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/10.0/sys/nfsclient/nfs_bio.c 249623 2013-04-18 13:09:04Z 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/rwlock.h>
49 #include <sys/vmmeter.h>
50 #include <sys/vnode.h>
51
52 #include <vm/vm.h>
53 #include <vm/vm_param.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 <nfs/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_WLOCK(object);
133 if (pages[ap->a_reqpage]->valid != 0) {
134 for (i = 0; i < npages; ++i) {
135 if (i != ap->a_reqpage) {
136 vm_page_lock(pages[i]);
137 vm_page_free(pages[i]);
138 vm_page_unlock(pages[i]);
139 }
140 }
141 VM_OBJECT_WUNLOCK(object);
142 return (0);
143 }
144 VM_OBJECT_WUNLOCK(object);
145
146 /*
147 * We use only the kva address for the buffer, but this is extremely
148 * convienient and fast.
149 */
150 bp = getpbuf(&nfs_pbuf_freecnt);
151
152 kva = (vm_offset_t) bp->b_data;
153 pmap_qenter(kva, pages, npages);
154 PCPU_INC(cnt.v_vnodein);
155 PCPU_ADD(cnt.v_vnodepgsin, npages);
156
157 iov.iov_base = (caddr_t) kva;
158 iov.iov_len = count;
159 uio.uio_iov = &iov;
160 uio.uio_iovcnt = 1;
161 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
162 uio.uio_resid = count;
163 uio.uio_segflg = UIO_SYSSPACE;
164 uio.uio_rw = UIO_READ;
165 uio.uio_td = td;
166
167 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
168 pmap_qremove(kva, npages);
169
170 relpbuf(bp, &nfs_pbuf_freecnt);
171
172 if (error && (uio.uio_resid == count)) {
173 nfs_printf("nfs_getpages: error %d\n", error);
174 VM_OBJECT_WLOCK(object);
175 for (i = 0; i < npages; ++i) {
176 if (i != ap->a_reqpage) {
177 vm_page_lock(pages[i]);
178 vm_page_free(pages[i]);
179 vm_page_unlock(pages[i]);
180 }
181 }
182 VM_OBJECT_WUNLOCK(object);
183 return (VM_PAGER_ERROR);
184 }
185
186 /*
187 * Calculate the number of bytes read and validate only that number
188 * of bytes. Note that due to pending writes, size may be 0. This
189 * does not mean that the remaining data is invalid!
190 */
191
192 size = count - uio.uio_resid;
193 VM_OBJECT_WLOCK(object);
194 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
195 vm_page_t m;
196 nextoff = toff + PAGE_SIZE;
197 m = pages[i];
198
199 if (nextoff <= size) {
200 /*
201 * Read operation filled an entire page
202 */
203 m->valid = VM_PAGE_BITS_ALL;
204 KASSERT(m->dirty == 0,
205 ("nfs_getpages: page %p is dirty", m));
206 } else if (size > toff) {
207 /*
208 * Read operation filled a partial page.
209 */
210 m->valid = 0;
211 vm_page_set_valid_range(m, 0, size - toff);
212 KASSERT(m->dirty == 0,
213 ("nfs_getpages: page %p is dirty", m));
214 } else {
215 /*
216 * Read operation was short. If no error
217 * occured we may have hit a zero-fill
218 * section. We leave valid set to 0, and page
219 * is freed by vm_page_readahead_finish() if
220 * its index is not equal to requested, or
221 * page is zeroed and set valid by
222 * vm_pager_get_pages() for requested page.
223 */
224 ;
225 }
226 if (i != ap->a_reqpage)
227 vm_page_readahead_finish(m);
228 }
229 VM_OBJECT_WUNLOCK(object);
230 return (0);
231 }
232
233 /*
234 * Vnode op for VM putpages.
235 */
236 int
237 nfs_putpages(struct vop_putpages_args *ap)
238 {
239 struct uio uio;
240 struct iovec iov;
241 vm_offset_t kva;
242 struct buf *bp;
243 int iomode, must_commit, i, error, npages, count;
244 off_t offset;
245 int *rtvals;
246 struct vnode *vp;
247 struct thread *td;
248 struct ucred *cred;
249 struct nfsmount *nmp;
250 struct nfsnode *np;
251 vm_page_t *pages;
252
253 vp = ap->a_vp;
254 np = VTONFS(vp);
255 td = curthread; /* XXX */
256 /* Set the cred to n_writecred for the write rpcs. */
257 if (np->n_writecred != NULL)
258 cred = crhold(np->n_writecred);
259 else
260 cred = crhold(curthread->td_ucred); /* XXX */
261 nmp = VFSTONFS(vp->v_mount);
262 pages = ap->a_m;
263 count = ap->a_count;
264 rtvals = ap->a_rtvals;
265 npages = btoc(count);
266 offset = IDX_TO_OFF(pages[0]->pindex);
267
268 mtx_lock(&nmp->nm_mtx);
269 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
270 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
271 mtx_unlock(&nmp->nm_mtx);
272 (void)nfs_fsinfo(nmp, vp, cred, td);
273 } else
274 mtx_unlock(&nmp->nm_mtx);
275
276 mtx_lock(&np->n_mtx);
277 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
278 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
279 mtx_unlock(&np->n_mtx);
280 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
281 mtx_lock(&np->n_mtx);
282 }
283
284 for (i = 0; i < npages; i++)
285 rtvals[i] = VM_PAGER_ERROR;
286
287 /*
288 * When putting pages, do not extend file past EOF.
289 */
290 if (offset + count > np->n_size) {
291 count = np->n_size - offset;
292 if (count < 0)
293 count = 0;
294 }
295 mtx_unlock(&np->n_mtx);
296
297 /*
298 * We use only the kva address for the buffer, but this is extremely
299 * convienient and fast.
300 */
301 bp = getpbuf(&nfs_pbuf_freecnt);
302
303 kva = (vm_offset_t) bp->b_data;
304 pmap_qenter(kva, pages, npages);
305 PCPU_INC(cnt.v_vnodeout);
306 PCPU_ADD(cnt.v_vnodepgsout, count);
307
308 iov.iov_base = (caddr_t) kva;
309 iov.iov_len = count;
310 uio.uio_iov = &iov;
311 uio.uio_iovcnt = 1;
312 uio.uio_offset = offset;
313 uio.uio_resid = count;
314 uio.uio_segflg = UIO_SYSSPACE;
315 uio.uio_rw = UIO_WRITE;
316 uio.uio_td = td;
317
318 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
319 iomode = NFSV3WRITE_UNSTABLE;
320 else
321 iomode = NFSV3WRITE_FILESYNC;
322
323 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
324 crfree(cred);
325
326 pmap_qremove(kva, npages);
327 relpbuf(bp, &nfs_pbuf_freecnt);
328
329 if (!error) {
330 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
331 if (must_commit) {
332 nfs_clearcommit(vp->v_mount);
333 }
334 }
335 return rtvals[0];
336 }
337
338 /*
339 * For nfs, cache consistency can only be maintained approximately.
340 * Although RFC1094 does not specify the criteria, the following is
341 * believed to be compatible with the reference port.
342 * For nfs:
343 * If the file's modify time on the server has changed since the
344 * last read rpc or you have written to the file,
345 * you may have lost data cache consistency with the
346 * server, so flush all of the file's data out of the cache.
347 * Then force a getattr rpc to ensure that you have up to date
348 * attributes.
349 * NB: This implies that cache data can be read when up to
350 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
351 * attributes this could be forced by setting n_attrstamp to 0 before
352 * the VOP_GETATTR() call.
353 */
354 static inline int
355 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
356 {
357 int error = 0;
358 struct vattr vattr;
359 struct nfsnode *np = VTONFS(vp);
360 int old_lock;
361 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
362
363 /*
364 * Grab the exclusive lock before checking whether the cache is
365 * consistent.
366 * XXX - We can make this cheaper later (by acquiring cheaper locks).
367 * But for now, this suffices.
368 */
369 old_lock = nfs_upgrade_vnlock(vp);
370 if (vp->v_iflag & VI_DOOMED) {
371 nfs_downgrade_vnlock(vp, old_lock);
372 return (EBADF);
373 }
374
375 mtx_lock(&np->n_mtx);
376 if (np->n_flag & NMODIFIED) {
377 mtx_unlock(&np->n_mtx);
378 if (vp->v_type != VREG) {
379 if (vp->v_type != VDIR)
380 panic("nfs: bioread, not dir");
381 (nmp->nm_rpcops->nr_invaldir)(vp);
382 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
383 if (error)
384 goto out;
385 }
386 np->n_attrstamp = 0;
387 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
388 error = VOP_GETATTR(vp, &vattr, cred);
389 if (error)
390 goto out;
391 mtx_lock(&np->n_mtx);
392 np->n_mtime = vattr.va_mtime;
393 mtx_unlock(&np->n_mtx);
394 } else {
395 mtx_unlock(&np->n_mtx);
396 error = VOP_GETATTR(vp, &vattr, cred);
397 if (error)
398 return (error);
399 mtx_lock(&np->n_mtx);
400 if ((np->n_flag & NSIZECHANGED)
401 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
402 mtx_unlock(&np->n_mtx);
403 if (vp->v_type == VDIR)
404 (nmp->nm_rpcops->nr_invaldir)(vp);
405 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
406 if (error)
407 goto out;
408 mtx_lock(&np->n_mtx);
409 np->n_mtime = vattr.va_mtime;
410 np->n_flag &= ~NSIZECHANGED;
411 }
412 mtx_unlock(&np->n_mtx);
413 }
414 out:
415 nfs_downgrade_vnlock(vp, old_lock);
416 return error;
417 }
418
419 /*
420 * Vnode op for read using bio
421 */
422 int
423 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
424 {
425 struct nfsnode *np = VTONFS(vp);
426 int biosize, i;
427 struct buf *bp, *rabp;
428 struct thread *td;
429 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
430 daddr_t lbn, rabn;
431 off_t end;
432 int bcount;
433 int seqcount;
434 int nra, error = 0, n = 0, on = 0;
435
436 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
437 if (uio->uio_resid == 0)
438 return (0);
439 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
440 return (EINVAL);
441 td = uio->uio_td;
442
443 mtx_lock(&nmp->nm_mtx);
444 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
445 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
446 mtx_unlock(&nmp->nm_mtx);
447 (void)nfs_fsinfo(nmp, vp, cred, td);
448 } else
449 mtx_unlock(&nmp->nm_mtx);
450
451 end = uio->uio_offset + uio->uio_resid;
452 if (vp->v_type != VDIR &&
453 (end > nmp->nm_maxfilesize || end < uio->uio_offset))
454 return (EFBIG);
455
456 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
457 /* No caching/ no readaheads. Just read data into the user buffer */
458 return nfs_readrpc(vp, uio, cred);
459
460 biosize = vp->v_bufobj.bo_bsize;
461 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
462
463 error = nfs_bioread_check_cons(vp, td, cred);
464 if (error)
465 return error;
466
467 do {
468 u_quad_t nsize;
469
470 mtx_lock(&np->n_mtx);
471 nsize = np->n_size;
472 mtx_unlock(&np->n_mtx);
473
474 switch (vp->v_type) {
475 case VREG:
476 nfsstats.biocache_reads++;
477 lbn = uio->uio_offset / biosize;
478 on = uio->uio_offset - (lbn * biosize);
479
480 /*
481 * Start the read ahead(s), as required.
482 */
483 if (nmp->nm_readahead > 0) {
484 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
485 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
486 rabn = lbn + 1 + nra;
487 if (incore(&vp->v_bufobj, rabn) == NULL) {
488 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
489 if (!rabp) {
490 error = nfs_sigintr(nmp, td);
491 return (error ? error : EINTR);
492 }
493 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
494 rabp->b_flags |= B_ASYNC;
495 rabp->b_iocmd = BIO_READ;
496 vfs_busy_pages(rabp, 0);
497 if (nfs_asyncio(nmp, rabp, cred, td)) {
498 rabp->b_flags |= B_INVAL;
499 rabp->b_ioflags |= BIO_ERROR;
500 vfs_unbusy_pages(rabp);
501 brelse(rabp);
502 break;
503 }
504 } else {
505 brelse(rabp);
506 }
507 }
508 }
509 }
510
511 /* Note that bcount is *not* DEV_BSIZE aligned. */
512 bcount = biosize;
513 if ((off_t)lbn * biosize >= nsize) {
514 bcount = 0;
515 } else if ((off_t)(lbn + 1) * biosize > nsize) {
516 bcount = nsize - (off_t)lbn * biosize;
517 }
518 bp = nfs_getcacheblk(vp, lbn, bcount, td);
519
520 if (!bp) {
521 error = nfs_sigintr(nmp, td);
522 return (error ? error : EINTR);
523 }
524
525 /*
526 * If B_CACHE is not set, we must issue the read. If this
527 * fails, we return an error.
528 */
529
530 if ((bp->b_flags & B_CACHE) == 0) {
531 bp->b_iocmd = BIO_READ;
532 vfs_busy_pages(bp, 0);
533 error = nfs_doio(vp, bp, cred, td);
534 if (error) {
535 brelse(bp);
536 return (error);
537 }
538 }
539
540 /*
541 * on is the offset into the current bp. Figure out how many
542 * bytes we can copy out of the bp. Note that bcount is
543 * NOT DEV_BSIZE aligned.
544 *
545 * Then figure out how many bytes we can copy into the uio.
546 */
547
548 n = 0;
549 if (on < bcount)
550 n = MIN((unsigned)(bcount - on), uio->uio_resid);
551 break;
552 case VLNK:
553 nfsstats.biocache_readlinks++;
554 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
555 if (!bp) {
556 error = nfs_sigintr(nmp, td);
557 return (error ? error : EINTR);
558 }
559 if ((bp->b_flags & B_CACHE) == 0) {
560 bp->b_iocmd = BIO_READ;
561 vfs_busy_pages(bp, 0);
562 error = nfs_doio(vp, bp, cred, td);
563 if (error) {
564 bp->b_ioflags |= BIO_ERROR;
565 brelse(bp);
566 return (error);
567 }
568 }
569 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
570 on = 0;
571 break;
572 case VDIR:
573 nfsstats.biocache_readdirs++;
574 if (np->n_direofoffset
575 && uio->uio_offset >= np->n_direofoffset) {
576 return (0);
577 }
578 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
579 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
580 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
581 if (!bp) {
582 error = nfs_sigintr(nmp, td);
583 return (error ? error : EINTR);
584 }
585 if ((bp->b_flags & B_CACHE) == 0) {
586 bp->b_iocmd = BIO_READ;
587 vfs_busy_pages(bp, 0);
588 error = nfs_doio(vp, bp, cred, td);
589 if (error) {
590 brelse(bp);
591 }
592 while (error == NFSERR_BAD_COOKIE) {
593 (nmp->nm_rpcops->nr_invaldir)(vp);
594 error = nfs_vinvalbuf(vp, 0, td, 1);
595 /*
596 * Yuck! The directory has been modified on the
597 * server. The only way to get the block is by
598 * reading from the beginning to get all the
599 * offset cookies.
600 *
601 * Leave the last bp intact unless there is an error.
602 * Loop back up to the while if the error is another
603 * NFSERR_BAD_COOKIE (double yuch!).
604 */
605 for (i = 0; i <= lbn && !error; i++) {
606 if (np->n_direofoffset
607 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
608 return (0);
609 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
610 if (!bp) {
611 error = nfs_sigintr(nmp, td);
612 return (error ? error : EINTR);
613 }
614 if ((bp->b_flags & B_CACHE) == 0) {
615 bp->b_iocmd = BIO_READ;
616 vfs_busy_pages(bp, 0);
617 error = nfs_doio(vp, bp, cred, td);
618 /*
619 * no error + B_INVAL == directory EOF,
620 * use the block.
621 */
622 if (error == 0 && (bp->b_flags & B_INVAL))
623 break;
624 }
625 /*
626 * An error will throw away the block and the
627 * for loop will break out. If no error and this
628 * is not the block we want, we throw away the
629 * block and go for the next one via the for loop.
630 */
631 if (error || i < lbn)
632 brelse(bp);
633 }
634 }
635 /*
636 * The above while is repeated if we hit another cookie
637 * error. If we hit an error and it wasn't a cookie error,
638 * we give up.
639 */
640 if (error)
641 return (error);
642 }
643
644 /*
645 * If not eof and read aheads are enabled, start one.
646 * (You need the current block first, so that you have the
647 * directory offset cookie of the next block.)
648 */
649 if (nmp->nm_readahead > 0 &&
650 (bp->b_flags & B_INVAL) == 0 &&
651 (np->n_direofoffset == 0 ||
652 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
653 incore(&vp->v_bufobj, lbn + 1) == NULL) {
654 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
655 if (rabp) {
656 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
657 rabp->b_flags |= B_ASYNC;
658 rabp->b_iocmd = BIO_READ;
659 vfs_busy_pages(rabp, 0);
660 if (nfs_asyncio(nmp, rabp, cred, td)) {
661 rabp->b_flags |= B_INVAL;
662 rabp->b_ioflags |= BIO_ERROR;
663 vfs_unbusy_pages(rabp);
664 brelse(rabp);
665 }
666 } else {
667 brelse(rabp);
668 }
669 }
670 }
671 /*
672 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
673 * chopped for the EOF condition, we cannot tell how large
674 * NFS directories are going to be until we hit EOF. So
675 * an NFS directory buffer is *not* chopped to its EOF. Now,
676 * it just so happens that b_resid will effectively chop it
677 * to EOF. *BUT* this information is lost if the buffer goes
678 * away and is reconstituted into a B_CACHE state ( due to
679 * being VMIO ) later. So we keep track of the directory eof
680 * in np->n_direofoffset and chop it off as an extra step
681 * right here.
682 */
683 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
684 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
685 n = np->n_direofoffset - uio->uio_offset;
686 break;
687 default:
688 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
689 bp = NULL;
690 break;
691 };
692
693 if (n > 0) {
694 error = uiomove(bp->b_data + on, (int)n, uio);
695 }
696 if (vp->v_type == VLNK)
697 n = 0;
698 if (bp != NULL)
699 brelse(bp);
700 } while (error == 0 && uio->uio_resid > 0 && n > 0);
701 return (error);
702 }
703
704 /*
705 * The NFS write path cannot handle iovecs with len > 1. So we need to
706 * break up iovecs accordingly (restricting them to wsize).
707 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
708 * For the ASYNC case, 2 copies are needed. The first a copy from the
709 * user buffer to a staging buffer and then a second copy from the staging
710 * buffer to mbufs. This can be optimized by copying from the user buffer
711 * directly into mbufs and passing the chain down, but that requires a
712 * fair amount of re-working of the relevant codepaths (and can be done
713 * later).
714 */
715 static int
716 nfs_directio_write(vp, uiop, cred, ioflag)
717 struct vnode *vp;
718 struct uio *uiop;
719 struct ucred *cred;
720 int ioflag;
721 {
722 int error;
723 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
724 struct thread *td = uiop->uio_td;
725 int size;
726 int wsize;
727
728 mtx_lock(&nmp->nm_mtx);
729 wsize = nmp->nm_wsize;
730 mtx_unlock(&nmp->nm_mtx);
731 if (ioflag & IO_SYNC) {
732 int iomode, must_commit;
733 struct uio uio;
734 struct iovec iov;
735 do_sync:
736 while (uiop->uio_resid > 0) {
737 size = MIN(uiop->uio_resid, wsize);
738 size = MIN(uiop->uio_iov->iov_len, size);
739 iov.iov_base = uiop->uio_iov->iov_base;
740 iov.iov_len = size;
741 uio.uio_iov = &iov;
742 uio.uio_iovcnt = 1;
743 uio.uio_offset = uiop->uio_offset;
744 uio.uio_resid = size;
745 uio.uio_segflg = UIO_USERSPACE;
746 uio.uio_rw = UIO_WRITE;
747 uio.uio_td = td;
748 iomode = NFSV3WRITE_FILESYNC;
749 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
750 &iomode, &must_commit);
751 KASSERT((must_commit == 0),
752 ("nfs_directio_write: Did not commit write"));
753 if (error)
754 return (error);
755 uiop->uio_offset += size;
756 uiop->uio_resid -= size;
757 if (uiop->uio_iov->iov_len <= size) {
758 uiop->uio_iovcnt--;
759 uiop->uio_iov++;
760 } else {
761 uiop->uio_iov->iov_base =
762 (char *)uiop->uio_iov->iov_base + size;
763 uiop->uio_iov->iov_len -= size;
764 }
765 }
766 } else {
767 struct uio *t_uio;
768 struct iovec *t_iov;
769 struct buf *bp;
770
771 /*
772 * Break up the write into blocksize chunks and hand these
773 * over to nfsiod's for write back.
774 * Unfortunately, this incurs a copy of the data. Since
775 * the user could modify the buffer before the write is
776 * initiated.
777 *
778 * The obvious optimization here is that one of the 2 copies
779 * in the async write path can be eliminated by copying the
780 * data here directly into mbufs and passing the mbuf chain
781 * down. But that will require a fair amount of re-working
782 * of the code and can be done if there's enough interest
783 * in NFS directio access.
784 */
785 while (uiop->uio_resid > 0) {
786 size = MIN(uiop->uio_resid, wsize);
787 size = MIN(uiop->uio_iov->iov_len, size);
788 bp = getpbuf(&nfs_pbuf_freecnt);
789 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
790 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
791 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
792 t_iov->iov_len = size;
793 t_uio->uio_iov = t_iov;
794 t_uio->uio_iovcnt = 1;
795 t_uio->uio_offset = uiop->uio_offset;
796 t_uio->uio_resid = size;
797 t_uio->uio_segflg = UIO_SYSSPACE;
798 t_uio->uio_rw = UIO_WRITE;
799 t_uio->uio_td = td;
800 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
801 uiop->uio_segflg == UIO_SYSSPACE,
802 ("nfs_directio_write: Bad uio_segflg"));
803 if (uiop->uio_segflg == UIO_USERSPACE) {
804 error = copyin(uiop->uio_iov->iov_base,
805 t_iov->iov_base, size);
806 if (error != 0)
807 goto err_free;
808 } else
809 /*
810 * UIO_SYSSPACE may never happen, but handle
811 * it just in case it does.
812 */
813 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
814 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 err_free:
826 if (error) {
827 free(t_iov->iov_base, M_NFSDIRECTIO);
828 free(t_iov, M_NFSDIRECTIO);
829 free(t_uio, M_NFSDIRECTIO);
830 bp->b_vp = NULL;
831 relpbuf(bp, &nfs_pbuf_freecnt);
832 if (error == EINTR)
833 return (error);
834 goto do_sync;
835 }
836 uiop->uio_offset += size;
837 uiop->uio_resid -= size;
838 if (uiop->uio_iov->iov_len <= size) {
839 uiop->uio_iovcnt--;
840 uiop->uio_iov++;
841 } else {
842 uiop->uio_iov->iov_base =
843 (char *)uiop->uio_iov->iov_base + size;
844 uiop->uio_iov->iov_len -= size;
845 }
846 }
847 }
848 return (0);
849 }
850
851 /*
852 * Vnode op for write using bio
853 */
854 int
855 nfs_write(struct vop_write_args *ap)
856 {
857 int biosize;
858 struct uio *uio = ap->a_uio;
859 struct thread *td = uio->uio_td;
860 struct vnode *vp = ap->a_vp;
861 struct nfsnode *np = VTONFS(vp);
862 struct ucred *cred = ap->a_cred;
863 int ioflag = ap->a_ioflag;
864 struct buf *bp;
865 struct vattr vattr;
866 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
867 daddr_t lbn;
868 off_t end;
869 int bcount;
870 int n, on, error = 0;
871
872 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode"));
873 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
874 ("nfs_write proc"));
875 if (vp->v_type != VREG)
876 return (EIO);
877 mtx_lock(&np->n_mtx);
878 if (np->n_flag & NWRITEERR) {
879 np->n_flag &= ~NWRITEERR;
880 mtx_unlock(&np->n_mtx);
881 return (np->n_error);
882 } else
883 mtx_unlock(&np->n_mtx);
884 mtx_lock(&nmp->nm_mtx);
885 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
886 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
887 mtx_unlock(&nmp->nm_mtx);
888 (void)nfs_fsinfo(nmp, vp, cred, td);
889 } else
890 mtx_unlock(&nmp->nm_mtx);
891
892 /*
893 * Synchronously flush pending buffers if we are in synchronous
894 * mode or if we are appending.
895 */
896 if (ioflag & (IO_APPEND | IO_SYNC)) {
897 mtx_lock(&np->n_mtx);
898 if (np->n_flag & NMODIFIED) {
899 mtx_unlock(&np->n_mtx);
900 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
901 /*
902 * Require non-blocking, synchronous writes to
903 * dirty files to inform the program it needs
904 * to fsync(2) explicitly.
905 */
906 if (ioflag & IO_NDELAY)
907 return (EAGAIN);
908 #endif
909 flush_and_restart:
910 np->n_attrstamp = 0;
911 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
912 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
913 if (error)
914 return (error);
915 } else
916 mtx_unlock(&np->n_mtx);
917 }
918
919 /*
920 * If IO_APPEND then load uio_offset. We restart here if we cannot
921 * get the append lock.
922 */
923 if (ioflag & IO_APPEND) {
924 np->n_attrstamp = 0;
925 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
926 error = VOP_GETATTR(vp, &vattr, cred);
927 if (error)
928 return (error);
929 mtx_lock(&np->n_mtx);
930 uio->uio_offset = np->n_size;
931 mtx_unlock(&np->n_mtx);
932 }
933
934 if (uio->uio_offset < 0)
935 return (EINVAL);
936 end = uio->uio_offset + uio->uio_resid;
937 if (end > nmp->nm_maxfilesize || end < uio->uio_offset)
938 return (EFBIG);
939 if (uio->uio_resid == 0)
940 return (0);
941
942 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
943 return nfs_directio_write(vp, uio, cred, ioflag);
944
945 /*
946 * Maybe this should be above the vnode op call, but so long as
947 * file servers have no limits, i don't think it matters
948 */
949 if (vn_rlimit_fsize(vp, uio, td))
950 return (EFBIG);
951
952 biosize = vp->v_bufobj.bo_bsize;
953 /*
954 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
955 * would exceed the local maximum per-file write commit size when
956 * combined with those, we must decide whether to flush,
957 * go synchronous, or return error. We don't bother checking
958 * IO_UNIT -- we just make all writes atomic anyway, as there's
959 * no point optimizing for something that really won't ever happen.
960 */
961 if (!(ioflag & IO_SYNC)) {
962 int nflag;
963
964 mtx_lock(&np->n_mtx);
965 nflag = np->n_flag;
966 mtx_unlock(&np->n_mtx);
967 int needrestart = 0;
968 if (nmp->nm_wcommitsize < uio->uio_resid) {
969 /*
970 * If this request could not possibly be completed
971 * without exceeding the maximum outstanding write
972 * commit size, see if we can convert it into a
973 * synchronous write operation.
974 */
975 if (ioflag & IO_NDELAY)
976 return (EAGAIN);
977 ioflag |= IO_SYNC;
978 if (nflag & NMODIFIED)
979 needrestart = 1;
980 } else if (nflag & NMODIFIED) {
981 int wouldcommit = 0;
982 BO_LOCK(&vp->v_bufobj);
983 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
984 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
985 b_bobufs) {
986 if (bp->b_flags & B_NEEDCOMMIT)
987 wouldcommit += bp->b_bcount;
988 }
989 }
990 BO_UNLOCK(&vp->v_bufobj);
991 /*
992 * Since we're not operating synchronously and
993 * bypassing the buffer cache, we are in a commit
994 * and holding all of these buffers whether
995 * transmitted or not. If not limited, this
996 * will lead to the buffer cache deadlocking,
997 * as no one else can flush our uncommitted buffers.
998 */
999 wouldcommit += uio->uio_resid;
1000 /*
1001 * If we would initially exceed the maximum
1002 * outstanding write commit size, flush and restart.
1003 */
1004 if (wouldcommit > nmp->nm_wcommitsize)
1005 needrestart = 1;
1006 }
1007 if (needrestart)
1008 goto flush_and_restart;
1009 }
1010
1011 do {
1012 nfsstats.biocache_writes++;
1013 lbn = uio->uio_offset / biosize;
1014 on = uio->uio_offset - (lbn * biosize);
1015 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1016 again:
1017 /*
1018 * Handle direct append and file extension cases, calculate
1019 * unaligned buffer size.
1020 */
1021 mtx_lock(&np->n_mtx);
1022 if (uio->uio_offset == np->n_size && n) {
1023 mtx_unlock(&np->n_mtx);
1024 /*
1025 * Get the buffer (in its pre-append state to maintain
1026 * B_CACHE if it was previously set). Resize the
1027 * nfsnode after we have locked the buffer to prevent
1028 * readers from reading garbage.
1029 */
1030 bcount = on;
1031 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1032
1033 if (bp != NULL) {
1034 long save;
1035
1036 mtx_lock(&np->n_mtx);
1037 np->n_size = uio->uio_offset + n;
1038 np->n_flag |= NMODIFIED;
1039 vnode_pager_setsize(vp, np->n_size);
1040 mtx_unlock(&np->n_mtx);
1041
1042 save = bp->b_flags & B_CACHE;
1043 bcount += n;
1044 allocbuf(bp, bcount);
1045 bp->b_flags |= save;
1046 }
1047 } else {
1048 /*
1049 * Obtain the locked cache block first, and then
1050 * adjust the file's size as appropriate.
1051 */
1052 bcount = on + n;
1053 if ((off_t)lbn * biosize + bcount < np->n_size) {
1054 if ((off_t)(lbn + 1) * biosize < np->n_size)
1055 bcount = biosize;
1056 else
1057 bcount = np->n_size - (off_t)lbn * biosize;
1058 }
1059 mtx_unlock(&np->n_mtx);
1060 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1061 mtx_lock(&np->n_mtx);
1062 if (uio->uio_offset + n > np->n_size) {
1063 np->n_size = uio->uio_offset + n;
1064 np->n_flag |= NMODIFIED;
1065 vnode_pager_setsize(vp, np->n_size);
1066 }
1067 mtx_unlock(&np->n_mtx);
1068 }
1069
1070 if (!bp) {
1071 error = nfs_sigintr(nmp, td);
1072 if (!error)
1073 error = EINTR;
1074 break;
1075 }
1076
1077 /*
1078 * Issue a READ if B_CACHE is not set. In special-append
1079 * mode, B_CACHE is based on the buffer prior to the write
1080 * op and is typically set, avoiding the read. If a read
1081 * is required in special append mode, the server will
1082 * probably send us a short-read since we extended the file
1083 * on our end, resulting in b_resid == 0 and, thusly,
1084 * B_CACHE getting set.
1085 *
1086 * We can also avoid issuing the read if the write covers
1087 * the entire buffer. We have to make sure the buffer state
1088 * is reasonable in this case since we will not be initiating
1089 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1090 * more information.
1091 *
1092 * B_CACHE may also be set due to the buffer being cached
1093 * normally.
1094 */
1095
1096 if (on == 0 && n == bcount) {
1097 bp->b_flags |= B_CACHE;
1098 bp->b_flags &= ~B_INVAL;
1099 bp->b_ioflags &= ~BIO_ERROR;
1100 }
1101
1102 if ((bp->b_flags & B_CACHE) == 0) {
1103 bp->b_iocmd = BIO_READ;
1104 vfs_busy_pages(bp, 0);
1105 error = nfs_doio(vp, bp, cred, td);
1106 if (error) {
1107 brelse(bp);
1108 break;
1109 }
1110 }
1111 if (bp->b_wcred == NOCRED)
1112 bp->b_wcred = crhold(cred);
1113 mtx_lock(&np->n_mtx);
1114 np->n_flag |= NMODIFIED;
1115 mtx_unlock(&np->n_mtx);
1116
1117 /*
1118 * If dirtyend exceeds file size, chop it down. This should
1119 * not normally occur but there is an append race where it
1120 * might occur XXX, so we log it.
1121 *
1122 * If the chopping creates a reverse-indexed or degenerate
1123 * situation with dirtyoff/end, we 0 both of them.
1124 */
1125
1126 if (bp->b_dirtyend > bcount) {
1127 nfs_printf("NFS append race @%lx:%d\n",
1128 (long)bp->b_blkno * DEV_BSIZE,
1129 bp->b_dirtyend - bcount);
1130 bp->b_dirtyend = bcount;
1131 }
1132
1133 if (bp->b_dirtyoff >= bp->b_dirtyend)
1134 bp->b_dirtyoff = bp->b_dirtyend = 0;
1135
1136 /*
1137 * If the new write will leave a contiguous dirty
1138 * area, just update the b_dirtyoff and b_dirtyend,
1139 * otherwise force a write rpc of the old dirty area.
1140 *
1141 * While it is possible to merge discontiguous writes due to
1142 * our having a B_CACHE buffer ( and thus valid read data
1143 * for the hole), we don't because it could lead to
1144 * significant cache coherency problems with multiple clients,
1145 * especially if locking is implemented later on.
1146 *
1147 * as an optimization we could theoretically maintain
1148 * a linked list of discontinuous areas, but we would still
1149 * have to commit them separately so there isn't much
1150 * advantage to it except perhaps a bit of asynchronization.
1151 */
1152
1153 if (bp->b_dirtyend > 0 &&
1154 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1155 if (bwrite(bp) == EINTR) {
1156 error = EINTR;
1157 break;
1158 }
1159 goto again;
1160 }
1161
1162 error = uiomove((char *)bp->b_data + on, n, uio);
1163
1164 /*
1165 * Since this block is being modified, it must be written
1166 * again and not just committed. Since write clustering does
1167 * not work for the stage 1 data write, only the stage 2
1168 * commit rpc, we have to clear B_CLUSTEROK as well.
1169 */
1170 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1171
1172 if (error) {
1173 bp->b_ioflags |= BIO_ERROR;
1174 brelse(bp);
1175 break;
1176 }
1177
1178 /*
1179 * Only update dirtyoff/dirtyend if not a degenerate
1180 * condition.
1181 */
1182 if (n) {
1183 if (bp->b_dirtyend > 0) {
1184 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1185 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1186 } else {
1187 bp->b_dirtyoff = on;
1188 bp->b_dirtyend = on + n;
1189 }
1190 vfs_bio_set_valid(bp, on, n);
1191 }
1192
1193 /*
1194 * If IO_SYNC do bwrite().
1195 *
1196 * IO_INVAL appears to be unused. The idea appears to be
1197 * to turn off caching in this case. Very odd. XXX
1198 */
1199 if ((ioflag & IO_SYNC)) {
1200 if (ioflag & IO_INVAL)
1201 bp->b_flags |= B_NOCACHE;
1202 error = bwrite(bp);
1203 if (error)
1204 break;
1205 } else if ((n + on) == biosize) {
1206 bp->b_flags |= B_ASYNC;
1207 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1208 } else {
1209 bdwrite(bp);
1210 }
1211 } while (uio->uio_resid > 0 && n > 0);
1212
1213 return (error);
1214 }
1215
1216 /*
1217 * Get an nfs cache block.
1218 *
1219 * Allocate a new one if the block isn't currently in the cache
1220 * and return the block marked busy. If the calling process is
1221 * interrupted by a signal for an interruptible mount point, return
1222 * NULL.
1223 *
1224 * The caller must carefully deal with the possible B_INVAL state of
1225 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1226 * indirectly), so synchronous reads can be issued without worrying about
1227 * the B_INVAL state. We have to be a little more careful when dealing
1228 * with writes (see comments in nfs_write()) when extending a file past
1229 * its EOF.
1230 */
1231 static struct buf *
1232 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1233 {
1234 struct buf *bp;
1235 struct mount *mp;
1236 struct nfsmount *nmp;
1237
1238 mp = vp->v_mount;
1239 nmp = VFSTONFS(mp);
1240
1241 if (nmp->nm_flag & NFSMNT_INT) {
1242 sigset_t oldset;
1243
1244 nfs_set_sigmask(td, &oldset);
1245 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1246 nfs_restore_sigmask(td, &oldset);
1247 while (bp == NULL) {
1248 if (nfs_sigintr(nmp, td))
1249 return (NULL);
1250 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1251 }
1252 } else {
1253 bp = getblk(vp, bn, size, 0, 0, 0);
1254 }
1255
1256 if (vp->v_type == VREG)
1257 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1258 return (bp);
1259 }
1260
1261 /*
1262 * Flush and invalidate all dirty buffers. If another process is already
1263 * doing the flush, just wait for completion.
1264 */
1265 int
1266 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1267 {
1268 struct nfsnode *np = VTONFS(vp);
1269 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1270 int error = 0, slpflag, slptimeo;
1271 int old_lock = 0;
1272
1273 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1274
1275 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1276 intrflg = 0;
1277 if (intrflg) {
1278 slpflag = PCATCH;
1279 slptimeo = 2 * hz;
1280 } else {
1281 slpflag = 0;
1282 slptimeo = 0;
1283 }
1284
1285 old_lock = nfs_upgrade_vnlock(vp);
1286 if (vp->v_iflag & VI_DOOMED) {
1287 /*
1288 * Since vgonel() uses the generic vinvalbuf() to flush
1289 * dirty buffers and it does not call this function, it
1290 * is safe to just return OK when VI_DOOMED is set.
1291 */
1292 nfs_downgrade_vnlock(vp, old_lock);
1293 return (0);
1294 }
1295
1296 /*
1297 * Now, flush as required.
1298 */
1299 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1300 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1301 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1302 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1303 /*
1304 * If the page clean was interrupted, fail the invalidation.
1305 * Not doing so, we run the risk of losing dirty pages in the
1306 * vinvalbuf() call below.
1307 */
1308 if (intrflg && (error = nfs_sigintr(nmp, td)))
1309 goto out;
1310 }
1311
1312 error = vinvalbuf(vp, flags, slpflag, 0);
1313 while (error) {
1314 if (intrflg && (error = nfs_sigintr(nmp, td)))
1315 goto out;
1316 error = vinvalbuf(vp, flags, 0, slptimeo);
1317 }
1318 mtx_lock(&np->n_mtx);
1319 if (np->n_directio_asyncwr == 0)
1320 np->n_flag &= ~NMODIFIED;
1321 mtx_unlock(&np->n_mtx);
1322 out:
1323 nfs_downgrade_vnlock(vp, old_lock);
1324 return error;
1325 }
1326
1327 /*
1328 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1329 * This is mainly to avoid queueing async I/O requests when the nfsiods
1330 * are all hung on a dead server.
1331 *
1332 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1333 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1334 */
1335 int
1336 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1337 {
1338 int iod;
1339 int gotiod;
1340 int slpflag = 0;
1341 int slptimeo = 0;
1342 int error, error2;
1343
1344 /*
1345 * Commits are usually short and sweet so lets save some cpu and
1346 * leave the async daemons for more important rpc's (such as reads
1347 * and writes).
1348 *
1349 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1350 * in the directory in order to update attributes. This can deadlock
1351 * with another thread that is waiting for async I/O to be done by
1352 * an nfsiod thread while holding a lock on one of these vnodes.
1353 * To avoid this deadlock, don't allow the async nfsiod threads to
1354 * perform Readdirplus RPCs.
1355 */
1356 mtx_lock(&nfs_iod_mtx);
1357 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1358 (nmp->nm_bufqiods > nfs_numasync / 2)) ||
1359 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1360 mtx_unlock(&nfs_iod_mtx);
1361 return(EIO);
1362 }
1363 again:
1364 if (nmp->nm_flag & NFSMNT_INT)
1365 slpflag = PCATCH;
1366 gotiod = FALSE;
1367
1368 /*
1369 * Find a free iod to process this request.
1370 */
1371 for (iod = 0; iod < nfs_numasync; iod++)
1372 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1373 gotiod = TRUE;
1374 break;
1375 }
1376
1377 /*
1378 * Try to create one if none are free.
1379 */
1380 if (!gotiod)
1381 nfs_nfsiodnew();
1382 else {
1383 /*
1384 * Found one, so wake it up and tell it which
1385 * mount to process.
1386 */
1387 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1388 iod, nmp));
1389 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1390 nfs_iodmount[iod] = nmp;
1391 nmp->nm_bufqiods++;
1392 wakeup(&nfs_iodwant[iod]);
1393 }
1394
1395 /*
1396 * If none are free, we may already have an iod working on this mount
1397 * point. If so, it will process our request.
1398 */
1399 if (!gotiod) {
1400 if (nmp->nm_bufqiods > 0) {
1401 NFS_DPF(ASYNCIO,
1402 ("nfs_asyncio: %d iods are already processing mount %p\n",
1403 nmp->nm_bufqiods, nmp));
1404 gotiod = TRUE;
1405 }
1406 }
1407
1408 /*
1409 * If we have an iod which can process the request, then queue
1410 * the buffer.
1411 */
1412 if (gotiod) {
1413 /*
1414 * Ensure that the queue never grows too large. We still want
1415 * to asynchronize so we block rather then return EIO.
1416 */
1417 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1418 NFS_DPF(ASYNCIO,
1419 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1420 nmp->nm_bufqwant = TRUE;
1421 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1422 slpflag | PRIBIO,
1423 "nfsaio", slptimeo);
1424 if (error) {
1425 error2 = nfs_sigintr(nmp, td);
1426 if (error2) {
1427 mtx_unlock(&nfs_iod_mtx);
1428 return (error2);
1429 }
1430 if (slpflag == PCATCH) {
1431 slpflag = 0;
1432 slptimeo = 2 * hz;
1433 }
1434 }
1435 /*
1436 * We might have lost our iod while sleeping,
1437 * so check and loop if nescessary.
1438 */
1439 goto again;
1440 }
1441
1442 /* We might have lost our nfsiod */
1443 if (nmp->nm_bufqiods == 0) {
1444 NFS_DPF(ASYNCIO,
1445 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1446 goto again;
1447 }
1448
1449 if (bp->b_iocmd == BIO_READ) {
1450 if (bp->b_rcred == NOCRED && cred != NOCRED)
1451 bp->b_rcred = crhold(cred);
1452 } else {
1453 if (bp->b_wcred == NOCRED && cred != NOCRED)
1454 bp->b_wcred = crhold(cred);
1455 }
1456
1457 if (bp->b_flags & B_REMFREE)
1458 bremfreef(bp);
1459 BUF_KERNPROC(bp);
1460 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1461 nmp->nm_bufqlen++;
1462 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1463 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1464 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1465 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1466 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1467 }
1468 mtx_unlock(&nfs_iod_mtx);
1469 return (0);
1470 }
1471
1472 mtx_unlock(&nfs_iod_mtx);
1473
1474 /*
1475 * All the iods are busy on other mounts, so return EIO to
1476 * force the caller to process the i/o synchronously.
1477 */
1478 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1479 return (EIO);
1480 }
1481
1482 void
1483 nfs_doio_directwrite(struct buf *bp)
1484 {
1485 int iomode, must_commit;
1486 struct uio *uiop = (struct uio *)bp->b_caller1;
1487 char *iov_base = uiop->uio_iov->iov_base;
1488 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1489
1490 iomode = NFSV3WRITE_FILESYNC;
1491 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1492 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1493 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1494 free(iov_base, M_NFSDIRECTIO);
1495 free(uiop->uio_iov, M_NFSDIRECTIO);
1496 free(uiop, M_NFSDIRECTIO);
1497 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1498 struct nfsnode *np = VTONFS(bp->b_vp);
1499 mtx_lock(&np->n_mtx);
1500 np->n_directio_asyncwr--;
1501 if (np->n_directio_asyncwr == 0) {
1502 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1503 if ((np->n_flag & NFSYNCWAIT)) {
1504 np->n_flag &= ~NFSYNCWAIT;
1505 wakeup((caddr_t)&np->n_directio_asyncwr);
1506 }
1507 }
1508 mtx_unlock(&np->n_mtx);
1509 }
1510 bp->b_vp = NULL;
1511 relpbuf(bp, &nfs_pbuf_freecnt);
1512 }
1513
1514 /*
1515 * Do an I/O operation to/from a cache block. This may be called
1516 * synchronously or from an nfsiod.
1517 */
1518 int
1519 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1520 {
1521 struct uio *uiop;
1522 struct nfsnode *np;
1523 struct nfsmount *nmp;
1524 int error = 0, iomode, must_commit = 0;
1525 struct uio uio;
1526 struct iovec io;
1527 struct proc *p = td ? td->td_proc : NULL;
1528 uint8_t iocmd;
1529
1530 np = VTONFS(vp);
1531 nmp = VFSTONFS(vp->v_mount);
1532 uiop = &uio;
1533 uiop->uio_iov = &io;
1534 uiop->uio_iovcnt = 1;
1535 uiop->uio_segflg = UIO_SYSSPACE;
1536 uiop->uio_td = td;
1537
1538 /*
1539 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1540 * do this here so we do not have to do it in all the code that
1541 * calls us.
1542 */
1543 bp->b_flags &= ~B_INVAL;
1544 bp->b_ioflags &= ~BIO_ERROR;
1545
1546 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1547 iocmd = bp->b_iocmd;
1548 if (iocmd == BIO_READ) {
1549 io.iov_len = uiop->uio_resid = bp->b_bcount;
1550 io.iov_base = bp->b_data;
1551 uiop->uio_rw = UIO_READ;
1552
1553 switch (vp->v_type) {
1554 case VREG:
1555 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1556 nfsstats.read_bios++;
1557 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1558
1559 if (!error) {
1560 if (uiop->uio_resid) {
1561 /*
1562 * If we had a short read with no error, we must have
1563 * hit a file hole. We should zero-fill the remainder.
1564 * This can also occur if the server hits the file EOF.
1565 *
1566 * Holes used to be able to occur due to pending
1567 * writes, but that is not possible any longer.
1568 */
1569 int nread = bp->b_bcount - uiop->uio_resid;
1570 int left = uiop->uio_resid;
1571
1572 if (left > 0)
1573 bzero((char *)bp->b_data + nread, left);
1574 uiop->uio_resid = 0;
1575 }
1576 }
1577 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1578 if (p && (vp->v_vflag & VV_TEXT)) {
1579 mtx_lock(&np->n_mtx);
1580 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1581 mtx_unlock(&np->n_mtx);
1582 PROC_LOCK(p);
1583 killproc(p, "text file modification");
1584 PROC_UNLOCK(p);
1585 } else
1586 mtx_unlock(&np->n_mtx);
1587 }
1588 break;
1589 case VLNK:
1590 uiop->uio_offset = (off_t)0;
1591 nfsstats.readlink_bios++;
1592 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1593 break;
1594 case VDIR:
1595 nfsstats.readdir_bios++;
1596 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1597 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1598 error = nfs_readdirplusrpc(vp, uiop, cr);
1599 if (error == NFSERR_NOTSUPP)
1600 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1601 }
1602 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1603 error = nfs_readdirrpc(vp, uiop, cr);
1604 /*
1605 * end-of-directory sets B_INVAL but does not generate an
1606 * error.
1607 */
1608 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1609 bp->b_flags |= B_INVAL;
1610 break;
1611 default:
1612 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1613 break;
1614 };
1615 if (error) {
1616 bp->b_ioflags |= BIO_ERROR;
1617 bp->b_error = error;
1618 }
1619 } else {
1620 /*
1621 * If we only need to commit, try to commit
1622 */
1623 if (bp->b_flags & B_NEEDCOMMIT) {
1624 int retv;
1625 off_t off;
1626
1627 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1628 retv = (nmp->nm_rpcops->nr_commit)(
1629 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1630 bp->b_wcred, td);
1631 if (retv == 0) {
1632 bp->b_dirtyoff = bp->b_dirtyend = 0;
1633 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1634 bp->b_resid = 0;
1635 bufdone(bp);
1636 return (0);
1637 }
1638 if (retv == NFSERR_STALEWRITEVERF) {
1639 nfs_clearcommit(vp->v_mount);
1640 }
1641 }
1642
1643 /*
1644 * Setup for actual write
1645 */
1646 mtx_lock(&np->n_mtx);
1647 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1648 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1649 mtx_unlock(&np->n_mtx);
1650
1651 if (bp->b_dirtyend > bp->b_dirtyoff) {
1652 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1653 - bp->b_dirtyoff;
1654 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1655 + bp->b_dirtyoff;
1656 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1657 uiop->uio_rw = UIO_WRITE;
1658 nfsstats.write_bios++;
1659
1660 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1661 iomode = NFSV3WRITE_UNSTABLE;
1662 else
1663 iomode = NFSV3WRITE_FILESYNC;
1664
1665 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1666
1667 /*
1668 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1669 * to cluster the buffers needing commit. This will allow
1670 * the system to submit a single commit rpc for the whole
1671 * cluster. We can do this even if the buffer is not 100%
1672 * dirty (relative to the NFS blocksize), so we optimize the
1673 * append-to-file-case.
1674 *
1675 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1676 * cleared because write clustering only works for commit
1677 * rpc's, not for the data portion of the write).
1678 */
1679
1680 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1681 bp->b_flags |= B_NEEDCOMMIT;
1682 if (bp->b_dirtyoff == 0
1683 && bp->b_dirtyend == bp->b_bcount)
1684 bp->b_flags |= B_CLUSTEROK;
1685 } else {
1686 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1687 }
1688
1689 /*
1690 * For an interrupted write, the buffer is still valid
1691 * and the write hasn't been pushed to the server yet,
1692 * so we can't set BIO_ERROR and report the interruption
1693 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1694 * is not relevant, so the rpc attempt is essentially
1695 * a noop. For the case of a V3 write rpc not being
1696 * committed to stable storage, the block is still
1697 * dirty and requires either a commit rpc or another
1698 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1699 * the block is reused. This is indicated by setting
1700 * the B_DELWRI and B_NEEDCOMMIT flags.
1701 *
1702 * If the buffer is marked B_PAGING, it does not reside on
1703 * the vp's paging queues so we cannot call bdirty(). The
1704 * bp in this case is not an NFS cache block so we should
1705 * be safe. XXX
1706 *
1707 * The logic below breaks up errors into recoverable and
1708 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1709 * and keep the buffer around for potential write retries.
1710 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1711 * and save the error in the nfsnode. This is less than ideal
1712 * but necessary. Keeping such buffers around could potentially
1713 * cause buffer exhaustion eventually (they can never be written
1714 * out, so will get constantly be re-dirtied). It also causes
1715 * all sorts of vfs panics. For non-recoverable write errors,
1716 * also invalidate the attrcache, so we'll be forced to go over
1717 * the wire for this object, returning an error to user on next
1718 * call (most of the time).
1719 */
1720 if (error == EINTR || error == EIO || error == ETIMEDOUT
1721 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1722 int s;
1723
1724 s = splbio();
1725 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1726 if ((bp->b_flags & B_PAGING) == 0) {
1727 bdirty(bp);
1728 bp->b_flags &= ~B_DONE;
1729 }
1730 if (error && (bp->b_flags & B_ASYNC) == 0)
1731 bp->b_flags |= B_EINTR;
1732 splx(s);
1733 } else {
1734 if (error) {
1735 bp->b_ioflags |= BIO_ERROR;
1736 bp->b_flags |= B_INVAL;
1737 bp->b_error = np->n_error = error;
1738 mtx_lock(&np->n_mtx);
1739 np->n_flag |= NWRITEERR;
1740 np->n_attrstamp = 0;
1741 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1742 mtx_unlock(&np->n_mtx);
1743 }
1744 bp->b_dirtyoff = bp->b_dirtyend = 0;
1745 }
1746 } else {
1747 bp->b_resid = 0;
1748 bufdone(bp);
1749 return (0);
1750 }
1751 }
1752 bp->b_resid = uiop->uio_resid;
1753 if (must_commit)
1754 nfs_clearcommit(vp->v_mount);
1755 bufdone(bp);
1756 return (error);
1757 }
1758
1759 /*
1760 * Used to aid in handling ftruncate() operations on the NFS client side.
1761 * Truncation creates a number of special problems for NFS. We have to
1762 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1763 * we have to properly handle VM pages or (potentially dirty) buffers
1764 * that straddle the truncation point.
1765 */
1766
1767 int
1768 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1769 {
1770 struct nfsnode *np = VTONFS(vp);
1771 u_quad_t tsize;
1772 int biosize = vp->v_bufobj.bo_bsize;
1773 int error = 0;
1774
1775 mtx_lock(&np->n_mtx);
1776 tsize = np->n_size;
1777 np->n_size = nsize;
1778 mtx_unlock(&np->n_mtx);
1779
1780 if (nsize < tsize) {
1781 struct buf *bp;
1782 daddr_t lbn;
1783 int bufsize;
1784
1785 /*
1786 * vtruncbuf() doesn't get the buffer overlapping the
1787 * truncation point. We may have a B_DELWRI and/or B_CACHE
1788 * buffer that now needs to be truncated.
1789 */
1790 error = vtruncbuf(vp, cred, nsize, biosize);
1791 lbn = nsize / biosize;
1792 bufsize = nsize - (lbn * biosize);
1793 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1794 if (!bp)
1795 return EINTR;
1796 if (bp->b_dirtyoff > bp->b_bcount)
1797 bp->b_dirtyoff = bp->b_bcount;
1798 if (bp->b_dirtyend > bp->b_bcount)
1799 bp->b_dirtyend = bp->b_bcount;
1800 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1801 brelse(bp);
1802 } else {
1803 vnode_pager_setsize(vp, nsize);
1804 }
1805 return(error);
1806 }
1807
Cache object: 417b53d9c88df2c112f6f44033ab9869
|