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;
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 sptlrpc_request_out_callback(req);
63 spin_lock(&req->rq_lock);
64 req->rq_real_sent = ktime_get_real_seconds();
65 req->rq_req_unlinked = 1;
66 /* reply_in_callback happened before request_out_callback? */
67 if (req->rq_reply_unlinked)
70 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
71 /* Failed send: make it seem like the reply timed out, just
72 * like failing sends in client.c does currently... */
78 ptlrpc_client_wake_req(req);
80 spin_unlock(&req->rq_lock);
82 ptlrpc_req_finished(req);
87 * Client's incoming reply callback
89 void reply_in_callback(struct lnet_event *ev)
91 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
92 struct ptlrpc_request *req = cbid->cbid_arg;
95 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
97 LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
98 LASSERT (ev->md.start == req->rq_repbuf);
99 LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
100 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
101 for adaptive timeouts' early reply. */
102 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
104 spin_lock(&req->rq_lock);
106 req->rq_receiving_reply = 0;
109 req->rq_reply_unlinked = 1;
114 if (ev->type == LNET_EVENT_UNLINK) {
115 LASSERT(ev->unlinked);
116 DEBUG_REQ(D_NET, req, "unlink");
120 if (ev->mlength < ev->rlength ) {
121 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
122 req->rq_replen, ev->rlength, ev->offset);
123 req->rq_reply_truncated = 1;
125 req->rq_status = -EOVERFLOW;
126 req->rq_nob_received = ev->rlength + ev->offset;
130 if ((ev->offset == 0) &&
131 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
133 DEBUG_REQ(D_ADAPTTO, req,
134 "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
135 ev->mlength, ev->offset,
136 req->rq_replen, req->rq_replied, ev->unlinked);
138 req->rq_early_count++; /* number received, client side */
140 /* already got the real reply or buffers are already unlinked */
141 if (req->rq_replied ||
142 req->rq_reply_unlinked == 1)
146 req->rq_reply_off = ev->offset;
147 req->rq_nob_received = ev->mlength;
148 /* And we're still receiving */
149 req->rq_receiving_reply = 1;
152 req->rq_rep_swab_mask = 0;
154 /* Got reply, no resend required */
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 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
167 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
170 /* NB don't unlock till after wakeup; req can disappear under us
171 * since we don't have our own ref */
172 ptlrpc_client_wake_req(req);
173 spin_unlock(&req->rq_lock);
178 * Client's bulk has been written/read
180 void client_bulk_callback(struct lnet_event *ev)
182 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
183 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
184 struct ptlrpc_request *req;
187 LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
188 ev->type == LNET_EVENT_PUT) ||
189 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
190 ev->type == LNET_EVENT_GET) ||
191 ev->type == LNET_EVENT_UNLINK);
192 LASSERT(ev->unlinked);
194 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
197 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
200 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
201 "event type %d, status %d, desc %p\n",
202 ev->type, ev->status, desc);
204 spin_lock(&desc->bd_lock);
206 LASSERT(desc->bd_md_count > 0);
209 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
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);
220 desc->bd_failure = 1;
222 /* NB don't unlock till after wakeup; desc can disappear under us
224 if (desc->bd_md_count == 0)
225 ptlrpc_client_wake_req(desc->bd_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_nsec / NSEC_PER_USEC / 16; /* 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 list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
288 * Server's incoming request callback
290 void request_in_callback(struct lnet_event *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 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
325 CERROR("Can't allocate incoming request descriptor: "
326 "Dropping %s RPC from %s\n",
328 libcfs_id2str(ev->initiator));
333 ptlrpc_srv_req_init(req);
334 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
335 * flags are reset and scalars are zero. We only set the message
336 * size to non-zero if this was a successful receive. */
337 req->rq_xid = ev->match_bits;
338 req->rq_reqbuf = ev->md.start + ev->offset;
339 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
340 req->rq_reqdata_len = ev->mlength;
341 ktime_get_real_ts64(&req->rq_arrival_time);
342 /* Multi-Rail: keep track of both initiator and source NID. */
343 req->rq_peer = ev->initiator;
344 req->rq_source = ev->source;
345 req->rq_self = ev->target.nid;
347 req->rq_phase = RQ_PHASE_NEW;
348 if (ev->type == LNET_EVENT_PUT)
349 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
350 req, req->rq_xid, ev->mlength);
352 CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
353 libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source));
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 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 wake_up(&svcpt->scp_waitq);
385 spin_unlock(&svcpt->scp_lock);
390 * Server's outgoing reply callback
392 void reply_out_callback(struct lnet_event *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 list_empty(&rs->rs_obd_list))
425 ptlrpc_schedule_difficult_reply(rs);
427 spin_unlock(&rs->rs_lock);
428 spin_unlock(&svcpt->scp_rep_lock);
433 #ifdef HAVE_SERVER_SUPPORT
435 * Server's bulk completion callback
437 void server_bulk_callback(struct lnet_event *ev)
439 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
440 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
443 LASSERT(ev->type == LNET_EVENT_SEND ||
444 ev->type == LNET_EVENT_UNLINK ||
445 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
446 ev->type == LNET_EVENT_ACK) ||
447 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
448 ev->type == LNET_EVENT_REPLY));
450 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
451 "event type %d, status %d, desc %p\n",
452 ev->type, ev->status, desc);
454 spin_lock(&desc->bd_lock);
456 LASSERT(desc->bd_md_count > 0);
458 if ((ev->type == LNET_EVENT_ACK ||
459 ev->type == LNET_EVENT_REPLY) &&
461 /* We heard back from the peer, so even if we get this
462 * before the SENT event (oh yes we can), we know we
463 * read/wrote the peer buffer and how much... */
464 desc->bd_nob_transferred += ev->mlength;
465 desc->bd_sender = ev->sender;
469 desc->bd_failure = 1;
473 /* This is the last callback no matter what... */
474 if (desc->bd_md_count == 0)
475 wake_up(&desc->bd_waitq);
478 spin_unlock(&desc->bd_lock);
483 static void ptlrpc_master_callback(struct lnet_event *ev)
485 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
486 void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
488 /* Honestly, it's best to find out early. */
489 LASSERT(cbid->cbid_arg != LP_POISON);
490 LASSERT(callback == request_out_callback ||
491 callback == reply_in_callback ||
492 callback == client_bulk_callback ||
493 callback == request_in_callback ||
494 callback == reply_out_callback
495 #ifdef HAVE_SERVER_SUPPORT
496 || callback == server_bulk_callback
502 percpu_ref_put(&ptlrpc_pending);
505 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
506 struct lnet_process_id *peer, lnet_nid_t *self)
509 __u32 best_order = 0;
517 peer->pid = LNET_PID_LUSTRE;
519 /* Choose the matching UUID that's closest */
520 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
521 if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 &&
522 LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid))
525 dist = LNetDist(dst_nid, &src_nid, &order);
529 if (dist == 0) { /* local! use loopback LND */
530 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
537 (dist == best_dist && order < best_order)) {
547 CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
551 static struct completion ptlrpc_done;
553 static void ptlrpc_release(struct percpu_ref *ref)
555 complete(&ptlrpc_done);
558 static void ptlrpc_ni_fini(void)
560 /* Wait for the event queue to become idle since there may still be
561 * messages in flight with pending events (i.e. the fire-and-forget
562 * messages == client requests and "non-difficult" server
565 init_completion(&ptlrpc_done);
566 percpu_ref_kill(&ptlrpc_pending);
567 wait_for_completion(&ptlrpc_done);
569 LNetEQFree(ptlrpc_eq);
573 lnet_pid_t ptl_get_pid(void)
575 return LNET_PID_LUSTRE;
578 int ptlrpc_ni_init(void)
584 CDEBUG(D_NET, "My pid is: %x\n", pid);
586 /* We're not passing any limits yet... */
587 rc = LNetNIInit(pid);
589 CDEBUG(D_NET, "ptlrpc: Can't init network interface: rc = %d\n",
594 rc = percpu_ref_init(&ptlrpc_pending, ptlrpc_release, 0, GFP_KERNEL);
596 CERROR("ptlrpc: Can't init percpu refcount: rc = %d\n", rc);
599 /* CAVEAT EMPTOR: how we process portals events is _radically_
600 * 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)