xref: /netbsd-src/sys/uvm/uvm_pglist.c (revision a8c74629f602faa0ccf8a463757d7baf858bbf3a)
1 /*	$NetBSD: uvm_pglist.c,v 1.86 2020/10/07 17:51:50 chs Exp $	*/
2 
3 /*-
4  * Copyright (c) 1997, 2019 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9  * NASA Ames Research Center, and by Andrew Doran.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30  * POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 /*
34  * uvm_pglist.c: pglist functions
35  */
36 
37 #include <sys/cdefs.h>
38 __KERNEL_RCSID(0, "$NetBSD: uvm_pglist.c,v 1.86 2020/10/07 17:51:50 chs Exp $");
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/cpu.h>
43 
44 #include <uvm/uvm.h>
45 #include <uvm/uvm_pdpolicy.h>
46 #include <uvm/uvm_pgflcache.h>
47 
48 #ifdef VM_PAGE_ALLOC_MEMORY_STATS
49 #define	STAT_INCR(v)	(v)++
50 #define	STAT_DECR(v)	do { \
51 		if ((v) == 0) \
52 			printf("%s:%d -- Already 0!\n", __FILE__, __LINE__); \
53 		else \
54 			(v)--; \
55 	} while (/*CONSTCOND*/ 0)
56 u_long	uvm_pglistalloc_npages;
57 #else
58 #define	STAT_INCR(v)
59 #define	STAT_DECR(v)
60 #endif
61 
62 kmutex_t uvm_pglistalloc_contig_lock;
63 
64 /*
65  * uvm_pglistalloc: allocate a list of pages
66  *
67  * => allocated pages are placed onto an rlist.  rlist is
68  *    initialized by uvm_pglistalloc.
69  * => returns 0 on success or errno on failure
70  * => implementation allocates a single segment if any constraints are
71  *	imposed by call arguments.
72  * => doesn't take into account clean non-busy pages on inactive list
73  *	that could be used(?)
74  * => params:
75  *	size		the size of the allocation, rounded to page size.
76  *	low		the low address of the allowed allocation range.
77  *	high		the high address of the allowed allocation range.
78  *	alignment	memory must be aligned to this power-of-two boundary.
79  *	boundary	no segment in the allocation may cross this
80  *			power-of-two boundary (relative to zero).
81  */
82 
83 static void
84 uvm_pglist_add(struct vm_page *pg, struct pglist *rlist)
85 {
86 	struct pgfreelist *pgfl;
87 	struct pgflbucket *pgb;
88 
89 	pgfl = &uvm.page_free[uvm_page_get_freelist(pg)];
90 	pgb = pgfl->pgfl_buckets[uvm_page_get_bucket(pg)];
91 
92 #ifdef UVMDEBUG
93 	struct vm_page *tp;
94 	LIST_FOREACH(tp, &pgb->pgb_colors[VM_PGCOLOR(pg)], pageq.list) {
95 		if (tp == pg)
96 			break;
97 	}
98 	if (tp == NULL)
99 		panic("uvm_pglistalloc: page not on freelist");
100 #endif
101 	LIST_REMOVE(pg, pageq.list);
102 	pgb->pgb_nfree--;
103     	CPU_COUNT(CPU_COUNT_FREEPAGES, -1);
104 	pg->flags = PG_CLEAN;
105 	pg->uobject = NULL;
106 	pg->uanon = NULL;
107 	TAILQ_INSERT_TAIL(rlist, pg, pageq.queue);
108 	STAT_INCR(uvm_pglistalloc_npages);
109 }
110 
111 static int
112 uvm_pglistalloc_c_ps(uvm_physseg_t psi, int num, paddr_t low, paddr_t high,
113     paddr_t alignment, paddr_t boundary, struct pglist *rlist)
114 {
115 	signed int candidate, limit, candidateidx, end, idx, skip;
116 	int pagemask;
117 	bool second_pass;
118 #ifdef DEBUG
119 	paddr_t idxpa, lastidxpa;
120 	paddr_t cidx = 0;	/* XXX: GCC */
121 #endif
122 #ifdef PGALLOC_VERBOSE
123 	printf("pgalloc: contig %d pgs from psi %d\n", num, psi);
124 #endif
125 
126 	low = atop(low);
127 	high = atop(high);
128 	alignment = atop(alignment);
129 
130 	/*
131 	 * Make sure that physseg falls within with range to be allocated from.
132 	 */
133 	if (high <= uvm_physseg_get_avail_start(psi) || low >= uvm_physseg_get_avail_end(psi))
134 		return 0;
135 
136 	/*
137 	 * We start our search at the just after where the last allocation
138 	 * succeeded.
139 	 */
140 	candidate = roundup2(uimax(low, uvm_physseg_get_avail_start(psi) +
141 		uvm_physseg_get_start_hint(psi)), alignment);
142 	limit = uimin(high, uvm_physseg_get_avail_end(psi));
143 	pagemask = ~((boundary >> PAGE_SHIFT) - 1);
144 	skip = 0;
145 	second_pass = false;
146 
147 	for (;;) {
148 		bool ok = true;
149 		signed int cnt;
150 
151 		if (candidate + num > limit) {
152 			if (uvm_physseg_get_start_hint(psi) == 0 || second_pass) {
153 				/*
154 				 * We've run past the allowable range.
155 				 */
156 				return 0; /* FAIL = 0 pages*/
157 			}
158 			/*
159 			 * We've wrapped around the end of this segment
160 			 * so restart at the beginning but now our limit
161 			 * is were we started.
162 			 */
163 			second_pass = true;
164 			candidate = roundup2(uimax(low, uvm_physseg_get_avail_start(psi)), alignment);
165 			limit = uimin(limit, uvm_physseg_get_avail_start(psi) +
166 			    uvm_physseg_get_start_hint(psi));
167 			skip = 0;
168 			continue;
169 		}
170 		if (boundary != 0 &&
171 		    ((candidate ^ (candidate + num - 1)) & pagemask) != 0) {
172 			/*
173 			 * Region crosses boundary. Jump to the boundary
174 			 * just crossed and ensure alignment.
175 			 */
176 			candidate = (candidate + num - 1) & pagemask;
177 			candidate = roundup2(candidate, alignment);
178 			skip = 0;
179 			continue;
180 		}
181 #ifdef DEBUG
182 		/*
183 		 * Make sure this is a managed physical page.
184 		 */
185 
186 		if (uvm_physseg_find(candidate, &cidx) != psi)
187 			panic("pgalloc contig: botch1");
188 		if (cidx != candidate - uvm_physseg_get_start(psi))
189 			panic("pgalloc contig: botch2");
190 		if (uvm_physseg_find(candidate + num - 1, &cidx) != psi)
191 			panic("pgalloc contig: botch3");
192 		if (cidx != candidate - uvm_physseg_get_start(psi) + num - 1)
193 			panic("pgalloc contig: botch4");
194 #endif
195 		candidateidx = candidate - uvm_physseg_get_start(psi);
196 		end = candidateidx + num;
197 
198 		/*
199 		 * Found a suitable starting page.  See if the range is free.
200 		 */
201 #ifdef PGALLOC_VERBOSE
202 		printf("%s: psi=%d candidate=%#x end=%#x skip=%#x, align=%#"PRIxPADDR,
203 		    __func__, psi, candidateidx, end, skip, alignment);
204 #endif
205 		/*
206 		 * We start at the end and work backwards since if we find a
207 		 * non-free page, it makes no sense to continue.
208 		 *
209 		 * But on the plus size we have "vetted" some number of free
210 		 * pages.  If this iteration fails, we may be able to skip
211 		 * testing most of those pages again in the next pass.
212 		 */
213 		for (idx = end - 1; idx >= candidateidx + skip; idx--) {
214 			if (VM_PAGE_IS_FREE(uvm_physseg_get_pg(psi, idx)) == 0) {
215 				ok = false;
216 				break;
217 			}
218 
219 #ifdef DEBUG
220 			if (idx > candidateidx) {
221 				idxpa = VM_PAGE_TO_PHYS(uvm_physseg_get_pg(psi, idx));
222 				lastidxpa = VM_PAGE_TO_PHYS(uvm_physseg_get_pg(psi, idx - 1));
223 				if ((lastidxpa + PAGE_SIZE) != idxpa) {
224 					/*
225 					 * Region not contiguous.
226 					 */
227 					panic("pgalloc contig: botch5");
228 				}
229 				if (boundary != 0 &&
230 				    ((lastidxpa ^ idxpa) & ~(boundary - 1))
231 				    != 0) {
232 					/*
233 					 * Region crosses boundary.
234 					 */
235 					panic("pgalloc contig: botch6");
236 				}
237 			}
238 #endif
239 		}
240 
241 		if (ok) {
242 			while (skip-- > 0) {
243 				KDASSERT(VM_PAGE_IS_FREE(uvm_physseg_get_pg(psi, candidateidx + skip)));
244 			}
245 #ifdef PGALLOC_VERBOSE
246 			printf(": ok\n");
247 #endif
248 			break;
249 		}
250 
251 #ifdef PGALLOC_VERBOSE
252 		printf(": non-free at %#x\n", idx - candidateidx);
253 #endif
254 		/*
255 		 * count the number of pages we can advance
256 		 * since we know they aren't all free.
257 		 */
258 		cnt = idx + 1 - candidateidx;
259 		/*
260 		 * now round up that to the needed alignment.
261 		 */
262 		cnt = roundup2(cnt, alignment);
263 		/*
264 		 * The number of pages we can skip checking
265 		 * (might be 0 if cnt > num).
266 		 */
267 		skip = uimax(num - cnt, 0);
268 		candidate += cnt;
269 	}
270 
271 	/*
272 	 * we have a chunk of memory that conforms to the requested constraints.
273 	 */
274 	for (idx = candidateidx; idx < end; idx++)
275 		uvm_pglist_add(uvm_physseg_get_pg(psi, idx), rlist);
276 
277 	/*
278 	 * the next time we need to search this segment, start after this
279 	 * chunk of pages we just allocated.
280 	 */
281 	uvm_physseg_set_start_hint(psi, candidate + num -
282 	    uvm_physseg_get_avail_start(psi));
283 	KASSERTMSG(uvm_physseg_get_start_hint(psi) <=
284 	    uvm_physseg_get_avail_end(psi) - uvm_physseg_get_avail_start(psi),
285 	    "%x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
286 	    candidate + num,
287 	    uvm_physseg_get_start_hint(psi), uvm_physseg_get_start_hint(psi),
288 	    uvm_physseg_get_avail_end(psi), uvm_physseg_get_avail_start(psi),
289 	    uvm_physseg_get_avail_end(psi) - uvm_physseg_get_avail_start(psi));
290 
291 #ifdef PGALLOC_VERBOSE
292 	printf("got %d pgs\n", num);
293 #endif
294 	return num; /* number of pages allocated */
295 }
296 
297 static int
298 uvm_pglistalloc_contig_aggressive(int num, paddr_t low, paddr_t high,
299     paddr_t alignment, paddr_t boundary, struct pglist *rlist)
300 {
301 	struct vm_page *pg;
302 	struct pglist tmp;
303 	paddr_t pa, off, spa, amask, bmask, rlo, rhi;
304 	uvm_physseg_t upm;
305 	int error, i, run, acnt;
306 
307 	/*
308 	 * Allocate pages the normal way and for each new page, check if
309 	 * the page completes a range satisfying the request.
310 	 * The pagedaemon will evict pages as we go and we are very likely
311 	 * to get compatible pages eventually.
312 	 */
313 
314 	error = ENOMEM;
315 	TAILQ_INIT(&tmp);
316 	acnt = atop(alignment);
317 	amask = ~(alignment - 1);
318 	bmask = ~(boundary - 1);
319 	KASSERT(bmask <= amask);
320 	mutex_enter(&uvm_pglistalloc_contig_lock);
321 	while (uvm_reclaimable()) {
322 		pg = uvm_pagealloc(NULL, 0, NULL, 0);
323 		if (pg == NULL) {
324 			uvm_wait("pglac2");
325 			continue;
326 		}
327 		pg->flags |= PG_PGLCA;
328 		TAILQ_INSERT_HEAD(&tmp, pg, pageq.queue);
329 
330 		pa = VM_PAGE_TO_PHYS(pg);
331 		if (pa < low || pa >= high) {
332 			continue;
333 		}
334 
335 		upm = uvm_physseg_find(atop(pa), &off);
336 		KASSERT(uvm_physseg_valid_p(upm));
337 
338 		spa = pa & amask;
339 
340 		/*
341 		 * Look backward for at most num - 1 pages, back to
342 		 * the highest of:
343 		 *  - the first page in the physseg
344 		 *  - the specified low address
345 		 *  - num-1 pages before the one we just allocated
346 		 *  - the start of the boundary range containing pa
347 		 * all rounded up to alignment.
348 		 */
349 
350 		rlo = roundup2(ptoa(uvm_physseg_get_avail_start(upm)), alignment);
351 		rlo = MAX(rlo, roundup2(low, alignment));
352 		rlo = MAX(rlo, roundup2(pa - ptoa(num - 1), alignment));
353 		if (boundary) {
354 			rlo = MAX(rlo, spa & bmask);
355 		}
356 
357 		/*
358 		 * Look forward as far as the lowest of:
359 		 *  - the last page of the physseg
360 		 *  - the specified high address
361 		 *  - the boundary after pa
362 		 */
363 
364 		rhi = ptoa(uvm_physseg_get_avail_end(upm));
365 		rhi = MIN(rhi, high);
366 		if (boundary) {
367 			rhi = MIN(rhi, rounddown2(pa, boundary) + boundary);
368 		}
369 
370 		/*
371 		 * Make sure our range to consider is big enough.
372 		 */
373 
374 		if (rhi - rlo < ptoa(num)) {
375 			continue;
376 		}
377 
378 		run = 0;
379 		while (spa > rlo) {
380 
381 			/*
382 			 * Examine pages before spa in groups of acnt.
383 			 * If all the pages in a group are marked then add
384 			 * these pages to the run.
385 			 */
386 
387 			for (i = 0; i < acnt; i++) {
388 				pg = PHYS_TO_VM_PAGE(spa - alignment + ptoa(i));
389 				if ((pg->flags & PG_PGLCA) == 0) {
390 					break;
391 				}
392 			}
393 			if (i < acnt) {
394 				break;
395 			}
396 			spa -= alignment;
397 			run += acnt;
398 		}
399 
400 		/*
401 		 * Look forward for any remaining pages.
402 		 */
403 
404 		for (; run < num; run++) {
405 			pg = PHYS_TO_VM_PAGE(spa + ptoa(run));
406 			if ((pg->flags & PG_PGLCA) == 0) {
407 				break;
408 			}
409 		}
410 		if (run < num) {
411 			continue;
412 		}
413 
414 		/*
415 		 * We found a match.  Move these pages from the tmp list to
416 		 * the caller's list.
417 		 */
418 
419 		for (i = 0; i < num; i++) {
420 			pg = PHYS_TO_VM_PAGE(spa + ptoa(i));
421 			TAILQ_REMOVE(&tmp, pg, pageq.queue);
422 			pg->flags &= ~PG_PGLCA;
423 			TAILQ_INSERT_TAIL(rlist, pg, pageq.queue);
424 			STAT_INCR(uvm_pglistalloc_npages);
425 		}
426 
427 		error = 0;
428 		break;
429 	}
430 
431 	/*
432 	 * Free all the pages that we didn't need.
433 	 */
434 
435 	while (!TAILQ_EMPTY(&tmp)) {
436 		pg = TAILQ_FIRST(&tmp);
437 		TAILQ_REMOVE(&tmp, pg, pageq.queue);
438 		pg->flags &= ~PG_PGLCA;
439 		uvm_pagefree(pg);
440 	}
441 	mutex_exit(&uvm_pglistalloc_contig_lock);
442 	return error;
443 }
444 
445 static int
446 uvm_pglistalloc_contig(int num, paddr_t low, paddr_t high, paddr_t alignment,
447     paddr_t boundary, struct pglist *rlist, int waitok)
448 {
449 	int fl;
450 	int error;
451 	uvm_physseg_t psi;
452 
453 	/* Default to "lose". */
454 	error = ENOMEM;
455 
456 	/*
457 	 * Block all memory allocation and lock the free list.
458 	 */
459 	uvm_pgfl_lock();
460 
461 	/* Are there even any free pages? */
462 	if (uvm_availmem(false) <=
463 	    (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
464 		goto out;
465 
466 	for (fl = 0; fl < VM_NFREELIST; fl++) {
467 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
468 		for (psi = uvm_physseg_get_last(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_prev(psi))
469 #else
470 		for (psi = uvm_physseg_get_first(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_next(psi))
471 #endif
472 		{
473 			if (uvm_physseg_get_free_list(psi) != fl)
474 				continue;
475 
476 			num -= uvm_pglistalloc_c_ps(psi, num, low, high,
477 						    alignment, boundary, rlist);
478 			if (num == 0) {
479 #ifdef PGALLOC_VERBOSE
480 				printf("pgalloc: %"PRIxMAX"-%"PRIxMAX"\n",
481 				       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
482 				       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
483 #endif
484 				error = 0;
485 				goto out;
486 			}
487 		}
488 	}
489 
490 out:
491 	uvm_pgfl_unlock();
492 
493 	/*
494 	 * If that didn't work, try the more aggressive approach.
495 	 */
496 
497 	if (error) {
498 		if (waitok) {
499 			error = uvm_pglistalloc_contig_aggressive(num, low, high,
500 			    alignment, boundary, rlist);
501 		} else {
502 			uvm_pglistfree(rlist);
503 			uvm_kick_pdaemon();
504 		}
505 	}
506 	return error;
507 }
508 
509 static int
510 uvm_pglistalloc_s_ps(uvm_physseg_t psi, int num, paddr_t low, paddr_t high,
511     struct pglist *rlist)
512 {
513 	int todo, limit, candidate;
514 	struct vm_page *pg;
515 	bool second_pass;
516 #ifdef PGALLOC_VERBOSE
517 	printf("pgalloc: simple %d pgs from psi %d\n", num, psi);
518 #endif
519 
520 	KASSERT(uvm_physseg_get_start(psi) <= uvm_physseg_get_avail_start(psi));
521 	KASSERT(uvm_physseg_get_start(psi) <= uvm_physseg_get_avail_end(psi));
522 	KASSERT(uvm_physseg_get_avail_start(psi) <= uvm_physseg_get_end(psi));
523 	KASSERT(uvm_physseg_get_avail_end(psi) <= uvm_physseg_get_end(psi));
524 
525 	low = atop(low);
526 	high = atop(high);
527 	todo = num;
528 	candidate = uimax(low, uvm_physseg_get_avail_start(psi) +
529 	    uvm_physseg_get_start_hint(psi));
530 	limit = uimin(high, uvm_physseg_get_avail_end(psi));
531 	pg = uvm_physseg_get_pg(psi, candidate - uvm_physseg_get_start(psi));
532 	second_pass = false;
533 
534 	/*
535 	 * Make sure that physseg falls within with range to be allocated from.
536 	 */
537 	if (high <= uvm_physseg_get_avail_start(psi) ||
538 	    low >= uvm_physseg_get_avail_end(psi))
539 		return 0;
540 
541 again:
542 	for (;; candidate++, pg++) {
543 		if (candidate >= limit) {
544 			if (uvm_physseg_get_start_hint(psi) == 0 || second_pass) {
545 				candidate = limit - 1;
546 				break;
547 			}
548 			second_pass = true;
549 			candidate = uimax(low, uvm_physseg_get_avail_start(psi));
550 			limit = uimin(limit, uvm_physseg_get_avail_start(psi) +
551 			    uvm_physseg_get_start_hint(psi));
552 			pg = uvm_physseg_get_pg(psi, candidate - uvm_physseg_get_start(psi));
553 			goto again;
554 		}
555 #if defined(DEBUG)
556 		{
557 			paddr_t cidx = 0;
558 			const uvm_physseg_t bank = uvm_physseg_find(candidate, &cidx);
559 			KDASSERTMSG(bank == psi,
560 			    "uvm_physseg_find(%#x) (%"PRIxPHYSSEG ") != psi %"PRIxPHYSSEG,
561 			     candidate, bank, psi);
562 			KDASSERTMSG(cidx == candidate - uvm_physseg_get_start(psi),
563 			    "uvm_physseg_find(%#x): %#"PRIxPADDR" != off %"PRIxPADDR,
564 			     candidate, cidx, candidate - uvm_physseg_get_start(psi));
565 		}
566 #endif
567 		if (VM_PAGE_IS_FREE(pg) == 0)
568 			continue;
569 
570 		uvm_pglist_add(pg, rlist);
571 		if (--todo == 0) {
572 			break;
573 		}
574 	}
575 
576 	/*
577 	 * The next time we need to search this segment,
578 	 * start just after the pages we just allocated.
579 	 */
580 	uvm_physseg_set_start_hint(psi, candidate + 1 - uvm_physseg_get_avail_start(psi));
581 	KASSERTMSG(uvm_physseg_get_start_hint(psi) <= uvm_physseg_get_avail_end(psi) -
582 	    uvm_physseg_get_avail_start(psi),
583 	    "%#x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
584 	    candidate + 1,
585 	    uvm_physseg_get_start_hint(psi),
586 	    uvm_physseg_get_start_hint(psi),
587 	    uvm_physseg_get_avail_end(psi),
588 	    uvm_physseg_get_avail_start(psi),
589 	    uvm_physseg_get_avail_end(psi) - uvm_physseg_get_avail_start(psi));
590 
591 #ifdef PGALLOC_VERBOSE
592 	printf("got %d pgs\n", num - todo);
593 #endif
594 	return (num - todo); /* number of pages allocated */
595 }
596 
597 static int
598 uvm_pglistalloc_simple(int num, paddr_t low, paddr_t high,
599     struct pglist *rlist, int waitok)
600 {
601 	int fl, error;
602 	uvm_physseg_t psi;
603 	int count = 0;
604 
605 	/* Default to "lose". */
606 	error = ENOMEM;
607 
608 again:
609 	/*
610 	 * Block all memory allocation and lock the free list.
611 	 */
612 	uvm_pgfl_lock();
613 	count++;
614 
615 	/* Are there even any free pages? */
616 	if (uvm_availmem(false) <=
617 	    (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
618 		goto out;
619 
620 	for (fl = 0; fl < VM_NFREELIST; fl++) {
621 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
622 		for (psi = uvm_physseg_get_last(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_prev(psi))
623 #else
624 		for (psi = uvm_physseg_get_first(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_next(psi))
625 #endif
626 		{
627 			if (uvm_physseg_get_free_list(psi) != fl)
628 				continue;
629 
630 			num -= uvm_pglistalloc_s_ps(psi, num, low, high, rlist);
631 			if (num == 0) {
632 				error = 0;
633 				goto out;
634 			}
635 		}
636 
637 	}
638 
639 out:
640 	/*
641 	 * check to see if we need to generate some free pages waking
642 	 * the pagedaemon.
643 	 */
644 
645 	uvm_pgfl_unlock();
646 	uvm_kick_pdaemon();
647 
648 	if (error) {
649 		if (waitok) {
650 			uvm_wait("pglalloc");
651 			goto again;
652 		} else
653 			uvm_pglistfree(rlist);
654 	}
655 #ifdef PGALLOC_VERBOSE
656 	if (!error)
657 		printf("pgalloc: %"PRIxMAX"..%"PRIxMAX"\n",
658 		       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
659 		       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
660 #endif
661 	return (error);
662 }
663 
664 int
665 uvm_pglistalloc(psize_t size, paddr_t low, paddr_t high, paddr_t alignment,
666     paddr_t boundary, struct pglist *rlist, int nsegs, int waitok)
667 {
668 	int num, res;
669 
670 	KASSERT(!cpu_intr_p());
671 	KASSERT(!cpu_softintr_p());
672 	KASSERT((alignment & (alignment - 1)) == 0);
673 	KASSERT((boundary & (boundary - 1)) == 0);
674 
675 	/*
676 	 * Our allocations are always page granularity, so our alignment
677 	 * must be, too.
678 	 */
679 	if (alignment < PAGE_SIZE)
680 		alignment = PAGE_SIZE;
681 	if (boundary != 0 && boundary < size)
682 		return (EINVAL);
683 	num = atop(round_page(size));
684 	low = roundup2(low, alignment);
685 
686 	TAILQ_INIT(rlist);
687 
688 	/*
689 	 * Turn off the caching of free pages - we need everything to be on
690 	 * the global freelists.
691 	 */
692 	uvm_pgflcache_pause();
693 
694 	if (nsegs < num || alignment != PAGE_SIZE || boundary != 0)
695 		res = uvm_pglistalloc_contig(num, low, high, alignment,
696 					     boundary, rlist, waitok);
697 	else
698 		res = uvm_pglistalloc_simple(num, low, high, rlist, waitok);
699 
700 	uvm_pgflcache_resume();
701 
702 	return (res);
703 }
704 
705 /*
706  * uvm_pglistfree: free a list of pages
707  *
708  * => pages should already be unmapped
709  */
710 
711 void
712 uvm_pglistfree(struct pglist *list)
713 {
714 	struct vm_page *pg;
715 
716 	KASSERT(!cpu_intr_p());
717 	KASSERT(!cpu_softintr_p());
718 
719 	while ((pg = TAILQ_FIRST(list)) != NULL) {
720 		TAILQ_REMOVE(list, pg, pageq.queue);
721 		uvm_pagefree(pg);
722 		STAT_DECR(uvm_pglistalloc_npages);
723 	}
724 }
725 
726 void
727 uvm_pglistalloc_init(void)
728 {
729 
730 	mutex_init(&uvm_pglistalloc_contig_lock, MUTEX_DEFAULT, IPL_NONE);
731 }
732