1 /* $NetBSD: nfs_clbio.c,v 1.7 2021/03/29 02:13:37 simonb Exp $ */
2 /*-
3 * Copyright (c) 1989, 1993
4 * The Regents of the University of California. All rights reserved.
5 *
6 * This code is derived from software contributed to Berkeley by
7 * Rick Macklem at The University of Guelph.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 *
33 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
34 */
35
36 #include <sys/cdefs.h>
37 /* __FBSDID("FreeBSD: head/sys/fs/nfsclient/nfs_clbio.c 304026 2016-08-12 22:44:59Z rmacklem "); */
38 __RCSID("$NetBSD: nfs_clbio.c,v 1.7 2021/03/29 02:13:37 simonb Exp $");
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/buf.h>
43 #include <sys/kernel.h>
44 #include <sys/mount.h>
45 #include <sys/rwlock.h>
46 #include <sys/vmmeter.h>
47 #include <sys/vnode.h>
48
49 #include <fs/nfs/common/nfsport.h>
50 #include <fs/nfs/client/nfsmount.h>
51 #include <fs/nfs/client/nfs.h>
52 #include <fs/nfs/client/nfsnode.h>
53 #include <fs/nfs/client/nfs_kdtrace.h>
54
55 extern int newnfs_directio_allow_mmap;
56 extern struct nfsstatsv1 nfsstatsv1;
57 extern struct mtx ncl_iod_mutex;
58 extern int ncl_numasync;
59 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
60 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
61 extern int newnfs_directio_enable;
62 extern int nfs_keep_dirty_on_error;
63
64 int ncl_pbuf_freecnt = -1; /* start out unlimited */
65
66 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
67 struct thread *td);
68 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
69 struct ucred *cred, int ioflag);
70
71 /*
72 * Vnode op for VM getpages.
73 */
74 int
ncl_getpages(struct vop_getpages_args * ap)75 ncl_getpages(struct vop_getpages_args *ap)
76 {
77 int i, error, nextoff, size, toff, count, npages;
78 struct uio uio;
79 struct iovec iov;
80 vaddr_t kva;
81 struct buf *bp;
82 struct vnode *vp;
83 struct thread *td;
84 struct ucred *cred;
85 struct nfsmount *nmp;
86 vm_object_t object;
87 vm_page_t *pages;
88 struct nfsnode *np;
89
90 vp = ap->a_vp;
91 np = VTONFS(vp);
92 td = curthread; /* XXX */
93 cred = curthread->td_ucred; /* XXX */
94 nmp = VFSTONFS(vp->v_mount);
95 pages = ap->a_m;
96 npages = ap->a_count;
97
98 if ((object = vp->v_object) == NULL) {
99 printf("ncl_getpages: called with non-merged cache vnode\n");
100 return (VM_PAGER_ERROR);
101 }
102
103 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
104 mtx_lock(&np->n_mtx);
105 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
106 mtx_unlock(&np->n_mtx);
107 printf("ncl_getpages: called on non-cacheable vnode\n");
108 return (VM_PAGER_ERROR);
109 } else
110 mtx_unlock(&np->n_mtx);
111 }
112
113 mtx_lock(&nmp->nm_mtx);
114 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
115 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
116 mtx_unlock(&nmp->nm_mtx);
117 /* We'll never get here for v4, because we always have fsinfo */
118 (void)ncl_fsinfo(nmp, vp, cred, td);
119 } else
120 mtx_unlock(&nmp->nm_mtx);
121
122 /*
123 * If the requested page is partially valid, just return it and
124 * allow the pager to zero-out the blanks. Partially valid pages
125 * can only occur at the file EOF.
126 *
127 * XXXGL: is that true for NFS, where short read can occur???
128 */
129 VM_OBJECT_WLOCK(object);
130 if (pages[npages - 1]->valid != 0 && --npages == 0)
131 goto out;
132 VM_OBJECT_WUNLOCK(object);
133
134 /*
135 * We use only the kva address for the buffer, but this is extremely
136 * convenient and fast.
137 */
138 bp = getpbuf(&ncl_pbuf_freecnt);
139
140 kva = (vaddr_t) bp->b_data;
141 pmap_qenter(kva, pages, npages);
142 PCPU_INC(cnt.v_vnodein);
143 PCPU_ADD(cnt.v_vnodepgsin, npages);
144
145 count = npages << PAGE_SHIFT;
146 iov.iov_base = (caddr_t) kva;
147 iov.iov_len = count;
148 uio.uio_iov = &iov;
149 uio.uio_iovcnt = 1;
150 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
151 uio.uio_resid = count;
152 uio.uio_segflg = UIO_SYSSPACE;
153 uio.uio_rw = UIO_READ;
154 uio.uio_td = td;
155
156 error = ncl_readrpc(vp, &uio, cred);
157 pmap_qremove(kva, npages);
158
159 relpbuf(bp, &ncl_pbuf_freecnt);
160
161 if (error && (uio.uio_resid == count)) {
162 printf("ncl_getpages: error %d\n", error);
163 return (VM_PAGER_ERROR);
164 }
165
166 /*
167 * Calculate the number of bytes read and validate only that number
168 * of bytes. Note that due to pending writes, size may be 0. This
169 * does not mean that the remaining data is invalid!
170 */
171
172 size = count - uio.uio_resid;
173 VM_OBJECT_WLOCK(object);
174 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
175 vm_page_t m;
176 nextoff = toff + PAGE_SIZE;
177 m = pages[i];
178
179 if (nextoff <= size) {
180 /*
181 * Read operation filled an entire page
182 */
183 m->valid = VM_PAGE_BITS_ALL;
184 KASSERT(m->dirty == 0,
185 ("nfs_getpages: page %p is dirty", m));
186 } else if (size > toff) {
187 /*
188 * Read operation filled a partial page.
189 */
190 m->valid = 0;
191 vm_page_set_valid_range(m, 0, size - toff);
192 KASSERT(m->dirty == 0,
193 ("nfs_getpages: page %p is dirty", m));
194 } else {
195 /*
196 * Read operation was short. If no error
197 * occurred we may have hit a zero-fill
198 * section. We leave valid set to 0, and page
199 * is freed by vm_page_readahead_finish() if
200 * its index is not equal to requested, or
201 * page is zeroed and set valid by
202 * vm_pager_get_pages() for requested page.
203 */
204 ;
205 }
206 }
207 out:
208 VM_OBJECT_WUNLOCK(object);
209 if (ap->a_rbehind)
210 *ap->a_rbehind = 0;
211 if (ap->a_rahead)
212 *ap->a_rahead = 0;
213 return (VM_PAGER_OK);
214 }
215
216 /*
217 * Vnode op for VM putpages.
218 */
219 int
ncl_putpages(struct vop_putpages_args * ap)220 ncl_putpages(struct vop_putpages_args *ap)
221 {
222 struct uio uio;
223 struct iovec iov;
224 vaddr_t kva;
225 struct buf *bp;
226 int iomode, must_commit, i, error, npages, count;
227 off_t offset;
228 int *rtvals;
229 struct vnode *vp;
230 struct thread *td;
231 struct ucred *cred;
232 struct nfsmount *nmp;
233 struct nfsnode *np;
234 vm_page_t *pages;
235
236 vp = ap->a_vp;
237 np = VTONFS(vp);
238 td = curthread; /* XXX */
239 /* Set the cred to n_writecred for the write rpcs. */
240 if (np->n_writecred != NULL)
241 cred = crhold(np->n_writecred);
242 else
243 cred = crhold(curthread->td_ucred); /* XXX */
244 nmp = VFSTONFS(vp->v_mount);
245 pages = ap->a_m;
246 count = ap->a_count;
247 rtvals = ap->a_rtvals;
248 npages = btoc(count);
249 offset = IDX_TO_OFF(pages[0]->pindex);
250
251 mtx_lock(&nmp->nm_mtx);
252 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
253 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
254 mtx_unlock(&nmp->nm_mtx);
255 (void)ncl_fsinfo(nmp, vp, cred, td);
256 } else
257 mtx_unlock(&nmp->nm_mtx);
258
259 mtx_lock(&np->n_mtx);
260 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
261 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
262 mtx_unlock(&np->n_mtx);
263 printf("ncl_putpages: called on noncache-able vnode\n");
264 mtx_lock(&np->n_mtx);
265 }
266
267 for (i = 0; i < npages; i++)
268 rtvals[i] = VM_PAGER_ERROR;
269
270 /*
271 * When putting pages, do not extend file past EOF.
272 */
273 if (offset + count > np->n_size) {
274 count = np->n_size - offset;
275 if (count < 0)
276 count = 0;
277 }
278 mtx_unlock(&np->n_mtx);
279
280 /*
281 * We use only the kva address for the buffer, but this is extremely
282 * convenient and fast.
283 */
284 bp = getpbuf(&ncl_pbuf_freecnt);
285
286 kva = (vaddr_t) bp->b_data;
287 pmap_qenter(kva, pages, npages);
288 PCPU_INC(cnt.v_vnodeout);
289 PCPU_ADD(cnt.v_vnodepgsout, count);
290
291 iov.iov_base = (caddr_t) kva;
292 iov.iov_len = count;
293 uio.uio_iov = &iov;
294 uio.uio_iovcnt = 1;
295 uio.uio_offset = offset;
296 uio.uio_resid = count;
297 uio.uio_segflg = UIO_SYSSPACE;
298 uio.uio_rw = UIO_WRITE;
299 uio.uio_td = td;
300
301 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
302 iomode = NFSWRITE_UNSTABLE;
303 else
304 iomode = NFSWRITE_FILESYNC;
305
306 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
307 crfree(cred);
308
309 pmap_qremove(kva, npages);
310 relpbuf(bp, &ncl_pbuf_freecnt);
311
312 if (error == 0 || !nfs_keep_dirty_on_error) {
313 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
314 if (must_commit)
315 ncl_clearcommit(vp->v_mount);
316 }
317 return rtvals[0];
318 }
319
320 /*
321 * For nfs, cache consistency can only be maintained approximately.
322 * Although RFC1094 does not specify the criteria, the following is
323 * believed to be compatible with the reference port.
324 * For nfs:
325 * If the file's modify time on the server has changed since the
326 * last read rpc or you have written to the file,
327 * you may have lost data cache consistency with the
328 * server, so flush all of the file's data out of the cache.
329 * Then force a getattr rpc to ensure that you have up to date
330 * attributes.
331 * NB: This implies that cache data can be read when up to
332 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
333 * attributes this could be forced by setting n_attrstamp to 0 before
334 * the VOP_GETATTR() call.
335 */
336 static inline int
nfs_bioread_check_cons(struct vnode * vp,struct thread * td,struct ucred * cred)337 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
338 {
339 int error = 0;
340 struct vattr vattr;
341 struct nfsnode *np = VTONFS(vp);
342 int old_lock;
343
344 /*
345 * Grab the exclusive lock before checking whether the cache is
346 * consistent.
347 * XXX - We can make this cheaper later (by acquiring cheaper locks).
348 * But for now, this suffices.
349 */
350 old_lock = ncl_upgrade_vnlock(vp);
351 if (vp->v_iflag & VI_DOOMED) {
352 ncl_downgrade_vnlock(vp, old_lock);
353 return (EBADF);
354 }
355
356 mtx_lock(&np->n_mtx);
357 if (np->n_flag & NMODIFIED) {
358 mtx_unlock(&np->n_mtx);
359 if (vp->v_type != VREG) {
360 if (vp->v_type != VDIR)
361 panic("nfs: bioread, not dir");
362 ncl_invaldir(vp);
363 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
364 if (error)
365 goto out;
366 }
367 np->n_attrstamp = 0;
368 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
369 error = VOP_GETATTR(vp, &vattr, cred);
370 if (error)
371 goto out;
372 mtx_lock(&np->n_mtx);
373 np->n_mtime = vattr.va_mtime;
374 mtx_unlock(&np->n_mtx);
375 } else {
376 mtx_unlock(&np->n_mtx);
377 error = VOP_GETATTR(vp, &vattr, cred);
378 if (error)
379 return (error);
380 mtx_lock(&np->n_mtx);
381 if ((np->n_flag & NSIZECHANGED)
382 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
383 mtx_unlock(&np->n_mtx);
384 if (vp->v_type == VDIR)
385 ncl_invaldir(vp);
386 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
387 if (error)
388 goto out;
389 mtx_lock(&np->n_mtx);
390 np->n_mtime = vattr.va_mtime;
391 np->n_flag &= ~NSIZECHANGED;
392 }
393 mtx_unlock(&np->n_mtx);
394 }
395 out:
396 ncl_downgrade_vnlock(vp, old_lock);
397 return error;
398 }
399
400 /*
401 * Vnode op for read using bio
402 */
403 int
ncl_bioread(struct vnode * vp,struct uio * uio,int ioflag,struct ucred * cred)404 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
405 {
406 struct nfsnode *np = VTONFS(vp);
407 int biosize, i;
408 struct buf *bp, *rabp;
409 struct thread *td;
410 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
411 daddr_t lbn, rabn;
412 int bcount;
413 int seqcount;
414 int nra, error = 0, n = 0, on = 0;
415 off_t tmp_off;
416
417 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
418 if (uio->uio_resid == 0)
419 return (0);
420 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
421 return (EINVAL);
422 td = uio->uio_td;
423
424 mtx_lock(&nmp->nm_mtx);
425 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
426 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
427 mtx_unlock(&nmp->nm_mtx);
428 (void)ncl_fsinfo(nmp, vp, cred, td);
429 mtx_lock(&nmp->nm_mtx);
430 }
431 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
432 (void) newnfs_iosize(nmp);
433
434 tmp_off = uio->uio_offset + uio->uio_resid;
435 if (vp->v_type != VDIR &&
436 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
437 mtx_unlock(&nmp->nm_mtx);
438 return (EFBIG);
439 }
440 mtx_unlock(&nmp->nm_mtx);
441
442 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
443 /* No caching/ no readaheads. Just read data into the user buffer */
444 return ncl_readrpc(vp, uio, cred);
445
446 biosize = vp->v_bufobj.bo_bsize;
447 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
448
449 error = nfs_bioread_check_cons(vp, td, cred);
450 if (error)
451 return error;
452
453 do {
454 u_quad_t nsize;
455
456 mtx_lock(&np->n_mtx);
457 nsize = np->n_size;
458 mtx_unlock(&np->n_mtx);
459
460 switch (vp->v_type) {
461 case VREG:
462 NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
463 lbn = uio->uio_offset / biosize;
464 on = uio->uio_offset - (lbn * biosize);
465
466 /*
467 * Start the read ahead(s), as required.
468 */
469 if (nmp->nm_readahead > 0) {
470 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
471 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
472 rabn = lbn + 1 + nra;
473 if (incore(&vp->v_bufobj, rabn) == NULL) {
474 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
475 if (!rabp) {
476 error = newnfs_sigintr(nmp, td);
477 return (error ? error : EINTR);
478 }
479 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
480 rabp->b_flags |= B_ASYNC;
481 rabp->b_iocmd = BIO_READ;
482 vfs_busy_pages(rabp, 0);
483 if (ncl_asyncio(nmp, rabp, cred, td)) {
484 rabp->b_flags |= B_INVAL;
485 rabp->b_ioflags |= BIO_ERROR;
486 vfs_unbusy_pages(rabp);
487 brelse(rabp);
488 break;
489 }
490 } else {
491 brelse(rabp);
492 }
493 }
494 }
495 }
496
497 /* Note that bcount is *not* DEV_BSIZE aligned. */
498 bcount = biosize;
499 if ((off_t)lbn * biosize >= nsize) {
500 bcount = 0;
501 } else if ((off_t)(lbn + 1) * biosize > nsize) {
502 bcount = nsize - (off_t)lbn * biosize;
503 }
504 bp = nfs_getcacheblk(vp, lbn, bcount, td);
505
506 if (!bp) {
507 error = newnfs_sigintr(nmp, td);
508 return (error ? error : EINTR);
509 }
510
511 /*
512 * If B_CACHE is not set, we must issue the read. If this
513 * fails, we return an error.
514 */
515
516 if ((bp->b_flags & B_CACHE) == 0) {
517 bp->b_iocmd = BIO_READ;
518 vfs_busy_pages(bp, 0);
519 error = ncl_doio(vp, bp, cred, td, 0);
520 if (error) {
521 brelse(bp);
522 return (error);
523 }
524 }
525
526 /*
527 * on is the offset into the current bp. Figure out how many
528 * bytes we can copy out of the bp. Note that bcount is
529 * NOT DEV_BSIZE aligned.
530 *
531 * Then figure out how many bytes we can copy into the uio.
532 */
533
534 n = 0;
535 if (on < bcount)
536 n = MIN((unsigned)(bcount - on), uio->uio_resid);
537 break;
538 case VLNK:
539 NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
540 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
541 if (!bp) {
542 error = newnfs_sigintr(nmp, td);
543 return (error ? error : EINTR);
544 }
545 if ((bp->b_flags & B_CACHE) == 0) {
546 bp->b_iocmd = BIO_READ;
547 vfs_busy_pages(bp, 0);
548 error = ncl_doio(vp, bp, cred, td, 0);
549 if (error) {
550 bp->b_ioflags |= BIO_ERROR;
551 brelse(bp);
552 return (error);
553 }
554 }
555 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
556 on = 0;
557 break;
558 case VDIR:
559 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
560 if (np->n_direofoffset
561 && uio->uio_offset >= np->n_direofoffset) {
562 return (0);
563 }
564 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
565 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
566 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
567 if (!bp) {
568 error = newnfs_sigintr(nmp, td);
569 return (error ? error : EINTR);
570 }
571 if ((bp->b_flags & B_CACHE) == 0) {
572 bp->b_iocmd = BIO_READ;
573 vfs_busy_pages(bp, 0);
574 error = ncl_doio(vp, bp, cred, td, 0);
575 if (error) {
576 brelse(bp);
577 }
578 while (error == NFSERR_BAD_COOKIE) {
579 ncl_invaldir(vp);
580 error = ncl_vinvalbuf(vp, 0, td, 1);
581 /*
582 * Yuck! The directory has been modified on the
583 * server. The only way to get the block is by
584 * reading from the beginning to get all the
585 * offset cookies.
586 *
587 * Leave the last bp intact unless there is an error.
588 * Loop back up to the while if the error is another
589 * NFSERR_BAD_COOKIE (double yuch!).
590 */
591 for (i = 0; i <= lbn && !error; i++) {
592 if (np->n_direofoffset
593 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
594 return (0);
595 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
596 if (!bp) {
597 error = newnfs_sigintr(nmp, td);
598 return (error ? error : EINTR);
599 }
600 if ((bp->b_flags & B_CACHE) == 0) {
601 bp->b_iocmd = BIO_READ;
602 vfs_busy_pages(bp, 0);
603 error = ncl_doio(vp, bp, cred, td, 0);
604 /*
605 * no error + B_INVAL == directory EOF,
606 * use the block.
607 */
608 if (error == 0 && (bp->b_flags & B_INVAL))
609 break;
610 }
611 /*
612 * An error will throw away the block and the
613 * for loop will break out. If no error and this
614 * is not the block we want, we throw away the
615 * block and go for the next one via the for loop.
616 */
617 if (error || i < lbn)
618 brelse(bp);
619 }
620 }
621 /*
622 * The above while is repeated if we hit another cookie
623 * error. If we hit an error and it wasn't a cookie error,
624 * we give up.
625 */
626 if (error)
627 return (error);
628 }
629
630 /*
631 * If not eof and read aheads are enabled, start one.
632 * (You need the current block first, so that you have the
633 * directory offset cookie of the next block.)
634 */
635 if (nmp->nm_readahead > 0 &&
636 (bp->b_flags & B_INVAL) == 0 &&
637 (np->n_direofoffset == 0 ||
638 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
639 incore(&vp->v_bufobj, lbn + 1) == NULL) {
640 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
641 if (rabp) {
642 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
643 rabp->b_flags |= B_ASYNC;
644 rabp->b_iocmd = BIO_READ;
645 vfs_busy_pages(rabp, 0);
646 if (ncl_asyncio(nmp, rabp, cred, td)) {
647 rabp->b_flags |= B_INVAL;
648 rabp->b_ioflags |= BIO_ERROR;
649 vfs_unbusy_pages(rabp);
650 brelse(rabp);
651 }
652 } else {
653 brelse(rabp);
654 }
655 }
656 }
657 /*
658 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
659 * chopped for the EOF condition, we cannot tell how large
660 * NFS directories are going to be until we hit EOF. So
661 * an NFS directory buffer is *not* chopped to its EOF. Now,
662 * it just so happens that b_resid will effectively chop it
663 * to EOF. *BUT* this information is lost if the buffer goes
664 * away and is reconstituted into a B_CACHE state ( due to
665 * being VMIO ) later. So we keep track of the directory eof
666 * in np->n_direofoffset and chop it off as an extra step
667 * right here.
668 */
669 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
670 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
671 n = np->n_direofoffset - uio->uio_offset;
672 break;
673 default:
674 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
675 bp = NULL;
676 break;
677 }
678
679 if (n > 0) {
680 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
681 }
682 if (vp->v_type == VLNK)
683 n = 0;
684 if (bp != NULL)
685 brelse(bp);
686 } while (error == 0 && uio->uio_resid > 0 && n > 0);
687 return (error);
688 }
689
690 /*
691 * The NFS write path cannot handle iovecs with len > 1. So we need to
692 * break up iovecs accordingly (restricting them to wsize).
693 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
694 * For the ASYNC case, 2 copies are needed. The first a copy from the
695 * user buffer to a staging buffer and then a second copy from the staging
696 * buffer to mbufs. This can be optimized by copying from the user buffer
697 * directly into mbufs and passing the chain down, but that requires a
698 * fair amount of re-working of the relevant codepaths (and can be done
699 * later).
700 */
701 static int
nfs_directio_write(vp,uiop,cred,ioflag)702 nfs_directio_write(vp, uiop, cred, ioflag)
703 struct vnode *vp;
704 struct uio *uiop;
705 struct ucred *cred;
706 int ioflag;
707 {
708 int error;
709 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
710 struct thread *td = uiop->uio_td;
711 int size;
712 int wsize;
713
714 mtx_lock(&nmp->nm_mtx);
715 wsize = nmp->nm_wsize;
716 mtx_unlock(&nmp->nm_mtx);
717 if (ioflag & IO_SYNC) {
718 int iomode, must_commit;
719 struct uio uio;
720 struct iovec iov;
721 do_sync:
722 while (uiop->uio_resid > 0) {
723 size = MIN(uiop->uio_resid, wsize);
724 size = MIN(uiop->uio_iov->iov_len, size);
725 iov.iov_base = uiop->uio_iov->iov_base;
726 iov.iov_len = size;
727 uio.uio_iov = &iov;
728 uio.uio_iovcnt = 1;
729 uio.uio_offset = uiop->uio_offset;
730 uio.uio_resid = size;
731 uio.uio_segflg = UIO_USERSPACE;
732 uio.uio_rw = UIO_WRITE;
733 uio.uio_td = td;
734 iomode = NFSWRITE_FILESYNC;
735 error = ncl_writerpc(vp, &uio, cred, &iomode,
736 &must_commit, 0);
737 KASSERT((must_commit == 0),
738 ("ncl_directio_write: Did not commit write"));
739 if (error)
740 return (error);
741 uiop->uio_offset += size;
742 uiop->uio_resid -= size;
743 if (uiop->uio_iov->iov_len <= size) {
744 uiop->uio_iovcnt--;
745 uiop->uio_iov++;
746 } else {
747 uiop->uio_iov->iov_base =
748 (char *)uiop->uio_iov->iov_base + size;
749 uiop->uio_iov->iov_len -= size;
750 }
751 }
752 } else {
753 struct uio *t_uio;
754 struct iovec *t_iov;
755 struct buf *bp;
756
757 /*
758 * Break up the write into blocksize chunks and hand these
759 * over to nfsiod's for write back.
760 * Unfortunately, this incurs a copy of the data. Since
761 * the user could modify the buffer before the write is
762 * initiated.
763 *
764 * The obvious optimization here is that one of the 2 copies
765 * in the async write path can be eliminated by copying the
766 * data here directly into mbufs and passing the mbuf chain
767 * down. But that will require a fair amount of re-working
768 * of the code and can be done if there's enough interest
769 * in NFS directio access.
770 */
771 while (uiop->uio_resid > 0) {
772 size = MIN(uiop->uio_resid, wsize);
773 size = MIN(uiop->uio_iov->iov_len, size);
774 bp = getpbuf(&ncl_pbuf_freecnt);
775 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
776 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
777 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
778 t_iov->iov_len = size;
779 t_uio->uio_iov = t_iov;
780 t_uio->uio_iovcnt = 1;
781 t_uio->uio_offset = uiop->uio_offset;
782 t_uio->uio_resid = size;
783 t_uio->uio_segflg = UIO_SYSSPACE;
784 t_uio->uio_rw = UIO_WRITE;
785 t_uio->uio_td = td;
786 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
787 uiop->uio_segflg == UIO_SYSSPACE,
788 ("nfs_directio_write: Bad uio_segflg"));
789 if (uiop->uio_segflg == UIO_USERSPACE) {
790 error = copyin(uiop->uio_iov->iov_base,
791 t_iov->iov_base, size);
792 if (error != 0)
793 goto err_free;
794 } else
795 /*
796 * UIO_SYSSPACE may never happen, but handle
797 * it just in case it does.
798 */
799 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
800 size);
801 bp->b_flags |= B_DIRECT;
802 bp->b_iocmd = BIO_WRITE;
803 if (cred != NOCRED) {
804 crhold(cred);
805 bp->b_wcred = cred;
806 } else
807 bp->b_wcred = NOCRED;
808 bp->b_caller1 = (void *)t_uio;
809 bp->b_vp = vp;
810 error = ncl_asyncio(nmp, bp, NOCRED, td);
811 err_free:
812 if (error) {
813 free(t_iov->iov_base, M_NFSDIRECTIO);
814 free(t_iov, M_NFSDIRECTIO);
815 free(t_uio, M_NFSDIRECTIO);
816 bp->b_vp = NULL;
817 relpbuf(bp, &ncl_pbuf_freecnt);
818 if (error == EINTR)
819 return (error);
820 goto do_sync;
821 }
822 uiop->uio_offset += size;
823 uiop->uio_resid -= size;
824 if (uiop->uio_iov->iov_len <= size) {
825 uiop->uio_iovcnt--;
826 uiop->uio_iov++;
827 } else {
828 uiop->uio_iov->iov_base =
829 (char *)uiop->uio_iov->iov_base + size;
830 uiop->uio_iov->iov_len -= size;
831 }
832 }
833 }
834 return (0);
835 }
836
837 /*
838 * Vnode op for write using bio
839 */
840 int
ncl_write(struct vop_write_args * ap)841 ncl_write(struct vop_write_args *ap)
842 {
843 int biosize;
844 struct uio *uio = ap->a_uio;
845 struct thread *td = uio->uio_td;
846 struct vnode *vp = ap->a_vp;
847 struct nfsnode *np = VTONFS(vp);
848 struct ucred *cred = ap->a_cred;
849 int ioflag = ap->a_ioflag;
850 struct buf *bp;
851 struct vattr vattr;
852 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
853 daddr_t lbn;
854 int bcount, noncontig_write, obcount;
855 int bp_cached, n, on, error = 0, error1, wouldcommit;
856 size_t orig_resid, local_resid;
857 off_t orig_size, tmp_off;
858
859 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
860 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
861 ("ncl_write proc"));
862 if (vp->v_type != VREG)
863 return (EIO);
864 mtx_lock(&np->n_mtx);
865 if (np->n_flag & NWRITEERR) {
866 np->n_flag &= ~NWRITEERR;
867 mtx_unlock(&np->n_mtx);
868 return (np->n_error);
869 } else
870 mtx_unlock(&np->n_mtx);
871 mtx_lock(&nmp->nm_mtx);
872 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
873 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
874 mtx_unlock(&nmp->nm_mtx);
875 (void)ncl_fsinfo(nmp, vp, cred, td);
876 mtx_lock(&nmp->nm_mtx);
877 }
878 if (nmp->nm_wsize == 0)
879 (void) newnfs_iosize(nmp);
880 mtx_unlock(&nmp->nm_mtx);
881
882 /*
883 * Synchronously flush pending buffers if we are in synchronous
884 * mode or if we are appending.
885 */
886 if (ioflag & (IO_APPEND | IO_SYNC)) {
887 mtx_lock(&np->n_mtx);
888 if (np->n_flag & NMODIFIED) {
889 mtx_unlock(&np->n_mtx);
890 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
891 /*
892 * Require non-blocking, synchronous writes to
893 * dirty files to inform the program it needs
894 * to fsync(2) explicitly.
895 */
896 if (ioflag & IO_NDELAY)
897 return (EAGAIN);
898 #endif
899 np->n_attrstamp = 0;
900 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
901 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
902 if (error)
903 return (error);
904 } else
905 mtx_unlock(&np->n_mtx);
906 }
907
908 orig_resid = uio->uio_resid;
909 mtx_lock(&np->n_mtx);
910 orig_size = np->n_size;
911 mtx_unlock(&np->n_mtx);
912
913 /*
914 * If IO_APPEND then load uio_offset. We restart here if we cannot
915 * get the append lock.
916 */
917 if (ioflag & IO_APPEND) {
918 np->n_attrstamp = 0;
919 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
920 error = VOP_GETATTR(vp, &vattr, cred);
921 if (error)
922 return (error);
923 mtx_lock(&np->n_mtx);
924 uio->uio_offset = np->n_size;
925 mtx_unlock(&np->n_mtx);
926 }
927
928 if (uio->uio_offset < 0)
929 return (EINVAL);
930 tmp_off = uio->uio_offset + uio->uio_resid;
931 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
932 return (EFBIG);
933 if (uio->uio_resid == 0)
934 return (0);
935
936 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
937 return nfs_directio_write(vp, uio, cred, ioflag);
938
939 /*
940 * Maybe this should be above the vnode op call, but so long as
941 * file servers have no limits, i don't think it matters
942 */
943 if (vn_rlimit_fsize(vp, uio, td))
944 return (EFBIG);
945
946 biosize = vp->v_bufobj.bo_bsize;
947 /*
948 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
949 * would exceed the local maximum per-file write commit size when
950 * combined with those, we must decide whether to flush,
951 * go synchronous, or return error. We don't bother checking
952 * IO_UNIT -- we just make all writes atomic anyway, as there's
953 * no point optimizing for something that really won't ever happen.
954 */
955 wouldcommit = 0;
956 if (!(ioflag & IO_SYNC)) {
957 int nflag;
958
959 mtx_lock(&np->n_mtx);
960 nflag = np->n_flag;
961 mtx_unlock(&np->n_mtx);
962 if (nflag & NMODIFIED) {
963 BO_LOCK(&vp->v_bufobj);
964 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
965 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
966 b_bobufs) {
967 if (bp->b_flags & B_NEEDCOMMIT)
968 wouldcommit += bp->b_bcount;
969 }
970 }
971 BO_UNLOCK(&vp->v_bufobj);
972 }
973 }
974
975 do {
976 if (!(ioflag & IO_SYNC)) {
977 wouldcommit += biosize;
978 if (wouldcommit > nmp->nm_wcommitsize) {
979 np->n_attrstamp = 0;
980 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
981 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
982 if (error)
983 return (error);
984 wouldcommit = biosize;
985 }
986 }
987
988 NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
989 lbn = uio->uio_offset / biosize;
990 on = uio->uio_offset - (lbn * biosize);
991 n = MIN((unsigned)(biosize - on), uio->uio_resid);
992 again:
993 /*
994 * Handle direct append and file extension cases, calculate
995 * unaligned buffer size.
996 */
997 mtx_lock(&np->n_mtx);
998 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
999 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1000 noncontig_write = 1;
1001 else
1002 noncontig_write = 0;
1003 if ((uio->uio_offset == np->n_size ||
1004 (noncontig_write != 0 &&
1005 lbn == (np->n_size / biosize) &&
1006 uio->uio_offset + n > np->n_size)) && n) {
1007 mtx_unlock(&np->n_mtx);
1008 /*
1009 * Get the buffer (in its pre-append state to maintain
1010 * B_CACHE if it was previously set). Resize the
1011 * nfsnode after we have locked the buffer to prevent
1012 * readers from reading garbage.
1013 */
1014 obcount = np->n_size - (lbn * biosize);
1015 bp = nfs_getcacheblk(vp, lbn, obcount, td);
1016
1017 if (bp != NULL) {
1018 long save;
1019
1020 mtx_lock(&np->n_mtx);
1021 np->n_size = uio->uio_offset + n;
1022 np->n_flag |= NMODIFIED;
1023 vnode_pager_setsize(vp, np->n_size);
1024 mtx_unlock(&np->n_mtx);
1025
1026 save = bp->b_flags & B_CACHE;
1027 bcount = on + n;
1028 allocbuf(bp, bcount);
1029 bp->b_flags |= save;
1030 if (noncontig_write != 0 && on > obcount)
1031 vfs_bio_bzero_buf(bp, obcount, on -
1032 obcount);
1033 }
1034 } else {
1035 /*
1036 * Obtain the locked cache block first, and then
1037 * adjust the file's size as appropriate.
1038 */
1039 bcount = on + n;
1040 if ((off_t)lbn * biosize + bcount < np->n_size) {
1041 if ((off_t)(lbn + 1) * biosize < np->n_size)
1042 bcount = biosize;
1043 else
1044 bcount = np->n_size - (off_t)lbn * biosize;
1045 }
1046 mtx_unlock(&np->n_mtx);
1047 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1048 mtx_lock(&np->n_mtx);
1049 if (uio->uio_offset + n > np->n_size) {
1050 np->n_size = uio->uio_offset + n;
1051 np->n_flag |= NMODIFIED;
1052 vnode_pager_setsize(vp, np->n_size);
1053 }
1054 mtx_unlock(&np->n_mtx);
1055 }
1056
1057 if (!bp) {
1058 error = newnfs_sigintr(nmp, td);
1059 if (!error)
1060 error = EINTR;
1061 break;
1062 }
1063
1064 /*
1065 * Issue a READ if B_CACHE is not set. In special-append
1066 * mode, B_CACHE is based on the buffer prior to the write
1067 * op and is typically set, avoiding the read. If a read
1068 * is required in special append mode, the server will
1069 * probably send us a short-read since we extended the file
1070 * on our end, resulting in b_resid == 0 and, thusly,
1071 * B_CACHE getting set.
1072 *
1073 * We can also avoid issuing the read if the write covers
1074 * the entire buffer. We have to make sure the buffer state
1075 * is reasonable in this case since we will not be initiating
1076 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1077 * more information.
1078 *
1079 * B_CACHE may also be set due to the buffer being cached
1080 * normally.
1081 */
1082
1083 bp_cached = 1;
1084 if (on == 0 && n == bcount) {
1085 if ((bp->b_flags & B_CACHE) == 0)
1086 bp_cached = 0;
1087 bp->b_flags |= B_CACHE;
1088 bp->b_flags &= ~B_INVAL;
1089 bp->b_ioflags &= ~BIO_ERROR;
1090 }
1091
1092 if ((bp->b_flags & B_CACHE) == 0) {
1093 bp->b_iocmd = BIO_READ;
1094 vfs_busy_pages(bp, 0);
1095 error = ncl_doio(vp, bp, cred, td, 0);
1096 if (error) {
1097 brelse(bp);
1098 break;
1099 }
1100 }
1101 if (bp->b_wcred == NOCRED)
1102 bp->b_wcred = crhold(cred);
1103 mtx_lock(&np->n_mtx);
1104 np->n_flag |= NMODIFIED;
1105 mtx_unlock(&np->n_mtx);
1106
1107 /*
1108 * If dirtyend exceeds file size, chop it down. This should
1109 * not normally occur but there is an append race where it
1110 * might occur XXX, so we log it.
1111 *
1112 * If the chopping creates a reverse-indexed or degenerate
1113 * situation with dirtyoff/end, we 0 both of them.
1114 */
1115
1116 if (bp->b_dirtyend > bcount) {
1117 printf("NFS append race @%lx:%d\n",
1118 (long)bp->b_blkno * DEV_BSIZE,
1119 bp->b_dirtyend - bcount);
1120 bp->b_dirtyend = bcount;
1121 }
1122
1123 if (bp->b_dirtyoff >= bp->b_dirtyend)
1124 bp->b_dirtyoff = bp->b_dirtyend = 0;
1125
1126 /*
1127 * If the new write will leave a contiguous dirty
1128 * area, just update the b_dirtyoff and b_dirtyend,
1129 * otherwise force a write rpc of the old dirty area.
1130 *
1131 * If there has been a file lock applied to this file
1132 * or vfs.nfs.old_noncontig_writing is set, do the following:
1133 * While it is possible to merge discontiguous writes due to
1134 * our having a B_CACHE buffer ( and thus valid read data
1135 * for the hole), we don't because it could lead to
1136 * significant cache coherency problems with multiple clients,
1137 * especially if locking is implemented later on.
1138 *
1139 * If vfs.nfs.old_noncontig_writing is not set and there has
1140 * not been file locking done on this file:
1141 * Relax coherency a bit for the sake of performance and
1142 * expand the current dirty region to contain the new
1143 * write even if it means we mark some non-dirty data as
1144 * dirty.
1145 */
1146
1147 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1148 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1149 if (bwrite(bp) == EINTR) {
1150 error = EINTR;
1151 break;
1152 }
1153 goto again;
1154 }
1155
1156 local_resid = uio->uio_resid;
1157 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1158
1159 if (error != 0 && !bp_cached) {
1160 /*
1161 * This block has no other content than what
1162 * possibly was written by the faulty uiomove.
1163 * Release it, forgetting the data pages, to
1164 * prevent the leak of uninitialized data to
1165 * usermode.
1166 */
1167 bp->b_ioflags |= BIO_ERROR;
1168 brelse(bp);
1169 uio->uio_offset -= local_resid - uio->uio_resid;
1170 uio->uio_resid = local_resid;
1171 break;
1172 }
1173
1174 /*
1175 * Since this block is being modified, it must be written
1176 * again and not just committed. Since write clustering does
1177 * not work for the stage 1 data write, only the stage 2
1178 * commit rpc, we have to clear B_CLUSTEROK as well.
1179 */
1180 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1181
1182 /*
1183 * Get the partial update on the progress made from
1184 * uiomove, if an error occurred.
1185 */
1186 if (error != 0)
1187 n = local_resid - uio->uio_resid;
1188
1189 /*
1190 * Only update dirtyoff/dirtyend if not a degenerate
1191 * condition.
1192 */
1193 if (n > 0) {
1194 if (bp->b_dirtyend > 0) {
1195 bp->b_dirtyoff = uimin(on, bp->b_dirtyoff);
1196 bp->b_dirtyend = uimax((on + n), bp->b_dirtyend);
1197 } else {
1198 bp->b_dirtyoff = on;
1199 bp->b_dirtyend = on + n;
1200 }
1201 vfs_bio_set_valid(bp, on, n);
1202 }
1203
1204 /*
1205 * If IO_SYNC do bwrite().
1206 *
1207 * IO_INVAL appears to be unused. The idea appears to be
1208 * to turn off caching in this case. Very odd. XXX
1209 */
1210 if ((ioflag & IO_SYNC)) {
1211 if (ioflag & IO_INVAL)
1212 bp->b_flags |= B_NOCACHE;
1213 error1 = bwrite(bp);
1214 if (error1 != 0) {
1215 if (error == 0)
1216 error = error1;
1217 break;
1218 }
1219 } else if ((n + on) == biosize) {
1220 bp->b_flags |= B_ASYNC;
1221 (void) ncl_writebp(bp, 0, NULL);
1222 } else {
1223 bdwrite(bp);
1224 }
1225
1226 if (error != 0)
1227 break;
1228 } while (uio->uio_resid > 0 && n > 0);
1229
1230 if (error != 0) {
1231 if (ioflag & IO_UNIT) {
1232 VATTR_NULL(&vattr);
1233 vattr.va_size = orig_size;
1234 /* IO_SYNC is handled implicitly */
1235 (void)VOP_SETATTR(vp, &vattr, cred);
1236 uio->uio_offset -= orig_resid - uio->uio_resid;
1237 uio->uio_resid = orig_resid;
1238 }
1239 }
1240
1241 return (error);
1242 }
1243
1244 /*
1245 * Get an nfs cache block.
1246 *
1247 * Allocate a new one if the block isn't currently in the cache
1248 * and return the block marked busy. If the calling process is
1249 * interrupted by a signal for an interruptible mount point, return
1250 * NULL.
1251 *
1252 * The caller must carefully deal with the possible B_INVAL state of
1253 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1254 * indirectly), so synchronous reads can be issued without worrying about
1255 * the B_INVAL state. We have to be a little more careful when dealing
1256 * with writes (see comments in nfs_write()) when extending a file past
1257 * its EOF.
1258 */
1259 static struct buf *
nfs_getcacheblk(struct vnode * vp,daddr_t bn,int size,struct thread * td)1260 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1261 {
1262 struct buf *bp;
1263 struct mount *mp;
1264 struct nfsmount *nmp;
1265
1266 mp = vp->v_mount;
1267 nmp = VFSTONFS(mp);
1268
1269 if (nmp->nm_flag & NFSMNT_INT) {
1270 sigset_t oldset;
1271
1272 newnfs_set_sigmask(td, &oldset);
1273 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1274 newnfs_restore_sigmask(td, &oldset);
1275 while (bp == NULL) {
1276 if (newnfs_sigintr(nmp, td))
1277 return (NULL);
1278 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1279 }
1280 } else {
1281 bp = getblk(vp, bn, size, 0, 0, 0);
1282 }
1283
1284 if (vp->v_type == VREG)
1285 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1286 return (bp);
1287 }
1288
1289 /*
1290 * Flush and invalidate all dirty buffers. If another process is already
1291 * doing the flush, just wait for completion.
1292 */
1293 int
ncl_vinvalbuf(struct vnode * vp,int flags,struct thread * td,int intrflg)1294 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1295 {
1296 struct nfsnode *np = VTONFS(vp);
1297 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1298 int error = 0, slpflag, slptimeo;
1299 int old_lock = 0;
1300
1301 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1302
1303 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1304 intrflg = 0;
1305 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1306 intrflg = 1;
1307 if (intrflg) {
1308 slpflag = PCATCH;
1309 slptimeo = 2 * hz;
1310 } else {
1311 slpflag = 0;
1312 slptimeo = 0;
1313 }
1314
1315 old_lock = ncl_upgrade_vnlock(vp);
1316 if (vp->v_iflag & VI_DOOMED) {
1317 /*
1318 * Since vgonel() uses the generic vinvalbuf() to flush
1319 * dirty buffers and it does not call this function, it
1320 * is safe to just return OK when VI_DOOMED is set.
1321 */
1322 ncl_downgrade_vnlock(vp, old_lock);
1323 return (0);
1324 }
1325
1326 /*
1327 * Now, flush as required.
1328 */
1329 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1330 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1331 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1332 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1333 /*
1334 * If the page clean was interrupted, fail the invalidation.
1335 * Not doing so, we run the risk of losing dirty pages in the
1336 * vinvalbuf() call below.
1337 */
1338 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1339 goto out;
1340 }
1341
1342 error = vinvalbuf(vp, flags, slpflag, 0);
1343 while (error) {
1344 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1345 goto out;
1346 error = vinvalbuf(vp, flags, 0, slptimeo);
1347 }
1348 if (NFSHASPNFS(nmp)) {
1349 nfscl_layoutcommit(vp, td);
1350 /*
1351 * Invalidate the attribute cache, since writes to a DS
1352 * won't update the size attribute.
1353 */
1354 mtx_lock(&np->n_mtx);
1355 np->n_attrstamp = 0;
1356 } else
1357 mtx_lock(&np->n_mtx);
1358 if (np->n_directio_asyncwr == 0)
1359 np->n_flag &= ~NMODIFIED;
1360 mtx_unlock(&np->n_mtx);
1361 out:
1362 ncl_downgrade_vnlock(vp, old_lock);
1363 return error;
1364 }
1365
1366 /*
1367 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1368 * This is mainly to avoid queueing async I/O requests when the nfsiods
1369 * are all hung on a dead server.
1370 *
1371 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1372 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1373 */
1374 int
ncl_asyncio(struct nfsmount * nmp,struct buf * bp,struct ucred * cred,struct thread * td)1375 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1376 {
1377 int iod;
1378 int gotiod;
1379 int slpflag = 0;
1380 int slptimeo = 0;
1381 int error, error2;
1382
1383 /*
1384 * Commits are usually short and sweet so lets save some cpu and
1385 * leave the async daemons for more important rpc's (such as reads
1386 * and writes).
1387 *
1388 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1389 * in the directory in order to update attributes. This can deadlock
1390 * with another thread that is waiting for async I/O to be done by
1391 * an nfsiod thread while holding a lock on one of these vnodes.
1392 * To avoid this deadlock, don't allow the async nfsiod threads to
1393 * perform Readdirplus RPCs.
1394 */
1395 mtx_lock(&ncl_iod_mutex);
1396 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1397 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1398 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1399 mtx_unlock(&ncl_iod_mutex);
1400 return(EIO);
1401 }
1402 again:
1403 if (nmp->nm_flag & NFSMNT_INT)
1404 slpflag = PCATCH;
1405 gotiod = FALSE;
1406
1407 /*
1408 * Find a free iod to process this request.
1409 */
1410 for (iod = 0; iod < ncl_numasync; iod++)
1411 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1412 gotiod = TRUE;
1413 break;
1414 }
1415
1416 /*
1417 * Try to create one if none are free.
1418 */
1419 if (!gotiod)
1420 ncl_nfsiodnew();
1421 else {
1422 /*
1423 * Found one, so wake it up and tell it which
1424 * mount to process.
1425 */
1426 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1427 iod, nmp));
1428 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1429 ncl_iodmount[iod] = nmp;
1430 nmp->nm_bufqiods++;
1431 wakeup(&ncl_iodwant[iod]);
1432 }
1433
1434 /*
1435 * If none are free, we may already have an iod working on this mount
1436 * point. If so, it will process our request.
1437 */
1438 if (!gotiod) {
1439 if (nmp->nm_bufqiods > 0) {
1440 NFS_DPF(ASYNCIO,
1441 ("ncl_asyncio: %d iods are already processing mount %p\n",
1442 nmp->nm_bufqiods, nmp));
1443 gotiod = TRUE;
1444 }
1445 }
1446
1447 /*
1448 * If we have an iod which can process the request, then queue
1449 * the buffer.
1450 */
1451 if (gotiod) {
1452 /*
1453 * Ensure that the queue never grows too large. We still want
1454 * to asynchronize so we block rather than return EIO.
1455 */
1456 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1457 NFS_DPF(ASYNCIO,
1458 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1459 nmp->nm_bufqwant = TRUE;
1460 error = newnfs_msleep(td, &nmp->nm_bufq,
1461 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1462 slptimeo);
1463 if (error) {
1464 error2 = newnfs_sigintr(nmp, td);
1465 if (error2) {
1466 mtx_unlock(&ncl_iod_mutex);
1467 return (error2);
1468 }
1469 if (slpflag == PCATCH) {
1470 slpflag = 0;
1471 slptimeo = 2 * hz;
1472 }
1473 }
1474 /*
1475 * We might have lost our iod while sleeping,
1476 * so check and loop if necessary.
1477 */
1478 goto again;
1479 }
1480
1481 /* We might have lost our nfsiod */
1482 if (nmp->nm_bufqiods == 0) {
1483 NFS_DPF(ASYNCIO,
1484 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1485 goto again;
1486 }
1487
1488 if (bp->b_iocmd == BIO_READ) {
1489 if (bp->b_rcred == NOCRED && cred != NOCRED)
1490 bp->b_rcred = crhold(cred);
1491 } else {
1492 if (bp->b_wcred == NOCRED && cred != NOCRED)
1493 bp->b_wcred = crhold(cred);
1494 }
1495
1496 if (bp->b_flags & B_REMFREE)
1497 bremfreef(bp);
1498 BUF_KERNPROC(bp);
1499 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1500 nmp->nm_bufqlen++;
1501 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1502 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1503 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1504 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1505 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1506 }
1507 mtx_unlock(&ncl_iod_mutex);
1508 return (0);
1509 }
1510
1511 mtx_unlock(&ncl_iod_mutex);
1512
1513 /*
1514 * All the iods are busy on other mounts, so return EIO to
1515 * force the caller to process the i/o synchronously.
1516 */
1517 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1518 return (EIO);
1519 }
1520
1521 void
ncl_doio_directwrite(struct buf * bp)1522 ncl_doio_directwrite(struct buf *bp)
1523 {
1524 int iomode, must_commit;
1525 struct uio *uiop = (struct uio *)bp->b_caller1;
1526 char *iov_base = uiop->uio_iov->iov_base;
1527
1528 iomode = NFSWRITE_FILESYNC;
1529 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1530 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1531 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1532 free(iov_base, M_NFSDIRECTIO);
1533 free(uiop->uio_iov, M_NFSDIRECTIO);
1534 free(uiop, M_NFSDIRECTIO);
1535 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1536 struct nfsnode *np = VTONFS(bp->b_vp);
1537 mtx_lock(&np->n_mtx);
1538 if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1539 /*
1540 * Invalidate the attribute cache, since writes to a DS
1541 * won't update the size attribute.
1542 */
1543 np->n_attrstamp = 0;
1544 }
1545 np->n_directio_asyncwr--;
1546 if (np->n_directio_asyncwr == 0) {
1547 np->n_flag &= ~NMODIFIED;
1548 if ((np->n_flag & NFSYNCWAIT)) {
1549 np->n_flag &= ~NFSYNCWAIT;
1550 wakeup((caddr_t)&np->n_directio_asyncwr);
1551 }
1552 }
1553 mtx_unlock(&np->n_mtx);
1554 }
1555 bp->b_vp = NULL;
1556 relpbuf(bp, &ncl_pbuf_freecnt);
1557 }
1558
1559 /*
1560 * Do an I/O operation to/from a cache block. This may be called
1561 * synchronously or from an nfsiod.
1562 */
1563 int
ncl_doio(struct vnode * vp,struct buf * bp,struct ucred * cr,struct thread * td,int called_from_strategy)1564 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1565 int called_from_strategy)
1566 {
1567 struct uio *uiop;
1568 struct nfsnode *np;
1569 struct nfsmount *nmp;
1570 int error = 0, iomode, must_commit = 0;
1571 struct uio uio;
1572 struct iovec io;
1573 struct proc *p = td ? td->td_proc : NULL;
1574 uint8_t iocmd;
1575
1576 np = VTONFS(vp);
1577 nmp = VFSTONFS(vp->v_mount);
1578 uiop = &uio;
1579 uiop->uio_iov = &io;
1580 uiop->uio_iovcnt = 1;
1581 uiop->uio_segflg = UIO_SYSSPACE;
1582 uiop->uio_td = td;
1583
1584 /*
1585 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1586 * do this here so we do not have to do it in all the code that
1587 * calls us.
1588 */
1589 bp->b_flags &= ~B_INVAL;
1590 bp->b_ioflags &= ~BIO_ERROR;
1591
1592 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1593 iocmd = bp->b_iocmd;
1594 if (iocmd == BIO_READ) {
1595 io.iov_len = uiop->uio_resid = bp->b_bcount;
1596 io.iov_base = bp->b_data;
1597 uiop->uio_rw = UIO_READ;
1598
1599 switch (vp->v_type) {
1600 case VREG:
1601 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1602 NFSINCRGLOBAL(nfsstatsv1.read_bios);
1603 error = ncl_readrpc(vp, uiop, cr);
1604
1605 if (!error) {
1606 if (uiop->uio_resid) {
1607 /*
1608 * If we had a short read with no error, we must have
1609 * hit a file hole. We should zero-fill the remainder.
1610 * This can also occur if the server hits the file EOF.
1611 *
1612 * Holes used to be able to occur due to pending
1613 * writes, but that is not possible any longer.
1614 */
1615 int nread = bp->b_bcount - uiop->uio_resid;
1616 ssize_t left = uiop->uio_resid;
1617
1618 if (left > 0)
1619 bzero((char *)bp->b_data + nread, left);
1620 uiop->uio_resid = 0;
1621 }
1622 }
1623 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1624 if (p && (vp->v_vflag & VV_TEXT)) {
1625 mtx_lock(&np->n_mtx);
1626 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1627 mtx_unlock(&np->n_mtx);
1628 PROC_LOCK(p);
1629 killproc(p, "text file modification");
1630 PROC_UNLOCK(p);
1631 } else
1632 mtx_unlock(&np->n_mtx);
1633 }
1634 break;
1635 case VLNK:
1636 uiop->uio_offset = (off_t)0;
1637 NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1638 error = ncl_readlinkrpc(vp, uiop, cr);
1639 break;
1640 case VDIR:
1641 NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1642 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1643 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1644 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1645 if (error == NFSERR_NOTSUPP)
1646 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1647 }
1648 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1649 error = ncl_readdirrpc(vp, uiop, cr, td);
1650 /*
1651 * end-of-directory sets B_INVAL but does not generate an
1652 * error.
1653 */
1654 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1655 bp->b_flags |= B_INVAL;
1656 break;
1657 default:
1658 printf("ncl_doio: type %x unexpected\n", vp->v_type);
1659 break;
1660 }
1661 if (error) {
1662 bp->b_ioflags |= BIO_ERROR;
1663 bp->b_error = error;
1664 }
1665 } else {
1666 /*
1667 * If we only need to commit, try to commit
1668 */
1669 if (bp->b_flags & B_NEEDCOMMIT) {
1670 int retv;
1671 off_t off;
1672
1673 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1674 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1675 bp->b_wcred, td);
1676 if (retv == 0) {
1677 bp->b_dirtyoff = bp->b_dirtyend = 0;
1678 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1679 bp->b_resid = 0;
1680 bufdone(bp);
1681 return (0);
1682 }
1683 if (retv == NFSERR_STALEWRITEVERF) {
1684 ncl_clearcommit(vp->v_mount);
1685 }
1686 }
1687
1688 /*
1689 * Setup for actual write
1690 */
1691 mtx_lock(&np->n_mtx);
1692 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1693 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1694 mtx_unlock(&np->n_mtx);
1695
1696 if (bp->b_dirtyend > bp->b_dirtyoff) {
1697 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1698 - bp->b_dirtyoff;
1699 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1700 + bp->b_dirtyoff;
1701 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1702 uiop->uio_rw = UIO_WRITE;
1703 NFSINCRGLOBAL(nfsstatsv1.write_bios);
1704
1705 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1706 iomode = NFSWRITE_UNSTABLE;
1707 else
1708 iomode = NFSWRITE_FILESYNC;
1709
1710 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1711 called_from_strategy);
1712
1713 /*
1714 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1715 * to cluster the buffers needing commit. This will allow
1716 * the system to submit a single commit rpc for the whole
1717 * cluster. We can do this even if the buffer is not 100%
1718 * dirty (relative to the NFS blocksize), so we optimize the
1719 * append-to-file-case.
1720 *
1721 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1722 * cleared because write clustering only works for commit
1723 * rpc's, not for the data portion of the write).
1724 */
1725
1726 if (!error && iomode == NFSWRITE_UNSTABLE) {
1727 bp->b_flags |= B_NEEDCOMMIT;
1728 if (bp->b_dirtyoff == 0
1729 && bp->b_dirtyend == bp->b_bcount)
1730 bp->b_flags |= B_CLUSTEROK;
1731 } else {
1732 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1733 }
1734
1735 /*
1736 * For an interrupted write, the buffer is still valid
1737 * and the write hasn't been pushed to the server yet,
1738 * so we can't set BIO_ERROR and report the interruption
1739 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1740 * is not relevant, so the rpc attempt is essentially
1741 * a noop. For the case of a V3 write rpc not being
1742 * committed to stable storage, the block is still
1743 * dirty and requires either a commit rpc or another
1744 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1745 * the block is reused. This is indicated by setting
1746 * the B_DELWRI and B_NEEDCOMMIT flags.
1747 *
1748 * EIO is returned by ncl_writerpc() to indicate a recoverable
1749 * write error and is handled as above, except that
1750 * B_EINTR isn't set. One cause of this is a stale stateid
1751 * error for the RPC that indicates recovery is required,
1752 * when called with called_from_strategy != 0.
1753 *
1754 * If the buffer is marked B_PAGING, it does not reside on
1755 * the vp's paging queues so we cannot call bdirty(). The
1756 * bp in this case is not an NFS cache block so we should
1757 * be safe. XXX
1758 *
1759 * The logic below breaks up errors into recoverable and
1760 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1761 * and keep the buffer around for potential write retries.
1762 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1763 * and save the error in the nfsnode. This is less than ideal
1764 * but necessary. Keeping such buffers around could potentially
1765 * cause buffer exhaustion eventually (they can never be written
1766 * out, so will get constantly be re-dirtied). It also causes
1767 * all sorts of vfs panics. For non-recoverable write errors,
1768 * also invalidate the attrcache, so we'll be forced to go over
1769 * the wire for this object, returning an error to user on next
1770 * call (most of the time).
1771 */
1772 if (error == EINTR || error == EIO || error == ETIMEDOUT
1773 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1774 int s;
1775
1776 s = splbio();
1777 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1778 if ((bp->b_flags & B_PAGING) == 0) {
1779 bdirty(bp);
1780 bp->b_flags &= ~B_DONE;
1781 }
1782 if ((error == EINTR || error == ETIMEDOUT) &&
1783 (bp->b_flags & B_ASYNC) == 0)
1784 bp->b_flags |= B_EINTR;
1785 splx(s);
1786 } else {
1787 if (error) {
1788 bp->b_ioflags |= BIO_ERROR;
1789 bp->b_flags |= B_INVAL;
1790 bp->b_error = np->n_error = error;
1791 mtx_lock(&np->n_mtx);
1792 np->n_flag |= NWRITEERR;
1793 np->n_attrstamp = 0;
1794 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1795 mtx_unlock(&np->n_mtx);
1796 }
1797 bp->b_dirtyoff = bp->b_dirtyend = 0;
1798 }
1799 } else {
1800 bp->b_resid = 0;
1801 bufdone(bp);
1802 return (0);
1803 }
1804 }
1805 bp->b_resid = uiop->uio_resid;
1806 if (must_commit)
1807 ncl_clearcommit(vp->v_mount);
1808 bufdone(bp);
1809 return (error);
1810 }
1811
1812 /*
1813 * Used to aid in handling ftruncate() operations on the NFS client side.
1814 * Truncation creates a number of special problems for NFS. We have to
1815 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1816 * we have to properly handle VM pages or (potentially dirty) buffers
1817 * that straddle the truncation point.
1818 */
1819
1820 int
ncl_meta_setsize(struct vnode * vp,struct ucred * cred,struct thread * td,u_quad_t nsize)1821 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1822 {
1823 struct nfsnode *np = VTONFS(vp);
1824 u_quad_t tsize;
1825 int biosize = vp->v_bufobj.bo_bsize;
1826 int error = 0;
1827
1828 mtx_lock(&np->n_mtx);
1829 tsize = np->n_size;
1830 np->n_size = nsize;
1831 mtx_unlock(&np->n_mtx);
1832
1833 if (nsize < tsize) {
1834 struct buf *bp;
1835 daddr_t lbn;
1836 int bufsize;
1837
1838 /*
1839 * vtruncbuf() doesn't get the buffer overlapping the
1840 * truncation point. We may have a B_DELWRI and/or B_CACHE
1841 * buffer that now needs to be truncated.
1842 */
1843 error = vtruncbuf(vp, cred, nsize, biosize);
1844 lbn = nsize / biosize;
1845 bufsize = nsize - (lbn * biosize);
1846 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1847 if (!bp)
1848 return EINTR;
1849 if (bp->b_dirtyoff > bp->b_bcount)
1850 bp->b_dirtyoff = bp->b_bcount;
1851 if (bp->b_dirtyend > bp->b_bcount)
1852 bp->b_dirtyend = bp->b_bcount;
1853 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1854 brelse(bp);
1855 } else {
1856 vnode_pager_setsize(vp, nsize);
1857 }
1858 return(error);
1859 }
1860
1861