1// SPDX-License-Identifier: GPL-2.0 2/* 3 * GPL HEADER START 4 * 5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 6 * 7 * This program is free software; you can redistribute it and/or modify 8 * it under the terms of the GNU General Public License version 2 only, 9 * as published by the Free Software Foundation. 10 * 11 * This program is distributed in the hope that it will be useful, but 12 * WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 14 * General Public License version 2 for more details (a copy is included 15 * in the LICENSE file that accompanied this code). 16 * 17 * You should have received a copy of the GNU General Public License 18 * version 2 along with this program; If not, see 19 * http://www.gnu.org/licenses/gpl-2.0.html 20 * 21 * GPL HEADER END 22 */ 23/* 24 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. 25 * Use is subject to license terms. 26 * 27 * Copyright (c) 2010, 2017, Intel Corporation. 28 */ 29/* 30 * This file is part of Lustre, http://www.lustre.org/ 31 * Lustre is a trademark of Sun Microsystems, Inc. 32 * 33 * lustre/ldlm/ldlm_pool.c 34 * 35 * Author: Yury Umanets <umka@clusterfs.com> 36 */ 37 38/* 39 * Idea of this code is rather simple. Each second, for each server namespace 40 * we have SLV - server lock volume which is calculated on current number of 41 * granted locks, grant speed for past period, etc - that is, locking load. 42 * This SLV number may be thought as a flow definition for simplicity. It is 43 * sent to clients with each occasion to let them know what is current load 44 * situation on the server. By default, at the beginning, SLV on server is 45 * set max value which is calculated as the following: allow to one client 46 * have all locks of limit ->pl_limit for 10h. 47 * 48 * Next, on clients, number of cached locks is not limited artificially in any 49 * way as it was before. Instead, client calculates CLV, that is, client lock 50 * volume for each lock and compares it with last SLV from the server. CLV is 51 * calculated as the number of locks in LRU * lock live time in seconds. If 52 * CLV > SLV - lock is canceled. 53 * 54 * Client has LVF, that is, lock volume factor which regulates how much 55 * sensitive client should be about last SLV from server. The higher LVF is the 56 * more locks will be canceled on client. Default value for it is 1. Setting LVF 57 * to 2 means that client will cancel locks 2 times faster. 58 * 59 * Locks on a client will be canceled more intensively in these cases: 60 * (1) if SLV is smaller, that is, load is higher on the server; 61 * (2) client has a lot of locks (the more locks are held by client, the bigger 62 * chances that some of them should be canceled); 63 * (3) client has old locks (taken some time ago); 64 * 65 * Thus, according to flow paradigm that we use for better understanding SLV, 66 * CLV is the volume of particle in flow described by SLV. According to this, 67 * if flow is getting thinner, more and more particles become outside of it and 68 * as particles are locks, they should be canceled. 69 * 70 * General idea of this belongs to Vitaly Fertman (vitaly@clusterfs.com). 71 * Andreas Dilger (adilger@clusterfs.com) proposed few nice ideas like using 72 * LVF and many cleanups. Flow definition to allow more easy understanding of 73 * the logic belongs to Nikita Danilov (nikita@clusterfs.com) as well as many 74 * cleanups and fixes. And design and implementation are done by Yury Umanets 75 * (umka@clusterfs.com). 76 * 77 * Glossary for terms used: 78 * 79 * pl_limit - Number of allowed locks in pool. Applies to server and client 80 * side (tunable); 81 * 82 * pl_granted - Number of granted locks (calculated); 83 * pl_grant_rate - Number of granted locks for last T (calculated); 84 * pl_cancel_rate - Number of canceled locks for last T (calculated); 85 * pl_grant_speed - Grant speed (GR - CR) for last T (calculated); 86 * pl_grant_plan - Planned number of granted locks for next T (calculated); 87 * pl_server_lock_volume - Current server lock volume (calculated); 88 * 89 * As it may be seen from list above, we have few possible tunables which may 90 * affect behavior much. They all may be modified via sysfs. However, they also 91 * give a possibility for constructing few pre-defined behavior policies. If 92 * none of predefines is suitable for a working pattern being used, new one may 93 * be "constructed" via sysfs tunables. 94 */ 95 96#define DEBUG_SUBSYSTEM S_LDLM 97 98#include <lustre_dlm.h> 99#include <cl_object.h> 100#include <obd_class.h> 101#include <obd_support.h> 102#include "ldlm_internal.h" 103 104/* 105 * 50 ldlm locks for 1MB of RAM. 106 */ 107#define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_SHIFT)) * 50) 108 109/* 110 * Maximal possible grant step plan in %. 111 */ 112#define LDLM_POOL_MAX_GSP (30) 113 114/* 115 * Minimal possible grant step plan in %. 116 */ 117#define LDLM_POOL_MIN_GSP (1) 118 119/* 120 * This controls the speed of reaching LDLM_POOL_MAX_GSP 121 * with increasing thread period. 122 */ 123#define LDLM_POOL_GSP_STEP_SHIFT (2) 124 125/* 126 * LDLM_POOL_GSP% of all locks is default GP. 127 */ 128#define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100) 129 130/* 131 * Max age for locks on clients. 132 */ 133#define LDLM_POOL_MAX_AGE (36000) 134 135/* 136 * The granularity of SLV calculation. 137 */ 138#define LDLM_POOL_SLV_SHIFT (10) 139 140static inline u64 dru(u64 val, u32 shift, int round_up) 141{ 142 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift; 143} 144 145static inline u64 ldlm_pool_slv_max(u32 L) 146{ 147 /* 148 * Allow to have all locks for 1 client for 10 hrs. 149 * Formula is the following: limit * 10h / 1 client. 150 */ 151 u64 lim = (u64)L * LDLM_POOL_MAX_AGE / 1; 152 return lim; 153} 154 155static inline u64 ldlm_pool_slv_min(u32 L) 156{ 157 return 1; 158} 159 160enum { 161 LDLM_POOL_FIRST_STAT = 0, 162 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT, 163 LDLM_POOL_GRANT_STAT, 164 LDLM_POOL_CANCEL_STAT, 165 LDLM_POOL_GRANT_RATE_STAT, 166 LDLM_POOL_CANCEL_RATE_STAT, 167 LDLM_POOL_GRANT_PLAN_STAT, 168 LDLM_POOL_SLV_STAT, 169 LDLM_POOL_SHRINK_REQTD_STAT, 170 LDLM_POOL_SHRINK_FREED_STAT, 171 LDLM_POOL_RECALC_STAT, 172 LDLM_POOL_TIMING_STAT, 173 LDLM_POOL_LAST_STAT 174}; 175 176/** 177 * Calculates suggested grant_step in % of available locks for passed 178 * @period. This is later used in grant_plan calculations. 179 */ 180static inline int ldlm_pool_t2gsp(unsigned int t) 181{ 182 /* 183 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP 184 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP. 185 * 186 * How this will affect execution is the following: 187 * 188 * - for thread period 1s we will have grant_step 1% which good from 189 * pov of taking some load off from server and push it out to clients. 190 * This is like that because 1% for grant_step means that server will 191 * not allow clients to get lots of locks in short period of time and 192 * keep all old locks in their caches. Clients will always have to 193 * get some locks back if they want to take some new; 194 * 195 * - for thread period 10s (which is default) we will have 23% which 196 * means that clients will have enough of room to take some new locks 197 * without getting some back. All locks from this 23% which were not 198 * taken by clients in current period will contribute in SLV growing. 199 * SLV growing means more locks cached on clients until limit or grant 200 * plan is reached. 201 */ 202 return LDLM_POOL_MAX_GSP - 203 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >> 204 (t >> LDLM_POOL_GSP_STEP_SHIFT)); 205} 206 207/** 208 * Recalculates next stats on passed @pl. 209 * 210 * \pre ->pl_lock is locked. 211 */ 212static void ldlm_pool_recalc_stats(struct ldlm_pool *pl, timeout_t period) 213{ 214 int grant_plan = pl->pl_grant_plan; 215 u64 slv = pl->pl_server_lock_volume; 216 int granted = atomic_read(&pl->pl_granted); 217 int grant_rate = atomic_read(&pl->pl_grant_rate) / period; 218 int cancel_rate = atomic_read(&pl->pl_cancel_rate) / period; 219 220 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, 221 slv); 222 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT, 223 granted); 224 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, 225 grant_rate); 226 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, 227 grant_plan); 228 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, 229 cancel_rate); 230} 231 232/** 233 * Sets SLV and Limit from container_of(pl, struct ldlm_namespace, 234 * ns_pool)->ns_obd tp passed @pl. 235 */ 236static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl) 237{ 238 struct obd_device *obd; 239 240 /* 241 * Get new SLV and Limit from obd which is updated with coming 242 * RPCs. 243 */ 244 obd = container_of(pl, struct ldlm_namespace, 245 ns_pool)->ns_obd; 246 read_lock(&obd->obd_pool_lock); 247 pl->pl_server_lock_volume = obd->obd_pool_slv; 248 atomic_set(&pl->pl_limit, obd->obd_pool_limit); 249 read_unlock(&obd->obd_pool_lock); 250} 251 252/** 253 * Recalculates client size pool @pl according to current SLV and Limit. 254 */ 255static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) 256{ 257 timeout_t recalc_interval_sec; 258 int ret; 259 260 recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; 261 if (recalc_interval_sec < pl->pl_recalc_period) 262 return 0; 263 264 spin_lock(&pl->pl_lock); 265 /* 266 * Check if we need to recalc lists now. 267 */ 268 recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; 269 if (recalc_interval_sec < pl->pl_recalc_period) { 270 spin_unlock(&pl->pl_lock); 271 return 0; 272<<<<<<< found 273 } 274||||||| expected 275 */ 276 ldlm_pool_recalc_grant_plan(pl); 277 278 pl->pl_recalc_time = ktime_get_real_seconds(); 279 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, 280 recalc_interval_sec); 281 spin_unlock(&pl->pl_lock); 282======= 283 */ 284 ldlm_pool_recalc_grant_plan(pl); 285 286 pl->pl_recalc_time = ktime_get_seconds(); 287 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, 288 recalc_interval_sec); 289 spin_unlock(&pl->pl_lock); 290>>>>>>> replacement 291<<<<<<< found 292||||||| expected 293 */ 294static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) 295{ 296 time64_t recalc_interval_sec; 297 int ret; 298======= 299 */ 300static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) 301{ 302 timeout_t recalc_interval_sec; 303 int ret; 304>>>>>>> replacement 305 306<<<<<<< found 307 /* 308 * Make sure that pool knows last SLV and Limit from obd. 309 */ 310 ldlm_cli_pool_pop_slv(pl); 311 312 spin_unlock(&pl->pl_lock); 313||||||| expected 314 ENTRY; 315 316 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; 317 if (recalc_interval_sec < pl->pl_recalc_period) 318 return 0; 319 320======= 321 ENTRY; 322 323 recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; 324 if (recalc_interval_sec < pl->pl_recalc_period) 325 return 0; 326 327>>>>>>> replacement 328<<<<<<< found 329||||||| expected 330 /* 331 * Check if we need to recalc lists now. 332 */ 333 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; 334 if (recalc_interval_sec < pl->pl_recalc_period) { 335 spin_unlock(&pl->pl_lock); 336======= 337 /* 338 * Check if we need to recalc lists now. 339 */ 340 recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; 341 if (recalc_interval_sec < pl->pl_recalc_period) { 342 spin_unlock(&pl->pl_lock); 343>>>>>>> replacement 344 345 /* 346 * In the time of canceling locks on client we do not need to maintain 347 * sharp timing, we only want to cancel locks asap according to new SLV. 348 * It may be called when SLV has changed much, this is why we do not 349 * take into account pl->pl_recalc_time here. 350 */ 351 ret = ldlm_cancel_lru(container_of(pl, struct ldlm_namespace, ns_pool), 352 0, LCF_ASYNC, 0); 353 354 spin_lock(&pl->pl_lock); 355 /* 356 * Time of LRU resizing might be longer than period, 357 * so update after LRU resizing rather than before it. 358 */ 359 pl->pl_recalc_time = ktime_get_seconds(); 360 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, 361 recalc_interval_sec); 362 spin_unlock(&pl->pl_lock); 363 return ret; 364} 365 366/** 367 * This function is main entry point for memory pressure handling on client 368 * side. Main goal of this function is to cancel some number of locks on 369 * passed @pl according to @nr and @gfp_mask. 370 */ 371static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, 372 int nr, gfp_t gfp_mask) 373{ 374 struct ldlm_namespace *ns; 375 int unused; 376 377 ns = container_of(pl, struct ldlm_namespace, ns_pool); 378 379 /* 380 * Do not cancel locks in case lru resize is disabled for this ns. 381 */ 382 if (!ns_connect_lru_resize(ns)) 383 return 0; 384 385 /* 386 * Make sure that pool knows last SLV and Limit from obd. 387 */ 388 spin_lock(&pl->pl_lock); 389 ldlm_cli_pool_pop_slv(pl); 390 spin_unlock(&pl->pl_lock); 391 392 spin_lock(&ns->ns_lock); 393 unused = ns->ns_nr_unused; 394 spin_unlock(&ns->ns_lock); 395 396 if (nr == 0) 397 return (unused / 100) * sysctl_vfs_cache_pressure; 398 else 399 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, 0); 400} 401 402static const struct ldlm_pool_ops ldlm_cli_pool_ops = { 403 .po_recalc = ldlm_cli_pool_recalc, 404 .po_shrink = ldlm_cli_pool_shrink 405}; 406 407/** 408 * Pool recalc wrapper. Will call either client or server pool recalc callback 409 * depending what pool @pl is used. 410 * 411 * \retval time in seconds for the next recalc of this pool 412 */ 413<<<<<<< found 414static int ldlm_pool_recalc(struct ldlm_pool *pl) 415{ 416 u32 recalc_interval_sec; 417||||||| expected 418time64_t ldlm_pool_recalc(struct ldlm_pool *pl) 419{ 420 time64_t recalc_interval_sec; 421======= 422time64_t ldlm_pool_recalc(struct ldlm_pool *pl) 423{ 424 timeout_t recalc_interval_sec; 425>>>>>>> replacement 426 int count; 427 428 recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; 429 if (recalc_interval_sec > 0) { 430 spin_lock(&pl->pl_lock); 431<<<<<<< found 432 recalc_interval_sec = ktime_get_real_seconds() - 433 pl->pl_recalc_time; 434||||||| expected 435 recalc_interval_sec = ktime_get_real_seconds() - 436 pl->pl_recalc_time; 437======= 438 recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; 439>>>>>>> replacement 440 441 if (recalc_interval_sec > 0) { 442 /* 443 * Update pool statistics every recalc interval. 444 */ 445 ldlm_pool_recalc_stats(pl, recalc_interval_sec); 446 447 /* 448 * Zero out all rates and speed for the last period. 449 */ 450 atomic_set(&pl->pl_grant_rate, 0); 451 atomic_set(&pl->pl_cancel_rate, 0); 452 } 453 spin_unlock(&pl->pl_lock); 454 } 455 456 if (pl->pl_ops->po_recalc) { 457 count = pl->pl_ops->po_recalc(pl); 458 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT, 459 count); 460 } 461 462<<<<<<< found 463 recalc_interval_sec = pl->pl_recalc_time - ktime_get_real_seconds() + 464 pl->pl_recalc_period; 465 if (recalc_interval_sec <= 0) { 466 /* DEBUG: should be re-removed after LU-4536 is fixed */ 467 CDEBUG(D_DLMTRACE, 468 "%s: Negative interval(%ld), too short period(%ld)\n", 469 pl->pl_name, (long)recalc_interval_sec, 470 (long)pl->pl_recalc_period); 471 472 /* Prevent too frequent recalculation. */ 473 recalc_interval_sec = 1; 474 } 475 476 return recalc_interval_sec; 477||||||| expected 478 recalc_interval_sec = pl->pl_recalc_time - ktime_get_real_seconds() + 479 pl->pl_recalc_period; 480 if (recalc_interval_sec <= 0) { 481 /* DEBUG: should be re-removed after LU-4536 is fixed */ 482 CDEBUG(D_DLMTRACE, "%s: Negative interval(%lld), too short period(%lld)\n", 483 pl->pl_name, recalc_interval_sec, 484 (s64)pl->pl_recalc_period); 485 486 /* Prevent too frequent recalculation. */ 487 recalc_interval_sec = 1; 488 } 489 490 return recalc_interval_sec; 491======= 492 return pl->pl_recalc_time + pl->pl_recalc_period; 493>>>>>>> replacement 494} 495 496/* 497 * Pool shrink wrapper. Will call either client or server pool recalc callback 498 * depending what pool pl is used. When nr == 0, just return the number of 499 * freeable locks. Otherwise, return the number of canceled locks. 500 */ 501static int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask) 502{ 503 int cancel = 0; 504 505 if (pl->pl_ops->po_shrink) { 506 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); 507 if (nr > 0) { 508 lprocfs_counter_add(pl->pl_stats, 509 LDLM_POOL_SHRINK_REQTD_STAT, 510 nr); 511 lprocfs_counter_add(pl->pl_stats, 512 LDLM_POOL_SHRINK_FREED_STAT, 513 cancel); 514 CDEBUG(D_DLMTRACE, 515 "%s: request to shrink %d locks, shrunk %d\n", 516 pl->pl_name, nr, cancel); 517 } 518 } 519 return cancel; 520} 521 522static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused) 523{ 524 int granted, grant_rate, cancel_rate; 525 int grant_speed, lvf; 526 struct ldlm_pool *pl = m->private; 527 timeout_t period; 528 u64 slv, clv; 529 u32 limit; 530 531 spin_lock(&pl->pl_lock); 532 slv = pl->pl_server_lock_volume; 533 clv = pl->pl_client_lock_volume; 534 limit = atomic_read(&pl->pl_limit); 535 granted = atomic_read(&pl->pl_granted); 536 period = ktime_get_seconds() - pl->pl_recalc_time; 537 if (period <= 0) 538 period = 1; 539 grant_rate = atomic_read(&pl->pl_grant_rate) / period; 540 cancel_rate = atomic_read(&pl->pl_cancel_rate) / period; 541 grant_speed = grant_rate - cancel_rate; 542 lvf = atomic_read(&pl->pl_lock_volume_factor); 543 spin_unlock(&pl->pl_lock); 544 545 seq_printf(m, "LDLM pool state (%s):\n" 546 " SLV: %llu\n" 547 " CLV: %llu\n" 548 " LVF: %d\n", 549 pl->pl_name, slv, clv, (lvf * 100) >> 8); 550 551 seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n" 552 " G: %d\n L: %d\n", 553 grant_rate, cancel_rate, grant_speed, 554 granted, limit); 555 556 return 0; 557} 558 559LDEBUGFS_SEQ_FOPS_RO(lprocfs_pool_state); 560 561static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr, 562 char *buf) 563{ 564 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, 565 pl_kobj); 566 int grant_speed; 567 timeout_t period; 568 569 spin_lock(&pl->pl_lock); 570 /* serialize with ldlm_pool_recalc */ 571 period = ktime_get_seconds() - pl->pl_recalc_time; 572 if (period <= 0) 573 period = 1; 574 grant_speed = (atomic_read(&pl->pl_grant_rate) - 575 atomic_read(&pl->pl_cancel_rate)) / period; 576 spin_unlock(&pl->pl_lock); 577 return sprintf(buf, "%d\n", grant_speed); 578} 579LUSTRE_RO_ATTR(grant_speed); 580 581LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int); 582LUSTRE_RO_ATTR(grant_plan); 583 584LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int); 585LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int); 586LUSTRE_RW_ATTR(recalc_period); 587 588LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64); 589LUSTRE_RO_ATTR(server_lock_volume); 590 591LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(client_lock_volume, u64); 592LUSTRE_RO_ATTR(client_lock_volume); 593 594LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic); 595LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic); 596LUSTRE_RW_ATTR(limit); 597 598LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic); 599LUSTRE_RO_ATTR(granted); 600 601LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic); 602LUSTRE_RO_ATTR(cancel_rate); 603 604LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic); 605LUSTRE_RO_ATTR(grant_rate); 606 607static ssize_t lock_volume_factor_show(struct kobject *kobj, 608 struct attribute *attr, 609 char *buf) 610{ 611 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, pl_kobj); 612 unsigned long tmp; 613 614 tmp = (atomic_read(&pl->pl_lock_volume_factor) * 100) >> 8; 615 return sprintf(buf, "%lu\n", tmp); 616} 617 618static ssize_t lock_volume_factor_store(struct kobject *kobj, 619 struct attribute *attr, 620 const char *buffer, 621 size_t count) 622{ 623 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, pl_kobj); 624 unsigned long tmp; 625 int rc; 626 627 rc = kstrtoul(buffer, 10, &tmp); 628 if (rc < 0) { 629 return rc; 630 } 631 632 tmp = (tmp << 8) / 100; 633 atomic_set(&pl->pl_lock_volume_factor, tmp); 634 635 return count; 636 637} 638LUSTRE_RW_ATTR(lock_volume_factor); 639 640static ssize_t recalc_time_show(struct kobject *kobj, 641 struct attribute *attr, 642 char *buf) 643{ 644 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, pl_kobj); 645 646 return snprintf(buf, PAGE_SIZE, "%llu\n", 647 ktime_get_seconds() - pl->pl_recalc_time); 648} 649LUSTRE_RO_ATTR(recalc_time); 650 651/* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */ 652static struct attribute *ldlm_pl_attrs[] = { 653 &lustre_attr_grant_speed.attr, 654 &lustre_attr_grant_plan.attr, 655 &lustre_attr_recalc_period.attr, 656 &lustre_attr_server_lock_volume.attr, 657 &lustre_attr_client_lock_volume.attr, 658 &lustre_attr_recalc_time.attr, 659 &lustre_attr_limit.attr, 660 &lustre_attr_granted.attr, 661 &lustre_attr_cancel_rate.attr, 662 &lustre_attr_grant_rate.attr, 663 &lustre_attr_lock_volume_factor.attr, 664 NULL, 665}; 666 667static void ldlm_pl_release(struct kobject *kobj) 668{ 669 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, 670 pl_kobj); 671 complete(&pl->pl_kobj_unregister); 672} 673 674static struct kobj_type ldlm_pl_ktype = { 675 .default_attrs = ldlm_pl_attrs, 676 .sysfs_ops = &lustre_sysfs_ops, 677 .release = ldlm_pl_release, 678}; 679 680static int ldlm_pool_sysfs_init(struct ldlm_pool *pl) 681{ 682 struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, 683 ns_pool); 684 int err; 685 686 init_completion(&pl->pl_kobj_unregister); 687 err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj, 688 "pool"); 689 690 return err; 691} 692 693static int ldlm_pool_debugfs_init(struct ldlm_pool *pl) 694{ 695 struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, 696 ns_pool); 697 struct dentry *debugfs_ns_parent; 698 struct ldebugfs_vars pool_vars[2]; 699 int rc = 0; 700 701 debugfs_ns_parent = ns->ns_debugfs_entry; 702 if (IS_ERR_OR_NULL(debugfs_ns_parent)) { 703 CERROR("%s: debugfs entry is not initialized\n", 704 ldlm_ns_name(ns)); 705 rc = -EINVAL; 706 goto out; 707 } 708 pl->pl_debugfs_entry = debugfs_create_dir("pool", debugfs_ns_parent); 709 710 memset(pool_vars, 0, sizeof(pool_vars)); 711 712 ldlm_add_var(&pool_vars[0], pl->pl_debugfs_entry, "state", pl, 713 &lprocfs_pool_state_fops); 714 715 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT - 716 LDLM_POOL_FIRST_STAT, 0); 717 if (!pl->pl_stats) { 718 rc = -ENOMEM; 719 goto out; 720 } 721 722 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, 723 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 724 "granted", "locks"); 725 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, 726 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 727 "grant", "locks"); 728 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, 729 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 730 "cancel", "locks"); 731 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, 732 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 733 "grant_rate", "locks/s"); 734 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, 735 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 736 "cancel_rate", "locks/s"); 737 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, 738 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 739 "grant_plan", "locks/s"); 740 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT, 741 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 742 "slv", "slv"); 743 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT, 744 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 745 "shrink_request", "locks"); 746 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT, 747 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 748 "shrink_freed", "locks"); 749 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT, 750 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 751 "recalc_freed", "locks"); 752 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT, 753 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, 754 "recalc_timing", "sec"); 755 debugfs_create_file("stats", 0644, pl->pl_debugfs_entry, pl->pl_stats, 756 &ldebugfs_stats_seq_fops); 757 758out: 759 return rc; 760} 761 762static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl) 763{ 764 kobject_put(&pl->pl_kobj); 765 wait_for_completion(&pl->pl_kobj_unregister); 766} 767 768static void ldlm_pool_debugfs_fini(struct ldlm_pool *pl) 769{ 770 if (pl->pl_stats) { 771 lprocfs_free_stats(&pl->pl_stats); 772 pl->pl_stats = NULL; 773 } 774 debugfs_remove_recursive(pl->pl_debugfs_entry); 775} 776 777int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, 778 int idx, enum ldlm_side client) 779{ 780 int rc; 781 782 spin_lock_init(&pl->pl_lock); 783 atomic_set(&pl->pl_granted, 0); 784 pl->pl_recalc_time = ktime_get_seconds(); 785 atomic_set(&pl->pl_lock_volume_factor, 1 << 8); 786 787 atomic_set(&pl->pl_grant_rate, 0); 788 atomic_set(&pl->pl_cancel_rate, 0); 789 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L); 790 791 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", 792 ldlm_ns_name(ns), idx); 793 794 atomic_set(&pl->pl_limit, 1); 795 pl->pl_server_lock_volume = 0; 796 pl->pl_ops = &ldlm_cli_pool_ops; 797 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD; 798 pl->pl_client_lock_volume = 0; 799 rc = ldlm_pool_debugfs_init(pl); 800 if (rc) 801 return rc; 802 803 rc = ldlm_pool_sysfs_init(pl); 804 if (rc) 805 return rc; 806 807 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name); 808 809 return rc; 810} 811 812void ldlm_pool_fini(struct ldlm_pool *pl) 813{ 814 ldlm_pool_sysfs_fini(pl); 815 ldlm_pool_debugfs_fini(pl); 816 817 /* 818 * Pool should not be used after this point. We can't free it here as 819 * it lives in struct ldlm_namespace, but still interested in catching 820 * any abnormal using cases. 821 */ 822 POISON(pl, 0x5a, sizeof(*pl)); 823} 824 825/** 826 * Add new taken ldlm lock @lock into pool @pl accounting. 827 */ 828void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) 829{ 830 /* 831 * FLOCK locks are special in a sense that they are almost never 832 * cancelled, instead special kind of lock is used to drop them. 833 * also there is no LRU for flock locks, so no point in tracking 834 * them anyway. 835 */ 836 if (lock->l_resource->lr_type == LDLM_FLOCK) 837 return; 838 839 atomic_inc(&pl->pl_granted); 840 atomic_inc(&pl->pl_grant_rate); 841 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); 842 /* 843 * Do not do pool recalc for client side as all locks which 844 * potentially may be canceled has already been packed into 845 * enqueue/cancel rpc. Also we do not want to run out of stack 846 * with too long call paths. 847 */ 848} 849 850/** 851 * Remove ldlm lock @lock from pool @pl accounting. 852 */ 853void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock) 854{ 855 /* 856 * Filter out FLOCK locks. Read above comment in ldlm_pool_add(). 857 */ 858 if (lock->l_resource->lr_type == LDLM_FLOCK) 859 return; 860 861 LASSERT(atomic_read(&pl->pl_granted) > 0); 862 atomic_dec(&pl->pl_granted); 863 atomic_inc(&pl->pl_cancel_rate); 864 865 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); 866} 867 868/** 869 * Returns current @pl SLV. 870 * 871 * \pre ->pl_lock is not locked. 872 */ 873u64 ldlm_pool_get_slv(struct ldlm_pool *pl) 874{ 875 u64 slv; 876 877 spin_lock(&pl->pl_lock); 878 slv = pl->pl_server_lock_volume; 879 spin_unlock(&pl->pl_lock); 880 return slv; 881} 882 883/** 884 * Sets passed @clv to @pl. 885 * 886 * \pre ->pl_lock is not locked. 887 */ 888void ldlm_pool_set_clv(struct ldlm_pool *pl, u64 clv) 889{ 890 spin_lock(&pl->pl_lock); 891 pl->pl_client_lock_volume = clv; 892 spin_unlock(&pl->pl_lock); 893} 894 895/** 896 * Returns current LVF from @pl. 897 */ 898u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) 899{ 900 return atomic_read(&pl->pl_lock_volume_factor); 901} 902 903static int ldlm_pool_granted(struct ldlm_pool *pl) 904{ 905 return atomic_read(&pl->pl_granted); 906} 907 908/* 909 * count locks from all namespaces (if possible). Returns number of 910 * cached locks. 911 */ 912static unsigned long ldlm_pools_count(enum ldlm_side client, gfp_t gfp_mask) 913{ 914 unsigned long total = 0; 915 int nr_ns; 916 struct ldlm_namespace *ns; 917 struct ldlm_namespace *ns_old = NULL; /* loop detection */ 918 919 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) 920 return 0; 921 922 /* 923 * Find out how many resources we may release. 924 */ 925 for (nr_ns = ldlm_namespace_nr_read(client); 926 nr_ns > 0; nr_ns--) { 927 mutex_lock(ldlm_namespace_lock(client)); 928 if (list_empty(ldlm_namespace_list(client))) { 929 mutex_unlock(ldlm_namespace_lock(client)); 930 return 0; 931 } 932 ns = ldlm_namespace_first_locked(client); 933 934 if (ns == ns_old) { 935 mutex_unlock(ldlm_namespace_lock(client)); 936 break; 937 } 938 939 if (ldlm_ns_empty(ns)) { 940 ldlm_namespace_move_to_inactive_locked(ns, client); 941 mutex_unlock(ldlm_namespace_lock(client)); 942 continue; 943 } 944 945 if (!ns_old) 946 ns_old = ns; 947 948 ldlm_namespace_get(ns); 949 ldlm_namespace_move_to_active_locked(ns, client); 950 mutex_unlock(ldlm_namespace_lock(client)); 951 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); 952 ldlm_namespace_put(ns); 953 } 954 955 return total; 956} 957 958static unsigned long ldlm_pools_scan(enum ldlm_side client, int nr, 959 gfp_t gfp_mask) 960{ 961 unsigned long freed = 0; 962 int tmp, nr_ns; 963 struct ldlm_namespace *ns; 964 965 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) 966 return -1; 967 968 /* 969 * Shrink at least ldlm_namespace_nr_read(client) namespaces. 970 */ 971 for (tmp = nr_ns = ldlm_namespace_nr_read(client); 972 tmp > 0; tmp--) { 973 int cancel, nr_locks; 974 975 /* 976 * Do not call shrink under ldlm_namespace_lock(client) 977 */ 978 mutex_lock(ldlm_namespace_lock(client)); 979 if (list_empty(ldlm_namespace_list(client))) { 980 mutex_unlock(ldlm_namespace_lock(client)); 981 break; 982 } 983 ns = ldlm_namespace_first_locked(client); 984 ldlm_namespace_get(ns); 985 ldlm_namespace_move_to_active_locked(ns, client); 986 mutex_unlock(ldlm_namespace_lock(client)); 987 988 nr_locks = ldlm_pool_granted(&ns->ns_pool); 989 /* 990 * We use to shrink propotionally but with new shrinker API, 991 * we lost the total number of freeable locks. 992 */ 993 cancel = 1 + min_t(int, nr_locks, nr / nr_ns); 994 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask); 995 ldlm_namespace_put(ns); 996 } 997 /* 998 * we only decrease the SLV in server pools shrinker, return 999 * SHRINK_STOP to kernel to avoid needless loop. LU-1128 1000 */ 1001 return freed; 1002} 1003 1004static unsigned long ldlm_pools_cli_count(struct shrinker *s, 1005 struct shrink_control *sc) 1006{ 1007 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask); 1008} 1009 1010static unsigned long ldlm_pools_cli_scan(struct shrinker *s, 1011 struct shrink_control *sc) 1012{ 1013 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan, 1014 sc->gfp_mask); 1015} 1016 1017static void ldlm_pools_recalc(struct work_struct *ws); 1018static DECLARE_DELAYED_WORK(ldlm_recalc_pools, ldlm_pools_recalc); 1019 1020static void ldlm_pools_recalc(struct work_struct *ws) 1021{ 1022 enum ldlm_side client = LDLM_NAMESPACE_CLIENT; 1023 struct ldlm_namespace *ns; 1024 struct ldlm_namespace *ns_old = NULL; 1025 /* seconds of sleep if no active namespaces */ 1026<<<<<<< found 1027 time64_t time = LDLM_POOL_CLI_DEF_RECALC_PERIOD; 1028||||||| expected 1029 time64_t delay = side == LDLM_NAMESPACE_SERVER ? 1030 LDLM_POOL_SRV_DEF_RECALC_PERIOD : 1031 LDLM_POOL_CLI_DEF_RECALC_PERIOD; 1032======= 1033 time64_t delay = ktime_get_seconds() + 1034 (side == LDLM_NAMESPACE_SERVER ? 1035 LDLM_POOL_SRV_DEF_RECALC_PERIOD : 1036 LDLM_POOL_CLI_DEF_RECALC_PERIOD); 1037>>>>>>> replacement 1038 int nr; 1039 1040 /* 1041 * Recalc at least ldlm_namespace_nr_read(client) namespaces. 1042 */ 1043 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) { 1044 int skip; 1045 /* 1046 * Lock the list, get first @ns in the list, getref, move it 1047 * to the tail, unlock and call pool recalc. This way we avoid 1048 * calling recalc under @ns lock what is really good as we get 1049 * rid of potential deadlock on client nodes when canceling 1050 * locks synchronously. 1051 */ 1052 mutex_lock(ldlm_namespace_lock(client)); 1053 if (list_empty(ldlm_namespace_list(client))) { 1054 mutex_unlock(ldlm_namespace_lock(client)); 1055 break; 1056 } 1057 ns = ldlm_namespace_first_locked(client); 1058 1059 if (ns_old == ns) { /* Full pass complete */ 1060 mutex_unlock(ldlm_namespace_lock(client)); 1061 break; 1062 } 1063 1064 /* We got an empty namespace, need to move it back to inactive 1065 * list. 1066 * The race with parallel resource creation is fine: 1067 * - If they do namespace_get before our check, we fail the 1068 * check and they move this item to the end of the list anyway 1069 * - If we do the check and then they do namespace_get, then 1070 * we move the namespace to inactive and they will move 1071 * it back to active (synchronised by the lock, so no clash 1072 * there). 1073 */ 1074 if (ldlm_ns_empty(ns)) { 1075 ldlm_namespace_move_to_inactive_locked(ns, client); 1076 mutex_unlock(ldlm_namespace_lock(client)); 1077 continue; 1078 } 1079 1080 if (!ns_old) 1081 ns_old = ns; 1082 1083 spin_lock(&ns->ns_lock); 1084 /* 1085 * skip ns which is being freed, and we don't want to increase 1086 * its refcount again, not even temporarily. bz21519 & LU-499. 1087 */ 1088 if (ns->ns_stopping) { 1089 skip = 1; 1090 } else { 1091 skip = 0; 1092 ldlm_namespace_get(ns); 1093 } 1094 spin_unlock(&ns->ns_lock); 1095 1096 ldlm_namespace_move_to_active_locked(ns, client); 1097 mutex_unlock(ldlm_namespace_lock(client)); 1098 1099 /* 1100 * After setup is done - recalc the pool. 1101 */ 1102 if (!skip) { 1103 time64_t ttime = ldlm_pool_recalc(&ns->ns_pool); 1104 1105 if (ttime < time) 1106 time = ttime; 1107 1108 ldlm_namespace_put(ns); 1109 } 1110 } 1111 1112 /* Wake up the blocking threads from time to time. */ 1113 ldlm_bl_thread_wakeup(); 1114 1115 delay -= ktime_get_seconds(); 1116 if (delay <= 0) { 1117 /* Prevent too frequent recalculation. */ 1118 CDEBUG(D_DLMTRACE, "Negative interval(%lld)\n", delay); 1119 delay = 1; 1120 } 1121 1122 schedule_delayed_work(&ldlm_recalc_pools, time * HZ); 1123} 1124 1125static int ldlm_pools_thread_start(void) 1126{ 1127 time64_t delay; 1128 1129 schedule_delayed_work(&ldlm_recalc_pools, 0); 1130 1131<<<<<<< found 1132 return 0; 1133} 1134 1135static void ldlm_pools_thread_stop(void) 1136{ 1137||||||| expected 1138 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink, 1139 ldlm_pools_cli_count, ldlm_pools_cli_scan); 1140 1141 schedule_delayed_work(&ldlm_pools_recalc_work, 1142 LDLM_POOL_CLI_DEF_RECALC_PERIOD); 1143 ldlm_pools_srv_shrinker = set_shrinker(DEFAULT_SEEKS, &shsvar); 1144======= 1145 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink, 1146 ldlm_pools_cli_count, ldlm_pools_cli_scan); 1147 1148#ifdef HAVE_SERVER_SUPPORT 1149 delay = min(LDLM_POOL_SRV_DEF_RECALC_PERIOD, 1150 LDLM_POOL_CLI_DEF_RECALC_PERIOD); 1151#else 1152 delay = LDLM_POOL_CLI_DEF_RECALC_PERIOD; 1153#endif 1154 1155 schedule_delayed_work(&ldlm_pools_recalc_work, delay); 1156 ldlm_pools_srv_shrinker = set_shrinker(DEFAULT_SEEKS, &shsvar); 1157>>>>>>> replacement 1158 cancel_delayed_work_sync(&ldlm_recalc_pools); 1159} 1160 1161static struct shrinker ldlm_pools_cli_shrinker = { 1162 .count_objects = ldlm_pools_cli_count, 1163 .scan_objects = ldlm_pools_cli_scan, 1164 .seeks = DEFAULT_SEEKS, 1165}; 1166 1167int ldlm_pools_init(void) 1168{ 1169 int rc; 1170 1171 rc = ldlm_pools_thread_start(); 1172 if (!rc) 1173 rc = register_shrinker(&ldlm_pools_cli_shrinker); 1174 1175 return rc; 1176} 1177 1178void ldlm_pools_fini(void) 1179{ 1180 unregister_shrinker(&ldlm_pools_cli_shrinker); 1181 1182 ldlm_pools_thread_stop(); 1183} 1184