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
  35 #include <obd_support.h>
  36 #include <obd_class.h>
  37 #include <lustre_net.h>
  38 #include <lu_object.h>
  39 #include <uapi/linux/lnet/lnet-types.h>
  40 #include "ptlrpc_internal.h"
  41
  42 /* The following are visible and mutable through /sys/module/ptlrpc */
  43 int test_req_buffer_pressure;
  44 module_param(test_req_buffer_pressure, int, 0444);
  45 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
  46 module_param(at_min, int, 0644);
  47 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
  48 module_param(at_max, int, 0644);
  49 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
  50 module_param(at_history, int, 0644);
  51 MODULE_PARM_DESC(at_history,
  52                  "Adaptive timeouts remember the slowest event that took place within this period (sec)");
  53 module_param(at_early_margin, int, 0644);
  54 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
  55 module_param(at_extra, int, 0644);
  56 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
  57
  58 /* forward ref */
  59 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
  60 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
  61 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
  62
  63 /** Holds a list of all PTLRPC services */
  64 LIST_HEAD(ptlrpc_all_services);
  65 /** Used to protect the \e ptlrpc_all_services list */
  66 struct mutex ptlrpc_all_services_mutex;
  67
  68 static struct ptlrpc_request_buffer_desc *
  69 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
  70 {
  71         struct ptlrpc_service *svc = svcpt->scp_service;
  72         struct ptlrpc_request_buffer_desc *rqbd;
  73
  74         rqbd = kzalloc_node(sizeof(*rqbd), GFP_NOFS,
  75                             cfs_cpt_spread_node(svc->srv_cptable,
  76                                                 svcpt->scp_cpt));
  77         if (!rqbd)
  78                 return NULL;
  79
  80         rqbd->rqbd_svcpt = svcpt;
  81         rqbd->rqbd_refcount = 0;
  82         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
  83         rqbd->rqbd_cbid.cbid_arg = rqbd;
  84         INIT_LIST_HEAD(&rqbd->rqbd_reqs);
  85         rqbd->rqbd_buffer = libcfs_kvzalloc_cpt(svc->srv_cptable,
  86                                                 svcpt->scp_cpt,
  87                                                 svc->srv_buf_size,
  88                                                 GFP_KERNEL);
  89         if (!rqbd->rqbd_buffer) {
  90                 kfree(rqbd);
  91                 return NULL;
  92         }
  93
  94         spin_lock(&svcpt->scp_lock);
  95         list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
  96         svcpt->scp_nrqbds_total++;
  97         spin_unlock(&svcpt->scp_lock);
  98
  99         return rqbd;
 100 }
 101
 102 static void
 103 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 104 {
 105         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
 106
 107         LASSERT(rqbd->rqbd_refcount == 0);
 108         LASSERT(list_empty(&rqbd->rqbd_reqs));
 109
 110         spin_lock(&svcpt->scp_lock);
 111         list_del(&rqbd->rqbd_list);
 112         svcpt->scp_nrqbds_total--;
 113         spin_unlock(&svcpt->scp_lock);
 114
 115         kvfree(rqbd->rqbd_buffer);
 116         kfree(rqbd);
 117 }
 118
 119 static int
 120 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
 121 {
 122         struct ptlrpc_service *svc = svcpt->scp_service;
 123         struct ptlrpc_request_buffer_desc *rqbd;
 124         int rc = 0;
 125         int i;
 126
 127         if (svcpt->scp_rqbd_allocating)
 128                 goto try_post;
 129
 130         spin_lock(&svcpt->scp_lock);
 131         /* check again with lock */
 132         if (svcpt->scp_rqbd_allocating) {
 133                 /* NB: we might allow more than one thread in the future */
 134                 LASSERT(svcpt->scp_rqbd_allocating == 1);
 135                 spin_unlock(&svcpt->scp_lock);
 136                 goto try_post;
 137         }
 138
 139         svcpt->scp_rqbd_allocating++;
 140         spin_unlock(&svcpt->scp_lock);
 141
 142         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 143                 /* NB: another thread might have recycled enough rqbds, we
 144                  * need to make sure it wouldn't over-allocate, see LU-1212.
 145                  */
 146                 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
 147                         break;
 148
 149                 rqbd = ptlrpc_alloc_rqbd(svcpt);
 150
 151                 if (!rqbd) {
 152                         CERROR("%s: Can't allocate request buffer\n",
 153                                svc->srv_name);
 154                         rc = -ENOMEM;
 155                         break;
 156                 }
 157         }
 158
 159         spin_lock(&svcpt->scp_lock);
 160
 161         LASSERT(svcpt->scp_rqbd_allocating == 1);
 162         svcpt->scp_rqbd_allocating--;
 163
 164         spin_unlock(&svcpt->scp_lock);
 165
 166         CDEBUG(D_RPCTRACE,
 167                "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
 168                svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
 169                svcpt->scp_nrqbds_total, rc);
 170
 171  try_post:
 172         if (post && rc == 0)
 173                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
 174
 175         return rc;
 176 }
 177
 178 struct ptlrpc_hr_partition;
 179
 180 struct ptlrpc_hr_thread {
 181         int                             hrt_id;         /* thread ID */
 182         spinlock_t                      hrt_lock;
 183         wait_queue_head_t                       hrt_waitq;
 184         struct list_head                        hrt_queue;      /* RS queue */
 185         struct ptlrpc_hr_partition      *hrt_partition;
 186 };
 187
 188 struct ptlrpc_hr_partition {
 189         /* # of started threads */
 190         atomic_t                        hrp_nstarted;
 191         /* # of stopped threads */
 192         atomic_t                        hrp_nstopped;
 193         /* cpu partition id */
 194         int                             hrp_cpt;
 195         /* round-robin rotor for choosing thread */
 196         int                             hrp_rotor;
 197         /* total number of threads on this partition */
 198         int                             hrp_nthrs;
 199         /* threads table */
 200         struct ptlrpc_hr_thread         *hrp_thrs;
 201 };
 202
 203 #define HRT_RUNNING 0
 204 #define HRT_STOPPING 1
 205
 206 struct ptlrpc_hr_service {
 207         /* CPU partition table, it's just cfs_cpt_table for now */
 208         struct cfs_cpt_table            *hr_cpt_table;
 209         /** controller sleep waitq */
 210         wait_queue_head_t                       hr_waitq;
 211         unsigned int                    hr_stopping;
 212         /** roundrobin rotor for non-affinity service */
 213         unsigned int                    hr_rotor;
 214         /* partition data */
 215         struct ptlrpc_hr_partition      **hr_partitions;
 216 };
 217
 218 /** reply handling service. */
 219 static struct ptlrpc_hr_service         ptlrpc_hr;
 220
 221 /**
 222  * Choose an hr thread to dispatch requests to.
 223  */
 224 static struct ptlrpc_hr_thread *
 225 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
 226 {
 227         struct ptlrpc_hr_partition *hrp;
 228         unsigned int rotor;
 229
 230         if (svcpt->scp_cpt >= 0 &&
 231             svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
 232                 /* directly match partition */
 233                 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
 234
 235         } else {
 236                 rotor = ptlrpc_hr.hr_rotor++;
 237                 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
 238
 239                 hrp = ptlrpc_hr.hr_partitions[rotor];
 240         }
 241
 242         rotor = hrp->hrp_rotor++;
 243         return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
 244 }
 245
 246 /**
 247  * Put reply state into a queue for processing because we received
 248  * ACK from the client
 249  */
 250 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
 251 {
 252         struct ptlrpc_hr_thread *hrt;
 253
 254         LASSERT(list_empty(&rs->rs_list));
 255
 256         hrt = ptlrpc_hr_select(rs->rs_svcpt);
 257
 258         spin_lock(&hrt->hrt_lock);
 259         list_add_tail(&rs->rs_list, &hrt->hrt_queue);
 260         spin_unlock(&hrt->hrt_lock);
 261
 262         wake_up(&hrt->hrt_waitq);
 263 }
 264
 265 void
 266 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
 267 {
 268         assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
 269         assert_spin_locked(&rs->rs_lock);
 270         LASSERT(rs->rs_difficult);
 271         rs->rs_scheduled_ever = 1;  /* flag any notification attempt */
 272
 273         if (rs->rs_scheduled) {     /* being set up or already notified */
 274                 return;
 275         }
 276
 277         rs->rs_scheduled = 1;
 278         list_del_init(&rs->rs_list);
 279         ptlrpc_dispatch_difficult_reply(rs);
 280 }
 281 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
 282
 283 static int
 284 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
 285 {
 286         struct ptlrpc_request_buffer_desc *rqbd;
 287         int rc;
 288         int posted = 0;
 289
 290         for (;;) {
 291                 spin_lock(&svcpt->scp_lock);
 292
 293                 if (list_empty(&svcpt->scp_rqbd_idle)) {
 294                         spin_unlock(&svcpt->scp_lock);
 295                         return posted;
 296                 }
 297
 298                 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
 299                                   struct ptlrpc_request_buffer_desc,
 300                                   rqbd_list);
 301                 list_del(&rqbd->rqbd_list);
 302
 303                 /* assume we will post successfully */
 304                 svcpt->scp_nrqbds_posted++;
 305                 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
 306
 307                 spin_unlock(&svcpt->scp_lock);
 308
 309                 rc = ptlrpc_register_rqbd(rqbd);
 310                 if (rc != 0)
 311                         break;
 312
 313                 posted = 1;
 314         }
 315
 316         spin_lock(&svcpt->scp_lock);
 317
 318         svcpt->scp_nrqbds_posted--;
 319         list_del(&rqbd->rqbd_list);
 320         list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
 321
 322         /* Don't complain if no request buffers are posted right now; LNET
 323          * won't drop requests because we set the portal lazy!
 324          */
 325
 326         spin_unlock(&svcpt->scp_lock);
 327
 328         return -1;
 329 }
 330
 331 static void ptlrpc_at_timer(unsigned long castmeharder)
 332 {
 333         struct ptlrpc_service_part *svcpt;
 334
 335         svcpt = (struct ptlrpc_service_part *)castmeharder;
 336
 337         svcpt->scp_at_check = 1;
 338         svcpt->scp_at_checktime = cfs_time_current();
 339         wake_up(&svcpt->scp_waitq);
 340 }
 341
 342 static void
 343 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
 344                              struct ptlrpc_service_conf *conf)
 345 {
 346         struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
 347         unsigned int init;
 348         unsigned int total;
 349         unsigned int nthrs;
 350         int weight;
 351
 352         /*
 353          * Common code for estimating & validating threads number.
 354          * CPT affinity service could have percpt thread-pool instead
 355          * of a global thread-pool, which means user might not always
 356          * get the threads number they give it in conf::tc_nthrs_user
 357          * even they did set. It's because we need to validate threads
 358          * number for each CPT to guarantee each pool will have enough
 359          * threads to keep the service healthy.
 360          */
 361         init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
 362         init = max_t(int, init, tc->tc_nthrs_init);
 363
 364         /* NB: please see comments in lustre_lnet.h for definition
 365          * details of these members
 366          */
 367         LASSERT(tc->tc_nthrs_max != 0);
 368
 369         if (tc->tc_nthrs_user != 0) {
 370                 /* In case there is a reason to test a service with many
 371                  * threads, we give a less strict check here, it can
 372                  * be up to 8 * nthrs_max
 373                  */
 374                 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
 375                 nthrs = total / svc->srv_ncpts;
 376                 init = max(init, nthrs);
 377                 goto out;
 378         }
 379
 380         total = tc->tc_nthrs_max;
 381         if (tc->tc_nthrs_base == 0) {
 382                 /* don't care about base threads number per partition,
 383                  * this is most for non-affinity service
 384                  */
 385                 nthrs = total / svc->srv_ncpts;
 386                 goto out;
 387         }
 388
 389         nthrs = tc->tc_nthrs_base;
 390         if (svc->srv_ncpts == 1) {
 391                 int i;
 392
 393                 /* NB: Increase the base number if it's single partition
 394                  * and total number of cores/HTs is larger or equal to 4.
 395                  * result will always < 2 * nthrs_base
 396                  */
 397                 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
 398                 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
 399                             (tc->tc_nthrs_base >> i) != 0; i++)
 400                         nthrs += tc->tc_nthrs_base >> i;
 401         }
 402
 403         if (tc->tc_thr_factor != 0) {
 404                 int factor = tc->tc_thr_factor;
 405                 const int fade = 4;
 406
 407                 /*
 408                  * User wants to increase number of threads with for
 409                  * each CPU core/HT, most likely the factor is larger then
 410                  * one thread/core because service threads are supposed to
 411                  * be blocked by lock or wait for IO.
 412                  */
 413                 /*
 414                  * Amdahl's law says that adding processors wouldn't give
 415                  * a linear increasing of parallelism, so it's nonsense to
 416                  * have too many threads no matter how many cores/HTs
 417                  * there are.
 418                  */
 419                 /* weight is # of HTs */
 420                 if (cpumask_weight(topology_sibling_cpumask(0)) > 1) {
 421                         /* depress thread factor for hyper-thread */
 422                         factor = factor - (factor >> 1) + (factor >> 3);
 423                 }
 424
 425                 weight = cfs_cpt_weight(svc->srv_cptable, 0);
 426                 LASSERT(weight > 0);
 427
 428                 for (; factor > 0 && weight > 0; factor--, weight -= fade)
 429                         nthrs += min(weight, fade) * factor;
 430         }
 431
 432         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
 433                 nthrs = max(tc->tc_nthrs_base,
 434                             tc->tc_nthrs_max / svc->srv_ncpts);
 435         }
 436  out:
 437         nthrs = max(nthrs, tc->tc_nthrs_init);
 438         svc->srv_nthrs_cpt_limit = nthrs;
 439         svc->srv_nthrs_cpt_init = init;
 440
 441         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
 442                 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
 443                        svc->srv_name, nthrs * svc->srv_ncpts,
 444                        tc->tc_nthrs_max);
 445         }
 446 }
 447
 448 /**
 449  * Initialize percpt data for a service
 450  */
 451 static int
 452 ptlrpc_service_part_init(struct ptlrpc_service *svc,
 453                          struct ptlrpc_service_part *svcpt, int cpt)
 454 {
 455         struct ptlrpc_at_array  *array;
 456         int size;
 457         int index;
 458         int rc;
 459
 460         svcpt->scp_cpt = cpt;
 461         INIT_LIST_HEAD(&svcpt->scp_threads);
 462
 463         /* rqbd and incoming request queue */
 464         spin_lock_init(&svcpt->scp_lock);
 465         INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
 466         INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
 467         INIT_LIST_HEAD(&svcpt->scp_req_incoming);
 468         init_waitqueue_head(&svcpt->scp_waitq);
 469         /* history request & rqbd list */
 470         INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
 471         INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
 472
 473         /* active requests and hp requests */
 474         spin_lock_init(&svcpt->scp_req_lock);
 475
 476         /* reply states */
 477         spin_lock_init(&svcpt->scp_rep_lock);
 478         INIT_LIST_HEAD(&svcpt->scp_rep_active);
 479         INIT_LIST_HEAD(&svcpt->scp_rep_idle);
 480         init_waitqueue_head(&svcpt->scp_rep_waitq);
 481         atomic_set(&svcpt->scp_nreps_difficult, 0);
 482
 483         /* adaptive timeout */
 484         spin_lock_init(&svcpt->scp_at_lock);
 485         array = &svcpt->scp_at_array;
 486
 487         size = at_est2timeout(at_max);
 488         array->paa_size = size;
 489         array->paa_count = 0;
 490         array->paa_deadline = -1;
 491
 492         /* allocate memory for scp_at_array (ptlrpc_at_array) */
 493         array->paa_reqs_array =
 494                 kzalloc_node(sizeof(struct list_head) * size, GFP_NOFS,
 495                              cfs_cpt_spread_node(svc->srv_cptable, cpt));
 496         if (!array->paa_reqs_array)
 497                 return -ENOMEM;
 498
 499         for (index = 0; index < size; index++)
 500                 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
 501
 502         array->paa_reqs_count =
 503                 kzalloc_node(sizeof(__u32) * size, GFP_NOFS,
 504                              cfs_cpt_spread_node(svc->srv_cptable, cpt));
 505         if (!array->paa_reqs_count)
 506                 goto free_reqs_array;
 507
 508         setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
 509                     (unsigned long)svcpt);
 510
 511         /* At SOW, service time should be quick; 10s seems generous. If client
 512          * timeout is less than this, we'll be sending an early reply.
 513          */
 514         at_init(&svcpt->scp_at_estimate, 10, 0);
 515
 516         /* assign this before call ptlrpc_grow_req_bufs */
 517         svcpt->scp_service = svc;
 518         /* Now allocate the request buffers, but don't post them now */
 519         rc = ptlrpc_grow_req_bufs(svcpt, 0);
 520         /* We shouldn't be under memory pressure at startup, so
 521          * fail if we can't allocate all our buffers at this time.
 522          */
 523         if (rc != 0)
 524                 goto free_reqs_count;
 525
 526         return 0;
 527
 528 free_reqs_count:
 529         kfree(array->paa_reqs_count);
 530         array->paa_reqs_count = NULL;
 531 free_reqs_array:
 532         kfree(array->paa_reqs_array);
 533         array->paa_reqs_array = NULL;
 534
 535         return -ENOMEM;
 536 }
 537
 538 /**
 539  * Initialize service on a given portal.
 540  * This includes starting serving threads , allocating and posting rqbds and
 541  * so on.
 542  */
 543 struct ptlrpc_service *
 544 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
 545                         struct kset *parent,
 546                         struct dentry *debugfs_entry)
 547 {
 548         struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
 549         struct ptlrpc_service *service;
 550         struct ptlrpc_service_part *svcpt;
 551         struct cfs_cpt_table *cptable;
 552         __u32 *cpts = NULL;
 553         int ncpts;
 554         int cpt;
 555         int rc;
 556         int i;
 557
 558         LASSERT(conf->psc_buf.bc_nbufs > 0);
 559         LASSERT(conf->psc_buf.bc_buf_size >=
 560                 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
 561         LASSERT(conf->psc_thr.tc_ctx_tags != 0);
 562
 563         cptable = cconf->cc_cptable;
 564         if (!cptable)
 565                 cptable = cfs_cpt_table;
 566
 567         if (!conf->psc_thr.tc_cpu_affinity) {
 568                 ncpts = 1;
 569         } else {
 570                 ncpts = cfs_cpt_number(cptable);
 571                 if (cconf->cc_pattern) {
 572                         struct cfs_expr_list *el;
 573
 574                         rc = cfs_expr_list_parse(cconf->cc_pattern,
 575                                                  strlen(cconf->cc_pattern),
 576                                                  0, ncpts - 1, &el);
 577                         if (rc != 0) {
 578                                 CERROR("%s: invalid CPT pattern string: %s",
 579                                        conf->psc_name, cconf->cc_pattern);
 580                                 return ERR_PTR(-EINVAL);
 581                         }
 582
 583                         rc = cfs_expr_list_values(el, ncpts, &cpts);
 584                         cfs_expr_list_free(el);
 585                         if (rc <= 0) {
 586                                 CERROR("%s: failed to parse CPT array %s: %d\n",
 587                                        conf->psc_name, cconf->cc_pattern, rc);
 588                                 kfree(cpts);
 589                                 return ERR_PTR(rc < 0 ? rc : -EINVAL);
 590                         }
 591                         ncpts = rc;
 592                 }
 593         }
 594
 595         service = kzalloc(offsetof(struct ptlrpc_service, srv_parts[ncpts]),
 596                           GFP_NOFS);
 597         if (!service) {
 598                 kfree(cpts);
 599                 return ERR_PTR(-ENOMEM);
 600         }
 601
 602         service->srv_cptable = cptable;
 603         service->srv_cpts = cpts;
 604         service->srv_ncpts = ncpts;
 605
 606         service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
 607         while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
 608                 service->srv_cpt_bits++;
 609
 610         /* public members */
 611         spin_lock_init(&service->srv_lock);
 612         service->srv_name = conf->psc_name;
 613         service->srv_watchdog_factor = conf->psc_watchdog_factor;
 614         INIT_LIST_HEAD(&service->srv_list); /* for safety of cleanup */
 615
 616         /* buffer configuration */
 617         service->srv_nbuf_per_group = test_req_buffer_pressure ?
 618                                           1 : conf->psc_buf.bc_nbufs;
 619         service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
 620                                           SPTLRPC_MAX_PAYLOAD;
 621         service->srv_buf_size = conf->psc_buf.bc_buf_size;
 622         service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
 623         service->srv_req_portal = conf->psc_buf.bc_req_portal;
 624
 625         /* Increase max reply size to next power of two */
 626         service->srv_max_reply_size = 1;
 627         while (service->srv_max_reply_size <
 628                conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
 629                 service->srv_max_reply_size <<= 1;
 630
 631         service->srv_thread_name = conf->psc_thr.tc_thr_name;
 632         service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
 633         service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
 634         service->srv_ops = conf->psc_ops;
 635
 636         for (i = 0; i < ncpts; i++) {
 637                 if (!conf->psc_thr.tc_cpu_affinity)
 638                         cpt = CFS_CPT_ANY;
 639                 else
 640                         cpt = cpts ? cpts[i] : i;
 641
 642                 svcpt = kzalloc_node(sizeof(*svcpt), GFP_NOFS,
 643                                      cfs_cpt_spread_node(cptable, cpt));
 644                 if (!svcpt) {
 645                         rc = -ENOMEM;
 646                         goto failed;
 647                 }
 648
 649                 service->srv_parts[i] = svcpt;
 650                 rc = ptlrpc_service_part_init(service, svcpt, cpt);
 651                 if (rc != 0)
 652                         goto failed;
 653         }
 654
 655         ptlrpc_server_nthreads_check(service, conf);
 656
 657         rc = LNetSetLazyPortal(service->srv_req_portal);
 658         LASSERT(rc == 0);
 659
 660         mutex_lock(&ptlrpc_all_services_mutex);
 661         list_add(&service->srv_list, &ptlrpc_all_services);
 662         mutex_unlock(&ptlrpc_all_services_mutex);
 663
 664         if (parent) {
 665                 rc = ptlrpc_sysfs_register_service(parent, service);
 666                 if (rc)
 667                         goto failed;
 668         }
 669
 670         if (!IS_ERR_OR_NULL(debugfs_entry))
 671                 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
 672
 673         rc = ptlrpc_service_nrs_setup(service);
 674         if (rc != 0)
 675                 goto failed;
 676
 677         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 678                service->srv_name, service->srv_req_portal);
 679
 680         rc = ptlrpc_start_threads(service);
 681         if (rc != 0) {
 682                 CERROR("Failed to start threads for service %s: %d\n",
 683                        service->srv_name, rc);
 684                 goto failed;
 685         }
 686
 687         return service;
 688 failed:
 689         ptlrpc_unregister_service(service);
 690         return ERR_PTR(rc);
 691 }
 692 EXPORT_SYMBOL(ptlrpc_register_service);
 693
 694 /**
 695  * to actually free the request, must be called without holding svc_lock.
 696  * note it's caller's responsibility to unlink req->rq_list.
 697  */
 698 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
 699 {
 700         LASSERT(atomic_read(&req->rq_refcount) == 0);
 701         LASSERT(list_empty(&req->rq_timed_list));
 702
 703          /* DEBUG_REQ() assumes the reply state of a request with a valid
 704           * ref will not be destroyed until that reference is dropped.
 705           */
 706         ptlrpc_req_drop_rs(req);
 707
 708         sptlrpc_svc_ctx_decref(req);
 709
 710         if (req != &req->rq_rqbd->rqbd_req) {
 711                 /* NB request buffers use an embedded
 712                  * req if the incoming req unlinked the
 713                  * MD; this isn't one of them!
 714                  */
 715                 ptlrpc_request_cache_free(req);
 716         }
 717 }
 718
 719 /**
 720  * drop a reference count of the request. if it reaches 0, we either
 721  * put it into history list, or free it immediately.
 722  */
 723 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
 724 {
 725         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 726         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
 727         struct ptlrpc_service *svc = svcpt->scp_service;
 728         int refcount;
 729         struct list_head *tmp;
 730         struct list_head *nxt;
 731
 732         if (!atomic_dec_and_test(&req->rq_refcount))
 733                 return;
 734
 735         if (req->rq_at_linked) {
 736                 spin_lock(&svcpt->scp_at_lock);
 737                 /* recheck with lock, in case it's unlinked by
 738                  * ptlrpc_at_check_timed()
 739                  */
 740                 if (likely(req->rq_at_linked))
 741                         ptlrpc_at_remove_timed(req);
 742                 spin_unlock(&svcpt->scp_at_lock);
 743         }
 744
 745         LASSERT(list_empty(&req->rq_timed_list));
 746
 747         /* finalize request */
 748         if (req->rq_export) {
 749                 class_export_put(req->rq_export);
 750                 req->rq_export = NULL;
 751         }
 752
 753         spin_lock(&svcpt->scp_lock);
 754
 755         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 756
 757         refcount = --(rqbd->rqbd_refcount);
 758         if (refcount == 0) {
 759                 /* request buffer is now idle: add to history */
 760                 list_del(&rqbd->rqbd_list);
 761
 762                 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
 763                 svcpt->scp_hist_nrqbds++;
 764
 765                 /* cull some history?
 766                  * I expect only about 1 or 2 rqbds need to be recycled here
 767                  */
 768                 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
 769                         rqbd = list_entry(svcpt->scp_hist_rqbds.next,
 770                                           struct ptlrpc_request_buffer_desc,
 771                                           rqbd_list);
 772
 773                         list_del(&rqbd->rqbd_list);
 774                         svcpt->scp_hist_nrqbds--;
 775
 776                         /* remove rqbd's reqs from svc's req history while
 777                          * I've got the service lock
 778                          */
 779                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 780                                 req = list_entry(tmp, struct ptlrpc_request,
 781                                                  rq_list);
 782                                 /* Track the highest culled req seq */
 783                                 if (req->rq_history_seq >
 784                                     svcpt->scp_hist_seq_culled) {
 785                                         svcpt->scp_hist_seq_culled =
 786                                                 req->rq_history_seq;
 787                                 }
 788                                 list_del(&req->rq_history_list);
 789                         }
 790
 791                         spin_unlock(&svcpt->scp_lock);
 792
 793                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 794                                 req = list_entry(rqbd->rqbd_reqs.next,
 795                                                  struct ptlrpc_request,
 796                                                  rq_list);
 797                                 list_del(&req->rq_list);
 798                                 ptlrpc_server_free_request(req);
 799                         }
 800
 801                         spin_lock(&svcpt->scp_lock);
 802                         /*
 803                          * now all reqs including the embedded req has been
 804                          * disposed, schedule request buffer for re-use.
 805                          */
 806                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) ==
 807                                 0);
 808                         list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
 809                 }
 810
 811                 spin_unlock(&svcpt->scp_lock);
 812         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 813                 /* If we are low on memory, we are not interested in history */
 814                 list_del(&req->rq_list);
 815                 list_del_init(&req->rq_history_list);
 816
 817                 /* Track the highest culled req seq */
 818                 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
 819                         svcpt->scp_hist_seq_culled = req->rq_history_seq;
 820
 821                 spin_unlock(&svcpt->scp_lock);
 822
 823                 ptlrpc_server_free_request(req);
 824         } else {
 825                 spin_unlock(&svcpt->scp_lock);
 826         }
 827 }
 828
 829 /**
 830  * to finish a request: stop sending more early replies, and release
 831  * the request.
 832  */
 833 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
 834                                          struct ptlrpc_request *req)
 835 {
 836         ptlrpc_server_hpreq_fini(req);
 837
 838         if (req->rq_session.lc_thread) {
 839                 lu_context_exit(&req->rq_session);
 840                 lu_context_fini(&req->rq_session);
 841         }
 842
 843         ptlrpc_server_drop_request(req);
 844 }
 845
 846 /**
 847  * to finish a active request: stop sending more early replies, and release
 848  * the request. should be called after we finished handling the request.
 849  */
 850 static void ptlrpc_server_finish_active_request(
 851                                         struct ptlrpc_service_part *svcpt,
 852                                         struct ptlrpc_request *req)
 853 {
 854         spin_lock(&svcpt->scp_req_lock);
 855         ptlrpc_nrs_req_stop_nolock(req);
 856         svcpt->scp_nreqs_active--;
 857         if (req->rq_hp)
 858                 svcpt->scp_nhreqs_active--;
 859         spin_unlock(&svcpt->scp_req_lock);
 860
 861         ptlrpc_nrs_req_finalize(req);
 862
 863         if (req->rq_export)
 864                 class_export_rpc_dec(req->rq_export);
 865
 866         ptlrpc_server_finish_request(svcpt, req);
 867 }
 868
 869 /**
 870  * Sanity check request \a req.
 871  * Return 0 if all is ok, error code otherwise.
 872  */
 873 static int ptlrpc_check_req(struct ptlrpc_request *req)
 874 {
 875         struct obd_device *obd = req->rq_export->exp_obd;
 876         int rc = 0;
 877
 878         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 879                      req->rq_export->exp_conn_cnt)) {
 880                 DEBUG_REQ(D_RPCTRACE, req,
 881                           "DROPPING req from old connection %d < %d",
 882                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 883                           req->rq_export->exp_conn_cnt);
 884                 return -EEXIST;
 885         }
 886         if (unlikely(!obd || obd->obd_fail)) {
 887                 /*
 888                  * Failing over, don't handle any more reqs, send
 889                  * error response instead.
 890                  */
 891                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 892                        req, obd ? obd->obd_name : "unknown");
 893                 rc = -ENODEV;
 894         } else if (lustre_msg_get_flags(req->rq_reqmsg) &
 895                    (MSG_REPLAY | MSG_REQ_REPLAY_DONE)) {
 896                 DEBUG_REQ(D_ERROR, req, "Invalid replay without recovery");
 897                 class_fail_export(req->rq_export);
 898                 rc = -ENODEV;
 899         } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0) {
 900                 DEBUG_REQ(D_ERROR, req,
 901                           "Invalid req with transno %llu without recovery",
 902                           lustre_msg_get_transno(req->rq_reqmsg));
 903                 class_fail_export(req->rq_export);
 904                 rc = -ENODEV;
 905         }
 906
 907         if (unlikely(rc < 0)) {
 908                 req->rq_status = rc;
 909                 ptlrpc_error(req);
 910         }
 911         return rc;
 912 }
 913
 914 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
 915 {
 916         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
 917         __s32 next;
 918
 919         if (array->paa_count == 0) {
 920                 del_timer(&svcpt->scp_at_timer);
 921                 return;
 922         }
 923
 924         /* Set timer for closest deadline */
 925         next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
 926                        at_early_margin);
 927         if (next <= 0) {
 928                 ptlrpc_at_timer((unsigned long)svcpt);
 929         } else {
 930                 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
 931                 CDEBUG(D_INFO, "armed %s at %+ds\n",
 932                        svcpt->scp_service->srv_name, next);
 933         }
 934 }
 935
 936 /* Add rpc to early reply check list */
 937 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 938 {
 939         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
 940         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
 941         struct ptlrpc_request *rq = NULL;
 942         __u32 index;
 943
 944         if (AT_OFF)
 945                 return 0;
 946
 947         if (req->rq_no_reply)
 948                 return 0;
 949
 950         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 951                 return -ENOSYS;
 952
 953         spin_lock(&svcpt->scp_at_lock);
 954         LASSERT(list_empty(&req->rq_timed_list));
 955
 956         div_u64_rem(req->rq_deadline, array->paa_size, &index);
 957         if (array->paa_reqs_count[index] > 0) {
 958                 /* latest rpcs will have the latest deadlines in the list,
 959                  * so search backward.
 960                  */
 961                 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
 962                                             rq_timed_list) {
 963                         if (req->rq_deadline >= rq->rq_deadline) {
 964                                 list_add(&req->rq_timed_list,
 965                                          &rq->rq_timed_list);
 966                                 break;
 967                         }
 968                 }
 969         }
 970
 971         /* Add the request at the head of the list */
 972         if (list_empty(&req->rq_timed_list))
 973                 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
 974
 975         spin_lock(&req->rq_lock);
 976         req->rq_at_linked = 1;
 977         spin_unlock(&req->rq_lock);
 978         req->rq_at_index = index;
 979         array->paa_reqs_count[index]++;
 980         array->paa_count++;
 981         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
 982                 array->paa_deadline = req->rq_deadline;
 983                 ptlrpc_at_set_timer(svcpt);
 984         }
 985         spin_unlock(&svcpt->scp_at_lock);
 986
 987         return 0;
 988 }
 989
 990 static void
 991 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
 992 {
 993         struct ptlrpc_at_array *array;
 994
 995         array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
 996
 997         /* NB: must call with hold svcpt::scp_at_lock */
 998         LASSERT(!list_empty(&req->rq_timed_list));
 999         list_del_init(&req->rq_timed_list);
1000
1001         spin_lock(&req->rq_lock);
1002         req->rq_at_linked = 0;
1003         spin_unlock(&req->rq_lock);
1004
1005         array->paa_reqs_count[req->rq_at_index]--;
1006         array->paa_count--;
1007 }
1008
1009 /*
1010  * Attempt to extend the request deadline by sending an early reply to the
1011  * client.
1012  */
1013 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1014 {
1015         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1016         struct ptlrpc_request *reqcopy;
1017         struct lustre_msg *reqmsg;
1018         long olddl = req->rq_deadline - ktime_get_real_seconds();
1019         time64_t newdl;
1020         int rc;
1021
1022         /* deadline is when the client expects us to reply, margin is the
1023          * difference between clients' and servers' expectations
1024          */
1025         DEBUG_REQ(D_ADAPTTO, req,
1026                   "%ssending early reply (deadline %+lds, margin %+lds) for %d+%d",
1027                   AT_OFF ? "AT off - not " : "",
1028                   olddl, olddl - at_get(&svcpt->scp_at_estimate),
1029                   at_get(&svcpt->scp_at_estimate), at_extra);
1030
1031         if (AT_OFF)
1032                 return 0;
1033
1034         if (olddl < 0) {
1035                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), not sending early reply. Consider increasing at_early_margin (%d)?",
1036                           olddl, at_early_margin);
1037
1038                 /* Return an error so we're not re-added to the timed list. */
1039                 return -ETIMEDOUT;
1040         }
1041
1042         if (!(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
1043                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, but no AT support");
1044                 return -ENOSYS;
1045         }
1046
1047         /*
1048          * We want to extend the request deadline by at_extra seconds,
1049          * so we set our service estimate to reflect how much time has
1050          * passed since this request arrived plus an additional
1051          * at_extra seconds. The client will calculate the new deadline
1052          * based on this service estimate (plus some additional time to
1053          * account for network latency). See ptlrpc_at_recv_early_reply
1054          */
1055         at_measured(&svcpt->scp_at_estimate, at_extra +
1056                     ktime_get_real_seconds() - req->rq_arrival_time.tv_sec);
1057         newdl = req->rq_arrival_time.tv_sec + at_get(&svcpt->scp_at_estimate);
1058
1059         /* Check to see if we've actually increased the deadline -
1060          * we may be past adaptive_max
1061          */
1062         if (req->rq_deadline >= newdl) {
1063                 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%ld/%lld), not sending early reply\n",
1064                           olddl, newdl - ktime_get_real_seconds());
1065                 return -ETIMEDOUT;
1066         }
1067
1068         reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1069         if (!reqcopy)
1070                 return -ENOMEM;
1071         reqmsg = libcfs_kvzalloc(req->rq_reqlen, GFP_NOFS);
1072         if (!reqmsg) {
1073                 rc = -ENOMEM;
1074                 goto out_free;
1075         }
1076
1077         *reqcopy = *req;
1078         reqcopy->rq_reply_state = NULL;
1079         reqcopy->rq_rep_swab_mask = 0;
1080         reqcopy->rq_pack_bulk = 0;
1081         reqcopy->rq_pack_udesc = 0;
1082         reqcopy->rq_packed_final = 0;
1083         sptlrpc_svc_ctx_addref(reqcopy);
1084         /* We only need the reqmsg for the magic */
1085         reqcopy->rq_reqmsg = reqmsg;
1086         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1087
1088         LASSERT(atomic_read(&req->rq_refcount));
1089         /** if it is last refcount then early reply isn't needed */
1090         if (atomic_read(&req->rq_refcount) == 1) {
1091                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, abort sending early reply\n");
1092                 rc = -EINVAL;
1093                 goto out;
1094         }
1095
1096         /* Connection ref */
1097         reqcopy->rq_export = class_conn2export(
1098                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
1099         if (!reqcopy->rq_export) {
1100                 rc = -ENODEV;
1101                 goto out;
1102         }
1103
1104         /* RPC ref */
1105         class_export_rpc_inc(reqcopy->rq_export);
1106         if (reqcopy->rq_export->exp_obd &&
1107             reqcopy->rq_export->exp_obd->obd_fail) {
1108                 rc = -ENODEV;
1109                 goto out_put;
1110         }
1111
1112         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1113         if (rc)
1114                 goto out_put;
1115
1116         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1117
1118         if (!rc) {
1119                 /* Adjust our own deadline to what we told the client */
1120                 req->rq_deadline = newdl;
1121                 req->rq_early_count++; /* number sent, server side */
1122         } else {
1123                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1124         }
1125
1126         /* Free the (early) reply state from lustre_pack_reply.
1127          * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1128          */
1129         ptlrpc_req_drop_rs(reqcopy);
1130
1131 out_put:
1132         class_export_rpc_dec(reqcopy->rq_export);
1133         class_export_put(reqcopy->rq_export);
1134 out:
1135         sptlrpc_svc_ctx_decref(reqcopy);
1136         kvfree(reqmsg);
1137 out_free:
1138         ptlrpc_request_cache_free(reqcopy);
1139         return rc;
1140 }
1141
1142 /* Send early replies to everybody expiring within at_early_margin
1143  * asking for at_extra time
1144  */
1145 static void ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1146 {
1147         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1148         struct ptlrpc_request *rq, *n;
1149         struct list_head work_list;
1150         __u32 index, count;
1151         time64_t deadline;
1152         time64_t now = ktime_get_real_seconds();
1153         long delay;
1154         int first, counter = 0;
1155
1156         spin_lock(&svcpt->scp_at_lock);
1157         if (svcpt->scp_at_check == 0) {
1158                 spin_unlock(&svcpt->scp_at_lock);
1159                 return;
1160         }
1161         delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1162         svcpt->scp_at_check = 0;
1163
1164         if (array->paa_count == 0) {
1165                 spin_unlock(&svcpt->scp_at_lock);
1166                 return;
1167         }
1168
1169         /* The timer went off, but maybe the nearest rpc already completed. */
1170         first = array->paa_deadline - now;
1171         if (first > at_early_margin) {
1172                 /* We've still got plenty of time.  Reset the timer. */
1173                 ptlrpc_at_set_timer(svcpt);
1174                 spin_unlock(&svcpt->scp_at_lock);
1175                 return;
1176         }
1177
1178         /* We're close to a timeout, and we don't know how much longer the
1179          * server will take. Send early replies to everyone expiring soon.
1180          */
1181         INIT_LIST_HEAD(&work_list);
1182         deadline = -1;
1183         div_u64_rem(array->paa_deadline, array->paa_size, &index);
1184         count = array->paa_count;
1185         while (count > 0) {
1186                 count -= array->paa_reqs_count[index];
1187                 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
1188                                          rq_timed_list) {
1189                         if (rq->rq_deadline > now + at_early_margin) {
1190                                 /* update the earliest deadline */
1191                                 if (deadline == -1 ||
1192                                     rq->rq_deadline < deadline)
1193                                         deadline = rq->rq_deadline;
1194                                 break;
1195                         }
1196
1197                         ptlrpc_at_remove_timed(rq);
1198                         /**
1199                          * ptlrpc_server_drop_request() may drop
1200                          * refcount to 0 already. Let's check this and
1201                          * don't add entry to work_list
1202                          */
1203                         if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1204                                 list_add(&rq->rq_timed_list, &work_list);
1205                         counter++;
1206                 }
1207
1208                 if (++index >= array->paa_size)
1209                         index = 0;
1210         }
1211         array->paa_deadline = deadline;
1212         /* we have a new earliest deadline, restart the timer */
1213         ptlrpc_at_set_timer(svcpt);
1214
1215         spin_unlock(&svcpt->scp_at_lock);
1216
1217         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early replies\n",
1218                first, at_extra, counter);
1219         if (first < 0) {
1220                 /* We're already past request deadlines before we even get a
1221                  * chance to send early replies
1222                  */
1223                 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1224                               svcpt->scp_service->srv_name);
1225                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1226                       counter, svcpt->scp_nreqs_incoming,
1227                       svcpt->scp_nreqs_active,
1228                       at_get(&svcpt->scp_at_estimate), delay);
1229         }
1230
1231         /* we took additional refcount so entries can't be deleted from list, no
1232          * locking is needed
1233          */
1234         while (!list_empty(&work_list)) {
1235                 rq = list_entry(work_list.next, struct ptlrpc_request,
1236                                 rq_timed_list);
1237                 list_del_init(&rq->rq_timed_list);
1238
1239                 if (ptlrpc_at_send_early_reply(rq) == 0)
1240                         ptlrpc_at_add_timed(rq);
1241
1242                 ptlrpc_server_drop_request(rq);
1243         }
1244 }
1245
1246 /**
1247  * Put the request to the export list if the request may become
1248  * a high priority one.
1249  */
1250 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1251                                     struct ptlrpc_request *req)
1252 {
1253         int rc = 0;
1254
1255         if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1256                 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1257                 if (rc < 0)
1258                         return rc;
1259                 LASSERT(rc == 0);
1260         }
1261         if (req->rq_export && req->rq_ops) {
1262                 /* Perform request specific check. We should do this check
1263                  * before the request is added into exp_hp_rpcs list otherwise
1264                  * it may hit swab race at LU-1044.
1265                  */
1266                 if (req->rq_ops->hpreq_check) {
1267                         rc = req->rq_ops->hpreq_check(req);
1268                         if (rc == -ESTALE) {
1269                                 req->rq_status = rc;
1270                                 ptlrpc_error(req);
1271                         }
1272                         /** can only return error,
1273                          * 0 for normal request,
1274                          *  or 1 for high priority request
1275                          */
1276                         LASSERT(rc <= 1);
1277                 }
1278
1279                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1280                 list_add(&req->rq_exp_list, &req->rq_export->exp_hp_rpcs);
1281                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1282         }
1283
1284         ptlrpc_nrs_req_initialize(svcpt, req, rc);
1285
1286         return rc;
1287 }
1288
1289 /** Remove the request from the export list. */
1290 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1291 {
1292         if (req->rq_export && req->rq_ops) {
1293                 /* refresh lock timeout again so that client has more
1294                  * room to send lock cancel RPC.
1295                  */
1296                 if (req->rq_ops->hpreq_fini)
1297                         req->rq_ops->hpreq_fini(req);
1298
1299                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1300                 list_del_init(&req->rq_exp_list);
1301                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1302         }
1303 }
1304
1305 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1306                                      struct ptlrpc_request *req)
1307 {
1308         int     rc;
1309
1310         rc = ptlrpc_server_hpreq_init(svcpt, req);
1311         if (rc < 0)
1312                 return rc;
1313
1314         ptlrpc_nrs_req_add(svcpt, req, !!rc);
1315
1316         return 0;
1317 }
1318
1319 /**
1320  * Allow to handle high priority request
1321  * User can call it w/o any lock but need to hold
1322  * ptlrpc_service_part::scp_req_lock to get reliable result
1323  */
1324 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1325                                      bool force)
1326 {
1327         int running = svcpt->scp_nthrs_running;
1328
1329         if (!nrs_svcpt_has_hp(svcpt))
1330                 return false;
1331
1332         if (force)
1333                 return true;
1334
1335         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1336                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1337                 /* leave just 1 thread for normal RPCs */
1338                 running = PTLRPC_NTHRS_INIT;
1339                 if (svcpt->scp_service->srv_ops.so_hpreq_handler)
1340                         running += 1;
1341         }
1342
1343         if (svcpt->scp_nreqs_active >= running - 1)
1344                 return false;
1345
1346         if (svcpt->scp_nhreqs_active == 0)
1347                 return true;
1348
1349         return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1350                svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1351 }
1352
1353 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1354                                        bool force)
1355 {
1356         return ptlrpc_server_allow_high(svcpt, force) &&
1357                ptlrpc_nrs_req_pending_nolock(svcpt, true);
1358 }
1359
1360 /**
1361  * Only allow normal priority requests on a service that has a high-priority
1362  * queue if forced (i.e. cleanup), if there are other high priority requests
1363  * already being processed (i.e. those threads can service more high-priority
1364  * requests), or if there are enough idle threads that a later thread can do
1365  * a high priority request.
1366  * User can call it w/o any lock but need to hold
1367  * ptlrpc_service_part::scp_req_lock to get reliable result
1368  */
1369 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1370                                        bool force)
1371 {
1372         int running = svcpt->scp_nthrs_running;
1373
1374         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1375                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1376                 /* leave just 1 thread for normal RPCs */
1377                 running = PTLRPC_NTHRS_INIT;
1378                 if (svcpt->scp_service->srv_ops.so_hpreq_handler)
1379                         running += 1;
1380         }
1381
1382         if (force ||
1383             svcpt->scp_nreqs_active < running - 2)
1384                 return true;
1385
1386         if (svcpt->scp_nreqs_active >= running - 1)
1387                 return false;
1388
1389         return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1390 }
1391
1392 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1393                                          bool force)
1394 {
1395         return ptlrpc_server_allow_normal(svcpt, force) &&
1396                ptlrpc_nrs_req_pending_nolock(svcpt, false);
1397 }
1398
1399 /**
1400  * Returns true if there are requests available in incoming
1401  * request queue for processing and it is allowed to fetch them.
1402  * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1403  * to get reliable result
1404  * \see ptlrpc_server_allow_normal
1405  * \see ptlrpc_server_allow high
1406  */
1407 static inline bool
1408 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1409 {
1410         return ptlrpc_server_high_pending(svcpt, force) ||
1411                ptlrpc_server_normal_pending(svcpt, force);
1412 }
1413
1414 /**
1415  * Fetch a request for processing from queue of unprocessed requests.
1416  * Favors high-priority requests.
1417  * Returns a pointer to fetched request.
1418  */
1419 static struct ptlrpc_request *
1420 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1421 {
1422         struct ptlrpc_request *req = NULL;
1423
1424         spin_lock(&svcpt->scp_req_lock);
1425
1426         if (ptlrpc_server_high_pending(svcpt, force)) {
1427                 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1428                 if (req) {
1429                         svcpt->scp_hreq_count++;
1430                         goto got_request;
1431                 }
1432         }
1433
1434         if (ptlrpc_server_normal_pending(svcpt, force)) {
1435                 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1436                 if (req) {
1437                         svcpt->scp_hreq_count = 0;
1438                         goto got_request;
1439                 }
1440         }
1441
1442         spin_unlock(&svcpt->scp_req_lock);
1443         return NULL;
1444
1445 got_request:
1446         svcpt->scp_nreqs_active++;
1447         if (req->rq_hp)
1448                 svcpt->scp_nhreqs_active++;
1449
1450         spin_unlock(&svcpt->scp_req_lock);
1451
1452         if (likely(req->rq_export))
1453                 class_export_rpc_inc(req->rq_export);
1454
1455         return req;
1456 }
1457
1458 /**
1459  * Handle freshly incoming reqs, add to timed early reply list,
1460  * pass on to regular request queue.
1461  * All incoming requests pass through here before getting into
1462  * ptlrpc_server_handle_req later on.
1463  */
1464 static int
1465 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1466                             struct ptlrpc_thread *thread)
1467 {
1468         struct ptlrpc_service *svc = svcpt->scp_service;
1469         struct ptlrpc_request *req;
1470         __u32 deadline;
1471         int rc;
1472
1473         spin_lock(&svcpt->scp_lock);
1474         if (list_empty(&svcpt->scp_req_incoming)) {
1475                 spin_unlock(&svcpt->scp_lock);
1476                 return 0;
1477         }
1478
1479         req = list_entry(svcpt->scp_req_incoming.next,
1480                          struct ptlrpc_request, rq_list);
1481         list_del_init(&req->rq_list);
1482         svcpt->scp_nreqs_incoming--;
1483         /* Consider this still a "queued" request as far as stats are
1484          * concerned
1485          */
1486         spin_unlock(&svcpt->scp_lock);
1487
1488         /* go through security check/transform */
1489         rc = sptlrpc_svc_unwrap_request(req);
1490         switch (rc) {
1491         case SECSVC_OK:
1492                 break;
1493         case SECSVC_COMPLETE:
1494                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1495                 goto err_req;
1496         case SECSVC_DROP:
1497                 goto err_req;
1498         default:
1499                 LBUG();
1500         }
1501
1502         /*
1503          * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1504          * redo it wouldn't be harmful.
1505          */
1506         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1507                 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1508                 if (rc != 0) {
1509                         CERROR("error unpacking request: ptl %d from %s x%llu\n",
1510                                svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1511                                req->rq_xid);
1512                         goto err_req;
1513                 }
1514         }
1515
1516         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1517         if (rc) {
1518                 CERROR("error unpacking ptlrpc body: ptl %d from %s x%llu\n",
1519                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1520                        req->rq_xid);
1521                 goto err_req;
1522         }
1523
1524         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1525             lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1526                 CERROR("drop incoming rpc opc %u, x%llu\n",
1527                        cfs_fail_val, req->rq_xid);
1528                 goto err_req;
1529         }
1530
1531         rc = -EINVAL;
1532         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1533                 CERROR("wrong packet type received (type=%u) from %s\n",
1534                        lustre_msg_get_type(req->rq_reqmsg),
1535                        libcfs_id2str(req->rq_peer));
1536                 goto err_req;
1537         }
1538
1539         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1540         case MDS_WRITEPAGE:
1541         case OST_WRITE:
1542                 req->rq_bulk_write = 1;
1543                 break;
1544         case MDS_READPAGE:
1545         case OST_READ:
1546         case MGS_CONFIG_READ:
1547                 req->rq_bulk_read = 1;
1548                 break;
1549         }
1550
1551         CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1552
1553         req->rq_export = class_conn2export(
1554                 lustre_msg_get_handle(req->rq_reqmsg));
1555         if (req->rq_export) {
1556                 rc = ptlrpc_check_req(req);
1557                 if (rc == 0) {
1558                         rc = sptlrpc_target_export_check(req->rq_export, req);
1559                         if (rc)
1560                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with illegal security flavor,");
1561                 }
1562
1563                 if (rc)
1564                         goto err_req;
1565         }
1566
1567         /* req_in handling should/must be fast */
1568         if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1569                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
1570                           (s64)(ktime_get_real_seconds() -
1571                                 req->rq_arrival_time.tv_sec));
1572
1573         /* Set rpc server deadline and add it to the timed list */
1574         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1575                     MSGHDR_AT_SUPPORT) ?
1576                    /* The max time the client expects us to take */
1577                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1578         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1579         if (unlikely(deadline == 0)) {
1580                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1581                 goto err_req;
1582         }
1583
1584         req->rq_svc_thread = thread;
1585         if (thread) {
1586                 /* initialize request session, it is needed for request
1587                  * processing by target
1588                  */
1589                 rc = lu_context_init(&req->rq_session,
1590                                      LCT_SERVER_SESSION | LCT_NOREF);
1591                 if (rc) {
1592                         CERROR("%s: failure to initialize session: rc = %d\n",
1593                                thread->t_name, rc);
1594                         goto err_req;
1595                 }
1596                 req->rq_session.lc_thread = thread;
1597                 lu_context_enter(&req->rq_session);
1598                 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1599         }
1600
1601         ptlrpc_at_add_timed(req);
1602
1603         /* Move it over to the request processing queue */
1604         rc = ptlrpc_server_request_add(svcpt, req);
1605         if (rc)
1606                 goto err_req;
1607
1608         wake_up(&svcpt->scp_waitq);
1609         return 1;
1610
1611 err_req:
1612         ptlrpc_server_finish_request(svcpt, req);
1613
1614         return 1;
1615 }
1616
1617 /**
1618  * Main incoming request handling logic.
1619  * Calls handler function from service to do actual processing.
1620  */
1621 static int
1622 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1623                              struct ptlrpc_thread *thread)
1624 {
1625         struct ptlrpc_service *svc = svcpt->scp_service;
1626         struct ptlrpc_request *request;
1627         struct timespec64 work_start;
1628         struct timespec64 work_end;
1629         struct timespec64 timediff;
1630         struct timespec64 arrived;
1631         unsigned long timediff_usecs;
1632         unsigned long arrived_usecs;
1633         int fail_opc = 0;
1634
1635         request = ptlrpc_server_request_get(svcpt, false);
1636         if (!request)
1637                 return 0;
1638
1639         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1640                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1641         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1642                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1643
1644         if (unlikely(fail_opc)) {
1645                 if (request->rq_export && request->rq_ops)
1646                         OBD_FAIL_TIMEOUT(fail_opc, 4);
1647         }
1648
1649         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1650
1651         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1652                 libcfs_debug_dumplog();
1653
1654         ktime_get_real_ts64(&work_start);
1655         timediff = timespec64_sub(work_start, request->rq_arrival_time);
1656         timediff_usecs = timediff.tv_sec * USEC_PER_SEC +
1657                          timediff.tv_nsec / NSEC_PER_USEC;
1658         if (likely(svc->srv_stats)) {
1659                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1660                                     timediff_usecs);
1661                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1662                                     svcpt->scp_nreqs_incoming);
1663                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1664                                     svcpt->scp_nreqs_active);
1665                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1666                                     at_get(&svcpt->scp_at_estimate));
1667         }
1668
1669         if (likely(request->rq_export)) {
1670                 if (unlikely(ptlrpc_check_req(request)))
1671                         goto put_conn;
1672         }
1673
1674         /* Discard requests queued for longer than the deadline.
1675          * The deadline is increased if we send an early reply.
1676          */
1677         if (ktime_get_real_seconds() > request->rq_deadline) {
1678                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
1679                           libcfs_id2str(request->rq_peer),
1680                           request->rq_deadline -
1681                           request->rq_arrival_time.tv_sec,
1682                           ktime_get_real_seconds() - 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                 if (hrp->hrp_nthrs == 0)
2540                         hrp->hrp_nthrs = 1;
2541
2542                 hrp->hrp_thrs =
2543                         kzalloc_node(hrp->hrp_nthrs * sizeof(*hrt), GFP_NOFS,
2544                                      cfs_cpt_spread_node(ptlrpc_hr.hr_cpt_table,
2545                                                          i));
2546                 if (!hrp->hrp_thrs) {
2547                         rc = -ENOMEM;
2548                         goto out;
2549                 }
2550
2551                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2552                         hrt = &hrp->hrp_thrs[j];
2553
2554                         hrt->hrt_id = j;
2555                         hrt->hrt_partition = hrp;
2556                         init_waitqueue_head(&hrt->hrt_waitq);
2557                         spin_lock_init(&hrt->hrt_lock);
2558                         INIT_LIST_HEAD(&hrt->hrt_queue);
2559                 }
2560         }
2561
2562         rc = ptlrpc_start_hr_threads();
2563 out:
2564         if (rc != 0)
2565                 ptlrpc_hr_fini();
2566         return rc;
2567 }
2568
2569 void ptlrpc_hr_fini(void)
2570 {
2571         struct ptlrpc_hr_partition *hrp;
2572         int i;
2573
2574         if (!ptlrpc_hr.hr_partitions)
2575                 return;
2576
2577         ptlrpc_stop_hr_threads();
2578
2579         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2580                 kfree(hrp->hrp_thrs);
2581         }
2582
2583         cfs_percpt_free(ptlrpc_hr.hr_partitions);
2584         ptlrpc_hr.hr_partitions = NULL;
2585 }
2586
2587 /**
2588  * Wait until all already scheduled replies are processed.
2589  */
2590 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2591 {
2592         while (1) {
2593                 int rc;
2594                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2595                                                      NULL, NULL);
2596
2597                 rc = l_wait_event(svcpt->scp_waitq,
2598                                   atomic_read(&svcpt->scp_nreps_difficult) == 0,
2599                                   &lwi);
2600                 if (rc == 0)
2601                         break;
2602                 CWARN("Unexpectedly long timeout %s %p\n",
2603                       svcpt->scp_service->srv_name, svcpt->scp_service);
2604         }
2605 }
2606
2607 static void
2608 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2609 {
2610         struct ptlrpc_service_part *svcpt;
2611         int i;
2612
2613         /* early disarm AT timer... */
2614         ptlrpc_service_for_each_part(svcpt, i, svc) {
2615                 if (svcpt->scp_service)
2616                         del_timer(&svcpt->scp_at_timer);
2617         }
2618 }
2619
2620 static void
2621 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2622 {
2623         struct ptlrpc_service_part *svcpt;
2624         struct ptlrpc_request_buffer_desc *rqbd;
2625         struct l_wait_info lwi;
2626         int rc;
2627         int i;
2628
2629         /* All history will be culled when the next request buffer is
2630          * freed in ptlrpc_service_purge_all()
2631          */
2632         svc->srv_hist_nrqbds_cpt_max = 0;
2633
2634         rc = LNetClearLazyPortal(svc->srv_req_portal);
2635         LASSERT(rc == 0);
2636
2637         ptlrpc_service_for_each_part(svcpt, i, svc) {
2638                 if (!svcpt->scp_service)
2639                         break;
2640
2641                 /* Unlink all the request buffers.  This forces a 'final'
2642                  * event with its 'unlink' flag set for each posted rqbd
2643                  */
2644                 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2645                                     rqbd_list) {
2646                         rc = LNetMDUnlink(rqbd->rqbd_md_h);
2647                         LASSERT(rc == 0 || rc == -ENOENT);
2648                 }
2649         }
2650
2651         ptlrpc_service_for_each_part(svcpt, i, svc) {
2652                 if (!svcpt->scp_service)
2653                         break;
2654
2655                 /* Wait for the network to release any buffers
2656                  * it's currently filling
2657                  */
2658                 spin_lock(&svcpt->scp_lock);
2659                 while (svcpt->scp_nrqbds_posted != 0) {
2660                         spin_unlock(&svcpt->scp_lock);
2661                         /* Network access will complete in finite time but
2662                          * the HUGE timeout lets us CWARN for visibility
2663                          * of sluggish LNDs
2664                          */
2665                         lwi = LWI_TIMEOUT_INTERVAL(
2666                                         cfs_time_seconds(LONG_UNLINK),
2667                                         cfs_time_seconds(1), NULL, NULL);
2668                         rc = l_wait_event(svcpt->scp_waitq,
2669                                           svcpt->scp_nrqbds_posted == 0, &lwi);
2670                         if (rc == -ETIMEDOUT) {
2671                                 CWARN("Service %s waiting for request buffers\n",
2672                                       svcpt->scp_service->srv_name);
2673                         }
2674                         spin_lock(&svcpt->scp_lock);
2675                 }
2676                 spin_unlock(&svcpt->scp_lock);
2677         }
2678 }
2679
2680 static void
2681 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2682 {
2683         struct ptlrpc_service_part *svcpt;
2684         struct ptlrpc_request_buffer_desc *rqbd;
2685         struct ptlrpc_request *req;
2686         struct ptlrpc_reply_state *rs;
2687         int i;
2688
2689         ptlrpc_service_for_each_part(svcpt, i, svc) {
2690                 if (!svcpt->scp_service)
2691                         break;
2692
2693                 spin_lock(&svcpt->scp_rep_lock);
2694                 while (!list_empty(&svcpt->scp_rep_active)) {
2695                         rs = list_entry(svcpt->scp_rep_active.next,
2696                                         struct ptlrpc_reply_state, rs_list);
2697                         spin_lock(&rs->rs_lock);
2698                         ptlrpc_schedule_difficult_reply(rs);
2699                         spin_unlock(&rs->rs_lock);
2700                 }
2701                 spin_unlock(&svcpt->scp_rep_lock);
2702
2703                 /* purge the request queue.  NB No new replies (rqbds
2704                  * all unlinked) and no service threads, so I'm the only
2705                  * thread noodling the request queue now
2706                  */
2707                 while (!list_empty(&svcpt->scp_req_incoming)) {
2708                         req = list_entry(svcpt->scp_req_incoming.next,
2709                                          struct ptlrpc_request, rq_list);
2710
2711                         list_del(&req->rq_list);
2712                         svcpt->scp_nreqs_incoming--;
2713                         ptlrpc_server_finish_request(svcpt, req);
2714                 }
2715
2716                 while (ptlrpc_server_request_pending(svcpt, true)) {
2717                         req = ptlrpc_server_request_get(svcpt, true);
2718                         ptlrpc_server_finish_active_request(svcpt, req);
2719                 }
2720
2721                 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
2722                 LASSERT(svcpt->scp_nreqs_incoming == 0);
2723                 LASSERT(svcpt->scp_nreqs_active == 0);
2724                 /* history should have been culled by
2725                  * ptlrpc_server_finish_request
2726                  */
2727                 LASSERT(svcpt->scp_hist_nrqbds == 0);
2728
2729                 /* Now free all the request buffers since nothing
2730                  * references them any more...
2731                  */
2732
2733                 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2734                         rqbd = list_entry(svcpt->scp_rqbd_idle.next,
2735                                           struct ptlrpc_request_buffer_desc,
2736                                           rqbd_list);
2737                         ptlrpc_free_rqbd(rqbd);
2738                 }
2739                 ptlrpc_wait_replies(svcpt);
2740
2741                 while (!list_empty(&svcpt->scp_rep_idle)) {
2742                         rs = list_entry(svcpt->scp_rep_idle.next,
2743                                         struct ptlrpc_reply_state,
2744                                         rs_list);
2745                         list_del(&rs->rs_list);
2746                         kvfree(rs);
2747                 }
2748         }
2749 }
2750
2751 static void
2752 ptlrpc_service_free(struct ptlrpc_service *svc)
2753 {
2754         struct ptlrpc_service_part *svcpt;
2755         struct ptlrpc_at_array *array;
2756         int i;
2757
2758         ptlrpc_service_for_each_part(svcpt, i, svc) {
2759                 if (!svcpt->scp_service)
2760                         break;
2761
2762                 /* In case somebody rearmed this in the meantime */
2763                 del_timer(&svcpt->scp_at_timer);
2764                 array = &svcpt->scp_at_array;
2765
2766                 kfree(array->paa_reqs_array);
2767                 array->paa_reqs_array = NULL;
2768                 kfree(array->paa_reqs_count);
2769                 array->paa_reqs_count = NULL;
2770         }
2771
2772         ptlrpc_service_for_each_part(svcpt, i, svc)
2773                 kfree(svcpt);
2774
2775         if (svc->srv_cpts)
2776                 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
2777
2778         kfree(svc);
2779 }
2780
2781 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2782 {
2783         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2784
2785         service->srv_is_stopping = 1;
2786
2787         mutex_lock(&ptlrpc_all_services_mutex);
2788         list_del_init(&service->srv_list);
2789         mutex_unlock(&ptlrpc_all_services_mutex);
2790
2791         ptlrpc_service_del_atimer(service);
2792         ptlrpc_stop_all_threads(service);
2793
2794         ptlrpc_service_unlink_rqbd(service);
2795         ptlrpc_service_purge_all(service);
2796         ptlrpc_service_nrs_cleanup(service);
2797
2798         ptlrpc_lprocfs_unregister_service(service);
2799         ptlrpc_sysfs_unregister_service(service);
2800
2801         ptlrpc_service_free(service);
2802
2803         return 0;
2804 }
2805 EXPORT_SYMBOL(ptlrpc_unregister_service);