xref: /netbsd-src/sys/uvm/uvm_pglist.c (revision 1ad9454efb13a65cd7535ccf867508cb14d9d30e)
1 /*	$NetBSD: uvm_pglist.c,v 1.36 2006/09/15 15:51:13 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 1.36 2006/09/15 15:51:13 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 #include <uvm/uvm_pdpolicy.h>
54 
55 #ifdef VM_PAGE_ALLOC_MEMORY_STATS
56 #define	STAT_INCR(v)	(v)++
57 #define	STAT_DECR(v)	do { \
58 		if ((v) == 0) \
59 			printf("%s:%d -- Already 0!\n", __FILE__, __LINE__); \
60 		else \
61 			(v)--; \
62 	} while (/*CONSTCOND*/ 0)
63 u_long	uvm_pglistalloc_npages;
64 #else
65 #define	STAT_INCR(v)
66 #define	STAT_DECR(v)
67 #endif
68 
69 /*
70  * uvm_pglistalloc: allocate a list of pages
71  *
72  * => allocated pages are placed onto an rlist.  rlist is
73  *    initialized by uvm_pglistalloc.
74  * => returns 0 on success or errno on failure
75  * => implementation allocates a single segment if any constraints are
76  *	imposed by call arguments.
77  * => doesn't take into account clean non-busy pages on inactive list
78  *	that could be used(?)
79  * => params:
80  *	size		the size of the allocation, rounded to page size.
81  *	low		the low address of the allowed allocation range.
82  *	high		the high address of the allowed allocation range.
83  *	alignment	memory must be aligned to this power-of-two boundary.
84  *	boundary	no segment in the allocation may cross this
85  *			power-of-two boundary (relative to zero).
86  */
87 
88 static void
89 uvm_pglist_add(struct vm_page *pg, struct pglist *rlist)
90 {
91 	int free_list, color, pgflidx;
92 #ifdef DEBUG
93 	struct vm_page *tp;
94 #endif
95 
96 #if PGFL_NQUEUES != 2
97 #error uvm_pglistalloc needs to be updated
98 #endif
99 
100 	free_list = uvm_page_lookup_freelist(pg);
101 	color = VM_PGCOLOR_BUCKET(pg);
102 	pgflidx = (pg->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN;
103 #ifdef DEBUG
104 	for (tp = TAILQ_FIRST(&uvm.page_free[
105 		free_list].pgfl_buckets[color].pgfl_queues[pgflidx]);
106 	     tp != NULL;
107 	     tp = TAILQ_NEXT(tp, pageq)) {
108 		if (tp == pg)
109 			break;
110 	}
111 	if (tp == NULL)
112 		panic("uvm_pglistalloc: page not on freelist");
113 #endif
114 	TAILQ_REMOVE(&uvm.page_free[free_list].pgfl_buckets[
115 			color].pgfl_queues[pgflidx], pg, pageq);
116 	uvmexp.free--;
117 	if (pg->flags & PG_ZERO)
118 		uvmexp.zeropages--;
119 	pg->flags = PG_CLEAN;
120 	pg->pqflags = 0;
121 	pg->uobject = NULL;
122 	pg->uanon = NULL;
123 	TAILQ_INSERT_TAIL(rlist, pg, pageq);
124 	STAT_INCR(uvm_pglistalloc_npages);
125 }
126 
127 static int
128 uvm_pglistalloc_c_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
129     paddr_t alignment, paddr_t boundary, struct pglist *rlist)
130 {
131 	int try, limit, tryidx, end, idx;
132 	struct vm_page *pgs;
133 	int pagemask;
134 #ifdef DEBUG
135 	paddr_t idxpa, lastidxpa;
136 	int cidx = 0;	/* XXX: GCC */
137 #endif
138 #ifdef PGALLOC_VERBOSE
139 	printf("pgalloc: contig %d pgs from psi %ld\n", num,
140 	(long)(ps - vm_physmem));
141 #endif
142 
143 	try = roundup(max(atop(low), ps->avail_start), atop(alignment));
144 	limit = min(atop(high), ps->avail_end);
145 	pagemask = ~((boundary >> PAGE_SHIFT) - 1);
146 
147 	for (;;) {
148 		if (try + num > limit) {
149 			/*
150 			 * We've run past the allowable range.
151 			 */
152 			return (0); /* FAIL */
153 		}
154 		if (boundary != 0 &&
155 		    ((try ^ (try + num - 1)) & pagemask) != 0) {
156 			/*
157 			 * Region crosses boundary. Jump to the boundary
158 			 * just crossed and ensure alignment.
159 			 */
160 			try = (try + num - 1) & pagemask;
161 			try = roundup(try, atop(alignment));
162 			continue;
163 		}
164 #ifdef DEBUG
165 		/*
166 		 * Make sure this is a managed physical page.
167 		 */
168 
169 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
170 			panic("pgalloc contig: botch1");
171 		if (cidx != try - ps->start)
172 			panic("pgalloc contig: botch2");
173 		if (vm_physseg_find(try + num - 1, &cidx) != ps - vm_physmem)
174 			panic("pgalloc contig: botch3");
175 		if (cidx != try - ps->start + num - 1)
176 			panic("pgalloc contig: botch4");
177 #endif
178 		tryidx = try - ps->start;
179 		end = tryidx + num;
180 		pgs = ps->pgs;
181 
182 		/*
183 		 * Found a suitable starting page.  See if the range is free.
184 		 */
185 		for (idx = tryidx; idx < end; idx++) {
186 			if (VM_PAGE_IS_FREE(&pgs[idx]) == 0)
187 				break;
188 
189 #ifdef DEBUG
190 			idxpa = VM_PAGE_TO_PHYS(&pgs[idx]);
191 			if (idx > tryidx) {
192 				lastidxpa = VM_PAGE_TO_PHYS(&pgs[idx - 1]);
193 				if ((lastidxpa + PAGE_SIZE) != idxpa) {
194 					/*
195 					 * Region not contiguous.
196 					 */
197 					panic("pgalloc contig: botch5");
198 				}
199 				if (boundary != 0 &&
200 				    ((lastidxpa ^ idxpa) & ~(boundary - 1))
201 				    != 0) {
202 					/*
203 					 * Region crosses boundary.
204 					 */
205 					panic("pgalloc contig: botch6");
206 				}
207 			}
208 #endif
209 		}
210 		if (idx == end)
211 			break;
212 
213 		try += atop(alignment);
214 	}
215 
216 	/*
217 	 * we have a chunk of memory that conforms to the requested constraints.
218 	 */
219 	idx = tryidx;
220 	while (idx < end)
221 		uvm_pglist_add(&pgs[idx++], rlist);
222 
223 #ifdef PGALLOC_VERBOSE
224 	printf("got %d pgs\n", num);
225 #endif
226 	return (num); /* number of pages allocated */
227 }
228 
229 static int
230 uvm_pglistalloc_contig(int num, paddr_t low, paddr_t high, paddr_t alignment,
231     paddr_t boundary, struct pglist *rlist)
232 {
233 	int fl, psi;
234 	struct vm_physseg *ps;
235 	int s, error;
236 
237 	/* Default to "lose". */
238 	error = ENOMEM;
239 
240 	/*
241 	 * Block all memory allocation and lock the free list.
242 	 */
243 	s = uvm_lock_fpageq();
244 
245 	/* Are there even any free pages? */
246 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
247 		goto out;
248 
249 	for (fl = 0; fl < VM_NFREELIST; fl++) {
250 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
251 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
252 #else
253 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
254 #endif
255 		{
256 			ps = &vm_physmem[psi];
257 
258 			if (ps->free_list != fl)
259 				continue;
260 
261 			num -= uvm_pglistalloc_c_ps(ps, num, low, high,
262 						    alignment, boundary, rlist);
263 			if (num == 0) {
264 #ifdef PGALLOC_VERBOSE
265 				printf("pgalloc: %lx-%lx\n",
266 				       VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
267 				       VM_PAGE_TO_PHYS(TAILQ_LAST(rlist)));
268 #endif
269 				error = 0;
270 				goto out;
271 			}
272 		}
273 	}
274 
275 out:
276 	/*
277 	 * check to see if we need to generate some free pages waking
278 	 * the pagedaemon.
279 	 */
280 
281 	uvm_kick_pdaemon();
282 	uvm_unlock_fpageq(s);
283 	return (error);
284 }
285 
286 static int
287 uvm_pglistalloc_s_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
288     struct pglist *rlist)
289 {
290 	int todo, limit, try;
291 	struct vm_page *pg;
292 #ifdef DEBUG
293 	int cidx = 0;	/* XXX: GCC */
294 #endif
295 #ifdef PGALLOC_VERBOSE
296 	printf("pgalloc: simple %d pgs from psi %ld\n", num,
297 	    (long)(ps - vm_physmem));
298 #endif
299 
300 	todo = num;
301 	limit = min(atop(high), ps->avail_end);
302 
303 	for (try = max(atop(low), ps->avail_start);
304 	     try < limit; try ++) {
305 #ifdef DEBUG
306 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
307 			panic("pgalloc simple: botch1");
308 		if (cidx != (try - ps->start))
309 			panic("pgalloc simple: botch2");
310 #endif
311 		pg = &ps->pgs[try - ps->start];
312 		if (VM_PAGE_IS_FREE(pg) == 0)
313 			continue;
314 
315 		uvm_pglist_add(pg, rlist);
316 		if (--todo == 0)
317 			break;
318 	}
319 
320 #ifdef PGALLOC_VERBOSE
321 	printf("got %d pgs\n", num - todo);
322 #endif
323 	return (num - todo); /* number of pages allocated */
324 }
325 
326 static int
327 uvm_pglistalloc_simple(int num, paddr_t low, paddr_t high,
328     struct pglist *rlist, int waitok)
329 {
330 	int fl, psi, s, error;
331 	struct vm_physseg *ps;
332 
333 	/* Default to "lose". */
334 	error = ENOMEM;
335 
336 again:
337 	/*
338 	 * Block all memory allocation and lock the free list.
339 	 */
340 	s = uvm_lock_fpageq();
341 
342 	/* Are there even any free pages? */
343 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
344 		goto out;
345 
346 	for (fl = 0; fl < VM_NFREELIST; fl++) {
347 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
348 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
349 #else
350 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
351 #endif
352 		{
353 			ps = &vm_physmem[psi];
354 
355 			if (ps->free_list != fl)
356 				continue;
357 
358 			num -= uvm_pglistalloc_s_ps(ps, num, low, high, rlist);
359 			if (num == 0) {
360 				error = 0;
361 				goto out;
362 			}
363 		}
364 
365 	}
366 
367 out:
368 	/*
369 	 * check to see if we need to generate some free pages waking
370 	 * the pagedaemon.
371 	 */
372 
373 	uvm_kick_pdaemon();
374 	uvm_unlock_fpageq(s);
375 	if (error) {
376 		if (waitok) {
377 			/* XXX perhaps some time limitation? */
378 #ifdef DEBUG
379 			printf("pglistalloc waiting\n");
380 #endif
381 			uvm_wait("pglalloc");
382 			goto again;
383 		} else
384 			uvm_pglistfree(rlist);
385 	}
386 #ifdef PGALLOC_VERBOSE
387 	if (!error)
388 		printf("pgalloc: %lx..%lx\n",
389 		       VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
390 		       VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
391 #endif
392 	return (error);
393 }
394 
395 int
396 uvm_pglistalloc(psize_t size, paddr_t low, paddr_t high, paddr_t alignment,
397     paddr_t boundary, struct pglist *rlist, int nsegs, int waitok)
398 {
399 	int num, res;
400 
401 	KASSERT((alignment & (alignment - 1)) == 0);
402 	KASSERT((boundary & (boundary - 1)) == 0);
403 
404 	/*
405 	 * Our allocations are always page granularity, so our alignment
406 	 * must be, too.
407 	 */
408 	if (alignment < PAGE_SIZE)
409 		alignment = PAGE_SIZE;
410 	if (boundary != 0 && boundary < size)
411 		return (EINVAL);
412 	num = atop(round_page(size));
413 	low = roundup(low, alignment);
414 
415 	TAILQ_INIT(rlist);
416 
417 	if ((nsegs < size >> PAGE_SHIFT) || (alignment != PAGE_SIZE) ||
418 	    (boundary != 0))
419 		res = uvm_pglistalloc_contig(num, low, high, alignment,
420 					     boundary, rlist);
421 	else
422 		res = uvm_pglistalloc_simple(num, low, high, rlist, waitok);
423 
424 	return (res);
425 }
426 
427 /*
428  * uvm_pglistfree: free a list of pages
429  *
430  * => pages should already be unmapped
431  */
432 
433 void
434 uvm_pglistfree(struct pglist *list)
435 {
436 	struct vm_page *pg;
437 	int s;
438 
439 	/*
440 	 * Lock the free list and free each page.
441 	 */
442 
443 	s = uvm_lock_fpageq();
444 	while ((pg = TAILQ_FIRST(list)) != NULL) {
445 		boolean_t iszero;
446 
447 		KASSERT(!uvmpdpol_pageisqueued_p(pg));
448 		TAILQ_REMOVE(list, pg, pageq);
449 		iszero = (pg->flags & PG_ZERO);
450 		pg->pqflags = PQ_FREE;
451 #ifdef DEBUG
452 		pg->uobject = (void *)0xdeadbeef;
453 		pg->offset = 0xdeadbeef;
454 		pg->uanon = (void *)0xdeadbeef;
455 #endif /* DEBUG */
456 #ifdef DEBUG
457 		if (iszero)
458 			uvm_pagezerocheck(pg);
459 #endif /* DEBUG */
460 		TAILQ_INSERT_HEAD(&uvm.page_free[uvm_page_lookup_freelist(pg)].
461 		    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
462 		    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN], pg, pageq);
463 		uvmexp.free++;
464 		if (iszero)
465 			uvmexp.zeropages++;
466 		if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
467 			uvm.page_idle_zero = vm_page_zero_enable;
468 		STAT_DECR(uvm_pglistalloc_npages);
469 	}
470 	uvm_unlock_fpageq(s);
471 }
472