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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
35 #include <libcfs/libcfs.h>
36 #include <linux/kernel.h>
37 #include <linux/delay.h>
38 #include <obd_class.h>
39 #include <lustre_net.h>
40 #include <lustre_sec.h>
41 #include "ptlrpc_internal.h"
43 struct lnet_eq *ptlrpc_eq;
46 * Client's outgoing request callback
48 void request_out_callback(struct lnet_event *ev)
50 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
51 struct ptlrpc_request *req = cbid->cbid_arg;
55 LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK);
56 LASSERT(ev->unlinked);
58 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
60 sptlrpc_request_out_callback(req);
62 spin_lock(&req->rq_lock);
63 req->rq_real_sent = ktime_get_real_seconds();
64 req->rq_req_unlinked = 1;
65 /* reply_in_callback happened before request_out_callback? */
66 if (req->rq_reply_unlinked)
69 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
70 /* Failed send: make it seem like the reply timed out, just
71 * like failing sends in client.c does currently... */
77 ptlrpc_client_wake_req(req);
79 spin_unlock(&req->rq_lock);
81 ptlrpc_req_finished(req);
86 * Client's incoming reply callback
88 void reply_in_callback(struct lnet_event *ev)
90 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
91 struct ptlrpc_request *req = cbid->cbid_arg;
94 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
96 LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
97 LASSERT (ev->md.start == req->rq_repbuf);
98 LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
99 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
100 for adaptive timeouts' early reply. */
101 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
103 spin_lock(&req->rq_lock);
105 req->rq_receiving_reply = 0;
108 req->rq_reply_unlinked = 1;
113 if (ev->type == LNET_EVENT_UNLINK) {
114 LASSERT(ev->unlinked);
115 DEBUG_REQ(D_NET, req, "unlink");
119 if (ev->mlength < ev->rlength ) {
120 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
121 req->rq_replen, ev->rlength, ev->offset);
122 req->rq_reply_truncated = 1;
124 req->rq_status = -EOVERFLOW;
125 req->rq_nob_received = ev->rlength + ev->offset;
129 if ((ev->offset == 0) &&
130 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
132 DEBUG_REQ(D_ADAPTTO, req,
133 "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
134 ev->mlength, ev->offset,
135 req->rq_replen, req->rq_replied, ev->unlinked);
137 req->rq_early_count++; /* number received, client side */
139 /* already got the real reply or buffers are already unlinked */
140 if (req->rq_replied ||
141 req->rq_reply_unlinked == 1)
145 req->rq_reply_off = ev->offset;
146 req->rq_nob_received = ev->mlength;
147 /* And we're still receiving */
148 req->rq_receiving_reply = 1;
151 req->rq_rep_swab_mask = 0;
153 /* Got reply, no resend required */
155 req->rq_reply_off = ev->offset;
156 req->rq_nob_received = ev->mlength;
157 /* LNetMDUnlink can't be called under the LNET_LOCK,
158 so we must unlink in ptlrpc_unregister_reply */
159 DEBUG_REQ(D_INFO, req,
160 "reply in flags=%x mlen=%u offset=%d replen=%d",
161 lustre_msg_get_flags(req->rq_reqmsg),
162 ev->mlength, ev->offset, req->rq_replen);
165 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
166 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
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(struct lnet_event *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((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
187 ev->type == LNET_EVENT_PUT) ||
188 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
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_md_count > 0);
208 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
209 desc->bd_nob_transferred += ev->mlength;
210 desc->bd_sender = ev->sender;
212 /* start reconnect and resend if network error hit */
213 spin_lock(&req->rq_lock);
215 spin_unlock(&req->rq_lock);
219 desc->bd_failure = 1;
221 /* NB don't unlock till after wakeup; desc can disappear under us
223 if (desc->bd_md_count == 0)
224 ptlrpc_client_wake_req(desc->bd_req);
226 spin_unlock(&desc->bd_lock);
231 * We will have percpt request history list for ptlrpc service in upcoming
232 * patches because we don't want to be serialized by current per-service
233 * history operations. So we require history ID can (somehow) show arriving
234 * order w/o grabbing global lock, and user can sort them in userspace.
236 * This is how we generate history ID for ptlrpc_request:
237 * ----------------------------------------------------
238 * | 32 bits | 16 bits | (16 - X)bits | X bits |
239 * ----------------------------------------------------
240 * | seconds | usec / 16 | sequence | CPT id |
241 * ----------------------------------------------------
243 * it might not be precise but should be good enough.
246 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
248 #define REQS_SEC_SHIFT 32
249 #define REQS_USEC_SHIFT 16
250 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
252 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
253 struct ptlrpc_request *req)
255 u64 sec = req->rq_arrival_time.tv_sec;
256 u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
259 /* set sequence ID for request and add it to history list,
260 * it must be called with hold svcpt::scp_lock */
262 new_seq = (sec << REQS_SEC_SHIFT) |
263 (usec << REQS_USEC_SHIFT) |
264 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
266 if (new_seq > svcpt->scp_hist_seq) {
267 /* This handles the initial case of scp_hist_seq == 0 or
268 * we just jumped into a new time window */
269 svcpt->scp_hist_seq = new_seq;
271 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
272 /* NB: increase sequence number in current usec bucket,
273 * however, it's possible that we used up all bits for
274 * sequence and jumped into the next usec bucket (future time),
275 * then we hope there will be less RPCs per bucket at some
276 * point, and sequence will catch up again */
277 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
278 new_seq = svcpt->scp_hist_seq;
281 req->rq_history_seq = new_seq;
283 list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
287 * Server's incoming request callback
289 void request_in_callback(struct lnet_event *ev)
291 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
292 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
293 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
294 struct ptlrpc_service *service = svcpt->scp_service;
295 struct ptlrpc_request *req;
298 LASSERT (ev->type == LNET_EVENT_PUT ||
299 ev->type == LNET_EVENT_UNLINK);
300 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
301 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
302 rqbd->rqbd_buffer + service->srv_buf_size);
304 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
305 "event type %d, status %d, service %s\n",
306 ev->type, ev->status, service->srv_name);
309 /* If this is the last request message to fit in the
310 * request buffer we can use the request object embedded in
311 * rqbd. Note that if we failed to allocate a request,
312 * we'd have to re-post the rqbd, which we can't do in this
314 req = &rqbd->rqbd_req;
315 memset(req, 0, sizeof (*req));
317 LASSERT (ev->type == LNET_EVENT_PUT);
318 if (ev->status != 0) {
319 /* We moaned above already... */
322 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
324 CERROR("Can't allocate incoming request descriptor: "
325 "Dropping %s RPC from %s\n",
327 libcfs_id2str(ev->initiator));
332 ptlrpc_srv_req_init(req);
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 ktime_get_real_ts64(&req->rq_arrival_time);
341 /* Multi-Rail: keep track of both initiator and source NID. */
342 req->rq_peer = ev->initiator;
343 req->rq_source = ev->source;
344 req->rq_self = ev->target.nid;
346 req->rq_phase = RQ_PHASE_NEW;
347 if (ev->type == LNET_EVENT_PUT)
348 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
349 req, req->rq_xid, ev->mlength);
351 CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
352 libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source));
354 spin_lock(&svcpt->scp_lock);
356 ptlrpc_req_add_history(svcpt, req);
359 svcpt->scp_nrqbds_posted--;
360 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
361 svcpt->scp_nrqbds_posted);
363 /* Normally, don't complain about 0 buffers posted; LNET won't
364 * drop incoming reqs since we set the portal lazy */
365 if (test_req_buffer_pressure &&
366 ev->type != LNET_EVENT_UNLINK &&
367 svcpt->scp_nrqbds_posted == 0)
368 CWARN("All %s request buffers busy\n",
371 /* req takes over the network's ref on rqbd */
373 /* req takes a ref on rqbd */
374 rqbd->rqbd_refcount++;
377 list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
378 svcpt->scp_nreqs_incoming++;
380 /* NB everything can disappear under us once the request
381 * has been queued and we unlock, so do the wake now... */
382 wake_up(&svcpt->scp_waitq);
384 spin_unlock(&svcpt->scp_lock);
389 * Server's outgoing reply callback
391 void reply_out_callback(struct lnet_event *ev)
393 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
394 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
395 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
398 LASSERT (ev->type == LNET_EVENT_SEND ||
399 ev->type == LNET_EVENT_ACK ||
400 ev->type == LNET_EVENT_UNLINK);
402 if (!rs->rs_difficult) {
403 /* 'Easy' replies have no further processing so I drop the
404 * net's ref on 'rs' */
405 LASSERT (ev->unlinked);
406 ptlrpc_rs_decref(rs);
411 LASSERT (rs->rs_on_net);
414 /* Last network callback. The net's ref on 'rs' stays put
415 * until ptlrpc_handle_rs() is done with it */
416 spin_lock(&svcpt->scp_rep_lock);
417 spin_lock(&rs->rs_lock);
420 if (!rs->rs_no_ack ||
422 rs->rs_export->exp_obd->obd_last_committed ||
423 list_empty(&rs->rs_obd_list))
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(struct lnet_event *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 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
445 ev->type == LNET_EVENT_ACK) ||
446 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
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 LASSERT(desc->bd_md_count > 0);
457 if ((ev->type == LNET_EVENT_ACK ||
458 ev->type == LNET_EVENT_REPLY) &&
460 /* We heard back from the peer, so even if we get this
461 * before the SENT event (oh yes we can), we know we
462 * read/wrote the peer buffer and how much... */
463 desc->bd_nob_transferred += ev->mlength;
464 desc->bd_sender = ev->sender;
468 desc->bd_failure = 1;
472 /* This is the last callback no matter what... */
473 if (desc->bd_md_count == 0)
474 wake_up(&desc->bd_waitq);
477 spin_unlock(&desc->bd_lock);
482 static void ptlrpc_master_callback(struct lnet_event *ev)
484 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
485 void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
487 /* Honestly, it's best to find out early. */
488 LASSERT (cbid->cbid_arg != LP_POISON);
489 LASSERT (callback == request_out_callback ||
490 callback == reply_in_callback ||
491 callback == client_bulk_callback ||
492 callback == request_in_callback ||
493 callback == reply_out_callback
494 #ifdef HAVE_SERVER_SUPPORT
495 || callback == server_bulk_callback
502 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
503 struct lnet_process_id *peer, lnet_nid_t *self)
506 __u32 best_order = 0;
514 peer->pid = LNET_PID_LUSTRE;
516 /* Choose the matching UUID that's closest */
517 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
518 if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 &&
519 LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid))
522 dist = LNetDist(dst_nid, &src_nid, &order);
526 if (dist == 0) { /* local! use loopback LND */
527 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
534 (dist == best_dist && order < best_order)) {
544 CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
548 void ptlrpc_ni_fini(void)
553 /* Wait for the event queue to become idle since there may still be
554 * messages in flight with pending events (i.e. the fire-and-forget
555 * messages == client requests and "non-difficult" server
558 for (retries = 0;; retries++) {
559 rc = LNetEQFree(ptlrpc_eq);
570 CWARN("Event queue still busy\n");
580 lnet_pid_t ptl_get_pid(void)
582 return LNET_PID_LUSTRE;
585 int ptlrpc_ni_init(void)
591 CDEBUG(D_NET, "My pid is: %x\n", pid);
593 /* We're not passing any limits yet... */
594 rc = LNetNIInit(pid);
596 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
600 /* CAVEAT EMPTOR: how we process portals events is _radically_
601 * different depending on... */
602 /* kernel LNet calls our master callback when there are new event,
603 * because we are guaranteed to get every event via callback,
604 * so we just set EQ size to 0 to avoid overhread of serializing
605 * enqueue/dequeue operations in LNet. */
606 ptlrpc_eq = LNetEQAlloc(ptlrpc_master_callback);
607 if (!IS_ERR(ptlrpc_eq))
610 rc = PTR_ERR(ptlrpc_eq);
611 CERROR("Failed to allocate event queue: %d\n", rc);
617 int ptlrpc_init_portals(void)
619 int rc = ptlrpc_ni_init();
622 CERROR("network initialisation failed\n");
625 rc = ptlrpcd_addref();
629 CERROR("rpcd initialisation failed\n");
634 void ptlrpc_exit_portals(void)