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) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/ptlrpc/ptlrpcd.c
34 /** \defgroup ptlrpcd PortalRPC daemon
36 * ptlrpcd is a special thread with its own set where other user might add
37 * requests when they don't want to wait for their completion.
38 * PtlRPCD will take care of sending such requests and then processing their
39 * replies and calling completion callbacks as necessary.
40 * The callbacks are called directly from ptlrpcd context.
41 * It is important to never significantly block (esp. on RPCs!) within such
42 * completion handler or a deadlock might occur where ptlrpcd enters some
43 * callback that attempts to send another RPC and wait for it to return,
44 * during which time ptlrpcd is completely blocked, so e.g. if import
45 * fails, recovery cannot progress because connection requests are also
51 #define DEBUG_SUBSYSTEM S_RPC
53 #include <linux/kthread.h>
54 #include <libcfs/libcfs.h>
55 #include <lustre_net.h>
56 #include <lustre_lib.h>
57 #include <lustre_ha.h>
58 #include <obd_class.h> /* for obd_zombie */
59 #include <obd_support.h> /* for OBD_FAIL_CHECK */
60 #include <cl_object.h> /* cl_env_{get,put}() */
61 #include <lprocfs_status.h>
63 #include "ptlrpc_internal.h"
65 /* One of these per CPT. */
73 struct ptlrpcd_ctl pd_threads[0];
77 * max_ptlrpcds is obsolete, but retained to ensure that the kernel
78 * module will load on a system where it has been tuned.
79 * A value other than 0 implies it was tuned, in which case the value
80 * is used to derive a setting for ptlrpcd_per_cpt_max.
82 static int max_ptlrpcds;
83 module_param(max_ptlrpcds, int, 0644);
84 MODULE_PARM_DESC(max_ptlrpcds,
85 "Max ptlrpcd thread count to be started (obsolete).");
88 * ptlrpcd_bind_policy is obsolete, but retained to ensure that
89 * the kernel module will load on a system where it has been tuned.
90 * A value other than 0 implies it was tuned, in which case the value
91 * is used to derive a setting for ptlrpcd_partner_group_size.
93 static int ptlrpcd_bind_policy;
94 module_param(ptlrpcd_bind_policy, int, 0644);
95 MODULE_PARM_DESC(ptlrpcd_bind_policy,
96 "Ptlrpcd threads binding mode (obsolete).");
99 * ptlrpcd_per_cpt_max: The maximum number of ptlrpcd threads to run
102 static int ptlrpcd_per_cpt_max;
103 module_param(ptlrpcd_per_cpt_max, int, 0644);
104 MODULE_PARM_DESC(ptlrpcd_per_cpt_max,
105 "Max ptlrpcd thread count to be started per CPT.");
108 * ptlrpcd_partner_group_size: The desired number of threads in each
109 * ptlrpcd partner thread group. Default is 2, corresponding to the
110 * old PDB_POLICY_PAIR. A negative value makes all ptlrpcd threads in
111 * a CPT partners of each other.
113 static int ptlrpcd_partner_group_size;
114 module_param(ptlrpcd_partner_group_size, int, 0644);
115 MODULE_PARM_DESC(ptlrpcd_partner_group_size,
116 "Number of ptlrpcd threads in a partner group.");
119 * ptlrpcd_cpts: A CPT string describing the CPU partitions that
120 * ptlrpcd threads should run on. Used to make ptlrpcd threads run on
121 * a subset of all CPTs.
125 * run ptlrpcd threads only on CPT 2.
129 * run ptlrpcd threads on CPTs 0, 1, 2, and 3.
131 * ptlrpcd_cpts=[0-3,5,7]
132 * run ptlrpcd threads on CPTS 0, 1, 2, 3, 5, and 7.
134 static char *ptlrpcd_cpts;
135 module_param(ptlrpcd_cpts, charp, 0644);
136 MODULE_PARM_DESC(ptlrpcd_cpts,
137 "CPU partitions ptlrpcd threads should run in");
139 /* ptlrpcds_cpt_idx maps cpt numbers to an index in the ptlrpcds array. */
140 static int *ptlrpcds_cpt_idx;
142 /* ptlrpcds_num is the number of entries in the ptlrpcds array. */
143 static int ptlrpcds_num;
144 static struct ptlrpcd **ptlrpcds;
147 * In addition to the regular thread pool above, there is a single
148 * global recovery thread. Recovery isn't critical for performance,
149 * and doesn't block, but must always be able to proceed, and it is
150 * possible that all normal ptlrpcd threads are blocked. Hence the
151 * need for a dedicated thread.
153 static struct ptlrpcd_ctl ptlrpcd_rcv;
155 struct mutex ptlrpcd_mutex;
156 static int ptlrpcd_users = 0;
158 void ptlrpcd_wake(struct ptlrpc_request *req)
160 struct ptlrpc_request_set *set = req->rq_set;
162 LASSERT(set != NULL);
163 wake_up(&set->set_waitq);
165 EXPORT_SYMBOL(ptlrpcd_wake);
167 static struct ptlrpcd_ctl *
168 ptlrpcd_select_pc(struct ptlrpc_request *req)
174 if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
177 cpt = cfs_cpt_current(cfs_cpt_tab, 1);
178 if (ptlrpcds_cpt_idx == NULL)
181 idx = ptlrpcds_cpt_idx[cpt];
184 /* We do not care whether it is strict load balance. */
186 if (++idx == pd->pd_nthreads)
190 return &pd->pd_threads[idx];
194 * Move all request from an existing request set to the ptlrpcd queue.
195 * All requests from the set must be in phase RQ_PHASE_NEW.
197 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
199 struct list_head *tmp, *pos;
200 struct ptlrpcd_ctl *pc;
201 struct ptlrpc_request_set *new;
204 pc = ptlrpcd_select_pc(NULL);
207 list_for_each_safe(pos, tmp, &set->set_requests) {
208 struct ptlrpc_request *req =
209 list_entry(pos, struct ptlrpc_request,
212 LASSERT(req->rq_phase == RQ_PHASE_NEW);
214 req->rq_queued_time = ktime_get_seconds();
217 spin_lock(&new->set_new_req_lock);
218 list_splice_init(&set->set_requests, &new->set_new_requests);
219 i = atomic_read(&set->set_remaining);
220 count = atomic_add_return(i, &new->set_new_count);
221 atomic_set(&set->set_remaining, 0);
222 spin_unlock(&new->set_new_req_lock);
224 wake_up(&new->set_waitq);
227 * XXX: It maybe unnecessary to wakeup all the partners. But to
228 * guarantee the async RPC can be processed ASAP, we have
229 * no other better choice. It maybe fixed in future.
231 for (i = 0; i < pc->pc_npartners; i++)
232 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
237 * Return transferred RPCs count.
239 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
240 struct ptlrpc_request_set *src)
242 struct ptlrpc_request *req;
245 spin_lock(&src->set_new_req_lock);
246 if (likely(!list_empty(&src->set_new_requests))) {
247 list_for_each_entry(req, &src->set_new_requests, rq_set_chain)
250 list_splice_init(&src->set_new_requests,
252 rc = atomic_read(&src->set_new_count);
253 atomic_add(rc, &des->set_remaining);
254 atomic_set(&src->set_new_count, 0);
256 spin_unlock(&src->set_new_req_lock);
261 * Requests that are added to the ptlrpcd queue are sent via
262 * ptlrpcd_check->ptlrpc_check_set().
264 void ptlrpcd_add_req(struct ptlrpc_request *req)
266 struct ptlrpcd_ctl *pc;
269 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
271 spin_lock(&req->rq_lock);
272 if (req->rq_invalid_rqset) {
273 req->rq_invalid_rqset = 0;
274 spin_unlock(&req->rq_lock);
275 if (wait_event_idle_timeout(req->rq_set_waitq,
277 cfs_time_seconds(5)) == 0)
278 l_wait_event_abortable(req->rq_set_waitq,
279 req->rq_set == NULL);
280 } else if (req->rq_set) {
282 * If we have a vaid "rq_set", just reuse it to avoid double
285 LASSERT(req->rq_phase == RQ_PHASE_NEW);
286 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
288 /* ptlrpc_check_set will decrease the count */
289 atomic_inc(&req->rq_set->set_remaining);
290 spin_unlock(&req->rq_lock);
291 wake_up(&req->rq_set->set_waitq);
294 spin_unlock(&req->rq_lock);
297 pc = ptlrpcd_select_pc(req);
299 DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s+%d]",
300 req, pc->pc_name, pc->pc_index);
302 ptlrpc_set_add_new_req(pc, req);
304 EXPORT_SYMBOL(ptlrpcd_add_req);
306 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
308 atomic_inc(&set->set_refcount);
312 * Check if there is more work to do on ptlrpcd set.
315 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
317 struct ptlrpc_request *req, *tmp;
318 struct ptlrpc_request_set *set = pc->pc_set;
324 if (atomic_read(&set->set_new_count)) {
325 spin_lock(&set->set_new_req_lock);
326 if (likely(!list_empty(&set->set_new_requests))) {
327 list_splice_init(&set->set_new_requests,
329 atomic_add(atomic_read(&set->set_new_count),
330 &set->set_remaining);
331 atomic_set(&set->set_new_count, 0);
333 * Need to calculate its timeout.
337 spin_unlock(&set->set_new_req_lock);
341 * We should call lu_env_refill() before handling new requests to make
342 * sure that env key the requests depending on really exists.
344 rc2 = lu_env_refill(env);
347 * XXX This is very awkward situation, because
348 * execution can neither continue (request
349 * interpreters assume that env is set up), nor repeat
350 * the loop (as this potentially results in a tight
351 * loop of -ENOMEM's).
353 * Fortunately, refill only ever does something when
354 * new modules are loaded, i.e., early during boot up.
356 CERROR("Failure to refill session: %d\n", rc2);
360 if (atomic_read(&set->set_remaining))
361 rc |= ptlrpc_check_set(env, set);
364 * NB: ptlrpc_check_set has already moved complted request at the
365 * head of seq::set_requests
367 list_for_each_entry_safe(req, tmp, &set->set_requests, rq_set_chain) {
368 if (req->rq_phase != RQ_PHASE_COMPLETE)
371 list_del_init(&req->rq_set_chain);
373 ptlrpc_req_finished(req);
378 * If new requests have been added, make sure to wake up.
380 rc = atomic_read(&set->set_new_count);
383 * If we have nothing to do, check whether we can take some
384 * work from our partner threads.
386 if (rc == 0 && pc->pc_npartners > 0) {
387 struct ptlrpcd_ctl *partner;
388 struct ptlrpc_request_set *ps;
389 int first = pc->pc_cursor;
392 partner = pc->pc_partners[pc->pc_cursor++];
393 if (pc->pc_cursor >= pc->pc_npartners)
398 spin_lock(&partner->pc_lock);
399 ps = partner->pc_set;
401 spin_unlock(&partner->pc_lock);
405 ptlrpc_reqset_get(ps);
406 spin_unlock(&partner->pc_lock);
408 if (atomic_read(&ps->set_new_count)) {
409 rc = ptlrpcd_steal_rqset(set, ps);
412 "transfer %d async RPCs [%d->%d]\n",
413 rc, partner->pc_index,
416 ptlrpc_reqset_put(ps);
417 } while (rc == 0 && pc->pc_cursor != first);
421 RETURN(rc || test_bit(LIOD_STOP, &pc->pc_flags));
425 * Main ptlrpcd thread.
426 * ptlrpc's code paths like to execute in process context, so we have this
427 * thread which spins on a set which contains the rpcs and sends them.
429 static int ptlrpcd(void *arg)
431 struct ptlrpcd_ctl *pc = arg;
432 struct ptlrpc_request_set *set;
433 struct lu_context ses = { 0 };
434 struct lu_env env = { .le_ses = &ses };
439 if (cfs_cpt_bind(cfs_cpt_tab, pc->pc_cpt) != 0)
440 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
443 * Allocate the request set after the thread has been bound
444 * above. This is safe because no requests will be queued
445 * until all ptlrpcd threads have confirmed that they have
446 * successfully started.
448 set = ptlrpc_prep_set();
450 GOTO(failed, rc = -ENOMEM);
451 spin_lock(&pc->pc_lock);
453 spin_unlock(&pc->pc_lock);
455 /* Both client and server (MDT/OST) may use the environment. */
456 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD |
463 rc = lu_context_init(env.le_ses, LCT_SESSION |
467 lu_context_fini(&env.le_ctx);
471 complete(&pc->pc_starting);
474 * This mainloop strongly resembles ptlrpc_set_wait() except that our
475 * set never completes. ptlrpcd_check() calls ptlrpc_check_set() when
476 * there are requests in the set. New requests come in on the set's
477 * new_req_list and ptlrpcd_check() moves them into the set.
480 DEFINE_WAIT_FUNC(wait, woken_wake_function);
483 timeout = ptlrpc_set_next_timeout(set);
485 lu_context_enter(&env.le_ctx);
486 lu_context_enter(env.le_ses);
488 add_wait_queue(&set->set_waitq, &wait);
489 while (!ptlrpcd_check(&env, pc)) {
493 ret = wait_woken(&wait, TASK_IDLE,
494 MAX_SCHEDULE_TIMEOUT);
496 ret = wait_woken(&wait, TASK_IDLE,
497 cfs_time_seconds(timeout));
501 ptlrpc_expired_set(set);
504 remove_wait_queue(&set->set_waitq, &wait);
506 lu_context_exit(&env.le_ctx);
507 lu_context_exit(env.le_ses);
510 * Abort inflight rpcs for forced stop case.
512 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
513 if (test_bit(LIOD_FORCE, &pc->pc_flags))
514 ptlrpc_abort_set(set);
519 * Let's make one more loop to make sure that ptlrpcd_check()
520 * copied all raced new rpcs into the set so we can kill them.
525 * Wait for inflight requests to drain.
527 if (!list_empty(&set->set_requests))
528 ptlrpc_set_wait(&env, set);
529 lu_context_fini(&env.le_ctx);
530 lu_context_fini(env.le_ses);
532 complete(&pc->pc_finishing);
538 complete(&pc->pc_starting);
542 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
546 pc->pc_index = index;
548 init_completion(&pc->pc_starting);
549 init_completion(&pc->pc_finishing);
550 spin_lock_init(&pc->pc_lock);
553 /* Recovery thread. */
554 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
556 /* Regular thread. */
557 snprintf(pc->pc_name, sizeof(pc->pc_name),
558 "ptlrpcd_%02d_%02d", cpt, index);
564 /* XXX: We want multiple CPU cores to share the async RPC load. So we
565 * start many ptlrpcd threads. We also want to reduce the ptlrpcd
566 * overhead caused by data transfer cross-CPU cores. So we bind
567 * all ptlrpcd threads to a CPT, in the expectation that CPTs
568 * will be defined in a way that matches these boundaries. Within
569 * a CPT a ptlrpcd thread can be scheduled on any available core.
571 * Each ptlrpcd thread has its own request queue. This can cause
572 * response delay if the thread is already busy. To help with
573 * this we define partner threads: these are other threads bound
574 * to the same CPT which will check for work in each other's
575 * request queues if they have no work to do.
577 * The desired number of partner threads can be tuned by setting
578 * ptlrpcd_partner_group_size. The default is to create pairs of
581 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
583 struct ptlrpcd_ctl *pc;
584 struct ptlrpcd_ctl **ppc;
591 LASSERT(index >= 0 && index < pd->pd_nthreads);
592 pc = &pd->pd_threads[index];
593 pc->pc_npartners = pd->pd_groupsize - 1;
595 if (pc->pc_npartners <= 0)
598 OBD_CPT_ALLOC(pc->pc_partners, cfs_cpt_tab, pc->pc_cpt,
599 sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
600 if (pc->pc_partners == NULL) {
601 pc->pc_npartners = 0;
602 GOTO(out, rc = -ENOMEM);
605 first = index - index % pd->pd_groupsize;
606 ppc = pc->pc_partners;
607 for (i = first; i < first + pd->pd_groupsize; i++) {
609 *ppc++ = &pd->pd_threads[i];
615 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
617 struct task_struct *task;
623 * Do not allow starting a second thread for one pc.
625 if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
626 CWARN("Starting second thread (%s) for same pc %p\n",
631 task = kthread_run(ptlrpcd, pc, "%s", pc->pc_name);
633 GOTO(out_set, rc = PTR_ERR(task));
635 wait_for_completion(&pc->pc_starting);
643 if (pc->pc_set != NULL) {
644 struct ptlrpc_request_set *set = pc->pc_set;
646 spin_lock(&pc->pc_lock);
648 spin_unlock(&pc->pc_lock);
649 ptlrpc_set_destroy(set);
651 clear_bit(LIOD_START, &pc->pc_flags);
655 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
659 if (!test_bit(LIOD_START, &pc->pc_flags)) {
660 CWARN("Thread for pc %p was not started\n", pc);
664 set_bit(LIOD_STOP, &pc->pc_flags);
666 set_bit(LIOD_FORCE, &pc->pc_flags);
667 wake_up(&pc->pc_set->set_waitq);
673 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
675 struct ptlrpc_request_set *set = pc->pc_set;
679 if (!test_bit(LIOD_START, &pc->pc_flags)) {
680 CWARN("Thread for pc %p was not started\n", pc);
684 wait_for_completion(&pc->pc_finishing);
686 spin_lock(&pc->pc_lock);
688 spin_unlock(&pc->pc_lock);
689 ptlrpc_set_destroy(set);
691 clear_bit(LIOD_START, &pc->pc_flags);
692 clear_bit(LIOD_STOP, &pc->pc_flags);
693 clear_bit(LIOD_FORCE, &pc->pc_flags);
696 if (pc->pc_npartners > 0) {
697 LASSERT(pc->pc_partners != NULL);
699 OBD_FREE_PTR_ARRAY(pc->pc_partners, pc->pc_npartners);
700 pc->pc_partners = NULL;
702 pc->pc_npartners = 0;
707 static void ptlrpcd_fini(void)
715 if (ptlrpcds != NULL) {
716 for (i = 0; i < ptlrpcds_num; i++) {
717 if (ptlrpcds[i] == NULL)
719 for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
720 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
721 for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
722 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
723 OBD_FREE(ptlrpcds[i], ptlrpcds[i]->pd_size);
726 OBD_FREE_PTR_ARRAY(ptlrpcds, ptlrpcds_num);
730 ptlrpcd_stop(&ptlrpcd_rcv, 0);
731 ptlrpcd_free(&ptlrpcd_rcv);
733 if (ptlrpcds_cpt_idx != NULL) {
734 ncpts = cfs_cpt_number(cfs_cpt_tab);
735 OBD_FREE_PTR_ARRAY(ptlrpcds_cpt_idx, ncpts);
736 ptlrpcds_cpt_idx = NULL;
742 static int ptlrpcd_init(void)
750 struct cfs_cpt_table *cptable;
759 * Determine the CPTs that ptlrpcd threads will run on.
761 cptable = cfs_cpt_tab;
762 ncpts = cfs_cpt_number(cptable);
763 if (ptlrpcd_cpts != NULL) {
764 struct cfs_expr_list *el;
766 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
767 OBD_ALLOC(ptlrpcds_cpt_idx, size);
768 if (ptlrpcds_cpt_idx == NULL)
769 GOTO(out, rc = -ENOMEM);
771 rc = cfs_expr_list_parse(ptlrpcd_cpts,
772 strlen(ptlrpcd_cpts),
775 CERROR("%s: invalid CPT pattern string: %s",
776 "ptlrpcd_cpts", ptlrpcd_cpts);
777 GOTO(out, rc = -EINVAL);
780 rc = cfs_expr_list_values(el, ncpts, &cpts);
781 cfs_expr_list_free(el);
783 CERROR("%s: failed to parse CPT array %s: %d\n",
784 "ptlrpcd_cpts", ptlrpcd_cpts, rc);
791 * Create the cpt-to-index map. When there is no match
792 * in the cpt table, pick a cpt at random. This could
793 * be changed to take the topology of the system into
796 for (cpt = 0; cpt < ncpts; cpt++) {
797 for (i = 0; i < rc; i++)
802 ptlrpcds_cpt_idx[cpt] = i;
805 cfs_expr_list_values_free(cpts, rc);
808 ptlrpcds_num = ncpts;
810 size = ncpts * sizeof(ptlrpcds[0]);
811 OBD_ALLOC(ptlrpcds, size);
812 if (ptlrpcds == NULL)
813 GOTO(out, rc = -ENOMEM);
816 * The max_ptlrpcds parameter is obsolete, but do something
817 * sane if it has been tuned, and complain if
818 * ptlrpcd_per_cpt_max has also been tuned.
820 if (max_ptlrpcds != 0) {
821 CWARN("max_ptlrpcds is obsolete.\n");
822 if (ptlrpcd_per_cpt_max == 0) {
823 ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
824 /* Round up if there is a remainder. */
825 if (max_ptlrpcds % ncpts != 0)
826 ptlrpcd_per_cpt_max++;
827 CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
828 ptlrpcd_per_cpt_max);
830 CWARN("ptlrpd_per_cpt_max is also set!\n");
835 * The ptlrpcd_bind_policy parameter is obsolete, but do
836 * something sane if it has been tuned, and complain if
837 * ptlrpcd_partner_group_size is also tuned.
839 if (ptlrpcd_bind_policy != 0) {
840 CWARN("ptlrpcd_bind_policy is obsolete.\n");
841 if (ptlrpcd_partner_group_size == 0) {
842 switch (ptlrpcd_bind_policy) {
843 case 1: /* PDB_POLICY_NONE */
844 case 2: /* PDB_POLICY_FULL */
845 ptlrpcd_partner_group_size = 1;
847 case 3: /* PDB_POLICY_PAIR */
848 ptlrpcd_partner_group_size = 2;
850 case 4: /* PDB_POLICY_NEIGHBOR */
852 ptlrpcd_partner_group_size = -1; /* CPT */
854 ptlrpcd_partner_group_size = 3; /* Triplets */
857 default: /* Illegal value, use the default. */
858 ptlrpcd_partner_group_size = 2;
861 CWARN("Setting ptlrpcd_partner_group_size = %d\n",
862 ptlrpcd_partner_group_size);
864 CWARN("ptlrpcd_partner_group_size is also set!\n");
868 if (ptlrpcd_partner_group_size == 0)
869 ptlrpcd_partner_group_size = 2;
870 else if (ptlrpcd_partner_group_size < 0)
871 ptlrpcd_partner_group_size = -1;
872 else if (ptlrpcd_per_cpt_max > 0 &&
873 ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
874 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
877 * Start the recovery thread first.
879 set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
880 ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
881 rc = ptlrpcd_start(&ptlrpcd_rcv);
885 for (i = 0; i < ncpts; i++) {
891 nthreads = cfs_cpt_weight(cptable, cpt);
892 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
893 nthreads = ptlrpcd_per_cpt_max;
897 if (ptlrpcd_partner_group_size <= 0) {
898 groupsize = nthreads;
899 } else if (nthreads <= ptlrpcd_partner_group_size) {
900 groupsize = nthreads;
902 groupsize = ptlrpcd_partner_group_size;
903 if (nthreads % groupsize != 0)
904 nthreads += groupsize - (nthreads % groupsize);
907 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
908 OBD_CPT_ALLOC(pd, cptable, cpt, size);
911 GOTO(out, rc = -ENOMEM);
916 pd->pd_nthreads = nthreads;
917 pd->pd_groupsize = groupsize;
921 * The ptlrpcd threads in a partner group can access
922 * each other's struct ptlrpcd_ctl, so these must be
923 * initialized before any thead is started.
925 for (j = 0; j < nthreads; j++) {
926 ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
927 rc = ptlrpcd_partners(pd, j);
932 /* XXX: We start nthreads ptlrpc daemons on this cpt.
933 * Each of them can process any non-recovery
934 * async RPC to improve overall async RPC
937 * But there are some issues with async I/O RPCs
938 * and async non-I/O RPCs processed in the same
939 * set under some cases. The ptlrpcd may be
940 * blocked by some async I/O RPC(s), then will
941 * cause other async non-I/O RPC(s) can not be
944 * Maybe we should distinguish blocked async RPCs
945 * from non-blocked async RPCs, and process them
946 * in different ptlrpcd sets to avoid unnecessary
947 * dependency. But how to distribute async RPCs
948 * load among all the ptlrpc daemons becomes
951 for (j = 0; j < nthreads; j++) {
952 rc = ptlrpcd_start(&pd->pd_threads[j]);
964 int ptlrpcd_addref(void)
970 mutex_lock(&ptlrpcd_mutex);
971 if (++ptlrpcd_users == 1) {
976 mutex_unlock(&ptlrpcd_mutex);
979 EXPORT_SYMBOL(ptlrpcd_addref);
981 void ptlrpcd_decref(void)
983 mutex_lock(&ptlrpcd_mutex);
984 if (--ptlrpcd_users == 0)
986 mutex_unlock(&ptlrpcd_mutex);
988 EXPORT_SYMBOL(ptlrpcd_decref);