xref: /freebsd-14.2/sys/fs/nfsclient/nfs_clbio.c (revision f148aa57)
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  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
35  */
36 
37 #include <sys/cdefs.h>
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/bio.h>
41 #include <sys/buf.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/rwlock.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 
48 #include <vm/vm.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_extern.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_pager.h>
54 #include <vm/vnode_pager.h>
55 
56 #include <fs/nfs/nfsport.h>
57 #include <fs/nfsclient/nfsmount.h>
58 #include <fs/nfsclient/nfs.h>
59 #include <fs/nfsclient/nfsnode.h>
60 #include <fs/nfsclient/nfs_kdtrace.h>
61 
62 extern int newnfs_directio_allow_mmap;
63 extern struct nfsstatsv1 nfsstatsv1;
64 extern struct mtx ncl_iod_mutex;
65 extern int ncl_numasync;
66 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
67 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
68 extern int newnfs_directio_enable;
69 extern int nfs_keep_dirty_on_error;
70 
71 uma_zone_t ncl_pbuf_zone;
72 
73 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
74     struct thread *td);
75 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
76     struct ucred *cred, int ioflag);
77 
78 /*
79  * Vnode op for VM getpages.
80  */
81 SYSCTL_DECL(_vfs_nfs);
82 static int use_buf_pager = 1;
83 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
84     &use_buf_pager, 0,
85     "Use buffer pager instead of direct readrpc call");
86 
87 static daddr_t
ncl_gbp_getblkno(struct vnode * vp,vm_ooffset_t off)88 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
89 {
90 
91 	return (off / vp->v_bufobj.bo_bsize);
92 }
93 
94 static int
ncl_gbp_getblksz(struct vnode * vp,daddr_t lbn,long * sz)95 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn, long *sz)
96 {
97 	struct nfsnode *np;
98 	u_quad_t nsize;
99 	int biosize, bcount;
100 
101 	np = VTONFS(vp);
102 	NFSLOCKNODE(np);
103 	nsize = np->n_size;
104 	NFSUNLOCKNODE(np);
105 
106 	biosize = vp->v_bufobj.bo_bsize;
107 	bcount = biosize;
108 	if ((off_t)lbn * biosize >= nsize)
109 		bcount = 0;
110 	else if ((off_t)(lbn + 1) * biosize > nsize)
111 		bcount = nsize - (off_t)lbn * biosize;
112 	*sz = bcount;
113 	return (0);
114 }
115 
116 int
ncl_getpages(struct vop_getpages_args * ap)117 ncl_getpages(struct vop_getpages_args *ap)
118 {
119 	int i, error, nextoff, size, toff, count, npages;
120 	struct uio uio;
121 	struct iovec iov;
122 	vm_offset_t kva;
123 	struct buf *bp;
124 	struct vnode *vp;
125 	struct thread *td;
126 	struct ucred *cred;
127 	struct nfsmount *nmp;
128 	vm_object_t object;
129 	vm_page_t *pages;
130 	struct nfsnode *np;
131 
132 	vp = ap->a_vp;
133 	np = VTONFS(vp);
134 	td = curthread;
135 	cred = curthread->td_ucred;
136 	nmp = VFSTONFS(vp->v_mount);
137 	pages = ap->a_m;
138 	npages = ap->a_count;
139 
140 	if ((object = vp->v_object) == NULL) {
141 		printf("ncl_getpages: called with non-merged cache vnode\n");
142 		return (VM_PAGER_ERROR);
143 	}
144 
145 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
146 		NFSLOCKNODE(np);
147 		if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
148 			NFSUNLOCKNODE(np);
149 			printf("ncl_getpages: called on non-cacheable vnode\n");
150 			return (VM_PAGER_ERROR);
151 		} else
152 			NFSUNLOCKNODE(np);
153 	}
154 
155 	mtx_lock(&nmp->nm_mtx);
156 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
157 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
158 		mtx_unlock(&nmp->nm_mtx);
159 		/* We'll never get here for v4, because we always have fsinfo */
160 		(void)ncl_fsinfo(nmp, vp, cred, td);
161 	} else
162 		mtx_unlock(&nmp->nm_mtx);
163 
164 	if (use_buf_pager)
165 		return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
166 		    ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
167 
168 	/*
169 	 * If the requested page is partially valid, just return it and
170 	 * allow the pager to zero-out the blanks.  Partially valid pages
171 	 * can only occur at the file EOF.
172 	 *
173 	 * XXXGL: is that true for NFS, where short read can occur???
174 	 */
175 	VM_OBJECT_WLOCK(object);
176 	if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
177 		goto out;
178 	VM_OBJECT_WUNLOCK(object);
179 
180 	/*
181 	 * We use only the kva address for the buffer, but this is extremely
182 	 * convenient and fast.
183 	 */
184 	bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
185 
186 	kva = (vm_offset_t) bp->b_data;
187 	pmap_qenter(kva, pages, npages);
188 	VM_CNT_INC(v_vnodein);
189 	VM_CNT_ADD(v_vnodepgsin, npages);
190 
191 	count = npages << PAGE_SHIFT;
192 	iov.iov_base = (caddr_t) kva;
193 	iov.iov_len = count;
194 	uio.uio_iov = &iov;
195 	uio.uio_iovcnt = 1;
196 	uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
197 	uio.uio_resid = count;
198 	uio.uio_segflg = UIO_SYSSPACE;
199 	uio.uio_rw = UIO_READ;
200 	uio.uio_td = td;
201 
202 	error = ncl_readrpc(vp, &uio, cred);
203 	pmap_qremove(kva, npages);
204 
205 	uma_zfree(ncl_pbuf_zone, bp);
206 
207 	if (error && (uio.uio_resid == count)) {
208 		printf("ncl_getpages: error %d\n", error);
209 		return (VM_PAGER_ERROR);
210 	}
211 
212 	/*
213 	 * Calculate the number of bytes read and validate only that number
214 	 * of bytes.  Note that due to pending writes, size may be 0.  This
215 	 * does not mean that the remaining data is invalid!
216 	 */
217 
218 	size = count - uio.uio_resid;
219 	VM_OBJECT_WLOCK(object);
220 	for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
221 		vm_page_t m;
222 		nextoff = toff + PAGE_SIZE;
223 		m = pages[i];
224 
225 		if (nextoff <= size) {
226 			/*
227 			 * Read operation filled an entire page
228 			 */
229 			vm_page_valid(m);
230 			KASSERT(m->dirty == 0,
231 			    ("nfs_getpages: page %p is dirty", m));
232 		} else if (size > toff) {
233 			/*
234 			 * Read operation filled a partial page.
235 			 */
236 			vm_page_invalid(m);
237 			vm_page_set_valid_range(m, 0, size - toff);
238 			KASSERT(m->dirty == 0,
239 			    ("nfs_getpages: page %p is dirty", m));
240 		} else {
241 			/*
242 			 * Read operation was short.  If no error
243 			 * occurred we may have hit a zero-fill
244 			 * section.  We leave valid set to 0, and page
245 			 * is freed by vm_page_readahead_finish() if
246 			 * its index is not equal to requested, or
247 			 * page is zeroed and set valid by
248 			 * vm_pager_get_pages() for requested page.
249 			 */
250 			;
251 		}
252 	}
253 out:
254 	VM_OBJECT_WUNLOCK(object);
255 	if (ap->a_rbehind)
256 		*ap->a_rbehind = 0;
257 	if (ap->a_rahead)
258 		*ap->a_rahead = 0;
259 	return (VM_PAGER_OK);
260 }
261 
262 /*
263  * Vnode op for VM putpages.
264  */
265 int
ncl_putpages(struct vop_putpages_args * ap)266 ncl_putpages(struct vop_putpages_args *ap)
267 {
268 	struct uio uio;
269 	struct iovec iov;
270 	int i, error, npages, count;
271 	off_t offset;
272 	int *rtvals;
273 	struct vnode *vp;
274 	struct thread *td;
275 	struct ucred *cred;
276 	struct nfsmount *nmp;
277 	struct nfsnode *np;
278 	vm_page_t *pages;
279 
280 	vp = ap->a_vp;
281 	np = VTONFS(vp);
282 	td = curthread;				/* XXX */
283 	/* Set the cred to n_writecred for the write rpcs. */
284 	if (np->n_writecred != NULL)
285 		cred = crhold(np->n_writecred);
286 	else
287 		cred = crhold(curthread->td_ucred);	/* XXX */
288 	nmp = VFSTONFS(vp->v_mount);
289 	pages = ap->a_m;
290 	count = ap->a_count;
291 	rtvals = ap->a_rtvals;
292 	npages = btoc(count);
293 	offset = IDX_TO_OFF(pages[0]->pindex);
294 
295 	mtx_lock(&nmp->nm_mtx);
296 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
297 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
298 		mtx_unlock(&nmp->nm_mtx);
299 		(void)ncl_fsinfo(nmp, vp, cred, td);
300 	} else
301 		mtx_unlock(&nmp->nm_mtx);
302 
303 	NFSLOCKNODE(np);
304 	if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
305 	    (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
306 		NFSUNLOCKNODE(np);
307 		printf("ncl_putpages: called on noncache-able vnode\n");
308 		NFSLOCKNODE(np);
309 	}
310 	/*
311 	 * When putting pages, do not extend file past EOF.
312 	 */
313 	if (offset + count > np->n_size) {
314 		count = np->n_size - offset;
315 		if (count < 0)
316 			count = 0;
317 	}
318 	NFSUNLOCKNODE(np);
319 
320 	for (i = 0; i < npages; i++)
321 		rtvals[i] = VM_PAGER_ERROR;
322 
323 	VM_CNT_INC(v_vnodeout);
324 	VM_CNT_ADD(v_vnodepgsout, count);
325 
326 	iov.iov_base = unmapped_buf;
327 	iov.iov_len = count;
328 	uio.uio_iov = &iov;
329 	uio.uio_iovcnt = 1;
330 	uio.uio_offset = offset;
331 	uio.uio_resid = count;
332 	uio.uio_segflg = UIO_NOCOPY;
333 	uio.uio_rw = UIO_WRITE;
334 	uio.uio_td = td;
335 
336 	error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
337 	    cred);
338 	crfree(cred);
339 
340 	if (error == 0 || !nfs_keep_dirty_on_error) {
341 		vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
342 		    np->n_size - offset, npages * PAGE_SIZE);
343 	}
344 	return (rtvals[0]);
345 }
346 
347 /*
348  * For nfs, cache consistency can only be maintained approximately.
349  * Although RFC1094 does not specify the criteria, the following is
350  * believed to be compatible with the reference port.
351  * For nfs:
352  * If the file's modify time on the server has changed since the
353  * last read rpc or you have written to the file,
354  * you may have lost data cache consistency with the
355  * server, so flush all of the file's data out of the cache.
356  * Then force a getattr rpc to ensure that you have up to date
357  * attributes.
358  * NB: This implies that cache data can be read when up to
359  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
360  * attributes this could be forced by setting n_attrstamp to 0 before
361  * the VOP_GETATTR() call.
362  */
363 static inline int
nfs_bioread_check_cons(struct vnode * vp,struct thread * td,struct ucred * cred)364 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
365 {
366 	int error = 0;
367 	struct vattr vattr;
368 	struct nfsnode *np = VTONFS(vp);
369 	bool old_lock;
370 
371 	/*
372 	 * Ensure the exclusive access to the node before checking
373 	 * whether the cache is consistent.
374 	 */
375 	old_lock = ncl_excl_start(vp);
376 	NFSLOCKNODE(np);
377 	if (np->n_flag & NMODIFIED) {
378 		NFSUNLOCKNODE(np);
379 		if (vp->v_type != VREG) {
380 			if (vp->v_type != VDIR)
381 				panic("nfs: bioread, not dir");
382 			ncl_invaldir(vp);
383 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
384 			if (error != 0)
385 				goto out;
386 		}
387 		np->n_attrstamp = 0;
388 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
389 		error = VOP_GETATTR(vp, &vattr, cred);
390 		if (error)
391 			goto out;
392 		NFSLOCKNODE(np);
393 		np->n_mtime = vattr.va_mtime;
394 		NFSUNLOCKNODE(np);
395 	} else {
396 		NFSUNLOCKNODE(np);
397 		error = VOP_GETATTR(vp, &vattr, cred);
398 		if (error)
399 			goto out;
400 		NFSLOCKNODE(np);
401 		if ((np->n_flag & NSIZECHANGED)
402 		    || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
403 			NFSUNLOCKNODE(np);
404 			if (vp->v_type == VDIR)
405 				ncl_invaldir(vp);
406 			error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
407 			if (error != 0)
408 				goto out;
409 			NFSLOCKNODE(np);
410 			np->n_mtime = vattr.va_mtime;
411 			np->n_flag &= ~NSIZECHANGED;
412 		}
413 		NFSUNLOCKNODE(np);
414 	}
415 out:
416 	ncl_excl_finish(vp, old_lock);
417 	return (error);
418 }
419 
420 static bool
ncl_bioread_dora(struct vnode * vp)421 ncl_bioread_dora(struct vnode *vp)
422 {
423 	vm_object_t obj;
424 
425 	obj = vp->v_object;
426 	if (obj == NULL)
427 		return (true);
428 	return (!vm_object_mightbedirty(vp->v_object) &&
429 	    vp->v_object->un_pager.vnp.writemappings == 0);
430 }
431 
432 /*
433  * Vnode op for read using bio
434  */
435 int
ncl_bioread(struct vnode * vp,struct uio * uio,int ioflag,struct ucred * cred)436 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
437 {
438 	struct nfsnode *np = VTONFS(vp);
439 	struct buf *bp, *rabp;
440 	struct thread *td;
441 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
442 	daddr_t lbn, rabn;
443 	int biosize, bcount, error, i, n, nra, on, save2, seqcount;
444 	off_t tmp_off;
445 
446 	KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
447 	if (uio->uio_resid == 0)
448 		return (0);
449 	if (uio->uio_offset < 0)	/* XXX VDIR cookies can be negative */
450 		return (EINVAL);
451 	td = uio->uio_td;
452 
453 	mtx_lock(&nmp->nm_mtx);
454 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
455 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
456 		mtx_unlock(&nmp->nm_mtx);
457 		(void)ncl_fsinfo(nmp, vp, cred, td);
458 		mtx_lock(&nmp->nm_mtx);
459 	}
460 	if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
461 		(void) newnfs_iosize(nmp);
462 
463 	tmp_off = uio->uio_offset + uio->uio_resid;
464 	if (vp->v_type != VDIR &&
465 	    (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
466 		mtx_unlock(&nmp->nm_mtx);
467 		return (EFBIG);
468 	}
469 	mtx_unlock(&nmp->nm_mtx);
470 
471 	if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
472 		/* No caching/ no readaheads. Just read data into the user buffer */
473 		return ncl_readrpc(vp, uio, cred);
474 
475 	n = 0;
476 	on = 0;
477 	biosize = vp->v_bufobj.bo_bsize;
478 	seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
479 
480 	error = nfs_bioread_check_cons(vp, td, cred);
481 	if (error)
482 		return error;
483 
484 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
485 	do {
486 	    u_quad_t nsize;
487 
488 	    NFSLOCKNODE(np);
489 	    nsize = np->n_size;
490 	    NFSUNLOCKNODE(np);
491 
492 	    switch (vp->v_type) {
493 	    case VREG:
494 		NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
495 		lbn = uio->uio_offset / biosize;
496 		on = uio->uio_offset - (lbn * biosize);
497 
498 		/*
499 		 * Start the read ahead(s), as required.  Do not do
500 		 * read-ahead if there are writeable mappings, since
501 		 * unlocked read by nfsiod could obliterate changes
502 		 * done by userspace.
503 		 */
504 		if (nmp->nm_readahead > 0 && ncl_bioread_dora(vp)) {
505 		    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
506 			(off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
507 			rabn = lbn + 1 + nra;
508 			if (incore(&vp->v_bufobj, rabn) == NULL) {
509 			    rabp = nfs_getcacheblk(vp, rabn, biosize, td);
510 			    if (!rabp) {
511 				error = newnfs_sigintr(nmp, td);
512 				if (error == 0)
513 					error = EINTR;
514 				goto out;
515 			    }
516 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
517 				rabp->b_flags |= B_ASYNC;
518 				rabp->b_iocmd = BIO_READ;
519 				vfs_busy_pages(rabp, 0);
520 				if (ncl_asyncio(nmp, rabp, cred, td)) {
521 				    rabp->b_flags |= B_INVAL;
522 				    rabp->b_ioflags |= BIO_ERROR;
523 				    vfs_unbusy_pages(rabp);
524 				    brelse(rabp);
525 				    break;
526 				}
527 			    } else {
528 				brelse(rabp);
529 			    }
530 			}
531 		    }
532 		}
533 
534 		/* Note that bcount is *not* DEV_BSIZE aligned. */
535 		bcount = biosize;
536 		if ((off_t)lbn * biosize >= nsize) {
537 			bcount = 0;
538 		} else if ((off_t)(lbn + 1) * biosize > nsize) {
539 			bcount = nsize - (off_t)lbn * biosize;
540 		}
541 		bp = nfs_getcacheblk(vp, lbn, bcount, td);
542 
543 		if (!bp) {
544 			error = newnfs_sigintr(nmp, td);
545 			if (error == 0)
546 				error = EINTR;
547 			goto out;
548 		}
549 
550 		/*
551 		 * If B_CACHE is not set, we must issue the read.  If this
552 		 * fails, we return an error.
553 		 */
554 
555 		if ((bp->b_flags & B_CACHE) == 0) {
556 		    bp->b_iocmd = BIO_READ;
557 		    vfs_busy_pages(bp, 0);
558 		    error = ncl_doio(vp, bp, cred, td, 0);
559 		    if (error) {
560 			brelse(bp);
561 			goto out;
562 		    }
563 		}
564 
565 		/*
566 		 * on is the offset into the current bp.  Figure out how many
567 		 * bytes we can copy out of the bp.  Note that bcount is
568 		 * NOT DEV_BSIZE aligned.
569 		 *
570 		 * Then figure out how many bytes we can copy into the uio.
571 		 */
572 
573 		n = 0;
574 		if (on < bcount)
575 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
576 		break;
577 	    case VLNK:
578 		NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
579 		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
580 		if (!bp) {
581 			error = newnfs_sigintr(nmp, td);
582 			if (error == 0)
583 				error = EINTR;
584 			goto out;
585 		}
586 		if ((bp->b_flags & B_CACHE) == 0) {
587 		    bp->b_iocmd = BIO_READ;
588 		    vfs_busy_pages(bp, 0);
589 		    error = ncl_doio(vp, bp, cred, td, 0);
590 		    if (error) {
591 			bp->b_ioflags |= BIO_ERROR;
592 			brelse(bp);
593 			goto out;
594 		    }
595 		}
596 		n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
597 		on = 0;
598 		break;
599 	    case VDIR:
600 		NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
601 		NFSLOCKNODE(np);
602 		if (np->n_direofoffset
603 		    && uio->uio_offset >= np->n_direofoffset) {
604 			NFSUNLOCKNODE(np);
605 			error = 0;
606 			goto out;
607 		}
608 		NFSUNLOCKNODE(np);
609 		lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
610 		on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
611 		bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
612 		if (!bp) {
613 			error = newnfs_sigintr(nmp, td);
614 			if (error == 0)
615 				error = EINTR;
616 			goto out;
617 		}
618 		if ((bp->b_flags & B_CACHE) == 0) {
619 		    bp->b_iocmd = BIO_READ;
620 		    vfs_busy_pages(bp, 0);
621 		    error = ncl_doio(vp, bp, cred, td, 0);
622 		    if (error) {
623 			    brelse(bp);
624 		    }
625 		    while (error == NFSERR_BAD_COOKIE) {
626 			ncl_invaldir(vp);
627 			error = ncl_vinvalbuf(vp, 0, td, 1);
628 
629 			/*
630 			 * Yuck! The directory has been modified on the
631 			 * server. The only way to get the block is by
632 			 * reading from the beginning to get all the
633 			 * offset cookies.
634 			 *
635 			 * Leave the last bp intact unless there is an error.
636 			 * Loop back up to the while if the error is another
637 			 * NFSERR_BAD_COOKIE (double yuch!).
638 			 */
639 			for (i = 0; i <= lbn && !error; i++) {
640 			    NFSLOCKNODE(np);
641 			    if (np->n_direofoffset
642 				&& (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
643 				    NFSUNLOCKNODE(np);
644 				    error = 0;
645 				    goto out;
646 			    }
647 			    NFSUNLOCKNODE(np);
648 			    bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
649 			    if (!bp) {
650 				error = newnfs_sigintr(nmp, td);
651 				if (error == 0)
652 					error = EINTR;
653 				goto out;
654 			    }
655 			    if ((bp->b_flags & B_CACHE) == 0) {
656 				    bp->b_iocmd = BIO_READ;
657 				    vfs_busy_pages(bp, 0);
658 				    error = ncl_doio(vp, bp, cred, td, 0);
659 				    /*
660 				     * no error + B_INVAL == directory EOF,
661 				     * use the block.
662 				     */
663 				    if (error == 0 && (bp->b_flags & B_INVAL))
664 					    break;
665 			    }
666 			    /*
667 			     * An error will throw away the block and the
668 			     * for loop will break out.  If no error and this
669 			     * is not the block we want, we throw away the
670 			     * block and go for the next one via the for loop.
671 			     */
672 			    if (error || i < lbn)
673 				    brelse(bp);
674 			}
675 		    }
676 		    /*
677 		     * The above while is repeated if we hit another cookie
678 		     * error.  If we hit an error and it wasn't a cookie error,
679 		     * we give up.
680 		     */
681 		    if (error)
682 			    goto out;
683 		}
684 
685 		/*
686 		 * If not eof and read aheads are enabled, start one.
687 		 * (You need the current block first, so that you have the
688 		 *  directory offset cookie of the next block.)
689 		 */
690 		NFSLOCKNODE(np);
691 		if (nmp->nm_readahead > 0 && ncl_bioread_dora(vp) &&
692 		    (bp->b_flags & B_INVAL) == 0 &&
693 		    (np->n_direofoffset == 0 ||
694 		    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
695 		    incore(&vp->v_bufobj, lbn + 1) == NULL) {
696 			NFSUNLOCKNODE(np);
697 			rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
698 			if (rabp) {
699 			    if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
700 				rabp->b_flags |= B_ASYNC;
701 				rabp->b_iocmd = BIO_READ;
702 				vfs_busy_pages(rabp, 0);
703 				if (ncl_asyncio(nmp, rabp, cred, td)) {
704 				    rabp->b_flags |= B_INVAL;
705 				    rabp->b_ioflags |= BIO_ERROR;
706 				    vfs_unbusy_pages(rabp);
707 				    brelse(rabp);
708 				}
709 			    } else {
710 				brelse(rabp);
711 			    }
712 			}
713 			NFSLOCKNODE(np);
714 		}
715 		/*
716 		 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
717 		 * chopped for the EOF condition, we cannot tell how large
718 		 * NFS directories are going to be until we hit EOF.  So
719 		 * an NFS directory buffer is *not* chopped to its EOF.  Now,
720 		 * it just so happens that b_resid will effectively chop it
721 		 * to EOF.  *BUT* this information is lost if the buffer goes
722 		 * away and is reconstituted into a B_CACHE state ( due to
723 		 * being VMIO ) later.  So we keep track of the directory eof
724 		 * in np->n_direofoffset and chop it off as an extra step
725 		 * right here.
726 		 */
727 		n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
728 		if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
729 			n = np->n_direofoffset - uio->uio_offset;
730 		NFSUNLOCKNODE(np);
731 		break;
732 	    default:
733 		printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
734 		bp = NULL;
735 		break;
736 	    }
737 
738 	    if (n > 0) {
739 		    error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
740 	    }
741 	    if (vp->v_type == VLNK)
742 		n = 0;
743 	    if (bp != NULL)
744 		brelse(bp);
745 	} while (error == 0 && uio->uio_resid > 0 && n > 0);
746 out:
747 	curthread_pflags2_restore(save2);
748 	if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
749 		NFSLOCKNODE(np);
750 		ncl_pager_setsize(vp, NULL);
751 	}
752 	return (error);
753 }
754 
755 /*
756  * The NFS write path cannot handle iovecs with len > 1. So we need to
757  * break up iovecs accordingly (restricting them to wsize).
758  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
759  * For the ASYNC case, 2 copies are needed. The first a copy from the
760  * user buffer to a staging buffer and then a second copy from the staging
761  * buffer to mbufs. This can be optimized by copying from the user buffer
762  * directly into mbufs and passing the chain down, but that requires a
763  * fair amount of re-working of the relevant codepaths (and can be done
764  * later).
765  */
766 static int
nfs_directio_write(struct vnode * vp,struct uio * uiop,struct ucred * cred,int ioflag)767 nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred,
768     int ioflag)
769 {
770 	struct uio uio;
771 	struct iovec iov;
772 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
773 	struct thread *td = uiop->uio_td;
774 	int error, iomode, must_commit, size, wsize;
775 
776 	KASSERT((ioflag & IO_SYNC) != 0, ("nfs_directio_write: not sync"));
777 	mtx_lock(&nmp->nm_mtx);
778 	wsize = nmp->nm_wsize;
779 	mtx_unlock(&nmp->nm_mtx);
780 	while (uiop->uio_resid > 0) {
781 		size = MIN(uiop->uio_resid, wsize);
782 		size = MIN(uiop->uio_iov->iov_len, size);
783 		iov.iov_base = uiop->uio_iov->iov_base;
784 		iov.iov_len = size;
785 		uio.uio_iov = &iov;
786 		uio.uio_iovcnt = 1;
787 		uio.uio_offset = uiop->uio_offset;
788 		uio.uio_resid = size;
789 		uio.uio_segflg = uiop->uio_segflg;
790 		uio.uio_rw = UIO_WRITE;
791 		uio.uio_td = td;
792 		iomode = NFSWRITE_FILESYNC;
793 		/*
794 		 * When doing direct I/O we do not care if the
795 		 * server's write verifier has changed, but we
796 		 * do not want to update the verifier if it has
797 		 * changed, since that hides the change from
798 		 * writes being done through the buffer cache.
799 		 * By passing must_commit in set to two, the code
800 		 * in nfsrpc_writerpc() will not update the
801 		 * verifier on the mount point.
802 		 */
803 		must_commit = 2;
804 		error = ncl_writerpc(vp, &uio, cred, &iomode,
805 		    &must_commit, 0, ioflag);
806 		KASSERT(must_commit == 2,
807 		    ("ncl_directio_write: Updated write verifier"));
808 		if (error != 0)
809 			return (error);
810 		if (iomode != NFSWRITE_FILESYNC)
811 			printf("nfs_directio_write: Broken server "
812 			    "did not reply FILE_SYNC\n");
813 		uiop->uio_offset += size;
814 		uiop->uio_resid -= size;
815 		if (uiop->uio_iov->iov_len <= size) {
816 			uiop->uio_iovcnt--;
817 			uiop->uio_iov++;
818 		} else {
819 			uiop->uio_iov->iov_base =
820 				(char *)uiop->uio_iov->iov_base + size;
821 			uiop->uio_iov->iov_len -= size;
822 		}
823 	}
824 	return (0);
825 }
826 
827 /*
828  * Vnode op for write using bio
829  */
830 int
ncl_write(struct vop_write_args * ap)831 ncl_write(struct vop_write_args *ap)
832 {
833 	int biosize;
834 	struct uio *uio = ap->a_uio;
835 	struct thread *td = uio->uio_td;
836 	struct vnode *vp = ap->a_vp;
837 	struct nfsnode *np = VTONFS(vp);
838 	struct ucred *cred = ap->a_cred;
839 	int ioflag = ap->a_ioflag;
840 	struct buf *bp;
841 	struct vattr vattr;
842 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
843 	daddr_t lbn;
844 	int bcount, noncontig_write, obcount;
845 	int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
846 	size_t orig_resid, local_resid;
847 	off_t orig_size, tmp_off;
848 	struct timespec ts;
849 
850 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
851 	KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
852 	    ("ncl_write proc"));
853 	if (vp->v_type != VREG)
854 		return (EIO);
855 	NFSLOCKNODE(np);
856 	if (np->n_flag & NWRITEERR) {
857 		np->n_flag &= ~NWRITEERR;
858 		NFSUNLOCKNODE(np);
859 		return (np->n_error);
860 	} else
861 		NFSUNLOCKNODE(np);
862 	mtx_lock(&nmp->nm_mtx);
863 	if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
864 	    (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
865 		mtx_unlock(&nmp->nm_mtx);
866 		(void)ncl_fsinfo(nmp, vp, cred, td);
867 		mtx_lock(&nmp->nm_mtx);
868 	}
869 	if (nmp->nm_wsize == 0)
870 		(void) newnfs_iosize(nmp);
871 	mtx_unlock(&nmp->nm_mtx);
872 
873 	/*
874 	 * Synchronously flush pending buffers if we are in synchronous
875 	 * mode or if we are appending.
876 	 */
877 	if ((ioflag & IO_APPEND) || ((ioflag & IO_SYNC) && (np->n_flag &
878 	    NMODIFIED))) {
879 		/*
880 		 * For the case where IO_APPEND is being done using a
881 		 * direct output (to the NFS server) RPC and
882 		 * newnfs_directio_enable is 0, all buffer cache buffers,
883 		 * including ones not modified, must be invalidated.
884 		 * This ensures that stale data is not read out of the
885 		 * buffer cache.  The call also invalidates all mapped
886 		 * pages and, since the exclusive lock is held on the vnode,
887 		 * new pages cannot be faulted in.
888 		 *
889 		 * For the case where newnfs_directio_enable is set
890 		 * (which is not the default), it is not obvious that
891 		 * stale data should be left in the buffer cache, but
892 		 * the code has been this way for over a decade without
893 		 * complaints.  Note that, unlike doing IO_APPEND via
894 		 * a direct write RPC when newnfs_directio_enable is not set,
895 		 * when newnfs_directio_enable is set, reading is done via
896 		 * direct to NFS server RPCs as well.
897 		 */
898 		np->n_attrstamp = 0;
899 		KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
900 		error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
901 		    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
902 		if (error != 0)
903 			return (error);
904 	}
905 
906 	orig_resid = uio->uio_resid;
907 	NFSLOCKNODE(np);
908 	orig_size = np->n_size;
909 	NFSUNLOCKNODE(np);
910 
911 	/*
912 	 * If IO_APPEND then load uio_offset.  We restart here if we cannot
913 	 * get the append lock.
914 	 */
915 	if (ioflag & IO_APPEND) {
916 		/*
917 		 * For NFSv4, the AppendWrite will Verify the size against
918 		 * the file's size on the server.  If not the same, the
919 		 * write will then be retried, using the file size returned
920 		 * by the AppendWrite.  However, for NFSv2 and NFSv3, the
921 		 * size must be acquired here via a Getattr RPC.
922 		 * The AppendWrite is not done for a pNFS mount.
923 		 */
924 		if (!NFSHASNFSV4(nmp) || NFSHASPNFS(nmp)) {
925 			np->n_attrstamp = 0;
926 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
927 			error = VOP_GETATTR(vp, &vattr, cred);
928 			if (error)
929 				return (error);
930 		}
931 		NFSLOCKNODE(np);
932 		uio->uio_offset = np->n_size;
933 		NFSUNLOCKNODE(np);
934 	}
935 
936 	if (uio->uio_offset < 0)
937 		return (EINVAL);
938 	tmp_off = uio->uio_offset + uio->uio_resid;
939 	if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
940 		return (EFBIG);
941 	if (uio->uio_resid == 0)
942 		return (0);
943 
944 	/*
945 	 * Do IO_APPEND writing via a synchronous direct write.
946 	 * This can result in a significant performance improvement.
947 	 */
948 	if ((newnfs_directio_enable && (ioflag & IO_DIRECT)) ||
949 	    (ioflag & IO_APPEND)) {
950 		/*
951 		 * Direct writes to the server must be done NFSWRITE_FILESYNC,
952 		 * because the write data is not cached and, therefore, the
953 		 * write cannot be redone after a server reboot.
954 		 * Set IO_SYNC to make this happen.
955 		 */
956 		ioflag |= IO_SYNC;
957 		return (nfs_directio_write(vp, uio, cred, ioflag));
958 	}
959 
960 	/*
961 	 * Maybe this should be above the vnode op call, but so long as
962 	 * file servers have no limits, i don't think it matters
963 	 */
964 	error = vn_rlimit_fsize(vp, uio, td);
965 	if (error != 0)
966 		return (error);
967 
968 	save2 = curthread_pflags2_set(TDP2_SBPAGES);
969 	biosize = vp->v_bufobj.bo_bsize;
970 	/*
971 	 * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
972 	 * would exceed the local maximum per-file write commit size when
973 	 * combined with those, we must decide whether to flush,
974 	 * go synchronous, or return error.  We don't bother checking
975 	 * IO_UNIT -- we just make all writes atomic anyway, as there's
976 	 * no point optimizing for something that really won't ever happen.
977 	 */
978 	wouldcommit = 0;
979 	if (!(ioflag & IO_SYNC)) {
980 		int nflag;
981 
982 		NFSLOCKNODE(np);
983 		nflag = np->n_flag;
984 		NFSUNLOCKNODE(np);
985 		if (nflag & NMODIFIED) {
986 			BO_LOCK(&vp->v_bufobj);
987 			if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
988 				TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
989 				    b_bobufs) {
990 					if (bp->b_flags & B_NEEDCOMMIT)
991 						wouldcommit += bp->b_bcount;
992 				}
993 			}
994 			BO_UNLOCK(&vp->v_bufobj);
995 		}
996 	}
997 
998 	do {
999 		if (!(ioflag & IO_SYNC)) {
1000 			wouldcommit += biosize;
1001 			if (wouldcommit > nmp->nm_wcommitsize) {
1002 				np->n_attrstamp = 0;
1003 				KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1004 				error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1005 				    IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1006 				if (error != 0)
1007 					goto out;
1008 				wouldcommit = biosize;
1009 			}
1010 		}
1011 
1012 		NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1013 		lbn = uio->uio_offset / biosize;
1014 		on = uio->uio_offset - (lbn * biosize);
1015 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
1016 again:
1017 		/*
1018 		 * Handle direct append and file extension cases, calculate
1019 		 * unaligned buffer size.
1020 		 */
1021 		NFSLOCKNODE(np);
1022 		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1023 		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1024 			noncontig_write = 1;
1025 		else
1026 			noncontig_write = 0;
1027 		if ((uio->uio_offset == np->n_size ||
1028 		    (noncontig_write != 0 &&
1029 		    lbn == (np->n_size / biosize) &&
1030 		    uio->uio_offset + n > np->n_size)) && n) {
1031 			NFSUNLOCKNODE(np);
1032 			/*
1033 			 * Get the buffer (in its pre-append state to maintain
1034 			 * B_CACHE if it was previously set).  Resize the
1035 			 * nfsnode after we have locked the buffer to prevent
1036 			 * readers from reading garbage.
1037 			 */
1038 			obcount = np->n_size - (lbn * biosize);
1039 			bp = nfs_getcacheblk(vp, lbn, obcount, td);
1040 
1041 			if (bp != NULL) {
1042 				long save;
1043 
1044 				NFSLOCKNODE(np);
1045 				np->n_size = uio->uio_offset + n;
1046 				np->n_flag |= NMODIFIED;
1047 				np->n_flag &= ~NVNSETSZSKIP;
1048 				vnode_pager_setsize(vp, np->n_size);
1049 				NFSUNLOCKNODE(np);
1050 
1051 				save = bp->b_flags & B_CACHE;
1052 				bcount = on + n;
1053 				allocbuf(bp, bcount);
1054 				bp->b_flags |= save;
1055 				if (noncontig_write != 0 && on > obcount)
1056 					vfs_bio_bzero_buf(bp, obcount, on -
1057 					    obcount);
1058 			}
1059 		} else {
1060 			/*
1061 			 * Obtain the locked cache block first, and then
1062 			 * adjust the file's size as appropriate.
1063 			 */
1064 			bcount = on + n;
1065 			if ((off_t)lbn * biosize + bcount < np->n_size) {
1066 				if ((off_t)(lbn + 1) * biosize < np->n_size)
1067 					bcount = biosize;
1068 				else
1069 					bcount = np->n_size - (off_t)lbn * biosize;
1070 			}
1071 			NFSUNLOCKNODE(np);
1072 			bp = nfs_getcacheblk(vp, lbn, bcount, td);
1073 			NFSLOCKNODE(np);
1074 			if (uio->uio_offset + n > np->n_size) {
1075 				np->n_size = uio->uio_offset + n;
1076 				np->n_flag |= NMODIFIED;
1077 				np->n_flag &= ~NVNSETSZSKIP;
1078 				vnode_pager_setsize(vp, np->n_size);
1079 			}
1080 			NFSUNLOCKNODE(np);
1081 		}
1082 
1083 		if (!bp) {
1084 			error = newnfs_sigintr(nmp, td);
1085 			if (!error)
1086 				error = EINTR;
1087 			break;
1088 		}
1089 
1090 		/*
1091 		 * Issue a READ if B_CACHE is not set.  In special-append
1092 		 * mode, B_CACHE is based on the buffer prior to the write
1093 		 * op and is typically set, avoiding the read.  If a read
1094 		 * is required in special append mode, the server will
1095 		 * probably send us a short-read since we extended the file
1096 		 * on our end, resulting in b_resid == 0 and, thusly,
1097 		 * B_CACHE getting set.
1098 		 *
1099 		 * We can also avoid issuing the read if the write covers
1100 		 * the entire buffer.  We have to make sure the buffer state
1101 		 * is reasonable in this case since we will not be initiating
1102 		 * I/O.  See the comments in kern/vfs_bio.c's getblk() for
1103 		 * more information.
1104 		 *
1105 		 * B_CACHE may also be set due to the buffer being cached
1106 		 * normally.
1107 		 */
1108 
1109 		bp_cached = 1;
1110 		if (on == 0 && n == bcount) {
1111 			if ((bp->b_flags & B_CACHE) == 0)
1112 				bp_cached = 0;
1113 			bp->b_flags |= B_CACHE;
1114 			bp->b_flags &= ~B_INVAL;
1115 			bp->b_ioflags &= ~BIO_ERROR;
1116 		}
1117 
1118 		if ((bp->b_flags & B_CACHE) == 0) {
1119 			bp->b_iocmd = BIO_READ;
1120 			vfs_busy_pages(bp, 0);
1121 			error = ncl_doio(vp, bp, cred, td, 0);
1122 			if (error) {
1123 				brelse(bp);
1124 				break;
1125 			}
1126 		}
1127 		if (bp->b_wcred == NOCRED)
1128 			bp->b_wcred = crhold(cred);
1129 		NFSLOCKNODE(np);
1130 		np->n_flag |= NMODIFIED;
1131 		NFSUNLOCKNODE(np);
1132 
1133 		/*
1134 		 * If dirtyend exceeds file size, chop it down.  This should
1135 		 * not normally occur but there is an append race where it
1136 		 * might occur XXX, so we log it.
1137 		 *
1138 		 * If the chopping creates a reverse-indexed or degenerate
1139 		 * situation with dirtyoff/end, we 0 both of them.
1140 		 */
1141 
1142 		if (bp->b_dirtyend > bcount) {
1143 			printf("NFS append race @%lx:%d\n",
1144 			    (long)bp->b_blkno * DEV_BSIZE,
1145 			    bp->b_dirtyend - bcount);
1146 			bp->b_dirtyend = bcount;
1147 		}
1148 
1149 		if (bp->b_dirtyoff >= bp->b_dirtyend)
1150 			bp->b_dirtyoff = bp->b_dirtyend = 0;
1151 
1152 		/*
1153 		 * If the new write will leave a contiguous dirty
1154 		 * area, just update the b_dirtyoff and b_dirtyend,
1155 		 * otherwise force a write rpc of the old dirty area.
1156 		 *
1157 		 * If there has been a file lock applied to this file
1158 		 * or vfs.nfs.old_noncontig_writing is set, do the following:
1159 		 * While it is possible to merge discontiguous writes due to
1160 		 * our having a B_CACHE buffer ( and thus valid read data
1161 		 * for the hole), we don't because it could lead to
1162 		 * significant cache coherency problems with multiple clients,
1163 		 * especially if locking is implemented later on.
1164 		 *
1165 		 * If vfs.nfs.old_noncontig_writing is not set and there has
1166 		 * not been file locking done on this file:
1167 		 * Relax coherency a bit for the sake of performance and
1168 		 * expand the current dirty region to contain the new
1169 		 * write even if it means we mark some non-dirty data as
1170 		 * dirty.
1171 		 */
1172 
1173 		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1174 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1175 			if (bwrite(bp) == EINTR) {
1176 				error = EINTR;
1177 				break;
1178 			}
1179 			goto again;
1180 		}
1181 
1182 		local_resid = uio->uio_resid;
1183 		error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1184 
1185 		if (error != 0 && !bp_cached) {
1186 			/*
1187 			 * This block has no other content then what
1188 			 * possibly was written by the faulty uiomove.
1189 			 * Release it, forgetting the data pages, to
1190 			 * prevent the leak of uninitialized data to
1191 			 * usermode.
1192 			 */
1193 			bp->b_ioflags |= BIO_ERROR;
1194 			brelse(bp);
1195 			uio->uio_offset -= local_resid - uio->uio_resid;
1196 			uio->uio_resid = local_resid;
1197 			break;
1198 		}
1199 
1200 		/*
1201 		 * Since this block is being modified, it must be written
1202 		 * again and not just committed.  Since write clustering does
1203 		 * not work for the stage 1 data write, only the stage 2
1204 		 * commit rpc, we have to clear B_CLUSTEROK as well.
1205 		 */
1206 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1207 
1208 		/*
1209 		 * Get the partial update on the progress made from
1210 		 * uiomove, if an error occurred.
1211 		 */
1212 		if (error != 0)
1213 			n = local_resid - uio->uio_resid;
1214 
1215 		/*
1216 		 * Only update dirtyoff/dirtyend if not a degenerate
1217 		 * condition.
1218 		 */
1219 		if (n > 0) {
1220 			if (bp->b_dirtyend > 0) {
1221 				bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1222 				bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1223 			} else {
1224 				bp->b_dirtyoff = on;
1225 				bp->b_dirtyend = on + n;
1226 			}
1227 			vfs_bio_set_valid(bp, on, n);
1228 		}
1229 
1230 		/*
1231 		 * If IO_SYNC do bwrite().
1232 		 *
1233 		 * IO_INVAL appears to be unused.  The idea appears to be
1234 		 * to turn off caching in this case.  Very odd.  XXX
1235 		 */
1236 		if ((ioflag & IO_SYNC)) {
1237 			if (ioflag & IO_INVAL)
1238 				bp->b_flags |= B_NOCACHE;
1239 			error1 = bwrite(bp);
1240 			if (error1 != 0) {
1241 				if (error == 0)
1242 					error = error1;
1243 				break;
1244 			}
1245 		} else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1246 			bp->b_flags |= B_ASYNC;
1247 			(void) bwrite(bp);
1248 		} else {
1249 			bdwrite(bp);
1250 		}
1251 
1252 		if (error != 0)
1253 			break;
1254 	} while (uio->uio_resid > 0 && n > 0);
1255 
1256 	if (error == 0) {
1257 		nanouptime(&ts);
1258 		NFSLOCKNODE(np);
1259 		np->n_localmodtime = ts;
1260 		NFSUNLOCKNODE(np);
1261 	} else {
1262 		if (ioflag & IO_UNIT) {
1263 			VATTR_NULL(&vattr);
1264 			vattr.va_size = orig_size;
1265 			/* IO_SYNC is handled implicitely */
1266 			(void)VOP_SETATTR(vp, &vattr, cred);
1267 			uio->uio_offset -= orig_resid - uio->uio_resid;
1268 			uio->uio_resid = orig_resid;
1269 		}
1270 	}
1271 
1272 out:
1273 	curthread_pflags2_restore(save2);
1274 	return (error);
1275 }
1276 
1277 /*
1278  * Get an nfs cache block.
1279  *
1280  * Allocate a new one if the block isn't currently in the cache
1281  * and return the block marked busy. If the calling process is
1282  * interrupted by a signal for an interruptible mount point, return
1283  * NULL.
1284  *
1285  * The caller must carefully deal with the possible B_INVAL state of
1286  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1287  * indirectly), so synchronous reads can be issued without worrying about
1288  * the B_INVAL state.  We have to be a little more careful when dealing
1289  * with writes (see comments in nfs_write()) when extending a file past
1290  * its EOF.
1291  */
1292 static struct buf *
nfs_getcacheblk(struct vnode * vp,daddr_t bn,int size,struct thread * td)1293 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1294 {
1295 	struct buf *bp;
1296 	struct mount *mp;
1297 	struct nfsmount *nmp;
1298 
1299 	mp = vp->v_mount;
1300 	nmp = VFSTONFS(mp);
1301 
1302 	if (nmp->nm_flag & NFSMNT_INT) {
1303 		sigset_t oldset;
1304 
1305 		newnfs_set_sigmask(td, &oldset);
1306 		bp = getblk(vp, bn, size, PCATCH, 0, 0);
1307 		newnfs_restore_sigmask(td, &oldset);
1308 		while (bp == NULL) {
1309 			if (newnfs_sigintr(nmp, td))
1310 				return (NULL);
1311 			bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1312 		}
1313 	} else {
1314 		bp = getblk(vp, bn, size, 0, 0, 0);
1315 	}
1316 
1317 	if (vp->v_type == VREG)
1318 		bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1319 	return (bp);
1320 }
1321 
1322 /*
1323  * Flush and invalidate all dirty buffers. If another process is already
1324  * doing the flush, just wait for completion.
1325  */
1326 int
ncl_vinvalbuf(struct vnode * vp,int flags,struct thread * td,int intrflg)1327 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1328 {
1329 	struct nfsnode *np = VTONFS(vp);
1330 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1331 	int error = 0, slpflag, slptimeo;
1332 	bool old_lock;
1333 	struct timespec ts;
1334 
1335 	ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1336 
1337 	if ((nmp->nm_flag & NFSMNT_INT) == 0)
1338 		intrflg = 0;
1339 	if (NFSCL_FORCEDISM(nmp->nm_mountp))
1340 		intrflg = 1;
1341 	if (intrflg) {
1342 		slpflag = PCATCH;
1343 		slptimeo = 2 * hz;
1344 	} else {
1345 		slpflag = 0;
1346 		slptimeo = 0;
1347 	}
1348 
1349 	old_lock = ncl_excl_start(vp);
1350 	if (old_lock)
1351 		flags |= V_ALLOWCLEAN;
1352 
1353 	/*
1354 	 * Now, flush as required.
1355 	 */
1356 	if ((flags & (V_SAVE | V_VMIO)) == V_SAVE) {
1357 		vnode_pager_clean_sync(vp);
1358 
1359 		/*
1360 		 * If the page clean was interrupted, fail the invalidation.
1361 		 * Not doing so, we run the risk of losing dirty pages in the
1362 		 * vinvalbuf() call below.
1363 		 */
1364 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1365 			goto out;
1366 	}
1367 
1368 	error = vinvalbuf(vp, flags, slpflag, 0);
1369 	while (error) {
1370 		if (intrflg && (error = newnfs_sigintr(nmp, td)))
1371 			goto out;
1372 		error = vinvalbuf(vp, flags, 0, slptimeo);
1373 	}
1374 	if (NFSHASPNFS(nmp)) {
1375 		nfscl_layoutcommit(vp, td);
1376 		nanouptime(&ts);
1377 		/*
1378 		 * Invalidate the attribute cache, since writes to a DS
1379 		 * won't update the size attribute.
1380 		 */
1381 		NFSLOCKNODE(np);
1382 		np->n_attrstamp = 0;
1383 	} else {
1384 		nanouptime(&ts);
1385 		NFSLOCKNODE(np);
1386 	}
1387 	if ((np->n_flag & NMODIFIED) != 0) {
1388 		np->n_localmodtime = ts;
1389 		np->n_flag &= ~NMODIFIED;
1390 	}
1391 	NFSUNLOCKNODE(np);
1392 out:
1393 	ncl_excl_finish(vp, old_lock);
1394 	return error;
1395 }
1396 
1397 /*
1398  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1399  * This is mainly to avoid queueing async I/O requests when the nfsiods
1400  * are all hung on a dead server.
1401  *
1402  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1403  * is eventually dequeued by the async daemon, ncl_doio() *will*.
1404  */
1405 int
ncl_asyncio(struct nfsmount * nmp,struct buf * bp,struct ucred * cred,struct thread * td)1406 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1407 {
1408 	int iod;
1409 	int gotiod;
1410 	int slpflag = 0;
1411 	int slptimeo = 0;
1412 	int error, error2;
1413 
1414 	/*
1415 	 * Commits are usually short and sweet so lets save some cpu and
1416 	 * leave the async daemons for more important rpc's (such as reads
1417 	 * and writes).
1418 	 *
1419 	 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1420 	 * in the directory in order to update attributes. This can deadlock
1421 	 * with another thread that is waiting for async I/O to be done by
1422 	 * an nfsiod thread while holding a lock on one of these vnodes.
1423 	 * To avoid this deadlock, don't allow the async nfsiod threads to
1424 	 * perform Readdirplus RPCs.
1425 	 */
1426 	NFSLOCKIOD();
1427 	if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1428 	     (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1429 	    (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1430 		NFSUNLOCKIOD();
1431 		return(EIO);
1432 	}
1433 again:
1434 	if (nmp->nm_flag & NFSMNT_INT)
1435 		slpflag = PCATCH;
1436 	gotiod = FALSE;
1437 
1438 	/*
1439 	 * Find a free iod to process this request.
1440 	 */
1441 	for (iod = 0; iod < ncl_numasync; iod++)
1442 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1443 			gotiod = TRUE;
1444 			break;
1445 		}
1446 
1447 	/*
1448 	 * Try to create one if none are free.
1449 	 */
1450 	if (!gotiod)
1451 		ncl_nfsiodnew();
1452 	else {
1453 		/*
1454 		 * Found one, so wake it up and tell it which
1455 		 * mount to process.
1456 		 */
1457 		NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1458 		    iod, nmp));
1459 		ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1460 		ncl_iodmount[iod] = nmp;
1461 		nmp->nm_bufqiods++;
1462 		wakeup(&ncl_iodwant[iod]);
1463 	}
1464 
1465 	/*
1466 	 * If none are free, we may already have an iod working on this mount
1467 	 * point.  If so, it will process our request.
1468 	 */
1469 	if (!gotiod) {
1470 		if (nmp->nm_bufqiods > 0) {
1471 			NFS_DPF(ASYNCIO,
1472 				("ncl_asyncio: %d iods are already processing mount %p\n",
1473 				 nmp->nm_bufqiods, nmp));
1474 			gotiod = TRUE;
1475 		}
1476 	}
1477 
1478 	/*
1479 	 * If we have an iod which can process the request, then queue
1480 	 * the buffer.
1481 	 */
1482 	if (gotiod) {
1483 		/*
1484 		 * Ensure that the queue never grows too large.  We still want
1485 		 * to asynchronize so we block rather then return EIO.
1486 		 */
1487 		while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1488 			NFS_DPF(ASYNCIO,
1489 				("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1490 			nmp->nm_bufqwant = TRUE;
1491 			error = newnfs_msleep(td, &nmp->nm_bufq,
1492 			    &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1493 			   slptimeo);
1494 			if (error) {
1495 				error2 = newnfs_sigintr(nmp, td);
1496 				if (error2) {
1497 					NFSUNLOCKIOD();
1498 					return (error2);
1499 				}
1500 				if (slpflag == PCATCH) {
1501 					slpflag = 0;
1502 					slptimeo = 2 * hz;
1503 				}
1504 			}
1505 			/*
1506 			 * We might have lost our iod while sleeping,
1507 			 * so check and loop if necessary.
1508 			 */
1509 			goto again;
1510 		}
1511 
1512 		/* We might have lost our nfsiod */
1513 		if (nmp->nm_bufqiods == 0) {
1514 			NFS_DPF(ASYNCIO,
1515 				("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1516 			goto again;
1517 		}
1518 
1519 		if (bp->b_iocmd == BIO_READ) {
1520 			if (bp->b_rcred == NOCRED && cred != NOCRED)
1521 				bp->b_rcred = crhold(cred);
1522 		} else {
1523 			if (bp->b_wcred == NOCRED && cred != NOCRED)
1524 				bp->b_wcred = crhold(cred);
1525 		}
1526 
1527 		if (bp->b_flags & B_REMFREE)
1528 			bremfreef(bp);
1529 		BUF_KERNPROC(bp);
1530 		TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1531 		nmp->nm_bufqlen++;
1532 		KASSERT((bp->b_flags & B_DIRECT) == 0,
1533 		    ("ncl_asyncio: B_DIRECT set"));
1534 		NFSUNLOCKIOD();
1535 		return (0);
1536 	}
1537 
1538 	NFSUNLOCKIOD();
1539 
1540 	/*
1541 	 * All the iods are busy on other mounts, so return EIO to
1542 	 * force the caller to process the i/o synchronously.
1543 	 */
1544 	NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1545 	return (EIO);
1546 }
1547 
1548 /*
1549  * Do an I/O operation to/from a cache block. This may be called
1550  * synchronously or from an nfsiod.
1551  */
1552 int
ncl_doio(struct vnode * vp,struct buf * bp,struct ucred * cr,struct thread * td,int called_from_strategy)1553 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1554     int called_from_strategy)
1555 {
1556 	struct uio *uiop;
1557 	struct nfsnode *np;
1558 	struct nfsmount *nmp;
1559 	int error = 0, iomode, must_commit = 0;
1560 	struct uio uio;
1561 	struct iovec io;
1562 	struct proc *p = td ? td->td_proc : NULL;
1563 	uint8_t	iocmd;
1564 
1565 	np = VTONFS(vp);
1566 	nmp = VFSTONFS(vp->v_mount);
1567 	uiop = &uio;
1568 	uiop->uio_iov = &io;
1569 	uiop->uio_iovcnt = 1;
1570 	uiop->uio_segflg = UIO_SYSSPACE;
1571 	uiop->uio_td = td;
1572 
1573 	/*
1574 	 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
1575 	 * do this here so we do not have to do it in all the code that
1576 	 * calls us.
1577 	 */
1578 	bp->b_flags &= ~B_INVAL;
1579 	bp->b_ioflags &= ~BIO_ERROR;
1580 
1581 	KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1582 	iocmd = bp->b_iocmd;
1583 	if (iocmd == BIO_READ) {
1584 	    io.iov_len = uiop->uio_resid = bp->b_bcount;
1585 	    io.iov_base = bp->b_data;
1586 	    uiop->uio_rw = UIO_READ;
1587 
1588 	    switch (vp->v_type) {
1589 	    case VREG:
1590 		uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1591 		NFSINCRGLOBAL(nfsstatsv1.read_bios);
1592 		error = ncl_readrpc(vp, uiop, cr);
1593 
1594 		if (!error) {
1595 		    if (uiop->uio_resid) {
1596 			/*
1597 			 * If we had a short read with no error, we must have
1598 			 * hit a file hole.  We should zero-fill the remainder.
1599 			 * This can also occur if the server hits the file EOF.
1600 			 *
1601 			 * Holes used to be able to occur due to pending
1602 			 * writes, but that is not possible any longer.
1603 			 */
1604 			int nread = bp->b_bcount - uiop->uio_resid;
1605 			ssize_t left = uiop->uio_resid;
1606 
1607 			if (left > 0)
1608 				bzero((char *)bp->b_data + nread, left);
1609 			uiop->uio_resid = 0;
1610 		    }
1611 		}
1612 		/* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1613 		if (p && vp->v_writecount <= -1) {
1614 			NFSLOCKNODE(np);
1615 			if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1616 				NFSUNLOCKNODE(np);
1617 				PROC_LOCK(p);
1618 				killproc(p, "text file modification");
1619 				PROC_UNLOCK(p);
1620 			} else
1621 				NFSUNLOCKNODE(np);
1622 		}
1623 		break;
1624 	    case VLNK:
1625 		uiop->uio_offset = (off_t)0;
1626 		NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1627 		error = ncl_readlinkrpc(vp, uiop, cr);
1628 		break;
1629 	    case VDIR:
1630 		NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1631 		uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1632 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1633 			error = ncl_readdirplusrpc(vp, uiop, cr, td);
1634 			if (error == NFSERR_NOTSUPP)
1635 				nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1636 		}
1637 		if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1638 			error = ncl_readdirrpc(vp, uiop, cr, td);
1639 		/*
1640 		 * end-of-directory sets B_INVAL but does not generate an
1641 		 * error.
1642 		 */
1643 		if (error == 0 && uiop->uio_resid == bp->b_bcount)
1644 			bp->b_flags |= B_INVAL;
1645 		break;
1646 	    default:
1647 		printf("ncl_doio:  type %x unexpected\n", vp->v_type);
1648 		break;
1649 	    }
1650 	    if (error) {
1651 		bp->b_ioflags |= BIO_ERROR;
1652 		bp->b_error = error;
1653 	    }
1654 	} else {
1655 	    /*
1656 	     * If we only need to commit, try to commit
1657 	     */
1658 	    if (bp->b_flags & B_NEEDCOMMIT) {
1659 		    int retv;
1660 		    off_t off;
1661 
1662 		    off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1663 		    retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1664 			bp->b_wcred, td);
1665 		    if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
1666 			    bp->b_dirtyoff = bp->b_dirtyend = 0;
1667 			    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1668 			    bp->b_resid = 0;
1669 			    bufdone(bp);
1670 			    return (0);
1671 		    }
1672 		    if (retv == NFSERR_STALEWRITEVERF) {
1673 			    ncl_clearcommit(vp->v_mount);
1674 		    }
1675 	    }
1676 
1677 	    /*
1678 	     * Setup for actual write
1679 	     */
1680 	    NFSLOCKNODE(np);
1681 	    if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1682 		bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1683 	    NFSUNLOCKNODE(np);
1684 
1685 	    if (bp->b_dirtyend > bp->b_dirtyoff) {
1686 		io.iov_len = uiop->uio_resid = bp->b_dirtyend
1687 		    - bp->b_dirtyoff;
1688 		uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1689 		    + bp->b_dirtyoff;
1690 		io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1691 		uiop->uio_rw = UIO_WRITE;
1692 		NFSINCRGLOBAL(nfsstatsv1.write_bios);
1693 
1694 		if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1695 		    iomode = NFSWRITE_UNSTABLE;
1696 		else
1697 		    iomode = NFSWRITE_FILESYNC;
1698 
1699 		error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1700 		    called_from_strategy, 0);
1701 
1702 		/*
1703 		 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1704 		 * to cluster the buffers needing commit.  This will allow
1705 		 * the system to submit a single commit rpc for the whole
1706 		 * cluster.  We can do this even if the buffer is not 100%
1707 		 * dirty (relative to the NFS blocksize), so we optimize the
1708 		 * append-to-file-case.
1709 		 *
1710 		 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1711 		 * cleared because write clustering only works for commit
1712 		 * rpc's, not for the data portion of the write).
1713 		 */
1714 
1715 		if (!error && iomode == NFSWRITE_UNSTABLE) {
1716 		    bp->b_flags |= B_NEEDCOMMIT;
1717 		    if (bp->b_dirtyoff == 0
1718 			&& bp->b_dirtyend == bp->b_bcount)
1719 			bp->b_flags |= B_CLUSTEROK;
1720 		} else {
1721 		    bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1722 		}
1723 
1724 		/*
1725 		 * For an interrupted write, the buffer is still valid
1726 		 * and the write hasn't been pushed to the server yet,
1727 		 * so we can't set BIO_ERROR and report the interruption
1728 		 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1729 		 * is not relevant, so the rpc attempt is essentially
1730 		 * a noop.  For the case of a V3 write rpc not being
1731 		 * committed to stable storage, the block is still
1732 		 * dirty and requires either a commit rpc or another
1733 		 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1734 		 * the block is reused. This is indicated by setting
1735 		 * the B_DELWRI and B_NEEDCOMMIT flags.
1736 		 *
1737 		 * EIO is returned by ncl_writerpc() to indicate a recoverable
1738 		 * write error and is handled as above, except that
1739 		 * B_EINTR isn't set. One cause of this is a stale stateid
1740 		 * error for the RPC that indicates recovery is required,
1741 		 * when called with called_from_strategy != 0.
1742 		 *
1743 		 * If the buffer is marked B_PAGING, it does not reside on
1744 		 * the vp's paging queues so we cannot call bdirty().  The
1745 		 * bp in this case is not an NFS cache block so we should
1746 		 * be safe. XXX
1747 		 *
1748 		 * The logic below breaks up errors into recoverable and
1749 		 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1750 		 * and keep the buffer around for potential write retries.
1751 		 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1752 		 * and save the error in the nfsnode. This is less than ideal
1753 		 * but necessary. Keeping such buffers around could potentially
1754 		 * cause buffer exhaustion eventually (they can never be written
1755 		 * out, so will get constantly be re-dirtied). It also causes
1756 		 * all sorts of vfs panics. For non-recoverable write errors,
1757 		 * also invalidate the attrcache, so we'll be forced to go over
1758 		 * the wire for this object, returning an error to user on next
1759 		 * call (most of the time).
1760 		 */
1761 		if (error == EINTR || error == EIO || error == ETIMEDOUT
1762 		    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1763 			bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1764 			if ((bp->b_flags & B_PAGING) == 0) {
1765 			    bdirty(bp);
1766 			    bp->b_flags &= ~B_DONE;
1767 			}
1768 			if ((error == EINTR || error == ETIMEDOUT) &&
1769 			    (bp->b_flags & B_ASYNC) == 0)
1770 			    bp->b_flags |= B_EINTR;
1771 		} else {
1772 		    if (error) {
1773 			bp->b_ioflags |= BIO_ERROR;
1774 			bp->b_flags |= B_INVAL;
1775 			bp->b_error = np->n_error = error;
1776 			NFSLOCKNODE(np);
1777 			np->n_flag |= NWRITEERR;
1778 			np->n_attrstamp = 0;
1779 			KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1780 			NFSUNLOCKNODE(np);
1781 		    }
1782 		    bp->b_dirtyoff = bp->b_dirtyend = 0;
1783 		}
1784 	    } else {
1785 		bp->b_resid = 0;
1786 		bufdone(bp);
1787 		return (0);
1788 	    }
1789 	}
1790 	bp->b_resid = uiop->uio_resid;
1791 	if (must_commit == 1)
1792 	    ncl_clearcommit(vp->v_mount);
1793 	bufdone(bp);
1794 	return (error);
1795 }
1796 
1797 /*
1798  * Used to aid in handling ftruncate() operations on the NFS client side.
1799  * Truncation creates a number of special problems for NFS.  We have to
1800  * throw away VM pages and buffer cache buffers that are beyond EOF, and
1801  * we have to properly handle VM pages or (potentially dirty) buffers
1802  * that straddle the truncation point.
1803  */
1804 
1805 int
ncl_meta_setsize(struct vnode * vp,struct thread * td,u_quad_t nsize)1806 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1807 {
1808 	struct nfsnode *np = VTONFS(vp);
1809 	u_quad_t tsize;
1810 	int biosize = vp->v_bufobj.bo_bsize;
1811 	int error = 0;
1812 
1813 	NFSLOCKNODE(np);
1814 	tsize = np->n_size;
1815 	np->n_size = nsize;
1816 	NFSUNLOCKNODE(np);
1817 
1818 	if (nsize < tsize) {
1819 		struct buf *bp;
1820 		daddr_t lbn;
1821 		int bufsize;
1822 
1823 		/*
1824 		 * vtruncbuf() doesn't get the buffer overlapping the
1825 		 * truncation point.  We may have a B_DELWRI and/or B_CACHE
1826 		 * buffer that now needs to be truncated.
1827 		 */
1828 		error = vtruncbuf(vp, nsize, biosize);
1829 		lbn = nsize / biosize;
1830 		bufsize = nsize - (lbn * biosize);
1831 		bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1832 		if (!bp)
1833 			return EINTR;
1834 		if (bp->b_dirtyoff > bp->b_bcount)
1835 			bp->b_dirtyoff = bp->b_bcount;
1836 		if (bp->b_dirtyend > bp->b_bcount)
1837 			bp->b_dirtyend = bp->b_bcount;
1838 		bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
1839 		brelse(bp);
1840 	} else {
1841 		vnode_pager_setsize(vp, nsize);
1842 	}
1843 	return(error);
1844 }
1845