1 /*
2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
37 */
38
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD: releng/5.1/sys/nfsclient/nfs_bio.c 115041 2003-05-15 21:12:08Z rwatson $");
41
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/bio.h>
45 #include <sys/buf.h>
46 #include <sys/kernel.h>
47 #include <sys/mount.h>
48 #include <sys/proc.h>
49 #include <sys/resourcevar.h>
50 #include <sys/signalvar.h>
51 #include <sys/vmmeter.h>
52 #include <sys/vnode.h>
53
54 #include <vm/vm.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_pager.h>
59 #include <vm/vnode_pager.h>
60
61 #include <nfs/rpcv2.h>
62 #include <nfs/nfsproto.h>
63 #include <nfsclient/nfs.h>
64 #include <nfsclient/nfsmount.h>
65 #include <nfsclient/nfsnode.h>
66
67 /*
68 * Just call nfs_writebp() with the force argument set to 1.
69 *
70 * NOTE: B_DONE may or may not be set in a_bp on call.
71 */
72 static int
73 nfs_bwrite(struct buf *bp)
74 {
75
76 return (nfs_writebp(bp, 1, curthread));
77 }
78
79 struct buf_ops buf_ops_nfs = {
80 "buf_ops_nfs",
81 nfs_bwrite
82 };
83
84 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
85 struct thread *td);
86
87 /*
88 * Vnode op for VM getpages.
89 */
90 int
91 nfs_getpages(struct vop_getpages_args *ap)
92 {
93 int i, error, nextoff, size, toff, count, npages;
94 struct uio uio;
95 struct iovec iov;
96 vm_offset_t kva;
97 struct buf *bp;
98 struct vnode *vp;
99 struct thread *td;
100 struct ucred *cred;
101 struct nfsmount *nmp;
102 vm_page_t *pages;
103
104 GIANT_REQUIRED;
105
106 vp = ap->a_vp;
107 td = curthread; /* XXX */
108 cred = curthread->td_ucred; /* XXX */
109 nmp = VFSTONFS(vp->v_mount);
110 pages = ap->a_m;
111 count = ap->a_count;
112
113 if (vp->v_object == NULL) {
114 printf("nfs_getpages: called with non-merged cache vnode??\n");
115 return VM_PAGER_ERROR;
116 }
117
118 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
119 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
120 (void)nfs_fsinfo(nmp, vp, cred, td);
121 }
122
123 npages = btoc(count);
124
125 /*
126 * If the requested page is partially valid, just return it and
127 * allow the pager to zero-out the blanks. Partially valid pages
128 * can only occur at the file EOF.
129 */
130
131 {
132 vm_page_t m = pages[ap->a_reqpage];
133
134 vm_page_lock_queues();
135 if (m->valid != 0) {
136 /* handled by vm_fault now */
137 /* vm_page_zero_invalid(m, TRUE); */
138 for (i = 0; i < npages; ++i) {
139 if (i != ap->a_reqpage)
140 vm_page_free(pages[i]);
141 }
142 vm_page_unlock_queues();
143 return(0);
144 }
145 vm_page_unlock_queues();
146 }
147
148 /*
149 * We use only the kva address for the buffer, but this is extremely
150 * convienient and fast.
151 */
152 bp = getpbuf(&nfs_pbuf_freecnt);
153
154 kva = (vm_offset_t) bp->b_data;
155 pmap_qenter(kva, pages, npages);
156 cnt.v_vnodein++;
157 cnt.v_vnodepgsin += npages;
158
159 iov.iov_base = (caddr_t) kva;
160 iov.iov_len = count;
161 uio.uio_iov = &iov;
162 uio.uio_iovcnt = 1;
163 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
164 uio.uio_resid = count;
165 uio.uio_segflg = UIO_SYSSPACE;
166 uio.uio_rw = UIO_READ;
167 uio.uio_td = td;
168
169 error = nfs_readrpc(vp, &uio, cred);
170 pmap_qremove(kva, npages);
171
172 relpbuf(bp, &nfs_pbuf_freecnt);
173
174 if (error && (uio.uio_resid == count)) {
175 printf("nfs_getpages: error %d\n", error);
176 vm_page_lock_queues();
177 for (i = 0; i < npages; ++i) {
178 if (i != ap->a_reqpage)
179 vm_page_free(pages[i]);
180 }
181 vm_page_unlock_queues();
182 return VM_PAGER_ERROR;
183 }
184
185 /*
186 * Calculate the number of bytes read and validate only that number
187 * of bytes. Note that due to pending writes, size may be 0. This
188 * does not mean that the remaining data is invalid!
189 */
190
191 size = count - uio.uio_resid;
192 vm_page_lock_queues();
193 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
194 vm_page_t m;
195 nextoff = toff + PAGE_SIZE;
196 m = pages[i];
197
198 m->flags &= ~PG_ZERO;
199
200 if (nextoff <= size) {
201 /*
202 * Read operation filled an entire page
203 */
204 m->valid = VM_PAGE_BITS_ALL;
205 vm_page_undirty(m);
206 } else if (size > toff) {
207 /*
208 * Read operation filled a partial page.
209 */
210 m->valid = 0;
211 vm_page_set_validclean(m, 0, size - toff);
212 /* handled by vm_fault now */
213 /* vm_page_zero_invalid(m, TRUE); */
214 } else {
215 /*
216 * Read operation was short. If no error occured
217 * we may have hit a zero-fill section. We simply
218 * leave valid set to 0.
219 */
220 ;
221 }
222 if (i != ap->a_reqpage) {
223 /*
224 * Whether or not to leave the page activated is up in
225 * the air, but we should put the page on a page queue
226 * somewhere (it already is in the object). Result:
227 * It appears that emperical results show that
228 * deactivating pages is best.
229 */
230
231 /*
232 * Just in case someone was asking for this page we
233 * now tell them that it is ok to use.
234 */
235 if (!error) {
236 if (m->flags & PG_WANTED)
237 vm_page_activate(m);
238 else
239 vm_page_deactivate(m);
240 vm_page_wakeup(m);
241 } else {
242 vm_page_free(m);
243 }
244 }
245 }
246 vm_page_unlock_queues();
247 return 0;
248 }
249
250 /*
251 * Vnode op for VM putpages.
252 */
253 int
254 nfs_putpages(struct vop_putpages_args *ap)
255 {
256 struct uio uio;
257 struct iovec iov;
258 vm_offset_t kva;
259 struct buf *bp;
260 int iomode, must_commit, i, error, npages, count;
261 off_t offset;
262 int *rtvals;
263 struct vnode *vp;
264 struct thread *td;
265 struct ucred *cred;
266 struct nfsmount *nmp;
267 struct nfsnode *np;
268 vm_page_t *pages;
269
270 GIANT_REQUIRED;
271
272 vp = ap->a_vp;
273 np = VTONFS(vp);
274 td = curthread; /* XXX */
275 cred = curthread->td_ucred; /* XXX */
276 nmp = VFSTONFS(vp->v_mount);
277 pages = ap->a_m;
278 count = ap->a_count;
279 rtvals = ap->a_rtvals;
280 npages = btoc(count);
281 offset = IDX_TO_OFF(pages[0]->pindex);
282
283 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
284 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
285 (void)nfs_fsinfo(nmp, vp, cred, td);
286 }
287
288 for (i = 0; i < npages; i++)
289 rtvals[i] = VM_PAGER_AGAIN;
290
291 /*
292 * When putting pages, do not extend file past EOF.
293 */
294
295 if (offset + count > np->n_size) {
296 count = np->n_size - offset;
297 if (count < 0)
298 count = 0;
299 }
300
301 /*
302 * We use only the kva address for the buffer, but this is extremely
303 * convienient and fast.
304 */
305 bp = getpbuf(&nfs_pbuf_freecnt);
306
307 kva = (vm_offset_t) bp->b_data;
308 pmap_qenter(kva, pages, npages);
309 cnt.v_vnodeout++;
310 cnt.v_vnodepgsout += count;
311
312 iov.iov_base = (caddr_t) kva;
313 iov.iov_len = count;
314 uio.uio_iov = &iov;
315 uio.uio_iovcnt = 1;
316 uio.uio_offset = offset;
317 uio.uio_resid = count;
318 uio.uio_segflg = UIO_SYSSPACE;
319 uio.uio_rw = UIO_WRITE;
320 uio.uio_td = td;
321
322 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
323 iomode = NFSV3WRITE_UNSTABLE;
324 else
325 iomode = NFSV3WRITE_FILESYNC;
326
327 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit);
328
329 pmap_qremove(kva, npages);
330 relpbuf(bp, &nfs_pbuf_freecnt);
331
332 if (!error) {
333 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
334 for (i = 0; i < nwritten; i++) {
335 rtvals[i] = VM_PAGER_OK;
336 vm_page_undirty(pages[i]);
337 }
338 if (must_commit) {
339 nfs_clearcommit(vp->v_mount);
340 }
341 }
342 return rtvals[0];
343 }
344
345 /*
346 * Vnode op for read using bio
347 */
348 int
349 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
350 {
351 struct nfsnode *np = VTONFS(vp);
352 int biosize, i;
353 struct buf *bp = 0, *rabp;
354 struct vattr vattr;
355 struct thread *td;
356 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
357 daddr_t lbn, rabn;
358 int bcount;
359 int seqcount;
360 int nra, error = 0, n = 0, on = 0;
361
362 #ifdef DIAGNOSTIC
363 if (uio->uio_rw != UIO_READ)
364 panic("nfs_read mode");
365 #endif
366 if (uio->uio_resid == 0)
367 return (0);
368 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
369 return (EINVAL);
370 td = uio->uio_td;
371
372 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
373 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
374 (void)nfs_fsinfo(nmp, vp, cred, td);
375 if (vp->v_type != VDIR &&
376 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
377 return (EFBIG);
378 biosize = vp->v_mount->mnt_stat.f_iosize;
379 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
380 /*
381 * For nfs, cache consistency can only be maintained approximately.
382 * Although RFC1094 does not specify the criteria, the following is
383 * believed to be compatible with the reference port.
384 * For nfs:
385 * If the file's modify time on the server has changed since the
386 * last read rpc or you have written to the file,
387 * you may have lost data cache consistency with the
388 * server, so flush all of the file's data out of the cache.
389 * Then force a getattr rpc to ensure that you have up to date
390 * attributes.
391 * NB: This implies that cache data can be read when up to
392 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
393 * attributes this could be forced by setting n_attrstamp to 0 before
394 * the VOP_GETATTR() call.
395 */
396 if (np->n_flag & NMODIFIED) {
397 if (vp->v_type != VREG) {
398 if (vp->v_type != VDIR)
399 panic("nfs: bioread, not dir");
400 nfs_invaldir(vp);
401 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
402 if (error)
403 return (error);
404 }
405 np->n_attrstamp = 0;
406 error = VOP_GETATTR(vp, &vattr, cred, td);
407 if (error)
408 return (error);
409 np->n_mtime = vattr.va_mtime.tv_sec;
410 } else {
411 error = VOP_GETATTR(vp, &vattr, cred, td);
412 if (error)
413 return (error);
414 if (np->n_mtime != vattr.va_mtime.tv_sec) {
415 if (vp->v_type == VDIR)
416 nfs_invaldir(vp);
417 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
418 if (error)
419 return (error);
420 np->n_mtime = vattr.va_mtime.tv_sec;
421 }
422 }
423 do {
424 switch (vp->v_type) {
425 case VREG:
426 nfsstats.biocache_reads++;
427 lbn = uio->uio_offset / biosize;
428 on = uio->uio_offset & (biosize - 1);
429
430 /*
431 * Start the read ahead(s), as required.
432 */
433 if (nmp->nm_readahead > 0) {
434 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
435 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
436 rabn = lbn + 1 + nra;
437 if (incore(vp, rabn) == NULL) {
438 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
439 if (!rabp)
440 return (EINTR);
441 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
442 rabp->b_flags |= B_ASYNC;
443 rabp->b_iocmd = BIO_READ;
444 vfs_busy_pages(rabp, 0);
445 if (nfs_asyncio(rabp, cred, td)) {
446 rabp->b_flags |= B_INVAL;
447 rabp->b_ioflags |= BIO_ERROR;
448 vfs_unbusy_pages(rabp);
449 brelse(rabp);
450 break;
451 }
452 } else {
453 brelse(rabp);
454 }
455 }
456 }
457 }
458
459 /*
460 * Obtain the buffer cache block. Figure out the buffer size
461 * when we are at EOF. If we are modifying the size of the
462 * buffer based on an EOF condition we need to hold
463 * nfs_rslock() through obtaining the buffer to prevent
464 * a potential writer-appender from messing with n_size.
465 * Otherwise we may accidently truncate the buffer and
466 * lose dirty data.
467 *
468 * Note that bcount is *not* DEV_BSIZE aligned.
469 */
470
471 again:
472 bcount = biosize;
473 if ((off_t)lbn * biosize >= np->n_size) {
474 bcount = 0;
475 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
476 bcount = np->n_size - (off_t)lbn * biosize;
477 }
478 if (bcount != biosize) {
479 switch(nfs_rslock(np, td)) {
480 case ENOLCK:
481 goto again;
482 /* not reached */
483 case EINTR:
484 case ERESTART:
485 return(EINTR);
486 /* not reached */
487 default:
488 break;
489 }
490 }
491
492 bp = nfs_getcacheblk(vp, lbn, bcount, td);
493
494 if (bcount != biosize)
495 nfs_rsunlock(np, td);
496 if (!bp)
497 return (EINTR);
498
499 /*
500 * If B_CACHE is not set, we must issue the read. If this
501 * fails, we return an error.
502 */
503
504 if ((bp->b_flags & B_CACHE) == 0) {
505 bp->b_iocmd = BIO_READ;
506 vfs_busy_pages(bp, 0);
507 error = nfs_doio(bp, cred, td);
508 if (error) {
509 brelse(bp);
510 return (error);
511 }
512 }
513
514 /*
515 * on is the offset into the current bp. Figure out how many
516 * bytes we can copy out of the bp. Note that bcount is
517 * NOT DEV_BSIZE aligned.
518 *
519 * Then figure out how many bytes we can copy into the uio.
520 */
521
522 n = 0;
523 if (on < bcount)
524 n = min((unsigned)(bcount - on), uio->uio_resid);
525 break;
526 case VLNK:
527 nfsstats.biocache_readlinks++;
528 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
529 if (!bp)
530 return (EINTR);
531 if ((bp->b_flags & B_CACHE) == 0) {
532 bp->b_iocmd = BIO_READ;
533 vfs_busy_pages(bp, 0);
534 error = nfs_doio(bp, cred, td);
535 if (error) {
536 bp->b_ioflags |= BIO_ERROR;
537 brelse(bp);
538 return (error);
539 }
540 }
541 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
542 on = 0;
543 break;
544 case VDIR:
545 nfsstats.biocache_readdirs++;
546 if (np->n_direofoffset
547 && uio->uio_offset >= np->n_direofoffset) {
548 return (0);
549 }
550 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
551 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
552 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
553 if (!bp)
554 return (EINTR);
555 if ((bp->b_flags & B_CACHE) == 0) {
556 bp->b_iocmd = BIO_READ;
557 vfs_busy_pages(bp, 0);
558 error = nfs_doio(bp, cred, td);
559 if (error) {
560 brelse(bp);
561 }
562 while (error == NFSERR_BAD_COOKIE) {
563 printf("got bad cookie vp %p bp %p\n", vp, bp);
564 nfs_invaldir(vp);
565 error = nfs_vinvalbuf(vp, 0, cred, td, 1);
566 /*
567 * Yuck! The directory has been modified on the
568 * server. The only way to get the block is by
569 * reading from the beginning to get all the
570 * offset cookies.
571 *
572 * Leave the last bp intact unless there is an error.
573 * Loop back up to the while if the error is another
574 * NFSERR_BAD_COOKIE (double yuch!).
575 */
576 for (i = 0; i <= lbn && !error; i++) {
577 if (np->n_direofoffset
578 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
579 return (0);
580 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
581 if (!bp)
582 return (EINTR);
583 if ((bp->b_flags & B_CACHE) == 0) {
584 bp->b_iocmd = BIO_READ;
585 vfs_busy_pages(bp, 0);
586 error = nfs_doio(bp, cred, td);
587 /*
588 * no error + B_INVAL == directory EOF,
589 * use the block.
590 */
591 if (error == 0 && (bp->b_flags & B_INVAL))
592 break;
593 }
594 /*
595 * An error will throw away the block and the
596 * for loop will break out. If no error and this
597 * is not the block we want, we throw away the
598 * block and go for the next one via the for loop.
599 */
600 if (error || i < lbn)
601 brelse(bp);
602 }
603 }
604 /*
605 * The above while is repeated if we hit another cookie
606 * error. If we hit an error and it wasn't a cookie error,
607 * we give up.
608 */
609 if (error)
610 return (error);
611 }
612
613 /*
614 * If not eof and read aheads are enabled, start one.
615 * (You need the current block first, so that you have the
616 * directory offset cookie of the next block.)
617 */
618 if (nmp->nm_readahead > 0 &&
619 (bp->b_flags & B_INVAL) == 0 &&
620 (np->n_direofoffset == 0 ||
621 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
622 incore(vp, lbn + 1) == NULL) {
623 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
624 if (rabp) {
625 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
626 rabp->b_flags |= B_ASYNC;
627 rabp->b_iocmd = BIO_READ;
628 vfs_busy_pages(rabp, 0);
629 if (nfs_asyncio(rabp, cred, td)) {
630 rabp->b_flags |= B_INVAL;
631 rabp->b_ioflags |= BIO_ERROR;
632 vfs_unbusy_pages(rabp);
633 brelse(rabp);
634 }
635 } else {
636 brelse(rabp);
637 }
638 }
639 }
640 /*
641 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
642 * chopped for the EOF condition, we cannot tell how large
643 * NFS directories are going to be until we hit EOF. So
644 * an NFS directory buffer is *not* chopped to its EOF. Now,
645 * it just so happens that b_resid will effectively chop it
646 * to EOF. *BUT* this information is lost if the buffer goes
647 * away and is reconstituted into a B_CACHE state ( due to
648 * being VMIO ) later. So we keep track of the directory eof
649 * in np->n_direofoffset and chop it off as an extra step
650 * right here.
651 */
652 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
653 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
654 n = np->n_direofoffset - uio->uio_offset;
655 break;
656 default:
657 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
658 break;
659 };
660
661 if (n > 0) {
662 error = uiomove(bp->b_data + on, (int)n, uio);
663 }
664 switch (vp->v_type) {
665 case VREG:
666 break;
667 case VLNK:
668 n = 0;
669 break;
670 case VDIR:
671 break;
672 default:
673 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
674 }
675 brelse(bp);
676 } while (error == 0 && uio->uio_resid > 0 && n > 0);
677 return (error);
678 }
679
680 /*
681 * Vnode op for write using bio
682 */
683 int
684 nfs_write(struct vop_write_args *ap)
685 {
686 int biosize;
687 struct uio *uio = ap->a_uio;
688 struct thread *td = uio->uio_td;
689 struct vnode *vp = ap->a_vp;
690 struct nfsnode *np = VTONFS(vp);
691 struct ucred *cred = ap->a_cred;
692 int ioflag = ap->a_ioflag;
693 struct buf *bp;
694 struct vattr vattr;
695 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
696 daddr_t lbn;
697 int bcount;
698 int n, on, error = 0;
699 int haverslock = 0;
700 struct proc *p = td?td->td_proc:NULL;
701
702 GIANT_REQUIRED;
703
704 #ifdef DIAGNOSTIC
705 if (uio->uio_rw != UIO_WRITE)
706 panic("nfs_write mode");
707 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
708 panic("nfs_write proc");
709 #endif
710 if (vp->v_type != VREG)
711 return (EIO);
712 if (np->n_flag & NWRITEERR) {
713 np->n_flag &= ~NWRITEERR;
714 return (np->n_error);
715 }
716 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
717 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
718 (void)nfs_fsinfo(nmp, vp, cred, td);
719
720 /*
721 * Synchronously flush pending buffers if we are in synchronous
722 * mode or if we are appending.
723 */
724 if (ioflag & (IO_APPEND | IO_SYNC)) {
725 if (np->n_flag & NMODIFIED) {
726 np->n_attrstamp = 0;
727 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
728 if (error)
729 return (error);
730 }
731 }
732
733 /*
734 * If IO_APPEND then load uio_offset. We restart here if we cannot
735 * get the append lock.
736 */
737 restart:
738 if (ioflag & IO_APPEND) {
739 np->n_attrstamp = 0;
740 error = VOP_GETATTR(vp, &vattr, cred, td);
741 if (error)
742 return (error);
743 uio->uio_offset = np->n_size;
744 }
745
746 if (uio->uio_offset < 0)
747 return (EINVAL);
748 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
749 return (EFBIG);
750 if (uio->uio_resid == 0)
751 return (0);
752
753 /*
754 * We need to obtain the rslock if we intend to modify np->n_size
755 * in order to guarentee the append point with multiple contending
756 * writers, to guarentee that no other appenders modify n_size
757 * while we are trying to obtain a truncated buffer (i.e. to avoid
758 * accidently truncating data written by another appender due to
759 * the race), and to ensure that the buffer is populated prior to
760 * our extending of the file. We hold rslock through the entire
761 * operation.
762 *
763 * Note that we do not synchronize the case where someone truncates
764 * the file while we are appending to it because attempting to lock
765 * this case may deadlock other parts of the system unexpectedly.
766 */
767 if ((ioflag & IO_APPEND) ||
768 uio->uio_offset + uio->uio_resid > np->n_size) {
769 switch(nfs_rslock(np, td)) {
770 case ENOLCK:
771 goto restart;
772 /* not reached */
773 case EINTR:
774 case ERESTART:
775 return(EINTR);
776 /* not reached */
777 default:
778 break;
779 }
780 haverslock = 1;
781 }
782
783 /*
784 * Maybe this should be above the vnode op call, but so long as
785 * file servers have no limits, i don't think it matters
786 */
787 if (p && uio->uio_offset + uio->uio_resid >
788 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
789 PROC_LOCK(p);
790 psignal(p, SIGXFSZ);
791 PROC_UNLOCK(p);
792 if (haverslock)
793 nfs_rsunlock(np, td);
794 return (EFBIG);
795 }
796
797 biosize = vp->v_mount->mnt_stat.f_iosize;
798
799 do {
800 nfsstats.biocache_writes++;
801 lbn = uio->uio_offset / biosize;
802 on = uio->uio_offset & (biosize-1);
803 n = min((unsigned)(biosize - on), uio->uio_resid);
804 again:
805 /*
806 * Handle direct append and file extension cases, calculate
807 * unaligned buffer size.
808 */
809
810 if (uio->uio_offset == np->n_size && n) {
811 /*
812 * Get the buffer (in its pre-append state to maintain
813 * B_CACHE if it was previously set). Resize the
814 * nfsnode after we have locked the buffer to prevent
815 * readers from reading garbage.
816 */
817 bcount = on;
818 bp = nfs_getcacheblk(vp, lbn, bcount, td);
819
820 if (bp != NULL) {
821 long save;
822
823 np->n_size = uio->uio_offset + n;
824 np->n_flag |= NMODIFIED;
825 vnode_pager_setsize(vp, np->n_size);
826
827 save = bp->b_flags & B_CACHE;
828 bcount += n;
829 allocbuf(bp, bcount);
830 bp->b_flags |= save;
831 bp->b_magic = B_MAGIC_NFS;
832 bp->b_op = &buf_ops_nfs;
833 }
834 } else {
835 /*
836 * Obtain the locked cache block first, and then
837 * adjust the file's size as appropriate.
838 */
839 bcount = on + n;
840 if ((off_t)lbn * biosize + bcount < np->n_size) {
841 if ((off_t)(lbn + 1) * biosize < np->n_size)
842 bcount = biosize;
843 else
844 bcount = np->n_size - (off_t)lbn * biosize;
845 }
846 bp = nfs_getcacheblk(vp, lbn, bcount, td);
847 if (uio->uio_offset + n > np->n_size) {
848 np->n_size = uio->uio_offset + n;
849 np->n_flag |= NMODIFIED;
850 vnode_pager_setsize(vp, np->n_size);
851 }
852 }
853
854 if (!bp) {
855 error = EINTR;
856 break;
857 }
858
859 /*
860 * Issue a READ if B_CACHE is not set. In special-append
861 * mode, B_CACHE is based on the buffer prior to the write
862 * op and is typically set, avoiding the read. If a read
863 * is required in special append mode, the server will
864 * probably send us a short-read since we extended the file
865 * on our end, resulting in b_resid == 0 and, thusly,
866 * B_CACHE getting set.
867 *
868 * We can also avoid issuing the read if the write covers
869 * the entire buffer. We have to make sure the buffer state
870 * is reasonable in this case since we will not be initiating
871 * I/O. See the comments in kern/vfs_bio.c's getblk() for
872 * more information.
873 *
874 * B_CACHE may also be set due to the buffer being cached
875 * normally.
876 */
877
878 if (on == 0 && n == bcount) {
879 bp->b_flags |= B_CACHE;
880 bp->b_flags &= ~B_INVAL;
881 bp->b_ioflags &= ~BIO_ERROR;
882 }
883
884 if ((bp->b_flags & B_CACHE) == 0) {
885 bp->b_iocmd = BIO_READ;
886 vfs_busy_pages(bp, 0);
887 error = nfs_doio(bp, cred, td);
888 if (error) {
889 brelse(bp);
890 break;
891 }
892 }
893 if (!bp) {
894 error = EINTR;
895 break;
896 }
897 if (bp->b_wcred == NOCRED)
898 bp->b_wcred = crhold(cred);
899 np->n_flag |= NMODIFIED;
900
901 /*
902 * If dirtyend exceeds file size, chop it down. This should
903 * not normally occur but there is an append race where it
904 * might occur XXX, so we log it.
905 *
906 * If the chopping creates a reverse-indexed or degenerate
907 * situation with dirtyoff/end, we 0 both of them.
908 */
909
910 if (bp->b_dirtyend > bcount) {
911 printf("NFS append race @%lx:%d\n",
912 (long)bp->b_blkno * DEV_BSIZE,
913 bp->b_dirtyend - bcount);
914 bp->b_dirtyend = bcount;
915 }
916
917 if (bp->b_dirtyoff >= bp->b_dirtyend)
918 bp->b_dirtyoff = bp->b_dirtyend = 0;
919
920 /*
921 * If the new write will leave a contiguous dirty
922 * area, just update the b_dirtyoff and b_dirtyend,
923 * otherwise force a write rpc of the old dirty area.
924 *
925 * While it is possible to merge discontiguous writes due to
926 * our having a B_CACHE buffer ( and thus valid read data
927 * for the hole), we don't because it could lead to
928 * significant cache coherency problems with multiple clients,
929 * especially if locking is implemented later on.
930 *
931 * as an optimization we could theoretically maintain
932 * a linked list of discontinuous areas, but we would still
933 * have to commit them separately so there isn't much
934 * advantage to it except perhaps a bit of asynchronization.
935 */
936
937 if (bp->b_dirtyend > 0 &&
938 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
939 if (BUF_WRITE(bp) == EINTR) {
940 error = EINTR;
941 break;
942 }
943 goto again;
944 }
945
946 error = uiomove((char *)bp->b_data + on, n, uio);
947
948 /*
949 * Since this block is being modified, it must be written
950 * again and not just committed. Since write clustering does
951 * not work for the stage 1 data write, only the stage 2
952 * commit rpc, we have to clear B_CLUSTEROK as well.
953 */
954 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
955
956 if (error) {
957 bp->b_ioflags |= BIO_ERROR;
958 brelse(bp);
959 break;
960 }
961
962 /*
963 * Only update dirtyoff/dirtyend if not a degenerate
964 * condition.
965 */
966 if (n) {
967 if (bp->b_dirtyend > 0) {
968 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
969 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
970 } else {
971 bp->b_dirtyoff = on;
972 bp->b_dirtyend = on + n;
973 }
974 vfs_bio_set_validclean(bp, on, n);
975 }
976 /*
977 * If IO_NOWDRAIN then set B_NOWDRAIN (nfs-backed MD
978 * filesystem)
979 */
980 if (ioflag & IO_NOWDRAIN)
981 bp->b_flags |= B_NOWDRAIN;
982
983 /*
984 * If IO_SYNC do bwrite().
985 *
986 * IO_INVAL appears to be unused. The idea appears to be
987 * to turn off caching in this case. Very odd. XXX
988 */
989 if ((ioflag & IO_SYNC)) {
990 if (ioflag & IO_INVAL)
991 bp->b_flags |= B_NOCACHE;
992 error = BUF_WRITE(bp);
993 if (error)
994 break;
995 } else if ((n + on) == biosize) {
996 bp->b_flags |= B_ASYNC;
997 (void)nfs_writebp(bp, 0, 0);
998 } else {
999 bdwrite(bp);
1000 }
1001 } while (uio->uio_resid > 0 && n > 0);
1002
1003 if (haverslock)
1004 nfs_rsunlock(np, td);
1005
1006 return (error);
1007 }
1008
1009 /*
1010 * Get an nfs cache block.
1011 *
1012 * Allocate a new one if the block isn't currently in the cache
1013 * and return the block marked busy. If the calling process is
1014 * interrupted by a signal for an interruptible mount point, return
1015 * NULL.
1016 *
1017 * The caller must carefully deal with the possible B_INVAL state of
1018 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1019 * indirectly), so synchronous reads can be issued without worrying about
1020 * the B_INVAL state. We have to be a little more careful when dealing
1021 * with writes (see comments in nfs_write()) when extending a file past
1022 * its EOF.
1023 */
1024 static struct buf *
1025 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1026 {
1027 struct buf *bp;
1028 struct mount *mp;
1029 struct nfsmount *nmp;
1030
1031 mp = vp->v_mount;
1032 nmp = VFSTONFS(mp);
1033
1034 if (nmp->nm_flag & NFSMNT_INT) {
1035 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1036 while (bp == NULL) {
1037 if (nfs_sigintr(nmp, NULL, td))
1038 return (NULL);
1039 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1040 }
1041 } else {
1042 bp = getblk(vp, bn, size, 0, 0, 0);
1043 }
1044
1045 if (vp->v_type == VREG) {
1046 int biosize;
1047
1048 biosize = mp->mnt_stat.f_iosize;
1049 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1050 }
1051 return (bp);
1052 }
1053
1054 /*
1055 * Flush and invalidate all dirty buffers. If another process is already
1056 * doing the flush, just wait for completion.
1057 */
1058 int
1059 nfs_vinvalbuf(struct vnode *vp, int flags, struct ucred *cred,
1060 struct thread *td, int intrflg)
1061 {
1062 struct nfsnode *np = VTONFS(vp);
1063 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1064 int error = 0, slpflag, slptimeo;
1065
1066 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1067
1068 VI_LOCK(vp);
1069 if (vp->v_iflag & VI_XLOCK) {
1070 /* XXX Should we wait here? */
1071 VI_UNLOCK(vp);
1072 return (0);
1073 }
1074 VI_UNLOCK(vp);
1075
1076 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1077 intrflg = 0;
1078 if (intrflg) {
1079 slpflag = PCATCH;
1080 slptimeo = 2 * hz;
1081 } else {
1082 slpflag = 0;
1083 slptimeo = 0;
1084 }
1085 /*
1086 * First wait for any other process doing a flush to complete.
1087 */
1088 while (np->n_flag & NFLUSHINPROG) {
1089 np->n_flag |= NFLUSHWANT;
1090 error = tsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
1091 slptimeo);
1092 if (error && intrflg &&
1093 nfs_sigintr(nmp, NULL, td))
1094 return (EINTR);
1095 }
1096
1097 /*
1098 * Now, flush as required.
1099 */
1100 np->n_flag |= NFLUSHINPROG;
1101 error = vinvalbuf(vp, flags, cred, td, slpflag, 0);
1102 while (error) {
1103 if (intrflg &&
1104 nfs_sigintr(nmp, NULL, td)) {
1105 np->n_flag &= ~NFLUSHINPROG;
1106 if (np->n_flag & NFLUSHWANT) {
1107 np->n_flag &= ~NFLUSHWANT;
1108 wakeup(&np->n_flag);
1109 }
1110 return (EINTR);
1111 }
1112 error = vinvalbuf(vp, flags, cred, td, 0, slptimeo);
1113 }
1114 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1115 if (np->n_flag & NFLUSHWANT) {
1116 np->n_flag &= ~NFLUSHWANT;
1117 wakeup(&np->n_flag);
1118 }
1119 return (0);
1120 }
1121
1122 /*
1123 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1124 * This is mainly to avoid queueing async I/O requests when the nfsiods
1125 * are all hung on a dead server.
1126 *
1127 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1128 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1129 */
1130 int
1131 nfs_asyncio(struct buf *bp, struct ucred *cred, struct thread *td)
1132 {
1133 struct nfsmount *nmp;
1134 int iod;
1135 int gotiod;
1136 int slpflag = 0;
1137 int slptimeo = 0;
1138 int error;
1139
1140 nmp = VFSTONFS(bp->b_vp->v_mount);
1141
1142 /*
1143 * Commits are usually short and sweet so lets save some cpu and
1144 * leave the async daemons for more important rpc's (such as reads
1145 * and writes).
1146 */
1147 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1148 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1149 return(EIO);
1150 }
1151
1152 again:
1153 if (nmp->nm_flag & NFSMNT_INT)
1154 slpflag = PCATCH;
1155 gotiod = FALSE;
1156
1157 /*
1158 * Find a free iod to process this request.
1159 */
1160 for (iod = 0; iod < nfs_numasync; iod++)
1161 if (nfs_iodwant[iod]) {
1162 gotiod = TRUE;
1163 break;
1164 }
1165
1166 /*
1167 * Try to create one if none are free.
1168 */
1169 if (!gotiod) {
1170 iod = nfs_nfsiodnew();
1171 if (iod != -1)
1172 gotiod = TRUE;
1173 }
1174
1175 if (gotiod) {
1176 /*
1177 * Found one, so wake it up and tell it which
1178 * mount to process.
1179 */
1180 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1181 iod, nmp));
1182 nfs_iodwant[iod] = NULL;
1183 nfs_iodmount[iod] = nmp;
1184 nmp->nm_bufqiods++;
1185 wakeup(&nfs_iodwant[iod]);
1186 }
1187
1188 /*
1189 * If none are free, we may already have an iod working on this mount
1190 * point. If so, it will process our request.
1191 */
1192 if (!gotiod) {
1193 if (nmp->nm_bufqiods > 0) {
1194 NFS_DPF(ASYNCIO,
1195 ("nfs_asyncio: %d iods are already processing mount %p\n",
1196 nmp->nm_bufqiods, nmp));
1197 gotiod = TRUE;
1198 }
1199 }
1200
1201 /*
1202 * If we have an iod which can process the request, then queue
1203 * the buffer.
1204 */
1205 if (gotiod) {
1206 /*
1207 * Ensure that the queue never grows too large. We still want
1208 * to asynchronize so we block rather then return EIO.
1209 */
1210 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1211 NFS_DPF(ASYNCIO,
1212 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1213 nmp->nm_bufqwant = TRUE;
1214 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
1215 "nfsaio", slptimeo);
1216 if (error) {
1217 if (nfs_sigintr(nmp, NULL, td))
1218 return (EINTR);
1219 if (slpflag == PCATCH) {
1220 slpflag = 0;
1221 slptimeo = 2 * hz;
1222 }
1223 }
1224 /*
1225 * We might have lost our iod while sleeping,
1226 * so check and loop if nescessary.
1227 */
1228 if (nmp->nm_bufqiods == 0) {
1229 NFS_DPF(ASYNCIO,
1230 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1231 goto again;
1232 }
1233 }
1234
1235 if (bp->b_iocmd == BIO_READ) {
1236 if (bp->b_rcred == NOCRED && cred != NOCRED)
1237 bp->b_rcred = crhold(cred);
1238 } else {
1239 bp->b_flags |= B_WRITEINPROG;
1240 if (bp->b_wcred == NOCRED && cred != NOCRED)
1241 bp->b_wcred = crhold(cred);
1242 }
1243
1244 BUF_KERNPROC(bp);
1245 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1246 nmp->nm_bufqlen++;
1247 return (0);
1248 }
1249
1250 /*
1251 * All the iods are busy on other mounts, so return EIO to
1252 * force the caller to process the i/o synchronously.
1253 */
1254 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1255 return (EIO);
1256 }
1257
1258 /*
1259 * Do an I/O operation to/from a cache block. This may be called
1260 * synchronously or from an nfsiod.
1261 */
1262 int
1263 nfs_doio(struct buf *bp, struct ucred *cr, struct thread *td)
1264 {
1265 struct uio *uiop;
1266 struct vnode *vp;
1267 struct nfsnode *np;
1268 struct nfsmount *nmp;
1269 int error = 0, iomode, must_commit = 0;
1270 struct uio uio;
1271 struct iovec io;
1272 struct proc *p = td ? td->td_proc : NULL;
1273
1274 vp = bp->b_vp;
1275 np = VTONFS(vp);
1276 nmp = VFSTONFS(vp->v_mount);
1277 uiop = &uio;
1278 uiop->uio_iov = &io;
1279 uiop->uio_iovcnt = 1;
1280 uiop->uio_segflg = UIO_SYSSPACE;
1281 uiop->uio_td = td;
1282
1283 /*
1284 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1285 * do this here so we do not have to do it in all the code that
1286 * calls us.
1287 */
1288 bp->b_flags &= ~B_INVAL;
1289 bp->b_ioflags &= ~BIO_ERROR;
1290
1291 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1292
1293 /*
1294 * Historically, paging was done with physio, but no more.
1295 */
1296 if (bp->b_flags & B_PHYS) {
1297 /*
1298 * ...though reading /dev/drum still gets us here.
1299 */
1300 io.iov_len = uiop->uio_resid = bp->b_bcount;
1301 /* mapping was done by vmapbuf() */
1302 io.iov_base = bp->b_data;
1303 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1304 if (bp->b_iocmd == BIO_READ) {
1305 uiop->uio_rw = UIO_READ;
1306 nfsstats.read_physios++;
1307 error = nfs_readrpc(vp, uiop, cr);
1308 } else {
1309 int com;
1310
1311 iomode = NFSV3WRITE_DATASYNC;
1312 uiop->uio_rw = UIO_WRITE;
1313 nfsstats.write_physios++;
1314 error = nfs_writerpc(vp, uiop, cr, &iomode, &com);
1315 }
1316 if (error) {
1317 bp->b_ioflags |= BIO_ERROR;
1318 bp->b_error = error;
1319 }
1320 } else if (bp->b_iocmd == BIO_READ) {
1321 io.iov_len = uiop->uio_resid = bp->b_bcount;
1322 io.iov_base = bp->b_data;
1323 uiop->uio_rw = UIO_READ;
1324
1325 switch (vp->v_type) {
1326 case VREG:
1327 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1328 nfsstats.read_bios++;
1329 error = nfs_readrpc(vp, uiop, cr);
1330
1331 if (!error) {
1332 if (uiop->uio_resid) {
1333 /*
1334 * If we had a short read with no error, we must have
1335 * hit a file hole. We should zero-fill the remainder.
1336 * This can also occur if the server hits the file EOF.
1337 *
1338 * Holes used to be able to occur due to pending
1339 * writes, but that is not possible any longer.
1340 */
1341 int nread = bp->b_bcount - uiop->uio_resid;
1342 int left = uiop->uio_resid;
1343
1344 if (left > 0)
1345 bzero((char *)bp->b_data + nread, left);
1346 uiop->uio_resid = 0;
1347 }
1348 }
1349 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1350 if (p && (vp->v_vflag & VV_TEXT) &&
1351 (np->n_mtime != np->n_vattr.va_mtime.tv_sec)) {
1352 uprintf("Process killed due to text file modification\n");
1353 PROC_LOCK(p);
1354 psignal(p, SIGKILL);
1355 _PHOLD(p);
1356 PROC_UNLOCK(p);
1357 }
1358 break;
1359 case VLNK:
1360 uiop->uio_offset = (off_t)0;
1361 nfsstats.readlink_bios++;
1362 error = nfs_readlinkrpc(vp, uiop, cr);
1363 break;
1364 case VDIR:
1365 nfsstats.readdir_bios++;
1366 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1367 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1368 error = nfs_readdirplusrpc(vp, uiop, cr);
1369 if (error == NFSERR_NOTSUPP)
1370 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1371 }
1372 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1373 error = nfs_readdirrpc(vp, uiop, cr);
1374 /*
1375 * end-of-directory sets B_INVAL but does not generate an
1376 * error.
1377 */
1378 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1379 bp->b_flags |= B_INVAL;
1380 break;
1381 default:
1382 printf("nfs_doio: type %x unexpected\n", vp->v_type);
1383 break;
1384 };
1385 if (error) {
1386 bp->b_ioflags |= BIO_ERROR;
1387 bp->b_error = error;
1388 }
1389 } else {
1390 /*
1391 * If we only need to commit, try to commit
1392 */
1393 if (bp->b_flags & B_NEEDCOMMIT) {
1394 int retv;
1395 off_t off;
1396
1397 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1398 bp->b_flags |= B_WRITEINPROG;
1399 retv = nfs_commit(
1400 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1401 bp->b_wcred, td);
1402 bp->b_flags &= ~B_WRITEINPROG;
1403 if (retv == 0) {
1404 bp->b_dirtyoff = bp->b_dirtyend = 0;
1405 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1406 bp->b_resid = 0;
1407 bufdone(bp);
1408 return (0);
1409 }
1410 if (retv == NFSERR_STALEWRITEVERF) {
1411 nfs_clearcommit(bp->b_vp->v_mount);
1412 }
1413 }
1414
1415 /*
1416 * Setup for actual write
1417 */
1418
1419 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1420 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1421
1422 if (bp->b_dirtyend > bp->b_dirtyoff) {
1423 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1424 - bp->b_dirtyoff;
1425 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1426 + bp->b_dirtyoff;
1427 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1428 uiop->uio_rw = UIO_WRITE;
1429 nfsstats.write_bios++;
1430
1431 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1432 iomode = NFSV3WRITE_UNSTABLE;
1433 else
1434 iomode = NFSV3WRITE_FILESYNC;
1435
1436 bp->b_flags |= B_WRITEINPROG;
1437 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1438
1439 /*
1440 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1441 * to cluster the buffers needing commit. This will allow
1442 * the system to submit a single commit rpc for the whole
1443 * cluster. We can do this even if the buffer is not 100%
1444 * dirty (relative to the NFS blocksize), so we optimize the
1445 * append-to-file-case.
1446 *
1447 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1448 * cleared because write clustering only works for commit
1449 * rpc's, not for the data portion of the write).
1450 */
1451
1452 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1453 bp->b_flags |= B_NEEDCOMMIT;
1454 if (bp->b_dirtyoff == 0
1455 && bp->b_dirtyend == bp->b_bcount)
1456 bp->b_flags |= B_CLUSTEROK;
1457 } else {
1458 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1459 }
1460 bp->b_flags &= ~B_WRITEINPROG;
1461
1462 /*
1463 * For an interrupted write, the buffer is still valid
1464 * and the write hasn't been pushed to the server yet,
1465 * so we can't set BIO_ERROR and report the interruption
1466 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1467 * is not relevant, so the rpc attempt is essentially
1468 * a noop. For the case of a V3 write rpc not being
1469 * committed to stable storage, the block is still
1470 * dirty and requires either a commit rpc or another
1471 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1472 * the block is reused. This is indicated by setting
1473 * the B_DELWRI and B_NEEDCOMMIT flags.
1474 *
1475 * If the buffer is marked B_PAGING, it does not reside on
1476 * the vp's paging queues so we cannot call bdirty(). The
1477 * bp in this case is not an NFS cache block so we should
1478 * be safe. XXX
1479 */
1480 if (error == EINTR
1481 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1482 int s;
1483
1484 s = splbio();
1485 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1486 if ((bp->b_flags & B_PAGING) == 0) {
1487 bdirty(bp);
1488 bp->b_flags &= ~B_DONE;
1489 }
1490 if (error && (bp->b_flags & B_ASYNC) == 0)
1491 bp->b_flags |= B_EINTR;
1492 splx(s);
1493 } else {
1494 if (error) {
1495 bp->b_ioflags |= BIO_ERROR;
1496 bp->b_error = np->n_error = error;
1497 np->n_flag |= NWRITEERR;
1498 }
1499 bp->b_dirtyoff = bp->b_dirtyend = 0;
1500 }
1501 } else {
1502 bp->b_resid = 0;
1503 bufdone(bp);
1504 return (0);
1505 }
1506 }
1507 bp->b_resid = uiop->uio_resid;
1508 if (must_commit)
1509 nfs_clearcommit(vp->v_mount);
1510 bufdone(bp);
1511 return (error);
1512 }
1513
1514 /*
1515 * Used to aid in handling ftruncate() operations on the NFS client side.
1516 * Truncation creates a number of special problems for NFS. We have to
1517 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1518 * we have to properly handle VM pages or (potentially dirty) buffers
1519 * that straddle the truncation point.
1520 */
1521
1522 int
1523 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1524 {
1525 struct nfsnode *np = VTONFS(vp);
1526 u_quad_t tsize = np->n_size;
1527 int biosize = vp->v_mount->mnt_stat.f_iosize;
1528 int error = 0;
1529
1530 np->n_size = nsize;
1531
1532 if (np->n_size < tsize) {
1533 struct buf *bp;
1534 daddr_t lbn;
1535 int bufsize;
1536
1537 /*
1538 * vtruncbuf() doesn't get the buffer overlapping the
1539 * truncation point. We may have a B_DELWRI and/or B_CACHE
1540 * buffer that now needs to be truncated.
1541 */
1542 error = vtruncbuf(vp, cred, td, nsize, biosize);
1543 lbn = nsize / biosize;
1544 bufsize = nsize & (biosize - 1);
1545 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1546 if (bp->b_dirtyoff > bp->b_bcount)
1547 bp->b_dirtyoff = bp->b_bcount;
1548 if (bp->b_dirtyend > bp->b_bcount)
1549 bp->b_dirtyend = bp->b_bcount;
1550 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1551 brelse(bp);
1552 } else {
1553 vnode_pager_setsize(vp, nsize);
1554 }
1555 return(error);
1556 }
1557
Cache object: cac6ed202d97cd10f5affd7fb1fe1fbe
|