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 <obd_class.h>
38 #include <lustre_net.h>
39 #include <lustre_sec.h>
40 #include "ptlrpc_internal.h"
42 struct lnet_handle_eq ptlrpc_eq_h;
45 * Client's outgoing request callback
47 void request_out_callback(struct lnet_event *ev)
49 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
50 struct ptlrpc_request *req = cbid->cbid_arg;
54 LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK);
55 LASSERT(ev->unlinked);
57 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
59 sptlrpc_request_out_callback(req);
61 spin_lock(&req->rq_lock);
62 req->rq_real_sent = ktime_get_real_seconds();
63 req->rq_req_unlinked = 1;
64 /* reply_in_callback happened before request_out_callback? */
65 if (req->rq_reply_unlinked)
68 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
69 /* Failed send: make it seem like the reply timed out, just
70 * like failing sends in client.c does currently... */
76 ptlrpc_client_wake_req(req);
78 spin_unlock(&req->rq_lock);
80 ptlrpc_req_finished(req);
85 * Client's incoming reply callback
87 void reply_in_callback(struct lnet_event *ev)
89 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
90 struct ptlrpc_request *req = cbid->cbid_arg;
93 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
95 LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
96 LASSERT (ev->md.start == req->rq_repbuf);
97 LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
98 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
99 for adaptive timeouts' early reply. */
100 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
102 spin_lock(&req->rq_lock);
104 req->rq_receiving_reply = 0;
107 req->rq_reply_unlinked = 1;
112 if (ev->type == LNET_EVENT_UNLINK) {
113 LASSERT(ev->unlinked);
114 DEBUG_REQ(D_NET, req, "unlink");
118 if (ev->mlength < ev->rlength ) {
119 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
120 req->rq_replen, ev->rlength, ev->offset);
121 req->rq_reply_truncated = 1;
123 req->rq_status = -EOVERFLOW;
124 req->rq_nob_received = ev->rlength + ev->offset;
128 if ((ev->offset == 0) &&
129 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
131 DEBUG_REQ(D_ADAPTTO, req,
132 "Early reply received: mlen=%u offset=%d replen=%d "
133 "replied=%d unlinked=%d", ev->mlength, ev->offset,
134 req->rq_replen, req->rq_replied, ev->unlinked);
136 req->rq_early_count++; /* number received, client side */
138 /* already got the real reply or buffers are already unlinked */
139 if (req->rq_replied ||
140 req->rq_reply_unlinked == 1)
144 req->rq_reply_off = ev->offset;
145 req->rq_nob_received = ev->mlength;
146 /* And we're still receiving */
147 req->rq_receiving_reply = 1;
150 req->rq_rep_swab_mask = 0;
152 /* Got reply, no resend required */
154 req->rq_reply_off = ev->offset;
155 req->rq_nob_received = ev->mlength;
156 /* LNetMDUnlink can't be called under the LNET_LOCK,
157 so we must unlink in ptlrpc_unregister_reply */
158 DEBUG_REQ(D_INFO, req,
159 "reply in flags=%x mlen=%u offset=%d replen=%d",
160 lustre_msg_get_flags(req->rq_reqmsg),
161 ev->mlength, ev->offset, req->rq_replen);
164 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
165 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
168 /* NB don't unlock till after wakeup; req can disappear under us
169 * since we don't have our own ref */
170 ptlrpc_client_wake_req(req);
171 spin_unlock(&req->rq_lock);
176 * Client's bulk has been written/read
178 void client_bulk_callback(struct lnet_event *ev)
180 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
181 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
182 struct ptlrpc_request *req;
185 LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
186 ev->type == LNET_EVENT_PUT) ||
187 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
188 ev->type == LNET_EVENT_GET) ||
189 ev->type == LNET_EVENT_UNLINK);
190 LASSERT(ev->unlinked);
192 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
195 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
198 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
199 "event type %d, status %d, desc %p\n",
200 ev->type, ev->status, desc);
202 spin_lock(&desc->bd_lock);
204 LASSERT(desc->bd_md_count > 0);
207 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
208 desc->bd_nob_transferred += ev->mlength;
209 desc->bd_sender = ev->sender;
211 /* start reconnect and resend if network error hit */
212 spin_lock(&req->rq_lock);
214 spin_unlock(&req->rq_lock);
218 desc->bd_failure = 1;
220 /* NB don't unlock till after wakeup; desc can disappear under us
222 if (desc->bd_md_count == 0)
223 ptlrpc_client_wake_req(desc->bd_req);
225 spin_unlock(&desc->bd_lock);
230 * We will have percpt request history list for ptlrpc service in upcoming
231 * patches because we don't want to be serialized by current per-service
232 * history operations. So we require history ID can (somehow) show arriving
233 * order w/o grabbing global lock, and user can sort them in userspace.
235 * This is how we generate history ID for ptlrpc_request:
236 * ----------------------------------------------------
237 * | 32 bits | 16 bits | (16 - X)bits | X bits |
238 * ----------------------------------------------------
239 * | seconds | usec / 16 | sequence | CPT id |
240 * ----------------------------------------------------
242 * it might not be precise but should be good enough.
245 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
247 #define REQS_SEC_SHIFT 32
248 #define REQS_USEC_SHIFT 16
249 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
251 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
252 struct ptlrpc_request *req)
254 u64 sec = req->rq_arrival_time.tv_sec;
255 u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
258 /* set sequence ID for request and add it to history list,
259 * it must be called with hold svcpt::scp_lock */
261 new_seq = (sec << REQS_SEC_SHIFT) |
262 (usec << REQS_USEC_SHIFT) |
263 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
265 if (new_seq > svcpt->scp_hist_seq) {
266 /* This handles the initial case of scp_hist_seq == 0 or
267 * we just jumped into a new time window */
268 svcpt->scp_hist_seq = new_seq;
270 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
271 /* NB: increase sequence number in current usec bucket,
272 * however, it's possible that we used up all bits for
273 * sequence and jumped into the next usec bucket (future time),
274 * then we hope there will be less RPCs per bucket at some
275 * point, and sequence will catch up again */
276 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
277 new_seq = svcpt->scp_hist_seq;
280 req->rq_history_seq = new_seq;
282 list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
286 * Server's incoming request callback
288 void request_in_callback(struct lnet_event *ev)
290 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
291 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
292 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
293 struct ptlrpc_service *service = svcpt->scp_service;
294 struct ptlrpc_request *req;
297 LASSERT (ev->type == LNET_EVENT_PUT ||
298 ev->type == LNET_EVENT_UNLINK);
299 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
300 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
301 rqbd->rqbd_buffer + service->srv_buf_size);
303 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
304 "event type %d, status %d, service %s\n",
305 ev->type, ev->status, service->srv_name);
308 /* If this is the last request message to fit in the
309 * request buffer we can use the request object embedded in
310 * rqbd. Note that if we failed to allocate a request,
311 * we'd have to re-post the rqbd, which we can't do in this
313 req = &rqbd->rqbd_req;
314 memset(req, 0, sizeof (*req));
316 LASSERT (ev->type == LNET_EVENT_PUT);
317 if (ev->status != 0) {
318 /* We moaned above already... */
321 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
323 CERROR("Can't allocate incoming request descriptor: "
324 "Dropping %s RPC from %s\n",
326 libcfs_id2str(ev->initiator));
331 ptlrpc_srv_req_init(req);
332 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
333 * flags are reset and scalars are zero. We only set the message
334 * size to non-zero if this was a successful receive. */
335 req->rq_xid = ev->match_bits;
336 req->rq_reqbuf = ev->md.start + ev->offset;
337 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
338 req->rq_reqdata_len = ev->mlength;
339 ktime_get_real_ts64(&req->rq_arrival_time);
340 /* Multi-Rail: keep track of both initiator and source NID. */
341 req->rq_peer = ev->initiator;
342 req->rq_source = ev->source;
343 req->rq_self = ev->target.nid;
345 req->rq_phase = RQ_PHASE_NEW;
346 if (ev->type == LNET_EVENT_PUT)
347 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
348 req, req->rq_xid, ev->mlength);
350 CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
351 libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source));
353 spin_lock(&svcpt->scp_lock);
355 ptlrpc_req_add_history(svcpt, req);
358 svcpt->scp_nrqbds_posted--;
359 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
360 svcpt->scp_nrqbds_posted);
362 /* Normally, don't complain about 0 buffers posted; LNET won't
363 * drop incoming reqs since we set the portal lazy */
364 if (test_req_buffer_pressure &&
365 ev->type != LNET_EVENT_UNLINK &&
366 svcpt->scp_nrqbds_posted == 0)
367 CWARN("All %s request buffers busy\n",
370 /* req takes over the network's ref on rqbd */
372 /* req takes a ref on rqbd */
373 rqbd->rqbd_refcount++;
376 list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
377 svcpt->scp_nreqs_incoming++;
379 /* NB everything can disappear under us once the request
380 * has been queued and we unlock, so do the wake now... */
381 wake_up(&svcpt->scp_waitq);
383 spin_unlock(&svcpt->scp_lock);
388 * Server's outgoing reply callback
390 void reply_out_callback(struct lnet_event *ev)
392 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
393 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
394 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
397 LASSERT (ev->type == LNET_EVENT_SEND ||
398 ev->type == LNET_EVENT_ACK ||
399 ev->type == LNET_EVENT_UNLINK);
401 if (!rs->rs_difficult) {
402 /* 'Easy' replies have no further processing so I drop the
403 * net's ref on 'rs' */
404 LASSERT (ev->unlinked);
405 ptlrpc_rs_decref(rs);
410 LASSERT (rs->rs_on_net);
413 /* Last network callback. The net's ref on 'rs' stays put
414 * until ptlrpc_handle_rs() is done with it */
415 spin_lock(&svcpt->scp_rep_lock);
416 spin_lock(&rs->rs_lock);
419 if (!rs->rs_no_ack ||
421 rs->rs_export->exp_obd->obd_last_committed ||
422 list_empty(&rs->rs_obd_list))
423 ptlrpc_schedule_difficult_reply(rs);
425 spin_unlock(&rs->rs_lock);
426 spin_unlock(&svcpt->scp_rep_lock);
431 #ifdef HAVE_SERVER_SUPPORT
433 * Server's bulk completion callback
435 void server_bulk_callback(struct lnet_event *ev)
437 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
438 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
441 LASSERT(ev->type == LNET_EVENT_SEND ||
442 ev->type == LNET_EVENT_UNLINK ||
443 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
444 ev->type == LNET_EVENT_ACK) ||
445 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
446 ev->type == LNET_EVENT_REPLY));
448 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
449 "event type %d, status %d, desc %p\n",
450 ev->type, ev->status, desc);
452 spin_lock(&desc->bd_lock);
454 LASSERT(desc->bd_md_count > 0);
456 if ((ev->type == LNET_EVENT_ACK ||
457 ev->type == LNET_EVENT_REPLY) &&
459 /* We heard back from the peer, so even if we get this
460 * before the SENT event (oh yes we can), we know we
461 * read/wrote the peer buffer and how much... */
462 desc->bd_nob_transferred += ev->mlength;
463 desc->bd_sender = ev->sender;
467 desc->bd_failure = 1;
471 /* This is the last callback no matter what... */
472 if (desc->bd_md_count == 0)
473 wake_up(&desc->bd_waitq);
476 spin_unlock(&desc->bd_lock);
481 static void ptlrpc_master_callback(struct lnet_event *ev)
483 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
484 void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
486 /* Honestly, it's best to find out early. */
487 LASSERT (cbid->cbid_arg != LP_POISON);
488 LASSERT (callback == request_out_callback ||
489 callback == reply_in_callback ||
490 callback == client_bulk_callback ||
491 callback == request_in_callback ||
492 callback == reply_out_callback
493 #ifdef HAVE_SERVER_SUPPORT
494 || callback == server_bulk_callback
501 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
502 struct lnet_process_id *peer, lnet_nid_t *self)
505 __u32 best_order = 0;
513 peer->pid = LNET_PID_LUSTRE;
515 /* Choose the matching UUID that's closest */
516 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
517 if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 &&
518 LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid))
521 dist = LNetDist(dst_nid, &src_nid, &order);
525 if (dist == 0) { /* local! use loopback LND */
526 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
533 (dist == best_dist && order < best_order)) {
543 CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
547 void ptlrpc_ni_fini(void)
549 wait_queue_head_t waitq;
550 struct l_wait_info lwi;
554 /* Wait for the event queue to become idle since there may still be
555 * messages in flight with pending events (i.e. the fire-and-forget
556 * messages == client requests and "non-difficult" server
559 for (retries = 0;; retries++) {
560 rc = LNetEQFree(ptlrpc_eq_h);
571 CWARN("Event queue still busy\n");
574 init_waitqueue_head(&waitq);
575 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
576 l_wait_event(waitq, 0, &lwi);
583 lnet_pid_t ptl_get_pid(void)
585 return LNET_PID_LUSTRE;
588 int ptlrpc_ni_init(void)
594 CDEBUG(D_NET, "My pid is: %x\n", pid);
596 /* We're not passing any limits yet... */
597 rc = LNetNIInit(pid);
599 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
603 /* CAVEAT EMPTOR: how we process portals events is _radically_
604 * different depending on... */
605 /* kernel LNet calls our master callback when there are new event,
606 * because we are guaranteed to get every event via callback,
607 * so we just set EQ size to 0 to avoid overhread of serializing
608 * enqueue/dequeue operations in LNet. */
609 rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
613 CERROR ("Failed to allocate event queue: %d\n", rc);
620 int ptlrpc_init_portals(void)
622 int rc = ptlrpc_ni_init();
625 CERROR("network initialisation failed\n");
628 rc = ptlrpcd_addref();
632 CERROR("rpcd initialisation failed\n");
637 void ptlrpc_exit_portals(void)