1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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>
41 #define DEBUG_SUBSYSTEM S_LNET
54 struct smoketest_rpc {
55 cfs_spinlock_t rpc_glock; /* global lock */
56 srpc_service_t *rpc_services[SRPC_SERVICE_MAX_ID + 1];
57 lnet_handle_eq_t rpc_lnet_eq; /* _the_ LNet event queue */
58 srpc_state_t rpc_state;
59 srpc_counters_t rpc_counters;
60 __u64 rpc_matchbits; /* matchbits counter */
64 int srpc_handle_rpc (swi_workitem_t *wi);
66 void srpc_get_counters (srpc_counters_t *cnt)
68 cfs_spin_lock(&srpc_data.rpc_glock);
69 *cnt = srpc_data.rpc_counters;
70 cfs_spin_unlock(&srpc_data.rpc_glock);
73 void srpc_set_counters (const srpc_counters_t *cnt)
75 cfs_spin_lock(&srpc_data.rpc_glock);
76 srpc_data.rpc_counters = *cnt;
77 cfs_spin_unlock(&srpc_data.rpc_glock);
81 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
83 LASSERT (i >= 0 && i < bk->bk_niov);
86 bk->bk_iovs[i].kiov_offset = 0;
87 bk->bk_iovs[i].kiov_page = pg;
88 bk->bk_iovs[i].kiov_len = CFS_PAGE_SIZE;
90 LASSERT (bk->bk_pages != NULL);
93 bk->bk_iovs[i].iov_len = CFS_PAGE_SIZE;
94 bk->bk_iovs[i].iov_base = cfs_page_address(pg);
100 srpc_free_bulk (srpc_bulk_t *bk)
105 LASSERT (bk != NULL);
107 LASSERT (bk->bk_pages != NULL);
110 for (i = 0; i < bk->bk_niov; i++) {
112 pg = bk->bk_iovs[i].kiov_page;
114 pg = bk->bk_pages[i];
116 if (pg == NULL) break;
122 LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
124 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
129 srpc_alloc_bulk (int npages, int sink)
135 LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
137 LIBCFS_ALLOC(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
139 CERROR ("Can't allocate descriptor for %d pages\n", npages);
143 memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
145 bk->bk_niov = npages;
146 bk->bk_len = npages * CFS_PAGE_SIZE;
148 LIBCFS_ALLOC(pages, sizeof(cfs_page_t *) * npages);
150 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
151 CERROR ("Can't allocate page array for %d pages\n", npages);
155 memset(pages, 0, sizeof(cfs_page_t *) * npages);
156 bk->bk_pages = pages;
161 for (i = 0; i < npages; i++) {
162 cfs_page_t *pg = cfs_alloc_page(CFS_ALLOC_STD);
165 CERROR ("Can't allocate page %d of %d\n", i, npages);
170 srpc_add_bulk_page(bk, pg, i);
181 cfs_spin_lock(&srpc_data.rpc_glock);
182 id = srpc_data.rpc_matchbits++;
183 cfs_spin_unlock(&srpc_data.rpc_glock);
188 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
189 srpc_service_t *sv, srpc_buffer_t *buffer)
191 memset(rpc, 0, sizeof(*rpc));
192 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
193 sv->sv_id <= SRPC_FRAMEWORK_SERVICE_MAX_ID ?
194 CFS_WI_SCHED_SERIAL : CFS_WI_SCHED_ANY);
196 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
198 rpc->srpc_service = sv;
199 rpc->srpc_reqstbuf = buffer;
200 rpc->srpc_peer = buffer->buf_peer;
201 rpc->srpc_self = buffer->buf_self;
202 LNetInvalidateHandle(&rpc->srpc_replymdh);
206 srpc_add_service (srpc_service_t *sv)
210 srpc_server_rpc_t *rpc;
212 LASSERT (sv->sv_concur > 0);
213 LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
215 cfs_spin_lock(&srpc_data.rpc_glock);
217 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
219 if (srpc_data.rpc_services[id] != NULL) {
220 cfs_spin_unlock(&srpc_data.rpc_glock);
224 srpc_data.rpc_services[id] = sv;
225 cfs_spin_unlock(&srpc_data.rpc_glock);
228 sv->sv_nposted_msg = 0;
229 sv->sv_shuttingdown = 0;
230 cfs_spin_lock_init(&sv->sv_lock);
231 CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
232 CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
233 CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
234 CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
236 sv->sv_ev.ev_data = sv;
237 sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
239 for (i = 0; i < sv->sv_concur; i++) {
240 LIBCFS_ALLOC(rpc, sizeof(*rpc));
241 if (rpc == NULL) goto enomem;
243 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
246 CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
247 id, sv->sv_name, sv->sv_concur);
251 while (!cfs_list_empty(&sv->sv_free_rpcq)) {
252 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
253 srpc_server_rpc_t, srpc_list);
254 cfs_list_del(&rpc->srpc_list);
255 LIBCFS_FREE(rpc, sizeof(*rpc));
258 cfs_spin_lock(&srpc_data.rpc_glock);
259 srpc_data.rpc_services[id] = NULL;
260 cfs_spin_unlock(&srpc_data.rpc_glock);
265 srpc_remove_service (srpc_service_t *sv)
269 cfs_spin_lock(&srpc_data.rpc_glock);
271 if (srpc_data.rpc_services[id] != sv) {
272 cfs_spin_unlock(&srpc_data.rpc_glock);
276 srpc_data.rpc_services[id] = NULL;
277 cfs_spin_unlock(&srpc_data.rpc_glock);
282 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
283 int len, int options, lnet_process_id_t peer,
284 lnet_handle_md_t *mdh, srpc_event_t *ev)
288 lnet_handle_me_t meh;
290 rc = LNetMEAttach(portal, peer, matchbits, 0,
291 LNET_UNLINK, LNET_INS_AFTER, &meh);
293 CERROR ("LNetMEAttach failed: %d\n", rc);
294 LASSERT (rc == -ENOMEM);
302 md.options = options;
303 md.eq_handle = srpc_data.rpc_lnet_eq;
305 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
307 CERROR ("LNetMDAttach failed: %d\n", rc);
308 LASSERT (rc == -ENOMEM);
310 rc = LNetMEUnlink(meh);
316 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
317 libcfs_id2str(peer), portal, matchbits);
322 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
323 int options, lnet_process_id_t peer, lnet_nid_t self,
324 lnet_handle_md_t *mdh, srpc_event_t *ev)
332 md.eq_handle = srpc_data.rpc_lnet_eq;
333 md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
334 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
336 rc = LNetMDBind(md, LNET_UNLINK, mdh);
338 CERROR ("LNetMDBind failed: %d\n", rc);
339 LASSERT (rc == -ENOMEM);
343 /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
344 * they're only meaningful for MDs attached to an ME (i.e. passive
346 if ((options & LNET_MD_OP_PUT) != 0) {
347 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
348 portal, matchbits, 0, 0);
350 LASSERT ((options & LNET_MD_OP_GET) != 0);
352 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
356 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
357 ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
358 libcfs_id2str(peer), portal, matchbits, rc);
360 /* The forthcoming unlink event will complete this operation
361 * with failure, so fall through and return success here.
363 rc = LNetMDUnlink(*mdh);
367 "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
368 libcfs_id2str(peer), portal, matchbits);
374 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
375 int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
380 if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
381 portal = SRPC_REQUEST_PORTAL;
383 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
385 rc = srpc_post_active_rdma(portal, service, buf, len,
386 LNET_MD_OP_PUT, peer,
387 LNET_NID_ANY, mdh, ev);
392 srpc_post_passive_rqtbuf(int service, void *buf, int len,
393 lnet_handle_md_t *mdh, srpc_event_t *ev)
397 lnet_process_id_t any = {0};
399 any.nid = LNET_NID_ANY;
400 any.pid = LNET_PID_ANY;
402 if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
403 portal = SRPC_REQUEST_PORTAL;
405 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
407 rc = srpc_post_passive_rdma(portal, service, buf, len,
408 LNET_MD_OP_PUT, any, mdh, ev);
413 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
415 srpc_msg_t *msg = &buf->buf_msg;
418 LASSERT (!sv->sv_shuttingdown);
420 LNetInvalidateHandle(&buf->buf_mdh);
421 cfs_list_add(&buf->buf_list, &sv->sv_posted_msgq);
422 sv->sv_nposted_msg++;
423 cfs_spin_unlock(&sv->sv_lock);
425 rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
426 &buf->buf_mdh, &sv->sv_ev);
428 /* At this point, a RPC (new or delayed) may have arrived in
429 * msg and its event handler has been called. So we must add
430 * buf to sv_posted_msgq _before_ dropping sv_lock */
432 cfs_spin_lock(&sv->sv_lock);
435 if (sv->sv_shuttingdown) {
436 cfs_spin_unlock(&sv->sv_lock);
438 /* srpc_shutdown_service might have tried to unlink me
439 * when my buf_mdh was still invalid */
440 LNetMDUnlink(buf->buf_mdh);
442 cfs_spin_lock(&sv->sv_lock);
447 sv->sv_nposted_msg--;
448 if (sv->sv_shuttingdown) return rc;
450 cfs_list_del(&buf->buf_list);
452 cfs_spin_unlock(&sv->sv_lock);
453 LIBCFS_FREE(buf, sizeof(*buf));
454 cfs_spin_lock(&sv->sv_lock);
459 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
465 LASSERTF (nbuffer > 0,
466 "nbuffer must be positive: %d\n", nbuffer);
468 for (posted = 0; posted < nbuffer; posted++) {
469 LIBCFS_ALLOC(buf, sizeof(*buf));
470 if (buf == NULL) break;
472 cfs_spin_lock(&sv->sv_lock);
473 rc = srpc_service_post_buffer(sv, buf);
474 cfs_spin_unlock(&sv->sv_lock);
483 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
485 LASSERTF (nbuffer > 0,
486 "nbuffer must be positive: %d\n", nbuffer);
488 cfs_spin_lock(&sv->sv_lock);
490 LASSERT (sv->sv_nprune >= 0);
491 LASSERT (!sv->sv_shuttingdown);
493 sv->sv_nprune += nbuffer;
495 cfs_spin_unlock(&sv->sv_lock);
499 /* returns 1 if sv has finished, otherwise 0 */
501 srpc_finish_service (srpc_service_t *sv)
503 srpc_server_rpc_t *rpc;
506 cfs_spin_lock(&sv->sv_lock);
508 LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
510 if (sv->sv_nposted_msg != 0 || !cfs_list_empty(&sv->sv_active_rpcq)) {
512 "waiting for %d posted buffers to unlink and "
513 "in-flight RPCs to die.\n",
516 if (!cfs_list_empty(&sv->sv_active_rpcq)) {
517 rpc = cfs_list_entry(sv->sv_active_rpcq.next,
518 srpc_server_rpc_t, srpc_list);
519 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
520 "wi %s scheduled %d running %d, "
521 "ev fired %d type %d status %d lnet %d\n",
522 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
523 swi_state2str(rpc->srpc_wi.swi_state),
524 rpc->srpc_wi.swi_workitem.wi_scheduled,
525 rpc->srpc_wi.swi_workitem.wi_running,
526 rpc->srpc_ev.ev_fired,
527 rpc->srpc_ev.ev_type,
528 rpc->srpc_ev.ev_status,
529 rpc->srpc_ev.ev_lnet);
532 cfs_spin_unlock(&sv->sv_lock);
536 cfs_spin_unlock(&sv->sv_lock); /* no lock needed from now on */
541 if (!cfs_list_empty(&sv->sv_posted_msgq))
542 q = &sv->sv_posted_msgq;
543 else if (!cfs_list_empty(&sv->sv_blocked_msgq))
544 q = &sv->sv_blocked_msgq;
548 buf = cfs_list_entry(q->next, srpc_buffer_t, buf_list);
549 cfs_list_del(&buf->buf_list);
551 LIBCFS_FREE(buf, sizeof(*buf));
554 while (!cfs_list_empty(&sv->sv_free_rpcq)) {
555 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
556 srpc_server_rpc_t, srpc_list);
557 cfs_list_del(&rpc->srpc_list);
558 LIBCFS_FREE(rpc, sizeof(*rpc));
564 /* called with sv->sv_lock held */
566 srpc_service_recycle_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
568 if (sv->sv_shuttingdown) goto free;
570 if (sv->sv_nprune == 0) {
571 if (srpc_service_post_buffer(sv, buf) != 0)
572 CWARN ("Failed to post %s buffer\n", sv->sv_name);
578 cfs_spin_unlock(&sv->sv_lock);
579 LIBCFS_FREE(buf, sizeof(*buf));
580 cfs_spin_lock(&sv->sv_lock);
583 /* called with srpc_service_t::sv_lock held */
585 srpc_schedule_server_rpc (srpc_server_rpc_t *rpc)
587 swi_schedule_workitem(&rpc->srpc_wi);
591 srpc_abort_service (srpc_service_t *sv)
593 srpc_server_rpc_t *rpc;
595 cfs_spin_lock(&sv->sv_lock);
597 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
598 sv->sv_id, sv->sv_name);
600 /* schedule in-flight RPCs to notice the abort, NB:
601 * racing with incoming RPCs; complete fix should make test
602 * RPCs carry session ID in its headers */
603 cfs_list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
604 rpc->srpc_aborted = 1;
605 srpc_schedule_server_rpc(rpc);
608 cfs_spin_unlock(&sv->sv_lock);
613 srpc_shutdown_service (srpc_service_t *sv)
615 srpc_server_rpc_t *rpc;
618 cfs_spin_lock(&sv->sv_lock);
620 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
621 sv->sv_id, sv->sv_name);
623 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
625 /* schedule in-flight RPCs to notice the shutdown */
626 cfs_list_for_each_entry_typed (rpc, &sv->sv_active_rpcq,
627 srpc_server_rpc_t, srpc_list) {
628 srpc_schedule_server_rpc(rpc);
631 cfs_spin_unlock(&sv->sv_lock);
633 /* OK to traverse sv_posted_msgq without lock, since no one
634 * touches sv_posted_msgq now */
635 cfs_list_for_each_entry_typed (buf, &sv->sv_posted_msgq,
636 srpc_buffer_t, buf_list)
637 LNetMDUnlink(buf->buf_mdh);
643 srpc_send_request (srpc_client_rpc_t *rpc)
645 srpc_event_t *ev = &rpc->crpc_reqstev;
650 ev->ev_type = SRPC_REQUEST_SENT;
652 rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
653 &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
654 &rpc->crpc_reqstmdh, ev);
656 LASSERT (rc == -ENOMEM);
657 ev->ev_fired = 1; /* no more event expected */
663 srpc_prepare_reply (srpc_client_rpc_t *rpc)
665 srpc_event_t *ev = &rpc->crpc_replyev;
666 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
671 ev->ev_type = SRPC_REPLY_RCVD;
673 *id = srpc_next_id();
675 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
676 &rpc->crpc_replymsg, sizeof(srpc_msg_t),
677 LNET_MD_OP_PUT, rpc->crpc_dest,
678 &rpc->crpc_replymdh, ev);
680 LASSERT (rc == -ENOMEM);
681 ev->ev_fired = 1; /* no more event expected */
687 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
689 srpc_bulk_t *bk = &rpc->crpc_bulk;
690 srpc_event_t *ev = &rpc->crpc_bulkev;
691 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
695 LASSERT (bk->bk_niov <= LNET_MAX_IOV);
697 if (bk->bk_niov == 0) return 0; /* nothing to do */
699 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
703 opt |= LNET_MD_IOVEC;
708 ev->ev_type = SRPC_BULK_REQ_RCVD;
710 *id = srpc_next_id();
712 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
713 &bk->bk_iovs[0], bk->bk_niov, opt,
714 rpc->crpc_dest, &bk->bk_mdh, ev);
716 LASSERT (rc == -ENOMEM);
717 ev->ev_fired = 1; /* no more event expected */
723 srpc_do_bulk (srpc_server_rpc_t *rpc)
725 srpc_event_t *ev = &rpc->srpc_ev;
726 srpc_bulk_t *bk = rpc->srpc_bulk;
727 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
731 LASSERT (bk != NULL);
733 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
737 opt |= LNET_MD_IOVEC;
742 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
744 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
745 &bk->bk_iovs[0], bk->bk_niov, opt,
746 rpc->srpc_peer, rpc->srpc_self,
749 ev->ev_fired = 1; /* no more event expected */
753 /* only called from srpc_handle_rpc */
755 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
757 srpc_service_t *sv = rpc->srpc_service;
758 srpc_buffer_t *buffer;
760 LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
762 rpc->srpc_status = status;
764 CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
765 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
766 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
767 swi_state2str(rpc->srpc_wi.swi_state), status);
770 cfs_spin_lock(&srpc_data.rpc_glock);
771 srpc_data.rpc_counters.rpcs_dropped++;
772 cfs_spin_unlock(&srpc_data.rpc_glock);
775 if (rpc->srpc_done != NULL)
776 (*rpc->srpc_done) (rpc);
777 LASSERT (rpc->srpc_bulk == NULL);
779 cfs_spin_lock(&sv->sv_lock);
781 if (rpc->srpc_reqstbuf != NULL) {
782 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
783 * sv won't go away for sv_active_rpcq must not be empty */
784 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
785 rpc->srpc_reqstbuf = NULL;
788 cfs_list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
791 * No one can schedule me now since:
792 * - I'm not on sv_active_rpcq.
793 * - all LNet events have been fired.
794 * Cancel pending schedules and prevent future schedule attempts:
796 LASSERT (rpc->srpc_ev.ev_fired);
797 swi_kill_workitem(&rpc->srpc_wi);
799 if (!sv->sv_shuttingdown && !cfs_list_empty(&sv->sv_blocked_msgq)) {
800 buffer = cfs_list_entry(sv->sv_blocked_msgq.next,
801 srpc_buffer_t, buf_list);
802 cfs_list_del(&buffer->buf_list);
804 srpc_init_server_rpc(rpc, sv, buffer);
805 cfs_list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
806 srpc_schedule_server_rpc(rpc);
808 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
811 cfs_spin_unlock(&sv->sv_lock);
815 /* handles an incoming RPC */
817 srpc_handle_rpc (swi_workitem_t *wi)
819 srpc_server_rpc_t *rpc = wi->swi_workitem.wi_data;
820 srpc_service_t *sv = rpc->srpc_service;
821 srpc_event_t *ev = &rpc->srpc_ev;
824 LASSERT (wi == &rpc->srpc_wi);
826 cfs_spin_lock(&sv->sv_lock);
828 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
829 cfs_spin_unlock(&sv->sv_lock);
831 if (rpc->srpc_bulk != NULL)
832 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
833 LNetMDUnlink(rpc->srpc_replymdh);
835 if (ev->ev_fired) { /* no more event, OK to finish */
836 srpc_server_rpc_done(rpc, -ESHUTDOWN);
842 cfs_spin_unlock(&sv->sv_lock);
844 switch (wi->swi_state) {
847 case SWI_STATE_NEWBORN: {
849 srpc_generic_reply_t *reply;
851 msg = &rpc->srpc_reqstbuf->buf_msg;
852 reply = &rpc->srpc_replymsg.msg_body.reply;
854 if (msg->msg_magic == 0) {
855 /* moaned already in srpc_lnet_ev_handler */
857 } else if (msg->msg_version != SRPC_MSG_VERSION &&
858 msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
859 CWARN ("Version mismatch: %u, %u expected, from %s\n",
860 msg->msg_version, SRPC_MSG_VERSION,
861 libcfs_id2str(rpc->srpc_peer));
862 reply->status = EPROTO;
865 rc = (*sv->sv_handler) (rpc);
866 LASSERT (reply->status == 0 || !rpc->srpc_bulk);
870 srpc_server_rpc_done(rpc, rc);
874 wi->swi_state = SWI_STATE_BULK_STARTED;
876 if (rpc->srpc_bulk != NULL) {
877 rc = srpc_do_bulk(rpc);
879 return 0; /* wait for bulk */
881 LASSERT (ev->ev_fired);
885 case SWI_STATE_BULK_STARTED:
886 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
888 if (rpc->srpc_bulk != NULL) {
891 if (sv->sv_bulk_ready != NULL)
892 rc = (*sv->sv_bulk_ready) (rpc, rc);
895 srpc_server_rpc_done(rpc, rc);
900 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
901 rc = srpc_send_reply(rpc);
903 return 0; /* wait for reply */
904 srpc_server_rpc_done(rpc, rc);
907 case SWI_STATE_REPLY_SUBMITTED:
909 CERROR("RPC %p: bulk %p, service %d\n",
910 rpc, rpc->srpc_bulk, rpc->srpc_service->sv_id);
911 CERROR("Event: status %d, type %d, lnet %d\n",
912 ev->ev_status, ev->ev_type, ev->ev_lnet);
913 LASSERT (ev->ev_fired);
916 wi->swi_state = SWI_STATE_DONE;
917 srpc_server_rpc_done(rpc, ev->ev_status);
925 srpc_client_rpc_expired (void *data)
927 srpc_client_rpc_t *rpc = data;
929 CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
930 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
933 cfs_spin_lock(&rpc->crpc_lock);
935 rpc->crpc_timeout = 0;
936 srpc_abort_rpc(rpc, -ETIMEDOUT);
938 cfs_spin_unlock(&rpc->crpc_lock);
940 cfs_spin_lock(&srpc_data.rpc_glock);
941 srpc_data.rpc_counters.rpcs_expired++;
942 cfs_spin_unlock(&srpc_data.rpc_glock);
947 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
949 stt_timer_t *timer = &rpc->crpc_timer;
951 if (rpc->crpc_timeout == 0) return;
953 CFS_INIT_LIST_HEAD(&timer->stt_list);
954 timer->stt_data = rpc;
955 timer->stt_func = srpc_client_rpc_expired;
956 timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
957 cfs_time_current_sec());
958 stt_add_timer(timer);
963 * Called with rpc->crpc_lock held.
965 * Upon exit the RPC expiry timer is not queued and the handler is not
966 * running on any CPU. */
968 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
970 /* timer not planted or already exploded */
971 if (rpc->crpc_timeout == 0) return;
973 /* timer sucessfully defused */
974 if (stt_del_timer(&rpc->crpc_timer)) return;
977 /* timer detonated, wait for it to explode */
978 while (rpc->crpc_timeout != 0) {
979 cfs_spin_unlock(&rpc->crpc_lock);
983 cfs_spin_lock(&rpc->crpc_lock);
986 LBUG(); /* impossible in single-threaded runtime */
992 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
994 swi_workitem_t *wi = &rpc->crpc_wi;
996 LASSERT (status != 0 || wi->swi_state == SWI_STATE_DONE);
998 cfs_spin_lock(&rpc->crpc_lock);
1000 rpc->crpc_closed = 1;
1001 if (rpc->crpc_status == 0)
1002 rpc->crpc_status = status;
1004 srpc_del_client_rpc_timer(rpc);
1006 CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1007 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1008 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1009 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1012 * No one can schedule me now since:
1013 * - RPC timer has been defused.
1014 * - all LNet events have been fired.
1015 * - crpc_closed has been set, preventing srpc_abort_rpc from
1017 * Cancel pending schedules and prevent future schedule attempts:
1019 LASSERT (!srpc_event_pending(rpc));
1020 swi_kill_workitem(wi);
1022 cfs_spin_unlock(&rpc->crpc_lock);
1024 (*rpc->crpc_done) (rpc);
1028 /* sends an outgoing RPC */
1030 srpc_send_rpc (swi_workitem_t *wi)
1033 srpc_client_rpc_t *rpc = wi->swi_workitem.wi_data;
1034 srpc_msg_t *reply = &rpc->crpc_replymsg;
1035 int do_bulk = rpc->crpc_bulk.bk_niov > 0;
1037 LASSERT (rpc != NULL);
1038 LASSERT (wi == &rpc->crpc_wi);
1040 cfs_spin_lock(&rpc->crpc_lock);
1042 if (rpc->crpc_aborted) {
1043 cfs_spin_unlock(&rpc->crpc_lock);
1047 cfs_spin_unlock(&rpc->crpc_lock);
1049 switch (wi->swi_state) {
1052 case SWI_STATE_NEWBORN:
1053 LASSERT (!srpc_event_pending(rpc));
1055 rc = srpc_prepare_reply(rpc);
1057 srpc_client_rpc_done(rpc, rc);
1061 rc = srpc_prepare_bulk(rpc);
1064 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1065 rc = srpc_send_request(rpc);
1068 case SWI_STATE_REQUEST_SUBMITTED:
1069 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1070 * order; however, they're processed in a strict order:
1071 * rqt, rpy, and bulk. */
1072 if (!rpc->crpc_reqstev.ev_fired) break;
1074 rc = rpc->crpc_reqstev.ev_status;
1077 wi->swi_state = SWI_STATE_REQUEST_SENT;
1078 /* perhaps more events, fall thru */
1079 case SWI_STATE_REQUEST_SENT: {
1080 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1082 if (!rpc->crpc_replyev.ev_fired) break;
1084 rc = rpc->crpc_replyev.ev_status;
1087 if ((reply->msg_type != type &&
1088 reply->msg_type != __swab32(type)) ||
1089 (reply->msg_magic != SRPC_MSG_MAGIC &&
1090 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1091 CWARN ("Bad message from %s: type %u (%d expected),"
1092 " magic %u (%d expected).\n",
1093 libcfs_id2str(rpc->crpc_dest),
1094 reply->msg_type, type,
1095 reply->msg_magic, SRPC_MSG_MAGIC);
1100 if (do_bulk && reply->msg_body.reply.status != 0) {
1101 CWARN ("Remote error %d at %s, unlink bulk buffer in "
1102 "case peer didn't initiate bulk transfer\n",
1103 reply->msg_body.reply.status,
1104 libcfs_id2str(rpc->crpc_dest));
1105 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1108 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1110 case SWI_STATE_REPLY_RECEIVED:
1111 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1113 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1115 /* Bulk buffer was unlinked due to remote error. Clear error
1116 * since reply buffer still contains valid data.
1117 * NB rpc->crpc_done shouldn't look into bulk data in case of
1119 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1120 rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1123 wi->swi_state = SWI_STATE_DONE;
1124 srpc_client_rpc_done(rpc, rc);
1129 cfs_spin_lock(&rpc->crpc_lock);
1130 srpc_abort_rpc(rpc, rc);
1131 cfs_spin_unlock(&rpc->crpc_lock);
1135 if (rpc->crpc_aborted) {
1136 LNetMDUnlink(rpc->crpc_reqstmdh);
1137 LNetMDUnlink(rpc->crpc_replymdh);
1138 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1140 if (!srpc_event_pending(rpc)) {
1141 srpc_client_rpc_done(rpc, -EINTR);
1149 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1150 int nbulkiov, int bulklen,
1151 void (*rpc_done)(srpc_client_rpc_t *),
1152 void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1154 srpc_client_rpc_t *rpc;
1156 LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1157 crpc_bulk.bk_iovs[nbulkiov]));
1161 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1162 bulklen, rpc_done, rpc_fini, priv);
1166 /* called with rpc->crpc_lock held */
1168 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1172 if (rpc->crpc_aborted || /* already aborted */
1173 rpc->crpc_closed) /* callback imminent */
1177 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1178 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1179 swi_state2str(rpc->crpc_wi.swi_state), why);
1181 rpc->crpc_aborted = 1;
1182 rpc->crpc_status = why;
1183 swi_schedule_workitem(&rpc->crpc_wi);
1187 /* called with rpc->crpc_lock held */
1189 srpc_post_rpc (srpc_client_rpc_t *rpc)
1191 LASSERT (!rpc->crpc_aborted);
1192 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1193 LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1195 CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1196 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1199 srpc_add_client_rpc_timer(rpc);
1200 swi_schedule_workitem(&rpc->crpc_wi);
1206 srpc_send_reply (srpc_server_rpc_t *rpc)
1208 srpc_event_t *ev = &rpc->srpc_ev;
1209 srpc_msg_t *msg = &rpc->srpc_replymsg;
1210 srpc_buffer_t *buffer = rpc->srpc_reqstbuf;
1211 srpc_service_t *sv = rpc->srpc_service;
1215 LASSERT (buffer != NULL);
1216 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1218 cfs_spin_lock(&sv->sv_lock);
1220 if (!sv->sv_shuttingdown &&
1221 sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1222 /* Repost buffer before replying since test client
1223 * might send me another RPC once it gets the reply */
1224 if (srpc_service_post_buffer(sv, buffer) != 0)
1225 CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1226 rpc->srpc_reqstbuf = NULL;
1229 cfs_spin_unlock(&sv->sv_lock);
1233 ev->ev_type = SRPC_REPLY_SENT;
1235 msg->msg_magic = SRPC_MSG_MAGIC;
1236 msg->msg_version = SRPC_MSG_VERSION;
1237 msg->msg_type = srpc_service2reply(sv->sv_id);
1239 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1240 sizeof(*msg), LNET_MD_OP_PUT,
1241 rpc->srpc_peer, rpc->srpc_self,
1242 &rpc->srpc_replymdh, ev);
1244 ev->ev_fired = 1; /* no more event expected */
1248 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1250 srpc_lnet_ev_handler (lnet_event_t *ev)
1252 srpc_event_t *rpcev = ev->md.user_ptr;
1253 srpc_client_rpc_t *crpc;
1254 srpc_server_rpc_t *srpc;
1255 srpc_buffer_t *buffer;
1258 srpc_msg_type_t type;
1260 LASSERT (!cfs_in_interrupt());
1262 if (ev->status != 0) {
1263 cfs_spin_lock(&srpc_data.rpc_glock);
1264 srpc_data.rpc_counters.errors++;
1265 cfs_spin_unlock(&srpc_data.rpc_glock);
1268 rpcev->ev_lnet = ev->type;
1270 switch (rpcev->ev_type) {
1272 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1273 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1275 case SRPC_REQUEST_SENT:
1276 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1277 cfs_spin_lock(&srpc_data.rpc_glock);
1278 srpc_data.rpc_counters.rpcs_sent++;
1279 cfs_spin_unlock(&srpc_data.rpc_glock);
1281 case SRPC_REPLY_RCVD:
1282 case SRPC_BULK_REQ_RCVD:
1283 crpc = rpcev->ev_data;
1285 if (rpcev != &crpc->crpc_reqstev &&
1286 rpcev != &crpc->crpc_replyev &&
1287 rpcev != &crpc->crpc_bulkev) {
1288 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1289 rpcev, crpc, &crpc->crpc_reqstev,
1290 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1291 CERROR("Bad event: status %d, type %d, lnet %d\n",
1292 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1296 cfs_spin_lock(&crpc->crpc_lock);
1298 LASSERT (rpcev->ev_fired == 0);
1299 rpcev->ev_fired = 1;
1300 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1301 -EINTR : ev->status;
1302 swi_schedule_workitem(&crpc->crpc_wi);
1304 cfs_spin_unlock(&crpc->crpc_lock);
1307 case SRPC_REQUEST_RCVD:
1308 sv = rpcev->ev_data;
1310 LASSERT (rpcev == &sv->sv_ev);
1312 cfs_spin_lock(&sv->sv_lock);
1314 LASSERT (ev->unlinked);
1315 LASSERT (ev->type == LNET_EVENT_PUT ||
1316 ev->type == LNET_EVENT_UNLINK);
1317 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1318 sv->sv_shuttingdown);
1320 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1321 buffer->buf_peer = ev->initiator;
1322 buffer->buf_self = ev->target.nid;
1324 sv->sv_nposted_msg--;
1325 LASSERT (sv->sv_nposted_msg >= 0);
1327 if (sv->sv_shuttingdown) {
1328 /* Leave buffer on sv->sv_posted_msgq since
1329 * srpc_finish_service needs to traverse it. */
1330 cfs_spin_unlock(&sv->sv_lock);
1334 cfs_list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1335 msg = &buffer->buf_msg;
1336 type = srpc_service2request(sv->sv_id);
1338 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1339 (msg->msg_type != type &&
1340 msg->msg_type != __swab32(type)) ||
1341 (msg->msg_magic != SRPC_MSG_MAGIC &&
1342 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1343 CERROR ("Dropping RPC (%s) from %s: "
1344 "status %d mlength %d type %u magic %u.\n",
1345 sv->sv_name, libcfs_id2str(ev->initiator),
1346 ev->status, ev->mlength,
1347 msg->msg_type, msg->msg_magic);
1349 /* NB can't call srpc_service_recycle_buffer here since
1350 * it may call LNetM[DE]Attach. The invalid magic tells
1351 * srpc_handle_rpc to drop this RPC */
1355 if (!cfs_list_empty(&sv->sv_free_rpcq)) {
1356 srpc = cfs_list_entry(sv->sv_free_rpcq.next,
1357 srpc_server_rpc_t, srpc_list);
1358 cfs_list_del(&srpc->srpc_list);
1360 srpc_init_server_rpc(srpc, sv, buffer);
1361 cfs_list_add_tail(&srpc->srpc_list,
1362 &sv->sv_active_rpcq);
1363 srpc_schedule_server_rpc(srpc);
1365 cfs_list_add_tail(&buffer->buf_list,
1366 &sv->sv_blocked_msgq);
1369 cfs_spin_unlock(&sv->sv_lock);
1371 cfs_spin_lock(&srpc_data.rpc_glock);
1372 srpc_data.rpc_counters.rpcs_rcvd++;
1373 cfs_spin_unlock(&srpc_data.rpc_glock);
1376 case SRPC_BULK_GET_RPLD:
1377 LASSERT (ev->type == LNET_EVENT_SEND ||
1378 ev->type == LNET_EVENT_REPLY ||
1379 ev->type == LNET_EVENT_UNLINK);
1382 break; /* wait for final event */
1384 case SRPC_BULK_PUT_SENT:
1385 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1386 cfs_spin_lock(&srpc_data.rpc_glock);
1388 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1389 srpc_data.rpc_counters.bulk_get += ev->mlength;
1391 srpc_data.rpc_counters.bulk_put += ev->mlength;
1393 cfs_spin_unlock(&srpc_data.rpc_glock);
1395 case SRPC_REPLY_SENT:
1396 srpc = rpcev->ev_data;
1397 sv = srpc->srpc_service;
1399 LASSERT (rpcev == &srpc->srpc_ev);
1401 cfs_spin_lock(&sv->sv_lock);
1403 rpcev->ev_fired = 1;
1404 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1405 -EINTR : ev->status;
1406 srpc_schedule_server_rpc(srpc);
1408 cfs_spin_unlock(&sv->sv_lock);
1418 srpc_check_event (int timeout)
1424 rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1425 timeout * 1000, &ev, &i);
1426 if (rc == 0) return 0;
1428 LASSERT (rc == -EOVERFLOW || rc == 1);
1430 /* We can't affort to miss any events... */
1431 if (rc == -EOVERFLOW) {
1432 CERROR ("Dropped an event!!!\n");
1436 srpc_lnet_ev_handler(&ev);
1447 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1448 cfs_spin_lock_init(&srpc_data.rpc_glock);
1450 /* 1 second pause to avoid timestamp reuse */
1451 cfs_pause(cfs_time_seconds(1));
1452 srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1454 srpc_data.rpc_state = SRPC_STATE_NONE;
1457 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1459 if (the_lnet.ln_server_mode_flag)
1460 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1462 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1465 CERROR ("LNetNIInit() has failed: %d\n", rc);
1469 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1471 LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1473 rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1475 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1478 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1482 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1485 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1493 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1499 srpc_shutdown (void)
1505 state = srpc_data.rpc_state;
1506 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1511 case SRPC_STATE_RUNNING:
1512 cfs_spin_lock(&srpc_data.rpc_glock);
1514 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1515 srpc_service_t *sv = srpc_data.rpc_services[i];
1517 LASSERTF (sv == NULL,
1518 "service not empty: id %d, name %s\n",
1522 cfs_spin_unlock(&srpc_data.rpc_glock);
1526 case SRPC_STATE_EQ_INIT:
1527 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1529 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1530 LASSERT (rc == 0); /* the EQ should have no user by now */
1532 case SRPC_STATE_NI_INIT: