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