xref: /netbsd-src/sys/uvm/uvm_pglist.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: uvm_pglist.c,v 1.32 2004/09/17 20:46:03 yamt Exp $	*/
2 
3 /*-
4  * Copyright (c) 1997 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.
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  * 3. All advertising materials mentioning features or use of this software
20  *    must display the following acknowledgement:
21  *      This product includes software developed by the NetBSD
22  *      Foundation, Inc. and its contributors.
23  * 4. Neither the name of The NetBSD Foundation nor the names of its
24  *    contributors may be used to endorse or promote products derived
25  *    from this software without specific prior written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
28  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
29  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
30  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
31  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
32  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
33  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
34  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
35  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
36  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
37  * POSSIBILITY OF SUCH DAMAGE.
38  */
39 
40 /*
41  * uvm_pglist.c: pglist functions
42  */
43 
44 #include <sys/cdefs.h>
45 __KERNEL_RCSID(0, "$NetBSD: uvm_pglist.c,v 1.32 2004/09/17 20:46:03 yamt Exp $");
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/malloc.h>
50 #include <sys/proc.h>
51 
52 #include <uvm/uvm.h>
53 
54 #ifdef VM_PAGE_ALLOC_MEMORY_STATS
55 #define	STAT_INCR(v)	(v)++
56 #define	STAT_DECR(v)	do { \
57 		if ((v) == 0) \
58 			printf("%s:%d -- Already 0!\n", __FILE__, __LINE__); \
59 		else \
60 			(v)--; \
61 	} while (/*CONSTCOND*/ 0)
62 u_long	uvm_pglistalloc_npages;
63 #else
64 #define	STAT_INCR(v)
65 #define	STAT_DECR(v)
66 #endif
67 
68 /*
69  * uvm_pglistalloc: allocate a list of pages
70  *
71  * => allocated pages are placed onto an rlist.  rlist is
72  *    initialized by uvm_pglistalloc.
73  * => returns 0 on success or errno on failure
74  * => implementation allocates a single segment if any constraints are
75  *	imposed by call arguments.
76  * => doesn't take into account clean non-busy pages on inactive list
77  *	that could be used(?)
78  * => params:
79  *	size		the size of the allocation, rounded to page size.
80  *	low		the low address of the allowed allocation range.
81  *	high		the high address of the allowed allocation range.
82  *	alignment	memory must be aligned to this power-of-two boundary.
83  *	boundary	no segment in the allocation may cross this
84  *			power-of-two boundary (relative to zero).
85  */
86 
87 static void uvm_pglist_add(struct vm_page *, struct pglist *);
88 static int uvm_pglistalloc_c_ps(struct vm_physseg *, int, paddr_t, paddr_t,
89 				paddr_t, paddr_t, struct pglist *);
90 static int uvm_pglistalloc_contig(int, paddr_t, paddr_t, paddr_t, paddr_t,
91 				  struct pglist *);
92 static int uvm_pglistalloc_s_ps(struct vm_physseg *, int, paddr_t, paddr_t,
93 				struct pglist *);
94 static int uvm_pglistalloc_simple(int, paddr_t, paddr_t,
95 				  struct pglist *, int);
96 
97 static void
98 uvm_pglist_add(pg, rlist)
99 	struct vm_page *pg;
100 	struct pglist *rlist;
101 {
102 	int free_list, color, pgflidx;
103 #ifdef DEBUG
104 	struct vm_page *tp;
105 #endif
106 
107 #if PGFL_NQUEUES != 2
108 #error uvm_pglistalloc needs to be updated
109 #endif
110 
111 	free_list = uvm_page_lookup_freelist(pg);
112 	color = VM_PGCOLOR_BUCKET(pg);
113 	pgflidx = (pg->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN;
114 #ifdef DEBUG
115 	for (tp = TAILQ_FIRST(&uvm.page_free[
116 		free_list].pgfl_buckets[color].pgfl_queues[pgflidx]);
117 	     tp != NULL;
118 	     tp = TAILQ_NEXT(tp, pageq)) {
119 		if (tp == pg)
120 			break;
121 	}
122 	if (tp == NULL)
123 		panic("uvm_pglistalloc: page not on freelist");
124 #endif
125 	TAILQ_REMOVE(&uvm.page_free[free_list].pgfl_buckets[
126 			color].pgfl_queues[pgflidx], pg, pageq);
127 	uvmexp.free--;
128 	if (pg->flags & PG_ZERO)
129 		uvmexp.zeropages--;
130 	pg->flags = PG_CLEAN;
131 	pg->pqflags = 0;
132 	pg->uobject = NULL;
133 	pg->uanon = NULL;
134 	TAILQ_INSERT_TAIL(rlist, pg, pageq);
135 	STAT_INCR(uvm_pglistalloc_npages);
136 }
137 
138 static int
139 uvm_pglistalloc_c_ps(ps, num, low, high, alignment, boundary, rlist)
140 	struct vm_physseg *ps;
141 	int num;
142 	paddr_t low, high, alignment, boundary;
143 	struct pglist *rlist;
144 {
145 	int try, limit, tryidx, end, idx;
146 	struct vm_page *pgs;
147 	int pagemask;
148 #ifdef DEBUG
149 	paddr_t idxpa, lastidxpa;
150 	int cidx;
151 #endif
152 #ifdef PGALLOC_VERBOSE
153 	printf("pgalloc: contig %d pgs from psi %ld\n", num,
154 	(long)(ps - vm_physmem));
155 #endif
156 
157 	try = roundup(max(atop(low), ps->avail_start), atop(alignment));
158 	limit = min(atop(high), ps->avail_end);
159 	pagemask = ~((boundary >> PAGE_SHIFT) - 1);
160 
161 	for (;;) {
162 		if (try + num > limit) {
163 			/*
164 			 * We've run past the allowable range.
165 			 */
166 			return (0); /* FAIL */
167 		}
168 		if (boundary != 0 &&
169 		    ((try ^ (try + num - 1)) & pagemask) != 0) {
170 			/*
171 			 * Region crosses boundary. Jump to the boundary
172 			 * just crossed and ensure alignment.
173 			 */
174 			try = (try + num - 1) & pagemask;
175 			try = roundup(try, atop(alignment));
176 			continue;
177 		}
178 #ifdef DEBUG
179 		/*
180 		 * Make sure this is a managed physical page.
181 		 */
182 
183 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
184 			panic("pgalloc contig: botch1");
185 		if (cidx != try - ps->start)
186 			panic("pgalloc contig: botch2");
187 		if (vm_physseg_find(try + num - 1, &cidx) != ps - vm_physmem)
188 			panic("pgalloc contig: botch3");
189 		if (cidx != try - ps->start + num - 1)
190 			panic("pgalloc contig: botch4");
191 #endif
192 		tryidx = try - ps->start;
193 		end = tryidx + num;
194 		pgs = ps->pgs;
195 
196 		/*
197 		 * Found a suitable starting page.  See if the range is free.
198 		 */
199 		for (idx = tryidx; idx < end; idx++) {
200 			if (VM_PAGE_IS_FREE(&pgs[idx]) == 0)
201 				break;
202 
203 #ifdef DEBUG
204 			idxpa = VM_PAGE_TO_PHYS(&pgs[idx]);
205 			if (idx > tryidx) {
206 				lastidxpa = VM_PAGE_TO_PHYS(&pgs[idx - 1]);
207 				if ((lastidxpa + PAGE_SIZE) != idxpa) {
208 					/*
209 					 * Region not contiguous.
210 					 */
211 					panic("pgalloc contig: botch5");
212 				}
213 				if (boundary != 0 &&
214 				    ((lastidxpa ^ idxpa) & ~(boundary - 1))
215 				    != 0) {
216 					/*
217 					 * Region crosses boundary.
218 					 */
219 					panic("pgalloc contig: botch6");
220 				}
221 			}
222 #endif
223 		}
224 		if (idx == end)
225 			break;
226 
227 		try += atop(alignment);
228 	}
229 
230 	/*
231 	 * we have a chunk of memory that conforms to the requested constraints.
232 	 */
233 	idx = tryidx;
234 	while (idx < end)
235 		uvm_pglist_add(&pgs[idx++], rlist);
236 
237 #ifdef PGALLOC_VERBOSE
238 	printf("got %d pgs\n", num);
239 #endif
240 	return (num); /* number of pages allocated */
241 }
242 
243 static int
244 uvm_pglistalloc_contig(num, low, high, alignment, boundary, rlist)
245 	int num;
246 	paddr_t low, high, alignment, boundary;
247 	struct pglist *rlist;
248 {
249 	int fl, psi;
250 	struct vm_physseg *ps;
251 	int s, error;
252 
253 	/* Default to "lose". */
254 	error = ENOMEM;
255 
256 	/*
257 	 * Block all memory allocation and lock the free list.
258 	 */
259 	s = uvm_lock_fpageq();
260 
261 	/* Are there even any free pages? */
262 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
263 		goto out;
264 
265 	for (fl = 0; fl < VM_NFREELIST; fl++) {
266 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
267 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
268 #else
269 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
270 #endif
271 		{
272 			ps = &vm_physmem[psi];
273 
274 			if (ps->free_list != fl)
275 				continue;
276 
277 			num -= uvm_pglistalloc_c_ps(ps, num, low, high,
278 						    alignment, boundary, rlist);
279 			if (num == 0) {
280 #ifdef PGALLOC_VERBOSE
281 				printf("pgalloc: %lx-%lx\n",
282 				       VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
283 				       VM_PAGE_TO_PHYS(TAILQ_LAST(rlist)));
284 #endif
285 				error = 0;
286 				goto out;
287 			}
288 		}
289 	}
290 
291 out:
292 	/*
293 	 * check to see if we need to generate some free pages waking
294 	 * the pagedaemon.
295 	 */
296 
297 	UVM_KICK_PDAEMON();
298 	uvm_unlock_fpageq(s);
299 	return (error);
300 }
301 
302 static int
303 uvm_pglistalloc_s_ps(ps, num, low, high, rlist)
304 	struct vm_physseg *ps;
305 	int num;
306 	paddr_t low, high;
307 	struct pglist *rlist;
308 {
309 	int todo, limit, try;
310 	struct vm_page *pg;
311 #ifdef DEBUG
312 	int cidx;
313 #endif
314 #ifdef PGALLOC_VERBOSE
315 	printf("pgalloc: simple %d pgs from psi %ld\n", num,
316 	    (long)(ps - vm_physmem));
317 #endif
318 
319 	todo = num;
320 	limit = min(atop(high), ps->avail_end);
321 
322 	for (try = max(atop(low), ps->avail_start);
323 	     try < limit; try ++) {
324 #ifdef DEBUG
325 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
326 			panic("pgalloc simple: botch1");
327 		if (cidx != (try - ps->start))
328 			panic("pgalloc simple: botch2");
329 #endif
330 		pg = &ps->pgs[try - ps->start];
331 		if (VM_PAGE_IS_FREE(pg) == 0)
332 			continue;
333 
334 		uvm_pglist_add(pg, rlist);
335 		if (--todo == 0)
336 			break;
337 	}
338 
339 #ifdef PGALLOC_VERBOSE
340 	printf("got %d pgs\n", num - todo);
341 #endif
342 	return (num - todo); /* number of pages allocated */
343 }
344 
345 static int
346 uvm_pglistalloc_simple(num, low, high, rlist, waitok)
347 	int num;
348 	paddr_t low, high;
349 	struct pglist *rlist;
350 	int waitok;
351 {
352 	int fl, psi, s, error;
353 	struct vm_physseg *ps;
354 
355 	/* Default to "lose". */
356 	error = ENOMEM;
357 
358 again:
359 	/*
360 	 * Block all memory allocation and lock the free list.
361 	 */
362 	s = uvm_lock_fpageq();
363 
364 	/* Are there even any free pages? */
365 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
366 		goto out;
367 
368 	for (fl = 0; fl < VM_NFREELIST; fl++) {
369 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
370 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
371 #else
372 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
373 #endif
374 		{
375 			ps = &vm_physmem[psi];
376 
377 			if (ps->free_list != fl)
378 				continue;
379 
380 			num -= uvm_pglistalloc_s_ps(ps, num, low, high, rlist);
381 			if (num == 0) {
382 				error = 0;
383 				goto out;
384 			}
385 		}
386 
387 	}
388 
389 out:
390 	/*
391 	 * check to see if we need to generate some free pages waking
392 	 * the pagedaemon.
393 	 */
394 
395 	UVM_KICK_PDAEMON();
396 	uvm_unlock_fpageq(s);
397 	if (error) {
398 		if (waitok) {
399 			/* XXX perhaps some time limitation? */
400 #ifdef DEBUG
401 			printf("pglistalloc waiting\n");
402 #endif
403 			uvm_wait("pglalloc");
404 			goto again;
405 		} else
406 			uvm_pglistfree(rlist);
407 	}
408 #ifdef PGALLOC_VERBOSE
409 	if (!error)
410 		printf("pgalloc: %lx..%lx\n",
411 		       VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
412 		       VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
413 #endif
414 	return (error);
415 }
416 
417 int
418 uvm_pglistalloc(size, low, high, alignment, boundary, rlist, nsegs, waitok)
419 	psize_t size;
420 	paddr_t low, high, alignment, boundary;
421 	struct pglist *rlist;
422 	int nsegs, waitok;
423 {
424 	int num, res;
425 
426 	KASSERT((alignment & (alignment - 1)) == 0);
427 	KASSERT((boundary & (boundary - 1)) == 0);
428 
429 	/*
430 	 * Our allocations are always page granularity, so our alignment
431 	 * must be, too.
432 	 */
433 	if (alignment < PAGE_SIZE)
434 		alignment = PAGE_SIZE;
435 	if (boundary != 0 && boundary < size)
436 		return (EINVAL);
437 	num = atop(round_page(size));
438 	low = roundup(low, alignment);
439 
440 	TAILQ_INIT(rlist);
441 
442 	if ((nsegs < size >> PAGE_SHIFT) || (alignment != PAGE_SIZE) ||
443 	    (boundary != 0))
444 		res = uvm_pglistalloc_contig(num, low, high, alignment,
445 					     boundary, rlist);
446 	else
447 		res = uvm_pglistalloc_simple(num, low, high, rlist, waitok);
448 
449 	return (res);
450 }
451 
452 /*
453  * uvm_pglistfree: free a list of pages
454  *
455  * => pages should already be unmapped
456  */
457 
458 void
459 uvm_pglistfree(list)
460 	struct pglist *list;
461 {
462 	struct vm_page *pg;
463 	int s;
464 
465 	/*
466 	 * Lock the free list and free each page.
467 	 */
468 
469 	s = uvm_lock_fpageq();
470 	while ((pg = TAILQ_FIRST(list)) != NULL) {
471 		boolean_t iszero;
472 
473 		KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
474 		TAILQ_REMOVE(list, pg, pageq);
475 		iszero = (pg->flags & PG_ZERO);
476 		pg->pqflags = PQ_FREE;
477 #ifdef DEBUG
478 		pg->uobject = (void *)0xdeadbeef;
479 		pg->offset = 0xdeadbeef;
480 		pg->uanon = (void *)0xdeadbeef;
481 #endif /* DEBUG */
482 #ifdef DEBUG
483 		if (iszero)
484 			uvm_pagezerocheck(pg);
485 #endif /* DEBUG */
486 		TAILQ_INSERT_HEAD(&uvm.page_free[uvm_page_lookup_freelist(pg)].
487 		    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
488 		    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN], pg, pageq);
489 		uvmexp.free++;
490 		if (iszero)
491 			uvmexp.zeropages++;
492 		if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
493 			uvm.page_idle_zero = vm_page_zero_enable;
494 		STAT_DECR(uvm_pglistalloc_npages);
495 	}
496 	uvm_unlock_fpageq(s);
497 }
498