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  78  * max_ptlrpcds is obsolete, but retained to ensure that the kernel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  89  * ptlrpcd_bind_policy is obsolete, but retained to ensure that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