4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
35 #include <linux/kthread.h>
36 #include <obd_support.h>
37 #include <obd_class.h>
38 #include <lustre_net.h>
39 #include <lu_object.h>
40 #include <uapi/linux/lnet/lnet-types.h>
41 #include "ptlrpc_internal.h"
43 /* The following are visible and mutable through /sys/module/ptlrpc */
44 int test_req_buffer_pressure = 0;
45 module_param(test_req_buffer_pressure, int, 0444);
46 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
47 module_param(at_min, int, 0644);
48 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
49 module_param(at_max, int, 0644);
50 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
51 module_param(at_history, int, 0644);
52 MODULE_PARM_DESC(at_history,
53 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
54 module_param(at_early_margin, int, 0644);
55 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
56 module_param(at_extra, int, 0644);
57 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
60 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
61 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
62 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
64 /** Holds a list of all PTLRPC services */
65 struct list_head ptlrpc_all_services;
66 /** Used to protect the \e ptlrpc_all_services list */
67 struct mutex ptlrpc_all_services_mutex;
69 static struct ptlrpc_request_buffer_desc *
70 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
72 struct ptlrpc_service *svc = svcpt->scp_service;
73 struct ptlrpc_request_buffer_desc *rqbd;
75 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
79 rqbd->rqbd_svcpt = svcpt;
80 rqbd->rqbd_refcount = 0;
81 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
82 rqbd->rqbd_cbid.cbid_arg = rqbd;
83 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
84 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
85 svcpt->scp_cpt, svc->srv_buf_size);
86 if (rqbd->rqbd_buffer == NULL) {
91 spin_lock(&svcpt->scp_lock);
92 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
93 svcpt->scp_nrqbds_total++;
94 spin_unlock(&svcpt->scp_lock);
100 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
102 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
104 LASSERT(rqbd->rqbd_refcount == 0);
105 LASSERT(list_empty(&rqbd->rqbd_reqs));
107 spin_lock(&svcpt->scp_lock);
108 list_del(&rqbd->rqbd_list);
109 svcpt->scp_nrqbds_total--;
110 spin_unlock(&svcpt->scp_lock);
112 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
117 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
119 struct ptlrpc_service *svc = svcpt->scp_service;
120 struct ptlrpc_request_buffer_desc *rqbd;
124 if (svcpt->scp_rqbd_allocating)
127 spin_lock(&svcpt->scp_lock);
128 /* check again with lock */
129 if (svcpt->scp_rqbd_allocating) {
130 /* NB: we might allow more than one thread in the future */
131 LASSERT(svcpt->scp_rqbd_allocating == 1);
132 spin_unlock(&svcpt->scp_lock);
136 svcpt->scp_rqbd_allocating++;
137 spin_unlock(&svcpt->scp_lock);
140 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
141 /* NB: another thread might have recycled enough rqbds, we
142 * need to make sure it wouldn't over-allocate, see LU-1212. */
143 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
144 (svc->srv_nrqbds_max != 0 &&
145 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
148 rqbd = ptlrpc_alloc_rqbd(svcpt);
151 CERROR("%s: Can't allocate request buffer\n",
158 spin_lock(&svcpt->scp_lock);
160 LASSERT(svcpt->scp_rqbd_allocating == 1);
161 svcpt->scp_rqbd_allocating--;
163 spin_unlock(&svcpt->scp_lock);
166 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
167 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
168 svcpt->scp_nrqbds_total, rc);
172 rc = ptlrpc_server_post_idle_rqbds(svcpt);
178 * Part of Rep-Ack logic.
179 * Puts a lock and its mode into reply state assotiated to request reply.
182 ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
183 int mode, bool no_ack, bool convert_lock)
185 struct ptlrpc_reply_state *rs = req->rq_reply_state;
189 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
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;
198 EXPORT_SYMBOL(ptlrpc_save_lock);
201 struct ptlrpc_hr_partition;
203 struct ptlrpc_hr_thread {
204 int hrt_id; /* thread ID */
206 wait_queue_head_t hrt_waitq;
207 struct list_head hrt_queue;
208 struct ptlrpc_hr_partition *hrt_partition;
211 struct ptlrpc_hr_partition {
212 /* # of started threads */
213 atomic_t hrp_nstarted;
214 /* # of stopped threads */
215 atomic_t hrp_nstopped;
216 /* cpu partition id */
218 /* round-robin rotor for choosing thread */
220 /* total number of threads on this partition */
223 struct ptlrpc_hr_thread *hrp_thrs;
226 #define HRT_RUNNING 0
227 #define HRT_STOPPING 1
229 struct ptlrpc_hr_service {
230 /* CPU partition table, it's just cfs_cpt_table for now */
231 struct cfs_cpt_table *hr_cpt_table;
232 /** controller sleep waitq */
233 wait_queue_head_t hr_waitq;
234 unsigned int hr_stopping;
235 /** roundrobin rotor for non-affinity service */
236 unsigned int hr_rotor;
238 struct ptlrpc_hr_partition **hr_partitions;
242 struct list_head rsb_replies;
243 unsigned int rsb_n_replies;
244 struct ptlrpc_service_part *rsb_svcpt;
247 /** reply handling service. */
248 static struct ptlrpc_hr_service ptlrpc_hr;
251 * maximum mumber of replies scheduled in one batch
253 #define MAX_SCHEDULED 256
256 * Initialize a reply batch.
260 static void rs_batch_init(struct rs_batch *b)
262 memset(b, 0, sizeof *b);
263 INIT_LIST_HEAD(&b->rsb_replies);
267 * Choose an hr thread to dispatch requests to.
269 static struct ptlrpc_hr_thread *
270 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
272 struct ptlrpc_hr_partition *hrp;
275 if (svcpt->scp_cpt >= 0 &&
276 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
277 /* directly match partition */
278 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
281 rotor = ptlrpc_hr.hr_rotor++;
282 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
284 hrp = ptlrpc_hr.hr_partitions[rotor];
287 rotor = hrp->hrp_rotor++;
288 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
292 * Dispatch all replies accumulated in the batch to one from
293 * dedicated reply handling threads.
297 static void rs_batch_dispatch(struct rs_batch *b)
299 if (b->rsb_n_replies != 0) {
300 struct ptlrpc_hr_thread *hrt;
302 hrt = ptlrpc_hr_select(b->rsb_svcpt);
304 spin_lock(&hrt->hrt_lock);
305 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
306 spin_unlock(&hrt->hrt_lock);
308 wake_up(&hrt->hrt_waitq);
309 b->rsb_n_replies = 0;
314 * Add a reply to a batch.
315 * Add one reply object to a batch, schedule batched replies if overload.
320 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
322 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
324 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
325 if (b->rsb_svcpt != NULL) {
326 rs_batch_dispatch(b);
327 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
329 spin_lock(&svcpt->scp_rep_lock);
330 b->rsb_svcpt = svcpt;
332 spin_lock(&rs->rs_lock);
333 rs->rs_scheduled_ever = 1;
334 if (rs->rs_scheduled == 0) {
335 list_move(&rs->rs_list, &b->rsb_replies);
336 rs->rs_scheduled = 1;
339 rs->rs_committed = 1;
340 spin_unlock(&rs->rs_lock);
344 * Reply batch finalization.
345 * Dispatch remaining replies from the batch
346 * and release remaining spinlock.
350 static void rs_batch_fini(struct rs_batch *b)
352 if (b->rsb_svcpt != NULL) {
353 rs_batch_dispatch(b);
354 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
358 #define DECLARE_RS_BATCH(b) struct rs_batch b
362 * Put reply state into a queue for processing because we received
363 * ACK from the client
365 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
367 struct ptlrpc_hr_thread *hrt;
370 LASSERT(list_empty(&rs->rs_list));
372 hrt = ptlrpc_hr_select(rs->rs_svcpt);
374 spin_lock(&hrt->hrt_lock);
375 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
376 spin_unlock(&hrt->hrt_lock);
378 wake_up(&hrt->hrt_waitq);
383 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
387 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
388 assert_spin_locked(&rs->rs_lock);
389 LASSERT (rs->rs_difficult);
390 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
392 if (rs->rs_scheduled) { /* being set up or already notified */
397 rs->rs_scheduled = 1;
398 list_del_init(&rs->rs_list);
399 ptlrpc_dispatch_difficult_reply(rs);
402 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
404 void ptlrpc_commit_replies(struct obd_export *exp)
406 struct ptlrpc_reply_state *rs, *nxt;
407 DECLARE_RS_BATCH(batch);
410 rs_batch_init(&batch);
411 /* Find any replies that have been committed and get their service
412 * to attend to complete them. */
414 /* CAVEAT EMPTOR: spinlock ordering!!! */
415 spin_lock(&exp->exp_uncommitted_replies_lock);
416 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
418 LASSERT (rs->rs_difficult);
419 /* VBR: per-export last_committed */
420 LASSERT(rs->rs_export);
421 if (rs->rs_transno <= exp->exp_last_committed) {
422 list_del_init(&rs->rs_obd_list);
423 rs_batch_add(&batch, rs);
426 spin_unlock(&exp->exp_uncommitted_replies_lock);
427 rs_batch_fini(&batch);
432 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
434 struct ptlrpc_request_buffer_desc *rqbd;
439 spin_lock(&svcpt->scp_lock);
441 if (list_empty(&svcpt->scp_rqbd_idle)) {
442 spin_unlock(&svcpt->scp_lock);
446 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
447 struct ptlrpc_request_buffer_desc,
449 list_del(&rqbd->rqbd_list);
451 /* assume we will post successfully */
452 svcpt->scp_nrqbds_posted++;
453 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
455 spin_unlock(&svcpt->scp_lock);
457 rc = ptlrpc_register_rqbd(rqbd);
464 spin_lock(&svcpt->scp_lock);
466 svcpt->scp_nrqbds_posted--;
467 list_del(&rqbd->rqbd_list);
468 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
470 /* Don't complain if no request buffers are posted right now; LNET
471 * won't drop requests because we set the portal lazy! */
473 spin_unlock(&svcpt->scp_lock);
478 static void ptlrpc_at_timer(unsigned long castmeharder)
480 struct ptlrpc_service_part *svcpt;
482 svcpt = (struct ptlrpc_service_part *)castmeharder;
484 svcpt->scp_at_check = 1;
485 svcpt->scp_at_checktime = ktime_get();
486 wake_up(&svcpt->scp_waitq);
490 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
491 struct ptlrpc_service_conf *conf)
493 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
500 * Common code for estimating & validating threads number.
501 * CPT affinity service could have percpt thread-pool instead
502 * of a global thread-pool, which means user might not always
503 * get the threads number they give it in conf::tc_nthrs_user
504 * even they did set. It's because we need to validate threads
505 * number for each CPT to guarantee each pool will have enough
506 * threads to keep the service healthy.
508 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
509 init = max_t(int, init, tc->tc_nthrs_init);
511 /* NB: please see comments in lustre_lnet.h for definition
512 * details of these members */
513 LASSERT(tc->tc_nthrs_max != 0);
515 if (tc->tc_nthrs_user != 0) {
516 /* In case there is a reason to test a service with many
517 * threads, we give a less strict check here, it can
518 * be up to 8 * nthrs_max */
519 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
520 nthrs = total / svc->srv_ncpts;
521 init = max(init, nthrs);
525 total = tc->tc_nthrs_max;
526 if (tc->tc_nthrs_base == 0) {
527 /* don't care about base threads number per partition,
528 * this is most for non-affinity service */
529 nthrs = total / svc->srv_ncpts;
533 nthrs = tc->tc_nthrs_base;
534 if (svc->srv_ncpts == 1) {
537 /* NB: Increase the base number if it's single partition
538 * and total number of cores/HTs is larger or equal to 4.
539 * result will always < 2 * nthrs_base */
540 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
541 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
542 (tc->tc_nthrs_base >> i) != 0; i++)
543 nthrs += tc->tc_nthrs_base >> i;
546 if (tc->tc_thr_factor != 0) {
547 int factor = tc->tc_thr_factor;
551 * User wants to increase number of threads with for
552 * each CPU core/HT, most likely the factor is larger than
553 * one thread/core because service threads are supposed to
554 * be blocked by lock or wait for IO.
557 * Amdahl's law says that adding processors wouldn't give
558 * a linear increasing of parallelism, so it's nonsense to
559 * have too many threads no matter how many cores/HTs
562 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
563 /* weight is # of HTs */
564 /* depress thread factor for hyper-thread */
565 factor = factor - (factor >> 1) + (factor >> 3);
568 weight = cfs_cpt_weight(svc->srv_cptable, 0);
570 for (; factor > 0 && weight > 0; factor--, weight -= fade)
571 nthrs += min(weight, fade) * factor;
574 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
575 nthrs = max(tc->tc_nthrs_base,
576 tc->tc_nthrs_max / svc->srv_ncpts);
579 nthrs = max(nthrs, tc->tc_nthrs_init);
580 svc->srv_nthrs_cpt_limit = nthrs;
581 svc->srv_nthrs_cpt_init = init;
583 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
584 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
585 "than the given soft limit (%d)\n",
586 svc->srv_name, nthrs * svc->srv_ncpts,
592 * Initialize percpt data for a service
595 ptlrpc_service_part_init(struct ptlrpc_service *svc,
596 struct ptlrpc_service_part *svcpt, int cpt)
598 struct ptlrpc_at_array *array;
603 svcpt->scp_cpt = cpt;
604 INIT_LIST_HEAD(&svcpt->scp_threads);
606 /* rqbd and incoming request queue */
607 spin_lock_init(&svcpt->scp_lock);
608 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
609 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
610 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
611 init_waitqueue_head(&svcpt->scp_waitq);
612 /* history request & rqbd list */
613 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
614 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
616 /* acitve requests and hp requests */
617 spin_lock_init(&svcpt->scp_req_lock);
620 spin_lock_init(&svcpt->scp_rep_lock);
621 INIT_LIST_HEAD(&svcpt->scp_rep_active);
622 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
623 init_waitqueue_head(&svcpt->scp_rep_waitq);
624 atomic_set(&svcpt->scp_nreps_difficult, 0);
626 /* adaptive timeout */
627 spin_lock_init(&svcpt->scp_at_lock);
628 array = &svcpt->scp_at_array;
630 size = at_est2timeout(at_max);
631 array->paa_size = size;
632 array->paa_count = 0;
633 array->paa_deadline = -1;
635 /* allocate memory for scp_at_array (ptlrpc_at_array) */
636 OBD_CPT_ALLOC(array->paa_reqs_array,
637 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
638 if (array->paa_reqs_array == NULL)
641 for (index = 0; index < size; index++)
642 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
644 OBD_CPT_ALLOC(array->paa_reqs_count,
645 svc->srv_cptable, cpt, sizeof(__u32) * size);
646 if (array->paa_reqs_count == NULL)
649 setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
650 (unsigned long)svcpt);
652 /* At SOW, service time should be quick; 10s seems generous. If client
653 * timeout is less than this, we'll be sending an early reply. */
654 at_init(&svcpt->scp_at_estimate, 10, 0);
656 /* assign this before call ptlrpc_grow_req_bufs */
657 svcpt->scp_service = svc;
658 /* Now allocate the request buffers, but don't post them now */
659 rc = ptlrpc_grow_req_bufs(svcpt, 0);
660 /* We shouldn't be under memory pressure at startup, so
661 * fail if we can't allocate all our buffers at this time. */
668 if (array->paa_reqs_count != NULL) {
669 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
670 array->paa_reqs_count = NULL;
673 if (array->paa_reqs_array != NULL) {
674 OBD_FREE(array->paa_reqs_array,
675 sizeof(struct list_head) * array->paa_size);
676 array->paa_reqs_array = NULL;
683 * Initialize service on a given portal.
684 * This includes starting serving threads , allocating and posting rqbds and
687 struct ptlrpc_service *
688 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
690 struct dentry *debugfs_entry)
692 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
693 struct ptlrpc_service *service;
694 struct ptlrpc_service_part *svcpt;
695 struct cfs_cpt_table *cptable;
703 LASSERT(conf->psc_buf.bc_nbufs > 0);
704 LASSERT(conf->psc_buf.bc_buf_size >=
705 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
706 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
708 cptable = cconf->cc_cptable;
710 cptable = cfs_cpt_table;
712 if (!conf->psc_thr.tc_cpu_affinity) {
715 ncpts = cfs_cpt_number(cptable);
716 if (cconf->cc_pattern != NULL) {
717 struct cfs_expr_list *el;
719 rc = cfs_expr_list_parse(cconf->cc_pattern,
720 strlen(cconf->cc_pattern),
723 CERROR("%s: invalid CPT pattern string: %s",
724 conf->psc_name, cconf->cc_pattern);
725 RETURN(ERR_PTR(-EINVAL));
728 rc = cfs_expr_list_values(el, ncpts, &cpts);
729 cfs_expr_list_free(el);
731 CERROR("%s: failed to parse CPT array %s: %d\n",
732 conf->psc_name, cconf->cc_pattern, rc);
734 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
735 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
741 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
742 if (service == NULL) {
744 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
745 RETURN(ERR_PTR(-ENOMEM));
748 service->srv_cptable = cptable;
749 service->srv_cpts = cpts;
750 service->srv_ncpts = ncpts;
752 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
753 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
754 service->srv_cpt_bits++;
757 spin_lock_init(&service->srv_lock);
758 service->srv_name = conf->psc_name;
759 service->srv_watchdog_factor = conf->psc_watchdog_factor;
760 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
762 /* buffer configuration */
763 service->srv_nbuf_per_group = test_req_buffer_pressure ?
764 1 : conf->psc_buf.bc_nbufs;
765 /* do not limit max number of rqbds by default */
766 service->srv_nrqbds_max = 0;
768 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
770 service->srv_buf_size = conf->psc_buf.bc_buf_size;
771 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
772 service->srv_req_portal = conf->psc_buf.bc_req_portal;
774 /* Increase max reply size to next power of two */
775 service->srv_max_reply_size = 1;
776 while (service->srv_max_reply_size <
777 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
778 service->srv_max_reply_size <<= 1;
780 service->srv_thread_name = conf->psc_thr.tc_thr_name;
781 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
782 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
783 service->srv_ops = conf->psc_ops;
785 for (i = 0; i < ncpts; i++) {
786 if (!conf->psc_thr.tc_cpu_affinity)
789 cpt = cpts != NULL ? cpts[i] : i;
791 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
793 GOTO(failed, rc = -ENOMEM);
795 service->srv_parts[i] = svcpt;
796 rc = ptlrpc_service_part_init(service, svcpt, cpt);
801 ptlrpc_server_nthreads_check(service, conf);
803 rc = LNetSetLazyPortal(service->srv_req_portal);
806 mutex_lock(&ptlrpc_all_services_mutex);
807 list_add(&service->srv_list, &ptlrpc_all_services);
808 mutex_unlock(&ptlrpc_all_services_mutex);
811 rc = ptlrpc_sysfs_register_service(parent, service);
816 if (debugfs_entry != NULL)
817 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
819 rc = ptlrpc_service_nrs_setup(service);
823 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
824 service->srv_name, service->srv_req_portal);
826 rc = ptlrpc_start_threads(service);
828 CERROR("Failed to start threads for service %s: %d\n",
829 service->srv_name, rc);
835 ptlrpc_unregister_service(service);
838 EXPORT_SYMBOL(ptlrpc_register_service);
841 * to actually free the request, must be called without holding svc_lock.
842 * note it's caller's responsibility to unlink req->rq_list.
844 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
846 LASSERT(atomic_read(&req->rq_refcount) == 0);
847 LASSERT(list_empty(&req->rq_timed_list));
849 /* DEBUG_REQ() assumes the reply state of a request with a valid
850 * ref will not be destroyed until that reference is dropped. */
851 ptlrpc_req_drop_rs(req);
853 sptlrpc_svc_ctx_decref(req);
855 if (req != &req->rq_rqbd->rqbd_req) {
856 /* NB request buffers use an embedded
857 * req if the incoming req unlinked the
858 * MD; this isn't one of them! */
859 ptlrpc_request_cache_free(req);
864 * drop a reference count of the request. if it reaches 0, we either
865 * put it into history list, or free it immediately.
867 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
869 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
870 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
871 struct ptlrpc_service *svc = svcpt->scp_service;
873 struct list_head *tmp;
874 struct list_head *nxt;
876 if (!atomic_dec_and_test(&req->rq_refcount))
879 if (req->rq_session.lc_state == LCS_ENTERED) {
880 lu_context_exit(&req->rq_session);
881 lu_context_fini(&req->rq_session);
884 if (req->rq_at_linked) {
885 spin_lock(&svcpt->scp_at_lock);
886 /* recheck with lock, in case it's unlinked by
887 * ptlrpc_at_check_timed() */
888 if (likely(req->rq_at_linked))
889 ptlrpc_at_remove_timed(req);
890 spin_unlock(&svcpt->scp_at_lock);
893 LASSERT(list_empty(&req->rq_timed_list));
895 /* finalize request */
896 if (req->rq_export) {
897 class_export_put(req->rq_export);
898 req->rq_export = NULL;
901 spin_lock(&svcpt->scp_lock);
903 list_add(&req->rq_list, &rqbd->rqbd_reqs);
905 refcount = --(rqbd->rqbd_refcount);
907 /* request buffer is now idle: add to history */
908 list_del(&rqbd->rqbd_list);
910 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
911 svcpt->scp_hist_nrqbds++;
913 /* cull some history?
914 * I expect only about 1 or 2 rqbds need to be recycled here */
915 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
916 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
917 struct ptlrpc_request_buffer_desc,
920 list_del(&rqbd->rqbd_list);
921 svcpt->scp_hist_nrqbds--;
923 /* remove rqbd's reqs from svc's req history while
924 * I've got the service lock */
925 list_for_each(tmp, &rqbd->rqbd_reqs) {
926 req = list_entry(tmp, struct ptlrpc_request,
928 /* Track the highest culled req seq */
929 if (req->rq_history_seq >
930 svcpt->scp_hist_seq_culled) {
931 svcpt->scp_hist_seq_culled =
934 list_del(&req->rq_history_list);
937 spin_unlock(&svcpt->scp_lock);
939 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
940 req = list_entry(rqbd->rqbd_reqs.next,
941 struct ptlrpc_request,
943 list_del(&req->rq_list);
944 ptlrpc_server_free_request(req);
947 spin_lock(&svcpt->scp_lock);
949 * now all reqs including the embedded req has been
950 * disposed, schedule request buffer for re-use
951 * or free it to drain some in excess.
953 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
954 if (svcpt->scp_nrqbds_posted >=
955 svc->srv_nbuf_per_group ||
956 (svc->srv_nrqbds_max != 0 &&
957 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
958 test_req_buffer_pressure) {
959 /* like in ptlrpc_free_rqbd() */
960 svcpt->scp_nrqbds_total--;
961 OBD_FREE_LARGE(rqbd->rqbd_buffer,
965 list_add_tail(&rqbd->rqbd_list,
966 &svcpt->scp_rqbd_idle);
970 spin_unlock(&svcpt->scp_lock);
971 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
972 /* If we are low on memory, we are not interested in history */
973 list_del(&req->rq_list);
974 list_del_init(&req->rq_history_list);
976 /* Track the highest culled req seq */
977 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
978 svcpt->scp_hist_seq_culled = req->rq_history_seq;
980 spin_unlock(&svcpt->scp_lock);
982 ptlrpc_server_free_request(req);
984 spin_unlock(&svcpt->scp_lock);
988 /** Change request export and move hp request from old export to new */
989 void ptlrpc_request_change_export(struct ptlrpc_request *req,
990 struct obd_export *export)
992 if (req->rq_export != NULL) {
993 LASSERT(!list_empty(&req->rq_exp_list));
994 /* remove rq_exp_list from last export */
995 spin_lock(&req->rq_export->exp_rpc_lock);
996 list_del_init(&req->rq_exp_list);
997 spin_unlock(&req->rq_export->exp_rpc_lock);
998 /* export has one reference already, so it`s safe to
999 * add req to export queue here and get another
1000 * reference for request later */
1001 spin_lock(&export->exp_rpc_lock);
1002 if (req->rq_ops != NULL) /* hp request */
1003 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1005 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1006 spin_unlock(&export->exp_rpc_lock);
1008 class_export_rpc_dec(req->rq_export);
1009 class_export_put(req->rq_export);
1012 /* request takes one export refcount */
1013 req->rq_export = class_export_get(export);
1014 class_export_rpc_inc(export);
1020 * to finish a request: stop sending more early replies, and release
1023 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1024 struct ptlrpc_request *req)
1026 ptlrpc_server_hpreq_fini(req);
1028 ptlrpc_server_drop_request(req);
1032 * to finish an active request: stop sending more early replies, and release
1033 * the request. should be called after we finished handling the request.
1035 static void ptlrpc_server_finish_active_request(
1036 struct ptlrpc_service_part *svcpt,
1037 struct ptlrpc_request *req)
1039 spin_lock(&svcpt->scp_req_lock);
1040 ptlrpc_nrs_req_stop_nolock(req);
1041 svcpt->scp_nreqs_active--;
1043 svcpt->scp_nhreqs_active--;
1044 spin_unlock(&svcpt->scp_req_lock);
1046 ptlrpc_nrs_req_finalize(req);
1048 if (req->rq_export != NULL)
1049 class_export_rpc_dec(req->rq_export);
1051 ptlrpc_server_finish_request(svcpt, req);
1055 * This function makes sure dead exports are evicted in a timely manner.
1056 * This function is only called when some export receives a message (i.e.,
1057 * the network is up.)
1059 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1061 struct obd_export *oldest_exp;
1062 time64_t oldest_time, new_time;
1068 /* Compensate for slow machines, etc, by faking our request time
1069 into the future. Although this can break the strict time-ordering
1070 of the list, we can be really lazy here - we don't have to evict
1071 at the exact right moment. Eventually, all silent exports
1072 will make it to the top of the list. */
1074 /* Do not pay attention on 1sec or smaller renewals. */
1075 new_time = ktime_get_real_seconds() + extra_delay;
1076 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1079 exp->exp_last_request_time = new_time;
1081 /* exports may get disconnected from the chain even though the
1082 export has references, so we must keep the spin lock while
1083 manipulating the lists */
1084 spin_lock(&exp->exp_obd->obd_dev_lock);
1086 if (list_empty(&exp->exp_obd_chain_timed)) {
1087 /* this one is not timed */
1088 spin_unlock(&exp->exp_obd->obd_dev_lock);
1092 list_move_tail(&exp->exp_obd_chain_timed,
1093 &exp->exp_obd->obd_exports_timed);
1095 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1096 struct obd_export, exp_obd_chain_timed);
1097 oldest_time = oldest_exp->exp_last_request_time;
1098 spin_unlock(&exp->exp_obd->obd_dev_lock);
1100 if (exp->exp_obd->obd_recovering) {
1101 /* be nice to everyone during recovery */
1106 /* Note - racing to start/reset the obd_eviction timer is safe */
1107 if (exp->exp_obd->obd_eviction_timer == 0) {
1108 /* Check if the oldest entry is expired. */
1109 if (ktime_get_real_seconds() >
1110 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1111 /* We need a second timer, in case the net was down and
1112 * it just came back. Since the pinger may skip every
1113 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1114 * we better wait for 3.
1116 exp->exp_obd->obd_eviction_timer =
1117 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1118 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1119 exp->exp_obd->obd_name,
1120 obd_export_nid2str(oldest_exp), oldest_time);
1123 if (ktime_get_real_seconds() >
1124 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1125 /* The evictor won't evict anyone who we've heard from
1126 * recently, so we don't have to check before we start
1129 if (!ping_evictor_wake(exp))
1130 exp->exp_obd->obd_eviction_timer = 0;
1138 * Sanity check request \a req.
1139 * Return 0 if all is ok, error code otherwise.
1141 static int ptlrpc_check_req(struct ptlrpc_request *req)
1143 struct obd_device *obd = req->rq_export->exp_obd;
1146 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1147 req->rq_export->exp_conn_cnt)) {
1148 DEBUG_REQ(D_RPCTRACE, req,
1149 "DROPPING req from old connection %d < %d",
1150 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1151 req->rq_export->exp_conn_cnt);
1154 if (unlikely(obd == NULL || obd->obd_fail)) {
1155 /* Failing over, don't handle any more reqs,
1156 * send error response instead. */
1157 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1158 req, (obd != NULL) ? obd->obd_name : "unknown");
1160 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1161 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1162 !obd->obd_recovering) {
1163 DEBUG_REQ(D_ERROR, req,
1164 "Invalid replay without recovery");
1165 class_fail_export(req->rq_export);
1167 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1168 !obd->obd_recovering) {
1169 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1170 "%llu without recovery",
1171 lustre_msg_get_transno(req->rq_reqmsg));
1172 class_fail_export(req->rq_export);
1176 if (unlikely(rc < 0)) {
1177 req->rq_status = rc;
1183 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1185 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1188 if (array->paa_count == 0) {
1189 del_timer(&svcpt->scp_at_timer);
1193 /* Set timer for closest deadline */
1194 next = array->paa_deadline - ktime_get_real_seconds() -
1197 ptlrpc_at_timer((unsigned long)svcpt);
1199 mod_timer(&svcpt->scp_at_timer,
1200 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1201 CDEBUG(D_INFO, "armed %s at %+llds\n",
1202 svcpt->scp_service->srv_name, next);
1206 /* Add rpc to early reply check list */
1207 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1209 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1210 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1211 struct ptlrpc_request *rq = NULL;
1217 if (req->rq_no_reply)
1220 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1223 spin_lock(&svcpt->scp_at_lock);
1224 LASSERT(list_empty(&req->rq_timed_list));
1226 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1227 if (array->paa_reqs_count[index] > 0) {
1228 /* latest rpcs will have the latest deadlines in the list,
1229 * so search backward. */
1230 list_for_each_entry_reverse(rq,
1231 &array->paa_reqs_array[index],
1233 if (req->rq_deadline >= rq->rq_deadline) {
1234 list_add(&req->rq_timed_list,
1235 &rq->rq_timed_list);
1241 /* Add the request at the head of the list */
1242 if (list_empty(&req->rq_timed_list))
1243 list_add(&req->rq_timed_list,
1244 &array->paa_reqs_array[index]);
1246 spin_lock(&req->rq_lock);
1247 req->rq_at_linked = 1;
1248 spin_unlock(&req->rq_lock);
1249 req->rq_at_index = index;
1250 array->paa_reqs_count[index]++;
1252 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1253 array->paa_deadline = req->rq_deadline;
1254 ptlrpc_at_set_timer(svcpt);
1256 spin_unlock(&svcpt->scp_at_lock);
1262 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1264 struct ptlrpc_at_array *array;
1266 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1268 /* NB: must call with hold svcpt::scp_at_lock */
1269 LASSERT(!list_empty(&req->rq_timed_list));
1270 list_del_init(&req->rq_timed_list);
1272 spin_lock(&req->rq_lock);
1273 req->rq_at_linked = 0;
1274 spin_unlock(&req->rq_lock);
1276 array->paa_reqs_count[req->rq_at_index]--;
1281 * Attempt to extend the request deadline by sending an early reply to the
1284 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1286 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1287 struct ptlrpc_request *reqcopy;
1288 struct lustre_msg *reqmsg;
1289 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1295 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1296 /* don't send early reply */
1300 /* deadline is when the client expects us to reply, margin is the
1301 difference between clients' and servers' expectations */
1302 DEBUG_REQ(D_ADAPTTO, req,
1303 "%ssending early reply (deadline %+llds, margin %+llds) for "
1304 "%d+%d", AT_OFF ? "AT off - not " : "",
1305 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1306 at_get(&svcpt->scp_at_estimate), at_extra);
1312 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1313 "not sending early reply. Consider increasing "
1314 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1316 /* Return an error so we're not re-added to the timed list. */
1320 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1321 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1322 "but no AT support");
1326 if (req->rq_export &&
1327 lustre_msg_get_flags(req->rq_reqmsg) &
1328 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1329 struct obd_device *obd_exp = req->rq_export->exp_obd;
1331 /* During recovery, we don't want to send too many early
1332 * replies, but on the other hand we want to make sure the
1333 * client has enough time to resend if the rpc is lost. So
1334 * during the recovery period send at least 4 early replies,
1335 * spacing them every at_extra if we can. at_estimate should
1336 * always equal this fixed value during recovery.
1338 /* Don't account request processing time into AT history
1339 * during recovery, it is not service time we need but
1340 * includes also waiting time for recovering clients
1342 newdl = min_t(time64_t, at_extra,
1343 obd_exp->obd_recovery_timeout / 4) +
1344 ktime_get_real_seconds();
1346 /* We want to extend the request deadline by at_extra seconds,
1347 * so we set our service estimate to reflect how much time has
1348 * passed since this request arrived plus an additional
1349 * at_extra seconds. The client will calculate the new deadline
1350 * based on this service estimate (plus some additional time to
1351 * account for network latency). See ptlrpc_at_recv_early_reply
1353 at_measured(&svcpt->scp_at_estimate, at_extra +
1354 ktime_get_real_seconds() -
1355 req->rq_arrival_time.tv_sec);
1356 newdl = req->rq_arrival_time.tv_sec +
1357 at_get(&svcpt->scp_at_estimate);
1360 /* Check to see if we've actually increased the deadline -
1361 * we may be past adaptive_max */
1362 if (req->rq_deadline >= newdl) {
1363 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1364 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1368 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1369 if (reqcopy == NULL)
1371 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1373 GOTO(out_free, rc = -ENOMEM);
1376 reqcopy->rq_reply_state = NULL;
1377 reqcopy->rq_rep_swab_mask = 0;
1378 reqcopy->rq_pack_bulk = 0;
1379 reqcopy->rq_pack_udesc = 0;
1380 reqcopy->rq_packed_final = 0;
1381 sptlrpc_svc_ctx_addref(reqcopy);
1382 /* We only need the reqmsg for the magic */
1383 reqcopy->rq_reqmsg = reqmsg;
1384 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1387 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1388 * Without this check, we would fail the rq_no_reply assertion in
1389 * ptlrpc_send_reply().
1391 if (reqcopy->rq_no_reply)
1392 GOTO(out, rc = -ETIMEDOUT);
1394 LASSERT(atomic_read(&req->rq_refcount));
1395 /** if it is last refcount then early reply isn't needed */
1396 if (atomic_read(&req->rq_refcount) == 1) {
1397 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1398 "abort sending early reply\n");
1399 GOTO(out, rc = -EINVAL);
1402 /* Connection ref */
1403 reqcopy->rq_export = class_conn2export(
1404 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1405 if (reqcopy->rq_export == NULL)
1406 GOTO(out, rc = -ENODEV);
1409 class_export_rpc_inc(reqcopy->rq_export);
1410 if (reqcopy->rq_export->exp_obd &&
1411 reqcopy->rq_export->exp_obd->obd_fail)
1412 GOTO(out_put, rc = -ENODEV);
1414 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1418 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1421 /* Adjust our own deadline to what we told the client */
1422 req->rq_deadline = newdl;
1423 req->rq_early_count++; /* number sent, server side */
1425 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1428 /* Free the (early) reply state from lustre_pack_reply.
1429 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1430 ptlrpc_req_drop_rs(reqcopy);
1433 class_export_rpc_dec(reqcopy->rq_export);
1434 class_export_put(reqcopy->rq_export);
1436 sptlrpc_svc_ctx_decref(reqcopy);
1437 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1439 ptlrpc_request_cache_free(reqcopy);
1443 /* Send early replies to everybody expiring within at_early_margin
1444 asking for at_extra time */
1445 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1447 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1448 struct ptlrpc_request *rq, *n;
1449 struct list_head work_list;
1452 time64_t now = ktime_get_real_seconds();
1454 int first, counter = 0;
1457 spin_lock(&svcpt->scp_at_lock);
1458 if (svcpt->scp_at_check == 0) {
1459 spin_unlock(&svcpt->scp_at_lock);
1462 delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1463 svcpt->scp_at_check = 0;
1465 if (array->paa_count == 0) {
1466 spin_unlock(&svcpt->scp_at_lock);
1470 /* The timer went off, but maybe the nearest rpc already completed. */
1471 first = array->paa_deadline - now;
1472 if (first > at_early_margin) {
1473 /* We've still got plenty of time. Reset the timer. */
1474 ptlrpc_at_set_timer(svcpt);
1475 spin_unlock(&svcpt->scp_at_lock);
1479 /* We're close to a timeout, and we don't know how much longer the
1480 server will take. Send early replies to everyone expiring soon. */
1481 INIT_LIST_HEAD(&work_list);
1483 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1484 count = array->paa_count;
1486 count -= array->paa_reqs_count[index];
1487 list_for_each_entry_safe(rq, n,
1488 &array->paa_reqs_array[index],
1490 if (rq->rq_deadline > now + at_early_margin) {
1491 /* update the earliest deadline */
1492 if (deadline == -1 ||
1493 rq->rq_deadline < deadline)
1494 deadline = rq->rq_deadline;
1498 ptlrpc_at_remove_timed(rq);
1500 * ptlrpc_server_drop_request() may drop
1501 * refcount to 0 already. Let's check this and
1502 * don't add entry to work_list
1504 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1505 list_add(&rq->rq_timed_list, &work_list);
1509 if (++index >= array->paa_size)
1512 array->paa_deadline = deadline;
1513 /* we have a new earliest deadline, restart the timer */
1514 ptlrpc_at_set_timer(svcpt);
1516 spin_unlock(&svcpt->scp_at_lock);
1518 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1519 "replies\n", first, at_extra, counter);
1521 /* We're already past request deadlines before we even get a
1522 chance to send early replies */
1523 LCONSOLE_WARN("%s: This server is not able to keep up with "
1524 "request traffic (cpu-bound).\n",
1525 svcpt->scp_service->srv_name);
1526 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1527 counter, svcpt->scp_nreqs_incoming,
1528 svcpt->scp_nreqs_active,
1529 at_get(&svcpt->scp_at_estimate), delay);
1532 /* we took additional refcount so entries can't be deleted from list, no
1533 * locking is needed */
1534 while (!list_empty(&work_list)) {
1535 rq = list_entry(work_list.next, struct ptlrpc_request,
1537 list_del_init(&rq->rq_timed_list);
1539 if (ptlrpc_at_send_early_reply(rq) == 0)
1540 ptlrpc_at_add_timed(rq);
1542 ptlrpc_server_drop_request(rq);
1545 RETURN(1); /* return "did_something" for liblustre */
1548 /* Check if we are already handling earlier incarnation of this request.
1549 * Called under &req->rq_export->exp_rpc_lock locked */
1550 static struct ptlrpc_request*
1551 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1553 struct ptlrpc_request *tmp = NULL;
1555 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1556 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1559 /* bulk request are aborted upon reconnect, don't try to
1561 if (req->rq_bulk_write || req->rq_bulk_read)
1564 /* This list should not be longer than max_requests in
1565 * flights on the client, so it is not all that long.
1566 * Also we only hit this codepath in case of a resent
1567 * request which makes it even more rarely hit */
1568 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1570 /* Found duplicate one */
1571 if (tmp->rq_xid == req->rq_xid)
1574 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1576 /* Found duplicate one */
1577 if (tmp->rq_xid == req->rq_xid)
1583 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1584 DEBUG_REQ(D_HA, tmp, "Request being processed");
1589 * Check if a request should be assigned with a high priority.
1591 * \retval < 0: error occurred
1592 * 0: normal RPC request
1593 * +1: high priority request
1595 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1596 struct ptlrpc_request *req)
1601 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1602 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1609 if (req->rq_export != NULL && req->rq_ops != NULL) {
1610 /* Perform request specific check. We should do this
1611 * check before the request is added into exp_hp_rpcs
1612 * list otherwise it may hit swab race at LU-1044. */
1613 if (req->rq_ops->hpreq_check != NULL) {
1614 rc = req->rq_ops->hpreq_check(req);
1615 if (rc == -ESTALE) {
1616 req->rq_status = rc;
1619 /** can only return error,
1620 * 0 for normal request,
1621 * or 1 for high priority request */
1629 /** Remove the request from the export list. */
1630 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1633 if (req->rq_export) {
1634 /* refresh lock timeout again so that client has more
1635 * room to send lock cancel RPC. */
1636 if (req->rq_ops && req->rq_ops->hpreq_fini)
1637 req->rq_ops->hpreq_fini(req);
1639 spin_lock(&req->rq_export->exp_rpc_lock);
1640 list_del_init(&req->rq_exp_list);
1641 spin_unlock(&req->rq_export->exp_rpc_lock);
1646 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1651 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1652 .hpreq_check = ptlrpc_hpreq_check,
1655 /* Hi-Priority RPC check by RPC operation code. */
1656 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1658 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1660 /* Check for export to let only reconnects for not yet evicted
1661 * export to become a HP rpc. */
1662 if ((req->rq_export != NULL) &&
1663 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1664 req->rq_ops = &ptlrpc_hpreq_common;
1668 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1670 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1671 struct ptlrpc_request *req)
1675 struct ptlrpc_request *orig;
1678 rc = ptlrpc_server_hpreq_init(svcpt, req);
1683 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1685 if (req->rq_export != NULL) {
1686 struct obd_export *exp = req->rq_export;
1688 /* do search for duplicated xid and the adding to the list
1690 spin_lock_bh(&exp->exp_rpc_lock);
1691 orig = ptlrpc_server_check_resend_in_progress(req);
1692 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1695 spin_unlock_bh(&exp->exp_rpc_lock);
1698 * When the client resend request and the server has
1699 * the previous copy of it, we need to update deadlines,
1700 * to be sure that the client and the server have equal
1701 * request deadlines.
1704 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1705 linked = orig->rq_at_linked;
1707 ptlrpc_at_remove_timed(orig);
1708 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1709 orig->rq_deadline = req->rq_deadline;
1711 ptlrpc_at_add_timed(orig);
1712 ptlrpc_server_drop_request(orig);
1713 ptlrpc_nrs_req_finalize(req);
1717 if (hp || req->rq_ops != NULL)
1718 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1720 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1721 spin_unlock_bh(&exp->exp_rpc_lock);
1724 /* the current thread is not the processing thread for this request
1725 * since that, but request is in exp_hp_list and can be find there.
1726 * Remove all relations between request and old thread. */
1727 req->rq_svc_thread->t_env->le_ses = NULL;
1728 req->rq_svc_thread = NULL;
1729 req->rq_session.lc_thread = NULL;
1731 ptlrpc_nrs_req_add(svcpt, req, hp);
1737 * Allow to handle high priority request
1738 * User can call it w/o any lock but need to hold
1739 * ptlrpc_service_part::scp_req_lock to get reliable result
1741 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1744 int running = svcpt->scp_nthrs_running;
1746 if (!nrs_svcpt_has_hp(svcpt))
1752 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1755 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1756 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1757 /* leave just 1 thread for normal RPCs */
1758 running = PTLRPC_NTHRS_INIT;
1759 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1763 if (svcpt->scp_nreqs_active >= running - 1)
1766 if (svcpt->scp_nhreqs_active == 0)
1769 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1770 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1773 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1776 return ptlrpc_server_allow_high(svcpt, force) &&
1777 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1781 * Only allow normal priority requests on a service that has a high-priority
1782 * queue if forced (i.e. cleanup), if there are other high priority requests
1783 * already being processed (i.e. those threads can service more high-priority
1784 * requests), or if there are enough idle threads that a later thread can do
1785 * a high priority request.
1786 * User can call it w/o any lock but need to hold
1787 * ptlrpc_service_part::scp_req_lock to get reliable result
1789 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1792 int running = svcpt->scp_nthrs_running;
1793 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1794 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1795 /* leave just 1 thread for normal RPCs */
1796 running = PTLRPC_NTHRS_INIT;
1797 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1804 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1807 if (svcpt->scp_nreqs_active < running - 2)
1810 if (svcpt->scp_nreqs_active >= running - 1)
1813 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1816 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1819 return ptlrpc_server_allow_normal(svcpt, force) &&
1820 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1824 * Returns true if there are requests available in incoming
1825 * request queue for processing and it is allowed to fetch them.
1826 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1827 * to get reliable result
1828 * \see ptlrpc_server_allow_normal
1829 * \see ptlrpc_server_allow high
1832 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1834 return ptlrpc_server_high_pending(svcpt, force) ||
1835 ptlrpc_server_normal_pending(svcpt, force);
1839 * Fetch a request for processing from queue of unprocessed requests.
1840 * Favors high-priority requests.
1841 * Returns a pointer to fetched request.
1843 static struct ptlrpc_request *
1844 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1846 struct ptlrpc_request *req = NULL;
1849 spin_lock(&svcpt->scp_req_lock);
1851 if (ptlrpc_server_high_pending(svcpt, force)) {
1852 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1854 svcpt->scp_hreq_count++;
1859 if (ptlrpc_server_normal_pending(svcpt, force)) {
1860 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1862 svcpt->scp_hreq_count = 0;
1867 spin_unlock(&svcpt->scp_req_lock);
1871 svcpt->scp_nreqs_active++;
1873 svcpt->scp_nhreqs_active++;
1875 spin_unlock(&svcpt->scp_req_lock);
1877 if (likely(req->rq_export))
1878 class_export_rpc_inc(req->rq_export);
1884 * Handle freshly incoming reqs, add to timed early reply list,
1885 * pass on to regular request queue.
1886 * All incoming requests pass through here before getting into
1887 * ptlrpc_server_handle_req later on.
1890 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1891 struct ptlrpc_thread *thread)
1893 struct ptlrpc_service *svc = svcpt->scp_service;
1894 struct ptlrpc_request *req;
1899 spin_lock(&svcpt->scp_lock);
1900 if (list_empty(&svcpt->scp_req_incoming)) {
1901 spin_unlock(&svcpt->scp_lock);
1905 req = list_entry(svcpt->scp_req_incoming.next,
1906 struct ptlrpc_request, rq_list);
1907 list_del_init(&req->rq_list);
1908 svcpt->scp_nreqs_incoming--;
1909 /* Consider this still a "queued" request as far as stats are
1911 spin_unlock(&svcpt->scp_lock);
1913 /* go through security check/transform */
1914 rc = sptlrpc_svc_unwrap_request(req);
1918 case SECSVC_COMPLETE:
1919 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1928 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1929 * redo it wouldn't be harmful.
1931 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1932 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1934 CERROR("error unpacking request: ptl %d from %s "
1935 "x%llu\n", svc->srv_req_portal,
1936 libcfs_id2str(req->rq_peer), req->rq_xid);
1941 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1943 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1944 "%llu\n", svc->srv_req_portal,
1945 libcfs_id2str(req->rq_peer), req->rq_xid);
1949 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1950 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1951 CERROR("drop incoming rpc opc %u, x%llu\n",
1952 cfs_fail_val, req->rq_xid);
1957 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1958 CERROR("wrong packet type received (type=%u) from %s\n",
1959 lustre_msg_get_type(req->rq_reqmsg),
1960 libcfs_id2str(req->rq_peer));
1964 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1968 req->rq_bulk_write = 1;
1972 case MGS_CONFIG_READ:
1973 req->rq_bulk_read = 1;
1977 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1979 req->rq_export = class_conn2export(
1980 lustre_msg_get_handle(req->rq_reqmsg));
1981 if (req->rq_export) {
1982 rc = ptlrpc_check_req(req);
1984 rc = sptlrpc_target_export_check(req->rq_export, req);
1986 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1987 "illegal security flavor,");
1992 ptlrpc_update_export_timer(req->rq_export, 0);
1995 /* req_in handling should/must be fast */
1996 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1997 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
1998 (s64)(ktime_get_real_seconds() -
1999 req->rq_arrival_time.tv_sec));
2001 /* Set rpc server deadline and add it to the timed list */
2002 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2003 MSGHDR_AT_SUPPORT) ?
2004 /* The max time the client expects us to take */
2005 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2007 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2008 if (unlikely(deadline == 0)) {
2009 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2013 /* Skip early reply */
2014 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2015 req->rq_deadline += obd_timeout;
2017 req->rq_svc_thread = thread;
2018 if (thread != NULL) {
2019 /* initialize request session, it is needed for request
2020 * processing by target */
2021 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2024 CERROR("%s: failure to initialize session: rc = %d\n",
2025 thread->t_name, rc);
2028 req->rq_session.lc_thread = thread;
2029 lu_context_enter(&req->rq_session);
2030 thread->t_env->le_ses = &req->rq_session;
2033 ptlrpc_at_add_timed(req);
2035 /* Move it over to the request processing queue */
2036 rc = ptlrpc_server_request_add(svcpt, req);
2040 wake_up(&svcpt->scp_waitq);
2044 ptlrpc_server_finish_request(svcpt, req);
2050 * Main incoming request handling logic.
2051 * Calls handler function from service to do actual processing.
2054 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2055 struct ptlrpc_thread *thread)
2057 struct ptlrpc_service *svc = svcpt->scp_service;
2058 struct ptlrpc_request *request;
2068 request = ptlrpc_server_request_get(svcpt, false);
2069 if (request == NULL)
2072 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2073 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2074 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2075 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2077 if (unlikely(fail_opc)) {
2078 if (request->rq_export && request->rq_ops)
2079 OBD_FAIL_TIMEOUT(fail_opc, 4);
2082 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2084 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2085 libcfs_debug_dumplog();
2087 work_start = ktime_get_real();
2088 arrived = timespec64_to_ktime(request->rq_arrival_time);
2089 timediff_usecs = ktime_us_delta(work_start, arrived);
2090 if (likely(svc->srv_stats != NULL)) {
2091 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2093 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2094 svcpt->scp_nreqs_incoming);
2095 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2096 svcpt->scp_nreqs_active);
2097 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2098 at_get(&svcpt->scp_at_estimate));
2101 if (likely(request->rq_export)) {
2102 if (unlikely(ptlrpc_check_req(request)))
2104 ptlrpc_update_export_timer(request->rq_export,
2105 timediff_usecs / (USEC_PER_SEC / 2));
2108 /* Discard requests queued for longer than the deadline.
2109 The deadline is increased if we send an early reply. */
2110 if (ktime_get_real_seconds() > request->rq_deadline) {
2111 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2112 libcfs_id2str(request->rq_peer),
2113 request->rq_deadline -
2114 request->rq_arrival_time.tv_sec,
2115 ktime_get_real_seconds() - request->rq_deadline);
2119 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2120 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2121 (request->rq_export ?
2122 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2123 (request->rq_export ?
2124 atomic_read(&request->rq_export->exp_refcount) : -99),
2125 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2126 libcfs_id2str(request->rq_peer),
2127 lustre_msg_get_opc(request->rq_reqmsg));
2129 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2130 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2132 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2134 /* re-assign request and sesson thread to the current one */
2135 request->rq_svc_thread = thread;
2136 if (thread != NULL) {
2137 LASSERT(request->rq_session.lc_thread == NULL);
2138 request->rq_session.lc_thread = thread;
2139 thread->t_env->le_ses = &request->rq_session;
2141 svc->srv_ops.so_req_handler(request);
2143 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2146 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2147 DEBUG_REQ(D_WARNING, request,
2148 "Request took longer than estimated (%lld:%llds); "
2149 "client may timeout.",
2150 request->rq_deadline -
2151 request->rq_arrival_time.tv_sec,
2152 ktime_get_real_seconds() - request->rq_deadline);
2155 work_end = ktime_get_real();
2156 timediff_usecs = ktime_us_delta(work_end, work_start);
2157 arrived_usecs = ktime_us_delta(work_end, arrived);
2158 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2159 "%s:%s+%d:%d:x%llu:%s:%d Request processed in %lldus "
2160 "(%lldus total) trans %llu rc %d/%d\n",
2162 (request->rq_export ?
2163 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2164 (request->rq_export ?
2165 atomic_read(&request->rq_export->exp_refcount) : -99),
2166 lustre_msg_get_status(request->rq_reqmsg),
2168 libcfs_id2str(request->rq_peer),
2169 lustre_msg_get_opc(request->rq_reqmsg),
2172 (request->rq_repmsg ?
2173 lustre_msg_get_transno(request->rq_repmsg) :
2174 request->rq_transno),
2176 (request->rq_repmsg ?
2177 lustre_msg_get_status(request->rq_repmsg) : -999));
2178 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2179 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2180 int opc = opcode_offset(op);
2181 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2182 LASSERT(opc < LUSTRE_MAX_OPCODES);
2183 lprocfs_counter_add(svc->srv_stats,
2184 opc + EXTRA_MAX_OPCODES,
2188 if (unlikely(request->rq_early_count)) {
2189 DEBUG_REQ(D_ADAPTTO, request,
2190 "sent %d early replies before finishing in %llds",
2191 request->rq_early_count,
2192 arrived_usecs / USEC_PER_SEC);
2195 ptlrpc_server_finish_active_request(svcpt, request);
2201 * An internal function to process a single reply state object.
2204 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2206 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2207 struct ptlrpc_service *svc = svcpt->scp_service;
2208 struct obd_export *exp;
2213 exp = rs->rs_export;
2215 LASSERT(rs->rs_difficult);
2216 LASSERT(rs->rs_scheduled);
2217 LASSERT(list_empty(&rs->rs_list));
2219 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2220 * iterates over newly committed replies, removing them from
2221 * exp_uncommitted_replies. It then drops this lock and schedules the
2222 * replies it found for handling here.
2224 * We can avoid contention for exp_uncommitted_replies_lock between the
2225 * HRT threads and further commit callbacks by checking rs_committed
2226 * which is set in the commit callback while it holds both
2227 * rs_lock and exp_uncommitted_reples.
2229 * If we see rs_committed clear, the commit callback _may_ not have
2230 * handled this reply yet and we race with it to grab
2231 * exp_uncommitted_replies_lock before removing the reply from
2232 * exp_uncommitted_replies. Note that if we lose the race and the
2233 * reply has already been removed, list_del_init() is a noop.
2235 * If we see rs_committed set, we know the commit callback is handling,
2236 * or has handled this reply since store reordering might allow us to
2237 * see rs_committed set out of sequence. But since this is done
2238 * holding rs_lock, we can be sure it has all completed once we hold
2239 * rs_lock, which we do right next.
2241 if (!rs->rs_committed) {
2242 /* if rs was commited, no need to convert locks, don't check
2243 * rs_committed here because rs may never be added into
2244 * exp_uncommitted_replies and this flag never be set, see
2245 * target_send_reply() */
2246 if (rs->rs_convert_lock &&
2247 rs->rs_transno > exp->exp_last_committed) {
2248 struct ldlm_lock *lock;
2249 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2251 spin_lock(&rs->rs_lock);
2252 if (rs->rs_convert_lock &&
2253 rs->rs_transno > exp->exp_last_committed) {
2254 nlocks = rs->rs_nlocks;
2255 while (nlocks-- > 0) {
2257 * NB don't assume rs is always handled
2258 * by the same service thread (see
2259 * ptlrpc_hr_select, so REP-ACK hr may
2260 * race with trans commit, while the
2261 * latter will release locks, get locks
2262 * here early to convert to COS mode
2265 lock = ldlm_handle2lock(
2266 &rs->rs_locks[nlocks]);
2268 ack_locks[nlocks] = lock;
2269 rs->rs_modes[nlocks] = LCK_COS;
2271 nlocks = rs->rs_nlocks;
2272 rs->rs_convert_lock = 0;
2273 /* clear rs_scheduled so that commit callback
2274 * can schedule again */
2275 rs->rs_scheduled = 0;
2276 spin_unlock(&rs->rs_lock);
2278 while (nlocks-- > 0) {
2279 lock = ack_locks[nlocks];
2280 ldlm_lock_mode_downgrade(lock, LCK_COS);
2281 LDLM_LOCK_PUT(lock);
2285 spin_unlock(&rs->rs_lock);
2288 spin_lock(&exp->exp_uncommitted_replies_lock);
2289 list_del_init(&rs->rs_obd_list);
2290 spin_unlock(&exp->exp_uncommitted_replies_lock);
2293 spin_lock(&exp->exp_lock);
2294 /* Noop if removed already */
2295 list_del_init(&rs->rs_exp_list);
2296 spin_unlock(&exp->exp_lock);
2298 spin_lock(&rs->rs_lock);
2300 been_handled = rs->rs_handled;
2303 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2304 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2306 if (nlocks == 0 && !been_handled) {
2307 /* If we see this, we should already have seen the warning
2308 * in mds_steal_ack_locks() */
2309 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2312 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2313 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2316 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2317 spin_unlock(&rs->rs_lock);
2319 if (!been_handled && rs->rs_on_net) {
2320 LNetMDUnlink(rs->rs_md_h);
2321 /* Ignore return code; we're racing with completion */
2324 while (nlocks-- > 0)
2325 ldlm_lock_decref(&rs->rs_locks[nlocks],
2326 rs->rs_modes[nlocks]);
2328 spin_lock(&rs->rs_lock);
2331 rs->rs_scheduled = 0;
2332 rs->rs_convert_lock = 0;
2334 if (!rs->rs_on_net) {
2336 spin_unlock(&rs->rs_lock);
2338 class_export_put (exp);
2339 rs->rs_export = NULL;
2340 ptlrpc_rs_decref(rs);
2341 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2342 svc->srv_is_stopping)
2343 wake_up_all(&svcpt->scp_waitq);
2347 /* still on the net; callback will schedule */
2348 spin_unlock(&rs->rs_lock);
2354 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2356 int avail = svcpt->scp_nrqbds_posted;
2357 int low_water = test_req_buffer_pressure ? 0 :
2358 svcpt->scp_service->srv_nbuf_per_group / 2;
2360 /* NB I'm not locking; just looking. */
2362 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2363 * allowed the request history to grow out of control. We could put a
2364 * sanity check on that here and cull some history if we need the
2367 if (avail <= low_water)
2368 ptlrpc_grow_req_bufs(svcpt, 1);
2370 if (svcpt->scp_service->srv_stats) {
2371 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2372 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2377 ptlrpc_retry_rqbds(void *arg)
2379 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2381 svcpt->scp_rqbd_timeout = 0;
2386 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2388 return svcpt->scp_nreqs_active <
2389 svcpt->scp_nthrs_running - 1 -
2390 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2394 * allowed to create more threads
2395 * user can call it w/o any lock but need to hold
2396 * ptlrpc_service_part::scp_lock to get reliable result
2399 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2401 return svcpt->scp_nthrs_running +
2402 svcpt->scp_nthrs_starting <
2403 svcpt->scp_service->srv_nthrs_cpt_limit;
2407 * too many requests and allowed to create more threads
2410 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2412 return !ptlrpc_threads_enough(svcpt) &&
2413 ptlrpc_threads_increasable(svcpt);
2417 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2419 return thread_is_stopping(thread) ||
2420 thread->t_svcpt->scp_service->srv_is_stopping;
2424 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2426 return !list_empty(&svcpt->scp_rqbd_idle) &&
2427 svcpt->scp_rqbd_timeout == 0;
2431 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2433 return svcpt->scp_at_check;
2437 * requests wait on preprocessing
2438 * user can call it w/o any lock but need to hold
2439 * ptlrpc_service_part::scp_lock to get reliable result
2442 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2444 return !list_empty(&svcpt->scp_req_incoming);
2447 static __attribute__((__noinline__)) int
2448 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2449 struct ptlrpc_thread *thread)
2451 /* Don't exit while there are replies to be handled */
2452 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2453 ptlrpc_retry_rqbds, svcpt);
2455 lc_watchdog_disable(thread->t_watchdog);
2459 l_wait_event_exclusive_head(svcpt->scp_waitq,
2460 ptlrpc_thread_stopping(thread) ||
2461 ptlrpc_server_request_incoming(svcpt) ||
2462 ptlrpc_server_request_pending(svcpt, false) ||
2463 ptlrpc_rqbd_pending(svcpt) ||
2464 ptlrpc_at_check(svcpt), &lwi);
2466 if (ptlrpc_thread_stopping(thread))
2469 lc_watchdog_touch(thread->t_watchdog,
2470 ptlrpc_server_get_timeout(svcpt));
2475 * Main thread body for service threads.
2476 * Waits in a loop waiting for new requests to process to appear.
2477 * Every time an incoming requests is added to its queue, a waitq
2478 * is woken up and one of the threads will handle it.
2480 static int ptlrpc_main(void *arg)
2482 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2483 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2484 struct ptlrpc_service *svc = svcpt->scp_service;
2485 struct ptlrpc_reply_state *rs;
2486 struct group_info *ginfo = NULL;
2488 int counter = 0, rc = 0;
2491 thread->t_pid = current_pid();
2492 unshare_fs_struct();
2494 /* NB: we will call cfs_cpt_bind() for all threads, because we
2495 * might want to run lustre server only on a subset of system CPUs,
2496 * in that case ->scp_cpt is CFS_CPT_ANY */
2497 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2499 CWARN("%s: failed to bind %s on CPT %d\n",
2500 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2503 ginfo = groups_alloc(0);
2509 set_current_groups(ginfo);
2510 put_group_info(ginfo);
2512 if (svc->srv_ops.so_thr_init != NULL) {
2513 rc = svc->srv_ops.so_thr_init(thread);
2524 rc = lu_context_init(&env->le_ctx,
2525 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2529 thread->t_env = env;
2530 env->le_ctx.lc_thread = thread;
2531 env->le_ctx.lc_cookie = 0x6;
2533 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2534 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2538 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2539 svc->srv_name, svcpt->scp_cpt, rc);
2543 /* Alloc reply state structure for this one */
2544 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2550 spin_lock(&svcpt->scp_lock);
2552 LASSERT(thread_is_starting(thread));
2553 thread_clear_flags(thread, SVC_STARTING);
2555 LASSERT(svcpt->scp_nthrs_starting == 1);
2556 svcpt->scp_nthrs_starting--;
2558 /* SVC_STOPPING may already be set here if someone else is trying
2559 * to stop the service while this new thread has been dynamically
2560 * forked. We still set SVC_RUNNING to let our creator know that
2561 * we are now running, however we will exit as soon as possible */
2562 thread_add_flags(thread, SVC_RUNNING);
2563 svcpt->scp_nthrs_running++;
2564 spin_unlock(&svcpt->scp_lock);
2566 /* wake up our creator in case he's still waiting. */
2567 wake_up(&thread->t_ctl_waitq);
2569 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2572 spin_lock(&svcpt->scp_rep_lock);
2573 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2574 wake_up(&svcpt->scp_rep_waitq);
2575 spin_unlock(&svcpt->scp_rep_lock);
2577 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2578 svcpt->scp_nthrs_running);
2580 /* XXX maintain a list of all managed devices: insert here */
2581 while (!ptlrpc_thread_stopping(thread)) {
2582 if (ptlrpc_wait_event(svcpt, thread))
2585 ptlrpc_check_rqbd_pool(svcpt);
2587 if (ptlrpc_threads_need_create(svcpt)) {
2588 /* Ignore return code - we tried... */
2589 ptlrpc_start_thread(svcpt, 0);
2592 /* reset le_ses to initial state */
2594 /* Process all incoming reqs before handling any */
2595 if (ptlrpc_server_request_incoming(svcpt)) {
2596 lu_context_enter(&env->le_ctx);
2597 ptlrpc_server_handle_req_in(svcpt, thread);
2598 lu_context_exit(&env->le_ctx);
2600 /* but limit ourselves in case of flood */
2601 if (counter++ < 100)
2606 if (ptlrpc_at_check(svcpt))
2607 ptlrpc_at_check_timed(svcpt);
2609 if (ptlrpc_server_request_pending(svcpt, false)) {
2610 lu_context_enter(&env->le_ctx);
2611 ptlrpc_server_handle_request(svcpt, thread);
2612 lu_context_exit(&env->le_ctx);
2615 if (ptlrpc_rqbd_pending(svcpt) &&
2616 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2617 /* I just failed to repost request buffers.
2618 * Wait for a timeout (unless something else
2619 * happens) before I try again */
2620 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2621 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2622 svcpt->scp_nrqbds_posted);
2626 lc_watchdog_delete(thread->t_watchdog);
2627 thread->t_watchdog = NULL;
2631 * deconstruct service specific state created by ptlrpc_start_thread()
2633 if (svc->srv_ops.so_thr_done != NULL)
2634 svc->srv_ops.so_thr_done(thread);
2637 lu_context_fini(&env->le_ctx);
2641 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2642 thread, thread->t_pid, thread->t_id, rc);
2644 spin_lock(&svcpt->scp_lock);
2645 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2646 svcpt->scp_nthrs_starting--;
2648 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2649 /* must know immediately */
2650 svcpt->scp_nthrs_running--;
2654 thread_add_flags(thread, SVC_STOPPED);
2656 wake_up(&thread->t_ctl_waitq);
2657 spin_unlock(&svcpt->scp_lock);
2662 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2663 struct list_head *replies)
2667 spin_lock(&hrt->hrt_lock);
2669 list_splice_init(&hrt->hrt_queue, replies);
2670 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2672 spin_unlock(&hrt->hrt_lock);
2677 * Main body of "handle reply" function.
2678 * It processes acked reply states
2680 static int ptlrpc_hr_main(void *arg)
2682 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2683 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2684 struct list_head replies;
2687 INIT_LIST_HEAD(&replies);
2688 unshare_fs_struct();
2690 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2692 char threadname[20];
2694 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2695 hrp->hrp_cpt, hrt->hrt_id);
2696 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2697 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2700 atomic_inc(&hrp->hrp_nstarted);
2701 wake_up(&ptlrpc_hr.hr_waitq);
2703 while (!ptlrpc_hr.hr_stopping) {
2704 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2706 while (!list_empty(&replies)) {
2707 struct ptlrpc_reply_state *rs;
2709 rs = list_entry(replies.prev,
2710 struct ptlrpc_reply_state,
2712 list_del_init(&rs->rs_list);
2713 ptlrpc_handle_rs(rs);
2717 atomic_inc(&hrp->hrp_nstopped);
2718 wake_up(&ptlrpc_hr.hr_waitq);
2723 static void ptlrpc_stop_hr_threads(void)
2725 struct ptlrpc_hr_partition *hrp;
2729 ptlrpc_hr.hr_stopping = 1;
2731 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2732 if (hrp->hrp_thrs == NULL)
2733 continue; /* uninitialized */
2734 for (j = 0; j < hrp->hrp_nthrs; j++)
2735 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2738 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2739 if (hrp->hrp_thrs == NULL)
2740 continue; /* uninitialized */
2741 wait_event(ptlrpc_hr.hr_waitq,
2742 atomic_read(&hrp->hrp_nstopped) ==
2743 atomic_read(&hrp->hrp_nstarted));
2747 static int ptlrpc_start_hr_threads(void)
2749 struct ptlrpc_hr_partition *hrp;
2754 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2757 for (j = 0; j < hrp->hrp_nthrs; j++) {
2758 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2759 struct task_struct *task;
2761 task = kthread_run(ptlrpc_hr_main,
2763 "ptlrpc_hr%02d_%03d",
2772 wait_event(ptlrpc_hr.hr_waitq,
2773 atomic_read(&hrp->hrp_nstarted) == j);
2776 CERROR("cannot start reply handler thread %d:%d: "
2777 "rc = %d\n", i, j, rc);
2778 ptlrpc_stop_hr_threads();
2786 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2788 struct l_wait_info lwi = { 0 };
2789 struct ptlrpc_thread *thread;
2790 struct list_head zombie;
2794 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2795 svcpt->scp_service->srv_name);
2797 INIT_LIST_HEAD(&zombie);
2798 spin_lock(&svcpt->scp_lock);
2799 /* let the thread know that we would like it to stop asap */
2800 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2801 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2802 svcpt->scp_service->srv_thread_name, thread->t_id);
2803 thread_add_flags(thread, SVC_STOPPING);
2806 wake_up_all(&svcpt->scp_waitq);
2808 while (!list_empty(&svcpt->scp_threads)) {
2809 thread = list_entry(svcpt->scp_threads.next,
2810 struct ptlrpc_thread, t_link);
2811 if (thread_is_stopped(thread)) {
2812 list_del(&thread->t_link);
2813 list_add(&thread->t_link, &zombie);
2816 spin_unlock(&svcpt->scp_lock);
2818 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2819 svcpt->scp_service->srv_thread_name, thread->t_id);
2820 l_wait_event(thread->t_ctl_waitq,
2821 thread_is_stopped(thread), &lwi);
2823 spin_lock(&svcpt->scp_lock);
2826 spin_unlock(&svcpt->scp_lock);
2828 while (!list_empty(&zombie)) {
2829 thread = list_entry(zombie.next,
2830 struct ptlrpc_thread, t_link);
2831 list_del(&thread->t_link);
2832 OBD_FREE_PTR(thread);
2838 * Stops all threads of a particular service \a svc
2840 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2842 struct ptlrpc_service_part *svcpt;
2846 ptlrpc_service_for_each_part(svcpt, i, svc) {
2847 if (svcpt->scp_service != NULL)
2848 ptlrpc_svcpt_stop_threads(svcpt);
2854 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2861 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2862 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2864 for (i = 0; i < svc->srv_ncpts; i++) {
2865 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2866 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2872 /* We have enough threads, don't start more. b=15759 */
2879 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2880 svc->srv_thread_name, i, j, rc);
2881 ptlrpc_stop_all_threads(svc);
2885 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2887 struct l_wait_info lwi = { 0 };
2888 struct ptlrpc_thread *thread;
2889 struct ptlrpc_service *svc;
2890 struct task_struct *task;
2894 LASSERT(svcpt != NULL);
2896 svc = svcpt->scp_service;
2898 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2899 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2900 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2903 if (unlikely(svc->srv_is_stopping))
2906 if (!ptlrpc_threads_increasable(svcpt) ||
2907 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2908 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2911 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2914 init_waitqueue_head(&thread->t_ctl_waitq);
2916 spin_lock(&svcpt->scp_lock);
2917 if (!ptlrpc_threads_increasable(svcpt)) {
2918 spin_unlock(&svcpt->scp_lock);
2919 OBD_FREE_PTR(thread);
2923 if (svcpt->scp_nthrs_starting != 0) {
2924 /* serialize starting because some modules (obdfilter)
2925 * might require unique and contiguous t_id */
2926 LASSERT(svcpt->scp_nthrs_starting == 1);
2927 spin_unlock(&svcpt->scp_lock);
2928 OBD_FREE_PTR(thread);
2930 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2931 svc->srv_thread_name, svcpt->scp_thr_nextid);
2936 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2937 svc->srv_thread_name, svcpt->scp_thr_nextid);
2941 svcpt->scp_nthrs_starting++;
2942 thread->t_id = svcpt->scp_thr_nextid++;
2943 thread_add_flags(thread, SVC_STARTING);
2944 thread->t_svcpt = svcpt;
2946 list_add(&thread->t_link, &svcpt->scp_threads);
2947 spin_unlock(&svcpt->scp_lock);
2949 if (svcpt->scp_cpt >= 0) {
2950 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2951 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2953 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2954 svc->srv_thread_name, thread->t_id);
2957 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2958 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2961 CERROR("cannot start thread '%s': rc = %d\n",
2962 thread->t_name, rc);
2963 spin_lock(&svcpt->scp_lock);
2964 --svcpt->scp_nthrs_starting;
2965 if (thread_is_stopping(thread)) {
2966 /* this ptlrpc_thread is being hanled
2967 * by ptlrpc_svcpt_stop_threads now
2969 thread_add_flags(thread, SVC_STOPPED);
2970 wake_up(&thread->t_ctl_waitq);
2971 spin_unlock(&svcpt->scp_lock);
2973 list_del(&thread->t_link);
2974 spin_unlock(&svcpt->scp_lock);
2975 OBD_FREE_PTR(thread);
2983 l_wait_event(thread->t_ctl_waitq,
2984 thread_is_running(thread) || thread_is_stopped(thread),
2987 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2991 int ptlrpc_hr_init(void)
2993 struct ptlrpc_hr_partition *hrp;
2994 struct ptlrpc_hr_thread *hrt;
3001 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3002 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3004 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3006 if (ptlrpc_hr.hr_partitions == NULL)
3009 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3011 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3013 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3016 atomic_set(&hrp->hrp_nstarted, 0);
3017 atomic_set(&hrp->hrp_nstopped, 0);
3019 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3020 hrp->hrp_nthrs /= weight;
3021 if (hrp->hrp_nthrs == 0)
3024 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3025 hrp->hrp_nthrs * sizeof(*hrt));
3026 if (hrp->hrp_thrs == NULL)
3027 GOTO(out, rc = -ENOMEM);
3029 for (i = 0; i < hrp->hrp_nthrs; i++) {
3030 hrt = &hrp->hrp_thrs[i];
3033 hrt->hrt_partition = hrp;
3034 init_waitqueue_head(&hrt->hrt_waitq);
3035 spin_lock_init(&hrt->hrt_lock);
3036 INIT_LIST_HEAD(&hrt->hrt_queue);
3040 rc = ptlrpc_start_hr_threads();
3047 void ptlrpc_hr_fini(void)
3049 struct ptlrpc_hr_partition *hrp;
3052 if (ptlrpc_hr.hr_partitions == NULL)
3055 ptlrpc_stop_hr_threads();
3057 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3058 if (hrp->hrp_thrs != NULL) {
3059 OBD_FREE(hrp->hrp_thrs,
3060 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3064 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3065 ptlrpc_hr.hr_partitions = NULL;
3070 * Wait until all already scheduled replies are processed.
3072 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3076 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3079 rc = l_wait_event(svcpt->scp_waitq,
3080 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3083 CWARN("Unexpectedly long timeout %s %p\n",
3084 svcpt->scp_service->srv_name, svcpt->scp_service);
3089 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3091 struct ptlrpc_service_part *svcpt;
3094 /* early disarm AT timer... */
3095 ptlrpc_service_for_each_part(svcpt, i, svc) {
3096 if (svcpt->scp_service != NULL)
3097 del_timer(&svcpt->scp_at_timer);
3102 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3104 struct ptlrpc_service_part *svcpt;
3105 struct ptlrpc_request_buffer_desc *rqbd;
3106 struct l_wait_info lwi;
3110 /* All history will be culled when the next request buffer is
3111 * freed in ptlrpc_service_purge_all() */
3112 svc->srv_hist_nrqbds_cpt_max = 0;
3114 rc = LNetClearLazyPortal(svc->srv_req_portal);
3117 ptlrpc_service_for_each_part(svcpt, i, svc) {
3118 if (svcpt->scp_service == NULL)
3121 /* Unlink all the request buffers. This forces a 'final'
3122 * event with its 'unlink' flag set for each posted rqbd */
3123 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3125 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3126 LASSERT(rc == 0 || rc == -ENOENT);
3130 ptlrpc_service_for_each_part(svcpt, i, svc) {
3131 if (svcpt->scp_service == NULL)
3134 /* Wait for the network to release any buffers
3135 * it's currently filling */
3136 spin_lock(&svcpt->scp_lock);
3137 while (svcpt->scp_nrqbds_posted != 0) {
3138 spin_unlock(&svcpt->scp_lock);
3139 /* Network access will complete in finite time but
3140 * the HUGE timeout lets us CWARN for visibility
3141 * of sluggish NALs */
3142 lwi = LWI_TIMEOUT_INTERVAL(
3143 cfs_time_seconds(LONG_UNLINK),
3144 cfs_time_seconds(1), NULL, NULL);
3145 rc = l_wait_event(svcpt->scp_waitq,
3146 svcpt->scp_nrqbds_posted == 0, &lwi);
3147 if (rc == -ETIMEDOUT) {
3148 CWARN("Service %s waiting for "
3149 "request buffers\n",
3150 svcpt->scp_service->srv_name);
3152 spin_lock(&svcpt->scp_lock);
3154 spin_unlock(&svcpt->scp_lock);
3159 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3161 struct ptlrpc_service_part *svcpt;
3162 struct ptlrpc_request_buffer_desc *rqbd;
3163 struct ptlrpc_request *req;
3164 struct ptlrpc_reply_state *rs;
3167 ptlrpc_service_for_each_part(svcpt, i, svc) {
3168 if (svcpt->scp_service == NULL)
3171 spin_lock(&svcpt->scp_rep_lock);
3172 while (!list_empty(&svcpt->scp_rep_active)) {
3173 rs = list_entry(svcpt->scp_rep_active.next,
3174 struct ptlrpc_reply_state, rs_list);
3175 spin_lock(&rs->rs_lock);
3176 ptlrpc_schedule_difficult_reply(rs);
3177 spin_unlock(&rs->rs_lock);
3179 spin_unlock(&svcpt->scp_rep_lock);
3181 /* purge the request queue. NB No new replies (rqbds
3182 * all unlinked) and no service threads, so I'm the only
3183 * thread noodling the request queue now */
3184 while (!list_empty(&svcpt->scp_req_incoming)) {
3185 req = list_entry(svcpt->scp_req_incoming.next,
3186 struct ptlrpc_request, rq_list);
3188 list_del(&req->rq_list);
3189 svcpt->scp_nreqs_incoming--;
3190 ptlrpc_server_finish_request(svcpt, req);
3193 while (ptlrpc_server_request_pending(svcpt, true)) {
3194 req = ptlrpc_server_request_get(svcpt, true);
3195 ptlrpc_server_finish_active_request(svcpt, req);
3198 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3199 LASSERT(svcpt->scp_nreqs_incoming == 0);
3200 LASSERT(svcpt->scp_nreqs_active == 0);
3201 /* history should have been culled by
3202 * ptlrpc_server_finish_request */
3203 LASSERT(svcpt->scp_hist_nrqbds == 0);
3205 /* Now free all the request buffers since nothing
3206 * references them any more... */
3208 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3209 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3210 struct ptlrpc_request_buffer_desc,
3212 ptlrpc_free_rqbd(rqbd);
3214 ptlrpc_wait_replies(svcpt);
3216 while (!list_empty(&svcpt->scp_rep_idle)) {
3217 rs = list_entry(svcpt->scp_rep_idle.next,
3218 struct ptlrpc_reply_state,
3220 list_del(&rs->rs_list);
3221 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3227 ptlrpc_service_free(struct ptlrpc_service *svc)
3229 struct ptlrpc_service_part *svcpt;
3230 struct ptlrpc_at_array *array;
3233 ptlrpc_service_for_each_part(svcpt, i, svc) {
3234 if (svcpt->scp_service == NULL)
3237 /* In case somebody rearmed this in the meantime */
3238 del_timer(&svcpt->scp_at_timer);
3239 array = &svcpt->scp_at_array;
3241 if (array->paa_reqs_array != NULL) {
3242 OBD_FREE(array->paa_reqs_array,
3243 sizeof(struct list_head) * array->paa_size);
3244 array->paa_reqs_array = NULL;
3247 if (array->paa_reqs_count != NULL) {
3248 OBD_FREE(array->paa_reqs_count,
3249 sizeof(__u32) * array->paa_size);
3250 array->paa_reqs_count = NULL;
3254 ptlrpc_service_for_each_part(svcpt, i, svc)
3255 OBD_FREE_PTR(svcpt);
3257 if (svc->srv_cpts != NULL)
3258 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3260 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3261 srv_parts[svc->srv_ncpts]));
3264 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3268 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3270 service->srv_is_stopping = 1;
3272 mutex_lock(&ptlrpc_all_services_mutex);
3273 list_del_init(&service->srv_list);
3274 mutex_unlock(&ptlrpc_all_services_mutex);
3276 ptlrpc_service_del_atimer(service);
3277 ptlrpc_stop_all_threads(service);
3279 ptlrpc_service_unlink_rqbd(service);
3280 ptlrpc_service_purge_all(service);
3281 ptlrpc_service_nrs_cleanup(service);
3283 ptlrpc_lprocfs_unregister_service(service);
3284 ptlrpc_sysfs_unregister_service(service);
3286 ptlrpc_service_free(service);
3290 EXPORT_SYMBOL(ptlrpc_unregister_service);
3293 * Returns 0 if the service is healthy.
3295 * Right now, it just checks to make sure that requests aren't languishing
3296 * in the queue. We'll use this health check to govern whether a node needs
3297 * to be shot, so it's intentionally non-aggressive. */
3298 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3300 struct ptlrpc_request *request = NULL;
3301 struct timespec64 right_now;
3302 struct timespec64 timediff;
3304 ktime_get_real_ts64(&right_now);
3306 spin_lock(&svcpt->scp_req_lock);
3307 /* How long has the next entry been waiting? */
3308 if (ptlrpc_server_high_pending(svcpt, true))
3309 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3310 else if (ptlrpc_server_normal_pending(svcpt, true))
3311 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3313 if (request == NULL) {
3314 spin_unlock(&svcpt->scp_req_lock);
3318 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3319 spin_unlock(&svcpt->scp_req_lock);
3321 if ((timediff.tv_sec) >
3322 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3323 CERROR("%s: unhealthy - request has been waiting %llds\n",
3324 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3332 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3334 struct ptlrpc_service_part *svcpt;
3340 ptlrpc_service_for_each_part(svcpt, i, svc) {
3341 int rc = ptlrpc_svcpt_health_check(svcpt);
3348 EXPORT_SYMBOL(ptlrpc_service_health_check);