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/
32 #define DEBUG_SUBSYSTEM S_RPC
34 #include <libcfs/libcfs.h>
35 #include <linux/kernel.h>
36 #include <linux/delay.h>
37 #include <obd_class.h>
38 #include <lustre_net.h>
39 #include <lustre_sec.h>
40 #include "ptlrpc_internal.h"
42 lnet_handler_t ptlrpc_handler;
43 struct percpu_ref ptlrpc_pending;
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 /* Do not update imp_next_ping for connection request */
61 if (lustre_msg_get_opc(req->rq_reqmsg) !=
62 req->rq_import->imp_connect_op)
63 ptlrpc_pinger_sending_on_import(req->rq_import);
65 sptlrpc_request_out_callback(req);
67 spin_lock(&req->rq_lock);
68 req->rq_real_sent = ktime_get_real_seconds();
69 req->rq_req_unlinked = 1;
70 /* reply_in_callback happened before request_out_callback? */
71 if (req->rq_reply_unlinked)
74 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
75 /* Failed send: make it seem like the reply timed out, just
76 * like failing sends in client.c does currently... */
82 ptlrpc_client_wake_req(req);
84 spin_unlock(&req->rq_lock);
86 ptlrpc_req_finished(req);
91 * Client's incoming reply callback
93 void reply_in_callback(struct lnet_event *ev)
95 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
96 struct ptlrpc_request *req = cbid->cbid_arg;
99 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
101 LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
102 LASSERT(ev->md_start == req->rq_repbuf);
103 LASSERT(ev->offset + ev->mlength <= req->rq_repbuf_len);
104 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
105 * for adaptive timeouts' early reply.
107 LASSERT((ev->md_options & LNET_MD_MANAGE_REMOTE) != 0);
109 spin_lock(&req->rq_lock);
111 req->rq_receiving_reply = 0;
114 req->rq_reply_unlinked = 1;
119 if (ev->type == LNET_EVENT_UNLINK) {
120 LASSERT(ev->unlinked);
121 DEBUG_REQ(D_NET, req, "unlink");
125 if (ev->mlength < ev->rlength ) {
126 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
127 req->rq_replen, ev->rlength, ev->offset);
128 req->rq_reply_truncated = 1;
130 req->rq_status = -EOVERFLOW;
131 req->rq_nob_received = ev->rlength + ev->offset;
135 if ((ev->offset == 0) &&
136 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
138 DEBUG_REQ(D_ADAPTTO, req,
139 "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
140 ev->mlength, ev->offset,
141 req->rq_replen, req->rq_replied, ev->unlinked);
143 req->rq_early_count++; /* number received, client side */
145 /* already got the real reply or buffers are already unlinked */
146 if (req->rq_replied ||
147 req->rq_reply_unlinked == 1)
151 req->rq_reply_off = ev->offset;
152 req->rq_nob_received = ev->mlength;
153 /* And we're still receiving */
154 req->rq_receiving_reply = 1;
157 req->rq_rep_swab_mask = 0;
159 /* Got reply, no resend required */
161 req->rq_reply_off = ev->offset;
162 req->rq_nob_received = ev->mlength;
163 /* LNetMDUnlink can't be called under the LNET_LOCK,
164 so we must unlink in ptlrpc_unregister_reply */
165 DEBUG_REQ(D_INFO, req,
166 "reply in flags=%x mlen=%u offset=%d replen=%d",
167 lustre_msg_get_flags(req->rq_reqmsg),
168 ev->mlength, ev->offset, req->rq_replen);
171 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
172 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
174 if (req->rq_xid > req->rq_import->imp_highest_replied_xid)
175 req->rq_import->imp_highest_replied_xid = req->rq_xid;
178 /* NB don't unlock till after wakeup; req can disappear under us
179 * since we don't have our own ref */
180 ptlrpc_client_wake_req(req);
181 spin_unlock(&req->rq_lock);
186 * Client's bulk has been written/read
188 void client_bulk_callback(struct lnet_event *ev)
190 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
191 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
192 struct ptlrpc_request *req;
195 LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
196 ev->type == LNET_EVENT_PUT) ||
197 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
198 ev->type == LNET_EVENT_GET) ||
199 ev->type == LNET_EVENT_UNLINK);
200 LASSERT(ev->unlinked);
202 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
205 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
208 CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
209 "event type %d, status %d, desc %p\n",
210 ev->type, ev->status, desc);
212 spin_lock(&desc->bd_lock);
214 LASSERT(desc->bd_refs > 0);
217 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
218 desc->bd_nob_transferred += ev->mlength;
219 desc->bd_sender = ev->sender;
221 /* start reconnect and resend if network error hit */
222 spin_lock(&req->rq_lock);
224 spin_unlock(&req->rq_lock);
225 desc->bd_failure = 1;
229 /* NB don't unlock till after wakeup; desc can disappear under us
231 if (desc->bd_refs == 0)
232 ptlrpc_client_wake_req(desc->bd_req);
234 spin_unlock(&desc->bd_lock);
239 * We will have percpt request history list for ptlrpc service in upcoming
240 * patches because we don't want to be serialized by current per-service
241 * history operations. So we require history ID can (somehow) show arriving
242 * order w/o grabbing global lock, and user can sort them in userspace.
244 * This is how we generate history ID for ptlrpc_request:
245 * ----------------------------------------------------
246 * | 32 bits | 16 bits | (16 - X)bits | X bits |
247 * ----------------------------------------------------
248 * | seconds | usec / 16 | sequence | CPT id |
249 * ----------------------------------------------------
251 * it might not be precise but should be good enough.
254 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
256 #define REQS_SEC_SHIFT 32
257 #define REQS_USEC_SHIFT 16
258 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
260 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
261 struct ptlrpc_request *req)
263 u64 sec = req->rq_arrival_time.tv_sec;
264 u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
267 /* set sequence ID for request and add it to history list,
268 * it must be called with hold svcpt::scp_lock */
270 new_seq = (sec << REQS_SEC_SHIFT) |
271 (usec << REQS_USEC_SHIFT) |
272 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
274 if (new_seq > svcpt->scp_hist_seq) {
275 /* This handles the initial case of scp_hist_seq == 0 or
276 * we just jumped into a new time window */
277 svcpt->scp_hist_seq = new_seq;
279 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
280 /* NB: increase sequence number in current usec bucket,
281 * however, it's possible that we used up all bits for
282 * sequence and jumped into the next usec bucket (future time),
283 * then we hope there will be less RPCs per bucket at some
284 * point, and sequence will catch up again */
285 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
286 new_seq = svcpt->scp_hist_seq;
289 req->rq_history_seq = new_seq;
291 list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
295 * Server's incoming request callback
297 void request_in_callback(struct lnet_event *ev)
299 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
300 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
301 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
302 struct ptlrpc_service *service = svcpt->scp_service;
303 struct ptlrpc_request *req;
306 LASSERT(ev->type == LNET_EVENT_PUT ||
307 ev->type == LNET_EVENT_UNLINK);
308 LASSERT((char *)ev->md_start >= rqbd->rqbd_buffer);
309 LASSERT((char *)ev->md_start + ev->offset + ev->mlength <=
310 rqbd->rqbd_buffer + service->srv_buf_size);
312 CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
313 "event type %d, status %d, service %s\n",
314 ev->type, ev->status, service->srv_name);
317 /* If this is the last request message to fit in the
318 * request buffer we can use the request object embedded in
319 * rqbd. Note that if we failed to allocate a request,
320 * we'd have to re-post the rqbd, which we can't do in this
323 req = &rqbd->rqbd_req;
324 memset(req, 0, sizeof(*req));
326 LASSERT(ev->type == LNET_EVENT_PUT);
327 if (ev->status != 0) /* We moaned above already... */
329 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
331 CERROR("Can't allocate incoming request descriptor: Dropping %s RPC from %s\n",
333 libcfs_idstr(&ev->initiator));
338 ptlrpc_srv_req_init(req);
339 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
340 * flags are reset and scalars are zero. We only set the message
341 * size to non-zero if this was a successful receive. */
342 req->rq_xid = ev->match_bits;
343 req->rq_reqbuf = ev->md_start + ev->offset;
344 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
345 req->rq_reqdata_len = ev->mlength;
346 ktime_get_real_ts64(&req->rq_arrival_time);
347 /* Multi-Rail: keep track of both initiator and source NID. */
348 req->rq_peer = ev->initiator;
349 req->rq_source = ev->source;
350 req->rq_self = ev->target.nid;
352 req->rq_phase = RQ_PHASE_NEW;
353 if (ev->type == LNET_EVENT_PUT)
354 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
355 req, req->rq_xid, ev->mlength);
357 CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
358 libcfs_idstr(&req->rq_peer), libcfs_idstr(&req->rq_source));
360 spin_lock(&svcpt->scp_lock);
362 ptlrpc_req_add_history(svcpt, req);
365 svcpt->scp_nrqbds_posted--;
366 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
367 svcpt->scp_nrqbds_posted);
369 /* Normally, don't complain about 0 buffers posted; LNET won't
370 * drop incoming reqs since we set the portal lazy */
371 if (test_req_buffer_pressure &&
372 ev->type != LNET_EVENT_UNLINK &&
373 svcpt->scp_nrqbds_posted == 0)
374 CWARN("All %s request buffers busy\n",
377 /* req takes over the network's ref on rqbd */
379 /* req takes a ref on rqbd */
380 rqbd->rqbd_refcount++;
383 list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
384 svcpt->scp_nreqs_incoming++;
386 /* NB everything can disappear under us once the request
387 * has been queued and we unlock, so do the wake now... */
388 wake_up(&svcpt->scp_waitq);
390 spin_unlock(&svcpt->scp_lock);
395 * Server's outgoing reply callback
397 void reply_out_callback(struct lnet_event *ev)
399 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
400 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
401 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
402 bool need_schedule = false;
406 LASSERT(ev->type == LNET_EVENT_SEND ||
407 ev->type == LNET_EVENT_ACK ||
408 ev->type == LNET_EVENT_UNLINK);
410 if (!rs->rs_difficult) {
411 /* 'Easy' replies have no further processing so I drop the
414 LASSERT(ev->unlinked);
415 ptlrpc_rs_decref(rs);
420 if (ev->type == LNET_EVENT_SEND) {
421 spin_lock(&rs->rs_lock);
423 /* If transaction was committed before the SEND, and the ACK
424 * is lost, then we need to schedule so ptlrpc_hr can unlink
428 need_schedule = true;
429 spin_unlock(&rs->rs_lock);
432 if (ev->unlinked || need_schedule) {
433 LASSERT(rs->rs_sent);
435 /* Last network callback. The net's ref on 'rs' stays put
436 * until ptlrpc_handle_rs() is done with it
438 spin_lock(&svcpt->scp_rep_lock);
439 spin_lock(&rs->rs_lock);
441 rs->rs_unlinked = ev->unlinked;
442 if (!rs->rs_no_ack ||
444 rs->rs_export->exp_obd->obd_last_committed ||
445 list_empty(&rs->rs_obd_list))
446 ptlrpc_schedule_difficult_reply(rs);
448 spin_unlock(&rs->rs_lock);
449 spin_unlock(&svcpt->scp_rep_lock);
454 #ifdef HAVE_SERVER_SUPPORT
456 * Server's bulk completion callback
458 void server_bulk_callback(struct lnet_event *ev)
460 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
461 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
464 LASSERT(ev->type == LNET_EVENT_SEND ||
465 ev->type == LNET_EVENT_UNLINK ||
466 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
467 ev->type == LNET_EVENT_ACK) ||
468 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
469 ev->type == LNET_EVENT_REPLY));
471 CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
472 "event type %d, status %d, desc %p\n",
473 ev->type, ev->status, desc);
475 spin_lock(&desc->bd_lock);
477 LASSERT(desc->bd_refs > 0);
479 if ((ev->type == LNET_EVENT_ACK ||
480 ev->type == LNET_EVENT_REPLY) &&
482 /* We heard back from the peer, so even if we get this
483 * before the SENT event (oh yes we can), we know we
484 * read/wrote the peer buffer and how much... */
485 desc->bd_nob_transferred += ev->mlength;
486 desc->bd_sender = ev->sender;
490 desc->bd_failure = 1;
494 /* This is the last callback no matter what... */
495 if (desc->bd_refs == 0)
496 wake_up(&desc->bd_waitq);
499 spin_unlock(&desc->bd_lock);
504 static void ptlrpc_master_callback(struct lnet_event *ev)
506 struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
507 void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
509 /* Honestly, it's best to find out early. */
510 LASSERT(cbid->cbid_arg != LP_POISON);
511 LASSERT(callback == request_out_callback ||
512 callback == reply_in_callback ||
513 callback == client_bulk_callback ||
514 callback == request_in_callback ||
515 callback == reply_out_callback
516 #ifdef HAVE_SERVER_SUPPORT
517 || callback == server_bulk_callback
523 percpu_ref_put(&ptlrpc_pending);
526 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
527 struct lnet_processid *peer,
528 struct lnet_nid *self,
532 __u32 best_order = 0;
537 struct lnet_nid dst_nid;
538 struct lnet_nid src_nid;
540 peer->pid = LNET_PID_LUSTRE;
542 /* Choose the matching UUID that's closest */
543 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
544 if (refnet != LNET_NET_ANY &&
545 LNET_NID_NET(&dst_nid) != refnet)
548 dist = LNetDist(&dst_nid, &src_nid, &order);
552 if (dist == 0) { /* local! use loopback LND */
553 lnet_nid4_to_nid(LNET_NID_LO_0, self);
561 (dist == best_dist && order < best_order)) {
571 CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_idstr(peer));
575 static struct completion ptlrpc_done;
577 static void ptlrpc_release(struct percpu_ref *ref)
579 complete(&ptlrpc_done);
582 static void ptlrpc_ni_fini(void)
584 /* Wait for the event queue to become idle since there may still be
585 * messages in flight with pending events (i.e. the fire-and-forget
586 * messages == client requests and "non-difficult" server
589 init_completion(&ptlrpc_done);
590 percpu_ref_kill(&ptlrpc_pending);
591 wait_for_completion(&ptlrpc_done);
593 lnet_assert_handler_unused(ptlrpc_handler);
597 lnet_pid_t ptl_get_pid(void)
599 return LNET_PID_LUSTRE;
602 int ptlrpc_ni_init(void)
608 CDEBUG(D_NET, "My pid is: %x\n", pid);
610 /* We're not passing any limits yet... */
611 rc = LNetNIInit(pid);
613 CDEBUG(D_NET, "ptlrpc: Can't init network interface: rc = %d\n",
618 rc = percpu_ref_init(&ptlrpc_pending, ptlrpc_release, 0, GFP_KERNEL);
620 CERROR("ptlrpc: Can't init percpu refcount: rc = %d\n", rc);
623 /* CAVEAT EMPTOR: how we process portals events is _radically_
624 * different depending on...
626 /* kernel LNet calls our master callback when there are new event,
627 * because we are guaranteed to get every event via callback,
628 * so we just set EQ size to 0 to avoid overhread of serializing
629 * enqueue/dequeue operations in LNet. */
630 ptlrpc_handler = ptlrpc_master_callback;
634 int ptlrpc_init_portals(void)
636 int rc = ptlrpc_ni_init();
639 CERROR("network initialisation failed\n");
642 rc = ptlrpcd_addref();
646 CERROR("rpcd initialisation failed\n");
651 void ptlrpc_exit_portals(void)