lustre/ptlrpc/ptlrpcd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/ptlrpc/ptlrpcd.c
  40  */
  41
  42 /** \defgroup ptlrpcd PortalRPC daemon
  43  *
  44  * ptlrpcd is a special thread with its own set where other user might add
  45  * requests when they don't want to wait for their completion.
  46  * PtlRPCD will take care of sending such requests and then processing their
  47  * replies and calling completion callbacks as necessary.
  48  * The callbacks are called directly from ptlrpcd context.
  49  * It is important to never significantly block (esp. on RPCs!) within such
  50  * completion handler or a deadlock might occur where ptlrpcd enters some
  51  * callback that attempts to send another RPC and wait for it to return,
  52  * during which time ptlrpcd is completely blocked, so e.g. if import
  53  * fails, recovery cannot progress because connection requests are also
  54  * sent by ptlrpcd.
  55  *
  56  * @{
  57  */
  58
  59 #define DEBUG_SUBSYSTEM S_RPC
  60
  61 #ifdef __KERNEL__
  62 # include <libcfs/libcfs.h>
  63 #else /* __KERNEL__ */
  64 # include <liblustre.h>
  65 # include <ctype.h>
  66 #endif
  67
  68 #include <lustre_net.h>
  69 # include <lustre_lib.h>
  70
  71 #include <lustre_ha.h>
  72 #include <obd_class.h>   /* for obd_zombie */
  73 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  74 #include <cl_object.h> /* cl_env_{get,put}() */
  75 #include <lprocfs_status.h>
  76
  77 #include "ptlrpc_internal.h"
  78
  79 struct ptlrpcd {
  80         int                pd_size;
  81         int                pd_index;
  82         int                pd_nthreads;
  83         struct ptlrpcd_ctl pd_thread_rcv;
  84         struct ptlrpcd_ctl pd_threads[0];
  85 };
  86
  87 #ifdef __KERNEL__
  88 static int max_ptlrpcds;
  89 CFS_MODULE_PARM(max_ptlrpcds, "i", int, 0644,
  90                 "Max ptlrpcd thread count to be started.");
  91
  92 static int ptlrpcd_bind_policy = PDB_POLICY_PAIR;
  93 CFS_MODULE_PARM(ptlrpcd_bind_policy, "i", int, 0644,
  94                 "Ptlrpcd threads binding mode.");
  95 #endif
  96 static struct ptlrpcd *ptlrpcds;
  97
  98 cfs_semaphore_t ptlrpcd_sem;
  99 static int ptlrpcd_users = 0;
 100
 101 void ptlrpcd_wake(struct ptlrpc_request *req)
 102 {
 103         struct ptlrpc_request_set *rq_set = req->rq_set;
 104
 105         LASSERT(rq_set != NULL);
 106
 107         cfs_waitq_signal(&rq_set->set_waitq);
 108 }
 109
 110 static struct ptlrpcd_ctl *
 111 ptlrpcd_select_pc(struct ptlrpc_request *req, pdl_policy_t policy, int index)
 112 {
 113         int idx = 0;
 114
 115         if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
 116                 return &ptlrpcds->pd_thread_rcv;
 117
 118 #ifdef __KERNEL__
 119         switch (policy) {
 120         case PDL_POLICY_SAME:
 121                 idx = cfs_smp_processor_id() % ptlrpcds->pd_nthreads;
 122                 break;
 123         case PDL_POLICY_LOCAL:
 124                 /* Before CPU partition patches available, process it the same
 125                  * as "PDL_POLICY_ROUND". */
 126 # ifdef CFS_CPU_MODE_NUMA
 127 # warning "fix this code to use new CPU partition APIs"
 128 # endif
 129                 /* Fall through to PDL_POLICY_ROUND until the CPU
 130                  * CPU partition patches are available. */
 131                 index = -1;
 132         case PDL_POLICY_PREFERRED:
 133                 if (index >= 0 && index < cfs_num_online_cpus()) {
 134                         idx = index % ptlrpcds->pd_nthreads;
 135                         break;
 136                 }
 137                 /* Fall through to PDL_POLICY_ROUND for bad index. */
 138         default:
 139                 /* Fall through to PDL_POLICY_ROUND for unknown policy. */
 140         case PDL_POLICY_ROUND:
 141                 /* We do not care whether it is strict load balance. */
 142                 idx = ptlrpcds->pd_index + 1;
 143                 if (idx == cfs_smp_processor_id())
 144                         idx++;
 145                 idx %= ptlrpcds->pd_nthreads;
 146                 ptlrpcds->pd_index = idx;
 147                 break;
 148         }
 149 #endif /* __KERNEL__ */
 150
 151         return &ptlrpcds->pd_threads[idx];
 152 }
 153
 154 /**
 155  * Move all request from an existing request set to the ptlrpcd queue.
 156  * All requests from the set must be in phase RQ_PHASE_NEW.
 157  */
 158 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 159 {
 160         cfs_list_t *tmp, *pos;
 161 #ifdef __KERNEL__
 162         struct ptlrpcd_ctl *pc;
 163         struct ptlrpc_request_set *new;
 164         int count, i;
 165
 166         pc = ptlrpcd_select_pc(NULL, PDL_POLICY_LOCAL, -1);
 167         new = pc->pc_set;
 168 #endif
 169
 170         cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 171                 struct ptlrpc_request *req =
 172                         cfs_list_entry(pos, struct ptlrpc_request,
 173                                        rq_set_chain);
 174
 175                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 176 #ifdef __KERNEL__
 177                 req->rq_set = new;
 178                 req->rq_queued_time = cfs_time_current();
 179 #else
 180                 cfs_list_del_init(&req->rq_set_chain);
 181                 req->rq_set = NULL;
 182                 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
 183                 cfs_atomic_dec(&set->set_remaining);
 184 #endif
 185         }
 186
 187 #ifdef __KERNEL__
 188         cfs_spin_lock(&new->set_new_req_lock);
 189         cfs_list_splice_init(&set->set_requests, &new->set_new_requests);
 190         i = cfs_atomic_read(&set->set_remaining);
 191         count = cfs_atomic_add_return(i, &new->set_new_count);
 192         cfs_atomic_set(&set->set_remaining, 0);
 193         cfs_spin_unlock(&new->set_new_req_lock);
 194         if (count == i) {
 195                 cfs_waitq_signal(&new->set_waitq);
 196
 197                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
 198                  *      guarantee the async RPC can be processed ASAP, we have
 199                  *      no other better choice. It maybe fixed in future. */
 200                 for (i = 0; i < pc->pc_npartners; i++)
 201                         cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
 202         }
 203 #endif
 204 }
 205 EXPORT_SYMBOL(ptlrpcd_add_rqset);
 206
 207 #ifdef __KERNEL__
 208 /**
 209  * Return transferred RPCs count.
 210  */
 211 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 212                                struct ptlrpc_request_set *src)
 213 {
 214         cfs_list_t *tmp, *pos;
 215         struct ptlrpc_request *req;
 216         int rc = 0;
 217
 218         cfs_spin_lock(&src->set_new_req_lock);
 219         if (likely(!cfs_list_empty(&src->set_new_requests))) {
 220                 cfs_list_for_each_safe(pos, tmp, &src->set_new_requests) {
 221                         req = cfs_list_entry(pos, struct ptlrpc_request,
 222                                              rq_set_chain);
 223                         req->rq_set = des;
 224                 }
 225                 cfs_list_splice_init(&src->set_new_requests,
 226                                      &des->set_requests);
 227                 rc = cfs_atomic_read(&src->set_new_count);
 228                 cfs_atomic_add(rc, &des->set_remaining);
 229                 cfs_atomic_set(&src->set_new_count, 0);
 230         }
 231         cfs_spin_unlock(&src->set_new_req_lock);
 232         return rc;
 233 }
 234 #endif
 235
 236 /**
 237  * Requests that are added to the ptlrpcd queue are sent via
 238  * ptlrpcd_check->ptlrpc_check_set().
 239  */
 240 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx)
 241 {
 242         struct ptlrpcd_ctl *pc;
 243
 244         cfs_spin_lock(&req->rq_lock);
 245         if (req->rq_invalid_rqset) {
 246                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 247                                                      back_to_sleep, NULL);
 248
 249                 req->rq_invalid_rqset = 0;
 250                 cfs_spin_unlock(&req->rq_lock);
 251                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 252         } else if (req->rq_set) {
 253                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 254                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 255
 256                 /* ptlrpc_check_set will decrease the count */
 257                 cfs_atomic_inc(&req->rq_set->set_remaining);
 258                 cfs_spin_unlock(&req->rq_lock);
 259                 cfs_waitq_signal(&req->rq_set->set_waitq);
 260                 return;
 261         } else {
 262                 cfs_spin_unlock(&req->rq_lock);
 263         }
 264
 265         pc = ptlrpcd_select_pc(req, policy, idx);
 266
 267         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 268                   req, pc->pc_name, pc->pc_index);
 269
 270         ptlrpc_set_add_new_req(pc, req);
 271 }
 272
 273 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 274 {
 275         cfs_atomic_inc(&set->set_refcount);
 276 }
 277
 278 /**
 279  * Check if there is more work to do on ptlrpcd set.
 280  * Returns 1 if yes.
 281  */
 282 static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc)
 283 {
 284         cfs_list_t *tmp, *pos;
 285         struct ptlrpc_request *req;
 286         struct ptlrpc_request_set *set = pc->pc_set;
 287         int rc = 0;
 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         if (cfs_atomic_read(&set->set_remaining))
 307                 rc |= ptlrpc_check_set(env, set);
 308
 309         if (!cfs_list_empty(&set->set_requests)) {
 310                 /*
 311                  * XXX: our set never completes, so we prune the completed
 312                  * reqs after each iteration. boy could this be smarter.
 313                  */
 314                 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 315                         req = cfs_list_entry(pos, struct ptlrpc_request,
 316                                              rq_set_chain);
 317                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 318                                 continue;
 319
 320                         cfs_list_del_init(&req->rq_set_chain);
 321                         req->rq_set = NULL;
 322                         ptlrpc_req_finished(req);
 323                 }
 324         }
 325
 326         if (rc == 0) {
 327                 /*
 328                  * If new requests have been added, make sure to wake up.
 329                  */
 330                 rc = cfs_atomic_read(&set->set_new_count);
 331
 332 #ifdef __KERNEL__
 333                 if (rc == 0 && pc->pc_npartners > 0) {
 334                         struct ptlrpcd_ctl *partner;
 335                         struct ptlrpc_request_set *ps;
 336                         int first = pc->pc_cursor;
 337
 338                         do {
 339                                 partner = pc->pc_partners[pc->pc_cursor++];
 340                                 if (pc->pc_cursor >= pc->pc_npartners)
 341                                         pc->pc_cursor = 0;
 342                                 if (partner == NULL)
 343                                         continue;
 344
 345                                 cfs_spin_lock(&partner->pc_lock);
 346                                 ps = partner->pc_set;
 347                                 if (ps == NULL) {
 348                                         cfs_spin_unlock(&partner->pc_lock);
 349                                         continue;
 350                                 }
 351
 352                                 ptlrpc_reqset_get(ps);
 353                                 cfs_spin_unlock(&partner->pc_lock);
 354
 355                                 if (cfs_atomic_read(&ps->set_new_count)) {
 356                                         rc = ptlrpcd_steal_rqset(set, ps);
 357                                         if (rc > 0)
 358                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 359                                                        " async RPCs [%d->%d]\n",
 360                                                         rc, pc->pc_index,
 361                                                         partner->pc_index);
 362                                 }
 363                                 ptlrpc_reqset_put(ps);
 364                         } while (rc == 0 && pc->pc_cursor != first);
 365                 }
 366 #endif
 367         }
 368
 369         RETURN(rc);
 370 }
 371
 372 #ifdef __KERNEL__
 373 /**
 374  * Main ptlrpcd thread.
 375  * ptlrpc's code paths like to execute in process context, so we have this
 376  * thread which spins on a set which contains the rpcs and sends them.
 377  *
 378  */
 379 static int ptlrpcd(void *arg)
 380 {
 381         struct ptlrpcd_ctl *pc = arg;
 382         struct ptlrpc_request_set *set = pc->pc_set;
 383         struct lu_env env = { .le_ses = NULL };
 384         int rc, exit = 0;
 385         ENTRY;
 386
 387         cfs_daemonize_ctxt(pc->pc_name);
 388 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
 389         if (cfs_test_bit(LIOD_BIND, &pc->pc_flags)) {
 390                 int index = pc->pc_index;
 391
 392                 if (index >= 0 && index < cfs_num_possible_cpus()) {
 393                         while (!cfs_cpu_online(index)) {
 394                                 if (++index >= cfs_num_possible_cpus())
 395                                         index = 0;
 396                         }
 397                         cfs_set_cpus_allowed(cfs_current(),
 398                                      node_to_cpumask(cpu_to_node(index)));
 399                 }
 400         }
 401 #endif
 402         /*
 403          * XXX So far only "client" ptlrpcd uses an environment. In
 404          * the future, ptlrpcd thread (or a thread-set) has to given
 405          * an argument, describing its "scope".
 406          */
 407         rc = lu_context_init(&env.le_ctx,
 408                              LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 409         cfs_complete(&pc->pc_starting);
 410
 411         if (rc != 0)
 412                 RETURN(rc);
 413
 414         /*
 415          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 416          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 417          * there are requests in the set. New requests come in on the set's
 418          * new_req_list and ptlrpcd_check() moves them into the set.
 419          */
 420         do {
 421                 struct l_wait_info lwi;
 422                 int timeout;
 423
 424                 rc = lu_env_refill(&env);
 425                 if (rc != 0) {
 426                         /*
 427                          * XXX This is very awkward situation, because
 428                          * execution can neither continue (request
 429                          * interpreters assume that env is set up), nor repeat
 430                          * the loop (as this potentially results in a tight
 431                          * loop of -ENOMEM's).
 432                          *
 433                          * Fortunately, refill only ever does something when
 434                          * new modules are loaded, i.e., early during boot up.
 435                          */
 436                         CERROR("Failure to refill session: %d\n", rc);
 437                         continue;
 438                 }
 439
 440                 timeout = ptlrpc_set_next_timeout(set);
 441                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 442                                   ptlrpc_expired_set, set);
 443
 444                 lu_context_enter(&env.le_ctx);
 445                 l_wait_event(set->set_waitq,
 446                              ptlrpcd_check(&env, pc), &lwi);
 447                 lu_context_exit(&env.le_ctx);
 448
 449                 /*
 450                  * Abort inflight rpcs for forced stop case.
 451                  */
 452                 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
 453                         if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
 454                                 ptlrpc_abort_set(set);
 455                         exit++;
 456                 }
 457
 458                 /*
 459                  * Let's make one more loop to make sure that ptlrpcd_check()
 460                  * copied all raced new rpcs into the set so we can kill them.
 461                  */
 462         } while (exit < 2);
 463
 464         /*
 465          * Wait for inflight requests to drain.
 466          */
 467         if (!cfs_list_empty(&set->set_requests))
 468                 ptlrpc_set_wait(set);
 469         lu_context_fini(&env.le_ctx);
 470         cfs_complete(&pc->pc_finishing);
 471
 472         cfs_clear_bit(LIOD_START, &pc->pc_flags);
 473         cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
 474         cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
 475         cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 476         return 0;
 477 }
 478
 479 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 480  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 481  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 482  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 483  *      because of some CPU core(s) busy with other loads.
 484  *
 485  *      For example: "ls -l", some async RPCs for statahead are assigned to
 486  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 487  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 488  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 489  *      such case, the statahead async RPCs can not be processed in time, it is
 490  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 491  *      be better. But it breaks former data transfer policy.
 492  *
 493  *      So we shouldn't be blind for avoiding the data transfer. We make some
 494  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 495  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 496  *      core. The other part is for free mode, all the ptlrpcd threads in the
 497  *      part can be scheduled on any CPU core. We specify some partnership
 498  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 499  *      and the async RPC load within the partners are shared.
 500  *
 501  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 502  *      thread can be scheduled in time), and try to guarantee the async RPC
 503  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 504  *      on any CPU core).
 505  *
 506  *      As for how to specify the partnership between bound mode ptlrpcd
 507  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 508  *      <free bound> pair. In future, we can specify some more complex
 509  *      partnership based on the patches for CPU partition. But before such
 510  *      patches are available, we prefer to use the simplest one.
 511  */
 512 # ifdef CFS_CPU_MODE_NUMA
 513 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 514 # endif
 515 static int ptlrpcd_bind(int index, int max)
 516 {
 517         struct ptlrpcd_ctl *pc;
 518         int rc = 0;
 519         ENTRY;
 520
 521         LASSERT(index <= max - 1);
 522         pc = &ptlrpcds->pd_threads[index];
 523         switch (ptlrpcd_bind_policy) {
 524         case PDB_POLICY_NONE:
 525                 pc->pc_npartners = -1;
 526                 break;
 527         case PDB_POLICY_FULL:
 528                 pc->pc_npartners = 0;
 529                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 530                 break;
 531         case PDB_POLICY_PAIR:
 532                 LASSERT(max % 2 == 0);
 533                 pc->pc_npartners = 1;
 534                 break;
 535         case PDB_POLICY_NEIGHBOR:
 536                 LASSERT(max >= 3);
 537                 pc->pc_npartners = 2;
 538                 break;
 539         default:
 540                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 541                 rc = -EINVAL;
 542         }
 543
 544         if (rc == 0 && pc->pc_npartners > 0) {
 545                 OBD_ALLOC(pc->pc_partners,
 546                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 547                 if (pc->pc_partners == NULL) {
 548                         pc->pc_npartners = 0;
 549                         rc = -ENOMEM;
 550                 } else {
 551                         if (index & 0x1)
 552                                 cfs_set_bit(LIOD_BIND, &pc->pc_flags);
 553
 554                         switch (ptlrpcd_bind_policy) {
 555                         case PDB_POLICY_PAIR:
 556                                 if (index & 0x1) {
 557                                         pc->pc_partners[0] = &ptlrpcds->
 558                                                 pd_threads[index - 1];
 559                                         ptlrpcds->pd_threads[index - 1].
 560                                                 pc_partners[0] = pc;
 561                                 }
 562                                 break;
 563                         case PDB_POLICY_NEIGHBOR:
 564                                 if (index > 0) {
 565                                         pc->pc_partners[0] = &ptlrpcds->
 566                                                 pd_threads[index - 1];
 567                                         ptlrpcds->pd_threads[index - 1].
 568                                                 pc_partners[1] = pc;
 569                                         if (index == max - 1) {
 570                                                 pc->pc_partners[1] =
 571                                                 &ptlrpcds->pd_threads[0];
 572                                                 ptlrpcds->pd_threads[0].
 573                                                 pc_partners[0] = pc;
 574                                         }
 575                                 }
 576                                 break;
 577                         }
 578                 }
 579         }
 580
 581         RETURN(rc);
 582 }
 583
 584 #else /* !__KERNEL__ */
 585
 586 /**
 587  * In liblustre we do not have separate threads, so this function
 588  * is called from time to time all across common code to see
 589  * if something needs to be processed on ptlrpcd set.
 590  */
 591 int ptlrpcd_check_async_rpcs(void *arg)
 592 {
 593         struct ptlrpcd_ctl *pc = arg;
 594         int                 rc = 0;
 595
 596         /*
 597          * Single threaded!!
 598          */
 599         pc->pc_recurred++;
 600
 601         if (pc->pc_recurred == 1) {
 602                 rc = lu_env_refill(&pc->pc_env);
 603                 if (rc == 0) {
 604                         lu_context_enter(&pc->pc_env.le_ctx);
 605                         rc = ptlrpcd_check(&pc->pc_env, pc);
 606                         if (!rc)
 607                                 ptlrpc_expired_set(pc->pc_set);
 608                         /*
 609                          * XXX: send replay requests.
 610                          */
 611                         if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
 612                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 613                         lu_context_exit(&pc->pc_env.le_ctx);
 614                 }
 615         }
 616
 617         pc->pc_recurred--;
 618         return rc;
 619 }
 620
 621 int ptlrpcd_idle(void *arg)
 622 {
 623         struct ptlrpcd_ctl *pc = arg;
 624
 625         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
 626                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 627 }
 628
 629 #endif
 630
 631 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 632 {
 633         int rc;
 634         int env = 0;
 635         ENTRY;
 636
 637         /*
 638          * Do not allow start second thread for one pc.
 639          */
 640         if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 641                 CWARN("Starting second thread (%s) for same pc %p\n",
 642                        name, pc);
 643                 RETURN(0);
 644         }
 645
 646         pc->pc_index = index;
 647         cfs_init_completion(&pc->pc_starting);
 648         cfs_init_completion(&pc->pc_finishing);
 649         cfs_spin_lock_init(&pc->pc_lock);
 650         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 651         pc->pc_set = ptlrpc_prep_set();
 652         if (pc->pc_set == NULL)
 653                 GOTO(out, rc = -ENOMEM);
 654         /*
 655          * So far only "client" ptlrpcd uses an environment. In the future,
 656          * ptlrpcd thread (or a thread-set) has to be given an argument,
 657          * describing its "scope".
 658          */
 659         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 660         if (rc != 0)
 661                 GOTO(out, rc);
 662
 663         env = 1;
 664 #ifdef __KERNEL__
 665         if (index >= 0) {
 666                 rc = ptlrpcd_bind(index, max);
 667                 if (rc < 0)
 668                         GOTO(out, rc);
 669         }
 670
 671         rc = cfs_create_thread(ptlrpcd, pc, 0);
 672         if (rc < 0)
 673                 GOTO(out, rc);
 674
 675         rc = 0;
 676         cfs_wait_for_completion(&pc->pc_starting);
 677 #else
 678         pc->pc_wait_callback =
 679                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 680                                                  &ptlrpcd_check_async_rpcs, pc);
 681         pc->pc_idle_callback =
 682                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 683                                                  &ptlrpcd_idle, pc);
 684 #endif
 685 out:
 686         if (rc) {
 687 #ifdef __KERNEL__
 688                 if (pc->pc_set != NULL) {
 689                         struct ptlrpc_request_set *set = pc->pc_set;
 690
 691                         cfs_spin_lock(&pc->pc_lock);
 692                         pc->pc_set = NULL;
 693                         cfs_spin_unlock(&pc->pc_lock);
 694                         ptlrpc_set_destroy(set);
 695                 }
 696                 if (env != 0)
 697                         lu_context_fini(&pc->pc_env.le_ctx);
 698                 cfs_clear_bit(LIOD_BIND, &pc->pc_flags);
 699 #endif
 700                 cfs_clear_bit(LIOD_START, &pc->pc_flags);
 701         }
 702         RETURN(rc);
 703 }
 704
 705 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 706 {
 707        struct ptlrpc_request_set *set = pc->pc_set;
 708         ENTRY;
 709
 710         if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
 711                 CWARN("Thread for pc %p was not started\n", pc);
 712                 goto out;
 713         }
 714
 715         cfs_set_bit(LIOD_STOP, &pc->pc_flags);
 716         if (force)
 717                 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
 718         cfs_waitq_signal(&pc->pc_set->set_waitq);
 719 #ifdef __KERNEL__
 720         cfs_wait_for_completion(&pc->pc_finishing);
 721 #else
 722         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 723         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 724 #endif
 725         lu_context_fini(&pc->pc_env.le_ctx);
 726
 727         cfs_spin_lock(&pc->pc_lock);
 728         pc->pc_set = NULL;
 729         cfs_spin_unlock(&pc->pc_lock);
 730         ptlrpc_set_destroy(set);
 731
 732 out:
 733 #ifdef __KERNEL__
 734         if (pc->pc_npartners > 0) {
 735                 LASSERT(pc->pc_partners != NULL);
 736
 737                 OBD_FREE(pc->pc_partners,
 738                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 739                 pc->pc_partners = NULL;
 740         }
 741         pc->pc_npartners = 0;
 742 #endif
 743         EXIT;
 744 }
 745
 746 static void ptlrpcd_fini(void)
 747 {
 748         int i;
 749         ENTRY;
 750
 751         if (ptlrpcds != NULL) {
 752                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 753                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 754                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 755                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 756                 ptlrpcds = NULL;
 757         }
 758
 759         EXIT;
 760 }
 761
 762 static int ptlrpcd_init(void)
 763 {
 764         int nthreads = cfs_num_online_cpus();
 765         char name[16];
 766         int size, i = -1, j, rc = 0;
 767         ENTRY;
 768
 769 #ifdef __KERNEL__
 770         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 771                 nthreads = max_ptlrpcds;
 772         if (nthreads < 2)
 773                 nthreads = 2;
 774         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 775                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 776         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 777                 nthreads &= ~1; /* make sure it is even */
 778 #else
 779         nthreads = 1;
 780 #endif
 781
 782         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 783         OBD_ALLOC(ptlrpcds, size);
 784         if (ptlrpcds == NULL)
 785                 GOTO(out, rc = -ENOMEM);
 786
 787         snprintf(name, 15, "ptlrpcd_rcv");
 788         cfs_set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 789         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 790         if (rc < 0)
 791                 GOTO(out, rc);
 792
 793         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 794          *      non-recovery async RPC to improve overall async RPC efficiency.
 795          *
 796          *      But there are some issues with async I/O RPCs and async non-I/O
 797          *      RPCs processed in the same set under some cases. The ptlrpcd may
 798          *      be blocked by some async I/O RPC(s), then will cause other async
 799          *      non-I/O RPC(s) can not be processed in time.
 800          *
 801          *      Maybe we should distinguish blocked async RPCs from non-blocked
 802          *      async RPCs, and process them in different ptlrpcd sets to avoid
 803          *      unnecessary dependency. But how to distribute async RPCs load
 804          *      among all the ptlrpc daemons becomes another trouble. */
 805         for (i = 0; i < nthreads; i++) {
 806                 snprintf(name, 15, "ptlrpcd_%d", i);
 807                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 808                 if (rc < 0)
 809                         GOTO(out, rc);
 810         }
 811
 812         ptlrpcds->pd_size = size;
 813         ptlrpcds->pd_index = 0;
 814         ptlrpcds->pd_nthreads = nthreads;
 815
 816 out:
 817         if (rc != 0 && ptlrpcds != NULL) {
 818                 for (j = 0; j <= i; j++)
 819                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 820                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 821                 OBD_FREE(ptlrpcds, size);
 822                 ptlrpcds = NULL;
 823         }
 824
 825         RETURN(0);
 826 }
 827
 828 int ptlrpcd_addref(void)
 829 {
 830         int rc = 0;
 831         ENTRY;
 832
 833         cfs_mutex_down(&ptlrpcd_sem);
 834         if (++ptlrpcd_users == 1)
 835                 rc = ptlrpcd_init();
 836         cfs_mutex_up(&ptlrpcd_sem);
 837         RETURN(rc);
 838 }
 839
 840 void ptlrpcd_decref(void)
 841 {
 842         cfs_mutex_down(&ptlrpcd_sem);
 843         if (--ptlrpcd_users == 0)
 844                 ptlrpcd_fini();
 845         cfs_mutex_up(&ptlrpcd_sem);
 846 }
 847 /** @} ptlrpcd */