Whamcloud - gitweb
LU-12368 ptlrpc: make DEBUG_REQ messages consistent
[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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2010, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 #define DEBUG_SUBSYSTEM S_RPC
34
35 #include <linux/kthread.h>
36 #include <linux/ratelimit.h>
37
38 #include <obd_support.h>
39 #include <obd_class.h>
40 #include <lustre_net.h>
41 #include <lu_object.h>
42 #include <uapi/linux/lnet/lnet-types.h>
43 #include "ptlrpc_internal.h"
44
45 /* The following are visible and mutable through /sys/module/ptlrpc */
46 int test_req_buffer_pressure = 0;
47 module_param(test_req_buffer_pressure, int, 0444);
48 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
49 module_param(at_min, int, 0644);
50 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
51 module_param(at_max, int, 0644);
52 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
53 module_param(at_history, int, 0644);
54 MODULE_PARM_DESC(at_history,
55                  "Adaptive timeouts remember the slowest event that took place within this period (sec)");
56 module_param(at_early_margin, int, 0644);
57 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
58 module_param(at_extra, int, 0644);
59 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
60
61 /* forward ref */
62 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
63 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
64 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
65
66 /** Holds a list of all PTLRPC services */
67 struct list_head ptlrpc_all_services;
68 /** Used to protect the \e ptlrpc_all_services list */
69 struct mutex ptlrpc_all_services_mutex;
70
71 static struct ptlrpc_request_buffer_desc *
72 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
73 {
74         struct ptlrpc_service             *svc = svcpt->scp_service;
75         struct ptlrpc_request_buffer_desc *rqbd;
76
77         OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
78         if (rqbd == NULL)
79                 return NULL;
80
81         rqbd->rqbd_svcpt = svcpt;
82         rqbd->rqbd_refcount = 0;
83         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
84         rqbd->rqbd_cbid.cbid_arg = rqbd;
85         INIT_LIST_HEAD(&rqbd->rqbd_reqs);
86         OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
87                             svcpt->scp_cpt, svc->srv_buf_size);
88         if (rqbd->rqbd_buffer == NULL) {
89                 OBD_FREE_PTR(rqbd);
90                 return NULL;
91         }
92
93         spin_lock(&svcpt->scp_lock);
94         list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
95         svcpt->scp_nrqbds_total++;
96         spin_unlock(&svcpt->scp_lock);
97
98         return rqbd;
99 }
100
101 static void ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
102 {
103         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
104
105         LASSERT(rqbd->rqbd_refcount == 0);
106         LASSERT(list_empty(&rqbd->rqbd_reqs));
107
108         spin_lock(&svcpt->scp_lock);
109         list_del(&rqbd->rqbd_list);
110         svcpt->scp_nrqbds_total--;
111         spin_unlock(&svcpt->scp_lock);
112
113         OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
114         OBD_FREE_PTR(rqbd);
115 }
116
117 static int ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
118 {
119         struct ptlrpc_service *svc = svcpt->scp_service;
120         struct ptlrpc_request_buffer_desc *rqbd;
121         int rc = 0;
122         int i;
123
124         if (svcpt->scp_rqbd_allocating)
125                 goto try_post;
126
127         spin_lock(&svcpt->scp_lock);
128         /* check again with lock */
129         if (svcpt->scp_rqbd_allocating) {
130                 /* NB: we might allow more than one thread in the future */
131                 LASSERT(svcpt->scp_rqbd_allocating == 1);
132                 spin_unlock(&svcpt->scp_lock);
133                 goto try_post;
134         }
135
136         svcpt->scp_rqbd_allocating++;
137         spin_unlock(&svcpt->scp_lock);
138
139
140         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
141                 /*
142                  * NB: another thread might have recycled enough rqbds, we
143                  * need to make sure it wouldn't over-allocate, see LU-1212.
144                  */
145                 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
146                     (svc->srv_nrqbds_max != 0 &&
147                      svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
148                         break;
149
150                 rqbd = ptlrpc_alloc_rqbd(svcpt);
151
152                 if (rqbd == NULL) {
153                         CERROR("%s: Can't allocate request buffer\n",
154                                svc->srv_name);
155                         rc = -ENOMEM;
156                         break;
157                 }
158         }
159
160         spin_lock(&svcpt->scp_lock);
161
162         LASSERT(svcpt->scp_rqbd_allocating == 1);
163         svcpt->scp_rqbd_allocating--;
164
165         spin_unlock(&svcpt->scp_lock);
166
167         CDEBUG(D_RPCTRACE,
168                "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
169                svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
170                svcpt->scp_nrqbds_total, rc);
171
172  try_post:
173         if (post && rc == 0)
174                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
175
176         return rc;
177 }
178
179 /**
180  * Part of Rep-Ack logic.
181  * Puts a lock and its mode into reply state assotiated to request reply.
182  */
183 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
184                       int mode, bool no_ack, bool convert_lock)
185 {
186         struct ptlrpc_reply_state *rs = req->rq_reply_state;
187         int idx;
188
189         LASSERT(rs != NULL);
190         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
191
192         idx = rs->rs_nlocks++;
193         rs->rs_locks[idx] = *lock;
194         rs->rs_modes[idx] = mode;
195         rs->rs_difficult = 1;
196         rs->rs_no_ack = no_ack;
197         rs->rs_convert_lock = convert_lock;
198 }
199 EXPORT_SYMBOL(ptlrpc_save_lock);
200
201
202 struct ptlrpc_hr_partition;
203
204 struct ptlrpc_hr_thread {
205         int                             hrt_id;         /* thread ID */
206         spinlock_t                      hrt_lock;
207         wait_queue_head_t               hrt_waitq;
208         struct list_head                hrt_queue;
209         struct ptlrpc_hr_partition      *hrt_partition;
210 };
211
212 struct ptlrpc_hr_partition {
213         /* # of started threads */
214         atomic_t                        hrp_nstarted;
215         /* # of stopped threads */
216         atomic_t                        hrp_nstopped;
217         /* cpu partition id */
218         int                             hrp_cpt;
219         /* round-robin rotor for choosing thread */
220         int                             hrp_rotor;
221         /* total number of threads on this partition */
222         int                             hrp_nthrs;
223         /* threads table */
224         struct ptlrpc_hr_thread         *hrp_thrs;
225 };
226
227 #define HRT_RUNNING 0
228 #define HRT_STOPPING 1
229
230 struct ptlrpc_hr_service {
231         /* CPU partition table, it's just cfs_cpt_table for now */
232         struct cfs_cpt_table            *hr_cpt_table;
233         /** controller sleep waitq */
234         wait_queue_head_t               hr_waitq;
235         unsigned int                    hr_stopping;
236         /** roundrobin rotor for non-affinity service */
237         unsigned int                    hr_rotor;
238         /* partition data */
239         struct ptlrpc_hr_partition      **hr_partitions;
240 };
241
242 struct rs_batch {
243         struct list_head                        rsb_replies;
244         unsigned int                    rsb_n_replies;
245         struct ptlrpc_service_part      *rsb_svcpt;
246 };
247
248 /** reply handling service. */
249 static struct ptlrpc_hr_service         ptlrpc_hr;
250
251 /**
252  * maximum mumber of replies scheduled in one batch
253  */
254 #define MAX_SCHEDULED 256
255
256 /**
257  * Initialize a reply batch.
258  *
259  * \param b batch
260  */
261 static void rs_batch_init(struct rs_batch *b)
262 {
263         memset(b, 0, sizeof(*b));
264         INIT_LIST_HEAD(&b->rsb_replies);
265 }
266
267 /**
268  * Choose an hr thread to dispatch requests to.
269  */
270 static
271 struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
272 {
273         struct ptlrpc_hr_partition      *hrp;
274         unsigned int                    rotor;
275
276         if (svcpt->scp_cpt >= 0 &&
277             svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
278                 /* directly match partition */
279                 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
280
281         } else {
282                 rotor = ptlrpc_hr.hr_rotor++;
283                 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
284
285                 hrp = ptlrpc_hr.hr_partitions[rotor];
286         }
287
288         rotor = hrp->hrp_rotor++;
289         return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
290 }
291
292 /**
293  * Dispatch all replies accumulated in the batch to one from
294  * dedicated reply handling threads.
295  *
296  * \param b batch
297  */
298 static void rs_batch_dispatch(struct rs_batch *b)
299 {
300         if (b->rsb_n_replies != 0) {
301                 struct ptlrpc_hr_thread *hrt;
302
303                 hrt = ptlrpc_hr_select(b->rsb_svcpt);
304
305                 spin_lock(&hrt->hrt_lock);
306                 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
307                 spin_unlock(&hrt->hrt_lock);
308
309                 wake_up(&hrt->hrt_waitq);
310                 b->rsb_n_replies = 0;
311         }
312 }
313
314 /**
315  * Add a reply to a batch.
316  * Add one reply object to a batch, schedule batched replies if overload.
317  *
318  * \param b batch
319  * \param rs reply
320  */
321 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
322 {
323         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
324
325         if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
326                 if (b->rsb_svcpt != NULL) {
327                         rs_batch_dispatch(b);
328                         spin_unlock(&b->rsb_svcpt->scp_rep_lock);
329                 }
330                 spin_lock(&svcpt->scp_rep_lock);
331                 b->rsb_svcpt = svcpt;
332         }
333         spin_lock(&rs->rs_lock);
334         rs->rs_scheduled_ever = 1;
335         if (rs->rs_scheduled == 0) {
336                 list_move(&rs->rs_list, &b->rsb_replies);
337                 rs->rs_scheduled = 1;
338                 b->rsb_n_replies++;
339         }
340         rs->rs_committed = 1;
341         spin_unlock(&rs->rs_lock);
342 }
343
344 /**
345  * Reply batch finalization.
346  * Dispatch remaining replies from the batch
347  * and release remaining spinlock.
348  *
349  * \param b batch
350  */
351 static void rs_batch_fini(struct rs_batch *b)
352 {
353         if (b->rsb_svcpt != NULL) {
354                 rs_batch_dispatch(b);
355                 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
356         }
357 }
358
359 #define DECLARE_RS_BATCH(b)     struct rs_batch b
360
361
362 /**
363  * Put reply state into a queue for processing because we received
364  * ACK from the client
365  */
366 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
367 {
368         struct ptlrpc_hr_thread *hrt;
369
370         ENTRY;
371
372         LASSERT(list_empty(&rs->rs_list));
373
374         hrt = ptlrpc_hr_select(rs->rs_svcpt);
375
376         spin_lock(&hrt->hrt_lock);
377         list_add_tail(&rs->rs_list, &hrt->hrt_queue);
378         spin_unlock(&hrt->hrt_lock);
379
380         wake_up(&hrt->hrt_waitq);
381         EXIT;
382 }
383
384 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
385 {
386         ENTRY;
387
388         assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
389         assert_spin_locked(&rs->rs_lock);
390         LASSERT(rs->rs_difficult);
391         rs->rs_scheduled_ever = 1;  /* flag any notification attempt */
392
393         if (rs->rs_scheduled) {     /* being set up or already notified */
394                 EXIT;
395                 return;
396         }
397
398         rs->rs_scheduled = 1;
399         list_del_init(&rs->rs_list);
400         ptlrpc_dispatch_difficult_reply(rs);
401         EXIT;
402 }
403 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
404
405 void ptlrpc_commit_replies(struct obd_export *exp)
406 {
407         struct ptlrpc_reply_state *rs, *nxt;
408         DECLARE_RS_BATCH(batch);
409
410         ENTRY;
411
412         rs_batch_init(&batch);
413         /*
414          * Find any replies that have been committed and get their service
415          * to attend to complete them.
416          */
417
418         /* CAVEAT EMPTOR: spinlock ordering!!! */
419         spin_lock(&exp->exp_uncommitted_replies_lock);
420         list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
421                                  rs_obd_list) {
422                 LASSERT(rs->rs_difficult);
423                 /* VBR: per-export last_committed */
424                 LASSERT(rs->rs_export);
425                 if (rs->rs_transno <= exp->exp_last_committed) {
426                         list_del_init(&rs->rs_obd_list);
427                         rs_batch_add(&batch, rs);
428                 }
429         }
430         spin_unlock(&exp->exp_uncommitted_replies_lock);
431         rs_batch_fini(&batch);
432         EXIT;
433 }
434
435 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
436 {
437         struct ptlrpc_request_buffer_desc *rqbd;
438         int rc;
439         int posted = 0;
440
441         for (;;) {
442                 spin_lock(&svcpt->scp_lock);
443
444                 if (list_empty(&svcpt->scp_rqbd_idle)) {
445                         spin_unlock(&svcpt->scp_lock);
446                         return posted;
447                 }
448
449                 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
450                                       struct ptlrpc_request_buffer_desc,
451                                       rqbd_list);
452
453                 /* assume we will post successfully */
454                 svcpt->scp_nrqbds_posted++;
455                 list_move(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
456
457                 spin_unlock(&svcpt->scp_lock);
458
459                 rc = ptlrpc_register_rqbd(rqbd);
460                 if (rc != 0)
461                         break;
462
463                 posted = 1;
464         }
465
466         spin_lock(&svcpt->scp_lock);
467
468         svcpt->scp_nrqbds_posted--;
469         list_move_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
470
471         /*
472          * Don't complain if no request buffers are posted right now; LNET
473          * won't drop requests because we set the portal lazy!
474          */
475
476         spin_unlock(&svcpt->scp_lock);
477
478         return -1;
479 }
480
481 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
482 {
483         struct ptlrpc_service_part *svcpt;
484
485         svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
486
487         svcpt->scp_at_check = 1;
488         svcpt->scp_at_checktime = ktime_get();
489         wake_up(&svcpt->scp_waitq);
490 }
491
492 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
493                                          struct ptlrpc_service_conf *conf)
494 {
495         struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
496         unsigned int init;
497         unsigned int total;
498         unsigned int nthrs;
499         int weight;
500
501         /*
502          * Common code for estimating & validating threads number.
503          * CPT affinity service could have percpt thread-pool instead
504          * of a global thread-pool, which means user might not always
505          * get the threads number they give it in conf::tc_nthrs_user
506          * even they did set. It's because we need to validate threads
507          * number for each CPT to guarantee each pool will have enough
508          * threads to keep the service healthy.
509          */
510         init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
511         init = max_t(int, init, tc->tc_nthrs_init);
512
513         /*
514          * NB: please see comments in lustre_lnet.h for definition
515          * details of these members
516          */
517         LASSERT(tc->tc_nthrs_max != 0);
518
519         if (tc->tc_nthrs_user != 0) {
520                 /*
521                  * In case there is a reason to test a service with many
522                  * threads, we give a less strict check here, it can
523                  * be up to 8 * nthrs_max
524                  */
525                 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
526                 nthrs = total / svc->srv_ncpts;
527                 init  = max(init, nthrs);
528                 goto out;
529         }
530
531         total = tc->tc_nthrs_max;
532         if (tc->tc_nthrs_base == 0) {
533                 /*
534                  * don't care about base threads number per partition,
535                  * this is most for non-affinity service
536                  */
537                 nthrs = total / svc->srv_ncpts;
538                 goto out;
539         }
540
541         nthrs = tc->tc_nthrs_base;
542         if (svc->srv_ncpts == 1) {
543                 int     i;
544
545                 /*
546                  * NB: Increase the base number if it's single partition
547                  * and total number of cores/HTs is larger or equal to 4.
548                  * result will always < 2 * nthrs_base
549                  */
550                 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
551                 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
552                             (tc->tc_nthrs_base >> i) != 0; i++)
553                         nthrs += tc->tc_nthrs_base >> i;
554         }
555
556         if (tc->tc_thr_factor != 0) {
557                 int       factor = tc->tc_thr_factor;
558                 const int fade = 4;
559
560                 /*
561                  * User wants to increase number of threads with for
562                  * each CPU core/HT, most likely the factor is larger than
563                  * one thread/core because service threads are supposed to
564                  * be blocked by lock or wait for IO.
565                  */
566                 /*
567                  * Amdahl's law says that adding processors wouldn't give
568                  * a linear increasing of parallelism, so it's nonsense to
569                  * have too many threads no matter how many cores/HTs
570                  * there are.
571                  */
572                 preempt_disable();
573                 if (cpumask_weight
574                     (topology_sibling_cpumask(smp_processor_id())) > 1) {
575                         /* weight is # of HTs */
576                         /* depress thread factor for hyper-thread */
577                         factor = factor - (factor >> 1) + (factor >> 3);
578                 }
579                 preempt_enable();
580
581                 weight = cfs_cpt_weight(svc->srv_cptable, 0);
582
583                 for (; factor > 0 && weight > 0; factor--, weight -= fade)
584                         nthrs += min(weight, fade) * factor;
585         }
586
587         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
588                 nthrs = max(tc->tc_nthrs_base,
589                             tc->tc_nthrs_max / svc->srv_ncpts);
590         }
591  out:
592         nthrs = max(nthrs, tc->tc_nthrs_init);
593         svc->srv_nthrs_cpt_limit = nthrs;
594         svc->srv_nthrs_cpt_init = init;
595
596         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
597                 CDEBUG(D_OTHER,
598                        "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
599                        svc->srv_name, nthrs * svc->srv_ncpts,
600                        tc->tc_nthrs_max);
601         }
602 }
603
604 /**
605  * Initialize percpt data for a service
606  */
607 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
608                                     struct ptlrpc_service_part *svcpt, int cpt)
609 {
610         struct ptlrpc_at_array *array;
611         int size;
612         int index;
613         int rc;
614
615         svcpt->scp_cpt = cpt;
616         INIT_LIST_HEAD(&svcpt->scp_threads);
617
618         /* rqbd and incoming request queue */
619         spin_lock_init(&svcpt->scp_lock);
620         mutex_init(&svcpt->scp_mutex);
621         INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
622         INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
623         INIT_LIST_HEAD(&svcpt->scp_req_incoming);
624         init_waitqueue_head(&svcpt->scp_waitq);
625         /* history request & rqbd list */
626         INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
627         INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
628
629         /* acitve requests and hp requests */
630         spin_lock_init(&svcpt->scp_req_lock);
631
632         /* reply states */
633         spin_lock_init(&svcpt->scp_rep_lock);
634         INIT_LIST_HEAD(&svcpt->scp_rep_active);
635         INIT_LIST_HEAD(&svcpt->scp_rep_idle);
636         init_waitqueue_head(&svcpt->scp_rep_waitq);
637         atomic_set(&svcpt->scp_nreps_difficult, 0);
638
639         /* adaptive timeout */
640         spin_lock_init(&svcpt->scp_at_lock);
641         array = &svcpt->scp_at_array;
642
643         size = at_est2timeout(at_max);
644         array->paa_size     = size;
645         array->paa_count    = 0;
646         array->paa_deadline = -1;
647
648         /* allocate memory for scp_at_array (ptlrpc_at_array) */
649         OBD_CPT_ALLOC(array->paa_reqs_array,
650                       svc->srv_cptable, cpt, sizeof(struct list_head) * size);
651         if (array->paa_reqs_array == NULL)
652                 return -ENOMEM;
653
654         for (index = 0; index < size; index++)
655                 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
656
657         OBD_CPT_ALLOC(array->paa_reqs_count,
658                       svc->srv_cptable, cpt, sizeof(__u32) * size);
659         if (array->paa_reqs_count == NULL)
660                 goto failed;
661
662         cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
663                         (unsigned long)svcpt, 0);
664
665         /*
666          * At SOW, service time should be quick; 10s seems generous. If client
667          * timeout is less than this, we'll be sending an early reply.
668          */
669         at_init(&svcpt->scp_at_estimate, 10, 0);
670
671         /* assign this before call ptlrpc_grow_req_bufs */
672         svcpt->scp_service = svc;
673         /* Now allocate the request buffers, but don't post them now */
674         rc = ptlrpc_grow_req_bufs(svcpt, 0);
675         /*
676          * We shouldn't be under memory pressure at startup, so
677          * fail if we can't allocate all our buffers at this time.
678          */
679         if (rc != 0)
680                 goto failed;
681
682         return 0;
683
684  failed:
685         if (array->paa_reqs_count != NULL) {
686                 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
687                 array->paa_reqs_count = NULL;
688         }
689
690         if (array->paa_reqs_array != NULL) {
691                 OBD_FREE(array->paa_reqs_array,
692                          sizeof(struct list_head) * array->paa_size);
693                 array->paa_reqs_array = NULL;
694         }
695
696         return -ENOMEM;
697 }
698
699 /**
700  * Initialize service on a given portal.
701  * This includes starting serving threads , allocating and posting rqbds and
702  * so on.
703  */
704 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
705                                                struct kset *parent,
706                                                struct dentry *debugfs_entry)
707 {
708         struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
709         struct ptlrpc_service *service;
710         struct ptlrpc_service_part *svcpt;
711         struct cfs_cpt_table *cptable;
712         __u32 *cpts = NULL;
713         int ncpts;
714         int cpt;
715         int rc;
716         int i;
717
718         ENTRY;
719
720         LASSERT(conf->psc_buf.bc_nbufs > 0);
721         LASSERT(conf->psc_buf.bc_buf_size >=
722                 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
723         LASSERT(conf->psc_thr.tc_ctx_tags != 0);
724
725         cptable = cconf->cc_cptable;
726         if (cptable == NULL)
727                 cptable = cfs_cpt_table;
728
729         if (conf->psc_thr.tc_cpu_bind > 1) {
730                 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
731                        conf->psc_name, conf->psc_thr.tc_cpu_bind);
732                 RETURN(ERR_PTR(-EINVAL));
733         }
734
735         if (!cconf->cc_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         service->srv_cpt_bind           = conf->psc_thr.tc_cpu_bind;
775
776         service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
777         while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
778                 service->srv_cpt_bits++;
779
780         /* public members */
781         spin_lock_init(&service->srv_lock);
782         service->srv_name               = conf->psc_name;
783         service->srv_watchdog_factor    = conf->psc_watchdog_factor;
784         INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
785
786         /* buffer configuration */
787         service->srv_nbuf_per_group     = test_req_buffer_pressure ?
788                                           1 : conf->psc_buf.bc_nbufs;
789         /* do not limit max number of rqbds by default */
790         service->srv_nrqbds_max         = 0;
791
792         service->srv_max_req_size       = conf->psc_buf.bc_req_max_size +
793                                           SPTLRPC_MAX_PAYLOAD;
794         service->srv_buf_size           = conf->psc_buf.bc_buf_size;
795         service->srv_rep_portal         = conf->psc_buf.bc_rep_portal;
796         service->srv_req_portal         = conf->psc_buf.bc_req_portal;
797
798         /* With slab/alloc_pages buffer size will be rounded up to 2^n */
799         if (service->srv_buf_size & (service->srv_buf_size - 1)) {
800                 int round = size_roundup_power2(service->srv_buf_size);
801
802                 service->srv_buf_size = round;
803         }
804
805         /* Increase max reply size to next power of two */
806         service->srv_max_reply_size = 1;
807         while (service->srv_max_reply_size <
808                conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
809                 service->srv_max_reply_size <<= 1;
810
811         service->srv_thread_name        = conf->psc_thr.tc_thr_name;
812         service->srv_ctx_tags           = conf->psc_thr.tc_ctx_tags;
813         service->srv_hpreq_ratio        = PTLRPC_SVC_HP_RATIO;
814         service->srv_ops                = conf->psc_ops;
815
816         for (i = 0; i < ncpts; i++) {
817                 if (!cconf->cc_affinity)
818                         cpt = CFS_CPT_ANY;
819                 else
820                         cpt = cpts != NULL ? cpts[i] : i;
821
822                 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
823                 if (svcpt == NULL)
824                         GOTO(failed, rc = -ENOMEM);
825
826                 service->srv_parts[i] = svcpt;
827                 rc = ptlrpc_service_part_init(service, svcpt, cpt);
828                 if (rc != 0)
829                         GOTO(failed, rc);
830         }
831
832         ptlrpc_server_nthreads_check(service, conf);
833
834         rc = LNetSetLazyPortal(service->srv_req_portal);
835         LASSERT(rc == 0);
836
837         mutex_lock(&ptlrpc_all_services_mutex);
838         list_add(&service->srv_list, &ptlrpc_all_services);
839         mutex_unlock(&ptlrpc_all_services_mutex);
840
841         if (parent) {
842                 rc = ptlrpc_sysfs_register_service(parent, service);
843                 if (rc)
844                         GOTO(failed, rc);
845         }
846
847         if (debugfs_entry != NULL)
848                 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
849
850         rc = ptlrpc_service_nrs_setup(service);
851         if (rc != 0)
852                 GOTO(failed, rc);
853
854         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
855                service->srv_name, service->srv_req_portal);
856
857         rc = ptlrpc_start_threads(service);
858         if (rc != 0) {
859                 CERROR("Failed to start threads for service %s: %d\n",
860                        service->srv_name, rc);
861                 GOTO(failed, rc);
862         }
863
864         RETURN(service);
865 failed:
866         ptlrpc_unregister_service(service);
867         RETURN(ERR_PTR(rc));
868 }
869 EXPORT_SYMBOL(ptlrpc_register_service);
870
871 /**
872  * to actually free the request, must be called without holding svc_lock.
873  * note it's caller's responsibility to unlink req->rq_list.
874  */
875 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
876 {
877         LASSERT(atomic_read(&req->rq_refcount) == 0);
878         LASSERT(list_empty(&req->rq_timed_list));
879
880         /*
881          * DEBUG_REQ() assumes the reply state of a request with a valid
882          * ref will not be destroyed until that reference is dropped.
883          */
884         ptlrpc_req_drop_rs(req);
885
886         sptlrpc_svc_ctx_decref(req);
887
888         if (req != &req->rq_rqbd->rqbd_req) {
889                 /*
890                  * NB request buffers use an embedded
891                  * req if the incoming req unlinked the
892                  * MD; this isn't one of them!
893                  */
894                 ptlrpc_request_cache_free(req);
895         }
896 }
897
898 /**
899  * drop a reference count of the request. if it reaches 0, we either
900  * put it into history list, or free it immediately.
901  */
902 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
903 {
904         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
905         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
906         struct ptlrpc_service             *svc = svcpt->scp_service;
907         int                                refcount;
908         struct list_head                          *tmp;
909         struct list_head                          *nxt;
910
911         if (!atomic_dec_and_test(&req->rq_refcount))
912                 return;
913
914         if (req->rq_session.lc_state == LCS_ENTERED) {
915                 lu_context_exit(&req->rq_session);
916                 lu_context_fini(&req->rq_session);
917         }
918
919         if (req->rq_at_linked) {
920                 spin_lock(&svcpt->scp_at_lock);
921                 /*
922                  * recheck with lock, in case it's unlinked by
923                  * ptlrpc_at_check_timed()
924                  */
925                 if (likely(req->rq_at_linked))
926                         ptlrpc_at_remove_timed(req);
927                 spin_unlock(&svcpt->scp_at_lock);
928         }
929
930         LASSERT(list_empty(&req->rq_timed_list));
931
932         /* finalize request */
933         if (req->rq_export) {
934                 class_export_put(req->rq_export);
935                 req->rq_export = NULL;
936         }
937
938         spin_lock(&svcpt->scp_lock);
939
940         list_add(&req->rq_list, &rqbd->rqbd_reqs);
941
942         refcount = --(rqbd->rqbd_refcount);
943         if (refcount == 0) {
944                 /* request buffer is now idle: add to history */
945                 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
946                 svcpt->scp_hist_nrqbds++;
947
948                 /*
949                  * cull some history?
950                  * I expect only about 1 or 2 rqbds need to be recycled here
951                  */
952                 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
953                         rqbd = list_entry(svcpt->scp_hist_rqbds.next,
954                                           struct ptlrpc_request_buffer_desc,
955                                           rqbd_list);
956
957                         list_del(&rqbd->rqbd_list);
958                         svcpt->scp_hist_nrqbds--;
959
960                         /*
961                          * remove rqbd's reqs from svc's req history while
962                          * I've got the service lock
963                          */
964                         list_for_each(tmp, &rqbd->rqbd_reqs) {
965                                 req = list_entry(tmp, struct ptlrpc_request,
966                                                  rq_list);
967                                 /* Track the highest culled req seq */
968                                 if (req->rq_history_seq >
969                                     svcpt->scp_hist_seq_culled) {
970                                         svcpt->scp_hist_seq_culled =
971                                                 req->rq_history_seq;
972                                 }
973                                 list_del(&req->rq_history_list);
974                         }
975
976                         spin_unlock(&svcpt->scp_lock);
977
978                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
979                                 req = list_entry(rqbd->rqbd_reqs.next,
980                                                  struct ptlrpc_request,
981                                                  rq_list);
982                                 list_del(&req->rq_list);
983                                 ptlrpc_server_free_request(req);
984                         }
985
986                         spin_lock(&svcpt->scp_lock);
987                         /*
988                          * now all reqs including the embedded req has been
989                          * disposed, schedule request buffer for re-use
990                          * or free it to drain some in excess.
991                          */
992                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
993                         if (svcpt->scp_nrqbds_posted >=
994                             svc->srv_nbuf_per_group ||
995                             (svc->srv_nrqbds_max != 0 &&
996                              svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
997                             test_req_buffer_pressure) {
998                                 /* like in ptlrpc_free_rqbd() */
999                                 svcpt->scp_nrqbds_total--;
1000                                 OBD_FREE_LARGE(rqbd->rqbd_buffer,
1001                                                svc->srv_buf_size);
1002                                 OBD_FREE_PTR(rqbd);
1003                         } else {
1004                                 list_add_tail(&rqbd->rqbd_list,
1005                                               &svcpt->scp_rqbd_idle);
1006                         }
1007                 }
1008
1009                 spin_unlock(&svcpt->scp_lock);
1010         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1011                 /* If we are low on memory, we are not interested in history */
1012                 list_del(&req->rq_list);
1013                 list_del_init(&req->rq_history_list);
1014
1015                 /* Track the highest culled req seq */
1016                 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1017                         svcpt->scp_hist_seq_culled = req->rq_history_seq;
1018
1019                 spin_unlock(&svcpt->scp_lock);
1020
1021                 ptlrpc_server_free_request(req);
1022         } else {
1023                 spin_unlock(&svcpt->scp_lock);
1024         }
1025 }
1026
1027 /** Change request export and move hp request from old export to new */
1028 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1029                                   struct obd_export *export)
1030 {
1031         if (req->rq_export != NULL) {
1032                 LASSERT(!list_empty(&req->rq_exp_list));
1033                 /* remove rq_exp_list from last export */
1034                 spin_lock(&req->rq_export->exp_rpc_lock);
1035                 list_del_init(&req->rq_exp_list);
1036                 spin_unlock(&req->rq_export->exp_rpc_lock);
1037                 /*
1038                  * export has one reference already, so it`s safe to
1039                  * add req to export queue here and get another
1040                  * reference for request later
1041                  */
1042                 spin_lock(&export->exp_rpc_lock);
1043                 if (req->rq_ops != NULL) /* hp request */
1044                         list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1045                 else
1046                         list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1047                 spin_unlock(&export->exp_rpc_lock);
1048
1049                 class_export_rpc_dec(req->rq_export);
1050                 class_export_put(req->rq_export);
1051         }
1052
1053         /* request takes one export refcount */
1054         req->rq_export = class_export_get(export);
1055         class_export_rpc_inc(export);
1056
1057         return;
1058 }
1059
1060 /**
1061  * to finish a request: stop sending more early replies, and release
1062  * the request.
1063  */
1064 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1065                                          struct ptlrpc_request *req)
1066 {
1067         ptlrpc_server_hpreq_fini(req);
1068
1069         ptlrpc_server_drop_request(req);
1070 }
1071
1072 /**
1073  * to finish an active request: stop sending more early replies, and release
1074  * the request. should be called after we finished handling the request.
1075  */
1076 static void ptlrpc_server_finish_active_request(
1077                                         struct ptlrpc_service_part *svcpt,
1078                                         struct ptlrpc_request *req)
1079 {
1080         spin_lock(&svcpt->scp_req_lock);
1081         ptlrpc_nrs_req_stop_nolock(req);
1082         svcpt->scp_nreqs_active--;
1083         if (req->rq_hp)
1084                 svcpt->scp_nhreqs_active--;
1085         spin_unlock(&svcpt->scp_req_lock);
1086
1087         ptlrpc_nrs_req_finalize(req);
1088
1089         if (req->rq_export != NULL)
1090                 class_export_rpc_dec(req->rq_export);
1091
1092         ptlrpc_server_finish_request(svcpt, req);
1093 }
1094
1095 /**
1096  * This function makes sure dead exports are evicted in a timely manner.
1097  * This function is only called when some export receives a message (i.e.,
1098  * the network is up.)
1099  */
1100 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1101 {
1102         struct obd_export *oldest_exp;
1103         time64_t oldest_time, new_time;
1104
1105         ENTRY;
1106
1107         LASSERT(exp);
1108
1109         /*
1110          * Compensate for slow machines, etc, by faking our request time
1111          * into the future.  Although this can break the strict time-ordering
1112          * of the list, we can be really lazy here - we don't have to evict
1113          * at the exact right moment.  Eventually, all silent exports
1114          * will make it to the top of the list.
1115          */
1116
1117         /* Do not pay attention on 1sec or smaller renewals. */
1118         new_time = ktime_get_real_seconds() + extra_delay;
1119         if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1120                 RETURN_EXIT;
1121
1122         exp->exp_last_request_time = new_time;
1123
1124         /*
1125          * exports may get disconnected from the chain even though the
1126          * export has references, so we must keep the spin lock while
1127          * manipulating the lists
1128          */
1129         spin_lock(&exp->exp_obd->obd_dev_lock);
1130
1131         if (list_empty(&exp->exp_obd_chain_timed)) {
1132                 /* this one is not timed */
1133                 spin_unlock(&exp->exp_obd->obd_dev_lock);
1134                 RETURN_EXIT;
1135         }
1136
1137         list_move_tail(&exp->exp_obd_chain_timed,
1138                        &exp->exp_obd->obd_exports_timed);
1139
1140         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1141                                 struct obd_export, exp_obd_chain_timed);
1142         oldest_time = oldest_exp->exp_last_request_time;
1143         spin_unlock(&exp->exp_obd->obd_dev_lock);
1144
1145         if (exp->exp_obd->obd_recovering) {
1146                 /* be nice to everyone during recovery */
1147                 EXIT;
1148                 return;
1149         }
1150
1151         /* Note - racing to start/reset the obd_eviction timer is safe */
1152         if (exp->exp_obd->obd_eviction_timer == 0) {
1153                 /* Check if the oldest entry is expired. */
1154                 if (ktime_get_real_seconds() >
1155                     oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1156                         /*
1157                          * We need a second timer, in case the net was down and
1158                          * it just came back. Since the pinger may skip every
1159                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1160                          * we better wait for 3.
1161                          */
1162                         exp->exp_obd->obd_eviction_timer =
1163                                 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1164                         CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1165                                exp->exp_obd->obd_name,
1166                                obd_export_nid2str(oldest_exp), oldest_time);
1167                 }
1168         } else {
1169                 if (ktime_get_real_seconds() >
1170                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1171                         /*
1172                          * The evictor won't evict anyone who we've heard from
1173                          * recently, so we don't have to check before we start
1174                          * it.
1175                          */
1176                         if (!ping_evictor_wake(exp))
1177                                 exp->exp_obd->obd_eviction_timer = 0;
1178                 }
1179         }
1180
1181         EXIT;
1182 }
1183
1184 /**
1185  * Sanity check request \a req.
1186  * Return 0 if all is ok, error code otherwise.
1187  */
1188 static int ptlrpc_check_req(struct ptlrpc_request *req)
1189 {
1190         struct obd_device *obd = req->rq_export->exp_obd;
1191         int rc = 0;
1192
1193         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1194                      req->rq_export->exp_conn_cnt)) {
1195                 DEBUG_REQ(D_RPCTRACE, req,
1196                           "DROPPING req from old connection %d < %d",
1197                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
1198                           req->rq_export->exp_conn_cnt);
1199                 return -EEXIST;
1200         }
1201         if (unlikely(obd == NULL || obd->obd_fail)) {
1202                 /*
1203                  * Failing over, don't handle any more reqs,
1204                  * send error response instead.
1205                  */
1206                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1207                         req, (obd != NULL) ? obd->obd_name : "unknown");
1208                 rc = -ENODEV;
1209         } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1210                    (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1211                    !obd->obd_recovering) {
1212                 DEBUG_REQ(D_ERROR, req,
1213                           "Invalid replay without recovery");
1214                 class_fail_export(req->rq_export);
1215                 rc = -ENODEV;
1216         } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1217                    !obd->obd_recovering) {
1218                 DEBUG_REQ(D_ERROR, req,
1219                           "Invalid req with transno %llu without recovery",
1220                           lustre_msg_get_transno(req->rq_reqmsg));
1221                 class_fail_export(req->rq_export);
1222                 rc = -ENODEV;
1223         }
1224
1225         if (unlikely(rc < 0)) {
1226                 req->rq_status = rc;
1227                 ptlrpc_error(req);
1228         }
1229         return rc;
1230 }
1231
1232 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1233 {
1234         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1235         time64_t next;
1236
1237         if (array->paa_count == 0) {
1238                 del_timer(&svcpt->scp_at_timer);
1239                 return;
1240         }
1241
1242         /* Set timer for closest deadline */
1243         next = array->paa_deadline - ktime_get_real_seconds() -
1244                at_early_margin;
1245         if (next <= 0) {
1246                 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1247         } else {
1248                 mod_timer(&svcpt->scp_at_timer,
1249                           jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1250                 CDEBUG(D_INFO, "armed %s at %+llds\n",
1251                        svcpt->scp_service->srv_name, next);
1252         }
1253 }
1254
1255 /* Add rpc to early reply check list */
1256 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1257 {
1258         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1259         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1260         struct ptlrpc_request *rq = NULL;
1261         __u32 index;
1262
1263         if (AT_OFF)
1264                 return(0);
1265
1266         if (req->rq_no_reply)
1267                 return 0;
1268
1269         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1270                 return(-ENOSYS);
1271
1272         spin_lock(&svcpt->scp_at_lock);
1273         LASSERT(list_empty(&req->rq_timed_list));
1274
1275         div_u64_rem(req->rq_deadline, array->paa_size, &index);
1276         if (array->paa_reqs_count[index] > 0) {
1277                 /*
1278                  * latest rpcs will have the latest deadlines in the list,
1279                  * so search backward.
1280                  */
1281                 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1282                                             rq_timed_list) {
1283                         if (req->rq_deadline >= rq->rq_deadline) {
1284                                 list_add(&req->rq_timed_list,
1285                                          &rq->rq_timed_list);
1286                                 break;
1287                         }
1288                 }
1289         }
1290
1291         /* Add the request at the head of the list */
1292         if (list_empty(&req->rq_timed_list))
1293                 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1294
1295         spin_lock(&req->rq_lock);
1296         req->rq_at_linked = 1;
1297         spin_unlock(&req->rq_lock);
1298         req->rq_at_index = index;
1299         array->paa_reqs_count[index]++;
1300         array->paa_count++;
1301         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1302                 array->paa_deadline = req->rq_deadline;
1303                 ptlrpc_at_set_timer(svcpt);
1304         }
1305         spin_unlock(&svcpt->scp_at_lock);
1306
1307         return 0;
1308 }
1309
1310 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1311 {
1312         struct ptlrpc_at_array *array;
1313
1314         array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1315
1316         /* NB: must call with hold svcpt::scp_at_lock */
1317         LASSERT(!list_empty(&req->rq_timed_list));
1318         list_del_init(&req->rq_timed_list);
1319
1320         spin_lock(&req->rq_lock);
1321         req->rq_at_linked = 0;
1322         spin_unlock(&req->rq_lock);
1323
1324         array->paa_reqs_count[req->rq_at_index]--;
1325         array->paa_count--;
1326 }
1327
1328 /*
1329  * Attempt to extend the request deadline by sending an early reply to the
1330  * client.
1331  */
1332 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1333 {
1334         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1335         struct ptlrpc_request *reqcopy;
1336         struct lustre_msg *reqmsg;
1337         time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1338         time64_t newdl;
1339         int rc;
1340
1341         ENTRY;
1342
1343         if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1344                 /* don't send early reply */
1345                 RETURN(1);
1346         }
1347
1348         /*
1349          * deadline is when the client expects us to reply, margin is the
1350          * difference between clients' and servers' expectations
1351          */
1352         DEBUG_REQ(D_ADAPTTO, req,
1353                   "%ssending early reply (deadline %+llds, margin %+llds) for %d+%d",
1354                   AT_OFF ? "AT off - not " : "",
1355                   (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1356                   at_get(&svcpt->scp_at_estimate), at_extra);
1357
1358         if (AT_OFF)
1359                 RETURN(0);
1360
1361         if (olddl < 0) {
1362                 /* below message is checked in replay-ost-single.sh test_9 */
1363                 DEBUG_REQ(D_WARNING, req,
1364                           "Already past deadline (%+llds), not sending early reply. Consider increasing at_early_margin (%d)?",
1365                           (s64)olddl, at_early_margin);
1366
1367                 /* Return an error so we're not re-added to the timed list. */
1368                 RETURN(-ETIMEDOUT);
1369         }
1370
1371         if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1372              MSGHDR_AT_SUPPORT) == 0) {
1373                 DEBUG_REQ(D_INFO, req,
1374                           "Wanted to ask client for more time, but no AT support");
1375                 RETURN(-ENOSYS);
1376         }
1377
1378         if (req->rq_export &&
1379             lustre_msg_get_flags(req->rq_reqmsg) &
1380             (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1381                 struct obd_device *obd_exp = req->rq_export->exp_obd;
1382
1383                 /*
1384                  * During recovery, we don't want to send too many early
1385                  * replies, but on the other hand we want to make sure the
1386                  * client has enough time to resend if the rpc is lost. So
1387                  * during the recovery period send at least 4 early replies,
1388                  * spacing them every at_extra if we can. at_estimate should
1389                  * always equal this fixed value during recovery.
1390                  */
1391
1392                 /*
1393                  * Don't account request processing time into AT history
1394                  * during recovery, it is not service time we need but
1395                  * includes also waiting time for recovering clients
1396                  */
1397                 newdl = min_t(time64_t, at_extra,
1398                               obd_exp->obd_recovery_timeout / 4) +
1399                         ktime_get_real_seconds();
1400         } else {
1401                 /*
1402                  * We want to extend the request deadline by at_extra seconds,
1403                  * so we set our service estimate to reflect how much time has
1404                  * passed since this request arrived plus an additional
1405                  * at_extra seconds. The client will calculate the new deadline
1406                  * based on this service estimate (plus some additional time to
1407                  * account for network latency). See ptlrpc_at_recv_early_reply
1408                  */
1409                 at_measured(&svcpt->scp_at_estimate, at_extra +
1410                             ktime_get_real_seconds() -
1411                             req->rq_arrival_time.tv_sec);
1412                 newdl = req->rq_arrival_time.tv_sec +
1413                         at_get(&svcpt->scp_at_estimate);
1414         }
1415
1416         /*
1417          * Check to see if we've actually increased the deadline -
1418          * we may be past adaptive_max
1419          */
1420         if (req->rq_deadline >= newdl) {
1421                 DEBUG_REQ(D_WARNING, req,
1422                           "Could not add any time (%lld/%lld), not sending early reply",
1423                           (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1424                 RETURN(-ETIMEDOUT);
1425         }
1426
1427         reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1428         if (reqcopy == NULL)
1429                 RETURN(-ENOMEM);
1430         OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1431         if (!reqmsg)
1432                 GOTO(out_free, rc = -ENOMEM);
1433
1434         *reqcopy = *req;
1435         reqcopy->rq_reply_state = NULL;
1436         reqcopy->rq_rep_swab_mask = 0;
1437         reqcopy->rq_pack_bulk = 0;
1438         reqcopy->rq_pack_udesc = 0;
1439         reqcopy->rq_packed_final = 0;
1440         sptlrpc_svc_ctx_addref(reqcopy);
1441         /* We only need the reqmsg for the magic */
1442         reqcopy->rq_reqmsg = reqmsg;
1443         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1444
1445         /*
1446          * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1447          * Without this check, we would fail the rq_no_reply assertion in
1448          * ptlrpc_send_reply().
1449          */
1450         if (reqcopy->rq_no_reply)
1451                 GOTO(out, rc = -ETIMEDOUT);
1452
1453         LASSERT(atomic_read(&req->rq_refcount));
1454         /* if it is last refcount then early reply isn't needed */
1455         if (atomic_read(&req->rq_refcount) == 1) {
1456                 DEBUG_REQ(D_ADAPTTO, reqcopy,
1457                           "Normal reply already sent, abort early reply");
1458                 GOTO(out, rc = -EINVAL);
1459         }
1460
1461         /* Connection ref */
1462         reqcopy->rq_export = class_conn2export(
1463                         lustre_msg_get_handle(reqcopy->rq_reqmsg));
1464         if (reqcopy->rq_export == NULL)
1465                 GOTO(out, rc = -ENODEV);
1466
1467         /* RPC ref */
1468         class_export_rpc_inc(reqcopy->rq_export);
1469         if (reqcopy->rq_export->exp_obd &&
1470             reqcopy->rq_export->exp_obd->obd_fail)
1471                 GOTO(out_put, rc = -ENODEV);
1472
1473         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1474         if (rc)
1475                 GOTO(out_put, rc);
1476
1477         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1478
1479         if (!rc) {
1480                 /* Adjust our own deadline to what we told the client */
1481                 req->rq_deadline = newdl;
1482                 req->rq_early_count++; /* number sent, server side */
1483         } else {
1484                 DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
1485         }
1486
1487         /*
1488          * Free the (early) reply state from lustre_pack_reply.
1489          * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1490          */
1491         ptlrpc_req_drop_rs(reqcopy);
1492
1493 out_put:
1494         class_export_rpc_dec(reqcopy->rq_export);
1495         class_export_put(reqcopy->rq_export);
1496 out:
1497         sptlrpc_svc_ctx_decref(reqcopy);
1498         OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1499 out_free:
1500         ptlrpc_request_cache_free(reqcopy);
1501         RETURN(rc);
1502 }
1503
1504 /*
1505  * Send early replies to everybody expiring within at_early_margin
1506  * asking for at_extra time
1507  */
1508 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1509 {
1510         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1511         struct ptlrpc_request *rq, *n;
1512         struct list_head work_list;
1513         __u32 index, count;
1514         time64_t deadline;
1515         time64_t now = ktime_get_real_seconds();
1516         s64 delay;
1517         int first, counter = 0;
1518
1519         ENTRY;
1520         spin_lock(&svcpt->scp_at_lock);
1521         if (svcpt->scp_at_check == 0) {
1522                 spin_unlock(&svcpt->scp_at_lock);
1523                 RETURN(0);
1524         }
1525         delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1526         svcpt->scp_at_check = 0;
1527
1528         if (array->paa_count == 0) {
1529                 spin_unlock(&svcpt->scp_at_lock);
1530                 RETURN(0);
1531         }
1532
1533         /* The timer went off, but maybe the nearest rpc already completed. */
1534         first = array->paa_deadline - now;
1535         if (first > at_early_margin) {
1536                 /* We've still got plenty of time.  Reset the timer. */
1537                 ptlrpc_at_set_timer(svcpt);
1538                 spin_unlock(&svcpt->scp_at_lock);
1539                 RETURN(0);
1540         }
1541
1542         /*
1543          * We're close to a timeout, and we don't know how much longer the
1544          * server will take. Send early replies to everyone expiring soon.
1545          */
1546         INIT_LIST_HEAD(&work_list);
1547         deadline = -1;
1548         div_u64_rem(array->paa_deadline, array->paa_size, &index);
1549         count = array->paa_count;
1550         while (count > 0) {
1551                 count -= array->paa_reqs_count[index];
1552                 list_for_each_entry_safe(rq, n,
1553                                          &array->paa_reqs_array[index],
1554                                          rq_timed_list) {
1555                         if (rq->rq_deadline > now + at_early_margin) {
1556                                 /* update the earliest deadline */
1557                                 if (deadline == -1 ||
1558                                     rq->rq_deadline < deadline)
1559                                         deadline = rq->rq_deadline;
1560                                 break;
1561                         }
1562
1563                         /**
1564                          * ptlrpc_server_drop_request() may drop
1565                          * refcount to 0 already. Let's check this and
1566                          * don't add entry to work_list
1567                          */
1568                         if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1569                                 ptlrpc_at_remove_timed(rq);
1570                                 list_add(&rq->rq_timed_list, &work_list);
1571                         } else {
1572                                 ptlrpc_at_remove_timed(rq);
1573                         }
1574
1575                         counter++;
1576                 }
1577
1578                 if (++index >= array->paa_size)
1579                         index = 0;
1580         }
1581         array->paa_deadline = deadline;
1582         /* we have a new earliest deadline, restart the timer */
1583         ptlrpc_at_set_timer(svcpt);
1584
1585         spin_unlock(&svcpt->scp_at_lock);
1586
1587         CDEBUG(D_ADAPTTO,
1588                "timeout in %+ds, asking for %d secs on %d early replies\n",
1589                first, at_extra, counter);
1590         if (first < 0) {
1591                 /*
1592                  * We're already past request deadlines before we even get a
1593                  * chance to send early replies
1594                  */
1595                 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1596                               svcpt->scp_service->srv_name);
1597                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1598                       counter, svcpt->scp_nreqs_incoming,
1599                       svcpt->scp_nreqs_active,
1600                       at_get(&svcpt->scp_at_estimate), delay);
1601         }
1602
1603         /*
1604          * we took additional refcount so entries can't be deleted from list, no
1605          * locking is needed
1606          */
1607         while (!list_empty(&work_list)) {
1608                 rq = list_entry(work_list.next, struct ptlrpc_request,
1609                                 rq_timed_list);
1610                 list_del_init(&rq->rq_timed_list);
1611
1612                 if (ptlrpc_at_send_early_reply(rq) == 0)
1613                         ptlrpc_at_add_timed(rq);
1614
1615                 ptlrpc_server_drop_request(rq);
1616         }
1617
1618         RETURN(1); /* return "did_something" for liblustre */
1619 }
1620
1621 /*
1622  * Check if we are already handling earlier incarnation of this request.
1623  * Called under &req->rq_export->exp_rpc_lock locked
1624  */
1625 static struct ptlrpc_request*
1626 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1627 {
1628         struct ptlrpc_request *tmp = NULL;
1629
1630         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1631             (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1632                 return NULL;
1633
1634         /*
1635          * bulk request are aborted upon reconnect, don't try to
1636          * find a match
1637          */
1638         if (req->rq_bulk_write || req->rq_bulk_read)
1639                 return NULL;
1640
1641         /*
1642          * This list should not be longer than max_requests in
1643          * flights on the client, so it is not all that long.
1644          * Also we only hit this codepath in case of a resent
1645          * request which makes it even more rarely hit
1646          */
1647         list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1648                                 rq_exp_list) {
1649                 /* Found duplicate one */
1650                 if (tmp->rq_xid == req->rq_xid)
1651                         goto found;
1652         }
1653         list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1654                                 rq_exp_list) {
1655                 /* Found duplicate one */
1656                 if (tmp->rq_xid == req->rq_xid)
1657                         goto found;
1658         }
1659         return NULL;
1660
1661 found:
1662         DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1663         DEBUG_REQ(D_HA, tmp, "Request being processed");
1664         return tmp;
1665 }
1666
1667 /**
1668  * Check if a request should be assigned with a high priority.
1669  *
1670  * \retval      < 0: error occurred
1671  *                0: normal RPC request
1672  *               +1: high priority request
1673  */
1674 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1675                                     struct ptlrpc_request *req)
1676 {
1677         int rc = 0;
1678
1679         ENTRY;
1680         if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1681                 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1682                 if (rc < 0)
1683                         RETURN(rc);
1684
1685                 LASSERT(rc == 0);
1686         }
1687
1688         if (req->rq_export != NULL && req->rq_ops != NULL) {
1689                 /*
1690                  * Perform request specific check. We should do this
1691                  * check before the request is added into exp_hp_rpcs
1692                  * list otherwise it may hit swab race at LU-1044.
1693                  */
1694                 if (req->rq_ops->hpreq_check != NULL) {
1695                         rc = req->rq_ops->hpreq_check(req);
1696                         if (rc == -ESTALE) {
1697                                 req->rq_status = rc;
1698                                 ptlrpc_error(req);
1699                         }
1700                         /*
1701                          * can only return error,
1702                          * 0 for normal request,
1703                          * or 1 for high priority request
1704                          */
1705                         LASSERT(rc <= 1);
1706                 }
1707         }
1708
1709         RETURN(rc);
1710 }
1711
1712 /** Remove the request from the export list. */
1713 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1714 {
1715         ENTRY;
1716         if (req->rq_export) {
1717                 /*
1718                  * refresh lock timeout again so that client has more
1719                  * room to send lock cancel RPC.
1720                  */
1721                 if (req->rq_ops && req->rq_ops->hpreq_fini)
1722                         req->rq_ops->hpreq_fini(req);
1723
1724                 spin_lock(&req->rq_export->exp_rpc_lock);
1725                 list_del_init(&req->rq_exp_list);
1726                 spin_unlock(&req->rq_export->exp_rpc_lock);
1727         }
1728         EXIT;
1729 }
1730
1731 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1732 {
1733         return 1;
1734 }
1735
1736 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1737         .hpreq_check       = ptlrpc_hpreq_check,
1738 };
1739
1740 /* Hi-Priority RPC check by RPC operation code. */
1741 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1742 {
1743         int opc = lustre_msg_get_opc(req->rq_reqmsg);
1744
1745         /*
1746          * Check for export to let only reconnects for not yet evicted
1747          * export to become a HP rpc.
1748          */
1749         if ((req->rq_export != NULL) &&
1750             (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1751                 req->rq_ops = &ptlrpc_hpreq_common;
1752
1753         return 0;
1754 }
1755 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1756
1757 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1758                                      struct ptlrpc_request *req)
1759 {
1760         int rc;
1761         bool hp;
1762         struct ptlrpc_request *orig;
1763
1764         ENTRY;
1765
1766         rc = ptlrpc_server_hpreq_init(svcpt, req);
1767         if (rc < 0)
1768                 RETURN(rc);
1769
1770         hp = rc > 0;
1771         ptlrpc_nrs_req_initialize(svcpt, req, hp);
1772
1773         if (req->rq_export != NULL) {
1774                 struct obd_export *exp = req->rq_export;
1775
1776                 /*
1777                  * do search for duplicated xid and the adding to the list
1778                  * atomically
1779                  */
1780                 spin_lock_bh(&exp->exp_rpc_lock);
1781                 orig = ptlrpc_server_check_resend_in_progress(req);
1782                 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1783                         bool linked;
1784
1785                         spin_unlock_bh(&exp->exp_rpc_lock);
1786
1787                         /*
1788                          * When the client resend request and the server has
1789                          * the previous copy of it, we need to update deadlines,
1790                          * to be sure that the client and the server have equal
1791                          *  request deadlines.
1792                          */
1793
1794                         spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1795                         linked = orig->rq_at_linked;
1796                         if (likely(linked))
1797                                 ptlrpc_at_remove_timed(orig);
1798                         spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1799                         orig->rq_deadline = req->rq_deadline;
1800                         if (likely(linked))
1801                                 ptlrpc_at_add_timed(orig);
1802                         ptlrpc_server_drop_request(orig);
1803                         ptlrpc_nrs_req_finalize(req);
1804                         RETURN(-EBUSY);
1805                 }
1806
1807                 if (hp || req->rq_ops != NULL)
1808                         list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1809                 else
1810                         list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1811                 spin_unlock_bh(&exp->exp_rpc_lock);
1812         }
1813
1814         /*
1815          * the current thread is not the processing thread for this request
1816          * since that, but request is in exp_hp_list and can be find there.
1817          * Remove all relations between request and old thread.
1818          */
1819         req->rq_svc_thread->t_env->le_ses = NULL;
1820         req->rq_svc_thread = NULL;
1821         req->rq_session.lc_thread = NULL;
1822
1823         ptlrpc_nrs_req_add(svcpt, req, hp);
1824
1825         RETURN(0);
1826 }
1827
1828 /**
1829  * Allow to handle high priority request
1830  * User can call it w/o any lock but need to hold
1831  * ptlrpc_service_part::scp_req_lock to get reliable result
1832  */
1833 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1834                                      bool force)
1835 {
1836         int running = svcpt->scp_nthrs_running;
1837
1838         if (!nrs_svcpt_has_hp(svcpt))
1839                 return false;
1840
1841         if (force)
1842                 return true;
1843
1844         if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1845                 return false;
1846
1847         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1848                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1849                 /* leave just 1 thread for normal RPCs */
1850                 running = PTLRPC_NTHRS_INIT;
1851                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1852                         running += 1;
1853         }
1854
1855         if (svcpt->scp_nreqs_active >= running - 1)
1856                 return false;
1857
1858         if (svcpt->scp_nhreqs_active == 0)
1859                 return true;
1860
1861         return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1862                svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1863 }
1864
1865 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1866                                        bool force)
1867 {
1868         return ptlrpc_server_allow_high(svcpt, force) &&
1869                ptlrpc_nrs_req_pending_nolock(svcpt, true);
1870 }
1871
1872 /**
1873  * Only allow normal priority requests on a service that has a high-priority
1874  * queue if forced (i.e. cleanup), if there are other high priority requests
1875  * already being processed (i.e. those threads can service more high-priority
1876  * requests), or if there are enough idle threads that a later thread can do
1877  * a high priority request.
1878  * User can call it w/o any lock but need to hold
1879  * ptlrpc_service_part::scp_req_lock to get reliable result
1880  */
1881 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1882                                        bool force)
1883 {
1884         int running = svcpt->scp_nthrs_running;
1885
1886         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1887                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1888                 /* leave just 1 thread for normal RPCs */
1889                 running = PTLRPC_NTHRS_INIT;
1890                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1891                         running += 1;
1892         }
1893
1894         if (force)
1895                 return true;
1896
1897         if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1898                 return false;
1899
1900         if (svcpt->scp_nreqs_active < running - 2)
1901                 return true;
1902
1903         if (svcpt->scp_nreqs_active >= running - 1)
1904                 return false;
1905
1906         return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1907 }
1908
1909 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1910                                          bool force)
1911 {
1912         return ptlrpc_server_allow_normal(svcpt, force) &&
1913                ptlrpc_nrs_req_pending_nolock(svcpt, false);
1914 }
1915
1916 /**
1917  * Returns true if there are requests available in incoming
1918  * request queue for processing and it is allowed to fetch them.
1919  * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1920  * to get reliable result
1921  * \see ptlrpc_server_allow_normal
1922  * \see ptlrpc_server_allow high
1923  */
1924 static inline
1925 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1926                                    bool force)
1927 {
1928         return ptlrpc_server_high_pending(svcpt, force) ||
1929                ptlrpc_server_normal_pending(svcpt, force);
1930 }
1931
1932 /**
1933  * Fetch a request for processing from queue of unprocessed requests.
1934  * Favors high-priority requests.
1935  * Returns a pointer to fetched request.
1936  */
1937 static struct ptlrpc_request *
1938 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1939 {
1940         struct ptlrpc_request *req = NULL;
1941
1942         ENTRY;
1943
1944         spin_lock(&svcpt->scp_req_lock);
1945
1946         if (ptlrpc_server_high_pending(svcpt, force)) {
1947                 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1948                 if (req != NULL) {
1949                         svcpt->scp_hreq_count++;
1950                         goto got_request;
1951                 }
1952         }
1953
1954         if (ptlrpc_server_normal_pending(svcpt, force)) {
1955                 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1956                 if (req != NULL) {
1957                         svcpt->scp_hreq_count = 0;
1958                         goto got_request;
1959                 }
1960         }
1961
1962         spin_unlock(&svcpt->scp_req_lock);
1963         RETURN(NULL);
1964
1965 got_request:
1966         svcpt->scp_nreqs_active++;
1967         if (req->rq_hp)
1968                 svcpt->scp_nhreqs_active++;
1969
1970         spin_unlock(&svcpt->scp_req_lock);
1971
1972         if (likely(req->rq_export))
1973                 class_export_rpc_inc(req->rq_export);
1974
1975         RETURN(req);
1976 }
1977
1978 /**
1979  * Handle freshly incoming reqs, add to timed early reply list,
1980  * pass on to regular request queue.
1981  * All incoming requests pass through here before getting into
1982  * ptlrpc_server_handle_req later on.
1983  */
1984 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1985                                        struct ptlrpc_thread *thread)
1986 {
1987         struct ptlrpc_service *svc = svcpt->scp_service;
1988         struct ptlrpc_request *req;
1989         __u32 deadline;
1990         int rc;
1991
1992         ENTRY;
1993
1994         spin_lock(&svcpt->scp_lock);
1995         if (list_empty(&svcpt->scp_req_incoming)) {
1996                 spin_unlock(&svcpt->scp_lock);
1997                 RETURN(0);
1998         }
1999
2000         req = list_entry(svcpt->scp_req_incoming.next,
2001                              struct ptlrpc_request, rq_list);
2002         list_del_init(&req->rq_list);
2003         svcpt->scp_nreqs_incoming--;
2004         /*
2005          * Consider this still a "queued" request as far as stats are
2006          * concerned
2007          */
2008         spin_unlock(&svcpt->scp_lock);
2009
2010         /* go through security check/transform */
2011         rc = sptlrpc_svc_unwrap_request(req);
2012         switch (rc) {
2013         case SECSVC_OK:
2014                 break;
2015         case SECSVC_COMPLETE:
2016                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2017                 goto err_req;
2018         case SECSVC_DROP:
2019                 goto err_req;
2020         default:
2021                 LBUG();
2022         }
2023
2024         /*
2025          * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2026          * redo it wouldn't be harmful.
2027          */
2028         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2029                 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2030                 if (rc != 0) {
2031                         CERROR("error unpacking request: ptl %d from %s x%llu\n",
2032                                svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2033                                req->rq_xid);
2034                         goto err_req;
2035                 }
2036         }
2037
2038         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2039         if (rc) {
2040                 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2041                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2042                        req->rq_xid);
2043                 goto err_req;
2044         }
2045
2046         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2047             lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
2048                 CERROR("drop incoming rpc opc %u, x%llu\n",
2049                        cfs_fail_val, req->rq_xid);
2050                 goto err_req;
2051         }
2052
2053         rc = -EINVAL;
2054         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2055                 CERROR("wrong packet type received (type=%u) from %s\n",
2056                        lustre_msg_get_type(req->rq_reqmsg),
2057                        libcfs_id2str(req->rq_peer));
2058                 goto err_req;
2059         }
2060
2061         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2062         case MDS_WRITEPAGE:
2063         case OST_WRITE:
2064         case OUT_UPDATE:
2065                 req->rq_bulk_write = 1;
2066                 break;
2067         case MDS_READPAGE:
2068         case OST_READ:
2069         case MGS_CONFIG_READ:
2070                 req->rq_bulk_read = 1;
2071                 break;
2072         }
2073
2074         CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2075
2076         req->rq_export = class_conn2export(
2077                 lustre_msg_get_handle(req->rq_reqmsg));
2078         if (req->rq_export) {
2079                 rc = ptlrpc_check_req(req);
2080                 if (rc == 0) {
2081                         rc = sptlrpc_target_export_check(req->rq_export, req);
2082                         if (rc)
2083                                 DEBUG_REQ(D_ERROR, req,
2084                                           "DROPPING req with illegal security flavor");
2085                 }
2086
2087                 if (rc)
2088                         goto err_req;
2089                 ptlrpc_update_export_timer(req->rq_export, 0);
2090         }
2091
2092         /* req_in handling should/must be fast */
2093         if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2094                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2095                           (s64)(ktime_get_real_seconds() -
2096                                 req->rq_arrival_time.tv_sec));
2097
2098         /* Set rpc server deadline and add it to the timed list */
2099         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2100                     MSGHDR_AT_SUPPORT) ?
2101                     /* The max time the client expects us to take */
2102                     lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2103
2104         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2105         if (unlikely(deadline == 0)) {
2106                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2107                 goto err_req;
2108         }
2109
2110         /* Skip early reply */
2111         if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2112                 req->rq_deadline += obd_timeout;
2113
2114         req->rq_svc_thread = thread;
2115         if (thread != NULL) {
2116                 /*
2117                  * initialize request session, it is needed for request
2118                  * processing by target
2119                  */
2120                 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2121                                                        LCT_NOREF);
2122                 if (rc) {
2123                         CERROR("%s: failure to initialize session: rc = %d\n",
2124                                thread->t_name, rc);
2125                         goto err_req;
2126                 }
2127                 req->rq_session.lc_thread = thread;
2128                 lu_context_enter(&req->rq_session);
2129                 thread->t_env->le_ses = &req->rq_session;
2130         }
2131
2132         ptlrpc_at_add_timed(req);
2133
2134         /* Move it over to the request processing queue */
2135         rc = ptlrpc_server_request_add(svcpt, req);
2136         if (rc)
2137                 GOTO(err_req, rc);
2138
2139         wake_up(&svcpt->scp_waitq);
2140         RETURN(1);
2141
2142 err_req:
2143         ptlrpc_server_finish_request(svcpt, req);
2144
2145         RETURN(1);
2146 }
2147
2148 /**
2149  * Main incoming request handling logic.
2150  * Calls handler function from service to do actual processing.
2151  */
2152 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2153                                         struct ptlrpc_thread *thread)
2154 {
2155         struct ptlrpc_service *svc = svcpt->scp_service;
2156         struct ptlrpc_request *request;
2157         ktime_t work_start;
2158         ktime_t work_end;
2159         ktime_t arrived;
2160         s64 timediff_usecs;
2161         s64 arrived_usecs;
2162         int fail_opc = 0;
2163
2164         ENTRY;
2165
2166         request = ptlrpc_server_request_get(svcpt, false);
2167         if (request == NULL)
2168                 RETURN(0);
2169
2170         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2171                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2172         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2173                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2174
2175         if (unlikely(fail_opc)) {
2176                 if (request->rq_export && request->rq_ops)
2177                         OBD_FAIL_TIMEOUT(fail_opc, 4);
2178         }
2179
2180         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2181
2182         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2183                 libcfs_debug_dumplog();
2184
2185         work_start = ktime_get_real();
2186         arrived = timespec64_to_ktime(request->rq_arrival_time);
2187         timediff_usecs = ktime_us_delta(work_start, arrived);
2188         if (likely(svc->srv_stats != NULL)) {
2189                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2190                                     timediff_usecs);
2191                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2192                                     svcpt->scp_nreqs_incoming);
2193                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2194                                     svcpt->scp_nreqs_active);
2195                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2196                                     at_get(&svcpt->scp_at_estimate));
2197         }
2198
2199         if (likely(request->rq_export)) {
2200                 if (unlikely(ptlrpc_check_req(request)))
2201                         goto put_conn;
2202                 ptlrpc_update_export_timer(request->rq_export,
2203                                            div_u64(timediff_usecs,
2204                                                    USEC_PER_SEC / 2));
2205         }
2206
2207         /*
2208          * Discard requests queued for longer than the deadline.
2209          * The deadline is increased if we send an early reply.
2210          */
2211         if (ktime_get_real_seconds() > request->rq_deadline) {
2212                 DEBUG_REQ(D_ERROR, request,
2213                           "Dropping timed-out request from %s: deadline %lld/%llds ago",
2214                           libcfs_id2str(request->rq_peer),
2215                           request->rq_deadline -
2216                           request->rq_arrival_time.tv_sec,
2217                           ktime_get_real_seconds() - request->rq_deadline);
2218                 goto put_conn;
2219         }
2220
2221         CDEBUG(D_RPCTRACE,
2222                "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2223                request, current_comm(),
2224                (request->rq_export ?
2225                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2226                (request->rq_export ?
2227                 atomic_read(&request->rq_export->exp_refcount) : -99),
2228                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2229                libcfs_id2str(request->rq_peer),
2230                lustre_msg_get_opc(request->rq_reqmsg),
2231                lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2232
2233         if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2234                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2235
2236         CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2237
2238         /* re-assign request and sesson thread to the current one */
2239         request->rq_svc_thread = thread;
2240         if (thread != NULL) {
2241                 LASSERT(request->rq_session.lc_thread == NULL);
2242                 request->rq_session.lc_thread = thread;
2243                 thread->t_env->le_ses = &request->rq_session;
2244         }
2245         svc->srv_ops.so_req_handler(request);
2246
2247         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2248
2249 put_conn:
2250         if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2251                 DEBUG_REQ(D_WARNING, request,
2252                           "Request took longer than estimated (%lld/%llds); client may timeout",
2253                           request->rq_deadline -
2254                           request->rq_arrival_time.tv_sec,
2255                           ktime_get_real_seconds() - request->rq_deadline);
2256         }
2257
2258         work_end = ktime_get_real();
2259         timediff_usecs = ktime_us_delta(work_end, work_start);
2260         arrived_usecs = ktime_us_delta(work_end, arrived);
2261         CDEBUG(D_RPCTRACE,
2262                "Handled RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s Request processed in %lldus (%lldus total) trans %llu rc %d/%d\n",
2263                request, current_comm(),
2264                (request->rq_export ?
2265                (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2266                (request->rq_export ?
2267                atomic_read(&request->rq_export->exp_refcount) : -99),
2268                lustre_msg_get_status(request->rq_reqmsg),
2269                request->rq_xid,
2270                libcfs_id2str(request->rq_peer),
2271                lustre_msg_get_opc(request->rq_reqmsg),
2272                lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2273                timediff_usecs,
2274                arrived_usecs,
2275                (request->rq_repmsg ?
2276                lustre_msg_get_transno(request->rq_repmsg) :
2277                request->rq_transno),
2278                request->rq_status,
2279                (request->rq_repmsg ?
2280                lustre_msg_get_status(request->rq_repmsg) : -999));
2281         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2282                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2283                 int opc = opcode_offset(op);
2284
2285                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2286                         LASSERT(opc < LUSTRE_MAX_OPCODES);
2287                         lprocfs_counter_add(svc->srv_stats,
2288                                             opc + EXTRA_MAX_OPCODES,
2289                                             timediff_usecs);
2290                 }
2291         }
2292         if (unlikely(request->rq_early_count)) {
2293                 DEBUG_REQ(D_ADAPTTO, request,
2294                           "sent %d early replies before finishing in %llds",
2295                           request->rq_early_count,
2296                           div_u64(arrived_usecs, USEC_PER_SEC));
2297         }
2298
2299         ptlrpc_server_finish_active_request(svcpt, request);
2300
2301         RETURN(1);
2302 }
2303
2304 /**
2305  * An internal function to process a single reply state object.
2306  */
2307 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2308 {
2309         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2310         struct ptlrpc_service *svc = svcpt->scp_service;
2311         struct obd_export *exp;
2312         int nlocks;
2313         int been_handled;
2314
2315         ENTRY;
2316
2317         exp = rs->rs_export;
2318
2319         LASSERT(rs->rs_difficult);
2320         LASSERT(rs->rs_scheduled);
2321         LASSERT(list_empty(&rs->rs_list));
2322
2323         /*
2324          * The disk commit callback holds exp_uncommitted_replies_lock while it
2325          * iterates over newly committed replies, removing them from
2326          * exp_uncommitted_replies.  It then drops this lock and schedules the
2327          * replies it found for handling here.
2328          *
2329          * We can avoid contention for exp_uncommitted_replies_lock between the
2330          * HRT threads and further commit callbacks by checking rs_committed
2331          * which is set in the commit callback while it holds both
2332          * rs_lock and exp_uncommitted_reples.
2333          *
2334          * If we see rs_committed clear, the commit callback _may_ not have
2335          * handled this reply yet and we race with it to grab
2336          * exp_uncommitted_replies_lock before removing the reply from
2337          * exp_uncommitted_replies.  Note that if we lose the race and the
2338          * reply has already been removed, list_del_init() is a noop.
2339          *
2340          * If we see rs_committed set, we know the commit callback is handling,
2341          * or has handled this reply since store reordering might allow us to
2342          * see rs_committed set out of sequence.  But since this is done
2343          * holding rs_lock, we can be sure it has all completed once we hold
2344          * rs_lock, which we do right next.
2345          */
2346         if (!rs->rs_committed) {
2347                 /*
2348                  * if rs was commited, no need to convert locks, don't check
2349                  * rs_committed here because rs may never be added into
2350                  * exp_uncommitted_replies and this flag never be set, see
2351                  * target_send_reply()
2352                  */
2353                 if (rs->rs_convert_lock &&
2354                     rs->rs_transno > exp->exp_last_committed) {
2355                         struct ldlm_lock *lock;
2356                         struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2357
2358                         spin_lock(&rs->rs_lock);
2359                         if (rs->rs_convert_lock &&
2360                             rs->rs_transno > exp->exp_last_committed) {
2361                                 nlocks = rs->rs_nlocks;
2362                                 while (nlocks-- > 0) {
2363                                         /*
2364                                          * NB don't assume rs is always handled
2365                                          * by the same service thread (see
2366                                          * ptlrpc_hr_select, so REP-ACK hr may
2367                                          * race with trans commit, while the
2368                                          * latter will release locks, get locks
2369                                          * here early to convert to COS mode
2370                                          * safely.
2371                                          */
2372                                         lock = ldlm_handle2lock(
2373                                                         &rs->rs_locks[nlocks]);
2374                                         LASSERT(lock);
2375                                         ack_locks[nlocks] = lock;
2376                                         rs->rs_modes[nlocks] = LCK_COS;
2377                                 }
2378                                 nlocks = rs->rs_nlocks;
2379                                 rs->rs_convert_lock = 0;
2380                                 /*
2381                                  * clear rs_scheduled so that commit callback
2382                                  * can schedule again
2383                                  */
2384                                 rs->rs_scheduled = 0;
2385                                 spin_unlock(&rs->rs_lock);
2386
2387                                 while (nlocks-- > 0) {
2388                                         lock = ack_locks[nlocks];
2389                                         ldlm_lock_mode_downgrade(lock, LCK_COS);
2390                                         LDLM_LOCK_PUT(lock);
2391                                 }
2392                                 RETURN(0);
2393                         }
2394                         spin_unlock(&rs->rs_lock);
2395                 }
2396
2397                 spin_lock(&exp->exp_uncommitted_replies_lock);
2398                 list_del_init(&rs->rs_obd_list);
2399                 spin_unlock(&exp->exp_uncommitted_replies_lock);
2400         }
2401
2402         spin_lock(&exp->exp_lock);
2403         /* Noop if removed already */
2404         list_del_init(&rs->rs_exp_list);
2405         spin_unlock(&exp->exp_lock);
2406
2407         spin_lock(&rs->rs_lock);
2408
2409         been_handled = rs->rs_handled;
2410         rs->rs_handled = 1;
2411
2412         nlocks = rs->rs_nlocks; /* atomic "steal", but */
2413         rs->rs_nlocks = 0; /* locks still on rs_locks! */
2414
2415         if (nlocks == 0 && !been_handled) {
2416                 /*
2417                  * If we see this, we should already have seen the warning
2418                  * in mds_steal_ack_locks()
2419                  */
2420                 CDEBUG(D_HA,
2421                        "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2422                        rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2423                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
2424         }
2425
2426         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2427                 spin_unlock(&rs->rs_lock);
2428
2429                 if (!been_handled && rs->rs_on_net) {
2430                         LNetMDUnlink(rs->rs_md_h);
2431                         /* Ignore return code; we're racing with completion */
2432                 }
2433
2434                 while (nlocks-- > 0)
2435                         ldlm_lock_decref(&rs->rs_locks[nlocks],
2436                                          rs->rs_modes[nlocks]);
2437
2438                 spin_lock(&rs->rs_lock);
2439         }
2440
2441         rs->rs_scheduled = 0;
2442         rs->rs_convert_lock = 0;
2443
2444         if (!rs->rs_on_net) {
2445                 /* Off the net */
2446                 spin_unlock(&rs->rs_lock);
2447
2448                 class_export_put(exp);
2449                 rs->rs_export = NULL;
2450                 ptlrpc_rs_decref(rs);
2451                 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2452                     svc->srv_is_stopping)
2453                         wake_up_all(&svcpt->scp_waitq);
2454                 RETURN(1);
2455         }
2456
2457         /* still on the net; callback will schedule */
2458         spin_unlock(&rs->rs_lock);
2459         RETURN(1);
2460 }
2461
2462
2463 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2464 {
2465         int avail = svcpt->scp_nrqbds_posted;
2466         int low_water = test_req_buffer_pressure ? 0 :
2467                         svcpt->scp_service->srv_nbuf_per_group / 2;
2468
2469         /* NB I'm not locking; just looking. */
2470
2471         /*
2472          * CAVEAT EMPTOR: We might be allocating buffers here because we've
2473          * allowed the request history to grow out of control.  We could put a
2474          * sanity check on that here and cull some history if we need the
2475          * space.
2476          */
2477
2478         if (avail <= low_water)
2479                 ptlrpc_grow_req_bufs(svcpt, 1);
2480
2481         if (svcpt->scp_service->srv_stats) {
2482                 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2483                                     PTLRPC_REQBUF_AVAIL_CNTR, avail);
2484         }
2485 }
2486
2487 static int ptlrpc_retry_rqbds(void *arg)
2488 {
2489         struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2490
2491         svcpt->scp_rqbd_timeout = 0;
2492         return -ETIMEDOUT;
2493 }
2494
2495 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2496 {
2497         return svcpt->scp_nreqs_active <
2498                svcpt->scp_nthrs_running - 1 -
2499                (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2500 }
2501
2502 /**
2503  * allowed to create more threads
2504  * user can call it w/o any lock but need to hold
2505  * ptlrpc_service_part::scp_lock to get reliable result
2506  */
2507 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2508 {
2509         return svcpt->scp_nthrs_running +
2510                svcpt->scp_nthrs_starting <
2511                svcpt->scp_service->srv_nthrs_cpt_limit;
2512 }
2513
2514 /**
2515  * too many requests and allowed to create more threads
2516  */
2517 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2518 {
2519         return !ptlrpc_threads_enough(svcpt) &&
2520                 ptlrpc_threads_increasable(svcpt);
2521 }
2522
2523 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2524 {
2525         return thread_is_stopping(thread) ||
2526                thread->t_svcpt->scp_service->srv_is_stopping;
2527 }
2528
2529 /* stop the highest numbered thread if there are too many threads running */
2530 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2531 {
2532         struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2533
2534         return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2535                 thread->t_id == svcpt->scp_thr_nextid - 1;
2536 }
2537
2538 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2539 {
2540         CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2541                thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2542         thread_add_flags(thread, SVC_STOPPING);
2543 }
2544
2545 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2546 {
2547         struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2548
2549         spin_lock(&svcpt->scp_lock);
2550         if (ptlrpc_thread_should_stop(thread)) {
2551                 ptlrpc_stop_thread(thread);
2552                 svcpt->scp_thr_nextid--;
2553         }
2554         spin_unlock(&svcpt->scp_lock);
2555 }
2556
2557 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2558 {
2559         return !list_empty(&svcpt->scp_rqbd_idle) &&
2560                svcpt->scp_rqbd_timeout == 0;
2561 }
2562
2563 static inline int
2564 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2565 {
2566         return svcpt->scp_at_check;
2567 }
2568
2569 /*
2570  * If a thread runs too long or spends to much time on a single request,
2571  * we want to know about it, so we set up a delayed work item as a watchdog.
2572  * If it fires, we display a stack trace of the delayed thread,
2573  * providing we aren't rate-limited
2574  *
2575  * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2576  * seconds
2577  */
2578 static struct ratelimit_state watchdog_limit;
2579
2580 static void ptlrpc_watchdog_fire(struct work_struct *w)
2581 {
2582         struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2583                                                     t_watchdog.work);
2584         u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2585         u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2586
2587         if (!__ratelimit(&watchdog_limit)) {
2588                 /* below message is checked in sanity-quota.sh test_6,18 */
2589                 LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. The thread might be hung, or it might only be slow and will resume later. Dumping the stack trace for debugging purposes:\n",
2590                               thread->t_task->comm, thread->t_task->pid,
2591                               ms_lapse, ms_frac);
2592
2593                 libcfs_debug_dumpstack(thread->t_task);
2594         } else {
2595                 /* below message is checked in sanity-quota.sh test_6,18 */
2596                 LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. Watchdog stack traces are limited to 3 per %u seconds, skipping this one.\n",
2597                               thread->t_task->comm, thread->t_task->pid,
2598                               ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2599         }
2600 }
2601
2602 static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
2603 {
2604         INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2605         schedule_delayed_work(work, cfs_time_seconds(time));
2606 }
2607
2608 static void ptlrpc_watchdog_disable(struct delayed_work *work)
2609 {
2610         cancel_delayed_work_sync(work);
2611 }
2612
2613 static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
2614 {
2615         struct ptlrpc_thread *thread = container_of(&work->work,
2616                                                     struct ptlrpc_thread,
2617                                                     t_watchdog.work);
2618         thread->t_touched = ktime_get();
2619         mod_delayed_work(system_wq, work, cfs_time_seconds(time));
2620 }
2621
2622 /**
2623  * requests wait on preprocessing
2624  * user can call it w/o any lock but need to hold
2625  * ptlrpc_service_part::scp_lock to get reliable result
2626  */
2627 static inline int
2628 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2629 {
2630         return !list_empty(&svcpt->scp_req_incoming);
2631 }
2632
2633 static __attribute__((__noinline__)) int
2634 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2635                   struct ptlrpc_thread *thread)
2636 {
2637         /* Don't exit while there are replies to be handled */
2638         struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2639                                              ptlrpc_retry_rqbds, svcpt);
2640
2641         ptlrpc_watchdog_disable(&thread->t_watchdog);
2642
2643         cond_resched();
2644
2645         l_wait_event_exclusive_head(svcpt->scp_waitq,
2646                                 ptlrpc_thread_stopping(thread) ||
2647                                 ptlrpc_server_request_incoming(svcpt) ||
2648                                 ptlrpc_server_request_pending(svcpt, false) ||
2649                                 ptlrpc_rqbd_pending(svcpt) ||
2650                                 ptlrpc_at_check(svcpt), &lwi);
2651
2652         if (ptlrpc_thread_stopping(thread))
2653                 return -EINTR;
2654
2655         ptlrpc_watchdog_touch(&thread->t_watchdog,
2656                               ptlrpc_server_get_timeout(svcpt));
2657         return 0;
2658 }
2659
2660 /**
2661  * Main thread body for service threads.
2662  * Waits in a loop waiting for new requests to process to appear.
2663  * Every time an incoming requests is added to its queue, a waitq
2664  * is woken up and one of the threads will handle it.
2665  */
2666 static int ptlrpc_main(void *arg)
2667 {
2668         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2669         struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2670         struct ptlrpc_service *svc = svcpt->scp_service;
2671         struct ptlrpc_reply_state *rs;
2672         struct group_info *ginfo = NULL;
2673         struct lu_env *env;
2674         int counter = 0, rc = 0;
2675
2676         ENTRY;
2677
2678         thread->t_task = current;
2679         thread->t_pid = current_pid();
2680         unshare_fs_struct();
2681
2682         if (svc->srv_cpt_bind) {
2683                 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2684                 if (rc != 0) {
2685                         CWARN("%s: failed to bind %s on CPT %d\n",
2686                               svc->srv_name, thread->t_name, svcpt->scp_cpt);
2687                 }
2688         }
2689
2690         ginfo = groups_alloc(0);
2691         if (!ginfo)
2692                 GOTO(out, rc = -ENOMEM);
2693
2694         set_current_groups(ginfo);
2695         put_group_info(ginfo);
2696
2697         if (svc->srv_ops.so_thr_init != NULL) {
2698                 rc = svc->srv_ops.so_thr_init(thread);
2699                 if (rc)
2700                         GOTO(out, rc);
2701         }
2702
2703         OBD_ALLOC_PTR(env);
2704         if (env == NULL)
2705                 GOTO(out_srv_fini, rc = -ENOMEM);
2706         rc = lu_env_add(env);
2707         if (rc)
2708                 GOTO(out_env, rc);
2709
2710         rc = lu_context_init(&env->le_ctx,
2711                              svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2712         if (rc)
2713                 GOTO(out_env_remove, rc);
2714
2715         thread->t_env = env;
2716         env->le_ctx.lc_thread = thread;
2717         env->le_ctx.lc_cookie = 0x6;
2718
2719         while (!list_empty(&svcpt->scp_rqbd_idle)) {
2720                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2721                 if (rc >= 0)
2722                         continue;
2723
2724                 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2725                         svc->srv_name, svcpt->scp_cpt, rc);
2726                 GOTO(out_ctx_fini, rc);
2727         }
2728
2729         /* Alloc reply state structure for this one */
2730         OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2731         if (!rs)
2732                 GOTO(out_ctx_fini, rc = -ENOMEM);
2733
2734         spin_lock(&svcpt->scp_lock);
2735
2736         LASSERT(thread_is_starting(thread));
2737         thread_clear_flags(thread, SVC_STARTING);
2738
2739         LASSERT(svcpt->scp_nthrs_starting == 1);
2740         svcpt->scp_nthrs_starting--;
2741
2742         /*
2743          * SVC_STOPPING may already be set here if someone else is trying
2744          * to stop the service while this new thread has been dynamically
2745          * forked. We still set SVC_RUNNING to let our creator know that
2746          * we are now running, however we will exit as soon as possible
2747          */
2748         thread_add_flags(thread, SVC_RUNNING);
2749         svcpt->scp_nthrs_running++;
2750         spin_unlock(&svcpt->scp_lock);
2751
2752         /* wake up our creator in case he's still waiting. */
2753         wake_up(&thread->t_ctl_waitq);
2754
2755         thread->t_touched = ktime_get();
2756         ptlrpc_watchdog_init(&thread->t_watchdog,
2757                          ptlrpc_server_get_timeout(svcpt));
2758
2759         spin_lock(&svcpt->scp_rep_lock);
2760         list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2761         wake_up(&svcpt->scp_rep_waitq);
2762         spin_unlock(&svcpt->scp_rep_lock);
2763
2764         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2765                svcpt->scp_nthrs_running);
2766
2767         /* XXX maintain a list of all managed devices: insert here */
2768         while (!ptlrpc_thread_stopping(thread)) {
2769                 if (ptlrpc_wait_event(svcpt, thread))
2770                         break;
2771
2772                 ptlrpc_check_rqbd_pool(svcpt);
2773
2774                 if (ptlrpc_threads_need_create(svcpt)) {
2775                         /* Ignore return code - we tried... */
2776                         ptlrpc_start_thread(svcpt, 0);
2777                 }
2778
2779                 /* reset le_ses to initial state */
2780                 env->le_ses = NULL;
2781                 /* Process all incoming reqs before handling any */
2782                 if (ptlrpc_server_request_incoming(svcpt)) {
2783                         lu_context_enter(&env->le_ctx);
2784                         ptlrpc_server_handle_req_in(svcpt, thread);
2785                         lu_context_exit(&env->le_ctx);
2786
2787                         /* but limit ourselves in case of flood */
2788                         if (counter++ < 100)
2789                                 continue;
2790                         counter = 0;
2791                 }
2792
2793                 if (ptlrpc_at_check(svcpt))
2794                         ptlrpc_at_check_timed(svcpt);
2795
2796                 if (ptlrpc_server_request_pending(svcpt, false)) {
2797                         lu_context_enter(&env->le_ctx);
2798                         ptlrpc_server_handle_request(svcpt, thread);
2799                         lu_context_exit(&env->le_ctx);
2800                 }
2801
2802                 if (ptlrpc_rqbd_pending(svcpt) &&
2803                     ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2804                         /*
2805                          * I just failed to repost request buffers.
2806                          * Wait for a timeout (unless something else
2807                          * happens) before I try again
2808                          */
2809                         svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2810                         CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2811                                svcpt->scp_nrqbds_posted);
2812                 }
2813                 /*
2814                  * If the number of threads has been tuned downward and this
2815                  * thread should be stopped, then stop in reverse order so the
2816                  * the threads always have contiguous thread index values.
2817                  */
2818                 if (unlikely(ptlrpc_thread_should_stop(thread)))
2819                         ptlrpc_thread_stop(thread);
2820         }
2821
2822         ptlrpc_watchdog_disable(&thread->t_watchdog);
2823
2824 out_ctx_fini:
2825         lu_context_fini(&env->le_ctx);
2826 out_env_remove:
2827         lu_env_remove(env);
2828 out_env:
2829         OBD_FREE_PTR(env);
2830 out_srv_fini:
2831         /* deconstruct service thread state created by ptlrpc_start_thread() */
2832         if (svc->srv_ops.so_thr_done != NULL)
2833                 svc->srv_ops.so_thr_done(thread);
2834 out:
2835         CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2836                thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2837         spin_lock(&svcpt->scp_lock);
2838         if (thread_test_and_clear_flags(thread, SVC_STARTING))
2839                 svcpt->scp_nthrs_starting--;
2840
2841         if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2842                 /* must know immediately */
2843                 svcpt->scp_nthrs_running--;
2844         }
2845
2846         thread->t_id = rc;
2847         thread_add_flags(thread, SVC_STOPPED);
2848
2849         wake_up(&thread->t_ctl_waitq);
2850         spin_unlock(&svcpt->scp_lock);
2851
2852         return rc;
2853 }
2854
2855 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2856                           struct list_head *replies)
2857 {
2858         int result;
2859
2860         spin_lock(&hrt->hrt_lock);
2861
2862         list_splice_init(&hrt->hrt_queue, replies);
2863         result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2864
2865         spin_unlock(&hrt->hrt_lock);
2866         return result;
2867 }
2868
2869 /**
2870  * Main body of "handle reply" function.
2871  * It processes acked reply states
2872  */
2873 static int ptlrpc_hr_main(void *arg)
2874 {
2875         struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2876         struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2877         struct list_head replies;
2878         struct lu_env *env;
2879         int rc;
2880
2881         OBD_ALLOC_PTR(env);
2882         if (env == NULL)
2883                 RETURN(-ENOMEM);
2884
2885         INIT_LIST_HEAD(&replies);
2886         unshare_fs_struct();
2887
2888         rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2889         if (rc != 0) {
2890                 char threadname[20];
2891
2892                 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2893                          hrp->hrp_cpt, hrt->hrt_id);
2894                 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2895                       threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2896         }
2897
2898         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2899                              LCT_REMEMBER | LCT_NOREF);
2900         if (rc)
2901                 GOTO(out_env, rc);
2902
2903         rc = lu_env_add(env);
2904         if (rc)
2905                 GOTO(out_ctx_fini, rc);
2906
2907         atomic_inc(&hrp->hrp_nstarted);
2908         wake_up(&ptlrpc_hr.hr_waitq);
2909
2910         while (!ptlrpc_hr.hr_stopping) {
2911                 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2912
2913                 while (!list_empty(&replies)) {
2914                         struct ptlrpc_reply_state *rs;
2915
2916                         rs = list_entry(replies.prev,
2917                                         struct ptlrpc_reply_state,
2918                                         rs_list);
2919                         list_del_init(&rs->rs_list);
2920                         /* refill keys if needed */
2921                         lu_env_refill(env);
2922                         lu_context_enter(&env->le_ctx);
2923                         ptlrpc_handle_rs(rs);
2924                         lu_context_exit(&env->le_ctx);
2925                 }
2926         }
2927
2928         atomic_inc(&hrp->hrp_nstopped);
2929         wake_up(&ptlrpc_hr.hr_waitq);
2930
2931         lu_env_remove(env);
2932 out_ctx_fini:
2933         lu_context_fini(&env->le_ctx);
2934 out_env:
2935         OBD_FREE_PTR(env);
2936         return 0;
2937 }
2938
2939 static void ptlrpc_stop_hr_threads(void)
2940 {
2941         struct ptlrpc_hr_partition *hrp;
2942         int i;
2943         int j;
2944
2945         ptlrpc_hr.hr_stopping = 1;
2946
2947         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2948                 if (hrp->hrp_thrs == NULL)
2949                         continue; /* uninitialized */
2950                 for (j = 0; j < hrp->hrp_nthrs; j++)
2951                         wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2952         }
2953
2954         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2955                 if (hrp->hrp_thrs == NULL)
2956                         continue; /* uninitialized */
2957                 wait_event(ptlrpc_hr.hr_waitq,
2958                                atomic_read(&hrp->hrp_nstopped) ==
2959                                atomic_read(&hrp->hrp_nstarted));
2960         }
2961 }
2962
2963 static int ptlrpc_start_hr_threads(void)
2964 {
2965         struct ptlrpc_hr_partition *hrp;
2966         int i;
2967         int j;
2968
2969         ENTRY;
2970
2971         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2972                 int     rc = 0;
2973
2974                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2975                         struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2976                         struct task_struct *task;
2977
2978                         task = kthread_run(ptlrpc_hr_main,
2979                                            &hrp->hrp_thrs[j],
2980                                            "ptlrpc_hr%02d_%03d",
2981                                            hrp->hrp_cpt,
2982                                            hrt->hrt_id);
2983                         if (IS_ERR(task)) {
2984                                 rc = PTR_ERR(task);
2985                                 break;
2986                         }
2987                 }
2988
2989                 wait_event(ptlrpc_hr.hr_waitq,
2990                            atomic_read(&hrp->hrp_nstarted) == j);
2991
2992                 if (rc < 0) {
2993                         CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
2994                                i, j, rc);
2995                         ptlrpc_stop_hr_threads();
2996                         RETURN(rc);
2997                 }
2998         }
2999
3000         RETURN(0);
3001 }
3002
3003 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3004 {
3005         struct l_wait_info lwi = { 0 };
3006         struct ptlrpc_thread *thread;
3007         struct list_head zombie;
3008
3009         ENTRY;
3010
3011         CDEBUG(D_INFO, "Stopping threads for service %s\n",
3012                svcpt->scp_service->srv_name);
3013
3014         INIT_LIST_HEAD(&zombie);
3015         spin_lock(&svcpt->scp_lock);
3016         /* let the thread know that we would like it to stop asap */
3017         list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3018                 ptlrpc_stop_thread(thread);
3019
3020         wake_up_all(&svcpt->scp_waitq);
3021
3022         while (!list_empty(&svcpt->scp_threads)) {
3023                 thread = list_entry(svcpt->scp_threads.next,
3024                                         struct ptlrpc_thread, t_link);
3025                 if (thread_is_stopped(thread)) {
3026                         list_move(&thread->t_link, &zombie);
3027                         continue;
3028                 }
3029                 spin_unlock(&svcpt->scp_lock);
3030
3031                 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3032                        svcpt->scp_service->srv_thread_name, thread->t_id);
3033                 l_wait_event(thread->t_ctl_waitq,
3034                              thread_is_stopped(thread), &lwi);
3035
3036                 spin_lock(&svcpt->scp_lock);
3037         }
3038
3039         spin_unlock(&svcpt->scp_lock);
3040
3041         while (!list_empty(&zombie)) {
3042                 thread = list_entry(zombie.next,
3043                                         struct ptlrpc_thread, t_link);
3044                 list_del(&thread->t_link);
3045                 OBD_FREE_PTR(thread);
3046         }
3047         EXIT;
3048 }
3049
3050 /**
3051  * Stops all threads of a particular service \a svc
3052  */
3053 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3054 {
3055         struct ptlrpc_service_part *svcpt;
3056         int i;
3057
3058         ENTRY;
3059
3060         ptlrpc_service_for_each_part(svcpt, i, svc) {
3061                 if (svcpt->scp_service != NULL)
3062                         ptlrpc_svcpt_stop_threads(svcpt);
3063         }
3064
3065         EXIT;
3066 }
3067
3068 int ptlrpc_start_threads(struct ptlrpc_service *svc)
3069 {
3070         int rc = 0;
3071         int i;
3072         int j;
3073
3074         ENTRY;
3075
3076         /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3077         LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3078
3079         for (i = 0; i < svc->srv_ncpts; i++) {
3080                 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3081                         rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3082                         if (rc == 0)
3083                                 continue;
3084
3085                         if (rc != -EMFILE)
3086                                 goto failed;
3087                         /* We have enough threads, don't start more. b=15759 */
3088                         break;
3089                 }
3090         }
3091
3092         RETURN(0);
3093  failed:
3094         CERROR("cannot start %s thread #%d_%d: rc %d\n",
3095                svc->srv_thread_name, i, j, rc);
3096         ptlrpc_stop_all_threads(svc);
3097         RETURN(rc);
3098 }
3099
3100 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3101 {
3102         struct l_wait_info lwi = { 0 };
3103         struct ptlrpc_thread *thread;
3104         struct ptlrpc_service *svc;
3105         struct task_struct *task;
3106         int rc;
3107
3108         ENTRY;
3109
3110         LASSERT(svcpt != NULL);
3111
3112         svc = svcpt->scp_service;
3113
3114         CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3115                svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3116                svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3117
3118  again:
3119         if (unlikely(svc->srv_is_stopping))
3120                 RETURN(-ESRCH);
3121
3122         if (!ptlrpc_threads_increasable(svcpt) ||
3123             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3124              svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3125                 RETURN(-EMFILE);
3126
3127         OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3128         if (thread == NULL)
3129                 RETURN(-ENOMEM);
3130         init_waitqueue_head(&thread->t_ctl_waitq);
3131
3132         spin_lock(&svcpt->scp_lock);
3133         if (!ptlrpc_threads_increasable(svcpt)) {
3134                 spin_unlock(&svcpt->scp_lock);
3135                 OBD_FREE_PTR(thread);
3136                 RETURN(-EMFILE);
3137         }
3138
3139         if (svcpt->scp_nthrs_starting != 0) {
3140                 /*
3141                  * serialize starting because some modules (obdfilter)
3142                  * might require unique and contiguous t_id
3143                  */
3144                 LASSERT(svcpt->scp_nthrs_starting == 1);
3145                 spin_unlock(&svcpt->scp_lock);
3146                 OBD_FREE_PTR(thread);
3147                 if (wait) {
3148                         CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3149                                svc->srv_thread_name, svcpt->scp_thr_nextid);
3150                         schedule();
3151                         goto again;
3152                 }
3153
3154                 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3155                        svc->srv_thread_name, svcpt->scp_thr_nextid);
3156                 RETURN(-EAGAIN);
3157         }
3158
3159         svcpt->scp_nthrs_starting++;
3160         thread->t_id = svcpt->scp_thr_nextid++;
3161         thread_add_flags(thread, SVC_STARTING);
3162         thread->t_svcpt = svcpt;
3163
3164         list_add(&thread->t_link, &svcpt->scp_threads);
3165         spin_unlock(&svcpt->scp_lock);
3166
3167         if (svcpt->scp_cpt >= 0) {
3168                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3169                          svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3170         } else {
3171                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3172                          svc->srv_thread_name, thread->t_id);
3173         }
3174
3175         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3176         task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3177         if (IS_ERR(task)) {
3178                 rc = PTR_ERR(task);
3179                 CERROR("cannot start thread '%s': rc = %d\n",
3180                        thread->t_name, rc);
3181                 spin_lock(&svcpt->scp_lock);
3182                 --svcpt->scp_nthrs_starting;
3183                 if (thread_is_stopping(thread)) {
3184                         /*
3185                          * this ptlrpc_thread is being hanled
3186                          * by ptlrpc_svcpt_stop_threads now
3187                          */
3188                         thread_add_flags(thread, SVC_STOPPED);
3189                         wake_up(&thread->t_ctl_waitq);
3190                         spin_unlock(&svcpt->scp_lock);
3191                 } else {
3192                         list_del(&thread->t_link);
3193                         spin_unlock(&svcpt->scp_lock);
3194                         OBD_FREE_PTR(thread);
3195                 }
3196                 RETURN(rc);
3197         }
3198
3199         if (!wait)
3200                 RETURN(0);
3201
3202         l_wait_event(thread->t_ctl_waitq,
3203                      thread_is_running(thread) || thread_is_stopped(thread),
3204                      &lwi);
3205
3206         rc = thread_is_stopped(thread) ? thread->t_id : 0;
3207         RETURN(rc);
3208 }
3209
3210 int ptlrpc_hr_init(void)
3211 {
3212         struct ptlrpc_hr_partition *hrp;
3213         struct ptlrpc_hr_thread *hrt;
3214         int rc;
3215         int cpt;
3216         int i;
3217         int weight;
3218
3219         ENTRY;
3220
3221         memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3222         ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3223
3224         ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3225                                                    sizeof(*hrp));
3226         if (ptlrpc_hr.hr_partitions == NULL)
3227                 RETURN(-ENOMEM);
3228
3229         ratelimit_state_init(&watchdog_limit,
3230                              cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3231
3232         init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3233
3234         preempt_disable();
3235         weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3236         preempt_enable();
3237
3238         cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3239                 hrp->hrp_cpt = cpt;
3240
3241                 atomic_set(&hrp->hrp_nstarted, 0);
3242                 atomic_set(&hrp->hrp_nstopped, 0);
3243
3244                 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3245                 hrp->hrp_nthrs /= weight;
3246                 if (hrp->hrp_nthrs == 0)
3247                         hrp->hrp_nthrs = 1;
3248
3249                 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3250                               hrp->hrp_nthrs * sizeof(*hrt));
3251                 if (hrp->hrp_thrs == NULL)
3252                         GOTO(out, rc = -ENOMEM);
3253
3254                 for (i = 0; i < hrp->hrp_nthrs; i++) {
3255                         hrt = &hrp->hrp_thrs[i];
3256
3257                         hrt->hrt_id = i;
3258                         hrt->hrt_partition = hrp;
3259                         init_waitqueue_head(&hrt->hrt_waitq);
3260                         spin_lock_init(&hrt->hrt_lock);
3261                         INIT_LIST_HEAD(&hrt->hrt_queue);
3262                 }
3263         }
3264
3265         rc = ptlrpc_start_hr_threads();
3266 out:
3267         if (rc != 0)
3268                 ptlrpc_hr_fini();
3269         RETURN(rc);
3270 }