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: stable/9/sys/nfsclient/nfs_bio.c 250257 2013-05-04 21:56:39Z rmacklem $");
37
38 #include "opt_kdtrace.h"
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/bio.h>
43 #include <sys/buf.h>
44 #include <sys/kernel.h>
45 #include <sys/mbuf.h>
46 #include <sys/mount.h>
47 #include <sys/proc.h>
48 #include <sys/vmmeter.h>
49 #include <sys/vnode.h>
50
51 #include <vm/vm.h>
52 #include <vm/vm_param.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_object.h>
56 #include <vm/vm_pager.h>
57 #include <vm/vnode_pager.h>
58
59 #include <nfs/nfsproto.h>
60 #include <nfsclient/nfs.h>
61 #include <nfsclient/nfsmount.h>
62 #include <nfsclient/nfsnode.h>
63 #include <nfs/nfs_kdtrace.h>
64
65 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
66 struct thread *td);
67 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
68 struct ucred *cred, int ioflag);
69
70 extern int nfs_directio_enable;
71 extern int nfs_directio_allow_mmap;
72
73 /*
74 * Vnode op for VM getpages.
75 */
76 int
nfs_getpages(struct vop_getpages_args * ap)77 nfs_getpages(struct vop_getpages_args *ap)
78 {
79 int i, error, nextoff, size, toff, count, npages;
80 struct uio uio;
81 struct iovec iov;
82 vm_offset_t kva;
83 struct buf *bp;
84 struct vnode *vp;
85 struct thread *td;
86 struct ucred *cred;
87 struct nfsmount *nmp;
88 vm_object_t object;
89 vm_page_t *pages;
90 struct nfsnode *np;
91
92 vp = ap->a_vp;
93 np = VTONFS(vp);
94 td = curthread; /* XXX */
95 cred = curthread->td_ucred; /* XXX */
96 nmp = VFSTONFS(vp->v_mount);
97 pages = ap->a_m;
98 count = ap->a_count;
99
100 if ((object = vp->v_object) == NULL) {
101 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n");
102 return (VM_PAGER_ERROR);
103 }
104
105 if (nfs_directio_enable && !nfs_directio_allow_mmap) {
106 mtx_lock(&np->n_mtx);
107 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
108 mtx_unlock(&np->n_mtx);
109 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n");
110 return (VM_PAGER_ERROR);
111 } else
112 mtx_unlock(&np->n_mtx);
113 }
114
115 mtx_lock(&nmp->nm_mtx);
116 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
117 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
118 mtx_unlock(&nmp->nm_mtx);
119 /* We'll never get here for v4, because we always have fsinfo */
120 (void)nfs_fsinfo(nmp, vp, cred, td);
121 } else
122 mtx_unlock(&nmp->nm_mtx);
123
124 npages = btoc(count);
125
126 /*
127 * If the requested page is partially valid, just return it and
128 * allow the pager to zero-out the blanks. Partially valid pages
129 * can only occur at the file EOF.
130 */
131 VM_OBJECT_LOCK(object);
132 if (pages[ap->a_reqpage]->valid != 0) {
133 for (i = 0; i < npages; ++i) {
134 if (i != ap->a_reqpage) {
135 vm_page_lock(pages[i]);
136 vm_page_free(pages[i]);
137 vm_page_unlock(pages[i]);
138 }
139 }
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 for (i = 0; i < npages; ++i) {
175 if (i != ap->a_reqpage) {
176 vm_page_lock(pages[i]);
177 vm_page_free(pages[i]);
178 vm_page_unlock(pages[i]);
179 }
180 }
181 VM_OBJECT_UNLOCK(object);
182 return (VM_PAGER_ERROR);
183 }
184
185 /*
186 * Calculate the number of bytes read and validate only that number
187 * of bytes. Note that due to pending writes, size may be 0. This
188 * does not mean that the remaining data is invalid!
189 */
190
191 size = count - uio.uio_resid;
192 VM_OBJECT_LOCK(object);
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
216 * occured we may have hit a zero-fill
217 * section. We leave valid set to 0, and page
218 * is freed by vm_page_readahead_finish() if
219 * its index is not equal to requested, or
220 * page is zeroed and set valid by
221 * vm_pager_get_pages() for requested page.
222 */
223 ;
224 }
225 if (i != ap->a_reqpage)
226 vm_page_readahead_finish(m);
227 }
228 VM_OBJECT_UNLOCK(object);
229 return (0);
230 }
231
232 /*
233 * Vnode op for VM putpages.
234 */
235 int
nfs_putpages(struct vop_putpages_args * ap)236 nfs_putpages(struct vop_putpages_args *ap)
237 {
238 struct uio uio;
239 struct iovec iov;
240 vm_offset_t kva;
241 struct buf *bp;
242 int iomode, must_commit, i, error, npages, count;
243 off_t offset;
244 int *rtvals;
245 struct vnode *vp;
246 struct thread *td;
247 struct ucred *cred;
248 struct nfsmount *nmp;
249 struct nfsnode *np;
250 vm_page_t *pages;
251
252 vp = ap->a_vp;
253 np = VTONFS(vp);
254 td = curthread; /* XXX */
255 /* Set the cred to n_writecred for the write rpcs. */
256 if (np->n_writecred != NULL)
257 cred = crhold(np->n_writecred);
258 else
259 cred = crhold(curthread->td_ucred); /* XXX */
260 nmp = VFSTONFS(vp->v_mount);
261 pages = ap->a_m;
262 count = ap->a_count;
263 rtvals = ap->a_rtvals;
264 npages = btoc(count);
265 offset = IDX_TO_OFF(pages[0]->pindex);
266
267 mtx_lock(&nmp->nm_mtx);
268 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
269 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
270 mtx_unlock(&nmp->nm_mtx);
271 (void)nfs_fsinfo(nmp, vp, cred, td);
272 } else
273 mtx_unlock(&nmp->nm_mtx);
274
275 mtx_lock(&np->n_mtx);
276 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
277 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
278 mtx_unlock(&np->n_mtx);
279 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
280 mtx_lock(&np->n_mtx);
281 }
282
283 for (i = 0; i < npages; i++)
284 rtvals[i] = VM_PAGER_ERROR;
285
286 /*
287 * When putting pages, do not extend file past EOF.
288 */
289 if (offset + count > np->n_size) {
290 count = np->n_size - offset;
291 if (count < 0)
292 count = 0;
293 }
294 mtx_unlock(&np->n_mtx);
295
296 /*
297 * We use only the kva address for the buffer, but this is extremely
298 * convienient and fast.
299 */
300 bp = getpbuf(&nfs_pbuf_freecnt);
301
302 kva = (vm_offset_t) bp->b_data;
303 pmap_qenter(kva, pages, npages);
304 PCPU_INC(cnt.v_vnodeout);
305 PCPU_ADD(cnt.v_vnodepgsout, count);
306
307 iov.iov_base = (caddr_t) kva;
308 iov.iov_len = count;
309 uio.uio_iov = &iov;
310 uio.uio_iovcnt = 1;
311 uio.uio_offset = offset;
312 uio.uio_resid = count;
313 uio.uio_segflg = UIO_SYSSPACE;
314 uio.uio_rw = UIO_WRITE;
315 uio.uio_td = td;
316
317 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
318 iomode = NFSV3WRITE_UNSTABLE;
319 else
320 iomode = NFSV3WRITE_FILESYNC;
321
322 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
323 crfree(cred);
324
325 pmap_qremove(kva, npages);
326 relpbuf(bp, &nfs_pbuf_freecnt);
327
328 if (!error) {
329 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
330 if (must_commit) {
331 nfs_clearcommit(vp->v_mount);
332 }
333 }
334 return rtvals[0];
335 }
336
337 /*
338 * For nfs, cache consistency can only be maintained approximately.
339 * Although RFC1094 does not specify the criteria, the following is
340 * believed to be compatible with the reference port.
341 * For nfs:
342 * If the file's modify time on the server has changed since the
343 * last read rpc or you have written to the file,
344 * you may have lost data cache consistency with the
345 * server, so flush all of the file's data out of the cache.
346 * Then force a getattr rpc to ensure that you have up to date
347 * attributes.
348 * NB: This implies that cache data can be read when up to
349 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
350 * attributes this could be forced by setting n_attrstamp to 0 before
351 * the VOP_GETATTR() call.
352 */
353 static inline int
nfs_bioread_check_cons(struct vnode * vp,struct thread * td,struct ucred * cred)354 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
355 {
356 int error = 0;
357 struct vattr vattr;
358 struct nfsnode *np = VTONFS(vp);
359 int old_lock;
360 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
361
362 /*
363 * Grab the exclusive lock before checking whether the cache is
364 * consistent.
365 * XXX - We can make this cheaper later (by acquiring cheaper locks).
366 * But for now, this suffices.
367 */
368 old_lock = nfs_upgrade_vnlock(vp);
369 if (vp->v_iflag & VI_DOOMED) {
370 nfs_downgrade_vnlock(vp, old_lock);
371 return (EBADF);
372 }
373
374 mtx_lock(&np->n_mtx);
375 if (np->n_flag & NMODIFIED) {
376 mtx_unlock(&np->n_mtx);
377 if (vp->v_type != VREG) {
378 if (vp->v_type != VDIR)
379 panic("nfs: bioread, not dir");
380 (nmp->nm_rpcops->nr_invaldir)(vp);
381 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
382 if (error)
383 goto out;
384 }
385 np->n_attrstamp = 0;
386 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
387 error = VOP_GETATTR(vp, &vattr, cred);
388 if (error)
389 goto out;
390 mtx_lock(&np->n_mtx);
391 np->n_mtime = vattr.va_mtime;
392 mtx_unlock(&np->n_mtx);
393 } else {
394 mtx_unlock(&np->n_mtx);
395 error = VOP_GETATTR(vp, &vattr, cred);
396 if (error)
397 return (error);
398 mtx_lock(&np->n_mtx);
399 if ((np->n_flag & NSIZECHANGED)
400 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
401 mtx_unlock(&np->n_mtx);
402 if (vp->v_type == VDIR)
403 (nmp->nm_rpcops->nr_invaldir)(vp);
404 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
405 if (error)
406 goto out;
407 mtx_lock(&np->n_mtx);
408 np->n_mtime = vattr.va_mtime;
409 np->n_flag &= ~NSIZECHANGED;
410 }
411 mtx_unlock(&np->n_mtx);
412 }
413 out:
414 nfs_downgrade_vnlock(vp, old_lock);
415 return error;
416 }
417
418 /*
419 * Vnode op for read using bio
420 */
421 int
nfs_bioread(struct vnode * vp,struct uio * uio,int ioflag,struct ucred * cred)422 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
423 {
424 struct nfsnode *np = VTONFS(vp);
425 int biosize, i;
426 struct buf *bp, *rabp;
427 struct thread *td;
428 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
429 daddr_t lbn, rabn;
430 off_t end;
431 int bcount;
432 int seqcount;
433 int nra, error = 0, n = 0, on = 0;
434
435 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
436 if (uio->uio_resid == 0)
437 return (0);
438 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
439 return (EINVAL);
440 td = uio->uio_td;
441
442 mtx_lock(&nmp->nm_mtx);
443 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
444 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
445 mtx_unlock(&nmp->nm_mtx);
446 (void)nfs_fsinfo(nmp, vp, cred, td);
447 } else
448 mtx_unlock(&nmp->nm_mtx);
449
450 end = uio->uio_offset + uio->uio_resid;
451 if (vp->v_type != VDIR &&
452 (end > nmp->nm_maxfilesize || end < uio->uio_offset))
453 return (EFBIG);
454
455 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
456 /* No caching/ no readaheads. Just read data into the user buffer */
457 return nfs_readrpc(vp, uio, cred);
458
459 biosize = vp->v_bufobj.bo_bsize;
460 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
461
462 error = nfs_bioread_check_cons(vp, td, cred);
463 if (error)
464 return error;
465
466 do {
467 u_quad_t nsize;
468
469 mtx_lock(&np->n_mtx);
470 nsize = np->n_size;
471 mtx_unlock(&np->n_mtx);
472
473 switch (vp->v_type) {
474 case VREG:
475 nfsstats.biocache_reads++;
476 lbn = uio->uio_offset / biosize;
477 on = uio->uio_offset & (biosize - 1);
478
479 /*
480 * Start the read ahead(s), as required.
481 */
482 if (nmp->nm_readahead > 0) {
483 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
484 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
485 rabn = lbn + 1 + nra;
486 if (incore(&vp->v_bufobj, rabn) == NULL) {
487 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
488 if (!rabp) {
489 error = nfs_sigintr(nmp, td);
490 return (error ? error : EINTR);
491 }
492 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
493 rabp->b_flags |= B_ASYNC;
494 rabp->b_iocmd = BIO_READ;
495 vfs_busy_pages(rabp, 0);
496 if (nfs_asyncio(nmp, rabp, cred, td)) {
497 rabp->b_flags |= B_INVAL;
498 rabp->b_ioflags |= BIO_ERROR;
499 vfs_unbusy_pages(rabp);
500 brelse(rabp);
501 break;
502 }
503 } else {
504 brelse(rabp);
505 }
506 }
507 }
508 }
509
510 /* Note that bcount is *not* DEV_BSIZE aligned. */
511 bcount = biosize;
512 if ((off_t)lbn * biosize >= nsize) {
513 bcount = 0;
514 } else if ((off_t)(lbn + 1) * biosize > nsize) {
515 bcount = nsize - (off_t)lbn * biosize;
516 }
517 bp = nfs_getcacheblk(vp, lbn, bcount, td);
518
519 if (!bp) {
520 error = nfs_sigintr(nmp, td);
521 return (error ? error : EINTR);
522 }
523
524 /*
525 * If B_CACHE is not set, we must issue the read. If this
526 * fails, we return an error.
527 */
528
529 if ((bp->b_flags & B_CACHE) == 0) {
530 bp->b_iocmd = BIO_READ;
531 vfs_busy_pages(bp, 0);
532 error = nfs_doio(vp, bp, cred, td);
533 if (error) {
534 brelse(bp);
535 return (error);
536 }
537 }
538
539 /*
540 * on is the offset into the current bp. Figure out how many
541 * bytes we can copy out of the bp. Note that bcount is
542 * NOT DEV_BSIZE aligned.
543 *
544 * Then figure out how many bytes we can copy into the uio.
545 */
546
547 n = 0;
548 if (on < bcount)
549 n = MIN((unsigned)(bcount - on), uio->uio_resid);
550 break;
551 case VLNK:
552 nfsstats.biocache_readlinks++;
553 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
554 if (!bp) {
555 error = nfs_sigintr(nmp, td);
556 return (error ? error : EINTR);
557 }
558 if ((bp->b_flags & B_CACHE) == 0) {
559 bp->b_iocmd = BIO_READ;
560 vfs_busy_pages(bp, 0);
561 error = nfs_doio(vp, bp, cred, td);
562 if (error) {
563 bp->b_ioflags |= BIO_ERROR;
564 brelse(bp);
565 return (error);
566 }
567 }
568 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
569 on = 0;
570 break;
571 case VDIR:
572 nfsstats.biocache_readdirs++;
573 if (np->n_direofoffset
574 && uio->uio_offset >= np->n_direofoffset) {
575 return (0);
576 }
577 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
578 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
579 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
580 if (!bp) {
581 error = nfs_sigintr(nmp, td);
582 return (error ? error : EINTR);
583 }
584 if ((bp->b_flags & B_CACHE) == 0) {
585 bp->b_iocmd = BIO_READ;
586 vfs_busy_pages(bp, 0);
587 error = nfs_doio(vp, bp, cred, td);
588 if (error) {
589 brelse(bp);
590 }
591 while (error == NFSERR_BAD_COOKIE) {
592 (nmp->nm_rpcops->nr_invaldir)(vp);
593 error = nfs_vinvalbuf(vp, 0, td, 1);
594 /*
595 * Yuck! The directory has been modified on the
596 * server. The only way to get the block is by
597 * reading from the beginning to get all the
598 * offset cookies.
599 *
600 * Leave the last bp intact unless there is an error.
601 * Loop back up to the while if the error is another
602 * NFSERR_BAD_COOKIE (double yuch!).
603 */
604 for (i = 0; i <= lbn && !error; i++) {
605 if (np->n_direofoffset
606 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
607 return (0);
608 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
609 if (!bp) {
610 error = nfs_sigintr(nmp, td);
611 return (error ? error : EINTR);
612 }
613 if ((bp->b_flags & B_CACHE) == 0) {
614 bp->b_iocmd = BIO_READ;
615 vfs_busy_pages(bp, 0);
616 error = nfs_doio(vp, bp, cred, td);
617 /*
618 * no error + B_INVAL == directory EOF,
619 * use the block.
620 */
621 if (error == 0 && (bp->b_flags & B_INVAL))
622 break;
623 }
624 /*
625 * An error will throw away the block and the
626 * for loop will break out. If no error and this
627 * is not the block we want, we throw away the
628 * block and go for the next one via the for loop.
629 */
630 if (error || i < lbn)
631 brelse(bp);
632 }
633 }
634 /*
635 * The above while is repeated if we hit another cookie
636 * error. If we hit an error and it wasn't a cookie error,
637 * we give up.
638 */
639 if (error)
640 return (error);
641 }
642
643 /*
644 * If not eof and read aheads are enabled, start one.
645 * (You need the current block first, so that you have the
646 * directory offset cookie of the next block.)
647 */
648 if (nmp->nm_readahead > 0 &&
649 (bp->b_flags & B_INVAL) == 0 &&
650 (np->n_direofoffset == 0 ||
651 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
652 incore(&vp->v_bufobj, lbn + 1) == NULL) {
653 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
654 if (rabp) {
655 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
656 rabp->b_flags |= B_ASYNC;
657 rabp->b_iocmd = BIO_READ;
658 vfs_busy_pages(rabp, 0);
659 if (nfs_asyncio(nmp, rabp, cred, td)) {
660 rabp->b_flags |= B_INVAL;
661 rabp->b_ioflags |= BIO_ERROR;
662 vfs_unbusy_pages(rabp);
663 brelse(rabp);
664 }
665 } else {
666 brelse(rabp);
667 }
668 }
669 }
670 /*
671 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
672 * chopped for the EOF condition, we cannot tell how large
673 * NFS directories are going to be until we hit EOF. So
674 * an NFS directory buffer is *not* chopped to its EOF. Now,
675 * it just so happens that b_resid will effectively chop it
676 * to EOF. *BUT* this information is lost if the buffer goes
677 * away and is reconstituted into a B_CACHE state ( due to
678 * being VMIO ) later. So we keep track of the directory eof
679 * in np->n_direofoffset and chop it off as an extra step
680 * right here.
681 */
682 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
683 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
684 n = np->n_direofoffset - uio->uio_offset;
685 break;
686 default:
687 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
688 bp = NULL;
689 break;
690 };
691
692 if (n > 0) {
693 error = uiomove(bp->b_data + on, (int)n, uio);
694 }
695 if (vp->v_type == VLNK)
696 n = 0;
697 if (bp != NULL)
698 brelse(bp);
699 } while (error == 0 && uio->uio_resid > 0 && n > 0);
700 return (error);
701 }
702
703 /*
704 * The NFS write path cannot handle iovecs with len > 1. So we need to
705 * break up iovecs accordingly (restricting them to wsize).
706 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
707 * For the ASYNC case, 2 copies are needed. The first a copy from the
708 * user buffer to a staging buffer and then a second copy from the staging
709 * buffer to mbufs. This can be optimized by copying from the user buffer
710 * directly into mbufs and passing the chain down, but that requires a
711 * fair amount of re-working of the relevant codepaths (and can be done
712 * later).
713 */
714 static int
nfs_directio_write(vp,uiop,cred,ioflag)715 nfs_directio_write(vp, uiop, cred, ioflag)
716 struct vnode *vp;
717 struct uio *uiop;
718 struct ucred *cred;
719 int ioflag;
720 {
721 int error;
722 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
723 struct thread *td = uiop->uio_td;
724 int size;
725 int wsize;
726
727 mtx_lock(&nmp->nm_mtx);
728 wsize = nmp->nm_wsize;
729 mtx_unlock(&nmp->nm_mtx);
730 if (ioflag & IO_SYNC) {
731 int iomode, must_commit;
732 struct uio uio;
733 struct iovec iov;
734 do_sync:
735 while (uiop->uio_resid > 0) {
736 size = MIN(uiop->uio_resid, wsize);
737 size = MIN(uiop->uio_iov->iov_len, size);
738 iov.iov_base = uiop->uio_iov->iov_base;
739 iov.iov_len = size;
740 uio.uio_iov = &iov;
741 uio.uio_iovcnt = 1;
742 uio.uio_offset = uiop->uio_offset;
743 uio.uio_resid = size;
744 uio.uio_segflg = UIO_USERSPACE;
745 uio.uio_rw = UIO_WRITE;
746 uio.uio_td = td;
747 iomode = NFSV3WRITE_FILESYNC;
748 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
749 &iomode, &must_commit);
750 KASSERT((must_commit == 0),
751 ("nfs_directio_write: Did not commit write"));
752 if (error)
753 return (error);
754 uiop->uio_offset += size;
755 uiop->uio_resid -= size;
756 if (uiop->uio_iov->iov_len <= size) {
757 uiop->uio_iovcnt--;
758 uiop->uio_iov++;
759 } else {
760 uiop->uio_iov->iov_base =
761 (char *)uiop->uio_iov->iov_base + size;
762 uiop->uio_iov->iov_len -= size;
763 }
764 }
765 } else {
766 struct uio *t_uio;
767 struct iovec *t_iov;
768 struct buf *bp;
769
770 /*
771 * Break up the write into blocksize chunks and hand these
772 * over to nfsiod's for write back.
773 * Unfortunately, this incurs a copy of the data. Since
774 * the user could modify the buffer before the write is
775 * initiated.
776 *
777 * The obvious optimization here is that one of the 2 copies
778 * in the async write path can be eliminated by copying the
779 * data here directly into mbufs and passing the mbuf chain
780 * down. But that will require a fair amount of re-working
781 * of the code and can be done if there's enough interest
782 * in NFS directio access.
783 */
784 while (uiop->uio_resid > 0) {
785 size = MIN(uiop->uio_resid, wsize);
786 size = MIN(uiop->uio_iov->iov_len, size);
787 bp = getpbuf(&nfs_pbuf_freecnt);
788 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
789 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
790 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
791 t_iov->iov_len = size;
792 t_uio->uio_iov = t_iov;
793 t_uio->uio_iovcnt = 1;
794 t_uio->uio_offset = uiop->uio_offset;
795 t_uio->uio_resid = size;
796 t_uio->uio_segflg = UIO_SYSSPACE;
797 t_uio->uio_rw = UIO_WRITE;
798 t_uio->uio_td = td;
799 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
800 uiop->uio_segflg == UIO_SYSSPACE,
801 ("nfs_directio_write: Bad uio_segflg"));
802 if (uiop->uio_segflg == UIO_USERSPACE) {
803 error = copyin(uiop->uio_iov->iov_base,
804 t_iov->iov_base, size);
805 if (error != 0)
806 goto err_free;
807 } else
808 /*
809 * UIO_SYSSPACE may never happen, but handle
810 * it just in case it does.
811 */
812 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
813 size);
814 bp->b_flags |= B_DIRECT;
815 bp->b_iocmd = BIO_WRITE;
816 if (cred != NOCRED) {
817 crhold(cred);
818 bp->b_wcred = cred;
819 } else
820 bp->b_wcred = NOCRED;
821 bp->b_caller1 = (void *)t_uio;
822 bp->b_vp = vp;
823 error = nfs_asyncio(nmp, bp, NOCRED, td);
824 err_free:
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
nfs_write(struct vop_write_args * ap)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 off_t end;
868 int bcount;
869 int n, on, error = 0;
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 end = uio->uio_offset + uio->uio_resid;
936 if (end > nmp->nm_maxfilesize || end < uio->uio_offset)
937 return (EFBIG);
938 if (uio->uio_resid == 0)
939 return (0);
940
941 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
942 return nfs_directio_write(vp, uio, cred, ioflag);
943
944 /*
945 * Maybe this should be above the vnode op call, but so long as
946 * file servers have no limits, i don't think it matters
947 */
948 if (vn_rlimit_fsize(vp, uio, td))
949 return (EFBIG);
950
951 biosize = vp->v_bufobj.bo_bsize;
952 /*
953 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
954 * would exceed the local maximum per-file write commit size when
955 * combined with those, we must decide whether to flush,
956 * go synchronous, or return error. We don't bother checking
957 * IO_UNIT -- we just make all writes atomic anyway, as there's
958 * no point optimizing for something that really won't ever happen.
959 */
960 if (!(ioflag & IO_SYNC)) {
961 int nflag;
962
963 mtx_lock(&np->n_mtx);
964 nflag = np->n_flag;
965 mtx_unlock(&np->n_mtx);
966 int needrestart = 0;
967 if (nmp->nm_wcommitsize < uio->uio_resid) {
968 /*
969 * If this request could not possibly be completed
970 * without exceeding the maximum outstanding write
971 * commit size, see if we can convert it into a
972 * synchronous write operation.
973 */
974 if (ioflag & IO_NDELAY)
975 return (EAGAIN);
976 ioflag |= IO_SYNC;
977 if (nflag & NMODIFIED)
978 needrestart = 1;
979 } else if (nflag & NMODIFIED) {
980 int wouldcommit = 0;
981 BO_LOCK(&vp->v_bufobj);
982 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
983 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
984 b_bobufs) {
985 if (bp->b_flags & B_NEEDCOMMIT)
986 wouldcommit += bp->b_bcount;
987 }
988 }
989 BO_UNLOCK(&vp->v_bufobj);
990 /*
991 * Since we're not operating synchronously and
992 * bypassing the buffer cache, we are in a commit
993 * and holding all of these buffers whether
994 * transmitted or not. If not limited, this
995 * will lead to the buffer cache deadlocking,
996 * as no one else can flush our uncommitted buffers.
997 */
998 wouldcommit += uio->uio_resid;
999 /*
1000 * If we would initially exceed the maximum
1001 * outstanding write commit size, flush and restart.
1002 */
1003 if (wouldcommit > nmp->nm_wcommitsize)
1004 needrestart = 1;
1005 }
1006 if (needrestart)
1007 goto flush_and_restart;
1008 }
1009
1010 do {
1011 nfsstats.biocache_writes++;
1012 lbn = uio->uio_offset / biosize;
1013 on = uio->uio_offset & (biosize-1);
1014 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1015 again:
1016 /*
1017 * Handle direct append and file extension cases, calculate
1018 * unaligned buffer size.
1019 */
1020 mtx_lock(&np->n_mtx);
1021 if (uio->uio_offset == np->n_size && n) {
1022 mtx_unlock(&np->n_mtx);
1023 /*
1024 * Get the buffer (in its pre-append state to maintain
1025 * B_CACHE if it was previously set). Resize the
1026 * nfsnode after we have locked the buffer to prevent
1027 * readers from reading garbage.
1028 */
1029 bcount = on;
1030 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1031
1032 if (bp != NULL) {
1033 long save;
1034
1035 mtx_lock(&np->n_mtx);
1036 np->n_size = uio->uio_offset + n;
1037 np->n_flag |= NMODIFIED;
1038 vnode_pager_setsize(vp, np->n_size);
1039 mtx_unlock(&np->n_mtx);
1040
1041 save = bp->b_flags & B_CACHE;
1042 bcount += n;
1043 allocbuf(bp, bcount);
1044 bp->b_flags |= save;
1045 }
1046 } else {
1047 /*
1048 * Obtain the locked cache block first, and then
1049 * adjust the file's size as appropriate.
1050 */
1051 bcount = on + n;
1052 if ((off_t)lbn * biosize + bcount < np->n_size) {
1053 if ((off_t)(lbn + 1) * biosize < np->n_size)
1054 bcount = biosize;
1055 else
1056 bcount = np->n_size - (off_t)lbn * biosize;
1057 }
1058 mtx_unlock(&np->n_mtx);
1059 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1060 mtx_lock(&np->n_mtx);
1061 if (uio->uio_offset + n > np->n_size) {
1062 np->n_size = uio->uio_offset + n;
1063 np->n_flag |= NMODIFIED;
1064 vnode_pager_setsize(vp, np->n_size);
1065 }
1066 mtx_unlock(&np->n_mtx);
1067 }
1068
1069 if (!bp) {
1070 error = nfs_sigintr(nmp, td);
1071 if (!error)
1072 error = EINTR;
1073 break;
1074 }
1075
1076 /*
1077 * Issue a READ if B_CACHE is not set. In special-append
1078 * mode, B_CACHE is based on the buffer prior to the write
1079 * op and is typically set, avoiding the read. If a read
1080 * is required in special append mode, the server will
1081 * probably send us a short-read since we extended the file
1082 * on our end, resulting in b_resid == 0 and, thusly,
1083 * B_CACHE getting set.
1084 *
1085 * We can also avoid issuing the read if the write covers
1086 * the entire buffer. We have to make sure the buffer state
1087 * is reasonable in this case since we will not be initiating
1088 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1089 * more information.
1090 *
1091 * B_CACHE may also be set due to the buffer being cached
1092 * normally.
1093 */
1094
1095 if (on == 0 && n == bcount) {
1096 bp->b_flags |= B_CACHE;
1097 bp->b_flags &= ~B_INVAL;
1098 bp->b_ioflags &= ~BIO_ERROR;
1099 }
1100
1101 if ((bp->b_flags & B_CACHE) == 0) {
1102 bp->b_iocmd = BIO_READ;
1103 vfs_busy_pages(bp, 0);
1104 error = nfs_doio(vp, bp, cred, td);
1105 if (error) {
1106 brelse(bp);
1107 break;
1108 }
1109 }
1110 if (bp->b_wcred == NOCRED)
1111 bp->b_wcred = crhold(cred);
1112 mtx_lock(&np->n_mtx);
1113 np->n_flag |= NMODIFIED;
1114 mtx_unlock(&np->n_mtx);
1115
1116 /*
1117 * If dirtyend exceeds file size, chop it down. This should
1118 * not normally occur but there is an append race where it
1119 * might occur XXX, so we log it.
1120 *
1121 * If the chopping creates a reverse-indexed or degenerate
1122 * situation with dirtyoff/end, we 0 both of them.
1123 */
1124
1125 if (bp->b_dirtyend > bcount) {
1126 nfs_printf("NFS append race @%lx:%d\n",
1127 (long)bp->b_blkno * DEV_BSIZE,
1128 bp->b_dirtyend - bcount);
1129 bp->b_dirtyend = bcount;
1130 }
1131
1132 if (bp->b_dirtyoff >= bp->b_dirtyend)
1133 bp->b_dirtyoff = bp->b_dirtyend = 0;
1134
1135 /*
1136 * If the new write will leave a contiguous dirty
1137 * area, just update the b_dirtyoff and b_dirtyend,
1138 * otherwise force a write rpc of the old dirty area.
1139 *
1140 * While it is possible to merge discontiguous writes due to
1141 * our having a B_CACHE buffer ( and thus valid read data
1142 * for the hole), we don't because it could lead to
1143 * significant cache coherency problems with multiple clients,
1144 * especially if locking is implemented later on.
1145 *
1146 * as an optimization we could theoretically maintain
1147 * a linked list of discontinuous areas, but we would still
1148 * have to commit them separately so there isn't much
1149 * advantage to it except perhaps a bit of asynchronization.
1150 */
1151
1152 if (bp->b_dirtyend > 0 &&
1153 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1154 if (bwrite(bp) == EINTR) {
1155 error = EINTR;
1156 break;
1157 }
1158 goto again;
1159 }
1160
1161 error = uiomove((char *)bp->b_data + on, n, uio);
1162
1163 /*
1164 * Since this block is being modified, it must be written
1165 * again and not just committed. Since write clustering does
1166 * not work for the stage 1 data write, only the stage 2
1167 * commit rpc, we have to clear B_CLUSTEROK as well.
1168 */
1169 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1170
1171 if (error) {
1172 bp->b_ioflags |= BIO_ERROR;
1173 brelse(bp);
1174 break;
1175 }
1176
1177 /*
1178 * Only update dirtyoff/dirtyend if not a degenerate
1179 * condition.
1180 */
1181 if (n) {
1182 if (bp->b_dirtyend > 0) {
1183 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1184 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1185 } else {
1186 bp->b_dirtyoff = on;
1187 bp->b_dirtyend = on + n;
1188 }
1189 vfs_bio_set_valid(bp, on, n);
1190 }
1191
1192 /*
1193 * If IO_SYNC do bwrite().
1194 *
1195 * IO_INVAL appears to be unused. The idea appears to be
1196 * to turn off caching in this case. Very odd. XXX
1197 */
1198 if ((ioflag & IO_SYNC)) {
1199 if (ioflag & IO_INVAL)
1200 bp->b_flags |= B_NOCACHE;
1201 error = bwrite(bp);
1202 if (error)
1203 break;
1204 } else if ((n + on) == biosize) {
1205 bp->b_flags |= B_ASYNC;
1206 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1207 } else {
1208 bdwrite(bp);
1209 }
1210 } while (uio->uio_resid > 0 && n > 0);
1211
1212 return (error);
1213 }
1214
1215 /*
1216 * Get an nfs cache block.
1217 *
1218 * Allocate a new one if the block isn't currently in the cache
1219 * and return the block marked busy. If the calling process is
1220 * interrupted by a signal for an interruptible mount point, return
1221 * NULL.
1222 *
1223 * The caller must carefully deal with the possible B_INVAL state of
1224 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1225 * indirectly), so synchronous reads can be issued without worrying about
1226 * the B_INVAL state. We have to be a little more careful when dealing
1227 * with writes (see comments in nfs_write()) when extending a file past
1228 * its EOF.
1229 */
1230 static struct buf *
nfs_getcacheblk(struct vnode * vp,daddr_t bn,int size,struct thread * td)1231 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1232 {
1233 struct buf *bp;
1234 struct mount *mp;
1235 struct nfsmount *nmp;
1236
1237 mp = vp->v_mount;
1238 nmp = VFSTONFS(mp);
1239
1240 if (nmp->nm_flag & NFSMNT_INT) {
1241 sigset_t oldset;
1242
1243 nfs_set_sigmask(td, &oldset);
1244 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1245 nfs_restore_sigmask(td, &oldset);
1246 while (bp == NULL) {
1247 if (nfs_sigintr(nmp, td))
1248 return (NULL);
1249 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1250 }
1251 } else {
1252 bp = getblk(vp, bn, size, 0, 0, 0);
1253 }
1254
1255 if (vp->v_type == VREG)
1256 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1257 return (bp);
1258 }
1259
1260 /*
1261 * Flush and invalidate all dirty buffers. If another process is already
1262 * doing the flush, just wait for completion.
1263 */
1264 int
nfs_vinvalbuf(struct vnode * vp,int flags,struct thread * td,int intrflg)1265 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1266 {
1267 struct nfsnode *np = VTONFS(vp);
1268 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1269 int error = 0, slpflag, slptimeo;
1270 int old_lock = 0;
1271
1272 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1273
1274 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1275 intrflg = 0;
1276 if (intrflg) {
1277 slpflag = NFS_PCATCH;
1278 slptimeo = 2 * hz;
1279 } else {
1280 slpflag = 0;
1281 slptimeo = 0;
1282 }
1283
1284 old_lock = nfs_upgrade_vnlock(vp);
1285 if (vp->v_iflag & VI_DOOMED) {
1286 /*
1287 * Since vgonel() uses the generic vinvalbuf() to flush
1288 * dirty buffers and it does not call this function, it
1289 * is safe to just return OK when VI_DOOMED is set.
1290 */
1291 nfs_downgrade_vnlock(vp, old_lock);
1292 return (0);
1293 }
1294
1295 /*
1296 * Now, flush as required.
1297 */
1298 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1299 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1300 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1301 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1302 /*
1303 * If the page clean was interrupted, fail the invalidation.
1304 * Not doing so, we run the risk of losing dirty pages in the
1305 * vinvalbuf() call below.
1306 */
1307 if (intrflg && (error = nfs_sigintr(nmp, td)))
1308 goto out;
1309 }
1310
1311 error = vinvalbuf(vp, flags, slpflag, 0);
1312 while (error) {
1313 if (intrflg && (error = nfs_sigintr(nmp, td)))
1314 goto out;
1315 error = vinvalbuf(vp, flags, 0, slptimeo);
1316 }
1317 mtx_lock(&np->n_mtx);
1318 if (np->n_directio_asyncwr == 0)
1319 np->n_flag &= ~NMODIFIED;
1320 mtx_unlock(&np->n_mtx);
1321 out:
1322 nfs_downgrade_vnlock(vp, old_lock);
1323 return error;
1324 }
1325
1326 /*
1327 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1328 * This is mainly to avoid queueing async I/O requests when the nfsiods
1329 * are all hung on a dead server.
1330 *
1331 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1332 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1333 */
1334 int
nfs_asyncio(struct nfsmount * nmp,struct buf * bp,struct ucred * cred,struct thread * td)1335 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1336 {
1337 int iod;
1338 int gotiod;
1339 int slpflag = 0;
1340 int slptimeo = 0;
1341 int error, error2;
1342
1343 /*
1344 * Commits are usually short and sweet so lets save some cpu and
1345 * leave the async daemons for more important rpc's (such as reads
1346 * and writes).
1347 *
1348 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1349 * in the directory in order to update attributes. This can deadlock
1350 * with another thread that is waiting for async I/O to be done by
1351 * an nfsiod thread while holding a lock on one of these vnodes.
1352 * To avoid this deadlock, don't allow the async nfsiod threads to
1353 * perform Readdirplus RPCs.
1354 */
1355 mtx_lock(&nfs_iod_mtx);
1356 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1357 (nmp->nm_bufqiods > nfs_numasync / 2)) ||
1358 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1359 mtx_unlock(&nfs_iod_mtx);
1360 return(EIO);
1361 }
1362 again:
1363 if (nmp->nm_flag & NFSMNT_INT)
1364 slpflag = NFS_PCATCH;
1365 gotiod = FALSE;
1366
1367 /*
1368 * Find a free iod to process this request.
1369 */
1370 for (iod = 0; iod < nfs_numasync; iod++)
1371 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1372 gotiod = TRUE;
1373 break;
1374 }
1375
1376 /*
1377 * Try to create one if none are free.
1378 */
1379 if (!gotiod)
1380 nfs_nfsiodnew();
1381 else {
1382 /*
1383 * Found one, so wake it up and tell it which
1384 * mount to process.
1385 */
1386 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1387 iod, nmp));
1388 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1389 nfs_iodmount[iod] = nmp;
1390 nmp->nm_bufqiods++;
1391 wakeup(&nfs_iodwant[iod]);
1392 }
1393
1394 /*
1395 * If none are free, we may already have an iod working on this mount
1396 * point. If so, it will process our request.
1397 */
1398 if (!gotiod) {
1399 if (nmp->nm_bufqiods > 0) {
1400 NFS_DPF(ASYNCIO,
1401 ("nfs_asyncio: %d iods are already processing mount %p\n",
1402 nmp->nm_bufqiods, nmp));
1403 gotiod = TRUE;
1404 }
1405 }
1406
1407 /*
1408 * If we have an iod which can process the request, then queue
1409 * the buffer.
1410 */
1411 if (gotiod) {
1412 /*
1413 * Ensure that the queue never grows too large. We still want
1414 * to asynchronize so we block rather then return EIO.
1415 */
1416 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1417 NFS_DPF(ASYNCIO,
1418 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1419 nmp->nm_bufqwant = TRUE;
1420 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1421 slpflag | PRIBIO,
1422 "nfsaio", slptimeo);
1423 if (error) {
1424 error2 = nfs_sigintr(nmp, td);
1425 if (error2) {
1426 mtx_unlock(&nfs_iod_mtx);
1427 return (error2);
1428 }
1429 if (slpflag == NFS_PCATCH) {
1430 slpflag = 0;
1431 slptimeo = 2 * hz;
1432 }
1433 }
1434 /*
1435 * We might have lost our iod while sleeping,
1436 * so check and loop if nescessary.
1437 */
1438 goto again;
1439 }
1440
1441 /* We might have lost our nfsiod */
1442 if (nmp->nm_bufqiods == 0) {
1443 NFS_DPF(ASYNCIO,
1444 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1445 goto again;
1446 }
1447
1448 if (bp->b_iocmd == BIO_READ) {
1449 if (bp->b_rcred == NOCRED && cred != NOCRED)
1450 bp->b_rcred = crhold(cred);
1451 } else {
1452 if (bp->b_wcred == NOCRED && cred != NOCRED)
1453 bp->b_wcred = crhold(cred);
1454 }
1455
1456 if (bp->b_flags & B_REMFREE)
1457 bremfreef(bp);
1458 BUF_KERNPROC(bp);
1459 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1460 nmp->nm_bufqlen++;
1461 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1462 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1463 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1464 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1465 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1466 }
1467 mtx_unlock(&nfs_iod_mtx);
1468 return (0);
1469 }
1470
1471 mtx_unlock(&nfs_iod_mtx);
1472
1473 /*
1474 * All the iods are busy on other mounts, so return EIO to
1475 * force the caller to process the i/o synchronously.
1476 */
1477 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1478 return (EIO);
1479 }
1480
1481 void
nfs_doio_directwrite(struct buf * bp)1482 nfs_doio_directwrite(struct buf *bp)
1483 {
1484 int iomode, must_commit;
1485 struct uio *uiop = (struct uio *)bp->b_caller1;
1486 char *iov_base = uiop->uio_iov->iov_base;
1487 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1488
1489 iomode = NFSV3WRITE_FILESYNC;
1490 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1491 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1492 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1493 free(iov_base, M_NFSDIRECTIO);
1494 free(uiop->uio_iov, M_NFSDIRECTIO);
1495 free(uiop, M_NFSDIRECTIO);
1496 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1497 struct nfsnode *np = VTONFS(bp->b_vp);
1498 mtx_lock(&np->n_mtx);
1499 np->n_directio_asyncwr--;
1500 if (np->n_directio_asyncwr == 0) {
1501 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1502 if ((np->n_flag & NFSYNCWAIT)) {
1503 np->n_flag &= ~NFSYNCWAIT;
1504 wakeup((caddr_t)&np->n_directio_asyncwr);
1505 }
1506 }
1507 mtx_unlock(&np->n_mtx);
1508 }
1509 bp->b_vp = NULL;
1510 relpbuf(bp, &nfs_pbuf_freecnt);
1511 }
1512
1513 /*
1514 * Do an I/O operation to/from a cache block. This may be called
1515 * synchronously or from an nfsiod.
1516 */
1517 int
nfs_doio(struct vnode * vp,struct buf * bp,struct ucred * cr,struct thread * td)1518 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1519 {
1520 struct uio *uiop;
1521 struct nfsnode *np;
1522 struct nfsmount *nmp;
1523 int error = 0, iomode, must_commit = 0;
1524 struct uio uio;
1525 struct iovec io;
1526 struct proc *p = td ? td->td_proc : NULL;
1527 uint8_t iocmd;
1528
1529 np = VTONFS(vp);
1530 nmp = VFSTONFS(vp->v_mount);
1531 uiop = &uio;
1532 uiop->uio_iov = &io;
1533 uiop->uio_iovcnt = 1;
1534 uiop->uio_segflg = UIO_SYSSPACE;
1535 uiop->uio_td = td;
1536
1537 /*
1538 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1539 * do this here so we do not have to do it in all the code that
1540 * calls us.
1541 */
1542 bp->b_flags &= ~B_INVAL;
1543 bp->b_ioflags &= ~BIO_ERROR;
1544
1545 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1546 iocmd = bp->b_iocmd;
1547 if (iocmd == BIO_READ) {
1548 io.iov_len = uiop->uio_resid = bp->b_bcount;
1549 io.iov_base = bp->b_data;
1550 uiop->uio_rw = UIO_READ;
1551
1552 switch (vp->v_type) {
1553 case VREG:
1554 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1555 nfsstats.read_bios++;
1556 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1557
1558 if (!error) {
1559 if (uiop->uio_resid) {
1560 /*
1561 * If we had a short read with no error, we must have
1562 * hit a file hole. We should zero-fill the remainder.
1563 * This can also occur if the server hits the file EOF.
1564 *
1565 * Holes used to be able to occur due to pending
1566 * writes, but that is not possible any longer.
1567 */
1568 int nread = bp->b_bcount - uiop->uio_resid;
1569 int left = uiop->uio_resid;
1570
1571 if (left > 0)
1572 bzero((char *)bp->b_data + nread, left);
1573 uiop->uio_resid = 0;
1574 }
1575 }
1576 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1577 if (p && (vp->v_vflag & VV_TEXT)) {
1578 mtx_lock(&np->n_mtx);
1579 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1580 mtx_unlock(&np->n_mtx);
1581 PROC_LOCK(p);
1582 killproc(p, "text file modification");
1583 PROC_UNLOCK(p);
1584 } else
1585 mtx_unlock(&np->n_mtx);
1586 }
1587 break;
1588 case VLNK:
1589 uiop->uio_offset = (off_t)0;
1590 nfsstats.readlink_bios++;
1591 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1592 break;
1593 case VDIR:
1594 nfsstats.readdir_bios++;
1595 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1596 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1597 error = nfs_readdirplusrpc(vp, uiop, cr);
1598 if (error == NFSERR_NOTSUPP)
1599 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1600 }
1601 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1602 error = nfs_readdirrpc(vp, uiop, cr);
1603 /*
1604 * end-of-directory sets B_INVAL but does not generate an
1605 * error.
1606 */
1607 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1608 bp->b_flags |= B_INVAL;
1609 break;
1610 default:
1611 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1612 break;
1613 };
1614 if (error) {
1615 bp->b_ioflags |= BIO_ERROR;
1616 bp->b_error = error;
1617 }
1618 } else {
1619 /*
1620 * If we only need to commit, try to commit
1621 */
1622 if (bp->b_flags & B_NEEDCOMMIT) {
1623 int retv;
1624 off_t off;
1625
1626 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1627 retv = (nmp->nm_rpcops->nr_commit)(
1628 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1629 bp->b_wcred, td);
1630 if (retv == 0) {
1631 bp->b_dirtyoff = bp->b_dirtyend = 0;
1632 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1633 bp->b_resid = 0;
1634 bufdone(bp);
1635 return (0);
1636 }
1637 if (retv == NFSERR_STALEWRITEVERF) {
1638 nfs_clearcommit(vp->v_mount);
1639 }
1640 }
1641
1642 /*
1643 * Setup for actual write
1644 */
1645 mtx_lock(&np->n_mtx);
1646 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1647 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1648 mtx_unlock(&np->n_mtx);
1649
1650 if (bp->b_dirtyend > bp->b_dirtyoff) {
1651 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1652 - bp->b_dirtyoff;
1653 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1654 + bp->b_dirtyoff;
1655 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1656 uiop->uio_rw = UIO_WRITE;
1657 nfsstats.write_bios++;
1658
1659 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1660 iomode = NFSV3WRITE_UNSTABLE;
1661 else
1662 iomode = NFSV3WRITE_FILESYNC;
1663
1664 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1665
1666 /*
1667 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1668 * to cluster the buffers needing commit. This will allow
1669 * the system to submit a single commit rpc for the whole
1670 * cluster. We can do this even if the buffer is not 100%
1671 * dirty (relative to the NFS blocksize), so we optimize the
1672 * append-to-file-case.
1673 *
1674 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1675 * cleared because write clustering only works for commit
1676 * rpc's, not for the data portion of the write).
1677 */
1678
1679 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1680 bp->b_flags |= B_NEEDCOMMIT;
1681 if (bp->b_dirtyoff == 0
1682 && bp->b_dirtyend == bp->b_bcount)
1683 bp->b_flags |= B_CLUSTEROK;
1684 } else {
1685 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1686 }
1687
1688 /*
1689 * For an interrupted write, the buffer is still valid
1690 * and the write hasn't been pushed to the server yet,
1691 * so we can't set BIO_ERROR and report the interruption
1692 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1693 * is not relevant, so the rpc attempt is essentially
1694 * a noop. For the case of a V3 write rpc not being
1695 * committed to stable storage, the block is still
1696 * dirty and requires either a commit rpc or another
1697 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1698 * the block is reused. This is indicated by setting
1699 * the B_DELWRI and B_NEEDCOMMIT flags.
1700 *
1701 * If the buffer is marked B_PAGING, it does not reside on
1702 * the vp's paging queues so we cannot call bdirty(). The
1703 * bp in this case is not an NFS cache block so we should
1704 * be safe. XXX
1705 *
1706 * The logic below breaks up errors into recoverable and
1707 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1708 * and keep the buffer around for potential write retries.
1709 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1710 * and save the error in the nfsnode. This is less than ideal
1711 * but necessary. Keeping such buffers around could potentially
1712 * cause buffer exhaustion eventually (they can never be written
1713 * out, so will get constantly be re-dirtied). It also causes
1714 * all sorts of vfs panics. For non-recoverable write errors,
1715 * also invalidate the attrcache, so we'll be forced to go over
1716 * the wire for this object, returning an error to user on next
1717 * call (most of the time).
1718 */
1719 if (error == EINTR || error == EIO || error == ETIMEDOUT
1720 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1721 int s;
1722
1723 s = splbio();
1724 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1725 if ((bp->b_flags & B_PAGING) == 0) {
1726 bdirty(bp);
1727 bp->b_flags &= ~B_DONE;
1728 }
1729 if (error && (bp->b_flags & B_ASYNC) == 0)
1730 bp->b_flags |= B_EINTR;
1731 splx(s);
1732 } else {
1733 if (error) {
1734 bp->b_ioflags |= BIO_ERROR;
1735 bp->b_flags |= B_INVAL;
1736 bp->b_error = np->n_error = error;
1737 mtx_lock(&np->n_mtx);
1738 np->n_flag |= NWRITEERR;
1739 np->n_attrstamp = 0;
1740 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1741 mtx_unlock(&np->n_mtx);
1742 }
1743 bp->b_dirtyoff = bp->b_dirtyend = 0;
1744 }
1745 } else {
1746 bp->b_resid = 0;
1747 bufdone(bp);
1748 return (0);
1749 }
1750 }
1751 bp->b_resid = uiop->uio_resid;
1752 if (must_commit)
1753 nfs_clearcommit(vp->v_mount);
1754 bufdone(bp);
1755 return (error);
1756 }
1757
1758 /*
1759 * Used to aid in handling ftruncate() operations on the NFS client side.
1760 * Truncation creates a number of special problems for NFS. We have to
1761 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1762 * we have to properly handle VM pages or (potentially dirty) buffers
1763 * that straddle the truncation point.
1764 */
1765
1766 int
nfs_meta_setsize(struct vnode * vp,struct ucred * cred,struct thread * td,u_quad_t nsize)1767 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1768 {
1769 struct nfsnode *np = VTONFS(vp);
1770 u_quad_t tsize;
1771 int biosize = vp->v_bufobj.bo_bsize;
1772 int error = 0;
1773
1774 mtx_lock(&np->n_mtx);
1775 tsize = np->n_size;
1776 np->n_size = nsize;
1777 mtx_unlock(&np->n_mtx);
1778
1779 if (nsize < tsize) {
1780 struct buf *bp;
1781 daddr_t lbn;
1782 int bufsize;
1783
1784 /*
1785 * vtruncbuf() doesn't get the buffer overlapping the
1786 * truncation point. We may have a B_DELWRI and/or B_CACHE
1787 * buffer that now needs to be truncated.
1788 */
1789 error = vtruncbuf(vp, cred, td, nsize, biosize);
1790 lbn = nsize / biosize;
1791 bufsize = nsize & (biosize - 1);
1792 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1793 if (!bp)
1794 return EINTR;
1795 if (bp->b_dirtyoff > bp->b_bcount)
1796 bp->b_dirtyoff = bp->b_bcount;
1797 if (bp->b_dirtyend > bp->b_bcount)
1798 bp->b_dirtyend = bp->b_bcount;
1799 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1800 brelse(bp);
1801 } else {
1802 vnode_pager_setsize(vp, nsize);
1803 }
1804 return(error);
1805 }
1806
1807