xref: /netbsd-src/sys/uvm/uvm_pglist.c (revision e6c7e151de239c49d2e38720a061ed9d1fa99309)
1 /*	$NetBSD: uvm_pglist.c,v 1.81 2020/03/01 21:43:56 ad 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 1.81 2020/03/01 21:43:56 ad 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 /*
63  * uvm_pglistalloc: allocate a list of pages
64  *
65  * => allocated pages are placed onto an rlist.  rlist is
66  *    initialized by uvm_pglistalloc.
67  * => returns 0 on success or errno on failure
68  * => implementation allocates a single segment if any constraints are
69  *	imposed by call arguments.
70  * => doesn't take into account clean non-busy pages on inactive list
71  *	that could be used(?)
72  * => params:
73  *	size		the size of the allocation, rounded to page size.
74  *	low		the low address of the allowed allocation range.
75  *	high		the high address of the allowed allocation range.
76  *	alignment	memory must be aligned to this power-of-two boundary.
77  *	boundary	no segment in the allocation may cross this
78  *			power-of-two boundary (relative to zero).
79  */
80 
81 static void
82 uvm_pglist_add(struct vm_page *pg, struct pglist *rlist)
83 {
84 	struct pgfreelist *pgfl;
85 	struct pgflbucket *pgb;
86 
87 	pgfl = &uvm.page_free[uvm_page_get_freelist(pg)];
88 	pgb = pgfl->pgfl_buckets[uvm_page_get_bucket(pg)];
89 
90 #ifdef UVMDEBUG
91 	struct vm_page *tp;
92 	LIST_FOREACH(tp, &pgb->pgb_colors[VM_PGCOLOR(pg)], pageq.list) {
93 		if (tp == pg)
94 			break;
95 	}
96 	if (tp == NULL)
97 		panic("uvm_pglistalloc: page not on freelist");
98 #endif
99 	LIST_REMOVE(pg, pageq.list);
100 	pgb->pgb_nfree--;
101 	if (pg->flags & PG_ZERO)
102 		CPU_COUNT(CPU_COUNT_ZEROPAGES, -1);
103 	pg->flags = PG_CLEAN;
104 	pg->uobject = NULL;
105 	pg->uanon = NULL;
106 	TAILQ_INSERT_TAIL(rlist, pg, pageq.queue);
107 	STAT_INCR(uvm_pglistalloc_npages);
108 }
109 
110 static int
111 uvm_pglistalloc_c_ps(uvm_physseg_t psi, int num, paddr_t low, paddr_t high,
112     paddr_t alignment, paddr_t boundary, struct pglist *rlist)
113 {
114 	signed int candidate, limit, candidateidx, end, idx, skip;
115 	int pagemask;
116 	bool second_pass;
117 #ifdef DEBUG
118 	paddr_t idxpa, lastidxpa;
119 	paddr_t cidx = 0;	/* XXX: GCC */
120 #endif
121 #ifdef PGALLOC_VERBOSE
122 	printf("pgalloc: contig %d pgs from psi %d\n", num, psi);
123 #endif
124 
125 	low = atop(low);
126 	high = atop(high);
127 	alignment = atop(alignment);
128 
129 	/*
130 	 * Make sure that physseg falls within with range to be allocated from.
131 	 */
132 	if (high <= uvm_physseg_get_avail_start(psi) || low >= uvm_physseg_get_avail_end(psi))
133 		return 0;
134 
135 	/*
136 	 * We start our search at the just after where the last allocation
137 	 * succeeded.
138 	 */
139 	candidate = roundup2(uimax(low, uvm_physseg_get_avail_start(psi) +
140 		uvm_physseg_get_start_hint(psi)), alignment);
141 	limit = uimin(high, uvm_physseg_get_avail_end(psi));
142 	pagemask = ~((boundary >> PAGE_SHIFT) - 1);
143 	skip = 0;
144 	second_pass = false;
145 
146 	for (;;) {
147 		bool ok = true;
148 		signed int cnt;
149 
150 		if (candidate + num > limit) {
151 			if (uvm_physseg_get_start_hint(psi) == 0 || second_pass) {
152 				/*
153 				 * We've run past the allowable range.
154 				 */
155 				return 0; /* FAIL = 0 pages*/
156 			}
157 			/*
158 			 * We've wrapped around the end of this segment
159 			 * so restart at the beginning but now our limit
160 			 * is were we started.
161 			 */
162 			second_pass = true;
163 			candidate = roundup2(uimax(low, uvm_physseg_get_avail_start(psi)), alignment);
164 			limit = uimin(limit, uvm_physseg_get_avail_start(psi) +
165 			    uvm_physseg_get_start_hint(psi));
166 			skip = 0;
167 			continue;
168 		}
169 		if (boundary != 0 &&
170 		    ((candidate ^ (candidate + num - 1)) & pagemask) != 0) {
171 			/*
172 			 * Region crosses boundary. Jump to the boundary
173 			 * just crossed and ensure alignment.
174 			 */
175 			candidate = (candidate + num - 1) & pagemask;
176 			candidate = roundup2(candidate, alignment);
177 			skip = 0;
178 			continue;
179 		}
180 #ifdef DEBUG
181 		/*
182 		 * Make sure this is a managed physical page.
183 		 */
184 
185 		if (uvm_physseg_find(candidate, &cidx) != psi)
186 			panic("pgalloc contig: botch1");
187 		if (cidx != candidate - uvm_physseg_get_start(psi))
188 			panic("pgalloc contig: botch2");
189 		if (uvm_physseg_find(candidate + num - 1, &cidx) != psi)
190 			panic("pgalloc contig: botch3");
191 		if (cidx != candidate - uvm_physseg_get_start(psi) + num - 1)
192 			panic("pgalloc contig: botch4");
193 #endif
194 		candidateidx = candidate - uvm_physseg_get_start(psi);
195 		end = candidateidx + num;
196 
197 		/*
198 		 * Found a suitable starting page.  See if the range is free.
199 		 */
200 #ifdef PGALLOC_VERBOSE
201 		printf("%s: psi=%d candidate=%#x end=%#x skip=%#x, align=%#"PRIxPADDR,
202 		    __func__, psi, candidateidx, end, skip, alignment);
203 #endif
204 		/*
205 		 * We start at the end and work backwards since if we find a
206 		 * non-free page, it makes no sense to continue.
207 		 *
208 		 * But on the plus size we have "vetted" some number of free
209 		 * pages.  If this iteration fails, we may be able to skip
210 		 * testing most of those pages again in the next pass.
211 		 */
212 		for (idx = end - 1; idx >= candidateidx + skip; idx--) {
213 			if (VM_PAGE_IS_FREE(uvm_physseg_get_pg(psi, idx)) == 0) {
214 				ok = false;
215 				break;
216 			}
217 
218 #ifdef DEBUG
219 			if (idx > candidateidx) {
220 				idxpa = VM_PAGE_TO_PHYS(uvm_physseg_get_pg(psi, idx));
221 				lastidxpa = VM_PAGE_TO_PHYS(uvm_physseg_get_pg(psi, idx - 1));
222 				if ((lastidxpa + PAGE_SIZE) != idxpa) {
223 					/*
224 					 * Region not contiguous.
225 					 */
226 					panic("pgalloc contig: botch5");
227 				}
228 				if (boundary != 0 &&
229 				    ((lastidxpa ^ idxpa) & ~(boundary - 1))
230 				    != 0) {
231 					/*
232 					 * Region crosses boundary.
233 					 */
234 					panic("pgalloc contig: botch6");
235 				}
236 			}
237 #endif
238 		}
239 
240 		if (ok) {
241 			while (skip-- > 0) {
242 				KDASSERT(VM_PAGE_IS_FREE(uvm_physseg_get_pg(psi, candidateidx + skip)));
243 			}
244 #ifdef PGALLOC_VERBOSE
245 			printf(": ok\n");
246 #endif
247 			break;
248 		}
249 
250 #ifdef PGALLOC_VERBOSE
251 		printf(": non-free at %#x\n", idx - candidateidx);
252 #endif
253 		/*
254 		 * count the number of pages we can advance
255 		 * since we know they aren't all free.
256 		 */
257 		cnt = idx + 1 - candidateidx;
258 		/*
259 		 * now round up that to the needed alignment.
260 		 */
261 		cnt = roundup2(cnt, alignment);
262 		/*
263 		 * The number of pages we can skip checking
264 		 * (might be 0 if cnt > num).
265 		 */
266 		skip = uimax(num - cnt, 0);
267 		candidate += cnt;
268 	}
269 
270 	/*
271 	 * we have a chunk of memory that conforms to the requested constraints.
272 	 */
273 	for (idx = candidateidx; idx < end; idx++)
274 		uvm_pglist_add(uvm_physseg_get_pg(psi, idx), rlist);
275 
276 	/*
277 	 * the next time we need to search this segment, start after this
278 	 * chunk of pages we just allocated.
279 	 */
280 	uvm_physseg_set_start_hint(psi, candidate + num -
281 	    uvm_physseg_get_avail_start(psi));
282 	KASSERTMSG(uvm_physseg_get_start_hint(psi) <=
283 	    uvm_physseg_get_avail_end(psi) - uvm_physseg_get_avail_start(psi),
284 	    "%x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
285 	    candidate + num,
286 	    uvm_physseg_get_start_hint(psi), uvm_physseg_get_start_hint(psi),
287 	    uvm_physseg_get_avail_end(psi), uvm_physseg_get_avail_start(psi),
288 	    uvm_physseg_get_avail_end(psi) - uvm_physseg_get_avail_start(psi));
289 
290 #ifdef PGALLOC_VERBOSE
291 	printf("got %d pgs\n", num);
292 #endif
293 	return num; /* number of pages allocated */
294 }
295 
296 static int
297 uvm_pglistalloc_contig(int num, paddr_t low, paddr_t high, paddr_t alignment,
298     paddr_t boundary, struct pglist *rlist)
299 {
300 	int fl;
301 	int error;
302 
303 	uvm_physseg_t psi;
304 	/* Default to "lose". */
305 	error = ENOMEM;
306 
307 	/*
308 	 * Block all memory allocation and lock the free list.
309 	 */
310 	uvm_pgfl_lock();
311 
312 	/* Are there even any free pages? */
313 	if (uvm_availmem() <=
314 	    (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
315 		goto out;
316 
317 	for (fl = 0; fl < VM_NFREELIST; fl++) {
318 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
319 		for (psi = uvm_physseg_get_last(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_prev(psi))
320 #else
321 		for (psi = uvm_physseg_get_first(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_next(psi))
322 #endif
323 		{
324 			if (uvm_physseg_get_free_list(psi) != fl)
325 				continue;
326 
327 			num -= uvm_pglistalloc_c_ps(psi, num, low, high,
328 						    alignment, boundary, rlist);
329 			if (num == 0) {
330 #ifdef PGALLOC_VERBOSE
331 				printf("pgalloc: %"PRIxMAX"-%"PRIxMAX"\n",
332 				       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
333 				       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
334 #endif
335 				error = 0;
336 				goto out;
337 			}
338 		}
339 	}
340 
341 out:
342 	/*
343 	 * check to see if we need to generate some free pages waking
344 	 * the pagedaemon.
345 	 */
346 
347 	uvm_pgfl_unlock();
348 	uvm_kick_pdaemon();
349 	return (error);
350 }
351 
352 static int
353 uvm_pglistalloc_s_ps(uvm_physseg_t psi, int num, paddr_t low, paddr_t high,
354     struct pglist *rlist)
355 {
356 	int todo, limit, candidate;
357 	struct vm_page *pg;
358 	bool second_pass;
359 #ifdef PGALLOC_VERBOSE
360 	printf("pgalloc: simple %d pgs from psi %zd\n", num, psi);
361 #endif
362 
363 	KASSERT(uvm_physseg_get_start(psi) <= uvm_physseg_get_avail_start(psi));
364 	KASSERT(uvm_physseg_get_start(psi) <= uvm_physseg_get_avail_end(psi));
365 	KASSERT(uvm_physseg_get_avail_start(psi) <= uvm_physseg_get_end(psi));
366 	KASSERT(uvm_physseg_get_avail_end(psi) <= uvm_physseg_get_end(psi));
367 
368 	low = atop(low);
369 	high = atop(high);
370 	todo = num;
371 	candidate = uimax(low, uvm_physseg_get_avail_start(psi) +
372 	    uvm_physseg_get_start_hint(psi));
373 	limit = uimin(high, uvm_physseg_get_avail_end(psi));
374 	pg = uvm_physseg_get_pg(psi, candidate - uvm_physseg_get_start(psi));
375 	second_pass = false;
376 
377 	/*
378 	 * Make sure that physseg falls within with range to be allocated from.
379 	 */
380 	if (high <= uvm_physseg_get_avail_start(psi) ||
381 	    low >= uvm_physseg_get_avail_end(psi))
382 		return 0;
383 
384 again:
385 	for (;; candidate++, pg++) {
386 		if (candidate >= limit) {
387 			if (uvm_physseg_get_start_hint(psi) == 0 || second_pass) {
388 				candidate = limit - 1;
389 				break;
390 			}
391 			second_pass = true;
392 			candidate = uimax(low, uvm_physseg_get_avail_start(psi));
393 			limit = uimin(limit, uvm_physseg_get_avail_start(psi) +
394 			    uvm_physseg_get_start_hint(psi));
395 			pg = uvm_physseg_get_pg(psi, candidate - uvm_physseg_get_start(psi));
396 			goto again;
397 		}
398 #if defined(DEBUG)
399 		{
400 			paddr_t cidx = 0;
401 			const uvm_physseg_t bank = uvm_physseg_find(candidate, &cidx);
402 			KDASSERTMSG(bank == psi,
403 			    "uvm_physseg_find(%#x) (%"PRIxPHYSSEG ") != psi %"PRIxPHYSSEG,
404 			     candidate, bank, psi);
405 			KDASSERTMSG(cidx == candidate - uvm_physseg_get_start(psi),
406 			    "uvm_physseg_find(%#x): %#"PRIxPADDR" != off %"PRIxPADDR,
407 			     candidate, cidx, candidate - uvm_physseg_get_start(psi));
408 		}
409 #endif
410 		if (VM_PAGE_IS_FREE(pg) == 0)
411 			continue;
412 
413 		uvm_pglist_add(pg, rlist);
414 		if (--todo == 0) {
415 			break;
416 		}
417 	}
418 
419 	/*
420 	 * The next time we need to search this segment,
421 	 * start just after the pages we just allocated.
422 	 */
423 	uvm_physseg_set_start_hint(psi, candidate + 1 - uvm_physseg_get_avail_start(psi));
424 	KASSERTMSG(uvm_physseg_get_start_hint(psi) <= uvm_physseg_get_avail_end(psi) -
425 	    uvm_physseg_get_avail_start(psi),
426 	    "%#x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
427 	    candidate + 1,
428 	    uvm_physseg_get_start_hint(psi),
429 	    uvm_physseg_get_start_hint(psi),
430 	    uvm_physseg_get_avail_end(psi),
431 	    uvm_physseg_get_avail_start(psi),
432 	    uvm_physseg_get_avail_end(psi) - uvm_physseg_get_avail_start(psi));
433 
434 #ifdef PGALLOC_VERBOSE
435 	printf("got %d pgs\n", num - todo);
436 #endif
437 	return (num - todo); /* number of pages allocated */
438 }
439 
440 static int
441 uvm_pglistalloc_simple(int num, paddr_t low, paddr_t high,
442     struct pglist *rlist, int waitok)
443 {
444 	int fl, error;
445 	uvm_physseg_t psi;
446 	int count = 0;
447 
448 	/* Default to "lose". */
449 	error = ENOMEM;
450 
451 again:
452 	/*
453 	 * Block all memory allocation and lock the free list.
454 	 */
455 	uvm_pgfl_lock();
456 	count++;
457 
458 	/* Are there even any free pages? */
459 	if (uvm_availmem() <=
460 	    (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
461 		goto out;
462 
463 	for (fl = 0; fl < VM_NFREELIST; fl++) {
464 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
465 		for (psi = uvm_physseg_get_last(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_prev(psi))
466 #else
467 		for (psi = uvm_physseg_get_first(); uvm_physseg_valid_p(psi); psi = uvm_physseg_get_next(psi))
468 #endif
469 		{
470 			if (uvm_physseg_get_free_list(psi) != fl)
471 				continue;
472 
473 			num -= uvm_pglistalloc_s_ps(psi, num, low, high, rlist);
474 			if (num == 0) {
475 				error = 0;
476 				goto out;
477 			}
478 		}
479 
480 	}
481 
482 out:
483 	/*
484 	 * check to see if we need to generate some free pages waking
485 	 * the pagedaemon.
486 	 */
487 
488 	uvm_pgfl_unlock();
489 	uvm_kick_pdaemon();
490 
491 	if (error) {
492 		if (waitok) {
493 			/* XXX perhaps some time limitation? */
494 #ifdef DEBUG
495 			if (count == 1)
496 				printf("pglistalloc waiting\n");
497 #endif
498 			uvm_wait("pglalloc");
499 			goto again;
500 		} else
501 			uvm_pglistfree(rlist);
502 	}
503 #ifdef PGALLOC_VERBOSE
504 	if (!error)
505 		printf("pgalloc: %"PRIxMAX"..%"PRIxMAX"\n",
506 		       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
507 		       (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
508 #endif
509 	return (error);
510 }
511 
512 int
513 uvm_pglistalloc(psize_t size, paddr_t low, paddr_t high, paddr_t alignment,
514     paddr_t boundary, struct pglist *rlist, int nsegs, int waitok)
515 {
516 	int num, res;
517 
518 	KASSERT(!cpu_intr_p());
519 	KASSERT(!cpu_softintr_p());
520 	KASSERT((alignment & (alignment - 1)) == 0);
521 	KASSERT((boundary & (boundary - 1)) == 0);
522 
523 	/*
524 	 * Our allocations are always page granularity, so our alignment
525 	 * must be, too.
526 	 */
527 	if (alignment < PAGE_SIZE)
528 		alignment = PAGE_SIZE;
529 	if (boundary != 0 && boundary < size)
530 		return (EINVAL);
531 	num = atop(round_page(size));
532 	low = roundup2(low, alignment);
533 
534 	TAILQ_INIT(rlist);
535 
536 	/*
537 	 * Turn off the caching of free pages - we need everything to be on
538 	 * the global freelists.
539 	 */
540 	uvm_pgflcache_pause();
541 
542 	if ((nsegs < size >> PAGE_SHIFT) || (alignment != PAGE_SIZE) ||
543 	    (boundary != 0))
544 		res = uvm_pglistalloc_contig(num, low, high, alignment,
545 					     boundary, rlist);
546 	else
547 		res = uvm_pglistalloc_simple(num, low, high, rlist, waitok);
548 
549 	uvm_pgflcache_resume();
550 
551 	return (res);
552 }
553 
554 /*
555  * uvm_pglistfree: free a list of pages
556  *
557  * => pages should already be unmapped
558  */
559 
560 void
561 uvm_pglistfree(struct pglist *list)
562 {
563 	struct pgfreelist *pgfl;
564 	struct pgflbucket *pgb;
565 	struct vm_page *pg;
566 	int c, b;
567 
568 	KASSERT(!cpu_intr_p());
569 	KASSERT(!cpu_softintr_p());
570 
571 	/*
572 	 * Lock the free list and free each page.
573 	 */
574 
575 	uvm_pgfl_lock();
576 	while ((pg = TAILQ_FIRST(list)) != NULL) {
577 		TAILQ_REMOVE(list, pg, pageq.queue);
578 		pg->flags = (pg->flags & PG_ZERO) | PG_FREE;
579 #ifdef DEBUG
580 		pg->uobject = (void *)0xdeadbeef;
581 		pg->uanon = (void *)0xdeadbeef;
582 		if (pg->flags & PG_ZERO)
583 			uvm_pagezerocheck(pg);
584 #endif /* DEBUG */
585 		c = VM_PGCOLOR(pg);
586 		b = uvm_page_get_bucket(pg);
587 		pgfl = &uvm.page_free[uvm_page_get_freelist(pg)];
588 		pgb = pgfl->pgfl_buckets[b];
589 		if (pg->flags & PG_ZERO)
590 			CPU_COUNT(CPU_COUNT_ZEROPAGES, 1);
591 		pgb->pgb_nfree++;
592 		LIST_INSERT_HEAD(&pgb->pgb_colors[c], pg, pageq.list);
593 		STAT_DECR(uvm_pglistalloc_npages);
594 	}
595 	uvm_pgfl_unlock();
596 }
597