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: head/sys/fs/nfsclient/nfs_clbio.c 234605 2012-04-23 13:21:28Z trasz $");
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/mount.h>
46#include <sys/vmmeter.h>
47#include <sys/vnode.h>
48
49#include <vm/vm.h>
50#include <vm/vm_extern.h>
51#include <vm/vm_page.h>
52#include <vm/vm_object.h>
53#include <vm/vm_pager.h>
54#include <vm/vnode_pager.h>
55
56#include <fs/nfs/nfsport.h>
57#include <fs/nfsclient/nfsmount.h>
58#include <fs/nfsclient/nfs.h>
59#include <fs/nfsclient/nfsnode.h>
60#include <fs/nfsclient/nfs_kdtrace.h>
61
62extern int newnfs_directio_allow_mmap;
63extern struct nfsstats newnfsstats;
64extern struct mtx ncl_iod_mutex;
65extern int ncl_numasync;
66extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
67extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
68extern int newnfs_directio_enable;
69extern int nfs_keep_dirty_on_error;
70
71int ncl_pbuf_freecnt = -1; /* start out unlimited */
72
73static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
74 struct thread *td);
75static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
76 struct ucred *cred, int ioflag);
77
78/*
79 * Vnode op for VM getpages.
80 */
81int
82ncl_getpages(struct vop_getpages_args *ap)
83{
84 int i, error, nextoff, size, toff, count, npages;
85 struct uio uio;
86 struct iovec iov;
87 vm_offset_t kva;
88 struct buf *bp;
89 struct vnode *vp;
90 struct thread *td;
91 struct ucred *cred;
92 struct nfsmount *nmp;
93 vm_object_t object;
94 vm_page_t *pages;
95 struct nfsnode *np;
96
97 vp = ap->a_vp;
98 np = VTONFS(vp);
99 td = curthread; /* XXX */
100 cred = curthread->td_ucred; /* XXX */
101 nmp = VFSTONFS(vp->v_mount);
102 pages = ap->a_m;
103 count = ap->a_count;
104
105 if ((object = vp->v_object) == NULL) {
106 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
107 return (VM_PAGER_ERROR);
108 }
109
110 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
111 mtx_lock(&np->n_mtx);
112 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
113 mtx_unlock(&np->n_mtx);
114 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n");
115 return (VM_PAGER_ERROR);
116 } else
117 mtx_unlock(&np->n_mtx);
118 }
119
120 mtx_lock(&nmp->nm_mtx);
121 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
122 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
123 mtx_unlock(&nmp->nm_mtx);
124 /* We'll never get here for v4, because we always have fsinfo */
125 (void)ncl_fsinfo(nmp, vp, cred, td);
126 } else
127 mtx_unlock(&nmp->nm_mtx);
128
129 npages = btoc(count);
130
131 /*
132 * If the requested page is partially valid, just return it and
133 * allow the pager to zero-out the blanks. Partially valid pages
134 * can only occur at the file EOF.
135 */
136 VM_OBJECT_LOCK(object);
137 if (pages[ap->a_reqpage]->valid != 0) {
138 for (i = 0; i < npages; ++i) {
139 if (i != ap->a_reqpage) {
140 vm_page_lock(pages[i]);
141 vm_page_free(pages[i]);
142 vm_page_unlock(pages[i]);
143 }
144 }
145 VM_OBJECT_UNLOCK(object);
146 return (0);
147 }
148 VM_OBJECT_UNLOCK(object);
149
150 /*
151 * We use only the kva address for the buffer, but this is extremely
152 * convienient and fast.
153 */
154 bp = getpbuf(&ncl_pbuf_freecnt);
155
156 kva = (vm_offset_t) bp->b_data;
157 pmap_qenter(kva, pages, npages);
158 PCPU_INC(cnt.v_vnodein);
159 PCPU_ADD(cnt.v_vnodepgsin, npages);
160
161 iov.iov_base = (caddr_t) kva;
162 iov.iov_len = count;
163 uio.uio_iov = &iov;
164 uio.uio_iovcnt = 1;
165 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
166 uio.uio_resid = count;
167 uio.uio_segflg = UIO_SYSSPACE;
168 uio.uio_rw = UIO_READ;
169 uio.uio_td = td;
170
171 error = ncl_readrpc(vp, &uio, cred);
172 pmap_qremove(kva, npages);
173
174 relpbuf(bp, &ncl_pbuf_freecnt);
175
176 if (error && (uio.uio_resid == count)) {
177 ncl_printf("nfs_getpages: error %d\n", error);
178 VM_OBJECT_LOCK(object);
179 for (i = 0; i < npages; ++i) {
180 if (i != ap->a_reqpage) {
181 vm_page_lock(pages[i]);
182 vm_page_free(pages[i]);
183 vm_page_unlock(pages[i]);
184 }
185 }
186 VM_OBJECT_UNLOCK(object);
187 return (VM_PAGER_ERROR);
188 }
189
190 /*
191 * Calculate the number of bytes read and validate only that number
192 * of bytes. Note that due to pending writes, size may be 0. This
193 * does not mean that the remaining data is invalid!
194 */
195
196 size = count - uio.uio_resid;
197 VM_OBJECT_LOCK(object);
198 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
199 vm_page_t m;
200 nextoff = toff + PAGE_SIZE;
201 m = pages[i];
202
203 if (nextoff <= size) {
204 /*
205 * Read operation filled an entire page
206 */
207 m->valid = VM_PAGE_BITS_ALL;
208 KASSERT(m->dirty == 0,
209 ("nfs_getpages: page %p is dirty", m));
210 } else if (size > toff) {
211 /*
212 * Read operation filled a partial page.
213 */
214 m->valid = 0;
215 vm_page_set_valid_range(m, 0, size - toff);
216 KASSERT(m->dirty == 0,
217 ("nfs_getpages: page %p is dirty", m));
218 } else {
219 /*
220 * Read operation was short. If no error occured
221 * we may have hit a zero-fill section. We simply
222 * leave valid set to 0.
223 */
224 ;
225 }
226 if (i != ap->a_reqpage) {
227 /*
228 * Whether or not to leave the page activated is up in
229 * the air, but we should put the page on a page queue
230 * somewhere (it already is in the object). Result:
231 * It appears that emperical results show that
232 * deactivating pages is best.
233 */
234
235 /*
236 * Just in case someone was asking for this page we
237 * now tell them that it is ok to use.
238 */
239 if (!error) {
240 if (m->oflags & VPO_WANTED) {
241 vm_page_lock(m);
242 vm_page_activate(m);
243 vm_page_unlock(m);
244 } else {
245 vm_page_lock(m);
246 vm_page_deactivate(m);
247 vm_page_unlock(m);
248 }
249 vm_page_wakeup(m);
250 } else {
251 vm_page_lock(m);
252 vm_page_free(m);
253 vm_page_unlock(m);
254 }
255 }
256 }
257 VM_OBJECT_UNLOCK(object);
258 return (0);
259}
260
261/*
262 * Vnode op for VM putpages.
263 */
264int
265ncl_putpages(struct vop_putpages_args *ap)
266{
267 struct uio uio;
268 struct iovec iov;
269 vm_offset_t kva;
270 struct buf *bp;
271 int iomode, must_commit, i, error, npages, count;
272 off_t offset;
273 int *rtvals;
274 struct vnode *vp;
275 struct thread *td;
276 struct ucred *cred;
277 struct nfsmount *nmp;
278 struct nfsnode *np;
279 vm_page_t *pages;
280
281 vp = ap->a_vp;
282 np = VTONFS(vp);
283 td = curthread; /* XXX */
284 cred = curthread->td_ucred; /* XXX */
285 nmp = VFSTONFS(vp->v_mount);
286 pages = ap->a_m;
287 count = ap->a_count;
288 rtvals = ap->a_rtvals;
289 npages = btoc(count);
290 offset = IDX_TO_OFF(pages[0]->pindex);
291
292 mtx_lock(&nmp->nm_mtx);
293 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
294 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
295 mtx_unlock(&nmp->nm_mtx);
296 (void)ncl_fsinfo(nmp, vp, cred, td);
297 } else
298 mtx_unlock(&nmp->nm_mtx);
299
300 mtx_lock(&np->n_mtx);
301 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
302 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
303 mtx_unlock(&np->n_mtx);
304 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
305 mtx_lock(&np->n_mtx);
306 }
307
308 for (i = 0; i < npages; i++)
309 rtvals[i] = VM_PAGER_ERROR;
310
311 /*
312 * When putting pages, do not extend file past EOF.
313 */
314 if (offset + count > np->n_size) {
315 count = np->n_size - offset;
316 if (count < 0)
317 count = 0;
318 }
319 mtx_unlock(&np->n_mtx);
320
321 /*
322 * We use only the kva address for the buffer, but this is extremely
323 * convienient and fast.
324 */
325 bp = getpbuf(&ncl_pbuf_freecnt);
326
327 kva = (vm_offset_t) bp->b_data;
328 pmap_qenter(kva, pages, npages);
329 PCPU_INC(cnt.v_vnodeout);
330 PCPU_ADD(cnt.v_vnodepgsout, count);
331
332 iov.iov_base = (caddr_t) kva;
333 iov.iov_len = count;
334 uio.uio_iov = &iov;
335 uio.uio_iovcnt = 1;
336 uio.uio_offset = offset;
337 uio.uio_resid = count;
338 uio.uio_segflg = UIO_SYSSPACE;
339 uio.uio_rw = UIO_WRITE;
340 uio.uio_td = td;
341
342 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
343 iomode = NFSWRITE_UNSTABLE;
344 else
345 iomode = NFSWRITE_FILESYNC;
346
347 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
348
349 pmap_qremove(kva, npages);
350 relpbuf(bp, &ncl_pbuf_freecnt);
351
352 if (error == 0 || !nfs_keep_dirty_on_error) {
353 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
354 if (must_commit)
355 ncl_clearcommit(vp->v_mount);
356 }
357 return rtvals[0];
358}
359
360/*
361 * For nfs, cache consistency can only be maintained approximately.
362 * Although RFC1094 does not specify the criteria, the following is
363 * believed to be compatible with the reference port.
364 * For nfs:
365 * If the file's modify time on the server has changed since the
366 * last read rpc or you have written to the file,
367 * you may have lost data cache consistency with the
368 * server, so flush all of the file's data out of the cache.
369 * Then force a getattr rpc to ensure that you have up to date
370 * attributes.
371 * NB: This implies that cache data can be read when up to
372 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
373 * attributes this could be forced by setting n_attrstamp to 0 before
374 * the VOP_GETATTR() call.
375 */
376static inline int
377nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
378{
379 int error = 0;
380 struct vattr vattr;
381 struct nfsnode *np = VTONFS(vp);
382 int old_lock;
383
384 /*
385 * Grab the exclusive lock before checking whether the cache is
386 * consistent.
387 * XXX - We can make this cheaper later (by acquiring cheaper locks).
388 * But for now, this suffices.
389 */
390 old_lock = ncl_upgrade_vnlock(vp);
391 if (vp->v_iflag & VI_DOOMED) {
392 ncl_downgrade_vnlock(vp, old_lock);
393 return (EBADF);
394 }
395
396 mtx_lock(&np->n_mtx);
397 if (np->n_flag & NMODIFIED) {
398 mtx_unlock(&np->n_mtx);
399 if (vp->v_type != VREG) {
400 if (vp->v_type != VDIR)
401 panic("nfs: bioread, not dir");
402 ncl_invaldir(vp);
403 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
404 if (error)
405 goto out;
406 }
407 np->n_attrstamp = 0;
408 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
409 error = VOP_GETATTR(vp, &vattr, cred);
410 if (error)
411 goto out;
412 mtx_lock(&np->n_mtx);
413 np->n_mtime = vattr.va_mtime;
414 mtx_unlock(&np->n_mtx);
415 } else {
416 mtx_unlock(&np->n_mtx);
417 error = VOP_GETATTR(vp, &vattr, cred);
418 if (error)
419 return (error);
420 mtx_lock(&np->n_mtx);
421 if ((np->n_flag & NSIZECHANGED)
422 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
423 mtx_unlock(&np->n_mtx);
424 if (vp->v_type == VDIR)
425 ncl_invaldir(vp);
426 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
427 if (error)
428 goto out;
429 mtx_lock(&np->n_mtx);
430 np->n_mtime = vattr.va_mtime;
431 np->n_flag &= ~NSIZECHANGED;
432 }
433 mtx_unlock(&np->n_mtx);
434 }
435out:
436 ncl_downgrade_vnlock(vp, old_lock);
437 return error;
438}
439
440/*
441 * Vnode op for read using bio
442 */
443int
444ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
445{
446 struct nfsnode *np = VTONFS(vp);
447 int biosize, i;
448 struct buf *bp, *rabp;
449 struct thread *td;
450 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
451 daddr_t lbn, rabn;
452 int bcount;
453 int seqcount;
454 int nra, error = 0, n = 0, on = 0;
455 off_t tmp_off;
456
457 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
458 if (uio->uio_resid == 0)
459 return (0);
460 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
461 return (EINVAL);
462 td = uio->uio_td;
463
464 mtx_lock(&nmp->nm_mtx);
465 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
466 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
467 mtx_unlock(&nmp->nm_mtx);
468 (void)ncl_fsinfo(nmp, vp, cred, td);
469 mtx_lock(&nmp->nm_mtx);
470 }
471 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
472 (void) newnfs_iosize(nmp);
473
474 tmp_off = uio->uio_offset + uio->uio_resid;
475 if (vp->v_type != VDIR &&
476 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
477 mtx_unlock(&nmp->nm_mtx);
478 return (EFBIG);
479 }
480 mtx_unlock(&nmp->nm_mtx);
481
482 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
483 /* No caching/ no readaheads. Just read data into the user buffer */
484 return ncl_readrpc(vp, uio, cred);
485
486 biosize = vp->v_bufobj.bo_bsize;
487 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
488
489 error = nfs_bioread_check_cons(vp, td, cred);
490 if (error)
491 return error;
492
493 do {
494 u_quad_t nsize;
495
496 mtx_lock(&np->n_mtx);
497 nsize = np->n_size;
498 mtx_unlock(&np->n_mtx);
499
500 switch (vp->v_type) {
501 case VREG:
502 NFSINCRGLOBAL(newnfsstats.biocache_reads);
503 lbn = uio->uio_offset / biosize;
504 on = uio->uio_offset & (biosize - 1);
505
506 /*
507 * Start the read ahead(s), as required.
508 */
509 if (nmp->nm_readahead > 0) {
510 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
511 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
512 rabn = lbn + 1 + nra;
513 if (incore(&vp->v_bufobj, rabn) == NULL) {
514 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
515 if (!rabp) {
516 error = newnfs_sigintr(nmp, td);
517 return (error ? error : EINTR);
518 }
519 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
520 rabp->b_flags |= B_ASYNC;
521 rabp->b_iocmd = BIO_READ;
522 vfs_busy_pages(rabp, 0);
523 if (ncl_asyncio(nmp, rabp, cred, td)) {
524 rabp->b_flags |= B_INVAL;
525 rabp->b_ioflags |= BIO_ERROR;
526 vfs_unbusy_pages(rabp);
527 brelse(rabp);
528 break;
529 }
530 } else {
531 brelse(rabp);
532 }
533 }
534 }
535 }
536
537 /* Note that bcount is *not* DEV_BSIZE aligned. */
538 bcount = biosize;
539 if ((off_t)lbn * biosize >= nsize) {
540 bcount = 0;
541 } else if ((off_t)(lbn + 1) * biosize > nsize) {
542 bcount = nsize - (off_t)lbn * biosize;
543 }
544 bp = nfs_getcacheblk(vp, lbn, bcount, td);
545
546 if (!bp) {
547 error = newnfs_sigintr(nmp, td);
548 return (error ? error : EINTR);
549 }
550
551 /*
552 * If B_CACHE is not set, we must issue the read. If this
553 * fails, we return an error.
554 */
555
556 if ((bp->b_flags & B_CACHE) == 0) {
557 bp->b_iocmd = BIO_READ;
558 vfs_busy_pages(bp, 0);
559 error = ncl_doio(vp, bp, cred, td, 0);
560 if (error) {
561 brelse(bp);
562 return (error);
563 }
564 }
565
566 /*
567 * on is the offset into the current bp. Figure out how many
568 * bytes we can copy out of the bp. Note that bcount is
569 * NOT DEV_BSIZE aligned.
570 *
571 * Then figure out how many bytes we can copy into the uio.
572 */
573
574 n = 0;
575 if (on < bcount)
576 n = MIN((unsigned)(bcount - on), uio->uio_resid);
577 break;
578 case VLNK:
579 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
580 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
581 if (!bp) {
582 error = newnfs_sigintr(nmp, td);
583 return (error ? error : EINTR);
584 }
585 if ((bp->b_flags & B_CACHE) == 0) {
586 bp->b_iocmd = BIO_READ;
587 vfs_busy_pages(bp, 0);
588 error = ncl_doio(vp, bp, cred, td, 0);
589 if (error) {
590 bp->b_ioflags |= BIO_ERROR;
591 brelse(bp);
592 return (error);
593 }
594 }
595 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
596 on = 0;
597 break;
598 case VDIR:
599 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
600 if (np->n_direofoffset
601 && uio->uio_offset >= np->n_direofoffset) {
602 return (0);
603 }
604 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
605 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
606 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
607 if (!bp) {
608 error = newnfs_sigintr(nmp, td);
609 return (error ? error : EINTR);
610 }
611 if ((bp->b_flags & B_CACHE) == 0) {
612 bp->b_iocmd = BIO_READ;
613 vfs_busy_pages(bp, 0);
614 error = ncl_doio(vp, bp, cred, td, 0);
615 if (error) {
616 brelse(bp);
617 }
618 while (error == NFSERR_BAD_COOKIE) {
619 ncl_invaldir(vp);
620 error = ncl_vinvalbuf(vp, 0, td, 1);
621 /*
622 * Yuck! The directory has been modified on the
623 * server. The only way to get the block is by
624 * reading from the beginning to get all the
625 * offset cookies.
626 *
627 * Leave the last bp intact unless there is an error.
628 * Loop back up to the while if the error is another
629 * NFSERR_BAD_COOKIE (double yuch!).
630 */
631 for (i = 0; i <= lbn && !error; i++) {
632 if (np->n_direofoffset
633 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
634 return (0);
635 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
636 if (!bp) {
637 error = newnfs_sigintr(nmp, td);
638 return (error ? error : EINTR);
639 }
640 if ((bp->b_flags & B_CACHE) == 0) {
641 bp->b_iocmd = BIO_READ;
642 vfs_busy_pages(bp, 0);
643 error = ncl_doio(vp, bp, cred, td, 0);
644 /*
645 * no error + B_INVAL == directory EOF,
646 * use the block.
647 */
648 if (error == 0 && (bp->b_flags & B_INVAL))
649 break;
650 }
651 /*
652 * An error will throw away the block and the
653 * for loop will break out. If no error and this
654 * is not the block we want, we throw away the
655 * block and go for the next one via the for loop.
656 */
657 if (error || i < lbn)
658 brelse(bp);
659 }
660 }
661 /*
662 * The above while is repeated if we hit another cookie
663 * error. If we hit an error and it wasn't a cookie error,
664 * we give up.
665 */
666 if (error)
667 return (error);
668 }
669
670 /*
671 * If not eof and read aheads are enabled, start one.
672 * (You need the current block first, so that you have the
673 * directory offset cookie of the next block.)
674 */
675 if (nmp->nm_readahead > 0 &&
676 (bp->b_flags & B_INVAL) == 0 &&
677 (np->n_direofoffset == 0 ||
678 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
679 incore(&vp->v_bufobj, lbn + 1) == NULL) {
680 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
681 if (rabp) {
682 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
683 rabp->b_flags |= B_ASYNC;
684 rabp->b_iocmd = BIO_READ;
685 vfs_busy_pages(rabp, 0);
686 if (ncl_asyncio(nmp, rabp, cred, td)) {
687 rabp->b_flags |= B_INVAL;
688 rabp->b_ioflags |= BIO_ERROR;
689 vfs_unbusy_pages(rabp);
690 brelse(rabp);
691 }
692 } else {
693 brelse(rabp);
694 }
695 }
696 }
697 /*
698 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
699 * chopped for the EOF condition, we cannot tell how large
700 * NFS directories are going to be until we hit EOF. So
701 * an NFS directory buffer is *not* chopped to its EOF. Now,
702 * it just so happens that b_resid will effectively chop it
703 * to EOF. *BUT* this information is lost if the buffer goes
704 * away and is reconstituted into a B_CACHE state ( due to
705 * being VMIO ) later. So we keep track of the directory eof
706 * in np->n_direofoffset and chop it off as an extra step
707 * right here.
708 */
709 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
710 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
711 n = np->n_direofoffset - uio->uio_offset;
712 break;
713 default:
714 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
715 bp = NULL;
716 break;
717 };
718
719 if (n > 0) {
720 error = uiomove(bp->b_data + on, (int)n, uio);
721 }
722 if (vp->v_type == VLNK)
723 n = 0;
724 if (bp != NULL)
725 brelse(bp);
726 } while (error == 0 && uio->uio_resid > 0 && n > 0);
727 return (error);
728}
729
730/*
731 * The NFS write path cannot handle iovecs with len > 1. So we need to
732 * break up iovecs accordingly (restricting them to wsize).
733 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
734 * For the ASYNC case, 2 copies are needed. The first a copy from the
735 * user buffer to a staging buffer and then a second copy from the staging
736 * buffer to mbufs. This can be optimized by copying from the user buffer
737 * directly into mbufs and passing the chain down, but that requires a
738 * fair amount of re-working of the relevant codepaths (and can be done
739 * later).
740 */
741static int
742nfs_directio_write(vp, uiop, cred, ioflag)
743 struct vnode *vp;
744 struct uio *uiop;
745 struct ucred *cred;
746 int ioflag;
747{
748 int error;
749 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
750 struct thread *td = uiop->uio_td;
751 int size;
752 int wsize;
753
754 mtx_lock(&nmp->nm_mtx);
755 wsize = nmp->nm_wsize;
756 mtx_unlock(&nmp->nm_mtx);
757 if (ioflag & IO_SYNC) {
758 int iomode, must_commit;
759 struct uio uio;
760 struct iovec iov;
761do_sync:
762 while (uiop->uio_resid > 0) {
763 size = MIN(uiop->uio_resid, wsize);
764 size = MIN(uiop->uio_iov->iov_len, size);
765 iov.iov_base = uiop->uio_iov->iov_base;
766 iov.iov_len = size;
767 uio.uio_iov = &iov;
768 uio.uio_iovcnt = 1;
769 uio.uio_offset = uiop->uio_offset;
770 uio.uio_resid = size;
771 uio.uio_segflg = UIO_USERSPACE;
772 uio.uio_rw = UIO_WRITE;
773 uio.uio_td = td;
774 iomode = NFSWRITE_FILESYNC;
775 error = ncl_writerpc(vp, &uio, cred, &iomode,
776 &must_commit, 0);
777 KASSERT((must_commit == 0),
778 ("ncl_directio_write: Did not commit write"));
779 if (error)
780 return (error);
781 uiop->uio_offset += size;
782 uiop->uio_resid -= size;
783 if (uiop->uio_iov->iov_len <= size) {
784 uiop->uio_iovcnt--;
785 uiop->uio_iov++;
786 } else {
787 uiop->uio_iov->iov_base =
788 (char *)uiop->uio_iov->iov_base + size;
789 uiop->uio_iov->iov_len -= size;
790 }
791 }
792 } else {
793 struct uio *t_uio;
794 struct iovec *t_iov;
795 struct buf *bp;
796
797 /*
798 * Break up the write into blocksize chunks and hand these
799 * over to nfsiod's for write back.
800 * Unfortunately, this incurs a copy of the data. Since
801 * the user could modify the buffer before the write is
802 * initiated.
803 *
804 * The obvious optimization here is that one of the 2 copies
805 * in the async write path can be eliminated by copying the
806 * data here directly into mbufs and passing the mbuf chain
807 * down. But that will require a fair amount of re-working
808 * of the code and can be done if there's enough interest
809 * in NFS directio access.
810 */
811 while (uiop->uio_resid > 0) {
812 size = MIN(uiop->uio_resid, wsize);
813 size = MIN(uiop->uio_iov->iov_len, size);
814 bp = getpbuf(&ncl_pbuf_freecnt);
815 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
816 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
817 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
818 t_iov->iov_len = size;
819 t_uio->uio_iov = t_iov;
820 t_uio->uio_iovcnt = 1;
821 t_uio->uio_offset = uiop->uio_offset;
822 t_uio->uio_resid = size;
823 t_uio->uio_segflg = UIO_SYSSPACE;
824 t_uio->uio_rw = UIO_WRITE;
825 t_uio->uio_td = td;
826 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
827 uiop->uio_segflg == UIO_SYSSPACE,
828 ("nfs_directio_write: Bad uio_segflg"));
829 if (uiop->uio_segflg == UIO_USERSPACE) {
830 error = copyin(uiop->uio_iov->iov_base,
831 t_iov->iov_base, size);
832 if (error != 0)
833 goto err_free;
834 } else
835 /*
836 * UIO_SYSSPACE may never happen, but handle
837 * it just in case it does.
838 */
839 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
840 size);
841 bp->b_flags |= B_DIRECT;
842 bp->b_iocmd = BIO_WRITE;
843 if (cred != NOCRED) {
844 crhold(cred);
845 bp->b_wcred = cred;
846 } else
847 bp->b_wcred = NOCRED;
848 bp->b_caller1 = (void *)t_uio;
849 bp->b_vp = vp;
850 error = ncl_asyncio(nmp, bp, NOCRED, td);
851err_free:
852 if (error) {
853 free(t_iov->iov_base, M_NFSDIRECTIO);
854 free(t_iov, M_NFSDIRECTIO);
855 free(t_uio, M_NFSDIRECTIO);
856 bp->b_vp = NULL;
857 relpbuf(bp, &ncl_pbuf_freecnt);
858 if (error == EINTR)
859 return (error);
860 goto do_sync;
861 }
862 uiop->uio_offset += size;
863 uiop->uio_resid -= size;
864 if (uiop->uio_iov->iov_len <= size) {
865 uiop->uio_iovcnt--;
866 uiop->uio_iov++;
867 } else {
868 uiop->uio_iov->iov_base =
869 (char *)uiop->uio_iov->iov_base + size;
870 uiop->uio_iov->iov_len -= size;
871 }
872 }
873 }
874 return (0);
875}
876
877/*
878 * Vnode op for write using bio
879 */
880int
881ncl_write(struct vop_write_args *ap)
882{
883 int biosize;
884 struct uio *uio = ap->a_uio;
885 struct thread *td = uio->uio_td;
886 struct vnode *vp = ap->a_vp;
887 struct nfsnode *np = VTONFS(vp);
888 struct ucred *cred = ap->a_cred;
889 int ioflag = ap->a_ioflag;
890 struct buf *bp;
891 struct vattr vattr;
892 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
893 daddr_t lbn;
894 int bcount;
895 int n, on, error = 0;
896 off_t tmp_off;
897
898 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
899 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
900 ("ncl_write proc"));
901 if (vp->v_type != VREG)
902 return (EIO);
903 mtx_lock(&np->n_mtx);
904 if (np->n_flag & NWRITEERR) {
905 np->n_flag &= ~NWRITEERR;
906 mtx_unlock(&np->n_mtx);
907 return (np->n_error);
908 } else
909 mtx_unlock(&np->n_mtx);
910 mtx_lock(&nmp->nm_mtx);
911 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
912 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
913 mtx_unlock(&nmp->nm_mtx);
914 (void)ncl_fsinfo(nmp, vp, cred, td);
915 mtx_lock(&nmp->nm_mtx);
916 }
917 if (nmp->nm_wsize == 0)
918 (void) newnfs_iosize(nmp);
919 mtx_unlock(&nmp->nm_mtx);
920
921 /*
922 * Synchronously flush pending buffers if we are in synchronous
923 * mode or if we are appending.
924 */
925 if (ioflag & (IO_APPEND | IO_SYNC)) {
926 mtx_lock(&np->n_mtx);
927 if (np->n_flag & NMODIFIED) {
928 mtx_unlock(&np->n_mtx);
929#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
930 /*
931 * Require non-blocking, synchronous writes to
932 * dirty files to inform the program it needs
933 * to fsync(2) explicitly.
934 */
935 if (ioflag & IO_NDELAY)
936 return (EAGAIN);
937#endif
938flush_and_restart:
939 np->n_attrstamp = 0;
940 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
941 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
942 if (error)
943 return (error);
944 } else
945 mtx_unlock(&np->n_mtx);
946 }
947
948 /*
949 * If IO_APPEND then load uio_offset. We restart here if we cannot
950 * get the append lock.
951 */
952 if (ioflag & IO_APPEND) {
953 np->n_attrstamp = 0;
954 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
955 error = VOP_GETATTR(vp, &vattr, cred);
956 if (error)
957 return (error);
958 mtx_lock(&np->n_mtx);
959 uio->uio_offset = np->n_size;
960 mtx_unlock(&np->n_mtx);
961 }
962
963 if (uio->uio_offset < 0)
964 return (EINVAL);
965 tmp_off = uio->uio_offset + uio->uio_resid;
966 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
967 return (EFBIG);
968 if (uio->uio_resid == 0)
969 return (0);
970
971 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
972 return nfs_directio_write(vp, uio, cred, ioflag);
973
974 /*
975 * Maybe this should be above the vnode op call, but so long as
976 * file servers have no limits, i don't think it matters
977 */
978 if (vn_rlimit_fsize(vp, uio, td))
979 return (EFBIG);
980
981 biosize = vp->v_bufobj.bo_bsize;
982 /*
983 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
984 * would exceed the local maximum per-file write commit size when
985 * combined with those, we must decide whether to flush,
986 * go synchronous, or return error. We don't bother checking
987 * IO_UNIT -- we just make all writes atomic anyway, as there's
988 * no point optimizing for something that really won't ever happen.
989 */
990 if (!(ioflag & IO_SYNC)) {
991 int nflag;
992
993 mtx_lock(&np->n_mtx);
994 nflag = np->n_flag;
995 mtx_unlock(&np->n_mtx);
996 int needrestart = 0;
997 if (nmp->nm_wcommitsize < uio->uio_resid) {
998 /*
999 * If this request could not possibly be completed
1000 * without exceeding the maximum outstanding write
1001 * commit size, see if we can convert it into a
1002 * synchronous write operation.
1003 */
1004 if (ioflag & IO_NDELAY)
1005 return (EAGAIN);
1006 ioflag |= IO_SYNC;
1007 if (nflag & NMODIFIED)
1008 needrestart = 1;
1009 } else if (nflag & NMODIFIED) {
1010 int wouldcommit = 0;
1011 BO_LOCK(&vp->v_bufobj);
1012 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1013 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1014 b_bobufs) {
1015 if (bp->b_flags & B_NEEDCOMMIT)
1016 wouldcommit += bp->b_bcount;
1017 }
1018 }
1019 BO_UNLOCK(&vp->v_bufobj);
1020 /*
1021 * Since we're not operating synchronously and
1022 * bypassing the buffer cache, we are in a commit
1023 * and holding all of these buffers whether
1024 * transmitted or not. If not limited, this
1025 * will lead to the buffer cache deadlocking,
1026 * as no one else can flush our uncommitted buffers.
1027 */
1028 wouldcommit += uio->uio_resid;
1029 /*
1030 * If we would initially exceed the maximum
1031 * outstanding write commit size, flush and restart.
1032 */
1033 if (wouldcommit > nmp->nm_wcommitsize)
1034 needrestart = 1;
1035 }
1036 if (needrestart)
1037 goto flush_and_restart;
1038 }
1039
1040 do {
1041 NFSINCRGLOBAL(newnfsstats.biocache_writes);
1042 lbn = uio->uio_offset / biosize;
1043 on = uio->uio_offset & (biosize-1);
1044 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1045again:
1046 /*
1047 * Handle direct append and file extension cases, calculate
1048 * unaligned buffer size.
1049 */
1050 mtx_lock(&np->n_mtx);
1051 if (uio->uio_offset == np->n_size && n) {
1052 mtx_unlock(&np->n_mtx);
1053 /*
1054 * Get the buffer (in its pre-append state to maintain
1055 * B_CACHE if it was previously set). Resize the
1056 * nfsnode after we have locked the buffer to prevent
1057 * readers from reading garbage.
1058 */
1059 bcount = on;
1060 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1061
1062 if (bp != NULL) {
1063 long save;
1064
1065 mtx_lock(&np->n_mtx);
1066 np->n_size = uio->uio_offset + n;
1067 np->n_flag |= NMODIFIED;
1068 vnode_pager_setsize(vp, np->n_size);
1069 mtx_unlock(&np->n_mtx);
1070
1071 save = bp->b_flags & B_CACHE;
1072 bcount += n;
1073 allocbuf(bp, bcount);
1074 bp->b_flags |= save;
1075 }
1076 } else {
1077 /*
1078 * Obtain the locked cache block first, and then
1079 * adjust the file's size as appropriate.
1080 */
1081 bcount = on + n;
1082 if ((off_t)lbn * biosize + bcount < np->n_size) {
1083 if ((off_t)(lbn + 1) * biosize < np->n_size)
1084 bcount = biosize;
1085 else
1086 bcount = np->n_size - (off_t)lbn * biosize;
1087 }
1088 mtx_unlock(&np->n_mtx);
1089 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1090 mtx_lock(&np->n_mtx);
1091 if (uio->uio_offset + n > np->n_size) {
1092 np->n_size = uio->uio_offset + n;
1093 np->n_flag |= NMODIFIED;
1094 vnode_pager_setsize(vp, np->n_size);
1095 }
1096 mtx_unlock(&np->n_mtx);
1097 }
1098
1099 if (!bp) {
1100 error = newnfs_sigintr(nmp, td);
1101 if (!error)
1102 error = EINTR;
1103 break;
1104 }
1105
1106 /*
1107 * Issue a READ if B_CACHE is not set. In special-append
1108 * mode, B_CACHE is based on the buffer prior to the write
1109 * op and is typically set, avoiding the read. If a read
1110 * is required in special append mode, the server will
1111 * probably send us a short-read since we extended the file
1112 * on our end, resulting in b_resid == 0 and, thusly,
1113 * B_CACHE getting set.
1114 *
1115 * We can also avoid issuing the read if the write covers
1116 * the entire buffer. We have to make sure the buffer state
1117 * is reasonable in this case since we will not be initiating
1118 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1119 * more information.
1120 *
1121 * B_CACHE may also be set due to the buffer being cached
1122 * normally.
1123 */
1124
1125 if (on == 0 && n == bcount) {
1126 bp->b_flags |= B_CACHE;
1127 bp->b_flags &= ~B_INVAL;
1128 bp->b_ioflags &= ~BIO_ERROR;
1129 }
1130
1131 if ((bp->b_flags & B_CACHE) == 0) {
1132 bp->b_iocmd = BIO_READ;
1133 vfs_busy_pages(bp, 0);
1134 error = ncl_doio(vp, bp, cred, td, 0);
1135 if (error) {
1136 brelse(bp);
1137 break;
1138 }
1139 }
1140 if (bp->b_wcred == NOCRED)
1141 bp->b_wcred = crhold(cred);
1142 mtx_lock(&np->n_mtx);
1143 np->n_flag |= NMODIFIED;
1144 mtx_unlock(&np->n_mtx);
1145
1146 /*
1147 * If dirtyend exceeds file size, chop it down. This should
1148 * not normally occur but there is an append race where it
1149 * might occur XXX, so we log it.
1150 *
1151 * If the chopping creates a reverse-indexed or degenerate
1152 * situation with dirtyoff/end, we 0 both of them.
1153 */
1154
1155 if (bp->b_dirtyend > bcount) {
1156 ncl_printf("NFS append race @%lx:%d\n",
1157 (long)bp->b_blkno * DEV_BSIZE,
1158 bp->b_dirtyend - bcount);
1159 bp->b_dirtyend = bcount;
1160 }
1161
1162 if (bp->b_dirtyoff >= bp->b_dirtyend)
1163 bp->b_dirtyoff = bp->b_dirtyend = 0;
1164
1165 /*
1166 * If the new write will leave a contiguous dirty
1167 * area, just update the b_dirtyoff and b_dirtyend,
1168 * otherwise force a write rpc of the old dirty area.
1169 *
1170 * While it is possible to merge discontiguous writes due to
1171 * our having a B_CACHE buffer ( and thus valid read data
1172 * for the hole), we don't because it could lead to
1173 * significant cache coherency problems with multiple clients,
1174 * especially if locking is implemented later on.
1175 *
1176 * as an optimization we could theoretically maintain
1177 * a linked list of discontinuous areas, but we would still
1178 * have to commit them separately so there isn't much
1179 * advantage to it except perhaps a bit of asynchronization.
1180 */
1181
1182 if (bp->b_dirtyend > 0 &&
1183 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1184 if (bwrite(bp) == EINTR) {
1185 error = EINTR;
1186 break;
1187 }
1188 goto again;
1189 }
1190
1191 error = uiomove((char *)bp->b_data + on, n, uio);
1192
1193 /*
1194 * Since this block is being modified, it must be written
1195 * again and not just committed. Since write clustering does
1196 * not work for the stage 1 data write, only the stage 2
1197 * commit rpc, we have to clear B_CLUSTEROK as well.
1198 */
1199 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1200
1201 if (error) {
1202 bp->b_ioflags |= BIO_ERROR;
1203 brelse(bp);
1204 break;
1205 }
1206
1207 /*
1208 * Only update dirtyoff/dirtyend if not a degenerate
1209 * condition.
1210 */
1211 if (n) {
1212 if (bp->b_dirtyend > 0) {
1213 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1214 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1215 } else {
1216 bp->b_dirtyoff = on;
1217 bp->b_dirtyend = on + n;
1218 }
1219 vfs_bio_set_valid(bp, on, n);
1220 }
1221
1222 /*
1223 * If IO_SYNC do bwrite().
1224 *
1225 * IO_INVAL appears to be unused. The idea appears to be
1226 * to turn off caching in this case. Very odd. XXX
1227 */
1228 if ((ioflag & IO_SYNC)) {
1229 if (ioflag & IO_INVAL)
1230 bp->b_flags |= B_NOCACHE;
1231 error = bwrite(bp);
1232 if (error)
1233 break;
1234 } else if ((n + on) == biosize) {
1235 bp->b_flags |= B_ASYNC;
1236 (void) ncl_writebp(bp, 0, NULL);
1237 } else {
1238 bdwrite(bp);
1239 }
1240 } while (uio->uio_resid > 0 && n > 0);
1241
1242 return (error);
1243}
1244
1245/*
1246 * Get an nfs cache block.
1247 *
1248 * Allocate a new one if the block isn't currently in the cache
1249 * and return the block marked busy. If the calling process is
1250 * interrupted by a signal for an interruptible mount point, return
1251 * NULL.
1252 *
1253 * The caller must carefully deal with the possible B_INVAL state of
1254 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1255 * indirectly), so synchronous reads can be issued without worrying about
1256 * the B_INVAL state. We have to be a little more careful when dealing
1257 * with writes (see comments in nfs_write()) when extending a file past
1258 * its EOF.
1259 */
1260static struct buf *
1261nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1262{
1263 struct buf *bp;
1264 struct mount *mp;
1265 struct nfsmount *nmp;
1266
1267 mp = vp->v_mount;
1268 nmp = VFSTONFS(mp);
1269
1270 if (nmp->nm_flag & NFSMNT_INT) {
1271 sigset_t oldset;
1272
1273 newnfs_set_sigmask(td, &oldset);
1274 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1275 newnfs_restore_sigmask(td, &oldset);
1276 while (bp == NULL) {
1277 if (newnfs_sigintr(nmp, td))
1278 return (NULL);
1279 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1280 }
1281 } else {
1282 bp = getblk(vp, bn, size, 0, 0, 0);
1283 }
1284
1285 if (vp->v_type == VREG)
1286 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1287 return (bp);
1288}
1289
1290/*
1291 * Flush and invalidate all dirty buffers. If another process is already
1292 * doing the flush, just wait for completion.
1293 */
1294int
1295ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1296{
1297 struct nfsnode *np = VTONFS(vp);
1298 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1299 int error = 0, slpflag, slptimeo;
1300 int old_lock = 0;
1301
1302 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1303
1304 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1305 intrflg = 0;
1306 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1307 intrflg = 1;
1308 if (intrflg) {
1309 slpflag = NFS_PCATCH;
1310 slptimeo = 2 * hz;
1311 } else {
1312 slpflag = 0;
1313 slptimeo = 0;
1314 }
1315
1316 old_lock = ncl_upgrade_vnlock(vp);
1317 if (vp->v_iflag & VI_DOOMED) {
1318 /*
1319 * Since vgonel() uses the generic vinvalbuf() to flush
1320 * dirty buffers and it does not call this function, it
1321 * is safe to just return OK when VI_DOOMED is set.
1322 */
1323 ncl_downgrade_vnlock(vp, old_lock);
1324 return (0);
1325 }
1326
1327 /*
1328 * Now, flush as required.
1329 */
1330 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1331 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1332 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1333 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1334 /*
1335 * If the page clean was interrupted, fail the invalidation.
1336 * Not doing so, we run the risk of losing dirty pages in the
1337 * vinvalbuf() call below.
1338 */
1339 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1340 goto out;
1341 }
1342
1343 error = vinvalbuf(vp, flags, slpflag, 0);
1344 while (error) {
1345 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1346 goto out;
1347 error = vinvalbuf(vp, flags, 0, slptimeo);
1348 }
1349 mtx_lock(&np->n_mtx);
1350 if (np->n_directio_asyncwr == 0)
1351 np->n_flag &= ~NMODIFIED;
1352 mtx_unlock(&np->n_mtx);
1353out:
1354 ncl_downgrade_vnlock(vp, old_lock);
1355 return error;
1356}
1357
1358/*
1359 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1360 * This is mainly to avoid queueing async I/O requests when the nfsiods
1361 * are all hung on a dead server.
1362 *
1363 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1364 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1365 */
1366int
1367ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1368{
1369 int iod;
1370 int gotiod;
1371 int slpflag = 0;
1372 int slptimeo = 0;
1373 int error, error2;
1374
1375 /*
1376 * Commits are usually short and sweet so lets save some cpu and
1377 * leave the async daemons for more important rpc's (such as reads
1378 * and writes).
1379 */
1380 mtx_lock(&ncl_iod_mutex);
1381 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1382 (nmp->nm_bufqiods > ncl_numasync / 2)) {
1383 mtx_unlock(&ncl_iod_mutex);
1384 return(EIO);
1385 }
1386again:
1387 if (nmp->nm_flag & NFSMNT_INT)
1388 slpflag = NFS_PCATCH;
1389 gotiod = FALSE;
1390
1391 /*
1392 * Find a free iod to process this request.
1393 */
1394 for (iod = 0; iod < ncl_numasync; iod++)
1395 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1396 gotiod = TRUE;
1397 break;
1398 }
1399
1400 /*
1401 * Try to create one if none are free.
1402 */
1403 if (!gotiod)
1404 ncl_nfsiodnew();
1405 else {
1406 /*
1407 * Found one, so wake it up and tell it which
1408 * mount to process.
1409 */
1410 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1411 iod, nmp));
1412 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1413 ncl_iodmount[iod] = nmp;
1414 nmp->nm_bufqiods++;
1415 wakeup(&ncl_iodwant[iod]);
1416 }
1417
1418 /*
1419 * If none are free, we may already have an iod working on this mount
1420 * point. If so, it will process our request.
1421 */
1422 if (!gotiod) {
1423 if (nmp->nm_bufqiods > 0) {
1424 NFS_DPF(ASYNCIO,
1425 ("ncl_asyncio: %d iods are already processing mount %p\n",
1426 nmp->nm_bufqiods, nmp));
1427 gotiod = TRUE;
1428 }
1429 }
1430
1431 /*
1432 * If we have an iod which can process the request, then queue
1433 * the buffer.
1434 */
1435 if (gotiod) {
1436 /*
1437 * Ensure that the queue never grows too large. We still want
1438 * to asynchronize so we block rather then return EIO.
1439 */
1440 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1441 NFS_DPF(ASYNCIO,
1442 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1443 nmp->nm_bufqwant = TRUE;
1444 error = newnfs_msleep(td, &nmp->nm_bufq,
1445 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1446 slptimeo);
1447 if (error) {
1448 error2 = newnfs_sigintr(nmp, td);
1449 if (error2) {
1450 mtx_unlock(&ncl_iod_mutex);
1451 return (error2);
1452 }
1453 if (slpflag == NFS_PCATCH) {
1454 slpflag = 0;
1455 slptimeo = 2 * hz;
1456 }
1457 }
1458 /*
1459 * We might have lost our iod while sleeping,
1460 * so check and loop if nescessary.
1461 */
1462 goto again;
1463 }
1464
1465 /* We might have lost our nfsiod */
1466 if (nmp->nm_bufqiods == 0) {
1467 NFS_DPF(ASYNCIO,
1468 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1469 goto again;
1470 }
1471
1472 if (bp->b_iocmd == BIO_READ) {
1473 if (bp->b_rcred == NOCRED && cred != NOCRED)
1474 bp->b_rcred = crhold(cred);
1475 } else {
1476 if (bp->b_wcred == NOCRED && cred != NOCRED)
1477 bp->b_wcred = crhold(cred);
1478 }
1479
1480 if (bp->b_flags & B_REMFREE)
1481 bremfreef(bp);
1482 BUF_KERNPROC(bp);
1483 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1484 nmp->nm_bufqlen++;
1485 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1486 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1487 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1488 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1489 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1490 }
1491 mtx_unlock(&ncl_iod_mutex);
1492 return (0);
1493 }
1494
1495 mtx_unlock(&ncl_iod_mutex);
1496
1497 /*
1498 * All the iods are busy on other mounts, so return EIO to
1499 * force the caller to process the i/o synchronously.
1500 */
1501 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1502 return (EIO);
1503}
1504
1505void
1506ncl_doio_directwrite(struct buf *bp)
1507{
1508 int iomode, must_commit;
1509 struct uio *uiop = (struct uio *)bp->b_caller1;
1510 char *iov_base = uiop->uio_iov->iov_base;
1511
1512 iomode = NFSWRITE_FILESYNC;
1513 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1514 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1515 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1516 free(iov_base, M_NFSDIRECTIO);
1517 free(uiop->uio_iov, M_NFSDIRECTIO);
1518 free(uiop, M_NFSDIRECTIO);
1519 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1520 struct nfsnode *np = VTONFS(bp->b_vp);
1521 mtx_lock(&np->n_mtx);
1522 np->n_directio_asyncwr--;
1523 if (np->n_directio_asyncwr == 0) {
1524 np->n_flag &= ~NMODIFIED;
1525 if ((np->n_flag & NFSYNCWAIT)) {
1526 np->n_flag &= ~NFSYNCWAIT;
1527 wakeup((caddr_t)&np->n_directio_asyncwr);
1528 }
1529 }
1530 mtx_unlock(&np->n_mtx);
1531 }
1532 bp->b_vp = NULL;
1533 relpbuf(bp, &ncl_pbuf_freecnt);
1534}
1535
1536/*
1537 * Do an I/O operation to/from a cache block. This may be called
1538 * synchronously or from an nfsiod.
1539 */
1540int
1541ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1542 int called_from_strategy)
1543{
1544 struct uio *uiop;
1545 struct nfsnode *np;
1546 struct nfsmount *nmp;
1547 int error = 0, iomode, must_commit = 0;
1548 struct uio uio;
1549 struct iovec io;
1550 struct proc *p = td ? td->td_proc : NULL;
1551 uint8_t iocmd;
1552
1553 np = VTONFS(vp);
1554 nmp = VFSTONFS(vp->v_mount);
1555 uiop = &uio;
1556 uiop->uio_iov = &io;
1557 uiop->uio_iovcnt = 1;
1558 uiop->uio_segflg = UIO_SYSSPACE;
1559 uiop->uio_td = td;
1560
1561 /*
1562 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1563 * do this here so we do not have to do it in all the code that
1564 * calls us.
1565 */
1566 bp->b_flags &= ~B_INVAL;
1567 bp->b_ioflags &= ~BIO_ERROR;
1568
1569 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1570 iocmd = bp->b_iocmd;
1571 if (iocmd == BIO_READ) {
1572 io.iov_len = uiop->uio_resid = bp->b_bcount;
1573 io.iov_base = bp->b_data;
1574 uiop->uio_rw = UIO_READ;
1575
1576 switch (vp->v_type) {
1577 case VREG:
1578 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1579 NFSINCRGLOBAL(newnfsstats.read_bios);
1580 error = ncl_readrpc(vp, uiop, cr);
1581
1582 if (!error) {
1583 if (uiop->uio_resid) {
1584 /*
1585 * If we had a short read with no error, we must have
1586 * hit a file hole. We should zero-fill the remainder.
1587 * This can also occur if the server hits the file EOF.
1588 *
1589 * Holes used to be able to occur due to pending
1590 * writes, but that is not possible any longer.
1591 */
1592 int nread = bp->b_bcount - uiop->uio_resid;
1593 ssize_t left = uiop->uio_resid;
1594
1595 if (left > 0)
1596 bzero((char *)bp->b_data + nread, left);
1597 uiop->uio_resid = 0;
1598 }
1599 }
1600 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1601 if (p && (vp->v_vflag & VV_TEXT)) {
1602 mtx_lock(&np->n_mtx);
1603 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1604 mtx_unlock(&np->n_mtx);
1605 PROC_LOCK(p);
1606 killproc(p, "text file modification");
1607 PROC_UNLOCK(p);
1608 } else
1609 mtx_unlock(&np->n_mtx);
1610 }
1611 break;
1612 case VLNK:
1613 uiop->uio_offset = (off_t)0;
1614 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1615 error = ncl_readlinkrpc(vp, uiop, cr);
1616 break;
1617 case VDIR:
1618 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1619 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1620 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1621 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1622 if (error == NFSERR_NOTSUPP)
1623 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1624 }
1625 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1626 error = ncl_readdirrpc(vp, uiop, cr, td);
1627 /*
1628 * end-of-directory sets B_INVAL but does not generate an
1629 * error.
1630 */
1631 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1632 bp->b_flags |= B_INVAL;
1633 break;
1634 default:
1635 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1636 break;
1637 };
1638 if (error) {
1639 bp->b_ioflags |= BIO_ERROR;
1640 bp->b_error = error;
1641 }
1642 } else {
1643 /*
1644 * If we only need to commit, try to commit
1645 */
1646 if (bp->b_flags & B_NEEDCOMMIT) {
1647 int retv;
1648 off_t off;
1649
1650 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1651 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1652 bp->b_wcred, td);
1653 if (retv == 0) {
1654 bp->b_dirtyoff = bp->b_dirtyend = 0;
1655 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1656 bp->b_resid = 0;
1657 bufdone(bp);
1658 return (0);
1659 }
1660 if (retv == NFSERR_STALEWRITEVERF) {
1661 ncl_clearcommit(vp->v_mount);
1662 }
1663 }
1664
1665 /*
1666 * Setup for actual write
1667 */
1668 mtx_lock(&np->n_mtx);
1669 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1670 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1671 mtx_unlock(&np->n_mtx);
1672
1673 if (bp->b_dirtyend > bp->b_dirtyoff) {
1674 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1675 - bp->b_dirtyoff;
1676 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1677 + bp->b_dirtyoff;
1678 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1679 uiop->uio_rw = UIO_WRITE;
1680 NFSINCRGLOBAL(newnfsstats.write_bios);
1681
1682 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1683 iomode = NFSWRITE_UNSTABLE;
1684 else
1685 iomode = NFSWRITE_FILESYNC;
1686
1687 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1688 called_from_strategy);
1689
1690 /*
1691 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1692 * to cluster the buffers needing commit. This will allow
1693 * the system to submit a single commit rpc for the whole
1694 * cluster. We can do this even if the buffer is not 100%
1695 * dirty (relative to the NFS blocksize), so we optimize the
1696 * append-to-file-case.
1697 *
1698 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1699 * cleared because write clustering only works for commit
1700 * rpc's, not for the data portion of the write).
1701 */
1702
1703 if (!error && iomode == NFSWRITE_UNSTABLE) {
1704 bp->b_flags |= B_NEEDCOMMIT;
1705 if (bp->b_dirtyoff == 0
1706 && bp->b_dirtyend == bp->b_bcount)
1707 bp->b_flags |= B_CLUSTEROK;
1708 } else {
1709 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1710 }
1711
1712 /*
1713 * For an interrupted write, the buffer is still valid
1714 * and the write hasn't been pushed to the server yet,
1715 * so we can't set BIO_ERROR and report the interruption
1716 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1717 * is not relevant, so the rpc attempt is essentially
1718 * a noop. For the case of a V3 write rpc not being
1719 * committed to stable storage, the block is still
1720 * dirty and requires either a commit rpc or another
1721 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1722 * the block is reused. This is indicated by setting
1723 * the B_DELWRI and B_NEEDCOMMIT flags.
1724 *
1725 * EIO is returned by ncl_writerpc() to indicate a recoverable
1726 * write error and is handled as above, except that
1727 * B_EINTR isn't set. One cause of this is a stale stateid
1728 * error for the RPC that indicates recovery is required,
1729 * when called with called_from_strategy != 0.
1730 *
1731 * If the buffer is marked B_PAGING, it does not reside on
1732 * the vp's paging queues so we cannot call bdirty(). The
1733 * bp in this case is not an NFS cache block so we should
1734 * be safe. XXX
1735 *
1736 * The logic below breaks up errors into recoverable and
1737 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1738 * and keep the buffer around for potential write retries.
1739 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1740 * and save the error in the nfsnode. This is less than ideal
1741 * but necessary. Keeping such buffers around could potentially
1742 * cause buffer exhaustion eventually (they can never be written
1743 * out, so will get constantly be re-dirtied). It also causes
1744 * all sorts of vfs panics. For non-recoverable write errors,
1745 * also invalidate the attrcache, so we'll be forced to go over
1746 * the wire for this object, returning an error to user on next
1747 * call (most of the time).
1748 */
1749 if (error == EINTR || error == EIO || error == ETIMEDOUT
1750 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1751 int s;
1752
1753 s = splbio();
1754 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1755 if ((bp->b_flags & B_PAGING) == 0) {
1756 bdirty(bp);
1757 bp->b_flags &= ~B_DONE;
1758 }
1759 if ((error == EINTR || error == ETIMEDOUT) &&
1760 (bp->b_flags & B_ASYNC) == 0)
1761 bp->b_flags |= B_EINTR;
1762 splx(s);
1763 } else {
1764 if (error) {
1765 bp->b_ioflags |= BIO_ERROR;
1766 bp->b_flags |= B_INVAL;
1767 bp->b_error = np->n_error = error;
1768 mtx_lock(&np->n_mtx);
1769 np->n_flag |= NWRITEERR;
1770 np->n_attrstamp = 0;
1771 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1772 mtx_unlock(&np->n_mtx);
1773 }
1774 bp->b_dirtyoff = bp->b_dirtyend = 0;
1775 }
1776 } else {
1777 bp->b_resid = 0;
1778 bufdone(bp);
1779 return (0);
1780 }
1781 }
1782 bp->b_resid = uiop->uio_resid;
1783 if (must_commit)
1784 ncl_clearcommit(vp->v_mount);
1785 bufdone(bp);
1786 return (error);
1787}
1788
1789/*
1790 * Used to aid in handling ftruncate() operations on the NFS client side.
1791 * Truncation creates a number of special problems for NFS. We have to
1792 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1793 * we have to properly handle VM pages or (potentially dirty) buffers
1794 * that straddle the truncation point.
1795 */
1796
1797int
1798ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1799{
1800 struct nfsnode *np = VTONFS(vp);
1801 u_quad_t tsize;
1802 int biosize = vp->v_bufobj.bo_bsize;
1803 int error = 0;
1804
1805 mtx_lock(&np->n_mtx);
1806 tsize = np->n_size;
1807 np->n_size = nsize;
1808 mtx_unlock(&np->n_mtx);
1809
1810 if (nsize < tsize) {
1811 struct buf *bp;
1812 daddr_t lbn;
1813 int bufsize;
1814
1815 /*
1816 * vtruncbuf() doesn't get the buffer overlapping the
1817 * truncation point. We may have a B_DELWRI and/or B_CACHE
1818 * buffer that now needs to be truncated.
1819 */
1820 error = vtruncbuf(vp, cred, nsize, biosize);
1821 lbn = nsize / biosize;
1822 bufsize = nsize & (biosize - 1);
1823 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1824 if (!bp)
1825 return EINTR;
1826 if (bp->b_dirtyoff > bp->b_bcount)
1827 bp->b_dirtyoff = bp->b_bcount;
1828 if (bp->b_dirtyend > bp->b_bcount)
1829 bp->b_dirtyend = bp->b_bcount;
1830 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1831 brelse(bp);
1832 } else {
1833 vnode_pager_setsize(vp, nsize);
1834 }
1835 return(error);
1836}
1837
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: head/sys/fs/nfsclient/nfs_clbio.c 234605 2012-04-23 13:21:28Z trasz $");
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/mount.h>
46#include <sys/vmmeter.h>
47#include <sys/vnode.h>
48
49#include <vm/vm.h>
50#include <vm/vm_extern.h>
51#include <vm/vm_page.h>
52#include <vm/vm_object.h>
53#include <vm/vm_pager.h>
54#include <vm/vnode_pager.h>
55
56#include <fs/nfs/nfsport.h>
57#include <fs/nfsclient/nfsmount.h>
58#include <fs/nfsclient/nfs.h>
59#include <fs/nfsclient/nfsnode.h>
60#include <fs/nfsclient/nfs_kdtrace.h>
61
62extern int newnfs_directio_allow_mmap;
63extern struct nfsstats newnfsstats;
64extern struct mtx ncl_iod_mutex;
65extern int ncl_numasync;
66extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
67extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
68extern int newnfs_directio_enable;
69extern int nfs_keep_dirty_on_error;
70
71int ncl_pbuf_freecnt = -1; /* start out unlimited */
72
73static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
74 struct thread *td);
75static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
76 struct ucred *cred, int ioflag);
77
78/*
79 * Vnode op for VM getpages.
80 */
81int
82ncl_getpages(struct vop_getpages_args *ap)
83{
84 int i, error, nextoff, size, toff, count, npages;
85 struct uio uio;
86 struct iovec iov;
87 vm_offset_t kva;
88 struct buf *bp;
89 struct vnode *vp;
90 struct thread *td;
91 struct ucred *cred;
92 struct nfsmount *nmp;
93 vm_object_t object;
94 vm_page_t *pages;
95 struct nfsnode *np;
96
97 vp = ap->a_vp;
98 np = VTONFS(vp);
99 td = curthread; /* XXX */
100 cred = curthread->td_ucred; /* XXX */
101 nmp = VFSTONFS(vp->v_mount);
102 pages = ap->a_m;
103 count = ap->a_count;
104
105 if ((object = vp->v_object) == NULL) {
106 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
107 return (VM_PAGER_ERROR);
108 }
109
110 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
111 mtx_lock(&np->n_mtx);
112 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
113 mtx_unlock(&np->n_mtx);
114 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n");
115 return (VM_PAGER_ERROR);
116 } else
117 mtx_unlock(&np->n_mtx);
118 }
119
120 mtx_lock(&nmp->nm_mtx);
121 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
122 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
123 mtx_unlock(&nmp->nm_mtx);
124 /* We'll never get here for v4, because we always have fsinfo */
125 (void)ncl_fsinfo(nmp, vp, cred, td);
126 } else
127 mtx_unlock(&nmp->nm_mtx);
128
129 npages = btoc(count);
130
131 /*
132 * If the requested page is partially valid, just return it and
133 * allow the pager to zero-out the blanks. Partially valid pages
134 * can only occur at the file EOF.
135 */
136 VM_OBJECT_LOCK(object);
137 if (pages[ap->a_reqpage]->valid != 0) {
138 for (i = 0; i < npages; ++i) {
139 if (i != ap->a_reqpage) {
140 vm_page_lock(pages[i]);
141 vm_page_free(pages[i]);
142 vm_page_unlock(pages[i]);
143 }
144 }
145 VM_OBJECT_UNLOCK(object);
146 return (0);
147 }
148 VM_OBJECT_UNLOCK(object);
149
150 /*
151 * We use only the kva address for the buffer, but this is extremely
152 * convienient and fast.
153 */
154 bp = getpbuf(&ncl_pbuf_freecnt);
155
156 kva = (vm_offset_t) bp->b_data;
157 pmap_qenter(kva, pages, npages);
158 PCPU_INC(cnt.v_vnodein);
159 PCPU_ADD(cnt.v_vnodepgsin, npages);
160
161 iov.iov_base = (caddr_t) kva;
162 iov.iov_len = count;
163 uio.uio_iov = &iov;
164 uio.uio_iovcnt = 1;
165 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
166 uio.uio_resid = count;
167 uio.uio_segflg = UIO_SYSSPACE;
168 uio.uio_rw = UIO_READ;
169 uio.uio_td = td;
170
171 error = ncl_readrpc(vp, &uio, cred);
172 pmap_qremove(kva, npages);
173
174 relpbuf(bp, &ncl_pbuf_freecnt);
175
176 if (error && (uio.uio_resid == count)) {
177 ncl_printf("nfs_getpages: error %d\n", error);
178 VM_OBJECT_LOCK(object);
179 for (i = 0; i < npages; ++i) {
180 if (i != ap->a_reqpage) {
181 vm_page_lock(pages[i]);
182 vm_page_free(pages[i]);
183 vm_page_unlock(pages[i]);
184 }
185 }
186 VM_OBJECT_UNLOCK(object);
187 return (VM_PAGER_ERROR);
188 }
189
190 /*
191 * Calculate the number of bytes read and validate only that number
192 * of bytes. Note that due to pending writes, size may be 0. This
193 * does not mean that the remaining data is invalid!
194 */
195
196 size = count - uio.uio_resid;
197 VM_OBJECT_LOCK(object);
198 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
199 vm_page_t m;
200 nextoff = toff + PAGE_SIZE;
201 m = pages[i];
202
203 if (nextoff <= size) {
204 /*
205 * Read operation filled an entire page
206 */
207 m->valid = VM_PAGE_BITS_ALL;
208 KASSERT(m->dirty == 0,
209 ("nfs_getpages: page %p is dirty", m));
210 } else if (size > toff) {
211 /*
212 * Read operation filled a partial page.
213 */
214 m->valid = 0;
215 vm_page_set_valid_range(m, 0, size - toff);
216 KASSERT(m->dirty == 0,
217 ("nfs_getpages: page %p is dirty", m));
218 } else {
219 /*
220 * Read operation was short. If no error occured
221 * we may have hit a zero-fill section. We simply
222 * leave valid set to 0.
223 */
224 ;
225 }
226 if (i != ap->a_reqpage) {
227 /*
228 * Whether or not to leave the page activated is up in
229 * the air, but we should put the page on a page queue
230 * somewhere (it already is in the object). Result:
231 * It appears that emperical results show that
232 * deactivating pages is best.
233 */
234
235 /*
236 * Just in case someone was asking for this page we
237 * now tell them that it is ok to use.
238 */
239 if (!error) {
240 if (m->oflags & VPO_WANTED) {
241 vm_page_lock(m);
242 vm_page_activate(m);
243 vm_page_unlock(m);
244 } else {
245 vm_page_lock(m);
246 vm_page_deactivate(m);
247 vm_page_unlock(m);
248 }
249 vm_page_wakeup(m);
250 } else {
251 vm_page_lock(m);
252 vm_page_free(m);
253 vm_page_unlock(m);
254 }
255 }
256 }
257 VM_OBJECT_UNLOCK(object);
258 return (0);
259}
260
261/*
262 * Vnode op for VM putpages.
263 */
264int
265ncl_putpages(struct vop_putpages_args *ap)
266{
267 struct uio uio;
268 struct iovec iov;
269 vm_offset_t kva;
270 struct buf *bp;
271 int iomode, must_commit, i, error, npages, count;
272 off_t offset;
273 int *rtvals;
274 struct vnode *vp;
275 struct thread *td;
276 struct ucred *cred;
277 struct nfsmount *nmp;
278 struct nfsnode *np;
279 vm_page_t *pages;
280
281 vp = ap->a_vp;
282 np = VTONFS(vp);
283 td = curthread; /* XXX */
284 cred = curthread->td_ucred; /* XXX */
285 nmp = VFSTONFS(vp->v_mount);
286 pages = ap->a_m;
287 count = ap->a_count;
288 rtvals = ap->a_rtvals;
289 npages = btoc(count);
290 offset = IDX_TO_OFF(pages[0]->pindex);
291
292 mtx_lock(&nmp->nm_mtx);
293 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
294 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
295 mtx_unlock(&nmp->nm_mtx);
296 (void)ncl_fsinfo(nmp, vp, cred, td);
297 } else
298 mtx_unlock(&nmp->nm_mtx);
299
300 mtx_lock(&np->n_mtx);
301 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
302 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
303 mtx_unlock(&np->n_mtx);
304 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
305 mtx_lock(&np->n_mtx);
306 }
307
308 for (i = 0; i < npages; i++)
309 rtvals[i] = VM_PAGER_ERROR;
310
311 /*
312 * When putting pages, do not extend file past EOF.
313 */
314 if (offset + count > np->n_size) {
315 count = np->n_size - offset;
316 if (count < 0)
317 count = 0;
318 }
319 mtx_unlock(&np->n_mtx);
320
321 /*
322 * We use only the kva address for the buffer, but this is extremely
323 * convienient and fast.
324 */
325 bp = getpbuf(&ncl_pbuf_freecnt);
326
327 kva = (vm_offset_t) bp->b_data;
328 pmap_qenter(kva, pages, npages);
329 PCPU_INC(cnt.v_vnodeout);
330 PCPU_ADD(cnt.v_vnodepgsout, count);
331
332 iov.iov_base = (caddr_t) kva;
333 iov.iov_len = count;
334 uio.uio_iov = &iov;
335 uio.uio_iovcnt = 1;
336 uio.uio_offset = offset;
337 uio.uio_resid = count;
338 uio.uio_segflg = UIO_SYSSPACE;
339 uio.uio_rw = UIO_WRITE;
340 uio.uio_td = td;
341
342 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
343 iomode = NFSWRITE_UNSTABLE;
344 else
345 iomode = NFSWRITE_FILESYNC;
346
347 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
348
349 pmap_qremove(kva, npages);
350 relpbuf(bp, &ncl_pbuf_freecnt);
351
352 if (error == 0 || !nfs_keep_dirty_on_error) {
353 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
354 if (must_commit)
355 ncl_clearcommit(vp->v_mount);
356 }
357 return rtvals[0];
358}
359
360/*
361 * For nfs, cache consistency can only be maintained approximately.
362 * Although RFC1094 does not specify the criteria, the following is
363 * believed to be compatible with the reference port.
364 * For nfs:
365 * If the file's modify time on the server has changed since the
366 * last read rpc or you have written to the file,
367 * you may have lost data cache consistency with the
368 * server, so flush all of the file's data out of the cache.
369 * Then force a getattr rpc to ensure that you have up to date
370 * attributes.
371 * NB: This implies that cache data can be read when up to
372 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
373 * attributes this could be forced by setting n_attrstamp to 0 before
374 * the VOP_GETATTR() call.
375 */
376static inline int
377nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
378{
379 int error = 0;
380 struct vattr vattr;
381 struct nfsnode *np = VTONFS(vp);
382 int old_lock;
383
384 /*
385 * Grab the exclusive lock before checking whether the cache is
386 * consistent.
387 * XXX - We can make this cheaper later (by acquiring cheaper locks).
388 * But for now, this suffices.
389 */
390 old_lock = ncl_upgrade_vnlock(vp);
391 if (vp->v_iflag & VI_DOOMED) {
392 ncl_downgrade_vnlock(vp, old_lock);
393 return (EBADF);
394 }
395
396 mtx_lock(&np->n_mtx);
397 if (np->n_flag & NMODIFIED) {
398 mtx_unlock(&np->n_mtx);
399 if (vp->v_type != VREG) {
400 if (vp->v_type != VDIR)
401 panic("nfs: bioread, not dir");
402 ncl_invaldir(vp);
403 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
404 if (error)
405 goto out;
406 }
407 np->n_attrstamp = 0;
408 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
409 error = VOP_GETATTR(vp, &vattr, cred);
410 if (error)
411 goto out;
412 mtx_lock(&np->n_mtx);
413 np->n_mtime = vattr.va_mtime;
414 mtx_unlock(&np->n_mtx);
415 } else {
416 mtx_unlock(&np->n_mtx);
417 error = VOP_GETATTR(vp, &vattr, cred);
418 if (error)
419 return (error);
420 mtx_lock(&np->n_mtx);
421 if ((np->n_flag & NSIZECHANGED)
422 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
423 mtx_unlock(&np->n_mtx);
424 if (vp->v_type == VDIR)
425 ncl_invaldir(vp);
426 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
427 if (error)
428 goto out;
429 mtx_lock(&np->n_mtx);
430 np->n_mtime = vattr.va_mtime;
431 np->n_flag &= ~NSIZECHANGED;
432 }
433 mtx_unlock(&np->n_mtx);
434 }
435out:
436 ncl_downgrade_vnlock(vp, old_lock);
437 return error;
438}
439
440/*
441 * Vnode op for read using bio
442 */
443int
444ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
445{
446 struct nfsnode *np = VTONFS(vp);
447 int biosize, i;
448 struct buf *bp, *rabp;
449 struct thread *td;
450 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
451 daddr_t lbn, rabn;
452 int bcount;
453 int seqcount;
454 int nra, error = 0, n = 0, on = 0;
455 off_t tmp_off;
456
457 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
458 if (uio->uio_resid == 0)
459 return (0);
460 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
461 return (EINVAL);
462 td = uio->uio_td;
463
464 mtx_lock(&nmp->nm_mtx);
465 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
466 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
467 mtx_unlock(&nmp->nm_mtx);
468 (void)ncl_fsinfo(nmp, vp, cred, td);
469 mtx_lock(&nmp->nm_mtx);
470 }
471 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
472 (void) newnfs_iosize(nmp);
473
474 tmp_off = uio->uio_offset + uio->uio_resid;
475 if (vp->v_type != VDIR &&
476 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
477 mtx_unlock(&nmp->nm_mtx);
478 return (EFBIG);
479 }
480 mtx_unlock(&nmp->nm_mtx);
481
482 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
483 /* No caching/ no readaheads. Just read data into the user buffer */
484 return ncl_readrpc(vp, uio, cred);
485
486 biosize = vp->v_bufobj.bo_bsize;
487 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
488
489 error = nfs_bioread_check_cons(vp, td, cred);
490 if (error)
491 return error;
492
493 do {
494 u_quad_t nsize;
495
496 mtx_lock(&np->n_mtx);
497 nsize = np->n_size;
498 mtx_unlock(&np->n_mtx);
499
500 switch (vp->v_type) {
501 case VREG:
502 NFSINCRGLOBAL(newnfsstats.biocache_reads);
503 lbn = uio->uio_offset / biosize;
504 on = uio->uio_offset & (biosize - 1);
505
506 /*
507 * Start the read ahead(s), as required.
508 */
509 if (nmp->nm_readahead > 0) {
510 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
511 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
512 rabn = lbn + 1 + nra;
513 if (incore(&vp->v_bufobj, rabn) == NULL) {
514 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
515 if (!rabp) {
516 error = newnfs_sigintr(nmp, td);
517 return (error ? error : EINTR);
518 }
519 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
520 rabp->b_flags |= B_ASYNC;
521 rabp->b_iocmd = BIO_READ;
522 vfs_busy_pages(rabp, 0);
523 if (ncl_asyncio(nmp, rabp, cred, td)) {
524 rabp->b_flags |= B_INVAL;
525 rabp->b_ioflags |= BIO_ERROR;
526 vfs_unbusy_pages(rabp);
527 brelse(rabp);
528 break;
529 }
530 } else {
531 brelse(rabp);
532 }
533 }
534 }
535 }
536
537 /* Note that bcount is *not* DEV_BSIZE aligned. */
538 bcount = biosize;
539 if ((off_t)lbn * biosize >= nsize) {
540 bcount = 0;
541 } else if ((off_t)(lbn + 1) * biosize > nsize) {
542 bcount = nsize - (off_t)lbn * biosize;
543 }
544 bp = nfs_getcacheblk(vp, lbn, bcount, td);
545
546 if (!bp) {
547 error = newnfs_sigintr(nmp, td);
548 return (error ? error : EINTR);
549 }
550
551 /*
552 * If B_CACHE is not set, we must issue the read. If this
553 * fails, we return an error.
554 */
555
556 if ((bp->b_flags & B_CACHE) == 0) {
557 bp->b_iocmd = BIO_READ;
558 vfs_busy_pages(bp, 0);
559 error = ncl_doio(vp, bp, cred, td, 0);
560 if (error) {
561 brelse(bp);
562 return (error);
563 }
564 }
565
566 /*
567 * on is the offset into the current bp. Figure out how many
568 * bytes we can copy out of the bp. Note that bcount is
569 * NOT DEV_BSIZE aligned.
570 *
571 * Then figure out how many bytes we can copy into the uio.
572 */
573
574 n = 0;
575 if (on < bcount)
576 n = MIN((unsigned)(bcount - on), uio->uio_resid);
577 break;
578 case VLNK:
579 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
580 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
581 if (!bp) {
582 error = newnfs_sigintr(nmp, td);
583 return (error ? error : EINTR);
584 }
585 if ((bp->b_flags & B_CACHE) == 0) {
586 bp->b_iocmd = BIO_READ;
587 vfs_busy_pages(bp, 0);
588 error = ncl_doio(vp, bp, cred, td, 0);
589 if (error) {
590 bp->b_ioflags |= BIO_ERROR;
591 brelse(bp);
592 return (error);
593 }
594 }
595 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
596 on = 0;
597 break;
598 case VDIR:
599 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
600 if (np->n_direofoffset
601 && uio->uio_offset >= np->n_direofoffset) {
602 return (0);
603 }
604 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
605 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
606 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
607 if (!bp) {
608 error = newnfs_sigintr(nmp, td);
609 return (error ? error : EINTR);
610 }
611 if ((bp->b_flags & B_CACHE) == 0) {
612 bp->b_iocmd = BIO_READ;
613 vfs_busy_pages(bp, 0);
614 error = ncl_doio(vp, bp, cred, td, 0);
615 if (error) {
616 brelse(bp);
617 }
618 while (error == NFSERR_BAD_COOKIE) {
619 ncl_invaldir(vp);
620 error = ncl_vinvalbuf(vp, 0, td, 1);
621 /*
622 * Yuck! The directory has been modified on the
623 * server. The only way to get the block is by
624 * reading from the beginning to get all the
625 * offset cookies.
626 *
627 * Leave the last bp intact unless there is an error.
628 * Loop back up to the while if the error is another
629 * NFSERR_BAD_COOKIE (double yuch!).
630 */
631 for (i = 0; i <= lbn && !error; i++) {
632 if (np->n_direofoffset
633 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
634 return (0);
635 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
636 if (!bp) {
637 error = newnfs_sigintr(nmp, td);
638 return (error ? error : EINTR);
639 }
640 if ((bp->b_flags & B_CACHE) == 0) {
641 bp->b_iocmd = BIO_READ;
642 vfs_busy_pages(bp, 0);
643 error = ncl_doio(vp, bp, cred, td, 0);
644 /*
645 * no error + B_INVAL == directory EOF,
646 * use the block.
647 */
648 if (error == 0 && (bp->b_flags & B_INVAL))
649 break;
650 }
651 /*
652 * An error will throw away the block and the
653 * for loop will break out. If no error and this
654 * is not the block we want, we throw away the
655 * block and go for the next one via the for loop.
656 */
657 if (error || i < lbn)
658 brelse(bp);
659 }
660 }
661 /*
662 * The above while is repeated if we hit another cookie
663 * error. If we hit an error and it wasn't a cookie error,
664 * we give up.
665 */
666 if (error)
667 return (error);
668 }
669
670 /*
671 * If not eof and read aheads are enabled, start one.
672 * (You need the current block first, so that you have the
673 * directory offset cookie of the next block.)
674 */
675 if (nmp->nm_readahead > 0 &&
676 (bp->b_flags & B_INVAL) == 0 &&
677 (np->n_direofoffset == 0 ||
678 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
679 incore(&vp->v_bufobj, lbn + 1) == NULL) {
680 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
681 if (rabp) {
682 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
683 rabp->b_flags |= B_ASYNC;
684 rabp->b_iocmd = BIO_READ;
685 vfs_busy_pages(rabp, 0);
686 if (ncl_asyncio(nmp, rabp, cred, td)) {
687 rabp->b_flags |= B_INVAL;
688 rabp->b_ioflags |= BIO_ERROR;
689 vfs_unbusy_pages(rabp);
690 brelse(rabp);
691 }
692 } else {
693 brelse(rabp);
694 }
695 }
696 }
697 /*
698 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
699 * chopped for the EOF condition, we cannot tell how large
700 * NFS directories are going to be until we hit EOF. So
701 * an NFS directory buffer is *not* chopped to its EOF. Now,
702 * it just so happens that b_resid will effectively chop it
703 * to EOF. *BUT* this information is lost if the buffer goes
704 * away and is reconstituted into a B_CACHE state ( due to
705 * being VMIO ) later. So we keep track of the directory eof
706 * in np->n_direofoffset and chop it off as an extra step
707 * right here.
708 */
709 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
710 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
711 n = np->n_direofoffset - uio->uio_offset;
712 break;
713 default:
714 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
715 bp = NULL;
716 break;
717 };
718
719 if (n > 0) {
720 error = uiomove(bp->b_data + on, (int)n, uio);
721 }
722 if (vp->v_type == VLNK)
723 n = 0;
724 if (bp != NULL)
725 brelse(bp);
726 } while (error == 0 && uio->uio_resid > 0 && n > 0);
727 return (error);
728}
729
730/*
731 * The NFS write path cannot handle iovecs with len > 1. So we need to
732 * break up iovecs accordingly (restricting them to wsize).
733 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
734 * For the ASYNC case, 2 copies are needed. The first a copy from the
735 * user buffer to a staging buffer and then a second copy from the staging
736 * buffer to mbufs. This can be optimized by copying from the user buffer
737 * directly into mbufs and passing the chain down, but that requires a
738 * fair amount of re-working of the relevant codepaths (and can be done
739 * later).
740 */
741static int
742nfs_directio_write(vp, uiop, cred, ioflag)
743 struct vnode *vp;
744 struct uio *uiop;
745 struct ucred *cred;
746 int ioflag;
747{
748 int error;
749 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
750 struct thread *td = uiop->uio_td;
751 int size;
752 int wsize;
753
754 mtx_lock(&nmp->nm_mtx);
755 wsize = nmp->nm_wsize;
756 mtx_unlock(&nmp->nm_mtx);
757 if (ioflag & IO_SYNC) {
758 int iomode, must_commit;
759 struct uio uio;
760 struct iovec iov;
761do_sync:
762 while (uiop->uio_resid > 0) {
763 size = MIN(uiop->uio_resid, wsize);
764 size = MIN(uiop->uio_iov->iov_len, size);
765 iov.iov_base = uiop->uio_iov->iov_base;
766 iov.iov_len = size;
767 uio.uio_iov = &iov;
768 uio.uio_iovcnt = 1;
769 uio.uio_offset = uiop->uio_offset;
770 uio.uio_resid = size;
771 uio.uio_segflg = UIO_USERSPACE;
772 uio.uio_rw = UIO_WRITE;
773 uio.uio_td = td;
774 iomode = NFSWRITE_FILESYNC;
775 error = ncl_writerpc(vp, &uio, cred, &iomode,
776 &must_commit, 0);
777 KASSERT((must_commit == 0),
778 ("ncl_directio_write: Did not commit write"));
779 if (error)
780 return (error);
781 uiop->uio_offset += size;
782 uiop->uio_resid -= size;
783 if (uiop->uio_iov->iov_len <= size) {
784 uiop->uio_iovcnt--;
785 uiop->uio_iov++;
786 } else {
787 uiop->uio_iov->iov_base =
788 (char *)uiop->uio_iov->iov_base + size;
789 uiop->uio_iov->iov_len -= size;
790 }
791 }
792 } else {
793 struct uio *t_uio;
794 struct iovec *t_iov;
795 struct buf *bp;
796
797 /*
798 * Break up the write into blocksize chunks and hand these
799 * over to nfsiod's for write back.
800 * Unfortunately, this incurs a copy of the data. Since
801 * the user could modify the buffer before the write is
802 * initiated.
803 *
804 * The obvious optimization here is that one of the 2 copies
805 * in the async write path can be eliminated by copying the
806 * data here directly into mbufs and passing the mbuf chain
807 * down. But that will require a fair amount of re-working
808 * of the code and can be done if there's enough interest
809 * in NFS directio access.
810 */
811 while (uiop->uio_resid > 0) {
812 size = MIN(uiop->uio_resid, wsize);
813 size = MIN(uiop->uio_iov->iov_len, size);
814 bp = getpbuf(&ncl_pbuf_freecnt);
815 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
816 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
817 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
818 t_iov->iov_len = size;
819 t_uio->uio_iov = t_iov;
820 t_uio->uio_iovcnt = 1;
821 t_uio->uio_offset = uiop->uio_offset;
822 t_uio->uio_resid = size;
823 t_uio->uio_segflg = UIO_SYSSPACE;
824 t_uio->uio_rw = UIO_WRITE;
825 t_uio->uio_td = td;
826 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
827 uiop->uio_segflg == UIO_SYSSPACE,
828 ("nfs_directio_write: Bad uio_segflg"));
829 if (uiop->uio_segflg == UIO_USERSPACE) {
830 error = copyin(uiop->uio_iov->iov_base,
831 t_iov->iov_base, size);
832 if (error != 0)
833 goto err_free;
834 } else
835 /*
836 * UIO_SYSSPACE may never happen, but handle
837 * it just in case it does.
838 */
839 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
840 size);
841 bp->b_flags |= B_DIRECT;
842 bp->b_iocmd = BIO_WRITE;
843 if (cred != NOCRED) {
844 crhold(cred);
845 bp->b_wcred = cred;
846 } else
847 bp->b_wcred = NOCRED;
848 bp->b_caller1 = (void *)t_uio;
849 bp->b_vp = vp;
850 error = ncl_asyncio(nmp, bp, NOCRED, td);
851err_free:
852 if (error) {
853 free(t_iov->iov_base, M_NFSDIRECTIO);
854 free(t_iov, M_NFSDIRECTIO);
855 free(t_uio, M_NFSDIRECTIO);
856 bp->b_vp = NULL;
857 relpbuf(bp, &ncl_pbuf_freecnt);
858 if (error == EINTR)
859 return (error);
860 goto do_sync;
861 }
862 uiop->uio_offset += size;
863 uiop->uio_resid -= size;
864 if (uiop->uio_iov->iov_len <= size) {
865 uiop->uio_iovcnt--;
866 uiop->uio_iov++;
867 } else {
868 uiop->uio_iov->iov_base =
869 (char *)uiop->uio_iov->iov_base + size;
870 uiop->uio_iov->iov_len -= size;
871 }
872 }
873 }
874 return (0);
875}
876
877/*
878 * Vnode op for write using bio
879 */
880int
881ncl_write(struct vop_write_args *ap)
882{
883 int biosize;
884 struct uio *uio = ap->a_uio;
885 struct thread *td = uio->uio_td;
886 struct vnode *vp = ap->a_vp;
887 struct nfsnode *np = VTONFS(vp);
888 struct ucred *cred = ap->a_cred;
889 int ioflag = ap->a_ioflag;
890 struct buf *bp;
891 struct vattr vattr;
892 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
893 daddr_t lbn;
894 int bcount;
895 int n, on, error = 0;
896 off_t tmp_off;
897
898 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
899 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
900 ("ncl_write proc"));
901 if (vp->v_type != VREG)
902 return (EIO);
903 mtx_lock(&np->n_mtx);
904 if (np->n_flag & NWRITEERR) {
905 np->n_flag &= ~NWRITEERR;
906 mtx_unlock(&np->n_mtx);
907 return (np->n_error);
908 } else
909 mtx_unlock(&np->n_mtx);
910 mtx_lock(&nmp->nm_mtx);
911 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
912 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
913 mtx_unlock(&nmp->nm_mtx);
914 (void)ncl_fsinfo(nmp, vp, cred, td);
915 mtx_lock(&nmp->nm_mtx);
916 }
917 if (nmp->nm_wsize == 0)
918 (void) newnfs_iosize(nmp);
919 mtx_unlock(&nmp->nm_mtx);
920
921 /*
922 * Synchronously flush pending buffers if we are in synchronous
923 * mode or if we are appending.
924 */
925 if (ioflag & (IO_APPEND | IO_SYNC)) {
926 mtx_lock(&np->n_mtx);
927 if (np->n_flag & NMODIFIED) {
928 mtx_unlock(&np->n_mtx);
929#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
930 /*
931 * Require non-blocking, synchronous writes to
932 * dirty files to inform the program it needs
933 * to fsync(2) explicitly.
934 */
935 if (ioflag & IO_NDELAY)
936 return (EAGAIN);
937#endif
938flush_and_restart:
939 np->n_attrstamp = 0;
940 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
941 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
942 if (error)
943 return (error);
944 } else
945 mtx_unlock(&np->n_mtx);
946 }
947
948 /*
949 * If IO_APPEND then load uio_offset. We restart here if we cannot
950 * get the append lock.
951 */
952 if (ioflag & IO_APPEND) {
953 np->n_attrstamp = 0;
954 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
955 error = VOP_GETATTR(vp, &vattr, cred);
956 if (error)
957 return (error);
958 mtx_lock(&np->n_mtx);
959 uio->uio_offset = np->n_size;
960 mtx_unlock(&np->n_mtx);
961 }
962
963 if (uio->uio_offset < 0)
964 return (EINVAL);
965 tmp_off = uio->uio_offset + uio->uio_resid;
966 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
967 return (EFBIG);
968 if (uio->uio_resid == 0)
969 return (0);
970
971 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
972 return nfs_directio_write(vp, uio, cred, ioflag);
973
974 /*
975 * Maybe this should be above the vnode op call, but so long as
976 * file servers have no limits, i don't think it matters
977 */
978 if (vn_rlimit_fsize(vp, uio, td))
979 return (EFBIG);
980
981 biosize = vp->v_bufobj.bo_bsize;
982 /*
983 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
984 * would exceed the local maximum per-file write commit size when
985 * combined with those, we must decide whether to flush,
986 * go synchronous, or return error. We don't bother checking
987 * IO_UNIT -- we just make all writes atomic anyway, as there's
988 * no point optimizing for something that really won't ever happen.
989 */
990 if (!(ioflag & IO_SYNC)) {
991 int nflag;
992
993 mtx_lock(&np->n_mtx);
994 nflag = np->n_flag;
995 mtx_unlock(&np->n_mtx);
996 int needrestart = 0;
997 if (nmp->nm_wcommitsize < uio->uio_resid) {
998 /*
999 * If this request could not possibly be completed
1000 * without exceeding the maximum outstanding write
1001 * commit size, see if we can convert it into a
1002 * synchronous write operation.
1003 */
1004 if (ioflag & IO_NDELAY)
1005 return (EAGAIN);
1006 ioflag |= IO_SYNC;
1007 if (nflag & NMODIFIED)
1008 needrestart = 1;
1009 } else if (nflag & NMODIFIED) {
1010 int wouldcommit = 0;
1011 BO_LOCK(&vp->v_bufobj);
1012 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1013 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1014 b_bobufs) {
1015 if (bp->b_flags & B_NEEDCOMMIT)
1016 wouldcommit += bp->b_bcount;
1017 }
1018 }
1019 BO_UNLOCK(&vp->v_bufobj);
1020 /*
1021 * Since we're not operating synchronously and
1022 * bypassing the buffer cache, we are in a commit
1023 * and holding all of these buffers whether
1024 * transmitted or not. If not limited, this
1025 * will lead to the buffer cache deadlocking,
1026 * as no one else can flush our uncommitted buffers.
1027 */
1028 wouldcommit += uio->uio_resid;
1029 /*
1030 * If we would initially exceed the maximum
1031 * outstanding write commit size, flush and restart.
1032 */
1033 if (wouldcommit > nmp->nm_wcommitsize)
1034 needrestart = 1;
1035 }
1036 if (needrestart)
1037 goto flush_and_restart;
1038 }
1039
1040 do {
1041 NFSINCRGLOBAL(newnfsstats.biocache_writes);
1042 lbn = uio->uio_offset / biosize;
1043 on = uio->uio_offset & (biosize-1);
1044 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1045again:
1046 /*
1047 * Handle direct append and file extension cases, calculate
1048 * unaligned buffer size.
1049 */
1050 mtx_lock(&np->n_mtx);
1051 if (uio->uio_offset == np->n_size && n) {
1052 mtx_unlock(&np->n_mtx);
1053 /*
1054 * Get the buffer (in its pre-append state to maintain
1055 * B_CACHE if it was previously set). Resize the
1056 * nfsnode after we have locked the buffer to prevent
1057 * readers from reading garbage.
1058 */
1059 bcount = on;
1060 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1061
1062 if (bp != NULL) {
1063 long save;
1064
1065 mtx_lock(&np->n_mtx);
1066 np->n_size = uio->uio_offset + n;
1067 np->n_flag |= NMODIFIED;
1068 vnode_pager_setsize(vp, np->n_size);
1069 mtx_unlock(&np->n_mtx);
1070
1071 save = bp->b_flags & B_CACHE;
1072 bcount += n;
1073 allocbuf(bp, bcount);
1074 bp->b_flags |= save;
1075 }
1076 } else {
1077 /*
1078 * Obtain the locked cache block first, and then
1079 * adjust the file's size as appropriate.
1080 */
1081 bcount = on + n;
1082 if ((off_t)lbn * biosize + bcount < np->n_size) {
1083 if ((off_t)(lbn + 1) * biosize < np->n_size)
1084 bcount = biosize;
1085 else
1086 bcount = np->n_size - (off_t)lbn * biosize;
1087 }
1088 mtx_unlock(&np->n_mtx);
1089 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1090 mtx_lock(&np->n_mtx);
1091 if (uio->uio_offset + n > np->n_size) {
1092 np->n_size = uio->uio_offset + n;
1093 np->n_flag |= NMODIFIED;
1094 vnode_pager_setsize(vp, np->n_size);
1095 }
1096 mtx_unlock(&np->n_mtx);
1097 }
1098
1099 if (!bp) {
1100 error = newnfs_sigintr(nmp, td);
1101 if (!error)
1102 error = EINTR;
1103 break;
1104 }
1105
1106 /*
1107 * Issue a READ if B_CACHE is not set. In special-append
1108 * mode, B_CACHE is based on the buffer prior to the write
1109 * op and is typically set, avoiding the read. If a read
1110 * is required in special append mode, the server will
1111 * probably send us a short-read since we extended the file
1112 * on our end, resulting in b_resid == 0 and, thusly,
1113 * B_CACHE getting set.
1114 *
1115 * We can also avoid issuing the read if the write covers
1116 * the entire buffer. We have to make sure the buffer state
1117 * is reasonable in this case since we will not be initiating
1118 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1119 * more information.
1120 *
1121 * B_CACHE may also be set due to the buffer being cached
1122 * normally.
1123 */
1124
1125 if (on == 0 && n == bcount) {
1126 bp->b_flags |= B_CACHE;
1127 bp->b_flags &= ~B_INVAL;
1128 bp->b_ioflags &= ~BIO_ERROR;
1129 }
1130
1131 if ((bp->b_flags & B_CACHE) == 0) {
1132 bp->b_iocmd = BIO_READ;
1133 vfs_busy_pages(bp, 0);
1134 error = ncl_doio(vp, bp, cred, td, 0);
1135 if (error) {
1136 brelse(bp);
1137 break;
1138 }
1139 }
1140 if (bp->b_wcred == NOCRED)
1141 bp->b_wcred = crhold(cred);
1142 mtx_lock(&np->n_mtx);
1143 np->n_flag |= NMODIFIED;
1144 mtx_unlock(&np->n_mtx);
1145
1146 /*
1147 * If dirtyend exceeds file size, chop it down. This should
1148 * not normally occur but there is an append race where it
1149 * might occur XXX, so we log it.
1150 *
1151 * If the chopping creates a reverse-indexed or degenerate
1152 * situation with dirtyoff/end, we 0 both of them.
1153 */
1154
1155 if (bp->b_dirtyend > bcount) {
1156 ncl_printf("NFS append race @%lx:%d\n",
1157 (long)bp->b_blkno * DEV_BSIZE,
1158 bp->b_dirtyend - bcount);
1159 bp->b_dirtyend = bcount;
1160 }
1161
1162 if (bp->b_dirtyoff >= bp->b_dirtyend)
1163 bp->b_dirtyoff = bp->b_dirtyend = 0;
1164
1165 /*
1166 * If the new write will leave a contiguous dirty
1167 * area, just update the b_dirtyoff and b_dirtyend,
1168 * otherwise force a write rpc of the old dirty area.
1169 *
1170 * While it is possible to merge discontiguous writes due to
1171 * our having a B_CACHE buffer ( and thus valid read data
1172 * for the hole), we don't because it could lead to
1173 * significant cache coherency problems with multiple clients,
1174 * especially if locking is implemented later on.
1175 *
1176 * as an optimization we could theoretically maintain
1177 * a linked list of discontinuous areas, but we would still
1178 * have to commit them separately so there isn't much
1179 * advantage to it except perhaps a bit of asynchronization.
1180 */
1181
1182 if (bp->b_dirtyend > 0 &&
1183 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1184 if (bwrite(bp) == EINTR) {
1185 error = EINTR;
1186 break;
1187 }
1188 goto again;
1189 }
1190
1191 error = uiomove((char *)bp->b_data + on, n, uio);
1192
1193 /*
1194 * Since this block is being modified, it must be written
1195 * again and not just committed. Since write clustering does
1196 * not work for the stage 1 data write, only the stage 2
1197 * commit rpc, we have to clear B_CLUSTEROK as well.
1198 */
1199 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1200
1201 if (error) {
1202 bp->b_ioflags |= BIO_ERROR;
1203 brelse(bp);
1204 break;
1205 }
1206
1207 /*
1208 * Only update dirtyoff/dirtyend if not a degenerate
1209 * condition.
1210 */
1211 if (n) {
1212 if (bp->b_dirtyend > 0) {
1213 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1214 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1215 } else {
1216 bp->b_dirtyoff = on;
1217 bp->b_dirtyend = on + n;
1218 }
1219 vfs_bio_set_valid(bp, on, n);
1220 }
1221
1222 /*
1223 * If IO_SYNC do bwrite().
1224 *
1225 * IO_INVAL appears to be unused. The idea appears to be
1226 * to turn off caching in this case. Very odd. XXX
1227 */
1228 if ((ioflag & IO_SYNC)) {
1229 if (ioflag & IO_INVAL)
1230 bp->b_flags |= B_NOCACHE;
1231 error = bwrite(bp);
1232 if (error)
1233 break;
1234 } else if ((n + on) == biosize) {
1235 bp->b_flags |= B_ASYNC;
1236 (void) ncl_writebp(bp, 0, NULL);
1237 } else {
1238 bdwrite(bp);
1239 }
1240 } while (uio->uio_resid > 0 && n > 0);
1241
1242 return (error);
1243}
1244
1245/*
1246 * Get an nfs cache block.
1247 *
1248 * Allocate a new one if the block isn't currently in the cache
1249 * and return the block marked busy. If the calling process is
1250 * interrupted by a signal for an interruptible mount point, return
1251 * NULL.
1252 *
1253 * The caller must carefully deal with the possible B_INVAL state of
1254 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1255 * indirectly), so synchronous reads can be issued without worrying about
1256 * the B_INVAL state. We have to be a little more careful when dealing
1257 * with writes (see comments in nfs_write()) when extending a file past
1258 * its EOF.
1259 */
1260static struct buf *
1261nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1262{
1263 struct buf *bp;
1264 struct mount *mp;
1265 struct nfsmount *nmp;
1266
1267 mp = vp->v_mount;
1268 nmp = VFSTONFS(mp);
1269
1270 if (nmp->nm_flag & NFSMNT_INT) {
1271 sigset_t oldset;
1272
1273 newnfs_set_sigmask(td, &oldset);
1274 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1275 newnfs_restore_sigmask(td, &oldset);
1276 while (bp == NULL) {
1277 if (newnfs_sigintr(nmp, td))
1278 return (NULL);
1279 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1280 }
1281 } else {
1282 bp = getblk(vp, bn, size, 0, 0, 0);
1283 }
1284
1285 if (vp->v_type == VREG)
1286 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1287 return (bp);
1288}
1289
1290/*
1291 * Flush and invalidate all dirty buffers. If another process is already
1292 * doing the flush, just wait for completion.
1293 */
1294int
1295ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1296{
1297 struct nfsnode *np = VTONFS(vp);
1298 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1299 int error = 0, slpflag, slptimeo;
1300 int old_lock = 0;
1301
1302 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1303
1304 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1305 intrflg = 0;
1306 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1307 intrflg = 1;
1308 if (intrflg) {
1309 slpflag = NFS_PCATCH;
1310 slptimeo = 2 * hz;
1311 } else {
1312 slpflag = 0;
1313 slptimeo = 0;
1314 }
1315
1316 old_lock = ncl_upgrade_vnlock(vp);
1317 if (vp->v_iflag & VI_DOOMED) {
1318 /*
1319 * Since vgonel() uses the generic vinvalbuf() to flush
1320 * dirty buffers and it does not call this function, it
1321 * is safe to just return OK when VI_DOOMED is set.
1322 */
1323 ncl_downgrade_vnlock(vp, old_lock);
1324 return (0);
1325 }
1326
1327 /*
1328 * Now, flush as required.
1329 */
1330 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1331 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1332 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1333 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1334 /*
1335 * If the page clean was interrupted, fail the invalidation.
1336 * Not doing so, we run the risk of losing dirty pages in the
1337 * vinvalbuf() call below.
1338 */
1339 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1340 goto out;
1341 }
1342
1343 error = vinvalbuf(vp, flags, slpflag, 0);
1344 while (error) {
1345 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1346 goto out;
1347 error = vinvalbuf(vp, flags, 0, slptimeo);
1348 }
1349 mtx_lock(&np->n_mtx);
1350 if (np->n_directio_asyncwr == 0)
1351 np->n_flag &= ~NMODIFIED;
1352 mtx_unlock(&np->n_mtx);
1353out:
1354 ncl_downgrade_vnlock(vp, old_lock);
1355 return error;
1356}
1357
1358/*
1359 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1360 * This is mainly to avoid queueing async I/O requests when the nfsiods
1361 * are all hung on a dead server.
1362 *
1363 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1364 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1365 */
1366int
1367ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1368{
1369 int iod;
1370 int gotiod;
1371 int slpflag = 0;
1372 int slptimeo = 0;
1373 int error, error2;
1374
1375 /*
1376 * Commits are usually short and sweet so lets save some cpu and
1377 * leave the async daemons for more important rpc's (such as reads
1378 * and writes).
1379 */
1380 mtx_lock(&ncl_iod_mutex);
1381 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1382 (nmp->nm_bufqiods > ncl_numasync / 2)) {
1383 mtx_unlock(&ncl_iod_mutex);
1384 return(EIO);
1385 }
1386again:
1387 if (nmp->nm_flag & NFSMNT_INT)
1388 slpflag = NFS_PCATCH;
1389 gotiod = FALSE;
1390
1391 /*
1392 * Find a free iod to process this request.
1393 */
1394 for (iod = 0; iod < ncl_numasync; iod++)
1395 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1396 gotiod = TRUE;
1397 break;
1398 }
1399
1400 /*
1401 * Try to create one if none are free.
1402 */
1403 if (!gotiod)
1404 ncl_nfsiodnew();
1405 else {
1406 /*
1407 * Found one, so wake it up and tell it which
1408 * mount to process.
1409 */
1410 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1411 iod, nmp));
1412 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1413 ncl_iodmount[iod] = nmp;
1414 nmp->nm_bufqiods++;
1415 wakeup(&ncl_iodwant[iod]);
1416 }
1417
1418 /*
1419 * If none are free, we may already have an iod working on this mount
1420 * point. If so, it will process our request.
1421 */
1422 if (!gotiod) {
1423 if (nmp->nm_bufqiods > 0) {
1424 NFS_DPF(ASYNCIO,
1425 ("ncl_asyncio: %d iods are already processing mount %p\n",
1426 nmp->nm_bufqiods, nmp));
1427 gotiod = TRUE;
1428 }
1429 }
1430
1431 /*
1432 * If we have an iod which can process the request, then queue
1433 * the buffer.
1434 */
1435 if (gotiod) {
1436 /*
1437 * Ensure that the queue never grows too large. We still want
1438 * to asynchronize so we block rather then return EIO.
1439 */
1440 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1441 NFS_DPF(ASYNCIO,
1442 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1443 nmp->nm_bufqwant = TRUE;
1444 error = newnfs_msleep(td, &nmp->nm_bufq,
1445 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1446 slptimeo);
1447 if (error) {
1448 error2 = newnfs_sigintr(nmp, td);
1449 if (error2) {
1450 mtx_unlock(&ncl_iod_mutex);
1451 return (error2);
1452 }
1453 if (slpflag == NFS_PCATCH) {
1454 slpflag = 0;
1455 slptimeo = 2 * hz;
1456 }
1457 }
1458 /*
1459 * We might have lost our iod while sleeping,
1460 * so check and loop if nescessary.
1461 */
1462 goto again;
1463 }
1464
1465 /* We might have lost our nfsiod */
1466 if (nmp->nm_bufqiods == 0) {
1467 NFS_DPF(ASYNCIO,
1468 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1469 goto again;
1470 }
1471
1472 if (bp->b_iocmd == BIO_READ) {
1473 if (bp->b_rcred == NOCRED && cred != NOCRED)
1474 bp->b_rcred = crhold(cred);
1475 } else {
1476 if (bp->b_wcred == NOCRED && cred != NOCRED)
1477 bp->b_wcred = crhold(cred);
1478 }
1479
1480 if (bp->b_flags & B_REMFREE)
1481 bremfreef(bp);
1482 BUF_KERNPROC(bp);
1483 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1484 nmp->nm_bufqlen++;
1485 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1486 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1487 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1488 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1489 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1490 }
1491 mtx_unlock(&ncl_iod_mutex);
1492 return (0);
1493 }
1494
1495 mtx_unlock(&ncl_iod_mutex);
1496
1497 /*
1498 * All the iods are busy on other mounts, so return EIO to
1499 * force the caller to process the i/o synchronously.
1500 */
1501 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1502 return (EIO);
1503}
1504
1505void
1506ncl_doio_directwrite(struct buf *bp)
1507{
1508 int iomode, must_commit;
1509 struct uio *uiop = (struct uio *)bp->b_caller1;
1510 char *iov_base = uiop->uio_iov->iov_base;
1511
1512 iomode = NFSWRITE_FILESYNC;
1513 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1514 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1515 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1516 free(iov_base, M_NFSDIRECTIO);
1517 free(uiop->uio_iov, M_NFSDIRECTIO);
1518 free(uiop, M_NFSDIRECTIO);
1519 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1520 struct nfsnode *np = VTONFS(bp->b_vp);
1521 mtx_lock(&np->n_mtx);
1522 np->n_directio_asyncwr--;
1523 if (np->n_directio_asyncwr == 0) {
1524 np->n_flag &= ~NMODIFIED;
1525 if ((np->n_flag & NFSYNCWAIT)) {
1526 np->n_flag &= ~NFSYNCWAIT;
1527 wakeup((caddr_t)&np->n_directio_asyncwr);
1528 }
1529 }
1530 mtx_unlock(&np->n_mtx);
1531 }
1532 bp->b_vp = NULL;
1533 relpbuf(bp, &ncl_pbuf_freecnt);
1534}
1535
1536/*
1537 * Do an I/O operation to/from a cache block. This may be called
1538 * synchronously or from an nfsiod.
1539 */
1540int
1541ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1542 int called_from_strategy)
1543{
1544 struct uio *uiop;
1545 struct nfsnode *np;
1546 struct nfsmount *nmp;
1547 int error = 0, iomode, must_commit = 0;
1548 struct uio uio;
1549 struct iovec io;
1550 struct proc *p = td ? td->td_proc : NULL;
1551 uint8_t iocmd;
1552
1553 np = VTONFS(vp);
1554 nmp = VFSTONFS(vp->v_mount);
1555 uiop = &uio;
1556 uiop->uio_iov = &io;
1557 uiop->uio_iovcnt = 1;
1558 uiop->uio_segflg = UIO_SYSSPACE;
1559 uiop->uio_td = td;
1560
1561 /*
1562 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1563 * do this here so we do not have to do it in all the code that
1564 * calls us.
1565 */
1566 bp->b_flags &= ~B_INVAL;
1567 bp->b_ioflags &= ~BIO_ERROR;
1568
1569 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1570 iocmd = bp->b_iocmd;
1571 if (iocmd == BIO_READ) {
1572 io.iov_len = uiop->uio_resid = bp->b_bcount;
1573 io.iov_base = bp->b_data;
1574 uiop->uio_rw = UIO_READ;
1575
1576 switch (vp->v_type) {
1577 case VREG:
1578 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1579 NFSINCRGLOBAL(newnfsstats.read_bios);
1580 error = ncl_readrpc(vp, uiop, cr);
1581
1582 if (!error) {
1583 if (uiop->uio_resid) {
1584 /*
1585 * If we had a short read with no error, we must have
1586 * hit a file hole. We should zero-fill the remainder.
1587 * This can also occur if the server hits the file EOF.
1588 *
1589 * Holes used to be able to occur due to pending
1590 * writes, but that is not possible any longer.
1591 */
1592 int nread = bp->b_bcount - uiop->uio_resid;
1593 ssize_t left = uiop->uio_resid;
1594
1595 if (left > 0)
1596 bzero((char *)bp->b_data + nread, left);
1597 uiop->uio_resid = 0;
1598 }
1599 }
1600 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1601 if (p && (vp->v_vflag & VV_TEXT)) {
1602 mtx_lock(&np->n_mtx);
1603 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1604 mtx_unlock(&np->n_mtx);
1605 PROC_LOCK(p);
1606 killproc(p, "text file modification");
1607 PROC_UNLOCK(p);
1608 } else
1609 mtx_unlock(&np->n_mtx);
1610 }
1611 break;
1612 case VLNK:
1613 uiop->uio_offset = (off_t)0;
1614 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1615 error = ncl_readlinkrpc(vp, uiop, cr);
1616 break;
1617 case VDIR:
1618 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1619 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1620 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1621 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1622 if (error == NFSERR_NOTSUPP)
1623 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1624 }
1625 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1626 error = ncl_readdirrpc(vp, uiop, cr, td);
1627 /*
1628 * end-of-directory sets B_INVAL but does not generate an
1629 * error.
1630 */
1631 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1632 bp->b_flags |= B_INVAL;
1633 break;
1634 default:
1635 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1636 break;
1637 };
1638 if (error) {
1639 bp->b_ioflags |= BIO_ERROR;
1640 bp->b_error = error;
1641 }
1642 } else {
1643 /*
1644 * If we only need to commit, try to commit
1645 */
1646 if (bp->b_flags & B_NEEDCOMMIT) {
1647 int retv;
1648 off_t off;
1649
1650 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1651 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1652 bp->b_wcred, td);
1653 if (retv == 0) {
1654 bp->b_dirtyoff = bp->b_dirtyend = 0;
1655 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1656 bp->b_resid = 0;
1657 bufdone(bp);
1658 return (0);
1659 }
1660 if (retv == NFSERR_STALEWRITEVERF) {
1661 ncl_clearcommit(vp->v_mount);
1662 }
1663 }
1664
1665 /*
1666 * Setup for actual write
1667 */
1668 mtx_lock(&np->n_mtx);
1669 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1670 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1671 mtx_unlock(&np->n_mtx);
1672
1673 if (bp->b_dirtyend > bp->b_dirtyoff) {
1674 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1675 - bp->b_dirtyoff;
1676 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1677 + bp->b_dirtyoff;
1678 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1679 uiop->uio_rw = UIO_WRITE;
1680 NFSINCRGLOBAL(newnfsstats.write_bios);
1681
1682 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1683 iomode = NFSWRITE_UNSTABLE;
1684 else
1685 iomode = NFSWRITE_FILESYNC;
1686
1687 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1688 called_from_strategy);
1689
1690 /*
1691 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1692 * to cluster the buffers needing commit. This will allow
1693 * the system to submit a single commit rpc for the whole
1694 * cluster. We can do this even if the buffer is not 100%
1695 * dirty (relative to the NFS blocksize), so we optimize the
1696 * append-to-file-case.
1697 *
1698 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1699 * cleared because write clustering only works for commit
1700 * rpc's, not for the data portion of the write).
1701 */
1702
1703 if (!error && iomode == NFSWRITE_UNSTABLE) {
1704 bp->b_flags |= B_NEEDCOMMIT;
1705 if (bp->b_dirtyoff == 0
1706 && bp->b_dirtyend == bp->b_bcount)
1707 bp->b_flags |= B_CLUSTEROK;
1708 } else {
1709 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1710 }
1711
1712 /*
1713 * For an interrupted write, the buffer is still valid
1714 * and the write hasn't been pushed to the server yet,
1715 * so we can't set BIO_ERROR and report the interruption
1716 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1717 * is not relevant, so the rpc attempt is essentially
1718 * a noop. For the case of a V3 write rpc not being
1719 * committed to stable storage, the block is still
1720 * dirty and requires either a commit rpc or another
1721 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1722 * the block is reused. This is indicated by setting
1723 * the B_DELWRI and B_NEEDCOMMIT flags.
1724 *
1725 * EIO is returned by ncl_writerpc() to indicate a recoverable
1726 * write error and is handled as above, except that
1727 * B_EINTR isn't set. One cause of this is a stale stateid
1728 * error for the RPC that indicates recovery is required,
1729 * when called with called_from_strategy != 0.
1730 *
1731 * If the buffer is marked B_PAGING, it does not reside on
1732 * the vp's paging queues so we cannot call bdirty(). The
1733 * bp in this case is not an NFS cache block so we should
1734 * be safe. XXX
1735 *
1736 * The logic below breaks up errors into recoverable and
1737 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1738 * and keep the buffer around for potential write retries.
1739 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1740 * and save the error in the nfsnode. This is less than ideal
1741 * but necessary. Keeping such buffers around could potentially
1742 * cause buffer exhaustion eventually (they can never be written
1743 * out, so will get constantly be re-dirtied). It also causes
1744 * all sorts of vfs panics. For non-recoverable write errors,
1745 * also invalidate the attrcache, so we'll be forced to go over
1746 * the wire for this object, returning an error to user on next
1747 * call (most of the time).
1748 */
1749 if (error == EINTR || error == EIO || error == ETIMEDOUT
1750 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1751 int s;
1752
1753 s = splbio();
1754 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1755 if ((bp->b_flags & B_PAGING) == 0) {
1756 bdirty(bp);
1757 bp->b_flags &= ~B_DONE;
1758 }
1759 if ((error == EINTR || error == ETIMEDOUT) &&
1760 (bp->b_flags & B_ASYNC) == 0)
1761 bp->b_flags |= B_EINTR;
1762 splx(s);
1763 } else {
1764 if (error) {
1765 bp->b_ioflags |= BIO_ERROR;
1766 bp->b_flags |= B_INVAL;
1767 bp->b_error = np->n_error = error;
1768 mtx_lock(&np->n_mtx);
1769 np->n_flag |= NWRITEERR;
1770 np->n_attrstamp = 0;
1771 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1772 mtx_unlock(&np->n_mtx);
1773 }
1774 bp->b_dirtyoff = bp->b_dirtyend = 0;
1775 }
1776 } else {
1777 bp->b_resid = 0;
1778 bufdone(bp);
1779 return (0);
1780 }
1781 }
1782 bp->b_resid = uiop->uio_resid;
1783 if (must_commit)
1784 ncl_clearcommit(vp->v_mount);
1785 bufdone(bp);
1786 return (error);
1787}
1788
1789/*
1790 * Used to aid in handling ftruncate() operations on the NFS client side.
1791 * Truncation creates a number of special problems for NFS. We have to
1792 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1793 * we have to properly handle VM pages or (potentially dirty) buffers
1794 * that straddle the truncation point.
1795 */
1796
1797int
1798ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1799{
1800 struct nfsnode *np = VTONFS(vp);
1801 u_quad_t tsize;
1802 int biosize = vp->v_bufobj.bo_bsize;
1803 int error = 0;
1804
1805 mtx_lock(&np->n_mtx);
1806 tsize = np->n_size;
1807 np->n_size = nsize;
1808 mtx_unlock(&np->n_mtx);
1809
1810 if (nsize < tsize) {
1811 struct buf *bp;
1812 daddr_t lbn;
1813 int bufsize;
1814
1815 /*
1816 * vtruncbuf() doesn't get the buffer overlapping the
1817 * truncation point. We may have a B_DELWRI and/or B_CACHE
1818 * buffer that now needs to be truncated.
1819 */
1820 error = vtruncbuf(vp, cred, nsize, biosize);
1821 lbn = nsize / biosize;
1822 bufsize = nsize & (biosize - 1);
1823 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1824 if (!bp)
1825 return EINTR;
1826 if (bp->b_dirtyoff > bp->b_bcount)
1827 bp->b_dirtyoff = bp->b_bcount;
1828 if (bp->b_dirtyend > bp->b_bcount)
1829 bp->b_dirtyend = bp->b_bcount;
1830 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1831 brelse(bp);
1832 } else {
1833 vnode_pager_setsize(vp, nsize);
1834 }
1835 return(error);
1836}
1837