1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 *
7 * This code is derived from software contributed to Berkeley by
8 * Rick Macklem at The University of Guelph.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 */
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/bio.h>
38 #include <sys/buf.h>
39 #include <sys/kernel.h>
40 #include <sys/mount.h>
41 #include <sys/rwlock.h>
42 #include <sys/vmmeter.h>
43 #include <sys/vnode.h>
44
45 #include <vm/vm.h>
46 #include <vm/vm_param.h>
47 #include <vm/vm_extern.h>
48 #include <vm/vm_page.h>
49 #include <vm/vm_object.h>
50 #include <vm/vm_pager.h>
51 #include <vm/vnode_pager.h>
52
53 #include <fs/nfs/nfsport.h>
54 #include <fs/nfsclient/nfsmount.h>
55 #include <fs/nfsclient/nfs.h>
56 #include <fs/nfsclient/nfsnode.h>
57 #include <fs/nfsclient/nfs_kdtrace.h>
58
59 extern int newnfs_directio_allow_mmap;
60 extern struct nfsstatsv1 nfsstatsv1;
61 extern struct mtx ncl_iod_mutex;
62 extern int ncl_numasync;
63 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
64 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
65 extern int newnfs_directio_enable;
66 extern int nfs_keep_dirty_on_error;
67
68 uma_zone_t ncl_pbuf_zone;
69
70 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
71 struct thread *td);
72 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
73 struct ucred *cred, int ioflag);
74
75 /*
76 * Vnode op for VM getpages.
77 */
78 SYSCTL_DECL(_vfs_nfs);
79 static int use_buf_pager = 1;
80 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
81 &use_buf_pager, 0,
82 "Use buffer pager instead of direct readrpc call");
83
84 static daddr_t
ncl_gbp_getblkno(struct vnode * vp,vm_ooffset_t off)85 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
86 {
87
88 return (off / vp->v_bufobj.bo_bsize);
89 }
90
91 static int
ncl_gbp_getblksz(struct vnode * vp,daddr_t lbn,long * sz)92 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn, long *sz)
93 {
94 struct nfsnode *np;
95 u_quad_t nsize;
96 int biosize, bcount;
97
98 np = VTONFS(vp);
99 NFSLOCKNODE(np);
100 nsize = np->n_size;
101 NFSUNLOCKNODE(np);
102
103 biosize = vp->v_bufobj.bo_bsize;
104 bcount = biosize;
105 if ((off_t)lbn * biosize >= nsize)
106 bcount = 0;
107 else if ((off_t)(lbn + 1) * biosize > nsize)
108 bcount = nsize - (off_t)lbn * biosize;
109 *sz = bcount;
110 return (0);
111 }
112
113 int
ncl_getpages(struct vop_getpages_args * ap)114 ncl_getpages(struct vop_getpages_args *ap)
115 {
116 int i, error, nextoff, size, toff, count, npages;
117 struct uio uio;
118 struct iovec iov;
119 vm_offset_t kva;
120 struct buf *bp;
121 struct vnode *vp;
122 struct thread *td;
123 struct ucred *cred;
124 struct nfsmount *nmp;
125 vm_object_t object;
126 vm_page_t *pages;
127 struct nfsnode *np;
128
129 vp = ap->a_vp;
130 np = VTONFS(vp);
131 td = curthread;
132 cred = curthread->td_ucred;
133 nmp = VFSTONFS(vp->v_mount);
134 pages = ap->a_m;
135 npages = ap->a_count;
136
137 if ((object = vp->v_object) == NULL) {
138 printf("ncl_getpages: called with non-merged cache vnode\n");
139 return (VM_PAGER_ERROR);
140 }
141
142 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
143 NFSLOCKNODE(np);
144 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
145 NFSUNLOCKNODE(np);
146 printf("ncl_getpages: called on non-cacheable vnode\n");
147 return (VM_PAGER_ERROR);
148 } else
149 NFSUNLOCKNODE(np);
150 }
151
152 mtx_lock(&nmp->nm_mtx);
153 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
154 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
155 mtx_unlock(&nmp->nm_mtx);
156 /* We'll never get here for v4, because we always have fsinfo */
157 (void)ncl_fsinfo(nmp, vp, cred, td);
158 } else
159 mtx_unlock(&nmp->nm_mtx);
160
161 if (use_buf_pager)
162 return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
163 ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
164
165 /*
166 * If the requested page is partially valid, just return it and
167 * allow the pager to zero-out the blanks. Partially valid pages
168 * can only occur at the file EOF.
169 *
170 * XXXGL: is that true for NFS, where short read can occur???
171 */
172 VM_OBJECT_WLOCK(object);
173 if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
174 goto out;
175 VM_OBJECT_WUNLOCK(object);
176
177 /*
178 * We use only the kva address for the buffer, but this is extremely
179 * convenient and fast.
180 */
181 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
182
183 kva = (vm_offset_t) bp->b_data;
184 pmap_qenter(kva, pages, npages);
185 VM_CNT_INC(v_vnodein);
186 VM_CNT_ADD(v_vnodepgsin, npages);
187
188 count = npages << PAGE_SHIFT;
189 iov.iov_base = (caddr_t) kva;
190 iov.iov_len = count;
191 uio.uio_iov = &iov;
192 uio.uio_iovcnt = 1;
193 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
194 uio.uio_resid = count;
195 uio.uio_segflg = UIO_SYSSPACE;
196 uio.uio_rw = UIO_READ;
197 uio.uio_td = td;
198
199 error = ncl_readrpc(vp, &uio, cred);
200 pmap_qremove(kva, npages);
201
202 uma_zfree(ncl_pbuf_zone, bp);
203
204 if (error && (uio.uio_resid == count)) {
205 printf("ncl_getpages: error %d\n", error);
206 return (VM_PAGER_ERROR);
207 }
208
209 /*
210 * Calculate the number of bytes read and validate only that number
211 * of bytes. Note that due to pending writes, size may be 0. This
212 * does not mean that the remaining data is invalid!
213 */
214
215 size = count - uio.uio_resid;
216 VM_OBJECT_WLOCK(object);
217 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
218 vm_page_t m;
219 nextoff = toff + PAGE_SIZE;
220 m = pages[i];
221
222 if (nextoff <= size) {
223 /*
224 * Read operation filled an entire page
225 */
226 vm_page_valid(m);
227 KASSERT(m->dirty == 0,
228 ("nfs_getpages: page %p is dirty", m));
229 } else if (size > toff) {
230 /*
231 * Read operation filled a partial page.
232 */
233 vm_page_invalid(m);
234 vm_page_set_valid_range(m, 0, size - toff);
235 KASSERT(m->dirty == 0,
236 ("nfs_getpages: page %p is dirty", m));
237 } else {
238 /*
239 * Read operation was short. If no error
240 * occurred we may have hit a zero-fill
241 * section. We leave valid set to 0, and page
242 * is freed by vm_page_readahead_finish() if
243 * its index is not equal to requested, or
244 * page is zeroed and set valid by
245 * vm_pager_get_pages() for requested page.
246 */
247 ;
248 }
249 }
250 out:
251 VM_OBJECT_WUNLOCK(object);
252 if (ap->a_rbehind)
253 *ap->a_rbehind = 0;
254 if (ap->a_rahead)
255 *ap->a_rahead = 0;
256 return (VM_PAGER_OK);
257 }
258
259 /*
260 * Vnode op for VM putpages.
261 */
262 int
ncl_putpages(struct vop_putpages_args * ap)263 ncl_putpages(struct vop_putpages_args *ap)
264 {
265 struct uio uio;
266 struct iovec iov;
267 int i, error, npages, count;
268 off_t offset;
269 int *rtvals;
270 struct vnode *vp;
271 struct thread *td;
272 struct ucred *cred;
273 struct nfsmount *nmp;
274 struct nfsnode *np;
275 vm_page_t *pages;
276
277 vp = ap->a_vp;
278 np = VTONFS(vp);
279 td = curthread; /* XXX */
280 /* Set the cred to n_writecred for the write rpcs. */
281 if (np->n_writecred != NULL)
282 cred = crhold(np->n_writecred);
283 else
284 cred = crhold(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 NFSLOCKNODE(np);
301 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
302 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
303 NFSUNLOCKNODE(np);
304 printf("ncl_putpages: called on noncache-able vnode\n");
305 NFSLOCKNODE(np);
306 }
307 /*
308 * When putting pages, do not extend file past EOF.
309 */
310 if (offset + count > np->n_size) {
311 count = np->n_size - offset;
312 if (count < 0)
313 count = 0;
314 }
315 NFSUNLOCKNODE(np);
316
317 for (i = 0; i < npages; i++)
318 rtvals[i] = VM_PAGER_ERROR;
319
320 VM_CNT_INC(v_vnodeout);
321 VM_CNT_ADD(v_vnodepgsout, count);
322
323 iov.iov_base = unmapped_buf;
324 iov.iov_len = count;
325 uio.uio_iov = &iov;
326 uio.uio_iovcnt = 1;
327 uio.uio_offset = offset;
328 uio.uio_resid = count;
329 uio.uio_segflg = UIO_NOCOPY;
330 uio.uio_rw = UIO_WRITE;
331 uio.uio_td = td;
332
333 error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
334 cred);
335 crfree(cred);
336
337 if (error == 0 || !nfs_keep_dirty_on_error) {
338 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
339 np->n_size - offset, npages * PAGE_SIZE);
340 }
341 return (rtvals[0]);
342 }
343
344 /*
345 * For nfs, cache consistency can only be maintained approximately.
346 * Although RFC1094 does not specify the criteria, the following is
347 * believed to be compatible with the reference port.
348 * For nfs:
349 * If the file's modify time on the server has changed since the
350 * last read rpc or you have written to the file,
351 * you may have lost data cache consistency with the
352 * server, so flush all of the file's data out of the cache.
353 * Then force a getattr rpc to ensure that you have up to date
354 * attributes.
355 * NB: This implies that cache data can be read when up to
356 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
357 * attributes this could be forced by setting n_attrstamp to 0 before
358 * the VOP_GETATTR() call.
359 */
360 static inline int
nfs_bioread_check_cons(struct vnode * vp,struct thread * td,struct ucred * cred)361 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
362 {
363 int error = 0;
364 struct vattr vattr;
365 struct nfsnode *np = VTONFS(vp);
366 bool old_lock;
367
368 /*
369 * Ensure the exclusive access to the node before checking
370 * whether the cache is consistent.
371 */
372 old_lock = ncl_excl_start(vp);
373 NFSLOCKNODE(np);
374 if (np->n_flag & NMODIFIED) {
375 NFSUNLOCKNODE(np);
376 if (vp->v_type != VREG) {
377 if (vp->v_type != VDIR)
378 panic("nfs: bioread, not dir");
379 ncl_invaldir(vp);
380 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
381 if (error != 0)
382 goto out;
383 }
384 np->n_attrstamp = 0;
385 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
386 error = VOP_GETATTR(vp, &vattr, cred);
387 if (error)
388 goto out;
389 NFSLOCKNODE(np);
390 np->n_mtime = vattr.va_mtime;
391 NFSUNLOCKNODE(np);
392 } else {
393 NFSUNLOCKNODE(np);
394 error = VOP_GETATTR(vp, &vattr, cred);
395 if (error)
396 goto out;
397 NFSLOCKNODE(np);
398 if ((np->n_flag & NSIZECHANGED)
399 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
400 NFSUNLOCKNODE(np);
401 if (vp->v_type == VDIR)
402 ncl_invaldir(vp);
403 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
404 if (error != 0)
405 goto out;
406 NFSLOCKNODE(np);
407 np->n_mtime = vattr.va_mtime;
408 np->n_flag &= ~NSIZECHANGED;
409 }
410 NFSUNLOCKNODE(np);
411 }
412 out:
413 ncl_excl_finish(vp, old_lock);
414 return (error);
415 }
416
417 static bool
ncl_bioread_dora(struct vnode * vp)418 ncl_bioread_dora(struct vnode *vp)
419 {
420 vm_object_t obj;
421
422 obj = vp->v_object;
423 if (obj == NULL)
424 return (true);
425 return (!vm_object_mightbedirty(vp->v_object) &&
426 vp->v_object->un_pager.vnp.writemappings == 0);
427 }
428
429 /*
430 * Vnode op for read using bio
431 */
432 int
ncl_bioread(struct vnode * vp,struct uio * uio,int ioflag,struct ucred * cred)433 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
434 {
435 struct nfsnode *np = VTONFS(vp);
436 struct buf *bp, *rabp;
437 struct thread *td;
438 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
439 daddr_t lbn, rabn;
440 int biosize, bcount, error, i, n, nra, on, save2, seqcount;
441 off_t tmp_off;
442
443 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
444 if (uio->uio_resid == 0)
445 return (0);
446 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
447 return (EINVAL);
448 td = uio->uio_td;
449
450 mtx_lock(&nmp->nm_mtx);
451 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
452 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
453 mtx_unlock(&nmp->nm_mtx);
454 (void)ncl_fsinfo(nmp, vp, cred, td);
455 mtx_lock(&nmp->nm_mtx);
456 }
457 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
458 (void) newnfs_iosize(nmp);
459
460 tmp_off = uio->uio_offset + uio->uio_resid;
461 if (vp->v_type != VDIR &&
462 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
463 mtx_unlock(&nmp->nm_mtx);
464 return (EFBIG);
465 }
466 mtx_unlock(&nmp->nm_mtx);
467
468 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
469 /* No caching/ no readaheads. Just read data into the user buffer */
470 return ncl_readrpc(vp, uio, cred);
471
472 n = 0;
473 on = 0;
474 biosize = vp->v_bufobj.bo_bsize;
475 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
476
477 error = nfs_bioread_check_cons(vp, td, cred);
478 if (error)
479 return error;
480
481 save2 = curthread_pflags2_set(TDP2_SBPAGES);
482 do {
483 u_quad_t nsize;
484
485 NFSLOCKNODE(np);
486 nsize = np->n_size;
487 NFSUNLOCKNODE(np);
488
489 switch (vp->v_type) {
490 case VREG:
491 NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
492 lbn = uio->uio_offset / biosize;
493 on = uio->uio_offset - (lbn * biosize);
494
495 /*
496 * Start the read ahead(s), as required. Do not do
497 * read-ahead if there are writeable mappings, since
498 * unlocked read by nfsiod could obliterate changes
499 * done by userspace.
500 */
501 if (nmp->nm_readahead > 0 && ncl_bioread_dora(vp)) {
502 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
503 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
504 rabn = lbn + 1 + nra;
505 if (incore(&vp->v_bufobj, rabn) == NULL) {
506 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
507 if (!rabp) {
508 error = newnfs_sigintr(nmp, td);
509 if (error == 0)
510 error = EINTR;
511 goto out;
512 }
513 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
514 rabp->b_flags |= B_ASYNC;
515 rabp->b_iocmd = BIO_READ;
516 vfs_busy_pages(rabp, 0);
517 if (ncl_asyncio(nmp, rabp, cred, td)) {
518 rabp->b_flags |= B_INVAL;
519 rabp->b_ioflags |= BIO_ERROR;
520 vfs_unbusy_pages(rabp);
521 brelse(rabp);
522 break;
523 }
524 } else {
525 brelse(rabp);
526 }
527 }
528 }
529 }
530
531 /* Note that bcount is *not* DEV_BSIZE aligned. */
532 bcount = biosize;
533 if ((off_t)lbn * biosize >= nsize) {
534 bcount = 0;
535 } else if ((off_t)(lbn + 1) * biosize > nsize) {
536 bcount = nsize - (off_t)lbn * biosize;
537 }
538 bp = nfs_getcacheblk(vp, lbn, bcount, td);
539
540 if (!bp) {
541 error = newnfs_sigintr(nmp, td);
542 if (error == 0)
543 error = EINTR;
544 goto out;
545 }
546
547 /*
548 * If B_CACHE is not set, we must issue the read. If this
549 * fails, we return an error.
550 */
551
552 if ((bp->b_flags & B_CACHE) == 0) {
553 bp->b_iocmd = BIO_READ;
554 vfs_busy_pages(bp, 0);
555 error = ncl_doio(vp, bp, cred, td, 0);
556 if (error) {
557 brelse(bp);
558 goto out;
559 }
560 }
561
562 /*
563 * on is the offset into the current bp. Figure out how many
564 * bytes we can copy out of the bp. Note that bcount is
565 * NOT DEV_BSIZE aligned.
566 *
567 * Then figure out how many bytes we can copy into the uio.
568 */
569
570 n = 0;
571 if (on < bcount)
572 n = MIN((unsigned)(bcount - on), uio->uio_resid);
573 break;
574 case VLNK:
575 NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
576 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
577 if (!bp) {
578 error = newnfs_sigintr(nmp, td);
579 if (error == 0)
580 error = EINTR;
581 goto out;
582 }
583 if ((bp->b_flags & B_CACHE) == 0) {
584 bp->b_iocmd = BIO_READ;
585 vfs_busy_pages(bp, 0);
586 error = ncl_doio(vp, bp, cred, td, 0);
587 if (error) {
588 bp->b_ioflags |= BIO_ERROR;
589 brelse(bp);
590 goto out;
591 }
592 }
593 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
594 on = 0;
595 break;
596 case VDIR:
597 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
598 NFSLOCKNODE(np);
599 if (np->n_direofoffset
600 && uio->uio_offset >= np->n_direofoffset) {
601 NFSUNLOCKNODE(np);
602 error = 0;
603 goto out;
604 }
605 NFSUNLOCKNODE(np);
606 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
607 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
608 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
609 if (!bp) {
610 error = newnfs_sigintr(nmp, td);
611 if (error == 0)
612 error = EINTR;
613 goto out;
614 }
615 if ((bp->b_flags & B_CACHE) == 0) {
616 bp->b_iocmd = BIO_READ;
617 vfs_busy_pages(bp, 0);
618 error = ncl_doio(vp, bp, cred, td, 0);
619 if (error) {
620 brelse(bp);
621 }
622 while (error == NFSERR_BAD_COOKIE) {
623 ncl_invaldir(vp);
624 error = ncl_vinvalbuf(vp, 0, td, 1);
625
626 /*
627 * Yuck! The directory has been modified on the
628 * server. The only way to get the block is by
629 * reading from the beginning to get all the
630 * offset cookies.
631 *
632 * Leave the last bp intact unless there is an error.
633 * Loop back up to the while if the error is another
634 * NFSERR_BAD_COOKIE (double yuch!).
635 */
636 for (i = 0; i <= lbn && !error; i++) {
637 NFSLOCKNODE(np);
638 if (np->n_direofoffset
639 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
640 NFSUNLOCKNODE(np);
641 error = 0;
642 goto out;
643 }
644 NFSUNLOCKNODE(np);
645 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
646 if (!bp) {
647 error = newnfs_sigintr(nmp, td);
648 if (error == 0)
649 error = EINTR;
650 goto out;
651 }
652 if ((bp->b_flags & B_CACHE) == 0) {
653 bp->b_iocmd = BIO_READ;
654 vfs_busy_pages(bp, 0);
655 error = ncl_doio(vp, bp, cred, td, 0);
656 /*
657 * no error + B_INVAL == directory EOF,
658 * use the block.
659 */
660 if (error == 0 && (bp->b_flags & B_INVAL))
661 break;
662 }
663 /*
664 * An error will throw away the block and the
665 * for loop will break out. If no error and this
666 * is not the block we want, we throw away the
667 * block and go for the next one via the for loop.
668 */
669 if (error || i < lbn)
670 brelse(bp);
671 }
672 }
673 /*
674 * The above while is repeated if we hit another cookie
675 * error. If we hit an error and it wasn't a cookie error,
676 * we give up.
677 */
678 if (error)
679 goto out;
680 }
681
682 /*
683 * If not eof and read aheads are enabled, start one.
684 * (You need the current block first, so that you have the
685 * directory offset cookie of the next block.)
686 */
687 NFSLOCKNODE(np);
688 if (nmp->nm_readahead > 0 && ncl_bioread_dora(vp) &&
689 (bp->b_flags & B_INVAL) == 0 &&
690 (np->n_direofoffset == 0 ||
691 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
692 incore(&vp->v_bufobj, lbn + 1) == NULL) {
693 NFSUNLOCKNODE(np);
694 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
695 if (rabp) {
696 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
697 rabp->b_flags |= B_ASYNC;
698 rabp->b_iocmd = BIO_READ;
699 vfs_busy_pages(rabp, 0);
700 if (ncl_asyncio(nmp, rabp, cred, td)) {
701 rabp->b_flags |= B_INVAL;
702 rabp->b_ioflags |= BIO_ERROR;
703 vfs_unbusy_pages(rabp);
704 brelse(rabp);
705 }
706 } else {
707 brelse(rabp);
708 }
709 }
710 NFSLOCKNODE(np);
711 }
712 /*
713 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
714 * chopped for the EOF condition, we cannot tell how large
715 * NFS directories are going to be until we hit EOF. So
716 * an NFS directory buffer is *not* chopped to its EOF. Now,
717 * it just so happens that b_resid will effectively chop it
718 * to EOF. *BUT* this information is lost if the buffer goes
719 * away and is reconstituted into a B_CACHE state ( due to
720 * being VMIO ) later. So we keep track of the directory eof
721 * in np->n_direofoffset and chop it off as an extra step
722 * right here.
723 */
724 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
725 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
726 n = np->n_direofoffset - uio->uio_offset;
727 NFSUNLOCKNODE(np);
728 break;
729 default:
730 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
731 bp = NULL;
732 break;
733 }
734
735 if (n > 0) {
736 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
737 }
738 if (vp->v_type == VLNK)
739 n = 0;
740 if (bp != NULL)
741 brelse(bp);
742 } while (error == 0 && uio->uio_resid > 0 && n > 0);
743 out:
744 curthread_pflags2_restore(save2);
745 if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
746 NFSLOCKNODE(np);
747 ncl_pager_setsize(vp, NULL);
748 }
749 return (error);
750 }
751
752 /*
753 * The NFS write path cannot handle iovecs with len > 1. So we need to
754 * break up iovecs accordingly (restricting them to wsize).
755 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
756 * For the ASYNC case, 2 copies are needed. The first a copy from the
757 * user buffer to a staging buffer and then a second copy from the staging
758 * buffer to mbufs. This can be optimized by copying from the user buffer
759 * directly into mbufs and passing the chain down, but that requires a
760 * fair amount of re-working of the relevant codepaths (and can be done
761 * later).
762 */
763 static int
nfs_directio_write(struct vnode * vp,struct uio * uiop,struct ucred * cred,int ioflag)764 nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred,
765 int ioflag)
766 {
767 struct uio uio;
768 struct iovec iov;
769 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
770 struct thread *td = uiop->uio_td;
771 int error, iomode, must_commit, size, wsize;
772
773 KASSERT((ioflag & IO_SYNC) != 0, ("nfs_directio_write: not sync"));
774 mtx_lock(&nmp->nm_mtx);
775 wsize = nmp->nm_wsize;
776 mtx_unlock(&nmp->nm_mtx);
777 while (uiop->uio_resid > 0) {
778 size = MIN(uiop->uio_resid, wsize);
779 size = MIN(uiop->uio_iov->iov_len, size);
780 iov.iov_base = uiop->uio_iov->iov_base;
781 iov.iov_len = size;
782 uio.uio_iov = &iov;
783 uio.uio_iovcnt = 1;
784 uio.uio_offset = uiop->uio_offset;
785 uio.uio_resid = size;
786 uio.uio_segflg = uiop->uio_segflg;
787 uio.uio_rw = UIO_WRITE;
788 uio.uio_td = td;
789 iomode = NFSWRITE_FILESYNC;
790 /*
791 * When doing direct I/O we do not care if the
792 * server's write verifier has changed, but we
793 * do not want to update the verifier if it has
794 * changed, since that hides the change from
795 * writes being done through the buffer cache.
796 * By passing must_commit in set to two, the code
797 * in nfsrpc_writerpc() will not update the
798 * verifier on the mount point.
799 */
800 must_commit = 2;
801 error = ncl_writerpc(vp, &uio, cred, &iomode,
802 &must_commit, 0, ioflag);
803 KASSERT(must_commit == 2,
804 ("ncl_directio_write: Updated write verifier"));
805 if (error != 0)
806 return (error);
807 if (iomode != NFSWRITE_FILESYNC)
808 printf("nfs_directio_write: Broken server "
809 "did not reply FILE_SYNC\n");
810 uiop->uio_offset += size;
811 uiop->uio_resid -= size;
812 if (uiop->uio_iov->iov_len <= size) {
813 uiop->uio_iovcnt--;
814 uiop->uio_iov++;
815 } else {
816 uiop->uio_iov->iov_base =
817 (char *)uiop->uio_iov->iov_base + size;
818 uiop->uio_iov->iov_len -= size;
819 }
820 }
821 return (0);
822 }
823
824 /*
825 * Vnode op for write using bio
826 */
827 int
ncl_write(struct vop_write_args * ap)828 ncl_write(struct vop_write_args *ap)
829 {
830 int biosize;
831 struct uio *uio = ap->a_uio;
832 struct thread *td = uio->uio_td;
833 struct vnode *vp = ap->a_vp;
834 struct nfsnode *np = VTONFS(vp);
835 struct ucred *cred = ap->a_cred;
836 int ioflag = ap->a_ioflag;
837 struct buf *bp;
838 struct vattr vattr;
839 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
840 daddr_t lbn;
841 int bcount, noncontig_write, obcount;
842 int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
843 size_t orig_resid, local_resid;
844 off_t orig_size, tmp_off;
845 struct timespec ts;
846
847 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
848 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
849 ("ncl_write proc"));
850 if (vp->v_type != VREG)
851 return (EIO);
852 NFSLOCKNODE(np);
853 if (np->n_flag & NWRITEERR) {
854 np->n_flag &= ~NWRITEERR;
855 NFSUNLOCKNODE(np);
856 return (np->n_error);
857 } else
858 NFSUNLOCKNODE(np);
859 mtx_lock(&nmp->nm_mtx);
860 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
861 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
862 mtx_unlock(&nmp->nm_mtx);
863 (void)ncl_fsinfo(nmp, vp, cred, td);
864 mtx_lock(&nmp->nm_mtx);
865 }
866 if (nmp->nm_wsize == 0)
867 (void) newnfs_iosize(nmp);
868 mtx_unlock(&nmp->nm_mtx);
869
870 /*
871 * Synchronously flush pending buffers if we are in synchronous
872 * mode or if we are appending.
873 */
874 if ((ioflag & IO_APPEND) || ((ioflag & IO_SYNC) && (np->n_flag &
875 NMODIFIED))) {
876 /*
877 * For the case where IO_APPEND is being done using a
878 * direct output (to the NFS server) RPC and
879 * newnfs_directio_enable is 0, all buffer cache buffers,
880 * including ones not modified, must be invalidated.
881 * This ensures that stale data is not read out of the
882 * buffer cache. The call also invalidates all mapped
883 * pages and, since the exclusive lock is held on the vnode,
884 * new pages cannot be faulted in.
885 *
886 * For the case where newnfs_directio_enable is set
887 * (which is not the default), it is not obvious that
888 * stale data should be left in the buffer cache, but
889 * the code has been this way for over a decade without
890 * complaints. Note that, unlike doing IO_APPEND via
891 * a direct write RPC when newnfs_directio_enable is not set,
892 * when newnfs_directio_enable is set, reading is done via
893 * direct to NFS server RPCs as well.
894 */
895 np->n_attrstamp = 0;
896 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
897 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
898 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
899 if (error != 0)
900 return (error);
901 }
902
903 orig_resid = uio->uio_resid;
904 NFSLOCKNODE(np);
905 orig_size = np->n_size;
906 NFSUNLOCKNODE(np);
907
908 /*
909 * If IO_APPEND then load uio_offset. We restart here if we cannot
910 * get the append lock.
911 */
912 if (ioflag & IO_APPEND) {
913 /*
914 * For NFSv4, the AppendWrite will Verify the size against
915 * the file's size on the server. If not the same, the
916 * write will then be retried, using the file size returned
917 * by the AppendWrite. However, for NFSv2 and NFSv3, the
918 * size must be acquired here via a Getattr RPC.
919 * The AppendWrite is not done for a pNFS mount.
920 */
921 if (!NFSHASNFSV4(nmp) || NFSHASPNFS(nmp)) {
922 np->n_attrstamp = 0;
923 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
924 error = VOP_GETATTR(vp, &vattr, cred);
925 if (error)
926 return (error);
927 }
928 NFSLOCKNODE(np);
929 uio->uio_offset = np->n_size;
930 NFSUNLOCKNODE(np);
931 }
932
933 if (uio->uio_offset < 0)
934 return (EINVAL);
935 tmp_off = uio->uio_offset + uio->uio_resid;
936 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
937 return (EFBIG);
938 if (uio->uio_resid == 0)
939 return (0);
940
941 /*
942 * Do IO_APPEND writing via a synchronous direct write.
943 * This can result in a significant performance improvement.
944 */
945 if ((newnfs_directio_enable && (ioflag & IO_DIRECT)) ||
946 (ioflag & IO_APPEND)) {
947 /*
948 * Direct writes to the server must be done NFSWRITE_FILESYNC,
949 * because the write data is not cached and, therefore, the
950 * write cannot be redone after a server reboot.
951 * Set IO_SYNC to make this happen.
952 */
953 ioflag |= IO_SYNC;
954 return (nfs_directio_write(vp, uio, cred, ioflag));
955 }
956
957 /*
958 * Maybe this should be above the vnode op call, but so long as
959 * file servers have no limits, i don't think it matters
960 */
961 error = vn_rlimit_fsize(vp, uio, td);
962 if (error != 0)
963 return (error);
964
965 save2 = curthread_pflags2_set(TDP2_SBPAGES);
966 biosize = vp->v_bufobj.bo_bsize;
967 /*
968 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
969 * would exceed the local maximum per-file write commit size when
970 * combined with those, we must decide whether to flush,
971 * go synchronous, or return error. We don't bother checking
972 * IO_UNIT -- we just make all writes atomic anyway, as there's
973 * no point optimizing for something that really won't ever happen.
974 */
975 wouldcommit = 0;
976 if (!(ioflag & IO_SYNC)) {
977 int nflag;
978
979 NFSLOCKNODE(np);
980 nflag = np->n_flag;
981 NFSUNLOCKNODE(np);
982 if (nflag & NMODIFIED) {
983 BO_LOCK(&vp->v_bufobj);
984 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
985 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
986 b_bobufs) {
987 if (bp->b_flags & B_NEEDCOMMIT)
988 wouldcommit += bp->b_bcount;
989 }
990 }
991 BO_UNLOCK(&vp->v_bufobj);
992 }
993 }
994
995 do {
996 if (!(ioflag & IO_SYNC)) {
997 wouldcommit += biosize;
998 if (wouldcommit > nmp->nm_wcommitsize) {
999 np->n_attrstamp = 0;
1000 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1001 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1002 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1003 if (error != 0)
1004 goto out;
1005 wouldcommit = biosize;
1006 }
1007 }
1008
1009 NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1010 lbn = uio->uio_offset / biosize;
1011 on = uio->uio_offset - (lbn * biosize);
1012 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1013 again:
1014 /*
1015 * Handle direct append and file extension cases, calculate
1016 * unaligned buffer size.
1017 */
1018 NFSLOCKNODE(np);
1019 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1020 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1021 noncontig_write = 1;
1022 else
1023 noncontig_write = 0;
1024 if ((uio->uio_offset == np->n_size ||
1025 (noncontig_write != 0 &&
1026 lbn == (np->n_size / biosize) &&
1027 uio->uio_offset + n > np->n_size)) && n) {
1028 NFSUNLOCKNODE(np);
1029 /*
1030 * Get the buffer (in its pre-append state to maintain
1031 * B_CACHE if it was previously set). Resize the
1032 * nfsnode after we have locked the buffer to prevent
1033 * readers from reading garbage.
1034 */
1035 obcount = np->n_size - (lbn * biosize);
1036 bp = nfs_getcacheblk(vp, lbn, obcount, td);
1037
1038 if (bp != NULL) {
1039 long save;
1040
1041 NFSLOCKNODE(np);
1042 np->n_size = uio->uio_offset + n;
1043 np->n_flag |= NMODIFIED;
1044 np->n_flag &= ~NVNSETSZSKIP;
1045 vnode_pager_setsize(vp, np->n_size);
1046 NFSUNLOCKNODE(np);
1047
1048 save = bp->b_flags & B_CACHE;
1049 bcount = on + n;
1050 allocbuf(bp, bcount);
1051 bp->b_flags |= save;
1052 if (noncontig_write != 0 && on > obcount)
1053 vfs_bio_bzero_buf(bp, obcount, on -
1054 obcount);
1055 }
1056 } else {
1057 /*
1058 * Obtain the locked cache block first, and then
1059 * adjust the file's size as appropriate.
1060 */
1061 bcount = on + n;
1062 if ((off_t)lbn * biosize + bcount < np->n_size) {
1063 if ((off_t)(lbn + 1) * biosize < np->n_size)
1064 bcount = biosize;
1065 else
1066 bcount = np->n_size - (off_t)lbn * biosize;
1067 }
1068 NFSUNLOCKNODE(np);
1069 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1070 NFSLOCKNODE(np);
1071 if (uio->uio_offset + n > np->n_size) {
1072 np->n_size = uio->uio_offset + n;
1073 np->n_flag |= NMODIFIED;
1074 np->n_flag &= ~NVNSETSZSKIP;
1075 vnode_pager_setsize(vp, np->n_size);
1076 }
1077 NFSUNLOCKNODE(np);
1078 }
1079
1080 if (!bp) {
1081 error = newnfs_sigintr(nmp, td);
1082 if (!error)
1083 error = EINTR;
1084 break;
1085 }
1086
1087 /*
1088 * Issue a READ if B_CACHE is not set. In special-append
1089 * mode, B_CACHE is based on the buffer prior to the write
1090 * op and is typically set, avoiding the read. If a read
1091 * is required in special append mode, the server will
1092 * probably send us a short-read since we extended the file
1093 * on our end, resulting in b_resid == 0 and, thusly,
1094 * B_CACHE getting set.
1095 *
1096 * We can also avoid issuing the read if the write covers
1097 * the entire buffer. We have to make sure the buffer state
1098 * is reasonable in this case since we will not be initiating
1099 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1100 * more information.
1101 *
1102 * B_CACHE may also be set due to the buffer being cached
1103 * normally.
1104 */
1105
1106 bp_cached = 1;
1107 if (on == 0 && n == bcount) {
1108 if ((bp->b_flags & B_CACHE) == 0)
1109 bp_cached = 0;
1110 bp->b_flags |= B_CACHE;
1111 bp->b_flags &= ~B_INVAL;
1112 bp->b_ioflags &= ~BIO_ERROR;
1113 }
1114
1115 if ((bp->b_flags & B_CACHE) == 0) {
1116 bp->b_iocmd = BIO_READ;
1117 vfs_busy_pages(bp, 0);
1118 error = ncl_doio(vp, bp, cred, td, 0);
1119 if (error) {
1120 brelse(bp);
1121 break;
1122 }
1123 }
1124 if (bp->b_wcred == NOCRED)
1125 bp->b_wcred = crhold(cred);
1126 NFSLOCKNODE(np);
1127 np->n_flag |= NMODIFIED;
1128 NFSUNLOCKNODE(np);
1129
1130 /*
1131 * If dirtyend exceeds file size, chop it down. This should
1132 * not normally occur but there is an append race where it
1133 * might occur XXX, so we log it.
1134 *
1135 * If the chopping creates a reverse-indexed or degenerate
1136 * situation with dirtyoff/end, we 0 both of them.
1137 */
1138
1139 if (bp->b_dirtyend > bcount) {
1140 printf("NFS append race @%lx:%d\n",
1141 (long)bp->b_blkno * DEV_BSIZE,
1142 bp->b_dirtyend - bcount);
1143 bp->b_dirtyend = bcount;
1144 }
1145
1146 if (bp->b_dirtyoff >= bp->b_dirtyend)
1147 bp->b_dirtyoff = bp->b_dirtyend = 0;
1148
1149 /*
1150 * If the new write will leave a contiguous dirty
1151 * area, just update the b_dirtyoff and b_dirtyend,
1152 * otherwise force a write rpc of the old dirty area.
1153 *
1154 * If there has been a file lock applied to this file
1155 * or vfs.nfs.old_noncontig_writing is set, do the following:
1156 * While it is possible to merge discontiguous writes due to
1157 * our having a B_CACHE buffer ( and thus valid read data
1158 * for the hole), we don't because it could lead to
1159 * significant cache coherency problems with multiple clients,
1160 * especially if locking is implemented later on.
1161 *
1162 * If vfs.nfs.old_noncontig_writing is not set and there has
1163 * not been file locking done on this file:
1164 * Relax coherency a bit for the sake of performance and
1165 * expand the current dirty region to contain the new
1166 * write even if it means we mark some non-dirty data as
1167 * dirty.
1168 */
1169
1170 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1171 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1172 if (bwrite(bp) == EINTR) {
1173 error = EINTR;
1174 break;
1175 }
1176 goto again;
1177 }
1178
1179 local_resid = uio->uio_resid;
1180 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1181
1182 if (error != 0 && !bp_cached) {
1183 /*
1184 * This block has no other content then what
1185 * possibly was written by the faulty uiomove.
1186 * Release it, forgetting the data pages, to
1187 * prevent the leak of uninitialized data to
1188 * usermode.
1189 */
1190 bp->b_ioflags |= BIO_ERROR;
1191 brelse(bp);
1192 uio->uio_offset -= local_resid - uio->uio_resid;
1193 uio->uio_resid = local_resid;
1194 break;
1195 }
1196
1197 /*
1198 * Since this block is being modified, it must be written
1199 * again and not just committed. Since write clustering does
1200 * not work for the stage 1 data write, only the stage 2
1201 * commit rpc, we have to clear B_CLUSTEROK as well.
1202 */
1203 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1204
1205 /*
1206 * Get the partial update on the progress made from
1207 * uiomove, if an error occurred.
1208 */
1209 if (error != 0)
1210 n = local_resid - uio->uio_resid;
1211
1212 /*
1213 * Only update dirtyoff/dirtyend if not a degenerate
1214 * condition.
1215 */
1216 if (n > 0) {
1217 if (bp->b_dirtyend > 0) {
1218 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1219 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1220 } else {
1221 bp->b_dirtyoff = on;
1222 bp->b_dirtyend = on + n;
1223 }
1224 vfs_bio_set_valid(bp, on, n);
1225 }
1226
1227 /*
1228 * If IO_SYNC do bwrite().
1229 *
1230 * IO_INVAL appears to be unused. The idea appears to be
1231 * to turn off caching in this case. Very odd. XXX
1232 */
1233 if ((ioflag & IO_SYNC)) {
1234 if (ioflag & IO_INVAL)
1235 bp->b_flags |= B_NOCACHE;
1236 error1 = bwrite(bp);
1237 if (error1 != 0) {
1238 if (error == 0)
1239 error = error1;
1240 break;
1241 }
1242 } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1243 bp->b_flags |= B_ASYNC;
1244 (void) bwrite(bp);
1245 } else {
1246 bdwrite(bp);
1247 }
1248
1249 if (error != 0)
1250 break;
1251 } while (uio->uio_resid > 0 && n > 0);
1252
1253 if (error == 0) {
1254 nanouptime(&ts);
1255 NFSLOCKNODE(np);
1256 np->n_localmodtime = ts;
1257 NFSUNLOCKNODE(np);
1258 } else {
1259 if (ioflag & IO_UNIT) {
1260 VATTR_NULL(&vattr);
1261 vattr.va_size = orig_size;
1262 /* IO_SYNC is handled implicitely */
1263 (void)VOP_SETATTR(vp, &vattr, cred);
1264 uio->uio_offset -= orig_resid - uio->uio_resid;
1265 uio->uio_resid = orig_resid;
1266 }
1267 }
1268
1269 out:
1270 curthread_pflags2_restore(save2);
1271 return (error);
1272 }
1273
1274 /*
1275 * Get an nfs cache block.
1276 *
1277 * Allocate a new one if the block isn't currently in the cache
1278 * and return the block marked busy. If the calling process is
1279 * interrupted by a signal for an interruptible mount point, return
1280 * NULL.
1281 *
1282 * The caller must carefully deal with the possible B_INVAL state of
1283 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1284 * indirectly), so synchronous reads can be issued without worrying about
1285 * the B_INVAL state. We have to be a little more careful when dealing
1286 * with writes (see comments in nfs_write()) when extending a file past
1287 * its EOF.
1288 */
1289 static struct buf *
nfs_getcacheblk(struct vnode * vp,daddr_t bn,int size,struct thread * td)1290 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1291 {
1292 struct buf *bp;
1293 struct mount *mp;
1294 struct nfsmount *nmp;
1295
1296 mp = vp->v_mount;
1297 nmp = VFSTONFS(mp);
1298
1299 if (nmp->nm_flag & NFSMNT_INT) {
1300 sigset_t oldset;
1301
1302 newnfs_set_sigmask(td, &oldset);
1303 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1304 newnfs_restore_sigmask(td, &oldset);
1305 while (bp == NULL) {
1306 if (newnfs_sigintr(nmp, td))
1307 return (NULL);
1308 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1309 }
1310 } else {
1311 bp = getblk(vp, bn, size, 0, 0, 0);
1312 }
1313
1314 if (vp->v_type == VREG)
1315 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1316 return (bp);
1317 }
1318
1319 /*
1320 * Flush and invalidate all dirty buffers. If another process is already
1321 * doing the flush, just wait for completion.
1322 */
1323 int
ncl_vinvalbuf(struct vnode * vp,int flags,struct thread * td,int intrflg)1324 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1325 {
1326 struct nfsnode *np = VTONFS(vp);
1327 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1328 int error = 0, slpflag, slptimeo;
1329 bool old_lock;
1330 struct timespec ts;
1331
1332 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1333
1334 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1335 intrflg = 0;
1336 if (NFSCL_FORCEDISM(nmp->nm_mountp))
1337 intrflg = 1;
1338 if (intrflg) {
1339 slpflag = PCATCH;
1340 slptimeo = 2 * hz;
1341 } else {
1342 slpflag = 0;
1343 slptimeo = 0;
1344 }
1345
1346 old_lock = ncl_excl_start(vp);
1347 if (old_lock)
1348 flags |= V_ALLOWCLEAN;
1349
1350 /*
1351 * Now, flush as required.
1352 */
1353 if ((flags & (V_SAVE | V_VMIO)) == V_SAVE) {
1354 vnode_pager_clean_sync(vp);
1355
1356 /*
1357 * If the page clean was interrupted, fail the invalidation.
1358 * Not doing so, we run the risk of losing dirty pages in the
1359 * vinvalbuf() call below.
1360 */
1361 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1362 goto out;
1363 }
1364
1365 error = vinvalbuf(vp, flags, slpflag, 0);
1366 while (error) {
1367 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1368 goto out;
1369 error = vinvalbuf(vp, flags, 0, slptimeo);
1370 }
1371 if (NFSHASPNFS(nmp)) {
1372 nfscl_layoutcommit(vp, td);
1373 nanouptime(&ts);
1374 /*
1375 * Invalidate the attribute cache, since writes to a DS
1376 * won't update the size attribute.
1377 */
1378 NFSLOCKNODE(np);
1379 np->n_attrstamp = 0;
1380 } else {
1381 nanouptime(&ts);
1382 NFSLOCKNODE(np);
1383 }
1384 if ((np->n_flag & NMODIFIED) != 0) {
1385 np->n_localmodtime = ts;
1386 np->n_flag &= ~NMODIFIED;
1387 }
1388 NFSUNLOCKNODE(np);
1389 out:
1390 ncl_excl_finish(vp, old_lock);
1391 return error;
1392 }
1393
1394 /*
1395 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1396 * This is mainly to avoid queueing async I/O requests when the nfsiods
1397 * are all hung on a dead server.
1398 *
1399 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1400 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1401 */
1402 int
ncl_asyncio(struct nfsmount * nmp,struct buf * bp,struct ucred * cred,struct thread * td)1403 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1404 {
1405 int iod;
1406 int gotiod;
1407 int slpflag = 0;
1408 int slptimeo = 0;
1409 int error, error2;
1410
1411 /*
1412 * Commits are usually short and sweet so lets save some cpu and
1413 * leave the async daemons for more important rpc's (such as reads
1414 * and writes).
1415 *
1416 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1417 * in the directory in order to update attributes. This can deadlock
1418 * with another thread that is waiting for async I/O to be done by
1419 * an nfsiod thread while holding a lock on one of these vnodes.
1420 * To avoid this deadlock, don't allow the async nfsiod threads to
1421 * perform Readdirplus RPCs.
1422 */
1423 NFSLOCKIOD();
1424 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1425 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1426 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1427 NFSUNLOCKIOD();
1428 return(EIO);
1429 }
1430 again:
1431 if (nmp->nm_flag & NFSMNT_INT)
1432 slpflag = PCATCH;
1433 gotiod = FALSE;
1434
1435 /*
1436 * Find a free iod to process this request.
1437 */
1438 for (iod = 0; iod < ncl_numasync; iod++)
1439 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1440 gotiod = TRUE;
1441 break;
1442 }
1443
1444 /*
1445 * Try to create one if none are free.
1446 */
1447 if (!gotiod)
1448 ncl_nfsiodnew();
1449 else {
1450 /*
1451 * Found one, so wake it up and tell it which
1452 * mount to process.
1453 */
1454 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1455 iod, nmp));
1456 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1457 ncl_iodmount[iod] = nmp;
1458 nmp->nm_bufqiods++;
1459 wakeup(&ncl_iodwant[iod]);
1460 }
1461
1462 /*
1463 * If none are free, we may already have an iod working on this mount
1464 * point. If so, it will process our request.
1465 */
1466 if (!gotiod) {
1467 if (nmp->nm_bufqiods > 0) {
1468 NFS_DPF(ASYNCIO,
1469 ("ncl_asyncio: %d iods are already processing mount %p\n",
1470 nmp->nm_bufqiods, nmp));
1471 gotiod = TRUE;
1472 }
1473 }
1474
1475 /*
1476 * If we have an iod which can process the request, then queue
1477 * the buffer.
1478 */
1479 if (gotiod) {
1480 /*
1481 * Ensure that the queue never grows too large. We still want
1482 * to asynchronize so we block rather then return EIO.
1483 */
1484 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1485 NFS_DPF(ASYNCIO,
1486 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1487 nmp->nm_bufqwant = TRUE;
1488 error = newnfs_msleep(td, &nmp->nm_bufq,
1489 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1490 slptimeo);
1491 if (error) {
1492 error2 = newnfs_sigintr(nmp, td);
1493 if (error2) {
1494 NFSUNLOCKIOD();
1495 return (error2);
1496 }
1497 if (slpflag == PCATCH) {
1498 slpflag = 0;
1499 slptimeo = 2 * hz;
1500 }
1501 }
1502 /*
1503 * We might have lost our iod while sleeping,
1504 * so check and loop if necessary.
1505 */
1506 goto again;
1507 }
1508
1509 /* We might have lost our nfsiod */
1510 if (nmp->nm_bufqiods == 0) {
1511 NFS_DPF(ASYNCIO,
1512 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1513 goto again;
1514 }
1515
1516 if (bp->b_iocmd == BIO_READ) {
1517 if (bp->b_rcred == NOCRED && cred != NOCRED)
1518 bp->b_rcred = crhold(cred);
1519 } else {
1520 if (bp->b_wcred == NOCRED && cred != NOCRED)
1521 bp->b_wcred = crhold(cred);
1522 }
1523
1524 if (bp->b_flags & B_REMFREE)
1525 bremfreef(bp);
1526 BUF_KERNPROC(bp);
1527 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1528 nmp->nm_bufqlen++;
1529 KASSERT((bp->b_flags & B_DIRECT) == 0,
1530 ("ncl_asyncio: B_DIRECT set"));
1531 NFSUNLOCKIOD();
1532 return (0);
1533 }
1534
1535 NFSUNLOCKIOD();
1536
1537 /*
1538 * All the iods are busy on other mounts, so return EIO to
1539 * force the caller to process the i/o synchronously.
1540 */
1541 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1542 return (EIO);
1543 }
1544
1545 /*
1546 * Do an I/O operation to/from a cache block. This may be called
1547 * synchronously or from an nfsiod.
1548 */
1549 int
ncl_doio(struct vnode * vp,struct buf * bp,struct ucred * cr,struct thread * td,int called_from_strategy)1550 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1551 int called_from_strategy)
1552 {
1553 struct uio *uiop;
1554 struct nfsnode *np;
1555 struct nfsmount *nmp;
1556 int error = 0, iomode, must_commit = 0;
1557 struct uio uio;
1558 struct iovec io;
1559 struct proc *p = td ? td->td_proc : NULL;
1560 uint8_t iocmd;
1561
1562 np = VTONFS(vp);
1563 nmp = VFSTONFS(vp->v_mount);
1564 uiop = &uio;
1565 uiop->uio_iov = &io;
1566 uiop->uio_iovcnt = 1;
1567 uiop->uio_segflg = UIO_SYSSPACE;
1568 uiop->uio_td = td;
1569
1570 /*
1571 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1572 * do this here so we do not have to do it in all the code that
1573 * calls us.
1574 */
1575 bp->b_flags &= ~B_INVAL;
1576 bp->b_ioflags &= ~BIO_ERROR;
1577
1578 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1579 iocmd = bp->b_iocmd;
1580 if (iocmd == BIO_READ) {
1581 io.iov_len = uiop->uio_resid = bp->b_bcount;
1582 io.iov_base = bp->b_data;
1583 uiop->uio_rw = UIO_READ;
1584
1585 switch (vp->v_type) {
1586 case VREG:
1587 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1588 NFSINCRGLOBAL(nfsstatsv1.read_bios);
1589 error = ncl_readrpc(vp, uiop, cr);
1590
1591 if (!error) {
1592 if (uiop->uio_resid) {
1593 /*
1594 * If we had a short read with no error, we must have
1595 * hit a file hole. We should zero-fill the remainder.
1596 * This can also occur if the server hits the file EOF.
1597 *
1598 * Holes used to be able to occur due to pending
1599 * writes, but that is not possible any longer.
1600 */
1601 int nread = bp->b_bcount - uiop->uio_resid;
1602 ssize_t left = uiop->uio_resid;
1603
1604 if (left > 0)
1605 bzero((char *)bp->b_data + nread, left);
1606 uiop->uio_resid = 0;
1607 }
1608 }
1609 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1610 if (p && vp->v_writecount <= -1) {
1611 NFSLOCKNODE(np);
1612 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1613 NFSUNLOCKNODE(np);
1614 PROC_LOCK(p);
1615 killproc(p, "text file modification");
1616 PROC_UNLOCK(p);
1617 } else
1618 NFSUNLOCKNODE(np);
1619 }
1620 break;
1621 case VLNK:
1622 uiop->uio_offset = (off_t)0;
1623 NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1624 error = ncl_readlinkrpc(vp, uiop, cr);
1625 break;
1626 case VDIR:
1627 NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1628 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1629 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1630 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1631 if (error == NFSERR_NOTSUPP)
1632 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1633 }
1634 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1635 error = ncl_readdirrpc(vp, uiop, cr, td);
1636 /*
1637 * end-of-directory sets B_INVAL but does not generate an
1638 * error.
1639 */
1640 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1641 bp->b_flags |= B_INVAL;
1642 break;
1643 default:
1644 printf("ncl_doio: type %x unexpected\n", vp->v_type);
1645 break;
1646 }
1647 if (error) {
1648 bp->b_ioflags |= BIO_ERROR;
1649 bp->b_error = error;
1650 }
1651 } else {
1652 /*
1653 * If we only need to commit, try to commit
1654 */
1655 if (bp->b_flags & B_NEEDCOMMIT) {
1656 int retv;
1657 off_t off;
1658
1659 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1660 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1661 bp->b_wcred, td);
1662 if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
1663 bp->b_dirtyoff = bp->b_dirtyend = 0;
1664 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1665 bp->b_resid = 0;
1666 bufdone(bp);
1667 return (0);
1668 }
1669 if (retv == NFSERR_STALEWRITEVERF) {
1670 ncl_clearcommit(vp->v_mount);
1671 }
1672 }
1673
1674 /*
1675 * Setup for actual write
1676 */
1677 NFSLOCKNODE(np);
1678 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1679 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1680 NFSUNLOCKNODE(np);
1681
1682 if (bp->b_dirtyend > bp->b_dirtyoff) {
1683 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1684 - bp->b_dirtyoff;
1685 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1686 + bp->b_dirtyoff;
1687 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1688 uiop->uio_rw = UIO_WRITE;
1689 NFSINCRGLOBAL(nfsstatsv1.write_bios);
1690
1691 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1692 iomode = NFSWRITE_UNSTABLE;
1693 else
1694 iomode = NFSWRITE_FILESYNC;
1695
1696 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1697 called_from_strategy, 0);
1698
1699 /*
1700 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1701 * to cluster the buffers needing commit. This will allow
1702 * the system to submit a single commit rpc for the whole
1703 * cluster. We can do this even if the buffer is not 100%
1704 * dirty (relative to the NFS blocksize), so we optimize the
1705 * append-to-file-case.
1706 *
1707 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1708 * cleared because write clustering only works for commit
1709 * rpc's, not for the data portion of the write).
1710 */
1711
1712 if (!error && iomode == NFSWRITE_UNSTABLE) {
1713 bp->b_flags |= B_NEEDCOMMIT;
1714 if (bp->b_dirtyoff == 0
1715 && bp->b_dirtyend == bp->b_bcount)
1716 bp->b_flags |= B_CLUSTEROK;
1717 } else {
1718 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1719 }
1720
1721 /*
1722 * For an interrupted write, the buffer is still valid
1723 * and the write hasn't been pushed to the server yet,
1724 * so we can't set BIO_ERROR and report the interruption
1725 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1726 * is not relevant, so the rpc attempt is essentially
1727 * a noop. For the case of a V3 write rpc not being
1728 * committed to stable storage, the block is still
1729 * dirty and requires either a commit rpc or another
1730 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1731 * the block is reused. This is indicated by setting
1732 * the B_DELWRI and B_NEEDCOMMIT flags.
1733 *
1734 * EIO is returned by ncl_writerpc() to indicate a recoverable
1735 * write error and is handled as above, except that
1736 * B_EINTR isn't set. One cause of this is a stale stateid
1737 * error for the RPC that indicates recovery is required,
1738 * when called with called_from_strategy != 0.
1739 *
1740 * If the buffer is marked B_PAGING, it does not reside on
1741 * the vp's paging queues so we cannot call bdirty(). The
1742 * bp in this case is not an NFS cache block so we should
1743 * be safe. XXX
1744 *
1745 * The logic below breaks up errors into recoverable and
1746 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1747 * and keep the buffer around for potential write retries.
1748 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1749 * and save the error in the nfsnode. This is less than ideal
1750 * but necessary. Keeping such buffers around could potentially
1751 * cause buffer exhaustion eventually (they can never be written
1752 * out, so will get constantly be re-dirtied). It also causes
1753 * all sorts of vfs panics. For non-recoverable write errors,
1754 * also invalidate the attrcache, so we'll be forced to go over
1755 * the wire for this object, returning an error to user on next
1756 * call (most of the time).
1757 */
1758 if (error == EINTR || error == EIO || error == ETIMEDOUT
1759 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1760 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1761 if ((bp->b_flags & B_PAGING) == 0) {
1762 bdirty(bp);
1763 bp->b_flags &= ~B_DONE;
1764 }
1765 if ((error == EINTR || error == ETIMEDOUT) &&
1766 (bp->b_flags & B_ASYNC) == 0)
1767 bp->b_flags |= B_EINTR;
1768 } else {
1769 if (error) {
1770 bp->b_ioflags |= BIO_ERROR;
1771 bp->b_flags |= B_INVAL;
1772 bp->b_error = np->n_error = error;
1773 NFSLOCKNODE(np);
1774 np->n_flag |= NWRITEERR;
1775 np->n_attrstamp = 0;
1776 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1777 NFSUNLOCKNODE(np);
1778 }
1779 bp->b_dirtyoff = bp->b_dirtyend = 0;
1780 }
1781 } else {
1782 bp->b_resid = 0;
1783 bufdone(bp);
1784 return (0);
1785 }
1786 }
1787 bp->b_resid = uiop->uio_resid;
1788 if (must_commit == 1)
1789 ncl_clearcommit(vp->v_mount);
1790 bufdone(bp);
1791 return (error);
1792 }
1793
1794 /*
1795 * Used to aid in handling ftruncate() operations on the NFS client side.
1796 * Truncation creates a number of special problems for NFS. We have to
1797 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1798 * we have to properly handle VM pages or (potentially dirty) buffers
1799 * that straddle the truncation point.
1800 */
1801
1802 int
ncl_meta_setsize(struct vnode * vp,struct thread * td,u_quad_t nsize)1803 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1804 {
1805 struct nfsnode *np = VTONFS(vp);
1806 u_quad_t tsize;
1807 int biosize = vp->v_bufobj.bo_bsize;
1808 int error = 0;
1809
1810 NFSLOCKNODE(np);
1811 tsize = np->n_size;
1812 np->n_size = nsize;
1813 NFSUNLOCKNODE(np);
1814
1815 if (nsize < tsize) {
1816 struct buf *bp;
1817 daddr_t lbn;
1818 int bufsize;
1819
1820 /*
1821 * vtruncbuf() doesn't get the buffer overlapping the
1822 * truncation point. We may have a B_DELWRI and/or B_CACHE
1823 * buffer that now needs to be truncated.
1824 */
1825 error = vtruncbuf(vp, nsize, biosize);
1826 lbn = nsize / biosize;
1827 bufsize = nsize - (lbn * biosize);
1828 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1829 if (!bp)
1830 return EINTR;
1831 if (bp->b_dirtyoff > bp->b_bcount)
1832 bp->b_dirtyoff = bp->b_bcount;
1833 if (bp->b_dirtyend > bp->b_bcount)
1834 bp->b_dirtyend = bp->b_bcount;
1835 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1836 brelse(bp);
1837 } else {
1838 vnode_pager_setsize(vp, nsize);
1839 }
1840 return(error);
1841 }
1842