1 /*-
2 * Copyright (c) 2008 Isilon Inc http://www.isilon.com/
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
12 *
13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23 * SUCH DAMAGE.
24 */
25
26 #include <sys/cdefs.h>
27 __FBSDID("$FreeBSD: releng/8.0/sys/nfsserver/nfs_fha.c 195202 2009-06-30 19:03:27Z dfr $");
28
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/sysproto.h>
32 #include <sys/kernel.h>
33 #include <sys/sysctl.h>
34 #include <sys/vnode.h>
35 #include <sys/malloc.h>
36 #include <sys/mount.h>
37 #include <sys/mbuf.h>
38 #include <sys/sbuf.h>
39
40 #include <rpc/rpc.h>
41 #include <nfs/xdr_subs.h>
42 #include <nfs/nfsproto.h>
43 #include <nfsserver/nfs.h>
44 #include <nfsserver/nfsm_subs.h>
45 #include <nfsserver/nfs_fha.h>
46
47 static MALLOC_DEFINE(M_NFS_FHA, "NFS FHA", "NFS FHA");
48
49 /* Sysctl defaults. */
50 #define DEF_BIN_SHIFT 18 /* 256k */
51 #define DEF_MAX_NFSDS_PER_FH 8
52 #define DEF_MAX_REQS_PER_NFSD 4
53
54 struct fha_ctls {
55 u_int32_t bin_shift;
56 u_int32_t max_nfsds_per_fh;
57 u_int32_t max_reqs_per_nfsd;
58 } fha_ctls;
59
60 struct sysctl_ctx_list fha_clist;
61
62 SYSCTL_DECL(_vfs_nfsrv);
63 SYSCTL_DECL(_vfs_nfsrv_fha);
64
65 /* Static sysctl node for the fha from the top-level vfs_nfsrv node. */
66 SYSCTL_NODE(_vfs_nfsrv, OID_AUTO, fha, CTLFLAG_RD, 0, "fha node");
67
68 /* This is the global structure that represents the state of the fha system. */
69 static struct fha_global {
70 struct fha_hash_entry_list *hashtable;
71 u_long hashmask;
72 } g_fha;
73
74 /*
75 * These are the entries in the filehandle hash. They talk about a specific
76 * file, requests against which are being handled by one or more nfsds. We keep
77 * a chain of nfsds against the file. We only have more than one if reads are
78 * ongoing, and then only if the reads affect disparate regions of the file.
79 *
80 * In general, we want to assign a new request to an existing nfsd if it is
81 * going to contend with work happening already on that nfsd, or if the
82 * operation is a read and the nfsd is already handling a proximate read. We
83 * do this to avoid jumping around in the read stream unnecessarily, and to
84 * avoid contention between threads over single files.
85 */
86 struct fha_hash_entry {
87 LIST_ENTRY(fha_hash_entry) link;
88 u_int64_t fh;
89 u_int16_t num_reads;
90 u_int16_t num_writes;
91 u_int8_t num_threads;
92 struct svcthread_list threads;
93 };
94 LIST_HEAD(fha_hash_entry_list, fha_hash_entry);
95
96 /* A structure used for passing around data internally. */
97 struct fha_info {
98 u_int64_t fh;
99 off_t offset;
100 int locktype;
101 };
102
103 static int fhe_stats_sysctl(SYSCTL_HANDLER_ARGS);
104
105 static void
106 nfs_fha_init(void *foo)
107 {
108
109 /*
110 * A small hash table to map filehandles to fha_hash_entry
111 * structures.
112 */
113 g_fha.hashtable = hashinit(256, M_NFS_FHA, &g_fha.hashmask);
114
115 /*
116 * Initialize the sysctl context list for the fha module.
117 */
118 sysctl_ctx_init(&fha_clist);
119
120 fha_ctls.bin_shift = DEF_BIN_SHIFT;
121 fha_ctls.max_nfsds_per_fh = DEF_MAX_NFSDS_PER_FH;
122 fha_ctls.max_reqs_per_nfsd = DEF_MAX_REQS_PER_NFSD;
123
124 SYSCTL_ADD_UINT(&fha_clist, SYSCTL_STATIC_CHILDREN(_vfs_nfsrv_fha),
125 OID_AUTO, "bin_shift", CTLFLAG_RW,
126 &fha_ctls.bin_shift, 0, "For FHA reads, no two requests will "
127 "contend if they're 2^(bin_shift) bytes apart");
128
129 SYSCTL_ADD_UINT(&fha_clist, SYSCTL_STATIC_CHILDREN(_vfs_nfsrv_fha),
130 OID_AUTO, "max_nfsds_per_fh", CTLFLAG_RW,
131 &fha_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(&fha_clist, SYSCTL_STATIC_CHILDREN(_vfs_nfsrv_fha),
135 OID_AUTO, "max_reqs_per_nfsd", CTLFLAG_RW,
136 &fha_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(&fha_clist, SYSCTL_STATIC_CHILDREN(_vfs_nfsrv_fha),
140 OID_AUTO, "fhe_stats", CTLTYPE_STRING | CTLFLAG_RD, 0, 0,
141 fhe_stats_sysctl, "A", "");
142 }
143
144 static void
145 nfs_fha_uninit(void *foo)
146 {
147
148 hashdestroy(g_fha.hashtable, M_NFS_FHA, g_fha.hashmask);
149 }
150
151 SYSINIT(nfs_fha, SI_SUB_ROOT_CONF, SI_ORDER_ANY, nfs_fha_init, NULL);
152 SYSUNINIT(nfs_fha, SI_SUB_ROOT_CONF, SI_ORDER_ANY, nfs_fha_uninit, NULL);
153
154 /*
155 * This just specifies that offsets should obey affinity when within
156 * the same 1Mbyte (1<<20) chunk for the file (reads only for now).
157 */
158 static void
159 fha_extract_info(struct svc_req *req, struct fha_info *i)
160 {
161 struct mbuf *md = req->rq_args;
162 nfsfh_t fh;
163 caddr_t dpos = mtod(md, caddr_t);
164 static u_int64_t random_fh = 0;
165 int error;
166 int v3 = (req->rq_vers == 3);
167 u_int32_t *tl;
168 rpcproc_t procnum;
169
170 /*
171 * We start off with a random fh. If we get a reasonable
172 * procnum, we set the fh. If there's a concept of offset
173 * that we're interested in, we set that.
174 */
175 i->fh = ++random_fh;
176 i->offset = 0;
177 i->locktype = LK_EXCLUSIVE;
178
179 /*
180 * Extract the procnum and convert to v3 form if necessary,
181 * taking care to deal with out-of-range procnums. Caller will
182 * ensure that rq_vers is either 2 or 3.
183 */
184 procnum = req->rq_proc;
185 if (!v3) {
186 if (procnum > NFSV2PROC_STATFS)
187 goto out;
188 procnum = nfsrv_nfsv3_procid[procnum];
189 }
190
191 /*
192 * We do affinity for most. However, we divide a realm of affinity
193 * by file offset so as to allow for concurrent random access. We
194 * only do this for reads today, but this may change when IFS supports
195 * efficient concurrent writes.
196 */
197 if (procnum == NFSPROC_FSSTAT ||
198 procnum == NFSPROC_FSINFO ||
199 procnum == NFSPROC_PATHCONF ||
200 procnum == NFSPROC_NOOP ||
201 procnum == NFSPROC_NULL)
202 goto out;
203
204 /* Grab the filehandle. */
205 error = nfsm_srvmtofh_xx(&fh.fh_generic, v3, &md, &dpos);
206 if (error)
207 goto out;
208
209 i->fh = *(const u_int64_t *)(fh.fh_generic.fh_fid.fid_data);
210
211 /* Content ourselves with zero offset for all but reads. */
212 if (procnum != NFSPROC_READ)
213 goto out;
214
215 if (v3) {
216 tl = nfsm_dissect_xx_nonblock(2 * NFSX_UNSIGNED, &md, &dpos);
217 if (tl == NULL)
218 goto out;
219 i->offset = fxdr_hyper(tl);
220 } else {
221 tl = nfsm_dissect_xx_nonblock(NFSX_UNSIGNED, &md, &dpos);
222 if (tl == NULL)
223 goto out;
224 i->offset = fxdr_unsigned(u_int32_t, *tl);
225 }
226 out:
227 switch (procnum) {
228 case NFSPROC_NULL:
229 case NFSPROC_GETATTR:
230 case NFSPROC_LOOKUP:
231 case NFSPROC_ACCESS:
232 case NFSPROC_READLINK:
233 case NFSPROC_READ:
234 case NFSPROC_READDIR:
235 case NFSPROC_READDIRPLUS:
236 i->locktype = LK_SHARED;
237 break;
238 case NFSPROC_SETATTR:
239 case NFSPROC_WRITE:
240 case NFSPROC_CREATE:
241 case NFSPROC_MKDIR:
242 case NFSPROC_SYMLINK:
243 case NFSPROC_MKNOD:
244 case NFSPROC_REMOVE:
245 case NFSPROC_RMDIR:
246 case NFSPROC_RENAME:
247 case NFSPROC_LINK:
248 case NFSPROC_FSSTAT:
249 case NFSPROC_FSINFO:
250 case NFSPROC_PATHCONF:
251 case NFSPROC_COMMIT:
252 case NFSPROC_NOOP:
253 i->locktype = LK_EXCLUSIVE;
254 break;
255 }
256 }
257
258 static struct fha_hash_entry *
259 fha_hash_entry_new(u_int64_t fh)
260 {
261 struct fha_hash_entry *e;
262
263 e = malloc(sizeof(*e), M_NFS_FHA, M_WAITOK);
264 e->fh = fh;
265 e->num_reads = 0;
266 e->num_writes = 0;
267 e->num_threads = 0;
268 LIST_INIT(&e->threads);
269
270 return e;
271 }
272
273 static void
274 fha_hash_entry_destroy(struct fha_hash_entry *e)
275 {
276
277 if (e->num_reads + e->num_writes)
278 panic("nonempty fhe");
279 free(e, M_NFS_FHA);
280 }
281
282 static void
283 fha_hash_entry_remove(struct fha_hash_entry *e)
284 {
285
286 LIST_REMOVE(e, link);
287 fha_hash_entry_destroy(e);
288 }
289
290 static struct fha_hash_entry *
291 fha_hash_entry_lookup(SVCPOOL *pool, u_int64_t fh)
292 {
293 struct fha_hash_entry *fhe, *new_fhe;
294
295 LIST_FOREACH(fhe, &g_fha.hashtable[fh % g_fha.hashmask], link) {
296 if (fhe->fh == fh)
297 break;
298 }
299
300 if (!fhe) {
301 /* Allocate a new entry. */
302 mtx_unlock(&pool->sp_lock);
303 new_fhe = fha_hash_entry_new(fh);
304 mtx_lock(&pool->sp_lock);
305
306 /* Double-check to make sure we still need the new entry. */
307 LIST_FOREACH(fhe, &g_fha.hashtable[fh % g_fha.hashmask], link) {
308 if (fhe->fh == fh)
309 break;
310 }
311 if (!fhe) {
312 fhe = new_fhe;
313 LIST_INSERT_HEAD(&g_fha.hashtable[fh % g_fha.hashmask],
314 fhe, link);
315 } else {
316 fha_hash_entry_destroy(new_fhe);
317 }
318 }
319
320 return fhe;
321 }
322
323 static void
324 fha_hash_entry_add_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread)
325 {
326 LIST_INSERT_HEAD(&fhe->threads, thread, st_alink);
327 fhe->num_threads++;
328 }
329
330 static void
331 fha_hash_entry_remove_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread)
332 {
333
334 LIST_REMOVE(thread, st_alink);
335 fhe->num_threads--;
336 }
337
338 /*
339 * Account for an ongoing operation associated with this file.
340 */
341 static void
342 fha_hash_entry_add_op(struct fha_hash_entry *fhe, int locktype, int count)
343 {
344
345 if (LK_EXCLUSIVE == locktype)
346 fhe->num_writes += count;
347 else
348 fhe->num_reads += count;
349 }
350
351 static SVCTHREAD *
352 get_idle_thread(SVCPOOL *pool)
353 {
354 SVCTHREAD *st;
355
356 LIST_FOREACH(st, &pool->sp_idlethreads, st_ilink) {
357 if (st->st_xprt == NULL && STAILQ_EMPTY(&st->st_reqs))
358 return (st);
359 }
360 return (NULL);
361 }
362
363
364 /*
365 * Get the service thread currently associated with the fhe that is
366 * appropriate to handle this operation.
367 */
368 SVCTHREAD *
369 fha_hash_entry_choose_thread(SVCPOOL *pool, struct fha_hash_entry *fhe,
370 struct fha_info *i, SVCTHREAD *this_thread);
371
372 SVCTHREAD *
373 fha_hash_entry_choose_thread(SVCPOOL *pool, struct fha_hash_entry *fhe,
374 struct fha_info *i, SVCTHREAD *this_thread)
375 {
376 SVCTHREAD *thread, *min_thread = NULL;
377 int req_count, min_count = 0;
378 off_t offset1, offset2;
379
380 LIST_FOREACH(thread, &fhe->threads, st_alink) {
381 req_count = thread->st_reqcount;
382
383 /* If there are any writes in progress, use the first thread. */
384 if (fhe->num_writes) {
385 #if 0
386 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
387 "fha: %p(%d)w", thread, req_count);
388 #endif
389 return (thread);
390 }
391
392 /*
393 * Check for read locality, making sure that we won't
394 * exceed our per-thread load limit in the process.
395 */
396 offset1 = i->offset >> fha_ctls.bin_shift;
397 offset2 = STAILQ_FIRST(&thread->st_reqs)->rq_p3
398 >> fha_ctls.bin_shift;
399 if (offset1 == offset2) {
400 if ((fha_ctls.max_reqs_per_nfsd == 0) ||
401 (req_count < fha_ctls.max_reqs_per_nfsd)) {
402 #if 0
403 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
404 "fha: %p(%d)r", thread, req_count);
405 #endif
406 return (thread);
407 }
408 }
409
410 /*
411 * We don't have a locality match, so skip this thread,
412 * but keep track of the most attractive thread in case
413 * we need to come back to it later.
414 */
415 #if 0
416 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
417 "fha: %p(%d)s off1 %llu off2 %llu", thread,
418 req_count, offset1, offset2);
419 #endif
420 if ((min_thread == NULL) || (req_count < min_count)) {
421 min_count = req_count;
422 min_thread = thread;
423 }
424 }
425
426 /*
427 * We didn't find a good match yet. See if we can add
428 * a new thread to this file handle entry's thread list.
429 */
430 if ((fha_ctls.max_nfsds_per_fh == 0) ||
431 (fhe->num_threads < fha_ctls.max_nfsds_per_fh)) {
432 /*
433 * We can add a new thread, so try for an idle thread
434 * first, and fall back to this_thread if none are idle.
435 */
436 if (STAILQ_EMPTY(&this_thread->st_reqs)) {
437 thread = this_thread;
438 #if 0
439 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
440 "fha: %p(%d)t", thread, thread->st_reqcount);
441 #endif
442 } else if ((thread = get_idle_thread(pool))) {
443 #if 0
444 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
445 "fha: %p(%d)i", thread, thread->st_reqcount);
446 #endif
447 } else {
448 thread = this_thread;
449 #if 0
450 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO,
451 "fha: %p(%d)b", thread, thread->st_reqcount);
452 #endif
453 }
454 fha_hash_entry_add_thread(fhe, thread);
455 } else {
456 /*
457 * We don't want to use any more threads for this file, so
458 * go back to the most attractive nfsd we're already using.
459 */
460 thread = min_thread;
461 }
462
463 return (thread);
464 }
465
466 /*
467 * After getting a request, try to assign it to some thread. Usually we
468 * handle it ourselves.
469 */
470 SVCTHREAD *
471 fha_assign(SVCTHREAD *this_thread, struct svc_req *req)
472 {
473 SVCPOOL *pool;
474 SVCTHREAD *thread;
475 struct fha_info i;
476 struct fha_hash_entry *fhe;
477
478 /*
479 * Only do placement if this is an NFS request.
480 */
481 if (req->rq_prog != NFS_PROG)
482 return (this_thread);
483
484 if (req->rq_vers != 2 && req->rq_vers != 3)
485 return (this_thread);
486
487 pool = req->rq_xprt->xp_pool;
488 fha_extract_info(req, &i);
489
490 /*
491 * We save the offset associated with this request for later
492 * nfsd matching.
493 */
494 fhe = fha_hash_entry_lookup(pool, i.fh);
495 req->rq_p1 = fhe;
496 req->rq_p2 = i.locktype;
497 req->rq_p3 = i.offset;
498
499 /*
500 * Choose a thread, taking into consideration locality, thread load,
501 * and the number of threads already working on this file.
502 */
503 thread = fha_hash_entry_choose_thread(pool, fhe, &i, this_thread);
504 KASSERT(thread, ("fha_assign: NULL thread!"));
505 fha_hash_entry_add_op(fhe, i.locktype, 1);
506
507 return (thread);
508 }
509
510 /*
511 * Called when we're done with an operation. The request has already
512 * been de-queued.
513 */
514 void
515 fha_nd_complete(SVCTHREAD *thread, struct svc_req *req)
516 {
517 struct fha_hash_entry *fhe = req->rq_p1;
518
519 /*
520 * This may be called for reqs that didn't go through
521 * fha_assign (e.g. extra NULL ops used for RPCSEC_GSS.
522 */
523 if (!fhe)
524 return;
525
526 fha_hash_entry_add_op(fhe, req->rq_p2, -1);
527
528 if (thread->st_reqcount == 0) {
529 fha_hash_entry_remove_thread(fhe, thread);
530 if (0 == fhe->num_reads + fhe->num_writes)
531 fha_hash_entry_remove(fhe);
532 }
533 }
534
535 extern SVCPOOL *nfsrv_pool;
536
537 static int
538 fhe_stats_sysctl(SYSCTL_HANDLER_ARGS)
539 {
540 int error, count, i;
541 struct sbuf sb;
542 struct fha_hash_entry *fhe;
543 bool_t first = TRUE;
544 SVCTHREAD *thread;
545
546 sbuf_new(&sb, NULL, 4096, SBUF_FIXEDLEN);
547
548 if (!nfsrv_pool) {
549 sbuf_printf(&sb, "NFSD not running\n");
550 goto out;
551 }
552
553 mtx_lock(&nfsrv_pool->sp_lock);
554 count = 0;
555 for (i = 0; i <= g_fha.hashmask; i++)
556 if (!LIST_EMPTY(&g_fha.hashtable[i]))
557 count++;
558
559 if (count == 0) {
560 sbuf_printf(&sb, "No file handle entries.\n");
561 goto out;
562 }
563
564 for (i = 0; i <= g_fha.hashmask; i++) {
565 LIST_FOREACH(fhe, &g_fha.hashtable[i], link) {
566 sbuf_printf(&sb, "%sfhe %p: {\n", first ? "" : ", ", fhe);
567
568 sbuf_printf(&sb, " fh: %ju\n", (uintmax_t) fhe->fh);
569 sbuf_printf(&sb, " num_reads: %d\n", fhe->num_reads);
570 sbuf_printf(&sb, " num_writes: %d\n", fhe->num_writes);
571 sbuf_printf(&sb, " num_threads: %d\n", fhe->num_threads);
572
573 LIST_FOREACH(thread, &fhe->threads, st_alink) {
574 sbuf_printf(&sb, " thread %p (count %d)\n",
575 thread, thread->st_reqcount);
576 }
577
578 sbuf_printf(&sb, "}");
579 first = FALSE;
580
581 /* Limit the output. */
582 if (++count > 128) {
583 sbuf_printf(&sb, "...");
584 break;
585 }
586 }
587 }
588
589 out:
590 if (nfsrv_pool)
591 mtx_unlock(&nfsrv_pool->sp_lock);
592 sbuf_trim(&sb);
593 sbuf_finish(&sb);
594 error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
595 sbuf_delete(&sb);
596 return (error);
597 }
Cache object: 10534b641a4e5e49b69045c65c724492
|