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, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * 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_table, 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                 /* 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                 for (i = 0; i < pc->pc_npartners; i++)
 231                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
 232         }
 233 }
 234
 235 /**
 236  * Return transferred RPCs count.
 237  */
 238 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 239                                struct ptlrpc_request_set *src)
 240 {
 241         struct list_head *tmp, *pos;
 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_safe(pos, tmp, &src->set_new_requests) {
 248                         req = list_entry(pos, struct ptlrpc_request,
 249                                          rq_set_chain);
 250                         req->rq_set = des;
 251                 }
 252                 list_splice_init(&src->set_new_requests,
 253                                  &des->set_requests);
 254                 rc = atomic_read(&src->set_new_count);
 255                 atomic_add(rc, &des->set_remaining);
 256                 atomic_set(&src->set_new_count, 0);
 257         }
 258         spin_unlock(&src->set_new_req_lock);
 259         return rc;
 260 }
 261
 262 /**
 263  * Requests that are added to the ptlrpcd queue are sent via
 264  * ptlrpcd_check->ptlrpc_check_set().
 265  */
 266 void ptlrpcd_add_req(struct ptlrpc_request *req)
 267 {
 268         struct ptlrpcd_ctl *pc;
 269
 270         if (req->rq_reqmsg)
 271                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
 272
 273         spin_lock(&req->rq_lock);
 274         if (req->rq_invalid_rqset) {
 275                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 276                                                      back_to_sleep, NULL);
 277
 278                 req->rq_invalid_rqset = 0;
 279                 spin_unlock(&req->rq_lock);
 280                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 281         } else if (req->rq_set) {
 282                 /* If we have a vaid "rq_set", just reuse it to avoid double
 283                  * linked. */
 284                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 285                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 286
 287                 /* ptlrpc_check_set will decrease the count */
 288                 atomic_inc(&req->rq_set->set_remaining);
 289                 spin_unlock(&req->rq_lock);
 290                 wake_up(&req->rq_set->set_waitq);
 291                 return;
 292         } else {
 293                 spin_unlock(&req->rq_lock);
 294         }
 295
 296         pc = ptlrpcd_select_pc(req);
 297
 298         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 299                   req, pc->pc_name, pc->pc_index);
 300
 301         ptlrpc_set_add_new_req(pc, req);
 302 }
 303 EXPORT_SYMBOL(ptlrpcd_add_req);
 304
 305 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 306 {
 307         atomic_inc(&set->set_refcount);
 308 }
 309
 310 /**
 311  * Check if there is more work to do on ptlrpcd set.
 312  * Returns 1 if yes.
 313  */
 314 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 315 {
 316         struct list_head *tmp, *pos;
 317         struct ptlrpc_request *req;
 318         struct ptlrpc_request_set *set = pc->pc_set;
 319         int rc = 0;
 320         int rc2;
 321         ENTRY;
 322
 323         if (atomic_read(&set->set_new_count)) {
 324                 spin_lock(&set->set_new_req_lock);
 325                 if (likely(!list_empty(&set->set_new_requests))) {
 326                         list_splice_init(&set->set_new_requests,
 327                                              &set->set_requests);
 328                         atomic_add(atomic_read(&set->set_new_count),
 329                                    &set->set_remaining);
 330                         atomic_set(&set->set_new_count, 0);
 331                         /*
 332                          * Need to calculate its timeout.
 333                          */
 334                         rc = 1;
 335                 }
 336                 spin_unlock(&set->set_new_req_lock);
 337         }
 338
 339         /* We should call lu_env_refill() before handling new requests to make
 340          * sure that env key the requests depending on really exists.
 341          */
 342         rc2 = lu_env_refill(env);
 343         if (rc2 != 0) {
 344                 /*
 345                  * XXX This is very awkward situation, because
 346                  * execution can neither continue (request
 347                  * interpreters assume that env is set up), nor repeat
 348                  * the loop (as this potentially results in a tight
 349                  * loop of -ENOMEM's).
 350                  *
 351                  * Fortunately, refill only ever does something when
 352                  * new modules are loaded, i.e., early during boot up.
 353                  */
 354                 CERROR("Failure to refill session: %d\n", rc2);
 355                 RETURN(rc);
 356         }
 357
 358         if (atomic_read(&set->set_remaining))
 359                 rc |= ptlrpc_check_set(env, set);
 360
 361         /* NB: ptlrpc_check_set has already moved complted request at the
 362          * head of seq::set_requests */
 363         list_for_each_safe(pos, tmp, &set->set_requests) {
 364                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 365                 if (req->rq_phase != RQ_PHASE_COMPLETE)
 366                         break;
 367
 368                 list_del_init(&req->rq_set_chain);
 369                 req->rq_set = NULL;
 370                 ptlrpc_req_finished(req);
 371         }
 372
 373         if (rc == 0) {
 374                 /*
 375                  * If new requests have been added, make sure to wake up.
 376                  */
 377                 rc = atomic_read(&set->set_new_count);
 378
 379                 /* If we have nothing to do, check whether we can take some
 380                  * work from our partner threads. */
 381                 if (rc == 0 && pc->pc_npartners > 0) {
 382                         struct ptlrpcd_ctl *partner;
 383                         struct ptlrpc_request_set *ps;
 384                         int first = pc->pc_cursor;
 385
 386                         do {
 387                                 partner = pc->pc_partners[pc->pc_cursor++];
 388                                 if (pc->pc_cursor >= pc->pc_npartners)
 389                                         pc->pc_cursor = 0;
 390                                 if (partner == NULL)
 391                                         continue;
 392
 393                                 spin_lock(&partner->pc_lock);
 394                                 ps = partner->pc_set;
 395                                 if (ps == NULL) {
 396                                         spin_unlock(&partner->pc_lock);
 397                                         continue;
 398                                 }
 399
 400                                 ptlrpc_reqset_get(ps);
 401                                 spin_unlock(&partner->pc_lock);
 402
 403                                 if (atomic_read(&ps->set_new_count)) {
 404                                         rc = ptlrpcd_steal_rqset(set, ps);
 405                                         if (rc > 0)
 406                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 407                                                        " async RPCs [%d->%d]\n",
 408                                                        rc, partner->pc_index,
 409                                                        pc->pc_index);
 410                                 }
 411                                 ptlrpc_reqset_put(ps);
 412                         } while (rc == 0 && pc->pc_cursor != first);
 413                 }
 414         }
 415
 416         RETURN(rc || test_bit(LIOD_STOP, &pc->pc_flags));
 417 }
 418
 419 /**
 420  * Main ptlrpcd thread.
 421  * ptlrpc's code paths like to execute in process context, so we have this
 422  * thread which spins on a set which contains the rpcs and sends them.
 423  *
 424  */
 425 static int ptlrpcd(void *arg)
 426 {
 427         struct ptlrpcd_ctl              *pc = arg;
 428         struct ptlrpc_request_set       *set;
 429         struct lu_context               ses = { 0 };
 430         struct lu_env                   env = { .le_ses = &ses };
 431         int                             rc = 0;
 432         int                             exit = 0;
 433         ENTRY;
 434
 435         unshare_fs_struct();
 436
 437         if (cfs_cpt_bind(cfs_cpt_table, pc->pc_cpt) != 0)
 438                 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
 439
 440         /*
 441          * Allocate the request set after the thread has been bound
 442          * above. This is safe because no requests will be queued
 443          * until all ptlrpcd threads have confirmed that they have
 444          * successfully started.
 445          */
 446         set = ptlrpc_prep_set();
 447         if (set == NULL)
 448                 GOTO(failed, rc = -ENOMEM);
 449         spin_lock(&pc->pc_lock);
 450         pc->pc_set = set;
 451         spin_unlock(&pc->pc_lock);
 452
 453         /* Both client and server (MDT/OST) may use the environment. */
 454         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD |
 455                                           LCT_DT_THREAD |
 456                                           LCT_CL_THREAD |
 457                                           LCT_REMEMBER  |
 458                                           LCT_NOREF);
 459         if (rc != 0)
 460                 GOTO(failed, rc);
 461         rc = lu_context_init(env.le_ses, LCT_SESSION  |
 462                                          LCT_REMEMBER |
 463                                          LCT_NOREF);
 464         if (rc != 0) {
 465                 lu_context_fini(&env.le_ctx);
 466                 GOTO(failed, rc);
 467         }
 468
 469         complete(&pc->pc_starting);
 470
 471         /*
 472          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 473          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 474          * there are requests in the set. New requests come in on the set's
 475          * new_req_list and ptlrpcd_check() moves them into the set.
 476          */
 477         do {
 478                 struct l_wait_info lwi;
 479                 time64_t timeout;
 480
 481                 timeout = ptlrpc_set_next_timeout(set);
 482                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout),
 483                                 ptlrpc_expired_set, set);
 484
 485                 lu_context_enter(&env.le_ctx);
 486                 lu_context_enter(env.le_ses);
 487                 l_wait_event(set->set_waitq, ptlrpcd_check(&env, pc), &lwi);
 488                 lu_context_exit(&env.le_ctx);
 489                 lu_context_exit(env.le_ses);
 490
 491                 /*
 492                  * Abort inflight rpcs for forced stop case.
 493                  */
 494                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 495                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 496                                 ptlrpc_abort_set(set);
 497                         exit++;
 498                 }
 499
 500                 /*
 501                  * Let's make one more loop to make sure that ptlrpcd_check()
 502                  * copied all raced new rpcs into the set so we can kill them.
 503                  */
 504         } while (exit < 2);
 505
 506         /*
 507          * Wait for inflight requests to drain.
 508          */
 509         if (!list_empty(&set->set_requests))
 510                 ptlrpc_set_wait(set);
 511         lu_context_fini(&env.le_ctx);
 512         lu_context_fini(env.le_ses);
 513
 514         complete(&pc->pc_finishing);
 515
 516         return 0;
 517
 518 failed:
 519         pc->pc_error = rc;
 520         complete(&pc->pc_starting);
 521         RETURN(rc);
 522 }
 523
 524 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
 525 {
 526         ENTRY;
 527
 528         pc->pc_index = index;
 529         pc->pc_cpt = cpt;
 530         init_completion(&pc->pc_starting);
 531         init_completion(&pc->pc_finishing);
 532         spin_lock_init(&pc->pc_lock);
 533
 534         if (index < 0) {
 535                 /* Recovery thread. */
 536                 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
 537         } else {
 538                 /* Regular thread. */
 539                 snprintf(pc->pc_name, sizeof(pc->pc_name),
 540                          "ptlrpcd_%02d_%02d", cpt, index);
 541         }
 542
 543         EXIT;
 544 }
 545
 546 /* XXX: We want multiple CPU cores to share the async RPC load. So we
 547  *      start many ptlrpcd threads. We also want to reduce the ptlrpcd
 548  *      overhead caused by data transfer cross-CPU cores. So we bind
 549  *      all ptlrpcd threads to a CPT, in the expectation that CPTs
 550  *      will be defined in a way that matches these boundaries. Within
 551  *      a CPT a ptlrpcd thread can be scheduled on any available core.
 552  *
 553  *      Each ptlrpcd thread has its own request queue. This can cause
 554  *      response delay if the thread is already busy. To help with
 555  *      this we define partner threads: these are other threads bound
 556  *      to the same CPT which will check for work in each other's
 557  *      request queues if they have no work to do.
 558  *
 559  *      The desired number of partner threads can be tuned by setting
 560  *      ptlrpcd_partner_group_size. The default is to create pairs of
 561  *      partner threads.
 562  */
 563 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
 564 {
 565         struct ptlrpcd_ctl      *pc;
 566         struct ptlrpcd_ctl      **ppc;
 567         int                     first;
 568         int                     i;
 569         int                     rc = 0;
 570         ENTRY;
 571
 572         LASSERT(index >= 0 && index < pd->pd_nthreads);
 573         pc = &pd->pd_threads[index];
 574         pc->pc_npartners = pd->pd_groupsize - 1;
 575
 576         if (pc->pc_npartners <= 0)
 577                 GOTO(out, rc);
 578
 579         OBD_CPT_ALLOC(pc->pc_partners, cfs_cpt_table, pc->pc_cpt,
 580                       sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 581         if (pc->pc_partners == NULL) {
 582                 pc->pc_npartners = 0;
 583                 GOTO(out, rc = -ENOMEM);
 584         }
 585
 586         first = index - index % pd->pd_groupsize;
 587         ppc = pc->pc_partners;
 588         for (i = first; i < first + pd->pd_groupsize; i++) {
 589                 if (i != index)
 590                         *ppc++ = &pd->pd_threads[i];
 591         }
 592 out:
 593         RETURN(rc);
 594 }
 595
 596 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
 597 {
 598         struct task_struct      *task;
 599         int                     rc = 0;
 600         ENTRY;
 601
 602         /*
 603          * Do not allow starting a second thread for one pc.
 604          */
 605         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 606                 CWARN("Starting second thread (%s) for same pc %p\n",
 607                       pc->pc_name, pc);
 608                 RETURN(0);
 609         }
 610
 611         task = kthread_run(ptlrpcd, pc, pc->pc_name);
 612         if (IS_ERR(task))
 613                 GOTO(out_set, rc = PTR_ERR(task));
 614
 615         wait_for_completion(&pc->pc_starting);
 616         rc = pc->pc_error;
 617         if (rc != 0)
 618                 GOTO(out_set, rc);
 619
 620         RETURN(0);
 621
 622 out_set:
 623         if (pc->pc_set != NULL) {
 624                 struct ptlrpc_request_set *set = pc->pc_set;
 625
 626                 spin_lock(&pc->pc_lock);
 627                 pc->pc_set = NULL;
 628                 spin_unlock(&pc->pc_lock);
 629                 ptlrpc_set_destroy(set);
 630         }
 631         clear_bit(LIOD_START, &pc->pc_flags);
 632         RETURN(rc);
 633 }
 634
 635 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 636 {
 637         ENTRY;
 638
 639         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 640                 CWARN("Thread for pc %p was not started\n", pc);
 641                 goto out;
 642         }
 643
 644         set_bit(LIOD_STOP, &pc->pc_flags);
 645         if (force)
 646                 set_bit(LIOD_FORCE, &pc->pc_flags);
 647         wake_up(&pc->pc_set->set_waitq);
 648
 649 out:
 650         EXIT;
 651 }
 652
 653 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 654 {
 655         struct ptlrpc_request_set *set = pc->pc_set;
 656         ENTRY;
 657
 658         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 659                 CWARN("Thread for pc %p was not started\n", pc);
 660                 goto out;
 661         }
 662
 663         wait_for_completion(&pc->pc_finishing);
 664
 665         spin_lock(&pc->pc_lock);
 666         pc->pc_set = NULL;
 667         spin_unlock(&pc->pc_lock);
 668         ptlrpc_set_destroy(set);
 669
 670         clear_bit(LIOD_START, &pc->pc_flags);
 671         clear_bit(LIOD_STOP, &pc->pc_flags);
 672         clear_bit(LIOD_FORCE, &pc->pc_flags);
 673
 674 out:
 675         if (pc->pc_npartners > 0) {
 676                 LASSERT(pc->pc_partners != NULL);
 677
 678                 OBD_FREE(pc->pc_partners,
 679                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 680                 pc->pc_partners = NULL;
 681         }
 682         pc->pc_npartners = 0;
 683         pc->pc_error = 0;
 684         EXIT;
 685 }
 686
 687 static void ptlrpcd_fini(void)
 688 {
 689         int     i;
 690         int     j;
 691         int     ncpts;
 692         ENTRY;
 693
 694         if (ptlrpcds != NULL) {
 695                 for (i = 0; i < ptlrpcds_num; i++) {
 696                         if (ptlrpcds[i] == NULL)
 697                                 break;
 698                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 699                                 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
 700                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 701                                 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
 702                         OBD_FREE(ptlrpcds[i], ptlrpcds[i]->pd_size);
 703                         ptlrpcds[i] = NULL;
 704                 }
 705                 OBD_FREE(ptlrpcds, sizeof(ptlrpcds[0]) * ptlrpcds_num);
 706         }
 707         ptlrpcds_num = 0;
 708
 709         ptlrpcd_stop(&ptlrpcd_rcv, 0);
 710         ptlrpcd_free(&ptlrpcd_rcv);
 711
 712         if (ptlrpcds_cpt_idx != NULL) {
 713                 ncpts = cfs_cpt_number(cfs_cpt_table);
 714                 OBD_FREE(ptlrpcds_cpt_idx, ncpts * sizeof(ptlrpcds_cpt_idx[0]));
 715                 ptlrpcds_cpt_idx = NULL;
 716         }
 717
 718         EXIT;
 719 }
 720
 721 static int ptlrpcd_init(void)
 722 {
 723         int                     nthreads;
 724         int                     groupsize;
 725         int                     size;
 726         int                     i;
 727         int                     j;
 728         int                     rc = 0;
 729         struct cfs_cpt_table    *cptable;
 730         __u32                   *cpts = NULL;
 731         int                     ncpts;
 732         int                     cpt;
 733         struct ptlrpcd          *pd;
 734         ENTRY;
 735
 736         /*
 737          * Determine the CPTs that ptlrpcd threads will run on.
 738          */
 739         cptable = cfs_cpt_table;
 740         ncpts = cfs_cpt_number(cptable);
 741         if (ptlrpcd_cpts != NULL) {
 742                 struct cfs_expr_list    *el;
 743
 744                 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
 745                 OBD_ALLOC(ptlrpcds_cpt_idx, size);
 746                 if (ptlrpcds_cpt_idx == NULL)
 747                         GOTO(out, rc = -ENOMEM);
 748
 749                 rc = cfs_expr_list_parse(ptlrpcd_cpts,
 750                                          strlen(ptlrpcd_cpts),
 751                                          0, ncpts - 1, &el);
 752                 if (rc != 0) {
 753                         CERROR("%s: invalid CPT pattern string: %s",
 754                                "ptlrpcd_cpts", ptlrpcd_cpts);
 755                         GOTO(out, rc = -EINVAL);
 756                 }
 757
 758                 rc = cfs_expr_list_values(el, ncpts, &cpts);
 759                 cfs_expr_list_free(el);
 760                 if (rc <= 0) {
 761                         CERROR("%s: failed to parse CPT array %s: %d\n",
 762                                "ptlrpcd_cpts", ptlrpcd_cpts, rc);
 763                         if (rc == 0)
 764                                 rc = -EINVAL;
 765                         GOTO(out, rc);
 766                 }
 767
 768                 /*
 769                  * Create the cpt-to-index map. When there is no match
 770                  * in the cpt table, pick a cpt at random. This could
 771                  * be changed to take the topology of the system into
 772                  * account.
 773                  */
 774                 for (cpt = 0; cpt < ncpts; cpt++) {
 775                         for (i = 0; i < rc; i++)
 776                                 if (cpts[i] == cpt)
 777                                         break;
 778                         if (i >= rc)
 779                                 i = cpt % rc;
 780                         ptlrpcds_cpt_idx[cpt] = i;
 781                 }
 782
 783                 cfs_expr_list_values_free(cpts, rc);
 784                 ncpts = rc;
 785         }
 786         ptlrpcds_num = ncpts;
 787
 788         size = ncpts * sizeof(ptlrpcds[0]);
 789         OBD_ALLOC(ptlrpcds, size);
 790         if (ptlrpcds == NULL)
 791                 GOTO(out, rc = -ENOMEM);
 792
 793         /*
 794          * The max_ptlrpcds parameter is obsolete, but do something
 795          * sane if it has been tuned, and complain if
 796          * ptlrpcd_per_cpt_max has also been tuned.
 797          */
 798         if (max_ptlrpcds != 0) {
 799                 CWARN("max_ptlrpcds is obsolete.\n");
 800                 if (ptlrpcd_per_cpt_max == 0) {
 801                         ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
 802                         /* Round up if there is a remainder. */
 803                         if (max_ptlrpcds % ncpts != 0)
 804                                 ptlrpcd_per_cpt_max++;
 805                         CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
 806                               ptlrpcd_per_cpt_max);
 807                 } else {
 808                         CWARN("ptlrpd_per_cpt_max is also set!\n");
 809                 }
 810         }
 811
 812         /*
 813          * The ptlrpcd_bind_policy parameter is obsolete, but do
 814          * something sane if it has been tuned, and complain if
 815          * ptlrpcd_partner_group_size is also tuned.
 816          */
 817         if (ptlrpcd_bind_policy != 0) {
 818                 CWARN("ptlrpcd_bind_policy is obsolete.\n");
 819                 if (ptlrpcd_partner_group_size == 0) {
 820                         switch (ptlrpcd_bind_policy) {
 821                         case 1: /* PDB_POLICY_NONE */
 822                         case 2: /* PDB_POLICY_FULL */
 823                                 ptlrpcd_partner_group_size = 1;
 824                                 break;
 825                         case 3: /* PDB_POLICY_PAIR */
 826                                 ptlrpcd_partner_group_size = 2;
 827                                 break;
 828                         case 4: /* PDB_POLICY_NEIGHBOR */
 829 #ifdef CONFIG_NUMA
 830                                 ptlrpcd_partner_group_size = -1; /* CPT */
 831 #else
 832                                 ptlrpcd_partner_group_size = 3; /* Triplets */
 833 #endif
 834                                 break;
 835                         default: /* Illegal value, use the default. */
 836                                 ptlrpcd_partner_group_size = 2;
 837                                 break;
 838                         }
 839                         CWARN("Setting ptlrpcd_partner_group_size = %d\n",
 840                               ptlrpcd_partner_group_size);
 841                 } else {
 842                         CWARN("ptlrpcd_partner_group_size is also set!\n");
 843                 }
 844         }
 845
 846         if (ptlrpcd_partner_group_size == 0)
 847                 ptlrpcd_partner_group_size = 2;
 848         else if (ptlrpcd_partner_group_size < 0)
 849                 ptlrpcd_partner_group_size = -1;
 850         else if (ptlrpcd_per_cpt_max > 0 &&
 851                  ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
 852                 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
 853
 854         /*
 855          * Start the recovery thread first.
 856          */
 857         set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
 858         ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
 859         rc = ptlrpcd_start(&ptlrpcd_rcv);
 860         if (rc < 0)
 861                 GOTO(out, rc);
 862
 863         for (i = 0; i < ncpts; i++) {
 864                 if (cpts == NULL)
 865                         cpt = i;
 866                 else
 867                         cpt = cpts[i];
 868
 869                 nthreads = cfs_cpt_weight(cptable, cpt);
 870                 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
 871                         nthreads = ptlrpcd_per_cpt_max;
 872                 if (nthreads < 2)
 873                         nthreads = 2;
 874
 875                 if (ptlrpcd_partner_group_size <= 0) {
 876                         groupsize = nthreads;
 877                 } else if (nthreads <= ptlrpcd_partner_group_size) {
 878                         groupsize = nthreads;
 879                 } else {
 880                         groupsize = ptlrpcd_partner_group_size;
 881                         if (nthreads % groupsize != 0)
 882                                 nthreads += groupsize - (nthreads % groupsize);
 883                 }
 884
 885                 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 886                 OBD_CPT_ALLOC(pd, cptable, cpt, size);
 887                 if (!pd)
 888                         GOTO(out, rc = -ENOMEM);
 889                 pd->pd_size      = size;
 890                 pd->pd_index     = i;
 891                 pd->pd_cpt       = cpt;
 892                 pd->pd_cursor    = 0;
 893                 pd->pd_nthreads  = nthreads;
 894                 pd->pd_groupsize = groupsize;
 895                 ptlrpcds[i] = pd;
 896
 897                 /*
 898                  * The ptlrpcd threads in a partner group can access
 899                  * each other's struct ptlrpcd_ctl, so these must be
 900                  * initialized before any thead is started.
 901                  */
 902                 for (j = 0; j < nthreads; j++) {
 903                         ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
 904                         rc = ptlrpcd_partners(pd, j);
 905                         if (rc < 0)
 906                                 GOTO(out, rc);
 907                 }
 908
 909                 /* XXX: We start nthreads ptlrpc daemons on this cpt.
 910                  *      Each of them can process any non-recovery
 911                  *      async RPC to improve overall async RPC
 912                  *      efficiency.
 913                  *
 914                  *      But there are some issues with async I/O RPCs
 915                  *      and async non-I/O RPCs processed in the same
 916                  *      set under some cases. The ptlrpcd may be
 917                  *      blocked by some async I/O RPC(s), then will
 918                  *      cause other async non-I/O RPC(s) can not be
 919                  *      processed in time.
 920                  *
 921                  *      Maybe we should distinguish blocked async RPCs
 922                  *      from non-blocked async RPCs, and process them
 923                  *      in different ptlrpcd sets to avoid unnecessary
 924                  *      dependency. But how to distribute async RPCs
 925                  *      load among all the ptlrpc daemons becomes
 926                  *      another trouble.
 927                  */
 928                 for (j = 0; j < nthreads; j++) {
 929                         rc = ptlrpcd_start(&pd->pd_threads[j]);
 930                         if (rc < 0)
 931                                 GOTO(out, rc);
 932                 }
 933         }
 934 out:
 935         if (rc != 0)
 936                 ptlrpcd_fini();
 937
 938         RETURN(rc);
 939 }
 940
 941 int ptlrpcd_addref(void)
 942 {
 943         int rc = 0;
 944         ENTRY;
 945
 946         mutex_lock(&ptlrpcd_mutex);
 947         if (++ptlrpcd_users == 1) {
 948                 rc = ptlrpcd_init();
 949                 if (rc < 0)
 950                         ptlrpcd_users--;
 951         }
 952         mutex_unlock(&ptlrpcd_mutex);
 953         RETURN(rc);
 954 }
 955 EXPORT_SYMBOL(ptlrpcd_addref);
 956
 957 void ptlrpcd_decref(void)
 958 {
 959         mutex_lock(&ptlrpcd_mutex);
 960         if (--ptlrpcd_users == 0)
 961                 ptlrpcd_fini();
 962         mutex_unlock(&ptlrpcd_mutex);
 963 }
 964 EXPORT_SYMBOL(ptlrpcd_decref);
 965 /** @} ptlrpcd */