Whamcloud - gitweb
LU-7434 ptlrpc: Early Reply vs Reply MDunlink
[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, 2015, 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         bool                   wakeup = false;
56         ENTRY;
57
58         LASSERT(ev->type == LNET_EVENT_SEND || 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
65         spin_lock(&req->rq_lock);
66         req->rq_real_sent = cfs_time_current_sec();
67         req->rq_req_unlinked = 1;
68         /* reply_in_callback happened before request_out_callback? */
69         if (req->rq_reply_unlinked)
70                 wakeup = true;
71
72         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
73                 /* Failed send: make it seem like the reply timed out, just
74                  * like failing sends in client.c does currently...  */
75                 req->rq_net_err = 1;
76                 wakeup = true;
77         }
78
79         if (wakeup)
80                 ptlrpc_client_wake_req(req);
81
82         spin_unlock(&req->rq_lock);
83
84         ptlrpc_req_finished(req);
85         EXIT;
86 }
87
88 /*
89  * Client's incoming reply callback
90  */
91 void reply_in_callback(lnet_event_t *ev)
92 {
93         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
94         struct ptlrpc_request *req = cbid->cbid_arg;
95         ENTRY;
96
97         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
98
99         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
100         LASSERT (ev->md.start == req->rq_repbuf);
101         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
102         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
103            for adaptive timeouts' early reply. */
104         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
105
106         spin_lock(&req->rq_lock);
107
108         req->rq_receiving_reply = 0;
109         req->rq_early = 0;
110         if (ev->unlinked)
111                 req->rq_reply_unlinked = 1;
112
113         if (ev->status)
114                 goto out_wake;
115
116         if (ev->type == LNET_EVENT_UNLINK) {
117                 LASSERT(ev->unlinked);
118                 DEBUG_REQ(D_NET, req, "unlink");
119                 goto out_wake;
120         }
121
122         if (ev->mlength < ev->rlength ) {
123                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
124                        req->rq_replen, ev->rlength, ev->offset);
125                 req->rq_reply_truncated = 1;
126                 req->rq_replied = 1;
127                 req->rq_status = -EOVERFLOW;
128                 req->rq_nob_received = ev->rlength + ev->offset;
129                 goto out_wake;
130         }
131
132         if ((ev->offset == 0) &&
133             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
134                 /* Early reply */
135                 DEBUG_REQ(D_ADAPTTO, req,
136                           "Early reply received: mlen=%u offset=%d replen=%d "
137                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
138                           req->rq_replen, req->rq_replied, ev->unlinked);
139
140                 req->rq_early_count++; /* number received, client side */
141
142                 /* already got the real reply or buffers are already unlinked */
143                 if (req->rq_replied ||
144                     req->rq_reply_unlinked == 1)
145                         goto out_wake;
146
147                 req->rq_early = 1;
148                 req->rq_reply_off = ev->offset;
149                 req->rq_nob_received = ev->mlength;
150                 /* And we're still receiving */
151                 req->rq_receiving_reply = 1;
152         } else {
153                 /* Real reply */
154                 req->rq_rep_swab_mask = 0;
155                 req->rq_replied = 1;
156                 /* Got reply, no resend required */
157                 req->rq_resend = 0;
158                 req->rq_reply_off = ev->offset;
159                 req->rq_nob_received = ev->mlength;
160                 /* LNetMDUnlink can't be called under the LNET_LOCK,
161                    so we must unlink in ptlrpc_unregister_reply */
162                 DEBUG_REQ(D_INFO, req,
163                           "reply in flags=%x mlen=%u offset=%d replen=%d",
164                           lustre_msg_get_flags(req->rq_reqmsg),
165                           ev->mlength, ev->offset, req->rq_replen);
166         }
167
168         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
169
170 out_wake:
171         /* NB don't unlock till after wakeup; req can disappear under us
172          * since we don't have our own ref */
173         ptlrpc_client_wake_req(req);
174         spin_unlock(&req->rq_lock);
175         EXIT;
176 }
177
178 /*
179  * Client's bulk has been written/read
180  */
181 void client_bulk_callback (lnet_event_t *ev)
182 {
183         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
184         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
185         struct ptlrpc_request   *req;
186         ENTRY;
187
188         LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
189                  ev->type == LNET_EVENT_PUT) ||
190                 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
191                  ev->type == LNET_EVENT_GET) ||
192                 ev->type == LNET_EVENT_UNLINK);
193         LASSERT(ev->unlinked);
194
195         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
196                 ev->status = -EIO;
197
198         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
199                 ev->status = -EIO;
200
201         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
202                "event type %d, status %d, desc %p\n",
203                ev->type, ev->status, desc);
204
205         spin_lock(&desc->bd_lock);
206         req = desc->bd_req;
207         LASSERT(desc->bd_md_count > 0);
208         desc->bd_md_count--;
209
210         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
211                 desc->bd_nob_transferred += ev->mlength;
212                 desc->bd_sender = ev->sender;
213         } else {
214                 /* start reconnect and resend if network error hit */
215                 spin_lock(&req->rq_lock);
216                 req->rq_net_err = 1;
217                 spin_unlock(&req->rq_lock);
218         }
219
220         if (ev->status != 0)
221                 desc->bd_failure = 1;
222
223         /* NB don't unlock till after wakeup; desc can disappear under us
224          * otherwise */
225         if (desc->bd_md_count == 0)
226                 ptlrpc_client_wake_req(desc->bd_req);
227
228         spin_unlock(&desc->bd_lock);
229         EXIT;
230 }
231
232 /*
233  * We will have percpt request history list for ptlrpc service in upcoming
234  * patches because we don't want to be serialized by current per-service
235  * history operations. So we require history ID can (somehow) show arriving
236  * order w/o grabbing global lock, and user can sort them in userspace.
237  *
238  * This is how we generate history ID for ptlrpc_request:
239  * ----------------------------------------------------
240  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
241  * ----------------------------------------------------
242  * |  seconds  | usec / 16 |   sequence    | CPT id   |
243  * ----------------------------------------------------
244  *
245  * it might not be precise but should be good enough.
246  */
247
248 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
249
250 #define REQS_SEC_SHIFT          32
251 #define REQS_USEC_SHIFT         16
252 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
253
254 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
255                                    struct ptlrpc_request *req)
256 {
257         __u64   sec = req->rq_arrival_time.tv_sec;
258         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
259         __u64   new_seq;
260
261         /* set sequence ID for request and add it to history list,
262          * it must be called with hold svcpt::scp_lock */
263
264         new_seq = (sec << REQS_SEC_SHIFT) |
265                   (usec << REQS_USEC_SHIFT) |
266                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
267
268         if (new_seq > svcpt->scp_hist_seq) {
269                 /* This handles the initial case of scp_hist_seq == 0 or
270                  * we just jumped into a new time window */
271                 svcpt->scp_hist_seq = new_seq;
272         } else {
273                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
274                 /* NB: increase sequence number in current usec bucket,
275                  * however, it's possible that we used up all bits for
276                  * sequence and jumped into the next usec bucket (future time),
277                  * then we hope there will be less RPCs per bucket at some
278                  * point, and sequence will catch up again */
279                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
280                 new_seq = svcpt->scp_hist_seq;
281         }
282
283         req->rq_history_seq = new_seq;
284
285         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
286 }
287
288 /*
289  * Server's incoming request callback
290  */
291 void request_in_callback(lnet_event_t *ev)
292 {
293         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
294         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
295         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
296         struct ptlrpc_service             *service = svcpt->scp_service;
297         struct ptlrpc_request             *req;
298         ENTRY;
299
300         LASSERT (ev->type == LNET_EVENT_PUT ||
301                  ev->type == LNET_EVENT_UNLINK);
302         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
303         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
304                  rqbd->rqbd_buffer + service->srv_buf_size);
305
306         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
307                "event type %d, status %d, service %s\n",
308                ev->type, ev->status, service->srv_name);
309
310         if (ev->unlinked) {
311                 /* If this is the last request message to fit in the
312                  * request buffer we can use the request object embedded in
313                  * rqbd.  Note that if we failed to allocate a request,
314                  * we'd have to re-post the rqbd, which we can't do in this
315                  * context. */
316                 req = &rqbd->rqbd_req;
317                 memset(req, 0, sizeof (*req));
318         } else {
319                 LASSERT (ev->type == LNET_EVENT_PUT);
320                 if (ev->status != 0) {
321                         /* We moaned above already... */
322                         return;
323                 }
324                 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
325                 if (req == NULL) {
326                         CERROR("Can't allocate incoming request descriptor: "
327                                "Dropping %s RPC from %s\n",
328                                service->srv_name,
329                                libcfs_id2str(ev->initiator));
330                         return;
331                 }
332         }
333
334         ptlrpc_srv_req_init(req);
335         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
336          * flags are reset and scalars are zero.  We only set the message
337          * size to non-zero if this was a successful receive. */
338         req->rq_xid = ev->match_bits;
339         req->rq_reqbuf = ev->md.start + ev->offset;
340         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
341                 req->rq_reqdata_len = ev->mlength;
342         do_gettimeofday(&req->rq_arrival_time);
343         req->rq_peer = ev->initiator;
344         req->rq_self = ev->target.nid;
345         req->rq_rqbd = rqbd;
346         req->rq_phase = RQ_PHASE_NEW;
347         if (ev->type == LNET_EVENT_PUT)
348                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
349                        req, req->rq_xid, ev->mlength);
350
351         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
352
353         spin_lock(&svcpt->scp_lock);
354
355         ptlrpc_req_add_history(svcpt, req);
356
357         if (ev->unlinked) {
358                 svcpt->scp_nrqbds_posted--;
359                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
360                        svcpt->scp_nrqbds_posted);
361
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",
368                               service->srv_name);
369
370                 /* req takes over the network's ref on rqbd */
371         } else {
372                 /* req takes a ref on rqbd */
373                 rqbd->rqbd_refcount++;
374         }
375
376         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
377         svcpt->scp_nreqs_incoming++;
378
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);
382
383         spin_unlock(&svcpt->scp_lock);
384         EXIT;
385 }
386
387 /*
388  *  Server's outgoing reply callback
389  */
390 void reply_out_callback(lnet_event_t *ev)
391 {
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;
395         ENTRY;
396
397         LASSERT (ev->type == LNET_EVENT_SEND ||
398                  ev->type == LNET_EVENT_ACK ||
399                  ev->type == LNET_EVENT_UNLINK);
400
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);
406                 EXIT;
407                 return;
408         }
409
410         LASSERT (rs->rs_on_net);
411
412         if (ev->unlinked) {
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);
417
418                 rs->rs_on_net = 0;
419                 if (!rs->rs_no_ack ||
420                     rs->rs_transno <=
421                     rs->rs_export->exp_obd->obd_last_committed)
422                         ptlrpc_schedule_difficult_reply(rs);
423
424                 spin_unlock(&rs->rs_lock);
425                 spin_unlock(&svcpt->scp_rep_lock);
426         }
427         EXIT;
428 }
429
430 #ifdef HAVE_SERVER_SUPPORT
431 /*
432  * Server's bulk completion callback
433  */
434 void server_bulk_callback (lnet_event_t *ev)
435 {
436         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
437         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
438         ENTRY;
439
440         LASSERT(ev->type == LNET_EVENT_SEND ||
441                 ev->type == LNET_EVENT_UNLINK ||
442                 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
443                  ev->type == LNET_EVENT_ACK) ||
444                 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
445                  ev->type == LNET_EVENT_REPLY));
446
447         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
448                "event type %d, status %d, desc %p\n",
449                ev->type, ev->status, desc);
450
451         spin_lock(&desc->bd_lock);
452
453         LASSERT(desc->bd_md_count > 0);
454
455         if ((ev->type == LNET_EVENT_ACK ||
456              ev->type == LNET_EVENT_REPLY) &&
457             ev->status == 0) {
458                 /* We heard back from the peer, so even if we get this
459                  * before the SENT event (oh yes we can), we know we
460                  * read/wrote the peer buffer and how much... */
461                 desc->bd_nob_transferred += ev->mlength;
462                 desc->bd_sender = ev->sender;
463         }
464
465         if (ev->status != 0)
466                 desc->bd_failure = 1;
467
468         if (ev->unlinked) {
469                 desc->bd_md_count--;
470                 /* This is the last callback no matter what... */
471                 if (desc->bd_md_count == 0)
472                         wake_up(&desc->bd_waitq);
473         }
474
475         spin_unlock(&desc->bd_lock);
476         EXIT;
477 }
478 #endif
479
480 static void ptlrpc_master_callback(lnet_event_t *ev)
481 {
482         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
483         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
484
485         /* Honestly, it's best to find out early. */
486         LASSERT (cbid->cbid_arg != LP_POISON);
487         LASSERT (callback == request_out_callback ||
488                  callback == reply_in_callback ||
489                  callback == client_bulk_callback ||
490                  callback == request_in_callback ||
491                  callback == reply_out_callback
492 #ifdef HAVE_SERVER_SUPPORT
493                  || callback == server_bulk_callback
494 #endif
495                  );
496
497         callback (ev);
498 }
499
500 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
501                          lnet_process_id_t *peer, lnet_nid_t *self)
502 {
503         int               best_dist = 0;
504         __u32             best_order = 0;
505         int               count = 0;
506         int               rc = -ENOENT;
507         int               dist;
508         __u32             order;
509         lnet_nid_t        dst_nid;
510         lnet_nid_t        src_nid;
511
512         peer->pid = LNET_PID_LUSTRE;
513
514         /* Choose the matching UUID that's closest */
515         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
516                 dist = LNetDist(dst_nid, &src_nid, &order);
517                 if (dist < 0)
518                         continue;
519
520                 if (dist == 0) {                /* local! use loopback LND */
521                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
522                         rc = 0;
523                         break;
524                 }
525
526                 if (rc < 0 ||
527                     dist < best_dist ||
528                     (dist == best_dist && order < best_order)) {
529                         best_dist = dist;
530                         best_order = order;
531
532                         peer->nid = dst_nid;
533                         *self = src_nid;
534                         rc = 0;
535                 }
536         }
537
538         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
539         return rc;
540 }
541
542 void ptlrpc_ni_fini(void)
543 {
544         wait_queue_head_t         waitq;
545         struct l_wait_info  lwi;
546         int                 rc;
547         int                 retries;
548
549         /* Wait for the event queue to become idle since there may still be
550          * messages in flight with pending events (i.e. the fire-and-forget
551          * messages == client requests and "non-difficult" server
552          * replies */
553
554         for (retries = 0;; retries++) {
555                 rc = LNetEQFree(ptlrpc_eq_h);
556                 switch (rc) {
557                 default:
558                         LBUG();
559
560                 case 0:
561                         LNetNIFini();
562                         return;
563
564                 case -EBUSY:
565                         if (retries != 0)
566                                 CWARN("Event queue still busy\n");
567
568                         /* Wait for a bit */
569                         init_waitqueue_head(&waitq);
570                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
571                         l_wait_event(waitq, 0, &lwi);
572                         break;
573                 }
574         }
575         /* notreached */
576 }
577
578 lnet_pid_t ptl_get_pid(void)
579 {
580         return LNET_PID_LUSTRE;
581 }
582
583 int ptlrpc_ni_init(void)
584 {
585         int              rc;
586         lnet_pid_t       pid;
587
588         pid = ptl_get_pid();
589         CDEBUG(D_NET, "My pid is: %x\n", pid);
590
591         /* We're not passing any limits yet... */
592         rc = LNetNIInit(pid);
593         if (rc < 0) {
594                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
595                 return rc;
596         }
597
598         /* CAVEAT EMPTOR: how we process portals events is _radically_
599          * different depending on... */
600         /* kernel LNet calls our master callback when there are new event,
601          * because we are guaranteed to get every event via callback,
602          * so we just set EQ size to 0 to avoid overhread of serializing
603          * enqueue/dequeue operations in LNet. */
604         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
605         if (rc == 0)
606                 return 0;
607
608         CERROR ("Failed to allocate event queue: %d\n", rc);
609         LNetNIFini();
610
611         return rc;
612 }
613
614
615 int ptlrpc_init_portals(void)
616 {
617         int   rc = ptlrpc_ni_init();
618
619         if (rc != 0) {
620                 CERROR("network initialisation failed\n");
621                 return rc;
622         }
623         rc = ptlrpcd_addref();
624         if (rc == 0)
625                 return 0;
626
627         CERROR("rpcd initialisation failed\n");
628         ptlrpc_ni_fini();
629         return rc;
630 }
631
632 void ptlrpc_exit_portals(void)
633 {
634         ptlrpcd_decref();
635         ptlrpc_ni_fini();
636 }