drivers/staging/lustre/lustre/ptlrpc/service.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_RPC
  34 #include "../include/obd_support.h"
  35 #include "../include/obd_class.h"
  36 #include "../include/lustre_net.h"
  37 #include "../include/lu_object.h"
  38 #include "../../include/linux/lnet/types.h"
  39 #include "ptlrpc_internal.h"
  40
  41 /* The following are visible and mutable through /sys/module/ptlrpc */
  42 int test_req_buffer_pressure;
  43 module_param(test_req_buffer_pressure, int, 0444);
  44 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
  45 module_param(at_min, int, 0644);
  46 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
  47 module_param(at_max, int, 0644);
  48 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
  49 module_param(at_history, int, 0644);
  50 MODULE_PARM_DESC(at_history,
  51                  "Adaptive timeouts remember the slowest event that took place within this period (sec)");
  52 module_param(at_early_margin, int, 0644);
  53 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
  54 module_param(at_extra, int, 0644);
  55 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
  56
  57 /* forward ref */
  58 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
  59 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
  60 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
  61
  62 /** Holds a list of all PTLRPC services */
  63 LIST_HEAD(ptlrpc_all_services);
  64 /** Used to protect the \e ptlrpc_all_services list */
  65 struct mutex ptlrpc_all_services_mutex;
  66
  67 static struct ptlrpc_request_buffer_desc *
  68 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
  69 {
  70         struct ptlrpc_service *svc = svcpt->scp_service;
  71         struct ptlrpc_request_buffer_desc *rqbd;
  72
  73         rqbd = kzalloc_node(sizeof(*rqbd), GFP_NOFS,
  74                             cfs_cpt_spread_node(svc->srv_cptable,
  75                                                 svcpt->scp_cpt));
  76         if (!rqbd)
  77                 return NULL;
  78
  79         rqbd->rqbd_svcpt = svcpt;
  80         rqbd->rqbd_refcount = 0;
  81         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
  82         rqbd->rqbd_cbid.cbid_arg = rqbd;
  83         INIT_LIST_HEAD(&rqbd->rqbd_reqs);
  84         rqbd->rqbd_buffer = libcfs_kvzalloc_cpt(svc->srv_cptable,
  85                                                 svcpt->scp_cpt,
  86                                                 svc->srv_buf_size,
  87                                                 GFP_KERNEL);
  88         if (!rqbd->rqbd_buffer) {
  89                 kfree(rqbd);
  90                 return NULL;
  91         }
  92
  93         spin_lock(&svcpt->scp_lock);
  94         list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
  95         svcpt->scp_nrqbds_total++;
  96         spin_unlock(&svcpt->scp_lock);
  97
  98         return rqbd;
  99 }
 100
 101 static void
 102 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 103 {
 104         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
 105
 106         LASSERT(rqbd->rqbd_refcount == 0);
 107         LASSERT(list_empty(&rqbd->rqbd_reqs));
 108
 109         spin_lock(&svcpt->scp_lock);
 110         list_del(&rqbd->rqbd_list);
 111         svcpt->scp_nrqbds_total--;
 112         spin_unlock(&svcpt->scp_lock);
 113
 114         kvfree(rqbd->rqbd_buffer);
 115         kfree(rqbd);
 116 }
 117
 118 static int
 119 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
 120 {
 121         struct ptlrpc_service *svc = svcpt->scp_service;
 122         struct ptlrpc_request_buffer_desc *rqbd;
 123         int rc = 0;
 124         int i;
 125
 126         if (svcpt->scp_rqbd_allocating)
 127                 goto try_post;
 128
 129         spin_lock(&svcpt->scp_lock);
 130         /* check again with lock */
 131         if (svcpt->scp_rqbd_allocating) {
 132                 /* NB: we might allow more than one thread in the future */
 133                 LASSERT(svcpt->scp_rqbd_allocating == 1);
 134                 spin_unlock(&svcpt->scp_lock);
 135                 goto try_post;
 136         }
 137
 138         svcpt->scp_rqbd_allocating++;
 139         spin_unlock(&svcpt->scp_lock);
 140
 141         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 142                 /* NB: another thread might have recycled enough rqbds, we
 143                  * need to make sure it wouldn't over-allocate, see LU-1212.
 144                  */
 145                 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
 146                         break;
 147
 148                 rqbd = ptlrpc_alloc_rqbd(svcpt);
 149
 150                 if (!rqbd) {
 151                         CERROR("%s: Can't allocate request buffer\n",
 152                                svc->srv_name);
 153                         rc = -ENOMEM;
 154                         break;
 155                 }
 156         }
 157
 158         spin_lock(&svcpt->scp_lock);
 159
 160         LASSERT(svcpt->scp_rqbd_allocating == 1);
 161         svcpt->scp_rqbd_allocating--;
 162
 163         spin_unlock(&svcpt->scp_lock);
 164
 165         CDEBUG(D_RPCTRACE,
 166                "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
 167                svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
 168                svcpt->scp_nrqbds_total, rc);
 169
 170  try_post:
 171         if (post && rc == 0)
 172                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
 173
 174         return rc;
 175 }
 176
 177 struct ptlrpc_hr_partition;
 178
 179 struct ptlrpc_hr_thread {
 180         int                             hrt_id;         /* thread ID */
 181         spinlock_t                      hrt_lock;
 182         wait_queue_head_t                       hrt_waitq;
 183         struct list_head                        hrt_queue;      /* RS queue */
 184         struct ptlrpc_hr_partition      *hrt_partition;
 185 };
 186
 187 struct ptlrpc_hr_partition {
 188         /* # of started threads */
 189         atomic_t                        hrp_nstarted;
 190         /* # of stopped threads */
 191         atomic_t                        hrp_nstopped;
 192         /* cpu partition id */
 193         int                             hrp_cpt;
 194         /* round-robin rotor for choosing thread */
 195         int                             hrp_rotor;
 196         /* total number of threads on this partition */
 197         int                             hrp_nthrs;
 198         /* threads table */
 199         struct ptlrpc_hr_thread         *hrp_thrs;
 200 };
 201
 202 #define HRT_RUNNING 0
 203 #define HRT_STOPPING 1
 204
 205 struct ptlrpc_hr_service {
 206         /* CPU partition table, it's just cfs_cpt_table for now */
 207         struct cfs_cpt_table            *hr_cpt_table;
 208         /** controller sleep waitq */
 209         wait_queue_head_t                       hr_waitq;
 210         unsigned int                    hr_stopping;
 211         /** roundrobin rotor for non-affinity service */
 212         unsigned int                    hr_rotor;
 213         /* partition data */
 214         struct ptlrpc_hr_partition      **hr_partitions;
 215 };
 216
 217 /** reply handling service. */
 218 static struct ptlrpc_hr_service         ptlrpc_hr;
 219
 220 /**
 221  * Choose an hr thread to dispatch requests to.
 222  */
 223 static struct ptlrpc_hr_thread *
 224 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
 225 {
 226         struct ptlrpc_hr_partition *hrp;
 227         unsigned int rotor;
 228
 229         if (svcpt->scp_cpt >= 0 &&
 230             svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
 231                 /* directly match partition */
 232                 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
 233
 234         } else {
 235                 rotor = ptlrpc_hr.hr_rotor++;
 236                 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
 237
 238                 hrp = ptlrpc_hr.hr_partitions[rotor];
 239         }
 240
 241         rotor = hrp->hrp_rotor++;
 242         return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
 243 }
 244
 245 /**
 246  * Put reply state into a queue for processing because we received
 247  * ACK from the client
 248  */
 249 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
 250 {
 251         struct ptlrpc_hr_thread *hrt;
 252
 253         LASSERT(list_empty(&rs->rs_list));
 254
 255         hrt = ptlrpc_hr_select(rs->rs_svcpt);
 256
 257         spin_lock(&hrt->hrt_lock);
 258         list_add_tail(&rs->rs_list, &hrt->hrt_queue);
 259         spin_unlock(&hrt->hrt_lock);
 260
 261         wake_up(&hrt->hrt_waitq);
 262 }
 263
 264 void
 265 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
 266 {
 267         assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
 268         assert_spin_locked(&rs->rs_lock);
 269         LASSERT(rs->rs_difficult);
 270         rs->rs_scheduled_ever = 1;  /* flag any notification attempt */
 271
 272         if (rs->rs_scheduled) {     /* being set up or already notified */
 273                 return;
 274         }
 275
 276         rs->rs_scheduled = 1;
 277         list_del_init(&rs->rs_list);
 278         ptlrpc_dispatch_difficult_reply(rs);
 279 }
 280 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
 281
 282 static int
 283 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
 284 {
 285         struct ptlrpc_request_buffer_desc *rqbd;
 286         int rc;
 287         int posted = 0;
 288
 289         for (;;) {
 290                 spin_lock(&svcpt->scp_lock);
 291
 292                 if (list_empty(&svcpt->scp_rqbd_idle)) {
 293                         spin_unlock(&svcpt->scp_lock);
 294                         return posted;
 295                 }
 296
 297                 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
 298                                   struct ptlrpc_request_buffer_desc,
 299                                   rqbd_list);
 300                 list_del(&rqbd->rqbd_list);
 301
 302                 /* assume we will post successfully */
 303                 svcpt->scp_nrqbds_posted++;
 304                 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
 305
 306                 spin_unlock(&svcpt->scp_lock);
 307
 308                 rc = ptlrpc_register_rqbd(rqbd);
 309                 if (rc != 0)
 310                         break;
 311
 312                 posted = 1;
 313         }
 314
 315         spin_lock(&svcpt->scp_lock);
 316
 317         svcpt->scp_nrqbds_posted--;
 318         list_del(&rqbd->rqbd_list);
 319         list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
 320
 321         /* Don't complain if no request buffers are posted right now; LNET
 322          * won't drop requests because we set the portal lazy!
 323          */
 324
 325         spin_unlock(&svcpt->scp_lock);
 326
 327         return -1;
 328 }
 329
 330 static void ptlrpc_at_timer(unsigned long castmeharder)
 331 {
 332         struct ptlrpc_service_part *svcpt;
 333
 334         svcpt = (struct ptlrpc_service_part *)castmeharder;
 335
 336         svcpt->scp_at_check = 1;
 337         svcpt->scp_at_checktime = cfs_time_current();
 338         wake_up(&svcpt->scp_waitq);
 339 }
 340
 341 static void
 342 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
 343                              struct ptlrpc_service_conf *conf)
 344 {
 345         struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
 346         unsigned init;
 347         unsigned total;
 348         unsigned nthrs;
 349         int weight;
 350
 351         /*
 352          * Common code for estimating & validating threads number.
 353          * CPT affinity service could have percpt thread-pool instead
 354          * of a global thread-pool, which means user might not always
 355          * get the threads number they give it in conf::tc_nthrs_user
 356          * even they did set. It's because we need to validate threads
 357          * number for each CPT to guarantee each pool will have enough
 358          * threads to keep the service healthy.
 359          */
 360         init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
 361         init = max_t(int, init, tc->tc_nthrs_init);
 362
 363         /* NB: please see comments in lustre_lnet.h for definition
 364          * details of these members
 365          */
 366         LASSERT(tc->tc_nthrs_max != 0);
 367
 368         if (tc->tc_nthrs_user != 0) {
 369                 /* In case there is a reason to test a service with many
 370                  * threads, we give a less strict check here, it can
 371                  * be up to 8 * nthrs_max
 372                  */
 373                 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
 374                 nthrs = total / svc->srv_ncpts;
 375                 init = max(init, nthrs);
 376                 goto out;
 377         }
 378
 379         total = tc->tc_nthrs_max;
 380         if (tc->tc_nthrs_base == 0) {
 381                 /* don't care about base threads number per partition,
 382                  * this is most for non-affinity service
 383                  */
 384                 nthrs = total / svc->srv_ncpts;
 385                 goto out;
 386         }
 387
 388         nthrs = tc->tc_nthrs_base;
 389         if (svc->srv_ncpts == 1) {
 390                 int i;
 391
 392                 /* NB: Increase the base number if it's single partition
 393                  * and total number of cores/HTs is larger or equal to 4.
 394                  * result will always < 2 * nthrs_base
 395                  */
 396                 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
 397                 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
 398                             (tc->tc_nthrs_base >> i) != 0; i++)
 399                         nthrs += tc->tc_nthrs_base >> i;
 400         }
 401
 402         if (tc->tc_thr_factor != 0) {
 403                 int factor = tc->tc_thr_factor;
 404                 const int fade = 4;
 405
 406                 /*
 407                  * User wants to increase number of threads with for
 408                  * each CPU core/HT, most likely the factor is larger then
 409                  * one thread/core because service threads are supposed to
 410                  * be blocked by lock or wait for IO.
 411                  */
 412                 /*
 413                  * Amdahl's law says that adding processors wouldn't give
 414                  * a linear increasing of parallelism, so it's nonsense to
 415                  * have too many threads no matter how many cores/HTs
 416                  * there are.
 417                  */
 418                 /* weight is # of HTs */
 419                 if (cpumask_weight(topology_sibling_cpumask(0)) > 1) {
 420                         /* depress thread factor for hyper-thread */
 421                         factor = factor - (factor >> 1) + (factor >> 3);
 422                 }
 423
 424                 weight = cfs_cpt_weight(svc->srv_cptable, 0);
 425                 LASSERT(weight > 0);
 426
 427                 for (; factor > 0 && weight > 0; factor--, weight -= fade)
 428                         nthrs += min(weight, fade) * factor;
 429         }
 430
 431         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
 432                 nthrs = max(tc->tc_nthrs_base,
 433                             tc->tc_nthrs_max / svc->srv_ncpts);
 434         }
 435  out:
 436         nthrs = max(nthrs, tc->tc_nthrs_init);
 437         svc->srv_nthrs_cpt_limit = nthrs;
 438         svc->srv_nthrs_cpt_init = init;
 439
 440         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
 441                 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
 442                        svc->srv_name, nthrs * svc->srv_ncpts,
 443                        tc->tc_nthrs_max);
 444         }
 445 }
 446
 447 /**
 448  * Initialize percpt data for a service
 449  */
 450 static int
 451 ptlrpc_service_part_init(struct ptlrpc_service *svc,
 452                          struct ptlrpc_service_part *svcpt, int cpt)
 453 {
 454         struct ptlrpc_at_array  *array;
 455         int size;
 456         int index;
 457         int rc;
 458
 459         svcpt->scp_cpt = cpt;
 460         INIT_LIST_HEAD(&svcpt->scp_threads);
 461
 462         /* rqbd and incoming request queue */
 463         spin_lock_init(&svcpt->scp_lock);
 464         INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
 465         INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
 466         INIT_LIST_HEAD(&svcpt->scp_req_incoming);
 467         init_waitqueue_head(&svcpt->scp_waitq);
 468         /* history request & rqbd list */
 469         INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
 470         INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
 471
 472         /* active requests and hp requests */
 473         spin_lock_init(&svcpt->scp_req_lock);
 474
 475         /* reply states */
 476         spin_lock_init(&svcpt->scp_rep_lock);
 477         INIT_LIST_HEAD(&svcpt->scp_rep_active);
 478         INIT_LIST_HEAD(&svcpt->scp_rep_idle);
 479         init_waitqueue_head(&svcpt->scp_rep_waitq);
 480         atomic_set(&svcpt->scp_nreps_difficult, 0);
 481
 482         /* adaptive timeout */
 483         spin_lock_init(&svcpt->scp_at_lock);
 484         array = &svcpt->scp_at_array;
 485
 486         size = at_est2timeout(at_max);
 487         array->paa_size = size;
 488         array->paa_count = 0;
 489         array->paa_deadline = -1;
 490
 491         /* allocate memory for scp_at_array (ptlrpc_at_array) */
 492         array->paa_reqs_array =
 493                 kzalloc_node(sizeof(struct list_head) * size, GFP_NOFS,
 494                              cfs_cpt_spread_node(svc->srv_cptable, cpt));
 495         if (!array->paa_reqs_array)
 496                 return -ENOMEM;
 497
 498         for (index = 0; index < size; index++)
 499                 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
 500
 501         array->paa_reqs_count =
 502                 kzalloc_node(sizeof(__u32) * size, GFP_NOFS,
 503                              cfs_cpt_spread_node(svc->srv_cptable, cpt));
 504         if (!array->paa_reqs_count)
 505                 goto free_reqs_array;
 506
 507         setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
 508                     (unsigned long)svcpt);
 509
 510         /* At SOW, service time should be quick; 10s seems generous. If client
 511          * timeout is less than this, we'll be sending an early reply.
 512          */
 513         at_init(&svcpt->scp_at_estimate, 10, 0);
 514
 515         /* assign this before call ptlrpc_grow_req_bufs */
 516         svcpt->scp_service = svc;
 517         /* Now allocate the request buffers, but don't post them now */
 518         rc = ptlrpc_grow_req_bufs(svcpt, 0);
 519         /* We shouldn't be under memory pressure at startup, so
 520          * fail if we can't allocate all our buffers at this time.
 521          */
 522         if (rc != 0)
 523                 goto free_reqs_count;
 524
 525         return 0;
 526
 527 free_reqs_count:
 528         kfree(array->paa_reqs_count);
 529         array->paa_reqs_count = NULL;
 530 free_reqs_array:
 531         kfree(array->paa_reqs_array);
 532         array->paa_reqs_array = NULL;
 533
 534         return -ENOMEM;
 535 }
 536
 537 /**
 538  * Initialize service on a given portal.
 539  * This includes starting serving threads , allocating and posting rqbds and
 540  * so on.
 541  */
 542 struct ptlrpc_service *
 543 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
 544                         struct kset *parent,
 545                         struct dentry *debugfs_entry)
 546 {
 547         struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
 548         struct ptlrpc_service *service;
 549         struct ptlrpc_service_part *svcpt;
 550         struct cfs_cpt_table *cptable;
 551         __u32 *cpts = NULL;
 552         int ncpts;
 553         int cpt;
 554         int rc;
 555         int i;
 556
 557         LASSERT(conf->psc_buf.bc_nbufs > 0);
 558         LASSERT(conf->psc_buf.bc_buf_size >=
 559                 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
 560         LASSERT(conf->psc_thr.tc_ctx_tags != 0);
 561
 562         cptable = cconf->cc_cptable;
 563         if (!cptable)
 564                 cptable = cfs_cpt_table;
 565
 566         if (!conf->psc_thr.tc_cpu_affinity) {
 567                 ncpts = 1;
 568         } else {
 569                 ncpts = cfs_cpt_number(cptable);
 570                 if (cconf->cc_pattern) {
 571                         struct cfs_expr_list *el;
 572
 573                         rc = cfs_expr_list_parse(cconf->cc_pattern,
 574                                                  strlen(cconf->cc_pattern),
 575                                                  0, ncpts - 1, &el);
 576                         if (rc != 0) {
 577                                 CERROR("%s: invalid CPT pattern string: %s",
 578                                        conf->psc_name, cconf->cc_pattern);
 579                                 return ERR_PTR(-EINVAL);
 580                         }
 581
 582                         rc = cfs_expr_list_values(el, ncpts, &cpts);
 583                         cfs_expr_list_free(el);
 584                         if (rc <= 0) {
 585                                 CERROR("%s: failed to parse CPT array %s: %d\n",
 586                                        conf->psc_name, cconf->cc_pattern, rc);
 587                                 kfree(cpts);
 588                                 return ERR_PTR(rc < 0 ? rc : -EINVAL);
 589                         }
 590                         ncpts = rc;
 591                 }
 592         }
 593
 594         service = kzalloc(offsetof(struct ptlrpc_service, srv_parts[ncpts]),
 595                           GFP_NOFS);
 596         if (!service) {
 597                 kfree(cpts);
 598                 return ERR_PTR(-ENOMEM);
 599         }
 600
 601         service->srv_cptable = cptable;
 602         service->srv_cpts = cpts;
 603         service->srv_ncpts = ncpts;
 604
 605         service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
 606         while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
 607                 service->srv_cpt_bits++;
 608
 609         /* public members */
 610         spin_lock_init(&service->srv_lock);
 611         service->srv_name = conf->psc_name;
 612         service->srv_watchdog_factor = conf->psc_watchdog_factor;
 613         INIT_LIST_HEAD(&service->srv_list); /* for safety of cleanup */
 614
 615         /* buffer configuration */
 616         service->srv_nbuf_per_group = test_req_buffer_pressure ?
 617                                           1 : conf->psc_buf.bc_nbufs;
 618         service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
 619                                           SPTLRPC_MAX_PAYLOAD;
 620         service->srv_buf_size = conf->psc_buf.bc_buf_size;
 621         service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
 622         service->srv_req_portal = conf->psc_buf.bc_req_portal;
 623
 624         /* Increase max reply size to next power of two */
 625         service->srv_max_reply_size = 1;
 626         while (service->srv_max_reply_size <
 627                conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
 628                 service->srv_max_reply_size <<= 1;
 629
 630         service->srv_thread_name = conf->psc_thr.tc_thr_name;
 631         service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
 632         service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
 633         service->srv_ops = conf->psc_ops;
 634
 635         for (i = 0; i < ncpts; i++) {
 636                 if (!conf->psc_thr.tc_cpu_affinity)
 637                         cpt = CFS_CPT_ANY;
 638                 else
 639                         cpt = cpts ? cpts[i] : i;
 640
 641                 svcpt = kzalloc_node(sizeof(*svcpt), GFP_NOFS,
 642                                      cfs_cpt_spread_node(cptable, cpt));
 643                 if (!svcpt) {
 644                         rc = -ENOMEM;
 645                         goto failed;
 646                 }
 647
 648                 service->srv_parts[i] = svcpt;
 649                 rc = ptlrpc_service_part_init(service, svcpt, cpt);
 650                 if (rc != 0)
 651                         goto failed;
 652         }
 653
 654         ptlrpc_server_nthreads_check(service, conf);
 655
 656         rc = LNetSetLazyPortal(service->srv_req_portal);
 657         LASSERT(rc == 0);
 658
 659         mutex_lock(&ptlrpc_all_services_mutex);
 660         list_add(&service->srv_list, &ptlrpc_all_services);
 661         mutex_unlock(&ptlrpc_all_services_mutex);
 662
 663         if (parent) {
 664                 rc = ptlrpc_sysfs_register_service(parent, service);
 665                 if (rc)
 666                         goto failed;
 667         }
 668
 669         if (!IS_ERR_OR_NULL(debugfs_entry))
 670                 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
 671
 672         rc = ptlrpc_service_nrs_setup(service);
 673         if (rc != 0)
 674                 goto failed;
 675
 676         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 677                service->srv_name, service->srv_req_portal);
 678
 679         rc = ptlrpc_start_threads(service);
 680         if (rc != 0) {
 681                 CERROR("Failed to start threads for service %s: %d\n",
 682                        service->srv_name, rc);
 683                 goto failed;
 684         }
 685
 686         return service;
 687 failed:
 688         ptlrpc_unregister_service(service);
 689         return ERR_PTR(rc);
 690 }
 691 EXPORT_SYMBOL(ptlrpc_register_service);
 692
 693 /**
 694  * to actually free the request, must be called without holding svc_lock.
 695  * note it's caller's responsibility to unlink req->rq_list.
 696  */
 697 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
 698 {
 699         LASSERT(atomic_read(&req->rq_refcount) == 0);
 700         LASSERT(list_empty(&req->rq_timed_list));
 701
 702          /* DEBUG_REQ() assumes the reply state of a request with a valid
 703           * ref will not be destroyed until that reference is dropped.
 704           */
 705         ptlrpc_req_drop_rs(req);
 706
 707         sptlrpc_svc_ctx_decref(req);
 708
 709         if (req != &req->rq_rqbd->rqbd_req) {
 710                 /* NB request buffers use an embedded
 711                  * req if the incoming req unlinked the
 712                  * MD; this isn't one of them!
 713                  */
 714                 ptlrpc_request_cache_free(req);
 715         }
 716 }
 717
 718 /**
 719  * drop a reference count of the request. if it reaches 0, we either
 720  * put it into history list, or free it immediately.
 721  */
 722 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
 723 {
 724         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 725         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
 726         struct ptlrpc_service *svc = svcpt->scp_service;
 727         int refcount;
 728         struct list_head *tmp;
 729         struct list_head *nxt;
 730
 731         if (!atomic_dec_and_test(&req->rq_refcount))
 732                 return;
 733
 734         if (req->rq_at_linked) {
 735                 spin_lock(&svcpt->scp_at_lock);
 736                 /* recheck with lock, in case it's unlinked by
 737                  * ptlrpc_at_check_timed()
 738                  */
 739                 if (likely(req->rq_at_linked))
 740                         ptlrpc_at_remove_timed(req);
 741                 spin_unlock(&svcpt->scp_at_lock);
 742         }
 743
 744         LASSERT(list_empty(&req->rq_timed_list));
 745
 746         /* finalize request */
 747         if (req->rq_export) {
 748                 class_export_put(req->rq_export);
 749                 req->rq_export = NULL;
 750         }
 751
 752         spin_lock(&svcpt->scp_lock);
 753
 754         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 755
 756         refcount = --(rqbd->rqbd_refcount);
 757         if (refcount == 0) {
 758                 /* request buffer is now idle: add to history */
 759                 list_del(&rqbd->rqbd_list);
 760
 761                 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
 762                 svcpt->scp_hist_nrqbds++;
 763
 764                 /* cull some history?
 765                  * I expect only about 1 or 2 rqbds need to be recycled here
 766                  */
 767                 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
 768                         rqbd = list_entry(svcpt->scp_hist_rqbds.next,
 769                                           struct ptlrpc_request_buffer_desc,
 770                                           rqbd_list);
 771
 772                         list_del(&rqbd->rqbd_list);
 773                         svcpt->scp_hist_nrqbds--;
 774
 775                         /* remove rqbd's reqs from svc's req history while
 776                          * I've got the service lock
 777                          */
 778                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 779                                 req = list_entry(tmp, struct ptlrpc_request,
 780                                                  rq_list);
 781                                 /* Track the highest culled req seq */
 782                                 if (req->rq_history_seq >
 783                                     svcpt->scp_hist_seq_culled) {
 784                                         svcpt->scp_hist_seq_culled =
 785                                                 req->rq_history_seq;
 786                                 }
 787                                 list_del(&req->rq_history_list);
 788                         }
 789
 790                         spin_unlock(&svcpt->scp_lock);
 791
 792                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 793                                 req = list_entry(rqbd->rqbd_reqs.next,
 794                                                  struct ptlrpc_request,
 795                                                  rq_list);
 796                                 list_del(&req->rq_list);
 797                                 ptlrpc_server_free_request(req);
 798                         }
 799
 800                         spin_lock(&svcpt->scp_lock);
 801                         /*
 802                          * now all reqs including the embedded req has been
 803                          * disposed, schedule request buffer for re-use.
 804                          */
 805                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) ==
 806                                 0);
 807                         list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
 808                 }
 809
 810                 spin_unlock(&svcpt->scp_lock);
 811         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 812                 /* If we are low on memory, we are not interested in history */
 813                 list_del(&req->rq_list);
 814                 list_del_init(&req->rq_history_list);
 815
 816                 /* Track the highest culled req seq */
 817                 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
 818                         svcpt->scp_hist_seq_culled = req->rq_history_seq;
 819
 820                 spin_unlock(&svcpt->scp_lock);
 821
 822                 ptlrpc_server_free_request(req);
 823         } else {
 824                 spin_unlock(&svcpt->scp_lock);
 825         }
 826 }
 827
 828 /**
 829  * to finish a request: stop sending more early replies, and release
 830  * the request.
 831  */
 832 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
 833                                          struct ptlrpc_request *req)
 834 {
 835         ptlrpc_server_hpreq_fini(req);
 836
 837         if (req->rq_session.lc_thread) {
 838                 lu_context_exit(&req->rq_session);
 839                 lu_context_fini(&req->rq_session);
 840         }
 841
 842         ptlrpc_server_drop_request(req);
 843 }
 844
 845 /**
 846  * to finish a active request: stop sending more early replies, and release
 847  * the request. should be called after we finished handling the request.
 848  */
 849 static void ptlrpc_server_finish_active_request(
 850                                         struct ptlrpc_service_part *svcpt,
 851                                         struct ptlrpc_request *req)
 852 {
 853         spin_lock(&svcpt->scp_req_lock);
 854         ptlrpc_nrs_req_stop_nolock(req);
 855         svcpt->scp_nreqs_active--;
 856         if (req->rq_hp)
 857                 svcpt->scp_nhreqs_active--;
 858         spin_unlock(&svcpt->scp_req_lock);
 859
 860         ptlrpc_nrs_req_finalize(req);
 861
 862         if (req->rq_export)
 863                 class_export_rpc_dec(req->rq_export);
 864
 865         ptlrpc_server_finish_request(svcpt, req);
 866 }
 867
 868 /**
 869  * Sanity check request \a req.
 870  * Return 0 if all is ok, error code otherwise.
 871  */
 872 static int ptlrpc_check_req(struct ptlrpc_request *req)
 873 {
 874         struct obd_device *obd = req->rq_export->exp_obd;
 875         int rc = 0;
 876
 877         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 878                      req->rq_export->exp_conn_cnt)) {
 879                 DEBUG_REQ(D_RPCTRACE, req,
 880                           "DROPPING req from old connection %d < %d",
 881                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 882                           req->rq_export->exp_conn_cnt);
 883                 return -EEXIST;
 884         }
 885         if (unlikely(!obd || obd->obd_fail)) {
 886                 /*
 887                  * Failing over, don't handle any more reqs, send
 888                  * error response instead.
 889                  */
 890                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 891                        req, obd ? obd->obd_name : "unknown");
 892                 rc = -ENODEV;
 893         } else if (lustre_msg_get_flags(req->rq_reqmsg) &
 894                    (MSG_REPLAY | MSG_REQ_REPLAY_DONE)) {
 895                 DEBUG_REQ(D_ERROR, req, "Invalid replay without recovery");
 896                 class_fail_export(req->rq_export);
 897                 rc = -ENODEV;
 898         } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0) {
 899                 DEBUG_REQ(D_ERROR, req,
 900                           "Invalid req with transno %llu without recovery",
 901                           lustre_msg_get_transno(req->rq_reqmsg));
 902                 class_fail_export(req->rq_export);
 903                 rc = -ENODEV;
 904         }
 905
 906         if (unlikely(rc < 0)) {
 907                 req->rq_status = rc;
 908                 ptlrpc_error(req);
 909         }
 910         return rc;
 911 }
 912
 913 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
 914 {
 915         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
 916         __s32 next;
 917
 918         if (array->paa_count == 0) {
 919                 del_timer(&svcpt->scp_at_timer);
 920                 return;
 921         }
 922
 923         /* Set timer for closest deadline */
 924         next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
 925                        at_early_margin);
 926         if (next <= 0) {
 927                 ptlrpc_at_timer((unsigned long)svcpt);
 928         } else {
 929                 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
 930                 CDEBUG(D_INFO, "armed %s at %+ds\n",
 931                        svcpt->scp_service->srv_name, next);
 932         }
 933 }
 934
 935 /* Add rpc to early reply check list */
 936 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 937 {
 938         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
 939         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
 940         struct ptlrpc_request *rq = NULL;
 941         __u32 index;
 942
 943         if (AT_OFF)
 944                 return 0;
 945
 946         if (req->rq_no_reply)
 947                 return 0;
 948
 949         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 950                 return -ENOSYS;
 951
 952         spin_lock(&svcpt->scp_at_lock);
 953         LASSERT(list_empty(&req->rq_timed_list));
 954
 955         div_u64_rem(req->rq_deadline, array->paa_size, &index);
 956         if (array->paa_reqs_count[index] > 0) {
 957                 /* latest rpcs will have the latest deadlines in the list,
 958                  * so search backward.
 959                  */
 960                 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
 961                                             rq_timed_list) {
 962                         if (req->rq_deadline >= rq->rq_deadline) {
 963                                 list_add(&req->rq_timed_list,
 964                                          &rq->rq_timed_list);
 965                                 break;
 966                         }
 967                 }
 968         }
 969
 970         /* Add the request at the head of the list */
 971         if (list_empty(&req->rq_timed_list))
 972                 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
 973
 974         spin_lock(&req->rq_lock);
 975         req->rq_at_linked = 1;
 976         spin_unlock(&req->rq_lock);
 977         req->rq_at_index = index;
 978         array->paa_reqs_count[index]++;
 979         array->paa_count++;
 980         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
 981                 array->paa_deadline = req->rq_deadline;
 982                 ptlrpc_at_set_timer(svcpt);
 983         }
 984         spin_unlock(&svcpt->scp_at_lock);
 985
 986         return 0;
 987 }
 988
 989 static void
 990 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
 991 {
 992         struct ptlrpc_at_array *array;
 993
 994         array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
 995
 996         /* NB: must call with hold svcpt::scp_at_lock */
 997         LASSERT(!list_empty(&req->rq_timed_list));
 998         list_del_init(&req->rq_timed_list);
 999
1000         spin_lock(&req->rq_lock);
1001         req->rq_at_linked = 0;
1002         spin_unlock(&req->rq_lock);
1003
1004         array->paa_reqs_count[req->rq_at_index]--;
1005         array->paa_count--;
1006 }
1007
1008 /*
1009  * Attempt to extend the request deadline by sending an early reply to the
1010  * client.
1011  */
1012 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1013 {
1014         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1015         struct ptlrpc_request *reqcopy;
1016         struct lustre_msg *reqmsg;
1017         long olddl = req->rq_deadline - ktime_get_real_seconds();
1018         time64_t newdl;
1019         int rc;
1020
1021         /* deadline is when the client expects us to reply, margin is the
1022          * difference between clients' and servers' expectations
1023          */
1024         DEBUG_REQ(D_ADAPTTO, req,
1025                   "%ssending early reply (deadline %+lds, margin %+lds) for %d+%d",
1026                   AT_OFF ? "AT off - not " : "",
1027                   olddl, olddl - at_get(&svcpt->scp_at_estimate),
1028                   at_get(&svcpt->scp_at_estimate), at_extra);
1029
1030         if (AT_OFF)
1031                 return 0;
1032
1033         if (olddl < 0) {
1034                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), not sending early reply. Consider increasing at_early_margin (%d)?",
1035                           olddl, at_early_margin);
1036
1037                 /* Return an error so we're not re-added to the timed list. */
1038                 return -ETIMEDOUT;
1039         }
1040
1041         if (!(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
1042                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, but no AT support");
1043                 return -ENOSYS;
1044         }
1045
1046         /*
1047          * We want to extend the request deadline by at_extra seconds,
1048          * so we set our service estimate to reflect how much time has
1049          * passed since this request arrived plus an additional
1050          * at_extra seconds. The client will calculate the new deadline
1051          * based on this service estimate (plus some additional time to
1052          * account for network latency). See ptlrpc_at_recv_early_reply
1053          */
1054         at_measured(&svcpt->scp_at_estimate, at_extra +
1055                     ktime_get_real_seconds() - req->rq_arrival_time.tv_sec);
1056         newdl = req->rq_arrival_time.tv_sec + at_get(&svcpt->scp_at_estimate);
1057
1058         /* Check to see if we've actually increased the deadline -
1059          * we may be past adaptive_max
1060          */
1061         if (req->rq_deadline >= newdl) {
1062                 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%ld/%lld), not sending early reply\n",
1063                           olddl, newdl - ktime_get_real_seconds());
1064                 return -ETIMEDOUT;
1065         }
1066
1067         reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1068         if (!reqcopy)
1069                 return -ENOMEM;
1070         reqmsg = libcfs_kvzalloc(req->rq_reqlen, GFP_NOFS);
1071         if (!reqmsg) {
1072                 rc = -ENOMEM;
1073                 goto out_free;
1074         }
1075
1076         *reqcopy = *req;
1077         reqcopy->rq_reply_state = NULL;
1078         reqcopy->rq_rep_swab_mask = 0;
1079         reqcopy->rq_pack_bulk = 0;
1080         reqcopy->rq_pack_udesc = 0;
1081         reqcopy->rq_packed_final = 0;
1082         sptlrpc_svc_ctx_addref(reqcopy);
1083         /* We only need the reqmsg for the magic */
1084         reqcopy->rq_reqmsg = reqmsg;
1085         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1086
1087         LASSERT(atomic_read(&req->rq_refcount));
1088         /** if it is last refcount then early reply isn't needed */
1089         if (atomic_read(&req->rq_refcount) == 1) {
1090                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, abort sending early reply\n");
1091                 rc = -EINVAL;
1092                 goto out;
1093         }
1094
1095         /* Connection ref */
1096         reqcopy->rq_export = class_conn2export(
1097                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
1098         if (!reqcopy->rq_export) {
1099                 rc = -ENODEV;
1100                 goto out;
1101         }
1102
1103         /* RPC ref */
1104         class_export_rpc_inc(reqcopy->rq_export);
1105         if (reqcopy->rq_export->exp_obd &&
1106             reqcopy->rq_export->exp_obd->obd_fail) {
1107                 rc = -ENODEV;
1108                 goto out_put;
1109         }
1110
1111         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1112         if (rc)
1113                 goto out_put;
1114
1115         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1116
1117         if (!rc) {
1118                 /* Adjust our own deadline to what we told the client */
1119                 req->rq_deadline = newdl;
1120                 req->rq_early_count++; /* number sent, server side */
1121         } else {
1122                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1123         }
1124
1125         /* Free the (early) reply state from lustre_pack_reply.
1126          * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1127          */
1128         ptlrpc_req_drop_rs(reqcopy);
1129
1130 out_put:
1131         class_export_rpc_dec(reqcopy->rq_export);
1132         class_export_put(reqcopy->rq_export);
1133 out:
1134         sptlrpc_svc_ctx_decref(reqcopy);
1135         kvfree(reqmsg);
1136 out_free:
1137         ptlrpc_request_cache_free(reqcopy);
1138         return rc;
1139 }
1140
1141 /* Send early replies to everybody expiring within at_early_margin
1142  * asking for at_extra time
1143  */
1144 static void ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1145 {
1146         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1147         struct ptlrpc_request *rq, *n;
1148         struct list_head work_list;
1149         __u32 index, count;
1150         time64_t deadline;
1151         time64_t now = ktime_get_real_seconds();
1152         long delay;
1153         int first, counter = 0;
1154
1155         spin_lock(&svcpt->scp_at_lock);
1156         if (svcpt->scp_at_check == 0) {
1157                 spin_unlock(&svcpt->scp_at_lock);
1158                 return;
1159         }
1160         delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1161         svcpt->scp_at_check = 0;
1162
1163         if (array->paa_count == 0) {
1164                 spin_unlock(&svcpt->scp_at_lock);
1165                 return;
1166         }
1167
1168         /* The timer went off, but maybe the nearest rpc already completed. */
1169         first = array->paa_deadline - now;
1170         if (first > at_early_margin) {
1171                 /* We've still got plenty of time.  Reset the timer. */
1172                 ptlrpc_at_set_timer(svcpt);
1173                 spin_unlock(&svcpt->scp_at_lock);
1174                 return;
1175         }
1176
1177         /* We're close to a timeout, and we don't know how much longer the
1178          * server will take. Send early replies to everyone expiring soon.
1179          */
1180         INIT_LIST_HEAD(&work_list);
1181         deadline = -1;
1182         div_u64_rem(array->paa_deadline, array->paa_size, &index);
1183         count = array->paa_count;
1184         while (count > 0) {
1185                 count -= array->paa_reqs_count[index];
1186                 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
1187                                          rq_timed_list) {
1188                         if (rq->rq_deadline > now + at_early_margin) {
1189                                 /* update the earliest deadline */
1190                                 if (deadline == -1 ||
1191                                     rq->rq_deadline < deadline)
1192                                         deadline = rq->rq_deadline;
1193                                 break;
1194                         }
1195
1196                         ptlrpc_at_remove_timed(rq);
1197                         /**
1198                          * ptlrpc_server_drop_request() may drop
1199                          * refcount to 0 already. Let's check this and
1200                          * don't add entry to work_list
1201                          */
1202                         if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1203                                 list_add(&rq->rq_timed_list, &work_list);
1204                         counter++;
1205                 }
1206
1207                 if (++index >= array->paa_size)
1208                         index = 0;
1209         }
1210         array->paa_deadline = deadline;
1211         /* we have a new earliest deadline, restart the timer */
1212         ptlrpc_at_set_timer(svcpt);
1213
1214         spin_unlock(&svcpt->scp_at_lock);
1215
1216         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early replies\n",
1217                first, at_extra, counter);
1218         if (first < 0) {
1219                 /* We're already past request deadlines before we even get a
1220                  * chance to send early replies
1221                  */
1222                 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1223                               svcpt->scp_service->srv_name);
1224                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1225                       counter, svcpt->scp_nreqs_incoming,
1226                       svcpt->scp_nreqs_active,
1227                       at_get(&svcpt->scp_at_estimate), delay);
1228         }
1229
1230         /* we took additional refcount so entries can't be deleted from list, no
1231          * locking is needed
1232          */
1233         while (!list_empty(&work_list)) {
1234                 rq = list_entry(work_list.next, struct ptlrpc_request,
1235                                 rq_timed_list);
1236                 list_del_init(&rq->rq_timed_list);
1237
1238                 if (ptlrpc_at_send_early_reply(rq) == 0)
1239                         ptlrpc_at_add_timed(rq);
1240
1241                 ptlrpc_server_drop_request(rq);
1242         }
1243 }
1244
1245 /**
1246  * Put the request to the export list if the request may become
1247  * a high priority one.
1248  */
1249 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1250                                     struct ptlrpc_request *req)
1251 {
1252         int rc = 0;
1253
1254         if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1255                 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1256                 if (rc < 0)
1257                         return rc;
1258                 LASSERT(rc == 0);
1259         }
1260         if (req->rq_export && req->rq_ops) {
1261                 /* Perform request specific check. We should do this check
1262                  * before the request is added into exp_hp_rpcs list otherwise
1263                  * it may hit swab race at LU-1044.
1264                  */
1265                 if (req->rq_ops->hpreq_check) {
1266                         rc = req->rq_ops->hpreq_check(req);
1267                         if (rc == -ESTALE) {
1268                                 req->rq_status = rc;
1269                                 ptlrpc_error(req);
1270                         }
1271                         /** can only return error,
1272                          * 0 for normal request,
1273                          *  or 1 for high priority request
1274                          */
1275                         LASSERT(rc <= 1);
1276                 }
1277
1278                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1279                 list_add(&req->rq_exp_list, &req->rq_export->exp_hp_rpcs);
1280                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1281         }
1282
1283         ptlrpc_nrs_req_initialize(svcpt, req, rc);
1284
1285         return rc;
1286 }
1287
1288 /** Remove the request from the export list. */
1289 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1290 {
1291         if (req->rq_export && req->rq_ops) {
1292                 /* refresh lock timeout again so that client has more
1293                  * room to send lock cancel RPC.
1294                  */
1295                 if (req->rq_ops->hpreq_fini)
1296                         req->rq_ops->hpreq_fini(req);
1297
1298                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1299                 list_del_init(&req->rq_exp_list);
1300                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1301         }
1302 }
1303
1304 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1305                                      struct ptlrpc_request *req)
1306 {
1307         int     rc;
1308
1309         rc = ptlrpc_server_hpreq_init(svcpt, req);
1310         if (rc < 0)
1311                 return rc;
1312
1313         ptlrpc_nrs_req_add(svcpt, req, !!rc);
1314
1315         return 0;
1316 }
1317
1318 /**
1319  * Allow to handle high priority request
1320  * User can call it w/o any lock but need to hold
1321  * ptlrpc_service_part::scp_req_lock to get reliable result
1322  */
1323 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1324                                      bool force)
1325 {
1326         int running = svcpt->scp_nthrs_running;
1327
1328         if (!nrs_svcpt_has_hp(svcpt))
1329                 return false;
1330
1331         if (force)
1332                 return true;
1333
1334         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1335                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1336                 /* leave just 1 thread for normal RPCs */
1337                 running = PTLRPC_NTHRS_INIT;
1338                 if (svcpt->scp_service->srv_ops.so_hpreq_handler)
1339                         running += 1;
1340         }
1341
1342         if (svcpt->scp_nreqs_active >= running - 1)
1343                 return false;
1344
1345         if (svcpt->scp_nhreqs_active == 0)
1346                 return true;
1347
1348         return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1349                svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1350 }
1351
1352 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1353                                        bool force)
1354 {
1355         return ptlrpc_server_allow_high(svcpt, force) &&
1356                ptlrpc_nrs_req_pending_nolock(svcpt, true);
1357 }
1358
1359 /**
1360  * Only allow normal priority requests on a service that has a high-priority
1361  * queue if forced (i.e. cleanup), if there are other high priority requests
1362  * already being processed (i.e. those threads can service more high-priority
1363  * requests), or if there are enough idle threads that a later thread can do
1364  * a high priority request.
1365  * User can call it w/o any lock but need to hold
1366  * ptlrpc_service_part::scp_req_lock to get reliable result
1367  */
1368 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1369                                        bool force)
1370 {
1371         int running = svcpt->scp_nthrs_running;
1372
1373         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1374                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1375                 /* leave just 1 thread for normal RPCs */
1376                 running = PTLRPC_NTHRS_INIT;
1377                 if (svcpt->scp_service->srv_ops.so_hpreq_handler)
1378                         running += 1;
1379         }
1380
1381         if (force ||
1382             svcpt->scp_nreqs_active < running - 2)
1383                 return true;
1384
1385         if (svcpt->scp_nreqs_active >= running - 1)
1386                 return false;
1387
1388         return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1389 }
1390
1391 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1392                                          bool force)
1393 {
1394         return ptlrpc_server_allow_normal(svcpt, force) &&
1395                ptlrpc_nrs_req_pending_nolock(svcpt, false);
1396 }
1397
1398 /**
1399  * Returns true if there are requests available in incoming
1400  * request queue for processing and it is allowed to fetch them.
1401  * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1402  * to get reliable result
1403  * \see ptlrpc_server_allow_normal
1404  * \see ptlrpc_server_allow high
1405  */
1406 static inline bool
1407 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1408 {
1409         return ptlrpc_server_high_pending(svcpt, force) ||
1410                ptlrpc_server_normal_pending(svcpt, force);
1411 }
1412
1413 /**
1414  * Fetch a request for processing from queue of unprocessed requests.
1415  * Favors high-priority requests.
1416  * Returns a pointer to fetched request.
1417  */
1418 static struct ptlrpc_request *
1419 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1420 {
1421         struct ptlrpc_request *req = NULL;
1422
1423         spin_lock(&svcpt->scp_req_lock);
1424
1425         if (ptlrpc_server_high_pending(svcpt, force)) {
1426                 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1427                 if (req) {
1428                         svcpt->scp_hreq_count++;
1429                         goto got_request;
1430                 }
1431         }
1432
1433         if (ptlrpc_server_normal_pending(svcpt, force)) {
1434                 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1435                 if (req) {
1436                         svcpt->scp_hreq_count = 0;
1437                         goto got_request;
1438                 }
1439         }
1440
1441         spin_unlock(&svcpt->scp_req_lock);
1442         return NULL;
1443
1444 got_request:
1445         svcpt->scp_nreqs_active++;
1446         if (req->rq_hp)
1447                 svcpt->scp_nhreqs_active++;
1448
1449         spin_unlock(&svcpt->scp_req_lock);
1450
1451         if (likely(req->rq_export))
1452                 class_export_rpc_inc(req->rq_export);
1453
1454         return req;
1455 }
1456
1457 /**
1458  * Handle freshly incoming reqs, add to timed early reply list,
1459  * pass on to regular request queue.
1460  * All incoming requests pass through here before getting into
1461  * ptlrpc_server_handle_req later on.
1462  */
1463 static int
1464 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1465                             struct ptlrpc_thread *thread)
1466 {
1467         struct ptlrpc_service *svc = svcpt->scp_service;
1468         struct ptlrpc_request *req;
1469         __u32 deadline;
1470         int rc;
1471
1472         spin_lock(&svcpt->scp_lock);
1473         if (list_empty(&svcpt->scp_req_incoming)) {
1474                 spin_unlock(&svcpt->scp_lock);
1475                 return 0;
1476         }
1477
1478         req = list_entry(svcpt->scp_req_incoming.next,
1479                          struct ptlrpc_request, rq_list);
1480         list_del_init(&req->rq_list);
1481         svcpt->scp_nreqs_incoming--;
1482         /* Consider this still a "queued" request as far as stats are
1483          * concerned
1484          */
1485         spin_unlock(&svcpt->scp_lock);
1486
1487         /* go through security check/transform */
1488         rc = sptlrpc_svc_unwrap_request(req);
1489         switch (rc) {
1490         case SECSVC_OK:
1491                 break;
1492         case SECSVC_COMPLETE:
1493                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1494                 goto err_req;
1495         case SECSVC_DROP:
1496                 goto err_req;
1497         default:
1498                 LBUG();
1499         }
1500
1501         /*
1502          * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1503          * redo it wouldn't be harmful.
1504          */
1505         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1506                 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1507                 if (rc != 0) {
1508                         CERROR("error unpacking request: ptl %d from %s x%llu\n",
1509                                svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1510                                req->rq_xid);
1511                         goto err_req;
1512                 }
1513         }
1514
1515         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1516         if (rc) {
1517                 CERROR("error unpacking ptlrpc body: ptl %d from %s x%llu\n",
1518                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1519                        req->rq_xid);
1520                 goto err_req;
1521         }
1522
1523         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1524             lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1525                 CERROR("drop incoming rpc opc %u, x%llu\n",
1526                        cfs_fail_val, req->rq_xid);
1527                 goto err_req;
1528         }
1529
1530         rc = -EINVAL;
1531         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1532                 CERROR("wrong packet type received (type=%u) from %s\n",
1533                        lustre_msg_get_type(req->rq_reqmsg),
1534                        libcfs_id2str(req->rq_peer));
1535                 goto err_req;
1536         }
1537
1538         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1539         case MDS_WRITEPAGE:
1540         case OST_WRITE:
1541                 req->rq_bulk_write = 1;
1542                 break;
1543         case MDS_READPAGE:
1544         case OST_READ:
1545         case MGS_CONFIG_READ:
1546                 req->rq_bulk_read = 1;
1547                 break;
1548         }
1549
1550         CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1551
1552         req->rq_export = class_conn2export(
1553                 lustre_msg_get_handle(req->rq_reqmsg));
1554         if (req->rq_export) {
1555                 rc = ptlrpc_check_req(req);
1556                 if (rc == 0) {
1557                         rc = sptlrpc_target_export_check(req->rq_export, req);
1558                         if (rc)
1559                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with illegal security flavor,");
1560                 }
1561
1562                 if (rc)
1563                         goto err_req;
1564         }
1565
1566         /* req_in handling should/must be fast */
1567         if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1568                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1569                           (long)(ktime_get_real_seconds() -
1570                                  req->rq_arrival_time.tv_sec));
1571
1572         /* Set rpc server deadline and add it to the timed list */
1573         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1574                     MSGHDR_AT_SUPPORT) ?
1575                    /* The max time the client expects us to take */
1576                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1577         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1578         if (unlikely(deadline == 0)) {
1579                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1580                 goto err_req;
1581         }
1582
1583         req->rq_svc_thread = thread;
1584         if (thread) {
1585                 /* initialize request session, it is needed for request
1586                  * processing by target
1587                  */
1588                 rc = lu_context_init(&req->rq_session,
1589                                      LCT_SERVER_SESSION | LCT_NOREF);
1590                 if (rc) {
1591                         CERROR("%s: failure to initialize session: rc = %d\n",
1592                                thread->t_name, rc);
1593                         goto err_req;
1594                 }
1595                 req->rq_session.lc_thread = thread;
1596                 lu_context_enter(&req->rq_session);
1597                 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1598         }
1599
1600         ptlrpc_at_add_timed(req);
1601
1602         /* Move it over to the request processing queue */
1603         rc = ptlrpc_server_request_add(svcpt, req);
1604         if (rc)
1605                 goto err_req;
1606
1607         wake_up(&svcpt->scp_waitq);
1608         return 1;
1609
1610 err_req:
1611         ptlrpc_server_finish_request(svcpt, req);
1612
1613         return 1;
1614 }
1615
1616 /**
1617  * Main incoming request handling logic.
1618  * Calls handler function from service to do actual processing.
1619  */
1620 static int
1621 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1622                              struct ptlrpc_thread *thread)
1623 {
1624         struct ptlrpc_service *svc = svcpt->scp_service;
1625         struct ptlrpc_request *request;
1626         struct timespec64 work_start;
1627         struct timespec64 work_end;
1628         struct timespec64 timediff;
1629         struct timespec64 arrived;
1630         unsigned long timediff_usecs;
1631         unsigned long arrived_usecs;
1632         int fail_opc = 0;
1633
1634         request = ptlrpc_server_request_get(svcpt, false);
1635         if (!request)
1636                 return 0;
1637
1638         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1639                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1640         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1641                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1642
1643         if (unlikely(fail_opc)) {
1644                 if (request->rq_export && request->rq_ops)
1645                         OBD_FAIL_TIMEOUT(fail_opc, 4);
1646         }
1647
1648         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1649
1650         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1651                 libcfs_debug_dumplog();
1652
1653         ktime_get_real_ts64(&work_start);
1654         timediff = timespec64_sub(work_start, request->rq_arrival_time);
1655         timediff_usecs = timediff.tv_sec * USEC_PER_SEC +
1656                          timediff.tv_nsec / NSEC_PER_USEC;
1657         if (likely(svc->srv_stats)) {
1658                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1659                                     timediff_usecs);
1660                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1661                                     svcpt->scp_nreqs_incoming);
1662                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1663                                     svcpt->scp_nreqs_active);
1664                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1665                                     at_get(&svcpt->scp_at_estimate));
1666         }
1667
1668         if (likely(request->rq_export)) {
1669                 if (unlikely(ptlrpc_check_req(request)))
1670                         goto put_conn;
1671         }
1672
1673         /* Discard requests queued for longer than the deadline.
1674          * The deadline is increased if we send an early reply.
1675          */
1676         if (ktime_get_real_seconds() > request->rq_deadline) {
1677                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline " CFS_DURATION_T ":" CFS_DURATION_T "s ago\n",
1678                           libcfs_id2str(request->rq_peer),
1679                           (long)(request->rq_deadline -
1680                                  request->rq_arrival_time.tv_sec),
1681                           (long)(ktime_get_real_seconds() -
1682                                  request->rq_deadline));
1683                 goto put_conn;
1684         }
1685
1686         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d\n",
1687                current_comm(),
1688                (request->rq_export ?
1689                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1690                (request->rq_export ?
1691                 atomic_read(&request->rq_export->exp_refcount) : -99),
1692                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1693                libcfs_id2str(request->rq_peer),
1694                lustre_msg_get_opc(request->rq_reqmsg));
1695
1696         if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1697                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1698
1699         CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
1700
1701         /* re-assign request and sesson thread to the current one */
1702         request->rq_svc_thread = thread;
1703         if (thread) {
1704                 LASSERT(request->rq_session.lc_thread);
1705                 request->rq_session.lc_thread = thread;
1706                 request->rq_session.lc_cookie = 0x55;
1707                 thread->t_env->le_ses = &request->rq_session;
1708         }
1709         svc->srv_ops.so_req_handler(request);
1710
1711         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1712
1713 put_conn:
1714         if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
1715                 DEBUG_REQ(D_WARNING, request,
1716                           "Request took longer than estimated (%lld:%llds); "
1717                           "client may timeout.",
1718                           (s64)request->rq_deadline -
1719                                request->rq_arrival_time.tv_sec,
1720                           (s64)ktime_get_real_seconds() - request->rq_deadline);
1721         }
1722
1723         ktime_get_real_ts64(&work_end);
1724         timediff = timespec64_sub(work_end, work_start);
1725         timediff_usecs = timediff.tv_sec * USEC_PER_SEC +
1726                          timediff.tv_nsec / NSEC_PER_USEC;
1727         arrived = timespec64_sub(work_end, request->rq_arrival_time);
1728         arrived_usecs = arrived.tv_sec * USEC_PER_SEC +
1729                          arrived.tv_nsec / NSEC_PER_USEC;
1730         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d Request processed in %ldus (%ldus total) trans %llu rc %d/%d\n",
1731                current_comm(),
1732                (request->rq_export ?
1733                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1734                (request->rq_export ?
1735                 atomic_read(&request->rq_export->exp_refcount) : -99),
1736                lustre_msg_get_status(request->rq_reqmsg),
1737                request->rq_xid,
1738                libcfs_id2str(request->rq_peer),
1739                lustre_msg_get_opc(request->rq_reqmsg),
1740                timediff_usecs,
1741                arrived_usecs,
1742                (request->rq_repmsg ?
1743                 lustre_msg_get_transno(request->rq_repmsg) :
1744                 request->rq_transno),
1745                request->rq_status,
1746                (request->rq_repmsg ?
1747                 lustre_msg_get_status(request->rq_repmsg) : -999));
1748         if (likely(svc->srv_stats && request->rq_reqmsg)) {
1749                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1750                 int opc = opcode_offset(op);
1751
1752                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1753                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1754                         lprocfs_counter_add(svc->srv_stats,
1755                                             opc + EXTRA_MAX_OPCODES,
1756                                             timediff_usecs);
1757                 }
1758         }
1759         if (unlikely(request->rq_early_count)) {
1760                 DEBUG_REQ(D_ADAPTTO, request,
1761                           "sent %d early replies before finishing in %llds",
1762                           request->rq_early_count,
1763                           (s64)work_end.tv_sec -
1764                           request->rq_arrival_time.tv_sec);
1765         }
1766
1767         ptlrpc_server_finish_active_request(svcpt, request);
1768
1769         return 1;
1770 }
1771
1772 /**
1773  * An internal function to process a single reply state object.
1774  */
1775 static int
1776 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
1777 {
1778         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
1779         struct ptlrpc_service *svc = svcpt->scp_service;
1780         struct obd_export *exp;
1781         int nlocks;
1782         int been_handled;
1783
1784         exp = rs->rs_export;
1785
1786         LASSERT(rs->rs_difficult);
1787         LASSERT(rs->rs_scheduled);
1788         LASSERT(list_empty(&rs->rs_list));
1789
1790         spin_lock(&exp->exp_lock);
1791         /* Noop if removed already */
1792         list_del_init(&rs->rs_exp_list);
1793         spin_unlock(&exp->exp_lock);
1794
1795         /* The disk commit callback holds exp_uncommitted_replies_lock while it
1796          * iterates over newly committed replies, removing them from
1797          * exp_uncommitted_replies.  It then drops this lock and schedules the
1798          * replies it found for handling here.
1799          *
1800          * We can avoid contention for exp_uncommitted_replies_lock between the
1801          * HRT threads and further commit callbacks by checking rs_committed
1802          * which is set in the commit callback while it holds both
1803          * rs_lock and exp_uncommitted_reples.
1804          *
1805          * If we see rs_committed clear, the commit callback _may_ not have
1806          * handled this reply yet and we race with it to grab
1807          * exp_uncommitted_replies_lock before removing the reply from
1808          * exp_uncommitted_replies.  Note that if we lose the race and the
1809          * reply has already been removed, list_del_init() is a noop.
1810          *
1811          * If we see rs_committed set, we know the commit callback is handling,
1812          * or has handled this reply since store reordering might allow us to
1813          * see rs_committed set out of sequence.  But since this is done
1814          * holding rs_lock, we can be sure it has all completed once we hold
1815          * rs_lock, which we do right next.
1816          */
1817         if (!rs->rs_committed) {
1818                 spin_lock(&exp->exp_uncommitted_replies_lock);
1819                 list_del_init(&rs->rs_obd_list);
1820                 spin_unlock(&exp->exp_uncommitted_replies_lock);
1821         }
1822
1823         spin_lock(&rs->rs_lock);
1824
1825         been_handled = rs->rs_handled;
1826         rs->rs_handled = 1;
1827
1828         nlocks = rs->rs_nlocks;          /* atomic "steal", but */
1829         rs->rs_nlocks = 0;                    /* locks still on rs_locks! */
1830
1831         if (nlocks == 0 && !been_handled) {
1832                 /* If we see this, we should already have seen the warning
1833                  * in mds_steal_ack_locks()
1834                  */
1835                 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
1836                        rs,
1837                        rs->rs_xid, rs->rs_transno, rs->rs_opc,
1838                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
1839         }
1840
1841         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1842                 spin_unlock(&rs->rs_lock);
1843
1844                 if (!been_handled && rs->rs_on_net) {
1845                         LNetMDUnlink(rs->rs_md_h);
1846                         /* Ignore return code; we're racing with completion */
1847                 }
1848
1849                 while (nlocks-- > 0)
1850                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1851                                          rs->rs_modes[nlocks]);
1852
1853                 spin_lock(&rs->rs_lock);
1854         }
1855
1856         rs->rs_scheduled = 0;
1857
1858         if (!rs->rs_on_net) {
1859                 /* Off the net */
1860                 spin_unlock(&rs->rs_lock);
1861
1862                 class_export_put(exp);
1863                 rs->rs_export = NULL;
1864                 ptlrpc_rs_decref(rs);
1865                 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
1866                     svc->srv_is_stopping)
1867                         wake_up_all(&svcpt->scp_waitq);
1868                 return 1;
1869         }
1870
1871         /* still on the net; callback will schedule */
1872         spin_unlock(&rs->rs_lock);
1873         return 1;
1874 }
1875
1876 static void
1877 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
1878 {
1879         int avail = svcpt->scp_nrqbds_posted;
1880         int low_water = test_req_buffer_pressure ? 0 :
1881                         svcpt->scp_service->srv_nbuf_per_group / 2;
1882
1883         /* NB I'm not locking; just looking. */
1884
1885         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1886          * allowed the request history to grow out of control.  We could put a
1887          * sanity check on that here and cull some history if we need the
1888          * space.
1889          */
1890
1891         if (avail <= low_water)
1892                 ptlrpc_grow_req_bufs(svcpt, 1);
1893
1894         if (svcpt->scp_service->srv_stats) {
1895                 lprocfs_counter_add(svcpt->scp_service->srv_stats,
1896                                     PTLRPC_REQBUF_AVAIL_CNTR, avail);
1897         }
1898 }
1899
1900 static int
1901 ptlrpc_retry_rqbds(void *arg)
1902 {
1903         struct ptlrpc_service_part *svcpt = arg;
1904
1905         svcpt->scp_rqbd_timeout = 0;
1906         return -ETIMEDOUT;
1907 }
1908
1909 static inline int
1910 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
1911 {
1912         return svcpt->scp_nreqs_active <
1913                svcpt->scp_nthrs_running - 1 -
1914                (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
1915 }
1916
1917 /**
1918  * allowed to create more threads
1919  * user can call it w/o any lock but need to hold
1920  * ptlrpc_service_part::scp_lock to get reliable result
1921  */
1922 static inline int
1923 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
1924 {
1925         return svcpt->scp_nthrs_running +
1926                svcpt->scp_nthrs_starting <
1927                svcpt->scp_service->srv_nthrs_cpt_limit;
1928 }
1929
1930 /**
1931  * too many requests and allowed to create more threads
1932  */
1933 static inline int
1934 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
1935 {
1936         return !ptlrpc_threads_enough(svcpt) &&
1937                 ptlrpc_threads_increasable(svcpt);
1938 }
1939
1940 static inline int
1941 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
1942 {
1943         return thread_is_stopping(thread) ||
1944                thread->t_svcpt->scp_service->srv_is_stopping;
1945 }
1946
1947 static inline int
1948 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
1949 {
1950         return !list_empty(&svcpt->scp_rqbd_idle) &&
1951                svcpt->scp_rqbd_timeout == 0;
1952 }
1953
1954 static inline int
1955 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
1956 {
1957         return svcpt->scp_at_check;
1958 }
1959
1960 /**
1961  * requests wait on preprocessing
1962  * user can call it w/o any lock but need to hold
1963  * ptlrpc_service_part::scp_lock to get reliable result
1964  */
1965 static inline int
1966 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
1967 {
1968         return !list_empty(&svcpt->scp_req_incoming);
1969 }
1970
1971 static __attribute__((__noinline__)) int
1972 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
1973                   struct ptlrpc_thread *thread)
1974 {
1975         /* Don't exit while there are replies to be handled */
1976         struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
1977                                              ptlrpc_retry_rqbds, svcpt);
1978
1979         /* XXX: Add this back when libcfs watchdog is merged upstream
1980         lc_watchdog_disable(thread->t_watchdog);
1981          */
1982
1983         cond_resched();
1984
1985         l_wait_event_exclusive_head(svcpt->scp_waitq,
1986                                     ptlrpc_thread_stopping(thread) ||
1987                                     ptlrpc_server_request_incoming(svcpt) ||
1988                                     ptlrpc_server_request_pending(svcpt,
1989                                                                   false) ||
1990                                     ptlrpc_rqbd_pending(svcpt) ||
1991                                     ptlrpc_at_check(svcpt), &lwi);
1992
1993         if (ptlrpc_thread_stopping(thread))
1994                 return -EINTR;
1995
1996         /*
1997         lc_watchdog_touch(thread->t_watchdog,
1998                           ptlrpc_server_get_timeout(svcpt));
1999          */
2000         return 0;
2001 }
2002
2003 /**
2004  * Main thread body for service threads.
2005  * Waits in a loop waiting for new requests to process to appear.
2006  * Every time an incoming requests is added to its queue, a waitq
2007  * is woken up and one of the threads will handle it.
2008  */
2009 static int ptlrpc_main(void *arg)
2010 {
2011         struct ptlrpc_thread *thread = arg;
2012         struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2013         struct ptlrpc_service *svc = svcpt->scp_service;
2014         struct ptlrpc_reply_state *rs;
2015         struct group_info *ginfo = NULL;
2016         struct lu_env *env;
2017         int counter = 0, rc = 0;
2018
2019         thread->t_pid = current_pid();
2020         unshare_fs_struct();
2021
2022         /* NB: we will call cfs_cpt_bind() for all threads, because we
2023          * might want to run lustre server only on a subset of system CPUs,
2024          * in that case ->scp_cpt is CFS_CPT_ANY
2025          */
2026         rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2027         if (rc != 0) {
2028                 CWARN("%s: failed to bind %s on CPT %d\n",
2029                       svc->srv_name, thread->t_name, svcpt->scp_cpt);
2030         }
2031
2032         ginfo = groups_alloc(0);
2033         if (!ginfo) {
2034                 rc = -ENOMEM;
2035                 goto out;
2036         }
2037
2038         set_current_groups(ginfo);
2039         put_group_info(ginfo);
2040
2041         if (svc->srv_ops.so_thr_init) {
2042                 rc = svc->srv_ops.so_thr_init(thread);
2043                 if (rc)
2044                         goto out;
2045         }
2046
2047         env = kzalloc(sizeof(*env), GFP_NOFS);
2048         if (!env) {
2049                 rc = -ENOMEM;
2050                 goto out_srv_fini;
2051         }
2052
2053         rc = lu_context_init(&env->le_ctx,
2054                              svc->srv_ctx_tags | LCT_REMEMBER | LCT_NOREF);
2055         if (rc)
2056                 goto out_srv_fini;
2057
2058         thread->t_env = env;
2059         env->le_ctx.lc_thread = thread;
2060         env->le_ctx.lc_cookie = 0x6;
2061
2062         while (!list_empty(&svcpt->scp_rqbd_idle)) {
2063                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2064                 if (rc >= 0)
2065                         continue;
2066
2067                 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2068                        svc->srv_name, svcpt->scp_cpt, rc);
2069                 goto out_srv_fini;
2070         }
2071
2072         /* Alloc reply state structure for this one */
2073         rs = libcfs_kvzalloc(svc->srv_max_reply_size, GFP_NOFS);
2074         if (!rs) {
2075                 rc = -ENOMEM;
2076                 goto out_srv_fini;
2077         }
2078
2079         spin_lock(&svcpt->scp_lock);
2080
2081         LASSERT(thread_is_starting(thread));
2082         thread_clear_flags(thread, SVC_STARTING);
2083
2084         LASSERT(svcpt->scp_nthrs_starting == 1);
2085         svcpt->scp_nthrs_starting--;
2086
2087         /* SVC_STOPPING may already be set here if someone else is trying
2088          * to stop the service while this new thread has been dynamically
2089          * forked. We still set SVC_RUNNING to let our creator know that
2090          * we are now running, however we will exit as soon as possible
2091          */
2092         thread_add_flags(thread, SVC_RUNNING);
2093         svcpt->scp_nthrs_running++;
2094         spin_unlock(&svcpt->scp_lock);
2095
2096         /* wake up our creator in case he's still waiting. */
2097         wake_up(&thread->t_ctl_waitq);
2098
2099         /*
2100         thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2101                                              NULL, NULL);
2102          */
2103
2104         spin_lock(&svcpt->scp_rep_lock);
2105         list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2106         wake_up(&svcpt->scp_rep_waitq);
2107         spin_unlock(&svcpt->scp_rep_lock);
2108
2109         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2110                svcpt->scp_nthrs_running);
2111
2112         /* XXX maintain a list of all managed devices: insert here */
2113         while (!ptlrpc_thread_stopping(thread)) {
2114                 if (ptlrpc_wait_event(svcpt, thread))
2115                         break;
2116
2117                 ptlrpc_check_rqbd_pool(svcpt);
2118
2119                 if (ptlrpc_threads_need_create(svcpt)) {
2120                         /* Ignore return code - we tried... */
2121                         ptlrpc_start_thread(svcpt, 0);
2122                 }
2123
2124                 /* Process all incoming reqs before handling any */
2125                 if (ptlrpc_server_request_incoming(svcpt)) {
2126                         lu_context_enter(&env->le_ctx);
2127                         env->le_ses = NULL;
2128                         ptlrpc_server_handle_req_in(svcpt, thread);
2129                         lu_context_exit(&env->le_ctx);
2130
2131                         /* but limit ourselves in case of flood */
2132                         if (counter++ < 100)
2133                                 continue;
2134                         counter = 0;
2135                 }
2136
2137                 if (ptlrpc_at_check(svcpt))
2138                         ptlrpc_at_check_timed(svcpt);
2139
2140                 if (ptlrpc_server_request_pending(svcpt, false)) {
2141                         lu_context_enter(&env->le_ctx);
2142                         ptlrpc_server_handle_request(svcpt, thread);
2143                         lu_context_exit(&env->le_ctx);
2144                 }
2145
2146                 if (ptlrpc_rqbd_pending(svcpt) &&
2147                     ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2148                         /* I just failed to repost request buffers.
2149                          * Wait for a timeout (unless something else
2150                          * happens) before I try again
2151                          */
2152                         svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2153                         CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2154                                svcpt->scp_nrqbds_posted);
2155                 }
2156         }
2157
2158         /*
2159         lc_watchdog_delete(thread->t_watchdog);
2160         thread->t_watchdog = NULL;
2161         */
2162
2163 out_srv_fini:
2164         /*
2165          * deconstruct service specific state created by ptlrpc_start_thread()
2166          */
2167         if (svc->srv_ops.so_thr_done)
2168                 svc->srv_ops.so_thr_done(thread);
2169
2170         if (env) {
2171                 lu_context_fini(&env->le_ctx);
2172                 kfree(env);
2173         }
2174 out:
2175         CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2176                thread, thread->t_pid, thread->t_id, rc);
2177
2178         spin_lock(&svcpt->scp_lock);
2179         if (thread_test_and_clear_flags(thread, SVC_STARTING))
2180                 svcpt->scp_nthrs_starting--;
2181
2182         if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2183                 /* must know immediately */
2184                 svcpt->scp_nthrs_running--;
2185         }
2186
2187         thread->t_id = rc;
2188         thread_add_flags(thread, SVC_STOPPED);
2189
2190         wake_up(&thread->t_ctl_waitq);
2191         spin_unlock(&svcpt->scp_lock);
2192
2193         return rc;
2194 }
2195
2196 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2197                           struct list_head *replies)
2198 {
2199         int result;
2200
2201         spin_lock(&hrt->hrt_lock);
2202
2203         list_splice_init(&hrt->hrt_queue, replies);
2204         result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2205
2206         spin_unlock(&hrt->hrt_lock);
2207         return result;
2208 }
2209
2210 /**
2211  * Main body of "handle reply" function.
2212  * It processes acked reply states
2213  */
2214 static int ptlrpc_hr_main(void *arg)
2215 {
2216         struct ptlrpc_hr_thread *hrt = arg;
2217         struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2218         LIST_HEAD(replies);
2219         char threadname[20];
2220         int rc;
2221
2222         snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2223                  hrp->hrp_cpt, hrt->hrt_id);
2224         unshare_fs_struct();
2225
2226         rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2227         if (rc != 0) {
2228                 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2229                       threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2230         }
2231
2232         atomic_inc(&hrp->hrp_nstarted);
2233         wake_up(&ptlrpc_hr.hr_waitq);
2234
2235         while (!ptlrpc_hr.hr_stopping) {
2236                 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2237
2238                 while (!list_empty(&replies)) {
2239                         struct ptlrpc_reply_state *rs;
2240
2241                         rs = list_entry(replies.prev, struct ptlrpc_reply_state,
2242                                         rs_list);
2243                         list_del_init(&rs->rs_list);
2244                         ptlrpc_handle_rs(rs);
2245                 }
2246         }
2247
2248         atomic_inc(&hrp->hrp_nstopped);
2249         wake_up(&ptlrpc_hr.hr_waitq);
2250
2251         return 0;
2252 }
2253
2254 static void ptlrpc_stop_hr_threads(void)
2255 {
2256         struct ptlrpc_hr_partition *hrp;
2257         int i;
2258         int j;
2259
2260         ptlrpc_hr.hr_stopping = 1;
2261
2262         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2263                 if (!hrp->hrp_thrs)
2264                         continue; /* uninitialized */
2265                 for (j = 0; j < hrp->hrp_nthrs; j++)
2266                         wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2267         }
2268
2269         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2270                 if (!hrp->hrp_thrs)
2271                         continue; /* uninitialized */
2272                 wait_event(ptlrpc_hr.hr_waitq,
2273                            atomic_read(&hrp->hrp_nstopped) ==
2274                            atomic_read(&hrp->hrp_nstarted));
2275         }
2276 }
2277
2278 static int ptlrpc_start_hr_threads(void)
2279 {
2280         struct ptlrpc_hr_partition *hrp;
2281         int i;
2282         int j;
2283
2284         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2285                 int rc = 0;
2286
2287                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2288                         struct  ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2289                         struct task_struct *task;
2290
2291                         task = kthread_run(ptlrpc_hr_main,
2292                                            &hrp->hrp_thrs[j],
2293                                            "ptlrpc_hr%02d_%03d",
2294                                            hrp->hrp_cpt, hrt->hrt_id);
2295                         if (IS_ERR(task)) {
2296                                 rc = PTR_ERR(task);
2297                                 break;
2298                         }
2299                 }
2300                 wait_event(ptlrpc_hr.hr_waitq,
2301                            atomic_read(&hrp->hrp_nstarted) == j);
2302
2303                 if (rc < 0) {
2304                         CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
2305                                i, j, rc);
2306                         ptlrpc_stop_hr_threads();
2307                         return rc;
2308                 }
2309         }
2310         return 0;
2311 }
2312
2313 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2314 {
2315         struct l_wait_info lwi = { 0 };
2316         struct ptlrpc_thread *thread;
2317         LIST_HEAD(zombie);
2318
2319         CDEBUG(D_INFO, "Stopping threads for service %s\n",
2320                svcpt->scp_service->srv_name);
2321
2322         spin_lock(&svcpt->scp_lock);
2323         /* let the thread know that we would like it to stop asap */
2324         list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2325                 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2326                        svcpt->scp_service->srv_thread_name, thread->t_id);
2327                 thread_add_flags(thread, SVC_STOPPING);
2328         }
2329
2330         wake_up_all(&svcpt->scp_waitq);
2331
2332         while (!list_empty(&svcpt->scp_threads)) {
2333                 thread = list_entry(svcpt->scp_threads.next,
2334                                     struct ptlrpc_thread, t_link);
2335                 if (thread_is_stopped(thread)) {
2336                         list_del(&thread->t_link);
2337                         list_add(&thread->t_link, &zombie);
2338                         continue;
2339                 }
2340                 spin_unlock(&svcpt->scp_lock);
2341
2342                 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2343                        svcpt->scp_service->srv_thread_name, thread->t_id);
2344                 l_wait_event(thread->t_ctl_waitq,
2345                              thread_is_stopped(thread), &lwi);
2346
2347                 spin_lock(&svcpt->scp_lock);
2348         }
2349
2350         spin_unlock(&svcpt->scp_lock);
2351
2352         while (!list_empty(&zombie)) {
2353                 thread = list_entry(zombie.next,
2354                                     struct ptlrpc_thread, t_link);
2355                 list_del(&thread->t_link);
2356                 kfree(thread);
2357         }
2358 }
2359
2360 /**
2361  * Stops all threads of a particular service \a svc
2362  */
2363 static void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2364 {
2365         struct ptlrpc_service_part *svcpt;
2366         int i;
2367
2368         ptlrpc_service_for_each_part(svcpt, i, svc) {
2369                 if (svcpt->scp_service)
2370                         ptlrpc_svcpt_stop_threads(svcpt);
2371         }
2372 }
2373
2374 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2375 {
2376         int rc = 0;
2377         int i;
2378         int j;
2379
2380         /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2381         LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2382
2383         for (i = 0; i < svc->srv_ncpts; i++) {
2384                 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2385                         rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2386                         if (rc == 0)
2387                                 continue;
2388
2389                         if (rc != -EMFILE)
2390                                 goto failed;
2391                         /* We have enough threads, don't start more. b=15759 */
2392                         break;
2393                 }
2394         }
2395
2396         return 0;
2397  failed:
2398         CERROR("cannot start %s thread #%d_%d: rc %d\n",
2399                svc->srv_thread_name, i, j, rc);
2400         ptlrpc_stop_all_threads(svc);
2401         return rc;
2402 }
2403
2404 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2405 {
2406         struct l_wait_info lwi = { 0 };
2407         struct ptlrpc_thread *thread;
2408         struct ptlrpc_service *svc;
2409         struct task_struct *task;
2410         int rc;
2411
2412         svc = svcpt->scp_service;
2413
2414         CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2415                svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2416                svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2417
2418  again:
2419         if (unlikely(svc->srv_is_stopping))
2420                 return -ESRCH;
2421
2422         if (!ptlrpc_threads_increasable(svcpt) ||
2423             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2424              svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2425                 return -EMFILE;
2426
2427         thread = kzalloc_node(sizeof(*thread), GFP_NOFS,
2428                               cfs_cpt_spread_node(svc->srv_cptable,
2429                                                   svcpt->scp_cpt));
2430         if (!thread)
2431                 return -ENOMEM;
2432         init_waitqueue_head(&thread->t_ctl_waitq);
2433
2434         spin_lock(&svcpt->scp_lock);
2435         if (!ptlrpc_threads_increasable(svcpt)) {
2436                 spin_unlock(&svcpt->scp_lock);
2437                 kfree(thread);
2438                 return -EMFILE;
2439         }
2440
2441         if (svcpt->scp_nthrs_starting != 0) {
2442                 /* serialize starting because some modules (obdfilter)
2443                  * might require unique and contiguous t_id
2444                  */
2445                 LASSERT(svcpt->scp_nthrs_starting == 1);
2446                 spin_unlock(&svcpt->scp_lock);
2447                 kfree(thread);
2448                 if (wait) {
2449                         CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2450                                svc->srv_thread_name, svcpt->scp_thr_nextid);
2451                         schedule();
2452                         goto again;
2453                 }
2454
2455                 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2456                        svc->srv_thread_name, svcpt->scp_thr_nextid);
2457                 return -EAGAIN;
2458         }
2459
2460         svcpt->scp_nthrs_starting++;
2461         thread->t_id = svcpt->scp_thr_nextid++;
2462         thread_add_flags(thread, SVC_STARTING);
2463         thread->t_svcpt = svcpt;
2464
2465         list_add(&thread->t_link, &svcpt->scp_threads);
2466         spin_unlock(&svcpt->scp_lock);
2467
2468         if (svcpt->scp_cpt >= 0) {
2469                 snprintf(thread->t_name, sizeof(thread->t_name), "%s%02d_%03d",
2470                          svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2471         } else {
2472                 snprintf(thread->t_name, sizeof(thread->t_name), "%s_%04d",
2473                          svc->srv_thread_name, thread->t_id);
2474         }
2475
2476         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2477         task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2478         if (IS_ERR(task)) {
2479                 rc = PTR_ERR(task);
2480                 CERROR("cannot start thread '%s': rc = %d\n",
2481                        thread->t_name, rc);
2482                 spin_lock(&svcpt->scp_lock);
2483                 --svcpt->scp_nthrs_starting;
2484                 if (thread_is_stopping(thread)) {
2485                         /* this ptlrpc_thread is being handled
2486                          * by ptlrpc_svcpt_stop_threads now
2487                          */
2488                         thread_add_flags(thread, SVC_STOPPED);
2489                         wake_up(&thread->t_ctl_waitq);
2490                         spin_unlock(&svcpt->scp_lock);
2491                 } else {
2492                         list_del(&thread->t_link);
2493                         spin_unlock(&svcpt->scp_lock);
2494                         kfree(thread);
2495                 }
2496                 return rc;
2497         }
2498
2499         if (!wait)
2500                 return 0;
2501
2502         l_wait_event(thread->t_ctl_waitq,
2503                      thread_is_running(thread) || thread_is_stopped(thread),
2504                      &lwi);
2505
2506         rc = thread_is_stopped(thread) ? thread->t_id : 0;
2507         return rc;
2508 }
2509
2510 int ptlrpc_hr_init(void)
2511 {
2512         struct ptlrpc_hr_partition *hrp;
2513         struct ptlrpc_hr_thread *hrt;
2514         int rc;
2515         int i;
2516         int j;
2517         int weight;
2518
2519         memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2520         ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2521
2522         ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2523                                                    sizeof(*hrp));
2524         if (!ptlrpc_hr.hr_partitions)
2525                 return -ENOMEM;
2526
2527         init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2528
2529         weight = cpumask_weight(topology_sibling_cpumask(0));
2530
2531         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2532                 hrp->hrp_cpt = i;
2533
2534                 atomic_set(&hrp->hrp_nstarted, 0);
2535                 atomic_set(&hrp->hrp_nstopped, 0);
2536
2537                 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2538                 hrp->hrp_nthrs /= weight;
2539
2540                 LASSERT(hrp->hrp_nthrs > 0);
2541                 hrp->hrp_thrs =
2542                         kzalloc_node(hrp->hrp_nthrs * sizeof(*hrt), GFP_NOFS,
2543                                      cfs_cpt_spread_node(ptlrpc_hr.hr_cpt_table,
2544                                                          i));
2545                 if (!hrp->hrp_thrs) {
2546                         rc = -ENOMEM;
2547                         goto out;
2548                 }
2549
2550                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2551                         hrt = &hrp->hrp_thrs[j];
2552
2553                         hrt->hrt_id = j;
2554                         hrt->hrt_partition = hrp;
2555                         init_waitqueue_head(&hrt->hrt_waitq);
2556                         spin_lock_init(&hrt->hrt_lock);
2557                         INIT_LIST_HEAD(&hrt->hrt_queue);
2558                 }
2559         }
2560
2561         rc = ptlrpc_start_hr_threads();
2562 out:
2563         if (rc != 0)
2564                 ptlrpc_hr_fini();
2565         return rc;
2566 }
2567
2568 void ptlrpc_hr_fini(void)
2569 {
2570         struct ptlrpc_hr_partition *hrp;
2571         int i;
2572
2573         if (!ptlrpc_hr.hr_partitions)
2574                 return;
2575
2576         ptlrpc_stop_hr_threads();
2577
2578         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2579                 kfree(hrp->hrp_thrs);
2580         }
2581
2582         cfs_percpt_free(ptlrpc_hr.hr_partitions);
2583         ptlrpc_hr.hr_partitions = NULL;
2584 }
2585
2586 /**
2587  * Wait until all already scheduled replies are processed.
2588  */
2589 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2590 {
2591         while (1) {
2592                 int rc;
2593                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2594                                                      NULL, NULL);
2595
2596                 rc = l_wait_event(svcpt->scp_waitq,
2597                                   atomic_read(&svcpt->scp_nreps_difficult) == 0,
2598                                   &lwi);
2599                 if (rc == 0)
2600                         break;
2601                 CWARN("Unexpectedly long timeout %s %p\n",
2602                       svcpt->scp_service->srv_name, svcpt->scp_service);
2603         }
2604 }
2605
2606 static void
2607 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2608 {
2609         struct ptlrpc_service_part *svcpt;
2610         int i;
2611
2612         /* early disarm AT timer... */
2613         ptlrpc_service_for_each_part(svcpt, i, svc) {
2614                 if (svcpt->scp_service)
2615                         del_timer(&svcpt->scp_at_timer);
2616         }
2617 }
2618
2619 static void
2620 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2621 {
2622         struct ptlrpc_service_part *svcpt;
2623         struct ptlrpc_request_buffer_desc *rqbd;
2624         struct l_wait_info lwi;
2625         int rc;
2626         int i;
2627
2628         /* All history will be culled when the next request buffer is
2629          * freed in ptlrpc_service_purge_all()
2630          */
2631         svc->srv_hist_nrqbds_cpt_max = 0;
2632
2633         rc = LNetClearLazyPortal(svc->srv_req_portal);
2634         LASSERT(rc == 0);
2635
2636         ptlrpc_service_for_each_part(svcpt, i, svc) {
2637                 if (!svcpt->scp_service)
2638                         break;
2639
2640                 /* Unlink all the request buffers.  This forces a 'final'
2641                  * event with its 'unlink' flag set for each posted rqbd
2642                  */
2643                 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2644                                     rqbd_list) {
2645                         rc = LNetMDUnlink(rqbd->rqbd_md_h);
2646                         LASSERT(rc == 0 || rc == -ENOENT);
2647                 }
2648         }
2649
2650         ptlrpc_service_for_each_part(svcpt, i, svc) {
2651                 if (!svcpt->scp_service)
2652                         break;
2653
2654                 /* Wait for the network to release any buffers
2655                  * it's currently filling
2656                  */
2657                 spin_lock(&svcpt->scp_lock);
2658                 while (svcpt->scp_nrqbds_posted != 0) {
2659                         spin_unlock(&svcpt->scp_lock);
2660                         /* Network access will complete in finite time but
2661                          * the HUGE timeout lets us CWARN for visibility
2662                          * of sluggish LNDs
2663                          */
2664                         lwi = LWI_TIMEOUT_INTERVAL(
2665                                         cfs_time_seconds(LONG_UNLINK),
2666                                         cfs_time_seconds(1), NULL, NULL);
2667                         rc = l_wait_event(svcpt->scp_waitq,
2668                                           svcpt->scp_nrqbds_posted == 0, &lwi);
2669                         if (rc == -ETIMEDOUT) {
2670                                 CWARN("Service %s waiting for request buffers\n",
2671                                       svcpt->scp_service->srv_name);
2672                         }
2673                         spin_lock(&svcpt->scp_lock);
2674                 }
2675                 spin_unlock(&svcpt->scp_lock);
2676         }
2677 }
2678
2679 static void
2680 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2681 {
2682         struct ptlrpc_service_part *svcpt;
2683         struct ptlrpc_request_buffer_desc *rqbd;
2684         struct ptlrpc_request *req;
2685         struct ptlrpc_reply_state *rs;
2686         int i;
2687
2688         ptlrpc_service_for_each_part(svcpt, i, svc) {
2689                 if (!svcpt->scp_service)
2690                         break;
2691
2692                 spin_lock(&svcpt->scp_rep_lock);
2693                 while (!list_empty(&svcpt->scp_rep_active)) {
2694                         rs = list_entry(svcpt->scp_rep_active.next,
2695                                         struct ptlrpc_reply_state, rs_list);
2696                         spin_lock(&rs->rs_lock);
2697                         ptlrpc_schedule_difficult_reply(rs);
2698                         spin_unlock(&rs->rs_lock);
2699                 }
2700                 spin_unlock(&svcpt->scp_rep_lock);
2701
2702                 /* purge the request queue.  NB No new replies (rqbds
2703                  * all unlinked) and no service threads, so I'm the only
2704                  * thread noodling the request queue now
2705                  */
2706                 while (!list_empty(&svcpt->scp_req_incoming)) {
2707                         req = list_entry(svcpt->scp_req_incoming.next,
2708                                          struct ptlrpc_request, rq_list);
2709
2710                         list_del(&req->rq_list);
2711                         svcpt->scp_nreqs_incoming--;
2712                         ptlrpc_server_finish_request(svcpt, req);
2713                 }
2714
2715                 while (ptlrpc_server_request_pending(svcpt, true)) {
2716                         req = ptlrpc_server_request_get(svcpt, true);
2717                         ptlrpc_server_finish_active_request(svcpt, req);
2718                 }
2719
2720                 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
2721                 LASSERT(svcpt->scp_nreqs_incoming == 0);
2722                 LASSERT(svcpt->scp_nreqs_active == 0);
2723                 /* history should have been culled by
2724                  * ptlrpc_server_finish_request
2725                  */
2726                 LASSERT(svcpt->scp_hist_nrqbds == 0);
2727
2728                 /* Now free all the request buffers since nothing
2729                  * references them any more...
2730                  */
2731
2732                 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2733                         rqbd = list_entry(svcpt->scp_rqbd_idle.next,
2734                                           struct ptlrpc_request_buffer_desc,
2735                                           rqbd_list);
2736                         ptlrpc_free_rqbd(rqbd);
2737                 }
2738                 ptlrpc_wait_replies(svcpt);
2739
2740                 while (!list_empty(&svcpt->scp_rep_idle)) {
2741                         rs = list_entry(svcpt->scp_rep_idle.next,
2742                                         struct ptlrpc_reply_state,
2743                                         rs_list);
2744                         list_del(&rs->rs_list);
2745                         kvfree(rs);
2746                 }
2747         }
2748 }
2749
2750 static void
2751 ptlrpc_service_free(struct ptlrpc_service *svc)
2752 {
2753         struct ptlrpc_service_part *svcpt;
2754         struct ptlrpc_at_array *array;
2755         int i;
2756
2757         ptlrpc_service_for_each_part(svcpt, i, svc) {
2758                 if (!svcpt->scp_service)
2759                         break;
2760
2761                 /* In case somebody rearmed this in the meantime */
2762                 del_timer(&svcpt->scp_at_timer);
2763                 array = &svcpt->scp_at_array;
2764
2765                 kfree(array->paa_reqs_array);
2766                 array->paa_reqs_array = NULL;
2767                 kfree(array->paa_reqs_count);
2768                 array->paa_reqs_count = NULL;
2769         }
2770
2771         ptlrpc_service_for_each_part(svcpt, i, svc)
2772                 kfree(svcpt);
2773
2774         if (svc->srv_cpts)
2775                 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
2776
2777         kfree(svc);
2778 }
2779
2780 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2781 {
2782         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2783
2784         service->srv_is_stopping = 1;
2785
2786         mutex_lock(&ptlrpc_all_services_mutex);
2787         list_del_init(&service->srv_list);
2788         mutex_unlock(&ptlrpc_all_services_mutex);
2789
2790         ptlrpc_service_del_atimer(service);
2791         ptlrpc_stop_all_threads(service);
2792
2793         ptlrpc_service_unlink_rqbd(service);
2794         ptlrpc_service_purge_all(service);
2795         ptlrpc_service_nrs_cleanup(service);
2796
2797         ptlrpc_lprocfs_unregister_service(service);
2798         ptlrpc_sysfs_unregister_service(service);
2799
2800         ptlrpc_service_free(service);
2801
2802         return 0;
2803 }
2804 EXPORT_SYMBOL(ptlrpc_unregister_service);