lustre/ptlrpc/ptlrpcd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/ptlrpc/ptlrpcd.c
  40  */
  41
  42 /** \defgroup ptlrpcd PortalRPC daemon
  43  *
  44  * ptlrpcd is a special thread with its own set where other user might add
  45  * requests when they don't want to wait for their completion.
  46  * PtlRPCD will take care of sending such requests and then processing their
  47  * replies and calling completion callbacks as necessary.
  48  * The callbacks are called directly from ptlrpcd context.
  49  * It is important to never significantly block (esp. on RPCs!) within such
  50  * completion handler or a deadlock might occur where ptlrpcd enters some
  51  * callback that attempts to send another RPC and wait for it to return,
  52  * during which time ptlrpcd is completely blocked, so e.g. if import
  53  * fails, recovery cannot progress because connection requests are also
  54  * sent by ptlrpcd.
  55  *
  56  * @{
  57  */
  58
  59 #define DEBUG_SUBSYSTEM S_RPC
  60
  61 #ifdef __KERNEL__
  62 # include <libcfs/libcfs.h>
  63 #else /* __KERNEL__ */
  64 # include <liblustre.h>
  65 # include <ctype.h>
  66 #endif
  67
  68 #include <lustre_net.h>
  69 # include <lustre_lib.h>
  70
  71 #include <lustre_ha.h>
  72 #include <obd_class.h>   /* for obd_zombie */
  73 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  74 #include <cl_object.h> /* cl_env_{get,put}() */
  75 #include <lprocfs_status.h>
  76
  77 #include "ptlrpc_internal.h"
  78
  79 struct ptlrpcd {
  80         int                pd_size;
  81         int                pd_index;
  82         int                pd_nthreads;
  83         struct ptlrpcd_ctl pd_thread_rcv;
  84         struct ptlrpcd_ctl pd_threads[0];
  85 };
  86
  87 #ifdef __KERNEL__
  88 static int max_ptlrpcds;
  89 CFS_MODULE_PARM(max_ptlrpcds, "i", int, 0644,
  90                 "Max ptlrpcd thread count to be started.");
  91
  92 static int ptlrpcd_bind_policy = PDB_POLICY_PAIR;
  93 CFS_MODULE_PARM(ptlrpcd_bind_policy, "i", int, 0644,
  94                 "Ptlrpcd threads binding mode.");
  95 #endif
  96 static struct ptlrpcd *ptlrpcds;
  97
  98 cfs_semaphore_t ptlrpcd_sem;
  99 static int ptlrpcd_users = 0;
 100
 101 void ptlrpcd_wake(struct ptlrpc_request *req)
 102 {
 103         struct ptlrpc_request_set *rq_set = req->rq_set;
 104
 105         LASSERT(rq_set != NULL);
 106
 107         cfs_waitq_signal(&rq_set->set_waitq);
 108 }
 109
 110 static struct ptlrpcd_ctl *
 111 ptlrpcd_select_pc(struct ptlrpc_request *req, pdl_policy_t policy, int index)
 112 {
 113         int idx = 0;
 114
 115         if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
 116                 return &ptlrpcds->pd_thread_rcv;
 117
 118 #ifdef __KERNEL__
 119         switch (policy) {
 120         case PDL_POLICY_SAME:
 121                 idx = cfs_smp_processor_id() % ptlrpcds->pd_nthreads;
 122                 break;
 123         case PDL_POLICY_LOCAL:
 124                 /* Before CPU partition patches available, process it the same
 125                  * as "PDL_POLICY_ROUND". */
 126 # ifdef CFS_CPU_MODE_NUMA
 127 # warning "fix this code to use new CPU partition APIs"
 128 # endif
 129                 /* Fall through to PDL_POLICY_ROUND until the CPU
 130                  * CPU partition patches are available. */
 131                 index = -1;
 132         case PDL_POLICY_PREFERRED:
 133                 if (index >= 0 && index < cfs_num_online_cpus()) {
 134                         idx = index % ptlrpcds->pd_nthreads;
 135                         break;
 136                 }
 137                 /* Fall through to PDL_POLICY_ROUND for bad index. */
 138         default:
 139                 /* Fall through to PDL_POLICY_ROUND for unknown policy. */
 140         case PDL_POLICY_ROUND:
 141                 /* We do not care whether it is strict load balance. */
 142                 idx = ptlrpcds->pd_index + 1;
 143                 if (idx == cfs_smp_processor_id())
 144                         idx++;
 145                 idx %= ptlrpcds->pd_nthreads;
 146                 ptlrpcds->pd_index = idx;
 147                 break;
 148         }
 149 #endif /* __KERNEL__ */
 150
 151         return &ptlrpcds->pd_threads[idx];
 152 }
 153
 154 /**
 155  * Move all request from an existing request set to the ptlrpcd queue.
 156  * All requests from the set must be in phase RQ_PHASE_NEW.
 157  */
 158 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 159 {
 160         cfs_list_t *tmp, *pos;
 161 #ifdef __KERNEL__
 162         struct ptlrpcd_ctl *pc;
 163         struct ptlrpc_request_set *new;
 164         int count, i;
 165
 166         pc = ptlrpcd_select_pc(NULL, PDL_POLICY_LOCAL, -1);
 167         new = pc->pc_set;
 168 #endif
 169
 170         cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 171                 struct ptlrpc_request *req =
 172                         cfs_list_entry(pos, struct ptlrpc_request,
 173                                        rq_set_chain);
 174
 175                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 176 #ifdef __KERNEL__
 177                 req->rq_set = new;
 178                 req->rq_queued_time = cfs_time_current();
 179 #else
 180                 cfs_list_del_init(&req->rq_set_chain);
 181                 req->rq_set = NULL;
 182                 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
 183                 cfs_atomic_dec(&set->set_remaining);
 184 #endif
 185         }
 186
 187 #ifdef __KERNEL__
 188         cfs_spin_lock(&new->set_new_req_lock);
 189         cfs_list_splice_init(&set->set_requests, &new->set_new_requests);
 190         i = cfs_atomic_read(&set->set_remaining);
 191         count = cfs_atomic_add_return(i, &new->set_new_count);
 192         cfs_atomic_set(&set->set_remaining, 0);
 193         cfs_spin_unlock(&new->set_new_req_lock);
 194         if (count == i) {
 195                 cfs_waitq_signal(&new->set_waitq);
 196
 197                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
 198                  *      guarantee the async RPC can be processed ASAP, we have
 199                  *      no other better choice. It maybe fixed in future. */
 200                 for (i = 0; i < pc->pc_npartners; i++)
 201                         cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
 202         }
 203 #endif
 204 }
 205 EXPORT_SYMBOL(ptlrpcd_add_rqset);
 206
 207 #ifdef __KERNEL__
 208 /**
 209  * Return transferred RPCs count.
 210  */
 211 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 212                                struct ptlrpc_request_set *src)
 213 {
 214         cfs_list_t *tmp, *pos;
 215         struct ptlrpc_request *req;
 216         int rc = 0;
 217
 218         cfs_spin_lock(&src->set_new_req_lock);
 219         if (likely(!cfs_list_empty(&src->set_new_requests))) {
 220                 cfs_list_for_each_safe(pos, tmp, &src->set_new_requests) {
 221                         req = cfs_list_entry(pos, struct ptlrpc_request,
 222                                              rq_set_chain);
 223                         req->rq_set = des;
 224                 }
 225                 cfs_list_splice_init(&src->set_new_requests,
 226                                      &des->set_requests);
 227                 rc = cfs_atomic_read(&src->set_new_count);
 228                 cfs_atomic_add(rc, &des->set_remaining);
 229                 cfs_atomic_set(&src->set_new_count, 0);
 230         }
 231         cfs_spin_unlock(&src->set_new_req_lock);
 232         return rc;
 233 }
 234 #endif
 235
 236 /**
 237  * Requests that are added to the ptlrpcd queue are sent via
 238  * ptlrpcd_check->ptlrpc_check_set().
 239  */
 240 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx)
 241 {
 242         struct ptlrpcd_ctl *pc;
 243
 244         cfs_spin_lock(&req->rq_lock);
 245         if (req->rq_invalid_rqset) {
 246                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 247                                                      back_to_sleep, NULL);
 248
 249                 req->rq_invalid_rqset = 0;
 250                 cfs_spin_unlock(&req->rq_lock);
 251                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 252         } else if (req->rq_set) {
 253                 /* If we have a vaid "rq_set", just reuse it to avoid double
 254                  * linked. */
 255                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 256                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 257
 258                 /* ptlrpc_check_set will decrease the count */
 259                 cfs_atomic_inc(&req->rq_set->set_remaining);
 260                 cfs_spin_unlock(&req->rq_lock);
 261                 cfs_waitq_signal(&req->rq_set->set_waitq);
 262                 return;
 263         } else {
 264                 cfs_spin_unlock(&req->rq_lock);
 265         }
 266
 267         pc = ptlrpcd_select_pc(req, policy, idx);
 268
 269         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 270                   req, pc->pc_name, pc->pc_index);
 271
 272         ptlrpc_set_add_new_req(pc, req);
 273 }
 274
 275 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 276 {
 277         cfs_atomic_inc(&set->set_refcount);
 278 }
 279
 280 /**
 281  * Check if there is more work to do on ptlrpcd set.
 282  * Returns 1 if yes.
 283  */
 284 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 285 {
 286         cfs_list_t *tmp, *pos;
 287         struct ptlrpc_request *req;
 288         struct ptlrpc_request_set *set = pc->pc_set;
 289         int rc = 0;
 290         int rc2;
 291         ENTRY;
 292
 293         if (cfs_atomic_read(&set->set_new_count)) {
 294                 cfs_spin_lock(&set->set_new_req_lock);
 295                 if (likely(!cfs_list_empty(&set->set_new_requests))) {
 296                         cfs_list_splice_init(&set->set_new_requests,
 297                                              &set->set_requests);
 298                         cfs_atomic_add(cfs_atomic_read(&set->set_new_count),
 299                                        &set->set_remaining);
 300                         cfs_atomic_set(&set->set_new_count, 0);
 301                         /*
 302                          * Need to calculate its timeout.
 303                          */
 304                         rc = 1;
 305                 }
 306                 cfs_spin_unlock(&set->set_new_req_lock);
 307         }
 308
 309         /* We should call lu_env_refill() before handling new requests to make
 310          * sure that env key the requests depending on really exists.
 311          */
 312         rc2 = lu_env_refill(env);
 313         if (rc2 != 0) {
 314                 /*
 315                  * XXX This is very awkward situation, because
 316                  * execution can neither continue (request
 317                  * interpreters assume that env is set up), nor repeat
 318                  * the loop (as this potentially results in a tight
 319                  * loop of -ENOMEM's).
 320                  *
 321                  * Fortunately, refill only ever does something when
 322                  * new modules are loaded, i.e., early during boot up.
 323                  */
 324                 CERROR("Failure to refill session: %d\n", rc2);
 325                 RETURN(rc);
 326         }
 327
 328         if (cfs_atomic_read(&set->set_remaining))
 329                 rc |= ptlrpc_check_set(env, set);
 330
 331         if (!cfs_list_empty(&set->set_requests)) {
 332                 /*
 333                  * XXX: our set never completes, so we prune the completed
 334                  * reqs after each iteration. boy could this be smarter.
 335                  */
 336                 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 337                         req = cfs_list_entry(pos, struct ptlrpc_request,
 338                                              rq_set_chain);
 339                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 340                                 continue;
 341
 342                         cfs_list_del_init(&req->rq_set_chain);
 343                         req->rq_set = NULL;
 344                         ptlrpc_req_finished(req);
 345                 }
 346         }
 347
 348         if (rc == 0) {
 349                 /*
 350                  * If new requests have been added, make sure to wake up.
 351                  */
 352                 rc = cfs_atomic_read(&set->set_new_count);
 353
 354 #ifdef __KERNEL__
 355                 /* If we have nothing to do, check whether we can take some
 356                  * work from our partner threads. */
 357                 if (rc == 0 && pc->pc_npartners > 0) {
 358                         struct ptlrpcd_ctl *partner;
 359                         struct ptlrpc_request_set *ps;
 360                         int first = pc->pc_cursor;
 361
 362                         do {
 363                                 partner = pc->pc_partners[pc->pc_cursor++];
 364                                 if (pc->pc_cursor >= pc->pc_npartners)
 365                                         pc->pc_cursor = 0;
 366                                 if (partner == NULL)
 367                                         continue;
 368
 369                                 cfs_spin_lock(&partner->pc_lock);
 370                                 ps = partner->pc_set;
 371                                 if (ps == NULL) {
 372                                         cfs_spin_unlock(&partner->pc_lock);
 373                                         continue;
 374                                 }
 375
 376                                 ptlrpc_reqset_get(ps);
 377                                 cfs_spin_unlock(&partner->pc_lock);
 378
 379                                 if (cfs_atomic_read(&ps->set_new_count)) {
 380                                         rc = ptlrpcd_steal_rqset(set, ps);
 381                                         if (rc > 0)
 382                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 383                                                        " async RPCs [%d->%d]\n",
 384                                                         rc, pc->pc_index,
 385                                                         partner->pc_index);
 386                                 }
 387                                 ptlrpc_reqset_put(ps);
 388                         } while (rc == 0 && pc->pc_cursor != first);
 389                 }
 390 #endif
 391         }
 392
 393         RETURN(rc);
 394 }
 395
 396 #ifdef __KERNEL__
 397 /**
 398  * Main ptlrpcd thread.
 399  * ptlrpc's code paths like to execute in process context, so we have this
 400  * thread which spins on a set which contains the rpcs and sends them.
 401  *
 402  */
 403 static int ptlrpcd(void *arg)
 404 {
 405         struct ptlrpcd_ctl *pc = arg;
 406         struct ptlrpc_request_set *set = pc->pc_set;
 407         struct lu_env env = { .le_ses = NULL };
 408         int rc, exit = 0;
 409         ENTRY;
 410
 411         cfs_daemonize_ctxt(pc->pc_name);
 412 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
 413         if (cfs_test_bit(LIOD_BIND, &pc->pc_flags)) {
 414                 int index = pc->pc_index;
 415
 416                 if (index >= 0 && index < cfs_num_possible_cpus()) {
 417                         while (!cfs_cpu_online(index)) {
 418                                 if (++index >= cfs_num_possible_cpus())
 419                                         index = 0;
 420                         }
 421                         cfs_set_cpus_allowed(cfs_current(),
 422                                      node_to_cpumask(cpu_to_node(index)));
 423                 }
 424         }
 425 #endif
 426         /*
 427          * XXX So far only "client" ptlrpcd uses an environment. In
 428          * the future, ptlrpcd thread (or a thread-set) has to given
 429          * an argument, describing its "scope".
 430          */
 431         rc = lu_context_init(&env.le_ctx,
 432                              LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 433         cfs_complete(&pc->pc_starting);
 434
 435         if (rc != 0)
 436                 RETURN(rc);
 437
 438         /*
 439          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 440          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 441          * there are requests in the set. New requests come in on the set's
 442          * new_req_list and ptlrpcd_check() moves them into the set.
 443          */
 444         do {
 445                 struct l_wait_info lwi;
 446                 int timeout;
 447
 448                 timeout = ptlrpc_set_next_timeout(set);
 449                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 450                                   ptlrpc_expired_set, set);
 451
 452                 lu_context_enter(&env.le_ctx);
 453                 l_wait_event(set->set_waitq,
 454                              ptlrpcd_check(&env, pc), &lwi);
 455                 lu_context_exit(&env.le_ctx);
 456
 457                 /*
 458                  * Abort inflight rpcs for forced stop case.
 459                  */
 460                 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
 461                         if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
 462                                 ptlrpc_abort_set(set);
 463                         exit++;
 464                 }
 465
 466                 /*
 467                  * Let's make one more loop to make sure that ptlrpcd_check()
 468                  * copied all raced new rpcs into the set so we can kill them.
 469                  */
 470         } while (exit < 2);
 471
 472         /*
 473          * Wait for inflight requests to drain.
 474          */
 475         if (!cfs_list_empty(&set->set_requests))
 476                 ptlrpc_set_wait(set);
 477         lu_context_fini(&env.le_ctx);
 478         cfs_complete(&pc->pc_finishing);
 479
 480         cfs_clear_bit(LIOD_START, &pc->pc_flags);
 481         cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
 482         cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
 483         cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 484         return 0;
 485 }
 486
 487 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 488  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 489  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 490  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 491  *      because of some CPU core(s) busy with other loads.
 492  *
 493  *      For example: "ls -l", some async RPCs for statahead are assigned to
 494  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 495  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 496  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 497  *      such case, the statahead async RPCs can not be processed in time, it is
 498  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 499  *      be better. But it breaks former data transfer policy.
 500  *
 501  *      So we shouldn't be blind for avoiding the data transfer. We make some
 502  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 503  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 504  *      core. The other part is for free mode, all the ptlrpcd threads in the
 505  *      part can be scheduled on any CPU core. We specify some partnership
 506  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 507  *      and the async RPC load within the partners are shared.
 508  *
 509  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 510  *      thread can be scheduled in time), and try to guarantee the async RPC
 511  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 512  *      on any CPU core).
 513  *
 514  *      As for how to specify the partnership between bound mode ptlrpcd
 515  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 516  *      <free bound> pair. In future, we can specify some more complex
 517  *      partnership based on the patches for CPU partition. But before such
 518  *      patches are available, we prefer to use the simplest one.
 519  */
 520 # ifdef CFS_CPU_MODE_NUMA
 521 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 522 # endif
 523 static int ptlrpcd_bind(int index, int max)
 524 {
 525         struct ptlrpcd_ctl *pc;
 526         int rc = 0;
 527         ENTRY;
 528
 529         LASSERT(index <= max - 1);
 530         pc = &ptlrpcds->pd_threads[index];
 531         switch (ptlrpcd_bind_policy) {
 532         case PDB_POLICY_NONE:
 533                 pc->pc_npartners = -1;
 534                 break;
 535         case PDB_POLICY_FULL:
 536                 pc->pc_npartners = 0;
 537                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 538                 break;
 539         case PDB_POLICY_PAIR:
 540                 LASSERT(max % 2 == 0);
 541                 pc->pc_npartners = 1;
 542                 break;
 543         case PDB_POLICY_NEIGHBOR:
 544                 LASSERT(max >= 3);
 545                 pc->pc_npartners = 2;
 546                 break;
 547         default:
 548                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 549                 rc = -EINVAL;
 550         }
 551
 552         if (rc == 0 && pc->pc_npartners > 0) {
 553                 OBD_ALLOC(pc->pc_partners,
 554                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 555                 if (pc->pc_partners == NULL) {
 556                         pc->pc_npartners = 0;
 557                         rc = -ENOMEM;
 558                 } else {
 559                         if (index & 0x1)
 560                                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 561
 562                         switch (ptlrpcd_bind_policy) {
 563                         case PDB_POLICY_PAIR:
 564                                 if (index & 0x1) {
 565                                         pc->pc_partners[0] = &ptlrpcds->
 566                                                 pd_threads[index - 1];
 567                                         ptlrpcds->pd_threads[index - 1].
 568                                                 pc_partners[0] = pc;
 569                                 }
 570                                 break;
 571                         case PDB_POLICY_NEIGHBOR:
 572                                 if (index > 0) {
 573                                         pc->pc_partners[0] = &ptlrpcds->
 574                                                 pd_threads[index - 1];
 575                                         ptlrpcds->pd_threads[index - 1].
 576                                                 pc_partners[1] = pc;
 577                                         if (index == max - 1) {
 578                                                 pc->pc_partners[1] =
 579                                                 &ptlrpcds->pd_threads[0];
 580                                                 ptlrpcds->pd_threads[0].
 581                                                 pc_partners[0] = pc;
 582                                         }
 583                                 }
 584                                 break;
 585                         }
 586                 }
 587         }
 588
 589         RETURN(rc);
 590 }
 591
 592 #else /* !__KERNEL__ */
 593
 594 /**
 595  * In liblustre we do not have separate threads, so this function
 596  * is called from time to time all across common code to see
 597  * if something needs to be processed on ptlrpcd set.
 598  */
 599 int ptlrpcd_check_async_rpcs(void *arg)
 600 {
 601         struct ptlrpcd_ctl *pc = arg;
 602         int                 rc = 0;
 603
 604         /*
 605          * Single threaded!!
 606          */
 607         pc->pc_recurred++;
 608
 609         if (pc->pc_recurred == 1) {
 610                 rc = lu_env_refill(&pc->pc_env);
 611                 if (rc == 0) {
 612                         lu_context_enter(&pc->pc_env.le_ctx);
 613                         rc = ptlrpcd_check(&pc->pc_env, pc);
 614                         if (!rc)
 615                                 ptlrpc_expired_set(pc->pc_set);
 616                         /*
 617                          * XXX: send replay requests.
 618                          */
 619                         if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
 620                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 621                         lu_context_exit(&pc->pc_env.le_ctx);
 622                 }
 623         }
 624
 625         pc->pc_recurred--;
 626         return rc;
 627 }
 628
 629 int ptlrpcd_idle(void *arg)
 630 {
 631         struct ptlrpcd_ctl *pc = arg;
 632
 633         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
 634                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 635 }
 636
 637 #endif
 638
 639 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 640 {
 641         int rc;
 642         int env = 0;
 643         ENTRY;
 644
 645         /*
 646          * Do not allow start second thread for one pc.
 647          */
 648         if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 649                 CWARN("Starting second thread (%s) for same pc %p\n",
 650                        name, pc);
 651                 RETURN(0);
 652         }
 653
 654         pc->pc_index = index;
 655         cfs_init_completion(&pc->pc_starting);
 656         cfs_init_completion(&pc->pc_finishing);
 657         cfs_spin_lock_init(&pc->pc_lock);
 658         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 659         pc->pc_set = ptlrpc_prep_set();
 660         if (pc->pc_set == NULL)
 661                 GOTO(out, rc = -ENOMEM);
 662         /*
 663          * So far only "client" ptlrpcd uses an environment. In the future,
 664          * ptlrpcd thread (or a thread-set) has to be given an argument,
 665          * describing its "scope".
 666          */
 667         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 668         if (rc != 0)
 669                 GOTO(out, rc);
 670
 671         env = 1;
 672 #ifdef __KERNEL__
 673         if (index >= 0) {
 674                 rc = ptlrpcd_bind(index, max);
 675                 if (rc < 0)
 676                         GOTO(out, rc);
 677         }
 678
 679         rc = cfs_create_thread(ptlrpcd, pc, 0);
 680         if (rc < 0)
 681                 GOTO(out, rc);
 682
 683         rc = 0;
 684         cfs_wait_for_completion(&pc->pc_starting);
 685 #else
 686         pc->pc_wait_callback =
 687                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 688                                                  &ptlrpcd_check_async_rpcs, pc);
 689         pc->pc_idle_callback =
 690                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 691                                                  &ptlrpcd_idle, pc);
 692 #endif
 693 out:
 694         if (rc) {
 695 #ifdef __KERNEL__
 696                 if (pc->pc_set != NULL) {
 697                         struct ptlrpc_request_set *set = pc->pc_set;
 698
 699                         cfs_spin_lock(&pc->pc_lock);
 700                         pc->pc_set = NULL;
 701                         cfs_spin_unlock(&pc->pc_lock);
 702                         ptlrpc_set_destroy(set);
 703                 }
 704                 if (env != 0)
 705                         lu_context_fini(&pc->pc_env.le_ctx);
 706                 cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 707 #else
 708                 SET_BUT_UNUSED(env);
 709 #endif
 710                 cfs_clear_bit(LIOD_START, &pc->pc_flags);
 711         }
 712         RETURN(rc);
 713 }
 714
 715 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 716 {
 717        struct ptlrpc_request_set *set = pc->pc_set;
 718         ENTRY;
 719
 720         if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
 721                 CWARN("Thread for pc %p was not started\n", pc);
 722                 goto out;
 723         }
 724
 725         cfs_set_bit(LIOD_STOP, &pc->pc_flags);
 726         if (force)
 727                 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
 728         cfs_waitq_signal(&pc->pc_set->set_waitq);
 729 #ifdef __KERNEL__
 730         cfs_wait_for_completion(&pc->pc_finishing);
 731 #else
 732         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 733         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 734 #endif
 735         lu_context_fini(&pc->pc_env.le_ctx);
 736
 737         cfs_spin_lock(&pc->pc_lock);
 738         pc->pc_set = NULL;
 739         cfs_spin_unlock(&pc->pc_lock);
 740         ptlrpc_set_destroy(set);
 741
 742 out:
 743 #ifdef __KERNEL__
 744         if (pc->pc_npartners > 0) {
 745                 LASSERT(pc->pc_partners != NULL);
 746
 747                 OBD_FREE(pc->pc_partners,
 748                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 749                 pc->pc_partners = NULL;
 750         }
 751         pc->pc_npartners = 0;
 752 #endif
 753         EXIT;
 754 }
 755
 756 static void ptlrpcd_fini(void)
 757 {
 758         int i;
 759         ENTRY;
 760
 761         if (ptlrpcds != NULL) {
 762                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 763                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 764                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 765                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 766                 ptlrpcds = NULL;
 767         }
 768
 769         EXIT;
 770 }
 771
 772 static int ptlrpcd_init(void)
 773 {
 774         int nthreads = cfs_num_online_cpus();
 775         char name[16];
 776         int size, i = -1, j, rc = 0;
 777         ENTRY;
 778
 779 #ifdef __KERNEL__
 780         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 781                 nthreads = max_ptlrpcds;
 782         if (nthreads < 2)
 783                 nthreads = 2;
 784         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 785                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 786         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 787                 nthreads &= ~1; /* make sure it is even */
 788 #else
 789         nthreads = 1;
 790 #endif
 791
 792         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 793         OBD_ALLOC(ptlrpcds, size);
 794         if (ptlrpcds == NULL)
 795                 GOTO(out, rc = -ENOMEM);
 796
 797         snprintf(name, 15, "ptlrpcd_rcv");
 798         cfs_set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 799         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 800         if (rc < 0)
 801                 GOTO(out, rc);
 802
 803         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 804          *      non-recovery async RPC to improve overall async RPC efficiency.
 805          *
 806          *      But there are some issues with async I/O RPCs and async non-I/O
 807          *      RPCs processed in the same set under some cases. The ptlrpcd may
 808          *      be blocked by some async I/O RPC(s), then will cause other async
 809          *      non-I/O RPC(s) can not be processed in time.
 810          *
 811          *      Maybe we should distinguish blocked async RPCs from non-blocked
 812          *      async RPCs, and process them in different ptlrpcd sets to avoid
 813          *      unnecessary dependency. But how to distribute async RPCs load
 814          *      among all the ptlrpc daemons becomes another trouble. */
 815         for (i = 0; i < nthreads; i++) {
 816                 snprintf(name, 15, "ptlrpcd_%d", i);
 817                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 818                 if (rc < 0)
 819                         GOTO(out, rc);
 820         }
 821
 822         ptlrpcds->pd_size = size;
 823         ptlrpcds->pd_index = 0;
 824         ptlrpcds->pd_nthreads = nthreads;
 825
 826 out:
 827         if (rc != 0 && ptlrpcds != NULL) {
 828                 for (j = 0; j <= i; j++)
 829                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 830                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 831                 OBD_FREE(ptlrpcds, size);
 832                 ptlrpcds = NULL;
 833         }
 834
 835         RETURN(0);
 836 }
 837
 838 int ptlrpcd_addref(void)
 839 {
 840         int rc = 0;
 841         ENTRY;
 842
 843         cfs_mutex_down(&ptlrpcd_sem);
 844         if (++ptlrpcd_users == 1)
 845                 rc = ptlrpcd_init();
 846         cfs_mutex_up(&ptlrpcd_sem);
 847         RETURN(rc);
 848 }
 849
 850 void ptlrpcd_decref(void)
 851 {
 852         cfs_mutex_down(&ptlrpcd_sem);
 853         if (--ptlrpcd_users == 0)
 854                 ptlrpcd_fini();
 855         cfs_mutex_up(&ptlrpcd_sem);
 856 }
 857 /** @} ptlrpcd */