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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/ptlrpcd.c
39 /** \defgroup ptlrpcd PortalRPC daemon
41 * ptlrpcd is a special thread with its own set where other user might add
42 * requests when they don't want to wait for their completion.
43 * PtlRPCD will take care of sending such requests and then processing their
44 * replies and calling completion callbacks as necessary.
45 * The callbacks are called directly from ptlrpcd context.
46 * It is important to never significantly block (esp. on RPCs!) within such
47 * completion handler or a deadlock might occur where ptlrpcd enters some
48 * callback that attempts to send another RPC and wait for it to return,
49 * during which time ptlrpcd is completely blocked, so e.g. if import
50 * fails, recovery cannot progress because connection requests are also
56 #define DEBUG_SUBSYSTEM S_RPC
58 #include "../../include/linux/libcfs/libcfs.h"
60 #include "../include/lustre_net.h"
61 #include "../include/lustre_lib.h"
62 #include "../include/lustre_ha.h"
63 #include "../include/obd_class.h" /* for obd_zombie */
64 #include "../include/obd_support.h" /* for OBD_FAIL_CHECK */
65 #include "../include/cl_object.h" /* cl_env_{get,put}() */
66 #include "../include/lprocfs_status.h"
68 #include "ptlrpc_internal.h"
70 /* One of these per CPT. */
78 struct ptlrpcd_ctl pd_threads[0];
82 * max_ptlrpcds is obsolete, but retained to ensure that the kernel
83 * module will load on a system where it has been tuned.
84 * A value other than 0 implies it was tuned, in which case the value
85 * is used to derive a setting for ptlrpcd_per_cpt_max.
87 static int max_ptlrpcds;
88 module_param(max_ptlrpcds, int, 0644);
89 MODULE_PARM_DESC(max_ptlrpcds, "Max ptlrpcd thread count to be started.");
92 * ptlrpcd_bind_policy is obsolete, but retained to ensure that
93 * the kernel module will load on a system where it has been tuned.
94 * A value other than 0 implies it was tuned, in which case the value
95 * is used to derive a setting for ptlrpcd_partner_group_size.
97 static int ptlrpcd_bind_policy;
98 module_param(ptlrpcd_bind_policy, int, 0644);
99 MODULE_PARM_DESC(ptlrpcd_bind_policy,
100 "Ptlrpcd threads binding mode (obsolete).");
103 * ptlrpcd_per_cpt_max: The maximum number of ptlrpcd threads to run
106 static int ptlrpcd_per_cpt_max;
107 module_param(ptlrpcd_per_cpt_max, int, 0644);
108 MODULE_PARM_DESC(ptlrpcd_per_cpt_max,
109 "Max ptlrpcd thread count to be started per cpt.");
112 * ptlrpcd_partner_group_size: The desired number of threads in each
113 * ptlrpcd partner thread group. Default is 2, corresponding to the
114 * old PDB_POLICY_PAIR. A negative value makes all ptlrpcd threads in
115 * a CPT partners of each other.
117 static int ptlrpcd_partner_group_size;
118 module_param(ptlrpcd_partner_group_size, int, 0644);
119 MODULE_PARM_DESC(ptlrpcd_partner_group_size,
120 "Number of ptlrpcd threads in a partner group.");
123 * ptlrpcd_cpts: A CPT string describing the CPU partitions that
124 * ptlrpcd threads should run on. Used to make ptlrpcd threads run on
125 * a subset of all CPTs.
129 * run ptlrpcd threads only on CPT 2.
133 * run ptlrpcd threads on CPTs 0, 1, 2, and 3.
135 * ptlrpcd_cpts=[0-3,5,7]
136 * run ptlrpcd threads on CPTS 0, 1, 2, 3, 5, and 7.
138 static char *ptlrpcd_cpts;
139 module_param(ptlrpcd_cpts, charp, 0644);
140 MODULE_PARM_DESC(ptlrpcd_cpts,
141 "CPU partitions ptlrpcd threads should run in");
143 /* ptlrpcds_cpt_idx maps cpt numbers to an index in the ptlrpcds array. */
144 static int *ptlrpcds_cpt_idx;
146 /* ptlrpcds_num is the number of entries in the ptlrpcds array. */
147 static int ptlrpcds_num;
148 static struct ptlrpcd **ptlrpcds;
151 * In addition to the regular thread pool above, there is a single
152 * global recovery thread. Recovery isn't critical for performance,
153 * and doesn't block, but must always be able to proceed, and it is
154 * possible that all normal ptlrpcd threads are blocked. Hence the
155 * need for a dedicated thread.
157 static struct ptlrpcd_ctl ptlrpcd_rcv;
159 struct mutex ptlrpcd_mutex;
160 static int ptlrpcd_users;
162 void ptlrpcd_wake(struct ptlrpc_request *req)
164 struct ptlrpc_request_set *rq_set = req->rq_set;
166 wake_up(&rq_set->set_waitq);
168 EXPORT_SYMBOL(ptlrpcd_wake);
170 static struct ptlrpcd_ctl *
171 ptlrpcd_select_pc(struct ptlrpc_request *req)
177 if (req && req->rq_send_state != LUSTRE_IMP_FULL)
180 cpt = cfs_cpt_current(cfs_cpt_table, 1);
181 if (!ptlrpcds_cpt_idx)
184 idx = ptlrpcds_cpt_idx[cpt];
187 /* We do not care whether it is strict load balance. */
189 if (++idx == pd->pd_nthreads)
193 return &pd->pd_threads[idx];
197 * Return transferred RPCs count.
199 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
200 struct ptlrpc_request_set *src)
202 struct list_head *tmp, *pos;
203 struct ptlrpc_request *req;
206 spin_lock(&src->set_new_req_lock);
207 if (likely(!list_empty(&src->set_new_requests))) {
208 list_for_each_safe(pos, tmp, &src->set_new_requests) {
209 req = list_entry(pos, struct ptlrpc_request,
213 list_splice_init(&src->set_new_requests, &des->set_requests);
214 rc = atomic_read(&src->set_new_count);
215 atomic_add(rc, &des->set_remaining);
216 atomic_set(&src->set_new_count, 0);
218 spin_unlock(&src->set_new_req_lock);
223 * Requests that are added to the ptlrpcd queue are sent via
224 * ptlrpcd_check->ptlrpc_check_set().
226 void ptlrpcd_add_req(struct ptlrpc_request *req)
228 struct ptlrpcd_ctl *pc;
231 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
233 spin_lock(&req->rq_lock);
234 if (req->rq_invalid_rqset) {
235 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5),
236 back_to_sleep, NULL);
238 req->rq_invalid_rqset = 0;
239 spin_unlock(&req->rq_lock);
240 l_wait_event(req->rq_set_waitq, !req->rq_set, &lwi);
241 } else if (req->rq_set) {
242 /* If we have a valid "rq_set", just reuse it to avoid double
245 LASSERT(req->rq_phase == RQ_PHASE_NEW);
246 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
248 /* ptlrpc_check_set will decrease the count */
249 atomic_inc(&req->rq_set->set_remaining);
250 spin_unlock(&req->rq_lock);
251 wake_up(&req->rq_set->set_waitq);
254 spin_unlock(&req->rq_lock);
257 pc = ptlrpcd_select_pc(req);
259 DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]",
260 req, pc->pc_name, pc->pc_index);
262 ptlrpc_set_add_new_req(pc, req);
264 EXPORT_SYMBOL(ptlrpcd_add_req);
266 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
268 atomic_inc(&set->set_refcount);
272 * Check if there is more work to do on ptlrpcd set.
275 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
277 struct list_head *tmp, *pos;
278 struct ptlrpc_request *req;
279 struct ptlrpc_request_set *set = pc->pc_set;
283 if (atomic_read(&set->set_new_count)) {
284 spin_lock(&set->set_new_req_lock);
285 if (likely(!list_empty(&set->set_new_requests))) {
286 list_splice_init(&set->set_new_requests,
288 atomic_add(atomic_read(&set->set_new_count),
289 &set->set_remaining);
290 atomic_set(&set->set_new_count, 0);
292 * Need to calculate its timeout.
296 spin_unlock(&set->set_new_req_lock);
299 /* We should call lu_env_refill() before handling new requests to make
300 * sure that env key the requests depending on really exists.
302 rc2 = lu_env_refill(env);
305 * XXX This is very awkward situation, because
306 * execution can neither continue (request
307 * interpreters assume that env is set up), nor repeat
308 * the loop (as this potentially results in a tight
309 * loop of -ENOMEM's).
311 * Fortunately, refill only ever does something when
312 * new modules are loaded, i.e., early during boot up.
314 CERROR("Failure to refill session: %d\n", rc2);
318 if (atomic_read(&set->set_remaining))
319 rc |= ptlrpc_check_set(env, set);
321 /* NB: ptlrpc_check_set has already moved completed request at the
322 * head of seq::set_requests
324 list_for_each_safe(pos, tmp, &set->set_requests) {
325 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
326 if (req->rq_phase != RQ_PHASE_COMPLETE)
329 list_del_init(&req->rq_set_chain);
331 ptlrpc_req_finished(req);
336 * If new requests have been added, make sure to wake up.
338 rc = atomic_read(&set->set_new_count);
340 /* If we have nothing to do, check whether we can take some
341 * work from our partner threads.
343 if (rc == 0 && pc->pc_npartners > 0) {
344 struct ptlrpcd_ctl *partner;
345 struct ptlrpc_request_set *ps;
346 int first = pc->pc_cursor;
349 partner = pc->pc_partners[pc->pc_cursor++];
350 if (pc->pc_cursor >= pc->pc_npartners)
355 spin_lock(&partner->pc_lock);
356 ps = partner->pc_set;
358 spin_unlock(&partner->pc_lock);
362 ptlrpc_reqset_get(ps);
363 spin_unlock(&partner->pc_lock);
365 if (atomic_read(&ps->set_new_count)) {
366 rc = ptlrpcd_steal_rqset(set, ps);
368 CDEBUG(D_RPCTRACE, "transfer %d async RPCs [%d->%d]\n",
369 rc, partner->pc_index,
372 ptlrpc_reqset_put(ps);
373 } while (rc == 0 && pc->pc_cursor != first);
381 * Main ptlrpcd thread.
382 * ptlrpc's code paths like to execute in process context, so we have this
383 * thread which spins on a set which contains the rpcs and sends them.
386 static int ptlrpcd(void *arg)
388 struct ptlrpcd_ctl *pc = arg;
389 struct ptlrpc_request_set *set;
390 struct lu_context ses = { 0 };
391 struct lu_env env = { .le_ses = &ses };
396 if (cfs_cpt_bind(cfs_cpt_table, pc->pc_cpt) != 0)
397 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
400 * Allocate the request set after the thread has been bound
401 * above. This is safe because no requests will be queued
402 * until all ptlrpcd threads have confirmed that they have
403 * successfully started.
405 set = ptlrpc_prep_set();
410 spin_lock(&pc->pc_lock);
412 spin_unlock(&pc->pc_lock);
414 * XXX So far only "client" ptlrpcd uses an environment. In
415 * the future, ptlrpcd thread (or a thread-set) has to given
416 * an argument, describing its "scope".
418 rc = lu_context_init(&env.le_ctx,
419 LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
421 rc = lu_context_init(env.le_ses,
422 LCT_SESSION | LCT_REMEMBER | LCT_NOREF);
424 lu_context_fini(&env.le_ctx);
430 complete(&pc->pc_starting);
433 * This mainloop strongly resembles ptlrpc_set_wait() except that our
434 * set never completes. ptlrpcd_check() calls ptlrpc_check_set() when
435 * there are requests in the set. New requests come in on the set's
436 * new_req_list and ptlrpcd_check() moves them into the set.
439 struct l_wait_info lwi;
442 timeout = ptlrpc_set_next_timeout(set);
443 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
444 ptlrpc_expired_set, set);
446 lu_context_enter(&env.le_ctx);
447 lu_context_enter(env.le_ses);
448 l_wait_event(set->set_waitq, ptlrpcd_check(&env, pc), &lwi);
449 lu_context_exit(&env.le_ctx);
450 lu_context_exit(env.le_ses);
453 * Abort inflight rpcs for forced stop case.
455 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
456 if (test_bit(LIOD_FORCE, &pc->pc_flags))
457 ptlrpc_abort_set(set);
462 * Let's make one more loop to make sure that ptlrpcd_check()
463 * copied all raced new rpcs into the set so we can kill them.
468 * Wait for inflight requests to drain.
470 if (!list_empty(&set->set_requests))
471 ptlrpc_set_wait(set);
472 lu_context_fini(&env.le_ctx);
473 lu_context_fini(env.le_ses);
475 complete(&pc->pc_finishing);
480 complete(&pc->pc_starting);
484 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
486 pc->pc_index = index;
488 init_completion(&pc->pc_starting);
489 init_completion(&pc->pc_finishing);
490 spin_lock_init(&pc->pc_lock);
493 /* Recovery thread. */
494 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
496 /* Regular thread. */
497 snprintf(pc->pc_name, sizeof(pc->pc_name),
498 "ptlrpcd_%02d_%02d", cpt, index);
502 /* XXX: We want multiple CPU cores to share the async RPC load. So we
503 * start many ptlrpcd threads. We also want to reduce the ptlrpcd
504 * overhead caused by data transfer cross-CPU cores. So we bind
505 * all ptlrpcd threads to a CPT, in the expectation that CPTs
506 * will be defined in a way that matches these boundaries. Within
507 * a CPT a ptlrpcd thread can be scheduled on any available core.
509 * Each ptlrpcd thread has its own request queue. This can cause
510 * response delay if the thread is already busy. To help with
511 * this we define partner threads: these are other threads bound
512 * to the same CPT which will check for work in each other's
513 * request queues if they have no work to do.
515 * The desired number of partner threads can be tuned by setting
516 * ptlrpcd_partner_group_size. The default is to create pairs of
519 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
521 struct ptlrpcd_ctl *pc;
522 struct ptlrpcd_ctl **ppc;
528 LASSERT(index >= 0 && index < pd->pd_nthreads);
529 pc = &pd->pd_threads[index];
530 pc->pc_npartners = pd->pd_groupsize - 1;
532 if (pc->pc_npartners <= 0)
535 size = sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners;
536 pc->pc_partners = kzalloc_node(size, GFP_NOFS,
537 cfs_cpt_spread_node(cfs_cpt_table,
539 if (!pc->pc_partners) {
540 pc->pc_npartners = 0;
545 first = index - index % pd->pd_groupsize;
546 ppc = pc->pc_partners;
547 for (i = first; i < first + pd->pd_groupsize; i++) {
549 *ppc++ = &pd->pd_threads[i];
555 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
557 struct task_struct *task;
561 * Do not allow start second thread for one pc.
563 if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
564 CWARN("Starting second thread (%s) for same pc %p\n",
570 * So far only "client" ptlrpcd uses an environment. In the future,
571 * ptlrpcd thread (or a thread-set) has to be given an argument,
572 * describing its "scope".
574 rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
578 task = kthread_run(ptlrpcd, pc, "%s", pc->pc_name);
584 wait_for_completion(&pc->pc_starting);
593 struct ptlrpc_request_set *set = pc->pc_set;
595 spin_lock(&pc->pc_lock);
597 spin_unlock(&pc->pc_lock);
598 ptlrpc_set_destroy(set);
600 lu_context_fini(&pc->pc_env.le_ctx);
603 clear_bit(LIOD_START, &pc->pc_flags);
607 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
609 if (!test_bit(LIOD_START, &pc->pc_flags)) {
610 CWARN("Thread for pc %p was not started\n", pc);
614 set_bit(LIOD_STOP, &pc->pc_flags);
616 set_bit(LIOD_FORCE, &pc->pc_flags);
617 wake_up(&pc->pc_set->set_waitq);
620 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
622 struct ptlrpc_request_set *set = pc->pc_set;
624 if (!test_bit(LIOD_START, &pc->pc_flags)) {
625 CWARN("Thread for pc %p was not started\n", pc);
629 wait_for_completion(&pc->pc_finishing);
630 lu_context_fini(&pc->pc_env.le_ctx);
632 spin_lock(&pc->pc_lock);
634 spin_unlock(&pc->pc_lock);
635 ptlrpc_set_destroy(set);
637 clear_bit(LIOD_START, &pc->pc_flags);
638 clear_bit(LIOD_STOP, &pc->pc_flags);
639 clear_bit(LIOD_FORCE, &pc->pc_flags);
642 if (pc->pc_npartners > 0) {
643 LASSERT(pc->pc_partners);
645 kfree(pc->pc_partners);
646 pc->pc_partners = NULL;
648 pc->pc_npartners = 0;
652 static void ptlrpcd_fini(void)
658 for (i = 0; i < ptlrpcds_num; i++) {
661 for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
662 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
663 for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
664 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
672 ptlrpcd_stop(&ptlrpcd_rcv, 0);
673 ptlrpcd_free(&ptlrpcd_rcv);
675 kfree(ptlrpcds_cpt_idx);
676 ptlrpcds_cpt_idx = NULL;
679 static int ptlrpcd_init(void)
687 struct cfs_cpt_table *cptable;
694 * Determine the CPTs that ptlrpcd threads will run on.
696 cptable = cfs_cpt_table;
697 ncpts = cfs_cpt_number(cptable);
699 struct cfs_expr_list *el;
701 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
702 ptlrpcds_cpt_idx = kzalloc(size, GFP_KERNEL);
703 if (!ptlrpcds_cpt_idx) {
708 rc = cfs_expr_list_parse(ptlrpcd_cpts,
709 strlen(ptlrpcd_cpts),
713 CERROR("ptlrpcd_cpts: invalid CPT pattern string: %s",
719 rc = cfs_expr_list_values(el, ncpts, &cpts);
720 cfs_expr_list_free(el);
722 CERROR("ptlrpcd_cpts: failed to parse CPT array %s: %d\n",
730 * Create the cpt-to-index map. When there is no match
731 * in the cpt table, pick a cpt at random. This could
732 * be changed to take the topology of the system into
735 for (cpt = 0; cpt < ncpts; cpt++) {
736 for (i = 0; i < rc; i++)
741 ptlrpcds_cpt_idx[cpt] = i;
744 cfs_expr_list_values_free(cpts, rc);
747 ptlrpcds_num = ncpts;
749 size = ncpts * sizeof(ptlrpcds[0]);
750 ptlrpcds = kzalloc(size, GFP_KERNEL);
757 * The max_ptlrpcds parameter is obsolete, but do something
758 * sane if it has been tuned, and complain if
759 * ptlrpcd_per_cpt_max has also been tuned.
761 if (max_ptlrpcds != 0) {
762 CWARN("max_ptlrpcds is obsolete.\n");
763 if (ptlrpcd_per_cpt_max == 0) {
764 ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
765 /* Round up if there is a remainder. */
766 if (max_ptlrpcds % ncpts != 0)
767 ptlrpcd_per_cpt_max++;
768 CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
769 ptlrpcd_per_cpt_max);
771 CWARN("ptlrpd_per_cpt_max is also set!\n");
776 * The ptlrpcd_bind_policy parameter is obsolete, but do
777 * something sane if it has been tuned, and complain if
778 * ptlrpcd_partner_group_size is also tuned.
780 if (ptlrpcd_bind_policy != 0) {
781 CWARN("ptlrpcd_bind_policy is obsolete.\n");
782 if (ptlrpcd_partner_group_size == 0) {
783 switch (ptlrpcd_bind_policy) {
784 case 1: /* PDB_POLICY_NONE */
785 case 2: /* PDB_POLICY_FULL */
786 ptlrpcd_partner_group_size = 1;
788 case 3: /* PDB_POLICY_PAIR */
789 ptlrpcd_partner_group_size = 2;
791 case 4: /* PDB_POLICY_NEIGHBOR */
793 ptlrpcd_partner_group_size = -1; /* CPT */
795 ptlrpcd_partner_group_size = 3; /* Triplets */
798 default: /* Illegal value, use the default. */
799 ptlrpcd_partner_group_size = 2;
802 CWARN("Setting ptlrpcd_partner_group_size = %d\n",
803 ptlrpcd_partner_group_size);
805 CWARN("ptlrpcd_partner_group_size is also set!\n");
809 if (ptlrpcd_partner_group_size == 0)
810 ptlrpcd_partner_group_size = 2;
811 else if (ptlrpcd_partner_group_size < 0)
812 ptlrpcd_partner_group_size = -1;
813 else if (ptlrpcd_per_cpt_max > 0 &&
814 ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
815 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
818 * Start the recovery thread first.
820 set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
821 ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
822 rc = ptlrpcd_start(&ptlrpcd_rcv);
826 for (i = 0; i < ncpts; i++) {
832 nthreads = cfs_cpt_weight(cptable, cpt);
833 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
834 nthreads = ptlrpcd_per_cpt_max;
838 if (ptlrpcd_partner_group_size <= 0) {
839 groupsize = nthreads;
840 } else if (nthreads <= ptlrpcd_partner_group_size) {
841 groupsize = nthreads;
843 groupsize = ptlrpcd_partner_group_size;
844 if (nthreads % groupsize != 0)
845 nthreads += groupsize - (nthreads % groupsize);
848 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
849 pd = kzalloc_node(size, GFP_NOFS,
850 cfs_cpt_spread_node(cfs_cpt_table, cpt));
859 pd->pd_nthreads = nthreads;
860 pd->pd_groupsize = groupsize;
864 * The ptlrpcd threads in a partner group can access
865 * each other's struct ptlrpcd_ctl, so these must be
866 * initialized before any thread is started.
868 for (j = 0; j < nthreads; j++) {
869 ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
870 rc = ptlrpcd_partners(pd, j);
875 /* XXX: We start nthreads ptlrpc daemons.
876 * Each of them can process any non-recovery
877 * async RPC to improve overall async RPC
880 * But there are some issues with async I/O RPCs
881 * and async non-I/O RPCs processed in the same
882 * set under some cases. The ptlrpcd may be
883 * blocked by some async I/O RPC(s), then will
884 * cause other async non-I/O RPC(s) can not be
887 * Maybe we should distinguish blocked async RPCs
888 * from non-blocked async RPCs, and process them
889 * in different ptlrpcd sets to avoid unnecessary
890 * dependency. But how to distribute async RPCs
891 * load among all the ptlrpc daemons becomes
894 for (j = 0; j < nthreads; j++) {
895 rc = ptlrpcd_start(&pd->pd_threads[j]);
907 int ptlrpcd_addref(void)
911 mutex_lock(&ptlrpcd_mutex);
912 if (++ptlrpcd_users == 1) {
917 mutex_unlock(&ptlrpcd_mutex);
920 EXPORT_SYMBOL(ptlrpcd_addref);
922 void ptlrpcd_decref(void)
924 mutex_lock(&ptlrpcd_mutex);
925 if (--ptlrpcd_users == 0)
927 mutex_unlock(&ptlrpcd_mutex);
929 EXPORT_SYMBOL(ptlrpcd_decref);