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
 242         if (req->rq_reqmsg)
 243                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
 244
 245         cfs_spin_lock(&req->rq_lock);
 246         if (req->rq_invalid_rqset) {
 247                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 248                                                      back_to_sleep, NULL);
 249
 250                 req->rq_invalid_rqset = 0;
 251                 cfs_spin_unlock(&req->rq_lock);
 252                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 253         } else if (req->rq_set) {
 254                 /* If we have a vaid "rq_set", just reuse it to avoid double
 255                  * linked. */
 256                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 257                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 258
 259                 /* ptlrpc_check_set will decrease the count */
 260                 cfs_atomic_inc(&req->rq_set->set_remaining);
 261                 cfs_spin_unlock(&req->rq_lock);
 262                 cfs_waitq_signal(&req->rq_set->set_waitq);
 263                 return;
 264         } else {
 265                 cfs_spin_unlock(&req->rq_lock);
 266         }
 267
 268         pc = ptlrpcd_select_pc(req, policy, idx);
 269
 270         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 271                   req, pc->pc_name, pc->pc_index);
 272
 273         ptlrpc_set_add_new_req(pc, req);
 274 }
 275 EXPORT_SYMBOL(ptlrpcd_add_req);
 276
 277 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 278 {
 279         cfs_atomic_inc(&set->set_refcount);
 280 }
 281
 282 /**
 283  * Check if there is more work to do on ptlrpcd set.
 284  * Returns 1 if yes.
 285  */
 286 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 287 {
 288         cfs_list_t *tmp, *pos;
 289         struct ptlrpc_request *req;
 290         struct ptlrpc_request_set *set = pc->pc_set;
 291         int rc = 0;
 292         int rc2;
 293         ENTRY;
 294
 295         if (cfs_atomic_read(&set->set_new_count)) {
 296                 cfs_spin_lock(&set->set_new_req_lock);
 297                 if (likely(!cfs_list_empty(&set->set_new_requests))) {
 298                         cfs_list_splice_init(&set->set_new_requests,
 299                                              &set->set_requests);
 300                         cfs_atomic_add(cfs_atomic_read(&set->set_new_count),
 301                                        &set->set_remaining);
 302                         cfs_atomic_set(&set->set_new_count, 0);
 303                         /*
 304                          * Need to calculate its timeout.
 305                          */
 306                         rc = 1;
 307                 }
 308                 cfs_spin_unlock(&set->set_new_req_lock);
 309         }
 310
 311         /* We should call lu_env_refill() before handling new requests to make
 312          * sure that env key the requests depending on really exists.
 313          */
 314         rc2 = lu_env_refill(env);
 315         if (rc2 != 0) {
 316                 /*
 317                  * XXX This is very awkward situation, because
 318                  * execution can neither continue (request
 319                  * interpreters assume that env is set up), nor repeat
 320                  * the loop (as this potentially results in a tight
 321                  * loop of -ENOMEM's).
 322                  *
 323                  * Fortunately, refill only ever does something when
 324                  * new modules are loaded, i.e., early during boot up.
 325                  */
 326                 CERROR("Failure to refill session: %d\n", rc2);
 327                 RETURN(rc);
 328         }
 329
 330         if (cfs_atomic_read(&set->set_remaining))
 331                 rc |= ptlrpc_check_set(env, set);
 332
 333         if (!cfs_list_empty(&set->set_requests)) {
 334                 /*
 335                  * XXX: our set never completes, so we prune the completed
 336                  * reqs after each iteration. boy could this be smarter.
 337                  */
 338                 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 339                         req = cfs_list_entry(pos, struct ptlrpc_request,
 340                                              rq_set_chain);
 341                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 342                                 continue;
 343
 344                         cfs_list_del_init(&req->rq_set_chain);
 345                         req->rq_set = NULL;
 346                         ptlrpc_req_finished(req);
 347                 }
 348         }
 349
 350         if (rc == 0) {
 351                 /*
 352                  * If new requests have been added, make sure to wake up.
 353                  */
 354                 rc = cfs_atomic_read(&set->set_new_count);
 355
 356 #ifdef __KERNEL__
 357                 /* If we have nothing to do, check whether we can take some
 358                  * work from our partner threads. */
 359                 if (rc == 0 && pc->pc_npartners > 0) {
 360                         struct ptlrpcd_ctl *partner;
 361                         struct ptlrpc_request_set *ps;
 362                         int first = pc->pc_cursor;
 363
 364                         do {
 365                                 partner = pc->pc_partners[pc->pc_cursor++];
 366                                 if (pc->pc_cursor >= pc->pc_npartners)
 367                                         pc->pc_cursor = 0;
 368                                 if (partner == NULL)
 369                                         continue;
 370
 371                                 cfs_spin_lock(&partner->pc_lock);
 372                                 ps = partner->pc_set;
 373                                 if (ps == NULL) {
 374                                         cfs_spin_unlock(&partner->pc_lock);
 375                                         continue;
 376                                 }
 377
 378                                 ptlrpc_reqset_get(ps);
 379                                 cfs_spin_unlock(&partner->pc_lock);
 380
 381                                 if (cfs_atomic_read(&ps->set_new_count)) {
 382                                         rc = ptlrpcd_steal_rqset(set, ps);
 383                                         if (rc > 0)
 384                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 385                                                        " async RPCs [%d->%d]\n",
 386                                                         rc, partner->pc_index,
 387                                                         pc->pc_index);
 388                                 }
 389                                 ptlrpc_reqset_put(ps);
 390                         } while (rc == 0 && pc->pc_cursor != first);
 391                 }
 392 #endif
 393         }
 394
 395         RETURN(rc);
 396 }
 397
 398 #ifdef __KERNEL__
 399 /**
 400  * Main ptlrpcd thread.
 401  * ptlrpc's code paths like to execute in process context, so we have this
 402  * thread which spins on a set which contains the rpcs and sends them.
 403  *
 404  */
 405 static int ptlrpcd(void *arg)
 406 {
 407         struct ptlrpcd_ctl *pc = arg;
 408         struct ptlrpc_request_set *set = pc->pc_set;
 409         struct lu_env env = { .le_ses = NULL };
 410         int rc, exit = 0;
 411         ENTRY;
 412
 413         cfs_daemonize_ctxt(pc->pc_name);
 414 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
 415         if (cfs_test_bit(LIOD_BIND, &pc->pc_flags)) {
 416                 int index = pc->pc_index;
 417
 418                 if (index >= 0 && index < cfs_num_possible_cpus()) {
 419                         while (!cpu_online(index)) {
 420                                 if (++index >= cfs_num_possible_cpus())
 421                                         index = 0;
 422                         }
 423                         cfs_set_cpus_allowed(cfs_current(),
 424                                      node_to_cpumask(cpu_to_node(index)));
 425                 }
 426         }
 427 #endif
 428         /*
 429          * XXX So far only "client" ptlrpcd uses an environment. In
 430          * the future, ptlrpcd thread (or a thread-set) has to given
 431          * an argument, describing its "scope".
 432          */
 433         rc = lu_context_init(&env.le_ctx,
 434                              LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 435         cfs_complete(&pc->pc_starting);
 436
 437         if (rc != 0)
 438                 RETURN(rc);
 439
 440         /*
 441          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 442          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 443          * there are requests in the set. New requests come in on the set's
 444          * new_req_list and ptlrpcd_check() moves them into the set.
 445          */
 446         do {
 447                 struct l_wait_info lwi;
 448                 int timeout;
 449
 450                 timeout = ptlrpc_set_next_timeout(set);
 451                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 452                                   ptlrpc_expired_set, set);
 453
 454                 lu_context_enter(&env.le_ctx);
 455                 l_wait_event(set->set_waitq,
 456                              ptlrpcd_check(&env, pc), &lwi);
 457                 lu_context_exit(&env.le_ctx);
 458
 459                 /*
 460                  * Abort inflight rpcs for forced stop case.
 461                  */
 462                 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
 463                         if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
 464                                 ptlrpc_abort_set(set);
 465                         exit++;
 466                 }
 467
 468                 /*
 469                  * Let's make one more loop to make sure that ptlrpcd_check()
 470                  * copied all raced new rpcs into the set so we can kill them.
 471                  */
 472         } while (exit < 2);
 473
 474         /*
 475          * Wait for inflight requests to drain.
 476          */
 477         if (!cfs_list_empty(&set->set_requests))
 478                 ptlrpc_set_wait(set);
 479         lu_context_fini(&env.le_ctx);
 480         cfs_complete(&pc->pc_finishing);
 481
 482         cfs_clear_bit(LIOD_START, &pc->pc_flags);
 483         cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
 484         cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
 485         cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 486         return 0;
 487 }
 488
 489 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 490  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 491  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 492  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 493  *      because of some CPU core(s) busy with other loads.
 494  *
 495  *      For example: "ls -l", some async RPCs for statahead are assigned to
 496  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 497  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 498  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 499  *      such case, the statahead async RPCs can not be processed in time, it is
 500  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 501  *      be better. But it breaks former data transfer policy.
 502  *
 503  *      So we shouldn't be blind for avoiding the data transfer. We make some
 504  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 505  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 506  *      core. The other part is for free mode, all the ptlrpcd threads in the
 507  *      part can be scheduled on any CPU core. We specify some partnership
 508  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 509  *      and the async RPC load within the partners are shared.
 510  *
 511  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 512  *      thread can be scheduled in time), and try to guarantee the async RPC
 513  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 514  *      on any CPU core).
 515  *
 516  *      As for how to specify the partnership between bound mode ptlrpcd
 517  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 518  *      <free bound> pair. In future, we can specify some more complex
 519  *      partnership based on the patches for CPU partition. But before such
 520  *      patches are available, we prefer to use the simplest one.
 521  */
 522 # ifdef CFS_CPU_MODE_NUMA
 523 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 524 # endif
 525 static int ptlrpcd_bind(int index, int max)
 526 {
 527         struct ptlrpcd_ctl *pc;
 528         int rc = 0;
 529 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 530         struct ptlrpcd_ctl *ppc;
 531         int node, i, pidx;
 532         cpumask_t mask;
 533 #endif
 534         ENTRY;
 535
 536         LASSERT(index <= max - 1);
 537         pc = &ptlrpcds->pd_threads[index];
 538         switch (ptlrpcd_bind_policy) {
 539         case PDB_POLICY_NONE:
 540                 pc->pc_npartners = -1;
 541                 break;
 542         case PDB_POLICY_FULL:
 543                 pc->pc_npartners = 0;
 544                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 545                 break;
 546         case PDB_POLICY_PAIR:
 547                 LASSERT(max % 2 == 0);
 548                 pc->pc_npartners = 1;
 549                 break;
 550         case PDB_POLICY_NEIGHBOR:
 551 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 552                 node = cpu_to_node(index);
 553                 mask = node_to_cpumask(node);
 554                 for (i = max; i < cfs_num_online_cpus(); i++)
 555                         cpu_clear(i, mask);
 556                 pc->pc_npartners = cpus_weight(mask) - 1;
 557                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 558 #else
 559                 LASSERT(max >= 3);
 560                 pc->pc_npartners = 2;
 561 #endif
 562                 break;
 563         default:
 564                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 565                 rc = -EINVAL;
 566         }
 567
 568         if (rc == 0 && pc->pc_npartners > 0) {
 569                 OBD_ALLOC(pc->pc_partners,
 570                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 571                 if (pc->pc_partners == NULL) {
 572                         pc->pc_npartners = 0;
 573                         rc = -ENOMEM;
 574                 } else {
 575                         switch (ptlrpcd_bind_policy) {
 576                         case PDB_POLICY_PAIR:
 577                                 if (index & 0x1) {
 578                                         cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 579                                         pc->pc_partners[0] = &ptlrpcds->
 580                                                 pd_threads[index - 1];
 581                                         ptlrpcds->pd_threads[index - 1].
 582                                                 pc_partners[0] = pc;
 583                                 }
 584                                 break;
 585                         case PDB_POLICY_NEIGHBOR:
 586 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 587                                 /* partners are cores in the same NUMA node.
 588                                  * setup partnership only with ptlrpcd threads
 589                                  * that are already initialized
 590                                  */
 591                                 for (pidx = 0, i = 0; i < index; i++) {
 592                                         if (cpu_isset(i, mask)) {
 593                                                 ppc = &ptlrpcds->pd_threads[i];
 594                                                 pc->pc_partners[pidx++] = ppc;
 595                                                 ppc->pc_partners[ppc->
 596                                                           pc_npartners++] = pc;
 597                                         }
 598                                 }
 599                                 /* adjust number of partners to the number
 600                                  * of partnership really setup */
 601                                 pc->pc_npartners = pidx;
 602 #else
 603                                 if (index & 0x1)
 604                                         cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 605                                 if (index > 0) {
 606                                         pc->pc_partners[0] = &ptlrpcds->
 607                                                 pd_threads[index - 1];
 608                                         ptlrpcds->pd_threads[index - 1].
 609                                                 pc_partners[1] = pc;
 610                                         if (index == max - 1) {
 611                                                 pc->pc_partners[1] =
 612                                                 &ptlrpcds->pd_threads[0];
 613                                                 ptlrpcds->pd_threads[0].
 614                                                 pc_partners[0] = pc;
 615                                         }
 616                                 }
 617 #endif
 618                                 break;
 619                         }
 620                 }
 621         }
 622
 623         RETURN(rc);
 624 }
 625
 626 #else /* !__KERNEL__ */
 627
 628 /**
 629  * In liblustre we do not have separate threads, so this function
 630  * is called from time to time all across common code to see
 631  * if something needs to be processed on ptlrpcd set.
 632  */
 633 int ptlrpcd_check_async_rpcs(void *arg)
 634 {
 635         struct ptlrpcd_ctl *pc = arg;
 636         int                 rc = 0;
 637
 638         /*
 639          * Single threaded!!
 640          */
 641         pc->pc_recurred++;
 642
 643         if (pc->pc_recurred == 1) {
 644                 rc = lu_env_refill(&pc->pc_env);
 645                 if (rc == 0) {
 646                         lu_context_enter(&pc->pc_env.le_ctx);
 647                         rc = ptlrpcd_check(&pc->pc_env, pc);
 648                         if (!rc)
 649                                 ptlrpc_expired_set(pc->pc_set);
 650                         /*
 651                          * XXX: send replay requests.
 652                          */
 653                         if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
 654                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 655                         lu_context_exit(&pc->pc_env.le_ctx);
 656                 }
 657         }
 658
 659         pc->pc_recurred--;
 660         return rc;
 661 }
 662
 663 int ptlrpcd_idle(void *arg)
 664 {
 665         struct ptlrpcd_ctl *pc = arg;
 666
 667         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
 668                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 669 }
 670
 671 #endif
 672
 673 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 674 {
 675         int rc;
 676         int env = 0;
 677         ENTRY;
 678
 679         /*
 680          * Do not allow start second thread for one pc.
 681          */
 682         if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 683                 CWARN("Starting second thread (%s) for same pc %p\n",
 684                        name, pc);
 685                 RETURN(0);
 686         }
 687
 688         pc->pc_index = index;
 689         cfs_init_completion(&pc->pc_starting);
 690         cfs_init_completion(&pc->pc_finishing);
 691         cfs_spin_lock_init(&pc->pc_lock);
 692         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 693         pc->pc_set = ptlrpc_prep_set();
 694         if (pc->pc_set == NULL)
 695                 GOTO(out, rc = -ENOMEM);
 696         /*
 697          * So far only "client" ptlrpcd uses an environment. In the future,
 698          * ptlrpcd thread (or a thread-set) has to be given an argument,
 699          * describing its "scope".
 700          */
 701         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 702         if (rc != 0)
 703                 GOTO(out, rc);
 704
 705         env = 1;
 706 #ifdef __KERNEL__
 707         if (index >= 0) {
 708                 rc = ptlrpcd_bind(index, max);
 709                 if (rc < 0)
 710                         GOTO(out, rc);
 711         }
 712
 713         rc = cfs_create_thread(ptlrpcd, pc, 0);
 714         if (rc < 0)
 715                 GOTO(out, rc);
 716
 717         rc = 0;
 718         cfs_wait_for_completion(&pc->pc_starting);
 719 #else
 720         pc->pc_wait_callback =
 721                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 722                                                  &ptlrpcd_check_async_rpcs, pc);
 723         pc->pc_idle_callback =
 724                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 725                                                  &ptlrpcd_idle, pc);
 726 #endif
 727 out:
 728         if (rc) {
 729 #ifdef __KERNEL__
 730                 if (pc->pc_set != NULL) {
 731                         struct ptlrpc_request_set *set = pc->pc_set;
 732
 733                         cfs_spin_lock(&pc->pc_lock);
 734                         pc->pc_set = NULL;
 735                         cfs_spin_unlock(&pc->pc_lock);
 736                         ptlrpc_set_destroy(set);
 737                 }
 738                 if (env != 0)
 739                         lu_context_fini(&pc->pc_env.le_ctx);
 740                 cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 741 #else
 742                 SET_BUT_UNUSED(env);
 743 #endif
 744                 cfs_clear_bit(LIOD_START, &pc->pc_flags);
 745         }
 746         RETURN(rc);
 747 }
 748
 749 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 750 {
 751        struct ptlrpc_request_set *set = pc->pc_set;
 752         ENTRY;
 753
 754         if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
 755                 CWARN("Thread for pc %p was not started\n", pc);
 756                 goto out;
 757         }
 758
 759         cfs_set_bit(LIOD_STOP, &pc->pc_flags);
 760         if (force)
 761                 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
 762         cfs_waitq_signal(&pc->pc_set->set_waitq);
 763 #ifdef __KERNEL__
 764         cfs_wait_for_completion(&pc->pc_finishing);
 765 #else
 766         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 767         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 768 #endif
 769         lu_context_fini(&pc->pc_env.le_ctx);
 770
 771         cfs_spin_lock(&pc->pc_lock);
 772         pc->pc_set = NULL;
 773         cfs_spin_unlock(&pc->pc_lock);
 774         ptlrpc_set_destroy(set);
 775
 776 out:
 777 #ifdef __KERNEL__
 778         if (pc->pc_npartners > 0) {
 779                 LASSERT(pc->pc_partners != NULL);
 780
 781                 OBD_FREE(pc->pc_partners,
 782                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 783                 pc->pc_partners = NULL;
 784         }
 785         pc->pc_npartners = 0;
 786 #endif
 787         EXIT;
 788 }
 789
 790 static void ptlrpcd_fini(void)
 791 {
 792         int i;
 793         ENTRY;
 794
 795         if (ptlrpcds != NULL) {
 796                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 797                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 798                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 799                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 800                 ptlrpcds = NULL;
 801         }
 802
 803         EXIT;
 804 }
 805
 806 static int ptlrpcd_init(void)
 807 {
 808         int nthreads = cfs_num_online_cpus();
 809         char name[16];
 810         int size, i = -1, j, rc = 0;
 811         ENTRY;
 812
 813 #ifdef __KERNEL__
 814         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 815                 nthreads = max_ptlrpcds;
 816         if (nthreads < 2)
 817                 nthreads = 2;
 818         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 819                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 820         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 821                 nthreads &= ~1; /* make sure it is even */
 822 #else
 823         nthreads = 1;
 824 #endif
 825
 826         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 827         OBD_ALLOC(ptlrpcds, size);
 828         if (ptlrpcds == NULL)
 829                 GOTO(out, rc = -ENOMEM);
 830
 831         snprintf(name, 15, "ptlrpcd_rcv");
 832         cfs_set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 833         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 834         if (rc < 0)
 835                 GOTO(out, rc);
 836
 837         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 838          *      non-recovery async RPC to improve overall async RPC efficiency.
 839          *
 840          *      But there are some issues with async I/O RPCs and async non-I/O
 841          *      RPCs processed in the same set under some cases. The ptlrpcd may
 842          *      be blocked by some async I/O RPC(s), then will cause other async
 843          *      non-I/O RPC(s) can not be processed in time.
 844          *
 845          *      Maybe we should distinguish blocked async RPCs from non-blocked
 846          *      async RPCs, and process them in different ptlrpcd sets to avoid
 847          *      unnecessary dependency. But how to distribute async RPCs load
 848          *      among all the ptlrpc daemons becomes another trouble. */
 849         for (i = 0; i < nthreads; i++) {
 850                 snprintf(name, 15, "ptlrpcd_%d", i);
 851                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 852                 if (rc < 0)
 853                         GOTO(out, rc);
 854         }
 855
 856         ptlrpcds->pd_size = size;
 857         ptlrpcds->pd_index = 0;
 858         ptlrpcds->pd_nthreads = nthreads;
 859
 860 out:
 861         if (rc != 0 && ptlrpcds != NULL) {
 862                 for (j = 0; j <= i; j++)
 863                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 864                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 865                 OBD_FREE(ptlrpcds, size);
 866                 ptlrpcds = NULL;
 867         }
 868
 869         RETURN(0);
 870 }
 871
 872 int ptlrpcd_addref(void)
 873 {
 874         int rc = 0;
 875         ENTRY;
 876
 877         cfs_mutex_lock(&ptlrpcd_mutex);
 878         if (++ptlrpcd_users == 1)
 879                 rc = ptlrpcd_init();
 880         cfs_mutex_unlock(&ptlrpcd_mutex);
 881         RETURN(rc);
 882 }
 883 EXPORT_SYMBOL(ptlrpcd_addref);
 884
 885 void ptlrpcd_decref(void)
 886 {
 887         cfs_mutex_lock(&ptlrpcd_mutex);
 888         if (--ptlrpcd_users == 0)
 889                 ptlrpcd_fini();
 890         cfs_mutex_unlock(&ptlrpcd_mutex);
 891 }
 892 EXPORT_SYMBOL(ptlrpcd_decref);
 893 /** @} ptlrpcd */