xref: /netbsd-src/sys/uvm/uvm_pager.c (revision 7fa608457b817eca6e0977b37f758ae064f3c99c)
1 /*	$NetBSD: uvm_pager.c,v 1.87 2007/10/25 13:03:06 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 1.87 2007/10/25 13:03:06 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 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 < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
133 	    lcv++) {
134 		if (uvmpagerops[lcv]->pgo_init)
135 			uvmpagerops[lcv]->pgo_init();
136 	}
137 }
138 
139 /*
140  * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
141  *
142  * we basically just map in a blank map entry to reserve the space in the
143  * map and then use pmap_enter() to put the mappings in by hand.
144  */
145 
146 vaddr_t
147 uvm_pagermapin(struct vm_page **pps, int npages, int flags)
148 {
149 	vsize_t size;
150 	vaddr_t kva;
151 	vaddr_t cva;
152 	struct vm_page *pp;
153 	vm_prot_t prot;
154 	const bool pdaemon = curlwp == uvm.pagedaemon_lwp;
155 	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
156 
157 	UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d)", pps, npages,0,0);
158 
159 	/*
160 	 * compute protection.  outgoing I/O only needs read
161 	 * access to the page, whereas incoming needs read/write.
162 	 */
163 
164 	prot = VM_PROT_READ;
165 	if (flags & UVMPAGER_MAPIN_READ)
166 		prot |= VM_PROT_WRITE;
167 
168 ReStart:
169 	size = npages << PAGE_SHIFT;
170 	kva = 0;			/* let system choose VA */
171 
172 	if (uvm_map(pager_map, &kva, size, NULL, UVM_UNKNOWN_OFFSET, 0,
173 	    UVM_FLAG_NOMERGE | (pdaemon ? UVM_FLAG_NOWAIT : 0)) != 0) {
174 		if (pdaemon) {
175 			mutex_enter(&pager_map_wanted_lock);
176 			if (emerginuse) {
177 				mtsleep(&emergva, PVM | PNORELOCK, "emergva",
178 				    0, &pager_map_wanted_lock);
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 		mtsleep(pager_map, PVM | PNORELOCK, "pager_map", 0,
196 		    &pager_map_wanted_lock);
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  * => must be at splbio().
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 	mbp->b_resid -= bp->b_bcount;
273 	putiobuf(bp);
274 	if (mbp->b_resid == 0) {
275 		biodone(mbp);
276 	}
277 }
278 
279 /*
280  * interrupt-context iodone handler for single-buf i/os
281  * or the top-level buf of a nested-buf i/o.
282  *
283  * => must be at splbio().
284  */
285 
286 void
287 uvm_aio_biodone(struct buf *bp)
288 {
289 	/* reset b_iodone for when this is a single-buf i/o. */
290 	bp->b_iodone = uvm_aio_aiodone;
291 
292 	workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL);
293 }
294 
295 /*
296  * uvm_aio_aiodone: do iodone processing for async i/os.
297  * this should be called in thread context, not interrupt context.
298  */
299 
300 void
301 uvm_aio_aiodone(struct buf *bp)
302 {
303 	int npages = bp->b_bufsize >> PAGE_SHIFT;
304 	struct vm_page *pg, *pgs[npages];
305 	struct uvm_object *uobj;
306 	struct simplelock *slock;
307 	int s, i, error, swslot;
308 	bool write, swap;
309 	UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist);
310 	UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0);
311 
312 	error = bp->b_error;
313 	write = (bp->b_flags & B_READ) == 0;
314 	/* XXXUBC B_NOCACHE is for swap pager, should be done differently */
315 	if (write && !(bp->b_flags & B_NOCACHE) && bioopsp) {
316 		bioopsp->io_pageiodone(bp);
317 	}
318 
319 	uobj = NULL;
320 	for (i = 0; i < npages; i++) {
321 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
322 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0);
323 	}
324 	uvm_pagermapout((vaddr_t)bp->b_data, npages);
325 
326 	swslot = 0;
327 	slock = NULL;
328 	pg = pgs[0];
329 	swap = (pg->uanon != NULL && pg->uobject == NULL) ||
330 		(pg->pqflags & PQ_AOBJ) != 0;
331 	if (!swap) {
332 		uobj = pg->uobject;
333 		slock = &uobj->vmobjlock;
334 		simple_lock(slock);
335 		uvm_lock_pageq();
336 	} else {
337 #if defined(VMSWAP)
338 		if (error) {
339 			if (pg->uobject != NULL) {
340 				swslot = uao_find_swslot(pg->uobject,
341 				    pg->offset >> PAGE_SHIFT);
342 			} else {
343 				KASSERT(pg->uanon != NULL);
344 				swslot = pg->uanon->an_swslot;
345 			}
346 			KASSERT(swslot);
347 		}
348 #else /* defined(VMSWAP) */
349 		panic("%s: swap", __func__);
350 #endif /* defined(VMSWAP) */
351 	}
352 	for (i = 0; i < npages; i++) {
353 		pg = pgs[i];
354 		KASSERT(swap || pg->uobject == uobj);
355 		UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0);
356 
357 #if defined(VMSWAP)
358 		/*
359 		 * for swap i/os, lock each page's object (or anon)
360 		 * individually since each page may need a different lock.
361 		 */
362 
363 		if (swap) {
364 			if (pg->uobject != NULL) {
365 				slock = &pg->uobject->vmobjlock;
366 			} else {
367 				slock = &pg->uanon->an_lock;
368 			}
369 			simple_lock(slock);
370 			uvm_lock_pageq();
371 		}
372 #endif /* defined(VMSWAP) */
373 
374 		/*
375 		 * process errors.  for reads, just mark the page to be freed.
376 		 * for writes, if the error was ENOMEM, we assume this was
377 		 * a transient failure so we mark the page dirty so that
378 		 * we'll try to write it again later.  for all other write
379 		 * errors, we assume the error is permanent, thus the data
380 		 * in the page is lost.  bummer.
381 		 */
382 
383 		if (error) {
384 			int slot;
385 			if (!write) {
386 				pg->flags |= PG_RELEASED;
387 				continue;
388 			} else if (error == ENOMEM) {
389 				if (pg->flags & PG_PAGEOUT) {
390 					pg->flags &= ~PG_PAGEOUT;
391 					uvmexp.paging--;
392 				}
393 				pg->flags &= ~PG_CLEAN;
394 				uvm_pageactivate(pg);
395 				slot = 0;
396 			} else
397 				slot = SWSLOT_BAD;
398 
399 #if defined(VMSWAP)
400 			if (swap) {
401 				if (pg->uobject != NULL) {
402 					int oldslot;
403 					oldslot = uao_set_swslot(pg->uobject,
404 						pg->offset >> PAGE_SHIFT, slot);
405 					KASSERT(oldslot == swslot + i);
406 				} else {
407 					KASSERT(pg->uanon->an_swslot ==
408 						swslot + i);
409 					pg->uanon->an_swslot = slot;
410 				}
411 			}
412 #endif /* defined(VMSWAP) */
413 		}
414 
415 		/*
416 		 * if the page is PG_FAKE, this must have been a read to
417 		 * initialize the page.  clear PG_FAKE and activate the page.
418 		 * we must also clear the pmap "modified" flag since it may
419 		 * still be set from the page's previous identity.
420 		 */
421 
422 		if (pg->flags & PG_FAKE) {
423 			KASSERT(!write);
424 			pg->flags &= ~PG_FAKE;
425 #if defined(READAHEAD_STATS)
426 			pg->pqflags |= PQ_READAHEAD;
427 			uvm_ra_total.ev_count++;
428 #endif /* defined(READAHEAD_STATS) */
429 			KASSERT((pg->flags & PG_CLEAN) != 0);
430 			uvm_pageenqueue(pg);
431 			pmap_clear_modify(pg);
432 		}
433 
434 		/*
435 		 * do accounting for pagedaemon i/o and arrange to free
436 		 * the pages instead of just unbusying them.
437 		 */
438 
439 		if (pg->flags & PG_PAGEOUT) {
440 			pg->flags &= ~PG_PAGEOUT;
441 			uvmexp.paging--;
442 			uvmexp.pdfreed++;
443 			pg->flags |= PG_RELEASED;
444 		}
445 
446 #if defined(VMSWAP)
447 		/*
448 		 * for swap pages, unlock everything for this page now.
449 		 */
450 
451 		if (swap) {
452 			if (pg->uobject == NULL && pg->uanon->an_ref == 0 &&
453 			    (pg->flags & PG_RELEASED) != 0) {
454 				uvm_unlock_pageq();
455 				uvm_anon_release(pg->uanon);
456 			} else {
457 				uvm_page_unbusy(&pg, 1);
458 				uvm_unlock_pageq();
459 				simple_unlock(slock);
460 			}
461 		}
462 #endif /* defined(VMSWAP) */
463 	}
464 	if (!swap) {
465 		uvm_page_unbusy(pgs, npages);
466 		uvm_unlock_pageq();
467 		simple_unlock(slock);
468 	} else {
469 #if defined(VMSWAP)
470 		KASSERT(write);
471 
472 		/* these pages are now only in swap. */
473 		mutex_enter(&uvm_swap_data_lock);
474 		KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse);
475 		if (error != ENOMEM)
476 			uvmexp.swpgonly += npages;
477 		mutex_exit(&uvm_swap_data_lock);
478 		if (error) {
479 			if (error != ENOMEM)
480 				uvm_swap_markbad(swslot, npages);
481 			else
482 				uvm_swap_free(swslot, npages);
483 		}
484 		uvmexp.pdpending--;
485 #endif /* defined(VMSWAP) */
486 	}
487 	s = splbio();
488 	if (write && (bp->b_flags & B_AGE) != 0) {
489 		vwakeup(bp);
490 	}
491 	putiobuf(bp);
492 	splx(s);
493 }
494 
495 /*
496  * uvm_pageratop: convert KVAs in the pager map back to their page
497  * structures.
498  */
499 
500 struct vm_page *
501 uvm_pageratop(vaddr_t kva)
502 {
503 	struct vm_page *pg;
504 	paddr_t pa;
505 	bool rv;
506 
507 	rv = pmap_extract(pmap_kernel(), kva, &pa);
508 	KASSERT(rv);
509 	pg = PHYS_TO_VM_PAGE(pa);
510 	KASSERT(pg != NULL);
511 	return (pg);
512 }
513