4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
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 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(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 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 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;
183 struct ptlrpc_request *req;
186 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
187 ev->type == LNET_EVENT_PUT) ||
188 (desc->bd_type == BULK_GET_SOURCE &&
189 ev->type == LNET_EVENT_GET) ||
190 ev->type == LNET_EVENT_UNLINK);
191 LASSERT (ev->unlinked);
193 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
196 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
199 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
200 "event type %d, status %d, desc %p\n",
201 ev->type, ev->status, desc);
203 spin_lock(&desc->bd_lock);
205 LASSERT(desc->bd_network_rw);
206 desc->bd_network_rw = 0;
208 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
209 desc->bd_success = 1;
210 desc->bd_nob_transferred = ev->mlength;
211 desc->bd_sender = ev->sender;
213 /* start reconnect and resend if network error hit */
214 spin_lock(&req->rq_lock);
216 spin_unlock(&req->rq_lock);
219 /* release the encrypted pages for write */
220 if (desc->bd_req->rq_bulk_write)
221 sptlrpc_enc_pool_put_pages(desc);
223 /* NB don't unlock till after wakeup; desc can disappear under us
225 ptlrpc_client_wake_req(req);
227 spin_unlock(&desc->bd_lock);
232 * We will have percpt request history list for ptlrpc service in upcoming
233 * patches because we don't want to be serialized by current per-service
234 * history operations. So we require history ID can (somehow) show arriving
235 * order w/o grabbing global lock, and user can sort them in userspace.
237 * This is how we generate history ID for ptlrpc_request:
238 * ----------------------------------------------------
239 * | 32 bits | 16 bits | (16 - X)bits | X bits |
240 * ----------------------------------------------------
241 * | seconds | usec / 16 | sequence | CPT id |
242 * ----------------------------------------------------
244 * it might not be precise but should be good enough.
247 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
249 #define REQS_SEC_SHIFT 32
250 #define REQS_USEC_SHIFT 16
251 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
253 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
254 struct ptlrpc_request *req)
256 __u64 sec = req->rq_arrival_time.tv_sec;
257 __u32 usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
260 /* set sequence ID for request and add it to history list,
261 * it must be called with hold svcpt::scp_lock */
263 new_seq = (sec << REQS_SEC_SHIFT) |
264 (usec << REQS_USEC_SHIFT) |
265 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
267 if (new_seq > svcpt->scp_hist_seq) {
268 /* This handles the initial case of scp_hist_seq == 0 or
269 * we just jumped into a new time window */
270 svcpt->scp_hist_seq = new_seq;
272 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
273 /* NB: increase sequence number in current usec bucket,
274 * however, it's possible that we used up all bits for
275 * sequence and jumped into the next usec bucket (future time),
276 * then we hope there will be less RPCs per bucket at some
277 * point, and sequence will catch up again */
278 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
279 new_seq = svcpt->scp_hist_seq;
282 req->rq_history_seq = new_seq;
284 cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
288 * Server's incoming request callback
290 void request_in_callback(lnet_event_t *ev)
292 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
293 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
294 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
295 struct ptlrpc_service *service = svcpt->scp_service;
296 struct ptlrpc_request *req;
299 LASSERT (ev->type == LNET_EVENT_PUT ||
300 ev->type == LNET_EVENT_UNLINK);
301 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
302 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
303 rqbd->rqbd_buffer + service->srv_buf_size);
305 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
306 "event type %d, status %d, service %s\n",
307 ev->type, ev->status, service->srv_name);
310 /* If this is the last request message to fit in the
311 * request buffer we can use the request object embedded in
312 * rqbd. Note that if we failed to allocate a request,
313 * we'd have to re-post the rqbd, which we can't do in this
315 req = &rqbd->rqbd_req;
316 memset(req, 0, sizeof (*req));
318 LASSERT (ev->type == LNET_EVENT_PUT);
319 if (ev->status != 0) {
320 /* We moaned above already... */
323 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
325 CERROR("Can't allocate incoming request descriptor: "
326 "Dropping %s RPC from %s\n",
328 libcfs_id2str(ev->initiator));
333 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
334 * flags are reset and scalars are zero. We only set the message
335 * size to non-zero if this was a successful receive. */
336 req->rq_xid = ev->match_bits;
337 req->rq_reqbuf = ev->md.start + ev->offset;
338 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
339 req->rq_reqdata_len = ev->mlength;
340 cfs_gettimeofday(&req->rq_arrival_time);
341 req->rq_peer = ev->initiator;
342 req->rq_self = ev->target.nid;
344 req->rq_phase = RQ_PHASE_NEW;
345 spin_lock_init(&req->rq_lock);
346 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
347 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
348 cfs_atomic_set(&req->rq_refcount, 1);
349 if (ev->type == LNET_EVENT_PUT)
350 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
351 req, req->rq_xid, ev->mlength);
353 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
355 spin_lock(&svcpt->scp_lock);
357 ptlrpc_req_add_history(svcpt, req);
360 svcpt->scp_nrqbds_posted--;
361 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
362 svcpt->scp_nrqbds_posted);
364 /* Normally, don't complain about 0 buffers posted; LNET won't
365 * drop incoming reqs since we set the portal lazy */
366 if (test_req_buffer_pressure &&
367 ev->type != LNET_EVENT_UNLINK &&
368 svcpt->scp_nrqbds_posted == 0)
369 CWARN("All %s request buffers busy\n",
372 /* req takes over the network's ref on rqbd */
374 /* req takes a ref on rqbd */
375 rqbd->rqbd_refcount++;
378 cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
379 svcpt->scp_nreqs_incoming++;
381 /* NB everything can disappear under us once the request
382 * has been queued and we unlock, so do the wake now... */
383 cfs_waitq_signal(&svcpt->scp_waitq);
385 spin_unlock(&svcpt->scp_lock);
390 * Server's outgoing reply callback
392 void reply_out_callback(lnet_event_t *ev)
394 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
395 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
396 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
399 LASSERT (ev->type == LNET_EVENT_SEND ||
400 ev->type == LNET_EVENT_ACK ||
401 ev->type == LNET_EVENT_UNLINK);
403 if (!rs->rs_difficult) {
404 /* 'Easy' replies have no further processing so I drop the
405 * net's ref on 'rs' */
406 LASSERT (ev->unlinked);
407 ptlrpc_rs_decref(rs);
412 LASSERT (rs->rs_on_net);
415 /* Last network callback. The net's ref on 'rs' stays put
416 * until ptlrpc_handle_rs() is done with it */
417 spin_lock(&svcpt->scp_rep_lock);
418 spin_lock(&rs->rs_lock);
421 if (!rs->rs_no_ack ||
423 rs->rs_export->exp_obd->obd_last_committed)
424 ptlrpc_schedule_difficult_reply(rs);
426 spin_unlock(&rs->rs_lock);
427 spin_unlock(&svcpt->scp_rep_lock);
432 #ifdef HAVE_SERVER_SUPPORT
434 * Server's bulk completion callback
436 void server_bulk_callback (lnet_event_t *ev)
438 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
439 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
442 LASSERT (ev->type == LNET_EVENT_SEND ||
443 ev->type == LNET_EVENT_UNLINK ||
444 (desc->bd_type == BULK_PUT_SOURCE &&
445 ev->type == LNET_EVENT_ACK) ||
446 (desc->bd_type == BULK_GET_SINK &&
447 ev->type == LNET_EVENT_REPLY));
449 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
450 "event type %d, status %d, desc %p\n",
451 ev->type, ev->status, desc);
453 spin_lock(&desc->bd_lock);
455 if ((ev->type == LNET_EVENT_ACK ||
456 ev->type == LNET_EVENT_REPLY) &&
458 /* We heard back from the peer, so even if we get this
459 * before the SENT event (oh yes we can), we know we
460 * read/wrote the peer buffer and how much... */
461 desc->bd_success = 1;
462 desc->bd_nob_transferred = ev->mlength;
463 desc->bd_sender = ev->sender;
467 /* This is the last callback no matter what... */
468 desc->bd_network_rw = 0;
469 cfs_waitq_signal(&desc->bd_waitq);
472 spin_unlock(&desc->bd_lock);
477 static void ptlrpc_master_callback(lnet_event_t *ev)
479 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
480 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
482 /* Honestly, it's best to find out early. */
483 LASSERT (cbid->cbid_arg != LP_POISON);
484 LASSERT (callback == request_out_callback ||
485 callback == reply_in_callback ||
486 callback == client_bulk_callback ||
487 callback == request_in_callback ||
488 callback == reply_out_callback
489 #ifdef HAVE_SERVER_SUPPORT
490 || callback == server_bulk_callback
497 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
498 lnet_process_id_t *peer, lnet_nid_t *self)
501 __u32 best_order = 0;
504 int portals_compatibility;
510 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
512 peer->pid = LUSTRE_SRV_LNET_PID;
514 /* Choose the matching UUID that's closest */
515 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
516 dist = LNetDist(dst_nid, &src_nid, &order);
520 if (dist == 0) { /* local! use loopback LND */
521 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
528 (dist == best_dist && order < best_order)) {
532 if (portals_compatibility > 1) {
533 /* Strong portals compatibility: Zero the nid's
534 * NET, so if I'm reading new config logs, or
535 * getting configured by (new) lconf I can
536 * still talk to old servers. */
537 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
538 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
546 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
550 void ptlrpc_ni_fini(void)
553 struct l_wait_info lwi;
557 /* Wait for the event queue to become idle since there may still be
558 * messages in flight with pending events (i.e. the fire-and-forget
559 * messages == client requests and "non-difficult" server
562 for (retries = 0;; retries++) {
563 rc = LNetEQFree(ptlrpc_eq_h);
574 CWARN("Event queue still busy\n");
577 cfs_waitq_init(&waitq);
578 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
579 l_wait_event(waitq, 0, &lwi);
586 lnet_pid_t ptl_get_pid(void)
593 pid = LUSTRE_SRV_LNET_PID;
598 int ptlrpc_ni_init(void)
604 CDEBUG(D_NET, "My pid is: %x\n", pid);
606 /* We're not passing any limits yet... */
607 rc = LNetNIInit(pid);
609 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
613 /* CAVEAT EMPTOR: how we process portals events is _radically_
614 * different depending on... */
616 /* kernel LNet calls our master callback when there are new event,
617 * because we are guaranteed to get every event via callback,
618 * so we just set EQ size to 0 to avoid overhread of serializing
619 * enqueue/dequeue operations in LNet. */
620 rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
622 /* liblustre calls the master callback when it removes events from the
623 * event queue. The event queue has to be big enough not to drop
625 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
630 CERROR ("Failed to allocate event queue: %d\n", rc);
637 CFS_LIST_HEAD(liblustre_wait_callbacks);
638 CFS_LIST_HEAD(liblustre_idle_callbacks);
639 void *liblustre_services_callback;
642 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
644 int (*fn)(void *arg), void *arg)
646 struct liblustre_wait_callback *llwc;
648 OBD_ALLOC(llwc, sizeof(*llwc));
649 LASSERT (llwc != NULL);
651 llwc->llwc_name = name;
653 llwc->llwc_arg = arg;
654 cfs_list_add_tail(&llwc->llwc_list, callback_list);
660 liblustre_deregister_waitidle_callback (void *opaque)
662 struct liblustre_wait_callback *llwc = opaque;
664 cfs_list_del(&llwc->llwc_list);
665 OBD_FREE(llwc, sizeof(*llwc));
669 liblustre_register_wait_callback (const char *name,
670 int (*fn)(void *arg), void *arg)
672 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
677 liblustre_deregister_wait_callback (void *opaque)
679 liblustre_deregister_waitidle_callback(opaque);
683 liblustre_register_idle_callback (const char *name,
684 int (*fn)(void *arg), void *arg)
686 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
691 liblustre_deregister_idle_callback (void *opaque)
693 liblustre_deregister_waitidle_callback(opaque);
697 liblustre_check_events (int timeout)
704 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
708 LASSERT (rc == -EOVERFLOW || rc == 1);
710 /* liblustre: no asynch callback so we can't afford to miss any
712 if (rc == -EOVERFLOW) {
713 CERROR ("Dropped an event!!!\n");
717 ptlrpc_master_callback (&ev);
721 int liblustre_waiting = 0;
724 liblustre_wait_event (int timeout)
727 struct liblustre_wait_callback *llwc;
728 int found_something = 0;
730 /* single threaded recursion check... */
731 liblustre_waiting = 1;
734 /* Deal with all pending events */
735 while (liblustre_check_events(0))
738 /* Give all registered callbacks a bite at the cherry */
739 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
740 llwc = cfs_list_entry(tmp,
741 struct liblustre_wait_callback,
744 if (llwc->llwc_fn(llwc->llwc_arg))
748 if (found_something || timeout == 0)
751 /* Nothing so far, but I'm allowed to block... */
752 found_something = liblustre_check_events(timeout);
753 if (!found_something) /* still nothing */
754 break; /* I timed out */
757 liblustre_waiting = 0;
759 return found_something;
763 liblustre_wait_idle(void)
765 static int recursed = 0;
768 struct liblustre_wait_callback *llwc;
775 liblustre_wait_event(0);
779 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
780 llwc = cfs_list_entry(tmp,
781 struct liblustre_wait_callback,
784 if (!llwc->llwc_fn(llwc->llwc_arg)) {
795 #endif /* __KERNEL__ */
797 int ptlrpc_init_portals(void)
799 int rc = ptlrpc_ni_init();
802 CERROR("network initialisation failed\n");
806 liblustre_services_callback =
807 liblustre_register_wait_callback("liblustre_check_services",
808 &liblustre_check_services,
810 init_completion_module(liblustre_wait_event);
812 rc = ptlrpcd_addref();
816 CERROR("rpcd initialisation failed\n");
818 liblustre_deregister_wait_callback(liblustre_services_callback);
824 void ptlrpc_exit_portals(void)
827 liblustre_deregister_wait_callback(liblustre_services_callback);