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 #include <libcfs/libcfs.h>
  59 #include <lustre_net.h>
  60 #include <lustre_lib.h>
  61 #include <lustre_ha.h>
  62 #include <obd_class.h>   /* for obd_zombie */
  63 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  64 #include <cl_object.h> /* cl_env_{get,put}() */
  65 #include <lprocfs_status.h>
  66
  67 #include "ptlrpc_internal.h"
  68
  69 struct ptlrpcd {
  70         int                pd_size;
  71         int                pd_index;
  72         int                pd_nthreads;
  73         struct ptlrpcd_ctl pd_thread_rcv;
  74         struct ptlrpcd_ctl pd_threads[0];
  75 };
  76
  77 static int max_ptlrpcds;
  78 CFS_MODULE_PARM(max_ptlrpcds, "i", int, 0644,
  79                 "Max ptlrpcd thread count to be started.");
  80
  81 static int ptlrpcd_bind_policy = PDB_POLICY_PAIR;
  82 CFS_MODULE_PARM(ptlrpcd_bind_policy, "i", int, 0644,
  83                 "Ptlrpcd threads binding mode.");
  84 static struct ptlrpcd *ptlrpcds;
  85
  86 struct mutex ptlrpcd_mutex;
  87 static int ptlrpcd_users = 0;
  88
  89 void ptlrpcd_wake(struct ptlrpc_request *req)
  90 {
  91         struct ptlrpc_request_set *rq_set = req->rq_set;
  92
  93         LASSERT(rq_set != NULL);
  94
  95         wake_up(&rq_set->set_waitq);
  96 }
  97 EXPORT_SYMBOL(ptlrpcd_wake);
  98
  99 static struct ptlrpcd_ctl *
 100 ptlrpcd_select_pc(struct ptlrpc_request *req, pdl_policy_t policy, int index)
 101 {
 102         int idx = 0;
 103
 104         if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
 105                 return &ptlrpcds->pd_thread_rcv;
 106
 107         switch (policy) {
 108         case PDL_POLICY_SAME:
 109                 idx = smp_processor_id() % ptlrpcds->pd_nthreads;
 110                 break;
 111         case PDL_POLICY_LOCAL:
 112                 /* Before CPU partition patches available, process it the same
 113                  * as "PDL_POLICY_ROUND". */
 114 # ifdef CFS_CPU_MODE_NUMA
 115 # warning "fix this code to use new CPU partition APIs"
 116 # endif
 117                 /* Fall through to PDL_POLICY_ROUND until the CPU
 118                  * CPU partition patches are available. */
 119                 index = -1;
 120         case PDL_POLICY_PREFERRED:
 121                 if (index >= 0 && index < num_online_cpus()) {
 122                         idx = index % ptlrpcds->pd_nthreads;
 123                         break;
 124                 }
 125                 /* Fall through to PDL_POLICY_ROUND for bad index. */
 126         default:
 127                 /* Fall through to PDL_POLICY_ROUND for unknown policy. */
 128         case PDL_POLICY_ROUND:
 129                 /* We do not care whether it is strict load balance. */
 130                 idx = ptlrpcds->pd_index + 1;
 131                 if (idx == smp_processor_id())
 132                         idx++;
 133                 idx %= ptlrpcds->pd_nthreads;
 134                 ptlrpcds->pd_index = idx;
 135                 break;
 136         }
 137
 138         return &ptlrpcds->pd_threads[idx];
 139 }
 140
 141 /**
 142  * Move all request from an existing request set to the ptlrpcd queue.
 143  * All requests from the set must be in phase RQ_PHASE_NEW.
 144  */
 145 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 146 {
 147         struct list_head *tmp, *pos;
 148         struct ptlrpcd_ctl *pc;
 149         struct ptlrpc_request_set *new;
 150         int count, i;
 151
 152         pc = ptlrpcd_select_pc(NULL, PDL_POLICY_LOCAL, -1);
 153         new = pc->pc_set;
 154
 155         list_for_each_safe(pos, tmp, &set->set_requests) {
 156                 struct ptlrpc_request *req =
 157                         list_entry(pos, struct ptlrpc_request,
 158                                    rq_set_chain);
 159
 160                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 161                 req->rq_set = new;
 162                 req->rq_queued_time = cfs_time_current();
 163         }
 164
 165         spin_lock(&new->set_new_req_lock);
 166         list_splice_init(&set->set_requests, &new->set_new_requests);
 167         i = atomic_read(&set->set_remaining);
 168         count = atomic_add_return(i, &new->set_new_count);
 169         atomic_set(&set->set_remaining, 0);
 170         spin_unlock(&new->set_new_req_lock);
 171         if (count == i) {
 172                 wake_up(&new->set_waitq);
 173
 174                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
 175                  *      guarantee the async RPC can be processed ASAP, we have
 176                  *      no other better choice. It maybe fixed in future. */
 177                 for (i = 0; i < pc->pc_npartners; i++)
 178                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
 179         }
 180 }
 181 EXPORT_SYMBOL(ptlrpcd_add_rqset);
 182
 183 /**
 184  * Return transferred RPCs count.
 185  */
 186 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
 187                                struct ptlrpc_request_set *src)
 188 {
 189         struct list_head *tmp, *pos;
 190         struct ptlrpc_request *req;
 191         int rc = 0;
 192
 193         spin_lock(&src->set_new_req_lock);
 194         if (likely(!list_empty(&src->set_new_requests))) {
 195                 list_for_each_safe(pos, tmp, &src->set_new_requests) {
 196                         req = list_entry(pos, struct ptlrpc_request,
 197                                          rq_set_chain);
 198                         req->rq_set = des;
 199                 }
 200                 list_splice_init(&src->set_new_requests,
 201                                  &des->set_requests);
 202                 rc = atomic_read(&src->set_new_count);
 203                 atomic_add(rc, &des->set_remaining);
 204                 atomic_set(&src->set_new_count, 0);
 205         }
 206         spin_unlock(&src->set_new_req_lock);
 207         return rc;
 208 }
 209
 210 /**
 211  * Requests that are added to the ptlrpcd queue are sent via
 212  * ptlrpcd_check->ptlrpc_check_set().
 213  */
 214 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx)
 215 {
 216         struct ptlrpcd_ctl *pc;
 217
 218         if (req->rq_reqmsg)
 219                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
 220
 221         spin_lock(&req->rq_lock);
 222         if (req->rq_invalid_rqset) {
 223                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
 224                                                      back_to_sleep, NULL);
 225
 226                 req->rq_invalid_rqset = 0;
 227                 spin_unlock(&req->rq_lock);
 228                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 229         } else if (req->rq_set) {
 230                 /* If we have a vaid "rq_set", just reuse it to avoid double
 231                  * linked. */
 232                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 233                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 234
 235                 /* ptlrpc_check_set will decrease the count */
 236                 atomic_inc(&req->rq_set->set_remaining);
 237                 spin_unlock(&req->rq_lock);
 238                 wake_up(&req->rq_set->set_waitq);
 239                 return;
 240         } else {
 241                 spin_unlock(&req->rq_lock);
 242         }
 243
 244         pc = ptlrpcd_select_pc(req, policy, idx);
 245
 246         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
 247                   req, pc->pc_name, pc->pc_index);
 248
 249         ptlrpc_set_add_new_req(pc, req);
 250 }
 251 EXPORT_SYMBOL(ptlrpcd_add_req);
 252
 253 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
 254 {
 255         atomic_inc(&set->set_refcount);
 256 }
 257
 258 /**
 259  * Check if there is more work to do on ptlrpcd set.
 260  * Returns 1 if yes.
 261  */
 262 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
 263 {
 264         struct list_head *tmp, *pos;
 265         struct ptlrpc_request *req;
 266         struct ptlrpc_request_set *set = pc->pc_set;
 267         int rc = 0;
 268         int rc2;
 269         ENTRY;
 270
 271         if (atomic_read(&set->set_new_count)) {
 272                 spin_lock(&set->set_new_req_lock);
 273                 if (likely(!list_empty(&set->set_new_requests))) {
 274                         list_splice_init(&set->set_new_requests,
 275                                              &set->set_requests);
 276                         atomic_add(atomic_read(&set->set_new_count),
 277                                    &set->set_remaining);
 278                         atomic_set(&set->set_new_count, 0);
 279                         /*
 280                          * Need to calculate its timeout.
 281                          */
 282                         rc = 1;
 283                 }
 284                 spin_unlock(&set->set_new_req_lock);
 285         }
 286
 287         /* We should call lu_env_refill() before handling new requests to make
 288          * sure that env key the requests depending on really exists.
 289          */
 290         rc2 = lu_env_refill(env);
 291         if (rc2 != 0) {
 292                 /*
 293                  * XXX This is very awkward situation, because
 294                  * execution can neither continue (request
 295                  * interpreters assume that env is set up), nor repeat
 296                  * the loop (as this potentially results in a tight
 297                  * loop of -ENOMEM's).
 298                  *
 299                  * Fortunately, refill only ever does something when
 300                  * new modules are loaded, i.e., early during boot up.
 301                  */
 302                 CERROR("Failure to refill session: %d\n", rc2);
 303                 RETURN(rc);
 304         }
 305
 306         if (atomic_read(&set->set_remaining))
 307                 rc |= ptlrpc_check_set(env, set);
 308
 309         /* NB: ptlrpc_check_set has already moved complted request at the
 310          * head of seq::set_requests */
 311         list_for_each_safe(pos, tmp, &set->set_requests) {
 312                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 313                 if (req->rq_phase != RQ_PHASE_COMPLETE)
 314                         break;
 315
 316                 list_del_init(&req->rq_set_chain);
 317                 req->rq_set = NULL;
 318                 ptlrpc_req_finished(req);
 319         }
 320
 321         if (rc == 0) {
 322                 /*
 323                  * If new requests have been added, make sure to wake up.
 324                  */
 325                 rc = atomic_read(&set->set_new_count);
 326
 327                 /* If we have nothing to do, check whether we can take some
 328                  * work from our partner threads. */
 329                 if (rc == 0 && pc->pc_npartners > 0) {
 330                         struct ptlrpcd_ctl *partner;
 331                         struct ptlrpc_request_set *ps;
 332                         int first = pc->pc_cursor;
 333
 334                         do {
 335                                 partner = pc->pc_partners[pc->pc_cursor++];
 336                                 if (pc->pc_cursor >= pc->pc_npartners)
 337                                         pc->pc_cursor = 0;
 338                                 if (partner == NULL)
 339                                         continue;
 340
 341                                 spin_lock(&partner->pc_lock);
 342                                 ps = partner->pc_set;
 343                                 if (ps == NULL) {
 344                                         spin_unlock(&partner->pc_lock);
 345                                         continue;
 346                                 }
 347
 348                                 ptlrpc_reqset_get(ps);
 349                                 spin_unlock(&partner->pc_lock);
 350
 351                                 if (atomic_read(&ps->set_new_count)) {
 352                                         rc = ptlrpcd_steal_rqset(set, ps);
 353                                         if (rc > 0)
 354                                                 CDEBUG(D_RPCTRACE, "transfer %d"
 355                                                        " async RPCs [%d->%d]\n",
 356                                                        rc, partner->pc_index,
 357                                                        pc->pc_index);
 358                                 }
 359                                 ptlrpc_reqset_put(ps);
 360                         } while (rc == 0 && pc->pc_cursor != first);
 361                 }
 362         }
 363
 364         RETURN(rc);
 365 }
 366
 367 /**
 368  * Main ptlrpcd thread.
 369  * ptlrpc's code paths like to execute in process context, so we have this
 370  * thread which spins on a set which contains the rpcs and sends them.
 371  *
 372  */
 373 static int ptlrpcd(void *arg)
 374 {
 375         struct ptlrpcd_ctl *pc = arg;
 376         struct ptlrpc_request_set *set = pc->pc_set;
 377         struct lu_context ses = { 0 };
 378         struct lu_env env = { .le_ses = &ses };
 379         int rc, exit = 0;
 380         ENTRY;
 381
 382         unshare_fs_struct();
 383 #if defined(CONFIG_SMP)
 384         if (test_bit(LIOD_BIND, &pc->pc_flags)) {
 385                 int index = pc->pc_index;
 386
 387                 if (index >= 0 && index < num_possible_cpus()) {
 388                         while (!cpu_online(index)) {
 389                                 if (++index >= num_possible_cpus())
 390                                         index = 0;
 391                         }
 392                         set_cpus_allowed_ptr(current,
 393                                      cpumask_of_node(cpu_to_node(index)));
 394                 }
 395         }
 396 #endif
 397         /* Both client and server (MDT/OST) may use the environment. */
 398         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
 399                                           LCT_CL_THREAD | LCT_REMEMBER |
 400                                           LCT_NOREF);
 401         if (rc == 0) {
 402                 rc = lu_context_init(env.le_ses,
 403                                      LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
 404                 if (rc != 0)
 405                         lu_context_fini(&env.le_ctx);
 406         }
 407         complete(&pc->pc_starting);
 408
 409         if (rc != 0)
 410                 RETURN(rc);
 411
 412         /*
 413          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 414          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 415          * there are requests in the set. New requests come in on the set's
 416          * new_req_list and ptlrpcd_check() moves them into the set.
 417          */
 418         do {
 419                 struct l_wait_info lwi;
 420                 int timeout;
 421
 422                 timeout = ptlrpc_set_next_timeout(set);
 423                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 424                                   ptlrpc_expired_set, set);
 425
 426                 lu_context_enter(&env.le_ctx);
 427                 lu_context_enter(env.le_ses);
 428                 l_wait_event(set->set_waitq, ptlrpcd_check(&env, pc), &lwi);
 429                 lu_context_exit(&env.le_ctx);
 430                 lu_context_exit(env.le_ses);
 431
 432                 /*
 433                  * Abort inflight rpcs for forced stop case.
 434                  */
 435                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 436                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 437                                 ptlrpc_abort_set(set);
 438                         exit++;
 439                 }
 440
 441                 /*
 442                  * Let's make one more loop to make sure that ptlrpcd_check()
 443                  * copied all raced new rpcs into the set so we can kill them.
 444                  */
 445         } while (exit < 2);
 446
 447         /*
 448          * Wait for inflight requests to drain.
 449          */
 450         if (!list_empty(&set->set_requests))
 451                 ptlrpc_set_wait(set);
 452         lu_context_fini(&env.le_ctx);
 453         lu_context_fini(env.le_ses);
 454
 455         complete(&pc->pc_finishing);
 456
 457         return 0;
 458 }
 459
 460 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 461  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 462  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 463  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 464  *      because of some CPU core(s) busy with other loads.
 465  *
 466  *      For example: "ls -l", some async RPCs for statahead are assigned to
 467  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 468  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 469  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 470  *      such case, the statahead async RPCs can not be processed in time, it is
 471  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 472  *      be better. But it breaks former data transfer policy.
 473  *
 474  *      So we shouldn't be blind for avoiding the data transfer. We make some
 475  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 476  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 477  *      core. The other part is for free mode, all the ptlrpcd threads in the
 478  *      part can be scheduled on any CPU core. We specify some partnership
 479  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 480  *      and the async RPC load within the partners are shared.
 481  *
 482  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 483  *      thread can be scheduled in time), and try to guarantee the async RPC
 484  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 485  *      on any CPU core).
 486  *
 487  *      As for how to specify the partnership between bound mode ptlrpcd
 488  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 489  *      <free bound> pair. In future, we can specify some more complex
 490  *      partnership based on the patches for CPU partition. But before such
 491  *      patches are available, we prefer to use the simplest one.
 492  */
 493 # ifdef CFS_CPU_MODE_NUMA
 494 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 495 # endif
 496 static int ptlrpcd_bind(int index, int max)
 497 {
 498         struct ptlrpcd_ctl *pc;
 499         int rc = 0;
 500 #if defined(CONFIG_NUMA)
 501         cpumask_t mask;
 502 #endif
 503         ENTRY;
 504
 505         LASSERT(index <= max - 1);
 506         pc = &ptlrpcds->pd_threads[index];
 507         switch (ptlrpcd_bind_policy) {
 508         case PDB_POLICY_NONE:
 509                 pc->pc_npartners = -1;
 510                 break;
 511         case PDB_POLICY_FULL:
 512                 pc->pc_npartners = 0;
 513                 set_bit(LIOD_BIND, &pc->pc_flags);
 514                 break;
 515         case PDB_POLICY_PAIR:
 516                 LASSERT(max % 2 == 0);
 517                 pc->pc_npartners = 1;
 518                 break;
 519         case PDB_POLICY_NEIGHBOR:
 520 #if defined(CONFIG_NUMA)
 521         {
 522                 int i;
 523                 mask = *cpumask_of_node(cpu_to_node(index));
 524                 for (i = max; i < num_online_cpus(); i++)
 525                         cpu_clear(i, mask);
 526                 pc->pc_npartners = cpus_weight(mask) - 1;
 527                 set_bit(LIOD_BIND, &pc->pc_flags);
 528         }
 529 #else
 530                 LASSERT(max >= 3);
 531                 pc->pc_npartners = 2;
 532 #endif
 533                 break;
 534         default:
 535                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 536                 rc = -EINVAL;
 537         }
 538
 539         if (rc == 0 && pc->pc_npartners > 0) {
 540                 OBD_ALLOC(pc->pc_partners,
 541                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 542                 if (pc->pc_partners == NULL) {
 543                         pc->pc_npartners = 0;
 544                         rc = -ENOMEM;
 545                 } else {
 546                         switch (ptlrpcd_bind_policy) {
 547                         case PDB_POLICY_PAIR:
 548                                 if (index & 0x1) {
 549                                         set_bit(LIOD_BIND, &pc->pc_flags);
 550                                         pc->pc_partners[0] = &ptlrpcds->
 551                                                 pd_threads[index - 1];
 552                                         ptlrpcds->pd_threads[index - 1].
 553                                                 pc_partners[0] = pc;
 554                                 }
 555                                 break;
 556                         case PDB_POLICY_NEIGHBOR:
 557 #if defined(CONFIG_NUMA)
 558                         {
 559                                 struct ptlrpcd_ctl *ppc;
 560                                 int i, pidx;
 561                                 /* partners are cores in the same NUMA node.
 562                                  * setup partnership only with ptlrpcd threads
 563                                  * that are already initialized
 564                                  */
 565                                 for (pidx = 0, i = 0; i < index; i++) {
 566                                         if (cpu_isset(i, mask)) {
 567                                                 ppc = &ptlrpcds->pd_threads[i];
 568                                                 pc->pc_partners[pidx++] = ppc;
 569                                                 ppc->pc_partners[ppc->
 570                                                           pc_npartners++] = pc;
 571                                         }
 572                                 }
 573                                 /* adjust number of partners to the number
 574                                  * of partnership really setup */
 575                                 pc->pc_npartners = pidx;
 576                         }
 577 #else
 578                                 if (index & 0x1)
 579                                         set_bit(LIOD_BIND, &pc->pc_flags);
 580                                 if (index > 0) {
 581                                         pc->pc_partners[0] = &ptlrpcds->
 582                                                 pd_threads[index - 1];
 583                                         ptlrpcds->pd_threads[index - 1].
 584                                                 pc_partners[1] = pc;
 585                                         if (index == max - 1) {
 586                                                 pc->pc_partners[1] =
 587                                                 &ptlrpcds->pd_threads[0];
 588                                                 ptlrpcds->pd_threads[0].
 589                                                 pc_partners[0] = pc;
 590                                         }
 591                                 }
 592 #endif
 593                                 break;
 594                         }
 595                 }
 596         }
 597
 598         RETURN(rc);
 599 }
 600
 601
 602 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 603 {
 604         int rc;
 605         ENTRY;
 606
 607         /*
 608          * Do not allow start second thread for one pc.
 609          */
 610         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 611                 CWARN("Starting second thread (%s) for same pc %p\n",
 612                       name, pc);
 613                 RETURN(0);
 614         }
 615
 616         pc->pc_index = index;
 617         init_completion(&pc->pc_starting);
 618         init_completion(&pc->pc_finishing);
 619         spin_lock_init(&pc->pc_lock);
 620         strlcpy(pc->pc_name, name, sizeof(pc->pc_name));
 621         pc->pc_set = ptlrpc_prep_set();
 622         if (pc->pc_set == NULL)
 623                 GOTO(out, rc = -ENOMEM);
 624
 625         /*
 626          * So far only "client" ptlrpcd uses an environment. In the future,
 627          * ptlrpcd thread (or a thread-set) has to be given an argument,
 628          * describing its "scope".
 629          */
 630         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 631         if (rc != 0)
 632                 GOTO(out_set, rc);
 633
 634         {
 635                 struct task_struct *task;
 636                 if (index >= 0) {
 637                         rc = ptlrpcd_bind(index, max);
 638                         if (rc < 0)
 639                                 GOTO(out_env, rc);
 640                 }
 641
 642                 task = kthread_run(ptlrpcd, pc, pc->pc_name);
 643                 if (IS_ERR(task))
 644                         GOTO(out_env, rc = PTR_ERR(task));
 645
 646                 wait_for_completion(&pc->pc_starting);
 647         }
 648         RETURN(0);
 649
 650 out_env:
 651         lu_context_fini(&pc->pc_env.le_ctx);
 652
 653 out_set:
 654         if (pc->pc_set != NULL) {
 655                 struct ptlrpc_request_set *set = pc->pc_set;
 656
 657                 spin_lock(&pc->pc_lock);
 658                 pc->pc_set = NULL;
 659                 spin_unlock(&pc->pc_lock);
 660                 ptlrpc_set_destroy(set);
 661         }
 662         clear_bit(LIOD_BIND, &pc->pc_flags);
 663 out:
 664         clear_bit(LIOD_START, &pc->pc_flags);
 665         RETURN(rc);
 666 }
 667
 668 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 669 {
 670         ENTRY;
 671
 672         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 673                 CWARN("Thread for pc %p was not started\n", pc);
 674                 goto out;
 675         }
 676
 677         set_bit(LIOD_STOP, &pc->pc_flags);
 678         if (force)
 679                 set_bit(LIOD_FORCE, &pc->pc_flags);
 680         wake_up(&pc->pc_set->set_waitq);
 681
 682 out:
 683         EXIT;
 684 }
 685
 686 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 687 {
 688         struct ptlrpc_request_set *set = pc->pc_set;
 689         ENTRY;
 690
 691         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 692                 CWARN("Thread for pc %p was not started\n", pc);
 693                 goto out;
 694         }
 695
 696         wait_for_completion(&pc->pc_finishing);
 697         lu_context_fini(&pc->pc_env.le_ctx);
 698
 699         spin_lock(&pc->pc_lock);
 700         pc->pc_set = NULL;
 701         spin_unlock(&pc->pc_lock);
 702         ptlrpc_set_destroy(set);
 703
 704         clear_bit(LIOD_START, &pc->pc_flags);
 705         clear_bit(LIOD_STOP, &pc->pc_flags);
 706         clear_bit(LIOD_FORCE, &pc->pc_flags);
 707         clear_bit(LIOD_BIND, &pc->pc_flags);
 708
 709 out:
 710         if (pc->pc_npartners > 0) {
 711                 LASSERT(pc->pc_partners != NULL);
 712
 713                 OBD_FREE(pc->pc_partners,
 714                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 715                 pc->pc_partners = NULL;
 716         }
 717         pc->pc_npartners = 0;
 718         EXIT;
 719 }
 720
 721 static void ptlrpcd_fini(void)
 722 {
 723         int i;
 724         ENTRY;
 725
 726         if (ptlrpcds != NULL) {
 727                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 728                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 729                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 730                         ptlrpcd_free(&ptlrpcds->pd_threads[i]);
 731                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 732                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 733                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 734                 ptlrpcds = NULL;
 735         }
 736
 737         EXIT;
 738 }
 739
 740 static int ptlrpcd_init(void)
 741 {
 742         int     nthreads = num_online_cpus();
 743         char    name[16];
 744         int     size, i = -1, j, rc = 0;
 745         ENTRY;
 746
 747         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 748                 nthreads = max_ptlrpcds;
 749         if (nthreads < 2)
 750                 nthreads = 2;
 751         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 752                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 753         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 754                 nthreads &= ~1; /* make sure it is even */
 755
 756         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 757         OBD_ALLOC(ptlrpcds, size);
 758         if (ptlrpcds == NULL)
 759                 GOTO(out, rc = -ENOMEM);
 760
 761         snprintf(name, 15, "ptlrpcd_rcv");
 762         set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 763         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 764         if (rc < 0)
 765                 GOTO(out, rc);
 766
 767         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 768          *      non-recovery async RPC to improve overall async RPC efficiency.
 769          *
 770          *      But there are some issues with async I/O RPCs and async non-I/O
 771          *      RPCs processed in the same set under some cases. The ptlrpcd may
 772          *      be blocked by some async I/O RPC(s), then will cause other async
 773          *      non-I/O RPC(s) can not be processed in time.
 774          *
 775          *      Maybe we should distinguish blocked async RPCs from non-blocked
 776          *      async RPCs, and process them in different ptlrpcd sets to avoid
 777          *      unnecessary dependency. But how to distribute async RPCs load
 778          *      among all the ptlrpc daemons becomes another trouble. */
 779         for (i = 0; i < nthreads; i++) {
 780                 snprintf(name, 15, "ptlrpcd_%d", i);
 781                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 782                 if (rc < 0)
 783                         GOTO(out, rc);
 784         }
 785
 786         ptlrpcds->pd_size = size;
 787         ptlrpcds->pd_index = 0;
 788         ptlrpcds->pd_nthreads = nthreads;
 789
 790 out:
 791         if (rc != 0 && ptlrpcds != NULL) {
 792                 for (j = 0; j <= i; j++)
 793                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 794                 for (j = 0; j <= i; j++)
 795                         ptlrpcd_free(&ptlrpcds->pd_threads[j]);
 796                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 797                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 798                 OBD_FREE(ptlrpcds, size);
 799                 ptlrpcds = NULL;
 800         }
 801
 802         RETURN(rc);
 803 }
 804
 805 int ptlrpcd_addref(void)
 806 {
 807         int rc = 0;
 808         ENTRY;
 809
 810         mutex_lock(&ptlrpcd_mutex);
 811         if (++ptlrpcd_users == 1) {
 812                 rc = ptlrpcd_init();
 813                 if (rc < 0)
 814                         ptlrpcd_users--;
 815         }
 816         mutex_unlock(&ptlrpcd_mutex);
 817         RETURN(rc);
 818 }
 819 EXPORT_SYMBOL(ptlrpcd_addref);
 820
 821 void ptlrpcd_decref(void)
 822 {
 823         mutex_lock(&ptlrpcd_mutex);
 824         if (--ptlrpcd_users == 0)
 825                 ptlrpcd_fini();
 826         mutex_unlock(&ptlrpcd_mutex);
 827 }
 828 EXPORT_SYMBOL(ptlrpcd_decref);
 829 /** @} ptlrpcd */