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 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
33 */
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD: releng/6.2/sys/nfsclient/nfs_bio.c 157830 2006-04-18 05:31:58Z delphij $");
37
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/bio.h>
41 #include <sys/buf.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/proc.h>
45 #include <sys/resourcevar.h>
46 #include <sys/signalvar.h>
47 #include <sys/vmmeter.h>
48 #include <sys/vnode.h>
49
50 #include <vm/vm.h>
51 #include <vm/vm_extern.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_pager.h>
55 #include <vm/vnode_pager.h>
56
57 #include <rpc/rpcclnt.h>
58
59 #include <nfs/rpcv2.h>
60 #include <nfs/nfsproto.h>
61 #include <nfsclient/nfs.h>
62 #include <nfsclient/nfsmount.h>
63 #include <nfsclient/nfsnode.h>
64
65 #include <nfs4client/nfs4.h>
66
67 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
68 struct thread *td);
69 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
70 struct ucred *cred, int ioflag);
71
72 extern int nfs_directio_enable;
73 extern int nfs_directio_allow_mmap;
74 /*
75 * Vnode op for VM getpages.
76 */
77 int
78 nfs_getpages(struct vop_getpages_args *ap)
79 {
80 int i, error, nextoff, size, toff, count, npages;
81 struct uio uio;
82 struct iovec iov;
83 vm_offset_t kva;
84 struct buf *bp;
85 struct vnode *vp;
86 struct thread *td;
87 struct ucred *cred;
88 struct nfsmount *nmp;
89 vm_object_t object;
90 vm_page_t *pages;
91 struct nfsnode *np;
92
93 GIANT_REQUIRED;
94
95 vp = ap->a_vp;
96 np = VTONFS(vp);
97 td = curthread; /* XXX */
98 cred = curthread->td_ucred; /* XXX */
99 nmp = VFSTONFS(vp->v_mount);
100 pages = ap->a_m;
101 count = ap->a_count;
102
103 if ((object = vp->v_object) == NULL) {
104 printf("nfs_getpages: called with non-merged cache vnode??\n");
105 return VM_PAGER_ERROR;
106 }
107
108 if (!nfs_directio_allow_mmap && (np->n_flag & NNONCACHE) &&
109 (vp->v_type == VREG)) {
110 printf("nfs_getpages: called on non-cacheable vnode??\n");
111 return VM_PAGER_ERROR;
112 }
113
114 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
115 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
116 /* We'll never get here for v4, because we always have fsinfo */
117 (void)nfs_fsinfo(nmp, vp, cred, td);
118 }
119
120 npages = btoc(count);
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
128 {
129 vm_page_t m = pages[ap->a_reqpage];
130
131 VM_OBJECT_LOCK(object);
132 vm_page_lock_queues();
133 if (m->valid != 0) {
134 /* handled by vm_fault now */
135 /* vm_page_zero_invalid(m, TRUE); */
136 for (i = 0; i < npages; ++i) {
137 if (i != ap->a_reqpage)
138 vm_page_free(pages[i]);
139 }
140 vm_page_unlock_queues();
141 VM_OBJECT_UNLOCK(object);
142 return(0);
143 }
144 vm_page_unlock_queues();
145 VM_OBJECT_UNLOCK(object);
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 = (nmp->nm_rpcops->nr_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_OBJECT_LOCK(object);
177 vm_page_lock_queues();
178 for (i = 0; i < npages; ++i) {
179 if (i != ap->a_reqpage)
180 vm_page_free(pages[i]);
181 }
182 vm_page_unlock_queues();
183 VM_OBJECT_UNLOCK(object);
184 return VM_PAGER_ERROR;
185 }
186
187 /*
188 * Calculate the number of bytes read and validate only that number
189 * of bytes. Note that due to pending writes, size may be 0. This
190 * does not mean that the remaining data is invalid!
191 */
192
193 size = count - uio.uio_resid;
194 VM_OBJECT_LOCK(object);
195 vm_page_lock_queues();
196 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
197 vm_page_t m;
198 nextoff = toff + PAGE_SIZE;
199 m = pages[i];
200
201 if (nextoff <= size) {
202 /*
203 * Read operation filled an entire page
204 */
205 m->valid = VM_PAGE_BITS_ALL;
206 vm_page_undirty(m);
207 } else if (size > toff) {
208 /*
209 * Read operation filled a partial page.
210 */
211 m->valid = 0;
212 vm_page_set_validclean(m, 0, size - toff);
213 /* handled by vm_fault now */
214 /* vm_page_zero_invalid(m, TRUE); */
215 } else {
216 /*
217 * Read operation was short. If no error occured
218 * we may have hit a zero-fill section. We simply
219 * leave valid set to 0.
220 */
221 ;
222 }
223 if (i != ap->a_reqpage) {
224 /*
225 * Whether or not to leave the page activated is up in
226 * the air, but we should put the page on a page queue
227 * somewhere (it already is in the object). Result:
228 * It appears that emperical results show that
229 * deactivating pages is best.
230 */
231
232 /*
233 * Just in case someone was asking for this page we
234 * now tell them that it is ok to use.
235 */
236 if (!error) {
237 if (m->flags & PG_WANTED)
238 vm_page_activate(m);
239 else
240 vm_page_deactivate(m);
241 vm_page_wakeup(m);
242 } else {
243 vm_page_free(m);
244 }
245 }
246 }
247 vm_page_unlock_queues();
248 VM_OBJECT_UNLOCK(object);
249 return 0;
250 }
251
252 /*
253 * Vnode op for VM putpages.
254 */
255 int
256 nfs_putpages(struct vop_putpages_args *ap)
257 {
258 struct uio uio;
259 struct iovec iov;
260 vm_offset_t kva;
261 struct buf *bp;
262 int iomode, must_commit, i, error, npages, count;
263 off_t offset;
264 int *rtvals;
265 struct vnode *vp;
266 struct thread *td;
267 struct ucred *cred;
268 struct nfsmount *nmp;
269 struct nfsnode *np;
270 vm_page_t *pages;
271
272 GIANT_REQUIRED;
273
274 vp = ap->a_vp;
275 np = VTONFS(vp);
276 td = curthread; /* XXX */
277 cred = curthread->td_ucred; /* XXX */
278 nmp = VFSTONFS(vp->v_mount);
279 pages = ap->a_m;
280 count = ap->a_count;
281 rtvals = ap->a_rtvals;
282 npages = btoc(count);
283 offset = IDX_TO_OFF(pages[0]->pindex);
284
285 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
286 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
287 (void)nfs_fsinfo(nmp, vp, cred, td);
288 }
289
290 if (!nfs_directio_allow_mmap && (np->n_flag & NNONCACHE) &&
291 (vp->v_type == VREG))
292 printf("nfs_putpages: called on noncache-able vnode??\n");
293
294 for (i = 0; i < npages; i++)
295 rtvals[i] = VM_PAGER_AGAIN;
296
297 /*
298 * When putting pages, do not extend file past EOF.
299 */
300
301 if (offset + count > np->n_size) {
302 count = np->n_size - offset;
303 if (count < 0)
304 count = 0;
305 }
306
307 /*
308 * We use only the kva address for the buffer, but this is extremely
309 * convienient and fast.
310 */
311 bp = getpbuf(&nfs_pbuf_freecnt);
312
313 kva = (vm_offset_t) bp->b_data;
314 pmap_qenter(kva, pages, npages);
315 cnt.v_vnodeout++;
316 cnt.v_vnodepgsout += count;
317
318 iov.iov_base = (caddr_t) kva;
319 iov.iov_len = count;
320 uio.uio_iov = &iov;
321 uio.uio_iovcnt = 1;
322 uio.uio_offset = offset;
323 uio.uio_resid = count;
324 uio.uio_segflg = UIO_SYSSPACE;
325 uio.uio_rw = UIO_WRITE;
326 uio.uio_td = td;
327
328 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
329 iomode = NFSV3WRITE_UNSTABLE;
330 else
331 iomode = NFSV3WRITE_FILESYNC;
332
333 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
334
335 pmap_qremove(kva, npages);
336 relpbuf(bp, &nfs_pbuf_freecnt);
337
338 if (!error) {
339 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
340 for (i = 0; i < nwritten; i++) {
341 rtvals[i] = VM_PAGER_OK;
342 vm_page_undirty(pages[i]);
343 }
344 if (must_commit) {
345 nfs_clearcommit(vp->v_mount);
346 }
347 }
348 return rtvals[0];
349 }
350
351 /*
352 * Vnode op for read using bio
353 */
354 int
355 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
356 {
357 struct nfsnode *np = VTONFS(vp);
358 int biosize, i;
359 struct buf *bp, *rabp;
360 struct vattr vattr;
361 struct thread *td;
362 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
363 daddr_t lbn, rabn;
364 int bcount;
365 int seqcount;
366 int nra, error = 0, n = 0, on = 0;
367
368 #ifdef DIAGNOSTIC
369 if (uio->uio_rw != UIO_READ)
370 panic("nfs_read mode");
371 #endif
372 if (uio->uio_resid == 0)
373 return (0);
374 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
375 return (EINVAL);
376 td = uio->uio_td;
377
378 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
379 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
380 (void)nfs_fsinfo(nmp, vp, cred, td);
381 if (vp->v_type != VDIR &&
382 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
383 return (EFBIG);
384
385 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
386 /* No caching/ no readaheads. Just read data into the user buffer */
387 return nfs_readrpc(vp, uio, cred);
388
389 biosize = vp->v_mount->mnt_stat.f_iosize;
390 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
391 /*
392 * For nfs, cache consistency can only be maintained approximately.
393 * Although RFC1094 does not specify the criteria, the following is
394 * believed to be compatible with the reference port.
395 * For nfs:
396 * If the file's modify time on the server has changed since the
397 * last read rpc or you have written to the file,
398 * you may have lost data cache consistency with the
399 * server, so flush all of the file's data out of the cache.
400 * Then force a getattr rpc to ensure that you have up to date
401 * attributes.
402 * NB: This implies that cache data can be read when up to
403 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
404 * attributes this could be forced by setting n_attrstamp to 0 before
405 * the VOP_GETATTR() call.
406 */
407 if (np->n_flag & NMODIFIED) {
408 if (vp->v_type != VREG) {
409 if (vp->v_type != VDIR)
410 panic("nfs: bioread, not dir");
411 (nmp->nm_rpcops->nr_invaldir)(vp);
412 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
413 if (error)
414 return (error);
415 }
416 np->n_attrstamp = 0;
417 error = VOP_GETATTR(vp, &vattr, cred, td);
418 if (error)
419 return (error);
420 np->n_mtime = vattr.va_mtime;
421 } else {
422 error = VOP_GETATTR(vp, &vattr, cred, td);
423 if (error)
424 return (error);
425 if ((np->n_flag & NSIZECHANGED)
426 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
427 if (vp->v_type == VDIR)
428 (nmp->nm_rpcops->nr_invaldir)(vp);
429 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
430 if (error)
431 return (error);
432 np->n_mtime = vattr.va_mtime;
433 np->n_flag &= ~NSIZECHANGED;
434 }
435 }
436 do {
437 switch (vp->v_type) {
438 case VREG:
439 nfsstats.biocache_reads++;
440 lbn = uio->uio_offset / biosize;
441 on = uio->uio_offset & (biosize - 1);
442
443 /*
444 * Start the read ahead(s), as required.
445 * The readahead is kicked off only if sequential access
446 * is detected, based on the readahead hint (ra_expect_lbn).
447 */
448 if (nmp->nm_readahead > 0 && np->ra_expect_lbn == lbn) {
449 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
450 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
451 rabn = lbn + 1 + nra;
452 if (incore(&vp->v_bufobj, rabn) == NULL) {
453 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
454 if (!rabp) {
455 error = nfs_sigintr(nmp, NULL, td);
456 return (error ? error : EINTR);
457 }
458 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
459 rabp->b_flags |= B_ASYNC;
460 rabp->b_iocmd = BIO_READ;
461 vfs_busy_pages(rabp, 0);
462 if (nfs_asyncio(nmp, rabp, cred, td)) {
463 rabp->b_flags |= B_INVAL;
464 rabp->b_ioflags |= BIO_ERROR;
465 vfs_unbusy_pages(rabp);
466 brelse(rabp);
467 break;
468 }
469 } else {
470 brelse(rabp);
471 }
472 }
473 }
474 np->ra_expect_lbn = lbn + 1;
475 }
476
477 /*
478 * Obtain the buffer cache block. Figure out the buffer size
479 * when we are at EOF. If we are modifying the size of the
480 * buffer based on an EOF condition we need to hold
481 * nfs_rslock() through obtaining the buffer to prevent
482 * a potential writer-appender from messing with n_size.
483 * Otherwise we may accidently truncate the buffer and
484 * lose dirty data.
485 *
486 * Note that bcount is *not* DEV_BSIZE aligned.
487 */
488
489 again:
490 bcount = biosize;
491 if ((off_t)lbn * biosize >= np->n_size) {
492 bcount = 0;
493 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
494 bcount = np->n_size - (off_t)lbn * biosize;
495 }
496 if (bcount != biosize) {
497 switch(nfs_rslock(np, td)) {
498 case ENOLCK:
499 goto again;
500 /* not reached */
501 case EIO:
502 return (EIO);
503 case EINTR:
504 case ERESTART:
505 return(EINTR);
506 /* not reached */
507 default:
508 break;
509 }
510 }
511
512 bp = nfs_getcacheblk(vp, lbn, bcount, td);
513
514 if (bcount != biosize)
515 nfs_rsunlock(np, td);
516 if (!bp) {
517 error = nfs_sigintr(nmp, NULL, td);
518 return (error ? error : EINTR);
519 }
520
521 /*
522 * If B_CACHE is not set, we must issue the read. If this
523 * fails, we return an error.
524 */
525
526 if ((bp->b_flags & B_CACHE) == 0) {
527 bp->b_iocmd = BIO_READ;
528 vfs_busy_pages(bp, 0);
529 error = nfs_doio(vp, bp, cred, td);
530 if (error) {
531 brelse(bp);
532 return (error);
533 }
534 }
535
536 /*
537 * on is the offset into the current bp. Figure out how many
538 * bytes we can copy out of the bp. Note that bcount is
539 * NOT DEV_BSIZE aligned.
540 *
541 * Then figure out how many bytes we can copy into the uio.
542 */
543
544 n = 0;
545 if (on < bcount)
546 n = min((unsigned)(bcount - on), uio->uio_resid);
547 break;
548 case VLNK:
549 nfsstats.biocache_readlinks++;
550 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
551 if (!bp) {
552 error = nfs_sigintr(nmp, NULL, td);
553 return (error ? error : EINTR);
554 }
555 if ((bp->b_flags & B_CACHE) == 0) {
556 bp->b_iocmd = BIO_READ;
557 vfs_busy_pages(bp, 0);
558 error = nfs_doio(vp, bp, cred, td);
559 if (error) {
560 bp->b_ioflags |= BIO_ERROR;
561 brelse(bp);
562 return (error);
563 }
564 }
565 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
566 on = 0;
567 break;
568 case VDIR:
569 nfsstats.biocache_readdirs++;
570 if (np->n_direofoffset
571 && uio->uio_offset >= np->n_direofoffset) {
572 return (0);
573 }
574 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
575 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
576 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
577 if (!bp) {
578 error = nfs_sigintr(nmp, NULL, td);
579 return (error ? error : EINTR);
580 }
581 if ((bp->b_flags & B_CACHE) == 0) {
582 bp->b_iocmd = BIO_READ;
583 vfs_busy_pages(bp, 0);
584 error = nfs_doio(vp, bp, cred, td);
585 if (error) {
586 brelse(bp);
587 }
588 while (error == NFSERR_BAD_COOKIE) {
589 (nmp->nm_rpcops->nr_invaldir)(vp);
590 error = nfs_vinvalbuf(vp, 0, td, 1);
591 /*
592 * Yuck! The directory has been modified on the
593 * server. The only way to get the block is by
594 * reading from the beginning to get all the
595 * offset cookies.
596 *
597 * Leave the last bp intact unless there is an error.
598 * Loop back up to the while if the error is another
599 * NFSERR_BAD_COOKIE (double yuch!).
600 */
601 for (i = 0; i <= lbn && !error; i++) {
602 if (np->n_direofoffset
603 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
604 return (0);
605 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
606 if (!bp) {
607 error = nfs_sigintr(nmp, NULL, td);
608 return (error ? error : EINTR);
609 }
610 if ((bp->b_flags & B_CACHE) == 0) {
611 bp->b_iocmd = BIO_READ;
612 vfs_busy_pages(bp, 0);
613 error = nfs_doio(vp, bp, cred, td);
614 /*
615 * no error + B_INVAL == directory EOF,
616 * use the block.
617 */
618 if (error == 0 && (bp->b_flags & B_INVAL))
619 break;
620 }
621 /*
622 * An error will throw away the block and the
623 * for loop will break out. If no error and this
624 * is not the block we want, we throw away the
625 * block and go for the next one via the for loop.
626 */
627 if (error || i < lbn)
628 brelse(bp);
629 }
630 }
631 /*
632 * The above while is repeated if we hit another cookie
633 * error. If we hit an error and it wasn't a cookie error,
634 * we give up.
635 */
636 if (error)
637 return (error);
638 }
639
640 /*
641 * If not eof and read aheads are enabled, start one.
642 * (You need the current block first, so that you have the
643 * directory offset cookie of the next block.)
644 */
645 if (nmp->nm_readahead > 0 &&
646 (bp->b_flags & B_INVAL) == 0 &&
647 (np->n_direofoffset == 0 ||
648 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
649 incore(&vp->v_bufobj, lbn + 1) == NULL) {
650 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
651 if (rabp) {
652 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
653 rabp->b_flags |= B_ASYNC;
654 rabp->b_iocmd = BIO_READ;
655 vfs_busy_pages(rabp, 0);
656 if (nfs_asyncio(nmp, rabp, cred, td)) {
657 rabp->b_flags |= B_INVAL;
658 rabp->b_ioflags |= BIO_ERROR;
659 vfs_unbusy_pages(rabp);
660 brelse(rabp);
661 }
662 } else {
663 brelse(rabp);
664 }
665 }
666 }
667 /*
668 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
669 * chopped for the EOF condition, we cannot tell how large
670 * NFS directories are going to be until we hit EOF. So
671 * an NFS directory buffer is *not* chopped to its EOF. Now,
672 * it just so happens that b_resid will effectively chop it
673 * to EOF. *BUT* this information is lost if the buffer goes
674 * away and is reconstituted into a B_CACHE state ( due to
675 * being VMIO ) later. So we keep track of the directory eof
676 * in np->n_direofoffset and chop it off as an extra step
677 * right here.
678 */
679 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
680 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
681 n = np->n_direofoffset - uio->uio_offset;
682 break;
683 default:
684 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
685 bp = NULL;
686 break;
687 };
688
689 if (n > 0) {
690 error = uiomove(bp->b_data + on, (int)n, uio);
691 }
692 if (vp->v_type == VLNK)
693 n = 0;
694 if (bp != NULL)
695 brelse(bp);
696 } while (error == 0 && uio->uio_resid > 0 && n > 0);
697 return (error);
698 }
699
700 /*
701 * The NFS write path cannot handle iovecs with len > 1. So we need to
702 * break up iovecs accordingly (restricting them to wsize).
703 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
704 * For the ASYNC case, 2 copies are needed. The first a copy from the
705 * user buffer to a staging buffer and then a second copy from the staging
706 * buffer to mbufs. This can be optimized by copying from the user buffer
707 * directly into mbufs and passing the chain down, but that requires a
708 * fair amount of re-working of the relevant codepaths (and can be done
709 * later).
710 */
711 static int
712 nfs_directio_write(vp, uiop, cred, ioflag)
713 struct vnode *vp;
714 struct uio *uiop;
715 struct ucred *cred;
716 int ioflag;
717 {
718 int error;
719 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
720 struct thread *td = uiop->uio_td;
721 int size;
722
723 if (ioflag & IO_SYNC) {
724 int iomode, must_commit;
725 struct uio uio;
726 struct iovec iov;
727 do_sync:
728 while (uiop->uio_resid > 0) {
729 size = min(uiop->uio_resid, nmp->nm_wsize);
730 size = min(uiop->uio_iov->iov_len, size);
731 iov.iov_base = uiop->uio_iov->iov_base;
732 iov.iov_len = size;
733 uio.uio_iov = &iov;
734 uio.uio_iovcnt = 1;
735 uio.uio_offset = uiop->uio_offset;
736 uio.uio_resid = size;
737 uio.uio_segflg = UIO_USERSPACE;
738 uio.uio_rw = UIO_WRITE;
739 uio.uio_td = td;
740 iomode = NFSV3WRITE_FILESYNC;
741 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
742 &iomode, &must_commit);
743 KASSERT((must_commit == 0),
744 ("nfs_directio_write: Did not commit write"));
745 if (error)
746 return (error);
747 uiop->uio_offset += size;
748 uiop->uio_resid -= size;
749 if (uiop->uio_iov->iov_len <= size) {
750 uiop->uio_iovcnt--;
751 uiop->uio_iov++;
752 } else {
753 uiop->uio_iov->iov_base =
754 (char *)uiop->uio_iov->iov_base + size;
755 uiop->uio_iov->iov_len -= size;
756 }
757 }
758 } else {
759 struct uio *t_uio;
760 struct iovec *t_iov;
761 struct buf *bp;
762
763 /*
764 * Break up the write into blocksize chunks and hand these
765 * over to nfsiod's for write back.
766 * Unfortunately, this incurs a copy of the data. Since
767 * the user could modify the buffer before the write is
768 * initiated.
769 *
770 * The obvious optimization here is that one of the 2 copies
771 * in the async write path can be eliminated by copying the
772 * data here directly into mbufs and passing the mbuf chain
773 * down. But that will require a fair amount of re-working
774 * of the code and can be done if there's enough interest
775 * in NFS directio access.
776 */
777 while (uiop->uio_resid > 0) {
778 size = min(uiop->uio_resid, nmp->nm_wsize);
779 size = min(uiop->uio_iov->iov_len, size);
780 bp = getpbuf(&nfs_pbuf_freecnt);
781 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
782 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
783 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
784 t_iov->iov_len = size;
785 t_uio->uio_iov = t_iov;
786 t_uio->uio_iovcnt = 1;
787 t_uio->uio_offset = uiop->uio_offset;
788 t_uio->uio_resid = size;
789 t_uio->uio_segflg = UIO_SYSSPACE;
790 t_uio->uio_rw = UIO_WRITE;
791 t_uio->uio_td = td;
792 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size);
793 bp->b_flags |= B_DIRECT;
794 bp->b_iocmd = BIO_WRITE;
795 if (cred != NOCRED) {
796 crhold(cred);
797 bp->b_wcred = cred;
798 } else
799 bp->b_wcred = NOCRED;
800 bp->b_caller1 = (void *)t_uio;
801 bp->b_vp = vp;
802 vhold(vp);
803 error = nfs_asyncio(nmp, bp, NOCRED, td);
804 if (error) {
805 free(t_iov->iov_base, M_NFSDIRECTIO);
806 free(t_iov, M_NFSDIRECTIO);
807 free(t_uio, M_NFSDIRECTIO);
808 vdrop(bp->b_vp);
809 bp->b_vp = NULL;
810 relpbuf(bp, &nfs_pbuf_freecnt);
811 if (error == EINTR)
812 return (error);
813 goto do_sync;
814 }
815 uiop->uio_offset += size;
816 uiop->uio_resid -= size;
817 if (uiop->uio_iov->iov_len <= size) {
818 uiop->uio_iovcnt--;
819 uiop->uio_iov++;
820 } else {
821 uiop->uio_iov->iov_base =
822 (char *)uiop->uio_iov->iov_base + size;
823 uiop->uio_iov->iov_len -= size;
824 }
825 }
826 }
827 return (0);
828 }
829
830 /*
831 * Vnode op for write using bio
832 */
833 int
834 nfs_write(struct vop_write_args *ap)
835 {
836 int biosize;
837 struct uio *uio = ap->a_uio;
838 struct thread *td = uio->uio_td;
839 struct vnode *vp = ap->a_vp;
840 struct nfsnode *np = VTONFS(vp);
841 struct ucred *cred = ap->a_cred;
842 int ioflag = ap->a_ioflag;
843 struct buf *bp;
844 struct vattr vattr;
845 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
846 daddr_t lbn;
847 int bcount;
848 int n, on, error = 0;
849 int haverslock = 0;
850 struct proc *p = td?td->td_proc:NULL;
851
852 GIANT_REQUIRED;
853
854 #ifdef DIAGNOSTIC
855 if (uio->uio_rw != UIO_WRITE)
856 panic("nfs_write mode");
857 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
858 panic("nfs_write proc");
859 #endif
860 if (vp->v_type != VREG)
861 return (EIO);
862 if (np->n_flag & NWRITEERR) {
863 np->n_flag &= ~NWRITEERR;
864 return (np->n_error);
865 }
866 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
867 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
868 (void)nfs_fsinfo(nmp, vp, cred, td);
869
870 /*
871 * Synchronously flush pending buffers if we are in synchronous
872 * mode or if we are appending.
873 */
874 if (ioflag & (IO_APPEND | IO_SYNC)) {
875 if (np->n_flag & NMODIFIED) {
876 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
877 /*
878 * Require non-blocking, synchronous writes to
879 * dirty files to inform the program it needs
880 * to fsync(2) explicitly.
881 */
882 if (ioflag & IO_NDELAY)
883 return (EAGAIN);
884 #endif
885 flush_and_restart:
886 np->n_attrstamp = 0;
887 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
888 if (error)
889 return (error);
890 }
891 }
892
893 /*
894 * If IO_APPEND then load uio_offset. We restart here if we cannot
895 * get the append lock.
896 */
897 restart:
898 if (ioflag & IO_APPEND) {
899 np->n_attrstamp = 0;
900 error = VOP_GETATTR(vp, &vattr, cred, td);
901 if (error)
902 return (error);
903 uio->uio_offset = np->n_size;
904 }
905
906 if (uio->uio_offset < 0)
907 return (EINVAL);
908 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
909 return (EFBIG);
910 if (uio->uio_resid == 0)
911 return (0);
912
913 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
914 return nfs_directio_write(vp, uio, cred, ioflag);
915
916 /*
917 * We need to obtain the rslock if we intend to modify np->n_size
918 * in order to guarentee the append point with multiple contending
919 * writers, to guarentee that no other appenders modify n_size
920 * while we are trying to obtain a truncated buffer (i.e. to avoid
921 * accidently truncating data written by another appender due to
922 * the race), and to ensure that the buffer is populated prior to
923 * our extending of the file. We hold rslock through the entire
924 * operation.
925 *
926 * Note that we do not synchronize the case where someone truncates
927 * the file while we are appending to it because attempting to lock
928 * this case may deadlock other parts of the system unexpectedly.
929 */
930 if ((ioflag & IO_APPEND) ||
931 uio->uio_offset + uio->uio_resid > np->n_size) {
932 switch(nfs_rslock(np, td)) {
933 case ENOLCK:
934 goto restart;
935 /* not reached */
936 case EIO:
937 return (EIO);
938 case EINTR:
939 case ERESTART:
940 return(EINTR);
941 /* not reached */
942 default:
943 break;
944 }
945 haverslock = 1;
946 }
947
948 /*
949 * Maybe this should be above the vnode op call, but so long as
950 * file servers have no limits, i don't think it matters
951 */
952 if (p != NULL) {
953 PROC_LOCK(p);
954 if (uio->uio_offset + uio->uio_resid >
955 lim_cur(p, RLIMIT_FSIZE)) {
956 psignal(p, SIGXFSZ);
957 PROC_UNLOCK(p);
958 if (haverslock)
959 nfs_rsunlock(np, td);
960 return (EFBIG);
961 }
962 PROC_UNLOCK(p);
963 }
964
965 biosize = vp->v_mount->mnt_stat.f_iosize;
966 /*
967 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
968 * would exceed the local maximum per-file write commit size when
969 * combined with those, we must decide whether to flush,
970 * go synchronous, or return error. We don't bother checking
971 * IO_UNIT -- we just make all writes atomic anyway, as there's
972 * no point optimizing for something that really won't ever happen.
973 */
974 if (!(ioflag & IO_SYNC)) {
975 int needrestart = 0;
976 if (nmp->nm_wcommitsize < uio->uio_resid) {
977 /*
978 * If this request could not possibly be completed
979 * without exceeding the maximum outstanding write
980 * commit size, see if we can convert it into a
981 * synchronous write operation.
982 */
983 if (ioflag & IO_NDELAY)
984 return (EAGAIN);
985 ioflag |= IO_SYNC;
986 if (np->n_flag & NMODIFIED)
987 needrestart = 1;
988 } else if (np->n_flag & NMODIFIED) {
989 int wouldcommit = 0;
990 BO_LOCK(&vp->v_bufobj);
991 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
992 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
993 b_bobufs) {
994 if (bp->b_flags & B_NEEDCOMMIT)
995 wouldcommit += bp->b_bcount;
996 }
997 }
998 BO_UNLOCK(&vp->v_bufobj);
999 /*
1000 * Since we're not operating synchronously and
1001 * bypassing the buffer cache, we are in a commit
1002 * and holding all of these buffers whether
1003 * transmitted or not. If not limited, this
1004 * will lead to the buffer cache deadlocking,
1005 * as no one else can flush our uncommitted buffers.
1006 */
1007 wouldcommit += uio->uio_resid;
1008 /*
1009 * If we would initially exceed the maximum
1010 * outstanding write commit size, flush and restart.
1011 */
1012 if (wouldcommit > nmp->nm_wcommitsize)
1013 needrestart = 1;
1014 }
1015 if (needrestart) {
1016 if (haverslock) {
1017 nfs_rsunlock(np, td);
1018 haverslock = 0;
1019 }
1020 goto flush_and_restart;
1021 }
1022 }
1023
1024 do {
1025 nfsstats.biocache_writes++;
1026 lbn = uio->uio_offset / biosize;
1027 on = uio->uio_offset & (biosize-1);
1028 n = min((unsigned)(biosize - on), uio->uio_resid);
1029 again:
1030 /*
1031 * Handle direct append and file extension cases, calculate
1032 * unaligned buffer size.
1033 */
1034
1035 if (uio->uio_offset == np->n_size && n) {
1036 /*
1037 * Get the buffer (in its pre-append state to maintain
1038 * B_CACHE if it was previously set). Resize the
1039 * nfsnode after we have locked the buffer to prevent
1040 * readers from reading garbage.
1041 */
1042 bcount = on;
1043 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1044
1045 if (bp != NULL) {
1046 long save;
1047
1048 np->n_size = uio->uio_offset + n;
1049 np->n_flag |= NMODIFIED;
1050 vnode_pager_setsize(vp, np->n_size);
1051
1052 save = bp->b_flags & B_CACHE;
1053 bcount += n;
1054 allocbuf(bp, bcount);
1055 bp->b_flags |= save;
1056 }
1057 } else {
1058 /*
1059 * Obtain the locked cache block first, and then
1060 * adjust the file's size as appropriate.
1061 */
1062 bcount = on + n;
1063 if ((off_t)lbn * biosize + bcount < np->n_size) {
1064 if ((off_t)(lbn + 1) * biosize < np->n_size)
1065 bcount = biosize;
1066 else
1067 bcount = np->n_size - (off_t)lbn * biosize;
1068 }
1069 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1070 if (uio->uio_offset + n > np->n_size) {
1071 np->n_size = uio->uio_offset + n;
1072 np->n_flag |= NMODIFIED;
1073 vnode_pager_setsize(vp, np->n_size);
1074 }
1075 }
1076
1077 if (!bp) {
1078 error = nfs_sigintr(nmp, NULL, td);
1079 if (!error)
1080 error = EINTR;
1081 break;
1082 }
1083
1084 /*
1085 * Issue a READ if B_CACHE is not set. In special-append
1086 * mode, B_CACHE is based on the buffer prior to the write
1087 * op and is typically set, avoiding the read. If a read
1088 * is required in special append mode, the server will
1089 * probably send us a short-read since we extended the file
1090 * on our end, resulting in b_resid == 0 and, thusly,
1091 * B_CACHE getting set.
1092 *
1093 * We can also avoid issuing the read if the write covers
1094 * the entire buffer. We have to make sure the buffer state
1095 * is reasonable in this case since we will not be initiating
1096 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1097 * more information.
1098 *
1099 * B_CACHE may also be set due to the buffer being cached
1100 * normally.
1101 */
1102
1103 if (on == 0 && n == bcount) {
1104 bp->b_flags |= B_CACHE;
1105 bp->b_flags &= ~B_INVAL;
1106 bp->b_ioflags &= ~BIO_ERROR;
1107 }
1108
1109 if ((bp->b_flags & B_CACHE) == 0) {
1110 bp->b_iocmd = BIO_READ;
1111 vfs_busy_pages(bp, 0);
1112 error = nfs_doio(vp, bp, cred, td);
1113 if (error) {
1114 brelse(bp);
1115 break;
1116 }
1117 }
1118 if (bp->b_wcred == NOCRED)
1119 bp->b_wcred = crhold(cred);
1120 np->n_flag |= NMODIFIED;
1121
1122 /*
1123 * If dirtyend exceeds file size, chop it down. This should
1124 * not normally occur but there is an append race where it
1125 * might occur XXX, so we log it.
1126 *
1127 * If the chopping creates a reverse-indexed or degenerate
1128 * situation with dirtyoff/end, we 0 both of them.
1129 */
1130
1131 if (bp->b_dirtyend > bcount) {
1132 printf("NFS append race @%lx:%d\n",
1133 (long)bp->b_blkno * DEV_BSIZE,
1134 bp->b_dirtyend - bcount);
1135 bp->b_dirtyend = bcount;
1136 }
1137
1138 if (bp->b_dirtyoff >= bp->b_dirtyend)
1139 bp->b_dirtyoff = bp->b_dirtyend = 0;
1140
1141 /*
1142 * If the new write will leave a contiguous dirty
1143 * area, just update the b_dirtyoff and b_dirtyend,
1144 * otherwise force a write rpc of the old dirty area.
1145 *
1146 * While it is possible to merge discontiguous writes due to
1147 * our having a B_CACHE buffer ( and thus valid read data
1148 * for the hole), we don't because it could lead to
1149 * significant cache coherency problems with multiple clients,
1150 * especially if locking is implemented later on.
1151 *
1152 * as an optimization we could theoretically maintain
1153 * a linked list of discontinuous areas, but we would still
1154 * have to commit them separately so there isn't much
1155 * advantage to it except perhaps a bit of asynchronization.
1156 */
1157
1158 if (bp->b_dirtyend > 0 &&
1159 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1160 if (bwrite(bp) == EINTR) {
1161 error = EINTR;
1162 break;
1163 }
1164 goto again;
1165 }
1166
1167 error = uiomove((char *)bp->b_data + on, n, uio);
1168
1169 /*
1170 * Since this block is being modified, it must be written
1171 * again and not just committed. Since write clustering does
1172 * not work for the stage 1 data write, only the stage 2
1173 * commit rpc, we have to clear B_CLUSTEROK as well.
1174 */
1175 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1176
1177 if (error) {
1178 bp->b_ioflags |= BIO_ERROR;
1179 brelse(bp);
1180 break;
1181 }
1182
1183 /*
1184 * Only update dirtyoff/dirtyend if not a degenerate
1185 * condition.
1186 */
1187 if (n) {
1188 if (bp->b_dirtyend > 0) {
1189 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1190 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1191 } else {
1192 bp->b_dirtyoff = on;
1193 bp->b_dirtyend = on + n;
1194 }
1195 vfs_bio_set_validclean(bp, on, n);
1196 }
1197
1198 /*
1199 * If IO_SYNC do bwrite().
1200 *
1201 * IO_INVAL appears to be unused. The idea appears to be
1202 * to turn off caching in this case. Very odd. XXX
1203 */
1204 if ((ioflag & IO_SYNC)) {
1205 if (ioflag & IO_INVAL)
1206 bp->b_flags |= B_NOCACHE;
1207 error = bwrite(bp);
1208 if (error)
1209 break;
1210 } else if ((n + on) == biosize) {
1211 bp->b_flags |= B_ASYNC;
1212 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, 0);
1213 } else {
1214 bdwrite(bp);
1215 }
1216 } while (uio->uio_resid > 0 && n > 0);
1217
1218 if (haverslock)
1219 nfs_rsunlock(np, td);
1220
1221 return (error);
1222 }
1223
1224 /*
1225 * Get an nfs cache block.
1226 *
1227 * Allocate a new one if the block isn't currently in the cache
1228 * and return the block marked busy. If the calling process is
1229 * interrupted by a signal for an interruptible mount point, return
1230 * NULL.
1231 *
1232 * The caller must carefully deal with the possible B_INVAL state of
1233 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1234 * indirectly), so synchronous reads can be issued without worrying about
1235 * the B_INVAL state. We have to be a little more careful when dealing
1236 * with writes (see comments in nfs_write()) when extending a file past
1237 * its EOF.
1238 */
1239 static struct buf *
1240 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1241 {
1242 struct buf *bp;
1243 struct mount *mp;
1244 struct nfsmount *nmp;
1245
1246 mp = vp->v_mount;
1247 nmp = VFSTONFS(mp);
1248
1249 if (nmp->nm_flag & NFSMNT_INT) {
1250 sigset_t oldset;
1251
1252 nfs_set_sigmask(td, &oldset);
1253 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1254 nfs_restore_sigmask(td, &oldset);
1255 while (bp == NULL) {
1256 if (nfs_sigintr(nmp, NULL, td))
1257 return (NULL);
1258 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1259 }
1260 } else {
1261 bp = getblk(vp, bn, size, 0, 0, 0);
1262 }
1263
1264 if (vp->v_type == VREG) {
1265 int biosize;
1266
1267 biosize = mp->mnt_stat.f_iosize;
1268 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1269 }
1270 return (bp);
1271 }
1272
1273 /*
1274 * Flush and invalidate all dirty buffers. If another process is already
1275 * doing the flush, just wait for completion.
1276 */
1277 int
1278 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1279 {
1280 struct nfsnode *np = VTONFS(vp);
1281 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1282 int error = 0, slpflag, slptimeo;
1283 int old_lock = 0;
1284
1285 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1286
1287 /*
1288 * XXX This check stops us from needlessly doing a vinvalbuf when
1289 * being called through vclean(). It is not clear that this is
1290 * unsafe.
1291 */
1292 if (vp->v_iflag & VI_DOOMED)
1293 return (0);
1294
1295 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1296 intrflg = 0;
1297 if (intrflg) {
1298 slpflag = PCATCH;
1299 slptimeo = 2 * hz;
1300 } else {
1301 slpflag = 0;
1302 slptimeo = 0;
1303 }
1304
1305 if ((old_lock = VOP_ISLOCKED(vp, td)) != LK_EXCLUSIVE) {
1306 if (old_lock == LK_SHARED) {
1307 /* Upgrade to exclusive lock, this might block */
1308 vn_lock(vp, LK_UPGRADE | LK_RETRY, td);
1309 } else {
1310 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1311 }
1312 }
1313
1314 /*
1315 * Now, flush as required.
1316 */
1317 error = vinvalbuf(vp, flags, td, slpflag, 0);
1318 while (error) {
1319 if (intrflg && (error = nfs_sigintr(nmp, NULL, td)))
1320 goto out;
1321 error = vinvalbuf(vp, flags, td, 0, slptimeo);
1322 }
1323 np->n_flag &= ~NMODIFIED;
1324 out:
1325 if (old_lock != LK_EXCLUSIVE) {
1326 if (old_lock == LK_SHARED) {
1327 /* Downgrade from exclusive lock, this might block */
1328 vn_lock(vp, LK_DOWNGRADE, td);
1329 } else {
1330 VOP_UNLOCK(vp, 0, td);
1331 }
1332 }
1333 return error;
1334 }
1335
1336 /*
1337 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1338 * This is mainly to avoid queueing async I/O requests when the nfsiods
1339 * are all hung on a dead server.
1340 *
1341 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1342 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1343 */
1344 int
1345 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1346 {
1347 int iod;
1348 int gotiod;
1349 int slpflag = 0;
1350 int slptimeo = 0;
1351 int error, error2;
1352
1353 /*
1354 * Commits are usually short and sweet so lets save some cpu and
1355 * leave the async daemons for more important rpc's (such as reads
1356 * and writes).
1357 */
1358 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1359 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1360 return(EIO);
1361 }
1362
1363 again:
1364 if (nmp->nm_flag & NFSMNT_INT)
1365 slpflag = PCATCH;
1366 gotiod = FALSE;
1367
1368 /*
1369 * Find a free iod to process this request.
1370 */
1371 for (iod = 0; iod < nfs_numasync; iod++)
1372 if (nfs_iodwant[iod]) {
1373 gotiod = TRUE;
1374 break;
1375 }
1376
1377 /*
1378 * Try to create one if none are free.
1379 */
1380 if (!gotiod) {
1381 iod = nfs_nfsiodnew();
1382 if (iod != -1)
1383 gotiod = TRUE;
1384 }
1385
1386 if (gotiod) {
1387 /*
1388 * Found one, so wake it up and tell it which
1389 * mount to process.
1390 */
1391 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1392 iod, nmp));
1393 nfs_iodwant[iod] = NULL;
1394 nfs_iodmount[iod] = nmp;
1395 nmp->nm_bufqiods++;
1396 wakeup(&nfs_iodwant[iod]);
1397 }
1398
1399 /*
1400 * If none are free, we may already have an iod working on this mount
1401 * point. If so, it will process our request.
1402 */
1403 if (!gotiod) {
1404 if (nmp->nm_bufqiods > 0) {
1405 NFS_DPF(ASYNCIO,
1406 ("nfs_asyncio: %d iods are already processing mount %p\n",
1407 nmp->nm_bufqiods, nmp));
1408 gotiod = TRUE;
1409 }
1410 }
1411
1412 /*
1413 * If we have an iod which can process the request, then queue
1414 * the buffer.
1415 */
1416 if (gotiod) {
1417 /*
1418 * Ensure that the queue never grows too large. We still want
1419 * to asynchronize so we block rather then return EIO.
1420 */
1421 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1422 NFS_DPF(ASYNCIO,
1423 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1424 nmp->nm_bufqwant = TRUE;
1425 error = nfs_tsleep(td, &nmp->nm_bufq, slpflag | PRIBIO,
1426 "nfsaio", slptimeo);
1427 if (error) {
1428 error2 = nfs_sigintr(nmp, NULL, td);
1429 if (error2)
1430 return (error2);
1431 if (slpflag == PCATCH) {
1432 slpflag = 0;
1433 slptimeo = 2 * hz;
1434 }
1435 }
1436 /*
1437 * We might have lost our iod while sleeping,
1438 * so check and loop if nescessary.
1439 */
1440 if (nmp->nm_bufqiods == 0) {
1441 NFS_DPF(ASYNCIO,
1442 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1443 goto again;
1444 }
1445 }
1446
1447 if (bp->b_iocmd == BIO_READ) {
1448 if (bp->b_rcred == NOCRED && cred != NOCRED)
1449 bp->b_rcred = crhold(cred);
1450 } else {
1451 if (bp->b_wcred == NOCRED && cred != NOCRED)
1452 bp->b_wcred = crhold(cred);
1453 }
1454
1455 if (bp->b_flags & B_REMFREE)
1456 bremfreef(bp);
1457 BUF_KERNPROC(bp);
1458 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1459 nmp->nm_bufqlen++;
1460 return (0);
1461 }
1462
1463 /*
1464 * All the iods are busy on other mounts, so return EIO to
1465 * force the caller to process the i/o synchronously.
1466 */
1467 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1468 return (EIO);
1469 }
1470
1471 void
1472 nfs_doio_directwrite(struct buf *bp)
1473 {
1474 int iomode, must_commit;
1475 struct uio *uiop = (struct uio *)bp->b_caller1;
1476 char *iov_base = uiop->uio_iov->iov_base;
1477 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1478
1479 iomode = NFSV3WRITE_FILESYNC;
1480 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1481 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1482 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1483 free(iov_base, M_NFSDIRECTIO);
1484 free(uiop->uio_iov, M_NFSDIRECTIO);
1485 free(uiop, M_NFSDIRECTIO);
1486 vdrop(bp->b_vp);
1487 bp->b_vp = NULL;
1488 relpbuf(bp, &nfs_pbuf_freecnt);
1489 }
1490
1491 /*
1492 * Do an I/O operation to/from a cache block. This may be called
1493 * synchronously or from an nfsiod.
1494 */
1495 int
1496 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1497 {
1498 struct uio *uiop;
1499 struct nfsnode *np;
1500 struct nfsmount *nmp;
1501 int error = 0, iomode, must_commit = 0;
1502 struct uio uio;
1503 struct iovec io;
1504 struct proc *p = td ? td->td_proc : NULL;
1505
1506 np = VTONFS(vp);
1507 nmp = VFSTONFS(vp->v_mount);
1508 uiop = &uio;
1509 uiop->uio_iov = &io;
1510 uiop->uio_iovcnt = 1;
1511 uiop->uio_segflg = UIO_SYSSPACE;
1512 uiop->uio_td = td;
1513
1514 /*
1515 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1516 * do this here so we do not have to do it in all the code that
1517 * calls us.
1518 */
1519 bp->b_flags &= ~B_INVAL;
1520 bp->b_ioflags &= ~BIO_ERROR;
1521
1522 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1523
1524 if (bp->b_iocmd == BIO_READ) {
1525 io.iov_len = uiop->uio_resid = bp->b_bcount;
1526 io.iov_base = bp->b_data;
1527 uiop->uio_rw = UIO_READ;
1528
1529 switch (vp->v_type) {
1530 case VREG:
1531 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1532 nfsstats.read_bios++;
1533 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1534
1535 if (!error) {
1536 if (uiop->uio_resid) {
1537 /*
1538 * If we had a short read with no error, we must have
1539 * hit a file hole. We should zero-fill the remainder.
1540 * This can also occur if the server hits the file EOF.
1541 *
1542 * Holes used to be able to occur due to pending
1543 * writes, but that is not possible any longer.
1544 */
1545 int nread = bp->b_bcount - uiop->uio_resid;
1546 int left = uiop->uio_resid;
1547
1548 if (left > 0)
1549 bzero((char *)bp->b_data + nread, left);
1550 uiop->uio_resid = 0;
1551 }
1552 }
1553 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1554 if (p && (vp->v_vflag & VV_TEXT) &&
1555 (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime))) {
1556 PROC_LOCK(p);
1557 killproc(p, "text file modification");
1558 PROC_UNLOCK(p);
1559 }
1560 break;
1561 case VLNK:
1562 uiop->uio_offset = (off_t)0;
1563 nfsstats.readlink_bios++;
1564 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1565 break;
1566 case VDIR:
1567 nfsstats.readdir_bios++;
1568 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1569 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0)
1570 error = nfs4_readdirrpc(vp, uiop, cr);
1571 else {
1572 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1573 error = nfs_readdirplusrpc(vp, uiop, cr);
1574 if (error == NFSERR_NOTSUPP)
1575 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1576 }
1577 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1578 error = nfs_readdirrpc(vp, uiop, cr);
1579 }
1580 /*
1581 * end-of-directory sets B_INVAL but does not generate an
1582 * error.
1583 */
1584 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1585 bp->b_flags |= B_INVAL;
1586 break;
1587 default:
1588 printf("nfs_doio: type %x unexpected\n", vp->v_type);
1589 break;
1590 };
1591 if (error) {
1592 bp->b_ioflags |= BIO_ERROR;
1593 bp->b_error = error;
1594 }
1595 } else {
1596 /*
1597 * If we only need to commit, try to commit
1598 */
1599 if (bp->b_flags & B_NEEDCOMMIT) {
1600 int retv;
1601 off_t off;
1602
1603 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1604 retv = (nmp->nm_rpcops->nr_commit)(
1605 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1606 bp->b_wcred, td);
1607 if (retv == 0) {
1608 bp->b_dirtyoff = bp->b_dirtyend = 0;
1609 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1610 bp->b_resid = 0;
1611 bufdone(bp);
1612 return (0);
1613 }
1614 if (retv == NFSERR_STALEWRITEVERF) {
1615 nfs_clearcommit(vp->v_mount);
1616 }
1617 }
1618
1619 /*
1620 * Setup for actual write
1621 */
1622
1623 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1624 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1625
1626 if (bp->b_dirtyend > bp->b_dirtyoff) {
1627 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1628 - bp->b_dirtyoff;
1629 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1630 + bp->b_dirtyoff;
1631 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1632 uiop->uio_rw = UIO_WRITE;
1633 nfsstats.write_bios++;
1634
1635 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1636 iomode = NFSV3WRITE_UNSTABLE;
1637 else
1638 iomode = NFSV3WRITE_FILESYNC;
1639
1640 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1641
1642 /*
1643 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1644 * to cluster the buffers needing commit. This will allow
1645 * the system to submit a single commit rpc for the whole
1646 * cluster. We can do this even if the buffer is not 100%
1647 * dirty (relative to the NFS blocksize), so we optimize the
1648 * append-to-file-case.
1649 *
1650 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1651 * cleared because write clustering only works for commit
1652 * rpc's, not for the data portion of the write).
1653 */
1654
1655 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1656 bp->b_flags |= B_NEEDCOMMIT;
1657 if (bp->b_dirtyoff == 0
1658 && bp->b_dirtyend == bp->b_bcount)
1659 bp->b_flags |= B_CLUSTEROK;
1660 } else {
1661 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1662 }
1663
1664 /*
1665 * For an interrupted write, the buffer is still valid
1666 * and the write hasn't been pushed to the server yet,
1667 * so we can't set BIO_ERROR and report the interruption
1668 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1669 * is not relevant, so the rpc attempt is essentially
1670 * a noop. For the case of a V3 write rpc not being
1671 * committed to stable storage, the block is still
1672 * dirty and requires either a commit rpc or another
1673 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1674 * the block is reused. This is indicated by setting
1675 * the B_DELWRI and B_NEEDCOMMIT flags.
1676 *
1677 * If the buffer is marked B_PAGING, it does not reside on
1678 * the vp's paging queues so we cannot call bdirty(). The
1679 * bp in this case is not an NFS cache block so we should
1680 * be safe. XXX
1681 */
1682 if (error == EINTR || error == EIO || error == ETIMEDOUT
1683 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1684 int s;
1685
1686 s = splbio();
1687 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1688 if ((bp->b_flags & B_PAGING) == 0) {
1689 bdirty(bp);
1690 bp->b_flags &= ~B_DONE;
1691 }
1692 if (error && (bp->b_flags & B_ASYNC) == 0)
1693 bp->b_flags |= B_EINTR;
1694 splx(s);
1695 } else {
1696 if (error) {
1697 bp->b_ioflags |= BIO_ERROR;
1698 bp->b_error = np->n_error = error;
1699 np->n_flag |= NWRITEERR;
1700 }
1701 bp->b_dirtyoff = bp->b_dirtyend = 0;
1702 }
1703 } else {
1704 bp->b_resid = 0;
1705 bufdone(bp);
1706 return (0);
1707 }
1708 }
1709 bp->b_resid = uiop->uio_resid;
1710 if (must_commit)
1711 nfs_clearcommit(vp->v_mount);
1712 bufdone(bp);
1713 return (error);
1714 }
1715
1716 /*
1717 * Used to aid in handling ftruncate() operations on the NFS client side.
1718 * Truncation creates a number of special problems for NFS. We have to
1719 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1720 * we have to properly handle VM pages or (potentially dirty) buffers
1721 * that straddle the truncation point.
1722 */
1723
1724 int
1725 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1726 {
1727 struct nfsnode *np = VTONFS(vp);
1728 u_quad_t tsize = np->n_size;
1729 int biosize = vp->v_mount->mnt_stat.f_iosize;
1730 int error = 0;
1731
1732 np->n_size = nsize;
1733
1734 if (np->n_size < tsize) {
1735 struct buf *bp;
1736 daddr_t lbn;
1737 int bufsize;
1738
1739 /*
1740 * vtruncbuf() doesn't get the buffer overlapping the
1741 * truncation point. We may have a B_DELWRI and/or B_CACHE
1742 * buffer that now needs to be truncated.
1743 */
1744 error = vtruncbuf(vp, cred, td, nsize, biosize);
1745 lbn = nsize / biosize;
1746 bufsize = nsize & (biosize - 1);
1747 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1748 if (!bp)
1749 return EINTR;
1750 if (bp->b_dirtyoff > bp->b_bcount)
1751 bp->b_dirtyoff = bp->b_bcount;
1752 if (bp->b_dirtyend > bp->b_bcount)
1753 bp->b_dirtyend = bp->b_bcount;
1754 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1755 brelse(bp);
1756 } else {
1757 vnode_pager_setsize(vp, nsize);
1758 }
1759 return(error);
1760 }
1761
Cache object: b8890e41abc4ed3c1435c4c61bd9c791
|