xref: /netbsd-src/sys/nfs/nfs_bio.c (revision 7cc2f76925f078d01ddc9e640a98f4ccfc9f8c3b)
1 /*	$NetBSD: nfs_bio.c,v 1.57 2000/12/13 18:15:56 jdolecek Exp $	*/
2 
3 /*
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Rick Macklem at The University of Guelph.
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  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
39  */
40 
41 #include "opt_nfs.h"
42 #include "opt_ddb.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/proc.h>
49 #include <sys/buf.h>
50 #include <sys/vnode.h>
51 #include <sys/mount.h>
52 #include <sys/kernel.h>
53 #include <sys/namei.h>
54 #include <sys/dirent.h>
55 #include <sys/malloc.h>
56 
57 #include <uvm/uvm_extern.h>
58 #include <uvm/uvm.h>
59 
60 #include <nfs/rpcv2.h>
61 #include <nfs/nfsproto.h>
62 #include <nfs/nfs.h>
63 #include <nfs/nfsmount.h>
64 #include <nfs/nqnfs.h>
65 #include <nfs/nfsnode.h>
66 #include <nfs/nfs_var.h>
67 
68 extern int nfs_numasync;
69 extern struct nfsstats nfsstats;
70 
71 /*
72  * Vnode op for read using bio
73  * Any similarity to readip() is purely coincidental
74  */
75 int
76 nfs_bioread(vp, uio, ioflag, cred, cflag)
77 	struct vnode *vp;
78 	struct uio *uio;
79 	int ioflag, cflag;
80 	struct ucred *cred;
81 {
82 	struct nfsnode *np = VTONFS(vp);
83 	int biosize;
84 	struct buf *bp = NULL, *rabp;
85 	struct vattr vattr;
86 	struct proc *p;
87 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
88 	struct nfsdircache *ndp = NULL, *nndp = NULL;
89 	caddr_t baddr, ep, edp;
90 	int got_buf = 0, error = 0, n = 0, on = 0, en, enn;
91 	int enough = 0;
92 	struct dirent *dp, *pdp;
93 	off_t curoff = 0;
94 
95 #ifdef DIAGNOSTIC
96 	if (uio->uio_rw != UIO_READ)
97 		panic("nfs_read mode");
98 #endif
99 	if (uio->uio_resid == 0)
100 		return (0);
101 	if (vp->v_type != VDIR && uio->uio_offset < 0)
102 		return (EINVAL);
103 	p = uio->uio_procp;
104 #ifndef NFS_V2_ONLY
105 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
106 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
107 		(void)nfs_fsinfo(nmp, vp, cred, p);
108 #endif
109 	if (vp->v_type != VDIR &&
110 	    (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
111 		return (EFBIG);
112 	biosize = nmp->nm_rsize;
113 
114 	/*
115 	 * For nfs, cache consistency can only be maintained approximately.
116 	 * Although RFC1094 does not specify the criteria, the following is
117 	 * believed to be compatible with the reference port.
118 	 * For nqnfs, full cache consistency is maintained within the loop.
119 	 * For nfs:
120 	 * If the file's modify time on the server has changed since the
121 	 * last read rpc or you have written to the file,
122 	 * you may have lost data cache consistency with the
123 	 * server, so flush all of the file's data out of the cache.
124 	 * Then force a getattr rpc to ensure that you have up to date
125 	 * attributes.
126 	 * NB: This implies that cache data can be read when up to
127 	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
128 	 * attributes this could be forced by setting n_attrstamp to 0 before
129 	 * the VOP_GETATTR() call.
130 	 */
131 
132 	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
133 		if (np->n_flag & NMODIFIED) {
134 			if (vp->v_type != VREG) {
135 				if (vp->v_type != VDIR)
136 					panic("nfs: bioread, not dir");
137 				nfs_invaldircache(vp, 0);
138 				np->n_direofoffset = 0;
139 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
140 				if (error)
141 					return (error);
142 			}
143 			np->n_attrstamp = 0;
144 			error = VOP_GETATTR(vp, &vattr, cred, p);
145 			if (error)
146 				return (error);
147 			np->n_mtime = vattr.va_mtime.tv_sec;
148 		} else {
149 			error = VOP_GETATTR(vp, &vattr, cred, p);
150 			if (error)
151 				return (error);
152 			if (np->n_mtime != vattr.va_mtime.tv_sec) {
153 				if (vp->v_type == VDIR) {
154 					nfs_invaldircache(vp, 0);
155 					np->n_direofoffset = 0;
156 				}
157 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
158 				if (error)
159 					return (error);
160 				np->n_mtime = vattr.va_mtime.tv_sec;
161 			}
162 		}
163 	}
164 
165 	/*
166 	 * update the cached read creds for this node.
167 	 */
168 
169 	if (np->n_rcred) {
170 		crfree(np->n_rcred);
171 	}
172 	np->n_rcred = cred;
173 	crhold(cred);
174 
175 	do {
176 #ifndef NFS_V2_ONLY
177 	    /*
178 	     * Get a valid lease. If cached data is stale, flush it.
179 	     */
180 	    if (nmp->nm_flag & NFSMNT_NQNFS) {
181 		if (NQNFS_CKINVALID(vp, np, ND_READ)) {
182 		    do {
183 			error = nqnfs_getlease(vp, ND_READ, cred, p);
184 		    } while (error == NQNFS_EXPIRED);
185 		    if (error)
186 			return (error);
187 		    if (np->n_lrev != np->n_brev ||
188 			(np->n_flag & NQNFSNONCACHE) ||
189 			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
190 			if (vp->v_type == VDIR) {
191 				nfs_invaldircache(vp, 0);
192 				np->n_direofoffset = 0;
193 			}
194 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
195 			if (error)
196 			    return (error);
197 			np->n_brev = np->n_lrev;
198 		    }
199 		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
200 		    nfs_invaldircache(vp, 0);
201 		    error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
202 		    np->n_direofoffset = 0;
203 		    if (error)
204 			return (error);
205 		}
206 	    }
207 #endif
208 	    /*
209 	     * Don't cache symlinks.
210 	     */
211 	    if (np->n_flag & NQNFSNONCACHE
212 		|| ((vp->v_flag & VROOT) && vp->v_type == VLNK)) {
213 		switch (vp->v_type) {
214 		case VREG:
215 			return (nfs_readrpc(vp, uio));
216 		case VLNK:
217 			return (nfs_readlinkrpc(vp, uio, cred));
218 		case VDIR:
219 			break;
220 		default:
221 			printf(" NQNFSNONCACHE: type %x unexpected\n",
222 			    vp->v_type);
223 		};
224 	    }
225 	    baddr = (caddr_t)0;
226 	    switch (vp->v_type) {
227 	    case VREG:
228 		nfsstats.biocache_reads++;
229 
230 		error = 0;
231 		while (uio->uio_resid > 0) {
232 			void *win;
233 			vsize_t bytelen = min(np->n_size - uio->uio_offset,
234 					      uio->uio_resid);
235 
236 			if (bytelen == 0)
237 				break;
238 			win = ubc_alloc(&vp->v_uvm.u_obj, uio->uio_offset,
239 					&bytelen, UBC_READ);
240 			error = uiomove(win, bytelen, uio);
241 			ubc_release(win, 0);
242 			if (error) {
243 				break;
244 			}
245 		}
246 		n = 0;
247 		break;
248 
249 	    case VLNK:
250 		nfsstats.biocache_readlinks++;
251 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
252 		if (!bp)
253 			return (EINTR);
254 		if ((bp->b_flags & B_DONE) == 0) {
255 			bp->b_flags |= B_READ;
256 			error = nfs_doio(bp, p);
257 			if (error) {
258 				brelse(bp);
259 				return (error);
260 			}
261 		}
262 		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
263 		got_buf = 1;
264 		on = 0;
265 		break;
266 	    case VDIR:
267 diragain:
268 		nfsstats.biocache_readdirs++;
269 		ndp = nfs_searchdircache(vp, uio->uio_offset,
270 			(nmp->nm_flag & NFSMNT_XLATECOOKIE), 0);
271 		if (!ndp) {
272 			/*
273 			 * We've been handed a cookie that is not
274 			 * in the cache. If we're not translating
275 			 * 32 <-> 64, it may be a value that was
276 			 * flushed out of the cache because it grew
277 			 * too big. Let the server judge if it's
278 			 * valid or not. In the translation case,
279 			 * we have no way of validating this value,
280 			 * so punt.
281 			 */
282 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE)
283 				return (EINVAL);
284 			ndp = nfs_enterdircache(vp, uio->uio_offset,
285 				uio->uio_offset, 0, 0);
286 		}
287 
288 		if (uio->uio_offset != 0 &&
289 		    ndp->dc_cookie == np->n_direofoffset) {
290 			nfsstats.direofcache_hits++;
291 			return (0);
292 		}
293 
294 		bp = nfs_getcacheblk(vp, ndp->dc_blkno, NFS_DIRBLKSIZ, p);
295 		if (!bp)
296 		    return (EINTR);
297 		if ((bp->b_flags & B_DONE) == 0) {
298 		    bp->b_flags |= B_READ;
299 		    bp->b_dcookie = ndp->dc_blkcookie;
300 		    error = nfs_doio(bp, p);
301 		    if (error) {
302 			/*
303 			 * Yuck! The directory has been modified on the
304 			 * server. Punt and let the userland code
305 			 * deal with it.
306 			 */
307 			brelse(bp);
308 			if (error == NFSERR_BAD_COOKIE) {
309 			    nfs_invaldircache(vp, 0);
310 			    nfs_vinvalbuf(vp, 0, cred, p, 1);
311 			    error = EINVAL;
312 			}
313 			return (error);
314 		    }
315 		}
316 
317 		/*
318 		 * Just return if we hit EOF right away with this
319 		 * block. Always check here, because direofoffset
320 		 * may have been set by an nfsiod since the last
321 		 * check.
322 		 */
323 		if (np->n_direofoffset != 0 &&
324 			ndp->dc_blkcookie == np->n_direofoffset) {
325 			brelse(bp);
326 			return (0);
327 		}
328 
329 		/*
330 		 * Find the entry we were looking for in the block.
331 		 */
332 
333 		en = ndp->dc_entry;
334 
335 		pdp = dp = (struct dirent *)bp->b_data;
336 		edp = bp->b_data + bp->b_bcount;
337 		enn = 0;
338 		while (enn < en && (caddr_t)dp < edp) {
339 			pdp = dp;
340 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
341 			enn++;
342 		}
343 
344 		/*
345 		 * If the entry number was bigger than the number of
346 		 * entries in the block, or the cookie of the previous
347 		 * entry doesn't match, the directory cache is
348 		 * stale. Flush it and try again (i.e. go to
349 		 * the server).
350 		 */
351 		if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp ||
352 		    (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) {
353 #ifdef DEBUG
354 		    	printf("invalid cache: %p %p %p off %lx %lx\n",
355 				pdp, dp, edp,
356 				(unsigned long)uio->uio_offset,
357 				(unsigned long)NFS_GETCOOKIE(pdp));
358 #endif
359 			brelse(bp);
360 			nfs_invaldircache(vp, 0);
361 			nfs_vinvalbuf(vp, 0, cred, p, 0);
362 			goto diragain;
363 		}
364 
365 		on = (caddr_t)dp - bp->b_data;
366 
367 		/*
368 		 * Cache all entries that may be exported to the
369 		 * user, as they may be thrown back at us. The
370 		 * NFSBIO_CACHECOOKIES flag indicates that all
371 		 * entries are being 'exported', so cache them all.
372 		 */
373 
374 		if (en == 0 && pdp == dp) {
375 			dp = (struct dirent *)
376 			    ((caddr_t)dp + dp->d_reclen);
377 			enn++;
378 		}
379 
380 		if (uio->uio_resid < (bp->b_bcount - on)) {
381 			n = uio->uio_resid;
382 			enough = 1;
383 		} else
384 			n = bp->b_bcount - on;
385 
386 		ep = bp->b_data + on + n;
387 
388 		/*
389 		 * Find last complete entry to copy, caching entries
390 		 * (if requested) as we go.
391 		 */
392 
393 		while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) {
394 			if (cflag & NFSBIO_CACHECOOKIES) {
395 				nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp),
396 				    ndp->dc_blkcookie, enn, bp->b_lblkno);
397 				if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
398 					NFS_STASHCOOKIE32(pdp,
399 					    nndp->dc_cookie32);
400 				}
401 			}
402 			pdp = dp;
403 			dp = (struct dirent *)((caddr_t)dp + dp->d_reclen);
404 			enn++;
405 		}
406 
407 		/*
408 		 * If the last requested entry was not the last in the
409 		 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ),
410 		 * cache the cookie of the last requested one, and
411 		 * set of the offset to it.
412 		 */
413 
414 		if ((on + n) < bp->b_bcount) {
415 			curoff = NFS_GETCOOKIE(pdp);
416 			nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie,
417 			    enn, bp->b_lblkno);
418 			if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
419 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
420 				curoff = nndp->dc_cookie32;
421 			}
422 		} else
423 			curoff = bp->b_dcookie;
424 
425 		/*
426 		 * Always cache the entry for the next block,
427 		 * so that readaheads can use it.
428 		 */
429 		nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0);
430 		if (nmp->nm_flag & NFSMNT_XLATECOOKIE) {
431 			if (curoff == bp->b_dcookie) {
432 				NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32);
433 				curoff = nndp->dc_cookie32;
434 			}
435 		}
436 
437 		n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on);
438 
439 		/*
440 		 * If not eof and read aheads are enabled, start one.
441 		 * (You need the current block first, so that you have the
442 		 *  directory offset cookie of the next block.)
443 		 */
444 		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
445 		    np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) {
446 			rabp = nfs_getcacheblk(vp, nndp->dc_blkno,
447 						NFS_DIRBLKSIZ, p);
448 			if (rabp) {
449 			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
450 				rabp->b_dcookie = nndp->dc_cookie;
451 				rabp->b_flags |= (B_READ | B_ASYNC);
452 				if (nfs_asyncio(rabp)) {
453 				    rabp->b_flags |= B_INVAL;
454 				    brelse(rabp);
455 				}
456 			    } else
457 				brelse(rabp);
458 			}
459 		}
460 		got_buf = 1;
461 		break;
462 	    default:
463 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
464 		break;
465 	    }
466 
467 	    if (n > 0) {
468 		if (!baddr)
469 			baddr = bp->b_data;
470 		error = uiomove(baddr + on, (int)n, uio);
471 	    }
472 	    switch (vp->v_type) {
473 	    case VREG:
474 		break;
475 	    case VLNK:
476 		n = 0;
477 		break;
478 	    case VDIR:
479 		if (np->n_flag & NQNFSNONCACHE)
480 			bp->b_flags |= B_INVAL;
481 		uio->uio_offset = curoff;
482 		if (enough)
483 			n = 0;
484 		break;
485 	    default:
486 		printf(" nfsbioread: type %x unexpected\n",vp->v_type);
487 	    }
488 	    if (got_buf)
489 		brelse(bp);
490 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
491 	return (error);
492 }
493 
494 /*
495  * Vnode op for write using bio
496  */
497 int
498 nfs_write(v)
499 	void *v;
500 {
501 	struct vop_write_args /* {
502 		struct vnode *a_vp;
503 		struct uio *a_uio;
504 		int  a_ioflag;
505 		struct ucred *a_cred;
506 	} */ *ap = v;
507 	struct uio *uio = ap->a_uio;
508 	struct proc *p = uio->uio_procp;
509 	struct vnode *vp = ap->a_vp;
510 	struct nfsnode *np = VTONFS(vp);
511 	struct ucred *cred = ap->a_cred;
512 	int ioflag = ap->a_ioflag;
513 	struct vattr vattr;
514 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
515 	int error = 0, iomode, must_commit;
516 	int rv;
517 
518 #ifdef DIAGNOSTIC
519 	if (uio->uio_rw != UIO_WRITE)
520 		panic("nfs_write mode");
521 	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
522 		panic("nfs_write proc");
523 #endif
524 	if (vp->v_type != VREG)
525 		return (EIO);
526 	if (np->n_flag & NWRITEERR) {
527 		np->n_flag &= ~NWRITEERR;
528 		return (np->n_error);
529 	}
530 #ifndef NFS_V2_ONLY
531 	if ((nmp->nm_flag & NFSMNT_NFSV3) &&
532 	    !(nmp->nm_iflag & NFSMNT_GOTFSINFO))
533 		(void)nfs_fsinfo(nmp, vp, cred, p);
534 #endif
535 	if (ioflag & (IO_APPEND | IO_SYNC)) {
536 		if (np->n_flag & NMODIFIED) {
537 			np->n_attrstamp = 0;
538 			error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
539 			if (error)
540 				return (error);
541 		}
542 		if (ioflag & IO_APPEND) {
543 			np->n_attrstamp = 0;
544 			error = VOP_GETATTR(vp, &vattr, cred, p);
545 			if (error)
546 				return (error);
547 			uio->uio_offset = np->n_size;
548 		}
549 	}
550 	if (uio->uio_offset < 0)
551 		return (EINVAL);
552 	if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
553 		return (EFBIG);
554 	if (uio->uio_resid == 0)
555 		return (0);
556 	/*
557 	 * Maybe this should be above the vnode op call, but so long as
558 	 * file servers have no limits, i don't think it matters
559 	 */
560 	if (p && uio->uio_offset + uio->uio_resid >
561 	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
562 		psignal(p, SIGXFSZ);
563 		return (EFBIG);
564 	}
565 
566 	/*
567 	 * update the cached write creds for this node.
568 	 */
569 
570 	if (np->n_wcred) {
571 		crfree(np->n_wcred);
572 	}
573 	np->n_wcred = cred;
574 	crhold(cred);
575 
576 	if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
577 		iomode = NFSV3WRITE_FILESYNC;
578 		error = nfs_writerpc(vp, uio, &iomode, &must_commit);
579 		if (must_commit)
580 			nfs_clearcommit(vp->v_mount);
581 		return (error);
582 	}
583 
584 	do {
585 		void *win;
586 		voff_t oldoff = uio->uio_offset;
587 		vsize_t bytelen = uio->uio_resid;
588 
589 #ifndef NFS_V2_ONLY
590 		/*
591 		 * Check for a valid write lease.
592 		 */
593 		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
594 		    NQNFS_CKINVALID(vp, np, ND_WRITE)) {
595 			do {
596 				error = nqnfs_getlease(vp, ND_WRITE, cred, p);
597 			} while (error == NQNFS_EXPIRED);
598 			if (error)
599 				return (error);
600 			if (np->n_lrev != np->n_brev ||
601 			    (np->n_flag & NQNFSNONCACHE)) {
602 				error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
603 				if (error)
604 					return (error);
605 				np->n_brev = np->n_lrev;
606 			}
607 		}
608 #endif
609 		nfsstats.biocache_writes++;
610 
611 		np->n_flag |= NMODIFIED;
612 		if (np->n_size < uio->uio_offset + bytelen) {
613 			np->n_size = uio->uio_offset + bytelen;
614 			uvm_vnp_setsize(vp, np->n_size);
615 		}
616 		win = ubc_alloc(&vp->v_uvm.u_obj, uio->uio_offset, &bytelen,
617 				UBC_WRITE);
618 		error = uiomove(win, bytelen, uio);
619 		if (error) {
620 			memset((void *)trunc_page((vaddr_t)win), 0,
621 			       round_page((vaddr_t)win + bytelen) -
622 			       trunc_page((vaddr_t)win));
623 		}
624 		ubc_release(win, 0);
625 		rv = 1;
626 		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
627 			simple_lock(&vp->v_uvm.u_obj.vmobjlock);
628 			rv = vp->v_uvm.u_obj.pgops->pgo_flush(
629 			    &vp->v_uvm.u_obj,
630 			    oldoff & ~(nmp->nm_wsize - 1),
631 			    uio->uio_offset & ~(nmp->nm_wsize - 1),
632 			    PGO_CLEANIT|PGO_SYNCIO);
633 			simple_unlock(&vp->v_uvm.u_obj.vmobjlock);
634 		} else if ((oldoff & ~(nmp->nm_wsize - 1)) !=
635 		    (uio->uio_offset & ~(nmp->nm_wsize - 1))) {
636 			simple_lock(&vp->v_uvm.u_obj.vmobjlock);
637 			rv = vp->v_uvm.u_obj.pgops->pgo_flush(
638 			    &vp->v_uvm.u_obj,
639 			    oldoff & ~(nmp->nm_wsize - 1),
640 			    uio->uio_offset & ~(nmp->nm_wsize - 1),
641 			    PGO_CLEANIT|PGO_WEAK);
642 			simple_unlock(&vp->v_uvm.u_obj.vmobjlock);
643 		}
644 		if (!rv) {
645 			error = EIO;
646 			break;
647 		}
648 	} while (uio->uio_resid > 0);
649 	return error;
650 }
651 
652 /*
653  * Get an nfs cache block.
654  * Allocate a new one if the block isn't currently in the cache
655  * and return the block marked busy. If the calling process is
656  * interrupted by a signal for an interruptible mount point, return
657  * NULL.
658  */
659 struct buf *
660 nfs_getcacheblk(vp, bn, size, p)
661 	struct vnode *vp;
662 	daddr_t bn;
663 	int size;
664 	struct proc *p;
665 {
666 	struct buf *bp;
667 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
668 
669 	if (nmp->nm_flag & NFSMNT_INT) {
670 		bp = getblk(vp, bn, size, PCATCH, 0);
671 		while (bp == NULL) {
672 			if (nfs_sigintr(nmp, NULL, p))
673 				return (NULL);
674 			bp = getblk(vp, bn, size, 0, 2 * hz);
675 		}
676 	} else
677 		bp = getblk(vp, bn, size, 0, 0);
678 	return (bp);
679 }
680 
681 /*
682  * Flush and invalidate all dirty buffers. If another process is already
683  * doing the flush, just wait for completion.
684  */
685 int
686 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
687 	struct vnode *vp;
688 	int flags;
689 	struct ucred *cred;
690 	struct proc *p;
691 	int intrflg;
692 {
693 	struct nfsnode *np = VTONFS(vp);
694 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
695 	int error = 0, slpflag, slptimeo;
696 
697 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
698 		intrflg = 0;
699 	if (intrflg) {
700 		slpflag = PCATCH;
701 		slptimeo = 2 * hz;
702 	} else {
703 		slpflag = 0;
704 		slptimeo = 0;
705 	}
706 	/*
707 	 * First wait for any other process doing a flush to complete.
708 	 */
709 	while (np->n_flag & NFLUSHINPROG) {
710 		np->n_flag |= NFLUSHWANT;
711 		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
712 			slptimeo);
713 		if (error && intrflg && nfs_sigintr(nmp, NULL, p))
714 			return (EINTR);
715 	}
716 
717 	/*
718 	 * Now, flush as required.
719 	 */
720 	np->n_flag |= NFLUSHINPROG;
721 	error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
722 	while (error) {
723 		if (intrflg && nfs_sigintr(nmp, NULL, p)) {
724 			np->n_flag &= ~NFLUSHINPROG;
725 			if (np->n_flag & NFLUSHWANT) {
726 				np->n_flag &= ~NFLUSHWANT;
727 				wakeup((caddr_t)&np->n_flag);
728 			}
729 			return (EINTR);
730 		}
731 		error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
732 	}
733 	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
734 	if (np->n_flag & NFLUSHWANT) {
735 		np->n_flag &= ~NFLUSHWANT;
736 		wakeup((caddr_t)&np->n_flag);
737 	}
738 	return (0);
739 }
740 
741 /*
742  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
743  * This is mainly to avoid queueing async I/O requests when the nfsiods
744  * are all hung on a dead server.
745  */
746 int
747 nfs_asyncio(bp)
748 	struct buf *bp;
749 {
750 	int i;
751 	struct nfsmount *nmp;
752 	int gotiod, slpflag = 0, slptimeo = 0, error;
753 
754 	if (nfs_numasync == 0)
755 		return (EIO);
756 
757 
758 	nmp = VFSTONFS(bp->b_vp->v_mount);
759 again:
760 	if (nmp->nm_flag & NFSMNT_INT)
761 		slpflag = PCATCH;
762 	gotiod = FALSE;
763 
764 	/*
765 	 * Find a free iod to process this request.
766 	 */
767 
768 	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
769 		if (nfs_iodwant[i]) {
770 			/*
771 			 * Found one, so wake it up and tell it which
772 			 * mount to process.
773 			 */
774 			nfs_iodwant[i] = NULL;
775 			nfs_iodmount[i] = nmp;
776 			nmp->nm_bufqiods++;
777 			wakeup((caddr_t)&nfs_iodwant[i]);
778 			gotiod = TRUE;
779 			break;
780 		}
781 	/*
782 	 * If none are free, we may already have an iod working on this mount
783 	 * point.  If so, it will process our request.
784 	 */
785 	if (!gotiod && nmp->nm_bufqiods > 0)
786 		gotiod = TRUE;
787 
788 	/*
789 	 * If we have an iod which can process the request, then queue
790 	 * the buffer.
791 	 */
792 	if (gotiod) {
793 		/*
794 		 * Ensure that the queue never grows too large.
795 		 */
796 		while (nmp->nm_bufqlen >= 2*nfs_numasync) {
797 			nmp->nm_bufqwant = TRUE;
798 			error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
799 				"nfsaio", slptimeo);
800 			if (error) {
801 				if (nfs_sigintr(nmp, NULL, bp->b_proc))
802 					return (EINTR);
803 				if (slpflag == PCATCH) {
804 					slpflag = 0;
805 					slptimeo = 2 * hz;
806 				}
807 			}
808 			/*
809 			 * We might have lost our iod while sleeping,
810 			 * so check and loop if nescessary.
811 			 */
812 			if (nmp->nm_bufqiods == 0)
813 				goto again;
814 		}
815 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
816 		nmp->nm_bufqlen++;
817 		return (0);
818 	    }
819 
820 	/*
821 	 * All the iods are busy on other mounts, so return EIO to
822 	 * force the caller to process the i/o synchronously.
823 	 */
824 	return (EIO);
825 }
826 
827 /*
828  * Do an I/O operation to/from a cache block. This may be called
829  * synchronously or from an nfsiod.
830  */
831 int
832 nfs_doio(bp, p)
833 	struct buf *bp;
834 	struct proc *p;
835 {
836 	struct uio *uiop;
837 	struct vnode *vp;
838 	struct nfsnode *np;
839 	struct nfsmount *nmp;
840 	int error = 0, diff, len, iomode, must_commit = 0;
841 	struct uio uio;
842 	struct iovec io;
843 
844 	vp = bp->b_vp;
845 	np = VTONFS(vp);
846 	nmp = VFSTONFS(vp->v_mount);
847 	uiop = &uio;
848 	uiop->uio_iov = &io;
849 	uiop->uio_iovcnt = 1;
850 	uiop->uio_segflg = UIO_SYSSPACE;
851 	uiop->uio_procp = p;
852 
853 	/*
854 	 * Historically, paging was done with physio, but no more...
855 	 */
856 	if (bp->b_flags & B_PHYS) {
857 	    /*
858 	     * ...though reading /dev/drum still gets us here.
859 	     */
860 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
861 	    /* mapping was done by vmapbuf() */
862 	    io.iov_base = bp->b_data;
863 	    uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
864 	    if (bp->b_flags & B_READ) {
865 		uiop->uio_rw = UIO_READ;
866 		nfsstats.read_physios++;
867 		error = nfs_readrpc(vp, uiop);
868 	    } else {
869 		iomode = NFSV3WRITE_DATASYNC;
870 		uiop->uio_rw = UIO_WRITE;
871 		nfsstats.write_physios++;
872 		error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
873 	    }
874 	    if (error) {
875 		bp->b_flags |= B_ERROR;
876 		bp->b_error = error;
877 	    }
878 	} else if (bp->b_flags & B_READ) {
879 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
880 	    io.iov_base = bp->b_data;
881 	    uiop->uio_rw = UIO_READ;
882 	    switch (vp->v_type) {
883 	    case VREG:
884 		uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT;
885 		nfsstats.read_bios++;
886 		error = nfs_readrpc(vp, uiop);
887 		if (!error && uiop->uio_resid) {
888 
889 			/*
890 			 * If len > 0, there is a hole in the file and
891 			 * no writes after the hole have been pushed to
892 			 * the server yet.
893 			 * Just zero fill the rest of the valid area.
894 			 */
895 
896 			diff = bp->b_bcount - uiop->uio_resid;
897 			len = np->n_size - ((((off_t)bp->b_blkno) << DEV_BSHIFT)
898 				+ diff);
899 			if (len > 0) {
900 				len = min(len, uiop->uio_resid);
901 				memset((char *)bp->b_data + diff, 0, len);
902 			}
903 		}
904 		if (p && (vp->v_flag & VTEXT) &&
905 			(((nmp->nm_flag & NFSMNT_NQNFS) &&
906 			  NQNFS_CKINVALID(vp, np, ND_READ) &&
907 			  np->n_lrev != np->n_brev) ||
908 			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
909 			  np->n_mtime != np->n_vattr->va_mtime.tv_sec))) {
910 			uprintf("Process killed due to "
911 				"text file modification\n");
912 			psignal(p, SIGKILL);
913 			p->p_holdcnt++;
914 		}
915 		break;
916 	    case VLNK:
917 		uiop->uio_offset = (off_t)0;
918 		nfsstats.readlink_bios++;
919 		error = nfs_readlinkrpc(vp, uiop, curproc->p_ucred);
920 		break;
921 	    case VDIR:
922 		nfsstats.readdir_bios++;
923 		uiop->uio_offset = bp->b_dcookie;
924 		if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
925 			error = nfs_readdirplusrpc(vp, uiop, curproc->p_ucred);
926 			if (error == NFSERR_NOTSUPP)
927 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
928 		}
929 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
930 			error = nfs_readdirrpc(vp, uiop, curproc->p_ucred);
931 		if (!error) {
932 			bp->b_dcookie = uiop->uio_offset;
933 		}
934 		break;
935 	    default:
936 		printf("nfs_doio:  type %x unexpected\n",vp->v_type);
937 		break;
938 	    }
939 	    if (error) {
940 		bp->b_flags |= B_ERROR;
941 		bp->b_error = error;
942 	    }
943 	} else {
944 	    /*
945 	     * If B_NEEDCOMMIT is set, a commit rpc may do the trick. If not
946 	     * an actual write will have to be scheduled.
947 	     */
948 
949 	    io.iov_base = bp->b_data;
950 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
951 	    uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT);
952 	    uiop->uio_rw = UIO_WRITE;
953 	    nfsstats.write_bios++;
954 	    iomode = NFSV3WRITE_UNSTABLE;
955 	    error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
956 	}
957 	bp->b_resid = uiop->uio_resid;
958 	if (must_commit)
959 		nfs_clearcommit(vp->v_mount);
960 	biodone(bp);
961 	return (error);
962 }
963 
964 /*
965  * Vnode op for VM getpages.
966  */
967 int
968 nfs_getpages(v)
969 	void *v;
970 {
971 	struct vop_getpages_args /* {
972 		struct vnode *a_vp;
973 		voff_t a_offset;
974 		vm_page_t *a_m;
975 		int *a_count;
976 		int a_centeridx;
977 		vm_prot_t a_access_type;
978 		int a_advice;
979 		int a_flags;
980 	} */ *ap = v;
981 
982 	off_t eof, offset, origoffset, startoffset, endoffset;
983 	int s, i, error, npages, orignpages, npgs, ridx, pidx, pcount;
984 	vaddr_t kva;
985 	struct buf *bp, *mbp;
986 	struct vnode *vp = ap->a_vp;
987 	struct nfsnode *np = VTONFS(vp);
988 	struct uvm_object *uobj = &vp->v_uvm.u_obj;
989 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
990 	size_t bytes, iobytes, tailbytes, totalbytes, skipbytes;
991 	int flags = ap->a_flags;
992 	int bsize;
993 	struct vm_page *pgs[16];			/* XXXUBC 16 */
994 	boolean_t v3 = NFS_ISV3(vp);
995 	boolean_t async = (flags & PGO_SYNCIO) == 0;
996 	boolean_t write = (ap->a_access_type & VM_PROT_WRITE) != 0;
997 
998 	UVMHIST_FUNC("nfs_getpages"); UVMHIST_CALLED(ubchist);
999 	UVMHIST_LOG(ubchist, "vp %p off 0x%x count %d", vp, (int)ap->a_offset,
1000 		    *ap->a_count,0);
1001 
1002 #ifdef DIAGNOSTIC
1003 	if (ap->a_centeridx < 0 || ap->a_centeridx >= *ap->a_count) {
1004 		panic("nfs_getpages: centeridx %d out of range",
1005 		      ap->a_centeridx);
1006 	}
1007 #endif
1008 
1009 	error = 0;
1010 	origoffset = ap->a_offset;
1011 	eof = vp->v_uvm.u_size;
1012 	if (origoffset >= eof) {
1013 		if ((flags & PGO_LOCKED) == 0) {
1014 			simple_unlock(&uobj->vmobjlock);
1015 		}
1016 		UVMHIST_LOG(ubchist, "off 0x%x past EOF 0x%x",
1017 			    (int)origoffset, (int)eof,0,0);
1018 		return EINVAL;
1019 	}
1020 
1021 	if (flags & PGO_LOCKED) {
1022 		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m,
1023 			      UFP_NOWAIT|UFP_NOALLOC);
1024 		return 0;
1025 	}
1026 
1027 	/* vnode is VOP_LOCKed, uobj is locked */
1028 
1029 	bsize = nmp->nm_rsize;
1030 	orignpages = min(*ap->a_count,
1031 			 round_page(eof - origoffset) >> PAGE_SHIFT);
1032 	npages = orignpages;
1033 	startoffset = origoffset & ~(bsize - 1);
1034 	endoffset = round_page((origoffset + (npages << PAGE_SHIFT)
1035 				+ bsize - 1) & ~(bsize - 1));
1036 	endoffset = min(endoffset, round_page(eof));
1037 	ridx = (origoffset - startoffset) >> PAGE_SHIFT;
1038 
1039 	if (!async && !write) {
1040 		int rapages = max(PAGE_SIZE, nmp->nm_rsize) >> PAGE_SHIFT;
1041 
1042 		(void) VOP_GETPAGES(vp, endoffset, NULL, &rapages, 0,
1043 				    VM_PROT_READ, 0, 0);
1044 		simple_lock(&uobj->vmobjlock);
1045 	}
1046 
1047 	UVMHIST_LOG(ubchist, "npages %d offset 0x%x", npages,
1048 		    (int)origoffset, 0,0);
1049 	memset(pgs, 0, sizeof(pgs));
1050 	uvn_findpages(uobj, origoffset, &npages, &pgs[ridx], UFP_ALL);
1051 
1052 	if (flags & PGO_OVERWRITE) {
1053 		UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0);
1054 
1055 		/* XXXUBC for now, zero the page if we allocated it */
1056 		for (i = 0; i < npages; i++) {
1057 			struct vm_page *pg = pgs[ridx + i];
1058 
1059 			if (pg->flags & PG_FAKE) {
1060 				uvm_pagezero(pg);
1061 				pg->flags &= ~(PG_FAKE);
1062 			}
1063 		}
1064 		goto out;
1065 	}
1066 
1067 	/*
1068 	 * if the pages are already resident, just return them.
1069 	 */
1070 
1071 	for (i = 0; i < npages; i++) {
1072 		struct vm_page *pg = pgs[ridx + i];
1073 
1074 		if ((pg->flags & PG_FAKE) != 0 ||
1075 		    ((ap->a_access_type & VM_PROT_WRITE) &&
1076 		      (pg->flags & PG_RDONLY))) {
1077 			break;
1078 		}
1079 	}
1080 	if (i == npages) {
1081 		UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0);
1082 		goto out;
1083 	}
1084 
1085 	/*
1086 	 * the page wasn't resident and we're not overwriting,
1087 	 * so we're going to have to do some i/o.
1088 	 * find any additional pages needed to cover the expanded range.
1089 	 */
1090 
1091 	if (startoffset != origoffset ||
1092 	    startoffset + (npages << PAGE_SHIFT) != endoffset) {
1093 		UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x",
1094 			    (int)startoffset, (int)endoffset, 0,0);
1095 		npages = (endoffset - startoffset) >> PAGE_SHIFT;
1096 		KASSERT(npages != 0);
1097 		npgs = npages;
1098 		uvn_findpages(uobj, startoffset, &npgs, pgs, UFP_ALL);
1099 	}
1100 	simple_unlock(&uobj->vmobjlock);
1101 
1102 	/*
1103 	 * update the cached read creds for this node.
1104 	 */
1105 
1106 	if (np->n_rcred) {
1107 		crfree(np->n_rcred);
1108 	}
1109 	np->n_rcred = curproc->p_ucred;
1110 	crhold(np->n_rcred);
1111 
1112 	/*
1113 	 * read the desired page(s).
1114 	 */
1115 
1116 	totalbytes = npages << PAGE_SHIFT;
1117 	bytes = min(totalbytes, vp->v_uvm.u_size - startoffset);
1118 	tailbytes = totalbytes - bytes;
1119 	skipbytes = 0;
1120 
1121 	kva = uvm_pagermapin(pgs, npages, UVMPAGER_MAPIN_WAITOK |
1122 			     UVMPAGER_MAPIN_READ);
1123 
1124 	s = splbio();
1125 	mbp = pool_get(&bufpool, PR_WAITOK);
1126 	splx(s);
1127 	mbp->b_bufsize = totalbytes;
1128 	mbp->b_data = (void *)kva;
1129 	mbp->b_resid = mbp->b_bcount = bytes;
1130 	mbp->b_flags = B_BUSY|B_READ| (async ? B_CALL|B_ASYNC : 0);
1131 	mbp->b_iodone = uvm_aio_biodone;
1132 	mbp->b_vp = vp;
1133 	LIST_INIT(&mbp->b_dep);
1134 
1135 	/*
1136 	 * if EOF is in the middle of the last page, zero the part past EOF.
1137 	 */
1138 
1139 	if (tailbytes > 0 && (pgs[bytes >> PAGE_SHIFT]->flags & PG_FAKE)) {
1140 		memset((char *)kva + bytes, 0, tailbytes);
1141 	}
1142 
1143 	/*
1144 	 * now loop over the pages, reading as needed.
1145 	 */
1146 
1147 	bp = NULL;
1148 	for (offset = startoffset;
1149 	     bytes > 0;
1150 	     offset += iobytes, bytes -= iobytes) {
1151 
1152 		/*
1153 		 * skip pages which don't need to be read.
1154 		 */
1155 
1156 		pidx = (offset - startoffset) >> PAGE_SHIFT;
1157 		UVMHIST_LOG(ubchist, "pidx %d offset 0x%x startoffset 0x%x",
1158 			    pidx, (int)offset, (int)startoffset,0);
1159 		while ((pgs[pidx]->flags & PG_FAKE) == 0) {
1160 			size_t b;
1161 
1162 #ifdef DEBUG
1163 			if (offset & (PAGE_SIZE - 1)) {
1164 				panic("nfs_getpages: skipping from middle "
1165 				      "of page");
1166 			}
1167 #endif
1168 
1169 			b = min(PAGE_SIZE, bytes);
1170 			offset += b;
1171 			bytes -= b;
1172 			skipbytes += b;
1173 			pidx++;
1174 			UVMHIST_LOG(ubchist, "skipping, new offset 0x%x",
1175 				    (int)offset, 0,0,0);
1176 			if (bytes == 0) {
1177 				goto loopdone;
1178 			}
1179 		}
1180 
1181 		/*
1182 		 * see how many pages can be read with this i/o.
1183 		 * reduce the i/o size if necessary.
1184 		 */
1185 
1186 		iobytes = bytes;
1187 		if (offset + iobytes > round_page(offset)) {
1188 			pcount = 1;
1189 			while (pidx + pcount < npages &&
1190 			       pgs[pidx + pcount]->flags & PG_FAKE) {
1191 				pcount++;
1192 			}
1193 			iobytes = min(iobytes, (pcount << PAGE_SHIFT) -
1194 				      (offset - trunc_page(offset)));
1195 		}
1196 		iobytes = min(iobytes, nmp->nm_rsize);
1197 
1198 		/*
1199 		 * allocate a sub-buf for this piece of the i/o
1200 		 * (or just use mbp if there's only 1 piece),
1201 		 * and start it going.
1202 		 */
1203 
1204 		if (offset == startoffset && iobytes == bytes) {
1205 			bp = mbp;
1206 		} else {
1207 			s = splbio();
1208 			bp = pool_get(&bufpool, PR_WAITOK);
1209 			splx(s);
1210 			bp->b_data = (char *)kva + offset - startoffset;
1211 			bp->b_resid = bp->b_bcount = iobytes;
1212 			bp->b_flags = B_BUSY|B_READ|B_CALL|B_ASYNC;
1213 			bp->b_iodone = uvm_aio_biodone1;
1214 			bp->b_vp = vp;
1215 			LIST_INIT(&bp->b_dep);
1216 		}
1217 		bp->b_private = mbp;
1218 		bp->b_lblkno = bp->b_blkno = offset >> DEV_BSHIFT;
1219 
1220 		UVMHIST_LOG(ubchist, "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
1221 			    bp, offset, iobytes, bp->b_blkno);
1222 
1223 		VOP_STRATEGY(bp);
1224 	}
1225 
1226 loopdone:
1227 	if (skipbytes) {
1228 		s = splbio();
1229 		mbp->b_resid -= skipbytes;
1230 		if (mbp->b_resid == 0) {
1231 			biodone(mbp);
1232 		}
1233 		splx(s);
1234 	}
1235 	if (async) {
1236 		UVMHIST_LOG(ubchist, "returning PEND",0,0,0,0);
1237 		return EINPROGRESS;
1238 	}
1239 	if (bp != NULL) {
1240 		error = biowait(mbp);
1241 	}
1242 	s = splbio();
1243 	pool_put(&bufpool, mbp);
1244 	splx(s);
1245 	uvm_pagermapout(kva, npages);
1246 
1247 	if (write && v3) {
1248 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1249 		nfs_del_committed_range(vp, origoffset, npages);
1250 		nfs_del_tobecommitted_range(vp, origoffset, npages);
1251 		for (i = 0; i < npages; i++) {
1252 			if (pgs[i] == NULL) {
1253 				continue;
1254 			}
1255 			pgs[i]->flags &= ~(PG_NEEDCOMMIT|PG_RDONLY);
1256 		}
1257 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1258 	}
1259 
1260 	simple_lock(&uobj->vmobjlock);
1261 
1262 out:
1263 	uvm_lock_pageq();
1264 	if (error) {
1265 		for (i = 0; i < npages; i++) {
1266 			if (pgs[i] == NULL) {
1267 				continue;
1268 			}
1269 			UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
1270 				    pgs[i], pgs[i]->flags, 0,0);
1271 			if ((pgs[i]->flags & PG_FAKE) == 0) {
1272 				continue;
1273 			}
1274 			if (pgs[i]->flags & PG_WANTED) {
1275 				wakeup(pgs[i]);
1276 			}
1277 			uvm_pagefree(pgs[i]);
1278 		}
1279 		goto done;
1280 	}
1281 
1282 	UVMHIST_LOG(ubchist, "ridx %d count %d", ridx, npages, 0,0);
1283 	for (i = 0; i < npages; i++) {
1284 		if (pgs[i] == NULL) {
1285 			continue;
1286 		}
1287 		UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
1288 			    pgs[i], pgs[i]->flags, 0,0);
1289 		if (pgs[i]->flags & PG_FAKE) {
1290 			UVMHIST_LOG(ubchist, "unfaking pg %p offset 0x%x",
1291 				    pgs[i], (int)pgs[i]->offset,0,0);
1292 			pgs[i]->flags &= ~(PG_FAKE);
1293 			pmap_clear_modify(pgs[i]);
1294 			pmap_clear_reference(pgs[i]);
1295 		}
1296 		if (i < ridx || i >= ridx + orignpages || async) {
1297 			UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x",
1298 				    pgs[i], (int)pgs[i]->offset,0,0);
1299 			KASSERT((pgs[i]->flags & PG_RELEASED) == 0);
1300 			if (pgs[i]->flags & PG_WANTED) {
1301 				wakeup(pgs[i]);
1302 			}
1303 			if (pgs[i]->wire_count == 0) {
1304 				uvm_pageactivate(pgs[i]);
1305 			}
1306 			pgs[i]->flags &= ~(PG_WANTED|PG_BUSY);
1307 			UVM_PAGE_OWN(pgs[i], NULL);
1308 		}
1309 	}
1310 
1311 done:
1312 	uvm_unlock_pageq();
1313 	simple_unlock(&uobj->vmobjlock);
1314 	if (ap->a_m != NULL) {
1315 		memcpy(ap->a_m, &pgs[ridx],
1316 		       *ap->a_count * sizeof(struct vm_page *));
1317 	}
1318 
1319 	UVMHIST_LOG(ubchist, "done -> %d", error, 0,0,0);
1320 	return error;
1321 }
1322 
1323 /*
1324  * Vnode op for VM putpages.
1325  */
1326 int
1327 nfs_putpages(v)
1328 	void *v;
1329 {
1330 	struct vop_putpages_args /* {
1331 		struct vnode *a_vp;
1332 		struct vm_page **a_m;
1333 		int a_count;
1334 		int a_flags;
1335 		int *a_rtvals;
1336 	} */ *ap = v;
1337 
1338 	struct vnode *vp = ap->a_vp;
1339 	struct nfsnode *np = VTONFS(vp);
1340 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1341 	struct buf *bp, *mbp;
1342 	struct vm_page **pgs = ap->a_m;
1343 	int flags = ap->a_flags;
1344 	int npages = ap->a_count;
1345 	int s, error = 0, i;
1346 	size_t bytes, iobytes, skipbytes;
1347 	vaddr_t kva;
1348 	off_t offset, origoffset, commitoff;
1349 	uint32_t commitbytes;
1350 	boolean_t v3 = NFS_ISV3(vp);
1351 	boolean_t async = (flags & PGO_SYNCIO) == 0;
1352 	boolean_t weak = (flags & PGO_WEAK) && v3;
1353 	UVMHIST_FUNC("nfs_putpages"); UVMHIST_CALLED(ubchist);
1354 
1355 	UVMHIST_LOG(ubchist, "vp %p pgp %p count %d",
1356 		    vp, ap->a_m, ap->a_count,0);
1357 
1358 	simple_unlock(&vp->v_uvm.u_obj.vmobjlock);
1359 
1360 	origoffset = pgs[0]->offset;
1361 	bytes = min(ap->a_count << PAGE_SHIFT, vp->v_uvm.u_size - origoffset);
1362 	skipbytes = 0;
1363 
1364 	/*
1365 	 * if the range has been committed already, mark the pages thus.
1366 	 * if the range just needs to be committed, we're done
1367 	 * if it's a weak putpage, otherwise commit the range.
1368 	 */
1369 
1370 	if (v3) {
1371 		lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1372 		if (nfs_in_committed_range(vp, origoffset, bytes)) {
1373 			goto committed;
1374 		}
1375 		if (nfs_in_tobecommitted_range(vp, origoffset, bytes)) {
1376 			if (weak) {
1377 				lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1378 				return 0;
1379 			} else {
1380 				commitoff = np->n_pushlo;
1381 				commitbytes = (uint32_t)(np->n_pushhi -
1382 							 np->n_pushlo);
1383 				goto commit;
1384 			}
1385 		}
1386 		lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1387 	}
1388 
1389 	/*
1390 	 * otherwise write or commit all the pages.
1391 	 */
1392 
1393 	kva = uvm_pagermapin(pgs, ap->a_count, UVMPAGER_MAPIN_WAITOK|
1394 			     UVMPAGER_MAPIN_WRITE);
1395 
1396 	s = splbio();
1397 	vp->v_numoutput += 2;
1398 	mbp = pool_get(&bufpool, PR_WAITOK);
1399 	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
1400 		    vp, mbp, vp->v_numoutput, bytes);
1401 	splx(s);
1402 	mbp->b_bufsize = npages << PAGE_SHIFT;
1403 	mbp->b_data = (void *)kva;
1404 	mbp->b_resid = mbp->b_bcount = bytes;
1405 	mbp->b_flags = B_BUSY|B_WRITE|B_AGE |
1406 		(async ? B_CALL|B_ASYNC : 0) |
1407 		(curproc == uvm.pagedaemon_proc ? B_PDAEMON : 0);
1408 	mbp->b_iodone = uvm_aio_aiodone;
1409 	mbp->b_vp = vp;
1410 	LIST_INIT(&mbp->b_dep);
1411 
1412 	for (offset = origoffset;
1413 	     bytes > 0;
1414 	     offset += iobytes, bytes -= iobytes) {
1415 		iobytes = min(nmp->nm_wsize, bytes);
1416 
1417 		/*
1418 		 * skip writing any pages which only need a commit.
1419 		 */
1420 
1421 		if ((pgs[(offset - origoffset) >> PAGE_SHIFT]->flags &
1422 		     PG_NEEDCOMMIT) != 0) {
1423 			iobytes = PAGE_SIZE;
1424 			skipbytes += min(iobytes, vp->v_uvm.u_size - offset);
1425 			continue;
1426 		}
1427 
1428 		/* if it's really one i/o, don't make a second buf */
1429 		if (offset == origoffset && iobytes == bytes) {
1430 			bp = mbp;
1431 		} else {
1432 			s = splbio();
1433 			vp->v_numoutput++;
1434 			bp = pool_get(&bufpool, PR_WAITOK);
1435 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
1436 				    vp, bp, vp->v_numoutput, 0);
1437 			splx(s);
1438 			bp->b_data = (char *)kva + (offset - origoffset);
1439 			bp->b_resid = bp->b_bcount = iobytes;
1440 			bp->b_flags = B_BUSY|B_WRITE|B_CALL|B_ASYNC;
1441 			bp->b_iodone = uvm_aio_biodone1;
1442 			bp->b_vp = vp;
1443 			LIST_INIT(&bp->b_dep);
1444 		}
1445 		bp->b_private = mbp;
1446 		bp->b_lblkno = bp->b_blkno = (daddr_t)(offset >> DEV_BSHIFT);
1447 		UVMHIST_LOG(ubchist, "bp %p numout %d",
1448 			    bp, vp->v_numoutput,0,0);
1449 		VOP_STRATEGY(bp);
1450 	}
1451 	if (skipbytes) {
1452 		UVMHIST_LOG(ubchist, "skipbytes %d", bytes, 0,0,0);
1453 		s = splbio();
1454 		mbp->b_resid -= skipbytes;
1455 		if (mbp->b_resid == 0) {
1456 			biodone(mbp);
1457 		}
1458 		splx(s);
1459 	}
1460 	if (async) {
1461 		return EINPROGRESS;
1462 	}
1463 	error = biowait(mbp);
1464 
1465 	s = splbio();
1466 	vwakeup(mbp);
1467 	pool_put(&bufpool, mbp);
1468 	splx(s);
1469 
1470 	uvm_pagermapout(kva, ap->a_count);
1471 	if (error || !v3) {
1472 		UVMHIST_LOG(ubchist, "returning error %d", error, 0,0,0);
1473 		return error;
1474 	}
1475 
1476 	/*
1477 	 * for a weak put, mark the range as "to be committed"
1478 	 * and mark the pages read-only so that we will be notified
1479 	 * to remove the pages from the "to be committed" range
1480 	 * if they are made dirty again.
1481 	 * for a strong put, commit the pages and remove them from the
1482 	 * "to be committed" range.  also, mark them as writable
1483 	 * and not cleanable with just a commit.
1484 	 */
1485 
1486 	lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL);
1487 	if (weak) {
1488 		nfs_add_tobecommitted_range(vp, origoffset,
1489 					    npages << PAGE_SHIFT);
1490 		for (i = 0; i < npages; i++) {
1491 			pgs[i]->flags |= PG_NEEDCOMMIT|PG_RDONLY;
1492 		}
1493 	} else {
1494 		commitoff = origoffset;
1495 		commitbytes = npages << PAGE_SHIFT;
1496 commit:
1497 		error = nfs_commit(vp, commitoff, commitbytes, curproc);
1498 		nfs_del_tobecommitted_range(vp, commitoff, commitbytes);
1499 committed:
1500 		for (i = 0; i < npages; i++) {
1501 			pgs[i]->flags &= ~(PG_NEEDCOMMIT|PG_RDONLY);
1502 		}
1503 	}
1504 	lockmgr(&np->n_commitlock, LK_RELEASE, NULL);
1505 	return error;
1506 }
1507