nfs_fha.c revision 261054
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: stable/10/sys/nfs/nfs_fha.c 261054 2014-01-22 23:52:20Z mav $"); 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/nfs_fha.h> 42 43static MALLOC_DEFINE(M_NFS_FHA, "NFS FHA", "NFS FHA"); 44 45/* 46 * XXX need to commonize definitions between old and new NFS code. Define 47 * this here so we don't include one nfsproto.h over the other. 48 */ 49#define NFS_PROG 100003 50 51void 52fha_init(struct fha_params *softc) 53{ 54 char tmpstr[128]; 55 int i; 56 57 for (i = 0; i < FHA_HASH_SIZE; i++) 58 mtx_init(&softc->fha_hash[i].mtx, "fhalock", NULL, MTX_DEF); 59 60 /* 61 * Set the default tuning parameters. 62 */ 63 softc->ctls.enable = FHA_DEF_ENABLE; 64 softc->ctls.bin_shift = FHA_DEF_BIN_SHIFT; 65 softc->ctls.max_nfsds_per_fh = FHA_DEF_MAX_NFSDS_PER_FH; 66 softc->ctls.max_reqs_per_nfsd = FHA_DEF_MAX_REQS_PER_NFSD; 67 68 /* 69 * Allow the user to override the defaults at boot time with 70 * tunables. 71 */ 72 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.enable", 73 softc->server_name); 74 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.enable); 75 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.bin_shift", 76 softc->server_name); 77 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.bin_shift); 78 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.max_nfsds_per_fh", 79 softc->server_name); 80 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.max_nfsds_per_fh); 81 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.max_reqs_per_nfsd", 82 softc->server_name); 83 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.max_reqs_per_nfsd); 84 85 /* 86 * Add sysctls so the user can change the tuning parameters at 87 * runtime. 88 */ 89 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 90 OID_AUTO, "enable", CTLFLAG_RW, 91 &softc->ctls.enable, 0, "Enable NFS File Handle Affinity (FHA)"); 92 93 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 94 OID_AUTO, "bin_shift", CTLFLAG_RW, 95 &softc->ctls.bin_shift, 0, "For FHA reads, no two requests will " 96 "contend if they're 2^(bin_shift) bytes apart"); 97 98 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 99 OID_AUTO, "max_nfsds_per_fh", CTLFLAG_RW, 100 &softc->ctls.max_nfsds_per_fh, 0, "Maximum nfsd threads that " 101 "should be working on requests for the same file handle"); 102 103 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 104 OID_AUTO, "max_reqs_per_nfsd", CTLFLAG_RW, 105 &softc->ctls.max_reqs_per_nfsd, 0, "Maximum requests that " 106 "single nfsd thread should be working on at any time"); 107 108 SYSCTL_ADD_OID(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 109 OID_AUTO, "fhe_stats", CTLTYPE_STRING | CTLFLAG_RD, 0, 0, 110 softc->callbacks.fhe_stats_sysctl, "A", ""); 111 112} 113 114void 115fha_uninit(struct fha_params *softc) 116{ 117 int i; 118 119 sysctl_ctx_free(&softc->sysctl_ctx); 120 for (i = 0; i < FHA_HASH_SIZE; i++) 121 mtx_destroy(&softc->fha_hash[i].mtx); 122} 123 124/* 125 * This just specifies that offsets should obey affinity when within 126 * the same 1Mbyte (1<<20) chunk for the file (reads only for now). 127 */ 128static void 129fha_extract_info(struct svc_req *req, struct fha_info *i, 130 struct fha_callbacks *cb) 131{ 132 struct mbuf *md; 133 caddr_t dpos; 134 static u_int64_t random_fh = 0; 135 int error; 136 int v3 = (req->rq_vers == 3); 137 rpcproc_t procnum; 138 139 /* 140 * We start off with a random fh. If we get a reasonable 141 * procnum, we set the fh. If there's a concept of offset 142 * that we're interested in, we set that. 143 */ 144 i->fh = ++random_fh; 145 i->offset = 0; 146 i->locktype = LK_EXCLUSIVE; 147 148 /* 149 * Extract the procnum and convert to v3 form if necessary, 150 * taking care to deal with out-of-range procnums. Caller will 151 * ensure that rq_vers is either 2 or 3. 152 */ 153 procnum = req->rq_proc; 154 if (!v3) { 155 rpcproc_t tmp_procnum; 156 157 tmp_procnum = cb->get_procnum(procnum); 158 if (tmp_procnum == -1) 159 goto out; 160 procnum = tmp_procnum; 161 } 162 163 /* 164 * We do affinity for most. However, we divide a realm of affinity 165 * by file offset so as to allow for concurrent random access. We 166 * only do this for reads today, but this may change when IFS supports 167 * efficient concurrent writes. 168 */ 169 if (cb->no_offset(procnum)) 170 goto out; 171 172 error = cb->realign(&req->rq_args, M_NOWAIT); 173 if (error) 174 goto out; 175 md = req->rq_args; 176 dpos = mtod(md, caddr_t); 177 178 /* Grab the filehandle. */ 179 error = cb->get_fh(&i->fh, v3, &md, &dpos); 180 if (error) 181 goto out; 182 183 /* Content ourselves with zero offset for all but reads. */ 184 if (cb->is_read(procnum) || cb->is_write(procnum)) 185 cb->get_offset(&md, &dpos, v3, i); 186 187out: 188 cb->set_locktype(procnum, i); 189} 190 191static struct fha_hash_entry * 192fha_hash_entry_new(u_int64_t fh) 193{ 194 struct fha_hash_entry *e; 195 196 e = malloc(sizeof(*e), M_NFS_FHA, M_WAITOK); 197 e->fh = fh; 198 e->num_rw = 0; 199 e->num_exclusive = 0; 200 e->num_threads = 0; 201 LIST_INIT(&e->threads); 202 203 return (e); 204} 205 206static void 207fha_hash_entry_destroy(struct fha_hash_entry *e) 208{ 209 210 mtx_assert(e->mtx, MA_OWNED); 211 KASSERT(e->num_rw == 0, 212 ("%d reqs on destroyed fhe %p", e->num_rw, e)); 213 KASSERT(e->num_exclusive == 0, 214 ("%d exclusive reqs on destroyed fhe %p", e->num_exclusive, e)); 215 KASSERT(e->num_threads == 0, 216 ("%d threads on destroyed fhe %p", e->num_threads, e)); 217 free(e, M_NFS_FHA); 218} 219 220static void 221fha_hash_entry_remove(struct fha_hash_entry *e) 222{ 223 224 mtx_assert(e->mtx, MA_OWNED); 225 LIST_REMOVE(e, link); 226 fha_hash_entry_destroy(e); 227} 228 229static struct fha_hash_entry * 230fha_hash_entry_lookup(struct fha_params *softc, u_int64_t fh) 231{ 232 SVCPOOL *pool; 233 struct fha_hash_slot *fhs; 234 struct fha_hash_entry *fhe, *new_fhe; 235 236 pool = *softc->pool; 237 fhs = &softc->fha_hash[fh % FHA_HASH_SIZE]; 238 new_fhe = fha_hash_entry_new(fh); 239 new_fhe->mtx = &fhs->mtx; 240 mtx_lock(&fhs->mtx); 241 LIST_FOREACH(fhe, &fhs->list, link) 242 if (fhe->fh == fh) 243 break; 244 if (!fhe) { 245 fhe = new_fhe; 246 LIST_INSERT_HEAD(&fhs->list, fhe, link); 247 } else 248 fha_hash_entry_destroy(new_fhe); 249 return (fhe); 250} 251 252static void 253fha_hash_entry_add_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread) 254{ 255 256 mtx_assert(fhe->mtx, MA_OWNED); 257 thread->st_p2 = 0; 258 LIST_INSERT_HEAD(&fhe->threads, thread, st_alink); 259 fhe->num_threads++; 260} 261 262static void 263fha_hash_entry_remove_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread) 264{ 265 266 mtx_assert(fhe->mtx, MA_OWNED); 267 KASSERT(thread->st_p2 == 0, 268 ("%d reqs on removed thread %p", thread->st_p2, thread)); 269 LIST_REMOVE(thread, st_alink); 270 fhe->num_threads--; 271} 272 273/* 274 * Account for an ongoing operation associated with this file. 275 */ 276static void 277fha_hash_entry_add_op(struct fha_hash_entry *fhe, int locktype, int count) 278{ 279 280 mtx_assert(fhe->mtx, MA_OWNED); 281 if (LK_EXCLUSIVE == locktype) 282 fhe->num_exclusive += count; 283 else 284 fhe->num_rw += count; 285} 286 287/* 288 * Get the service thread currently associated with the fhe that is 289 * appropriate to handle this operation. 290 */ 291SVCTHREAD * 292fha_hash_entry_choose_thread(struct fha_params *softc, 293 struct fha_hash_entry *fhe, struct fha_info *i, SVCTHREAD *this_thread); 294 295SVCTHREAD * 296fha_hash_entry_choose_thread(struct fha_params *softc, 297 struct fha_hash_entry *fhe, struct fha_info *i, SVCTHREAD *this_thread) 298{ 299 SVCTHREAD *thread, *min_thread = NULL; 300 SVCPOOL *pool; 301 int req_count, min_count = 0; 302 off_t offset1, offset2; 303 304 pool = *softc->pool; 305 306 LIST_FOREACH(thread, &fhe->threads, st_alink) { 307 req_count = thread->st_p2; 308 309 /* If there are any writes in progress, use the first thread. */ 310 if (fhe->num_exclusive) { 311#if 0 312 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 313 "fha: %p(%d)w", thread, req_count); 314#endif 315 return (thread); 316 } 317 318 /* 319 * Check for read locality, making sure that we won't 320 * exceed our per-thread load limit in the process. 321 */ 322 offset1 = i->offset; 323 offset2 = thread->st_p3; 324 325 if (((offset1 >= offset2) 326 && ((offset1 - offset2) < (1 << softc->ctls.bin_shift))) 327 || ((offset2 > offset1) 328 && ((offset2 - offset1) < (1 << softc->ctls.bin_shift)))) { 329 if ((softc->ctls.max_reqs_per_nfsd == 0) || 330 (req_count < softc->ctls.max_reqs_per_nfsd)) { 331#if 0 332 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 333 "fha: %p(%d)r", thread, req_count); 334#endif 335 return (thread); 336 } 337 } 338 339 /* 340 * We don't have a locality match, so skip this thread, 341 * but keep track of the most attractive thread in case 342 * we need to come back to it later. 343 */ 344#if 0 345 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 346 "fha: %p(%d)s off1 %llu off2 %llu", thread, 347 req_count, offset1, offset2); 348#endif 349 if ((min_thread == NULL) || (req_count < min_count)) { 350 min_count = req_count; 351 min_thread = thread; 352 } 353 } 354 355 /* 356 * We didn't find a good match yet. See if we can add 357 * a new thread to this file handle entry's thread list. 358 */ 359 if ((softc->ctls.max_nfsds_per_fh == 0) || 360 (fhe->num_threads < softc->ctls.max_nfsds_per_fh)) { 361 thread = this_thread; 362#if 0 363 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 364 "fha: %p(%d)t", thread, thread->st_p2); 365#endif 366 fha_hash_entry_add_thread(fhe, thread); 367 } else { 368 /* 369 * We don't want to use any more threads for this file, so 370 * go back to the most attractive nfsd we're already using. 371 */ 372 thread = min_thread; 373 } 374 375 return (thread); 376} 377 378/* 379 * After getting a request, try to assign it to some thread. Usually we 380 * handle it ourselves. 381 */ 382SVCTHREAD * 383fha_assign(SVCTHREAD *this_thread, struct svc_req *req, 384 struct fha_params *softc) 385{ 386 SVCTHREAD *thread; 387 struct fha_info i; 388 struct fha_hash_entry *fhe; 389 struct fha_callbacks *cb; 390 391 cb = &softc->callbacks; 392 393 /* Check to see whether we're enabled. */ 394 if (softc->ctls.enable == 0) 395 goto thist; 396 397 /* 398 * Only do placement if this is an NFS request. 399 */ 400 if (req->rq_prog != NFS_PROG) 401 goto thist; 402 403 if (req->rq_vers != 2 && req->rq_vers != 3) 404 goto thist; 405 406 fha_extract_info(req, &i, cb); 407 408 /* 409 * We save the offset associated with this request for later 410 * nfsd matching. 411 */ 412 fhe = fha_hash_entry_lookup(softc, i.fh); 413 req->rq_p1 = fhe; 414 req->rq_p2 = i.locktype; 415 req->rq_p3 = i.offset; 416 417 /* 418 * Choose a thread, taking into consideration locality, thread load, 419 * and the number of threads already working on this file. 420 */ 421 thread = fha_hash_entry_choose_thread(softc, fhe, &i, this_thread); 422 KASSERT(thread, ("fha_assign: NULL thread!")); 423 fha_hash_entry_add_op(fhe, i.locktype, 1); 424 thread->st_p2++; 425 thread->st_p3 = i.offset; 426 427 /* 428 * Grab the pool lock here to not let chosen thread go away before 429 * the new request inserted to its queue while we drop fhe lock. 430 */ 431 mtx_lock(&(*softc->pool)->sp_lock); 432 mtx_unlock(fhe->mtx); 433 434 return (thread); 435thist: 436 req->rq_p1 = NULL; 437 mtx_lock(&(*softc->pool)->sp_lock); 438 return (this_thread); 439} 440 441/* 442 * Called when we're done with an operation. The request has already 443 * been de-queued. 444 */ 445void 446fha_nd_complete(SVCTHREAD *thread, struct svc_req *req) 447{ 448 struct fha_hash_entry *fhe = req->rq_p1; 449 struct mtx *mtx; 450 451 /* 452 * This may be called for reqs that didn't go through 453 * fha_assign (e.g. extra NULL ops used for RPCSEC_GSS. 454 */ 455 if (!fhe) 456 return; 457 458 mtx = fhe->mtx; 459 mtx_lock(mtx); 460 fha_hash_entry_add_op(fhe, req->rq_p2, -1); 461 thread->st_p2--; 462 KASSERT(thread->st_p2 >= 0, ("Negative request count %d on %p", 463 thread->st_p2, thread)); 464 if (thread->st_p2 == 0) { 465 fha_hash_entry_remove_thread(fhe, thread); 466 if (0 == fhe->num_rw + fhe->num_exclusive) 467 fha_hash_entry_remove(fhe); 468 } 469 mtx_unlock(mtx); 470} 471 472int 473fhe_stats_sysctl(SYSCTL_HANDLER_ARGS, struct fha_params *softc) 474{ 475 int error, count, i; 476 struct sbuf sb; 477 struct fha_hash_entry *fhe; 478 bool_t first = TRUE; 479 SVCTHREAD *thread; 480 SVCPOOL *pool; 481 482 sbuf_new(&sb, NULL, 4096, SBUF_FIXEDLEN); 483 484 pool = NULL; 485 486 if (!*softc->pool) { 487 sbuf_printf(&sb, "NFSD not running\n"); 488 goto out; 489 } 490 pool = *softc->pool; 491 492 count = 0; 493 for (i = 0; i < FHA_HASH_SIZE; i++) 494 if (!LIST_EMPTY(&softc->fha_hash[i].list)) 495 count++; 496 497 if (count == 0) { 498 sbuf_printf(&sb, "No file handle entries.\n"); 499 goto out; 500 } 501 502 for (i = 0; i < FHA_HASH_SIZE; i++) { 503 mtx_lock(&softc->fha_hash[i].mtx); 504 LIST_FOREACH(fhe, &softc->fha_hash[i].list, link) { 505 sbuf_printf(&sb, "%sfhe %p: {\n", first ? "" : ", ", fhe); 506 507 sbuf_printf(&sb, " fh: %ju\n", (uintmax_t) fhe->fh); 508 sbuf_printf(&sb, " num_rw: %d\n", fhe->num_rw); 509 sbuf_printf(&sb, " num_exclusive: %d\n", fhe->num_exclusive); 510 sbuf_printf(&sb, " num_threads: %d\n", fhe->num_threads); 511 512 LIST_FOREACH(thread, &fhe->threads, st_alink) { 513 sbuf_printf(&sb, " thread %p offset %ju " 514 "(count %d)\n", thread, 515 thread->st_p3, thread->st_p2); 516 } 517 518 sbuf_printf(&sb, "}"); 519 first = FALSE; 520 521 /* Limit the output. */ 522 if (++count > 128) { 523 sbuf_printf(&sb, "..."); 524 break; 525 } 526 } 527 mtx_unlock(&softc->fha_hash[i].mtx); 528 } 529 530 out: 531 sbuf_trim(&sb); 532 sbuf_finish(&sb); 533 error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); 534 sbuf_delete(&sb); 535 return (error); 536} 537