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.
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 <lustre_sec.h>
49 #include "ptlrpc_internal.h"
51 lnet_handle_eq_t ptlrpc_eq_h;
54 * Client's outgoing request callback
56 void request_out_callback(lnet_event_t *ev)
58 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
59 struct ptlrpc_request *req = cbid->cbid_arg;
62 LASSERT (ev->type == LNET_EVENT_SEND ||
63 ev->type == LNET_EVENT_UNLINK);
64 LASSERT (ev->unlinked);
66 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
67 "type %d, status %d", ev->type, ev->status);
69 sptlrpc_request_out_callback(req);
71 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
73 /* Failed send: make it seem like the reply timed out, just
74 * like failing sends in client.c does currently... */
76 spin_lock(&req->rq_lock);
78 spin_unlock(&req->rq_lock);
80 ptlrpc_wake_client_req(req);
83 ptlrpc_req_finished(req);
89 * Client's incoming reply callback
91 void reply_in_callback(lnet_event_t *ev)
93 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
94 struct ptlrpc_request *req = cbid->cbid_arg;
97 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
98 "type %d, status %d", ev->type, ev->status);
100 LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
101 LASSERT (ev->md.start == req->rq_repbuf);
102 LASSERT (ev->mlength <= req->rq_repbuf_len);
103 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
104 for adaptive timeouts' early reply. */
105 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
107 spin_lock(&req->rq_lock);
109 req->rq_receiving_reply = 0;
114 if (ev->type == LNET_EVENT_UNLINK) {
115 req->rq_must_unlink = 0;
116 DEBUG_REQ(D_RPCTRACE, req, "unlink");
120 if ((ev->offset == 0) &&
121 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
123 DEBUG_REQ(D_ADAPTTO, req,
124 "Early reply received: mlen=%u offset=%d replen=%d "
125 "replied=%d unlinked=%d", ev->mlength, ev->offset,
126 req->rq_replen, req->rq_replied, ev->unlinked);
128 req->rq_early_count++; /* number received, client side */
129 if (req->rq_replied) {
130 /* If we already got the real reply, then we need to
131 * check if lnet_finalize() unlinked the md. In that
132 * case, there will be no further callback of type
136 req->rq_must_unlink = 0;
138 DEBUG_REQ(D_RPCTRACE, req, "unlinked in reply");
142 req->rq_reply_off = ev->offset;
143 req->rq_nob_received = ev->mlength;
144 /* And we're still receiving */
145 req->rq_receiving_reply = 1;
149 req->rq_reply_off = ev->offset;
150 req->rq_nob_received = ev->mlength;
151 /* LNetMDUnlink can't be called under the LNET_LOCK,
152 so we must unlink in ptlrpc_unregister_reply */
153 DEBUG_REQ(D_INFO, req,
154 "reply in flags=%x mlen=%u offset=%d replen=%d",
155 lustre_msg_get_flags(req->rq_reqmsg),
156 ev->mlength, ev->offset, req->rq_replen);
159 req->rq_import->imp_last_reply_time = cfs_time_current_sec();
162 /* NB don't unlock till after wakeup; req can disappear under us
163 * since we don't have our own ref */
164 ptlrpc_wake_client_req(req);
165 spin_unlock(&req->rq_lock);
170 * Client's bulk has been written/read
172 void client_bulk_callback (lnet_event_t *ev)
174 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
175 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
178 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
179 ev->type == LNET_EVENT_PUT) ||
180 (desc->bd_type == BULK_GET_SOURCE &&
181 ev->type == LNET_EVENT_GET) ||
182 ev->type == LNET_EVENT_UNLINK);
183 LASSERT (ev->unlinked);
185 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
186 "event type %d, status %d, desc %p\n",
187 ev->type, ev->status, desc);
189 spin_lock(&desc->bd_lock);
191 LASSERT(desc->bd_network_rw);
192 desc->bd_network_rw = 0;
194 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
195 desc->bd_success = 1;
196 desc->bd_nob_transferred = ev->mlength;
197 desc->bd_sender = ev->sender;
200 sptlrpc_enc_pool_put_pages(desc);
202 /* NB don't unlock till after wakeup; desc can disappear under us
204 ptlrpc_wake_client_req(desc->bd_req);
206 spin_unlock(&desc->bd_lock);
211 * Server's incoming request callback
213 void request_in_callback(lnet_event_t *ev)
215 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
216 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
217 struct ptlrpc_service *service = rqbd->rqbd_service;
218 struct ptlrpc_request *req;
221 LASSERT (ev->type == LNET_EVENT_PUT ||
222 ev->type == LNET_EVENT_UNLINK);
223 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
224 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
225 rqbd->rqbd_buffer + service->srv_buf_size);
227 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
228 "event type %d, status %d, service %s\n",
229 ev->type, ev->status, service->srv_name);
232 /* If this is the last request message to fit in the
233 * request buffer we can use the request object embedded in
234 * rqbd. Note that if we failed to allocate a request,
235 * we'd have to re-post the rqbd, which we can't do in this
237 req = &rqbd->rqbd_req;
238 memset(req, 0, sizeof (*req));
240 LASSERT (ev->type == LNET_EVENT_PUT);
241 if (ev->status != 0) {
242 /* We moaned above already... */
245 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
247 CERROR("Can't allocate incoming request descriptor: "
248 "Dropping %s RPC from %s\n",
250 libcfs_id2str(ev->initiator));
255 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
256 * flags are reset and scalars are zero. We only set the message
257 * size to non-zero if this was a successful receive. */
258 req->rq_xid = ev->match_bits;
259 req->rq_reqbuf = ev->md.start + ev->offset;
260 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
261 req->rq_reqdata_len = ev->mlength;
262 do_gettimeofday(&req->rq_arrival_time);
263 req->rq_peer = ev->initiator;
264 req->rq_self = ev->target.nid;
266 req->rq_phase = RQ_PHASE_NEW;
268 req->rq_uid = ev->uid;
270 spin_lock_init(&req->rq_lock);
271 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
272 atomic_set(&req->rq_refcount, 1);
273 if (ev->type == LNET_EVENT_PUT)
274 DEBUG_REQ(D_RPCTRACE, req, "incoming req");
276 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
278 spin_lock(&service->srv_lock);
280 req->rq_history_seq = service->srv_request_seq++;
281 list_add_tail(&req->rq_history_list, &service->srv_request_history);
284 service->srv_nrqbd_receiving--;
285 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
286 service->srv_nrqbd_receiving);
288 /* Normally, don't complain about 0 buffers posted; LNET won't
289 * drop incoming reqs since we set the portal lazy */
290 if (test_req_buffer_pressure &&
291 ev->type != LNET_EVENT_UNLINK &&
292 service->srv_nrqbd_receiving == 0)
293 CWARN("All %s request buffers busy\n",
296 /* req takes over the network's ref on rqbd */
298 /* req takes a ref on rqbd */
299 rqbd->rqbd_refcount++;
302 list_add_tail(&req->rq_list, &service->srv_req_in_queue);
303 service->srv_n_queued_reqs++;
305 /* NB everything can disappear under us once the request
306 * has been queued and we unlock, so do the wake now... */
307 cfs_waitq_signal(&service->srv_waitq);
309 spin_unlock(&service->srv_lock);
314 * Server's outgoing reply callback
316 void reply_out_callback(lnet_event_t *ev)
318 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
319 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
320 struct ptlrpc_service *svc = rs->rs_service;
323 LASSERT (ev->type == LNET_EVENT_SEND ||
324 ev->type == LNET_EVENT_ACK ||
325 ev->type == LNET_EVENT_UNLINK);
327 if (!rs->rs_difficult) {
328 /* 'Easy' replies have no further processing so I drop the
329 * net's ref on 'rs' */
330 LASSERT (ev->unlinked);
331 ptlrpc_rs_decref(rs);
332 atomic_dec (&svc->srv_outstanding_replies);
337 LASSERT (rs->rs_on_net);
340 /* Last network callback. The net's ref on 'rs' stays put
341 * until ptlrpc_server_handle_reply() is done with it */
342 spin_lock(&svc->srv_lock);
344 ptlrpc_schedule_difficult_reply (rs);
345 spin_unlock(&svc->srv_lock);
352 * Server's bulk completion callback
354 void server_bulk_callback (lnet_event_t *ev)
356 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
357 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
360 LASSERT (ev->type == LNET_EVENT_SEND ||
361 ev->type == LNET_EVENT_UNLINK ||
362 (desc->bd_type == BULK_PUT_SOURCE &&
363 ev->type == LNET_EVENT_ACK) ||
364 (desc->bd_type == BULK_GET_SINK &&
365 ev->type == LNET_EVENT_REPLY));
367 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
368 "event type %d, status %d, desc %p\n",
369 ev->type, ev->status, desc);
371 spin_lock(&desc->bd_lock);
373 if ((ev->type == LNET_EVENT_ACK ||
374 ev->type == LNET_EVENT_REPLY) &&
376 /* We heard back from the peer, so even if we get this
377 * before the SENT event (oh yes we can), we know we
378 * read/wrote the peer buffer and how much... */
379 desc->bd_success = 1;
380 desc->bd_nob_transferred = ev->mlength;
381 desc->bd_sender = ev->sender;
385 /* This is the last callback no matter what... */
386 desc->bd_network_rw = 0;
387 cfs_waitq_signal(&desc->bd_waitq);
390 spin_unlock(&desc->bd_lock);
394 static void ptlrpc_master_callback(lnet_event_t *ev)
396 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
397 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
399 /* Honestly, it's best to find out early. */
400 LASSERT (cbid->cbid_arg != LP_POISON);
401 LASSERT (callback == request_out_callback ||
402 callback == reply_in_callback ||
403 callback == client_bulk_callback ||
404 callback == request_in_callback ||
405 callback == reply_out_callback ||
406 callback == server_bulk_callback);
411 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
412 lnet_process_id_t *peer, lnet_nid_t *self)
415 __u32 best_order = 0;
418 int portals_compatibility;
424 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
426 peer->pid = LUSTRE_SRV_LNET_PID;
428 /* Choose the matching UUID that's closest */
429 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
430 dist = LNetDist(dst_nid, &src_nid, &order);
434 if (dist == 0) { /* local! use loopback LND */
435 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
442 (dist == best_dist && order < best_order)) {
446 if (portals_compatibility > 1) {
447 /* Strong portals compatibility: Zero the nid's
448 * NET, so if I'm reading new config logs, or
449 * getting configured by (new) lconf I can
450 * still talk to old servers. */
451 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
452 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
460 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
462 CERROR("No NID found for %s\n", uuid->uuid);
466 void ptlrpc_ni_fini(void)
469 struct l_wait_info lwi;
473 /* Wait for the event queue to become idle since there may still be
474 * messages in flight with pending events (i.e. the fire-and-forget
475 * messages == client requests and "non-difficult" server
478 for (retries = 0;; retries++) {
479 rc = LNetEQFree(ptlrpc_eq_h);
490 CWARN("Event queue still busy\n");
493 cfs_waitq_init(&waitq);
494 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
495 l_wait_event(waitq, 0, &lwi);
502 lnet_pid_t ptl_get_pid(void)
509 pid = LUSTRE_SRV_LNET_PID;
514 int ptlrpc_ni_init(void)
520 CDEBUG(D_NET, "My pid is: %x\n", pid);
522 /* We're not passing any limits yet... */
523 rc = LNetNIInit(pid);
525 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
529 /* CAVEAT EMPTOR: how we process portals events is _radically_
530 * different depending on... */
532 /* kernel portals calls our master callback when events are added to
533 * the event queue. In fact lustre never pulls events off this queue,
534 * so it's only sized for some debug history. */
535 rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
537 /* liblustre calls the master callback when it removes events from the
538 * event queue. The event queue has to be big enough not to drop
540 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
545 CERROR ("Failed to allocate event queue: %d\n", rc);
552 CFS_LIST_HEAD(liblustre_wait_callbacks);
553 CFS_LIST_HEAD(liblustre_idle_callbacks);
554 void *liblustre_services_callback;
557 liblustre_register_waitidle_callback (struct list_head *callback_list,
559 int (*fn)(void *arg), void *arg)
561 struct liblustre_wait_callback *llwc;
563 OBD_ALLOC(llwc, sizeof(*llwc));
564 LASSERT (llwc != NULL);
566 llwc->llwc_name = name;
568 llwc->llwc_arg = arg;
569 list_add_tail(&llwc->llwc_list, callback_list);
575 liblustre_deregister_waitidle_callback (void *opaque)
577 struct liblustre_wait_callback *llwc = opaque;
579 list_del(&llwc->llwc_list);
580 OBD_FREE(llwc, sizeof(*llwc));
584 liblustre_register_wait_callback (const char *name,
585 int (*fn)(void *arg), void *arg)
587 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
592 liblustre_deregister_wait_callback (void *opaque)
594 liblustre_deregister_waitidle_callback(opaque);
598 liblustre_register_idle_callback (const char *name,
599 int (*fn)(void *arg), void *arg)
601 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
606 liblustre_deregister_idle_callback (void *opaque)
608 liblustre_deregister_waitidle_callback(opaque);
612 liblustre_check_events (int timeout)
619 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
623 LASSERT (rc == -EOVERFLOW || rc == 1);
625 /* liblustre: no asynch callback so we can't affort to miss any
627 if (rc == -EOVERFLOW) {
628 CERROR ("Dropped an event!!!\n");
632 ptlrpc_master_callback (&ev);
636 int liblustre_waiting = 0;
639 liblustre_wait_event (int timeout)
641 struct list_head *tmp;
642 struct liblustre_wait_callback *llwc;
643 int found_something = 0;
645 /* single threaded recursion check... */
646 liblustre_waiting = 1;
649 /* Deal with all pending events */
650 while (liblustre_check_events(0))
653 /* Give all registered callbacks a bite at the cherry */
654 list_for_each(tmp, &liblustre_wait_callbacks) {
655 llwc = list_entry(tmp, struct liblustre_wait_callback,
658 if (llwc->llwc_fn(llwc->llwc_arg))
662 if (found_something || timeout == 0)
665 /* Nothing so far, but I'm allowed to block... */
666 found_something = liblustre_check_events(timeout);
667 if (!found_something) /* still nothing */
668 break; /* I timed out */
671 liblustre_waiting = 0;
673 return found_something;
677 liblustre_wait_idle(void)
679 static int recursed = 0;
681 struct list_head *tmp;
682 struct liblustre_wait_callback *llwc;
689 liblustre_wait_event(0);
693 list_for_each(tmp, &liblustre_idle_callbacks) {
694 llwc = list_entry(tmp, struct liblustre_wait_callback,
697 if (!llwc->llwc_fn(llwc->llwc_arg)) {
708 #endif /* __KERNEL__ */
710 int ptlrpc_init_portals(void)
712 int rc = ptlrpc_ni_init();
715 CERROR("network initialisation failed\n");
719 liblustre_services_callback =
720 liblustre_register_wait_callback("liblustre_check_services",
721 &liblustre_check_services,
724 rc = ptlrpcd_addref();
728 CERROR("rpcd initialisation failed\n");
730 liblustre_deregister_wait_callback(liblustre_services_callback);
736 void ptlrpc_exit_portals(void)
739 liblustre_deregister_wait_callback(liblustre_services_callback);