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/8.3/sys/nfsclient/nfs_bio.c 231332 2012-02-10 04:01:17Z rmacklem $");
37
38 #include "opt_kdtrace.h"
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/bio.h>
43 #include <sys/buf.h>
44 #include <sys/kernel.h>
45 #include <sys/mbuf.h>
46 #include <sys/mount.h>
47 #include <sys/proc.h>
48 #include <sys/vmmeter.h>
49 #include <sys/vnode.h>
50
51 #include <vm/vm.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vnode_pager.h>
57
58 #include <nfs/nfsproto.h>
59 #include <nfsclient/nfs.h>
60 #include <nfsclient/nfsmount.h>
61 #include <nfsclient/nfsnode.h>
62 #include <nfsclient/nfs_kdtrace.h>
63
64 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
65 struct thread *td);
66 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
67 struct ucred *cred, int ioflag);
68
69 extern int nfs_directio_enable;
70 extern int nfs_directio_allow_mmap;
71
72 /*
73 * Vnode op for VM getpages.
74 */
75 int
76 nfs_getpages(struct vop_getpages_args *ap)
77 {
78 int i, error, nextoff, size, toff, count, npages;
79 struct uio uio;
80 struct iovec iov;
81 vm_offset_t kva;
82 struct buf *bp;
83 struct vnode *vp;
84 struct thread *td;
85 struct ucred *cred;
86 struct nfsmount *nmp;
87 vm_object_t object;
88 vm_page_t *pages;
89 struct nfsnode *np;
90
91 vp = ap->a_vp;
92 np = VTONFS(vp);
93 td = curthread; /* XXX */
94 cred = curthread->td_ucred; /* XXX */
95 nmp = VFSTONFS(vp->v_mount);
96 pages = ap->a_m;
97 count = ap->a_count;
98
99 if ((object = vp->v_object) == NULL) {
100 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n");
101 return (VM_PAGER_ERROR);
102 }
103
104 if (nfs_directio_enable && !nfs_directio_allow_mmap) {
105 mtx_lock(&np->n_mtx);
106 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
107 mtx_unlock(&np->n_mtx);
108 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n");
109 return (VM_PAGER_ERROR);
110 } else
111 mtx_unlock(&np->n_mtx);
112 }
113
114 mtx_lock(&nmp->nm_mtx);
115 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
116 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
117 mtx_unlock(&nmp->nm_mtx);
118 /* We'll never get here for v4, because we always have fsinfo */
119 (void)nfs_fsinfo(nmp, vp, cred, td);
120 } else
121 mtx_unlock(&nmp->nm_mtx);
122
123 npages = btoc(count);
124
125 /*
126 * If the requested page is partially valid, just return it and
127 * allow the pager to zero-out the blanks. Partially valid pages
128 * can only occur at the file EOF.
129 */
130 VM_OBJECT_LOCK(object);
131 if (pages[ap->a_reqpage]->valid != 0) {
132 vm_page_lock_queues();
133 for (i = 0; i < npages; ++i) {
134 if (i != ap->a_reqpage)
135 vm_page_free(pages[i]);
136 }
137 vm_page_unlock_queues();
138 VM_OBJECT_UNLOCK(object);
139 return (0);
140 }
141 VM_OBJECT_UNLOCK(object);
142
143 /*
144 * We use only the kva address for the buffer, but this is extremely
145 * convienient and fast.
146 */
147 bp = getpbuf(&nfs_pbuf_freecnt);
148
149 kva = (vm_offset_t) bp->b_data;
150 pmap_qenter(kva, pages, npages);
151 PCPU_INC(cnt.v_vnodein);
152 PCPU_ADD(cnt.v_vnodepgsin, npages);
153
154 iov.iov_base = (caddr_t) kva;
155 iov.iov_len = count;
156 uio.uio_iov = &iov;
157 uio.uio_iovcnt = 1;
158 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
159 uio.uio_resid = count;
160 uio.uio_segflg = UIO_SYSSPACE;
161 uio.uio_rw = UIO_READ;
162 uio.uio_td = td;
163
164 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
165 pmap_qremove(kva, npages);
166
167 relpbuf(bp, &nfs_pbuf_freecnt);
168
169 if (error && (uio.uio_resid == count)) {
170 nfs_printf("nfs_getpages: error %d\n", error);
171 VM_OBJECT_LOCK(object);
172 vm_page_lock_queues();
173 for (i = 0; i < npages; ++i) {
174 if (i != ap->a_reqpage)
175 vm_page_free(pages[i]);
176 }
177 vm_page_unlock_queues();
178 VM_OBJECT_UNLOCK(object);
179 return (VM_PAGER_ERROR);
180 }
181
182 /*
183 * Calculate the number of bytes read and validate only that number
184 * of bytes. Note that due to pending writes, size may be 0. This
185 * does not mean that the remaining data is invalid!
186 */
187
188 size = count - uio.uio_resid;
189 VM_OBJECT_LOCK(object);
190 vm_page_lock_queues();
191 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
192 vm_page_t m;
193 nextoff = toff + PAGE_SIZE;
194 m = pages[i];
195
196 if (nextoff <= size) {
197 /*
198 * Read operation filled an entire page
199 */
200 m->valid = VM_PAGE_BITS_ALL;
201 KASSERT(m->dirty == 0,
202 ("nfs_getpages: page %p is dirty", m));
203 } else if (size > toff) {
204 /*
205 * Read operation filled a partial page.
206 */
207 m->valid = 0;
208 vm_page_set_valid(m, 0, size - toff);
209 KASSERT(m->dirty == 0,
210 ("nfs_getpages: page %p is dirty", m));
211 } else {
212 /*
213 * Read operation was short. If no error occured
214 * we may have hit a zero-fill section. We simply
215 * leave valid set to 0.
216 */
217 ;
218 }
219 if (i != ap->a_reqpage) {
220 /*
221 * Whether or not to leave the page activated is up in
222 * the air, but we should put the page on a page queue
223 * somewhere (it already is in the object). Result:
224 * It appears that emperical results show that
225 * deactivating pages is best.
226 */
227
228 /*
229 * Just in case someone was asking for this page we
230 * now tell them that it is ok to use.
231 */
232 if (!error) {
233 if (m->oflags & VPO_WANTED)
234 vm_page_activate(m);
235 else
236 vm_page_deactivate(m);
237 vm_page_wakeup(m);
238 } else {
239 vm_page_free(m);
240 }
241 }
242 }
243 vm_page_unlock_queues();
244 VM_OBJECT_UNLOCK(object);
245 return (0);
246 }
247
248 /*
249 * Vnode op for VM putpages.
250 */
251 int
252 nfs_putpages(struct vop_putpages_args *ap)
253 {
254 struct uio uio;
255 struct iovec iov;
256 vm_offset_t kva;
257 struct buf *bp;
258 int iomode, must_commit, i, error, npages, count;
259 off_t offset;
260 int *rtvals;
261 struct vnode *vp;
262 struct thread *td;
263 struct ucred *cred;
264 struct nfsmount *nmp;
265 struct nfsnode *np;
266 vm_page_t *pages;
267
268 vp = ap->a_vp;
269 np = VTONFS(vp);
270 td = curthread; /* XXX */
271 cred = curthread->td_ucred; /* XXX */
272 nmp = VFSTONFS(vp->v_mount);
273 pages = ap->a_m;
274 count = ap->a_count;
275 rtvals = ap->a_rtvals;
276 npages = btoc(count);
277 offset = IDX_TO_OFF(pages[0]->pindex);
278
279 mtx_lock(&nmp->nm_mtx);
280 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
281 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
282 mtx_unlock(&nmp->nm_mtx);
283 (void)nfs_fsinfo(nmp, vp, cred, td);
284 } else
285 mtx_unlock(&nmp->nm_mtx);
286
287 mtx_lock(&np->n_mtx);
288 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
289 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
290 mtx_unlock(&np->n_mtx);
291 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
292 mtx_lock(&np->n_mtx);
293 }
294
295 for (i = 0; i < npages; i++)
296 rtvals[i] = VM_PAGER_ERROR;
297
298 /*
299 * When putting pages, do not extend file past EOF.
300 */
301 if (offset + count > np->n_size) {
302 count = np->n_size - offset;
303 if (count < 0)
304 count = 0;
305 }
306 mtx_unlock(&np->n_mtx);
307
308 /*
309 * We use only the kva address for the buffer, but this is extremely
310 * convienient and fast.
311 */
312 bp = getpbuf(&nfs_pbuf_freecnt);
313
314 kva = (vm_offset_t) bp->b_data;
315 pmap_qenter(kva, pages, npages);
316 PCPU_INC(cnt.v_vnodeout);
317 PCPU_ADD(cnt.v_vnodepgsout, count);
318
319 iov.iov_base = (caddr_t) kva;
320 iov.iov_len = count;
321 uio.uio_iov = &iov;
322 uio.uio_iovcnt = 1;
323 uio.uio_offset = offset;
324 uio.uio_resid = count;
325 uio.uio_segflg = UIO_SYSSPACE;
326 uio.uio_rw = UIO_WRITE;
327 uio.uio_td = td;
328
329 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
330 iomode = NFSV3WRITE_UNSTABLE;
331 else
332 iomode = NFSV3WRITE_FILESYNC;
333
334 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
335
336 pmap_qremove(kva, npages);
337 relpbuf(bp, &nfs_pbuf_freecnt);
338
339 if (!error) {
340 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
341 if (must_commit) {
342 nfs_clearcommit(vp->v_mount);
343 }
344 }
345 return rtvals[0];
346 }
347
348 /*
349 * For nfs, cache consistency can only be maintained approximately.
350 * Although RFC1094 does not specify the criteria, the following is
351 * believed to be compatible with the reference port.
352 * For nfs:
353 * If the file's modify time on the server has changed since the
354 * last read rpc or you have written to the file,
355 * you may have lost data cache consistency with the
356 * server, so flush all of the file's data out of the cache.
357 * Then force a getattr rpc to ensure that you have up to date
358 * attributes.
359 * NB: This implies that cache data can be read when up to
360 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
361 * attributes this could be forced by setting n_attrstamp to 0 before
362 * the VOP_GETATTR() call.
363 */
364 static inline int
365 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
366 {
367 int error = 0;
368 struct vattr vattr;
369 struct nfsnode *np = VTONFS(vp);
370 int old_lock;
371 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
372
373 /*
374 * Grab the exclusive lock before checking whether the cache is
375 * consistent.
376 * XXX - We can make this cheaper later (by acquiring cheaper locks).
377 * But for now, this suffices.
378 */
379 old_lock = nfs_upgrade_vnlock(vp);
380 if (vp->v_iflag & VI_DOOMED) {
381 nfs_downgrade_vnlock(vp, old_lock);
382 return (EBADF);
383 }
384
385 mtx_lock(&np->n_mtx);
386 if (np->n_flag & NMODIFIED) {
387 mtx_unlock(&np->n_mtx);
388 if (vp->v_type != VREG) {
389 if (vp->v_type != VDIR)
390 panic("nfs: bioread, not dir");
391 (nmp->nm_rpcops->nr_invaldir)(vp);
392 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
393 if (error)
394 goto out;
395 }
396 np->n_attrstamp = 0;
397 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
398 error = VOP_GETATTR(vp, &vattr, cred);
399 if (error)
400 goto out;
401 mtx_lock(&np->n_mtx);
402 np->n_mtime = vattr.va_mtime;
403 mtx_unlock(&np->n_mtx);
404 } else {
405 mtx_unlock(&np->n_mtx);
406 error = VOP_GETATTR(vp, &vattr, cred);
407 if (error)
408 return (error);
409 mtx_lock(&np->n_mtx);
410 if ((np->n_flag & NSIZECHANGED)
411 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
412 mtx_unlock(&np->n_mtx);
413 if (vp->v_type == VDIR)
414 (nmp->nm_rpcops->nr_invaldir)(vp);
415 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
416 if (error)
417 goto out;
418 mtx_lock(&np->n_mtx);
419 np->n_mtime = vattr.va_mtime;
420 np->n_flag &= ~NSIZECHANGED;
421 }
422 mtx_unlock(&np->n_mtx);
423 }
424 out:
425 nfs_downgrade_vnlock(vp, old_lock);
426 return error;
427 }
428
429 /*
430 * Vnode op for read using bio
431 */
432 int
433 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
434 {
435 struct nfsnode *np = VTONFS(vp);
436 int biosize, i;
437 struct buf *bp, *rabp;
438 struct thread *td;
439 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
440 daddr_t lbn, rabn;
441 off_t end;
442 int bcount;
443 int seqcount;
444 int nra, error = 0, n = 0, on = 0;
445
446 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
447 if (uio->uio_resid == 0)
448 return (0);
449 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
450 return (EINVAL);
451 td = uio->uio_td;
452
453 mtx_lock(&nmp->nm_mtx);
454 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
455 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
456 mtx_unlock(&nmp->nm_mtx);
457 (void)nfs_fsinfo(nmp, vp, cred, td);
458 } else
459 mtx_unlock(&nmp->nm_mtx);
460
461 end = uio->uio_offset + uio->uio_resid;
462 if (vp->v_type != VDIR &&
463 (end > nmp->nm_maxfilesize || end < uio->uio_offset))
464 return (EFBIG);
465
466 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
467 /* No caching/ no readaheads. Just read data into the user buffer */
468 return nfs_readrpc(vp, uio, cred);
469
470 biosize = vp->v_bufobj.bo_bsize;
471 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
472
473 error = nfs_bioread_check_cons(vp, td, cred);
474 if (error)
475 return error;
476
477 do {
478 u_quad_t nsize;
479
480 mtx_lock(&np->n_mtx);
481 nsize = np->n_size;
482 mtx_unlock(&np->n_mtx);
483
484 switch (vp->v_type) {
485 case VREG:
486 nfsstats.biocache_reads++;
487 lbn = uio->uio_offset / biosize;
488 on = uio->uio_offset & (biosize - 1);
489
490 /*
491 * Start the read ahead(s), as required.
492 */
493 if (nmp->nm_readahead > 0) {
494 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
495 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
496 rabn = lbn + 1 + nra;
497 if (incore(&vp->v_bufobj, rabn) == NULL) {
498 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
499 if (!rabp) {
500 error = nfs_sigintr(nmp, td);
501 return (error ? error : EINTR);
502 }
503 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
504 rabp->b_flags |= B_ASYNC;
505 rabp->b_iocmd = BIO_READ;
506 vfs_busy_pages(rabp, 0);
507 if (nfs_asyncio(nmp, rabp, cred, td)) {
508 rabp->b_flags |= B_INVAL;
509 rabp->b_ioflags |= BIO_ERROR;
510 vfs_unbusy_pages(rabp);
511 brelse(rabp);
512 break;
513 }
514 } else {
515 brelse(rabp);
516 }
517 }
518 }
519 }
520
521 /* Note that bcount is *not* DEV_BSIZE aligned. */
522 bcount = biosize;
523 if ((off_t)lbn * biosize >= nsize) {
524 bcount = 0;
525 } else if ((off_t)(lbn + 1) * biosize > nsize) {
526 bcount = nsize - (off_t)lbn * biosize;
527 }
528 bp = nfs_getcacheblk(vp, lbn, bcount, td);
529
530 if (!bp) {
531 error = nfs_sigintr(nmp, td);
532 return (error ? error : EINTR);
533 }
534
535 /*
536 * If B_CACHE is not set, we must issue the read. If this
537 * fails, we return an error.
538 */
539
540 if ((bp->b_flags & B_CACHE) == 0) {
541 bp->b_iocmd = BIO_READ;
542 vfs_busy_pages(bp, 0);
543 error = nfs_doio(vp, bp, cred, td);
544 if (error) {
545 brelse(bp);
546 return (error);
547 }
548 }
549
550 /*
551 * on is the offset into the current bp. Figure out how many
552 * bytes we can copy out of the bp. Note that bcount is
553 * NOT DEV_BSIZE aligned.
554 *
555 * Then figure out how many bytes we can copy into the uio.
556 */
557
558 n = 0;
559 if (on < bcount)
560 n = min((unsigned)(bcount - on), uio->uio_resid);
561 break;
562 case VLNK:
563 nfsstats.biocache_readlinks++;
564 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
565 if (!bp) {
566 error = nfs_sigintr(nmp, td);
567 return (error ? error : EINTR);
568 }
569 if ((bp->b_flags & B_CACHE) == 0) {
570 bp->b_iocmd = BIO_READ;
571 vfs_busy_pages(bp, 0);
572 error = nfs_doio(vp, bp, cred, td);
573 if (error) {
574 bp->b_ioflags |= BIO_ERROR;
575 brelse(bp);
576 return (error);
577 }
578 }
579 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
580 on = 0;
581 break;
582 case VDIR:
583 nfsstats.biocache_readdirs++;
584 if (np->n_direofoffset
585 && uio->uio_offset >= np->n_direofoffset) {
586 return (0);
587 }
588 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
589 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
590 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
591 if (!bp) {
592 error = nfs_sigintr(nmp, td);
593 return (error ? error : EINTR);
594 }
595 if ((bp->b_flags & B_CACHE) == 0) {
596 bp->b_iocmd = BIO_READ;
597 vfs_busy_pages(bp, 0);
598 error = nfs_doio(vp, bp, cred, td);
599 if (error) {
600 brelse(bp);
601 }
602 while (error == NFSERR_BAD_COOKIE) {
603 (nmp->nm_rpcops->nr_invaldir)(vp);
604 error = nfs_vinvalbuf(vp, 0, td, 1);
605 /*
606 * Yuck! The directory has been modified on the
607 * server. The only way to get the block is by
608 * reading from the beginning to get all the
609 * offset cookies.
610 *
611 * Leave the last bp intact unless there is an error.
612 * Loop back up to the while if the error is another
613 * NFSERR_BAD_COOKIE (double yuch!).
614 */
615 for (i = 0; i <= lbn && !error; i++) {
616 if (np->n_direofoffset
617 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
618 return (0);
619 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
620 if (!bp) {
621 error = nfs_sigintr(nmp, td);
622 return (error ? error : EINTR);
623 }
624 if ((bp->b_flags & B_CACHE) == 0) {
625 bp->b_iocmd = BIO_READ;
626 vfs_busy_pages(bp, 0);
627 error = nfs_doio(vp, bp, cred, td);
628 /*
629 * no error + B_INVAL == directory EOF,
630 * use the block.
631 */
632 if (error == 0 && (bp->b_flags & B_INVAL))
633 break;
634 }
635 /*
636 * An error will throw away the block and the
637 * for loop will break out. If no error and this
638 * is not the block we want, we throw away the
639 * block and go for the next one via the for loop.
640 */
641 if (error || i < lbn)
642 brelse(bp);
643 }
644 }
645 /*
646 * The above while is repeated if we hit another cookie
647 * error. If we hit an error and it wasn't a cookie error,
648 * we give up.
649 */
650 if (error)
651 return (error);
652 }
653
654 /*
655 * If not eof and read aheads are enabled, start one.
656 * (You need the current block first, so that you have the
657 * directory offset cookie of the next block.)
658 */
659 if (nmp->nm_readahead > 0 &&
660 (bp->b_flags & B_INVAL) == 0 &&
661 (np->n_direofoffset == 0 ||
662 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
663 incore(&vp->v_bufobj, lbn + 1) == NULL) {
664 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
665 if (rabp) {
666 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
667 rabp->b_flags |= B_ASYNC;
668 rabp->b_iocmd = BIO_READ;
669 vfs_busy_pages(rabp, 0);
670 if (nfs_asyncio(nmp, rabp, cred, td)) {
671 rabp->b_flags |= B_INVAL;
672 rabp->b_ioflags |= BIO_ERROR;
673 vfs_unbusy_pages(rabp);
674 brelse(rabp);
675 }
676 } else {
677 brelse(rabp);
678 }
679 }
680 }
681 /*
682 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
683 * chopped for the EOF condition, we cannot tell how large
684 * NFS directories are going to be until we hit EOF. So
685 * an NFS directory buffer is *not* chopped to its EOF. Now,
686 * it just so happens that b_resid will effectively chop it
687 * to EOF. *BUT* this information is lost if the buffer goes
688 * away and is reconstituted into a B_CACHE state ( due to
689 * being VMIO ) later. So we keep track of the directory eof
690 * in np->n_direofoffset and chop it off as an extra step
691 * right here.
692 */
693 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
694 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
695 n = np->n_direofoffset - uio->uio_offset;
696 break;
697 default:
698 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
699 bp = NULL;
700 break;
701 };
702
703 if (n > 0) {
704 error = uiomove(bp->b_data + on, (int)n, uio);
705 }
706 if (vp->v_type == VLNK)
707 n = 0;
708 if (bp != NULL)
709 brelse(bp);
710 } while (error == 0 && uio->uio_resid > 0 && n > 0);
711 return (error);
712 }
713
714 /*
715 * The NFS write path cannot handle iovecs with len > 1. So we need to
716 * break up iovecs accordingly (restricting them to wsize).
717 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
718 * For the ASYNC case, 2 copies are needed. The first a copy from the
719 * user buffer to a staging buffer and then a second copy from the staging
720 * buffer to mbufs. This can be optimized by copying from the user buffer
721 * directly into mbufs and passing the chain down, but that requires a
722 * fair amount of re-working of the relevant codepaths (and can be done
723 * later).
724 */
725 static int
726 nfs_directio_write(vp, uiop, cred, ioflag)
727 struct vnode *vp;
728 struct uio *uiop;
729 struct ucred *cred;
730 int ioflag;
731 {
732 int error;
733 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
734 struct thread *td = uiop->uio_td;
735 int size;
736 int wsize;
737
738 mtx_lock(&nmp->nm_mtx);
739 wsize = nmp->nm_wsize;
740 mtx_unlock(&nmp->nm_mtx);
741 if (ioflag & IO_SYNC) {
742 int iomode, must_commit;
743 struct uio uio;
744 struct iovec iov;
745 do_sync:
746 while (uiop->uio_resid > 0) {
747 size = min(uiop->uio_resid, wsize);
748 size = min(uiop->uio_iov->iov_len, size);
749 iov.iov_base = uiop->uio_iov->iov_base;
750 iov.iov_len = size;
751 uio.uio_iov = &iov;
752 uio.uio_iovcnt = 1;
753 uio.uio_offset = uiop->uio_offset;
754 uio.uio_resid = size;
755 uio.uio_segflg = UIO_USERSPACE;
756 uio.uio_rw = UIO_WRITE;
757 uio.uio_td = td;
758 iomode = NFSV3WRITE_FILESYNC;
759 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
760 &iomode, &must_commit);
761 KASSERT((must_commit == 0),
762 ("nfs_directio_write: Did not commit write"));
763 if (error)
764 return (error);
765 uiop->uio_offset += size;
766 uiop->uio_resid -= size;
767 if (uiop->uio_iov->iov_len <= size) {
768 uiop->uio_iovcnt--;
769 uiop->uio_iov++;
770 } else {
771 uiop->uio_iov->iov_base =
772 (char *)uiop->uio_iov->iov_base + size;
773 uiop->uio_iov->iov_len -= size;
774 }
775 }
776 } else {
777 struct uio *t_uio;
778 struct iovec *t_iov;
779 struct buf *bp;
780
781 /*
782 * Break up the write into blocksize chunks and hand these
783 * over to nfsiod's for write back.
784 * Unfortunately, this incurs a copy of the data. Since
785 * the user could modify the buffer before the write is
786 * initiated.
787 *
788 * The obvious optimization here is that one of the 2 copies
789 * in the async write path can be eliminated by copying the
790 * data here directly into mbufs and passing the mbuf chain
791 * down. But that will require a fair amount of re-working
792 * of the code and can be done if there's enough interest
793 * in NFS directio access.
794 */
795 while (uiop->uio_resid > 0) {
796 size = min(uiop->uio_resid, wsize);
797 size = min(uiop->uio_iov->iov_len, size);
798 bp = getpbuf(&nfs_pbuf_freecnt);
799 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
800 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
801 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
802 t_iov->iov_len = size;
803 t_uio->uio_iov = t_iov;
804 t_uio->uio_iovcnt = 1;
805 t_uio->uio_offset = uiop->uio_offset;
806 t_uio->uio_resid = size;
807 t_uio->uio_segflg = UIO_SYSSPACE;
808 t_uio->uio_rw = UIO_WRITE;
809 t_uio->uio_td = td;
810 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size);
811 bp->b_flags |= B_DIRECT;
812 bp->b_iocmd = BIO_WRITE;
813 if (cred != NOCRED) {
814 crhold(cred);
815 bp->b_wcred = cred;
816 } else
817 bp->b_wcred = NOCRED;
818 bp->b_caller1 = (void *)t_uio;
819 bp->b_vp = vp;
820 error = nfs_asyncio(nmp, bp, NOCRED, td);
821 if (error) {
822 free(t_iov->iov_base, M_NFSDIRECTIO);
823 free(t_iov, M_NFSDIRECTIO);
824 free(t_uio, M_NFSDIRECTIO);
825 bp->b_vp = NULL;
826 relpbuf(bp, &nfs_pbuf_freecnt);
827 if (error == EINTR)
828 return (error);
829 goto do_sync;
830 }
831 uiop->uio_offset += size;
832 uiop->uio_resid -= size;
833 if (uiop->uio_iov->iov_len <= size) {
834 uiop->uio_iovcnt--;
835 uiop->uio_iov++;
836 } else {
837 uiop->uio_iov->iov_base =
838 (char *)uiop->uio_iov->iov_base + size;
839 uiop->uio_iov->iov_len -= size;
840 }
841 }
842 }
843 return (0);
844 }
845
846 /*
847 * Vnode op for write using bio
848 */
849 int
850 nfs_write(struct vop_write_args *ap)
851 {
852 int biosize;
853 struct uio *uio = ap->a_uio;
854 struct thread *td = uio->uio_td;
855 struct vnode *vp = ap->a_vp;
856 struct nfsnode *np = VTONFS(vp);
857 struct ucred *cred = ap->a_cred;
858 int ioflag = ap->a_ioflag;
859 struct buf *bp;
860 struct vattr vattr;
861 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
862 daddr_t lbn;
863 off_t end;
864 int bcount;
865 int n, on, error = 0;
866
867 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode"));
868 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
869 ("nfs_write proc"));
870 if (vp->v_type != VREG)
871 return (EIO);
872 mtx_lock(&np->n_mtx);
873 if (np->n_flag & NWRITEERR) {
874 np->n_flag &= ~NWRITEERR;
875 mtx_unlock(&np->n_mtx);
876 return (np->n_error);
877 } else
878 mtx_unlock(&np->n_mtx);
879 mtx_lock(&nmp->nm_mtx);
880 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
881 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
882 mtx_unlock(&nmp->nm_mtx);
883 (void)nfs_fsinfo(nmp, vp, cred, td);
884 } else
885 mtx_unlock(&nmp->nm_mtx);
886
887 /*
888 * Synchronously flush pending buffers if we are in synchronous
889 * mode or if we are appending.
890 */
891 if (ioflag & (IO_APPEND | IO_SYNC)) {
892 mtx_lock(&np->n_mtx);
893 if (np->n_flag & NMODIFIED) {
894 mtx_unlock(&np->n_mtx);
895 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
896 /*
897 * Require non-blocking, synchronous writes to
898 * dirty files to inform the program it needs
899 * to fsync(2) explicitly.
900 */
901 if (ioflag & IO_NDELAY)
902 return (EAGAIN);
903 #endif
904 flush_and_restart:
905 np->n_attrstamp = 0;
906 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
907 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
908 if (error)
909 return (error);
910 } else
911 mtx_unlock(&np->n_mtx);
912 }
913
914 /*
915 * If IO_APPEND then load uio_offset. We restart here if we cannot
916 * get the append lock.
917 */
918 if (ioflag & IO_APPEND) {
919 np->n_attrstamp = 0;
920 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
921 error = VOP_GETATTR(vp, &vattr, cred);
922 if (error)
923 return (error);
924 mtx_lock(&np->n_mtx);
925 uio->uio_offset = np->n_size;
926 mtx_unlock(&np->n_mtx);
927 }
928
929 if (uio->uio_offset < 0)
930 return (EINVAL);
931 end = uio->uio_offset + uio->uio_resid;
932 if (end > nmp->nm_maxfilesize || end < uio->uio_offset)
933 return (EFBIG);
934 if (uio->uio_resid == 0)
935 return (0);
936
937 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
938 return nfs_directio_write(vp, uio, cred, ioflag);
939
940 /*
941 * Maybe this should be above the vnode op call, but so long as
942 * file servers have no limits, i don't think it matters
943 */
944 if (vn_rlimit_fsize(vp, uio, td))
945 return (EFBIG);
946
947 biosize = vp->v_bufobj.bo_bsize;
948 /*
949 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
950 * would exceed the local maximum per-file write commit size when
951 * combined with those, we must decide whether to flush,
952 * go synchronous, or return error. We don't bother checking
953 * IO_UNIT -- we just make all writes atomic anyway, as there's
954 * no point optimizing for something that really won't ever happen.
955 */
956 if (!(ioflag & IO_SYNC)) {
957 int nflag;
958
959 mtx_lock(&np->n_mtx);
960 nflag = np->n_flag;
961 mtx_unlock(&np->n_mtx);
962 int needrestart = 0;
963 if (nmp->nm_wcommitsize < uio->uio_resid) {
964 /*
965 * If this request could not possibly be completed
966 * without exceeding the maximum outstanding write
967 * commit size, see if we can convert it into a
968 * synchronous write operation.
969 */
970 if (ioflag & IO_NDELAY)
971 return (EAGAIN);
972 ioflag |= IO_SYNC;
973 if (nflag & NMODIFIED)
974 needrestart = 1;
975 } else if (nflag & NMODIFIED) {
976 int wouldcommit = 0;
977 BO_LOCK(&vp->v_bufobj);
978 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
979 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
980 b_bobufs) {
981 if (bp->b_flags & B_NEEDCOMMIT)
982 wouldcommit += bp->b_bcount;
983 }
984 }
985 BO_UNLOCK(&vp->v_bufobj);
986 /*
987 * Since we're not operating synchronously and
988 * bypassing the buffer cache, we are in a commit
989 * and holding all of these buffers whether
990 * transmitted or not. If not limited, this
991 * will lead to the buffer cache deadlocking,
992 * as no one else can flush our uncommitted buffers.
993 */
994 wouldcommit += uio->uio_resid;
995 /*
996 * If we would initially exceed the maximum
997 * outstanding write commit size, flush and restart.
998 */
999 if (wouldcommit > nmp->nm_wcommitsize)
1000 needrestart = 1;
1001 }
1002 if (needrestart)
1003 goto flush_and_restart;
1004 }
1005
1006 do {
1007 nfsstats.biocache_writes++;
1008 lbn = uio->uio_offset / biosize;
1009 on = uio->uio_offset & (biosize-1);
1010 n = min((unsigned)(biosize - on), uio->uio_resid);
1011 again:
1012 /*
1013 * Handle direct append and file extension cases, calculate
1014 * unaligned buffer size.
1015 */
1016 mtx_lock(&np->n_mtx);
1017 if (uio->uio_offset == np->n_size && n) {
1018 mtx_unlock(&np->n_mtx);
1019 /*
1020 * Get the buffer (in its pre-append state to maintain
1021 * B_CACHE if it was previously set). Resize the
1022 * nfsnode after we have locked the buffer to prevent
1023 * readers from reading garbage.
1024 */
1025 bcount = on;
1026 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1027
1028 if (bp != NULL) {
1029 long save;
1030
1031 mtx_lock(&np->n_mtx);
1032 np->n_size = uio->uio_offset + n;
1033 np->n_flag |= NMODIFIED;
1034 vnode_pager_setsize(vp, np->n_size);
1035 mtx_unlock(&np->n_mtx);
1036
1037 save = bp->b_flags & B_CACHE;
1038 bcount += n;
1039 allocbuf(bp, bcount);
1040 bp->b_flags |= save;
1041 }
1042 } else {
1043 /*
1044 * Obtain the locked cache block first, and then
1045 * adjust the file's size as appropriate.
1046 */
1047 bcount = on + n;
1048 if ((off_t)lbn * biosize + bcount < np->n_size) {
1049 if ((off_t)(lbn + 1) * biosize < np->n_size)
1050 bcount = biosize;
1051 else
1052 bcount = np->n_size - (off_t)lbn * biosize;
1053 }
1054 mtx_unlock(&np->n_mtx);
1055 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1056 mtx_lock(&np->n_mtx);
1057 if (uio->uio_offset + n > np->n_size) {
1058 np->n_size = uio->uio_offset + n;
1059 np->n_flag |= NMODIFIED;
1060 vnode_pager_setsize(vp, np->n_size);
1061 }
1062 mtx_unlock(&np->n_mtx);
1063 }
1064
1065 if (!bp) {
1066 error = nfs_sigintr(nmp, td);
1067 if (!error)
1068 error = EINTR;
1069 break;
1070 }
1071
1072 /*
1073 * Issue a READ if B_CACHE is not set. In special-append
1074 * mode, B_CACHE is based on the buffer prior to the write
1075 * op and is typically set, avoiding the read. If a read
1076 * is required in special append mode, the server will
1077 * probably send us a short-read since we extended the file
1078 * on our end, resulting in b_resid == 0 and, thusly,
1079 * B_CACHE getting set.
1080 *
1081 * We can also avoid issuing the read if the write covers
1082 * the entire buffer. We have to make sure the buffer state
1083 * is reasonable in this case since we will not be initiating
1084 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1085 * more information.
1086 *
1087 * B_CACHE may also be set due to the buffer being cached
1088 * normally.
1089 */
1090
1091 if (on == 0 && n == bcount) {
1092 bp->b_flags |= B_CACHE;
1093 bp->b_flags &= ~B_INVAL;
1094 bp->b_ioflags &= ~BIO_ERROR;
1095 }
1096
1097 if ((bp->b_flags & B_CACHE) == 0) {
1098 bp->b_iocmd = BIO_READ;
1099 vfs_busy_pages(bp, 0);
1100 error = nfs_doio(vp, bp, cred, td);
1101 if (error) {
1102 brelse(bp);
1103 break;
1104 }
1105 }
1106 if (bp->b_wcred == NOCRED)
1107 bp->b_wcred = crhold(cred);
1108 mtx_lock(&np->n_mtx);
1109 np->n_flag |= NMODIFIED;
1110 mtx_unlock(&np->n_mtx);
1111
1112 /*
1113 * If dirtyend exceeds file size, chop it down. This should
1114 * not normally occur but there is an append race where it
1115 * might occur XXX, so we log it.
1116 *
1117 * If the chopping creates a reverse-indexed or degenerate
1118 * situation with dirtyoff/end, we 0 both of them.
1119 */
1120
1121 if (bp->b_dirtyend > bcount) {
1122 nfs_printf("NFS append race @%lx:%d\n",
1123 (long)bp->b_blkno * DEV_BSIZE,
1124 bp->b_dirtyend - bcount);
1125 bp->b_dirtyend = bcount;
1126 }
1127
1128 if (bp->b_dirtyoff >= bp->b_dirtyend)
1129 bp->b_dirtyoff = bp->b_dirtyend = 0;
1130
1131 /*
1132 * If the new write will leave a contiguous dirty
1133 * area, just update the b_dirtyoff and b_dirtyend,
1134 * otherwise force a write rpc of the old dirty area.
1135 *
1136 * While it is possible to merge discontiguous writes due to
1137 * our having a B_CACHE buffer ( and thus valid read data
1138 * for the hole), we don't because it could lead to
1139 * significant cache coherency problems with multiple clients,
1140 * especially if locking is implemented later on.
1141 *
1142 * as an optimization we could theoretically maintain
1143 * a linked list of discontinuous areas, but we would still
1144 * have to commit them separately so there isn't much
1145 * advantage to it except perhaps a bit of asynchronization.
1146 */
1147
1148 if (bp->b_dirtyend > 0 &&
1149 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1150 if (bwrite(bp) == EINTR) {
1151 error = EINTR;
1152 break;
1153 }
1154 goto again;
1155 }
1156
1157 error = uiomove((char *)bp->b_data + on, n, uio);
1158
1159 /*
1160 * Since this block is being modified, it must be written
1161 * again and not just committed. Since write clustering does
1162 * not work for the stage 1 data write, only the stage 2
1163 * commit rpc, we have to clear B_CLUSTEROK as well.
1164 */
1165 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1166
1167 if (error) {
1168 bp->b_ioflags |= BIO_ERROR;
1169 brelse(bp);
1170 break;
1171 }
1172
1173 /*
1174 * Only update dirtyoff/dirtyend if not a degenerate
1175 * condition.
1176 */
1177 if (n) {
1178 if (bp->b_dirtyend > 0) {
1179 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1180 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1181 } else {
1182 bp->b_dirtyoff = on;
1183 bp->b_dirtyend = on + n;
1184 }
1185 vfs_bio_set_valid(bp, on, n);
1186 }
1187
1188 /*
1189 * If IO_SYNC do bwrite().
1190 *
1191 * IO_INVAL appears to be unused. The idea appears to be
1192 * to turn off caching in this case. Very odd. XXX
1193 */
1194 if ((ioflag & IO_SYNC)) {
1195 if (ioflag & IO_INVAL)
1196 bp->b_flags |= B_NOCACHE;
1197 error = bwrite(bp);
1198 if (error)
1199 break;
1200 } else if ((n + on) == biosize) {
1201 bp->b_flags |= B_ASYNC;
1202 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1203 } else {
1204 bdwrite(bp);
1205 }
1206 } while (uio->uio_resid > 0 && n > 0);
1207
1208 return (error);
1209 }
1210
1211 /*
1212 * Get an nfs cache block.
1213 *
1214 * Allocate a new one if the block isn't currently in the cache
1215 * and return the block marked busy. If the calling process is
1216 * interrupted by a signal for an interruptible mount point, return
1217 * NULL.
1218 *
1219 * The caller must carefully deal with the possible B_INVAL state of
1220 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1221 * indirectly), so synchronous reads can be issued without worrying about
1222 * the B_INVAL state. We have to be a little more careful when dealing
1223 * with writes (see comments in nfs_write()) when extending a file past
1224 * its EOF.
1225 */
1226 static struct buf *
1227 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1228 {
1229 struct buf *bp;
1230 struct mount *mp;
1231 struct nfsmount *nmp;
1232
1233 mp = vp->v_mount;
1234 nmp = VFSTONFS(mp);
1235
1236 if (nmp->nm_flag & NFSMNT_INT) {
1237 sigset_t oldset;
1238
1239 nfs_set_sigmask(td, &oldset);
1240 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1241 nfs_restore_sigmask(td, &oldset);
1242 while (bp == NULL) {
1243 if (nfs_sigintr(nmp, td))
1244 return (NULL);
1245 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1246 }
1247 } else {
1248 bp = getblk(vp, bn, size, 0, 0, 0);
1249 }
1250
1251 if (vp->v_type == VREG)
1252 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1253 return (bp);
1254 }
1255
1256 /*
1257 * Flush and invalidate all dirty buffers. If another process is already
1258 * doing the flush, just wait for completion.
1259 */
1260 int
1261 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1262 {
1263 struct nfsnode *np = VTONFS(vp);
1264 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1265 int error = 0, slpflag, slptimeo;
1266 int old_lock = 0;
1267
1268 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1269
1270 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1271 intrflg = 0;
1272 if (intrflg) {
1273 slpflag = NFS_PCATCH;
1274 slptimeo = 2 * hz;
1275 } else {
1276 slpflag = 0;
1277 slptimeo = 0;
1278 }
1279
1280 old_lock = nfs_upgrade_vnlock(vp);
1281 if (vp->v_iflag & VI_DOOMED) {
1282 /*
1283 * Since vgonel() uses the generic vinvalbuf() to flush
1284 * dirty buffers and it does not call this function, it
1285 * is safe to just return OK when VI_DOOMED is set.
1286 */
1287 nfs_downgrade_vnlock(vp, old_lock);
1288 return (0);
1289 }
1290
1291 /*
1292 * Now, flush as required.
1293 */
1294 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1295 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1296 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1297 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1298 /*
1299 * If the page clean was interrupted, fail the invalidation.
1300 * Not doing so, we run the risk of losing dirty pages in the
1301 * vinvalbuf() call below.
1302 */
1303 if (intrflg && (error = nfs_sigintr(nmp, td)))
1304 goto out;
1305 }
1306
1307 error = vinvalbuf(vp, flags, slpflag, 0);
1308 while (error) {
1309 if (intrflg && (error = nfs_sigintr(nmp, td)))
1310 goto out;
1311 error = vinvalbuf(vp, flags, 0, slptimeo);
1312 }
1313 mtx_lock(&np->n_mtx);
1314 if (np->n_directio_asyncwr == 0)
1315 np->n_flag &= ~NMODIFIED;
1316 mtx_unlock(&np->n_mtx);
1317 out:
1318 nfs_downgrade_vnlock(vp, old_lock);
1319 return error;
1320 }
1321
1322 /*
1323 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1324 * This is mainly to avoid queueing async I/O requests when the nfsiods
1325 * are all hung on a dead server.
1326 *
1327 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1328 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1329 */
1330 int
1331 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1332 {
1333 int iod;
1334 int gotiod;
1335 int slpflag = 0;
1336 int slptimeo = 0;
1337 int error, error2;
1338
1339 /*
1340 * Commits are usually short and sweet so lets save some cpu and
1341 * leave the async daemons for more important rpc's (such as reads
1342 * and writes).
1343 */
1344 mtx_lock(&nfs_iod_mtx);
1345 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1346 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1347 mtx_unlock(&nfs_iod_mtx);
1348 return(EIO);
1349 }
1350 again:
1351 if (nmp->nm_flag & NFSMNT_INT)
1352 slpflag = NFS_PCATCH;
1353 gotiod = FALSE;
1354
1355 /*
1356 * Find a free iod to process this request.
1357 */
1358 for (iod = 0; iod < nfs_numasync; iod++)
1359 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1360 gotiod = TRUE;
1361 break;
1362 }
1363
1364 /*
1365 * Try to create one if none are free.
1366 */
1367 if (!gotiod)
1368 nfs_nfsiodnew();
1369 else {
1370 /*
1371 * Found one, so wake it up and tell it which
1372 * mount to process.
1373 */
1374 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1375 iod, nmp));
1376 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1377 nfs_iodmount[iod] = nmp;
1378 nmp->nm_bufqiods++;
1379 wakeup(&nfs_iodwant[iod]);
1380 }
1381
1382 /*
1383 * If none are free, we may already have an iod working on this mount
1384 * point. If so, it will process our request.
1385 */
1386 if (!gotiod) {
1387 if (nmp->nm_bufqiods > 0) {
1388 NFS_DPF(ASYNCIO,
1389 ("nfs_asyncio: %d iods are already processing mount %p\n",
1390 nmp->nm_bufqiods, nmp));
1391 gotiod = TRUE;
1392 }
1393 }
1394
1395 /*
1396 * If we have an iod which can process the request, then queue
1397 * the buffer.
1398 */
1399 if (gotiod) {
1400 /*
1401 * Ensure that the queue never grows too large. We still want
1402 * to asynchronize so we block rather then return EIO.
1403 */
1404 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1405 NFS_DPF(ASYNCIO,
1406 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1407 nmp->nm_bufqwant = TRUE;
1408 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1409 slpflag | PRIBIO,
1410 "nfsaio", slptimeo);
1411 if (error) {
1412 error2 = nfs_sigintr(nmp, td);
1413 if (error2) {
1414 mtx_unlock(&nfs_iod_mtx);
1415 return (error2);
1416 }
1417 if (slpflag == NFS_PCATCH) {
1418 slpflag = 0;
1419 slptimeo = 2 * hz;
1420 }
1421 }
1422 /*
1423 * We might have lost our iod while sleeping,
1424 * so check and loop if nescessary.
1425 */
1426 goto again;
1427 }
1428
1429 /* We might have lost our nfsiod */
1430 if (nmp->nm_bufqiods == 0) {
1431 NFS_DPF(ASYNCIO,
1432 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1433 goto again;
1434 }
1435
1436 if (bp->b_iocmd == BIO_READ) {
1437 if (bp->b_rcred == NOCRED && cred != NOCRED)
1438 bp->b_rcred = crhold(cred);
1439 } else {
1440 if (bp->b_wcred == NOCRED && cred != NOCRED)
1441 bp->b_wcred = crhold(cred);
1442 }
1443
1444 if (bp->b_flags & B_REMFREE)
1445 bremfreef(bp);
1446 BUF_KERNPROC(bp);
1447 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1448 nmp->nm_bufqlen++;
1449 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1450 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1451 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1452 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1453 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1454 }
1455 mtx_unlock(&nfs_iod_mtx);
1456 return (0);
1457 }
1458
1459 mtx_unlock(&nfs_iod_mtx);
1460
1461 /*
1462 * All the iods are busy on other mounts, so return EIO to
1463 * force the caller to process the i/o synchronously.
1464 */
1465 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1466 return (EIO);
1467 }
1468
1469 void
1470 nfs_doio_directwrite(struct buf *bp)
1471 {
1472 int iomode, must_commit;
1473 struct uio *uiop = (struct uio *)bp->b_caller1;
1474 char *iov_base = uiop->uio_iov->iov_base;
1475 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1476
1477 iomode = NFSV3WRITE_FILESYNC;
1478 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1479 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1480 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1481 free(iov_base, M_NFSDIRECTIO);
1482 free(uiop->uio_iov, M_NFSDIRECTIO);
1483 free(uiop, M_NFSDIRECTIO);
1484 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1485 struct nfsnode *np = VTONFS(bp->b_vp);
1486 mtx_lock(&np->n_mtx);
1487 np->n_directio_asyncwr--;
1488 if (np->n_directio_asyncwr == 0) {
1489 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1490 if ((np->n_flag & NFSYNCWAIT)) {
1491 np->n_flag &= ~NFSYNCWAIT;
1492 wakeup((caddr_t)&np->n_directio_asyncwr);
1493 }
1494 }
1495 mtx_unlock(&np->n_mtx);
1496 }
1497 bp->b_vp = NULL;
1498 relpbuf(bp, &nfs_pbuf_freecnt);
1499 }
1500
1501 /*
1502 * Do an I/O operation to/from a cache block. This may be called
1503 * synchronously or from an nfsiod.
1504 */
1505 int
1506 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1507 {
1508 struct uio *uiop;
1509 struct nfsnode *np;
1510 struct nfsmount *nmp;
1511 int error = 0, iomode, must_commit = 0;
1512 struct uio uio;
1513 struct iovec io;
1514 struct proc *p = td ? td->td_proc : NULL;
1515 uint8_t iocmd;
1516
1517 np = VTONFS(vp);
1518 nmp = VFSTONFS(vp->v_mount);
1519 uiop = &uio;
1520 uiop->uio_iov = &io;
1521 uiop->uio_iovcnt = 1;
1522 uiop->uio_segflg = UIO_SYSSPACE;
1523 uiop->uio_td = td;
1524
1525 /*
1526 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1527 * do this here so we do not have to do it in all the code that
1528 * calls us.
1529 */
1530 bp->b_flags &= ~B_INVAL;
1531 bp->b_ioflags &= ~BIO_ERROR;
1532
1533 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1534 iocmd = bp->b_iocmd;
1535 if (iocmd == BIO_READ) {
1536 io.iov_len = uiop->uio_resid = bp->b_bcount;
1537 io.iov_base = bp->b_data;
1538 uiop->uio_rw = UIO_READ;
1539
1540 switch (vp->v_type) {
1541 case VREG:
1542 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1543 nfsstats.read_bios++;
1544 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1545
1546 if (!error) {
1547 if (uiop->uio_resid) {
1548 /*
1549 * If we had a short read with no error, we must have
1550 * hit a file hole. We should zero-fill the remainder.
1551 * This can also occur if the server hits the file EOF.
1552 *
1553 * Holes used to be able to occur due to pending
1554 * writes, but that is not possible any longer.
1555 */
1556 int nread = bp->b_bcount - uiop->uio_resid;
1557 int left = uiop->uio_resid;
1558
1559 if (left > 0)
1560 bzero((char *)bp->b_data + nread, left);
1561 uiop->uio_resid = 0;
1562 }
1563 }
1564 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1565 if (p && (vp->v_vflag & VV_TEXT)) {
1566 mtx_lock(&np->n_mtx);
1567 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1568 mtx_unlock(&np->n_mtx);
1569 PROC_LOCK(p);
1570 killproc(p, "text file modification");
1571 PROC_UNLOCK(p);
1572 } else
1573 mtx_unlock(&np->n_mtx);
1574 }
1575 break;
1576 case VLNK:
1577 uiop->uio_offset = (off_t)0;
1578 nfsstats.readlink_bios++;
1579 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1580 break;
1581 case VDIR:
1582 nfsstats.readdir_bios++;
1583 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1584 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1585 error = nfs_readdirplusrpc(vp, uiop, cr);
1586 if (error == NFSERR_NOTSUPP)
1587 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1588 }
1589 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1590 error = nfs_readdirrpc(vp, uiop, cr);
1591 /*
1592 * end-of-directory sets B_INVAL but does not generate an
1593 * error.
1594 */
1595 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1596 bp->b_flags |= B_INVAL;
1597 break;
1598 default:
1599 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1600 break;
1601 };
1602 if (error) {
1603 bp->b_ioflags |= BIO_ERROR;
1604 bp->b_error = error;
1605 }
1606 } else {
1607 /*
1608 * If we only need to commit, try to commit
1609 */
1610 if (bp->b_flags & B_NEEDCOMMIT) {
1611 int retv;
1612 off_t off;
1613
1614 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1615 retv = (nmp->nm_rpcops->nr_commit)(
1616 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1617 bp->b_wcred, td);
1618 if (retv == 0) {
1619 bp->b_dirtyoff = bp->b_dirtyend = 0;
1620 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1621 bp->b_resid = 0;
1622 bufdone(bp);
1623 return (0);
1624 }
1625 if (retv == NFSERR_STALEWRITEVERF) {
1626 nfs_clearcommit(vp->v_mount);
1627 }
1628 }
1629
1630 /*
1631 * Setup for actual write
1632 */
1633 mtx_lock(&np->n_mtx);
1634 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1635 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1636 mtx_unlock(&np->n_mtx);
1637
1638 if (bp->b_dirtyend > bp->b_dirtyoff) {
1639 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1640 - bp->b_dirtyoff;
1641 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1642 + bp->b_dirtyoff;
1643 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1644 uiop->uio_rw = UIO_WRITE;
1645 nfsstats.write_bios++;
1646
1647 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1648 iomode = NFSV3WRITE_UNSTABLE;
1649 else
1650 iomode = NFSV3WRITE_FILESYNC;
1651
1652 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1653
1654 /*
1655 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1656 * to cluster the buffers needing commit. This will allow
1657 * the system to submit a single commit rpc for the whole
1658 * cluster. We can do this even if the buffer is not 100%
1659 * dirty (relative to the NFS blocksize), so we optimize the
1660 * append-to-file-case.
1661 *
1662 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1663 * cleared because write clustering only works for commit
1664 * rpc's, not for the data portion of the write).
1665 */
1666
1667 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1668 bp->b_flags |= B_NEEDCOMMIT;
1669 if (bp->b_dirtyoff == 0
1670 && bp->b_dirtyend == bp->b_bcount)
1671 bp->b_flags |= B_CLUSTEROK;
1672 } else {
1673 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1674 }
1675
1676 /*
1677 * For an interrupted write, the buffer is still valid
1678 * and the write hasn't been pushed to the server yet,
1679 * so we can't set BIO_ERROR and report the interruption
1680 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1681 * is not relevant, so the rpc attempt is essentially
1682 * a noop. For the case of a V3 write rpc not being
1683 * committed to stable storage, the block is still
1684 * dirty and requires either a commit rpc or another
1685 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1686 * the block is reused. This is indicated by setting
1687 * the B_DELWRI and B_NEEDCOMMIT flags.
1688 *
1689 * If the buffer is marked B_PAGING, it does not reside on
1690 * the vp's paging queues so we cannot call bdirty(). The
1691 * bp in this case is not an NFS cache block so we should
1692 * be safe. XXX
1693 *
1694 * The logic below breaks up errors into recoverable and
1695 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1696 * and keep the buffer around for potential write retries.
1697 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1698 * and save the error in the nfsnode. This is less than ideal
1699 * but necessary. Keeping such buffers around could potentially
1700 * cause buffer exhaustion eventually (they can never be written
1701 * out, so will get constantly be re-dirtied). It also causes
1702 * all sorts of vfs panics. For non-recoverable write errors,
1703 * also invalidate the attrcache, so we'll be forced to go over
1704 * the wire for this object, returning an error to user on next
1705 * call (most of the time).
1706 */
1707 if (error == EINTR || error == EIO || error == ETIMEDOUT
1708 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1709 int s;
1710
1711 s = splbio();
1712 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1713 if ((bp->b_flags & B_PAGING) == 0) {
1714 bdirty(bp);
1715 bp->b_flags &= ~B_DONE;
1716 }
1717 if (error && (bp->b_flags & B_ASYNC) == 0)
1718 bp->b_flags |= B_EINTR;
1719 splx(s);
1720 } else {
1721 if (error) {
1722 bp->b_ioflags |= BIO_ERROR;
1723 bp->b_flags |= B_INVAL;
1724 bp->b_error = np->n_error = error;
1725 mtx_lock(&np->n_mtx);
1726 np->n_flag |= NWRITEERR;
1727 np->n_attrstamp = 0;
1728 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1729 mtx_unlock(&np->n_mtx);
1730 }
1731 bp->b_dirtyoff = bp->b_dirtyend = 0;
1732 }
1733 } else {
1734 bp->b_resid = 0;
1735 bufdone(bp);
1736 return (0);
1737 }
1738 }
1739 bp->b_resid = uiop->uio_resid;
1740 if (must_commit)
1741 nfs_clearcommit(vp->v_mount);
1742 bufdone(bp);
1743 return (error);
1744 }
1745
1746 /*
1747 * Used to aid in handling ftruncate() operations on the NFS client side.
1748 * Truncation creates a number of special problems for NFS. We have to
1749 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1750 * we have to properly handle VM pages or (potentially dirty) buffers
1751 * that straddle the truncation point.
1752 */
1753
1754 int
1755 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1756 {
1757 struct nfsnode *np = VTONFS(vp);
1758 u_quad_t tsize;
1759 int biosize = vp->v_bufobj.bo_bsize;
1760 int error = 0;
1761
1762 mtx_lock(&np->n_mtx);
1763 tsize = np->n_size;
1764 np->n_size = nsize;
1765 mtx_unlock(&np->n_mtx);
1766
1767 if (nsize < tsize) {
1768 struct buf *bp;
1769 daddr_t lbn;
1770 int bufsize;
1771
1772 /*
1773 * vtruncbuf() doesn't get the buffer overlapping the
1774 * truncation point. We may have a B_DELWRI and/or B_CACHE
1775 * buffer that now needs to be truncated.
1776 */
1777 error = vtruncbuf(vp, cred, td, nsize, biosize);
1778 lbn = nsize / biosize;
1779 bufsize = nsize & (biosize - 1);
1780 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1781 if (!bp)
1782 return EINTR;
1783 if (bp->b_dirtyoff > bp->b_bcount)
1784 bp->b_dirtyoff = bp->b_bcount;
1785 if (bp->b_dirtyend > bp->b_bcount)
1786 bp->b_dirtyend = bp->b_bcount;
1787 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1788 brelse(bp);
1789 } else {
1790 vnode_pager_setsize(vp, nsize);
1791 }
1792 return(error);
1793 }
1794
Cache object: b328dd21caba21a81c0d3572febf2910
|