xref: /netbsd-src/sys/uvm/uvm_pglist.c (revision e5548b402ae4c44fb816de42c7bba9581ce23ef5)
1 /*	$NetBSD: uvm_pglist.c,v 1.34 2005/12/11 12:25:29 christos 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 1.34 2005/12/11 12:25:29 christos 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
88 uvm_pglist_add(struct vm_page *pg, struct pglist *rlist)
89 {
90 	int free_list, color, pgflidx;
91 #ifdef DEBUG
92 	struct vm_page *tp;
93 #endif
94 
95 #if PGFL_NQUEUES != 2
96 #error uvm_pglistalloc needs to be updated
97 #endif
98 
99 	free_list = uvm_page_lookup_freelist(pg);
100 	color = VM_PGCOLOR_BUCKET(pg);
101 	pgflidx = (pg->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN;
102 #ifdef DEBUG
103 	for (tp = TAILQ_FIRST(&uvm.page_free[
104 		free_list].pgfl_buckets[color].pgfl_queues[pgflidx]);
105 	     tp != NULL;
106 	     tp = TAILQ_NEXT(tp, pageq)) {
107 		if (tp == pg)
108 			break;
109 	}
110 	if (tp == NULL)
111 		panic("uvm_pglistalloc: page not on freelist");
112 #endif
113 	TAILQ_REMOVE(&uvm.page_free[free_list].pgfl_buckets[
114 			color].pgfl_queues[pgflidx], pg, pageq);
115 	uvmexp.free--;
116 	if (pg->flags & PG_ZERO)
117 		uvmexp.zeropages--;
118 	pg->flags = PG_CLEAN;
119 	pg->pqflags = 0;
120 	pg->uobject = NULL;
121 	pg->uanon = NULL;
122 	TAILQ_INSERT_TAIL(rlist, pg, pageq);
123 	STAT_INCR(uvm_pglistalloc_npages);
124 }
125 
126 static int
127 uvm_pglistalloc_c_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
128     paddr_t alignment, paddr_t boundary, struct pglist *rlist)
129 {
130 	int try, limit, tryidx, end, idx;
131 	struct vm_page *pgs;
132 	int pagemask;
133 #ifdef DEBUG
134 	paddr_t idxpa, lastidxpa;
135 	int cidx;
136 #endif
137 #ifdef PGALLOC_VERBOSE
138 	printf("pgalloc: contig %d pgs from psi %ld\n", num,
139 	(long)(ps - vm_physmem));
140 #endif
141 
142 	try = roundup(max(atop(low), ps->avail_start), atop(alignment));
143 	limit = min(atop(high), ps->avail_end);
144 	pagemask = ~((boundary >> PAGE_SHIFT) - 1);
145 
146 	for (;;) {
147 		if (try + num > limit) {
148 			/*
149 			 * We've run past the allowable range.
150 			 */
151 			return (0); /* FAIL */
152 		}
153 		if (boundary != 0 &&
154 		    ((try ^ (try + num - 1)) & pagemask) != 0) {
155 			/*
156 			 * Region crosses boundary. Jump to the boundary
157 			 * just crossed and ensure alignment.
158 			 */
159 			try = (try + num - 1) & pagemask;
160 			try = roundup(try, atop(alignment));
161 			continue;
162 		}
163 #ifdef DEBUG
164 		/*
165 		 * Make sure this is a managed physical page.
166 		 */
167 
168 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
169 			panic("pgalloc contig: botch1");
170 		if (cidx != try - ps->start)
171 			panic("pgalloc contig: botch2");
172 		if (vm_physseg_find(try + num - 1, &cidx) != ps - vm_physmem)
173 			panic("pgalloc contig: botch3");
174 		if (cidx != try - ps->start + num - 1)
175 			panic("pgalloc contig: botch4");
176 #endif
177 		tryidx = try - ps->start;
178 		end = tryidx + num;
179 		pgs = ps->pgs;
180 
181 		/*
182 		 * Found a suitable starting page.  See if the range is free.
183 		 */
184 		for (idx = tryidx; idx < end; idx++) {
185 			if (VM_PAGE_IS_FREE(&pgs[idx]) == 0)
186 				break;
187 
188 #ifdef DEBUG
189 			idxpa = VM_PAGE_TO_PHYS(&pgs[idx]);
190 			if (idx > tryidx) {
191 				lastidxpa = VM_PAGE_TO_PHYS(&pgs[idx - 1]);
192 				if ((lastidxpa + PAGE_SIZE) != idxpa) {
193 					/*
194 					 * Region not contiguous.
195 					 */
196 					panic("pgalloc contig: botch5");
197 				}
198 				if (boundary != 0 &&
199 				    ((lastidxpa ^ idxpa) & ~(boundary - 1))
200 				    != 0) {
201 					/*
202 					 * Region crosses boundary.
203 					 */
204 					panic("pgalloc contig: botch6");
205 				}
206 			}
207 #endif
208 		}
209 		if (idx == end)
210 			break;
211 
212 		try += atop(alignment);
213 	}
214 
215 	/*
216 	 * we have a chunk of memory that conforms to the requested constraints.
217 	 */
218 	idx = tryidx;
219 	while (idx < end)
220 		uvm_pglist_add(&pgs[idx++], rlist);
221 
222 #ifdef PGALLOC_VERBOSE
223 	printf("got %d pgs\n", num);
224 #endif
225 	return (num); /* number of pages allocated */
226 }
227 
228 static int
229 uvm_pglistalloc_contig(int num, paddr_t low, paddr_t high, paddr_t alignment,
230     paddr_t boundary, struct pglist *rlist)
231 {
232 	int fl, psi;
233 	struct vm_physseg *ps;
234 	int s, error;
235 
236 	/* Default to "lose". */
237 	error = ENOMEM;
238 
239 	/*
240 	 * Block all memory allocation and lock the free list.
241 	 */
242 	s = uvm_lock_fpageq();
243 
244 	/* Are there even any free pages? */
245 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
246 		goto out;
247 
248 	for (fl = 0; fl < VM_NFREELIST; fl++) {
249 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
250 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
251 #else
252 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
253 #endif
254 		{
255 			ps = &vm_physmem[psi];
256 
257 			if (ps->free_list != fl)
258 				continue;
259 
260 			num -= uvm_pglistalloc_c_ps(ps, num, low, high,
261 						    alignment, boundary, rlist);
262 			if (num == 0) {
263 #ifdef PGALLOC_VERBOSE
264 				printf("pgalloc: %lx-%lx\n",
265 				       VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
266 				       VM_PAGE_TO_PHYS(TAILQ_LAST(rlist)));
267 #endif
268 				error = 0;
269 				goto out;
270 			}
271 		}
272 	}
273 
274 out:
275 	/*
276 	 * check to see if we need to generate some free pages waking
277 	 * the pagedaemon.
278 	 */
279 
280 	UVM_KICK_PDAEMON();
281 	uvm_unlock_fpageq(s);
282 	return (error);
283 }
284 
285 static int
286 uvm_pglistalloc_s_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
287     struct pglist *rlist)
288 {
289 	int todo, limit, try;
290 	struct vm_page *pg;
291 #ifdef DEBUG
292 	int cidx;
293 #endif
294 #ifdef PGALLOC_VERBOSE
295 	printf("pgalloc: simple %d pgs from psi %ld\n", num,
296 	    (long)(ps - vm_physmem));
297 #endif
298 
299 	todo = num;
300 	limit = min(atop(high), ps->avail_end);
301 
302 	for (try = max(atop(low), ps->avail_start);
303 	     try < limit; try ++) {
304 #ifdef DEBUG
305 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
306 			panic("pgalloc simple: botch1");
307 		if (cidx != (try - ps->start))
308 			panic("pgalloc simple: botch2");
309 #endif
310 		pg = &ps->pgs[try - ps->start];
311 		if (VM_PAGE_IS_FREE(pg) == 0)
312 			continue;
313 
314 		uvm_pglist_add(pg, rlist);
315 		if (--todo == 0)
316 			break;
317 	}
318 
319 #ifdef PGALLOC_VERBOSE
320 	printf("got %d pgs\n", num - todo);
321 #endif
322 	return (num - todo); /* number of pages allocated */
323 }
324 
325 static int
326 uvm_pglistalloc_simple(int num, paddr_t low, paddr_t high,
327     struct pglist *rlist, int waitok)
328 {
329 	int fl, psi, s, error;
330 	struct vm_physseg *ps;
331 
332 	/* Default to "lose". */
333 	error = ENOMEM;
334 
335 again:
336 	/*
337 	 * Block all memory allocation and lock the free list.
338 	 */
339 	s = uvm_lock_fpageq();
340 
341 	/* Are there even any free pages? */
342 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
343 		goto out;
344 
345 	for (fl = 0; fl < VM_NFREELIST; fl++) {
346 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
347 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
348 #else
349 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
350 #endif
351 		{
352 			ps = &vm_physmem[psi];
353 
354 			if (ps->free_list != fl)
355 				continue;
356 
357 			num -= uvm_pglistalloc_s_ps(ps, num, low, high, rlist);
358 			if (num == 0) {
359 				error = 0;
360 				goto out;
361 			}
362 		}
363 
364 	}
365 
366 out:
367 	/*
368 	 * check to see if we need to generate some free pages waking
369 	 * the pagedaemon.
370 	 */
371 
372 	UVM_KICK_PDAEMON();
373 	uvm_unlock_fpageq(s);
374 	if (error) {
375 		if (waitok) {
376 			/* XXX perhaps some time limitation? */
377 #ifdef DEBUG
378 			printf("pglistalloc waiting\n");
379 #endif
380 			uvm_wait("pglalloc");
381 			goto again;
382 		} else
383 			uvm_pglistfree(rlist);
384 	}
385 #ifdef PGALLOC_VERBOSE
386 	if (!error)
387 		printf("pgalloc: %lx..%lx\n",
388 		       VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
389 		       VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
390 #endif
391 	return (error);
392 }
393 
394 int
395 uvm_pglistalloc(psize_t size, paddr_t low, paddr_t high, paddr_t alignment,
396     paddr_t boundary, struct pglist *rlist, int nsegs, int waitok)
397 {
398 	int num, res;
399 
400 	KASSERT((alignment & (alignment - 1)) == 0);
401 	KASSERT((boundary & (boundary - 1)) == 0);
402 
403 	/*
404 	 * Our allocations are always page granularity, so our alignment
405 	 * must be, too.
406 	 */
407 	if (alignment < PAGE_SIZE)
408 		alignment = PAGE_SIZE;
409 	if (boundary != 0 && boundary < size)
410 		return (EINVAL);
411 	num = atop(round_page(size));
412 	low = roundup(low, alignment);
413 
414 	TAILQ_INIT(rlist);
415 
416 	if ((nsegs < size >> PAGE_SHIFT) || (alignment != PAGE_SIZE) ||
417 	    (boundary != 0))
418 		res = uvm_pglistalloc_contig(num, low, high, alignment,
419 					     boundary, rlist);
420 	else
421 		res = uvm_pglistalloc_simple(num, low, high, rlist, waitok);
422 
423 	return (res);
424 }
425 
426 /*
427  * uvm_pglistfree: free a list of pages
428  *
429  * => pages should already be unmapped
430  */
431 
432 void
433 uvm_pglistfree(struct pglist *list)
434 {
435 	struct vm_page *pg;
436 	int s;
437 
438 	/*
439 	 * Lock the free list and free each page.
440 	 */
441 
442 	s = uvm_lock_fpageq();
443 	while ((pg = TAILQ_FIRST(list)) != NULL) {
444 		boolean_t iszero;
445 
446 		KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
447 		TAILQ_REMOVE(list, pg, pageq);
448 		iszero = (pg->flags & PG_ZERO);
449 		pg->pqflags = PQ_FREE;
450 #ifdef DEBUG
451 		pg->uobject = (void *)0xdeadbeef;
452 		pg->offset = 0xdeadbeef;
453 		pg->uanon = (void *)0xdeadbeef;
454 #endif /* DEBUG */
455 #ifdef DEBUG
456 		if (iszero)
457 			uvm_pagezerocheck(pg);
458 #endif /* DEBUG */
459 		TAILQ_INSERT_HEAD(&uvm.page_free[uvm_page_lookup_freelist(pg)].
460 		    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
461 		    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN], pg, pageq);
462 		uvmexp.free++;
463 		if (iszero)
464 			uvmexp.zeropages++;
465 		if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
466 			uvm.page_idle_zero = vm_page_zero_enable;
467 		STAT_DECR(uvm_pglistalloc_npages);
468 	}
469 	uvm_unlock_fpageq(s);
470 }
471