xref: /plan9-contrib/sys/src/9/port/segment.c (revision 422e887e3271b1733565fbd5ed139f4b15a9ec7f)
1 #include	"u.h"
2 #include	"../port/lib.h"
3 #include	"mem.h"
4 #include	"dat.h"
5 #include	"fns.h"
6 #include	"../port/error.h"
7 
8 static void	imagereclaim(void);
9 static void	imagechanreclaim(void);
10 
11 #include "io.h"
12 
13 /*
14  * Attachable segment types
15  */
16 static Physseg physseg[10] = {
17 	{ SG_SHARED,	"shared",	0,	SEGMAXSIZE,	0, 	0 },
18 	{ SG_BSS,	"memory",	0,	SEGMAXSIZE,	0,	0 },
19 	{ 0,		0,		0,	0,		0,	0 },
20 };
21 
22 static Lock physseglock;
23 
24 #define NFREECHAN	64
25 #define IHASHSIZE	64
26 #define ihash(s)	imagealloc.hash[s%IHASHSIZE]
27 static struct Imagealloc
28 {
29 	Lock;
30 	Image	*free;
31 	Image	*hash[IHASHSIZE];
32 	QLock	ireclaim;	/* mutex on reclaiming free images */
33 
34 	Chan	**freechan;	/* free image channels */
35 	int	nfreechan;	/* number of free channels */
36 	int	szfreechan;	/* size of freechan array */
37 	QLock	fcreclaim;	/* mutex on reclaiming free channels */
38 }imagealloc;
39 
40 Segment* (*_globalsegattach)(Proc*, char*);
41 
42 void
initseg(void)43 initseg(void)
44 {
45 	Image *i, *ie;
46 
47 	imagealloc.free = xalloc(conf.nimage*sizeof(Image));
48 	if (imagealloc.free == nil)
49 		panic("initseg: no memory");
50 	ie = &imagealloc.free[conf.nimage-1];
51 	for(i = imagealloc.free; i < ie; i++)
52 		i->next = i+1;
53 	i->next = 0;
54 	imagealloc.freechan = malloc(NFREECHAN * sizeof(Chan*));
55 	imagealloc.szfreechan = NFREECHAN;
56 }
57 
58 Segment *
newseg(int type,ulong base,ulong size)59 newseg(int type, ulong base, ulong size)
60 {
61 	Segment *s;
62 	int mapsize;
63 
64 	if(size > (SEGMAPSIZE*PTEPERTAB))
65 		error(Enovmem);
66 
67 	s = smalloc(sizeof(Segment));
68 	s->ref = 1;
69 	s->type = type;
70 	s->base = base;
71 	s->top = base+(size*BY2PG);
72 	s->size = size;
73 	s->sema.prev = &s->sema;
74 	s->sema.next = &s->sema;
75 
76 	mapsize = ROUND(size, PTEPERTAB)/PTEPERTAB;
77 	if(mapsize > nelem(s->ssegmap)){
78 		mapsize *= 2;
79 		if(mapsize > SEGMAPSIZE)
80 			mapsize = SEGMAPSIZE;
81 		s->map = smalloc(mapsize*sizeof(Pte*));
82 		s->mapsize = mapsize;
83 	}
84 	else{
85 		s->map = s->ssegmap;
86 		s->mapsize = nelem(s->ssegmap);
87 	}
88 
89 	return s;
90 }
91 
92 void
putseg(Segment * s)93 putseg(Segment *s)
94 {
95 	Pte **pp, **emap;
96 	Image *i;
97 
98 	if(s == 0)
99 		return;
100 
101 	i = s->image;
102 	if(i != 0) {
103 		lock(i);
104 		lock(s);
105 		if(i->s == s && s->ref == 1)
106 			i->s = 0;
107 		unlock(i);
108 	}
109 	else
110 		lock(s);
111 
112 	s->ref--;
113 	if(s->ref != 0) {
114 		unlock(s);
115 		return;
116 	}
117 	unlock(s);
118 
119 	qlock(&s->lk);
120 	if(i)
121 		putimage(i);
122 
123 	emap = &s->map[s->mapsize];
124 	for(pp = s->map; pp < emap; pp++)
125 		if(*pp)
126 			freepte(s, *pp);
127 
128 	qunlock(&s->lk);
129 	if(s->map != s->ssegmap)
130 		free(s->map);
131 	if(s->profile != 0)
132 		free(s->profile);
133 	free(s);
134 }
135 
136 void
relocateseg(Segment * s,ulong offset)137 relocateseg(Segment *s, ulong offset)
138 {
139 	Page **pg, *x;
140 	Pte *pte, **p, **endpte;
141 
142 	endpte = &s->map[s->mapsize];
143 	for(p = s->map; p < endpte; p++) {
144 		if(*p == 0)
145 			continue;
146 		pte = *p;
147 		for(pg = pte->first; pg <= pte->last; pg++) {
148 			if(x = *pg)
149 				x->va += offset;
150 		}
151 	}
152 }
153 
154 Segment*
dupseg(Segment ** seg,int segno,int share)155 dupseg(Segment **seg, int segno, int share)
156 {
157 	int i, size;
158 	Pte *pte;
159 	Segment *n, *s;
160 
161 	SET(n);
162 	s = seg[segno];
163 
164 	qlock(&s->lk);
165 	if(waserror()){
166 		qunlock(&s->lk);
167 		nexterror();
168 	}
169 	switch(s->type&SG_TYPE) {
170 	case SG_TEXT:		/* New segment shares pte set */
171 	case SG_SHARED:
172 	case SG_PHYSICAL:
173 		goto sameseg;
174 
175 	case SG_STACK:
176 		n = newseg(s->type, s->base, s->size);
177 		break;
178 
179 	case SG_BSS:		/* Just copy on write */
180 		if(share)
181 			goto sameseg;
182 		n = newseg(s->type, s->base, s->size);
183 		break;
184 
185 	case SG_DATA:		/* Copy on write plus demand load info */
186 		if(segno == TSEG){
187 			poperror();
188 			qunlock(&s->lk);
189 			return data2txt(s);
190 		}
191 
192 		if(share)
193 			goto sameseg;
194 		n = newseg(s->type, s->base, s->size);
195 
196 		incref(s->image);
197 		n->image = s->image;
198 		n->fstart = s->fstart;
199 		n->flen = s->flen;
200 		break;
201 	}
202 	size = s->mapsize;
203 	for(i = 0; i < size; i++)
204 		if(pte = s->map[i])
205 			n->map[i] = ptecpy(pte);
206 
207 	n->flushme = s->flushme;
208 	if(s->ref > 1)
209 		procflushseg(s);
210 	poperror();
211 	qunlock(&s->lk);
212 	return n;
213 
214 sameseg:
215 	incref(s);
216 	poperror();
217 	qunlock(&s->lk);
218 	return s;
219 }
220 
221 void
segpage(Segment * s,Page * p)222 segpage(Segment *s, Page *p)
223 {
224 	Pte **pte;
225 	ulong off;
226 	Page **pg;
227 
228 	if(p->va < s->base || p->va >= s->top)
229 		panic("segpage");
230 
231 	off = p->va - s->base;
232 	pte = &s->map[off/PTEMAPMEM];
233 	if(*pte == 0)
234 		*pte = ptealloc();
235 
236 	pg = &(*pte)->pages[(off&(PTEMAPMEM-1))/BY2PG];
237 	*pg = p;
238 	if(pg < (*pte)->first)
239 		(*pte)->first = pg;
240 	if(pg > (*pte)->last)
241 		(*pte)->last = pg;
242 }
243 
244 Image*
attachimage(int type,Chan * c,ulong base,ulong len)245 attachimage(int type, Chan *c, ulong base, ulong len)
246 {
247 	Image *i, **l;
248 
249 	/* reclaim any free channels from reclaimed segments */
250 	if(imagealloc.nfreechan)
251 		imagechanreclaim();
252 
253 	lock(&imagealloc);
254 
255 	/*
256 	 * Search the image cache for remains of the text from a previous
257 	 * or currently running incarnation
258 	 */
259 	for(i = ihash(c->qid.path); i; i = i->hash) {
260 		if(c->qid.path == i->qid.path) {
261 			lock(i);
262 			if(eqqid(c->qid, i->qid) &&
263 			   eqqid(c->mqid, i->mqid) &&
264 			   c->mchan == i->mchan &&
265 			   c->type == i->type) {
266 				goto found;
267 			}
268 			unlock(i);
269 		}
270 	}
271 
272 	/*
273 	 * imagereclaim dumps pages from the free list which are cached by image
274 	 * structures. This should free some image structures.
275 	 */
276 	while(!(i = imagealloc.free)) {
277 		unlock(&imagealloc);
278 		imagereclaim();
279 		sched();
280 		lock(&imagealloc);
281 	}
282 
283 	imagealloc.free = i->next;
284 
285 	lock(i);
286 	incref(c);
287 	i->c = c;
288 	i->type = c->type;
289 	i->qid = c->qid;
290 	i->mqid = c->mqid;
291 	i->mchan = c->mchan;
292 	l = &ihash(c->qid.path);
293 	i->hash = *l;
294 	*l = i;
295 found:
296 	unlock(&imagealloc);
297 
298 	if(i->s == 0) {
299 		/* Disaster after commit in exec */
300 		if(waserror()) {
301 			unlock(i);
302 			pexit(Enovmem, 1);
303 		}
304 		i->s = newseg(type, base, len);
305 		i->s->image = i;
306 		i->ref++;
307 		poperror();
308 	}
309 	else
310 		incref(i->s);
311 
312 	return i;
313 }
314 
315 static struct {
316 	int	calls;			/* times imagereclaim was called */
317 	int	loops;			/* times the main loop was run */
318 	uvlong	ticks;			/* total time in the main loop */
319 	uvlong	maxt;			/* longest time in main loop */
320 } irstats;
321 
322 static void
imagereclaim(void)323 imagereclaim(void)
324 {
325 	int n;
326 	Page *p;
327 	uvlong ticks;
328 
329 	irstats.calls++;
330 	/* Somebody is already cleaning the page cache */
331 	if(!canqlock(&imagealloc.ireclaim))
332 		return;
333 
334 	lock(&palloc);
335 	ticks = fastticks(nil);
336 	n = 0;
337 	/*
338 	 * All the pages with images backing them are at the
339 	 * end of the list (see putpage) so start there and work
340 	 * backward.
341 	 */
342 	for(p = palloc.tail; p && p->image && n<1000; p = p->prev) {
343 		if(p->ref == 0 && canlock(p)) {
344 			if(p->ref == 0) {
345 				n++;
346 				uncachepage(p);
347 				pageunchain(p);
348 				pagechainhead(p);
349 			}
350 			unlock(p);
351 		}
352 	}
353 	ticks = fastticks(nil) - ticks;
354 	unlock(&palloc);
355 	irstats.loops++;
356 	irstats.ticks += ticks;
357 	if(ticks > irstats.maxt)
358 		irstats.maxt = ticks;
359 	//print("T%llud+", ticks);
360 	qunlock(&imagealloc.ireclaim);
361 }
362 
363 /*
364  *  since close can block, this has to be called outside of
365  *  spin locks.
366  */
367 static void
imagechanreclaim(void)368 imagechanreclaim(void)
369 {
370 	Chan *c;
371 
372 	/* Somebody is already cleaning the image chans */
373 	if(!canqlock(&imagealloc.fcreclaim))
374 		return;
375 
376 	/*
377 	 * We don't have to recheck that nfreechan > 0 after we
378 	 * acquire the lock, because we're the only ones who decrement
379 	 * it (the other lock contender increments it), and there's only
380 	 * one of us thanks to the qlock above.
381 	 */
382 	while(imagealloc.nfreechan > 0){
383 		lock(&imagealloc);
384 		imagealloc.nfreechan--;
385 		c = imagealloc.freechan[imagealloc.nfreechan];
386 		unlock(&imagealloc);
387 		cclose(c);
388 	}
389 
390 	qunlock(&imagealloc.fcreclaim);
391 }
392 
393 void
putimage(Image * i)394 putimage(Image *i)
395 {
396 	Chan *c, **cp;
397 	Image *f, **l;
398 
399 	if(i->notext)
400 		return;
401 
402 	lock(i);
403 	if(--i->ref == 0) {
404 		l = &ihash(i->qid.path);
405 		mkqid(&i->qid, ~0, ~0, QTFILE);
406 		unlock(i);
407 		c = i->c;
408 
409 		lock(&imagealloc);
410 		for(f = *l; f; f = f->hash) {
411 			if(f == i) {
412 				*l = i->hash;
413 				break;
414 			}
415 			l = &f->hash;
416 		}
417 
418 		i->next = imagealloc.free;
419 		imagealloc.free = i;
420 
421 		/* defer freeing channel till we're out of spin lock's */
422 		if(imagealloc.nfreechan == imagealloc.szfreechan){
423 			imagealloc.szfreechan += NFREECHAN;
424 			cp = malloc(imagealloc.szfreechan*sizeof(Chan*));
425 			if(cp == nil)
426 				panic("putimage");
427 			memmove(cp, imagealloc.freechan, imagealloc.nfreechan*sizeof(Chan*));
428 			free(imagealloc.freechan);
429 			imagealloc.freechan = cp;
430 		}
431 		imagealloc.freechan[imagealloc.nfreechan++] = c;
432 		unlock(&imagealloc);
433 
434 		return;
435 	}
436 	unlock(i);
437 }
438 
439 long
ibrk(ulong addr,int seg)440 ibrk(ulong addr, int seg)
441 {
442 	Segment *s, *ns;
443 	ulong newtop, newsize;
444 	int i, mapsize;
445 	Pte **map;
446 
447 	s = up->seg[seg];
448 	if(s == 0)
449 		error(Ebadarg);
450 
451 	if(addr == 0)
452 		return s->base;
453 
454 	qlock(&s->lk);
455 
456 	/* We may start with the bss overlapping the data */
457 	if(addr < s->base) {
458 		if(seg != BSEG || up->seg[DSEG] == 0 || addr < up->seg[DSEG]->base) {
459 			qunlock(&s->lk);
460 			error(Enovmem);
461 		}
462 		addr = s->base;
463 	}
464 
465 	newtop = PGROUND(addr);
466 	newsize = (newtop-s->base)/BY2PG;
467 	if(newtop < s->top) {
468 		/*
469 		 * do not shrink a segment shared with other procs, as the
470 		 * to-be-freed address space may have been passed to the kernel
471 		 * already by another proc and is past the validaddr stage.
472 		 */
473 		if(s->ref > 1){
474 			qunlock(&s->lk);
475 			error(Einuse);
476 		}
477 		mfreeseg(s, newtop, (s->top-newtop)/BY2PG);
478 		s->top = newtop;
479 		s->size = newsize;
480 		qunlock(&s->lk);
481 		flushmmu();
482 		return 0;
483 	}
484 
485 	for(i = 0; i < NSEG; i++) {
486 		ns = up->seg[i];
487 		if(ns == 0 || ns == s)
488 			continue;
489 		if(newtop > ns->base && s->base < ns->top) {
490 			qunlock(&s->lk);
491 			error(Esoverlap);
492 		}
493 	}
494 
495 	if(newsize > (SEGMAPSIZE*PTEPERTAB)) {
496 		qunlock(&s->lk);
497 		error(Enovmem);
498 	}
499 	mapsize = ROUND(newsize, PTEPERTAB)/PTEPERTAB;
500 	if(mapsize > s->mapsize){
501 		map = smalloc(mapsize*sizeof(Pte*));
502 		memmove(map, s->map, s->mapsize*sizeof(Pte*));
503 		if(s->map != s->ssegmap)
504 			free(s->map);
505 		s->map = map;
506 		s->mapsize = mapsize;
507 	}
508 
509 	s->top = newtop;
510 	s->size = newsize;
511 	qunlock(&s->lk);
512 	return 0;
513 }
514 
515 /*
516  *  called with s->lk locked
517  */
518 void
mfreeseg(Segment * s,ulong start,int pages)519 mfreeseg(Segment *s, ulong start, int pages)
520 {
521 	int i, j, size;
522 	ulong soff;
523 	Page *pg;
524 	Page *list;
525 
526 	soff = start-s->base;
527 	j = (soff&(PTEMAPMEM-1))/BY2PG;
528 
529 	size = s->mapsize;
530 	list = nil;
531 	for(i = soff/PTEMAPMEM; i < size; i++) {
532 		if(pages <= 0)
533 			break;
534 		if(s->map[i] == 0) {
535 			pages -= PTEPERTAB-j;
536 			j = 0;
537 			continue;
538 		}
539 		while(j < PTEPERTAB) {
540 			pg = s->map[i]->pages[j];
541 			/*
542 			 * We want to zero s->map[i]->page[j] and putpage(pg),
543 			 * but we have to make sure other processors flush the
544 			 * entry from their TLBs before the page is freed.
545 			 * We construct a list of the pages to be freed, zero
546 			 * the entries, then (below) call procflushseg, and call
547 			 * putpage on the whole list.
548 			 *
549 			 * Swapped-out pages don't appear in TLBs, so it's okay
550 			 * to putswap those pages before procflushseg.
551 			 */
552 			if(pg){
553 				if(onswap(pg))
554 					putswap(pg);
555 				else{
556 					pg->next = list;
557 					list = pg;
558 				}
559 				s->map[i]->pages[j] = 0;
560 			}
561 			if(--pages == 0)
562 				goto out;
563 			j++;
564 		}
565 		j = 0;
566 	}
567 out:
568 	/* flush this seg in all other processes */
569 	if(s->ref > 1)
570 		procflushseg(s);
571 
572 	/* free the pages */
573 	for(pg = list; pg != nil; pg = list){
574 		list = list->next;
575 		putpage(pg);
576 	}
577 }
578 
579 Segment*
isoverlap(Proc * p,ulong va,int len)580 isoverlap(Proc *p, ulong va, int len)
581 {
582 	int i;
583 	Segment *ns;
584 	ulong newtop;
585 
586 	newtop = va+len;
587 	for(i = 0; i < NSEG; i++) {
588 		ns = p->seg[i];
589 		if(ns == 0)
590 			continue;
591 		if(newtop > ns->base && va < ns->top)
592 			return ns;
593 	}
594 	return nil;
595 }
596 
597 int
addphysseg(Physseg * new)598 addphysseg(Physseg* new)
599 {
600 	Physseg *ps;
601 
602 	/*
603 	 * Check not already entered and there is room
604 	 * for a new entry and the terminating null entry.
605 	 */
606 	lock(&physseglock);
607 	for(ps = physseg; ps->name; ps++){
608 		if(strcmp(ps->name, new->name) == 0){
609 			unlock(&physseglock);
610 			return -1;
611 		}
612 	}
613 	if(ps-physseg >= nelem(physseg)-2){
614 		unlock(&physseglock);
615 		return -1;
616 	}
617 
618 	*ps = *new;
619 	unlock(&physseglock);
620 
621 	return 0;
622 }
623 
624 int
isphysseg(char * name)625 isphysseg(char *name)
626 {
627 	Physseg *ps;
628 	int rv = 0;
629 
630 	lock(&physseglock);
631 	for(ps = physseg; ps->name; ps++){
632 		if(strcmp(ps->name, name) == 0){
633 			rv = 1;
634 			break;
635 		}
636 	}
637 	unlock(&physseglock);
638 	return rv;
639 }
640 
641 ulong
segattach(Proc * p,ulong attr,char * name,ulong va,ulong len)642 segattach(Proc *p, ulong attr, char *name, ulong va, ulong len)
643 {
644 	int sno;
645 	Segment *s, *os;
646 	Physseg *ps;
647 
648 	if(va != 0 && va >= USTKTOP)
649 		error(Ebadarg);
650 
651 	validaddr((ulong)name, 1, 0);
652 	vmemchr(name, 0, ~0);
653 
654 	for(sno = 0; sno < NSEG; sno++)
655 		if(p->seg[sno] == nil && sno != ESEG)
656 			break;
657 
658 	if(sno == NSEG)
659 		error(Enovmem);
660 
661 	/*
662 	 *  first look for a global segment with the
663 	 *  same name
664 	 */
665 	if(_globalsegattach != nil){
666 		s = (*_globalsegattach)(p, name);
667 		if(s != nil){
668 			p->seg[sno] = s;
669 			return s->base;
670 		}
671 	}
672 
673 	len = PGROUND(len);
674 	if(len == 0)
675 		error(Ebadarg);
676 
677 	/*
678 	 * Find a hole in the address space.
679 	 * Starting at the lowest possible stack address - len,
680 	 * check for an overlapping segment, and repeat at the
681 	 * base of that segment - len until either a hole is found
682 	 * or the address space is exhausted.  Ensure that we don't
683 	 * map the zero page.
684 	 */
685 	if(va == 0) {
686 		for (os = p->seg[SSEG]; os != nil; os = isoverlap(p, va, len)) {
687 			va = os->base;
688 			if(len >= va)
689 				error(Enovmem);
690 			va -= len;
691 		}
692 		va &= ~(BY2PG-1);
693 	} else {
694 		va &= ~(BY2PG-1);
695 		if(va == 0 || va >= USTKTOP)
696 			error(Ebadarg);
697 	}
698 
699 	if(isoverlap(p, va, len) != nil)
700 		error(Esoverlap);
701 
702 	for(ps = physseg; ps->name; ps++)
703 		if(strcmp(name, ps->name) == 0)
704 			goto found;
705 
706 	error(Ebadarg);
707 found:
708 	if(len > ps->size)
709 		error(Enovmem);
710 
711 	attr &= ~SG_TYPE;		/* Turn off what is not allowed */
712 	attr |= ps->attr;		/* Copy in defaults */
713 
714 	s = newseg(attr, va, len/BY2PG);
715 	s->pseg = ps;
716 	p->seg[sno] = s;
717 
718 	return va;
719 }
720 
721 void
pteflush(Pte * pte,int s,int e)722 pteflush(Pte *pte, int s, int e)
723 {
724 	int i;
725 	Page *p;
726 
727 	for(i = s; i < e; i++) {
728 		p = pte->pages[i];
729 		if(pagedout(p) == 0)
730 			memset(p->cachectl, PG_TXTFLUSH, sizeof(p->cachectl));
731 	}
732 }
733 
734 long
syssegflush(ulong * arg)735 syssegflush(ulong *arg)
736 {
737 	Segment *s;
738 	ulong addr, l;
739 	Pte *pte;
740 	int chunk, ps, pe, len;
741 
742 	addr = arg[0];
743 	len = arg[1];
744 
745 	while(len > 0) {
746 		s = seg(up, addr, 1);
747 		if(s == 0)
748 			error(Ebadarg);
749 
750 		s->flushme = 1;
751 	more:
752 		l = len;
753 		if(addr+l > s->top)
754 			l = s->top - addr;
755 
756 		ps = addr-s->base;
757 		pte = s->map[ps/PTEMAPMEM];
758 		ps &= PTEMAPMEM-1;
759 		pe = PTEMAPMEM;
760 		if(pe-ps > l){
761 			pe = ps + l;
762 			pe = (pe+BY2PG-1)&~(BY2PG-1);
763 		}
764 		if(pe == ps) {
765 			qunlock(&s->lk);
766 			error(Ebadarg);
767 		}
768 
769 		if(pte)
770 			pteflush(pte, ps/BY2PG, pe/BY2PG);
771 
772 		chunk = pe-ps;
773 		len -= chunk;
774 		addr += chunk;
775 
776 		if(len > 0 && addr < s->top)
777 			goto more;
778 
779 		qunlock(&s->lk);
780 	}
781 	flushmmu();
782 	return 0;
783 }
784 
785 void
segclock(ulong pc)786 segclock(ulong pc)
787 {
788 	Segment *s;
789 
790 	s = up->seg[TSEG];
791 	if(s == 0 || s->profile == 0)
792 		return;
793 
794 	s->profile[0] += TK2MS(1);
795 	if(pc >= s->base && pc < s->top) {
796 		pc -= s->base;
797 		s->profile[pc>>LRESPROF] += TK2MS(1);
798 	}
799 }
800 
801