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