Whamcloud - gitweb
LU-793 ptlrpc: fix ptlrpc_request_change_export()
[fs/lustre-release.git] / lustre / ptlrpc / service.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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2013, 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
37 #define DEBUG_SUBSYSTEM S_RPC
38 #ifndef __KERNEL__
39 #include <liblustre.h>
40 #endif
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_net.h>
44 #include <lu_object.h>
45 #include <lnet/types.h>
46 #include "ptlrpc_internal.h"
47
48 /* The following are visible and mutable through /sys/module/ptlrpc */
49 int test_req_buffer_pressure = 0;
50 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
51                 "set non-zero to put pressure on request buffer pools");
52 CFS_MODULE_PARM(at_min, "i", int, 0644,
53                 "Adaptive timeout minimum (sec)");
54 CFS_MODULE_PARM(at_max, "i", int, 0644,
55                 "Adaptive timeout maximum (sec)");
56 CFS_MODULE_PARM(at_history, "i", int, 0644,
57                 "Adaptive timeouts remember the slowest event that took place "
58                 "within this period (sec)");
59 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
60                 "How soon before an RPC deadline to send an early reply");
61 CFS_MODULE_PARM(at_extra, "i", int, 0644,
62                 "How much extra time to give with each early reply");
63
64
65 /* forward ref */
66 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
67 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
68 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
69
70 /** Holds a list of all PTLRPC services */
71 CFS_LIST_HEAD(ptlrpc_all_services);
72 /** Used to protect the \e ptlrpc_all_services list */
73 struct mutex ptlrpc_all_services_mutex;
74
75 struct ptlrpc_request_buffer_desc *
76 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
77 {
78         struct ptlrpc_service             *svc = svcpt->scp_service;
79         struct ptlrpc_request_buffer_desc *rqbd;
80
81         OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
82         if (rqbd == NULL)
83                 return NULL;
84
85         rqbd->rqbd_svcpt = svcpt;
86         rqbd->rqbd_refcount = 0;
87         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
88         rqbd->rqbd_cbid.cbid_arg = rqbd;
89         CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
90         OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
91                             svcpt->scp_cpt, svc->srv_buf_size);
92         if (rqbd->rqbd_buffer == NULL) {
93                 OBD_FREE_PTR(rqbd);
94                 return NULL;
95         }
96
97         spin_lock(&svcpt->scp_lock);
98         cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
99         svcpt->scp_nrqbds_total++;
100         spin_unlock(&svcpt->scp_lock);
101
102         return rqbd;
103 }
104
105 void
106 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
107 {
108         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
109
110         LASSERT(rqbd->rqbd_refcount == 0);
111         LASSERT(cfs_list_empty(&rqbd->rqbd_reqs));
112
113         spin_lock(&svcpt->scp_lock);
114         cfs_list_del(&rqbd->rqbd_list);
115         svcpt->scp_nrqbds_total--;
116         spin_unlock(&svcpt->scp_lock);
117
118         OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
119         OBD_FREE_PTR(rqbd);
120 }
121
122 int
123 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
124 {
125         struct ptlrpc_service             *svc = svcpt->scp_service;
126         struct ptlrpc_request_buffer_desc *rqbd;
127         int                                rc = 0;
128         int                                i;
129
130         if (svcpt->scp_rqbd_allocating)
131                 goto try_post;
132
133         spin_lock(&svcpt->scp_lock);
134         /* check again with lock */
135         if (svcpt->scp_rqbd_allocating) {
136                 /* NB: we might allow more than one thread in the future */
137                 LASSERT(svcpt->scp_rqbd_allocating == 1);
138                 spin_unlock(&svcpt->scp_lock);
139                 goto try_post;
140         }
141
142         svcpt->scp_rqbd_allocating++;
143         spin_unlock(&svcpt->scp_lock);
144
145
146         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
147                 /* NB: another thread might have recycled enough rqbds, we
148                  * need to make sure it wouldn't over-allocate, see LU-1212. */
149                 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
150                         break;
151
152                 rqbd = ptlrpc_alloc_rqbd(svcpt);
153
154                 if (rqbd == NULL) {
155                         CERROR("%s: Can't allocate request buffer\n",
156                                svc->srv_name);
157                         rc = -ENOMEM;
158                         break;
159                 }
160         }
161
162         spin_lock(&svcpt->scp_lock);
163
164         LASSERT(svcpt->scp_rqbd_allocating == 1);
165         svcpt->scp_rqbd_allocating--;
166
167         spin_unlock(&svcpt->scp_lock);
168
169         CDEBUG(D_RPCTRACE,
170                "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
171                svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
172                svcpt->scp_nrqbds_total, rc);
173
174  try_post:
175         if (post && rc == 0)
176                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
177
178         return rc;
179 }
180
181 /**
182  * Part of Rep-Ack logic.
183  * Puts a lock and its mode into reply state assotiated to request reply.
184  */
185 void
186 ptlrpc_save_lock(struct ptlrpc_request *req,
187                  struct lustre_handle *lock, int mode, int no_ack)
188 {
189         struct ptlrpc_reply_state *rs = req->rq_reply_state;
190         int                        idx;
191
192         LASSERT(rs != NULL);
193         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
194
195         if (req->rq_export->exp_disconnected) {
196                 ldlm_lock_decref(lock, mode);
197         } else {
198                 idx = rs->rs_nlocks++;
199                 rs->rs_locks[idx] = *lock;
200                 rs->rs_modes[idx] = mode;
201                 rs->rs_difficult = 1;
202                 rs->rs_no_ack = !!no_ack;
203         }
204 }
205 EXPORT_SYMBOL(ptlrpc_save_lock);
206
207 #ifdef __KERNEL__
208
209 struct ptlrpc_hr_partition;
210
211 struct ptlrpc_hr_thread {
212         int                             hrt_id;         /* thread ID */
213         spinlock_t                      hrt_lock;
214         wait_queue_head_t               hrt_waitq;
215         cfs_list_t                      hrt_queue;      /* RS queue */
216         struct ptlrpc_hr_partition      *hrt_partition;
217 };
218
219 struct ptlrpc_hr_partition {
220         /* # of started threads */
221         cfs_atomic_t                    hrp_nstarted;
222         /* # of stopped threads */
223         cfs_atomic_t                    hrp_nstopped;
224         /* cpu partition id */
225         int                             hrp_cpt;
226         /* round-robin rotor for choosing thread */
227         int                             hrp_rotor;
228         /* total number of threads on this partition */
229         int                             hrp_nthrs;
230         /* threads table */
231         struct ptlrpc_hr_thread         *hrp_thrs;
232 };
233
234 #define HRT_RUNNING 0
235 #define HRT_STOPPING 1
236
237 struct ptlrpc_hr_service {
238         /* CPU partition table, it's just cfs_cpt_table for now */
239         struct cfs_cpt_table            *hr_cpt_table;
240         /** controller sleep waitq */
241         wait_queue_head_t               hr_waitq;
242         unsigned int                    hr_stopping;
243         /** roundrobin rotor for non-affinity service */
244         unsigned int                    hr_rotor;
245         /* partition data */
246         struct ptlrpc_hr_partition      **hr_partitions;
247 };
248
249 struct rs_batch {
250         cfs_list_t                      rsb_replies;
251         unsigned int                    rsb_n_replies;
252         struct ptlrpc_service_part      *rsb_svcpt;
253 };
254
255 /** reply handling service. */
256 static struct ptlrpc_hr_service         ptlrpc_hr;
257
258 /**
259  * maximum mumber of replies scheduled in one batch
260  */
261 #define MAX_SCHEDULED 256
262
263 /**
264  * Initialize a reply batch.
265  *
266  * \param b batch
267  */
268 static void rs_batch_init(struct rs_batch *b)
269 {
270         memset(b, 0, sizeof *b);
271         CFS_INIT_LIST_HEAD(&b->rsb_replies);
272 }
273
274 /**
275  * Choose an hr thread to dispatch requests to.
276  */
277 static struct ptlrpc_hr_thread *
278 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
279 {
280         struct ptlrpc_hr_partition      *hrp;
281         unsigned int                    rotor;
282
283         if (svcpt->scp_cpt >= 0 &&
284             svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
285                 /* directly match partition */
286                 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
287
288         } else {
289                 rotor = ptlrpc_hr.hr_rotor++;
290                 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
291
292                 hrp = ptlrpc_hr.hr_partitions[rotor];
293         }
294
295         rotor = hrp->hrp_rotor++;
296         return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
297 }
298
299 /**
300  * Dispatch all replies accumulated in the batch to one from
301  * dedicated reply handling threads.
302  *
303  * \param b batch
304  */
305 static void rs_batch_dispatch(struct rs_batch *b)
306 {
307         if (b->rsb_n_replies != 0) {
308                 struct ptlrpc_hr_thread *hrt;
309
310                 hrt = ptlrpc_hr_select(b->rsb_svcpt);
311
312                 spin_lock(&hrt->hrt_lock);
313                 cfs_list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
314                 spin_unlock(&hrt->hrt_lock);
315
316                 wake_up(&hrt->hrt_waitq);
317                 b->rsb_n_replies = 0;
318         }
319 }
320
321 /**
322  * Add a reply to a batch.
323  * Add one reply object to a batch, schedule batched replies if overload.
324  *
325  * \param b batch
326  * \param rs reply
327  */
328 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
329 {
330         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
331
332         if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
333                 if (b->rsb_svcpt != NULL) {
334                         rs_batch_dispatch(b);
335                         spin_unlock(&b->rsb_svcpt->scp_rep_lock);
336                 }
337                 spin_lock(&svcpt->scp_rep_lock);
338                 b->rsb_svcpt = svcpt;
339         }
340         spin_lock(&rs->rs_lock);
341         rs->rs_scheduled_ever = 1;
342         if (rs->rs_scheduled == 0) {
343                 cfs_list_move(&rs->rs_list, &b->rsb_replies);
344                 rs->rs_scheduled = 1;
345                 b->rsb_n_replies++;
346         }
347         rs->rs_committed = 1;
348         spin_unlock(&rs->rs_lock);
349 }
350
351 /**
352  * Reply batch finalization.
353  * Dispatch remaining replies from the batch
354  * and release remaining spinlock.
355  *
356  * \param b batch
357  */
358 static void rs_batch_fini(struct rs_batch *b)
359 {
360         if (b->rsb_svcpt != NULL) {
361                 rs_batch_dispatch(b);
362                 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
363         }
364 }
365
366 #define DECLARE_RS_BATCH(b)     struct rs_batch b
367
368 #else /* __KERNEL__ */
369
370 #define rs_batch_init(b)        do{}while(0)
371 #define rs_batch_fini(b)        do{}while(0)
372 #define rs_batch_add(b, r)      ptlrpc_schedule_difficult_reply(r)
373 #define DECLARE_RS_BATCH(b)
374
375 #endif /* __KERNEL__ */
376
377 /**
378  * Put reply state into a queue for processing because we received
379  * ACK from the client
380  */
381 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
382 {
383 #ifdef __KERNEL__
384         struct ptlrpc_hr_thread *hrt;
385         ENTRY;
386
387         LASSERT(cfs_list_empty(&rs->rs_list));
388
389         hrt = ptlrpc_hr_select(rs->rs_svcpt);
390
391         spin_lock(&hrt->hrt_lock);
392         cfs_list_add_tail(&rs->rs_list, &hrt->hrt_queue);
393         spin_unlock(&hrt->hrt_lock);
394
395         wake_up(&hrt->hrt_waitq);
396         EXIT;
397 #else
398         cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue);
399 #endif
400 }
401
402 void
403 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
404 {
405         ENTRY;
406
407         LASSERT(spin_is_locked(&rs->rs_svcpt->scp_rep_lock));
408         LASSERT(spin_is_locked(&rs->rs_lock));
409         LASSERT (rs->rs_difficult);
410         rs->rs_scheduled_ever = 1;  /* flag any notification attempt */
411
412         if (rs->rs_scheduled) {     /* being set up or already notified */
413                 EXIT;
414                 return;
415         }
416
417         rs->rs_scheduled = 1;
418         cfs_list_del_init(&rs->rs_list);
419         ptlrpc_dispatch_difficult_reply(rs);
420         EXIT;
421 }
422 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
423
424 void ptlrpc_commit_replies(struct obd_export *exp)
425 {
426         struct ptlrpc_reply_state *rs, *nxt;
427         DECLARE_RS_BATCH(batch);
428         ENTRY;
429
430         rs_batch_init(&batch);
431         /* Find any replies that have been committed and get their service
432          * to attend to complete them. */
433
434         /* CAVEAT EMPTOR: spinlock ordering!!! */
435         spin_lock(&exp->exp_uncommitted_replies_lock);
436         cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
437                                      rs_obd_list) {
438                 LASSERT (rs->rs_difficult);
439                 /* VBR: per-export last_committed */
440                 LASSERT(rs->rs_export);
441                 if (rs->rs_transno <= exp->exp_last_committed) {
442                         cfs_list_del_init(&rs->rs_obd_list);
443                         rs_batch_add(&batch, rs);
444                 }
445         }
446         spin_unlock(&exp->exp_uncommitted_replies_lock);
447         rs_batch_fini(&batch);
448         EXIT;
449 }
450 EXPORT_SYMBOL(ptlrpc_commit_replies);
451
452 static int
453 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
454 {
455         struct ptlrpc_request_buffer_desc *rqbd;
456         int                               rc;
457         int                               posted = 0;
458
459         for (;;) {
460                 spin_lock(&svcpt->scp_lock);
461
462                 if (cfs_list_empty(&svcpt->scp_rqbd_idle)) {
463                         spin_unlock(&svcpt->scp_lock);
464                         return posted;
465                 }
466
467                 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
468                                       struct ptlrpc_request_buffer_desc,
469                                       rqbd_list);
470                 cfs_list_del(&rqbd->rqbd_list);
471
472                 /* assume we will post successfully */
473                 svcpt->scp_nrqbds_posted++;
474                 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
475
476                 spin_unlock(&svcpt->scp_lock);
477
478                 rc = ptlrpc_register_rqbd(rqbd);
479                 if (rc != 0)
480                         break;
481
482                 posted = 1;
483         }
484
485         spin_lock(&svcpt->scp_lock);
486
487         svcpt->scp_nrqbds_posted--;
488         cfs_list_del(&rqbd->rqbd_list);
489         cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
490
491         /* Don't complain if no request buffers are posted right now; LNET
492          * won't drop requests because we set the portal lazy! */
493
494         spin_unlock(&svcpt->scp_lock);
495
496         return -1;
497 }
498
499 static void ptlrpc_at_timer(unsigned long castmeharder)
500 {
501         struct ptlrpc_service_part *svcpt;
502
503         svcpt = (struct ptlrpc_service_part *)castmeharder;
504
505         svcpt->scp_at_check = 1;
506         svcpt->scp_at_checktime = cfs_time_current();
507         wake_up(&svcpt->scp_waitq);
508 }
509
510 static void
511 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
512                              struct ptlrpc_service_conf *conf)
513 {
514 #ifdef __KERNEL__
515         struct ptlrpc_service_thr_conf  *tc = &conf->psc_thr;
516         unsigned                        init;
517         unsigned                        total;
518         unsigned                        nthrs;
519         int                             weight;
520
521         /*
522          * Common code for estimating & validating threads number.
523          * CPT affinity service could have percpt thread-pool instead
524          * of a global thread-pool, which means user might not always
525          * get the threads number they give it in conf::tc_nthrs_user
526          * even they did set. It's because we need to validate threads
527          * number for each CPT to guarantee each pool will have enough
528          * threads to keep the service healthy.
529          */
530         init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
531         init = max_t(int, init, tc->tc_nthrs_init);
532
533         /* NB: please see comments in lustre_lnet.h for definition
534          * details of these members */
535         LASSERT(tc->tc_nthrs_max != 0);
536
537         if (tc->tc_nthrs_user != 0) {
538                 /* In case there is a reason to test a service with many
539                  * threads, we give a less strict check here, it can
540                  * be up to 8 * nthrs_max */
541                 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
542                 nthrs = total / svc->srv_ncpts;
543                 init  = max(init, nthrs);
544                 goto out;
545         }
546
547         total = tc->tc_nthrs_max;
548         if (tc->tc_nthrs_base == 0) {
549                 /* don't care about base threads number per partition,
550                  * this is most for non-affinity service */
551                 nthrs = total / svc->srv_ncpts;
552                 goto out;
553         }
554
555         nthrs = tc->tc_nthrs_base;
556         if (svc->srv_ncpts == 1) {
557                 int     i;
558
559                 /* NB: Increase the base number if it's single partition
560                  * and total number of cores/HTs is larger or equal to 4.
561                  * result will always < 2 * nthrs_base */
562                 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
563                 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
564                             (tc->tc_nthrs_base >> i) != 0; i++)
565                         nthrs += tc->tc_nthrs_base >> i;
566         }
567
568         if (tc->tc_thr_factor != 0) {
569                 int       factor = tc->tc_thr_factor;
570                 const int fade = 4;
571
572                 /*
573                  * User wants to increase number of threads with for
574                  * each CPU core/HT, most likely the factor is larger then
575                  * one thread/core because service threads are supposed to
576                  * be blocked by lock or wait for IO.
577                  */
578                 /*
579                  * Amdahl's law says that adding processors wouldn't give
580                  * a linear increasing of parallelism, so it's nonsense to
581                  * have too many threads no matter how many cores/HTs
582                  * there are.
583                  */
584                 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
585                         /* depress thread factor for hyper-thread */
586                         factor = factor - (factor >> 1) + (factor >> 3);
587                 }
588
589                 weight = cfs_cpt_weight(svc->srv_cptable, 0);
590                 LASSERT(weight > 0);
591
592                 for (; factor > 0 && weight > 0; factor--, weight -= fade)
593                         nthrs += min(weight, fade) * factor;
594         }
595
596         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
597                 nthrs = max(tc->tc_nthrs_base,
598                             tc->tc_nthrs_max / svc->srv_ncpts);
599         }
600  out:
601         nthrs = max(nthrs, tc->tc_nthrs_init);
602         svc->srv_nthrs_cpt_limit = nthrs;
603         svc->srv_nthrs_cpt_init = init;
604
605         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
606                 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
607                        "than the given soft limit (%d)\n",
608                        svc->srv_name, nthrs * svc->srv_ncpts,
609                        tc->tc_nthrs_max);
610         }
611 #endif
612 }
613
614 /**
615  * Initialize percpt data for a service
616  */
617 static int
618 ptlrpc_service_part_init(struct ptlrpc_service *svc,
619                          struct ptlrpc_service_part *svcpt, int cpt)
620 {
621         struct ptlrpc_at_array  *array;
622         int                     size;
623         int                     index;
624         int                     rc;
625
626         svcpt->scp_cpt = cpt;
627         CFS_INIT_LIST_HEAD(&svcpt->scp_threads);
628
629         /* rqbd and incoming request queue */
630         spin_lock_init(&svcpt->scp_lock);
631         CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
632         CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
633         CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming);
634         init_waitqueue_head(&svcpt->scp_waitq);
635         /* history request & rqbd list */
636         CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
637         CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
638
639         /* acitve requests and hp requests */
640         spin_lock_init(&svcpt->scp_req_lock);
641
642         /* reply states */
643         spin_lock_init(&svcpt->scp_rep_lock);
644         CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active);
645 #ifndef __KERNEL__
646         CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue);
647 #endif
648         CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle);
649         init_waitqueue_head(&svcpt->scp_rep_waitq);
650         cfs_atomic_set(&svcpt->scp_nreps_difficult, 0);
651
652         /* adaptive timeout */
653         spin_lock_init(&svcpt->scp_at_lock);
654         array = &svcpt->scp_at_array;
655
656         size = at_est2timeout(at_max);
657         array->paa_size     = size;
658         array->paa_count    = 0;
659         array->paa_deadline = -1;
660
661         /* allocate memory for scp_at_array (ptlrpc_at_array) */
662         OBD_CPT_ALLOC(array->paa_reqs_array,
663                       svc->srv_cptable, cpt, sizeof(cfs_list_t) * size);
664         if (array->paa_reqs_array == NULL)
665                 return -ENOMEM;
666
667         for (index = 0; index < size; index++)
668                 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
669
670         OBD_CPT_ALLOC(array->paa_reqs_count,
671                       svc->srv_cptable, cpt, sizeof(__u32) * size);
672         if (array->paa_reqs_count == NULL)
673                 goto failed;
674
675         cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
676         /* At SOW, service time should be quick; 10s seems generous. If client
677          * timeout is less than this, we'll be sending an early reply. */
678         at_init(&svcpt->scp_at_estimate, 10, 0);
679
680         /* assign this before call ptlrpc_grow_req_bufs */
681         svcpt->scp_service = svc;
682         /* Now allocate the request buffers, but don't post them now */
683         rc = ptlrpc_grow_req_bufs(svcpt, 0);
684         /* We shouldn't be under memory pressure at startup, so
685          * fail if we can't allocate all our buffers at this time. */
686         if (rc != 0)
687                 goto failed;
688
689         return 0;
690
691  failed:
692         if (array->paa_reqs_count != NULL) {
693                 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
694                 array->paa_reqs_count = NULL;
695         }
696
697         if (array->paa_reqs_array != NULL) {
698                 OBD_FREE(array->paa_reqs_array,
699                          sizeof(cfs_list_t) * array->paa_size);
700                 array->paa_reqs_array = NULL;
701         }
702
703         return -ENOMEM;
704 }
705
706 /**
707  * Initialize service on a given portal.
708  * This includes starting serving threads , allocating and posting rqbds and
709  * so on.
710  */
711 struct ptlrpc_service *
712 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
713                         struct proc_dir_entry *proc_entry)
714 {
715         struct ptlrpc_service_cpt_conf  *cconf = &conf->psc_cpt;
716         struct ptlrpc_service           *service;
717         struct ptlrpc_service_part      *svcpt;
718         struct cfs_cpt_table            *cptable;
719         __u32                           *cpts = NULL;
720         int                             ncpts;
721         int                             cpt;
722         int                             rc;
723         int                             i;
724         ENTRY;
725
726         LASSERT(conf->psc_buf.bc_nbufs > 0);
727         LASSERT(conf->psc_buf.bc_buf_size >=
728                 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
729         LASSERT(conf->psc_thr.tc_ctx_tags != 0);
730
731         cptable = cconf->cc_cptable;
732         if (cptable == NULL)
733                 cptable = cfs_cpt_table;
734
735         if (!conf->psc_thr.tc_cpu_affinity) {
736                 ncpts = 1;
737         } else {
738                 ncpts = cfs_cpt_number(cptable);
739                 if (cconf->cc_pattern != NULL) {
740                         struct cfs_expr_list    *el;
741
742                         rc = cfs_expr_list_parse(cconf->cc_pattern,
743                                                  strlen(cconf->cc_pattern),
744                                                  0, ncpts - 1, &el);
745                         if (rc != 0) {
746                                 CERROR("%s: invalid CPT pattern string: %s",
747                                        conf->psc_name, cconf->cc_pattern);
748                                 RETURN(ERR_PTR(-EINVAL));
749                         }
750
751                         rc = cfs_expr_list_values(el, ncpts, &cpts);
752                         cfs_expr_list_free(el);
753                         if (rc <= 0) {
754                                 CERROR("%s: failed to parse CPT array %s: %d\n",
755                                        conf->psc_name, cconf->cc_pattern, rc);
756                                 if (cpts != NULL)
757                                         OBD_FREE(cpts, sizeof(*cpts) * ncpts);
758                                 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
759                         }
760                         ncpts = rc;
761                 }
762         }
763
764         OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
765         if (service == NULL) {
766                 if (cpts != NULL)
767                         OBD_FREE(cpts, sizeof(*cpts) * ncpts);
768                 RETURN(ERR_PTR(-ENOMEM));
769         }
770
771         service->srv_cptable            = cptable;
772         service->srv_cpts               = cpts;
773         service->srv_ncpts              = ncpts;
774
775         service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
776         while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
777                 service->srv_cpt_bits++;
778
779         /* public members */
780         spin_lock_init(&service->srv_lock);
781         service->srv_name               = conf->psc_name;
782         service->srv_watchdog_factor    = conf->psc_watchdog_factor;
783         CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
784
785         /* buffer configuration */
786         service->srv_nbuf_per_group     = test_req_buffer_pressure ?
787                                           1 : conf->psc_buf.bc_nbufs;
788         service->srv_max_req_size       = conf->psc_buf.bc_req_max_size +
789                                           SPTLRPC_MAX_PAYLOAD;
790         service->srv_buf_size           = conf->psc_buf.bc_buf_size;
791         service->srv_rep_portal         = conf->psc_buf.bc_rep_portal;
792         service->srv_req_portal         = conf->psc_buf.bc_req_portal;
793
794         /* Increase max reply size to next power of two */
795         service->srv_max_reply_size = 1;
796         while (service->srv_max_reply_size <
797                conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
798                 service->srv_max_reply_size <<= 1;
799
800         service->srv_thread_name        = conf->psc_thr.tc_thr_name;
801         service->srv_ctx_tags           = conf->psc_thr.tc_ctx_tags;
802         service->srv_hpreq_ratio        = PTLRPC_SVC_HP_RATIO;
803         service->srv_ops                = conf->psc_ops;
804
805         for (i = 0; i < ncpts; i++) {
806                 if (!conf->psc_thr.tc_cpu_affinity)
807                         cpt = CFS_CPT_ANY;
808                 else
809                         cpt = cpts != NULL ? cpts[i] : i;
810
811                 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
812                 if (svcpt == NULL)
813                         GOTO(failed, rc = -ENOMEM);
814
815                 service->srv_parts[i] = svcpt;
816                 rc = ptlrpc_service_part_init(service, svcpt, cpt);
817                 if (rc != 0)
818                         GOTO(failed, rc);
819         }
820
821         ptlrpc_server_nthreads_check(service, conf);
822
823         rc = LNetSetLazyPortal(service->srv_req_portal);
824         LASSERT(rc == 0);
825
826         mutex_lock(&ptlrpc_all_services_mutex);
827         cfs_list_add (&service->srv_list, &ptlrpc_all_services);
828         mutex_unlock(&ptlrpc_all_services_mutex);
829
830         if (proc_entry != NULL)
831                 ptlrpc_lprocfs_register_service(proc_entry, service);
832
833         rc = ptlrpc_service_nrs_setup(service);
834         if (rc != 0)
835                 GOTO(failed, rc);
836
837         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
838                service->srv_name, service->srv_req_portal);
839
840 #ifdef __KERNEL__
841         rc = ptlrpc_start_threads(service);
842         if (rc != 0) {
843                 CERROR("Failed to start threads for service %s: %d\n",
844                        service->srv_name, rc);
845                 GOTO(failed, rc);
846         }
847 #endif
848
849         RETURN(service);
850 failed:
851         ptlrpc_unregister_service(service);
852         RETURN(ERR_PTR(rc));
853 }
854 EXPORT_SYMBOL(ptlrpc_register_service);
855
856 /**
857  * to actually free the request, must be called without holding svc_lock.
858  * note it's caller's responsibility to unlink req->rq_list.
859  */
860 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
861 {
862         LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
863         LASSERT(cfs_list_empty(&req->rq_timed_list));
864
865          /* DEBUG_REQ() assumes the reply state of a request with a valid
866           * ref will not be destroyed until that reference is dropped. */
867         ptlrpc_req_drop_rs(req);
868
869         sptlrpc_svc_ctx_decref(req);
870
871         if (req != &req->rq_rqbd->rqbd_req) {
872                 /* NB request buffers use an embedded
873                  * req if the incoming req unlinked the
874                  * MD; this isn't one of them! */
875                 ptlrpc_request_cache_free(req);
876         }
877 }
878
879 /**
880  * drop a reference count of the request. if it reaches 0, we either
881  * put it into history list, or free it immediately.
882  */
883 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
884 {
885         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
886         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
887         struct ptlrpc_service             *svc = svcpt->scp_service;
888         int                                refcount;
889         cfs_list_t                        *tmp;
890         cfs_list_t                        *nxt;
891
892         if (!cfs_atomic_dec_and_test(&req->rq_refcount))
893                 return;
894
895         if (req->rq_at_linked) {
896                 spin_lock(&svcpt->scp_at_lock);
897                 /* recheck with lock, in case it's unlinked by
898                  * ptlrpc_at_check_timed() */
899                 if (likely(req->rq_at_linked))
900                         ptlrpc_at_remove_timed(req);
901                 spin_unlock(&svcpt->scp_at_lock);
902         }
903
904         LASSERT(cfs_list_empty(&req->rq_timed_list));
905
906         /* finalize request */
907         if (req->rq_export) {
908                 class_export_put(req->rq_export);
909                 req->rq_export = NULL;
910         }
911
912         spin_lock(&svcpt->scp_lock);
913
914         cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
915
916         refcount = --(rqbd->rqbd_refcount);
917         if (refcount == 0) {
918                 /* request buffer is now idle: add to history */
919                 cfs_list_del(&rqbd->rqbd_list);
920
921                 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
922                 svcpt->scp_hist_nrqbds++;
923
924                 /* cull some history?
925                  * I expect only about 1 or 2 rqbds need to be recycled here */
926                 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
927                         rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next,
928                                               struct ptlrpc_request_buffer_desc,
929                                               rqbd_list);
930
931                         cfs_list_del(&rqbd->rqbd_list);
932                         svcpt->scp_hist_nrqbds--;
933
934                         /* remove rqbd's reqs from svc's req history while
935                          * I've got the service lock */
936                         cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
937                                 req = cfs_list_entry(tmp, struct ptlrpc_request,
938                                                      rq_list);
939                                 /* Track the highest culled req seq */
940                                 if (req->rq_history_seq >
941                                     svcpt->scp_hist_seq_culled) {
942                                         svcpt->scp_hist_seq_culled =
943                                                 req->rq_history_seq;
944                                 }
945                                 cfs_list_del(&req->rq_history_list);
946                         }
947
948                         spin_unlock(&svcpt->scp_lock);
949
950                         cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
951                                 req = cfs_list_entry(rqbd->rqbd_reqs.next,
952                                                      struct ptlrpc_request,
953                                                      rq_list);
954                                 cfs_list_del(&req->rq_list);
955                                 ptlrpc_server_free_request(req);
956                         }
957
958                         spin_lock(&svcpt->scp_lock);
959                         /*
960                          * now all reqs including the embedded req has been
961                          * disposed, schedule request buffer for re-use.
962                          */
963                         LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
964                                 0);
965                         cfs_list_add_tail(&rqbd->rqbd_list,
966                                           &svcpt->scp_rqbd_idle);
967                 }
968
969                 spin_unlock(&svcpt->scp_lock);
970         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
971                 /* If we are low on memory, we are not interested in history */
972                 cfs_list_del(&req->rq_list);
973                 cfs_list_del_init(&req->rq_history_list);
974
975                 /* Track the highest culled req seq */
976                 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
977                         svcpt->scp_hist_seq_culled = req->rq_history_seq;
978
979                 spin_unlock(&svcpt->scp_lock);
980
981                 ptlrpc_server_free_request(req);
982         } else {
983                 spin_unlock(&svcpt->scp_lock);
984         }
985 }
986
987 /** Change request export and move hp request from old export to new */
988 void ptlrpc_request_change_export(struct ptlrpc_request *req,
989                                   struct obd_export *export)
990 {
991         if (req->rq_export != NULL) {
992                 LASSERT(!list_empty(&req->rq_exp_list));
993                 /* remove rq_exp_list from last export */
994                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
995                 list_del_init(&req->rq_exp_list);
996                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
997                 /* export has one reference already, so it`s safe to
998                  * add req to export queue here and get another
999                  * reference for request later */
1000                 spin_lock_bh(&export->exp_rpc_lock);
1001                 if (req->rq_ops != NULL) /* hp request */
1002                         list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1003                 else
1004                         list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1005                 spin_unlock_bh(&export->exp_rpc_lock);
1006
1007                 class_export_rpc_dec(req->rq_export);
1008                 class_export_put(req->rq_export);
1009         }
1010
1011         /* request takes one export refcount */
1012         req->rq_export = class_export_get(export);
1013         class_export_rpc_inc(export);
1014
1015         return;
1016 }
1017
1018 /**
1019  * to finish a request: stop sending more early replies, and release
1020  * the request.
1021  */
1022 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1023                                          struct ptlrpc_request *req)
1024 {
1025         ptlrpc_server_hpreq_fini(req);
1026
1027         if (req->rq_session.lc_thread != NULL) {
1028                 lu_context_exit(&req->rq_session);
1029                 lu_context_fini(&req->rq_session);
1030         }
1031
1032         ptlrpc_server_drop_request(req);
1033 }
1034
1035 /**
1036  * to finish a active request: stop sending more early replies, and release
1037  * the request. should be called after we finished handling the request.
1038  */
1039 static void ptlrpc_server_finish_active_request(
1040                                         struct ptlrpc_service_part *svcpt,
1041                                         struct ptlrpc_request *req)
1042 {
1043         spin_lock(&svcpt->scp_req_lock);
1044         ptlrpc_nrs_req_stop_nolock(req);
1045         svcpt->scp_nreqs_active--;
1046         if (req->rq_hp)
1047                 svcpt->scp_nhreqs_active--;
1048         spin_unlock(&svcpt->scp_req_lock);
1049
1050         ptlrpc_nrs_req_finalize(req);
1051
1052         if (req->rq_export != NULL)
1053                 class_export_rpc_dec(req->rq_export);
1054
1055         ptlrpc_server_finish_request(svcpt, req);
1056 }
1057
1058 /**
1059  * This function makes sure dead exports are evicted in a timely manner.
1060  * This function is only called when some export receives a message (i.e.,
1061  * the network is up.)
1062  */
1063 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1064 {
1065         struct obd_export *oldest_exp;
1066         time_t oldest_time, new_time;
1067
1068         ENTRY;
1069
1070         LASSERT(exp);
1071
1072         /* Compensate for slow machines, etc, by faking our request time
1073            into the future.  Although this can break the strict time-ordering
1074            of the list, we can be really lazy here - we don't have to evict
1075            at the exact right moment.  Eventually, all silent exports
1076            will make it to the top of the list. */
1077
1078         /* Do not pay attention on 1sec or smaller renewals. */
1079         new_time = cfs_time_current_sec() + extra_delay;
1080         if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1081                 RETURN_EXIT;
1082
1083         exp->exp_last_request_time = new_time;
1084         CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1085                exp->exp_client_uuid.uuid,
1086                exp->exp_last_request_time, exp);
1087
1088         /* exports may get disconnected from the chain even though the
1089            export has references, so we must keep the spin lock while
1090            manipulating the lists */
1091         spin_lock(&exp->exp_obd->obd_dev_lock);
1092
1093         if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
1094                 /* this one is not timed */
1095                 spin_unlock(&exp->exp_obd->obd_dev_lock);
1096                 RETURN_EXIT;
1097         }
1098
1099         cfs_list_move_tail(&exp->exp_obd_chain_timed,
1100                            &exp->exp_obd->obd_exports_timed);
1101
1102         oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
1103                                     struct obd_export, exp_obd_chain_timed);
1104         oldest_time = oldest_exp->exp_last_request_time;
1105         spin_unlock(&exp->exp_obd->obd_dev_lock);
1106
1107         if (exp->exp_obd->obd_recovering) {
1108                 /* be nice to everyone during recovery */
1109                 EXIT;
1110                 return;
1111         }
1112
1113         /* Note - racing to start/reset the obd_eviction timer is safe */
1114         if (exp->exp_obd->obd_eviction_timer == 0) {
1115                 /* Check if the oldest entry is expired. */
1116                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1117                                               extra_delay)) {
1118                         /* We need a second timer, in case the net was down and
1119                          * it just came back. Since the pinger may skip every
1120                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1121                          * we better wait for 3. */
1122                         exp->exp_obd->obd_eviction_timer =
1123                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
1124                         CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1125                                exp->exp_obd->obd_name, 
1126                                obd_export_nid2str(oldest_exp), oldest_time);
1127                 }
1128         } else {
1129                 if (cfs_time_current_sec() >
1130                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1131                         /* The evictor won't evict anyone who we've heard from
1132                          * recently, so we don't have to check before we start
1133                          * it. */
1134                         if (!ping_evictor_wake(exp))
1135                                 exp->exp_obd->obd_eviction_timer = 0;
1136                 }
1137         }
1138
1139         EXIT;
1140 }
1141
1142 /**
1143  * Sanity check request \a req.
1144  * Return 0 if all is ok, error code otherwise.
1145  */
1146 static int ptlrpc_check_req(struct ptlrpc_request *req)
1147 {
1148         int rc = 0;
1149
1150         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1151                      req->rq_export->exp_conn_cnt)) {
1152                 DEBUG_REQ(D_RPCTRACE, req,
1153                           "DROPPING req from old connection %d < %d",
1154                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
1155                           req->rq_export->exp_conn_cnt);
1156                 return -EEXIST;
1157         }
1158         if (unlikely(req->rq_export->exp_obd &&
1159                      req->rq_export->exp_obd->obd_fail)) {
1160              /* Failing over, don't handle any more reqs, send
1161                 error response instead. */
1162                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1163                        req, req->rq_export->exp_obd->obd_name);
1164                 rc = -ENODEV;
1165         } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1166                    (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1167                    !(req->rq_export->exp_obd->obd_recovering)) {
1168                         DEBUG_REQ(D_ERROR, req,
1169                                   "Invalid replay without recovery");
1170                         class_fail_export(req->rq_export);
1171                         rc = -ENODEV;
1172         } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1173                    !(req->rq_export->exp_obd->obd_recovering)) {
1174                         DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1175                                   LPU64" without recovery",
1176                                   lustre_msg_get_transno(req->rq_reqmsg));
1177                         class_fail_export(req->rq_export);
1178                         rc = -ENODEV;
1179         }
1180
1181         if (unlikely(rc < 0)) {
1182                 req->rq_status = rc;
1183                 ptlrpc_error(req);
1184         }
1185         return rc;
1186 }
1187
1188 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1189 {
1190         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1191         __s32 next;
1192
1193         if (array->paa_count == 0) {
1194                 cfs_timer_disarm(&svcpt->scp_at_timer);
1195                 return;
1196         }
1197
1198         /* Set timer for closest deadline */
1199         next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1200                        at_early_margin);
1201         if (next <= 0) {
1202                 ptlrpc_at_timer((unsigned long)svcpt);
1203         } else {
1204                 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1205                 CDEBUG(D_INFO, "armed %s at %+ds\n",
1206                        svcpt->scp_service->srv_name, next);
1207         }
1208 }
1209
1210 /* Add rpc to early reply check list */
1211 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1212 {
1213         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1214         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1215         struct ptlrpc_request *rq = NULL;
1216         __u32 index;
1217
1218         if (AT_OFF)
1219                 return(0);
1220
1221         if (req->rq_no_reply)
1222                 return 0;
1223
1224         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1225                 return(-ENOSYS);
1226
1227         spin_lock(&svcpt->scp_at_lock);
1228         LASSERT(cfs_list_empty(&req->rq_timed_list));
1229
1230         index = (unsigned long)req->rq_deadline % array->paa_size;
1231         if (array->paa_reqs_count[index] > 0) {
1232                 /* latest rpcs will have the latest deadlines in the list,
1233                  * so search backward. */
1234                 cfs_list_for_each_entry_reverse(rq,
1235                                                 &array->paa_reqs_array[index],
1236                                                 rq_timed_list) {
1237                         if (req->rq_deadline >= rq->rq_deadline) {
1238                                 cfs_list_add(&req->rq_timed_list,
1239                                              &rq->rq_timed_list);
1240                                 break;
1241                         }
1242                 }
1243         }
1244
1245         /* Add the request at the head of the list */
1246         if (cfs_list_empty(&req->rq_timed_list))
1247                 cfs_list_add(&req->rq_timed_list,
1248                              &array->paa_reqs_array[index]);
1249
1250         spin_lock(&req->rq_lock);
1251         req->rq_at_linked = 1;
1252         spin_unlock(&req->rq_lock);
1253         req->rq_at_index = index;
1254         array->paa_reqs_count[index]++;
1255         array->paa_count++;
1256         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1257                 array->paa_deadline = req->rq_deadline;
1258                 ptlrpc_at_set_timer(svcpt);
1259         }
1260         spin_unlock(&svcpt->scp_at_lock);
1261
1262         return 0;
1263 }
1264
1265 static void
1266 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1267 {
1268         struct ptlrpc_at_array *array;
1269
1270         array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1271
1272         /* NB: must call with hold svcpt::scp_at_lock */
1273         LASSERT(!cfs_list_empty(&req->rq_timed_list));
1274         cfs_list_del_init(&req->rq_timed_list);
1275
1276         spin_lock(&req->rq_lock);
1277         req->rq_at_linked = 0;
1278         spin_unlock(&req->rq_lock);
1279
1280         array->paa_reqs_count[req->rq_at_index]--;
1281         array->paa_count--;
1282 }
1283
1284 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1285 {
1286         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1287         struct ptlrpc_request *reqcopy;
1288         struct lustre_msg *reqmsg;
1289         cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1290         time_t newdl;
1291         int rc;
1292         ENTRY;
1293
1294         /* deadline is when the client expects us to reply, margin is the
1295            difference between clients' and servers' expectations */
1296         DEBUG_REQ(D_ADAPTTO, req,
1297                   "%ssending early reply (deadline %+lds, margin %+lds) for "
1298                   "%d+%d", AT_OFF ? "AT off - not " : "",
1299                   olddl, olddl - at_get(&svcpt->scp_at_estimate),
1300                   at_get(&svcpt->scp_at_estimate), at_extra);
1301
1302         if (AT_OFF)
1303                 RETURN(0);
1304
1305         if (olddl < 0) {
1306                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1307                           "not sending early reply. Consider increasing "
1308                           "at_early_margin (%d)?", olddl, at_early_margin);
1309
1310                 /* Return an error so we're not re-added to the timed list. */
1311                 RETURN(-ETIMEDOUT);
1312         }
1313
1314         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1315                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1316                           "but no AT support");
1317                 RETURN(-ENOSYS);
1318         }
1319
1320         if (req->rq_export &&
1321             lustre_msg_get_flags(req->rq_reqmsg) &
1322             (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1323                 /* During recovery, we don't want to send too many early
1324                  * replies, but on the other hand we want to make sure the
1325                  * client has enough time to resend if the rpc is lost. So
1326                  * during the recovery period send at least 4 early replies,
1327                  * spacing them every at_extra if we can. at_estimate should
1328                  * always equal this fixed value during recovery. */
1329                 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1330                             req->rq_export->exp_obd->obd_recovery_timeout / 4));
1331         } else {
1332                 /* Fake our processing time into the future to ask the clients
1333                  * for some extra amount of time */
1334                 at_measured(&svcpt->scp_at_estimate, at_extra +
1335                             cfs_time_current_sec() -
1336                             req->rq_arrival_time.tv_sec);
1337
1338                 /* Check to see if we've actually increased the deadline -
1339                  * we may be past adaptive_max */
1340                 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1341                     at_get(&svcpt->scp_at_estimate)) {
1342                         DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1343                                   "(%ld/%ld), not sending early reply\n",
1344                                   olddl, req->rq_arrival_time.tv_sec +
1345                                   at_get(&svcpt->scp_at_estimate) -
1346                                   cfs_time_current_sec());
1347                         RETURN(-ETIMEDOUT);
1348                 }
1349         }
1350         newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1351
1352         reqcopy = ptlrpc_request_cache_alloc(__GFP_IO);
1353         if (reqcopy == NULL)
1354                 RETURN(-ENOMEM);
1355         OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1356         if (!reqmsg)
1357                 GOTO(out_free, rc = -ENOMEM);
1358
1359         *reqcopy = *req;
1360         reqcopy->rq_reply_state = NULL;
1361         reqcopy->rq_rep_swab_mask = 0;
1362         reqcopy->rq_pack_bulk = 0;
1363         reqcopy->rq_pack_udesc = 0;
1364         reqcopy->rq_packed_final = 0;
1365         sptlrpc_svc_ctx_addref(reqcopy);
1366         /* We only need the reqmsg for the magic */
1367         reqcopy->rq_reqmsg = reqmsg;
1368         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1369
1370         LASSERT(cfs_atomic_read(&req->rq_refcount));
1371         /** if it is last refcount then early reply isn't needed */
1372         if (cfs_atomic_read(&req->rq_refcount) == 1) {
1373                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1374                           "abort sending early reply\n");
1375                 GOTO(out, rc = -EINVAL);
1376         }
1377
1378         /* Connection ref */
1379         reqcopy->rq_export = class_conn2export(
1380                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
1381         if (reqcopy->rq_export == NULL)
1382                 GOTO(out, rc = -ENODEV);
1383
1384         /* RPC ref */
1385         class_export_rpc_inc(reqcopy->rq_export);
1386         if (reqcopy->rq_export->exp_obd &&
1387             reqcopy->rq_export->exp_obd->obd_fail)
1388                 GOTO(out_put, rc = -ENODEV);
1389
1390         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1391         if (rc)
1392                 GOTO(out_put, rc);
1393
1394         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1395
1396         if (!rc) {
1397                 /* Adjust our own deadline to what we told the client */
1398                 req->rq_deadline = newdl;
1399                 req->rq_early_count++; /* number sent, server side */
1400         } else {
1401                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1402         }
1403
1404         /* Free the (early) reply state from lustre_pack_reply.
1405            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1406         ptlrpc_req_drop_rs(reqcopy);
1407
1408 out_put:
1409         class_export_rpc_dec(reqcopy->rq_export);
1410         class_export_put(reqcopy->rq_export);
1411 out:
1412         sptlrpc_svc_ctx_decref(reqcopy);
1413         OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1414 out_free:
1415         ptlrpc_request_cache_free(reqcopy);
1416         RETURN(rc);
1417 }
1418
1419 /* Send early replies to everybody expiring within at_early_margin
1420    asking for at_extra time */
1421 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1422 {
1423         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1424         struct ptlrpc_request *rq, *n;
1425         cfs_list_t work_list;
1426         __u32  index, count;
1427         time_t deadline;
1428         time_t now = cfs_time_current_sec();
1429         cfs_duration_t delay;
1430         int first, counter = 0;
1431         ENTRY;
1432
1433         spin_lock(&svcpt->scp_at_lock);
1434         if (svcpt->scp_at_check == 0) {
1435                 spin_unlock(&svcpt->scp_at_lock);
1436                 RETURN(0);
1437         }
1438         delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1439         svcpt->scp_at_check = 0;
1440
1441         if (array->paa_count == 0) {
1442                 spin_unlock(&svcpt->scp_at_lock);
1443                 RETURN(0);
1444         }
1445
1446         /* The timer went off, but maybe the nearest rpc already completed. */
1447         first = array->paa_deadline - now;
1448         if (first > at_early_margin) {
1449                 /* We've still got plenty of time.  Reset the timer. */
1450                 ptlrpc_at_set_timer(svcpt);
1451                 spin_unlock(&svcpt->scp_at_lock);
1452                 RETURN(0);
1453         }
1454
1455         /* We're close to a timeout, and we don't know how much longer the
1456            server will take. Send early replies to everyone expiring soon. */
1457         CFS_INIT_LIST_HEAD(&work_list);
1458         deadline = -1;
1459         index = (unsigned long)array->paa_deadline % array->paa_size;
1460         count = array->paa_count;
1461         while (count > 0) {
1462                 count -= array->paa_reqs_count[index];
1463                 cfs_list_for_each_entry_safe(rq, n,
1464                                              &array->paa_reqs_array[index],
1465                                              rq_timed_list) {
1466                         if (rq->rq_deadline > now + at_early_margin) {
1467                                 /* update the earliest deadline */
1468                                 if (deadline == -1 ||
1469                                     rq->rq_deadline < deadline)
1470                                         deadline = rq->rq_deadline;
1471                                 break;
1472                         }
1473
1474                         ptlrpc_at_remove_timed(rq);
1475                         /**
1476                          * ptlrpc_server_drop_request() may drop
1477                          * refcount to 0 already. Let's check this and
1478                          * don't add entry to work_list
1479                          */
1480                         if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1481                                 cfs_list_add(&rq->rq_timed_list, &work_list);
1482                         counter++;
1483                 }
1484
1485                 if (++index >= array->paa_size)
1486                         index = 0;
1487         }
1488         array->paa_deadline = deadline;
1489         /* we have a new earliest deadline, restart the timer */
1490         ptlrpc_at_set_timer(svcpt);
1491
1492         spin_unlock(&svcpt->scp_at_lock);
1493
1494         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1495                "replies\n", first, at_extra, counter);
1496         if (first < 0) {
1497                 /* We're already past request deadlines before we even get a
1498                    chance to send early replies */
1499                 LCONSOLE_WARN("%s: This server is not able to keep up with "
1500                               "request traffic (cpu-bound).\n",
1501                               svcpt->scp_service->srv_name);
1502                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1503                       "delay="CFS_DURATION_T"(jiff)\n",
1504                       counter, svcpt->scp_nreqs_incoming,
1505                       svcpt->scp_nreqs_active,
1506                       at_get(&svcpt->scp_at_estimate), delay);
1507         }
1508
1509         /* we took additional refcount so entries can't be deleted from list, no
1510          * locking is needed */
1511         while (!cfs_list_empty(&work_list)) {
1512                 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1513                                     rq_timed_list);
1514                 cfs_list_del_init(&rq->rq_timed_list);
1515
1516                 if (ptlrpc_at_send_early_reply(rq) == 0)
1517                         ptlrpc_at_add_timed(rq);
1518
1519                 ptlrpc_server_drop_request(rq);
1520         }
1521
1522         RETURN(1); /* return "did_something" for liblustre */
1523 }
1524
1525 /* Check if we are already handling earlier incarnation of this request.
1526  * Called under &req->rq_export->exp_rpc_lock locked */
1527 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1528 {
1529         struct ptlrpc_request   *tmp = NULL;
1530
1531         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1532             (cfs_atomic_read(&req->rq_export->exp_rpc_count) == 0))
1533                 return 0;
1534
1535         /* bulk request are aborted upon reconnect, don't try to
1536          * find a match */
1537         if (req->rq_bulk_write || req->rq_bulk_read)
1538                 return 0;
1539
1540         /* This list should not be longer than max_requests in
1541          * flights on the client, so it is not all that long.
1542          * Also we only hit this codepath in case of a resent
1543          * request which makes it even more rarely hit */
1544         cfs_list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1545                                 rq_exp_list) {
1546                 /* Found duplicate one */
1547                 if (tmp->rq_xid == req->rq_xid)
1548                         goto found;
1549         }
1550         cfs_list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1551                                 rq_exp_list) {
1552                 /* Found duplicate one */
1553                 if (tmp->rq_xid == req->rq_xid)
1554                         goto found;
1555         }
1556         return 0;
1557
1558 found:
1559         DEBUG_REQ(D_HA, req, "Found duplicate req in processing\n");
1560         DEBUG_REQ(D_HA, tmp, "Request being processed\n");
1561         return -EBUSY;
1562 }
1563
1564 /**
1565  * Put the request to the export list if the request may become
1566  * a high priority one.
1567  */
1568 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1569                                     struct ptlrpc_request *req)
1570 {
1571         cfs_list_t      *list;
1572         int              rc, hp = 0;
1573
1574         ENTRY;
1575
1576         if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1577                 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1578                 if (rc < 0)
1579                         RETURN(rc);
1580                 LASSERT(rc == 0);
1581         }
1582         if (req->rq_export) {
1583                 if (req->rq_ops) {
1584                         /* Perform request specific check. We should do this
1585                          * check before the request is added into exp_hp_rpcs
1586                          * list otherwise it may hit swab race at LU-1044. */
1587                         if (req->rq_ops->hpreq_check) {
1588                                 rc = req->rq_ops->hpreq_check(req);
1589                                 /**
1590                                  * XXX: Out of all current
1591                                  * ptlrpc_hpreq_ops::hpreq_check(), only
1592                                  * ldlm_cancel_hpreq_check() can return an
1593                                  * error code; other functions assert in
1594                                  * similar places, which seems odd.
1595                                  * What also does not seem right is that
1596                                  * handlers for those RPCs do not assert
1597                                  * on the same checks, but rather handle the
1598                                  * error cases. e.g. see ost_rw_hpreq_check(),
1599                                  * and ost_brw_read(), ost_brw_write().
1600                                  */
1601                                 if (rc < 0)
1602                                         RETURN(rc);
1603                                 LASSERT(rc == 0 || rc == 1);
1604                                 hp = rc;
1605                         }
1606                         list = &req->rq_export->exp_hp_rpcs;
1607                 } else {
1608                         list = &req->rq_export->exp_reg_rpcs;
1609                 }
1610
1611                 /* do search for duplicated xid and the adding to the list
1612                  * atomically */
1613                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1614                 rc = ptlrpc_server_check_resend_in_progress(req);
1615                 if (rc < 0) {
1616                         spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1617                         RETURN(rc);
1618                 }
1619                 cfs_list_add(&req->rq_exp_list, list);
1620                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1621         }
1622
1623         ptlrpc_nrs_req_initialize(svcpt, req, !!hp);
1624
1625         RETURN(hp);
1626 }
1627
1628 /** Remove the request from the export list. */
1629 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1630 {
1631         ENTRY;
1632         if (req->rq_export) {
1633                 /* refresh lock timeout again so that client has more
1634                  * room to send lock cancel RPC. */
1635                 if (req->rq_ops && req->rq_ops->hpreq_fini)
1636                         req->rq_ops->hpreq_fini(req);
1637
1638                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1639                 cfs_list_del_init(&req->rq_exp_list);
1640                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1641         }
1642         EXIT;
1643 }
1644
1645 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1646 {
1647         return 1;
1648 }
1649
1650 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1651         .hpreq_check       = ptlrpc_hpreq_check,
1652 };
1653
1654 /* Hi-Priority RPC check by RPC operation code. */
1655 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1656 {
1657         int opc = lustre_msg_get_opc(req->rq_reqmsg);
1658
1659         /* Check for export to let only reconnects for not yet evicted
1660          * export to become a HP rpc. */
1661         if ((req->rq_export != NULL) &&
1662             (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1663                 req->rq_ops = &ptlrpc_hpreq_common;
1664
1665         return 0;
1666 }
1667 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1668
1669 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1670                                      struct ptlrpc_request *req)
1671 {
1672         int     rc;
1673         ENTRY;
1674
1675         rc = ptlrpc_server_hpreq_init(svcpt, req);
1676         if (rc < 0)
1677                 RETURN(rc);
1678
1679         ptlrpc_nrs_req_add(svcpt, req, !!rc);
1680
1681         RETURN(0);
1682 }
1683
1684 /**
1685  * Allow to handle high priority request
1686  * User can call it w/o any lock but need to hold
1687  * ptlrpc_service_part::scp_req_lock to get reliable result
1688  */
1689 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1690                                      bool force)
1691 {
1692         int running = svcpt->scp_nthrs_running;
1693
1694         if (!nrs_svcpt_has_hp(svcpt))
1695                 return false;
1696
1697         if (force)
1698                 return true;
1699
1700         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1701                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1702                 /* leave just 1 thread for normal RPCs */
1703                 running = PTLRPC_NTHRS_INIT;
1704                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1705                         running += 1;
1706         }
1707
1708         if (svcpt->scp_nreqs_active >= running - 1)
1709                 return false;
1710
1711         if (svcpt->scp_nhreqs_active == 0)
1712                 return true;
1713
1714         return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1715                svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1716 }
1717
1718 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1719                                        bool force)
1720 {
1721         return ptlrpc_server_allow_high(svcpt, force) &&
1722                ptlrpc_nrs_req_pending_nolock(svcpt, true);
1723 }
1724
1725 /**
1726  * Only allow normal priority requests on a service that has a high-priority
1727  * queue if forced (i.e. cleanup), if there are other high priority requests
1728  * already being processed (i.e. those threads can service more high-priority
1729  * requests), or if there are enough idle threads that a later thread can do
1730  * a high priority request.
1731  * User can call it w/o any lock but need to hold
1732  * ptlrpc_service_part::scp_req_lock to get reliable result
1733  */
1734 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1735                                        bool force)
1736 {
1737         int running = svcpt->scp_nthrs_running;
1738 #ifndef __KERNEL__
1739         if (1) /* always allow to handle normal request for liblustre */
1740                 return true;
1741 #endif
1742         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1743                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1744                 /* leave just 1 thread for normal RPCs */
1745                 running = PTLRPC_NTHRS_INIT;
1746                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1747                         running += 1;
1748         }
1749
1750         if (force ||
1751             svcpt->scp_nreqs_active < running - 2)
1752                 return true;
1753
1754         if (svcpt->scp_nreqs_active >= running - 1)
1755                 return false;
1756
1757         return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1758 }
1759
1760 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1761                                          bool force)
1762 {
1763         return ptlrpc_server_allow_normal(svcpt, force) &&
1764                ptlrpc_nrs_req_pending_nolock(svcpt, false);
1765 }
1766
1767 /**
1768  * Returns true if there are requests available in incoming
1769  * request queue for processing and it is allowed to fetch them.
1770  * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1771  * to get reliable result
1772  * \see ptlrpc_server_allow_normal
1773  * \see ptlrpc_server_allow high
1774  */
1775 static inline bool
1776 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1777 {
1778         return ptlrpc_server_high_pending(svcpt, force) ||
1779                ptlrpc_server_normal_pending(svcpt, force);
1780 }
1781
1782 /**
1783  * Fetch a request for processing from queue of unprocessed requests.
1784  * Favors high-priority requests.
1785  * Returns a pointer to fetched request.
1786  */
1787 static struct ptlrpc_request *
1788 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1789 {
1790         struct ptlrpc_request *req = NULL;
1791         ENTRY;
1792
1793         spin_lock(&svcpt->scp_req_lock);
1794 #ifndef __KERNEL__
1795         /* !@%$# liblustre only has 1 thread */
1796         if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1797                 spin_unlock(&svcpt->scp_req_lock);
1798                 RETURN(NULL);
1799         }
1800 #endif
1801
1802         if (ptlrpc_server_high_pending(svcpt, force)) {
1803                 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1804                 if (req != NULL) {
1805                         svcpt->scp_hreq_count++;
1806                         goto got_request;
1807                 }
1808         }
1809
1810         if (ptlrpc_server_normal_pending(svcpt, force)) {
1811                 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1812                 if (req != NULL) {
1813                         svcpt->scp_hreq_count = 0;
1814                         goto got_request;
1815                 }
1816         }
1817
1818         spin_unlock(&svcpt->scp_req_lock);
1819         RETURN(NULL);
1820
1821 got_request:
1822         svcpt->scp_nreqs_active++;
1823         if (req->rq_hp)
1824                 svcpt->scp_nhreqs_active++;
1825
1826         spin_unlock(&svcpt->scp_req_lock);
1827
1828         if (likely(req->rq_export))
1829                 class_export_rpc_inc(req->rq_export);
1830
1831         RETURN(req);
1832 }
1833
1834 /**
1835  * Handle freshly incoming reqs, add to timed early reply list,
1836  * pass on to regular request queue.
1837  * All incoming requests pass through here before getting into
1838  * ptlrpc_server_handle_req later on.
1839  */
1840 static int
1841 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1842                             struct ptlrpc_thread *thread)
1843 {
1844         struct ptlrpc_service   *svc = svcpt->scp_service;
1845         struct ptlrpc_request   *req;
1846         __u32                   deadline;
1847         int                     rc;
1848         ENTRY;
1849
1850         spin_lock(&svcpt->scp_lock);
1851         if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1852                 spin_unlock(&svcpt->scp_lock);
1853                 RETURN(0);
1854         }
1855
1856         req = cfs_list_entry(svcpt->scp_req_incoming.next,
1857                              struct ptlrpc_request, rq_list);
1858         cfs_list_del_init(&req->rq_list);
1859         svcpt->scp_nreqs_incoming--;
1860         /* Consider this still a "queued" request as far as stats are
1861          * concerned */
1862         spin_unlock(&svcpt->scp_lock);
1863
1864         /* go through security check/transform */
1865         rc = sptlrpc_svc_unwrap_request(req);
1866         switch (rc) {
1867         case SECSVC_OK:
1868                 break;
1869         case SECSVC_COMPLETE:
1870                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1871                 goto err_req;
1872         case SECSVC_DROP:
1873                 goto err_req;
1874         default:
1875                 LBUG();
1876         }
1877
1878         /*
1879          * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1880          * redo it wouldn't be harmful.
1881          */
1882         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1883                 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1884                 if (rc != 0) {
1885                         CERROR("error unpacking request: ptl %d from %s "
1886                                "x"LPU64"\n", svc->srv_req_portal,
1887                                libcfs_id2str(req->rq_peer), req->rq_xid);
1888                         goto err_req;
1889                 }
1890         }
1891
1892         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1893         if (rc) {
1894                 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1895                         LPU64"\n", svc->srv_req_portal,
1896                         libcfs_id2str(req->rq_peer), req->rq_xid);
1897                 goto err_req;
1898         }
1899
1900         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1901             lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1902                 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1903                        cfs_fail_val, req->rq_xid);
1904                 goto err_req;
1905         }
1906
1907         rc = -EINVAL;
1908         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1909                 CERROR("wrong packet type received (type=%u) from %s\n",
1910                        lustre_msg_get_type(req->rq_reqmsg),
1911                        libcfs_id2str(req->rq_peer));
1912                 goto err_req;
1913         }
1914
1915         switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1916         case MDS_WRITEPAGE:
1917         case OST_WRITE:
1918                 req->rq_bulk_write = 1;
1919                 break;
1920         case MDS_READPAGE:
1921         case OST_READ:
1922         case MGS_CONFIG_READ:
1923                 req->rq_bulk_read = 1;
1924                 break;
1925         }
1926
1927         CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1928
1929         req->rq_export = class_conn2export(
1930                 lustre_msg_get_handle(req->rq_reqmsg));
1931         if (req->rq_export) {
1932                 rc = ptlrpc_check_req(req);
1933                 if (rc == 0) {
1934                         rc = sptlrpc_target_export_check(req->rq_export, req);
1935                         if (rc)
1936                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1937                                           "illegal security flavor,");
1938                 }
1939
1940                 if (rc)
1941                         goto err_req;
1942                 ptlrpc_update_export_timer(req->rq_export, 0);
1943         }
1944
1945         /* req_in handling should/must be fast */
1946         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1947                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1948                           cfs_time_sub(cfs_time_current_sec(),
1949                                        req->rq_arrival_time.tv_sec));
1950
1951         /* Set rpc server deadline and add it to the timed list */
1952         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1953                     MSGHDR_AT_SUPPORT) ?
1954                    /* The max time the client expects us to take */
1955                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1956         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1957         if (unlikely(deadline == 0)) {
1958                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1959                 goto err_req;
1960         }
1961
1962         req->rq_svc_thread = thread;
1963         if (thread != NULL) {
1964                 /* initialize request session, it is needed for request
1965                  * processing by target */
1966                 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1967                                                        LCT_NOREF);
1968                 if (rc) {
1969                         CERROR("%s: failure to initialize session: rc = %d\n",
1970                                thread->t_name, rc);
1971                         goto err_req;
1972                 }
1973                 req->rq_session.lc_thread = thread;
1974                 lu_context_enter(&req->rq_session);
1975                 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1976         }
1977
1978         ptlrpc_at_add_timed(req);
1979
1980         /* Move it over to the request processing queue */
1981         rc = ptlrpc_server_request_add(svcpt, req);
1982         if (rc)
1983                 GOTO(err_req, rc);
1984
1985         wake_up(&svcpt->scp_waitq);
1986         RETURN(1);
1987
1988 err_req:
1989         ptlrpc_server_finish_request(svcpt, req);
1990
1991         RETURN(1);
1992 }
1993
1994 /**
1995  * Main incoming request handling logic.
1996  * Calls handler function from service to do actual processing.
1997  */
1998 static int
1999 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2000                              struct ptlrpc_thread *thread)
2001 {
2002         struct ptlrpc_service   *svc = svcpt->scp_service;
2003         struct ptlrpc_request   *request;
2004         struct timeval           work_start;
2005         struct timeval           work_end;
2006         long                     timediff;
2007         int                      fail_opc = 0;
2008
2009         ENTRY;
2010
2011         request = ptlrpc_server_request_get(svcpt, false);
2012         if (request == NULL)
2013                 RETURN(0);
2014
2015         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2016                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2017         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2018                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2019
2020         if (unlikely(fail_opc)) {
2021                 if (request->rq_export && request->rq_ops)
2022                         OBD_FAIL_TIMEOUT(fail_opc, 4);
2023         }
2024
2025         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2026
2027         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2028                 libcfs_debug_dumplog();
2029
2030         do_gettimeofday(&work_start);
2031         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
2032         if (likely(svc->srv_stats != NULL)) {
2033                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2034                                     timediff);
2035                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2036                                     svcpt->scp_nreqs_incoming);
2037                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2038                                     svcpt->scp_nreqs_active);
2039                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2040                                     at_get(&svcpt->scp_at_estimate));
2041         }
2042
2043         if (likely(request->rq_export)) {
2044                 if (unlikely(ptlrpc_check_req(request)))
2045                         goto put_conn;
2046                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
2047         }
2048
2049         /* Discard requests queued for longer than the deadline.
2050            The deadline is increased if we send an early reply. */
2051         if (cfs_time_current_sec() > request->rq_deadline) {
2052                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
2053                           ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
2054                           libcfs_id2str(request->rq_peer),
2055                           cfs_time_sub(request->rq_deadline,
2056                           request->rq_arrival_time.tv_sec),
2057                           cfs_time_sub(cfs_time_current_sec(),
2058                           request->rq_deadline));
2059                 goto put_conn;
2060         }
2061
2062         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2063                "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
2064                (request->rq_export ?
2065                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2066                (request->rq_export ?
2067                 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2068                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2069                libcfs_id2str(request->rq_peer),
2070                lustre_msg_get_opc(request->rq_reqmsg));
2071
2072         if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2073                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2074
2075         CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
2076
2077         /* re-assign request and sesson thread to the current one */
2078         request->rq_svc_thread = thread;
2079         if (thread != NULL) {
2080                 LASSERT(request->rq_session.lc_thread != NULL);
2081                 request->rq_session.lc_thread = thread;
2082                 request->rq_session.lc_cookie = 0x55;
2083                 thread->t_env->le_ses = &request->rq_session;
2084         }
2085         svc->srv_ops.so_req_handler(request);
2086
2087         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2088
2089 put_conn:
2090         if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2091                      DEBUG_REQ(D_WARNING, request, "Request took longer "
2092                                "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2093                                " client may timeout.",
2094                                cfs_time_sub(request->rq_deadline,
2095                                             request->rq_arrival_time.tv_sec),
2096                                cfs_time_sub(cfs_time_current_sec(),
2097                                             request->rq_deadline));
2098         }
2099
2100         do_gettimeofday(&work_end);
2101         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2102         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2103                "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2104                "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2105                 current_comm(),
2106                 (request->rq_export ?
2107                  (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2108                 (request->rq_export ?
2109                  cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2110                 lustre_msg_get_status(request->rq_reqmsg),
2111                 request->rq_xid,
2112                 libcfs_id2str(request->rq_peer),
2113                 lustre_msg_get_opc(request->rq_reqmsg),
2114                 timediff,
2115                 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2116                 (request->rq_repmsg ?
2117                  lustre_msg_get_transno(request->rq_repmsg) :
2118                  request->rq_transno),
2119                 request->rq_status,
2120                 (request->rq_repmsg ?
2121                  lustre_msg_get_status(request->rq_repmsg) : -999));
2122         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2123                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2124                 int opc = opcode_offset(op);
2125                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2126                         LASSERT(opc < LUSTRE_MAX_OPCODES);
2127                         lprocfs_counter_add(svc->srv_stats,
2128                                             opc + EXTRA_MAX_OPCODES,
2129                                             timediff);
2130                 }
2131         }
2132         if (unlikely(request->rq_early_count)) {
2133                 DEBUG_REQ(D_ADAPTTO, request,
2134                           "sent %d early replies before finishing in "
2135                           CFS_DURATION_T"s",
2136                           request->rq_early_count,
2137                           cfs_time_sub(work_end.tv_sec,
2138                           request->rq_arrival_time.tv_sec));
2139         }
2140
2141         ptlrpc_server_finish_active_request(svcpt, request);
2142
2143         RETURN(1);
2144 }
2145
2146 /**
2147  * An internal function to process a single reply state object.
2148  */
2149 static int
2150 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2151 {
2152         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2153         struct ptlrpc_service     *svc = svcpt->scp_service;
2154         struct obd_export         *exp;
2155         int                        nlocks;
2156         int                        been_handled;
2157         ENTRY;
2158
2159         exp = rs->rs_export;
2160
2161         LASSERT (rs->rs_difficult);
2162         LASSERT (rs->rs_scheduled);
2163         LASSERT (cfs_list_empty(&rs->rs_list));
2164
2165         spin_lock(&exp->exp_lock);
2166         /* Noop if removed already */
2167         cfs_list_del_init (&rs->rs_exp_list);
2168         spin_unlock(&exp->exp_lock);
2169
2170         /* The disk commit callback holds exp_uncommitted_replies_lock while it
2171          * iterates over newly committed replies, removing them from
2172          * exp_uncommitted_replies.  It then drops this lock and schedules the
2173          * replies it found for handling here.
2174          *
2175          * We can avoid contention for exp_uncommitted_replies_lock between the
2176          * HRT threads and further commit callbacks by checking rs_committed
2177          * which is set in the commit callback while it holds both
2178          * rs_lock and exp_uncommitted_reples.
2179          *
2180          * If we see rs_committed clear, the commit callback _may_ not have
2181          * handled this reply yet and we race with it to grab
2182          * exp_uncommitted_replies_lock before removing the reply from
2183          * exp_uncommitted_replies.  Note that if we lose the race and the
2184          * reply has already been removed, list_del_init() is a noop.
2185          *
2186          * If we see rs_committed set, we know the commit callback is handling,
2187          * or has handled this reply since store reordering might allow us to
2188          * see rs_committed set out of sequence.  But since this is done
2189          * holding rs_lock, we can be sure it has all completed once we hold
2190          * rs_lock, which we do right next.
2191          */
2192         if (!rs->rs_committed) {
2193                 spin_lock(&exp->exp_uncommitted_replies_lock);
2194                 cfs_list_del_init(&rs->rs_obd_list);
2195                 spin_unlock(&exp->exp_uncommitted_replies_lock);
2196         }
2197
2198         spin_lock(&rs->rs_lock);
2199
2200         been_handled = rs->rs_handled;
2201         rs->rs_handled = 1;
2202
2203         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
2204         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
2205
2206         if (nlocks == 0 && !been_handled) {
2207                 /* If we see this, we should already have seen the warning
2208                  * in mds_steal_ack_locks()  */
2209                 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2210                        " o%d NID %s\n",
2211                        rs,
2212                        rs->rs_xid, rs->rs_transno, rs->rs_opc,
2213                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
2214         }
2215
2216         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2217                 spin_unlock(&rs->rs_lock);
2218
2219                 if (!been_handled && rs->rs_on_net) {
2220                         LNetMDUnlink(rs->rs_md_h);
2221                         /* Ignore return code; we're racing with completion */
2222                 }
2223
2224                 while (nlocks-- > 0)
2225                         ldlm_lock_decref(&rs->rs_locks[nlocks],
2226                                          rs->rs_modes[nlocks]);
2227
2228                 spin_lock(&rs->rs_lock);
2229         }
2230
2231         rs->rs_scheduled = 0;
2232
2233         if (!rs->rs_on_net) {
2234                 /* Off the net */
2235                 spin_unlock(&rs->rs_lock);
2236
2237                 class_export_put (exp);
2238                 rs->rs_export = NULL;
2239                 ptlrpc_rs_decref (rs);
2240                 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2241                     svc->srv_is_stopping)
2242                         wake_up_all(&svcpt->scp_waitq);
2243                 RETURN(1);
2244         }
2245
2246         /* still on the net; callback will schedule */
2247         spin_unlock(&rs->rs_lock);
2248         RETURN(1);
2249 }
2250
2251 #ifndef __KERNEL__
2252
2253 /**
2254  * Check whether given service has a reply available for processing
2255  * and process it.
2256  *
2257  * \param svc a ptlrpc service
2258  * \retval 0 no replies processed
2259  * \retval 1 one reply processed
2260  */
2261 static int
2262 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2263 {
2264         struct ptlrpc_reply_state *rs = NULL;
2265         ENTRY;
2266
2267         spin_lock(&svcpt->scp_rep_lock);
2268         if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2269                 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2270                                     struct ptlrpc_reply_state,
2271                                     rs_list);
2272                 cfs_list_del_init(&rs->rs_list);
2273         }
2274         spin_unlock(&svcpt->scp_rep_lock);
2275         if (rs != NULL)
2276                 ptlrpc_handle_rs(rs);
2277         RETURN(rs != NULL);
2278 }
2279
2280 /* FIXME make use of timeout later */
2281 int
2282 liblustre_check_services (void *arg)
2283 {
2284         int  did_something = 0;
2285         int  rc;
2286         cfs_list_t *tmp, *nxt;
2287         ENTRY;
2288
2289         /* I'm relying on being single threaded, not to have to lock
2290          * ptlrpc_all_services etc */
2291         cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2292                 struct ptlrpc_service *svc =
2293                         cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2294                 struct ptlrpc_service_part *svcpt;
2295
2296                 LASSERT(svc->srv_ncpts == 1);
2297                 svcpt = svc->srv_parts[0];
2298
2299                 if (svcpt->scp_nthrs_running != 0)     /* I've recursed */
2300                         continue;
2301
2302                 /* service threads can block for bulk, so this limits us
2303                  * (arbitrarily) to recursing 1 stack frame per service.
2304                  * Note that the problem with recursion is that we have to
2305                  * unwind completely before our caller can resume. */
2306
2307                 svcpt->scp_nthrs_running++;
2308
2309                 do {
2310                         rc = ptlrpc_server_handle_req_in(svcpt, NULL);
2311                         rc |= ptlrpc_server_handle_reply(svcpt);
2312                         rc |= ptlrpc_at_check_timed(svcpt);
2313                         rc |= ptlrpc_server_handle_request(svcpt, NULL);
2314                         rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2315                         did_something |= rc;
2316                 } while (rc);
2317
2318                 svcpt->scp_nthrs_running--;
2319         }
2320
2321         RETURN(did_something);
2322 }
2323 #define ptlrpc_stop_all_threads(s) do {} while (0)
2324
2325 #else /* __KERNEL__ */
2326
2327 static void
2328 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2329 {
2330         int avail = svcpt->scp_nrqbds_posted;
2331         int low_water = test_req_buffer_pressure ? 0 :
2332                         svcpt->scp_service->srv_nbuf_per_group / 2;
2333
2334         /* NB I'm not locking; just looking. */
2335
2336         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2337          * allowed the request history to grow out of control.  We could put a
2338          * sanity check on that here and cull some history if we need the
2339          * space. */
2340
2341         if (avail <= low_water)
2342                 ptlrpc_grow_req_bufs(svcpt, 1);
2343
2344         if (svcpt->scp_service->srv_stats) {
2345                 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2346                                     PTLRPC_REQBUF_AVAIL_CNTR, avail);
2347         }
2348 }
2349
2350 static int
2351 ptlrpc_retry_rqbds(void *arg)
2352 {
2353         struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2354
2355         svcpt->scp_rqbd_timeout = 0;
2356         return -ETIMEDOUT;
2357 }
2358
2359 static inline int
2360 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2361 {
2362         return svcpt->scp_nreqs_active <
2363                svcpt->scp_nthrs_running - 1 -
2364                (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2365 }
2366
2367 /**
2368  * allowed to create more threads
2369  * user can call it w/o any lock but need to hold
2370  * ptlrpc_service_part::scp_lock to get reliable result
2371  */
2372 static inline int
2373 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2374 {
2375         return svcpt->scp_nthrs_running +
2376                svcpt->scp_nthrs_starting <
2377                svcpt->scp_service->srv_nthrs_cpt_limit;
2378 }
2379
2380 /**
2381  * too many requests and allowed to create more threads
2382  */
2383 static inline int
2384 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2385 {
2386         return !ptlrpc_threads_enough(svcpt) &&
2387                 ptlrpc_threads_increasable(svcpt);
2388 }
2389
2390 static inline int
2391 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2392 {
2393         return thread_is_stopping(thread) ||
2394                thread->t_svcpt->scp_service->srv_is_stopping;
2395 }
2396
2397 static inline int
2398 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2399 {
2400         return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2401                svcpt->scp_rqbd_timeout == 0;
2402 }
2403
2404 static inline int
2405 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2406 {
2407         return svcpt->scp_at_check;
2408 }
2409
2410 /**
2411  * requests wait on preprocessing
2412  * user can call it w/o any lock but need to hold
2413  * ptlrpc_service_part::scp_lock to get reliable result
2414  */
2415 static inline int
2416 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2417 {
2418         return !cfs_list_empty(&svcpt->scp_req_incoming);
2419 }
2420
2421 static __attribute__((__noinline__)) int
2422 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2423                   struct ptlrpc_thread *thread)
2424 {
2425         /* Don't exit while there are replies to be handled */
2426         struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2427                                              ptlrpc_retry_rqbds, svcpt);
2428
2429         lc_watchdog_disable(thread->t_watchdog);
2430
2431         cond_resched();
2432
2433         l_wait_event_exclusive_head(svcpt->scp_waitq,
2434                                 ptlrpc_thread_stopping(thread) ||
2435                                 ptlrpc_server_request_incoming(svcpt) ||
2436                                 ptlrpc_server_request_pending(svcpt, false) ||
2437                                 ptlrpc_rqbd_pending(svcpt) ||
2438                                 ptlrpc_at_check(svcpt), &lwi);
2439
2440         if (ptlrpc_thread_stopping(thread))
2441                 return -EINTR;
2442
2443         lc_watchdog_touch(thread->t_watchdog,
2444                           ptlrpc_server_get_timeout(svcpt));
2445         return 0;
2446 }
2447
2448 /**
2449  * Main thread body for service threads.
2450  * Waits in a loop waiting for new requests to process to appear.
2451  * Every time an incoming requests is added to its queue, a waitq
2452  * is woken up and one of the threads will handle it.
2453  */
2454 static int ptlrpc_main(void *arg)
2455 {
2456         struct ptlrpc_thread            *thread = (struct ptlrpc_thread *)arg;
2457         struct ptlrpc_service_part      *svcpt = thread->t_svcpt;
2458         struct ptlrpc_service           *svc = svcpt->scp_service;
2459         struct ptlrpc_reply_state       *rs;
2460 #ifdef WITH_GROUP_INFO
2461         struct group_info *ginfo = NULL;
2462 #endif
2463         struct lu_env *env;
2464         int counter = 0, rc = 0;
2465         ENTRY;
2466
2467         thread->t_pid = current_pid();
2468         unshare_fs_struct();
2469
2470         /* NB: we will call cfs_cpt_bind() for all threads, because we
2471          * might want to run lustre server only on a subset of system CPUs,
2472          * in that case ->scp_cpt is CFS_CPT_ANY */
2473         rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2474         if (rc != 0) {
2475                 CWARN("%s: failed to bind %s on CPT %d\n",
2476                       svc->srv_name, thread->t_name, svcpt->scp_cpt);
2477         }
2478
2479 #ifdef WITH_GROUP_INFO
2480         ginfo = groups_alloc(0);
2481         if (!ginfo) {
2482                 rc = -ENOMEM;
2483                 goto out;
2484         }
2485
2486         set_current_groups(ginfo);
2487         put_group_info(ginfo);
2488 #endif
2489
2490         if (svc->srv_ops.so_thr_init != NULL) {
2491                 rc = svc->srv_ops.so_thr_init(thread);
2492                 if (rc)
2493                         goto out;
2494         }
2495
2496         OBD_ALLOC_PTR(env);
2497         if (env == NULL) {
2498                 rc = -ENOMEM;
2499                 goto out_srv_fini;
2500         }
2501
2502         rc = lu_context_init(&env->le_ctx,
2503                              svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2504         if (rc)
2505                 goto out_srv_fini;
2506
2507         thread->t_env = env;
2508         env->le_ctx.lc_thread = thread;
2509         env->le_ctx.lc_cookie = 0x6;
2510
2511         while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2512                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2513                 if (rc >= 0)
2514                         continue;
2515
2516                 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2517                         svc->srv_name, svcpt->scp_cpt, rc);
2518                 goto out_srv_fini;
2519         }
2520
2521         /* Alloc reply state structure for this one */
2522         OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2523         if (!rs) {
2524                 rc = -ENOMEM;
2525                 goto out_srv_fini;
2526         }
2527
2528         spin_lock(&svcpt->scp_lock);
2529
2530         LASSERT(thread_is_starting(thread));
2531         thread_clear_flags(thread, SVC_STARTING);
2532
2533         LASSERT(svcpt->scp_nthrs_starting == 1);
2534         svcpt->scp_nthrs_starting--;
2535
2536         /* SVC_STOPPING may already be set here if someone else is trying
2537          * to stop the service while this new thread has been dynamically
2538          * forked. We still set SVC_RUNNING to let our creator know that
2539          * we are now running, however we will exit as soon as possible */
2540         thread_add_flags(thread, SVC_RUNNING);
2541         svcpt->scp_nthrs_running++;
2542         spin_unlock(&svcpt->scp_lock);
2543
2544         /* wake up our creator in case he's still waiting. */
2545         wake_up(&thread->t_ctl_waitq);
2546
2547         thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2548                                              NULL, NULL);
2549
2550         spin_lock(&svcpt->scp_rep_lock);
2551         cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2552         wake_up(&svcpt->scp_rep_waitq);
2553         spin_unlock(&svcpt->scp_rep_lock);
2554
2555         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2556                svcpt->scp_nthrs_running);
2557
2558         /* XXX maintain a list of all managed devices: insert here */
2559         while (!ptlrpc_thread_stopping(thread)) {
2560                 if (ptlrpc_wait_event(svcpt, thread))
2561                         break;
2562
2563                 ptlrpc_check_rqbd_pool(svcpt);
2564
2565                 if (ptlrpc_threads_need_create(svcpt)) {
2566                         /* Ignore return code - we tried... */
2567                         ptlrpc_start_thread(svcpt, 0);
2568                 }
2569
2570                 /* Process all incoming reqs before handling any */
2571                 if (ptlrpc_server_request_incoming(svcpt)) {
2572                         lu_context_enter(&env->le_ctx);
2573                         env->le_ses = NULL;
2574                         ptlrpc_server_handle_req_in(svcpt, thread);
2575                         lu_context_exit(&env->le_ctx);
2576
2577                         /* but limit ourselves in case of flood */
2578                         if (counter++ < 100)
2579                                 continue;
2580                         counter = 0;
2581                 }
2582
2583                 if (ptlrpc_at_check(svcpt))
2584                         ptlrpc_at_check_timed(svcpt);
2585
2586                 if (ptlrpc_server_request_pending(svcpt, false)) {
2587                         lu_context_enter(&env->le_ctx);
2588                         ptlrpc_server_handle_request(svcpt, thread);
2589                         lu_context_exit(&env->le_ctx);
2590                 }
2591
2592                 if (ptlrpc_rqbd_pending(svcpt) &&
2593                     ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2594                         /* I just failed to repost request buffers.
2595                          * Wait for a timeout (unless something else
2596                          * happens) before I try again */
2597                         svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2598                         CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2599                                svcpt->scp_nrqbds_posted);
2600                 }
2601         }
2602
2603         lc_watchdog_delete(thread->t_watchdog);
2604         thread->t_watchdog = NULL;
2605
2606 out_srv_fini:
2607         /*
2608          * deconstruct service specific state created by ptlrpc_start_thread()
2609          */
2610         if (svc->srv_ops.so_thr_done != NULL)
2611                 svc->srv_ops.so_thr_done(thread);
2612
2613         if (env != NULL) {
2614                 lu_context_fini(&env->le_ctx);
2615                 OBD_FREE_PTR(env);
2616         }
2617 out:
2618         CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2619                thread, thread->t_pid, thread->t_id, rc);
2620
2621         spin_lock(&svcpt->scp_lock);
2622         if (thread_test_and_clear_flags(thread, SVC_STARTING))
2623                 svcpt->scp_nthrs_starting--;
2624
2625         if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2626                 /* must know immediately */
2627                 svcpt->scp_nthrs_running--;
2628         }
2629
2630         thread->t_id = rc;
2631         thread_add_flags(thread, SVC_STOPPED);
2632
2633         wake_up(&thread->t_ctl_waitq);
2634         spin_unlock(&svcpt->scp_lock);
2635
2636         return rc;
2637 }
2638
2639 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2640                           cfs_list_t *replies)
2641 {
2642         int result;
2643
2644         spin_lock(&hrt->hrt_lock);
2645
2646         cfs_list_splice_init(&hrt->hrt_queue, replies);
2647         result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2648
2649         spin_unlock(&hrt->hrt_lock);
2650         return result;
2651 }
2652
2653 /**
2654  * Main body of "handle reply" function.
2655  * It processes acked reply states
2656  */
2657 static int ptlrpc_hr_main(void *arg)
2658 {
2659         struct ptlrpc_hr_thread         *hrt = (struct ptlrpc_hr_thread *)arg;
2660         struct ptlrpc_hr_partition      *hrp = hrt->hrt_partition;
2661         CFS_LIST_HEAD                   (replies);
2662         char                            threadname[20];
2663         int                             rc;
2664
2665         snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2666                  hrp->hrp_cpt, hrt->hrt_id);
2667         unshare_fs_struct();
2668
2669         rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2670         if (rc != 0) {
2671                 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2672                       threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2673         }
2674
2675         cfs_atomic_inc(&hrp->hrp_nstarted);
2676         wake_up(&ptlrpc_hr.hr_waitq);
2677
2678         while (!ptlrpc_hr.hr_stopping) {
2679                 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2680
2681                 while (!cfs_list_empty(&replies)) {
2682                         struct ptlrpc_reply_state *rs;
2683
2684                         rs = cfs_list_entry(replies.prev,
2685                                             struct ptlrpc_reply_state,
2686                                             rs_list);
2687                         cfs_list_del_init(&rs->rs_list);
2688                         ptlrpc_handle_rs(rs);
2689                 }
2690         }
2691
2692         cfs_atomic_inc(&hrp->hrp_nstopped);
2693         wake_up(&ptlrpc_hr.hr_waitq);
2694
2695         return 0;
2696 }
2697
2698 static void ptlrpc_stop_hr_threads(void)
2699 {
2700         struct ptlrpc_hr_partition      *hrp;
2701         int                             i;
2702         int                             j;
2703
2704         ptlrpc_hr.hr_stopping = 1;
2705
2706         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2707                 if (hrp->hrp_thrs == NULL)
2708                         continue; /* uninitialized */
2709                 for (j = 0; j < hrp->hrp_nthrs; j++)
2710                         wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2711         }
2712
2713         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2714                 if (hrp->hrp_thrs == NULL)
2715                         continue; /* uninitialized */
2716                 wait_event(ptlrpc_hr.hr_waitq,
2717                                cfs_atomic_read(&hrp->hrp_nstopped) ==
2718                                cfs_atomic_read(&hrp->hrp_nstarted));
2719         }
2720 }
2721
2722 static int ptlrpc_start_hr_threads(void)
2723 {
2724         struct ptlrpc_hr_partition      *hrp;
2725         int                             i;
2726         int                             j;
2727         ENTRY;
2728
2729         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2730                 int     rc = 0;
2731
2732                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2733                         struct  ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2734                         rc = PTR_ERR(kthread_run(ptlrpc_hr_main,
2735                                                  &hrp->hrp_thrs[j],
2736                                                  "ptlrpc_hr%02d_%03d",
2737                                                  hrp->hrp_cpt,
2738                                                  hrt->hrt_id));
2739                         if (IS_ERR_VALUE(rc))
2740                                 break;
2741                 }
2742                 wait_event(ptlrpc_hr.hr_waitq,
2743                                cfs_atomic_read(&hrp->hrp_nstarted) == j);
2744                 if (!IS_ERR_VALUE(rc))
2745                         continue;
2746
2747                 CERROR("Reply handling thread %d:%d Failed on starting: "
2748                        "rc = %d\n", i, j, rc);
2749                 ptlrpc_stop_hr_threads();
2750                 RETURN(rc);
2751         }
2752         RETURN(0);
2753 }
2754
2755 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2756 {
2757         struct l_wait_info      lwi = { 0 };
2758         struct ptlrpc_thread    *thread;
2759         CFS_LIST_HEAD           (zombie);
2760
2761         ENTRY;
2762
2763         CDEBUG(D_INFO, "Stopping threads for service %s\n",
2764                svcpt->scp_service->srv_name);
2765
2766         spin_lock(&svcpt->scp_lock);
2767         /* let the thread know that we would like it to stop asap */
2768         list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2769                 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2770                        svcpt->scp_service->srv_thread_name, thread->t_id);
2771                 thread_add_flags(thread, SVC_STOPPING);
2772         }
2773
2774         wake_up_all(&svcpt->scp_waitq);
2775
2776         while (!cfs_list_empty(&svcpt->scp_threads)) {
2777                 thread = cfs_list_entry(svcpt->scp_threads.next,
2778                                         struct ptlrpc_thread, t_link);
2779                 if (thread_is_stopped(thread)) {
2780                         cfs_list_del(&thread->t_link);
2781                         cfs_list_add(&thread->t_link, &zombie);
2782                         continue;
2783                 }
2784                 spin_unlock(&svcpt->scp_lock);
2785
2786                 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2787                        svcpt->scp_service->srv_thread_name, thread->t_id);
2788                 l_wait_event(thread->t_ctl_waitq,
2789                              thread_is_stopped(thread), &lwi);
2790
2791                 spin_lock(&svcpt->scp_lock);
2792         }
2793
2794         spin_unlock(&svcpt->scp_lock);
2795
2796         while (!cfs_list_empty(&zombie)) {
2797                 thread = cfs_list_entry(zombie.next,
2798                                         struct ptlrpc_thread, t_link);
2799                 cfs_list_del(&thread->t_link);
2800                 OBD_FREE_PTR(thread);
2801         }
2802         EXIT;
2803 }
2804
2805 /**
2806  * Stops all threads of a particular service \a svc
2807  */
2808 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2809 {
2810         struct ptlrpc_service_part *svcpt;
2811         int                        i;
2812         ENTRY;
2813
2814         ptlrpc_service_for_each_part(svcpt, i, svc) {
2815                 if (svcpt->scp_service != NULL)
2816                         ptlrpc_svcpt_stop_threads(svcpt);
2817         }
2818
2819         EXIT;
2820 }
2821 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2822
2823 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2824 {
2825         int     rc = 0;
2826         int     i;
2827         int     j;
2828         ENTRY;
2829
2830         /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2831         LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2832
2833         for (i = 0; i < svc->srv_ncpts; i++) {
2834                 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2835                         rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2836                         if (rc == 0)
2837                                 continue;
2838
2839                         if (rc != -EMFILE)
2840                                 goto failed;
2841                         /* We have enough threads, don't start more. b=15759 */
2842                         break;
2843                 }
2844         }
2845
2846         RETURN(0);
2847  failed:
2848         CERROR("cannot start %s thread #%d_%d: rc %d\n",
2849                svc->srv_thread_name, i, j, rc);
2850         ptlrpc_stop_all_threads(svc);
2851         RETURN(rc);
2852 }
2853 EXPORT_SYMBOL(ptlrpc_start_threads);
2854
2855 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2856 {
2857         struct l_wait_info      lwi = { 0 };
2858         struct ptlrpc_thread    *thread;
2859         struct ptlrpc_service   *svc;
2860         int                     rc;
2861         ENTRY;
2862
2863         LASSERT(svcpt != NULL);
2864
2865         svc = svcpt->scp_service;
2866
2867         CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2868                svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2869                svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2870
2871  again:
2872         if (unlikely(svc->srv_is_stopping))
2873                 RETURN(-ESRCH);
2874
2875         if (!ptlrpc_threads_increasable(svcpt) ||
2876             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2877              svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2878                 RETURN(-EMFILE);
2879
2880         OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2881         if (thread == NULL)
2882                 RETURN(-ENOMEM);
2883         init_waitqueue_head(&thread->t_ctl_waitq);
2884
2885         spin_lock(&svcpt->scp_lock);
2886         if (!ptlrpc_threads_increasable(svcpt)) {
2887                 spin_unlock(&svcpt->scp_lock);
2888                 OBD_FREE_PTR(thread);
2889                 RETURN(-EMFILE);
2890         }
2891
2892         if (svcpt->scp_nthrs_starting != 0) {
2893                 /* serialize starting because some modules (obdfilter)
2894                  * might require unique and contiguous t_id */
2895                 LASSERT(svcpt->scp_nthrs_starting == 1);
2896                 spin_unlock(&svcpt->scp_lock);
2897                 OBD_FREE_PTR(thread);
2898                 if (wait) {
2899                         CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2900                                svc->srv_thread_name, svcpt->scp_thr_nextid);
2901                         schedule();
2902                         goto again;
2903                 }
2904
2905                 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2906                        svc->srv_thread_name, svcpt->scp_thr_nextid);
2907                 RETURN(-EAGAIN);
2908         }
2909
2910         svcpt->scp_nthrs_starting++;
2911         thread->t_id = svcpt->scp_thr_nextid++;
2912         thread_add_flags(thread, SVC_STARTING);
2913         thread->t_svcpt = svcpt;
2914
2915         cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2916         spin_unlock(&svcpt->scp_lock);
2917
2918         if (svcpt->scp_cpt >= 0) {
2919                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2920                          svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2921         } else {
2922                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2923                          svc->srv_thread_name, thread->t_id);
2924         }
2925
2926         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2927         rc = PTR_ERR(kthread_run(ptlrpc_main, thread, thread->t_name));
2928         if (IS_ERR_VALUE(rc)) {
2929                 CERROR("cannot start thread '%s': rc %d\n",
2930                        thread->t_name, rc);
2931                 spin_lock(&svcpt->scp_lock);
2932                 --svcpt->scp_nthrs_starting;
2933                 if (thread_is_stopping(thread)) {
2934                         /* this ptlrpc_thread is being hanled
2935                          * by ptlrpc_svcpt_stop_threads now
2936                          */
2937                         thread_add_flags(thread, SVC_STOPPED);
2938                         wake_up(&thread->t_ctl_waitq);
2939                         spin_unlock(&svcpt->scp_lock);
2940                 } else {
2941                         cfs_list_del(&thread->t_link);
2942                         spin_unlock(&svcpt->scp_lock);
2943                         OBD_FREE_PTR(thread);
2944                 }
2945                 RETURN(rc);
2946         }
2947
2948         if (!wait)
2949                 RETURN(0);
2950
2951         l_wait_event(thread->t_ctl_waitq,
2952                      thread_is_running(thread) || thread_is_stopped(thread),
2953                      &lwi);
2954
2955         rc = thread_is_stopped(thread) ? thread->t_id : 0;
2956         RETURN(rc);
2957 }
2958
2959 int ptlrpc_hr_init(void)
2960 {
2961         struct ptlrpc_hr_partition      *hrp;
2962         struct ptlrpc_hr_thread         *hrt;
2963         int                             rc;
2964         int                             i;
2965         int                             j;
2966         ENTRY;
2967
2968         memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2969         ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2970
2971         ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2972                                                    sizeof(*hrp));
2973         if (ptlrpc_hr.hr_partitions == NULL)
2974                 RETURN(-ENOMEM);
2975
2976         init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2977
2978         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2979                 hrp->hrp_cpt = i;
2980
2981                 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2982                 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2983
2984                 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2985                 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2986
2987                 LASSERT(hrp->hrp_nthrs > 0);
2988                 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2989                               hrp->hrp_nthrs * sizeof(*hrt));
2990                 if (hrp->hrp_thrs == NULL)
2991                         GOTO(out, rc = -ENOMEM);
2992
2993                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2994                         hrt = &hrp->hrp_thrs[j];
2995
2996                         hrt->hrt_id = j;
2997                         hrt->hrt_partition = hrp;
2998                         init_waitqueue_head(&hrt->hrt_waitq);
2999                         spin_lock_init(&hrt->hrt_lock);
3000                         CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
3001                 }
3002         }
3003
3004         rc = ptlrpc_start_hr_threads();
3005 out:
3006         if (rc != 0)
3007                 ptlrpc_hr_fini();
3008         RETURN(rc);
3009 }
3010
3011 void ptlrpc_hr_fini(void)
3012 {
3013         struct ptlrpc_hr_partition      *hrp;
3014         int                             i;
3015
3016         if (ptlrpc_hr.hr_partitions == NULL)
3017                 return;
3018
3019         ptlrpc_stop_hr_threads();
3020
3021         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3022                 if (hrp->hrp_thrs != NULL) {
3023                         OBD_FREE(hrp->hrp_thrs,
3024                                  hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3025                 }
3026         }
3027
3028         cfs_percpt_free(ptlrpc_hr.hr_partitions);
3029         ptlrpc_hr.hr_partitions = NULL;
3030 }
3031
3032 #endif /* __KERNEL__ */
3033
3034 /**
3035  * Wait until all already scheduled replies are processed.
3036  */
3037 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3038 {
3039         while (1) {
3040                 int rc;
3041                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3042                                                      NULL, NULL);
3043
3044                 rc = l_wait_event(svcpt->scp_waitq,
3045                      cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3046                 if (rc == 0)
3047                         break;
3048                 CWARN("Unexpectedly long timeout %s %p\n",
3049                       svcpt->scp_service->srv_name, svcpt->scp_service);
3050         }
3051 }
3052
3053 static void
3054 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3055 {
3056         struct ptlrpc_service_part      *svcpt;
3057         int                             i;
3058
3059         /* early disarm AT timer... */
3060         ptlrpc_service_for_each_part(svcpt, i, svc) {
3061                 if (svcpt->scp_service != NULL)
3062                         cfs_timer_disarm(&svcpt->scp_at_timer);
3063         }
3064 }
3065
3066 static void
3067 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3068 {
3069         struct ptlrpc_service_part        *svcpt;
3070         struct ptlrpc_request_buffer_desc *rqbd;
3071         struct l_wait_info                lwi;
3072         int                               rc;
3073         int                               i;
3074
3075         /* All history will be culled when the next request buffer is
3076          * freed in ptlrpc_service_purge_all() */
3077         svc->srv_hist_nrqbds_cpt_max = 0;
3078
3079         rc = LNetClearLazyPortal(svc->srv_req_portal);
3080         LASSERT(rc == 0);
3081
3082         ptlrpc_service_for_each_part(svcpt, i, svc) {
3083                 if (svcpt->scp_service == NULL)
3084                         break;
3085
3086                 /* Unlink all the request buffers.  This forces a 'final'
3087                  * event with its 'unlink' flag set for each posted rqbd */
3088                 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3089                                         rqbd_list) {
3090                         rc = LNetMDUnlink(rqbd->rqbd_md_h);
3091                         LASSERT(rc == 0 || rc == -ENOENT);
3092                 }
3093         }
3094
3095         ptlrpc_service_for_each_part(svcpt, i, svc) {
3096                 if (svcpt->scp_service == NULL)
3097                         break;
3098
3099                 /* Wait for the network to release any buffers
3100                  * it's currently filling */
3101                 spin_lock(&svcpt->scp_lock);
3102                 while (svcpt->scp_nrqbds_posted != 0) {
3103                         spin_unlock(&svcpt->scp_lock);
3104                         /* Network access will complete in finite time but
3105                          * the HUGE timeout lets us CWARN for visibility
3106                          * of sluggish NALs */
3107                         lwi = LWI_TIMEOUT_INTERVAL(
3108                                         cfs_time_seconds(LONG_UNLINK),
3109                                         cfs_time_seconds(1), NULL, NULL);
3110                         rc = l_wait_event(svcpt->scp_waitq,
3111                                           svcpt->scp_nrqbds_posted == 0, &lwi);
3112                         if (rc == -ETIMEDOUT) {
3113                                 CWARN("Service %s waiting for "
3114                                       "request buffers\n",
3115                                       svcpt->scp_service->srv_name);
3116                         }
3117                         spin_lock(&svcpt->scp_lock);
3118                 }
3119                 spin_unlock(&svcpt->scp_lock);
3120         }
3121 }
3122
3123 static void
3124 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3125 {
3126         struct ptlrpc_service_part              *svcpt;
3127         struct ptlrpc_request_buffer_desc       *rqbd;
3128         struct ptlrpc_request                   *req;
3129         struct ptlrpc_reply_state               *rs;
3130         int                                     i;
3131
3132         ptlrpc_service_for_each_part(svcpt, i, svc) {
3133                 if (svcpt->scp_service == NULL)
3134                         break;
3135
3136                 spin_lock(&svcpt->scp_rep_lock);
3137                 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
3138                         rs = cfs_list_entry(svcpt->scp_rep_active.next,
3139                                             struct ptlrpc_reply_state, rs_list);
3140                         spin_lock(&rs->rs_lock);
3141                         ptlrpc_schedule_difficult_reply(rs);
3142                         spin_unlock(&rs->rs_lock);
3143                 }
3144                 spin_unlock(&svcpt->scp_rep_lock);
3145
3146                 /* purge the request queue.  NB No new replies (rqbds
3147                  * all unlinked) and no service threads, so I'm the only
3148                  * thread noodling the request queue now */
3149                 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
3150                         req = cfs_list_entry(svcpt->scp_req_incoming.next,
3151                                              struct ptlrpc_request, rq_list);
3152
3153                         cfs_list_del(&req->rq_list);
3154                         svcpt->scp_nreqs_incoming--;
3155                         ptlrpc_server_finish_request(svcpt, req);
3156                 }
3157
3158                 while (ptlrpc_server_request_pending(svcpt, true)) {
3159                         req = ptlrpc_server_request_get(svcpt, true);
3160                         ptlrpc_server_finish_active_request(svcpt, req);
3161                 }
3162
3163                 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3164                 LASSERT(svcpt->scp_nreqs_incoming == 0);
3165                 LASSERT(svcpt->scp_nreqs_active == 0);
3166                 /* history should have been culled by
3167                  * ptlrpc_server_finish_request */
3168                 LASSERT(svcpt->scp_hist_nrqbds == 0);
3169
3170                 /* Now free all the request buffers since nothing
3171                  * references them any more... */
3172
3173                 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3174                         rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3175                                               struct ptlrpc_request_buffer_desc,
3176                                               rqbd_list);
3177                         ptlrpc_free_rqbd(rqbd);
3178                 }
3179                 ptlrpc_wait_replies(svcpt);
3180
3181                 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3182                         rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3183                                             struct ptlrpc_reply_state,
3184                                             rs_list);
3185                         cfs_list_del(&rs->rs_list);
3186                         OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3187                 }
3188         }
3189 }
3190
3191 static void
3192 ptlrpc_service_free(struct ptlrpc_service *svc)
3193 {
3194         struct ptlrpc_service_part      *svcpt;
3195         struct ptlrpc_at_array          *array;
3196         int                             i;
3197
3198         ptlrpc_service_for_each_part(svcpt, i, svc) {
3199                 if (svcpt->scp_service == NULL)
3200                         break;
3201
3202                 /* In case somebody rearmed this in the meantime */
3203                 cfs_timer_disarm(&svcpt->scp_at_timer);
3204                 array = &svcpt->scp_at_array;
3205
3206                 if (array->paa_reqs_array != NULL) {
3207                         OBD_FREE(array->paa_reqs_array,
3208                                  sizeof(cfs_list_t) * array->paa_size);
3209                         array->paa_reqs_array = NULL;
3210                 }
3211
3212                 if (array->paa_reqs_count != NULL) {
3213                         OBD_FREE(array->paa_reqs_count,
3214                                  sizeof(__u32) * array->paa_size);
3215                         array->paa_reqs_count = NULL;
3216                 }
3217         }
3218
3219         ptlrpc_service_for_each_part(svcpt, i, svc)
3220                 OBD_FREE_PTR(svcpt);
3221
3222         if (svc->srv_cpts != NULL)
3223                 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3224
3225         OBD_FREE(svc, offsetof(struct ptlrpc_service,
3226                                srv_parts[svc->srv_ncpts]));
3227 }
3228
3229 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3230 {
3231         ENTRY;
3232
3233         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3234
3235         service->srv_is_stopping = 1;
3236
3237         mutex_lock(&ptlrpc_all_services_mutex);
3238         cfs_list_del_init(&service->srv_list);
3239         mutex_unlock(&ptlrpc_all_services_mutex);
3240
3241         ptlrpc_service_del_atimer(service);
3242         ptlrpc_stop_all_threads(service);
3243
3244         ptlrpc_service_unlink_rqbd(service);
3245         ptlrpc_service_purge_all(service);
3246         ptlrpc_service_nrs_cleanup(service);
3247
3248         ptlrpc_lprocfs_unregister_service(service);
3249
3250         ptlrpc_service_free(service);