xref: /netbsd-src/sys/uvm/uvm_pager.c (revision d48f14661dda8638fee055ba15d35bdfb29b9fa8)
1 /*	$NetBSD: uvm_pager.c,v 1.77 2006/04/13 02:32:14 christos 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 1.77 2006/04/13 02:32:14 christos Exp $");
43 
44 #include "opt_uvmhist.h"
45 #include "opt_readahead.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/proc.h>
50 #include <sys/malloc.h>
51 #include <sys/pool.h>
52 #include <sys/vnode.h>
53 
54 #include <uvm/uvm.h>
55 
56 struct pool *uvm_aiobuf_pool;
57 
58 /*
59  * list of uvm pagers in the system
60  */
61 
62 struct uvm_pagerops * const uvmpagerops[] = {
63 	&aobj_pager,
64 	&uvm_deviceops,
65 	&uvm_vnodeops,
66 	&ubc_pager,
67 };
68 
69 /*
70  * the pager map: provides KVA for I/O
71  */
72 
73 struct vm_map *pager_map;		/* XXX */
74 struct simplelock pager_map_wanted_lock;
75 boolean_t pager_map_wanted;	/* locked by pager map */
76 static vaddr_t emergva;
77 static boolean_t emerginuse;
78 
79 /*
80  * uvm_pager_init: init pagers (at boot time)
81  */
82 
83 void
84 uvm_pager_init(void)
85 {
86 	u_int lcv;
87 	vaddr_t sva, eva;
88 
89 	/*
90 	 * init pager map
91 	 */
92 
93 	sva = 0;
94 	pager_map = uvm_km_suballoc(kernel_map, &sva, &eva, PAGER_MAP_SIZE, 0,
95 	    FALSE, NULL);
96 	simple_lock_init(&pager_map_wanted_lock);
97 	pager_map_wanted = FALSE;
98 	emergva = uvm_km_alloc(kernel_map, round_page(MAXPHYS), 0,
99 	    UVM_KMF_VAONLY);
100 #if defined(DEBUG)
101 	if (emergva == 0)
102 		panic("emergva");
103 #endif
104 	emerginuse = FALSE;
105 
106 	/*
107 	 * init ASYNC I/O queue
108 	 */
109 
110 	TAILQ_INIT(&uvm.aio_done);
111 
112 	/*
113 	 * call pager init functions
114 	 */
115 	for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
116 	    lcv++) {
117 		if (uvmpagerops[lcv]->pgo_init)
118 			uvmpagerops[lcv]->pgo_init();
119 	}
120 }
121 
122 /*
123  * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
124  *
125  * we basically just map in a blank map entry to reserve the space in the
126  * map and then use pmap_enter() to put the mappings in by hand.
127  */
128 
129 vaddr_t
130 uvm_pagermapin(struct vm_page **pps, int npages, int flags)
131 {
132 	vsize_t size;
133 	vaddr_t kva;
134 	vaddr_t cva;
135 	struct vm_page *pp;
136 	vm_prot_t prot;
137 	const boolean_t pdaemon = curproc == uvm.pagedaemon_proc;
138 	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
139 
140 	UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d)", pps, npages,0,0);
141 
142 	/*
143 	 * compute protection.  outgoing I/O only needs read
144 	 * access to the page, whereas incoming needs read/write.
145 	 */
146 
147 	prot = VM_PROT_READ;
148 	if (flags & UVMPAGER_MAPIN_READ)
149 		prot |= VM_PROT_WRITE;
150 
151 ReStart:
152 	size = npages << PAGE_SHIFT;
153 	kva = 0;			/* let system choose VA */
154 
155 	if (uvm_map(pager_map, &kva, size, NULL, UVM_UNKNOWN_OFFSET, 0,
156 	    UVM_FLAG_NOMERGE | (pdaemon ? UVM_FLAG_NOWAIT : 0)) != 0) {
157 		if (pdaemon) {
158 			simple_lock(&pager_map_wanted_lock);
159 			if (emerginuse) {
160 				UVM_UNLOCK_AND_WAIT(&emergva,
161 				    &pager_map_wanted_lock, FALSE,
162 				    "emergva", 0);
163 				goto ReStart;
164 			}
165 			emerginuse = TRUE;
166 			simple_unlock(&pager_map_wanted_lock);
167 			kva = emergva;
168 			/* The shift implicitly truncates to PAGE_SIZE */
169 			KASSERT(npages <= (MAXPHYS >> PAGE_SHIFT));
170 			goto enter;
171 		}
172 		if ((flags & UVMPAGER_MAPIN_WAITOK) == 0) {
173 			UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
174 			return(0);
175 		}
176 		simple_lock(&pager_map_wanted_lock);
177 		pager_map_wanted = TRUE;
178 		UVMHIST_LOG(maphist, "  SLEEPING on pager_map",0,0,0,0);
179 		UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, FALSE,
180 		    "pager_map", 0);
181 		goto ReStart;
182 	}
183 
184 enter:
185 	/* got it */
186 	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
187 		pp = *pps++;
188 		KASSERT(pp);
189 		KASSERT(pp->flags & PG_BUSY);
190 		pmap_kenter_pa(cva, VM_PAGE_TO_PHYS(pp), prot);
191 	}
192 	pmap_update(vm_map_pmap(pager_map));
193 
194 	UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
195 	return(kva);
196 }
197 
198 /*
199  * uvm_pagermapout: remove pager_map mapping
200  *
201  * we remove our mappings by hand and then remove the mapping (waking
202  * up anyone wanting space).
203  */
204 
205 void
206 uvm_pagermapout(vaddr_t kva, int npages)
207 {
208 	vsize_t size = npages << PAGE_SHIFT;
209 	struct vm_map_entry *entries;
210 	UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
211 
212 	UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
213 
214 	/*
215 	 * duplicate uvm_unmap, but add in pager_map_wanted handling.
216 	 */
217 
218 	pmap_kremove(kva, npages << PAGE_SHIFT);
219 	if (kva == emergva) {
220 		simple_lock(&pager_map_wanted_lock);
221 		emerginuse = FALSE;
222 		wakeup(&emergva);
223 		simple_unlock(&pager_map_wanted_lock);
224 		return;
225 	}
226 
227 	vm_map_lock(pager_map);
228 	uvm_unmap_remove(pager_map, kva, kva + size, &entries, NULL, 0);
229 	simple_lock(&pager_map_wanted_lock);
230 	if (pager_map_wanted) {
231 		pager_map_wanted = FALSE;
232 		wakeup(pager_map);
233 	}
234 	simple_unlock(&pager_map_wanted_lock);
235 	vm_map_unlock(pager_map);
236 	if (entries)
237 		uvm_unmap_detach(entries, 0);
238 	pmap_update(pmap_kernel());
239 	UVMHIST_LOG(maphist,"<- done",0,0,0,0);
240 }
241 
242 /*
243  * interrupt-context iodone handler for nested i/o bufs.
244  *
245  * => must be at splbio().
246  */
247 
248 void
249 uvm_aio_biodone1(struct buf *bp)
250 {
251 	struct buf *mbp = bp->b_private;
252 
253 	KASSERT(mbp != bp);
254 	if (bp->b_flags & B_ERROR) {
255 		mbp->b_flags |= B_ERROR;
256 		mbp->b_error = bp->b_error;
257 	}
258 	mbp->b_resid -= bp->b_bcount;
259 	putiobuf(bp);
260 	if (mbp->b_resid == 0) {
261 		biodone(mbp);
262 	}
263 }
264 
265 /*
266  * interrupt-context iodone handler for single-buf i/os
267  * or the top-level buf of a nested-buf i/o.
268  *
269  * => must be at splbio().
270  */
271 
272 void
273 uvm_aio_biodone(struct buf *bp)
274 {
275 	/* reset b_iodone for when this is a single-buf i/o. */
276 	bp->b_iodone = uvm_aio_aiodone;
277 
278 	simple_lock(&uvm.aiodoned_lock);	/* locks uvm.aio_done */
279 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
280 	wakeup(&uvm.aiodoned);
281 	simple_unlock(&uvm.aiodoned_lock);
282 }
283 
284 /*
285  * uvm_aio_aiodone: do iodone processing for async i/os.
286  * this should be called in thread context, not interrupt context.
287  */
288 
289 void
290 uvm_aio_aiodone(struct buf *bp)
291 {
292 	int npages = bp->b_bufsize >> PAGE_SHIFT;
293 	struct vm_page *pg, *pgs[npages];
294 	struct uvm_object *uobj;
295 	struct simplelock *slock;
296 	int s, i, error, swslot;
297 	boolean_t write, swap;
298 	UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist);
299 	UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0);
300 
301 	error = (bp->b_flags & B_ERROR) ? (bp->b_error ? bp->b_error : EIO) : 0;
302 	write = (bp->b_flags & B_READ) == 0;
303 	/* XXXUBC B_NOCACHE is for swap pager, should be done differently */
304 	if (write && !(bp->b_flags & B_NOCACHE) && bioops.io_pageiodone) {
305 		(*bioops.io_pageiodone)(bp);
306 	}
307 
308 	uobj = NULL;
309 	for (i = 0; i < npages; i++) {
310 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
311 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0);
312 	}
313 	uvm_pagermapout((vaddr_t)bp->b_data, npages);
314 
315 	swslot = 0;
316 	slock = NULL;
317 	pg = pgs[0];
318 	swap = (pg->uanon != NULL && pg->uobject == NULL) ||
319 		(pg->pqflags & PQ_AOBJ) != 0;
320 	if (!swap) {
321 		uobj = pg->uobject;
322 		slock = &uobj->vmobjlock;
323 		simple_lock(slock);
324 		uvm_lock_pageq();
325 	} else {
326 #if defined(VMSWAP)
327 		if (error) {
328 			if (pg->uobject != NULL) {
329 				swslot = uao_find_swslot(pg->uobject,
330 				    pg->offset >> PAGE_SHIFT);
331 			} else {
332 				KASSERT(pg->uanon != NULL);
333 				swslot = pg->uanon->an_swslot;
334 			}
335 			KASSERT(swslot);
336 		}
337 #else /* defined(VMSWAP) */
338 		panic("%s: swap", __func__);
339 #endif /* defined(VMSWAP) */
340 	}
341 	for (i = 0; i < npages; i++) {
342 		pg = pgs[i];
343 		KASSERT(swap || pg->uobject == uobj);
344 		UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0);
345 
346 #if defined(VMSWAP)
347 		/*
348 		 * for swap i/os, lock each page's object (or anon)
349 		 * individually since each page may need a different lock.
350 		 */
351 
352 		if (swap) {
353 			if (pg->uobject != NULL) {
354 				slock = &pg->uobject->vmobjlock;
355 			} else {
356 				slock = &pg->uanon->an_lock;
357 			}
358 			simple_lock(slock);
359 			uvm_lock_pageq();
360 		}
361 #endif /* defined(VMSWAP) */
362 
363 		/*
364 		 * process errors.  for reads, just mark the page to be freed.
365 		 * for writes, if the error was ENOMEM, we assume this was
366 		 * a transient failure so we mark the page dirty so that
367 		 * we'll try to write it again later.  for all other write
368 		 * errors, we assume the error is permanent, thus the data
369 		 * in the page is lost.  bummer.
370 		 */
371 
372 		if (error) {
373 			int slot;
374 			if (!write) {
375 				pg->flags |= PG_RELEASED;
376 				continue;
377 			} else if (error == ENOMEM) {
378 				if (pg->flags & PG_PAGEOUT) {
379 					pg->flags &= ~PG_PAGEOUT;
380 					uvmexp.paging--;
381 				}
382 				pg->flags &= ~PG_CLEAN;
383 				uvm_pageactivate(pg);
384 				slot = 0;
385 			} else
386 				slot = SWSLOT_BAD;
387 
388 #if defined(VMSWAP)
389 			if (swap) {
390 				if (pg->uobject != NULL) {
391 					int oldslot;
392 					oldslot = uao_set_swslot(pg->uobject,
393 						pg->offset >> PAGE_SHIFT, slot);
394 					KASSERT(oldslot == swslot + i);
395 				} else {
396 					KASSERT(pg->uanon->an_swslot ==
397 						swslot + i);
398 					pg->uanon->an_swslot = slot;
399 				}
400 			}
401 #endif /* defined(VMSWAP) */
402 		}
403 
404 		/*
405 		 * if the page is PG_FAKE, this must have been a read to
406 		 * initialize the page.  clear PG_FAKE and activate the page.
407 		 * we must also clear the pmap "modified" flag since it may
408 		 * still be set from the page's previous identity.
409 		 */
410 
411 		if (pg->flags & PG_FAKE) {
412 			KASSERT(!write);
413 			pg->flags &= ~PG_FAKE;
414 #if defined(READAHEAD_STATS)
415 			pg->flags |= PG_SPECULATIVE;
416 			uvm_ra_total.ev_count++;
417 #endif /* defined(READAHEAD_STATS) */
418 			uvm_pageactivate(pg);
419 			pmap_clear_modify(pg);
420 		}
421 
422 		/*
423 		 * do accounting for pagedaemon i/o and arrange to free
424 		 * the pages instead of just unbusying them.
425 		 */
426 
427 		if (pg->flags & PG_PAGEOUT) {
428 			pg->flags &= ~PG_PAGEOUT;
429 			uvmexp.paging--;
430 			uvmexp.pdfreed++;
431 			pg->flags |= PG_RELEASED;
432 		}
433 
434 #if defined(VMSWAP)
435 		/*
436 		 * for swap pages, unlock everything for this page now.
437 		 */
438 
439 		if (swap) {
440 			if (pg->uobject == NULL && pg->uanon->an_ref == 0 &&
441 			    (pg->flags & PG_RELEASED) != 0) {
442 				uvm_unlock_pageq();
443 				uvm_anon_release(pg->uanon);
444 			} else {
445 				uvm_page_unbusy(&pg, 1);
446 				uvm_unlock_pageq();
447 				simple_unlock(slock);
448 			}
449 		}
450 #endif /* defined(VMSWAP) */
451 	}
452 	if (!swap) {
453 		uvm_page_unbusy(pgs, npages);
454 		uvm_unlock_pageq();
455 		simple_unlock(slock);
456 	} else {
457 #if defined(VMSWAP)
458 		KASSERT(write);
459 
460 		/* these pages are now only in swap. */
461 		simple_lock(&uvm.swap_data_lock);
462 		KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse);
463 		if (error != ENOMEM)
464 			uvmexp.swpgonly += npages;
465 		simple_unlock(&uvm.swap_data_lock);
466 		if (error) {
467 			if (error != ENOMEM)
468 				uvm_swap_markbad(swslot, npages);
469 			else
470 				uvm_swap_free(swslot, npages);
471 		}
472 		uvmexp.pdpending--;
473 #endif /* defined(VMSWAP) */
474 	}
475 	s = splbio();
476 	if (write && (bp->b_flags & B_AGE) != 0) {
477 		vwakeup(bp);
478 	}
479 	putiobuf(bp);
480 	splx(s);
481 }
482 
483 /*
484  * uvm_pageratop: convert KVAs in the pager map back to their page
485  * structures.
486  */
487 
488 struct vm_page *
489 uvm_pageratop(vaddr_t kva)
490 {
491 	struct vm_page *pg;
492 	paddr_t pa;
493 	boolean_t rv;
494 
495 	rv = pmap_extract(pmap_kernel(), kva, &pa);
496 	KASSERT(rv);
497 	pg = PHYS_TO_VM_PAGE(pa);
498 	KASSERT(pg != NULL);
499 	return (pg);
500 }
501