uvm_pager.c revision 1.33 1 1.33 thorpej /* $NetBSD: uvm_pager.c,v 1.33 2000/09/13 15:00:25 thorpej Exp $ */
2 1.1 mrg
3 1.1 mrg /*
4 1.1 mrg *
5 1.1 mrg * Copyright (c) 1997 Charles D. Cranor and Washington University.
6 1.1 mrg * All rights reserved.
7 1.1 mrg *
8 1.1 mrg * Redistribution and use in source and binary forms, with or without
9 1.1 mrg * modification, are permitted provided that the following conditions
10 1.1 mrg * are met:
11 1.1 mrg * 1. Redistributions of source code must retain the above copyright
12 1.1 mrg * notice, this list of conditions and the following disclaimer.
13 1.1 mrg * 2. Redistributions in binary form must reproduce the above copyright
14 1.1 mrg * notice, this list of conditions and the following disclaimer in the
15 1.1 mrg * documentation and/or other materials provided with the distribution.
16 1.1 mrg * 3. All advertising materials mentioning features or use of this software
17 1.1 mrg * must display the following acknowledgement:
18 1.1 mrg * This product includes software developed by Charles D. Cranor and
19 1.1 mrg * Washington University.
20 1.1 mrg * 4. The name of the author may not be used to endorse or promote products
21 1.1 mrg * derived from this software without specific prior written permission.
22 1.1 mrg *
23 1.1 mrg * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24 1.1 mrg * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25 1.1 mrg * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26 1.1 mrg * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27 1.1 mrg * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28 1.1 mrg * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29 1.1 mrg * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30 1.1 mrg * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 1.1 mrg * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32 1.1 mrg * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 1.3 mrg *
34 1.3 mrg * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
35 1.1 mrg */
36 1.1 mrg
37 1.5 mrg #include "opt_uvmhist.h"
38 1.5 mrg
39 1.1 mrg /*
40 1.1 mrg * uvm_pager.c: generic functions used to assist the pagers.
41 1.1 mrg */
42 1.1 mrg
43 1.1 mrg #include <sys/param.h>
44 1.1 mrg #include <sys/systm.h>
45 1.1 mrg #include <sys/proc.h>
46 1.1 mrg #include <sys/malloc.h>
47 1.1 mrg
48 1.1 mrg #define UVM_PAGER
49 1.1 mrg #include <uvm/uvm.h>
50 1.1 mrg
51 1.1 mrg /*
52 1.1 mrg * list of uvm pagers in the system
53 1.1 mrg */
54 1.1 mrg
55 1.1 mrg extern struct uvm_pagerops uvm_deviceops;
56 1.1 mrg extern struct uvm_pagerops uvm_vnodeops;
57 1.1 mrg
58 1.1 mrg struct uvm_pagerops *uvmpagerops[] = {
59 1.10 thorpej &aobj_pager,
60 1.6 mrg &uvm_deviceops,
61 1.6 mrg &uvm_vnodeops,
62 1.1 mrg };
63 1.1 mrg
64 1.1 mrg /*
65 1.1 mrg * the pager map: provides KVA for I/O
66 1.1 mrg */
67 1.1 mrg
68 1.1 mrg #define PAGER_MAP_SIZE (4 * 1024 * 1024)
69 1.1 mrg vm_map_t pager_map; /* XXX */
70 1.1 mrg simple_lock_data_t pager_map_wanted_lock;
71 1.1 mrg boolean_t pager_map_wanted; /* locked by pager map */
72 1.1 mrg
73 1.1 mrg
74 1.1 mrg /*
75 1.1 mrg * uvm_pager_init: init pagers (at boot time)
76 1.1 mrg */
77 1.1 mrg
78 1.6 mrg void
79 1.6 mrg uvm_pager_init()
80 1.6 mrg {
81 1.6 mrg int lcv;
82 1.1 mrg
83 1.6 mrg /*
84 1.6 mrg * init pager map
85 1.6 mrg */
86 1.6 mrg
87 1.6 mrg pager_map = uvm_km_suballoc(kernel_map, &uvm.pager_sva, &uvm.pager_eva,
88 1.20 thorpej PAGER_MAP_SIZE, 0, FALSE, NULL);
89 1.6 mrg simple_lock_init(&pager_map_wanted_lock);
90 1.6 mrg pager_map_wanted = FALSE;
91 1.6 mrg
92 1.6 mrg /*
93 1.6 mrg * init ASYNC I/O queue
94 1.6 mrg */
95 1.6 mrg
96 1.6 mrg TAILQ_INIT(&uvm.aio_done);
97 1.1 mrg
98 1.6 mrg /*
99 1.6 mrg * call pager init functions
100 1.6 mrg */
101 1.6 mrg for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
102 1.6 mrg lcv++) {
103 1.6 mrg if (uvmpagerops[lcv]->pgo_init)
104 1.6 mrg uvmpagerops[lcv]->pgo_init();
105 1.6 mrg }
106 1.1 mrg }
107 1.1 mrg
108 1.1 mrg /*
109 1.1 mrg * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
110 1.1 mrg *
111 1.1 mrg * we basically just map in a blank map entry to reserve the space in the
112 1.1 mrg * map and then use pmap_enter() to put the mappings in by hand.
113 1.1 mrg */
114 1.1 mrg
115 1.9 eeh vaddr_t
116 1.29 thorpej uvm_pagermapin(pps, npages, aiop, flags)
117 1.6 mrg struct vm_page **pps;
118 1.6 mrg int npages;
119 1.6 mrg struct uvm_aiodesc **aiop; /* OUT */
120 1.29 thorpej int flags;
121 1.1 mrg {
122 1.9 eeh vsize_t size;
123 1.9 eeh vaddr_t kva;
124 1.6 mrg struct uvm_aiodesc *aio;
125 1.9 eeh vaddr_t cva;
126 1.6 mrg struct vm_page *pp;
127 1.29 thorpej vm_prot_t prot;
128 1.6 mrg UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
129 1.1 mrg
130 1.29 thorpej UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d, aiop=0x%x, flags=0x%x)",
131 1.29 thorpej pps, npages, aiop, flags);
132 1.29 thorpej
133 1.29 thorpej /*
134 1.29 thorpej * compute protection. outgoing I/O only needs read
135 1.29 thorpej * access to the page, whereas incoming needs read/write.
136 1.29 thorpej */
137 1.29 thorpej
138 1.29 thorpej prot = VM_PROT_READ;
139 1.29 thorpej if (flags & UVMPAGER_MAPIN_READ)
140 1.29 thorpej prot |= VM_PROT_WRITE;
141 1.1 mrg
142 1.1 mrg ReStart:
143 1.6 mrg if (aiop) {
144 1.29 thorpej MALLOC(aio, struct uvm_aiodesc *, sizeof(*aio), M_TEMP,
145 1.29 thorpej (flags & UVMPAGER_MAPIN_WAITOK));
146 1.6 mrg if (aio == NULL)
147 1.6 mrg return(0);
148 1.6 mrg *aiop = aio;
149 1.6 mrg } else {
150 1.6 mrg aio = NULL;
151 1.6 mrg }
152 1.1 mrg
153 1.12 chs size = npages << PAGE_SHIFT;
154 1.29 thorpej kva = 0; /* let system choose VA */
155 1.1 mrg
156 1.6 mrg if (uvm_map(pager_map, &kva, size, NULL,
157 1.33 thorpej UVM_UNKNOWN_OFFSET, 0, UVM_FLAG_NOMERGE) != KERN_SUCCESS) {
158 1.29 thorpej if ((flags & UVMPAGER_MAPIN_WAITOK) == 0) {
159 1.6 mrg if (aio)
160 1.6 mrg FREE(aio, M_TEMP);
161 1.6 mrg UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
162 1.29 thorpej return(0);
163 1.6 mrg }
164 1.6 mrg simple_lock(&pager_map_wanted_lock);
165 1.6 mrg pager_map_wanted = TRUE;
166 1.6 mrg UVMHIST_LOG(maphist, " SLEEPING on pager_map",0,0,0,0);
167 1.6 mrg UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, FALSE,
168 1.6 mrg "pager_map",0);
169 1.6 mrg goto ReStart;
170 1.6 mrg }
171 1.1 mrg
172 1.6 mrg /* got it */
173 1.6 mrg for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
174 1.6 mrg pp = *pps++;
175 1.1 mrg #ifdef DEBUG
176 1.6 mrg if ((pp->flags & PG_BUSY) == 0)
177 1.6 mrg panic("uvm_pagermapin: page not busy");
178 1.1 mrg #endif
179 1.6 mrg pmap_enter(vm_map_pmap(pager_map), cva, VM_PAGE_TO_PHYS(pp),
180 1.30 thorpej prot, PMAP_WIRED | prot);
181 1.6 mrg }
182 1.1 mrg
183 1.6 mrg UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
184 1.6 mrg return(kva);
185 1.1 mrg }
186 1.1 mrg
187 1.1 mrg /*
188 1.1 mrg * uvm_pagermapout: remove pager_map mapping
189 1.1 mrg *
190 1.1 mrg * we remove our mappings by hand and then remove the mapping (waking
191 1.1 mrg * up anyone wanting space).
192 1.1 mrg */
193 1.1 mrg
194 1.6 mrg void
195 1.6 mrg uvm_pagermapout(kva, npages)
196 1.9 eeh vaddr_t kva;
197 1.6 mrg int npages;
198 1.6 mrg {
199 1.12 chs vsize_t size = npages << PAGE_SHIFT;
200 1.6 mrg vm_map_entry_t entries;
201 1.6 mrg UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
202 1.6 mrg
203 1.6 mrg UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
204 1.1 mrg
205 1.6 mrg /*
206 1.6 mrg * duplicate uvm_unmap, but add in pager_map_wanted handling.
207 1.6 mrg */
208 1.6 mrg
209 1.6 mrg vm_map_lock(pager_map);
210 1.11 chuck (void) uvm_unmap_remove(pager_map, kva, kva + size, &entries);
211 1.6 mrg simple_lock(&pager_map_wanted_lock);
212 1.6 mrg if (pager_map_wanted) {
213 1.6 mrg pager_map_wanted = FALSE;
214 1.6 mrg wakeup(pager_map);
215 1.6 mrg }
216 1.6 mrg simple_unlock(&pager_map_wanted_lock);
217 1.6 mrg vm_map_unlock(pager_map);
218 1.6 mrg if (entries)
219 1.6 mrg uvm_unmap_detach(entries, 0);
220 1.1 mrg
221 1.6 mrg UVMHIST_LOG(maphist,"<- done",0,0,0,0);
222 1.1 mrg }
223 1.1 mrg
224 1.1 mrg /*
225 1.1 mrg * uvm_mk_pcluster
226 1.1 mrg *
227 1.1 mrg * generic "make 'pager put' cluster" function. a pager can either
228 1.1 mrg * [1] set pgo_mk_pcluster to NULL (never cluster), [2] set it to this
229 1.1 mrg * generic function, or [3] set it to a pager specific function.
230 1.1 mrg *
231 1.1 mrg * => caller must lock object _and_ pagequeues (since we need to look
232 1.1 mrg * at active vs. inactive bits, etc.)
233 1.1 mrg * => caller must make center page busy and write-protect it
234 1.1 mrg * => we mark all cluster pages busy for the caller
235 1.1 mrg * => the caller must unbusy all pages (and check wanted/released
236 1.1 mrg * status if it drops the object lock)
237 1.1 mrg * => flags:
238 1.1 mrg * PGO_ALLPAGES: all pages in object are valid targets
239 1.1 mrg * !PGO_ALLPAGES: use "lo" and "hi" to limit range of cluster
240 1.1 mrg * PGO_DOACTCLUST: include active pages in cluster.
241 1.1 mrg * NOTE: the caller should clear PG_CLEANCHK bits if PGO_DOACTCLUST.
242 1.1 mrg * PG_CLEANCHK is only a hint, but clearing will help reduce
243 1.1 mrg * the number of calls we make to the pmap layer.
244 1.1 mrg */
245 1.1 mrg
246 1.6 mrg struct vm_page **
247 1.6 mrg uvm_mk_pcluster(uobj, pps, npages, center, flags, mlo, mhi)
248 1.6 mrg struct uvm_object *uobj; /* IN */
249 1.6 mrg struct vm_page **pps, *center; /* IN/OUT, IN */
250 1.6 mrg int *npages, flags; /* IN/OUT, IN */
251 1.26 kleink voff_t mlo, mhi; /* IN (if !PGO_ALLPAGES) */
252 1.1 mrg {
253 1.6 mrg struct vm_page **ppsp, *pclust;
254 1.26 kleink voff_t lo, hi, curoff;
255 1.6 mrg int center_idx, forward;
256 1.6 mrg UVMHIST_FUNC("uvm_mk_pcluster"); UVMHIST_CALLED(maphist);
257 1.6 mrg
258 1.6 mrg /*
259 1.6 mrg * center page should already be busy and write protected. XXX:
260 1.6 mrg * suppose page is wired? if we lock, then a process could
261 1.6 mrg * fault/block on it. if we don't lock, a process could write the
262 1.6 mrg * pages in the middle of an I/O. (consider an msync()). let's
263 1.6 mrg * lock it for now (better to delay than corrupt data?).
264 1.6 mrg */
265 1.6 mrg
266 1.6 mrg /*
267 1.6 mrg * get cluster boundaries, check sanity, and apply our limits as well.
268 1.6 mrg */
269 1.6 mrg
270 1.6 mrg uobj->pgops->pgo_cluster(uobj, center->offset, &lo, &hi);
271 1.6 mrg if ((flags & PGO_ALLPAGES) == 0) {
272 1.6 mrg if (lo < mlo)
273 1.6 mrg lo = mlo;
274 1.6 mrg if (hi > mhi)
275 1.6 mrg hi = mhi;
276 1.6 mrg }
277 1.12 chs if ((hi - lo) >> PAGE_SHIFT > *npages) { /* pps too small, bail out! */
278 1.1 mrg #ifdef DIAGNOSTIC
279 1.6 mrg printf("uvm_mk_pcluster: provided page array too small (fixed)\n");
280 1.1 mrg #endif
281 1.6 mrg pps[0] = center;
282 1.6 mrg *npages = 1;
283 1.6 mrg return(pps);
284 1.6 mrg }
285 1.6 mrg
286 1.6 mrg /*
287 1.6 mrg * now determine the center and attempt to cluster around the
288 1.6 mrg * edges
289 1.6 mrg */
290 1.6 mrg
291 1.12 chs center_idx = (center->offset - lo) >> PAGE_SHIFT;
292 1.6 mrg pps[center_idx] = center; /* plug in the center page */
293 1.6 mrg ppsp = &pps[center_idx];
294 1.6 mrg *npages = 1;
295 1.6 mrg
296 1.6 mrg /*
297 1.6 mrg * attempt to cluster around the left [backward], and then
298 1.6 mrg * the right side [forward].
299 1.6 mrg *
300 1.6 mrg * note that for inactive pages (pages that have been deactivated)
301 1.6 mrg * there are no valid mappings and PG_CLEAN should be up to date.
302 1.6 mrg * [i.e. there is no need to query the pmap with pmap_is_modified
303 1.6 mrg * since there are no mappings].
304 1.6 mrg */
305 1.6 mrg
306 1.6 mrg for (forward = 0 ; forward <= 1 ; forward++) {
307 1.6 mrg
308 1.14 chs curoff = center->offset + (forward ? PAGE_SIZE : -PAGE_SIZE);
309 1.6 mrg for ( ;(forward == 0 && curoff >= lo) ||
310 1.12 chs (forward && curoff < hi);
311 1.12 chs curoff += (forward ? 1 : -1) << PAGE_SHIFT) {
312 1.6 mrg
313 1.6 mrg pclust = uvm_pagelookup(uobj, curoff); /* lookup page */
314 1.6 mrg if (pclust == NULL)
315 1.6 mrg break; /* no page */
316 1.6 mrg /* handle active pages */
317 1.6 mrg /* NOTE: inactive pages don't have pmap mappings */
318 1.6 mrg if ((pclust->pqflags & PQ_INACTIVE) == 0) {
319 1.6 mrg if ((flags & PGO_DOACTCLUST) == 0)
320 1.6 mrg /* dont want mapped pages at all */
321 1.6 mrg break;
322 1.6 mrg
323 1.6 mrg /* make sure "clean" bit is sync'd */
324 1.6 mrg if ((pclust->flags & PG_CLEANCHK) == 0) {
325 1.6 mrg if ((pclust->flags & (PG_CLEAN|PG_BUSY))
326 1.6 mrg == PG_CLEAN &&
327 1.23 chs pmap_is_modified(pclust))
328 1.23 chs pclust->flags &= ~PG_CLEAN;
329 1.23 chs
330 1.6 mrg /* now checked */
331 1.6 mrg pclust->flags |= PG_CLEANCHK;
332 1.6 mrg }
333 1.6 mrg }
334 1.6 mrg /* is page available for cleaning and does it need it */
335 1.6 mrg if ((pclust->flags & (PG_CLEAN|PG_BUSY)) != 0)
336 1.6 mrg break; /* page is already clean or is busy */
337 1.6 mrg
338 1.6 mrg /* yes! enroll the page in our array */
339 1.6 mrg pclust->flags |= PG_BUSY; /* busy! */
340 1.6 mrg UVM_PAGE_OWN(pclust, "uvm_mk_pcluster");
341 1.6 mrg /* XXX: protect wired page? see above comment. */
342 1.23 chs pmap_page_protect(pclust, VM_PROT_READ);
343 1.6 mrg if (!forward) {
344 1.6 mrg ppsp--; /* back up one page */
345 1.6 mrg *ppsp = pclust;
346 1.6 mrg } else {
347 1.6 mrg /* move forward one page */
348 1.6 mrg ppsp[*npages] = pclust;
349 1.6 mrg }
350 1.6 mrg *npages = *npages + 1;
351 1.6 mrg }
352 1.6 mrg }
353 1.6 mrg
354 1.6 mrg /*
355 1.6 mrg * done! return the cluster array to the caller!!!
356 1.6 mrg */
357 1.1 mrg
358 1.6 mrg UVMHIST_LOG(maphist, "<- done",0,0,0,0);
359 1.6 mrg return(ppsp);
360 1.1 mrg }
361 1.1 mrg
362 1.1 mrg /*
363 1.1 mrg * uvm_pager_put: high level pageout routine
364 1.1 mrg *
365 1.1 mrg * we want to pageout page "pg" to backing store, clustering if
366 1.1 mrg * possible.
367 1.1 mrg *
368 1.1 mrg * => page queues must be locked by caller
369 1.1 mrg * => if page is not swap-backed, then "uobj" points to the object
370 1.1 mrg * backing it. this object should be locked by the caller.
371 1.1 mrg * => if page is swap-backed, then "uobj" should be NULL.
372 1.1 mrg * => "pg" should be PG_BUSY (by caller), and !PG_CLEAN
373 1.1 mrg * for swap-backed memory, "pg" can be NULL if there is no page
374 1.1 mrg * of interest [sometimes the case for the pagedaemon]
375 1.1 mrg * => "ppsp_ptr" should point to an array of npages vm_page pointers
376 1.1 mrg * for possible cluster building
377 1.1 mrg * => flags (first two for non-swap-backed pages)
378 1.1 mrg * PGO_ALLPAGES: all pages in uobj are valid targets
379 1.1 mrg * PGO_DOACTCLUST: include "PQ_ACTIVE" pages as valid targets
380 1.1 mrg * PGO_SYNCIO: do SYNC I/O (no async)
381 1.1 mrg * PGO_PDFREECLUST: pagedaemon: drop cluster on successful I/O
382 1.1 mrg * => start/stop: if (uobj && !PGO_ALLPAGES) limit targets to this range
383 1.1 mrg * if (!uobj) start is the (daddr_t) of the starting swapblk
384 1.1 mrg * => return state:
385 1.1 mrg * 1. we return the VM_PAGER status code of the pageout
386 1.1 mrg * 2. we return with the page queues unlocked
387 1.1 mrg * 3. if (uobj != NULL) [!swap_backed] we return with
388 1.1 mrg * uobj locked _only_ if PGO_PDFREECLUST is set
389 1.1 mrg * AND result != VM_PAGER_PEND. in all other cases
390 1.1 mrg * we return with uobj unlocked. [this is a hack
391 1.1 mrg * that allows the pagedaemon to save one lock/unlock
392 1.1 mrg * pair in the !swap_backed case since we have to
393 1.1 mrg * lock the uobj to drop the cluster anyway]
394 1.1 mrg * 4. on errors we always drop the cluster. thus, if we return
395 1.1 mrg * !PEND, !OK, then the caller only has to worry about
396 1.1 mrg * un-busying the main page (not the cluster pages).
397 1.1 mrg * 5. on success, if !PGO_PDFREECLUST, we return the cluster
398 1.1 mrg * with all pages busy (caller must un-busy and check
399 1.1 mrg * wanted/released flags).
400 1.1 mrg */
401 1.1 mrg
402 1.6 mrg int
403 1.6 mrg uvm_pager_put(uobj, pg, ppsp_ptr, npages, flags, start, stop)
404 1.6 mrg struct uvm_object *uobj; /* IN */
405 1.6 mrg struct vm_page *pg, ***ppsp_ptr;/* IN, IN/OUT */
406 1.6 mrg int *npages; /* IN/OUT */
407 1.6 mrg int flags; /* IN */
408 1.26 kleink voff_t start, stop; /* IN, IN */
409 1.6 mrg {
410 1.6 mrg int result;
411 1.6 mrg daddr_t swblk;
412 1.6 mrg struct vm_page **ppsp = *ppsp_ptr;
413 1.6 mrg
414 1.6 mrg /*
415 1.6 mrg * note that uobj is null if we are doing a swap-backed pageout.
416 1.6 mrg * note that uobj is !null if we are doing normal object pageout.
417 1.6 mrg * note that the page queues must be locked to cluster.
418 1.6 mrg */
419 1.6 mrg
420 1.6 mrg if (uobj) { /* if !swap-backed */
421 1.6 mrg
422 1.6 mrg /*
423 1.6 mrg * attempt to build a cluster for pageout using its
424 1.6 mrg * make-put-cluster function (if it has one).
425 1.6 mrg */
426 1.6 mrg
427 1.6 mrg if (uobj->pgops->pgo_mk_pcluster) {
428 1.6 mrg ppsp = uobj->pgops->pgo_mk_pcluster(uobj, ppsp,
429 1.6 mrg npages, pg, flags, start, stop);
430 1.6 mrg *ppsp_ptr = ppsp; /* update caller's pointer */
431 1.6 mrg } else {
432 1.6 mrg ppsp[0] = pg;
433 1.6 mrg *npages = 1;
434 1.25 chs }
435 1.25 chs
436 1.6 mrg swblk = 0; /* XXX: keep gcc happy */
437 1.1 mrg
438 1.6 mrg } else {
439 1.1 mrg
440 1.6 mrg /*
441 1.6 mrg * for swap-backed pageout, the caller (the pagedaemon) has
442 1.6 mrg * already built the cluster for us. the starting swap
443 1.6 mrg * block we are writing to has been passed in as "start."
444 1.6 mrg * "pg" could be NULL if there is no page we are especially
445 1.6 mrg * interested in (in which case the whole cluster gets dropped
446 1.6 mrg * in the event of an error or a sync "done").
447 1.6 mrg */
448 1.6 mrg swblk = (daddr_t) start;
449 1.6 mrg /* ppsp and npages should be ok */
450 1.6 mrg }
451 1.1 mrg
452 1.6 mrg /* now that we've clustered we can unlock the page queues */
453 1.6 mrg uvm_unlock_pageq();
454 1.1 mrg
455 1.6 mrg /*
456 1.6 mrg * now attempt the I/O. if we have a failure and we are
457 1.6 mrg * clustered, we will drop the cluster and try again.
458 1.6 mrg */
459 1.1 mrg
460 1.1 mrg ReTry:
461 1.6 mrg if (uobj) {
462 1.6 mrg /* object is locked */
463 1.6 mrg result = uobj->pgops->pgo_put(uobj, ppsp, *npages,
464 1.6 mrg flags & PGO_SYNCIO);
465 1.6 mrg /* object is now unlocked */
466 1.6 mrg } else {
467 1.6 mrg /* nothing locked */
468 1.6 mrg result = uvm_swap_put(swblk, ppsp, *npages, flags & PGO_SYNCIO);
469 1.6 mrg /* nothing locked */
470 1.6 mrg }
471 1.6 mrg
472 1.6 mrg /*
473 1.6 mrg * we have attempted the I/O.
474 1.6 mrg *
475 1.6 mrg * if the I/O was a success then:
476 1.6 mrg * if !PGO_PDFREECLUST, we return the cluster to the
477 1.6 mrg * caller (who must un-busy all pages)
478 1.6 mrg * else we un-busy cluster pages for the pagedaemon
479 1.6 mrg *
480 1.6 mrg * if I/O is pending (async i/o) then we return the pending code.
481 1.6 mrg * [in this case the async i/o done function must clean up when
482 1.6 mrg * i/o is done...]
483 1.6 mrg */
484 1.6 mrg
485 1.6 mrg if (result == VM_PAGER_PEND || result == VM_PAGER_OK) {
486 1.6 mrg if (result == VM_PAGER_OK && (flags & PGO_PDFREECLUST)) {
487 1.6 mrg /*
488 1.6 mrg * drop cluster and relock object (only if I/O is
489 1.6 mrg * not pending)
490 1.6 mrg */
491 1.6 mrg if (uobj)
492 1.6 mrg /* required for dropcluster */
493 1.6 mrg simple_lock(&uobj->vmobjlock);
494 1.6 mrg if (*npages > 1 || pg == NULL)
495 1.6 mrg uvm_pager_dropcluster(uobj, pg, ppsp, npages,
496 1.25 chs PGO_PDFREECLUST);
497 1.6 mrg /* if (uobj): object still locked, as per
498 1.6 mrg * return-state item #3 */
499 1.6 mrg }
500 1.6 mrg return (result);
501 1.6 mrg }
502 1.6 mrg
503 1.6 mrg /*
504 1.25 chs * a pager error occured.
505 1.25 chs * for transient errors, drop to a cluster of 1 page ("pg")
506 1.25 chs * and try again. for hard errors, don't bother retrying.
507 1.6 mrg */
508 1.6 mrg
509 1.6 mrg if (*npages > 1 || pg == NULL) {
510 1.25 chs if (uobj) {
511 1.6 mrg simple_lock(&uobj->vmobjlock);
512 1.25 chs }
513 1.25 chs uvm_pager_dropcluster(uobj, pg, ppsp, npages, PGO_REALLOCSWAP);
514 1.25 chs
515 1.25 chs /*
516 1.25 chs * for failed swap-backed pageouts with a "pg",
517 1.25 chs * we need to reset pg's swslot to either:
518 1.25 chs * "swblk" (for transient errors, so we can retry),
519 1.25 chs * or 0 (for hard errors).
520 1.25 chs */
521 1.25 chs
522 1.25 chs if (uobj == NULL && pg != NULL) {
523 1.25 chs int nswblk = (result == VM_PAGER_AGAIN) ? swblk : 0;
524 1.25 chs if (pg->pqflags & PQ_ANON) {
525 1.25 chs simple_lock(&pg->uanon->an_lock);
526 1.25 chs pg->uanon->an_swslot = nswblk;
527 1.25 chs simple_unlock(&pg->uanon->an_lock);
528 1.25 chs } else {
529 1.25 chs simple_lock(&pg->uobject->vmobjlock);
530 1.25 chs uao_set_swslot(pg->uobject,
531 1.25 chs pg->offset >> PAGE_SHIFT,
532 1.25 chs nswblk);
533 1.25 chs simple_unlock(&pg->uobject->vmobjlock);
534 1.25 chs }
535 1.25 chs }
536 1.25 chs if (result == VM_PAGER_AGAIN) {
537 1.25 chs
538 1.25 chs /*
539 1.25 chs * for transient failures, free all the swslots that
540 1.25 chs * we're not going to retry with.
541 1.25 chs */
542 1.25 chs
543 1.25 chs if (uobj == NULL) {
544 1.25 chs if (pg) {
545 1.25 chs uvm_swap_free(swblk + 1, *npages - 1);
546 1.25 chs } else {
547 1.25 chs uvm_swap_free(swblk, *npages);
548 1.25 chs }
549 1.25 chs }
550 1.25 chs if (pg) {
551 1.25 chs ppsp[0] = pg;
552 1.25 chs *npages = 1;
553 1.25 chs goto ReTry;
554 1.25 chs }
555 1.25 chs } else if (uobj == NULL) {
556 1.25 chs
557 1.25 chs /*
558 1.25 chs * for hard errors on swap-backed pageouts,
559 1.25 chs * mark the swslots as bad. note that we do not
560 1.25 chs * free swslots that we mark bad.
561 1.25 chs */
562 1.25 chs
563 1.25 chs uvm_swap_markbad(swblk, *npages);
564 1.25 chs }
565 1.6 mrg }
566 1.6 mrg
567 1.6 mrg /*
568 1.6 mrg * a pager error occured (even after dropping the cluster, if there
569 1.6 mrg * was one). give up! the caller only has one page ("pg")
570 1.6 mrg * to worry about.
571 1.6 mrg */
572 1.6 mrg
573 1.6 mrg if (uobj && (flags & PGO_PDFREECLUST) != 0)
574 1.6 mrg simple_lock(&uobj->vmobjlock);
575 1.6 mrg return(result);
576 1.1 mrg }
577 1.1 mrg
578 1.1 mrg /*
579 1.1 mrg * uvm_pager_dropcluster: drop a cluster we have built (because we
580 1.1 mrg * got an error, or, if PGO_PDFREECLUST we are un-busying the
581 1.1 mrg * cluster pages on behalf of the pagedaemon).
582 1.1 mrg *
583 1.1 mrg * => uobj, if non-null, is a non-swap-backed object that is
584 1.1 mrg * locked by the caller. we return with this object still
585 1.1 mrg * locked.
586 1.1 mrg * => page queues are not locked
587 1.1 mrg * => pg is our page of interest (the one we clustered around, can be null)
588 1.1 mrg * => ppsp/npages is our current cluster
589 1.1 mrg * => flags: PGO_PDFREECLUST: pageout was a success: un-busy cluster
590 1.1 mrg * pages on behalf of the pagedaemon.
591 1.1 mrg * PGO_REALLOCSWAP: drop previously allocated swap slots for
592 1.1 mrg * clustered swap-backed pages (except for "pg" if !NULL)
593 1.1 mrg * "swblk" is the start of swap alloc (e.g. for ppsp[0])
594 1.1 mrg * [only meaningful if swap-backed (uobj == NULL)]
595 1.1 mrg */
596 1.1 mrg
597 1.21 thorpej void
598 1.25 chs uvm_pager_dropcluster(uobj, pg, ppsp, npages, flags)
599 1.21 thorpej struct uvm_object *uobj; /* IN */
600 1.21 thorpej struct vm_page *pg, **ppsp; /* IN, IN/OUT */
601 1.21 thorpej int *npages; /* IN/OUT */
602 1.21 thorpej int flags;
603 1.1 mrg {
604 1.6 mrg int lcv;
605 1.6 mrg boolean_t obj_is_alive;
606 1.7 chuck struct uvm_object *saved_uobj;
607 1.1 mrg
608 1.6 mrg /*
609 1.6 mrg * drop all pages but "pg"
610 1.6 mrg */
611 1.1 mrg
612 1.6 mrg for (lcv = 0 ; lcv < *npages ; lcv++) {
613 1.1 mrg
614 1.6 mrg if (ppsp[lcv] == pg) /* skip "pg" */
615 1.6 mrg continue;
616 1.1 mrg
617 1.6 mrg /*
618 1.6 mrg * if swap-backed, gain lock on object that owns page. note
619 1.6 mrg * that PQ_ANON bit can't change as long as we are holding
620 1.6 mrg * the PG_BUSY bit (so there is no need to lock the page
621 1.6 mrg * queues to test it).
622 1.6 mrg *
623 1.6 mrg * once we have the lock, dispose of the pointer to swap, if
624 1.6 mrg * requested
625 1.6 mrg */
626 1.6 mrg if (!uobj) {
627 1.6 mrg if (ppsp[lcv]->pqflags & PQ_ANON) {
628 1.6 mrg simple_lock(&ppsp[lcv]->uanon->an_lock);
629 1.6 mrg if (flags & PGO_REALLOCSWAP)
630 1.6 mrg /* zap swap block */
631 1.6 mrg ppsp[lcv]->uanon->an_swslot = 0;
632 1.6 mrg } else {
633 1.6 mrg simple_lock(&ppsp[lcv]->uobject->vmobjlock);
634 1.6 mrg if (flags & PGO_REALLOCSWAP)
635 1.6 mrg uao_set_swslot(ppsp[lcv]->uobject,
636 1.12 chs ppsp[lcv]->offset >> PAGE_SHIFT, 0);
637 1.6 mrg }
638 1.6 mrg }
639 1.6 mrg
640 1.6 mrg /* did someone want the page while we had it busy-locked? */
641 1.6 mrg if (ppsp[lcv]->flags & PG_WANTED)
642 1.6 mrg /* still holding obj lock */
643 1.22 thorpej wakeup(ppsp[lcv]);
644 1.6 mrg
645 1.6 mrg /* if page was released, release it. otherwise un-busy it */
646 1.6 mrg if (ppsp[lcv]->flags & PG_RELEASED) {
647 1.6 mrg
648 1.6 mrg if (ppsp[lcv]->pqflags & PQ_ANON) {
649 1.6 mrg /* so that anfree will free */
650 1.6 mrg ppsp[lcv]->flags &= ~(PG_BUSY);
651 1.6 mrg UVM_PAGE_OWN(ppsp[lcv], NULL);
652 1.6 mrg
653 1.23 chs pmap_page_protect(ppsp[lcv], VM_PROT_NONE);
654 1.13 chs simple_unlock(&ppsp[lcv]->uanon->an_lock);
655 1.6 mrg /* kills anon and frees pg */
656 1.13 chs uvm_anfree(ppsp[lcv]->uanon);
657 1.6 mrg
658 1.6 mrg continue;
659 1.6 mrg }
660 1.6 mrg
661 1.6 mrg /*
662 1.6 mrg * pgo_releasepg will dump the page for us
663 1.6 mrg */
664 1.1 mrg
665 1.1 mrg #ifdef DIAGNOSTIC
666 1.6 mrg if (ppsp[lcv]->uobject->pgops->pgo_releasepg == NULL)
667 1.6 mrg panic("uvm_pager_dropcluster: no releasepg "
668 1.6 mrg "function");
669 1.1 mrg #endif
670 1.7 chuck saved_uobj = ppsp[lcv]->uobject;
671 1.6 mrg obj_is_alive =
672 1.7 chuck saved_uobj->pgops->pgo_releasepg(ppsp[lcv], NULL);
673 1.6 mrg
674 1.1 mrg #ifdef DIAGNOSTIC
675 1.6 mrg /* for normal objects, "pg" is still PG_BUSY by us,
676 1.6 mrg * so obj can't die */
677 1.6 mrg if (uobj && !obj_is_alive)
678 1.6 mrg panic("uvm_pager_dropcluster: object died "
679 1.6 mrg "with active page");
680 1.1 mrg #endif
681 1.7 chuck /* only unlock the object if it is still alive... */
682 1.7 chuck if (obj_is_alive && saved_uobj != uobj)
683 1.7 chuck simple_unlock(&saved_uobj->vmobjlock);
684 1.7 chuck
685 1.7 chuck /*
686 1.7 chuck * XXXCDC: suppose uobj died in the pgo_releasepg?
687 1.7 chuck * how pass that
688 1.7 chuck * info up to caller. we are currently ignoring it...
689 1.7 chuck */
690 1.7 chuck
691 1.7 chuck continue; /* next page */
692 1.1 mrg
693 1.6 mrg } else {
694 1.6 mrg ppsp[lcv]->flags &= ~(PG_BUSY|PG_WANTED);
695 1.6 mrg UVM_PAGE_OWN(ppsp[lcv], NULL);
696 1.6 mrg }
697 1.6 mrg
698 1.6 mrg /*
699 1.6 mrg * if we are operating on behalf of the pagedaemon and we
700 1.6 mrg * had a successful pageout update the page!
701 1.6 mrg */
702 1.6 mrg if (flags & PGO_PDFREECLUST) {
703 1.23 chs pmap_clear_reference(ppsp[lcv]);
704 1.23 chs pmap_clear_modify(ppsp[lcv]);
705 1.6 mrg ppsp[lcv]->flags |= PG_CLEAN;
706 1.6 mrg }
707 1.6 mrg
708 1.6 mrg /* if anonymous cluster, unlock object and move on */
709 1.6 mrg if (!uobj) {
710 1.6 mrg if (ppsp[lcv]->pqflags & PQ_ANON)
711 1.6 mrg simple_unlock(&ppsp[lcv]->uanon->an_lock);
712 1.6 mrg else
713 1.6 mrg simple_unlock(&ppsp[lcv]->uobject->vmobjlock);
714 1.6 mrg }
715 1.6 mrg }
716 1.1 mrg }
717