xref: /titanic_41/usr/src/uts/common/fs/swapfs/swap_vnops.c (revision d58fda4376e4bf67072ce2e69f6f47036f9dbb68)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/types.h>
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/buf.h>
33 #include <sys/cred.h>
34 #include <sys/errno.h>
35 #include <sys/vnode.h>
36 #include <sys/cmn_err.h>
37 #include <sys/swap.h>
38 #include <sys/mman.h>
39 #include <sys/vmsystm.h>
40 #include <sys/vtrace.h>
41 #include <sys/debug.h>
42 #include <sys/sysmacros.h>
43 #include <sys/vm.h>
44 
45 #include <sys/fs/swapnode.h>
46 
47 #include <vm/seg.h>
48 #include <vm/page.h>
49 #include <vm/pvn.h>
50 #include <fs/fs_subr.h>
51 
52 #include <vm/seg_kp.h>
53 
54 /*
55  * Define the routines within this file.
56  */
57 static int	swap_getpage(struct vnode *vp, offset_t off, size_t len,
58     uint_t *protp, struct page **plarr, size_t plsz,
59     struct seg *seg, caddr_t addr, enum seg_rw rw, struct cred *cr);
60 static int	swap_putpage(struct vnode *vp, offset_t off, size_t len,
61     int flags, struct cred *cr);
62 static void	swap_inactive(struct vnode *vp, struct cred *cr);
63 static void	swap_dispose(vnode_t *vp, page_t *pp, int fl, int dn,
64     cred_t *cr);
65 
66 static int	swap_getapage(struct vnode *vp, u_offset_t off, size_t len,
67     uint_t *protp, page_t **plarr, size_t plsz,
68     struct seg *seg, caddr_t addr, enum seg_rw rw, struct cred *cr);
69 
70 int	swap_getconpage(struct vnode *vp, u_offset_t off, size_t len,
71     uint_t *protp, page_t **plarr, size_t plsz, page_t *conpp, spgcnt_t *nreloc,
72     struct seg *seg, caddr_t addr, enum seg_rw rw, struct cred *cr);
73 
74 static int 	swap_putapage(struct vnode *vp, page_t *pp, u_offset_t *off,
75     size_t *lenp, int flags, struct cred *cr);
76 
77 const fs_operation_def_t swap_vnodeops_template[] = {
78 	VOPNAME_INACTIVE, (fs_generic_func_p) swap_inactive,
79 	VOPNAME_GETPAGE, swap_getpage,
80 	VOPNAME_PUTPAGE, swap_putpage,
81 	VOPNAME_DISPOSE, (fs_generic_func_p) swap_dispose,
82 	VOPNAME_SETFL, fs_error,
83 	VOPNAME_POLL, fs_error,
84 	VOPNAME_PATHCONF, fs_error,
85 	VOPNAME_GETSECATTR, fs_error,
86 	VOPNAME_SHRLOCK, fs_error,
87 	NULL, NULL
88 };
89 
90 vnodeops_t *swap_vnodeops;
91 
92 /* ARGSUSED */
93 static void
94 swap_inactive(
95 	struct vnode *vp,
96 	struct cred *cr)
97 {
98 	SWAPFS_PRINT(SWAP_VOPS, "swap_inactive: vp %x\n", vp, 0, 0, 0, 0);
99 }
100 
101 /*
102  * Return all the pages from [off..off+len] in given file
103  */
104 static int
105 swap_getpage(
106 	struct vnode *vp,
107 	offset_t off,
108 	size_t len,
109 	uint_t *protp,
110 	page_t *pl[],
111 	size_t plsz,
112 	struct seg *seg,
113 	caddr_t addr,
114 	enum seg_rw rw,
115 	struct cred *cr)
116 {
117 	int err;
118 
119 	SWAPFS_PRINT(SWAP_VOPS, "swap_getpage: vp %p, off %llx, len %lx\n",
120 	    (void *)vp, off, len, 0, 0);
121 
122 	TRACE_3(TR_FAC_SWAPFS, TR_SWAPFS_GETPAGE,
123 	    "swapfs getpage:vp %p off %llx len %ld",
124 	    (void *)vp, off, len);
125 
126 	if (len <= PAGESIZE) {
127 		err = swap_getapage(vp, (u_offset_t)off, len, protp, pl, plsz,
128 		    seg, addr, rw, cr);
129 	} else {
130 		err = pvn_getpages(swap_getapage, vp, (u_offset_t)off, len,
131 		    protp, pl, plsz, seg, addr, rw, cr);
132 	}
133 
134 	return (err);
135 }
136 
137 /*
138  * Called from pvn_getpages or swap_getpage to get a particular page.
139  */
140 /*ARGSUSED*/
141 static int
142 swap_getapage(
143 	struct vnode *vp,
144 	u_offset_t off,
145 	size_t len,
146 	uint_t *protp,
147 	page_t *pl[],
148 	size_t plsz,
149 	struct seg *seg,
150 	caddr_t addr,
151 	enum seg_rw rw,
152 	struct cred *cr)
153 {
154 	struct page *pp, *rpp;
155 	int flags;
156 	int err = 0;
157 	struct vnode *pvp = NULL;
158 	u_offset_t poff;
159 	int flag_noreloc;
160 	se_t lock;
161 	extern int kcage_on;
162 	int upgrade = 0;
163 
164 	SWAPFS_PRINT(SWAP_VOPS, "swap_getapage: vp %p, off %llx, len %lx\n",
165 		vp, off, len, 0, 0);
166 
167 	/*
168 	 * Until there is a call-back mechanism to cause SEGKP
169 	 * pages to be unlocked, make them non-relocatable.
170 	 */
171 	if (SEG_IS_SEGKP(seg))
172 		flag_noreloc = PG_NORELOC;
173 	else
174 		flag_noreloc = 0;
175 
176 	if (protp != NULL)
177 		*protp = PROT_ALL;
178 
179 	lock = (rw == S_CREATE ? SE_EXCL : SE_SHARED);
180 
181 again:
182 	if (pp = page_lookup(vp, off, lock)) {
183 		/*
184 		 * In very rare instances, a segkp page may have been
185 		 * relocated outside of the kernel by the kernel cage
186 		 * due to the window between page_unlock() and
187 		 * VOP_PUTPAGE() in segkp_unlock().  Due to the
188 		 * rareness of these occurances, the solution is to
189 		 * relocate the page to a P_NORELOC page.
190 		 */
191 		if (flag_noreloc != 0) {
192 			if (!PP_ISNORELOC(pp) && kcage_on) {
193 				if (lock != SE_EXCL) {
194 					upgrade = 1;
195 					if (!page_tryupgrade(pp)) {
196 						page_unlock(pp);
197 						lock = SE_EXCL;
198 						goto again;
199 					}
200 				}
201 
202 				if (page_relocate_cage(&pp, &rpp) != 0)
203 					panic("swap_getapage: "
204 					    "page_relocate_cage failed");
205 
206 				pp = rpp;
207 			}
208 		}
209 
210 		if (pl) {
211 			if (upgrade)
212 				page_downgrade(pp);
213 
214 			pl[0] = pp;
215 			pl[1] = NULL;
216 		} else {
217 			page_unlock(pp);
218 		}
219 	} else {
220 		pp = page_create_va(vp, off, PAGESIZE,
221 		    PG_WAIT | PG_EXCL | flag_noreloc,
222 		    seg, addr);
223 		/*
224 		 * Someone raced in and created the page after we did the
225 		 * lookup but before we did the create, so go back and
226 		 * try to look it up again.
227 		 */
228 		if (pp == NULL)
229 			goto again;
230 		if (rw != S_CREATE) {
231 			err = swap_getphysname(vp, off, &pvp, &poff);
232 			if (pvp) {
233 				struct anon *ap;
234 				kmutex_t *ahm;
235 
236 				flags = (pl == NULL ? B_ASYNC|B_READ : B_READ);
237 				err = VOP_PAGEIO(pvp, pp, poff,
238 				    PAGESIZE, flags, cr);
239 
240 				if (!err) {
241 					ahm = &anonhash_lock[AH_LOCK(vp, off)];
242 					mutex_enter(ahm);
243 
244 					ap = swap_anon(vp, off);
245 					if (ap == NULL)
246 					    panic("swap_getapage: null anon");
247 
248 					if (ap->an_pvp == pvp &&
249 					    ap->an_poff == poff) {
250 						swap_phys_free(pvp, poff,
251 						    PAGESIZE);
252 						ap->an_pvp = NULL;
253 						ap->an_poff = NULL;
254 						hat_setmod(pp);
255 					}
256 
257 					mutex_exit(ahm);
258 				}
259 			} else {
260 				if (!err)
261 					pagezero(pp, 0, PAGESIZE);
262 
263 				/*
264 				 * If it's a fault ahead, release page_io_lock
265 				 * and SE_EXCL we grabbed in page_create_va
266 				 *
267 				 * If we are here, we haven't called VOP_PAGEIO
268 				 * and thus calling pvn_read_done(pp, B_READ)
269 				 * below may mislead that we tried i/o. Besides,
270 				 * in case of async, pvn_read_done() should
271 				 * not be called by *getpage()
272 				 */
273 				if (pl == NULL) {
274 					/*
275 					 * swap_getphysname can return error
276 					 * only when we are getting called from
277 					 * swapslot_free which passes non-NULL
278 					 * pl to VOP_GETPAGE.
279 					 */
280 					ASSERT(err == 0);
281 					page_io_unlock(pp);
282 					page_unlock(pp);
283 				}
284 			}
285 		}
286 
287 		ASSERT(pp != NULL);
288 
289 		if (err && pl)
290 			pvn_read_done(pp, B_ERROR);
291 
292 		if (!err && pl)
293 			pvn_plist_init(pp, pl, plsz, off, PAGESIZE, rw);
294 	}
295 	TRACE_3(TR_FAC_SWAPFS, TR_SWAPFS_GETAPAGE,
296 		"swapfs getapage:pp %p vp %p off %llx", pp, vp, off);
297 	return (err);
298 }
299 
300 /*
301  * Called from large page anon routines only! This is an ugly hack where
302  * the anon layer directly calls into swapfs with a preallocated large page.
303  * Another method would have been to change to VOP and add an extra arg for
304  * the preallocated large page. This all could be cleaned up later when we
305  * solve the anonymous naming problem and no longer need to loop across of
306  * the VOP in PAGESIZE increments to fill in or initialize a large page as
307  * is done today. I think the latter is better since it avoid a change to
308  * the VOP interface that could later be avoided.
309  */
310 int
311 swap_getconpage(
312 	struct vnode *vp,
313 	u_offset_t off,
314 	size_t len,
315 	uint_t *protp,
316 	page_t *pl[],
317 	size_t plsz,
318 	page_t	*conpp,
319 	spgcnt_t *nreloc,
320 	struct seg *seg,
321 	caddr_t addr,
322 	enum seg_rw rw,
323 	struct cred *cr)
324 {
325 	struct page	*pp;
326 	int 		err = 0;
327 	struct vnode	*pvp = NULL;
328 	u_offset_t	poff;
329 
330 	ASSERT(len == PAGESIZE);
331 	ASSERT(pl != NULL);
332 	ASSERT(plsz == PAGESIZE);
333 	ASSERT(protp == NULL);
334 	ASSERT(nreloc != NULL);
335 	ASSERT(!SEG_IS_SEGKP(seg)); /* XXX for now not supported */
336 	SWAPFS_PRINT(SWAP_VOPS, "swap_getconpage: vp %p, off %llx, len %lx\n",
337 		vp, off, len, 0, 0);
338 
339 	/*
340 	 * If we are not using a preallocated page then we know one already
341 	 * exists. So just let the old code handle it.
342 	 */
343 	if (conpp == NULL) {
344 		err = swap_getapage(vp, (u_offset_t)off, len, protp, pl, plsz,
345 		    seg, addr, rw, cr);
346 		return (err);
347 	}
348 	ASSERT(conpp->p_szc != 0);
349 	ASSERT(PAGE_EXCL(conpp));
350 
351 
352 	ASSERT(conpp->p_next == conpp);
353 	ASSERT(conpp->p_prev == conpp);
354 	ASSERT(!PP_ISAGED(conpp));
355 	ASSERT(!PP_ISFREE(conpp));
356 
357 	*nreloc = 0;
358 	pp = page_lookup_create(vp, off, SE_SHARED, conpp, nreloc, 0);
359 
360 	/*
361 	 * If existing page is found we may need to relocate.
362 	 */
363 	if (pp != conpp) {
364 		ASSERT(rw != S_CREATE);
365 		ASSERT(PAGE_SHARED(pp));
366 		if (pp->p_szc < conpp->p_szc) {
367 			page_unlock(pp);
368 			err = -1;
369 		} else if (pp->p_szc > conpp->p_szc) {
370 			page_unlock(pp);
371 			err = -2;
372 		} else {
373 			pl[0] = pp;
374 			pl[1] = NULL;
375 			if (page_pptonum(pp) &
376 			    (page_get_pagecnt(pp->p_szc) - 1))
377 			    cmn_err(CE_PANIC, "swap_getconpage: no root");
378 		}
379 		return (err);
380 	}
381 
382 	ASSERT(PAGE_EXCL(pp));
383 
384 	if (*nreloc != 0) {
385 		ASSERT(rw != S_CREATE);
386 		pl[0] = pp;
387 		pl[1] = NULL;
388 		return (0);
389 	}
390 
391 	*nreloc = 1;
392 
393 	/*
394 	 * If necessary do the page io.
395 	 */
396 	if (rw != S_CREATE) {
397 		/*
398 		 * Since we are only called now on behalf of an
399 		 * address space operation it's impossible for
400 		 * us to fail unlike swap_getapge() which
401 		 * also gets called from swapslot_free().
402 		 */
403 		if (swap_getphysname(vp, off, &pvp, &poff)) {
404 			cmn_err(CE_PANIC,
405 			    "swap_getconpage: swap_getphysname failed!");
406 		}
407 
408 		if (pvp) {
409 			err = VOP_PAGEIO(pvp, pp, poff, PAGESIZE, B_READ, cr);
410 		} else {
411 			pagezero(pp, 0, PAGESIZE);
412 		}
413 	}
414 
415 	/*
416 	 * Normally we would let pvn_read_done() destroy
417 	 * the page on IO error. But since this is a preallocated
418 	 * page we'll let the anon layer handle it.
419 	 */
420 	page_io_unlock(pp);
421 	if (err != 0)
422 		page_hashout(pp, NULL);
423 	ASSERT(pp->p_next == pp);
424 	ASSERT(pp->p_prev == pp);
425 
426 	TRACE_3(TR_FAC_SWAPFS, TR_SWAPFS_GETAPAGE,
427 		"swapfs getconpage:pp %p vp %p off %llx", pp, vp, off);
428 
429 	pl[0] = pp;
430 	pl[1] = NULL;
431 	return (err);
432 }
433 
434 /* Async putpage klustering stuff */
435 int sw_pending_size;
436 extern int klustsize;
437 extern struct async_reqs *sw_getreq();
438 extern void sw_putreq(struct async_reqs *);
439 extern void sw_putbackreq(struct async_reqs *);
440 extern struct async_reqs *sw_getfree();
441 extern void sw_putfree(struct async_reqs *);
442 
443 static size_t swap_putpagecnt, swap_pagespushed;
444 static size_t swap_otherfail, swap_otherpages;
445 static size_t swap_klustfail, swap_klustpages;
446 static size_t swap_getiofail, swap_getiopages;
447 
448 /*
449  * Flags are composed of {B_INVAL, B_DIRTY B_FREE, B_DONTNEED}.
450  * If len == 0, do from off to EOF.
451  */
452 static int swap_nopage = 0;	/* Don't do swap_putpage's if set */
453 
454 /* ARGSUSED */
455 static int
456 swap_putpage(
457 	struct vnode *vp,
458 	offset_t off,
459 	size_t len,
460 	int flags,
461 	struct cred *cr)
462 {
463 	page_t *pp;
464 	u_offset_t io_off;
465 	size_t io_len = 0;
466 	int err = 0;
467 	struct async_reqs *arg;
468 
469 	if (swap_nopage)
470 		return (0);
471 
472 	ASSERT(vp->v_count != 0);
473 
474 	SWAPFS_PRINT(SWAP_VOPS,
475 	    "swap_putpage: vp %p, off %llx len %lx, flags %x\n",
476 	    (void *)vp, off, len, flags, 0);
477 	TRACE_3(TR_FAC_SWAPFS, TR_SWAPFS_PUTPAGE,
478 	    "swapfs putpage:vp %p off %llx len %ld", (void *)vp, off, len);
479 
480 	if (vp->v_flag & VNOMAP)
481 		return (ENOSYS);
482 
483 	if (!vn_has_cached_data(vp))
484 		return (0);
485 
486 	if (len == 0) {
487 		if (curproc == proc_pageout)
488 			cmn_err(CE_PANIC, "swapfs: pageout can't block");
489 
490 		/* Search the entire vp list for pages >= off. */
491 		err = pvn_vplist_dirty(vp, (u_offset_t)off, swap_putapage,
492 		    flags, cr);
493 	} else {
494 		u_offset_t eoff;
495 
496 		/*
497 		 * Loop over all offsets in the range [off...off + len]
498 		 * looking for pages to deal with.
499 		 */
500 		eoff = off + len;
501 		for (io_off = (u_offset_t)off; io_off < eoff;
502 		    io_off += io_len) {
503 			/*
504 			 * If we run out of the async req slot, put the page
505 			 * now instead of queuing.
506 			 */
507 			if (flags == (B_ASYNC | B_FREE) &&
508 			    sw_pending_size < klustsize &&
509 			    (arg = sw_getfree())) {
510 				/*
511 				 * If we are clustering, we should allow
512 				 * pageout to feed us more pages because # of
513 				 * pushes is limited by # of I/Os, and one
514 				 * cluster is considered to be one I/O.
515 				 */
516 				if (pushes)
517 					pushes--;
518 
519 				arg->a_vp = vp;
520 				arg->a_off = io_off;
521 				arg->a_len = PAGESIZE;
522 				arg->a_flags = B_ASYNC | B_FREE;
523 				arg->a_cred = kcred;
524 				sw_putreq(arg);
525 				io_len = PAGESIZE;
526 				continue;
527 			}
528 			/*
529 			 * If we are not invalidating pages, use the
530 			 * routine page_lookup_nowait() to prevent
531 			 * reclaiming them from the free list.
532 			 */
533 			if ((flags & B_INVAL) ||
534 			    (flags & (B_ASYNC | B_FREE)) == B_FREE)
535 				pp = page_lookup(vp, io_off, SE_EXCL);
536 			else
537 				pp = page_lookup_nowait(vp, io_off,
538 					(flags & B_FREE) ? SE_EXCL : SE_SHARED);
539 
540 			if (pp == NULL || pvn_getdirty(pp, flags) == 0)
541 				io_len = PAGESIZE;
542 			else {
543 				err = swap_putapage(vp, pp, &io_off, &io_len,
544 				    flags, cr);
545 				if (err != 0)
546 					break;
547 			}
548 		}
549 	}
550 	/* If invalidating, verify all pages on vnode list are gone. */
551 	if (err == 0 && off == 0 && len == 0 &&
552 	    (flags & B_INVAL) && vn_has_cached_data(vp)) {
553 		cmn_err(CE_WARN,
554 		    "swap_putpage: B_INVAL, pages not gone");
555 	}
556 	return (err);
557 }
558 
559 /*
560  * Write out a single page.
561  * For swapfs this means choose a physical swap slot and write the page
562  * out using VOP_PAGEIO.
563  * In the (B_ASYNC | B_FREE) case we try to find a bunch of other dirty
564  * swapfs pages, a bunch of contiguous swap slots and then write them
565  * all out in one clustered i/o.
566  */
567 /*ARGSUSED*/
568 static int
569 swap_putapage(
570 	struct vnode *vp,
571 	page_t *pp,
572 	u_offset_t *offp,
573 	size_t *lenp,
574 	int flags,
575 	struct cred *cr)
576 {
577 	int err;
578 	struct vnode *pvp;
579 	u_offset_t poff, off;
580 	u_offset_t doff;
581 	size_t dlen;
582 	size_t klsz = 0;
583 	u_offset_t klstart = 0;
584 	struct vnode *klvp = NULL;
585 	page_t *pplist;
586 	se_t se;
587 	struct async_reqs *arg;
588 	size_t swap_klustsize;
589 
590 	/*
591 	 * This check is added for callers who access swap_putpage with len = 0.
592 	 * swap_putpage calls swap_putapage page-by-page via pvn_vplist_dirty.
593 	 * And it's necessary to do the same queuing if users have the same
594 	 * B_ASYNC|B_FREE flags on.
595 	 */
596 	if (flags == (B_ASYNC | B_FREE) &&
597 	    sw_pending_size < klustsize && (arg = sw_getfree())) {
598 
599 		hat_setmod(pp);
600 		page_io_unlock(pp);
601 		page_unlock(pp);
602 
603 		arg->a_vp = vp;
604 		arg->a_off = pp->p_offset;
605 		arg->a_len = PAGESIZE;
606 		arg->a_flags = B_ASYNC | B_FREE;
607 		arg->a_cred = kcred;
608 		sw_putreq(arg);
609 
610 		return (0);
611 	}
612 
613 	SWAPFS_PRINT(SWAP_PUTP,
614 		"swap_putapage: pp %p, vp %p, off %llx, flags %x\n",
615 		pp, vp, pp->p_offset, flags, 0);
616 
617 	ASSERT(PAGE_LOCKED(pp));
618 
619 	off = pp->p_offset;
620 
621 	doff = off;
622 	dlen = PAGESIZE;
623 
624 	if (err = swap_newphysname(vp, off, &doff, &dlen, &pvp, &poff)) {
625 		err = (flags == (B_ASYNC | B_FREE) ? ENOMEM : 0);
626 		hat_setmod(pp);
627 		page_io_unlock(pp);
628 		page_unlock(pp);
629 		goto out;
630 	}
631 
632 	klvp = pvp;
633 	klstart = poff;
634 	pplist = pp;
635 	/*
636 	 * If this is ASYNC | FREE and we've accumulated a bunch of such
637 	 * pending requests, kluster.
638 	 */
639 	if (flags == (B_ASYNC | B_FREE))
640 		swap_klustsize = klustsize;
641 	else
642 		swap_klustsize = PAGESIZE;
643 	se = (flags & B_FREE ? SE_EXCL : SE_SHARED);
644 	klsz = PAGESIZE;
645 	while (klsz < swap_klustsize) {
646 		if ((arg = sw_getreq()) == NULL) {
647 			swap_getiofail++;
648 			swap_getiopages += btop(klsz);
649 			break;
650 		}
651 		ASSERT(vn_matchops(arg->a_vp, swap_vnodeops));
652 		vp = arg->a_vp;
653 		off = arg->a_off;
654 
655 		if ((pp = page_lookup_nowait(vp, off, se)) == NULL) {
656 			swap_otherfail++;
657 			swap_otherpages += btop(klsz);
658 			sw_putfree(arg);
659 			break;
660 		}
661 		if (pvn_getdirty(pp, flags | B_DELWRI) == 0) {
662 			sw_putfree(arg);
663 			continue;
664 		}
665 		/* Get new physical backing store for the page */
666 		doff = off;
667 		dlen = PAGESIZE;
668 		if (err = swap_newphysname(vp, off, &doff, &dlen,
669 						&pvp, &poff)) {
670 			swap_otherfail++;
671 			swap_otherpages += btop(klsz);
672 			hat_setmod(pp);
673 			page_io_unlock(pp);
674 			page_unlock(pp);
675 			sw_putbackreq(arg);
676 			break;
677 		}
678 		/* Try to cluster new physical name with previous ones */
679 		if (klvp == pvp && poff == klstart + klsz) {
680 			klsz += PAGESIZE;
681 			page_add(&pplist, pp);
682 			pplist = pplist->p_next;
683 			sw_putfree(arg);
684 		} else if (klvp == pvp && poff == klstart - PAGESIZE) {
685 			klsz += PAGESIZE;
686 			klstart -= PAGESIZE;
687 			page_add(&pplist, pp);
688 			sw_putfree(arg);
689 		} else {
690 			swap_klustfail++;
691 			swap_klustpages += btop(klsz);
692 			hat_setmod(pp);
693 			page_io_unlock(pp);
694 			page_unlock(pp);
695 			sw_putbackreq(arg);
696 			break;
697 		}
698 	}
699 
700 	err = VOP_PAGEIO(klvp, pplist, klstart, klsz,
701 		    B_WRITE | flags, cr);
702 
703 	if ((flags & B_ASYNC) == 0)
704 		pvn_write_done(pp, ((err) ? B_ERROR : 0) | B_WRITE | flags);
705 
706 	/* Statistics */
707 	if (!err) {
708 		swap_putpagecnt++;
709 		swap_pagespushed += btop(klsz);
710 	}
711 out:
712 	TRACE_4(TR_FAC_SWAPFS, TR_SWAPFS_PUTAPAGE,
713 		"swapfs putapage:vp %p klvp %p, klstart %lx, klsz %lx",
714 		vp, klvp, klstart, klsz);
715 	if (err && err != ENOMEM)
716 		cmn_err(CE_WARN, "swapfs_putapage: err %d\n", err);
717 	if (lenp)
718 		*lenp = PAGESIZE;
719 	return (err);
720 }
721 
722 static void
723 swap_dispose(vnode_t *vp, page_t *pp, int fl, int dn, cred_t *cr)
724 {
725 	int err;
726 	u_offset_t off = pp->p_offset;
727 	vnode_t *pvp;
728 	u_offset_t poff;
729 
730 	ASSERT(PAGE_EXCL(pp));
731 
732 	/*
733 	 * The caller will free/invalidate large page in one shot instead of
734 	 * one small page at a time.
735 	 */
736 	if (pp->p_szc != 0) {
737 		page_unlock(pp);
738 		return;
739 	}
740 
741 	err = swap_getphysname(vp, off, &pvp, &poff);
742 	if (!err && pvp != NULL)
743 		VOP_DISPOSE(pvp, pp, fl, dn, cr);
744 	else
745 		fs_dispose(vp, pp, fl, dn, cr);
746 }
747