Line data Source code
1 : /*
2 : * SPDX-License-Identifier: MPL-2.0
3 : *
4 : * This Source Code Form is subject to the terms of the Mozilla Public
5 : * License, v. 2.0. If a copy of the MPL was not distributed with this
6 : * file, You can obtain one at http://mozilla.org/MPL/2.0/.
7 : *
8 : * Copyright 2024 MonetDB Foundation;
9 : * Copyright August 2008 - 2023 MonetDB B.V.;
10 : * Copyright 1997 - July 2008 CWI.
11 : */
12 :
13 : /*
14 : * (author) M Kersten, S Mullender
15 : * Dataflow processing only works on a code
16 : * sequence that does not include additional (implicit) flow of control
17 : * statements and, ideally, consist of expensive BAT operations.
18 : * The dataflow portion is identified as a guarded block,
19 : * whose entry is controlled by the function language.dataflow();
20 : *
21 : * The dataflow worker tries to follow the sequence of actions
22 : * as laid out in the plan, but abandon this track when it hits
23 : * a blocking operator, or an instruction for which not all arguments
24 : * are available or resources become scarce.
25 : *
26 : * The flow graphs is organized such that parallel threads can
27 : * access it mostly without expensive locking and dependent
28 : * variables are easy to find..
29 : */
30 : #include "monetdb_config.h"
31 : #include "mal_dataflow.h"
32 : #include "mal_exception.h"
33 : #include "mal_private.h"
34 : #include "mal_internal.h"
35 : #include "mal_runtime.h"
36 : #include "mal_resource.h"
37 : #include "mal_function.h"
38 :
39 : #define DFLOWpending 0 /* runnable */
40 : #define DFLOWrunning 1 /* currently in progress */
41 : #define DFLOWwrapup 2 /* done! */
42 : #define DFLOWretry 3 /* reschedule */
43 : #define DFLOWskipped 4 /* due to errors */
44 :
45 : /* The per instruction status of execution */
46 : typedef struct FLOWEVENT {
47 : struct DATAFLOW *flow; /* execution context */
48 : int pc; /* pc in underlying malblock */
49 : int blocks; /* awaiting for variables */
50 : sht state; /* of execution */
51 : lng clk;
52 : sht cost;
53 : lng hotclaim; /* memory foot print of result variables */
54 : lng argclaim; /* memory foot print of arguments */
55 : lng maxclaim; /* memory foot print of largest argument, could be used to indicate result size */
56 : struct FLOWEVENT *next; /* linked list for queues */
57 : } *FlowEvent, FlowEventRec;
58 :
59 : typedef struct queue {
60 : int exitcount; /* how many threads should exit */
61 : FlowEvent first, last; /* first and last element of the queue */
62 : MT_Lock l; /* it's a shared resource, ie we need locks */
63 : MT_Sema s; /* threads wait on empty queues */
64 : } Queue;
65 :
66 : /*
67 : * The dataflow dependency is administered in a graph list structure.
68 : * For each instruction we keep the list of instructions that
69 : * should be checked for eligibility once we are finished with it.
70 : */
71 : typedef struct DATAFLOW {
72 : Client cntxt; /* for debugging and client resolution */
73 : MalBlkPtr mb; /* carry the context */
74 : MalStkPtr stk;
75 : int start, stop; /* guarded block under consideration */
76 : FlowEvent status; /* status of each instruction */
77 : ATOMIC_PTR_TYPE error; /* error encountered */
78 : int *nodes; /* dependency graph nodes */
79 : int *edges; /* dependency graph */
80 : MT_Lock flowlock; /* lock to protect the above */
81 : Queue *done; /* instructions handled */
82 : bool set_qry_ctx;
83 : } *DataFlow, DataFlowRec;
84 :
85 : struct worker {
86 : MT_Id id;
87 : enum { WAITING, RUNNING, FREE, EXITED, FINISHING } flag;
88 : ATOMIC_PTR_TYPE cntxt; /* client we do work for (NULL -> any) */
89 : MT_Sema s;
90 : struct worker *next;
91 : char errbuf[GDKMAXERRLEN]; /* GDKerrbuf so that we can allocate before fork */
92 : };
93 : /* heads of three mutually exclusive linked lists, all using the .next
94 : * field in the worker struct */
95 : static struct worker *workers; /* "working" workers */
96 : static struct worker *exited_workers; /* to be joined threads (.flag==EXITED) */
97 : static struct worker *free_workers; /* free workers (.flag==FREE) */
98 : static int free_count = 0; /* number of free threads */
99 : static int free_max = 0; /* max number of spare free threads */
100 :
101 : static Queue *todo = 0; /* pending instructions */
102 :
103 : static ATOMIC_TYPE exiting = ATOMIC_VAR_INIT(0);
104 : static MT_Lock dataflowLock = MT_LOCK_INITIALIZER(dataflowLock);
105 :
106 : /*
107 : * Calculate the size of the dataflow dependency graph.
108 : */
109 : static int
110 : DFLOWgraphSize(MalBlkPtr mb, int start, int stop)
111 : {
112 : int cnt = 0;
113 :
114 14022179 : for (int i = start; i < stop; i++)
115 13876794 : cnt += getInstrPtr(mb, i)->argc;
116 145385 : return cnt;
117 : }
118 :
119 : /*
120 : * The dataflow execution is confined to a barrier block.
121 : * Within the block there are multiple flows, which, in principle,
122 : * can be executed in parallel.
123 : */
124 :
125 : static Queue *
126 145712 : q_create(const char *name)
127 : {
128 145712 : Queue *q = GDKzalloc(sizeof(Queue));
129 :
130 145712 : if (q == NULL)
131 : return NULL;
132 145712 : MT_lock_init(&q->l, name);
133 145712 : MT_sema_init(&q->s, 0, name);
134 145712 : return q;
135 : }
136 :
137 : static void
138 145383 : q_destroy(Queue *q)
139 : {
140 145383 : assert(q);
141 145383 : MT_lock_destroy(&q->l);
142 145383 : MT_sema_destroy(&q->s);
143 145380 : GDKfree(q);
144 145384 : }
145 :
146 : /* keep a simple LIFO queue. It won't be a large one, so shuffles of requeue is possible */
147 : /* we might actually sort it for better scheduling behavior */
148 : static void
149 21614068 : q_enqueue(Queue *q, FlowEvent d)
150 : {
151 21614068 : assert(q);
152 21614068 : assert(d);
153 21614068 : MT_lock_set(&q->l);
154 21636926 : if (q->first == NULL) {
155 5343050 : assert(q->last == NULL);
156 5343050 : q->first = q->last = d;
157 : } else {
158 16293876 : assert(q->last != NULL);
159 16293876 : q->last->next = d;
160 16293876 : q->last = d;
161 : }
162 21636926 : d->next = NULL;
163 21636926 : MT_lock_unset(&q->l);
164 21666779 : MT_sema_up(&q->s);
165 21620936 : }
166 :
167 : /*
168 : * A priority queue over the hot claims of memory may
169 : * be more effective. It prioritizes those instructions
170 : * that want to use a big recent result
171 : */
172 :
173 : static void
174 0 : q_requeue(Queue *q, FlowEvent d)
175 : {
176 0 : assert(q);
177 0 : assert(d);
178 0 : MT_lock_set(&q->l);
179 0 : if (q->first == NULL) {
180 0 : assert(q->last == NULL);
181 0 : q->first = q->last = d;
182 0 : d->next = NULL;
183 : } else {
184 0 : assert(q->last != NULL);
185 0 : d->next = q->first;
186 0 : q->first = d;
187 : }
188 0 : MT_lock_unset(&q->l);
189 0 : MT_sema_up(&q->s);
190 0 : }
191 :
192 : static FlowEvent
193 21868502 : q_dequeue(Queue *q, Client cntxt)
194 : {
195 21868502 : assert(q);
196 21868502 : MT_sema_down(&q->s);
197 21918932 : if (ATOMIC_GET(&exiting))
198 : return NULL;
199 21916517 : MT_lock_set(&q->l);
200 21900301 : if (cntxt == NULL && q->exitcount > 0) {
201 145385 : q->exitcount--;
202 145385 : MT_lock_unset(&q->l);
203 145385 : return NULL;
204 : }
205 :
206 21754916 : FlowEvent *dp = &q->first;
207 21754916 : FlowEvent pd = NULL;
208 : /* if cntxt == NULL, return the first event, if cntxt != NULL, find
209 : * the first event in the queue with matching cntxt value and return
210 : * that */
211 21754916 : if (cntxt != NULL) {
212 31349416 : while (*dp && (*dp)->flow->cntxt != cntxt) {
213 29738334 : pd = *dp;
214 29738334 : dp = &pd->next;
215 : }
216 : }
217 21754916 : FlowEvent d = *dp;
218 21754916 : if (d) {
219 21590481 : *dp = d->next;
220 21590481 : d->next = NULL;
221 21590481 : if (*dp == NULL)
222 5363792 : q->last = pd;
223 : }
224 21754916 : MT_lock_unset(&q->l);
225 21726165 : return d;
226 : }
227 :
228 : /*
229 : * We simply move an instruction into the front of the queue.
230 : * Beware, we assume that variables are assigned a value once, otherwise
231 : * the order may really create errors.
232 : * The order of the instructions should be retained as long as possible.
233 : * Delay processing when we run out of memory. Push the instruction back
234 : * on the end of queue, waiting for another attempt. Problem might become
235 : * that all threads but one are cycling through the queue, each time
236 : * finding an eligible instruction, but without enough space.
237 : * Therefore, we wait for a few milliseconds as an initial punishment.
238 : *
239 : * The process could be refined by checking for cheap operations,
240 : * i.e. those that would require no memory at all (aggr.count)
241 : * This, however, would lead to a dependency to the upper layers,
242 : * because in the kernel we don't know what routines are available
243 : * with this property. Nor do we maintain such properties.
244 : */
245 :
246 : static void
247 4847 : DFLOWworker(void *T)
248 : {
249 4847 : struct worker *t = (struct worker *) T;
250 4847 : bool locked = false;
251 : #ifdef _MSC_VER
252 : srand((unsigned int) GDKusec());
253 : #endif
254 4847 : GDKsetbuf(t->errbuf); /* where to leave errors */
255 4843 : snprintf(t->s.name, sizeof(t->s.name), "DFLOWsema%04zu", MT_getpid());
256 :
257 147630 : for (;;) {
258 147630 : DataFlow flow;
259 147630 : FlowEvent fe = 0, fnxt = 0;
260 147630 : str error = 0;
261 147630 : int i;
262 147630 : lng claim;
263 147630 : Client cntxt;
264 147630 : InstrPtr p;
265 :
266 147630 : GDKclrerr();
267 :
268 147598 : if (t->flag == WAITING) {
269 : /* wait until we are allowed to start working */
270 145373 : MT_sema_down(&t->s);
271 145377 : t->flag = RUNNING;
272 145377 : if (ATOMIC_GET(&exiting)) {
273 : break;
274 : }
275 : }
276 147602 : assert(t->flag == RUNNING);
277 147602 : cntxt = ATOMIC_PTR_GET(&t->cntxt);
278 14000266 : while (1) {
279 14000266 : MT_thread_set_qry_ctx(NULL);
280 13957295 : if (fnxt == 0) {
281 8233200 : MT_thread_setworking("waiting for work");
282 8233261 : cntxt = ATOMIC_PTR_GET(&t->cntxt);
283 8233261 : fe = q_dequeue(todo, cntxt);
284 8253390 : if (fe == NULL) {
285 311874 : if (cntxt) {
286 : /* we're not done yet with work for the current
287 : * client (as far as we know), so give up the CPU
288 : * and let the scheduler enter some more work, but
289 : * first compensate for the down we did in
290 : * dequeue */
291 164792 : MT_sema_up(&todo->s);
292 164787 : MT_sleep_ms(1);
293 14164856 : continue;
294 : }
295 : /* no more work to be done: exit */
296 147082 : break;
297 : }
298 7941516 : if (fe->flow->cntxt && fe->flow->cntxt->mythread)
299 7940679 : MT_thread_setworking(fe->flow->cntxt->mythread);
300 : } else
301 : fe = fnxt;
302 13664704 : if (ATOMIC_GET(&exiting)) {
303 : break;
304 : }
305 13664704 : fnxt = 0;
306 13664704 : assert(fe);
307 13664704 : flow = fe->flow;
308 13664704 : assert(flow);
309 13664704 : MT_thread_set_qry_ctx(flow->set_qry_ctx ? &flow->cntxt->
310 : qryctx : NULL);
311 :
312 : /* whenever we have a (concurrent) error, skip it */
313 13690784 : if (ATOMIC_PTR_GET(&flow->error)) {
314 4183 : q_enqueue(flow->done, fe);
315 4183 : continue;
316 : }
317 :
318 13686601 : p = getInstrPtr(flow->mb, fe->pc);
319 13686601 : claim = fe->argclaim;
320 27400752 : if (p->fcn != (MALfcn) deblockdataflow && /* never block on deblockdataflow() */
321 13686601 : !MALadmission_claim(flow->cntxt, flow->mb, flow->stk, p, claim)) {
322 0 : fe->hotclaim = 0; /* don't assume priority anymore */
323 0 : fe->maxclaim = 0;
324 0 : MT_lock_set(&todo->l);
325 0 : FlowEvent last = todo->last;
326 0 : MT_lock_unset(&todo->l);
327 0 : if (last == NULL)
328 0 : MT_sleep_ms(DELAYUNIT);
329 0 : q_requeue(todo, fe);
330 0 : continue;
331 : }
332 13714151 : ATOMIC_BASE_TYPE wrks = ATOMIC_INC(&flow->cntxt->workers);
333 13714151 : ATOMIC_BASE_TYPE mwrks = ATOMIC_GET(&flow->mb->workers);
334 13714570 : while (wrks > mwrks) {
335 34064 : if (ATOMIC_CAS(&flow->mb->workers, &mwrks, wrks))
336 : break;
337 : }
338 13714203 : error = runMALsequence(flow->cntxt, flow->mb, fe->pc, fe->pc + 1,
339 : flow->stk, 0, 0);
340 13545202 : ATOMIC_DEC(&flow->cntxt->workers);
341 : /* release the memory claim */
342 13545202 : MALadmission_release(flow->cntxt, flow->mb, flow->stk, p, claim);
343 :
344 13652297 : MT_lock_set(&flow->flowlock);
345 13723959 : fe->state = DFLOWwrapup;
346 13723959 : MT_lock_unset(&flow->flowlock);
347 13706707 : if (error) {
348 523 : void *null = NULL;
349 : /* only collect one error (from one thread, needed for stable testing) */
350 523 : if (!ATOMIC_PTR_CAS(&flow->error, &null, error))
351 47 : freeException(error);
352 : /* after an error we skip the rest of the block */
353 523 : q_enqueue(flow->done, fe);
354 523 : continue;
355 : }
356 :
357 : /* see if you can find an eligible instruction that uses the
358 : * result just produced. Then we can continue with it right away.
359 : * We are just looking forward for the last block, which means we
360 : * are safe from concurrent actions. No other thread can steal it,
361 : * because we hold the logical lock.
362 : * All eligible instructions are queued
363 : */
364 13706184 : p = getInstrPtr(flow->mb, fe->pc);
365 13706184 : assert(p);
366 13706184 : fe->hotclaim = 0;
367 13706184 : fe->maxclaim = 0;
368 :
369 28548046 : for (i = 0; i < p->retc; i++) {
370 14848998 : lng footprint;
371 14848998 : footprint = getMemoryClaim(flow->mb, flow->stk, p, i, FALSE);
372 14841862 : fe->hotclaim += footprint;
373 14841862 : if (footprint > fe->maxclaim)
374 5398715 : fe->maxclaim = footprint;
375 : }
376 :
377 : /* Try to get rid of the hot potato or locate an alternative to proceed.
378 : */
379 : #define HOTPOTATO
380 : #ifdef HOTPOTATO
381 : /* HOT potato choice */
382 13699048 : int last = 0, nxt = -1;
383 13699048 : lng nxtclaim = -1;
384 :
385 13699048 : MT_lock_set(&flow->flowlock);
386 13725343 : for (last = fe->pc - flow->start;
387 55150204 : last >= 0 && (i = flow->nodes[last]) > 0;
388 41424861 : last = flow->edges[last]) {
389 41424861 : if (flow->status[i].state == DFLOWpending
390 41420696 : && flow->status[i].blocks == 1) {
391 : /* find the one with the largest footprint */
392 11313197 : if (nxt == -1 || flow->status[i].argclaim > nxtclaim) {
393 6254465 : nxt = i;
394 6254465 : nxtclaim = flow->status[i].argclaim;
395 : }
396 : }
397 : }
398 : /* hot potato can not be removed, use alternative to proceed */
399 13725343 : if (nxt >= 0) {
400 5788281 : flow->status[nxt].state = DFLOWrunning;
401 5788281 : flow->status[nxt].blocks = 0;
402 5788281 : flow->status[nxt].hotclaim = fe->hotclaim;
403 5788281 : flow->status[nxt].argclaim += fe->hotclaim;
404 5788281 : if (flow->status[nxt].maxclaim < fe->maxclaim)
405 2250098 : flow->status[nxt].maxclaim = fe->maxclaim;
406 : fnxt = flow->status + nxt;
407 : }
408 13725343 : MT_lock_unset(&flow->flowlock);
409 : #endif
410 :
411 13711650 : q_enqueue(flow->done, fe);
412 13683368 : if (fnxt == 0 && profilerStatus) {
413 0 : profilerHeartbeatEvent("wait");
414 : }
415 : }
416 147082 : MT_lock_set(&dataflowLock);
417 147661 : if (GDKexiting() || ATOMIC_GET(&exiting) || free_count >= free_max) {
418 : locked = true;
419 : break;
420 : }
421 143412 : free_count++;
422 143412 : struct worker **tp = &workers;
423 890097 : while (*tp && *tp != t)
424 746685 : tp = &(*tp)->next;
425 143412 : assert(*tp && *tp == t);
426 143412 : *tp = t->next;
427 143412 : t->flag = FREE;
428 143412 : t->next = free_workers;
429 143412 : free_workers = t;
430 143412 : MT_lock_unset(&dataflowLock);
431 143412 : MT_thread_setworking("idle, waiting for new client");
432 143412 : MT_sema_down(&t->s);
433 143410 : if (GDKexiting() || ATOMIC_GET(&exiting))
434 : break;
435 142805 : assert(t->flag == WAITING);
436 : }
437 595 : if (!locked)
438 595 : MT_lock_set(&dataflowLock);
439 4844 : if (t->flag != FINISHING) {
440 4231 : struct worker **tp = t->flag == FREE ? &free_workers : &workers;
441 19530 : while (*tp && *tp != t)
442 15299 : tp = &(*tp)->next;
443 4231 : assert(*tp && *tp == t);
444 4231 : *tp = t->next;
445 4231 : t->flag = EXITED;
446 4231 : t->next = exited_workers;
447 4231 : exited_workers = t;
448 : }
449 4844 : MT_lock_unset(&dataflowLock);
450 4841 : GDKsetbuf(NULL);
451 4836 : }
452 :
453 : /*
454 : * Create an interpreter pool.
455 : * One worker will adaptively be available for each client.
456 : * The remainder are taken from the GDKnr_threads argument and
457 : * typically is equal to the number of cores
458 : * The workers are assembled in a local table to enable debugging.
459 : */
460 : static int
461 327 : DFLOWinitialize(void)
462 : {
463 327 : int limit;
464 327 : int created = 0;
465 :
466 327 : MT_lock_set(&mal_contextLock);
467 327 : MT_lock_set(&dataflowLock);
468 327 : if (todo) {
469 : /* somebody else beat us to it */
470 0 : MT_lock_unset(&dataflowLock);
471 0 : MT_lock_unset(&mal_contextLock);
472 0 : return 0;
473 : }
474 327 : free_max = GDKgetenv_int("dataflow_max_free",
475 : GDKnr_threads < 4 ? 4 : GDKnr_threads);
476 327 : todo = q_create("todo");
477 327 : if (todo == NULL) {
478 0 : MT_lock_unset(&dataflowLock);
479 0 : MT_lock_unset(&mal_contextLock);
480 0 : return -1;
481 : }
482 327 : limit = GDKnr_threads ? GDKnr_threads - 1 : 0;
483 2610 : while (limit > 0) {
484 2283 : limit--;
485 2283 : struct worker *t = GDKmalloc(sizeof(*t));
486 2283 : if (t == NULL) {
487 0 : TRC_CRITICAL(MAL_SERVER, "cannot allocate structure for worker");
488 0 : continue;
489 : }
490 2283 : *t = (struct worker) {
491 : .flag = RUNNING,
492 : .cntxt = ATOMIC_PTR_VAR_INIT(NULL),
493 : };
494 2283 : MT_sema_init(&t->s, 0, "DFLOWsema"); /* placeholder name */
495 2283 : if (MT_create_thread(&t->id, DFLOWworker, t,
496 : MT_THR_JOINABLE, "DFLOWworkerXXXX") < 0) {
497 0 : ATOMIC_PTR_DESTROY(&t->cntxt);
498 0 : MT_sema_destroy(&t->s);
499 0 : GDKfree(t);
500 : } else {
501 2283 : t->next = workers;
502 2283 : workers = t;
503 2283 : created++;
504 : }
505 : }
506 327 : if (created == 0) {
507 : /* no threads created */
508 0 : q_destroy(todo);
509 0 : todo = NULL;
510 0 : MT_lock_unset(&dataflowLock);
511 0 : MT_lock_unset(&mal_contextLock);
512 0 : return -1;
513 : }
514 327 : MT_lock_unset(&dataflowLock);
515 327 : MT_lock_unset(&mal_contextLock);
516 327 : return 0;
517 : }
518 :
519 : /*
520 : * The dataflow administration is based on administration of
521 : * how many variables are still missing before it can be executed.
522 : * For each instruction we keep a list of instructions whose
523 : * blocking counter should be decremented upon finishing it.
524 : */
525 : static str
526 145385 : DFLOWinitBlk(DataFlow flow, MalBlkPtr mb, int size)
527 : {
528 145385 : int pc, i, j, k, l, n, etop = 0;
529 145385 : int *assign;
530 145385 : InstrPtr p;
531 :
532 145385 : if (flow == NULL)
533 0 : throw(MAL, "dataflow", "DFLOWinitBlk(): Called with flow == NULL");
534 145385 : if (mb == NULL)
535 0 : throw(MAL, "dataflow", "DFLOWinitBlk(): Called with mb == NULL");
536 145385 : assign = (int *) GDKzalloc(mb->vtop * sizeof(int));
537 145385 : if (assign == NULL)
538 0 : throw(MAL, "dataflow", SQLSTATE(HY013) MAL_MALLOC_FAIL);
539 145385 : etop = flow->stop - flow->start;
540 13876393 : for (n = 0, pc = flow->start; pc < flow->stop; pc++, n++) {
541 13731009 : p = getInstrPtr(mb, pc);
542 13731009 : if (p == NULL) {
543 0 : GDKfree(assign);
544 0 : throw(MAL, "dataflow",
545 : "DFLOWinitBlk(): getInstrPtr() returned NULL");
546 : }
547 :
548 : /* initial state, ie everything can run */
549 13731009 : flow->status[n].flow = flow;
550 13731009 : flow->status[n].pc = pc;
551 13731009 : flow->status[n].state = DFLOWpending;
552 13731009 : flow->status[n].cost = -1;
553 13731009 : ATOMIC_PTR_SET(&flow->status[n].flow->error, NULL);
554 :
555 : /* administer flow dependencies */
556 64813444 : for (j = p->retc; j < p->argc; j++) {
557 : /* list of instructions that wake n-th instruction up */
558 51082436 : if (!isVarConstant(mb, getArg(p, j)) && (k = assign[getArg(p, j)])) {
559 25720176 : assert(k < pc); /* only dependencies on earlier instructions */
560 : /* add edge to the target instruction for wakeup call */
561 25720176 : k -= flow->start;
562 25720176 : if (flow->nodes[k]) {
563 : /* add wakeup to tail of list */
564 225080056 : for (i = k; flow->edges[i] > 0; i = flow->edges[i])
565 : ;
566 23137748 : flow->nodes[etop] = n;
567 23137748 : flow->edges[etop] = -1;
568 23137748 : flow->edges[i] = etop;
569 23137748 : etop++;
570 23137748 : (void) size;
571 23137748 : if (etop == size) {
572 213 : int *tmp;
573 : /* in case of realloc failure, the original
574 : * pointers will be freed by the caller */
575 213 : tmp = (int *) GDKrealloc(flow->nodes,
576 : sizeof(int) * 2 * size);
577 212 : if (tmp == NULL) {
578 0 : GDKfree(assign);
579 0 : throw(MAL, "dataflow",
580 : SQLSTATE(HY013) MAL_MALLOC_FAIL);
581 : }
582 212 : flow->nodes = tmp;
583 212 : tmp = (int *) GDKrealloc(flow->edges,
584 : sizeof(int) * 2 * size);
585 212 : if (tmp == NULL) {
586 0 : GDKfree(assign);
587 0 : throw(MAL, "dataflow",
588 : SQLSTATE(HY013) MAL_MALLOC_FAIL);
589 : }
590 212 : flow->edges = tmp;
591 212 : size *= 2;
592 : }
593 : } else {
594 2582428 : flow->nodes[k] = n;
595 2582428 : flow->edges[k] = -1;
596 : }
597 :
598 25720175 : flow->status[n].blocks++;
599 : }
600 :
601 : /* list of instructions to be woken up explicitly */
602 51082435 : if (!isVarConstant(mb, getArg(p, j))) {
603 : /* be careful, watch out for garbage collection interference */
604 : /* those should be scheduled after all its other uses */
605 27189421 : l = getEndScope(mb, getArg(p, j));
606 27189421 : if (l != pc && l < flow->stop && l > flow->start) {
607 : /* add edge to the target instruction for wakeup call */
608 15731627 : assert(pc < l); /* only dependencies on earlier instructions */
609 15731627 : l -= flow->start;
610 15731627 : if (flow->nodes[n]) {
611 : /* add wakeup to tail of list */
612 77899723 : for (i = n; flow->edges[i] > 0; i = flow->edges[i])
613 : ;
614 7688355 : flow->nodes[etop] = l;
615 7688355 : flow->edges[etop] = -1;
616 7688355 : flow->edges[i] = etop;
617 7688355 : etop++;
618 7688355 : if (etop == size) {
619 209 : int *tmp;
620 : /* in case of realloc failure, the original
621 : * pointers will be freed by the caller */
622 209 : tmp = (int *) GDKrealloc(flow->nodes,
623 : sizeof(int) * 2 * size);
624 209 : if (tmp == NULL) {
625 0 : GDKfree(assign);
626 0 : throw(MAL, "dataflow",
627 : SQLSTATE(HY013) MAL_MALLOC_FAIL);
628 : }
629 209 : flow->nodes = tmp;
630 209 : tmp = (int *) GDKrealloc(flow->edges,
631 : sizeof(int) * 2 * size);
632 209 : if (tmp == NULL) {
633 0 : GDKfree(assign);
634 0 : throw(MAL, "dataflow",
635 : SQLSTATE(HY013) MAL_MALLOC_FAIL);
636 : }
637 209 : flow->edges = tmp;
638 209 : size *= 2;
639 : }
640 : } else {
641 8043272 : flow->nodes[n] = l;
642 8043272 : flow->edges[n] = -1;
643 : }
644 15731627 : flow->status[l].blocks++;
645 : }
646 : }
647 : }
648 :
649 28609228 : for (j = 0; j < p->retc; j++)
650 14878220 : assign[getArg(p, j)] = pc; /* ensure recognition of dependency on first instruction and constant */
651 : }
652 145384 : GDKfree(assign);
653 :
654 145384 : return MAL_SUCCEED;
655 : }
656 :
657 : /*
658 : * Parallel processing is mostly driven by dataflow, but within this context
659 : * there may be different schemes to take instructions into execution.
660 : * The admission scheme (and wrapup) are the necessary scheduler hooks.
661 : * A scheduler registers the functions needed and should release them
662 : * at the end of the parallel block.
663 : * They take effect after we have ensured that the basic properties for
664 : * execution hold.
665 : */
666 : static str
667 145385 : DFLOWscheduler(DataFlow flow, struct worker *w)
668 : {
669 145385 : int last;
670 145385 : int i;
671 145385 : int j;
672 145385 : InstrPtr p;
673 145385 : int tasks = 0, actions = 0;
674 145385 : str ret = MAL_SUCCEED;
675 145385 : FlowEvent fe, f = 0;
676 :
677 145385 : if (flow == NULL)
678 0 : throw(MAL, "dataflow", "DFLOWscheduler(): Called with flow == NULL");
679 145385 : actions = flow->stop - flow->start;
680 145385 : if (actions == 0)
681 0 : throw(MAL, "dataflow", "Empty dataflow block");
682 : /* initialize the eligible statements */
683 145385 : fe = flow->status;
684 :
685 145385 : ATOMIC_DEC(&flow->cntxt->workers);
686 145385 : MT_lock_set(&flow->flowlock);
687 14022263 : for (i = 0; i < actions; i++)
688 13731495 : if (fe[i].blocks == 0) {
689 899216 : p = getInstrPtr(flow->mb, fe[i].pc);
690 899216 : if (p == NULL) {
691 0 : MT_lock_unset(&flow->flowlock);
692 0 : ATOMIC_INC(&flow->cntxt->workers);
693 0 : throw(MAL, "dataflow",
694 : "DFLOWscheduler(): getInstrPtr(flow->mb,fe[i].pc) returned NULL");
695 : }
696 899216 : fe[i].argclaim = 0;
697 5228685 : for (j = p->retc; j < p->argc; j++)
698 4329469 : fe[i].argclaim += getMemoryClaim(fe[0].flow->mb,
699 4329480 : fe[0].flow->stk, p, j, FALSE);
700 899205 : flow->status[i].state = DFLOWrunning;
701 899205 : q_enqueue(todo, flow->status + i);
702 : }
703 145384 : MT_lock_unset(&flow->flowlock);
704 145385 : MT_sema_up(&w->s);
705 :
706 145385 : while (actions != tasks) {
707 13730456 : f = q_dequeue(flow->done, NULL);
708 13730633 : if (ATOMIC_GET(&exiting))
709 : break;
710 13730633 : if (f == NULL) {
711 0 : ATOMIC_INC(&flow->cntxt->workers);
712 0 : throw(MAL, "dataflow",
713 : "DFLOWscheduler(): q_dequeue(flow->done) returned NULL");
714 : }
715 :
716 : /*
717 : * When an instruction is finished we have to reduce the blocked
718 : * counter for all dependent instructions. for those where it
719 : * drops to zero we can scheduler it we do it here instead of the scheduler
720 : */
721 :
722 13730633 : MT_lock_set(&flow->flowlock);
723 13730120 : tasks++;
724 13730120 : for (last = f->pc - flow->start;
725 55172750 : last >= 0 && (i = flow->nodes[last]) > 0; last = flow->edges[last])
726 41442835 : if (flow->status[i].state == DFLOWpending) {
727 35660897 : flow->status[i].argclaim += f->hotclaim;
728 35660897 : if (flow->status[i].blocks == 1) {
729 7043413 : flow->status[i].blocks--;
730 7043413 : flow->status[i].state = DFLOWrunning;
731 7043413 : q_enqueue(todo, flow->status + i);
732 : } else {
733 28617484 : flow->status[i].blocks--;
734 : }
735 : }
736 13875299 : MT_lock_unset(&flow->flowlock);
737 : }
738 : /* release the worker from its specific task (turn it into a
739 : * generic worker) */
740 145384 : ATOMIC_PTR_SET(&w->cntxt, NULL);
741 145384 : ATOMIC_INC(&flow->cntxt->workers);
742 : /* wrap up errors */
743 145384 : assert(flow->done->last == 0);
744 145384 : if ((ret = ATOMIC_PTR_XCG(&flow->error, NULL)) != NULL) {
745 476 : TRC_DEBUG(MAL_SERVER, "Errors encountered: %s\n", ret);
746 : }
747 : return ret;
748 : }
749 :
750 : /* called and returns with dataflowLock locked, temporarily unlocks
751 : * join the thread associated with the worker and destroy the structure */
752 : static inline void
753 4844 : finish_worker(struct worker *t)
754 : {
755 4844 : t->flag = FINISHING;
756 4844 : MT_lock_unset(&dataflowLock);
757 4844 : MT_join_thread(t->id);
758 4844 : MT_sema_destroy(&t->s);
759 4844 : ATOMIC_PTR_DESTROY(&t->cntxt);
760 4844 : GDKfree(t);
761 4844 : MT_lock_set(&dataflowLock);
762 4844 : }
763 :
764 : /* We create a pool of GDKnr_threads-1 generic workers, that is,
765 : * workers that will take on jobs from any clients. In addition, we
766 : * create a single specific worker per client (i.e. each time we enter
767 : * here). This specific worker will only do work for the client for
768 : * which it was started. In this way we can guarantee that there will
769 : * always be progress for the client, even if all other workers are
770 : * doing something big.
771 : *
772 : * When all jobs for a client have been done (there are no more
773 : * entries for the client in the queue), the specific worker turns
774 : * itself into a generic worker. At the same time, we signal that one
775 : * generic worker should exit and this function returns. In this way
776 : * we make sure that there are once again GDKnr_threads-1 generic
777 : * workers. */
778 : str
779 145380 : runMALdataflow(Client cntxt, MalBlkPtr mb, int startpc, int stoppc,
780 : MalStkPtr stk)
781 : {
782 145380 : DataFlow flow = NULL;
783 145380 : str msg = MAL_SUCCEED;
784 145380 : int size;
785 145380 : bit *ret;
786 145380 : struct worker *t;
787 :
788 145380 : if (stk == NULL)
789 0 : throw(MAL, "dataflow", "runMALdataflow(): Called with stk == NULL");
790 145380 : ret = getArgReference_bit(stk, getInstrPtr(mb, startpc), 0);
791 145380 : *ret = FALSE;
792 :
793 145380 : assert(stoppc > startpc);
794 :
795 : /* check existence of workers */
796 145380 : if (todo == NULL) {
797 : /* create thread pool */
798 327 : if (GDKnr_threads <= 1 || DFLOWinitialize() < 0) {
799 : /* no threads created, run serially */
800 0 : *ret = TRUE;
801 0 : return MAL_SUCCEED;
802 : }
803 : }
804 145380 : assert(todo);
805 : /* in addition, create one more worker that will only execute
806 : * tasks for the current client to compensate for our waiting
807 : * until all work is done */
808 145380 : MT_lock_set(&dataflowLock);
809 : /* join with already exited threads */
810 147358 : while (exited_workers != NULL) {
811 1973 : assert(exited_workers->flag == EXITED);
812 1973 : struct worker *t = exited_workers;
813 1973 : exited_workers = exited_workers->next;
814 1973 : finish_worker(t);
815 : }
816 145385 : assert(cntxt != NULL);
817 145385 : if (free_workers != NULL) {
818 142816 : t = free_workers;
819 142816 : assert(t->flag == FREE);
820 142816 : assert(free_count > 0);
821 142816 : free_count--;
822 142816 : free_workers = t->next;
823 142816 : t->next = workers;
824 142816 : workers = t;
825 142816 : t->flag = WAITING;
826 142816 : ATOMIC_PTR_SET(&t->cntxt, cntxt);
827 142816 : MT_sema_up(&t->s);
828 : } else {
829 2569 : t = GDKmalloc(sizeof(*t));
830 2569 : if (t != NULL) {
831 2569 : *t = (struct worker) {
832 : .flag = WAITING,
833 : .cntxt = ATOMIC_PTR_VAR_INIT(cntxt),
834 : };
835 2569 : MT_sema_init(&t->s, 0, "DFLOWsema"); /* placeholder name */
836 2569 : if (MT_create_thread(&t->id, DFLOWworker, t,
837 : MT_THR_JOINABLE, "DFLOWworkerXXXX") < 0) {
838 0 : ATOMIC_PTR_DESTROY(&t->cntxt);
839 0 : MT_sema_destroy(&t->s);
840 0 : GDKfree(t);
841 0 : t = NULL;
842 : } else {
843 2569 : t->next = workers;
844 2569 : workers = t;
845 : }
846 : }
847 2569 : if (t == NULL) {
848 : /* cannot start new thread, run serially */
849 0 : *ret = TRUE;
850 0 : MT_lock_unset(&dataflowLock);
851 0 : return MAL_SUCCEED;
852 : }
853 : }
854 145385 : MT_lock_unset(&dataflowLock);
855 :
856 145385 : flow = (DataFlow) GDKzalloc(sizeof(DataFlowRec));
857 145385 : if (flow == NULL)
858 0 : throw(MAL, "dataflow", SQLSTATE(HY013) MAL_MALLOC_FAIL);
859 :
860 145385 : size = DFLOWgraphSize(mb, startpc, stoppc);
861 145385 : size += stoppc - startpc;
862 :
863 290770 : *flow = (DataFlowRec) {
864 : .cntxt = cntxt,
865 : .mb = mb,
866 : .stk = stk,
867 145385 : .set_qry_ctx = MT_thread_get_qry_ctx() != NULL,
868 : /* keep real block count, exclude brackets */
869 145385 : .start = startpc + 1,
870 : .stop = stoppc,
871 145385 : .done = q_create("flow->done"),
872 145385 : .status = (FlowEvent) GDKzalloc((stoppc - startpc + 1) *
873 : sizeof(FlowEventRec)),
874 : .error = ATOMIC_PTR_VAR_INIT(NULL),
875 145385 : .nodes = (int *) GDKzalloc(sizeof(int) * size),
876 145385 : .edges = (int *) GDKzalloc(sizeof(int) * size),
877 : };
878 :
879 145385 : if (flow->done == NULL) {
880 0 : GDKfree(flow->status);
881 0 : GDKfree(flow->nodes);
882 0 : GDKfree(flow->edges);
883 0 : GDKfree(flow);
884 0 : throw(MAL, "dataflow",
885 : "runMALdataflow(): Failed to create flow->done queue");
886 : }
887 :
888 145385 : if (flow->status == NULL || flow->nodes == NULL || flow->edges == NULL) {
889 0 : q_destroy(flow->done);
890 0 : GDKfree(flow->status);
891 0 : GDKfree(flow->nodes);
892 0 : GDKfree(flow->edges);
893 0 : GDKfree(flow);
894 0 : throw(MAL, "dataflow", SQLSTATE(HY013) MAL_MALLOC_FAIL);
895 : }
896 :
897 145385 : MT_lock_init(&flow->flowlock, "flow->flowlock");
898 145385 : msg = DFLOWinitBlk(flow, mb, size);
899 :
900 145384 : if (msg == MAL_SUCCEED)
901 145384 : msg = DFLOWscheduler(flow, t);
902 :
903 145385 : GDKfree(flow->status);
904 145385 : GDKfree(flow->edges);
905 145385 : GDKfree(flow->nodes);
906 145385 : q_destroy(flow->done);
907 145385 : MT_lock_destroy(&flow->flowlock);
908 145385 : ATOMIC_PTR_DESTROY(&flow->error);
909 145385 : GDKfree(flow);
910 :
911 : /* we created one worker, now tell one worker to exit again */
912 145385 : MT_lock_set(&todo->l);
913 145385 : todo->exitcount++;
914 145385 : MT_lock_unset(&todo->l);
915 145385 : MT_sema_up(&todo->s);
916 :
917 145385 : return msg;
918 : }
919 :
920 : str
921 0 : deblockdataflow(Client cntxt, MalBlkPtr mb, MalStkPtr stk, InstrPtr pci)
922 : {
923 0 : int *ret = getArgReference_int(stk, pci, 0);
924 0 : int *val = getArgReference_int(stk, pci, 1);
925 0 : (void) cntxt;
926 0 : (void) mb;
927 0 : *ret = *val;
928 0 : return MAL_SUCCEED;
929 : }
930 :
931 : static void
932 326 : stopMALdataflow(void)
933 : {
934 326 : ATOMIC_SET(&exiting, 1);
935 326 : if (todo) {
936 326 : MT_lock_set(&dataflowLock);
937 : /* first wake up all running threads */
938 326 : int n = 0;
939 921 : for (struct worker *t = free_workers; t; t = t->next)
940 595 : n++;
941 2602 : for (struct worker *t = workers; t; t = t->next)
942 2276 : n++;
943 3197 : while (n-- > 0) {
944 : /* one UP for each thread we know about */
945 3197 : MT_sema_up(&todo->s);
946 : }
947 921 : while (free_workers) {
948 595 : struct worker *t = free_workers;
949 595 : assert(free_count > 0);
950 595 : free_count--;
951 595 : free_workers = free_workers->next;
952 595 : MT_sema_up(&t->s);
953 595 : finish_worker(t);
954 : }
955 344 : while (workers) {
956 18 : struct worker *t = workers;
957 18 : workers = workers->next;
958 18 : finish_worker(t);
959 : }
960 2584 : while (exited_workers) {
961 2258 : struct worker *t = exited_workers;
962 2258 : exited_workers = exited_workers->next;
963 2258 : finish_worker(t);
964 : }
965 326 : MT_lock_unset(&dataflowLock);
966 : }
967 326 : }
968 :
969 : void
970 326 : mal_dataflow_reset(void)
971 : {
972 326 : stopMALdataflow();
973 326 : workers = exited_workers = NULL;
974 326 : if (todo) {
975 326 : MT_lock_destroy(&todo->l);
976 326 : MT_sema_destroy(&todo->s);
977 326 : GDKfree(todo);
978 : }
979 326 : todo = 0; /* pending instructions */
980 326 : ATOMIC_SET(&exiting, 0);
981 326 : }
|
