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, 2014, 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 <linux/kthread.h>
  59 #include <libcfs/libcfs.h>
  60 #include <lustre_net.h>
  61 #include <lustre_lib.h>
  62 #include <lustre_ha.h>
  63 #include <obd_class.h>   /* for obd_zombie */
  64 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  65 #include <cl_object.h> /* cl_env_{get,put}() */
  66 #include <lprocfs_status.h>
  67
  68 #include "ptlrpc_internal.h"
  69
  70 struct ptlrpcd {
  71         int                pd_size;
  72         int                pd_index;
  73         int                pd_nthreads;
  74         struct ptlrpcd_ctl pd_thread_rcv;
  75         struct ptlrpcd_ctl pd_threads[0];
  76 };
  77
  78 static int max_ptlrpcds;
  79 CFS_MODULE_PARM(max_ptlrpcds, "i", int, 0644,
  80                 "Max ptlrpcd thread count to be started.");
  81
  82 static int ptlrpcd_bind_policy = PDB_POLICY_PAIR;
  83 CFS_MODULE_PARM(ptlrpcd_bind_policy, "i", int, 0644,
  84                 "Ptlrpcd threads binding mode.");
  85 static struct ptlrpcd *ptlrpcds;
  86
  87 struct mutex ptlrpcd_mutex;
  88 static int ptlrpcd_users = 0;
  89
  90 void ptlrpcd_wake(struct ptlrpc_request *req)
  91 {
  92         struct ptlrpc_request_set *set = req->rq_set;
  93
  94         LASSERT(set != NULL);
  95         wake_up(&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
 182 /**
 183  * Return transferred RPCs count.
 184  */
 185 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 186                                struct ptlrpc_request_set *src)
 187 {
 188         struct list_head *tmp, *pos;
 189         struct ptlrpc_request *req;
 190         int rc = 0;
 191
 192         spin_lock(&src->set_new_req_lock);
 193         if (likely(!list_empty(&src->set_new_requests))) {
 194                 list_for_each_safe(pos, tmp, &src->set_new_requests) {
 195                         req = list_entry(pos, struct ptlrpc_request,
 196                                          rq_set_chain);
 197                         req->rq_set = des;
 198                 }
 199                 list_splice_init(&src->set_new_requests,
 200                                  &des->set_requests);
 201                 rc = atomic_read(&src->set_new_count);
 202                 atomic_add(rc, &des->set_remaining);
 203                 atomic_set(&src->set_new_count, 0);
 204         }
 205         spin_unlock(&src->set_new_req_lock);
 206         return rc;
 207 }
 208
 209 /**
 210  * Requests that are added to the ptlrpcd queue are sent via
 211  * ptlrpcd_check->ptlrpc_check_set().
 212  */
 213 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx)
 214 {
 215         struct ptlrpcd_ctl *pc;
 216
 217         if (req->rq_reqmsg)
 218                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
 219
 220         spin_lock(&req->rq_lock);
 221         if (req->rq_invalid_rqset) {
 222                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 223                                                      back_to_sleep, NULL);
 224
 225                 req->rq_invalid_rqset = 0;
 226                 spin_unlock(&req->rq_lock);
 227                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 228         } else if (req->rq_set) {
 229                 /* If we have a vaid "rq_set", just reuse it to avoid double
 230                  * linked. */
 231                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 232                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 233
 234                 /* ptlrpc_check_set will decrease the count */
 235                 atomic_inc(&req->rq_set->set_remaining);
 236                 spin_unlock(&req->rq_lock);
 237                 wake_up(&req->rq_set->set_waitq);
 238                 return;
 239         } else {
 240                 spin_unlock(&req->rq_lock);
 241         }
 242
 243         pc = ptlrpcd_select_pc(req, policy, idx);
 244
 245         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 246                   req, pc->pc_name, pc->pc_index);
 247
 248         ptlrpc_set_add_new_req(pc, req);
 249 }
 250 EXPORT_SYMBOL(ptlrpcd_add_req);
 251
 252 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 253 {
 254         atomic_inc(&set->set_refcount);
 255 }
 256
 257 /**
 258  * Check if there is more work to do on ptlrpcd set.
 259  * Returns 1 if yes.
 260  */
 261 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 262 {
 263         struct list_head *tmp, *pos;
 264         struct ptlrpc_request *req;
 265         struct ptlrpc_request_set *set = pc->pc_set;
 266         int rc = 0;
 267         int rc2;
 268         ENTRY;
 269
 270         if (atomic_read(&set->set_new_count)) {
 271                 spin_lock(&set->set_new_req_lock);
 272                 if (likely(!list_empty(&set->set_new_requests))) {
 273                         list_splice_init(&set->set_new_requests,
 274                                              &set->set_requests);
 275                         atomic_add(atomic_read(&set->set_new_count),
 276                                    &set->set_remaining);
 277                         atomic_set(&set->set_new_count, 0);
 278                         /*
 279                          * Need to calculate its timeout.
 280                          */
 281                         rc = 1;
 282                 }
 283                 spin_unlock(&set->set_new_req_lock);
 284         }
 285
 286         /* We should call lu_env_refill() before handling new requests to make
 287          * sure that env key the requests depending on really exists.
 288          */
 289         rc2 = lu_env_refill(env);
 290         if (rc2 != 0) {
 291                 /*
 292                  * XXX This is very awkward situation, because
 293                  * execution can neither continue (request
 294                  * interpreters assume that env is set up), nor repeat
 295                  * the loop (as this potentially results in a tight
 296                  * loop of -ENOMEM's).
 297                  *
 298                  * Fortunately, refill only ever does something when
 299                  * new modules are loaded, i.e., early during boot up.
 300                  */
 301                 CERROR("Failure to refill session: %d\n", rc2);
 302                 RETURN(rc);
 303         }
 304
 305         if (atomic_read(&set->set_remaining))
 306                 rc |= ptlrpc_check_set(env, set);
 307
 308         /* NB: ptlrpc_check_set has already moved complted request at the
 309          * head of seq::set_requests */
 310         list_for_each_safe(pos, tmp, &set->set_requests) {
 311                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 312                 if (req->rq_phase != RQ_PHASE_COMPLETE)
 313                         break;
 314
 315                 list_del_init(&req->rq_set_chain);
 316                 req->rq_set = NULL;
 317                 ptlrpc_req_finished(req);
 318         }
 319
 320         if (rc == 0) {
 321                 /*
 322                  * If new requests have been added, make sure to wake up.
 323                  */
 324                 rc = atomic_read(&set->set_new_count);
 325
 326                 /* If we have nothing to do, check whether we can take some
 327                  * work from our partner threads. */
 328                 if (rc == 0 && pc->pc_npartners > 0) {
 329                         struct ptlrpcd_ctl *partner;
 330                         struct ptlrpc_request_set *ps;
 331                         int first = pc->pc_cursor;
 332
 333                         do {
 334                                 partner = pc->pc_partners[pc->pc_cursor++];
 335                                 if (pc->pc_cursor >= pc->pc_npartners)
 336                                         pc->pc_cursor = 0;
 337                                 if (partner == NULL)
 338                                         continue;
 339
 340                                 spin_lock(&partner->pc_lock);
 341                                 ps = partner->pc_set;
 342                                 if (ps == NULL) {
 343                                         spin_unlock(&partner->pc_lock);
 344                                         continue;
 345                                 }
 346
 347                                 ptlrpc_reqset_get(ps);
 348                                 spin_unlock(&partner->pc_lock);
 349
 350                                 if (atomic_read(&ps->set_new_count)) {
 351                                         rc = ptlrpcd_steal_rqset(set, ps);
 352                                         if (rc > 0)
 353                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 354                                                        " async RPCs [%d->%d]\n",
 355                                                        rc, partner->pc_index,
 356                                                        pc->pc_index);
 357                                 }
 358                                 ptlrpc_reqset_put(ps);
 359                         } while (rc == 0 && pc->pc_cursor != first);
 360                 }
 361         }
 362
 363         RETURN(rc);
 364 }
 365
 366 /**
 367  * Main ptlrpcd thread.
 368  * ptlrpc's code paths like to execute in process context, so we have this
 369  * thread which spins on a set which contains the rpcs and sends them.
 370  *
 371  */
 372 static int ptlrpcd(void *arg)
 373 {
 374         struct ptlrpcd_ctl *pc = arg;
 375         struct ptlrpc_request_set *set = pc->pc_set;
 376         struct lu_context ses = { 0 };
 377         struct lu_env env = { .le_ses = &ses };
 378         int rc, exit = 0;
 379         ENTRY;
 380
 381         unshare_fs_struct();
 382 #if defined(CONFIG_SMP)
 383         if (test_bit(LIOD_BIND, &pc->pc_flags)) {
 384                 int index = pc->pc_index;
 385
 386                 if (index >= 0 && index < num_possible_cpus()) {
 387                         while (!cpu_online(index)) {
 388                                 if (++index >= num_possible_cpus())
 389                                         index = 0;
 390                         }
 391                         set_cpus_allowed_ptr(current,
 392                                      cpumask_of_node(cpu_to_node(index)));
 393                 }
 394         }
 395 #endif
 396         /* Both client and server (MDT/OST) may use the environment. */
 397         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
 398                                           LCT_CL_THREAD | LCT_REMEMBER |
 399                                           LCT_NOREF);
 400         if (rc == 0) {
 401                 rc = lu_context_init(env.le_ses,
 402                                      LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
 403                 if (rc != 0)
 404                         lu_context_fini(&env.le_ctx);
 405         }
 406         complete(&pc->pc_starting);
 407
 408         if (rc != 0)
 409                 RETURN(rc);
 410
 411         /*
 412          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 413          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 414          * there are requests in the set. New requests come in on the set's
 415          * new_req_list and ptlrpcd_check() moves them into the set.
 416          */
 417         do {
 418                 struct l_wait_info lwi;
 419                 int timeout;
 420
 421                 timeout = ptlrpc_set_next_timeout(set);
 422                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 423                                   ptlrpc_expired_set, set);
 424
 425                 lu_context_enter(&env.le_ctx);
 426                 lu_context_enter(env.le_ses);
 427                 l_wait_event(set->set_waitq, ptlrpcd_check(&env, pc), &lwi);
 428                 lu_context_exit(&env.le_ctx);
 429                 lu_context_exit(env.le_ses);
 430
 431                 /*
 432                  * Abort inflight rpcs for forced stop case.
 433                  */
 434                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 435                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 436                                 ptlrpc_abort_set(set);
 437                         exit++;
 438                 }
 439
 440                 /*
 441                  * Let's make one more loop to make sure that ptlrpcd_check()
 442                  * copied all raced new rpcs into the set so we can kill them.
 443                  */
 444         } while (exit < 2);
 445
 446         /*
 447          * Wait for inflight requests to drain.
 448          */
 449         if (!list_empty(&set->set_requests))
 450                 ptlrpc_set_wait(set);
 451         lu_context_fini(&env.le_ctx);
 452         lu_context_fini(env.le_ses);
 453
 454         complete(&pc->pc_finishing);
 455
 456         return 0;
 457 }
 458
 459 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 460  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 461  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 462  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 463  *      because of some CPU core(s) busy with other loads.
 464  *
 465  *      For example: "ls -l", some async RPCs for statahead are assigned to
 466  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 467  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 468  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 469  *      such case, the statahead async RPCs can not be processed in time, it is
 470  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 471  *      be better. But it breaks former data transfer policy.
 472  *
 473  *      So we shouldn't be blind for avoiding the data transfer. We make some
 474  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 475  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 476  *      core. The other part is for free mode, all the ptlrpcd threads in the
 477  *      part can be scheduled on any CPU core. We specify some partnership
 478  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 479  *      and the async RPC load within the partners are shared.
 480  *
 481  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 482  *      thread can be scheduled in time), and try to guarantee the async RPC
 483  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 484  *      on any CPU core).
 485  *
 486  *      As for how to specify the partnership between bound mode ptlrpcd
 487  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 488  *      <free bound> pair. In future, we can specify some more complex
 489  *      partnership based on the patches for CPU partition. But before such
 490  *      patches are available, we prefer to use the simplest one.
 491  */
 492 # ifdef CFS_CPU_MODE_NUMA
 493 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 494 # endif
 495 static int ptlrpcd_bind(int index, int max)
 496 {
 497         struct ptlrpcd_ctl *pc;
 498         int rc = 0;
 499 #if defined(CONFIG_NUMA)
 500         cpumask_t mask;
 501 #endif
 502         ENTRY;
 503
 504         LASSERT(index <= max - 1);
 505         pc = &ptlrpcds->pd_threads[index];
 506         switch (ptlrpcd_bind_policy) {
 507         case PDB_POLICY_NONE:
 508                 pc->pc_npartners = -1;
 509                 break;
 510         case PDB_POLICY_FULL:
 511                 pc->pc_npartners = 0;
 512                 set_bit(LIOD_BIND, &pc->pc_flags);
 513                 break;
 514         case PDB_POLICY_PAIR:
 515                 LASSERT(max % 2 == 0);
 516                 pc->pc_npartners = 1;
 517                 break;
 518         case PDB_POLICY_NEIGHBOR:
 519 #if defined(CONFIG_NUMA)
 520         {
 521                 int i;
 522                 cpumask_copy(&mask, cpumask_of_node(cpu_to_node(index)));
 523                 for (i = max; i < num_online_cpus(); i++)
 524                         cpumask_clear_cpu(i, &mask);
 525                 pc->pc_npartners = cpumask_weight(&mask) - 1;
 526                 set_bit(LIOD_BIND, &pc->pc_flags);
 527         }
 528 #else
 529                 LASSERT(max >= 3);
 530                 pc->pc_npartners = 2;
 531 #endif
 532                 break;
 533         default:
 534                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 535                 rc = -EINVAL;
 536         }
 537
 538         if (rc == 0 && pc->pc_npartners > 0) {
 539                 OBD_ALLOC(pc->pc_partners,
 540                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 541                 if (pc->pc_partners == NULL) {
 542                         pc->pc_npartners = 0;
 543                         rc = -ENOMEM;
 544                 } else {
 545                         switch (ptlrpcd_bind_policy) {
 546                         case PDB_POLICY_PAIR:
 547                                 if (index & 0x1) {
 548                                         set_bit(LIOD_BIND, &pc->pc_flags);
 549                                         pc->pc_partners[0] = &ptlrpcds->
 550                                                 pd_threads[index - 1];
 551                                         ptlrpcds->pd_threads[index - 1].
 552                                                 pc_partners[0] = pc;
 553                                 }
 554                                 break;
 555                         case PDB_POLICY_NEIGHBOR:
 556 #if defined(CONFIG_NUMA)
 557                         {
 558                                 struct ptlrpcd_ctl *ppc;
 559                                 int i, pidx;
 560                                 /* partners are cores in the same NUMA node.
 561                                  * setup partnership only with ptlrpcd threads
 562                                  * that are already initialized
 563                                  */
 564                                 for (pidx = 0, i = 0; i < index; i++) {
 565                                         if (cpumask_test_cpu(i, &mask)) {
 566                                                 ppc = &ptlrpcds->pd_threads[i];
 567                                                 pc->pc_partners[pidx++] = ppc;
 568                                                 ppc->pc_partners[ppc->
 569                                                           pc_npartners++] = pc;
 570                                         }
 571                                 }
 572                                 /* adjust number of partners to the number
 573                                  * of partnership really setup */
 574                                 pc->pc_npartners = pidx;
 575                         }
 576 #else
 577                                 if (index & 0x1)
 578                                         set_bit(LIOD_BIND, &pc->pc_flags);
 579                                 if (index > 0) {
 580                                         pc->pc_partners[0] = &ptlrpcds->
 581                                                 pd_threads[index - 1];
 582                                         ptlrpcds->pd_threads[index - 1].
 583                                                 pc_partners[1] = pc;
 584                                         if (index == max - 1) {
 585                                                 pc->pc_partners[1] =
 586                                                 &ptlrpcds->pd_threads[0];
 587                                                 ptlrpcds->pd_threads[0].
 588                                                 pc_partners[0] = pc;
 589                                         }
 590                                 }
 591 #endif
 592                                 break;
 593                         }
 594                 }
 595         }
 596
 597         RETURN(rc);
 598 }
 599
 600
 601 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 602 {
 603         int rc;
 604         ENTRY;
 605
 606         /*
 607          * Do not allow start second thread for one pc.
 608          */
 609         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 610                 CWARN("Starting second thread (%s) for same pc %p\n",
 611                       name, pc);
 612                 RETURN(0);
 613         }
 614
 615         pc->pc_index = index;
 616         init_completion(&pc->pc_starting);
 617         init_completion(&pc->pc_finishing);
 618         spin_lock_init(&pc->pc_lock);
 619         strlcpy(pc->pc_name, name, sizeof(pc->pc_name));
 620         pc->pc_set = ptlrpc_prep_set();
 621         if (pc->pc_set == NULL)
 622                 GOTO(out, rc = -ENOMEM);
 623
 624         /*
 625          * So far only "client" ptlrpcd uses an environment. In the future,
 626          * ptlrpcd thread (or a thread-set) has to be given an argument,
 627          * describing its "scope".
 628          */
 629         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 630         if (rc != 0)
 631                 GOTO(out_set, rc);
 632
 633         {
 634                 struct task_struct *task;
 635                 if (index >= 0) {
 636                         rc = ptlrpcd_bind(index, max);
 637                         if (rc < 0)
 638                                 GOTO(out_env, rc);
 639                 }
 640
 641                 task = kthread_run(ptlrpcd, pc, pc->pc_name);
 642                 if (IS_ERR(task))
 643                         GOTO(out_env, rc = PTR_ERR(task));
 644
 645                 wait_for_completion(&pc->pc_starting);
 646         }
 647         RETURN(0);
 648
 649 out_env:
 650         lu_context_fini(&pc->pc_env.le_ctx);
 651
 652 out_set:
 653         if (pc->pc_set != NULL) {
 654                 struct ptlrpc_request_set *set = pc->pc_set;
 655
 656                 spin_lock(&pc->pc_lock);
 657                 pc->pc_set = NULL;
 658                 spin_unlock(&pc->pc_lock);
 659                 ptlrpc_set_destroy(set);
 660         }
 661         clear_bit(LIOD_BIND, &pc->pc_flags);
 662 out:
 663         clear_bit(LIOD_START, &pc->pc_flags);
 664         RETURN(rc);
 665 }
 666
 667 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 668 {
 669         ENTRY;
 670
 671         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 672                 CWARN("Thread for pc %p was not started\n", pc);
 673                 goto out;
 674         }
 675
 676         set_bit(LIOD_STOP, &pc->pc_flags);
 677         if (force)
 678                 set_bit(LIOD_FORCE, &pc->pc_flags);
 679         wake_up(&pc->pc_set->set_waitq);
 680
 681 out:
 682         EXIT;
 683 }
 684
 685 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 686 {
 687         struct ptlrpc_request_set *set = pc->pc_set;
 688         ENTRY;
 689
 690         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 691                 CWARN("Thread for pc %p was not started\n", pc);
 692                 goto out;
 693         }
 694
 695         wait_for_completion(&pc->pc_finishing);
 696         lu_context_fini(&pc->pc_env.le_ctx);
 697
 698         spin_lock(&pc->pc_lock);
 699         pc->pc_set = NULL;
 700         spin_unlock(&pc->pc_lock);
 701         ptlrpc_set_destroy(set);
 702
 703         clear_bit(LIOD_START, &pc->pc_flags);
 704         clear_bit(LIOD_STOP, &pc->pc_flags);
 705         clear_bit(LIOD_FORCE, &pc->pc_flags);
 706         clear_bit(LIOD_BIND, &pc->pc_flags);
 707
 708 out:
 709         if (pc->pc_npartners > 0) {
 710                 LASSERT(pc->pc_partners != NULL);
 711
 712                 OBD_FREE(pc->pc_partners,
 713                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 714                 pc->pc_partners = NULL;
 715         }
 716         pc->pc_npartners = 0;
 717         EXIT;
 718 }
 719
 720 static void ptlrpcd_fini(void)
 721 {
 722         int i;
 723         ENTRY;
 724
 725         if (ptlrpcds != NULL) {
 726                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 727                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 728                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 729                         ptlrpcd_free(&ptlrpcds->pd_threads[i]);
 730                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 731                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 732                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 733                 ptlrpcds = NULL;
 734         }
 735
 736         EXIT;
 737 }
 738
 739 static int ptlrpcd_init(void)
 740 {
 741         int     nthreads = num_online_cpus();
 742         char    name[16];
 743         int     size, i = -1, j, rc = 0;
 744         ENTRY;
 745
 746         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 747                 nthreads = max_ptlrpcds;
 748         if (nthreads < 2)
 749                 nthreads = 2;
 750         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 751                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 752         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 753                 nthreads &= ~1; /* make sure it is even */
 754
 755         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 756         OBD_ALLOC(ptlrpcds, size);
 757         if (ptlrpcds == NULL)
 758                 GOTO(out, rc = -ENOMEM);
 759
 760         snprintf(name, 15, "ptlrpcd_rcv");
 761         set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 762         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 763         if (rc < 0)
 764                 GOTO(out, rc);
 765
 766         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 767          *      non-recovery async RPC to improve overall async RPC efficiency.
 768          *
 769          *      But there are some issues with async I/O RPCs and async non-I/O
 770          *      RPCs processed in the same set under some cases. The ptlrpcd may
 771          *      be blocked by some async I/O RPC(s), then will cause other async
 772          *      non-I/O RPC(s) can not be processed in time.
 773          *
 774          *      Maybe we should distinguish blocked async RPCs from non-blocked
 775          *      async RPCs, and process them in different ptlrpcd sets to avoid
 776          *      unnecessary dependency. But how to distribute async RPCs load
 777          *      among all the ptlrpc daemons becomes another trouble. */
 778         for (i = 0; i < nthreads; i++) {
 779                 snprintf(name, 15, "ptlrpcd_%d", i);
 780                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 781                 if (rc < 0)
 782                         GOTO(out, rc);
 783         }
 784
 785         ptlrpcds->pd_size = size;
 786         ptlrpcds->pd_index = 0;
 787         ptlrpcds->pd_nthreads = nthreads;
 788
 789 out:
 790         if (rc != 0 && ptlrpcds != NULL) {
 791                 for (j = 0; j <= i; j++)
 792                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 793                 for (j = 0; j <= i; j++)
 794                         ptlrpcd_free(&ptlrpcds->pd_threads[j]);
 795                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 796                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 797                 OBD_FREE(ptlrpcds, size);
 798                 ptlrpcds = NULL;
 799         }
 800
 801         RETURN(rc);
 802 }
 803
 804 int ptlrpcd_addref(void)
 805 {
 806         int rc = 0;
 807         ENTRY;
 808
 809         mutex_lock(&ptlrpcd_mutex);
 810         if (++ptlrpcd_users == 1) {
 811                 rc = ptlrpcd_init();
 812                 if (rc < 0)
 813                         ptlrpcd_users--;
 814         }
 815         mutex_unlock(&ptlrpcd_mutex);
 816         RETURN(rc);
 817 }
 818 EXPORT_SYMBOL(ptlrpcd_addref);
 819
 820 void ptlrpcd_decref(void)
 821 {
 822         mutex_lock(&ptlrpcd_mutex);
 823         if (--ptlrpcd_users == 0)
 824                 ptlrpcd_fini();
 825         mutex_unlock(&ptlrpcd_mutex);
 826 }
 827 EXPORT_SYMBOL(ptlrpcd_decref);
 828 /** @} ptlrpcd */