uvm_pglist.c revision 1.52.2.1 1 /* $NetBSD: uvm_pglist.c,v 1.52.2.1 2011/02/08 16:20:07 bouyer 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 1.52.2.1 2011/02/08 16:20:07 bouyer 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
86 KASSERT(mutex_owned(&uvm_fpageqlock));
87
88 #if PGFL_NQUEUES != 2
89 #error uvm_pglistalloc needs to be updated
90 #endif
91
92 free_list = uvm_page_lookup_freelist(pg);
93 color = VM_PGCOLOR_BUCKET(pg);
94 pgflidx = (pg->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN;
95 #ifdef NOT_DEBUG
96 struct vm_page *tp;
97 LIST_FOREACH(tp,
98 &uvm.page_free[free_list].pgfl_buckets[color].pgfl_queues[pgflidx],
99 pageq.list) {
100 if (tp == pg)
101 break;
102 }
103 if (tp == NULL)
104 panic("uvm_pglistalloc: page not on freelist");
105 #endif
106 LIST_REMOVE(pg, pageq.list); /* global */
107 LIST_REMOVE(pg, listq.list); /* cpu */
108 uvmexp.free--;
109 if (pg->flags & PG_ZERO)
110 uvmexp.zeropages--;
111 VM_FREE_PAGE_TO_CPU(pg)->pages[pgflidx]--;
112 pg->flags = PG_CLEAN;
113 pg->pqflags = 0;
114 pg->uobject = NULL;
115 pg->uanon = NULL;
116 TAILQ_INSERT_TAIL(rlist, pg, pageq.queue);
117 STAT_INCR(uvm_pglistalloc_npages);
118 }
119
120 static int
121 uvm_pglistalloc_c_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
122 paddr_t alignment, paddr_t boundary, struct pglist *rlist)
123 {
124 signed int try, limit, tryidx, end, idx, skip;
125 struct vm_page *pgs;
126 int pagemask;
127 bool second_pass;
128 #ifdef DEBUG
129 paddr_t idxpa, lastidxpa;
130 int cidx = 0; /* XXX: GCC */
131 #endif
132 #ifdef PGALLOC_VERBOSE
133 printf("pgalloc: contig %d pgs from psi %zd\n", num, ps - vm_physmem);
134 #endif
135
136 KASSERT(mutex_owned(&uvm_fpageqlock));
137
138 low = atop(low);
139 high = atop(high);
140 alignment = atop(alignment);
141
142 /*
143 * Make sure that physseg falls within with range to be allocated from.
144 */
145 if (high <= ps->avail_start || low >= ps->avail_end)
146 return 0;
147
148 /*
149 * We start our search at the just after where the last allocation
150 * succeeded.
151 */
152 try = roundup2(max(low, ps->avail_start + ps->start_hint), alignment);
153 limit = min(high, ps->avail_end);
154 pagemask = ~((boundary >> PAGE_SHIFT) - 1);
155 skip = 0;
156 second_pass = false;
157 pgs = ps->pgs;
158
159 for (;;) {
160 bool ok = true;
161 signed int cnt;
162
163 if (try + num > limit) {
164 if (ps->start_hint == 0 || second_pass) {
165 /*
166 * We've run past the allowable range.
167 */
168 return 0; /* FAIL = 0 pages*/
169 }
170 /*
171 * We've wrapped around the end of this segment
172 * so restart at the beginning but now our limit
173 * is were we started.
174 */
175 second_pass = true;
176 try = roundup2(max(low, ps->avail_start), alignment);
177 limit = min(limit, ps->avail_start + ps->start_hint);
178 skip = 0;
179 continue;
180 }
181 if (boundary != 0 &&
182 ((try ^ (try + num - 1)) & pagemask) != 0) {
183 /*
184 * Region crosses boundary. Jump to the boundary
185 * just crossed and ensure alignment.
186 */
187 try = (try + num - 1) & pagemask;
188 try = roundup2(try, alignment);
189 skip = 0;
190 continue;
191 }
192 #ifdef DEBUG
193 /*
194 * Make sure this is a managed physical page.
195 */
196
197 if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
198 panic("pgalloc contig: botch1");
199 if (cidx != try - ps->start)
200 panic("pgalloc contig: botch2");
201 if (vm_physseg_find(try + num - 1, &cidx) != ps - vm_physmem)
202 panic("pgalloc contig: botch3");
203 if (cidx != try - ps->start + num - 1)
204 panic("pgalloc contig: botch4");
205 #endif
206 tryidx = try - ps->start;
207 end = tryidx + num;
208
209 /*
210 * Found a suitable starting page. See if the range is free.
211 */
212 #ifdef PGALLOC_VERBOSE
213 printf("%s: ps=%p try=%#x end=%#x skip=%#x, align=%#"PRIxPADDR,
214 __func__, ps, tryidx, end, skip, alignment);
215 #endif
216 /*
217 * We start at the end and work backwards since if we find a
218 * non-free page, it makes no sense to continue.
219 *
220 * But on the plus size we have "vetted" some number of free
221 * pages. If this iteration fails, we may be able to skip
222 * testing most of those pages again in the next pass.
223 */
224 for (idx = end - 1; idx >= tryidx + skip; idx--) {
225 if (VM_PAGE_IS_FREE(&pgs[idx]) == 0) {
226 ok = false;
227 break;
228 }
229
230 #ifdef DEBUG
231 if (idx > tryidx) {
232 idxpa = VM_PAGE_TO_PHYS(&pgs[idx]);
233 lastidxpa = VM_PAGE_TO_PHYS(&pgs[idx - 1]);
234 if ((lastidxpa + PAGE_SIZE) != idxpa) {
235 /*
236 * Region not contiguous.
237 */
238 panic("pgalloc contig: botch5");
239 }
240 if (boundary != 0 &&
241 ((lastidxpa ^ idxpa) & ~(boundary - 1))
242 != 0) {
243 /*
244 * Region crosses boundary.
245 */
246 panic("pgalloc contig: botch6");
247 }
248 }
249 #endif
250 }
251
252 if (ok) {
253 while (skip-- > 0) {
254 KDASSERT(VM_PAGE_IS_FREE(&pgs[tryidx + skip]));
255 }
256 #ifdef PGALLOC_VERBOSE
257 printf(": ok\n");
258 #endif
259 break;
260 }
261
262 #ifdef PGALLOC_VERBOSE
263 printf(": non-free at %#x\n", idx - tryidx);
264 #endif
265 /*
266 * count the number of pages we can advance
267 * since we know they aren't all free.
268 */
269 cnt = idx + 1 - tryidx;
270 /*
271 * now round up that to the needed alignment.
272 */
273 cnt = roundup2(cnt, alignment);
274 /*
275 * The number of pages we can skip checking
276 * (might be 0 if cnt > num).
277 */
278 skip = max(num - cnt, 0);
279 try += cnt;
280 }
281
282 /*
283 * we have a chunk of memory that conforms to the requested constraints.
284 */
285 for (idx = tryidx, pgs += idx; idx < end; idx++, pgs++)
286 uvm_pglist_add(pgs, rlist);
287
288 /*
289 * the next time we need to search this segment, start after this
290 * chunk of pages we just allocated.
291 */
292 ps->start_hint = try + num - ps->avail_start;
293 KASSERTMSG(ps->start_hint <= ps->avail_end - ps->avail_start,
294 ("%x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
295 try + num,
296 ps->start_hint, ps->start_hint, ps->avail_end, ps->avail_start,
297 ps->avail_end - ps->avail_start));
298
299 #ifdef PGALLOC_VERBOSE
300 printf("got %d pgs\n", num);
301 #endif
302 return num; /* number of pages allocated */
303 }
304
305 static int
306 uvm_pglistalloc_contig(int num, paddr_t low, paddr_t high, paddr_t alignment,
307 paddr_t boundary, struct pglist *rlist)
308 {
309 int fl, psi;
310 struct vm_physseg *ps;
311 int error;
312
313 /* Default to "lose". */
314 error = ENOMEM;
315
316 /*
317 * Block all memory allocation and lock the free list.
318 */
319 mutex_spin_enter(&uvm_fpageqlock);
320
321 /* Are there even any free pages? */
322 if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
323 goto out;
324
325 for (fl = 0; fl < VM_NFREELIST; fl++) {
326 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
327 for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
328 #else
329 for (psi = 0 ; psi < vm_nphysseg ; psi++)
330 #endif
331 {
332 ps = &vm_physmem[psi];
333
334 if (ps->free_list != fl)
335 continue;
336
337 num -= uvm_pglistalloc_c_ps(ps, num, low, high,
338 alignment, boundary, rlist);
339 if (num == 0) {
340 #ifdef PGALLOC_VERBOSE
341 printf("pgalloc: %"PRIxMAX"-%"PRIxMAX"\n",
342 (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
343 (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
344 #endif
345 error = 0;
346 goto out;
347 }
348 }
349 }
350
351 out:
352 /*
353 * check to see if we need to generate some free pages waking
354 * the pagedaemon.
355 */
356
357 uvm_kick_pdaemon();
358 mutex_spin_exit(&uvm_fpageqlock);
359 return (error);
360 }
361
362 static int
363 uvm_pglistalloc_s_ps(struct vm_physseg *ps, int num, paddr_t low, paddr_t high,
364 struct pglist *rlist)
365 {
366 int todo, limit, try;
367 struct vm_page *pg;
368 bool second_pass;
369 #ifdef PGALLOC_VERBOSE
370 printf("pgalloc: simple %d pgs from psi %zd\n", num, ps - vm_physmem);
371 #endif
372
373 KASSERT(mutex_owned(&uvm_fpageqlock));
374 KASSERT(ps->start <= ps->avail_start);
375 KASSERT(ps->start <= ps->avail_end);
376 KASSERT(ps->avail_start <= ps->end);
377 KASSERT(ps->avail_end <= ps->end);
378
379 low = atop(low);
380 high = atop(high);
381 todo = num;
382 try = max(low, ps->avail_start + ps->start_hint);
383 limit = min(high, ps->avail_end);
384 pg = &ps->pgs[try - ps->start];
385 second_pass = false;
386
387 /*
388 * Make sure that physseg falls within with range to be allocated from.
389 */
390 if (high <= ps->avail_start || low >= ps->avail_end)
391 return 0;
392
393 again:
394 for (;; try++, pg++) {
395 if (try >= limit) {
396 if (ps->start_hint == 0 || second_pass) {
397 try = limit - 1;
398 break;
399 }
400 second_pass = true;
401 try = max(low, ps->avail_start);
402 limit = min(limit, ps->avail_start + ps->start_hint);
403 pg = &ps->pgs[try - ps->start];
404 goto again;
405 }
406 #if defined(DEBUG)
407 {
408 int cidx = 0;
409 const int bank = vm_physseg_find(try, &cidx);
410 KDASSERTMSG(bank == ps - vm_physmem,
411 ("vm_physseg_find(%#x) (%d) != ps %zd",
412 try, bank, ps - vm_physmem));
413 KDASSERTMSG(cidx == try - ps->start,
414 ("vm_physseg_find(%#x): %#x != off %"PRIxPADDR,
415 try, cidx, try - ps->start));
416 }
417 #endif
418 if (VM_PAGE_IS_FREE(pg) == 0)
419 continue;
420
421 uvm_pglist_add(pg, rlist);
422 if (--todo == 0) {
423 break;
424 }
425 }
426
427 /*
428 * The next time we need to search this segment,
429 * start just after the pages we just allocated.
430 */
431 ps->start_hint = try + 1 - ps->avail_start;
432 KASSERTMSG(ps->start_hint <= ps->avail_end - ps->avail_start,
433 ("%#x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
434 try + 1,
435 ps->start_hint, ps->start_hint, ps->avail_end, ps->avail_start,
436 ps->avail_end - ps->avail_start));
437
438 #ifdef PGALLOC_VERBOSE
439 printf("got %d pgs\n", num - todo);
440 #endif
441 return (num - todo); /* number of pages allocated */
442 }
443
444 static int
445 uvm_pglistalloc_simple(int num, paddr_t low, paddr_t high,
446 struct pglist *rlist, int waitok)
447 {
448 int fl, psi, error;
449 struct vm_physseg *ps;
450
451 /* Default to "lose". */
452 error = ENOMEM;
453
454 again:
455 /*
456 * Block all memory allocation and lock the free list.
457 */
458 mutex_spin_enter(&uvm_fpageqlock);
459
460 /* Are there even any free pages? */
461 if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
462 goto out;
463
464 for (fl = 0; fl < VM_NFREELIST; fl++) {
465 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
466 for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
467 #else
468 for (psi = 0 ; psi < vm_nphysseg ; psi++)
469 #endif
470 {
471 ps = &vm_physmem[psi];
472
473 if (ps->free_list != fl)
474 continue;
475
476 num -= uvm_pglistalloc_s_ps(ps, num, low, high, rlist);
477 if (num == 0) {
478 error = 0;
479 goto out;
480 }
481 }
482
483 }
484
485 out:
486 /*
487 * check to see if we need to generate some free pages waking
488 * the pagedaemon.
489 */
490
491 uvm_kick_pdaemon();
492 mutex_spin_exit(&uvm_fpageqlock);
493
494 if (error) {
495 if (waitok) {
496 /* XXX perhaps some time limitation? */
497 #ifdef DEBUG
498 printf("pglistalloc waiting\n");
499 #endif
500 uvm_wait("pglalloc");
501 goto again;
502 } else
503 uvm_pglistfree(rlist);
504 }
505 #ifdef PGALLOC_VERBOSE
506 if (!error)
507 printf("pgalloc: %"PRIxMAX"..%"PRIxMAX"\n",
508 (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_FIRST(rlist)),
509 (uintmax_t) VM_PAGE_TO_PHYS(TAILQ_LAST(rlist, pglist)));
510 #endif
511 return (error);
512 }
513
514 int
515 uvm_pglistalloc(psize_t size, paddr_t low, paddr_t high, paddr_t alignment,
516 paddr_t boundary, struct pglist *rlist, int nsegs, int waitok)
517 {
518 int num, res;
519
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 if ((nsegs < size >> PAGE_SHIFT) || (alignment != PAGE_SIZE) ||
537 (boundary != 0))
538 res = uvm_pglistalloc_contig(num, low, high, alignment,
539 boundary, rlist);
540 else
541 res = uvm_pglistalloc_simple(num, low, high, rlist, waitok);
542
543 return (res);
544 }
545
546 /*
547 * uvm_pglistfree: free a list of pages
548 *
549 * => pages should already be unmapped
550 */
551
552 void
553 uvm_pglistfree(struct pglist *list)
554 {
555 struct uvm_cpu *ucpu;
556 struct vm_page *pg;
557 int index, color, queue;
558 bool iszero;
559
560 /*
561 * Lock the free list and free each page.
562 */
563
564 mutex_spin_enter(&uvm_fpageqlock);
565 ucpu = curcpu()->ci_data.cpu_uvm;
566 while ((pg = TAILQ_FIRST(list)) != NULL) {
567 KASSERT(!uvmpdpol_pageisqueued_p(pg));
568 TAILQ_REMOVE(list, pg, pageq.queue);
569 iszero = (pg->flags & PG_ZERO);
570 pg->pqflags = PQ_FREE;
571 #ifdef DEBUG
572 pg->uobject = (void *)0xdeadbeef;
573 pg->uanon = (void *)0xdeadbeef;
574 #endif /* DEBUG */
575 #ifdef DEBUG
576 if (iszero)
577 uvm_pagezerocheck(pg);
578 #endif /* DEBUG */
579 index = uvm_page_lookup_freelist(pg);
580 color = VM_PGCOLOR_BUCKET(pg);
581 queue = iszero ? PGFL_ZEROS : PGFL_UNKNOWN;
582 pg->offset = (uintptr_t)ucpu;
583 LIST_INSERT_HEAD(&uvm.page_free[index].pgfl_buckets[color].
584 pgfl_queues[queue], pg, pageq.list);
585 LIST_INSERT_HEAD(&ucpu->page_free[index].pgfl_buckets[color].
586 pgfl_queues[queue], pg, listq.list);
587 uvmexp.free++;
588 if (iszero)
589 uvmexp.zeropages++;
590 ucpu->pages[queue]++;
591 STAT_DECR(uvm_pglistalloc_npages);
592 }
593 if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN])
594 ucpu->page_idle_zero = vm_page_zero_enable;
595 mutex_spin_exit(&uvm_fpageqlock);
596 }
597