Whamcloud - gitweb
LU-13005 lnet: remove 'eq' from names, use 'handler'
[fs/lustre-release.git] / lustre / ptlrpc / events.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 #define DEBUG_SUBSYSTEM S_RPC
34
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"
42
43 lnet_handler_t ptlrpc_handler;
44 struct percpu_ref ptlrpc_pending;
45
46 /*
47  *  Client's outgoing request callback
48  */
49 void request_out_callback(struct lnet_event *ev)
50 {
51         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
52         struct ptlrpc_request *req = cbid->cbid_arg;
53         bool                   wakeup = false;
54         ENTRY;
55
56         LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK);
57         LASSERT(ev->unlinked);
58
59         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
60
61         sptlrpc_request_out_callback(req);
62
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)
68                 wakeup = true;
69
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...  */
73                 req->rq_net_err = 1;
74                 wakeup = true;
75         }
76
77         if (wakeup)
78                 ptlrpc_client_wake_req(req);
79
80         spin_unlock(&req->rq_lock);
81
82         ptlrpc_req_finished(req);
83         EXIT;
84 }
85
86 /*
87  * Client's incoming reply callback
88  */
89 void reply_in_callback(struct lnet_event *ev)
90 {
91         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
92         struct ptlrpc_request *req = cbid->cbid_arg;
93         ENTRY;
94
95         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
96
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);
103
104         spin_lock(&req->rq_lock);
105
106         req->rq_receiving_reply = 0;
107         req->rq_early = 0;
108         if (ev->unlinked)
109                 req->rq_reply_unlinked = 1;
110
111         if (ev->status)
112                 goto out_wake;
113
114         if (ev->type == LNET_EVENT_UNLINK) {
115                 LASSERT(ev->unlinked);
116                 DEBUG_REQ(D_NET, req, "unlink");
117                 goto out_wake;
118         }
119
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;
124                 req->rq_replied = 1;
125                 req->rq_status = -EOVERFLOW;
126                 req->rq_nob_received = ev->rlength + ev->offset;
127                 goto out_wake;
128         }
129
130         if ((ev->offset == 0) &&
131             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
132                 /* Early reply */
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);
137
138                 req->rq_early_count++; /* number received, client side */
139
140                 /* already got the real reply or buffers are already unlinked */
141                 if (req->rq_replied ||
142                     req->rq_reply_unlinked == 1)
143                         goto out_wake;
144
145                 req->rq_early = 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;
150         } else {
151                 /* Real reply */
152                 req->rq_rep_swab_mask = 0;
153                 req->rq_replied = 1;
154                 /* Got reply, no resend required */
155                 req->rq_resend = 0;
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);
164         }
165
166         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
167                 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
168
169 out_wake:
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);
174         EXIT;
175 }
176
177 /*
178  * Client's bulk has been written/read
179  */
180 void client_bulk_callback(struct lnet_event *ev)
181 {
182         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
183         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
184         struct ptlrpc_request   *req;
185         ENTRY;
186
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);
193
194         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
195                 ev->status = -EIO;
196
197         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
198                 ev->status = -EIO;
199
200         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
201                "event type %d, status %d, desc %p\n",
202                ev->type, ev->status, desc);
203
204         spin_lock(&desc->bd_lock);
205         req = desc->bd_req;
206         LASSERT(desc->bd_md_count > 0);
207         desc->bd_md_count--;
208
209         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
210                 desc->bd_nob_transferred += ev->mlength;
211                 desc->bd_sender = ev->sender;
212         } else {
213                 /* start reconnect and resend if network error hit */
214                 spin_lock(&req->rq_lock);
215                 req->rq_net_err = 1;
216                 spin_unlock(&req->rq_lock);
217         }
218
219         if (ev->status != 0)
220                 desc->bd_failure = 1;
221
222         /* NB don't unlock till after wakeup; desc can disappear under us
223          * otherwise */
224         if (desc->bd_md_count == 0)
225                 ptlrpc_client_wake_req(desc->bd_req);
226
227         spin_unlock(&desc->bd_lock);
228         EXIT;
229 }
230
231 /*
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.
236  *
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  * ----------------------------------------------------
243  *
244  * it might not be precise but should be good enough.
245  */
246
247 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
248
249 #define REQS_SEC_SHIFT          32
250 #define REQS_USEC_SHIFT         16
251 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
252
253 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
254                                    struct ptlrpc_request *req)
255 {
256         u64 sec = req->rq_arrival_time.tv_sec;
257         u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
258         u64 new_seq;
259
260         /* set sequence ID for request and add it to history list,
261          * it must be called with hold svcpt::scp_lock */
262
263         new_seq = (sec << REQS_SEC_SHIFT) |
264                   (usec << REQS_USEC_SHIFT) |
265                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
266
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;
271         } else {
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;
280         }
281
282         req->rq_history_seq = new_seq;
283
284         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
285 }
286
287 /*
288  * Server's incoming request callback
289  */
290 void request_in_callback(struct lnet_event *ev)
291 {
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;
297         ENTRY;
298
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);
304
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);
308
309         if (ev->unlinked) {
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
314                  * context. */
315                 req = &rqbd->rqbd_req;
316                 memset(req, 0, sizeof (*req));
317         } else {
318                 LASSERT (ev->type == LNET_EVENT_PUT);
319                 if (ev->status != 0) {
320                         /* We moaned above already... */
321                         return;
322                 }
323                 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
324                 if (req == NULL) {
325                         CERROR("Can't allocate incoming request descriptor: "
326                                "Dropping %s RPC from %s\n",
327                                service->srv_name,
328                                libcfs_id2str(ev->initiator));
329                         return;
330                 }
331         }
332
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;
346         req->rq_rqbd = rqbd;
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);
351
352         CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
353                 libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source));
354
355         spin_lock(&svcpt->scp_lock);
356
357         ptlrpc_req_add_history(svcpt, req);
358
359         if (ev->unlinked) {
360                 svcpt->scp_nrqbds_posted--;
361                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
362                        svcpt->scp_nrqbds_posted);
363
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",
370                               service->srv_name);
371
372                 /* req takes over the network's ref on rqbd */
373         } else {
374                 /* req takes a ref on rqbd */
375                 rqbd->rqbd_refcount++;
376         }
377
378         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
379         svcpt->scp_nreqs_incoming++;
380
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);
384
385         spin_unlock(&svcpt->scp_lock);
386         EXIT;
387 }
388
389 /*
390  *  Server's outgoing reply callback
391  */
392 void reply_out_callback(struct lnet_event *ev)
393 {
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;
397         ENTRY;
398
399         LASSERT (ev->type == LNET_EVENT_SEND ||
400                  ev->type == LNET_EVENT_ACK ||
401                  ev->type == LNET_EVENT_UNLINK);
402
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);
408                 EXIT;
409                 return;
410         }
411
412         LASSERT (rs->rs_on_net);
413
414         if (ev->unlinked) {
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);
419
420                 rs->rs_on_net = 0;
421                 if (!rs->rs_no_ack ||
422                     rs->rs_transno <=
423                     rs->rs_export->exp_obd->obd_last_committed ||
424                     list_empty(&rs->rs_obd_list))
425                         ptlrpc_schedule_difficult_reply(rs);
426
427                 spin_unlock(&rs->rs_lock);
428                 spin_unlock(&svcpt->scp_rep_lock);
429         }
430         EXIT;
431 }
432
433 #ifdef HAVE_SERVER_SUPPORT
434 /*
435  * Server's bulk completion callback
436  */
437 void server_bulk_callback(struct lnet_event *ev)
438 {
439         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
440         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
441         ENTRY;
442
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));
449
450         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
451                "event type %d, status %d, desc %p\n",
452                ev->type, ev->status, desc);
453
454         spin_lock(&desc->bd_lock);
455
456         LASSERT(desc->bd_md_count > 0);
457
458         if ((ev->type == LNET_EVENT_ACK ||
459              ev->type == LNET_EVENT_REPLY) &&
460             ev->status == 0) {
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;
466         }
467
468         if (ev->status != 0)
469                 desc->bd_failure = 1;
470
471         if (ev->unlinked) {
472                 desc->bd_md_count--;
473                 /* This is the last callback no matter what... */
474                 if (desc->bd_md_count == 0)
475                         wake_up(&desc->bd_waitq);
476         }
477
478         spin_unlock(&desc->bd_lock);
479         EXIT;
480 }
481 #endif
482
483 static void ptlrpc_master_callback(struct lnet_event *ev)
484 {
485         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
486         void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
487
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
497 #endif
498                 );
499
500         callback(ev);
501         if (ev->unlinked)
502                 percpu_ref_put(&ptlrpc_pending);
503 }
504
505 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
506                         struct lnet_process_id *peer, lnet_nid_t *self)
507 {
508         int               best_dist = 0;
509         __u32             best_order = 0;
510         int               count = 0;
511         int               rc = -ENOENT;
512         int               dist;
513         __u32             order;
514         lnet_nid_t        dst_nid;
515         lnet_nid_t        src_nid;
516
517         peer->pid = LNET_PID_LUSTRE;
518
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))
523                         continue;
524
525                 dist = LNetDist(dst_nid, &src_nid, &order);
526                 if (dist < 0)
527                         continue;
528
529                 if (dist == 0) {                /* local! use loopback LND */
530                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
531                         rc = 0;
532                         break;
533                 }
534
535                 if (rc < 0 ||
536                     dist < best_dist ||
537                     (dist == best_dist && order < best_order)) {
538                         best_dist = dist;
539                         best_order = order;
540
541                         peer->nid = dst_nid;
542                         *self = src_nid;
543                         rc = 0;
544                 }
545         }
546
547         CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
548         return rc;
549 }
550
551 static struct completion ptlrpc_done;
552
553 static void ptlrpc_release(struct percpu_ref *ref)
554 {
555         complete(&ptlrpc_done);
556 }
557
558 static void ptlrpc_ni_fini(void)
559 {
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
563          * replies */
564
565         init_completion(&ptlrpc_done);
566         percpu_ref_kill(&ptlrpc_pending);
567         wait_for_completion(&ptlrpc_done);
568
569         LNetNIFini();
570 }
571
572 lnet_pid_t ptl_get_pid(void)
573 {
574         return LNET_PID_LUSTRE;
575 }
576
577 int ptlrpc_ni_init(void)
578 {
579         int rc;
580         lnet_pid_t pid;
581
582         pid = ptl_get_pid();
583         CDEBUG(D_NET, "My pid is: %x\n", pid);
584
585         /* We're not passing any limits yet... */
586         rc = LNetNIInit(pid);
587         if (rc < 0) {
588                 CDEBUG(D_NET, "ptlrpc: Can't init network interface: rc = %d\n",
589                        rc);
590                 return rc;
591         }
592
593         rc = percpu_ref_init(&ptlrpc_pending, ptlrpc_release, 0, GFP_KERNEL);
594         if (rc) {
595                 CERROR("ptlrpc: Can't init percpu refcount: rc = %d\n", rc);
596                 return rc;
597         }
598         /* CAVEAT EMPTOR: how we process portals events is _radically_
599          * different depending on...
600          */
601         /* kernel LNet calls our master callback when there are new event,
602          * because we are guaranteed to get every event via callback,
603          * so we just set EQ size to 0 to avoid overhread of serializing
604          * enqueue/dequeue operations in LNet. */
605         ptlrpc_handler = ptlrpc_master_callback;
606         return 0;
607 }
608
609 int ptlrpc_init_portals(void)
610 {
611         int   rc = ptlrpc_ni_init();
612
613         if (rc != 0) {
614                 CERROR("network initialisation failed\n");
615                 return rc;
616         }
617         rc = ptlrpcd_addref();
618         if (rc == 0)
619                 return 0;
620
621         CERROR("rpcd initialisation failed\n");
622         ptlrpc_ni_fini();
623         return rc;
624 }
625
626 void ptlrpc_exit_portals(void)
627 {
628         ptlrpcd_decref();
629         ptlrpc_ni_fini();
630 }