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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/ptlrpcd.c
  37  */
  38
  39 /** \defgroup ptlrpcd PortalRPC daemon
  40  *
  41  * ptlrpcd is a special thread with its own set where other user might add
  42  * requests when they don't want to wait for their completion.
  43  * PtlRPCD will take care of sending such requests and then processing their
  44  * replies and calling completion callbacks as necessary.
  45  * The callbacks are called directly from ptlrpcd context.
  46  * It is important to never significantly block (esp. on RPCs!) within such
  47  * completion handler or a deadlock might occur where ptlrpcd enters some
  48  * callback that attempts to send another RPC and wait for it to return,
  49  * during which time ptlrpcd is completely blocked, so e.g. if import
  50  * fails, recovery cannot progress because connection requests are also
  51  * sent by ptlrpcd.
  52  *
  53  * @{
  54  */
  55
  56 #define DEBUG_SUBSYSTEM S_RPC
  57
  58 #ifdef __KERNEL__
  59 # include <libcfs/libcfs.h>
  60 #else /* __KERNEL__ */
  61 # include <liblustre.h>
  62 # include <ctype.h>
  63 #endif
  64
  65 #include <lustre_net.h>
  66 # include <lustre_lib.h>
  67
  68 #include <lustre_ha.h>
  69 #include <obd_class.h>   /* for obd_zombie */
  70 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  71 #include <cl_object.h> /* cl_env_{get,put}() */
  72 #include <lprocfs_status.h>
  73
  74 #include "ptlrpc_internal.h"
  75
  76 struct ptlrpcd {
  77         int                pd_size;
  78         int                pd_index;
  79         int                pd_nthreads;
  80         struct ptlrpcd_ctl pd_thread_rcv;
  81         struct ptlrpcd_ctl pd_threads[0];
  82 };
  83
  84 #ifdef __KERNEL__
  85 static int max_ptlrpcds;
  86 CFS_MODULE_PARM(max_ptlrpcds, "i", int, 0644,
  87                 "Max ptlrpcd thread count to be started.");
  88
  89 static int ptlrpcd_bind_policy = PDB_POLICY_PAIR;
  90 CFS_MODULE_PARM(ptlrpcd_bind_policy, "i", int, 0644,
  91                 "Ptlrpcd threads binding mode.");
  92 #endif
  93 static struct ptlrpcd *ptlrpcds;
  94
  95 cfs_mutex_t ptlrpcd_mutex;
  96 static int ptlrpcd_users = 0;
  97
  98 void ptlrpcd_wake(struct ptlrpc_request *req)
  99 {
 100         struct ptlrpc_request_set *rq_set = req->rq_set;
 101
 102         LASSERT(rq_set != NULL);
 103
 104         cfs_waitq_signal(&rq_set->set_waitq);
 105 }
 106 EXPORT_SYMBOL(ptlrpcd_wake);
 107
 108 static struct ptlrpcd_ctl *
 109 ptlrpcd_select_pc(struct ptlrpc_request *req, pdl_policy_t policy, int index)
 110 {
 111         int idx = 0;
 112
 113         if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
 114                 return &ptlrpcds->pd_thread_rcv;
 115
 116 #ifdef __KERNEL__
 117         switch (policy) {
 118         case PDL_POLICY_SAME:
 119                 idx = cfs_smp_processor_id() % ptlrpcds->pd_nthreads;
 120                 break;
 121         case PDL_POLICY_LOCAL:
 122                 /* Before CPU partition patches available, process it the same
 123                  * as "PDL_POLICY_ROUND". */
 124 # ifdef CFS_CPU_MODE_NUMA
 125 # warning "fix this code to use new CPU partition APIs"
 126 # endif
 127                 /* Fall through to PDL_POLICY_ROUND until the CPU
 128                  * CPU partition patches are available. */
 129                 index = -1;
 130         case PDL_POLICY_PREFERRED:
 131                 if (index >= 0 && index < cfs_num_online_cpus()) {
 132                         idx = index % ptlrpcds->pd_nthreads;
 133                         break;
 134                 }
 135                 /* Fall through to PDL_POLICY_ROUND for bad index. */
 136         default:
 137                 /* Fall through to PDL_POLICY_ROUND for unknown policy. */
 138         case PDL_POLICY_ROUND:
 139                 /* We do not care whether it is strict load balance. */
 140                 idx = ptlrpcds->pd_index + 1;
 141                 if (idx == cfs_smp_processor_id())
 142                         idx++;
 143                 idx %= ptlrpcds->pd_nthreads;
 144                 ptlrpcds->pd_index = idx;
 145                 break;
 146         }
 147 #endif /* __KERNEL__ */
 148
 149         return &ptlrpcds->pd_threads[idx];
 150 }
 151
 152 /**
 153  * Move all request from an existing request set to the ptlrpcd queue.
 154  * All requests from the set must be in phase RQ_PHASE_NEW.
 155  */
 156 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 157 {
 158         cfs_list_t *tmp, *pos;
 159 #ifdef __KERNEL__
 160         struct ptlrpcd_ctl *pc;
 161         struct ptlrpc_request_set *new;
 162         int count, i;
 163
 164         pc = ptlrpcd_select_pc(NULL, PDL_POLICY_LOCAL, -1);
 165         new = pc->pc_set;
 166 #endif
 167
 168         cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 169                 struct ptlrpc_request *req =
 170                         cfs_list_entry(pos, struct ptlrpc_request,
 171                                        rq_set_chain);
 172
 173                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 174 #ifdef __KERNEL__
 175                 req->rq_set = new;
 176                 req->rq_queued_time = cfs_time_current();
 177 #else
 178                 cfs_list_del_init(&req->rq_set_chain);
 179                 req->rq_set = NULL;
 180                 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
 181                 cfs_atomic_dec(&set->set_remaining);
 182 #endif
 183         }
 184
 185 #ifdef __KERNEL__
 186         cfs_spin_lock(&new->set_new_req_lock);
 187         cfs_list_splice_init(&set->set_requests, &new->set_new_requests);
 188         i = cfs_atomic_read(&set->set_remaining);
 189         count = cfs_atomic_add_return(i, &new->set_new_count);
 190         cfs_atomic_set(&set->set_remaining, 0);
 191         cfs_spin_unlock(&new->set_new_req_lock);
 192         if (count == i) {
 193                 cfs_waitq_signal(&new->set_waitq);
 194
 195                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
 196                  *      guarantee the async RPC can be processed ASAP, we have
 197                  *      no other better choice. It maybe fixed in future. */
 198                 for (i = 0; i < pc->pc_npartners; i++)
 199                         cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
 200         }
 201 #endif
 202 }
 203 EXPORT_SYMBOL(ptlrpcd_add_rqset);
 204
 205 #ifdef __KERNEL__
 206 /**
 207  * Return transferred RPCs count.
 208  */
 209 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 210                                struct ptlrpc_request_set *src)
 211 {
 212         cfs_list_t *tmp, *pos;
 213         struct ptlrpc_request *req;
 214         int rc = 0;
 215
 216         cfs_spin_lock(&src->set_new_req_lock);
 217         if (likely(!cfs_list_empty(&src->set_new_requests))) {
 218                 cfs_list_for_each_safe(pos, tmp, &src->set_new_requests) {
 219                         req = cfs_list_entry(pos, struct ptlrpc_request,
 220                                              rq_set_chain);
 221                         req->rq_set = des;
 222                 }
 223                 cfs_list_splice_init(&src->set_new_requests,
 224                                      &des->set_requests);
 225                 rc = cfs_atomic_read(&src->set_new_count);
 226                 cfs_atomic_add(rc, &des->set_remaining);
 227                 cfs_atomic_set(&src->set_new_count, 0);
 228         }
 229         cfs_spin_unlock(&src->set_new_req_lock);
 230         return rc;
 231 }
 232 #endif
 233
 234 /**
 235  * Requests that are added to the ptlrpcd queue are sent via
 236  * ptlrpcd_check->ptlrpc_check_set().
 237  */
 238 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx)
 239 {
 240         struct ptlrpcd_ctl *pc;
 241         char jobid[JOBSTATS_JOBID_SIZE];
 242
 243         if (req->rq_reqmsg) {
 244                 lustre_get_jobid(jobid);
 245                 lustre_msg_set_jobid(req->rq_reqmsg, jobid);
 246         }
 247
 248         cfs_spin_lock(&req->rq_lock);
 249         if (req->rq_invalid_rqset) {
 250                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 251                                                      back_to_sleep, NULL);
 252
 253                 req->rq_invalid_rqset = 0;
 254                 cfs_spin_unlock(&req->rq_lock);
 255                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 256         } else if (req->rq_set) {
 257                 /* If we have a vaid "rq_set", just reuse it to avoid double
 258                  * linked. */
 259                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 260                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 261
 262                 /* ptlrpc_check_set will decrease the count */
 263                 cfs_atomic_inc(&req->rq_set->set_remaining);
 264                 cfs_spin_unlock(&req->rq_lock);
 265                 cfs_waitq_signal(&req->rq_set->set_waitq);
 266                 return;
 267         } else {
 268                 cfs_spin_unlock(&req->rq_lock);
 269         }
 270
 271         pc = ptlrpcd_select_pc(req, policy, idx);
 272
 273         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 274                   req, pc->pc_name, pc->pc_index);
 275
 276         ptlrpc_set_add_new_req(pc, req);
 277 }
 278 EXPORT_SYMBOL(ptlrpcd_add_req);
 279
 280 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 281 {
 282         cfs_atomic_inc(&set->set_refcount);
 283 }
 284
 285 /**
 286  * Check if there is more work to do on ptlrpcd set.
 287  * Returns 1 if yes.
 288  */
 289 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 290 {
 291         cfs_list_t *tmp, *pos;
 292         struct ptlrpc_request *req;
 293         struct ptlrpc_request_set *set = pc->pc_set;
 294         int rc = 0;
 295         int rc2;
 296         ENTRY;
 297
 298         if (cfs_atomic_read(&set->set_new_count)) {
 299                 cfs_spin_lock(&set->set_new_req_lock);
 300                 if (likely(!cfs_list_empty(&set->set_new_requests))) {
 301                         cfs_list_splice_init(&set->set_new_requests,
 302                                              &set->set_requests);
 303                         cfs_atomic_add(cfs_atomic_read(&set->set_new_count),
 304                                        &set->set_remaining);
 305                         cfs_atomic_set(&set->set_new_count, 0);
 306                         /*
 307                          * Need to calculate its timeout.
 308                          */
 309                         rc = 1;
 310                 }
 311                 cfs_spin_unlock(&set->set_new_req_lock);
 312         }
 313
 314         /* We should call lu_env_refill() before handling new requests to make
 315          * sure that env key the requests depending on really exists.
 316          */
 317         rc2 = lu_env_refill(env);
 318         if (rc2 != 0) {
 319                 /*
 320                  * XXX This is very awkward situation, because
 321                  * execution can neither continue (request
 322                  * interpreters assume that env is set up), nor repeat
 323                  * the loop (as this potentially results in a tight
 324                  * loop of -ENOMEM's).
 325                  *
 326                  * Fortunately, refill only ever does something when
 327                  * new modules are loaded, i.e., early during boot up.
 328                  */
 329                 CERROR("Failure to refill session: %d\n", rc2);
 330                 RETURN(rc);
 331         }
 332
 333         if (cfs_atomic_read(&set->set_remaining))
 334                 rc |= ptlrpc_check_set(env, set);
 335
 336         if (!cfs_list_empty(&set->set_requests)) {
 337                 /*
 338                  * XXX: our set never completes, so we prune the completed
 339                  * reqs after each iteration. boy could this be smarter.
 340                  */
 341                 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 342                         req = cfs_list_entry(pos, struct ptlrpc_request,
 343                                              rq_set_chain);
 344                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 345                                 continue;
 346
 347                         cfs_list_del_init(&req->rq_set_chain);
 348                         req->rq_set = NULL;
 349                         ptlrpc_req_finished(req);
 350                 }
 351         }
 352
 353         if (rc == 0) {
 354                 /*
 355                  * If new requests have been added, make sure to wake up.
 356                  */
 357                 rc = cfs_atomic_read(&set->set_new_count);
 358
 359 #ifdef __KERNEL__
 360                 /* If we have nothing to do, check whether we can take some
 361                  * work from our partner threads. */
 362                 if (rc == 0 && pc->pc_npartners > 0) {
 363                         struct ptlrpcd_ctl *partner;
 364                         struct ptlrpc_request_set *ps;
 365                         int first = pc->pc_cursor;
 366
 367                         do {
 368                                 partner = pc->pc_partners[pc->pc_cursor++];
 369                                 if (pc->pc_cursor >= pc->pc_npartners)
 370                                         pc->pc_cursor = 0;
 371                                 if (partner == NULL)
 372                                         continue;
 373
 374                                 cfs_spin_lock(&partner->pc_lock);
 375                                 ps = partner->pc_set;
 376                                 if (ps == NULL) {
 377                                         cfs_spin_unlock(&partner->pc_lock);
 378                                         continue;
 379                                 }
 380
 381                                 ptlrpc_reqset_get(ps);
 382                                 cfs_spin_unlock(&partner->pc_lock);
 383
 384                                 if (cfs_atomic_read(&ps->set_new_count)) {
 385                                         rc = ptlrpcd_steal_rqset(set, ps);
 386                                         if (rc > 0)
 387                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 388                                                        " async RPCs [%d->%d]\n",
 389                                                         rc, partner->pc_index,
 390                                                         pc->pc_index);
 391                                 }
 392                                 ptlrpc_reqset_put(ps);
 393                         } while (rc == 0 && pc->pc_cursor != first);
 394                 }
 395 #endif
 396         }
 397
 398         RETURN(rc);
 399 }
 400
 401 #ifdef __KERNEL__
 402 /**
 403  * Main ptlrpcd thread.
 404  * ptlrpc's code paths like to execute in process context, so we have this
 405  * thread which spins on a set which contains the rpcs and sends them.
 406  *
 407  */
 408 static int ptlrpcd(void *arg)
 409 {
 410         struct ptlrpcd_ctl *pc = arg;
 411         struct ptlrpc_request_set *set = pc->pc_set;
 412         struct lu_env env = { .le_ses = NULL };
 413         int rc, exit = 0;
 414         ENTRY;
 415
 416         cfs_daemonize_ctxt(pc->pc_name);
 417 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
 418         if (cfs_test_bit(LIOD_BIND, &pc->pc_flags)) {
 419                 int index = pc->pc_index;
 420
 421                 if (index >= 0 && index < cfs_num_possible_cpus()) {
 422                         while (!cpu_online(index)) {
 423                                 if (++index >= cfs_num_possible_cpus())
 424                                         index = 0;
 425                         }
 426                         cfs_set_cpus_allowed(cfs_current(),
 427                                      node_to_cpumask(cpu_to_node(index)));
 428                 }
 429         }
 430 #endif
 431         /*
 432          * XXX So far only "client" ptlrpcd uses an environment. In
 433          * the future, ptlrpcd thread (or a thread-set) has to given
 434          * an argument, describing its "scope".
 435          */
 436         rc = lu_context_init(&env.le_ctx,
 437                              LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 438         cfs_complete(&pc->pc_starting);
 439
 440         if (rc != 0)
 441                 RETURN(rc);
 442
 443         /*
 444          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 445          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 446          * there are requests in the set. New requests come in on the set's
 447          * new_req_list and ptlrpcd_check() moves them into the set.
 448          */
 449         do {
 450                 struct l_wait_info lwi;
 451                 int timeout;
 452
 453                 timeout = ptlrpc_set_next_timeout(set);
 454                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 455                                   ptlrpc_expired_set, set);
 456
 457                 lu_context_enter(&env.le_ctx);
 458                 l_wait_event(set->set_waitq,
 459                              ptlrpcd_check(&env, pc), &lwi);
 460                 lu_context_exit(&env.le_ctx);
 461
 462                 /*
 463                  * Abort inflight rpcs for forced stop case.
 464                  */
 465                 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
 466                         if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
 467                                 ptlrpc_abort_set(set);
 468                         exit++;
 469                 }
 470
 471                 /*
 472                  * Let's make one more loop to make sure that ptlrpcd_check()
 473                  * copied all raced new rpcs into the set so we can kill them.
 474                  */
 475         } while (exit < 2);
 476
 477         /*
 478          * Wait for inflight requests to drain.
 479          */
 480         if (!cfs_list_empty(&set->set_requests))
 481                 ptlrpc_set_wait(set);
 482         lu_context_fini(&env.le_ctx);
 483         cfs_complete(&pc->pc_finishing);
 484
 485         cfs_clear_bit(LIOD_START, &pc->pc_flags);
 486         cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
 487         cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
 488         cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 489         return 0;
 490 }
 491
 492 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 493  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 494  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 495  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 496  *      because of some CPU core(s) busy with other loads.
 497  *
 498  *      For example: "ls -l", some async RPCs for statahead are assigned to
 499  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 500  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 501  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 502  *      such case, the statahead async RPCs can not be processed in time, it is
 503  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 504  *      be better. But it breaks former data transfer policy.
 505  *
 506  *      So we shouldn't be blind for avoiding the data transfer. We make some
 507  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 508  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 509  *      core. The other part is for free mode, all the ptlrpcd threads in the
 510  *      part can be scheduled on any CPU core. We specify some partnership
 511  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 512  *      and the async RPC load within the partners are shared.
 513  *
 514  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 515  *      thread can be scheduled in time), and try to guarantee the async RPC
 516  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 517  *      on any CPU core).
 518  *
 519  *      As for how to specify the partnership between bound mode ptlrpcd
 520  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 521  *      <free bound> pair. In future, we can specify some more complex
 522  *      partnership based on the patches for CPU partition. But before such
 523  *      patches are available, we prefer to use the simplest one.
 524  */
 525 # ifdef CFS_CPU_MODE_NUMA
 526 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 527 # endif
 528 static int ptlrpcd_bind(int index, int max)
 529 {
 530         struct ptlrpcd_ctl *pc;
 531         int rc = 0;
 532 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 533         struct ptlrpcd_ctl *ppc;
 534         int node, i, pidx;
 535         cpumask_t mask;
 536 #endif
 537         ENTRY;
 538
 539         LASSERT(index <= max - 1);
 540         pc = &ptlrpcds->pd_threads[index];
 541         switch (ptlrpcd_bind_policy) {
 542         case PDB_POLICY_NONE:
 543                 pc->pc_npartners = -1;
 544                 break;
 545         case PDB_POLICY_FULL:
 546                 pc->pc_npartners = 0;
 547                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 548                 break;
 549         case PDB_POLICY_PAIR:
 550                 LASSERT(max % 2 == 0);
 551                 pc->pc_npartners = 1;
 552                 break;
 553         case PDB_POLICY_NEIGHBOR:
 554 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 555                 node = cpu_to_node(index);
 556                 mask = node_to_cpumask(node);
 557                 for (i = max; i < cfs_num_online_cpus(); i++)
 558                         cpu_clear(i, mask);
 559                 pc->pc_npartners = cpus_weight(mask) - 1;
 560                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 561 #else
 562                 LASSERT(max >= 3);
 563                 pc->pc_npartners = 2;
 564 #endif
 565                 break;
 566         default:
 567                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 568                 rc = -EINVAL;
 569         }
 570
 571         if (rc == 0 && pc->pc_npartners > 0) {
 572                 OBD_ALLOC(pc->pc_partners,
 573                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 574                 if (pc->pc_partners == NULL) {
 575                         pc->pc_npartners = 0;
 576                         rc = -ENOMEM;
 577                 } else {
 578                         switch (ptlrpcd_bind_policy) {
 579                         case PDB_POLICY_PAIR:
 580                                 if (index & 0x1) {
 581                                         cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 582                                         pc->pc_partners[0] = &ptlrpcds->
 583                                                 pd_threads[index - 1];
 584                                         ptlrpcds->pd_threads[index - 1].
 585                                                 pc_partners[0] = pc;
 586                                 }
 587                                 break;
 588                         case PDB_POLICY_NEIGHBOR:
 589 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 590                                 /* partners are cores in the same NUMA node.
 591                                  * setup partnership only with ptlrpcd threads
 592                                  * that are already initialized
 593                                  */
 594                                 for (pidx = 0, i = 0; i < index; i++) {
 595                                         if (cpu_isset(i, mask)) {
 596                                                 ppc = &ptlrpcds->pd_threads[i];
 597                                                 pc->pc_partners[pidx++] = ppc;
 598                                                 ppc->pc_partners[ppc->
 599                                                           pc_npartners++] = pc;
 600                                         }
 601                                 }
 602                                 /* adjust number of partners to the number
 603                                  * of partnership really setup */
 604                                 pc->pc_npartners = pidx;
 605 #else
 606                                 if (index & 0x1)
 607                                         cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 608                                 if (index > 0) {
 609                                         pc->pc_partners[0] = &ptlrpcds->
 610                                                 pd_threads[index - 1];
 611                                         ptlrpcds->pd_threads[index - 1].
 612                                                 pc_partners[1] = pc;
 613                                         if (index == max - 1) {
 614                                                 pc->pc_partners[1] =
 615                                                 &ptlrpcds->pd_threads[0];
 616                                                 ptlrpcds->pd_threads[0].
 617                                                 pc_partners[0] = pc;
 618                                         }
 619                                 }
 620 #endif
 621                                 break;
 622                         }
 623                 }
 624         }
 625
 626         RETURN(rc);
 627 }
 628
 629 #else /* !__KERNEL__ */
 630
 631 /**
 632  * In liblustre we do not have separate threads, so this function
 633  * is called from time to time all across common code to see
 634  * if something needs to be processed on ptlrpcd set.
 635  */
 636 int ptlrpcd_check_async_rpcs(void *arg)
 637 {
 638         struct ptlrpcd_ctl *pc = arg;
 639         int                 rc = 0;
 640
 641         /*
 642          * Single threaded!!
 643          */
 644         pc->pc_recurred++;
 645
 646         if (pc->pc_recurred == 1) {
 647                 rc = lu_env_refill(&pc->pc_env);
 648                 if (rc == 0) {
 649                         lu_context_enter(&pc->pc_env.le_ctx);
 650                         rc = ptlrpcd_check(&pc->pc_env, pc);
 651                         if (!rc)
 652                                 ptlrpc_expired_set(pc->pc_set);
 653                         /*
 654                          * XXX: send replay requests.
 655                          */
 656                         if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
 657                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 658                         lu_context_exit(&pc->pc_env.le_ctx);
 659                 }
 660         }
 661
 662         pc->pc_recurred--;
 663         return rc;
 664 }
 665
 666 int ptlrpcd_idle(void *arg)
 667 {
 668         struct ptlrpcd_ctl *pc = arg;
 669
 670         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
 671                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 672 }
 673
 674 #endif
 675
 676 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 677 {
 678         int rc;
 679         int env = 0;
 680         ENTRY;
 681
 682         /*
 683          * Do not allow start second thread for one pc.
 684          */
 685         if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 686                 CWARN("Starting second thread (%s) for same pc %p\n",
 687                        name, pc);
 688                 RETURN(0);
 689         }
 690
 691         pc->pc_index = index;
 692         cfs_init_completion(&pc->pc_starting);
 693         cfs_init_completion(&pc->pc_finishing);
 694         cfs_spin_lock_init(&pc->pc_lock);
 695         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 696         pc->pc_set = ptlrpc_prep_set();
 697         if (pc->pc_set == NULL)
 698                 GOTO(out, rc = -ENOMEM);
 699         /*
 700          * So far only "client" ptlrpcd uses an environment. In the future,
 701          * ptlrpcd thread (or a thread-set) has to be given an argument,
 702          * describing its "scope".
 703          */
 704         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 705         if (rc != 0)
 706                 GOTO(out, rc);
 707
 708         env = 1;
 709 #ifdef __KERNEL__
 710         if (index >= 0) {
 711                 rc = ptlrpcd_bind(index, max);
 712                 if (rc < 0)
 713                         GOTO(out, rc);
 714         }
 715
 716         rc = cfs_create_thread(ptlrpcd, pc, 0);
 717         if (rc < 0)
 718                 GOTO(out, rc);
 719
 720         rc = 0;
 721         cfs_wait_for_completion(&pc->pc_starting);
 722 #else
 723         pc->pc_wait_callback =
 724                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 725                                                  &ptlrpcd_check_async_rpcs, pc);
 726         pc->pc_idle_callback =
 727                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 728                                                  &ptlrpcd_idle, pc);
 729 #endif
 730 out:
 731         if (rc) {
 732 #ifdef __KERNEL__
 733                 if (pc->pc_set != NULL) {
 734                         struct ptlrpc_request_set *set = pc->pc_set;
 735
 736                         cfs_spin_lock(&pc->pc_lock);
 737                         pc->pc_set = NULL;
 738                         cfs_spin_unlock(&pc->pc_lock);
 739                         ptlrpc_set_destroy(set);
 740                 }
 741                 if (env != 0)
 742                         lu_context_fini(&pc->pc_env.le_ctx);
 743                 cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 744 #else
 745                 SET_BUT_UNUSED(env);
 746 #endif
 747                 cfs_clear_bit(LIOD_START, &pc->pc_flags);
 748         }
 749         RETURN(rc);
 750 }
 751
 752 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 753 {
 754        struct ptlrpc_request_set *set = pc->pc_set;
 755         ENTRY;
 756
 757         if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
 758                 CWARN("Thread for pc %p was not started\n", pc);
 759                 goto out;
 760         }
 761
 762         cfs_set_bit(LIOD_STOP, &pc->pc_flags);
 763         if (force)
 764                 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
 765         cfs_waitq_signal(&pc->pc_set->set_waitq);
 766 #ifdef __KERNEL__
 767         cfs_wait_for_completion(&pc->pc_finishing);
 768 #else
 769         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 770         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 771 #endif
 772         lu_context_fini(&pc->pc_env.le_ctx);
 773
 774         cfs_spin_lock(&pc->pc_lock);
 775         pc->pc_set = NULL;
 776         cfs_spin_unlock(&pc->pc_lock);
 777         ptlrpc_set_destroy(set);
 778
 779 out:
 780 #ifdef __KERNEL__
 781         if (pc->pc_npartners > 0) {
 782                 LASSERT(pc->pc_partners != NULL);
 783
 784                 OBD_FREE(pc->pc_partners,
 785                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 786                 pc->pc_partners = NULL;
 787         }
 788         pc->pc_npartners = 0;
 789 #endif
 790         EXIT;
 791 }
 792
 793 static void ptlrpcd_fini(void)
 794 {
 795         int i;
 796         ENTRY;
 797
 798         if (ptlrpcds != NULL) {
 799                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 800                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 801                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 802                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 803                 ptlrpcds = NULL;
 804         }
 805
 806         EXIT;
 807 }
 808
 809 static int ptlrpcd_init(void)
 810 {
 811         int nthreads = cfs_num_online_cpus();
 812         char name[16];
 813         int size, i = -1, j, rc = 0;
 814         ENTRY;
 815
 816 #ifdef __KERNEL__
 817         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 818                 nthreads = max_ptlrpcds;
 819         if (nthreads < 2)
 820                 nthreads = 2;
 821         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 822                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 823         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 824                 nthreads &= ~1; /* make sure it is even */
 825 #else
 826         nthreads = 1;
 827 #endif
 828
 829         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 830         OBD_ALLOC(ptlrpcds, size);
 831         if (ptlrpcds == NULL)
 832                 GOTO(out, rc = -ENOMEM);
 833
 834         snprintf(name, 15, "ptlrpcd_rcv");
 835         cfs_set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 836         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 837         if (rc < 0)
 838                 GOTO(out, rc);
 839
 840         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 841          *      non-recovery async RPC to improve overall async RPC efficiency.
 842          *
 843          *      But there are some issues with async I/O RPCs and async non-I/O
 844          *      RPCs processed in the same set under some cases. The ptlrpcd may
 845          *      be blocked by some async I/O RPC(s), then will cause other async
 846          *      non-I/O RPC(s) can not be processed in time.
 847          *
 848          *      Maybe we should distinguish blocked async RPCs from non-blocked
 849          *      async RPCs, and process them in different ptlrpcd sets to avoid
 850          *      unnecessary dependency. But how to distribute async RPCs load
 851          *      among all the ptlrpc daemons becomes another trouble. */
 852         for (i = 0; i < nthreads; i++) {
 853                 snprintf(name, 15, "ptlrpcd_%d", i);
 854                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 855                 if (rc < 0)
 856                         GOTO(out, rc);
 857         }
 858
 859         ptlrpcds->pd_size = size;
 860         ptlrpcds->pd_index = 0;
 861         ptlrpcds->pd_nthreads = nthreads;
 862
 863 out:
 864         if (rc != 0 && ptlrpcds != NULL) {
 865                 for (j = 0; j <= i; j++)
 866                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 867                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 868                 OBD_FREE(ptlrpcds, size);
 869                 ptlrpcds = NULL;
 870         }
 871
 872         RETURN(0);
 873 }
 874
 875 int ptlrpcd_addref(void)
 876 {
 877         int rc = 0;
 878         ENTRY;
 879
 880         cfs_mutex_lock(&ptlrpcd_mutex);
 881         if (++ptlrpcd_users == 1)
 882                 rc = ptlrpcd_init();
 883         cfs_mutex_unlock(&ptlrpcd_mutex);
 884         RETURN(rc);
 885 }
 886 EXPORT_SYMBOL(ptlrpcd_addref);
 887
 888 void ptlrpcd_decref(void)
 889 {
 890         cfs_mutex_lock(&ptlrpcd_mutex);
 891         if (--ptlrpcd_users == 0)
 892                 ptlrpcd_fini();
 893         cfs_mutex_unlock(&ptlrpcd_mutex);
 894 }
 895 EXPORT_SYMBOL(ptlrpcd_decref);
 896 /** @} ptlrpcd */