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         cfs_waitq_signal(&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 = cfs_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 < cfs_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 == cfs_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                 cfs_waitq_signal(&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                         cfs_waitq_signal(&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                 cfs_waitq_signal(&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_env env = { .le_ses = NULL };
 410         int rc, exit = 0;
 411         ENTRY;
 412
 413         unshare_fs_struct();
 414 #if defined(CONFIG_SMP) && \
 415 (defined(HAVE_CPUMASK_OF_NODE) || defined(HAVE_NODE_TO_CPUMASK))
 416         if (test_bit(LIOD_BIND, &pc->pc_flags)) {
 417                 int index = pc->pc_index;
 418
 419                 if (index >= 0 && index < cfs_num_possible_cpus()) {
 420                         while (!cpu_online(index)) {
 421                                 if (++index >= cfs_num_possible_cpus())
 422                                         index = 0;
 423                         }
 424                         cfs_set_cpus_allowed(cfs_current(),
 425                                      *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         complete(&pc->pc_starting);
 437
 438         if (rc != 0)
 439                 RETURN(rc);
 440
 441         /*
 442          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 443          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 444          * there are requests in the set. New requests come in on the set's
 445          * new_req_list and ptlrpcd_check() moves them into the set.
 446          */
 447         do {
 448                 struct l_wait_info lwi;
 449                 int timeout;
 450
 451                 timeout = ptlrpc_set_next_timeout(set);
 452                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 453                                   ptlrpc_expired_set, set);
 454
 455                 lu_context_enter(&env.le_ctx);
 456                 l_wait_event(set->set_waitq,
 457                              ptlrpcd_check(&env, pc), &lwi);
 458                 lu_context_exit(&env.le_ctx);
 459
 460                 /*
 461                  * Abort inflight rpcs for forced stop case.
 462                  */
 463                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 464                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 465                                 ptlrpc_abort_set(set);
 466                         exit++;
 467                 }
 468
 469                 /*
 470                  * Let's make one more loop to make sure that ptlrpcd_check()
 471                  * copied all raced new rpcs into the set so we can kill them.
 472                  */
 473         } while (exit < 2);
 474
 475         /*
 476          * Wait for inflight requests to drain.
 477          */
 478         if (!cfs_list_empty(&set->set_requests))
 479                 ptlrpc_set_wait(set);
 480         lu_context_fini(&env.le_ctx);
 481
 482         complete(&pc->pc_finishing);
 483
 484         return 0;
 485 }
 486
 487 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
 488  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
 489  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
 490  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
 491  *      because of some CPU core(s) busy with other loads.
 492  *
 493  *      For example: "ls -l", some async RPCs for statahead are assigned to
 494  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
 495  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
 496  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
 497  *      such case, the statahead async RPCs can not be processed in time, it is
 498  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
 499  *      be better. But it breaks former data transfer policy.
 500  *
 501  *      So we shouldn't be blind for avoiding the data transfer. We make some
 502  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
 503  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
 504  *      core. The other part is for free mode, all the ptlrpcd threads in the
 505  *      part can be scheduled on any CPU core. We specify some partnership
 506  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
 507  *      and the async RPC load within the partners are shared.
 508  *
 509  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
 510  *      thread can be scheduled in time), and try to guarantee the async RPC
 511  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
 512  *      on any CPU core).
 513  *
 514  *      As for how to specify the partnership between bound mode ptlrpcd
 515  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
 516  *      <free bound> pair. In future, we can specify some more complex
 517  *      partnership based on the patches for CPU partition. But before such
 518  *      patches are available, we prefer to use the simplest one.
 519  */
 520 # ifdef CFS_CPU_MODE_NUMA
 521 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
 522 # endif
 523 static int ptlrpcd_bind(int index, int max)
 524 {
 525         struct ptlrpcd_ctl *pc;
 526         int rc = 0;
 527 #if defined(CONFIG_NUMA) && \
 528 (defined(HAVE_CPUMASK_OF_NODE) || defined(HAVE_NODE_TO_CPUMASK))
 529         cpumask_t mask;
 530 #endif
 531         ENTRY;
 532
 533         LASSERT(index <= max - 1);
 534         pc = &ptlrpcds->pd_threads[index];
 535         switch (ptlrpcd_bind_policy) {
 536         case PDB_POLICY_NONE:
 537                 pc->pc_npartners = -1;
 538                 break;
 539         case PDB_POLICY_FULL:
 540                 pc->pc_npartners = 0;
 541                 set_bit(LIOD_BIND, &pc->pc_flags);
 542                 break;
 543         case PDB_POLICY_PAIR:
 544                 LASSERT(max % 2 == 0);
 545                 pc->pc_npartners = 1;
 546                 break;
 547         case PDB_POLICY_NEIGHBOR:
 548 #if defined(CONFIG_NUMA) && \
 549 (defined(HAVE_CPUMASK_OF_NODE) || defined(HAVE_NODE_TO_CPUMASK))
 550         {
 551                 int i;
 552                 mask = *cpumask_of_node(cpu_to_node(index));
 553                 for (i = max; i < cfs_num_online_cpus(); i++)
 554                         cpu_clear(i, mask);
 555                 pc->pc_npartners = cpus_weight(mask) - 1;
 556                 set_bit(LIOD_BIND, &pc->pc_flags);
 557         }
 558 #else
 559                 LASSERT(max >= 3);
 560                 pc->pc_npartners = 2;
 561 #endif
 562                 break;
 563         default:
 564                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
 565                 rc = -EINVAL;
 566         }
 567
 568         if (rc == 0 && pc->pc_npartners > 0) {
 569                 OBD_ALLOC(pc->pc_partners,
 570                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 571                 if (pc->pc_partners == NULL) {
 572                         pc->pc_npartners = 0;
 573                         rc = -ENOMEM;
 574                 } else {
 575                         switch (ptlrpcd_bind_policy) {
 576                         case PDB_POLICY_PAIR:
 577                                 if (index & 0x1) {
 578                                         set_bit(LIOD_BIND, &pc->pc_flags);
 579                                         pc->pc_partners[0] = &ptlrpcds->
 580                                                 pd_threads[index - 1];
 581                                         ptlrpcds->pd_threads[index - 1].
 582                                                 pc_partners[0] = pc;
 583                                 }
 584                                 break;
 585                         case PDB_POLICY_NEIGHBOR:
 586 #if defined(CONFIG_NUMA) && \
 587 (defined(HAVE_CPUMASK_OF_NODE) || defined(HAVE_NODE_TO_CPUMASK))
 588                         {
 589                                 struct ptlrpcd_ctl *ppc;
 590                                 int i, pidx;
 591                                 /* partners are cores in the same NUMA node.
 592                                  * setup partnership only with ptlrpcd threads
 593                                  * that are already initialized
 594                                  */
 595                                 for (pidx = 0, i = 0; i < index; i++) {
 596                                         if (cpu_isset(i, mask)) {
 597                                                 ppc = &ptlrpcds->pd_threads[i];
 598                                                 pc->pc_partners[pidx++] = ppc;
 599                                                 ppc->pc_partners[ppc->
 600                                                           pc_npartners++] = pc;
 601                                         }
 602                                 }
 603                                 /* adjust number of partners to the number
 604                                  * of partnership really setup */
 605                                 pc->pc_npartners = pidx;
 606                         }
 607 #else
 608                                 if (index & 0x1)
 609                                         set_bit(LIOD_BIND, &pc->pc_flags);
 610                                 if (index > 0) {
 611                                         pc->pc_partners[0] = &ptlrpcds->
 612                                                 pd_threads[index - 1];
 613                                         ptlrpcds->pd_threads[index - 1].
 614                                                 pc_partners[1] = pc;
 615                                         if (index == max - 1) {
 616                                                 pc->pc_partners[1] =
 617                                                 &ptlrpcds->pd_threads[0];
 618                                                 ptlrpcds->pd_threads[0].
 619                                                 pc_partners[0] = pc;
 620                                         }
 621                                 }
 622 #endif
 623                                 break;
 624                         }
 625                 }
 626         }
 627
 628         RETURN(rc);
 629 }
 630
 631 #else /* !__KERNEL__ */
 632
 633 /**
 634  * In liblustre we do not have separate threads, so this function
 635  * is called from time to time all across common code to see
 636  * if something needs to be processed on ptlrpcd set.
 637  */
 638 int ptlrpcd_check_async_rpcs(void *arg)
 639 {
 640         struct ptlrpcd_ctl *pc = arg;
 641         int                 rc = 0;
 642
 643         /*
 644          * Single threaded!!
 645          */
 646         pc->pc_recurred++;
 647
 648         if (pc->pc_recurred == 1) {
 649                 rc = lu_env_refill(&pc->pc_env);
 650                 if (rc == 0) {
 651                         lu_context_enter(&pc->pc_env.le_ctx);
 652                         rc = ptlrpcd_check(&pc->pc_env, pc);
 653                         if (!rc)
 654                                 ptlrpc_expired_set(pc->pc_set);
 655                         /*
 656                          * XXX: send replay requests.
 657                          */
 658                         if (test_bit(LIOD_RECOVERY, &pc->pc_flags))
 659                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 660                         lu_context_exit(&pc->pc_env.le_ctx);
 661                 }
 662         }
 663
 664         pc->pc_recurred--;
 665         return rc;
 666 }
 667
 668 int ptlrpcd_idle(void *arg)
 669 {
 670         struct ptlrpcd_ctl *pc = arg;
 671
 672         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
 673                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 674 }
 675
 676 #endif
 677
 678 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
 679 {
 680         int rc;
 681         int env = 0;
 682         ENTRY;
 683
 684         /*
 685          * Do not allow start second thread for one pc.
 686          */
 687         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 688                 CWARN("Starting second thread (%s) for same pc %p\n",
 689                       name, pc);
 690                 RETURN(0);
 691         }
 692
 693         pc->pc_index = index;
 694         init_completion(&pc->pc_starting);
 695         init_completion(&pc->pc_finishing);
 696         spin_lock_init(&pc->pc_lock);
 697         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 698         pc->pc_set = ptlrpc_prep_set();
 699         if (pc->pc_set == NULL)
 700                 GOTO(out, rc = -ENOMEM);
 701         /*
 702          * So far only "client" ptlrpcd uses an environment. In the future,
 703          * ptlrpcd thread (or a thread-set) has to be given an argument,
 704          * describing its "scope".
 705          */
 706         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 707         if (rc != 0)
 708                 GOTO(out, rc);
 709
 710         env = 1;
 711 #ifdef __KERNEL__
 712         {
 713                 cfs_task_t *task;
 714                 if (index >= 0) {
 715                         rc = ptlrpcd_bind(index, max);
 716                         if (rc < 0)
 717                                 GOTO(out, rc);
 718                 }
 719
 720                 task = kthread_run(ptlrpcd, pc, pc->pc_name);
 721                 if (IS_ERR(task))
 722                         GOTO(out, rc = PTR_ERR(task));
 723
 724                 rc = 0;
 725                 wait_for_completion(&pc->pc_starting);
 726         }
 727 #else
 728         pc->pc_wait_callback =
 729                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 730                                                  &ptlrpcd_check_async_rpcs, pc);
 731         pc->pc_idle_callback =
 732                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 733                                                  &ptlrpcd_idle, pc);
 734 #endif
 735 out:
 736         if (rc) {
 737 #ifdef __KERNEL__
 738                 if (pc->pc_set != NULL) {
 739                         struct ptlrpc_request_set *set = pc->pc_set;
 740
 741                         spin_lock(&pc->pc_lock);
 742                         pc->pc_set = NULL;
 743                         spin_unlock(&pc->pc_lock);
 744                         ptlrpc_set_destroy(set);
 745                 }
 746                 if (env != 0)
 747                         lu_context_fini(&pc->pc_env.le_ctx);
 748                 clear_bit(LIOD_BIND, &pc->pc_flags);
 749 #else
 750                 SET_BUT_UNUSED(env);
 751 #endif
 752                 clear_bit(LIOD_START, &pc->pc_flags);
 753         }
 754         RETURN(rc);
 755 }
 756
 757 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 758 {
 759         ENTRY;
 760
 761         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 762                 CWARN("Thread for pc %p was not started\n", pc);
 763                 goto out;
 764         }
 765
 766         set_bit(LIOD_STOP, &pc->pc_flags);
 767         if (force)
 768                 set_bit(LIOD_FORCE, &pc->pc_flags);
 769         cfs_waitq_signal(&pc->pc_set->set_waitq);
 770
 771 out:
 772         EXIT;
 773 }
 774
 775 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
 776 {
 777         struct ptlrpc_request_set *set = pc->pc_set;
 778         ENTRY;
 779
 780         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 781                 CWARN("Thread for pc %p was not started\n", pc);
 782                 goto out;
 783         }
 784
 785 #ifdef __KERNEL__
 786         wait_for_completion(&pc->pc_finishing);
 787 #else
 788         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 789         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 790 #endif
 791         lu_context_fini(&pc->pc_env.le_ctx);
 792
 793         spin_lock(&pc->pc_lock);
 794         pc->pc_set = NULL;
 795         spin_unlock(&pc->pc_lock);
 796         ptlrpc_set_destroy(set);
 797
 798         clear_bit(LIOD_START, &pc->pc_flags);
 799         clear_bit(LIOD_STOP, &pc->pc_flags);
 800         clear_bit(LIOD_FORCE, &pc->pc_flags);
 801         clear_bit(LIOD_BIND, &pc->pc_flags);
 802
 803 out:
 804 #ifdef __KERNEL__
 805         if (pc->pc_npartners > 0) {
 806                 LASSERT(pc->pc_partners != NULL);
 807
 808                 OBD_FREE(pc->pc_partners,
 809                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
 810                 pc->pc_partners = NULL;
 811         }
 812         pc->pc_npartners = 0;
 813 #endif
 814         EXIT;
 815 }
 816
 817 static void ptlrpcd_fini(void)
 818 {
 819         int i;
 820         ENTRY;
 821
 822         if (ptlrpcds != NULL) {
 823                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 824                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
 825                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
 826                         ptlrpcd_free(&ptlrpcds->pd_threads[i]);
 827                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 828                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 829                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
 830                 ptlrpcds = NULL;
 831         }
 832
 833         EXIT;
 834 }
 835
 836 static int ptlrpcd_init(void)
 837 {
 838         int nthreads = cfs_num_online_cpus();
 839         char name[16];
 840         int size, i = -1, j, rc = 0;
 841         ENTRY;
 842
 843 #ifdef __KERNEL__
 844         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
 845                 nthreads = max_ptlrpcds;
 846         if (nthreads < 2)
 847                 nthreads = 2;
 848         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
 849                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
 850         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
 851                 nthreads &= ~1; /* make sure it is even */
 852 #else
 853         nthreads = 1;
 854 #endif
 855
 856         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
 857         OBD_ALLOC(ptlrpcds, size);
 858         if (ptlrpcds == NULL)
 859                 GOTO(out, rc = -ENOMEM);
 860
 861         snprintf(name, 15, "ptlrpcd_rcv");
 862         set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
 863         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
 864         if (rc < 0)
 865                 GOTO(out, rc);
 866
 867         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
 868          *      non-recovery async RPC to improve overall async RPC efficiency.
 869          *
 870          *      But there are some issues with async I/O RPCs and async non-I/O
 871          *      RPCs processed in the same set under some cases. The ptlrpcd may
 872          *      be blocked by some async I/O RPC(s), then will cause other async
 873          *      non-I/O RPC(s) can not be processed in time.
 874          *
 875          *      Maybe we should distinguish blocked async RPCs from non-blocked
 876          *      async RPCs, and process them in different ptlrpcd sets to avoid
 877          *      unnecessary dependency. But how to distribute async RPCs load
 878          *      among all the ptlrpc daemons becomes another trouble. */
 879         for (i = 0; i < nthreads; i++) {
 880                 snprintf(name, 15, "ptlrpcd_%d", i);
 881                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
 882                 if (rc < 0)
 883                         GOTO(out, rc);
 884         }
 885
 886         ptlrpcds->pd_size = size;
 887         ptlrpcds->pd_index = 0;
 888         ptlrpcds->pd_nthreads = nthreads;
 889
 890 out:
 891         if (rc != 0 && ptlrpcds != NULL) {
 892                 for (j = 0; j <= i; j++)
 893                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
 894                 for (j = 0; j <= i; j++)
 895                         ptlrpcd_free(&ptlrpcds->pd_threads[j]);
 896                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
 897                 ptlrpcd_free(&ptlrpcds->pd_thread_rcv);
 898                 OBD_FREE(ptlrpcds, size);
 899                 ptlrpcds = NULL;
 900         }
 901
 902         RETURN(0);
 903 }
 904
 905 int ptlrpcd_addref(void)
 906 {
 907         int rc = 0;
 908         ENTRY;
 909
 910         mutex_lock(&ptlrpcd_mutex);
 911         if (++ptlrpcd_users == 1)
 912                 rc = ptlrpcd_init();
 913         mutex_unlock(&ptlrpcd_mutex);
 914         RETURN(rc);
 915 }
 916 EXPORT_SYMBOL(ptlrpcd_addref);
 917
 918 void ptlrpcd_decref(void)
 919 {
 920         mutex_lock(&ptlrpcd_mutex);
 921         if (--ptlrpcd_users == 0)
 922                 ptlrpcd_fini();
 923         mutex_unlock(&ptlrpcd_mutex);
 924 }
 925 EXPORT_SYMBOL(ptlrpcd_decref);
 926 /** @} ptlrpcd */