xref: /netbsd-src/sys/uvm/uvm_pager.c (revision e6d6e05cb173f30287ab619b21120b27baa66ad6)
1 /*	$NetBSD: uvm_pager.c,v 1.91 2008/02/29 20:35:23 yamt Exp $	*/
2 
3 /*
4  *
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
35  */
36 
37 /*
38  * uvm_pager.c: generic functions used to assist the pagers.
39  */
40 
41 #include <sys/cdefs.h>
42 __KERNEL_RCSID(0, "$NetBSD: uvm_pager.c,v 1.91 2008/02/29 20:35:23 yamt Exp $");
43 
44 #include "opt_uvmhist.h"
45 #include "opt_readahead.h"
46 #include "opt_pagermap.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/proc.h>
51 #include <sys/malloc.h>
52 #include <sys/pool.h>
53 #include <sys/vnode.h>
54 
55 #include <uvm/uvm.h>
56 
57 /*
58  * XXX
59  * this is needed until the device strategy interface
60  * is changed to do physically-addressed i/o.
61  */
62 
63 #ifndef PAGER_MAP_DEFAULT_SIZE
64 #define PAGER_MAP_DEFAULT_SIZE	(16 * 1024 * 1024)
65 #endif
66 
67 #ifndef PAGER_MAP_SIZE
68 #define PAGER_MAP_SIZE	PAGER_MAP_DEFAULT_SIZE
69 #endif
70 
71 size_t pager_map_size = PAGER_MAP_SIZE;
72 
73 struct pool *uvm_aiobuf_pool;
74 
75 /*
76  * list of uvm pagers in the system
77  */
78 
79 const struct uvm_pagerops * const uvmpagerops[] = {
80 	&aobj_pager,
81 	&uvm_deviceops,
82 	&uvm_vnodeops,
83 	&ubc_pager,
84 };
85 
86 /*
87  * the pager map: provides KVA for I/O
88  */
89 
90 struct vm_map *pager_map;		/* XXX */
91 kmutex_t pager_map_wanted_lock;
92 bool pager_map_wanted;	/* locked by pager map */
93 static vaddr_t emergva;
94 static bool emerginuse;
95 
96 /*
97  * uvm_pager_init: init pagers (at boot time)
98  */
99 
100 void
101 uvm_pager_init(void)
102 {
103 	u_int lcv;
104 	vaddr_t sva, eva;
105 
106 	/*
107 	 * init pager map
108 	 */
109 
110 	sva = 0;
111 	pager_map = uvm_km_suballoc(kernel_map, &sva, &eva, pager_map_size, 0,
112 	    false, NULL);
113 	mutex_init(&pager_map_wanted_lock, MUTEX_DEFAULT, IPL_NONE);
114 	pager_map_wanted = false;
115 	emergva = uvm_km_alloc(kernel_map, round_page(MAXPHYS), 0,
116 	    UVM_KMF_VAONLY);
117 #if defined(DEBUG)
118 	if (emergva == 0)
119 		panic("emergva");
120 #endif
121 	emerginuse = false;
122 
123 	/*
124 	 * init ASYNC I/O queue
125 	 */
126 
127 	TAILQ_INIT(&uvm.aio_done);
128 
129 	/*
130 	 * call pager init functions
131 	 */
132 	for (lcv = 0 ; lcv < __arraycount(uvmpagerops); lcv++) {
133 		if (uvmpagerops[lcv]->pgo_init)
134 			uvmpagerops[lcv]->pgo_init();
135 	}
136 }
137 
138 /*
139  * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
140  *
141  * we basically just map in a blank map entry to reserve the space in the
142  * map and then use pmap_enter() to put the mappings in by hand.
143  */
144 
145 vaddr_t
146 uvm_pagermapin(struct vm_page **pps, int npages, int flags)
147 {
148 	vsize_t size;
149 	vaddr_t kva;
150 	vaddr_t cva;
151 	struct vm_page *pp;
152 	vm_prot_t prot;
153 	const bool pdaemon = curlwp == uvm.pagedaemon_lwp;
154 	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
155 
156 	UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d)", pps, npages,0,0);
157 
158 	/*
159 	 * compute protection.  outgoing I/O only needs read
160 	 * access to the page, whereas incoming needs read/write.
161 	 */
162 
163 	prot = VM_PROT_READ;
164 	if (flags & UVMPAGER_MAPIN_READ)
165 		prot |= VM_PROT_WRITE;
166 
167 ReStart:
168 	size = npages << PAGE_SHIFT;
169 	kva = 0;			/* let system choose VA */
170 
171 	if (uvm_map(pager_map, &kva, size, NULL, UVM_UNKNOWN_OFFSET, 0,
172 	    UVM_FLAG_NOMERGE | (pdaemon ? UVM_FLAG_NOWAIT : 0)) != 0) {
173 		if (pdaemon) {
174 			mutex_enter(&pager_map_wanted_lock);
175 			if (emerginuse) {
176 				UVM_UNLOCK_AND_WAIT(&emergva,
177 				    &pager_map_wanted_lock, false,
178 				    "emergva", 0);
179 				goto ReStart;
180 			}
181 			emerginuse = true;
182 			mutex_exit(&pager_map_wanted_lock);
183 			kva = emergva;
184 			/* The shift implicitly truncates to PAGE_SIZE */
185 			KASSERT(npages <= (MAXPHYS >> PAGE_SHIFT));
186 			goto enter;
187 		}
188 		if ((flags & UVMPAGER_MAPIN_WAITOK) == 0) {
189 			UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
190 			return(0);
191 		}
192 		mutex_enter(&pager_map_wanted_lock);
193 		pager_map_wanted = true;
194 		UVMHIST_LOG(maphist, "  SLEEPING on pager_map",0,0,0,0);
195 		UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, false,
196 		    "pager_map", 0);
197 		goto ReStart;
198 	}
199 
200 enter:
201 	/* got it */
202 	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
203 		pp = *pps++;
204 		KASSERT(pp);
205 		KASSERT(pp->flags & PG_BUSY);
206 		pmap_kenter_pa(cva, VM_PAGE_TO_PHYS(pp), prot);
207 	}
208 	pmap_update(vm_map_pmap(pager_map));
209 
210 	UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
211 	return(kva);
212 }
213 
214 /*
215  * uvm_pagermapout: remove pager_map mapping
216  *
217  * we remove our mappings by hand and then remove the mapping (waking
218  * up anyone wanting space).
219  */
220 
221 void
222 uvm_pagermapout(vaddr_t kva, int npages)
223 {
224 	vsize_t size = npages << PAGE_SHIFT;
225 	struct vm_map_entry *entries;
226 	UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
227 
228 	UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
229 
230 	/*
231 	 * duplicate uvm_unmap, but add in pager_map_wanted handling.
232 	 */
233 
234 	pmap_kremove(kva, npages << PAGE_SHIFT);
235 	if (kva == emergva) {
236 		mutex_enter(&pager_map_wanted_lock);
237 		emerginuse = false;
238 		wakeup(&emergva);
239 		mutex_exit(&pager_map_wanted_lock);
240 		return;
241 	}
242 
243 	vm_map_lock(pager_map);
244 	uvm_unmap_remove(pager_map, kva, kva + size, &entries, NULL, 0);
245 	mutex_enter(&pager_map_wanted_lock);
246 	if (pager_map_wanted) {
247 		pager_map_wanted = false;
248 		wakeup(pager_map);
249 	}
250 	mutex_exit(&pager_map_wanted_lock);
251 	vm_map_unlock(pager_map);
252 	if (entries)
253 		uvm_unmap_detach(entries, 0);
254 	pmap_update(pmap_kernel());
255 	UVMHIST_LOG(maphist,"<- done",0,0,0,0);
256 }
257 
258 /*
259  * interrupt-context iodone handler for nested i/o bufs.
260  *
261  * => the buffer is private so need not be locked here
262  */
263 
264 void
265 uvm_aio_biodone1(struct buf *bp)
266 {
267 	struct buf *mbp = bp->b_private;
268 
269 	KASSERT(mbp != bp);
270 	if (bp->b_error != 0) {
271 		mbp->b_error = bp->b_error;
272 	}
273 	mbp->b_resid -= bp->b_bcount;
274 	putiobuf(bp);
275 	if (mbp->b_resid == 0) {
276 		biodone(mbp);
277 	}
278 }
279 
280 /*
281  * interrupt-context iodone handler for single-buf i/os
282  * or the top-level buf of a nested-buf i/o.
283  */
284 
285 void
286 uvm_aio_biodone(struct buf *bp)
287 {
288 	/* reset b_iodone for when this is a single-buf i/o. */
289 	bp->b_iodone = uvm_aio_aiodone;
290 
291 	workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL);
292 }
293 
294 void
295 uvm_aio_aiodone_pages(struct vm_page **pgs, int npages, bool write, int error)
296 {
297 	struct uvm_object *uobj;
298 	struct vm_page *pg;
299 	kmutex_t *slock;
300 	int pageout_done;
301 	int swslot;
302 	int i;
303 	bool swap;
304 
305 	swslot = 0;
306 	pageout_done = 0;
307 	slock = NULL;
308 	uobj = NULL;
309 	pg = pgs[0];
310 	swap = (pg->uanon != NULL && pg->uobject == NULL) ||
311 		(pg->pqflags & PQ_AOBJ) != 0;
312 	if (!swap) {
313 		uobj = pg->uobject;
314 		slock = &uobj->vmobjlock;
315 		mutex_enter(slock);
316 		mutex_enter(&uvm_pageqlock);
317 	} else {
318 #if defined(VMSWAP)
319 		if (error) {
320 			if (pg->uobject != NULL) {
321 				swslot = uao_find_swslot(pg->uobject,
322 				    pg->offset >> PAGE_SHIFT);
323 			} else {
324 				KASSERT(pg->uanon != NULL);
325 				swslot = pg->uanon->an_swslot;
326 			}
327 			KASSERT(swslot);
328 		}
329 #else /* defined(VMSWAP) */
330 		panic("%s: swap", __func__);
331 #endif /* defined(VMSWAP) */
332 	}
333 	for (i = 0; i < npages; i++) {
334 		pg = pgs[i];
335 		KASSERT(swap || pg->uobject == uobj);
336 		UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0);
337 
338 #if defined(VMSWAP)
339 		/*
340 		 * for swap i/os, lock each page's object (or anon)
341 		 * individually since each page may need a different lock.
342 		 */
343 
344 		if (swap) {
345 			if (pg->uobject != NULL) {
346 				slock = &pg->uobject->vmobjlock;
347 			} else {
348 				slock = &pg->uanon->an_lock;
349 			}
350 			mutex_enter(slock);
351 			mutex_enter(&uvm_pageqlock);
352 		}
353 #endif /* defined(VMSWAP) */
354 
355 		/*
356 		 * process errors.  for reads, just mark the page to be freed.
357 		 * for writes, if the error was ENOMEM, we assume this was
358 		 * a transient failure so we mark the page dirty so that
359 		 * we'll try to write it again later.  for all other write
360 		 * errors, we assume the error is permanent, thus the data
361 		 * in the page is lost.  bummer.
362 		 */
363 
364 		if (error) {
365 			int slot;
366 			if (!write) {
367 				pg->flags |= PG_RELEASED;
368 				continue;
369 			} else if (error == ENOMEM) {
370 				if (pg->flags & PG_PAGEOUT) {
371 					pg->flags &= ~PG_PAGEOUT;
372 					pageout_done++;
373 				}
374 				pg->flags &= ~PG_CLEAN;
375 				uvm_pageactivate(pg);
376 				slot = 0;
377 			} else
378 				slot = SWSLOT_BAD;
379 
380 #if defined(VMSWAP)
381 			if (swap) {
382 				if (pg->uobject != NULL) {
383 					int oldslot;
384 					oldslot = uao_set_swslot(pg->uobject,
385 						pg->offset >> PAGE_SHIFT, slot);
386 					KASSERT(oldslot == swslot + i);
387 				} else {
388 					KASSERT(pg->uanon->an_swslot ==
389 						swslot + i);
390 					pg->uanon->an_swslot = slot;
391 				}
392 			}
393 #endif /* defined(VMSWAP) */
394 		}
395 
396 		/*
397 		 * if the page is PG_FAKE, this must have been a read to
398 		 * initialize the page.  clear PG_FAKE and activate the page.
399 		 * we must also clear the pmap "modified" flag since it may
400 		 * still be set from the page's previous identity.
401 		 */
402 
403 		if (pg->flags & PG_FAKE) {
404 			KASSERT(!write);
405 			pg->flags &= ~PG_FAKE;
406 #if defined(READAHEAD_STATS)
407 			pg->pqflags |= PQ_READAHEAD;
408 			uvm_ra_total.ev_count++;
409 #endif /* defined(READAHEAD_STATS) */
410 			KASSERT((pg->flags & PG_CLEAN) != 0);
411 			uvm_pageenqueue(pg);
412 			pmap_clear_modify(pg);
413 		}
414 
415 		/*
416 		 * do accounting for pagedaemon i/o and arrange to free
417 		 * the pages instead of just unbusying them.
418 		 */
419 
420 		if (pg->flags & PG_PAGEOUT) {
421 			pg->flags &= ~PG_PAGEOUT;
422 			pageout_done++;
423 			uvmexp.pdfreed++;
424 			pg->flags |= PG_RELEASED;
425 		}
426 
427 #if defined(VMSWAP)
428 		/*
429 		 * for swap pages, unlock everything for this page now.
430 		 */
431 
432 		if (swap) {
433 			if (pg->uobject == NULL && pg->uanon->an_ref == 0 &&
434 			    (pg->flags & PG_RELEASED) != 0) {
435 				mutex_exit(&uvm_pageqlock);
436 				uvm_anon_release(pg->uanon);
437 			} else {
438 				uvm_page_unbusy(&pg, 1);
439 				mutex_exit(&uvm_pageqlock);
440 				mutex_exit(slock);
441 			}
442 		}
443 #endif /* defined(VMSWAP) */
444 	}
445 	uvm_pageout_done(pageout_done);
446 	if (!swap) {
447 		uvm_page_unbusy(pgs, npages);
448 		mutex_exit(&uvm_pageqlock);
449 		mutex_exit(slock);
450 	} else {
451 #if defined(VMSWAP)
452 		KASSERT(write);
453 
454 		/* these pages are now only in swap. */
455 		mutex_enter(&uvm_swap_data_lock);
456 		KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse);
457 		if (error != ENOMEM)
458 			uvmexp.swpgonly += npages;
459 		mutex_exit(&uvm_swap_data_lock);
460 		if (error) {
461 			if (error != ENOMEM)
462 				uvm_swap_markbad(swslot, npages);
463 			else
464 				uvm_swap_free(swslot, npages);
465 		}
466 		uvmexp.pdpending--;
467 #endif /* defined(VMSWAP) */
468 	}
469 }
470 
471 /*
472  * uvm_aio_aiodone: do iodone processing for async i/os.
473  * this should be called in thread context, not interrupt context.
474  */
475 
476 void
477 uvm_aio_aiodone(struct buf *bp)
478 {
479 	int npages = bp->b_bufsize >> PAGE_SHIFT;
480 	struct vm_page *pgs[npages];
481 	int i, error;
482 	bool write;
483 	UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist);
484 	UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0);
485 
486 	error = bp->b_error;
487 	write = (bp->b_flags & B_READ) == 0;
488 	/* XXXUBC BC_NOCACHE is for swap pager, should be done differently */
489 	if (write && !(bp->b_cflags & BC_NOCACHE) && bioopsp != NULL)
490 		(*bioopsp->io_pageiodone)(bp);
491 
492 	for (i = 0; i < npages; i++) {
493 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
494 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0);
495 	}
496 	uvm_pagermapout((vaddr_t)bp->b_data, npages);
497 
498 	uvm_aio_aiodone_pages(pgs, npages, write, error);
499 
500 	if (write && (bp->b_cflags & BC_AGE) != 0) {
501 		mutex_enter(bp->b_objlock);
502 		vwakeup(bp);
503 		mutex_exit(bp->b_objlock);
504 	}
505 	putiobuf(bp);
506 }
507 
508 /*
509  * uvm_pageratop: convert KVAs in the pager map back to their page
510  * structures.
511  */
512 
513 struct vm_page *
514 uvm_pageratop(vaddr_t kva)
515 {
516 	struct vm_page *pg;
517 	paddr_t pa;
518 	bool rv;
519 
520 	rv = pmap_extract(pmap_kernel(), kva, &pa);
521 	KASSERT(rv);
522 	pg = PHYS_TO_VM_PAGE(pa);
523 	KASSERT(pg != NULL);
524 	return (pg);
525 }
526