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 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val &&
60 CFS_FAIL_CHECK_RESET(OBD_FAIL_NET_ERROR_RPC,
61 OBD_FAIL_OSP_PRECREATE_PAUSE |
63 ev->status = -ECONNABORTED;
65 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
67 /* Do not update imp_next_ping for connection request */
68 if (lustre_msg_get_opc(req->rq_reqmsg) !=
69 req->rq_import->imp_connect_op)
70 ptlrpc_pinger_sending_on_import(req->rq_import);
72 sptlrpc_request_out_callback(req);
74 spin_lock(&req->rq_lock);
75 req->rq_real_sent = ktime_get_real_seconds();
76 req->rq_req_unlinked = 1;
77 /* reply_in_callback happened before request_out_callback? */
78 if (req->rq_reply_unlinked)
81 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
82 /* Failed send: make it seem like the reply timed out, just
83 * like failing sends in client.c does currently... */
89 ptlrpc_client_wake_req(req);
91 spin_unlock(&req->rq_lock);
93 ptlrpc_req_finished(req);
98 * Client's incoming reply callback
100 void reply_in_callback(struct lnet_event *ev)
102 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
103 struct ptlrpc_request *req = cbid->cbid_arg;
106 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
108 LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
109 LASSERT(ev->md_start == req->rq_repbuf);
110 LASSERT(ev->offset + ev->mlength <= req->rq_repbuf_len);
111 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
112 * for adaptive timeouts' early reply.
114 LASSERT((ev->md_options & LNET_MD_MANAGE_REMOTE) != 0);
116 spin_lock(&req->rq_lock);
118 req->rq_receiving_reply = 0;
121 req->rq_reply_unlinked = 1;
126 if (ev->type == LNET_EVENT_UNLINK) {
127 LASSERT(ev->unlinked);
128 DEBUG_REQ(D_NET, req, "unlink");
132 if (ev->mlength < ev->rlength ) {
133 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
134 req->rq_replen, ev->rlength, ev->offset);
135 req->rq_reply_truncated = 1;
137 req->rq_status = -EOVERFLOW;
138 req->rq_nob_received = ev->rlength + ev->offset;
142 if ((ev->offset == 0) &&
143 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
145 DEBUG_REQ(D_ADAPTTO, req,
146 "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
147 ev->mlength, ev->offset,
148 req->rq_replen, req->rq_replied, ev->unlinked);
150 req->rq_early_count++; /* number received, client side */
152 /* already got the real reply or buffers are already unlinked */
153 if (req->rq_replied ||
154 req->rq_reply_unlinked == 1)
158 req->rq_reply_off = ev->offset;
159 req->rq_nob_received = ev->mlength;
160 /* And we're still receiving */
161 req->rq_receiving_reply = 1;
164 req->rq_rep_swab_mask = 0;
166 /* Got reply, no resend required */
168 req->rq_reply_off = ev->offset;
169 req->rq_nob_received = ev->mlength;
170 /* LNetMDUnlink can't be called under the LNET_LOCK,
171 so we must unlink in ptlrpc_unregister_reply */
172 DEBUG_REQ(D_INFO, req,
173 "reply in flags=%x mlen=%u offset=%d replen=%d",
174 lustre_msg_get_flags(req->rq_reqmsg),
175 ev->mlength, ev->offset, req->rq_replen);
178 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
179 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
182 /* NB don't unlock till after wakeup; req can disappear under us
183 * since we don't have our own ref */
184 ptlrpc_client_wake_req(req);
185 spin_unlock(&req->rq_lock);
190 * Client's bulk has been written/read
192 void client_bulk_callback(struct lnet_event *ev)
194 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
195 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
196 struct ptlrpc_request *req;
199 LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
200 ev->type == LNET_EVENT_PUT) ||
201 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
202 ev->type == LNET_EVENT_GET) ||
203 ev->type == LNET_EVENT_UNLINK);
204 LASSERT(ev->unlinked);
206 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
209 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
212 CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
213 "event type %d, status %d, desc %p\n",
214 ev->type, ev->status, desc);
216 spin_lock(&desc->bd_lock);
218 LASSERT(desc->bd_refs > 0);
221 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
222 desc->bd_nob_transferred += ev->mlength;
223 desc->bd_sender = ev->sender;
225 /* start reconnect and resend if network error hit */
226 spin_lock(&req->rq_lock);
228 spin_unlock(&req->rq_lock);
232 desc->bd_failure = 1;
234 /* NB don't unlock till after wakeup; desc can disappear under us
236 if (desc->bd_refs == 0)
237 ptlrpc_client_wake_req(desc->bd_req);
239 spin_unlock(&desc->bd_lock);
244 * We will have percpt request history list for ptlrpc service in upcoming
245 * patches because we don't want to be serialized by current per-service
246 * history operations. So we require history ID can (somehow) show arriving
247 * order w/o grabbing global lock, and user can sort them in userspace.
249 * This is how we generate history ID for ptlrpc_request:
250 * ----------------------------------------------------
251 * | 32 bits | 16 bits | (16 - X)bits | X bits |
252 * ----------------------------------------------------
253 * | seconds | usec / 16 | sequence | CPT id |
254 * ----------------------------------------------------
256 * it might not be precise but should be good enough.
259 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
261 #define REQS_SEC_SHIFT 32
262 #define REQS_USEC_SHIFT 16
263 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
265 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
266 struct ptlrpc_request *req)
268 u64 sec = req->rq_arrival_time.tv_sec;
269 u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
272 /* set sequence ID for request and add it to history list,
273 * it must be called with hold svcpt::scp_lock */
275 new_seq = (sec << REQS_SEC_SHIFT) |
276 (usec << REQS_USEC_SHIFT) |
277 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
279 if (new_seq > svcpt->scp_hist_seq) {
280 /* This handles the initial case of scp_hist_seq == 0 or
281 * we just jumped into a new time window */
282 svcpt->scp_hist_seq = new_seq;
284 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
285 /* NB: increase sequence number in current usec bucket,
286 * however, it's possible that we used up all bits for
287 * sequence and jumped into the next usec bucket (future time),
288 * then we hope there will be less RPCs per bucket at some
289 * point, and sequence will catch up again */
290 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
291 new_seq = svcpt->scp_hist_seq;
294 req->rq_history_seq = new_seq;
296 list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
300 * Server's incoming request callback
302 void request_in_callback(struct lnet_event *ev)
304 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
305 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
306 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
307 struct ptlrpc_service *service = svcpt->scp_service;
308 struct ptlrpc_request *req;
311 LASSERT(ev->type == LNET_EVENT_PUT ||
312 ev->type == LNET_EVENT_UNLINK);
313 LASSERT((char *)ev->md_start >= rqbd->rqbd_buffer);
314 LASSERT((char *)ev->md_start + ev->offset + ev->mlength <=
315 rqbd->rqbd_buffer + service->srv_buf_size);
317 CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
318 "event type %d, status %d, service %s\n",
319 ev->type, ev->status, service->srv_name);
322 /* If this is the last request message to fit in the
323 * request buffer we can use the request object embedded in
324 * rqbd. Note that if we failed to allocate a request,
325 * we'd have to re-post the rqbd, which we can't do in this
328 req = &rqbd->rqbd_req;
329 memset(req, 0, sizeof(*req));
331 LASSERT(ev->type == LNET_EVENT_PUT);
332 if (ev->status != 0) /* We moaned above already... */
334 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
336 CERROR("Can't allocate incoming request descriptor: "
337 "Dropping %s RPC from %s\n",
339 libcfs_id2str(ev->initiator));
344 ptlrpc_srv_req_init(req);
345 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
346 * flags are reset and scalars are zero. We only set the message
347 * size to non-zero if this was a successful receive. */
348 req->rq_xid = ev->match_bits;
349 req->rq_reqbuf = ev->md_start + ev->offset;
350 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
351 req->rq_reqdata_len = ev->mlength;
352 ktime_get_real_ts64(&req->rq_arrival_time);
353 /* Multi-Rail: keep track of both initiator and source NID. */
354 req->rq_peer = ev->initiator;
355 req->rq_source = ev->source;
356 req->rq_self = ev->target.nid;
358 req->rq_phase = RQ_PHASE_NEW;
359 if (ev->type == LNET_EVENT_PUT)
360 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
361 req, req->rq_xid, ev->mlength);
363 CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
364 libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source));
366 spin_lock(&svcpt->scp_lock);
368 ptlrpc_req_add_history(svcpt, req);
371 svcpt->scp_nrqbds_posted--;
372 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
373 svcpt->scp_nrqbds_posted);
375 /* Normally, don't complain about 0 buffers posted; LNET won't
376 * drop incoming reqs since we set the portal lazy */
377 if (test_req_buffer_pressure &&
378 ev->type != LNET_EVENT_UNLINK &&
379 svcpt->scp_nrqbds_posted == 0)
380 CWARN("All %s request buffers busy\n",
383 /* req takes over the network's ref on rqbd */
385 /* req takes a ref on rqbd */
386 rqbd->rqbd_refcount++;
389 list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
390 svcpt->scp_nreqs_incoming++;
392 /* NB everything can disappear under us once the request
393 * has been queued and we unlock, so do the wake now... */
394 wake_up(&svcpt->scp_waitq);
396 spin_unlock(&svcpt->scp_lock);
401 * Server's outgoing reply callback
403 void reply_out_callback(struct lnet_event *ev)
405 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
406 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
407 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
410 LASSERT (ev->type == LNET_EVENT_SEND ||
411 ev->type == LNET_EVENT_ACK ||
412 ev->type == LNET_EVENT_UNLINK);
414 if (!rs->rs_difficult) {
415 /* 'Easy' replies have no further processing so I drop the
416 * net's ref on 'rs' */
417 LASSERT (ev->unlinked);
418 ptlrpc_rs_decref(rs);
423 LASSERT (rs->rs_on_net);
426 /* Last network callback. The net's ref on 'rs' stays put
427 * until ptlrpc_handle_rs() is done with it */
428 spin_lock(&svcpt->scp_rep_lock);
429 spin_lock(&rs->rs_lock);
432 if (!rs->rs_no_ack ||
434 rs->rs_export->exp_obd->obd_last_committed ||
435 list_empty(&rs->rs_obd_list))
436 ptlrpc_schedule_difficult_reply(rs);
438 spin_unlock(&rs->rs_lock);
439 spin_unlock(&svcpt->scp_rep_lock);
444 #ifdef HAVE_SERVER_SUPPORT
446 * Server's bulk completion callback
448 void server_bulk_callback(struct lnet_event *ev)
450 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
451 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
454 LASSERT(ev->type == LNET_EVENT_SEND ||
455 ev->type == LNET_EVENT_UNLINK ||
456 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
457 ev->type == LNET_EVENT_ACK) ||
458 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
459 ev->type == LNET_EVENT_REPLY));
461 CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
462 "event type %d, status %d, desc %p\n",
463 ev->type, ev->status, desc);
465 spin_lock(&desc->bd_lock);
467 LASSERT(desc->bd_refs > 0);
469 if ((ev->type == LNET_EVENT_ACK ||
470 ev->type == LNET_EVENT_REPLY) &&
472 /* We heard back from the peer, so even if we get this
473 * before the SENT event (oh yes we can), we know we
474 * read/wrote the peer buffer and how much... */
475 desc->bd_nob_transferred += ev->mlength;
476 desc->bd_sender = ev->sender;
480 desc->bd_failure = 1;
484 /* This is the last callback no matter what... */
485 if (desc->bd_refs == 0)
486 wake_up(&desc->bd_waitq);
489 spin_unlock(&desc->bd_lock);
494 static void ptlrpc_master_callback(struct lnet_event *ev)
496 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
497 void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
499 /* Honestly, it's best to find out early. */
500 LASSERT(cbid->cbid_arg != LP_POISON);
501 LASSERT(callback == request_out_callback ||
502 callback == reply_in_callback ||
503 callback == client_bulk_callback ||
504 callback == request_in_callback ||
505 callback == reply_out_callback
506 #ifdef HAVE_SERVER_SUPPORT
507 || callback == server_bulk_callback
513 percpu_ref_put(&ptlrpc_pending);
516 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
517 struct lnet_process_id *peer, lnet_nid_t *self)
520 __u32 best_order = 0;
528 peer->pid = LNET_PID_LUSTRE;
530 /* Choose the matching UUID that's closest */
531 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
532 if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 &&
533 LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid))
536 dist = LNetDist(dst_nid, &src_nid, &order);
540 if (dist == 0) { /* local! use loopback LND */
541 peer->nid = *self = LNET_NID_LO_0;
548 (dist == best_dist && order < best_order)) {
558 CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
562 static struct completion ptlrpc_done;
564 static void ptlrpc_release(struct percpu_ref *ref)
566 complete(&ptlrpc_done);
569 static void ptlrpc_ni_fini(void)
571 /* Wait for the event queue to become idle since there may still be
572 * messages in flight with pending events (i.e. the fire-and-forget
573 * messages == client requests and "non-difficult" server
576 init_completion(&ptlrpc_done);
577 percpu_ref_kill(&ptlrpc_pending);
578 wait_for_completion(&ptlrpc_done);
580 lnet_assert_handler_unused(ptlrpc_handler);
584 lnet_pid_t ptl_get_pid(void)
586 return LNET_PID_LUSTRE;
589 int ptlrpc_ni_init(void)
595 CDEBUG(D_NET, "My pid is: %x\n", pid);
597 /* We're not passing any limits yet... */
598 rc = LNetNIInit(pid);
600 CDEBUG(D_NET, "ptlrpc: Can't init network interface: rc = %d\n",
605 rc = percpu_ref_init(&ptlrpc_pending, ptlrpc_release, 0, GFP_KERNEL);
607 CERROR("ptlrpc: Can't init percpu refcount: rc = %d\n", rc);
610 /* CAVEAT EMPTOR: how we process portals events is _radically_
611 * different depending on...
613 /* kernel LNet calls our master callback when there are new event,
614 * because we are guaranteed to get every event via callback,
615 * so we just set EQ size to 0 to avoid overhread of serializing
616 * enqueue/dequeue operations in LNet. */
617 ptlrpc_handler = ptlrpc_master_callback;
621 int ptlrpc_init_portals(void)
623 int rc = ptlrpc_ni_init();
626 CERROR("network initialisation failed\n");
629 rc = ptlrpcd_addref();
633 CERROR("rpcd initialisation failed\n");
638 void ptlrpc_exit_portals(void)