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 lnet_handler_t ptlrpc_handler;
44 struct percpu_ref ptlrpc_pending;
47 * Client's outgoing request callback
49 void request_out_callback(struct lnet_event *ev)
51 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
52 struct ptlrpc_request *req = cbid->cbid_arg;
56 LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK);
57 LASSERT(ev->unlinked);
59 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
61 /* Do not update imp_next_ping for connection request */
62 if (lustre_msg_get_opc(req->rq_reqmsg) !=
63 req->rq_import->imp_connect_op)
64 ptlrpc_pinger_sending_on_import(req->rq_import);
66 sptlrpc_request_out_callback(req);
68 spin_lock(&req->rq_lock);
69 req->rq_real_sent = ktime_get_real_seconds();
70 req->rq_req_unlinked = 1;
71 /* reply_in_callback happened before request_out_callback? */
72 if (req->rq_reply_unlinked)
75 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
76 /* Failed send: make it seem like the reply timed out, just
77 * like failing sends in client.c does currently... */
83 ptlrpc_client_wake_req(req);
85 spin_unlock(&req->rq_lock);
87 ptlrpc_req_finished(req);
92 * Client's incoming reply callback
94 void reply_in_callback(struct lnet_event *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.
108 LASSERT((ev->md_options & LNET_MD_MANAGE_REMOTE) != 0);
110 spin_lock(&req->rq_lock);
112 req->rq_receiving_reply = 0;
115 req->rq_reply_unlinked = 1;
120 if (ev->type == LNET_EVENT_UNLINK) {
121 LASSERT(ev->unlinked);
122 DEBUG_REQ(D_NET, req, "unlink");
126 if (ev->mlength < ev->rlength ) {
127 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
128 req->rq_replen, ev->rlength, ev->offset);
129 req->rq_reply_truncated = 1;
131 req->rq_status = -EOVERFLOW;
132 req->rq_nob_received = ev->rlength + ev->offset;
136 if ((ev->offset == 0) &&
137 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
139 DEBUG_REQ(D_ADAPTTO, req,
140 "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
141 ev->mlength, ev->offset,
142 req->rq_replen, req->rq_replied, ev->unlinked);
144 req->rq_early_count++; /* number received, client side */
146 /* already got the real reply or buffers are already unlinked */
147 if (req->rq_replied ||
148 req->rq_reply_unlinked == 1)
152 req->rq_reply_off = ev->offset;
153 req->rq_nob_received = ev->mlength;
154 /* And we're still receiving */
155 req->rq_receiving_reply = 1;
158 req->rq_rep_swab_mask = 0;
160 /* Got reply, no resend required */
162 req->rq_reply_off = ev->offset;
163 req->rq_nob_received = ev->mlength;
164 /* LNetMDUnlink can't be called under the LNET_LOCK,
165 so we must unlink in ptlrpc_unregister_reply */
166 DEBUG_REQ(D_INFO, req,
167 "reply in flags=%x mlen=%u offset=%d replen=%d",
168 lustre_msg_get_flags(req->rq_reqmsg),
169 ev->mlength, ev->offset, req->rq_replen);
172 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
173 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
176 /* NB don't unlock till after wakeup; req can disappear under us
177 * since we don't have our own ref */
178 ptlrpc_client_wake_req(req);
179 spin_unlock(&req->rq_lock);
184 * Client's bulk has been written/read
186 void client_bulk_callback(struct lnet_event *ev)
188 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
189 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
190 struct ptlrpc_request *req;
193 LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
194 ev->type == LNET_EVENT_PUT) ||
195 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
196 ev->type == LNET_EVENT_GET) ||
197 ev->type == LNET_EVENT_UNLINK);
198 LASSERT(ev->unlinked);
200 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
203 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
206 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
207 "event type %d, status %d, desc %p\n",
208 ev->type, ev->status, desc);
210 spin_lock(&desc->bd_lock);
212 LASSERT(desc->bd_refs > 0);
215 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
216 desc->bd_nob_transferred += ev->mlength;
217 desc->bd_sender = ev->sender;
219 /* start reconnect and resend if network error hit */
220 spin_lock(&req->rq_lock);
222 spin_unlock(&req->rq_lock);
226 desc->bd_failure = 1;
228 /* NB don't unlock till after wakeup; desc can disappear under us
230 if (desc->bd_refs == 0)
231 ptlrpc_client_wake_req(desc->bd_req);
233 spin_unlock(&desc->bd_lock);
238 * We will have percpt request history list for ptlrpc service in upcoming
239 * patches because we don't want to be serialized by current per-service
240 * history operations. So we require history ID can (somehow) show arriving
241 * order w/o grabbing global lock, and user can sort them in userspace.
243 * This is how we generate history ID for ptlrpc_request:
244 * ----------------------------------------------------
245 * | 32 bits | 16 bits | (16 - X)bits | X bits |
246 * ----------------------------------------------------
247 * | seconds | usec / 16 | sequence | CPT id |
248 * ----------------------------------------------------
250 * it might not be precise but should be good enough.
253 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
255 #define REQS_SEC_SHIFT 32
256 #define REQS_USEC_SHIFT 16
257 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
259 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
260 struct ptlrpc_request *req)
262 u64 sec = req->rq_arrival_time.tv_sec;
263 u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
266 /* set sequence ID for request and add it to history list,
267 * it must be called with hold svcpt::scp_lock */
269 new_seq = (sec << REQS_SEC_SHIFT) |
270 (usec << REQS_USEC_SHIFT) |
271 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
273 if (new_seq > svcpt->scp_hist_seq) {
274 /* This handles the initial case of scp_hist_seq == 0 or
275 * we just jumped into a new time window */
276 svcpt->scp_hist_seq = new_seq;
278 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
279 /* NB: increase sequence number in current usec bucket,
280 * however, it's possible that we used up all bits for
281 * sequence and jumped into the next usec bucket (future time),
282 * then we hope there will be less RPCs per bucket at some
283 * point, and sequence will catch up again */
284 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
285 new_seq = svcpt->scp_hist_seq;
288 req->rq_history_seq = new_seq;
290 list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
294 * Server's incoming request callback
296 void request_in_callback(struct lnet_event *ev)
298 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
299 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
300 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
301 struct ptlrpc_service *service = svcpt->scp_service;
302 struct ptlrpc_request *req;
305 LASSERT(ev->type == LNET_EVENT_PUT ||
306 ev->type == LNET_EVENT_UNLINK);
307 LASSERT((char *)ev->md_start >= rqbd->rqbd_buffer);
308 LASSERT((char *)ev->md_start + ev->offset + ev->mlength <=
309 rqbd->rqbd_buffer + service->srv_buf_size);
311 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
312 "event type %d, status %d, service %s\n",
313 ev->type, ev->status, service->srv_name);
316 /* If this is the last request message to fit in the
317 * request buffer we can use the request object embedded in
318 * rqbd. Note that if we failed to allocate a request,
319 * we'd have to re-post the rqbd, which we can't do in this
321 req = &rqbd->rqbd_req;
322 memset(req, 0, sizeof (*req));
324 LASSERT (ev->type == LNET_EVENT_PUT);
325 if (ev->status != 0) {
326 /* We moaned above already... */
329 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
331 CERROR("Can't allocate incoming request descriptor: "
332 "Dropping %s RPC from %s\n",
334 libcfs_id2str(ev->initiator));
339 ptlrpc_srv_req_init(req);
340 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
341 * flags are reset and scalars are zero. We only set the message
342 * size to non-zero if this was a successful receive. */
343 req->rq_xid = ev->match_bits;
344 req->rq_reqbuf = ev->md_start + ev->offset;
345 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
346 req->rq_reqdata_len = ev->mlength;
347 ktime_get_real_ts64(&req->rq_arrival_time);
348 /* Multi-Rail: keep track of both initiator and source NID. */
349 req->rq_peer = ev->initiator;
350 req->rq_source = ev->source;
351 req->rq_self = ev->target.nid;
353 req->rq_phase = RQ_PHASE_NEW;
354 if (ev->type == LNET_EVENT_PUT)
355 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
356 req, req->rq_xid, ev->mlength);
358 CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
359 libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source));
361 spin_lock(&svcpt->scp_lock);
363 ptlrpc_req_add_history(svcpt, req);
366 svcpt->scp_nrqbds_posted--;
367 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
368 svcpt->scp_nrqbds_posted);
370 /* Normally, don't complain about 0 buffers posted; LNET won't
371 * drop incoming reqs since we set the portal lazy */
372 if (test_req_buffer_pressure &&
373 ev->type != LNET_EVENT_UNLINK &&
374 svcpt->scp_nrqbds_posted == 0)
375 CWARN("All %s request buffers busy\n",
378 /* req takes over the network's ref on rqbd */
380 /* req takes a ref on rqbd */
381 rqbd->rqbd_refcount++;
384 list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
385 svcpt->scp_nreqs_incoming++;
387 /* NB everything can disappear under us once the request
388 * has been queued and we unlock, so do the wake now... */
389 wake_up(&svcpt->scp_waitq);
391 spin_unlock(&svcpt->scp_lock);
396 * Server's outgoing reply callback
398 void reply_out_callback(struct lnet_event *ev)
400 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
401 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
402 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
405 LASSERT (ev->type == LNET_EVENT_SEND ||
406 ev->type == LNET_EVENT_ACK ||
407 ev->type == LNET_EVENT_UNLINK);
409 if (!rs->rs_difficult) {
410 /* 'Easy' replies have no further processing so I drop the
411 * net's ref on 'rs' */
412 LASSERT (ev->unlinked);
413 ptlrpc_rs_decref(rs);
418 LASSERT (rs->rs_on_net);
421 /* Last network callback. The net's ref on 'rs' stays put
422 * until ptlrpc_handle_rs() is done with it */
423 spin_lock(&svcpt->scp_rep_lock);
424 spin_lock(&rs->rs_lock);
427 if (!rs->rs_no_ack ||
429 rs->rs_export->exp_obd->obd_last_committed ||
430 list_empty(&rs->rs_obd_list))
431 ptlrpc_schedule_difficult_reply(rs);
433 spin_unlock(&rs->rs_lock);
434 spin_unlock(&svcpt->scp_rep_lock);
439 #ifdef HAVE_SERVER_SUPPORT
441 * Server's bulk completion callback
443 void server_bulk_callback(struct lnet_event *ev)
445 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
446 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
449 LASSERT(ev->type == LNET_EVENT_SEND ||
450 ev->type == LNET_EVENT_UNLINK ||
451 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
452 ev->type == LNET_EVENT_ACK) ||
453 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
454 ev->type == LNET_EVENT_REPLY));
456 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
457 "event type %d, status %d, desc %p\n",
458 ev->type, ev->status, desc);
460 spin_lock(&desc->bd_lock);
462 LASSERT(desc->bd_refs > 0);
464 if ((ev->type == LNET_EVENT_ACK ||
465 ev->type == LNET_EVENT_REPLY) &&
467 /* We heard back from the peer, so even if we get this
468 * before the SENT event (oh yes we can), we know we
469 * read/wrote the peer buffer and how much... */
470 desc->bd_nob_transferred += ev->mlength;
471 desc->bd_sender = ev->sender;
475 desc->bd_failure = 1;
479 /* This is the last callback no matter what... */
480 if (desc->bd_refs == 0)
481 wake_up(&desc->bd_waitq);
484 spin_unlock(&desc->bd_lock);
489 static void ptlrpc_master_callback(struct lnet_event *ev)
491 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
492 void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
494 /* Honestly, it's best to find out early. */
495 LASSERT(cbid->cbid_arg != LP_POISON);
496 LASSERT(callback == request_out_callback ||
497 callback == reply_in_callback ||
498 callback == client_bulk_callback ||
499 callback == request_in_callback ||
500 callback == reply_out_callback
501 #ifdef HAVE_SERVER_SUPPORT
502 || callback == server_bulk_callback
508 percpu_ref_put(&ptlrpc_pending);
511 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
512 struct lnet_process_id *peer, lnet_nid_t *self)
515 __u32 best_order = 0;
523 peer->pid = LNET_PID_LUSTRE;
525 /* Choose the matching UUID that's closest */
526 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
527 if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 &&
528 LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid))
531 dist = LNetDist(dst_nid, &src_nid, &order);
535 if (dist == 0) { /* local! use loopback LND */
536 peer->nid = *self = LNET_NID_LO_0;
543 (dist == best_dist && order < best_order)) {
553 CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
557 static struct completion ptlrpc_done;
559 static void ptlrpc_release(struct percpu_ref *ref)
561 complete(&ptlrpc_done);
564 static void ptlrpc_ni_fini(void)
566 /* Wait for the event queue to become idle since there may still be
567 * messages in flight with pending events (i.e. the fire-and-forget
568 * messages == client requests and "non-difficult" server
571 init_completion(&ptlrpc_done);
572 percpu_ref_kill(&ptlrpc_pending);
573 wait_for_completion(&ptlrpc_done);
575 lnet_assert_handler_unused(ptlrpc_handler);
579 lnet_pid_t ptl_get_pid(void)
581 return LNET_PID_LUSTRE;
584 int ptlrpc_ni_init(void)
590 CDEBUG(D_NET, "My pid is: %x\n", pid);
592 /* We're not passing any limits yet... */
593 rc = LNetNIInit(pid);
595 CDEBUG(D_NET, "ptlrpc: Can't init network interface: rc = %d\n",
600 rc = percpu_ref_init(&ptlrpc_pending, ptlrpc_release, 0, GFP_KERNEL);
602 CERROR("ptlrpc: Can't init percpu refcount: rc = %d\n", rc);
605 /* CAVEAT EMPTOR: how we process portals events is _radically_
606 * different depending on...
608 /* kernel LNet calls our master callback when there are new event,
609 * because we are guaranteed to get every event via callback,
610 * so we just set EQ size to 0 to avoid overhread of serializing
611 * enqueue/dequeue operations in LNet. */
612 ptlrpc_handler = ptlrpc_master_callback;
616 int ptlrpc_init_portals(void)
618 int rc = ptlrpc_ni_init();
621 CERROR("network initialisation failed\n");
624 rc = ptlrpcd_addref();
628 CERROR("rpcd initialisation failed\n");
633 void ptlrpc_exit_portals(void)