VFS messaging/interfacing work stage 5/99. Start work on the new
[dragonfly.git] / sys / vfs / nfs / nfs_bio.c
CommitLineData
984263bc
MD
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 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
79e5012e 37 * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $
ce71fd38 38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.17 2004/06/08 02:58:52 hmp Exp $
984263bc
MD
39 */
40
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/resourcevar.h>
45#include <sys/signalvar.h>
46#include <sys/proc.h>
47#include <sys/buf.h>
48#include <sys/vnode.h>
49#include <sys/mount.h>
50#include <sys/kernel.h>
6e1e09e4 51#include <sys/buf2.h>
ce71fd38 52#include <sys/msfbuf.h>
984263bc
MD
53
54#include <vm/vm.h>
55#include <vm/vm_extern.h>
56#include <vm/vm_page.h>
57#include <vm/vm_object.h>
58#include <vm/vm_pager.h>
59#include <vm/vnode_pager.h>
60
1f2de5d4
MD
61#include "rpcv2.h"
62#include "nfsproto.h"
63#include "nfs.h"
64#include "nfsmount.h"
65#include "nqnfs.h"
66#include "nfsnode.h"
984263bc 67
a6ee311a
RG
68static struct buf *nfs_getcacheblk (struct vnode *vp, daddr_t bn, int size,
69 struct thread *td);
984263bc
MD
70
71extern int nfs_numasync;
72extern int nfs_pbuf_freecnt;
73extern struct nfsstats nfsstats;
74
75/*
76 * Vnode op for VM getpages.
e851b29e
CP
77 *
78 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
79 * int a_reqpage, vm_ooffset_t a_offset)
984263bc
MD
80 */
81int
e851b29e 82nfs_getpages(struct vop_getpages_args *ap)
984263bc 83{
dadab5e9 84 struct thread *td = curthread; /* XXX */
984263bc
MD
85 int i, error, nextoff, size, toff, count, npages;
86 struct uio uio;
87 struct iovec iov;
88 vm_offset_t kva;
984263bc 89 struct vnode *vp;
984263bc
MD
90 struct nfsmount *nmp;
91 vm_page_t *pages;
8d429613 92 vm_page_t m;
ce71fd38 93 struct msf_buf *msf;
984263bc
MD
94
95 vp = ap->a_vp;
984263bc
MD
96 nmp = VFSTONFS(vp->v_mount);
97 pages = ap->a_m;
98 count = ap->a_count;
99
100 if (vp->v_object == NULL) {
101 printf("nfs_getpages: called with non-merged cache vnode??\n");
102 return VM_PAGER_ERROR;
103 }
104
105 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
106 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
3b568787 107 (void)nfs_fsinfo(nmp, vp, td);
984263bc
MD
108
109 npages = btoc(count);
110
111 /*
8d429613
MD
112 * NOTE that partially valid pages may occur in cases other
113 * then file EOF, such as when a file is partially written and
114 * ftruncate()-extended to a larger size. It is also possible
115 * for the valid bits to be set on garbage beyond the file EOF and
116 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
117 * occur due to vtruncbuf() and the buffer cache's handling of
118 * pages which 'straddle' buffers or when b_bufsize is not a
119 * multiple of PAGE_SIZE.... the buffer cache cannot normally
120 * clear the extra bits. This kind of situation occurs when you
121 * make a small write() (m->valid == 0x03) and then mmap() and
122 * fault in the buffer(m->valid = 0xFF). When NFS flushes the
123 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
124 *
125 * This is combined with the possibility that the pages are partially
126 * dirty or that there is a buffer backing the pages that is dirty
127 * (even if m->dirty is 0).
128 *
129 * To solve this problem several hacks have been made: (1) NFS
130 * guarentees that the IO block size is a multiple of PAGE_SIZE and
131 * (2) The buffer cache, when invalidating an NFS buffer, will
132 * disregard the buffer's fragmentory b_bufsize and invalidate
133 * the whole page rather then just the piece the buffer owns.
134 *
135 * This allows us to assume that a partially valid page found here
136 * is fully valid (vm_fault will zero'd out areas of the page not
137 * marked as valid).
984263bc 138 */
8d429613
MD
139 m = pages[ap->a_reqpage];
140 if (m->valid != 0) {
141 for (i = 0; i < npages; ++i) {
142 if (i != ap->a_reqpage)
143 vnode_pager_freepage(pages[i]);
984263bc 144 }
8d429613 145 return(0);
984263bc
MD
146 }
147
148 /*
ce71fd38 149 * Use an MSF_BUF as a medium to retrieve data from the pages.
984263bc 150 */
ce71fd38
HP
151 msf = msf_buf_alloc(pages, npages, 0);
152 kva = msf_buf_kva(msf);
984263bc
MD
153
154 iov.iov_base = (caddr_t) kva;
155 iov.iov_len = count;
156 uio.uio_iov = &iov;
157 uio.uio_iovcnt = 1;
158 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
159 uio.uio_resid = count;
160 uio.uio_segflg = UIO_SYSSPACE;
161 uio.uio_rw = UIO_READ;
dadab5e9 162 uio.uio_td = td;
984263bc 163
3b568787 164 error = nfs_readrpc(vp, &uio);
ce71fd38 165 msf_buf_free(msf);
984263bc
MD
166
167 if (error && (uio.uio_resid == count)) {
168 printf("nfs_getpages: error %d\n", error);
169 for (i = 0; i < npages; ++i) {
170 if (i != ap->a_reqpage)
171 vnode_pager_freepage(pages[i]);
172 }
173 return VM_PAGER_ERROR;
174 }
175
176 /*
177 * Calculate the number of bytes read and validate only that number
178 * of bytes. Note that due to pending writes, size may be 0. This
179 * does not mean that the remaining data is invalid!
180 */
181
182 size = count - uio.uio_resid;
183
184 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
984263bc
MD
185 nextoff = toff + PAGE_SIZE;
186 m = pages[i];
187
188 m->flags &= ~PG_ZERO;
189
190 if (nextoff <= size) {
191 /*
192 * Read operation filled an entire page
193 */
194 m->valid = VM_PAGE_BITS_ALL;
195 vm_page_undirty(m);
196 } else if (size > toff) {
197 /*
198 * Read operation filled a partial page.
199 */
200 m->valid = 0;
201 vm_page_set_validclean(m, 0, size - toff);
202 /* handled by vm_fault now */
203 /* vm_page_zero_invalid(m, TRUE); */
204 } else {
205 /*
206 * Read operation was short. If no error occured
207 * we may have hit a zero-fill section. We simply
208 * leave valid set to 0.
209 */
210 ;
211 }
212 if (i != ap->a_reqpage) {
213 /*
214 * Whether or not to leave the page activated is up in
215 * the air, but we should put the page on a page queue
216 * somewhere (it already is in the object). Result:
217 * It appears that emperical results show that
218 * deactivating pages is best.
219 */
220
221 /*
222 * Just in case someone was asking for this page we
223 * now tell them that it is ok to use.
224 */
225 if (!error) {
226 if (m->flags & PG_WANTED)
227 vm_page_activate(m);
228 else
229 vm_page_deactivate(m);
230 vm_page_wakeup(m);
231 } else {
232 vnode_pager_freepage(m);
233 }
234 }
235 }
236 return 0;
237}
238
239/*
240 * Vnode op for VM putpages.
e851b29e
CP
241 *
242 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
243 * int *a_rtvals, vm_ooffset_t a_offset)
984263bc
MD
244 */
245int
e851b29e 246nfs_putpages(struct vop_putpages_args *ap)
984263bc 247{
dadab5e9 248 struct thread *td = curthread;
984263bc
MD
249 struct uio uio;
250 struct iovec iov;
251 vm_offset_t kva;
984263bc
MD
252 int iomode, must_commit, i, error, npages, count;
253 off_t offset;
254 int *rtvals;
255 struct vnode *vp;
984263bc
MD
256 struct nfsmount *nmp;
257 struct nfsnode *np;
258 vm_page_t *pages;
ce71fd38 259 struct msf_buf *msf;
984263bc
MD
260
261 vp = ap->a_vp;
262 np = VTONFS(vp);
984263bc
MD
263 nmp = VFSTONFS(vp->v_mount);
264 pages = ap->a_m;
265 count = ap->a_count;
266 rtvals = ap->a_rtvals;
267 npages = btoc(count);
268 offset = IDX_TO_OFF(pages[0]->pindex);
269
270 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
271 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
3b568787 272 (void)nfs_fsinfo(nmp, vp, td);
984263bc
MD
273
274 for (i = 0; i < npages; i++) {
275 rtvals[i] = VM_PAGER_AGAIN;
276 }
277
278 /*
279 * When putting pages, do not extend file past EOF.
280 */
281
282 if (offset + count > np->n_size) {
283 count = np->n_size - offset;
284 if (count < 0)
285 count = 0;
286 }
287
288 /*
ce71fd38 289 * Use an MSF_BUF as a medium to retrieve data from the pages.
984263bc 290 */
ce71fd38
HP
291 msf = msf_buf_alloc(pages, npages, 0);
292 kva = msf_buf_kva(msf);
984263bc
MD
293
294 iov.iov_base = (caddr_t) kva;
295 iov.iov_len = count;
296 uio.uio_iov = &iov;
297 uio.uio_iovcnt = 1;
298 uio.uio_offset = offset;
299 uio.uio_resid = count;
300 uio.uio_segflg = UIO_SYSSPACE;
301 uio.uio_rw = UIO_WRITE;
dadab5e9 302 uio.uio_td = td;
984263bc
MD
303
304 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
305 iomode = NFSV3WRITE_UNSTABLE;
306 else
307 iomode = NFSV3WRITE_FILESYNC;
308
3b568787 309 error = nfs_writerpc(vp, &uio, &iomode, &must_commit);
984263bc 310
ce71fd38 311 msf_buf_free(msf);
984263bc
MD
312
313 if (!error) {
314 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
315 for (i = 0; i < nwritten; i++) {
316 rtvals[i] = VM_PAGER_OK;
317 vm_page_undirty(pages[i]);
318 }
319 if (must_commit)
320 nfs_clearcommit(vp->v_mount);
321 }
322 return rtvals[0];
323}
324
325/*
326 * Vnode op for read using bio
327 */
328int
3b568787 329nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
984263bc 330{
dadab5e9
MD
331 struct nfsnode *np = VTONFS(vp);
332 int biosize, i;
984263bc
MD
333 struct buf *bp = 0, *rabp;
334 struct vattr vattr;
dadab5e9 335 struct thread *td;
984263bc
MD
336 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
337 daddr_t lbn, rabn;
338 int bcount;
339 int seqcount;
340 int nra, error = 0, n = 0, on = 0;
341
342#ifdef DIAGNOSTIC
343 if (uio->uio_rw != UIO_READ)
344 panic("nfs_read mode");
345#endif
346 if (uio->uio_resid == 0)
347 return (0);
348 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
349 return (EINVAL);
dadab5e9 350 td = uio->uio_td;
984263bc
MD
351
352 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
353 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
3b568787 354 (void)nfs_fsinfo(nmp, vp, td);
984263bc
MD
355 if (vp->v_type != VDIR &&
356 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
357 return (EFBIG);
358 biosize = vp->v_mount->mnt_stat.f_iosize;
359 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
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 nqnfs, full cache consistency is maintained within the loop.
365 * For nfs:
366 * If the file's modify time on the server has changed since the
367 * last read rpc or you have written to the file,
368 * you may have lost data cache consistency with the
369 * server, so flush all of the file's data out of the cache.
370 * Then force a getattr rpc to ensure that you have up to date
371 * attributes.
372 * NB: This implies that cache data can be read when up to
373 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
374 * attributes this could be forced by setting n_attrstamp to 0 before
375 * the VOP_GETATTR() call.
376 */
377 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
378 if (np->n_flag & NMODIFIED) {
379 if (vp->v_type != VREG) {
380 if (vp->v_type != VDIR)
381 panic("nfs: bioread, not dir");
382 nfs_invaldir(vp);
3b568787 383 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
384 if (error)
385 return (error);
386 }
387 np->n_attrstamp = 0;
3b568787 388 error = VOP_GETATTR(vp, &vattr, td);
984263bc
MD
389 if (error)
390 return (error);
391 np->n_mtime = vattr.va_mtime.tv_sec;
392 } else {
3b568787 393 error = VOP_GETATTR(vp, &vattr, td);
984263bc
MD
394 if (error)
395 return (error);
79e5012e
MD
396 if ((np->n_flag & NSIZECHANGED)
397 || np->n_mtime != vattr.va_mtime.tv_sec) {
984263bc
MD
398 if (vp->v_type == VDIR)
399 nfs_invaldir(vp);
3b568787 400 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
401 if (error)
402 return (error);
403 np->n_mtime = vattr.va_mtime.tv_sec;
79e5012e 404 np->n_flag &= ~NSIZECHANGED;
984263bc
MD
405 }
406 }
407 }
408 do {
409
410 /*
411 * Get a valid lease. If cached data is stale, flush it.
412 */
413 if (nmp->nm_flag & NFSMNT_NQNFS) {
414 if (NQNFS_CKINVALID(vp, np, ND_READ)) {
415 do {
3b568787 416 error = nqnfs_getlease(vp, ND_READ, td);
984263bc
MD
417 } while (error == NQNFS_EXPIRED);
418 if (error)
419 return (error);
420 if (np->n_lrev != np->n_brev ||
421 (np->n_flag & NQNFSNONCACHE) ||
422 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
423 if (vp->v_type == VDIR)
424 nfs_invaldir(vp);
3b568787 425 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
426 if (error)
427 return (error);
428 np->n_brev = np->n_lrev;
429 }
430 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
431 nfs_invaldir(vp);
3b568787 432 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
433 if (error)
434 return (error);
435 }
436 }
437 if (np->n_flag & NQNFSNONCACHE) {
438 switch (vp->v_type) {
439 case VREG:
3b568787 440 return (nfs_readrpc(vp, uio));
984263bc 441 case VLNK:
3b568787 442 return (nfs_readlinkrpc(vp, uio));
984263bc
MD
443 case VDIR:
444 break;
445 default:
446 printf(" NQNFSNONCACHE: type %x unexpected\n",
447 vp->v_type);
448 };
449 }
450 switch (vp->v_type) {
451 case VREG:
452 nfsstats.biocache_reads++;
453 lbn = uio->uio_offset / biosize;
454 on = uio->uio_offset & (biosize - 1);
455
456 /*
457 * Start the read ahead(s), as required.
458 */
459 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
460 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
461 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
462 rabn = lbn + 1 + nra;
463 if (!incore(vp, rabn)) {
dadab5e9 464 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
984263bc
MD
465 if (!rabp)
466 return (EINTR);
467 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
468 rabp->b_flags |= (B_READ | B_ASYNC);
469 vfs_busy_pages(rabp, 0);
3b568787 470 if (nfs_asyncio(rabp, td)) {
984263bc
MD
471 rabp->b_flags |= B_INVAL|B_ERROR;
472 vfs_unbusy_pages(rabp);
473 brelse(rabp);
474 break;
475 }
476 } else {
477 brelse(rabp);
478 }
479 }
480 }
481 }
482
483 /*
484 * Obtain the buffer cache block. Figure out the buffer size
485 * when we are at EOF. If we are modifying the size of the
486 * buffer based on an EOF condition we need to hold
487 * nfs_rslock() through obtaining the buffer to prevent
488 * a potential writer-appender from messing with n_size.
489 * Otherwise we may accidently truncate the buffer and
490 * lose dirty data.
491 *
492 * Note that bcount is *not* DEV_BSIZE aligned.
493 */
494
495again:
496 bcount = biosize;
497 if ((off_t)lbn * biosize >= np->n_size) {
498 bcount = 0;
499 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
500 bcount = np->n_size - (off_t)lbn * biosize;
501 }
502 if (bcount != biosize) {
dadab5e9 503 switch(nfs_rslock(np, td)) {
984263bc
MD
504 case ENOLCK:
505 goto again;
506 /* not reached */
507 case EINTR:
508 case ERESTART:
509 return(EINTR);
510 /* not reached */
511 default:
512 break;
513 }
514 }
515
dadab5e9 516 bp = nfs_getcacheblk(vp, lbn, bcount, td);
984263bc
MD
517
518 if (bcount != biosize)
dadab5e9 519 nfs_rsunlock(np, td);
984263bc
MD
520 if (!bp)
521 return (EINTR);
522
523 /*
524 * If B_CACHE is not set, we must issue the read. If this
525 * fails, we return an error.
526 */
527
528 if ((bp->b_flags & B_CACHE) == 0) {
529 bp->b_flags |= B_READ;
530 vfs_busy_pages(bp, 0);
3b568787 531 error = nfs_doio(bp, td);
984263bc
MD
532 if (error) {
533 brelse(bp);
534 return (error);
535 }
536 }
537
538 /*
539 * on is the offset into the current bp. Figure out how many
540 * bytes we can copy out of the bp. Note that bcount is
541 * NOT DEV_BSIZE aligned.
542 *
543 * Then figure out how many bytes we can copy into the uio.
544 */
545
546 n = 0;
547 if (on < bcount)
548 n = min((unsigned)(bcount - on), uio->uio_resid);
549 break;
550 case VLNK:
551 nfsstats.biocache_readlinks++;
dadab5e9 552 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
984263bc
MD
553 if (!bp)
554 return (EINTR);
555 if ((bp->b_flags & B_CACHE) == 0) {
556 bp->b_flags |= B_READ;
557 vfs_busy_pages(bp, 0);
3b568787 558 error = nfs_doio(bp, td);
984263bc
MD
559 if (error) {
560 bp->b_flags |= B_ERROR;
561 brelse(bp);
562 return (error);
563 }
564 }
565 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
566 on = 0;
567 break;
568 case VDIR:
569 nfsstats.biocache_readdirs++;
570 if (np->n_direofoffset
571 && uio->uio_offset >= np->n_direofoffset) {
572 return (0);
573 }
574 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
575 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
dadab5e9 576 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
984263bc
MD
577 if (!bp)
578 return (EINTR);
579 if ((bp->b_flags & B_CACHE) == 0) {
580 bp->b_flags |= B_READ;
581 vfs_busy_pages(bp, 0);
3b568787 582 error = nfs_doio(bp, td);
984263bc
MD
583 if (error) {
584 brelse(bp);
585 }
586 while (error == NFSERR_BAD_COOKIE) {
587 printf("got bad cookie vp %p bp %p\n", vp, bp);
588 nfs_invaldir(vp);
3b568787 589 error = nfs_vinvalbuf(vp, 0, td, 1);
984263bc
MD
590 /*
591 * Yuck! The directory has been modified on the
592 * server. The only way to get the block is by
593 * reading from the beginning to get all the
594 * offset cookies.
595 *
596 * Leave the last bp intact unless there is an error.
597 * Loop back up to the while if the error is another
598 * NFSERR_BAD_COOKIE (double yuch!).
599 */
600 for (i = 0; i <= lbn && !error; i++) {
601 if (np->n_direofoffset
602 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
603 return (0);
dadab5e9 604 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
984263bc
MD
605 if (!bp)
606 return (EINTR);
607 if ((bp->b_flags & B_CACHE) == 0) {
608 bp->b_flags |= B_READ;
609 vfs_busy_pages(bp, 0);
3b568787 610 error = nfs_doio(bp, td);
984263bc
MD
611 /*
612 * no error + B_INVAL == directory EOF,
613 * use the block.
614 */
615 if (error == 0 && (bp->b_flags & B_INVAL))
616 break;
617 }
618 /*
619 * An error will throw away the block and the
620 * for loop will break out. If no error and this
621 * is not the block we want, we throw away the
622 * block and go for the next one via the for loop.
623 */
624 if (error || i < lbn)
625 brelse(bp);
626 }
627 }
628 /*
629 * The above while is repeated if we hit another cookie
630 * error. If we hit an error and it wasn't a cookie error,
631 * we give up.
632 */
633 if (error)
634 return (error);
635 }
636
637 /*
638 * If not eof and read aheads are enabled, start one.
639 * (You need the current block first, so that you have the
640 * directory offset cookie of the next block.)
641 */
642 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
643 (bp->b_flags & B_INVAL) == 0 &&
644 (np->n_direofoffset == 0 ||
645 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
646 !(np->n_flag & NQNFSNONCACHE) &&
647 !incore(vp, lbn + 1)) {
dadab5e9 648 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
984263bc
MD
649 if (rabp) {
650 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
651 rabp->b_flags |= (B_READ | B_ASYNC);
652 vfs_busy_pages(rabp, 0);
3b568787 653 if (nfs_asyncio(rabp, td)) {
984263bc
MD
654 rabp->b_flags |= B_INVAL|B_ERROR;
655 vfs_unbusy_pages(rabp);
656 brelse(rabp);
657 }
658 } else {
659 brelse(rabp);
660 }
661 }
662 }
663 /*
664 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
665 * chopped for the EOF condition, we cannot tell how large
666 * NFS directories are going to be until we hit EOF. So
667 * an NFS directory buffer is *not* chopped to its EOF. Now,
668 * it just so happens that b_resid will effectively chop it
669 * to EOF. *BUT* this information is lost if the buffer goes
670 * away and is reconstituted into a B_CACHE state ( due to
671 * being VMIO ) later. So we keep track of the directory eof
672 * in np->n_direofoffset and chop it off as an extra step
673 * right here.
674 */
675 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
676 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
677 n = np->n_direofoffset - uio->uio_offset;
678 break;
679 default:
680 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
681 break;
682 };
683
684 if (n > 0) {
685 error = uiomove(bp->b_data + on, (int)n, uio);
686 }
687 switch (vp->v_type) {
688 case VREG:
689 break;
690 case VLNK:
691 n = 0;
692 break;
693 case VDIR:
694 /*
695 * Invalidate buffer if caching is disabled, forcing a
696 * re-read from the remote later.
697 */
698 if (np->n_flag & NQNFSNONCACHE)
699 bp->b_flags |= B_INVAL;
700 break;
701 default:
702 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
703 }
704 brelse(bp);
705 } while (error == 0 && uio->uio_resid > 0 && n > 0);
706 return (error);
707}
708
709/*
710 * Vnode op for write using bio
e851b29e
CP
711 *
712 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
713 * struct ucred *a_cred)
984263bc
MD
714 */
715int
e851b29e 716nfs_write(struct vop_write_args *ap)
984263bc
MD
717{
718 int biosize;
719 struct uio *uio = ap->a_uio;
dadab5e9 720 struct thread *td = uio->uio_td;
984263bc
MD
721 struct vnode *vp = ap->a_vp;
722 struct nfsnode *np = VTONFS(vp);
984263bc
MD
723 int ioflag = ap->a_ioflag;
724 struct buf *bp;
725 struct vattr vattr;
726 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
727 daddr_t lbn;
728 int bcount;
729 int n, on, error = 0, iomode, must_commit;
730 int haverslock = 0;
731
732#ifdef DIAGNOSTIC
733 if (uio->uio_rw != UIO_WRITE)
734 panic("nfs_write mode");
7b95be2a 735 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
984263bc
MD
736 panic("nfs_write proc");
737#endif
738 if (vp->v_type != VREG)
739 return (EIO);
740 if (np->n_flag & NWRITEERR) {
741 np->n_flag &= ~NWRITEERR;
742 return (np->n_error);
743 }
744 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
745 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
3b568787 746 (void)nfs_fsinfo(nmp, vp, td);
984263bc
MD
747
748 /*
749 * Synchronously flush pending buffers if we are in synchronous
750 * mode or if we are appending.
751 */
752 if (ioflag & (IO_APPEND | IO_SYNC)) {
753 if (np->n_flag & NMODIFIED) {
754 np->n_attrstamp = 0;
3b568787 755 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
756 if (error)
757 return (error);
758 }
759 }
760
761 /*
762 * If IO_APPEND then load uio_offset. We restart here if we cannot
763 * get the append lock.
764 */
765restart:
766 if (ioflag & IO_APPEND) {
767 np->n_attrstamp = 0;
3b568787 768 error = VOP_GETATTR(vp, &vattr, td);
984263bc
MD
769 if (error)
770 return (error);
771 uio->uio_offset = np->n_size;
772 }
773
774 if (uio->uio_offset < 0)
775 return (EINVAL);
776 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
777 return (EFBIG);
778 if (uio->uio_resid == 0)
779 return (0);
780
781 /*
782 * We need to obtain the rslock if we intend to modify np->n_size
783 * in order to guarentee the append point with multiple contending
784 * writers, to guarentee that no other appenders modify n_size
785 * while we are trying to obtain a truncated buffer (i.e. to avoid
786 * accidently truncating data written by another appender due to
787 * the race), and to ensure that the buffer is populated prior to
788 * our extending of the file. We hold rslock through the entire
789 * operation.
790 *
791 * Note that we do not synchronize the case where someone truncates
792 * the file while we are appending to it because attempting to lock
793 * this case may deadlock other parts of the system unexpectedly.
794 */
795 if ((ioflag & IO_APPEND) ||
796 uio->uio_offset + uio->uio_resid > np->n_size) {
dadab5e9 797 switch(nfs_rslock(np, td)) {
984263bc
MD
798 case ENOLCK:
799 goto restart;
800 /* not reached */
801 case EINTR:
802 case ERESTART:
803 return(EINTR);
804 /* not reached */
805 default:
806 break;
807 }
808 haverslock = 1;
809 }
810
811 /*
812 * Maybe this should be above the vnode op call, but so long as
813 * file servers have no limits, i don't think it matters
814 */
dadab5e9
MD
815 if (td->td_proc && uio->uio_offset + uio->uio_resid >
816 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
817 psignal(td->td_proc, SIGXFSZ);
984263bc 818 if (haverslock)
dadab5e9 819 nfs_rsunlock(np, td);
984263bc
MD
820 return (EFBIG);
821 }
822
823 biosize = vp->v_mount->mnt_stat.f_iosize;
824
825 do {
826 /*
827 * Check for a valid write lease.
828 */
829 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
830 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
831 do {
3b568787 832 error = nqnfs_getlease(vp, ND_WRITE, td);
984263bc
MD
833 } while (error == NQNFS_EXPIRED);
834 if (error)
835 break;
836 if (np->n_lrev != np->n_brev ||
837 (np->n_flag & NQNFSNONCACHE)) {
3b568787 838 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
839 if (error)
840 break;
841 np->n_brev = np->n_lrev;
842 }
843 }
844 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
845 iomode = NFSV3WRITE_FILESYNC;
3b568787 846 error = nfs_writerpc(vp, uio, &iomode, &must_commit);
984263bc
MD
847 if (must_commit)
848 nfs_clearcommit(vp->v_mount);
849 break;
850 }
851 nfsstats.biocache_writes++;
852 lbn = uio->uio_offset / biosize;
853 on = uio->uio_offset & (biosize-1);
854 n = min((unsigned)(biosize - on), uio->uio_resid);
855again:
856 /*
857 * Handle direct append and file extension cases, calculate
858 * unaligned buffer size.
859 */
860
861 if (uio->uio_offset == np->n_size && n) {
862 /*
863 * Get the buffer (in its pre-append state to maintain
864 * B_CACHE if it was previously set). Resize the
865 * nfsnode after we have locked the buffer to prevent
866 * readers from reading garbage.
867 */
868 bcount = on;
dadab5e9 869 bp = nfs_getcacheblk(vp, lbn, bcount, td);
984263bc
MD
870
871 if (bp != NULL) {
872 long save;
873
874 np->n_size = uio->uio_offset + n;
875 np->n_flag |= NMODIFIED;
876 vnode_pager_setsize(vp, np->n_size);
877
878 save = bp->b_flags & B_CACHE;
879 bcount += n;
880 allocbuf(bp, bcount);
881 bp->b_flags |= save;
882 }
883 } else {
884 /*
885 * Obtain the locked cache block first, and then
886 * adjust the file's size as appropriate.
887 */
888 bcount = on + n;
889 if ((off_t)lbn * biosize + bcount < np->n_size) {
890 if ((off_t)(lbn + 1) * biosize < np->n_size)
891 bcount = biosize;
892 else
893 bcount = np->n_size - (off_t)lbn * biosize;
894 }
dadab5e9 895 bp = nfs_getcacheblk(vp, lbn, bcount, td);
984263bc
MD
896 if (uio->uio_offset + n > np->n_size) {
897 np->n_size = uio->uio_offset + n;
898 np->n_flag |= NMODIFIED;
899 vnode_pager_setsize(vp, np->n_size);
900 }
901 }
902
903 if (!bp) {
904 error = EINTR;
905 break;
906 }
907
908 /*
909 * Issue a READ if B_CACHE is not set. In special-append
910 * mode, B_CACHE is based on the buffer prior to the write
911 * op and is typically set, avoiding the read. If a read
912 * is required in special append mode, the server will
913 * probably send us a short-read since we extended the file
914 * on our end, resulting in b_resid == 0 and, thusly,
915 * B_CACHE getting set.
916 *
917 * We can also avoid issuing the read if the write covers
918 * the entire buffer. We have to make sure the buffer state
919 * is reasonable in this case since we will not be initiating
920 * I/O. See the comments in kern/vfs_bio.c's getblk() for
921 * more information.
922 *
923 * B_CACHE may also be set due to the buffer being cached
924 * normally.
925 */
926
927 if (on == 0 && n == bcount) {
928 bp->b_flags |= B_CACHE;
929 bp->b_flags &= ~(B_ERROR | B_INVAL);
930 }
931
932 if ((bp->b_flags & B_CACHE) == 0) {
933 bp->b_flags |= B_READ;
934 vfs_busy_pages(bp, 0);
3b568787 935 error = nfs_doio(bp, td);
984263bc
MD
936 if (error) {
937 brelse(bp);
938 break;
939 }
940 }
941 if (!bp) {
942 error = EINTR;
943 break;
944 }
984263bc
MD
945 np->n_flag |= NMODIFIED;
946
947 /*
948 * If dirtyend exceeds file size, chop it down. This should
949 * not normally occur but there is an append race where it
950 * might occur XXX, so we log it.
951 *
952 * If the chopping creates a reverse-indexed or degenerate
953 * situation with dirtyoff/end, we 0 both of them.
954 */
955
956 if (bp->b_dirtyend > bcount) {
957 printf("NFS append race @%lx:%d\n",
958 (long)bp->b_blkno * DEV_BSIZE,
959 bp->b_dirtyend - bcount);
960 bp->b_dirtyend = bcount;
961 }
962
963 if (bp->b_dirtyoff >= bp->b_dirtyend)
964 bp->b_dirtyoff = bp->b_dirtyend = 0;
965
966 /*
967 * If the new write will leave a contiguous dirty
968 * area, just update the b_dirtyoff and b_dirtyend,
969 * otherwise force a write rpc of the old dirty area.
970 *
971 * While it is possible to merge discontiguous writes due to
972 * our having a B_CACHE buffer ( and thus valid read data
973 * for the hole), we don't because it could lead to
974 * significant cache coherency problems with multiple clients,
975 * especially if locking is implemented later on.
976 *
977 * as an optimization we could theoretically maintain
978 * a linked list of discontinuous areas, but we would still
979 * have to commit them separately so there isn't much
980 * advantage to it except perhaps a bit of asynchronization.
981 */
982
983 if (bp->b_dirtyend > 0 &&
984 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
985 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) {
986 error = EINTR;
987 break;
988 }
989 goto again;
990 }
991
992 /*
993 * Check for valid write lease and get one as required.
994 * In case getblk() and/or bwrite() delayed us.
995 */
996 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
997 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
998 do {
3b568787 999 error = nqnfs_getlease(vp, ND_WRITE, td);
984263bc
MD
1000 } while (error == NQNFS_EXPIRED);
1001 if (error) {
1002 brelse(bp);
1003 break;
1004 }
1005 if (np->n_lrev != np->n_brev ||
1006 (np->n_flag & NQNFSNONCACHE)) {
1007 brelse(bp);
3b568787 1008 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
1009 if (error)
1010 break;
1011 np->n_brev = np->n_lrev;
1012 goto again;
1013 }
1014 }
1015
1016 error = uiomove((char *)bp->b_data + on, n, uio);
1017
1018 /*
1019 * Since this block is being modified, it must be written
1020 * again and not just committed. Since write clustering does
1021 * not work for the stage 1 data write, only the stage 2
1022 * commit rpc, we have to clear B_CLUSTEROK as well.
1023 */
1024 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1025
1026 if (error) {
1027 bp->b_flags |= B_ERROR;
1028 brelse(bp);
1029 break;
1030 }
1031
1032 /*
1033 * Only update dirtyoff/dirtyend if not a degenerate
1034 * condition.
1035 */
1036 if (n) {
1037 if (bp->b_dirtyend > 0) {
1038 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1039 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1040 } else {
1041 bp->b_dirtyoff = on;
1042 bp->b_dirtyend = on + n;
1043 }
1044 vfs_bio_set_validclean(bp, on, n);
1045 }
1046 /*
1047 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1048 * filesystem). XXX also use for loopback NFS mounts.
1049 */
1050 if (ioflag & IO_NOWDRAIN)
1051 bp->b_flags |= B_NOWDRAIN;
1052
1053 /*
1054 * If the lease is non-cachable or IO_SYNC do bwrite().
1055 *
1056 * IO_INVAL appears to be unused. The idea appears to be
1057 * to turn off caching in this case. Very odd. XXX
1058 */
1059 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1060 if (ioflag & IO_INVAL)
1061 bp->b_flags |= B_NOCACHE;
1062 error = VOP_BWRITE(bp->b_vp, bp);
1063 if (error)
1064 break;
1065 if (np->n_flag & NQNFSNONCACHE) {
3b568787 1066 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
984263bc
MD
1067 if (error)
1068 break;
1069 }
1070 } else if ((n + on) == biosize &&
1071 (nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1072 bp->b_flags |= B_ASYNC;
1073 (void)nfs_writebp(bp, 0, 0);
1074 } else {
1075 bdwrite(bp);
1076 }
1077 } while (uio->uio_resid > 0 && n > 0);
1078
1079 if (haverslock)
dadab5e9 1080 nfs_rsunlock(np, td);
984263bc
MD
1081
1082 return (error);
1083}
1084
1085/*
1086 * Get an nfs cache block.
1087 *
1088 * Allocate a new one if the block isn't currently in the cache
1089 * and return the block marked busy. If the calling process is
1090 * interrupted by a signal for an interruptible mount point, return
1091 * NULL.
1092 *
1093 * The caller must carefully deal with the possible B_INVAL state of
1094 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1095 * indirectly), so synchronous reads can be issued without worrying about
1096 * the B_INVAL state. We have to be a little more careful when dealing
1097 * with writes (see comments in nfs_write()) when extending a file past
1098 * its EOF.
1099 */
1100static struct buf *
dadab5e9 1101nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
984263bc 1102{
40393ded 1103 struct buf *bp;
984263bc
MD
1104 struct mount *mp;
1105 struct nfsmount *nmp;
1106
1107 mp = vp->v_mount;
1108 nmp = VFSTONFS(mp);
1109
1110 if (nmp->nm_flag & NFSMNT_INT) {
1111 bp = getblk(vp, bn, size, PCATCH, 0);
1112 while (bp == (struct buf *)0) {
dadab5e9 1113 if (nfs_sigintr(nmp, (struct nfsreq *)0, td))
984263bc
MD
1114 return ((struct buf *)0);
1115 bp = getblk(vp, bn, size, 0, 2 * hz);
1116 }
1117 } else {
1118 bp = getblk(vp, bn, size, 0, 0);
1119 }
1120
1121 if (vp->v_type == VREG) {
1122 int biosize;
1123
1124 biosize = mp->mnt_stat.f_iosize;
1125 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1126 }
1127 return (bp);
1128}
1129
1130/*
1131 * Flush and invalidate all dirty buffers. If another process is already
1132 * doing the flush, just wait for completion.
1133 */
1134int
3b568787 1135nfs_vinvalbuf(struct vnode *vp, int flags,
e851b29e 1136 struct thread *td, int intrflg)
984263bc 1137{
40393ded 1138 struct nfsnode *np = VTONFS(vp);
984263bc
MD
1139 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1140 int error = 0, slpflag, slptimeo;
1141
1142 if (vp->v_flag & VXLOCK) {
1143 return (0);
1144 }
1145
1146 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1147 intrflg = 0;
1148 if (intrflg) {
1149 slpflag = PCATCH;
1150 slptimeo = 2 * hz;
1151 } else {
1152 slpflag = 0;
1153 slptimeo = 0;
1154 }
1155 /*
1156 * First wait for any other process doing a flush to complete.
1157 */
1158 while (np->n_flag & NFLUSHINPROG) {
1159 np->n_flag |= NFLUSHWANT;
377d4740 1160 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
dadab5e9 1161 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td))
984263bc
MD
1162 return (EINTR);
1163 }
1164
1165 /*
1166 * Now, flush as required.
1167 */
1168 np->n_flag |= NFLUSHINPROG;
3b568787 1169 error = vinvalbuf(vp, flags, td, slpflag, 0);
984263bc 1170 while (error) {
dadab5e9 1171 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) {
984263bc
MD
1172 np->n_flag &= ~NFLUSHINPROG;
1173 if (np->n_flag & NFLUSHWANT) {
1174 np->n_flag &= ~NFLUSHWANT;
1175 wakeup((caddr_t)&np->n_flag);
1176 }
1177 return (EINTR);
1178 }
3b568787 1179 error = vinvalbuf(vp, flags, td, 0, slptimeo);
984263bc
MD
1180 }
1181 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1182 if (np->n_flag & NFLUSHWANT) {
1183 np->n_flag &= ~NFLUSHWANT;
1184 wakeup((caddr_t)&np->n_flag);
1185 }
1186 return (0);
1187}
1188
1189/*
1190 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1191 * This is mainly to avoid queueing async I/O requests when the nfsiods
1192 * are all hung on a dead server.
1193 *
1194 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1195 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1196 */
1197int
3b568787 1198nfs_asyncio(struct buf *bp, struct thread *td)
984263bc
MD
1199{
1200 struct nfsmount *nmp;
1201 int i;
1202 int gotiod;
1203 int slpflag = 0;
1204 int slptimeo = 0;
1205 int error;
1206
1207 /*
1208 * If no async daemons then return EIO to force caller to run the rpc
1209 * synchronously.
1210 */
1211 if (nfs_numasync == 0)
1212 return (EIO);
1213
1214 nmp = VFSTONFS(bp->b_vp->v_mount);
1215
1216 /*
1217 * Commits are usually short and sweet so lets save some cpu and
1218 * leave the async daemons for more important rpc's (such as reads
1219 * and writes).
1220 */
1221 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT &&
1222 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1223 return(EIO);
1224 }
1225
1226again:
1227 if (nmp->nm_flag & NFSMNT_INT)
1228 slpflag = PCATCH;
1229 gotiod = FALSE;
1230
1231 /*
1232 * Find a free iod to process this request.
1233 */
1234 for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1235 if (nfs_iodwant[i]) {
1236 /*
1237 * Found one, so wake it up and tell it which
1238 * mount to process.
1239 */
1240 NFS_DPF(ASYNCIO,
1241 ("nfs_asyncio: waking iod %d for mount %p\n",
1242 i, nmp));
dadab5e9 1243 nfs_iodwant[i] = NULL;
984263bc
MD
1244 nfs_iodmount[i] = nmp;
1245 nmp->nm_bufqiods++;
1246 wakeup((caddr_t)&nfs_iodwant[i]);
1247 gotiod = TRUE;
1248 break;
1249 }
1250
1251 /*
1252 * If none are free, we may already have an iod working on this mount
1253 * point. If so, it will process our request.
1254 */
1255 if (!gotiod) {
1256 if (nmp->nm_bufqiods > 0) {
1257 NFS_DPF(ASYNCIO,
1258 ("nfs_asyncio: %d iods are already processing mount %p\n",
1259 nmp->nm_bufqiods, nmp));
1260 gotiod = TRUE;
1261 }
1262 }
1263
1264 /*
1265 * If we have an iod which can process the request, then queue
1266 * the buffer.
1267 */
1268 if (gotiod) {
1269 /*
1270 * Ensure that the queue never grows too large. We still want
1271 * to asynchronize so we block rather then return EIO.
1272 */
1273 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1274 NFS_DPF(ASYNCIO,
1275 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1276 nmp->nm_bufqwant = TRUE;
377d4740 1277 error = tsleep(&nmp->nm_bufq, slpflag,
984263bc
MD
1278 "nfsaio", slptimeo);
1279 if (error) {
dadab5e9 1280 if (nfs_sigintr(nmp, NULL, td))
984263bc
MD
1281 return (EINTR);
1282 if (slpflag == PCATCH) {
1283 slpflag = 0;
1284 slptimeo = 2 * hz;
1285 }
1286 }
1287 /*
1288 * We might have lost our iod while sleeping,
1289 * so check and loop if nescessary.
1290 */
1291 if (nmp->nm_bufqiods == 0) {
1292 NFS_DPF(ASYNCIO,
1293 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1294 goto again;
1295 }
1296 }
1297
3b568787 1298 if ((bp->b_flags & B_READ) == 0)
984263bc 1299 bp->b_flags |= B_WRITEINPROG;
984263bc
MD
1300
1301 BUF_KERNPROC(bp);
1302 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1303 nmp->nm_bufqlen++;
1304 return (0);
1305 }
1306
1307 /*
1308 * All the iods are busy on other mounts, so return EIO to
1309 * force the caller to process the i/o synchronously.
1310 */
1311 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1312 return (EIO);
1313}
1314
1315/*
1316 * Do an I/O operation to/from a cache block. This may be called
1317 * synchronously or from an nfsiod.
dadab5e9
MD
1318 *
1319 * NOTE! TD MIGHT BE NULL
984263bc
MD
1320 */
1321int
3b568787 1322nfs_doio(struct buf *bp, struct thread *td)
984263bc
MD
1323{
1324 struct uio *uiop;
1325 struct vnode *vp;
1326 struct nfsnode *np;
1327 struct nfsmount *nmp;
1328 int error = 0, iomode, must_commit = 0;
1329 struct uio uio;
1330 struct iovec io;
1331
1332 vp = bp->b_vp;
1333 np = VTONFS(vp);
1334 nmp = VFSTONFS(vp->v_mount);
1335 uiop = &uio;
1336 uiop->uio_iov = &io;
1337 uiop->uio_iovcnt = 1;
1338 uiop->uio_segflg = UIO_SYSSPACE;
dadab5e9 1339 uiop->uio_td = td;
984263bc
MD
1340
1341 /*
1342 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1343 * do this here so we do not have to do it in all the code that
1344 * calls us.
1345 */
1346 bp->b_flags &= ~(B_ERROR | B_INVAL);
1347
1348 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1349
1350 /*
1351 * Historically, paging was done with physio, but no more.
1352 */
1353 if (bp->b_flags & B_PHYS) {
1354 /*
1355 * ...though reading /dev/drum still gets us here.
1356 */
1357 io.iov_len = uiop->uio_resid = bp->b_bcount;
1358 /* mapping was done by vmapbuf() */
1359 io.iov_base = bp->b_data;
1360 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1361 if (bp->b_flags & B_READ) {
1362 uiop->uio_rw = UIO_READ;
1363 nfsstats.read_physios++;
3b568787 1364 error = nfs_readrpc(vp, uiop);
984263bc
MD
1365 } else {
1366 int com;
1367
1368 iomode = NFSV3WRITE_DATASYNC;
1369 uiop->uio_rw = UIO_WRITE;
1370 nfsstats.write_physios++;
3b568787 1371 error = nfs_writerpc(vp, uiop, &iomode, &com);
984263bc
MD
1372 }
1373 if (error) {
1374 bp->b_flags |= B_ERROR;
1375 bp->b_error = error;
1376 }
1377 } else if (bp->b_flags & B_READ) {
1378 io.iov_len = uiop->uio_resid = bp->b_bcount;
1379 io.iov_base = bp->b_data;
1380 uiop->uio_rw = UIO_READ;
1381
1382 switch (vp->v_type) {
1383 case VREG:
1384 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1385 nfsstats.read_bios++;
3b568787 1386 error = nfs_readrpc(vp, uiop);
984263bc
MD
1387
1388 if (!error) {
1389 if (uiop->uio_resid) {
1390 /*
1391 * If we had a short read with no error, we must have
1392 * hit a file hole. We should zero-fill the remainder.
1393 * This can also occur if the server hits the file EOF.
1394 *
1395 * Holes used to be able to occur due to pending
1396 * writes, but that is not possible any longer.
1397 */
1398 int nread = bp->b_bcount - uiop->uio_resid;
1399 int left = uiop->uio_resid;
1400
1401 if (left > 0)
1402 bzero((char *)bp->b_data + nread, left);
1403 uiop->uio_resid = 0;
1404 }
1405 }
dadab5e9 1406 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
984263bc
MD
1407 (((nmp->nm_flag & NFSMNT_NQNFS) &&
1408 NQNFS_CKINVALID(vp, np, ND_READ) &&
1409 np->n_lrev != np->n_brev) ||
1410 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1411 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1412 uprintf("Process killed due to text file modification\n");
dadab5e9
MD
1413 psignal(td->td_proc, SIGKILL);
1414 PHOLD(td->td_proc);
984263bc
MD
1415 }
1416 break;
1417 case VLNK:
1418 uiop->uio_offset = (off_t)0;
1419 nfsstats.readlink_bios++;
3b568787 1420 error = nfs_readlinkrpc(vp, uiop);
984263bc
MD
1421 break;
1422 case VDIR:
1423 nfsstats.readdir_bios++;
1424 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1425 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
3b568787 1426 error = nfs_readdirplusrpc(vp, uiop);
984263bc
MD
1427 if (error == NFSERR_NOTSUPP)
1428 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1429 }
1430 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
3b568787 1431 error = nfs_readdirrpc(vp, uiop);
984263bc
MD
1432 /*
1433 * end-of-directory sets B_INVAL but does not generate an
1434 * error.
1435 */
1436 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1437 bp->b_flags |= B_INVAL;
1438 break;
1439 default:
1440 printf("nfs_doio: type %x unexpected\n",vp->v_type);
1441 break;
1442 };
1443 if (error) {
1444 bp->b_flags |= B_ERROR;
1445 bp->b_error = error;
1446 }
1447 } else {
1448 /*
1449 * If we only need to commit, try to commit
1450 */
1451 if (bp->b_flags & B_NEEDCOMMIT) {
1452 int retv;
1453 off_t off;
1454
1455 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1456 bp->b_flags |= B_WRITEINPROG;
3b568787
MD
1457 retv = nfs_commit(bp->b_vp, off,
1458 bp->b_dirtyend - bp->b_dirtyoff, td);
984263bc
MD
1459 bp->b_flags &= ~B_WRITEINPROG;
1460 if (retv == 0) {
1461 bp->b_dirtyoff = bp->b_dirtyend = 0;
1462 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1463 bp->b_resid = 0;
1464 biodone(bp);
1465 return (0);
1466 }
1467 if (retv == NFSERR_STALEWRITEVERF) {
1468 nfs_clearcommit(bp->b_vp->v_mount);
1469 }
1470 }
1471
1472 /*
1473 * Setup for actual write
1474 */
1475
1476 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1477 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1478
1479 if (bp->b_dirtyend > bp->b_dirtyoff) {
1480 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1481 - bp->b_dirtyoff;
1482 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1483 + bp->b_dirtyoff;
1484 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1485 uiop->uio_rw = UIO_WRITE;
1486 nfsstats.write_bios++;
1487
1488 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1489 iomode = NFSV3WRITE_UNSTABLE;
1490 else
1491 iomode = NFSV3WRITE_FILESYNC;
1492
1493 bp->b_flags |= B_WRITEINPROG;
3b568787 1494 error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
984263bc
MD
1495
1496 /*
1497 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1498 * to cluster the buffers needing commit. This will allow
1499 * the system to submit a single commit rpc for the whole
1500 * cluster. We can do this even if the buffer is not 100%
1501 * dirty (relative to the NFS blocksize), so we optimize the
1502 * append-to-file-case.
1503 *
1504 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1505 * cleared because write clustering only works for commit
1506 * rpc's, not for the data portion of the write).
1507 */
1508
1509 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1510 bp->b_flags |= B_NEEDCOMMIT;
1511 if (bp->b_dirtyoff == 0
1512 && bp->b_dirtyend == bp->b_bcount)
1513 bp->b_flags |= B_CLUSTEROK;
1514 } else {
1515 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1516 }
1517 bp->b_flags &= ~B_WRITEINPROG;
1518
1519 /*
1520 * For an interrupted write, the buffer is still valid
1521 * and the write hasn't been pushed to the server yet,
1522 * so we can't set B_ERROR and report the interruption
1523 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1524 * is not relevant, so the rpc attempt is essentially
1525 * a noop. For the case of a V3 write rpc not being
1526 * committed to stable storage, the block is still
1527 * dirty and requires either a commit rpc or another
1528 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1529 * the block is reused. This is indicated by setting
1530 * the B_DELWRI and B_NEEDCOMMIT flags.
1531 *
1532 * If the buffer is marked B_PAGING, it does not reside on
1533 * the vp's paging queues so we cannot call bdirty(). The
1534 * bp in this case is not an NFS cache block so we should
1535 * be safe. XXX
1536 */
1537 if (error == EINTR
1538 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1539 int s;
1540
1541 s = splbio();
1542 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1543 if ((bp->b_flags & B_PAGING) == 0) {
1544 bdirty(bp);
1545 bp->b_flags &= ~B_DONE;
1546 }
1547 if (error && (bp->b_flags & B_ASYNC) == 0)
1548 bp->b_flags |= B_EINTR;
1549 splx(s);
1550 } else {
1551 if (error) {
1552 bp->b_flags |= B_ERROR;
1553 bp->b_error = np->n_error = error;
1554 np->n_flag |= NWRITEERR;
1555 }
1556 bp->b_dirtyoff = bp->b_dirtyend = 0;
1557 }
1558 } else {
1559 bp->b_resid = 0;
1560 biodone(bp);
1561 return (0);
1562 }
1563 }
1564 bp->b_resid = uiop->uio_resid;
1565 if (must_commit)
1566 nfs_clearcommit(vp->v_mount);
1567 biodone(bp);
1568 return (error);
1569}
1570
1571/*
1572 * Used to aid in handling ftruncate() operations on the NFS client side.
1573 * Truncation creates a number of special problems for NFS. We have to
1574 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1575 * we have to properly handle VM pages or (potentially dirty) buffers
1576 * that straddle the truncation point.
1577 */
1578
1579int
3b568787 1580nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
984263bc
MD
1581{
1582 struct nfsnode *np = VTONFS(vp);
1583 u_quad_t tsize = np->n_size;
1584 int biosize = vp->v_mount->mnt_stat.f_iosize;
1585 int error = 0;
1586
1587 np->n_size = nsize;
1588
1589 if (np->n_size < tsize) {
1590 struct buf *bp;
1591 daddr_t lbn;
1592 int bufsize;
1593
1594 /*
1595 * vtruncbuf() doesn't get the buffer overlapping the
1596 * truncation point. We may have a B_DELWRI and/or B_CACHE
1597 * buffer that now needs to be truncated.
1598 */
3b568787 1599 error = vtruncbuf(vp, td, nsize, biosize);
984263bc
MD
1600 lbn = nsize / biosize;
1601 bufsize = nsize & (biosize - 1);
dadab5e9 1602 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
984263bc
MD
1603 if (bp->b_dirtyoff > bp->b_bcount)
1604 bp->b_dirtyoff = bp->b_bcount;
1605 if (bp->b_dirtyend > bp->b_bcount)
1606 bp->b_dirtyend = bp->b_bcount;
1607 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1608 brelse(bp);
1609 } else {
1610 vnode_pager_setsize(vp, nsize);
1611 }
1612 return(error);
1613}
1614