xref: /netbsd-src/sys/ufs/lfs/lfs_pages.c (revision d909946ca08dceb44d7d0f22ec9488679695d976)
1 /*	$NetBSD: lfs_pages.c,v 1.8 2016/07/21 18:10:47 christos Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2001, 2002, 2003 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Konrad E. Schroder <perseant@hhhh.org>.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 /*
32  * Copyright (c) 1986, 1989, 1991, 1993, 1995
33  *	The Regents of the University of California.  All rights reserved.
34  *
35  * Redistribution and use in source and binary forms, with or without
36  * modification, are permitted provided that the following conditions
37  * are met:
38  * 1. Redistributions of source code must retain the above copyright
39  *    notice, this list of conditions and the following disclaimer.
40  * 2. Redistributions in binary form must reproduce the above copyright
41  *    notice, this list of conditions and the following disclaimer in the
42  *    documentation and/or other materials provided with the distribution.
43  * 3. Neither the name of the University nor the names of its contributors
44  *    may be used to endorse or promote products derived from this software
45  *    without specific prior written permission.
46  *
47  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
48  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
49  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
50  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
51  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
52  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
53  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
54  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
55  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
56  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
57  * SUCH DAMAGE.
58  *
59  *	@(#)lfs_vnops.c	8.13 (Berkeley) 6/10/95
60  */
61 
62 #include <sys/cdefs.h>
63 __KERNEL_RCSID(0, "$NetBSD: lfs_pages.c,v 1.8 2016/07/21 18:10:47 christos Exp $");
64 
65 #ifdef _KERNEL_OPT
66 #include "opt_compat_netbsd.h"
67 #include "opt_uvm_page_trkown.h"
68 #endif
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/namei.h>
73 #include <sys/resourcevar.h>
74 #include <sys/kernel.h>
75 #include <sys/file.h>
76 #include <sys/stat.h>
77 #include <sys/buf.h>
78 #include <sys/proc.h>
79 #include <sys/mount.h>
80 #include <sys/vnode.h>
81 #include <sys/pool.h>
82 #include <sys/signalvar.h>
83 #include <sys/kauth.h>
84 #include <sys/syslog.h>
85 #include <sys/fstrans.h>
86 
87 #include <miscfs/fifofs/fifo.h>
88 #include <miscfs/genfs/genfs.h>
89 #include <miscfs/specfs/specdev.h>
90 
91 #include <ufs/lfs/ulfs_inode.h>
92 #include <ufs/lfs/ulfsmount.h>
93 #include <ufs/lfs/ulfs_bswap.h>
94 #include <ufs/lfs/ulfs_extern.h>
95 
96 #include <uvm/uvm.h>
97 #include <uvm/uvm_pmap.h>
98 #include <uvm/uvm_stat.h>
99 #include <uvm/uvm_pager.h>
100 
101 #include <ufs/lfs/lfs.h>
102 #include <ufs/lfs/lfs_accessors.h>
103 #include <ufs/lfs/lfs_kernel.h>
104 #include <ufs/lfs/lfs_extern.h>
105 
106 extern pid_t lfs_writer_daemon;
107 
108 static int check_dirty(struct lfs *, struct vnode *, off_t, off_t, off_t, int, int, struct vm_page **);
109 
110 int
111 lfs_getpages(void *v)
112 {
113 	struct vop_getpages_args /* {
114 		struct vnode *a_vp;
115 		voff_t a_offset;
116 		struct vm_page **a_m;
117 		int *a_count;
118 		int a_centeridx;
119 		vm_prot_t a_access_type;
120 		int a_advice;
121 		int a_flags;
122 	} */ *ap = v;
123 
124 	if (VTOI(ap->a_vp)->i_number == LFS_IFILE_INUM &&
125 	    (ap->a_access_type & VM_PROT_WRITE) != 0) {
126 		return EPERM;
127 	}
128 	if ((ap->a_access_type & VM_PROT_WRITE) != 0) {
129 		mutex_enter(&lfs_lock);
130 		LFS_SET_UINO(VTOI(ap->a_vp), IN_MODIFIED);
131 		mutex_exit(&lfs_lock);
132 	}
133 
134 	/*
135 	 * we're relying on the fact that genfs_getpages() always read in
136 	 * entire filesystem blocks.
137 	 */
138 	return genfs_getpages(v);
139 }
140 
141 /*
142  * Wait for a page to become unbusy, possibly printing diagnostic messages
143  * as well.
144  *
145  * Called with vp->v_interlock held; return with it held.
146  */
147 static void
148 wait_for_page(struct vnode *vp, struct vm_page *pg, const char *label)
149 {
150 	KASSERT(mutex_owned(vp->v_interlock));
151 	if ((pg->flags & PG_BUSY) == 0)
152 		return;		/* Nothing to wait for! */
153 
154 #if defined(DEBUG) && defined(UVM_PAGE_TRKOWN)
155 	static struct vm_page *lastpg;
156 
157 	if (label != NULL && pg != lastpg) {
158 		if (pg->owner_tag) {
159 			printf("lfs_putpages[%d.%d]: %s: page %p owner %d.%d [%s]\n",
160 			       curproc->p_pid, curlwp->l_lid, label,
161 			       pg, pg->owner, pg->lowner, pg->owner_tag);
162 		} else {
163 			printf("lfs_putpages[%d.%d]: %s: page %p unowned?!\n",
164 			       curproc->p_pid, curlwp->l_lid, label, pg);
165 		}
166 	}
167 	lastpg = pg;
168 #endif
169 
170 	pg->flags |= PG_WANTED;
171 	UVM_UNLOCK_AND_WAIT(pg, vp->v_interlock, 0, "lfsput", 0);
172 	mutex_enter(vp->v_interlock);
173 }
174 
175 /*
176  * This routine is called by lfs_putpages() when it can't complete the
177  * write because a page is busy.  This means that either (1) someone,
178  * possibly the pagedaemon, is looking at this page, and will give it up
179  * presently; or (2) we ourselves are holding the page busy in the
180  * process of being written (either gathered or actually on its way to
181  * disk).  We don't need to give up the segment lock, but we might need
182  * to call lfs_writeseg() to expedite the page's journey to disk.
183  *
184  * Called with vp->v_interlock held; return with it held.
185  */
186 /* #define BUSYWAIT */
187 static void
188 write_and_wait(struct lfs *fs, struct vnode *vp, struct vm_page *pg,
189 	       int seglocked, const char *label)
190 {
191 	KASSERT(mutex_owned(vp->v_interlock));
192 #ifndef BUSYWAIT
193 	struct inode *ip = VTOI(vp);
194 	struct segment *sp = fs->lfs_sp;
195 	int count = 0;
196 
197 	if (pg == NULL)
198 		return;
199 
200 	while (pg->flags & PG_BUSY &&
201 	    pg->uobject == &vp->v_uobj) {
202 		mutex_exit(vp->v_interlock);
203 		if (sp->cbpp - sp->bpp > 1) {
204 			/* Write gathered pages */
205 			lfs_updatemeta(sp);
206 			lfs_release_finfo(fs);
207 			(void) lfs_writeseg(fs, sp);
208 
209 			/*
210 			 * Reinitialize FIP
211 			 */
212 			KASSERT(sp->vp == vp);
213 			lfs_acquire_finfo(fs, ip->i_number,
214 					  ip->i_gen);
215 		}
216 		++count;
217 		mutex_enter(vp->v_interlock);
218 		wait_for_page(vp, pg, label);
219 	}
220 	if (label != NULL && count > 1) {
221 		DLOG((DLOG_PAGE, "lfs_putpages[%d]: %s: %sn = %d\n",
222 		      curproc->p_pid, label, (count > 0 ? "looping, " : ""),
223 		      count));
224 	}
225 #else
226 	preempt(1);
227 #endif
228 	KASSERT(mutex_owned(vp->v_interlock));
229 }
230 
231 /*
232  * Make sure that for all pages in every block in the given range,
233  * either all are dirty or all are clean.  If any of the pages
234  * we've seen so far are dirty, put the vnode on the paging chain,
235  * and mark it IN_PAGING.
236  *
237  * If checkfirst != 0, don't check all the pages but return at the
238  * first dirty page.
239  */
240 static int
241 check_dirty(struct lfs *fs, struct vnode *vp,
242 	    off_t startoffset, off_t endoffset, off_t blkeof,
243 	    int flags, int checkfirst, struct vm_page **pgp)
244 {
245 	int by_list;
246 	struct vm_page *curpg = NULL; /* XXX: gcc */
247 	struct vm_page *pgs[MAXBSIZE /
248 	    (__builtin_constant_p(PAGE_SIZE) ? PAGE_SIZE : 1024)], *pg;
249 	off_t soff = 0; /* XXX: gcc */
250 	voff_t off;
251 	int i;
252 	int nonexistent;
253 	int any_dirty;	/* number of dirty pages */
254 	int dirty;	/* number of dirty pages in a block */
255 	int tdirty;
256 	int pages_per_block = lfs_sb_getbsize(fs) >> PAGE_SHIFT;
257 	int pagedaemon = (curlwp == uvm.pagedaemon_lwp);
258 
259 	KASSERT(mutex_owned(vp->v_interlock));
260 	ASSERT_MAYBE_SEGLOCK(fs);
261   top:
262 	by_list = (vp->v_uobj.uo_npages <=
263 		   ((endoffset - startoffset) >> PAGE_SHIFT) *
264 		   UVM_PAGE_TREE_PENALTY);
265 	any_dirty = 0;
266 
267 	if (by_list) {
268 		curpg = TAILQ_FIRST(&vp->v_uobj.memq);
269 	} else {
270 		soff = startoffset;
271 	}
272 	while (by_list || soff < MIN(blkeof, endoffset)) {
273 		if (by_list) {
274 			/*
275 			 * Find the first page in a block.  Skip
276 			 * blocks outside our area of interest or beyond
277 			 * the end of file.
278 			 */
279 			KASSERT(curpg == NULL
280 			    || (curpg->flags & PG_MARKER) == 0);
281 			if (pages_per_block > 1) {
282 				while (curpg &&
283 				    ((curpg->offset & lfs_sb_getbmask(fs)) ||
284 				    curpg->offset >= vp->v_size ||
285 				    curpg->offset >= endoffset)) {
286 					curpg = TAILQ_NEXT(curpg, listq.queue);
287 					KASSERT(curpg == NULL ||
288 					    (curpg->flags & PG_MARKER) == 0);
289 				}
290 			}
291 			if (curpg == NULL)
292 				break;
293 			soff = curpg->offset;
294 		}
295 
296 		/*
297 		 * Mark all pages in extended range busy; find out if any
298 		 * of them are dirty.
299 		 */
300 		nonexistent = dirty = 0;
301 		for (i = 0; i == 0 || i < pages_per_block; i++) {
302 			KASSERT(mutex_owned(vp->v_interlock));
303 			if (by_list && pages_per_block <= 1) {
304 				pgs[i] = pg = curpg;
305 			} else {
306 				off = soff + (i << PAGE_SHIFT);
307 				pgs[i] = pg = uvm_pagelookup(&vp->v_uobj, off);
308 				if (pg == NULL) {
309 					++nonexistent;
310 					continue;
311 				}
312 			}
313 			KASSERT(pg != NULL);
314 
315 			/*
316 			 * If we're holding the segment lock, we can deadlock
317 			 * against a process that has our page and is waiting
318 			 * for the cleaner, while the cleaner waits for the
319 			 * segment lock.  Just bail in that case.
320 			 */
321 			if ((pg->flags & PG_BUSY) &&
322 			    (pagedaemon || LFS_SEGLOCK_HELD(fs))) {
323 				if (i > 0)
324 					uvm_page_unbusy(pgs, i);
325 				DLOG((DLOG_PAGE, "lfs_putpages: avoiding 3-way or pagedaemon deadlock\n"));
326 				if (pgp)
327 					*pgp = pg;
328 				KASSERT(mutex_owned(vp->v_interlock));
329 				return -1;
330 			}
331 
332 			while (pg->flags & PG_BUSY) {
333 				wait_for_page(vp, pg, NULL);
334 				KASSERT(mutex_owned(vp->v_interlock));
335 				if (i > 0)
336 					uvm_page_unbusy(pgs, i);
337 				KASSERT(mutex_owned(vp->v_interlock));
338 				goto top;
339 			}
340 			pg->flags |= PG_BUSY;
341 			UVM_PAGE_OWN(pg, "lfs_putpages");
342 
343 			pmap_page_protect(pg, VM_PROT_NONE);
344 			tdirty = (pmap_clear_modify(pg) ||
345 				  (pg->flags & PG_CLEAN) == 0);
346 			dirty += tdirty;
347 		}
348 		if (pages_per_block > 0 && nonexistent >= pages_per_block) {
349 			if (by_list) {
350 				curpg = TAILQ_NEXT(curpg, listq.queue);
351 			} else {
352 				soff += lfs_sb_getbsize(fs);
353 			}
354 			continue;
355 		}
356 
357 		any_dirty += dirty;
358 		KASSERT(nonexistent == 0);
359 		KASSERT(mutex_owned(vp->v_interlock));
360 
361 		/*
362 		 * If any are dirty make all dirty; unbusy them,
363 		 * but if we were asked to clean, wire them so that
364 		 * the pagedaemon doesn't bother us about them while
365 		 * they're on their way to disk.
366 		 */
367 		for (i = 0; i == 0 || i < pages_per_block; i++) {
368 			KASSERT(mutex_owned(vp->v_interlock));
369 			pg = pgs[i];
370 			KASSERT(!((pg->flags & PG_CLEAN) && (pg->flags & PG_DELWRI)));
371 			KASSERT(pg->flags & PG_BUSY);
372 			if (dirty) {
373 				pg->flags &= ~PG_CLEAN;
374 				if (flags & PGO_FREE) {
375 					/*
376 					 * Wire the page so that
377 					 * pdaemon doesn't see it again.
378 					 */
379 					mutex_enter(&uvm_pageqlock);
380 					uvm_pagewire(pg);
381 					mutex_exit(&uvm_pageqlock);
382 
383 					/* Suspended write flag */
384 					pg->flags |= PG_DELWRI;
385 				}
386 			}
387 			if (pg->flags & PG_WANTED)
388 				wakeup(pg);
389 			pg->flags &= ~(PG_WANTED|PG_BUSY);
390 			UVM_PAGE_OWN(pg, NULL);
391 		}
392 
393 		if (checkfirst && any_dirty)
394 			break;
395 
396 		if (by_list) {
397 			curpg = TAILQ_NEXT(curpg, listq.queue);
398 		} else {
399 			soff += MAX(PAGE_SIZE, lfs_sb_getbsize(fs));
400 		}
401 	}
402 
403 	KASSERT(mutex_owned(vp->v_interlock));
404 	return any_dirty;
405 }
406 
407 /*
408  * lfs_putpages functions like genfs_putpages except that
409  *
410  * (1) It needs to bounds-check the incoming requests to ensure that
411  *     they are block-aligned; if they are not, expand the range and
412  *     do the right thing in case, e.g., the requested range is clean
413  *     but the expanded range is dirty.
414  *
415  * (2) It needs to explicitly send blocks to be written when it is done.
416  *     If VOP_PUTPAGES is called without the seglock held, we simply take
417  *     the seglock and let lfs_segunlock wait for us.
418  *     XXX There might be a bad situation if we have to flush a vnode while
419  *     XXX lfs_markv is in operation.  As of this writing we panic in this
420  *     XXX case.
421  *
422  * Assumptions:
423  *
424  * (1) The caller does not hold any pages in this vnode busy.  If it does,
425  *     there is a danger that when we expand the page range and busy the
426  *     pages we will deadlock.
427  *
428  * (2) We are called with vp->v_interlock held; we must return with it
429  *     released.
430  *
431  * (3) We don't absolutely have to free pages right away, provided that
432  *     the request does not have PGO_SYNCIO.  When the pagedaemon gives
433  *     us a request with PGO_FREE, we take the pages out of the paging
434  *     queue and wake up the writer, which will handle freeing them for us.
435  *
436  *     We ensure that for any filesystem block, all pages for that
437  *     block are either resident or not, even if those pages are higher
438  *     than EOF; that means that we will be getting requests to free
439  *     "unused" pages above EOF all the time, and should ignore them.
440  *
441  * (4) If we are called with PGO_LOCKED, the finfo array we are to write
442  *     into has been set up for us by lfs_writefile.  If not, we will
443  *     have to handle allocating and/or freeing an finfo entry.
444  *
445  * XXX note that we're (ab)using PGO_LOCKED as "seglock held".
446  */
447 
448 /* How many times to loop before we should start to worry */
449 #define TOOMANY 4
450 
451 int
452 lfs_putpages(void *v)
453 {
454 	int error;
455 	struct vop_putpages_args /* {
456 		struct vnode *a_vp;
457 		voff_t a_offlo;
458 		voff_t a_offhi;
459 		int a_flags;
460 	} */ *ap = v;
461 	struct vnode *vp;
462 	struct inode *ip;
463 	struct lfs *fs;
464 	struct segment *sp;
465 	off_t origoffset, startoffset, endoffset, origendoffset, blkeof;
466 	off_t off, max_endoffset;
467 	bool seglocked, sync, pagedaemon, reclaim;
468 	struct vm_page *pg, *busypg;
469 	UVMHIST_FUNC("lfs_putpages"); UVMHIST_CALLED(ubchist);
470 	int oreclaim = 0;
471 	int donewriting = 0;
472 #ifdef DEBUG
473 	int debug_n_again, debug_n_dirtyclean;
474 #endif
475 
476 	vp = ap->a_vp;
477 	ip = VTOI(vp);
478 	fs = ip->i_lfs;
479 	sync = (ap->a_flags & PGO_SYNCIO) != 0;
480 	reclaim = (ap->a_flags & PGO_RECLAIM) != 0;
481 	pagedaemon = (curlwp == uvm.pagedaemon_lwp);
482 
483 	KASSERT(mutex_owned(vp->v_interlock));
484 
485 	/* Putpages does nothing for metadata. */
486 	if (vp == fs->lfs_ivnode || vp->v_type != VREG) {
487 		mutex_exit(vp->v_interlock);
488 		return 0;
489 	}
490 
491 	/*
492 	 * If there are no pages, don't do anything.
493 	 */
494 	if (vp->v_uobj.uo_npages == 0) {
495 		if (TAILQ_EMPTY(&vp->v_uobj.memq) &&
496 		    (vp->v_iflag & VI_ONWORKLST) &&
497 		    LIST_FIRST(&vp->v_dirtyblkhd) == NULL) {
498 			vp->v_iflag &= ~VI_WRMAPDIRTY;
499 			vn_syncer_remove_from_worklist(vp);
500 		}
501 		mutex_exit(vp->v_interlock);
502 
503 		/* Remove us from paging queue, if we were on it */
504 		mutex_enter(&lfs_lock);
505 		if (ip->i_flags & IN_PAGING) {
506 			ip->i_flags &= ~IN_PAGING;
507 			TAILQ_REMOVE(&fs->lfs_pchainhd, ip, i_lfs_pchain);
508 		}
509 		mutex_exit(&lfs_lock);
510 
511 		KASSERT(!mutex_owned(vp->v_interlock));
512 		return 0;
513 	}
514 
515 	blkeof = lfs_blkroundup(fs, ip->i_size);
516 
517 	/*
518 	 * Ignore requests to free pages past EOF but in the same block
519 	 * as EOF, unless the vnode is being reclaimed or the request
520 	 * is synchronous.  (If the request is sync, it comes from
521 	 * lfs_truncate.)
522 	 *
523 	 * To avoid being flooded with this request, make these pages
524 	 * look "active".
525 	 */
526 	if (!sync && !reclaim &&
527 	    ap->a_offlo >= ip->i_size && ap->a_offlo < blkeof) {
528 		origoffset = ap->a_offlo;
529 		for (off = origoffset; off < blkeof; off += lfs_sb_getbsize(fs)) {
530 			pg = uvm_pagelookup(&vp->v_uobj, off);
531 			KASSERT(pg != NULL);
532 			while (pg->flags & PG_BUSY) {
533 				pg->flags |= PG_WANTED;
534 				UVM_UNLOCK_AND_WAIT(pg, vp->v_interlock, 0,
535 						    "lfsput2", 0);
536 				mutex_enter(vp->v_interlock);
537 			}
538 			mutex_enter(&uvm_pageqlock);
539 			uvm_pageactivate(pg);
540 			mutex_exit(&uvm_pageqlock);
541 		}
542 		ap->a_offlo = blkeof;
543 		if (ap->a_offhi > 0 && ap->a_offhi <= ap->a_offlo) {
544 			mutex_exit(vp->v_interlock);
545 			return 0;
546 		}
547 	}
548 
549 	/*
550 	 * Extend page range to start and end at block boundaries.
551 	 * (For the purposes of VOP_PUTPAGES, fragments don't exist.)
552 	 */
553 	origoffset = ap->a_offlo;
554 	origendoffset = ap->a_offhi;
555 	startoffset = origoffset & ~(lfs_sb_getbmask(fs));
556 	max_endoffset = (trunc_page(LLONG_MAX) >> lfs_sb_getbshift(fs))
557 					       << lfs_sb_getbshift(fs);
558 
559 	if (origendoffset == 0 || ap->a_flags & PGO_ALLPAGES) {
560 		endoffset = max_endoffset;
561 		origendoffset = endoffset;
562 	} else {
563 		origendoffset = round_page(ap->a_offhi);
564 		endoffset = round_page(lfs_blkroundup(fs, origendoffset));
565 	}
566 
567 	KASSERT(startoffset > 0 || endoffset >= startoffset);
568 	if (startoffset == endoffset) {
569 		/* Nothing to do, why were we called? */
570 		mutex_exit(vp->v_interlock);
571 		DLOG((DLOG_PAGE, "lfs_putpages: startoffset = endoffset = %"
572 		      PRId64 "\n", startoffset));
573 		return 0;
574 	}
575 
576 	ap->a_offlo = startoffset;
577 	ap->a_offhi = endoffset;
578 
579 	/*
580 	 * If not cleaning, just send the pages through genfs_putpages
581 	 * to be returned to the pool.
582 	 */
583 	if (!(ap->a_flags & PGO_CLEANIT)) {
584 		DLOG((DLOG_PAGE, "lfs_putpages: no cleanit vn %p ino %d (flags %x)\n",
585 		      vp, (int)ip->i_number, ap->a_flags));
586 		int r = genfs_putpages(v);
587 		KASSERT(!mutex_owned(vp->v_interlock));
588 		return r;
589 	}
590 
591 	/* Set PGO_BUSYFAIL to avoid deadlocks */
592 	ap->a_flags |= PGO_BUSYFAIL;
593 
594 	/*
595 	 * Likewise, if we are asked to clean but the pages are not
596 	 * dirty, we can just free them using genfs_putpages.
597 	 */
598 #ifdef DEBUG
599 	debug_n_dirtyclean = 0;
600 #endif
601 	do {
602 		int r;
603 		KASSERT(mutex_owned(vp->v_interlock));
604 
605 		/* Count the number of dirty pages */
606 		r = check_dirty(fs, vp, startoffset, endoffset, blkeof,
607 				ap->a_flags, 1, NULL);
608 		if (r < 0) {
609 			/* Pages are busy with another process */
610 			mutex_exit(vp->v_interlock);
611 			return EDEADLK;
612 		}
613 		if (r > 0) /* Some pages are dirty */
614 			break;
615 
616 		/*
617 		 * Sometimes pages are dirtied between the time that
618 		 * we check and the time we try to clean them.
619 		 * Instruct lfs_gop_write to return EDEADLK in this case
620 		 * so we can write them properly.
621 		 */
622 		ip->i_lfs_iflags |= LFSI_NO_GOP_WRITE;
623 		r = genfs_do_putpages(vp, startoffset, endoffset,
624 				       ap->a_flags & ~PGO_SYNCIO, &busypg);
625 		ip->i_lfs_iflags &= ~LFSI_NO_GOP_WRITE;
626 		if (r != EDEADLK) {
627 			KASSERT(!mutex_owned(vp->v_interlock));
628  			return r;
629 		}
630 
631 		/* One of the pages was busy.  Start over. */
632 		mutex_enter(vp->v_interlock);
633 		wait_for_page(vp, busypg, "dirtyclean");
634 #ifdef DEBUG
635 		++debug_n_dirtyclean;
636 #endif
637 	} while(1);
638 
639 #ifdef DEBUG
640 	if (debug_n_dirtyclean > TOOMANY)
641 		DLOG((DLOG_PAGE, "lfs_putpages: dirtyclean: looping, n = %d\n",
642 		      debug_n_dirtyclean));
643 #endif
644 
645 	/*
646 	 * Dirty and asked to clean.
647 	 *
648 	 * Pagedaemon can't actually write LFS pages; wake up
649 	 * the writer to take care of that.  The writer will
650 	 * notice the pager inode queue and act on that.
651 	 *
652 	 * XXX We must drop the vp->interlock before taking the lfs_lock or we
653 	 * get a nasty deadlock with lfs_flush_pchain().
654 	 */
655 	if (pagedaemon) {
656 		mutex_exit(vp->v_interlock);
657 		mutex_enter(&lfs_lock);
658 		if (!(ip->i_flags & IN_PAGING)) {
659 			ip->i_flags |= IN_PAGING;
660 			TAILQ_INSERT_TAIL(&fs->lfs_pchainhd, ip, i_lfs_pchain);
661 		}
662 		wakeup(&lfs_writer_daemon);
663 		mutex_exit(&lfs_lock);
664 		preempt();
665 		KASSERT(!mutex_owned(vp->v_interlock));
666 		return EWOULDBLOCK;
667 	}
668 
669 	/*
670 	 * If this is a file created in a recent dirop, we can't flush its
671 	 * inode until the dirop is complete.  Drain dirops, then flush the
672 	 * filesystem (taking care of any other pending dirops while we're
673 	 * at it).
674 	 */
675 	if ((ap->a_flags & (PGO_CLEANIT|PGO_LOCKED)) == PGO_CLEANIT &&
676 	    (vp->v_uflag & VU_DIROP)) {
677 		DLOG((DLOG_PAGE, "lfs_putpages: flushing VU_DIROP\n"));
678 
679  		lfs_writer_enter(fs, "ppdirop");
680 
681 		/* Note if we hold the vnode locked */
682 		if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
683 		{
684 		    DLOG((DLOG_PAGE, "lfs_putpages: dirop inode already locked\n"));
685 		} else {
686 		    DLOG((DLOG_PAGE, "lfs_putpages: dirop inode not locked\n"));
687 		}
688 		mutex_exit(vp->v_interlock);
689 
690 		mutex_enter(&lfs_lock);
691 		lfs_flush_fs(fs, sync ? SEGM_SYNC : 0);
692 		mutex_exit(&lfs_lock);
693 
694 		mutex_enter(vp->v_interlock);
695 		lfs_writer_leave(fs);
696 
697 		/* The flush will have cleaned out this vnode as well,
698 		   no need to do more to it. */
699 	}
700 
701 	/*
702 	 * This is it.	We are going to write some pages.  From here on
703 	 * down it's all just mechanics.
704 	 *
705 	 * Don't let genfs_putpages wait; lfs_segunlock will wait for us.
706 	 */
707 	ap->a_flags &= ~PGO_SYNCIO;
708 
709 	/*
710 	 * If we've already got the seglock, flush the node and return.
711 	 * The FIP has already been set up for us by lfs_writefile,
712 	 * and FIP cleanup and lfs_updatemeta will also be done there,
713 	 * unless genfs_putpages returns EDEADLK; then we must flush
714 	 * what we have, and correct FIP and segment header accounting.
715 	 */
716   get_seglock:
717 	/*
718 	 * If we are not called with the segment locked, lock it.
719 	 * Account for a new FIP in the segment header, and set sp->vp.
720 	 * (This should duplicate the setup at the top of lfs_writefile().)
721 	 */
722 	seglocked = (ap->a_flags & PGO_LOCKED) != 0;
723 	if (!seglocked) {
724 		mutex_exit(vp->v_interlock);
725 		error = lfs_seglock(fs, SEGM_PROT | (sync ? SEGM_SYNC : 0));
726 		if (error != 0) {
727 			KASSERT(!mutex_owned(vp->v_interlock));
728  			return error;
729 		}
730 		mutex_enter(vp->v_interlock);
731 		lfs_acquire_finfo(fs, ip->i_number, ip->i_gen);
732 	}
733 	sp = fs->lfs_sp;
734 	KASSERT(sp->vp == NULL);
735 	sp->vp = vp;
736 
737 	/* Note segments written by reclaim; only for debugging */
738 	if (vdead_check(vp, VDEAD_NOWAIT) != 0) {
739 		sp->seg_flags |= SEGM_RECLAIM;
740 		fs->lfs_reclino = ip->i_number;
741 	}
742 
743 	/*
744 	 * Ensure that the partial segment is marked SS_DIROP if this
745 	 * vnode is a DIROP.
746 	 */
747 	if (!seglocked && vp->v_uflag & VU_DIROP) {
748 		SEGSUM *ssp = sp->segsum;
749 
750 		lfs_ss_setflags(fs, ssp,
751 				lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT));
752 	}
753 
754 	/*
755 	 * Loop over genfs_putpages until all pages are gathered.
756 	 * genfs_putpages() drops the interlock, so reacquire it if necessary.
757 	 * Whenever we lose the interlock we have to rerun check_dirty, as
758 	 * well, since more pages might have been dirtied in our absence.
759 	 */
760 #ifdef DEBUG
761 	debug_n_again = 0;
762 #endif
763 	do {
764 		busypg = NULL;
765 		KASSERT(mutex_owned(vp->v_interlock));
766 		if (check_dirty(fs, vp, startoffset, endoffset, blkeof,
767 				ap->a_flags, 0, &busypg) < 0) {
768 			mutex_exit(vp->v_interlock);
769 			/* XXX why? --ks */
770 			mutex_enter(vp->v_interlock);
771 			write_and_wait(fs, vp, busypg, seglocked, NULL);
772 			if (!seglocked) {
773 				mutex_exit(vp->v_interlock);
774 				lfs_release_finfo(fs);
775 				lfs_segunlock(fs);
776 				mutex_enter(vp->v_interlock);
777 			}
778 			sp->vp = NULL;
779 			goto get_seglock;
780 		}
781 
782 		busypg = NULL;
783 		KASSERT(!mutex_owned(&uvm_pageqlock));
784 		oreclaim = (ap->a_flags & PGO_RECLAIM);
785 		ap->a_flags &= ~PGO_RECLAIM;
786 		error = genfs_do_putpages(vp, startoffset, endoffset,
787 					   ap->a_flags, &busypg);
788 		ap->a_flags |= oreclaim;
789 
790 		if (error == EDEADLK || error == EAGAIN) {
791 			DLOG((DLOG_PAGE, "lfs_putpages: genfs_putpages returned"
792 			      " %d ino %d off %jx (seg %d)\n", error,
793 			      ip->i_number, (uintmax_t)lfs_sb_getoffset(fs),
794 			      lfs_dtosn(fs, lfs_sb_getoffset(fs))));
795 
796 			if (oreclaim) {
797 				mutex_enter(vp->v_interlock);
798 				write_and_wait(fs, vp, busypg, seglocked, "again");
799 				mutex_exit(vp->v_interlock);
800 			} else {
801 				if ((sp->seg_flags & SEGM_SINGLE) &&
802 				    lfs_sb_getcurseg(fs) != fs->lfs_startseg)
803 					donewriting = 1;
804 			}
805 		} else if (error) {
806 			DLOG((DLOG_PAGE, "lfs_putpages: genfs_putpages returned"
807 			      " %d ino %d off %jx (seg %d)\n", error,
808 			      (int)ip->i_number, (uintmax_t)lfs_sb_getoffset(fs),
809 			      lfs_dtosn(fs, lfs_sb_getoffset(fs))));
810 		}
811 		/* genfs_do_putpages loses the interlock */
812 #ifdef DEBUG
813 		++debug_n_again;
814 #endif
815 		if (oreclaim && error == EAGAIN) {
816 			DLOG((DLOG_PAGE, "vp %p ino %d vi_flags %x a_flags %x avoiding vclean panic\n",
817 			      vp, (int)ip->i_number, vp->v_iflag, ap->a_flags));
818 			mutex_enter(vp->v_interlock);
819 		}
820 		if (error == EDEADLK)
821 			mutex_enter(vp->v_interlock);
822 	} while (error == EDEADLK || (oreclaim && error == EAGAIN));
823 #ifdef DEBUG
824 	if (debug_n_again > TOOMANY)
825 		DLOG((DLOG_PAGE, "lfs_putpages: again: looping, n = %d\n", debug_n_again));
826 #endif
827 
828 	KASSERT(sp != NULL && sp->vp == vp);
829 	if (!seglocked && !donewriting) {
830 		sp->vp = NULL;
831 
832 		/* Write indirect blocks as well */
833 		lfs_gather(fs, fs->lfs_sp, vp, lfs_match_indir);
834 		lfs_gather(fs, fs->lfs_sp, vp, lfs_match_dindir);
835 		lfs_gather(fs, fs->lfs_sp, vp, lfs_match_tindir);
836 
837 		KASSERT(sp->vp == NULL);
838 		sp->vp = vp;
839 	}
840 
841 	/*
842 	 * Blocks are now gathered into a segment waiting to be written.
843 	 * All that's left to do is update metadata, and write them.
844 	 */
845 	lfs_updatemeta(sp);
846 	KASSERT(sp->vp == vp);
847 	sp->vp = NULL;
848 
849 	/*
850 	 * If we were called from lfs_writefile, we don't need to clean up
851 	 * the FIP or unlock the segment lock.	We're done.
852 	 */
853 	if (seglocked) {
854 		KASSERT(!mutex_owned(vp->v_interlock));
855 		return error;
856 	}
857 
858 	/* Clean up FIP and send it to disk. */
859 	lfs_release_finfo(fs);
860 	lfs_writeseg(fs, fs->lfs_sp);
861 
862 	/*
863 	 * Remove us from paging queue if we wrote all our pages.
864 	 */
865 	if (origendoffset == 0 || ap->a_flags & PGO_ALLPAGES) {
866 		mutex_enter(&lfs_lock);
867 		if (ip->i_flags & IN_PAGING) {
868 			ip->i_flags &= ~IN_PAGING;
869 			TAILQ_REMOVE(&fs->lfs_pchainhd, ip, i_lfs_pchain);
870 		}
871 		mutex_exit(&lfs_lock);
872 	}
873 
874 	/*
875 	 * XXX - with the malloc/copy writeseg, the pages are freed by now
876 	 * even if we don't wait (e.g. if we hold a nested lock).  This
877 	 * will not be true if we stop using malloc/copy.
878 	 */
879 	KASSERT(fs->lfs_sp->seg_flags & SEGM_PROT);
880 	lfs_segunlock(fs);
881 
882 	/*
883 	 * Wait for v_numoutput to drop to zero.  The seglock should
884 	 * take care of this, but there is a slight possibility that
885 	 * aiodoned might not have got around to our buffers yet.
886 	 */
887 	if (sync) {
888 		mutex_enter(vp->v_interlock);
889 		while (vp->v_numoutput > 0) {
890 			DLOG((DLOG_PAGE, "lfs_putpages: ino %d sleeping on"
891 			      " num %d\n", ip->i_number, vp->v_numoutput));
892 			cv_wait(&vp->v_cv, vp->v_interlock);
893 		}
894 		mutex_exit(vp->v_interlock);
895 	}
896 	KASSERT(!mutex_owned(vp->v_interlock));
897 	return error;
898 }
899 
900