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>
43 # include <linux/kernel.h>
46 #include <obd_class.h>
47 #include <lustre_net.h>
48 #include "ptlrpc_internal.h"
50 lnet_handle_eq_t ptlrpc_eq_h;
53 * Client's outgoing request callback
55 void request_out_callback(lnet_event_t *ev)
57 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
58 struct ptlrpc_request *req = cbid->cbid_arg;
61 LASSERT (ev->type == LNET_EVENT_SEND ||
62 ev->type == LNET_EVENT_UNLINK);
63 LASSERT (ev->unlinked);
65 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
67 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
69 /* Failed send: make it seem like the reply timed out, just
70 * like failing sends in client.c does currently... */
72 spin_lock(&req->rq_lock);
74 spin_unlock(&req->rq_lock);
76 ptlrpc_client_wake_req(req);
79 ptlrpc_req_finished(req);
85 * Client's incoming reply callback
87 void reply_in_callback(lnet_event_t *ev)
89 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
90 struct ptlrpc_request *req = cbid->cbid_arg;
93 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
95 LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
96 LASSERT(ev->md.start == req->rq_repbuf);
97 LASSERT(ev->offset + ev->mlength <= req->rq_replen);
98 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
99 for adaptive timeouts' early reply. */
100 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
102 spin_lock(&req->rq_lock);
104 req->rq_receiving_reply = 0;
107 req->rq_must_unlink = 0;
112 if (ev->type == LNET_EVENT_UNLINK) {
113 LASSERT(ev->unlinked);
114 DEBUG_REQ(D_NET, req, "unlink");
118 if (ev->mlength < ev->rlength ) {
119 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
120 req->rq_replen, ev->rlength, ev->offset);
121 req->rq_reply_truncate = 1;
123 req->rq_status = -EOVERFLOW;
124 req->rq_nob_received = ev->rlength + ev->offset;
128 if ((ev->offset == 0) &&
129 (lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
131 DEBUG_REQ(D_ADAPTTO, req,
132 "Early reply received: mlen=%u offset=%d replen=%d "
133 "replied=%d unlinked=%d", ev->mlength, ev->offset,
134 req->rq_replen, req->rq_replied, ev->unlinked);
136 if (unlikely(ev->mlength != lustre_msg_early_size(req)))
137 CERROR("early reply sized %u, expect %u\n",
138 ev->mlength, lustre_msg_early_size(req));
140 req->rq_early_count++; /* number received, client side */
142 if (req->rq_replied) /* already got the real reply */
146 req->rq_nob_received = ev->mlength;
147 /* repmsg points to early reply */
148 req->rq_repmsg = req->rq_repbuf;
149 /* And we're still receiving */
150 req->rq_receiving_reply = 1;
153 req->rq_rep_swab_mask = 0;
155 req->rq_nob_received = ev->mlength;
156 /* repmsg points to real reply */
157 req->rq_repmsg = (struct lustre_msg *)((char *)req->rq_repbuf +
159 /* LNetMDUnlink can't be called under the LNET_LOCK,
160 so we must unlink in ptlrpc_unregister_reply */
161 DEBUG_REQ(D_INFO, req,
162 "reply in flags=%x mlen=%u offset=%d replen=%d",
163 lustre_msg_get_flags(req->rq_reqmsg),
164 ev->mlength, ev->offset, req->rq_replen);
167 req->rq_import->imp_last_reply_time = cfs_time_current_sec();
170 /* NB don't unlock till after wakeup; req can disappear under us
171 * since we don't have our own ref */
172 ptlrpc_client_wake_req(req);
173 spin_unlock(&req->rq_lock);
178 * Client's bulk has been written/read
180 void client_bulk_callback (lnet_event_t *ev)
182 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
183 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
186 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB))
189 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
190 ev->type == LNET_EVENT_PUT) ||
191 (desc->bd_type == BULK_GET_SOURCE &&
192 ev->type == LNET_EVENT_GET) ||
193 ev->type == LNET_EVENT_UNLINK);
194 LASSERT (ev->unlinked);
196 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
197 "event type %d, status %d, desc %p\n",
198 ev->type, ev->status, desc);
200 spin_lock(&desc->bd_lock);
202 LASSERT(desc->bd_network_rw);
203 desc->bd_network_rw = 0;
205 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
206 desc->bd_success = 1;
207 desc->bd_nob_transferred = ev->mlength;
208 desc->bd_sender = ev->sender;
211 /* NB don't unlock till after wakeup; desc can disappear under us
213 ptlrpc_client_wake_req(desc->bd_req);
215 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_reqmsg = ev->md.start + ev->offset;
269 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
270 req->rq_reqlen = ev->mlength;
271 do_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 spin_lock_init(&req->rq_lock);
280 CFS_INIT_LIST_HEAD(&req->rq_list);
281 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
282 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
283 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
284 CFS_INIT_LIST_HEAD(&req->rq_history_list);
285 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
286 atomic_set(&req->rq_refcount, 1);
287 if (ev->type == LNET_EVENT_PUT)
288 DEBUG_REQ(D_NET, req, "incoming req");
290 spin_lock(&service->srv_lock);
292 req->rq_history_seq = service->srv_request_seq++;
293 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 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 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 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_server_handle_reply() is done with it */
354 spin_lock(&svc->srv_lock);
356 ptlrpc_schedule_difficult_reply (rs);
357 spin_unlock(&svc->srv_lock);
364 * Server's bulk completion callback
366 void server_bulk_callback (lnet_event_t *ev)
368 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
369 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
372 LASSERT (ev->type == LNET_EVENT_SEND ||
373 ev->type == LNET_EVENT_UNLINK ||
374 (desc->bd_type == BULK_PUT_SOURCE &&
375 ev->type == LNET_EVENT_ACK) ||
376 (desc->bd_type == BULK_GET_SINK &&
377 ev->type == LNET_EVENT_REPLY));
379 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
380 "event type %d, status %d, desc %p\n",
381 ev->type, ev->status, desc);
383 spin_lock(&desc->bd_lock);
385 if ((ev->type == LNET_EVENT_ACK ||
386 ev->type == LNET_EVENT_REPLY) &&
388 /* We heard back from the peer, so even if we get this
389 * before the SENT event (oh yes we can), we know we
390 * read/wrote the peer buffer and how much... */
391 desc->bd_success = 1;
392 desc->bd_nob_transferred = ev->mlength;
393 desc->bd_sender = ev->sender;
397 /* This is the last callback no matter what... */
398 desc->bd_network_rw = 0;
399 cfs_waitq_signal(&desc->bd_waitq);
402 spin_unlock(&desc->bd_lock);
406 static void ptlrpc_master_callback(lnet_event_t *ev)
408 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
409 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
411 /* Honestly, it's best to find out early. */
412 LASSERT (cbid->cbid_arg != LP_POISON);
413 LASSERT (callback == request_out_callback ||
414 callback == reply_in_callback ||
415 callback == client_bulk_callback ||
416 callback == request_in_callback ||
417 callback == reply_out_callback ||
418 callback == server_bulk_callback);
423 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
424 lnet_process_id_t *peer, lnet_nid_t *self)
427 __u32 best_order = 0;
430 int portals_compatibility;
436 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
438 peer->pid = LUSTRE_SRV_LNET_PID;
440 /* Choose the matching UUID that's closest */
441 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
442 dist = LNetDist(dst_nid, &src_nid, &order);
446 if (dist == 0) { /* local! use loopback LND */
447 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
454 (dist == best_dist && order < best_order)) {
458 if (portals_compatibility > 1) {
459 /* Strong portals compatibility: Zero the nid's
460 * NET, so if I'm reading new config logs, or
461 * getting configured by (new) lconf I can
462 * still talk to old servers. */
463 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
464 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
472 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
476 void ptlrpc_ni_fini(void)
479 struct l_wait_info lwi;
483 /* Wait for the event queue to become idle since there may still be
484 * messages in flight with pending events (i.e. the fire-and-forget
485 * messages == client requests and "non-difficult" server
488 for (retries = 0;; retries++) {
489 rc = LNetEQFree(ptlrpc_eq_h);
500 CWARN("Event queue still busy\n");
503 cfs_waitq_init(&waitq);
504 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
505 l_wait_event(waitq, 0, &lwi);
512 lnet_pid_t ptl_get_pid(void)
519 pid = LUSTRE_SRV_LNET_PID;
524 int ptlrpc_ni_init(void)
530 CDEBUG(D_NET, "My pid is: %x\n", pid);
532 /* We're not passing any limits yet... */
533 rc = LNetNIInit(pid);
535 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
539 /* CAVEAT EMPTOR: how we process portals events is _radically_
540 * different depending on... */
542 /* kernel portals calls our master callback when events are added to
543 * the event queue. In fact lustre never pulls events off this queue,
544 * so it's only sized for some debug history. */
545 rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
547 /* liblustre calls the master callback when it removes events from the
548 * event queue. The event queue has to be big enough not to drop
550 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
555 CERROR ("Failed to allocate event queue: %d\n", rc);
562 CFS_LIST_HEAD(liblustre_wait_callbacks);
563 CFS_LIST_HEAD(liblustre_idle_callbacks);
564 void *liblustre_services_callback;
567 liblustre_register_waitidle_callback (struct list_head *callback_list,
569 int (*fn)(void *arg), void *arg)
571 struct liblustre_wait_callback *llwc;
573 OBD_ALLOC(llwc, sizeof(*llwc));
574 LASSERT (llwc != NULL);
576 llwc->llwc_name = name;
578 llwc->llwc_arg = arg;
579 list_add_tail(&llwc->llwc_list, callback_list);
585 liblustre_deregister_waitidle_callback (void *opaque)
587 struct liblustre_wait_callback *llwc = opaque;
589 list_del(&llwc->llwc_list);
590 OBD_FREE(llwc, sizeof(*llwc));
594 liblustre_register_wait_callback (const char *name,
595 int (*fn)(void *arg), void *arg)
597 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
602 liblustre_deregister_wait_callback (void *opaque)
604 liblustre_deregister_waitidle_callback(opaque);
608 liblustre_register_idle_callback (const char *name,
609 int (*fn)(void *arg), void *arg)
611 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
616 liblustre_deregister_idle_callback (void *opaque)
618 liblustre_deregister_waitidle_callback(opaque);
622 liblustre_check_events (int timeout)
629 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
633 LASSERT (rc == -EOVERFLOW || rc == 1);
635 /* liblustre: no asynch callback so we can't affort to miss any
637 if (rc == -EOVERFLOW) {
638 CERROR ("Dropped an event!!!\n");
642 ptlrpc_master_callback (&ev);
646 int liblustre_waiting = 0;
649 liblustre_wait_event (int timeout)
651 struct list_head *tmp;
652 struct liblustre_wait_callback *llwc;
653 int found_something = 0;
655 /* single threaded recursion check... */
656 liblustre_waiting = 1;
659 /* Deal with all pending events */
660 while (liblustre_check_events(0))
663 /* Give all registered callbacks a bite at the cherry */
664 list_for_each(tmp, &liblustre_wait_callbacks) {
665 llwc = list_entry(tmp, struct liblustre_wait_callback,
668 if (llwc->llwc_fn(llwc->llwc_arg))
672 if (found_something || timeout == 0)
675 /* Nothing so far, but I'm allowed to block... */
676 found_something = liblustre_check_events(timeout);
677 if (!found_something) /* still nothing */
678 break; /* I timed out */
681 liblustre_waiting = 0;
683 return found_something;
687 liblustre_wait_idle(void)
689 static int recursed = 0;
691 struct list_head *tmp;
692 struct liblustre_wait_callback *llwc;
699 liblustre_wait_event(0);
703 list_for_each(tmp, &liblustre_idle_callbacks) {
704 llwc = list_entry(tmp, struct liblustre_wait_callback,
707 if (!llwc->llwc_fn(llwc->llwc_arg)) {
718 #endif /* __KERNEL__ */
720 int ptlrpc_init_portals(void)
722 int rc = ptlrpc_ni_init();
725 CERROR("network initialisation failed\n");
729 liblustre_services_callback =
730 liblustre_register_wait_callback("liblustre_check_services",
731 &liblustre_check_services, NULL);
733 rc = ptlrpcd_addref();
737 CERROR("rpcd initialisation failed\n");
739 liblustre_deregister_wait_callback(liblustre_services_callback);
745 void ptlrpc_exit_portals(void)
748 liblustre_deregister_wait_callback(liblustre_services_callback);