lustre/ptlrpc/ptlrpcd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/ptlrpcd.c
  37  */
  38
  39 /** \defgroup ptlrpcd PortalRPC daemon
  40  *
  41  * ptlrpcd is a special thread with its own set where other user might add
  42  * requests when they don't want to wait for their completion.
  43  * PtlRPCD will take care of sending such requests and then processing their
  44  * replies and calling completion callbacks as necessary.
  45  * The callbacks are called directly from ptlrpcd context.
  46  * It is important to never significantly block (esp. on RPCs!) within such
  47  * completion handler or a deadlock might occur where ptlrpcd enters some
  48  * callback that attempts to send another RPC and wait for it to return,
  49  * during which time ptlrpcd is completely blocked, so e.g. if import
  50  * fails, recovery cannot progress because connection requests are also
  51  * sent by ptlrpcd.
  52  *
  53  * @{
  54  */
  55
  56 #define DEBUG_SUBSYSTEM S_RPC
  57
  58 #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 struct mutex 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         wake_up(&rq_set->set_waitq);
 105 }
 106 EXPORT_SYMBOL(ptlrpcd_wake);
 107
 108 static struct ptlrpcd_ctl *
 109 ptlrpcd_select_pc(struct ptlrpc_request *req, pdl_policy_t policy, int index)
 110 {
 111         int idx = 0;
 112
 113         if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
 114                 return &ptlrpcds->pd_thread_rcv;
 115
 116 #ifdef __KERNEL__
 117         switch (policy) {
 118         case PDL_POLICY_SAME:
 119                 idx = smp_processor_id() % ptlrpcds->pd_nthreads;
 120                 break;
 121         case PDL_POLICY_LOCAL:
 122                 /* Before CPU partition patches available, process it the same
 123                  * as "PDL_POLICY_ROUND". */
 124 # ifdef CFS_CPU_MODE_NUMA
 125 # warning "fix this code to use new CPU partition APIs"
 126 # endif
 127                 /* Fall through to PDL_POLICY_ROUND until the CPU
 128                  * CPU partition patches are available. */
 129                 index = -1;
 130         case PDL_POLICY_PREFERRED:
 131                 if (index >= 0 && index < num_online_cpus()) {
 132                         idx = index % ptlrpcds->pd_nthreads;
 133                         break;
 134                 }
 135                 /* Fall through to PDL_POLICY_ROUND for bad index. */
 136         default:
 137                 /* Fall through to PDL_POLICY_ROUND for unknown policy. */
 138         case PDL_POLICY_ROUND:
 139                 /* We do not care whether it is strict load balance. */
 140                 idx = ptlrpcds->pd_index + 1;
 141                 if (idx == smp_processor_id())
 142                         idx++;
 143                 idx %= ptlrpcds->pd_nthreads;
 144                 ptlrpcds->pd_index = idx;
 145                 break;
 146         }
 147 #endif /* __KERNEL__ */
 148
 149         return &ptlrpcds->pd_threads[idx];
 150 }
 151
 152 /**
 153  * Move all request from an existing request set to the ptlrpcd queue.
 154  * All requests from the set must be in phase RQ_PHASE_NEW.
 155  */
 156 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 157 {
 158         cfs_list_t *tmp, *pos;
 159 #ifdef __KERNEL__
 160         struct ptlrpcd_ctl *pc;
 161         struct ptlrpc_request_set *new;
 162         int count, i;
 163
 164         pc = ptlrpcd_select_pc(NULL, PDL_POLICY_LOCAL, -1);
 165         new = pc->pc_set;
 166 #endif
 167
 168         cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 169                 struct ptlrpc_request *req =
 170                         cfs_list_entry(pos, struct ptlrpc_request,
 171                                        rq_set_chain);
 172
 173                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 174 #ifdef __KERNEL__
 175                 req->rq_set = new;
 176                 req->rq_queued_time = cfs_time_current();
 177 #else
 178                 cfs_list_del_init(&req->rq_set_chain);
 179                 req->rq_set = NULL;
 180                 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
 181                 cfs_atomic_dec(&set->set_remaining);
 182 #endif
 183         }
 184
 185 #ifdef __KERNEL__
 186         spin_lock(&new->set_new_req_lock);
 187         cfs_list_splice_init(&set->set_requests, &new->set_new_requests);
 188         i = cfs_atomic_read(&set->set_remaining);
 189         count = cfs_atomic_add_return(i, &new->set_new_count);
 190         cfs_atomic_set(&set->set_remaining, 0);
 191         spin_unlock(&new->set_new_req_lock);
 192         if (count == i) {
 193                 wake_up(&new->set_waitq);
 194
 195                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
 196                  *      guarantee the async RPC can be processed ASAP, we have
 197                  *      no other better choice. It maybe fixed in future. */
 198                 for (i = 0; i < pc->pc_npartners; i++)
 199                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
 200         }
 201 #endif
 202 }
 203 EXPORT_SYMBOL(ptlrpcd_add_rqset);
 204
 205 #ifdef __KERNEL__
 206 /**
 207  * Return transferred RPCs count.
 208  */
 209 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 210                                struct ptlrpc_request_set *src)
 211 {
 212         cfs_list_t *tmp, *pos;
 213         struct ptlrpc_request *req;
 214         int rc = 0;
 215
 216         spin_lock(&src->set_new_req_lock);
 217         if (likely(!cfs_list_empty(&src->set_new_requests))) {
 218                 cfs_list_for_each_safe(pos, tmp, &src->set_new_requests) {
 219                         req = cfs_list_entry(pos, struct ptlrpc_request,
 220                                              rq_set_chain);
 221                         req->rq_set = des;
 222                 }
 223                 cfs_list_splice_init(&src->set_new_requests,
 224                                      &des->set_requests);
 225                 rc = cfs_atomic_read(&src->set_new_count);
 226                 cfs_atomic_add(rc, &des->set_remaining);
 227                 cfs_atomic_set(&src->set_new_count, 0);
 228         }
 229         spin_unlock(&src->set_new_req_lock);
 230         return rc;
 231 }
 232 #endif
 233
 234 /**
 235  * Requests that are added to the ptlrpcd queue are sent via
 236  * ptlrpcd_check->ptlrpc_check_set().
 237  */
 238 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx)
 239 {
 240         struct ptlrpcd_ctl *pc;
 241
 242         if (req->rq_reqmsg)
 243                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
 244
 245         spin_lock(&req->rq_lock);
 246         if (req->rq_invalid_rqset) {
 247                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 248                                                      back_to_sleep, NULL);
 249
 250                 req->rq_invalid_rqset = 0;
 251                 spin_unlock(&req->rq_lock);
 252                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 253         } else if (req->rq_set) {
 254                 /* If we have a vaid "rq_set", just reuse it to avoid double
 255                  * linked. */
 256                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 257                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 258
 259                 /* ptlrpc_check_set will decrease the count */
 260                 cfs_atomic_inc(&req->rq_set->set_remaining);
 261                 spin_unlock(&req->rq_lock);
 262                 wake_up(&req->rq_set->set_waitq);
 263                 return;
 264         } else {
 265                 spin_unlock(&req->rq_lock);
 266         }
 267
 268         pc = ptlrpcd_select_pc(req, policy, idx);
 269
 270         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 271                   req, pc->pc_name, pc->pc_index);
 272
 273         ptlrpc_set_add_new_req(pc, req);
 274 }
 275 EXPORT_SYMBOL(ptlrpcd_add_req);
 276
 277 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 278 {
 279         cfs_atomic_inc(&set->set_refcount);
 280 }
 281
 282 /**
 283  * Check if there is more work to do on ptlrpcd set.
 284  * Returns 1 if yes.
 285  */
 286 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 287 {
 288         cfs_list_t *tmp, *pos;
 289         struct ptlrpc_request *req;
 290         struct ptlrpc_request_set *set = pc->pc_set;
 291         int rc = 0;
 292         int rc2;
 293         ENTRY;
 294
 295         if (cfs_atomic_read(&set->set_new_count)) {
 296                 spin_lock(&set->set_new_req_lock);
 297                 if (likely(!cfs_list_empty(&set->set_new_requests))) {
 298                         cfs_list_splice_init(&set->set_new_requests,
 299                                              &set->set_requests);
 300                         cfs_atomic_add(cfs_atomic_read(&set->set_new_count),
 301                                        &set->set_remaining);
 302                         cfs_atomic_set(&set->set_new_count, 0);
 303                         /*
 304                          * Need to calculate its timeout.
 305                          */
 306                         rc = 1;
 307                 }
 308                 spin_unlock(&set->set_new_req_lock);
 309         }
 310
 311         /* We should call lu_env_refill() before handling new requests to make
 312          * sure that env key the requests depending on really exists.
 313          */
 314         rc2 = lu_env_refill(env);
 315         if (rc2 != 0) {
 316                 /*
 317                  * XXX This is very awkward situation, because
 318                  * execution can neither continue (request
 319                  * interpreters assume that env is set up), nor repeat
 320                  * the loop (as this potentially results in a tight
 321                  * loop of -ENOMEM's).
 322                  *
 323                  * Fortunately, refill only ever does something when
 324                  * new modules are loaded, i.e., early during boot up.
 325                  */
 326                 CERROR("Failure to refill session: %d\n", rc2);
 327                 RETURN(rc);
 328         }
 329
 330         if (cfs_atomic_read(&set->set_remaining))
 331                 rc |= ptlrpc_check_set(env, set);
 332
 333         if (!cfs_list_empty(&set->set_requests)) {
 334                 /*
 335                  * XXX: our set never completes, so we prune the completed
 336                  * reqs after each iteration. boy could this be smarter.
 337                  */
 338                 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 339                         req = cfs_list_entry(pos, struct ptlrpc_request,
 340                                              rq_set_chain);
 341                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 342                                 continue;
 343
 344                         cfs_list_del_init(&req->rq_set_chain);
 345                         req->rq_set = NULL;
 346                         ptlrpc_req_finished(req);
 347                 }
 348         }
 349
 350         if (rc == 0) {
 351                 /*
 352                  * If new requests have been added, make sure to wake up.
 353                  */
 354                 rc = cfs_atomic_read(&set->set_new_count);
 355
 356 #ifdef __KERNEL__
 357                 /* If we have nothing to do, check whether we can take some
 358                  * work from our partner threads. */
 359                 if (rc == 0 && pc->pc_npartners > 0) {
 360                         struct ptlrpcd_ctl *partner;
 361                         struct ptlrpc_request_set *ps;
 362                         int first = pc->pc_cursor;
 363
 364                         do {
 365                                 partner = pc->pc_partners[pc->pc_cursor++];
 366                                 if (pc->pc_cursor >= pc->pc_npartners)
 367                                         pc->pc_cursor = 0;
 368                                 if (partner == NULL)
 369                                         continue;
 370
 371                                 spin_lock(&partner->pc_lock);
 372                                 ps = partner->pc_set;
 373                                 if (ps == NULL) {
 374                                         spin_unlock(&partner->pc_lock);
 375                                         continue;
 376                                 }
 377
 378                                 ptlrpc_reqset_get(ps);
 379                                 spin_unlock(&partner->pc_lock);
 380
 381                                 if (cfs_atomic_read(&ps->set_new_count)) {
 382                                         rc = ptlrpcd_steal_rqset(set, ps);
 383                                         if (rc > 0)
 384                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 385                                                        " async RPCs [%d->%d]\n",
 386                                                         rc, partner->pc_index,
 387                                                         pc->pc_index);
 388                                 }
 389                                 ptlrpc_reqset_put(ps);
 390                         } while (rc == 0 && pc->pc_cursor != first);
 391                 }
 392 #endif
 393         }
 394
 395         RETURN(rc);
 396 }
 397
 398 #ifdef __KERNEL__
 399 /**
 400  * Main ptlrpcd thread.
 401  * ptlrpc's code paths like to execute in process context, so we have this
 402  * thread which spins on a set which contains the rpcs and sends them.
 403  *
 404  */
 405 static int ptlrpcd(void *arg)
 406 {
 407         struct ptlrpcd_ctl *pc = arg;
 408         struct ptlrpc_request_set *set = pc->pc_set;
 409         struct lu_context ses = { 0 };
 410         struct lu_env env = { .le_ses = &ses };
 411         int rc, exit = 0;
 412         ENTRY;
 413
 414         unshare_fs_struct();
 415 #if defined(CONFIG_SMP)
 416         if (test_bit(LIOD_BIND, &pc->pc_flags)) {
 417                 int index = pc->pc_index;
 418
 419                 if (index >= 0 && index < num_possible_cpus()) {
 420                         while (!cpu_online(index)) {
 421                                 if (++index >= num_possible_cpus())
 422                                         index = 0;
 423                         }
 424                         set_cpus_allowed_ptr(current,
 425                                      cpumask_of_node(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         if (rc == 0) {
 437                 rc = lu_context_init(env.le_ses,
 438                                      LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
 439                 if (rc != 0)
 440                         lu_context_fini(&env.le_ctx);
 441         }
 442         complete(&pc->pc_starting);
 443
 444         if (rc != 0)
 445                 RETURN(rc);
 446
 447         /*
 448          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 449          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 450          * there are requests in the set. New requests come in on the set's
 451          * new_req_list and ptlrpcd_check() moves them into the set.
 452          */
 453         do {
 454                 struct l_wait_info lwi;
 455                 int timeout;
 456
 457                 timeout = ptlrpc_set_next_timeout(set);
 458                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 459                                   ptlrpc_expired_set, set);
 460
 461                 lu_context_enter(&env.le_ctx);
 462                 lu_context_enter(env.le_ses);
 463                 l_wait_event(set->set_waitq, ptlrpcd_check(&env, pc), &lwi);
 464                 lu_context_exit(&env.le_ctx);
 465                 lu_context_exit(env.le_ses);
 466
 467                 /*
 468                  * Abort inflight rpcs for forced stop case.
 469                  */
 470                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 471                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 472                                 ptlrpc_abort_set(set);
 473                         exit++;
 474                 }
 475
 476                 /*
 477                  * Let's make one more loop to make sure that ptlrpcd_check()
 478                  * copied all raced new rpcs into the set so we can kill them.
 479                  */
 480         } while (exit < 2);
 481
 482         /*
 483          * Wait for inflight requests to drain.
 484          */
 485         if (!cfs_list_empty(&set->set_requests))
 486                 ptlrpc_set_wait(set);
 487         lu_context_fini(&env.le_ctx);
 488         lu_context_fini(env.le_ses);
 489
 490         complete(&pc->pc_finishing);
 491
 492         return 0;
 493 }
 494
 495 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 496  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 497  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 498  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 499  *      because of some CPU core(s) busy with other loads.
 500  *
 501  *      For example: "ls -l", some async RPCs for statahead are assigned to
 502  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 503  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 504  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 505  *      such case, the statahead async RPCs can not be processed in time, it is
 506  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 507  *      be better. But it breaks former data transfer policy.
 508  *
 509  *      So we shouldn't be blind for avoiding the data transfer. We make some
 510  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 511  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 512  *      core. The other part is for free mode, all the ptlrpcd threads in the
 513  *      part can be scheduled on any CPU core. We specify some partnership
 514  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 515  *      and the async RPC load within the partners are shared.
 516  *
 517  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 518  *      thread can be scheduled in time), and try to guarantee the async RPC
 519  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 520  *      on any CPU core).
 521  *
 522  *      As for how to specify the partnership between bound mode ptlrpcd
 523  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 524  *      <free bound> pair. In future, we can specify some more complex
 525  *      partnership based on the patches for CPU partition. But before such
 526  *      patches are available, we prefer to use the simplest one.
 527  */
 528 # ifdef CFS_CPU_MODE_NUMA
 529 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 530 # endif
 531 static int ptlrpcd_bind(int index, int max)
 532 {
 533         struct ptlrpcd_ctl *pc;
 534         int rc = 0;
 535 #if defined(CONFIG_NUMA)
 536         cpumask_t mask;
 537 #endif
 538         ENTRY;
 539
 540         LASSERT(index <= max - 1);
 541         pc = &ptlrpcds->pd_threads[index];
 542         switch (ptlrpcd_bind_policy) {
 543         case PDB_POLICY_NONE:
 544                 pc->pc_npartners = -1;
 545                 break;
 546         case PDB_POLICY_FULL:
 547                 pc->pc_npartners = 0;
 548                 set_bit(LIOD_BIND, &pc->pc_flags);
 549                 break;
 550         case PDB_POLICY_PAIR:
 551                 LASSERT(max % 2 == 0);
 552                 pc->pc_npartners = 1;
 553                 break;
 554         case PDB_POLICY_NEIGHBOR:
 555 #if defined(CONFIG_NUMA)
 556         {
 557                 int i;
 558                 mask = *cpumask_of_node(cpu_to_node(index));
 559                 for (i = max; i < num_online_cpus(); i++)
 560                         cpu_clear(i, mask);
 561                 pc->pc_npartners = cpus_weight(mask) - 1;
 562                 set_bit(LIOD_BIND, &pc->pc_flags);
 563         }
 564 #else
 565                 LASSERT(max >= 3);
 566                 pc->pc_npartners = 2;
 567 #endif
 568                 break;
 569         default:
 570                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 571                 rc = -EINVAL;
 572         }
 573
 574         if (rc == 0 && pc->pc_npartners > 0) {
 575                 OBD_ALLOC(pc->pc_partners,
 576                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 577                 if (pc->pc_partners == NULL) {
 578                         pc->pc_npartners = 0;
 579                         rc = -ENOMEM;
 580                 } else {
 581                         switch (ptlrpcd_bind_policy) {
 582                         case PDB_POLICY_PAIR:
 583                                 if (index & 0x1) {
 584                                         set_bit(LIOD_BIND, &pc->pc_flags);
 585                                         pc->pc_partners[0] = &ptlrpcds->
 586                                                 pd_threads[index - 1];
 587                                         ptlrpcds->pd_threads[index - 1].
 588                                                 pc_partners[0] = pc;
 589                                 }
 590                                 break;
 591                         case PDB_POLICY_NEIGHBOR:
 592 #if defined(CONFIG_NUMA)
 593                         {
 594                                 struct ptlrpcd_ctl *ppc;
 595                                 int i, pidx;
 596                                 /* partners are cores in the same NUMA node.
 597                                  * setup partnership only with ptlrpcd threads
 598                                  * that are already initialized
 599                                  */
 600                                 for (pidx = 0, i = 0; i < index; i++) {
 601                                         if (cpu_isset(i, mask)) {
 602                                                 ppc = &ptlrpcds->pd_threads[i];
 603                                                 pc->pc_partners[pidx++] = ppc;
 604                                                 ppc->pc_partners[ppc->
 605                                                           pc_npartners++] = pc;
 606                                         }
 607                                 }
 608                                 /* adjust number of partners to the number
 609                                  * of partnership really setup */
 610                                 pc->pc_npartners = pidx;
 611                         }
 612 #else
 613                                 if (index & 0x1)
 614                                         set_bit(LIOD_BIND, &pc->pc_flags);
 615                                 if (index > 0) {
 616                                         pc->pc_partners[0] = &ptlrpcds->
 617                                                 pd_threads[index - 1];
 618                                         ptlrpcds->pd_threads[index - 1].
 619                                                 pc_partners[1] = pc;
 620                                         if (index == max - 1) {
 621                                                 pc->pc_partners[1] =
 622                                                 &ptlrpcds->pd_threads[0];
 623                                                 ptlrpcds->pd_threads[0].
 624                                                 pc_partners[0] = pc;
 625                                         }
 626                                 }
 627 #endif
 628                                 break;
 629                         }
 630                 }
 631         }
 632
 633         RETURN(rc);
 634 }
 635
 636 #else /* !__KERNEL__ */
 637
 638 /**
 639  * In liblustre we do not have separate threads, so this function
 640  * is called from time to time all across common code to see
 641  * if something needs to be processed on ptlrpcd set.
 642  */
 643 int ptlrpcd_check_async_rpcs(void *arg)
 644 {
 645         struct ptlrpcd_ctl *pc = arg;
 646         int                 rc = 0;
 647
 648         /*
 649          * Single threaded!!
 650          */
 651         pc->pc_recurred++;
 652
 653         if (pc->pc_recurred == 1) {
 654                 rc = lu_env_refill(&pc->pc_env);
 655                 if (rc == 0) {
 656                         lu_context_enter(&pc->pc_env.le_ctx);
 657                         rc = ptlrpcd_check(&pc->pc_env, pc);
 658                         if (!rc)
 659                                 ptlrpc_expired_set(pc->pc_set);
 660                         /*
 661                          * XXX: send replay requests.
 662                          */
 663                         if (test_bit(LIOD_RECOVERY, &pc->pc_flags))
 664                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 665                         lu_context_exit(&pc->pc_env.le_ctx);
 666                 }
 667         }
 668
 669         pc->pc_recurred--;
 670         return rc;
 671 }
 672
 673 int ptlrpcd_idle(void *arg)
 674 {
 675         struct ptlrpcd_ctl *pc = arg;
 676
 677         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
 678                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 679 }
 680
 681 #endif
 682
 683 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 684 {
 685         int rc;
 686         ENTRY;
 687
 688         /*
 689          * Do not allow start second thread for one pc.
 690          */
 691         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 692                 CWARN("Starting second thread (%s) for same pc %p\n",
 693                       name, pc);
 694                 RETURN(0);
 695         }
 696
 697         pc->pc_index = index;
 698         init_completion(&pc->pc_starting);
 699         init_completion(&pc->pc_finishing);
 700         spin_lock_init(&pc->pc_lock);
 701         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 702         pc->pc_set = ptlrpc_prep_set();
 703         if (pc->pc_set == NULL)
 704                 GOTO(out, rc = -ENOMEM);
 705
 706 #ifndef __KERNEL__
 707         pc->pc_wait_callback =
 708                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 709                                                  &ptlrpcd_check_async_rpcs, pc);
 710         pc->pc_idle_callback =
 711                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 712                                                  &ptlrpcd_idle, pc);
 713         RETURN(0);
 714 #else
 715         /*
 716          * So far only "client" ptlrpcd uses an environment. In the future,
 717          * ptlrpcd thread (or a thread-set) has to be given an argument,
 718          * describing its "scope".
 719          */
 720         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 721         if (rc != 0)
 722                 GOTO(out_set, rc);
 723
 724         {
 725                 struct task_struct *task;
 726                 if (index >= 0) {
 727                         rc = ptlrpcd_bind(index, max);
 728                         if (rc < 0)
 729                                 GOTO(out_env, rc);
 730                 }
 731
 732                 task = kthread_run(ptlrpcd, pc, pc->pc_name);
 733                 if (IS_ERR(task))
 734                         GOTO(out_env, rc = PTR_ERR(task));
 735
 736                 wait_for_completion(&pc->pc_starting);
 737         }
 738         RETURN(0);
 739
 740 out_env:
 741         lu_context_fini(&pc->pc_env.le_ctx);
 742
 743 out_set:
 744         if (pc->pc_set != NULL) {
 745                 struct ptlrpc_request_set *set = pc->pc_set;
 746
 747                 spin_lock(&pc->pc_lock);
 748                 pc->pc_set = NULL;
 749                 spin_unlock(&pc->pc_lock);
 750                 ptlrpc_set_destroy(set);
 751         }
 752         clear_bit(LIOD_BIND, &pc->pc_flags);
 753 #endif
 754 out:
 755         clear_bit(LIOD_START, &pc->pc_flags);
 756         RETURN(rc);
 757 }
 758
 759 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 760 {
 761         ENTRY;
 762
 763         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 764                 CWARN("Thread for pc %p was not started\n", pc);
 765                 goto out;
 766         }
 767
 768         set_bit(LIOD_STOP, &pc->pc_flags);
 769         if (force)
 770                 set_bit(LIOD_FORCE, &pc->pc_flags);
 771         wake_up(&pc->pc_set->set_waitq);
 772
 773 out:
 774         EXIT;
 775 }
 776
 777 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 778 {
 779         struct ptlrpc_request_set *set = pc->pc_set;
 780         ENTRY;
 781
 782         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 783                 CWARN("Thread for pc %p was not started\n", pc);
 784                 goto out;
 785         }
 786
 787 #ifdef __KERNEL__
 788         wait_for_completion(&pc->pc_finishing);
 789 #else
 790         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 791         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 792 #endif
 793         lu_context_fini(&pc->pc_env.le_ctx);
 794
 795         spin_lock(&pc->pc_lock);
 796         pc->pc_set = NULL;
 797         spin_unlock(&pc->pc_lock);
 798         ptlrpc_set_destroy(set);
 799
 800         clear_bit(LIOD_START, &pc->pc_flags);
 801         clear_bit(LIOD_STOP, &pc->pc_flags);
 802         clear_bit(LIOD_FORCE, &pc->pc_flags);
 803         clear_bit(LIOD_BIND, &pc->pc_flags);
 804
 805 out:
 806 #ifdef __KERNEL__
 807         if (pc->pc_npartners > 0) {
 808                 LASSERT(pc->pc_partners != NULL);
 809
 810                 OBD_FREE(pc->pc_partners,
 811                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 812                 pc->pc_partners = NULL;
 813         }
 814         pc->pc_npartners = 0;
 815 #endif
 816         EXIT;
 817 }
 818
 819 static void ptlrpcd_fini(void)
 820 {
 821         int i;
 822         ENTRY;
 823
 824         if (ptlrpcds != NULL) {
 825                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 826                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 827                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 828                         ptlrpcd_free(&ptlrpcds->pd_threads[i]);
 829                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 830                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 831                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 832                 ptlrpcds = NULL;
 833         }
 834
 835         EXIT;
 836 }
 837
 838 static int ptlrpcd_init(void)
 839 {
 840         int     nthreads = num_online_cpus();
 841         char    name[16];
 842         int     size, i = -1, j, rc = 0;
 843         ENTRY;
 844
 845 #ifdef __KERNEL__
 846         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 847                 nthreads = max_ptlrpcds;
 848         if (nthreads < 2)
 849                 nthreads = 2;
 850         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 851                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 852         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 853                 nthreads &= ~1; /* make sure it is even */
 854 #else
 855         nthreads = 1;
 856 #endif
 857
 858         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 859         OBD_ALLOC(ptlrpcds, size);
 860         if (ptlrpcds == NULL)
 861                 GOTO(out, rc = -ENOMEM);
 862
 863         snprintf(name, 15, "ptlrpcd_rcv");
 864         set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 865         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 866         if (rc < 0)
 867                 GOTO(out, rc);
 868
 869         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 870          *      non-recovery async RPC to improve overall async RPC efficiency.
 871          *
 872          *      But there are some issues with async I/O RPCs and async non-I/O
 873          *      RPCs processed in the same set under some cases. The ptlrpcd may
 874          *      be blocked by some async I/O RPC(s), then will cause other async
 875          *      non-I/O RPC(s) can not be processed in time.
 876          *
 877          *      Maybe we should distinguish blocked async RPCs from non-blocked
 878          *      async RPCs, and process them in different ptlrpcd sets to avoid
 879          *      unnecessary dependency. But how to distribute async RPCs load
 880          *      among all the ptlrpc daemons becomes another trouble. */
 881         for (i = 0; i < nthreads; i++) {
 882                 snprintf(name, 15, "ptlrpcd_%d", i);
 883                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 884                 if (rc < 0)
 885                         GOTO(out, rc);
 886         }
 887
 888         ptlrpcds->pd_size = size;
 889         ptlrpcds->pd_index = 0;
 890         ptlrpcds->pd_nthreads = nthreads;
 891
 892 out:
 893         if (rc != 0 && ptlrpcds != NULL) {
 894                 for (j = 0; j <= i; j++)
 895                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 896                 for (j = 0; j <= i; j++)
 897                         ptlrpcd_free(&ptlrpcds->pd_threads[j]);
 898                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 899                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 900                 OBD_FREE(ptlrpcds, size);
 901                 ptlrpcds = NULL;
 902         }
 903
 904         RETURN(rc);
 905 }
 906
 907 int ptlrpcd_addref(void)
 908 {
 909         int rc = 0;
 910         ENTRY;
 911
 912         mutex_lock(&ptlrpcd_mutex);
 913         if (++ptlrpcd_users == 1) {
 914                 rc = ptlrpcd_init();
 915                 if (rc < 0)
 916                         ptlrpcd_users--;
 917         }
 918         mutex_unlock(&ptlrpcd_mutex);
 919         RETURN(rc);
 920 }
 921 EXPORT_SYMBOL(ptlrpcd_addref);
 922
 923 void ptlrpcd_decref(void)
 924 {
 925         mutex_lock(&ptlrpcd_mutex);
 926         if (--ptlrpcd_users == 0)
 927                 ptlrpcd_fini();
 928         mutex_unlock(&ptlrpcd_mutex);
 929 }
 930 EXPORT_SYMBOL(ptlrpcd_decref);
 931 /** @} ptlrpcd */