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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #include "../include/obd_support.h"
  39 #include "../include/obd_class.h"
  40 #include "../include/lustre_net.h"
  41 #include "../include/lu_object.h"
  42 #include "../../include/linux/lnet/types.h"
  43 #include "ptlrpc_internal.h"
  44
  45 /* The following are visible and mutable through /sys/module/ptlrpc */
  46 int test_req_buffer_pressure;
  47 module_param(test_req_buffer_pressure, int, 0444);
  48 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
  49 module_param(at_min, int, 0644);
  50 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
  51 module_param(at_max, int, 0644);
  52 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
  53 module_param(at_history, int, 0644);
  54 MODULE_PARM_DESC(at_history,
  55                  "Adaptive timeouts remember the slowest event that took place within this period (sec)");
  56 module_param(at_early_margin, int, 0644);
  57 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
  58 module_param(at_extra, int, 0644);
  59 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
  60
  61 /* forward ref */
  62 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
  63 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
  64 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
  65
  66 /** Holds a list of all PTLRPC services */
  67 LIST_HEAD(ptlrpc_all_services);
  68 /** Used to protect the \e ptlrpc_all_services list */
  69 struct mutex ptlrpc_all_services_mutex;
  70
  71 static struct ptlrpc_request_buffer_desc *
  72 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
  73 {
  74         struct ptlrpc_service *svc = svcpt->scp_service;
  75         struct ptlrpc_request_buffer_desc *rqbd;
  76
  77         rqbd = kzalloc_node(sizeof(*rqbd), GFP_NOFS,
  78                             cfs_cpt_spread_node(svc->srv_cptable,
  79                                                 svcpt->scp_cpt));
  80         if (rqbd == NULL)
  81                 return NULL;
  82
  83         rqbd->rqbd_svcpt = svcpt;
  84         rqbd->rqbd_refcount = 0;
  85         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
  86         rqbd->rqbd_cbid.cbid_arg = rqbd;
  87         INIT_LIST_HEAD(&rqbd->rqbd_reqs);
  88         rqbd->rqbd_buffer = libcfs_kvzalloc_cpt(svc->srv_cptable,
  89                                                 svcpt->scp_cpt,
  90                                                 svc->srv_buf_size,
  91                                                 GFP_KERNEL);
  92         if (rqbd->rqbd_buffer == NULL) {
  93                 kfree(rqbd);
  94                 return NULL;
  95         }
  96
  97         spin_lock(&svcpt->scp_lock);
  98         list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
  99         svcpt->scp_nrqbds_total++;
 100         spin_unlock(&svcpt->scp_lock);
 101
 102         return rqbd;
 103 }
 104
 105 static void
 106 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 107 {
 108         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
 109
 110         LASSERT(rqbd->rqbd_refcount == 0);
 111         LASSERT(list_empty(&rqbd->rqbd_reqs));
 112
 113         spin_lock(&svcpt->scp_lock);
 114         list_del(&rqbd->rqbd_list);
 115         svcpt->scp_nrqbds_total--;
 116         spin_unlock(&svcpt->scp_lock);
 117
 118         kvfree(rqbd->rqbd_buffer);
 119         kfree(rqbd);
 120 }
 121
 122 static int
 123 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
 124 {
 125         struct ptlrpc_service *svc = svcpt->scp_service;
 126         struct ptlrpc_request_buffer_desc *rqbd;
 127         int rc = 0;
 128         int i;
 129
 130         if (svcpt->scp_rqbd_allocating)
 131                 goto try_post;
 132
 133         spin_lock(&svcpt->scp_lock);
 134         /* check again with lock */
 135         if (svcpt->scp_rqbd_allocating) {
 136                 /* NB: we might allow more than one thread in the future */
 137                 LASSERT(svcpt->scp_rqbd_allocating == 1);
 138                 spin_unlock(&svcpt->scp_lock);
 139                 goto try_post;
 140         }
 141
 142         svcpt->scp_rqbd_allocating++;
 143         spin_unlock(&svcpt->scp_lock);
 144
 145         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 146                 /* NB: another thread might have recycled enough rqbds, we
 147                  * need to make sure it wouldn't over-allocate, see LU-1212. */
 148                 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
 149                         break;
 150
 151                 rqbd = ptlrpc_alloc_rqbd(svcpt);
 152
 153                 if (rqbd == NULL) {
 154                         CERROR("%s: Can't allocate request buffer\n",
 155                                svc->srv_name);
 156                         rc = -ENOMEM;
 157                         break;
 158                 }
 159         }
 160
 161         spin_lock(&svcpt->scp_lock);
 162
 163         LASSERT(svcpt->scp_rqbd_allocating == 1);
 164         svcpt->scp_rqbd_allocating--;
 165
 166         spin_unlock(&svcpt->scp_lock);
 167
 168         CDEBUG(D_RPCTRACE,
 169                "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
 170                svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
 171                svcpt->scp_nrqbds_total, rc);
 172
 173  try_post:
 174         if (post && rc == 0)
 175                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
 176
 177         return rc;
 178 }
 179
 180 struct ptlrpc_hr_partition;
 181
 182 struct ptlrpc_hr_thread {
 183         int                             hrt_id;         /* thread ID */
 184         spinlock_t                      hrt_lock;
 185         wait_queue_head_t                       hrt_waitq;
 186         struct list_head                        hrt_queue;      /* RS queue */
 187         struct ptlrpc_hr_partition      *hrt_partition;
 188 };
 189
 190 struct ptlrpc_hr_partition {
 191         /* # of started threads */
 192         atomic_t                        hrp_nstarted;
 193         /* # of stopped threads */
 194         atomic_t                        hrp_nstopped;
 195         /* cpu partition id */
 196         int                             hrp_cpt;
 197         /* round-robin rotor for choosing thread */
 198         int                             hrp_rotor;
 199         /* total number of threads on this partition */
 200         int                             hrp_nthrs;
 201         /* threads table */
 202         struct ptlrpc_hr_thread         *hrp_thrs;
 203 };
 204
 205 #define HRT_RUNNING 0
 206 #define HRT_STOPPING 1
 207
 208 struct ptlrpc_hr_service {
 209         /* CPU partition table, it's just cfs_cpt_table for now */
 210         struct cfs_cpt_table            *hr_cpt_table;
 211         /** controller sleep waitq */
 212         wait_queue_head_t                       hr_waitq;
 213         unsigned int                    hr_stopping;
 214         /** roundrobin rotor for non-affinity service */
 215         unsigned int                    hr_rotor;
 216         /* partition data */
 217         struct ptlrpc_hr_partition      **hr_partitions;
 218 };
 219
 220 /** reply handling service. */
 221 static struct ptlrpc_hr_service         ptlrpc_hr;
 222
 223 /**
 224  * Choose an hr thread to dispatch requests to.
 225  */
 226 static struct ptlrpc_hr_thread *
 227 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
 228 {
 229         struct ptlrpc_hr_partition *hrp;
 230         unsigned int rotor;
 231
 232         if (svcpt->scp_cpt >= 0 &&
 233             svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
 234                 /* directly match partition */
 235                 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
 236
 237         } else {
 238                 rotor = ptlrpc_hr.hr_rotor++;
 239                 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
 240
 241                 hrp = ptlrpc_hr.hr_partitions[rotor];
 242         }
 243
 244         rotor = hrp->hrp_rotor++;
 245         return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
 246 }
 247
 248 /**
 249  * Put reply state into a queue for processing because we received
 250  * ACK from the client
 251  */
 252 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
 253 {
 254         struct ptlrpc_hr_thread *hrt;
 255
 256         LASSERT(list_empty(&rs->rs_list));
 257
 258         hrt = ptlrpc_hr_select(rs->rs_svcpt);
 259
 260         spin_lock(&hrt->hrt_lock);
 261         list_add_tail(&rs->rs_list, &hrt->hrt_queue);
 262         spin_unlock(&hrt->hrt_lock);
 263
 264         wake_up(&hrt->hrt_waitq);
 265 }
 266
 267 void
 268 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
 269 {
 270         assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
 271         assert_spin_locked(&rs->rs_lock);
 272         LASSERT(rs->rs_difficult);
 273         rs->rs_scheduled_ever = 1;  /* flag any notification attempt */
 274
 275         if (rs->rs_scheduled) {     /* being set up or already notified */
 276                 return;
 277         }
 278
 279         rs->rs_scheduled = 1;
 280         list_del_init(&rs->rs_list);
 281         ptlrpc_dispatch_difficult_reply(rs);
 282 }
 283 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
 284
 285 static int
 286 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
 287 {
 288         struct ptlrpc_request_buffer_desc *rqbd;
 289         int rc;
 290         int posted = 0;
 291
 292         for (;;) {
 293                 spin_lock(&svcpt->scp_lock);
 294
 295                 if (list_empty(&svcpt->scp_rqbd_idle)) {
 296                         spin_unlock(&svcpt->scp_lock);
 297                         return posted;
 298                 }
 299
 300                 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
 301                                       struct ptlrpc_request_buffer_desc,
 302                                       rqbd_list);
 303                 list_del(&rqbd->rqbd_list);
 304
 305                 /* assume we will post successfully */
 306                 svcpt->scp_nrqbds_posted++;
 307                 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
 308
 309                 spin_unlock(&svcpt->scp_lock);
 310
 311                 rc = ptlrpc_register_rqbd(rqbd);
 312                 if (rc != 0)
 313                         break;
 314
 315                 posted = 1;
 316         }
 317
 318         spin_lock(&svcpt->scp_lock);
 319
 320         svcpt->scp_nrqbds_posted--;
 321         list_del(&rqbd->rqbd_list);
 322         list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
 323
 324         /* Don't complain if no request buffers are posted right now; LNET
 325          * won't drop requests because we set the portal lazy! */
 326
 327         spin_unlock(&svcpt->scp_lock);
 328
 329         return -1;
 330 }
 331
 332 static void ptlrpc_at_timer(unsigned long castmeharder)
 333 {
 334         struct ptlrpc_service_part *svcpt;
 335
 336         svcpt = (struct ptlrpc_service_part *)castmeharder;
 337
 338         svcpt->scp_at_check = 1;
 339         svcpt->scp_at_checktime = cfs_time_current();
 340         wake_up(&svcpt->scp_waitq);
 341 }
 342
 343 static void
 344 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
 345                              struct ptlrpc_service_conf *conf)
 346 {
 347         struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
 348         unsigned init;
 349         unsigned total;
 350         unsigned nthrs;
 351         int weight;
 352
 353         /*
 354          * Common code for estimating & validating threads number.
 355          * CPT affinity service could have percpt thread-pool instead
 356          * of a global thread-pool, which means user might not always
 357          * get the threads number they give it in conf::tc_nthrs_user
 358          * even they did set. It's because we need to validate threads
 359          * number for each CPT to guarantee each pool will have enough
 360          * threads to keep the service healthy.
 361          */
 362         init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
 363         init = max_t(int, init, tc->tc_nthrs_init);
 364
 365         /* NB: please see comments in lustre_lnet.h for definition
 366          * details of these members */
 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                 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
 374                 nthrs = total / svc->srv_ncpts;
 375                 init = max(init, nthrs);
 376                 goto out;
 377         }
 378
 379         total = tc->tc_nthrs_max;
 380         if (tc->tc_nthrs_base == 0) {
 381                 /* don't care about base threads number per partition,
 382                  * this is most for non-affinity service */
 383                 nthrs = total / svc->srv_ncpts;
 384                 goto out;
 385         }
 386
 387         nthrs = tc->tc_nthrs_base;
 388         if (svc->srv_ncpts == 1) {
 389                 int i;
 390
 391                 /* NB: Increase the base number if it's single partition
 392                  * and total number of cores/HTs is larger or equal to 4.
 393                  * result will always < 2 * nthrs_base */
 394                 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
 395                 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
 396                             (tc->tc_nthrs_base >> i) != 0; i++)
 397                         nthrs += tc->tc_nthrs_base >> i;
 398         }
 399
 400         if (tc->tc_thr_factor != 0) {
 401                 int factor = tc->tc_thr_factor;
 402                 const int fade = 4;
 403
 404                 /*
 405                  * User wants to increase number of threads with for
 406                  * each CPU core/HT, most likely the factor is larger then
 407                  * one thread/core because service threads are supposed to
 408                  * be blocked by lock or wait for IO.
 409                  */
 410                 /*
 411                  * Amdahl's law says that adding processors wouldn't give
 412                  * a linear increasing of parallelism, so it's nonsense to
 413                  * have too many threads no matter how many cores/HTs
 414                  * there are.
 415                  */
 416                 /* weight is # of HTs */
 417                 if (cpumask_weight(topology_sibling_cpumask(0)) > 1) {
 418                         /* depress thread factor for hyper-thread */
 419                         factor = factor - (factor >> 1) + (factor >> 3);
 420                 }
 421
 422                 weight = cfs_cpt_weight(svc->srv_cptable, 0);
 423                 LASSERT(weight > 0);
 424
 425                 for (; factor > 0 && weight > 0; factor--, weight -= fade)
 426                         nthrs += min(weight, fade) * factor;
 427         }
 428
 429         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
 430                 nthrs = max(tc->tc_nthrs_base,
 431                             tc->tc_nthrs_max / svc->srv_ncpts);
 432         }
 433  out:
 434         nthrs = max(nthrs, tc->tc_nthrs_init);
 435         svc->srv_nthrs_cpt_limit = nthrs;
 436         svc->srv_nthrs_cpt_init = init;
 437
 438         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
 439                 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
 440                        svc->srv_name, nthrs * svc->srv_ncpts,
 441                        tc->tc_nthrs_max);
 442         }
 443 }
 444
 445 /**
 446  * Initialize percpt data for a service
 447  */
 448 static int
 449 ptlrpc_service_part_init(struct ptlrpc_service *svc,
 450                          struct ptlrpc_service_part *svcpt, int cpt)
 451 {
 452         struct ptlrpc_at_array  *array;
 453         int size;
 454         int index;
 455         int rc;
 456
 457         svcpt->scp_cpt = cpt;
 458         INIT_LIST_HEAD(&svcpt->scp_threads);
 459
 460         /* rqbd and incoming request queue */
 461         spin_lock_init(&svcpt->scp_lock);
 462         INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
 463         INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
 464         INIT_LIST_HEAD(&svcpt->scp_req_incoming);
 465         init_waitqueue_head(&svcpt->scp_waitq);
 466         /* history request & rqbd list */
 467         INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
 468         INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
 469
 470         /* active requests and hp requests */
 471         spin_lock_init(&svcpt->scp_req_lock);
 472
 473         /* reply states */
 474         spin_lock_init(&svcpt->scp_rep_lock);
 475         INIT_LIST_HEAD(&svcpt->scp_rep_active);
 476         INIT_LIST_HEAD(&svcpt->scp_rep_idle);
 477         init_waitqueue_head(&svcpt->scp_rep_waitq);
 478         atomic_set(&svcpt->scp_nreps_difficult, 0);
 479
 480         /* adaptive timeout */
 481         spin_lock_init(&svcpt->scp_at_lock);
 482         array = &svcpt->scp_at_array;
 483
 484         size = at_est2timeout(at_max);
 485         array->paa_size = size;
 486         array->paa_count = 0;
 487         array->paa_deadline = -1;
 488
 489         /* allocate memory for scp_at_array (ptlrpc_at_array) */
 490         array->paa_reqs_array =
 491                 kzalloc_node(sizeof(struct list_head) * size, GFP_NOFS,
 492                              cfs_cpt_spread_node(svc->srv_cptable, cpt));
 493         if (array->paa_reqs_array == NULL)
 494                 return -ENOMEM;
 495
 496         for (index = 0; index < size; index++)
 497                 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
 498
 499         array->paa_reqs_count =
 500                 kzalloc_node(sizeof(__u32) * size, GFP_NOFS,
 501                              cfs_cpt_spread_node(svc->srv_cptable, cpt));
 502         if (array->paa_reqs_count == NULL)
 503                 goto free_reqs_array;
 504
 505         setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
 506                     (unsigned long)svcpt);
 507
 508         /* At SOW, service time should be quick; 10s seems generous. If client
 509          * timeout is less than this, we'll be sending an early reply. */
 510         at_init(&svcpt->scp_at_estimate, 10, 0);
 511
 512         /* assign this before call ptlrpc_grow_req_bufs */
 513         svcpt->scp_service = svc;
 514         /* Now allocate the request buffers, but don't post them now */
 515         rc = ptlrpc_grow_req_bufs(svcpt, 0);
 516         /* We shouldn't be under memory pressure at startup, so
 517          * fail if we can't allocate all our buffers at this time. */
 518         if (rc != 0)
 519                 goto free_reqs_count;
 520
 521         return 0;
 522
 523 free_reqs_count:
 524         kfree(array->paa_reqs_count);
 525         array->paa_reqs_count = NULL;
 526 free_reqs_array:
 527         kfree(array->paa_reqs_array);
 528         array->paa_reqs_array = NULL;
 529
 530         return -ENOMEM;
 531 }
 532
 533 /**
 534  * Initialize service on a given portal.
 535  * This includes starting serving threads , allocating and posting rqbds and
 536  * so on.
 537  */
 538 struct ptlrpc_service *
 539 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
 540                         struct kset *parent,
 541                         struct dentry *debugfs_entry)
 542 {
 543         struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
 544         struct ptlrpc_service *service;
 545         struct ptlrpc_service_part *svcpt;
 546         struct cfs_cpt_table *cptable;
 547         __u32 *cpts = NULL;
 548         int ncpts;
 549         int cpt;
 550         int rc;
 551         int i;
 552
 553         LASSERT(conf->psc_buf.bc_nbufs > 0);
 554         LASSERT(conf->psc_buf.bc_buf_size >=
 555                 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
 556         LASSERT(conf->psc_thr.tc_ctx_tags != 0);
 557
 558         cptable = cconf->cc_cptable;
 559         if (cptable == NULL)
 560                 cptable = cfs_cpt_table;
 561
 562         if (!conf->psc_thr.tc_cpu_affinity) {
 563                 ncpts = 1;
 564         } else {
 565                 ncpts = cfs_cpt_number(cptable);
 566                 if (cconf->cc_pattern != NULL) {
 567                         struct cfs_expr_list *el;
 568
 569                         rc = cfs_expr_list_parse(cconf->cc_pattern,
 570                                                  strlen(cconf->cc_pattern),
 571                                                  0, ncpts - 1, &el);
 572                         if (rc != 0) {
 573                                 CERROR("%s: invalid CPT pattern string: %s",
 574                                        conf->psc_name, cconf->cc_pattern);
 575                                 return ERR_PTR(-EINVAL);
 576                         }
 577
 578                         rc = cfs_expr_list_values(el, ncpts, &cpts);
 579                         cfs_expr_list_free(el);
 580                         if (rc <= 0) {
 581                                 CERROR("%s: failed to parse CPT array %s: %d\n",
 582                                        conf->psc_name, cconf->cc_pattern, rc);
 583                                 kfree(cpts);
 584                                 return ERR_PTR(rc < 0 ? rc : -EINVAL);
 585                         }
 586                         ncpts = rc;
 587                 }
 588         }
 589
 590         service = kzalloc(offsetof(struct ptlrpc_service, srv_parts[ncpts]),
 591                           GFP_NOFS);
 592         if (!service) {
 593                 kfree(cpts);
 594                 return ERR_PTR(-ENOMEM);
 595         }
 596
 597         service->srv_cptable = cptable;
 598         service->srv_cpts = cpts;
 599         service->srv_ncpts = ncpts;
 600
 601         service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
 602         while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
 603                 service->srv_cpt_bits++;
 604
 605         /* public members */
 606         spin_lock_init(&service->srv_lock);
 607         service->srv_name = conf->psc_name;
 608         service->srv_watchdog_factor = conf->psc_watchdog_factor;
 609         INIT_LIST_HEAD(&service->srv_list); /* for safety of cleanup */
 610
 611         /* buffer configuration */
 612         service->srv_nbuf_per_group = test_req_buffer_pressure ?
 613                                           1 : conf->psc_buf.bc_nbufs;
 614         service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
 615                                           SPTLRPC_MAX_PAYLOAD;
 616         service->srv_buf_size = conf->psc_buf.bc_buf_size;
 617         service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
 618         service->srv_req_portal = conf->psc_buf.bc_req_portal;
 619
 620         /* Increase max reply size to next power of two */
 621         service->srv_max_reply_size = 1;
 622         while (service->srv_max_reply_size <
 623                conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
 624                 service->srv_max_reply_size <<= 1;
 625
 626         service->srv_thread_name = conf->psc_thr.tc_thr_name;
 627         service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
 628         service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
 629         service->srv_ops = conf->psc_ops;
 630
 631         for (i = 0; i < ncpts; i++) {
 632                 if (!conf->psc_thr.tc_cpu_affinity)
 633                         cpt = CFS_CPT_ANY;
 634                 else
 635                         cpt = cpts != NULL ? cpts[i] : i;
 636
 637                 svcpt = kzalloc_node(sizeof(*svcpt), GFP_NOFS,
 638                                      cfs_cpt_spread_node(cptable, cpt));
 639                 if (svcpt == NULL) {
 640                         rc = -ENOMEM;
 641                         goto failed;
 642                 }
 643
 644                 service->srv_parts[i] = svcpt;
 645                 rc = ptlrpc_service_part_init(service, svcpt, cpt);
 646                 if (rc != 0)
 647                         goto failed;
 648         }
 649
 650         ptlrpc_server_nthreads_check(service, conf);
 651
 652         rc = LNetSetLazyPortal(service->srv_req_portal);
 653         LASSERT(rc == 0);
 654
 655         mutex_lock(&ptlrpc_all_services_mutex);
 656         list_add(&service->srv_list, &ptlrpc_all_services);
 657         mutex_unlock(&ptlrpc_all_services_mutex);
 658
 659         if (parent) {
 660                 rc = ptlrpc_sysfs_register_service(parent, service);
 661                 if (rc)
 662                         goto failed;
 663         }
 664
 665         if (!IS_ERR_OR_NULL(debugfs_entry))
 666                 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
 667
 668         rc = ptlrpc_service_nrs_setup(service);
 669         if (rc != 0)
 670                 goto failed;
 671
 672         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 673                service->srv_name, service->srv_req_portal);
 674
 675         rc = ptlrpc_start_threads(service);
 676         if (rc != 0) {
 677                 CERROR("Failed to start threads for service %s: %d\n",
 678                        service->srv_name, rc);
 679                 goto failed;
 680         }
 681
 682         return service;
 683 failed:
 684         ptlrpc_unregister_service(service);
 685         return ERR_PTR(rc);
 686 }
 687 EXPORT_SYMBOL(ptlrpc_register_service);
 688
 689 /**
 690  * to actually free the request, must be called without holding svc_lock.
 691  * note it's caller's responsibility to unlink req->rq_list.
 692  */
 693 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
 694 {
 695         LASSERT(atomic_read(&req->rq_refcount) == 0);
 696         LASSERT(list_empty(&req->rq_timed_list));
 697
 698          /* DEBUG_REQ() assumes the reply state of a request with a valid
 699           * ref will not be destroyed until that reference is dropped. */
 700         ptlrpc_req_drop_rs(req);
 701
 702         sptlrpc_svc_ctx_decref(req);
 703
 704         if (req != &req->rq_rqbd->rqbd_req) {
 705                 /* NB request buffers use an embedded
 706                  * req if the incoming req unlinked the
 707                  * MD; this isn't one of them! */
 708                 ptlrpc_request_cache_free(req);
 709         }
 710 }
 711
 712 /**
 713  * drop a reference count of the request. if it reaches 0, we either
 714  * put it into history list, or free it immediately.
 715  */
 716 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
 717 {
 718         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 719         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
 720         struct ptlrpc_service *svc = svcpt->scp_service;
 721         int refcount;
 722         struct list_head *tmp;
 723         struct list_head *nxt;
 724
 725         if (!atomic_dec_and_test(&req->rq_refcount))
 726                 return;
 727
 728         if (req->rq_at_linked) {
 729                 spin_lock(&svcpt->scp_at_lock);
 730                 /* recheck with lock, in case it's unlinked by
 731                  * ptlrpc_at_check_timed() */
 732                 if (likely(req->rq_at_linked))
 733                         ptlrpc_at_remove_timed(req);
 734                 spin_unlock(&svcpt->scp_at_lock);
 735         }
 736
 737         LASSERT(list_empty(&req->rq_timed_list));
 738
 739         /* finalize request */
 740         if (req->rq_export) {
 741                 class_export_put(req->rq_export);
 742                 req->rq_export = NULL;
 743         }
 744
 745         spin_lock(&svcpt->scp_lock);
 746
 747         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 748
 749         refcount = --(rqbd->rqbd_refcount);
 750         if (refcount == 0) {
 751                 /* request buffer is now idle: add to history */
 752                 list_del(&rqbd->rqbd_list);
 753
 754                 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
 755                 svcpt->scp_hist_nrqbds++;
 756
 757                 /* cull some history?
 758                  * I expect only about 1 or 2 rqbds need to be recycled here */
 759                 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
 760                         rqbd = list_entry(svcpt->scp_hist_rqbds.next,
 761                                               struct ptlrpc_request_buffer_desc,
 762                                               rqbd_list);
 763
 764                         list_del(&rqbd->rqbd_list);
 765                         svcpt->scp_hist_nrqbds--;
 766
 767                         /* remove rqbd's reqs from svc's req history while
 768                          * I've got the service lock */
 769                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 770                                 req = list_entry(tmp, struct ptlrpc_request,
 771                                                      rq_list);
 772                                 /* Track the highest culled req seq */
 773                                 if (req->rq_history_seq >
 774                                     svcpt->scp_hist_seq_culled) {
 775                                         svcpt->scp_hist_seq_culled =
 776                                                 req->rq_history_seq;
 777                                 }
 778                                 list_del(&req->rq_history_list);
 779                         }
 780
 781                         spin_unlock(&svcpt->scp_lock);
 782
 783                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 784                                 req = list_entry(rqbd->rqbd_reqs.next,
 785                                                      struct ptlrpc_request,
 786                                                      rq_list);
 787                                 list_del(&req->rq_list);
 788                                 ptlrpc_server_free_request(req);
 789                         }
 790
 791                         spin_lock(&svcpt->scp_lock);
 792                         /*
 793                          * now all reqs including the embedded req has been
 794                          * disposed, schedule request buffer for re-use.
 795                          */
 796                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) ==
 797                                 0);
 798                         list_add_tail(&rqbd->rqbd_list,
 799                                           &svcpt->scp_rqbd_idle);
 800                 }
 801
 802                 spin_unlock(&svcpt->scp_lock);
 803         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 804                 /* If we are low on memory, we are not interested in history */
 805                 list_del(&req->rq_list);
 806                 list_del_init(&req->rq_history_list);
 807
 808                 /* Track the highest culled req seq */
 809                 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
 810                         svcpt->scp_hist_seq_culled = req->rq_history_seq;
 811
 812                 spin_unlock(&svcpt->scp_lock);
 813
 814                 ptlrpc_server_free_request(req);
 815         } else {
 816                 spin_unlock(&svcpt->scp_lock);
 817         }
 818 }
 819
 820 /**
 821  * to finish a request: stop sending more early replies, and release
 822  * the request.
 823  */
 824 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
 825                                          struct ptlrpc_request *req)
 826 {
 827         ptlrpc_server_hpreq_fini(req);
 828
 829         ptlrpc_server_drop_request(req);
 830 }
 831
 832 /**
 833  * to finish a active request: stop sending more early replies, and release
 834  * the request. should be called after we finished handling the request.
 835  */
 836 static void ptlrpc_server_finish_active_request(
 837                                         struct ptlrpc_service_part *svcpt,
 838                                         struct ptlrpc_request *req)
 839 {
 840         spin_lock(&svcpt->scp_req_lock);
 841         ptlrpc_nrs_req_stop_nolock(req);
 842         svcpt->scp_nreqs_active--;
 843         if (req->rq_hp)
 844                 svcpt->scp_nhreqs_active--;
 845         spin_unlock(&svcpt->scp_req_lock);
 846
 847         ptlrpc_nrs_req_finalize(req);
 848
 849         if (req->rq_export != NULL)
 850                 class_export_rpc_dec(req->rq_export);
 851
 852         ptlrpc_server_finish_request(svcpt, req);
 853 }
 854
 855 /**
 856  * Sanity check request \a req.
 857  * Return 0 if all is ok, error code otherwise.
 858  */
 859 static int ptlrpc_check_req(struct ptlrpc_request *req)
 860 {
 861         struct obd_device *obd = req->rq_export->exp_obd;
 862         int rc = 0;
 863
 864         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 865                      req->rq_export->exp_conn_cnt)) {
 866                 DEBUG_REQ(D_RPCTRACE, req,
 867                           "DROPPING req from old connection %d < %d",
 868                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 869                           req->rq_export->exp_conn_cnt);
 870                 return -EEXIST;
 871         }
 872         if (unlikely(obd == NULL || obd->obd_fail)) {
 873                 /*
 874                  * Failing over, don't handle any more reqs, send
 875                  * error response instead.
 876                  */
 877                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 878                        req, (obd != NULL) ? obd->obd_name : "unknown");
 879                 rc = -ENODEV;
 880         } else if (lustre_msg_get_flags(req->rq_reqmsg) &
 881                    (MSG_REPLAY | MSG_REQ_REPLAY_DONE)) {
 882                 DEBUG_REQ(D_ERROR, req, "Invalid replay without recovery");
 883                 class_fail_export(req->rq_export);
 884                 rc = -ENODEV;
 885         } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0) {
 886                 DEBUG_REQ(D_ERROR, req,
 887                           "Invalid req with transno %llu without recovery",
 888                           lustre_msg_get_transno(req->rq_reqmsg));
 889                 class_fail_export(req->rq_export);
 890                 rc = -ENODEV;
 891         }
 892
 893         if (unlikely(rc < 0)) {
 894                 req->rq_status = rc;
 895                 ptlrpc_error(req);
 896         }
 897         return rc;
 898 }
 899
 900 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
 901 {
 902         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
 903         __s32 next;
 904
 905         if (array->paa_count == 0) {
 906                 del_timer(&svcpt->scp_at_timer);
 907                 return;
 908         }
 909
 910         /* Set timer for closest deadline */
 911         next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
 912                        at_early_margin);
 913         if (next <= 0) {
 914                 ptlrpc_at_timer((unsigned long)svcpt);
 915         } else {
 916                 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
 917                 CDEBUG(D_INFO, "armed %s at %+ds\n",
 918                        svcpt->scp_service->srv_name, next);
 919         }
 920 }
 921
 922 /* Add rpc to early reply check list */
 923 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 924 {
 925         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
 926         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
 927         struct ptlrpc_request *rq = NULL;
 928         __u32 index;
 929
 930         if (AT_OFF)
 931                 return 0;
 932
 933         if (req->rq_no_reply)
 934                 return 0;
 935
 936         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 937                 return -ENOSYS;
 938
 939         spin_lock(&svcpt->scp_at_lock);
 940         LASSERT(list_empty(&req->rq_timed_list));
 941
 942         div_u64_rem(req->rq_deadline, array->paa_size, &index);
 943         if (array->paa_reqs_count[index] > 0) {
 944                 /* latest rpcs will have the latest deadlines in the list,
 945                  * so search backward. */
 946                 list_for_each_entry_reverse(rq,
 947                                                 &array->paa_reqs_array[index],
 948                                                 rq_timed_list) {
 949                         if (req->rq_deadline >= rq->rq_deadline) {
 950                                 list_add(&req->rq_timed_list,
 951                                              &rq->rq_timed_list);
 952                                 break;
 953                         }
 954                 }
 955         }
 956
 957         /* Add the request at the head of the list */
 958         if (list_empty(&req->rq_timed_list))
 959                 list_add(&req->rq_timed_list,
 960                              &array->paa_reqs_array[index]);
 961
 962         spin_lock(&req->rq_lock);
 963         req->rq_at_linked = 1;
 964         spin_unlock(&req->rq_lock);
 965         req->rq_at_index = index;
 966         array->paa_reqs_count[index]++;
 967         array->paa_count++;
 968         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
 969                 array->paa_deadline = req->rq_deadline;
 970                 ptlrpc_at_set_timer(svcpt);
 971         }
 972         spin_unlock(&svcpt->scp_at_lock);
 973
 974         return 0;
 975 }
 976
 977 static void
 978 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
 979 {
 980         struct ptlrpc_at_array *array;
 981
 982         array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
 983
 984         /* NB: must call with hold svcpt::scp_at_lock */
 985         LASSERT(!list_empty(&req->rq_timed_list));
 986         list_del_init(&req->rq_timed_list);
 987
 988         spin_lock(&req->rq_lock);
 989         req->rq_at_linked = 0;
 990         spin_unlock(&req->rq_lock);
 991
 992         array->paa_reqs_count[req->rq_at_index]--;
 993         array->paa_count--;
 994 }
 995
 996 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
 997 {
 998         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
 999         struct ptlrpc_request *reqcopy;
1000         struct lustre_msg *reqmsg;
1001         long olddl = req->rq_deadline - ktime_get_real_seconds();
1002         time64_t newdl;
1003         int rc;
1004
1005         /* deadline is when the client expects us to reply, margin is the
1006            difference between clients' and servers' expectations */
1007         DEBUG_REQ(D_ADAPTTO, req,
1008                   "%ssending early reply (deadline %+lds, margin %+lds) for %d+%d",
1009                   AT_OFF ? "AT off - not " : "",
1010                   olddl, olddl - at_get(&svcpt->scp_at_estimate),
1011                   at_get(&svcpt->scp_at_estimate), at_extra);
1012
1013         if (AT_OFF)
1014                 return 0;
1015
1016         if (olddl < 0) {
1017                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), not sending early reply. Consider increasing at_early_margin (%d)?",
1018                           olddl, at_early_margin);
1019
1020                 /* Return an error so we're not re-added to the timed list. */
1021                 return -ETIMEDOUT;
1022         }
1023
1024         if (!(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
1025                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, but no AT support");
1026                 return -ENOSYS;
1027         }
1028
1029         /* Fake our processing time into the future to ask the clients
1030          * for some extra amount of time */
1031         at_measured(&svcpt->scp_at_estimate, at_extra +
1032                     ktime_get_real_seconds() - req->rq_arrival_time.tv_sec);
1033
1034         /* Check to see if we've actually increased the deadline -
1035          * we may be past adaptive_max */
1036         if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1037             at_get(&svcpt->scp_at_estimate)) {
1038                 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%ld/%lld), not sending early reply\n",
1039                           olddl, req->rq_arrival_time.tv_sec +
1040                           at_get(&svcpt->scp_at_estimate) -
1041                           ktime_get_real_seconds());
1042                 return -ETIMEDOUT;
1043         }
1044         newdl = ktime_get_real_seconds() + at_get(&svcpt->scp_at_estimate);
1045
1046         reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1047         if (reqcopy == NULL)
1048                 return -ENOMEM;
1049         reqmsg = libcfs_kvzalloc(req->rq_reqlen, GFP_NOFS);
1050         if (!reqmsg) {
1051                 rc = -ENOMEM;
1052                 goto out_free;
1053         }
1054
1055         *reqcopy = *req;
1056         reqcopy->rq_reply_state = NULL;
1057         reqcopy->rq_rep_swab_mask = 0;
1058         reqcopy->rq_pack_bulk = 0;
1059         reqcopy->rq_pack_udesc = 0;
1060         reqcopy->rq_packed_final = 0;
1061         sptlrpc_svc_ctx_addref(reqcopy);
1062         /* We only need the reqmsg for the magic */
1063         reqcopy->rq_reqmsg = reqmsg;
1064         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1065
1066         LASSERT(atomic_read(&req->rq_refcount));
1067         /** if it is last refcount then early reply isn't needed */
1068         if (atomic_read(&req->rq_refcount) == 1) {
1069                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, abort sending early reply\n");
1070                 rc = -EINVAL;
1071                 goto out;
1072         }
1073
1074         /* Connection ref */
1075         reqcopy->rq_export = class_conn2export(
1076                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
1077         if (reqcopy->rq_export == NULL) {
1078                 rc = -ENODEV;
1079                 goto out;
1080         }
1081
1082         /* RPC ref */
1083         class_export_rpc_inc(reqcopy->rq_export);
1084         if (reqcopy->rq_export->exp_obd &&
1085             reqcopy->rq_export->exp_obd->obd_fail) {
1086                 rc = -ENODEV;
1087                 goto out_put;
1088         }
1089
1090         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1091         if (rc)
1092                 goto out_put;
1093
1094         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1095
1096         if (!rc) {
1097                 /* Adjust our own deadline to what we told the client */
1098                 req->rq_deadline = newdl;
1099                 req->rq_early_count++; /* number sent, server side */
1100         } else {
1101                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1102         }
1103
1104         /* Free the (early) reply state from lustre_pack_reply.
1105            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1106         ptlrpc_req_drop_rs(reqcopy);
1107
1108 out_put:
1109         class_export_rpc_dec(reqcopy->rq_export);
1110         class_export_put(reqcopy->rq_export);
1111 out:
1112         sptlrpc_svc_ctx_decref(reqcopy);
1113         kvfree(reqmsg);
1114 out_free:
1115         ptlrpc_request_cache_free(reqcopy);
1116         return rc;
1117 }
1118
1119 /* Send early replies to everybody expiring within at_early_margin
1120    asking for at_extra time */
1121 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1122 {
1123         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1124         struct ptlrpc_request *rq, *n;
1125         struct list_head work_list;
1126         __u32 index, count;
1127         time64_t deadline;
1128         time64_t now = ktime_get_real_seconds();
1129         long delay;
1130         int first, counter = 0;
1131
1132         spin_lock(&svcpt->scp_at_lock);
1133         if (svcpt->scp_at_check == 0) {
1134                 spin_unlock(&svcpt->scp_at_lock);
1135                 return 0;
1136         }
1137         delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1138         svcpt->scp_at_check = 0;
1139
1140         if (array->paa_count == 0) {
1141                 spin_unlock(&svcpt->scp_at_lock);
1142                 return 0;
1143         }
1144
1145         /* The timer went off, but maybe the nearest rpc already completed. */
1146         first = array->paa_deadline - now;
1147         if (first > at_early_margin) {
1148                 /* We've still got plenty of time.  Reset the timer. */
1149                 ptlrpc_at_set_timer(svcpt);
1150                 spin_unlock(&svcpt->scp_at_lock);
1151                 return 0;
1152         }
1153
1154         /* We're close to a timeout, and we don't know how much longer the
1155            server will take. Send early replies to everyone expiring soon. */
1156         INIT_LIST_HEAD(&work_list);
1157         deadline = -1;
1158         div_u64_rem(array->paa_deadline, array->paa_size, &index);
1159         count = array->paa_count;
1160         while (count > 0) {
1161                 count -= array->paa_reqs_count[index];
1162                 list_for_each_entry_safe(rq, n,
1163                                              &array->paa_reqs_array[index],
1164                                              rq_timed_list) {
1165                         if (rq->rq_deadline > now + at_early_margin) {
1166                                 /* update the earliest deadline */
1167                                 if (deadline == -1 ||
1168                                     rq->rq_deadline < deadline)
1169                                         deadline = rq->rq_deadline;
1170                                 break;
1171                         }
1172
1173                         ptlrpc_at_remove_timed(rq);
1174                         /**
1175                          * ptlrpc_server_drop_request() may drop
1176                          * refcount to 0 already. Let's check this and
1177                          * don't add entry to work_list
1178                          */
1179                         if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1180                                 list_add(&rq->rq_timed_list, &work_list);
1181                         counter++;
1182                 }
1183
1184                 if (++index >= array->paa_size)
1185                         index = 0;
1186         }
1187         array->paa_deadline = deadline;
1188         /* we have a new earliest deadline, restart the timer */
1189         ptlrpc_at_set_timer(svcpt);
1190
1191         spin_unlock(&svcpt->scp_at_lock);
1192
1193         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early replies\n",
1194                first, at_extra, counter);
1195         if (first < 0) {
1196                 /* We're already past request deadlines before we even get a
1197                    chance to send early replies */
1198                 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1199                               svcpt->scp_service->srv_name);
1200                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1201                       counter, svcpt->scp_nreqs_incoming,
1202                       svcpt->scp_nreqs_active,
1203                       at_get(&svcpt->scp_at_estimate), delay);
1204         }
1205
1206         /* we took additional refcount so entries can't be deleted from list, no
1207          * locking is needed */
1208         while (!list_empty(&work_list)) {
1209                 rq = list_entry(work_list.next, struct ptlrpc_request,
1210                                     rq_timed_list);
1211                 list_del_init(&rq->rq_timed_list);
1212
1213                 if (ptlrpc_at_send_early_reply(rq) == 0)
1214                         ptlrpc_at_add_timed(rq);
1215
1216                 ptlrpc_server_drop_request(rq);
1217         }
1218
1219         return 1; /* return "did_something" for liblustre */
1220 }
1221
1222 /**
1223  * Put the request to the export list if the request may become
1224  * a high priority one.
1225  */
1226 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1227                                     struct ptlrpc_request *req)
1228 {
1229         int rc = 0;
1230
1231         if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1232                 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1233                 if (rc < 0)
1234                         return rc;
1235                 LASSERT(rc == 0);
1236         }
1237         if (req->rq_export && req->rq_ops) {
1238                 /* Perform request specific check. We should do this check
1239                  * before the request is added into exp_hp_rpcs list otherwise
1240                  * it may hit swab race at LU-1044. */
1241                 if (req->rq_ops->hpreq_check) {
1242                         rc = req->rq_ops->hpreq_check(req);
1243                         if (rc == -ESTALE) {
1244                                 req->rq_status = rc;
1245                                 ptlrpc_error(req);
1246                         }
1247                         /** can only return error,
1248                          * 0 for normal request,
1249                          *  or 1 for high priority request
1250                          */
1251                         LASSERT(rc <= 1);
1252                 }
1253
1254                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1255                 list_add(&req->rq_exp_list,
1256                              &req->rq_export->exp_hp_rpcs);
1257                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1258         }
1259
1260         ptlrpc_nrs_req_initialize(svcpt, req, rc);
1261
1262         return rc;
1263 }
1264
1265 /** Remove the request from the export list. */
1266 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1267 {
1268         if (req->rq_export && req->rq_ops) {
1269                 /* refresh lock timeout again so that client has more
1270                  * room to send lock cancel RPC. */
1271                 if (req->rq_ops->hpreq_fini)
1272                         req->rq_ops->hpreq_fini(req);
1273
1274                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1275                 list_del_init(&req->rq_exp_list);
1276                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1277         }
1278 }
1279
1280 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1281                                      struct ptlrpc_request *req)
1282 {
1283         int     rc;
1284
1285         rc = ptlrpc_server_hpreq_init(svcpt, req);
1286         if (rc < 0)
1287                 return rc;
1288
1289         ptlrpc_nrs_req_add(svcpt, req, !!rc);
1290
1291         return 0;
1292 }
1293
1294 /**
1295  * Allow to handle high priority request
1296  * User can call it w/o any lock but need to hold
1297  * ptlrpc_service_part::scp_req_lock to get reliable result
1298  */
1299 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1300                                      bool force)
1301 {
1302         int running = svcpt->scp_nthrs_running;
1303
1304         if (!nrs_svcpt_has_hp(svcpt))
1305                 return false;
1306
1307         if (force)
1308                 return true;
1309
1310         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1311                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1312                 /* leave just 1 thread for normal RPCs */
1313                 running = PTLRPC_NTHRS_INIT;
1314                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1315                         running += 1;
1316         }
1317
1318         if (svcpt->scp_nreqs_active >= running - 1)
1319                 return false;
1320
1321         if (svcpt->scp_nhreqs_active == 0)
1322                 return true;
1323
1324         return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1325                svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1326 }
1327
1328 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1329                                        bool force)
1330 {
1331         return ptlrpc_server_allow_high(svcpt, force) &&
1332                ptlrpc_nrs_req_pending_nolock(svcpt, true);
1333 }
1334
1335 /**
1336  * Only allow normal priority requests on a service that has a high-priority
1337  * queue if forced (i.e. cleanup), if there are other high priority requests
1338  * already being processed (i.e. those threads can service more high-priority
1339  * requests), or if there are enough idle threads that a later thread can do
1340  * a high priority request.
1341  * User can call it w/o any lock but need to hold
1342  * ptlrpc_service_part::scp_req_lock to get reliable result
1343  */
1344 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1345                                        bool force)
1346 {
1347         int running = svcpt->scp_nthrs_running;
1348
1349         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1350                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1351                 /* leave just 1 thread for normal RPCs */
1352                 running = PTLRPC_NTHRS_INIT;
1353                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1354                         running += 1;
1355         }
1356
1357         if (force ||
1358             svcpt->scp_nreqs_active < running - 2)
1359                 return true;
1360
1361         if (svcpt->scp_nreqs_active >= running - 1)
1362                 return false;
1363
1364         return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1365 }
1366
1367 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1368                                          bool force)
1369 {
1370         return ptlrpc_server_allow_normal(svcpt, force) &&
1371                ptlrpc_nrs_req_pending_nolock(svcpt, false);
1372 }
1373
1374 /**
1375  * Returns true if there are requests available in incoming
1376  * request queue for processing and it is allowed to fetch them.
1377  * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1378  * to get reliable result
1379  * \see ptlrpc_server_allow_normal
1380  * \see ptlrpc_server_allow high
1381  */
1382 static inline bool
1383 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1384 {
1385         return ptlrpc_server_high_pending(svcpt, force) ||
1386                ptlrpc_server_normal_pending(svcpt, force);
1387 }
1388
1389 /**
1390  * Fetch a request for processing from queue of unprocessed requests.
1391  * Favors high-priority requests.
1392  * Returns a pointer to fetched request.
1393  */
1394 static struct ptlrpc_request *
1395 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1396 {
1397         struct ptlrpc_request *req = NULL;
1398
1399         spin_lock(&svcpt->scp_req_lock);
1400
1401         if (ptlrpc_server_high_pending(svcpt, force)) {
1402                 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1403                 if (req != NULL) {
1404                         svcpt->scp_hreq_count++;
1405                         goto got_request;
1406                 }
1407         }
1408
1409         if (ptlrpc_server_normal_pending(svcpt, force)) {
1410                 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1411                 if (req != NULL) {
1412                         svcpt->scp_hreq_count = 0;
1413                         goto got_request;
1414                 }
1415         }
1416
1417         spin_unlock(&svcpt->scp_req_lock);
1418         return NULL;
1419
1420 got_request:
1421         svcpt->scp_nreqs_active++;
1422         if (req->rq_hp)
1423                 svcpt->scp_nhreqs_active++;
1424
1425         spin_unlock(&svcpt->scp_req_lock);
1426
1427         if (likely(req->rq_export))
1428                 class_export_rpc_inc(req->rq_export);
1429
1430         return req;
1431 }
1432
1433 /**
1434  * Handle freshly incoming reqs, add to timed early reply list,
1435  * pass on to regular request queue.
1436  * All incoming requests pass through here before getting into
1437  * ptlrpc_server_handle_req later on.
1438  */
1439 static int
1440 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1441                             struct ptlrpc_thread *thread)
1442 {
1443         struct ptlrpc_service *svc = svcpt->scp_service;
1444         struct ptlrpc_request *req;
1445         __u32 deadline;
1446         int rc;
1447
1448         spin_lock(&svcpt->scp_lock);
1449         if (list_empty(&svcpt->scp_req_incoming)) {
1450                 spin_unlock(&svcpt->scp_lock);
1451                 return 0;
1452         }
1453
1454         req = list_entry(svcpt->scp_req_incoming.next,
1455                              struct ptlrpc_request, rq_list);
1456         list_del_init(&req->rq_list);
1457         svcpt->scp_nreqs_incoming--;
1458         /* Consider this still a "queued" request as far as stats are
1459          * concerned */
1460         spin_unlock(&svcpt->scp_lock);
1461
1462         /* go through security check/transform */
1463         rc = sptlrpc_svc_unwrap_request(req);
1464         switch (rc) {
1465         case SECSVC_OK:
1466                 break;
1467         case SECSVC_COMPLETE:
1468                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1469                 goto err_req;
1470         case SECSVC_DROP:
1471                 goto err_req;
1472         default:
1473                 LBUG();
1474         }
1475
1476         /*
1477          * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1478          * redo it wouldn't be harmful.
1479          */
1480         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1481                 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1482                 if (rc != 0) {
1483                         CERROR("error unpacking request: ptl %d from %s x%llu\n",
1484                                svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1485                                req->rq_xid);
1486                         goto err_req;
1487                 }
1488         }
1489
1490         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1491         if (rc) {
1492                 CERROR("error unpacking ptlrpc body: ptl %d from %s x%llu\n",
1493                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1494                        req->rq_xid);
1495                 goto err_req;
1496         }
1497
1498         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1499             lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1500                 CERROR("drop incoming rpc opc %u, x%llu\n",
1501                        cfs_fail_val, req->rq_xid);
1502                 goto err_req;
1503         }
1504
1505         rc = -EINVAL;
1506         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1507                 CERROR("wrong packet type received (type=%u) from %s\n",
1508                        lustre_msg_get_type(req->rq_reqmsg),
1509                        libcfs_id2str(req->rq_peer));
1510                 goto err_req;
1511         }
1512
1513         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1514         case MDS_WRITEPAGE:
1515         case OST_WRITE:
1516                 req->rq_bulk_write = 1;
1517                 break;
1518         case MDS_READPAGE:
1519         case OST_READ:
1520         case MGS_CONFIG_READ:
1521                 req->rq_bulk_read = 1;
1522                 break;
1523         }
1524
1525         CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1526
1527         req->rq_export = class_conn2export(
1528                 lustre_msg_get_handle(req->rq_reqmsg));
1529         if (req->rq_export) {
1530                 rc = ptlrpc_check_req(req);
1531                 if (rc == 0) {
1532                         rc = sptlrpc_target_export_check(req->rq_export, req);
1533                         if (rc)
1534                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with illegal security flavor,");
1535                 }
1536
1537                 if (rc)
1538                         goto err_req;
1539         }
1540
1541         /* req_in handling should/must be fast */
1542         if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1543                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1544                           (long)(ktime_get_real_seconds() -
1545                                  req->rq_arrival_time.tv_sec));
1546
1547         /* Set rpc server deadline and add it to the timed list */
1548         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1549                     MSGHDR_AT_SUPPORT) ?
1550                    /* The max time the client expects us to take */
1551                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1552         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1553         if (unlikely(deadline == 0)) {
1554                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1555                 goto err_req;
1556         }
1557
1558         req->rq_svc_thread = thread;
1559
1560         ptlrpc_at_add_timed(req);
1561
1562         /* Move it over to the request processing queue */
1563         rc = ptlrpc_server_request_add(svcpt, req);
1564         if (rc)
1565                 goto err_req;
1566
1567         wake_up(&svcpt->scp_waitq);
1568         return 1;
1569
1570 err_req:
1571         ptlrpc_server_finish_request(svcpt, req);
1572
1573         return 1;
1574 }
1575
1576 /**
1577  * Main incoming request handling logic.
1578  * Calls handler function from service to do actual processing.
1579  */
1580 static int
1581 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1582                              struct ptlrpc_thread *thread)
1583 {
1584         struct ptlrpc_service *svc = svcpt->scp_service;
1585         struct ptlrpc_request *request;
1586         struct timespec64 work_start;
1587         struct timespec64 work_end;
1588         struct timespec64 timediff;
1589         struct timespec64 arrived;
1590         unsigned long timediff_usecs;
1591         unsigned long arrived_usecs;
1592         int rc;
1593         int fail_opc = 0;
1594
1595         request = ptlrpc_server_request_get(svcpt, false);
1596         if (request == NULL)
1597                 return 0;
1598
1599         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1600                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1601         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1602                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1603
1604         if (unlikely(fail_opc)) {
1605                 if (request->rq_export && request->rq_ops)
1606                         OBD_FAIL_TIMEOUT(fail_opc, 4);
1607         }
1608
1609         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1610
1611         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1612                 libcfs_debug_dumplog();
1613
1614         ktime_get_real_ts64(&work_start);
1615         timediff = timespec64_sub(work_start, request->rq_arrival_time);
1616         timediff_usecs = timediff.tv_sec * USEC_PER_SEC +
1617                          timediff.tv_nsec / NSEC_PER_USEC;
1618         if (likely(svc->srv_stats != NULL)) {
1619                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1620                                     timediff_usecs);
1621                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1622                                     svcpt->scp_nreqs_incoming);
1623                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1624                                     svcpt->scp_nreqs_active);
1625                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1626                                     at_get(&svcpt->scp_at_estimate));
1627         }
1628
1629         rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1630         if (rc) {
1631                 CERROR("Failure to initialize session: %d\n", rc);
1632                 goto out_req;
1633         }
1634         request->rq_session.lc_thread = thread;
1635         request->rq_session.lc_cookie = 0x5;
1636         lu_context_enter(&request->rq_session);
1637
1638         CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
1639
1640         request->rq_svc_thread = thread;
1641         if (thread)
1642                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1643
1644         if (likely(request->rq_export)) {
1645                 if (unlikely(ptlrpc_check_req(request)))
1646                         goto put_conn;
1647         }
1648
1649         /* Discard requests queued for longer than the deadline.
1650            The deadline is increased if we send an early reply. */
1651         if (ktime_get_real_seconds() > request->rq_deadline) {
1652                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline " CFS_DURATION_T ":" CFS_DURATION_T "s ago\n",
1653                           libcfs_id2str(request->rq_peer),
1654                           (long)(request->rq_deadline -
1655                                  request->rq_arrival_time.tv_sec),
1656                           (long)(ktime_get_real_seconds() -
1657                                  request->rq_deadline));
1658                 goto put_conn;
1659         }
1660
1661         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d\n",
1662                current_comm(),
1663                (request->rq_export ?
1664                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1665                (request->rq_export ?
1666                 atomic_read(&request->rq_export->exp_refcount) : -99),
1667                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1668                libcfs_id2str(request->rq_peer),
1669                lustre_msg_get_opc(request->rq_reqmsg));
1670
1671         if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1672                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1673
1674         rc = svc->srv_ops.so_req_handler(request);
1675
1676         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1677
1678 put_conn:
1679         lu_context_exit(&request->rq_session);
1680         lu_context_fini(&request->rq_session);
1681
1682         if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
1683                 DEBUG_REQ(D_WARNING, request,
1684                           "Request took longer than estimated (%lld:%llds); "
1685                           "client may timeout.",
1686                           (s64)request->rq_deadline -
1687                                request->rq_arrival_time.tv_sec,
1688                           (s64)ktime_get_real_seconds() - request->rq_deadline);
1689         }
1690
1691         ktime_get_real_ts64(&work_end);
1692         timediff = timespec64_sub(work_end, work_start);
1693         timediff_usecs = timediff.tv_sec * USEC_PER_SEC +
1694                          timediff.tv_nsec / NSEC_PER_USEC;
1695         arrived = timespec64_sub(work_end, request->rq_arrival_time);
1696         arrived_usecs = arrived.tv_sec * USEC_PER_SEC +
1697                          arrived.tv_nsec / NSEC_PER_USEC;
1698         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",
1699                current_comm(),
1700                (request->rq_export ?
1701                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1702                (request->rq_export ?
1703                 atomic_read(&request->rq_export->exp_refcount) : -99),
1704                lustre_msg_get_status(request->rq_reqmsg),
1705                request->rq_xid,
1706                libcfs_id2str(request->rq_peer),
1707                lustre_msg_get_opc(request->rq_reqmsg),
1708                timediff_usecs,
1709                arrived_usecs,
1710                (request->rq_repmsg ?
1711                 lustre_msg_get_transno(request->rq_repmsg) :
1712                 request->rq_transno),
1713                request->rq_status,
1714                (request->rq_repmsg ?
1715                 lustre_msg_get_status(request->rq_repmsg) : -999));
1716         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1717                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1718                 int opc = opcode_offset(op);
1719
1720                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1721                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1722                         lprocfs_counter_add(svc->srv_stats,
1723                                             opc + EXTRA_MAX_OPCODES,
1724                                             timediff_usecs);
1725                 }
1726         }
1727         if (unlikely(request->rq_early_count)) {
1728                 DEBUG_REQ(D_ADAPTTO, request,
1729                           "sent %d early replies before finishing in %llds",
1730                           request->rq_early_count,
1731                           (s64)work_end.tv_sec -
1732                           request->rq_arrival_time.tv_sec);
1733         }
1734
1735 out_req:
1736         ptlrpc_server_finish_active_request(svcpt, request);
1737
1738         return 1;
1739 }
1740
1741 /**
1742  * An internal function to process a single reply state object.
1743  */
1744 static int
1745 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
1746 {
1747         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
1748         struct ptlrpc_service *svc = svcpt->scp_service;
1749         struct obd_export *exp;
1750         int nlocks;
1751         int been_handled;
1752
1753         exp = rs->rs_export;
1754
1755         LASSERT(rs->rs_difficult);
1756         LASSERT(rs->rs_scheduled);
1757         LASSERT(list_empty(&rs->rs_list));
1758
1759         spin_lock(&exp->exp_lock);
1760         /* Noop if removed already */
1761         list_del_init(&rs->rs_exp_list);
1762         spin_unlock(&exp->exp_lock);
1763
1764         /* The disk commit callback holds exp_uncommitted_replies_lock while it
1765          * iterates over newly committed replies, removing them from
1766          * exp_uncommitted_replies.  It then drops this lock and schedules the
1767          * replies it found for handling here.
1768          *
1769          * We can avoid contention for exp_uncommitted_replies_lock between the
1770          * HRT threads and further commit callbacks by checking rs_committed
1771          * which is set in the commit callback while it holds both
1772          * rs_lock and exp_uncommitted_reples.
1773          *
1774          * If we see rs_committed clear, the commit callback _may_ not have
1775          * handled this reply yet and we race with it to grab
1776          * exp_uncommitted_replies_lock before removing the reply from
1777          * exp_uncommitted_replies.  Note that if we lose the race and the
1778          * reply has already been removed, list_del_init() is a noop.
1779          *
1780          * If we see rs_committed set, we know the commit callback is handling,
1781          * or has handled this reply since store reordering might allow us to
1782          * see rs_committed set out of sequence.  But since this is done
1783          * holding rs_lock, we can be sure it has all completed once we hold
1784          * rs_lock, which we do right next.
1785          */
1786         if (!rs->rs_committed) {
1787                 spin_lock(&exp->exp_uncommitted_replies_lock);
1788                 list_del_init(&rs->rs_obd_list);
1789                 spin_unlock(&exp->exp_uncommitted_replies_lock);
1790         }
1791
1792         spin_lock(&rs->rs_lock);
1793
1794         been_handled = rs->rs_handled;
1795         rs->rs_handled = 1;
1796
1797         nlocks = rs->rs_nlocks;          /* atomic "steal", but */
1798         rs->rs_nlocks = 0;                    /* locks still on rs_locks! */
1799
1800         if (nlocks == 0 && !been_handled) {
1801                 /* If we see this, we should already have seen the warning
1802                  * in mds_steal_ack_locks()  */
1803                 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
1804                        rs,
1805                        rs->rs_xid, rs->rs_transno, rs->rs_opc,
1806                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
1807         }
1808
1809         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1810                 spin_unlock(&rs->rs_lock);
1811
1812                 if (!been_handled && rs->rs_on_net) {
1813                         LNetMDUnlink(rs->rs_md_h);
1814                         /* Ignore return code; we're racing with completion */
1815                 }
1816
1817                 while (nlocks-- > 0)
1818                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1819                                          rs->rs_modes[nlocks]);
1820
1821                 spin_lock(&rs->rs_lock);
1822         }
1823
1824         rs->rs_scheduled = 0;
1825
1826         if (!rs->rs_on_net) {
1827                 /* Off the net */
1828                 spin_unlock(&rs->rs_lock);
1829
1830                 class_export_put(exp);
1831                 rs->rs_export = NULL;
1832                 ptlrpc_rs_decref(rs);
1833                 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
1834                     svc->srv_is_stopping)
1835                         wake_up_all(&svcpt->scp_waitq);
1836                 return 1;
1837         }
1838
1839         /* still on the net; callback will schedule */
1840         spin_unlock(&rs->rs_lock);
1841         return 1;
1842 }
1843
1844 static void
1845 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
1846 {
1847         int avail = svcpt->scp_nrqbds_posted;
1848         int low_water = test_req_buffer_pressure ? 0 :
1849                         svcpt->scp_service->srv_nbuf_per_group / 2;
1850
1851         /* NB I'm not locking; just looking. */
1852
1853         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1854          * allowed the request history to grow out of control.  We could put a
1855          * sanity check on that here and cull some history if we need the
1856          * space. */
1857
1858         if (avail <= low_water)
1859                 ptlrpc_grow_req_bufs(svcpt, 1);
1860
1861         if (svcpt->scp_service->srv_stats) {
1862                 lprocfs_counter_add(svcpt->scp_service->srv_stats,
1863                                     PTLRPC_REQBUF_AVAIL_CNTR, avail);
1864         }
1865 }
1866
1867 static int
1868 ptlrpc_retry_rqbds(void *arg)
1869 {
1870         struct ptlrpc_service_part *svcpt = arg;
1871
1872         svcpt->scp_rqbd_timeout = 0;
1873         return -ETIMEDOUT;
1874 }
1875
1876 static inline int
1877 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
1878 {
1879         return svcpt->scp_nreqs_active <
1880                svcpt->scp_nthrs_running - 1 -
1881                (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
1882 }
1883
1884 /**
1885  * allowed to create more threads
1886  * user can call it w/o any lock but need to hold
1887  * ptlrpc_service_part::scp_lock to get reliable result
1888  */
1889 static inline int
1890 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
1891 {
1892         return svcpt->scp_nthrs_running +
1893                svcpt->scp_nthrs_starting <
1894                svcpt->scp_service->srv_nthrs_cpt_limit;
1895 }
1896
1897 /**
1898  * too many requests and allowed to create more threads
1899  */
1900 static inline int
1901 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
1902 {
1903         return !ptlrpc_threads_enough(svcpt) &&
1904                 ptlrpc_threads_increasable(svcpt);
1905 }
1906
1907 static inline int
1908 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
1909 {
1910         return thread_is_stopping(thread) ||
1911                thread->t_svcpt->scp_service->srv_is_stopping;
1912 }
1913
1914 static inline int
1915 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
1916 {
1917         return !list_empty(&svcpt->scp_rqbd_idle) &&
1918                svcpt->scp_rqbd_timeout == 0;
1919 }
1920
1921 static inline int
1922 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
1923 {
1924         return svcpt->scp_at_check;
1925 }
1926
1927 /**
1928  * requests wait on preprocessing
1929  * user can call it w/o any lock but need to hold
1930  * ptlrpc_service_part::scp_lock to get reliable result
1931  */
1932 static inline int
1933 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
1934 {
1935         return !list_empty(&svcpt->scp_req_incoming);
1936 }
1937
1938 static __attribute__((__noinline__)) int
1939 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
1940                   struct ptlrpc_thread *thread)
1941 {
1942         /* Don't exit while there are replies to be handled */
1943         struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
1944                                              ptlrpc_retry_rqbds, svcpt);
1945
1946         /* XXX: Add this back when libcfs watchdog is merged upstream
1947         lc_watchdog_disable(thread->t_watchdog);
1948          */
1949
1950         cond_resched();
1951
1952         l_wait_event_exclusive_head(svcpt->scp_waitq,
1953                                 ptlrpc_thread_stopping(thread) ||
1954                                 ptlrpc_server_request_incoming(svcpt) ||
1955                                 ptlrpc_server_request_pending(svcpt, false) ||
1956                                 ptlrpc_rqbd_pending(svcpt) ||
1957                                 ptlrpc_at_check(svcpt), &lwi);
1958
1959         if (ptlrpc_thread_stopping(thread))
1960                 return -EINTR;
1961
1962         /*
1963         lc_watchdog_touch(thread->t_watchdog,
1964                           ptlrpc_server_get_timeout(svcpt));
1965          */
1966         return 0;
1967 }
1968
1969 /**
1970  * Main thread body for service threads.
1971  * Waits in a loop waiting for new requests to process to appear.
1972  * Every time an incoming requests is added to its queue, a waitq
1973  * is woken up and one of the threads will handle it.
1974  */
1975 static int ptlrpc_main(void *arg)
1976 {
1977         struct ptlrpc_thread *thread = arg;
1978         struct ptlrpc_service_part *svcpt = thread->t_svcpt;
1979         struct ptlrpc_service *svc = svcpt->scp_service;
1980         struct ptlrpc_reply_state *rs;
1981         struct group_info *ginfo = NULL;
1982         struct lu_env *env;
1983         int counter = 0, rc = 0;
1984
1985         thread->t_pid = current_pid();
1986         unshare_fs_struct();
1987
1988         /* NB: we will call cfs_cpt_bind() for all threads, because we
1989          * might want to run lustre server only on a subset of system CPUs,
1990          * in that case ->scp_cpt is CFS_CPT_ANY */
1991         rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
1992         if (rc != 0) {
1993                 CWARN("%s: failed to bind %s on CPT %d\n",
1994                       svc->srv_name, thread->t_name, svcpt->scp_cpt);
1995         }
1996
1997         ginfo = groups_alloc(0);
1998         if (!ginfo) {
1999                 rc = -ENOMEM;
2000                 goto out;
2001         }
2002
2003         set_current_groups(ginfo);
2004         put_group_info(ginfo);
2005
2006         if (svc->srv_ops.so_thr_init != NULL) {
2007                 rc = svc->srv_ops.so_thr_init(thread);
2008                 if (rc)
2009                         goto out;
2010         }
2011
2012         env = kzalloc(sizeof(*env), GFP_NOFS);
2013         if (!env) {
2014                 rc = -ENOMEM;
2015                 goto out_srv_fini;
2016         }
2017
2018         rc = lu_context_init(&env->le_ctx,
2019                              svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2020         if (rc)
2021                 goto out_srv_fini;
2022
2023         thread->t_env = env;
2024         env->le_ctx.lc_thread = thread;
2025         env->le_ctx.lc_cookie = 0x6;
2026
2027         while (!list_empty(&svcpt->scp_rqbd_idle)) {
2028                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2029                 if (rc >= 0)
2030                         continue;
2031
2032                 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2033                         svc->srv_name, svcpt->scp_cpt, rc);
2034                 goto out_srv_fini;
2035         }
2036
2037         /* Alloc reply state structure for this one */
2038         rs = libcfs_kvzalloc(svc->srv_max_reply_size, GFP_NOFS);
2039         if (!rs) {
2040                 rc = -ENOMEM;
2041                 goto out_srv_fini;
2042         }
2043
2044         spin_lock(&svcpt->scp_lock);
2045
2046         LASSERT(thread_is_starting(thread));
2047         thread_clear_flags(thread, SVC_STARTING);
2048
2049         LASSERT(svcpt->scp_nthrs_starting == 1);
2050         svcpt->scp_nthrs_starting--;
2051
2052         /* SVC_STOPPING may already be set here if someone else is trying
2053          * to stop the service while this new thread has been dynamically
2054          * forked. We still set SVC_RUNNING to let our creator know that
2055          * we are now running, however we will exit as soon as possible */
2056         thread_add_flags(thread, SVC_RUNNING);
2057         svcpt->scp_nthrs_running++;
2058         spin_unlock(&svcpt->scp_lock);
2059
2060         /* wake up our creator in case he's still waiting. */
2061         wake_up(&thread->t_ctl_waitq);
2062
2063         /*
2064         thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2065                                              NULL, NULL);
2066          */
2067
2068         spin_lock(&svcpt->scp_rep_lock);
2069         list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2070         wake_up(&svcpt->scp_rep_waitq);
2071         spin_unlock(&svcpt->scp_rep_lock);
2072
2073         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2074                svcpt->scp_nthrs_running);
2075
2076         /* XXX maintain a list of all managed devices: insert here */
2077         while (!ptlrpc_thread_stopping(thread)) {
2078                 if (ptlrpc_wait_event(svcpt, thread))
2079                         break;
2080
2081                 ptlrpc_check_rqbd_pool(svcpt);
2082
2083                 if (ptlrpc_threads_need_create(svcpt)) {
2084                         /* Ignore return code - we tried... */
2085                         ptlrpc_start_thread(svcpt, 0);
2086                 }
2087
2088                 /* Process all incoming reqs before handling any */
2089                 if (ptlrpc_server_request_incoming(svcpt)) {
2090                         lu_context_enter(&env->le_ctx);
2091                         env->le_ses = NULL;
2092                         ptlrpc_server_handle_req_in(svcpt, thread);
2093                         lu_context_exit(&env->le_ctx);
2094
2095                         /* but limit ourselves in case of flood */
2096                         if (counter++ < 100)
2097                                 continue;
2098                         counter = 0;
2099                 }
2100
2101                 if (ptlrpc_at_check(svcpt))
2102                         ptlrpc_at_check_timed(svcpt);
2103
2104                 if (ptlrpc_server_request_pending(svcpt, false)) {
2105                         lu_context_enter(&env->le_ctx);
2106                         ptlrpc_server_handle_request(svcpt, thread);
2107                         lu_context_exit(&env->le_ctx);
2108                 }
2109
2110                 if (ptlrpc_rqbd_pending(svcpt) &&
2111                     ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2112                         /* I just failed to repost request buffers.
2113                          * Wait for a timeout (unless something else
2114                          * happens) before I try again */
2115                         svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2116                         CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2117                                svcpt->scp_nrqbds_posted);
2118                 }
2119         }
2120
2121         /*
2122         lc_watchdog_delete(thread->t_watchdog);
2123         thread->t_watchdog = NULL;
2124         */
2125
2126 out_srv_fini:
2127         /*
2128          * deconstruct service specific state created by ptlrpc_start_thread()
2129          */
2130         if (svc->srv_ops.so_thr_done != NULL)
2131                 svc->srv_ops.so_thr_done(thread);
2132
2133         if (env != NULL) {
2134                 lu_context_fini(&env->le_ctx);
2135                 kfree(env);
2136         }
2137 out:
2138         CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2139                thread, thread->t_pid, thread->t_id, rc);
2140
2141         spin_lock(&svcpt->scp_lock);
2142         if (thread_test_and_clear_flags(thread, SVC_STARTING))
2143                 svcpt->scp_nthrs_starting--;
2144
2145         if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2146                 /* must know immediately */
2147                 svcpt->scp_nthrs_running--;
2148         }
2149
2150         thread->t_id = rc;
2151         thread_add_flags(thread, SVC_STOPPED);
2152
2153         wake_up(&thread->t_ctl_waitq);
2154         spin_unlock(&svcpt->scp_lock);
2155
2156         return rc;
2157 }
2158
2159 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2160                           struct list_head *replies)
2161 {
2162         int result;
2163
2164         spin_lock(&hrt->hrt_lock);
2165
2166         list_splice_init(&hrt->hrt_queue, replies);
2167         result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2168
2169         spin_unlock(&hrt->hrt_lock);
2170         return result;
2171 }
2172
2173 /**
2174  * Main body of "handle reply" function.
2175  * It processes acked reply states
2176  */
2177 static int ptlrpc_hr_main(void *arg)
2178 {
2179         struct ptlrpc_hr_thread *hrt = arg;
2180         struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2181         LIST_HEAD       (replies);
2182         char threadname[20];
2183         int rc;
2184
2185         snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2186                  hrp->hrp_cpt, hrt->hrt_id);
2187         unshare_fs_struct();
2188
2189         rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2190         if (rc != 0) {
2191                 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2192                       threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2193         }
2194
2195         atomic_inc(&hrp->hrp_nstarted);
2196         wake_up(&ptlrpc_hr.hr_waitq);
2197
2198         while (!ptlrpc_hr.hr_stopping) {
2199                 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2200
2201                 while (!list_empty(&replies)) {
2202                         struct ptlrpc_reply_state *rs;
2203
2204                         rs = list_entry(replies.prev,
2205                                             struct ptlrpc_reply_state,
2206                                             rs_list);
2207                         list_del_init(&rs->rs_list);
2208                         ptlrpc_handle_rs(rs);
2209                 }
2210         }
2211
2212         atomic_inc(&hrp->hrp_nstopped);
2213         wake_up(&ptlrpc_hr.hr_waitq);
2214
2215         return 0;
2216 }
2217
2218 static void ptlrpc_stop_hr_threads(void)
2219 {
2220         struct ptlrpc_hr_partition *hrp;
2221         int i;
2222         int j;
2223
2224         ptlrpc_hr.hr_stopping = 1;
2225
2226         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2227                 if (hrp->hrp_thrs == NULL)
2228                         continue; /* uninitialized */
2229                 for (j = 0; j < hrp->hrp_nthrs; j++)
2230                         wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2231         }
2232
2233         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2234                 if (hrp->hrp_thrs == NULL)
2235                         continue; /* uninitialized */
2236                 wait_event(ptlrpc_hr.hr_waitq,
2237                                atomic_read(&hrp->hrp_nstopped) ==
2238                                atomic_read(&hrp->hrp_nstarted));
2239         }
2240 }
2241
2242 static int ptlrpc_start_hr_threads(void)
2243 {
2244         struct ptlrpc_hr_partition *hrp;
2245         int i;
2246         int j;
2247
2248         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2249                 int rc = 0;
2250
2251                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2252                         struct  ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2253
2254                         rc = PTR_ERR(kthread_run(ptlrpc_hr_main,
2255                                                  &hrp->hrp_thrs[j],
2256                                                  "ptlrpc_hr%02d_%03d",
2257                                                  hrp->hrp_cpt,
2258                                                  hrt->hrt_id));
2259                         if (IS_ERR_VALUE(rc))
2260                                 break;
2261                 }
2262                 wait_event(ptlrpc_hr.hr_waitq,
2263                                atomic_read(&hrp->hrp_nstarted) == j);
2264                 if (!IS_ERR_VALUE(rc))
2265                         continue;
2266
2267                 CERROR("Reply handling thread %d:%d Failed on starting: rc = %d\n",
2268                        i, j, rc);
2269                 ptlrpc_stop_hr_threads();
2270                 return rc;
2271         }
2272         return 0;
2273 }
2274
2275 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2276 {
2277         struct l_wait_info lwi = { 0 };
2278         struct ptlrpc_thread *thread;
2279         LIST_HEAD       (zombie);
2280
2281         CDEBUG(D_INFO, "Stopping threads for service %s\n",
2282                svcpt->scp_service->srv_name);
2283
2284         spin_lock(&svcpt->scp_lock);
2285         /* let the thread know that we would like it to stop asap */
2286         list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2287                 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2288                        svcpt->scp_service->srv_thread_name, thread->t_id);
2289                 thread_add_flags(thread, SVC_STOPPING);
2290         }
2291
2292         wake_up_all(&svcpt->scp_waitq);
2293
2294         while (!list_empty(&svcpt->scp_threads)) {
2295                 thread = list_entry(svcpt->scp_threads.next,
2296                                         struct ptlrpc_thread, t_link);
2297                 if (thread_is_stopped(thread)) {
2298                         list_del(&thread->t_link);
2299                         list_add(&thread->t_link, &zombie);
2300                         continue;
2301                 }
2302                 spin_unlock(&svcpt->scp_lock);
2303
2304                 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2305                        svcpt->scp_service->srv_thread_name, thread->t_id);
2306                 l_wait_event(thread->t_ctl_waitq,
2307                              thread_is_stopped(thread), &lwi);
2308
2309                 spin_lock(&svcpt->scp_lock);
2310         }
2311
2312         spin_unlock(&svcpt->scp_lock);
2313
2314         while (!list_empty(&zombie)) {
2315                 thread = list_entry(zombie.next,
2316                                         struct ptlrpc_thread, t_link);
2317                 list_del(&thread->t_link);
2318                 kfree(thread);
2319         }
2320 }
2321
2322 /**
2323  * Stops all threads of a particular service \a svc
2324  */
2325 static void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2326 {
2327         struct ptlrpc_service_part *svcpt;
2328         int i;
2329
2330         ptlrpc_service_for_each_part(svcpt, i, svc) {
2331                 if (svcpt->scp_service != NULL)
2332                         ptlrpc_svcpt_stop_threads(svcpt);
2333         }
2334 }
2335
2336 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2337 {
2338         int rc = 0;
2339         int i;
2340         int j;
2341
2342         /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2343         LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2344
2345         for (i = 0; i < svc->srv_ncpts; i++) {
2346                 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2347                         rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2348                         if (rc == 0)
2349                                 continue;
2350
2351                         if (rc != -EMFILE)
2352                                 goto failed;
2353                         /* We have enough threads, don't start more. b=15759 */
2354                         break;
2355                 }
2356         }
2357
2358         return 0;
2359  failed:
2360         CERROR("cannot start %s thread #%d_%d: rc %d\n",
2361                svc->srv_thread_name, i, j, rc);
2362         ptlrpc_stop_all_threads(svc);
2363         return rc;
2364 }
2365 EXPORT_SYMBOL(ptlrpc_start_threads);
2366
2367 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2368 {
2369         struct l_wait_info lwi = { 0 };
2370         struct ptlrpc_thread *thread;
2371         struct ptlrpc_service *svc;
2372         int rc;
2373
2374         LASSERT(svcpt != NULL);
2375
2376         svc = svcpt->scp_service;
2377
2378         CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2379                svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2380                svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2381
2382  again:
2383         if (unlikely(svc->srv_is_stopping))
2384                 return -ESRCH;
2385
2386         if (!ptlrpc_threads_increasable(svcpt) ||
2387             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2388              svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2389                 return -EMFILE;
2390
2391         thread = kzalloc_node(sizeof(*thread), GFP_NOFS,
2392                               cfs_cpt_spread_node(svc->srv_cptable,
2393                                                   svcpt->scp_cpt));
2394         if (thread == NULL)
2395                 return -ENOMEM;
2396         init_waitqueue_head(&thread->t_ctl_waitq);
2397
2398         spin_lock(&svcpt->scp_lock);
2399         if (!ptlrpc_threads_increasable(svcpt)) {
2400                 spin_unlock(&svcpt->scp_lock);
2401                 kfree(thread);
2402                 return -EMFILE;
2403         }
2404
2405         if (svcpt->scp_nthrs_starting != 0) {
2406                 /* serialize starting because some modules (obdfilter)
2407                  * might require unique and contiguous t_id */
2408                 LASSERT(svcpt->scp_nthrs_starting == 1);
2409                 spin_unlock(&svcpt->scp_lock);
2410                 kfree(thread);
2411                 if (wait) {
2412                         CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2413                                svc->srv_thread_name, svcpt->scp_thr_nextid);
2414                         schedule();
2415                         goto again;
2416                 }
2417
2418                 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2419                        svc->srv_thread_name, svcpt->scp_thr_nextid);
2420                 return -EAGAIN;
2421         }
2422
2423         svcpt->scp_nthrs_starting++;
2424         thread->t_id = svcpt->scp_thr_nextid++;
2425         thread_add_flags(thread, SVC_STARTING);
2426         thread->t_svcpt = svcpt;
2427
2428         list_add(&thread->t_link, &svcpt->scp_threads);
2429         spin_unlock(&svcpt->scp_lock);
2430
2431         if (svcpt->scp_cpt >= 0) {
2432                 snprintf(thread->t_name, sizeof(thread->t_name), "%s%02d_%03d",
2433                          svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2434         } else {
2435                 snprintf(thread->t_name, sizeof(thread->t_name), "%s_%04d",
2436                          svc->srv_thread_name, thread->t_id);
2437         }
2438
2439         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2440         rc = PTR_ERR(kthread_run(ptlrpc_main, thread, "%s", thread->t_name));
2441         if (IS_ERR_VALUE(rc)) {
2442                 CERROR("cannot start thread '%s': rc %d\n",
2443                        thread->t_name, rc);
2444                 spin_lock(&svcpt->scp_lock);
2445                 --svcpt->scp_nthrs_starting;
2446                 if (thread_is_stopping(thread)) {
2447                         /* this ptlrpc_thread is being handled
2448                          * by ptlrpc_svcpt_stop_threads now
2449                          */
2450                         thread_add_flags(thread, SVC_STOPPED);
2451                         wake_up(&thread->t_ctl_waitq);
2452                         spin_unlock(&svcpt->scp_lock);
2453                 } else {
2454                         list_del(&thread->t_link);
2455                         spin_unlock(&svcpt->scp_lock);
2456                         kfree(thread);
2457                 }
2458                 return rc;
2459         }
2460
2461         if (!wait)
2462                 return 0;
2463
2464         l_wait_event(thread->t_ctl_waitq,
2465                      thread_is_running(thread) || thread_is_stopped(thread),
2466                      &lwi);
2467
2468         rc = thread_is_stopped(thread) ? thread->t_id : 0;
2469         return rc;
2470 }
2471
2472 int ptlrpc_hr_init(void)
2473 {
2474         struct ptlrpc_hr_partition *hrp;
2475         struct ptlrpc_hr_thread *hrt;
2476         int rc;
2477         int i;
2478         int j;
2479         int weight;
2480
2481         memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2482         ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2483
2484         ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2485                                                    sizeof(*hrp));
2486         if (ptlrpc_hr.hr_partitions == NULL)
2487                 return -ENOMEM;
2488
2489         init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2490
2491         weight = cpumask_weight(topology_sibling_cpumask(0));
2492
2493         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2494                 hrp->hrp_cpt = i;
2495
2496                 atomic_set(&hrp->hrp_nstarted, 0);
2497                 atomic_set(&hrp->hrp_nstopped, 0);
2498
2499                 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2500                 hrp->hrp_nthrs /= weight;
2501
2502                 LASSERT(hrp->hrp_nthrs > 0);
2503                 hrp->hrp_thrs =
2504                         kzalloc_node(hrp->hrp_nthrs * sizeof(*hrt), GFP_NOFS,
2505                                 cfs_cpt_spread_node(ptlrpc_hr.hr_cpt_table,
2506                                                     i));
2507                 if (hrp->hrp_thrs == NULL) {
2508                         rc = -ENOMEM;
2509                         goto out;
2510                 }
2511
2512                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2513                         hrt = &hrp->hrp_thrs[j];
2514
2515                         hrt->hrt_id = j;
2516                         hrt->hrt_partition = hrp;
2517                         init_waitqueue_head(&hrt->hrt_waitq);
2518                         spin_lock_init(&hrt->hrt_lock);
2519                         INIT_LIST_HEAD(&hrt->hrt_queue);
2520                 }
2521         }
2522
2523         rc = ptlrpc_start_hr_threads();
2524 out:
2525         if (rc != 0)
2526                 ptlrpc_hr_fini();
2527         return rc;
2528 }
2529
2530 void ptlrpc_hr_fini(void)
2531 {
2532         struct ptlrpc_hr_partition *hrp;
2533         int i;
2534
2535         if (ptlrpc_hr.hr_partitions == NULL)
2536                 return;
2537
2538         ptlrpc_stop_hr_threads();
2539
2540         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2541                 kfree(hrp->hrp_thrs);
2542         }
2543
2544         cfs_percpt_free(ptlrpc_hr.hr_partitions);
2545         ptlrpc_hr.hr_partitions = NULL;
2546 }
2547
2548 /**
2549  * Wait until all already scheduled replies are processed.
2550  */
2551 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2552 {
2553         while (1) {
2554                 int rc;
2555                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2556                                                      NULL, NULL);
2557
2558                 rc = l_wait_event(svcpt->scp_waitq,
2559                      atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2560                 if (rc == 0)
2561                         break;
2562                 CWARN("Unexpectedly long timeout %s %p\n",
2563                       svcpt->scp_service->srv_name, svcpt->scp_service);
2564         }
2565 }
2566
2567 static void
2568 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2569 {
2570         struct ptlrpc_service_part *svcpt;
2571         int i;
2572
2573         /* early disarm AT timer... */
2574         ptlrpc_service_for_each_part(svcpt, i, svc) {
2575                 if (svcpt->scp_service != NULL)
2576                         del_timer(&svcpt->scp_at_timer);
2577         }
2578 }
2579
2580 static void
2581 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2582 {
2583         struct ptlrpc_service_part *svcpt;
2584         struct ptlrpc_request_buffer_desc *rqbd;
2585         struct l_wait_info lwi;
2586         int rc;
2587         int i;
2588
2589         /* All history will be culled when the next request buffer is
2590          * freed in ptlrpc_service_purge_all() */
2591         svc->srv_hist_nrqbds_cpt_max = 0;
2592
2593         rc = LNetClearLazyPortal(svc->srv_req_portal);
2594         LASSERT(rc == 0);
2595
2596         ptlrpc_service_for_each_part(svcpt, i, svc) {
2597                 if (svcpt->scp_service == NULL)
2598                         break;
2599
2600                 /* Unlink all the request buffers.  This forces a 'final'
2601                  * event with its 'unlink' flag set for each posted rqbd */
2602                 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2603                                         rqbd_list) {
2604                         rc = LNetMDUnlink(rqbd->rqbd_md_h);
2605                         LASSERT(rc == 0 || rc == -ENOENT);
2606                 }
2607         }
2608
2609         ptlrpc_service_for_each_part(svcpt, i, svc) {
2610                 if (svcpt->scp_service == NULL)
2611                         break;
2612
2613                 /* Wait for the network to release any buffers
2614                  * it's currently filling */
2615                 spin_lock(&svcpt->scp_lock);
2616                 while (svcpt->scp_nrqbds_posted != 0) {
2617                         spin_unlock(&svcpt->scp_lock);
2618                         /* Network access will complete in finite time but
2619                          * the HUGE timeout lets us CWARN for visibility
2620                          * of sluggish NALs */
2621                         lwi = LWI_TIMEOUT_INTERVAL(
2622                                         cfs_time_seconds(LONG_UNLINK),
2623                                         cfs_time_seconds(1), NULL, NULL);
2624                         rc = l_wait_event(svcpt->scp_waitq,
2625                                           svcpt->scp_nrqbds_posted == 0, &lwi);
2626                         if (rc == -ETIMEDOUT) {
2627                                 CWARN("Service %s waiting for request buffers\n",
2628                                       svcpt->scp_service->srv_name);
2629                         }
2630                         spin_lock(&svcpt->scp_lock);
2631                 }
2632                 spin_unlock(&svcpt->scp_lock);
2633         }
2634 }
2635
2636 static void
2637 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2638 {
2639         struct ptlrpc_service_part *svcpt;
2640         struct ptlrpc_request_buffer_desc *rqbd;
2641         struct ptlrpc_request *req;
2642         struct ptlrpc_reply_state *rs;
2643         int i;
2644
2645         ptlrpc_service_for_each_part(svcpt, i, svc) {
2646                 if (svcpt->scp_service == NULL)
2647                         break;
2648
2649                 spin_lock(&svcpt->scp_rep_lock);
2650                 while (!list_empty(&svcpt->scp_rep_active)) {
2651                         rs = list_entry(svcpt->scp_rep_active.next,
2652                                             struct ptlrpc_reply_state, rs_list);
2653                         spin_lock(&rs->rs_lock);
2654                         ptlrpc_schedule_difficult_reply(rs);
2655                         spin_unlock(&rs->rs_lock);
2656                 }
2657                 spin_unlock(&svcpt->scp_rep_lock);
2658
2659                 /* purge the request queue.  NB No new replies (rqbds
2660                  * all unlinked) and no service threads, so I'm the only
2661                  * thread noodling the request queue now */
2662                 while (!list_empty(&svcpt->scp_req_incoming)) {
2663                         req = list_entry(svcpt->scp_req_incoming.next,
2664                                              struct ptlrpc_request, rq_list);
2665
2666                         list_del(&req->rq_list);
2667                         svcpt->scp_nreqs_incoming--;
2668                         ptlrpc_server_finish_request(svcpt, req);
2669                 }
2670
2671                 while (ptlrpc_server_request_pending(svcpt, true)) {
2672                         req = ptlrpc_server_request_get(svcpt, true);
2673                         ptlrpc_server_finish_active_request(svcpt, req);
2674                 }
2675
2676                 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
2677                 LASSERT(svcpt->scp_nreqs_incoming == 0);
2678                 LASSERT(svcpt->scp_nreqs_active == 0);
2679                 /* history should have been culled by
2680                  * ptlrpc_server_finish_request */
2681                 LASSERT(svcpt->scp_hist_nrqbds == 0);
2682
2683                 /* Now free all the request buffers since nothing
2684                  * references them any more... */
2685
2686                 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2687                         rqbd = list_entry(svcpt->scp_rqbd_idle.next,
2688                                               struct ptlrpc_request_buffer_desc,
2689                                               rqbd_list);
2690                         ptlrpc_free_rqbd(rqbd);
2691                 }
2692                 ptlrpc_wait_replies(svcpt);
2693
2694                 while (!list_empty(&svcpt->scp_rep_idle)) {
2695                         rs = list_entry(svcpt->scp_rep_idle.next,
2696                                             struct ptlrpc_reply_state,
2697                                             rs_list);
2698                         list_del(&rs->rs_list);
2699                         kvfree(rs);
2700                 }
2701         }
2702 }
2703
2704 static void
2705 ptlrpc_service_free(struct ptlrpc_service *svc)
2706 {
2707         struct ptlrpc_service_part *svcpt;
2708         struct ptlrpc_at_array *array;
2709         int i;
2710
2711         ptlrpc_service_for_each_part(svcpt, i, svc) {
2712                 if (svcpt->scp_service == NULL)
2713                         break;
2714
2715                 /* In case somebody rearmed this in the meantime */
2716                 del_timer(&svcpt->scp_at_timer);
2717                 array = &svcpt->scp_at_array;
2718
2719                 kfree(array->paa_reqs_array);
2720                 array->paa_reqs_array = NULL;
2721                 kfree(array->paa_reqs_count);
2722                 array->paa_reqs_count = NULL;
2723         }
2724
2725         ptlrpc_service_for_each_part(svcpt, i, svc)
2726                 kfree(svcpt);
2727
2728         if (svc->srv_cpts != NULL)
2729                 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
2730
2731         kfree(svc);
2732 }
2733
2734 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2735 {
2736         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2737
2738         service->srv_is_stopping = 1;
2739
2740         mutex_lock(&ptlrpc_all_services_mutex);
2741         list_del_init(&service->srv_list);
2742         mutex_unlock(&ptlrpc_all_services_mutex);
2743
2744         ptlrpc_service_del_atimer(service);
2745         ptlrpc_stop_all_threads(service);
2746
2747         ptlrpc_service_unlink_rqbd(service);
2748         ptlrpc_service_purge_all(service);
2749         ptlrpc_service_nrs_cleanup(service);
2750
2751         ptlrpc_lprocfs_unregister_service(service);
2752         ptlrpc_sysfs_unregister_service(service);
2753
2754         ptlrpc_service_free(service);
2755
2756         return 0;
2757 }
2758 EXPORT_SYMBOL(ptlrpc_unregister_service);