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/8.2/sys/nfsclient/nfs_bio.c 214977 2010-11-08 10:23:39Z kib $");
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/resourcevar.h>
49 #include <sys/signalvar.h>
50 #include <sys/vmmeter.h>
51 #include <sys/vnode.h>
52
53 #include <vm/vm.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_object.h>
57 #include <vm/vm_pager.h>
58 #include <vm/vnode_pager.h>
59
60 #include <nfs/nfsproto.h>
61 #include <nfsclient/nfs.h>
62 #include <nfsclient/nfsmount.h>
63 #include <nfsclient/nfsnode.h>
64 #include <nfsclient/nfs_kdtrace.h>
65
66 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
67 struct thread *td);
68 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
69 struct ucred *cred, int ioflag);
70
71 extern int nfs_directio_enable;
72 extern int nfs_directio_allow_mmap;
73
74 /*
75 * Vnode op for VM getpages.
76 */
77 int
78 nfs_getpages(struct vop_getpages_args *ap)
79 {
80 int i, error, nextoff, size, toff, count, npages;
81 struct uio uio;
82 struct iovec iov;
83 vm_offset_t kva;
84 struct buf *bp;
85 struct vnode *vp;
86 struct thread *td;
87 struct ucred *cred;
88 struct nfsmount *nmp;
89 vm_object_t object;
90 vm_page_t *pages;
91 struct nfsnode *np;
92
93 vp = ap->a_vp;
94 np = VTONFS(vp);
95 td = curthread; /* XXX */
96 cred = curthread->td_ucred; /* XXX */
97 nmp = VFSTONFS(vp->v_mount);
98 pages = ap->a_m;
99 count = ap->a_count;
100
101 if ((object = vp->v_object) == NULL) {
102 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n");
103 return (VM_PAGER_ERROR);
104 }
105
106 if (nfs_directio_enable && !nfs_directio_allow_mmap) {
107 mtx_lock(&np->n_mtx);
108 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
109 mtx_unlock(&np->n_mtx);
110 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n");
111 return (VM_PAGER_ERROR);
112 } else
113 mtx_unlock(&np->n_mtx);
114 }
115
116 mtx_lock(&nmp->nm_mtx);
117 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
118 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
119 mtx_unlock(&nmp->nm_mtx);
120 /* We'll never get here for v4, because we always have fsinfo */
121 (void)nfs_fsinfo(nmp, vp, cred, td);
122 } else
123 mtx_unlock(&nmp->nm_mtx);
124
125 npages = btoc(count);
126
127 /*
128 * If the requested page is partially valid, just return it and
129 * allow the pager to zero-out the blanks. Partially valid pages
130 * can only occur at the file EOF.
131 */
132 VM_OBJECT_LOCK(object);
133 if (pages[ap->a_reqpage]->valid != 0) {
134 vm_page_lock_queues();
135 for (i = 0; i < npages; ++i) {
136 if (i != ap->a_reqpage)
137 vm_page_free(pages[i]);
138 }
139 vm_page_unlock_queues();
140 VM_OBJECT_UNLOCK(object);
141 return (0);
142 }
143 VM_OBJECT_UNLOCK(object);
144
145 /*
146 * We use only the kva address for the buffer, but this is extremely
147 * convienient and fast.
148 */
149 bp = getpbuf(&nfs_pbuf_freecnt);
150
151 kva = (vm_offset_t) bp->b_data;
152 pmap_qenter(kva, pages, npages);
153 PCPU_INC(cnt.v_vnodein);
154 PCPU_ADD(cnt.v_vnodepgsin, npages);
155
156 iov.iov_base = (caddr_t) kva;
157 iov.iov_len = count;
158 uio.uio_iov = &iov;
159 uio.uio_iovcnt = 1;
160 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
161 uio.uio_resid = count;
162 uio.uio_segflg = UIO_SYSSPACE;
163 uio.uio_rw = UIO_READ;
164 uio.uio_td = td;
165
166 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
167 pmap_qremove(kva, npages);
168
169 relpbuf(bp, &nfs_pbuf_freecnt);
170
171 if (error && (uio.uio_resid == count)) {
172 nfs_printf("nfs_getpages: error %d\n", error);
173 VM_OBJECT_LOCK(object);
174 vm_page_lock_queues();
175 for (i = 0; i < npages; ++i) {
176 if (i != ap->a_reqpage)
177 vm_page_free(pages[i]);
178 }
179 vm_page_unlock_queues();
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 vm_page_lock_queues();
193 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
194 vm_page_t m;
195 nextoff = toff + PAGE_SIZE;
196 m = pages[i];
197
198 if (nextoff <= size) {
199 /*
200 * Read operation filled an entire page
201 */
202 m->valid = VM_PAGE_BITS_ALL;
203 KASSERT(m->dirty == 0,
204 ("nfs_getpages: page %p is dirty", m));
205 } else if (size > toff) {
206 /*
207 * Read operation filled a partial page.
208 */
209 m->valid = 0;
210 vm_page_set_valid(m, 0, size - toff);
211 KASSERT(m->dirty == 0,
212 ("nfs_getpages: page %p is dirty", m));
213 } else {
214 /*
215 * Read operation was short. If no error occured
216 * we may have hit a zero-fill section. We simply
217 * leave valid set to 0.
218 */
219 ;
220 }
221 if (i != ap->a_reqpage) {
222 /*
223 * Whether or not to leave the page activated is up in
224 * the air, but we should put the page on a page queue
225 * somewhere (it already is in the object). Result:
226 * It appears that emperical results show that
227 * deactivating pages is best.
228 */
229
230 /*
231 * Just in case someone was asking for this page we
232 * now tell them that it is ok to use.
233 */
234 if (!error) {
235 if (m->oflags & VPO_WANTED)
236 vm_page_activate(m);
237 else
238 vm_page_deactivate(m);
239 vm_page_wakeup(m);
240 } else {
241 vm_page_free(m);
242 }
243 }
244 }
245 vm_page_unlock_queues();
246 VM_OBJECT_UNLOCK(object);
247 return (0);
248 }
249
250 /*
251 * Vnode op for VM putpages.
252 */
253 int
254 nfs_putpages(struct vop_putpages_args *ap)
255 {
256 struct uio uio;
257 struct iovec iov;
258 vm_offset_t kva;
259 struct buf *bp;
260 int iomode, must_commit, i, error, npages, count;
261 off_t offset;
262 int *rtvals;
263 struct vnode *vp;
264 struct thread *td;
265 struct ucred *cred;
266 struct nfsmount *nmp;
267 struct nfsnode *np;
268 vm_page_t *pages;
269
270 vp = ap->a_vp;
271 np = VTONFS(vp);
272 td = curthread; /* XXX */
273 cred = curthread->td_ucred; /* XXX */
274 nmp = VFSTONFS(vp->v_mount);
275 pages = ap->a_m;
276 count = ap->a_count;
277 rtvals = ap->a_rtvals;
278 npages = btoc(count);
279 offset = IDX_TO_OFF(pages[0]->pindex);
280
281 mtx_lock(&nmp->nm_mtx);
282 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
283 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
284 mtx_unlock(&nmp->nm_mtx);
285 (void)nfs_fsinfo(nmp, vp, cred, td);
286 } else
287 mtx_unlock(&nmp->nm_mtx);
288
289 mtx_lock(&np->n_mtx);
290 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
291 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
292 mtx_unlock(&np->n_mtx);
293 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
294 mtx_lock(&np->n_mtx);
295 }
296
297 for (i = 0; i < npages; i++)
298 rtvals[i] = VM_PAGER_AGAIN;
299
300 /*
301 * When putting pages, do not extend file past EOF.
302 */
303 if (offset + count > np->n_size) {
304 count = np->n_size - offset;
305 if (count < 0)
306 count = 0;
307 }
308 mtx_unlock(&np->n_mtx);
309
310 /*
311 * We use only the kva address for the buffer, but this is extremely
312 * convienient and fast.
313 */
314 bp = getpbuf(&nfs_pbuf_freecnt);
315
316 kva = (vm_offset_t) bp->b_data;
317 pmap_qenter(kva, pages, npages);
318 PCPU_INC(cnt.v_vnodeout);
319 PCPU_ADD(cnt.v_vnodepgsout, count);
320
321 iov.iov_base = (caddr_t) kva;
322 iov.iov_len = count;
323 uio.uio_iov = &iov;
324 uio.uio_iovcnt = 1;
325 uio.uio_offset = offset;
326 uio.uio_resid = count;
327 uio.uio_segflg = UIO_SYSSPACE;
328 uio.uio_rw = UIO_WRITE;
329 uio.uio_td = td;
330
331 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
332 iomode = NFSV3WRITE_UNSTABLE;
333 else
334 iomode = NFSV3WRITE_FILESYNC;
335
336 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
337
338 pmap_qremove(kva, npages);
339 relpbuf(bp, &nfs_pbuf_freecnt);
340
341 if (!error) {
342 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
343 for (i = 0; i < nwritten; i++) {
344 rtvals[i] = VM_PAGER_OK;
345 vm_page_undirty(pages[i]);
346 }
347 if (must_commit) {
348 nfs_clearcommit(vp->v_mount);
349 }
350 }
351 return rtvals[0];
352 }
353
354 /*
355 * For nfs, cache consistency can only be maintained approximately.
356 * Although RFC1094 does not specify the criteria, the following is
357 * believed to be compatible with the reference port.
358 * For nfs:
359 * If the file's modify time on the server has changed since the
360 * last read rpc or you have written to the file,
361 * you may have lost data cache consistency with the
362 * server, so flush all of the file's data out of the cache.
363 * Then force a getattr rpc to ensure that you have up to date
364 * attributes.
365 * NB: This implies that cache data can be read when up to
366 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
367 * attributes this could be forced by setting n_attrstamp to 0 before
368 * the VOP_GETATTR() call.
369 */
370 static inline int
371 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
372 {
373 int error = 0;
374 struct vattr vattr;
375 struct nfsnode *np = VTONFS(vp);
376 int old_lock;
377 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
378
379 /*
380 * Grab the exclusive lock before checking whether the cache is
381 * consistent.
382 * XXX - We can make this cheaper later (by acquiring cheaper locks).
383 * But for now, this suffices.
384 */
385 old_lock = nfs_upgrade_vnlock(vp);
386 if (vp->v_iflag & VI_DOOMED) {
387 nfs_downgrade_vnlock(vp, old_lock);
388 return (EBADF);
389 }
390
391 mtx_lock(&np->n_mtx);
392 if (np->n_flag & NMODIFIED) {
393 mtx_unlock(&np->n_mtx);
394 if (vp->v_type != VREG) {
395 if (vp->v_type != VDIR)
396 panic("nfs: bioread, not dir");
397 (nmp->nm_rpcops->nr_invaldir)(vp);
398 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
399 if (error)
400 goto out;
401 }
402 np->n_attrstamp = 0;
403 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
404 error = VOP_GETATTR(vp, &vattr, cred);
405 if (error)
406 goto out;
407 mtx_lock(&np->n_mtx);
408 np->n_mtime = vattr.va_mtime;
409 mtx_unlock(&np->n_mtx);
410 } else {
411 mtx_unlock(&np->n_mtx);
412 error = VOP_GETATTR(vp, &vattr, cred);
413 if (error)
414 return (error);
415 mtx_lock(&np->n_mtx);
416 if ((np->n_flag & NSIZECHANGED)
417 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
418 mtx_unlock(&np->n_mtx);
419 if (vp->v_type == VDIR)
420 (nmp->nm_rpcops->nr_invaldir)(vp);
421 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
422 if (error)
423 goto out;
424 mtx_lock(&np->n_mtx);
425 np->n_mtime = vattr.va_mtime;
426 np->n_flag &= ~NSIZECHANGED;
427 }
428 mtx_unlock(&np->n_mtx);
429 }
430 out:
431 nfs_downgrade_vnlock(vp, old_lock);
432 return error;
433 }
434
435 /*
436 * Vnode op for read using bio
437 */
438 int
439 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
440 {
441 struct nfsnode *np = VTONFS(vp);
442 int biosize, i;
443 struct buf *bp, *rabp;
444 struct thread *td;
445 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
446 daddr_t lbn, rabn;
447 int bcount;
448 int seqcount;
449 int nra, error = 0, n = 0, on = 0;
450
451 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
452 if (uio->uio_resid == 0)
453 return (0);
454 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
455 return (EINVAL);
456 td = uio->uio_td;
457
458 mtx_lock(&nmp->nm_mtx);
459 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
460 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
461 mtx_unlock(&nmp->nm_mtx);
462 (void)nfs_fsinfo(nmp, vp, cred, td);
463 } else
464 mtx_unlock(&nmp->nm_mtx);
465
466 if (vp->v_type != VDIR &&
467 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
468 return (EFBIG);
469
470 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
471 /* No caching/ no readaheads. Just read data into the user buffer */
472 return nfs_readrpc(vp, uio, cred);
473
474 biosize = vp->v_mount->mnt_stat.f_iosize;
475 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
476
477 error = nfs_bioread_check_cons(vp, td, cred);
478 if (error)
479 return error;
480
481 do {
482 u_quad_t nsize;
483
484 mtx_lock(&np->n_mtx);
485 nsize = np->n_size;
486 mtx_unlock(&np->n_mtx);
487
488 switch (vp->v_type) {
489 case VREG:
490 nfsstats.biocache_reads++;
491 lbn = uio->uio_offset / biosize;
492 on = uio->uio_offset & (biosize - 1);
493
494 /*
495 * Start the read ahead(s), as required.
496 */
497 if (nmp->nm_readahead > 0) {
498 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
499 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
500 rabn = lbn + 1 + nra;
501 if (incore(&vp->v_bufobj, rabn) == NULL) {
502 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
503 if (!rabp) {
504 error = nfs_sigintr(nmp, td);
505 return (error ? error : EINTR);
506 }
507 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
508 rabp->b_flags |= B_ASYNC;
509 rabp->b_iocmd = BIO_READ;
510 vfs_busy_pages(rabp, 0);
511 if (nfs_asyncio(nmp, rabp, cred, td)) {
512 rabp->b_flags |= B_INVAL;
513 rabp->b_ioflags |= BIO_ERROR;
514 vfs_unbusy_pages(rabp);
515 brelse(rabp);
516 break;
517 }
518 } else {
519 brelse(rabp);
520 }
521 }
522 }
523 }
524
525 /* Note that bcount is *not* DEV_BSIZE aligned. */
526 bcount = biosize;
527 if ((off_t)lbn * biosize >= nsize) {
528 bcount = 0;
529 } else if ((off_t)(lbn + 1) * biosize > nsize) {
530 bcount = nsize - (off_t)lbn * biosize;
531 }
532 bp = nfs_getcacheblk(vp, lbn, bcount, td);
533
534 if (!bp) {
535 error = nfs_sigintr(nmp, td);
536 return (error ? error : EINTR);
537 }
538
539 /*
540 * If B_CACHE is not set, we must issue the read. If this
541 * fails, we return an error.
542 */
543
544 if ((bp->b_flags & B_CACHE) == 0) {
545 bp->b_iocmd = BIO_READ;
546 vfs_busy_pages(bp, 0);
547 error = nfs_doio(vp, bp, cred, td);
548 if (error) {
549 brelse(bp);
550 return (error);
551 }
552 }
553
554 /*
555 * on is the offset into the current bp. Figure out how many
556 * bytes we can copy out of the bp. Note that bcount is
557 * NOT DEV_BSIZE aligned.
558 *
559 * Then figure out how many bytes we can copy into the uio.
560 */
561
562 n = 0;
563 if (on < bcount)
564 n = min((unsigned)(bcount - on), uio->uio_resid);
565 break;
566 case VLNK:
567 nfsstats.biocache_readlinks++;
568 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
569 if (!bp) {
570 error = nfs_sigintr(nmp, td);
571 return (error ? error : EINTR);
572 }
573 if ((bp->b_flags & B_CACHE) == 0) {
574 bp->b_iocmd = BIO_READ;
575 vfs_busy_pages(bp, 0);
576 error = nfs_doio(vp, bp, cred, td);
577 if (error) {
578 bp->b_ioflags |= BIO_ERROR;
579 brelse(bp);
580 return (error);
581 }
582 }
583 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
584 on = 0;
585 break;
586 case VDIR:
587 nfsstats.biocache_readdirs++;
588 if (np->n_direofoffset
589 && uio->uio_offset >= np->n_direofoffset) {
590 return (0);
591 }
592 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
593 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
594 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
595 if (!bp) {
596 error = nfs_sigintr(nmp, td);
597 return (error ? error : EINTR);
598 }
599 if ((bp->b_flags & B_CACHE) == 0) {
600 bp->b_iocmd = BIO_READ;
601 vfs_busy_pages(bp, 0);
602 error = nfs_doio(vp, bp, cred, td);
603 if (error) {
604 brelse(bp);
605 }
606 while (error == NFSERR_BAD_COOKIE) {
607 (nmp->nm_rpcops->nr_invaldir)(vp);
608 error = nfs_vinvalbuf(vp, 0, td, 1);
609 /*
610 * Yuck! The directory has been modified on the
611 * server. The only way to get the block is by
612 * reading from the beginning to get all the
613 * offset cookies.
614 *
615 * Leave the last bp intact unless there is an error.
616 * Loop back up to the while if the error is another
617 * NFSERR_BAD_COOKIE (double yuch!).
618 */
619 for (i = 0; i <= lbn && !error; i++) {
620 if (np->n_direofoffset
621 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
622 return (0);
623 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
624 if (!bp) {
625 error = nfs_sigintr(nmp, td);
626 return (error ? error : EINTR);
627 }
628 if ((bp->b_flags & B_CACHE) == 0) {
629 bp->b_iocmd = BIO_READ;
630 vfs_busy_pages(bp, 0);
631 error = nfs_doio(vp, bp, cred, td);
632 /*
633 * no error + B_INVAL == directory EOF,
634 * use the block.
635 */
636 if (error == 0 && (bp->b_flags & B_INVAL))
637 break;
638 }
639 /*
640 * An error will throw away the block and the
641 * for loop will break out. If no error and this
642 * is not the block we want, we throw away the
643 * block and go for the next one via the for loop.
644 */
645 if (error || i < lbn)
646 brelse(bp);
647 }
648 }
649 /*
650 * The above while is repeated if we hit another cookie
651 * error. If we hit an error and it wasn't a cookie error,
652 * we give up.
653 */
654 if (error)
655 return (error);
656 }
657
658 /*
659 * If not eof and read aheads are enabled, start one.
660 * (You need the current block first, so that you have the
661 * directory offset cookie of the next block.)
662 */
663 if (nmp->nm_readahead > 0 &&
664 (bp->b_flags & B_INVAL) == 0 &&
665 (np->n_direofoffset == 0 ||
666 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
667 incore(&vp->v_bufobj, lbn + 1) == NULL) {
668 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
669 if (rabp) {
670 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
671 rabp->b_flags |= B_ASYNC;
672 rabp->b_iocmd = BIO_READ;
673 vfs_busy_pages(rabp, 0);
674 if (nfs_asyncio(nmp, rabp, cred, td)) {
675 rabp->b_flags |= B_INVAL;
676 rabp->b_ioflags |= BIO_ERROR;
677 vfs_unbusy_pages(rabp);
678 brelse(rabp);
679 }
680 } else {
681 brelse(rabp);
682 }
683 }
684 }
685 /*
686 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
687 * chopped for the EOF condition, we cannot tell how large
688 * NFS directories are going to be until we hit EOF. So
689 * an NFS directory buffer is *not* chopped to its EOF. Now,
690 * it just so happens that b_resid will effectively chop it
691 * to EOF. *BUT* this information is lost if the buffer goes
692 * away and is reconstituted into a B_CACHE state ( due to
693 * being VMIO ) later. So we keep track of the directory eof
694 * in np->n_direofoffset and chop it off as an extra step
695 * right here.
696 */
697 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
698 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
699 n = np->n_direofoffset - uio->uio_offset;
700 break;
701 default:
702 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
703 bp = NULL;
704 break;
705 };
706
707 if (n > 0) {
708 error = uiomove(bp->b_data + on, (int)n, uio);
709 }
710 if (vp->v_type == VLNK)
711 n = 0;
712 if (bp != NULL)
713 brelse(bp);
714 } while (error == 0 && uio->uio_resid > 0 && n > 0);
715 return (error);
716 }
717
718 /*
719 * The NFS write path cannot handle iovecs with len > 1. So we need to
720 * break up iovecs accordingly (restricting them to wsize).
721 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
722 * For the ASYNC case, 2 copies are needed. The first a copy from the
723 * user buffer to a staging buffer and then a second copy from the staging
724 * buffer to mbufs. This can be optimized by copying from the user buffer
725 * directly into mbufs and passing the chain down, but that requires a
726 * fair amount of re-working of the relevant codepaths (and can be done
727 * later).
728 */
729 static int
730 nfs_directio_write(vp, uiop, cred, ioflag)
731 struct vnode *vp;
732 struct uio *uiop;
733 struct ucred *cred;
734 int ioflag;
735 {
736 int error;
737 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
738 struct thread *td = uiop->uio_td;
739 int size;
740 int wsize;
741
742 mtx_lock(&nmp->nm_mtx);
743 wsize = nmp->nm_wsize;
744 mtx_unlock(&nmp->nm_mtx);
745 if (ioflag & IO_SYNC) {
746 int iomode, must_commit;
747 struct uio uio;
748 struct iovec iov;
749 do_sync:
750 while (uiop->uio_resid > 0) {
751 size = min(uiop->uio_resid, wsize);
752 size = min(uiop->uio_iov->iov_len, size);
753 iov.iov_base = uiop->uio_iov->iov_base;
754 iov.iov_len = size;
755 uio.uio_iov = &iov;
756 uio.uio_iovcnt = 1;
757 uio.uio_offset = uiop->uio_offset;
758 uio.uio_resid = size;
759 uio.uio_segflg = UIO_USERSPACE;
760 uio.uio_rw = UIO_WRITE;
761 uio.uio_td = td;
762 iomode = NFSV3WRITE_FILESYNC;
763 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
764 &iomode, &must_commit);
765 KASSERT((must_commit == 0),
766 ("nfs_directio_write: Did not commit write"));
767 if (error)
768 return (error);
769 uiop->uio_offset += size;
770 uiop->uio_resid -= size;
771 if (uiop->uio_iov->iov_len <= size) {
772 uiop->uio_iovcnt--;
773 uiop->uio_iov++;
774 } else {
775 uiop->uio_iov->iov_base =
776 (char *)uiop->uio_iov->iov_base + size;
777 uiop->uio_iov->iov_len -= size;
778 }
779 }
780 } else {
781 struct uio *t_uio;
782 struct iovec *t_iov;
783 struct buf *bp;
784
785 /*
786 * Break up the write into blocksize chunks and hand these
787 * over to nfsiod's for write back.
788 * Unfortunately, this incurs a copy of the data. Since
789 * the user could modify the buffer before the write is
790 * initiated.
791 *
792 * The obvious optimization here is that one of the 2 copies
793 * in the async write path can be eliminated by copying the
794 * data here directly into mbufs and passing the mbuf chain
795 * down. But that will require a fair amount of re-working
796 * of the code and can be done if there's enough interest
797 * in NFS directio access.
798 */
799 while (uiop->uio_resid > 0) {
800 size = min(uiop->uio_resid, wsize);
801 size = min(uiop->uio_iov->iov_len, size);
802 bp = getpbuf(&nfs_pbuf_freecnt);
803 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
804 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
805 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
806 t_iov->iov_len = size;
807 t_uio->uio_iov = t_iov;
808 t_uio->uio_iovcnt = 1;
809 t_uio->uio_offset = uiop->uio_offset;
810 t_uio->uio_resid = size;
811 t_uio->uio_segflg = UIO_SYSSPACE;
812 t_uio->uio_rw = UIO_WRITE;
813 t_uio->uio_td = td;
814 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size);
815 bp->b_flags |= B_DIRECT;
816 bp->b_iocmd = BIO_WRITE;
817 if (cred != NOCRED) {
818 crhold(cred);
819 bp->b_wcred = cred;
820 } else
821 bp->b_wcred = NOCRED;
822 bp->b_caller1 = (void *)t_uio;
823 bp->b_vp = vp;
824 error = nfs_asyncio(nmp, bp, NOCRED, td);
825 if (error) {
826 free(t_iov->iov_base, M_NFSDIRECTIO);
827 free(t_iov, M_NFSDIRECTIO);
828 free(t_uio, M_NFSDIRECTIO);
829 bp->b_vp = NULL;
830 relpbuf(bp, &nfs_pbuf_freecnt);
831 if (error == EINTR)
832 return (error);
833 goto do_sync;
834 }
835 uiop->uio_offset += size;
836 uiop->uio_resid -= size;
837 if (uiop->uio_iov->iov_len <= size) {
838 uiop->uio_iovcnt--;
839 uiop->uio_iov++;
840 } else {
841 uiop->uio_iov->iov_base =
842 (char *)uiop->uio_iov->iov_base + size;
843 uiop->uio_iov->iov_len -= size;
844 }
845 }
846 }
847 return (0);
848 }
849
850 /*
851 * Vnode op for write using bio
852 */
853 int
854 nfs_write(struct vop_write_args *ap)
855 {
856 int biosize;
857 struct uio *uio = ap->a_uio;
858 struct thread *td = uio->uio_td;
859 struct vnode *vp = ap->a_vp;
860 struct nfsnode *np = VTONFS(vp);
861 struct ucred *cred = ap->a_cred;
862 int ioflag = ap->a_ioflag;
863 struct buf *bp;
864 struct vattr vattr;
865 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
866 daddr_t lbn;
867 int bcount;
868 int n, on, error = 0;
869 struct proc *p = td?td->td_proc:NULL;
870
871 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode"));
872 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
873 ("nfs_write proc"));
874 if (vp->v_type != VREG)
875 return (EIO);
876 mtx_lock(&np->n_mtx);
877 if (np->n_flag & NWRITEERR) {
878 np->n_flag &= ~NWRITEERR;
879 mtx_unlock(&np->n_mtx);
880 return (np->n_error);
881 } else
882 mtx_unlock(&np->n_mtx);
883 mtx_lock(&nmp->nm_mtx);
884 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
885 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
886 mtx_unlock(&nmp->nm_mtx);
887 (void)nfs_fsinfo(nmp, vp, cred, td);
888 } else
889 mtx_unlock(&nmp->nm_mtx);
890
891 /*
892 * Synchronously flush pending buffers if we are in synchronous
893 * mode or if we are appending.
894 */
895 if (ioflag & (IO_APPEND | IO_SYNC)) {
896 mtx_lock(&np->n_mtx);
897 if (np->n_flag & NMODIFIED) {
898 mtx_unlock(&np->n_mtx);
899 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
900 /*
901 * Require non-blocking, synchronous writes to
902 * dirty files to inform the program it needs
903 * to fsync(2) explicitly.
904 */
905 if (ioflag & IO_NDELAY)
906 return (EAGAIN);
907 #endif
908 flush_and_restart:
909 np->n_attrstamp = 0;
910 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
911 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
912 if (error)
913 return (error);
914 } else
915 mtx_unlock(&np->n_mtx);
916 }
917
918 /*
919 * If IO_APPEND then load uio_offset. We restart here if we cannot
920 * get the append lock.
921 */
922 if (ioflag & IO_APPEND) {
923 np->n_attrstamp = 0;
924 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
925 error = VOP_GETATTR(vp, &vattr, cred);
926 if (error)
927 return (error);
928 mtx_lock(&np->n_mtx);
929 uio->uio_offset = np->n_size;
930 mtx_unlock(&np->n_mtx);
931 }
932
933 if (uio->uio_offset < 0)
934 return (EINVAL);
935 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
936 return (EFBIG);
937 if (uio->uio_resid == 0)
938 return (0);
939
940 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
941 return nfs_directio_write(vp, uio, cred, ioflag);
942
943 /*
944 * Maybe this should be above the vnode op call, but so long as
945 * file servers have no limits, i don't think it matters
946 */
947 if (p != NULL) {
948 PROC_LOCK(p);
949 if (uio->uio_offset + uio->uio_resid >
950 lim_cur(p, RLIMIT_FSIZE)) {
951 psignal(p, SIGXFSZ);
952 PROC_UNLOCK(p);
953 return (EFBIG);
954 }
955 PROC_UNLOCK(p);
956 }
957
958 biosize = vp->v_mount->mnt_stat.f_iosize;
959 /*
960 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
961 * would exceed the local maximum per-file write commit size when
962 * combined with those, we must decide whether to flush,
963 * go synchronous, or return error. We don't bother checking
964 * IO_UNIT -- we just make all writes atomic anyway, as there's
965 * no point optimizing for something that really won't ever happen.
966 */
967 if (!(ioflag & IO_SYNC)) {
968 int nflag;
969
970 mtx_lock(&np->n_mtx);
971 nflag = np->n_flag;
972 mtx_unlock(&np->n_mtx);
973 int needrestart = 0;
974 if (nmp->nm_wcommitsize < uio->uio_resid) {
975 /*
976 * If this request could not possibly be completed
977 * without exceeding the maximum outstanding write
978 * commit size, see if we can convert it into a
979 * synchronous write operation.
980 */
981 if (ioflag & IO_NDELAY)
982 return (EAGAIN);
983 ioflag |= IO_SYNC;
984 if (nflag & NMODIFIED)
985 needrestart = 1;
986 } else if (nflag & NMODIFIED) {
987 int wouldcommit = 0;
988 BO_LOCK(&vp->v_bufobj);
989 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
990 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
991 b_bobufs) {
992 if (bp->b_flags & B_NEEDCOMMIT)
993 wouldcommit += bp->b_bcount;
994 }
995 }
996 BO_UNLOCK(&vp->v_bufobj);
997 /*
998 * Since we're not operating synchronously and
999 * bypassing the buffer cache, we are in a commit
1000 * and holding all of these buffers whether
1001 * transmitted or not. If not limited, this
1002 * will lead to the buffer cache deadlocking,
1003 * as no one else can flush our uncommitted buffers.
1004 */
1005 wouldcommit += uio->uio_resid;
1006 /*
1007 * If we would initially exceed the maximum
1008 * outstanding write commit size, flush and restart.
1009 */
1010 if (wouldcommit > nmp->nm_wcommitsize)
1011 needrestart = 1;
1012 }
1013 if (needrestart)
1014 goto flush_and_restart;
1015 }
1016
1017 do {
1018 nfsstats.biocache_writes++;
1019 lbn = uio->uio_offset / biosize;
1020 on = uio->uio_offset & (biosize-1);
1021 n = min((unsigned)(biosize - on), uio->uio_resid);
1022 again:
1023 /*
1024 * Handle direct append and file extension cases, calculate
1025 * unaligned buffer size.
1026 */
1027 mtx_lock(&np->n_mtx);
1028 if (uio->uio_offset == np->n_size && n) {
1029 mtx_unlock(&np->n_mtx);
1030 /*
1031 * Get the buffer (in its pre-append state to maintain
1032 * B_CACHE if it was previously set). Resize the
1033 * nfsnode after we have locked the buffer to prevent
1034 * readers from reading garbage.
1035 */
1036 bcount = on;
1037 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1038
1039 if (bp != NULL) {
1040 long save;
1041
1042 mtx_lock(&np->n_mtx);
1043 np->n_size = uio->uio_offset + n;
1044 np->n_flag |= NMODIFIED;
1045 vnode_pager_setsize(vp, np->n_size);
1046 mtx_unlock(&np->n_mtx);
1047
1048 save = bp->b_flags & B_CACHE;
1049 bcount += n;
1050 allocbuf(bp, bcount);
1051 bp->b_flags |= save;
1052 }
1053 } else {
1054 /*
1055 * Obtain the locked cache block first, and then
1056 * adjust the file's size as appropriate.
1057 */
1058 bcount = on + n;
1059 if ((off_t)lbn * biosize + bcount < np->n_size) {
1060 if ((off_t)(lbn + 1) * biosize < np->n_size)
1061 bcount = biosize;
1062 else
1063 bcount = np->n_size - (off_t)lbn * biosize;
1064 }
1065 mtx_unlock(&np->n_mtx);
1066 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1067 mtx_lock(&np->n_mtx);
1068 if (uio->uio_offset + n > np->n_size) {
1069 np->n_size = uio->uio_offset + n;
1070 np->n_flag |= NMODIFIED;
1071 vnode_pager_setsize(vp, np->n_size);
1072 }
1073 mtx_unlock(&np->n_mtx);
1074 }
1075
1076 if (!bp) {
1077 error = nfs_sigintr(nmp, td);
1078 if (!error)
1079 error = EINTR;
1080 break;
1081 }
1082
1083 /*
1084 * Issue a READ if B_CACHE is not set. In special-append
1085 * mode, B_CACHE is based on the buffer prior to the write
1086 * op and is typically set, avoiding the read. If a read
1087 * is required in special append mode, the server will
1088 * probably send us a short-read since we extended the file
1089 * on our end, resulting in b_resid == 0 and, thusly,
1090 * B_CACHE getting set.
1091 *
1092 * We can also avoid issuing the read if the write covers
1093 * the entire buffer. We have to make sure the buffer state
1094 * is reasonable in this case since we will not be initiating
1095 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1096 * more information.
1097 *
1098 * B_CACHE may also be set due to the buffer being cached
1099 * normally.
1100 */
1101
1102 if (on == 0 && n == bcount) {
1103 bp->b_flags |= B_CACHE;
1104 bp->b_flags &= ~B_INVAL;
1105 bp->b_ioflags &= ~BIO_ERROR;
1106 }
1107
1108 if ((bp->b_flags & B_CACHE) == 0) {
1109 bp->b_iocmd = BIO_READ;
1110 vfs_busy_pages(bp, 0);
1111 error = nfs_doio(vp, bp, cred, td);
1112 if (error) {
1113 brelse(bp);
1114 break;
1115 }
1116 }
1117 if (bp->b_wcred == NOCRED)
1118 bp->b_wcred = crhold(cred);
1119 mtx_lock(&np->n_mtx);
1120 np->n_flag |= NMODIFIED;
1121 mtx_unlock(&np->n_mtx);
1122
1123 /*
1124 * If dirtyend exceeds file size, chop it down. This should
1125 * not normally occur but there is an append race where it
1126 * might occur XXX, so we log it.
1127 *
1128 * If the chopping creates a reverse-indexed or degenerate
1129 * situation with dirtyoff/end, we 0 both of them.
1130 */
1131
1132 if (bp->b_dirtyend > bcount) {
1133 nfs_printf("NFS append race @%lx:%d\n",
1134 (long)bp->b_blkno * DEV_BSIZE,
1135 bp->b_dirtyend - bcount);
1136 bp->b_dirtyend = bcount;
1137 }
1138
1139 if (bp->b_dirtyoff >= bp->b_dirtyend)
1140 bp->b_dirtyoff = bp->b_dirtyend = 0;
1141
1142 /*
1143 * If the new write will leave a contiguous dirty
1144 * area, just update the b_dirtyoff and b_dirtyend,
1145 * otherwise force a write rpc of the old dirty area.
1146 *
1147 * While it is possible to merge discontiguous writes due to
1148 * our having a B_CACHE buffer ( and thus valid read data
1149 * for the hole), we don't because it could lead to
1150 * significant cache coherency problems with multiple clients,
1151 * especially if locking is implemented later on.
1152 *
1153 * as an optimization we could theoretically maintain
1154 * a linked list of discontinuous areas, but we would still
1155 * have to commit them separately so there isn't much
1156 * advantage to it except perhaps a bit of asynchronization.
1157 */
1158
1159 if (bp->b_dirtyend > 0 &&
1160 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1161 if (bwrite(bp) == EINTR) {
1162 error = EINTR;
1163 break;
1164 }
1165 goto again;
1166 }
1167
1168 error = uiomove((char *)bp->b_data + on, n, uio);
1169
1170 /*
1171 * Since this block is being modified, it must be written
1172 * again and not just committed. Since write clustering does
1173 * not work for the stage 1 data write, only the stage 2
1174 * commit rpc, we have to clear B_CLUSTEROK as well.
1175 */
1176 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1177
1178 if (error) {
1179 bp->b_ioflags |= BIO_ERROR;
1180 brelse(bp);
1181 break;
1182 }
1183
1184 /*
1185 * Only update dirtyoff/dirtyend if not a degenerate
1186 * condition.
1187 */
1188 if (n) {
1189 if (bp->b_dirtyend > 0) {
1190 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1191 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1192 } else {
1193 bp->b_dirtyoff = on;
1194 bp->b_dirtyend = on + n;
1195 }
1196 vfs_bio_set_valid(bp, on, n);
1197 }
1198
1199 /*
1200 * If IO_SYNC do bwrite().
1201 *
1202 * IO_INVAL appears to be unused. The idea appears to be
1203 * to turn off caching in this case. Very odd. XXX
1204 */
1205 if ((ioflag & IO_SYNC)) {
1206 if (ioflag & IO_INVAL)
1207 bp->b_flags |= B_NOCACHE;
1208 error = bwrite(bp);
1209 if (error)
1210 break;
1211 } else if ((n + on) == biosize) {
1212 bp->b_flags |= B_ASYNC;
1213 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1214 } else {
1215 bdwrite(bp);
1216 }
1217 } while (uio->uio_resid > 0 && n > 0);
1218
1219 return (error);
1220 }
1221
1222 /*
1223 * Get an nfs cache block.
1224 *
1225 * Allocate a new one if the block isn't currently in the cache
1226 * and return the block marked busy. If the calling process is
1227 * interrupted by a signal for an interruptible mount point, return
1228 * NULL.
1229 *
1230 * The caller must carefully deal with the possible B_INVAL state of
1231 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1232 * indirectly), so synchronous reads can be issued without worrying about
1233 * the B_INVAL state. We have to be a little more careful when dealing
1234 * with writes (see comments in nfs_write()) when extending a file past
1235 * its EOF.
1236 */
1237 static struct buf *
1238 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1239 {
1240 struct buf *bp;
1241 struct mount *mp;
1242 struct nfsmount *nmp;
1243
1244 mp = vp->v_mount;
1245 nmp = VFSTONFS(mp);
1246
1247 if (nmp->nm_flag & NFSMNT_INT) {
1248 sigset_t oldset;
1249
1250 nfs_set_sigmask(td, &oldset);
1251 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1252 nfs_restore_sigmask(td, &oldset);
1253 while (bp == NULL) {
1254 if (nfs_sigintr(nmp, td))
1255 return (NULL);
1256 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1257 }
1258 } else {
1259 bp = getblk(vp, bn, size, 0, 0, 0);
1260 }
1261
1262 if (vp->v_type == VREG) {
1263 int biosize;
1264
1265 biosize = mp->mnt_stat.f_iosize;
1266 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1267 }
1268 return (bp);
1269 }
1270
1271 /*
1272 * Flush and invalidate all dirty buffers. If another process is already
1273 * doing the flush, just wait for completion.
1274 */
1275 int
1276 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1277 {
1278 struct nfsnode *np = VTONFS(vp);
1279 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1280 int error = 0, slpflag, slptimeo;
1281 int old_lock = 0;
1282
1283 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1284
1285 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1286 intrflg = 0;
1287 if (intrflg) {
1288 slpflag = NFS_PCATCH;
1289 slptimeo = 2 * hz;
1290 } else {
1291 slpflag = 0;
1292 slptimeo = 0;
1293 }
1294
1295 old_lock = nfs_upgrade_vnlock(vp);
1296 if (vp->v_iflag & VI_DOOMED) {
1297 /*
1298 * Since vgonel() uses the generic vinvalbuf() to flush
1299 * dirty buffers and it does not call this function, it
1300 * is safe to just return OK when VI_DOOMED is set.
1301 */
1302 nfs_downgrade_vnlock(vp, old_lock);
1303 return (0);
1304 }
1305
1306 /*
1307 * Now, flush as required.
1308 */
1309 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1310 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1311 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1312 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1313 /*
1314 * If the page clean was interrupted, fail the invalidation.
1315 * Not doing so, we run the risk of losing dirty pages in the
1316 * vinvalbuf() call below.
1317 */
1318 if (intrflg && (error = nfs_sigintr(nmp, td)))
1319 goto out;
1320 }
1321
1322 error = vinvalbuf(vp, flags, slpflag, 0);
1323 while (error) {
1324 if (intrflg && (error = nfs_sigintr(nmp, td)))
1325 goto out;
1326 error = vinvalbuf(vp, flags, 0, slptimeo);
1327 }
1328 mtx_lock(&np->n_mtx);
1329 if (np->n_directio_asyncwr == 0)
1330 np->n_flag &= ~NMODIFIED;
1331 mtx_unlock(&np->n_mtx);
1332 out:
1333 nfs_downgrade_vnlock(vp, old_lock);
1334 return error;
1335 }
1336
1337 /*
1338 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1339 * This is mainly to avoid queueing async I/O requests when the nfsiods
1340 * are all hung on a dead server.
1341 *
1342 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1343 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1344 */
1345 int
1346 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1347 {
1348 int iod;
1349 int gotiod;
1350 int slpflag = 0;
1351 int slptimeo = 0;
1352 int error, error2;
1353
1354 /*
1355 * Commits are usually short and sweet so lets save some cpu and
1356 * leave the async daemons for more important rpc's (such as reads
1357 * and writes).
1358 */
1359 mtx_lock(&nfs_iod_mtx);
1360 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1361 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1362 mtx_unlock(&nfs_iod_mtx);
1363 return(EIO);
1364 }
1365 again:
1366 if (nmp->nm_flag & NFSMNT_INT)
1367 slpflag = NFS_PCATCH;
1368 gotiod = FALSE;
1369
1370 /*
1371 * Find a free iod to process this request.
1372 */
1373 for (iod = 0; iod < nfs_numasync; iod++)
1374 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1375 gotiod = TRUE;
1376 break;
1377 }
1378
1379 /*
1380 * Try to create one if none are free.
1381 */
1382 if (!gotiod)
1383 nfs_nfsiodnew();
1384 else {
1385 /*
1386 * Found one, so wake it up and tell it which
1387 * mount to process.
1388 */
1389 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1390 iod, nmp));
1391 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1392 nfs_iodmount[iod] = nmp;
1393 nmp->nm_bufqiods++;
1394 wakeup(&nfs_iodwant[iod]);
1395 }
1396
1397 /*
1398 * If none are free, we may already have an iod working on this mount
1399 * point. If so, it will process our request.
1400 */
1401 if (!gotiod) {
1402 if (nmp->nm_bufqiods > 0) {
1403 NFS_DPF(ASYNCIO,
1404 ("nfs_asyncio: %d iods are already processing mount %p\n",
1405 nmp->nm_bufqiods, nmp));
1406 gotiod = TRUE;
1407 }
1408 }
1409
1410 /*
1411 * If we have an iod which can process the request, then queue
1412 * the buffer.
1413 */
1414 if (gotiod) {
1415 /*
1416 * Ensure that the queue never grows too large. We still want
1417 * to asynchronize so we block rather then return EIO.
1418 */
1419 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1420 NFS_DPF(ASYNCIO,
1421 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1422 nmp->nm_bufqwant = TRUE;
1423 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1424 slpflag | PRIBIO,
1425 "nfsaio", slptimeo);
1426 if (error) {
1427 error2 = nfs_sigintr(nmp, td);
1428 if (error2) {
1429 mtx_unlock(&nfs_iod_mtx);
1430 return (error2);
1431 }
1432 if (slpflag == NFS_PCATCH) {
1433 slpflag = 0;
1434 slptimeo = 2 * hz;
1435 }
1436 }
1437 /*
1438 * We might have lost our iod while sleeping,
1439 * so check and loop if nescessary.
1440 */
1441 goto again;
1442 }
1443
1444 /* We might have lost our nfsiod */
1445 if (nmp->nm_bufqiods == 0) {
1446 NFS_DPF(ASYNCIO,
1447 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1448 goto again;
1449 }
1450
1451 if (bp->b_iocmd == BIO_READ) {
1452 if (bp->b_rcred == NOCRED && cred != NOCRED)
1453 bp->b_rcred = crhold(cred);
1454 } else {
1455 if (bp->b_wcred == NOCRED && cred != NOCRED)
1456 bp->b_wcred = crhold(cred);
1457 }
1458
1459 if (bp->b_flags & B_REMFREE)
1460 bremfreef(bp);
1461 BUF_KERNPROC(bp);
1462 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1463 nmp->nm_bufqlen++;
1464 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1465 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1466 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1467 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1468 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1469 }
1470 mtx_unlock(&nfs_iod_mtx);
1471 return (0);
1472 }
1473
1474 mtx_unlock(&nfs_iod_mtx);
1475
1476 /*
1477 * All the iods are busy on other mounts, so return EIO to
1478 * force the caller to process the i/o synchronously.
1479 */
1480 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1481 return (EIO);
1482 }
1483
1484 void
1485 nfs_doio_directwrite(struct buf *bp)
1486 {
1487 int iomode, must_commit;
1488 struct uio *uiop = (struct uio *)bp->b_caller1;
1489 char *iov_base = uiop->uio_iov->iov_base;
1490 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1491
1492 iomode = NFSV3WRITE_FILESYNC;
1493 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1494 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1495 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1496 free(iov_base, M_NFSDIRECTIO);
1497 free(uiop->uio_iov, M_NFSDIRECTIO);
1498 free(uiop, M_NFSDIRECTIO);
1499 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1500 struct nfsnode *np = VTONFS(bp->b_vp);
1501 mtx_lock(&np->n_mtx);
1502 np->n_directio_asyncwr--;
1503 if (np->n_directio_asyncwr == 0) {
1504 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1505 if ((np->n_flag & NFSYNCWAIT)) {
1506 np->n_flag &= ~NFSYNCWAIT;
1507 wakeup((caddr_t)&np->n_directio_asyncwr);
1508 }
1509 }
1510 mtx_unlock(&np->n_mtx);
1511 }
1512 bp->b_vp = NULL;
1513 relpbuf(bp, &nfs_pbuf_freecnt);
1514 }
1515
1516 /*
1517 * Do an I/O operation to/from a cache block. This may be called
1518 * synchronously or from an nfsiod.
1519 */
1520 int
1521 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1522 {
1523 struct uio *uiop;
1524 struct nfsnode *np;
1525 struct nfsmount *nmp;
1526 int error = 0, iomode, must_commit = 0;
1527 struct uio uio;
1528 struct iovec io;
1529 struct proc *p = td ? td->td_proc : NULL;
1530 uint8_t iocmd;
1531
1532 np = VTONFS(vp);
1533 nmp = VFSTONFS(vp->v_mount);
1534 uiop = &uio;
1535 uiop->uio_iov = &io;
1536 uiop->uio_iovcnt = 1;
1537 uiop->uio_segflg = UIO_SYSSPACE;
1538 uiop->uio_td = td;
1539
1540 /*
1541 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1542 * do this here so we do not have to do it in all the code that
1543 * calls us.
1544 */
1545 bp->b_flags &= ~B_INVAL;
1546 bp->b_ioflags &= ~BIO_ERROR;
1547
1548 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1549 iocmd = bp->b_iocmd;
1550 if (iocmd == BIO_READ) {
1551 io.iov_len = uiop->uio_resid = bp->b_bcount;
1552 io.iov_base = bp->b_data;
1553 uiop->uio_rw = UIO_READ;
1554
1555 switch (vp->v_type) {
1556 case VREG:
1557 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1558 nfsstats.read_bios++;
1559 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1560
1561 if (!error) {
1562 if (uiop->uio_resid) {
1563 /*
1564 * If we had a short read with no error, we must have
1565 * hit a file hole. We should zero-fill the remainder.
1566 * This can also occur if the server hits the file EOF.
1567 *
1568 * Holes used to be able to occur due to pending
1569 * writes, but that is not possible any longer.
1570 */
1571 int nread = bp->b_bcount - uiop->uio_resid;
1572 int left = uiop->uio_resid;
1573
1574 if (left > 0)
1575 bzero((char *)bp->b_data + nread, left);
1576 uiop->uio_resid = 0;
1577 }
1578 }
1579 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1580 if (p && (vp->v_vflag & VV_TEXT)) {
1581 mtx_lock(&np->n_mtx);
1582 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1583 mtx_unlock(&np->n_mtx);
1584 PROC_LOCK(p);
1585 killproc(p, "text file modification");
1586 PROC_UNLOCK(p);
1587 } else
1588 mtx_unlock(&np->n_mtx);
1589 }
1590 break;
1591 case VLNK:
1592 uiop->uio_offset = (off_t)0;
1593 nfsstats.readlink_bios++;
1594 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1595 break;
1596 case VDIR:
1597 nfsstats.readdir_bios++;
1598 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1599 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1600 error = nfs_readdirplusrpc(vp, uiop, cr);
1601 if (error == NFSERR_NOTSUPP)
1602 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1603 }
1604 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1605 error = nfs_readdirrpc(vp, uiop, cr);
1606 /*
1607 * end-of-directory sets B_INVAL but does not generate an
1608 * error.
1609 */
1610 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1611 bp->b_flags |= B_INVAL;
1612 break;
1613 default:
1614 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1615 break;
1616 };
1617 if (error) {
1618 bp->b_ioflags |= BIO_ERROR;
1619 bp->b_error = error;
1620 }
1621 } else {
1622 /*
1623 * If we only need to commit, try to commit
1624 */
1625 if (bp->b_flags & B_NEEDCOMMIT) {
1626 int retv;
1627 off_t off;
1628
1629 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1630 retv = (nmp->nm_rpcops->nr_commit)(
1631 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1632 bp->b_wcred, td);
1633 if (retv == 0) {
1634 bp->b_dirtyoff = bp->b_dirtyend = 0;
1635 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1636 bp->b_resid = 0;
1637 bufdone(bp);
1638 return (0);
1639 }
1640 if (retv == NFSERR_STALEWRITEVERF) {
1641 nfs_clearcommit(vp->v_mount);
1642 }
1643 }
1644
1645 /*
1646 * Setup for actual write
1647 */
1648 mtx_lock(&np->n_mtx);
1649 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1650 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1651 mtx_unlock(&np->n_mtx);
1652
1653 if (bp->b_dirtyend > bp->b_dirtyoff) {
1654 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1655 - bp->b_dirtyoff;
1656 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1657 + bp->b_dirtyoff;
1658 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1659 uiop->uio_rw = UIO_WRITE;
1660 nfsstats.write_bios++;
1661
1662 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1663 iomode = NFSV3WRITE_UNSTABLE;
1664 else
1665 iomode = NFSV3WRITE_FILESYNC;
1666
1667 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1668
1669 /*
1670 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1671 * to cluster the buffers needing commit. This will allow
1672 * the system to submit a single commit rpc for the whole
1673 * cluster. We can do this even if the buffer is not 100%
1674 * dirty (relative to the NFS blocksize), so we optimize the
1675 * append-to-file-case.
1676 *
1677 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1678 * cleared because write clustering only works for commit
1679 * rpc's, not for the data portion of the write).
1680 */
1681
1682 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1683 bp->b_flags |= B_NEEDCOMMIT;
1684 if (bp->b_dirtyoff == 0
1685 && bp->b_dirtyend == bp->b_bcount)
1686 bp->b_flags |= B_CLUSTEROK;
1687 } else {
1688 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1689 }
1690
1691 /*
1692 * For an interrupted write, the buffer is still valid
1693 * and the write hasn't been pushed to the server yet,
1694 * so we can't set BIO_ERROR and report the interruption
1695 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1696 * is not relevant, so the rpc attempt is essentially
1697 * a noop. For the case of a V3 write rpc not being
1698 * committed to stable storage, the block is still
1699 * dirty and requires either a commit rpc or another
1700 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1701 * the block is reused. This is indicated by setting
1702 * the B_DELWRI and B_NEEDCOMMIT flags.
1703 *
1704 * If the buffer is marked B_PAGING, it does not reside on
1705 * the vp's paging queues so we cannot call bdirty(). The
1706 * bp in this case is not an NFS cache block so we should
1707 * be safe. XXX
1708 *
1709 * The logic below breaks up errors into recoverable and
1710 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1711 * and keep the buffer around for potential write retries.
1712 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1713 * and save the error in the nfsnode. This is less than ideal
1714 * but necessary. Keeping such buffers around could potentially
1715 * cause buffer exhaustion eventually (they can never be written
1716 * out, so will get constantly be re-dirtied). It also causes
1717 * all sorts of vfs panics. For non-recoverable write errors,
1718 * also invalidate the attrcache, so we'll be forced to go over
1719 * the wire for this object, returning an error to user on next
1720 * call (most of the time).
1721 */
1722 if (error == EINTR || error == EIO || error == ETIMEDOUT
1723 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1724 int s;
1725
1726 s = splbio();
1727 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1728 if ((bp->b_flags & B_PAGING) == 0) {
1729 bdirty(bp);
1730 bp->b_flags &= ~B_DONE;
1731 }
1732 if (error && (bp->b_flags & B_ASYNC) == 0)
1733 bp->b_flags |= B_EINTR;
1734 splx(s);
1735 } else {
1736 if (error) {
1737 bp->b_ioflags |= BIO_ERROR;
1738 bp->b_flags |= B_INVAL;
1739 bp->b_error = np->n_error = error;
1740 mtx_lock(&np->n_mtx);
1741 np->n_flag |= NWRITEERR;
1742 np->n_attrstamp = 0;
1743 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1744 mtx_unlock(&np->n_mtx);
1745 }
1746 bp->b_dirtyoff = bp->b_dirtyend = 0;
1747 }
1748 } else {
1749 bp->b_resid = 0;
1750 bufdone(bp);
1751 return (0);
1752 }
1753 }
1754 bp->b_resid = uiop->uio_resid;
1755 if (must_commit)
1756 nfs_clearcommit(vp->v_mount);
1757 bufdone(bp);
1758 return (error);
1759 }
1760
1761 /*
1762 * Used to aid in handling ftruncate() operations on the NFS client side.
1763 * Truncation creates a number of special problems for NFS. We have to
1764 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1765 * we have to properly handle VM pages or (potentially dirty) buffers
1766 * that straddle the truncation point.
1767 */
1768
1769 int
1770 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1771 {
1772 struct nfsnode *np = VTONFS(vp);
1773 u_quad_t tsize;
1774 int biosize = vp->v_mount->mnt_stat.f_iosize;
1775 int error = 0;
1776
1777 mtx_lock(&np->n_mtx);
1778 tsize = np->n_size;
1779 np->n_size = nsize;
1780 mtx_unlock(&np->n_mtx);
1781
1782 if (nsize < tsize) {
1783 struct buf *bp;
1784 daddr_t lbn;
1785 int bufsize;
1786
1787 /*
1788 * vtruncbuf() doesn't get the buffer overlapping the
1789 * truncation point. We may have a B_DELWRI and/or B_CACHE
1790 * buffer that now needs to be truncated.
1791 */
1792 error = vtruncbuf(vp, cred, td, nsize, biosize);
1793 lbn = nsize / biosize;
1794 bufsize = nsize & (biosize - 1);
1795 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1796 if (!bp)
1797 return EINTR;
1798 if (bp->b_dirtyoff > bp->b_bcount)
1799 bp->b_dirtyoff = bp->b_bcount;
1800 if (bp->b_dirtyend > bp->b_bcount)
1801 bp->b_dirtyend = bp->b_bcount;
1802 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1803 brelse(bp);
1804 } else {
1805 vnode_pager_setsize(vp, nsize);
1806 }
1807 return(error);
1808 }
1809
Cache object: cf1cfd2ea8497443b2c63ecf3c4bd97f
|