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/fs_struct.h>
  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         unshare_fs_struct();
 441         if (cfs_cpt_bind(cfs_cpt_tab, pc->pc_cpt) != 0)
 442                 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
 443
 444         /*
 445          * Allocate the request set after the thread has been bound
 446          * above. This is safe because no requests will be queued
 447          * until all ptlrpcd threads have confirmed that they have
 448          * successfully started.
 449          */
 450         set = ptlrpc_prep_set();
 451         if (set == NULL)
 452                 GOTO(failed, rc = -ENOMEM);
 453         spin_lock(&pc->pc_lock);
 454         pc->pc_set = set;
 455         spin_unlock(&pc->pc_lock);
 456
 457         /* Both client and server (MDT/OST) may use the environment. */
 458         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD |
 459                                           LCT_DT_THREAD |
 460                                           LCT_CL_THREAD |
 461                                           LCT_REMEMBER  |
 462                                           LCT_NOREF);
 463         if (rc != 0)
 464                 GOTO(failed, rc);
 465         rc = lu_context_init(env.le_ses, LCT_SESSION  |
 466                                          LCT_REMEMBER |
 467                                          LCT_NOREF);
 468         if (rc != 0) {
 469                 lu_context_fini(&env.le_ctx);
 470                 GOTO(failed, rc);
 471         }
 472
 473         complete(&pc->pc_starting);
 474
 475         /*
 476          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 477          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 478          * there are requests in the set. New requests come in on the set's
 479          * new_req_list and ptlrpcd_check() moves them into the set.
 480          */
 481         do {
 482                 DEFINE_WAIT_FUNC(wait, woken_wake_function);
 483                 time64_t timeout;
 484
 485                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(set));
 486
 487                 lu_context_enter(&env.le_ctx);
 488                 lu_context_enter(env.le_ses);
 489
 490                 add_wait_queue(&set->set_waitq, &wait);
 491                 while (!ptlrpcd_check(&env, pc)) {
 492                         int ret;
 493
 494                         if (timeout == 0)
 495                                 ret = wait_woken(&wait, TASK_IDLE,
 496                                                  MAX_SCHEDULE_TIMEOUT);
 497                         else {
 498                                 ret = wait_woken(&wait, TASK_IDLE, timeout);
 499                                 if (ret > 0)
 500                                         timeout = ret;
 501                         }
 502                         if (ret != 0)
 503                                 continue;
 504                         /* Timed out */
 505                         ptlrpc_expired_set(set);
 506                         break;
 507                 }
 508                 remove_wait_queue(&set->set_waitq, &wait);
 509
 510                 lu_context_exit(&env.le_ctx);
 511                 lu_context_exit(env.le_ses);
 512
 513                 /*
 514                  * Abort inflight rpcs for forced stop case.
 515                  */
 516                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 517                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 518                                 ptlrpc_abort_set(set);
 519                         exit++;
 520                 }
 521
 522                 /*
 523                  * Let's make one more loop to make sure that ptlrpcd_check()
 524                  * copied all raced new rpcs into the set so we can kill them.
 525                  */
 526         } while (exit < 2);
 527
 528         /*
 529          * Wait for inflight requests to drain.
 530          */
 531         if (!list_empty(&set->set_requests))
 532                 ptlrpc_set_wait(&env, set);
 533         lu_context_fini(&env.le_ctx);
 534         lu_context_fini(env.le_ses);
 535
 536         complete(&pc->pc_finishing);
 537
 538         return 0;
 539
 540 failed:
 541         pc->pc_error = rc;
 542         complete(&pc->pc_starting);
 543         RETURN(rc);
 544 }
 545
 546 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
 547 {
 548         ENTRY;
 549
 550         pc->pc_index = index;
 551         pc->pc_cpt = cpt;
 552         init_completion(&pc->pc_starting);
 553         init_completion(&pc->pc_finishing);
 554         spin_lock_init(&pc->pc_lock);
 555
 556         if (index < 0) {
 557                 /* Recovery thread. */
 558                 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
 559         } else {
 560                 /* Regular thread. */
 561                 snprintf(pc->pc_name, sizeof(pc->pc_name),
 562                          "ptlrpcd_%02d_%02d", cpt, index);
 563         }
 564
 565         EXIT;
 566 }
 567
 568 /* XXX: We want multiple CPU cores to share the async RPC load. So we
 569  *      start many ptlrpcd threads. We also want to reduce the ptlrpcd
 570  *      overhead caused by data transfer cross-CPU cores. So we bind
 571  *      all ptlrpcd threads to a CPT, in the expectation that CPTs
 572  *      will be defined in a way that matches these boundaries. Within
 573  *      a CPT a ptlrpcd thread can be scheduled on any available core.
 574  *
 575  *      Each ptlrpcd thread has its own request queue. This can cause
 576  *      response delay if the thread is already busy. To help with
 577  *      this we define partner threads: these are other threads bound
 578  *      to the same CPT which will check for work in each other's
 579  *      request queues if they have no work to do.
 580  *
 581  *      The desired number of partner threads can be tuned by setting
 582  *      ptlrpcd_partner_group_size. The default is to create pairs of
 583  *      partner threads.
 584  */
 585 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
 586 {
 587         struct ptlrpcd_ctl      *pc;
 588         struct ptlrpcd_ctl      **ppc;
 589         int                     first;
 590         int                     i;
 591         int                     rc = 0;
 592
 593         ENTRY;
 594
 595         LASSERT(index >= 0 && index < pd->pd_nthreads);
 596         pc = &pd->pd_threads[index];
 597         pc->pc_npartners = pd->pd_groupsize - 1;
 598
 599         if (pc->pc_npartners <= 0)
 600                 GOTO(out, rc);
 601
 602         OBD_CPT_ALLOC(pc->pc_partners, cfs_cpt_tab, pc->pc_cpt,
 603                       sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 604         if (pc->pc_partners == NULL) {
 605                 pc->pc_npartners = 0;
 606                 GOTO(out, rc = -ENOMEM);
 607         }
 608
 609         first = index - index % pd->pd_groupsize;
 610         ppc = pc->pc_partners;
 611         for (i = first; i < first + pd->pd_groupsize; i++) {
 612                 if (i != index)
 613                         *ppc++ = &pd->pd_threads[i];
 614         }
 615 out:
 616         RETURN(rc);
 617 }
 618
 619 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
 620 {
 621         struct task_struct      *task;
 622         int                     rc = 0;
 623
 624         ENTRY;
 625
 626         /*
 627          * Do not allow starting a second thread for one pc.
 628          */
 629         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 630                 CWARN("Starting second thread (%s) for same pc %p\n",
 631                       pc->pc_name, pc);
 632                 RETURN(0);
 633         }
 634
 635         task = kthread_run(ptlrpcd, pc, "%s", pc->pc_name);
 636         if (IS_ERR(task))
 637                 GOTO(out_set, rc = PTR_ERR(task));
 638
 639         wait_for_completion(&pc->pc_starting);
 640         rc = pc->pc_error;
 641         if (rc != 0)
 642                 GOTO(out_set, rc);
 643
 644         RETURN(0);
 645
 646 out_set:
 647         if (pc->pc_set != NULL) {
 648                 struct ptlrpc_request_set *set = pc->pc_set;
 649
 650                 spin_lock(&pc->pc_lock);
 651                 pc->pc_set = NULL;
 652                 spin_unlock(&pc->pc_lock);
 653                 ptlrpc_set_destroy(set);
 654         }
 655         clear_bit(LIOD_START, &pc->pc_flags);
 656         RETURN(rc);
 657 }
 658
 659 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 660 {
 661         ENTRY;
 662
 663         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 664                 CWARN("Thread for pc %p was not started\n", pc);
 665                 goto out;
 666         }
 667
 668         set_bit(LIOD_STOP, &pc->pc_flags);
 669         if (force)
 670                 set_bit(LIOD_FORCE, &pc->pc_flags);
 671         wake_up(&pc->pc_set->set_waitq);
 672
 673 out:
 674         EXIT;
 675 }
 676
 677 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 678 {
 679         struct ptlrpc_request_set *set = pc->pc_set;
 680
 681         ENTRY;
 682
 683         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 684                 CWARN("Thread for pc %p was not started\n", pc);
 685                 goto out;
 686         }
 687
 688         wait_for_completion(&pc->pc_finishing);
 689
 690         spin_lock(&pc->pc_lock);
 691         pc->pc_set = NULL;
 692         spin_unlock(&pc->pc_lock);
 693         ptlrpc_set_destroy(set);
 694
 695         clear_bit(LIOD_START, &pc->pc_flags);
 696         clear_bit(LIOD_STOP, &pc->pc_flags);
 697         clear_bit(LIOD_FORCE, &pc->pc_flags);
 698
 699 out:
 700         if (pc->pc_npartners > 0) {
 701                 LASSERT(pc->pc_partners != NULL);
 702
 703                 OBD_FREE_PTR_ARRAY(pc->pc_partners, pc->pc_npartners);
 704                 pc->pc_partners = NULL;
 705         }
 706         pc->pc_npartners = 0;
 707         pc->pc_error = 0;
 708         EXIT;
 709 }
 710
 711 static void ptlrpcd_fini(void)
 712 {
 713         int     i;
 714         int     j;
 715         int     ncpts;
 716
 717         ENTRY;
 718
 719         if (ptlrpcds != NULL) {
 720                 for (i = 0; i < ptlrpcds_num; i++) {
 721                         if (ptlrpcds[i] == NULL)
 722                                 break;
 723                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 724                                 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
 725                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 726                                 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
 727                         OBD_FREE(ptlrpcds[i], ptlrpcds[i]->pd_size);
 728                         ptlrpcds[i] = NULL;
 729                 }
 730                 OBD_FREE_PTR_ARRAY(ptlrpcds, ptlrpcds_num);
 731         }
 732         ptlrpcds_num = 0;
 733
 734         ptlrpcd_stop(&ptlrpcd_rcv, 0);
 735         ptlrpcd_free(&ptlrpcd_rcv);
 736
 737         if (ptlrpcds_cpt_idx != NULL) {
 738                 ncpts = cfs_cpt_number(cfs_cpt_tab);
 739                 OBD_FREE_PTR_ARRAY(ptlrpcds_cpt_idx, ncpts);
 740                 ptlrpcds_cpt_idx = NULL;
 741         }
 742
 743         EXIT;
 744 }
 745
 746 static int ptlrpcd_init(void)
 747 {
 748         int                     nthreads;
 749         int                     groupsize;
 750         int                     size;
 751         int                     i;
 752         int                     j;
 753         int                     rc = 0;
 754         struct cfs_cpt_table    *cptable;
 755         __u32                   *cpts = NULL;
 756         int                     ncpts;
 757         int                     cpt;
 758         struct ptlrpcd          *pd;
 759
 760         ENTRY;
 761
 762         /*
 763          * Determine the CPTs that ptlrpcd threads will run on.
 764          */
 765         cptable = cfs_cpt_tab;
 766         ncpts = cfs_cpt_number(cptable);
 767         if (ptlrpcd_cpts != NULL) {
 768                 struct cfs_expr_list *el;
 769
 770                 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
 771                 OBD_ALLOC(ptlrpcds_cpt_idx, size);
 772                 if (ptlrpcds_cpt_idx == NULL)
 773                         GOTO(out, rc = -ENOMEM);
 774
 775                 rc = cfs_expr_list_parse(ptlrpcd_cpts,
 776                                          strlen(ptlrpcd_cpts),
 777                                          0, ncpts - 1, &el);
 778                 if (rc != 0) {
 779                         CERROR("%s: invalid CPT pattern string: %s",
 780                                "ptlrpcd_cpts", ptlrpcd_cpts);
 781                         GOTO(out, rc = -EINVAL);
 782                 }
 783
 784                 rc = cfs_expr_list_values(el, ncpts, &cpts);
 785                 cfs_expr_list_free(el);
 786                 if (rc <= 0) {
 787                         CERROR("%s: failed to parse CPT array %s: %d\n",
 788                                "ptlrpcd_cpts", ptlrpcd_cpts, rc);
 789                         if (rc == 0)
 790                                 rc = -EINVAL;
 791                         GOTO(out, rc);
 792                 }
 793
 794                 /*
 795                  * Create the cpt-to-index map. When there is no match
 796                  * in the cpt table, pick a cpt at random. This could
 797                  * be changed to take the topology of the system into
 798                  * account.
 799                  */
 800                 for (cpt = 0; cpt < ncpts; cpt++) {
 801                         for (i = 0; i < rc; i++)
 802                                 if (cpts[i] == cpt)
 803                                         break;
 804                         if (i >= rc)
 805                                 i = cpt % rc;
 806                         ptlrpcds_cpt_idx[cpt] = i;
 807                 }
 808
 809                 cfs_expr_list_values_free(cpts, rc);
 810                 ncpts = rc;
 811         }
 812         ptlrpcds_num = ncpts;
 813
 814         size = ncpts * sizeof(ptlrpcds[0]);
 815         OBD_ALLOC(ptlrpcds, size);
 816         if (ptlrpcds == NULL)
 817                 GOTO(out, rc = -ENOMEM);
 818
 819         /*
 820          * The max_ptlrpcds parameter is obsolete, but do something
 821          * sane if it has been tuned, and complain if
 822          * ptlrpcd_per_cpt_max has also been tuned.
 823          */
 824         if (max_ptlrpcds != 0) {
 825                 CWARN("max_ptlrpcds is obsolete.\n");
 826                 if (ptlrpcd_per_cpt_max == 0) {
 827                         ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
 828                         /* Round up if there is a remainder. */
 829                         if (max_ptlrpcds % ncpts != 0)
 830                                 ptlrpcd_per_cpt_max++;
 831                         CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
 832                               ptlrpcd_per_cpt_max);
 833                 } else {
 834                         CWARN("ptlrpd_per_cpt_max is also set!\n");
 835                 }
 836         }
 837
 838         /*
 839          * The ptlrpcd_bind_policy parameter is obsolete, but do
 840          * something sane if it has been tuned, and complain if
 841          * ptlrpcd_partner_group_size is also tuned.
 842          */
 843         if (ptlrpcd_bind_policy != 0) {
 844                 CWARN("ptlrpcd_bind_policy is obsolete.\n");
 845                 if (ptlrpcd_partner_group_size == 0) {
 846                         switch (ptlrpcd_bind_policy) {
 847                         case 1: /* PDB_POLICY_NONE */
 848                         case 2: /* PDB_POLICY_FULL */
 849                                 ptlrpcd_partner_group_size = 1;
 850                                 break;
 851                         case 3: /* PDB_POLICY_PAIR */
 852                                 ptlrpcd_partner_group_size = 2;
 853                                 break;
 854                         case 4: /* PDB_POLICY_NEIGHBOR */
 855 #ifdef CONFIG_NUMA
 856                                 ptlrpcd_partner_group_size = -1; /* CPT */
 857 #else
 858                                 ptlrpcd_partner_group_size = 3; /* Triplets */
 859 #endif
 860                                 break;
 861                         default: /* Illegal value, use the default. */
 862                                 ptlrpcd_partner_group_size = 2;
 863                                 break;
 864                         }
 865                         CWARN("Setting ptlrpcd_partner_group_size = %d\n",
 866                               ptlrpcd_partner_group_size);
 867                 } else {
 868                         CWARN("ptlrpcd_partner_group_size is also set!\n");
 869                 }
 870         }
 871
 872         if (ptlrpcd_partner_group_size == 0)
 873                 ptlrpcd_partner_group_size = 2;
 874         else if (ptlrpcd_partner_group_size < 0)
 875                 ptlrpcd_partner_group_size = -1;
 876         else if (ptlrpcd_per_cpt_max > 0 &&
 877                  ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
 878                 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
 879
 880         /*
 881          * Start the recovery thread first.
 882          */
 883         set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
 884         ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
 885         rc = ptlrpcd_start(&ptlrpcd_rcv);
 886         if (rc < 0)
 887                 GOTO(out, rc);
 888
 889         for (i = 0; i < ncpts; i++) {
 890                 if (cpts == NULL)
 891                         cpt = i;
 892                 else
 893                         cpt = cpts[i];
 894
 895                 nthreads = cfs_cpt_weight(cptable, cpt);
 896                 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
 897                         nthreads = ptlrpcd_per_cpt_max;
 898                 if (nthreads < 2)
 899                         nthreads = 2;
 900
 901                 if (ptlrpcd_partner_group_size <= 0) {
 902                         groupsize = nthreads;
 903                 } else if (nthreads <= ptlrpcd_partner_group_size) {
 904                         groupsize = nthreads;
 905                 } else {
 906                         groupsize = ptlrpcd_partner_group_size;
 907                         if (nthreads % groupsize != 0)
 908                                 nthreads += groupsize - (nthreads % groupsize);
 909                 }
 910
 911                 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 912                 OBD_CPT_ALLOC(pd, cptable, cpt, size);
 913
 914                 if (!pd)
 915                         GOTO(out, rc = -ENOMEM);
 916                 pd->pd_size      = size;
 917                 pd->pd_index     = i;
 918                 pd->pd_cpt       = cpt;
 919                 pd->pd_cursor    = 0;
 920                 pd->pd_nthreads  = nthreads;
 921                 pd->pd_groupsize = groupsize;
 922                 ptlrpcds[i] = pd;
 923
 924                 /*
 925                  * The ptlrpcd threads in a partner group can access
 926                  * each other's struct ptlrpcd_ctl, so these must be
 927                  * initialized before any thead is started.
 928                  */
 929                 for (j = 0; j < nthreads; j++) {
 930                         ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
 931                         rc = ptlrpcd_partners(pd, j);
 932                         if (rc < 0)
 933                                 GOTO(out, rc);
 934                 }
 935
 936                 /* XXX: We start nthreads ptlrpc daemons on this cpt.
 937                  *      Each of them can process any non-recovery
 938                  *      async RPC to improve overall async RPC
 939                  *      efficiency.
 940                  *
 941                  *      But there are some issues with async I/O RPCs
 942                  *      and async non-I/O RPCs processed in the same
 943                  *      set under some cases. The ptlrpcd may be
 944                  *      blocked by some async I/O RPC(s), then will
 945                  *      cause other async non-I/O RPC(s) can not be
 946                  *      processed in time.
 947                  *
 948                  *      Maybe we should distinguish blocked async RPCs
 949                  *      from non-blocked async RPCs, and process them
 950                  *      in different ptlrpcd sets to avoid unnecessary
 951                  *      dependency. But how to distribute async RPCs
 952                  *      load among all the ptlrpc daemons becomes
 953                  *      another trouble.
 954                  */
 955                 for (j = 0; j < nthreads; j++) {
 956                         rc = ptlrpcd_start(&pd->pd_threads[j]);
 957                         if (rc < 0)
 958                                 GOTO(out, rc);
 959                 }
 960         }
 961 out:
 962         if (rc != 0)
 963                 ptlrpcd_fini();
 964
 965         RETURN(rc);
 966 }
 967
 968 int ptlrpcd_addref(void)
 969 {
 970         int rc = 0;
 971
 972         ENTRY;
 973
 974         mutex_lock(&ptlrpcd_mutex);
 975         if (++ptlrpcd_users == 1) {
 976                 rc = ptlrpcd_init();
 977                 if (rc < 0)
 978                         ptlrpcd_users--;
 979         }
 980         mutex_unlock(&ptlrpcd_mutex);
 981         RETURN(rc);
 982 }
 983 EXPORT_SYMBOL(ptlrpcd_addref);
 984
 985 void ptlrpcd_decref(void)
 986 {
 987         mutex_lock(&ptlrpcd_mutex);
 988         if (--ptlrpcd_users == 0)
 989                 ptlrpcd_fini();
 990         mutex_unlock(&ptlrpcd_mutex);
 991 }
 992 EXPORT_SYMBOL(ptlrpcd_decref);
 993 /** @} ptlrpcd */