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 2008 Sun Microsystems, Inc. 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
55 struct smoketest_rpc {
56 cfs_spinlock_t rpc_glock; /* global lock */
57 srpc_service_t *rpc_services[SRPC_SERVICE_MAX_ID + 1];
58 lnet_handle_eq_t rpc_lnet_eq; /* _the_ LNet event queue */
59 srpc_state_t rpc_state;
60 srpc_counters_t rpc_counters;
61 __u64 rpc_matchbits; /* matchbits counter */
65 int srpc_handle_rpc (swi_workitem_t *wi);
67 void srpc_get_counters (srpc_counters_t *cnt)
69 cfs_spin_lock(&srpc_data.rpc_glock);
70 *cnt = srpc_data.rpc_counters;
71 cfs_spin_unlock(&srpc_data.rpc_glock);
74 void srpc_set_counters (const srpc_counters_t *cnt)
76 cfs_spin_lock(&srpc_data.rpc_glock);
77 srpc_data.rpc_counters = *cnt;
78 cfs_spin_unlock(&srpc_data.rpc_glock);
82 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
84 LASSERT (i >= 0 && i < bk->bk_niov);
87 bk->bk_iovs[i].kiov_offset = 0;
88 bk->bk_iovs[i].kiov_page = pg;
89 bk->bk_iovs[i].kiov_len = CFS_PAGE_SIZE;
91 LASSERT (bk->bk_pages != NULL);
94 bk->bk_iovs[i].iov_len = CFS_PAGE_SIZE;
95 bk->bk_iovs[i].iov_base = cfs_page_address(pg);
101 srpc_free_bulk (srpc_bulk_t *bk)
106 LASSERT (bk != NULL);
108 LASSERT (bk->bk_pages != NULL);
111 for (i = 0; i < bk->bk_niov; i++) {
113 pg = bk->bk_iovs[i].kiov_page;
115 pg = bk->bk_pages[i];
117 if (pg == NULL) break;
123 LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
125 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
130 srpc_alloc_bulk (int npages, int sink)
136 LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
138 LIBCFS_ALLOC(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
140 CERROR ("Can't allocate descriptor for %d pages\n", npages);
144 memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
146 bk->bk_niov = npages;
147 bk->bk_len = npages * CFS_PAGE_SIZE;
149 LIBCFS_ALLOC(pages, sizeof(cfs_page_t *) * npages);
151 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
152 CERROR ("Can't allocate page array for %d pages\n", npages);
156 memset(pages, 0, sizeof(cfs_page_t *) * npages);
157 bk->bk_pages = pages;
162 for (i = 0; i < npages; i++) {
163 cfs_page_t *pg = cfs_alloc_page(CFS_ALLOC_STD);
166 CERROR ("Can't allocate page %d of %d\n", i, npages);
171 srpc_add_bulk_page(bk, pg, i);
182 cfs_spin_lock(&srpc_data.rpc_glock);
183 id = srpc_data.rpc_matchbits++;
184 cfs_spin_unlock(&srpc_data.rpc_glock);
189 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
190 srpc_service_t *sv, srpc_buffer_t *buffer)
192 memset(rpc, 0, sizeof(*rpc));
193 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc);
195 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
197 rpc->srpc_service = sv;
198 rpc->srpc_reqstbuf = buffer;
199 rpc->srpc_peer = buffer->buf_peer;
200 rpc->srpc_self = buffer->buf_self;
201 LNetInvalidateHandle(&rpc->srpc_replymdh);
205 srpc_add_service (srpc_service_t *sv)
209 srpc_server_rpc_t *rpc;
211 LASSERT (sv->sv_concur > 0);
212 LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
214 cfs_spin_lock(&srpc_data.rpc_glock);
216 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
218 if (srpc_data.rpc_services[id] != NULL) {
219 cfs_spin_unlock(&srpc_data.rpc_glock);
223 srpc_data.rpc_services[id] = sv;
224 cfs_spin_unlock(&srpc_data.rpc_glock);
227 sv->sv_nposted_msg = 0;
228 sv->sv_shuttingdown = 0;
229 cfs_spin_lock_init(&sv->sv_lock);
230 CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
231 CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
232 CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
233 CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
235 sv->sv_ev.ev_data = sv;
236 sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
238 for (i = 0; i < sv->sv_concur; i++) {
239 LIBCFS_ALLOC(rpc, sizeof(*rpc));
240 if (rpc == NULL) goto enomem;
242 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
245 CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
246 id, sv->sv_name, sv->sv_concur);
250 while (!cfs_list_empty(&sv->sv_free_rpcq)) {
251 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
252 srpc_server_rpc_t, srpc_list);
253 cfs_list_del(&rpc->srpc_list);
254 LIBCFS_FREE(rpc, sizeof(*rpc));
257 cfs_spin_lock(&srpc_data.rpc_glock);
258 srpc_data.rpc_services[id] = NULL;
259 cfs_spin_unlock(&srpc_data.rpc_glock);
264 srpc_remove_service (srpc_service_t *sv)
268 cfs_spin_lock(&srpc_data.rpc_glock);
270 if (srpc_data.rpc_services[id] != sv) {
271 cfs_spin_unlock(&srpc_data.rpc_glock);
275 srpc_data.rpc_services[id] = NULL;
276 cfs_spin_unlock(&srpc_data.rpc_glock);
281 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
282 int len, int options, lnet_process_id_t peer,
283 lnet_handle_md_t *mdh, srpc_event_t *ev)
287 lnet_handle_me_t meh;
289 rc = LNetMEAttach(portal, peer, matchbits, 0,
290 LNET_UNLINK, LNET_INS_AFTER, &meh);
292 CERROR ("LNetMEAttach failed: %d\n", rc);
293 LASSERT (rc == -ENOMEM);
301 md.options = options;
302 md.eq_handle = srpc_data.rpc_lnet_eq;
304 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
306 CERROR ("LNetMDAttach failed: %d\n", rc);
307 LASSERT (rc == -ENOMEM);
309 rc = LNetMEUnlink(meh);
315 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
316 libcfs_id2str(peer), portal, matchbits);
321 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
322 int options, lnet_process_id_t peer, lnet_nid_t self,
323 lnet_handle_md_t *mdh, srpc_event_t *ev)
331 md.eq_handle = srpc_data.rpc_lnet_eq;
332 md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
333 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
335 rc = LNetMDBind(md, LNET_UNLINK, mdh);
337 CERROR ("LNetMDBind failed: %d\n", rc);
338 LASSERT (rc == -ENOMEM);
342 /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
343 * they're only meaningful for MDs attached to an ME (i.e. passive
345 if ((options & LNET_MD_OP_PUT) != 0) {
346 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
347 portal, matchbits, 0, 0);
349 LASSERT ((options & LNET_MD_OP_GET) != 0);
351 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
355 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
356 ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
357 libcfs_id2str(peer), portal, matchbits, rc);
359 /* The forthcoming unlink event will complete this operation
360 * with failure, so fall through and return success here.
362 rc = LNetMDUnlink(*mdh);
366 "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
367 libcfs_id2str(peer), portal, matchbits);
373 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
374 int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
379 if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
380 portal = SRPC_REQUEST_PORTAL;
382 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
384 rc = srpc_post_active_rdma(portal, service, buf, len,
385 LNET_MD_OP_PUT, peer,
386 LNET_NID_ANY, mdh, ev);
391 srpc_post_passive_rqtbuf(int service, void *buf, int len,
392 lnet_handle_md_t *mdh, srpc_event_t *ev)
396 lnet_process_id_t any = {0};
398 any.nid = LNET_NID_ANY;
399 any.pid = LNET_PID_ANY;
401 if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
402 portal = SRPC_REQUEST_PORTAL;
404 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
406 rc = srpc_post_passive_rdma(portal, service, buf, len,
407 LNET_MD_OP_PUT, any, mdh, ev);
412 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
414 srpc_msg_t *msg = &buf->buf_msg;
417 LASSERT (!sv->sv_shuttingdown);
419 LNetInvalidateHandle(&buf->buf_mdh);
420 cfs_list_add(&buf->buf_list, &sv->sv_posted_msgq);
421 sv->sv_nposted_msg++;
422 cfs_spin_unlock(&sv->sv_lock);
424 rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
425 &buf->buf_mdh, &sv->sv_ev);
427 /* At this point, a RPC (new or delayed) may have arrived in
428 * msg and its event handler has been called. So we must add
429 * buf to sv_posted_msgq _before_ dropping sv_lock */
431 cfs_spin_lock(&sv->sv_lock);
434 if (sv->sv_shuttingdown) {
435 cfs_spin_unlock(&sv->sv_lock);
437 /* srpc_shutdown_service might have tried to unlink me
438 * when my buf_mdh was still invalid */
439 LNetMDUnlink(buf->buf_mdh);
441 cfs_spin_lock(&sv->sv_lock);
446 sv->sv_nposted_msg--;
447 if (sv->sv_shuttingdown) return rc;
449 cfs_list_del(&buf->buf_list);
451 cfs_spin_unlock(&sv->sv_lock);
452 LIBCFS_FREE(buf, sizeof(*buf));
453 cfs_spin_lock(&sv->sv_lock);
458 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
464 LASSERTF (nbuffer > 0,
465 "nbuffer must be positive: %d\n", nbuffer);
467 for (posted = 0; posted < nbuffer; posted++) {
468 LIBCFS_ALLOC(buf, sizeof(*buf));
469 if (buf == NULL) break;
471 cfs_spin_lock(&sv->sv_lock);
472 rc = srpc_service_post_buffer(sv, buf);
473 cfs_spin_unlock(&sv->sv_lock);
482 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
484 LASSERTF (nbuffer > 0,
485 "nbuffer must be positive: %d\n", nbuffer);
487 cfs_spin_lock(&sv->sv_lock);
489 LASSERT (sv->sv_nprune >= 0);
490 LASSERT (!sv->sv_shuttingdown);
492 sv->sv_nprune += nbuffer;
494 cfs_spin_unlock(&sv->sv_lock);
498 /* returns 1 if sv has finished, otherwise 0 */
500 srpc_finish_service (srpc_service_t *sv)
502 srpc_server_rpc_t *rpc;
505 cfs_spin_lock(&sv->sv_lock);
507 LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
509 if (sv->sv_nposted_msg != 0 || !cfs_list_empty(&sv->sv_active_rpcq)) {
511 "waiting for %d posted buffers to unlink and "
512 "in-flight RPCs to die.\n",
515 if (!cfs_list_empty(&sv->sv_active_rpcq)) {
516 rpc = cfs_list_entry(sv->sv_active_rpcq.next,
517 srpc_server_rpc_t, srpc_list);
519 "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.wi_state),
524 rpc->srpc_wi.wi_scheduled,
525 rpc->srpc_wi.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 srpc_service_t *sv = rpc->srpc_service;
589 if (sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID)
590 swi_schedule_workitem(&rpc->srpc_wi);
591 else /* framework RPCs are handled one by one */
592 swi_schedule_serial_workitem(&rpc->srpc_wi);
598 srpc_abort_service (srpc_service_t *sv)
600 srpc_server_rpc_t *rpc;
602 cfs_spin_lock(&sv->sv_lock);
604 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
605 sv->sv_id, sv->sv_name);
607 /* schedule in-flight RPCs to notice the abort, NB:
608 * racing with incoming RPCs; complete fix should make test
609 * RPCs carry session ID in its headers */
610 cfs_list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
611 rpc->srpc_aborted = 1;
612 srpc_schedule_server_rpc(rpc);
615 cfs_spin_unlock(&sv->sv_lock);
620 srpc_shutdown_service (srpc_service_t *sv)
622 srpc_server_rpc_t *rpc;
625 cfs_spin_lock(&sv->sv_lock);
627 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
628 sv->sv_id, sv->sv_name);
630 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
632 /* schedule in-flight RPCs to notice the shutdown */
633 cfs_list_for_each_entry_typed (rpc, &sv->sv_active_rpcq,
634 srpc_server_rpc_t, srpc_list) {
635 srpc_schedule_server_rpc(rpc);
638 cfs_spin_unlock(&sv->sv_lock);
640 /* OK to traverse sv_posted_msgq without lock, since no one
641 * touches sv_posted_msgq now */
642 cfs_list_for_each_entry_typed (buf, &sv->sv_posted_msgq,
643 srpc_buffer_t, buf_list)
644 LNetMDUnlink(buf->buf_mdh);
650 srpc_send_request (srpc_client_rpc_t *rpc)
652 srpc_event_t *ev = &rpc->crpc_reqstev;
657 ev->ev_type = SRPC_REQUEST_SENT;
659 rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
660 &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
661 &rpc->crpc_reqstmdh, ev);
663 LASSERT (rc == -ENOMEM);
664 ev->ev_fired = 1; /* no more event expected */
670 srpc_prepare_reply (srpc_client_rpc_t *rpc)
672 srpc_event_t *ev = &rpc->crpc_replyev;
673 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
678 ev->ev_type = SRPC_REPLY_RCVD;
680 *id = srpc_next_id();
682 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
683 &rpc->crpc_replymsg, sizeof(srpc_msg_t),
684 LNET_MD_OP_PUT, rpc->crpc_dest,
685 &rpc->crpc_replymdh, ev);
687 LASSERT (rc == -ENOMEM);
688 ev->ev_fired = 1; /* no more event expected */
694 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
696 srpc_bulk_t *bk = &rpc->crpc_bulk;
697 srpc_event_t *ev = &rpc->crpc_bulkev;
698 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
702 LASSERT (bk->bk_niov <= LNET_MAX_IOV);
704 if (bk->bk_niov == 0) return 0; /* nothing to do */
706 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
710 opt |= LNET_MD_IOVEC;
715 ev->ev_type = SRPC_BULK_REQ_RCVD;
717 *id = srpc_next_id();
719 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
720 &bk->bk_iovs[0], bk->bk_niov, opt,
721 rpc->crpc_dest, &bk->bk_mdh, ev);
723 LASSERT (rc == -ENOMEM);
724 ev->ev_fired = 1; /* no more event expected */
730 srpc_do_bulk (srpc_server_rpc_t *rpc)
732 srpc_event_t *ev = &rpc->srpc_ev;
733 srpc_bulk_t *bk = rpc->srpc_bulk;
734 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
738 LASSERT (bk != NULL);
740 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
744 opt |= LNET_MD_IOVEC;
749 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
751 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
752 &bk->bk_iovs[0], bk->bk_niov, opt,
753 rpc->srpc_peer, rpc->srpc_self,
756 ev->ev_fired = 1; /* no more event expected */
760 /* only called from srpc_handle_rpc */
762 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
764 srpc_service_t *sv = rpc->srpc_service;
765 srpc_buffer_t *buffer;
767 LASSERT (status != 0 || rpc->srpc_wi.wi_state == SWI_STATE_DONE);
769 rpc->srpc_status = status;
771 CDEBUG (status == 0 ? D_NET : D_NETERROR,
772 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
773 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
774 swi_state2str(rpc->srpc_wi.wi_state), status);
777 cfs_spin_lock(&srpc_data.rpc_glock);
778 srpc_data.rpc_counters.rpcs_dropped++;
779 cfs_spin_unlock(&srpc_data.rpc_glock);
782 if (rpc->srpc_done != NULL)
783 (*rpc->srpc_done) (rpc);
784 LASSERT (rpc->srpc_bulk == NULL);
786 cfs_spin_lock(&sv->sv_lock);
788 if (rpc->srpc_reqstbuf != NULL) {
789 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
790 * sv won't go away for sv_active_rpcq must not be empty */
791 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
792 rpc->srpc_reqstbuf = NULL;
795 cfs_list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
798 * No one can schedule me now since:
799 * - I'm not on sv_active_rpcq.
800 * - all LNet events have been fired.
801 * Cancel pending schedules and prevent future schedule attempts:
803 LASSERT (rpc->srpc_ev.ev_fired);
804 swi_kill_workitem(&rpc->srpc_wi);
806 if (!sv->sv_shuttingdown && !cfs_list_empty(&sv->sv_blocked_msgq)) {
807 buffer = cfs_list_entry(sv->sv_blocked_msgq.next,
808 srpc_buffer_t, buf_list);
809 cfs_list_del(&buffer->buf_list);
811 srpc_init_server_rpc(rpc, sv, buffer);
812 cfs_list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
813 srpc_schedule_server_rpc(rpc);
815 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
818 cfs_spin_unlock(&sv->sv_lock);
822 /* handles an incoming RPC */
824 srpc_handle_rpc (swi_workitem_t *wi)
826 srpc_server_rpc_t *rpc = wi->wi_data;
827 srpc_service_t *sv = rpc->srpc_service;
828 srpc_event_t *ev = &rpc->srpc_ev;
831 LASSERT (wi == &rpc->srpc_wi);
833 cfs_spin_lock(&sv->sv_lock);
835 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
836 cfs_spin_unlock(&sv->sv_lock);
838 if (rpc->srpc_bulk != NULL)
839 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
840 LNetMDUnlink(rpc->srpc_replymdh);
842 if (ev->ev_fired) { /* no more event, OK to finish */
843 srpc_server_rpc_done(rpc, -ESHUTDOWN);
849 cfs_spin_unlock(&sv->sv_lock);
851 switch (wi->wi_state) {
854 case SWI_STATE_NEWBORN: {
856 srpc_generic_reply_t *reply;
858 msg = &rpc->srpc_reqstbuf->buf_msg;
859 reply = &rpc->srpc_replymsg.msg_body.reply;
861 if (msg->msg_magic == 0) {
862 /* moaned already in srpc_lnet_ev_handler */
864 } else if (msg->msg_version != SRPC_MSG_VERSION &&
865 msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
866 CWARN ("Version mismatch: %u, %u expected, from %s\n",
867 msg->msg_version, SRPC_MSG_VERSION,
868 libcfs_id2str(rpc->srpc_peer));
869 reply->status = EPROTO;
872 rc = (*sv->sv_handler) (rpc);
873 LASSERT (reply->status == 0 || !rpc->srpc_bulk);
877 srpc_server_rpc_done(rpc, rc);
881 wi->wi_state = SWI_STATE_BULK_STARTED;
883 if (rpc->srpc_bulk != NULL) {
884 rc = srpc_do_bulk(rpc);
886 return 0; /* wait for bulk */
888 LASSERT (ev->ev_fired);
892 case SWI_STATE_BULK_STARTED:
893 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
895 if (rpc->srpc_bulk != NULL) {
898 if (sv->sv_bulk_ready != NULL)
899 rc = (*sv->sv_bulk_ready) (rpc, rc);
902 srpc_server_rpc_done(rpc, rc);
907 wi->wi_state = SWI_STATE_REPLY_SUBMITTED;
908 rc = srpc_send_reply(rpc);
910 return 0; /* wait for reply */
911 srpc_server_rpc_done(rpc, rc);
914 case SWI_STATE_REPLY_SUBMITTED:
916 CERROR("RPC %p: bulk %p, service %d\n",
917 rpc, rpc->srpc_bulk, rpc->srpc_service->sv_id);
918 CERROR("Event: status %d, type %d, lnet %d\n",
919 ev->ev_status, ev->ev_type, ev->ev_lnet);
920 LASSERT (ev->ev_fired);
923 wi->wi_state = SWI_STATE_DONE;
924 srpc_server_rpc_done(rpc, ev->ev_status);
932 srpc_client_rpc_expired (void *data)
934 srpc_client_rpc_t *rpc = data;
936 CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
937 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
940 cfs_spin_lock(&rpc->crpc_lock);
942 rpc->crpc_timeout = 0;
943 srpc_abort_rpc(rpc, -ETIMEDOUT);
945 cfs_spin_unlock(&rpc->crpc_lock);
947 cfs_spin_lock(&srpc_data.rpc_glock);
948 srpc_data.rpc_counters.rpcs_expired++;
949 cfs_spin_unlock(&srpc_data.rpc_glock);
954 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
956 stt_timer_t *timer = &rpc->crpc_timer;
958 if (rpc->crpc_timeout == 0) return;
960 CFS_INIT_LIST_HEAD(&timer->stt_list);
961 timer->stt_data = rpc;
962 timer->stt_func = srpc_client_rpc_expired;
963 timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
964 cfs_time_current_sec());
965 stt_add_timer(timer);
970 * Called with rpc->crpc_lock held.
972 * Upon exit the RPC expiry timer is not queued and the handler is not
973 * running on any CPU. */
975 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
977 /* timer not planted or already exploded */
978 if (rpc->crpc_timeout == 0) return;
980 /* timer sucessfully defused */
981 if (stt_del_timer(&rpc->crpc_timer)) return;
984 /* timer detonated, wait for it to explode */
985 while (rpc->crpc_timeout != 0) {
986 cfs_spin_unlock(&rpc->crpc_lock);
990 cfs_spin_lock(&rpc->crpc_lock);
993 LBUG(); /* impossible in single-threaded runtime */
999 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1001 swi_workitem_t *wi = &rpc->crpc_wi;
1003 LASSERT (status != 0 || wi->wi_state == SWI_STATE_DONE);
1005 cfs_spin_lock(&rpc->crpc_lock);
1007 rpc->crpc_closed = 1;
1008 if (rpc->crpc_status == 0)
1009 rpc->crpc_status = status;
1011 srpc_del_client_rpc_timer(rpc);
1013 CDEBUG ((status == 0) ? D_NET : D_NETERROR,
1014 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1015 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1016 swi_state2str(wi->wi_state), rpc->crpc_aborted, status);
1019 * No one can schedule me now since:
1020 * - RPC timer has been defused.
1021 * - all LNet events have been fired.
1022 * - crpc_closed has been set, preventing srpc_abort_rpc from
1024 * Cancel pending schedules and prevent future schedule attempts:
1026 LASSERT (!srpc_event_pending(rpc));
1027 swi_kill_workitem(wi);
1029 cfs_spin_unlock(&rpc->crpc_lock);
1031 (*rpc->crpc_done) (rpc);
1035 /* sends an outgoing RPC */
1037 srpc_send_rpc (swi_workitem_t *wi)
1040 srpc_client_rpc_t *rpc = wi->wi_data;
1041 srpc_msg_t *reply = &rpc->crpc_replymsg;
1042 int do_bulk = rpc->crpc_bulk.bk_niov > 0;
1044 LASSERT (rpc != NULL);
1045 LASSERT (wi == &rpc->crpc_wi);
1047 cfs_spin_lock(&rpc->crpc_lock);
1049 if (rpc->crpc_aborted) {
1050 cfs_spin_unlock(&rpc->crpc_lock);
1054 cfs_spin_unlock(&rpc->crpc_lock);
1056 switch (wi->wi_state) {
1059 case SWI_STATE_NEWBORN:
1060 LASSERT (!srpc_event_pending(rpc));
1062 rc = srpc_prepare_reply(rpc);
1064 srpc_client_rpc_done(rpc, rc);
1068 rc = srpc_prepare_bulk(rpc);
1071 wi->wi_state = SWI_STATE_REQUEST_SUBMITTED;
1072 rc = srpc_send_request(rpc);
1075 case SWI_STATE_REQUEST_SUBMITTED:
1076 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1077 * order; however, they're processed in a strict order:
1078 * rqt, rpy, and bulk. */
1079 if (!rpc->crpc_reqstev.ev_fired) break;
1081 rc = rpc->crpc_reqstev.ev_status;
1084 wi->wi_state = SWI_STATE_REQUEST_SENT;
1085 /* perhaps more events, fall thru */
1086 case SWI_STATE_REQUEST_SENT: {
1087 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1089 if (!rpc->crpc_replyev.ev_fired) break;
1091 rc = rpc->crpc_replyev.ev_status;
1094 if ((reply->msg_type != type &&
1095 reply->msg_type != __swab32(type)) ||
1096 (reply->msg_magic != SRPC_MSG_MAGIC &&
1097 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1098 CWARN ("Bad message from %s: type %u (%d expected),"
1099 " magic %u (%d expected).\n",
1100 libcfs_id2str(rpc->crpc_dest),
1101 reply->msg_type, type,
1102 reply->msg_magic, SRPC_MSG_MAGIC);
1107 if (do_bulk && reply->msg_body.reply.status != 0) {
1108 CWARN ("Remote error %d at %s, unlink bulk buffer in "
1109 "case peer didn't initiate bulk transfer\n",
1110 reply->msg_body.reply.status,
1111 libcfs_id2str(rpc->crpc_dest));
1112 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1115 wi->wi_state = SWI_STATE_REPLY_RECEIVED;
1117 case SWI_STATE_REPLY_RECEIVED:
1118 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1120 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1122 /* Bulk buffer was unlinked due to remote error. Clear error
1123 * since reply buffer still contains valid data.
1124 * NB rpc->crpc_done shouldn't look into bulk data in case of
1126 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1127 rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1130 wi->wi_state = SWI_STATE_DONE;
1131 srpc_client_rpc_done(rpc, rc);
1136 cfs_spin_lock(&rpc->crpc_lock);
1137 srpc_abort_rpc(rpc, rc);
1138 cfs_spin_unlock(&rpc->crpc_lock);
1142 if (rpc->crpc_aborted) {
1143 LNetMDUnlink(rpc->crpc_reqstmdh);
1144 LNetMDUnlink(rpc->crpc_replymdh);
1145 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1147 if (!srpc_event_pending(rpc)) {
1148 srpc_client_rpc_done(rpc, -EINTR);
1156 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1157 int nbulkiov, int bulklen,
1158 void (*rpc_done)(srpc_client_rpc_t *),
1159 void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1161 srpc_client_rpc_t *rpc;
1163 LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1164 crpc_bulk.bk_iovs[nbulkiov]));
1168 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1169 bulklen, rpc_done, rpc_fini, priv);
1173 /* called with rpc->crpc_lock held */
1175 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1179 if (rpc->crpc_aborted || /* already aborted */
1180 rpc->crpc_closed) /* callback imminent */
1184 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1185 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1186 swi_state2str(rpc->crpc_wi.wi_state), why);
1188 rpc->crpc_aborted = 1;
1189 rpc->crpc_status = why;
1190 swi_schedule_workitem(&rpc->crpc_wi);
1194 /* called with rpc->crpc_lock held */
1196 srpc_post_rpc (srpc_client_rpc_t *rpc)
1198 LASSERT (!rpc->crpc_aborted);
1199 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1200 LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1202 CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1203 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1206 srpc_add_client_rpc_timer(rpc);
1207 swi_schedule_workitem(&rpc->crpc_wi);
1213 srpc_send_reply (srpc_server_rpc_t *rpc)
1215 srpc_event_t *ev = &rpc->srpc_ev;
1216 srpc_msg_t *msg = &rpc->srpc_replymsg;
1217 srpc_buffer_t *buffer = rpc->srpc_reqstbuf;
1218 srpc_service_t *sv = rpc->srpc_service;
1222 LASSERT (buffer != NULL);
1223 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1225 cfs_spin_lock(&sv->sv_lock);
1227 if (!sv->sv_shuttingdown &&
1228 sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1229 /* Repost buffer before replying since test client
1230 * might send me another RPC once it gets the reply */
1231 if (srpc_service_post_buffer(sv, buffer) != 0)
1232 CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1233 rpc->srpc_reqstbuf = NULL;
1236 cfs_spin_unlock(&sv->sv_lock);
1240 ev->ev_type = SRPC_REPLY_SENT;
1242 msg->msg_magic = SRPC_MSG_MAGIC;
1243 msg->msg_version = SRPC_MSG_VERSION;
1244 msg->msg_type = srpc_service2reply(sv->sv_id);
1246 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1247 sizeof(*msg), LNET_MD_OP_PUT,
1248 rpc->srpc_peer, rpc->srpc_self,
1249 &rpc->srpc_replymdh, ev);
1251 ev->ev_fired = 1; /* no more event expected */
1255 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1257 srpc_lnet_ev_handler (lnet_event_t *ev)
1259 srpc_event_t *rpcev = ev->md.user_ptr;
1260 srpc_client_rpc_t *crpc;
1261 srpc_server_rpc_t *srpc;
1262 srpc_buffer_t *buffer;
1265 srpc_msg_type_t type;
1267 LASSERT (!cfs_in_interrupt());
1269 if (ev->status != 0) {
1270 cfs_spin_lock(&srpc_data.rpc_glock);
1271 srpc_data.rpc_counters.errors++;
1272 cfs_spin_unlock(&srpc_data.rpc_glock);
1275 rpcev->ev_lnet = ev->type;
1277 switch (rpcev->ev_type) {
1279 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1280 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1282 case SRPC_REQUEST_SENT:
1283 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1284 cfs_spin_lock(&srpc_data.rpc_glock);
1285 srpc_data.rpc_counters.rpcs_sent++;
1286 cfs_spin_unlock(&srpc_data.rpc_glock);
1288 case SRPC_REPLY_RCVD:
1289 case SRPC_BULK_REQ_RCVD:
1290 crpc = rpcev->ev_data;
1292 if (rpcev != &crpc->crpc_reqstev &&
1293 rpcev != &crpc->crpc_replyev &&
1294 rpcev != &crpc->crpc_bulkev) {
1295 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1296 rpcev, crpc, &crpc->crpc_reqstev,
1297 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1298 CERROR("Bad event: status %d, type %d, lnet %d\n",
1299 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1303 cfs_spin_lock(&crpc->crpc_lock);
1305 LASSERT (rpcev->ev_fired == 0);
1306 rpcev->ev_fired = 1;
1307 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1308 -EINTR : ev->status;
1309 swi_schedule_workitem(&crpc->crpc_wi);
1311 cfs_spin_unlock(&crpc->crpc_lock);
1314 case SRPC_REQUEST_RCVD:
1315 sv = rpcev->ev_data;
1317 LASSERT (rpcev == &sv->sv_ev);
1319 cfs_spin_lock(&sv->sv_lock);
1321 LASSERT (ev->unlinked);
1322 LASSERT (ev->type == LNET_EVENT_PUT ||
1323 ev->type == LNET_EVENT_UNLINK);
1324 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1325 sv->sv_shuttingdown);
1327 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1328 buffer->buf_peer = ev->initiator;
1329 buffer->buf_self = ev->target.nid;
1331 sv->sv_nposted_msg--;
1332 LASSERT (sv->sv_nposted_msg >= 0);
1334 if (sv->sv_shuttingdown) {
1335 /* Leave buffer on sv->sv_posted_msgq since
1336 * srpc_finish_service needs to traverse it. */
1337 cfs_spin_unlock(&sv->sv_lock);
1341 cfs_list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1342 msg = &buffer->buf_msg;
1343 type = srpc_service2request(sv->sv_id);
1345 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1346 (msg->msg_type != type &&
1347 msg->msg_type != __swab32(type)) ||
1348 (msg->msg_magic != SRPC_MSG_MAGIC &&
1349 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1350 CERROR ("Dropping RPC (%s) from %s: "
1351 "status %d mlength %d type %u magic %u.\n",
1352 sv->sv_name, libcfs_id2str(ev->initiator),
1353 ev->status, ev->mlength,
1354 msg->msg_type, msg->msg_magic);
1356 /* NB can't call srpc_service_recycle_buffer here since
1357 * it may call LNetM[DE]Attach. The invalid magic tells
1358 * srpc_handle_rpc to drop this RPC */
1362 if (!cfs_list_empty(&sv->sv_free_rpcq)) {
1363 srpc = cfs_list_entry(sv->sv_free_rpcq.next,
1364 srpc_server_rpc_t, srpc_list);
1365 cfs_list_del(&srpc->srpc_list);
1367 srpc_init_server_rpc(srpc, sv, buffer);
1368 cfs_list_add_tail(&srpc->srpc_list,
1369 &sv->sv_active_rpcq);
1370 srpc_schedule_server_rpc(srpc);
1372 cfs_list_add_tail(&buffer->buf_list,
1373 &sv->sv_blocked_msgq);
1376 cfs_spin_unlock(&sv->sv_lock);
1378 cfs_spin_lock(&srpc_data.rpc_glock);
1379 srpc_data.rpc_counters.rpcs_rcvd++;
1380 cfs_spin_unlock(&srpc_data.rpc_glock);
1383 case SRPC_BULK_GET_RPLD:
1384 LASSERT (ev->type == LNET_EVENT_SEND ||
1385 ev->type == LNET_EVENT_REPLY ||
1386 ev->type == LNET_EVENT_UNLINK);
1389 break; /* wait for final event */
1391 case SRPC_BULK_PUT_SENT:
1392 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1393 cfs_spin_lock(&srpc_data.rpc_glock);
1395 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1396 srpc_data.rpc_counters.bulk_get += ev->mlength;
1398 srpc_data.rpc_counters.bulk_put += ev->mlength;
1400 cfs_spin_unlock(&srpc_data.rpc_glock);
1402 case SRPC_REPLY_SENT:
1403 srpc = rpcev->ev_data;
1404 sv = srpc->srpc_service;
1406 LASSERT (rpcev == &srpc->srpc_ev);
1408 cfs_spin_lock(&sv->sv_lock);
1410 rpcev->ev_fired = 1;
1411 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1412 -EINTR : ev->status;
1413 srpc_schedule_server_rpc(srpc);
1415 cfs_spin_unlock(&sv->sv_lock);
1425 srpc_check_event (int timeout)
1431 rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1432 timeout * 1000, &ev, &i);
1433 if (rc == 0) return 0;
1435 LASSERT (rc == -EOVERFLOW || rc == 1);
1437 /* We can't affort to miss any events... */
1438 if (rc == -EOVERFLOW) {
1439 CERROR ("Dropped an event!!!\n");
1443 srpc_lnet_ev_handler(&ev);
1454 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1455 cfs_spin_lock_init(&srpc_data.rpc_glock);
1457 /* 1 second pause to avoid timestamp reuse */
1458 cfs_pause(cfs_time_seconds(1));
1459 srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1461 srpc_data.rpc_state = SRPC_STATE_NONE;
1464 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1466 if (the_lnet.ln_server_mode_flag)
1467 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1469 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1472 CERROR ("LNetNIInit() has failed: %d\n", rc);
1476 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1478 LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1480 rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1482 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1485 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1489 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1492 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1498 srpc_data.rpc_state = SRPC_STATE_WI_INIT;
1506 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1512 srpc_shutdown (void)
1518 state = srpc_data.rpc_state;
1519 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1524 case SRPC_STATE_RUNNING:
1525 cfs_spin_lock(&srpc_data.rpc_glock);
1527 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1528 srpc_service_t *sv = srpc_data.rpc_services[i];
1530 LASSERTF (sv == NULL,
1531 "service not empty: id %d, name %s\n",
1535 cfs_spin_unlock(&srpc_data.rpc_glock);
1539 case SRPC_STATE_WI_INIT:
1542 case SRPC_STATE_EQ_INIT:
1543 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1545 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1546 LASSERT (rc == 0); /* the EQ should have no user by now */
1548 case SRPC_STATE_NI_INIT: