xref: /netbsd-src/sys/uvm/uvm_pglist.c (revision cd22f25e6f6d1cc1f197fe8c5468a80f51d1c4e1)
1 /*	$NetBSD: uvm_pglist.c,v 1.40 2008/04/28 20:24:12 martin 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  *
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 1.40 2008/04/28 20:24:12 martin Exp $");
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/malloc.h>
43 #include <sys/proc.h>
44 
45 #include <uvm/uvm.h>
46 #include <uvm/uvm_pdpolicy.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 	int free_list, color, pgflidx;
85 #ifdef DEBUG
86 	struct vm_page *tp;
87 #endif
88 
89 	KASSERT(mutex_owned(&uvm_fpageqlock));
90 
91 #if PGFL_NQUEUES != 2
92 #error uvm_pglistalloc needs to be updated
93 #endif
94 
95 	free_list = uvm_page_lookup_freelist(pg);
96 	color = VM_PGCOLOR_BUCKET(pg);
97 	pgflidx = (pg->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN;
98 #ifdef DEBUG
99 	for (tp = TAILQ_FIRST(&uvm.page_free[
100 		free_list].pgfl_buckets[color].pgfl_queues[pgflidx]);
101 	     tp != NULL;
102 	     tp = TAILQ_NEXT(tp, pageq)) {
103 		if (tp == pg)
104 			break;
105 	}
106 	if (tp == NULL)
107 		panic("uvm_pglistalloc: page not on freelist");
108 #endif
109 	TAILQ_REMOVE(&uvm.page_free[free_list].pgfl_buckets[
110 			color].pgfl_queues[pgflidx], pg, pageq);
111 	uvmexp.free--;
112 	if (pg->flags & PG_ZERO)
113 		uvmexp.zeropages--;
114 	pg->flags = PG_CLEAN;
115 	pg->pqflags = 0;
116 	pg->uobject = NULL;
117 	pg->uanon = NULL;
118 	TAILQ_INSERT_TAIL(rlist, pg, pageq);
119 	STAT_INCR(uvm_pglistalloc_npages);
120 }
121 
122 static int
123 uvm_pglistalloc_c_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
124     paddr_t alignment, paddr_t boundary, struct pglist *rlist)
125 {
126 	int try, limit, tryidx, end, idx;
127 	struct vm_page *pgs;
128 	int pagemask;
129 #ifdef DEBUG
130 	paddr_t idxpa, lastidxpa;
131 	int cidx = 0;	/* XXX: GCC */
132 #endif
133 #ifdef PGALLOC_VERBOSE
134 	printf("pgalloc: contig %d pgs from psi %ld\n", num,
135 	(long)(ps - vm_physmem));
136 #endif
137 
138 	KASSERT(mutex_owned(&uvm_fpageqlock));
139 
140 	try = roundup(max(atop(low), ps->avail_start), atop(alignment));
141 	limit = min(atop(high), ps->avail_end);
142 	pagemask = ~((boundary >> PAGE_SHIFT) - 1);
143 
144 	for (;;) {
145 		if (try + num > limit) {
146 			/*
147 			 * We've run past the allowable range.
148 			 */
149 			return (0); /* FAIL */
150 		}
151 		if (boundary != 0 &&
152 		    ((try ^ (try + num - 1)) & pagemask) != 0) {
153 			/*
154 			 * Region crosses boundary. Jump to the boundary
155 			 * just crossed and ensure alignment.
156 			 */
157 			try = (try + num - 1) & pagemask;
158 			try = roundup(try, atop(alignment));
159 			continue;
160 		}
161 #ifdef DEBUG
162 		/*
163 		 * Make sure this is a managed physical page.
164 		 */
165 
166 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
167 			panic("pgalloc contig: botch1");
168 		if (cidx != try - ps->start)
169 			panic("pgalloc contig: botch2");
170 		if (vm_physseg_find(try + num - 1, &cidx) != ps - vm_physmem)
171 			panic("pgalloc contig: botch3");
172 		if (cidx != try - ps->start + num - 1)
173 			panic("pgalloc contig: botch4");
174 #endif
175 		tryidx = try - ps->start;
176 		end = tryidx + num;
177 		pgs = ps->pgs;
178 
179 		/*
180 		 * Found a suitable starting page.  See if the range is free.
181 		 */
182 		for (idx = tryidx; idx < end; idx++) {
183 			if (VM_PAGE_IS_FREE(&pgs[idx]) == 0)
184 				break;
185 
186 #ifdef DEBUG
187 			idxpa = VM_PAGE_TO_PHYS(&pgs[idx]);
188 			if (idx > tryidx) {
189 				lastidxpa = VM_PAGE_TO_PHYS(&pgs[idx - 1]);
190 				if ((lastidxpa + PAGE_SIZE) != idxpa) {
191 					/*
192 					 * Region not contiguous.
193 					 */
194 					panic("pgalloc contig: botch5");
195 				}
196 				if (boundary != 0 &&
197 				    ((lastidxpa ^ idxpa) & ~(boundary - 1))
198 				    != 0) {
199 					/*
200 					 * Region crosses boundary.
201 					 */
202 					panic("pgalloc contig: botch6");
203 				}
204 			}
205 #endif
206 		}
207 		if (idx == end)
208 			break;
209 
210 		try += atop(alignment);
211 	}
212 
213 	/*
214 	 * we have a chunk of memory that conforms to the requested constraints.
215 	 */
216 	idx = tryidx;
217 	while (idx < end)
218 		uvm_pglist_add(&pgs[idx++], rlist);
219 
220 #ifdef PGALLOC_VERBOSE
221 	printf("got %d pgs\n", num);
222 #endif
223 	return (num); /* number of pages allocated */
224 }
225 
226 static int
227 uvm_pglistalloc_contig(int num, paddr_t low, paddr_t high, paddr_t alignment,
228     paddr_t boundary, struct pglist *rlist)
229 {
230 	int fl, psi;
231 	struct vm_physseg *ps;
232 	int error;
233 
234 	/* Default to "lose". */
235 	error = ENOMEM;
236 
237 	/*
238 	 * Block all memory allocation and lock the free list.
239 	 */
240 	mutex_spin_enter(&uvm_fpageqlock);
241 
242 	/* Are there even any free pages? */
243 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
244 		goto out;
245 
246 	for (fl = 0; fl < VM_NFREELIST; fl++) {
247 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
248 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
249 #else
250 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
251 #endif
252 		{
253 			ps = &vm_physmem[psi];
254 
255 			if (ps->free_list != fl)
256 				continue;
257 
258 			num -= uvm_pglistalloc_c_ps(ps, num, low, high,
259 						    alignment, boundary, rlist);
260 			if (num == 0) {
261 #ifdef PGALLOC_VERBOSE
262 				printf("pgalloc: %lx-%lx\n",
263 				       VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
264 				       VM_PAGE_TO_PHYS(TAILQ_LAST(rlist)));
265 #endif
266 				error = 0;
267 				goto out;
268 			}
269 		}
270 	}
271 
272 out:
273 	/*
274 	 * check to see if we need to generate some free pages waking
275 	 * the pagedaemon.
276 	 */
277 
278 	uvm_kick_pdaemon();
279 	mutex_spin_exit(&uvm_fpageqlock);
280 	return (error);
281 }
282 
283 static int
284 uvm_pglistalloc_s_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
285     struct pglist *rlist)
286 {
287 	int todo, limit, try;
288 	struct vm_page *pg;
289 #ifdef DEBUG
290 	int cidx = 0;	/* XXX: GCC */
291 #endif
292 #ifdef PGALLOC_VERBOSE
293 	printf("pgalloc: simple %d pgs from psi %ld\n", num,
294 	    (long)(ps - vm_physmem));
295 #endif
296 
297 	KASSERT(mutex_owned(&uvm_fpageqlock));
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, 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 	mutex_spin_enter(&uvm_fpageqlock);
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 	mutex_spin_exit(&uvm_fpageqlock);
374 
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 
438 	/*
439 	 * Lock the free list and free each page.
440 	 */
441 
442 	mutex_spin_enter(&uvm_fpageqlock);
443 	while ((pg = TAILQ_FIRST(list)) != NULL) {
444 		bool iszero;
445 
446 		KASSERT(!uvmpdpol_pageisqueued_p(pg));
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 	mutex_spin_exit(&uvm_fpageqlock);
470 }
471