1/* $NetBSD: uvm_pager.c,v 1.110 2014/03/01 18:32:01 christos Exp $ */
2
3/*
4 * Copyright (c) 1997 Charles D. Cranor and Washington University.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 *
27 * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
28 */
29
30/*
31 * uvm_pager.c: generic functions used to assist the pagers.
32 */
33
34#include <sys/cdefs.h>
35__KERNEL_RCSID(0, "$NetBSD: uvm_pager.c,v 1.110 2014/03/01 18:32:01 christos Exp $");
36
37#include "opt_uvmhist.h"
38#include "opt_readahead.h"
39#include "opt_pagermap.h"
40
41#include <sys/param.h>
42#include <sys/systm.h>
43#include <sys/vnode.h>
44#include <sys/buf.h>
45
46#include <uvm/uvm.h>
47
48/*
49 * XXX
50 * this is needed until the device strategy interface
51 * is changed to do physically-addressed i/o.
52 */
53
54#ifndef PAGER_MAP_DEFAULT_SIZE
55#define PAGER_MAP_DEFAULT_SIZE (16 * 1024 * 1024)
56#endif
57
58#ifndef PAGER_MAP_SIZE
59#define PAGER_MAP_SIZE PAGER_MAP_DEFAULT_SIZE
60#endif
61
62size_t pager_map_size = PAGER_MAP_SIZE;
63
64/*
65 * list of uvm pagers in the system
66 */
67
68const struct uvm_pagerops * const uvmpagerops[] = {
69 &aobj_pager,
70 &uvm_deviceops,
71 &uvm_vnodeops,
72 &ubc_pager,
73};
74
75/*
76 * the pager map: provides KVA for I/O
77 */
78
79struct vm_map *pager_map; /* XXX */
80kmutex_t pager_map_wanted_lock;
81bool pager_map_wanted; /* locked by pager map */
82static vaddr_t emergva;
83static int emerg_ncolors;
84static bool emerginuse;
85
86void
87uvm_pager_realloc_emerg(void)
88{
89 vaddr_t new_emergva, old_emergva;
90 int old_emerg_ncolors;
91
92 if (__predict_true(emergva != 0 && emerg_ncolors >= uvmexp.ncolors))
93 return;
94
95 KASSERT(!emerginuse);
96
97 new_emergva = uvm_km_alloc(kernel_map,
98 round_page(MAXPHYS) + ptoa(uvmexp.ncolors), ptoa(uvmexp.ncolors),
99 UVM_KMF_VAONLY);
100
101 KASSERT(new_emergva != 0);
102
103 old_emergva = emergva;
104 old_emerg_ncolors = emerg_ncolors;
105
106 /*
107 * don't support re-color in late boot anyway.
108 */
109 if (0) /* XXX */
110 mutex_enter(&pager_map_wanted_lock);
111
112 emergva = new_emergva;
113 emerg_ncolors = uvmexp.ncolors;
114 wakeup(&old_emergva);
115
116 if (0) /* XXX */
117 mutex_exit(&pager_map_wanted_lock);
118
119 if (old_emergva)
120 uvm_km_free(kernel_map, old_emergva,
121 round_page(MAXPHYS) + ptoa(old_emerg_ncolors),
122 UVM_KMF_VAONLY);
123}
124
125/*
126 * uvm_pager_init: init pagers (at boot time)
127 */
128
129void
130uvm_pager_init(void)
131{
132 u_int lcv;
133 vaddr_t sva, eva;
134
135 /*
136 * init pager map
137 */
138
139 sva = 0;
140 pager_map = uvm_km_suballoc(kernel_map, &sva, &eva, pager_map_size, 0,
141 false, NULL);
142 mutex_init(&pager_map_wanted_lock, MUTEX_DEFAULT, IPL_NONE);
143 pager_map_wanted = false;
144
145 uvm_pager_realloc_emerg();
146
147 /*
148 * init ASYNC I/O queue
149 */
150
151 TAILQ_INIT(&uvm.aio_done);
152
153 /*
154 * call pager init functions
155 */
156 for (lcv = 0 ; lcv < __arraycount(uvmpagerops); lcv++) {
157 if (uvmpagerops[lcv]->pgo_init)
158 uvmpagerops[lcv]->pgo_init();
159 }
160}
161
162/*
163 * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
164 *
165 * we basically just map in a blank map entry to reserve the space in the
166 * map and then use pmap_enter() to put the mappings in by hand.
167 */
168
169vaddr_t
170uvm_pagermapin(struct vm_page **pps, int npages, int flags)
171{
172 vsize_t size;
173 vaddr_t kva;
174 vaddr_t cva;
175 struct vm_page *pp;
176 vm_prot_t prot;
177 const bool pdaemon = (curlwp == uvm.pagedaemon_lwp);
178 const u_int first_color = VM_PGCOLOR_BUCKET(*pps);
179 UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
180
181 UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d, first_color=%u)",
182 pps, npages, first_color, 0);
183
184 /*
185 * compute protection. outgoing I/O only needs read
186 * access to the page, whereas incoming needs read/write.
187 */
188
189 prot = VM_PROT_READ;
190 if (flags & UVMPAGER_MAPIN_READ)
191 prot |= VM_PROT_WRITE;
192
193ReStart:
194 size = ptoa(npages);
195 kva = 0; /* let system choose VA */
196
197 if (uvm_map(pager_map, &kva, size, NULL, UVM_UNKNOWN_OFFSET,
198 first_color, UVM_FLAG_COLORMATCH | UVM_FLAG_NOMERGE
199 | (pdaemon ? UVM_FLAG_NOWAIT : 0)) != 0) {
200 if (pdaemon) {
201 mutex_enter(&pager_map_wanted_lock);
202 if (emerginuse) {
203 UVM_UNLOCK_AND_WAIT(&emergva,
204 &pager_map_wanted_lock, false,
205 "emergva", 0);
206 goto ReStart;
207 }
208 emerginuse = true;
209 mutex_exit(&pager_map_wanted_lock);
210 kva = emergva + ptoa(first_color);
211 /* The shift implicitly truncates to PAGE_SIZE */
212 KASSERT(npages <= (MAXPHYS >> PAGE_SHIFT));
213 goto enter;
214 }
215 if ((flags & UVMPAGER_MAPIN_WAITOK) == 0) {
216 UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
217 return(0);
218 }
219 mutex_enter(&pager_map_wanted_lock);
220 pager_map_wanted = true;
221 UVMHIST_LOG(maphist, " SLEEPING on pager_map",0,0,0,0);
222 UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, false,
223 "pager_map", 0);
224 goto ReStart;
225 }
226
227enter:
228 /* got it */
229 for (cva = kva; npages != 0; npages--, cva += PAGE_SIZE) {
230 pp = *pps++;
231 KASSERT(pp);
232 // KASSERT(!((VM_PAGE_TO_PHYS(pp) ^ cva) & uvmexp.colormask));
233 KASSERT(pp->flags & PG_BUSY);
234 pmap_kenter_pa(cva, VM_PAGE_TO_PHYS(pp), prot, 0);
235 }
236 pmap_update(vm_map_pmap(pager_map));
237
238 UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
239 return(kva);
240}
241
242/*
243 * uvm_pagermapout: remove pager_map mapping
244 *
245 * we remove our mappings by hand and then remove the mapping (waking
246 * up anyone wanting space).
247 */
248
249void
250uvm_pagermapout(vaddr_t kva, int npages)
251{
252 vsize_t size = ptoa(npages);
253 struct vm_map_entry *entries;
254 UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
255
256 UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
257
258 /*
259 * duplicate uvm_unmap, but add in pager_map_wanted handling.
260 */
261
262 pmap_kremove(kva, size);
263 pmap_update(pmap_kernel());
264
265 if ((kva & ~ptoa(uvmexp.colormask)) == emergva) {
266 mutex_enter(&pager_map_wanted_lock);
267 KASSERT(emerginuse);
268 emerginuse = false;
269 wakeup(&emergva);
270 mutex_exit(&pager_map_wanted_lock);
271 return;
272 }
273
274 vm_map_lock(pager_map);
275 uvm_unmap_remove(pager_map, kva, kva + size, &entries, 0);
276 mutex_enter(&pager_map_wanted_lock);
277 if (pager_map_wanted) {
278 pager_map_wanted = false;
279 wakeup(pager_map);
280 }
281 mutex_exit(&pager_map_wanted_lock);
282 vm_map_unlock(pager_map);
283 if (entries)
284 uvm_unmap_detach(entries, 0);
285 UVMHIST_LOG(maphist,"<- done",0,0,0,0);
286}
287
288/*
289 * interrupt-context iodone handler for single-buf i/os
290 * or the top-level buf of a nested-buf i/o.
291 */
292
293void
294uvm_aio_biodone(struct buf *bp)
295{
296 /* reset b_iodone for when this is a single-buf i/o. */
297 bp->b_iodone = uvm_aio_aiodone;
298
299 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL);
300}
301
302void
303uvm_aio_aiodone_pages(struct vm_page **pgs, int npages, bool write, int error)
304{
305 struct uvm_object *uobj;
306 struct vm_page *pg;
307 kmutex_t *slock;
308 int pageout_done; /* number of PG_PAGEOUT pages processed */
309 int swslot;
310 int i;
311 bool swap;
312 UVMHIST_FUNC("uvm_aio_aiodone_pages"); UVMHIST_CALLED(ubchist);
313
314 swslot = 0;
315 pageout_done = 0;
316 slock = NULL;
317 uobj = NULL;
318 pg = pgs[0];
319 swap = (pg->uanon != NULL && pg->uobject == NULL) ||
320 (pg->pqflags & PQ_AOBJ) != 0;
321 if (!swap) {
322 uobj = pg->uobject;
323 slock = uobj->vmobjlock;
324 mutex_enter(slock);
325 mutex_enter(&uvm_pageqlock);
326 } else {
327#if defined(VMSWAP)
328 if (error) {
329 if (pg->uobject != NULL) {
330 swslot = uao_find_swslot(pg->uobject,
331 pg->offset >> PAGE_SHIFT);
332 } else {
333 KASSERT(pg->uanon != NULL);
334 swslot = pg->uanon->an_swslot;
335 }
336 KASSERT(swslot);
337 }
338#else /* defined(VMSWAP) */
339 panic("%s: swap", __func__);
340#endif /* defined(VMSWAP) */
341 }
342 for (i = 0; i < npages; i++) {
343#if defined(VMSWAP)
344 bool anon_disposed = false; /* XXX gcc */
345#endif /* defined(VMSWAP) */
346
347 pg = pgs[i];
348 KASSERT(swap || pg->uobject == uobj);
349 UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0);
350
351#if defined(VMSWAP)
352 /*
353 * for swap i/os, lock each page's object (or anon)
354 * individually since each page may need a different lock.
355 */
356
357 if (swap) {
358 if (pg->uobject != NULL) {
359 slock = pg->uobject->vmobjlock;
360 } else {
361 slock = pg->uanon->an_lock;
362 }
363 mutex_enter(slock);
364 mutex_enter(&uvm_pageqlock);
365 anon_disposed = (pg->flags & PG_RELEASED) != 0;
366 KASSERT(!anon_disposed || pg->uobject != NULL ||
367 pg->uanon->an_ref == 0);
368 }
369#endif /* defined(VMSWAP) */
370
371 /*
372 * process errors. for reads, just mark the page to be freed.
373 * for writes, if the error was ENOMEM, we assume this was
374 * a transient failure so we mark the page dirty so that
375 * we'll try to write it again later. for all other write
376 * errors, we assume the error is permanent, thus the data
377 * in the page is lost. bummer.
378 */
379
380 if (error) {
381 int slot;
382 if (!write) {
383 pg->flags |= PG_RELEASED;
384 continue;
385 } else if (error == ENOMEM) {
386 if (pg->flags & PG_PAGEOUT) {
387 pg->flags &= ~PG_PAGEOUT;
388 pageout_done++;
389 }
390 pg->flags &= ~PG_CLEAN;
391 uvm_pageactivate(pg);
392 slot = 0;
393 } else
394 slot = SWSLOT_BAD;
395
396#if defined(VMSWAP)
397 if (swap) {
398 if (pg->uobject != NULL) {
399 int oldslot __diagused;
400 oldslot = uao_set_swslot(pg->uobject,
401 pg->offset >> PAGE_SHIFT, slot);
402 KASSERT(oldslot == swslot + i);
403 } else {
404 KASSERT(pg->uanon->an_swslot ==
405 swslot + i);
406 pg->uanon->an_swslot = slot;
407 }
408 }
409#endif /* defined(VMSWAP) */
410 }
411
412 /*
413 * if the page is PG_FAKE, this must have been a read to
414 * initialize the page. clear PG_FAKE and activate the page.
415 * we must also clear the pmap "modified" flag since it may
416 * still be set from the page's previous identity.
417 */
418
419 if (pg->flags & PG_FAKE) {
420 KASSERT(!write);
421 pg->flags &= ~PG_FAKE;
422#if defined(READAHEAD_STATS)
423 pg->pqflags |= PQ_READAHEAD;
424 uvm_ra_total.ev_count++;
425#endif /* defined(READAHEAD_STATS) */
426 KASSERT((pg->flags & PG_CLEAN) != 0);
427 uvm_pageenqueue(pg);
428 pmap_clear_modify(pg);
429 }
430
431 /*
432 * do accounting for pagedaemon i/o and arrange to free
433 * the pages instead of just unbusying them.
434 */
435
436 if (pg->flags & PG_PAGEOUT) {
437 pg->flags &= ~PG_PAGEOUT;
438 pageout_done++;
439 uvmexp.pdfreed++;
440 pg->flags |= PG_RELEASED;
441 }
442
443#if defined(VMSWAP)
444 /*
445 * for swap pages, unlock everything for this page now.
446 */
447
448 if (swap) {
449 if (pg->uobject == NULL && anon_disposed) {
450 mutex_exit(&uvm_pageqlock);
451 uvm_anon_release(pg->uanon);
452 } else {
453 uvm_page_unbusy(&pg, 1);
454 mutex_exit(&uvm_pageqlock);
455 mutex_exit(slock);
456 }
457 }
458#endif /* defined(VMSWAP) */
459 }
460 uvm_pageout_done(pageout_done);
461 if (!swap) {
462 uvm_page_unbusy(pgs, npages);
463 mutex_exit(&uvm_pageqlock);
464 mutex_exit(slock);
465 } else {
466#if defined(VMSWAP)
467 KASSERT(write);
468
469 /* these pages are now only in swap. */
470 mutex_enter(&uvm_swap_data_lock);
471 if (error != ENOMEM) {
472 KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse);
473 uvmexp.swpgonly += npages;
474 }
475 mutex_exit(&uvm_swap_data_lock);
476 if (error) {
477 if (error != ENOMEM)
478 uvm_swap_markbad(swslot, npages);
479 else
480 uvm_swap_free(swslot, npages);
481 }
482 uvmexp.pdpending--;
483#endif /* defined(VMSWAP) */
484 }
485}
486
487/*
488 * uvm_aio_aiodone: do iodone processing for async i/os.
489 * this should be called in thread context, not interrupt context.
490 */
491
492void
493uvm_aio_aiodone(struct buf *bp)
494{
495 int npages = bp->b_bufsize >> PAGE_SHIFT;
496 struct vm_page *pgs[npages];
497 int i, error;
498 bool write;
499 UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist);
500 UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0);
501
502 error = bp->b_error;
503 write = (bp->b_flags & B_READ) == 0;
504
505 for (i = 0; i < npages; i++) {
506 pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
507 UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0);
508 }
509 uvm_pagermapout((vaddr_t)bp->b_data, npages);
510
511 uvm_aio_aiodone_pages(pgs, npages, write, error);
512
513 if (write && (bp->b_cflags & BC_AGE) != 0) {
514 mutex_enter(bp->b_objlock);
515 vwakeup(bp);
516 mutex_exit(bp->b_objlock);
517 }
518 putiobuf(bp);
519}
520
521/*
522 * uvm_pageratop: convert KVAs in the pager map back to their page
523 * structures.
524 */
525
526struct vm_page *
527uvm_pageratop(vaddr_t kva)
528{
529 struct vm_page *pg;
530 paddr_t pa;
531 bool rv __diagused;
532
533 rv = pmap_extract(pmap_kernel(), kva, &pa);
534 KASSERT(rv);
535 pg = PHYS_TO_VM_PAGE(pa);
536 KASSERT(pg != NULL);
537 return (pg);
538}
539