xref: /illumos-gate/usr/src/uts/common/fs/nfs/nfs_client.c (revision 0dfe541e)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 
26 /*
27  *	Copyright (c) 1983,1984,1985,1986,1987,1988,1989  AT&T.
28  *	All rights reserved.
29  */
30 
31 /*
32  * Copyright 2018 Nexenta Systems, Inc.
33  */
34 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/systm.h>
38 #include <sys/thread.h>
39 #include <sys/t_lock.h>
40 #include <sys/time.h>
41 #include <sys/vnode.h>
42 #include <sys/vfs.h>
43 #include <sys/errno.h>
44 #include <sys/buf.h>
45 #include <sys/stat.h>
46 #include <sys/cred.h>
47 #include <sys/kmem.h>
48 #include <sys/debug.h>
49 #include <sys/dnlc.h>
50 #include <sys/vmsystm.h>
51 #include <sys/flock.h>
52 #include <sys/share.h>
53 #include <sys/cmn_err.h>
54 #include <sys/tiuser.h>
55 #include <sys/sysmacros.h>
56 #include <sys/callb.h>
57 #include <sys/acl.h>
58 #include <sys/kstat.h>
59 #include <sys/signal.h>
60 #include <sys/list.h>
61 #include <sys/zone.h>
62 
63 #include <rpc/types.h>
64 #include <rpc/xdr.h>
65 #include <rpc/auth.h>
66 #include <rpc/clnt.h>
67 
68 #include <nfs/nfs.h>
69 #include <nfs/nfs_clnt.h>
70 #include <nfs/nfs_cmd.h>
71 
72 #include <nfs/rnode.h>
73 #include <nfs/nfs_acl.h>
74 #include <nfs/lm.h>
75 
76 #include <vm/hat.h>
77 #include <vm/as.h>
78 #include <vm/page.h>
79 #include <vm/pvn.h>
80 #include <vm/seg.h>
81 #include <vm/seg_map.h>
82 #include <vm/seg_vn.h>
83 
84 static void	nfs3_attr_cache(vnode_t *, vattr_t *, vattr_t *, hrtime_t,
85 			cred_t *);
86 static int	nfs_getattr_cache(vnode_t *, struct vattr *);
87 static int	nfs_remove_locking_id(vnode_t *, int, char *, char *, int *);
88 
89 struct mi_globals {
90 	kmutex_t	mig_lock;  /* lock protecting mig_list */
91 	list_t		mig_list;  /* list of NFS v2 or v3 mounts in zone */
92 	boolean_t	mig_destructor_called;
93 };
94 
95 static zone_key_t mi_list_key;
96 
97 /* Debugging flag for PC file shares. */
98 extern int	share_debug;
99 
100 /*
101  * Attributes caching:
102  *
103  * Attributes are cached in the rnode in struct vattr form.
104  * There is a time associated with the cached attributes (r_attrtime)
105  * which tells whether the attributes are valid. The time is initialized
106  * to the difference between current time and the modify time of the vnode
107  * when new attributes are cached. This allows the attributes for
108  * files that have changed recently to be timed out sooner than for files
109  * that have not changed for a long time. There are minimum and maximum
110  * timeout values that can be set per mount point.
111  */
112 
113 int
nfs_waitfor_purge_complete(vnode_t * vp)114 nfs_waitfor_purge_complete(vnode_t *vp)
115 {
116 	rnode_t *rp;
117 	k_sigset_t smask;
118 
119 	rp = VTOR(vp);
120 	if (rp->r_serial != NULL && rp->r_serial != curthread) {
121 		mutex_enter(&rp->r_statelock);
122 		sigintr(&smask, VTOMI(vp)->mi_flags & MI_INT);
123 		while (rp->r_serial != NULL) {
124 			if (!cv_wait_sig(&rp->r_cv, &rp->r_statelock)) {
125 				sigunintr(&smask);
126 				mutex_exit(&rp->r_statelock);
127 				return (EINTR);
128 			}
129 		}
130 		sigunintr(&smask);
131 		mutex_exit(&rp->r_statelock);
132 	}
133 	return (0);
134 }
135 
136 /*
137  * Validate caches by checking cached attributes. If the cached
138  * attributes have timed out, then get new attributes from the server.
139  * As a side affect, this will do cache invalidation if the attributes
140  * have changed.
141  *
142  * If the attributes have not timed out and if there is a cache
143  * invalidation being done by some other thread, then wait until that
144  * thread has completed the cache invalidation.
145  */
146 int
nfs_validate_caches(vnode_t * vp,cred_t * cr)147 nfs_validate_caches(vnode_t *vp, cred_t *cr)
148 {
149 	int error;
150 	struct vattr va;
151 
152 	if (ATTRCACHE_VALID(vp)) {
153 		error = nfs_waitfor_purge_complete(vp);
154 		if (error)
155 			return (error);
156 		return (0);
157 	}
158 
159 	va.va_mask = AT_ALL;
160 	return (nfs_getattr_otw(vp, &va, cr));
161 }
162 
163 /*
164  * Validate caches by checking cached attributes. If the cached
165  * attributes have timed out, then get new attributes from the server.
166  * As a side affect, this will do cache invalidation if the attributes
167  * have changed.
168  *
169  * If the attributes have not timed out and if there is a cache
170  * invalidation being done by some other thread, then wait until that
171  * thread has completed the cache invalidation.
172  */
173 int
nfs3_validate_caches(vnode_t * vp,cred_t * cr)174 nfs3_validate_caches(vnode_t *vp, cred_t *cr)
175 {
176 	int error;
177 	struct vattr va;
178 
179 	if (ATTRCACHE_VALID(vp)) {
180 		error = nfs_waitfor_purge_complete(vp);
181 		if (error)
182 			return (error);
183 		return (0);
184 	}
185 
186 	va.va_mask = AT_ALL;
187 	return (nfs3_getattr_otw(vp, &va, cr));
188 }
189 
190 /*
191  * Purge all of the various NFS `data' caches.
192  */
193 void
nfs_purge_caches(vnode_t * vp,int purge_dnlc,cred_t * cr)194 nfs_purge_caches(vnode_t *vp, int purge_dnlc, cred_t *cr)
195 {
196 	rnode_t *rp;
197 	char *contents;
198 	int size;
199 	int error;
200 
201 	/*
202 	 * Purge the DNLC for any entries which refer to this file.
203 	 * Avoid recursive entry into dnlc_purge_vp() in case of a directory.
204 	 */
205 	rp = VTOR(vp);
206 	mutex_enter(&rp->r_statelock);
207 	if (vp->v_count > 1 &&
208 	    (vp->v_type == VDIR || purge_dnlc == NFS_PURGE_DNLC) &&
209 	    !(rp->r_flags & RINDNLCPURGE)) {
210 		/*
211 		 * Set the RINDNLCPURGE flag to prevent recursive entry
212 		 * into dnlc_purge_vp()
213 		 */
214 		if (vp->v_type == VDIR)
215 			rp->r_flags |= RINDNLCPURGE;
216 		mutex_exit(&rp->r_statelock);
217 		dnlc_purge_vp(vp);
218 		mutex_enter(&rp->r_statelock);
219 		if (rp->r_flags & RINDNLCPURGE)
220 			rp->r_flags &= ~RINDNLCPURGE;
221 	}
222 
223 	/*
224 	 * Clear any readdir state bits and purge the readlink response cache.
225 	 */
226 	contents = rp->r_symlink.contents;
227 	size = rp->r_symlink.size;
228 	rp->r_symlink.contents = NULL;
229 	mutex_exit(&rp->r_statelock);
230 
231 	if (contents != NULL) {
232 
233 		kmem_free((void *)contents, size);
234 	}
235 
236 	/*
237 	 * Flush the page cache.
238 	 */
239 	if (vn_has_cached_data(vp)) {
240 		error = VOP_PUTPAGE(vp, (u_offset_t)0, 0, B_INVAL, cr, NULL);
241 		if (error && (error == ENOSPC || error == EDQUOT)) {
242 			mutex_enter(&rp->r_statelock);
243 			if (!rp->r_error)
244 				rp->r_error = error;
245 			mutex_exit(&rp->r_statelock);
246 		}
247 	}
248 
249 	/*
250 	 * Flush the readdir response cache.
251 	 */
252 	if (HAVE_RDDIR_CACHE(rp))
253 		nfs_purge_rddir_cache(vp);
254 }
255 
256 /*
257  * Purge the readdir cache of all entries
258  */
259 void
nfs_purge_rddir_cache(vnode_t * vp)260 nfs_purge_rddir_cache(vnode_t *vp)
261 {
262 	rnode_t *rp;
263 	rddir_cache *rdc;
264 	rddir_cache *nrdc;
265 
266 	rp = VTOR(vp);
267 top:
268 	mutex_enter(&rp->r_statelock);
269 	rp->r_direof = NULL;
270 	rp->r_flags &= ~RLOOKUP;
271 	rp->r_flags |= RREADDIRPLUS;
272 	rdc = avl_first(&rp->r_dir);
273 	while (rdc != NULL) {
274 		nrdc = AVL_NEXT(&rp->r_dir, rdc);
275 		avl_remove(&rp->r_dir, rdc);
276 		rddir_cache_rele(rdc);
277 		rdc = nrdc;
278 	}
279 	mutex_exit(&rp->r_statelock);
280 }
281 
282 /*
283  * Do a cache check based on the post-operation attributes.
284  * Then make them the new cached attributes.  If no attributes
285  * were returned, then mark the attributes as timed out.
286  */
287 void
nfs3_cache_post_op_attr(vnode_t * vp,post_op_attr * poap,hrtime_t t,cred_t * cr)288 nfs3_cache_post_op_attr(vnode_t *vp, post_op_attr *poap, hrtime_t t, cred_t *cr)
289 {
290 	vattr_t attr;
291 
292 	if (!poap->attributes) {
293 		PURGE_ATTRCACHE(vp);
294 		return;
295 	}
296 	(void) nfs3_cache_fattr3(vp, &poap->attr, &attr, t, cr);
297 }
298 
299 /*
300  * Same as above, but using a vattr
301  */
302 void
nfs3_cache_post_op_vattr(vnode_t * vp,post_op_vattr * poap,hrtime_t t,cred_t * cr)303 nfs3_cache_post_op_vattr(vnode_t *vp, post_op_vattr *poap, hrtime_t t,
304     cred_t *cr)
305 {
306 	if (!poap->attributes) {
307 		PURGE_ATTRCACHE(vp);
308 		return;
309 	}
310 	nfs_attr_cache(vp, poap->fres.vap, t, cr);
311 }
312 
313 /*
314  * Do a cache check based on the weak cache consistency attributes.
315  * These consist of a small set of pre-operation attributes and the
316  * full set of post-operation attributes.
317  *
318  * If we are given the pre-operation attributes, then use them to
319  * check the validity of the various caches.  Then, if we got the
320  * post-operation attributes, make them the new cached attributes.
321  * If we didn't get the post-operation attributes, then mark the
322  * attribute cache as timed out so that the next reference will
323  * cause a GETATTR to the server to refresh with the current
324  * attributes.
325  *
326  * Otherwise, if we didn't get the pre-operation attributes, but
327  * we did get the post-operation attributes, then use these
328  * attributes to check the validity of the various caches.  This
329  * will probably cause a flush of the caches because if the
330  * operation succeeded, the attributes of the object were changed
331  * in some way from the old post-operation attributes.  This
332  * should be okay because it is the safe thing to do.  After
333  * checking the data caches, then we make these the new cached
334  * attributes.
335  *
336  * Otherwise, we didn't get either the pre- or post-operation
337  * attributes.  Simply mark the attribute cache as timed out so
338  * the next reference will cause a GETATTR to the server to
339  * refresh with the current attributes.
340  *
341  * If an error occurred trying to convert the over the wire
342  * attributes to a vattr, then simply mark the attribute cache as
343  * timed out.
344  */
345 void
nfs3_cache_wcc_data(vnode_t * vp,wcc_data * wccp,hrtime_t t,cred_t * cr)346 nfs3_cache_wcc_data(vnode_t *vp, wcc_data *wccp, hrtime_t t, cred_t *cr)
347 {
348 	vattr_t bva;
349 	vattr_t ava;
350 
351 	if (wccp->after.attributes) {
352 		if (fattr3_to_vattr(vp, &wccp->after.attr, &ava)) {
353 			PURGE_ATTRCACHE(vp);
354 			return;
355 		}
356 		if (wccp->before.attributes) {
357 			bva.va_ctime.tv_sec = wccp->before.attr.ctime.seconds;
358 			bva.va_ctime.tv_nsec = wccp->before.attr.ctime.nseconds;
359 			bva.va_mtime.tv_sec = wccp->before.attr.mtime.seconds;
360 			bva.va_mtime.tv_nsec = wccp->before.attr.mtime.nseconds;
361 			bva.va_size = wccp->before.attr.size;
362 			nfs3_attr_cache(vp, &bva, &ava, t, cr);
363 		} else
364 			nfs_attr_cache(vp, &ava, t, cr);
365 	} else {
366 		PURGE_ATTRCACHE(vp);
367 	}
368 }
369 
370 /*
371  * Set attributes cache for given vnode using nfsattr.
372  *
373  * This routine does not do cache validation with the attributes.
374  *
375  * If an error occurred trying to convert the over the wire
376  * attributes to a vattr, then simply mark the attribute cache as
377  * timed out.
378  */
379 void
nfs_attrcache(vnode_t * vp,struct nfsfattr * na,hrtime_t t)380 nfs_attrcache(vnode_t *vp, struct nfsfattr *na, hrtime_t t)
381 {
382 	rnode_t *rp;
383 	struct vattr va;
384 
385 	if (!nattr_to_vattr(vp, na, &va)) {
386 		rp = VTOR(vp);
387 		mutex_enter(&rp->r_statelock);
388 		if (rp->r_mtime <= t)
389 			nfs_attrcache_va(vp, &va);
390 		mutex_exit(&rp->r_statelock);
391 	} else {
392 		PURGE_ATTRCACHE(vp);
393 	}
394 }
395 
396 /*
397  * Set attributes cache for given vnode using fattr3.
398  *
399  * This routine does not do cache validation with the attributes.
400  *
401  * If an error occurred trying to convert the over the wire
402  * attributes to a vattr, then simply mark the attribute cache as
403  * timed out.
404  */
405 void
nfs3_attrcache(vnode_t * vp,fattr3 * na,hrtime_t t)406 nfs3_attrcache(vnode_t *vp, fattr3 *na, hrtime_t t)
407 {
408 	rnode_t *rp;
409 	struct vattr va;
410 
411 	if (!fattr3_to_vattr(vp, na, &va)) {
412 		rp = VTOR(vp);
413 		mutex_enter(&rp->r_statelock);
414 		if (rp->r_mtime <= t)
415 			nfs_attrcache_va(vp, &va);
416 		mutex_exit(&rp->r_statelock);
417 	} else {
418 		PURGE_ATTRCACHE(vp);
419 	}
420 }
421 
422 /*
423  * Do a cache check based on attributes returned over the wire.  The
424  * new attributes are cached.
425  *
426  * If an error occurred trying to convert the over the wire attributes
427  * to a vattr, then just return that error.
428  *
429  * As a side affect, the vattr argument is filled in with the converted
430  * attributes.
431  */
432 int
nfs_cache_fattr(vnode_t * vp,struct nfsfattr * na,vattr_t * vap,hrtime_t t,cred_t * cr)433 nfs_cache_fattr(vnode_t *vp, struct nfsfattr *na, vattr_t *vap, hrtime_t t,
434     cred_t *cr)
435 {
436 	int error;
437 
438 	error = nattr_to_vattr(vp, na, vap);
439 	if (error)
440 		return (error);
441 	nfs_attr_cache(vp, vap, t, cr);
442 	return (0);
443 }
444 
445 /*
446  * Do a cache check based on attributes returned over the wire.  The
447  * new attributes are cached.
448  *
449  * If an error occurred trying to convert the over the wire attributes
450  * to a vattr, then just return that error.
451  *
452  * As a side affect, the vattr argument is filled in with the converted
453  * attributes.
454  */
455 int
nfs3_cache_fattr3(vnode_t * vp,fattr3 * na,vattr_t * vap,hrtime_t t,cred_t * cr)456 nfs3_cache_fattr3(vnode_t *vp, fattr3 *na, vattr_t *vap, hrtime_t t, cred_t *cr)
457 {
458 	int error;
459 
460 	error = fattr3_to_vattr(vp, na, vap);
461 	if (error)
462 		return (error);
463 	nfs_attr_cache(vp, vap, t, cr);
464 	return (0);
465 }
466 
467 /*
468  * Use the passed in virtual attributes to check to see whether the
469  * data and metadata caches are valid, cache the new attributes, and
470  * then do the cache invalidation if required.
471  *
472  * The cache validation and caching of the new attributes is done
473  * atomically via the use of the mutex, r_statelock.  If required,
474  * the cache invalidation is done atomically w.r.t. the cache
475  * validation and caching of the attributes via the pseudo lock,
476  * r_serial.
477  *
478  * This routine is used to do cache validation and attributes caching
479  * for operations with a single set of post operation attributes.
480  */
481 void
nfs_attr_cache(vnode_t * vp,vattr_t * vap,hrtime_t t,cred_t * cr)482 nfs_attr_cache(vnode_t *vp, vattr_t *vap, hrtime_t t, cred_t *cr)
483 {
484 	rnode_t *rp;
485 	int mtime_changed = 0;
486 	int ctime_changed = 0;
487 	vsecattr_t *vsp;
488 	int was_serial;
489 	len_t preattr_rsize;
490 	boolean_t writeattr_set = B_FALSE;
491 	boolean_t cachepurge_set = B_FALSE;
492 
493 	rp = VTOR(vp);
494 
495 	mutex_enter(&rp->r_statelock);
496 
497 	if (rp->r_serial != curthread) {
498 		klwp_t *lwp = ttolwp(curthread);
499 
500 		was_serial = 0;
501 		if (lwp != NULL)
502 			lwp->lwp_nostop++;
503 		while (rp->r_serial != NULL) {
504 			if (!cv_wait_sig(&rp->r_cv, &rp->r_statelock)) {
505 				mutex_exit(&rp->r_statelock);
506 				if (lwp != NULL)
507 					lwp->lwp_nostop--;
508 				return;
509 			}
510 		}
511 		if (lwp != NULL)
512 			lwp->lwp_nostop--;
513 	} else
514 		was_serial = 1;
515 
516 	if (rp->r_mtime > t) {
517 		if (!CACHE_VALID(rp, vap->va_mtime, vap->va_size))
518 			PURGE_ATTRCACHE_LOCKED(rp);
519 		mutex_exit(&rp->r_statelock);
520 		return;
521 	}
522 
523 	/*
524 	 * Write thread after writing data to file on remote server,
525 	 * will always set RWRITEATTR to indicate that file on remote
526 	 * server was modified with a WRITE operation and would have
527 	 * marked attribute cache as timed out. If RWRITEATTR
528 	 * is set, then do not check for mtime and ctime change.
529 	 */
530 	if (!(rp->r_flags & RWRITEATTR)) {
531 		if (!CACHE_VALID(rp, vap->va_mtime, vap->va_size))
532 			mtime_changed = 1;
533 
534 		if (rp->r_attr.va_ctime.tv_sec != vap->va_ctime.tv_sec ||
535 		    rp->r_attr.va_ctime.tv_nsec != vap->va_ctime.tv_nsec)
536 			ctime_changed = 1;
537 	} else {
538 		writeattr_set = B_TRUE;
539 	}
540 
541 	preattr_rsize = rp->r_size;
542 
543 	nfs_attrcache_va(vp, vap);
544 
545 	/*
546 	 * If we have updated filesize in nfs_attrcache_va, as soon as we
547 	 * drop statelock we will be in transition of purging all
548 	 * our caches and updating them. It is possible for another
549 	 * thread to pick this new file size and read in zeroed data.
550 	 * stall other threads till cache purge is complete.
551 	 */
552 	if ((vp->v_type == VREG) && (rp->r_size != preattr_rsize)) {
553 		/*
554 		 * If RWRITEATTR was set and we have updated the file
555 		 * size, Server's returned file size need not necessarily
556 		 * be because of this Client's WRITE. We need to purge
557 		 * all caches.
558 		 */
559 		if (writeattr_set)
560 			mtime_changed = 1;
561 
562 		if (mtime_changed && !(rp->r_flags & RINCACHEPURGE)) {
563 			rp->r_flags |= RINCACHEPURGE;
564 			cachepurge_set = B_TRUE;
565 		}
566 	}
567 
568 	if (!mtime_changed && !ctime_changed) {
569 		mutex_exit(&rp->r_statelock);
570 		return;
571 	}
572 
573 	rp->r_serial = curthread;
574 
575 	mutex_exit(&rp->r_statelock);
576 
577 	if (mtime_changed)
578 		nfs_purge_caches(vp, NFS_NOPURGE_DNLC, cr);
579 
580 	if ((rp->r_flags & RINCACHEPURGE) && cachepurge_set) {
581 		mutex_enter(&rp->r_statelock);
582 		rp->r_flags &= ~RINCACHEPURGE;
583 		cv_broadcast(&rp->r_cv);
584 		mutex_exit(&rp->r_statelock);
585 		cachepurge_set = B_FALSE;
586 	}
587 
588 	if (ctime_changed) {
589 		(void) nfs_access_purge_rp(rp);
590 		if (rp->r_secattr != NULL) {
591 			mutex_enter(&rp->r_statelock);
592 			vsp = rp->r_secattr;
593 			rp->r_secattr = NULL;
594 			mutex_exit(&rp->r_statelock);
595 			if (vsp != NULL)
596 				nfs_acl_free(vsp);
597 		}
598 	}
599 
600 	if (!was_serial) {
601 		mutex_enter(&rp->r_statelock);
602 		rp->r_serial = NULL;
603 		cv_broadcast(&rp->r_cv);
604 		mutex_exit(&rp->r_statelock);
605 	}
606 }
607 
608 /*
609  * Use the passed in "before" virtual attributes to check to see
610  * whether the data and metadata caches are valid, cache the "after"
611  * new attributes, and then do the cache invalidation if required.
612  *
613  * The cache validation and caching of the new attributes is done
614  * atomically via the use of the mutex, r_statelock.  If required,
615  * the cache invalidation is done atomically w.r.t. the cache
616  * validation and caching of the attributes via the pseudo lock,
617  * r_serial.
618  *
619  * This routine is used to do cache validation and attributes caching
620  * for operations with both pre operation attributes and post operation
621  * attributes.
622  */
623 static void
nfs3_attr_cache(vnode_t * vp,vattr_t * bvap,vattr_t * avap,hrtime_t t,cred_t * cr)624 nfs3_attr_cache(vnode_t *vp, vattr_t *bvap, vattr_t *avap, hrtime_t t,
625     cred_t *cr)
626 {
627 	rnode_t *rp;
628 	int mtime_changed = 0;
629 	int ctime_changed = 0;
630 	vsecattr_t *vsp;
631 	int was_serial;
632 	len_t preattr_rsize;
633 	boolean_t writeattr_set = B_FALSE;
634 	boolean_t cachepurge_set = B_FALSE;
635 
636 	rp = VTOR(vp);
637 
638 	mutex_enter(&rp->r_statelock);
639 
640 	if (rp->r_serial != curthread) {
641 		klwp_t *lwp = ttolwp(curthread);
642 
643 		was_serial = 0;
644 		if (lwp != NULL)
645 			lwp->lwp_nostop++;
646 		while (rp->r_serial != NULL) {
647 			if (!cv_wait_sig(&rp->r_cv, &rp->r_statelock)) {
648 				mutex_exit(&rp->r_statelock);
649 				if (lwp != NULL)
650 					lwp->lwp_nostop--;
651 				return;
652 			}
653 		}
654 		if (lwp != NULL)
655 			lwp->lwp_nostop--;
656 	} else
657 		was_serial = 1;
658 
659 	if (rp->r_mtime > t) {
660 		if (!CACHE_VALID(rp, avap->va_mtime, avap->va_size))
661 			PURGE_ATTRCACHE_LOCKED(rp);
662 		mutex_exit(&rp->r_statelock);
663 		return;
664 	}
665 
666 	/*
667 	 * Write thread after writing data to file on remote server,
668 	 * will always set RWRITEATTR to indicate that file on remote
669 	 * server was modified with a WRITE operation and would have
670 	 * marked attribute cache as timed out. If RWRITEATTR
671 	 * is set, then do not check for mtime and ctime change.
672 	 */
673 	if (!(rp->r_flags & RWRITEATTR)) {
674 		if (!CACHE_VALID(rp, bvap->va_mtime, bvap->va_size))
675 			mtime_changed = 1;
676 
677 		if (rp->r_attr.va_ctime.tv_sec != bvap->va_ctime.tv_sec ||
678 		    rp->r_attr.va_ctime.tv_nsec != bvap->va_ctime.tv_nsec)
679 			ctime_changed = 1;
680 	} else {
681 		writeattr_set = B_TRUE;
682 	}
683 
684 	preattr_rsize = rp->r_size;
685 
686 	nfs_attrcache_va(vp, avap);
687 
688 	/*
689 	 * If we have updated filesize in nfs_attrcache_va, as soon as we
690 	 * drop statelock we will be in transition of purging all
691 	 * our caches and updating them. It is possible for another
692 	 * thread to pick this new file size and read in zeroed data.
693 	 * stall other threads till cache purge is complete.
694 	 */
695 	if ((vp->v_type == VREG) && (rp->r_size != preattr_rsize)) {
696 		/*
697 		 * If RWRITEATTR was set and we have updated the file
698 		 * size, Server's returned file size need not necessarily
699 		 * be because of this Client's WRITE. We need to purge
700 		 * all caches.
701 		 */
702 		if (writeattr_set)
703 			mtime_changed = 1;
704 
705 		if (mtime_changed && !(rp->r_flags & RINCACHEPURGE)) {
706 			rp->r_flags |= RINCACHEPURGE;
707 			cachepurge_set = B_TRUE;
708 		}
709 	}
710 
711 	if (!mtime_changed && !ctime_changed) {
712 		mutex_exit(&rp->r_statelock);
713 		return;
714 	}
715 
716 	rp->r_serial = curthread;
717 
718 	mutex_exit(&rp->r_statelock);
719 
720 	if (mtime_changed)
721 		nfs_purge_caches(vp, NFS_NOPURGE_DNLC, cr);
722 
723 	if ((rp->r_flags & RINCACHEPURGE) && cachepurge_set) {
724 		mutex_enter(&rp->r_statelock);
725 		rp->r_flags &= ~RINCACHEPURGE;
726 		cv_broadcast(&rp->r_cv);
727 		mutex_exit(&rp->r_statelock);
728 		cachepurge_set = B_FALSE;
729 	}
730 
731 	if (ctime_changed) {
732 		(void) nfs_access_purge_rp(rp);
733 		if (rp->r_secattr != NULL) {
734 			mutex_enter(&rp->r_statelock);
735 			vsp = rp->r_secattr;
736 			rp->r_secattr = NULL;
737 			mutex_exit(&rp->r_statelock);
738 			if (vsp != NULL)
739 				nfs_acl_free(vsp);
740 		}
741 	}
742 
743 	if (!was_serial) {
744 		mutex_enter(&rp->r_statelock);
745 		rp->r_serial = NULL;
746 		cv_broadcast(&rp->r_cv);
747 		mutex_exit(&rp->r_statelock);
748 	}
749 }
750 
751 /*
752  * Set attributes cache for given vnode using virtual attributes.
753  *
754  * Set the timeout value on the attribute cache and fill it
755  * with the passed in attributes.
756  *
757  * The caller must be holding r_statelock.
758  */
759 void
nfs_attrcache_va(vnode_t * vp,struct vattr * va)760 nfs_attrcache_va(vnode_t *vp, struct vattr *va)
761 {
762 	rnode_t *rp;
763 	mntinfo_t *mi;
764 	hrtime_t delta;
765 	hrtime_t now;
766 
767 	rp = VTOR(vp);
768 
769 	ASSERT(MUTEX_HELD(&rp->r_statelock));
770 
771 	now = gethrtime();
772 
773 	mi = VTOMI(vp);
774 
775 	/*
776 	 * Delta is the number of nanoseconds that we will
777 	 * cache the attributes of the file.  It is based on
778 	 * the number of nanoseconds since the last time that
779 	 * we detected a change.  The assumption is that files
780 	 * that changed recently are likely to change again.
781 	 * There is a minimum and a maximum for regular files
782 	 * and for directories which is enforced though.
783 	 *
784 	 * Using the time since last change was detected
785 	 * eliminates direct comparison or calculation
786 	 * using mixed client and server times.  NFS does
787 	 * not make any assumptions regarding the client
788 	 * and server clocks being synchronized.
789 	 */
790 	if (va->va_mtime.tv_sec != rp->r_attr.va_mtime.tv_sec ||
791 	    va->va_mtime.tv_nsec != rp->r_attr.va_mtime.tv_nsec ||
792 	    va->va_size != rp->r_attr.va_size)
793 		rp->r_mtime = now;
794 
795 	if ((mi->mi_flags & MI_NOAC) || (vp->v_flag & VNOCACHE))
796 		delta = 0;
797 	else {
798 		delta = now - rp->r_mtime;
799 		if (vp->v_type == VDIR) {
800 			if (delta < mi->mi_acdirmin)
801 				delta = mi->mi_acdirmin;
802 			else if (delta > mi->mi_acdirmax)
803 				delta = mi->mi_acdirmax;
804 		} else {
805 			if (delta < mi->mi_acregmin)
806 				delta = mi->mi_acregmin;
807 			else if (delta > mi->mi_acregmax)
808 				delta = mi->mi_acregmax;
809 		}
810 	}
811 	rp->r_attrtime = now + delta;
812 	rp->r_attr = *va;
813 	/*
814 	 * Update the size of the file if there is no cached data or if
815 	 * the cached data is clean and there is no data being written
816 	 * out.
817 	 */
818 	if (rp->r_size != va->va_size &&
819 	    (!vn_has_cached_data(vp) ||
820 	    (!(rp->r_flags & RDIRTY) && rp->r_count == 0)))
821 		rp->r_size = va->va_size;
822 	nfs_setswaplike(vp, va);
823 	rp->r_flags &= ~RWRITEATTR;
824 }
825 
826 /*
827  * Fill in attribute from the cache.
828  * If valid, then return 0 to indicate that no error occurred,
829  * otherwise return 1 to indicate that an error occurred.
830  */
831 static int
nfs_getattr_cache(vnode_t * vp,struct vattr * vap)832 nfs_getattr_cache(vnode_t *vp, struct vattr *vap)
833 {
834 	rnode_t *rp;
835 	uint_t mask = vap->va_mask;
836 
837 	rp = VTOR(vp);
838 	mutex_enter(&rp->r_statelock);
839 	if (ATTRCACHE_VALID(vp)) {
840 		/*
841 		 * Cached attributes are valid
842 		 */
843 		*vap = rp->r_attr;
844 		/*
845 		 * Set the caller's va_mask to the set of attributes
846 		 * that were requested ANDed with the attributes that
847 		 * are available.  If attributes were requested that
848 		 * are not available, those bits must be turned off
849 		 * in the callers va_mask.
850 		 */
851 		vap->va_mask &= mask;
852 		mutex_exit(&rp->r_statelock);
853 		return (0);
854 	}
855 	mutex_exit(&rp->r_statelock);
856 	return (1);
857 }
858 
859 /*
860  * Get attributes over-the-wire and update attributes cache
861  * if no error occurred in the over-the-wire operation.
862  * Return 0 if successful, otherwise error.
863  */
864 int
nfs_getattr_otw(vnode_t * vp,struct vattr * vap,cred_t * cr)865 nfs_getattr_otw(vnode_t *vp, struct vattr *vap, cred_t *cr)
866 {
867 	int error;
868 	struct nfsattrstat ns;
869 	int douprintf;
870 	mntinfo_t *mi;
871 	failinfo_t fi;
872 	hrtime_t t;
873 
874 	mi = VTOMI(vp);
875 	fi.vp = vp;
876 	fi.fhp = NULL;		/* no need to update, filehandle not copied */
877 	fi.copyproc = nfscopyfh;
878 	fi.lookupproc = nfslookup;
879 	fi.xattrdirproc = acl_getxattrdir2;
880 
881 	if (mi->mi_flags & MI_ACL) {
882 		error = acl_getattr2_otw(vp, vap, cr);
883 		if (mi->mi_flags & MI_ACL)
884 			return (error);
885 	}
886 
887 	douprintf = 1;
888 
889 	t = gethrtime();
890 
891 	error = rfs2call(mi, RFS_GETATTR,
892 	    xdr_fhandle, (caddr_t)VTOFH(vp),
893 	    xdr_attrstat, (caddr_t)&ns, cr,
894 	    &douprintf, &ns.ns_status, 0, &fi);
895 
896 	if (!error) {
897 		error = geterrno(ns.ns_status);
898 		if (!error)
899 			error = nfs_cache_fattr(vp, &ns.ns_attr, vap, t, cr);
900 		else {
901 			PURGE_STALE_FH(error, vp, cr);
902 		}
903 	}
904 
905 	return (error);
906 }
907 
908 /*
909  * Return either cached ot remote attributes. If get remote attr
910  * use them to check and invalidate caches, then cache the new attributes.
911  */
912 int
nfsgetattr(vnode_t * vp,struct vattr * vap,cred_t * cr)913 nfsgetattr(vnode_t *vp, struct vattr *vap, cred_t *cr)
914 {
915 	int error;
916 	rnode_t *rp;
917 
918 	/*
919 	 * If we've got cached attributes, we're done, otherwise go
920 	 * to the server to get attributes, which will update the cache
921 	 * in the process.
922 	 */
923 	error = nfs_getattr_cache(vp, vap);
924 	if (error)
925 		error = nfs_getattr_otw(vp, vap, cr);
926 
927 	/* Return the client's view of file size */
928 	rp = VTOR(vp);
929 	mutex_enter(&rp->r_statelock);
930 	vap->va_size = rp->r_size;
931 	mutex_exit(&rp->r_statelock);
932 
933 	return (error);
934 }
935 
936 /*
937  * Get attributes over-the-wire and update attributes cache
938  * if no error occurred in the over-the-wire operation.
939  * Return 0 if successful, otherwise error.
940  */
941 int
nfs3_getattr_otw(vnode_t * vp,struct vattr * vap,cred_t * cr)942 nfs3_getattr_otw(vnode_t *vp, struct vattr *vap, cred_t *cr)
943 {
944 	int error;
945 	GETATTR3args args;
946 	GETATTR3vres res;
947 	int douprintf;
948 	failinfo_t fi;
949 	hrtime_t t;
950 
951 	args.object = *VTOFH3(vp);
952 	fi.vp = vp;
953 	fi.fhp = (caddr_t)&args.object;
954 	fi.copyproc = nfs3copyfh;
955 	fi.lookupproc = nfs3lookup;
956 	fi.xattrdirproc = acl_getxattrdir3;
957 	res.fres.vp = vp;
958 	res.fres.vap = vap;
959 
960 	douprintf = 1;
961 
962 	t = gethrtime();
963 
964 	error = rfs3call(VTOMI(vp), NFSPROC3_GETATTR,
965 	    xdr_nfs_fh3, (caddr_t)&args,
966 	    xdr_GETATTR3vres, (caddr_t)&res, cr,
967 	    &douprintf, &res.status, 0, &fi);
968 
969 	if (error)
970 		return (error);
971 
972 	error = geterrno3(res.status);
973 	if (error) {
974 		PURGE_STALE_FH(error, vp, cr);
975 		return (error);
976 	}
977 
978 	/*
979 	 * Catch status codes that indicate fattr3 to vattr translation failure
980 	 */
981 	if (res.fres.status)
982 		return (res.fres.status);
983 
984 	nfs_attr_cache(vp, vap, t, cr);
985 	return (0);
986 }
987 
988 /*
989  * Return either cached or remote attributes. If get remote attr
990  * use them to check and invalidate caches, then cache the new attributes.
991  */
992 int
nfs3getattr(vnode_t * vp,struct vattr * vap,cred_t * cr)993 nfs3getattr(vnode_t *vp, struct vattr *vap, cred_t *cr)
994 {
995 	int error;
996 	rnode_t *rp;
997 
998 	/*
999 	 * If we've got cached attributes, we're done, otherwise go
1000 	 * to the server to get attributes, which will update the cache
1001 	 * in the process.
1002 	 */
1003 	error = nfs_getattr_cache(vp, vap);
1004 	if (error)
1005 		error = nfs3_getattr_otw(vp, vap, cr);
1006 
1007 	/* Return the client's view of file size */
1008 	rp = VTOR(vp);
1009 	mutex_enter(&rp->r_statelock);
1010 	vap->va_size = rp->r_size;
1011 	mutex_exit(&rp->r_statelock);
1012 
1013 	return (error);
1014 }
1015 
1016 vtype_t nf_to_vt[] = {
1017 	VNON, VREG, VDIR, VBLK, VCHR, VLNK, VSOCK
1018 };
1019 /*
1020  * Convert NFS Version 2 over the network attributes to the local
1021  * virtual attributes.  The mapping between the UID_NOBODY/GID_NOBODY
1022  * network representation and the local representation is done here.
1023  * Returns 0 for success, error if failed due to overflow.
1024  */
1025 int
nattr_to_vattr(vnode_t * vp,struct nfsfattr * na,struct vattr * vap)1026 nattr_to_vattr(vnode_t *vp, struct nfsfattr *na, struct vattr *vap)
1027 {
1028 	/* overflow in time attributes? */
1029 #ifndef _LP64
1030 	if (!NFS2_FATTR_TIME_OK(na))
1031 		return (EOVERFLOW);
1032 #endif
1033 
1034 	vap->va_mask = AT_ALL;
1035 
1036 	if (na->na_type < NFNON || na->na_type > NFSOC)
1037 		vap->va_type = VBAD;
1038 	else
1039 		vap->va_type = nf_to_vt[na->na_type];
1040 	vap->va_mode = na->na_mode;
1041 	vap->va_uid = (na->na_uid == NFS_UID_NOBODY) ? UID_NOBODY : na->na_uid;
1042 	vap->va_gid = (na->na_gid == NFS_GID_NOBODY) ? GID_NOBODY : na->na_gid;
1043 	vap->va_fsid = vp->v_vfsp->vfs_dev;
1044 	vap->va_nodeid = na->na_nodeid;
1045 	vap->va_nlink = na->na_nlink;
1046 	vap->va_size = na->na_size;	/* keep for cache validation */
1047 	/*
1048 	 * nfs protocol defines times as unsigned so don't extend sign,
1049 	 * unless sysadmin set nfs_allow_preepoch_time.
1050 	 */
1051 	NFS_TIME_T_CONVERT(vap->va_atime.tv_sec, na->na_atime.tv_sec);
1052 	vap->va_atime.tv_nsec = (uint32_t)(na->na_atime.tv_usec * 1000);
1053 	NFS_TIME_T_CONVERT(vap->va_mtime.tv_sec, na->na_mtime.tv_sec);
1054 	vap->va_mtime.tv_nsec = (uint32_t)(na->na_mtime.tv_usec * 1000);
1055 	NFS_TIME_T_CONVERT(vap->va_ctime.tv_sec, na->na_ctime.tv_sec);
1056 	vap->va_ctime.tv_nsec = (uint32_t)(na->na_ctime.tv_usec * 1000);
1057 	/*
1058 	 * Shannon's law - uncompress the received dev_t
1059 	 * if the top half of is zero indicating a response
1060 	 * from an `older style' OS. Except for when it is a
1061 	 * `new style' OS sending the maj device of zero,
1062 	 * in which case the algorithm still works because the
1063 	 * fact that it is a new style server
1064 	 * is hidden by the minor device not being greater
1065 	 * than 255 (a requirement in this case).
1066 	 */
1067 	if ((na->na_rdev & 0xffff0000) == 0)
1068 		vap->va_rdev = nfsv2_expdev(na->na_rdev);
1069 	else
1070 		vap->va_rdev = expldev(na->na_rdev);
1071 
1072 	vap->va_nblocks = na->na_blocks;
1073 	switch (na->na_type) {
1074 	case NFBLK:
1075 		vap->va_blksize = DEV_BSIZE;
1076 		break;
1077 
1078 	case NFCHR:
1079 		vap->va_blksize = MAXBSIZE;
1080 		break;
1081 
1082 	case NFSOC:
1083 	default:
1084 		vap->va_blksize = na->na_blocksize;
1085 		break;
1086 	}
1087 	/*
1088 	 * This bit of ugliness is a hack to preserve the
1089 	 * over-the-wire protocols for named-pipe vnodes.
1090 	 * It remaps the special over-the-wire type to the
1091 	 * VFIFO type. (see note in nfs.h)
1092 	 */
1093 	if (NA_ISFIFO(na)) {
1094 		vap->va_type = VFIFO;
1095 		vap->va_mode = (vap->va_mode & ~S_IFMT) | S_IFIFO;
1096 		vap->va_rdev = 0;
1097 		vap->va_blksize = na->na_blocksize;
1098 	}
1099 	vap->va_seq = 0;
1100 	return (0);
1101 }
1102 
1103 /*
1104  * Convert NFS Version 3 over the network attributes to the local
1105  * virtual attributes.  The mapping between the UID_NOBODY/GID_NOBODY
1106  * network representation and the local representation is done here.
1107  */
1108 vtype_t nf3_to_vt[] = {
1109 	VBAD, VREG, VDIR, VBLK, VCHR, VLNK, VSOCK, VFIFO
1110 };
1111 
1112 int
fattr3_to_vattr(vnode_t * vp,fattr3 * na,struct vattr * vap)1113 fattr3_to_vattr(vnode_t *vp, fattr3 *na, struct vattr *vap)
1114 {
1115 
1116 #ifndef _LP64
1117 	/* overflow in time attributes? */
1118 	if (!NFS3_FATTR_TIME_OK(na))
1119 		return (EOVERFLOW);
1120 #endif
1121 	if (!NFS3_SIZE_OK(na->size))
1122 		/* file too big */
1123 		return (EFBIG);
1124 
1125 	vap->va_mask = AT_ALL;
1126 
1127 	if (na->type < NF3REG || na->type > NF3FIFO)
1128 		vap->va_type = VBAD;
1129 	else
1130 		vap->va_type = nf3_to_vt[na->type];
1131 	vap->va_mode = na->mode;
1132 	vap->va_uid = (na->uid == NFS_UID_NOBODY) ? UID_NOBODY : (uid_t)na->uid;
1133 	vap->va_gid = (na->gid == NFS_GID_NOBODY) ? GID_NOBODY : (gid_t)na->gid;
1134 	vap->va_fsid = vp->v_vfsp->vfs_dev;
1135 	vap->va_nodeid = na->fileid;
1136 	vap->va_nlink = na->nlink;
1137 	vap->va_size = na->size;
1138 
1139 	/*
1140 	 * nfs protocol defines times as unsigned so don't extend sign,
1141 	 * unless sysadmin set nfs_allow_preepoch_time.
1142 	 */
1143 	NFS_TIME_T_CONVERT(vap->va_atime.tv_sec, na->atime.seconds);
1144 	vap->va_atime.tv_nsec = (uint32_t)na->atime.nseconds;
1145 	NFS_TIME_T_CONVERT(vap->va_mtime.tv_sec, na->mtime.seconds);
1146 	vap->va_mtime.tv_nsec = (uint32_t)na->mtime.nseconds;
1147 	NFS_TIME_T_CONVERT(vap->va_ctime.tv_sec, na->ctime.seconds);
1148 	vap->va_ctime.tv_nsec = (uint32_t)na->ctime.nseconds;
1149 
1150 	switch (na->type) {
1151 	case NF3BLK:
1152 		vap->va_rdev = makedevice(na->rdev.specdata1,
1153 		    na->rdev.specdata2);
1154 		vap->va_blksize = DEV_BSIZE;
1155 		vap->va_nblocks = 0;
1156 		break;
1157 	case NF3CHR:
1158 		vap->va_rdev = makedevice(na->rdev.specdata1,
1159 		    na->rdev.specdata2);
1160 		vap->va_blksize = MAXBSIZE;
1161 		vap->va_nblocks = 0;
1162 		break;
1163 	case NF3REG:
1164 	case NF3DIR:
1165 	case NF3LNK:
1166 		vap->va_rdev = 0;
1167 		vap->va_blksize = MAXBSIZE;
1168 		vap->va_nblocks = (u_longlong_t)
1169 		    ((na->used + (size3)DEV_BSIZE - (size3)1) /
1170 		    (size3)DEV_BSIZE);
1171 		break;
1172 	case NF3SOCK:
1173 	case NF3FIFO:
1174 	default:
1175 		vap->va_rdev = 0;
1176 		vap->va_blksize = MAXBSIZE;
1177 		vap->va_nblocks = 0;
1178 		break;
1179 	}
1180 	vap->va_seq = 0;
1181 	return (0);
1182 }
1183 
1184 /*
1185  * Asynchronous I/O parameters.  nfs_async_threads is the high-water mark
1186  * for the demand-based allocation of async threads per-mount.  The
1187  * nfs_async_timeout is the amount of time a thread will live after it
1188  * becomes idle, unless new I/O requests are received before the thread
1189  * dies.  See nfs_async_putpage and nfs_async_start.
1190  */
1191 
1192 int nfs_async_timeout = -1;	/* uninitialized */
1193 
1194 static void	nfs_async_start(struct vfs *);
1195 static void	nfs_async_pgops_start(struct vfs *);
1196 static void	nfs_async_common_start(struct vfs *, int);
1197 
1198 static void
free_async_args(struct nfs_async_reqs * args)1199 free_async_args(struct nfs_async_reqs *args)
1200 {
1201 	rnode_t *rp;
1202 
1203 	if (args->a_io != NFS_INACTIVE) {
1204 		rp = VTOR(args->a_vp);
1205 		mutex_enter(&rp->r_statelock);
1206 		rp->r_count--;
1207 		if (args->a_io == NFS_PUTAPAGE ||
1208 		    args->a_io == NFS_PAGEIO)
1209 			rp->r_awcount--;
1210 		cv_broadcast(&rp->r_cv);
1211 		mutex_exit(&rp->r_statelock);
1212 		VN_RELE(args->a_vp);
1213 	}
1214 	crfree(args->a_cred);
1215 	kmem_free(args, sizeof (*args));
1216 }
1217 
1218 /*
1219  * Cross-zone thread creation and NFS access is disallowed, yet fsflush() and
1220  * pageout(), running in the global zone, have legitimate reasons to do
1221  * VOP_PUTPAGE(B_ASYNC) on other zones' NFS mounts.  We avoid the problem by
1222  * use of a a per-mount "asynchronous requests manager thread" which is
1223  * signaled by the various asynchronous work routines when there is
1224  * asynchronous work to be done.  It is responsible for creating new
1225  * worker threads if necessary, and notifying existing worker threads
1226  * that there is work to be done.
1227  *
1228  * In other words, it will "take the specifications from the customers and
1229  * give them to the engineers."
1230  *
1231  * Worker threads die off of their own accord if they are no longer
1232  * needed.
1233  *
1234  * This thread is killed when the zone is going away or the filesystem
1235  * is being unmounted.
1236  */
1237 void
nfs_async_manager(vfs_t * vfsp)1238 nfs_async_manager(vfs_t *vfsp)
1239 {
1240 	callb_cpr_t cprinfo;
1241 	mntinfo_t *mi;
1242 	uint_t max_threads;
1243 
1244 	mi = VFTOMI(vfsp);
1245 
1246 	CALLB_CPR_INIT(&cprinfo, &mi->mi_async_lock, callb_generic_cpr,
1247 	    "nfs_async_manager");
1248 
1249 	mutex_enter(&mi->mi_async_lock);
1250 	/*
1251 	 * We want to stash the max number of threads that this mount was
1252 	 * allowed so we can use it later when the variable is set to zero as
1253 	 * part of the zone/mount going away.
1254 	 *
1255 	 * We want to be able to create at least one thread to handle
1256 	 * asynchronous inactive calls.
1257 	 */
1258 	max_threads = MAX(mi->mi_max_threads, 1);
1259 	/*
1260 	 * We don't want to wait for mi_max_threads to go to zero, since that
1261 	 * happens as part of a failed unmount, but this thread should only
1262 	 * exit when the mount/zone is really going away.
1263 	 *
1264 	 * Once MI_ASYNC_MGR_STOP is set, no more async operations will be
1265 	 * attempted: the various _async_*() functions know to do things
1266 	 * inline if mi_max_threads == 0.  Henceforth we just drain out the
1267 	 * outstanding requests.
1268 	 *
1269 	 * Note that we still create zthreads even if we notice the zone is
1270 	 * shutting down (MI_ASYNC_MGR_STOP is set); this may cause the zone
1271 	 * shutdown sequence to take slightly longer in some cases, but
1272 	 * doesn't violate the protocol, as all threads will exit as soon as
1273 	 * they're done processing the remaining requests.
1274 	 */
1275 	for (;;) {
1276 		while (mi->mi_async_req_count > 0) {
1277 			/*
1278 			 * Paranoia: If the mount started out having
1279 			 * (mi->mi_max_threads == 0), and the value was
1280 			 * later changed (via a debugger or somesuch),
1281 			 * we could be confused since we will think we
1282 			 * can't create any threads, and the calling
1283 			 * code (which looks at the current value of
1284 			 * mi->mi_max_threads, now non-zero) thinks we
1285 			 * can.
1286 			 *
1287 			 * So, because we're paranoid, we create threads
1288 			 * up to the maximum of the original and the
1289 			 * current value. This means that future
1290 			 * (debugger-induced) lowerings of
1291 			 * mi->mi_max_threads are ignored for our
1292 			 * purposes, but who told them they could change
1293 			 * random values on a live kernel anyhow?
1294 			 */
1295 			if (mi->mi_threads[NFS_ASYNC_QUEUE] <
1296 			    MAX(mi->mi_max_threads, max_threads)) {
1297 				mi->mi_threads[NFS_ASYNC_QUEUE]++;
1298 				mutex_exit(&mi->mi_async_lock);
1299 				VFS_HOLD(vfsp);	/* hold for new thread */
1300 				(void) zthread_create(NULL, 0, nfs_async_start,
1301 				    vfsp, 0, minclsyspri);
1302 				mutex_enter(&mi->mi_async_lock);
1303 			} else if (mi->mi_threads[NFS_ASYNC_PGOPS_QUEUE] <
1304 			    NUM_ASYNC_PGOPS_THREADS) {
1305 				mi->mi_threads[NFS_ASYNC_PGOPS_QUEUE]++;
1306 				mutex_exit(&mi->mi_async_lock);
1307 				VFS_HOLD(vfsp); /* hold for new thread */
1308 				(void) zthread_create(NULL, 0,
1309 				    nfs_async_pgops_start, vfsp, 0,
1310 				    minclsyspri);
1311 				mutex_enter(&mi->mi_async_lock);
1312 			}
1313 			NFS_WAKE_ASYNC_WORKER(mi->mi_async_work_cv);
1314 			ASSERT(mi->mi_async_req_count != 0);
1315 			mi->mi_async_req_count--;
1316 		}
1317 
1318 		mutex_enter(&mi->mi_lock);
1319 		if (mi->mi_flags & MI_ASYNC_MGR_STOP) {
1320 			mutex_exit(&mi->mi_lock);
1321 			break;
1322 		}
1323 		mutex_exit(&mi->mi_lock);
1324 
1325 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
1326 		cv_wait(&mi->mi_async_reqs_cv, &mi->mi_async_lock);
1327 		CALLB_CPR_SAFE_END(&cprinfo, &mi->mi_async_lock);
1328 	}
1329 	/*
1330 	 * Let everyone know we're done.
1331 	 */
1332 	mi->mi_manager_thread = NULL;
1333 	cv_broadcast(&mi->mi_async_cv);
1334 
1335 	/*
1336 	 * There is no explicit call to mutex_exit(&mi->mi_async_lock)
1337 	 * since CALLB_CPR_EXIT is actually responsible for releasing
1338 	 * 'mi_async_lock'.
1339 	 */
1340 	CALLB_CPR_EXIT(&cprinfo);
1341 	VFS_RELE(vfsp);	/* release thread's hold */
1342 	zthread_exit();
1343 }
1344 
1345 /*
1346  * Signal (and wait for) the async manager thread to clean up and go away.
1347  */
1348 void
nfs_async_manager_stop(vfs_t * vfsp)1349 nfs_async_manager_stop(vfs_t *vfsp)
1350 {
1351 	mntinfo_t *mi = VFTOMI(vfsp);
1352 
1353 	mutex_enter(&mi->mi_async_lock);
1354 	mutex_enter(&mi->mi_lock);
1355 	mi->mi_flags |= MI_ASYNC_MGR_STOP;
1356 	mutex_exit(&mi->mi_lock);
1357 	cv_broadcast(&mi->mi_async_reqs_cv);
1358 	while (mi->mi_manager_thread != NULL)
1359 		cv_wait(&mi->mi_async_cv, &mi->mi_async_lock);
1360 	mutex_exit(&mi->mi_async_lock);
1361 }
1362 
1363 int
nfs_async_readahead(vnode_t * vp,u_offset_t blkoff,caddr_t addr,struct seg * seg,cred_t * cr,void (* readahead)(vnode_t *,u_offset_t,caddr_t,struct seg *,cred_t *))1364 nfs_async_readahead(vnode_t *vp, u_offset_t blkoff, caddr_t addr,
1365     struct seg *seg, cred_t *cr, void (*readahead)(vnode_t *,
1366     u_offset_t, caddr_t, struct seg *, cred_t *))
1367 {
1368 	rnode_t *rp;
1369 	mntinfo_t *mi;
1370 	struct nfs_async_reqs *args;
1371 
1372 	rp = VTOR(vp);
1373 	ASSERT(rp->r_freef == NULL);
1374 
1375 	mi = VTOMI(vp);
1376 
1377 	/*
1378 	 * If addr falls in a different segment, don't bother doing readahead.
1379 	 */
1380 	if (addr >= seg->s_base + seg->s_size)
1381 		return (-1);
1382 
1383 	/*
1384 	 * If we can't allocate a request structure, punt on the readahead.
1385 	 */
1386 	if ((args = kmem_alloc(sizeof (*args), KM_NOSLEEP)) == NULL)
1387 		return (-1);
1388 
1389 	/*
1390 	 * If a lock operation is pending, don't initiate any new
1391 	 * readaheads.  Otherwise, bump r_count to indicate the new
1392 	 * asynchronous I/O.
1393 	 */
1394 	if (!nfs_rw_tryenter(&rp->r_lkserlock, RW_READER)) {
1395 		kmem_free(args, sizeof (*args));
1396 		return (-1);
1397 	}
1398 	mutex_enter(&rp->r_statelock);
1399 	rp->r_count++;
1400 	mutex_exit(&rp->r_statelock);
1401 	nfs_rw_exit(&rp->r_lkserlock);
1402 
1403 	args->a_next = NULL;
1404 #ifdef DEBUG
1405 	args->a_queuer = curthread;
1406 #endif
1407 	VN_HOLD(vp);
1408 	args->a_vp = vp;
1409 	ASSERT(cr != NULL);
1410 	crhold(cr);
1411 	args->a_cred = cr;
1412 	args->a_io = NFS_READ_AHEAD;
1413 	args->a_nfs_readahead = readahead;
1414 	args->a_nfs_blkoff = blkoff;
1415 	args->a_nfs_seg = seg;
1416 	args->a_nfs_addr = addr;
1417 
1418 	mutex_enter(&mi->mi_async_lock);
1419 
1420 	/*
1421 	 * If asyncio has been disabled, don't bother readahead.
1422 	 */
1423 	if (mi->mi_max_threads == 0) {
1424 		mutex_exit(&mi->mi_async_lock);
1425 		goto noasync;
1426 	}
1427 
1428 	/*
1429 	 * Link request structure into the async list and
1430 	 * wakeup async thread to do the i/o.
1431 	 */
1432 	if (mi->mi_async_reqs[NFS_READ_AHEAD] == NULL) {
1433 		mi->mi_async_reqs[NFS_READ_AHEAD] = args;
1434 		mi->mi_async_tail[NFS_READ_AHEAD] = args;
1435 	} else {
1436 		mi->mi_async_tail[NFS_READ_AHEAD]->a_next = args;
1437 		mi->mi_async_tail[NFS_READ_AHEAD] = args;
1438 	}
1439 
1440 	if (mi->mi_io_kstats) {
1441 		mutex_enter(&mi->mi_lock);
1442 		kstat_waitq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1443 		mutex_exit(&mi->mi_lock);
1444 	}
1445 
1446 	mi->mi_async_req_count++;
1447 	ASSERT(mi->mi_async_req_count != 0);
1448 	cv_signal(&mi->mi_async_reqs_cv);
1449 	mutex_exit(&mi->mi_async_lock);
1450 	return (0);
1451 
1452 noasync:
1453 	mutex_enter(&rp->r_statelock);
1454 	rp->r_count--;
1455 	cv_broadcast(&rp->r_cv);
1456 	mutex_exit(&rp->r_statelock);
1457 	VN_RELE(vp);
1458 	crfree(cr);
1459 	kmem_free(args, sizeof (*args));
1460 	return (-1);
1461 }
1462 
1463 int
nfs_async_putapage(vnode_t * vp,page_t * pp,u_offset_t off,size_t len,int flags,cred_t * cr,int (* putapage)(vnode_t *,page_t *,u_offset_t,size_t,int,cred_t *))1464 nfs_async_putapage(vnode_t *vp, page_t *pp, u_offset_t off, size_t len,
1465     int flags, cred_t *cr, int (*putapage)(vnode_t *, page_t *,
1466     u_offset_t, size_t, int, cred_t *))
1467 {
1468 	rnode_t *rp;
1469 	mntinfo_t *mi;
1470 	struct nfs_async_reqs *args;
1471 
1472 	ASSERT(flags & B_ASYNC);
1473 	ASSERT(vp->v_vfsp != NULL);
1474 
1475 	rp = VTOR(vp);
1476 	ASSERT(rp->r_count > 0);
1477 
1478 	mi = VTOMI(vp);
1479 
1480 	/*
1481 	 * If we can't allocate a request structure, do the putpage
1482 	 * operation synchronously in this thread's context.
1483 	 */
1484 	if ((args = kmem_alloc(sizeof (*args), KM_NOSLEEP)) == NULL)
1485 		goto noasync;
1486 
1487 	args->a_next = NULL;
1488 #ifdef DEBUG
1489 	args->a_queuer = curthread;
1490 #endif
1491 	VN_HOLD(vp);
1492 	args->a_vp = vp;
1493 	ASSERT(cr != NULL);
1494 	crhold(cr);
1495 	args->a_cred = cr;
1496 	args->a_io = NFS_PUTAPAGE;
1497 	args->a_nfs_putapage = putapage;
1498 	args->a_nfs_pp = pp;
1499 	args->a_nfs_off = off;
1500 	args->a_nfs_len = (uint_t)len;
1501 	args->a_nfs_flags = flags;
1502 
1503 	mutex_enter(&mi->mi_async_lock);
1504 
1505 	/*
1506 	 * If asyncio has been disabled, then make a synchronous request.
1507 	 * This check is done a second time in case async io was diabled
1508 	 * while this thread was blocked waiting for memory pressure to
1509 	 * reduce or for the queue to drain.
1510 	 */
1511 	if (mi->mi_max_threads == 0) {
1512 		mutex_exit(&mi->mi_async_lock);
1513 		goto noasync;
1514 	}
1515 
1516 	/*
1517 	 * Link request structure into the async list and
1518 	 * wakeup async thread to do the i/o.
1519 	 */
1520 	if (mi->mi_async_reqs[NFS_PUTAPAGE] == NULL) {
1521 		mi->mi_async_reqs[NFS_PUTAPAGE] = args;
1522 		mi->mi_async_tail[NFS_PUTAPAGE] = args;
1523 	} else {
1524 		mi->mi_async_tail[NFS_PUTAPAGE]->a_next = args;
1525 		mi->mi_async_tail[NFS_PUTAPAGE] = args;
1526 	}
1527 
1528 	mutex_enter(&rp->r_statelock);
1529 	rp->r_count++;
1530 	rp->r_awcount++;
1531 	mutex_exit(&rp->r_statelock);
1532 
1533 	if (mi->mi_io_kstats) {
1534 		mutex_enter(&mi->mi_lock);
1535 		kstat_waitq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1536 		mutex_exit(&mi->mi_lock);
1537 	}
1538 
1539 	mi->mi_async_req_count++;
1540 	ASSERT(mi->mi_async_req_count != 0);
1541 	cv_signal(&mi->mi_async_reqs_cv);
1542 	mutex_exit(&mi->mi_async_lock);
1543 	return (0);
1544 
1545 noasync:
1546 	if (args != NULL) {
1547 		VN_RELE(vp);
1548 		crfree(cr);
1549 		kmem_free(args, sizeof (*args));
1550 	}
1551 
1552 	if (curproc == proc_pageout || curproc == proc_fsflush) {
1553 		/*
1554 		 * If we get here in the context of the pageout/fsflush,
1555 		 * we refuse to do a sync write, because this may hang
1556 		 * pageout (and the machine). In this case, we just
1557 		 * re-mark the page as dirty and punt on the page.
1558 		 *
1559 		 * Make sure B_FORCE isn't set.  We can re-mark the
1560 		 * pages as dirty and unlock the pages in one swoop by
1561 		 * passing in B_ERROR to pvn_write_done().  However,
1562 		 * we should make sure B_FORCE isn't set - we don't
1563 		 * want the page tossed before it gets written out.
1564 		 */
1565 		if (flags & B_FORCE)
1566 			flags &= ~(B_INVAL | B_FORCE);
1567 		pvn_write_done(pp, flags | B_ERROR);
1568 		return (0);
1569 	}
1570 	if (nfs_zone() != mi->mi_zone) {
1571 		/*
1572 		 * So this was a cross-zone sync putpage.  We pass in B_ERROR
1573 		 * to pvn_write_done() to re-mark the pages as dirty and unlock
1574 		 * them.
1575 		 *
1576 		 * We don't want to clear B_FORCE here as the caller presumably
1577 		 * knows what they're doing if they set it.
1578 		 */
1579 		pvn_write_done(pp, flags | B_ERROR);
1580 		return (EPERM);
1581 	}
1582 	return ((*putapage)(vp, pp, off, len, flags, cr));
1583 }
1584 
1585 int
nfs_async_pageio(vnode_t * vp,page_t * pp,u_offset_t io_off,size_t io_len,int flags,cred_t * cr,int (* pageio)(vnode_t *,page_t *,u_offset_t,size_t,int,cred_t *))1586 nfs_async_pageio(vnode_t *vp, page_t *pp, u_offset_t io_off, size_t io_len,
1587     int flags, cred_t *cr, int (*pageio)(vnode_t *, page_t *, u_offset_t,
1588     size_t, int, cred_t *))
1589 {
1590 	rnode_t *rp;
1591 	mntinfo_t *mi;
1592 	struct nfs_async_reqs *args;
1593 
1594 	ASSERT(flags & B_ASYNC);
1595 	ASSERT(vp->v_vfsp != NULL);
1596 
1597 	rp = VTOR(vp);
1598 	ASSERT(rp->r_count > 0);
1599 
1600 	mi = VTOMI(vp);
1601 
1602 	/*
1603 	 * If we can't allocate a request structure, do the pageio
1604 	 * request synchronously in this thread's context.
1605 	 */
1606 	if ((args = kmem_alloc(sizeof (*args), KM_NOSLEEP)) == NULL)
1607 		goto noasync;
1608 
1609 	args->a_next = NULL;
1610 #ifdef DEBUG
1611 	args->a_queuer = curthread;
1612 #endif
1613 	VN_HOLD(vp);
1614 	args->a_vp = vp;
1615 	ASSERT(cr != NULL);
1616 	crhold(cr);
1617 	args->a_cred = cr;
1618 	args->a_io = NFS_PAGEIO;
1619 	args->a_nfs_pageio = pageio;
1620 	args->a_nfs_pp = pp;
1621 	args->a_nfs_off = io_off;
1622 	args->a_nfs_len = (uint_t)io_len;
1623 	args->a_nfs_flags = flags;
1624 
1625 	mutex_enter(&mi->mi_async_lock);
1626 
1627 	/*
1628 	 * If asyncio has been disabled, then make a synchronous request.
1629 	 * This check is done a second time in case async io was diabled
1630 	 * while this thread was blocked waiting for memory pressure to
1631 	 * reduce or for the queue to drain.
1632 	 */
1633 	if (mi->mi_max_threads == 0) {
1634 		mutex_exit(&mi->mi_async_lock);
1635 		goto noasync;
1636 	}
1637 
1638 	/*
1639 	 * Link request structure into the async list and
1640 	 * wakeup async thread to do the i/o.
1641 	 */
1642 	if (mi->mi_async_reqs[NFS_PAGEIO] == NULL) {
1643 		mi->mi_async_reqs[NFS_PAGEIO] = args;
1644 		mi->mi_async_tail[NFS_PAGEIO] = args;
1645 	} else {
1646 		mi->mi_async_tail[NFS_PAGEIO]->a_next = args;
1647 		mi->mi_async_tail[NFS_PAGEIO] = args;
1648 	}
1649 
1650 	mutex_enter(&rp->r_statelock);
1651 	rp->r_count++;
1652 	rp->r_awcount++;
1653 	mutex_exit(&rp->r_statelock);
1654 
1655 	if (mi->mi_io_kstats) {
1656 		mutex_enter(&mi->mi_lock);
1657 		kstat_waitq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1658 		mutex_exit(&mi->mi_lock);
1659 	}
1660 
1661 	mi->mi_async_req_count++;
1662 	ASSERT(mi->mi_async_req_count != 0);
1663 	cv_signal(&mi->mi_async_reqs_cv);
1664 	mutex_exit(&mi->mi_async_lock);
1665 	return (0);
1666 
1667 noasync:
1668 	if (args != NULL) {
1669 		VN_RELE(vp);
1670 		crfree(cr);
1671 		kmem_free(args, sizeof (*args));
1672 	}
1673 
1674 	/*
1675 	 * If we can't do it ASYNC, for reads we do nothing (but cleanup
1676 	 * the page list), for writes we do it synchronously, except for
1677 	 * proc_pageout/proc_fsflush as described below.
1678 	 */
1679 	if (flags & B_READ) {
1680 		pvn_read_done(pp, flags | B_ERROR);
1681 		return (0);
1682 	}
1683 
1684 	if (curproc == proc_pageout || curproc == proc_fsflush) {
1685 		/*
1686 		 * If we get here in the context of the pageout/fsflush,
1687 		 * we refuse to do a sync write, because this may hang
1688 		 * pageout/fsflush (and the machine). In this case, we just
1689 		 * re-mark the page as dirty and punt on the page.
1690 		 *
1691 		 * Make sure B_FORCE isn't set.  We can re-mark the
1692 		 * pages as dirty and unlock the pages in one swoop by
1693 		 * passing in B_ERROR to pvn_write_done().  However,
1694 		 * we should make sure B_FORCE isn't set - we don't
1695 		 * want the page tossed before it gets written out.
1696 		 */
1697 		if (flags & B_FORCE)
1698 			flags &= ~(B_INVAL | B_FORCE);
1699 		pvn_write_done(pp, flags | B_ERROR);
1700 		return (0);
1701 	}
1702 
1703 	if (nfs_zone() != mi->mi_zone) {
1704 		/*
1705 		 * So this was a cross-zone sync pageio.  We pass in B_ERROR
1706 		 * to pvn_write_done() to re-mark the pages as dirty and unlock
1707 		 * them.
1708 		 *
1709 		 * We don't want to clear B_FORCE here as the caller presumably
1710 		 * knows what they're doing if they set it.
1711 		 */
1712 		pvn_write_done(pp, flags | B_ERROR);
1713 		return (EPERM);
1714 	}
1715 	return ((*pageio)(vp, pp, io_off, io_len, flags, cr));
1716 }
1717 
1718 void
nfs_async_readdir(vnode_t * vp,rddir_cache * rdc,cred_t * cr,int (* readdir)(vnode_t *,rddir_cache *,cred_t *))1719 nfs_async_readdir(vnode_t *vp, rddir_cache *rdc, cred_t *cr,
1720     int (*readdir)(vnode_t *, rddir_cache *, cred_t *))
1721 {
1722 	rnode_t *rp;
1723 	mntinfo_t *mi;
1724 	struct nfs_async_reqs *args;
1725 
1726 	rp = VTOR(vp);
1727 	ASSERT(rp->r_freef == NULL);
1728 
1729 	mi = VTOMI(vp);
1730 
1731 	/*
1732 	 * If we can't allocate a request structure, do the readdir
1733 	 * operation synchronously in this thread's context.
1734 	 */
1735 	if ((args = kmem_alloc(sizeof (*args), KM_NOSLEEP)) == NULL)
1736 		goto noasync;
1737 
1738 	args->a_next = NULL;
1739 #ifdef DEBUG
1740 	args->a_queuer = curthread;
1741 #endif
1742 	VN_HOLD(vp);
1743 	args->a_vp = vp;
1744 	ASSERT(cr != NULL);
1745 	crhold(cr);
1746 	args->a_cred = cr;
1747 	args->a_io = NFS_READDIR;
1748 	args->a_nfs_readdir = readdir;
1749 	args->a_nfs_rdc = rdc;
1750 
1751 	mutex_enter(&mi->mi_async_lock);
1752 
1753 	/*
1754 	 * If asyncio has been disabled, then make a synchronous request.
1755 	 */
1756 	if (mi->mi_max_threads == 0) {
1757 		mutex_exit(&mi->mi_async_lock);
1758 		goto noasync;
1759 	}
1760 
1761 	/*
1762 	 * Link request structure into the async list and
1763 	 * wakeup async thread to do the i/o.
1764 	 */
1765 	if (mi->mi_async_reqs[NFS_READDIR] == NULL) {
1766 		mi->mi_async_reqs[NFS_READDIR] = args;
1767 		mi->mi_async_tail[NFS_READDIR] = args;
1768 	} else {
1769 		mi->mi_async_tail[NFS_READDIR]->a_next = args;
1770 		mi->mi_async_tail[NFS_READDIR] = args;
1771 	}
1772 
1773 	mutex_enter(&rp->r_statelock);
1774 	rp->r_count++;
1775 	mutex_exit(&rp->r_statelock);
1776 
1777 	if (mi->mi_io_kstats) {
1778 		mutex_enter(&mi->mi_lock);
1779 		kstat_waitq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1780 		mutex_exit(&mi->mi_lock);
1781 	}
1782 
1783 	mi->mi_async_req_count++;
1784 	ASSERT(mi->mi_async_req_count != 0);
1785 	cv_signal(&mi->mi_async_reqs_cv);
1786 	mutex_exit(&mi->mi_async_lock);
1787 	return;
1788 
1789 noasync:
1790 	if (args != NULL) {
1791 		VN_RELE(vp);
1792 		crfree(cr);
1793 		kmem_free(args, sizeof (*args));
1794 	}
1795 
1796 	rdc->entries = NULL;
1797 	mutex_enter(&rp->r_statelock);
1798 	ASSERT(rdc->flags & RDDIR);
1799 	rdc->flags &= ~RDDIR;
1800 	rdc->flags |= RDDIRREQ;
1801 	/*
1802 	 * Check the flag to see if RDDIRWAIT is set. If RDDIRWAIT
1803 	 * is set, wakeup the thread sleeping in cv_wait_sig().
1804 	 * The woken up thread will reset the flag to RDDIR and will
1805 	 * continue with the readdir opeartion.
1806 	 */
1807 	if (rdc->flags & RDDIRWAIT) {
1808 		rdc->flags &= ~RDDIRWAIT;
1809 		cv_broadcast(&rdc->cv);
1810 	}
1811 	mutex_exit(&rp->r_statelock);
1812 	rddir_cache_rele(rdc);
1813 }
1814 
1815 void
nfs_async_commit(vnode_t * vp,page_t * plist,offset3 offset,count3 count,cred_t * cr,void (* commit)(vnode_t *,page_t *,offset3,count3,cred_t *))1816 nfs_async_commit(vnode_t *vp, page_t *plist, offset3 offset, count3 count,
1817     cred_t *cr, void (*commit)(vnode_t *, page_t *, offset3, count3, cred_t *))
1818 {
1819 	rnode_t *rp;
1820 	mntinfo_t *mi;
1821 	struct nfs_async_reqs *args;
1822 	page_t *pp;
1823 
1824 	rp = VTOR(vp);
1825 	mi = VTOMI(vp);
1826 
1827 	/*
1828 	 * If we can't allocate a request structure, do the commit
1829 	 * operation synchronously in this thread's context.
1830 	 */
1831 	if ((args = kmem_alloc(sizeof (*args), KM_NOSLEEP)) == NULL)
1832 		goto noasync;
1833 
1834 	args->a_next = NULL;
1835 #ifdef DEBUG
1836 	args->a_queuer = curthread;
1837 #endif
1838 	VN_HOLD(vp);
1839 	args->a_vp = vp;
1840 	ASSERT(cr != NULL);
1841 	crhold(cr);
1842 	args->a_cred = cr;
1843 	args->a_io = NFS_COMMIT;
1844 	args->a_nfs_commit = commit;
1845 	args->a_nfs_plist = plist;
1846 	args->a_nfs_offset = offset;
1847 	args->a_nfs_count = count;
1848 
1849 	mutex_enter(&mi->mi_async_lock);
1850 
1851 	/*
1852 	 * If asyncio has been disabled, then make a synchronous request.
1853 	 * This check is done a second time in case async io was diabled
1854 	 * while this thread was blocked waiting for memory pressure to
1855 	 * reduce or for the queue to drain.
1856 	 */
1857 	if (mi->mi_max_threads == 0) {
1858 		mutex_exit(&mi->mi_async_lock);
1859 		goto noasync;
1860 	}
1861 
1862 	/*
1863 	 * Link request structure into the async list and
1864 	 * wakeup async thread to do the i/o.
1865 	 */
1866 	if (mi->mi_async_reqs[NFS_COMMIT] == NULL) {
1867 		mi->mi_async_reqs[NFS_COMMIT] = args;
1868 		mi->mi_async_tail[NFS_COMMIT] = args;
1869 	} else {
1870 		mi->mi_async_tail[NFS_COMMIT]->a_next = args;
1871 		mi->mi_async_tail[NFS_COMMIT] = args;
1872 	}
1873 
1874 	mutex_enter(&rp->r_statelock);
1875 	rp->r_count++;
1876 	mutex_exit(&rp->r_statelock);
1877 
1878 	if (mi->mi_io_kstats) {
1879 		mutex_enter(&mi->mi_lock);
1880 		kstat_waitq_enter(KSTAT_IO_PTR(mi->mi_io_kstats));
1881 		mutex_exit(&mi->mi_lock);
1882 	}
1883 
1884 	mi->mi_async_req_count++;
1885 	ASSERT(mi->mi_async_req_count != 0);
1886 	cv_signal(&mi->mi_async_reqs_cv);
1887 	mutex_exit(&mi->mi_async_lock);
1888 	return;
1889 
1890 noasync:
1891 	if (args != NULL) {
1892 		VN_RELE(vp);
1893 		crfree(cr);
1894 		kmem_free(args, sizeof (*args));
1895 	}
1896 
1897 	if (curproc == proc_pageout || curproc == proc_fsflush ||
1898 	    nfs_zone() != mi->mi_zone) {
1899 		while (plist != NULL) {
1900 			pp = plist;
1901 			page_sub(&plist, pp);
1902 			pp->p_fsdata = C_COMMIT;
1903 			page_unlock(pp);
1904 		}
1905 		return;
1906 	}
1907 	(*commit)(vp, plist, offset, count, cr);
1908 }
1909 
1910 void
nfs_async_inactive(vnode_t * vp,cred_t * cr,void (* inactive)(vnode_t *,cred_t *,caller_context_t *))1911 nfs_async_inactive(vnode_t *vp, cred_t *cr,
1912     void (*inactive)(vnode_t *, cred_t *, caller_context_t *))
1913 {
1914 	mntinfo_t *mi;
1915 	struct nfs_async_reqs *args;
1916 
1917 	mi = VTOMI(vp);
1918 
1919 	args = kmem_alloc(sizeof (*args), KM_SLEEP);
1920 	args->a_next = NULL;
1921 #ifdef DEBUG
1922 	args->a_queuer = curthread;
1923 #endif
1924 	args->a_vp = vp;
1925 	ASSERT(cr != NULL);
1926 	crhold(cr);
1927 	args->a_cred = cr;
1928 	args->a_io = NFS_INACTIVE;
1929 	args->a_nfs_inactive = inactive;
1930 
1931 	/*
1932 	 * Note that we don't check mi->mi_max_threads here, since we
1933 	 * *need* to get rid of this vnode regardless of whether someone
1934 	 * set nfs3_max_threads/nfs_max_threads to zero in /etc/system.
1935 	 *
1936 	 * The manager thread knows about this and is willing to create
1937 	 * at least one thread to accommodate us.
1938 	 */
1939 	mutex_enter(&mi->mi_async_lock);
1940 	if (mi->mi_manager_thread == NULL) {
1941 		rnode_t *rp = VTOR(vp);
1942 
1943 		mutex_exit(&mi->mi_async_lock);
1944 		crfree(cr);	/* drop our reference */
1945 		kmem_free(args, sizeof (*args));
1946 		/*
1947 		 * We can't do an over-the-wire call since we're in the wrong
1948 		 * zone, so we need to clean up state as best we can and then
1949 		 * throw away the vnode.
1950 		 */
1951 		mutex_enter(&rp->r_statelock);
1952 		if (rp->r_unldvp != NULL) {
1953 			vnode_t *unldvp;
1954 			char *unlname;
1955 			cred_t *unlcred;
1956 
1957 			unldvp = rp->r_unldvp;
1958 			rp->r_unldvp = NULL;
1959 			unlname = rp->r_unlname;
1960 			rp->r_unlname = NULL;
1961 			unlcred = rp->r_unlcred;
1962 			rp->r_unlcred = NULL;
1963 			mutex_exit(&rp->r_statelock);
1964 
1965 			VN_RELE(unldvp);
1966 			kmem_free(unlname, MAXNAMELEN);
1967 			crfree(unlcred);
1968 		} else {
1969 			mutex_exit(&rp->r_statelock);
1970 		}
1971 		/*
1972 		 * No need to explicitly throw away any cached pages.  The
1973 		 * eventual rinactive() will attempt a synchronous
1974 		 * VOP_PUTPAGE() which will immediately fail since the request
1975 		 * is coming from the wrong zone, and then will proceed to call
1976 		 * nfs_invalidate_pages() which will clean things up for us.
1977 		 */
1978 		rp_addfree(VTOR(vp), cr);
1979 		return;
1980 	}
1981 
1982 	if (mi->mi_async_reqs[NFS_INACTIVE] == NULL) {
1983 		mi->mi_async_reqs[NFS_INACTIVE] = args;
1984 	} else {
1985 		mi->mi_async_tail[NFS_INACTIVE]->a_next = args;
1986 	}
1987 	mi->mi_async_tail[NFS_INACTIVE] = args;
1988 	/*
1989 	 * Don't increment r_count, since we're trying to get rid of the vnode.
1990 	 */
1991 
1992 	mi->mi_async_req_count++;
1993 	ASSERT(mi->mi_async_req_count != 0);
1994 	cv_signal(&mi->mi_async_reqs_cv);
1995 	mutex_exit(&mi->mi_async_lock);
1996 }
1997 
1998 static void
nfs_async_start(struct vfs * vfsp)1999 nfs_async_start(struct vfs *vfsp)
2000 {
2001 	nfs_async_common_start(vfsp, NFS_ASYNC_QUEUE);
2002 }
2003 
2004 static void
nfs_async_pgops_start(struct vfs * vfsp)2005 nfs_async_pgops_start(struct vfs *vfsp)
2006 {
2007 	nfs_async_common_start(vfsp, NFS_ASYNC_PGOPS_QUEUE);
2008 }
2009 
2010 /*
2011  * The async queues for each mounted file system are arranged as a
2012  * set of queues, one for each async i/o type.  Requests are taken
2013  * from the queues in a round-robin fashion.  A number of consecutive
2014  * requests are taken from each queue before moving on to the next
2015  * queue.  This functionality may allow the NFS Version 2 server to do
2016  * write clustering, even if the client is mixing writes and reads
2017  * because it will take multiple write requests from the queue
2018  * before processing any of the other async i/o types.
2019  *
2020  * XXX The nfs_async_common_start thread is unsafe in the light of the present
2021  * model defined by cpr to suspend the system. Specifically over the
2022  * wire calls are cpr-unsafe. The thread should be reevaluated in
2023  * case of future updates to the cpr model.
2024  */
2025 static void
nfs_async_common_start(struct vfs * vfsp,int async_queue)2026 nfs_async_common_start(struct vfs *vfsp, int async_queue)
2027 {
2028 	struct nfs_async_reqs *args;
2029 	mntinfo_t *mi = VFTOMI(vfsp);
2030 	clock_t time_left = 1;
2031 	callb_cpr_t cprinfo;
2032 	int i;
2033 	int async_types;
2034 	kcondvar_t *async_work_cv;
2035 
2036 	if (async_queue == NFS_ASYNC_QUEUE) {
2037 		async_types = NFS_ASYNC_TYPES;
2038 		async_work_cv = &mi->mi_async_work_cv[NFS_ASYNC_QUEUE];
2039 	} else {
2040 		async_types = NFS_ASYNC_PGOPS_TYPES;
2041 		async_work_cv = &mi->mi_async_work_cv[NFS_ASYNC_PGOPS_QUEUE];
2042 	}
2043 
2044 	/*
2045 	 * Dynamic initialization of nfs_async_timeout to allow nfs to be
2046 	 * built in an implementation independent manner.
2047 	 */
2048 	if (nfs_async_timeout == -1)
2049 		nfs_async_timeout = NFS_ASYNC_TIMEOUT;
2050 
2051 	CALLB_CPR_INIT(&cprinfo, &mi->mi_async_lock, callb_generic_cpr, "nas");
2052 
2053 	mutex_enter(&mi->mi_async_lock);
2054 	for (;;) {
2055 		/*
2056 		 * Find the next queue containing an entry.  We start
2057 		 * at the current queue pointer and then round robin
2058 		 * through all of them until we either find a non-empty
2059 		 * queue or have looked through all of them.
2060 		 */
2061 		for (i = 0; i < async_types; i++) {
2062 			args = *mi->mi_async_curr[async_queue];
2063 			if (args != NULL)
2064 				break;
2065 			mi->mi_async_curr[async_queue]++;
2066 			if (mi->mi_async_curr[async_queue] ==
2067 			    &mi->mi_async_reqs[async_types]) {
2068 				mi->mi_async_curr[async_queue] =
2069 				    &mi->mi_async_reqs[0];
2070 			}
2071 		}
2072 		/*
2073 		 * If we didn't find a entry, then block until woken up
2074 		 * again and then look through the queues again.
2075 		 */
2076 		if (args == NULL) {
2077 			/*
2078 			 * Exiting is considered to be safe for CPR as well
2079 			 */
2080 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
2081 
2082 			/*
2083 			 * Wakeup thread waiting to unmount the file
2084 			 * system only if all async threads are inactive.
2085 			 *
2086 			 * If we've timed-out and there's nothing to do,
2087 			 * then get rid of this thread.
2088 			 */
2089 			if (mi->mi_max_threads == 0 || time_left <= 0) {
2090 				--mi->mi_threads[async_queue];
2091 
2092 				if (mi->mi_threads[NFS_ASYNC_QUEUE] == 0 &&
2093 				    mi->mi_threads[NFS_ASYNC_PGOPS_QUEUE] == 0)
2094 					cv_signal(&mi->mi_async_cv);
2095 				CALLB_CPR_EXIT(&cprinfo);
2096 				VFS_RELE(vfsp);	/* release thread's hold */
2097 				zthread_exit();
2098 				/* NOTREACHED */
2099 			}
2100 			time_left = cv_reltimedwait(async_work_cv,
2101 			    &mi->mi_async_lock, nfs_async_timeout,
2102 			    TR_CLOCK_TICK);
2103 
2104 			CALLB_CPR_SAFE_END(&cprinfo, &mi->mi_async_lock);
2105 
2106 			continue;
2107 		}
2108 		time_left = 1;
2109 
2110 		/*
2111 		 * Remove the request from the async queue and then
2112 		 * update the current async request queue pointer.  If
2113 		 * the current queue is empty or we have removed enough
2114 		 * consecutive entries from it, then reset the counter
2115 		 * for this queue and then move the current pointer to
2116 		 * the next queue.
2117 		 */
2118 		*mi->mi_async_curr[async_queue] = args->a_next;
2119 		if (*mi->mi_async_curr[async_queue] == NULL ||
2120 		    --mi->mi_async_clusters[args->a_io] == 0) {
2121 			mi->mi_async_clusters[args->a_io] =
2122 			    mi->mi_async_init_clusters;
2123 			mi->mi_async_curr[async_queue]++;
2124 			if (mi->mi_async_curr[async_queue] ==
2125 			    &mi->mi_async_reqs[async_types]) {
2126 				mi->mi_async_curr[async_queue] =
2127 				    &mi->mi_async_reqs[0];
2128 			}
2129 		}
2130 
2131 		if (args->a_io != NFS_INACTIVE && mi->mi_io_kstats) {
2132 			mutex_enter(&mi->mi_lock);
2133 			kstat_waitq_exit(KSTAT_IO_PTR(mi->mi_io_kstats));
2134 			mutex_exit(&mi->mi_lock);
2135 		}
2136 
2137 		mutex_exit(&mi->mi_async_lock);
2138 
2139 		/*
2140 		 * Obtain arguments from the async request structure.
2141 		 */
2142 		if (args->a_io == NFS_READ_AHEAD && mi->mi_max_threads > 0) {
2143 			(*args->a_nfs_readahead)(args->a_vp, args->a_nfs_blkoff,
2144 			    args->a_nfs_addr, args->a_nfs_seg,
2145 			    args->a_cred);
2146 		} else if (args->a_io == NFS_PUTAPAGE) {
2147 			(void) (*args->a_nfs_putapage)(args->a_vp,
2148 			    args->a_nfs_pp, args->a_nfs_off,
2149 			    args->a_nfs_len, args->a_nfs_flags,
2150 			    args->a_cred);
2151 		} else if (args->a_io == NFS_PAGEIO) {
2152 			(void) (*args->a_nfs_pageio)(args->a_vp,
2153 			    args->a_nfs_pp, args->a_nfs_off,
2154 			    args->a_nfs_len, args->a_nfs_flags,
2155 			    args->a_cred);
2156 		} else if (args->a_io == NFS_READDIR) {
2157 			(void) ((*args->a_nfs_readdir)(args->a_vp,
2158 			    args->a_nfs_rdc, args->a_cred));
2159 		} else if (args->a_io == NFS_COMMIT) {
2160 			(*args->a_nfs_commit)(args->a_vp, args->a_nfs_plist,
2161 			    args->a_nfs_offset, args->a_nfs_count,
2162 			    args->a_cred);
2163 		} else if (args->a_io == NFS_INACTIVE) {
2164 			(*args->a_nfs_inactive)(args->a_vp, args->a_cred, NULL);
2165 		}
2166 
2167 		/*
2168 		 * Now, release the vnode and free the credentials
2169 		 * structure.
2170 		 */
2171 		free_async_args(args);
2172 		/*
2173 		 * Reacquire the mutex because it will be needed above.
2174 		 */
2175 		mutex_enter(&mi->mi_async_lock);
2176 	}
2177 }
2178 
2179 void
nfs_async_stop(struct vfs * vfsp)2180 nfs_async_stop(struct vfs *vfsp)
2181 {
2182 	mntinfo_t *mi = VFTOMI(vfsp);
2183 
2184 	/*
2185 	 * Wait for all outstanding async operations to complete and for the
2186 	 * worker threads to exit.
2187 	 */
2188 	mutex_enter(&mi->mi_async_lock);
2189 	mi->mi_max_threads = 0;
2190 	NFS_WAKEALL_ASYNC_WORKERS(mi->mi_async_work_cv);
2191 	while (mi->mi_threads[NFS_ASYNC_QUEUE] != 0 ||
2192 	    mi->mi_threads[NFS_ASYNC_PGOPS_QUEUE] != 0)
2193 		cv_wait(&mi->mi_async_cv, &mi->mi_async_lock);
2194 	mutex_exit(&mi->mi_async_lock);
2195 }
2196 
2197 /*
2198  * nfs_async_stop_sig:
2199  * Wait for all outstanding putpage operation to complete. If a signal
2200  * is deliver we will abort and return non-zero. If we can put all the
2201  * pages we will return 0. This routine is called from nfs_unmount and
2202  * nfs3_unmount to make these operations interruptible.
2203  */
2204 int
nfs_async_stop_sig(struct vfs * vfsp)2205 nfs_async_stop_sig(struct vfs *vfsp)
2206 {
2207 	mntinfo_t *mi = VFTOMI(vfsp);
2208 	ushort_t omax;
2209 	int rval;
2210 
2211 	/*
2212 	 * Wait for all outstanding async operations to complete and for the
2213 	 * worker threads to exit.
2214 	 */
2215 	mutex_enter(&mi->mi_async_lock);
2216 	omax = mi->mi_max_threads;
2217 	mi->mi_max_threads = 0;
2218 	/*
2219 	 * Tell all the worker threads to exit.
2220 	 */
2221 	NFS_WAKEALL_ASYNC_WORKERS(mi->mi_async_work_cv);
2222 	while (mi->mi_threads[NFS_ASYNC_QUEUE] != 0 ||
2223 	    mi->mi_threads[NFS_ASYNC_PGOPS_QUEUE] != 0) {
2224 		if (!cv_wait_sig(&mi->mi_async_cv, &mi->mi_async_lock))
2225 			break;
2226 	}
2227 	rval = (mi->mi_threads[NFS_ASYNC_QUEUE] != 0 ||
2228 	    mi->mi_threads[NFS_ASYNC_PGOPS_QUEUE]  != 0); /* Interrupted */
2229 	if (rval)
2230 		mi->mi_max_threads = omax;
2231 	mutex_exit(&mi->mi_async_lock);
2232 
2233 	return (rval);
2234 }
2235 
2236 int
writerp(rnode_t * rp,caddr_t base,int tcount,struct uio * uio,int pgcreated)2237 writerp(rnode_t *rp, caddr_t base, int tcount, struct uio *uio, int pgcreated)
2238 {
2239 	int pagecreate;
2240 	int n;
2241 	int saved_n;
2242 	caddr_t saved_base;
2243 	u_offset_t offset;
2244 	int error;
2245 	int sm_error;
2246 	vnode_t *vp = RTOV(rp);
2247 
2248 	ASSERT(tcount <= MAXBSIZE && tcount <= uio->uio_resid);
2249 	ASSERT(nfs_rw_lock_held(&rp->r_rwlock, RW_WRITER));
2250 	if (!vpm_enable) {
2251 		ASSERT(((uintptr_t)base & MAXBOFFSET) + tcount <= MAXBSIZE);
2252 	}
2253 
2254 	/*
2255 	 * Move bytes in at most PAGESIZE chunks. We must avoid
2256 	 * spanning pages in uiomove() because page faults may cause
2257 	 * the cache to be invalidated out from under us. The r_size is not
2258 	 * updated until after the uiomove. If we push the last page of a
2259 	 * file before r_size is correct, we will lose the data written past
2260 	 * the current (and invalid) r_size.
2261 	 */
2262 	do {
2263 		offset = uio->uio_loffset;
2264 		pagecreate = 0;
2265 
2266 		/*
2267 		 * n is the number of bytes required to satisfy the request
2268 		 *   or the number of bytes to fill out the page.
2269 		 */
2270 		n = (int)MIN((PAGESIZE - (offset & PAGEOFFSET)), tcount);
2271 
2272 		/*
2273 		 * Check to see if we can skip reading in the page
2274 		 * and just allocate the memory.  We can do this
2275 		 * if we are going to rewrite the entire mapping
2276 		 * or if we are going to write to or beyond the current
2277 		 * end of file from the beginning of the mapping.
2278 		 *
2279 		 * The read of r_size is now protected by r_statelock.
2280 		 */
2281 		mutex_enter(&rp->r_statelock);
2282 		/*
2283 		 * When pgcreated is nonzero the caller has already done
2284 		 * a segmap_getmapflt with forcefault 0 and S_WRITE. With
2285 		 * segkpm this means we already have at least one page
2286 		 * created and mapped at base.
2287 		 */
2288 		pagecreate = pgcreated ||
2289 		    ((offset & PAGEOFFSET) == 0 &&
2290 		    (n == PAGESIZE || ((offset + n) >= rp->r_size)));
2291 
2292 		mutex_exit(&rp->r_statelock);
2293 		if (!vpm_enable && pagecreate) {
2294 			/*
2295 			 * The last argument tells segmap_pagecreate() to
2296 			 * always lock the page, as opposed to sometimes
2297 			 * returning with the page locked. This way we avoid a
2298 			 * fault on the ensuing uiomove(), but also
2299 			 * more importantly (to fix bug 1094402) we can
2300 			 * call segmap_fault() to unlock the page in all
2301 			 * cases. An alternative would be to modify
2302 			 * segmap_pagecreate() to tell us when it is
2303 			 * locking a page, but that's a fairly major
2304 			 * interface change.
2305 			 */
2306 			if (pgcreated == 0)
2307 				(void) segmap_pagecreate(segkmap, base,
2308 				    (uint_t)n, 1);
2309 			saved_base = base;
2310 			saved_n = n;
2311 		}
2312 
2313 		/*
2314 		 * The number of bytes of data in the last page can not
2315 		 * be accurately be determined while page is being
2316 		 * uiomove'd to and the size of the file being updated.
2317 		 * Thus, inform threads which need to know accurately
2318 		 * how much data is in the last page of the file.  They
2319 		 * will not do the i/o immediately, but will arrange for
2320 		 * the i/o to happen later when this modify operation
2321 		 * will have finished.
2322 		 */
2323 		ASSERT(!(rp->r_flags & RMODINPROGRESS));
2324 		mutex_enter(&rp->r_statelock);
2325 		rp->r_flags |= RMODINPROGRESS;
2326 		rp->r_modaddr = (offset & MAXBMASK);
2327 		mutex_exit(&rp->r_statelock);
2328 
2329 		if (vpm_enable) {
2330 			/*
2331 			 * Copy data. If new pages are created, part of
2332 			 * the page that is not written will be initizliazed
2333 			 * with zeros.
2334 			 */
2335 			error = vpm_data_copy(vp, offset, n, uio,
2336 			    !pagecreate, NULL, 0, S_WRITE);
2337 		} else {
2338 			error = uiomove(base, n, UIO_WRITE, uio);
2339 		}
2340 
2341 		/*
2342 		 * r_size is the maximum number of
2343 		 * bytes known to be in the file.
2344 		 * Make sure it is at least as high as the
2345 		 * first unwritten byte pointed to by uio_loffset.
2346 		 */
2347 		mutex_enter(&rp->r_statelock);
2348 		if (rp->r_size < uio->uio_loffset)
2349 			rp->r_size = uio->uio_loffset;
2350 		rp->r_flags &= ~RMODINPROGRESS;
2351 		rp->r_flags |= RDIRTY;
2352 		mutex_exit(&rp->r_statelock);
2353 
2354 		/* n = # of bytes written */
2355 		n = (int)(uio->uio_loffset - offset);
2356 
2357 		if (!vpm_enable) {
2358 			base += n;
2359 		}
2360 		tcount -= n;
2361 		/*
2362 		 * If we created pages w/o initializing them completely,
2363 		 * we need to zero the part that wasn't set up.
2364 		 * This happens on a most EOF write cases and if
2365 		 * we had some sort of error during the uiomove.
2366 		 */
2367 		if (!vpm_enable && pagecreate) {
2368 			if ((uio->uio_loffset & PAGEOFFSET) || n == 0)
2369 				(void) kzero(base, PAGESIZE - n);
2370 
2371 			if (pgcreated) {
2372 				/*
2373 				 * Caller is responsible for this page,
2374 				 * it was not created in this loop.
2375 				 */
2376 				pgcreated = 0;
2377 			} else {
2378 				/*
2379 				 * For bug 1094402: segmap_pagecreate locks
2380 				 * page. Unlock it. This also unlocks the
2381 				 * pages allocated by page_create_va() in
2382 				 * segmap_pagecreate().
2383 				 */
2384 				sm_error = segmap_fault(kas.a_hat, segkmap,
2385 				    saved_base, saved_n,
2386 				    F_SOFTUNLOCK, S_WRITE);
2387 				if (error == 0)
2388 					error = sm_error;
2389 			}
2390 		}
2391 	} while (tcount > 0 && error == 0);
2392 
2393 	return (error);
2394 }
2395 
2396 int
nfs_putpages(vnode_t * vp,u_offset_t off,size_t len,int flags,cred_t * cr)2397 nfs_putpages(vnode_t *vp, u_offset_t off, size_t len, int flags, cred_t *cr)
2398 {
2399 	rnode_t *rp;
2400 	page_t *pp;
2401 	u_offset_t eoff;
2402 	u_offset_t io_off;
2403 	size_t io_len;
2404 	int error;
2405 	int rdirty;
2406 	int err;
2407 
2408 	rp = VTOR(vp);
2409 	ASSERT(rp->r_count > 0);
2410 
2411 	if (!vn_has_cached_data(vp))
2412 		return (0);
2413 
2414 	ASSERT(vp->v_type != VCHR);
2415 
2416 	/*
2417 	 * If ROUTOFSPACE is set, then all writes turn into B_INVAL
2418 	 * writes.  B_FORCE is set to force the VM system to actually
2419 	 * invalidate the pages, even if the i/o failed.  The pages
2420 	 * need to get invalidated because they can't be written out
2421 	 * because there isn't any space left on either the server's
2422 	 * file system or in the user's disk quota.  The B_FREE bit
2423 	 * is cleared to avoid confusion as to whether this is a
2424 	 * request to place the page on the freelist or to destroy
2425 	 * it.
2426 	 */
2427 	if ((rp->r_flags & ROUTOFSPACE) ||
2428 	    (vp->v_vfsp->vfs_flag & VFS_UNMOUNTED))
2429 		flags = (flags & ~B_FREE) | B_INVAL | B_FORCE;
2430 
2431 	if (len == 0) {
2432 		/*
2433 		 * If doing a full file synchronous operation, then clear
2434 		 * the RDIRTY bit.  If a page gets dirtied while the flush
2435 		 * is happening, then RDIRTY will get set again.  The
2436 		 * RDIRTY bit must get cleared before the flush so that
2437 		 * we don't lose this information.
2438 		 *
2439 		 * If there are no full file async write operations
2440 		 * pending and RDIRTY bit is set, clear it.
2441 		 */
2442 		if (off == (u_offset_t)0 &&
2443 		    !(flags & B_ASYNC) &&
2444 		    (rp->r_flags & RDIRTY)) {
2445 			mutex_enter(&rp->r_statelock);
2446 			rdirty = (rp->r_flags & RDIRTY);
2447 			rp->r_flags &= ~RDIRTY;
2448 			mutex_exit(&rp->r_statelock);
2449 		} else if (flags & B_ASYNC && off == (u_offset_t)0) {
2450 			mutex_enter(&rp->r_statelock);
2451 			if (rp->r_flags & RDIRTY && rp->r_awcount == 0) {
2452 				rdirty = (rp->r_flags & RDIRTY);
2453 				rp->r_flags &= ~RDIRTY;
2454 			}
2455 			mutex_exit(&rp->r_statelock);
2456 		} else
2457 			rdirty = 0;
2458 
2459 		/*
2460 		 * Search the entire vp list for pages >= off, and flush
2461 		 * the dirty pages.
2462 		 */
2463 		error = pvn_vplist_dirty(vp, off, rp->r_putapage,
2464 		    flags, cr);
2465 
2466 		/*
2467 		 * If an error occurred and the file was marked as dirty
2468 		 * before and we aren't forcibly invalidating pages, then
2469 		 * reset the RDIRTY flag.
2470 		 */
2471 		if (error && rdirty &&
2472 		    (flags & (B_INVAL | B_FORCE)) != (B_INVAL | B_FORCE)) {
2473 			mutex_enter(&rp->r_statelock);
2474 			rp->r_flags |= RDIRTY;
2475 			mutex_exit(&rp->r_statelock);
2476 		}
2477 	} else {
2478 		/*
2479 		 * Do a range from [off...off + len) looking for pages
2480 		 * to deal with.
2481 		 */
2482 		error = 0;
2483 #ifdef lint
2484 		io_len = 0;
2485 #endif
2486 		eoff = off + len;
2487 		mutex_enter(&rp->r_statelock);
2488 		for (io_off = off; io_off < eoff && io_off < rp->r_size;
2489 		    io_off += io_len) {
2490 			mutex_exit(&rp->r_statelock);
2491 			/*
2492 			 * If we are not invalidating, synchronously
2493 			 * freeing or writing pages use the routine
2494 			 * page_lookup_nowait() to prevent reclaiming
2495 			 * them from the free list.
2496 			 */
2497 			if ((flags & B_INVAL) || !(flags & B_ASYNC)) {
2498 				pp = page_lookup(vp, io_off,
2499 				    (flags & (B_INVAL | B_FREE)) ?
2500 				    SE_EXCL : SE_SHARED);
2501 			} else {
2502 				pp = page_lookup_nowait(vp, io_off,
2503 				    (flags & B_FREE) ? SE_EXCL : SE_SHARED);
2504 			}
2505 
2506 			if (pp == NULL || !pvn_getdirty(pp, flags))
2507 				io_len = PAGESIZE;
2508 			else {
2509 				err = (*rp->r_putapage)(vp, pp, &io_off,
2510 				    &io_len, flags, cr);
2511 				if (!error)
2512 					error = err;
2513 				/*
2514 				 * "io_off" and "io_len" are returned as
2515 				 * the range of pages we actually wrote.
2516 				 * This allows us to skip ahead more quickly
2517 				 * since several pages may've been dealt
2518 				 * with by this iteration of the loop.
2519 				 */
2520 			}
2521 			mutex_enter(&rp->r_statelock);
2522 		}
2523 		mutex_exit(&rp->r_statelock);
2524 	}
2525 
2526 	return (error);
2527 }
2528 
2529 void
nfs_invalidate_pages(vnode_t * vp,u_offset_t off,cred_t * cr)2530 nfs_invalidate_pages(vnode_t *vp, u_offset_t off, cred_t *cr)
2531 {
2532 	rnode_t *rp;
2533 
2534 	rp = VTOR(vp);
2535 	mutex_enter(&rp->r_statelock);
2536 	while (rp->r_flags & RTRUNCATE)
2537 		cv_wait(&rp->r_cv, &rp->r_statelock);
2538 	rp->r_flags |= RTRUNCATE;
2539 	if (off == (u_offset_t)0) {
2540 		rp->r_flags &= ~RDIRTY;
2541 		if (!(rp->r_flags & RSTALE))
2542 			rp->r_error = 0;
2543 	}
2544 	rp->r_truncaddr = off;
2545 	mutex_exit(&rp->r_statelock);
2546 	(void) pvn_vplist_dirty(vp, off, rp->r_putapage,
2547 	    B_INVAL | B_TRUNC, cr);
2548 	mutex_enter(&rp->r_statelock);
2549 	rp->r_flags &= ~RTRUNCATE;
2550 	cv_broadcast(&rp->r_cv);
2551 	mutex_exit(&rp->r_statelock);
2552 }
2553 
2554 static int nfs_write_error_to_cons_only = 0;
2555 #define	MSG(x)	(nfs_write_error_to_cons_only ? (x) : (x) + 1)
2556 
2557 /*
2558  * Print a file handle
2559  */
2560 void
nfs_printfhandle(nfs_fhandle * fhp)2561 nfs_printfhandle(nfs_fhandle *fhp)
2562 {
2563 	int *ip;
2564 	char *buf;
2565 	size_t bufsize;
2566 	char *cp;
2567 
2568 	/*
2569 	 * 13 == "(file handle:"
2570 	 * maximum of NFS_FHANDLE / sizeof (*ip) elements in fh_buf times
2571 	 *	1 == ' '
2572 	 *	8 == maximum strlen of "%x"
2573 	 * 3 == ")\n\0"
2574 	 */
2575 	bufsize = 13 + ((NFS_FHANDLE_LEN / sizeof (*ip)) * (1 + 8)) + 3;
2576 	buf = kmem_alloc(bufsize, KM_NOSLEEP);
2577 	if (buf == NULL)
2578 		return;
2579 
2580 	cp = buf;
2581 	(void) strcpy(cp, "(file handle:");
2582 	while (*cp != '\0')
2583 		cp++;
2584 	for (ip = (int *)fhp->fh_buf;
2585 	    ip < (int *)&fhp->fh_buf[fhp->fh_len];
2586 	    ip++) {
2587 		(void) sprintf(cp, " %x", *ip);
2588 		while (*cp != '\0')
2589 			cp++;
2590 	}
2591 	(void) strcpy(cp, ")\n");
2592 
2593 	zcmn_err(getzoneid(), CE_CONT, MSG("^%s"), buf);
2594 
2595 	kmem_free(buf, bufsize);
2596 }
2597 
2598 /*
2599  * Notify the system administrator that an NFS write error has
2600  * occurred.
2601  */
2602 
2603 /* seconds between ENOSPC/EDQUOT messages */
2604 clock_t nfs_write_error_interval = 5;
2605 
2606 void
nfs_write_error(vnode_t * vp,int error,cred_t * cr)2607 nfs_write_error(vnode_t *vp, int error, cred_t *cr)
2608 {
2609 	mntinfo_t *mi;
2610 	clock_t now;
2611 
2612 	mi = VTOMI(vp);
2613 	/*
2614 	 * In case of forced unmount or zone shutdown, do not print any
2615 	 * messages since it can flood the console with error messages.
2616 	 */
2617 	if (FS_OR_ZONE_GONE(mi->mi_vfsp))
2618 		return;
2619 
2620 	/*
2621 	 * No use in flooding the console with ENOSPC
2622 	 * messages from the same file system.
2623 	 */
2624 	now = ddi_get_lbolt();
2625 	if ((error != ENOSPC && error != EDQUOT) ||
2626 	    now - mi->mi_printftime > 0) {
2627 		zoneid_t zoneid = mi->mi_zone->zone_id;
2628 
2629 #ifdef DEBUG
2630 		nfs_perror(error, "NFS%ld write error on host %s: %m.\n",
2631 		    mi->mi_vers, VTOR(vp)->r_server->sv_hostname, NULL);
2632 #else
2633 		nfs_perror(error, "NFS write error on host %s: %m.\n",
2634 		    VTOR(vp)->r_server->sv_hostname, NULL);
2635 #endif
2636 		if (error == ENOSPC || error == EDQUOT) {
2637 			zcmn_err(zoneid, CE_CONT,
2638 			    MSG("^File: userid=%d, groupid=%d\n"),
2639 			    crgetuid(cr), crgetgid(cr));
2640 			if (crgetuid(CRED()) != crgetuid(cr) ||
2641 			    crgetgid(CRED()) != crgetgid(cr)) {
2642 				zcmn_err(zoneid, CE_CONT,
2643 				    MSG("^User: userid=%d, groupid=%d\n"),
2644 				    crgetuid(CRED()), crgetgid(CRED()));
2645 			}
2646 			mi->mi_printftime = now +
2647 			    nfs_write_error_interval * hz;
2648 		}
2649 		nfs_printfhandle(&VTOR(vp)->r_fh);
2650 #ifdef DEBUG
2651 		if (error == EACCES) {
2652 			zcmn_err(zoneid, CE_CONT,
2653 			    MSG("^nfs_bio: cred is%s kcred\n"),
2654 			    cr == kcred ? "" : " not");
2655 		}
2656 #endif
2657 	}
2658 }
2659 
2660 /* ARGSUSED */
2661 static void *
nfs_mi_init(zoneid_t zoneid)2662 nfs_mi_init(zoneid_t zoneid)
2663 {
2664 	struct mi_globals *mig;
2665 
2666 	mig = kmem_alloc(sizeof (*mig), KM_SLEEP);
2667 	mutex_init(&mig->mig_lock, NULL, MUTEX_DEFAULT, NULL);
2668 	list_create(&mig->mig_list, sizeof (mntinfo_t),
2669 	    offsetof(mntinfo_t, mi_zone_node));
2670 	mig->mig_destructor_called = B_FALSE;
2671 	return (mig);
2672 }
2673 
2674 /*
2675  * Callback routine to tell all NFS mounts in the zone to stop creating new
2676  * threads.  Existing threads should exit.
2677  */
2678 /* ARGSUSED */
2679 static void
nfs_mi_shutdown(zoneid_t zoneid,void * data)2680 nfs_mi_shutdown(zoneid_t zoneid, void *data)
2681 {
2682 	struct mi_globals *mig = data;
2683 	mntinfo_t *mi;
2684 
2685 	ASSERT(mig != NULL);
2686 again:
2687 	mutex_enter(&mig->mig_lock);
2688 	for (mi = list_head(&mig->mig_list); mi != NULL;
2689 	    mi = list_next(&mig->mig_list, mi)) {
2690 
2691 		/*
2692 		 * If we've done the shutdown work for this FS, skip.
2693 		 * Once we go off the end of the list, we're done.
2694 		 */
2695 		if (mi->mi_flags & MI_DEAD)
2696 			continue;
2697 
2698 		/*
2699 		 * We will do work, so not done.  Get a hold on the FS.
2700 		 */
2701 		VFS_HOLD(mi->mi_vfsp);
2702 
2703 		/*
2704 		 * purge the DNLC for this filesystem
2705 		 */
2706 		(void) dnlc_purge_vfsp(mi->mi_vfsp, 0);
2707 
2708 		mutex_enter(&mi->mi_async_lock);
2709 		/*
2710 		 * Tell existing async worker threads to exit.
2711 		 */
2712 		mi->mi_max_threads = 0;
2713 		NFS_WAKEALL_ASYNC_WORKERS(mi->mi_async_work_cv);
2714 		/*
2715 		 * Set MI_ASYNC_MGR_STOP so the async manager thread starts
2716 		 * getting ready to exit when it's done with its current work.
2717 		 * Also set MI_DEAD to note we've acted on this FS.
2718 		 */
2719 		mutex_enter(&mi->mi_lock);
2720 		mi->mi_flags |= (MI_ASYNC_MGR_STOP|MI_DEAD);
2721 		mutex_exit(&mi->mi_lock);
2722 		/*
2723 		 * Wake up the async manager thread.
2724 		 */
2725 		cv_broadcast(&mi->mi_async_reqs_cv);
2726 		mutex_exit(&mi->mi_async_lock);
2727 
2728 		/*
2729 		 * Drop lock and release FS, which may change list, then repeat.
2730 		 * We're done when every mi has been done or the list is empty.
2731 		 */
2732 		mutex_exit(&mig->mig_lock);
2733 		VFS_RELE(mi->mi_vfsp);
2734 		goto again;
2735 	}
2736 	mutex_exit(&mig->mig_lock);
2737 }
2738 
2739 static void
nfs_mi_free_globals(struct mi_globals * mig)2740 nfs_mi_free_globals(struct mi_globals *mig)
2741 {
2742 	list_destroy(&mig->mig_list);	/* makes sure the list is empty */
2743 	mutex_destroy(&mig->mig_lock);
2744 	kmem_free(mig, sizeof (*mig));
2745 
2746 }
2747 
2748 /* ARGSUSED */
2749 static void
nfs_mi_destroy(zoneid_t zoneid,void * data)2750 nfs_mi_destroy(zoneid_t zoneid, void *data)
2751 {
2752 	struct mi_globals *mig = data;
2753 
2754 	ASSERT(mig != NULL);
2755 	mutex_enter(&mig->mig_lock);
2756 	if (list_head(&mig->mig_list) != NULL) {
2757 		/* Still waiting for VFS_FREEVFS() */
2758 		mig->mig_destructor_called = B_TRUE;
2759 		mutex_exit(&mig->mig_lock);
2760 		return;
2761 	}
2762 	nfs_mi_free_globals(mig);
2763 }
2764 
2765 /*
2766  * Add an NFS mount to the per-zone list of NFS mounts.
2767  */
2768 void
nfs_mi_zonelist_add(mntinfo_t * mi)2769 nfs_mi_zonelist_add(mntinfo_t *mi)
2770 {
2771 	struct mi_globals *mig;
2772 
2773 	mig = zone_getspecific(mi_list_key, mi->mi_zone);
2774 	mutex_enter(&mig->mig_lock);
2775 	list_insert_head(&mig->mig_list, mi);
2776 	mutex_exit(&mig->mig_lock);
2777 }
2778 
2779 /*
2780  * Remove an NFS mount from the per-zone list of NFS mounts.
2781  */
2782 static void
nfs_mi_zonelist_remove(mntinfo_t * mi)2783 nfs_mi_zonelist_remove(mntinfo_t *mi)
2784 {
2785 	struct mi_globals *mig;
2786 
2787 	mig = zone_getspecific(mi_list_key, mi->mi_zone);
2788 	mutex_enter(&mig->mig_lock);
2789 	list_remove(&mig->mig_list, mi);
2790 	/*
2791 	 * We can be called asynchronously by VFS_FREEVFS() after the zone
2792 	 * shutdown/destroy callbacks have executed; if so, clean up the zone's
2793 	 * mi globals.
2794 	 */
2795 	if (list_head(&mig->mig_list) == NULL &&
2796 	    mig->mig_destructor_called == B_TRUE) {
2797 		nfs_mi_free_globals(mig);
2798 		return;
2799 	}
2800 	mutex_exit(&mig->mig_lock);
2801 }
2802 
2803 /*
2804  * NFS Client initialization routine.  This routine should only be called
2805  * once.  It performs the following tasks:
2806  *	- Initalize all global locks
2807  *	- Call sub-initialization routines (localize access to variables)
2808  */
2809 int
nfs_clntinit(void)2810 nfs_clntinit(void)
2811 {
2812 #ifdef DEBUG
2813 	static boolean_t nfs_clntup = B_FALSE;
2814 #endif
2815 	int error;
2816 
2817 #ifdef DEBUG
2818 	ASSERT(nfs_clntup == B_FALSE);
2819 #endif
2820 
2821 	error = nfs_subrinit();
2822 	if (error)
2823 		return (error);
2824 
2825 	error = nfs_vfsinit();
2826 	if (error) {
2827 		/*
2828 		 * Cleanup nfs_subrinit() work
2829 		 */
2830 		nfs_subrfini();
2831 		return (error);
2832 	}
2833 	zone_key_create(&mi_list_key, nfs_mi_init, nfs_mi_shutdown,
2834 	    nfs_mi_destroy);
2835 
2836 	nfs4_clnt_init();
2837 
2838 	nfscmd_init();
2839 
2840 #ifdef DEBUG
2841 	nfs_clntup = B_TRUE;
2842 #endif
2843 
2844 	return (0);
2845 }
2846 
2847 /*
2848  * This routine is only called if the NFS Client has been initialized but
2849  * the module failed to be installed. This routine will cleanup the previously
2850  * allocated/initialized work.
2851  */
2852 void
nfs_clntfini(void)2853 nfs_clntfini(void)
2854 {
2855 	(void) zone_key_delete(mi_list_key);
2856 	nfs_subrfini();
2857 	nfs_vfsfini();
2858 	nfs4_clnt_fini();
2859 	nfscmd_fini();
2860 }
2861 
2862 /*
2863  * nfs_lockrelease:
2864  *
2865  * Release any locks on the given vnode that are held by the current
2866  * process.
2867  */
2868 void
nfs_lockrelease(vnode_t * vp,int flag,offset_t offset,cred_t * cr)2869 nfs_lockrelease(vnode_t *vp, int flag, offset_t offset, cred_t *cr)
2870 {
2871 	flock64_t ld;
2872 	struct shrlock shr;
2873 	char *buf;
2874 	int remote_lock_possible;
2875 	int ret;
2876 
2877 	ASSERT((uintptr_t)vp > KERNELBASE);
2878 
2879 	/*
2880 	 * Generate an explicit unlock operation for the entire file.  As a
2881 	 * partial optimization, only generate the unlock if there is a
2882 	 * lock registered for the file.  We could check whether this
2883 	 * particular process has any locks on the file, but that would
2884 	 * require the local locking code to provide yet another query
2885 	 * routine.  Note that no explicit synchronization is needed here.
2886 	 * At worst, flk_has_remote_locks() will return a false positive,
2887 	 * in which case the unlock call wastes time but doesn't harm
2888 	 * correctness.
2889 	 *
2890 	 * In addition, an unlock request is generated if the process
2891 	 * is listed as possibly having a lock on the file because the
2892 	 * server and client lock managers may have gotten out of sync.
2893 	 * N.B. It is important to make sure nfs_remove_locking_id() is
2894 	 * called here even if flk_has_remote_locks(vp) reports true.
2895 	 * If it is not called and there is an entry on the process id
2896 	 * list, that entry will never get removed.
2897 	 */
2898 	remote_lock_possible = nfs_remove_locking_id(vp, RLMPL_PID,
2899 	    (char *)&(ttoproc(curthread)->p_pid), NULL, NULL);
2900 	if (remote_lock_possible || flk_has_remote_locks(vp)) {
2901 		ld.l_type = F_UNLCK;	/* set to unlock entire file */
2902 		ld.l_whence = 0;	/* unlock from start of file */
2903 		ld.l_start = 0;
2904 		ld.l_len = 0;		/* do entire file */
2905 		ret = VOP_FRLOCK(vp, F_SETLK, &ld, flag, offset, NULL, cr,
2906 		    NULL);
2907 
2908 		if (ret != 0) {
2909 			/*
2910 			 * If VOP_FRLOCK fails, make sure we unregister
2911 			 * local locks before we continue.
2912 			 */
2913 			ld.l_pid = ttoproc(curthread)->p_pid;
2914 			lm_register_lock_locally(vp, NULL, &ld, flag, offset);
2915 #ifdef DEBUG
2916 			nfs_perror(ret,
2917 			    "NFS lock release error on vp %p: %m.\n",
2918 			    (void *)vp, NULL);
2919 #endif
2920 		}
2921 
2922 		/*
2923 		 * The call to VOP_FRLOCK may put the pid back on the
2924 		 * list.  We need to remove it.
2925 		 */
2926 		(void) nfs_remove_locking_id(vp, RLMPL_PID,
2927 		    (char *)&(ttoproc(curthread)->p_pid), NULL, NULL);
2928 	}
2929 
2930 	/*
2931 	 * As long as the vp has a share matching our pid,
2932 	 * pluck it off and unshare it.  There are circumstances in
2933 	 * which the call to nfs_remove_locking_id() may put the
2934 	 * owner back on the list, in which case we simply do a
2935 	 * redundant and harmless unshare.
2936 	 */
2937 	buf = kmem_alloc(MAX_SHR_OWNER_LEN, KM_SLEEP);
2938 	while (nfs_remove_locking_id(vp, RLMPL_OWNER,
2939 	    (char *)NULL, buf, &shr.s_own_len)) {
2940 		shr.s_owner = buf;
2941 		shr.s_access = 0;
2942 		shr.s_deny = 0;
2943 		shr.s_sysid = 0;
2944 		shr.s_pid = curproc->p_pid;
2945 
2946 		ret = VOP_SHRLOCK(vp, F_UNSHARE, &shr, flag, cr, NULL);
2947 #ifdef DEBUG
2948 		if (ret != 0) {
2949 			nfs_perror(ret,
2950 			    "NFS share release error on vp %p: %m.\n",
2951 			    (void *)vp, NULL);
2952 		}
2953 #endif
2954 	}
2955 	kmem_free(buf, MAX_SHR_OWNER_LEN);
2956 }
2957 
2958 /*
2959  * nfs_lockcompletion:
2960  *
2961  * If the vnode has a lock that makes it unsafe to cache the file, mark it
2962  * as non cachable (set VNOCACHE bit).
2963  */
2964 
2965 void
nfs_lockcompletion(vnode_t * vp,int cmd)2966 nfs_lockcompletion(vnode_t *vp, int cmd)
2967 {
2968 #ifdef DEBUG
2969 	rnode_t *rp = VTOR(vp);
2970 
2971 	ASSERT(nfs_rw_lock_held(&rp->r_lkserlock, RW_WRITER));
2972 #endif
2973 
2974 	if (cmd == F_SETLK || cmd == F_SETLKW) {
2975 		if (!lm_safemap(vp)) {
2976 			mutex_enter(&vp->v_lock);
2977 			vp->v_flag |= VNOCACHE;
2978 			mutex_exit(&vp->v_lock);
2979 		} else {
2980 			mutex_enter(&vp->v_lock);
2981 			vp->v_flag &= ~VNOCACHE;
2982 			mutex_exit(&vp->v_lock);
2983 		}
2984 	}
2985 	/*
2986 	 * The cached attributes of the file are stale after acquiring
2987 	 * the lock on the file. They were updated when the file was
2988 	 * opened, but not updated when the lock was acquired. Therefore the
2989 	 * cached attributes are invalidated after the lock is obtained.
2990 	 */
2991 	PURGE_ATTRCACHE(vp);
2992 }
2993 
2994 /*
2995  * The lock manager holds state making it possible for the client
2996  * and server to be out of sync.  For example, if the response from
2997  * the server granting a lock request is lost, the server will think
2998  * the lock is granted and the client will think the lock is lost.
2999  * The client can tell when it is not positive if it is in sync with
3000  * the server.
3001  *
3002  * To deal with this, a list of processes for which the client is
3003  * not sure if the server holds a lock is attached to the rnode.
3004  * When such a process closes the rnode, an unlock request is sent
3005  * to the server to unlock the entire file.
3006  *
3007  * The list is kept as a singularly linked NULL terminated list.
3008  * Because it is only added to under extreme error conditions, the
3009  * list shouldn't get very big.  DEBUG kernels print a message if
3010  * the list gets bigger than nfs_lmpl_high_water.  This is arbitrarily
3011  * choosen to be 8, but can be tuned at runtime.
3012  */
3013 #ifdef DEBUG
3014 /* int nfs_lmpl_high_water = 8; */
3015 int nfs_lmpl_high_water = 128;
3016 int nfs_cnt_add_locking_id = 0;
3017 int nfs_len_add_locking_id = 0;
3018 #endif /* DEBUG */
3019 
3020 /*
3021  * Record that the nfs lock manager server may be holding a lock on
3022  * a vnode for a process.
3023  *
3024  * Because the nfs lock manager server holds state, it is possible
3025  * for the server to get out of sync with the client.  This routine is called
3026  * from the client when it is no longer sure if the server is in sync
3027  * with the client.  nfs_lockrelease() will then notice this and send
3028  * an unlock request when the file is closed
3029  */
3030 void
nfs_add_locking_id(vnode_t * vp,pid_t pid,int type,char * id,int len)3031 nfs_add_locking_id(vnode_t *vp, pid_t pid, int type, char *id, int len)
3032 {
3033 	rnode_t *rp;
3034 	lmpl_t *new;
3035 	lmpl_t *cur;
3036 	lmpl_t **lmplp;
3037 #ifdef DEBUG
3038 	int list_len = 1;
3039 #endif /* DEBUG */
3040 
3041 #ifdef DEBUG
3042 	++nfs_cnt_add_locking_id;
3043 #endif /* DEBUG */
3044 	/*
3045 	 * allocate new lmpl_t now so we don't sleep
3046 	 * later after grabbing mutexes
3047 	 */
3048 	ASSERT(len < MAX_SHR_OWNER_LEN);
3049 	new = kmem_alloc(sizeof (*new), KM_SLEEP);
3050 	new->lmpl_type = type;
3051 	new->lmpl_pid = pid;
3052 	new->lmpl_owner = kmem_alloc(len, KM_SLEEP);
3053 	bcopy(id, new->lmpl_owner, len);
3054 	new->lmpl_own_len = len;
3055 	new->lmpl_next = (lmpl_t *)NULL;
3056 #ifdef DEBUG
3057 	if (type == RLMPL_PID) {
3058 		ASSERT(len == sizeof (pid_t));
3059 		ASSERT(pid == *(pid_t *)new->lmpl_owner);
3060 	} else {
3061 		ASSERT(type == RLMPL_OWNER);
3062 	}
3063 #endif
3064 
3065 	rp = VTOR(vp);
3066 	mutex_enter(&rp->r_statelock);
3067 
3068 	/*
3069 	 * Add this id to the list for this rnode only if the
3070 	 * rnode is active and the id is not already there.
3071 	 */
3072 	ASSERT(rp->r_flags & RHASHED);
3073 	lmplp = &(rp->r_lmpl);
3074 	for (cur = rp->r_lmpl; cur != (lmpl_t *)NULL; cur = cur->lmpl_next) {
3075 		if (cur->lmpl_pid == pid &&
3076 		    cur->lmpl_type == type &&
3077 		    cur->lmpl_own_len == len &&
3078 		    bcmp(cur->lmpl_owner, new->lmpl_owner, len) == 0) {
3079 			kmem_free(new->lmpl_owner, len);
3080 			kmem_free(new, sizeof (*new));
3081 			break;
3082 		}
3083 		lmplp = &cur->lmpl_next;
3084 #ifdef DEBUG
3085 		++list_len;
3086 #endif /* DEBUG */
3087 	}
3088 	if (cur == (lmpl_t *)NULL) {
3089 		*lmplp = new;
3090 #ifdef DEBUG
3091 		if (list_len > nfs_len_add_locking_id) {
3092 			nfs_len_add_locking_id = list_len;
3093 		}
3094 		if (list_len > nfs_lmpl_high_water) {
3095 			cmn_err(CE_WARN, "nfs_add_locking_id: long list "
3096 			    "vp=%p is %d", (void *)vp, list_len);
3097 		}
3098 #endif /* DEBUG */
3099 	}
3100 
3101 #ifdef DEBUG
3102 	if (share_debug) {
3103 		int nitems = 0;
3104 		int npids = 0;
3105 		int nowners = 0;
3106 
3107 		/*
3108 		 * Count the number of things left on r_lmpl after the remove.
3109 		 */
3110 		for (cur = rp->r_lmpl; cur != (lmpl_t *)NULL;
3111 		    cur = cur->lmpl_next) {
3112 			nitems++;
3113 			if (cur->lmpl_type == RLMPL_PID) {
3114 				npids++;
3115 			} else if (cur->lmpl_type == RLMPL_OWNER) {
3116 				nowners++;
3117 			} else {
3118 				cmn_err(CE_PANIC, "nfs_add_locking_id: "
3119 				    "unrecognized lmpl_type %d",
3120 				    cur->lmpl_type);
3121 			}
3122 		}
3123 
3124 		cmn_err(CE_CONT, "nfs_add_locking_id(%s): %d PIDs + %d "
3125 		    "OWNs = %d items left on r_lmpl\n",
3126 		    (type == RLMPL_PID) ? "P" : "O", npids, nowners, nitems);
3127 	}
3128 #endif
3129 
3130 	mutex_exit(&rp->r_statelock);
3131 }
3132 
3133 /*
3134  * Remove an id from the lock manager id list.
3135  *
3136  * If the id is not in the list return 0.  If it was found and
3137  * removed, return 1.
3138  */
3139 static int
nfs_remove_locking_id(vnode_t * vp,int type,char * id,char * rid,int * rlen)3140 nfs_remove_locking_id(vnode_t *vp, int type, char *id, char *rid, int *rlen)
3141 {
3142 	lmpl_t *cur;
3143 	lmpl_t **lmplp;
3144 	rnode_t *rp;
3145 	int rv = 0;
3146 
3147 	ASSERT(type == RLMPL_PID || type == RLMPL_OWNER);
3148 
3149 	rp = VTOR(vp);
3150 
3151 	mutex_enter(&rp->r_statelock);
3152 	ASSERT(rp->r_flags & RHASHED);
3153 	lmplp = &(rp->r_lmpl);
3154 
3155 	/*
3156 	 * Search through the list and remove the entry for this id
3157 	 * if it is there.  The special case id == NULL allows removal
3158 	 * of the first share on the r_lmpl list belonging to the
3159 	 * current process (if any), without regard to further details
3160 	 * of its identity.
3161 	 */
3162 	for (cur = rp->r_lmpl; cur != (lmpl_t *)NULL; cur = cur->lmpl_next) {
3163 		if (cur->lmpl_type == type &&
3164 		    cur->lmpl_pid == curproc->p_pid &&
3165 		    (id == (char *)NULL ||
3166 		    bcmp(cur->lmpl_owner, id, cur->lmpl_own_len) == 0)) {
3167 			*lmplp = cur->lmpl_next;
3168 			ASSERT(cur->lmpl_own_len < MAX_SHR_OWNER_LEN);
3169 			if (rid != NULL) {
3170 				bcopy(cur->lmpl_owner, rid, cur->lmpl_own_len);
3171 				*rlen = cur->lmpl_own_len;
3172 			}
3173 			kmem_free(cur->lmpl_owner, cur->lmpl_own_len);
3174 			kmem_free(cur, sizeof (*cur));
3175 			rv = 1;
3176 			break;
3177 		}
3178 		lmplp = &cur->lmpl_next;
3179 	}
3180 
3181 #ifdef DEBUG
3182 	if (share_debug) {
3183 		int nitems = 0;
3184 		int npids = 0;
3185 		int nowners = 0;
3186 
3187 		/*
3188 		 * Count the number of things left on r_lmpl after the remove.
3189 		 */
3190 		for (cur = rp->r_lmpl; cur != (lmpl_t *)NULL;
3191 		    cur = cur->lmpl_next) {
3192 			nitems++;
3193 			if (cur->lmpl_type == RLMPL_PID) {
3194 				npids++;
3195 			} else if (cur->lmpl_type == RLMPL_OWNER) {
3196 				nowners++;
3197 			} else {
3198 				cmn_err(CE_PANIC,
3199 				    "nrli: unrecognized lmpl_type %d",
3200 				    cur->lmpl_type);
3201 			}
3202 		}
3203 
3204 		cmn_err(CE_CONT,
3205 		"nrli(%s): %d PIDs + %d OWNs = %d items left on r_lmpl\n",
3206 		    (type == RLMPL_PID) ? "P" : "O",
3207 		    npids,
3208 		    nowners,
3209 		    nitems);
3210 	}
3211 #endif
3212 
3213 	mutex_exit(&rp->r_statelock);
3214 	return (rv);
3215 }
3216 
3217 void
nfs_free_mi(mntinfo_t * mi)3218 nfs_free_mi(mntinfo_t *mi)
3219 {
3220 	ASSERT(mi->mi_flags & MI_ASYNC_MGR_STOP);
3221 	ASSERT(mi->mi_manager_thread == NULL);
3222 	ASSERT(mi->mi_threads[NFS_ASYNC_QUEUE] == 0 &&
3223 	    mi->mi_threads[NFS_ASYNC_PGOPS_QUEUE] == 0);
3224 
3225 	/*
3226 	 * Remove the node from the global list before we start tearing it down.
3227 	 */
3228 	nfs_mi_zonelist_remove(mi);
3229 	if (mi->mi_klmconfig) {
3230 		lm_free_config(mi->mi_klmconfig);
3231 		kmem_free(mi->mi_klmconfig, sizeof (struct knetconfig));
3232 	}
3233 	mutex_destroy(&mi->mi_lock);
3234 	mutex_destroy(&mi->mi_remap_lock);
3235 	mutex_destroy(&mi->mi_async_lock);
3236 	mutex_destroy(&mi->mi_rnodes_lock);
3237 	cv_destroy(&mi->mi_failover_cv);
3238 	cv_destroy(&mi->mi_async_work_cv[NFS_ASYNC_QUEUE]);
3239 	cv_destroy(&mi->mi_async_work_cv[NFS_ASYNC_PGOPS_QUEUE]);
3240 	cv_destroy(&mi->mi_async_reqs_cv);
3241 	cv_destroy(&mi->mi_async_cv);
3242 	list_destroy(&mi->mi_rnodes);
3243 	zone_rele_ref(&mi->mi_zone_ref, ZONE_REF_NFS);
3244 	kmem_free(mi, sizeof (*mi));
3245 }
3246 
3247 static int
mnt_kstat_update(kstat_t * ksp,int rw)3248 mnt_kstat_update(kstat_t *ksp, int rw)
3249 {
3250 	mntinfo_t *mi;
3251 	struct mntinfo_kstat *mik;
3252 	vfs_t *vfsp;
3253 	int i;
3254 
3255 	/* this is a read-only kstat. Bail out on a write */
3256 	if (rw == KSTAT_WRITE)
3257 		return (EACCES);
3258 
3259 	/*
3260 	 * We don't want to wait here as kstat_chain_lock could be held by
3261 	 * dounmount(). dounmount() takes vfs_reflock before the chain lock
3262 	 * and thus could lead to a deadlock.
3263 	 */
3264 	vfsp = (struct vfs *)ksp->ks_private;
3265 
3266 
3267 	mi = VFTOMI(vfsp);
3268 
3269 	mik = (struct mntinfo_kstat *)ksp->ks_data;
3270 
3271 	(void) strcpy(mik->mik_proto, mi->mi_curr_serv->sv_knconf->knc_proto);
3272 	mik->mik_vers = (uint32_t)mi->mi_vers;
3273 	mik->mik_flags = mi->mi_flags;
3274 	mik->mik_secmod = mi->mi_curr_serv->sv_secdata->secmod;
3275 	mik->mik_curread = (uint32_t)mi->mi_curread;
3276 	mik->mik_curwrite = (uint32_t)mi->mi_curwrite;
3277 	mik->mik_retrans = mi->mi_retrans;
3278 	mik->mik_timeo = mi->mi_timeo;
3279 	mik->mik_acregmin = HR2SEC(mi->mi_acregmin);
3280 	mik->mik_acregmax = HR2SEC(mi->mi_acregmax);
3281 	mik->mik_acdirmin = HR2SEC(mi->mi_acdirmin);
3282 	mik->mik_acdirmax = HR2SEC(mi->mi_acdirmax);
3283 	for (i = 0; i < NFS_CALLTYPES + 1; i++) {
3284 		mik->mik_timers[i].srtt = (uint32_t)mi->mi_timers[i].rt_srtt;
3285 		mik->mik_timers[i].deviate =
3286 		    (uint32_t)mi->mi_timers[i].rt_deviate;
3287 		mik->mik_timers[i].rtxcur =
3288 		    (uint32_t)mi->mi_timers[i].rt_rtxcur;
3289 	}
3290 	mik->mik_noresponse = (uint32_t)mi->mi_noresponse;
3291 	mik->mik_failover = (uint32_t)mi->mi_failover;
3292 	mik->mik_remap = (uint32_t)mi->mi_remap;
3293 	(void) strcpy(mik->mik_curserver, mi->mi_curr_serv->sv_hostname);
3294 
3295 	return (0);
3296 }
3297 
3298 void
nfs_mnt_kstat_init(struct vfs * vfsp)3299 nfs_mnt_kstat_init(struct vfs *vfsp)
3300 {
3301 	mntinfo_t *mi = VFTOMI(vfsp);
3302 
3303 	/*
3304 	 * Create the version specific kstats.
3305 	 *
3306 	 * PSARC 2001/697 Contract Private Interface
3307 	 * All nfs kstats are under SunMC contract
3308 	 * Please refer to the PSARC listed above and contact
3309 	 * SunMC before making any changes!
3310 	 *
3311 	 * Changes must be reviewed by Solaris File Sharing
3312 	 * Changes must be communicated to contract-2001-697@sun.com
3313 	 *
3314 	 */
3315 
3316 	mi->mi_io_kstats = kstat_create_zone("nfs", getminor(vfsp->vfs_dev),
3317 	    NULL, "nfs", KSTAT_TYPE_IO, 1, 0, mi->mi_zone->zone_id);
3318 	if (mi->mi_io_kstats) {
3319 		if (mi->mi_zone->zone_id != GLOBAL_ZONEID)
3320 			kstat_zone_add(mi->mi_io_kstats, GLOBAL_ZONEID);
3321 		mi->mi_io_kstats->ks_lock = &mi->mi_lock;
3322 		kstat_install(mi->mi_io_kstats);
3323 	}
3324 
3325 	if ((mi->mi_ro_kstats = kstat_create_zone("nfs",
3326 	    getminor(vfsp->vfs_dev), "mntinfo", "misc", KSTAT_TYPE_RAW,
3327 	    sizeof (struct mntinfo_kstat), 0, mi->mi_zone->zone_id)) != NULL) {
3328 		if (mi->mi_zone->zone_id != GLOBAL_ZONEID)
3329 			kstat_zone_add(mi->mi_ro_kstats, GLOBAL_ZONEID);
3330 		mi->mi_ro_kstats->ks_update = mnt_kstat_update;
3331 		mi->mi_ro_kstats->ks_private = (void *)vfsp;
3332 		kstat_install(mi->mi_ro_kstats);
3333 	}
3334 }
3335 
3336 nfs_delmapcall_t *
nfs_init_delmapcall()3337 nfs_init_delmapcall()
3338 {
3339 	nfs_delmapcall_t	*delmap_call;
3340 
3341 	delmap_call = kmem_alloc(sizeof (nfs_delmapcall_t), KM_SLEEP);
3342 	delmap_call->call_id = curthread;
3343 	delmap_call->error = 0;
3344 
3345 	return (delmap_call);
3346 }
3347 
3348 void
nfs_free_delmapcall(nfs_delmapcall_t * delmap_call)3349 nfs_free_delmapcall(nfs_delmapcall_t *delmap_call)
3350 {
3351 	kmem_free(delmap_call, sizeof (nfs_delmapcall_t));
3352 }
3353 
3354 /*
3355  * Searches for the current delmap caller (based on curthread) in the list of
3356  * callers.  If it is found, we remove it and free the delmap caller.
3357  * Returns:
3358  *	0 if the caller wasn't found
3359  *	1 if the caller was found, removed and freed.  *errp is set to what
3360  *	the result of the delmap was.
3361  */
3362 int
nfs_find_and_delete_delmapcall(rnode_t * rp,int * errp)3363 nfs_find_and_delete_delmapcall(rnode_t *rp, int *errp)
3364 {
3365 	nfs_delmapcall_t	*delmap_call;
3366 
3367 	/*
3368 	 * If the list doesn't exist yet, we create it and return
3369 	 * that the caller wasn't found.  No list = no callers.
3370 	 */
3371 	mutex_enter(&rp->r_statelock);
3372 	if (!(rp->r_flags & RDELMAPLIST)) {
3373 		/* The list does not exist */
3374 		list_create(&rp->r_indelmap, sizeof (nfs_delmapcall_t),
3375 		    offsetof(nfs_delmapcall_t, call_node));
3376 		rp->r_flags |= RDELMAPLIST;
3377 		mutex_exit(&rp->r_statelock);
3378 		return (0);
3379 	} else {
3380 		/* The list exists so search it */
3381 		for (delmap_call = list_head(&rp->r_indelmap);
3382 		    delmap_call != NULL;
3383 		    delmap_call = list_next(&rp->r_indelmap, delmap_call)) {
3384 			if (delmap_call->call_id == curthread) {
3385 				/* current caller is in the list */
3386 				*errp = delmap_call->error;
3387 				list_remove(&rp->r_indelmap, delmap_call);
3388 				mutex_exit(&rp->r_statelock);
3389 				nfs_free_delmapcall(delmap_call);
3390 				return (1);
3391 			}
3392 		}
3393 	}
3394 	mutex_exit(&rp->r_statelock);
3395 	return (0);
3396 }
3397