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>
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);
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 spin_lock(&req->rq_lock);
80 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((ev->status == 0) ? D_NET : D_ERROR, req,
100 "type %d, status %d", ev->type, ev->status);
102 LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
103 LASSERT (ev->md.start == req->rq_repbuf);
104 LASSERT (ev->mlength <= req->rq_repbuf_len);
105 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
106 for adaptive timeouts' early reply. */
107 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
109 spin_lock(&req->rq_lock);
111 req->rq_receiving_reply = 0;
114 req->rq_must_unlink = 0;
118 if (ev->type == LNET_EVENT_UNLINK) {
119 LASSERT(ev->unlinked);
120 DEBUG_REQ(D_RPCTRACE, req, "unlink");
124 if ((ev->offset == 0) &&
125 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
127 DEBUG_REQ(D_ADAPTTO, req,
128 "Early reply received: mlen=%u offset=%d replen=%d "
129 "replied=%d unlinked=%d", ev->mlength, ev->offset,
130 req->rq_replen, req->rq_replied, ev->unlinked);
132 req->rq_early_count++; /* number received, client side */
134 if (req->rq_replied) /* already got the real reply */
138 req->rq_reply_off = ev->offset;
139 req->rq_nob_received = ev->mlength;
140 /* And we're still receiving */
141 req->rq_receiving_reply = 1;
144 req->rq_rep_swab_mask = 0;
146 req->rq_reply_off = ev->offset;
147 req->rq_nob_received = ev->mlength;
148 /* LNetMDUnlink can't be called under the LNET_LOCK,
149 so we must unlink in ptlrpc_unregister_reply */
150 DEBUG_REQ(D_INFO, req,
151 "reply in flags=%x mlen=%u offset=%d replen=%d",
152 lustre_msg_get_flags(req->rq_reqmsg),
153 ev->mlength, ev->offset, req->rq_replen);
156 req->rq_import->imp_last_reply_time = cfs_time_current_sec();
159 /* NB don't unlock till after wakeup; req can disappear under us
160 * since we don't have our own ref */
161 ptlrpc_client_wake_req(req);
162 spin_unlock(&req->rq_lock);
167 * Client's bulk has been written/read
169 void client_bulk_callback (lnet_event_t *ev)
171 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
172 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
175 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
176 ev->type == LNET_EVENT_PUT) ||
177 (desc->bd_type == BULK_GET_SOURCE &&
178 ev->type == LNET_EVENT_GET) ||
179 ev->type == LNET_EVENT_UNLINK);
180 LASSERT (ev->unlinked);
182 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
183 "event type %d, status %d, desc %p\n",
184 ev->type, ev->status, desc);
186 spin_lock(&desc->bd_lock);
188 LASSERT(desc->bd_network_rw);
189 desc->bd_network_rw = 0;
191 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
192 desc->bd_success = 1;
193 desc->bd_nob_transferred = ev->mlength;
194 desc->bd_sender = ev->sender;
197 /* release the encrypted pages for write */
198 if (desc->bd_req->rq_bulk_write)
199 sptlrpc_enc_pool_put_pages(desc);
201 /* NB don't unlock till after wakeup; desc can disappear under us
203 ptlrpc_client_wake_req(desc->bd_req);
205 spin_unlock(&desc->bd_lock);
210 * Server's incoming request callback
212 void request_in_callback(lnet_event_t *ev)
214 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
215 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
216 struct ptlrpc_service *service = rqbd->rqbd_service;
217 struct ptlrpc_request *req;
220 LASSERT (ev->type == LNET_EVENT_PUT ||
221 ev->type == LNET_EVENT_UNLINK);
222 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
223 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
224 rqbd->rqbd_buffer + service->srv_buf_size);
226 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
227 "event type %d, status %d, service %s\n",
228 ev->type, ev->status, service->srv_name);
231 /* If this is the last request message to fit in the
232 * request buffer we can use the request object embedded in
233 * rqbd. Note that if we failed to allocate a request,
234 * we'd have to re-post the rqbd, which we can't do in this
236 req = &rqbd->rqbd_req;
237 memset(req, 0, sizeof (*req));
239 LASSERT (ev->type == LNET_EVENT_PUT);
240 if (ev->status != 0) {
241 /* We moaned above already... */
244 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
246 CERROR("Can't allocate incoming request descriptor: "
247 "Dropping %s RPC from %s\n",
249 libcfs_id2str(ev->initiator));
254 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
255 * flags are reset and scalars are zero. We only set the message
256 * size to non-zero if this was a successful receive. */
257 req->rq_xid = ev->match_bits;
258 req->rq_reqbuf = ev->md.start + ev->offset;
259 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
260 req->rq_reqdata_len = ev->mlength;
261 do_gettimeofday(&req->rq_arrival_time);
262 req->rq_peer = ev->initiator;
263 req->rq_self = ev->target.nid;
265 req->rq_phase = RQ_PHASE_NEW;
267 req->rq_uid = ev->uid;
269 spin_lock_init(&req->rq_lock);
270 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
271 atomic_set(&req->rq_refcount, 1);
272 if (ev->type == LNET_EVENT_PUT)
273 DEBUG_REQ(D_RPCTRACE, req, "incoming req");
275 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
277 spin_lock(&service->srv_lock);
279 req->rq_history_seq = service->srv_request_seq++;
280 list_add_tail(&req->rq_history_list, &service->srv_request_history);
283 service->srv_nrqbd_receiving--;
284 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
285 service->srv_nrqbd_receiving);
287 /* Normally, don't complain about 0 buffers posted; LNET won't
288 * drop incoming reqs since we set the portal lazy */
289 if (test_req_buffer_pressure &&
290 ev->type != LNET_EVENT_UNLINK &&
291 service->srv_nrqbd_receiving == 0)
292 CWARN("All %s request buffers busy\n",
295 /* req takes over the network's ref on rqbd */
297 /* req takes a ref on rqbd */
298 rqbd->rqbd_refcount++;
301 list_add_tail(&req->rq_list, &service->srv_req_in_queue);
302 service->srv_n_queued_reqs++;
304 /* NB everything can disappear under us once the request
305 * has been queued and we unlock, so do the wake now... */
306 cfs_waitq_signal(&service->srv_waitq);
308 spin_unlock(&service->srv_lock);
313 * Server's outgoing reply callback
315 void reply_out_callback(lnet_event_t *ev)
317 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
318 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
319 struct ptlrpc_service *svc = rs->rs_service;
322 LASSERT (ev->type == LNET_EVENT_SEND ||
323 ev->type == LNET_EVENT_ACK ||
324 ev->type == LNET_EVENT_UNLINK);
326 if (!rs->rs_difficult) {
327 /* 'Easy' replies have no further processing so I drop the
328 * net's ref on 'rs' */
329 LASSERT (ev->unlinked);
330 ptlrpc_rs_decref(rs);
331 atomic_dec (&svc->srv_outstanding_replies);
336 LASSERT (rs->rs_on_net);
339 /* Last network callback. The net's ref on 'rs' stays put
340 * until ptlrpc_server_handle_reply() is done with it */
341 spin_lock(&svc->srv_lock);
342 spin_lock(&rs->rs_lock);
344 if (!rs->rs_no_ack ||
345 rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
346 ptlrpc_schedule_difficult_reply (rs);
347 spin_unlock(&rs->rs_lock);
348 spin_unlock(&svc->srv_lock);
355 * Server's bulk completion callback
357 void server_bulk_callback (lnet_event_t *ev)
359 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
360 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
363 LASSERT (ev->type == LNET_EVENT_SEND ||
364 ev->type == LNET_EVENT_UNLINK ||
365 (desc->bd_type == BULK_PUT_SOURCE &&
366 ev->type == LNET_EVENT_ACK) ||
367 (desc->bd_type == BULK_GET_SINK &&
368 ev->type == LNET_EVENT_REPLY));
370 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
371 "event type %d, status %d, desc %p\n",
372 ev->type, ev->status, desc);
374 spin_lock(&desc->bd_lock);
376 if ((ev->type == LNET_EVENT_ACK ||
377 ev->type == LNET_EVENT_REPLY) &&
379 /* We heard back from the peer, so even if we get this
380 * before the SENT event (oh yes we can), we know we
381 * read/wrote the peer buffer and how much... */
382 desc->bd_success = 1;
383 desc->bd_nob_transferred = ev->mlength;
384 desc->bd_sender = ev->sender;
388 /* This is the last callback no matter what... */
389 desc->bd_network_rw = 0;
390 cfs_waitq_signal(&desc->bd_waitq);
393 spin_unlock(&desc->bd_lock);
397 static void ptlrpc_master_callback(lnet_event_t *ev)
399 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
400 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
402 /* Honestly, it's best to find out early. */
403 LASSERT (cbid->cbid_arg != LP_POISON);
404 LASSERT (callback == request_out_callback ||
405 callback == reply_in_callback ||
406 callback == client_bulk_callback ||
407 callback == request_in_callback ||
408 callback == reply_out_callback ||
409 callback == server_bulk_callback);
414 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
415 lnet_process_id_t *peer, lnet_nid_t *self)
418 __u32 best_order = 0;
421 int portals_compatibility;
427 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
429 peer->pid = LUSTRE_SRV_LNET_PID;
431 /* Choose the matching UUID that's closest */
432 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
433 dist = LNetDist(dst_nid, &src_nid, &order);
437 if (dist == 0) { /* local! use loopback LND */
438 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
445 (dist == best_dist && order < best_order)) {
449 if (portals_compatibility > 1) {
450 /* Strong portals compatibility: Zero the nid's
451 * NET, so if I'm reading new config logs, or
452 * getting configured by (new) lconf I can
453 * still talk to old servers. */
454 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
455 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
463 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
465 CERROR("No NID found for %s\n", uuid->uuid);
469 void ptlrpc_ni_fini(void)
472 struct l_wait_info lwi;
476 /* Wait for the event queue to become idle since there may still be
477 * messages in flight with pending events (i.e. the fire-and-forget
478 * messages == client requests and "non-difficult" server
481 for (retries = 0;; retries++) {
482 rc = LNetEQFree(ptlrpc_eq_h);
493 CWARN("Event queue still busy\n");
496 cfs_waitq_init(&waitq);
497 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
498 l_wait_event(waitq, 0, &lwi);
505 lnet_pid_t ptl_get_pid(void)
512 pid = LUSTRE_SRV_LNET_PID;
517 int ptlrpc_ni_init(void)
523 CDEBUG(D_NET, "My pid is: %x\n", pid);
525 /* We're not passing any limits yet... */
526 rc = LNetNIInit(pid);
528 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
532 /* CAVEAT EMPTOR: how we process portals events is _radically_
533 * different depending on... */
535 /* kernel portals calls our master callback when events are added to
536 * the event queue. In fact lustre never pulls events off this queue,
537 * so it's only sized for some debug history. */
538 rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
540 /* liblustre calls the master callback when it removes events from the
541 * event queue. The event queue has to be big enough not to drop
543 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
548 CERROR ("Failed to allocate event queue: %d\n", rc);
555 CFS_LIST_HEAD(liblustre_wait_callbacks);
556 CFS_LIST_HEAD(liblustre_idle_callbacks);
557 void *liblustre_services_callback;
560 liblustre_register_waitidle_callback (struct list_head *callback_list,
562 int (*fn)(void *arg), void *arg)
564 struct liblustre_wait_callback *llwc;
566 OBD_ALLOC(llwc, sizeof(*llwc));
567 LASSERT (llwc != NULL);
569 llwc->llwc_name = name;
571 llwc->llwc_arg = arg;
572 list_add_tail(&llwc->llwc_list, callback_list);
578 liblustre_deregister_waitidle_callback (void *opaque)
580 struct liblustre_wait_callback *llwc = opaque;
582 list_del(&llwc->llwc_list);
583 OBD_FREE(llwc, sizeof(*llwc));
587 liblustre_register_wait_callback (const char *name,
588 int (*fn)(void *arg), void *arg)
590 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
595 liblustre_deregister_wait_callback (void *opaque)
597 liblustre_deregister_waitidle_callback(opaque);
601 liblustre_register_idle_callback (const char *name,
602 int (*fn)(void *arg), void *arg)
604 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
609 liblustre_deregister_idle_callback (void *opaque)
611 liblustre_deregister_waitidle_callback(opaque);
615 liblustre_check_events (int timeout)
622 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
626 LASSERT (rc == -EOVERFLOW || rc == 1);
628 /* liblustre: no asynch callback so we can't affort to miss any
630 if (rc == -EOVERFLOW) {
631 CERROR ("Dropped an event!!!\n");
635 ptlrpc_master_callback (&ev);
639 int liblustre_waiting = 0;
642 liblustre_wait_event (int timeout)
644 struct list_head *tmp;
645 struct liblustre_wait_callback *llwc;
646 int found_something = 0;
648 /* single threaded recursion check... */
649 liblustre_waiting = 1;
652 /* Deal with all pending events */
653 while (liblustre_check_events(0))
656 /* Give all registered callbacks a bite at the cherry */
657 list_for_each(tmp, &liblustre_wait_callbacks) {
658 llwc = list_entry(tmp, struct liblustre_wait_callback,
661 if (llwc->llwc_fn(llwc->llwc_arg))
665 if (found_something || timeout == 0)
668 /* Nothing so far, but I'm allowed to block... */
669 found_something = liblustre_check_events(timeout);
670 if (!found_something) /* still nothing */
671 break; /* I timed out */
674 liblustre_waiting = 0;
676 return found_something;
680 liblustre_wait_idle(void)
682 static int recursed = 0;
684 struct list_head *tmp;
685 struct liblustre_wait_callback *llwc;
692 liblustre_wait_event(0);
696 list_for_each(tmp, &liblustre_idle_callbacks) {
697 llwc = list_entry(tmp, struct liblustre_wait_callback,
700 if (!llwc->llwc_fn(llwc->llwc_arg)) {
711 #endif /* __KERNEL__ */
713 int ptlrpc_init_portals(void)
715 int rc = ptlrpc_ni_init();
718 CERROR("network initialisation failed\n");
722 liblustre_services_callback =
723 liblustre_register_wait_callback("liblustre_check_services",
724 &liblustre_check_services,
726 init_completion_module(liblustre_wait_event);
728 rc = ptlrpcd_addref();
732 CERROR("rpcd initialisation failed\n");
734 liblustre_deregister_wait_callback(liblustre_services_callback);
740 void ptlrpc_exit_portals(void)
743 liblustre_deregister_wait_callback(liblustre_services_callback);