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         char jobid[JOBSTATS_JOBID_SIZE];
 241
 242         if (req->rq_reqmsg) {
 243                 lustre_get_jobid(jobid);
 244                 lustre_msg_set_jobid(req->rq_reqmsg, jobid);
 245         }
 246
 247         cfs_spin_lock(&req->rq_lock);
 248         if (req->rq_invalid_rqset) {
 249                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 250                                                      back_to_sleep, NULL);
 251
 252                 req->rq_invalid_rqset = 0;
 253                 cfs_spin_unlock(&req->rq_lock);
 254                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 255         } else if (req->rq_set) {
 256                 /* If we have a vaid "rq_set", just reuse it to avoid double
 257                  * linked. */
 258                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 259                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 260
 261                 /* ptlrpc_check_set will decrease the count */
 262                 cfs_atomic_inc(&req->rq_set->set_remaining);
 263                 cfs_spin_unlock(&req->rq_lock);
 264                 cfs_waitq_signal(&req->rq_set->set_waitq);
 265                 return;
 266         } else {
 267                 cfs_spin_unlock(&req->rq_lock);
 268         }
 269
 270         pc = ptlrpcd_select_pc(req, policy, idx);
 271
 272         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 273                   req, pc->pc_name, pc->pc_index);
 274
 275         ptlrpc_set_add_new_req(pc, req);
 276 }
 277
 278 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 279 {
 280         cfs_atomic_inc(&set->set_refcount);
 281 }
 282
 283 /**
 284  * Check if there is more work to do on ptlrpcd set.
 285  * Returns 1 if yes.
 286  */
 287 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 288 {
 289         cfs_list_t *tmp, *pos;
 290         struct ptlrpc_request *req;
 291         struct ptlrpc_request_set *set = pc->pc_set;
 292         int rc = 0;
 293         int rc2;
 294         ENTRY;
 295
 296         if (cfs_atomic_read(&set->set_new_count)) {
 297                 cfs_spin_lock(&set->set_new_req_lock);
 298                 if (likely(!cfs_list_empty(&set->set_new_requests))) {
 299                         cfs_list_splice_init(&set->set_new_requests,
 300                                              &set->set_requests);
 301                         cfs_atomic_add(cfs_atomic_read(&set->set_new_count),
 302                                        &set->set_remaining);
 303                         cfs_atomic_set(&set->set_new_count, 0);
 304                         /*
 305                          * Need to calculate its timeout.
 306                          */
 307                         rc = 1;
 308                 }
 309                 cfs_spin_unlock(&set->set_new_req_lock);
 310         }
 311
 312         /* We should call lu_env_refill() before handling new requests to make
 313          * sure that env key the requests depending on really exists.
 314          */
 315         rc2 = lu_env_refill(env);
 316         if (rc2 != 0) {
 317                 /*
 318                  * XXX This is very awkward situation, because
 319                  * execution can neither continue (request
 320                  * interpreters assume that env is set up), nor repeat
 321                  * the loop (as this potentially results in a tight
 322                  * loop of -ENOMEM's).
 323                  *
 324                  * Fortunately, refill only ever does something when
 325                  * new modules are loaded, i.e., early during boot up.
 326                  */
 327                 CERROR("Failure to refill session: %d\n", rc2);
 328                 RETURN(rc);
 329         }
 330
 331         if (cfs_atomic_read(&set->set_remaining))
 332                 rc |= ptlrpc_check_set(env, set);
 333
 334         if (!cfs_list_empty(&set->set_requests)) {
 335                 /*
 336                  * XXX: our set never completes, so we prune the completed
 337                  * reqs after each iteration. boy could this be smarter.
 338                  */
 339                 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 340                         req = cfs_list_entry(pos, struct ptlrpc_request,
 341                                              rq_set_chain);
 342                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 343                                 continue;
 344
 345                         cfs_list_del_init(&req->rq_set_chain);
 346                         req->rq_set = NULL;
 347                         ptlrpc_req_finished(req);
 348                 }
 349         }
 350
 351         if (rc == 0) {
 352                 /*
 353                  * If new requests have been added, make sure to wake up.
 354                  */
 355                 rc = cfs_atomic_read(&set->set_new_count);
 356
 357 #ifdef __KERNEL__
 358                 /* If we have nothing to do, check whether we can take some
 359                  * work from our partner threads. */
 360                 if (rc == 0 && pc->pc_npartners > 0) {
 361                         struct ptlrpcd_ctl *partner;
 362                         struct ptlrpc_request_set *ps;
 363                         int first = pc->pc_cursor;
 364
 365                         do {
 366                                 partner = pc->pc_partners[pc->pc_cursor++];
 367                                 if (pc->pc_cursor >= pc->pc_npartners)
 368                                         pc->pc_cursor = 0;
 369                                 if (partner == NULL)
 370                                         continue;
 371
 372                                 cfs_spin_lock(&partner->pc_lock);
 373                                 ps = partner->pc_set;
 374                                 if (ps == NULL) {
 375                                         cfs_spin_unlock(&partner->pc_lock);
 376                                         continue;
 377                                 }
 378
 379                                 ptlrpc_reqset_get(ps);
 380                                 cfs_spin_unlock(&partner->pc_lock);
 381
 382                                 if (cfs_atomic_read(&ps->set_new_count)) {
 383                                         rc = ptlrpcd_steal_rqset(set, ps);
 384                                         if (rc > 0)
 385                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 386                                                        " async RPCs [%d->%d]\n",
 387                                                         rc, partner->pc_index,
 388                                                         pc->pc_index);
 389                                 }
 390                                 ptlrpc_reqset_put(ps);
 391                         } while (rc == 0 && pc->pc_cursor != first);
 392                 }
 393 #endif
 394         }
 395
 396         RETURN(rc);
 397 }
 398
 399 #ifdef __KERNEL__
 400 /**
 401  * Main ptlrpcd thread.
 402  * ptlrpc's code paths like to execute in process context, so we have this
 403  * thread which spins on a set which contains the rpcs and sends them.
 404  *
 405  */
 406 static int ptlrpcd(void *arg)
 407 {
 408         struct ptlrpcd_ctl *pc = arg;
 409         struct ptlrpc_request_set *set = pc->pc_set;
 410         struct lu_env env = { .le_ses = NULL };
 411         int rc, exit = 0;
 412         ENTRY;
 413
 414         cfs_daemonize_ctxt(pc->pc_name);
 415 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
 416         if (cfs_test_bit(LIOD_BIND, &pc->pc_flags)) {
 417                 int index = pc->pc_index;
 418
 419                 if (index >= 0 && index < cfs_num_possible_cpus()) {
 420                         while (!cpu_online(index)) {
 421                                 if (++index >= cfs_num_possible_cpus())
 422                                         index = 0;
 423                         }
 424                         cfs_set_cpus_allowed(cfs_current(),
 425                                      node_to_cpumask(cpu_to_node(index)));
 426                 }
 427         }
 428 #endif
 429         /*
 430          * XXX So far only "client" ptlrpcd uses an environment. In
 431          * the future, ptlrpcd thread (or a thread-set) has to given
 432          * an argument, describing its "scope".
 433          */
 434         rc = lu_context_init(&env.le_ctx,
 435                              LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 436         cfs_complete(&pc->pc_starting);
 437
 438         if (rc != 0)
 439                 RETURN(rc);
 440
 441         /*
 442          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 443          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 444          * there are requests in the set. New requests come in on the set's
 445          * new_req_list and ptlrpcd_check() moves them into the set.
 446          */
 447         do {
 448                 struct l_wait_info lwi;
 449                 int timeout;
 450
 451                 timeout = ptlrpc_set_next_timeout(set);
 452                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 453                                   ptlrpc_expired_set, set);
 454
 455                 lu_context_enter(&env.le_ctx);
 456                 l_wait_event(set->set_waitq,
 457                              ptlrpcd_check(&env, pc), &lwi);
 458                 lu_context_exit(&env.le_ctx);
 459
 460                 /*
 461                  * Abort inflight rpcs for forced stop case.
 462                  */
 463                 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
 464                         if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
 465                                 ptlrpc_abort_set(set);
 466                         exit++;
 467                 }
 468
 469                 /*
 470                  * Let's make one more loop to make sure that ptlrpcd_check()
 471                  * copied all raced new rpcs into the set so we can kill them.
 472                  */
 473         } while (exit < 2);
 474
 475         /*
 476          * Wait for inflight requests to drain.
 477          */
 478         if (!cfs_list_empty(&set->set_requests))
 479                 ptlrpc_set_wait(set);
 480         lu_context_fini(&env.le_ctx);
 481         cfs_complete(&pc->pc_finishing);
 482
 483         cfs_clear_bit(LIOD_START, &pc->pc_flags);
 484         cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
 485         cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
 486         cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 487         return 0;
 488 }
 489
 490 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 491  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 492  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 493  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 494  *      because of some CPU core(s) busy with other loads.
 495  *
 496  *      For example: "ls -l", some async RPCs for statahead are assigned to
 497  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 498  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 499  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 500  *      such case, the statahead async RPCs can not be processed in time, it is
 501  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 502  *      be better. But it breaks former data transfer policy.
 503  *
 504  *      So we shouldn't be blind for avoiding the data transfer. We make some
 505  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 506  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 507  *      core. The other part is for free mode, all the ptlrpcd threads in the
 508  *      part can be scheduled on any CPU core. We specify some partnership
 509  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 510  *      and the async RPC load within the partners are shared.
 511  *
 512  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 513  *      thread can be scheduled in time), and try to guarantee the async RPC
 514  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 515  *      on any CPU core).
 516  *
 517  *      As for how to specify the partnership between bound mode ptlrpcd
 518  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 519  *      <free bound> pair. In future, we can specify some more complex
 520  *      partnership based on the patches for CPU partition. But before such
 521  *      patches are available, we prefer to use the simplest one.
 522  */
 523 # ifdef CFS_CPU_MODE_NUMA
 524 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 525 # endif
 526 static int ptlrpcd_bind(int index, int max)
 527 {
 528         struct ptlrpcd_ctl *pc;
 529         int rc = 0;
 530 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 531         struct ptlrpcd_ctl *ppc;
 532         int node, i, pidx;
 533         cpumask_t mask;
 534 #endif
 535         ENTRY;
 536
 537         LASSERT(index <= max - 1);
 538         pc = &ptlrpcds->pd_threads[index];
 539         switch (ptlrpcd_bind_policy) {
 540         case PDB_POLICY_NONE:
 541                 pc->pc_npartners = -1;
 542                 break;
 543         case PDB_POLICY_FULL:
 544                 pc->pc_npartners = 0;
 545                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 546                 break;
 547         case PDB_POLICY_PAIR:
 548                 LASSERT(max % 2 == 0);
 549                 pc->pc_npartners = 1;
 550                 break;
 551         case PDB_POLICY_NEIGHBOR:
 552 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 553                 node = cpu_to_node(index);
 554                 mask = node_to_cpumask(node);
 555                 for (i = max; i < cfs_num_online_cpus(); i++)
 556                         cpu_clear(i, mask);
 557                 pc->pc_npartners = cpus_weight(mask) - 1;
 558                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 559 #else
 560                 LASSERT(max >= 3);
 561                 pc->pc_npartners = 2;
 562 #endif
 563                 break;
 564         default:
 565                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 566                 rc = -EINVAL;
 567         }
 568
 569         if (rc == 0 && pc->pc_npartners > 0) {
 570                 OBD_ALLOC(pc->pc_partners,
 571                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 572                 if (pc->pc_partners == NULL) {
 573                         pc->pc_npartners = 0;
 574                         rc = -ENOMEM;
 575                 } else {
 576                         switch (ptlrpcd_bind_policy) {
 577                         case PDB_POLICY_PAIR:
 578                                 if (index & 0x1) {
 579                                         cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 580                                         pc->pc_partners[0] = &ptlrpcds->
 581                                                 pd_threads[index - 1];
 582                                         ptlrpcds->pd_threads[index - 1].
 583                                                 pc_partners[0] = pc;
 584                                 }
 585                                 break;
 586                         case PDB_POLICY_NEIGHBOR:
 587 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
 588                                 /* partners are cores in the same NUMA node.
 589                                  * setup partnership only with ptlrpcd threads
 590                                  * that are already initialized
 591                                  */
 592                                 for (pidx = 0, i = 0; i < index; i++) {
 593                                         if (cpu_isset(i, mask)) {
 594                                                 ppc = &ptlrpcds->pd_threads[i];
 595                                                 pc->pc_partners[pidx++] = ppc;
 596                                                 ppc->pc_partners[ppc->
 597                                                           pc_npartners++] = pc;
 598                                         }
 599                                 }
 600                                 /* adjust number of partners to the number
 601                                  * of partnership really setup */
 602                                 pc->pc_npartners = pidx;
 603 #else
 604                                 if (index & 0x1)
 605                                         cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 606                                 if (index > 0) {
 607                                         pc->pc_partners[0] = &ptlrpcds->
 608                                                 pd_threads[index - 1];
 609                                         ptlrpcds->pd_threads[index - 1].
 610                                                 pc_partners[1] = pc;
 611                                         if (index == max - 1) {
 612                                                 pc->pc_partners[1] =
 613                                                 &ptlrpcds->pd_threads[0];
 614                                                 ptlrpcds->pd_threads[0].
 615                                                 pc_partners[0] = pc;
 616                                         }
 617                                 }
 618 #endif
 619                                 break;
 620                         }
 621                 }
 622         }
 623
 624         RETURN(rc);
 625 }
 626
 627 #else /* !__KERNEL__ */
 628
 629 /**
 630  * In liblustre we do not have separate threads, so this function
 631  * is called from time to time all across common code to see
 632  * if something needs to be processed on ptlrpcd set.
 633  */
 634 int ptlrpcd_check_async_rpcs(void *arg)
 635 {
 636         struct ptlrpcd_ctl *pc = arg;
 637         int                 rc = 0;
 638
 639         /*
 640          * Single threaded!!
 641          */
 642         pc->pc_recurred++;
 643
 644         if (pc->pc_recurred == 1) {
 645                 rc = lu_env_refill(&pc->pc_env);
 646                 if (rc == 0) {
 647                         lu_context_enter(&pc->pc_env.le_ctx);
 648                         rc = ptlrpcd_check(&pc->pc_env, pc);
 649                         if (!rc)
 650                                 ptlrpc_expired_set(pc->pc_set);
 651                         /*
 652                          * XXX: send replay requests.
 653                          */
 654                         if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
 655                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 656                         lu_context_exit(&pc->pc_env.le_ctx);
 657                 }
 658         }
 659
 660         pc->pc_recurred--;
 661         return rc;
 662 }
 663
 664 int ptlrpcd_idle(void *arg)
 665 {
 666         struct ptlrpcd_ctl *pc = arg;
 667
 668         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
 669                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 670 }
 671
 672 #endif
 673
 674 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 675 {
 676         int rc;
 677         int env = 0;
 678         ENTRY;
 679
 680         /*
 681          * Do not allow start second thread for one pc.
 682          */
 683         if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 684                 CWARN("Starting second thread (%s) for same pc %p\n",
 685                        name, pc);
 686                 RETURN(0);
 687         }
 688
 689         pc->pc_index = index;
 690         cfs_init_completion(&pc->pc_starting);
 691         cfs_init_completion(&pc->pc_finishing);
 692         cfs_spin_lock_init(&pc->pc_lock);
 693         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 694         pc->pc_set = ptlrpc_prep_set();
 695         if (pc->pc_set == NULL)
 696                 GOTO(out, rc = -ENOMEM);
 697         /*
 698          * So far only "client" ptlrpcd uses an environment. In the future,
 699          * ptlrpcd thread (or a thread-set) has to be given an argument,
 700          * describing its "scope".
 701          */
 702         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 703         if (rc != 0)
 704                 GOTO(out, rc);
 705
 706         env = 1;
 707 #ifdef __KERNEL__
 708         if (index >= 0) {
 709                 rc = ptlrpcd_bind(index, max);
 710                 if (rc < 0)
 711                         GOTO(out, rc);
 712         }
 713
 714         rc = cfs_create_thread(ptlrpcd, pc, 0);
 715         if (rc < 0)
 716                 GOTO(out, rc);
 717
 718         rc = 0;
 719         cfs_wait_for_completion(&pc->pc_starting);
 720 #else
 721         pc->pc_wait_callback =
 722                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 723                                                  &ptlrpcd_check_async_rpcs, pc);
 724         pc->pc_idle_callback =
 725                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 726                                                  &ptlrpcd_idle, pc);
 727 #endif
 728 out:
 729         if (rc) {
 730 #ifdef __KERNEL__
 731                 if (pc->pc_set != NULL) {
 732                         struct ptlrpc_request_set *set = pc->pc_set;
 733
 734                         cfs_spin_lock(&pc->pc_lock);
 735                         pc->pc_set = NULL;
 736                         cfs_spin_unlock(&pc->pc_lock);
 737                         ptlrpc_set_destroy(set);
 738                 }
 739                 if (env != 0)
 740                         lu_context_fini(&pc->pc_env.le_ctx);
 741                 cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 742 #else
 743                 SET_BUT_UNUSED(env);
 744 #endif
 745                 cfs_clear_bit(LIOD_START, &pc->pc_flags);
 746         }
 747         RETURN(rc);
 748 }
 749
 750 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 751 {
 752        struct ptlrpc_request_set *set = pc->pc_set;
 753         ENTRY;
 754
 755         if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
 756                 CWARN("Thread for pc %p was not started\n", pc);
 757                 goto out;
 758         }
 759
 760         cfs_set_bit(LIOD_STOP, &pc->pc_flags);
 761         if (force)
 762                 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
 763         cfs_waitq_signal(&pc->pc_set->set_waitq);
 764 #ifdef __KERNEL__
 765         cfs_wait_for_completion(&pc->pc_finishing);
 766 #else
 767         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 768         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 769 #endif
 770         lu_context_fini(&pc->pc_env.le_ctx);
 771
 772         cfs_spin_lock(&pc->pc_lock);
 773         pc->pc_set = NULL;
 774         cfs_spin_unlock(&pc->pc_lock);
 775         ptlrpc_set_destroy(set);
 776
 777 out:
 778 #ifdef __KERNEL__
 779         if (pc->pc_npartners > 0) {
 780                 LASSERT(pc->pc_partners != NULL);
 781
 782                 OBD_FREE(pc->pc_partners,
 783                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 784                 pc->pc_partners = NULL;
 785         }
 786         pc->pc_npartners = 0;
 787 #endif
 788         EXIT;
 789 }
 790
 791 static void ptlrpcd_fini(void)
 792 {
 793         int i;
 794         ENTRY;
 795
 796         if (ptlrpcds != NULL) {
 797                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 798                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 799                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 800                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 801                 ptlrpcds = NULL;
 802         }
 803
 804         EXIT;
 805 }
 806
 807 static int ptlrpcd_init(void)
 808 {
 809         int nthreads = cfs_num_online_cpus();
 810         char name[16];
 811         int size, i = -1, j, rc = 0;
 812         ENTRY;
 813
 814 #ifdef __KERNEL__
 815         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 816                 nthreads = max_ptlrpcds;
 817         if (nthreads < 2)
 818                 nthreads = 2;
 819         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 820                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 821         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 822                 nthreads &= ~1; /* make sure it is even */
 823 #else
 824         nthreads = 1;
 825 #endif
 826
 827         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 828         OBD_ALLOC(ptlrpcds, size);
 829         if (ptlrpcds == NULL)
 830                 GOTO(out, rc = -ENOMEM);
 831
 832         snprintf(name, 15, "ptlrpcd_rcv");
 833         cfs_set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 834         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 835         if (rc < 0)
 836                 GOTO(out, rc);
 837
 838         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 839          *      non-recovery async RPC to improve overall async RPC efficiency.
 840          *
 841          *      But there are some issues with async I/O RPCs and async non-I/O
 842          *      RPCs processed in the same set under some cases. The ptlrpcd may
 843          *      be blocked by some async I/O RPC(s), then will cause other async
 844          *      non-I/O RPC(s) can not be processed in time.
 845          *
 846          *      Maybe we should distinguish blocked async RPCs from non-blocked
 847          *      async RPCs, and process them in different ptlrpcd sets to avoid
 848          *      unnecessary dependency. But how to distribute async RPCs load
 849          *      among all the ptlrpc daemons becomes another trouble. */
 850         for (i = 0; i < nthreads; i++) {
 851                 snprintf(name, 15, "ptlrpcd_%d", i);
 852                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 853                 if (rc < 0)
 854                         GOTO(out, rc);
 855         }
 856
 857         ptlrpcds->pd_size = size;
 858         ptlrpcds->pd_index = 0;
 859         ptlrpcds->pd_nthreads = nthreads;
 860
 861 out:
 862         if (rc != 0 && ptlrpcds != NULL) {
 863                 for (j = 0; j <= i; j++)
 864                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 865                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 866                 OBD_FREE(ptlrpcds, size);
 867                 ptlrpcds = NULL;
 868         }
 869
 870         RETURN(0);
 871 }
 872
 873 int ptlrpcd_addref(void)
 874 {
 875         int rc = 0;
 876         ENTRY;
 877
 878         cfs_mutex_lock(&ptlrpcd_mutex);
 879         if (++ptlrpcd_users == 1)
 880                 rc = ptlrpcd_init();
 881         cfs_mutex_unlock(&ptlrpcd_mutex);
 882         RETURN(rc);
 883 }
 884
 885 void ptlrpcd_decref(void)
 886 {
 887         cfs_mutex_lock(&ptlrpcd_mutex);
 888         if (--ptlrpcd_users == 0)
 889                 ptlrpcd_fini();
 890         cfs_mutex_unlock(&ptlrpcd_mutex);
 891 }
 892 /** @} ptlrpcd */