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, 2013, 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 spin_lock(&req->rq_lock);
72 req->rq_real_sent = cfs_time_current_sec();
74 req->rq_req_unlink = 0;
76 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
78 /* Failed send: make it seem like the reply timed out, just
79 * like failing sends in client.c does currently... */
82 ptlrpc_client_wake_req(req);
84 spin_unlock(&req->rq_lock);
86 ptlrpc_req_finished(req);
92 * Client's incoming reply callback
94 void reply_in_callback(lnet_event_t *ev)
96 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
97 struct ptlrpc_request *req = cbid->cbid_arg;
100 DEBUG_REQ(D_NET, req, "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->offset + 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_reply_unlink = 0;
119 if (ev->type == LNET_EVENT_UNLINK) {
120 LASSERT(ev->unlinked);
121 DEBUG_REQ(D_NET, req, "unlink");
125 if (ev->mlength < ev->rlength ) {
126 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
127 req->rq_replen, ev->rlength, ev->offset);
128 req->rq_reply_truncate = 1;
130 req->rq_status = -EOVERFLOW;
131 req->rq_nob_received = ev->rlength + ev->offset;
135 if ((ev->offset == 0) &&
136 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
138 DEBUG_REQ(D_ADAPTTO, req,
139 "Early reply received: mlen=%u offset=%d replen=%d "
140 "replied=%d unlinked=%d", ev->mlength, ev->offset,
141 req->rq_replen, req->rq_replied, ev->unlinked);
143 req->rq_early_count++; /* number received, client side */
145 if (req->rq_replied) /* already got the real reply */
149 req->rq_reply_off = ev->offset;
150 req->rq_nob_received = ev->mlength;
151 /* And we're still receiving */
152 req->rq_receiving_reply = 1;
155 req->rq_rep_swab_mask = 0;
157 /* Got reply, no resend required */
159 req->rq_reply_off = ev->offset;
160 req->rq_nob_received = ev->mlength;
161 /* LNetMDUnlink can't be called under the LNET_LOCK,
162 so we must unlink in ptlrpc_unregister_reply */
163 DEBUG_REQ(D_INFO, req,
164 "reply in flags=%x mlen=%u offset=%d replen=%d",
165 lustre_msg_get_flags(req->rq_reqmsg),
166 ev->mlength, ev->offset, req->rq_replen);
169 req->rq_import->imp_last_reply_time = cfs_time_current_sec();
172 /* NB don't unlock till after wakeup; req can disappear under us
173 * since we don't have our own ref */
174 ptlrpc_client_wake_req(req);
175 spin_unlock(&req->rq_lock);
180 * Client's bulk has been written/read
182 void client_bulk_callback (lnet_event_t *ev)
184 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
185 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
186 struct ptlrpc_request *req;
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 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
199 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
202 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
203 "event type %d, status %d, desc %p\n",
204 ev->type, ev->status, desc);
206 spin_lock(&desc->bd_lock);
208 LASSERT(desc->bd_md_count > 0);
211 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
212 desc->bd_nob_transferred += ev->mlength;
213 desc->bd_sender = ev->sender;
215 /* start reconnect and resend if network error hit */
216 spin_lock(&req->rq_lock);
218 spin_unlock(&req->rq_lock);
222 desc->bd_failure = 1;
224 /* NB don't unlock till after wakeup; desc can disappear under us
226 if (desc->bd_md_count == 0)
227 ptlrpc_client_wake_req(desc->bd_req);
229 spin_unlock(&desc->bd_lock);
234 * We will have percpt request history list for ptlrpc service in upcoming
235 * patches because we don't want to be serialized by current per-service
236 * history operations. So we require history ID can (somehow) show arriving
237 * order w/o grabbing global lock, and user can sort them in userspace.
239 * This is how we generate history ID for ptlrpc_request:
240 * ----------------------------------------------------
241 * | 32 bits | 16 bits | (16 - X)bits | X bits |
242 * ----------------------------------------------------
243 * | seconds | usec / 16 | sequence | CPT id |
244 * ----------------------------------------------------
246 * it might not be precise but should be good enough.
249 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
251 #define REQS_SEC_SHIFT 32
252 #define REQS_USEC_SHIFT 16
253 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
255 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
256 struct ptlrpc_request *req)
258 __u64 sec = req->rq_arrival_time.tv_sec;
259 __u32 usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
262 /* set sequence ID for request and add it to history list,
263 * it must be called with hold svcpt::scp_lock */
265 new_seq = (sec << REQS_SEC_SHIFT) |
266 (usec << REQS_USEC_SHIFT) |
267 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
269 if (new_seq > svcpt->scp_hist_seq) {
270 /* This handles the initial case of scp_hist_seq == 0 or
271 * we just jumped into a new time window */
272 svcpt->scp_hist_seq = new_seq;
274 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
275 /* NB: increase sequence number in current usec bucket,
276 * however, it's possible that we used up all bits for
277 * sequence and jumped into the next usec bucket (future time),
278 * then we hope there will be less RPCs per bucket at some
279 * point, and sequence will catch up again */
280 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
281 new_seq = svcpt->scp_hist_seq;
284 req->rq_history_seq = new_seq;
286 cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
290 * Server's incoming request callback
292 void request_in_callback(lnet_event_t *ev)
294 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
295 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
296 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
297 struct ptlrpc_service *service = svcpt->scp_service;
298 struct ptlrpc_request *req;
301 LASSERT (ev->type == LNET_EVENT_PUT ||
302 ev->type == LNET_EVENT_UNLINK);
303 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
304 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
305 rqbd->rqbd_buffer + service->srv_buf_size);
307 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
308 "event type %d, status %d, service %s\n",
309 ev->type, ev->status, service->srv_name);
312 /* If this is the last request message to fit in the
313 * request buffer we can use the request object embedded in
314 * rqbd. Note that if we failed to allocate a request,
315 * we'd have to re-post the rqbd, which we can't do in this
317 req = &rqbd->rqbd_req;
318 memset(req, 0, sizeof (*req));
320 LASSERT (ev->type == LNET_EVENT_PUT);
321 if (ev->status != 0) {
322 /* We moaned above already... */
325 req = ptlrpc_request_cache_alloc(ALLOC_ATOMIC_TRY);
327 CERROR("Can't allocate incoming request descriptor: "
328 "Dropping %s RPC from %s\n",
330 libcfs_id2str(ev->initiator));
335 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
336 * flags are reset and scalars are zero. We only set the message
337 * size to non-zero if this was a successful receive. */
338 req->rq_xid = ev->match_bits;
339 req->rq_reqbuf = ev->md.start + ev->offset;
340 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
341 req->rq_reqdata_len = ev->mlength;
342 do_gettimeofday(&req->rq_arrival_time);
343 req->rq_peer = ev->initiator;
344 req->rq_self = ev->target.nid;
346 req->rq_phase = RQ_PHASE_NEW;
347 spin_lock_init(&req->rq_lock);
348 CFS_INIT_LIST_HEAD(&req->rq_timed_list);
349 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
350 atomic_set(&req->rq_refcount, 1);
351 if (ev->type == LNET_EVENT_PUT)
352 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
353 req, req->rq_xid, ev->mlength);
355 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
357 spin_lock(&svcpt->scp_lock);
359 ptlrpc_req_add_history(svcpt, req);
362 svcpt->scp_nrqbds_posted--;
363 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
364 svcpt->scp_nrqbds_posted);
366 /* Normally, don't complain about 0 buffers posted; LNET won't
367 * drop incoming reqs since we set the portal lazy */
368 if (test_req_buffer_pressure &&
369 ev->type != LNET_EVENT_UNLINK &&
370 svcpt->scp_nrqbds_posted == 0)
371 CWARN("All %s request buffers busy\n",
374 /* req takes over the network's ref on rqbd */
376 /* req takes a ref on rqbd */
377 rqbd->rqbd_refcount++;
380 cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
381 svcpt->scp_nreqs_incoming++;
383 /* NB everything can disappear under us once the request
384 * has been queued and we unlock, so do the wake now... */
385 wake_up(&svcpt->scp_waitq);
387 spin_unlock(&svcpt->scp_lock);
392 * Server's outgoing reply callback
394 void reply_out_callback(lnet_event_t *ev)
396 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
397 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
398 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
401 LASSERT (ev->type == LNET_EVENT_SEND ||
402 ev->type == LNET_EVENT_ACK ||
403 ev->type == LNET_EVENT_UNLINK);
405 if (!rs->rs_difficult) {
406 /* 'Easy' replies have no further processing so I drop the
407 * net's ref on 'rs' */
408 LASSERT (ev->unlinked);
409 ptlrpc_rs_decref(rs);
414 LASSERT (rs->rs_on_net);
417 /* Last network callback. The net's ref on 'rs' stays put
418 * until ptlrpc_handle_rs() is done with it */
419 spin_lock(&svcpt->scp_rep_lock);
420 spin_lock(&rs->rs_lock);
423 if (!rs->rs_no_ack ||
425 rs->rs_export->exp_obd->obd_last_committed)
426 ptlrpc_schedule_difficult_reply(rs);
428 spin_unlock(&rs->rs_lock);
429 spin_unlock(&svcpt->scp_rep_lock);
434 #ifdef HAVE_SERVER_SUPPORT
436 * Server's bulk completion callback
438 void server_bulk_callback (lnet_event_t *ev)
440 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
441 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
444 LASSERT(ev->type == LNET_EVENT_SEND ||
445 ev->type == LNET_EVENT_UNLINK ||
446 (desc->bd_type == BULK_PUT_SOURCE &&
447 ev->type == LNET_EVENT_ACK) ||
448 (desc->bd_type == BULK_GET_SINK &&
449 ev->type == LNET_EVENT_REPLY));
451 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
452 "event type %d, status %d, desc %p\n",
453 ev->type, ev->status, desc);
455 spin_lock(&desc->bd_lock);
457 LASSERT(desc->bd_md_count > 0);
459 if ((ev->type == LNET_EVENT_ACK ||
460 ev->type == LNET_EVENT_REPLY) &&
462 /* We heard back from the peer, so even if we get this
463 * before the SENT event (oh yes we can), we know we
464 * read/wrote the peer buffer and how much... */
465 desc->bd_nob_transferred += ev->mlength;
466 desc->bd_sender = ev->sender;
470 desc->bd_failure = 1;
474 /* This is the last callback no matter what... */
475 if (desc->bd_md_count == 0)
476 wake_up(&desc->bd_waitq);
479 spin_unlock(&desc->bd_lock);
484 static void ptlrpc_master_callback(lnet_event_t *ev)
486 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
487 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
489 /* Honestly, it's best to find out early. */
490 LASSERT (cbid->cbid_arg != LP_POISON);
491 LASSERT (callback == request_out_callback ||
492 callback == reply_in_callback ||
493 callback == client_bulk_callback ||
494 callback == request_in_callback ||
495 callback == reply_out_callback
496 #ifdef HAVE_SERVER_SUPPORT
497 || callback == server_bulk_callback
504 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
505 lnet_process_id_t *peer, lnet_nid_t *self)
508 __u32 best_order = 0;
511 int portals_compatibility;
517 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
519 peer->pid = LUSTRE_SRV_LNET_PID;
521 /* Choose the matching UUID that's closest */
522 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
523 dist = LNetDist(dst_nid, &src_nid, &order);
527 if (dist == 0) { /* local! use loopback LND */
528 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
535 (dist == best_dist && order < best_order)) {
539 if (portals_compatibility > 1) {
540 /* Strong portals compatibility: Zero the nid's
541 * NET, so if I'm reading new config logs, or
542 * getting configured by (new) lconf I can
543 * still talk to old servers. */
544 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
545 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
553 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
557 void ptlrpc_ni_fini(void)
559 wait_queue_head_t waitq;
560 struct l_wait_info lwi;
564 /* Wait for the event queue to become idle since there may still be
565 * messages in flight with pending events (i.e. the fire-and-forget
566 * messages == client requests and "non-difficult" server
569 for (retries = 0;; retries++) {
570 rc = LNetEQFree(ptlrpc_eq_h);
581 CWARN("Event queue still busy\n");
584 init_waitqueue_head(&waitq);
585 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
586 l_wait_event(waitq, 0, &lwi);
593 lnet_pid_t ptl_get_pid(void)
600 pid = LUSTRE_SRV_LNET_PID;
605 int ptlrpc_ni_init(void)
611 CDEBUG(D_NET, "My pid is: %x\n", pid);
613 /* We're not passing any limits yet... */
614 rc = LNetNIInit(pid);
616 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
620 /* CAVEAT EMPTOR: how we process portals events is _radically_
621 * different depending on... */
623 /* kernel LNet calls our master callback when there are new event,
624 * because we are guaranteed to get every event via callback,
625 * so we just set EQ size to 0 to avoid overhread of serializing
626 * enqueue/dequeue operations in LNet. */
627 rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
629 /* liblustre calls the master callback when it removes events from the
630 * event queue. The event queue has to be big enough not to drop
632 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
637 CERROR ("Failed to allocate event queue: %d\n", rc);
644 CFS_LIST_HEAD(liblustre_wait_callbacks);
645 CFS_LIST_HEAD(liblustre_idle_callbacks);
646 void *liblustre_services_callback;
649 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
651 int (*fn)(void *arg), void *arg)
653 struct liblustre_wait_callback *llwc;
655 OBD_ALLOC(llwc, sizeof(*llwc));
656 LASSERT (llwc != NULL);
658 llwc->llwc_name = name;
660 llwc->llwc_arg = arg;
661 cfs_list_add_tail(&llwc->llwc_list, callback_list);
667 liblustre_deregister_waitidle_callback (void *opaque)
669 struct liblustre_wait_callback *llwc = opaque;
671 cfs_list_del(&llwc->llwc_list);
672 OBD_FREE(llwc, sizeof(*llwc));
676 liblustre_register_wait_callback (const char *name,
677 int (*fn)(void *arg), void *arg)
679 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
684 liblustre_deregister_wait_callback (void *opaque)
686 liblustre_deregister_waitidle_callback(opaque);
690 liblustre_register_idle_callback (const char *name,
691 int (*fn)(void *arg), void *arg)
693 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
698 liblustre_deregister_idle_callback (void *opaque)
700 liblustre_deregister_waitidle_callback(opaque);
704 liblustre_check_events (int timeout)
711 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
715 LASSERT (rc == -EOVERFLOW || rc == 1);
717 /* liblustre: no asynch callback so we can't afford to miss any
719 if (rc == -EOVERFLOW) {
720 CERROR ("Dropped an event!!!\n");
724 ptlrpc_master_callback (&ev);
728 int liblustre_waiting = 0;
731 liblustre_wait_event (int timeout)
734 struct liblustre_wait_callback *llwc;
735 int found_something = 0;
737 /* single threaded recursion check... */
738 liblustre_waiting = 1;
741 /* Deal with all pending events */
742 while (liblustre_check_events(0))
745 /* Give all registered callbacks a bite at the cherry */
746 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
747 llwc = cfs_list_entry(tmp,
748 struct liblustre_wait_callback,
751 if (llwc->llwc_fn(llwc->llwc_arg))
755 if (found_something || timeout == 0)
758 /* Nothing so far, but I'm allowed to block... */
759 found_something = liblustre_check_events(timeout);
760 if (!found_something) /* still nothing */
761 break; /* I timed out */
764 liblustre_waiting = 0;
766 return found_something;
770 liblustre_wait_idle(void)
772 static int recursed = 0;
775 struct liblustre_wait_callback *llwc;
782 liblustre_wait_event(0);
786 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
787 llwc = cfs_list_entry(tmp,
788 struct liblustre_wait_callback,
791 if (!llwc->llwc_fn(llwc->llwc_arg)) {
802 #endif /* __KERNEL__ */
804 int ptlrpc_init_portals(void)
806 int rc = ptlrpc_ni_init();
809 CERROR("network initialisation failed\n");
813 liblustre_services_callback =
814 liblustre_register_wait_callback("liblustre_check_services",
815 &liblustre_check_services,
817 init_completion_module(liblustre_wait_event);
819 rc = ptlrpcd_addref();
823 CERROR("rpcd initialisation failed\n");
825 liblustre_deregister_wait_callback(liblustre_services_callback);
831 void ptlrpc_exit_portals(void)
834 liblustre_deregister_wait_callback(liblustre_services_callback);