1 /*-
2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
33 */
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD: releng/8.3/sys/fs/nfsclient/nfs_clbio.c 231332 2012-02-10 04:01:17Z rmacklem $");
37
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/bio.h>
41 #include <sys/buf.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/vmmeter.h>
45 #include <sys/vnode.h>
46
47 #include <vm/vm.h>
48 #include <vm/vm_extern.h>
49 #include <vm/vm_page.h>
50 #include <vm/vm_object.h>
51 #include <vm/vm_pager.h>
52 #include <vm/vnode_pager.h>
53
54 #include <fs/nfs/nfsport.h>
55 #include <fs/nfsclient/nfsmount.h>
56 #include <fs/nfsclient/nfs.h>
57 #include <fs/nfsclient/nfsnode.h>
58
59 extern int newnfs_directio_allow_mmap;
60 extern struct nfsstats newnfsstats;
61 extern struct mtx ncl_iod_mutex;
62 extern int ncl_numasync;
63 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
64 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
65 extern int newnfs_directio_enable;
66
67 int ncl_pbuf_freecnt = -1; /* start out unlimited */
68
69 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
70 struct thread *td);
71 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
72 struct ucred *cred, int ioflag);
73
74 /*
75 * Vnode op for VM getpages.
76 */
77 int
78 ncl_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 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
103 return (VM_PAGER_ERROR);
104 }
105
106 if (newnfs_directio_enable && !newnfs_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 ncl_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)ncl_fsinfo(nmp, vp, cred, td);
122 } else
123 mtx_unlock(&nmp->nm_mtx);
124
125 npages = btoc(count);
126
127 /*
128 * If the requested page is partially valid, just return it and
129 * allow the pager to zero-out the blanks. Partially valid pages
130 * can only occur at the file EOF.
131 */
132 VM_OBJECT_LOCK(object);
133 if (pages[ap->a_reqpage]->valid != 0) {
134 vm_page_lock_queues();
135 for (i = 0; i < npages; ++i) {
136 if (i != ap->a_reqpage)
137 vm_page_free(pages[i]);
138 }
139 vm_page_unlock_queues();
140 VM_OBJECT_UNLOCK(object);
141 return (0);
142 }
143 VM_OBJECT_UNLOCK(object);
144
145 /*
146 * We use only the kva address for the buffer, but this is extremely
147 * convienient and fast.
148 */
149 bp = getpbuf(&ncl_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 = ncl_readrpc(vp, &uio, cred);
167 pmap_qremove(kva, npages);
168
169 relpbuf(bp, &ncl_pbuf_freecnt);
170
171 if (error && (uio.uio_resid == count)) {
172 ncl_printf("nfs_getpages: error %d\n", error);
173 VM_OBJECT_LOCK(object);
174 vm_page_lock_queues();
175 for (i = 0; i < npages; ++i) {
176 if (i != ap->a_reqpage)
177 vm_page_free(pages[i]);
178 }
179 vm_page_unlock_queues();
180 VM_OBJECT_UNLOCK(object);
181 return (VM_PAGER_ERROR);
182 }
183
184 /*
185 * Calculate the number of bytes read and validate only that number
186 * of bytes. Note that due to pending writes, size may be 0. This
187 * does not mean that the remaining data is invalid!
188 */
189
190 size = count - uio.uio_resid;
191 VM_OBJECT_LOCK(object);
192 vm_page_lock_queues();
193 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
194 vm_page_t m;
195 nextoff = toff + PAGE_SIZE;
196 m = pages[i];
197
198 if (nextoff <= size) {
199 /*
200 * Read operation filled an entire page
201 */
202 m->valid = VM_PAGE_BITS_ALL;
203 KASSERT(m->dirty == 0,
204 ("nfs_getpages: page %p is dirty", m));
205 } else if (size > toff) {
206 /*
207 * Read operation filled a partial page.
208 */
209 m->valid = 0;
210 vm_page_set_valid(m, 0, size - toff);
211 KASSERT(m->dirty == 0,
212 ("nfs_getpages: page %p is dirty", m));
213 } else {
214 /*
215 * Read operation was short. If no error occured
216 * we may have hit a zero-fill section. We simply
217 * leave valid set to 0.
218 */
219 ;
220 }
221 if (i != ap->a_reqpage) {
222 /*
223 * Whether or not to leave the page activated is up in
224 * the air, but we should put the page on a page queue
225 * somewhere (it already is in the object). Result:
226 * It appears that emperical results show that
227 * deactivating pages is best.
228 */
229
230 /*
231 * Just in case someone was asking for this page we
232 * now tell them that it is ok to use.
233 */
234 if (!error) {
235 if (m->oflags & VPO_WANTED)
236 vm_page_activate(m);
237 else
238 vm_page_deactivate(m);
239 vm_page_wakeup(m);
240 } else {
241 vm_page_free(m);
242 }
243 }
244 }
245 vm_page_unlock_queues();
246 VM_OBJECT_UNLOCK(object);
247 return (0);
248 }
249
250 /*
251 * Vnode op for VM putpages.
252 */
253 int
254 ncl_putpages(struct vop_putpages_args *ap)
255 {
256 struct uio uio;
257 struct iovec iov;
258 vm_offset_t kva;
259 struct buf *bp;
260 int iomode, must_commit, i, error, npages, count;
261 off_t offset;
262 int *rtvals;
263 struct vnode *vp;
264 struct thread *td;
265 struct ucred *cred;
266 struct nfsmount *nmp;
267 struct nfsnode *np;
268 vm_page_t *pages;
269
270 vp = ap->a_vp;
271 np = VTONFS(vp);
272 td = curthread; /* XXX */
273 cred = curthread->td_ucred; /* XXX */
274 nmp = VFSTONFS(vp->v_mount);
275 pages = ap->a_m;
276 count = ap->a_count;
277 rtvals = ap->a_rtvals;
278 npages = btoc(count);
279 offset = IDX_TO_OFF(pages[0]->pindex);
280
281 mtx_lock(&nmp->nm_mtx);
282 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
283 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
284 mtx_unlock(&nmp->nm_mtx);
285 (void)ncl_fsinfo(nmp, vp, cred, td);
286 } else
287 mtx_unlock(&nmp->nm_mtx);
288
289 mtx_lock(&np->n_mtx);
290 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
291 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
292 mtx_unlock(&np->n_mtx);
293 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
294 mtx_lock(&np->n_mtx);
295 }
296
297 for (i = 0; i < npages; i++)
298 rtvals[i] = VM_PAGER_ERROR;
299
300 /*
301 * When putting pages, do not extend file past EOF.
302 */
303 if (offset + count > np->n_size) {
304 count = np->n_size - offset;
305 if (count < 0)
306 count = 0;
307 }
308 mtx_unlock(&np->n_mtx);
309
310 /*
311 * We use only the kva address for the buffer, but this is extremely
312 * convienient and fast.
313 */
314 bp = getpbuf(&ncl_pbuf_freecnt);
315
316 kva = (vm_offset_t) bp->b_data;
317 pmap_qenter(kva, pages, npages);
318 PCPU_INC(cnt.v_vnodeout);
319 PCPU_ADD(cnt.v_vnodepgsout, count);
320
321 iov.iov_base = (caddr_t) kva;
322 iov.iov_len = count;
323 uio.uio_iov = &iov;
324 uio.uio_iovcnt = 1;
325 uio.uio_offset = offset;
326 uio.uio_resid = count;
327 uio.uio_segflg = UIO_SYSSPACE;
328 uio.uio_rw = UIO_WRITE;
329 uio.uio_td = td;
330
331 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
332 iomode = NFSWRITE_UNSTABLE;
333 else
334 iomode = NFSWRITE_FILESYNC;
335
336 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
337
338 pmap_qremove(kva, npages);
339 relpbuf(bp, &ncl_pbuf_freecnt);
340
341 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
342 if (must_commit)
343 ncl_clearcommit(vp->v_mount);
344 return rtvals[0];
345 }
346
347 /*
348 * For nfs, cache consistency can only be maintained approximately.
349 * Although RFC1094 does not specify the criteria, the following is
350 * believed to be compatible with the reference port.
351 * For nfs:
352 * If the file's modify time on the server has changed since the
353 * last read rpc or you have written to the file,
354 * you may have lost data cache consistency with the
355 * server, so flush all of the file's data out of the cache.
356 * Then force a getattr rpc to ensure that you have up to date
357 * attributes.
358 * NB: This implies that cache data can be read when up to
359 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
360 * attributes this could be forced by setting n_attrstamp to 0 before
361 * the VOP_GETATTR() call.
362 */
363 static inline int
364 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
365 {
366 int error = 0;
367 struct vattr vattr;
368 struct nfsnode *np = VTONFS(vp);
369 int old_lock;
370
371 /*
372 * Grab the exclusive lock before checking whether the cache is
373 * consistent.
374 * XXX - We can make this cheaper later (by acquiring cheaper locks).
375 * But for now, this suffices.
376 */
377 old_lock = ncl_upgrade_vnlock(vp);
378 if (vp->v_iflag & VI_DOOMED) {
379 ncl_downgrade_vnlock(vp, old_lock);
380 return (EBADF);
381 }
382
383 mtx_lock(&np->n_mtx);
384 if (np->n_flag & NMODIFIED) {
385 mtx_unlock(&np->n_mtx);
386 if (vp->v_type != VREG) {
387 if (vp->v_type != VDIR)
388 panic("nfs: bioread, not dir");
389 ncl_invaldir(vp);
390 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
391 if (error)
392 goto out;
393 }
394 np->n_attrstamp = 0;
395 error = VOP_GETATTR(vp, &vattr, cred);
396 if (error)
397 goto out;
398 mtx_lock(&np->n_mtx);
399 np->n_mtime = vattr.va_mtime;
400 mtx_unlock(&np->n_mtx);
401 } else {
402 mtx_unlock(&np->n_mtx);
403 error = VOP_GETATTR(vp, &vattr, cred);
404 if (error)
405 return (error);
406 mtx_lock(&np->n_mtx);
407 if ((np->n_flag & NSIZECHANGED)
408 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
409 mtx_unlock(&np->n_mtx);
410 if (vp->v_type == VDIR)
411 ncl_invaldir(vp);
412 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
413 if (error)
414 goto out;
415 mtx_lock(&np->n_mtx);
416 np->n_mtime = vattr.va_mtime;
417 np->n_flag &= ~NSIZECHANGED;
418 }
419 mtx_unlock(&np->n_mtx);
420 }
421 out:
422 ncl_downgrade_vnlock(vp, old_lock);
423 return error;
424 }
425
426 /*
427 * Vnode op for read using bio
428 */
429 int
430 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
431 {
432 struct nfsnode *np = VTONFS(vp);
433 int biosize, i;
434 struct buf *bp, *rabp;
435 struct thread *td;
436 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
437 daddr_t lbn, rabn;
438 int bcount;
439 int seqcount;
440 int nra, error = 0, n = 0, on = 0;
441 off_t tmp_off;
442
443 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
444 if (uio->uio_resid == 0)
445 return (0);
446 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
447 return (EINVAL);
448 td = uio->uio_td;
449
450 mtx_lock(&nmp->nm_mtx);
451 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
452 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
453 mtx_unlock(&nmp->nm_mtx);
454 (void)ncl_fsinfo(nmp, vp, cred, td);
455 mtx_lock(&nmp->nm_mtx);
456 }
457 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
458 (void) newnfs_iosize(nmp);
459
460 tmp_off = uio->uio_offset + uio->uio_resid;
461 if (vp->v_type != VDIR &&
462 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
463 mtx_unlock(&nmp->nm_mtx);
464 return (EFBIG);
465 }
466 mtx_unlock(&nmp->nm_mtx);
467
468 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
469 /* No caching/ no readaheads. Just read data into the user buffer */
470 return ncl_readrpc(vp, uio, cred);
471
472 biosize = vp->v_bufobj.bo_bsize;
473 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
474
475 error = nfs_bioread_check_cons(vp, td, cred);
476 if (error)
477 return error;
478
479 do {
480 u_quad_t nsize;
481
482 mtx_lock(&np->n_mtx);
483 nsize = np->n_size;
484 mtx_unlock(&np->n_mtx);
485
486 switch (vp->v_type) {
487 case VREG:
488 NFSINCRGLOBAL(newnfsstats.biocache_reads);
489 lbn = uio->uio_offset / biosize;
490 on = uio->uio_offset & (biosize - 1);
491
492 /*
493 * Start the read ahead(s), as required.
494 */
495 if (nmp->nm_readahead > 0) {
496 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
497 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
498 rabn = lbn + 1 + nra;
499 if (incore(&vp->v_bufobj, rabn) == NULL) {
500 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
501 if (!rabp) {
502 error = newnfs_sigintr(nmp, td);
503 return (error ? error : EINTR);
504 }
505 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
506 rabp->b_flags |= B_ASYNC;
507 rabp->b_iocmd = BIO_READ;
508 vfs_busy_pages(rabp, 0);
509 if (ncl_asyncio(nmp, rabp, cred, td)) {
510 rabp->b_flags |= B_INVAL;
511 rabp->b_ioflags |= BIO_ERROR;
512 vfs_unbusy_pages(rabp);
513 brelse(rabp);
514 break;
515 }
516 } else {
517 brelse(rabp);
518 }
519 }
520 }
521 }
522
523 /* Note that bcount is *not* DEV_BSIZE aligned. */
524 bcount = biosize;
525 if ((off_t)lbn * biosize >= nsize) {
526 bcount = 0;
527 } else if ((off_t)(lbn + 1) * biosize > nsize) {
528 bcount = nsize - (off_t)lbn * biosize;
529 }
530 bp = nfs_getcacheblk(vp, lbn, bcount, td);
531
532 if (!bp) {
533 error = newnfs_sigintr(nmp, td);
534 return (error ? error : EINTR);
535 }
536
537 /*
538 * If B_CACHE is not set, we must issue the read. If this
539 * fails, we return an error.
540 */
541
542 if ((bp->b_flags & B_CACHE) == 0) {
543 bp->b_iocmd = BIO_READ;
544 vfs_busy_pages(bp, 0);
545 error = ncl_doio(vp, bp, cred, td, 0);
546 if (error) {
547 brelse(bp);
548 return (error);
549 }
550 }
551
552 /*
553 * on is the offset into the current bp. Figure out how many
554 * bytes we can copy out of the bp. Note that bcount is
555 * NOT DEV_BSIZE aligned.
556 *
557 * Then figure out how many bytes we can copy into the uio.
558 */
559
560 n = 0;
561 if (on < bcount)
562 n = min((unsigned)(bcount - on), uio->uio_resid);
563 break;
564 case VLNK:
565 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
566 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
567 if (!bp) {
568 error = newnfs_sigintr(nmp, td);
569 return (error ? error : EINTR);
570 }
571 if ((bp->b_flags & B_CACHE) == 0) {
572 bp->b_iocmd = BIO_READ;
573 vfs_busy_pages(bp, 0);
574 error = ncl_doio(vp, bp, cred, td, 0);
575 if (error) {
576 bp->b_ioflags |= BIO_ERROR;
577 brelse(bp);
578 return (error);
579 }
580 }
581 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
582 on = 0;
583 break;
584 case VDIR:
585 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
586 if (np->n_direofoffset
587 && uio->uio_offset >= np->n_direofoffset) {
588 return (0);
589 }
590 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
591 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
592 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
593 if (!bp) {
594 error = newnfs_sigintr(nmp, td);
595 return (error ? error : EINTR);
596 }
597 if ((bp->b_flags & B_CACHE) == 0) {
598 bp->b_iocmd = BIO_READ;
599 vfs_busy_pages(bp, 0);
600 error = ncl_doio(vp, bp, cred, td, 0);
601 if (error) {
602 brelse(bp);
603 }
604 while (error == NFSERR_BAD_COOKIE) {
605 ncl_invaldir(vp);
606 error = ncl_vinvalbuf(vp, 0, td, 1);
607 /*
608 * Yuck! The directory has been modified on the
609 * server. The only way to get the block is by
610 * reading from the beginning to get all the
611 * offset cookies.
612 *
613 * Leave the last bp intact unless there is an error.
614 * Loop back up to the while if the error is another
615 * NFSERR_BAD_COOKIE (double yuch!).
616 */
617 for (i = 0; i <= lbn && !error; i++) {
618 if (np->n_direofoffset
619 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
620 return (0);
621 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
622 if (!bp) {
623 error = newnfs_sigintr(nmp, td);
624 return (error ? error : EINTR);
625 }
626 if ((bp->b_flags & B_CACHE) == 0) {
627 bp->b_iocmd = BIO_READ;
628 vfs_busy_pages(bp, 0);
629 error = ncl_doio(vp, bp, cred, td, 0);
630 /*
631 * no error + B_INVAL == directory EOF,
632 * use the block.
633 */
634 if (error == 0 && (bp->b_flags & B_INVAL))
635 break;
636 }
637 /*
638 * An error will throw away the block and the
639 * for loop will break out. If no error and this
640 * is not the block we want, we throw away the
641 * block and go for the next one via the for loop.
642 */
643 if (error || i < lbn)
644 brelse(bp);
645 }
646 }
647 /*
648 * The above while is repeated if we hit another cookie
649 * error. If we hit an error and it wasn't a cookie error,
650 * we give up.
651 */
652 if (error)
653 return (error);
654 }
655
656 /*
657 * If not eof and read aheads are enabled, start one.
658 * (You need the current block first, so that you have the
659 * directory offset cookie of the next block.)
660 */
661 if (nmp->nm_readahead > 0 &&
662 (bp->b_flags & B_INVAL) == 0 &&
663 (np->n_direofoffset == 0 ||
664 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
665 incore(&vp->v_bufobj, lbn + 1) == NULL) {
666 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
667 if (rabp) {
668 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
669 rabp->b_flags |= B_ASYNC;
670 rabp->b_iocmd = BIO_READ;
671 vfs_busy_pages(rabp, 0);
672 if (ncl_asyncio(nmp, rabp, cred, td)) {
673 rabp->b_flags |= B_INVAL;
674 rabp->b_ioflags |= BIO_ERROR;
675 vfs_unbusy_pages(rabp);
676 brelse(rabp);
677 }
678 } else {
679 brelse(rabp);
680 }
681 }
682 }
683 /*
684 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
685 * chopped for the EOF condition, we cannot tell how large
686 * NFS directories are going to be until we hit EOF. So
687 * an NFS directory buffer is *not* chopped to its EOF. Now,
688 * it just so happens that b_resid will effectively chop it
689 * to EOF. *BUT* this information is lost if the buffer goes
690 * away and is reconstituted into a B_CACHE state ( due to
691 * being VMIO ) later. So we keep track of the directory eof
692 * in np->n_direofoffset and chop it off as an extra step
693 * right here.
694 */
695 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
696 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
697 n = np->n_direofoffset - uio->uio_offset;
698 break;
699 default:
700 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
701 bp = NULL;
702 break;
703 };
704
705 if (n > 0) {
706 error = uiomove(bp->b_data + on, (int)n, uio);
707 }
708 if (vp->v_type == VLNK)
709 n = 0;
710 if (bp != NULL)
711 brelse(bp);
712 } while (error == 0 && uio->uio_resid > 0 && n > 0);
713 return (error);
714 }
715
716 /*
717 * The NFS write path cannot handle iovecs with len > 1. So we need to
718 * break up iovecs accordingly (restricting them to wsize).
719 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
720 * For the ASYNC case, 2 copies are needed. The first a copy from the
721 * user buffer to a staging buffer and then a second copy from the staging
722 * buffer to mbufs. This can be optimized by copying from the user buffer
723 * directly into mbufs and passing the chain down, but that requires a
724 * fair amount of re-working of the relevant codepaths (and can be done
725 * later).
726 */
727 static int
728 nfs_directio_write(vp, uiop, cred, ioflag)
729 struct vnode *vp;
730 struct uio *uiop;
731 struct ucred *cred;
732 int ioflag;
733 {
734 int error;
735 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
736 struct thread *td = uiop->uio_td;
737 int size;
738 int wsize;
739
740 mtx_lock(&nmp->nm_mtx);
741 wsize = nmp->nm_wsize;
742 mtx_unlock(&nmp->nm_mtx);
743 if (ioflag & IO_SYNC) {
744 int iomode, must_commit;
745 struct uio uio;
746 struct iovec iov;
747 do_sync:
748 while (uiop->uio_resid > 0) {
749 size = min(uiop->uio_resid, wsize);
750 size = min(uiop->uio_iov->iov_len, size);
751 iov.iov_base = uiop->uio_iov->iov_base;
752 iov.iov_len = size;
753 uio.uio_iov = &iov;
754 uio.uio_iovcnt = 1;
755 uio.uio_offset = uiop->uio_offset;
756 uio.uio_resid = size;
757 uio.uio_segflg = UIO_USERSPACE;
758 uio.uio_rw = UIO_WRITE;
759 uio.uio_td = td;
760 iomode = NFSWRITE_FILESYNC;
761 error = ncl_writerpc(vp, &uio, cred, &iomode,
762 &must_commit, 0);
763 KASSERT((must_commit == 0),
764 ("ncl_directio_write: Did not commit write"));
765 if (error)
766 return (error);
767 uiop->uio_offset += size;
768 uiop->uio_resid -= size;
769 if (uiop->uio_iov->iov_len <= size) {
770 uiop->uio_iovcnt--;
771 uiop->uio_iov++;
772 } else {
773 uiop->uio_iov->iov_base =
774 (char *)uiop->uio_iov->iov_base + size;
775 uiop->uio_iov->iov_len -= size;
776 }
777 }
778 } else {
779 struct uio *t_uio;
780 struct iovec *t_iov;
781 struct buf *bp;
782
783 /*
784 * Break up the write into blocksize chunks and hand these
785 * over to nfsiod's for write back.
786 * Unfortunately, this incurs a copy of the data. Since
787 * the user could modify the buffer before the write is
788 * initiated.
789 *
790 * The obvious optimization here is that one of the 2 copies
791 * in the async write path can be eliminated by copying the
792 * data here directly into mbufs and passing the mbuf chain
793 * down. But that will require a fair amount of re-working
794 * of the code and can be done if there's enough interest
795 * in NFS directio access.
796 */
797 while (uiop->uio_resid > 0) {
798 size = min(uiop->uio_resid, wsize);
799 size = min(uiop->uio_iov->iov_len, size);
800 bp = getpbuf(&ncl_pbuf_freecnt);
801 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
802 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
803 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
804 t_iov->iov_len = size;
805 t_uio->uio_iov = t_iov;
806 t_uio->uio_iovcnt = 1;
807 t_uio->uio_offset = uiop->uio_offset;
808 t_uio->uio_resid = size;
809 t_uio->uio_segflg = UIO_SYSSPACE;
810 t_uio->uio_rw = UIO_WRITE;
811 t_uio->uio_td = td;
812 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size);
813 bp->b_flags |= B_DIRECT;
814 bp->b_iocmd = BIO_WRITE;
815 if (cred != NOCRED) {
816 crhold(cred);
817 bp->b_wcred = cred;
818 } else
819 bp->b_wcred = NOCRED;
820 bp->b_caller1 = (void *)t_uio;
821 bp->b_vp = vp;
822 error = ncl_asyncio(nmp, bp, NOCRED, td);
823 if (error) {
824 free(t_iov->iov_base, M_NFSDIRECTIO);
825 free(t_iov, M_NFSDIRECTIO);
826 free(t_uio, M_NFSDIRECTIO);
827 bp->b_vp = NULL;
828 relpbuf(bp, &ncl_pbuf_freecnt);
829 if (error == EINTR)
830 return (error);
831 goto do_sync;
832 }
833 uiop->uio_offset += size;
834 uiop->uio_resid -= size;
835 if (uiop->uio_iov->iov_len <= size) {
836 uiop->uio_iovcnt--;
837 uiop->uio_iov++;
838 } else {
839 uiop->uio_iov->iov_base =
840 (char *)uiop->uio_iov->iov_base + size;
841 uiop->uio_iov->iov_len -= size;
842 }
843 }
844 }
845 return (0);
846 }
847
848 /*
849 * Vnode op for write using bio
850 */
851 int
852 ncl_write(struct vop_write_args *ap)
853 {
854 int biosize;
855 struct uio *uio = ap->a_uio;
856 struct thread *td = uio->uio_td;
857 struct vnode *vp = ap->a_vp;
858 struct nfsnode *np = VTONFS(vp);
859 struct ucred *cred = ap->a_cred;
860 int ioflag = ap->a_ioflag;
861 struct buf *bp;
862 struct vattr vattr;
863 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
864 daddr_t lbn;
865 int bcount;
866 int n, on, error = 0;
867 off_t tmp_off;
868
869 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
870 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
871 ("ncl_write proc"));
872 if (vp->v_type != VREG)
873 return (EIO);
874 mtx_lock(&np->n_mtx);
875 if (np->n_flag & NWRITEERR) {
876 np->n_flag &= ~NWRITEERR;
877 mtx_unlock(&np->n_mtx);
878 return (np->n_error);
879 } else
880 mtx_unlock(&np->n_mtx);
881 mtx_lock(&nmp->nm_mtx);
882 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
883 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
884 mtx_unlock(&nmp->nm_mtx);
885 (void)ncl_fsinfo(nmp, vp, cred, td);
886 mtx_lock(&nmp->nm_mtx);
887 }
888 if (nmp->nm_wsize == 0)
889 (void) newnfs_iosize(nmp);
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 error = ncl_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 error = VOP_GETATTR(vp, &vattr, cred);
925 if (error)
926 return (error);
927 mtx_lock(&np->n_mtx);
928 uio->uio_offset = np->n_size;
929 mtx_unlock(&np->n_mtx);
930 }
931
932 if (uio->uio_offset < 0)
933 return (EINVAL);
934 tmp_off = uio->uio_offset + uio->uio_resid;
935 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
936 return (EFBIG);
937 if (uio->uio_resid == 0)
938 return (0);
939
940 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
941 return nfs_directio_write(vp, uio, cred, ioflag);
942
943 /*
944 * Maybe this should be above the vnode op call, but so long as
945 * file servers have no limits, i don't think it matters
946 */
947 if (vn_rlimit_fsize(vp, uio, td))
948 return (EFBIG);
949
950 biosize = vp->v_bufobj.bo_bsize;
951 /*
952 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
953 * would exceed the local maximum per-file write commit size when
954 * combined with those, we must decide whether to flush,
955 * go synchronous, or return error. We don't bother checking
956 * IO_UNIT -- we just make all writes atomic anyway, as there's
957 * no point optimizing for something that really won't ever happen.
958 */
959 if (!(ioflag & IO_SYNC)) {
960 int nflag;
961
962 mtx_lock(&np->n_mtx);
963 nflag = np->n_flag;
964 mtx_unlock(&np->n_mtx);
965 int needrestart = 0;
966 if (nmp->nm_wcommitsize < uio->uio_resid) {
967 /*
968 * If this request could not possibly be completed
969 * without exceeding the maximum outstanding write
970 * commit size, see if we can convert it into a
971 * synchronous write operation.
972 */
973 if (ioflag & IO_NDELAY)
974 return (EAGAIN);
975 ioflag |= IO_SYNC;
976 if (nflag & NMODIFIED)
977 needrestart = 1;
978 } else if (nflag & NMODIFIED) {
979 int wouldcommit = 0;
980 BO_LOCK(&vp->v_bufobj);
981 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
982 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
983 b_bobufs) {
984 if (bp->b_flags & B_NEEDCOMMIT)
985 wouldcommit += bp->b_bcount;
986 }
987 }
988 BO_UNLOCK(&vp->v_bufobj);
989 /*
990 * Since we're not operating synchronously and
991 * bypassing the buffer cache, we are in a commit
992 * and holding all of these buffers whether
993 * transmitted or not. If not limited, this
994 * will lead to the buffer cache deadlocking,
995 * as no one else can flush our uncommitted buffers.
996 */
997 wouldcommit += uio->uio_resid;
998 /*
999 * If we would initially exceed the maximum
1000 * outstanding write commit size, flush and restart.
1001 */
1002 if (wouldcommit > nmp->nm_wcommitsize)
1003 needrestart = 1;
1004 }
1005 if (needrestart)
1006 goto flush_and_restart;
1007 }
1008
1009 do {
1010 NFSINCRGLOBAL(newnfsstats.biocache_writes);
1011 lbn = uio->uio_offset / biosize;
1012 on = uio->uio_offset & (biosize-1);
1013 n = min((unsigned)(biosize - on), uio->uio_resid);
1014 again:
1015 /*
1016 * Handle direct append and file extension cases, calculate
1017 * unaligned buffer size.
1018 */
1019 mtx_lock(&np->n_mtx);
1020 if (uio->uio_offset == np->n_size && n) {
1021 mtx_unlock(&np->n_mtx);
1022 /*
1023 * Get the buffer (in its pre-append state to maintain
1024 * B_CACHE if it was previously set). Resize the
1025 * nfsnode after we have locked the buffer to prevent
1026 * readers from reading garbage.
1027 */
1028 bcount = on;
1029 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1030
1031 if (bp != NULL) {
1032 long save;
1033
1034 mtx_lock(&np->n_mtx);
1035 np->n_size = uio->uio_offset + n;
1036 np->n_flag |= NMODIFIED;
1037 vnode_pager_setsize(vp, np->n_size);
1038 mtx_unlock(&np->n_mtx);
1039
1040 save = bp->b_flags & B_CACHE;
1041 bcount += n;
1042 allocbuf(bp, bcount);
1043 bp->b_flags |= save;
1044 }
1045 } else {
1046 /*
1047 * Obtain the locked cache block first, and then
1048 * adjust the file's size as appropriate.
1049 */
1050 bcount = on + n;
1051 if ((off_t)lbn * biosize + bcount < np->n_size) {
1052 if ((off_t)(lbn + 1) * biosize < np->n_size)
1053 bcount = biosize;
1054 else
1055 bcount = np->n_size - (off_t)lbn * biosize;
1056 }
1057 mtx_unlock(&np->n_mtx);
1058 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1059 mtx_lock(&np->n_mtx);
1060 if (uio->uio_offset + n > np->n_size) {
1061 np->n_size = uio->uio_offset + n;
1062 np->n_flag |= NMODIFIED;
1063 vnode_pager_setsize(vp, np->n_size);
1064 }
1065 mtx_unlock(&np->n_mtx);
1066 }
1067
1068 if (!bp) {
1069 error = newnfs_sigintr(nmp, td);
1070 if (!error)
1071 error = EINTR;
1072 break;
1073 }
1074
1075 /*
1076 * Issue a READ if B_CACHE is not set. In special-append
1077 * mode, B_CACHE is based on the buffer prior to the write
1078 * op and is typically set, avoiding the read. If a read
1079 * is required in special append mode, the server will
1080 * probably send us a short-read since we extended the file
1081 * on our end, resulting in b_resid == 0 and, thusly,
1082 * B_CACHE getting set.
1083 *
1084 * We can also avoid issuing the read if the write covers
1085 * the entire buffer. We have to make sure the buffer state
1086 * is reasonable in this case since we will not be initiating
1087 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1088 * more information.
1089 *
1090 * B_CACHE may also be set due to the buffer being cached
1091 * normally.
1092 */
1093
1094 if (on == 0 && n == bcount) {
1095 bp->b_flags |= B_CACHE;
1096 bp->b_flags &= ~B_INVAL;
1097 bp->b_ioflags &= ~BIO_ERROR;
1098 }
1099
1100 if ((bp->b_flags & B_CACHE) == 0) {
1101 bp->b_iocmd = BIO_READ;
1102 vfs_busy_pages(bp, 0);
1103 error = ncl_doio(vp, bp, cred, td, 0);
1104 if (error) {
1105 brelse(bp);
1106 break;
1107 }
1108 }
1109 if (bp->b_wcred == NOCRED)
1110 bp->b_wcred = crhold(cred);
1111 mtx_lock(&np->n_mtx);
1112 np->n_flag |= NMODIFIED;
1113 mtx_unlock(&np->n_mtx);
1114
1115 /*
1116 * If dirtyend exceeds file size, chop it down. This should
1117 * not normally occur but there is an append race where it
1118 * might occur XXX, so we log it.
1119 *
1120 * If the chopping creates a reverse-indexed or degenerate
1121 * situation with dirtyoff/end, we 0 both of them.
1122 */
1123
1124 if (bp->b_dirtyend > bcount) {
1125 ncl_printf("NFS append race @%lx:%d\n",
1126 (long)bp->b_blkno * DEV_BSIZE,
1127 bp->b_dirtyend - bcount);
1128 bp->b_dirtyend = bcount;
1129 }
1130
1131 if (bp->b_dirtyoff >= bp->b_dirtyend)
1132 bp->b_dirtyoff = bp->b_dirtyend = 0;
1133
1134 /*
1135 * If the new write will leave a contiguous dirty
1136 * area, just update the b_dirtyoff and b_dirtyend,
1137 * otherwise force a write rpc of the old dirty area.
1138 *
1139 * While it is possible to merge discontiguous writes due to
1140 * our having a B_CACHE buffer ( and thus valid read data
1141 * for the hole), we don't because it could lead to
1142 * significant cache coherency problems with multiple clients,
1143 * especially if locking is implemented later on.
1144 *
1145 * as an optimization we could theoretically maintain
1146 * a linked list of discontinuous areas, but we would still
1147 * have to commit them separately so there isn't much
1148 * advantage to it except perhaps a bit of asynchronization.
1149 */
1150
1151 if (bp->b_dirtyend > 0 &&
1152 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1153 if (bwrite(bp) == EINTR) {
1154 error = EINTR;
1155 break;
1156 }
1157 goto again;
1158 }
1159
1160 error = uiomove((char *)bp->b_data + on, n, uio);
1161
1162 /*
1163 * Since this block is being modified, it must be written
1164 * again and not just committed. Since write clustering does
1165 * not work for the stage 1 data write, only the stage 2
1166 * commit rpc, we have to clear B_CLUSTEROK as well.
1167 */
1168 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1169
1170 if (error) {
1171 bp->b_ioflags |= BIO_ERROR;
1172 brelse(bp);
1173 break;
1174 }
1175
1176 /*
1177 * Only update dirtyoff/dirtyend if not a degenerate
1178 * condition.
1179 */
1180 if (n) {
1181 if (bp->b_dirtyend > 0) {
1182 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1183 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1184 } else {
1185 bp->b_dirtyoff = on;
1186 bp->b_dirtyend = on + n;
1187 }
1188 vfs_bio_set_valid(bp, on, n);
1189 }
1190
1191 /*
1192 * If IO_SYNC do bwrite().
1193 *
1194 * IO_INVAL appears to be unused. The idea appears to be
1195 * to turn off caching in this case. Very odd. XXX
1196 */
1197 if ((ioflag & IO_SYNC)) {
1198 if (ioflag & IO_INVAL)
1199 bp->b_flags |= B_NOCACHE;
1200 error = bwrite(bp);
1201 if (error)
1202 break;
1203 } else if ((n + on) == biosize) {
1204 bp->b_flags |= B_ASYNC;
1205 (void) ncl_writebp(bp, 0, NULL);
1206 } else {
1207 bdwrite(bp);
1208 }
1209 } while (uio->uio_resid > 0 && n > 0);
1210
1211 return (error);
1212 }
1213
1214 /*
1215 * Get an nfs cache block.
1216 *
1217 * Allocate a new one if the block isn't currently in the cache
1218 * and return the block marked busy. If the calling process is
1219 * interrupted by a signal for an interruptible mount point, return
1220 * NULL.
1221 *
1222 * The caller must carefully deal with the possible B_INVAL state of
1223 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1224 * indirectly), so synchronous reads can be issued without worrying about
1225 * the B_INVAL state. We have to be a little more careful when dealing
1226 * with writes (see comments in nfs_write()) when extending a file past
1227 * its EOF.
1228 */
1229 static struct buf *
1230 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1231 {
1232 struct buf *bp;
1233 struct mount *mp;
1234 struct nfsmount *nmp;
1235
1236 mp = vp->v_mount;
1237 nmp = VFSTONFS(mp);
1238
1239 if (nmp->nm_flag & NFSMNT_INT) {
1240 sigset_t oldset;
1241
1242 newnfs_set_sigmask(td, &oldset);
1243 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1244 newnfs_restore_sigmask(td, &oldset);
1245 while (bp == NULL) {
1246 if (newnfs_sigintr(nmp, td))
1247 return (NULL);
1248 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1249 }
1250 } else {
1251 bp = getblk(vp, bn, size, 0, 0, 0);
1252 }
1253
1254 if (vp->v_type == VREG)
1255 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1256 return (bp);
1257 }
1258
1259 /*
1260 * Flush and invalidate all dirty buffers. If another process is already
1261 * doing the flush, just wait for completion.
1262 */
1263 int
1264 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1265 {
1266 struct nfsnode *np = VTONFS(vp);
1267 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1268 int error = 0, slpflag, slptimeo;
1269 int old_lock = 0;
1270
1271 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1272
1273 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1274 intrflg = 0;
1275 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1276 intrflg = 1;
1277 if (intrflg) {
1278 slpflag = NFS_PCATCH;
1279 slptimeo = 2 * hz;
1280 } else {
1281 slpflag = 0;
1282 slptimeo = 0;
1283 }
1284
1285 old_lock = ncl_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 ncl_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_LOCK(vp->v_bufobj.bo_object);
1301 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1302 VM_OBJECT_UNLOCK(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 = newnfs_sigintr(nmp, td)))
1309 goto out;
1310 }
1311
1312 error = vinvalbuf(vp, flags, slpflag, 0);
1313 while (error) {
1314 if (intrflg && (error = newnfs_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 ncl_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: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1333 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1334 */
1335 int
1336 ncl_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 mtx_lock(&ncl_iod_mutex);
1350 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1351 (nmp->nm_bufqiods > ncl_numasync / 2)) {
1352 mtx_unlock(&ncl_iod_mutex);
1353 return(EIO);
1354 }
1355 again:
1356 if (nmp->nm_flag & NFSMNT_INT)
1357 slpflag = NFS_PCATCH;
1358 gotiod = FALSE;
1359
1360 /*
1361 * Find a free iod to process this request.
1362 */
1363 for (iod = 0; iod < ncl_numasync; iod++)
1364 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1365 gotiod = TRUE;
1366 break;
1367 }
1368
1369 /*
1370 * Try to create one if none are free.
1371 */
1372 if (!gotiod)
1373 ncl_nfsiodnew();
1374 else {
1375 /*
1376 * Found one, so wake it up and tell it which
1377 * mount to process.
1378 */
1379 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1380 iod, nmp));
1381 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1382 ncl_iodmount[iod] = nmp;
1383 nmp->nm_bufqiods++;
1384 wakeup(&ncl_iodwant[iod]);
1385 }
1386
1387 /*
1388 * If none are free, we may already have an iod working on this mount
1389 * point. If so, it will process our request.
1390 */
1391 if (!gotiod) {
1392 if (nmp->nm_bufqiods > 0) {
1393 NFS_DPF(ASYNCIO,
1394 ("ncl_asyncio: %d iods are already processing mount %p\n",
1395 nmp->nm_bufqiods, nmp));
1396 gotiod = TRUE;
1397 }
1398 }
1399
1400 /*
1401 * If we have an iod which can process the request, then queue
1402 * the buffer.
1403 */
1404 if (gotiod) {
1405 /*
1406 * Ensure that the queue never grows too large. We still want
1407 * to asynchronize so we block rather then return EIO.
1408 */
1409 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1410 NFS_DPF(ASYNCIO,
1411 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1412 nmp->nm_bufqwant = TRUE;
1413 error = newnfs_msleep(td, &nmp->nm_bufq,
1414 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1415 slptimeo);
1416 if (error) {
1417 error2 = newnfs_sigintr(nmp, td);
1418 if (error2) {
1419 mtx_unlock(&ncl_iod_mutex);
1420 return (error2);
1421 }
1422 if (slpflag == NFS_PCATCH) {
1423 slpflag = 0;
1424 slptimeo = 2 * hz;
1425 }
1426 }
1427 /*
1428 * We might have lost our iod while sleeping,
1429 * so check and loop if nescessary.
1430 */
1431 goto again;
1432 }
1433
1434 /* We might have lost our nfsiod */
1435 if (nmp->nm_bufqiods == 0) {
1436 NFS_DPF(ASYNCIO,
1437 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1438 goto again;
1439 }
1440
1441 if (bp->b_iocmd == BIO_READ) {
1442 if (bp->b_rcred == NOCRED && cred != NOCRED)
1443 bp->b_rcred = crhold(cred);
1444 } else {
1445 if (bp->b_wcred == NOCRED && cred != NOCRED)
1446 bp->b_wcred = crhold(cred);
1447 }
1448
1449 if (bp->b_flags & B_REMFREE)
1450 bremfreef(bp);
1451 BUF_KERNPROC(bp);
1452 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1453 nmp->nm_bufqlen++;
1454 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1455 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1456 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1457 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1458 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1459 }
1460 mtx_unlock(&ncl_iod_mutex);
1461 return (0);
1462 }
1463
1464 mtx_unlock(&ncl_iod_mutex);
1465
1466 /*
1467 * All the iods are busy on other mounts, so return EIO to
1468 * force the caller to process the i/o synchronously.
1469 */
1470 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1471 return (EIO);
1472 }
1473
1474 void
1475 ncl_doio_directwrite(struct buf *bp)
1476 {
1477 int iomode, must_commit;
1478 struct uio *uiop = (struct uio *)bp->b_caller1;
1479 char *iov_base = uiop->uio_iov->iov_base;
1480
1481 iomode = NFSWRITE_FILESYNC;
1482 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1483 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1484 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1485 free(iov_base, M_NFSDIRECTIO);
1486 free(uiop->uio_iov, M_NFSDIRECTIO);
1487 free(uiop, M_NFSDIRECTIO);
1488 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1489 struct nfsnode *np = VTONFS(bp->b_vp);
1490 mtx_lock(&np->n_mtx);
1491 np->n_directio_asyncwr--;
1492 if (np->n_directio_asyncwr == 0) {
1493 np->n_flag &= ~NMODIFIED;
1494 if ((np->n_flag & NFSYNCWAIT)) {
1495 np->n_flag &= ~NFSYNCWAIT;
1496 wakeup((caddr_t)&np->n_directio_asyncwr);
1497 }
1498 }
1499 mtx_unlock(&np->n_mtx);
1500 }
1501 bp->b_vp = NULL;
1502 relpbuf(bp, &ncl_pbuf_freecnt);
1503 }
1504
1505 /*
1506 * Do an I/O operation to/from a cache block. This may be called
1507 * synchronously or from an nfsiod.
1508 */
1509 int
1510 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1511 int called_from_strategy)
1512 {
1513 struct uio *uiop;
1514 struct nfsnode *np;
1515 struct nfsmount *nmp;
1516 int error = 0, iomode, must_commit = 0;
1517 struct uio uio;
1518 struct iovec io;
1519 struct proc *p = td ? td->td_proc : NULL;
1520 uint8_t iocmd;
1521
1522 np = VTONFS(vp);
1523 nmp = VFSTONFS(vp->v_mount);
1524 uiop = &uio;
1525 uiop->uio_iov = &io;
1526 uiop->uio_iovcnt = 1;
1527 uiop->uio_segflg = UIO_SYSSPACE;
1528 uiop->uio_td = td;
1529
1530 /*
1531 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1532 * do this here so we do not have to do it in all the code that
1533 * calls us.
1534 */
1535 bp->b_flags &= ~B_INVAL;
1536 bp->b_ioflags &= ~BIO_ERROR;
1537
1538 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1539 iocmd = bp->b_iocmd;
1540 if (iocmd == BIO_READ) {
1541 io.iov_len = uiop->uio_resid = bp->b_bcount;
1542 io.iov_base = bp->b_data;
1543 uiop->uio_rw = UIO_READ;
1544
1545 switch (vp->v_type) {
1546 case VREG:
1547 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1548 NFSINCRGLOBAL(newnfsstats.read_bios);
1549 error = ncl_readrpc(vp, uiop, cr);
1550
1551 if (!error) {
1552 if (uiop->uio_resid) {
1553 /*
1554 * If we had a short read with no error, we must have
1555 * hit a file hole. We should zero-fill the remainder.
1556 * This can also occur if the server hits the file EOF.
1557 *
1558 * Holes used to be able to occur due to pending
1559 * writes, but that is not possible any longer.
1560 */
1561 int nread = bp->b_bcount - uiop->uio_resid;
1562 int left = uiop->uio_resid;
1563
1564 if (left > 0)
1565 bzero((char *)bp->b_data + nread, left);
1566 uiop->uio_resid = 0;
1567 }
1568 }
1569 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1570 if (p && (vp->v_vflag & VV_TEXT)) {
1571 mtx_lock(&np->n_mtx);
1572 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1573 mtx_unlock(&np->n_mtx);
1574 PROC_LOCK(p);
1575 killproc(p, "text file modification");
1576 PROC_UNLOCK(p);
1577 } else
1578 mtx_unlock(&np->n_mtx);
1579 }
1580 break;
1581 case VLNK:
1582 uiop->uio_offset = (off_t)0;
1583 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1584 error = ncl_readlinkrpc(vp, uiop, cr);
1585 break;
1586 case VDIR:
1587 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1588 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1589 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1590 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1591 if (error == NFSERR_NOTSUPP)
1592 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1593 }
1594 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1595 error = ncl_readdirrpc(vp, uiop, cr, td);
1596 /*
1597 * end-of-directory sets B_INVAL but does not generate an
1598 * error.
1599 */
1600 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1601 bp->b_flags |= B_INVAL;
1602 break;
1603 default:
1604 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1605 break;
1606 };
1607 if (error) {
1608 bp->b_ioflags |= BIO_ERROR;
1609 bp->b_error = error;
1610 }
1611 } else {
1612 /*
1613 * If we only need to commit, try to commit
1614 */
1615 if (bp->b_flags & B_NEEDCOMMIT) {
1616 int retv;
1617 off_t off;
1618
1619 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1620 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1621 bp->b_wcred, td);
1622 if (retv == 0) {
1623 bp->b_dirtyoff = bp->b_dirtyend = 0;
1624 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1625 bp->b_resid = 0;
1626 bufdone(bp);
1627 return (0);
1628 }
1629 if (retv == NFSERR_STALEWRITEVERF) {
1630 ncl_clearcommit(vp->v_mount);
1631 }
1632 }
1633
1634 /*
1635 * Setup for actual write
1636 */
1637 mtx_lock(&np->n_mtx);
1638 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1639 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1640 mtx_unlock(&np->n_mtx);
1641
1642 if (bp->b_dirtyend > bp->b_dirtyoff) {
1643 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1644 - bp->b_dirtyoff;
1645 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1646 + bp->b_dirtyoff;
1647 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1648 uiop->uio_rw = UIO_WRITE;
1649 NFSINCRGLOBAL(newnfsstats.write_bios);
1650
1651 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1652 iomode = NFSWRITE_UNSTABLE;
1653 else
1654 iomode = NFSWRITE_FILESYNC;
1655
1656 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1657 called_from_strategy);
1658
1659 /*
1660 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1661 * to cluster the buffers needing commit. This will allow
1662 * the system to submit a single commit rpc for the whole
1663 * cluster. We can do this even if the buffer is not 100%
1664 * dirty (relative to the NFS blocksize), so we optimize the
1665 * append-to-file-case.
1666 *
1667 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1668 * cleared because write clustering only works for commit
1669 * rpc's, not for the data portion of the write).
1670 */
1671
1672 if (!error && iomode == NFSWRITE_UNSTABLE) {
1673 bp->b_flags |= B_NEEDCOMMIT;
1674 if (bp->b_dirtyoff == 0
1675 && bp->b_dirtyend == bp->b_bcount)
1676 bp->b_flags |= B_CLUSTEROK;
1677 } else {
1678 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1679 }
1680
1681 /*
1682 * For an interrupted write, the buffer is still valid
1683 * and the write hasn't been pushed to the server yet,
1684 * so we can't set BIO_ERROR and report the interruption
1685 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1686 * is not relevant, so the rpc attempt is essentially
1687 * a noop. For the case of a V3 write rpc not being
1688 * committed to stable storage, the block is still
1689 * dirty and requires either a commit rpc or another
1690 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1691 * the block is reused. This is indicated by setting
1692 * the B_DELWRI and B_NEEDCOMMIT flags.
1693 *
1694 * EIO is returned by ncl_writerpc() to indicate a recoverable
1695 * write error and is handled as above, except that
1696 * B_EINTR isn't set. One cause of this is a stale stateid
1697 * error for the RPC that indicates recovery is required,
1698 * when called with called_from_strategy != 0.
1699 *
1700 * If the buffer is marked B_PAGING, it does not reside on
1701 * the vp's paging queues so we cannot call bdirty(). The
1702 * bp in this case is not an NFS cache block so we should
1703 * be safe. XXX
1704 *
1705 * The logic below breaks up errors into recoverable and
1706 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1707 * and keep the buffer around for potential write retries.
1708 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1709 * and save the error in the nfsnode. This is less than ideal
1710 * but necessary. Keeping such buffers around could potentially
1711 * cause buffer exhaustion eventually (they can never be written
1712 * out, so will get constantly be re-dirtied). It also causes
1713 * all sorts of vfs panics. For non-recoverable write errors,
1714 * also invalidate the attrcache, so we'll be forced to go over
1715 * the wire for this object, returning an error to user on next
1716 * call (most of the time).
1717 */
1718 if (error == EINTR || error == EIO || error == ETIMEDOUT
1719 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1720 int s;
1721
1722 s = splbio();
1723 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1724 if ((bp->b_flags & B_PAGING) == 0) {
1725 bdirty(bp);
1726 bp->b_flags &= ~B_DONE;
1727 }
1728 if ((error == EINTR || error == ETIMEDOUT) &&
1729 (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 mtx_unlock(&np->n_mtx);
1741 }
1742 bp->b_dirtyoff = bp->b_dirtyend = 0;
1743 }
1744 } else {
1745 bp->b_resid = 0;
1746 bufdone(bp);
1747 return (0);
1748 }
1749 }
1750 bp->b_resid = uiop->uio_resid;
1751 if (must_commit)
1752 ncl_clearcommit(vp->v_mount);
1753 bufdone(bp);
1754 return (error);
1755 }
1756
1757 /*
1758 * Used to aid in handling ftruncate() operations on the NFS client side.
1759 * Truncation creates a number of special problems for NFS. We have to
1760 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1761 * we have to properly handle VM pages or (potentially dirty) buffers
1762 * that straddle the truncation point.
1763 */
1764
1765 int
1766 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1767 {
1768 struct nfsnode *np = VTONFS(vp);
1769 u_quad_t tsize;
1770 int biosize = vp->v_bufobj.bo_bsize;
1771 int error = 0;
1772
1773 mtx_lock(&np->n_mtx);
1774 tsize = np->n_size;
1775 np->n_size = nsize;
1776 mtx_unlock(&np->n_mtx);
1777
1778 if (nsize < tsize) {
1779 struct buf *bp;
1780 daddr_t lbn;
1781 int bufsize;
1782
1783 /*
1784 * vtruncbuf() doesn't get the buffer overlapping the
1785 * truncation point. We may have a B_DELWRI and/or B_CACHE
1786 * buffer that now needs to be truncated.
1787 */
1788 error = vtruncbuf(vp, cred, td, nsize, biosize);
1789 lbn = nsize / biosize;
1790 bufsize = nsize & (biosize - 1);
1791 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1792 if (!bp)
1793 return EINTR;
1794 if (bp->b_dirtyoff > bp->b_bcount)
1795 bp->b_dirtyoff = bp->b_bcount;
1796 if (bp->b_dirtyend > bp->b_bcount)
1797 bp->b_dirtyend = bp->b_bcount;
1798 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1799 brelse(bp);
1800 } else {
1801 vnode_pager_setsize(vp, nsize);
1802 }
1803 return(error);
1804 }
1805
Cache object: d6dd42c908cf1125e22f0f5c5ffef239
|