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 12832598 : for (int i = start; i < stop; i++)
115 12677406 : cnt += getInstrPtr(mb, i)->argc;
116 155192 : 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 155521 : q_create(const char *name)
127 : {
128 155521 : Queue *q = GDKzalloc(sizeof(Queue));
129 :
130 155521 : if (q == NULL)
131 : return NULL;
132 155521 : MT_lock_init(&q->l, name);
133 155521 : MT_sema_init(&q->s, 0, name);
134 155521 : return q;
135 : }
136 :
137 : static void
138 155191 : q_destroy(Queue *q)
139 : {
140 155191 : assert(q);
141 155191 : MT_lock_destroy(&q->l);
142 155187 : MT_sema_destroy(&q->s);
143 155185 : GDKfree(q);
144 155189 : }
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 19682460 : q_enqueue(Queue *q, FlowEvent d)
150 : {
151 19682460 : assert(q);
152 19682460 : assert(d);
153 19682460 : MT_lock_set(&q->l);
154 19704182 : if (q->first == NULL) {
155 5059024 : assert(q->last == NULL);
156 5059024 : q->first = q->last = d;
157 : } else {
158 14645158 : assert(q->last != NULL);
159 14645158 : q->last->next = d;
160 14645158 : q->last = d;
161 : }
162 19704182 : d->next = NULL;
163 19704182 : MT_lock_unset(&q->l);
164 19725820 : MT_sema_up(&q->s);
165 19687200 : }
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 19915060 : q_dequeue(Queue *q, Client cntxt)
194 : {
195 19915060 : assert(q);
196 19915060 : MT_sema_down(&q->s);
197 19964914 : if (ATOMIC_GET(&exiting))
198 : return NULL;
199 19962524 : MT_lock_set(&q->l);
200 19955717 : if (cntxt == NULL && q->exitcount > 0) {
201 155192 : q->exitcount--;
202 155192 : MT_lock_unset(&q->l);
203 155192 : return NULL;
204 : }
205 :
206 19800525 : FlowEvent *dp = &q->first;
207 19800525 : 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 19800525 : if (cntxt != NULL) {
212 19784784 : while (*dp && (*dp)->flow->cntxt != cntxt) {
213 18384937 : pd = *dp;
214 18384937 : dp = &pd->next;
215 : }
216 : }
217 19800525 : FlowEvent d = *dp;
218 19800525 : if (d) {
219 19663718 : *dp = d->next;
220 19663718 : d->next = NULL;
221 19663718 : if (*dp == NULL)
222 5078034 : q->last = pd;
223 : }
224 19800525 : MT_lock_unset(&q->l);
225 19777276 : 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 4604 : DFLOWworker(void *T)
248 : {
249 4604 : struct worker *t = (struct worker *) T;
250 4604 : bool locked = false;
251 : #ifdef _MSC_VER
252 : srand((unsigned int) GDKusec());
253 : #endif
254 4604 : GDKsetbuf(t->errbuf); /* where to leave errors */
255 4596 : snprintf(t->s.name, sizeof(t->s.name), "DFLOWsema%04zu", MT_getpid());
256 :
257 157454 : for (;;) {
258 157454 : DataFlow flow;
259 157454 : FlowEvent fe = 0, fnxt = 0;
260 157454 : str error = 0;
261 157454 : int i;
262 157454 : lng claim;
263 157454 : Client cntxt;
264 157454 : InstrPtr p;
265 :
266 157454 : GDKclrerr();
267 :
268 157399 : if (t->flag == WAITING) {
269 : /* wait until we are allowed to start working */
270 155176 : MT_sema_down(&t->s);
271 155186 : t->flag = RUNNING;
272 155186 : if (ATOMIC_GET(&exiting)) {
273 : break;
274 : }
275 : }
276 157409 : assert(t->flag == RUNNING);
277 157409 : cntxt = ATOMIC_PTR_GET(&t->cntxt);
278 12777658 : while (1) {
279 12777658 : MT_thread_set_qry_ctx(NULL);
280 12749300 : if (fnxt == 0) {
281 7487768 : MT_thread_setworking("waiting for work");
282 7487155 : cntxt = ATOMIC_PTR_GET(&t->cntxt);
283 7487155 : fe = q_dequeue(todo, cntxt);
284 7503252 : if (fe == NULL) {
285 294173 : 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 137155 : MT_sema_up(&todo->s);
292 137150 : MT_sleep_ms(1);
293 12914659 : continue;
294 : }
295 : /* no more work to be done: exit */
296 157018 : break;
297 : }
298 7209079 : if (fe->flow->cntxt && fe->flow->cntxt->mythread)
299 7208230 : MT_thread_setworking(fe->flow->cntxt->mythread);
300 : } else
301 : fe = fnxt;
302 12470115 : if (ATOMIC_GET(&exiting)) {
303 : break;
304 : }
305 12470115 : fnxt = 0;
306 12470115 : assert(fe);
307 12470115 : flow = fe->flow;
308 12470115 : assert(flow);
309 12470115 : 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 12490731 : if (ATOMIC_PTR_GET(&flow->error)) {
314 3688 : q_enqueue(flow->done, fe);
315 3679 : continue;
316 : }
317 :
318 12487043 : p = getInstrPtr(flow->mb, fe->pc);
319 12487043 : claim = fe->argclaim;
320 24995897 : if (p->fcn != (MALfcn) deblockdataflow && /* never block on deblockdataflow() */
321 12487043 : !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 12508854 : ATOMIC_BASE_TYPE wrks = ATOMIC_INC(&flow->cntxt->workers);
333 12508854 : ATOMIC_BASE_TYPE mwrks = ATOMIC_GET(&flow->mb->workers);
334 12509151 : while (wrks > mwrks) {
335 34532 : if (ATOMIC_CAS(&flow->mb->workers, &mwrks, wrks))
336 : break;
337 : }
338 12508878 : error = runMALsequence(flow->cntxt, flow->mb, fe->pc, fe->pc + 1,
339 : flow->stk, 0, 0);
340 12352553 : ATOMIC_DEC(&flow->cntxt->workers);
341 : /* release the memory claim */
342 12352553 : MALadmission_release(flow->cntxt, flow->mb, flow->stk, p, claim);
343 :
344 12444298 : MT_lock_set(&flow->flowlock);
345 12515918 : fe->state = DFLOWwrapup;
346 12515918 : MT_lock_unset(&flow->flowlock);
347 12500963 : if (error) {
348 484 : void *null = NULL;
349 : /* only collect one error (from one thread, needed for stable testing) */
350 484 : if (!ATOMIC_PTR_CAS(&flow->error, &null, error))
351 55 : freeException(error);
352 : /* after an error we skip the rest of the block */
353 484 : q_enqueue(flow->done, fe);
354 484 : 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 12500479 : p = getInstrPtr(flow->mb, fe->pc);
365 12500479 : assert(p);
366 12500479 : fe->hotclaim = 0;
367 12500479 : fe->maxclaim = 0;
368 :
369 26064603 : for (i = 0; i < p->retc; i++) {
370 13569258 : lng footprint;
371 13569258 : footprint = getMemoryClaim(flow->mb, flow->stk, p, i, FALSE);
372 13564124 : fe->hotclaim += footprint;
373 13564124 : if (footprint > fe->maxclaim)
374 5292627 : 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 12495345 : int last = 0, nxt = -1;
383 12495345 : lng nxtclaim = -1;
384 :
385 12495345 : MT_lock_set(&flow->flowlock);
386 12516942 : for (last = fe->pc - flow->start;
387 50572268 : last >= 0 && (i = flow->nodes[last]) > 0;
388 38055326 : last = flow->edges[last]) {
389 38055326 : if (flow->status[i].state == DFLOWpending
390 38050267 : && flow->status[i].blocks == 1) {
391 : /* find the one with the largest footprint */
392 10152965 : if (nxt == -1 || flow->status[i].argclaim > nxtclaim) {
393 5725276 : nxt = i;
394 5725276 : nxtclaim = flow->status[i].argclaim;
395 : }
396 : }
397 : }
398 : /* hot potato can not be removed, use alternative to proceed */
399 12516942 : if (nxt >= 0) {
400 5311553 : flow->status[nxt].state = DFLOWrunning;
401 5311553 : flow->status[nxt].blocks = 0;
402 5311553 : flow->status[nxt].hotclaim = fe->hotclaim;
403 5311553 : flow->status[nxt].argclaim += fe->hotclaim;
404 5311553 : if (flow->status[nxt].maxclaim < fe->maxclaim)
405 2311550 : flow->status[nxt].maxclaim = fe->maxclaim;
406 : fnxt = flow->status + nxt;
407 : }
408 12516942 : MT_lock_unset(&flow->flowlock);
409 : #endif
410 :
411 12501935 : q_enqueue(flow->done, fe);
412 12479085 : if (fnxt == 0 && profilerStatus) {
413 0 : profilerHeartbeatEvent("wait");
414 : }
415 : }
416 157018 : MT_lock_set(&dataflowLock);
417 157482 : if (GDKexiting() || ATOMIC_GET(&exiting) || free_count >= free_max) {
418 : locked = true;
419 : break;
420 : }
421 153476 : free_count++;
422 153476 : struct worker **tp = &workers;
423 926846 : while (*tp && *tp != t)
424 773370 : tp = &(*tp)->next;
425 153476 : assert(*tp && *tp == t);
426 153476 : *tp = t->next;
427 153476 : t->flag = FREE;
428 153476 : t->next = free_workers;
429 153476 : free_workers = t;
430 153476 : MT_lock_unset(&dataflowLock);
431 153476 : MT_thread_setworking("idle, waiting for new client");
432 153476 : MT_sema_down(&t->s);
433 153470 : if (GDKexiting() || ATOMIC_GET(&exiting))
434 : break;
435 152868 : assert(t->flag == WAITING);
436 : }
437 597 : if (!locked)
438 597 : MT_lock_set(&dataflowLock);
439 4603 : if (t->flag != FINISHING) {
440 3985 : struct worker **tp = t->flag == FREE ? &free_workers : &workers;
441 17337 : while (*tp && *tp != t)
442 13352 : tp = &(*tp)->next;
443 3985 : assert(*tp && *tp == t);
444 3985 : *tp = t->next;
445 3985 : t->flag = EXITED;
446 3985 : t->next = exited_workers;
447 3985 : exited_workers = t;
448 : }
449 4603 : MT_lock_unset(&dataflowLock);
450 4597 : GDKsetbuf(NULL);
451 4597 : }
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
461 329 : DFLOWinitialize(void)
462 : {
463 329 : int limit;
464 329 : int created = 0;
465 :
466 329 : MT_lock_set(&mal_contextLock);
467 329 : MT_lock_set(&dataflowLock);
468 329 : 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 : }
474 329 : free_max = GDKgetenv_int("dataflow_max_free",
475 : GDKnr_threads < 4 ? 4 : GDKnr_threads);
476 329 : todo = q_create("todo");
477 329 : if (todo == NULL) {
478 0 : MT_lock_unset(&dataflowLock);
479 0 : MT_lock_unset(&mal_contextLock);
480 0 : return -1;
481 : }
482 329 : limit = GDKnr_threads ? GDKnr_threads - 1 : 0;
483 2626 : while (limit > 0) {
484 2297 : limit--;
485 2297 : struct worker *t = GDKmalloc(sizeof(*t));
486 2297 : if (t == NULL) {
487 0 : TRC_CRITICAL(MAL_SERVER, "cannot allocate structure for worker");
488 0 : continue;
489 : }
490 2297 : *t = (struct worker) {
491 : .flag = RUNNING,
492 : .cntxt = ATOMIC_PTR_VAR_INIT(NULL),
493 : };
494 2297 : MT_sema_init(&t->s, 0, "DFLOWsema"); /* placeholder name */
495 2297 : 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 2297 : t->next = workers;
502 2297 : workers = t;
503 2297 : created++;
504 : }
505 : }
506 329 : 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 : }
514 329 : MT_lock_unset(&dataflowLock);
515 329 : MT_lock_unset(&mal_contextLock);
516 329 : 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 155191 : DFLOWinitBlk(DataFlow flow, MalBlkPtr mb, int size)
527 : {
528 155191 : int pc, i, j, k, l, n, etop = 0;
529 155191 : int *assign;
530 155191 : InstrPtr p;
531 :
532 155191 : if (flow == NULL)
533 0 : throw(MAL, "dataflow", "DFLOWinitBlk(): Called with flow == NULL");
534 155191 : if (mb == NULL)
535 0 : throw(MAL, "dataflow", "DFLOWinitBlk(): Called with mb == NULL");
536 155191 : assign = (int *) GDKzalloc(mb->vtop * sizeof(int));
537 155191 : if (assign == NULL)
538 0 : throw(MAL, "dataflow", SQLSTATE(HY013) MAL_MALLOC_FAIL);
539 155191 : etop = flow->stop - flow->start;
540 12677257 : for (n = 0, pc = flow->start; pc < flow->stop; pc++, n++) {
541 12522067 : p = getInstrPtr(mb, pc);
542 12522067 : 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 12522067 : flow->status[n].flow = flow;
550 12522067 : flow->status[n].pc = pc;
551 12522067 : flow->status[n].state = DFLOWpending;
552 12522067 : flow->status[n].cost = -1;
553 12522067 : ATOMIC_PTR_SET(&flow->status[n].flow->error, NULL);
554 :
555 : /* administer flow dependencies */
556 60139612 : for (j = p->retc; j < p->argc; j++) {
557 : /* list of instructions that wake n-th instruction up */
558 47617546 : if (!isVarConstant(mb, getArg(p, j)) && (k = assign[getArg(p, j)])) {
559 23410068 : assert(k < pc); /* only dependencies on earlier instructions */
560 : /* add edge to the target instruction for wakeup call */
561 23410068 : k -= flow->start;
562 23410068 : if (flow->nodes[k]) {
563 : /* add wakeup to tail of list */
564 211027428 : for (i = k; flow->edges[i] > 0; i = flow->edges[i])
565 : ;
566 20702285 : flow->nodes[etop] = n;
567 20702285 : flow->edges[etop] = -1;
568 20702285 : flow->edges[i] = etop;
569 20702285 : etop++;
570 20702285 : (void) size;
571 20702285 : 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 2707783 : flow->nodes[k] = n;
595 2707783 : flow->edges[k] = -1;
596 : }
597 :
598 23410067 : flow->status[n].blocks++;
599 : }
600 :
601 : /* list of instructions to be woken up explicitly */
602 47617545 : 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 25167630 : l = getEndScope(mb, getArg(p, j));
606 25167630 : if (l != pc && l < flow->stop && l > flow->start) {
607 : /* add edge to the target instruction for wakeup call */
608 14666277 : assert(pc < l); /* only dependencies on earlier instructions */
609 14666277 : l -= flow->start;
610 14666277 : if (flow->nodes[n]) {
611 : /* add wakeup to tail of list */
612 77783586 : for (i = n; flow->edges[i] > 0; i = flow->edges[i])
613 : ;
614 7565669 : flow->nodes[etop] = l;
615 7565669 : flow->edges[etop] = -1;
616 7565669 : flow->edges[i] = etop;
617 7565669 : etop++;
618 7565669 : 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 7100608 : flow->nodes[n] = l;
642 7100608 : flow->edges[n] = -1;
643 : }
644 14666277 : flow->status[l].blocks++;
645 : }
646 : }
647 : }
648 :
649 26118940 : for (j = 0; j < p->retc; j++)
650 13596874 : assign[getArg(p, j)] = pc; /* ensure recognition of dependency on first instruction and constant */
651 : }
652 155190 : GDKfree(assign);
653 :
654 155190 : 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 155192 : DFLOWscheduler(DataFlow flow, struct worker *w)
668 : {
669 155192 : int last;
670 155192 : int i;
671 155192 : int j;
672 155192 : InstrPtr p;
673 155192 : int tasks = 0, actions = 0;
674 155192 : str ret = MAL_SUCCEED;
675 155192 : FlowEvent fe, f = 0;
676 :
677 155192 : if (flow == NULL)
678 0 : throw(MAL, "dataflow", "DFLOWscheduler(): Called with flow == NULL");
679 155192 : actions = flow->stop - flow->start;
680 155192 : if (actions == 0)
681 0 : throw(MAL, "dataflow", "Empty dataflow block");
682 : /* initialize the eligible statements */
683 155192 : fe = flow->status;
684 :
685 155192 : ATOMIC_DEC(&flow->cntxt->workers);
686 155192 : MT_lock_set(&flow->flowlock);
687 12832762 : for (i = 0; i < actions; i++)
688 12522378 : if (fe[i].blocks == 0) {
689 978721 : p = getInstrPtr(flow->mb, fe[i].pc);
690 978721 : 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 978721 : fe[i].argclaim = 0;
697 5895915 : for (j = p->retc; j < p->argc; j++)
698 4917194 : fe[i].argclaim += getMemoryClaim(fe[0].flow->mb,
699 4917196 : fe[0].flow->stk, p, j, FALSE);
700 978719 : flow->status[i].state = DFLOWrunning;
701 978719 : q_enqueue(todo, flow->status + i);
702 : }
703 155192 : MT_lock_unset(&flow->flowlock);
704 155192 : MT_sema_up(&w->s);
705 :
706 155192 : while (actions != tasks) {
707 12521450 : f = q_dequeue(flow->done, NULL);
708 12521635 : if (ATOMIC_GET(&exiting))
709 : break;
710 12521635 : 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 12521635 : MT_lock_set(&flow->flowlock);
723 12521255 : tasks++;
724 12521255 : for (last = f->pc - flow->start;
725 50589877 : last >= 0 && (i = flow->nodes[last]) > 0; last = flow->edges[last])
726 38068773 : if (flow->status[i].state == DFLOWpending) {
727 32762350 : flow->status[i].argclaim += f->hotclaim;
728 32762350 : if (flow->status[i].blocks == 1) {
729 6231575 : flow->status[i].blocks--;
730 6231575 : flow->status[i].state = DFLOWrunning;
731 6231575 : q_enqueue(todo, flow->status + i);
732 : } else {
733 26530775 : flow->status[i].blocks--;
734 : }
735 : }
736 12676294 : MT_lock_unset(&flow->flowlock);
737 : }
738 : /* release the worker from its specific task (turn it into a
739 : * generic worker) */
740 155192 : ATOMIC_PTR_SET(&w->cntxt, NULL);
741 155192 : ATOMIC_INC(&flow->cntxt->workers);
742 : /* wrap up errors */
743 155192 : assert(flow->done->last == 0);
744 155192 : if ((ret = ATOMIC_PTR_XCG(&flow->error, NULL)) != NULL) {
745 429 : 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 4603 : finish_worker(struct worker *t)
754 : {
755 4603 : t->flag = FINISHING;
756 4603 : MT_lock_unset(&dataflowLock);
757 4603 : MT_join_thread(t->id);
758 4603 : MT_sema_destroy(&t->s);
759 4603 : ATOMIC_PTR_DESTROY(&t->cntxt);
760 4603 : GDKfree(t);
761 4603 : MT_lock_set(&dataflowLock);
762 4603 : }
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 155186 : runMALdataflow(Client cntxt, MalBlkPtr mb, int startpc, int stoppc,
780 : MalStkPtr stk)
781 : {
782 155186 : DataFlow flow = NULL;
783 155186 : str msg = MAL_SUCCEED;
784 155186 : int size;
785 155186 : bit *ret;
786 155186 : struct worker *t;
787 :
788 155186 : if (stk == NULL)
789 0 : throw(MAL, "dataflow", "runMALdataflow(): Called with stk == NULL");
790 155186 : ret = getArgReference_bit(stk, getInstrPtr(mb, startpc), 0);
791 155186 : *ret = FALSE;
792 :
793 155186 : assert(stoppc > startpc);
794 :
795 : /* check existence of workers */
796 155186 : if (todo == NULL) {
797 : /* create thread pool */
798 329 : if (GDKnr_threads <= 1 || DFLOWinitialize() < 0) {
799 : /* no threads created, run serially */
800 0 : *ret = TRUE;
801 0 : return MAL_SUCCEED;
802 : }
803 : }
804 155186 : 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 155186 : MT_lock_set(&dataflowLock);
809 : /* join with already exited threads */
810 156908 : while (exited_workers != NULL) {
811 1716 : assert(exited_workers->flag == EXITED);
812 1716 : struct worker *t = exited_workers;
813 1716 : exited_workers = exited_workers->next;
814 1716 : finish_worker(t);
815 : }
816 155192 : assert(cntxt != NULL);
817 155192 : if (free_workers != NULL) {
818 152878 : t = free_workers;
819 152878 : assert(t->flag == FREE);
820 152878 : assert(free_count > 0);
821 152878 : free_count--;
822 152878 : free_workers = t->next;
823 152878 : t->next = workers;
824 152878 : workers = t;
825 152878 : t->flag = WAITING;
826 152878 : ATOMIC_PTR_SET(&t->cntxt, cntxt);
827 152878 : MT_sema_up(&t->s);
828 : } else {
829 2314 : t = GDKmalloc(sizeof(*t));
830 2314 : if (t != NULL) {
831 2314 : *t = (struct worker) {
832 : .flag = WAITING,
833 : .cntxt = ATOMIC_PTR_VAR_INIT(cntxt),
834 : };
835 2314 : MT_sema_init(&t->s, 0, "DFLOWsema"); /* placeholder name */
836 2314 : 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 2314 : t->next = workers;
844 2314 : workers = t;
845 : }
846 : }
847 2314 : 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 155192 : MT_lock_unset(&dataflowLock);
855 :
856 155192 : flow = (DataFlow) GDKzalloc(sizeof(DataFlowRec));
857 155192 : if (flow == NULL)
858 0 : throw(MAL, "dataflow", SQLSTATE(HY013) MAL_MALLOC_FAIL);
859 :
860 155192 : size = DFLOWgraphSize(mb, startpc, stoppc);
861 155192 : size += stoppc - startpc;
862 :
863 310383 : *flow = (DataFlowRec) {
864 : .cntxt = cntxt,
865 : .mb = mb,
866 : .stk = stk,
867 155192 : .set_qry_ctx = MT_thread_get_qry_ctx() != NULL,
868 : /* keep real block count, exclude brackets */
869 155192 : .start = startpc + 1,
870 : .stop = stoppc,
871 155192 : .done = q_create("flow->done"),
872 155192 : .status = (FlowEvent) GDKzalloc((stoppc - startpc + 1) *
873 : sizeof(FlowEventRec)),
874 : .error = ATOMIC_PTR_VAR_INIT(NULL),
875 155192 : .nodes = (int *) GDKzalloc(sizeof(int) * size),
876 155192 : .edges = (int *) GDKzalloc(sizeof(int) * size),
877 : };
878 :
879 155191 : 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 155191 : 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 155191 : MT_lock_init(&flow->flowlock, "flow->flowlock");
898 155190 : msg = DFLOWinitBlk(flow, mb, size);
899 :
900 155192 : if (msg == MAL_SUCCEED)
901 155192 : msg = DFLOWscheduler(flow, t);
902 :
903 155192 : GDKfree(flow->status);
904 155192 : GDKfree(flow->edges);
905 155192 : GDKfree(flow->nodes);
906 155192 : q_destroy(flow->done);
907 155190 : MT_lock_destroy(&flow->flowlock);
908 155191 : ATOMIC_PTR_DESTROY(&flow->error);
909 155191 : GDKfree(flow);
910 :
911 : /* we created one worker, now tell one worker to exit again */
912 155192 : MT_lock_set(&todo->l);
913 155192 : todo->exitcount++;
914 155192 : MT_lock_unset(&todo->l);
915 155192 : MT_sema_up(&todo->s);
916 :
917 155192 : 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
932 328 : stopMALdataflow(void)
933 : {
934 328 : ATOMIC_SET(&exiting, 1);
935 328 : if (todo) {
936 328 : MT_lock_set(&dataflowLock);
937 : /* first wake up all running threads */
938 328 : int n = 0;
939 925 : for (struct worker *t = free_workers; t; t = t->next)
940 597 : n++;
941 2618 : for (struct worker *t = workers; t; t = t->next)
942 2290 : n++;
943 3215 : while (n-- > 0) {
944 : /* one UP for each thread we know about */
945 3215 : MT_sema_up(&todo->s);
946 : }
947 925 : while (free_workers) {
948 597 : struct worker *t = free_workers;
949 597 : assert(free_count > 0);
950 597 : free_count--;
951 597 : free_workers = free_workers->next;
952 597 : MT_sema_up(&t->s);
953 597 : finish_worker(t);
954 : }
955 349 : while (workers) {
956 21 : struct worker *t = workers;
957 21 : workers = workers->next;
958 21 : finish_worker(t);
959 : }
960 2597 : while (exited_workers) {
961 2269 : struct worker *t = exited_workers;
962 2269 : exited_workers = exited_workers->next;
963 2269 : finish_worker(t);
964 : }
965 328 : MT_lock_unset(&dataflowLock);
966 : }
967 328 : }
968 :
969 : void
970 328 : mal_dataflow_reset(void)
971 : {
972 328 : stopMALdataflow();
973 328 : workers = exited_workers = NULL;
974 328 : if (todo) {
975 328 : MT_lock_destroy(&todo->l);
976 328 : MT_sema_destroy(&todo->s);
977 328 : GDKfree(todo);
978 : }
979 328 : todo = 0; /* pending instructions */
980 328 : ATOMIC_SET(&exiting, 0);
981 328 : }
|
