Whamcloud - gitweb
e65bcaed38b3078f829c422a0045602990a16842
[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, 2016, 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 #include <linux/kthread.h>
35 #include <obd_support.h>
36 #include <obd_class.h>
37 #include <lustre_net.h>
38 #include <lu_object.h>
39 #include <lnet/types.h>
40 #include "ptlrpc_internal.h"
41
42 /* The following are visible and mutable through /sys/module/ptlrpc */
43 int test_req_buffer_pressure = 0;
44 module_param(test_req_buffer_pressure, int, 0444);
45 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
46 module_param(at_min, int, 0644);
47 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
48 module_param(at_max, int, 0644);
49 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
50 module_param(at_history, int, 0644);
51 MODULE_PARM_DESC(at_history,
52                  "Adaptive timeouts remember the slowest event that took place within this period (sec)");
53 module_param(at_early_margin, int, 0644);
54 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
55 module_param(at_extra, int, 0644);
56 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
57
58 /* forward ref */
59 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
60 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
61 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
62
63 /** Holds a list of all PTLRPC services */
64 struct list_head ptlrpc_all_services;
65 /** Used to protect the \e ptlrpc_all_services list */
66 struct mutex ptlrpc_all_services_mutex;
67
68 static struct ptlrpc_request_buffer_desc *
69 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
70 {
71         struct ptlrpc_service             *svc = svcpt->scp_service;
72         struct ptlrpc_request_buffer_desc *rqbd;
73
74         OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
75         if (rqbd == NULL)
76                 return NULL;
77
78         rqbd->rqbd_svcpt = svcpt;
79         rqbd->rqbd_refcount = 0;
80         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
81         rqbd->rqbd_cbid.cbid_arg = rqbd;
82         INIT_LIST_HEAD(&rqbd->rqbd_reqs);
83         OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
84                             svcpt->scp_cpt, svc->srv_buf_size);
85         if (rqbd->rqbd_buffer == NULL) {
86                 OBD_FREE_PTR(rqbd);
87                 return NULL;
88         }
89
90         spin_lock(&svcpt->scp_lock);
91         list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
92         svcpt->scp_nrqbds_total++;
93         spin_unlock(&svcpt->scp_lock);
94
95         return rqbd;
96 }
97
98 static void
99 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
100 {
101         struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
102
103         LASSERT(rqbd->rqbd_refcount == 0);
104         LASSERT(list_empty(&rqbd->rqbd_reqs));
105
106         spin_lock(&svcpt->scp_lock);
107         list_del(&rqbd->rqbd_list);
108         svcpt->scp_nrqbds_total--;
109         spin_unlock(&svcpt->scp_lock);
110
111         OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
112         OBD_FREE_PTR(rqbd);
113 }
114
115 static int
116 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
117 {
118         struct ptlrpc_service             *svc = svcpt->scp_service;
119         struct ptlrpc_request_buffer_desc *rqbd;
120         int                                rc = 0;
121         int                                i;
122
123         if (svcpt->scp_rqbd_allocating)
124                 goto try_post;
125
126         spin_lock(&svcpt->scp_lock);
127         /* check again with lock */
128         if (svcpt->scp_rqbd_allocating) {
129                 /* NB: we might allow more than one thread in the future */
130                 LASSERT(svcpt->scp_rqbd_allocating == 1);
131                 spin_unlock(&svcpt->scp_lock);
132                 goto try_post;
133         }
134
135         svcpt->scp_rqbd_allocating++;
136         spin_unlock(&svcpt->scp_lock);
137
138
139         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
140                 /* NB: another thread might have recycled enough rqbds, we
141                  * need to make sure it wouldn't over-allocate, see LU-1212. */
142                 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
143                         break;
144
145                 rqbd = ptlrpc_alloc_rqbd(svcpt);
146
147                 if (rqbd == NULL) {
148                         CERROR("%s: Can't allocate request buffer\n",
149                                svc->srv_name);
150                         rc = -ENOMEM;
151                         break;
152                 }
153         }
154
155         spin_lock(&svcpt->scp_lock);
156
157         LASSERT(svcpt->scp_rqbd_allocating == 1);
158         svcpt->scp_rqbd_allocating--;
159
160         spin_unlock(&svcpt->scp_lock);
161
162         CDEBUG(D_RPCTRACE,
163                "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
164                svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
165                svcpt->scp_nrqbds_total, rc);
166
167  try_post:
168         if (post && rc == 0)
169                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
170
171         return rc;
172 }
173
174 /**
175  * Part of Rep-Ack logic.
176  * Puts a lock and its mode into reply state assotiated to request reply.
177  */
178 void
179 ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
180                  int mode, bool no_ack, bool convert_lock)
181 {
182         struct ptlrpc_reply_state *rs = req->rq_reply_state;
183         int idx;
184
185         LASSERT(rs != NULL);
186         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
187
188         if (req->rq_export->exp_disconnected) {
189                 ldlm_lock_decref(lock, mode);
190         } else {
191                 idx = rs->rs_nlocks++;
192                 rs->rs_locks[idx] = *lock;
193                 rs->rs_modes[idx] = mode;
194                 rs->rs_difficult = 1;
195                 rs->rs_no_ack = no_ack;
196                 rs->rs_convert_lock = convert_lock;
197         }
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 struct ptlrpc_hr_thread *
271 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         ENTRY;
370
371         LASSERT(list_empty(&rs->rs_list));
372
373         hrt = ptlrpc_hr_select(rs->rs_svcpt);
374
375         spin_lock(&hrt->hrt_lock);
376         list_add_tail(&rs->rs_list, &hrt->hrt_queue);
377         spin_unlock(&hrt->hrt_lock);
378
379         wake_up(&hrt->hrt_waitq);
380         EXIT;
381 }
382
383 void
384 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         ENTRY;
410
411         rs_batch_init(&batch);
412         /* Find any replies that have been committed and get their service
413          * to attend to complete them. */
414
415         /* CAVEAT EMPTOR: spinlock ordering!!! */
416         spin_lock(&exp->exp_uncommitted_replies_lock);
417         list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
418                                      rs_obd_list) {
419                 LASSERT (rs->rs_difficult);
420                 /* VBR: per-export last_committed */
421                 LASSERT(rs->rs_export);
422                 if (rs->rs_transno <= exp->exp_last_committed) {
423                         list_del_init(&rs->rs_obd_list);
424                         rs_batch_add(&batch, rs);
425                 }
426         }
427         spin_unlock(&exp->exp_uncommitted_replies_lock);
428         rs_batch_fini(&batch);
429         EXIT;
430 }
431
432 static int
433 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
434 {
435         struct ptlrpc_request_buffer_desc *rqbd;
436         int                               rc;
437         int                               posted = 0;
438
439         for (;;) {
440                 spin_lock(&svcpt->scp_lock);
441
442                 if (list_empty(&svcpt->scp_rqbd_idle)) {
443                         spin_unlock(&svcpt->scp_lock);
444                         return posted;
445                 }
446
447                 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
448                                       struct ptlrpc_request_buffer_desc,
449                                       rqbd_list);
450                 list_del(&rqbd->rqbd_list);
451
452                 /* assume we will post successfully */
453                 svcpt->scp_nrqbds_posted++;
454                 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
455
456                 spin_unlock(&svcpt->scp_lock);
457
458                 rc = ptlrpc_register_rqbd(rqbd);
459                 if (rc != 0)
460                         break;
461
462                 posted = 1;
463         }
464
465         spin_lock(&svcpt->scp_lock);
466
467         svcpt->scp_nrqbds_posted--;
468         list_del(&rqbd->rqbd_list);
469         list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
470
471         /* Don't complain if no request buffers are posted right now; LNET
472          * won't drop requests because we set the portal lazy! */
473
474         spin_unlock(&svcpt->scp_lock);
475
476         return -1;
477 }
478
479 static void ptlrpc_at_timer(unsigned long castmeharder)
480 {
481         struct ptlrpc_service_part *svcpt;
482
483         svcpt = (struct ptlrpc_service_part *)castmeharder;
484
485         svcpt->scp_at_check = 1;
486         svcpt->scp_at_checktime = cfs_time_current();
487         wake_up(&svcpt->scp_waitq);
488 }
489
490 static void
491 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
492                              struct ptlrpc_service_conf *conf)
493 {
494         struct ptlrpc_service_thr_conf  *tc = &conf->psc_thr;
495         unsigned                        init;
496         unsigned                        total;
497         unsigned                        nthrs;
498         int                             weight;
499
500         /*
501          * Common code for estimating & validating threads number.
502          * CPT affinity service could have percpt thread-pool instead
503          * of a global thread-pool, which means user might not always
504          * get the threads number they give it in conf::tc_nthrs_user
505          * even they did set. It's because we need to validate threads
506          * number for each CPT to guarantee each pool will have enough
507          * threads to keep the service healthy.
508          */
509         init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
510         init = max_t(int, init, tc->tc_nthrs_init);
511
512         /* NB: please see comments in lustre_lnet.h for definition
513          * details of these members */
514         LASSERT(tc->tc_nthrs_max != 0);
515
516         if (tc->tc_nthrs_user != 0) {
517                 /* In case there is a reason to test a service with many
518                  * threads, we give a less strict check here, it can
519                  * be up to 8 * nthrs_max */
520                 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
521                 nthrs = total / svc->srv_ncpts;
522                 init  = max(init, nthrs);
523                 goto out;
524         }
525
526         total = tc->tc_nthrs_max;
527         if (tc->tc_nthrs_base == 0) {
528                 /* don't care about base threads number per partition,
529                  * this is most for non-affinity service */
530                 nthrs = total / svc->srv_ncpts;
531                 goto out;
532         }
533
534         nthrs = tc->tc_nthrs_base;
535         if (svc->srv_ncpts == 1) {
536                 int     i;
537
538                 /* NB: Increase the base number if it's single partition
539                  * and total number of cores/HTs is larger or equal to 4.
540                  * result will always < 2 * nthrs_base */
541                 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
542                 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
543                             (tc->tc_nthrs_base >> i) != 0; i++)
544                         nthrs += tc->tc_nthrs_base >> i;
545         }
546
547         if (tc->tc_thr_factor != 0) {
548                 int       factor = tc->tc_thr_factor;
549                 const int fade = 4;
550
551                 /*
552                  * User wants to increase number of threads with for
553                  * each CPU core/HT, most likely the factor is larger than
554                  * one thread/core because service threads are supposed to
555                  * be blocked by lock or wait for IO.
556                  */
557                 /*
558                  * Amdahl's law says that adding processors wouldn't give
559                  * a linear increasing of parallelism, so it's nonsense to
560                  * have too many threads no matter how many cores/HTs
561                  * there are.
562                  */
563                 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
564                         /* weight is # of HTs */
565                         /* depress thread factor for hyper-thread */
566                         factor = factor - (factor >> 1) + (factor >> 3);
567                 }
568
569                 weight = cfs_cpt_weight(svc->srv_cptable, 0);
570                 LASSERT(weight > 0);
571
572                 for (; factor > 0 && weight > 0; factor--, weight -= fade)
573                         nthrs += min(weight, fade) * factor;
574         }
575
576         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
577                 nthrs = max(tc->tc_nthrs_base,
578                             tc->tc_nthrs_max / svc->srv_ncpts);
579         }
580  out:
581         nthrs = max(nthrs, tc->tc_nthrs_init);
582         svc->srv_nthrs_cpt_limit = nthrs;
583         svc->srv_nthrs_cpt_init = init;
584
585         if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
586                 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
587                        "than the given soft limit (%d)\n",
588                        svc->srv_name, nthrs * svc->srv_ncpts,
589                        tc->tc_nthrs_max);
590         }
591 }
592
593 /**
594  * Initialize percpt data for a service
595  */
596 static int
597 ptlrpc_service_part_init(struct ptlrpc_service *svc,
598                          struct ptlrpc_service_part *svcpt, int cpt)
599 {
600         struct ptlrpc_at_array  *array;
601         int                     size;
602         int                     index;
603         int                     rc;
604
605         svcpt->scp_cpt = cpt;
606         INIT_LIST_HEAD(&svcpt->scp_threads);
607
608         /* rqbd and incoming request queue */
609         spin_lock_init(&svcpt->scp_lock);
610         INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
611         INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
612         INIT_LIST_HEAD(&svcpt->scp_req_incoming);
613         init_waitqueue_head(&svcpt->scp_waitq);
614         /* history request & rqbd list */
615         INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
616         INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
617
618         /* acitve requests and hp requests */
619         spin_lock_init(&svcpt->scp_req_lock);
620
621         /* reply states */
622         spin_lock_init(&svcpt->scp_rep_lock);
623         INIT_LIST_HEAD(&svcpt->scp_rep_active);
624         INIT_LIST_HEAD(&svcpt->scp_rep_idle);
625         init_waitqueue_head(&svcpt->scp_rep_waitq);
626         atomic_set(&svcpt->scp_nreps_difficult, 0);
627
628         /* adaptive timeout */
629         spin_lock_init(&svcpt->scp_at_lock);
630         array = &svcpt->scp_at_array;
631
632         size = at_est2timeout(at_max);
633         array->paa_size     = size;
634         array->paa_count    = 0;
635         array->paa_deadline = -1;
636
637         /* allocate memory for scp_at_array (ptlrpc_at_array) */
638         OBD_CPT_ALLOC(array->paa_reqs_array,
639                       svc->srv_cptable, cpt, sizeof(struct list_head) * size);
640         if (array->paa_reqs_array == NULL)
641                 return -ENOMEM;
642
643         for (index = 0; index < size; index++)
644                 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
645
646         OBD_CPT_ALLOC(array->paa_reqs_count,
647                       svc->srv_cptable, cpt, sizeof(__u32) * size);
648         if (array->paa_reqs_count == NULL)
649                 goto failed;
650
651         setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
652                     (unsigned long)svcpt);
653
654         /* At SOW, service time should be quick; 10s seems generous. If client
655          * timeout is less than this, we'll be sending an early reply. */
656         at_init(&svcpt->scp_at_estimate, 10, 0);
657
658         /* assign this before call ptlrpc_grow_req_bufs */
659         svcpt->scp_service = svc;
660         /* Now allocate the request buffers, but don't post them now */
661         rc = ptlrpc_grow_req_bufs(svcpt, 0);
662         /* We shouldn't be under memory pressure at startup, so
663          * fail if we can't allocate all our buffers at this time. */
664         if (rc != 0)
665                 goto failed;
666
667         return 0;
668
669  failed:
670         if (array->paa_reqs_count != NULL) {
671                 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
672                 array->paa_reqs_count = NULL;
673         }
674
675         if (array->paa_reqs_array != NULL) {
676                 OBD_FREE(array->paa_reqs_array,
677                          sizeof(struct list_head) * array->paa_size);
678                 array->paa_reqs_array = NULL;
679         }
680
681         return -ENOMEM;
682 }
683
684 /**
685  * Initialize service on a given portal.
686  * This includes starting serving threads , allocating and posting rqbds and
687  * so on.
688  */
689 struct ptlrpc_service *
690 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
691                         struct proc_dir_entry *proc_entry)
692 {
693         struct ptlrpc_service_cpt_conf  *cconf = &conf->psc_cpt;
694         struct ptlrpc_service           *service;
695         struct ptlrpc_service_part      *svcpt;
696         struct cfs_cpt_table            *cptable;
697         __u32                           *cpts = NULL;
698         int                             ncpts;
699         int                             cpt;
700         int                             rc;
701         int                             i;
702         ENTRY;
703
704         LASSERT(conf->psc_buf.bc_nbufs > 0);
705         LASSERT(conf->psc_buf.bc_buf_size >=
706                 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
707         LASSERT(conf->psc_thr.tc_ctx_tags != 0);
708
709         cptable = cconf->cc_cptable;
710         if (cptable == NULL)
711                 cptable = cfs_cpt_table;
712
713         if (!conf->psc_thr.tc_cpu_affinity) {
714                 ncpts = 1;
715         } else {
716                 ncpts = cfs_cpt_number(cptable);
717                 if (cconf->cc_pattern != NULL) {
718                         struct cfs_expr_list    *el;
719
720                         rc = cfs_expr_list_parse(cconf->cc_pattern,
721                                                  strlen(cconf->cc_pattern),
722                                                  0, ncpts - 1, &el);
723                         if (rc != 0) {
724                                 CERROR("%s: invalid CPT pattern string: %s",
725                                        conf->psc_name, cconf->cc_pattern);
726                                 RETURN(ERR_PTR(-EINVAL));
727                         }
728
729                         rc = cfs_expr_list_values(el, ncpts, &cpts);
730                         cfs_expr_list_free(el);
731                         if (rc <= 0) {
732                                 CERROR("%s: failed to parse CPT array %s: %d\n",
733                                        conf->psc_name, cconf->cc_pattern, rc);
734                                 if (cpts != NULL)
735                                         OBD_FREE(cpts, sizeof(*cpts) * ncpts);
736                                 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
737                         }
738                         ncpts = rc;
739                 }
740         }
741
742         OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
743         if (service == NULL) {
744                 if (cpts != NULL)
745                         OBD_FREE(cpts, sizeof(*cpts) * ncpts);
746                 RETURN(ERR_PTR(-ENOMEM));
747         }
748
749         service->srv_cptable            = cptable;
750         service->srv_cpts               = cpts;
751         service->srv_ncpts              = ncpts;
752
753         service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
754         while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
755                 service->srv_cpt_bits++;
756
757         /* public members */
758         spin_lock_init(&service->srv_lock);
759         service->srv_name               = conf->psc_name;
760         service->srv_watchdog_factor    = conf->psc_watchdog_factor;
761         INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
762
763         /* buffer configuration */
764         service->srv_nbuf_per_group     = test_req_buffer_pressure ?
765                                           1 : conf->psc_buf.bc_nbufs;
766         service->srv_max_req_size       = conf->psc_buf.bc_req_max_size +
767                                           SPTLRPC_MAX_PAYLOAD;
768         service->srv_buf_size           = conf->psc_buf.bc_buf_size;
769         service->srv_rep_portal         = conf->psc_buf.bc_rep_portal;
770         service->srv_req_portal         = conf->psc_buf.bc_req_portal;
771
772         /* Increase max reply size to next power of two */
773         service->srv_max_reply_size = 1;
774         while (service->srv_max_reply_size <
775                conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
776                 service->srv_max_reply_size <<= 1;
777
778         service->srv_thread_name        = conf->psc_thr.tc_thr_name;
779         service->srv_ctx_tags           = conf->psc_thr.tc_ctx_tags;
780         service->srv_hpreq_ratio        = PTLRPC_SVC_HP_RATIO;
781         service->srv_ops                = conf->psc_ops;
782
783         for (i = 0; i < ncpts; i++) {
784                 if (!conf->psc_thr.tc_cpu_affinity)
785                         cpt = CFS_CPT_ANY;
786                 else
787                         cpt = cpts != NULL ? cpts[i] : i;
788
789                 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
790                 if (svcpt == NULL)
791                         GOTO(failed, rc = -ENOMEM);
792
793                 service->srv_parts[i] = svcpt;
794                 rc = ptlrpc_service_part_init(service, svcpt, cpt);
795                 if (rc != 0)
796                         GOTO(failed, rc);
797         }
798
799         ptlrpc_server_nthreads_check(service, conf);
800
801         rc = LNetSetLazyPortal(service->srv_req_portal);
802         LASSERT(rc == 0);
803
804         mutex_lock(&ptlrpc_all_services_mutex);
805         list_add(&service->srv_list, &ptlrpc_all_services);
806         mutex_unlock(&ptlrpc_all_services_mutex);
807
808         if (proc_entry != NULL)
809                 ptlrpc_lprocfs_register_service(proc_entry, service);
810
811         rc = ptlrpc_service_nrs_setup(service);
812         if (rc != 0)
813                 GOTO(failed, rc);
814
815         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
816                service->srv_name, service->srv_req_portal);
817
818         rc = ptlrpc_start_threads(service);
819         if (rc != 0) {
820                 CERROR("Failed to start threads for service %s: %d\n",
821                        service->srv_name, rc);
822                 GOTO(failed, rc);
823         }
824
825         RETURN(service);
826 failed:
827         ptlrpc_unregister_service(service);
828         RETURN(ERR_PTR(rc));
829 }
830 EXPORT_SYMBOL(ptlrpc_register_service);
831
832 /**
833  * to actually free the request, must be called without holding svc_lock.
834  * note it's caller's responsibility to unlink req->rq_list.
835  */
836 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
837 {
838         LASSERT(atomic_read(&req->rq_refcount) == 0);
839         LASSERT(list_empty(&req->rq_timed_list));
840
841         /* DEBUG_REQ() assumes the reply state of a request with a valid
842          * ref will not be destroyed until that reference is dropped. */
843         ptlrpc_req_drop_rs(req);
844
845         sptlrpc_svc_ctx_decref(req);
846
847         if (req != &req->rq_rqbd->rqbd_req) {
848                 /* NB request buffers use an embedded
849                  * req if the incoming req unlinked the
850                  * MD; this isn't one of them! */
851                 ptlrpc_request_cache_free(req);
852         }
853 }
854
855 /**
856  * drop a reference count of the request. if it reaches 0, we either
857  * put it into history list, or free it immediately.
858  */
859 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
860 {
861         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
862         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
863         struct ptlrpc_service             *svc = svcpt->scp_service;
864         int                                refcount;
865         struct list_head                          *tmp;
866         struct list_head                          *nxt;
867
868         if (!atomic_dec_and_test(&req->rq_refcount))
869                 return;
870
871         if (req->rq_session.lc_state == LCS_ENTERED) {
872                 lu_context_exit(&req->rq_session);
873                 lu_context_fini(&req->rq_session);
874         }
875
876         if (req->rq_at_linked) {
877                 spin_lock(&svcpt->scp_at_lock);
878                 /* recheck with lock, in case it's unlinked by
879                  * ptlrpc_at_check_timed() */
880                 if (likely(req->rq_at_linked))
881                         ptlrpc_at_remove_timed(req);
882                 spin_unlock(&svcpt->scp_at_lock);
883         }
884
885         LASSERT(list_empty(&req->rq_timed_list));
886
887         /* finalize request */
888         if (req->rq_export) {
889                 class_export_put(req->rq_export);
890                 req->rq_export = NULL;
891         }
892
893         spin_lock(&svcpt->scp_lock);
894
895         list_add(&req->rq_list, &rqbd->rqbd_reqs);
896
897         refcount = --(rqbd->rqbd_refcount);
898         if (refcount == 0) {
899                 /* request buffer is now idle: add to history */
900                 list_del(&rqbd->rqbd_list);
901
902                 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
903                 svcpt->scp_hist_nrqbds++;
904
905                 /* cull some history?
906                  * I expect only about 1 or 2 rqbds need to be recycled here */
907                 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
908                         rqbd = list_entry(svcpt->scp_hist_rqbds.next,
909                                           struct ptlrpc_request_buffer_desc,
910                                           rqbd_list);
911
912                         list_del(&rqbd->rqbd_list);
913                         svcpt->scp_hist_nrqbds--;
914
915                         /* remove rqbd's reqs from svc's req history while
916                          * I've got the service lock */
917                         list_for_each(tmp, &rqbd->rqbd_reqs) {
918                                 req = list_entry(tmp, struct ptlrpc_request,
919                                                  rq_list);
920                                 /* Track the highest culled req seq */
921                                 if (req->rq_history_seq >
922                                     svcpt->scp_hist_seq_culled) {
923                                         svcpt->scp_hist_seq_culled =
924                                                 req->rq_history_seq;
925                                 }
926                                 list_del(&req->rq_history_list);
927                         }
928
929                         spin_unlock(&svcpt->scp_lock);
930
931                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
932                                 req = list_entry(rqbd->rqbd_reqs.next,
933                                                  struct ptlrpc_request,
934                                                  rq_list);
935                                 list_del(&req->rq_list);
936                                 ptlrpc_server_free_request(req);
937                         }
938
939                         spin_lock(&svcpt->scp_lock);
940                         /*
941                          * now all reqs including the embedded req has been
942                          * disposed, schedule request buffer for re-use
943                          * or free it to drain some in excess.
944                          */
945                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
946                         if (svcpt->scp_nrqbds_posted >=
947                             svc->srv_nbuf_per_group &&
948                             !test_req_buffer_pressure) {
949                                 /* like in ptlrpc_free_rqbd() */
950                                 svcpt->scp_nrqbds_total--;
951                                 OBD_FREE_LARGE(rqbd->rqbd_buffer,
952                                                svc->srv_buf_size);
953                                 OBD_FREE_PTR(rqbd);
954                         } else {
955                                 list_add_tail(&rqbd->rqbd_list,
956                                               &svcpt->scp_rqbd_idle);
957                         }
958                 }
959
960                 spin_unlock(&svcpt->scp_lock);
961         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
962                 /* If we are low on memory, we are not interested in history */
963                 list_del(&req->rq_list);
964                 list_del_init(&req->rq_history_list);
965
966                 /* Track the highest culled req seq */
967                 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
968                         svcpt->scp_hist_seq_culled = req->rq_history_seq;
969
970                 spin_unlock(&svcpt->scp_lock);
971
972                 ptlrpc_server_free_request(req);
973         } else {
974                 spin_unlock(&svcpt->scp_lock);
975         }
976 }
977
978 /** Change request export and move hp request from old export to new */
979 void ptlrpc_request_change_export(struct ptlrpc_request *req,
980                                   struct obd_export *export)
981 {
982         if (req->rq_export != NULL) {
983                 LASSERT(!list_empty(&req->rq_exp_list));
984                 /* remove rq_exp_list from last export */
985                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
986                 list_del_init(&req->rq_exp_list);
987                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
988                 /* export has one reference already, so it`s safe to
989                  * add req to export queue here and get another
990                  * reference for request later */
991                 spin_lock_bh(&export->exp_rpc_lock);
992                 if (req->rq_ops != NULL) /* hp request */
993                         list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
994                 else
995                         list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
996                 spin_unlock_bh(&export->exp_rpc_lock);
997
998                 class_export_rpc_dec(req->rq_export);
999                 class_export_put(req->rq_export);
1000         }
1001
1002         /* request takes one export refcount */
1003         req->rq_export = class_export_get(export);
1004         class_export_rpc_inc(export);
1005
1006         return;
1007 }
1008
1009 /**
1010  * to finish a request: stop sending more early replies, and release
1011  * the request.
1012  */
1013 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1014                                          struct ptlrpc_request *req)
1015 {
1016         ptlrpc_server_hpreq_fini(req);
1017
1018         ptlrpc_server_drop_request(req);
1019 }
1020
1021 /**
1022  * to finish an active request: stop sending more early replies, and release
1023  * the request. should be called after we finished handling the request.
1024  */
1025 static void ptlrpc_server_finish_active_request(
1026                                         struct ptlrpc_service_part *svcpt,
1027                                         struct ptlrpc_request *req)
1028 {
1029         spin_lock(&svcpt->scp_req_lock);
1030         ptlrpc_nrs_req_stop_nolock(req);
1031         svcpt->scp_nreqs_active--;
1032         if (req->rq_hp)
1033                 svcpt->scp_nhreqs_active--;
1034         spin_unlock(&svcpt->scp_req_lock);
1035
1036         ptlrpc_nrs_req_finalize(req);
1037
1038         if (req->rq_export != NULL)
1039                 class_export_rpc_dec(req->rq_export);
1040
1041         ptlrpc_server_finish_request(svcpt, req);
1042 }
1043
1044 /**
1045  * This function makes sure dead exports are evicted in a timely manner.
1046  * This function is only called when some export receives a message (i.e.,
1047  * the network is up.)
1048  */
1049 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1050 {
1051         struct obd_export *oldest_exp;
1052         time_t oldest_time, new_time;
1053
1054         ENTRY;
1055
1056         LASSERT(exp);
1057
1058         /* Compensate for slow machines, etc, by faking our request time
1059            into the future.  Although this can break the strict time-ordering
1060            of the list, we can be really lazy here - we don't have to evict
1061            at the exact right moment.  Eventually, all silent exports
1062            will make it to the top of the list. */
1063
1064         /* Do not pay attention on 1sec or smaller renewals. */
1065         new_time = cfs_time_current_sec() + extra_delay;
1066         if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1067                 RETURN_EXIT;
1068
1069         exp->exp_last_request_time = new_time;
1070
1071         /* exports may get disconnected from the chain even though the
1072            export has references, so we must keep the spin lock while
1073            manipulating the lists */
1074         spin_lock(&exp->exp_obd->obd_dev_lock);
1075
1076         if (list_empty(&exp->exp_obd_chain_timed)) {
1077                 /* this one is not timed */
1078                 spin_unlock(&exp->exp_obd->obd_dev_lock);
1079                 RETURN_EXIT;
1080         }
1081
1082         list_move_tail(&exp->exp_obd_chain_timed,
1083                        &exp->exp_obd->obd_exports_timed);
1084
1085         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1086                                 struct obd_export, exp_obd_chain_timed);
1087         oldest_time = oldest_exp->exp_last_request_time;
1088         spin_unlock(&exp->exp_obd->obd_dev_lock);
1089
1090         if (exp->exp_obd->obd_recovering) {
1091                 /* be nice to everyone during recovery */
1092                 EXIT;
1093                 return;
1094         }
1095
1096         /* Note - racing to start/reset the obd_eviction timer is safe */
1097         if (exp->exp_obd->obd_eviction_timer == 0) {
1098                 /* Check if the oldest entry is expired. */
1099                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1100                                               extra_delay)) {
1101                         /* We need a second timer, in case the net was down and
1102                          * it just came back. Since the pinger may skip every
1103                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1104                          * we better wait for 3. */
1105                         exp->exp_obd->obd_eviction_timer =
1106                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
1107                         CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
1108                                exp->exp_obd->obd_name,
1109                                obd_export_nid2str(oldest_exp), oldest_time);
1110                 }
1111         } else {
1112                 if (cfs_time_current_sec() >
1113                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1114                         /* The evictor won't evict anyone who we've heard from
1115                          * recently, so we don't have to check before we start
1116                          * it. */
1117                         if (!ping_evictor_wake(exp))
1118                                 exp->exp_obd->obd_eviction_timer = 0;
1119                 }
1120         }
1121
1122         EXIT;
1123 }
1124
1125 /**
1126  * Sanity check request \a req.
1127  * Return 0 if all is ok, error code otherwise.
1128  */
1129 static int ptlrpc_check_req(struct ptlrpc_request *req)
1130 {
1131         struct obd_device *obd = req->rq_export->exp_obd;
1132         int rc = 0;
1133
1134         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1135                      req->rq_export->exp_conn_cnt)) {
1136                 DEBUG_REQ(D_RPCTRACE, req,
1137                           "DROPPING req from old connection %d < %d",
1138                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
1139                           req->rq_export->exp_conn_cnt);
1140                 return -EEXIST;
1141         }
1142         if (unlikely(obd == NULL || obd->obd_fail)) {
1143                 /* Failing over, don't handle any more reqs,
1144                  * send error response instead. */
1145                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1146                         req, (obd != NULL) ? obd->obd_name : "unknown");
1147                 rc = -ENODEV;
1148         } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1149                    (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1150                    !obd->obd_recovering) {
1151                         DEBUG_REQ(D_ERROR, req,
1152                                   "Invalid replay without recovery");
1153                         class_fail_export(req->rq_export);
1154                         rc = -ENODEV;
1155         } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1156                    !obd->obd_recovering) {
1157                         DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1158                                   "%llu without recovery",
1159                                   lustre_msg_get_transno(req->rq_reqmsg));
1160                         class_fail_export(req->rq_export);
1161                         rc = -ENODEV;
1162         }
1163
1164         if (unlikely(rc < 0)) {
1165                 req->rq_status = rc;
1166                 ptlrpc_error(req);
1167         }
1168         return rc;
1169 }
1170
1171 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1172 {
1173         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1174         __s32 next;
1175
1176         if (array->paa_count == 0) {
1177                 del_timer(&svcpt->scp_at_timer);
1178                 return;
1179         }
1180
1181         /* Set timer for closest deadline */
1182         next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
1183                        at_early_margin);
1184         if (next <= 0) {
1185                 ptlrpc_at_timer((unsigned long)svcpt);
1186         } else {
1187                 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1188                 CDEBUG(D_INFO, "armed %s at %+ds\n",
1189                        svcpt->scp_service->srv_name, next);
1190         }
1191 }
1192
1193 /* Add rpc to early reply check list */
1194 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1195 {
1196         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1197         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1198         struct ptlrpc_request *rq = NULL;
1199         __u32 index;
1200
1201         if (AT_OFF)
1202                 return(0);
1203
1204         if (req->rq_no_reply)
1205                 return 0;
1206
1207         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1208                 return(-ENOSYS);
1209
1210         spin_lock(&svcpt->scp_at_lock);
1211         LASSERT(list_empty(&req->rq_timed_list));
1212
1213         div_u64_rem(req->rq_deadline, array->paa_size, &index);
1214         if (array->paa_reqs_count[index] > 0) {
1215                 /* latest rpcs will have the latest deadlines in the list,
1216                  * so search backward. */
1217                 list_for_each_entry_reverse(rq,
1218                                                 &array->paa_reqs_array[index],
1219                                                 rq_timed_list) {
1220                         if (req->rq_deadline >= rq->rq_deadline) {
1221                                 list_add(&req->rq_timed_list,
1222                                              &rq->rq_timed_list);
1223                                 break;
1224                         }
1225                 }
1226         }
1227
1228         /* Add the request at the head of the list */
1229         if (list_empty(&req->rq_timed_list))
1230                 list_add(&req->rq_timed_list,
1231                              &array->paa_reqs_array[index]);
1232
1233         spin_lock(&req->rq_lock);
1234         req->rq_at_linked = 1;
1235         spin_unlock(&req->rq_lock);
1236         req->rq_at_index = index;
1237         array->paa_reqs_count[index]++;
1238         array->paa_count++;
1239         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1240                 array->paa_deadline = req->rq_deadline;
1241                 ptlrpc_at_set_timer(svcpt);
1242         }
1243         spin_unlock(&svcpt->scp_at_lock);
1244
1245         return 0;
1246 }
1247
1248 static void
1249 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1250 {
1251         struct ptlrpc_at_array *array;
1252
1253         array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1254
1255         /* NB: must call with hold svcpt::scp_at_lock */
1256         LASSERT(!list_empty(&req->rq_timed_list));
1257         list_del_init(&req->rq_timed_list);
1258
1259         spin_lock(&req->rq_lock);
1260         req->rq_at_linked = 0;
1261         spin_unlock(&req->rq_lock);
1262
1263         array->paa_reqs_count[req->rq_at_index]--;
1264         array->paa_count--;
1265 }
1266
1267 /*
1268  * Attempt to extend the request deadline by sending an early reply to the
1269  * client.
1270  */
1271 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1272 {
1273         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1274         struct ptlrpc_request *reqcopy;
1275         struct lustre_msg *reqmsg;
1276         time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1277         time64_t newdl;
1278         int rc;
1279
1280         ENTRY;
1281
1282         if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1283                 /* don't send early reply */
1284                 RETURN(1);
1285         }
1286
1287         /* deadline is when the client expects us to reply, margin is the
1288            difference between clients' and servers' expectations */
1289         DEBUG_REQ(D_ADAPTTO, req,
1290                   "%ssending early reply (deadline %+llds, margin %+llds) for "
1291                   "%d+%d", AT_OFF ? "AT off - not " : "",
1292                   (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1293                   at_get(&svcpt->scp_at_estimate), at_extra);
1294
1295         if (AT_OFF)
1296                 RETURN(0);
1297
1298         if (olddl < 0) {
1299                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1300                           "not sending early reply. Consider increasing "
1301                           "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1302
1303                 /* Return an error so we're not re-added to the timed list. */
1304                 RETURN(-ETIMEDOUT);
1305         }
1306
1307         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1308                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1309                           "but no AT support");
1310                 RETURN(-ENOSYS);
1311         }
1312
1313         if (req->rq_export &&
1314             lustre_msg_get_flags(req->rq_reqmsg) &
1315             (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1316                 struct obd_device *obd_exp = req->rq_export->exp_obd;
1317
1318                 /* During recovery, we don't want to send too many early
1319                  * replies, but on the other hand we want to make sure the
1320                  * client has enough time to resend if the rpc is lost. So
1321                  * during the recovery period send at least 4 early replies,
1322                  * spacing them every at_extra if we can. at_estimate should
1323                  * always equal this fixed value during recovery.
1324                  */
1325                 /* Don't account request processing time into AT history
1326                  * during recovery, it is not service time we need but
1327                  * includes also waiting time for recovering clients
1328                  */
1329                 newdl = min_t(time64_t, at_extra,
1330                               obd_exp->obd_recovery_timeout / 4) +
1331                         ktime_get_real_seconds();
1332         } else {
1333                 /* We want to extend the request deadline by at_extra seconds,
1334                  * so we set our service estimate to reflect how much time has
1335                  * passed since this request arrived plus an additional
1336                  * at_extra seconds. The client will calculate the new deadline
1337                  * based on this service estimate (plus some additional time to
1338                  * account for network latency). See ptlrpc_at_recv_early_reply
1339                  */
1340                 at_measured(&svcpt->scp_at_estimate, at_extra +
1341                             ktime_get_real_seconds() -
1342                             req->rq_arrival_time.tv_sec);
1343                 newdl = req->rq_arrival_time.tv_sec +
1344                         at_get(&svcpt->scp_at_estimate);
1345         }
1346
1347         /* Check to see if we've actually increased the deadline -
1348          * we may be past adaptive_max */
1349         if (req->rq_deadline >= newdl) {
1350                 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1351                           (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1352                 RETURN(-ETIMEDOUT);
1353         }
1354
1355         reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1356         if (reqcopy == NULL)
1357                 RETURN(-ENOMEM);
1358         OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1359         if (!reqmsg)
1360                 GOTO(out_free, rc = -ENOMEM);
1361
1362         *reqcopy = *req;
1363         reqcopy->rq_reply_state = NULL;
1364         reqcopy->rq_rep_swab_mask = 0;
1365         reqcopy->rq_pack_bulk = 0;
1366         reqcopy->rq_pack_udesc = 0;
1367         reqcopy->rq_packed_final = 0;
1368         sptlrpc_svc_ctx_addref(reqcopy);
1369         /* We only need the reqmsg for the magic */
1370         reqcopy->rq_reqmsg = reqmsg;
1371         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1372
1373         /*
1374          * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1375          * Without this check, we would fail the rq_no_reply assertion in
1376          * ptlrpc_send_reply().
1377          */
1378         if (reqcopy->rq_no_reply)
1379                 GOTO(out, rc = -ETIMEDOUT);
1380
1381         LASSERT(atomic_read(&req->rq_refcount));
1382         /** if it is last refcount then early reply isn't needed */
1383         if (atomic_read(&req->rq_refcount) == 1) {
1384                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1385                           "abort sending early reply\n");
1386                 GOTO(out, rc = -EINVAL);
1387         }
1388
1389         /* Connection ref */
1390         reqcopy->rq_export = class_conn2export(
1391                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
1392         if (reqcopy->rq_export == NULL)
1393                 GOTO(out, rc = -ENODEV);
1394
1395         /* RPC ref */
1396         class_export_rpc_inc(reqcopy->rq_export);
1397         if (reqcopy->rq_export->exp_obd &&
1398             reqcopy->rq_export->exp_obd->obd_fail)
1399                 GOTO(out_put, rc = -ENODEV);
1400
1401         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1402         if (rc)
1403                 GOTO(out_put, rc);
1404
1405         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1406
1407         if (!rc) {
1408                 /* Adjust our own deadline to what we told the client */
1409                 req->rq_deadline = newdl;
1410                 req->rq_early_count++; /* number sent, server side */
1411         } else {
1412                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1413         }
1414
1415         /* Free the (early) reply state from lustre_pack_reply.
1416            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1417         ptlrpc_req_drop_rs(reqcopy);
1418
1419 out_put:
1420         class_export_rpc_dec(reqcopy->rq_export);
1421         class_export_put(reqcopy->rq_export);
1422 out:
1423         sptlrpc_svc_ctx_decref(reqcopy);
1424         OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1425 out_free:
1426         ptlrpc_request_cache_free(reqcopy);
1427         RETURN(rc);
1428 }
1429
1430 /* Send early replies to everybody expiring within at_early_margin
1431    asking for at_extra time */
1432 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1433 {
1434         struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1435         struct ptlrpc_request *rq, *n;
1436         struct list_head work_list;
1437         __u32  index, count;
1438         time64_t deadline;
1439         time64_t now = ktime_get_real_seconds();
1440         cfs_duration_t delay;
1441         int first, counter = 0;
1442         ENTRY;
1443
1444         spin_lock(&svcpt->scp_at_lock);
1445         if (svcpt->scp_at_check == 0) {
1446                 spin_unlock(&svcpt->scp_at_lock);
1447                 RETURN(0);
1448         }
1449         delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1450         svcpt->scp_at_check = 0;
1451
1452         if (array->paa_count == 0) {
1453                 spin_unlock(&svcpt->scp_at_lock);
1454                 RETURN(0);
1455         }
1456
1457         /* The timer went off, but maybe the nearest rpc already completed. */
1458         first = array->paa_deadline - now;
1459         if (first > at_early_margin) {
1460                 /* We've still got plenty of time.  Reset the timer. */
1461                 ptlrpc_at_set_timer(svcpt);
1462                 spin_unlock(&svcpt->scp_at_lock);
1463                 RETURN(0);
1464         }
1465
1466         /* We're close to a timeout, and we don't know how much longer the
1467            server will take. Send early replies to everyone expiring soon. */
1468         INIT_LIST_HEAD(&work_list);
1469         deadline = -1;
1470         div_u64_rem(array->paa_deadline, array->paa_size, &index);
1471         count = array->paa_count;
1472         while (count > 0) {
1473                 count -= array->paa_reqs_count[index];
1474                 list_for_each_entry_safe(rq, n,
1475                                          &array->paa_reqs_array[index],
1476                                          rq_timed_list) {
1477                         if (rq->rq_deadline > now + at_early_margin) {
1478                                 /* update the earliest deadline */
1479                                 if (deadline == -1 ||
1480                                     rq->rq_deadline < deadline)
1481                                         deadline = rq->rq_deadline;
1482                                 break;
1483                         }
1484
1485                         ptlrpc_at_remove_timed(rq);
1486                         /**
1487                          * ptlrpc_server_drop_request() may drop
1488                          * refcount to 0 already. Let's check this and
1489                          * don't add entry to work_list
1490                          */
1491                         if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1492                                 list_add(&rq->rq_timed_list, &work_list);
1493                         counter++;
1494                 }
1495
1496                 if (++index >= array->paa_size)
1497                         index = 0;
1498         }
1499         array->paa_deadline = deadline;
1500         /* we have a new earliest deadline, restart the timer */
1501         ptlrpc_at_set_timer(svcpt);
1502
1503         spin_unlock(&svcpt->scp_at_lock);
1504
1505         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1506                "replies\n", first, at_extra, counter);
1507         if (first < 0) {
1508                 /* We're already past request deadlines before we even get a
1509                    chance to send early replies */
1510                 LCONSOLE_WARN("%s: This server is not able to keep up with "
1511                               "request traffic (cpu-bound).\n",
1512                               svcpt->scp_service->srv_name);
1513                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1514                       counter, svcpt->scp_nreqs_incoming,
1515                       svcpt->scp_nreqs_active,
1516                       at_get(&svcpt->scp_at_estimate), delay);
1517         }
1518
1519         /* we took additional refcount so entries can't be deleted from list, no
1520          * locking is needed */
1521         while (!list_empty(&work_list)) {
1522                 rq = list_entry(work_list.next, struct ptlrpc_request,
1523                                     rq_timed_list);
1524                 list_del_init(&rq->rq_timed_list);
1525
1526                 if (ptlrpc_at_send_early_reply(rq) == 0)
1527                         ptlrpc_at_add_timed(rq);
1528
1529                 ptlrpc_server_drop_request(rq);
1530         }
1531
1532         RETURN(1); /* return "did_something" for liblustre */
1533 }
1534
1535 /* Check if we are already handling earlier incarnation of this request.
1536  * Called under &req->rq_export->exp_rpc_lock locked */
1537 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1538 {
1539         struct ptlrpc_request   *tmp = NULL;
1540
1541         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1542             (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1543                 return 0;
1544
1545         /* bulk request are aborted upon reconnect, don't try to
1546          * find a match */
1547         if (req->rq_bulk_write || req->rq_bulk_read)
1548                 return 0;
1549
1550         /* This list should not be longer than max_requests in
1551          * flights on the client, so it is not all that long.
1552          * Also we only hit this codepath in case of a resent
1553          * request which makes it even more rarely hit */
1554         list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1555                                 rq_exp_list) {
1556                 /* Found duplicate one */
1557                 if (tmp->rq_xid == req->rq_xid)
1558                         goto found;
1559         }
1560         list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1561                                 rq_exp_list) {
1562                 /* Found duplicate one */
1563                 if (tmp->rq_xid == req->rq_xid)
1564                         goto found;
1565         }
1566         return 0;
1567
1568 found:
1569         DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1570         DEBUG_REQ(D_HA, tmp, "Request being processed");
1571         return -EBUSY;
1572 }
1573
1574 /**
1575  * Check if a request should be assigned with a high priority.
1576  *
1577  * \retval      < 0: error occurred
1578  *                0: normal RPC request
1579  *               +1: high priority request
1580  */
1581 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1582                                     struct ptlrpc_request *req)
1583 {
1584         int rc = 0;
1585         ENTRY;
1586
1587         if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1588                 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1589                 if (rc < 0)
1590                         RETURN(rc);
1591
1592                 LASSERT(rc == 0);
1593         }
1594
1595         if (req->rq_export != NULL && req->rq_ops != NULL) {
1596                 /* Perform request specific check. We should do this
1597                  * check before the request is added into exp_hp_rpcs
1598                  * list otherwise it may hit swab race at LU-1044. */
1599                 if (req->rq_ops->hpreq_check != NULL) {
1600                         rc = req->rq_ops->hpreq_check(req);
1601                         if (rc == -ESTALE) {
1602                                 req->rq_status = rc;
1603                                 ptlrpc_error(req);
1604                         }
1605                         /** can only return error,
1606                          * 0 for normal request,
1607                          *  or 1 for high priority request */
1608                         LASSERT(rc <= 1);
1609                 }
1610         }
1611
1612         RETURN(rc);
1613 }
1614
1615 /** Remove the request from the export list. */
1616 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1617 {
1618         ENTRY;
1619         if (req->rq_export) {
1620                 /* refresh lock timeout again so that client has more
1621                  * room to send lock cancel RPC. */
1622                 if (req->rq_ops && req->rq_ops->hpreq_fini)
1623                         req->rq_ops->hpreq_fini(req);
1624
1625                 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1626                 list_del_init(&req->rq_exp_list);
1627                 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1628         }
1629         EXIT;
1630 }
1631
1632 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1633 {
1634         return 1;
1635 }
1636
1637 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1638         .hpreq_check       = ptlrpc_hpreq_check,
1639 };
1640
1641 /* Hi-Priority RPC check by RPC operation code. */
1642 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1643 {
1644         int opc = lustre_msg_get_opc(req->rq_reqmsg);
1645
1646         /* Check for export to let only reconnects for not yet evicted
1647          * export to become a HP rpc. */
1648         if ((req->rq_export != NULL) &&
1649             (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1650                 req->rq_ops = &ptlrpc_hpreq_common;
1651
1652         return 0;
1653 }
1654 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1655
1656 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1657                                      struct ptlrpc_request *req)
1658 {
1659         int rc;
1660         bool hp;
1661         ENTRY;
1662
1663         rc = ptlrpc_server_hpreq_init(svcpt, req);
1664         if (rc < 0)
1665                 RETURN(rc);
1666
1667         hp = rc > 0;
1668         ptlrpc_nrs_req_initialize(svcpt, req, hp);
1669
1670         if (req->rq_export != NULL) {
1671                 struct obd_export *exp = req->rq_export;
1672
1673                 /* do search for duplicated xid and the adding to the list
1674                  * atomically */
1675                 spin_lock_bh(&exp->exp_rpc_lock);
1676                 rc = ptlrpc_server_check_resend_in_progress(req);
1677                 if (rc < 0) {
1678                         spin_unlock_bh(&exp->exp_rpc_lock);
1679
1680                         ptlrpc_nrs_req_finalize(req);
1681                         RETURN(rc);
1682                 }
1683
1684                 if (hp || req->rq_ops != NULL)
1685                         list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1686                 else
1687                         list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1688                 spin_unlock_bh(&exp->exp_rpc_lock);
1689         }
1690
1691         /* the current thread is not the processing thread for this request
1692          * since that, but request is in exp_hp_list and can be find there.
1693          * Remove all relations between request and old thread. */
1694         req->rq_svc_thread->t_env->le_ses = NULL;
1695         req->rq_svc_thread = NULL;
1696         req->rq_session.lc_thread = NULL;
1697
1698         ptlrpc_nrs_req_add(svcpt, req, hp);
1699
1700         RETURN(0);
1701 }
1702
1703 /**
1704  * Allow to handle high priority request
1705  * User can call it w/o any lock but need to hold
1706  * ptlrpc_service_part::scp_req_lock to get reliable result
1707  */
1708 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1709                                      bool force)
1710 {
1711         int running = svcpt->scp_nthrs_running;
1712
1713         if (!nrs_svcpt_has_hp(svcpt))
1714                 return false;
1715
1716         if (force)
1717                 return true;
1718
1719         if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1720                 return false;
1721
1722         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1723                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1724                 /* leave just 1 thread for normal RPCs */
1725                 running = PTLRPC_NTHRS_INIT;
1726                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1727                         running += 1;
1728         }
1729
1730         if (svcpt->scp_nreqs_active >= running - 1)
1731                 return false;
1732
1733         if (svcpt->scp_nhreqs_active == 0)
1734                 return true;
1735
1736         return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1737                svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1738 }
1739
1740 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1741                                        bool force)
1742 {
1743         return ptlrpc_server_allow_high(svcpt, force) &&
1744                ptlrpc_nrs_req_pending_nolock(svcpt, true);
1745 }
1746
1747 /**
1748  * Only allow normal priority requests on a service that has a high-priority
1749  * queue if forced (i.e. cleanup), if there are other high priority requests
1750  * already being processed (i.e. those threads can service more high-priority
1751  * requests), or if there are enough idle threads that a later thread can do
1752  * a high priority request.
1753  * User can call it w/o any lock but need to hold
1754  * ptlrpc_service_part::scp_req_lock to get reliable result
1755  */
1756 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1757                                        bool force)
1758 {
1759         int running = svcpt->scp_nthrs_running;
1760         if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1761                      CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1762                 /* leave just 1 thread for normal RPCs */
1763                 running = PTLRPC_NTHRS_INIT;
1764                 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1765                         running += 1;
1766         }
1767
1768         if (force)
1769                 return true;
1770
1771         if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1772                 return false;
1773
1774         if (svcpt->scp_nreqs_active < running - 2)
1775                 return true;
1776
1777         if (svcpt->scp_nreqs_active >= running - 1)
1778                 return false;
1779
1780         return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1781 }
1782
1783 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1784                                          bool force)
1785 {
1786         return ptlrpc_server_allow_normal(svcpt, force) &&
1787                ptlrpc_nrs_req_pending_nolock(svcpt, false);
1788 }
1789
1790 /**
1791  * Returns true if there are requests available in incoming
1792  * request queue for processing and it is allowed to fetch them.
1793  * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1794  * to get reliable result
1795  * \see ptlrpc_server_allow_normal
1796  * \see ptlrpc_server_allow high
1797  */
1798 static inline bool
1799 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1800 {
1801         return ptlrpc_server_high_pending(svcpt, force) ||
1802                ptlrpc_server_normal_pending(svcpt, force);
1803 }
1804
1805 /**
1806  * Fetch a request for processing from queue of unprocessed requests.
1807  * Favors high-priority requests.
1808  * Returns a pointer to fetched request.
1809  */
1810 static struct ptlrpc_request *
1811 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1812 {
1813         struct ptlrpc_request *req = NULL;
1814         ENTRY;
1815
1816         spin_lock(&svcpt->scp_req_lock);
1817
1818         if (ptlrpc_server_high_pending(svcpt, force)) {
1819                 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1820                 if (req != NULL) {
1821                         svcpt->scp_hreq_count++;
1822                         goto got_request;
1823                 }
1824         }
1825
1826         if (ptlrpc_server_normal_pending(svcpt, force)) {
1827                 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1828                 if (req != NULL) {
1829                         svcpt->scp_hreq_count = 0;
1830                         goto got_request;
1831                 }
1832         }
1833
1834         spin_unlock(&svcpt->scp_req_lock);
1835         RETURN(NULL);
1836
1837 got_request:
1838         svcpt->scp_nreqs_active++;
1839         if (req->rq_hp)
1840                 svcpt->scp_nhreqs_active++;
1841
1842         spin_unlock(&svcpt->scp_req_lock);
1843
1844         if (likely(req->rq_export))
1845                 class_export_rpc_inc(req->rq_export);
1846
1847         RETURN(req);
1848 }
1849
1850 /**
1851  * Handle freshly incoming reqs, add to timed early reply list,
1852  * pass on to regular request queue.
1853  * All incoming requests pass through here before getting into
1854  * ptlrpc_server_handle_req later on.
1855  */
1856 static int
1857 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1858                             struct ptlrpc_thread *thread)
1859 {
1860         struct ptlrpc_service   *svc = svcpt->scp_service;
1861         struct ptlrpc_request   *req;
1862         __u32                   deadline;
1863         int                     rc;
1864         ENTRY;
1865
1866         spin_lock(&svcpt->scp_lock);
1867         if (list_empty(&svcpt->scp_req_incoming)) {
1868                 spin_unlock(&svcpt->scp_lock);
1869                 RETURN(0);
1870         }
1871
1872         req = list_entry(svcpt->scp_req_incoming.next,
1873                              struct ptlrpc_request, rq_list);
1874         list_del_init(&req->rq_list);
1875         svcpt->scp_nreqs_incoming--;
1876         /* Consider this still a "queued" request as far as stats are
1877          * concerned */
1878         spin_unlock(&svcpt->scp_lock);
1879
1880         /* go through security check/transform */
1881         rc = sptlrpc_svc_unwrap_request(req);
1882         switch (rc) {
1883         case SECSVC_OK:
1884                 break;
1885         case SECSVC_COMPLETE:
1886                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1887                 goto err_req;
1888         case SECSVC_DROP:
1889                 goto err_req;
1890         default:
1891                 LBUG();
1892         }
1893
1894         /*
1895          * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1896          * redo it wouldn't be harmful.
1897          */
1898         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1899                 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1900                 if (rc != 0) {
1901                         CERROR("error unpacking request: ptl %d from %s "
1902                                "x%llu\n", svc->srv_req_portal,
1903                                libcfs_id2str(req->rq_peer), req->rq_xid);
1904                         goto err_req;
1905                 }
1906         }
1907
1908         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1909         if (rc) {
1910                 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1911                         "%llu\n", svc->srv_req_portal,
1912                         libcfs_id2str(req->rq_peer), req->rq_xid);
1913                 goto err_req;
1914         }
1915
1916         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1917             lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1918                 CERROR("drop incoming rpc opc %u, x%llu\n",
1919                        cfs_fail_val, req->rq_xid);
1920                 goto err_req;
1921         }
1922
1923         rc = -EINVAL;
1924         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1925                 CERROR("wrong packet type received (type=%u) from %s\n",
1926                        lustre_msg_get_type(req->rq_reqmsg),
1927                        libcfs_id2str(req->rq_peer));
1928                 goto err_req;
1929         }
1930
1931         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1932         case MDS_WRITEPAGE:
1933         case OST_WRITE:
1934         case OUT_UPDATE:
1935                 req->rq_bulk_write = 1;
1936                 break;
1937         case MDS_READPAGE:
1938         case OST_READ:
1939         case MGS_CONFIG_READ:
1940                 req->rq_bulk_read = 1;
1941                 break;
1942         }
1943
1944         CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1945
1946         req->rq_export = class_conn2export(
1947                 lustre_msg_get_handle(req->rq_reqmsg));
1948         if (req->rq_export) {
1949                 rc = ptlrpc_check_req(req);
1950                 if (rc == 0) {
1951                         rc = sptlrpc_target_export_check(req->rq_export, req);
1952                         if (rc)
1953                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1954                                           "illegal security flavor,");
1955                 }
1956
1957                 if (rc)
1958                         goto err_req;
1959                 ptlrpc_update_export_timer(req->rq_export, 0);
1960         }
1961
1962         /* req_in handling should/must be fast */
1963         if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1964                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
1965                           (s64)(ktime_get_real_seconds() -
1966                                 req->rq_arrival_time.tv_sec));
1967
1968         /* Set rpc server deadline and add it to the timed list */
1969         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1970                     MSGHDR_AT_SUPPORT) ?
1971                    /* The max time the client expects us to take */
1972                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1973
1974         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1975         if (unlikely(deadline == 0)) {
1976                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1977                 goto err_req;
1978         }
1979
1980         /* Skip early reply */
1981         if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1982                 req->rq_deadline += obd_timeout;
1983
1984         req->rq_svc_thread = thread;
1985         if (thread != NULL) {
1986                 /* initialize request session, it is needed for request
1987                  * processing by target */
1988                 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1989                                                        LCT_NOREF);
1990                 if (rc) {
1991                         CERROR("%s: failure to initialize session: rc = %d\n",
1992                                thread->t_name, rc);
1993                         goto err_req;
1994                 }
1995                 req->rq_session.lc_thread = thread;
1996                 lu_context_enter(&req->rq_session);
1997                 thread->t_env->le_ses = &req->rq_session;
1998         }
1999
2000         ptlrpc_at_add_timed(req);
2001
2002         /* Move it over to the request processing queue */
2003         rc = ptlrpc_server_request_add(svcpt, req);
2004         if (rc)
2005                 GOTO(err_req, rc);
2006
2007         wake_up(&svcpt->scp_waitq);
2008         RETURN(1);
2009
2010 err_req:
2011         ptlrpc_server_finish_request(svcpt, req);
2012
2013         RETURN(1);
2014 }
2015
2016 /**
2017  * Main incoming request handling logic.
2018  * Calls handler function from service to do actual processing.
2019  */
2020 static int
2021 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2022                              struct ptlrpc_thread *thread)
2023 {
2024         struct ptlrpc_service *svc = svcpt->scp_service;
2025         struct ptlrpc_request *request;
2026         ktime_t work_start;
2027         ktime_t work_end;
2028         ktime_t arrived;
2029         s64 timediff_usecs;
2030         s64 arrived_usecs;
2031         int fail_opc = 0;
2032
2033         ENTRY;
2034
2035         request = ptlrpc_server_request_get(svcpt, false);
2036         if (request == NULL)
2037                 RETURN(0);
2038
2039         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2040                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2041         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2042                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2043
2044         if (unlikely(fail_opc)) {
2045                 if (request->rq_export && request->rq_ops)
2046                         OBD_FAIL_TIMEOUT(fail_opc, 4);
2047         }
2048
2049         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2050
2051         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2052                 libcfs_debug_dumplog();
2053
2054         work_start = ktime_get_real();
2055         arrived = timespec64_to_ktime(request->rq_arrival_time);
2056         timediff_usecs = ktime_us_delta(arrived, work_start);
2057         if (likely(svc->srv_stats != NULL)) {
2058                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2059                                     timediff_usecs);
2060                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2061                                     svcpt->scp_nreqs_incoming);
2062                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2063                                     svcpt->scp_nreqs_active);
2064                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2065                                     at_get(&svcpt->scp_at_estimate));
2066         }
2067
2068         if (likely(request->rq_export)) {
2069                 if (unlikely(ptlrpc_check_req(request)))
2070                         goto put_conn;
2071                 ptlrpc_update_export_timer(request->rq_export,
2072                                            timediff_usecs >> 19);
2073         }
2074
2075         /* Discard requests queued for longer than the deadline.
2076            The deadline is increased if we send an early reply. */
2077         if (ktime_get_real_seconds() > request->rq_deadline) {
2078                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2079                           libcfs_id2str(request->rq_peer),
2080                           request->rq_deadline -
2081                           request->rq_arrival_time.tv_sec,
2082                           ktime_get_real_seconds() - request->rq_deadline);
2083                 goto put_conn;
2084         }
2085
2086         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2087                "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2088                (request->rq_export ?
2089                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2090                (request->rq_export ?
2091                 atomic_read(&request->rq_export->exp_refcount) : -99),
2092                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2093                libcfs_id2str(request->rq_peer),
2094                lustre_msg_get_opc(request->rq_reqmsg));
2095
2096         if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2097                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2098
2099         CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2100
2101         /* re-assign request and sesson thread to the current one */
2102         request->rq_svc_thread = thread;
2103         if (thread != NULL) {
2104                 LASSERT(request->rq_session.lc_thread == NULL);
2105                 request->rq_session.lc_thread = thread;
2106                 thread->t_env->le_ses = &request->rq_session;
2107         }
2108         svc->srv_ops.so_req_handler(request);
2109
2110         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2111
2112 put_conn:
2113         if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2114                 DEBUG_REQ(D_WARNING, request, "Request took longer than estimated (%lld:%llds); client may timeout.",
2115                           request->rq_deadline -
2116                           request->rq_arrival_time.tv_sec,
2117                           ktime_get_real_seconds() - request->rq_deadline);
2118         }
2119
2120         work_end = ktime_get_real();
2121         timediff_usecs = ktime_us_delta(work_end, work_start);
2122         arrived_usecs = ktime_us_delta(work_end, arrived);
2123         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d Request procesed in %lldus (%lldus total) trans %llu rc %d/%d\n",
2124                 current_comm(),
2125                 (request->rq_export ?
2126                  (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2127                 (request->rq_export ?
2128                  atomic_read(&request->rq_export->exp_refcount) : -99),
2129                 lustre_msg_get_status(request->rq_reqmsg),
2130                 request->rq_xid,
2131                 libcfs_id2str(request->rq_peer),
2132                 lustre_msg_get_opc(request->rq_reqmsg),
2133                 timediff_usecs,
2134                 arrived_usecs,
2135                 (request->rq_repmsg ?
2136                  lustre_msg_get_transno(request->rq_repmsg) :
2137                  request->rq_transno),
2138                 request->rq_status,
2139                 (request->rq_repmsg ?
2140                  lustre_msg_get_status(request->rq_repmsg) : -999));
2141         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2142                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2143                 int opc = opcode_offset(op);
2144                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2145                         LASSERT(opc < LUSTRE_MAX_OPCODES);
2146                         lprocfs_counter_add(svc->srv_stats,
2147                                             opc + EXTRA_MAX_OPCODES,
2148                                             timediff_usecs);
2149                 }
2150         }
2151         if (unlikely(request->rq_early_count)) {
2152                 DEBUG_REQ(D_ADAPTTO, request,
2153                           "sent %d early replies before finishing in %llds",
2154                           request->rq_early_count,
2155                           arrived_usecs / USEC_PER_SEC);
2156         }
2157
2158         ptlrpc_server_finish_active_request(svcpt, request);
2159
2160         RETURN(1);
2161 }
2162
2163 /**
2164  * An internal function to process a single reply state object.
2165  */
2166 static int
2167 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2168 {
2169         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2170         struct ptlrpc_service     *svc = svcpt->scp_service;
2171         struct obd_export         *exp;
2172         int                        nlocks;
2173         int                        been_handled;
2174         ENTRY;
2175
2176         exp = rs->rs_export;
2177
2178         LASSERT(rs->rs_difficult);
2179         LASSERT(rs->rs_scheduled);
2180         LASSERT(list_empty(&rs->rs_list));
2181
2182         /* The disk commit callback holds exp_uncommitted_replies_lock while it
2183          * iterates over newly committed replies, removing them from
2184          * exp_uncommitted_replies.  It then drops this lock and schedules the
2185          * replies it found for handling here.
2186          *
2187          * We can avoid contention for exp_uncommitted_replies_lock between the
2188          * HRT threads and further commit callbacks by checking rs_committed
2189          * which is set in the commit callback while it holds both
2190          * rs_lock and exp_uncommitted_reples.
2191          *
2192          * If we see rs_committed clear, the commit callback _may_ not have
2193          * handled this reply yet and we race with it to grab
2194          * exp_uncommitted_replies_lock before removing the reply from
2195          * exp_uncommitted_replies.  Note that if we lose the race and the
2196          * reply has already been removed, list_del_init() is a noop.
2197          *
2198          * If we see rs_committed set, we know the commit callback is handling,
2199          * or has handled this reply since store reordering might allow us to
2200          * see rs_committed set out of sequence.  But since this is done
2201          * holding rs_lock, we can be sure it has all completed once we hold
2202          * rs_lock, which we do right next.
2203          */
2204         if (!rs->rs_committed) {
2205                 /* if rs was commited, no need to convert locks, don't check
2206                  * rs_committed here because rs may never be added into
2207                  * exp_uncommitted_replies and this flag never be set, see
2208                  * target_send_reply() */
2209                 if (rs->rs_convert_lock &&
2210                     rs->rs_transno > exp->exp_last_committed) {
2211                         struct ldlm_lock *lock;
2212
2213                         spin_lock(&rs->rs_lock);
2214                         if (rs->rs_convert_lock &&
2215                             rs->rs_transno > exp->exp_last_committed) {
2216                                 nlocks = rs->rs_nlocks;
2217                                 while (nlocks-- > 0)
2218                                         rs->rs_modes[nlocks] = LCK_COS;
2219                                 nlocks = rs->rs_nlocks;
2220                                 rs->rs_convert_lock = 0;
2221                                 /* clear rs_scheduled so that commit callback
2222                                  * can schedule again */
2223                                 rs->rs_scheduled = 0;
2224                                 spin_unlock(&rs->rs_lock);
2225
2226                                 while (nlocks-- > 0) {
2227                                         lock = ldlm_handle2lock(
2228                                                         &rs->rs_locks[nlocks]);
2229                                         LASSERT(lock != NULL);
2230                                         ldlm_lock_downgrade(lock, LCK_COS);
2231                                         LDLM_LOCK_PUT(lock);
2232                                 }
2233                                 RETURN(0);
2234                         }
2235                         spin_unlock(&rs->rs_lock);
2236                 }
2237
2238                 spin_lock(&exp->exp_uncommitted_replies_lock);
2239                 list_del_init(&rs->rs_obd_list);
2240                 spin_unlock(&exp->exp_uncommitted_replies_lock);
2241         }
2242
2243         spin_lock(&exp->exp_lock);
2244         /* Noop if removed already */
2245         list_del_init(&rs->rs_exp_list);
2246         spin_unlock(&exp->exp_lock);
2247
2248         spin_lock(&rs->rs_lock);
2249
2250         been_handled = rs->rs_handled;
2251         rs->rs_handled = 1;
2252
2253         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
2254         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
2255
2256         if (nlocks == 0 && !been_handled) {
2257                 /* If we see this, we should already have seen the warning
2258                  * in mds_steal_ack_locks()  */
2259                 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2260                        " o%d NID %s\n",
2261                        rs,
2262                        rs->rs_xid, rs->rs_transno, rs->rs_opc,
2263                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
2264         }
2265
2266         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2267                 spin_unlock(&rs->rs_lock);
2268
2269                 if (!been_handled && rs->rs_on_net) {
2270                         LNetMDUnlink(rs->rs_md_h);
2271                         /* Ignore return code; we're racing with completion */
2272                 }
2273
2274                 while (nlocks-- > 0)
2275                         ldlm_lock_decref(&rs->rs_locks[nlocks],
2276                                          rs->rs_modes[nlocks]);
2277
2278                 spin_lock(&rs->rs_lock);
2279         }
2280
2281         rs->rs_scheduled = 0;
2282         rs->rs_convert_lock = 0;
2283
2284         if (!rs->rs_on_net) {
2285                 /* Off the net */
2286                 spin_unlock(&rs->rs_lock);
2287
2288                 class_export_put (exp);
2289                 rs->rs_export = NULL;
2290                 ptlrpc_rs_decref(rs);
2291                 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2292                     svc->srv_is_stopping)
2293                         wake_up_all(&svcpt->scp_waitq);
2294                 RETURN(1);
2295         }
2296
2297         /* still on the net; callback will schedule */
2298         spin_unlock(&rs->rs_lock);
2299         RETURN(1);
2300 }
2301
2302
2303 static void
2304 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2305 {
2306         int avail = svcpt->scp_nrqbds_posted;
2307         int low_water = test_req_buffer_pressure ? 0 :
2308                         svcpt->scp_service->srv_nbuf_per_group / 2;
2309
2310         /* NB I'm not locking; just looking. */
2311
2312         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2313          * allowed the request history to grow out of control.  We could put a
2314          * sanity check on that here and cull some history if we need the
2315          * space. */
2316
2317         if (avail <= low_water)
2318                 ptlrpc_grow_req_bufs(svcpt, 1);
2319
2320         if (svcpt->scp_service->srv_stats) {
2321                 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2322                                     PTLRPC_REQBUF_AVAIL_CNTR, avail);
2323         }
2324 }
2325
2326 static int
2327 ptlrpc_retry_rqbds(void *arg)
2328 {
2329         struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2330
2331         svcpt->scp_rqbd_timeout = 0;
2332         return -ETIMEDOUT;
2333 }
2334
2335 static inline int
2336 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2337 {
2338         return svcpt->scp_nreqs_active <
2339                svcpt->scp_nthrs_running - 1 -
2340                (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2341 }
2342
2343 /**
2344  * allowed to create more threads
2345  * user can call it w/o any lock but need to hold
2346  * ptlrpc_service_part::scp_lock to get reliable result
2347  */
2348 static inline int
2349 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2350 {
2351         return svcpt->scp_nthrs_running +
2352                svcpt->scp_nthrs_starting <
2353                svcpt->scp_service->srv_nthrs_cpt_limit;
2354 }
2355
2356 /**
2357  * too many requests and allowed to create more threads
2358  */
2359 static inline int
2360 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2361 {
2362         return !ptlrpc_threads_enough(svcpt) &&
2363                 ptlrpc_threads_increasable(svcpt);
2364 }
2365
2366 static inline int
2367 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2368 {
2369         return thread_is_stopping(thread) ||
2370                thread->t_svcpt->scp_service->srv_is_stopping;
2371 }
2372
2373 static inline int
2374 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2375 {
2376         return !list_empty(&svcpt->scp_rqbd_idle) &&
2377                svcpt->scp_rqbd_timeout == 0;
2378 }
2379
2380 static inline int
2381 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2382 {
2383         return svcpt->scp_at_check;
2384 }
2385
2386 /**
2387  * requests wait on preprocessing
2388  * user can call it w/o any lock but need to hold
2389  * ptlrpc_service_part::scp_lock to get reliable result
2390  */
2391 static inline int
2392 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2393 {
2394         return !list_empty(&svcpt->scp_req_incoming);
2395 }
2396
2397 static __attribute__((__noinline__)) int
2398 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2399                   struct ptlrpc_thread *thread)
2400 {
2401         /* Don't exit while there are replies to be handled */
2402         struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2403                                              ptlrpc_retry_rqbds, svcpt);
2404
2405         lc_watchdog_disable(thread->t_watchdog);
2406
2407         cond_resched();
2408
2409         l_wait_event_exclusive_head(svcpt->scp_waitq,
2410                                 ptlrpc_thread_stopping(thread) ||
2411                                 ptlrpc_server_request_incoming(svcpt) ||
2412                                 ptlrpc_server_request_pending(svcpt, false) ||
2413                                 ptlrpc_rqbd_pending(svcpt) ||
2414                                 ptlrpc_at_check(svcpt), &lwi);
2415
2416         if (ptlrpc_thread_stopping(thread))
2417                 return -EINTR;
2418
2419         lc_watchdog_touch(thread->t_watchdog,
2420                           ptlrpc_server_get_timeout(svcpt));
2421         return 0;
2422 }
2423
2424 /**
2425  * Main thread body for service threads.
2426  * Waits in a loop waiting for new requests to process to appear.
2427  * Every time an incoming requests is added to its queue, a waitq
2428  * is woken up and one of the threads will handle it.
2429  */
2430 static int ptlrpc_main(void *arg)
2431 {
2432         struct ptlrpc_thread            *thread = (struct ptlrpc_thread *)arg;
2433         struct ptlrpc_service_part      *svcpt = thread->t_svcpt;
2434         struct ptlrpc_service           *svc = svcpt->scp_service;
2435         struct ptlrpc_reply_state       *rs;
2436         struct group_info *ginfo = NULL;
2437         struct lu_env *env;
2438         int counter = 0, rc = 0;
2439         ENTRY;
2440
2441         thread->t_pid = current_pid();
2442         unshare_fs_struct();
2443
2444         /* NB: we will call cfs_cpt_bind() for all threads, because we
2445          * might want to run lustre server only on a subset of system CPUs,
2446          * in that case ->scp_cpt is CFS_CPT_ANY */
2447         rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2448         if (rc != 0) {
2449                 CWARN("%s: failed to bind %s on CPT %d\n",
2450                       svc->srv_name, thread->t_name, svcpt->scp_cpt);
2451         }
2452
2453         ginfo = groups_alloc(0);
2454         if (!ginfo) {
2455                 rc = -ENOMEM;
2456                 goto out;
2457         }
2458
2459         set_current_groups(ginfo);
2460         put_group_info(ginfo);
2461
2462         if (svc->srv_ops.so_thr_init != NULL) {
2463                 rc = svc->srv_ops.so_thr_init(thread);
2464                 if (rc)
2465                         goto out;
2466         }
2467
2468         OBD_ALLOC_PTR(env);
2469         if (env == NULL) {
2470                 rc = -ENOMEM;
2471                 goto out_srv_fini;
2472         }
2473
2474         rc = lu_context_init(&env->le_ctx,
2475                              svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2476         if (rc)
2477                 goto out_srv_fini;
2478
2479         thread->t_env = env;
2480         env->le_ctx.lc_thread = thread;
2481         env->le_ctx.lc_cookie = 0x6;
2482
2483         while (!list_empty(&svcpt->scp_rqbd_idle)) {
2484                 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2485                 if (rc >= 0)
2486                         continue;
2487
2488                 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2489                         svc->srv_name, svcpt->scp_cpt, rc);
2490                 goto out_srv_fini;
2491         }
2492
2493         /* Alloc reply state structure for this one */
2494         OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2495         if (!rs) {
2496                 rc = -ENOMEM;
2497                 goto out_srv_fini;
2498         }
2499
2500         spin_lock(&svcpt->scp_lock);
2501
2502         LASSERT(thread_is_starting(thread));
2503         thread_clear_flags(thread, SVC_STARTING);
2504
2505         LASSERT(svcpt->scp_nthrs_starting == 1);
2506         svcpt->scp_nthrs_starting--;
2507
2508         /* SVC_STOPPING may already be set here if someone else is trying
2509          * to stop the service while this new thread has been dynamically
2510          * forked. We still set SVC_RUNNING to let our creator know that
2511          * we are now running, however we will exit as soon as possible */
2512         thread_add_flags(thread, SVC_RUNNING);
2513         svcpt->scp_nthrs_running++;
2514         spin_unlock(&svcpt->scp_lock);
2515
2516         /* wake up our creator in case he's still waiting. */
2517         wake_up(&thread->t_ctl_waitq);
2518
2519         thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2520                                              NULL, NULL);
2521
2522         spin_lock(&svcpt->scp_rep_lock);
2523         list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2524         wake_up(&svcpt->scp_rep_waitq);
2525         spin_unlock(&svcpt->scp_rep_lock);
2526
2527         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2528                svcpt->scp_nthrs_running);
2529
2530         /* XXX maintain a list of all managed devices: insert here */
2531         while (!ptlrpc_thread_stopping(thread)) {
2532                 if (ptlrpc_wait_event(svcpt, thread))
2533                         break;
2534
2535                 ptlrpc_check_rqbd_pool(svcpt);
2536
2537                 if (ptlrpc_threads_need_create(svcpt)) {
2538                         /* Ignore return code - we tried... */
2539                         ptlrpc_start_thread(svcpt, 0);
2540                 }
2541
2542                 /* reset le_ses to initial state */
2543                 env->le_ses = NULL;
2544                 /* Process all incoming reqs before handling any */
2545                 if (ptlrpc_server_request_incoming(svcpt)) {
2546                         lu_context_enter(&env->le_ctx);
2547                         ptlrpc_server_handle_req_in(svcpt, thread);
2548                         lu_context_exit(&env->le_ctx);
2549
2550                         /* but limit ourselves in case of flood */
2551                         if (counter++ < 100)
2552                                 continue;
2553                         counter = 0;
2554                 }
2555
2556                 if (ptlrpc_at_check(svcpt))
2557                         ptlrpc_at_check_timed(svcpt);
2558
2559                 if (ptlrpc_server_request_pending(svcpt, false)) {
2560                         lu_context_enter(&env->le_ctx);
2561                         ptlrpc_server_handle_request(svcpt, thread);
2562                         lu_context_exit(&env->le_ctx);
2563                 }
2564
2565                 if (ptlrpc_rqbd_pending(svcpt) &&
2566                     ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2567                         /* I just failed to repost request buffers.
2568                          * Wait for a timeout (unless something else
2569                          * happens) before I try again */
2570                         svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2571                         CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2572                                svcpt->scp_nrqbds_posted);
2573                 }
2574         }
2575
2576         lc_watchdog_delete(thread->t_watchdog);
2577         thread->t_watchdog = NULL;
2578
2579 out_srv_fini:
2580         /*
2581          * deconstruct service specific state created by ptlrpc_start_thread()
2582          */
2583         if (svc->srv_ops.so_thr_done != NULL)
2584                 svc->srv_ops.so_thr_done(thread);
2585
2586         if (env != NULL) {
2587                 lu_context_fini(&env->le_ctx);
2588                 OBD_FREE_PTR(env);
2589         }
2590 out:
2591         CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2592                thread, thread->t_pid, thread->t_id, rc);
2593
2594         spin_lock(&svcpt->scp_lock);
2595         if (thread_test_and_clear_flags(thread, SVC_STARTING))
2596                 svcpt->scp_nthrs_starting--;
2597
2598         if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2599                 /* must know immediately */
2600                 svcpt->scp_nthrs_running--;
2601         }
2602
2603         thread->t_id = rc;
2604         thread_add_flags(thread, SVC_STOPPED);
2605
2606         wake_up(&thread->t_ctl_waitq);
2607         spin_unlock(&svcpt->scp_lock);
2608
2609         return rc;
2610 }
2611
2612 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2613                           struct list_head *replies)
2614 {
2615         int result;
2616
2617         spin_lock(&hrt->hrt_lock);
2618
2619         list_splice_init(&hrt->hrt_queue, replies);
2620         result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2621
2622         spin_unlock(&hrt->hrt_lock);
2623         return result;
2624 }
2625
2626 /**
2627  * Main body of "handle reply" function.
2628  * It processes acked reply states
2629  */
2630 static int ptlrpc_hr_main(void *arg)
2631 {
2632         struct ptlrpc_hr_thread         *hrt = (struct ptlrpc_hr_thread *)arg;
2633         struct ptlrpc_hr_partition      *hrp = hrt->hrt_partition;
2634         struct list_head                replies;
2635         int                             rc;
2636
2637         INIT_LIST_HEAD(&replies);
2638         unshare_fs_struct();
2639
2640         rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2641         if (rc != 0) {
2642                 char threadname[20];
2643
2644                 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2645                          hrp->hrp_cpt, hrt->hrt_id);
2646                 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2647                       threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2648         }
2649
2650         atomic_inc(&hrp->hrp_nstarted);
2651         wake_up(&ptlrpc_hr.hr_waitq);
2652
2653         while (!ptlrpc_hr.hr_stopping) {
2654                 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2655
2656                 while (!list_empty(&replies)) {
2657                         struct ptlrpc_reply_state *rs;
2658
2659                         rs = list_entry(replies.prev,
2660                                         struct ptlrpc_reply_state,
2661                                         rs_list);
2662                         list_del_init(&rs->rs_list);
2663                         ptlrpc_handle_rs(rs);
2664                 }
2665         }
2666
2667         atomic_inc(&hrp->hrp_nstopped);
2668         wake_up(&ptlrpc_hr.hr_waitq);
2669
2670         return 0;
2671 }
2672
2673 static void ptlrpc_stop_hr_threads(void)
2674 {
2675         struct ptlrpc_hr_partition      *hrp;
2676         int                             i;
2677         int                             j;
2678
2679         ptlrpc_hr.hr_stopping = 1;
2680
2681         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2682                 if (hrp->hrp_thrs == NULL)
2683                         continue; /* uninitialized */
2684                 for (j = 0; j < hrp->hrp_nthrs; j++)
2685                         wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2686         }
2687
2688         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2689                 if (hrp->hrp_thrs == NULL)
2690                         continue; /* uninitialized */
2691                 wait_event(ptlrpc_hr.hr_waitq,
2692                                atomic_read(&hrp->hrp_nstopped) ==
2693                                atomic_read(&hrp->hrp_nstarted));
2694         }
2695 }
2696
2697 static int ptlrpc_start_hr_threads(void)
2698 {
2699         struct ptlrpc_hr_partition      *hrp;
2700         int                             i;
2701         int                             j;
2702         ENTRY;
2703
2704         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2705                 int     rc = 0;
2706
2707                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2708                         struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2709                         struct task_struct *task;
2710
2711                         task = kthread_run(ptlrpc_hr_main,
2712                                            &hrp->hrp_thrs[j],
2713                                            "ptlrpc_hr%02d_%03d",
2714                                            hrp->hrp_cpt,
2715                                            hrt->hrt_id);
2716                         if (IS_ERR(task)) {
2717                                 rc = PTR_ERR(task);
2718                                 break;
2719                         }
2720                 }
2721
2722                 wait_event(ptlrpc_hr.hr_waitq,
2723                            atomic_read(&hrp->hrp_nstarted) == j);
2724
2725                 if (rc < 0) {
2726                         CERROR("cannot start reply handler thread %d:%d: "
2727                                "rc = %d\n", i, j, rc);
2728                         ptlrpc_stop_hr_threads();
2729                         RETURN(rc);
2730                 }
2731         }
2732
2733         RETURN(0);
2734 }
2735
2736 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2737 {
2738         struct l_wait_info      lwi = { 0 };
2739         struct ptlrpc_thread    *thread;
2740         struct list_head        zombie;
2741
2742         ENTRY;
2743
2744         CDEBUG(D_INFO, "Stopping threads for service %s\n",
2745                svcpt->scp_service->srv_name);
2746
2747         INIT_LIST_HEAD(&zombie);
2748         spin_lock(&svcpt->scp_lock);
2749         /* let the thread know that we would like it to stop asap */
2750         list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2751                 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2752                        svcpt->scp_service->srv_thread_name, thread->t_id);
2753                 thread_add_flags(thread, SVC_STOPPING);
2754         }
2755
2756         wake_up_all(&svcpt->scp_waitq);
2757
2758         while (!list_empty(&svcpt->scp_threads)) {
2759                 thread = list_entry(svcpt->scp_threads.next,
2760                                         struct ptlrpc_thread, t_link);
2761                 if (thread_is_stopped(thread)) {
2762                         list_del(&thread->t_link);
2763                         list_add(&thread->t_link, &zombie);
2764                         continue;
2765                 }
2766                 spin_unlock(&svcpt->scp_lock);
2767
2768                 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2769                        svcpt->scp_service->srv_thread_name, thread->t_id);
2770                 l_wait_event(thread->t_ctl_waitq,
2771                              thread_is_stopped(thread), &lwi);
2772
2773                 spin_lock(&svcpt->scp_lock);
2774         }
2775
2776         spin_unlock(&svcpt->scp_lock);
2777
2778         while (!list_empty(&zombie)) {
2779                 thread = list_entry(zombie.next,
2780                                         struct ptlrpc_thread, t_link);
2781                 list_del(&thread->t_link);
2782                 OBD_FREE_PTR(thread);
2783         }
2784         EXIT;
2785 }
2786
2787 /**
2788  * Stops all threads of a particular service \a svc
2789  */
2790 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2791 {
2792         struct ptlrpc_service_part *svcpt;
2793         int                        i;
2794         ENTRY;
2795
2796         ptlrpc_service_for_each_part(svcpt, i, svc) {
2797                 if (svcpt->scp_service != NULL)
2798                         ptlrpc_svcpt_stop_threads(svcpt);
2799         }
2800
2801         EXIT;
2802 }
2803
2804 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2805 {
2806         int     rc = 0;
2807         int     i;
2808         int     j;
2809         ENTRY;
2810
2811         /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2812         LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2813
2814         for (i = 0; i < svc->srv_ncpts; i++) {
2815                 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2816                         rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2817                         if (rc == 0)
2818                                 continue;
2819
2820                         if (rc != -EMFILE)
2821                                 goto failed;
2822                         /* We have enough threads, don't start more. b=15759 */
2823                         break;
2824                 }
2825         }
2826
2827         RETURN(0);
2828  failed:
2829         CERROR("cannot start %s thread #%d_%d: rc %d\n",
2830                svc->srv_thread_name, i, j, rc);
2831         ptlrpc_stop_all_threads(svc);
2832         RETURN(rc);
2833 }
2834
2835 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2836 {
2837         struct l_wait_info      lwi = { 0 };
2838         struct ptlrpc_thread    *thread;
2839         struct ptlrpc_service   *svc;
2840         struct task_struct      *task;
2841         int                     rc;
2842         ENTRY;
2843
2844         LASSERT(svcpt != NULL);
2845
2846         svc = svcpt->scp_service;
2847
2848         CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2849                svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2850                svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2851
2852  again:
2853         if (unlikely(svc->srv_is_stopping))
2854                 RETURN(-ESRCH);
2855
2856         if (!ptlrpc_threads_increasable(svcpt) ||
2857             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2858              svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2859                 RETURN(-EMFILE);
2860
2861         OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2862         if (thread == NULL)
2863                 RETURN(-ENOMEM);
2864         init_waitqueue_head(&thread->t_ctl_waitq);
2865
2866         spin_lock(&svcpt->scp_lock);
2867         if (!ptlrpc_threads_increasable(svcpt)) {
2868                 spin_unlock(&svcpt->scp_lock);
2869                 OBD_FREE_PTR(thread);
2870                 RETURN(-EMFILE);
2871         }
2872
2873         if (svcpt->scp_nthrs_starting != 0) {
2874                 /* serialize starting because some modules (obdfilter)
2875                  * might require unique and contiguous t_id */
2876                 LASSERT(svcpt->scp_nthrs_starting == 1);
2877                 spin_unlock(&svcpt->scp_lock);
2878                 OBD_FREE_PTR(thread);
2879                 if (wait) {
2880                         CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2881                                svc->srv_thread_name, svcpt->scp_thr_nextid);
2882                         schedule();
2883                         goto again;
2884                 }
2885
2886                 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2887                        svc->srv_thread_name, svcpt->scp_thr_nextid);
2888                 RETURN(-EAGAIN);
2889         }
2890
2891         svcpt->scp_nthrs_starting++;
2892         thread->t_id = svcpt->scp_thr_nextid++;
2893         thread_add_flags(thread, SVC_STARTING);
2894         thread->t_svcpt = svcpt;
2895
2896         list_add(&thread->t_link, &svcpt->scp_threads);
2897         spin_unlock(&svcpt->scp_lock);
2898
2899         if (svcpt->scp_cpt >= 0) {
2900                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2901                          svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2902         } else {
2903                 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2904                          svc->srv_thread_name, thread->t_id);
2905         }
2906
2907         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2908         task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2909         if (IS_ERR(task)) {
2910                 rc = PTR_ERR(task);
2911                 CERROR("cannot start thread '%s': rc = %d\n",
2912                        thread->t_name, rc);
2913                 spin_lock(&svcpt->scp_lock);
2914                 --svcpt->scp_nthrs_starting;
2915                 if (thread_is_stopping(thread)) {
2916                         /* this ptlrpc_thread is being hanled
2917                          * by ptlrpc_svcpt_stop_threads now
2918                          */
2919                         thread_add_flags(thread, SVC_STOPPED);
2920                         wake_up(&thread->t_ctl_waitq);
2921                         spin_unlock(&svcpt->scp_lock);
2922                 } else {
2923                         list_del(&thread->t_link);
2924                         spin_unlock(&svcpt->scp_lock);
2925                         OBD_FREE_PTR(thread);
2926                 }
2927                 RETURN(rc);
2928         }
2929
2930         if (!wait)
2931                 RETURN(0);
2932
2933         l_wait_event(thread->t_ctl_waitq,
2934                      thread_is_running(thread) || thread_is_stopped(thread),
2935                      &lwi);
2936
2937         rc = thread_is_stopped(thread) ? thread->t_id : 0;
2938         RETURN(rc);
2939 }
2940
2941 int ptlrpc_hr_init(void)
2942 {
2943         struct ptlrpc_hr_partition      *hrp;
2944         struct ptlrpc_hr_thread         *hrt;
2945         int                             rc;
2946         int                             i;
2947         int                             j;
2948         int                             weight;
2949         ENTRY;
2950
2951         memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2952         ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2953
2954         ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2955                                                    sizeof(*hrp));
2956         if (ptlrpc_hr.hr_partitions == NULL)
2957                 RETURN(-ENOMEM);
2958
2959         init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2960
2961         weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
2962
2963         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2964                 hrp->hrp_cpt = i;
2965
2966                 atomic_set(&hrp->hrp_nstarted, 0);
2967                 atomic_set(&hrp->hrp_nstopped, 0);
2968
2969                 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2970
2971                 hrp->hrp_nthrs /= weight;
2972                 if (hrp->hrp_nthrs == 0)
2973                         hrp->hrp_nthrs = 1;
2974
2975                 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2976                               hrp->hrp_nthrs * sizeof(*hrt));
2977                 if (hrp->hrp_thrs == NULL)
2978                         GOTO(out, rc = -ENOMEM);
2979
2980                 for (j = 0; j < hrp->hrp_nthrs; j++) {
2981                         hrt = &hrp->hrp_thrs[j];
2982
2983                         hrt->hrt_id = j;
2984                         hrt->hrt_partition = hrp;
2985                         init_waitqueue_head(&hrt->hrt_waitq);
2986                         spin_lock_init(&hrt->hrt_lock);
2987                         INIT_LIST_HEAD(&hrt->hrt_queue);
2988                 }
2989         }
2990
2991         rc = ptlrpc_start_hr_threads();
2992 out:
2993         if (rc != 0)
2994                 ptlrpc_hr_fini();
2995         RETURN(rc);
2996 }
2997
2998 void ptlrpc_hr_fini(void)
2999 {
3000         struct ptlrpc_hr_partition      *hrp;
3001         int                             i;
3002
3003         if (ptlrpc_hr.hr_partitions == NULL)
3004                 return;
3005
3006         ptlrpc_stop_hr_threads();
3007
3008         cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3009                 if (hrp->hrp_thrs != NULL) {
3010                         OBD_FREE(hrp->hrp_thrs,
3011                                  hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3012                 }
3013         }
3014
3015         cfs_percpt_free(ptlrpc_hr.hr_partitions);
3016         ptlrpc_hr.hr_partitions = NULL;
3017 }
3018
3019
3020 /**
3021  * Wait until all already scheduled replies are processed.
3022  */
3023 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3024 {
3025         while (1) {
3026                 int rc;
3027                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3028                                                      NULL, NULL);
3029
3030                 rc = l_wait_event(svcpt->scp_waitq,
3031                      atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3032                 if (rc == 0)
3033                         break;
3034                 CWARN("Unexpectedly long timeout %s %p\n",
3035                       svcpt->scp_service->srv_name, svcpt->scp_service);
3036         }
3037 }
3038
3039 static void
3040 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3041 {
3042         struct ptlrpc_service_part      *svcpt;
3043         int                             i;
3044
3045         /* early disarm AT timer... */
3046         ptlrpc_service_for_each_part(svcpt, i, svc) {
3047                 if (svcpt->scp_service != NULL)
3048                         del_timer(&svcpt->scp_at_timer);
3049         }
3050 }
3051
3052 static void
3053 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3054 {
3055         struct ptlrpc_service_part        *svcpt;
3056         struct ptlrpc_request_buffer_desc *rqbd;
3057         struct l_wait_info                lwi;
3058         int                               rc;
3059         int                               i;
3060
3061         /* All history will be culled when the next request buffer is
3062          * freed in ptlrpc_service_purge_all() */
3063         svc->srv_hist_nrqbds_cpt_max = 0;
3064
3065         rc = LNetClearLazyPortal(svc->srv_req_portal);
3066         LASSERT(rc == 0);
3067
3068         ptlrpc_service_for_each_part(svcpt, i, svc) {
3069                 if (svcpt->scp_service == NULL)
3070                         break;
3071
3072                 /* Unlink all the request buffers.  This forces a 'final'
3073                  * event with its 'unlink' flag set for each posted rqbd */
3074                 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3075                                         rqbd_list) {
3076                         rc = LNetMDUnlink(rqbd->rqbd_md_h);
3077                         LASSERT(rc == 0 || rc == -ENOENT);
3078                 }
3079         }
3080
3081         ptlrpc_service_for_each_part(svcpt, i, svc) {
3082                 if (svcpt->scp_service == NULL)
3083                         break;
3084
3085                 /* Wait for the network to release any buffers
3086                  * it's currently filling */
3087                 spin_lock(&svcpt->scp_lock);
3088                 while (svcpt->scp_nrqbds_posted != 0) {
3089                         spin_unlock(&svcpt->scp_lock);
3090                         /* Network access will complete in finite time but
3091                          * the HUGE timeout lets us CWARN for visibility
3092                          * of sluggish NALs */
3093                         lwi = LWI_TIMEOUT_INTERVAL(
3094                                         cfs_time_seconds(LONG_UNLINK),
3095                                         cfs_time_seconds(1), NULL, NULL);
3096                         rc = l_wait_event(svcpt->scp_waitq,
3097                                           svcpt->scp_nrqbds_posted == 0, &lwi);
3098                         if (rc == -ETIMEDOUT) {
3099                                 CWARN("Service %s waiting for "
3100                                       "request buffers\n",
3101                                       svcpt->scp_service->srv_name);
3102                         }
3103                         spin_lock(&svcpt->scp_lock);
3104                 }
3105                 spin_unlock(&svcpt->scp_lock);
3106         }
3107 }
3108
3109 static void
3110 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3111 {
3112         struct ptlrpc_service_part              *svcpt;
3113         struct ptlrpc_request_buffer_desc       *rqbd;
3114         struct ptlrpc_request                   *req;
3115         struct ptlrpc_reply_state               *rs;
3116         int                                     i;
3117
3118         ptlrpc_service_for_each_part(svcpt, i, svc) {
3119                 if (svcpt->scp_service == NULL)
3120                         break;
3121
3122                 spin_lock(&svcpt->scp_rep_lock);
3123                 while (!list_empty(&svcpt->scp_rep_active)) {
3124                         rs = list_entry(svcpt->scp_rep_active.next,
3125                                             struct ptlrpc_reply_state, rs_list);
3126                         spin_lock(&rs->rs_lock);
3127                         ptlrpc_schedule_difficult_reply(rs);
3128                         spin_unlock(&rs->rs_lock);
3129                 }
3130                 spin_unlock(&svcpt->scp_rep_lock);
3131
3132                 /* purge the request queue.  NB No new replies (rqbds
3133                  * all unlinked) and no service threads, so I'm the only
3134                  * thread noodling the request queue now */
3135                 while (!list_empty(&svcpt->scp_req_incoming)) {
3136                         req = list_entry(svcpt->scp_req_incoming.next,
3137                                              struct ptlrpc_request, rq_list);
3138
3139                         list_del(&req->rq_list);
3140                         svcpt->scp_nreqs_incoming--;
3141                         ptlrpc_server_finish_request(svcpt, req);
3142                 }
3143
3144                 while (ptlrpc_server_request_pending(svcpt, true)) {
3145                         req = ptlrpc_server_request_get(svcpt, true);
3146                         ptlrpc_server_finish_active_request(svcpt, req);
3147                 }
3148
3149                 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3150                 LASSERT(svcpt->scp_nreqs_incoming == 0);
3151                 LASSERT(svcpt->scp_nreqs_active == 0);
3152                 /* history should have been culled by
3153                  * ptlrpc_server_finish_request */
3154                 LASSERT(svcpt->scp_hist_nrqbds == 0);
3155
3156                 /* Now free all the request buffers since nothing
3157                  * references them any more... */
3158
3159                 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3160                         rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3161                                               struct ptlrpc_request_buffer_desc,
3162                                               rqbd_list);
3163                         ptlrpc_free_rqbd(rqbd);
3164                 }
3165                 ptlrpc_wait_replies(svcpt);
3166
3167                 while (!list_empty(&svcpt->scp_rep_idle)) {
3168                         rs = list_entry(svcpt->scp_rep_idle.next,
3169                                             struct ptlrpc_reply_state,
3170                                             rs_list);
3171                         list_del(&rs->rs_list);
3172                         OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3173                 }
3174         }
3175 }
3176
3177 static void
3178 ptlrpc_service_free(struct ptlrpc_service *svc)
3179 {
3180         struct ptlrpc_service_part      *svcpt;
3181         struct ptlrpc_at_array          *array;
3182         int                             i;
3183
3184         ptlrpc_service_for_each_part(svcpt, i, svc) {
3185                 if (svcpt->scp_service == NULL)
3186                         break;
3187
3188                 /* In case somebody rearmed this in the meantime */
3189                 del_timer(&svcpt->scp_at_timer);
3190                 array = &svcpt->scp_at_array;
3191
3192                 if (array->paa_reqs_array != NULL) {
3193                         OBD_FREE(array->paa_reqs_array,
3194                                  sizeof(struct list_head) * array->paa_size);
3195                         array->paa_reqs_array = NULL;
3196                 }
3197
3198                 if (array->paa_reqs_count != NULL) {
3199                         OBD_FREE(array->paa_reqs_count,
3200                                  sizeof(__u32) * array->paa_size);
3201                         array->paa_reqs_count = NULL;
3202                 }
3203         }
3204
3205         ptlrpc_service_for_each_part(svcpt, i, svc)
3206                 OBD_FREE_PTR(svcpt);
3207
3208         if (svc->srv_cpts != NULL)
3209                 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3210
3211         OBD_FREE(svc, offsetof(struct ptlrpc_service,
3212                                srv_parts[svc->srv_ncpts]));
3213 }
3214
3215 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3216 {
3217         ENTRY;
3218
3219         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3220
3221         service->srv_is_stopping = 1;
3222
3223         mutex_lock(&ptlrpc_all_services_mutex);
3224         list_del_init(&service->srv_list);
3225         mutex_unlock(&ptlrpc_all_services_mutex);
3226
3227         ptlrpc_service_del_atimer(service);
3228         ptlrpc_stop_all_threads(service);
3229
3230         ptlrpc_service_unlink_rqbd(service);
3231         ptlrpc_service_purge_all(service);
3232         ptlrpc_service_nrs_cleanup(service);
3233
3234         ptlrpc_lprocfs_unregister_service(service);
3235
3236         ptlrpc_service_free(service);
3237
3238         RETURN(0);
3239 }
3240 EXPORT_SYMBOL(ptlrpc_unregister_service);
3241
3242 /**
3243  * Returns 0 if the service is healthy.
3244  *
3245  * Right now, it just checks to make sure that requests aren't languishing
3246  * in the queue.  We'll use this health check to govern whether a node needs
3247  * to be shot, so it's intentionally non-aggressive. */
3248 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3249 {
3250         struct ptlrpc_request *request = NULL;
3251         struct timespec64 right_now;
3252         struct timespec64 timediff;
3253
3254         ktime_get_real_ts64(&right_now);
3255
3256         spin_lock(&svcpt->scp_req_lock);
3257         /* How long has the next entry been waiting? */
3258         if (ptlrpc_server_high_pending(svcpt, true))
3259                 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3260         else if (ptlrpc_server_normal_pending(svcpt, true))
3261                 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3262
3263         if (request == NULL) {
3264                 spin_unlock(&svcpt->scp_req_lock);
3265                 return 0;
3266         }
3267
3268         timediff = timespec64_sub(right_now, request->rq_arrival_time);
3269         spin_unlock(&svcpt->scp_req_lock);
3270
3271         if ((timediff.tv_sec) >
3272             (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3273                 CERROR("%s: unhealthy - request has been waiting %llds\n",
3274                        svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3275                 return -1;
3276         }
3277
3278         return 0;
3279 }
3280
3281 int
3282 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3283 {
3284         struct ptlrpc_service_part      *svcpt;
3285         int                             i;
3286
3287         if (svc == NULL)
3288                 return 0;
3289
3290         ptlrpc_service_for_each_part(svcpt, i, svc) {
3291                 int rc = ptlrpc_svcpt_health_check(svcpt);
3292
3293                 if (rc != 0)
3294                         return rc;
3295         }
3296         return 0;
3297 }
3298 EXPORT_SYMBOL(ptlrpc_service_health_check);