Whamcloud - gitweb
ca4b0e613e94aef06c20dd635896146ba7b9e907
[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.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2012, 2013, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 #define DEBUG_SUBSYSTEM S_RPC
38
39 #include <libcfs/libcfs.h>
40 #include <linux/kernel.h>
41 #include <obd_class.h>
42 #include <lustre_net.h>
43 #include <lustre_sec.h>
44 #include "ptlrpc_internal.h"
45
46 lnet_handle_eq_t   ptlrpc_eq_h;
47
48 /*
49  *  Client's outgoing request callback
50  */
51 void request_out_callback(lnet_event_t *ev)
52 {
53         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
54         struct ptlrpc_request *req = cbid->cbid_arg;
55         ENTRY;
56
57         LASSERT (ev->type == LNET_EVENT_SEND ||
58                  ev->type == LNET_EVENT_UNLINK);
59         LASSERT (ev->unlinked);
60
61         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
62
63         sptlrpc_request_out_callback(req);
64         spin_lock(&req->rq_lock);
65         req->rq_real_sent = cfs_time_current_sec();
66         if (ev->unlinked)
67                 req->rq_req_unlink = 0;
68
69         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
70
71                 /* Failed send: make it seem like the reply timed out, just
72                  * like failing sends in client.c does currently...  */
73
74                 req->rq_net_err = 1;
75                 ptlrpc_client_wake_req(req);
76         }
77         spin_unlock(&req->rq_lock);
78
79         ptlrpc_req_finished(req);
80
81         EXIT;
82 }
83
84 /*
85  * Client's incoming reply callback
86  */
87 void reply_in_callback(lnet_event_t *ev)
88 {
89         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
90         struct ptlrpc_request *req = cbid->cbid_arg;
91         ENTRY;
92
93         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
94
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);
101
102         spin_lock(&req->rq_lock);
103
104         req->rq_receiving_reply = 0;
105         req->rq_early = 0;
106         if (ev->unlinked)
107                 req->rq_reply_unlink = 0;
108
109         if (ev->status)
110                 goto out_wake;
111
112         if (ev->type == LNET_EVENT_UNLINK) {
113                 LASSERT(ev->unlinked);
114                 DEBUG_REQ(D_NET, req, "unlink");
115                 goto out_wake;
116         }
117
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_truncate = 1;
122                 req->rq_replied = 1;
123                 req->rq_status = -EOVERFLOW;
124                 req->rq_nob_received = ev->rlength + ev->offset;
125                 goto out_wake;
126         }
127
128         if ((ev->offset == 0) &&
129             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
130                 /* Early reply */
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);
135
136                 req->rq_early_count++; /* number received, client side */
137
138                 if (req->rq_replied)   /* already got the real reply */
139                         goto out_wake;
140
141                 req->rq_early = 1;
142                 req->rq_reply_off = ev->offset;
143                 req->rq_nob_received = ev->mlength;
144                 /* And we're still receiving */
145                 req->rq_receiving_reply = 1;
146         } else {
147                 /* Real reply */
148                 req->rq_rep_swab_mask = 0;
149                 req->rq_replied = 1;
150                 /* Got reply, no resend required */
151                 req->rq_resend = 0;
152                 req->rq_reply_off = ev->offset;
153                 req->rq_nob_received = ev->mlength;
154                 /* LNetMDUnlink can't be called under the LNET_LOCK,
155                    so we must unlink in ptlrpc_unregister_reply */
156                 DEBUG_REQ(D_INFO, req,
157                           "reply in flags=%x mlen=%u offset=%d replen=%d",
158                           lustre_msg_get_flags(req->rq_reqmsg),
159                           ev->mlength, ev->offset, req->rq_replen);
160         }
161
162         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
163
164 out_wake:
165         /* NB don't unlock till after wakeup; req can disappear under us
166          * since we don't have our own ref */
167         ptlrpc_client_wake_req(req);
168         spin_unlock(&req->rq_lock);
169         EXIT;
170 }
171
172 /*
173  * Client's bulk has been written/read
174  */
175 void client_bulk_callback (lnet_event_t *ev)
176 {
177         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
178         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
179         struct ptlrpc_request   *req;
180         ENTRY;
181
182         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
183                   ev->type == LNET_EVENT_PUT) ||
184                  (desc->bd_type == BULK_GET_SOURCE &&
185                   ev->type == LNET_EVENT_GET) ||
186                  ev->type == LNET_EVENT_UNLINK);
187         LASSERT (ev->unlinked);
188
189         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
190                 ev->status = -EIO;
191
192         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
193                 ev->status = -EIO;
194
195         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
196                "event type %d, status %d, desc %p\n",
197                ev->type, ev->status, desc);
198
199         spin_lock(&desc->bd_lock);
200         req = desc->bd_req;
201         LASSERT(desc->bd_md_count > 0);
202         desc->bd_md_count--;
203
204         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
205                 desc->bd_nob_transferred += ev->mlength;
206                 desc->bd_sender = ev->sender;
207         } else {
208                 /* start reconnect and resend if network error hit */
209                 spin_lock(&req->rq_lock);
210                 req->rq_net_err = 1;
211                 spin_unlock(&req->rq_lock);
212         }
213
214         if (ev->status != 0)
215                 desc->bd_failure = 1;
216
217         /* NB don't unlock till after wakeup; desc can disappear under us
218          * otherwise */
219         if (desc->bd_md_count == 0)
220                 ptlrpc_client_wake_req(desc->bd_req);
221
222         spin_unlock(&desc->bd_lock);
223         EXIT;
224 }
225
226 /*
227  * We will have percpt request history list for ptlrpc service in upcoming
228  * patches because we don't want to be serialized by current per-service
229  * history operations. So we require history ID can (somehow) show arriving
230  * order w/o grabbing global lock, and user can sort them in userspace.
231  *
232  * This is how we generate history ID for ptlrpc_request:
233  * ----------------------------------------------------
234  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
235  * ----------------------------------------------------
236  * |  seconds  | usec / 16 |   sequence    | CPT id   |
237  * ----------------------------------------------------
238  *
239  * it might not be precise but should be good enough.
240  */
241
242 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
243
244 #define REQS_SEC_SHIFT          32
245 #define REQS_USEC_SHIFT         16
246 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
247
248 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
249                                    struct ptlrpc_request *req)
250 {
251         __u64   sec = req->rq_arrival_time.tv_sec;
252         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
253         __u64   new_seq;
254
255         /* set sequence ID for request and add it to history list,
256          * it must be called with hold svcpt::scp_lock */
257
258         new_seq = (sec << REQS_SEC_SHIFT) |
259                   (usec << REQS_USEC_SHIFT) |
260                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
261
262         if (new_seq > svcpt->scp_hist_seq) {
263                 /* This handles the initial case of scp_hist_seq == 0 or
264                  * we just jumped into a new time window */
265                 svcpt->scp_hist_seq = new_seq;
266         } else {
267                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
268                 /* NB: increase sequence number in current usec bucket,
269                  * however, it's possible that we used up all bits for
270                  * sequence and jumped into the next usec bucket (future time),
271                  * then we hope there will be less RPCs per bucket at some
272                  * point, and sequence will catch up again */
273                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
274                 new_seq = svcpt->scp_hist_seq;
275         }
276
277         req->rq_history_seq = new_seq;
278
279         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
280 }
281
282 /*
283  * Server's incoming request callback
284  */
285 void request_in_callback(lnet_event_t *ev)
286 {
287         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
288         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
289         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
290         struct ptlrpc_service             *service = svcpt->scp_service;
291         struct ptlrpc_request             *req;
292         ENTRY;
293
294         LASSERT (ev->type == LNET_EVENT_PUT ||
295                  ev->type == LNET_EVENT_UNLINK);
296         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
297         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
298                  rqbd->rqbd_buffer + service->srv_buf_size);
299
300         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
301                "event type %d, status %d, service %s\n",
302                ev->type, ev->status, service->srv_name);
303
304         if (ev->unlinked) {
305                 /* If this is the last request message to fit in the
306                  * request buffer we can use the request object embedded in
307                  * rqbd.  Note that if we failed to allocate a request,
308                  * we'd have to re-post the rqbd, which we can't do in this
309                  * context. */
310                 req = &rqbd->rqbd_req;
311                 memset(req, 0, sizeof (*req));
312         } else {
313                 LASSERT (ev->type == LNET_EVENT_PUT);
314                 if (ev->status != 0) {
315                         /* We moaned above already... */
316                         return;
317                 }
318                 req = ptlrpc_request_cache_alloc(ALLOC_ATOMIC_TRY);
319                 if (req == NULL) {
320                         CERROR("Can't allocate incoming request descriptor: "
321                                "Dropping %s RPC from %s\n",
322                                service->srv_name,
323                                libcfs_id2str(ev->initiator));
324                         return;
325                 }
326         }
327
328         ptlrpc_srv_req_init(req);
329         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
330          * flags are reset and scalars are zero.  We only set the message
331          * size to non-zero if this was a successful receive. */
332         req->rq_xid = ev->match_bits;
333         req->rq_reqbuf = ev->md.start + ev->offset;
334         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
335                 req->rq_reqdata_len = ev->mlength;
336         do_gettimeofday(&req->rq_arrival_time);
337         req->rq_peer = ev->initiator;
338         req->rq_self = ev->target.nid;
339         req->rq_rqbd = rqbd;
340         req->rq_phase = RQ_PHASE_NEW;
341         if (ev->type == LNET_EVENT_PUT)
342                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
343                        req, req->rq_xid, ev->mlength);
344
345         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
346
347         spin_lock(&svcpt->scp_lock);
348
349         ptlrpc_req_add_history(svcpt, req);
350
351         if (ev->unlinked) {
352                 svcpt->scp_nrqbds_posted--;
353                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
354                        svcpt->scp_nrqbds_posted);
355
356                 /* Normally, don't complain about 0 buffers posted; LNET won't
357                  * drop incoming reqs since we set the portal lazy */
358                 if (test_req_buffer_pressure &&
359                     ev->type != LNET_EVENT_UNLINK &&
360                     svcpt->scp_nrqbds_posted == 0)
361                         CWARN("All %s request buffers busy\n",
362                               service->srv_name);
363
364                 /* req takes over the network's ref on rqbd */
365         } else {
366                 /* req takes a ref on rqbd */
367                 rqbd->rqbd_refcount++;
368         }
369
370         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
371         svcpt->scp_nreqs_incoming++;
372
373         /* NB everything can disappear under us once the request
374          * has been queued and we unlock, so do the wake now... */
375         wake_up(&svcpt->scp_waitq);
376
377         spin_unlock(&svcpt->scp_lock);
378         EXIT;
379 }
380
381 /*
382  *  Server's outgoing reply callback
383  */
384 void reply_out_callback(lnet_event_t *ev)
385 {
386         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
387         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
388         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
389         ENTRY;
390
391         LASSERT (ev->type == LNET_EVENT_SEND ||
392                  ev->type == LNET_EVENT_ACK ||
393                  ev->type == LNET_EVENT_UNLINK);
394
395         if (!rs->rs_difficult) {
396                 /* 'Easy' replies have no further processing so I drop the
397                  * net's ref on 'rs' */
398                 LASSERT (ev->unlinked);
399                 ptlrpc_rs_decref(rs);
400                 EXIT;
401                 return;
402         }
403
404         LASSERT (rs->rs_on_net);
405
406         if (ev->unlinked) {
407                 /* Last network callback. The net's ref on 'rs' stays put
408                  * until ptlrpc_handle_rs() is done with it */
409                 spin_lock(&svcpt->scp_rep_lock);
410                 spin_lock(&rs->rs_lock);
411
412                 rs->rs_on_net = 0;
413                 if (!rs->rs_no_ack ||
414                     rs->rs_transno <=
415                     rs->rs_export->exp_obd->obd_last_committed)
416                         ptlrpc_schedule_difficult_reply(rs);
417
418                 spin_unlock(&rs->rs_lock);
419                 spin_unlock(&svcpt->scp_rep_lock);
420         }
421         EXIT;
422 }
423
424 #ifdef HAVE_SERVER_SUPPORT
425 /*
426  * Server's bulk completion callback
427  */
428 void server_bulk_callback (lnet_event_t *ev)
429 {
430         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
431         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
432         ENTRY;
433
434         LASSERT(ev->type == LNET_EVENT_SEND ||
435                 ev->type == LNET_EVENT_UNLINK ||
436                 (desc->bd_type == BULK_PUT_SOURCE &&
437                  ev->type == LNET_EVENT_ACK) ||
438                 (desc->bd_type == BULK_GET_SINK &&
439                  ev->type == LNET_EVENT_REPLY));
440
441         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
442                "event type %d, status %d, desc %p\n",
443                ev->type, ev->status, desc);
444
445         spin_lock(&desc->bd_lock);
446
447         LASSERT(desc->bd_md_count > 0);
448
449         if ((ev->type == LNET_EVENT_ACK ||
450              ev->type == LNET_EVENT_REPLY) &&
451             ev->status == 0) {
452                 /* We heard back from the peer, so even if we get this
453                  * before the SENT event (oh yes we can), we know we
454                  * read/wrote the peer buffer and how much... */
455                 desc->bd_nob_transferred += ev->mlength;
456                 desc->bd_sender = ev->sender;
457         }
458
459         if (ev->status != 0)
460                 desc->bd_failure = 1;
461
462         if (ev->unlinked) {
463                 desc->bd_md_count--;
464                 /* This is the last callback no matter what... */
465                 if (desc->bd_md_count == 0)
466                         wake_up(&desc->bd_waitq);
467         }
468
469         spin_unlock(&desc->bd_lock);
470         EXIT;
471 }
472 #endif
473
474 static void ptlrpc_master_callback(lnet_event_t *ev)
475 {
476         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
477         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
478
479         /* Honestly, it's best to find out early. */
480         LASSERT (cbid->cbid_arg != LP_POISON);
481         LASSERT (callback == request_out_callback ||
482                  callback == reply_in_callback ||
483                  callback == client_bulk_callback ||
484                  callback == request_in_callback ||
485                  callback == reply_out_callback
486 #ifdef HAVE_SERVER_SUPPORT
487                  || callback == server_bulk_callback
488 #endif
489                  );
490
491         callback (ev);
492 }
493
494 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
495                          lnet_process_id_t *peer, lnet_nid_t *self)
496 {
497         int               best_dist = 0;
498         __u32             best_order = 0;
499         int               count = 0;
500         int               rc = -ENOENT;
501         int               portals_compatibility;
502         int               dist;
503         __u32             order;
504         lnet_nid_t        dst_nid;
505         lnet_nid_t        src_nid;
506
507         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
508
509         peer->pid = LNET_PID_LUSTRE;
510
511         /* Choose the matching UUID that's closest */
512         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
513                 dist = LNetDist(dst_nid, &src_nid, &order);
514                 if (dist < 0)
515                         continue;
516
517                 if (dist == 0) {                /* local! use loopback LND */
518                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
519                         rc = 0;
520                         break;
521                 }
522
523                 if (rc < 0 ||
524                     dist < best_dist ||
525                     (dist == best_dist && order < best_order)) {
526                         best_dist = dist;
527                         best_order = order;
528
529                         if (portals_compatibility > 1) {
530                                 /* Strong portals compatibility: Zero the nid's
531                                  * NET, so if I'm reading new config logs, or
532                                  * getting configured by (new) lconf I can
533                                  * still talk to old servers. */
534                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
535                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
536                         }
537                         peer->nid = dst_nid;
538                         *self = src_nid;
539                         rc = 0;
540                 }
541         }
542
543         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
544         return rc;
545 }
546
547 void ptlrpc_ni_fini(void)
548 {
549         wait_queue_head_t         waitq;
550         struct l_wait_info  lwi;
551         int                 rc;
552         int                 retries;
553
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
557          * replies */
558
559         for (retries = 0;; retries++) {
560                 rc = LNetEQFree(ptlrpc_eq_h);
561                 switch (rc) {
562                 default:
563                         LBUG();
564
565                 case 0:
566                         LNetNIFini();
567                         return;
568
569                 case -EBUSY:
570                         if (retries != 0)
571                                 CWARN("Event queue still busy\n");
572
573                         /* Wait for a bit */
574                         init_waitqueue_head(&waitq);
575                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
576                         l_wait_event(waitq, 0, &lwi);
577                         break;
578                 }
579         }
580         /* notreached */
581 }
582
583 lnet_pid_t ptl_get_pid(void)
584 {
585         return LNET_PID_LUSTRE;
586 }
587
588 int ptlrpc_ni_init(void)
589 {
590         int              rc;
591         lnet_pid_t       pid;
592
593         pid = ptl_get_pid();
594         CDEBUG(D_NET, "My pid is: %x\n", pid);
595
596         /* We're not passing any limits yet... */
597         rc = LNetNIInit(pid);
598         if (rc < 0) {
599                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
600                 return rc;
601         }
602
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);
610         if (rc == 0)
611                 return 0;
612
613         CERROR ("Failed to allocate event queue: %d\n", rc);
614         LNetNIFini();
615
616         return rc;
617 }
618
619
620 int ptlrpc_init_portals(void)
621 {
622         int   rc = ptlrpc_ni_init();
623
624         if (rc != 0) {
625                 CERROR("network initialisation failed\n");
626                 return rc;
627         }
628         rc = ptlrpcd_addref();
629         if (rc == 0)
630                 return 0;
631
632         CERROR("rpcd initialisation failed\n");
633         ptlrpc_ni_fini();
634         return rc;
635 }
636
637 void ptlrpc_exit_portals(void)
638 {
639         ptlrpcd_decref();
640         ptlrpc_ni_fini();
641 }