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