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/9.1/sys/nfsclient/nfs_bio.c 236096 2012-05-26 13:12:14Z 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 <nfs/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 for (i = 0; i < npages; ++i) {
133 if (i != ap->a_reqpage) {
134 vm_page_lock(pages[i]);
135 vm_page_free(pages[i]);
136 vm_page_unlock(pages[i]);
137 }
138 }
139 VM_OBJECT_UNLOCK(object);
140 return (0);
141 }
142 VM_OBJECT_UNLOCK(object);
143
144 /*
145 * We use only the kva address for the buffer, but this is extremely
146 * convienient and fast.
147 */
148 bp = getpbuf(&nfs_pbuf_freecnt);
149
150 kva = (vm_offset_t) bp->b_data;
151 pmap_qenter(kva, pages, npages);
152 PCPU_INC(cnt.v_vnodein);
153 PCPU_ADD(cnt.v_vnodepgsin, npages);
154
155 iov.iov_base = (caddr_t) kva;
156 iov.iov_len = count;
157 uio.uio_iov = &iov;
158 uio.uio_iovcnt = 1;
159 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
160 uio.uio_resid = count;
161 uio.uio_segflg = UIO_SYSSPACE;
162 uio.uio_rw = UIO_READ;
163 uio.uio_td = td;
164
165 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
166 pmap_qremove(kva, npages);
167
168 relpbuf(bp, &nfs_pbuf_freecnt);
169
170 if (error && (uio.uio_resid == count)) {
171 nfs_printf("nfs_getpages: error %d\n", error);
172 VM_OBJECT_LOCK(object);
173 for (i = 0; i < npages; ++i) {
174 if (i != ap->a_reqpage) {
175 vm_page_lock(pages[i]);
176 vm_page_free(pages[i]);
177 vm_page_unlock(pages[i]);
178 }
179 }
180 VM_OBJECT_UNLOCK(object);
181 return (VM_PAGER_ERROR);
182 }
183
184 /*
185 * Calculate the number of bytes read and validate only that number
186 * of bytes. Note that due to pending writes, size may be 0. This
187 * does not mean that the remaining data is invalid!
188 */
189
190 size = count - uio.uio_resid;
191 VM_OBJECT_LOCK(object);
192 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
193 vm_page_t m;
194 nextoff = toff + PAGE_SIZE;
195 m = pages[i];
196
197 if (nextoff <= size) {
198 /*
199 * Read operation filled an entire page
200 */
201 m->valid = VM_PAGE_BITS_ALL;
202 KASSERT(m->dirty == 0,
203 ("nfs_getpages: page %p is dirty", m));
204 } else if (size > toff) {
205 /*
206 * Read operation filled a partial page.
207 */
208 m->valid = 0;
209 vm_page_set_valid(m, 0, size - toff);
210 KASSERT(m->dirty == 0,
211 ("nfs_getpages: page %p is dirty", m));
212 } else {
213 /*
214 * Read operation was short. If no error occured
215 * we may have hit a zero-fill section. We simply
216 * leave valid set to 0.
217 */
218 ;
219 }
220 if (i != ap->a_reqpage) {
221 /*
222 * Whether or not to leave the page activated is up in
223 * the air, but we should put the page on a page queue
224 * somewhere (it already is in the object). Result:
225 * It appears that emperical results show that
226 * deactivating pages is best.
227 */
228
229 /*
230 * Just in case someone was asking for this page we
231 * now tell them that it is ok to use.
232 */
233 if (!error) {
234 if (m->oflags & VPO_WANTED) {
235 vm_page_lock(m);
236 vm_page_activate(m);
237 vm_page_unlock(m);
238 } else {
239 vm_page_lock(m);
240 vm_page_deactivate(m);
241 vm_page_unlock(m);
242 }
243 vm_page_wakeup(m);
244 } else {
245 vm_page_lock(m);
246 vm_page_free(m);
247 vm_page_unlock(m);
248 }
249 }
250 }
251 VM_OBJECT_UNLOCK(object);
252 return (0);
253 }
254
255 /*
256 * Vnode op for VM putpages.
257 */
258 int
259 nfs_putpages(struct vop_putpages_args *ap)
260 {
261 struct uio uio;
262 struct iovec iov;
263 vm_offset_t kva;
264 struct buf *bp;
265 int iomode, must_commit, i, error, npages, count;
266 off_t offset;
267 int *rtvals;
268 struct vnode *vp;
269 struct thread *td;
270 struct ucred *cred;
271 struct nfsmount *nmp;
272 struct nfsnode *np;
273 vm_page_t *pages;
274
275 vp = ap->a_vp;
276 np = VTONFS(vp);
277 td = curthread; /* XXX */
278 /* Set the cred to n_writecred for the write rpcs. */
279 if (np->n_writecred != NULL)
280 cred = crhold(np->n_writecred);
281 else
282 cred = crhold(curthread->td_ucred); /* XXX */
283 nmp = VFSTONFS(vp->v_mount);
284 pages = ap->a_m;
285 count = ap->a_count;
286 rtvals = ap->a_rtvals;
287 npages = btoc(count);
288 offset = IDX_TO_OFF(pages[0]->pindex);
289
290 mtx_lock(&nmp->nm_mtx);
291 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
292 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
293 mtx_unlock(&nmp->nm_mtx);
294 (void)nfs_fsinfo(nmp, vp, cred, td);
295 } else
296 mtx_unlock(&nmp->nm_mtx);
297
298 mtx_lock(&np->n_mtx);
299 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
300 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
301 mtx_unlock(&np->n_mtx);
302 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
303 mtx_lock(&np->n_mtx);
304 }
305
306 for (i = 0; i < npages; i++)
307 rtvals[i] = VM_PAGER_ERROR;
308
309 /*
310 * When putting pages, do not extend file past EOF.
311 */
312 if (offset + count > np->n_size) {
313 count = np->n_size - offset;
314 if (count < 0)
315 count = 0;
316 }
317 mtx_unlock(&np->n_mtx);
318
319 /*
320 * We use only the kva address for the buffer, but this is extremely
321 * convienient and fast.
322 */
323 bp = getpbuf(&nfs_pbuf_freecnt);
324
325 kva = (vm_offset_t) bp->b_data;
326 pmap_qenter(kva, pages, npages);
327 PCPU_INC(cnt.v_vnodeout);
328 PCPU_ADD(cnt.v_vnodepgsout, count);
329
330 iov.iov_base = (caddr_t) kva;
331 iov.iov_len = count;
332 uio.uio_iov = &iov;
333 uio.uio_iovcnt = 1;
334 uio.uio_offset = offset;
335 uio.uio_resid = count;
336 uio.uio_segflg = UIO_SYSSPACE;
337 uio.uio_rw = UIO_WRITE;
338 uio.uio_td = td;
339
340 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
341 iomode = NFSV3WRITE_UNSTABLE;
342 else
343 iomode = NFSV3WRITE_FILESYNC;
344
345 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
346 crfree(cred);
347
348 pmap_qremove(kva, npages);
349 relpbuf(bp, &nfs_pbuf_freecnt);
350
351 if (!error) {
352 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
353 if (must_commit) {
354 nfs_clearcommit(vp->v_mount);
355 }
356 }
357 return rtvals[0];
358 }
359
360 /*
361 * For nfs, cache consistency can only be maintained approximately.
362 * Although RFC1094 does not specify the criteria, the following is
363 * believed to be compatible with the reference port.
364 * For nfs:
365 * If the file's modify time on the server has changed since the
366 * last read rpc or you have written to the file,
367 * you may have lost data cache consistency with the
368 * server, so flush all of the file's data out of the cache.
369 * Then force a getattr rpc to ensure that you have up to date
370 * attributes.
371 * NB: This implies that cache data can be read when up to
372 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
373 * attributes this could be forced by setting n_attrstamp to 0 before
374 * the VOP_GETATTR() call.
375 */
376 static inline int
377 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
378 {
379 int error = 0;
380 struct vattr vattr;
381 struct nfsnode *np = VTONFS(vp);
382 int old_lock;
383 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
384
385 /*
386 * Grab the exclusive lock before checking whether the cache is
387 * consistent.
388 * XXX - We can make this cheaper later (by acquiring cheaper locks).
389 * But for now, this suffices.
390 */
391 old_lock = nfs_upgrade_vnlock(vp);
392 if (vp->v_iflag & VI_DOOMED) {
393 nfs_downgrade_vnlock(vp, old_lock);
394 return (EBADF);
395 }
396
397 mtx_lock(&np->n_mtx);
398 if (np->n_flag & NMODIFIED) {
399 mtx_unlock(&np->n_mtx);
400 if (vp->v_type != VREG) {
401 if (vp->v_type != VDIR)
402 panic("nfs: bioread, not dir");
403 (nmp->nm_rpcops->nr_invaldir)(vp);
404 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
405 if (error)
406 goto out;
407 }
408 np->n_attrstamp = 0;
409 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
410 error = VOP_GETATTR(vp, &vattr, cred);
411 if (error)
412 goto out;
413 mtx_lock(&np->n_mtx);
414 np->n_mtime = vattr.va_mtime;
415 mtx_unlock(&np->n_mtx);
416 } else {
417 mtx_unlock(&np->n_mtx);
418 error = VOP_GETATTR(vp, &vattr, cred);
419 if (error)
420 return (error);
421 mtx_lock(&np->n_mtx);
422 if ((np->n_flag & NSIZECHANGED)
423 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
424 mtx_unlock(&np->n_mtx);
425 if (vp->v_type == VDIR)
426 (nmp->nm_rpcops->nr_invaldir)(vp);
427 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
428 if (error)
429 goto out;
430 mtx_lock(&np->n_mtx);
431 np->n_mtime = vattr.va_mtime;
432 np->n_flag &= ~NSIZECHANGED;
433 }
434 mtx_unlock(&np->n_mtx);
435 }
436 out:
437 nfs_downgrade_vnlock(vp, old_lock);
438 return error;
439 }
440
441 /*
442 * Vnode op for read using bio
443 */
444 int
445 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
446 {
447 struct nfsnode *np = VTONFS(vp);
448 int biosize, i;
449 struct buf *bp, *rabp;
450 struct thread *td;
451 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
452 daddr_t lbn, rabn;
453 off_t end;
454 int bcount;
455 int seqcount;
456 int nra, error = 0, n = 0, on = 0;
457
458 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
459 if (uio->uio_resid == 0)
460 return (0);
461 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
462 return (EINVAL);
463 td = uio->uio_td;
464
465 mtx_lock(&nmp->nm_mtx);
466 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
467 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
468 mtx_unlock(&nmp->nm_mtx);
469 (void)nfs_fsinfo(nmp, vp, cred, td);
470 } else
471 mtx_unlock(&nmp->nm_mtx);
472
473 end = uio->uio_offset + uio->uio_resid;
474 if (vp->v_type != VDIR &&
475 (end > nmp->nm_maxfilesize || end < uio->uio_offset))
476 return (EFBIG);
477
478 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
479 /* No caching/ no readaheads. Just read data into the user buffer */
480 return nfs_readrpc(vp, uio, cred);
481
482 biosize = vp->v_bufobj.bo_bsize;
483 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
484
485 error = nfs_bioread_check_cons(vp, td, cred);
486 if (error)
487 return error;
488
489 do {
490 u_quad_t nsize;
491
492 mtx_lock(&np->n_mtx);
493 nsize = np->n_size;
494 mtx_unlock(&np->n_mtx);
495
496 switch (vp->v_type) {
497 case VREG:
498 nfsstats.biocache_reads++;
499 lbn = uio->uio_offset / biosize;
500 on = uio->uio_offset & (biosize - 1);
501
502 /*
503 * Start the read ahead(s), as required.
504 */
505 if (nmp->nm_readahead > 0) {
506 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
507 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
508 rabn = lbn + 1 + nra;
509 if (incore(&vp->v_bufobj, rabn) == NULL) {
510 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
511 if (!rabp) {
512 error = nfs_sigintr(nmp, td);
513 return (error ? error : EINTR);
514 }
515 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
516 rabp->b_flags |= B_ASYNC;
517 rabp->b_iocmd = BIO_READ;
518 vfs_busy_pages(rabp, 0);
519 if (nfs_asyncio(nmp, rabp, cred, td)) {
520 rabp->b_flags |= B_INVAL;
521 rabp->b_ioflags |= BIO_ERROR;
522 vfs_unbusy_pages(rabp);
523 brelse(rabp);
524 break;
525 }
526 } else {
527 brelse(rabp);
528 }
529 }
530 }
531 }
532
533 /* Note that bcount is *not* DEV_BSIZE aligned. */
534 bcount = biosize;
535 if ((off_t)lbn * biosize >= nsize) {
536 bcount = 0;
537 } else if ((off_t)(lbn + 1) * biosize > nsize) {
538 bcount = nsize - (off_t)lbn * biosize;
539 }
540 bp = nfs_getcacheblk(vp, lbn, bcount, td);
541
542 if (!bp) {
543 error = nfs_sigintr(nmp, td);
544 return (error ? error : EINTR);
545 }
546
547 /*
548 * If B_CACHE is not set, we must issue the read. If this
549 * fails, we return an error.
550 */
551
552 if ((bp->b_flags & B_CACHE) == 0) {
553 bp->b_iocmd = BIO_READ;
554 vfs_busy_pages(bp, 0);
555 error = nfs_doio(vp, bp, cred, td);
556 if (error) {
557 brelse(bp);
558 return (error);
559 }
560 }
561
562 /*
563 * on is the offset into the current bp. Figure out how many
564 * bytes we can copy out of the bp. Note that bcount is
565 * NOT DEV_BSIZE aligned.
566 *
567 * Then figure out how many bytes we can copy into the uio.
568 */
569
570 n = 0;
571 if (on < bcount)
572 n = MIN((unsigned)(bcount - on), uio->uio_resid);
573 break;
574 case VLNK:
575 nfsstats.biocache_readlinks++;
576 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
577 if (!bp) {
578 error = nfs_sigintr(nmp, td);
579 return (error ? error : EINTR);
580 }
581 if ((bp->b_flags & B_CACHE) == 0) {
582 bp->b_iocmd = BIO_READ;
583 vfs_busy_pages(bp, 0);
584 error = nfs_doio(vp, bp, cred, td);
585 if (error) {
586 bp->b_ioflags |= BIO_ERROR;
587 brelse(bp);
588 return (error);
589 }
590 }
591 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
592 on = 0;
593 break;
594 case VDIR:
595 nfsstats.biocache_readdirs++;
596 if (np->n_direofoffset
597 && uio->uio_offset >= np->n_direofoffset) {
598 return (0);
599 }
600 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
601 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
602 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
603 if (!bp) {
604 error = nfs_sigintr(nmp, td);
605 return (error ? error : EINTR);
606 }
607 if ((bp->b_flags & B_CACHE) == 0) {
608 bp->b_iocmd = BIO_READ;
609 vfs_busy_pages(bp, 0);
610 error = nfs_doio(vp, bp, cred, td);
611 if (error) {
612 brelse(bp);
613 }
614 while (error == NFSERR_BAD_COOKIE) {
615 (nmp->nm_rpcops->nr_invaldir)(vp);
616 error = nfs_vinvalbuf(vp, 0, td, 1);
617 /*
618 * Yuck! The directory has been modified on the
619 * server. The only way to get the block is by
620 * reading from the beginning to get all the
621 * offset cookies.
622 *
623 * Leave the last bp intact unless there is an error.
624 * Loop back up to the while if the error is another
625 * NFSERR_BAD_COOKIE (double yuch!).
626 */
627 for (i = 0; i <= lbn && !error; i++) {
628 if (np->n_direofoffset
629 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
630 return (0);
631 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
632 if (!bp) {
633 error = nfs_sigintr(nmp, td);
634 return (error ? error : EINTR);
635 }
636 if ((bp->b_flags & B_CACHE) == 0) {
637 bp->b_iocmd = BIO_READ;
638 vfs_busy_pages(bp, 0);
639 error = nfs_doio(vp, bp, cred, td);
640 /*
641 * no error + B_INVAL == directory EOF,
642 * use the block.
643 */
644 if (error == 0 && (bp->b_flags & B_INVAL))
645 break;
646 }
647 /*
648 * An error will throw away the block and the
649 * for loop will break out. If no error and this
650 * is not the block we want, we throw away the
651 * block and go for the next one via the for loop.
652 */
653 if (error || i < lbn)
654 brelse(bp);
655 }
656 }
657 /*
658 * The above while is repeated if we hit another cookie
659 * error. If we hit an error and it wasn't a cookie error,
660 * we give up.
661 */
662 if (error)
663 return (error);
664 }
665
666 /*
667 * If not eof and read aheads are enabled, start one.
668 * (You need the current block first, so that you have the
669 * directory offset cookie of the next block.)
670 */
671 if (nmp->nm_readahead > 0 &&
672 (bp->b_flags & B_INVAL) == 0 &&
673 (np->n_direofoffset == 0 ||
674 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
675 incore(&vp->v_bufobj, lbn + 1) == NULL) {
676 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
677 if (rabp) {
678 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
679 rabp->b_flags |= B_ASYNC;
680 rabp->b_iocmd = BIO_READ;
681 vfs_busy_pages(rabp, 0);
682 if (nfs_asyncio(nmp, rabp, cred, td)) {
683 rabp->b_flags |= B_INVAL;
684 rabp->b_ioflags |= BIO_ERROR;
685 vfs_unbusy_pages(rabp);
686 brelse(rabp);
687 }
688 } else {
689 brelse(rabp);
690 }
691 }
692 }
693 /*
694 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
695 * chopped for the EOF condition, we cannot tell how large
696 * NFS directories are going to be until we hit EOF. So
697 * an NFS directory buffer is *not* chopped to its EOF. Now,
698 * it just so happens that b_resid will effectively chop it
699 * to EOF. *BUT* this information is lost if the buffer goes
700 * away and is reconstituted into a B_CACHE state ( due to
701 * being VMIO ) later. So we keep track of the directory eof
702 * in np->n_direofoffset and chop it off as an extra step
703 * right here.
704 */
705 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
706 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
707 n = np->n_direofoffset - uio->uio_offset;
708 break;
709 default:
710 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
711 bp = NULL;
712 break;
713 };
714
715 if (n > 0) {
716 error = uiomove(bp->b_data + on, (int)n, uio);
717 }
718 if (vp->v_type == VLNK)
719 n = 0;
720 if (bp != NULL)
721 brelse(bp);
722 } while (error == 0 && uio->uio_resid > 0 && n > 0);
723 return (error);
724 }
725
726 /*
727 * The NFS write path cannot handle iovecs with len > 1. So we need to
728 * break up iovecs accordingly (restricting them to wsize).
729 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
730 * For the ASYNC case, 2 copies are needed. The first a copy from the
731 * user buffer to a staging buffer and then a second copy from the staging
732 * buffer to mbufs. This can be optimized by copying from the user buffer
733 * directly into mbufs and passing the chain down, but that requires a
734 * fair amount of re-working of the relevant codepaths (and can be done
735 * later).
736 */
737 static int
738 nfs_directio_write(vp, uiop, cred, ioflag)
739 struct vnode *vp;
740 struct uio *uiop;
741 struct ucred *cred;
742 int ioflag;
743 {
744 int error;
745 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
746 struct thread *td = uiop->uio_td;
747 int size;
748 int wsize;
749
750 mtx_lock(&nmp->nm_mtx);
751 wsize = nmp->nm_wsize;
752 mtx_unlock(&nmp->nm_mtx);
753 if (ioflag & IO_SYNC) {
754 int iomode, must_commit;
755 struct uio uio;
756 struct iovec iov;
757 do_sync:
758 while (uiop->uio_resid > 0) {
759 size = MIN(uiop->uio_resid, wsize);
760 size = MIN(uiop->uio_iov->iov_len, size);
761 iov.iov_base = uiop->uio_iov->iov_base;
762 iov.iov_len = size;
763 uio.uio_iov = &iov;
764 uio.uio_iovcnt = 1;
765 uio.uio_offset = uiop->uio_offset;
766 uio.uio_resid = size;
767 uio.uio_segflg = UIO_USERSPACE;
768 uio.uio_rw = UIO_WRITE;
769 uio.uio_td = td;
770 iomode = NFSV3WRITE_FILESYNC;
771 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
772 &iomode, &must_commit);
773 KASSERT((must_commit == 0),
774 ("nfs_directio_write: Did not commit write"));
775 if (error)
776 return (error);
777 uiop->uio_offset += size;
778 uiop->uio_resid -= size;
779 if (uiop->uio_iov->iov_len <= size) {
780 uiop->uio_iovcnt--;
781 uiop->uio_iov++;
782 } else {
783 uiop->uio_iov->iov_base =
784 (char *)uiop->uio_iov->iov_base + size;
785 uiop->uio_iov->iov_len -= size;
786 }
787 }
788 } else {
789 struct uio *t_uio;
790 struct iovec *t_iov;
791 struct buf *bp;
792
793 /*
794 * Break up the write into blocksize chunks and hand these
795 * over to nfsiod's for write back.
796 * Unfortunately, this incurs a copy of the data. Since
797 * the user could modify the buffer before the write is
798 * initiated.
799 *
800 * The obvious optimization here is that one of the 2 copies
801 * in the async write path can be eliminated by copying the
802 * data here directly into mbufs and passing the mbuf chain
803 * down. But that will require a fair amount of re-working
804 * of the code and can be done if there's enough interest
805 * in NFS directio access.
806 */
807 while (uiop->uio_resid > 0) {
808 size = MIN(uiop->uio_resid, wsize);
809 size = MIN(uiop->uio_iov->iov_len, size);
810 bp = getpbuf(&nfs_pbuf_freecnt);
811 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
812 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
813 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
814 t_iov->iov_len = size;
815 t_uio->uio_iov = t_iov;
816 t_uio->uio_iovcnt = 1;
817 t_uio->uio_offset = uiop->uio_offset;
818 t_uio->uio_resid = size;
819 t_uio->uio_segflg = UIO_SYSSPACE;
820 t_uio->uio_rw = UIO_WRITE;
821 t_uio->uio_td = td;
822 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
823 uiop->uio_segflg == UIO_SYSSPACE,
824 ("nfs_directio_write: Bad uio_segflg"));
825 if (uiop->uio_segflg == UIO_USERSPACE) {
826 error = copyin(uiop->uio_iov->iov_base,
827 t_iov->iov_base, size);
828 if (error != 0)
829 goto err_free;
830 } else
831 /*
832 * UIO_SYSSPACE may never happen, but handle
833 * it just in case it does.
834 */
835 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
836 size);
837 bp->b_flags |= B_DIRECT;
838 bp->b_iocmd = BIO_WRITE;
839 if (cred != NOCRED) {
840 crhold(cred);
841 bp->b_wcred = cred;
842 } else
843 bp->b_wcred = NOCRED;
844 bp->b_caller1 = (void *)t_uio;
845 bp->b_vp = vp;
846 error = nfs_asyncio(nmp, bp, NOCRED, td);
847 err_free:
848 if (error) {
849 free(t_iov->iov_base, M_NFSDIRECTIO);
850 free(t_iov, M_NFSDIRECTIO);
851 free(t_uio, M_NFSDIRECTIO);
852 bp->b_vp = NULL;
853 relpbuf(bp, &nfs_pbuf_freecnt);
854 if (error == EINTR)
855 return (error);
856 goto do_sync;
857 }
858 uiop->uio_offset += size;
859 uiop->uio_resid -= size;
860 if (uiop->uio_iov->iov_len <= size) {
861 uiop->uio_iovcnt--;
862 uiop->uio_iov++;
863 } else {
864 uiop->uio_iov->iov_base =
865 (char *)uiop->uio_iov->iov_base + size;
866 uiop->uio_iov->iov_len -= size;
867 }
868 }
869 }
870 return (0);
871 }
872
873 /*
874 * Vnode op for write using bio
875 */
876 int
877 nfs_write(struct vop_write_args *ap)
878 {
879 int biosize;
880 struct uio *uio = ap->a_uio;
881 struct thread *td = uio->uio_td;
882 struct vnode *vp = ap->a_vp;
883 struct nfsnode *np = VTONFS(vp);
884 struct ucred *cred = ap->a_cred;
885 int ioflag = ap->a_ioflag;
886 struct buf *bp;
887 struct vattr vattr;
888 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
889 daddr_t lbn;
890 off_t end;
891 int bcount;
892 int n, on, error = 0;
893
894 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode"));
895 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
896 ("nfs_write proc"));
897 if (vp->v_type != VREG)
898 return (EIO);
899 mtx_lock(&np->n_mtx);
900 if (np->n_flag & NWRITEERR) {
901 np->n_flag &= ~NWRITEERR;
902 mtx_unlock(&np->n_mtx);
903 return (np->n_error);
904 } else
905 mtx_unlock(&np->n_mtx);
906 mtx_lock(&nmp->nm_mtx);
907 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
908 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
909 mtx_unlock(&nmp->nm_mtx);
910 (void)nfs_fsinfo(nmp, vp, cred, td);
911 } else
912 mtx_unlock(&nmp->nm_mtx);
913
914 /*
915 * Synchronously flush pending buffers if we are in synchronous
916 * mode or if we are appending.
917 */
918 if (ioflag & (IO_APPEND | IO_SYNC)) {
919 mtx_lock(&np->n_mtx);
920 if (np->n_flag & NMODIFIED) {
921 mtx_unlock(&np->n_mtx);
922 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
923 /*
924 * Require non-blocking, synchronous writes to
925 * dirty files to inform the program it needs
926 * to fsync(2) explicitly.
927 */
928 if (ioflag & IO_NDELAY)
929 return (EAGAIN);
930 #endif
931 flush_and_restart:
932 np->n_attrstamp = 0;
933 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
934 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
935 if (error)
936 return (error);
937 } else
938 mtx_unlock(&np->n_mtx);
939 }
940
941 /*
942 * If IO_APPEND then load uio_offset. We restart here if we cannot
943 * get the append lock.
944 */
945 if (ioflag & IO_APPEND) {
946 np->n_attrstamp = 0;
947 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
948 error = VOP_GETATTR(vp, &vattr, cred);
949 if (error)
950 return (error);
951 mtx_lock(&np->n_mtx);
952 uio->uio_offset = np->n_size;
953 mtx_unlock(&np->n_mtx);
954 }
955
956 if (uio->uio_offset < 0)
957 return (EINVAL);
958 end = uio->uio_offset + uio->uio_resid;
959 if (end > nmp->nm_maxfilesize || end < uio->uio_offset)
960 return (EFBIG);
961 if (uio->uio_resid == 0)
962 return (0);
963
964 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
965 return nfs_directio_write(vp, uio, cred, ioflag);
966
967 /*
968 * Maybe this should be above the vnode op call, but so long as
969 * file servers have no limits, i don't think it matters
970 */
971 if (vn_rlimit_fsize(vp, uio, td))
972 return (EFBIG);
973
974 biosize = vp->v_bufobj.bo_bsize;
975 /*
976 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
977 * would exceed the local maximum per-file write commit size when
978 * combined with those, we must decide whether to flush,
979 * go synchronous, or return error. We don't bother checking
980 * IO_UNIT -- we just make all writes atomic anyway, as there's
981 * no point optimizing for something that really won't ever happen.
982 */
983 if (!(ioflag & IO_SYNC)) {
984 int nflag;
985
986 mtx_lock(&np->n_mtx);
987 nflag = np->n_flag;
988 mtx_unlock(&np->n_mtx);
989 int needrestart = 0;
990 if (nmp->nm_wcommitsize < uio->uio_resid) {
991 /*
992 * If this request could not possibly be completed
993 * without exceeding the maximum outstanding write
994 * commit size, see if we can convert it into a
995 * synchronous write operation.
996 */
997 if (ioflag & IO_NDELAY)
998 return (EAGAIN);
999 ioflag |= IO_SYNC;
1000 if (nflag & NMODIFIED)
1001 needrestart = 1;
1002 } else if (nflag & NMODIFIED) {
1003 int wouldcommit = 0;
1004 BO_LOCK(&vp->v_bufobj);
1005 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1006 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1007 b_bobufs) {
1008 if (bp->b_flags & B_NEEDCOMMIT)
1009 wouldcommit += bp->b_bcount;
1010 }
1011 }
1012 BO_UNLOCK(&vp->v_bufobj);
1013 /*
1014 * Since we're not operating synchronously and
1015 * bypassing the buffer cache, we are in a commit
1016 * and holding all of these buffers whether
1017 * transmitted or not. If not limited, this
1018 * will lead to the buffer cache deadlocking,
1019 * as no one else can flush our uncommitted buffers.
1020 */
1021 wouldcommit += uio->uio_resid;
1022 /*
1023 * If we would initially exceed the maximum
1024 * outstanding write commit size, flush and restart.
1025 */
1026 if (wouldcommit > nmp->nm_wcommitsize)
1027 needrestart = 1;
1028 }
1029 if (needrestart)
1030 goto flush_and_restart;
1031 }
1032
1033 do {
1034 nfsstats.biocache_writes++;
1035 lbn = uio->uio_offset / biosize;
1036 on = uio->uio_offset & (biosize-1);
1037 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1038 again:
1039 /*
1040 * Handle direct append and file extension cases, calculate
1041 * unaligned buffer size.
1042 */
1043 mtx_lock(&np->n_mtx);
1044 if (uio->uio_offset == np->n_size && n) {
1045 mtx_unlock(&np->n_mtx);
1046 /*
1047 * Get the buffer (in its pre-append state to maintain
1048 * B_CACHE if it was previously set). Resize the
1049 * nfsnode after we have locked the buffer to prevent
1050 * readers from reading garbage.
1051 */
1052 bcount = on;
1053 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1054
1055 if (bp != NULL) {
1056 long save;
1057
1058 mtx_lock(&np->n_mtx);
1059 np->n_size = uio->uio_offset + n;
1060 np->n_flag |= NMODIFIED;
1061 vnode_pager_setsize(vp, np->n_size);
1062 mtx_unlock(&np->n_mtx);
1063
1064 save = bp->b_flags & B_CACHE;
1065 bcount += n;
1066 allocbuf(bp, bcount);
1067 bp->b_flags |= save;
1068 }
1069 } else {
1070 /*
1071 * Obtain the locked cache block first, and then
1072 * adjust the file's size as appropriate.
1073 */
1074 bcount = on + n;
1075 if ((off_t)lbn * biosize + bcount < np->n_size) {
1076 if ((off_t)(lbn + 1) * biosize < np->n_size)
1077 bcount = biosize;
1078 else
1079 bcount = np->n_size - (off_t)lbn * biosize;
1080 }
1081 mtx_unlock(&np->n_mtx);
1082 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1083 mtx_lock(&np->n_mtx);
1084 if (uio->uio_offset + n > np->n_size) {
1085 np->n_size = uio->uio_offset + n;
1086 np->n_flag |= NMODIFIED;
1087 vnode_pager_setsize(vp, np->n_size);
1088 }
1089 mtx_unlock(&np->n_mtx);
1090 }
1091
1092 if (!bp) {
1093 error = nfs_sigintr(nmp, td);
1094 if (!error)
1095 error = EINTR;
1096 break;
1097 }
1098
1099 /*
1100 * Issue a READ if B_CACHE is not set. In special-append
1101 * mode, B_CACHE is based on the buffer prior to the write
1102 * op and is typically set, avoiding the read. If a read
1103 * is required in special append mode, the server will
1104 * probably send us a short-read since we extended the file
1105 * on our end, resulting in b_resid == 0 and, thusly,
1106 * B_CACHE getting set.
1107 *
1108 * We can also avoid issuing the read if the write covers
1109 * the entire buffer. We have to make sure the buffer state
1110 * is reasonable in this case since we will not be initiating
1111 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1112 * more information.
1113 *
1114 * B_CACHE may also be set due to the buffer being cached
1115 * normally.
1116 */
1117
1118 if (on == 0 && n == bcount) {
1119 bp->b_flags |= B_CACHE;
1120 bp->b_flags &= ~B_INVAL;
1121 bp->b_ioflags &= ~BIO_ERROR;
1122 }
1123
1124 if ((bp->b_flags & B_CACHE) == 0) {
1125 bp->b_iocmd = BIO_READ;
1126 vfs_busy_pages(bp, 0);
1127 error = nfs_doio(vp, bp, cred, td);
1128 if (error) {
1129 brelse(bp);
1130 break;
1131 }
1132 }
1133 if (bp->b_wcred == NOCRED)
1134 bp->b_wcred = crhold(cred);
1135 mtx_lock(&np->n_mtx);
1136 np->n_flag |= NMODIFIED;
1137 mtx_unlock(&np->n_mtx);
1138
1139 /*
1140 * If dirtyend exceeds file size, chop it down. This should
1141 * not normally occur but there is an append race where it
1142 * might occur XXX, so we log it.
1143 *
1144 * If the chopping creates a reverse-indexed or degenerate
1145 * situation with dirtyoff/end, we 0 both of them.
1146 */
1147
1148 if (bp->b_dirtyend > bcount) {
1149 nfs_printf("NFS append race @%lx:%d\n",
1150 (long)bp->b_blkno * DEV_BSIZE,
1151 bp->b_dirtyend - bcount);
1152 bp->b_dirtyend = bcount;
1153 }
1154
1155 if (bp->b_dirtyoff >= bp->b_dirtyend)
1156 bp->b_dirtyoff = bp->b_dirtyend = 0;
1157
1158 /*
1159 * If the new write will leave a contiguous dirty
1160 * area, just update the b_dirtyoff and b_dirtyend,
1161 * otherwise force a write rpc of the old dirty area.
1162 *
1163 * While it is possible to merge discontiguous writes due to
1164 * our having a B_CACHE buffer ( and thus valid read data
1165 * for the hole), we don't because it could lead to
1166 * significant cache coherency problems with multiple clients,
1167 * especially if locking is implemented later on.
1168 *
1169 * as an optimization we could theoretically maintain
1170 * a linked list of discontinuous areas, but we would still
1171 * have to commit them separately so there isn't much
1172 * advantage to it except perhaps a bit of asynchronization.
1173 */
1174
1175 if (bp->b_dirtyend > 0 &&
1176 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1177 if (bwrite(bp) == EINTR) {
1178 error = EINTR;
1179 break;
1180 }
1181 goto again;
1182 }
1183
1184 error = uiomove((char *)bp->b_data + on, n, uio);
1185
1186 /*
1187 * Since this block is being modified, it must be written
1188 * again and not just committed. Since write clustering does
1189 * not work for the stage 1 data write, only the stage 2
1190 * commit rpc, we have to clear B_CLUSTEROK as well.
1191 */
1192 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1193
1194 if (error) {
1195 bp->b_ioflags |= BIO_ERROR;
1196 brelse(bp);
1197 break;
1198 }
1199
1200 /*
1201 * Only update dirtyoff/dirtyend if not a degenerate
1202 * condition.
1203 */
1204 if (n) {
1205 if (bp->b_dirtyend > 0) {
1206 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1207 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1208 } else {
1209 bp->b_dirtyoff = on;
1210 bp->b_dirtyend = on + n;
1211 }
1212 vfs_bio_set_valid(bp, on, n);
1213 }
1214
1215 /*
1216 * If IO_SYNC do bwrite().
1217 *
1218 * IO_INVAL appears to be unused. The idea appears to be
1219 * to turn off caching in this case. Very odd. XXX
1220 */
1221 if ((ioflag & IO_SYNC)) {
1222 if (ioflag & IO_INVAL)
1223 bp->b_flags |= B_NOCACHE;
1224 error = bwrite(bp);
1225 if (error)
1226 break;
1227 } else if ((n + on) == biosize) {
1228 bp->b_flags |= B_ASYNC;
1229 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1230 } else {
1231 bdwrite(bp);
1232 }
1233 } while (uio->uio_resid > 0 && n > 0);
1234
1235 return (error);
1236 }
1237
1238 /*
1239 * Get an nfs cache block.
1240 *
1241 * Allocate a new one if the block isn't currently in the cache
1242 * and return the block marked busy. If the calling process is
1243 * interrupted by a signal for an interruptible mount point, return
1244 * NULL.
1245 *
1246 * The caller must carefully deal with the possible B_INVAL state of
1247 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1248 * indirectly), so synchronous reads can be issued without worrying about
1249 * the B_INVAL state. We have to be a little more careful when dealing
1250 * with writes (see comments in nfs_write()) when extending a file past
1251 * its EOF.
1252 */
1253 static struct buf *
1254 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1255 {
1256 struct buf *bp;
1257 struct mount *mp;
1258 struct nfsmount *nmp;
1259
1260 mp = vp->v_mount;
1261 nmp = VFSTONFS(mp);
1262
1263 if (nmp->nm_flag & NFSMNT_INT) {
1264 sigset_t oldset;
1265
1266 nfs_set_sigmask(td, &oldset);
1267 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1268 nfs_restore_sigmask(td, &oldset);
1269 while (bp == NULL) {
1270 if (nfs_sigintr(nmp, td))
1271 return (NULL);
1272 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1273 }
1274 } else {
1275 bp = getblk(vp, bn, size, 0, 0, 0);
1276 }
1277
1278 if (vp->v_type == VREG)
1279 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1280 return (bp);
1281 }
1282
1283 /*
1284 * Flush and invalidate all dirty buffers. If another process is already
1285 * doing the flush, just wait for completion.
1286 */
1287 int
1288 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1289 {
1290 struct nfsnode *np = VTONFS(vp);
1291 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1292 int error = 0, slpflag, slptimeo;
1293 int old_lock = 0;
1294
1295 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1296
1297 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1298 intrflg = 0;
1299 if (intrflg) {
1300 slpflag = NFS_PCATCH;
1301 slptimeo = 2 * hz;
1302 } else {
1303 slpflag = 0;
1304 slptimeo = 0;
1305 }
1306
1307 old_lock = nfs_upgrade_vnlock(vp);
1308 if (vp->v_iflag & VI_DOOMED) {
1309 /*
1310 * Since vgonel() uses the generic vinvalbuf() to flush
1311 * dirty buffers and it does not call this function, it
1312 * is safe to just return OK when VI_DOOMED is set.
1313 */
1314 nfs_downgrade_vnlock(vp, old_lock);
1315 return (0);
1316 }
1317
1318 /*
1319 * Now, flush as required.
1320 */
1321 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1322 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1323 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1324 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1325 /*
1326 * If the page clean was interrupted, fail the invalidation.
1327 * Not doing so, we run the risk of losing dirty pages in the
1328 * vinvalbuf() call below.
1329 */
1330 if (intrflg && (error = nfs_sigintr(nmp, td)))
1331 goto out;
1332 }
1333
1334 error = vinvalbuf(vp, flags, slpflag, 0);
1335 while (error) {
1336 if (intrflg && (error = nfs_sigintr(nmp, td)))
1337 goto out;
1338 error = vinvalbuf(vp, flags, 0, slptimeo);
1339 }
1340 mtx_lock(&np->n_mtx);
1341 if (np->n_directio_asyncwr == 0)
1342 np->n_flag &= ~NMODIFIED;
1343 mtx_unlock(&np->n_mtx);
1344 out:
1345 nfs_downgrade_vnlock(vp, old_lock);
1346 return error;
1347 }
1348
1349 /*
1350 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1351 * This is mainly to avoid queueing async I/O requests when the nfsiods
1352 * are all hung on a dead server.
1353 *
1354 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1355 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1356 */
1357 int
1358 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1359 {
1360 int iod;
1361 int gotiod;
1362 int slpflag = 0;
1363 int slptimeo = 0;
1364 int error, error2;
1365
1366 /*
1367 * Commits are usually short and sweet so lets save some cpu and
1368 * leave the async daemons for more important rpc's (such as reads
1369 * and writes).
1370 */
1371 mtx_lock(&nfs_iod_mtx);
1372 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1373 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1374 mtx_unlock(&nfs_iod_mtx);
1375 return(EIO);
1376 }
1377 again:
1378 if (nmp->nm_flag & NFSMNT_INT)
1379 slpflag = NFS_PCATCH;
1380 gotiod = FALSE;
1381
1382 /*
1383 * Find a free iod to process this request.
1384 */
1385 for (iod = 0; iod < nfs_numasync; iod++)
1386 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1387 gotiod = TRUE;
1388 break;
1389 }
1390
1391 /*
1392 * Try to create one if none are free.
1393 */
1394 if (!gotiod)
1395 nfs_nfsiodnew();
1396 else {
1397 /*
1398 * Found one, so wake it up and tell it which
1399 * mount to process.
1400 */
1401 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1402 iod, nmp));
1403 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1404 nfs_iodmount[iod] = nmp;
1405 nmp->nm_bufqiods++;
1406 wakeup(&nfs_iodwant[iod]);
1407 }
1408
1409 /*
1410 * If none are free, we may already have an iod working on this mount
1411 * point. If so, it will process our request.
1412 */
1413 if (!gotiod) {
1414 if (nmp->nm_bufqiods > 0) {
1415 NFS_DPF(ASYNCIO,
1416 ("nfs_asyncio: %d iods are already processing mount %p\n",
1417 nmp->nm_bufqiods, nmp));
1418 gotiod = TRUE;
1419 }
1420 }
1421
1422 /*
1423 * If we have an iod which can process the request, then queue
1424 * the buffer.
1425 */
1426 if (gotiod) {
1427 /*
1428 * Ensure that the queue never grows too large. We still want
1429 * to asynchronize so we block rather then return EIO.
1430 */
1431 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1432 NFS_DPF(ASYNCIO,
1433 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1434 nmp->nm_bufqwant = TRUE;
1435 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1436 slpflag | PRIBIO,
1437 "nfsaio", slptimeo);
1438 if (error) {
1439 error2 = nfs_sigintr(nmp, td);
1440 if (error2) {
1441 mtx_unlock(&nfs_iod_mtx);
1442 return (error2);
1443 }
1444 if (slpflag == NFS_PCATCH) {
1445 slpflag = 0;
1446 slptimeo = 2 * hz;
1447 }
1448 }
1449 /*
1450 * We might have lost our iod while sleeping,
1451 * so check and loop if nescessary.
1452 */
1453 goto again;
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_RDIRPLUS) != 0) {
1612 error = nfs_readdirplusrpc(vp, uiop, cr);
1613 if (error == NFSERR_NOTSUPP)
1614 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1615 }
1616 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1617 error = nfs_readdirrpc(vp, uiop, cr);
1618 /*
1619 * end-of-directory sets B_INVAL but does not generate an
1620 * error.
1621 */
1622 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1623 bp->b_flags |= B_INVAL;
1624 break;
1625 default:
1626 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1627 break;
1628 };
1629 if (error) {
1630 bp->b_ioflags |= BIO_ERROR;
1631 bp->b_error = error;
1632 }
1633 } else {
1634 /*
1635 * If we only need to commit, try to commit
1636 */
1637 if (bp->b_flags & B_NEEDCOMMIT) {
1638 int retv;
1639 off_t off;
1640
1641 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1642 retv = (nmp->nm_rpcops->nr_commit)(
1643 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1644 bp->b_wcred, td);
1645 if (retv == 0) {
1646 bp->b_dirtyoff = bp->b_dirtyend = 0;
1647 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1648 bp->b_resid = 0;
1649 bufdone(bp);
1650 return (0);
1651 }
1652 if (retv == NFSERR_STALEWRITEVERF) {
1653 nfs_clearcommit(vp->v_mount);
1654 }
1655 }
1656
1657 /*
1658 * Setup for actual write
1659 */
1660 mtx_lock(&np->n_mtx);
1661 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1662 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1663 mtx_unlock(&np->n_mtx);
1664
1665 if (bp->b_dirtyend > bp->b_dirtyoff) {
1666 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1667 - bp->b_dirtyoff;
1668 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1669 + bp->b_dirtyoff;
1670 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1671 uiop->uio_rw = UIO_WRITE;
1672 nfsstats.write_bios++;
1673
1674 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1675 iomode = NFSV3WRITE_UNSTABLE;
1676 else
1677 iomode = NFSV3WRITE_FILESYNC;
1678
1679 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1680
1681 /*
1682 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1683 * to cluster the buffers needing commit. This will allow
1684 * the system to submit a single commit rpc for the whole
1685 * cluster. We can do this even if the buffer is not 100%
1686 * dirty (relative to the NFS blocksize), so we optimize the
1687 * append-to-file-case.
1688 *
1689 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1690 * cleared because write clustering only works for commit
1691 * rpc's, not for the data portion of the write).
1692 */
1693
1694 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1695 bp->b_flags |= B_NEEDCOMMIT;
1696 if (bp->b_dirtyoff == 0
1697 && bp->b_dirtyend == bp->b_bcount)
1698 bp->b_flags |= B_CLUSTEROK;
1699 } else {
1700 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1701 }
1702
1703 /*
1704 * For an interrupted write, the buffer is still valid
1705 * and the write hasn't been pushed to the server yet,
1706 * so we can't set BIO_ERROR and report the interruption
1707 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1708 * is not relevant, so the rpc attempt is essentially
1709 * a noop. For the case of a V3 write rpc not being
1710 * committed to stable storage, the block is still
1711 * dirty and requires either a commit rpc or another
1712 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1713 * the block is reused. This is indicated by setting
1714 * the B_DELWRI and B_NEEDCOMMIT flags.
1715 *
1716 * If the buffer is marked B_PAGING, it does not reside on
1717 * the vp's paging queues so we cannot call bdirty(). The
1718 * bp in this case is not an NFS cache block so we should
1719 * be safe. XXX
1720 *
1721 * The logic below breaks up errors into recoverable and
1722 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1723 * and keep the buffer around for potential write retries.
1724 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1725 * and save the error in the nfsnode. This is less than ideal
1726 * but necessary. Keeping such buffers around could potentially
1727 * cause buffer exhaustion eventually (they can never be written
1728 * out, so will get constantly be re-dirtied). It also causes
1729 * all sorts of vfs panics. For non-recoverable write errors,
1730 * also invalidate the attrcache, so we'll be forced to go over
1731 * the wire for this object, returning an error to user on next
1732 * call (most of the time).
1733 */
1734 if (error == EINTR || error == EIO || error == ETIMEDOUT
1735 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1736 int s;
1737
1738 s = splbio();
1739 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1740 if ((bp->b_flags & B_PAGING) == 0) {
1741 bdirty(bp);
1742 bp->b_flags &= ~B_DONE;
1743 }
1744 if (error && (bp->b_flags & B_ASYNC) == 0)
1745 bp->b_flags |= B_EINTR;
1746 splx(s);
1747 } else {
1748 if (error) {
1749 bp->b_ioflags |= BIO_ERROR;
1750 bp->b_flags |= B_INVAL;
1751 bp->b_error = np->n_error = error;
1752 mtx_lock(&np->n_mtx);
1753 np->n_flag |= NWRITEERR;
1754 np->n_attrstamp = 0;
1755 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1756 mtx_unlock(&np->n_mtx);
1757 }
1758 bp->b_dirtyoff = bp->b_dirtyend = 0;
1759 }
1760 } else {
1761 bp->b_resid = 0;
1762 bufdone(bp);
1763 return (0);
1764 }
1765 }
1766 bp->b_resid = uiop->uio_resid;
1767 if (must_commit)
1768 nfs_clearcommit(vp->v_mount);
1769 bufdone(bp);
1770 return (error);
1771 }
1772
1773 /*
1774 * Used to aid in handling ftruncate() operations on the NFS client side.
1775 * Truncation creates a number of special problems for NFS. We have to
1776 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1777 * we have to properly handle VM pages or (potentially dirty) buffers
1778 * that straddle the truncation point.
1779 */
1780
1781 int
1782 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1783 {
1784 struct nfsnode *np = VTONFS(vp);
1785 u_quad_t tsize;
1786 int biosize = vp->v_bufobj.bo_bsize;
1787 int error = 0;
1788
1789 mtx_lock(&np->n_mtx);
1790 tsize = np->n_size;
1791 np->n_size = nsize;
1792 mtx_unlock(&np->n_mtx);
1793
1794 if (nsize < tsize) {
1795 struct buf *bp;
1796 daddr_t lbn;
1797 int bufsize;
1798
1799 /*
1800 * vtruncbuf() doesn't get the buffer overlapping the
1801 * truncation point. We may have a B_DELWRI and/or B_CACHE
1802 * buffer that now needs to be truncated.
1803 */
1804 error = vtruncbuf(vp, cred, td, nsize, biosize);
1805 lbn = nsize / biosize;
1806 bufsize = nsize & (biosize - 1);
1807 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1808 if (!bp)
1809 return EINTR;
1810 if (bp->b_dirtyoff > bp->b_bcount)
1811 bp->b_dirtyoff = bp->b_bcount;
1812 if (bp->b_dirtyend > bp->b_bcount)
1813 bp->b_dirtyend = bp->b_bcount;
1814 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1815 brelse(bp);
1816 } else {
1817 vnode_pager_setsize(vp, nsize);
1818 }
1819 return(error);
1820 }
1821
Cache object: d0450d2f5afea28b980819ac88e69c5d
|