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_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                 /*
 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 list_head *tmp, *pos;
 244         struct ptlrpc_request *req;
 245         int rc = 0;
 246
 247         spin_lock(&src->set_new_req_lock);
 248         if (likely(!list_empty(&src->set_new_requests))) {
 249                 list_for_each_safe(pos, tmp, &src->set_new_requests) {
 250                         req = list_entry(pos, struct ptlrpc_request,
 251                                          rq_set_chain);
 252                         req->rq_set = des;
 253                 }
 254                 list_splice_init(&src->set_new_requests,
 255                                  &des->set_requests);
 256                 rc = atomic_read(&src->set_new_count);
 257                 atomic_add(rc, &des->set_remaining);
 258                 atomic_set(&src->set_new_count, 0);
 259         }
 260         spin_unlock(&src->set_new_req_lock);
 261         return rc;
 262 }
 263
 264 /**
 265  * Requests that are added to the ptlrpcd queue are sent via
 266  * ptlrpcd_check->ptlrpc_check_set().
 267  */
 268 void ptlrpcd_add_req(struct ptlrpc_request *req)
 269 {
 270         struct ptlrpcd_ctl *pc;
 271
 272         if (req->rq_reqmsg)
 273                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
 274
 275         spin_lock(&req->rq_lock);
 276         if (req->rq_invalid_rqset) {
 277                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 278                                                      back_to_sleep, NULL);
 279
 280                 req->rq_invalid_rqset = 0;
 281                 spin_unlock(&req->rq_lock);
 282                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 283         } else if (req->rq_set) {
 284                 /*
 285                  * If we have a vaid "rq_set", just reuse it to avoid double
 286                  * linked.
 287                  */
 288                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 289                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 290
 291                 /* ptlrpc_check_set will decrease the count */
 292                 atomic_inc(&req->rq_set->set_remaining);
 293                 spin_unlock(&req->rq_lock);
 294                 wake_up(&req->rq_set->set_waitq);
 295                 return;
 296         } else {
 297                 spin_unlock(&req->rq_lock);
 298         }
 299
 300         pc = ptlrpcd_select_pc(req);
 301
 302         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 303                   req, pc->pc_name, pc->pc_index);
 304
 305         ptlrpc_set_add_new_req(pc, req);
 306 }
 307 EXPORT_SYMBOL(ptlrpcd_add_req);
 308
 309 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 310 {
 311         atomic_inc(&set->set_refcount);
 312 }
 313
 314 /**
 315  * Check if there is more work to do on ptlrpcd set.
 316  * Returns 1 if yes.
 317  */
 318 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 319 {
 320         struct list_head *tmp, *pos;
 321         struct ptlrpc_request *req;
 322         struct ptlrpc_request_set *set = pc->pc_set;
 323         int rc = 0;
 324         int rc2;
 325
 326         ENTRY;
 327
 328         if (atomic_read(&set->set_new_count)) {
 329                 spin_lock(&set->set_new_req_lock);
 330                 if (likely(!list_empty(&set->set_new_requests))) {
 331                         list_splice_init(&set->set_new_requests,
 332                                              &set->set_requests);
 333                         atomic_add(atomic_read(&set->set_new_count),
 334                                    &set->set_remaining);
 335                         atomic_set(&set->set_new_count, 0);
 336                         /*
 337                          * Need to calculate its timeout.
 338                          */
 339                         rc = 1;
 340                 }
 341                 spin_unlock(&set->set_new_req_lock);
 342         }
 343
 344         /*
 345          * We should call lu_env_refill() before handling new requests to make
 346          * sure that env key the requests depending on really exists.
 347          */
 348         rc2 = lu_env_refill(env);
 349         if (rc2 != 0) {
 350                 /*
 351                  * XXX This is very awkward situation, because
 352                  * execution can neither continue (request
 353                  * interpreters assume that env is set up), nor repeat
 354                  * the loop (as this potentially results in a tight
 355                  * loop of -ENOMEM's).
 356                  *
 357                  * Fortunately, refill only ever does something when
 358                  * new modules are loaded, i.e., early during boot up.
 359                  */
 360                 CERROR("Failure to refill session: %d\n", rc2);
 361                 RETURN(rc);
 362         }
 363
 364         if (atomic_read(&set->set_remaining))
 365                 rc |= ptlrpc_check_set(env, set);
 366
 367         /*
 368          * NB: ptlrpc_check_set has already moved complted request at the
 369          * head of seq::set_requests
 370          */
 371         list_for_each_safe(pos, tmp, &set->set_requests) {
 372                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 373                 if (req->rq_phase != RQ_PHASE_COMPLETE)
 374                         break;
 375
 376                 list_del_init(&req->rq_set_chain);
 377                 req->rq_set = NULL;
 378                 ptlrpc_req_finished(req);
 379         }
 380
 381         if (rc == 0) {
 382                 /*
 383                  * If new requests have been added, make sure to wake up.
 384                  */
 385                 rc = atomic_read(&set->set_new_count);
 386
 387                 /*
 388                  * If we have nothing to do, check whether we can take some
 389                  * work from our partner threads.
 390                  */
 391                 if (rc == 0 && pc->pc_npartners > 0) {
 392                         struct ptlrpcd_ctl *partner;
 393                         struct ptlrpc_request_set *ps;
 394                         int first = pc->pc_cursor;
 395
 396                         do {
 397                                 partner = pc->pc_partners[pc->pc_cursor++];
 398                                 if (pc->pc_cursor >= pc->pc_npartners)
 399                                         pc->pc_cursor = 0;
 400                                 if (partner == NULL)
 401                                         continue;
 402
 403                                 spin_lock(&partner->pc_lock);
 404                                 ps = partner->pc_set;
 405                                 if (ps == NULL) {
 406                                         spin_unlock(&partner->pc_lock);
 407                                         continue;
 408                                 }
 409
 410                                 ptlrpc_reqset_get(ps);
 411                                 spin_unlock(&partner->pc_lock);
 412
 413                                 if (atomic_read(&ps->set_new_count)) {
 414                                         rc = ptlrpcd_steal_rqset(set, ps);
 415                                         if (rc > 0)
 416                                                 CDEBUG(D_RPCTRACE,
 417                                                        "transfer %d async RPCs [%d->%d]\n",
 418                                                        rc, partner->pc_index,
 419                                                        pc->pc_index);
 420                                 }
 421                                 ptlrpc_reqset_put(ps);
 422                         } while (rc == 0 && pc->pc_cursor != first);
 423                 }
 424         }
 425
 426         RETURN(rc || test_bit(LIOD_STOP, &pc->pc_flags));
 427 }
 428
 429 /**
 430  * Main ptlrpcd thread.
 431  * ptlrpc's code paths like to execute in process context, so we have this
 432  * thread which spins on a set which contains the rpcs and sends them.
 433  */
 434 static int ptlrpcd(void *arg)
 435 {
 436         struct ptlrpcd_ctl              *pc = arg;
 437         struct ptlrpc_request_set       *set;
 438         struct lu_context               ses = { 0 };
 439         struct lu_env                   env = { .le_ses = &ses };
 440         int                             rc = 0;
 441         int                             exit = 0;
 442
 443         ENTRY;
 444
 445         unshare_fs_struct();
 446
 447         if (cfs_cpt_bind(cfs_cpt_table, pc->pc_cpt) != 0)
 448                 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
 449
 450         /*
 451          * Allocate the request set after the thread has been bound
 452          * above. This is safe because no requests will be queued
 453          * until all ptlrpcd threads have confirmed that they have
 454          * successfully started.
 455          */
 456         set = ptlrpc_prep_set();
 457         if (set == NULL)
 458                 GOTO(failed, rc = -ENOMEM);
 459         spin_lock(&pc->pc_lock);
 460         pc->pc_set = set;
 461         spin_unlock(&pc->pc_lock);
 462
 463         /* Both client and server (MDT/OST) may use the environment. */
 464         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD |
 465                                           LCT_DT_THREAD |
 466                                           LCT_CL_THREAD |
 467                                           LCT_REMEMBER  |
 468                                           LCT_NOREF);
 469         if (rc != 0)
 470                 GOTO(failed, rc);
 471         rc = lu_context_init(env.le_ses, LCT_SESSION  |
 472                                          LCT_REMEMBER |
 473                                          LCT_NOREF);
 474         if (rc != 0) {
 475                 lu_context_fini(&env.le_ctx);
 476                 GOTO(failed, rc);
 477         }
 478
 479         complete(&pc->pc_starting);
 480
 481         /*
 482          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 483          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 484          * there are requests in the set. New requests come in on the set's
 485          * new_req_list and ptlrpcd_check() moves them into the set.
 486          */
 487         do {
 488                 struct l_wait_info lwi;
 489                 time64_t timeout;
 490
 491                 timeout = ptlrpc_set_next_timeout(set);
 492                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout),
 493                                 ptlrpc_expired_set, set);
 494
 495                 lu_context_enter(&env.le_ctx);
 496                 lu_context_enter(env.le_ses);
 497                 l_wait_event(set->set_waitq, ptlrpcd_check(&env, pc), &lwi);
 498                 lu_context_exit(&env.le_ctx);
 499                 lu_context_exit(env.le_ses);
 500
 501                 /*
 502                  * Abort inflight rpcs for forced stop case.
 503                  */
 504                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 505                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 506                                 ptlrpc_abort_set(set);
 507                         exit++;
 508                 }
 509
 510                 /*
 511                  * Let's make one more loop to make sure that ptlrpcd_check()
 512                  * copied all raced new rpcs into the set so we can kill them.
 513                  */
 514         } while (exit < 2);
 515
 516         /*
 517          * Wait for inflight requests to drain.
 518          */
 519         if (!list_empty(&set->set_requests))
 520                 ptlrpc_set_wait(&env, set);
 521         lu_context_fini(&env.le_ctx);
 522         lu_context_fini(env.le_ses);
 523
 524         complete(&pc->pc_finishing);
 525
 526         return 0;
 527
 528 failed:
 529         pc->pc_error = rc;
 530         complete(&pc->pc_starting);
 531         RETURN(rc);
 532 }
 533
 534 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
 535 {
 536         ENTRY;
 537
 538         pc->pc_index = index;
 539         pc->pc_cpt = cpt;
 540         init_completion(&pc->pc_starting);
 541         init_completion(&pc->pc_finishing);
 542         spin_lock_init(&pc->pc_lock);
 543
 544         if (index < 0) {
 545                 /* Recovery thread. */
 546                 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
 547         } else {
 548                 /* Regular thread. */
 549                 snprintf(pc->pc_name, sizeof(pc->pc_name),
 550                          "ptlrpcd_%02d_%02d", cpt, index);
 551         }
 552
 553         EXIT;
 554 }
 555
 556 /* XXX: We want multiple CPU cores to share the async RPC load. So we
 557  *      start many ptlrpcd threads. We also want to reduce the ptlrpcd
 558  *      overhead caused by data transfer cross-CPU cores. So we bind
 559  *      all ptlrpcd threads to a CPT, in the expectation that CPTs
 560  *      will be defined in a way that matches these boundaries. Within
 561  *      a CPT a ptlrpcd thread can be scheduled on any available core.
 562  *
 563  *      Each ptlrpcd thread has its own request queue. This can cause
 564  *      response delay if the thread is already busy. To help with
 565  *      this we define partner threads: these are other threads bound
 566  *      to the same CPT which will check for work in each other's
 567  *      request queues if they have no work to do.
 568  *
 569  *      The desired number of partner threads can be tuned by setting
 570  *      ptlrpcd_partner_group_size. The default is to create pairs of
 571  *      partner threads.
 572  */
 573 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
 574 {
 575         struct ptlrpcd_ctl      *pc;
 576         struct ptlrpcd_ctl      **ppc;
 577         int                     first;
 578         int                     i;
 579         int                     rc = 0;
 580
 581         ENTRY;
 582
 583         LASSERT(index >= 0 && index < pd->pd_nthreads);
 584         pc = &pd->pd_threads[index];
 585         pc->pc_npartners = pd->pd_groupsize - 1;
 586
 587         if (pc->pc_npartners <= 0)
 588                 GOTO(out, rc);
 589
 590         OBD_CPT_ALLOC(pc->pc_partners, cfs_cpt_table, pc->pc_cpt,
 591                       sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 592         if (pc->pc_partners == NULL) {
 593                 pc->pc_npartners = 0;
 594                 GOTO(out, rc = -ENOMEM);
 595         }
 596
 597         first = index - index % pd->pd_groupsize;
 598         ppc = pc->pc_partners;
 599         for (i = first; i < first + pd->pd_groupsize; i++) {
 600                 if (i != index)
 601                         *ppc++ = &pd->pd_threads[i];
 602         }
 603 out:
 604         RETURN(rc);
 605 }
 606
 607 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
 608 {
 609         struct task_struct      *task;
 610         int                     rc = 0;
 611
 612         ENTRY;
 613
 614         /*
 615          * Do not allow starting a second thread for one pc.
 616          */
 617         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 618                 CWARN("Starting second thread (%s) for same pc %p\n",
 619                       pc->pc_name, pc);
 620                 RETURN(0);
 621         }
 622
 623         task = kthread_run(ptlrpcd, pc, pc->pc_name);
 624         if (IS_ERR(task))
 625                 GOTO(out_set, rc = PTR_ERR(task));
 626
 627         wait_for_completion(&pc->pc_starting);
 628         rc = pc->pc_error;
 629         if (rc != 0)
 630                 GOTO(out_set, rc);
 631
 632         RETURN(0);
 633
 634 out_set:
 635         if (pc->pc_set != NULL) {
 636                 struct ptlrpc_request_set *set = pc->pc_set;
 637
 638                 spin_lock(&pc->pc_lock);
 639                 pc->pc_set = NULL;
 640                 spin_unlock(&pc->pc_lock);
 641                 ptlrpc_set_destroy(set);
 642         }
 643         clear_bit(LIOD_START, &pc->pc_flags);
 644         RETURN(rc);
 645 }
 646
 647 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 648 {
 649         ENTRY;
 650
 651         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 652                 CWARN("Thread for pc %p was not started\n", pc);
 653                 goto out;
 654         }
 655
 656         set_bit(LIOD_STOP, &pc->pc_flags);
 657         if (force)
 658                 set_bit(LIOD_FORCE, &pc->pc_flags);
 659         wake_up(&pc->pc_set->set_waitq);
 660
 661 out:
 662         EXIT;
 663 }
 664
 665 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 666 {
 667         struct ptlrpc_request_set *set = pc->pc_set;
 668
 669         ENTRY;
 670
 671         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 672                 CWARN("Thread for pc %p was not started\n", pc);
 673                 goto out;
 674         }
 675
 676         wait_for_completion(&pc->pc_finishing);
 677
 678         spin_lock(&pc->pc_lock);
 679         pc->pc_set = NULL;
 680         spin_unlock(&pc->pc_lock);
 681         ptlrpc_set_destroy(set);
 682
 683         clear_bit(LIOD_START, &pc->pc_flags);
 684         clear_bit(LIOD_STOP, &pc->pc_flags);
 685         clear_bit(LIOD_FORCE, &pc->pc_flags);
 686
 687 out:
 688         if (pc->pc_npartners > 0) {
 689                 LASSERT(pc->pc_partners != NULL);
 690
 691                 OBD_FREE(pc->pc_partners,
 692                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 693                 pc->pc_partners = NULL;
 694         }
 695         pc->pc_npartners = 0;
 696         pc->pc_error = 0;
 697         EXIT;
 698 }
 699
 700 static void ptlrpcd_fini(void)
 701 {
 702         int     i;
 703         int     j;
 704         int     ncpts;
 705
 706         ENTRY;
 707
 708         if (ptlrpcds != NULL) {
 709                 for (i = 0; i < ptlrpcds_num; i++) {
 710                         if (ptlrpcds[i] == NULL)
 711                                 break;
 712                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 713                                 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
 714                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
 715                                 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
 716                         OBD_FREE(ptlrpcds[i], ptlrpcds[i]->pd_size);
 717                         ptlrpcds[i] = NULL;
 718                 }
 719                 OBD_FREE(ptlrpcds, sizeof(ptlrpcds[0]) * ptlrpcds_num);
 720         }
 721         ptlrpcds_num = 0;
 722
 723         ptlrpcd_stop(&ptlrpcd_rcv, 0);
 724         ptlrpcd_free(&ptlrpcd_rcv);
 725
 726         if (ptlrpcds_cpt_idx != NULL) {
 727                 ncpts = cfs_cpt_number(cfs_cpt_table);
 728                 OBD_FREE(ptlrpcds_cpt_idx, ncpts * sizeof(ptlrpcds_cpt_idx[0]));
 729                 ptlrpcds_cpt_idx = NULL;
 730         }
 731
 732         EXIT;
 733 }
 734
 735 static int ptlrpcd_init(void)
 736 {
 737         int                     nthreads;
 738         int                     groupsize;
 739         int                     size;
 740         int                     i;
 741         int                     j;
 742         int                     rc = 0;
 743         struct cfs_cpt_table    *cptable;
 744         __u32                   *cpts = NULL;
 745         int                     ncpts;
 746         int                     cpt;
 747         struct ptlrpcd          *pd;
 748
 749         ENTRY;
 750
 751         /*
 752          * Determine the CPTs that ptlrpcd threads will run on.
 753          */
 754         cptable = cfs_cpt_table;
 755         ncpts = cfs_cpt_number(cptable);
 756         if (ptlrpcd_cpts != NULL) {
 757                 struct cfs_expr_list *el;
 758
 759                 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
 760                 OBD_ALLOC(ptlrpcds_cpt_idx, size);
 761                 if (ptlrpcds_cpt_idx == NULL)
 762                         GOTO(out, rc = -ENOMEM);
 763
 764                 rc = cfs_expr_list_parse(ptlrpcd_cpts,
 765                                          strlen(ptlrpcd_cpts),
 766                                          0, ncpts - 1, &el);
 767                 if (rc != 0) {
 768                         CERROR("%s: invalid CPT pattern string: %s",
 769                                "ptlrpcd_cpts", ptlrpcd_cpts);
 770                         GOTO(out, rc = -EINVAL);
 771                 }
 772
 773                 rc = cfs_expr_list_values(el, ncpts, &cpts);
 774                 cfs_expr_list_free(el);
 775                 if (rc <= 0) {
 776                         CERROR("%s: failed to parse CPT array %s: %d\n",
 777                                "ptlrpcd_cpts", ptlrpcd_cpts, rc);
 778                         if (rc == 0)
 779                                 rc = -EINVAL;
 780                         GOTO(out, rc);
 781                 }
 782
 783                 /*
 784                  * Create the cpt-to-index map. When there is no match
 785                  * in the cpt table, pick a cpt at random. This could
 786                  * be changed to take the topology of the system into
 787                  * account.
 788                  */
 789                 for (cpt = 0; cpt < ncpts; cpt++) {
 790                         for (i = 0; i < rc; i++)
 791                                 if (cpts[i] == cpt)
 792                                         break;
 793                         if (i >= rc)
 794                                 i = cpt % rc;
 795                         ptlrpcds_cpt_idx[cpt] = i;
 796                 }
 797
 798                 cfs_expr_list_values_free(cpts, rc);
 799                 ncpts = rc;
 800         }
 801         ptlrpcds_num = ncpts;
 802
 803         size = ncpts * sizeof(ptlrpcds[0]);
 804         OBD_ALLOC(ptlrpcds, size);
 805         if (ptlrpcds == NULL)
 806                 GOTO(out, rc = -ENOMEM);
 807
 808         /*
 809          * The max_ptlrpcds parameter is obsolete, but do something
 810          * sane if it has been tuned, and complain if
 811          * ptlrpcd_per_cpt_max has also been tuned.
 812          */
 813         if (max_ptlrpcds != 0) {
 814                 CWARN("max_ptlrpcds is obsolete.\n");
 815                 if (ptlrpcd_per_cpt_max == 0) {
 816                         ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
 817                         /* Round up if there is a remainder. */
 818                         if (max_ptlrpcds % ncpts != 0)
 819                                 ptlrpcd_per_cpt_max++;
 820                         CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
 821                               ptlrpcd_per_cpt_max);
 822                 } else {
 823                         CWARN("ptlrpd_per_cpt_max is also set!\n");
 824                 }
 825         }
 826
 827         /*
 828          * The ptlrpcd_bind_policy parameter is obsolete, but do
 829          * something sane if it has been tuned, and complain if
 830          * ptlrpcd_partner_group_size is also tuned.
 831          */
 832         if (ptlrpcd_bind_policy != 0) {
 833                 CWARN("ptlrpcd_bind_policy is obsolete.\n");
 834                 if (ptlrpcd_partner_group_size == 0) {
 835                         switch (ptlrpcd_bind_policy) {
 836                         case 1: /* PDB_POLICY_NONE */
 837                         case 2: /* PDB_POLICY_FULL */
 838                                 ptlrpcd_partner_group_size = 1;
 839                                 break;
 840                         case 3: /* PDB_POLICY_PAIR */
 841                                 ptlrpcd_partner_group_size = 2;
 842                                 break;
 843                         case 4: /* PDB_POLICY_NEIGHBOR */
 844 #ifdef CONFIG_NUMA
 845                                 ptlrpcd_partner_group_size = -1; /* CPT */
 846 #else
 847                                 ptlrpcd_partner_group_size = 3; /* Triplets */
 848 #endif
 849                                 break;
 850                         default: /* Illegal value, use the default. */
 851                                 ptlrpcd_partner_group_size = 2;
 852                                 break;
 853                         }
 854                         CWARN("Setting ptlrpcd_partner_group_size = %d\n",
 855                               ptlrpcd_partner_group_size);
 856                 } else {
 857                         CWARN("ptlrpcd_partner_group_size is also set!\n");
 858                 }
 859         }
 860
 861         if (ptlrpcd_partner_group_size == 0)
 862                 ptlrpcd_partner_group_size = 2;
 863         else if (ptlrpcd_partner_group_size < 0)
 864                 ptlrpcd_partner_group_size = -1;
 865         else if (ptlrpcd_per_cpt_max > 0 &&
 866                  ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
 867                 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
 868
 869         /*
 870          * Start the recovery thread first.
 871          */
 872         set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
 873         ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
 874         rc = ptlrpcd_start(&ptlrpcd_rcv);
 875         if (rc < 0)
 876                 GOTO(out, rc);
 877
 878         for (i = 0; i < ncpts; i++) {
 879                 if (cpts == NULL)
 880                         cpt = i;
 881                 else
 882                         cpt = cpts[i];
 883
 884                 nthreads = cfs_cpt_weight(cptable, cpt);
 885                 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
 886                         nthreads = ptlrpcd_per_cpt_max;
 887                 if (nthreads < 2)
 888                         nthreads = 2;
 889
 890                 if (ptlrpcd_partner_group_size <= 0) {
 891                         groupsize = nthreads;
 892                 } else if (nthreads <= ptlrpcd_partner_group_size) {
 893                         groupsize = nthreads;
 894                 } else {
 895                         groupsize = ptlrpcd_partner_group_size;
 896                         if (nthreads % groupsize != 0)
 897                                 nthreads += groupsize - (nthreads % groupsize);
 898                 }
 899
 900                 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 901                 OBD_CPT_ALLOC(pd, cptable, cpt, size);
 902
 903                 if (!pd)
 904                         GOTO(out, rc = -ENOMEM);
 905                 pd->pd_size      = size;
 906                 pd->pd_index     = i;
 907                 pd->pd_cpt       = cpt;
 908                 pd->pd_cursor    = 0;
 909                 pd->pd_nthreads  = nthreads;
 910                 pd->pd_groupsize = groupsize;
 911                 ptlrpcds[i] = pd;
 912
 913                 /*
 914                  * The ptlrpcd threads in a partner group can access
 915                  * each other's struct ptlrpcd_ctl, so these must be
 916                  * initialized before any thead is started.
 917                  */
 918                 for (j = 0; j < nthreads; j++) {
 919                         ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
 920                         rc = ptlrpcd_partners(pd, j);
 921                         if (rc < 0)
 922                                 GOTO(out, rc);
 923                 }
 924
 925                 /* XXX: We start nthreads ptlrpc daemons on this cpt.
 926                  *      Each of them can process any non-recovery
 927                  *      async RPC to improve overall async RPC
 928                  *      efficiency.
 929                  *
 930                  *      But there are some issues with async I/O RPCs
 931                  *      and async non-I/O RPCs processed in the same
 932                  *      set under some cases. The ptlrpcd may be
 933                  *      blocked by some async I/O RPC(s), then will
 934                  *      cause other async non-I/O RPC(s) can not be
 935                  *      processed in time.
 936                  *
 937                  *      Maybe we should distinguish blocked async RPCs
 938                  *      from non-blocked async RPCs, and process them
 939                  *      in different ptlrpcd sets to avoid unnecessary
 940                  *      dependency. But how to distribute async RPCs
 941                  *      load among all the ptlrpc daemons becomes
 942                  *      another trouble.
 943                  */
 944                 for (j = 0; j < nthreads; j++) {
 945                         rc = ptlrpcd_start(&pd->pd_threads[j]);
 946                         if (rc < 0)
 947                                 GOTO(out, rc);
 948                 }
 949         }
 950 out:
 951         if (rc != 0)
 952                 ptlrpcd_fini();
 953
 954         RETURN(rc);
 955 }
 956
 957 int ptlrpcd_addref(void)
 958 {
 959         int rc = 0;
 960
 961         ENTRY;
 962
 963         mutex_lock(&ptlrpcd_mutex);
 964         if (++ptlrpcd_users == 1) {
 965                 rc = ptlrpcd_init();
 966                 if (rc < 0)
 967                         ptlrpcd_users--;
 968         }
 969         mutex_unlock(&ptlrpcd_mutex);
 970         RETURN(rc);
 971 }
 972 EXPORT_SYMBOL(ptlrpcd_addref);
 973
 974 void ptlrpcd_decref(void)
 975 {
 976         mutex_lock(&ptlrpcd_mutex);
 977         if (--ptlrpcd_users == 0)
 978                 ptlrpcd_fini();
 979         mutex_unlock(&ptlrpcd_mutex);
 980 }
 981 EXPORT_SYMBOL(ptlrpcd_decref);
 982 /** @} ptlrpcd */