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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * Author: Isaac Huang <isaac@clusterfs.com>
36 * 2012-05-13: Liang Zhen <liang@whamcloud.com>
37 * - percpt data for service to improve smp performance
41 #define DEBUG_SUBSYSTEM S_LNET
53 static struct smoketest_rpc {
54 spinlock_t rpc_glock; /* global lock */
55 struct srpc_service *rpc_services[SRPC_SERVICE_MAX_ID + 1];
56 struct lnet_handle_eq rpc_lnet_eq; /* _the_ LNet event queue */
57 enum srpc_state rpc_state;
58 struct srpc_counters rpc_counters;
59 __u64 rpc_matchbits; /* matchbits counter */
63 srpc_serv_portal(int svc_id)
65 return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
66 SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
70 static int srpc_handle_rpc(struct swi_workitem *wi);
72 void srpc_get_counters(struct srpc_counters *cnt)
74 spin_lock(&srpc_data.rpc_glock);
75 *cnt = srpc_data.rpc_counters;
76 spin_unlock(&srpc_data.rpc_glock);
79 void srpc_set_counters(const struct srpc_counters *cnt)
81 spin_lock(&srpc_data.rpc_glock);
82 srpc_data.rpc_counters = *cnt;
83 spin_unlock(&srpc_data.rpc_glock);
87 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
90 LASSERT(off < PAGE_SIZE);
91 LASSERT(nob > 0 && nob <= PAGE_SIZE);
93 bk->bk_iovs[i].kiov_offset = off;
94 bk->bk_iovs[i].kiov_page = pg;
95 bk->bk_iovs[i].kiov_len = nob;
100 srpc_free_bulk(struct srpc_bulk *bk)
105 LASSERT (bk != NULL);
107 for (i = 0; i < bk->bk_niov; i++) {
108 pg = bk->bk_iovs[i].kiov_page;
109 if (pg == NULL) break;
114 LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
119 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
120 unsigned bulk_len, int sink)
122 struct srpc_bulk *bk;
125 LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
127 LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
128 offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
130 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
134 memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
136 bk->bk_len = bulk_len;
137 bk->bk_niov = bulk_npg;
139 for (i = 0; i < bulk_npg; i++) {
143 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
145 CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
150 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
153 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
166 spin_lock(&srpc_data.rpc_glock);
167 id = srpc_data.rpc_matchbits++;
168 spin_unlock(&srpc_data.rpc_glock);
173 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
174 struct srpc_service_cd *scd,
175 struct srpc_buffer *buffer)
177 memset(rpc, 0, sizeof(*rpc));
178 swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
179 srpc_serv_is_framework(scd->scd_svc) ?
180 lst_sched_serial : lst_sched_test[scd->scd_cpt]);
182 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
185 rpc->srpc_reqstbuf = buffer;
186 rpc->srpc_peer = buffer->buf_peer;
187 rpc->srpc_self = buffer->buf_self;
188 LNetInvalidateMDHandle(&rpc->srpc_replymdh);
192 srpc_service_fini(struct srpc_service *svc)
194 struct srpc_service_cd *scd;
195 struct srpc_server_rpc *rpc;
196 struct srpc_buffer *buf;
200 if (svc->sv_cpt_data == NULL)
203 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
205 if (!list_empty(&scd->scd_buf_posted))
206 q = &scd->scd_buf_posted;
207 else if (!list_empty(&scd->scd_buf_blocked))
208 q = &scd->scd_buf_blocked;
212 while (!list_empty(q)) {
213 buf = list_entry(q->next,
216 list_del(&buf->buf_list);
217 LIBCFS_FREE(buf, sizeof(*buf));
221 LASSERT(list_empty(&scd->scd_rpc_active));
223 while (!list_empty(&scd->scd_rpc_free)) {
224 rpc = list_entry(scd->scd_rpc_free.next,
225 struct srpc_server_rpc,
227 list_del(&rpc->srpc_list);
228 LIBCFS_FREE(rpc, sizeof(*rpc));
232 cfs_percpt_free(svc->sv_cpt_data);
233 svc->sv_cpt_data = NULL;
237 srpc_service_nrpcs(struct srpc_service *svc)
239 int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
241 return srpc_serv_is_framework(svc) ?
242 max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
245 int srpc_add_buffer(struct swi_workitem *wi);
248 srpc_service_init(struct srpc_service *svc)
250 struct srpc_service_cd *scd;
251 struct srpc_server_rpc *rpc;
256 svc->sv_shuttingdown = 0;
258 svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
259 sizeof(struct srpc_service_cd));
260 if (svc->sv_cpt_data == NULL)
263 svc->sv_ncpts = srpc_serv_is_framework(svc) ?
264 1 : cfs_cpt_number(lnet_cpt_table());
265 nrpcs = srpc_service_nrpcs(svc);
267 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
270 spin_lock_init(&scd->scd_lock);
271 INIT_LIST_HEAD(&scd->scd_rpc_free);
272 INIT_LIST_HEAD(&scd->scd_rpc_active);
273 INIT_LIST_HEAD(&scd->scd_buf_posted);
274 INIT_LIST_HEAD(&scd->scd_buf_blocked);
276 scd->scd_ev.ev_data = scd;
277 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
279 /* NB: don't use lst_sched_serial for adding buffer,
280 * see details in srpc_service_add_buffers() */
281 swi_init_workitem(&scd->scd_buf_wi,
282 srpc_add_buffer, lst_sched_test[i]);
284 if (i != 0 && srpc_serv_is_framework(svc)) {
285 /* NB: framework service only needs srpc_service_cd for
286 * one partition, but we allocate for all to make
287 * it easier to implement, it will waste a little
288 * memory but nobody should care about this */
292 for (j = 0; j < nrpcs; j++) {
293 LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
296 srpc_service_fini(svc);
299 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
307 srpc_add_service(struct srpc_service *sv)
311 LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
313 if (srpc_service_init(sv) != 0)
316 spin_lock(&srpc_data.rpc_glock);
318 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
320 if (srpc_data.rpc_services[id] != NULL) {
321 spin_unlock(&srpc_data.rpc_glock);
325 srpc_data.rpc_services[id] = sv;
326 spin_unlock(&srpc_data.rpc_glock);
328 CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
332 srpc_service_fini(sv);
337 srpc_remove_service(struct srpc_service *sv)
341 spin_lock(&srpc_data.rpc_glock);
343 if (srpc_data.rpc_services[id] != sv) {
344 spin_unlock(&srpc_data.rpc_glock);
348 srpc_data.rpc_services[id] = NULL;
349 spin_unlock(&srpc_data.rpc_glock);
354 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
355 int len, int options, struct lnet_process_id peer,
356 struct lnet_handle_md *mdh, struct srpc_event *ev)
360 struct lnet_handle_me meh;
362 rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
363 local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
365 CERROR ("LNetMEAttach failed: %d\n", rc);
366 LASSERT (rc == -ENOMEM);
374 md.options = options;
375 md.eq_handle = srpc_data.rpc_lnet_eq;
377 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
379 CERROR ("LNetMDAttach failed: %d\n", rc);
380 LASSERT (rc == -ENOMEM);
382 rc = LNetMEUnlink(meh);
388 "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
389 libcfs_id2str(peer), portal, matchbits);
394 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
395 int options, struct lnet_process_id peer,
396 lnet_nid_t self, struct lnet_handle_md *mdh,
397 struct srpc_event *ev)
405 md.eq_handle = srpc_data.rpc_lnet_eq;
406 md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
407 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
409 rc = LNetMDBind(md, LNET_UNLINK, mdh);
411 CERROR ("LNetMDBind failed: %d\n", rc);
412 LASSERT (rc == -ENOMEM);
416 /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
417 * they're only meaningful for MDs attached to an ME (i.e. passive
419 if ((options & LNET_MD_OP_PUT) != 0) {
420 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
421 portal, matchbits, 0, 0);
423 LASSERT ((options & LNET_MD_OP_GET) != 0);
425 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0, false);
429 CERROR ("LNet%s(%s, %d, %lld) failed: %d\n",
430 ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
431 libcfs_id2str(peer), portal, matchbits, rc);
433 /* The forthcoming unlink event will complete this operation
434 * with failure, so fall through and return success here.
436 rc = LNetMDUnlink(*mdh);
440 "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
441 libcfs_id2str(peer), portal, matchbits);
447 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
448 struct lnet_handle_md *mdh, struct srpc_event *ev)
450 struct lnet_process_id any = {0};
452 any.nid = LNET_NID_ANY;
453 any.pid = LNET_PID_ANY;
455 return srpc_post_passive_rdma(srpc_serv_portal(service),
456 local, service, buf, len,
457 LNET_MD_OP_PUT, any, mdh, ev);
461 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
462 __must_hold(&scd->scd_lock)
464 struct srpc_service *sv = scd->scd_svc;
465 struct srpc_msg *msg = &buf->buf_msg;
468 LNetInvalidateMDHandle(&buf->buf_mdh);
469 list_add(&buf->buf_list, &scd->scd_buf_posted);
470 scd->scd_buf_nposted++;
471 spin_unlock(&scd->scd_lock);
473 rc = srpc_post_passive_rqtbuf(sv->sv_id,
474 !srpc_serv_is_framework(sv),
475 msg, sizeof(*msg), &buf->buf_mdh,
478 /* At this point, a RPC (new or delayed) may have arrived in
479 * msg and its event handler has been called. So we must add
480 * buf to scd_buf_posted _before_ dropping scd_lock */
482 spin_lock(&scd->scd_lock);
485 if (!sv->sv_shuttingdown)
488 spin_unlock(&scd->scd_lock);
489 /* srpc_shutdown_service might have tried to unlink me
490 * when my buf_mdh was still invalid */
491 LNetMDUnlink(buf->buf_mdh);
492 spin_lock(&scd->scd_lock);
496 scd->scd_buf_nposted--;
497 if (sv->sv_shuttingdown)
498 return rc; /* don't allow to change scd_buf_posted */
500 list_del(&buf->buf_list);
501 spin_unlock(&scd->scd_lock);
503 LIBCFS_FREE(buf, sizeof(*buf));
505 spin_lock(&scd->scd_lock);
510 srpc_add_buffer(struct swi_workitem *wi)
512 struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd, scd_buf_wi);
513 struct srpc_buffer *buf;
516 /* it's called by workitem scheduler threads, these threads
517 * should have been set CPT affinity, so buffers will be posted
518 * on CPT local list of Portal */
519 spin_lock(&scd->scd_lock);
521 while (scd->scd_buf_adjust > 0 &&
522 !scd->scd_svc->sv_shuttingdown) {
523 scd->scd_buf_adjust--; /* consume it */
524 scd->scd_buf_posting++;
526 spin_unlock(&scd->scd_lock);
528 LIBCFS_ALLOC(buf, sizeof(*buf));
530 CERROR("Failed to add new buf to service: %s\n",
531 scd->scd_svc->sv_name);
532 spin_lock(&scd->scd_lock);
537 spin_lock(&scd->scd_lock);
538 if (scd->scd_svc->sv_shuttingdown) {
539 spin_unlock(&scd->scd_lock);
540 LIBCFS_FREE(buf, sizeof(*buf));
542 spin_lock(&scd->scd_lock);
547 rc = srpc_service_post_buffer(scd, buf);
549 break; /* buf has been freed inside */
551 LASSERT(scd->scd_buf_posting > 0);
552 scd->scd_buf_posting--;
553 scd->scd_buf_total++;
554 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
558 scd->scd_buf_err_stamp = ktime_get_real_seconds();
559 scd->scd_buf_err = rc;
561 LASSERT(scd->scd_buf_posting > 0);
562 scd->scd_buf_posting--;
565 spin_unlock(&scd->scd_lock);
570 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
572 struct srpc_service_cd *scd;
576 LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
578 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
579 spin_lock(&scd->scd_lock);
581 scd->scd_buf_err = 0;
582 scd->scd_buf_err_stamp = 0;
583 scd->scd_buf_posting = 0;
584 scd->scd_buf_adjust = nbuffer;
585 /* start to post buffers */
586 swi_schedule_workitem(&scd->scd_buf_wi);
587 spin_unlock(&scd->scd_lock);
589 /* framework service only post buffer for one partition */
590 if (srpc_serv_is_framework(sv))
594 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
595 spin_lock(&scd->scd_lock);
597 * NB: srpc_service_add_buffers() can be called inside
598 * thread context of lst_sched_serial, and we don't normally
599 * allow to sleep inside thread context of WI scheduler
600 * because it will block current scheduler thread from doing
601 * anything else, even worse, it could deadlock if it's
602 * waiting on result from another WI of the same scheduler.
603 * However, it's safe at here because scd_buf_wi is scheduled
604 * by thread in a different WI scheduler (lst_sched_test),
605 * so we don't have any risk of deadlock, though this could
606 * block all WIs pending on lst_sched_serial for a moment
607 * which is not good but not fatal.
609 lst_wait_until(scd->scd_buf_err != 0 ||
610 (scd->scd_buf_adjust == 0 &&
611 scd->scd_buf_posting == 0),
612 scd->scd_lock, "waiting for adding buffer\n");
614 if (scd->scd_buf_err != 0 && rc == 0)
615 rc = scd->scd_buf_err;
617 spin_unlock(&scd->scd_lock);
624 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
626 struct srpc_service_cd *scd;
630 LASSERT(!sv->sv_shuttingdown);
632 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
633 spin_lock(&scd->scd_lock);
635 num = scd->scd_buf_total + scd->scd_buf_posting;
636 scd->scd_buf_adjust -= min(nbuffer, num);
638 spin_unlock(&scd->scd_lock);
642 /* returns 1 if sv has finished, otherwise 0 */
644 srpc_finish_service(struct srpc_service *sv)
646 struct srpc_service_cd *scd;
647 struct srpc_server_rpc *rpc;
650 LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
652 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
653 spin_lock(&scd->scd_lock);
654 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
655 spin_unlock(&scd->scd_lock);
659 if (scd->scd_buf_nposted > 0) {
660 CDEBUG(D_NET, "waiting for %d posted buffers to "
661 "unlink\n", scd->scd_buf_nposted);
662 spin_unlock(&scd->scd_lock);
666 if (list_empty(&scd->scd_rpc_active)) {
667 spin_unlock(&scd->scd_lock);
671 rpc = list_entry(scd->scd_rpc_active.next,
672 struct srpc_server_rpc, srpc_list);
673 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
674 "wi %s scheduled %d running %d, "
675 "ev fired %d type %d status %d lnet %d\n",
676 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
677 swi_state2str(rpc->srpc_wi.swi_state),
678 rpc->srpc_wi.swi_workitem.wi_scheduled,
679 rpc->srpc_wi.swi_workitem.wi_running,
680 rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
681 rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
682 spin_unlock(&scd->scd_lock);
686 /* no lock needed from now on */
687 srpc_service_fini(sv);
691 /* called with sv->sv_lock held */
693 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
694 struct srpc_buffer *buf)
695 __must_hold(&scd->scd_lock)
697 if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
698 if (srpc_service_post_buffer(scd, buf) != 0) {
699 CWARN("Failed to post %s buffer\n",
700 scd->scd_svc->sv_name);
705 /* service is shutting down, or we want to recycle some buffers */
706 scd->scd_buf_total--;
708 if (scd->scd_buf_adjust < 0) {
709 scd->scd_buf_adjust++;
710 if (scd->scd_buf_adjust < 0 &&
711 scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
713 "Try to recyle %d buffers but nothing left\n",
714 scd->scd_buf_adjust);
715 scd->scd_buf_adjust = 0;
719 spin_unlock(&scd->scd_lock);
720 LIBCFS_FREE(buf, sizeof(*buf));
721 spin_lock(&scd->scd_lock);
725 srpc_abort_service(struct srpc_service *sv)
727 struct srpc_service_cd *scd;
728 struct srpc_server_rpc *rpc;
731 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
732 sv->sv_id, sv->sv_name);
734 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
735 spin_lock(&scd->scd_lock);
737 /* schedule in-flight RPCs to notice the abort, NB:
738 * racing with incoming RPCs; complete fix should make test
739 * RPCs carry session ID in its headers */
740 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
741 rpc->srpc_aborted = 1;
742 swi_schedule_workitem(&rpc->srpc_wi);
745 spin_unlock(&scd->scd_lock);
750 srpc_shutdown_service(struct srpc_service *sv)
752 struct srpc_service_cd *scd;
753 struct srpc_server_rpc *rpc;
754 struct srpc_buffer *buf;
757 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
758 sv->sv_id, sv->sv_name);
760 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
761 spin_lock(&scd->scd_lock);
763 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
765 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
766 spin_unlock(&scd->scd_lock);
768 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
769 spin_lock(&scd->scd_lock);
771 /* schedule in-flight RPCs to notice the shutdown */
772 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
773 swi_schedule_workitem(&rpc->srpc_wi);
775 spin_unlock(&scd->scd_lock);
777 /* OK to traverse scd_buf_posted without lock, since no one
778 * touches scd_buf_posted now */
779 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
780 LNetMDUnlink(buf->buf_mdh);
785 srpc_send_request(struct srpc_client_rpc *rpc)
787 struct srpc_event *ev = &rpc->crpc_reqstev;
792 ev->ev_type = SRPC_REQUEST_SENT;
794 rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
795 rpc->crpc_service, &rpc->crpc_reqstmsg,
796 sizeof(struct srpc_msg), LNET_MD_OP_PUT,
797 rpc->crpc_dest, LNET_NID_ANY,
798 &rpc->crpc_reqstmdh, ev);
800 LASSERT (rc == -ENOMEM);
801 ev->ev_fired = 1; /* no more event expected */
807 srpc_prepare_reply(struct srpc_client_rpc *rpc)
809 struct srpc_event *ev = &rpc->crpc_replyev;
810 u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
815 ev->ev_type = SRPC_REPLY_RCVD;
817 *id = srpc_next_id();
819 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
821 sizeof(struct srpc_msg),
822 LNET_MD_OP_PUT, rpc->crpc_dest,
823 &rpc->crpc_replymdh, ev);
825 LASSERT (rc == -ENOMEM);
826 ev->ev_fired = 1; /* no more event expected */
832 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
834 struct srpc_bulk *bk = &rpc->crpc_bulk;
835 struct srpc_event *ev = &rpc->crpc_bulkev;
836 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
840 LASSERT (bk->bk_niov <= LNET_MAX_IOV);
842 if (bk->bk_niov == 0) return 0; /* nothing to do */
844 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
849 ev->ev_type = SRPC_BULK_REQ_RCVD;
851 *id = srpc_next_id();
853 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
854 &bk->bk_iovs[0], bk->bk_niov, opt,
855 rpc->crpc_dest, &bk->bk_mdh, ev);
857 LASSERT (rc == -ENOMEM);
858 ev->ev_fired = 1; /* no more event expected */
864 srpc_do_bulk(struct srpc_server_rpc *rpc)
866 struct srpc_event *ev = &rpc->srpc_ev;
867 struct srpc_bulk *bk = rpc->srpc_bulk;
868 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
872 LASSERT (bk != NULL);
874 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
879 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
881 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
882 &bk->bk_iovs[0], bk->bk_niov, opt,
883 rpc->srpc_peer, rpc->srpc_self,
886 ev->ev_fired = 1; /* no more event expected */
890 /* only called from srpc_handle_rpc */
892 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
894 struct srpc_service_cd *scd = rpc->srpc_scd;
895 struct srpc_service *sv = scd->scd_svc;
896 struct srpc_buffer *buffer;
898 LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
900 rpc->srpc_status = status;
902 CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
903 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
904 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
905 swi_state2str(rpc->srpc_wi.swi_state), status);
908 spin_lock(&srpc_data.rpc_glock);
909 srpc_data.rpc_counters.rpcs_dropped++;
910 spin_unlock(&srpc_data.rpc_glock);
913 if (rpc->srpc_done != NULL)
914 (*rpc->srpc_done) (rpc);
915 LASSERT(rpc->srpc_bulk == NULL);
917 spin_lock(&scd->scd_lock);
919 if (rpc->srpc_reqstbuf != NULL) {
920 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
921 * sv won't go away for scd_rpc_active must not be empty */
922 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
923 rpc->srpc_reqstbuf = NULL;
926 list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
929 * No one can schedule me now since:
930 * - I'm not on scd_rpc_active.
931 * - all LNet events have been fired.
932 * Cancel pending schedules and prevent future schedule attempts:
934 LASSERT(rpc->srpc_ev.ev_fired);
935 swi_exit_workitem(&rpc->srpc_wi);
937 if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
938 buffer = list_entry(scd->scd_buf_blocked.next,
939 struct srpc_buffer, buf_list);
940 list_del(&buffer->buf_list);
942 srpc_init_server_rpc(rpc, scd, buffer);
943 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
944 swi_schedule_workitem(&rpc->srpc_wi);
946 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
949 spin_unlock(&scd->scd_lock);
953 /* handles an incoming RPC */
954 static int srpc_handle_rpc(struct swi_workitem *wi)
956 struct srpc_server_rpc *rpc = container_of(wi, struct srpc_server_rpc, srpc_wi);
957 struct srpc_service_cd *scd = rpc->srpc_scd;
958 struct srpc_service *sv = scd->scd_svc;
959 struct srpc_event *ev = &rpc->srpc_ev;
962 LASSERT(wi == &rpc->srpc_wi);
964 spin_lock(&scd->scd_lock);
966 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
967 spin_unlock(&scd->scd_lock);
969 if (rpc->srpc_bulk != NULL)
970 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
971 LNetMDUnlink(rpc->srpc_replymdh);
973 if (ev->ev_fired) { /* no more event, OK to finish */
974 srpc_server_rpc_done(rpc, -ESHUTDOWN);
980 spin_unlock(&scd->scd_lock);
982 switch (wi->swi_state) {
985 case SWI_STATE_NEWBORN: {
986 struct srpc_msg *msg;
987 struct srpc_generic_reply *reply;
989 msg = &rpc->srpc_reqstbuf->buf_msg;
990 reply = &rpc->srpc_replymsg.msg_body.reply;
992 if (msg->msg_magic == 0) {
993 /* moaned already in srpc_lnet_ev_handler */
994 srpc_server_rpc_done(rpc, EBADMSG);
998 srpc_unpack_msg_hdr(msg);
999 if (msg->msg_version != SRPC_MSG_VERSION) {
1000 CWARN("Version mismatch: %u, %u expected, from %s\n",
1001 msg->msg_version, SRPC_MSG_VERSION,
1002 libcfs_id2str(rpc->srpc_peer));
1003 reply->status = EPROTO;
1004 /* drop through and send reply */
1007 rc = (*sv->sv_handler)(rpc);
1008 LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1010 srpc_server_rpc_done(rpc, rc);
1015 wi->swi_state = SWI_STATE_BULK_STARTED;
1017 if (rpc->srpc_bulk != NULL) {
1018 rc = srpc_do_bulk(rpc);
1020 return 0; /* wait for bulk */
1022 LASSERT (ev->ev_fired);
1026 case SWI_STATE_BULK_STARTED:
1027 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1029 if (rpc->srpc_bulk != NULL) {
1032 if (sv->sv_bulk_ready != NULL)
1033 rc = (*sv->sv_bulk_ready) (rpc, rc);
1036 srpc_server_rpc_done(rpc, rc);
1041 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1042 rc = srpc_send_reply(rpc);
1044 return 0; /* wait for reply */
1045 srpc_server_rpc_done(rpc, rc);
1048 case SWI_STATE_REPLY_SUBMITTED:
1049 if (!ev->ev_fired) {
1050 CERROR("RPC %p: bulk %p, service %d\n",
1051 rpc, rpc->srpc_bulk, sv->sv_id);
1052 CERROR("Event: status %d, type %d, lnet %d\n",
1053 ev->ev_status, ev->ev_type, ev->ev_lnet);
1054 LASSERT (ev->ev_fired);
1057 wi->swi_state = SWI_STATE_DONE;
1058 srpc_server_rpc_done(rpc, ev->ev_status);
1066 srpc_client_rpc_expired (void *data)
1068 struct srpc_client_rpc *rpc = data;
1070 CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1071 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1074 spin_lock(&rpc->crpc_lock);
1076 rpc->crpc_timeout = 0;
1077 srpc_abort_rpc(rpc, -ETIMEDOUT);
1079 spin_unlock(&rpc->crpc_lock);
1081 spin_lock(&srpc_data.rpc_glock);
1082 srpc_data.rpc_counters.rpcs_expired++;
1083 spin_unlock(&srpc_data.rpc_glock);
1087 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1089 struct stt_timer *timer = &rpc->crpc_timer;
1091 if (rpc->crpc_timeout == 0)
1094 INIT_LIST_HEAD(&timer->stt_list);
1095 timer->stt_data = rpc;
1096 timer->stt_func = srpc_client_rpc_expired;
1097 timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1098 stt_add_timer(timer);
1103 * Called with rpc->crpc_lock held.
1105 * Upon exit the RPC expiry timer is not queued and the handler is not
1106 * running on any CPU. */
1108 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1110 /* timer not planted or already exploded */
1111 if (rpc->crpc_timeout == 0)
1114 /* timer successfully defused */
1115 if (stt_del_timer(&rpc->crpc_timer))
1118 /* timer detonated, wait for it to explode */
1119 while (rpc->crpc_timeout != 0) {
1120 spin_unlock(&rpc->crpc_lock);
1124 spin_lock(&rpc->crpc_lock);
1129 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1131 struct swi_workitem *wi = &rpc->crpc_wi;
1133 LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1135 spin_lock(&rpc->crpc_lock);
1137 rpc->crpc_closed = 1;
1138 if (rpc->crpc_status == 0)
1139 rpc->crpc_status = status;
1141 srpc_del_client_rpc_timer(rpc);
1143 CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1144 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1145 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1146 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1149 * No one can schedule me now since:
1150 * - RPC timer has been defused.
1151 * - all LNet events have been fired.
1152 * - crpc_closed has been set, preventing srpc_abort_rpc from
1154 * Cancel pending schedules and prevent future schedule attempts:
1156 LASSERT (!srpc_event_pending(rpc));
1157 swi_exit_workitem(wi);
1159 spin_unlock(&rpc->crpc_lock);
1161 (*rpc->crpc_done)(rpc);
1165 /* sends an outgoing RPC */
1167 srpc_send_rpc(struct swi_workitem *wi)
1170 struct srpc_client_rpc *rpc;
1171 struct srpc_msg *reply;
1174 LASSERT(wi != NULL);
1176 rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1178 LASSERT (rpc != NULL);
1179 LASSERT (wi == &rpc->crpc_wi);
1181 reply = &rpc->crpc_replymsg;
1182 do_bulk = rpc->crpc_bulk.bk_niov > 0;
1184 spin_lock(&rpc->crpc_lock);
1186 if (rpc->crpc_aborted) {
1187 spin_unlock(&rpc->crpc_lock);
1191 spin_unlock(&rpc->crpc_lock);
1193 switch (wi->swi_state) {
1196 case SWI_STATE_NEWBORN:
1197 LASSERT (!srpc_event_pending(rpc));
1199 rc = srpc_prepare_reply(rpc);
1201 srpc_client_rpc_done(rpc, rc);
1205 rc = srpc_prepare_bulk(rpc);
1208 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1209 rc = srpc_send_request(rpc);
1212 case SWI_STATE_REQUEST_SUBMITTED:
1213 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1214 * order; however, they're processed in a strict order:
1215 * rqt, rpy, and bulk. */
1216 if (!rpc->crpc_reqstev.ev_fired) break;
1218 rc = rpc->crpc_reqstev.ev_status;
1221 wi->swi_state = SWI_STATE_REQUEST_SENT;
1222 /* perhaps more events, fall thru */
1223 case SWI_STATE_REQUEST_SENT: {
1224 enum srpc_msg_type type;
1226 type = srpc_service2reply(rpc->crpc_service);
1228 if (!rpc->crpc_replyev.ev_fired) break;
1230 rc = rpc->crpc_replyev.ev_status;
1233 srpc_unpack_msg_hdr(reply);
1234 if (reply->msg_type != type ||
1235 (reply->msg_magic != SRPC_MSG_MAGIC &&
1236 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1237 CWARN ("Bad message from %s: type %u (%d expected),"
1238 " magic %u (%d expected).\n",
1239 libcfs_id2str(rpc->crpc_dest),
1240 reply->msg_type, type,
1241 reply->msg_magic, SRPC_MSG_MAGIC);
1246 if (do_bulk && reply->msg_body.reply.status != 0) {
1247 CWARN ("Remote error %d at %s, unlink bulk buffer in "
1248 "case peer didn't initiate bulk transfer\n",
1249 reply->msg_body.reply.status,
1250 libcfs_id2str(rpc->crpc_dest));
1251 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1254 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1256 case SWI_STATE_REPLY_RECEIVED:
1257 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1259 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1261 /* Bulk buffer was unlinked due to remote error. Clear error
1262 * since reply buffer still contains valid data.
1263 * NB rpc->crpc_done shouldn't look into bulk data in case of
1265 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1266 rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1269 wi->swi_state = SWI_STATE_DONE;
1270 srpc_client_rpc_done(rpc, rc);
1275 spin_lock(&rpc->crpc_lock);
1276 srpc_abort_rpc(rpc, rc);
1277 spin_unlock(&rpc->crpc_lock);
1281 if (rpc->crpc_aborted) {
1282 LNetMDUnlink(rpc->crpc_reqstmdh);
1283 LNetMDUnlink(rpc->crpc_replymdh);
1284 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1286 if (!srpc_event_pending(rpc)) {
1287 srpc_client_rpc_done(rpc, -EINTR);
1294 struct srpc_client_rpc *
1295 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1296 int nbulkiov, int bulklen,
1297 void (*rpc_done)(struct srpc_client_rpc *),
1298 void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1300 struct srpc_client_rpc *rpc;
1302 LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1303 crpc_bulk.bk_iovs[nbulkiov]));
1307 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1308 bulklen, rpc_done, rpc_fini, priv);
1312 /* called with rpc->crpc_lock held */
1314 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1318 if (rpc->crpc_aborted || /* already aborted */
1319 rpc->crpc_closed) /* callback imminent */
1323 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1324 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1325 swi_state2str(rpc->crpc_wi.swi_state), why);
1327 rpc->crpc_aborted = 1;
1328 rpc->crpc_status = why;
1329 swi_schedule_workitem(&rpc->crpc_wi);
1333 /* called with rpc->crpc_lock held */
1335 srpc_post_rpc(struct srpc_client_rpc *rpc)
1337 LASSERT (!rpc->crpc_aborted);
1338 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1340 CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1341 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1344 srpc_add_client_rpc_timer(rpc);
1345 swi_schedule_workitem(&rpc->crpc_wi);
1351 srpc_send_reply(struct srpc_server_rpc *rpc)
1353 struct srpc_event *ev = &rpc->srpc_ev;
1354 struct srpc_msg *msg = &rpc->srpc_replymsg;
1355 struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1356 struct srpc_service_cd *scd = rpc->srpc_scd;
1357 struct srpc_service *sv = scd->scd_svc;
1361 LASSERT(buffer != NULL);
1362 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1364 spin_lock(&scd->scd_lock);
1366 if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1367 /* Repost buffer before replying since test client
1368 * might send me another RPC once it gets the reply */
1369 if (srpc_service_post_buffer(scd, buffer) != 0)
1370 CWARN("Failed to repost %s buffer\n", sv->sv_name);
1371 rpc->srpc_reqstbuf = NULL;
1374 spin_unlock(&scd->scd_lock);
1378 ev->ev_type = SRPC_REPLY_SENT;
1380 msg->msg_magic = SRPC_MSG_MAGIC;
1381 msg->msg_version = SRPC_MSG_VERSION;
1382 msg->msg_type = srpc_service2reply(sv->sv_id);
1384 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1385 sizeof(*msg), LNET_MD_OP_PUT,
1386 rpc->srpc_peer, rpc->srpc_self,
1387 &rpc->srpc_replymdh, ev);
1389 ev->ev_fired = 1; /* no more event expected */
1393 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1395 srpc_lnet_ev_handler(struct lnet_event *ev)
1397 struct srpc_service_cd *scd;
1398 struct srpc_event *rpcev = ev->md.user_ptr;
1399 struct srpc_client_rpc *crpc;
1400 struct srpc_server_rpc *srpc;
1401 struct srpc_buffer *buffer;
1402 struct srpc_service *sv;
1403 struct srpc_msg *msg;
1404 enum srpc_msg_type type;
1406 LASSERT (!in_interrupt());
1408 if (ev->status != 0) {
1411 spin_lock(&srpc_data.rpc_glock);
1412 if (ev->status != -ECANCELED) /* cancellation is not error */
1413 srpc_data.rpc_counters.errors++;
1414 errors = srpc_data.rpc_counters.errors;
1415 spin_unlock(&srpc_data.rpc_glock);
1417 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1418 ev->status, ev->type, errors);
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, struct srpc_buffer,
1476 buffer->buf_peer = ev->source;
1477 buffer->buf_self = ev->target.nid;
1479 LASSERT(scd->scd_buf_nposted > 0);
1480 scd->scd_buf_nposted--;
1482 if (sv->sv_shuttingdown) {
1483 /* Leave buffer on scd->scd_buf_nposted since
1484 * srpc_finish_service needs to traverse it. */
1485 spin_unlock(&scd->scd_lock);
1489 if (scd->scd_buf_err_stamp != 0 &&
1490 scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1491 /* re-enable adding buffer */
1492 scd->scd_buf_err_stamp = 0;
1493 scd->scd_buf_err = 0;
1496 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1497 scd->scd_buf_adjust == 0 &&
1498 scd->scd_buf_nposted < scd->scd_buf_low) {
1499 scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1501 swi_schedule_workitem(&scd->scd_buf_wi);
1504 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1505 msg = &buffer->buf_msg;
1506 type = srpc_service2request(sv->sv_id);
1508 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1509 (msg->msg_type != type &&
1510 msg->msg_type != __swab32(type)) ||
1511 (msg->msg_magic != SRPC_MSG_MAGIC &&
1512 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1513 CERROR ("Dropping RPC (%s) from %s: "
1514 "status %d mlength %d type %u magic %u.\n",
1515 sv->sv_name, libcfs_id2str(ev->initiator),
1516 ev->status, ev->mlength,
1517 msg->msg_type, msg->msg_magic);
1519 /* NB can't call srpc_service_recycle_buffer here since
1520 * it may call LNetM[DE]Attach. The invalid magic tells
1521 * srpc_handle_rpc to drop this RPC */
1525 if (!list_empty(&scd->scd_rpc_free)) {
1526 srpc = list_entry(scd->scd_rpc_free.next,
1527 struct srpc_server_rpc,
1529 list_del(&srpc->srpc_list);
1531 srpc_init_server_rpc(srpc, scd, buffer);
1532 list_add_tail(&srpc->srpc_list,
1533 &scd->scd_rpc_active);
1534 swi_schedule_workitem(&srpc->srpc_wi);
1536 list_add_tail(&buffer->buf_list,
1537 &scd->scd_buf_blocked);
1540 spin_unlock(&scd->scd_lock);
1542 spin_lock(&srpc_data.rpc_glock);
1543 srpc_data.rpc_counters.rpcs_rcvd++;
1544 spin_unlock(&srpc_data.rpc_glock);
1547 case SRPC_BULK_GET_RPLD:
1548 LASSERT (ev->type == LNET_EVENT_SEND ||
1549 ev->type == LNET_EVENT_REPLY ||
1550 ev->type == LNET_EVENT_UNLINK);
1553 break; /* wait for final event */
1555 case SRPC_BULK_PUT_SENT:
1556 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1557 spin_lock(&srpc_data.rpc_glock);
1559 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1560 srpc_data.rpc_counters.bulk_get += ev->mlength;
1562 srpc_data.rpc_counters.bulk_put += ev->mlength;
1564 spin_unlock(&srpc_data.rpc_glock);
1566 case SRPC_REPLY_SENT:
1567 srpc = rpcev->ev_data;
1568 scd = srpc->srpc_scd;
1570 LASSERT(rpcev == &srpc->srpc_ev);
1572 spin_lock(&scd->scd_lock);
1574 rpcev->ev_fired = 1;
1575 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1576 -EINTR : ev->status;
1577 swi_schedule_workitem(&srpc->srpc_wi);
1579 spin_unlock(&scd->scd_lock);
1590 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1591 spin_lock_init(&srpc_data.rpc_glock);
1593 /* 1 second pause to avoid timestamp reuse */
1594 set_current_state(TASK_UNINTERRUPTIBLE);
1595 schedule_timeout(cfs_time_seconds(1));
1596 srpc_data.rpc_matchbits = ((__u64) ktime_get_real_seconds()) << 48;
1598 srpc_data.rpc_state = SRPC_STATE_NONE;
1600 rc = LNetNIInit(LNET_PID_LUSTRE);
1602 CERROR ("LNetNIInit() has failed: %d\n", rc);
1606 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1608 LNetInvalidateEQHandle(&srpc_data.rpc_lnet_eq);
1609 rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1611 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1615 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1617 rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1620 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1628 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1634 srpc_shutdown (void)
1640 state = srpc_data.rpc_state;
1641 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1646 case SRPC_STATE_RUNNING:
1647 spin_lock(&srpc_data.rpc_glock);
1649 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1650 struct srpc_service *sv = srpc_data.rpc_services[i];
1652 LASSERTF (sv == NULL,
1653 "service not empty: id %d, name %s\n",
1657 spin_unlock(&srpc_data.rpc_glock);
1661 case SRPC_STATE_EQ_INIT:
1662 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1663 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1665 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1666 LASSERT (rc == 0); /* the EQ should have no user by now */
1668 case SRPC_STATE_NI_INIT: