nfs_fha.c revision 261049
1146998Sdes/*- 2146998Sdes * Copyright (c) 2008 Isilon Inc http://www.isilon.com/ 3146998Sdes * 4146998Sdes * Redistribution and use in source and binary forms, with or without 5146998Sdes * modification, are permitted provided that the following conditions 6146998Sdes * are met: 7146998Sdes * 1. Redistributions of source code must retain the above copyright 8146998Sdes * notice, this list of conditions and the following disclaimer. 9146998Sdes * 2. Redistributions in binary form must reproduce the above copyright 10146998Sdes * notice, this list of conditions and the following disclaimer in the 11146998Sdes * documentation and/or other materials provided with the distribution. 12146998Sdes * 13146998Sdes * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14146998Sdes * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15146998Sdes * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16146998Sdes * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 17146998Sdes * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18146998Sdes * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19146998Sdes * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20146998Sdes * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21146998Sdes * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22146998Sdes * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23146998Sdes * SUCH DAMAGE. 24146998Sdes */ 25146998Sdes 26146998Sdes#include <sys/cdefs.h> 27146998Sdes__FBSDID("$FreeBSD: stable/10/sys/nfs/nfs_fha.c 261049 2014-01-22 23:48:15Z mav $"); 28146998Sdes 29146998Sdes#include <sys/param.h> 30146998Sdes#include <sys/systm.h> 31146998Sdes#include <sys/sysproto.h> 32146998Sdes#include <sys/kernel.h> 33146998Sdes#include <sys/sysctl.h> 34146998Sdes#include <sys/vnode.h> 35146998Sdes#include <sys/malloc.h> 36146998Sdes#include <sys/mount.h> 37146998Sdes#include <sys/mbuf.h> 38146998Sdes#include <sys/sbuf.h> 39146998Sdes 40146998Sdes#include <rpc/rpc.h> 41146998Sdes#include <nfs/nfs_fha.h> 42146998Sdes 43146998Sdesstatic MALLOC_DEFINE(M_NFS_FHA, "NFS FHA", "NFS FHA"); 44146998Sdes 45146998Sdes/* 46146998Sdes * XXX need to commonize definitions between old and new NFS code. Define 47146998Sdes * this here so we don't include one nfsproto.h over the other. 48146998Sdes */ 49146998Sdes#define NFS_PROG 100003 50146998Sdes 51146998Sdesvoid 52146998Sdesfha_init(struct fha_params *softc) 53146998Sdes{ 54146998Sdes char tmpstr[128]; 55146998Sdes 56146998Sdes /* 57146998Sdes * A small hash table to map filehandles to fha_hash_entry 58146998Sdes * structures. 59146998Sdes */ 60146998Sdes softc->g_fha.hashtable = hashinit(256, M_NFS_FHA, 61146998Sdes &softc->g_fha.hashmask); 62146998Sdes 63146998Sdes /* 64146998Sdes * Set the default tuning parameters. 65146998Sdes */ 66146998Sdes softc->ctls.enable = FHA_DEF_ENABLE; 67146998Sdes softc->ctls.bin_shift = FHA_DEF_BIN_SHIFT; 68146998Sdes softc->ctls.max_nfsds_per_fh = FHA_DEF_MAX_NFSDS_PER_FH; 69146998Sdes softc->ctls.max_reqs_per_nfsd = FHA_DEF_MAX_REQS_PER_NFSD; 70 71 /* 72 * Allow the user to override the defaults at boot time with 73 * tunables. 74 */ 75 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.enable", 76 softc->server_name); 77 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.enable); 78 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.bin_shift", 79 softc->server_name); 80 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.bin_shift); 81 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.max_nfsds_per_fh", 82 softc->server_name); 83 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.max_nfsds_per_fh); 84 snprintf(tmpstr, sizeof(tmpstr), "vfs.%s.fha.max_reqs_per_nfsd", 85 softc->server_name); 86 TUNABLE_INT_FETCH(tmpstr, &softc->ctls.max_reqs_per_nfsd); 87 88 /* 89 * Add sysctls so the user can change the tuning parameters at 90 * runtime. 91 */ 92 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 93 OID_AUTO, "enable", CTLFLAG_RW, 94 &softc->ctls.enable, 0, "Enable NFS File Handle Affinity (FHA)"); 95 96 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 97 OID_AUTO, "bin_shift", CTLFLAG_RW, 98 &softc->ctls.bin_shift, 0, "For FHA reads, no two requests will " 99 "contend if they're 2^(bin_shift) bytes apart"); 100 101 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 102 OID_AUTO, "max_nfsds_per_fh", CTLFLAG_RW, 103 &softc->ctls.max_nfsds_per_fh, 0, "Maximum nfsd threads that " 104 "should be working on requests for the same file handle"); 105 106 SYSCTL_ADD_UINT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 107 OID_AUTO, "max_reqs_per_nfsd", CTLFLAG_RW, 108 &softc->ctls.max_reqs_per_nfsd, 0, "Maximum requests that " 109 "single nfsd thread should be working on at any time"); 110 111 SYSCTL_ADD_OID(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 112 OID_AUTO, "fhe_stats", CTLTYPE_STRING | CTLFLAG_RD, 0, 0, 113 softc->callbacks.fhe_stats_sysctl, "A", ""); 114 115} 116 117void 118fha_uninit(struct fha_params *softc) 119{ 120 sysctl_ctx_free(&softc->sysctl_ctx); 121 hashdestroy(softc->g_fha.hashtable, M_NFS_FHA, softc->g_fha.hashmask); 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 if (e->num_rw + e->num_exclusive) 211 panic("nonempty fhe"); 212 free(e, M_NFS_FHA); 213} 214 215static void 216fha_hash_entry_remove(struct fha_hash_entry *e) 217{ 218 219 LIST_REMOVE(e, link); 220 fha_hash_entry_destroy(e); 221} 222 223static struct fha_hash_entry * 224fha_hash_entry_lookup(struct fha_params *softc, u_int64_t fh) 225{ 226 SVCPOOL *pool; 227 228 pool = *softc->pool; 229 230 struct fha_hash_entry *fhe, *new_fhe; 231 232 LIST_FOREACH(fhe, &softc->g_fha.hashtable[fh % softc->g_fha.hashmask], 233 link) 234 if (fhe->fh == fh) 235 break; 236 237 if (!fhe) { 238 /* Allocate a new entry. */ 239 mtx_unlock(&pool->sp_lock); 240 new_fhe = fha_hash_entry_new(fh); 241 mtx_lock(&pool->sp_lock); 242 243 /* Double-check to make sure we still need the new entry. */ 244 LIST_FOREACH(fhe, 245 &softc->g_fha.hashtable[fh % softc->g_fha.hashmask], link) 246 if (fhe->fh == fh) 247 break; 248 if (!fhe) { 249 fhe = new_fhe; 250 LIST_INSERT_HEAD( 251 &softc->g_fha.hashtable[fh % softc->g_fha.hashmask], 252 fhe, link); 253 } else 254 fha_hash_entry_destroy(new_fhe); 255 } 256 257 return (fhe); 258} 259 260static void 261fha_hash_entry_add_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread) 262{ 263 264 LIST_INSERT_HEAD(&fhe->threads, thread, st_alink); 265 fhe->num_threads++; 266} 267 268static void 269fha_hash_entry_remove_thread(struct fha_hash_entry *fhe, SVCTHREAD *thread) 270{ 271 272 LIST_REMOVE(thread, st_alink); 273 fhe->num_threads--; 274} 275 276/* 277 * Account for an ongoing operation associated with this file. 278 */ 279static void 280fha_hash_entry_add_op(struct fha_hash_entry *fhe, int locktype, int count) 281{ 282 283 if (LK_EXCLUSIVE == locktype) 284 fhe->num_exclusive += count; 285 else 286 fhe->num_rw += count; 287} 288 289/* 290 * Get the service thread currently associated with the fhe that is 291 * appropriate to handle this operation. 292 */ 293SVCTHREAD * 294fha_hash_entry_choose_thread(struct fha_params *softc, 295 struct fha_hash_entry *fhe, struct fha_info *i, SVCTHREAD *this_thread); 296 297SVCTHREAD * 298fha_hash_entry_choose_thread(struct fha_params *softc, 299 struct fha_hash_entry *fhe, struct fha_info *i, SVCTHREAD *this_thread) 300{ 301 SVCTHREAD *thread, *min_thread = NULL; 302 SVCPOOL *pool; 303 int req_count, min_count = 0; 304 off_t offset1, offset2; 305 306 pool = *softc->pool; 307 308 LIST_FOREACH(thread, &fhe->threads, st_alink) { 309 req_count = thread->st_reqcount; 310 311 /* If there are any writes in progress, use the first thread. */ 312 if (fhe->num_exclusive) { 313#if 0 314 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 315 "fha: %p(%d)w", thread, req_count); 316#endif 317 return (thread); 318 } 319 320 /* 321 * Check for read locality, making sure that we won't 322 * exceed our per-thread load limit in the process. 323 */ 324 offset1 = i->offset; 325 offset2 = STAILQ_FIRST(&thread->st_reqs)->rq_p3; 326 327 if (((offset1 >= offset2) 328 && ((offset1 - offset2) < (1 << softc->ctls.bin_shift))) 329 || ((offset2 > offset1) 330 && ((offset2 - offset1) < (1 << softc->ctls.bin_shift)))) { 331 if ((softc->ctls.max_reqs_per_nfsd == 0) || 332 (req_count < softc->ctls.max_reqs_per_nfsd)) { 333#if 0 334 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 335 "fha: %p(%d)r", thread, req_count); 336#endif 337 return (thread); 338 } 339 } 340 341 /* 342 * We don't have a locality match, so skip this thread, 343 * but keep track of the most attractive thread in case 344 * we need to come back to it later. 345 */ 346#if 0 347 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 348 "fha: %p(%d)s off1 %llu off2 %llu", thread, 349 req_count, offset1, offset2); 350#endif 351 if ((min_thread == NULL) || (req_count < min_count)) { 352 min_count = req_count; 353 min_thread = thread; 354 } 355 } 356 357 /* 358 * We didn't find a good match yet. See if we can add 359 * a new thread to this file handle entry's thread list. 360 */ 361 if ((softc->ctls.max_nfsds_per_fh == 0) || 362 (fhe->num_threads < softc->ctls.max_nfsds_per_fh)) { 363 /* 364 * We can add a new thread, so try for an idle thread 365 * first, and fall back to this_thread if none are idle. 366 */ 367 if (STAILQ_EMPTY(&this_thread->st_reqs)) { 368 thread = this_thread; 369#if 0 370 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 371 "fha: %p(%d)t", thread, thread->st_reqcount); 372#endif 373 } else if ((thread = LIST_FIRST(&pool->sp_idlethreads))) { 374#if 0 375 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 376 "fha: %p(%d)i", thread, thread->st_reqcount); 377#endif 378 } else { 379 thread = this_thread; 380#if 0 381 ITRACE_CURPROC(ITRACE_NFS, ITRACE_INFO, 382 "fha: %p(%d)b", thread, thread->st_reqcount); 383#endif 384 } 385 fha_hash_entry_add_thread(fhe, thread); 386 } else { 387 /* 388 * We don't want to use any more threads for this file, so 389 * go back to the most attractive nfsd we're already using. 390 */ 391 thread = min_thread; 392 } 393 394 return (thread); 395} 396 397/* 398 * After getting a request, try to assign it to some thread. Usually we 399 * handle it ourselves. 400 */ 401SVCTHREAD * 402fha_assign(SVCTHREAD *this_thread, struct svc_req *req, 403 struct fha_params *softc) 404{ 405 SVCTHREAD *thread; 406 struct fha_info i; 407 struct fha_hash_entry *fhe; 408 struct fha_callbacks *cb; 409 410 cb = &softc->callbacks; 411 412 /* Check to see whether we're enabled. */ 413 if (softc->ctls.enable == 0) 414 return (this_thread); 415 416 /* 417 * Only do placement if this is an NFS request. 418 */ 419 if (req->rq_prog != NFS_PROG) 420 return (this_thread); 421 422 if (req->rq_vers != 2 && req->rq_vers != 3) 423 return (this_thread); 424 425 fha_extract_info(req, &i, cb); 426 427 /* 428 * We save the offset associated with this request for later 429 * nfsd matching. 430 */ 431 fhe = fha_hash_entry_lookup(softc, i.fh); 432 req->rq_p1 = fhe; 433 req->rq_p2 = i.locktype; 434 req->rq_p3 = i.offset; 435 436 /* 437 * Choose a thread, taking into consideration locality, thread load, 438 * and the number of threads already working on this file. 439 */ 440 thread = fha_hash_entry_choose_thread(softc, fhe, &i, this_thread); 441 KASSERT(thread, ("fha_assign: NULL thread!")); 442 fha_hash_entry_add_op(fhe, i.locktype, 1); 443 444 return (thread); 445} 446 447/* 448 * Called when we're done with an operation. The request has already 449 * been de-queued. 450 */ 451void 452fha_nd_complete(SVCTHREAD *thread, struct svc_req *req) 453{ 454 struct fha_hash_entry *fhe = req->rq_p1; 455 456 /* 457 * This may be called for reqs that didn't go through 458 * fha_assign (e.g. extra NULL ops used for RPCSEC_GSS. 459 */ 460 if (!fhe) 461 return; 462 463 fha_hash_entry_add_op(fhe, req->rq_p2, -1); 464 465 if (thread->st_reqcount == 0) { 466 fha_hash_entry_remove_thread(fhe, thread); 467 if (0 == fhe->num_rw + fhe->num_exclusive) 468 fha_hash_entry_remove(fhe); 469 } 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 mtx_lock(&pool->sp_lock); 493 count = 0; 494 for (i = 0; i <= softc->g_fha.hashmask; i++) 495 if (!LIST_EMPTY(&softc->g_fha.hashtable[i])) 496 count++; 497 498 if (count == 0) { 499 sbuf_printf(&sb, "No file handle entries.\n"); 500 goto out; 501 } 502 503 for (i = 0; i <= softc->g_fha.hashmask; i++) { 504 LIST_FOREACH(fhe, &softc->g_fha.hashtable[i], 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 STAILQ_FIRST(&thread->st_reqs)->rq_p3, 516 thread->st_reqcount); 517 } 518 519 sbuf_printf(&sb, "}"); 520 first = FALSE; 521 522 /* Limit the output. */ 523 if (++count > 128) { 524 sbuf_printf(&sb, "..."); 525 break; 526 } 527 } 528 } 529 530 out: 531 if (pool) 532 mtx_unlock(&pool->sp_lock); 533 sbuf_trim(&sb); 534 sbuf_finish(&sb); 535 error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); 536 sbuf_delete(&sb); 537 return (error); 538} 539