xref: /freebsd/sys/fs/nfsserver/nfs_fha_new.c (revision 8b04c1cbfc1cb71a1ce53b3a7855f1d45866fcfb)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008 Isilon Inc http://www.isilon.com/
5  * Copyright (c) 2013 Spectra Logic Corporation
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/types.h>
33 #include <sys/mbuf.h>
34 #include <sys/sbuf.h>
35 
36 #include <fs/nfs/nfsport.h>
37 #include <fs/nfsserver/nfs_fha_new.h>
38 
39 #include <rpc/rpc.h>
40 
41 static MALLOC_DEFINE(M_NFS_FHA, "NFS FHA", "NFS FHA");
42 
43 static void		fhanew_init(void *foo);
44 static void		fhanew_uninit(void *foo);
45 static rpcproc_t	fhanew_get_procnum(rpcproc_t procnum);
46 static int		fhanew_get_fh(uint64_t *fh, int v3, struct mbuf **md,
47 			    caddr_t *dpos);
48 static int		fhanew_is_read(rpcproc_t procnum);
49 static int		fhanew_is_write(rpcproc_t procnum);
50 static int		fhanew_get_offset(struct mbuf **md, caddr_t *dpos,
51 			    int v3, struct fha_info *info);
52 static int		fhanew_no_offset(rpcproc_t procnum);
53 static void		fhanew_set_locktype(rpcproc_t procnum,
54 			    struct fha_info *info);
55 static int		fhenew_stats_sysctl(SYSCTL_HANDLER_ARGS);
56 static void		fha_extract_info(struct svc_req *req,
57 			    struct fha_info *i);
58 
59 NFSD_VNET_DEFINE_STATIC(struct fha_params *, fhanew_softc);
60 
61 SYSCTL_DECL(_vfs_nfsd);
62 
63 extern int newnfs_nfsv3_procid[];
64 
65 VNET_SYSINIT(nfs_fhanew, SI_SUB_VNET_DONE, SI_ORDER_ANY, fhanew_init, NULL);
66 VNET_SYSUNINIT(nfs_fhanew, SI_SUB_VNET_DONE, SI_ORDER_ANY, fhanew_uninit, NULL);
67 
68 static void
69 fhanew_init(void *foo)
70 {
71 	struct fha_params *softc;
72 	int i;
73 
74 	NFSD_VNET(fhanew_softc) = malloc(sizeof(struct fha_params), M_TEMP,
75 	    M_WAITOK | M_ZERO);
76 	softc = NFSD_VNET(fhanew_softc);
77 
78 	snprintf(softc->server_name, sizeof(softc->server_name),
79 	    FHANEW_SERVER_NAME);
80 
81 	/*
82 	 * Initialize the sysctl context list for the fha module.
83 	 */
84 	sysctl_ctx_init(&softc->sysctl_ctx);
85 	if (IS_DEFAULT_VNET(curvnet)) {
86 		softc->sysctl_tree = SYSCTL_ADD_NODE(&softc->sysctl_ctx,
87 		    SYSCTL_STATIC_CHILDREN(_vfs_nfsd), OID_AUTO, "fha",
88 		    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "NFS File Handle Affinity (FHA)");
89 		if (softc->sysctl_tree == NULL) {
90 			printf("%s: unable to allocate sysctl tree\n", __func__);
91 			return;
92 		}
93 	}
94 
95 	for (i = 0; i < FHA_HASH_SIZE; i++)
96 		mtx_init(&softc->fha_hash[i].mtx, "fhalock", NULL, MTX_DEF);
97 
98 	/*
99 	 * Set the default tuning parameters.
100 	 */
101 	softc->ctls.enable = FHA_DEF_ENABLE;
102 	softc->ctls.read = FHA_DEF_READ;
103 	softc->ctls.write = FHA_DEF_WRITE;
104 	softc->ctls.bin_shift = FHA_DEF_BIN_SHIFT;
105 	softc->ctls.max_nfsds_per_fh = FHA_DEF_MAX_NFSDS_PER_FH;
106 	softc->ctls.max_reqs_per_nfsd = FHA_DEF_MAX_REQS_PER_NFSD;
107 
108 	/*
109 	 * Add sysctls so the user can change the tuning parameters.
110 	 */
111 	if (IS_DEFAULT_VNET(curvnet)) {
112 		SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
113 		    OID_AUTO, "enable", CTLFLAG_RWTUN,
114 		    &softc->ctls.enable, 0, "Enable NFS File Handle Affinity (FHA)");
115 
116 		SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
117 		    OID_AUTO, "read", CTLFLAG_RWTUN,
118 		    &softc->ctls.read, 0, "Enable NFS FHA read locality");
119 
120 		SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
121 		    OID_AUTO, "write", CTLFLAG_RWTUN,
122 		    &softc->ctls.write, 0, "Enable NFS FHA write locality");
123 
124 		SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
125 		    OID_AUTO, "bin_shift", CTLFLAG_RWTUN,
126 		    &softc->ctls.bin_shift, 0,
127 		    "Maximum locality distance 2^(bin_shift) bytes");
128 
129 		SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
130 		    OID_AUTO, "max_nfsds_per_fh", CTLFLAG_RWTUN,
131 		    &softc->ctls.max_nfsds_per_fh, 0, "Maximum nfsd threads that "
132 		    "should be working on requests for the same file handle");
133 
134 		SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
135 		    OID_AUTO, "max_reqs_per_nfsd", CTLFLAG_RWTUN,
136 		    &softc->ctls.max_reqs_per_nfsd, 0, "Maximum requests that "
137 		    "single nfsd thread should be working on at any time");
138 
139 		SYSCTL_ADD_OID(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
140 		    OID_AUTO, "fhe_stats", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
141 		    0, 0, fhenew_stats_sysctl, "A", "");
142 	}
143 }
144 
145 static void
146 fhanew_uninit(void *foo)
147 {
148 	struct fha_params *softc;
149 	int i;
150 
151 	softc = NFSD_VNET(fhanew_softc);
152 
153 	sysctl_ctx_free(&softc->sysctl_ctx);
154 	for (i = 0; i < FHA_HASH_SIZE; i++)
155 		mtx_destroy(&softc->fha_hash[i].mtx);
156 	free(softc, M_TEMP);
157 }
158 
159 static rpcproc_t
160 fhanew_get_procnum(rpcproc_t procnum)
161 {
162 	if (procnum > NFSV2PROC_STATFS)
163 		return (-1);
164 
165 	return (newnfs_nfsv3_procid[procnum]);
166 }
167 
168 static int
169 fhanew_get_fh(uint64_t *fh, int v3, struct mbuf **md, caddr_t *dpos)
170 {
171 	struct nfsrv_descript lnd, *nd;
172 	uint32_t *tl;
173 	uint8_t *buf;
174 	uint64_t t;
175 	int error, len, i;
176 
177 	error = 0;
178 	len = 0;
179 	nd = &lnd;
180 
181 	nd->nd_md = *md;
182 	nd->nd_dpos = *dpos;
183 
184 	if (v3) {
185 		NFSM_DISSECT_NONBLOCK(tl, uint32_t *, NFSX_UNSIGNED);
186 		if ((len = fxdr_unsigned(int, *tl)) <= 0 || len > NFSX_FHMAX) {
187 			error = EBADRPC;
188 			goto nfsmout;
189 		}
190 	} else {
191 		len = NFSX_V2FH;
192 	}
193 
194 	t = 0;
195 	if (len != 0) {
196 		NFSM_DISSECT_NONBLOCK(buf, uint8_t *, len);
197 		for (i = 0; i < len; i++)
198 			t ^= ((uint64_t)buf[i] << (i & 7) * 8);
199 	}
200 	*fh = t;
201 
202 nfsmout:
203 	*md = nd->nd_md;
204 	*dpos = nd->nd_dpos;
205 
206 	return (error);
207 }
208 
209 static int
210 fhanew_is_read(rpcproc_t procnum)
211 {
212 	if (procnum == NFSPROC_READ)
213 		return (1);
214 	else
215 		return (0);
216 }
217 
218 static int
219 fhanew_is_write(rpcproc_t procnum)
220 {
221 	if (procnum == NFSPROC_WRITE)
222 		return (1);
223 	else
224 		return (0);
225 }
226 
227 static int
228 fhanew_get_offset(struct mbuf **md, caddr_t *dpos, int v3,
229     struct fha_info *info)
230 {
231 	struct nfsrv_descript lnd, *nd;
232 	uint32_t *tl;
233 	int error;
234 
235 	error = 0;
236 
237 	nd = &lnd;
238 	nd->nd_md = *md;
239 	nd->nd_dpos = *dpos;
240 
241 	if (v3) {
242 		NFSM_DISSECT_NONBLOCK(tl, uint32_t *, 2 * NFSX_UNSIGNED);
243 		info->offset = fxdr_hyper(tl);
244 	} else {
245 		NFSM_DISSECT_NONBLOCK(tl, uint32_t *, NFSX_UNSIGNED);
246 		info->offset = fxdr_unsigned(uint32_t, *tl);
247 	}
248 
249 nfsmout:
250 	*md = nd->nd_md;
251 	*dpos = nd->nd_dpos;
252 
253 	return (error);
254 }
255 
256 static int
257 fhanew_no_offset(rpcproc_t procnum)
258 {
259 	if (procnum == NFSPROC_FSSTAT ||
260 	    procnum == NFSPROC_FSINFO ||
261 	    procnum == NFSPROC_PATHCONF ||
262 	    procnum == NFSPROC_NOOP ||
263 	    procnum == NFSPROC_NULL)
264 		return (1);
265 	else
266 		return (0);
267 }
268 
269 static void
270 fhanew_set_locktype(rpcproc_t procnum, struct fha_info *info)
271 {
272 	switch (procnum) {
273 	case NFSPROC_NULL:
274 	case NFSPROC_GETATTR:
275 	case NFSPROC_LOOKUP:
276 	case NFSPROC_ACCESS:
277 	case NFSPROC_READLINK:
278 	case NFSPROC_READ:
279 	case NFSPROC_READDIR:
280 	case NFSPROC_READDIRPLUS:
281 	case NFSPROC_WRITE:
282 		info->locktype = LK_SHARED;
283 		break;
284 	case NFSPROC_SETATTR:
285 	case NFSPROC_CREATE:
286 	case NFSPROC_MKDIR:
287 	case NFSPROC_SYMLINK:
288 	case NFSPROC_MKNOD:
289 	case NFSPROC_REMOVE:
290 	case NFSPROC_RMDIR:
291 	case NFSPROC_RENAME:
292 	case NFSPROC_LINK:
293 	case NFSPROC_FSSTAT:
294 	case NFSPROC_FSINFO:
295 	case NFSPROC_PATHCONF:
296 	case NFSPROC_COMMIT:
297 	case NFSPROC_NOOP:
298 		info->locktype = LK_EXCLUSIVE;
299 		break;
300 	}
301 }
302 
303 /*
304  * This just specifies that offsets should obey affinity when within
305  * the same 1Mbyte (1<<20) chunk for the file (reads only for now).
306  */
307 static void
308 fha_extract_info(struct svc_req *req, struct fha_info *i)
309 {
310 	struct mbuf *md;
311 	caddr_t dpos;
312 	static u_int64_t random_fh = 0;
313 	int error;
314 	int v3 = (req->rq_vers == 3);
315 	rpcproc_t procnum;
316 
317 	/*
318 	 * We start off with a random fh.  If we get a reasonable
319 	 * procnum, we set the fh.  If there's a concept of offset
320 	 * that we're interested in, we set that.
321 	 */
322 	i->fh = ++random_fh;
323 	i->offset = 0;
324 	i->locktype = LK_EXCLUSIVE;
325 	i->read = i->write = 0;
326 
327 	/*
328 	 * Extract the procnum and convert to v3 form if necessary,
329 	 * taking care to deal with out-of-range procnums.  Caller will
330 	 * ensure that rq_vers is either 2 or 3.
331 	 */
332 	procnum = req->rq_proc;
333 	if (!v3) {
334 		rpcproc_t tmp_procnum;
335 
336 		tmp_procnum = fhanew_get_procnum(procnum);
337 		if (tmp_procnum == -1)
338 			goto out;
339 		procnum = tmp_procnum;
340 	}
341 
342 	/*
343 	 * We do affinity for most.  However, we divide a realm of affinity
344 	 * by file offset so as to allow for concurrent random access.  We
345 	 * only do this for reads today, but this may change when IFS supports
346 	 * efficient concurrent writes.
347 	 */
348 	if (fhanew_no_offset(procnum))
349 		goto out;
350 
351 	i->read = fhanew_is_read(procnum);
352 	i->write = fhanew_is_write(procnum);
353 
354 	error = newnfs_realign(&req->rq_args, M_NOWAIT);
355 	if (error)
356 		goto out;
357 	md = req->rq_args;
358 	dpos = mtod(md, caddr_t);
359 
360 	/* Grab the filehandle. */
361 	error = fhanew_get_fh(&i->fh, v3, &md, &dpos);
362 	if (error)
363 		goto out;
364 
365 	/* Content ourselves with zero offset for all but reads. */
366 	if (i->read || i->write)
367 		fhanew_get_offset(&md, &dpos, v3, i);
368 
369 out:
370 	fhanew_set_locktype(procnum, i);
371 }
372 
373 static struct fha_hash_entry *
374 fha_hash_entry_new(u_int64_t fh)
375 {
376 	struct fha_hash_entry *e;
377 
378 	e = malloc(sizeof(*e), M_NFS_FHA, M_WAITOK);
379 	e->fh = fh;
380 	e->num_rw = 0;
381 	e->num_exclusive = 0;
382 	e->num_threads = 0;
383 	LIST_INIT(&e->threads);
384 
385 	return (e);
386 }
387 
388 static void
389 fha_hash_entry_destroy(struct fha_hash_entry *e)
390 {
391 
392 	mtx_assert(e->mtx, MA_OWNED);
393 	KASSERT(e->num_rw == 0,
394 	    ("%d reqs on destroyed fhe %p", e->num_rw, e));
395 	KASSERT(e->num_exclusive == 0,
396 	    ("%d exclusive reqs on destroyed fhe %p", e->num_exclusive, e));
397 	KASSERT(e->num_threads == 0,
398 	    ("%d threads on destroyed fhe %p", e->num_threads, e));
399 	free(e, M_NFS_FHA);
400 }
401 
402 static void
403 fha_hash_entry_remove(struct fha_hash_entry *e)
404 {
405 
406 	mtx_assert(e->mtx, MA_OWNED);
407 	LIST_REMOVE(e, link);
408 	fha_hash_entry_destroy(e);
409 }
410 
411 static struct fha_hash_entry *
412 fha_hash_entry_lookup(struct fha_params *softc, u_int64_t fh)
413 {
414 	struct fha_hash_slot *fhs;
415 	struct fha_hash_entry *fhe, *new_fhe;
416 
417 	fhs = &softc->fha_hash[fh % FHA_HASH_SIZE];
418 	new_fhe = fha_hash_entry_new(fh);
419 	new_fhe->mtx = &fhs->mtx;
420 	mtx_lock(&fhs->mtx);
421 	LIST_FOREACH(fhe, &fhs->list, link)
422 		if (fhe->fh == fh)
423 			break;
424 	if (!fhe) {
425 		fhe = new_fhe;
426 		LIST_INSERT_HEAD(&fhs->list, fhe, link);
427 	} else
428 		fha_hash_entry_destroy(new_fhe);
429 	return (fhe);
430 }
431 
432 static void
433 fha_hash_entry_add_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread)
434 {
435 
436 	mtx_assert(fhe->mtx, MA_OWNED);
437 	thread->st_p2 = 0;
438 	LIST_INSERT_HEAD(&fhe->threads, thread, st_alink);
439 	fhe->num_threads++;
440 }
441 
442 static void
443 fha_hash_entry_remove_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread)
444 {
445 
446 	mtx_assert(fhe->mtx, MA_OWNED);
447 	KASSERT(thread->st_p2 == 0,
448 	    ("%d reqs on removed thread %p", thread->st_p2, thread));
449 	LIST_REMOVE(thread, st_alink);
450 	fhe->num_threads--;
451 }
452 
453 /*
454  * Account for an ongoing operation associated with this file.
455  */
456 static void
457 fha_hash_entry_add_op(struct fha_hash_entry *fhe, int locktype, int count)
458 {
459 
460 	mtx_assert(fhe->mtx, MA_OWNED);
461 	if (LK_EXCLUSIVE == locktype)
462 		fhe->num_exclusive += count;
463 	else
464 		fhe->num_rw += count;
465 }
466 
467 /*
468  * Get the service thread currently associated with the fhe that is
469  * appropriate to handle this operation.
470  */
471 static SVCTHREAD *
472 fha_hash_entry_choose_thread(struct fha_params *softc,
473     struct fha_hash_entry *fhe, struct fha_info *i, SVCTHREAD *this_thread)
474 {
475 	SVCTHREAD *thread, *min_thread = NULL;
476 	int req_count, min_count = 0;
477 	off_t offset1, offset2;
478 
479 	LIST_FOREACH(thread, &fhe->threads, st_alink) {
480 		req_count = thread->st_p2;
481 
482 		/* If there are any writes in progress, use the first thread. */
483 		if (fhe->num_exclusive) {
484 #if 0
485 			ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
486 			    "fha: %p(%d)w", thread, req_count);
487 #endif
488 			return (thread);
489 		}
490 
491 		/* Check whether we should consider locality. */
492 		if ((i->read && !softc->ctls.read) ||
493 		    (i->write && !softc->ctls.write))
494 			goto noloc;
495 
496 		/*
497 		 * Check for locality, making sure that we won't
498 		 * exceed our per-thread load limit in the process.
499 		 */
500 		offset1 = i->offset;
501 		offset2 = thread->st_p3;
502 
503 		if (((offset1 >= offset2)
504 		  && ((offset1 - offset2) < (1 << softc->ctls.bin_shift)))
505 		 || ((offset2 > offset1)
506 		  && ((offset2 - offset1) < (1 << softc->ctls.bin_shift)))) {
507 			if ((softc->ctls.max_reqs_per_nfsd == 0) ||
508 			    (req_count < softc->ctls.max_reqs_per_nfsd)) {
509 #if 0
510 				ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
511 				    "fha: %p(%d)r", thread, req_count);
512 #endif
513 				return (thread);
514 			}
515 		}
516 
517 noloc:
518 		/*
519 		 * We don't have a locality match, so skip this thread,
520 		 * but keep track of the most attractive thread in case
521 		 * we need to come back to it later.
522 		 */
523 #if 0
524 		ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
525 		    "fha: %p(%d)s off1 %llu off2 %llu", thread,
526 		    req_count, offset1, offset2);
527 #endif
528 		if ((min_thread == NULL) || (req_count < min_count)) {
529 			min_count = req_count;
530 			min_thread = thread;
531 		}
532 	}
533 
534 	/*
535 	 * We didn't find a good match yet.  See if we can add
536 	 * a new thread to this file handle entry's thread list.
537 	 */
538 	if ((softc->ctls.max_nfsds_per_fh == 0) ||
539 	    (fhe->num_threads < softc->ctls.max_nfsds_per_fh)) {
540 		thread = this_thread;
541 #if 0
542 		ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
543 		    "fha: %p(%d)t", thread, thread->st_p2);
544 #endif
545 		fha_hash_entry_add_thread(fhe, thread);
546 	} else {
547 		/*
548 		 * We don't want to use any more threads for this file, so
549 		 * go back to the most attractive nfsd we're already using.
550 		 */
551 		thread = min_thread;
552 	}
553 
554 	return (thread);
555 }
556 
557 /*
558  * After getting a request, try to assign it to some thread.  Usually we
559  * handle it ourselves.
560  */
561 SVCTHREAD *
562 fhanew_assign(SVCTHREAD *this_thread, struct svc_req *req)
563 {
564 	struct fha_params *softc;
565 	SVCTHREAD *thread;
566 	struct fha_info i;
567 	struct fha_hash_entry *fhe;
568 
569 	NFSD_CURVNET_SET(NFSD_TD_TO_VNET(curthread));
570 	softc = NFSD_VNET(fhanew_softc);
571 	/* Check to see whether we're enabled. */
572 	if (softc->ctls.enable == 0)
573 		goto thist;
574 
575 	/*
576 	 * Only do placement if this is an NFS request.
577 	 */
578 	if (req->rq_prog != NFS_PROG)
579 		goto thist;
580 
581 	if (req->rq_vers != 2 && req->rq_vers != 3)
582 		goto thist;
583 
584 	fha_extract_info(req, &i);
585 
586 	/*
587 	 * We save the offset associated with this request for later
588 	 * nfsd matching.
589 	 */
590 	fhe = fha_hash_entry_lookup(softc, i.fh);
591 	req->rq_p1 = fhe;
592 	req->rq_p2 = i.locktype;
593 	req->rq_p3 = i.offset;
594 
595 	/*
596 	 * Choose a thread, taking into consideration locality, thread load,
597 	 * and the number of threads already working on this file.
598 	 */
599 	thread = fha_hash_entry_choose_thread(softc, fhe, &i, this_thread);
600 	KASSERT(thread, ("fha_assign: NULL thread!"));
601 	fha_hash_entry_add_op(fhe, i.locktype, 1);
602 	thread->st_p2++;
603 	thread->st_p3 = i.offset;
604 
605 	/*
606 	 * Grab the pool lock here to not let chosen thread go away before
607 	 * the new request inserted to its queue while we drop fhe lock.
608 	 */
609 	mtx_lock(&thread->st_lock);
610 	mtx_unlock(fhe->mtx);
611 
612 	NFSD_CURVNET_RESTORE();
613 	return (thread);
614 thist:
615 	req->rq_p1 = NULL;
616 	NFSD_CURVNET_RESTORE();
617 	mtx_lock(&this_thread->st_lock);
618 	return (this_thread);
619 }
620 
621 /*
622  * Called when we're done with an operation.  The request has already
623  * been de-queued.
624  */
625 void
626 fhanew_nd_complete(SVCTHREAD *thread, struct svc_req *req)
627 {
628 	struct fha_hash_entry *fhe = req->rq_p1;
629 	struct mtx *mtx;
630 
631 	NFSD_CURVNET_SET(NFSD_TD_TO_VNET(curthread));
632 	/*
633 	 * This may be called for reqs that didn't go through
634 	 * fha_assign (e.g. extra NULL ops used for RPCSEC_GSS.
635 	 */
636 	if (!fhe) {
637 		NFSD_CURVNET_RESTORE();
638 		return;
639 	}
640 
641 	mtx = fhe->mtx;
642 	mtx_lock(mtx);
643 	fha_hash_entry_add_op(fhe, req->rq_p2, -1);
644 	thread->st_p2--;
645 	KASSERT(thread->st_p2 >= 0, ("Negative request count %d on %p",
646 	    thread->st_p2, thread));
647 	if (thread->st_p2 == 0) {
648 		fha_hash_entry_remove_thread(fhe, thread);
649 		if (0 == fhe->num_rw + fhe->num_exclusive)
650 			fha_hash_entry_remove(fhe);
651 	}
652 	mtx_unlock(mtx);
653 	NFSD_CURVNET_RESTORE();
654 }
655 
656 static int
657 fhenew_stats_sysctl(SYSCTL_HANDLER_ARGS)
658 {
659 	struct fha_params *softc;
660 	int error, i;
661 	struct sbuf sb;
662 	struct fha_hash_entry *fhe;
663 	bool_t first, hfirst;
664 	SVCTHREAD *thread;
665 
666 	sbuf_new(&sb, NULL, 65536, SBUF_FIXEDLEN);
667 
668 	NFSD_CURVNET_SET(NFSD_TD_TO_VNET(curthread));
669 	softc = NFSD_VNET(fhanew_softc);
670 	for (i = 0; i < FHA_HASH_SIZE; i++)
671 		if (!LIST_EMPTY(&softc->fha_hash[i].list))
672 			break;
673 
674 	if (i == FHA_HASH_SIZE) {
675 		sbuf_printf(&sb, "No file handle entries.\n");
676 		goto out;
677 	}
678 
679 	hfirst = TRUE;
680 	for (; i < FHA_HASH_SIZE; i++) {
681 		mtx_lock(&softc->fha_hash[i].mtx);
682 		if (LIST_EMPTY(&softc->fha_hash[i].list)) {
683 			mtx_unlock(&softc->fha_hash[i].mtx);
684 			continue;
685 		}
686 		sbuf_printf(&sb, "%shash %d: {\n", hfirst ? "" : ", ", i);
687 		first = TRUE;
688 		LIST_FOREACH(fhe, &softc->fha_hash[i].list, link) {
689 			sbuf_printf(&sb, "%sfhe %p: {\n", first ? "  " : ", ",
690 			    fhe);
691 			sbuf_printf(&sb, "    fh: %ju\n", (uintmax_t) fhe->fh);
692 			sbuf_printf(&sb, "    num_rw/exclusive: %d/%d\n",
693 			    fhe->num_rw, fhe->num_exclusive);
694 			sbuf_printf(&sb, "    num_threads: %d\n",
695 			    fhe->num_threads);
696 
697 			LIST_FOREACH(thread, &fhe->threads, st_alink) {
698 				sbuf_printf(&sb, "      thread %p offset %ju "
699 				    "reqs %d\n", thread,
700 				    thread->st_p3, thread->st_p2);
701 			}
702 
703 			sbuf_printf(&sb, "  }");
704 			first = FALSE;
705 		}
706 		sbuf_printf(&sb, "\n}");
707 		mtx_unlock(&softc->fha_hash[i].mtx);
708 		hfirst = FALSE;
709 	}
710 
711  out:
712 	NFSD_CURVNET_RESTORE();
713 	sbuf_trim(&sb);
714 	sbuf_finish(&sb);
715 	error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
716 	sbuf_delete(&sb);
717 	return (error);
718 }
719