uvm_pager.c revision 1.8 1 1.8 kleink /* $NetBSD: uvm_pager.c,v 1.8 1998/05/05 20:51:07 kleink Exp $ */
2 1.1 mrg
3 1.1 mrg /*
4 1.1 mrg * XXXCDC: "ROUGH DRAFT" QUALITY UVM PRE-RELEASE FILE!
5 1.1 mrg * >>>USE AT YOUR OWN RISK, WORK IS NOT FINISHED<<<
6 1.1 mrg */
7 1.1 mrg /*
8 1.1 mrg *
9 1.1 mrg * Copyright (c) 1997 Charles D. Cranor and Washington University.
10 1.1 mrg * All rights reserved.
11 1.1 mrg *
12 1.1 mrg * Redistribution and use in source and binary forms, with or without
13 1.1 mrg * modification, are permitted provided that the following conditions
14 1.1 mrg * are met:
15 1.1 mrg * 1. Redistributions of source code must retain the above copyright
16 1.1 mrg * notice, this list of conditions and the following disclaimer.
17 1.1 mrg * 2. Redistributions in binary form must reproduce the above copyright
18 1.1 mrg * notice, this list of conditions and the following disclaimer in the
19 1.1 mrg * documentation and/or other materials provided with the distribution.
20 1.1 mrg * 3. All advertising materials mentioning features or use of this software
21 1.1 mrg * must display the following acknowledgement:
22 1.1 mrg * This product includes software developed by Charles D. Cranor and
23 1.1 mrg * Washington University.
24 1.1 mrg * 4. The name of the author may not be used to endorse or promote products
25 1.1 mrg * derived from this software without specific prior written permission.
26 1.1 mrg *
27 1.1 mrg * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
28 1.1 mrg * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
29 1.1 mrg * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
30 1.1 mrg * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
31 1.1 mrg * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
32 1.1 mrg * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
33 1.1 mrg * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
34 1.1 mrg * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
35 1.1 mrg * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
36 1.1 mrg * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37 1.3 mrg *
38 1.3 mrg * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
39 1.1 mrg */
40 1.1 mrg
41 1.5 mrg #include "opt_uvmhist.h"
42 1.5 mrg #include "opt_pmap_new.h"
43 1.5 mrg
44 1.1 mrg /*
45 1.1 mrg * uvm_pager.c: generic functions used to assist the pagers.
46 1.1 mrg */
47 1.1 mrg
48 1.1 mrg #include <sys/param.h>
49 1.1 mrg #include <sys/systm.h>
50 1.1 mrg #include <sys/proc.h>
51 1.1 mrg #include <sys/malloc.h>
52 1.1 mrg
53 1.1 mrg #include <vm/vm.h>
54 1.1 mrg #include <vm/vm_page.h>
55 1.1 mrg #include <vm/vm_kern.h>
56 1.1 mrg
57 1.1 mrg #define UVM_PAGER
58 1.1 mrg #include <uvm/uvm.h>
59 1.1 mrg
60 1.1 mrg /*
61 1.1 mrg * list of uvm pagers in the system
62 1.1 mrg */
63 1.1 mrg
64 1.1 mrg extern struct uvm_pagerops uvm_deviceops;
65 1.1 mrg extern struct uvm_pagerops uvm_vnodeops;
66 1.1 mrg
67 1.1 mrg struct uvm_pagerops *uvmpagerops[] = {
68 1.6 mrg &uvm_deviceops,
69 1.6 mrg &uvm_vnodeops,
70 1.1 mrg };
71 1.1 mrg
72 1.1 mrg /*
73 1.1 mrg * the pager map: provides KVA for I/O
74 1.1 mrg */
75 1.1 mrg
76 1.1 mrg #define PAGER_MAP_SIZE (4 * 1024 * 1024)
77 1.1 mrg vm_map_t pager_map; /* XXX */
78 1.1 mrg simple_lock_data_t pager_map_wanted_lock;
79 1.1 mrg boolean_t pager_map_wanted; /* locked by pager map */
80 1.1 mrg
81 1.1 mrg
82 1.1 mrg /*
83 1.1 mrg * uvm_pager_init: init pagers (at boot time)
84 1.1 mrg */
85 1.1 mrg
86 1.6 mrg void
87 1.6 mrg uvm_pager_init()
88 1.6 mrg {
89 1.6 mrg int lcv;
90 1.1 mrg
91 1.6 mrg /*
92 1.6 mrg * init pager map
93 1.6 mrg */
94 1.6 mrg
95 1.6 mrg pager_map = uvm_km_suballoc(kernel_map, &uvm.pager_sva, &uvm.pager_eva,
96 1.6 mrg PAGER_MAP_SIZE, FALSE, FALSE, NULL);
97 1.6 mrg simple_lock_init(&pager_map_wanted_lock);
98 1.6 mrg pager_map_wanted = FALSE;
99 1.6 mrg
100 1.6 mrg /*
101 1.6 mrg * init ASYNC I/O queue
102 1.6 mrg */
103 1.6 mrg
104 1.6 mrg TAILQ_INIT(&uvm.aio_done);
105 1.1 mrg
106 1.6 mrg /*
107 1.6 mrg * call pager init functions
108 1.6 mrg */
109 1.6 mrg for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
110 1.6 mrg lcv++) {
111 1.6 mrg if (uvmpagerops[lcv]->pgo_init)
112 1.6 mrg uvmpagerops[lcv]->pgo_init();
113 1.6 mrg }
114 1.1 mrg }
115 1.1 mrg
116 1.1 mrg /*
117 1.1 mrg * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
118 1.1 mrg *
119 1.1 mrg * we basically just map in a blank map entry to reserve the space in the
120 1.1 mrg * map and then use pmap_enter() to put the mappings in by hand.
121 1.1 mrg */
122 1.1 mrg
123 1.6 mrg vm_offset_t
124 1.6 mrg uvm_pagermapin(pps, npages, aiop, waitf)
125 1.6 mrg struct vm_page **pps;
126 1.6 mrg int npages;
127 1.6 mrg struct uvm_aiodesc **aiop; /* OUT */
128 1.6 mrg int waitf;
129 1.1 mrg {
130 1.6 mrg vm_size_t size;
131 1.6 mrg vm_offset_t kva;
132 1.6 mrg struct uvm_aiodesc *aio;
133 1.1 mrg #if !defined(PMAP_NEW)
134 1.6 mrg vm_offset_t cva;
135 1.6 mrg struct vm_page *pp;
136 1.1 mrg #endif
137 1.6 mrg UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
138 1.1 mrg
139 1.6 mrg UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d, aiop=0x%x, waitf=%d)",
140 1.1 mrg pps, npages, aiop, waitf);
141 1.1 mrg
142 1.1 mrg ReStart:
143 1.6 mrg if (aiop) {
144 1.6 mrg MALLOC(aio, struct uvm_aiodesc *, sizeof(*aio), M_TEMP, waitf);
145 1.6 mrg if (aio == NULL)
146 1.6 mrg return(0);
147 1.6 mrg *aiop = aio;
148 1.6 mrg } else {
149 1.6 mrg aio = NULL;
150 1.6 mrg }
151 1.1 mrg
152 1.6 mrg size = npages * PAGE_SIZE;
153 1.6 mrg kva = NULL; /* let system choose VA */
154 1.1 mrg
155 1.6 mrg if (uvm_map(pager_map, &kva, size, NULL,
156 1.1 mrg UVM_UNKNOWN_OFFSET, UVM_FLAG_NOMERGE) != KERN_SUCCESS) {
157 1.6 mrg if (waitf == M_NOWAIT) {
158 1.6 mrg if (aio)
159 1.6 mrg FREE(aio, M_TEMP);
160 1.6 mrg UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
161 1.6 mrg return(NULL);
162 1.6 mrg }
163 1.6 mrg simple_lock(&pager_map_wanted_lock);
164 1.6 mrg pager_map_wanted = TRUE;
165 1.6 mrg UVMHIST_LOG(maphist, " SLEEPING on pager_map",0,0,0,0);
166 1.6 mrg UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, FALSE,
167 1.6 mrg "pager_map",0);
168 1.6 mrg goto ReStart;
169 1.6 mrg }
170 1.1 mrg
171 1.1 mrg #if defined(PMAP_NEW)
172 1.6 mrg /*
173 1.6 mrg * XXX: (ab)using the pmap module to store state info for us.
174 1.6 mrg * (pmap stores the PAs... we fetch them back later and convert back
175 1.6 mrg * to pages with PHYS_TO_VM_PAGE).
176 1.6 mrg */
177 1.6 mrg pmap_kenter_pgs(kva, pps, npages);
178 1.1 mrg
179 1.1 mrg #else /* PMAP_NEW */
180 1.1 mrg
181 1.6 mrg /* got it */
182 1.6 mrg for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
183 1.6 mrg pp = *pps++;
184 1.1 mrg #ifdef DEBUG
185 1.6 mrg if ((pp->flags & PG_BUSY) == 0)
186 1.6 mrg panic("uvm_pagermapin: page not busy");
187 1.1 mrg #endif
188 1.1 mrg
189 1.6 mrg pmap_enter(vm_map_pmap(pager_map), cva, VM_PAGE_TO_PHYS(pp),
190 1.6 mrg VM_PROT_DEFAULT, TRUE);
191 1.6 mrg }
192 1.1 mrg
193 1.1 mrg #endif /* PMAP_NEW */
194 1.1 mrg
195 1.6 mrg UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
196 1.6 mrg return(kva);
197 1.1 mrg }
198 1.1 mrg
199 1.1 mrg /*
200 1.1 mrg * uvm_pagermapout: remove pager_map mapping
201 1.1 mrg *
202 1.1 mrg * we remove our mappings by hand and then remove the mapping (waking
203 1.1 mrg * up anyone wanting space).
204 1.1 mrg */
205 1.1 mrg
206 1.6 mrg void
207 1.6 mrg uvm_pagermapout(kva, npages)
208 1.6 mrg vm_offset_t kva;
209 1.6 mrg int npages;
210 1.6 mrg {
211 1.6 mrg vm_size_t size = npages * PAGE_SIZE;
212 1.6 mrg vm_map_entry_t entries;
213 1.6 mrg UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
214 1.6 mrg
215 1.6 mrg UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
216 1.1 mrg
217 1.6 mrg /*
218 1.6 mrg * duplicate uvm_unmap, but add in pager_map_wanted handling.
219 1.6 mrg */
220 1.6 mrg
221 1.6 mrg vm_map_lock(pager_map);
222 1.6 mrg (void) uvm_unmap_remove(pager_map, kva, kva + size, 0, &entries);
223 1.6 mrg simple_lock(&pager_map_wanted_lock);
224 1.6 mrg if (pager_map_wanted) {
225 1.6 mrg pager_map_wanted = FALSE;
226 1.6 mrg wakeup(pager_map);
227 1.6 mrg }
228 1.6 mrg simple_unlock(&pager_map_wanted_lock);
229 1.6 mrg vm_map_unlock(pager_map);
230 1.6 mrg if (entries)
231 1.6 mrg uvm_unmap_detach(entries, 0);
232 1.1 mrg
233 1.6 mrg UVMHIST_LOG(maphist,"<- done",0,0,0,0);
234 1.1 mrg }
235 1.1 mrg
236 1.1 mrg /*
237 1.1 mrg * uvm_mk_pcluster
238 1.1 mrg *
239 1.1 mrg * generic "make 'pager put' cluster" function. a pager can either
240 1.1 mrg * [1] set pgo_mk_pcluster to NULL (never cluster), [2] set it to this
241 1.1 mrg * generic function, or [3] set it to a pager specific function.
242 1.1 mrg *
243 1.1 mrg * => caller must lock object _and_ pagequeues (since we need to look
244 1.1 mrg * at active vs. inactive bits, etc.)
245 1.1 mrg * => caller must make center page busy and write-protect it
246 1.1 mrg * => we mark all cluster pages busy for the caller
247 1.1 mrg * => the caller must unbusy all pages (and check wanted/released
248 1.1 mrg * status if it drops the object lock)
249 1.1 mrg * => flags:
250 1.1 mrg * PGO_ALLPAGES: all pages in object are valid targets
251 1.1 mrg * !PGO_ALLPAGES: use "lo" and "hi" to limit range of cluster
252 1.1 mrg * PGO_DOACTCLUST: include active pages in cluster.
253 1.1 mrg * NOTE: the caller should clear PG_CLEANCHK bits if PGO_DOACTCLUST.
254 1.1 mrg * PG_CLEANCHK is only a hint, but clearing will help reduce
255 1.1 mrg * the number of calls we make to the pmap layer.
256 1.1 mrg */
257 1.1 mrg
258 1.6 mrg struct vm_page **
259 1.6 mrg uvm_mk_pcluster(uobj, pps, npages, center, flags, mlo, mhi)
260 1.6 mrg struct uvm_object *uobj; /* IN */
261 1.6 mrg struct vm_page **pps, *center; /* IN/OUT, IN */
262 1.6 mrg int *npages, flags; /* IN/OUT, IN */
263 1.6 mrg vm_offset_t mlo, mhi; /* IN (if !PGO_ALLPAGES) */
264 1.1 mrg {
265 1.6 mrg struct vm_page **ppsp, *pclust;
266 1.6 mrg vm_offset_t lo, hi, curoff;
267 1.6 mrg int center_idx, forward;
268 1.6 mrg UVMHIST_FUNC("uvm_mk_pcluster"); UVMHIST_CALLED(maphist);
269 1.6 mrg
270 1.6 mrg /*
271 1.6 mrg * center page should already be busy and write protected. XXX:
272 1.6 mrg * suppose page is wired? if we lock, then a process could
273 1.6 mrg * fault/block on it. if we don't lock, a process could write the
274 1.6 mrg * pages in the middle of an I/O. (consider an msync()). let's
275 1.6 mrg * lock it for now (better to delay than corrupt data?).
276 1.6 mrg */
277 1.6 mrg
278 1.6 mrg /*
279 1.6 mrg * get cluster boundaries, check sanity, and apply our limits as well.
280 1.6 mrg */
281 1.6 mrg
282 1.6 mrg uobj->pgops->pgo_cluster(uobj, center->offset, &lo, &hi);
283 1.6 mrg if ((flags & PGO_ALLPAGES) == 0) {
284 1.6 mrg if (lo < mlo)
285 1.6 mrg lo = mlo;
286 1.6 mrg if (hi > mhi)
287 1.6 mrg hi = mhi;
288 1.6 mrg }
289 1.6 mrg if ((hi - lo) / PAGE_SIZE > *npages) { /* pps too small, bail out! */
290 1.1 mrg #ifdef DIAGNOSTIC
291 1.6 mrg printf("uvm_mk_pcluster: provided page array too small (fixed)\n");
292 1.1 mrg #endif
293 1.6 mrg pps[0] = center;
294 1.6 mrg *npages = 1;
295 1.6 mrg return(pps);
296 1.6 mrg }
297 1.6 mrg
298 1.6 mrg /*
299 1.6 mrg * now determine the center and attempt to cluster around the
300 1.6 mrg * edges
301 1.6 mrg */
302 1.6 mrg
303 1.6 mrg center_idx = (center->offset - lo) / PAGE_SIZE;
304 1.6 mrg pps[center_idx] = center; /* plug in the center page */
305 1.6 mrg ppsp = &pps[center_idx];
306 1.6 mrg *npages = 1;
307 1.6 mrg
308 1.6 mrg /*
309 1.6 mrg * attempt to cluster around the left [backward], and then
310 1.6 mrg * the right side [forward].
311 1.6 mrg *
312 1.6 mrg * note that for inactive pages (pages that have been deactivated)
313 1.6 mrg * there are no valid mappings and PG_CLEAN should be up to date.
314 1.6 mrg * [i.e. there is no need to query the pmap with pmap_is_modified
315 1.6 mrg * since there are no mappings].
316 1.6 mrg */
317 1.6 mrg
318 1.6 mrg for (forward = 0 ; forward <= 1 ; forward++) {
319 1.6 mrg
320 1.6 mrg curoff = center->offset + PAGE_SIZE * (forward) ? 1 : -1;
321 1.6 mrg for ( ;(forward == 0 && curoff >= lo) ||
322 1.6 mrg (forward && curoff < hi); curoff +=
323 1.6 mrg PAGE_SIZE * (forward) ? 1 : -1) {
324 1.6 mrg
325 1.6 mrg pclust = uvm_pagelookup(uobj, curoff); /* lookup page */
326 1.6 mrg if (pclust == NULL)
327 1.6 mrg break; /* no page */
328 1.6 mrg /* handle active pages */
329 1.6 mrg /* NOTE: inactive pages don't have pmap mappings */
330 1.6 mrg if ((pclust->pqflags & PQ_INACTIVE) == 0) {
331 1.6 mrg if ((flags & PGO_DOACTCLUST) == 0)
332 1.6 mrg /* dont want mapped pages at all */
333 1.6 mrg break;
334 1.6 mrg
335 1.6 mrg /* make sure "clean" bit is sync'd */
336 1.6 mrg if ((pclust->flags & PG_CLEANCHK) == 0) {
337 1.6 mrg if ((pclust->flags & (PG_CLEAN|PG_BUSY))
338 1.6 mrg == PG_CLEAN &&
339 1.6 mrg pmap_is_modified(PMAP_PGARG(pclust)))
340 1.6 mrg pclust->flags &= ~PG_CLEAN;
341 1.6 mrg /* now checked */
342 1.6 mrg pclust->flags |= PG_CLEANCHK;
343 1.6 mrg }
344 1.6 mrg }
345 1.6 mrg /* is page available for cleaning and does it need it */
346 1.6 mrg if ((pclust->flags & (PG_CLEAN|PG_BUSY)) != 0)
347 1.6 mrg break; /* page is already clean or is busy */
348 1.6 mrg
349 1.6 mrg /* yes! enroll the page in our array */
350 1.6 mrg pclust->flags |= PG_BUSY; /* busy! */
351 1.6 mrg UVM_PAGE_OWN(pclust, "uvm_mk_pcluster");
352 1.6 mrg /* XXX: protect wired page? see above comment. */
353 1.6 mrg pmap_page_protect(PMAP_PGARG(pclust), VM_PROT_READ);
354 1.6 mrg if (!forward) {
355 1.6 mrg ppsp--; /* back up one page */
356 1.6 mrg *ppsp = pclust;
357 1.6 mrg } else {
358 1.6 mrg /* move forward one page */
359 1.6 mrg ppsp[*npages] = pclust;
360 1.6 mrg }
361 1.6 mrg *npages = *npages + 1;
362 1.6 mrg }
363 1.6 mrg }
364 1.6 mrg
365 1.6 mrg /*
366 1.6 mrg * done! return the cluster array to the caller!!!
367 1.6 mrg */
368 1.1 mrg
369 1.6 mrg UVMHIST_LOG(maphist, "<- done",0,0,0,0);
370 1.6 mrg return(ppsp);
371 1.1 mrg }
372 1.1 mrg
373 1.1 mrg
374 1.1 mrg /*
375 1.1 mrg * uvm_shareprot: generic share protect routine
376 1.1 mrg *
377 1.1 mrg * => caller must lock map entry's map
378 1.1 mrg * => caller must lock object pointed to by map entry
379 1.1 mrg */
380 1.1 mrg
381 1.6 mrg void
382 1.6 mrg uvm_shareprot(entry, prot)
383 1.6 mrg vm_map_entry_t entry;
384 1.6 mrg vm_prot_t prot;
385 1.1 mrg {
386 1.6 mrg struct uvm_object *uobj = entry->object.uvm_obj;
387 1.6 mrg struct vm_page *pp;
388 1.6 mrg vm_offset_t start, stop;
389 1.6 mrg UVMHIST_FUNC("uvm_shareprot"); UVMHIST_CALLED(maphist);
390 1.6 mrg
391 1.6 mrg if (UVM_ET_ISMAP(entry))
392 1.6 mrg panic("uvm_shareprot: non-object attached");
393 1.6 mrg
394 1.6 mrg start = entry->offset;
395 1.6 mrg stop = start + (entry->end - entry->start);
396 1.6 mrg
397 1.6 mrg /*
398 1.6 mrg * traverse list of pages in object. if page in range, pmap_prot it
399 1.6 mrg */
400 1.6 mrg
401 1.6 mrg for (pp = uobj->memq.tqh_first ; pp != NULL ; pp = pp->listq.tqe_next) {
402 1.6 mrg if (pp->offset >= start && pp->offset < stop)
403 1.6 mrg pmap_page_protect(PMAP_PGARG(pp), prot);
404 1.6 mrg }
405 1.6 mrg UVMHIST_LOG(maphist, "<- done",0,0,0,0);
406 1.1 mrg }
407 1.1 mrg
408 1.1 mrg /*
409 1.1 mrg * uvm_pager_put: high level pageout routine
410 1.1 mrg *
411 1.1 mrg * we want to pageout page "pg" to backing store, clustering if
412 1.1 mrg * possible.
413 1.1 mrg *
414 1.1 mrg * => page queues must be locked by caller
415 1.1 mrg * => if page is not swap-backed, then "uobj" points to the object
416 1.1 mrg * backing it. this object should be locked by the caller.
417 1.1 mrg * => if page is swap-backed, then "uobj" should be NULL.
418 1.1 mrg * => "pg" should be PG_BUSY (by caller), and !PG_CLEAN
419 1.1 mrg * for swap-backed memory, "pg" can be NULL if there is no page
420 1.1 mrg * of interest [sometimes the case for the pagedaemon]
421 1.1 mrg * => "ppsp_ptr" should point to an array of npages vm_page pointers
422 1.1 mrg * for possible cluster building
423 1.1 mrg * => flags (first two for non-swap-backed pages)
424 1.1 mrg * PGO_ALLPAGES: all pages in uobj are valid targets
425 1.1 mrg * PGO_DOACTCLUST: include "PQ_ACTIVE" pages as valid targets
426 1.1 mrg * PGO_SYNCIO: do SYNC I/O (no async)
427 1.1 mrg * PGO_PDFREECLUST: pagedaemon: drop cluster on successful I/O
428 1.1 mrg * => start/stop: if (uobj && !PGO_ALLPAGES) limit targets to this range
429 1.1 mrg * if (!uobj) start is the (daddr_t) of the starting swapblk
430 1.1 mrg * => return state:
431 1.1 mrg * 1. we return the VM_PAGER status code of the pageout
432 1.1 mrg * 2. we return with the page queues unlocked
433 1.1 mrg * 3. if (uobj != NULL) [!swap_backed] we return with
434 1.1 mrg * uobj locked _only_ if PGO_PDFREECLUST is set
435 1.1 mrg * AND result != VM_PAGER_PEND. in all other cases
436 1.1 mrg * we return with uobj unlocked. [this is a hack
437 1.1 mrg * that allows the pagedaemon to save one lock/unlock
438 1.1 mrg * pair in the !swap_backed case since we have to
439 1.1 mrg * lock the uobj to drop the cluster anyway]
440 1.1 mrg * 4. on errors we always drop the cluster. thus, if we return
441 1.1 mrg * !PEND, !OK, then the caller only has to worry about
442 1.1 mrg * un-busying the main page (not the cluster pages).
443 1.1 mrg * 5. on success, if !PGO_PDFREECLUST, we return the cluster
444 1.1 mrg * with all pages busy (caller must un-busy and check
445 1.1 mrg * wanted/released flags).
446 1.1 mrg */
447 1.1 mrg
448 1.6 mrg int
449 1.6 mrg uvm_pager_put(uobj, pg, ppsp_ptr, npages, flags, start, stop)
450 1.6 mrg struct uvm_object *uobj; /* IN */
451 1.6 mrg struct vm_page *pg, ***ppsp_ptr;/* IN, IN/OUT */
452 1.6 mrg int *npages; /* IN/OUT */
453 1.6 mrg int flags; /* IN */
454 1.6 mrg vm_offset_t start, stop; /* IN, IN */
455 1.6 mrg {
456 1.6 mrg int result;
457 1.6 mrg daddr_t swblk;
458 1.6 mrg struct vm_page **ppsp = *ppsp_ptr;
459 1.6 mrg
460 1.6 mrg /*
461 1.6 mrg * note that uobj is null if we are doing a swap-backed pageout.
462 1.6 mrg * note that uobj is !null if we are doing normal object pageout.
463 1.6 mrg * note that the page queues must be locked to cluster.
464 1.6 mrg */
465 1.6 mrg
466 1.6 mrg if (uobj) { /* if !swap-backed */
467 1.6 mrg
468 1.6 mrg /*
469 1.6 mrg * attempt to build a cluster for pageout using its
470 1.6 mrg * make-put-cluster function (if it has one).
471 1.6 mrg */
472 1.6 mrg
473 1.6 mrg if (uobj->pgops->pgo_mk_pcluster) {
474 1.6 mrg ppsp = uobj->pgops->pgo_mk_pcluster(uobj, ppsp,
475 1.6 mrg npages, pg, flags, start, stop);
476 1.6 mrg *ppsp_ptr = ppsp; /* update caller's pointer */
477 1.6 mrg } else {
478 1.6 mrg ppsp[0] = pg;
479 1.6 mrg *npages = 1;
480 1.6 mrg }
481 1.6 mrg
482 1.6 mrg swblk = 0; /* XXX: keep gcc happy */
483 1.1 mrg
484 1.6 mrg } else {
485 1.1 mrg
486 1.6 mrg /*
487 1.6 mrg * for swap-backed pageout, the caller (the pagedaemon) has
488 1.6 mrg * already built the cluster for us. the starting swap
489 1.6 mrg * block we are writing to has been passed in as "start."
490 1.6 mrg * "pg" could be NULL if there is no page we are especially
491 1.6 mrg * interested in (in which case the whole cluster gets dropped
492 1.6 mrg * in the event of an error or a sync "done").
493 1.6 mrg */
494 1.6 mrg swblk = (daddr_t) start;
495 1.6 mrg /* ppsp and npages should be ok */
496 1.6 mrg }
497 1.1 mrg
498 1.6 mrg /* now that we've clustered we can unlock the page queues */
499 1.6 mrg uvm_unlock_pageq();
500 1.1 mrg
501 1.6 mrg /*
502 1.6 mrg * now attempt the I/O. if we have a failure and we are
503 1.6 mrg * clustered, we will drop the cluster and try again.
504 1.6 mrg */
505 1.1 mrg
506 1.1 mrg ReTry:
507 1.6 mrg if (uobj) {
508 1.6 mrg /* object is locked */
509 1.6 mrg result = uobj->pgops->pgo_put(uobj, ppsp, *npages,
510 1.6 mrg flags & PGO_SYNCIO);
511 1.6 mrg /* object is now unlocked */
512 1.6 mrg } else {
513 1.6 mrg /* nothing locked */
514 1.6 mrg result = uvm_swap_put(swblk, ppsp, *npages, flags & PGO_SYNCIO);
515 1.6 mrg /* nothing locked */
516 1.6 mrg }
517 1.6 mrg
518 1.6 mrg /*
519 1.6 mrg * we have attempted the I/O.
520 1.6 mrg *
521 1.6 mrg * if the I/O was a success then:
522 1.6 mrg * if !PGO_PDFREECLUST, we return the cluster to the
523 1.6 mrg * caller (who must un-busy all pages)
524 1.6 mrg * else we un-busy cluster pages for the pagedaemon
525 1.6 mrg *
526 1.6 mrg * if I/O is pending (async i/o) then we return the pending code.
527 1.6 mrg * [in this case the async i/o done function must clean up when
528 1.6 mrg * i/o is done...]
529 1.6 mrg */
530 1.6 mrg
531 1.6 mrg if (result == VM_PAGER_PEND || result == VM_PAGER_OK) {
532 1.6 mrg if (result == VM_PAGER_OK && (flags & PGO_PDFREECLUST)) {
533 1.6 mrg /*
534 1.6 mrg * drop cluster and relock object (only if I/O is
535 1.6 mrg * not pending)
536 1.6 mrg */
537 1.6 mrg if (uobj)
538 1.6 mrg /* required for dropcluster */
539 1.6 mrg simple_lock(&uobj->vmobjlock);
540 1.6 mrg if (*npages > 1 || pg == NULL)
541 1.6 mrg uvm_pager_dropcluster(uobj, pg, ppsp, npages,
542 1.6 mrg PGO_PDFREECLUST, 0);
543 1.6 mrg /* if (uobj): object still locked, as per
544 1.6 mrg * return-state item #3 */
545 1.6 mrg }
546 1.6 mrg return (result);
547 1.6 mrg }
548 1.6 mrg
549 1.6 mrg /*
550 1.6 mrg * a pager error occured. if we have clustered, we drop the
551 1.6 mrg * cluster and try again.
552 1.6 mrg */
553 1.6 mrg
554 1.6 mrg if (*npages > 1 || pg == NULL) {
555 1.6 mrg if (uobj)
556 1.6 mrg simple_lock(&uobj->vmobjlock);
557 1.6 mrg uvm_pager_dropcluster(uobj, pg, ppsp, npages, PGO_REALLOCSWAP,
558 1.6 mrg swblk);
559 1.6 mrg if (pg != NULL)
560 1.6 mrg goto ReTry;
561 1.6 mrg }
562 1.6 mrg
563 1.6 mrg /*
564 1.6 mrg * a pager error occured (even after dropping the cluster, if there
565 1.6 mrg * was one). give up! the caller only has one page ("pg")
566 1.6 mrg * to worry about.
567 1.6 mrg */
568 1.6 mrg
569 1.6 mrg if (uobj && (flags & PGO_PDFREECLUST) != 0)
570 1.6 mrg simple_lock(&uobj->vmobjlock);
571 1.6 mrg return(result);
572 1.1 mrg }
573 1.1 mrg
574 1.1 mrg /*
575 1.1 mrg * uvm_pager_dropcluster: drop a cluster we have built (because we
576 1.1 mrg * got an error, or, if PGO_PDFREECLUST we are un-busying the
577 1.1 mrg * cluster pages on behalf of the pagedaemon).
578 1.1 mrg *
579 1.1 mrg * => uobj, if non-null, is a non-swap-backed object that is
580 1.1 mrg * locked by the caller. we return with this object still
581 1.1 mrg * locked.
582 1.1 mrg * => page queues are not locked
583 1.1 mrg * => pg is our page of interest (the one we clustered around, can be null)
584 1.1 mrg * => ppsp/npages is our current cluster
585 1.1 mrg * => flags: PGO_PDFREECLUST: pageout was a success: un-busy cluster
586 1.1 mrg * pages on behalf of the pagedaemon.
587 1.1 mrg * PGO_REALLOCSWAP: drop previously allocated swap slots for
588 1.1 mrg * clustered swap-backed pages (except for "pg" if !NULL)
589 1.1 mrg * "swblk" is the start of swap alloc (e.g. for ppsp[0])
590 1.1 mrg * [only meaningful if swap-backed (uobj == NULL)]
591 1.1 mrg */
592 1.1 mrg
593 1.1 mrg
594 1.1 mrg void uvm_pager_dropcluster(uobj, pg, ppsp, npages, flags, swblk)
595 1.1 mrg
596 1.1 mrg struct uvm_object *uobj; /* IN */
597 1.1 mrg struct vm_page *pg, **ppsp; /* IN, IN/OUT */
598 1.1 mrg int *npages; /* IN/OUT */
599 1.1 mrg int flags;
600 1.1 mrg int swblk; /* valid if (uobj == NULL && PGO_REALLOCSWAP) */
601 1.1 mrg
602 1.1 mrg {
603 1.6 mrg int lcv;
604 1.6 mrg boolean_t obj_is_alive;
605 1.7 chuck struct uvm_object *saved_uobj;
606 1.1 mrg
607 1.6 mrg /*
608 1.6 mrg * if we need to reallocate swap space for the cluster we are dropping
609 1.6 mrg * (true if swap-backed and PGO_REALLOCSWAP) then free the old
610 1.6 mrg * allocation now. save a block for "pg" if it is non-NULL.
611 1.6 mrg *
612 1.6 mrg * note that we will zap the object's pointer to swap in the "for" loop
613 1.6 mrg * below...
614 1.6 mrg */
615 1.6 mrg
616 1.6 mrg if (uobj == NULL && (flags & PGO_REALLOCSWAP)) {
617 1.6 mrg if (pg)
618 1.6 mrg uvm_swap_free(swblk + 1, *npages - 1);
619 1.6 mrg else
620 1.6 mrg uvm_swap_free(swblk, *npages);
621 1.6 mrg }
622 1.6 mrg
623 1.6 mrg /*
624 1.6 mrg * drop all pages but "pg"
625 1.6 mrg */
626 1.1 mrg
627 1.6 mrg for (lcv = 0 ; lcv < *npages ; lcv++) {
628 1.1 mrg
629 1.6 mrg if (ppsp[lcv] == pg) /* skip "pg" */
630 1.6 mrg continue;
631 1.1 mrg
632 1.6 mrg /*
633 1.6 mrg * if swap-backed, gain lock on object that owns page. note
634 1.6 mrg * that PQ_ANON bit can't change as long as we are holding
635 1.6 mrg * the PG_BUSY bit (so there is no need to lock the page
636 1.6 mrg * queues to test it).
637 1.6 mrg *
638 1.6 mrg * once we have the lock, dispose of the pointer to swap, if
639 1.6 mrg * requested
640 1.6 mrg */
641 1.6 mrg if (!uobj) {
642 1.6 mrg if (ppsp[lcv]->pqflags & PQ_ANON) {
643 1.6 mrg simple_lock(&ppsp[lcv]->uanon->an_lock);
644 1.6 mrg if (flags & PGO_REALLOCSWAP)
645 1.6 mrg /* zap swap block */
646 1.6 mrg ppsp[lcv]->uanon->an_swslot = 0;
647 1.6 mrg } else {
648 1.6 mrg simple_lock(&ppsp[lcv]->uobject->vmobjlock);
649 1.6 mrg if (flags & PGO_REALLOCSWAP)
650 1.6 mrg uao_set_swslot(ppsp[lcv]->uobject,
651 1.6 mrg ppsp[lcv]->offset / PAGE_SIZE, 0);
652 1.6 mrg }
653 1.6 mrg }
654 1.6 mrg
655 1.6 mrg /* did someone want the page while we had it busy-locked? */
656 1.6 mrg if (ppsp[lcv]->flags & PG_WANTED)
657 1.6 mrg /* still holding obj lock */
658 1.6 mrg thread_wakeup(ppsp[lcv]);
659 1.6 mrg
660 1.6 mrg /* if page was released, release it. otherwise un-busy it */
661 1.6 mrg if (ppsp[lcv]->flags & PG_RELEASED) {
662 1.6 mrg
663 1.6 mrg if (ppsp[lcv]->pqflags & PQ_ANON) {
664 1.6 mrg /* so that anfree will free */
665 1.6 mrg ppsp[lcv]->flags &= ~(PG_BUSY);
666 1.6 mrg UVM_PAGE_OWN(ppsp[lcv], NULL);
667 1.6 mrg
668 1.6 mrg pmap_page_protect(PMAP_PGARG(ppsp[lcv]),
669 1.6 mrg VM_PROT_NONE); /* be safe */
670 1.6 mrg /* kills anon and frees pg */
671 1.6 mrg uvm_anfree(ppsp[lcv]->uanon);
672 1.6 mrg
673 1.6 mrg continue;
674 1.6 mrg }
675 1.6 mrg
676 1.6 mrg /*
677 1.6 mrg * pgo_releasepg will dump the page for us
678 1.6 mrg */
679 1.1 mrg
680 1.1 mrg #ifdef DIAGNOSTIC
681 1.6 mrg if (ppsp[lcv]->uobject->pgops->pgo_releasepg == NULL)
682 1.6 mrg panic("uvm_pager_dropcluster: no releasepg "
683 1.6 mrg "function");
684 1.1 mrg #endif
685 1.7 chuck saved_uobj = ppsp[lcv]->uobject;
686 1.6 mrg obj_is_alive =
687 1.7 chuck saved_uobj->pgops->pgo_releasepg(ppsp[lcv], NULL);
688 1.6 mrg
689 1.1 mrg #ifdef DIAGNOSTIC
690 1.6 mrg /* for normal objects, "pg" is still PG_BUSY by us,
691 1.6 mrg * so obj can't die */
692 1.6 mrg if (uobj && !obj_is_alive)
693 1.6 mrg panic("uvm_pager_dropcluster: object died "
694 1.6 mrg "with active page");
695 1.1 mrg #endif
696 1.7 chuck /* only unlock the object if it is still alive... */
697 1.7 chuck if (obj_is_alive && saved_uobj != uobj)
698 1.7 chuck simple_unlock(&saved_uobj->vmobjlock);
699 1.7 chuck
700 1.7 chuck /*
701 1.7 chuck * XXXCDC: suppose uobj died in the pgo_releasepg?
702 1.7 chuck * how pass that
703 1.7 chuck * info up to caller. we are currently ignoring it...
704 1.7 chuck */
705 1.7 chuck
706 1.7 chuck continue; /* next page */
707 1.1 mrg
708 1.6 mrg } else {
709 1.6 mrg ppsp[lcv]->flags &= ~(PG_BUSY|PG_WANTED);
710 1.6 mrg UVM_PAGE_OWN(ppsp[lcv], NULL);
711 1.6 mrg }
712 1.6 mrg
713 1.6 mrg /*
714 1.6 mrg * if we are operating on behalf of the pagedaemon and we
715 1.6 mrg * had a successful pageout update the page!
716 1.6 mrg */
717 1.6 mrg if (flags & PGO_PDFREECLUST) {
718 1.6 mrg /* XXX: with PMAP_NEW ref should already be clear,
719 1.6 mrg * but don't trust! */
720 1.6 mrg pmap_clear_reference(PMAP_PGARG(ppsp[lcv]));
721 1.6 mrg pmap_clear_modify(PMAP_PGARG(ppsp[lcv]));
722 1.6 mrg ppsp[lcv]->flags |= PG_CLEAN;
723 1.6 mrg }
724 1.6 mrg
725 1.6 mrg /* if anonymous cluster, unlock object and move on */
726 1.6 mrg if (!uobj) {
727 1.6 mrg if (ppsp[lcv]->pqflags & PQ_ANON)
728 1.6 mrg simple_unlock(&ppsp[lcv]->uanon->an_lock);
729 1.6 mrg else
730 1.6 mrg simple_unlock(&ppsp[lcv]->uobject->vmobjlock);
731 1.6 mrg }
732 1.1 mrg
733 1.6 mrg }
734 1.6 mrg
735 1.6 mrg /*
736 1.6 mrg * drop to a cluster of 1 page ("pg") if requested
737 1.6 mrg */
738 1.6 mrg
739 1.6 mrg if (pg && (flags & PGO_PDFREECLUST) == 0) {
740 1.6 mrg /*
741 1.6 mrg * if we are not a successful pageout, we make a 1 page cluster.
742 1.6 mrg */
743 1.6 mrg ppsp[0] = pg;
744 1.6 mrg *npages = 1;
745 1.6 mrg
746 1.6 mrg /*
747 1.6 mrg * assign new swap block to new cluster, if anon backed
748 1.6 mrg */
749 1.6 mrg if (uobj == NULL && (flags & PGO_REALLOCSWAP)) {
750 1.6 mrg if (pg->pqflags & PQ_ANON) {
751 1.6 mrg simple_lock(&pg->uanon->an_lock);
752 1.6 mrg pg->uanon->an_swslot = swblk; /* reassign */
753 1.6 mrg simple_unlock(&pg->uanon->an_lock);
754 1.6 mrg } else {
755 1.6 mrg simple_lock(&pg->uobject->vmobjlock);
756 1.6 mrg uao_set_swslot(pg->uobject,
757 1.6 mrg pg->offset / PAGE_SIZE, swblk);
758 1.6 mrg simple_unlock(&pg->uobject->vmobjlock);
759 1.6 mrg }
760 1.6 mrg }
761 1.6 mrg }
762 1.1 mrg }
763