lustre/ptlrpc/ptlrpcd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/ptlrpc/ptlrpcd.c
  32  */
  33
  34 /** \defgroup ptlrpcd PortalRPC daemon
  35  *
  36  * ptlrpcd is a special thread with its own set where other user might add
  37  * requests when they don't want to wait for their completion.
  38  * PtlRPCD will take care of sending such requests and then processing their
  39  * replies and calling completion callbacks as necessary.
  40  * The callbacks are called directly from ptlrpcd context.
  41  * It is important to never significantly block (esp. on RPCs!) within such
  42  * completion handler or a deadlock might occur where ptlrpcd enters some
  43  * callback that attempts to send another RPC and wait for it to return,
  44  * during which time ptlrpcd is completely blocked, so e.g. if import
  45  * fails, recovery cannot progress because connection requests are also
  46  * sent by ptlrpcd.
  47  *
  48  * @{
  49  */
  50
  51 #define DEBUG_SUBSYSTEM S_RPC
  52
  53 #include <linux/kthread.h>
  54 #include <libcfs/libcfs.h>
  55 #include <lustre_net.h>
  56 #include <lustre_lib.h>
  57 #include <lustre_ha.h>
  58 #include <obd_class.h>   /* for obd_zombie */
  59 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  60 #include <cl_object.h> /* cl_env_{get,put}() */
  61 #include <lprocfs_status.h>
  62
  63 #include "ptlrpc_internal.h"
  64
  65 /* One of these per CPT. */
  66 struct ptlrpcd {
  67         int                     pd_size;
  68         int                     pd_index;
  69         int                     pd_cpt;
  70         int                     pd_cursor;
  71         int                     pd_nthreads;
  72         int                     pd_groupsize;
  73         struct ptlrpcd_ctl      pd_threads[0];
  74 };
  75
  76 /*
  77  * max_ptlrpcds is obsolete, but retained to ensure that the kernel
  78  * module will load on a system where it has been tuned.
  79  * A value other than 0 implies it was tuned, in which case the value
  80  * is used to derive a setting for ptlrpcd_per_cpt_max.
  81  */
  82 static int max_ptlrpcds;
  83 module_param(max_ptlrpcds, int, 0644);
  84 MODULE_PARM_DESC(max_ptlrpcds,
  85                  "Max ptlrpcd thread count to be started (obsolete).");
  86
  87 /*
  88  * ptlrpcd_bind_policy is obsolete, but retained to ensure that
  89  * the kernel module will load on a system where it has been tuned.
  90  * A value other than 0 implies it was tuned, in which case the value
  91  * is used to derive a setting for ptlrpcd_partner_group_size.
  92  */
  93 static int ptlrpcd_bind_policy;
  94 module_param(ptlrpcd_bind_policy, int, 0644);
  95 MODULE_PARM_DESC(ptlrpcd_bind_policy,
  96                  "Ptlrpcd threads binding mode (obsolete).");
  97
  98 /*
  99  * ptlrpcd_per_cpt_max: The maximum number of ptlrpcd threads to run
 100  * in a CPT.
 101  */
 102 static int ptlrpcd_per_cpt_max;
 103 module_param(ptlrpcd_per_cpt_max, int, 0644);
 104 MODULE_PARM_DESC(ptlrpcd_per_cpt_max,
 105                  "Max ptlrpcd thread count to be started per CPT.");
 106
 107 /*
 108  * ptlrpcd_partner_group_size: The desired number of threads in each
 109  * ptlrpcd partner thread group. Default is 2, corresponding to the
 110  * old PDB_POLICY_PAIR. A negative value makes all ptlrpcd threads in
 111  * a CPT partners of each other.
 112  */
 113 static int ptlrpcd_partner_group_size;
 114 module_param(ptlrpcd_partner_group_size, int, 0644);
 115 MODULE_PARM_DESC(ptlrpcd_partner_group_size,
 116                  "Number of ptlrpcd threads in a partner group.");
 117
 118 /*
 119  * ptlrpcd_cpts: A CPT string describing the CPU partitions that
 120  * ptlrpcd threads should run on. Used to make ptlrpcd threads run on
 121  * a subset of all CPTs.
 122  *
 123  * ptlrpcd_cpts=2
 124  * ptlrpcd_cpts=[2]
 125  *   run ptlrpcd threads only on CPT 2.
 126  *
 127  * ptlrpcd_cpts=0-3
 128  * ptlrpcd_cpts=[0-3]
 129  *   run ptlrpcd threads on CPTs 0, 1, 2, and 3.
 130  *
 131  * ptlrpcd_cpts=[0-3,5,7]
 132  *   run ptlrpcd threads on CPTS 0, 1, 2, 3, 5, and 7.
 133  */
 134 static char *ptlrpcd_cpts;
 135 module_param(ptlrpcd_cpts, charp, 0644);
 136 MODULE_PARM_DESC(ptlrpcd_cpts,
 137                  "CPU partitions ptlrpcd threads should run in");
 138
 139 /* ptlrpcds_cpt_idx maps cpt numbers to an index in the ptlrpcds array. */
 140 static int              *ptlrpcds_cpt_idx;
 141
 142 /* ptlrpcds_num is the number of entries in the ptlrpcds array. */
 143 static int              ptlrpcds_num;
 144 static struct ptlrpcd   **ptlrpcds;
 145
 146 /*
 147  * In addition to the regular thread pool above, there is a single
 148  * global recovery thread. Recovery isn't critical for performance,
 149  * and doesn't block, but must always be able to proceed, and it is
 150  * possible that all normal ptlrpcd threads are blocked. Hence the
 151  * need for a dedicated thread.
 152  */
 153 static struct ptlrpcd_ctl ptlrpcd_rcv;
 154
 155 struct mutex ptlrpcd_mutex;
 156 static int ptlrpcd_users = 0;
 157
 158 void ptlrpcd_wake(struct ptlrpc_request *req)
 159 {
 160         struct ptlrpc_request_set *set = req->rq_set;
 161
 162         LASSERT(set != NULL);
 163         wake_up(&set->set_waitq);
 164 }
 165 EXPORT_SYMBOL(ptlrpcd_wake);
 166
 167 static struct ptlrpcd_ctl *
 168 ptlrpcd_select_pc(struct ptlrpc_request *req)
 169 {
 170         struct ptlrpcd  *pd;
 171         int             cpt;
 172         int             idx;
 173
 174         if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
 175                 return &ptlrpcd_rcv;
 176
 177         cpt = cfs_cpt_current(cfs_cpt_tab, 1);
 178         if (ptlrpcds_cpt_idx == NULL)
 179                 idx = cpt;
 180         else
 181                 idx = ptlrpcds_cpt_idx[cpt];
 182         pd = ptlrpcds[idx];
 183
 184         /* We do not care whether it is strict load balance. */
 185         idx = pd->pd_cursor;
 186         if (++idx == pd->pd_nthreads)
 187                 idx = 0;
 188         pd->pd_cursor = idx;
 189
 190         return &pd->pd_threads[idx];
 191 }
 192
 193 /**
 194  * Move all request from an existing request set to the ptlrpcd queue.
 195  * All requests from the set must be in phase RQ_PHASE_NEW.
 196  */
 197 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 198 {
 199         struct list_head *tmp, *pos;
 200         struct ptlrpcd_ctl *pc;
 201         struct ptlrpc_request_set *new;
 202         int count, i;
 203
 204         pc = ptlrpcd_select_pc(NULL);
 205         new = pc->pc_set;
 206
 207         list_for_each_safe(pos, tmp, &set->set_requests) {
 208                 struct ptlrpc_request *req =
 209                         list_entry(pos, struct ptlrpc_request,
 210                                    rq_set_chain);
 211
 212                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 213                 req->rq_set = new;
 214                 req->rq_queued_time = ktime_get_seconds();
 215         }
 216
 217         spin_lock(&new->set_new_req_lock);
 218         list_splice_init(&set->set_requests, &new->set_new_requests);
 219         i = atomic_read(&set->set_remaining);
 220         count = atomic_add_return(i, &new->set_new_count);
 221         atomic_set(&set->set_remaining, 0);
 222         spin_unlock(&new->set_new_req_lock);
 223         if (count == i) {
 224                 wake_up(&new->set_waitq);
 225
 226                 /*
 227                  * XXX: It maybe unnecessary to wakeup all the partners. But to
 228                  *      guarantee the async RPC can be processed ASAP, we have
 229                  *      no other better choice. It maybe fixed in future.
 230                  */
 231                 for (i = 0; i < pc->pc_npartners; i++)
 232                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
 233         }
 234 }
 235
 236 /**
 237  * Return transferred RPCs count.
 238  */
 239 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 240                                struct ptlrpc_request_set *src)
 241 {
 242         struct ptlrpc_request *req;
 243         int rc = 0;
 244
 245         spin_lock(&src->set_new_req_lock);
 246         if (likely(!list_empty(&src->set_new_requests))) {
 247                 list_for_each_entry(req, &src->set_new_requests, rq_set_chain)
 248                         req->rq_set = des;
 249
 250                 list_splice_init(&src->set_new_requests,
 251                                  &des->set_requests);
 252                 rc = atomic_read(&src->set_new_count);
 253                 atomic_add(rc, &des->set_remaining);
 254                 atomic_set(&src->set_new_count, 0);
 255         }
 256         spin_unlock(&src->set_new_req_lock);
 257         return rc;
 258 }
 259
 260 /**
 261  * Requests that are added to the ptlrpcd queue are sent via
 262  * ptlrpcd_check->ptlrpc_check_set().
 263  */
 264 void ptlrpcd_add_req(struct ptlrpc_request *req)
 265 {
 266         struct ptlrpcd_ctl *pc;
 267
 268         if (req->rq_reqmsg)
 269                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
 270
 271         spin_lock(&req->rq_lock);
 272         if (req->rq_invalid_rqset) {
 273                 req->rq_invalid_rqset = 0;
 274                 spin_unlock(&req->rq_lock);
 275                 if (wait_event_idle_timeout(req->rq_set_waitq,
 276                                             req->rq_set == NULL,
 277                                             cfs_time_seconds(5)) == 0)
 278                         l_wait_event_abortable(req->rq_set_waitq,
 279                                                req->rq_set == NULL);
 280         } else if (req->rq_set) {
 281                 /*
 282                  * If we have a vaid "rq_set", just reuse it to avoid double
 283                  * linked.
 284                  */
 285                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 286                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 287
 288                 /* ptlrpc_check_set will decrease the count */
 289                 atomic_inc(&req->rq_set->set_remaining);
 290                 spin_unlock(&req->rq_lock);
 291                 wake_up(&req->rq_set->set_waitq);
 292                 return;
 293         } else {
 294                 spin_unlock(&req->rq_lock);
 295         }
 296
 297         pc = ptlrpcd_select_pc(req);
 298
 299         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s+%d]",
 300                   req, pc->pc_name, pc->pc_index);
 301
 302         ptlrpc_set_add_new_req(pc, req);
 303 }
 304 EXPORT_SYMBOL(ptlrpcd_add_req);
 305
 306 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 307 {
 308         atomic_inc(&set->set_refcount);
 309 }
 310
 311 /**
 312  * Check if there is more work to do on ptlrpcd set.
 313  * Returns 1 if yes.
 314  */
 315 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 316 {
 317         struct ptlrpc_request *req, *tmp;
 318         struct ptlrpc_request_set *set = pc->pc_set;
 319         int rc = 0;
 320         int rc2;
 321
 322         ENTRY;
 323
 324         if (atomic_read(&set->set_new_count)) {
 325                 spin_lock(&set->set_new_req_lock);
 326                 if (likely(!list_empty(&set->set_new_requests))) {
 327                         list_splice_init(&set->set_new_requests,
 328                                              &set->set_requests);
 329                         atomic_add(atomic_read(&set->set_new_count),
 330                                    &set->set_remaining);
 331                         atomic_set(&set->set_new_count, 0);
 332                         /*
 333                          * Need to calculate its timeout.
 334                          */
 335                         rc = 1;
 336                 }
 337                 spin_unlock(&set->set_new_req_lock);
 338         }
 339
 340         /*
 341          * We should call lu_env_refill() before handling new requests to make
 342          * sure that env key the requests depending on really exists.
 343          */
 344         rc2 = lu_env_refill(env);
 345         if (rc2 != 0) {
 346                 /*
 347                  * XXX This is very awkward situation, because
 348                  * execution can neither continue (request
 349                  * interpreters assume that env is set up), nor repeat
 350                  * the loop (as this potentially results in a tight
 351                  * loop of -ENOMEM's).
 352                  *
 353                  * Fortunately, refill only ever does something when
 354                  * new modules are loaded, i.e., early during boot up.
 355                  */
 356                 CERROR("Failure to refill session: %d\n", rc2);
 357                 RETURN(rc);
 358         }
 359
 360         if (atomic_read(&set->set_remaining))
 361                 rc |= ptlrpc_check_set(env, set);
 362
 363         /*
 364          * NB: ptlrpc_check_set has already moved complted request at the
 365          * head of seq::set_requests
 366          */
 367         list_for_each_entry_safe(req, tmp, &set->set_requests, rq_set_chain) {
 368                 if (req->rq_phase != RQ_PHASE_COMPLETE)
 369                         break;
 370
 371                 list_del_init(&req->rq_set_chain);
 372                 req->rq_set = NULL;
 373                 ptlrpc_req_finished(req);
 374         }
 375
 376         if (rc == 0) {
 377                 /*
 378                  * If new requests have been added, make sure to wake up.
 379                  */
 380                 rc = atomic_read(&set->set_new_count);
 381
 382                 /*
 383                  * If we have nothing to do, check whether we can take some
 384                  * work from our partner threads.
 385                  */
 386                 if (rc == 0 && pc->pc_npartners > 0) {
 387                         struct ptlrpcd_ctl *partner;
 388                         struct ptlrpc_request_set *ps;
 389                         int first = pc->pc_cursor;
 390
 391                         do {
 392                                 partner = pc->pc_partners[pc->pc_cursor++];
 393                                 if (pc->pc_cursor >= pc->pc_npartners)
 394                                         pc->pc_cursor = 0;
 395                                 if (partner == NULL)
 396                                         continue;
 397
 398                                 spin_lock(&partner->pc_lock);
 399                                 ps = partner->pc_set;
 400                                 if (ps == NULL) {
 401                                         spin_unlock(&partner->pc_lock);
 402                                         continue;
 403                                 }
 404
 405                                 ptlrpc_reqset_get(ps);
 406                                 spin_unlock(&partner->pc_lock);
 407
 408                                 if (atomic_read(&ps->set_new_count)) {
 409                                         rc = ptlrpcd_steal_rqset(set, ps);
 410                                         if (rc > 0)
 411                                                 CDEBUG(D_RPCTRACE,
 412                                                        "transfer %d async RPCs [%d->%d]\n",
 413                                                        rc, partner->pc_index,
 414                                                        pc->pc_index);
 415                                 }
 416                                 ptlrpc_reqset_put(ps);
 417                         } while (rc == 0 && pc->pc_cursor != first);
 418                 }
 419         }
 420
 421         RETURN(rc || test_bit(LIOD_STOP, &pc->pc_flags));
 422 }
 423
 424 /**
 425  * Main ptlrpcd thread.
 426  * ptlrpc's code paths like to execute in process context, so we have this
 427  * thread which spins on a set which contains the rpcs and sends them.
 428  */
 429 static int ptlrpcd(void *arg)
 430 {
 431         struct ptlrpcd_ctl              *pc = arg;
 432         struct ptlrpc_request_set       *set;
 433         struct lu_context               ses = { 0 };
 434         struct lu_env                   env = { .le_ses = &ses };
 435         int                             rc = 0;
 436         int                             exit = 0;
 437
 438         ENTRY;
 439         if (cfs_cpt_bind(cfs_cpt_tab, pc->pc_cpt) != 0)
 440                 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
 441
 442         /*
 443          * Allocate the request set after the thread has been bound
 444          * above. This is safe because no requests will be queued
 445          * until all ptlrpcd threads have confirmed that they have
 446          * successfully started.
 447          */
 448         set = ptlrpc_prep_set();
 449         if (set == NULL)
 450                 GOTO(failed, rc = -ENOMEM);
 451         spin_lock(&pc->pc_lock);
 452         pc->pc_set = set;
 453         spin_unlock(&pc->pc_lock);
 454
 455         /* Both client and server (MDT/OST) may use the environment. */
 456         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD |
 457                                           LCT_DT_THREAD |
 458                                           LCT_CL_THREAD |
 459                                           LCT_REMEMBER  |
 460                                           LCT_NOREF);
 461         if (rc != 0)
 462                 GOTO(failed, rc);
 463         rc = lu_context_init(env.le_ses, LCT_SESSION  |
 464                                          LCT_REMEMBER |
 465                                          LCT_NOREF);
 466         if (rc != 0) {
 467                 lu_context_fini(&env.le_ctx);
 468                 GOTO(failed, rc);
 469         }
 470
 471         complete(&pc->pc_starting);
 472
 473         /*
 474          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 475          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 476          * there are requests in the set. New requests come in on the set's
 477          * new_req_list and ptlrpcd_check() moves them into the set.
 478          */
 479         do {
 480                 time64_t timeout;
 481
 482                 timeout = ptlrpc_set_next_timeout(set);
 483
 484                 lu_context_enter(&env.le_ctx);
 485                 lu_context_enter(env.le_ses);
 486                 if (timeout == 0)
 487                         wait_event_idle(set->set_waitq,
 488                                         ptlrpcd_check(&env, pc));
 489                 else if (wait_event_idle_timeout(set->set_waitq,
 490                                                  ptlrpcd_check(&env, pc),
 491                                                  cfs_time_seconds(timeout))
 492                          == 0)
 493                         ptlrpc_expired_set(set);
 494                 lu_context_exit(&env.le_ctx);
 495                 lu_context_exit(env.le_ses);
 496
 497                 /*
 498                  * Abort inflight rpcs for forced stop case.
 499                  */
 500                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 501                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 502                                 ptlrpc_abort_set(set);
 503                         exit++;
 504                 }
 505
 506                 /*
 507                  * Let's make one more loop to make sure that ptlrpcd_check()
 508                  * copied all raced new rpcs into the set so we can kill them.
 509                  */
 510         } while (exit < 2);
 511
 512         /*
 513          * Wait for inflight requests to drain.
 514          */
 515         if (!list_empty(&set->set_requests))
 516                 ptlrpc_set_wait(&env, set);
 517         lu_context_fini(&env.le_ctx);
 518         lu_context_fini(env.le_ses);
 519
 520         complete(&pc->pc_finishing);
 521
 522         return 0;
 523
 524 failed:
 525         pc->pc_error = rc;
 526         complete(&pc->pc_starting);
 527         RETURN(rc);
 528 }
 529
 530 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
 531 {
 532         ENTRY;
 533
 534         pc->pc_index = index;
 535         pc->pc_cpt = cpt;
 536         init_completion(&pc->pc_starting);
 537         init_completion(&pc->pc_finishing);
 538         spin_lock_init(&pc->pc_lock);
 539
 540         if (index < 0) {
 541                 /* Recovery thread. */
 542                 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
 543         } else {
 544                 /* Regular thread. */
 545                 snprintf(pc->pc_name, sizeof(pc->pc_name),
 546                          "ptlrpcd_%02d_%02d", cpt, index);
 547         }
 548
 549         EXIT;
 550 }
 551
 552 /* XXX: We want multiple CPU cores to share the async RPC load. So we
 553  *      start many ptlrpcd threads. We also want to reduce the ptlrpcd
 554  *      overhead caused by data transfer cross-CPU cores. So we bind
 555  *      all ptlrpcd threads to a CPT, in the expectation that CPTs
 556  *      will be defined in a way that matches these boundaries. Within
 557  *      a CPT a ptlrpcd thread can be scheduled on any available core.
 558  *
 559  *      Each ptlrpcd thread has its own request queue. This can cause
 560  *      response delay if the thread is already busy. To help with
 561  *      this we define partner threads: these are other threads bound
 562  *      to the same CPT which will check for work in each other's
 563  *      request queues if they have no work to do.
 564  *
 565  *      The desired number of partner threads can be tuned by setting
 566  *      ptlrpcd_partner_group_size. The default is to create pairs of
 567  *      partner threads.
 568  */
 569 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
 570 {
 571         struct ptlrpcd_ctl      *pc;
 572         struct ptlrpcd_ctl      **ppc;
 573         int                     first;
 574         int                     i;
 575         int                     rc = 0;
 576
 577         ENTRY;
 578
 579         LASSERT(index >= 0 && index < pd->pd_nthreads);
 580         pc = &pd->pd_threads[index];
 581         pc->pc_npartners = pd->pd_groupsize - 1;
 582
 583         if (pc->pc_npartners <= 0)
 584                 GOTO(out, rc);
 585
 586         OBD_CPT_ALLOC(pc->pc_partners, cfs_cpt_tab, pc->pc_cpt,
 587                       sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 588         if (pc->pc_partners == NULL) {
 589                 pc->pc_npartners = 0;
 590                 GOTO(out, rc = -ENOMEM);
 591         }
 592
 593         first = index - index % pd->pd_groupsize;
 594         ppc = pc->pc_partners;
 595         for (i = first; i < first + pd->pd_groupsize; i++) {
 596                 if (i != index)
 597                         *ppc++ = &pd->pd_threads[i];
 598         }
 599 out:
 600         RETURN(rc);
 601 }
 602
 603 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
 604 {
 605         struct task_struct      *task;
 606         int                     rc = 0;
 607
 608         ENTRY;
 609
 610         /*
 611          * Do not allow starting a second thread for one pc.
 612          */
 613         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 614                 CWARN("Starting second thread (%s) for same pc %p\n",
 615                       pc->pc_name, pc);
 616                 RETURN(0);
 617         }
 618
 619         task = kthread_run(ptlrpcd, pc, "%s", pc->pc_name);
 620         if (IS_ERR(task))
 621                 GOTO(out_set, rc = PTR_ERR(task));
 622
 623         wait_for_completion(&pc->pc_starting);
 624         rc = pc->pc_error;
 625         if (rc != 0)
 626                 GOTO(out_set, rc);
 627
 628         RETURN(0);
 629
 630 out_set:
 631         if (pc->pc_set != NULL) {
 632                 struct ptlrpc_request_set *set = pc->pc_set;
 633
 634                 spin_lock(&pc->pc_lock);
 635                 pc->pc_set = NULL;
 636                 spin_unlock(&pc->pc_lock);
 637                 ptlrpc_set_destroy(set);
 638         }
 639         clear_bit(LIOD_START, &pc->pc_flags);
 640         RETURN(rc);
 641 }
 642
 643 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 644 {
 645         ENTRY;
 646
 647         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 648                 CWARN("Thread for pc %p was not started\n", pc);
 649                 goto out;
 650         }
 651
 652         set_bit(LIOD_STOP, &pc->pc_flags);
 653         if (force)
 654                 set_bit(LIOD_FORCE, &pc->pc_flags);
 655         wake_up(&pc->pc_set->set_waitq);
 656
 657 out:
 658         EXIT;
 659 }
 660
 661 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 662 {
 663         struct ptlrpc_request_set *set = pc->pc_set;
 664
 665         ENTRY;
 666
 667         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 668                 CWARN("Thread for pc %p was not started\n", pc);
 669                 goto out;
 670         }
 671
 672         wait_for_completion(&pc->pc_finishing);
 673
 674         spin_lock(&pc->pc_lock);
 675         pc->pc_set = NULL;
 676         spin_unlock(&pc->pc_lock);
 677         ptlrpc_set_destroy(set);
 678
 679         clear_bit(LIOD_START, &pc->pc_flags);
 680         clear_bit(LIOD_STOP, &pc->pc_flags);
 681         clear_bit(LIOD_FORCE, &pc->pc_flags);
 682
 683 out:
 684         if (pc->pc_npartners > 0) {
 685                 LASSERT(pc->pc_partners != NULL);
 686
 687                 OBD_FREE_PTR_ARRAY(pc->pc_partners, pc->pc_npartners);
 688                 pc->pc_partners = NULL;
 689         }
 690         pc->pc_npartners = 0;
 691         pc->pc_error = 0;
 692         EXIT;
 693 }
 694
 695 static void ptlrpcd_fini(void)
 696 {
 697         int     i;
 698         int     j;
 699         int     ncpts;
 700
 701         ENTRY;
 702
 703         if (ptlrpcds != NULL) {
 704                 for (i = 0; i < ptlrpcds_num; i++) {
 705                         if (ptlrpcds[i] == NULL)
 706                                 break;
 707                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 708                                 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
 709                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 710                                 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
 711                         OBD_FREE(ptlrpcds[i], ptlrpcds[i]->pd_size);
 712                         ptlrpcds[i] = NULL;
 713                 }
 714                 OBD_FREE_PTR_ARRAY(ptlrpcds, ptlrpcds_num);
 715         }
 716         ptlrpcds_num = 0;
 717
 718         ptlrpcd_stop(&ptlrpcd_rcv, 0);
 719         ptlrpcd_free(&ptlrpcd_rcv);
 720
 721         if (ptlrpcds_cpt_idx != NULL) {
 722                 ncpts = cfs_cpt_number(cfs_cpt_tab);
 723                 OBD_FREE_PTR_ARRAY(ptlrpcds_cpt_idx, ncpts);
 724                 ptlrpcds_cpt_idx = NULL;
 725         }
 726
 727         EXIT;
 728 }
 729
 730 static int ptlrpcd_init(void)
 731 {
 732         int                     nthreads;
 733         int                     groupsize;
 734         int                     size;
 735         int                     i;
 736         int                     j;
 737         int                     rc = 0;
 738         struct cfs_cpt_table    *cptable;
 739         __u32                   *cpts = NULL;
 740         int                     ncpts;
 741         int                     cpt;
 742         struct ptlrpcd          *pd;
 743
 744         ENTRY;
 745
 746         /*
 747          * Determine the CPTs that ptlrpcd threads will run on.
 748          */
 749         cptable = cfs_cpt_tab;
 750         ncpts = cfs_cpt_number(cptable);
 751         if (ptlrpcd_cpts != NULL) {
 752                 struct cfs_expr_list *el;
 753
 754                 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
 755                 OBD_ALLOC(ptlrpcds_cpt_idx, size);
 756                 if (ptlrpcds_cpt_idx == NULL)
 757                         GOTO(out, rc = -ENOMEM);
 758
 759                 rc = cfs_expr_list_parse(ptlrpcd_cpts,
 760                                          strlen(ptlrpcd_cpts),
 761                                          0, ncpts - 1, &el);
 762                 if (rc != 0) {
 763                         CERROR("%s: invalid CPT pattern string: %s",
 764                                "ptlrpcd_cpts", ptlrpcd_cpts);
 765                         GOTO(out, rc = -EINVAL);
 766                 }
 767
 768                 rc = cfs_expr_list_values(el, ncpts, &cpts);
 769                 cfs_expr_list_free(el);
 770                 if (rc <= 0) {
 771                         CERROR("%s: failed to parse CPT array %s: %d\n",
 772                                "ptlrpcd_cpts", ptlrpcd_cpts, rc);
 773                         if (rc == 0)
 774                                 rc = -EINVAL;
 775                         GOTO(out, rc);
 776                 }
 777
 778                 /*
 779                  * Create the cpt-to-index map. When there is no match
 780                  * in the cpt table, pick a cpt at random. This could
 781                  * be changed to take the topology of the system into
 782                  * account.
 783                  */
 784                 for (cpt = 0; cpt < ncpts; cpt++) {
 785                         for (i = 0; i < rc; i++)
 786                                 if (cpts[i] == cpt)
 787                                         break;
 788                         if (i >= rc)
 789                                 i = cpt % rc;
 790                         ptlrpcds_cpt_idx[cpt] = i;
 791                 }
 792
 793                 cfs_expr_list_values_free(cpts, rc);
 794                 ncpts = rc;
 795         }
 796         ptlrpcds_num = ncpts;
 797
 798         size = ncpts * sizeof(ptlrpcds[0]);
 799         OBD_ALLOC(ptlrpcds, size);
 800         if (ptlrpcds == NULL)
 801                 GOTO(out, rc = -ENOMEM);
 802
 803         /*
 804          * The max_ptlrpcds parameter is obsolete, but do something
 805          * sane if it has been tuned, and complain if
 806          * ptlrpcd_per_cpt_max has also been tuned.
 807          */
 808         if (max_ptlrpcds != 0) {
 809                 CWARN("max_ptlrpcds is obsolete.\n");
 810                 if (ptlrpcd_per_cpt_max == 0) {
 811                         ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
 812                         /* Round up if there is a remainder. */
 813                         if (max_ptlrpcds % ncpts != 0)
 814                                 ptlrpcd_per_cpt_max++;
 815                         CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
 816                               ptlrpcd_per_cpt_max);
 817                 } else {
 818                         CWARN("ptlrpd_per_cpt_max is also set!\n");
 819                 }
 820         }
 821
 822         /*
 823          * The ptlrpcd_bind_policy parameter is obsolete, but do
 824          * something sane if it has been tuned, and complain if
 825          * ptlrpcd_partner_group_size is also tuned.
 826          */
 827         if (ptlrpcd_bind_policy != 0) {
 828                 CWARN("ptlrpcd_bind_policy is obsolete.\n");
 829                 if (ptlrpcd_partner_group_size == 0) {
 830                         switch (ptlrpcd_bind_policy) {
 831                         case 1: /* PDB_POLICY_NONE */
 832                         case 2: /* PDB_POLICY_FULL */
 833                                 ptlrpcd_partner_group_size = 1;
 834                                 break;
 835                         case 3: /* PDB_POLICY_PAIR */
 836                                 ptlrpcd_partner_group_size = 2;
 837                                 break;
 838                         case 4: /* PDB_POLICY_NEIGHBOR */
 839 #ifdef CONFIG_NUMA
 840                                 ptlrpcd_partner_group_size = -1; /* CPT */
 841 #else
 842                                 ptlrpcd_partner_group_size = 3; /* Triplets */
 843 #endif
 844                                 break;
 845                         default: /* Illegal value, use the default. */
 846                                 ptlrpcd_partner_group_size = 2;
 847                                 break;
 848                         }
 849                         CWARN("Setting ptlrpcd_partner_group_size = %d\n",
 850                               ptlrpcd_partner_group_size);
 851                 } else {
 852                         CWARN("ptlrpcd_partner_group_size is also set!\n");
 853                 }
 854         }
 855
 856         if (ptlrpcd_partner_group_size == 0)
 857                 ptlrpcd_partner_group_size = 2;
 858         else if (ptlrpcd_partner_group_size < 0)
 859                 ptlrpcd_partner_group_size = -1;
 860         else if (ptlrpcd_per_cpt_max > 0 &&
 861                  ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
 862                 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
 863
 864         /*
 865          * Start the recovery thread first.
 866          */
 867         set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
 868         ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
 869         rc = ptlrpcd_start(&ptlrpcd_rcv);
 870         if (rc < 0)
 871                 GOTO(out, rc);
 872
 873         for (i = 0; i < ncpts; i++) {
 874                 if (cpts == NULL)
 875                         cpt = i;
 876                 else
 877                         cpt = cpts[i];
 878
 879                 nthreads = cfs_cpt_weight(cptable, cpt);
 880                 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
 881                         nthreads = ptlrpcd_per_cpt_max;
 882                 if (nthreads < 2)
 883                         nthreads = 2;
 884
 885                 if (ptlrpcd_partner_group_size <= 0) {
 886                         groupsize = nthreads;
 887                 } else if (nthreads <= ptlrpcd_partner_group_size) {
 888                         groupsize = nthreads;
 889                 } else {
 890                         groupsize = ptlrpcd_partner_group_size;
 891                         if (nthreads % groupsize != 0)
 892                                 nthreads += groupsize - (nthreads % groupsize);
 893                 }
 894
 895                 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 896                 OBD_CPT_ALLOC(pd, cptable, cpt, size);
 897
 898                 if (!pd)
 899                         GOTO(out, rc = -ENOMEM);
 900                 pd->pd_size      = size;
 901                 pd->pd_index     = i;
 902                 pd->pd_cpt       = cpt;
 903                 pd->pd_cursor    = 0;
 904                 pd->pd_nthreads  = nthreads;
 905                 pd->pd_groupsize = groupsize;
 906                 ptlrpcds[i] = pd;
 907
 908                 /*
 909                  * The ptlrpcd threads in a partner group can access
 910                  * each other's struct ptlrpcd_ctl, so these must be
 911                  * initialized before any thead is started.
 912                  */
 913                 for (j = 0; j < nthreads; j++) {
 914                         ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
 915                         rc = ptlrpcd_partners(pd, j);
 916                         if (rc < 0)
 917                                 GOTO(out, rc);
 918                 }
 919
 920                 /* XXX: We start nthreads ptlrpc daemons on this cpt.
 921                  *      Each of them can process any non-recovery
 922                  *      async RPC to improve overall async RPC
 923                  *      efficiency.
 924                  *
 925                  *      But there are some issues with async I/O RPCs
 926                  *      and async non-I/O RPCs processed in the same
 927                  *      set under some cases. The ptlrpcd may be
 928                  *      blocked by some async I/O RPC(s), then will
 929                  *      cause other async non-I/O RPC(s) can not be
 930                  *      processed in time.
 931                  *
 932                  *      Maybe we should distinguish blocked async RPCs
 933                  *      from non-blocked async RPCs, and process them
 934                  *      in different ptlrpcd sets to avoid unnecessary
 935                  *      dependency. But how to distribute async RPCs
 936                  *      load among all the ptlrpc daemons becomes
 937                  *      another trouble.
 938                  */
 939                 for (j = 0; j < nthreads; j++) {
 940                         rc = ptlrpcd_start(&pd->pd_threads[j]);
 941                         if (rc < 0)
 942                                 GOTO(out, rc);
 943                 }
 944         }
 945 out:
 946         if (rc != 0)
 947                 ptlrpcd_fini();
 948
 949         RETURN(rc);
 950 }
 951
 952 int ptlrpcd_addref(void)
 953 {
 954         int rc = 0;
 955
 956         ENTRY;
 957
 958         mutex_lock(&ptlrpcd_mutex);
 959         if (++ptlrpcd_users == 1) {
 960                 rc = ptlrpcd_init();
 961                 if (rc < 0)
 962                         ptlrpcd_users--;
 963         }
 964         mutex_unlock(&ptlrpcd_mutex);
 965         RETURN(rc);
 966 }
 967 EXPORT_SYMBOL(ptlrpcd_addref);
 968
 969 void ptlrpcd_decref(void)
 970 {
 971         mutex_lock(&ptlrpcd_mutex);
 972         if (--ptlrpcd_users == 0)
 973                 ptlrpcd_fini();
 974         mutex_unlock(&ptlrpcd_mutex);
 975 }
 976 EXPORT_SYMBOL(ptlrpcd_decref);
 977 /** @} ptlrpcd */