xref: /freebsd/sys/fs/nfsserver/nfs_fha_new.c (revision fdafd315ad0d0f28a11b9fb4476a9ab059c62b92)
1  /*-
2   * SPDX-License-Identifier: BSD-2-Clause
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/types.h>
30  #include <sys/mbuf.h>
31  #include <sys/sbuf.h>
32  
33  #include <fs/nfs/nfsport.h>
34  #include <fs/nfsserver/nfs_fha_new.h>
35  
36  #include <rpc/rpc.h>
37  
38  static MALLOC_DEFINE(M_NFS_FHA, "NFS FHA", "NFS FHA");
39  
40  static void		fhanew_init(void *foo);
41  static void		fhanew_uninit(void *foo);
42  static rpcproc_t	fhanew_get_procnum(rpcproc_t procnum);
43  static int		fhanew_get_fh(uint64_t *fh, int v3, struct mbuf **md,
44  			    caddr_t *dpos);
45  static int		fhanew_is_read(rpcproc_t procnum);
46  static int		fhanew_is_write(rpcproc_t procnum);
47  static int		fhanew_get_offset(struct mbuf **md, caddr_t *dpos,
48  			    int v3, struct fha_info *info);
49  static int		fhanew_no_offset(rpcproc_t procnum);
50  static void		fhanew_set_locktype(rpcproc_t procnum,
51  			    struct fha_info *info);
52  static int		fhenew_stats_sysctl(SYSCTL_HANDLER_ARGS);
53  static void		fha_extract_info(struct svc_req *req,
54  			    struct fha_info *i);
55  
56  NFSD_VNET_DEFINE_STATIC(struct fha_params *, fhanew_softc);
57  NFSD_VNET_DEFINE_STATIC(struct fha_ctls, nfsfha_ctls);
58  
59  SYSCTL_DECL(_vfs_nfsd);
60  SYSCTL_NODE(_vfs_nfsd, OID_AUTO, fha, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
61      "NFS File Handle Affinity (FHA)");
62  
63  SYSCTL_UINT(_vfs_nfsd_fha,
64      OID_AUTO, enable, CTLFLAG_NFSD_VNET | CTLFLAG_RWTUN,
65      &NFSD_VNET_NAME(nfsfha_ctls).enable, 0,
66      "Enable NFS File Handle Affinity (FHA)");
67  
68  SYSCTL_UINT(_vfs_nfsd_fha,
69      OID_AUTO, read, CTLFLAG_NFSD_VNET | CTLFLAG_RWTUN,
70      &NFSD_VNET_NAME(nfsfha_ctls).read, 0,
71      "Enable NFS FHA read locality");
72  
73  SYSCTL_UINT(_vfs_nfsd_fha,
74      OID_AUTO, write, CTLFLAG_NFSD_VNET | CTLFLAG_RWTUN,
75      &NFSD_VNET_NAME(nfsfha_ctls).write, 0,
76      "Enable NFS FHA write locality");
77  
78  SYSCTL_UINT(_vfs_nfsd_fha,
79      OID_AUTO, bin_shift, CTLFLAG_NFSD_VNET | CTLFLAG_RWTUN,
80      &NFSD_VNET_NAME(nfsfha_ctls).bin_shift, 0,
81      "Maximum locality distance 2^(bin_shift) bytes");
82  
83  SYSCTL_UINT(_vfs_nfsd_fha,
84      OID_AUTO, max_nfsds_per_fh, CTLFLAG_NFSD_VNET | CTLFLAG_RWTUN,
85      &NFSD_VNET_NAME(nfsfha_ctls).max_nfsds_per_fh, 0,
86      "Maximum nfsd threads that "
87      "should be working on requests for the same file handle");
88  
89  SYSCTL_UINT(_vfs_nfsd_fha,
90      OID_AUTO, max_reqs_per_nfsd, CTLFLAG_NFSD_VNET | CTLFLAG_RWTUN,
91      &NFSD_VNET_NAME(nfsfha_ctls).max_reqs_per_nfsd, 0, "Maximum requests that "
92      "single nfsd thread should be working on at any time");
93  
94  SYSCTL_PROC(_vfs_nfsd_fha, OID_AUTO, fhe_stats,
95      CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
96      fhenew_stats_sysctl, "A", "");
97  
98  extern int newnfs_nfsv3_procid[];
99  
100  VNET_SYSINIT(nfs_fhanew, SI_SUB_VNET_DONE, SI_ORDER_ANY, fhanew_init, NULL);
101  VNET_SYSUNINIT(nfs_fhanew, SI_SUB_VNET_DONE, SI_ORDER_ANY, fhanew_uninit, NULL);
102  
103  static void
fhanew_init(void * foo)104  fhanew_init(void *foo)
105  {
106  	struct fha_params *softc;
107  	int i;
108  
109  	NFSD_VNET(fhanew_softc) = malloc(sizeof(struct fha_params), M_TEMP,
110  	    M_WAITOK | M_ZERO);
111  	softc = NFSD_VNET(fhanew_softc);
112  
113  	snprintf(softc->server_name, sizeof(softc->server_name),
114  	    FHANEW_SERVER_NAME);
115  
116  	for (i = 0; i < FHA_HASH_SIZE; i++)
117  		mtx_init(&softc->fha_hash[i].mtx, "fhalock", NULL, MTX_DEF);
118  
119  	/*
120  	 * Set the default tuning parameters.
121  	 */
122  	NFSD_VNET(nfsfha_ctls).enable = FHA_DEF_ENABLE;
123  	NFSD_VNET(nfsfha_ctls).read = FHA_DEF_READ;
124  	NFSD_VNET(nfsfha_ctls).write = FHA_DEF_WRITE;
125  	NFSD_VNET(nfsfha_ctls).bin_shift = FHA_DEF_BIN_SHIFT;
126  	NFSD_VNET(nfsfha_ctls).max_nfsds_per_fh = FHA_DEF_MAX_NFSDS_PER_FH;
127  	NFSD_VNET(nfsfha_ctls).max_reqs_per_nfsd = FHA_DEF_MAX_REQS_PER_NFSD;
128  
129  }
130  
131  static void
fhanew_uninit(void * foo)132  fhanew_uninit(void *foo)
133  {
134  	struct fha_params *softc;
135  	int i;
136  
137  	softc = NFSD_VNET(fhanew_softc);
138  
139  	for (i = 0; i < FHA_HASH_SIZE; i++)
140  		mtx_destroy(&softc->fha_hash[i].mtx);
141  	free(softc, M_TEMP);
142  }
143  
144  static rpcproc_t
fhanew_get_procnum(rpcproc_t procnum)145  fhanew_get_procnum(rpcproc_t procnum)
146  {
147  	if (procnum > NFSV2PROC_STATFS)
148  		return (-1);
149  
150  	return (newnfs_nfsv3_procid[procnum]);
151  }
152  
153  static int
fhanew_get_fh(uint64_t * fh,int v3,struct mbuf ** md,caddr_t * dpos)154  fhanew_get_fh(uint64_t *fh, int v3, struct mbuf **md, caddr_t *dpos)
155  {
156  	struct nfsrv_descript lnd, *nd;
157  	uint32_t *tl;
158  	uint8_t *buf;
159  	uint64_t t;
160  	int error, len, i;
161  
162  	error = 0;
163  	len = 0;
164  	nd = &lnd;
165  
166  	nd->nd_md = *md;
167  	nd->nd_dpos = *dpos;
168  
169  	if (v3) {
170  		NFSM_DISSECT_NONBLOCK(tl, uint32_t *, NFSX_UNSIGNED);
171  		if ((len = fxdr_unsigned(int, *tl)) <= 0 || len > NFSX_FHMAX) {
172  			error = EBADRPC;
173  			goto nfsmout;
174  		}
175  	} else {
176  		len = NFSX_V2FH;
177  	}
178  
179  	t = 0;
180  	if (len != 0) {
181  		NFSM_DISSECT_NONBLOCK(buf, uint8_t *, len);
182  		for (i = 0; i < len; i++)
183  			t ^= ((uint64_t)buf[i] << (i & 7) * 8);
184  	}
185  	*fh = t;
186  
187  nfsmout:
188  	*md = nd->nd_md;
189  	*dpos = nd->nd_dpos;
190  
191  	return (error);
192  }
193  
194  static int
fhanew_is_read(rpcproc_t procnum)195  fhanew_is_read(rpcproc_t procnum)
196  {
197  	if (procnum == NFSPROC_READ)
198  		return (1);
199  	else
200  		return (0);
201  }
202  
203  static int
fhanew_is_write(rpcproc_t procnum)204  fhanew_is_write(rpcproc_t procnum)
205  {
206  	if (procnum == NFSPROC_WRITE)
207  		return (1);
208  	else
209  		return (0);
210  }
211  
212  static int
fhanew_get_offset(struct mbuf ** md,caddr_t * dpos,int v3,struct fha_info * info)213  fhanew_get_offset(struct mbuf **md, caddr_t *dpos, int v3,
214      struct fha_info *info)
215  {
216  	struct nfsrv_descript lnd, *nd;
217  	uint32_t *tl;
218  	int error;
219  
220  	error = 0;
221  
222  	nd = &lnd;
223  	nd->nd_md = *md;
224  	nd->nd_dpos = *dpos;
225  
226  	if (v3) {
227  		NFSM_DISSECT_NONBLOCK(tl, uint32_t *, 2 * NFSX_UNSIGNED);
228  		info->offset = fxdr_hyper(tl);
229  	} else {
230  		NFSM_DISSECT_NONBLOCK(tl, uint32_t *, NFSX_UNSIGNED);
231  		info->offset = fxdr_unsigned(uint32_t, *tl);
232  	}
233  
234  nfsmout:
235  	*md = nd->nd_md;
236  	*dpos = nd->nd_dpos;
237  
238  	return (error);
239  }
240  
241  static int
fhanew_no_offset(rpcproc_t procnum)242  fhanew_no_offset(rpcproc_t procnum)
243  {
244  	if (procnum == NFSPROC_FSSTAT ||
245  	    procnum == NFSPROC_FSINFO ||
246  	    procnum == NFSPROC_PATHCONF ||
247  	    procnum == NFSPROC_NOOP ||
248  	    procnum == NFSPROC_NULL)
249  		return (1);
250  	else
251  		return (0);
252  }
253  
254  static void
fhanew_set_locktype(rpcproc_t procnum,struct fha_info * info)255  fhanew_set_locktype(rpcproc_t procnum, struct fha_info *info)
256  {
257  	switch (procnum) {
258  	case NFSPROC_NULL:
259  	case NFSPROC_GETATTR:
260  	case NFSPROC_LOOKUP:
261  	case NFSPROC_ACCESS:
262  	case NFSPROC_READLINK:
263  	case NFSPROC_READ:
264  	case NFSPROC_READDIR:
265  	case NFSPROC_READDIRPLUS:
266  	case NFSPROC_WRITE:
267  		info->locktype = LK_SHARED;
268  		break;
269  	case NFSPROC_SETATTR:
270  	case NFSPROC_CREATE:
271  	case NFSPROC_MKDIR:
272  	case NFSPROC_SYMLINK:
273  	case NFSPROC_MKNOD:
274  	case NFSPROC_REMOVE:
275  	case NFSPROC_RMDIR:
276  	case NFSPROC_RENAME:
277  	case NFSPROC_LINK:
278  	case NFSPROC_FSSTAT:
279  	case NFSPROC_FSINFO:
280  	case NFSPROC_PATHCONF:
281  	case NFSPROC_COMMIT:
282  	case NFSPROC_NOOP:
283  		info->locktype = LK_EXCLUSIVE;
284  		break;
285  	}
286  }
287  
288  /*
289   * This just specifies that offsets should obey affinity when within
290   * the same 1Mbyte (1<<20) chunk for the file (reads only for now).
291   */
292  static void
fha_extract_info(struct svc_req * req,struct fha_info * i)293  fha_extract_info(struct svc_req *req, struct fha_info *i)
294  {
295  	struct mbuf *md;
296  	caddr_t dpos;
297  	static u_int64_t random_fh = 0;
298  	int error;
299  	int v3 = (req->rq_vers == 3);
300  	rpcproc_t procnum;
301  
302  	/*
303  	 * We start off with a random fh.  If we get a reasonable
304  	 * procnum, we set the fh.  If there's a concept of offset
305  	 * that we're interested in, we set that.
306  	 */
307  	i->fh = ++random_fh;
308  	i->offset = 0;
309  	i->locktype = LK_EXCLUSIVE;
310  	i->read = i->write = 0;
311  
312  	/*
313  	 * Extract the procnum and convert to v3 form if necessary,
314  	 * taking care to deal with out-of-range procnums.  Caller will
315  	 * ensure that rq_vers is either 2 or 3.
316  	 */
317  	procnum = req->rq_proc;
318  	if (!v3) {
319  		rpcproc_t tmp_procnum;
320  
321  		tmp_procnum = fhanew_get_procnum(procnum);
322  		if (tmp_procnum == -1)
323  			goto out;
324  		procnum = tmp_procnum;
325  	}
326  
327  	/*
328  	 * We do affinity for most.  However, we divide a realm of affinity
329  	 * by file offset so as to allow for concurrent random access.  We
330  	 * only do this for reads today, but this may change when IFS supports
331  	 * efficient concurrent writes.
332  	 */
333  	if (fhanew_no_offset(procnum))
334  		goto out;
335  
336  	i->read = fhanew_is_read(procnum);
337  	i->write = fhanew_is_write(procnum);
338  
339  	error = newnfs_realign(&req->rq_args, M_NOWAIT);
340  	if (error)
341  		goto out;
342  	md = req->rq_args;
343  	dpos = mtod(md, caddr_t);
344  
345  	/* Grab the filehandle. */
346  	error = fhanew_get_fh(&i->fh, v3, &md, &dpos);
347  	if (error)
348  		goto out;
349  
350  	/* Content ourselves with zero offset for all but reads. */
351  	if (i->read || i->write)
352  		fhanew_get_offset(&md, &dpos, v3, i);
353  
354  out:
355  	fhanew_set_locktype(procnum, i);
356  }
357  
358  static struct fha_hash_entry *
fha_hash_entry_new(u_int64_t fh)359  fha_hash_entry_new(u_int64_t fh)
360  {
361  	struct fha_hash_entry *e;
362  
363  	e = malloc(sizeof(*e), M_NFS_FHA, M_WAITOK);
364  	e->fh = fh;
365  	e->num_rw = 0;
366  	e->num_exclusive = 0;
367  	e->num_threads = 0;
368  	LIST_INIT(&e->threads);
369  
370  	return (e);
371  }
372  
373  static void
fha_hash_entry_destroy(struct fha_hash_entry * e)374  fha_hash_entry_destroy(struct fha_hash_entry *e)
375  {
376  
377  	mtx_assert(e->mtx, MA_OWNED);
378  	KASSERT(e->num_rw == 0,
379  	    ("%d reqs on destroyed fhe %p", e->num_rw, e));
380  	KASSERT(e->num_exclusive == 0,
381  	    ("%d exclusive reqs on destroyed fhe %p", e->num_exclusive, e));
382  	KASSERT(e->num_threads == 0,
383  	    ("%d threads on destroyed fhe %p", e->num_threads, e));
384  	free(e, M_NFS_FHA);
385  }
386  
387  static void
fha_hash_entry_remove(struct fha_hash_entry * e)388  fha_hash_entry_remove(struct fha_hash_entry *e)
389  {
390  
391  	mtx_assert(e->mtx, MA_OWNED);
392  	LIST_REMOVE(e, link);
393  	fha_hash_entry_destroy(e);
394  }
395  
396  static struct fha_hash_entry *
fha_hash_entry_lookup(struct fha_params * softc,u_int64_t fh)397  fha_hash_entry_lookup(struct fha_params *softc, u_int64_t fh)
398  {
399  	struct fha_hash_slot *fhs;
400  	struct fha_hash_entry *fhe, *new_fhe;
401  
402  	fhs = &softc->fha_hash[fh % FHA_HASH_SIZE];
403  	new_fhe = fha_hash_entry_new(fh);
404  	new_fhe->mtx = &fhs->mtx;
405  	mtx_lock(&fhs->mtx);
406  	LIST_FOREACH(fhe, &fhs->list, link)
407  		if (fhe->fh == fh)
408  			break;
409  	if (!fhe) {
410  		fhe = new_fhe;
411  		LIST_INSERT_HEAD(&fhs->list, fhe, link);
412  	} else
413  		fha_hash_entry_destroy(new_fhe);
414  	return (fhe);
415  }
416  
417  static void
fha_hash_entry_add_thread(struct fha_hash_entry * fhe,SVCTHREAD * thread)418  fha_hash_entry_add_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread)
419  {
420  
421  	mtx_assert(fhe->mtx, MA_OWNED);
422  	thread->st_p2 = 0;
423  	LIST_INSERT_HEAD(&fhe->threads, thread, st_alink);
424  	fhe->num_threads++;
425  }
426  
427  static void
fha_hash_entry_remove_thread(struct fha_hash_entry * fhe,SVCTHREAD * thread)428  fha_hash_entry_remove_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread)
429  {
430  
431  	mtx_assert(fhe->mtx, MA_OWNED);
432  	KASSERT(thread->st_p2 == 0,
433  	    ("%d reqs on removed thread %p", thread->st_p2, thread));
434  	LIST_REMOVE(thread, st_alink);
435  	fhe->num_threads--;
436  }
437  
438  /*
439   * Account for an ongoing operation associated with this file.
440   */
441  static void
fha_hash_entry_add_op(struct fha_hash_entry * fhe,int locktype,int count)442  fha_hash_entry_add_op(struct fha_hash_entry *fhe, int locktype, int count)
443  {
444  
445  	mtx_assert(fhe->mtx, MA_OWNED);
446  	if (LK_EXCLUSIVE == locktype)
447  		fhe->num_exclusive += count;
448  	else
449  		fhe->num_rw += count;
450  }
451  
452  /*
453   * Get the service thread currently associated with the fhe that is
454   * appropriate to handle this operation.
455   */
456  static SVCTHREAD *
fha_hash_entry_choose_thread(struct fha_params * softc,struct fha_hash_entry * fhe,struct fha_info * i,SVCTHREAD * this_thread)457  fha_hash_entry_choose_thread(struct fha_params *softc,
458      struct fha_hash_entry *fhe, struct fha_info *i, SVCTHREAD *this_thread)
459  {
460  	SVCTHREAD *thread, *min_thread = NULL;
461  	int req_count, min_count = 0;
462  	off_t offset1, offset2;
463  
464  	LIST_FOREACH(thread, &fhe->threads, st_alink) {
465  		req_count = thread->st_p2;
466  
467  		/* If there are any writes in progress, use the first thread. */
468  		if (fhe->num_exclusive) {
469  #if 0
470  			ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
471  			    "fha: %p(%d)w", thread, req_count);
472  #endif
473  			return (thread);
474  		}
475  
476  		/* Check whether we should consider locality. */
477  		if ((i->read && !NFSD_VNET(nfsfha_ctls).read) ||
478  		    (i->write && !NFSD_VNET(nfsfha_ctls).write))
479  			goto noloc;
480  
481  		/*
482  		 * Check for locality, making sure that we won't
483  		 * exceed our per-thread load limit in the process.
484  		 */
485  		offset1 = i->offset;
486  		offset2 = thread->st_p3;
487  
488  		if (((offset1 >= offset2)
489  		  && ((offset1 - offset2) < (1 << NFSD_VNET(nfsfha_ctls).bin_shift)))
490  		 || ((offset2 > offset1)
491  		  && ((offset2 - offset1) < (1 << NFSD_VNET(nfsfha_ctls).bin_shift)))) {
492  			if ((NFSD_VNET(nfsfha_ctls).max_reqs_per_nfsd == 0) ||
493  			    (req_count < NFSD_VNET(nfsfha_ctls).max_reqs_per_nfsd)) {
494  #if 0
495  				ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
496  				    "fha: %p(%d)r", thread, req_count);
497  #endif
498  				return (thread);
499  			}
500  		}
501  
502  noloc:
503  		/*
504  		 * We don't have a locality match, so skip this thread,
505  		 * but keep track of the most attractive thread in case
506  		 * we need to come back to it later.
507  		 */
508  #if 0
509  		ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
510  		    "fha: %p(%d)s off1 %llu off2 %llu", thread,
511  		    req_count, offset1, offset2);
512  #endif
513  		if ((min_thread == NULL) || (req_count < min_count)) {
514  			min_count = req_count;
515  			min_thread = thread;
516  		}
517  	}
518  
519  	/*
520  	 * We didn't find a good match yet.  See if we can add
521  	 * a new thread to this file handle entry's thread list.
522  	 */
523  	if ((NFSD_VNET(nfsfha_ctls).max_nfsds_per_fh == 0) ||
524  	    (fhe->num_threads < NFSD_VNET(nfsfha_ctls).max_nfsds_per_fh)) {
525  		thread = this_thread;
526  #if 0
527  		ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
528  		    "fha: %p(%d)t", thread, thread->st_p2);
529  #endif
530  		fha_hash_entry_add_thread(fhe, thread);
531  	} else {
532  		/*
533  		 * We don't want to use any more threads for this file, so
534  		 * go back to the most attractive nfsd we're already using.
535  		 */
536  		thread = min_thread;
537  	}
538  
539  	return (thread);
540  }
541  
542  /*
543   * After getting a request, try to assign it to some thread.  Usually we
544   * handle it ourselves.
545   */
546  SVCTHREAD *
fhanew_assign(SVCTHREAD * this_thread,struct svc_req * req)547  fhanew_assign(SVCTHREAD *this_thread, struct svc_req *req)
548  {
549  	struct fha_params *softc;
550  	SVCTHREAD *thread;
551  	struct fha_info i;
552  	struct fha_hash_entry *fhe;
553  
554  	NFSD_CURVNET_SET(NFSD_TD_TO_VNET(curthread));
555  	softc = NFSD_VNET(fhanew_softc);
556  	/* Check to see whether we're enabled. */
557  	if (NFSD_VNET(nfsfha_ctls).enable == 0)
558  		goto thist;
559  
560  	/*
561  	 * Only do placement if this is an NFS request.
562  	 */
563  	if (req->rq_prog != NFS_PROG)
564  		goto thist;
565  
566  	if (req->rq_vers != 2 && req->rq_vers != 3)
567  		goto thist;
568  
569  	fha_extract_info(req, &i);
570  
571  	/*
572  	 * We save the offset associated with this request for later
573  	 * nfsd matching.
574  	 */
575  	fhe = fha_hash_entry_lookup(softc, i.fh);
576  	req->rq_p1 = fhe;
577  	req->rq_p2 = i.locktype;
578  	req->rq_p3 = i.offset;
579  
580  	/*
581  	 * Choose a thread, taking into consideration locality, thread load,
582  	 * and the number of threads already working on this file.
583  	 */
584  	thread = fha_hash_entry_choose_thread(softc, fhe, &i, this_thread);
585  	KASSERT(thread, ("fha_assign: NULL thread!"));
586  	fha_hash_entry_add_op(fhe, i.locktype, 1);
587  	thread->st_p2++;
588  	thread->st_p3 = i.offset;
589  
590  	/*
591  	 * Grab the pool lock here to not let chosen thread go away before
592  	 * the new request inserted to its queue while we drop fhe lock.
593  	 */
594  	mtx_lock(&thread->st_lock);
595  	mtx_unlock(fhe->mtx);
596  
597  	NFSD_CURVNET_RESTORE();
598  	return (thread);
599  thist:
600  	req->rq_p1 = NULL;
601  	NFSD_CURVNET_RESTORE();
602  	mtx_lock(&this_thread->st_lock);
603  	return (this_thread);
604  }
605  
606  /*
607   * Called when we're done with an operation.  The request has already
608   * been de-queued.
609   */
610  void
fhanew_nd_complete(SVCTHREAD * thread,struct svc_req * req)611  fhanew_nd_complete(SVCTHREAD *thread, struct svc_req *req)
612  {
613  	struct fha_hash_entry *fhe = req->rq_p1;
614  	struct mtx *mtx;
615  
616  	NFSD_CURVNET_SET(NFSD_TD_TO_VNET(curthread));
617  	/*
618  	 * This may be called for reqs that didn't go through
619  	 * fha_assign (e.g. extra NULL ops used for RPCSEC_GSS.
620  	 */
621  	if (!fhe) {
622  		NFSD_CURVNET_RESTORE();
623  		return;
624  	}
625  
626  	mtx = fhe->mtx;
627  	mtx_lock(mtx);
628  	fha_hash_entry_add_op(fhe, req->rq_p2, -1);
629  	thread->st_p2--;
630  	KASSERT(thread->st_p2 >= 0, ("Negative request count %d on %p",
631  	    thread->st_p2, thread));
632  	if (thread->st_p2 == 0) {
633  		fha_hash_entry_remove_thread(fhe, thread);
634  		if (0 == fhe->num_rw + fhe->num_exclusive)
635  			fha_hash_entry_remove(fhe);
636  	}
637  	mtx_unlock(mtx);
638  	NFSD_CURVNET_RESTORE();
639  }
640  
641  static int
fhenew_stats_sysctl(SYSCTL_HANDLER_ARGS)642  fhenew_stats_sysctl(SYSCTL_HANDLER_ARGS)
643  {
644  	struct fha_params *softc;
645  	int error, i;
646  	struct sbuf sb;
647  	struct fha_hash_entry *fhe;
648  	bool_t first, hfirst;
649  	SVCTHREAD *thread;
650  
651  	sbuf_new(&sb, NULL, 65536, SBUF_FIXEDLEN);
652  
653  	NFSD_CURVNET_SET(NFSD_TD_TO_VNET(curthread));
654  	softc = NFSD_VNET(fhanew_softc);
655  	for (i = 0; i < FHA_HASH_SIZE; i++)
656  		if (!LIST_EMPTY(&softc->fha_hash[i].list))
657  			break;
658  
659  	if (i == FHA_HASH_SIZE) {
660  		sbuf_printf(&sb, "No file handle entries.\n");
661  		goto out;
662  	}
663  
664  	hfirst = TRUE;
665  	for (; i < FHA_HASH_SIZE; i++) {
666  		mtx_lock(&softc->fha_hash[i].mtx);
667  		if (LIST_EMPTY(&softc->fha_hash[i].list)) {
668  			mtx_unlock(&softc->fha_hash[i].mtx);
669  			continue;
670  		}
671  		sbuf_printf(&sb, "%shash %d: {\n", hfirst ? "" : ", ", i);
672  		first = TRUE;
673  		LIST_FOREACH(fhe, &softc->fha_hash[i].list, link) {
674  			sbuf_printf(&sb, "%sfhe %p: {\n", first ? "  " : ", ",
675  			    fhe);
676  			sbuf_printf(&sb, "    fh: %ju\n", (uintmax_t) fhe->fh);
677  			sbuf_printf(&sb, "    num_rw/exclusive: %d/%d\n",
678  			    fhe->num_rw, fhe->num_exclusive);
679  			sbuf_printf(&sb, "    num_threads: %d\n",
680  			    fhe->num_threads);
681  
682  			LIST_FOREACH(thread, &fhe->threads, st_alink) {
683  				sbuf_printf(&sb, "      thread %p offset %ju "
684  				    "reqs %d\n", thread,
685  				    thread->st_p3, thread->st_p2);
686  			}
687  
688  			sbuf_printf(&sb, "  }");
689  			first = FALSE;
690  		}
691  		sbuf_printf(&sb, "\n}");
692  		mtx_unlock(&softc->fha_hash[i].mtx);
693  		hfirst = FALSE;
694  	}
695  
696   out:
697  	NFSD_CURVNET_RESTORE();
698  	sbuf_trim(&sb);
699  	sbuf_finish(&sb);
700  	error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
701  	sbuf_delete(&sb);
702  	return (error);
703  }
704