Initial import from FreeBSD RELENG_4:
[dragonfly.git] / sys / vfs / nfs / nfs_bio.c
CommitLineData
984263bc
MD
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 * $FreeBSD: src/sys/nfs/nfs_bio.c,v 1.83.2.4 2002/12/29 18:19:53 dillon Exp $
38 */
39
40
41#include <sys/param.h>
42#include <sys/systm.h>
43#include <sys/resourcevar.h>
44#include <sys/signalvar.h>
45#include <sys/proc.h>
46#include <sys/buf.h>
47#include <sys/vnode.h>
48#include <sys/mount.h>
49#include <sys/kernel.h>
50
51#include <vm/vm.h>
52#include <vm/vm_extern.h>
53#include <vm/vm_page.h>
54#include <vm/vm_object.h>
55#include <vm/vm_pager.h>
56#include <vm/vnode_pager.h>
57
58#include <nfs/rpcv2.h>
59#include <nfs/nfsproto.h>
60#include <nfs/nfs.h>
61#include <nfs/nfsmount.h>
62#include <nfs/nqnfs.h>
63#include <nfs/nfsnode.h>
64
65static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size,
66 struct proc *p));
67
68extern int nfs_numasync;
69extern int nfs_pbuf_freecnt;
70extern struct nfsstats nfsstats;
71
72/*
73 * Vnode op for VM getpages.
74 */
75int
76nfs_getpages(ap)
77 struct vop_getpages_args /* {
78 struct vnode *a_vp;
79 vm_page_t *a_m;
80 int a_count;
81 int a_reqpage;
82 vm_ooffset_t a_offset;
83 } */ *ap;
84{
85 int i, error, nextoff, size, toff, count, npages;
86 struct uio uio;
87 struct iovec iov;
88 vm_offset_t kva;
89 struct buf *bp;
90 struct vnode *vp;
91 struct proc *p;
92 struct ucred *cred;
93 struct nfsmount *nmp;
94 vm_page_t *pages;
95
96 vp = ap->a_vp;
97 p = curproc; /* XXX */
98 cred = curproc->p_ucred; /* XXX */
99 nmp = VFSTONFS(vp->v_mount);
100 pages = ap->a_m;
101 count = ap->a_count;
102
103 if (vp->v_object == NULL) {
104 printf("nfs_getpages: called with non-merged cache vnode??\n");
105 return VM_PAGER_ERROR;
106 }
107
108 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
109 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
110 (void)nfs_fsinfo(nmp, vp, cred, p);
111
112 npages = btoc(count);
113
114 /*
115 * If the requested page is partially valid, just return it and
116 * allow the pager to zero-out the blanks. Partially valid pages
117 * can only occur at the file EOF.
118 */
119
120 {
121 vm_page_t m = pages[ap->a_reqpage];
122
123 if (m->valid != 0) {
124 /* handled by vm_fault now */
125 /* vm_page_zero_invalid(m, TRUE); */
126 for (i = 0; i < npages; ++i) {
127 if (i != ap->a_reqpage)
128 vnode_pager_freepage(pages[i]);
129 }
130 return(0);
131 }
132 }
133
134 /*
135 * We use only the kva address for the buffer, but this is extremely
136 * convienient and fast.
137 */
138 bp = getpbuf(&nfs_pbuf_freecnt);
139
140 kva = (vm_offset_t) bp->b_data;
141 pmap_qenter(kva, pages, npages);
142
143 iov.iov_base = (caddr_t) kva;
144 iov.iov_len = count;
145 uio.uio_iov = &iov;
146 uio.uio_iovcnt = 1;
147 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
148 uio.uio_resid = count;
149 uio.uio_segflg = UIO_SYSSPACE;
150 uio.uio_rw = UIO_READ;
151 uio.uio_procp = p;
152
153 error = nfs_readrpc(vp, &uio, cred);
154 pmap_qremove(kva, npages);
155
156 relpbuf(bp, &nfs_pbuf_freecnt);
157
158 if (error && (uio.uio_resid == count)) {
159 printf("nfs_getpages: error %d\n", error);
160 for (i = 0; i < npages; ++i) {
161 if (i != ap->a_reqpage)
162 vnode_pager_freepage(pages[i]);
163 }
164 return VM_PAGER_ERROR;
165 }
166
167 /*
168 * Calculate the number of bytes read and validate only that number
169 * of bytes. Note that due to pending writes, size may be 0. This
170 * does not mean that the remaining data is invalid!
171 */
172
173 size = count - uio.uio_resid;
174
175 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
176 vm_page_t m;
177 nextoff = toff + PAGE_SIZE;
178 m = pages[i];
179
180 m->flags &= ~PG_ZERO;
181
182 if (nextoff <= size) {
183 /*
184 * Read operation filled an entire page
185 */
186 m->valid = VM_PAGE_BITS_ALL;
187 vm_page_undirty(m);
188 } else if (size > toff) {
189 /*
190 * Read operation filled a partial page.
191 */
192 m->valid = 0;
193 vm_page_set_validclean(m, 0, size - toff);
194 /* handled by vm_fault now */
195 /* vm_page_zero_invalid(m, TRUE); */
196 } else {
197 /*
198 * Read operation was short. If no error occured
199 * we may have hit a zero-fill section. We simply
200 * leave valid set to 0.
201 */
202 ;
203 }
204 if (i != ap->a_reqpage) {
205 /*
206 * Whether or not to leave the page activated is up in
207 * the air, but we should put the page on a page queue
208 * somewhere (it already is in the object). Result:
209 * It appears that emperical results show that
210 * deactivating pages is best.
211 */
212
213 /*
214 * Just in case someone was asking for this page we
215 * now tell them that it is ok to use.
216 */
217 if (!error) {
218 if (m->flags & PG_WANTED)
219 vm_page_activate(m);
220 else
221 vm_page_deactivate(m);
222 vm_page_wakeup(m);
223 } else {
224 vnode_pager_freepage(m);
225 }
226 }
227 }
228 return 0;
229}
230
231/*
232 * Vnode op for VM putpages.
233 */
234int
235nfs_putpages(ap)
236 struct vop_putpages_args /* {
237 struct vnode *a_vp;
238 vm_page_t *a_m;
239 int a_count;
240 int a_sync;
241 int *a_rtvals;
242 vm_ooffset_t a_offset;
243 } */ *ap;
244{
245 struct uio uio;
246 struct iovec iov;
247 vm_offset_t kva;
248 struct buf *bp;
249 int iomode, must_commit, i, error, npages, count;
250 off_t offset;
251 int *rtvals;
252 struct vnode *vp;
253 struct proc *p;
254 struct ucred *cred;
255 struct nfsmount *nmp;
256 struct nfsnode *np;
257 vm_page_t *pages;
258
259 vp = ap->a_vp;
260 np = VTONFS(vp);
261 p = curproc; /* XXX */
262 cred = curproc->p_ucred; /* XXX */
263 nmp = VFSTONFS(vp->v_mount);
264 pages = ap->a_m;
265 count = ap->a_count;
266 rtvals = ap->a_rtvals;
267 npages = btoc(count);
268 offset = IDX_TO_OFF(pages[0]->pindex);
269
270 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
271 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
272 (void)nfs_fsinfo(nmp, vp, cred, p);
273
274 for (i = 0; i < npages; i++) {
275 rtvals[i] = VM_PAGER_AGAIN;
276 }
277
278 /*
279 * When putting pages, do not extend file past EOF.
280 */
281
282 if (offset + count > np->n_size) {
283 count = np->n_size - offset;
284 if (count < 0)
285 count = 0;
286 }
287
288 /*
289 * We use only the kva address for the buffer, but this is extremely
290 * convienient and fast.
291 */
292 bp = getpbuf(&nfs_pbuf_freecnt);
293
294 kva = (vm_offset_t) bp->b_data;
295 pmap_qenter(kva, pages, npages);
296
297 iov.iov_base = (caddr_t) kva;
298 iov.iov_len = count;
299 uio.uio_iov = &iov;
300 uio.uio_iovcnt = 1;
301 uio.uio_offset = offset;
302 uio.uio_resid = count;
303 uio.uio_segflg = UIO_SYSSPACE;
304 uio.uio_rw = UIO_WRITE;
305 uio.uio_procp = p;
306
307 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
308 iomode = NFSV3WRITE_UNSTABLE;
309 else
310 iomode = NFSV3WRITE_FILESYNC;
311
312 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit);
313
314 pmap_qremove(kva, npages);
315 relpbuf(bp, &nfs_pbuf_freecnt);
316
317 if (!error) {
318 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
319 for (i = 0; i < nwritten; i++) {
320 rtvals[i] = VM_PAGER_OK;
321 vm_page_undirty(pages[i]);
322 }
323 if (must_commit)
324 nfs_clearcommit(vp->v_mount);
325 }
326 return rtvals[0];
327}
328
329/*
330 * Vnode op for read using bio
331 */
332int
333nfs_bioread(vp, uio, ioflag, cred)
334 register struct vnode *vp;
335 register struct uio *uio;
336 int ioflag;
337 struct ucred *cred;
338{
339 register struct nfsnode *np = VTONFS(vp);
340 register int biosize, i;
341 struct buf *bp = 0, *rabp;
342 struct vattr vattr;
343 struct proc *p;
344 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
345 daddr_t lbn, rabn;
346 int bcount;
347 int seqcount;
348 int nra, error = 0, n = 0, on = 0;
349
350#ifdef DIAGNOSTIC
351 if (uio->uio_rw != UIO_READ)
352 panic("nfs_read mode");
353#endif
354 if (uio->uio_resid == 0)
355 return (0);
356 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
357 return (EINVAL);
358 p = uio->uio_procp;
359
360 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
361 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
362 (void)nfs_fsinfo(nmp, vp, cred, p);
363 if (vp->v_type != VDIR &&
364 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
365 return (EFBIG);
366 biosize = vp->v_mount->mnt_stat.f_iosize;
367 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
368 /*
369 * For nfs, cache consistency can only be maintained approximately.
370 * Although RFC1094 does not specify the criteria, the following is
371 * believed to be compatible with the reference port.
372 * For nqnfs, full cache consistency is maintained within the loop.
373 * For nfs:
374 * If the file's modify time on the server has changed since the
375 * last read rpc or you have written to the file,
376 * you may have lost data cache consistency with the
377 * server, so flush all of the file's data out of the cache.
378 * Then force a getattr rpc to ensure that you have up to date
379 * attributes.
380 * NB: This implies that cache data can be read when up to
381 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
382 * attributes this could be forced by setting n_attrstamp to 0 before
383 * the VOP_GETATTR() call.
384 */
385 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
386 if (np->n_flag & NMODIFIED) {
387 if (vp->v_type != VREG) {
388 if (vp->v_type != VDIR)
389 panic("nfs: bioread, not dir");
390 nfs_invaldir(vp);
391 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
392 if (error)
393 return (error);
394 }
395 np->n_attrstamp = 0;
396 error = VOP_GETATTR(vp, &vattr, cred, p);
397 if (error)
398 return (error);
399 np->n_mtime = vattr.va_mtime.tv_sec;
400 } else {
401 error = VOP_GETATTR(vp, &vattr, cred, p);
402 if (error)
403 return (error);
404 if (np->n_mtime != vattr.va_mtime.tv_sec) {
405 if (vp->v_type == VDIR)
406 nfs_invaldir(vp);
407 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
408 if (error)
409 return (error);
410 np->n_mtime = vattr.va_mtime.tv_sec;
411 }
412 }
413 }
414 do {
415
416 /*
417 * Get a valid lease. If cached data is stale, flush it.
418 */
419 if (nmp->nm_flag & NFSMNT_NQNFS) {
420 if (NQNFS_CKINVALID(vp, np, ND_READ)) {
421 do {
422 error = nqnfs_getlease(vp, ND_READ, cred, p);
423 } while (error == NQNFS_EXPIRED);
424 if (error)
425 return (error);
426 if (np->n_lrev != np->n_brev ||
427 (np->n_flag & NQNFSNONCACHE) ||
428 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
429 if (vp->v_type == VDIR)
430 nfs_invaldir(vp);
431 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
432 if (error)
433 return (error);
434 np->n_brev = np->n_lrev;
435 }
436 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
437 nfs_invaldir(vp);
438 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
439 if (error)
440 return (error);
441 }
442 }
443 if (np->n_flag & NQNFSNONCACHE) {
444 switch (vp->v_type) {
445 case VREG:
446 return (nfs_readrpc(vp, uio, cred));
447 case VLNK:
448 return (nfs_readlinkrpc(vp, uio, cred));
449 case VDIR:
450 break;
451 default:
452 printf(" NQNFSNONCACHE: type %x unexpected\n",
453 vp->v_type);
454 };
455 }
456 switch (vp->v_type) {
457 case VREG:
458 nfsstats.biocache_reads++;
459 lbn = uio->uio_offset / biosize;
460 on = uio->uio_offset & (biosize - 1);
461
462 /*
463 * Start the read ahead(s), as required.
464 */
465 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
466 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
467 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
468 rabn = lbn + 1 + nra;
469 if (!incore(vp, rabn)) {
470 rabp = nfs_getcacheblk(vp, rabn, biosize, p);
471 if (!rabp)
472 return (EINTR);
473 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
474 rabp->b_flags |= (B_READ | B_ASYNC);
475 vfs_busy_pages(rabp, 0);
476 if (nfs_asyncio(rabp, cred, p)) {
477 rabp->b_flags |= B_INVAL|B_ERROR;
478 vfs_unbusy_pages(rabp);
479 brelse(rabp);
480 break;
481 }
482 } else {
483 brelse(rabp);
484 }
485 }
486 }
487 }
488
489 /*
490 * Obtain the buffer cache block. Figure out the buffer size
491 * when we are at EOF. If we are modifying the size of the
492 * buffer based on an EOF condition we need to hold
493 * nfs_rslock() through obtaining the buffer to prevent
494 * a potential writer-appender from messing with n_size.
495 * Otherwise we may accidently truncate the buffer and
496 * lose dirty data.
497 *
498 * Note that bcount is *not* DEV_BSIZE aligned.
499 */
500
501again:
502 bcount = biosize;
503 if ((off_t)lbn * biosize >= np->n_size) {
504 bcount = 0;
505 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
506 bcount = np->n_size - (off_t)lbn * biosize;
507 }
508 if (bcount != biosize) {
509 switch(nfs_rslock(np, p)) {
510 case ENOLCK:
511 goto again;
512 /* not reached */
513 case EINTR:
514 case ERESTART:
515 return(EINTR);
516 /* not reached */
517 default:
518 break;
519 }
520 }
521
522 bp = nfs_getcacheblk(vp, lbn, bcount, p);
523
524 if (bcount != biosize)
525 nfs_rsunlock(np, p);
526 if (!bp)
527 return (EINTR);
528
529 /*
530 * If B_CACHE is not set, we must issue the read. If this
531 * fails, we return an error.
532 */
533
534 if ((bp->b_flags & B_CACHE) == 0) {
535 bp->b_flags |= B_READ;
536 vfs_busy_pages(bp, 0);
537 error = nfs_doio(bp, cred, p);
538 if (error) {
539 brelse(bp);
540 return (error);
541 }
542 }
543
544 /*
545 * on is the offset into the current bp. Figure out how many
546 * bytes we can copy out of the bp. Note that bcount is
547 * NOT DEV_BSIZE aligned.
548 *
549 * Then figure out how many bytes we can copy into the uio.
550 */
551
552 n = 0;
553 if (on < bcount)
554 n = min((unsigned)(bcount - on), uio->uio_resid);
555 break;
556 case VLNK:
557 nfsstats.biocache_readlinks++;
558 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
559 if (!bp)
560 return (EINTR);
561 if ((bp->b_flags & B_CACHE) == 0) {
562 bp->b_flags |= B_READ;
563 vfs_busy_pages(bp, 0);
564 error = nfs_doio(bp, cred, p);
565 if (error) {
566 bp->b_flags |= B_ERROR;
567 brelse(bp);
568 return (error);
569 }
570 }
571 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
572 on = 0;
573 break;
574 case VDIR:
575 nfsstats.biocache_readdirs++;
576 if (np->n_direofoffset
577 && uio->uio_offset >= np->n_direofoffset) {
578 return (0);
579 }
580 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
581 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
582 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p);
583 if (!bp)
584 return (EINTR);
585 if ((bp->b_flags & B_CACHE) == 0) {
586 bp->b_flags |= B_READ;
587 vfs_busy_pages(bp, 0);
588 error = nfs_doio(bp, cred, p);
589 if (error) {
590 brelse(bp);
591 }
592 while (error == NFSERR_BAD_COOKIE) {
593 printf("got bad cookie vp %p bp %p\n", vp, bp);
594 nfs_invaldir(vp);
595 error = nfs_vinvalbuf(vp, 0, cred, p, 1);
596 /*
597 * Yuck! The directory has been modified on the
598 * server. The only way to get the block is by
599 * reading from the beginning to get all the
600 * offset cookies.
601 *
602 * Leave the last bp intact unless there is an error.
603 * Loop back up to the while if the error is another
604 * NFSERR_BAD_COOKIE (double yuch!).
605 */
606 for (i = 0; i <= lbn && !error; i++) {
607 if (np->n_direofoffset
608 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
609 return (0);
610 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p);
611 if (!bp)
612 return (EINTR);
613 if ((bp->b_flags & B_CACHE) == 0) {
614 bp->b_flags |= B_READ;
615 vfs_busy_pages(bp, 0);
616 error = nfs_doio(bp, cred, p);
617 /*
618 * no error + B_INVAL == directory EOF,
619 * use the block.
620 */
621 if (error == 0 && (bp->b_flags & B_INVAL))
622 break;
623 }
624 /*
625 * An error will throw away the block and the
626 * for loop will break out. If no error and this
627 * is not the block we want, we throw away the
628 * block and go for the next one via the for loop.
629 */
630 if (error || i < lbn)
631 brelse(bp);
632 }
633 }
634 /*
635 * The above while is repeated if we hit another cookie
636 * error. If we hit an error and it wasn't a cookie error,
637 * we give up.
638 */
639 if (error)
640 return (error);
641 }
642
643 /*
644 * If not eof and read aheads are enabled, start one.
645 * (You need the current block first, so that you have the
646 * directory offset cookie of the next block.)
647 */
648 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
649 (bp->b_flags & B_INVAL) == 0 &&
650 (np->n_direofoffset == 0 ||
651 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
652 !(np->n_flag & NQNFSNONCACHE) &&
653 !incore(vp, lbn + 1)) {
654 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p);
655 if (rabp) {
656 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
657 rabp->b_flags |= (B_READ | B_ASYNC);
658 vfs_busy_pages(rabp, 0);
659 if (nfs_asyncio(rabp, cred, p)) {
660 rabp->b_flags |= B_INVAL|B_ERROR;
661 vfs_unbusy_pages(rabp);
662 brelse(rabp);
663 }
664 } else {
665 brelse(rabp);
666 }
667 }
668 }
669 /*
670 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
671 * chopped for the EOF condition, we cannot tell how large
672 * NFS directories are going to be until we hit EOF. So
673 * an NFS directory buffer is *not* chopped to its EOF. Now,
674 * it just so happens that b_resid will effectively chop it
675 * to EOF. *BUT* this information is lost if the buffer goes
676 * away and is reconstituted into a B_CACHE state ( due to
677 * being VMIO ) later. So we keep track of the directory eof
678 * in np->n_direofoffset and chop it off as an extra step
679 * right here.
680 */
681 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
682 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
683 n = np->n_direofoffset - uio->uio_offset;
684 break;
685 default:
686 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
687 break;
688 };
689
690 if (n > 0) {
691 error = uiomove(bp->b_data + on, (int)n, uio);
692 }
693 switch (vp->v_type) {
694 case VREG:
695 break;
696 case VLNK:
697 n = 0;
698 break;
699 case VDIR:
700 /*
701 * Invalidate buffer if caching is disabled, forcing a
702 * re-read from the remote later.
703 */
704 if (np->n_flag & NQNFSNONCACHE)
705 bp->b_flags |= B_INVAL;
706 break;
707 default:
708 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
709 }
710 brelse(bp);
711 } while (error == 0 && uio->uio_resid > 0 && n > 0);
712 return (error);
713}
714
715/*
716 * Vnode op for write using bio
717 */
718int
719nfs_write(ap)
720 struct vop_write_args /* {
721 struct vnode *a_vp;
722 struct uio *a_uio;
723 int a_ioflag;
724 struct ucred *a_cred;
725 } */ *ap;
726{
727 int biosize;
728 struct uio *uio = ap->a_uio;
729 struct proc *p = uio->uio_procp;
730 struct vnode *vp = ap->a_vp;
731 struct nfsnode *np = VTONFS(vp);
732 struct ucred *cred = ap->a_cred;
733 int ioflag = ap->a_ioflag;
734 struct buf *bp;
735 struct vattr vattr;
736 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
737 daddr_t lbn;
738 int bcount;
739 int n, on, error = 0, iomode, must_commit;
740 int haverslock = 0;
741
742#ifdef DIAGNOSTIC
743 if (uio->uio_rw != UIO_WRITE)
744 panic("nfs_write mode");
745 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
746 panic("nfs_write proc");
747#endif
748 if (vp->v_type != VREG)
749 return (EIO);
750 if (np->n_flag & NWRITEERR) {
751 np->n_flag &= ~NWRITEERR;
752 return (np->n_error);
753 }
754 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
755 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
756 (void)nfs_fsinfo(nmp, vp, cred, p);
757
758 /*
759 * Synchronously flush pending buffers if we are in synchronous
760 * mode or if we are appending.
761 */
762 if (ioflag & (IO_APPEND | IO_SYNC)) {
763 if (np->n_flag & NMODIFIED) {
764 np->n_attrstamp = 0;
765 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
766 if (error)
767 return (error);
768 }
769 }
770
771 /*
772 * If IO_APPEND then load uio_offset. We restart here if we cannot
773 * get the append lock.
774 */
775restart:
776 if (ioflag & IO_APPEND) {
777 np->n_attrstamp = 0;
778 error = VOP_GETATTR(vp, &vattr, cred, p);
779 if (error)
780 return (error);
781 uio->uio_offset = np->n_size;
782 }
783
784 if (uio->uio_offset < 0)
785 return (EINVAL);
786 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
787 return (EFBIG);
788 if (uio->uio_resid == 0)
789 return (0);
790
791 /*
792 * We need to obtain the rslock if we intend to modify np->n_size
793 * in order to guarentee the append point with multiple contending
794 * writers, to guarentee that no other appenders modify n_size
795 * while we are trying to obtain a truncated buffer (i.e. to avoid
796 * accidently truncating data written by another appender due to
797 * the race), and to ensure that the buffer is populated prior to
798 * our extending of the file. We hold rslock through the entire
799 * operation.
800 *
801 * Note that we do not synchronize the case where someone truncates
802 * the file while we are appending to it because attempting to lock
803 * this case may deadlock other parts of the system unexpectedly.
804 */
805 if ((ioflag & IO_APPEND) ||
806 uio->uio_offset + uio->uio_resid > np->n_size) {
807 switch(nfs_rslock(np, p)) {
808 case ENOLCK:
809 goto restart;
810 /* not reached */
811 case EINTR:
812 case ERESTART:
813 return(EINTR);
814 /* not reached */
815 default:
816 break;
817 }
818 haverslock = 1;
819 }
820
821 /*
822 * Maybe this should be above the vnode op call, but so long as
823 * file servers have no limits, i don't think it matters
824 */
825 if (p && uio->uio_offset + uio->uio_resid >
826 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
827 psignal(p, SIGXFSZ);
828 if (haverslock)
829 nfs_rsunlock(np, p);
830 return (EFBIG);
831 }
832
833 biosize = vp->v_mount->mnt_stat.f_iosize;
834
835 do {
836 /*
837 * Check for a valid write lease.
838 */
839 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
840 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
841 do {
842 error = nqnfs_getlease(vp, ND_WRITE, cred, p);
843 } while (error == NQNFS_EXPIRED);
844 if (error)
845 break;
846 if (np->n_lrev != np->n_brev ||
847 (np->n_flag & NQNFSNONCACHE)) {
848 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
849 if (error)
850 break;
851 np->n_brev = np->n_lrev;
852 }
853 }
854 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
855 iomode = NFSV3WRITE_FILESYNC;
856 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit);
857 if (must_commit)
858 nfs_clearcommit(vp->v_mount);
859 break;
860 }
861 nfsstats.biocache_writes++;
862 lbn = uio->uio_offset / biosize;
863 on = uio->uio_offset & (biosize-1);
864 n = min((unsigned)(biosize - on), uio->uio_resid);
865again:
866 /*
867 * Handle direct append and file extension cases, calculate
868 * unaligned buffer size.
869 */
870
871 if (uio->uio_offset == np->n_size && n) {
872 /*
873 * Get the buffer (in its pre-append state to maintain
874 * B_CACHE if it was previously set). Resize the
875 * nfsnode after we have locked the buffer to prevent
876 * readers from reading garbage.
877 */
878 bcount = on;
879 bp = nfs_getcacheblk(vp, lbn, bcount, p);
880
881 if (bp != NULL) {
882 long save;
883
884 np->n_size = uio->uio_offset + n;
885 np->n_flag |= NMODIFIED;
886 vnode_pager_setsize(vp, np->n_size);
887
888 save = bp->b_flags & B_CACHE;
889 bcount += n;
890 allocbuf(bp, bcount);
891 bp->b_flags |= save;
892 }
893 } else {
894 /*
895 * Obtain the locked cache block first, and then
896 * adjust the file's size as appropriate.
897 */
898 bcount = on + n;
899 if ((off_t)lbn * biosize + bcount < np->n_size) {
900 if ((off_t)(lbn + 1) * biosize < np->n_size)
901 bcount = biosize;
902 else
903 bcount = np->n_size - (off_t)lbn * biosize;
904 }
905 bp = nfs_getcacheblk(vp, lbn, bcount, p);
906 if (uio->uio_offset + n > np->n_size) {
907 np->n_size = uio->uio_offset + n;
908 np->n_flag |= NMODIFIED;
909 vnode_pager_setsize(vp, np->n_size);
910 }
911 }
912
913 if (!bp) {
914 error = EINTR;
915 break;
916 }
917
918 /*
919 * Issue a READ if B_CACHE is not set. In special-append
920 * mode, B_CACHE is based on the buffer prior to the write
921 * op and is typically set, avoiding the read. If a read
922 * is required in special append mode, the server will
923 * probably send us a short-read since we extended the file
924 * on our end, resulting in b_resid == 0 and, thusly,
925 * B_CACHE getting set.
926 *
927 * We can also avoid issuing the read if the write covers
928 * the entire buffer. We have to make sure the buffer state
929 * is reasonable in this case since we will not be initiating
930 * I/O. See the comments in kern/vfs_bio.c's getblk() for
931 * more information.
932 *
933 * B_CACHE may also be set due to the buffer being cached
934 * normally.
935 */
936
937 if (on == 0 && n == bcount) {
938 bp->b_flags |= B_CACHE;
939 bp->b_flags &= ~(B_ERROR | B_INVAL);
940 }
941
942 if ((bp->b_flags & B_CACHE) == 0) {
943 bp->b_flags |= B_READ;
944 vfs_busy_pages(bp, 0);
945 error = nfs_doio(bp, cred, p);
946 if (error) {
947 brelse(bp);
948 break;
949 }
950 }
951 if (!bp) {
952 error = EINTR;
953 break;
954 }
955 if (bp->b_wcred == NOCRED) {
956 crhold(cred);
957 bp->b_wcred = cred;
958 }
959 np->n_flag |= NMODIFIED;
960
961 /*
962 * If dirtyend exceeds file size, chop it down. This should
963 * not normally occur but there is an append race where it
964 * might occur XXX, so we log it.
965 *
966 * If the chopping creates a reverse-indexed or degenerate
967 * situation with dirtyoff/end, we 0 both of them.
968 */
969
970 if (bp->b_dirtyend > bcount) {
971 printf("NFS append race @%lx:%d\n",
972 (long)bp->b_blkno * DEV_BSIZE,
973 bp->b_dirtyend - bcount);
974 bp->b_dirtyend = bcount;
975 }
976
977 if (bp->b_dirtyoff >= bp->b_dirtyend)
978 bp->b_dirtyoff = bp->b_dirtyend = 0;
979
980 /*
981 * If the new write will leave a contiguous dirty
982 * area, just update the b_dirtyoff and b_dirtyend,
983 * otherwise force a write rpc of the old dirty area.
984 *
985 * While it is possible to merge discontiguous writes due to
986 * our having a B_CACHE buffer ( and thus valid read data
987 * for the hole), we don't because it could lead to
988 * significant cache coherency problems with multiple clients,
989 * especially if locking is implemented later on.
990 *
991 * as an optimization we could theoretically maintain
992 * a linked list of discontinuous areas, but we would still
993 * have to commit them separately so there isn't much
994 * advantage to it except perhaps a bit of asynchronization.
995 */
996
997 if (bp->b_dirtyend > 0 &&
998 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
999 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) {
1000 error = EINTR;
1001 break;
1002 }
1003 goto again;
1004 }
1005
1006 /*
1007 * Check for valid write lease and get one as required.
1008 * In case getblk() and/or bwrite() delayed us.
1009 */
1010 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
1011 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
1012 do {
1013 error = nqnfs_getlease(vp, ND_WRITE, cred, p);
1014 } while (error == NQNFS_EXPIRED);
1015 if (error) {
1016 brelse(bp);
1017 break;
1018 }
1019 if (np->n_lrev != np->n_brev ||
1020 (np->n_flag & NQNFSNONCACHE)) {
1021 brelse(bp);
1022 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
1023 if (error)
1024 break;
1025 np->n_brev = np->n_lrev;
1026 goto again;
1027 }
1028 }
1029
1030 error = uiomove((char *)bp->b_data + on, n, uio);
1031
1032 /*
1033 * Since this block is being modified, it must be written
1034 * again and not just committed. Since write clustering does
1035 * not work for the stage 1 data write, only the stage 2
1036 * commit rpc, we have to clear B_CLUSTEROK as well.
1037 */
1038 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1039
1040 if (error) {
1041 bp->b_flags |= B_ERROR;
1042 brelse(bp);
1043 break;
1044 }
1045
1046 /*
1047 * Only update dirtyoff/dirtyend if not a degenerate
1048 * condition.
1049 */
1050 if (n) {
1051 if (bp->b_dirtyend > 0) {
1052 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1053 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1054 } else {
1055 bp->b_dirtyoff = on;
1056 bp->b_dirtyend = on + n;
1057 }
1058 vfs_bio_set_validclean(bp, on, n);
1059 }
1060 /*
1061 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1062 * filesystem). XXX also use for loopback NFS mounts.
1063 */
1064 if (ioflag & IO_NOWDRAIN)
1065 bp->b_flags |= B_NOWDRAIN;
1066
1067 /*
1068 * If the lease is non-cachable or IO_SYNC do bwrite().
1069 *
1070 * IO_INVAL appears to be unused. The idea appears to be
1071 * to turn off caching in this case. Very odd. XXX
1072 */
1073 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1074 if (ioflag & IO_INVAL)
1075 bp->b_flags |= B_NOCACHE;
1076 error = VOP_BWRITE(bp->b_vp, bp);
1077 if (error)
1078 break;
1079 if (np->n_flag & NQNFSNONCACHE) {
1080 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
1081 if (error)
1082 break;
1083 }
1084 } else if ((n + on) == biosize &&
1085 (nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1086 bp->b_flags |= B_ASYNC;
1087 (void)nfs_writebp(bp, 0, 0);
1088 } else {
1089 bdwrite(bp);
1090 }
1091 } while (uio->uio_resid > 0 && n > 0);
1092
1093 if (haverslock)
1094 nfs_rsunlock(np, p);
1095
1096 return (error);
1097}
1098
1099/*
1100 * Get an nfs cache block.
1101 *
1102 * Allocate a new one if the block isn't currently in the cache
1103 * and return the block marked busy. If the calling process is
1104 * interrupted by a signal for an interruptible mount point, return
1105 * NULL.
1106 *
1107 * The caller must carefully deal with the possible B_INVAL state of
1108 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1109 * indirectly), so synchronous reads can be issued without worrying about
1110 * the B_INVAL state. We have to be a little more careful when dealing
1111 * with writes (see comments in nfs_write()) when extending a file past
1112 * its EOF.
1113 */
1114static struct buf *
1115nfs_getcacheblk(vp, bn, size, p)
1116 struct vnode *vp;
1117 daddr_t bn;
1118 int size;
1119 struct proc *p;
1120{
1121 register struct buf *bp;
1122 struct mount *mp;
1123 struct nfsmount *nmp;
1124
1125 mp = vp->v_mount;
1126 nmp = VFSTONFS(mp);
1127
1128 if (nmp->nm_flag & NFSMNT_INT) {
1129 bp = getblk(vp, bn, size, PCATCH, 0);
1130 while (bp == (struct buf *)0) {
1131 if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
1132 return ((struct buf *)0);
1133 bp = getblk(vp, bn, size, 0, 2 * hz);
1134 }
1135 } else {
1136 bp = getblk(vp, bn, size, 0, 0);
1137 }
1138
1139 if (vp->v_type == VREG) {
1140 int biosize;
1141
1142 biosize = mp->mnt_stat.f_iosize;
1143 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1144 }
1145 return (bp);
1146}
1147
1148/*
1149 * Flush and invalidate all dirty buffers. If another process is already
1150 * doing the flush, just wait for completion.
1151 */
1152int
1153nfs_vinvalbuf(vp, flags, cred, p, intrflg)
1154 struct vnode *vp;
1155 int flags;
1156 struct ucred *cred;
1157 struct proc *p;
1158 int intrflg;
1159{
1160 register struct nfsnode *np = VTONFS(vp);
1161 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1162 int error = 0, slpflag, slptimeo;
1163
1164 if (vp->v_flag & VXLOCK) {
1165 return (0);
1166 }
1167
1168 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1169 intrflg = 0;
1170 if (intrflg) {
1171 slpflag = PCATCH;
1172 slptimeo = 2 * hz;
1173 } else {
1174 slpflag = 0;
1175 slptimeo = 0;
1176 }
1177 /*
1178 * First wait for any other process doing a flush to complete.
1179 */
1180 while (np->n_flag & NFLUSHINPROG) {
1181 np->n_flag |= NFLUSHWANT;
1182 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
1183 slptimeo);
1184 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
1185 return (EINTR);
1186 }
1187
1188 /*
1189 * Now, flush as required.
1190 */
1191 np->n_flag |= NFLUSHINPROG;
1192 error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
1193 while (error) {
1194 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
1195 np->n_flag &= ~NFLUSHINPROG;
1196 if (np->n_flag & NFLUSHWANT) {
1197 np->n_flag &= ~NFLUSHWANT;
1198 wakeup((caddr_t)&np->n_flag);
1199 }
1200 return (EINTR);
1201 }
1202 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
1203 }
1204 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1205 if (np->n_flag & NFLUSHWANT) {
1206 np->n_flag &= ~NFLUSHWANT;
1207 wakeup((caddr_t)&np->n_flag);
1208 }
1209 return (0);
1210}
1211
1212/*
1213 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1214 * This is mainly to avoid queueing async I/O requests when the nfsiods
1215 * are all hung on a dead server.
1216 *
1217 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1218 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1219 */
1220int
1221nfs_asyncio(bp, cred, procp)
1222 register struct buf *bp;
1223 struct ucred *cred;
1224 struct proc *procp;
1225{
1226 struct nfsmount *nmp;
1227 int i;
1228 int gotiod;
1229 int slpflag = 0;
1230 int slptimeo = 0;
1231 int error;
1232
1233 /*
1234 * If no async daemons then return EIO to force caller to run the rpc
1235 * synchronously.
1236 */
1237 if (nfs_numasync == 0)
1238 return (EIO);
1239
1240 nmp = VFSTONFS(bp->b_vp->v_mount);
1241
1242 /*
1243 * Commits are usually short and sweet so lets save some cpu and
1244 * leave the async daemons for more important rpc's (such as reads
1245 * and writes).
1246 */
1247 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT &&
1248 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1249 return(EIO);
1250 }
1251
1252again:
1253 if (nmp->nm_flag & NFSMNT_INT)
1254 slpflag = PCATCH;
1255 gotiod = FALSE;
1256
1257 /*
1258 * Find a free iod to process this request.
1259 */
1260 for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1261 if (nfs_iodwant[i]) {
1262 /*
1263 * Found one, so wake it up and tell it which
1264 * mount to process.
1265 */
1266 NFS_DPF(ASYNCIO,
1267 ("nfs_asyncio: waking iod %d for mount %p\n",
1268 i, nmp));
1269 nfs_iodwant[i] = (struct proc *)0;
1270 nfs_iodmount[i] = nmp;
1271 nmp->nm_bufqiods++;
1272 wakeup((caddr_t)&nfs_iodwant[i]);
1273 gotiod = TRUE;
1274 break;
1275 }
1276
1277 /*
1278 * If none are free, we may already have an iod working on this mount
1279 * point. If so, it will process our request.
1280 */
1281 if (!gotiod) {
1282 if (nmp->nm_bufqiods > 0) {
1283 NFS_DPF(ASYNCIO,
1284 ("nfs_asyncio: %d iods are already processing mount %p\n",
1285 nmp->nm_bufqiods, nmp));
1286 gotiod = TRUE;
1287 }
1288 }
1289
1290 /*
1291 * If we have an iod which can process the request, then queue
1292 * the buffer.
1293 */
1294 if (gotiod) {
1295 /*
1296 * Ensure that the queue never grows too large. We still want
1297 * to asynchronize so we block rather then return EIO.
1298 */
1299 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1300 NFS_DPF(ASYNCIO,
1301 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1302 nmp->nm_bufqwant = TRUE;
1303 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
1304 "nfsaio", slptimeo);
1305 if (error) {
1306 if (nfs_sigintr(nmp, NULL, procp))
1307 return (EINTR);
1308 if (slpflag == PCATCH) {
1309 slpflag = 0;
1310 slptimeo = 2 * hz;
1311 }
1312 }
1313 /*
1314 * We might have lost our iod while sleeping,
1315 * so check and loop if nescessary.
1316 */
1317 if (nmp->nm_bufqiods == 0) {
1318 NFS_DPF(ASYNCIO,
1319 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1320 goto again;
1321 }
1322 }
1323
1324 if (bp->b_flags & B_READ) {
1325 if (bp->b_rcred == NOCRED && cred != NOCRED) {
1326 crhold(cred);
1327 bp->b_rcred = cred;
1328 }
1329 } else {
1330 bp->b_flags |= B_WRITEINPROG;
1331 if (bp->b_wcred == NOCRED && cred != NOCRED) {
1332 crhold(cred);
1333 bp->b_wcred = cred;
1334 }
1335 }
1336
1337 BUF_KERNPROC(bp);
1338 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1339 nmp->nm_bufqlen++;
1340 return (0);
1341 }
1342
1343 /*
1344 * All the iods are busy on other mounts, so return EIO to
1345 * force the caller to process the i/o synchronously.
1346 */
1347 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1348 return (EIO);
1349}
1350
1351/*
1352 * Do an I/O operation to/from a cache block. This may be called
1353 * synchronously or from an nfsiod.
1354 */
1355int
1356nfs_doio(bp, cr, p)
1357 struct buf *bp;
1358 struct ucred *cr;
1359 struct proc *p;
1360{
1361 struct uio *uiop;
1362 struct vnode *vp;
1363 struct nfsnode *np;
1364 struct nfsmount *nmp;
1365 int error = 0, iomode, must_commit = 0;
1366 struct uio uio;
1367 struct iovec io;
1368
1369 vp = bp->b_vp;
1370 np = VTONFS(vp);
1371 nmp = VFSTONFS(vp->v_mount);
1372 uiop = &uio;
1373 uiop->uio_iov = &io;
1374 uiop->uio_iovcnt = 1;
1375 uiop->uio_segflg = UIO_SYSSPACE;
1376 uiop->uio_procp = p;
1377
1378 /*
1379 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1380 * do this here so we do not have to do it in all the code that
1381 * calls us.
1382 */
1383 bp->b_flags &= ~(B_ERROR | B_INVAL);
1384
1385 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1386
1387 /*
1388 * Historically, paging was done with physio, but no more.
1389 */
1390 if (bp->b_flags & B_PHYS) {
1391 /*
1392 * ...though reading /dev/drum still gets us here.
1393 */
1394 io.iov_len = uiop->uio_resid = bp->b_bcount;
1395 /* mapping was done by vmapbuf() */
1396 io.iov_base = bp->b_data;
1397 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1398 if (bp->b_flags & B_READ) {
1399 uiop->uio_rw = UIO_READ;
1400 nfsstats.read_physios++;
1401 error = nfs_readrpc(vp, uiop, cr);
1402 } else {
1403 int com;
1404
1405 iomode = NFSV3WRITE_DATASYNC;
1406 uiop->uio_rw = UIO_WRITE;
1407 nfsstats.write_physios++;
1408 error = nfs_writerpc(vp, uiop, cr, &iomode, &com);
1409 }
1410 if (error) {
1411 bp->b_flags |= B_ERROR;
1412 bp->b_error = error;
1413 }
1414 } else if (bp->b_flags & B_READ) {
1415 io.iov_len = uiop->uio_resid = bp->b_bcount;
1416 io.iov_base = bp->b_data;
1417 uiop->uio_rw = UIO_READ;
1418
1419 switch (vp->v_type) {
1420 case VREG:
1421 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1422 nfsstats.read_bios++;
1423 error = nfs_readrpc(vp, uiop, cr);
1424
1425 if (!error) {
1426 if (uiop->uio_resid) {
1427 /*
1428 * If we had a short read with no error, we must have
1429 * hit a file hole. We should zero-fill the remainder.
1430 * This can also occur if the server hits the file EOF.
1431 *
1432 * Holes used to be able to occur due to pending
1433 * writes, but that is not possible any longer.
1434 */
1435 int nread = bp->b_bcount - uiop->uio_resid;
1436 int left = uiop->uio_resid;
1437
1438 if (left > 0)
1439 bzero((char *)bp->b_data + nread, left);
1440 uiop->uio_resid = 0;
1441 }
1442 }
1443 if (p && (vp->v_flag & VTEXT) &&
1444 (((nmp->nm_flag & NFSMNT_NQNFS) &&
1445 NQNFS_CKINVALID(vp, np, ND_READ) &&
1446 np->n_lrev != np->n_brev) ||
1447 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1448 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1449 uprintf("Process killed due to text file modification\n");
1450 psignal(p, SIGKILL);
1451 PHOLD(p);
1452 }
1453 break;
1454 case VLNK:
1455 uiop->uio_offset = (off_t)0;
1456 nfsstats.readlink_bios++;
1457 error = nfs_readlinkrpc(vp, uiop, cr);
1458 break;
1459 case VDIR:
1460 nfsstats.readdir_bios++;
1461 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1462 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1463 error = nfs_readdirplusrpc(vp, uiop, cr);
1464 if (error == NFSERR_NOTSUPP)
1465 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1466 }
1467 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1468 error = nfs_readdirrpc(vp, uiop, cr);
1469 /*
1470 * end-of-directory sets B_INVAL but does not generate an
1471 * error.
1472 */
1473 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1474 bp->b_flags |= B_INVAL;
1475 break;
1476 default:
1477 printf("nfs_doio: type %x unexpected\n",vp->v_type);
1478 break;
1479 };
1480 if (error) {
1481 bp->b_flags |= B_ERROR;
1482 bp->b_error = error;
1483 }
1484 } else {
1485 /*
1486 * If we only need to commit, try to commit
1487 */
1488 if (bp->b_flags & B_NEEDCOMMIT) {
1489 int retv;
1490 off_t off;
1491
1492 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1493 bp->b_flags |= B_WRITEINPROG;
1494 retv = nfs_commit(
1495 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1496 bp->b_wcred, p);
1497 bp->b_flags &= ~B_WRITEINPROG;
1498 if (retv == 0) {
1499 bp->b_dirtyoff = bp->b_dirtyend = 0;
1500 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1501 bp->b_resid = 0;
1502 biodone(bp);
1503 return (0);
1504 }
1505 if (retv == NFSERR_STALEWRITEVERF) {
1506 nfs_clearcommit(bp->b_vp->v_mount);
1507 }
1508 }
1509
1510 /*
1511 * Setup for actual write
1512 */
1513
1514 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1515 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1516
1517 if (bp->b_dirtyend > bp->b_dirtyoff) {
1518 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1519 - bp->b_dirtyoff;
1520 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1521 + bp->b_dirtyoff;
1522 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1523 uiop->uio_rw = UIO_WRITE;
1524 nfsstats.write_bios++;
1525
1526 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1527 iomode = NFSV3WRITE_UNSTABLE;
1528 else
1529 iomode = NFSV3WRITE_FILESYNC;
1530
1531 bp->b_flags |= B_WRITEINPROG;
1532 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1533
1534 /*
1535 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1536 * to cluster the buffers needing commit. This will allow
1537 * the system to submit a single commit rpc for the whole
1538 * cluster. We can do this even if the buffer is not 100%
1539 * dirty (relative to the NFS blocksize), so we optimize the
1540 * append-to-file-case.
1541 *
1542 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1543 * cleared because write clustering only works for commit
1544 * rpc's, not for the data portion of the write).
1545 */
1546
1547 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1548 bp->b_flags |= B_NEEDCOMMIT;
1549 if (bp->b_dirtyoff == 0
1550 && bp->b_dirtyend == bp->b_bcount)
1551 bp->b_flags |= B_CLUSTEROK;
1552 } else {
1553 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1554 }
1555 bp->b_flags &= ~B_WRITEINPROG;
1556
1557 /*
1558 * For an interrupted write, the buffer is still valid
1559 * and the write hasn't been pushed to the server yet,
1560 * so we can't set B_ERROR and report the interruption
1561 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1562 * is not relevant, so the rpc attempt is essentially
1563 * a noop. For the case of a V3 write rpc not being
1564 * committed to stable storage, the block is still
1565 * dirty and requires either a commit rpc or another
1566 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1567 * the block is reused. This is indicated by setting
1568 * the B_DELWRI and B_NEEDCOMMIT flags.
1569 *
1570 * If the buffer is marked B_PAGING, it does not reside on
1571 * the vp's paging queues so we cannot call bdirty(). The
1572 * bp in this case is not an NFS cache block so we should
1573 * be safe. XXX
1574 */
1575 if (error == EINTR
1576 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1577 int s;
1578
1579 s = splbio();
1580 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1581 if ((bp->b_flags & B_PAGING) == 0) {
1582 bdirty(bp);
1583 bp->b_flags &= ~B_DONE;
1584 }
1585 if (error && (bp->b_flags & B_ASYNC) == 0)
1586 bp->b_flags |= B_EINTR;
1587 splx(s);
1588 } else {
1589 if (error) {
1590 bp->b_flags |= B_ERROR;
1591 bp->b_error = np->n_error = error;
1592 np->n_flag |= NWRITEERR;
1593 }
1594 bp->b_dirtyoff = bp->b_dirtyend = 0;
1595 }
1596 } else {
1597 bp->b_resid = 0;
1598 biodone(bp);
1599 return (0);
1600 }
1601 }
1602 bp->b_resid = uiop->uio_resid;
1603 if (must_commit)
1604 nfs_clearcommit(vp->v_mount);
1605 biodone(bp);
1606 return (error);
1607}
1608
1609/*
1610 * Used to aid in handling ftruncate() operations on the NFS client side.
1611 * Truncation creates a number of special problems for NFS. We have to
1612 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1613 * we have to properly handle VM pages or (potentially dirty) buffers
1614 * that straddle the truncation point.
1615 */
1616
1617int
1618nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct proc *p, u_quad_t nsize)
1619{
1620 struct nfsnode *np = VTONFS(vp);
1621 u_quad_t tsize = np->n_size;
1622 int biosize = vp->v_mount->mnt_stat.f_iosize;
1623 int error = 0;
1624
1625 np->n_size = nsize;
1626
1627 if (np->n_size < tsize) {
1628 struct buf *bp;
1629 daddr_t lbn;
1630 int bufsize;
1631
1632 /*
1633 * vtruncbuf() doesn't get the buffer overlapping the
1634 * truncation point. We may have a B_DELWRI and/or B_CACHE
1635 * buffer that now needs to be truncated.
1636 */
1637 error = vtruncbuf(vp, cred, p, nsize, biosize);
1638 lbn = nsize / biosize;
1639 bufsize = nsize & (biosize - 1);
1640 bp = nfs_getcacheblk(vp, lbn, bufsize, p);
1641 if (bp->b_dirtyoff > bp->b_bcount)
1642 bp->b_dirtyoff = bp->b_bcount;
1643 if (bp->b_dirtyend > bp->b_bcount)
1644 bp->b_dirtyend = bp->b_bcount;
1645 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1646 brelse(bp);
1647 } else {
1648 vnode_pager_setsize(vp, nsize);
1649 }
1650 return(error);
1651}
1652