Whamcloud - gitweb
LU-398 ptlrpc: Add the NRS ORR and TRR policies
[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, 2012, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
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         cfs_waitq_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         cfs_waitq_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                 cfs_waitq_signal(&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         cfs_waitq_signal(&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_LOCKED(&rs->rs_svcpt->scp_rep_lock);
408         LASSERT_SPIN_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         cfs_waitq_signal(&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         cfs_waitq_init(&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         cfs_waitq_init(&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                         cfs_proc_dir_entry_t *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                 OBD_FREE(req, sizeof(*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                 if (!cfs_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                         cfs_list_del_init(&req->rq_exp_list);
996                         spin_unlock_bh(&req->rq_export->exp_rpc_lock);
997
998                         /* export has one reference already, so it`s safe to
999                          * add req to export queue here and get another
1000                          * reference for request later */
1001                         spin_lock_bh(&export->exp_rpc_lock);
1002                         cfs_list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1003                         spin_unlock_bh(&export->exp_rpc_lock);
1004                 }
1005                 class_export_rpc_dec(req->rq_export);
1006                 class_export_put(req->rq_export);
1007         }
1008
1009         /* request takes one export refcount */
1010         req->rq_export = class_export_get(export);
1011         class_export_rpc_inc(export);
1012
1013         return;
1014 }
1015
1016 /**
1017  * to finish a request: stop sending more early replies, and release
1018  * the request.
1019  */
1020 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1021                                          struct ptlrpc_request *req)
1022 {
1023         ptlrpc_server_hpreq_fini(req);
1024
1025         ptlrpc_server_drop_request(req);
1026 }
1027
1028 /**
1029  * to finish a active request: stop sending more early replies, and release
1030  * the request. should be called after we finished handling the request.
1031  */
1032 static void ptlrpc_server_finish_active_request(
1033                                         struct ptlrpc_service_part *svcpt,
1034                                         struct ptlrpc_request *req)
1035 {
1036         spin_lock(&svcpt->scp_req_lock);
1037         ptlrpc_nrs_req_stop_nolock(req);
1038         svcpt->scp_nreqs_active--;
1039         if (req->rq_hp)
1040                 svcpt->scp_nhreqs_active--;
1041         spin_unlock(&svcpt->scp_req_lock);
1042
1043         ptlrpc_nrs_req_finalize(req);
1044
1045         if (req->rq_export != NULL)
1046                 class_export_rpc_dec(req->rq_export);
1047
1048         ptlrpc_server_finish_request(svcpt, req);
1049 }
1050
1051 /**
1052  * This function makes sure dead exports are evicted in a timely manner.
1053  * This function is only called when some export receives a message (i.e.,
1054  * the network is up.)
1055  */
1056 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1057 {
1058         struct obd_export *oldest_exp;
1059         time_t oldest_time, new_time;
1060
1061         ENTRY;
1062
1063         LASSERT(exp);
1064
1065         /* Compensate for slow machines, etc, by faking our request time
1066            into the future.  Although this can break the strict time-ordering
1067            of the list, we can be really lazy here - we don't have to evict
1068            at the exact right moment.  Eventually, all silent exports
1069            will make it to the top of the list. */
1070
1071         /* Do not pay attention on 1sec or smaller renewals. */
1072         new_time = cfs_time_current_sec() + extra_delay;
1073         if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1074                 RETURN_EXIT;
1075
1076         exp->exp_last_request_time = new_time;
1077         CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1078                exp->exp_client_uuid.uuid,
1079                exp->exp_last_request_time, exp);
1080
1081         /* exports may get disconnected from the chain even though the
1082            export has references, so we must keep the spin lock while
1083            manipulating the lists */
1084         spin_lock(&exp->exp_obd->obd_dev_lock);
1085
1086         if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
1087                 /* this one is not timed */
1088                 spin_unlock(&exp->exp_obd->obd_dev_lock);
1089                 RETURN_EXIT;
1090         }
1091
1092         cfs_list_move_tail(&exp->exp_obd_chain_timed,
1093                            &exp->exp_obd->obd_exports_timed);
1094
1095         oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
1096                                     struct obd_export, exp_obd_chain_timed);
1097         oldest_time = oldest_exp->exp_last_request_time;
1098         spin_unlock(&exp->exp_obd->obd_dev_lock);
1099
1100         if (exp->exp_obd->obd_recovering) {
1101                 /* be nice to everyone during recovery */
1102                 EXIT;
1103                 return;
1104         }
1105
1106         /* Note - racing to start/reset the obd_eviction timer is safe */
1107         if (exp->exp_obd->obd_eviction_timer == 0) {
1108                 /* Check if the oldest entry is expired. */
1109                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1110                                               extra_delay)) {
1111                         /* We need a second timer, in case the net was down and
1112                          * it just came back. Since the pinger may skip every
1113                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1114                          * we better wait for 3. */
1115                         exp->exp_obd->obd_eviction_timer =
1116                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
1117                         CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1118                                exp->exp_obd->obd_name, 
1119                                obd_export_nid2str(oldest_exp), oldest_time);
1120                 }
1121         } else {
1122                 if (cfs_time_current_sec() >
1123                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1124                         /* The evictor won't evict anyone who we've heard from
1125                          * recently, so we don't have to check before we start
1126                          * it. */
1127                         if (!ping_evictor_wake(exp))
1128                                 exp->exp_obd->obd_eviction_timer = 0;
1129                 }
1130         }
1131
1132         EXIT;
1133 }
1134
1135 /**
1136  * Sanity check request \a req.
1137  * Return 0 if all is ok, error code otherwise.
1138  */
1139 static int ptlrpc_check_req(struct ptlrpc_request *req)
1140 {
1141         int rc = 0;
1142
1143         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1144                      req->rq_export->exp_conn_cnt)) {
1145                 DEBUG_REQ(D_RPCTRACE, req,
1146                           "DROPPING req from old connection %d < %d",
1147                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
1148                           req->rq_export->exp_conn_cnt);
1149                 return -EEXIST;
1150         }
1151         if (unlikely(req->rq_export->exp_obd &&
1152                      req->rq_export->exp_obd->obd_fail)) {
1153              /* Failing over, don't handle any more reqs, send
1154                 error response instead. */
1155                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1156                        req, req->rq_export->exp_obd->obd_name);
1157                 rc = -ENODEV;
1158         } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1159                    (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1160                    !(req->rq_export->exp_obd->obd_recovering)) {
1161                         DEBUG_REQ(D_ERROR, req,
1162                                   "Invalid replay without recovery");
1163                         class_fail_export(req->rq_export);
1164                         rc = -ENODEV;
1165         } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1166                    !(req->rq_export->exp_obd->obd_recovering)) {
1167                         DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1168                                   LPU64" without recovery",
1169                                   lustre_msg_get_transno(req->rq_reqmsg));
1170                         class_fail_export(req->rq_export);
1171                         rc = -ENODEV;
1172         }
1173
1174         if (unlikely(rc < 0)) {
1175                 req->rq_status = rc;
1176                 ptlrpc_error(req);
1177         }
1178         return rc;
1179 }
1180
1181 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1182 {
1183         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1184         __s32 next;
1185
1186         if (array->paa_count == 0) {
1187                 cfs_timer_disarm(&svcpt->scp_at_timer);
1188                 return;
1189         }
1190
1191         /* Set timer for closest deadline */
1192         next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1193                        at_early_margin);
1194         if (next <= 0) {
1195                 ptlrpc_at_timer((unsigned long)svcpt);
1196         } else {
1197                 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1198                 CDEBUG(D_INFO, "armed %s at %+ds\n",
1199                        svcpt->scp_service->srv_name, next);
1200         }
1201 }
1202
1203 /* Add rpc to early reply check list */
1204 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1205 {
1206         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1207         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1208         struct ptlrpc_request *rq = NULL;
1209         __u32 index;
1210
1211         if (AT_OFF)
1212                 return(0);
1213
1214         if (req->rq_no_reply)
1215                 return 0;
1216
1217         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1218                 return(-ENOSYS);
1219
1220         spin_lock(&svcpt->scp_at_lock);
1221         LASSERT(cfs_list_empty(&req->rq_timed_list));
1222
1223         index = (unsigned long)req->rq_deadline % array->paa_size;
1224         if (array->paa_reqs_count[index] > 0) {
1225                 /* latest rpcs will have the latest deadlines in the list,
1226                  * so search backward. */
1227                 cfs_list_for_each_entry_reverse(rq,
1228                                                 &array->paa_reqs_array[index],
1229                                                 rq_timed_list) {
1230                         if (req->rq_deadline >= rq->rq_deadline) {
1231                                 cfs_list_add(&req->rq_timed_list,
1232                                              &rq->rq_timed_list);
1233                                 break;
1234                         }
1235                 }
1236         }
1237
1238         /* Add the request at the head of the list */
1239         if (cfs_list_empty(&req->rq_timed_list))
1240                 cfs_list_add(&req->rq_timed_list,
1241                              &array->paa_reqs_array[index]);
1242
1243         spin_lock(&req->rq_lock);
1244         req->rq_at_linked = 1;
1245         spin_unlock(&req->rq_lock);
1246         req->rq_at_index = index;
1247         array->paa_reqs_count[index]++;
1248         array->paa_count++;
1249         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1250                 array->paa_deadline = req->rq_deadline;
1251                 ptlrpc_at_set_timer(svcpt);
1252         }
1253         spin_unlock(&svcpt->scp_at_lock);
1254
1255         return 0;
1256 }
1257
1258 static void
1259 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1260 {
1261         struct ptlrpc_at_array *array;
1262
1263         array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1264
1265         /* NB: must call with hold svcpt::scp_at_lock */
1266         LASSERT(!cfs_list_empty(&req->rq_timed_list));
1267         cfs_list_del_init(&req->rq_timed_list);
1268
1269         spin_lock(&req->rq_lock);
1270         req->rq_at_linked = 0;
1271         spin_unlock(&req->rq_lock);
1272
1273         array->paa_reqs_count[req->rq_at_index]--;
1274         array->paa_count--;
1275 }
1276
1277 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1278 {
1279         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1280         struct ptlrpc_request *reqcopy;
1281         struct lustre_msg *reqmsg;
1282         cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1283         time_t newdl;
1284         int rc;
1285         ENTRY;
1286
1287         /* deadline is when the client expects us to reply, margin is the
1288            difference between clients' and servers' expectations */
1289         DEBUG_REQ(D_ADAPTTO, req,
1290                   "%ssending early reply (deadline %+lds, margin %+lds) for "
1291                   "%d+%d", AT_OFF ? "AT off - not " : "",
1292                   olddl, olddl - at_get(&svcpt->scp_at_estimate),
1293                   at_get(&svcpt->scp_at_estimate), at_extra);
1294
1295         if (AT_OFF)
1296                 RETURN(0);
1297
1298         if (olddl < 0) {
1299                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1300                           "not sending early reply. Consider increasing "
1301                           "at_early_margin (%d)?", olddl, at_early_margin);
1302
1303                 /* Return an error so we're not re-added to the timed list. */
1304                 RETURN(-ETIMEDOUT);
1305         }
1306
1307         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1308                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1309                           "but no AT support");
1310                 RETURN(-ENOSYS);
1311         }
1312
1313         if (req->rq_export &&
1314             lustre_msg_get_flags(req->rq_reqmsg) &
1315             (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1316                 /* During recovery, we don't want to send too many early
1317                  * replies, but on the other hand we want to make sure the
1318                  * client has enough time to resend if the rpc is lost. So
1319                  * during the recovery period send at least 4 early replies,
1320                  * spacing them every at_extra if we can. at_estimate should
1321                  * always equal this fixed value during recovery. */
1322                 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1323                             req->rq_export->exp_obd->obd_recovery_timeout / 4));
1324         } else {
1325                 /* Fake our processing time into the future to ask the clients
1326                  * for some extra amount of time */
1327                 at_measured(&svcpt->scp_at_estimate, at_extra +
1328                             cfs_time_current_sec() -
1329                             req->rq_arrival_time.tv_sec);
1330
1331                 /* Check to see if we've actually increased the deadline -
1332                  * we may be past adaptive_max */
1333                 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1334                     at_get(&svcpt->scp_at_estimate)) {
1335                         DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1336                                   "(%ld/%ld), not sending early reply\n",
1337                                   olddl, req->rq_arrival_time.tv_sec +
1338                                   at_get(&svcpt->scp_at_estimate) -
1339                                   cfs_time_current_sec());
1340                         RETURN(-ETIMEDOUT);
1341                 }
1342         }
1343         newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1344
1345         OBD_ALLOC(reqcopy, sizeof *reqcopy);
1346         if (reqcopy == NULL)
1347                 RETURN(-ENOMEM);
1348         OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1349         if (!reqmsg) {
1350                 OBD_FREE(reqcopy, sizeof *reqcopy);
1351                 RETURN(-ENOMEM);
1352         }
1353
1354         *reqcopy = *req;
1355         reqcopy->rq_reply_state = NULL;
1356         reqcopy->rq_rep_swab_mask = 0;
1357         reqcopy->rq_pack_bulk = 0;
1358         reqcopy->rq_pack_udesc = 0;
1359         reqcopy->rq_packed_final = 0;
1360         sptlrpc_svc_ctx_addref(reqcopy);
1361         /* We only need the reqmsg for the magic */
1362         reqcopy->rq_reqmsg = reqmsg;
1363         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1364
1365         LASSERT(cfs_atomic_read(&req->rq_refcount));
1366         /** if it is last refcount then early reply isn't needed */
1367         if (cfs_atomic_read(&req->rq_refcount) == 1) {
1368                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1369                           "abort sending early reply\n");
1370                 GOTO(out, rc = -EINVAL);
1371         }
1372
1373         /* Connection ref */
1374         reqcopy->rq_export = class_conn2export(
1375                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
1376         if (reqcopy->rq_export == NULL)
1377                 GOTO(out, rc = -ENODEV);
1378
1379         /* RPC ref */
1380         class_export_rpc_inc(reqcopy->rq_export);
1381         if (reqcopy->rq_export->exp_obd &&
1382             reqcopy->rq_export->exp_obd->obd_fail)
1383                 GOTO(out_put, rc = -ENODEV);
1384
1385         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1386         if (rc)
1387                 GOTO(out_put, rc);
1388
1389         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1390
1391         if (!rc) {
1392                 /* Adjust our own deadline to what we told the client */
1393                 req->rq_deadline = newdl;
1394                 req->rq_early_count++; /* number sent, server side */
1395         } else {
1396                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1397         }
1398
1399         /* Free the (early) reply state from lustre_pack_reply.
1400            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1401         ptlrpc_req_drop_rs(reqcopy);
1402
1403 out_put:
1404         class_export_rpc_dec(reqcopy->rq_export);
1405         class_export_put(reqcopy->rq_export);
1406 out:
1407         sptlrpc_svc_ctx_decref(reqcopy);
1408         OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1409         OBD_FREE(reqcopy, sizeof *reqcopy);
1410         RETURN(rc);
1411 }
1412
1413 /* Send early replies to everybody expiring within at_early_margin
1414    asking for at_extra time */
1415 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1416 {
1417         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1418         struct ptlrpc_request *rq, *n;
1419         cfs_list_t work_list;
1420         __u32  index, count;
1421         time_t deadline;
1422         time_t now = cfs_time_current_sec();
1423         cfs_duration_t delay;
1424         int first, counter = 0;
1425         ENTRY;
1426
1427         spin_lock(&svcpt->scp_at_lock);
1428         if (svcpt->scp_at_check == 0) {
1429                 spin_unlock(&svcpt->scp_at_lock);
1430                 RETURN(0);
1431         }
1432         delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1433         svcpt->scp_at_check = 0;
1434
1435         if (array->paa_count == 0) {
1436                 spin_unlock(&svcpt->scp_at_lock);
1437                 RETURN(0);
1438         }
1439
1440         /* The timer went off, but maybe the nearest rpc already completed. */
1441         first = array->paa_deadline - now;
1442         if (first > at_early_margin) {
1443                 /* We've still got plenty of time.  Reset the timer. */
1444                 ptlrpc_at_set_timer(svcpt);
1445                 spin_unlock(&svcpt->scp_at_lock);
1446                 RETURN(0);
1447         }
1448
1449         /* We're close to a timeout, and we don't know how much longer the
1450            server will take. Send early replies to everyone expiring soon. */
1451         CFS_INIT_LIST_HEAD(&work_list);
1452         deadline = -1;
1453         index = (unsigned long)array->paa_deadline % array->paa_size;
1454         count = array->paa_count;
1455         while (count > 0) {
1456                 count -= array->paa_reqs_count[index];
1457                 cfs_list_for_each_entry_safe(rq, n,
1458                                              &array->paa_reqs_array[index],
1459                                              rq_timed_list) {
1460                         if (rq->rq_deadline > now + at_early_margin) {
1461                                 /* update the earliest deadline */
1462                                 if (deadline == -1 ||
1463                                     rq->rq_deadline < deadline)
1464                                         deadline = rq->rq_deadline;
1465                                 break;
1466                         }
1467
1468                         ptlrpc_at_remove_timed(rq);
1469                         /**
1470                          * ptlrpc_server_drop_request() may drop
1471                          * refcount to 0 already. Let's check this and
1472                          * don't add entry to work_list
1473                          */
1474                         if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1475                                 cfs_list_add(&rq->rq_timed_list, &work_list);
1476                         counter++;
1477                 }
1478
1479                 if (++index >= array->paa_size)
1480                         index = 0;
1481         }
1482         array->paa_deadline = deadline;
1483         /* we have a new earliest deadline, restart the timer */
1484         ptlrpc_at_set_timer(svcpt);
1485
1486         spin_unlock(&svcpt->scp_at_lock);
1487
1488         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1489                "replies\n", first, at_extra, counter);
1490         if (first < 0) {
1491                 /* We're already past request deadlines before we even get a
1492                    chance to send early replies */
1493                 LCONSOLE_WARN("%s: This server is not able to keep up with "
1494                               "request traffic (cpu-bound).\n",
1495                               svcpt->scp_service->srv_name);
1496                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1497                       "delay="CFS_DURATION_T"(jiff)\n",
1498                       counter, svcpt->scp_nreqs_incoming,
1499                       svcpt->scp_nreqs_active,
1500                       at_get(&svcpt->scp_at_estimate), delay);
1501         }
1502
1503         /* we took additional refcount so entries can't be deleted from list, no
1504          * locking is needed */
1505         while (!cfs_list_empty(&work_list)) {
1506                 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1507                                     rq_timed_list);
1508                 cfs_list_del_init(&rq->rq_timed_list);
1509
1510                 if (ptlrpc_at_send_early_reply(rq) == 0)
1511                         ptlrpc_at_add_timed(rq);
1512
1513                 ptlrpc_server_drop_request(rq);
1514         }
1515
1516         RETURN(1); /* return "did_something" for liblustre */
1517 }
1518
1519 /**
1520  * Put the request to the export list if the request may become
1521  * a high priority one.
1522  */
1523 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1524                                     struct ptlrpc_request *req)
1525 {
1526         int rc = 0;
1527         ENTRY;
1528
1529         if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1530                 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1531                 if (rc < 0)
1532                         RETURN(rc);
1533                 LASSERT(rc == 0);
1534         }
1535         if (req->rq_export && req->rq_ops) {
1536                 /* Perform request specific check. We should do this check
1537                  * before the request is added into exp_hp_rpcs list otherwise
1538                  * it may hit swab race at LU-1044. */
1539                 if (req->rq_ops->hpreq_check) {
1540                         rc = req->rq_ops->hpreq_check(req);
1541                         /**
1542                          * XXX: Out of all current
1543                          * ptlrpc_hpreq_ops::hpreq_check(), only
1544                          * ldlm_cancel_hpreq_check() can return an error code;
1545                          * other functions assert in similar places, which seems
1546                          * odd. What also does not seem right is that handlers
1547                          * for those RPCs do not assert on the same checks, but
1548                          * rather handle the error cases. e.g. see
1549                          * ost_rw_hpreq_check(), and ost_brw_read(),
1550                          * ost_brw_write().
1551                          */
1552                         if (rc < 0)
1553                                 RETURN(rc);
1554                         LASSERT(rc == 0 || rc == 1);
1555                 }
1556
1557                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1558                 cfs_list_add(&req->rq_exp_list,
1559                              &req->rq_export->exp_hp_rpcs);
1560                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1561         }
1562
1563         ptlrpc_nrs_req_initialize(svcpt, req, rc);
1564
1565         RETURN(rc);
1566 }
1567
1568 /** Remove the request from the export list. */
1569 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1570 {
1571         ENTRY;
1572         if (req->rq_export && req->rq_ops) {
1573                 /* refresh lock timeout again so that client has more
1574                  * room to send lock cancel RPC. */
1575                 if (req->rq_ops->hpreq_fini)
1576                         req->rq_ops->hpreq_fini(req);
1577
1578                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1579                 cfs_list_del_init(&req->rq_exp_list);
1580                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1581         }
1582         EXIT;
1583 }
1584
1585 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1586 {
1587         return 1;
1588 }
1589
1590 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1591         .hpreq_check       = ptlrpc_hpreq_check,
1592 };
1593
1594 /* Hi-Priority RPC check by RPC operation code. */
1595 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1596 {
1597         int opc = lustre_msg_get_opc(req->rq_reqmsg);
1598
1599         /* Check for export to let only reconnects for not yet evicted
1600          * export to become a HP rpc. */
1601         if ((req->rq_export != NULL) &&
1602             (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1603                 req->rq_ops = &ptlrpc_hpreq_common;
1604
1605         return 0;
1606 }
1607 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1608
1609 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1610                                      struct ptlrpc_request *req)
1611 {
1612         int     rc;
1613         ENTRY;
1614
1615         rc = ptlrpc_server_hpreq_init(svcpt, req);
1616         if (rc < 0)
1617                 RETURN(rc);
1618
1619         ptlrpc_nrs_req_add(svcpt, req, !!rc);
1620
1621         RETURN(0);
1622 }
1623
1624 /**
1625  * Allow to handle high priority request
1626  * User can call it w/o any lock but need to hold
1627  * ptlrpc_service_part::scp_req_lock to get reliable result
1628  */
1629 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1630                                      bool force)
1631 {
1632         int running = svcpt->scp_nthrs_running;
1633
1634         if (!nrs_svcpt_has_hp(svcpt))
1635                 return false;
1636
1637         if (force)
1638                 return true;
1639
1640         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1641                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1642                 /* leave just 1 thread for normal RPCs */
1643                 running = PTLRPC_NTHRS_INIT;
1644                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1645                         running += 1;
1646         }
1647
1648         if (svcpt->scp_nreqs_active >= running - 1)
1649                 return false;
1650
1651         if (svcpt->scp_nhreqs_active == 0)
1652                 return true;
1653
1654         return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1655                svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1656 }
1657
1658 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1659                                        bool force)
1660 {
1661         return ptlrpc_server_allow_high(svcpt, force) &&
1662                ptlrpc_nrs_req_pending_nolock(svcpt, true);
1663 }
1664
1665 /**
1666  * Only allow normal priority requests on a service that has a high-priority
1667  * queue if forced (i.e. cleanup), if there are other high priority requests
1668  * already being processed (i.e. those threads can service more high-priority
1669  * requests), or if there are enough idle threads that a later thread can do
1670  * a high priority request.
1671  * User can call it w/o any lock but need to hold
1672  * ptlrpc_service_part::scp_req_lock to get reliable result
1673  */
1674 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1675                                        bool force)
1676 {
1677         int running = svcpt->scp_nthrs_running;
1678 #ifndef __KERNEL__
1679         if (1) /* always allow to handle normal request for liblustre */
1680                 return true;
1681 #endif
1682         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1683                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1684                 /* leave just 1 thread for normal RPCs */
1685                 running = PTLRPC_NTHRS_INIT;
1686                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1687                         running += 1;
1688         }
1689
1690         if (force ||
1691             svcpt->scp_nreqs_active < running - 2)
1692                 return true;
1693
1694         if (svcpt->scp_nreqs_active >= running - 1)
1695                 return false;
1696
1697         return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1698 }
1699
1700 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1701                                          bool force)
1702 {
1703         return ptlrpc_server_allow_normal(svcpt, force) &&
1704                ptlrpc_nrs_req_pending_nolock(svcpt, false);
1705 }
1706
1707 /**
1708  * Returns true if there are requests available in incoming
1709  * request queue for processing and it is allowed to fetch them.
1710  * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1711  * to get reliable result
1712  * \see ptlrpc_server_allow_normal
1713  * \see ptlrpc_server_allow high
1714  */
1715 static inline bool
1716 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1717 {
1718         return ptlrpc_server_high_pending(svcpt, force) ||
1719                ptlrpc_server_normal_pending(svcpt, force);
1720 }
1721
1722 /**
1723  * Fetch a request for processing from queue of unprocessed requests.
1724  * Favors high-priority requests.
1725  * Returns a pointer to fetched request.
1726  */
1727 static struct ptlrpc_request *
1728 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1729 {
1730         struct ptlrpc_request *req = NULL;
1731         ENTRY;
1732
1733         spin_lock(&svcpt->scp_req_lock);
1734 #ifndef __KERNEL__
1735         /* !@%$# liblustre only has 1 thread */
1736         if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1737                 spin_unlock(&svcpt->scp_req_lock);
1738                 RETURN(NULL);
1739         }
1740 #endif
1741
1742         if (ptlrpc_server_high_pending(svcpt, force)) {
1743                 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1744                 if (req != NULL) {
1745                         svcpt->scp_hreq_count++;
1746                         goto got_request;
1747                 }
1748         }
1749
1750         if (ptlrpc_server_normal_pending(svcpt, force)) {
1751                 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1752                 if (req != NULL) {
1753                         svcpt->scp_hreq_count = 0;
1754                         goto got_request;
1755                 }
1756         }
1757
1758         spin_unlock(&svcpt->scp_req_lock);
1759         RETURN(NULL);
1760
1761 got_request:
1762         svcpt->scp_nreqs_active++;
1763         if (req->rq_hp)
1764                 svcpt->scp_nhreqs_active++;
1765
1766         spin_unlock(&svcpt->scp_req_lock);
1767
1768         if (likely(req->rq_export))
1769                 class_export_rpc_inc(req->rq_export);
1770
1771         RETURN(req);
1772 }
1773
1774 /**
1775  * Handle freshly incoming reqs, add to timed early reply list,
1776  * pass on to regular request queue.
1777  * All incoming requests pass through here before getting into
1778  * ptlrpc_server_handle_req later on.
1779  */
1780 static int
1781 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1782                             struct ptlrpc_thread *thread)
1783 {
1784         struct ptlrpc_service   *svc = svcpt->scp_service;
1785         struct ptlrpc_request   *req;
1786         __u32                   deadline;
1787         int                     rc;
1788         ENTRY;
1789
1790         spin_lock(&svcpt->scp_lock);
1791         if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1792                 spin_unlock(&svcpt->scp_lock);
1793                 RETURN(0);
1794         }
1795
1796         req = cfs_list_entry(svcpt->scp_req_incoming.next,
1797                              struct ptlrpc_request, rq_list);
1798         cfs_list_del_init(&req->rq_list);
1799         svcpt->scp_nreqs_incoming--;
1800         /* Consider this still a "queued" request as far as stats are
1801          * concerned */
1802         spin_unlock(&svcpt->scp_lock);
1803
1804         /* go through security check/transform */
1805         rc = sptlrpc_svc_unwrap_request(req);
1806         switch (rc) {
1807         case SECSVC_OK:
1808                 break;
1809         case SECSVC_COMPLETE:
1810                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1811                 goto err_req;
1812         case SECSVC_DROP:
1813                 goto err_req;
1814         default:
1815                 LBUG();
1816         }
1817
1818         /*
1819          * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1820          * redo it wouldn't be harmful.
1821          */
1822         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1823                 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1824                 if (rc != 0) {
1825                         CERROR("error unpacking request: ptl %d from %s "
1826                                "x"LPU64"\n", svc->srv_req_portal,
1827                                libcfs_id2str(req->rq_peer), req->rq_xid);
1828                         goto err_req;
1829                 }
1830         }
1831
1832         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1833         if (rc) {
1834                 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1835                         LPU64"\n", svc->srv_req_portal,
1836                         libcfs_id2str(req->rq_peer), req->rq_xid);
1837                 goto err_req;
1838         }
1839
1840         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1841             lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1842                 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1843                        cfs_fail_val, req->rq_xid);
1844                 goto err_req;
1845         }
1846
1847         rc = -EINVAL;
1848         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1849                 CERROR("wrong packet type received (type=%u) from %s\n",
1850                        lustre_msg_get_type(req->rq_reqmsg),
1851                        libcfs_id2str(req->rq_peer));
1852                 goto err_req;
1853         }
1854
1855         switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1856         case MDS_WRITEPAGE:
1857         case OST_WRITE:
1858                 req->rq_bulk_write = 1;
1859                 break;
1860         case MDS_READPAGE:
1861         case OST_READ:
1862         case MGS_CONFIG_READ:
1863                 req->rq_bulk_read = 1;
1864                 break;
1865         }
1866
1867         CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1868
1869         req->rq_export = class_conn2export(
1870                 lustre_msg_get_handle(req->rq_reqmsg));
1871         if (req->rq_export) {
1872                 rc = ptlrpc_check_req(req);
1873                 if (rc == 0) {
1874                         rc = sptlrpc_target_export_check(req->rq_export, req);
1875                         if (rc)
1876                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1877                                           "illegal security flavor,");
1878                 }
1879
1880                 if (rc)
1881                         goto err_req;
1882                 ptlrpc_update_export_timer(req->rq_export, 0);
1883         }
1884
1885         /* req_in handling should/must be fast */
1886         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1887                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1888                           cfs_time_sub(cfs_time_current_sec(),
1889                                        req->rq_arrival_time.tv_sec));
1890
1891         /* Set rpc server deadline and add it to the timed list */
1892         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1893                     MSGHDR_AT_SUPPORT) ?
1894                    /* The max time the client expects us to take */
1895                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1896         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1897         if (unlikely(deadline == 0)) {
1898                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1899                 goto err_req;
1900         }
1901
1902         req->rq_svc_thread = thread;
1903
1904         ptlrpc_at_add_timed(req);
1905
1906         /* Move it over to the request processing queue */
1907         rc = ptlrpc_server_request_add(svcpt, req);
1908         if (rc)
1909                 GOTO(err_req, rc);
1910
1911         cfs_waitq_signal(&svcpt->scp_waitq);
1912         RETURN(1);
1913
1914 err_req:
1915         ptlrpc_server_finish_request(svcpt, req);
1916
1917         RETURN(1);
1918 }
1919
1920 /**
1921  * Main incoming request handling logic.
1922  * Calls handler function from service to do actual processing.
1923  */
1924 static int
1925 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1926                              struct ptlrpc_thread *thread)
1927 {
1928         struct ptlrpc_service *svc = svcpt->scp_service;
1929         struct ptlrpc_request *request;
1930         struct timeval         work_start;
1931         struct timeval         work_end;
1932         long                   timediff;
1933         int                    rc;
1934         int                    fail_opc = 0;
1935         ENTRY;
1936
1937         request = ptlrpc_server_request_get(svcpt, false);
1938         if (request == NULL)
1939                 RETURN(0);
1940
1941         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1942                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1943         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1944                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1945
1946         if (unlikely(fail_opc)) {
1947                 if (request->rq_export && request->rq_ops)
1948                         OBD_FAIL_TIMEOUT(fail_opc, 4);
1949         }
1950
1951         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1952
1953         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1954                 libcfs_debug_dumplog();
1955
1956         cfs_gettimeofday(&work_start);
1957         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1958         if (likely(svc->srv_stats != NULL)) {
1959                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1960                                     timediff);
1961                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1962                                     svcpt->scp_nreqs_incoming);
1963                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1964                                     svcpt->scp_nreqs_active);
1965                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1966                                     at_get(&svcpt->scp_at_estimate));
1967         }
1968
1969         rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1970         if (rc) {
1971                 CERROR("Failure to initialize session: %d\n", rc);
1972                 goto out_req;
1973         }
1974         request->rq_session.lc_thread = thread;
1975         request->rq_session.lc_cookie = 0x5;
1976         lu_context_enter(&request->rq_session);
1977
1978         CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1979
1980         request->rq_svc_thread = thread;
1981         if (thread)
1982                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1983
1984         if (likely(request->rq_export)) {
1985                 if (unlikely(ptlrpc_check_req(request)))
1986                         goto put_conn;
1987                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1988         }
1989
1990         /* Discard requests queued for longer than the deadline.
1991            The deadline is increased if we send an early reply. */
1992         if (cfs_time_current_sec() > request->rq_deadline) {
1993                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1994                           ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1995                           libcfs_id2str(request->rq_peer),
1996                           cfs_time_sub(request->rq_deadline,
1997                           request->rq_arrival_time.tv_sec),
1998                           cfs_time_sub(cfs_time_current_sec(),
1999                           request->rq_deadline));
2000                 goto put_conn;
2001         }
2002
2003         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2004                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
2005                (request->rq_export ?
2006                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2007                (request->rq_export ?
2008                 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2009                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2010                libcfs_id2str(request->rq_peer),
2011                lustre_msg_get_opc(request->rq_reqmsg));
2012
2013         if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2014                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2015
2016         rc = svc->srv_ops.so_req_handler(request);
2017
2018         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2019
2020 put_conn:
2021         lu_context_exit(&request->rq_session);
2022         lu_context_fini(&request->rq_session);
2023
2024         if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2025                      DEBUG_REQ(D_WARNING, request, "Request took longer "
2026                                "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2027                                " client may timeout.",
2028                                cfs_time_sub(request->rq_deadline,
2029                                             request->rq_arrival_time.tv_sec),
2030                                cfs_time_sub(cfs_time_current_sec(),
2031                                             request->rq_deadline));
2032         }
2033
2034         cfs_gettimeofday(&work_end);
2035         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2036         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2037                "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2038                "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2039                 cfs_curproc_comm(),
2040                 (request->rq_export ?
2041                  (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2042                 (request->rq_export ?
2043                  cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2044                 lustre_msg_get_status(request->rq_reqmsg),
2045                 request->rq_xid,
2046                 libcfs_id2str(request->rq_peer),
2047                 lustre_msg_get_opc(request->rq_reqmsg),
2048                 timediff,
2049                 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2050                 (request->rq_repmsg ?
2051                  lustre_msg_get_transno(request->rq_repmsg) :
2052                  request->rq_transno),
2053                 request->rq_status,
2054                 (request->rq_repmsg ?
2055                  lustre_msg_get_status(request->rq_repmsg) : -999));
2056         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2057                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2058                 int opc = opcode_offset(op);
2059                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2060                         LASSERT(opc < LUSTRE_MAX_OPCODES);
2061                         lprocfs_counter_add(svc->srv_stats,
2062                                             opc + EXTRA_MAX_OPCODES,
2063                                             timediff);
2064                 }
2065         }
2066         if (unlikely(request->rq_early_count)) {
2067                 DEBUG_REQ(D_ADAPTTO, request,
2068                           "sent %d early replies before finishing in "
2069                           CFS_DURATION_T"s",
2070                           request->rq_early_count,
2071                           cfs_time_sub(work_end.tv_sec,
2072                           request->rq_arrival_time.tv_sec));
2073         }
2074
2075 out_req:
2076         ptlrpc_server_finish_active_request(svcpt, request);
2077
2078         RETURN(1);
2079 }
2080
2081 /**
2082  * An internal function to process a single reply state object.
2083  */
2084 static int
2085 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2086 {
2087         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2088         struct ptlrpc_service     *svc = svcpt->scp_service;
2089         struct obd_export         *exp;
2090         int                        nlocks;
2091         int                        been_handled;
2092         ENTRY;
2093
2094         exp = rs->rs_export;
2095
2096         LASSERT (rs->rs_difficult);
2097         LASSERT (rs->rs_scheduled);
2098         LASSERT (cfs_list_empty(&rs->rs_list));
2099
2100         spin_lock(&exp->exp_lock);
2101         /* Noop if removed already */
2102         cfs_list_del_init (&rs->rs_exp_list);
2103         spin_unlock(&exp->exp_lock);
2104
2105         /* The disk commit callback holds exp_uncommitted_replies_lock while it
2106          * iterates over newly committed replies, removing them from
2107          * exp_uncommitted_replies.  It then drops this lock and schedules the
2108          * replies it found for handling here.
2109          *
2110          * We can avoid contention for exp_uncommitted_replies_lock between the
2111          * HRT threads and further commit callbacks by checking rs_committed
2112          * which is set in the commit callback while it holds both
2113          * rs_lock and exp_uncommitted_reples.
2114          *
2115          * If we see rs_committed clear, the commit callback _may_ not have
2116          * handled this reply yet and we race with it to grab
2117          * exp_uncommitted_replies_lock before removing the reply from
2118          * exp_uncommitted_replies.  Note that if we lose the race and the
2119          * reply has already been removed, list_del_init() is a noop.
2120          *
2121          * If we see rs_committed set, we know the commit callback is handling,
2122          * or has handled this reply since store reordering might allow us to
2123          * see rs_committed set out of sequence.  But since this is done
2124          * holding rs_lock, we can be sure it has all completed once we hold
2125          * rs_lock, which we do right next.
2126          */
2127         if (!rs->rs_committed) {
2128                 spin_lock(&exp->exp_uncommitted_replies_lock);
2129                 cfs_list_del_init(&rs->rs_obd_list);
2130                 spin_unlock(&exp->exp_uncommitted_replies_lock);
2131         }
2132
2133         spin_lock(&rs->rs_lock);
2134
2135         been_handled = rs->rs_handled;
2136         rs->rs_handled = 1;
2137
2138         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
2139         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
2140
2141         if (nlocks == 0 && !been_handled) {
2142                 /* If we see this, we should already have seen the warning
2143                  * in mds_steal_ack_locks()  */
2144                 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2145                        " o%d NID %s\n",
2146                        rs,
2147                        rs->rs_xid, rs->rs_transno, rs->rs_opc,
2148                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
2149         }
2150
2151         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2152                 spin_unlock(&rs->rs_lock);
2153
2154                 if (!been_handled && rs->rs_on_net) {
2155                         LNetMDUnlink(rs->rs_md_h);
2156                         /* Ignore return code; we're racing with completion */
2157                 }
2158
2159                 while (nlocks-- > 0)
2160                         ldlm_lock_decref(&rs->rs_locks[nlocks],
2161                                          rs->rs_modes[nlocks]);
2162
2163                 spin_lock(&rs->rs_lock);
2164         }
2165
2166         rs->rs_scheduled = 0;
2167
2168         if (!rs->rs_on_net) {
2169                 /* Off the net */
2170                 spin_unlock(&rs->rs_lock);
2171
2172                 class_export_put (exp);
2173                 rs->rs_export = NULL;
2174                 ptlrpc_rs_decref (rs);
2175                 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2176                     svc->srv_is_stopping)
2177                         cfs_waitq_broadcast(&svcpt->scp_waitq);
2178                 RETURN(1);
2179         }
2180
2181         /* still on the net; callback will schedule */
2182         spin_unlock(&rs->rs_lock);
2183         RETURN(1);
2184 }
2185
2186 #ifndef __KERNEL__
2187
2188 /**
2189  * Check whether given service has a reply available for processing
2190  * and process it.
2191  *
2192  * \param svc a ptlrpc service
2193  * \retval 0 no replies processed
2194  * \retval 1 one reply processed
2195  */
2196 static int
2197 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2198 {
2199         struct ptlrpc_reply_state *rs = NULL;
2200         ENTRY;
2201
2202         spin_lock(&svcpt->scp_rep_lock);
2203         if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2204                 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2205                                     struct ptlrpc_reply_state,
2206                                     rs_list);
2207                 cfs_list_del_init(&rs->rs_list);
2208         }
2209         spin_unlock(&svcpt->scp_rep_lock);
2210         if (rs != NULL)
2211                 ptlrpc_handle_rs(rs);
2212         RETURN(rs != NULL);
2213 }
2214
2215 /* FIXME make use of timeout later */
2216 int
2217 liblustre_check_services (void *arg)
2218 {
2219         int  did_something = 0;
2220         int  rc;
2221         cfs_list_t *tmp, *nxt;
2222         ENTRY;
2223
2224         /* I'm relying on being single threaded, not to have to lock
2225          * ptlrpc_all_services etc */
2226         cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2227                 struct ptlrpc_service *svc =
2228                         cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2229                 struct ptlrpc_service_part *svcpt;
2230
2231                 LASSERT(svc->srv_ncpts == 1);
2232                 svcpt = svc->srv_parts[0];
2233
2234                 if (svcpt->scp_nthrs_running != 0)     /* I've recursed */
2235                         continue;
2236
2237                 /* service threads can block for bulk, so this limits us
2238                  * (arbitrarily) to recursing 1 stack frame per service.
2239                  * Note that the problem with recursion is that we have to
2240                  * unwind completely before our caller can resume. */
2241
2242                 svcpt->scp_nthrs_running++;
2243
2244                 do {
2245                         rc = ptlrpc_server_handle_req_in(svcpt, NULL);
2246                         rc |= ptlrpc_server_handle_reply(svcpt);
2247                         rc |= ptlrpc_at_check_timed(svcpt);
2248                         rc |= ptlrpc_server_handle_request(svcpt, NULL);
2249                         rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2250                         did_something |= rc;
2251                 } while (rc);
2252
2253                 svcpt->scp_nthrs_running--;
2254         }
2255
2256         RETURN(did_something);
2257 }
2258 #define ptlrpc_stop_all_threads(s) do {} while (0)
2259
2260 #else /* __KERNEL__ */
2261
2262 static void
2263 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2264 {
2265         int avail = svcpt->scp_nrqbds_posted;
2266         int low_water = test_req_buffer_pressure ? 0 :
2267                         svcpt->scp_service->srv_nbuf_per_group / 2;
2268
2269         /* NB I'm not locking; just looking. */
2270
2271         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2272          * allowed the request history to grow out of control.  We could put a
2273          * sanity check on that here and cull some history if we need the
2274          * space. */
2275
2276         if (avail <= low_water)
2277                 ptlrpc_grow_req_bufs(svcpt, 1);
2278
2279         if (svcpt->scp_service->srv_stats) {
2280                 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2281                                     PTLRPC_REQBUF_AVAIL_CNTR, avail);
2282         }
2283 }
2284
2285 static int
2286 ptlrpc_retry_rqbds(void *arg)
2287 {
2288         struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2289
2290         svcpt->scp_rqbd_timeout = 0;
2291         return -ETIMEDOUT;
2292 }
2293
2294 static inline int
2295 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2296 {
2297         return svcpt->scp_nreqs_active <
2298                svcpt->scp_nthrs_running - 1 -
2299                (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2300 }
2301
2302 /**
2303  * allowed to create more threads
2304  * user can call it w/o any lock but need to hold
2305  * ptlrpc_service_part::scp_lock to get reliable result
2306  */
2307 static inline int
2308 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2309 {
2310         return svcpt->scp_nthrs_running +
2311                svcpt->scp_nthrs_starting <
2312                svcpt->scp_service->srv_nthrs_cpt_limit;
2313 }
2314
2315 /**
2316  * too many requests and allowed to create more threads
2317  */
2318 static inline int
2319 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2320 {
2321         return !ptlrpc_threads_enough(svcpt) &&
2322                 ptlrpc_threads_increasable(svcpt);
2323 }
2324
2325 static inline int
2326 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2327 {
2328         return thread_is_stopping(thread) ||
2329                thread->t_svcpt->scp_service->srv_is_stopping;
2330 }
2331
2332 static inline int
2333 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2334 {
2335         return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2336                svcpt->scp_rqbd_timeout == 0;
2337 }
2338
2339 static inline int
2340 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2341 {
2342         return svcpt->scp_at_check;
2343 }
2344
2345 /**
2346  * requests wait on preprocessing
2347  * user can call it w/o any lock but need to hold
2348  * ptlrpc_service_part::scp_lock to get reliable result
2349  */
2350 static inline int
2351 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2352 {
2353         return !cfs_list_empty(&svcpt->scp_req_incoming);
2354 }
2355
2356 static __attribute__((__noinline__)) int
2357 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2358                   struct ptlrpc_thread *thread)
2359 {
2360         /* Don't exit while there are replies to be handled */
2361         struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2362                                              ptlrpc_retry_rqbds, svcpt);
2363
2364         lc_watchdog_disable(thread->t_watchdog);
2365
2366         cfs_cond_resched();
2367
2368         l_wait_event_exclusive_head(svcpt->scp_waitq,
2369                                 ptlrpc_thread_stopping(thread) ||
2370                                 ptlrpc_server_request_incoming(svcpt) ||
2371                                 ptlrpc_server_request_pending(svcpt, false) ||
2372                                 ptlrpc_rqbd_pending(svcpt) ||
2373                                 ptlrpc_at_check(svcpt), &lwi);
2374
2375         if (ptlrpc_thread_stopping(thread))
2376                 return -EINTR;
2377
2378         lc_watchdog_touch(thread->t_watchdog,
2379                           ptlrpc_server_get_timeout(svcpt));
2380         return 0;
2381 }
2382
2383 /**
2384  * Main thread body for service threads.
2385  * Waits in a loop waiting for new requests to process to appear.
2386  * Every time an incoming requests is added to its queue, a waitq
2387  * is woken up and one of the threads will handle it.
2388  */
2389 static int ptlrpc_main(void *arg)
2390 {
2391         struct ptlrpc_thread            *thread = (struct ptlrpc_thread *)arg;
2392         struct ptlrpc_service_part      *svcpt = thread->t_svcpt;
2393         struct ptlrpc_service           *svc = svcpt->scp_service;
2394         struct ptlrpc_reply_state       *rs;
2395 #ifdef WITH_GROUP_INFO
2396         cfs_group_info_t *ginfo = NULL;
2397 #endif
2398         struct lu_env *env;
2399         int counter = 0, rc = 0;
2400         ENTRY;
2401
2402         thread->t_pid = cfs_curproc_pid();
2403         cfs_daemonize_ctxt(thread->t_name);
2404
2405         /* NB: we will call cfs_cpt_bind() for all threads, because we
2406          * might want to run lustre server only on a subset of system CPUs,
2407          * in that case ->scp_cpt is CFS_CPT_ANY */
2408         rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2409         if (rc != 0) {
2410                 CWARN("%s: failed to bind %s on CPT %d\n",
2411                       svc->srv_name, thread->t_name, svcpt->scp_cpt);
2412         }
2413
2414 #ifdef WITH_GROUP_INFO
2415         ginfo = cfs_groups_alloc(0);
2416         if (!ginfo) {
2417                 rc = -ENOMEM;
2418                 goto out;
2419         }
2420
2421         cfs_set_current_groups(ginfo);
2422         cfs_put_group_info(ginfo);
2423 #endif
2424
2425         if (svc->srv_ops.so_thr_init != NULL) {
2426                 rc = svc->srv_ops.so_thr_init(thread);
2427                 if (rc)
2428                         goto out;
2429         }
2430
2431         OBD_ALLOC_PTR(env);
2432         if (env == NULL) {
2433                 rc = -ENOMEM;
2434                 goto out_srv_fini;
2435         }
2436
2437         rc = lu_context_init(&env->le_ctx,
2438                              svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2439         if (rc)
2440                 goto out_srv_fini;
2441
2442         thread->t_env = env;
2443         env->le_ctx.lc_thread = thread;
2444         env->le_ctx.lc_cookie = 0x6;
2445
2446         while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2447                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2448                 if (rc >= 0)
2449                         continue;
2450
2451                 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2452                         svc->srv_name, svcpt->scp_cpt, rc);
2453                 goto out_srv_fini;
2454         }
2455
2456         /* Alloc reply state structure for this one */
2457         OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2458         if (!rs) {
2459                 rc = -ENOMEM;
2460                 goto out_srv_fini;
2461         }
2462
2463         spin_lock(&svcpt->scp_lock);
2464
2465         LASSERT(thread_is_starting(thread));
2466         thread_clear_flags(thread, SVC_STARTING);
2467
2468         LASSERT(svcpt->scp_nthrs_starting == 1);
2469         svcpt->scp_nthrs_starting--;
2470
2471         /* SVC_STOPPING may already be set here if someone else is trying
2472          * to stop the service while this new thread has been dynamically
2473          * forked. We still set SVC_RUNNING to let our creator know that
2474          * we are now running, however we will exit as soon as possible */
2475         thread_add_flags(thread, SVC_RUNNING);
2476         svcpt->scp_nthrs_running++;
2477         spin_unlock(&svcpt->scp_lock);
2478
2479         /* wake up our creator in case he's still waiting. */
2480         cfs_waitq_signal(&thread->t_ctl_waitq);
2481
2482         thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2483                                              NULL, NULL);
2484
2485         spin_lock(&svcpt->scp_rep_lock);
2486         cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2487         cfs_waitq_signal(&svcpt->scp_rep_waitq);
2488         spin_unlock(&svcpt->scp_rep_lock);
2489
2490         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2491                svcpt->scp_nthrs_running);
2492
2493         /* XXX maintain a list of all managed devices: insert here */
2494         while (!ptlrpc_thread_stopping(thread)) {
2495                 if (ptlrpc_wait_event(svcpt, thread))
2496                         break;
2497
2498                 ptlrpc_check_rqbd_pool(svcpt);
2499
2500                 if (ptlrpc_threads_need_create(svcpt)) {
2501                         /* Ignore return code - we tried... */
2502                         ptlrpc_start_thread(svcpt, 0);
2503                 }
2504
2505                 /* Process all incoming reqs before handling any */
2506                 if (ptlrpc_server_request_incoming(svcpt)) {
2507                         lu_context_enter(&env->le_ctx);
2508                         ptlrpc_server_handle_req_in(svcpt, thread);
2509                         lu_context_exit(&env->le_ctx);
2510
2511                         /* but limit ourselves in case of flood */
2512                         if (counter++ < 100)
2513                                 continue;
2514                         counter = 0;
2515                 }
2516
2517                 if (ptlrpc_at_check(svcpt))
2518                         ptlrpc_at_check_timed(svcpt);
2519
2520                 if (ptlrpc_server_request_pending(svcpt, false)) {
2521                         lu_context_enter(&env->le_ctx);
2522                         ptlrpc_server_handle_request(svcpt, thread);
2523                         lu_context_exit(&env->le_ctx);
2524                 }
2525
2526                 if (ptlrpc_rqbd_pending(svcpt) &&
2527                     ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2528                         /* I just failed to repost request buffers.
2529                          * Wait for a timeout (unless something else
2530                          * happens) before I try again */
2531                         svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2532                         CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2533                                svcpt->scp_nrqbds_posted);
2534                 }
2535         }
2536
2537         lc_watchdog_delete(thread->t_watchdog);
2538         thread->t_watchdog = NULL;
2539
2540 out_srv_fini:
2541         /*
2542          * deconstruct service specific state created by ptlrpc_start_thread()
2543          */
2544         if (svc->srv_ops.so_thr_done != NULL)
2545                 svc->srv_ops.so_thr_done(thread);
2546
2547         if (env != NULL) {
2548                 lu_context_fini(&env->le_ctx);
2549                 OBD_FREE_PTR(env);
2550         }
2551 out:
2552         CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2553                thread, thread->t_pid, thread->t_id, rc);
2554
2555         spin_lock(&svcpt->scp_lock);
2556         if (thread_test_and_clear_flags(thread, SVC_STARTING))
2557                 svcpt->scp_nthrs_starting--;
2558
2559         if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2560                 /* must know immediately */
2561                 svcpt->scp_nthrs_running--;
2562         }
2563
2564         thread->t_id = rc;
2565         thread_add_flags(thread, SVC_STOPPED);
2566
2567         cfs_waitq_signal(&thread->t_ctl_waitq);
2568         spin_unlock(&svcpt->scp_lock);
2569
2570         return rc;
2571 }
2572
2573 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2574                           cfs_list_t *replies)
2575 {
2576         int result;
2577
2578         spin_lock(&hrt->hrt_lock);
2579
2580         cfs_list_splice_init(&hrt->hrt_queue, replies);
2581         result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2582
2583         spin_unlock(&hrt->hrt_lock);
2584         return result;
2585 }
2586
2587 /**
2588  * Main body of "handle reply" function.
2589  * It processes acked reply states
2590  */
2591 static int ptlrpc_hr_main(void *arg)
2592 {
2593         struct ptlrpc_hr_thread         *hrt = (struct ptlrpc_hr_thread *)arg;
2594         struct ptlrpc_hr_partition      *hrp = hrt->hrt_partition;
2595         CFS_LIST_HEAD                   (replies);
2596         char                            threadname[20];
2597         int                             rc;
2598
2599         snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2600                  hrp->hrp_cpt, hrt->hrt_id);
2601         cfs_daemonize_ctxt(threadname);
2602
2603         rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2604         if (rc != 0) {
2605                 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2606                       threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2607         }
2608
2609         cfs_atomic_inc(&hrp->hrp_nstarted);
2610         cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2611
2612         while (!ptlrpc_hr.hr_stopping) {
2613                 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2614
2615                 while (!cfs_list_empty(&replies)) {
2616                         struct ptlrpc_reply_state *rs;
2617
2618                         rs = cfs_list_entry(replies.prev,
2619                                             struct ptlrpc_reply_state,
2620                                             rs_list);
2621                         cfs_list_del_init(&rs->rs_list);
2622                         ptlrpc_handle_rs(rs);
2623                 }
2624         }
2625
2626         cfs_atomic_inc(&hrp->hrp_nstopped);
2627         cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2628
2629         return 0;
2630 }
2631
2632 static void ptlrpc_stop_hr_threads(void)
2633 {
2634         struct ptlrpc_hr_partition      *hrp;
2635         int                             i;
2636         int                             j;
2637
2638         ptlrpc_hr.hr_stopping = 1;
2639
2640         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2641                 if (hrp->hrp_thrs == NULL)
2642                         continue; /* uninitialized */
2643                 for (j = 0; j < hrp->hrp_nthrs; j++)
2644                         cfs_waitq_broadcast(&hrp->hrp_thrs[j].hrt_waitq);
2645         }
2646
2647         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2648                 if (hrp->hrp_thrs == NULL)
2649                         continue; /* uninitialized */
2650                 cfs_wait_event(ptlrpc_hr.hr_waitq,
2651                                cfs_atomic_read(&hrp->hrp_nstopped) ==
2652                                cfs_atomic_read(&hrp->hrp_nstarted));
2653         }
2654 }
2655
2656 static int ptlrpc_start_hr_threads(void)
2657 {
2658         struct ptlrpc_hr_partition      *hrp;
2659         int                             i;
2660         int                             j;
2661         ENTRY;
2662
2663         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2664                 int     rc = 0;
2665
2666                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2667                         rc = cfs_create_thread(ptlrpc_hr_main,
2668                                                &hrp->hrp_thrs[j],
2669                                                CLONE_VM | CLONE_FILES);
2670                         if (rc < 0)
2671                                 break;
2672                 }
2673                 cfs_wait_event(ptlrpc_hr.hr_waitq,
2674                                cfs_atomic_read(&hrp->hrp_nstarted) == j);
2675                 if (rc >= 0)
2676                         continue;
2677
2678                 CERROR("Reply handling thread %d:%d Failed on starting: "
2679                        "rc = %d\n", i, j, rc);
2680                 ptlrpc_stop_hr_threads();
2681                 RETURN(rc);
2682         }
2683         RETURN(0);
2684 }
2685
2686 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2687 {
2688         struct l_wait_info      lwi = { 0 };
2689         struct ptlrpc_thread    *thread;
2690         CFS_LIST_HEAD           (zombie);
2691
2692         ENTRY;
2693
2694         CDEBUG(D_INFO, "Stopping threads for service %s\n",
2695                svcpt->scp_service->srv_name);
2696
2697         spin_lock(&svcpt->scp_lock);
2698         /* let the thread know that we would like it to stop asap */
2699         list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2700                 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2701                        svcpt->scp_service->srv_thread_name, thread->t_id);
2702                 thread_add_flags(thread, SVC_STOPPING);
2703         }
2704
2705         cfs_waitq_broadcast(&svcpt->scp_waitq);
2706
2707         while (!cfs_list_empty(&svcpt->scp_threads)) {
2708                 thread = cfs_list_entry(svcpt->scp_threads.next,
2709                                         struct ptlrpc_thread, t_link);
2710                 if (thread_is_stopped(thread)) {
2711                         cfs_list_del(&thread->t_link);
2712                         cfs_list_add(&thread->t_link, &zombie);
2713                         continue;
2714                 }
2715                 spin_unlock(&svcpt->scp_lock);
2716
2717                 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2718                        svcpt->scp_service->srv_thread_name, thread->t_id);
2719                 l_wait_event(thread->t_ctl_waitq,
2720                              thread_is_stopped(thread), &lwi);
2721
2722                 spin_lock(&svcpt->scp_lock);
2723         }
2724
2725         spin_unlock(&svcpt->scp_lock);
2726
2727         while (!cfs_list_empty(&zombie)) {
2728                 thread = cfs_list_entry(zombie.next,
2729                                         struct ptlrpc_thread, t_link);
2730                 cfs_list_del(&thread->t_link);
2731                 OBD_FREE_PTR(thread);
2732         }
2733         EXIT;
2734 }
2735
2736 /**
2737  * Stops all threads of a particular service \a svc
2738  */
2739 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2740 {
2741         struct ptlrpc_service_part *svcpt;
2742         int                        i;
2743         ENTRY;
2744
2745         ptlrpc_service_for_each_part(svcpt, i, svc) {
2746                 if (svcpt->scp_service != NULL)
2747                         ptlrpc_svcpt_stop_threads(svcpt);
2748         }
2749
2750         EXIT;
2751 }
2752 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2753
2754 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2755 {
2756         int     rc = 0;
2757         int     i;
2758         int     j;
2759         ENTRY;
2760
2761         /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2762         LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2763
2764         for (i = 0; i < svc->srv_ncpts; i++) {
2765                 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2766                         rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2767                         if (rc == 0)
2768                                 continue;
2769
2770                         if (rc != -EMFILE)
2771                                 goto failed;
2772                         /* We have enough threads, don't start more. b=15759 */
2773                         break;
2774                 }
2775         }
2776
2777         RETURN(0);
2778  failed:
2779         CERROR("cannot start %s thread #%d_%d: rc %d\n",
2780                svc->srv_thread_name, i, j, rc);
2781         ptlrpc_stop_all_threads(svc);
2782         RETURN(rc);
2783 }
2784 EXPORT_SYMBOL(ptlrpc_start_threads);
2785
2786 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2787 {
2788         struct l_wait_info      lwi = { 0 };
2789         struct ptlrpc_thread    *thread;
2790         struct ptlrpc_service   *svc;
2791         int                     rc;
2792         ENTRY;
2793
2794         LASSERT(svcpt != NULL);
2795
2796         svc = svcpt->scp_service;
2797
2798         CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2799                svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2800                svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2801
2802  again:
2803         if (unlikely(svc->srv_is_stopping))
2804                 RETURN(-ESRCH);
2805
2806         if (!ptlrpc_threads_increasable(svcpt) ||
2807             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2808              svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2809                 RETURN(-EMFILE);
2810
2811         OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2812         if (thread == NULL)
2813                 RETURN(-ENOMEM);
2814         cfs_waitq_init(&thread->t_ctl_waitq);
2815
2816         spin_lock(&svcpt->scp_lock);
2817         if (!ptlrpc_threads_increasable(svcpt)) {
2818                 spin_unlock(&svcpt->scp_lock);
2819                 OBD_FREE_PTR(thread);
2820                 RETURN(-EMFILE);
2821         }
2822
2823         if (svcpt->scp_nthrs_starting != 0) {
2824                 /* serialize starting because some modules (obdfilter)
2825                  * might require unique and contiguous t_id */
2826                 LASSERT(svcpt->scp_nthrs_starting == 1);
2827                 spin_unlock(&svcpt->scp_lock);
2828                 OBD_FREE_PTR(thread);
2829                 if (wait) {
2830                         CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2831                                svc->srv_thread_name, svcpt->scp_thr_nextid);
2832                         cfs_schedule();
2833                         goto again;
2834                 }
2835
2836                 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2837                        svc->srv_thread_name, svcpt->scp_thr_nextid);
2838                 RETURN(-EAGAIN);
2839         }
2840
2841         svcpt->scp_nthrs_starting++;
2842         thread->t_id = svcpt->scp_thr_nextid++;
2843         thread_add_flags(thread, SVC_STARTING);
2844         thread->t_svcpt = svcpt;
2845
2846         cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2847         spin_unlock(&svcpt->scp_lock);
2848
2849         if (svcpt->scp_cpt >= 0) {
2850                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2851                          svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2852         } else {
2853                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2854                          svc->srv_thread_name, thread->t_id);
2855         }
2856
2857         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2858         /*
2859          * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2860          * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2861          */
2862         rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS);
2863         if (rc < 0) {
2864                 CERROR("cannot start thread '%s': rc %d\n",
2865                        thread->t_name, rc);
2866                 spin_lock(&svcpt->scp_lock);
2867                 cfs_list_del(&thread->t_link);
2868                 --svcpt->scp_nthrs_starting;
2869                 spin_unlock(&svcpt->scp_lock);
2870
2871                 OBD_FREE(thread, sizeof(*thread));
2872                 RETURN(rc);
2873         }
2874
2875         if (!wait)
2876                 RETURN(0);
2877
2878         l_wait_event(thread->t_ctl_waitq,
2879                      thread_is_running(thread) || thread_is_stopped(thread),
2880                      &lwi);
2881
2882         rc = thread_is_stopped(thread) ? thread->t_id : 0;
2883         RETURN(rc);
2884 }
2885
2886 int ptlrpc_hr_init(void)
2887 {
2888         struct ptlrpc_hr_partition      *hrp;
2889         struct ptlrpc_hr_thread         *hrt;
2890         int                             rc;
2891         int                             i;
2892         int                             j;
2893         ENTRY;
2894
2895         memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2896         ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2897
2898         ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2899                                                    sizeof(*hrp));
2900         if (ptlrpc_hr.hr_partitions == NULL)
2901                 RETURN(-ENOMEM);
2902
2903         cfs_waitq_init(&ptlrpc_hr.hr_waitq);
2904
2905         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2906                 hrp->hrp_cpt = i;
2907
2908                 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2909                 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2910
2911                 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2912                 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2913
2914                 LASSERT(hrp->hrp_nthrs > 0);
2915                 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2916                               hrp->hrp_nthrs * sizeof(*hrt));
2917                 if (hrp->hrp_thrs == NULL)
2918                         GOTO(out, rc = -ENOMEM);
2919
2920                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2921                         hrt = &hrp->hrp_thrs[j];
2922
2923                         hrt->hrt_id = j;
2924                         hrt->hrt_partition = hrp;
2925                         cfs_waitq_init(&hrt->hrt_waitq);
2926                         spin_lock_init(&hrt->hrt_lock);
2927                         CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
2928                 }
2929         }
2930
2931         rc = ptlrpc_start_hr_threads();
2932 out:
2933         if (rc != 0)
2934                 ptlrpc_hr_fini();
2935         RETURN(rc);
2936 }
2937
2938 void ptlrpc_hr_fini(void)
2939 {
2940         struct ptlrpc_hr_partition      *hrp;
2941         int                             i;
2942
2943         if (ptlrpc_hr.hr_partitions == NULL)
2944                 return;
2945
2946         ptlrpc_stop_hr_threads();
2947
2948         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2949                 if (hrp->hrp_thrs != NULL) {
2950                         OBD_FREE(hrp->hrp_thrs,
2951                                  hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2952                 }
2953         }
2954
2955         cfs_percpt_free(ptlrpc_hr.hr_partitions);
2956         ptlrpc_hr.hr_partitions = NULL;
2957 }
2958
2959 #endif /* __KERNEL__ */
2960
2961 /**
2962  * Wait until all already scheduled replies are processed.
2963  */
2964 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2965 {
2966         while (1) {
2967                 int rc;
2968                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2969                                                      NULL, NULL);
2970
2971                 rc = l_wait_event(svcpt->scp_waitq,
2972                      cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2973                 if (rc == 0)
2974                         break;
2975                 CWARN("Unexpectedly long timeout %s %p\n",
2976                       svcpt->scp_service->srv_name, svcpt->scp_service);
2977         }
2978 }
2979
2980 static void
2981 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2982 {
2983         struct ptlrpc_service_part      *svcpt;
2984         int                             i;
2985
2986         /* early disarm AT timer... */
2987         ptlrpc_service_for_each_part(svcpt, i, svc) {
2988                 if (svcpt->scp_service != NULL)
2989                         cfs_timer_disarm(&svcpt->scp_at_timer);
2990         }
2991 }
2992
2993 static void
2994 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2995 {
2996         struct ptlrpc_service_part        *svcpt;
2997         struct ptlrpc_request_buffer_desc *rqbd;
2998         struct l_wait_info                lwi;
2999         int                               rc;
3000         int                               i;
3001
3002         /* All history will be culled when the next request buffer is
3003          * freed in ptlrpc_service_purge_all() */
3004         svc->srv_hist_nrqbds_cpt_max = 0;
3005
3006         rc = LNetClearLazyPortal(svc->srv_req_portal);
3007         LASSERT(rc == 0);
3008
3009         ptlrpc_service_for_each_part(svcpt, i, svc) {
3010                 if (svcpt->scp_service == NULL)
3011                         break;
3012
3013                 /* Unlink all the request buffers.  This forces a 'final'
3014                  * event with its 'unlink' flag set for each posted rqbd */
3015                 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3016                                         rqbd_list) {
3017                         rc = LNetMDUnlink(rqbd->rqbd_md_h);
3018                         LASSERT(rc == 0 || rc == -ENOENT);
3019                 }
3020         }
3021
3022         ptlrpc_service_for_each_part(svcpt, i, svc) {
3023                 if (svcpt->scp_service == NULL)
3024                         break;
3025
3026                 /* Wait for the network to release any buffers
3027                  * it's currently filling */
3028                 spin_lock(&svcpt->scp_lock);
3029                 while (svcpt->scp_nrqbds_posted != 0) {
3030                         spin_unlock(&svcpt->scp_lock);
3031                         /* Network access will complete in finite time but
3032                          * the HUGE timeout lets us CWARN for visibility
3033                          * of sluggish NALs */
3034                         lwi = LWI_TIMEOUT_INTERVAL(
3035                                         cfs_time_seconds(LONG_UNLINK),
3036                                         cfs_time_seconds(1), NULL, NULL);
3037                         rc = l_wait_event(svcpt->scp_waitq,
3038                                           svcpt->scp_nrqbds_posted == 0, &lwi);
3039                         if (rc == -ETIMEDOUT) {
3040                                 CWARN("Service %s waiting for "
3041                                       "request buffers\n",
3042                                       svcpt->scp_service->srv_name);
3043                         }
3044                         spin_lock(&svcpt->scp_lock);
3045                 }
3046                 spin_unlock(&svcpt->scp_lock);
3047         }
3048 }
3049
3050 static void
3051 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3052 {
3053         struct ptlrpc_service_part              *svcpt;
3054         struct ptlrpc_request_buffer_desc       *rqbd;
3055         struct ptlrpc_request                   *req;
3056         struct ptlrpc_reply_state               *rs;
3057         int                                     i;
3058
3059         ptlrpc_service_for_each_part(svcpt, i, svc) {
3060                 if (svcpt->scp_service == NULL)
3061                         break;
3062
3063                 spin_lock(&svcpt->scp_rep_lock);
3064                 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
3065                         rs = cfs_list_entry(svcpt->scp_rep_active.next,
3066                                             struct ptlrpc_reply_state, rs_list);
3067                         spin_lock(&rs->rs_lock);
3068                         ptlrpc_schedule_difficult_reply(rs);
3069                         spin_unlock(&rs->rs_lock);
3070                 }
3071                 spin_unlock(&svcpt->scp_rep_lock);
3072
3073                 /* purge the request queue.  NB No new replies (rqbds
3074                  * all unlinked) and no service threads, so I'm the only
3075                  * thread noodling the request queue now */
3076                 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
3077                         req = cfs_list_entry(svcpt->scp_req_incoming.next,
3078                                              struct ptlrpc_request, rq_list);
3079
3080                         cfs_list_del(&req->rq_list);
3081                         svcpt->scp_nreqs_incoming--;
3082                         ptlrpc_server_finish_request(svcpt, req);
3083                 }
3084
3085                 while (ptlrpc_server_request_pending(svcpt, true)) {
3086                         req = ptlrpc_server_request_get(svcpt, true);
3087                         ptlrpc_server_finish_active_request(svcpt, req);
3088                 }
3089
3090                 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3091                 LASSERT(svcpt->scp_nreqs_incoming == 0);
3092                 LASSERT(svcpt->scp_nreqs_active == 0);
3093                 /* history should have been culled by
3094                  * ptlrpc_server_finish_request */
3095                 LASSERT(svcpt->scp_hist_nrqbds == 0);
3096
3097                 /* Now free all the request buffers since nothing
3098                  * references them any more... */
3099
3100                 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3101                         rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3102                                               struct ptlrpc_request_buffer_desc,
3103                                               rqbd_list);
3104                         ptlrpc_free_rqbd(rqbd);
3105                 }
3106                 ptlrpc_wait_replies(svcpt);
3107
3108                 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3109                         rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3110                                             struct ptlrpc_reply_state,
3111                                             rs_list);
3112                         cfs_list_del(&rs->rs_list);
3113                         OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3114                 }
3115         }
3116 }
3117
3118 static void
3119 ptlrpc_service_free(struct ptlrpc_service *svc)
3120 {
3121         struct ptlrpc_service_part      *svcpt;
3122         struct ptlrpc_at_array          *array;
3123         int                             i;
3124
3125         ptlrpc_service_for_each_part(svcpt, i, svc) {
3126                 if (svcpt->scp_service == NULL)
3127                         break;
3128
3129                 /* In case somebody rearmed this in the meantime */
3130                 cfs_timer_disarm(&svcpt->scp_at_timer);
3131                 array = &svcpt->scp_at_array;
3132
3133                 if (array->paa_reqs_array != NULL) {
3134                         OBD_FREE(array->paa_reqs_array,
3135                                  sizeof(cfs_list_t) * array->paa_size);
3136                         array->paa_reqs_array = NULL;
3137                 }
3138
3139                 if (array->paa_reqs_count != NULL) {
3140                         OBD_FREE(array->paa_reqs_count,
3141                                  sizeof(__u32) * array->paa_size);
3142                         array->paa_reqs_count = NULL;
3143                 }
3144         }
3145
3146         ptlrpc_service_for_each_part(svcpt, i, svc)
3147                 OBD_FREE_PTR(svcpt);
3148
3149         if (svc->srv_cpts != NULL)
3150                 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3151
3152         OBD_FREE(svc, offsetof(struct ptlrpc_service,
3153                                srv_parts[svc->srv_ncpts]));
3154 }
3155
3156 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3157 {
3158         ENTRY;
3159
3160         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3161
3162         service->srv_is_stopping = 1;
3163
3164         mutex_lock(&ptlrpc_all_services_mutex);
3165         cfs_list_del_init(&service->srv_list);
3166         mutex_unlock(&ptlrpc_all_services_mutex);
3167
3168         ptlrpc_service_del_atimer(service);
3169         ptlrpc_stop_all_threads(service);
3170
3171         ptlrpc_service_unlink_rqbd(service);
3172         ptlrpc_service_purge_all(service);
3173         ptlrpc_service_nrs_cleanup(service);
3174
3175         ptlrpc_lprocfs_unregister_service(service);
3176
3177         ptlrpc_service_free(service);
3178
3179         RETURN(0);
3180 }
3181 EXPORT_SYMBOL(ptlrpc_unregister_service);
3182
3183 /**
3184  * Returns 0 if the service is healthy.
3185  *
3186  * Right now, it just checks to make sure that requests aren't languishing
3187  * in the queue.  We'll use this health check to govern whether a node needs
3188  * to be shot, so it's intentionally non-aggressive. */
3189 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3190 {
3191         struct ptlrpc_request           *request = NULL;
3192         struct timeval                  right_now;
3193         long                            timediff;
3194
3195         cfs_gettimeofday(&right_now);
3196
3197         spin_lock(&svcpt->scp_req_lock);
3198         /* How long has the next entry been waiting? */
3199         if (ptlrpc_server_high_pending(svcpt, true))
3200                 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3201         else if (ptlrpc_server_normal_pending(svcpt, true))
3202                 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3203
3204         if (request == NULL) {
3205                 spin_unlock(&svcpt->scp_req_lock);
3206                 return 0;
3207         }
3208
3209         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3210         spin_unlock(&svcpt->scp_req_lock);
3211
3212         if ((timediff / ONE_MILLION) >
3213             (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3214                 CERROR("%s: unhealthy - request has been waiting %lds\n",
3215                        svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3216                 return -1;
3217         }
3218
3219         return 0;
3220 }
3221
3222 int
3223 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3224 {
3225         struct ptlrpc_service_part      *svcpt;
3226         int                             i;
3227
3228         if (svc == NULL)
3229                 return 0;
3230
3231         ptlrpc_service_for_each_part(svcpt, i, svc) {
3232                 int rc = ptlrpc_svcpt_health_check(svcpt);
3233
3234                 if (rc != 0)
3235                         return rc;
3236         }
3237         return 0;
3238 }
3239 EXPORT_SYMBOL(ptlrpc_service_health_check);