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) 2002, 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.
37 #define DEBUG_SUBSYSTEM S_RPC
40 # include <liblustre.h>
42 # include <libcfs/libcfs.h>
44 # include <linux/kernel.h>
48 #include <obd_class.h>
49 #include <lustre_net.h>
50 #include <lustre_sec.h>
51 #include "ptlrpc_internal.h"
53 lnet_handle_eq_t ptlrpc_eq_h;
56 * Client's outgoing request callback
58 void request_out_callback(lnet_event_t *ev)
60 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
61 struct ptlrpc_request *req = cbid->cbid_arg;
64 LASSERT (ev->type == LNET_EVENT_SEND ||
65 ev->type == LNET_EVENT_UNLINK);
66 LASSERT (ev->unlinked);
68 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
69 "type %d, status %d", ev->type, ev->status);
71 sptlrpc_request_out_callback(req);
72 req->rq_real_sent = cfs_time_current_sec();
74 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
76 /* Failed send: make it seem like the reply timed out, just
77 * like failing sends in client.c does currently... */
79 cfs_spin_lock(&req->rq_lock);
81 cfs_spin_unlock(&req->rq_lock);
83 ptlrpc_client_wake_req(req);
86 ptlrpc_req_finished(req);
92 * Client's incoming reply callback
94 void reply_in_callback(lnet_event_t *ev)
96 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
97 struct ptlrpc_request *req = cbid->cbid_arg;
100 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
101 "type %d, status %d", ev->type, ev->status);
103 LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
104 LASSERT (ev->md.start == req->rq_repbuf);
105 LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
106 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
107 for adaptive timeouts' early reply. */
108 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
110 cfs_spin_lock(&req->rq_lock);
112 req->rq_receiving_reply = 0;
115 req->rq_must_unlink = 0;
120 if (ev->type == LNET_EVENT_UNLINK) {
121 LASSERT(ev->unlinked);
122 DEBUG_REQ(D_RPCTRACE, req, "unlink");
126 if (ev->mlength < ev->rlength ) {
127 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
128 req->rq_replen, ev->rlength, ev->offset);
129 req->rq_reply_truncate = 1;
131 req->rq_status = -EOVERFLOW;
132 req->rq_nob_received = ev->rlength + ev->offset;
136 if ((ev->offset == 0) &&
137 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
139 DEBUG_REQ(D_ADAPTTO, req,
140 "Early reply received: mlen=%u offset=%d replen=%d "
141 "replied=%d unlinked=%d", ev->mlength, ev->offset,
142 req->rq_replen, req->rq_replied, ev->unlinked);
144 req->rq_early_count++; /* number received, client side */
146 if (req->rq_replied) /* already got the real reply */
150 req->rq_reply_off = ev->offset;
151 req->rq_nob_received = ev->mlength;
152 /* And we're still receiving */
153 req->rq_receiving_reply = 1;
156 req->rq_rep_swab_mask = 0;
158 req->rq_reply_off = ev->offset;
159 req->rq_nob_received = ev->mlength;
160 /* LNetMDUnlink can't be called under the LNET_LOCK,
161 so we must unlink in ptlrpc_unregister_reply */
162 DEBUG_REQ(D_INFO, req,
163 "reply in flags=%x mlen=%u offset=%d replen=%d",
164 lustre_msg_get_flags(req->rq_reqmsg),
165 ev->mlength, ev->offset, req->rq_replen);
168 req->rq_import->imp_last_reply_time = cfs_time_current_sec();
171 /* NB don't unlock till after wakeup; req can disappear under us
172 * since we don't have our own ref */
173 ptlrpc_client_wake_req(req);
174 cfs_spin_unlock(&req->rq_lock);
179 * Client's bulk has been written/read
181 void client_bulk_callback (lnet_event_t *ev)
183 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
184 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
187 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
188 ev->type == LNET_EVENT_PUT) ||
189 (desc->bd_type == BULK_GET_SOURCE &&
190 ev->type == LNET_EVENT_GET) ||
191 ev->type == LNET_EVENT_UNLINK);
192 LASSERT (ev->unlinked);
194 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
195 "event type %d, status %d, desc %p\n",
196 ev->type, ev->status, desc);
198 cfs_spin_lock(&desc->bd_lock);
200 LASSERT(desc->bd_network_rw);
201 desc->bd_network_rw = 0;
203 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
204 desc->bd_success = 1;
205 desc->bd_nob_transferred = ev->mlength;
206 desc->bd_sender = ev->sender;
209 /* release the encrypted pages for write */
210 if (desc->bd_req->rq_bulk_write)
211 sptlrpc_enc_pool_put_pages(desc);
213 /* NB don't unlock till after wakeup; desc can disappear under us
215 ptlrpc_client_wake_req(desc->bd_req);
217 cfs_spin_unlock(&desc->bd_lock);
222 * Server's incoming request callback
224 void request_in_callback(lnet_event_t *ev)
226 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
227 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
228 struct ptlrpc_service *service = rqbd->rqbd_service;
229 struct ptlrpc_request *req;
232 LASSERT (ev->type == LNET_EVENT_PUT ||
233 ev->type == LNET_EVENT_UNLINK);
234 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
235 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
236 rqbd->rqbd_buffer + service->srv_buf_size);
238 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
239 "event type %d, status %d, service %s\n",
240 ev->type, ev->status, service->srv_name);
243 /* If this is the last request message to fit in the
244 * request buffer we can use the request object embedded in
245 * rqbd. Note that if we failed to allocate a request,
246 * we'd have to re-post the rqbd, which we can't do in this
248 req = &rqbd->rqbd_req;
249 memset(req, 0, sizeof (*req));
251 LASSERT (ev->type == LNET_EVENT_PUT);
252 if (ev->status != 0) {
253 /* We moaned above already... */
256 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
258 CERROR("Can't allocate incoming request descriptor: "
259 "Dropping %s RPC from %s\n",
261 libcfs_id2str(ev->initiator));
266 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
267 * flags are reset and scalars are zero. We only set the message
268 * size to non-zero if this was a successful receive. */
269 req->rq_xid = ev->match_bits;
270 req->rq_reqbuf = ev->md.start + ev->offset;
271 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
272 req->rq_reqdata_len = ev->mlength;
273 cfs_gettimeofday(&req->rq_arrival_time);
274 req->rq_peer = ev->initiator;
275 req->rq_self = ev->target.nid;
277 req->rq_phase = RQ_PHASE_NEW;
279 req->rq_uid = ev->uid;
281 cfs_spin_lock_init(&req->rq_lock);
282 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
283 cfs_atomic_set(&req->rq_refcount, 1);
284 if (ev->type == LNET_EVENT_PUT)
285 CDEBUG(D_RPCTRACE, "incoming req@%p x"LPU64" msgsize %u\n",
286 req, req->rq_xid, ev->mlength);
288 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
290 cfs_spin_lock(&service->srv_lock);
292 req->rq_history_seq = service->srv_request_seq++;
293 cfs_list_add_tail(&req->rq_history_list, &service->srv_request_history);
296 service->srv_nrqbd_receiving--;
297 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
298 service->srv_nrqbd_receiving);
300 /* Normally, don't complain about 0 buffers posted; LNET won't
301 * drop incoming reqs since we set the portal lazy */
302 if (test_req_buffer_pressure &&
303 ev->type != LNET_EVENT_UNLINK &&
304 service->srv_nrqbd_receiving == 0)
305 CWARN("All %s request buffers busy\n",
308 /* req takes over the network's ref on rqbd */
310 /* req takes a ref on rqbd */
311 rqbd->rqbd_refcount++;
314 cfs_list_add_tail(&req->rq_list, &service->srv_req_in_queue);
315 service->srv_n_queued_reqs++;
317 /* NB everything can disappear under us once the request
318 * has been queued and we unlock, so do the wake now... */
319 cfs_waitq_signal(&service->srv_waitq);
321 cfs_spin_unlock(&service->srv_lock);
326 * Server's outgoing reply callback
328 void reply_out_callback(lnet_event_t *ev)
330 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
331 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
332 struct ptlrpc_service *svc = rs->rs_service;
335 LASSERT (ev->type == LNET_EVENT_SEND ||
336 ev->type == LNET_EVENT_ACK ||
337 ev->type == LNET_EVENT_UNLINK);
339 if (!rs->rs_difficult) {
340 /* 'Easy' replies have no further processing so I drop the
341 * net's ref on 'rs' */
342 LASSERT (ev->unlinked);
343 ptlrpc_rs_decref(rs);
344 cfs_atomic_dec (&svc->srv_outstanding_replies);
349 LASSERT (rs->rs_on_net);
352 /* Last network callback. The net's ref on 'rs' stays put
353 * until ptlrpc_handle_rs() is done with it */
354 cfs_spin_lock(&svc->srv_lock);
355 cfs_spin_lock(&rs->rs_lock);
357 if (!rs->rs_no_ack ||
358 rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
359 ptlrpc_schedule_difficult_reply (rs);
360 cfs_spin_unlock(&rs->rs_lock);
361 cfs_spin_unlock(&svc->srv_lock);
368 * Server's bulk completion callback
370 void server_bulk_callback (lnet_event_t *ev)
372 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
373 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
376 LASSERT (ev->type == LNET_EVENT_SEND ||
377 ev->type == LNET_EVENT_UNLINK ||
378 (desc->bd_type == BULK_PUT_SOURCE &&
379 ev->type == LNET_EVENT_ACK) ||
380 (desc->bd_type == BULK_GET_SINK &&
381 ev->type == LNET_EVENT_REPLY));
383 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
384 "event type %d, status %d, desc %p\n",
385 ev->type, ev->status, desc);
387 cfs_spin_lock(&desc->bd_lock);
389 if ((ev->type == LNET_EVENT_ACK ||
390 ev->type == LNET_EVENT_REPLY) &&
392 /* We heard back from the peer, so even if we get this
393 * before the SENT event (oh yes we can), we know we
394 * read/wrote the peer buffer and how much... */
395 desc->bd_success = 1;
396 desc->bd_nob_transferred = ev->mlength;
397 desc->bd_sender = ev->sender;
401 /* This is the last callback no matter what... */
402 desc->bd_network_rw = 0;
403 cfs_waitq_signal(&desc->bd_waitq);
406 cfs_spin_unlock(&desc->bd_lock);
410 static void ptlrpc_master_callback(lnet_event_t *ev)
412 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
413 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
415 /* Honestly, it's best to find out early. */
416 LASSERT (cbid->cbid_arg != LP_POISON);
417 LASSERT (callback == request_out_callback ||
418 callback == reply_in_callback ||
419 callback == client_bulk_callback ||
420 callback == request_in_callback ||
421 callback == reply_out_callback ||
422 callback == server_bulk_callback);
427 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
428 lnet_process_id_t *peer, lnet_nid_t *self)
431 __u32 best_order = 0;
434 int portals_compatibility;
440 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
442 peer->pid = LUSTRE_SRV_LNET_PID;
444 /* Choose the matching UUID that's closest */
445 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
446 dist = LNetDist(dst_nid, &src_nid, &order);
450 if (dist == 0) { /* local! use loopback LND */
451 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
458 (dist == best_dist && order < best_order)) {
462 if (portals_compatibility > 1) {
463 /* Strong portals compatibility: Zero the nid's
464 * NET, so if I'm reading new config logs, or
465 * getting configured by (new) lconf I can
466 * still talk to old servers. */
467 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
468 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
476 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
480 void ptlrpc_ni_fini(void)
483 struct l_wait_info lwi;
487 /* Wait for the event queue to become idle since there may still be
488 * messages in flight with pending events (i.e. the fire-and-forget
489 * messages == client requests and "non-difficult" server
492 for (retries = 0;; retries++) {
493 rc = LNetEQFree(ptlrpc_eq_h);
504 CWARN("Event queue still busy\n");
507 cfs_waitq_init(&waitq);
508 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
509 l_wait_event(waitq, 0, &lwi);
516 lnet_pid_t ptl_get_pid(void)
523 pid = LUSTRE_SRV_LNET_PID;
528 int ptlrpc_ni_init(void)
534 CDEBUG(D_NET, "My pid is: %x\n", pid);
536 /* We're not passing any limits yet... */
537 rc = LNetNIInit(pid);
539 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
543 /* CAVEAT EMPTOR: how we process portals events is _radically_
544 * different depending on... */
546 /* kernel portals calls our master callback when events are added to
547 * the event queue. In fact lustre never pulls events off this queue,
548 * so it's only sized for some debug history. */
549 rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
551 /* liblustre calls the master callback when it removes events from the
552 * event queue. The event queue has to be big enough not to drop
554 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
559 CERROR ("Failed to allocate event queue: %d\n", rc);
566 CFS_LIST_HEAD(liblustre_wait_callbacks);
567 CFS_LIST_HEAD(liblustre_idle_callbacks);
568 void *liblustre_services_callback;
571 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
573 int (*fn)(void *arg), void *arg)
575 struct liblustre_wait_callback *llwc;
577 OBD_ALLOC(llwc, sizeof(*llwc));
578 LASSERT (llwc != NULL);
580 llwc->llwc_name = name;
582 llwc->llwc_arg = arg;
583 cfs_list_add_tail(&llwc->llwc_list, callback_list);
589 liblustre_deregister_waitidle_callback (void *opaque)
591 struct liblustre_wait_callback *llwc = opaque;
593 cfs_list_del(&llwc->llwc_list);
594 OBD_FREE(llwc, sizeof(*llwc));
598 liblustre_register_wait_callback (const char *name,
599 int (*fn)(void *arg), void *arg)
601 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
606 liblustre_deregister_wait_callback (void *opaque)
608 liblustre_deregister_waitidle_callback(opaque);
612 liblustre_register_idle_callback (const char *name,
613 int (*fn)(void *arg), void *arg)
615 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
620 liblustre_deregister_idle_callback (void *opaque)
622 liblustre_deregister_waitidle_callback(opaque);
626 liblustre_check_events (int timeout)
633 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
637 LASSERT (rc == -EOVERFLOW || rc == 1);
639 /* liblustre: no asynch callback so we can't afford to miss any
641 if (rc == -EOVERFLOW) {
642 CERROR ("Dropped an event!!!\n");
646 ptlrpc_master_callback (&ev);
650 int liblustre_waiting = 0;
653 liblustre_wait_event (int timeout)
656 struct liblustre_wait_callback *llwc;
657 int found_something = 0;
659 /* single threaded recursion check... */
660 liblustre_waiting = 1;
663 /* Deal with all pending events */
664 while (liblustre_check_events(0))
667 /* Give all registered callbacks a bite at the cherry */
668 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
669 llwc = cfs_list_entry(tmp,
670 struct liblustre_wait_callback,
673 if (llwc->llwc_fn(llwc->llwc_arg))
677 if (found_something || timeout == 0)
680 /* Nothing so far, but I'm allowed to block... */
681 found_something = liblustre_check_events(timeout);
682 if (!found_something) /* still nothing */
683 break; /* I timed out */
686 liblustre_waiting = 0;
688 return found_something;
692 liblustre_wait_idle(void)
694 static int recursed = 0;
697 struct liblustre_wait_callback *llwc;
704 liblustre_wait_event(0);
708 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
709 llwc = cfs_list_entry(tmp,
710 struct liblustre_wait_callback,
713 if (!llwc->llwc_fn(llwc->llwc_arg)) {
724 #endif /* __KERNEL__ */
726 int ptlrpc_init_portals(void)
728 int rc = ptlrpc_ni_init();
731 CERROR("network initialisation failed\n");
735 liblustre_services_callback =
736 liblustre_register_wait_callback("liblustre_check_services",
737 &liblustre_check_services,
739 cfs_init_completion_module(liblustre_wait_event);
741 rc = ptlrpcd_addref();
745 CERROR("rpcd initialisation failed\n");
747 liblustre_deregister_wait_callback(liblustre_services_callback);
753 void ptlrpc_exit_portals(void)
756 liblustre_deregister_wait_callback(liblustre_services_callback);