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;
116 if (ev->type == LNET_EVENT_UNLINK) {
117 req->rq_must_unlink = 0;
118 DEBUG_REQ(D_RPCTRACE, req, "unlink");
122 if ((ev->offset == 0) &&
123 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
125 DEBUG_REQ(D_ADAPTTO, req,
126 "Early reply received: mlen=%u offset=%d replen=%d "
127 "replied=%d unlinked=%d", ev->mlength, ev->offset,
128 req->rq_replen, req->rq_replied, ev->unlinked);
130 req->rq_early_count++; /* number received, client side */
131 if (req->rq_replied) {
132 /* If we already got the real reply, then we need to
133 * check if lnet_finalize() unlinked the md. In that
134 * case, there will be no further callback of type
138 req->rq_must_unlink = 0;
140 DEBUG_REQ(D_RPCTRACE, req, "unlinked in reply");
144 req->rq_reply_off = ev->offset;
145 req->rq_nob_received = ev->mlength;
146 /* And we're still receiving */
147 req->rq_receiving_reply = 1;
151 req->rq_reply_off = ev->offset;
152 req->rq_nob_received = ev->mlength;
153 /* LNetMDUnlink can't be called under the LNET_LOCK,
154 so we must unlink in ptlrpc_unregister_reply */
155 DEBUG_REQ(D_INFO, req,
156 "reply in flags=%x mlen=%u offset=%d replen=%d",
157 lustre_msg_get_flags(req->rq_reqmsg),
158 ev->mlength, ev->offset, req->rq_replen);
161 req->rq_import->imp_last_reply_time = cfs_time_current_sec();
164 /* NB don't unlock till after wakeup; req can disappear under us
165 * since we don't have our own ref */
166 ptlrpc_client_wake_req(req);
167 spin_unlock(&req->rq_lock);
172 * Client's bulk has been written/read
174 void client_bulk_callback (lnet_event_t *ev)
176 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
177 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
180 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
181 ev->type == LNET_EVENT_PUT) ||
182 (desc->bd_type == BULK_GET_SOURCE &&
183 ev->type == LNET_EVENT_GET) ||
184 ev->type == LNET_EVENT_UNLINK);
185 LASSERT (ev->unlinked);
187 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
188 "event type %d, status %d, desc %p\n",
189 ev->type, ev->status, desc);
191 spin_lock(&desc->bd_lock);
193 LASSERT(desc->bd_network_rw);
194 desc->bd_network_rw = 0;
196 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
197 desc->bd_success = 1;
198 desc->bd_nob_transferred = ev->mlength;
199 desc->bd_sender = ev->sender;
202 sptlrpc_enc_pool_put_pages(desc);
204 /* NB don't unlock till after wakeup; desc can disappear under us
206 ptlrpc_client_wake_req(desc->bd_req);
208 spin_unlock(&desc->bd_lock);
213 * Server's incoming request callback
215 void request_in_callback(lnet_event_t *ev)
217 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
218 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
219 struct ptlrpc_service *service = rqbd->rqbd_service;
220 struct ptlrpc_request *req;
223 LASSERT (ev->type == LNET_EVENT_PUT ||
224 ev->type == LNET_EVENT_UNLINK);
225 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
226 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
227 rqbd->rqbd_buffer + service->srv_buf_size);
229 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
230 "event type %d, status %d, service %s\n",
231 ev->type, ev->status, service->srv_name);
234 /* If this is the last request message to fit in the
235 * request buffer we can use the request object embedded in
236 * rqbd. Note that if we failed to allocate a request,
237 * we'd have to re-post the rqbd, which we can't do in this
239 req = &rqbd->rqbd_req;
240 memset(req, 0, sizeof (*req));
242 LASSERT (ev->type == LNET_EVENT_PUT);
243 if (ev->status != 0) {
244 /* We moaned above already... */
247 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
249 CERROR("Can't allocate incoming request descriptor: "
250 "Dropping %s RPC from %s\n",
252 libcfs_id2str(ev->initiator));
257 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
258 * flags are reset and scalars are zero. We only set the message
259 * size to non-zero if this was a successful receive. */
260 req->rq_xid = ev->match_bits;
261 req->rq_reqbuf = ev->md.start + ev->offset;
262 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
263 req->rq_reqdata_len = ev->mlength;
264 do_gettimeofday(&req->rq_arrival_time);
265 req->rq_peer = ev->initiator;
266 req->rq_self = ev->target.nid;
268 req->rq_phase = RQ_PHASE_NEW;
270 req->rq_uid = ev->uid;
272 spin_lock_init(&req->rq_lock);
273 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
274 atomic_set(&req->rq_refcount, 1);
275 if (ev->type == LNET_EVENT_PUT)
276 DEBUG_REQ(D_RPCTRACE, req, "incoming req");
278 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
280 spin_lock(&service->srv_lock);
282 req->rq_history_seq = service->srv_request_seq++;
283 list_add_tail(&req->rq_history_list, &service->srv_request_history);
286 service->srv_nrqbd_receiving--;
287 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
288 service->srv_nrqbd_receiving);
290 /* Normally, don't complain about 0 buffers posted; LNET won't
291 * drop incoming reqs since we set the portal lazy */
292 if (test_req_buffer_pressure &&
293 ev->type != LNET_EVENT_UNLINK &&
294 service->srv_nrqbd_receiving == 0)
295 CWARN("All %s request buffers busy\n",
298 /* req takes over the network's ref on rqbd */
300 /* req takes a ref on rqbd */
301 rqbd->rqbd_refcount++;
304 list_add_tail(&req->rq_list, &service->srv_req_in_queue);
305 service->srv_n_queued_reqs++;
307 /* NB everything can disappear under us once the request
308 * has been queued and we unlock, so do the wake now... */
309 cfs_waitq_signal(&service->srv_waitq);
311 spin_unlock(&service->srv_lock);
316 * Server's outgoing reply callback
318 void reply_out_callback(lnet_event_t *ev)
320 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
321 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
322 struct ptlrpc_service *svc = rs->rs_service;
325 LASSERT (ev->type == LNET_EVENT_SEND ||
326 ev->type == LNET_EVENT_ACK ||
327 ev->type == LNET_EVENT_UNLINK);
329 if (!rs->rs_difficult) {
330 /* 'Easy' replies have no further processing so I drop the
331 * net's ref on 'rs' */
332 LASSERT (ev->unlinked);
333 ptlrpc_rs_decref(rs);
334 atomic_dec (&svc->srv_outstanding_replies);
339 LASSERT (rs->rs_on_net);
342 /* Last network callback. The net's ref on 'rs' stays put
343 * until ptlrpc_server_handle_reply() is done with it */
344 spin_lock(&svc->srv_lock);
345 spin_lock(&rs->rs_lock);
347 if (!rs->rs_no_ack ||
348 rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
349 ptlrpc_schedule_difficult_reply (rs);
350 spin_unlock(&rs->rs_lock);
351 spin_unlock(&svc->srv_lock);
358 * Server's bulk completion callback
360 void server_bulk_callback (lnet_event_t *ev)
362 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
363 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
366 LASSERT (ev->type == LNET_EVENT_SEND ||
367 ev->type == LNET_EVENT_UNLINK ||
368 (desc->bd_type == BULK_PUT_SOURCE &&
369 ev->type == LNET_EVENT_ACK) ||
370 (desc->bd_type == BULK_GET_SINK &&
371 ev->type == LNET_EVENT_REPLY));
373 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
374 "event type %d, status %d, desc %p\n",
375 ev->type, ev->status, desc);
377 spin_lock(&desc->bd_lock);
379 if ((ev->type == LNET_EVENT_ACK ||
380 ev->type == LNET_EVENT_REPLY) &&
382 /* We heard back from the peer, so even if we get this
383 * before the SENT event (oh yes we can), we know we
384 * read/wrote the peer buffer and how much... */
385 desc->bd_success = 1;
386 desc->bd_nob_transferred = ev->mlength;
387 desc->bd_sender = ev->sender;
391 /* This is the last callback no matter what... */
392 desc->bd_network_rw = 0;
393 cfs_waitq_signal(&desc->bd_waitq);
396 spin_unlock(&desc->bd_lock);
400 static void ptlrpc_master_callback(lnet_event_t *ev)
402 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
403 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
405 /* Honestly, it's best to find out early. */
406 LASSERT (cbid->cbid_arg != LP_POISON);
407 LASSERT (callback == request_out_callback ||
408 callback == reply_in_callback ||
409 callback == client_bulk_callback ||
410 callback == request_in_callback ||
411 callback == reply_out_callback ||
412 callback == server_bulk_callback);
417 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
418 lnet_process_id_t *peer, lnet_nid_t *self)
421 __u32 best_order = 0;
424 int portals_compatibility;
430 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
432 peer->pid = LUSTRE_SRV_LNET_PID;
434 /* Choose the matching UUID that's closest */
435 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
436 dist = LNetDist(dst_nid, &src_nid, &order);
440 if (dist == 0) { /* local! use loopback LND */
441 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
448 (dist == best_dist && order < best_order)) {
452 if (portals_compatibility > 1) {
453 /* Strong portals compatibility: Zero the nid's
454 * NET, so if I'm reading new config logs, or
455 * getting configured by (new) lconf I can
456 * still talk to old servers. */
457 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
458 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
466 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
468 CERROR("No NID found for %s\n", uuid->uuid);
472 void ptlrpc_ni_fini(void)
475 struct l_wait_info lwi;
479 /* Wait for the event queue to become idle since there may still be
480 * messages in flight with pending events (i.e. the fire-and-forget
481 * messages == client requests and "non-difficult" server
484 for (retries = 0;; retries++) {
485 rc = LNetEQFree(ptlrpc_eq_h);
496 CWARN("Event queue still busy\n");
499 cfs_waitq_init(&waitq);
500 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
501 l_wait_event(waitq, 0, &lwi);
508 lnet_pid_t ptl_get_pid(void)
515 pid = LUSTRE_SRV_LNET_PID;
520 int ptlrpc_ni_init(void)
526 CDEBUG(D_NET, "My pid is: %x\n", pid);
528 /* We're not passing any limits yet... */
529 rc = LNetNIInit(pid);
531 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
535 /* CAVEAT EMPTOR: how we process portals events is _radically_
536 * different depending on... */
538 /* kernel portals calls our master callback when events are added to
539 * the event queue. In fact lustre never pulls events off this queue,
540 * so it's only sized for some debug history. */
541 rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
543 /* liblustre calls the master callback when it removes events from the
544 * event queue. The event queue has to be big enough not to drop
546 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
551 CERROR ("Failed to allocate event queue: %d\n", rc);
558 CFS_LIST_HEAD(liblustre_wait_callbacks);
559 CFS_LIST_HEAD(liblustre_idle_callbacks);
560 void *liblustre_services_callback;
563 liblustre_register_waitidle_callback (struct list_head *callback_list,
565 int (*fn)(void *arg), void *arg)
567 struct liblustre_wait_callback *llwc;
569 OBD_ALLOC(llwc, sizeof(*llwc));
570 LASSERT (llwc != NULL);
572 llwc->llwc_name = name;
574 llwc->llwc_arg = arg;
575 list_add_tail(&llwc->llwc_list, callback_list);
581 liblustre_deregister_waitidle_callback (void *opaque)
583 struct liblustre_wait_callback *llwc = opaque;
585 list_del(&llwc->llwc_list);
586 OBD_FREE(llwc, sizeof(*llwc));
590 liblustre_register_wait_callback (const char *name,
591 int (*fn)(void *arg), void *arg)
593 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
598 liblustre_deregister_wait_callback (void *opaque)
600 liblustre_deregister_waitidle_callback(opaque);
604 liblustre_register_idle_callback (const char *name,
605 int (*fn)(void *arg), void *arg)
607 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
612 liblustre_deregister_idle_callback (void *opaque)
614 liblustre_deregister_waitidle_callback(opaque);
618 liblustre_check_events (int timeout)
625 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
629 LASSERT (rc == -EOVERFLOW || rc == 1);
631 /* liblustre: no asynch callback so we can't affort to miss any
633 if (rc == -EOVERFLOW) {
634 CERROR ("Dropped an event!!!\n");
638 ptlrpc_master_callback (&ev);
642 int liblustre_waiting = 0;
645 liblustre_wait_event (int timeout)
647 struct list_head *tmp;
648 struct liblustre_wait_callback *llwc;
649 int found_something = 0;
651 /* single threaded recursion check... */
652 liblustre_waiting = 1;
655 /* Deal with all pending events */
656 while (liblustre_check_events(0))
659 /* Give all registered callbacks a bite at the cherry */
660 list_for_each(tmp, &liblustre_wait_callbacks) {
661 llwc = list_entry(tmp, struct liblustre_wait_callback,
664 if (llwc->llwc_fn(llwc->llwc_arg))
668 if (found_something || timeout == 0)
671 /* Nothing so far, but I'm allowed to block... */
672 found_something = liblustre_check_events(timeout);
673 if (!found_something) /* still nothing */
674 break; /* I timed out */
677 liblustre_waiting = 0;
679 return found_something;
683 liblustre_wait_idle(void)
685 static int recursed = 0;
687 struct list_head *tmp;
688 struct liblustre_wait_callback *llwc;
695 liblustre_wait_event(0);
699 list_for_each(tmp, &liblustre_idle_callbacks) {
700 llwc = list_entry(tmp, struct liblustre_wait_callback,
703 if (!llwc->llwc_fn(llwc->llwc_arg)) {
714 #endif /* __KERNEL__ */
716 int ptlrpc_init_portals(void)
718 int rc = ptlrpc_ni_init();
721 CERROR("network initialisation failed\n");
725 liblustre_services_callback =
726 liblustre_register_wait_callback("liblustre_check_services",
727 &liblustre_check_services,
729 init_completion_module(liblustre_wait_event);
731 rc = ptlrpcd_addref();
735 CERROR("rpcd initialisation failed\n");
737 liblustre_deregister_wait_callback(liblustre_services_callback);
743 void ptlrpc_exit_portals(void)
746 liblustre_deregister_wait_callback(liblustre_services_callback);