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/7.3/sys/nfsclient/nfs_bio.c 172324 2007-09-25 21:08:49Z mohans $");
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 /*
76 * Vnode op for VM getpages.
77 */
78 int
79 nfs_getpages(struct vop_getpages_args *ap)
80 {
81 int i, error, nextoff, size, toff, count, npages;
82 struct uio uio;
83 struct iovec iov;
84 vm_offset_t kva;
85 struct buf *bp;
86 struct vnode *vp;
87 struct thread *td;
88 struct ucred *cred;
89 struct nfsmount *nmp;
90 vm_object_t object;
91 vm_page_t *pages;
92 struct nfsnode *np;
93
94 vp = ap->a_vp;
95 np = VTONFS(vp);
96 td = curthread; /* XXX */
97 cred = curthread->td_ucred; /* XXX */
98 nmp = VFSTONFS(vp->v_mount);
99 pages = ap->a_m;
100 count = ap->a_count;
101
102 if ((object = vp->v_object) == NULL) {
103 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n");
104 return VM_PAGER_ERROR;
105 }
106
107 if (nfs_directio_enable && !nfs_directio_allow_mmap) {
108 mtx_lock(&np->n_mtx);
109 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
110 mtx_unlock(&np->n_mtx);
111 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n");
112 return VM_PAGER_ERROR;
113 } else
114 mtx_unlock(&np->n_mtx);
115 }
116
117 mtx_lock(&nmp->nm_mtx);
118 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
119 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
120 mtx_unlock(&nmp->nm_mtx);
121 /* We'll never get here for v4, because we always have fsinfo */
122 (void)nfs_fsinfo(nmp, vp, cred, td);
123 } else
124 mtx_unlock(&nmp->nm_mtx);
125
126 npages = btoc(count);
127
128 /*
129 * If the requested page is partially valid, just return it and
130 * allow the pager to zero-out the blanks. Partially valid pages
131 * can only occur at the file EOF.
132 */
133
134 {
135 vm_page_t m = pages[ap->a_reqpage];
136
137 VM_OBJECT_LOCK(object);
138 vm_page_lock_queues();
139 if (m->valid != 0) {
140 /* handled by vm_fault now */
141 /* vm_page_zero_invalid(m, TRUE); */
142 for (i = 0; i < npages; ++i) {
143 if (i != ap->a_reqpage)
144 vm_page_free(pages[i]);
145 }
146 vm_page_unlock_queues();
147 VM_OBJECT_UNLOCK(object);
148 return(0);
149 }
150 vm_page_unlock_queues();
151 VM_OBJECT_UNLOCK(object);
152 }
153
154 /*
155 * We use only the kva address for the buffer, but this is extremely
156 * convienient and fast.
157 */
158 bp = getpbuf(&nfs_pbuf_freecnt);
159
160 kva = (vm_offset_t) bp->b_data;
161 pmap_qenter(kva, pages, npages);
162 PCPU_INC(cnt.v_vnodein);
163 PCPU_ADD(cnt.v_vnodepgsin, npages);
164
165 iov.iov_base = (caddr_t) kva;
166 iov.iov_len = count;
167 uio.uio_iov = &iov;
168 uio.uio_iovcnt = 1;
169 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
170 uio.uio_resid = count;
171 uio.uio_segflg = UIO_SYSSPACE;
172 uio.uio_rw = UIO_READ;
173 uio.uio_td = td;
174
175 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
176 pmap_qremove(kva, npages);
177
178 relpbuf(bp, &nfs_pbuf_freecnt);
179
180 if (error && (uio.uio_resid == count)) {
181 nfs_printf("nfs_getpages: error %d\n", error);
182 VM_OBJECT_LOCK(object);
183 vm_page_lock_queues();
184 for (i = 0; i < npages; ++i) {
185 if (i != ap->a_reqpage)
186 vm_page_free(pages[i]);
187 }
188 vm_page_unlock_queues();
189 VM_OBJECT_UNLOCK(object);
190 return VM_PAGER_ERROR;
191 }
192
193 /*
194 * Calculate the number of bytes read and validate only that number
195 * of bytes. Note that due to pending writes, size may be 0. This
196 * does not mean that the remaining data is invalid!
197 */
198
199 size = count - uio.uio_resid;
200 VM_OBJECT_LOCK(object);
201 vm_page_lock_queues();
202 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
203 vm_page_t m;
204 nextoff = toff + PAGE_SIZE;
205 m = pages[i];
206
207 if (nextoff <= size) {
208 /*
209 * Read operation filled an entire page
210 */
211 m->valid = VM_PAGE_BITS_ALL;
212 vm_page_undirty(m);
213 } else if (size > toff) {
214 /*
215 * Read operation filled a partial page.
216 */
217 m->valid = 0;
218 vm_page_set_validclean(m, 0, size - toff);
219 /* handled by vm_fault now */
220 /* vm_page_zero_invalid(m, TRUE); */
221 } else {
222 /*
223 * Read operation was short. If no error occured
224 * we may have hit a zero-fill section. We simply
225 * leave valid set to 0.
226 */
227 ;
228 }
229 if (i != ap->a_reqpage) {
230 /*
231 * Whether or not to leave the page activated is up in
232 * the air, but we should put the page on a page queue
233 * somewhere (it already is in the object). Result:
234 * It appears that emperical results show that
235 * deactivating pages is best.
236 */
237
238 /*
239 * Just in case someone was asking for this page we
240 * now tell them that it is ok to use.
241 */
242 if (!error) {
243 if (m->oflags & VPO_WANTED)
244 vm_page_activate(m);
245 else
246 vm_page_deactivate(m);
247 vm_page_wakeup(m);
248 } else {
249 vm_page_free(m);
250 }
251 }
252 }
253 vm_page_unlock_queues();
254 VM_OBJECT_UNLOCK(object);
255 return 0;
256 }
257
258 /*
259 * Vnode op for VM putpages.
260 */
261 int
262 nfs_putpages(struct vop_putpages_args *ap)
263 {
264 struct uio uio;
265 struct iovec iov;
266 vm_offset_t kva;
267 struct buf *bp;
268 int iomode, must_commit, i, error, npages, count;
269 off_t offset;
270 int *rtvals;
271 struct vnode *vp;
272 struct thread *td;
273 struct ucred *cred;
274 struct nfsmount *nmp;
275 struct nfsnode *np;
276 vm_page_t *pages;
277
278 vp = ap->a_vp;
279 np = VTONFS(vp);
280 td = curthread; /* XXX */
281 cred = curthread->td_ucred; /* XXX */
282 nmp = VFSTONFS(vp->v_mount);
283 pages = ap->a_m;
284 count = ap->a_count;
285 rtvals = ap->a_rtvals;
286 npages = btoc(count);
287 offset = IDX_TO_OFF(pages[0]->pindex);
288
289 mtx_lock(&nmp->nm_mtx);
290 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
291 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
292 mtx_unlock(&nmp->nm_mtx);
293 (void)nfs_fsinfo(nmp, vp, cred, td);
294 } else
295 mtx_unlock(&nmp->nm_mtx);
296
297 mtx_lock(&np->n_mtx);
298 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
299 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
300 mtx_unlock(&np->n_mtx);
301 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
302 mtx_lock(&np->n_mtx);
303 }
304
305 for (i = 0; i < npages; i++)
306 rtvals[i] = VM_PAGER_AGAIN;
307
308 /*
309 * When putting pages, do not extend file past EOF.
310 */
311 if (offset + count > np->n_size) {
312 count = np->n_size - offset;
313 if (count < 0)
314 count = 0;
315 }
316 mtx_unlock(&np->n_mtx);
317
318 /*
319 * We use only the kva address for the buffer, but this is extremely
320 * convienient and fast.
321 */
322 bp = getpbuf(&nfs_pbuf_freecnt);
323
324 kva = (vm_offset_t) bp->b_data;
325 pmap_qenter(kva, pages, npages);
326 PCPU_INC(cnt.v_vnodeout);
327 PCPU_ADD(cnt.v_vnodepgsout, count);
328
329 iov.iov_base = (caddr_t) kva;
330 iov.iov_len = count;
331 uio.uio_iov = &iov;
332 uio.uio_iovcnt = 1;
333 uio.uio_offset = offset;
334 uio.uio_resid = count;
335 uio.uio_segflg = UIO_SYSSPACE;
336 uio.uio_rw = UIO_WRITE;
337 uio.uio_td = td;
338
339 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
340 iomode = NFSV3WRITE_UNSTABLE;
341 else
342 iomode = NFSV3WRITE_FILESYNC;
343
344 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
345
346 pmap_qremove(kva, npages);
347 relpbuf(bp, &nfs_pbuf_freecnt);
348
349 if (!error) {
350 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
351 for (i = 0; i < nwritten; i++) {
352 rtvals[i] = VM_PAGER_OK;
353 vm_page_undirty(pages[i]);
354 }
355 if (must_commit) {
356 nfs_clearcommit(vp->v_mount);
357 }
358 }
359 return rtvals[0];
360 }
361
362 /*
363 * For nfs, cache consistency can only be maintained approximately.
364 * Although RFC1094 does not specify the criteria, the following is
365 * believed to be compatible with the reference port.
366 * For nfs:
367 * If the file's modify time on the server has changed since the
368 * last read rpc or you have written to the file,
369 * you may have lost data cache consistency with the
370 * server, so flush all of the file's data out of the cache.
371 * Then force a getattr rpc to ensure that you have up to date
372 * attributes.
373 * NB: This implies that cache data can be read when up to
374 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
375 * attributes this could be forced by setting n_attrstamp to 0 before
376 * the VOP_GETATTR() call.
377 */
378 static inline int
379 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
380 {
381 int error = 0;
382 struct vattr vattr;
383 struct nfsnode *np = VTONFS(vp);
384 int old_lock;
385 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
386
387 /*
388 * Grab the exclusive lock before checking whether the cache is
389 * consistent.
390 * XXX - We can make this cheaper later (by acquiring cheaper locks).
391 * But for now, this suffices.
392 */
393 old_lock = nfs_upgrade_vnlock(vp, td);
394 mtx_lock(&np->n_mtx);
395 if (np->n_flag & NMODIFIED) {
396 mtx_unlock(&np->n_mtx);
397 if (vp->v_type != VREG) {
398 if (vp->v_type != VDIR)
399 panic("nfs: bioread, not dir");
400 (nmp->nm_rpcops->nr_invaldir)(vp);
401 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
402 if (error)
403 goto out;
404 }
405 np->n_attrstamp = 0;
406 error = VOP_GETATTR(vp, &vattr, cred, td);
407 if (error)
408 goto out;
409 mtx_lock(&np->n_mtx);
410 np->n_mtime = vattr.va_mtime;
411 mtx_unlock(&np->n_mtx);
412 } else {
413 mtx_unlock(&np->n_mtx);
414 error = VOP_GETATTR(vp, &vattr, cred, td);
415 if (error)
416 return (error);
417 mtx_lock(&np->n_mtx);
418 if ((np->n_flag & NSIZECHANGED)
419 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
420 mtx_unlock(&np->n_mtx);
421 if (vp->v_type == VDIR)
422 (nmp->nm_rpcops->nr_invaldir)(vp);
423 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
424 if (error)
425 goto out;
426 mtx_lock(&np->n_mtx);
427 np->n_mtime = vattr.va_mtime;
428 np->n_flag &= ~NSIZECHANGED;
429 }
430 mtx_unlock(&np->n_mtx);
431 }
432 out:
433 nfs_downgrade_vnlock(vp, td, old_lock);
434 return error;
435 }
436
437 /*
438 * Vnode op for read using bio
439 */
440 int
441 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
442 {
443 struct nfsnode *np = VTONFS(vp);
444 int biosize, i;
445 struct buf *bp, *rabp;
446 struct thread *td;
447 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
448 daddr_t lbn, rabn;
449 int bcount;
450 int seqcount;
451 int nra, error = 0, n = 0, on = 0;
452
453 #ifdef DIAGNOSTIC
454 if (uio->uio_rw != UIO_READ)
455 panic("nfs_read mode");
456 #endif
457 if (uio->uio_resid == 0)
458 return (0);
459 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
460 return (EINVAL);
461 td = uio->uio_td;
462
463 mtx_lock(&nmp->nm_mtx);
464 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
465 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
466 mtx_unlock(&nmp->nm_mtx);
467 (void)nfs_fsinfo(nmp, vp, cred, td);
468 } else
469 mtx_unlock(&nmp->nm_mtx);
470
471 if (vp->v_type != VDIR &&
472 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
473 return (EFBIG);
474
475 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
476 /* No caching/ no readaheads. Just read data into the user buffer */
477 return nfs_readrpc(vp, uio, cred);
478
479 biosize = vp->v_mount->mnt_stat.f_iosize;
480 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
481
482 error = nfs_bioread_check_cons(vp, td, cred);
483 if (error)
484 return error;
485
486 do {
487 u_quad_t nsize;
488
489 mtx_lock(&np->n_mtx);
490 nsize = np->n_size;
491 mtx_unlock(&np->n_mtx);
492
493 switch (vp->v_type) {
494 case VREG:
495 nfsstats.biocache_reads++;
496 lbn = uio->uio_offset / biosize;
497 on = uio->uio_offset & (biosize - 1);
498
499 /*
500 * Start the read ahead(s), as required.
501 */
502 if (nmp->nm_readahead > 0) {
503 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
504 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
505 rabn = lbn + 1 + nra;
506 if (incore(&vp->v_bufobj, rabn) == NULL) {
507 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
508 if (!rabp) {
509 error = nfs_sigintr(nmp, NULL, td);
510 return (error ? error : EINTR);
511 }
512 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
513 rabp->b_flags |= B_ASYNC;
514 rabp->b_iocmd = BIO_READ;
515 vfs_busy_pages(rabp, 0);
516 if (nfs_asyncio(nmp, rabp, cred, td)) {
517 rabp->b_flags |= B_INVAL;
518 rabp->b_ioflags |= BIO_ERROR;
519 vfs_unbusy_pages(rabp);
520 brelse(rabp);
521 break;
522 }
523 } else {
524 brelse(rabp);
525 }
526 }
527 }
528 }
529
530 /* Note that bcount is *not* DEV_BSIZE aligned. */
531 bcount = biosize;
532 if ((off_t)lbn * biosize >= nsize) {
533 bcount = 0;
534 } else if ((off_t)(lbn + 1) * biosize > nsize) {
535 bcount = nsize - (off_t)lbn * biosize;
536 }
537 bp = nfs_getcacheblk(vp, lbn, bcount, td);
538
539 if (!bp) {
540 error = nfs_sigintr(nmp, NULL, td);
541 return (error ? error : EINTR);
542 }
543
544 /*
545 * If B_CACHE is not set, we must issue the read. If this
546 * fails, we return an error.
547 */
548
549 if ((bp->b_flags & B_CACHE) == 0) {
550 bp->b_iocmd = BIO_READ;
551 vfs_busy_pages(bp, 0);
552 error = nfs_doio(vp, bp, cred, td);
553 if (error) {
554 brelse(bp);
555 return (error);
556 }
557 }
558
559 /*
560 * on is the offset into the current bp. Figure out how many
561 * bytes we can copy out of the bp. Note that bcount is
562 * NOT DEV_BSIZE aligned.
563 *
564 * Then figure out how many bytes we can copy into the uio.
565 */
566
567 n = 0;
568 if (on < bcount)
569 n = min((unsigned)(bcount - on), uio->uio_resid);
570 break;
571 case VLNK:
572 nfsstats.biocache_readlinks++;
573 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
574 if (!bp) {
575 error = nfs_sigintr(nmp, NULL, td);
576 return (error ? error : EINTR);
577 }
578 if ((bp->b_flags & B_CACHE) == 0) {
579 bp->b_iocmd = BIO_READ;
580 vfs_busy_pages(bp, 0);
581 error = nfs_doio(vp, bp, cred, td);
582 if (error) {
583 bp->b_ioflags |= BIO_ERROR;
584 brelse(bp);
585 return (error);
586 }
587 }
588 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
589 on = 0;
590 break;
591 case VDIR:
592 nfsstats.biocache_readdirs++;
593 if (np->n_direofoffset
594 && uio->uio_offset >= np->n_direofoffset) {
595 return (0);
596 }
597 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
598 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
599 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
600 if (!bp) {
601 error = nfs_sigintr(nmp, NULL, td);
602 return (error ? error : EINTR);
603 }
604 if ((bp->b_flags & B_CACHE) == 0) {
605 bp->b_iocmd = BIO_READ;
606 vfs_busy_pages(bp, 0);
607 error = nfs_doio(vp, bp, cred, td);
608 if (error) {
609 brelse(bp);
610 }
611 while (error == NFSERR_BAD_COOKIE) {
612 (nmp->nm_rpcops->nr_invaldir)(vp);
613 error = nfs_vinvalbuf(vp, 0, td, 1);
614 /*
615 * Yuck! The directory has been modified on the
616 * server. The only way to get the block is by
617 * reading from the beginning to get all the
618 * offset cookies.
619 *
620 * Leave the last bp intact unless there is an error.
621 * Loop back up to the while if the error is another
622 * NFSERR_BAD_COOKIE (double yuch!).
623 */
624 for (i = 0; i <= lbn && !error; i++) {
625 if (np->n_direofoffset
626 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
627 return (0);
628 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
629 if (!bp) {
630 error = nfs_sigintr(nmp, NULL, td);
631 return (error ? error : EINTR);
632 }
633 if ((bp->b_flags & B_CACHE) == 0) {
634 bp->b_iocmd = BIO_READ;
635 vfs_busy_pages(bp, 0);
636 error = nfs_doio(vp, bp, cred, td);
637 /*
638 * no error + B_INVAL == directory EOF,
639 * use the block.
640 */
641 if (error == 0 && (bp->b_flags & B_INVAL))
642 break;
643 }
644 /*
645 * An error will throw away the block and the
646 * for loop will break out. If no error and this
647 * is not the block we want, we throw away the
648 * block and go for the next one via the for loop.
649 */
650 if (error || i < lbn)
651 brelse(bp);
652 }
653 }
654 /*
655 * The above while is repeated if we hit another cookie
656 * error. If we hit an error and it wasn't a cookie error,
657 * we give up.
658 */
659 if (error)
660 return (error);
661 }
662
663 /*
664 * If not eof and read aheads are enabled, start one.
665 * (You need the current block first, so that you have the
666 * directory offset cookie of the next block.)
667 */
668 if (nmp->nm_readahead > 0 &&
669 (bp->b_flags & B_INVAL) == 0 &&
670 (np->n_direofoffset == 0 ||
671 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
672 incore(&vp->v_bufobj, lbn + 1) == NULL) {
673 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
674 if (rabp) {
675 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
676 rabp->b_flags |= B_ASYNC;
677 rabp->b_iocmd = BIO_READ;
678 vfs_busy_pages(rabp, 0);
679 if (nfs_asyncio(nmp, rabp, cred, td)) {
680 rabp->b_flags |= B_INVAL;
681 rabp->b_ioflags |= BIO_ERROR;
682 vfs_unbusy_pages(rabp);
683 brelse(rabp);
684 }
685 } else {
686 brelse(rabp);
687 }
688 }
689 }
690 /*
691 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
692 * chopped for the EOF condition, we cannot tell how large
693 * NFS directories are going to be until we hit EOF. So
694 * an NFS directory buffer is *not* chopped to its EOF. Now,
695 * it just so happens that b_resid will effectively chop it
696 * to EOF. *BUT* this information is lost if the buffer goes
697 * away and is reconstituted into a B_CACHE state ( due to
698 * being VMIO ) later. So we keep track of the directory eof
699 * in np->n_direofoffset and chop it off as an extra step
700 * right here.
701 */
702 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
703 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
704 n = np->n_direofoffset - uio->uio_offset;
705 break;
706 default:
707 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
708 bp = NULL;
709 break;
710 };
711
712 if (n > 0) {
713 error = uiomove(bp->b_data + on, (int)n, uio);
714 }
715 if (vp->v_type == VLNK)
716 n = 0;
717 if (bp != NULL)
718 brelse(bp);
719 } while (error == 0 && uio->uio_resid > 0 && n > 0);
720 return (error);
721 }
722
723 /*
724 * The NFS write path cannot handle iovecs with len > 1. So we need to
725 * break up iovecs accordingly (restricting them to wsize).
726 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
727 * For the ASYNC case, 2 copies are needed. The first a copy from the
728 * user buffer to a staging buffer and then a second copy from the staging
729 * buffer to mbufs. This can be optimized by copying from the user buffer
730 * directly into mbufs and passing the chain down, but that requires a
731 * fair amount of re-working of the relevant codepaths (and can be done
732 * later).
733 */
734 static int
735 nfs_directio_write(vp, uiop, cred, ioflag)
736 struct vnode *vp;
737 struct uio *uiop;
738 struct ucred *cred;
739 int ioflag;
740 {
741 int error;
742 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
743 struct thread *td = uiop->uio_td;
744 int size;
745 int wsize;
746
747 mtx_lock(&nmp->nm_mtx);
748 wsize = nmp->nm_wsize;
749 mtx_unlock(&nmp->nm_mtx);
750 if (ioflag & IO_SYNC) {
751 int iomode, must_commit;
752 struct uio uio;
753 struct iovec iov;
754 do_sync:
755 while (uiop->uio_resid > 0) {
756 size = min(uiop->uio_resid, wsize);
757 size = min(uiop->uio_iov->iov_len, size);
758 iov.iov_base = uiop->uio_iov->iov_base;
759 iov.iov_len = size;
760 uio.uio_iov = &iov;
761 uio.uio_iovcnt = 1;
762 uio.uio_offset = uiop->uio_offset;
763 uio.uio_resid = size;
764 uio.uio_segflg = UIO_USERSPACE;
765 uio.uio_rw = UIO_WRITE;
766 uio.uio_td = td;
767 iomode = NFSV3WRITE_FILESYNC;
768 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
769 &iomode, &must_commit);
770 KASSERT((must_commit == 0),
771 ("nfs_directio_write: Did not commit write"));
772 if (error)
773 return (error);
774 uiop->uio_offset += size;
775 uiop->uio_resid -= size;
776 if (uiop->uio_iov->iov_len <= size) {
777 uiop->uio_iovcnt--;
778 uiop->uio_iov++;
779 } else {
780 uiop->uio_iov->iov_base =
781 (char *)uiop->uio_iov->iov_base + size;
782 uiop->uio_iov->iov_len -= size;
783 }
784 }
785 } else {
786 struct uio *t_uio;
787 struct iovec *t_iov;
788 struct buf *bp;
789
790 /*
791 * Break up the write into blocksize chunks and hand these
792 * over to nfsiod's for write back.
793 * Unfortunately, this incurs a copy of the data. Since
794 * the user could modify the buffer before the write is
795 * initiated.
796 *
797 * The obvious optimization here is that one of the 2 copies
798 * in the async write path can be eliminated by copying the
799 * data here directly into mbufs and passing the mbuf chain
800 * down. But that will require a fair amount of re-working
801 * of the code and can be done if there's enough interest
802 * in NFS directio access.
803 */
804 while (uiop->uio_resid > 0) {
805 size = min(uiop->uio_resid, wsize);
806 size = min(uiop->uio_iov->iov_len, size);
807 bp = getpbuf(&nfs_pbuf_freecnt);
808 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
809 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
810 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
811 t_iov->iov_len = size;
812 t_uio->uio_iov = t_iov;
813 t_uio->uio_iovcnt = 1;
814 t_uio->uio_offset = uiop->uio_offset;
815 t_uio->uio_resid = size;
816 t_uio->uio_segflg = UIO_SYSSPACE;
817 t_uio->uio_rw = UIO_WRITE;
818 t_uio->uio_td = td;
819 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size);
820 bp->b_flags |= B_DIRECT;
821 bp->b_iocmd = BIO_WRITE;
822 if (cred != NOCRED) {
823 crhold(cred);
824 bp->b_wcred = cred;
825 } else
826 bp->b_wcred = NOCRED;
827 bp->b_caller1 = (void *)t_uio;
828 bp->b_vp = vp;
829 error = nfs_asyncio(nmp, bp, NOCRED, td);
830 if (error) {
831 free(t_iov->iov_base, M_NFSDIRECTIO);
832 free(t_iov, M_NFSDIRECTIO);
833 free(t_uio, M_NFSDIRECTIO);
834 bp->b_vp = NULL;
835 relpbuf(bp, &nfs_pbuf_freecnt);
836 if (error == EINTR)
837 return (error);
838 goto do_sync;
839 }
840 uiop->uio_offset += size;
841 uiop->uio_resid -= size;
842 if (uiop->uio_iov->iov_len <= size) {
843 uiop->uio_iovcnt--;
844 uiop->uio_iov++;
845 } else {
846 uiop->uio_iov->iov_base =
847 (char *)uiop->uio_iov->iov_base + size;
848 uiop->uio_iov->iov_len -= size;
849 }
850 }
851 }
852 return (0);
853 }
854
855 /*
856 * Vnode op for write using bio
857 */
858 int
859 nfs_write(struct vop_write_args *ap)
860 {
861 int biosize;
862 struct uio *uio = ap->a_uio;
863 struct thread *td = uio->uio_td;
864 struct vnode *vp = ap->a_vp;
865 struct nfsnode *np = VTONFS(vp);
866 struct ucred *cred = ap->a_cred;
867 int ioflag = ap->a_ioflag;
868 struct buf *bp;
869 struct vattr vattr;
870 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
871 daddr_t lbn;
872 int bcount;
873 int n, on, error = 0;
874 struct proc *p = td?td->td_proc:NULL;
875
876 #ifdef DIAGNOSTIC
877 if (uio->uio_rw != UIO_WRITE)
878 panic("nfs_write mode");
879 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
880 panic("nfs_write proc");
881 #endif
882 if (vp->v_type != VREG)
883 return (EIO);
884 mtx_lock(&np->n_mtx);
885 if (np->n_flag & NWRITEERR) {
886 np->n_flag &= ~NWRITEERR;
887 mtx_unlock(&np->n_mtx);
888 return (np->n_error);
889 } else
890 mtx_unlock(&np->n_mtx);
891 mtx_lock(&nmp->nm_mtx);
892 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
893 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
894 mtx_unlock(&nmp->nm_mtx);
895 (void)nfs_fsinfo(nmp, vp, cred, td);
896 } else
897 mtx_unlock(&nmp->nm_mtx);
898
899 /*
900 * Synchronously flush pending buffers if we are in synchronous
901 * mode or if we are appending.
902 */
903 if (ioflag & (IO_APPEND | IO_SYNC)) {
904 mtx_lock(&np->n_mtx);
905 if (np->n_flag & NMODIFIED) {
906 mtx_unlock(&np->n_mtx);
907 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
908 /*
909 * Require non-blocking, synchronous writes to
910 * dirty files to inform the program it needs
911 * to fsync(2) explicitly.
912 */
913 if (ioflag & IO_NDELAY)
914 return (EAGAIN);
915 #endif
916 flush_and_restart:
917 np->n_attrstamp = 0;
918 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
919 if (error)
920 return (error);
921 } else
922 mtx_unlock(&np->n_mtx);
923 }
924
925 /*
926 * If IO_APPEND then load uio_offset. We restart here if we cannot
927 * get the append lock.
928 */
929 if (ioflag & IO_APPEND) {
930 np->n_attrstamp = 0;
931 error = VOP_GETATTR(vp, &vattr, cred, td);
932 if (error)
933 return (error);
934 mtx_lock(&np->n_mtx);
935 uio->uio_offset = np->n_size;
936 mtx_unlock(&np->n_mtx);
937 }
938
939 if (uio->uio_offset < 0)
940 return (EINVAL);
941 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
942 return (EFBIG);
943 if (uio->uio_resid == 0)
944 return (0);
945
946 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
947 return nfs_directio_write(vp, uio, cred, ioflag);
948
949 /*
950 * Maybe this should be above the vnode op call, but so long as
951 * file servers have no limits, i don't think it matters
952 */
953 if (p != NULL) {
954 PROC_LOCK(p);
955 if (uio->uio_offset + uio->uio_resid >
956 lim_cur(p, RLIMIT_FSIZE)) {
957 psignal(p, SIGXFSZ);
958 PROC_UNLOCK(p);
959 return (EFBIG);
960 }
961 PROC_UNLOCK(p);
962 }
963
964 biosize = vp->v_mount->mnt_stat.f_iosize;
965 /*
966 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
967 * would exceed the local maximum per-file write commit size when
968 * combined with those, we must decide whether to flush,
969 * go synchronous, or return error. We don't bother checking
970 * IO_UNIT -- we just make all writes atomic anyway, as there's
971 * no point optimizing for something that really won't ever happen.
972 */
973 if (!(ioflag & IO_SYNC)) {
974 int nflag;
975
976 mtx_lock(&np->n_mtx);
977 nflag = np->n_flag;
978 mtx_unlock(&np->n_mtx);
979 int needrestart = 0;
980 if (nmp->nm_wcommitsize < uio->uio_resid) {
981 /*
982 * If this request could not possibly be completed
983 * without exceeding the maximum outstanding write
984 * commit size, see if we can convert it into a
985 * synchronous write operation.
986 */
987 if (ioflag & IO_NDELAY)
988 return (EAGAIN);
989 ioflag |= IO_SYNC;
990 if (nflag & NMODIFIED)
991 needrestart = 1;
992 } else if (nflag & NMODIFIED) {
993 int wouldcommit = 0;
994 BO_LOCK(&vp->v_bufobj);
995 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
996 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
997 b_bobufs) {
998 if (bp->b_flags & B_NEEDCOMMIT)
999 wouldcommit += bp->b_bcount;
1000 }
1001 }
1002 BO_UNLOCK(&vp->v_bufobj);
1003 /*
1004 * Since we're not operating synchronously and
1005 * bypassing the buffer cache, we are in a commit
1006 * and holding all of these buffers whether
1007 * transmitted or not. If not limited, this
1008 * will lead to the buffer cache deadlocking,
1009 * as no one else can flush our uncommitted buffers.
1010 */
1011 wouldcommit += uio->uio_resid;
1012 /*
1013 * If we would initially exceed the maximum
1014 * outstanding write commit size, flush and restart.
1015 */
1016 if (wouldcommit > nmp->nm_wcommitsize)
1017 needrestart = 1;
1018 }
1019 if (needrestart)
1020 goto flush_and_restart;
1021 }
1022
1023 do {
1024 nfsstats.biocache_writes++;
1025 lbn = uio->uio_offset / biosize;
1026 on = uio->uio_offset & (biosize-1);
1027 n = min((unsigned)(biosize - on), uio->uio_resid);
1028 again:
1029 /*
1030 * Handle direct append and file extension cases, calculate
1031 * unaligned buffer size.
1032 */
1033 mtx_lock(&np->n_mtx);
1034 if (uio->uio_offset == np->n_size && n) {
1035 mtx_unlock(&np->n_mtx);
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 mtx_lock(&np->n_mtx);
1049 np->n_size = uio->uio_offset + n;
1050 np->n_flag |= NMODIFIED;
1051 vnode_pager_setsize(vp, np->n_size);
1052 mtx_unlock(&np->n_mtx);
1053
1054 save = bp->b_flags & B_CACHE;
1055 bcount += n;
1056 allocbuf(bp, bcount);
1057 bp->b_flags |= save;
1058 }
1059 } else {
1060 /*
1061 * Obtain the locked cache block first, and then
1062 * adjust the file's size as appropriate.
1063 */
1064 bcount = on + n;
1065 if ((off_t)lbn * biosize + bcount < np->n_size) {
1066 if ((off_t)(lbn + 1) * biosize < np->n_size)
1067 bcount = biosize;
1068 else
1069 bcount = np->n_size - (off_t)lbn * biosize;
1070 }
1071 mtx_unlock(&np->n_mtx);
1072 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1073 mtx_lock(&np->n_mtx);
1074 if (uio->uio_offset + n > np->n_size) {
1075 np->n_size = uio->uio_offset + n;
1076 np->n_flag |= NMODIFIED;
1077 vnode_pager_setsize(vp, np->n_size);
1078 }
1079 mtx_unlock(&np->n_mtx);
1080 }
1081
1082 if (!bp) {
1083 error = nfs_sigintr(nmp, NULL, td);
1084 if (!error)
1085 error = EINTR;
1086 break;
1087 }
1088
1089 /*
1090 * Issue a READ if B_CACHE is not set. In special-append
1091 * mode, B_CACHE is based on the buffer prior to the write
1092 * op and is typically set, avoiding the read. If a read
1093 * is required in special append mode, the server will
1094 * probably send us a short-read since we extended the file
1095 * on our end, resulting in b_resid == 0 and, thusly,
1096 * B_CACHE getting set.
1097 *
1098 * We can also avoid issuing the read if the write covers
1099 * the entire buffer. We have to make sure the buffer state
1100 * is reasonable in this case since we will not be initiating
1101 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1102 * more information.
1103 *
1104 * B_CACHE may also be set due to the buffer being cached
1105 * normally.
1106 */
1107
1108 if (on == 0 && n == bcount) {
1109 bp->b_flags |= B_CACHE;
1110 bp->b_flags &= ~B_INVAL;
1111 bp->b_ioflags &= ~BIO_ERROR;
1112 }
1113
1114 if ((bp->b_flags & B_CACHE) == 0) {
1115 bp->b_iocmd = BIO_READ;
1116 vfs_busy_pages(bp, 0);
1117 error = nfs_doio(vp, bp, cred, td);
1118 if (error) {
1119 brelse(bp);
1120 break;
1121 }
1122 }
1123 if (bp->b_wcred == NOCRED)
1124 bp->b_wcred = crhold(cred);
1125 mtx_lock(&np->n_mtx);
1126 np->n_flag |= NMODIFIED;
1127 mtx_unlock(&np->n_mtx);
1128
1129 /*
1130 * If dirtyend exceeds file size, chop it down. This should
1131 * not normally occur but there is an append race where it
1132 * might occur XXX, so we log it.
1133 *
1134 * If the chopping creates a reverse-indexed or degenerate
1135 * situation with dirtyoff/end, we 0 both of them.
1136 */
1137
1138 if (bp->b_dirtyend > bcount) {
1139 nfs_printf("NFS append race @%lx:%d\n",
1140 (long)bp->b_blkno * DEV_BSIZE,
1141 bp->b_dirtyend - bcount);
1142 bp->b_dirtyend = bcount;
1143 }
1144
1145 if (bp->b_dirtyoff >= bp->b_dirtyend)
1146 bp->b_dirtyoff = bp->b_dirtyend = 0;
1147
1148 /*
1149 * If the new write will leave a contiguous dirty
1150 * area, just update the b_dirtyoff and b_dirtyend,
1151 * otherwise force a write rpc of the old dirty area.
1152 *
1153 * While it is possible to merge discontiguous writes due to
1154 * our having a B_CACHE buffer ( and thus valid read data
1155 * for the hole), we don't because it could lead to
1156 * significant cache coherency problems with multiple clients,
1157 * especially if locking is implemented later on.
1158 *
1159 * as an optimization we could theoretically maintain
1160 * a linked list of discontinuous areas, but we would still
1161 * have to commit them separately so there isn't much
1162 * advantage to it except perhaps a bit of asynchronization.
1163 */
1164
1165 if (bp->b_dirtyend > 0 &&
1166 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1167 if (bwrite(bp) == EINTR) {
1168 error = EINTR;
1169 break;
1170 }
1171 goto again;
1172 }
1173
1174 error = uiomove((char *)bp->b_data + on, n, uio);
1175
1176 /*
1177 * Since this block is being modified, it must be written
1178 * again and not just committed. Since write clustering does
1179 * not work for the stage 1 data write, only the stage 2
1180 * commit rpc, we have to clear B_CLUSTEROK as well.
1181 */
1182 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1183
1184 if (error) {
1185 bp->b_ioflags |= BIO_ERROR;
1186 brelse(bp);
1187 break;
1188 }
1189
1190 /*
1191 * Only update dirtyoff/dirtyend if not a degenerate
1192 * condition.
1193 */
1194 if (n) {
1195 if (bp->b_dirtyend > 0) {
1196 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1197 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1198 } else {
1199 bp->b_dirtyoff = on;
1200 bp->b_dirtyend = on + n;
1201 }
1202 vfs_bio_set_validclean(bp, on, n);
1203 }
1204
1205 /*
1206 * If IO_SYNC do bwrite().
1207 *
1208 * IO_INVAL appears to be unused. The idea appears to be
1209 * to turn off caching in this case. Very odd. XXX
1210 */
1211 if ((ioflag & IO_SYNC)) {
1212 if (ioflag & IO_INVAL)
1213 bp->b_flags |= B_NOCACHE;
1214 error = bwrite(bp);
1215 if (error)
1216 break;
1217 } else if ((n + on) == biosize) {
1218 bp->b_flags |= B_ASYNC;
1219 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1220 } else {
1221 bdwrite(bp);
1222 }
1223 } while (uio->uio_resid > 0 && n > 0);
1224
1225 return (error);
1226 }
1227
1228 /*
1229 * Get an nfs cache block.
1230 *
1231 * Allocate a new one if the block isn't currently in the cache
1232 * and return the block marked busy. If the calling process is
1233 * interrupted by a signal for an interruptible mount point, return
1234 * NULL.
1235 *
1236 * The caller must carefully deal with the possible B_INVAL state of
1237 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1238 * indirectly), so synchronous reads can be issued without worrying about
1239 * the B_INVAL state. We have to be a little more careful when dealing
1240 * with writes (see comments in nfs_write()) when extending a file past
1241 * its EOF.
1242 */
1243 static struct buf *
1244 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1245 {
1246 struct buf *bp;
1247 struct mount *mp;
1248 struct nfsmount *nmp;
1249
1250 mp = vp->v_mount;
1251 nmp = VFSTONFS(mp);
1252
1253 if (nmp->nm_flag & NFSMNT_INT) {
1254 sigset_t oldset;
1255
1256 nfs_set_sigmask(td, &oldset);
1257 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1258 nfs_restore_sigmask(td, &oldset);
1259 while (bp == NULL) {
1260 if (nfs_sigintr(nmp, NULL, td))
1261 return (NULL);
1262 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1263 }
1264 } else {
1265 bp = getblk(vp, bn, size, 0, 0, 0);
1266 }
1267
1268 if (vp->v_type == VREG) {
1269 int biosize;
1270
1271 biosize = mp->mnt_stat.f_iosize;
1272 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1273 }
1274 return (bp);
1275 }
1276
1277 /*
1278 * Flush and invalidate all dirty buffers. If another process is already
1279 * doing the flush, just wait for completion.
1280 */
1281 int
1282 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1283 {
1284 struct nfsnode *np = VTONFS(vp);
1285 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1286 int error = 0, slpflag, slptimeo;
1287 int old_lock = 0;
1288
1289 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1290
1291 /*
1292 * XXX This check stops us from needlessly doing a vinvalbuf when
1293 * being called through vclean(). It is not clear that this is
1294 * unsafe.
1295 */
1296 if (vp->v_iflag & VI_DOOMED)
1297 return (0);
1298
1299 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1300 intrflg = 0;
1301 if (intrflg) {
1302 slpflag = PCATCH;
1303 slptimeo = 2 * hz;
1304 } else {
1305 slpflag = 0;
1306 slptimeo = 0;
1307 }
1308
1309 old_lock = nfs_upgrade_vnlock(vp, td);
1310 /*
1311 * Now, flush as required.
1312 */
1313 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1314 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1315 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1316 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1317 /*
1318 * If the page clean was interrupted, fail the invalidation.
1319 * Not doing so, we run the risk of losing dirty pages in the
1320 * vinvalbuf() call below.
1321 */
1322 if (intrflg && (error = nfs_sigintr(nmp, NULL, td)))
1323 goto out;
1324 }
1325
1326 error = vinvalbuf(vp, flags, td, slpflag, 0);
1327 while (error) {
1328 if (intrflg && (error = nfs_sigintr(nmp, NULL, td)))
1329 goto out;
1330 error = vinvalbuf(vp, flags, td, 0, slptimeo);
1331 }
1332 mtx_lock(&np->n_mtx);
1333 if (np->n_directio_asyncwr == 0)
1334 np->n_flag &= ~NMODIFIED;
1335 mtx_unlock(&np->n_mtx);
1336 out:
1337 nfs_downgrade_vnlock(vp, td, old_lock);
1338 return error;
1339 }
1340
1341 /*
1342 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1343 * This is mainly to avoid queueing async I/O requests when the nfsiods
1344 * are all hung on a dead server.
1345 *
1346 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1347 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1348 */
1349 int
1350 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1351 {
1352 int iod;
1353 int gotiod;
1354 int slpflag = 0;
1355 int slptimeo = 0;
1356 int error, error2;
1357
1358 /*
1359 * Commits are usually short and sweet so lets save some cpu and
1360 * leave the async daemons for more important rpc's (such as reads
1361 * and writes).
1362 */
1363 mtx_lock(&nfs_iod_mtx);
1364 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1365 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1366 mtx_unlock(&nfs_iod_mtx);
1367 return(EIO);
1368 }
1369 again:
1370 if (nmp->nm_flag & NFSMNT_INT)
1371 slpflag = PCATCH;
1372 gotiod = FALSE;
1373
1374 /*
1375 * Find a free iod to process this request.
1376 */
1377 for (iod = 0; iod < nfs_numasync; iod++)
1378 if (nfs_iodwant[iod]) {
1379 gotiod = TRUE;
1380 break;
1381 }
1382
1383 /*
1384 * Try to create one if none are free.
1385 */
1386 if (!gotiod) {
1387 iod = nfs_nfsiodnew();
1388 if (iod != -1)
1389 gotiod = TRUE;
1390 }
1391
1392 if (gotiod) {
1393 /*
1394 * Found one, so wake it up and tell it which
1395 * mount to process.
1396 */
1397 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1398 iod, nmp));
1399 nfs_iodwant[iod] = NULL;
1400 nfs_iodmount[iod] = nmp;
1401 nmp->nm_bufqiods++;
1402 wakeup(&nfs_iodwant[iod]);
1403 }
1404
1405 /*
1406 * If none are free, we may already have an iod working on this mount
1407 * point. If so, it will process our request.
1408 */
1409 if (!gotiod) {
1410 if (nmp->nm_bufqiods > 0) {
1411 NFS_DPF(ASYNCIO,
1412 ("nfs_asyncio: %d iods are already processing mount %p\n",
1413 nmp->nm_bufqiods, nmp));
1414 gotiod = TRUE;
1415 }
1416 }
1417
1418 /*
1419 * If we have an iod which can process the request, then queue
1420 * the buffer.
1421 */
1422 if (gotiod) {
1423 /*
1424 * Ensure that the queue never grows too large. We still want
1425 * to asynchronize so we block rather then return EIO.
1426 */
1427 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1428 NFS_DPF(ASYNCIO,
1429 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1430 nmp->nm_bufqwant = TRUE;
1431 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1432 slpflag | PRIBIO,
1433 "nfsaio", slptimeo);
1434 if (error) {
1435 error2 = nfs_sigintr(nmp, NULL, td);
1436 if (error2) {
1437 mtx_unlock(&nfs_iod_mtx);
1438 return (error2);
1439 }
1440 if (slpflag == PCATCH) {
1441 slpflag = 0;
1442 slptimeo = 2 * hz;
1443 }
1444 }
1445 /*
1446 * We might have lost our iod while sleeping,
1447 * so check and loop if nescessary.
1448 */
1449 if (nmp->nm_bufqiods == 0) {
1450 NFS_DPF(ASYNCIO,
1451 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1452 goto again;
1453 }
1454 }
1455
1456 /* We might have lost our nfsiod */
1457 if (nmp->nm_bufqiods == 0) {
1458 NFS_DPF(ASYNCIO,
1459 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1460 goto again;
1461 }
1462
1463 if (bp->b_iocmd == BIO_READ) {
1464 if (bp->b_rcred == NOCRED && cred != NOCRED)
1465 bp->b_rcred = crhold(cred);
1466 } else {
1467 if (bp->b_wcred == NOCRED && cred != NOCRED)
1468 bp->b_wcred = crhold(cred);
1469 }
1470
1471 if (bp->b_flags & B_REMFREE)
1472 bremfreef(bp);
1473 BUF_KERNPROC(bp);
1474 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1475 nmp->nm_bufqlen++;
1476 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1477 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1478 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1479 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1480 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1481 }
1482 mtx_unlock(&nfs_iod_mtx);
1483 return (0);
1484 }
1485
1486 mtx_unlock(&nfs_iod_mtx);
1487
1488 /*
1489 * All the iods are busy on other mounts, so return EIO to
1490 * force the caller to process the i/o synchronously.
1491 */
1492 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1493 return (EIO);
1494 }
1495
1496 void
1497 nfs_doio_directwrite(struct buf *bp)
1498 {
1499 int iomode, must_commit;
1500 struct uio *uiop = (struct uio *)bp->b_caller1;
1501 char *iov_base = uiop->uio_iov->iov_base;
1502 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1503
1504 iomode = NFSV3WRITE_FILESYNC;
1505 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1506 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1507 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1508 free(iov_base, M_NFSDIRECTIO);
1509 free(uiop->uio_iov, M_NFSDIRECTIO);
1510 free(uiop, M_NFSDIRECTIO);
1511 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1512 struct nfsnode *np = VTONFS(bp->b_vp);
1513 mtx_lock(&np->n_mtx);
1514 np->n_directio_asyncwr--;
1515 if (np->n_directio_asyncwr == 0) {
1516 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1517 if ((np->n_flag & NFSYNCWAIT)) {
1518 np->n_flag &= ~NFSYNCWAIT;
1519 wakeup((caddr_t)&np->n_directio_asyncwr);
1520 }
1521 }
1522 mtx_unlock(&np->n_mtx);
1523 }
1524 bp->b_vp = NULL;
1525 relpbuf(bp, &nfs_pbuf_freecnt);
1526 }
1527
1528 /*
1529 * Do an I/O operation to/from a cache block. This may be called
1530 * synchronously or from an nfsiod.
1531 */
1532 int
1533 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1534 {
1535 struct uio *uiop;
1536 struct nfsnode *np;
1537 struct nfsmount *nmp;
1538 int error = 0, iomode, must_commit = 0;
1539 struct uio uio;
1540 struct iovec io;
1541 struct proc *p = td ? td->td_proc : NULL;
1542 uint8_t iocmd;
1543
1544 np = VTONFS(vp);
1545 nmp = VFSTONFS(vp->v_mount);
1546 uiop = &uio;
1547 uiop->uio_iov = &io;
1548 uiop->uio_iovcnt = 1;
1549 uiop->uio_segflg = UIO_SYSSPACE;
1550 uiop->uio_td = td;
1551
1552 /*
1553 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1554 * do this here so we do not have to do it in all the code that
1555 * calls us.
1556 */
1557 bp->b_flags &= ~B_INVAL;
1558 bp->b_ioflags &= ~BIO_ERROR;
1559
1560 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1561 iocmd = bp->b_iocmd;
1562 if (iocmd == BIO_READ) {
1563 io.iov_len = uiop->uio_resid = bp->b_bcount;
1564 io.iov_base = bp->b_data;
1565 uiop->uio_rw = UIO_READ;
1566
1567 switch (vp->v_type) {
1568 case VREG:
1569 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1570 nfsstats.read_bios++;
1571 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1572
1573 if (!error) {
1574 if (uiop->uio_resid) {
1575 /*
1576 * If we had a short read with no error, we must have
1577 * hit a file hole. We should zero-fill the remainder.
1578 * This can also occur if the server hits the file EOF.
1579 *
1580 * Holes used to be able to occur due to pending
1581 * writes, but that is not possible any longer.
1582 */
1583 int nread = bp->b_bcount - uiop->uio_resid;
1584 int left = uiop->uio_resid;
1585
1586 if (left > 0)
1587 bzero((char *)bp->b_data + nread, left);
1588 uiop->uio_resid = 0;
1589 }
1590 }
1591 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1592 if (p && (vp->v_vflag & VV_TEXT)) {
1593 mtx_lock(&np->n_mtx);
1594 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1595 mtx_unlock(&np->n_mtx);
1596 PROC_LOCK(p);
1597 killproc(p, "text file modification");
1598 PROC_UNLOCK(p);
1599 } else
1600 mtx_unlock(&np->n_mtx);
1601 }
1602 break;
1603 case VLNK:
1604 uiop->uio_offset = (off_t)0;
1605 nfsstats.readlink_bios++;
1606 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1607 break;
1608 case VDIR:
1609 nfsstats.readdir_bios++;
1610 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1611 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0)
1612 error = nfs4_readdirrpc(vp, uiop, cr);
1613 else {
1614 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1615 error = nfs_readdirplusrpc(vp, uiop, cr);
1616 if (error == NFSERR_NOTSUPP)
1617 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1618 }
1619 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1620 error = nfs_readdirrpc(vp, uiop, cr);
1621 }
1622 /*
1623 * end-of-directory sets B_INVAL but does not generate an
1624 * error.
1625 */
1626 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1627 bp->b_flags |= B_INVAL;
1628 break;
1629 default:
1630 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1631 break;
1632 };
1633 if (error) {
1634 bp->b_ioflags |= BIO_ERROR;
1635 bp->b_error = error;
1636 }
1637 } else {
1638 /*
1639 * If we only need to commit, try to commit
1640 */
1641 if (bp->b_flags & B_NEEDCOMMIT) {
1642 int retv;
1643 off_t off;
1644
1645 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1646 retv = (nmp->nm_rpcops->nr_commit)(
1647 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1648 bp->b_wcred, td);
1649 if (retv == 0) {
1650 bp->b_dirtyoff = bp->b_dirtyend = 0;
1651 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1652 bp->b_resid = 0;
1653 bufdone(bp);
1654 return (0);
1655 }
1656 if (retv == NFSERR_STALEWRITEVERF) {
1657 nfs_clearcommit(vp->v_mount);
1658 }
1659 }
1660
1661 /*
1662 * Setup for actual write
1663 */
1664 mtx_lock(&np->n_mtx);
1665 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1666 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1667 mtx_unlock(&np->n_mtx);
1668
1669 if (bp->b_dirtyend > bp->b_dirtyoff) {
1670 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1671 - bp->b_dirtyoff;
1672 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1673 + bp->b_dirtyoff;
1674 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1675 uiop->uio_rw = UIO_WRITE;
1676 nfsstats.write_bios++;
1677
1678 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1679 iomode = NFSV3WRITE_UNSTABLE;
1680 else
1681 iomode = NFSV3WRITE_FILESYNC;
1682
1683 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1684
1685 /*
1686 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1687 * to cluster the buffers needing commit. This will allow
1688 * the system to submit a single commit rpc for the whole
1689 * cluster. We can do this even if the buffer is not 100%
1690 * dirty (relative to the NFS blocksize), so we optimize the
1691 * append-to-file-case.
1692 *
1693 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1694 * cleared because write clustering only works for commit
1695 * rpc's, not for the data portion of the write).
1696 */
1697
1698 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1699 bp->b_flags |= B_NEEDCOMMIT;
1700 if (bp->b_dirtyoff == 0
1701 && bp->b_dirtyend == bp->b_bcount)
1702 bp->b_flags |= B_CLUSTEROK;
1703 } else {
1704 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1705 }
1706
1707 /*
1708 * For an interrupted write, the buffer is still valid
1709 * and the write hasn't been pushed to the server yet,
1710 * so we can't set BIO_ERROR and report the interruption
1711 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1712 * is not relevant, so the rpc attempt is essentially
1713 * a noop. For the case of a V3 write rpc not being
1714 * committed to stable storage, the block is still
1715 * dirty and requires either a commit rpc or another
1716 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1717 * the block is reused. This is indicated by setting
1718 * the B_DELWRI and B_NEEDCOMMIT flags.
1719 *
1720 * If the buffer is marked B_PAGING, it does not reside on
1721 * the vp's paging queues so we cannot call bdirty(). The
1722 * bp in this case is not an NFS cache block so we should
1723 * be safe. XXX
1724 *
1725 * The logic below breaks up errors into recoverable and
1726 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1727 * and keep the buffer around for potential write retries.
1728 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1729 * and save the error in the nfsnode. This is less than ideal
1730 * but necessary. Keeping such buffers around could potentially
1731 * cause buffer exhaustion eventually (they can never be written
1732 * out, so will get constantly be re-dirtied). It also causes
1733 * all sorts of vfs panics. For non-recoverable write errors,
1734 * also invalidate the attrcache, so we'll be forced to go over
1735 * the wire for this object, returning an error to user on next
1736 * call (most of the time).
1737 */
1738 if (error == EINTR || error == EIO || error == ETIMEDOUT
1739 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1740 int s;
1741
1742 s = splbio();
1743 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1744 if ((bp->b_flags & B_PAGING) == 0) {
1745 bdirty(bp);
1746 bp->b_flags &= ~B_DONE;
1747 }
1748 if (error && (bp->b_flags & B_ASYNC) == 0)
1749 bp->b_flags |= B_EINTR;
1750 splx(s);
1751 } else {
1752 if (error) {
1753 bp->b_ioflags |= BIO_ERROR;
1754 bp->b_flags |= B_INVAL;
1755 bp->b_error = np->n_error = error;
1756 mtx_lock(&np->n_mtx);
1757 np->n_flag |= NWRITEERR;
1758 np->n_attrstamp = 0;
1759 mtx_unlock(&np->n_mtx);
1760 }
1761 bp->b_dirtyoff = bp->b_dirtyend = 0;
1762 }
1763 } else {
1764 bp->b_resid = 0;
1765 bufdone(bp);
1766 return (0);
1767 }
1768 }
1769 bp->b_resid = uiop->uio_resid;
1770 if (must_commit)
1771 nfs_clearcommit(vp->v_mount);
1772 bufdone(bp);
1773 return (error);
1774 }
1775
1776 /*
1777 * Used to aid in handling ftruncate() operations on the NFS client side.
1778 * Truncation creates a number of special problems for NFS. We have to
1779 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1780 * we have to properly handle VM pages or (potentially dirty) buffers
1781 * that straddle the truncation point.
1782 */
1783
1784 int
1785 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1786 {
1787 struct nfsnode *np = VTONFS(vp);
1788 u_quad_t tsize;
1789 int biosize = vp->v_mount->mnt_stat.f_iosize;
1790 int error = 0;
1791
1792 mtx_lock(&np->n_mtx);
1793 tsize = np->n_size;
1794 np->n_size = nsize;
1795 mtx_unlock(&np->n_mtx);
1796
1797 if (nsize < tsize) {
1798 struct buf *bp;
1799 daddr_t lbn;
1800 int bufsize;
1801
1802 /*
1803 * vtruncbuf() doesn't get the buffer overlapping the
1804 * truncation point. We may have a B_DELWRI and/or B_CACHE
1805 * buffer that now needs to be truncated.
1806 */
1807 error = vtruncbuf(vp, cred, td, nsize, biosize);
1808 lbn = nsize / biosize;
1809 bufsize = nsize & (biosize - 1);
1810 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1811 if (!bp)
1812 return EINTR;
1813 if (bp->b_dirtyoff > bp->b_bcount)
1814 bp->b_dirtyoff = bp->b_bcount;
1815 if (bp->b_dirtyend > bp->b_bcount)
1816 bp->b_dirtyend = bp->b_bcount;
1817 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1818 brelse(bp);
1819 } else {
1820 vnode_pager_setsize(vp, nsize);
1821 }
1822 return(error);
1823 }
1824
Cache object: 5cbe308a9fae453891415e09e78cef00
|