Whamcloud - gitweb
LU-1346 libcfs: replace libcfs wrappers with kernel API
[fs/lustre-release.git] / lustre / ptlrpc / ptlrpcd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/ptlrpc/ptlrpcd.c
37  */
38
39 /** \defgroup ptlrpcd PortalRPC daemon
40  *
41  * ptlrpcd is a special thread with its own set where other user might add
42  * requests when they don't want to wait for their completion.
43  * PtlRPCD will take care of sending such requests and then processing their
44  * replies and calling completion callbacks as necessary.
45  * The callbacks are called directly from ptlrpcd context.
46  * It is important to never significantly block (esp. on RPCs!) within such
47  * completion handler or a deadlock might occur where ptlrpcd enters some
48  * callback that attempts to send another RPC and wait for it to return,
49  * during which time ptlrpcd is completely blocked, so e.g. if import
50  * fails, recovery cannot progress because connection requests are also
51  * sent by ptlrpcd.
52  *
53  * @{
54  */
55
56 #define DEBUG_SUBSYSTEM S_RPC
57
58 #ifdef __KERNEL__
59 # include <libcfs/libcfs.h>
60 #else /* __KERNEL__ */
61 # include <liblustre.h>
62 # include <ctype.h>
63 #endif
64
65 #include <lustre_net.h>
66 # include <lustre_lib.h>
67
68 #include <lustre_ha.h>
69 #include <obd_class.h>   /* for obd_zombie */
70 #include <obd_support.h> /* for OBD_FAIL_CHECK */
71 #include <cl_object.h> /* cl_env_{get,put}() */
72 #include <lprocfs_status.h>
73
74 #include "ptlrpc_internal.h"
75
76 struct ptlrpcd {
77         int                pd_size;
78         int                pd_index;
79         int                pd_nthreads;
80         struct ptlrpcd_ctl pd_thread_rcv;
81         struct ptlrpcd_ctl pd_threads[0];
82 };
83
84 #ifdef __KERNEL__
85 static int max_ptlrpcds;
86 CFS_MODULE_PARM(max_ptlrpcds, "i", int, 0644,
87                 "Max ptlrpcd thread count to be started.");
88
89 static int ptlrpcd_bind_policy = PDB_POLICY_PAIR;
90 CFS_MODULE_PARM(ptlrpcd_bind_policy, "i", int, 0644,
91                 "Ptlrpcd threads binding mode.");
92 #endif
93 static struct ptlrpcd *ptlrpcds;
94
95 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         cfs_daemonize_ctxt(pc->pc_name);
414 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
415         if (test_bit(LIOD_BIND, &pc->pc_flags)) {
416                 int index = pc->pc_index;
417
418                 if (index >= 0 && index < cfs_num_possible_cpus()) {
419                         while (!cpu_online(index)) {
420                                 if (++index >= cfs_num_possible_cpus())
421                                         index = 0;
422                         }
423                         cfs_set_cpus_allowed(cfs_current(),
424                                      node_to_cpumask(cpu_to_node(index)));
425                 }
426         }
427 #endif
428         /*
429          * XXX So far only "client" ptlrpcd uses an environment. In
430          * the future, ptlrpcd thread (or a thread-set) has to given
431          * an argument, describing its "scope".
432          */
433         rc = lu_context_init(&env.le_ctx,
434                              LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
435         complete(&pc->pc_starting);
436
437         if (rc != 0)
438                 RETURN(rc);
439
440         /*
441          * This mainloop strongly resembles ptlrpc_set_wait() except that our
442          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
443          * there are requests in the set. New requests come in on the set's
444          * new_req_list and ptlrpcd_check() moves them into the set.
445          */
446         do {
447                 struct l_wait_info lwi;
448                 int timeout;
449
450                 timeout = ptlrpc_set_next_timeout(set);
451                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
452                                   ptlrpc_expired_set, set);
453
454                 lu_context_enter(&env.le_ctx);
455                 l_wait_event(set->set_waitq,
456                              ptlrpcd_check(&env, pc), &lwi);
457                 lu_context_exit(&env.le_ctx);
458
459                 /*
460                  * Abort inflight rpcs for forced stop case.
461                  */
462                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
463                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
464                                 ptlrpc_abort_set(set);
465                         exit++;
466                 }
467
468                 /*
469                  * Let's make one more loop to make sure that ptlrpcd_check()
470                  * copied all raced new rpcs into the set so we can kill them.
471                  */
472         } while (exit < 2);
473
474         /*
475          * Wait for inflight requests to drain.
476          */
477         if (!cfs_list_empty(&set->set_requests))
478                 ptlrpc_set_wait(set);
479         lu_context_fini(&env.le_ctx);
480
481         clear_bit(LIOD_START, &pc->pc_flags);
482         clear_bit(LIOD_STOP, &pc->pc_flags);
483         clear_bit(LIOD_FORCE, &pc->pc_flags);
484         clear_bit(LIOD_BIND, &pc->pc_flags);
485
486         complete(&pc->pc_finishing);
487
488         return 0;
489 }
490
491 /* XXX: We want multiple CPU cores to share the async RPC load. So we start many
492  *      ptlrpcd threads. We also want to reduce the ptlrpcd overhead caused by
493  *      data transfer cross-CPU cores. So we bind ptlrpcd thread to specified
494  *      CPU core. But binding all ptlrpcd threads maybe cause response delay
495  *      because of some CPU core(s) busy with other loads.
496  *
497  *      For example: "ls -l", some async RPCs for statahead are assigned to
498  *      ptlrpcd_0, and ptlrpcd_0 is bound to CPU_0, but CPU_0 may be quite busy
499  *      with other non-ptlrpcd, like "ls -l" itself (we want to the "ls -l"
500  *      thread, statahead thread, and ptlrpcd thread can run in parallel), under
501  *      such case, the statahead async RPCs can not be processed in time, it is
502  *      unexpected. If ptlrpcd_0 can be re-scheduled on other CPU core, it may
503  *      be better. But it breaks former data transfer policy.
504  *
505  *      So we shouldn't be blind for avoiding the data transfer. We make some
506  *      compromise: divide the ptlrpcd threds pool into two parts. One part is
507  *      for bound mode, each ptlrpcd thread in this part is bound to some CPU
508  *      core. The other part is for free mode, all the ptlrpcd threads in the
509  *      part can be scheduled on any CPU core. We specify some partnership
510  *      between bound mode ptlrpcd thread(s) and free mode ptlrpcd thread(s),
511  *      and the async RPC load within the partners are shared.
512  *
513  *      It can partly avoid data transfer cross-CPU (if the bound mode ptlrpcd
514  *      thread can be scheduled in time), and try to guarantee the async RPC
515  *      processed ASAP (as long as the free mode ptlrpcd thread can be scheduled
516  *      on any CPU core).
517  *
518  *      As for how to specify the partnership between bound mode ptlrpcd
519  *      thread(s) and free mode ptlrpcd thread(s), the simplest way is to use
520  *      <free bound> pair. In future, we can specify some more complex
521  *      partnership based on the patches for CPU partition. But before such
522  *      patches are available, we prefer to use the simplest one.
523  */
524 # ifdef CFS_CPU_MODE_NUMA
525 # warning "fix ptlrpcd_bind() to use new CPU partition APIs"
526 # endif
527 static int ptlrpcd_bind(int index, int max)
528 {
529         struct ptlrpcd_ctl *pc;
530         int rc = 0;
531 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
532         struct ptlrpcd_ctl *ppc;
533         int node, i, pidx;
534         cpumask_t mask;
535 #endif
536         ENTRY;
537
538         LASSERT(index <= max - 1);
539         pc = &ptlrpcds->pd_threads[index];
540         switch (ptlrpcd_bind_policy) {
541         case PDB_POLICY_NONE:
542                 pc->pc_npartners = -1;
543                 break;
544         case PDB_POLICY_FULL:
545                 pc->pc_npartners = 0;
546                 set_bit(LIOD_BIND, &pc->pc_flags);
547                 break;
548         case PDB_POLICY_PAIR:
549                 LASSERT(max % 2 == 0);
550                 pc->pc_npartners = 1;
551                 break;
552         case PDB_POLICY_NEIGHBOR:
553 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
554                 node = cpu_to_node(index);
555                 mask = node_to_cpumask(node);
556                 for (i = max; i < cfs_num_online_cpus(); i++)
557                         cpu_clear(i, mask);
558                 pc->pc_npartners = cpus_weight(mask) - 1;
559                 set_bit(LIOD_BIND, &pc->pc_flags);
560 #else
561                 LASSERT(max >= 3);
562                 pc->pc_npartners = 2;
563 #endif
564                 break;
565         default:
566                 CERROR("unknown ptlrpcd bind policy %d\n", ptlrpcd_bind_policy);
567                 rc = -EINVAL;
568         }
569
570         if (rc == 0 && pc->pc_npartners > 0) {
571                 OBD_ALLOC(pc->pc_partners,
572                           sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
573                 if (pc->pc_partners == NULL) {
574                         pc->pc_npartners = 0;
575                         rc = -ENOMEM;
576                 } else {
577                         switch (ptlrpcd_bind_policy) {
578                         case PDB_POLICY_PAIR:
579                                 if (index & 0x1) {
580                                         set_bit(LIOD_BIND, &pc->pc_flags);
581                                         pc->pc_partners[0] = &ptlrpcds->
582                                                 pd_threads[index - 1];
583                                         ptlrpcds->pd_threads[index - 1].
584                                                 pc_partners[0] = pc;
585                                 }
586                                 break;
587                         case PDB_POLICY_NEIGHBOR:
588 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
589                                 /* partners are cores in the same NUMA node.
590                                  * setup partnership only with ptlrpcd threads
591                                  * that are already initialized
592                                  */
593                                 for (pidx = 0, i = 0; i < index; i++) {
594                                         if (cpu_isset(i, mask)) {
595                                                 ppc = &ptlrpcds->pd_threads[i];
596                                                 pc->pc_partners[pidx++] = ppc;
597                                                 ppc->pc_partners[ppc->
598                                                           pc_npartners++] = pc;
599                                         }
600                                 }
601                                 /* adjust number of partners to the number
602                                  * of partnership really setup */
603                                 pc->pc_npartners = pidx;
604 #else
605                                 if (index & 0x1)
606                                         set_bit(LIOD_BIND, &pc->pc_flags);
607                                 if (index > 0) {
608                                         pc->pc_partners[0] = &ptlrpcds->
609                                                 pd_threads[index - 1];
610                                         ptlrpcds->pd_threads[index - 1].
611                                                 pc_partners[1] = pc;
612                                         if (index == max - 1) {
613                                                 pc->pc_partners[1] =
614                                                 &ptlrpcds->pd_threads[0];
615                                                 ptlrpcds->pd_threads[0].
616                                                 pc_partners[0] = pc;
617                                         }
618                                 }
619 #endif
620                                 break;
621                         }
622                 }
623         }
624
625         RETURN(rc);
626 }
627
628 #else /* !__KERNEL__ */
629
630 /**
631  * In liblustre we do not have separate threads, so this function
632  * is called from time to time all across common code to see
633  * if something needs to be processed on ptlrpcd set.
634  */
635 int ptlrpcd_check_async_rpcs(void *arg)
636 {
637         struct ptlrpcd_ctl *pc = arg;
638         int                 rc = 0;
639
640         /*
641          * Single threaded!!
642          */
643         pc->pc_recurred++;
644
645         if (pc->pc_recurred == 1) {
646                 rc = lu_env_refill(&pc->pc_env);
647                 if (rc == 0) {
648                         lu_context_enter(&pc->pc_env.le_ctx);
649                         rc = ptlrpcd_check(&pc->pc_env, pc);
650                         if (!rc)
651                                 ptlrpc_expired_set(pc->pc_set);
652                         /*
653                          * XXX: send replay requests.
654                          */
655                         if (test_bit(LIOD_RECOVERY, &pc->pc_flags))
656                                 rc = ptlrpcd_check(&pc->pc_env, pc);
657                         lu_context_exit(&pc->pc_env.le_ctx);
658                 }
659         }
660
661         pc->pc_recurred--;
662         return rc;
663 }
664
665 int ptlrpcd_idle(void *arg)
666 {
667         struct ptlrpcd_ctl *pc = arg;
668
669         return (cfs_atomic_read(&pc->pc_set->set_new_count) == 0 &&
670                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
671 }
672
673 #endif
674
675 int ptlrpcd_start(int index, int max, const char *name, struct ptlrpcd_ctl *pc)
676 {
677         int rc;
678         int env = 0;
679         ENTRY;
680
681         /*
682          * Do not allow start second thread for one pc.
683          */
684         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
685                 CWARN("Starting second thread (%s) for same pc %p\n",
686                       name, pc);
687                 RETURN(0);
688         }
689
690         pc->pc_index = index;
691         init_completion(&pc->pc_starting);
692         init_completion(&pc->pc_finishing);
693         spin_lock_init(&pc->pc_lock);
694         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
695         pc->pc_set = ptlrpc_prep_set();
696         if (pc->pc_set == NULL)
697                 GOTO(out, rc = -ENOMEM);
698         /*
699          * So far only "client" ptlrpcd uses an environment. In the future,
700          * ptlrpcd thread (or a thread-set) has to be given an argument,
701          * describing its "scope".
702          */
703         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
704         if (rc != 0)
705                 GOTO(out, rc);
706
707         env = 1;
708 #ifdef __KERNEL__
709         if (index >= 0) {
710                 rc = ptlrpcd_bind(index, max);
711                 if (rc < 0)
712                         GOTO(out, rc);
713         }
714
715         rc = cfs_create_thread(ptlrpcd, pc, 0);
716         if (rc < 0)
717                 GOTO(out, rc);
718
719         rc = 0;
720         wait_for_completion(&pc->pc_starting);
721 #else
722         pc->pc_wait_callback =
723                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
724                                                  &ptlrpcd_check_async_rpcs, pc);
725         pc->pc_idle_callback =
726                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
727                                                  &ptlrpcd_idle, pc);
728 #endif
729 out:
730         if (rc) {
731 #ifdef __KERNEL__
732                 if (pc->pc_set != NULL) {
733                         struct ptlrpc_request_set *set = pc->pc_set;
734
735                         spin_lock(&pc->pc_lock);
736                         pc->pc_set = NULL;
737                         spin_unlock(&pc->pc_lock);
738                         ptlrpc_set_destroy(set);
739                 }
740                 if (env != 0)
741                         lu_context_fini(&pc->pc_env.le_ctx);
742                 clear_bit(LIOD_BIND, &pc->pc_flags);
743 #else
744                 SET_BUT_UNUSED(env);
745 #endif
746                 clear_bit(LIOD_START, &pc->pc_flags);
747         }
748         RETURN(rc);
749 }
750
751 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
752 {
753        struct ptlrpc_request_set *set = pc->pc_set;
754         ENTRY;
755
756         if (!test_bit(LIOD_START, &pc->pc_flags)) {
757                 CWARN("Thread for pc %p was not started\n", pc);
758                 goto out;
759         }
760
761         set_bit(LIOD_STOP, &pc->pc_flags);
762         if (force)
763                 set_bit(LIOD_FORCE, &pc->pc_flags);
764         cfs_waitq_signal(&pc->pc_set->set_waitq);
765 #ifdef __KERNEL__
766         wait_for_completion(&pc->pc_finishing);
767 #else
768         liblustre_deregister_wait_callback(pc->pc_wait_callback);
769         liblustre_deregister_idle_callback(pc->pc_idle_callback);
770 #endif
771         lu_context_fini(&pc->pc_env.le_ctx);
772
773         spin_lock(&pc->pc_lock);
774         pc->pc_set = NULL;
775         spin_unlock(&pc->pc_lock);
776         ptlrpc_set_destroy(set);
777
778 out:
779 #ifdef __KERNEL__
780         if (pc->pc_npartners > 0) {
781                 LASSERT(pc->pc_partners != NULL);
782
783                 OBD_FREE(pc->pc_partners,
784                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
785                 pc->pc_partners = NULL;
786         }
787         pc->pc_npartners = 0;
788 #endif
789         EXIT;
790 }
791
792 static void ptlrpcd_fini(void)
793 {
794         int i;
795         ENTRY;
796
797         if (ptlrpcds != NULL) {
798                 for (i = 0; i < ptlrpcds->pd_nthreads; i++)
799                         ptlrpcd_stop(&ptlrpcds->pd_threads[i], 0);
800                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
801                 OBD_FREE(ptlrpcds, ptlrpcds->pd_size);
802                 ptlrpcds = NULL;
803         }
804
805         EXIT;
806 }
807
808 static int ptlrpcd_init(void)
809 {
810         int nthreads = cfs_num_online_cpus();
811         char name[16];
812         int size, i = -1, j, rc = 0;
813         ENTRY;
814
815 #ifdef __KERNEL__
816         if (max_ptlrpcds > 0 && max_ptlrpcds < nthreads)
817                 nthreads = max_ptlrpcds;
818         if (nthreads < 2)
819                 nthreads = 2;
820         if (nthreads < 3 && ptlrpcd_bind_policy == PDB_POLICY_NEIGHBOR)
821                 ptlrpcd_bind_policy = PDB_POLICY_PAIR;
822         else if (nthreads % 2 != 0 && ptlrpcd_bind_policy == PDB_POLICY_PAIR)
823                 nthreads &= ~1; /* make sure it is even */
824 #else
825         nthreads = 1;
826 #endif
827
828         size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
829         OBD_ALLOC(ptlrpcds, size);
830         if (ptlrpcds == NULL)
831                 GOTO(out, rc = -ENOMEM);
832
833         snprintf(name, 15, "ptlrpcd_rcv");
834         set_bit(LIOD_RECOVERY, &ptlrpcds->pd_thread_rcv.pc_flags);
835         rc = ptlrpcd_start(-1, nthreads, name, &ptlrpcds->pd_thread_rcv);
836         if (rc < 0)
837                 GOTO(out, rc);
838
839         /* XXX: We start nthreads ptlrpc daemons. Each of them can process any
840          *      non-recovery async RPC to improve overall async RPC efficiency.
841          *
842          *      But there are some issues with async I/O RPCs and async non-I/O
843          *      RPCs processed in the same set under some cases. The ptlrpcd may
844          *      be blocked by some async I/O RPC(s), then will cause other async
845          *      non-I/O RPC(s) can not be processed in time.
846          *
847          *      Maybe we should distinguish blocked async RPCs from non-blocked
848          *      async RPCs, and process them in different ptlrpcd sets to avoid
849          *      unnecessary dependency. But how to distribute async RPCs load
850          *      among all the ptlrpc daemons becomes another trouble. */
851         for (i = 0; i < nthreads; i++) {
852                 snprintf(name, 15, "ptlrpcd_%d", i);
853                 rc = ptlrpcd_start(i, nthreads, name, &ptlrpcds->pd_threads[i]);
854                 if (rc < 0)
855                         GOTO(out, rc);
856         }
857
858         ptlrpcds->pd_size = size;
859         ptlrpcds->pd_index = 0;
860         ptlrpcds->pd_nthreads = nthreads;
861
862 out:
863         if (rc != 0 && ptlrpcds != NULL) {
864                 for (j = 0; j <= i; j++)
865                         ptlrpcd_stop(&ptlrpcds->pd_threads[j], 0);
866                 ptlrpcd_stop(&ptlrpcds->pd_thread_rcv, 0);
867                 OBD_FREE(ptlrpcds, size);
868                 ptlrpcds = NULL;
869         }
870
871         RETURN(0);
872 }
873
874 int ptlrpcd_addref(void)
875 {
876         int rc = 0;
877         ENTRY;
878
879         mutex_lock(&ptlrpcd_mutex);
880         if (++ptlrpcd_users == 1)
881                 rc = ptlrpcd_init();
882         mutex_unlock(&ptlrpcd_mutex);
883         RETURN(rc);
884 }
885 EXPORT_SYMBOL(ptlrpcd_addref);
886
887 void ptlrpcd_decref(void)
888 {
889         mutex_lock(&ptlrpcd_mutex);
890         if (--ptlrpcd_users == 0)
891                 ptlrpcd_fini();
892         mutex_unlock(&ptlrpcd_mutex);
893 }
894 EXPORT_SYMBOL(ptlrpcd_decref);
895 /** @} ptlrpcd */