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 = cfs_time_current();
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 proc_dir_entry *proc_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 (proc_entry != NULL)
817 ptlrpc_lprocfs_register_service(proc_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 = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
1197 ptlrpc_at_timer((unsigned long)svcpt);
1199 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1200 CDEBUG(D_INFO, "armed %s at %+ds\n",
1201 svcpt->scp_service->srv_name, next);
1205 /* Add rpc to early reply check list */
1206 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1208 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1209 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1210 struct ptlrpc_request *rq = NULL;
1216 if (req->rq_no_reply)
1219 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1222 spin_lock(&svcpt->scp_at_lock);
1223 LASSERT(list_empty(&req->rq_timed_list));
1225 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1226 if (array->paa_reqs_count[index] > 0) {
1227 /* latest rpcs will have the latest deadlines in the list,
1228 * so search backward. */
1229 list_for_each_entry_reverse(rq,
1230 &array->paa_reqs_array[index],
1232 if (req->rq_deadline >= rq->rq_deadline) {
1233 list_add(&req->rq_timed_list,
1234 &rq->rq_timed_list);
1240 /* Add the request at the head of the list */
1241 if (list_empty(&req->rq_timed_list))
1242 list_add(&req->rq_timed_list,
1243 &array->paa_reqs_array[index]);
1245 spin_lock(&req->rq_lock);
1246 req->rq_at_linked = 1;
1247 spin_unlock(&req->rq_lock);
1248 req->rq_at_index = index;
1249 array->paa_reqs_count[index]++;
1251 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1252 array->paa_deadline = req->rq_deadline;
1253 ptlrpc_at_set_timer(svcpt);
1255 spin_unlock(&svcpt->scp_at_lock);
1261 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1263 struct ptlrpc_at_array *array;
1265 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1267 /* NB: must call with hold svcpt::scp_at_lock */
1268 LASSERT(!list_empty(&req->rq_timed_list));
1269 list_del_init(&req->rq_timed_list);
1271 spin_lock(&req->rq_lock);
1272 req->rq_at_linked = 0;
1273 spin_unlock(&req->rq_lock);
1275 array->paa_reqs_count[req->rq_at_index]--;
1280 * Attempt to extend the request deadline by sending an early reply to the
1283 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1285 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1286 struct ptlrpc_request *reqcopy;
1287 struct lustre_msg *reqmsg;
1288 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1294 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1295 /* don't send early reply */
1299 /* deadline is when the client expects us to reply, margin is the
1300 difference between clients' and servers' expectations */
1301 DEBUG_REQ(D_ADAPTTO, req,
1302 "%ssending early reply (deadline %+llds, margin %+llds) for "
1303 "%d+%d", AT_OFF ? "AT off - not " : "",
1304 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1305 at_get(&svcpt->scp_at_estimate), at_extra);
1311 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1312 "not sending early reply. Consider increasing "
1313 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1315 /* Return an error so we're not re-added to the timed list. */
1319 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1320 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1321 "but no AT support");
1325 if (req->rq_export &&
1326 lustre_msg_get_flags(req->rq_reqmsg) &
1327 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1328 struct obd_device *obd_exp = req->rq_export->exp_obd;
1330 /* During recovery, we don't want to send too many early
1331 * replies, but on the other hand we want to make sure the
1332 * client has enough time to resend if the rpc is lost. So
1333 * during the recovery period send at least 4 early replies,
1334 * spacing them every at_extra if we can. at_estimate should
1335 * always equal this fixed value during recovery.
1337 /* Don't account request processing time into AT history
1338 * during recovery, it is not service time we need but
1339 * includes also waiting time for recovering clients
1341 newdl = min_t(time64_t, at_extra,
1342 obd_exp->obd_recovery_timeout / 4) +
1343 ktime_get_real_seconds();
1345 /* We want to extend the request deadline by at_extra seconds,
1346 * so we set our service estimate to reflect how much time has
1347 * passed since this request arrived plus an additional
1348 * at_extra seconds. The client will calculate the new deadline
1349 * based on this service estimate (plus some additional time to
1350 * account for network latency). See ptlrpc_at_recv_early_reply
1352 at_measured(&svcpt->scp_at_estimate, at_extra +
1353 ktime_get_real_seconds() -
1354 req->rq_arrival_time.tv_sec);
1355 newdl = req->rq_arrival_time.tv_sec +
1356 at_get(&svcpt->scp_at_estimate);
1359 /* Check to see if we've actually increased the deadline -
1360 * we may be past adaptive_max */
1361 if (req->rq_deadline >= newdl) {
1362 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1363 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1367 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1368 if (reqcopy == NULL)
1370 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1372 GOTO(out_free, rc = -ENOMEM);
1375 reqcopy->rq_reply_state = NULL;
1376 reqcopy->rq_rep_swab_mask = 0;
1377 reqcopy->rq_pack_bulk = 0;
1378 reqcopy->rq_pack_udesc = 0;
1379 reqcopy->rq_packed_final = 0;
1380 sptlrpc_svc_ctx_addref(reqcopy);
1381 /* We only need the reqmsg for the magic */
1382 reqcopy->rq_reqmsg = reqmsg;
1383 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1386 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1387 * Without this check, we would fail the rq_no_reply assertion in
1388 * ptlrpc_send_reply().
1390 if (reqcopy->rq_no_reply)
1391 GOTO(out, rc = -ETIMEDOUT);
1393 LASSERT(atomic_read(&req->rq_refcount));
1394 /** if it is last refcount then early reply isn't needed */
1395 if (atomic_read(&req->rq_refcount) == 1) {
1396 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1397 "abort sending early reply\n");
1398 GOTO(out, rc = -EINVAL);
1401 /* Connection ref */
1402 reqcopy->rq_export = class_conn2export(
1403 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1404 if (reqcopy->rq_export == NULL)
1405 GOTO(out, rc = -ENODEV);
1408 class_export_rpc_inc(reqcopy->rq_export);
1409 if (reqcopy->rq_export->exp_obd &&
1410 reqcopy->rq_export->exp_obd->obd_fail)
1411 GOTO(out_put, rc = -ENODEV);
1413 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1417 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1420 /* Adjust our own deadline to what we told the client */
1421 req->rq_deadline = newdl;
1422 req->rq_early_count++; /* number sent, server side */
1424 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1427 /* Free the (early) reply state from lustre_pack_reply.
1428 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1429 ptlrpc_req_drop_rs(reqcopy);
1432 class_export_rpc_dec(reqcopy->rq_export);
1433 class_export_put(reqcopy->rq_export);
1435 sptlrpc_svc_ctx_decref(reqcopy);
1436 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1438 ptlrpc_request_cache_free(reqcopy);
1442 /* Send early replies to everybody expiring within at_early_margin
1443 asking for at_extra time */
1444 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1446 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1447 struct ptlrpc_request *rq, *n;
1448 struct list_head work_list;
1451 time64_t now = ktime_get_real_seconds();
1452 cfs_duration_t delay;
1453 int first, counter = 0;
1456 spin_lock(&svcpt->scp_at_lock);
1457 if (svcpt->scp_at_check == 0) {
1458 spin_unlock(&svcpt->scp_at_lock);
1461 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1462 svcpt->scp_at_check = 0;
1464 if (array->paa_count == 0) {
1465 spin_unlock(&svcpt->scp_at_lock);
1469 /* The timer went off, but maybe the nearest rpc already completed. */
1470 first = array->paa_deadline - now;
1471 if (first > at_early_margin) {
1472 /* We've still got plenty of time. Reset the timer. */
1473 ptlrpc_at_set_timer(svcpt);
1474 spin_unlock(&svcpt->scp_at_lock);
1478 /* We're close to a timeout, and we don't know how much longer the
1479 server will take. Send early replies to everyone expiring soon. */
1480 INIT_LIST_HEAD(&work_list);
1482 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1483 count = array->paa_count;
1485 count -= array->paa_reqs_count[index];
1486 list_for_each_entry_safe(rq, n,
1487 &array->paa_reqs_array[index],
1489 if (rq->rq_deadline > now + at_early_margin) {
1490 /* update the earliest deadline */
1491 if (deadline == -1 ||
1492 rq->rq_deadline < deadline)
1493 deadline = rq->rq_deadline;
1497 ptlrpc_at_remove_timed(rq);
1499 * ptlrpc_server_drop_request() may drop
1500 * refcount to 0 already. Let's check this and
1501 * don't add entry to work_list
1503 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1504 list_add(&rq->rq_timed_list, &work_list);
1508 if (++index >= array->paa_size)
1511 array->paa_deadline = deadline;
1512 /* we have a new earliest deadline, restart the timer */
1513 ptlrpc_at_set_timer(svcpt);
1515 spin_unlock(&svcpt->scp_at_lock);
1517 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1518 "replies\n", first, at_extra, counter);
1520 /* We're already past request deadlines before we even get a
1521 chance to send early replies */
1522 LCONSOLE_WARN("%s: This server is not able to keep up with "
1523 "request traffic (cpu-bound).\n",
1524 svcpt->scp_service->srv_name);
1525 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1526 counter, svcpt->scp_nreqs_incoming,
1527 svcpt->scp_nreqs_active,
1528 at_get(&svcpt->scp_at_estimate), delay);
1531 /* we took additional refcount so entries can't be deleted from list, no
1532 * locking is needed */
1533 while (!list_empty(&work_list)) {
1534 rq = list_entry(work_list.next, struct ptlrpc_request,
1536 list_del_init(&rq->rq_timed_list);
1538 if (ptlrpc_at_send_early_reply(rq) == 0)
1539 ptlrpc_at_add_timed(rq);
1541 ptlrpc_server_drop_request(rq);
1544 RETURN(1); /* return "did_something" for liblustre */
1547 /* Check if we are already handling earlier incarnation of this request.
1548 * Called under &req->rq_export->exp_rpc_lock locked */
1549 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1551 struct ptlrpc_request *tmp = NULL;
1553 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1554 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1557 /* bulk request are aborted upon reconnect, don't try to
1559 if (req->rq_bulk_write || req->rq_bulk_read)
1562 /* This list should not be longer than max_requests in
1563 * flights on the client, so it is not all that long.
1564 * Also we only hit this codepath in case of a resent
1565 * request which makes it even more rarely hit */
1566 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1568 /* Found duplicate one */
1569 if (tmp->rq_xid == req->rq_xid)
1572 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1574 /* Found duplicate one */
1575 if (tmp->rq_xid == req->rq_xid)
1581 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1582 DEBUG_REQ(D_HA, tmp, "Request being processed");
1587 * Check if a request should be assigned with a high priority.
1589 * \retval < 0: error occurred
1590 * 0: normal RPC request
1591 * +1: high priority request
1593 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1594 struct ptlrpc_request *req)
1599 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1600 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1607 if (req->rq_export != NULL && req->rq_ops != NULL) {
1608 /* Perform request specific check. We should do this
1609 * check before the request is added into exp_hp_rpcs
1610 * list otherwise it may hit swab race at LU-1044. */
1611 if (req->rq_ops->hpreq_check != NULL) {
1612 rc = req->rq_ops->hpreq_check(req);
1613 if (rc == -ESTALE) {
1614 req->rq_status = rc;
1617 /** can only return error,
1618 * 0 for normal request,
1619 * or 1 for high priority request */
1627 /** Remove the request from the export list. */
1628 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1631 if (req->rq_export) {
1632 /* refresh lock timeout again so that client has more
1633 * room to send lock cancel RPC. */
1634 if (req->rq_ops && req->rq_ops->hpreq_fini)
1635 req->rq_ops->hpreq_fini(req);
1637 spin_lock(&req->rq_export->exp_rpc_lock);
1638 list_del_init(&req->rq_exp_list);
1639 spin_unlock(&req->rq_export->exp_rpc_lock);
1644 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1649 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1650 .hpreq_check = ptlrpc_hpreq_check,
1653 /* Hi-Priority RPC check by RPC operation code. */
1654 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1656 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1658 /* Check for export to let only reconnects for not yet evicted
1659 * export to become a HP rpc. */
1660 if ((req->rq_export != NULL) &&
1661 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1662 req->rq_ops = &ptlrpc_hpreq_common;
1666 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1668 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1669 struct ptlrpc_request *req)
1675 rc = ptlrpc_server_hpreq_init(svcpt, req);
1680 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1682 if (req->rq_export != NULL) {
1683 struct obd_export *exp = req->rq_export;
1685 /* do search for duplicated xid and the adding to the list
1687 spin_lock_bh(&exp->exp_rpc_lock);
1688 rc = ptlrpc_server_check_resend_in_progress(req);
1690 spin_unlock_bh(&exp->exp_rpc_lock);
1692 ptlrpc_nrs_req_finalize(req);
1696 if (hp || req->rq_ops != NULL)
1697 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1699 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1700 spin_unlock_bh(&exp->exp_rpc_lock);
1703 /* the current thread is not the processing thread for this request
1704 * since that, but request is in exp_hp_list and can be find there.
1705 * Remove all relations between request and old thread. */
1706 req->rq_svc_thread->t_env->le_ses = NULL;
1707 req->rq_svc_thread = NULL;
1708 req->rq_session.lc_thread = NULL;
1710 ptlrpc_nrs_req_add(svcpt, req, hp);
1716 * Allow to handle high priority request
1717 * User can call it w/o any lock but need to hold
1718 * ptlrpc_service_part::scp_req_lock to get reliable result
1720 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1723 int running = svcpt->scp_nthrs_running;
1725 if (!nrs_svcpt_has_hp(svcpt))
1731 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1734 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1735 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1736 /* leave just 1 thread for normal RPCs */
1737 running = PTLRPC_NTHRS_INIT;
1738 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1742 if (svcpt->scp_nreqs_active >= running - 1)
1745 if (svcpt->scp_nhreqs_active == 0)
1748 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1749 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1752 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1755 return ptlrpc_server_allow_high(svcpt, force) &&
1756 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1760 * Only allow normal priority requests on a service that has a high-priority
1761 * queue if forced (i.e. cleanup), if there are other high priority requests
1762 * already being processed (i.e. those threads can service more high-priority
1763 * requests), or if there are enough idle threads that a later thread can do
1764 * a high priority request.
1765 * User can call it w/o any lock but need to hold
1766 * ptlrpc_service_part::scp_req_lock to get reliable result
1768 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1771 int running = svcpt->scp_nthrs_running;
1772 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1773 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1774 /* leave just 1 thread for normal RPCs */
1775 running = PTLRPC_NTHRS_INIT;
1776 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1783 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1786 if (svcpt->scp_nreqs_active < running - 2)
1789 if (svcpt->scp_nreqs_active >= running - 1)
1792 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1795 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1798 return ptlrpc_server_allow_normal(svcpt, force) &&
1799 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1803 * Returns true if there are requests available in incoming
1804 * request queue for processing and it is allowed to fetch them.
1805 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1806 * to get reliable result
1807 * \see ptlrpc_server_allow_normal
1808 * \see ptlrpc_server_allow high
1811 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1813 return ptlrpc_server_high_pending(svcpt, force) ||
1814 ptlrpc_server_normal_pending(svcpt, force);
1818 * Fetch a request for processing from queue of unprocessed requests.
1819 * Favors high-priority requests.
1820 * Returns a pointer to fetched request.
1822 static struct ptlrpc_request *
1823 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1825 struct ptlrpc_request *req = NULL;
1828 spin_lock(&svcpt->scp_req_lock);
1830 if (ptlrpc_server_high_pending(svcpt, force)) {
1831 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1833 svcpt->scp_hreq_count++;
1838 if (ptlrpc_server_normal_pending(svcpt, force)) {
1839 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1841 svcpt->scp_hreq_count = 0;
1846 spin_unlock(&svcpt->scp_req_lock);
1850 svcpt->scp_nreqs_active++;
1852 svcpt->scp_nhreqs_active++;
1854 spin_unlock(&svcpt->scp_req_lock);
1856 if (likely(req->rq_export))
1857 class_export_rpc_inc(req->rq_export);
1863 * Handle freshly incoming reqs, add to timed early reply list,
1864 * pass on to regular request queue.
1865 * All incoming requests pass through here before getting into
1866 * ptlrpc_server_handle_req later on.
1869 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1870 struct ptlrpc_thread *thread)
1872 struct ptlrpc_service *svc = svcpt->scp_service;
1873 struct ptlrpc_request *req;
1878 spin_lock(&svcpt->scp_lock);
1879 if (list_empty(&svcpt->scp_req_incoming)) {
1880 spin_unlock(&svcpt->scp_lock);
1884 req = list_entry(svcpt->scp_req_incoming.next,
1885 struct ptlrpc_request, rq_list);
1886 list_del_init(&req->rq_list);
1887 svcpt->scp_nreqs_incoming--;
1888 /* Consider this still a "queued" request as far as stats are
1890 spin_unlock(&svcpt->scp_lock);
1892 /* go through security check/transform */
1893 rc = sptlrpc_svc_unwrap_request(req);
1897 case SECSVC_COMPLETE:
1898 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1907 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1908 * redo it wouldn't be harmful.
1910 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1911 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1913 CERROR("error unpacking request: ptl %d from %s "
1914 "x%llu\n", svc->srv_req_portal,
1915 libcfs_id2str(req->rq_peer), req->rq_xid);
1920 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1922 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1923 "%llu\n", svc->srv_req_portal,
1924 libcfs_id2str(req->rq_peer), req->rq_xid);
1928 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1929 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1930 CERROR("drop incoming rpc opc %u, x%llu\n",
1931 cfs_fail_val, req->rq_xid);
1936 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1937 CERROR("wrong packet type received (type=%u) from %s\n",
1938 lustre_msg_get_type(req->rq_reqmsg),
1939 libcfs_id2str(req->rq_peer));
1943 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1947 req->rq_bulk_write = 1;
1951 case MGS_CONFIG_READ:
1952 req->rq_bulk_read = 1;
1956 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1958 req->rq_export = class_conn2export(
1959 lustre_msg_get_handle(req->rq_reqmsg));
1960 if (req->rq_export) {
1961 rc = ptlrpc_check_req(req);
1963 rc = sptlrpc_target_export_check(req->rq_export, req);
1965 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1966 "illegal security flavor,");
1971 ptlrpc_update_export_timer(req->rq_export, 0);
1974 /* req_in handling should/must be fast */
1975 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1976 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
1977 (s64)(ktime_get_real_seconds() -
1978 req->rq_arrival_time.tv_sec));
1980 /* Set rpc server deadline and add it to the timed list */
1981 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1982 MSGHDR_AT_SUPPORT) ?
1983 /* The max time the client expects us to take */
1984 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1986 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1987 if (unlikely(deadline == 0)) {
1988 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1992 /* Skip early reply */
1993 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1994 req->rq_deadline += obd_timeout;
1996 req->rq_svc_thread = thread;
1997 if (thread != NULL) {
1998 /* initialize request session, it is needed for request
1999 * processing by target */
2000 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2003 CERROR("%s: failure to initialize session: rc = %d\n",
2004 thread->t_name, rc);
2007 req->rq_session.lc_thread = thread;
2008 lu_context_enter(&req->rq_session);
2009 thread->t_env->le_ses = &req->rq_session;
2012 ptlrpc_at_add_timed(req);
2014 /* Move it over to the request processing queue */
2015 rc = ptlrpc_server_request_add(svcpt, req);
2019 wake_up(&svcpt->scp_waitq);
2023 ptlrpc_server_finish_request(svcpt, req);
2029 * Main incoming request handling logic.
2030 * Calls handler function from service to do actual processing.
2033 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2034 struct ptlrpc_thread *thread)
2036 struct ptlrpc_service *svc = svcpt->scp_service;
2037 struct ptlrpc_request *request;
2047 request = ptlrpc_server_request_get(svcpt, false);
2048 if (request == NULL)
2051 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2052 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2053 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2054 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2056 if (unlikely(fail_opc)) {
2057 if (request->rq_export && request->rq_ops)
2058 OBD_FAIL_TIMEOUT(fail_opc, 4);
2061 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2063 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2064 libcfs_debug_dumplog();
2066 work_start = ktime_get_real();
2067 arrived = timespec64_to_ktime(request->rq_arrival_time);
2068 timediff_usecs = ktime_us_delta(work_start, arrived);
2069 if (likely(svc->srv_stats != NULL)) {
2070 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2072 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2073 svcpt->scp_nreqs_incoming);
2074 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2075 svcpt->scp_nreqs_active);
2076 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2077 at_get(&svcpt->scp_at_estimate));
2080 if (likely(request->rq_export)) {
2081 if (unlikely(ptlrpc_check_req(request)))
2083 ptlrpc_update_export_timer(request->rq_export,
2084 timediff_usecs / (USEC_PER_SEC / 2));
2087 /* Discard requests queued for longer than the deadline.
2088 The deadline is increased if we send an early reply. */
2089 if (ktime_get_real_seconds() > request->rq_deadline) {
2090 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2091 libcfs_id2str(request->rq_peer),
2092 request->rq_deadline -
2093 request->rq_arrival_time.tv_sec,
2094 ktime_get_real_seconds() - request->rq_deadline);
2098 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2099 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2100 (request->rq_export ?
2101 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2102 (request->rq_export ?
2103 atomic_read(&request->rq_export->exp_refcount) : -99),
2104 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2105 libcfs_id2str(request->rq_peer),
2106 lustre_msg_get_opc(request->rq_reqmsg));
2108 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2109 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2111 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2113 /* re-assign request and sesson thread to the current one */
2114 request->rq_svc_thread = thread;
2115 if (thread != NULL) {
2116 LASSERT(request->rq_session.lc_thread == NULL);
2117 request->rq_session.lc_thread = thread;
2118 thread->t_env->le_ses = &request->rq_session;
2120 svc->srv_ops.so_req_handler(request);
2122 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2125 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2126 DEBUG_REQ(D_WARNING, request,
2127 "Request took longer than estimated (%lld:%llds); "
2128 "client may timeout.",
2129 request->rq_deadline -
2130 request->rq_arrival_time.tv_sec,
2131 ktime_get_real_seconds() - request->rq_deadline);
2134 work_end = ktime_get_real();
2135 timediff_usecs = ktime_us_delta(work_end, work_start);
2136 arrived_usecs = ktime_us_delta(work_end, arrived);
2137 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2138 "%s:%s+%d:%d:x%llu:%s:%d Request processed in %lldus "
2139 "(%lldus total) trans %llu rc %d/%d\n",
2141 (request->rq_export ?
2142 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2143 (request->rq_export ?
2144 atomic_read(&request->rq_export->exp_refcount) : -99),
2145 lustre_msg_get_status(request->rq_reqmsg),
2147 libcfs_id2str(request->rq_peer),
2148 lustre_msg_get_opc(request->rq_reqmsg),
2151 (request->rq_repmsg ?
2152 lustre_msg_get_transno(request->rq_repmsg) :
2153 request->rq_transno),
2155 (request->rq_repmsg ?
2156 lustre_msg_get_status(request->rq_repmsg) : -999));
2157 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2158 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2159 int opc = opcode_offset(op);
2160 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2161 LASSERT(opc < LUSTRE_MAX_OPCODES);
2162 lprocfs_counter_add(svc->srv_stats,
2163 opc + EXTRA_MAX_OPCODES,
2167 if (unlikely(request->rq_early_count)) {
2168 DEBUG_REQ(D_ADAPTTO, request,
2169 "sent %d early replies before finishing in %llds",
2170 request->rq_early_count,
2171 arrived_usecs / USEC_PER_SEC);
2174 ptlrpc_server_finish_active_request(svcpt, request);
2180 * An internal function to process a single reply state object.
2183 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2185 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2186 struct ptlrpc_service *svc = svcpt->scp_service;
2187 struct obd_export *exp;
2192 exp = rs->rs_export;
2194 LASSERT(rs->rs_difficult);
2195 LASSERT(rs->rs_scheduled);
2196 LASSERT(list_empty(&rs->rs_list));
2198 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2199 * iterates over newly committed replies, removing them from
2200 * exp_uncommitted_replies. It then drops this lock and schedules the
2201 * replies it found for handling here.
2203 * We can avoid contention for exp_uncommitted_replies_lock between the
2204 * HRT threads and further commit callbacks by checking rs_committed
2205 * which is set in the commit callback while it holds both
2206 * rs_lock and exp_uncommitted_reples.
2208 * If we see rs_committed clear, the commit callback _may_ not have
2209 * handled this reply yet and we race with it to grab
2210 * exp_uncommitted_replies_lock before removing the reply from
2211 * exp_uncommitted_replies. Note that if we lose the race and the
2212 * reply has already been removed, list_del_init() is a noop.
2214 * If we see rs_committed set, we know the commit callback is handling,
2215 * or has handled this reply since store reordering might allow us to
2216 * see rs_committed set out of sequence. But since this is done
2217 * holding rs_lock, we can be sure it has all completed once we hold
2218 * rs_lock, which we do right next.
2220 if (!rs->rs_committed) {
2221 /* if rs was commited, no need to convert locks, don't check
2222 * rs_committed here because rs may never be added into
2223 * exp_uncommitted_replies and this flag never be set, see
2224 * target_send_reply() */
2225 if (rs->rs_convert_lock &&
2226 rs->rs_transno > exp->exp_last_committed) {
2227 struct ldlm_lock *lock;
2228 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2230 spin_lock(&rs->rs_lock);
2231 if (rs->rs_convert_lock &&
2232 rs->rs_transno > exp->exp_last_committed) {
2233 nlocks = rs->rs_nlocks;
2234 while (nlocks-- > 0) {
2236 * NB don't assume rs is always handled
2237 * by the same service thread (see
2238 * ptlrpc_hr_select, so REP-ACK hr may
2239 * race with trans commit, while the
2240 * latter will release locks, get locks
2241 * here early to convert to COS mode
2244 lock = ldlm_handle2lock(
2245 &rs->rs_locks[nlocks]);
2247 ack_locks[nlocks] = lock;
2248 rs->rs_modes[nlocks] = LCK_COS;
2250 nlocks = rs->rs_nlocks;
2251 rs->rs_convert_lock = 0;
2252 /* clear rs_scheduled so that commit callback
2253 * can schedule again */
2254 rs->rs_scheduled = 0;
2255 spin_unlock(&rs->rs_lock);
2257 while (nlocks-- > 0) {
2258 lock = ack_locks[nlocks];
2259 ldlm_lock_mode_downgrade(lock, LCK_COS);
2260 LDLM_LOCK_PUT(lock);
2264 spin_unlock(&rs->rs_lock);
2267 spin_lock(&exp->exp_uncommitted_replies_lock);
2268 list_del_init(&rs->rs_obd_list);
2269 spin_unlock(&exp->exp_uncommitted_replies_lock);
2272 spin_lock(&exp->exp_lock);
2273 /* Noop if removed already */
2274 list_del_init(&rs->rs_exp_list);
2275 spin_unlock(&exp->exp_lock);
2277 spin_lock(&rs->rs_lock);
2279 been_handled = rs->rs_handled;
2282 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2283 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2285 if (nlocks == 0 && !been_handled) {
2286 /* If we see this, we should already have seen the warning
2287 * in mds_steal_ack_locks() */
2288 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2291 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2292 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2295 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2296 spin_unlock(&rs->rs_lock);
2298 if (!been_handled && rs->rs_on_net) {
2299 LNetMDUnlink(rs->rs_md_h);
2300 /* Ignore return code; we're racing with completion */
2303 while (nlocks-- > 0)
2304 ldlm_lock_decref(&rs->rs_locks[nlocks],
2305 rs->rs_modes[nlocks]);
2307 spin_lock(&rs->rs_lock);
2310 rs->rs_scheduled = 0;
2311 rs->rs_convert_lock = 0;
2313 if (!rs->rs_on_net) {
2315 spin_unlock(&rs->rs_lock);
2317 class_export_put (exp);
2318 rs->rs_export = NULL;
2319 ptlrpc_rs_decref(rs);
2320 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2321 svc->srv_is_stopping)
2322 wake_up_all(&svcpt->scp_waitq);
2326 /* still on the net; callback will schedule */
2327 spin_unlock(&rs->rs_lock);
2333 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2335 int avail = svcpt->scp_nrqbds_posted;
2336 int low_water = test_req_buffer_pressure ? 0 :
2337 svcpt->scp_service->srv_nbuf_per_group / 2;
2339 /* NB I'm not locking; just looking. */
2341 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2342 * allowed the request history to grow out of control. We could put a
2343 * sanity check on that here and cull some history if we need the
2346 if (avail <= low_water)
2347 ptlrpc_grow_req_bufs(svcpt, 1);
2349 if (svcpt->scp_service->srv_stats) {
2350 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2351 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2356 ptlrpc_retry_rqbds(void *arg)
2358 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2360 svcpt->scp_rqbd_timeout = 0;
2365 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2367 return svcpt->scp_nreqs_active <
2368 svcpt->scp_nthrs_running - 1 -
2369 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2373 * allowed to create more threads
2374 * user can call it w/o any lock but need to hold
2375 * ptlrpc_service_part::scp_lock to get reliable result
2378 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2380 return svcpt->scp_nthrs_running +
2381 svcpt->scp_nthrs_starting <
2382 svcpt->scp_service->srv_nthrs_cpt_limit;
2386 * too many requests and allowed to create more threads
2389 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2391 return !ptlrpc_threads_enough(svcpt) &&
2392 ptlrpc_threads_increasable(svcpt);
2396 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2398 return thread_is_stopping(thread) ||
2399 thread->t_svcpt->scp_service->srv_is_stopping;
2403 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2405 return !list_empty(&svcpt->scp_rqbd_idle) &&
2406 svcpt->scp_rqbd_timeout == 0;
2410 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2412 return svcpt->scp_at_check;
2416 * requests wait on preprocessing
2417 * user can call it w/o any lock but need to hold
2418 * ptlrpc_service_part::scp_lock to get reliable result
2421 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2423 return !list_empty(&svcpt->scp_req_incoming);
2426 static __attribute__((__noinline__)) int
2427 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2428 struct ptlrpc_thread *thread)
2430 /* Don't exit while there are replies to be handled */
2431 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2432 ptlrpc_retry_rqbds, svcpt);
2434 lc_watchdog_disable(thread->t_watchdog);
2438 l_wait_event_exclusive_head(svcpt->scp_waitq,
2439 ptlrpc_thread_stopping(thread) ||
2440 ptlrpc_server_request_incoming(svcpt) ||
2441 ptlrpc_server_request_pending(svcpt, false) ||
2442 ptlrpc_rqbd_pending(svcpt) ||
2443 ptlrpc_at_check(svcpt), &lwi);
2445 if (ptlrpc_thread_stopping(thread))
2448 lc_watchdog_touch(thread->t_watchdog,
2449 ptlrpc_server_get_timeout(svcpt));
2454 * Main thread body for service threads.
2455 * Waits in a loop waiting for new requests to process to appear.
2456 * Every time an incoming requests is added to its queue, a waitq
2457 * is woken up and one of the threads will handle it.
2459 static int ptlrpc_main(void *arg)
2461 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2462 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2463 struct ptlrpc_service *svc = svcpt->scp_service;
2464 struct ptlrpc_reply_state *rs;
2465 struct group_info *ginfo = NULL;
2467 int counter = 0, rc = 0;
2470 thread->t_pid = current_pid();
2471 unshare_fs_struct();
2473 /* NB: we will call cfs_cpt_bind() for all threads, because we
2474 * might want to run lustre server only on a subset of system CPUs,
2475 * in that case ->scp_cpt is CFS_CPT_ANY */
2476 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2478 CWARN("%s: failed to bind %s on CPT %d\n",
2479 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2482 ginfo = groups_alloc(0);
2488 set_current_groups(ginfo);
2489 put_group_info(ginfo);
2491 if (svc->srv_ops.so_thr_init != NULL) {
2492 rc = svc->srv_ops.so_thr_init(thread);
2503 rc = lu_context_init(&env->le_ctx,
2504 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2508 thread->t_env = env;
2509 env->le_ctx.lc_thread = thread;
2510 env->le_ctx.lc_cookie = 0x6;
2512 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2513 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2517 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2518 svc->srv_name, svcpt->scp_cpt, rc);
2522 /* Alloc reply state structure for this one */
2523 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2529 spin_lock(&svcpt->scp_lock);
2531 LASSERT(thread_is_starting(thread));
2532 thread_clear_flags(thread, SVC_STARTING);
2534 LASSERT(svcpt->scp_nthrs_starting == 1);
2535 svcpt->scp_nthrs_starting--;
2537 /* SVC_STOPPING may already be set here if someone else is trying
2538 * to stop the service while this new thread has been dynamically
2539 * forked. We still set SVC_RUNNING to let our creator know that
2540 * we are now running, however we will exit as soon as possible */
2541 thread_add_flags(thread, SVC_RUNNING);
2542 svcpt->scp_nthrs_running++;
2543 spin_unlock(&svcpt->scp_lock);
2545 /* wake up our creator in case he's still waiting. */
2546 wake_up(&thread->t_ctl_waitq);
2548 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2551 spin_lock(&svcpt->scp_rep_lock);
2552 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2553 wake_up(&svcpt->scp_rep_waitq);
2554 spin_unlock(&svcpt->scp_rep_lock);
2556 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2557 svcpt->scp_nthrs_running);
2559 /* XXX maintain a list of all managed devices: insert here */
2560 while (!ptlrpc_thread_stopping(thread)) {
2561 if (ptlrpc_wait_event(svcpt, thread))
2564 ptlrpc_check_rqbd_pool(svcpt);
2566 if (ptlrpc_threads_need_create(svcpt)) {
2567 /* Ignore return code - we tried... */
2568 ptlrpc_start_thread(svcpt, 0);
2571 /* reset le_ses to initial state */
2573 /* Process all incoming reqs before handling any */
2574 if (ptlrpc_server_request_incoming(svcpt)) {
2575 lu_context_enter(&env->le_ctx);
2576 ptlrpc_server_handle_req_in(svcpt, thread);
2577 lu_context_exit(&env->le_ctx);
2579 /* but limit ourselves in case of flood */
2580 if (counter++ < 100)
2585 if (ptlrpc_at_check(svcpt))
2586 ptlrpc_at_check_timed(svcpt);
2588 if (ptlrpc_server_request_pending(svcpt, false)) {
2589 lu_context_enter(&env->le_ctx);
2590 ptlrpc_server_handle_request(svcpt, thread);
2591 lu_context_exit(&env->le_ctx);
2594 if (ptlrpc_rqbd_pending(svcpt) &&
2595 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2596 /* I just failed to repost request buffers.
2597 * Wait for a timeout (unless something else
2598 * happens) before I try again */
2599 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2600 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2601 svcpt->scp_nrqbds_posted);
2605 lc_watchdog_delete(thread->t_watchdog);
2606 thread->t_watchdog = NULL;
2610 * deconstruct service specific state created by ptlrpc_start_thread()
2612 if (svc->srv_ops.so_thr_done != NULL)
2613 svc->srv_ops.so_thr_done(thread);
2616 lu_context_fini(&env->le_ctx);
2620 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2621 thread, thread->t_pid, thread->t_id, rc);
2623 spin_lock(&svcpt->scp_lock);
2624 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2625 svcpt->scp_nthrs_starting--;
2627 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2628 /* must know immediately */
2629 svcpt->scp_nthrs_running--;
2633 thread_add_flags(thread, SVC_STOPPED);
2635 wake_up(&thread->t_ctl_waitq);
2636 spin_unlock(&svcpt->scp_lock);
2641 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2642 struct list_head *replies)
2646 spin_lock(&hrt->hrt_lock);
2648 list_splice_init(&hrt->hrt_queue, replies);
2649 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2651 spin_unlock(&hrt->hrt_lock);
2656 * Main body of "handle reply" function.
2657 * It processes acked reply states
2659 static int ptlrpc_hr_main(void *arg)
2661 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2662 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2663 struct list_head replies;
2666 INIT_LIST_HEAD(&replies);
2667 unshare_fs_struct();
2669 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2671 char threadname[20];
2673 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2674 hrp->hrp_cpt, hrt->hrt_id);
2675 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2676 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2679 atomic_inc(&hrp->hrp_nstarted);
2680 wake_up(&ptlrpc_hr.hr_waitq);
2682 while (!ptlrpc_hr.hr_stopping) {
2683 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2685 while (!list_empty(&replies)) {
2686 struct ptlrpc_reply_state *rs;
2688 rs = list_entry(replies.prev,
2689 struct ptlrpc_reply_state,
2691 list_del_init(&rs->rs_list);
2692 ptlrpc_handle_rs(rs);
2696 atomic_inc(&hrp->hrp_nstopped);
2697 wake_up(&ptlrpc_hr.hr_waitq);
2702 static void ptlrpc_stop_hr_threads(void)
2704 struct ptlrpc_hr_partition *hrp;
2708 ptlrpc_hr.hr_stopping = 1;
2710 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2711 if (hrp->hrp_thrs == NULL)
2712 continue; /* uninitialized */
2713 for (j = 0; j < hrp->hrp_nthrs; j++)
2714 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2717 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2718 if (hrp->hrp_thrs == NULL)
2719 continue; /* uninitialized */
2720 wait_event(ptlrpc_hr.hr_waitq,
2721 atomic_read(&hrp->hrp_nstopped) ==
2722 atomic_read(&hrp->hrp_nstarted));
2726 static int ptlrpc_start_hr_threads(void)
2728 struct ptlrpc_hr_partition *hrp;
2733 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2736 for (j = 0; j < hrp->hrp_nthrs; j++) {
2737 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2738 struct task_struct *task;
2740 task = kthread_run(ptlrpc_hr_main,
2742 "ptlrpc_hr%02d_%03d",
2751 wait_event(ptlrpc_hr.hr_waitq,
2752 atomic_read(&hrp->hrp_nstarted) == j);
2755 CERROR("cannot start reply handler thread %d:%d: "
2756 "rc = %d\n", i, j, rc);
2757 ptlrpc_stop_hr_threads();
2765 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2767 struct l_wait_info lwi = { 0 };
2768 struct ptlrpc_thread *thread;
2769 struct list_head zombie;
2773 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2774 svcpt->scp_service->srv_name);
2776 INIT_LIST_HEAD(&zombie);
2777 spin_lock(&svcpt->scp_lock);
2778 /* let the thread know that we would like it to stop asap */
2779 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2780 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2781 svcpt->scp_service->srv_thread_name, thread->t_id);
2782 thread_add_flags(thread, SVC_STOPPING);
2785 wake_up_all(&svcpt->scp_waitq);
2787 while (!list_empty(&svcpt->scp_threads)) {
2788 thread = list_entry(svcpt->scp_threads.next,
2789 struct ptlrpc_thread, t_link);
2790 if (thread_is_stopped(thread)) {
2791 list_del(&thread->t_link);
2792 list_add(&thread->t_link, &zombie);
2795 spin_unlock(&svcpt->scp_lock);
2797 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2798 svcpt->scp_service->srv_thread_name, thread->t_id);
2799 l_wait_event(thread->t_ctl_waitq,
2800 thread_is_stopped(thread), &lwi);
2802 spin_lock(&svcpt->scp_lock);
2805 spin_unlock(&svcpt->scp_lock);
2807 while (!list_empty(&zombie)) {
2808 thread = list_entry(zombie.next,
2809 struct ptlrpc_thread, t_link);
2810 list_del(&thread->t_link);
2811 OBD_FREE_PTR(thread);
2817 * Stops all threads of a particular service \a svc
2819 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2821 struct ptlrpc_service_part *svcpt;
2825 ptlrpc_service_for_each_part(svcpt, i, svc) {
2826 if (svcpt->scp_service != NULL)
2827 ptlrpc_svcpt_stop_threads(svcpt);
2833 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2840 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2841 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2843 for (i = 0; i < svc->srv_ncpts; i++) {
2844 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2845 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2851 /* We have enough threads, don't start more. b=15759 */
2858 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2859 svc->srv_thread_name, i, j, rc);
2860 ptlrpc_stop_all_threads(svc);
2864 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2866 struct l_wait_info lwi = { 0 };
2867 struct ptlrpc_thread *thread;
2868 struct ptlrpc_service *svc;
2869 struct task_struct *task;
2873 LASSERT(svcpt != NULL);
2875 svc = svcpt->scp_service;
2877 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2878 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2879 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2882 if (unlikely(svc->srv_is_stopping))
2885 if (!ptlrpc_threads_increasable(svcpt) ||
2886 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2887 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2890 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2893 init_waitqueue_head(&thread->t_ctl_waitq);
2895 spin_lock(&svcpt->scp_lock);
2896 if (!ptlrpc_threads_increasable(svcpt)) {
2897 spin_unlock(&svcpt->scp_lock);
2898 OBD_FREE_PTR(thread);
2902 if (svcpt->scp_nthrs_starting != 0) {
2903 /* serialize starting because some modules (obdfilter)
2904 * might require unique and contiguous t_id */
2905 LASSERT(svcpt->scp_nthrs_starting == 1);
2906 spin_unlock(&svcpt->scp_lock);
2907 OBD_FREE_PTR(thread);
2909 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2910 svc->srv_thread_name, svcpt->scp_thr_nextid);
2915 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2916 svc->srv_thread_name, svcpt->scp_thr_nextid);
2920 svcpt->scp_nthrs_starting++;
2921 thread->t_id = svcpt->scp_thr_nextid++;
2922 thread_add_flags(thread, SVC_STARTING);
2923 thread->t_svcpt = svcpt;
2925 list_add(&thread->t_link, &svcpt->scp_threads);
2926 spin_unlock(&svcpt->scp_lock);
2928 if (svcpt->scp_cpt >= 0) {
2929 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2930 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2932 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2933 svc->srv_thread_name, thread->t_id);
2936 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2937 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2940 CERROR("cannot start thread '%s': rc = %d\n",
2941 thread->t_name, rc);
2942 spin_lock(&svcpt->scp_lock);
2943 --svcpt->scp_nthrs_starting;
2944 if (thread_is_stopping(thread)) {
2945 /* this ptlrpc_thread is being hanled
2946 * by ptlrpc_svcpt_stop_threads now
2948 thread_add_flags(thread, SVC_STOPPED);
2949 wake_up(&thread->t_ctl_waitq);
2950 spin_unlock(&svcpt->scp_lock);
2952 list_del(&thread->t_link);
2953 spin_unlock(&svcpt->scp_lock);
2954 OBD_FREE_PTR(thread);
2962 l_wait_event(thread->t_ctl_waitq,
2963 thread_is_running(thread) || thread_is_stopped(thread),
2966 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2970 int ptlrpc_hr_init(void)
2972 struct ptlrpc_hr_partition *hrp;
2973 struct ptlrpc_hr_thread *hrt;
2980 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2981 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2983 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2985 if (ptlrpc_hr.hr_partitions == NULL)
2988 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2990 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
2992 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
2995 atomic_set(&hrp->hrp_nstarted, 0);
2996 atomic_set(&hrp->hrp_nstopped, 0);
2998 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
2999 hrp->hrp_nthrs /= weight;
3000 if (hrp->hrp_nthrs == 0)
3003 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3004 hrp->hrp_nthrs * sizeof(*hrt));
3005 if (hrp->hrp_thrs == NULL)
3006 GOTO(out, rc = -ENOMEM);
3008 for (i = 0; i < hrp->hrp_nthrs; i++) {
3009 hrt = &hrp->hrp_thrs[i];
3012 hrt->hrt_partition = hrp;
3013 init_waitqueue_head(&hrt->hrt_waitq);
3014 spin_lock_init(&hrt->hrt_lock);
3015 INIT_LIST_HEAD(&hrt->hrt_queue);
3019 rc = ptlrpc_start_hr_threads();
3026 void ptlrpc_hr_fini(void)
3028 struct ptlrpc_hr_partition *hrp;
3031 if (ptlrpc_hr.hr_partitions == NULL)
3034 ptlrpc_stop_hr_threads();
3036 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3037 if (hrp->hrp_thrs != NULL) {
3038 OBD_FREE(hrp->hrp_thrs,
3039 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3043 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3044 ptlrpc_hr.hr_partitions = NULL;
3049 * Wait until all already scheduled replies are processed.
3051 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3055 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3058 rc = l_wait_event(svcpt->scp_waitq,
3059 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3062 CWARN("Unexpectedly long timeout %s %p\n",
3063 svcpt->scp_service->srv_name, svcpt->scp_service);
3068 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3070 struct ptlrpc_service_part *svcpt;
3073 /* early disarm AT timer... */
3074 ptlrpc_service_for_each_part(svcpt, i, svc) {
3075 if (svcpt->scp_service != NULL)
3076 del_timer(&svcpt->scp_at_timer);
3081 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3083 struct ptlrpc_service_part *svcpt;
3084 struct ptlrpc_request_buffer_desc *rqbd;
3085 struct l_wait_info lwi;
3089 /* All history will be culled when the next request buffer is
3090 * freed in ptlrpc_service_purge_all() */
3091 svc->srv_hist_nrqbds_cpt_max = 0;
3093 rc = LNetClearLazyPortal(svc->srv_req_portal);
3096 ptlrpc_service_for_each_part(svcpt, i, svc) {
3097 if (svcpt->scp_service == NULL)
3100 /* Unlink all the request buffers. This forces a 'final'
3101 * event with its 'unlink' flag set for each posted rqbd */
3102 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3104 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3105 LASSERT(rc == 0 || rc == -ENOENT);
3109 ptlrpc_service_for_each_part(svcpt, i, svc) {
3110 if (svcpt->scp_service == NULL)
3113 /* Wait for the network to release any buffers
3114 * it's currently filling */
3115 spin_lock(&svcpt->scp_lock);
3116 while (svcpt->scp_nrqbds_posted != 0) {
3117 spin_unlock(&svcpt->scp_lock);
3118 /* Network access will complete in finite time but
3119 * the HUGE timeout lets us CWARN for visibility
3120 * of sluggish NALs */
3121 lwi = LWI_TIMEOUT_INTERVAL(
3122 cfs_time_seconds(LONG_UNLINK),
3123 cfs_time_seconds(1), NULL, NULL);
3124 rc = l_wait_event(svcpt->scp_waitq,
3125 svcpt->scp_nrqbds_posted == 0, &lwi);
3126 if (rc == -ETIMEDOUT) {
3127 CWARN("Service %s waiting for "
3128 "request buffers\n",
3129 svcpt->scp_service->srv_name);
3131 spin_lock(&svcpt->scp_lock);
3133 spin_unlock(&svcpt->scp_lock);
3138 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3140 struct ptlrpc_service_part *svcpt;
3141 struct ptlrpc_request_buffer_desc *rqbd;
3142 struct ptlrpc_request *req;
3143 struct ptlrpc_reply_state *rs;
3146 ptlrpc_service_for_each_part(svcpt, i, svc) {
3147 if (svcpt->scp_service == NULL)
3150 spin_lock(&svcpt->scp_rep_lock);
3151 while (!list_empty(&svcpt->scp_rep_active)) {
3152 rs = list_entry(svcpt->scp_rep_active.next,
3153 struct ptlrpc_reply_state, rs_list);
3154 spin_lock(&rs->rs_lock);
3155 ptlrpc_schedule_difficult_reply(rs);
3156 spin_unlock(&rs->rs_lock);
3158 spin_unlock(&svcpt->scp_rep_lock);
3160 /* purge the request queue. NB No new replies (rqbds
3161 * all unlinked) and no service threads, so I'm the only
3162 * thread noodling the request queue now */
3163 while (!list_empty(&svcpt->scp_req_incoming)) {
3164 req = list_entry(svcpt->scp_req_incoming.next,
3165 struct ptlrpc_request, rq_list);
3167 list_del(&req->rq_list);
3168 svcpt->scp_nreqs_incoming--;
3169 ptlrpc_server_finish_request(svcpt, req);
3172 while (ptlrpc_server_request_pending(svcpt, true)) {
3173 req = ptlrpc_server_request_get(svcpt, true);
3174 ptlrpc_server_finish_active_request(svcpt, req);
3177 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3178 LASSERT(svcpt->scp_nreqs_incoming == 0);
3179 LASSERT(svcpt->scp_nreqs_active == 0);
3180 /* history should have been culled by
3181 * ptlrpc_server_finish_request */
3182 LASSERT(svcpt->scp_hist_nrqbds == 0);
3184 /* Now free all the request buffers since nothing
3185 * references them any more... */
3187 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3188 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3189 struct ptlrpc_request_buffer_desc,
3191 ptlrpc_free_rqbd(rqbd);
3193 ptlrpc_wait_replies(svcpt);
3195 while (!list_empty(&svcpt->scp_rep_idle)) {
3196 rs = list_entry(svcpt->scp_rep_idle.next,
3197 struct ptlrpc_reply_state,
3199 list_del(&rs->rs_list);
3200 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3206 ptlrpc_service_free(struct ptlrpc_service *svc)
3208 struct ptlrpc_service_part *svcpt;
3209 struct ptlrpc_at_array *array;
3212 ptlrpc_service_for_each_part(svcpt, i, svc) {
3213 if (svcpt->scp_service == NULL)
3216 /* In case somebody rearmed this in the meantime */
3217 del_timer(&svcpt->scp_at_timer);
3218 array = &svcpt->scp_at_array;
3220 if (array->paa_reqs_array != NULL) {
3221 OBD_FREE(array->paa_reqs_array,
3222 sizeof(struct list_head) * array->paa_size);
3223 array->paa_reqs_array = NULL;
3226 if (array->paa_reqs_count != NULL) {
3227 OBD_FREE(array->paa_reqs_count,
3228 sizeof(__u32) * array->paa_size);
3229 array->paa_reqs_count = NULL;
3233 ptlrpc_service_for_each_part(svcpt, i, svc)
3234 OBD_FREE_PTR(svcpt);
3236 if (svc->srv_cpts != NULL)
3237 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3239 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3240 srv_parts[svc->srv_ncpts]));
3243 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3247 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3249 service->srv_is_stopping = 1;
3251 mutex_lock(&ptlrpc_all_services_mutex);
3252 list_del_init(&service->srv_list);
3253 mutex_unlock(&ptlrpc_all_services_mutex);
3255 ptlrpc_service_del_atimer(service);
3256 ptlrpc_stop_all_threads(service);
3258 ptlrpc_service_unlink_rqbd(service);
3259 ptlrpc_service_purge_all(service);
3260 ptlrpc_service_nrs_cleanup(service);
3262 ptlrpc_lprocfs_unregister_service(service);
3263 ptlrpc_sysfs_unregister_service(service);
3265 ptlrpc_service_free(service);
3269 EXPORT_SYMBOL(ptlrpc_unregister_service);
3272 * Returns 0 if the service is healthy.
3274 * Right now, it just checks to make sure that requests aren't languishing
3275 * in the queue. We'll use this health check to govern whether a node needs
3276 * to be shot, so it's intentionally non-aggressive. */
3277 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3279 struct ptlrpc_request *request = NULL;
3280 struct timespec64 right_now;
3281 struct timespec64 timediff;
3283 ktime_get_real_ts64(&right_now);
3285 spin_lock(&svcpt->scp_req_lock);
3286 /* How long has the next entry been waiting? */
3287 if (ptlrpc_server_high_pending(svcpt, true))
3288 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3289 else if (ptlrpc_server_normal_pending(svcpt, true))
3290 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3292 if (request == NULL) {
3293 spin_unlock(&svcpt->scp_req_lock);
3297 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3298 spin_unlock(&svcpt->scp_req_lock);
3300 if ((timediff.tv_sec) >
3301 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3302 CERROR("%s: unhealthy - request has been waiting %llds\n",
3303 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3311 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3313 struct ptlrpc_service_part *svcpt;
3319 ptlrpc_service_for_each_part(svcpt, i, svc) {
3320 int rc = ptlrpc_svcpt_health_check(svcpt);
3327 EXPORT_SYMBOL(ptlrpc_service_health_check);