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, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ptlrpc/ptlrpcd.c
35 /** \defgroup ptlrpcd PortalRPC daemon
37 * ptlrpcd is a special thread with its own set where other user might add
38 * requests when they don't want to wait for their completion.
39 * PtlRPCD will take care of sending such requests and then processing their
40 * replies and calling completion callbacks as necessary.
41 * The callbacks are called directly from ptlrpcd context.
42 * It is important to never significantly block (esp. on RPCs!) within such
43 * completion handler or a deadlock might occur where ptlrpcd enters some
44 * callback that attempts to send another RPC and wait for it to return,
45 * during which time ptlrpcd is completely blocked, so e.g. if import
46 * fails, recovery cannot progress because connection requests are also
52 #define DEBUG_SUBSYSTEM S_RPC
54 #include "../../include/linux/libcfs/libcfs.h"
56 #include "../include/lustre_net.h"
57 #include "../include/lustre_lib.h"
58 #include "../include/lustre_ha.h"
59 #include "../include/obd_class.h" /* for obd_zombie */
60 #include "../include/obd_support.h" /* for OBD_FAIL_CHECK */
61 #include "../include/cl_object.h" /* cl_env_{get,put}() */
62 #include "../include/lprocfs_status.h"
64 #include "ptlrpc_internal.h"
66 /* One of these per CPT. */
74 struct ptlrpcd_ctl pd_threads[0];
78 * max_ptlrpcds is obsolete, but retained to ensure that the kernel
79 * module will load on a system where it has been tuned.
80 * A value other than 0 implies it was tuned, in which case the value
81 * is used to derive a setting for ptlrpcd_per_cpt_max.
83 static int max_ptlrpcds;
84 module_param(max_ptlrpcds, int, 0644);
85 MODULE_PARM_DESC(max_ptlrpcds, "Max ptlrpcd thread count to be started.");
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;
158 void ptlrpcd_wake(struct ptlrpc_request *req)
160 struct ptlrpc_request_set *set = req->rq_set;
162 wake_up(&set->set_waitq);
164 EXPORT_SYMBOL(ptlrpcd_wake);
166 static struct ptlrpcd_ctl *
167 ptlrpcd_select_pc(struct ptlrpc_request *req)
173 if (req && req->rq_send_state != LUSTRE_IMP_FULL)
176 cpt = cfs_cpt_current(cfs_cpt_table, 1);
177 if (!ptlrpcds_cpt_idx)
180 idx = ptlrpcds_cpt_idx[cpt];
183 /* We do not care whether it is strict load balance. */
185 if (++idx == pd->pd_nthreads)
189 return &pd->pd_threads[idx];
193 * Return transferred RPCs count.
195 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
196 struct ptlrpc_request_set *src)
198 struct list_head *tmp, *pos;
199 struct ptlrpc_request *req;
202 spin_lock(&src->set_new_req_lock);
203 if (likely(!list_empty(&src->set_new_requests))) {
204 list_for_each_safe(pos, tmp, &src->set_new_requests) {
205 req = list_entry(pos, struct ptlrpc_request,
209 list_splice_init(&src->set_new_requests, &des->set_requests);
210 rc = atomic_read(&src->set_new_count);
211 atomic_add(rc, &des->set_remaining);
212 atomic_set(&src->set_new_count, 0);
214 spin_unlock(&src->set_new_req_lock);
219 * Requests that are added to the ptlrpcd queue are sent via
220 * ptlrpcd_check->ptlrpc_check_set().
222 void ptlrpcd_add_req(struct ptlrpc_request *req)
224 struct ptlrpcd_ctl *pc;
227 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
229 spin_lock(&req->rq_lock);
230 if (req->rq_invalid_rqset) {
231 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
232 back_to_sleep, NULL);
234 req->rq_invalid_rqset = 0;
235 spin_unlock(&req->rq_lock);
236 l_wait_event(req->rq_set_waitq, !req->rq_set, &lwi);
237 } else if (req->rq_set) {
238 /* If we have a valid "rq_set", just reuse it to avoid double
241 LASSERT(req->rq_phase == RQ_PHASE_NEW);
242 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
244 /* ptlrpc_check_set will decrease the count */
245 atomic_inc(&req->rq_set->set_remaining);
246 spin_unlock(&req->rq_lock);
247 wake_up(&req->rq_set->set_waitq);
250 spin_unlock(&req->rq_lock);
253 pc = ptlrpcd_select_pc(req);
255 DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
256 req, pc->pc_name, pc->pc_index);
258 ptlrpc_set_add_new_req(pc, req);
260 EXPORT_SYMBOL(ptlrpcd_add_req);
262 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
264 atomic_inc(&set->set_refcount);
268 * Check if there is more work to do on ptlrpcd set.
271 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
273 struct list_head *tmp, *pos;
274 struct ptlrpc_request *req;
275 struct ptlrpc_request_set *set = pc->pc_set;
279 if (atomic_read(&set->set_new_count)) {
280 spin_lock(&set->set_new_req_lock);
281 if (likely(!list_empty(&set->set_new_requests))) {
282 list_splice_init(&set->set_new_requests,
284 atomic_add(atomic_read(&set->set_new_count),
285 &set->set_remaining);
286 atomic_set(&set->set_new_count, 0);
288 * Need to calculate its timeout.
292 spin_unlock(&set->set_new_req_lock);
295 /* We should call lu_env_refill() before handling new requests to make
296 * sure that env key the requests depending on really exists.
298 rc2 = lu_env_refill(env);
301 * XXX This is very awkward situation, because
302 * execution can neither continue (request
303 * interpreters assume that env is set up), nor repeat
304 * the loop (as this potentially results in a tight
305 * loop of -ENOMEM's).
307 * Fortunately, refill only ever does something when
308 * new modules are loaded, i.e., early during boot up.
310 CERROR("Failure to refill session: %d\n", rc2);
314 if (atomic_read(&set->set_remaining))
315 rc |= ptlrpc_check_set(env, set);
317 /* NB: ptlrpc_check_set has already moved completed request at the
318 * head of seq::set_requests
320 list_for_each_safe(pos, tmp, &set->set_requests) {
321 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
322 if (req->rq_phase != RQ_PHASE_COMPLETE)
325 list_del_init(&req->rq_set_chain);
327 ptlrpc_req_finished(req);
332 * If new requests have been added, make sure to wake up.
334 rc = atomic_read(&set->set_new_count);
336 /* If we have nothing to do, check whether we can take some
337 * work from our partner threads.
339 if (rc == 0 && pc->pc_npartners > 0) {
340 struct ptlrpcd_ctl *partner;
341 struct ptlrpc_request_set *ps;
342 int first = pc->pc_cursor;
345 partner = pc->pc_partners[pc->pc_cursor++];
346 if (pc->pc_cursor >= pc->pc_npartners)
351 spin_lock(&partner->pc_lock);
352 ps = partner->pc_set;
354 spin_unlock(&partner->pc_lock);
358 ptlrpc_reqset_get(ps);
359 spin_unlock(&partner->pc_lock);
361 if (atomic_read(&ps->set_new_count)) {
362 rc = ptlrpcd_steal_rqset(set, ps);
364 CDEBUG(D_RPCTRACE, "transfer %d async RPCs [%d->%d]\n",
365 rc, partner->pc_index,
368 ptlrpc_reqset_put(ps);
369 } while (rc == 0 && pc->pc_cursor != first);
377 * Main ptlrpcd thread.
378 * ptlrpc's code paths like to execute in process context, so we have this
379 * thread which spins on a set which contains the rpcs and sends them.
382 static int ptlrpcd(void *arg)
384 struct ptlrpcd_ctl *pc = arg;
385 struct ptlrpc_request_set *set;
386 struct lu_context ses = { 0 };
387 struct lu_env env = { .le_ses = &ses };
392 if (cfs_cpt_bind(cfs_cpt_table, pc->pc_cpt) != 0)
393 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
396 * Allocate the request set after the thread has been bound
397 * above. This is safe because no requests will be queued
398 * until all ptlrpcd threads have confirmed that they have
399 * successfully started.
401 set = ptlrpc_prep_set();
406 spin_lock(&pc->pc_lock);
408 spin_unlock(&pc->pc_lock);
410 * XXX So far only "client" ptlrpcd uses an environment. In
411 * the future, ptlrpcd thread (or a thread-set) has to given
412 * an argument, describing its "scope".
414 rc = lu_context_init(&env.le_ctx,
415 LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
417 rc = lu_context_init(env.le_ses,
418 LCT_SESSION | LCT_REMEMBER | LCT_NOREF);
420 lu_context_fini(&env.le_ctx);
426 complete(&pc->pc_starting);
429 * This mainloop strongly resembles ptlrpc_set_wait() except that our
430 * set never completes. ptlrpcd_check() calls ptlrpc_check_set() when
431 * there are requests in the set. New requests come in on the set's
432 * new_req_list and ptlrpcd_check() moves them into the set.
435 struct l_wait_info lwi;
438 timeout = ptlrpc_set_next_timeout(set);
439 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
440 ptlrpc_expired_set, set);
442 lu_context_enter(&env.le_ctx);
443 lu_context_enter(env.le_ses);
444 l_wait_event(set->set_waitq, ptlrpcd_check(&env, pc), &lwi);
445 lu_context_exit(&env.le_ctx);
446 lu_context_exit(env.le_ses);
449 * Abort inflight rpcs for forced stop case.
451 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
452 if (test_bit(LIOD_FORCE, &pc->pc_flags))
453 ptlrpc_abort_set(set);
458 * Let's make one more loop to make sure that ptlrpcd_check()
459 * copied all raced new rpcs into the set so we can kill them.
464 * Wait for inflight requests to drain.
466 if (!list_empty(&set->set_requests))
467 ptlrpc_set_wait(set);
468 lu_context_fini(&env.le_ctx);
469 lu_context_fini(env.le_ses);
471 complete(&pc->pc_finishing);
476 complete(&pc->pc_starting);
480 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
482 pc->pc_index = index;
484 init_completion(&pc->pc_starting);
485 init_completion(&pc->pc_finishing);
486 spin_lock_init(&pc->pc_lock);
489 /* Recovery thread. */
490 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
492 /* Regular thread. */
493 snprintf(pc->pc_name, sizeof(pc->pc_name),
494 "ptlrpcd_%02d_%02d", cpt, index);
498 /* XXX: We want multiple CPU cores to share the async RPC load. So we
499 * start many ptlrpcd threads. We also want to reduce the ptlrpcd
500 * overhead caused by data transfer cross-CPU cores. So we bind
501 * all ptlrpcd threads to a CPT, in the expectation that CPTs
502 * will be defined in a way that matches these boundaries. Within
503 * a CPT a ptlrpcd thread can be scheduled on any available core.
505 * Each ptlrpcd thread has its own request queue. This can cause
506 * response delay if the thread is already busy. To help with
507 * this we define partner threads: these are other threads bound
508 * to the same CPT which will check for work in each other's
509 * request queues if they have no work to do.
511 * The desired number of partner threads can be tuned by setting
512 * ptlrpcd_partner_group_size. The default is to create pairs of
515 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
517 struct ptlrpcd_ctl *pc;
518 struct ptlrpcd_ctl **ppc;
524 LASSERT(index >= 0 && index < pd->pd_nthreads);
525 pc = &pd->pd_threads[index];
526 pc->pc_npartners = pd->pd_groupsize - 1;
528 if (pc->pc_npartners <= 0)
531 size = sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners;
532 pc->pc_partners = kzalloc_node(size, GFP_NOFS,
533 cfs_cpt_spread_node(cfs_cpt_table,
535 if (!pc->pc_partners) {
536 pc->pc_npartners = 0;
541 first = index - index % pd->pd_groupsize;
542 ppc = pc->pc_partners;
543 for (i = first; i < first + pd->pd_groupsize; i++) {
545 *ppc++ = &pd->pd_threads[i];
551 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
553 struct task_struct *task;
557 * Do not allow start second thread for one pc.
559 if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
560 CWARN("Starting second thread (%s) for same pc %p\n",
566 * So far only "client" ptlrpcd uses an environment. In the future,
567 * ptlrpcd thread (or a thread-set) has to be given an argument,
568 * describing its "scope".
570 rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
574 task = kthread_run(ptlrpcd, pc, "%s", pc->pc_name);
580 wait_for_completion(&pc->pc_starting);
589 struct ptlrpc_request_set *set = pc->pc_set;
591 spin_lock(&pc->pc_lock);
593 spin_unlock(&pc->pc_lock);
594 ptlrpc_set_destroy(set);
596 lu_context_fini(&pc->pc_env.le_ctx);
599 clear_bit(LIOD_START, &pc->pc_flags);
603 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
605 if (!test_bit(LIOD_START, &pc->pc_flags)) {
606 CWARN("Thread for pc %p was not started\n", pc);
610 set_bit(LIOD_STOP, &pc->pc_flags);
612 set_bit(LIOD_FORCE, &pc->pc_flags);
613 wake_up(&pc->pc_set->set_waitq);
616 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
618 struct ptlrpc_request_set *set = pc->pc_set;
620 if (!test_bit(LIOD_START, &pc->pc_flags)) {
621 CWARN("Thread for pc %p was not started\n", pc);
625 wait_for_completion(&pc->pc_finishing);
626 lu_context_fini(&pc->pc_env.le_ctx);
628 spin_lock(&pc->pc_lock);
630 spin_unlock(&pc->pc_lock);
631 ptlrpc_set_destroy(set);
633 clear_bit(LIOD_START, &pc->pc_flags);
634 clear_bit(LIOD_STOP, &pc->pc_flags);
635 clear_bit(LIOD_FORCE, &pc->pc_flags);
638 if (pc->pc_npartners > 0) {
639 LASSERT(pc->pc_partners);
641 kfree(pc->pc_partners);
642 pc->pc_partners = NULL;
644 pc->pc_npartners = 0;
648 static void ptlrpcd_fini(void)
654 for (i = 0; i < ptlrpcds_num; i++) {
657 for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
658 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
659 for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
660 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
668 ptlrpcd_stop(&ptlrpcd_rcv, 0);
669 ptlrpcd_free(&ptlrpcd_rcv);
671 kfree(ptlrpcds_cpt_idx);
672 ptlrpcds_cpt_idx = NULL;
675 static int ptlrpcd_init(void)
683 struct cfs_cpt_table *cptable;
690 * Determine the CPTs that ptlrpcd threads will run on.
692 cptable = cfs_cpt_table;
693 ncpts = cfs_cpt_number(cptable);
695 struct cfs_expr_list *el;
697 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
698 ptlrpcds_cpt_idx = kzalloc(size, GFP_KERNEL);
699 if (!ptlrpcds_cpt_idx) {
704 rc = cfs_expr_list_parse(ptlrpcd_cpts,
705 strlen(ptlrpcd_cpts),
709 CERROR("ptlrpcd_cpts: invalid CPT pattern string: %s",
715 rc = cfs_expr_list_values(el, ncpts, &cpts);
716 cfs_expr_list_free(el);
718 CERROR("ptlrpcd_cpts: failed to parse CPT array %s: %d\n",
726 * Create the cpt-to-index map. When there is no match
727 * in the cpt table, pick a cpt at random. This could
728 * be changed to take the topology of the system into
731 for (cpt = 0; cpt < ncpts; cpt++) {
732 for (i = 0; i < rc; i++)
737 ptlrpcds_cpt_idx[cpt] = i;
740 cfs_expr_list_values_free(cpts, rc);
743 ptlrpcds_num = ncpts;
745 size = ncpts * sizeof(ptlrpcds[0]);
746 ptlrpcds = kzalloc(size, GFP_KERNEL);
753 * The max_ptlrpcds parameter is obsolete, but do something
754 * sane if it has been tuned, and complain if
755 * ptlrpcd_per_cpt_max has also been tuned.
757 if (max_ptlrpcds != 0) {
758 CWARN("max_ptlrpcds is obsolete.\n");
759 if (ptlrpcd_per_cpt_max == 0) {
760 ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
761 /* Round up if there is a remainder. */
762 if (max_ptlrpcds % ncpts != 0)
763 ptlrpcd_per_cpt_max++;
764 CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
765 ptlrpcd_per_cpt_max);
767 CWARN("ptlrpd_per_cpt_max is also set!\n");
772 * The ptlrpcd_bind_policy parameter is obsolete, but do
773 * something sane if it has been tuned, and complain if
774 * ptlrpcd_partner_group_size is also tuned.
776 if (ptlrpcd_bind_policy != 0) {
777 CWARN("ptlrpcd_bind_policy is obsolete.\n");
778 if (ptlrpcd_partner_group_size == 0) {
779 switch (ptlrpcd_bind_policy) {
780 case 1: /* PDB_POLICY_NONE */
781 case 2: /* PDB_POLICY_FULL */
782 ptlrpcd_partner_group_size = 1;
784 case 3: /* PDB_POLICY_PAIR */
785 ptlrpcd_partner_group_size = 2;
787 case 4: /* PDB_POLICY_NEIGHBOR */
789 ptlrpcd_partner_group_size = -1; /* CPT */
791 ptlrpcd_partner_group_size = 3; /* Triplets */
794 default: /* Illegal value, use the default. */
795 ptlrpcd_partner_group_size = 2;
798 CWARN("Setting ptlrpcd_partner_group_size = %d\n",
799 ptlrpcd_partner_group_size);
801 CWARN("ptlrpcd_partner_group_size is also set!\n");
805 if (ptlrpcd_partner_group_size == 0)
806 ptlrpcd_partner_group_size = 2;
807 else if (ptlrpcd_partner_group_size < 0)
808 ptlrpcd_partner_group_size = -1;
809 else if (ptlrpcd_per_cpt_max > 0 &&
810 ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
811 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
814 * Start the recovery thread first.
816 set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
817 ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
818 rc = ptlrpcd_start(&ptlrpcd_rcv);
822 for (i = 0; i < ncpts; i++) {
828 nthreads = cfs_cpt_weight(cptable, cpt);
829 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
830 nthreads = ptlrpcd_per_cpt_max;
834 if (ptlrpcd_partner_group_size <= 0) {
835 groupsize = nthreads;
836 } else if (nthreads <= ptlrpcd_partner_group_size) {
837 groupsize = nthreads;
839 groupsize = ptlrpcd_partner_group_size;
840 if (nthreads % groupsize != 0)
841 nthreads += groupsize - (nthreads % groupsize);
844 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
845 pd = kzalloc_node(size, GFP_NOFS,
846 cfs_cpt_spread_node(cfs_cpt_table, cpt));
855 pd->pd_nthreads = nthreads;
856 pd->pd_groupsize = groupsize;
860 * The ptlrpcd threads in a partner group can access
861 * each other's struct ptlrpcd_ctl, so these must be
862 * initialized before any thread is started.
864 for (j = 0; j < nthreads; j++) {
865 ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
866 rc = ptlrpcd_partners(pd, j);
871 /* XXX: We start nthreads ptlrpc daemons.
872 * Each of them can process any non-recovery
873 * async RPC to improve overall async RPC
876 * But there are some issues with async I/O RPCs
877 * and async non-I/O RPCs processed in the same
878 * set under some cases. The ptlrpcd may be
879 * blocked by some async I/O RPC(s), then will
880 * cause other async non-I/O RPC(s) can not be
883 * Maybe we should distinguish blocked async RPCs
884 * from non-blocked async RPCs, and process them
885 * in different ptlrpcd sets to avoid unnecessary
886 * dependency. But how to distribute async RPCs
887 * load among all the ptlrpc daemons becomes
890 for (j = 0; j < nthreads; j++) {
891 rc = ptlrpcd_start(&pd->pd_threads[j]);
903 int ptlrpcd_addref(void)
907 mutex_lock(&ptlrpcd_mutex);
908 if (++ptlrpcd_users == 1) {
913 mutex_unlock(&ptlrpcd_mutex);
916 EXPORT_SYMBOL(ptlrpcd_addref);
918 void ptlrpcd_decref(void)
920 mutex_lock(&ptlrpcd_mutex);
921 if (--ptlrpcd_users == 0)
923 mutex_unlock(&ptlrpcd_mutex);
925 EXPORT_SYMBOL(ptlrpcd_decref);