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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Author: Isaac Huang <isaac@clusterfs.com>
40 * 2012-05-13: Liang Zhen <liang@whamcloud.com>
41 * - percpt data for service to improve smp performance
45 #define DEBUG_SUBSYSTEM S_LNET
57 static struct smoketest_rpc {
58 spinlock_t rpc_glock; /* global lock */
59 srpc_service_t *rpc_services[SRPC_SERVICE_MAX_ID + 1];
60 lnet_handle_eq_t rpc_lnet_eq; /* _the_ LNet event queue */
61 srpc_state_t rpc_state;
62 srpc_counters_t rpc_counters;
63 __u64 rpc_matchbits; /* matchbits counter */
67 srpc_serv_portal(int svc_id)
69 return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
70 SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
74 int srpc_handle_rpc(swi_workitem_t *wi);
76 void srpc_get_counters(srpc_counters_t *cnt)
78 spin_lock(&srpc_data.rpc_glock);
79 *cnt = srpc_data.rpc_counters;
80 spin_unlock(&srpc_data.rpc_glock);
83 void srpc_set_counters(const srpc_counters_t *cnt)
85 spin_lock(&srpc_data.rpc_glock);
86 srpc_data.rpc_counters = *cnt;
87 spin_unlock(&srpc_data.rpc_glock);
91 srpc_add_bulk_page(srpc_bulk_t *bk, struct page *pg, int i, int nob)
93 nob = min_t(int, nob, PAGE_CACHE_SIZE);
96 LASSERT(i >= 0 && i < bk->bk_niov);
98 bk->bk_iovs[i].kiov_offset = 0;
99 bk->bk_iovs[i].kiov_page = pg;
100 bk->bk_iovs[i].kiov_len = nob;
105 srpc_free_bulk(srpc_bulk_t *bk)
112 for (i = 0; i < bk->bk_niov; i++) {
113 pg = bk->bk_iovs[i].kiov_page;
120 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
125 srpc_alloc_bulk(int cpt, unsigned bulk_npg, unsigned bulk_len, int sink)
130 LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
132 LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
133 offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
135 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
139 memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
141 bk->bk_len = bulk_len;
142 bk->bk_niov = bulk_npg;
144 for (i = 0; i < bulk_npg; i++) {
148 pg = alloc_pages_node(cfs_cpt_spread_node(lnet_cpt_table(), cpt),
151 CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
156 nob = srpc_add_bulk_page(bk, pg, i, bulk_len);
168 spin_lock(&srpc_data.rpc_glock);
169 id = srpc_data.rpc_matchbits++;
170 spin_unlock(&srpc_data.rpc_glock);
175 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
176 struct srpc_service_cd *scd,
177 struct srpc_buffer *buffer)
179 memset(rpc, 0, sizeof(*rpc));
180 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
181 srpc_serv_is_framework(scd->scd_svc) ?
182 lst_sched_serial : lst_sched_test[scd->scd_cpt]);
184 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
187 rpc->srpc_reqstbuf = buffer;
188 rpc->srpc_peer = buffer->buf_peer;
189 rpc->srpc_self = buffer->buf_self;
190 LNetInvalidateHandle(&rpc->srpc_replymdh);
194 srpc_service_fini(struct srpc_service *svc)
196 struct srpc_service_cd *scd;
197 struct srpc_server_rpc *rpc;
198 struct srpc_buffer *buf;
202 if (svc->sv_cpt_data == NULL)
205 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
207 if (!list_empty(&scd->scd_buf_posted))
208 q = &scd->scd_buf_posted;
209 else if (!list_empty(&scd->scd_buf_blocked))
210 q = &scd->scd_buf_blocked;
214 while (!list_empty(q)) {
215 buf = list_entry(q->next,
218 list_del(&buf->buf_list);
219 LIBCFS_FREE(buf, sizeof(*buf));
223 LASSERT(list_empty(&scd->scd_rpc_active));
225 while (!list_empty(&scd->scd_rpc_free)) {
226 rpc = list_entry(scd->scd_rpc_free.next,
227 struct srpc_server_rpc,
229 list_del(&rpc->srpc_list);
230 LIBCFS_FREE(rpc, sizeof(*rpc));
234 cfs_percpt_free(svc->sv_cpt_data);
235 svc->sv_cpt_data = NULL;
239 srpc_service_nrpcs(struct srpc_service *svc)
241 int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
243 return srpc_serv_is_framework(svc) ?
244 max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
247 int srpc_add_buffer(struct swi_workitem *wi);
250 srpc_service_init(struct srpc_service *svc)
252 struct srpc_service_cd *scd;
253 struct srpc_server_rpc *rpc;
258 svc->sv_shuttingdown = 0;
260 svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
261 sizeof(struct srpc_service_cd));
262 if (svc->sv_cpt_data == NULL)
265 svc->sv_ncpts = srpc_serv_is_framework(svc) ?
266 1 : cfs_cpt_number(lnet_cpt_table());
267 nrpcs = srpc_service_nrpcs(svc);
269 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
272 spin_lock_init(&scd->scd_lock);
273 INIT_LIST_HEAD(&scd->scd_rpc_free);
274 INIT_LIST_HEAD(&scd->scd_rpc_active);
275 INIT_LIST_HEAD(&scd->scd_buf_posted);
276 INIT_LIST_HEAD(&scd->scd_buf_blocked);
278 scd->scd_ev.ev_data = scd;
279 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
281 /* NB: don't use lst_sched_serial for adding buffer,
282 * see details in srpc_service_add_buffers() */
283 swi_init_workitem(&scd->scd_buf_wi, scd,
284 srpc_add_buffer, lst_sched_test[i]);
286 if (i != 0 && srpc_serv_is_framework(svc)) {
287 /* NB: framework service only needs srpc_service_cd for
288 * one partition, but we allocate for all to make
289 * it easier to implement, it will waste a little
290 * memory but nobody should care about this */
294 for (j = 0; j < nrpcs; j++) {
295 LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
298 srpc_service_fini(svc);
301 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
309 srpc_add_service(struct srpc_service *sv)
313 LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
315 if (srpc_service_init(sv) != 0)
318 spin_lock(&srpc_data.rpc_glock);
320 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
322 if (srpc_data.rpc_services[id] != NULL) {
323 spin_unlock(&srpc_data.rpc_glock);
327 srpc_data.rpc_services[id] = sv;
328 spin_unlock(&srpc_data.rpc_glock);
330 CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
334 srpc_service_fini(sv);
339 srpc_remove_service(srpc_service_t *sv)
343 spin_lock(&srpc_data.rpc_glock);
345 if (srpc_data.rpc_services[id] != sv) {
346 spin_unlock(&srpc_data.rpc_glock);
350 srpc_data.rpc_services[id] = NULL;
351 spin_unlock(&srpc_data.rpc_glock);
356 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
357 int len, int options, lnet_process_id_t peer,
358 lnet_handle_md_t *mdh, srpc_event_t *ev)
362 lnet_handle_me_t meh;
364 rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
365 local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
367 CERROR("LNetMEAttach failed: %d\n", rc);
368 LASSERT(rc == -ENOMEM);
376 md.options = options;
377 md.eq_handle = srpc_data.rpc_lnet_eq;
379 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
381 CERROR("LNetMDAttach failed: %d\n", rc);
382 LASSERT(rc == -ENOMEM);
384 rc = LNetMEUnlink(meh);
390 "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
391 libcfs_id2str(peer), portal, matchbits);
396 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
397 int options, lnet_process_id_t peer, lnet_nid_t self,
398 lnet_handle_md_t *mdh, srpc_event_t *ev)
406 md.eq_handle = srpc_data.rpc_lnet_eq;
407 md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
408 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
410 rc = LNetMDBind(md, LNET_UNLINK, mdh);
412 CERROR("LNetMDBind failed: %d\n", rc);
413 LASSERT(rc == -ENOMEM);
417 /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
418 * they're only meaningful for MDs attached to an ME (i.e. passive
420 if ((options & LNET_MD_OP_PUT) != 0) {
421 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
422 portal, matchbits, 0, 0);
424 LASSERT((options & LNET_MD_OP_GET) != 0);
426 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
430 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
431 ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
432 libcfs_id2str(peer), portal, matchbits, rc);
434 /* The forthcoming unlink event will complete this operation
435 * with failure, so fall through and return success here.
437 rc = LNetMDUnlink(*mdh);
441 "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
442 libcfs_id2str(peer), portal, matchbits);
448 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
449 int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
451 return srpc_post_active_rdma(srpc_serv_portal(service), service,
452 buf, len, LNET_MD_OP_PUT, peer,
453 LNET_NID_ANY, mdh, ev);
457 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
458 lnet_handle_md_t *mdh, srpc_event_t *ev)
460 lnet_process_id_t any = {0};
462 any.nid = LNET_NID_ANY;
463 any.pid = LNET_PID_ANY;
465 return srpc_post_passive_rdma(srpc_serv_portal(service),
466 local, service, buf, len,
467 LNET_MD_OP_PUT, any, mdh, ev);
471 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
472 __must_hold(&scd->scd_lock)
474 struct srpc_service *sv = scd->scd_svc;
475 struct srpc_msg *msg = &buf->buf_msg;
478 LNetInvalidateHandle(&buf->buf_mdh);
479 list_add(&buf->buf_list, &scd->scd_buf_posted);
480 scd->scd_buf_nposted++;
481 spin_unlock(&scd->scd_lock);
483 rc = srpc_post_passive_rqtbuf(sv->sv_id,
484 !srpc_serv_is_framework(sv),
485 msg, sizeof(*msg), &buf->buf_mdh,
488 /* At this point, a RPC (new or delayed) may have arrived in
489 * msg and its event handler has been called. So we must add
490 * buf to scd_buf_posted _before_ dropping scd_lock */
492 spin_lock(&scd->scd_lock);
495 if (!sv->sv_shuttingdown)
498 spin_unlock(&scd->scd_lock);
499 /* srpc_shutdown_service might have tried to unlink me
500 * when my buf_mdh was still invalid */
501 LNetMDUnlink(buf->buf_mdh);
502 spin_lock(&scd->scd_lock);
506 scd->scd_buf_nposted--;
507 if (sv->sv_shuttingdown)
508 return rc; /* don't allow to change scd_buf_posted */
510 list_del(&buf->buf_list);
511 spin_unlock(&scd->scd_lock);
513 LIBCFS_FREE(buf, sizeof(*buf));
515 spin_lock(&scd->scd_lock);
520 srpc_add_buffer(struct swi_workitem *wi)
522 struct srpc_service_cd *scd = wi->swi_workitem.wi_data;
523 struct srpc_buffer *buf;
526 /* it's called by workitem scheduler threads, these threads
527 * should have been set CPT affinity, so buffers will be posted
528 * on CPT local list of Portal */
529 spin_lock(&scd->scd_lock);
531 while (scd->scd_buf_adjust > 0 &&
532 !scd->scd_svc->sv_shuttingdown) {
533 scd->scd_buf_adjust--; /* consume it */
534 scd->scd_buf_posting++;
536 spin_unlock(&scd->scd_lock);
538 LIBCFS_ALLOC(buf, sizeof(*buf));
540 CERROR("Failed to add new buf to service: %s\n",
541 scd->scd_svc->sv_name);
542 spin_lock(&scd->scd_lock);
547 spin_lock(&scd->scd_lock);
548 if (scd->scd_svc->sv_shuttingdown) {
549 spin_unlock(&scd->scd_lock);
550 LIBCFS_FREE(buf, sizeof(*buf));
552 spin_lock(&scd->scd_lock);
557 rc = srpc_service_post_buffer(scd, buf);
559 break; /* buf has been freed inside */
561 LASSERT(scd->scd_buf_posting > 0);
562 scd->scd_buf_posting--;
563 scd->scd_buf_total++;
564 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
568 scd->scd_buf_err_stamp = ktime_get_real_seconds();
569 scd->scd_buf_err = rc;
571 LASSERT(scd->scd_buf_posting > 0);
572 scd->scd_buf_posting--;
575 spin_unlock(&scd->scd_lock);
580 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
582 struct srpc_service_cd *scd;
586 LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
588 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
589 spin_lock(&scd->scd_lock);
591 scd->scd_buf_err = 0;
592 scd->scd_buf_err_stamp = 0;
593 scd->scd_buf_posting = 0;
594 scd->scd_buf_adjust = nbuffer;
595 /* start to post buffers */
596 swi_schedule_workitem(&scd->scd_buf_wi);
597 spin_unlock(&scd->scd_lock);
599 /* framework service only post buffer for one partition */
600 if (srpc_serv_is_framework(sv))
604 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
605 spin_lock(&scd->scd_lock);
607 * NB: srpc_service_add_buffers() can be called inside
608 * thread context of lst_sched_serial, and we don't normally
609 * allow to sleep inside thread context of WI scheduler
610 * because it will block current scheduler thread from doing
611 * anything else, even worse, it could deadlock if it's
612 * waiting on result from another WI of the same scheduler.
613 * However, it's safe at here because scd_buf_wi is scheduled
614 * by thread in a different WI scheduler (lst_sched_test),
615 * so we don't have any risk of deadlock, though this could
616 * block all WIs pending on lst_sched_serial for a moment
617 * which is not good but not fatal.
619 lst_wait_until(scd->scd_buf_err != 0 ||
620 (scd->scd_buf_adjust == 0 &&
621 scd->scd_buf_posting == 0),
622 scd->scd_lock, "waiting for adding buffer\n");
624 if (scd->scd_buf_err != 0 && rc == 0)
625 rc = scd->scd_buf_err;
627 spin_unlock(&scd->scd_lock);
634 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
636 struct srpc_service_cd *scd;
640 LASSERT(!sv->sv_shuttingdown);
642 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
643 spin_lock(&scd->scd_lock);
645 num = scd->scd_buf_total + scd->scd_buf_posting;
646 scd->scd_buf_adjust -= min(nbuffer, num);
648 spin_unlock(&scd->scd_lock);
652 /* returns 1 if sv has finished, otherwise 0 */
654 srpc_finish_service(struct srpc_service *sv)
656 struct srpc_service_cd *scd;
657 struct srpc_server_rpc *rpc;
660 LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
662 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
663 spin_lock(&scd->scd_lock);
664 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
665 spin_unlock(&scd->scd_lock);
669 if (scd->scd_buf_nposted > 0) {
670 CDEBUG(D_NET, "waiting for %d posted buffers to unlink",
671 scd->scd_buf_nposted);
672 spin_unlock(&scd->scd_lock);
676 if (list_empty(&scd->scd_rpc_active)) {
677 spin_unlock(&scd->scd_lock);
681 rpc = list_entry(scd->scd_rpc_active.next,
682 struct srpc_server_rpc, srpc_list);
683 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, wi %s scheduled %d running %d, ev fired %d type %d status %d lnet %d\n",
684 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
685 swi_state2str(rpc->srpc_wi.swi_state),
686 rpc->srpc_wi.swi_workitem.wi_scheduled,
687 rpc->srpc_wi.swi_workitem.wi_running,
688 rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
689 rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
690 spin_unlock(&scd->scd_lock);
694 /* no lock needed from now on */
695 srpc_service_fini(sv);
699 /* called with sv->sv_lock held */
701 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
702 __must_hold(&scd->scd_lock)
704 if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
705 if (srpc_service_post_buffer(scd, buf) != 0) {
706 CWARN("Failed to post %s buffer\n",
707 scd->scd_svc->sv_name);
712 /* service is shutting down, or we want to recycle some buffers */
713 scd->scd_buf_total--;
715 if (scd->scd_buf_adjust < 0) {
716 scd->scd_buf_adjust++;
717 if (scd->scd_buf_adjust < 0 &&
718 scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
720 "Try to recycle %d buffers but nothing left\n",
721 scd->scd_buf_adjust);
722 scd->scd_buf_adjust = 0;
726 spin_unlock(&scd->scd_lock);
727 LIBCFS_FREE(buf, sizeof(*buf));
728 spin_lock(&scd->scd_lock);
732 srpc_abort_service(struct srpc_service *sv)
734 struct srpc_service_cd *scd;
735 struct srpc_server_rpc *rpc;
738 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
739 sv->sv_id, sv->sv_name);
741 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
742 spin_lock(&scd->scd_lock);
744 /* schedule in-flight RPCs to notice the abort, NB:
745 * racing with incoming RPCs; complete fix should make test
746 * RPCs carry session ID in its headers */
747 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
748 rpc->srpc_aborted = 1;
749 swi_schedule_workitem(&rpc->srpc_wi);
752 spin_unlock(&scd->scd_lock);
757 srpc_shutdown_service(srpc_service_t *sv)
759 struct srpc_service_cd *scd;
760 struct srpc_server_rpc *rpc;
764 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
765 sv->sv_id, sv->sv_name);
767 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
768 spin_lock(&scd->scd_lock);
770 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
772 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
773 spin_unlock(&scd->scd_lock);
775 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
776 spin_lock(&scd->scd_lock);
778 /* schedule in-flight RPCs to notice the shutdown */
779 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
780 swi_schedule_workitem(&rpc->srpc_wi);
782 spin_unlock(&scd->scd_lock);
784 /* OK to traverse scd_buf_posted without lock, since no one
785 * touches scd_buf_posted now */
786 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
787 LNetMDUnlink(buf->buf_mdh);
792 srpc_send_request(srpc_client_rpc_t *rpc)
794 srpc_event_t *ev = &rpc->crpc_reqstev;
799 ev->ev_type = SRPC_REQUEST_SENT;
801 rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
802 &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
803 &rpc->crpc_reqstmdh, ev);
805 LASSERT(rc == -ENOMEM);
806 ev->ev_fired = 1; /* no more event expected */
812 srpc_prepare_reply(srpc_client_rpc_t *rpc)
814 srpc_event_t *ev = &rpc->crpc_replyev;
815 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
820 ev->ev_type = SRPC_REPLY_RCVD;
822 *id = srpc_next_id();
824 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
825 &rpc->crpc_replymsg, sizeof(srpc_msg_t),
826 LNET_MD_OP_PUT, rpc->crpc_dest,
827 &rpc->crpc_replymdh, ev);
829 LASSERT(rc == -ENOMEM);
830 ev->ev_fired = 1; /* no more event expected */
836 srpc_prepare_bulk(srpc_client_rpc_t *rpc)
838 srpc_bulk_t *bk = &rpc->crpc_bulk;
839 srpc_event_t *ev = &rpc->crpc_bulkev;
840 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
844 LASSERT(bk->bk_niov <= LNET_MAX_IOV);
846 if (bk->bk_niov == 0)
847 return 0; /* nothing to do */
849 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
854 ev->ev_type = SRPC_BULK_REQ_RCVD;
856 *id = srpc_next_id();
858 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
859 &bk->bk_iovs[0], bk->bk_niov, opt,
860 rpc->crpc_dest, &bk->bk_mdh, ev);
862 LASSERT(rc == -ENOMEM);
863 ev->ev_fired = 1; /* no more event expected */
869 srpc_do_bulk(srpc_server_rpc_t *rpc)
871 srpc_event_t *ev = &rpc->srpc_ev;
872 srpc_bulk_t *bk = rpc->srpc_bulk;
873 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
879 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
884 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
886 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
887 &bk->bk_iovs[0], bk->bk_niov, opt,
888 rpc->srpc_peer, rpc->srpc_self,
891 ev->ev_fired = 1; /* no more event expected */
895 /* only called from srpc_handle_rpc */
897 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
899 struct srpc_service_cd *scd = rpc->srpc_scd;
900 struct srpc_service *sv = scd->scd_svc;
901 srpc_buffer_t *buffer;
903 LASSERT(status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
905 rpc->srpc_status = status;
907 CDEBUG_LIMIT(status == 0 ? D_NET : D_NETERROR,
908 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
909 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
910 swi_state2str(rpc->srpc_wi.swi_state), status);
913 spin_lock(&srpc_data.rpc_glock);
914 srpc_data.rpc_counters.rpcs_dropped++;
915 spin_unlock(&srpc_data.rpc_glock);
918 if (rpc->srpc_done != NULL)
919 (*rpc->srpc_done) (rpc);
920 LASSERT(rpc->srpc_bulk == NULL);
922 spin_lock(&scd->scd_lock);
924 if (rpc->srpc_reqstbuf != NULL) {
925 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
926 * sv won't go away for scd_rpc_active must not be empty */
927 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
928 rpc->srpc_reqstbuf = NULL;
931 list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
934 * No one can schedule me now since:
935 * - I'm not on scd_rpc_active.
936 * - all LNet events have been fired.
937 * Cancel pending schedules and prevent future schedule attempts:
939 LASSERT(rpc->srpc_ev.ev_fired);
940 swi_exit_workitem(&rpc->srpc_wi);
942 if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
943 buffer = list_entry(scd->scd_buf_blocked.next,
944 srpc_buffer_t, buf_list);
945 list_del(&buffer->buf_list);
947 srpc_init_server_rpc(rpc, scd, buffer);
948 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
949 swi_schedule_workitem(&rpc->srpc_wi);
951 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
954 spin_unlock(&scd->scd_lock);
958 /* handles an incoming RPC */
960 srpc_handle_rpc(swi_workitem_t *wi)
962 struct srpc_server_rpc *rpc = wi->swi_workitem.wi_data;
963 struct srpc_service_cd *scd = rpc->srpc_scd;
964 struct srpc_service *sv = scd->scd_svc;
965 srpc_event_t *ev = &rpc->srpc_ev;
968 LASSERT(wi == &rpc->srpc_wi);
970 spin_lock(&scd->scd_lock);
972 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
973 spin_unlock(&scd->scd_lock);
975 if (rpc->srpc_bulk != NULL)
976 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
977 LNetMDUnlink(rpc->srpc_replymdh);
979 if (ev->ev_fired) { /* no more event, OK to finish */
980 srpc_server_rpc_done(rpc, -ESHUTDOWN);
986 spin_unlock(&scd->scd_lock);
988 switch (wi->swi_state) {
991 case SWI_STATE_NEWBORN: {
993 srpc_generic_reply_t *reply;
995 msg = &rpc->srpc_reqstbuf->buf_msg;
996 reply = &rpc->srpc_replymsg.msg_body.reply;
998 if (msg->msg_magic == 0) {
999 /* moaned already in srpc_lnet_ev_handler */
1000 srpc_server_rpc_done(rpc, EBADMSG);
1004 srpc_unpack_msg_hdr(msg);
1005 if (msg->msg_version != SRPC_MSG_VERSION) {
1006 CWARN("Version mismatch: %u, %u expected, from %s\n",
1007 msg->msg_version, SRPC_MSG_VERSION,
1008 libcfs_id2str(rpc->srpc_peer));
1009 reply->status = EPROTO;
1010 /* drop through and send reply */
1013 rc = (*sv->sv_handler)(rpc);
1014 LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1016 srpc_server_rpc_done(rpc, rc);
1021 wi->swi_state = SWI_STATE_BULK_STARTED;
1023 if (rpc->srpc_bulk != NULL) {
1024 rc = srpc_do_bulk(rpc);
1026 return 0; /* wait for bulk */
1028 LASSERT(ev->ev_fired);
1032 case SWI_STATE_BULK_STARTED:
1033 LASSERT(rpc->srpc_bulk == NULL || ev->ev_fired);
1035 if (rpc->srpc_bulk != NULL) {
1038 if (sv->sv_bulk_ready != NULL)
1039 rc = (*sv->sv_bulk_ready) (rpc, rc);
1042 srpc_server_rpc_done(rpc, rc);
1047 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1048 rc = srpc_send_reply(rpc);
1050 return 0; /* wait for reply */
1051 srpc_server_rpc_done(rpc, rc);
1054 case SWI_STATE_REPLY_SUBMITTED:
1055 if (!ev->ev_fired) {
1056 CERROR("RPC %p: bulk %p, service %d\n",
1057 rpc, rpc->srpc_bulk, sv->sv_id);
1058 CERROR("Event: status %d, type %d, lnet %d\n",
1059 ev->ev_status, ev->ev_type, ev->ev_lnet);
1060 LASSERT(ev->ev_fired);
1063 wi->swi_state = SWI_STATE_DONE;
1064 srpc_server_rpc_done(rpc, ev->ev_status);
1072 srpc_client_rpc_expired(void *data)
1074 srpc_client_rpc_t *rpc = data;
1076 CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1077 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1080 spin_lock(&rpc->crpc_lock);
1082 rpc->crpc_timeout = 0;
1083 srpc_abort_rpc(rpc, -ETIMEDOUT);
1085 spin_unlock(&rpc->crpc_lock);
1087 spin_lock(&srpc_data.rpc_glock);
1088 srpc_data.rpc_counters.rpcs_expired++;
1089 spin_unlock(&srpc_data.rpc_glock);
1093 srpc_add_client_rpc_timer(srpc_client_rpc_t *rpc)
1095 stt_timer_t *timer = &rpc->crpc_timer;
1097 if (rpc->crpc_timeout == 0)
1100 INIT_LIST_HEAD(&timer->stt_list);
1101 timer->stt_data = rpc;
1102 timer->stt_func = srpc_client_rpc_expired;
1103 timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1104 stt_add_timer(timer);
1109 * Called with rpc->crpc_lock held.
1111 * Upon exit the RPC expiry timer is not queued and the handler is not
1112 * running on any CPU. */
1114 srpc_del_client_rpc_timer(srpc_client_rpc_t *rpc)
1116 /* timer not planted or already exploded */
1117 if (rpc->crpc_timeout == 0)
1120 /* timer successfully defused */
1121 if (stt_del_timer(&rpc->crpc_timer))
1124 /* timer detonated, wait for it to explode */
1125 while (rpc->crpc_timeout != 0) {
1126 spin_unlock(&rpc->crpc_lock);
1130 spin_lock(&rpc->crpc_lock);
1135 srpc_client_rpc_done(srpc_client_rpc_t *rpc, int status)
1137 swi_workitem_t *wi = &rpc->crpc_wi;
1139 LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1141 spin_lock(&rpc->crpc_lock);
1143 rpc->crpc_closed = 1;
1144 if (rpc->crpc_status == 0)
1145 rpc->crpc_status = status;
1147 srpc_del_client_rpc_timer(rpc);
1149 CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1150 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1151 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1152 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1155 * No one can schedule me now since:
1156 * - RPC timer has been defused.
1157 * - all LNet events have been fired.
1158 * - crpc_closed has been set, preventing srpc_abort_rpc from
1160 * Cancel pending schedules and prevent future schedule attempts:
1162 LASSERT(!srpc_event_pending(rpc));
1163 swi_exit_workitem(wi);
1165 spin_unlock(&rpc->crpc_lock);
1167 (*rpc->crpc_done)(rpc);
1171 /* sends an outgoing RPC */
1173 srpc_send_rpc(swi_workitem_t *wi)
1176 srpc_client_rpc_t *rpc;
1180 LASSERT(wi != NULL);
1182 rpc = wi->swi_workitem.wi_data;
1184 LASSERT(rpc != NULL);
1185 LASSERT(wi == &rpc->crpc_wi);
1187 reply = &rpc->crpc_replymsg;
1188 do_bulk = rpc->crpc_bulk.bk_niov > 0;
1190 spin_lock(&rpc->crpc_lock);
1192 if (rpc->crpc_aborted) {
1193 spin_unlock(&rpc->crpc_lock);
1197 spin_unlock(&rpc->crpc_lock);
1199 switch (wi->swi_state) {
1202 case SWI_STATE_NEWBORN:
1203 LASSERT(!srpc_event_pending(rpc));
1205 rc = srpc_prepare_reply(rpc);
1207 srpc_client_rpc_done(rpc, rc);
1211 rc = srpc_prepare_bulk(rpc);
1215 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1216 rc = srpc_send_request(rpc);
1219 case SWI_STATE_REQUEST_SUBMITTED:
1220 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1221 * order; however, they're processed in a strict order:
1222 * rqt, rpy, and bulk. */
1223 if (!rpc->crpc_reqstev.ev_fired)
1226 rc = rpc->crpc_reqstev.ev_status;
1230 wi->swi_state = SWI_STATE_REQUEST_SENT;
1231 /* perhaps more events, fall thru */
1232 case SWI_STATE_REQUEST_SENT: {
1233 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1235 if (!rpc->crpc_replyev.ev_fired)
1238 rc = rpc->crpc_replyev.ev_status;
1242 srpc_unpack_msg_hdr(reply);
1243 if (reply->msg_type != type ||
1244 (reply->msg_magic != SRPC_MSG_MAGIC &&
1245 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1246 CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1247 libcfs_id2str(rpc->crpc_dest),
1248 reply->msg_type, type,
1249 reply->msg_magic, SRPC_MSG_MAGIC);
1254 if (do_bulk && reply->msg_body.reply.status != 0) {
1255 CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1256 reply->msg_body.reply.status,
1257 libcfs_id2str(rpc->crpc_dest));
1258 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1261 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1263 case SWI_STATE_REPLY_RECEIVED:
1264 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1267 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1269 /* Bulk buffer was unlinked due to remote error. Clear error
1270 * since reply buffer still contains valid data.
1271 * NB rpc->crpc_done shouldn't look into bulk data in case of
1273 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1274 rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1277 wi->swi_state = SWI_STATE_DONE;
1278 srpc_client_rpc_done(rpc, rc);
1283 spin_lock(&rpc->crpc_lock);
1284 srpc_abort_rpc(rpc, rc);
1285 spin_unlock(&rpc->crpc_lock);
1289 if (rpc->crpc_aborted) {
1290 LNetMDUnlink(rpc->crpc_reqstmdh);
1291 LNetMDUnlink(rpc->crpc_replymdh);
1292 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1294 if (!srpc_event_pending(rpc)) {
1295 srpc_client_rpc_done(rpc, -EINTR);
1303 srpc_create_client_rpc(lnet_process_id_t peer, int service,
1304 int nbulkiov, int bulklen,
1305 void (*rpc_done)(srpc_client_rpc_t *),
1306 void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1308 srpc_client_rpc_t *rpc;
1310 LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1311 crpc_bulk.bk_iovs[nbulkiov]));
1315 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1316 bulklen, rpc_done, rpc_fini, priv);
1320 /* called with rpc->crpc_lock held */
1322 srpc_abort_rpc(srpc_client_rpc_t *rpc, int why)
1326 if (rpc->crpc_aborted || /* already aborted */
1327 rpc->crpc_closed) /* callback imminent */
1331 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1332 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1333 swi_state2str(rpc->crpc_wi.swi_state), why);
1335 rpc->crpc_aborted = 1;
1336 rpc->crpc_status = why;
1337 swi_schedule_workitem(&rpc->crpc_wi);
1341 /* called with rpc->crpc_lock held */
1343 srpc_post_rpc(srpc_client_rpc_t *rpc)
1345 LASSERT(!rpc->crpc_aborted);
1346 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1348 CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1349 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1352 srpc_add_client_rpc_timer(rpc);
1353 swi_schedule_workitem(&rpc->crpc_wi);
1358 srpc_send_reply(struct srpc_server_rpc *rpc)
1360 srpc_event_t *ev = &rpc->srpc_ev;
1361 struct srpc_msg *msg = &rpc->srpc_replymsg;
1362 struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1363 struct srpc_service_cd *scd = rpc->srpc_scd;
1364 struct srpc_service *sv = scd->scd_svc;
1368 LASSERT(buffer != NULL);
1369 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1371 spin_lock(&scd->scd_lock);
1373 if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1374 /* Repost buffer before replying since test client
1375 * might send me another RPC once it gets the reply */
1376 if (srpc_service_post_buffer(scd, buffer) != 0)
1377 CWARN("Failed to repost %s buffer\n", sv->sv_name);
1378 rpc->srpc_reqstbuf = NULL;
1381 spin_unlock(&scd->scd_lock);
1385 ev->ev_type = SRPC_REPLY_SENT;
1387 msg->msg_magic = SRPC_MSG_MAGIC;
1388 msg->msg_version = SRPC_MSG_VERSION;
1389 msg->msg_type = srpc_service2reply(sv->sv_id);
1391 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1392 sizeof(*msg), LNET_MD_OP_PUT,
1393 rpc->srpc_peer, rpc->srpc_self,
1394 &rpc->srpc_replymdh, ev);
1396 ev->ev_fired = 1; /* no more event expected */
1400 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1402 srpc_lnet_ev_handler(lnet_event_t *ev)
1404 struct srpc_service_cd *scd;
1405 srpc_event_t *rpcev = ev->md.user_ptr;
1406 srpc_client_rpc_t *crpc;
1407 srpc_server_rpc_t *srpc;
1408 srpc_buffer_t *buffer;
1411 srpc_msg_type_t type;
1413 LASSERT(!in_interrupt());
1415 if (ev->status != 0) {
1416 spin_lock(&srpc_data.rpc_glock);
1417 srpc_data.rpc_counters.errors++;
1418 spin_unlock(&srpc_data.rpc_glock);
1421 rpcev->ev_lnet = ev->type;
1423 switch (rpcev->ev_type) {
1425 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1426 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1428 case SRPC_REQUEST_SENT:
1429 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1430 spin_lock(&srpc_data.rpc_glock);
1431 srpc_data.rpc_counters.rpcs_sent++;
1432 spin_unlock(&srpc_data.rpc_glock);
1434 case SRPC_REPLY_RCVD:
1435 case SRPC_BULK_REQ_RCVD:
1436 crpc = rpcev->ev_data;
1438 if (rpcev != &crpc->crpc_reqstev &&
1439 rpcev != &crpc->crpc_replyev &&
1440 rpcev != &crpc->crpc_bulkev) {
1441 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1442 rpcev, crpc, &crpc->crpc_reqstev,
1443 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1444 CERROR("Bad event: status %d, type %d, lnet %d\n",
1445 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1449 spin_lock(&crpc->crpc_lock);
1451 LASSERT(rpcev->ev_fired == 0);
1452 rpcev->ev_fired = 1;
1453 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1454 -EINTR : ev->status;
1455 swi_schedule_workitem(&crpc->crpc_wi);
1457 spin_unlock(&crpc->crpc_lock);
1460 case SRPC_REQUEST_RCVD:
1461 scd = rpcev->ev_data;
1464 LASSERT(rpcev == &scd->scd_ev);
1466 spin_lock(&scd->scd_lock);
1468 LASSERT(ev->unlinked);
1469 LASSERT(ev->type == LNET_EVENT_PUT ||
1470 ev->type == LNET_EVENT_UNLINK);
1471 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1472 sv->sv_shuttingdown);
1474 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1475 buffer->buf_peer = ev->initiator;
1476 buffer->buf_self = ev->target.nid;
1478 LASSERT(scd->scd_buf_nposted > 0);
1479 scd->scd_buf_nposted--;
1481 if (sv->sv_shuttingdown) {
1482 /* Leave buffer on scd->scd_buf_nposted since
1483 * srpc_finish_service needs to traverse it. */
1484 spin_unlock(&scd->scd_lock);
1488 if (scd->scd_buf_err_stamp != 0 &&
1489 scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1490 /* re-enable adding buffer */
1491 scd->scd_buf_err_stamp = 0;
1492 scd->scd_buf_err = 0;
1495 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1496 scd->scd_buf_adjust == 0 &&
1497 scd->scd_buf_nposted < scd->scd_buf_low) {
1498 scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1500 swi_schedule_workitem(&scd->scd_buf_wi);
1503 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1504 msg = &buffer->buf_msg;
1505 type = srpc_service2request(sv->sv_id);
1507 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1508 (msg->msg_type != type &&
1509 msg->msg_type != __swab32(type)) ||
1510 (msg->msg_magic != SRPC_MSG_MAGIC &&
1511 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1512 CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1513 sv->sv_name, libcfs_id2str(ev->initiator),
1514 ev->status, ev->mlength,
1515 msg->msg_type, msg->msg_magic);
1517 /* NB can't call srpc_service_recycle_buffer here since
1518 * it may call LNetM[DE]Attach. The invalid magic tells
1519 * srpc_handle_rpc to drop this RPC */
1523 if (!list_empty(&scd->scd_rpc_free)) {
1524 srpc = list_entry(scd->scd_rpc_free.next,
1525 struct srpc_server_rpc,
1527 list_del(&srpc->srpc_list);
1529 srpc_init_server_rpc(srpc, scd, buffer);
1530 list_add_tail(&srpc->srpc_list,
1531 &scd->scd_rpc_active);
1532 swi_schedule_workitem(&srpc->srpc_wi);
1534 list_add_tail(&buffer->buf_list,
1535 &scd->scd_buf_blocked);
1538 spin_unlock(&scd->scd_lock);
1540 spin_lock(&srpc_data.rpc_glock);
1541 srpc_data.rpc_counters.rpcs_rcvd++;
1542 spin_unlock(&srpc_data.rpc_glock);
1545 case SRPC_BULK_GET_RPLD:
1546 LASSERT(ev->type == LNET_EVENT_SEND ||
1547 ev->type == LNET_EVENT_REPLY ||
1548 ev->type == LNET_EVENT_UNLINK);
1551 break; /* wait for final event */
1553 case SRPC_BULK_PUT_SENT:
1554 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1555 spin_lock(&srpc_data.rpc_glock);
1557 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1558 srpc_data.rpc_counters.bulk_get += ev->mlength;
1560 srpc_data.rpc_counters.bulk_put += ev->mlength;
1562 spin_unlock(&srpc_data.rpc_glock);
1564 case SRPC_REPLY_SENT:
1565 srpc = rpcev->ev_data;
1566 scd = srpc->srpc_scd;
1568 LASSERT(rpcev == &srpc->srpc_ev);
1570 spin_lock(&scd->scd_lock);
1572 rpcev->ev_fired = 1;
1573 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1574 -EINTR : ev->status;
1575 swi_schedule_workitem(&srpc->srpc_wi);
1577 spin_unlock(&scd->scd_lock);
1587 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1588 spin_lock_init(&srpc_data.rpc_glock);
1590 /* 1 second pause to avoid timestamp reuse */
1591 set_current_state(TASK_UNINTERRUPTIBLE);
1592 schedule_timeout(cfs_time_seconds(1));
1593 srpc_data.rpc_matchbits = ((__u64)ktime_get_real_seconds()) << 48;
1595 srpc_data.rpc_state = SRPC_STATE_NONE;
1597 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1599 CERROR("LNetNIInit() has failed: %d\n", rc);
1603 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1605 LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1606 rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1608 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1612 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1614 rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1617 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1625 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1637 state = srpc_data.rpc_state;
1638 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1643 case SRPC_STATE_RUNNING:
1644 spin_lock(&srpc_data.rpc_glock);
1646 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1647 srpc_service_t *sv = srpc_data.rpc_services[i];
1649 LASSERTF(sv == NULL,
1650 "service not empty: id %d, name %s\n",
1654 spin_unlock(&srpc_data.rpc_glock);
1658 case SRPC_STATE_EQ_INIT:
1659 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1660 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1662 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1663 LASSERT(rc == 0); /* the EQ should have no user by now */
1665 case SRPC_STATE_NI_INIT: