a);
1558 if (used + desc.size_a > al->pool_size) {
1559 // not enough space, need to go fallback path
1560 KMP_TEST_THEN_ADD64((kmp_int64 *)&al->pool_used, -desc.size_a);
1561 if (al->fb == omp_atv_default_mem_fb) {
1562 al = (kmp_allocator_t *)omp_default_mem_alloc;
1563 ptr = __kmp_thread_malloc(__kmp_thread_from_gtid(gtid), desc.size_a);
1564 } else if (al->fb == omp_atv_abort_fb) {
1565 KMP_ASSERT(0); // abort fallback requested
1566 } else if (al->fb == omp_atv_allocator_fb) {
1567 KMP_ASSERT(al != al->fb_data);
1568 al = al->fb_data;
1569 return __kmpc_alloc(gtid, size, (omp_allocator_handle_t)al);
1570 } // else ptr == NULL;
1571 } else {
1572 // pool has enough space
1573 ptr = __kmp_thread_malloc(__kmp_thread_from_gtid(gtid), desc.size_a);
1574 if (ptr == NULL && al->fb == omp_atv_abort_fb) {
1575 KMP_ASSERT(0); // abort fallback requested
1576 } // no sense to look for another fallback because of same internal alloc
1577 }
1578 } else {
1579 // custom allocator, pool size not requested
1580 ptr = __kmp_thread_malloc(__kmp_thread_from_gtid(gtid), desc.size_a);
1581 if (ptr == NULL && al->fb == omp_atv_abort_fb) {
1582 KMP_ASSERT(0); // abort fallback requested
1583 } // no sense to look for another fallback because of same internal alloc
1584 }
1585 KE_TRACE(10, ("__kmpc_alloc: T#%d %p=alloc(%d)\n", gtid, ptr, desc.size_a));
1586 if (ptr == NULL)
1587 return NULL;
1588
1589 addr = (kmp_uintptr_t)ptr;
1590 addr_align = (addr + sz_desc + align - 1) & ~(align - 1);
1591 addr_descr = addr_align - sz_desc;
1592
1593 desc.ptr_alloc = ptr;
1594 desc.ptr_align = (void *)addr_align;
1595 desc.allocator = al;
1596 *((kmp_mem_desc_t *)addr_descr) = desc; // save descriptor contents
1597 KMP_MB();
1598
1599 KE_TRACE(25, ("__kmpc_alloc returns %p, T#%d\n", desc.ptr_align, gtid));
1600 return desc.ptr_align;
1601 }
1602
__kmpc_calloc(int gtid,size_t nmemb,size_t size,omp_allocator_handle_t allocator)1603 void *__kmpc_calloc(int gtid, size_t nmemb, size_t size,
1604 omp_allocator_handle_t allocator) {
1605 void *ptr = NULL;
1606 kmp_allocator_t *al;
1607 KMP_DEBUG_ASSERT(__kmp_init_serial);
1608
1609 if (allocator == omp_null_allocator)
1610 allocator = __kmp_threads[gtid]->th.th_def_allocator;
1611
1612 KE_TRACE(25, ("__kmpc_calloc: T#%d (%d, %d, %p)\n", gtid, (int)nmemb,
1613 (int)size, allocator));
1614
1615 al = RCAST(kmp_allocator_t *, CCAST(omp_allocator_handle_t, allocator));
1616
1617 if (nmemb == 0 || size == 0)
1618 return ptr;
1619
1620 if ((SIZE_MAX - sizeof(kmp_mem_desc_t)) / size < nmemb) {
1621 if (al->fb == omp_atv_abort_fb) {
1622 KMP_ASSERT(0);
1623 }
1624 return ptr;
1625 }
1626
1627 ptr = __kmpc_alloc(gtid, nmemb * size, allocator);
1628
1629 if (ptr) {
1630 memset(ptr, 0x00, nmemb * size);
1631 }
1632 KE_TRACE(25, ("__kmpc_calloc returns %p, T#%d\n", ptr, gtid));
1633 return ptr;
1634 }
1635
__kmpc_realloc(int gtid,void * ptr,size_t size,omp_allocator_handle_t allocator,omp_allocator_handle_t free_allocator)1636 void *__kmpc_realloc(int gtid, void *ptr, size_t size,
1637 omp_allocator_handle_t allocator,
1638 omp_allocator_handle_t free_allocator) {
1639 void *nptr = NULL;
1640 KMP_DEBUG_ASSERT(__kmp_init_serial);
1641
1642 if (size == 0) {
1643 if (ptr != NULL)
1644 __kmpc_free(gtid, ptr, free_allocator);
1645 return nptr;
1646 }
1647
1648 KE_TRACE(25, ("__kmpc_realloc: T#%d (%p, %d, %p, %p)\n", gtid, ptr, (int)size,
1649 allocator, free_allocator));
1650
1651 nptr = __kmpc_alloc(gtid, size, allocator);
1652
1653 if (nptr != NULL && ptr != NULL) {
1654 kmp_mem_desc_t desc;
1655 kmp_uintptr_t addr_align; // address to return to caller
1656 kmp_uintptr_t addr_descr; // address of memory block descriptor
1657
1658 addr_align = (kmp_uintptr_t)ptr;
1659 addr_descr = addr_align - sizeof(kmp_mem_desc_t);
1660 desc = *((kmp_mem_desc_t *)addr_descr); // read descriptor
1661
1662 KMP_DEBUG_ASSERT(desc.ptr_align == ptr);
1663 KMP_DEBUG_ASSERT(desc.size_orig > 0);
1664 KMP_DEBUG_ASSERT(desc.size_orig < desc.size_a);
1665 KMP_MEMCPY((char *)nptr, (char *)ptr,
1666 (size_t)((size < desc.size_orig) ? size : desc.size_orig));
1667 }
1668
1669 if (nptr != NULL) {
1670 __kmpc_free(gtid, ptr, free_allocator);
1671 }
1672
1673 KE_TRACE(25, ("__kmpc_realloc returns %p, T#%d\n", nptr, gtid));
1674 return nptr;
1675 }
1676
__kmpc_free(int gtid,void * ptr,const omp_allocator_handle_t allocator)1677 void __kmpc_free(int gtid, void *ptr, const omp_allocator_handle_t allocator) {
1678 KE_TRACE(25, ("__kmpc_free: T#%d free(%p,%p)\n", gtid, ptr, allocator));
1679 if (ptr == NULL)
1680 return;
1681
1682 kmp_allocator_t *al;
1683 omp_allocator_handle_t oal;
1684 al = RCAST(kmp_allocator_t *, CCAST(omp_allocator_handle_t, allocator));
1685 kmp_mem_desc_t desc;
1686 kmp_uintptr_t addr_align; // address to return to caller
1687 kmp_uintptr_t addr_descr; // address of memory block descriptor
1688
1689 addr_align = (kmp_uintptr_t)ptr;
1690 addr_descr = addr_align - sizeof(kmp_mem_desc_t);
1691 desc = *((kmp_mem_desc_t *)addr_descr); // read descriptor
1692
1693 KMP_DEBUG_ASSERT(desc.ptr_align == ptr);
1694 if (allocator) {
1695 KMP_DEBUG_ASSERT(desc.allocator == al || desc.allocator == al->fb_data);
1696 }
1697 al = desc.allocator;
1698 oal = (omp_allocator_handle_t)al; // cast to void* for comparisons
1699 KMP_DEBUG_ASSERT(al);
1700
1701 if (__kmp_memkind_available) {
1702 if (oal < kmp_max_mem_alloc) {
1703 // pre-defined allocator
1704 if (oal == omp_high_bw_mem_alloc && mk_hbw_preferred) {
1705 kmp_mk_free(*mk_hbw_preferred, desc.ptr_alloc);
1706 } else if (oal == omp_large_cap_mem_alloc && mk_dax_kmem_all) {
1707 kmp_mk_free(*mk_dax_kmem_all, desc.ptr_alloc);
1708 } else {
1709 kmp_mk_free(*mk_default, desc.ptr_alloc);
1710 }
1711 } else {
1712 if (al->pool_size > 0) { // custom allocator with pool size requested
1713 kmp_uint64 used =
1714 KMP_TEST_THEN_ADD64((kmp_int64 *)&al->pool_used, -desc.size_a);
1715 (void)used; // to suppress compiler warning
1716 KMP_DEBUG_ASSERT(used >= desc.size_a);
1717 }
1718 kmp_mk_free(*al->memkind, desc.ptr_alloc);
1719 }
1720 } else {
1721 if (oal > kmp_max_mem_alloc && al->pool_size > 0) {
1722 kmp_uint64 used =
1723 KMP_TEST_THEN_ADD64((kmp_int64 *)&al->pool_used, -desc.size_a);
1724 (void)used; // to suppress compiler warning
1725 KMP_DEBUG_ASSERT(used >= desc.size_a);
1726 }
1727 __kmp_thread_free(__kmp_thread_from_gtid(gtid), desc.ptr_alloc);
1728 }
1729 KE_TRACE(10, ("__kmpc_free: T#%d freed %p (%p)\n", gtid, desc.ptr_alloc,
1730 allocator));
1731 }
1732
1733 /* If LEAK_MEMORY is defined, __kmp_free() will *not* free memory. It causes
1734 memory leaks, but it may be useful for debugging memory corruptions, used
1735 freed pointers, etc. */
1736 /* #define LEAK_MEMORY */
1737 struct kmp_mem_descr { // Memory block descriptor.
1738 void *ptr_allocated; // Pointer returned by malloc(), subject for free().
1739 size_t size_allocated; // Size of allocated memory block.
1740 void *ptr_aligned; // Pointer to aligned memory, to be used by client code.
1741 size_t size_aligned; // Size of aligned memory block.
1742 };
1743 typedef struct kmp_mem_descr kmp_mem_descr_t;
1744
1745 /* Allocate memory on requested boundary, fill allocated memory with 0x00.
1746 NULL is NEVER returned, __kmp_abort() is called in case of memory allocation
1747 error. Must use __kmp_free when freeing memory allocated by this routine! */
___kmp_allocate_align(size_t size,size_t alignment KMP_SRC_LOC_DECL)1748 static void *___kmp_allocate_align(size_t size,
1749 size_t alignment KMP_SRC_LOC_DECL) {
1750 /* __kmp_allocate() allocates (by call to malloc()) bigger memory block than
1751 requested to return properly aligned pointer. Original pointer returned
1752 by malloc() and size of allocated block is saved in descriptor just
1753 before the aligned pointer. This information used by __kmp_free() -- it
1754 has to pass to free() original pointer, not aligned one.
1755
1756 +---------+------------+-----------------------------------+---------+
1757 | padding | descriptor | aligned block | padding |
1758 +---------+------------+-----------------------------------+---------+
1759 ^ ^
1760 | |
1761 | +- Aligned pointer returned to caller
1762 +- Pointer returned by malloc()
1763
1764 Aligned block is filled with zeros, paddings are filled with 0xEF. */
1765
1766 kmp_mem_descr_t descr;
1767 kmp_uintptr_t addr_allocated; // Address returned by malloc().
1768 kmp_uintptr_t addr_aligned; // Aligned address to return to caller.
1769 kmp_uintptr_t addr_descr; // Address of memory block descriptor.
1770
1771 KE_TRACE(25, ("-> ___kmp_allocate_align( %d, %d ) called from %s:%d\n",
1772 (int)size, (int)alignment KMP_SRC_LOC_PARM));
1773
1774 KMP_DEBUG_ASSERT(alignment < 32 * 1024); // Alignment should not be too
1775 KMP_DEBUG_ASSERT(sizeof(void *) <= sizeof(kmp_uintptr_t));
1776 // Make sure kmp_uintptr_t is enough to store addresses.
1777
1778 descr.size_aligned = size;
1779 descr.size_allocated =
1780 descr.size_aligned + sizeof(kmp_mem_descr_t) + alignment;
1781
1782 #if KMP_DEBUG
1783 descr.ptr_allocated = _malloc_src_loc(descr.size_allocated, _file_, _line_);
1784 #else
1785 descr.ptr_allocated = malloc_src_loc(descr.size_allocated KMP_SRC_LOC_PARM);
1786 #endif
1787 KE_TRACE(10, (" malloc( %d ) returned %p\n", (int)descr.size_allocated,
1788 descr.ptr_allocated));
1789 if (descr.ptr_allocated == NULL) {
1790 KMP_FATAL(OutOfHeapMemory);
1791 }
1792
1793 addr_allocated = (kmp_uintptr_t)descr.ptr_allocated;
1794 addr_aligned =
1795 (addr_allocated + sizeof(kmp_mem_descr_t) + alignment) & ~(alignment - 1);
1796 addr_descr = addr_aligned - sizeof(kmp_mem_descr_t);
1797
1798 descr.ptr_aligned = (void *)addr_aligned;
1799
1800 KE_TRACE(26, (" ___kmp_allocate_align: "
1801 "ptr_allocated=%p, size_allocated=%d, "
1802 "ptr_aligned=%p, size_aligned=%d\n",
1803 descr.ptr_allocated, (int)descr.size_allocated,
1804 descr.ptr_aligned, (int)descr.size_aligned));
1805
1806 KMP_DEBUG_ASSERT(addr_allocated <= addr_descr);
1807 KMP_DEBUG_ASSERT(addr_descr + sizeof(kmp_mem_descr_t) == addr_aligned);
1808 KMP_DEBUG_ASSERT(addr_aligned + descr.size_aligned <=
1809 addr_allocated + descr.size_allocated);
1810 KMP_DEBUG_ASSERT(addr_aligned % alignment == 0);
1811 #ifdef KMP_DEBUG
1812 memset(descr.ptr_allocated, 0xEF, descr.size_allocated);
1813 // Fill allocated memory block with 0xEF.
1814 #endif
1815 memset(descr.ptr_aligned, 0x00, descr.size_aligned);
1816 // Fill the aligned memory block (which is intended for using by caller) with
1817 // 0x00. Do not
1818 // put this filling under KMP_DEBUG condition! Many callers expect zeroed
1819 // memory. (Padding
1820 // bytes remain filled with 0xEF in debugging library.)
1821 *((kmp_mem_descr_t *)addr_descr) = descr;
1822
1823 KMP_MB();
1824
1825 KE_TRACE(25, ("<- ___kmp_allocate_align() returns %p\n", descr.ptr_aligned));
1826 return descr.ptr_aligned;
1827 } // func ___kmp_allocate_align
1828
1829 /* Allocate memory on cache line boundary, fill allocated memory with 0x00.
1830 Do not call this func directly! Use __kmp_allocate macro instead.
1831 NULL is NEVER returned, __kmp_abort() is called in case of memory allocation
1832 error. Must use __kmp_free when freeing memory allocated by this routine! */
___kmp_allocate(size_t size KMP_SRC_LOC_DECL)1833 void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL) {
1834 void *ptr;
1835 KE_TRACE(25, ("-> __kmp_allocate( %d ) called from %s:%d\n",
1836 (int)size KMP_SRC_LOC_PARM));
1837 ptr = ___kmp_allocate_align(size, __kmp_align_alloc KMP_SRC_LOC_PARM);
1838 KE_TRACE(25, ("<- __kmp_allocate() returns %p\n", ptr));
1839 return ptr;
1840 } // func ___kmp_allocate
1841
1842 /* Allocate memory on page boundary, fill allocated memory with 0x00.
1843 Does not call this func directly! Use __kmp_page_allocate macro instead.
1844 NULL is NEVER returned, __kmp_abort() is called in case of memory allocation
1845 error. Must use __kmp_free when freeing memory allocated by this routine! */
___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL)1846 void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL) {
1847 int page_size = 8 * 1024;
1848 void *ptr;
1849
1850 KE_TRACE(25, ("-> __kmp_page_allocate( %d ) called from %s:%d\n",
1851 (int)size KMP_SRC_LOC_PARM));
1852 ptr = ___kmp_allocate_align(size, page_size KMP_SRC_LOC_PARM);
1853 KE_TRACE(25, ("<- __kmp_page_allocate( %d ) returns %p\n", (int)size, ptr));
1854 return ptr;
1855 } // ___kmp_page_allocate
1856
1857 /* Free memory allocated by __kmp_allocate() and __kmp_page_allocate().
1858 In debug mode, fill the memory block with 0xEF before call to free(). */
___kmp_free(void * ptr KMP_SRC_LOC_DECL)1859 void ___kmp_free(void *ptr KMP_SRC_LOC_DECL) {
1860 kmp_mem_descr_t descr;
1861 kmp_uintptr_t addr_allocated; // Address returned by malloc().
1862 kmp_uintptr_t addr_aligned; // Aligned address passed by caller.
1863
1864 KE_TRACE(25,
1865 ("-> __kmp_free( %p ) called from %s:%d\n", ptr KMP_SRC_LOC_PARM));
1866 KMP_ASSERT(ptr != NULL);
1867
1868 descr = *(kmp_mem_descr_t *)((kmp_uintptr_t)ptr - sizeof(kmp_mem_descr_t));
1869
1870 KE_TRACE(26, (" __kmp_free: "
1871 "ptr_allocated=%p, size_allocated=%d, "
1872 "ptr_aligned=%p, size_aligned=%d\n",
1873 descr.ptr_allocated, (int)descr.size_allocated,
1874 descr.ptr_aligned, (int)descr.size_aligned));
1875
1876 addr_allocated = (kmp_uintptr_t)descr.ptr_allocated;
1877 addr_aligned = (kmp_uintptr_t)descr.ptr_aligned;
1878
1879 KMP_DEBUG_ASSERT(addr_aligned % CACHE_LINE == 0);
1880 KMP_DEBUG_ASSERT(descr.ptr_aligned == ptr);
1881 KMP_DEBUG_ASSERT(addr_allocated + sizeof(kmp_mem_descr_t) <= addr_aligned);
1882 KMP_DEBUG_ASSERT(descr.size_aligned < descr.size_allocated);
1883 KMP_DEBUG_ASSERT(addr_aligned + descr.size_aligned <=
1884 addr_allocated + descr.size_allocated);
1885
1886 #ifdef KMP_DEBUG
1887 memset(descr.ptr_allocated, 0xEF, descr.size_allocated);
1888 // Fill memory block with 0xEF, it helps catch using freed memory.
1889 #endif
1890
1891 #ifndef LEAK_MEMORY
1892 KE_TRACE(10, (" free( %p )\n", descr.ptr_allocated));
1893 #ifdef KMP_DEBUG
1894 _free_src_loc(descr.ptr_allocated, _file_, _line_);
1895 #else
1896 free_src_loc(descr.ptr_allocated KMP_SRC_LOC_PARM);
1897 #endif
1898 #endif
1899 KMP_MB();
1900 KE_TRACE(25, ("<- __kmp_free() returns\n"));
1901 } // func ___kmp_free
1902
1903 #if USE_FAST_MEMORY == 3
1904 // Allocate fast memory by first scanning the thread's free lists
1905 // If a chunk the right size exists, grab it off the free list.
1906 // Otherwise allocate normally using kmp_thread_malloc.
1907
1908 // AC: How to choose the limit? Just get 16 for now...
1909 #define KMP_FREE_LIST_LIMIT 16
1910
1911 // Always use 128 bytes for determining buckets for caching memory blocks
1912 #define DCACHE_LINE 128
1913
___kmp_fast_allocate(kmp_info_t * this_thr,size_t size KMP_SRC_LOC_DECL)1914 void *___kmp_fast_allocate(kmp_info_t *this_thr, size_t size KMP_SRC_LOC_DECL) {
1915 void *ptr;
1916 size_t num_lines, idx;
1917 int index;
1918 void *alloc_ptr;
1919 size_t alloc_size;
1920 kmp_mem_descr_t *descr;
1921
1922 KE_TRACE(25, ("-> __kmp_fast_allocate( T#%d, %d ) called from %s:%d\n",
1923 __kmp_gtid_from_thread(this_thr), (int)size KMP_SRC_LOC_PARM));
1924
1925 num_lines = (size + DCACHE_LINE - 1) / DCACHE_LINE;
1926 idx = num_lines - 1;
1927 KMP_DEBUG_ASSERT(idx >= 0);
1928 if (idx < 2) {
1929 index = 0; // idx is [ 0, 1 ], use first free list
1930 num_lines = 2; // 1, 2 cache lines or less than cache line
1931 } else if ((idx >>= 2) == 0) {
1932 index = 1; // idx is [ 2, 3 ], use second free list
1933 num_lines = 4; // 3, 4 cache lines
1934 } else if ((idx >>= 2) == 0) {
1935 index = 2; // idx is [ 4, 15 ], use third free list
1936 num_lines = 16; // 5, 6, ..., 16 cache lines
1937 } else if ((idx >>= 2) == 0) {
1938 index = 3; // idx is [ 16, 63 ], use fourth free list
1939 num_lines = 64; // 17, 18, ..., 64 cache lines
1940 } else {
1941 goto alloc_call; // 65 or more cache lines ( > 8KB ), don't use free lists
1942 }
1943
1944 ptr = this_thr->th.th_free_lists[index].th_free_list_self;
1945 if (ptr != NULL) {
1946 // pop the head of no-sync free list
1947 this_thr->th.th_free_lists[index].th_free_list_self = *((void **)ptr);
1948 KMP_DEBUG_ASSERT(
1949 this_thr ==
1950 ((kmp_mem_descr_t *)((kmp_uintptr_t)ptr - sizeof(kmp_mem_descr_t)))
1951 ->ptr_aligned);
1952 goto end;
1953 }
1954 ptr = TCR_SYNC_PTR(this_thr->th.th_free_lists[index].th_free_list_sync);
1955 if (ptr != NULL) {
1956 // no-sync free list is empty, use sync free list (filled in by other
1957 // threads only)
1958 // pop the head of the sync free list, push NULL instead
1959 while (!KMP_COMPARE_AND_STORE_PTR(
1960 &this_thr->th.th_free_lists[index].th_free_list_sync, ptr, nullptr)) {
1961 KMP_CPU_PAUSE();
1962 ptr = TCR_SYNC_PTR(this_thr->th.th_free_lists[index].th_free_list_sync);
1963 }
1964 // push the rest of chain into no-sync free list (can be NULL if there was
1965 // the only block)
1966 this_thr->th.th_free_lists[index].th_free_list_self = *((void **)ptr);
1967 KMP_DEBUG_ASSERT(
1968 this_thr ==
1969 ((kmp_mem_descr_t *)((kmp_uintptr_t)ptr - sizeof(kmp_mem_descr_t)))
1970 ->ptr_aligned);
1971 goto end;
1972 }
1973
1974 alloc_call:
1975 // haven't found block in the free lists, thus allocate it
1976 size = num_lines * DCACHE_LINE;
1977
1978 alloc_size = size + sizeof(kmp_mem_descr_t) + DCACHE_LINE;
1979 KE_TRACE(25, ("__kmp_fast_allocate: T#%d Calling __kmp_thread_malloc with "
1980 "alloc_size %d\n",
1981 __kmp_gtid_from_thread(this_thr), alloc_size));
1982 alloc_ptr = bget(this_thr, (bufsize)alloc_size);
1983
1984 // align ptr to DCACHE_LINE
1985 ptr = (void *)((((kmp_uintptr_t)alloc_ptr) + sizeof(kmp_mem_descr_t) +
1986 DCACHE_LINE) &
1987 ~(DCACHE_LINE - 1));
1988 descr = (kmp_mem_descr_t *)(((kmp_uintptr_t)ptr) - sizeof(kmp_mem_descr_t));
1989
1990 descr->ptr_allocated = alloc_ptr; // remember allocated pointer
1991 // we don't need size_allocated
1992 descr->ptr_aligned = (void *)this_thr; // remember allocating thread
1993 // (it is already saved in bget buffer,
1994 // but we may want to use another allocator in future)
1995 descr->size_aligned = size;
1996
1997 end:
1998 KE_TRACE(25, ("<- __kmp_fast_allocate( T#%d ) returns %p\n",
1999 __kmp_gtid_from_thread(this_thr), ptr));
2000 return ptr;
2001 } // func __kmp_fast_allocate
2002
2003 // Free fast memory and place it on the thread's free list if it is of
2004 // the correct size.
___kmp_fast_free(kmp_info_t * this_thr,void * ptr KMP_SRC_LOC_DECL)2005 void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL) {
2006 kmp_mem_descr_t *descr;
2007 kmp_info_t *alloc_thr;
2008 size_t size;
2009 size_t idx;
2010 int index;
2011
2012 KE_TRACE(25, ("-> __kmp_fast_free( T#%d, %p ) called from %s:%d\n",
2013 __kmp_gtid_from_thread(this_thr), ptr KMP_SRC_LOC_PARM));
2014 KMP_ASSERT(ptr != NULL);
2015
2016 descr = (kmp_mem_descr_t *)(((kmp_uintptr_t)ptr) - sizeof(kmp_mem_descr_t));
2017
2018 KE_TRACE(26, (" __kmp_fast_free: size_aligned=%d\n",
2019 (int)descr->size_aligned));
2020
2021 size = descr->size_aligned; // 2, 4, 16, 64, 65, 66, ... cache lines
2022
2023 idx = DCACHE_LINE * 2; // 2 cache lines is minimal size of block
2024 if (idx == size) {
2025 index = 0; // 2 cache lines
2026 } else if ((idx <<= 1) == size) {
2027 index = 1; // 4 cache lines
2028 } else if ((idx <<= 2) == size) {
2029 index = 2; // 16 cache lines
2030 } else if ((idx <<= 2) == size) {
2031 index = 3; // 64 cache lines
2032 } else {
2033 KMP_DEBUG_ASSERT(size > DCACHE_LINE * 64);
2034 goto free_call; // 65 or more cache lines ( > 8KB )
2035 }
2036
2037 alloc_thr = (kmp_info_t *)descr->ptr_aligned; // get thread owning the block
2038 if (alloc_thr == this_thr) {
2039 // push block to self no-sync free list, linking previous head (LIFO)
2040 *((void **)ptr) = this_thr->th.th_free_lists[index].th_free_list_self;
2041 this_thr->th.th_free_lists[index].th_free_list_self = ptr;
2042 } else {
2043 void *head = this_thr->th.th_free_lists[index].th_free_list_other;
2044 if (head == NULL) {
2045 // Create new free list
2046 this_thr->th.th_free_lists[index].th_free_list_other = ptr;
2047 *((void **)ptr) = NULL; // mark the tail of the list
2048 descr->size_allocated = (size_t)1; // head of the list keeps its length
2049 } else {
2050 // need to check existed "other" list's owner thread and size of queue
2051 kmp_mem_descr_t *dsc =
2052 (kmp_mem_descr_t *)((char *)head - sizeof(kmp_mem_descr_t));
2053 // allocating thread, same for all queue nodes
2054 kmp_info_t *q_th = (kmp_info_t *)(dsc->ptr_aligned);
2055 size_t q_sz =
2056 dsc->size_allocated + 1; // new size in case we add current task
2057 if (q_th == alloc_thr && q_sz <= KMP_FREE_LIST_LIMIT) {
2058 // we can add current task to "other" list, no sync needed
2059 *((void **)ptr) = head;
2060 descr->size_allocated = q_sz;
2061 this_thr->th.th_free_lists[index].th_free_list_other = ptr;
2062 } else {
2063 // either queue blocks owner is changing or size limit exceeded
2064 // return old queue to allocating thread (q_th) synchronously,
2065 // and start new list for alloc_thr's tasks
2066 void *old_ptr;
2067 void *tail = head;
2068 void *next = *((void **)head);
2069 while (next != NULL) {
2070 KMP_DEBUG_ASSERT(
2071 // queue size should decrease by 1 each step through the list
2072 ((kmp_mem_descr_t *)((char *)next - sizeof(kmp_mem_descr_t)))
2073 ->size_allocated +
2074 1 ==
2075 ((kmp_mem_descr_t *)((char *)tail - sizeof(kmp_mem_descr_t)))
2076 ->size_allocated);
2077 tail = next; // remember tail node
2078 next = *((void **)next);
2079 }
2080 KMP_DEBUG_ASSERT(q_th != NULL);
2081 // push block to owner's sync free list
2082 old_ptr = TCR_PTR(q_th->th.th_free_lists[index].th_free_list_sync);
2083 /* the next pointer must be set before setting free_list to ptr to avoid
2084 exposing a broken list to other threads, even for an instant. */
2085 *((void **)tail) = old_ptr;
2086
2087 while (!KMP_COMPARE_AND_STORE_PTR(
2088 &q_th->th.th_free_lists[index].th_free_list_sync, old_ptr, head)) {
2089 KMP_CPU_PAUSE();
2090 old_ptr = TCR_PTR(q_th->th.th_free_lists[index].th_free_list_sync);
2091 *((void **)tail) = old_ptr;
2092 }
2093
2094 // start new list of not-selt tasks
2095 this_thr->th.th_free_lists[index].th_free_list_other = ptr;
2096 *((void **)ptr) = NULL;
2097 descr->size_allocated = (size_t)1; // head of queue keeps its length
2098 }
2099 }
2100 }
2101 goto end;
2102
2103 free_call:
2104 KE_TRACE(25, ("__kmp_fast_free: T#%d Calling __kmp_thread_free for size %d\n",
2105 __kmp_gtid_from_thread(this_thr), size));
2106 __kmp_bget_dequeue(this_thr); /* Release any queued buffers */
2107 brel(this_thr, descr->ptr_allocated);
2108
2109 end:
2110 KE_TRACE(25, ("<- __kmp_fast_free() returns\n"));
2111
2112 } // func __kmp_fast_free
2113
2114 // Initialize the thread free lists related to fast memory
2115 // Only do this when a thread is initially created.
__kmp_initialize_fast_memory(kmp_info_t * this_thr)2116 void __kmp_initialize_fast_memory(kmp_info_t *this_thr) {
2117 KE_TRACE(10, ("__kmp_initialize_fast_memory: Called from th %p\n", this_thr));
2118
2119 memset(this_thr->th.th_free_lists, 0, NUM_LISTS * sizeof(kmp_free_list_t));
2120 }
2121
2122 // Free the memory in the thread free lists related to fast memory
2123 // Only do this when a thread is being reaped (destroyed).
__kmp_free_fast_memory(kmp_info_t * th)2124 void __kmp_free_fast_memory(kmp_info_t *th) {
2125 // Suppose we use BGET underlying allocator, walk through its structures...
2126 int bin;
2127 thr_data_t *thr = get_thr_data(th);
2128 void **lst = NULL;
2129
2130 KE_TRACE(
2131 5, ("__kmp_free_fast_memory: Called T#%d\n", __kmp_gtid_from_thread(th)));
2132
2133 __kmp_bget_dequeue(th); // Release any queued buffers
2134
2135 // Dig through free lists and extract all allocated blocks
2136 for (bin = 0; bin < MAX_BGET_BINS; ++bin) {
2137 bfhead_t *b = thr->freelist[bin].ql.flink;
2138 while (b != &thr->freelist[bin]) {
2139 if ((kmp_uintptr_t)b->bh.bb.bthr & 1) { // the buffer is allocated address
2140 *((void **)b) =
2141 lst; // link the list (override bthr, but keep flink yet)
2142 lst = (void **)b; // push b into lst
2143 }
2144 b = b->ql.flink; // get next buffer
2145 }
2146 }
2147 while (lst != NULL) {
2148 void *next = *lst;
2149 KE_TRACE(10, ("__kmp_free_fast_memory: freeing %p, next=%p th %p (%d)\n",
2150 lst, next, th, __kmp_gtid_from_thread(th)));
2151 (*thr->relfcn)(lst);
2152 #if BufStats
2153 // count blocks to prevent problems in __kmp_finalize_bget()
2154 thr->numprel++; /* Nr of expansion block releases */
2155 thr->numpblk--; /* Total number of blocks */
2156 #endif
2157 lst = (void **)next;
2158 }
2159
2160 KE_TRACE(
2161 5, ("__kmp_free_fast_memory: Freed T#%d\n", __kmp_gtid_from_thread(th)));
2162 }
2163
2164 #endif // USE_FAST_MEMORY
2165