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(D_NET, req, "type %d, status %d", ev->type, ev->status);
70 sptlrpc_request_out_callback(req);
71 req->rq_real_sent = cfs_time_current_sec();
73 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
75 /* Failed send: make it seem like the reply timed out, just
76 * like failing sends in client.c does currently... */
78 cfs_spin_lock(&req->rq_lock);
80 cfs_spin_unlock(&req->rq_lock);
82 ptlrpc_client_wake_req(req);
85 ptlrpc_req_finished(req);
91 * Client's incoming reply callback
93 void reply_in_callback(lnet_event_t *ev)
95 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
96 struct ptlrpc_request *req = cbid->cbid_arg;
99 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
101 LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
102 LASSERT (ev->md.start == req->rq_repbuf);
103 LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
104 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
105 for adaptive timeouts' early reply. */
106 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
108 cfs_spin_lock(&req->rq_lock);
110 req->rq_receiving_reply = 0;
113 req->rq_must_unlink = 0;
118 if (ev->type == LNET_EVENT_UNLINK) {
119 LASSERT(ev->unlinked);
120 DEBUG_REQ(D_NET, req, "unlink");
124 if (ev->mlength < ev->rlength ) {
125 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
126 req->rq_replen, ev->rlength, ev->offset);
127 req->rq_reply_truncate = 1;
129 req->rq_status = -EOVERFLOW;
130 req->rq_nob_received = ev->rlength + ev->offset;
134 if ((ev->offset == 0) &&
135 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
137 DEBUG_REQ(D_ADAPTTO, req,
138 "Early reply received: mlen=%u offset=%d replen=%d "
139 "replied=%d unlinked=%d", ev->mlength, ev->offset,
140 req->rq_replen, req->rq_replied, ev->unlinked);
142 req->rq_early_count++; /* number received, client side */
144 if (req->rq_replied) /* already got the real reply */
148 req->rq_reply_off = ev->offset;
149 req->rq_nob_received = ev->mlength;
150 /* And we're still receiving */
151 req->rq_receiving_reply = 1;
154 req->rq_rep_swab_mask = 0;
156 req->rq_reply_off = ev->offset;
157 req->rq_nob_received = ev->mlength;
158 /* LNetMDUnlink can't be called under the LNET_LOCK,
159 so we must unlink in ptlrpc_unregister_reply */
160 DEBUG_REQ(D_INFO, req,
161 "reply in flags=%x mlen=%u offset=%d replen=%d",
162 lustre_msg_get_flags(req->rq_reqmsg),
163 ev->mlength, ev->offset, req->rq_replen);
166 req->rq_import->imp_last_reply_time = cfs_time_current_sec();
169 /* NB don't unlock till after wakeup; req can disappear under us
170 * since we don't have our own ref */
171 ptlrpc_client_wake_req(req);
172 cfs_spin_unlock(&req->rq_lock);
177 * Client's bulk has been written/read
179 void client_bulk_callback (lnet_event_t *ev)
181 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
182 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
185 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
186 ev->type == LNET_EVENT_PUT) ||
187 (desc->bd_type == BULK_GET_SOURCE &&
188 ev->type == LNET_EVENT_GET) ||
189 ev->type == LNET_EVENT_UNLINK);
190 LASSERT (ev->unlinked);
192 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
193 "event type %d, status %d, desc %p\n",
194 ev->type, ev->status, desc);
196 cfs_spin_lock(&desc->bd_lock);
198 LASSERT(desc->bd_network_rw);
199 desc->bd_network_rw = 0;
201 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
202 desc->bd_success = 1;
203 desc->bd_nob_transferred = ev->mlength;
204 desc->bd_sender = ev->sender;
207 /* release the encrypted pages for write */
208 if (desc->bd_req->rq_bulk_write)
209 sptlrpc_enc_pool_put_pages(desc);
211 /* NB don't unlock till after wakeup; desc can disappear under us
213 ptlrpc_client_wake_req(desc->bd_req);
215 cfs_spin_unlock(&desc->bd_lock);
220 * Server's incoming request callback
222 void request_in_callback(lnet_event_t *ev)
224 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
225 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
226 struct ptlrpc_service *service = rqbd->rqbd_service;
227 struct ptlrpc_request *req;
230 LASSERT (ev->type == LNET_EVENT_PUT ||
231 ev->type == LNET_EVENT_UNLINK);
232 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
233 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
234 rqbd->rqbd_buffer + service->srv_buf_size);
236 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
237 "event type %d, status %d, service %s\n",
238 ev->type, ev->status, service->srv_name);
241 /* If this is the last request message to fit in the
242 * request buffer we can use the request object embedded in
243 * rqbd. Note that if we failed to allocate a request,
244 * we'd have to re-post the rqbd, which we can't do in this
246 req = &rqbd->rqbd_req;
247 memset(req, 0, sizeof (*req));
249 LASSERT (ev->type == LNET_EVENT_PUT);
250 if (ev->status != 0) {
251 /* We moaned above already... */
254 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
256 CERROR("Can't allocate incoming request descriptor: "
257 "Dropping %s RPC from %s\n",
259 libcfs_id2str(ev->initiator));
264 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
265 * flags are reset and scalars are zero. We only set the message
266 * size to non-zero if this was a successful receive. */
267 req->rq_xid = ev->match_bits;
268 req->rq_reqbuf = ev->md.start + ev->offset;
269 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
270 req->rq_reqdata_len = ev->mlength;
271 cfs_gettimeofday(&req->rq_arrival_time);
272 req->rq_peer = ev->initiator;
273 req->rq_self = ev->target.nid;
275 req->rq_phase = RQ_PHASE_NEW;
277 req->rq_uid = ev->uid;
279 cfs_spin_lock_init(&req->rq_lock);
280 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
281 cfs_atomic_set(&req->rq_refcount, 1);
282 if (ev->type == LNET_EVENT_PUT)
283 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
284 req, req->rq_xid, ev->mlength);
286 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
288 cfs_spin_lock(&service->srv_lock);
290 req->rq_history_seq = service->srv_request_seq++;
291 cfs_list_add_tail(&req->rq_history_list, &service->srv_request_history);
294 service->srv_nrqbd_receiving--;
295 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
296 service->srv_nrqbd_receiving);
298 /* Normally, don't complain about 0 buffers posted; LNET won't
299 * drop incoming reqs since we set the portal lazy */
300 if (test_req_buffer_pressure &&
301 ev->type != LNET_EVENT_UNLINK &&
302 service->srv_nrqbd_receiving == 0)
303 CWARN("All %s request buffers busy\n",
306 /* req takes over the network's ref on rqbd */
308 /* req takes a ref on rqbd */
309 rqbd->rqbd_refcount++;
312 cfs_list_add_tail(&req->rq_list, &service->srv_req_in_queue);
313 service->srv_n_queued_reqs++;
315 /* NB everything can disappear under us once the request
316 * has been queued and we unlock, so do the wake now... */
317 cfs_waitq_signal(&service->srv_waitq);
319 cfs_spin_unlock(&service->srv_lock);
324 * Server's outgoing reply callback
326 void reply_out_callback(lnet_event_t *ev)
328 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
329 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
330 struct ptlrpc_service *svc = rs->rs_service;
333 LASSERT (ev->type == LNET_EVENT_SEND ||
334 ev->type == LNET_EVENT_ACK ||
335 ev->type == LNET_EVENT_UNLINK);
337 if (!rs->rs_difficult) {
338 /* 'Easy' replies have no further processing so I drop the
339 * net's ref on 'rs' */
340 LASSERT (ev->unlinked);
341 ptlrpc_rs_decref(rs);
342 cfs_atomic_dec (&svc->srv_outstanding_replies);
347 LASSERT (rs->rs_on_net);
350 /* Last network callback. The net's ref on 'rs' stays put
351 * until ptlrpc_handle_rs() is done with it */
352 cfs_spin_lock(&svc->srv_lock);
353 cfs_spin_lock(&rs->rs_lock);
355 if (!rs->rs_no_ack ||
356 rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
357 ptlrpc_schedule_difficult_reply (rs);
358 cfs_spin_unlock(&rs->rs_lock);
359 cfs_spin_unlock(&svc->srv_lock);
366 * Server's bulk completion callback
368 void server_bulk_callback (lnet_event_t *ev)
370 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
371 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
374 LASSERT (ev->type == LNET_EVENT_SEND ||
375 ev->type == LNET_EVENT_UNLINK ||
376 (desc->bd_type == BULK_PUT_SOURCE &&
377 ev->type == LNET_EVENT_ACK) ||
378 (desc->bd_type == BULK_GET_SINK &&
379 ev->type == LNET_EVENT_REPLY));
381 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
382 "event type %d, status %d, desc %p\n",
383 ev->type, ev->status, desc);
385 cfs_spin_lock(&desc->bd_lock);
387 if ((ev->type == LNET_EVENT_ACK ||
388 ev->type == LNET_EVENT_REPLY) &&
390 /* We heard back from the peer, so even if we get this
391 * before the SENT event (oh yes we can), we know we
392 * read/wrote the peer buffer and how much... */
393 desc->bd_success = 1;
394 desc->bd_nob_transferred = ev->mlength;
395 desc->bd_sender = ev->sender;
399 /* This is the last callback no matter what... */
400 desc->bd_network_rw = 0;
401 cfs_waitq_signal(&desc->bd_waitq);
404 cfs_spin_unlock(&desc->bd_lock);
408 static void ptlrpc_master_callback(lnet_event_t *ev)
410 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
411 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
413 /* Honestly, it's best to find out early. */
414 LASSERT (cbid->cbid_arg != LP_POISON);
415 LASSERT (callback == request_out_callback ||
416 callback == reply_in_callback ||
417 callback == client_bulk_callback ||
418 callback == request_in_callback ||
419 callback == reply_out_callback ||
420 callback == server_bulk_callback);
425 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
426 lnet_process_id_t *peer, lnet_nid_t *self)
429 __u32 best_order = 0;
432 int portals_compatibility;
438 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
440 peer->pid = LUSTRE_SRV_LNET_PID;
442 /* Choose the matching UUID that's closest */
443 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
444 dist = LNetDist(dst_nid, &src_nid, &order);
448 if (dist == 0) { /* local! use loopback LND */
449 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
456 (dist == best_dist && order < best_order)) {
460 if (portals_compatibility > 1) {
461 /* Strong portals compatibility: Zero the nid's
462 * NET, so if I'm reading new config logs, or
463 * getting configured by (new) lconf I can
464 * still talk to old servers. */
465 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
466 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
474 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
478 void ptlrpc_ni_fini(void)
481 struct l_wait_info lwi;
485 /* Wait for the event queue to become idle since there may still be
486 * messages in flight with pending events (i.e. the fire-and-forget
487 * messages == client requests and "non-difficult" server
490 for (retries = 0;; retries++) {
491 rc = LNetEQFree(ptlrpc_eq_h);
502 CWARN("Event queue still busy\n");
505 cfs_waitq_init(&waitq);
506 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
507 l_wait_event(waitq, 0, &lwi);
514 lnet_pid_t ptl_get_pid(void)
521 pid = LUSTRE_SRV_LNET_PID;
526 int ptlrpc_ni_init(void)
532 CDEBUG(D_NET, "My pid is: %x\n", pid);
534 /* We're not passing any limits yet... */
535 rc = LNetNIInit(pid);
537 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
541 /* CAVEAT EMPTOR: how we process portals events is _radically_
542 * different depending on... */
544 /* kernel portals calls our master callback when events are added to
545 * the event queue. In fact lustre never pulls events off this queue,
546 * so it's only sized for some debug history. */
547 rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
549 /* liblustre calls the master callback when it removes events from the
550 * event queue. The event queue has to be big enough not to drop
552 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
557 CERROR ("Failed to allocate event queue: %d\n", rc);
564 CFS_LIST_HEAD(liblustre_wait_callbacks);
565 CFS_LIST_HEAD(liblustre_idle_callbacks);
566 void *liblustre_services_callback;
569 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
571 int (*fn)(void *arg), void *arg)
573 struct liblustre_wait_callback *llwc;
575 OBD_ALLOC(llwc, sizeof(*llwc));
576 LASSERT (llwc != NULL);
578 llwc->llwc_name = name;
580 llwc->llwc_arg = arg;
581 cfs_list_add_tail(&llwc->llwc_list, callback_list);
587 liblustre_deregister_waitidle_callback (void *opaque)
589 struct liblustre_wait_callback *llwc = opaque;
591 cfs_list_del(&llwc->llwc_list);
592 OBD_FREE(llwc, sizeof(*llwc));
596 liblustre_register_wait_callback (const char *name,
597 int (*fn)(void *arg), void *arg)
599 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
604 liblustre_deregister_wait_callback (void *opaque)
606 liblustre_deregister_waitidle_callback(opaque);
610 liblustre_register_idle_callback (const char *name,
611 int (*fn)(void *arg), void *arg)
613 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
618 liblustre_deregister_idle_callback (void *opaque)
620 liblustre_deregister_waitidle_callback(opaque);
624 liblustre_check_events (int timeout)
631 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
635 LASSERT (rc == -EOVERFLOW || rc == 1);
637 /* liblustre: no asynch callback so we can't afford to miss any
639 if (rc == -EOVERFLOW) {
640 CERROR ("Dropped an event!!!\n");
644 ptlrpc_master_callback (&ev);
648 int liblustre_waiting = 0;
651 liblustre_wait_event (int timeout)
654 struct liblustre_wait_callback *llwc;
655 int found_something = 0;
657 /* single threaded recursion check... */
658 liblustre_waiting = 1;
661 /* Deal with all pending events */
662 while (liblustre_check_events(0))
665 /* Give all registered callbacks a bite at the cherry */
666 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
667 llwc = cfs_list_entry(tmp,
668 struct liblustre_wait_callback,
671 if (llwc->llwc_fn(llwc->llwc_arg))
675 if (found_something || timeout == 0)
678 /* Nothing so far, but I'm allowed to block... */
679 found_something = liblustre_check_events(timeout);
680 if (!found_something) /* still nothing */
681 break; /* I timed out */
684 liblustre_waiting = 0;
686 return found_something;
690 liblustre_wait_idle(void)
692 static int recursed = 0;
695 struct liblustre_wait_callback *llwc;
702 liblustre_wait_event(0);
706 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
707 llwc = cfs_list_entry(tmp,
708 struct liblustre_wait_callback,
711 if (!llwc->llwc_fn(llwc->llwc_arg)) {
722 #endif /* __KERNEL__ */
724 int ptlrpc_init_portals(void)
726 int rc = ptlrpc_ni_init();
729 CERROR("network initialisation failed\n");
733 liblustre_services_callback =
734 liblustre_register_wait_callback("liblustre_check_services",
735 &liblustre_check_services,
737 cfs_init_completion_module(liblustre_wait_event);
739 rc = ptlrpcd_addref();
743 CERROR("rpcd initialisation failed\n");
745 liblustre_deregister_wait_callback(liblustre_services_callback);
751 void ptlrpc_exit_portals(void)
754 liblustre_deregister_wait_callback(liblustre_services_callback);