Whamcloud - gitweb
64dfa70b0dac6804ee87c6a00dc3ff6392df4035
[fs/lustre-release.git] / lustre / ptlrpc / niobuf.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  */
31
32 #define DEBUG_SUBSYSTEM S_RPC
33 #include <libcfs/linux/linux-mem.h>
34 #include <obd_support.h>
35 #include <lustre_net.h>
36 #include <lustre_lib.h>
37 #include <obd.h>
38 #include <obd_class.h>
39 #include "ptlrpc_internal.h"
40 #include <lnet/lib-lnet.h> /* for CFS_FAIL_PTLRPC_OST_BULK_CB2 */
41
42 /**
43  * Helper function. Sends \a len bytes from \a base at offset \a offset
44  * over \a conn connection to portal \a portal.
45  * Returns 0 on success or error code.
46  */
47 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
48                         enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
49                         struct lnet_nid *self, struct lnet_processid *peer_id,
50                         int portal, __u64 xid, unsigned int offset,
51                         struct lnet_handle_md *bulk_cookie)
52 {
53         int rc;
54         struct lnet_md md;
55         ENTRY;
56
57         LASSERT(portal != 0);
58         CDEBUG(D_INFO, "peer_id %s\n", libcfs_idstr(peer_id));
59         md.start     = base;
60         md.length    = len;
61         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
62         md.options   = PTLRPC_MD_OPTIONS;
63         md.user_ptr  = cbid;
64         md.handler   = ptlrpc_handler;
65         LNetInvalidateMDHandle(&md.bulk_handle);
66
67         if (bulk_cookie) {
68                 md.bulk_handle = *bulk_cookie;
69                 md.options |= LNET_MD_BULK_HANDLE;
70         }
71
72         if (unlikely(ack == LNET_ACK_REQ &&
73                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
74                 /* don't ask for the ack to simulate failing client */
75                 ack = LNET_NOACK_REQ;
76         }
77
78         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
79         if (unlikely(rc != 0)) {
80                 CERROR ("LNetMDBind failed: %d\n", rc);
81                 LASSERT (rc == -ENOMEM);
82                 RETURN (-ENOMEM);
83         }
84
85         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
86                len, portal, xid, offset);
87
88         percpu_ref_get(&ptlrpc_pending);
89
90         rc = LNetPut(self, *mdh, ack,
91                      peer_id, portal, xid, offset, 0);
92         if (unlikely(rc != 0)) {
93                 int rc2;
94                 /* We're going to get an UNLINK event when I unlink below,
95                  * which will complete just like any other failed send, so
96                  * I fall through and return success here! */
97                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
98                        libcfs_idstr(peer_id), portal, xid, rc);
99                 rc2 = LNetMDUnlink(*mdh);
100                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
101         }
102
103         RETURN (0);
104 }
105
106 #define mdunlink_iterate_helper(mds, count) \
107                 __mdunlink_iterate_helper(mds, count, false) 
108 static void __mdunlink_iterate_helper(struct lnet_handle_md *bd_mds,
109                                       int count, bool discard)
110 {
111         int i;
112
113         for (i = 0; i < count; i++)
114                 __LNetMDUnlink(bd_mds[i], discard);
115 }
116
117 #ifdef HAVE_SERVER_SUPPORT
118 /**
119  * Prepare bulk descriptor for specified incoming request \a req that
120  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
121  * the bulk to be sent. Used on server-side after request was already
122  * received.
123  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
124  * error.
125  */
126 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
127                                               unsigned nfrags, unsigned max_brw,
128                                               unsigned int type,
129                                               unsigned portal,
130                                               const struct ptlrpc_bulk_frag_ops
131                                                 *ops)
132 {
133         struct obd_export *exp = req->rq_export;
134         struct ptlrpc_bulk_desc *desc;
135
136         ENTRY;
137         LASSERT(ptlrpc_is_bulk_op_active(type));
138
139         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
140         if (desc == NULL)
141                 RETURN(NULL);
142
143         desc->bd_export = class_export_get(exp);
144         desc->bd_req = req;
145
146         desc->bd_cbid.cbid_fn  = server_bulk_callback;
147         desc->bd_cbid.cbid_arg = desc;
148
149         /* NB we don't assign rq_bulk here; server-side requests are
150          * re-used, and the handler frees the bulk desc explicitly. */
151
152         return desc;
153 }
154 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
155
156 /**
157  * Starts bulk transfer for descriptor \a desc on the server.
158  * Returns 0 on success or error code.
159  */
160 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
161 {
162         struct obd_export       *exp = desc->bd_export;
163         struct lnet_nid          self_nid;
164         struct lnet_processid    peer_id;
165         int                      rc = 0;
166         __u64                    mbits;
167         int                      posted_md;
168         int                      total_md;
169         struct lnet_md           md;
170         ENTRY;
171
172         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
173                 RETURN(0);
174
175         /* NB no locking required until desc is on the network */
176         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
177
178         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
179         LASSERT(desc->bd_cbid.cbid_arg == desc);
180
181         /*
182          * Multi-Rail: get the preferred self and peer NIDs from the
183          * request, so they are based on the route taken by the
184          * message.
185          */
186         self_nid = desc->bd_req->rq_self;
187         peer_id = desc->bd_req->rq_source;
188
189         /* NB total length may be 0 for a read past EOF, so we send 0
190          * length bulks, since the client expects bulk events.
191          *
192          * The client may not need all of the bulk mbits for the RPC. The RPC
193          * used the mbits of the highest bulk mbits needed, and the server masks
194          * off high bits to get bulk count for this RPC. LU-1431 */
195         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
196         total_md = desc->bd_req->rq_mbits - mbits + 1;
197         desc->bd_refs = total_md;
198         desc->bd_failure = 0;
199
200         md.user_ptr = &desc->bd_cbid;
201         md.handler = ptlrpc_handler;
202         md.threshold = 2; /* SENT and ACK/REPLY */
203
204         for (posted_md = 0; posted_md < total_md; mbits++) {
205                 md.options = PTLRPC_MD_OPTIONS;
206
207                 /* NB it's assumed that source and sink buffer frags are
208                  * page-aligned. Otherwise we'd have to send client bulk
209                  * sizes over and split server buffer accordingly */
210                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
211                 rc = LNetMDBind(&md, LNET_UNLINK, &desc->bd_mds[posted_md]);
212                 if (rc != 0) {
213                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
214                                exp->exp_obd->obd_name, posted_md, rc);
215                         LASSERT(rc == -ENOMEM);
216                         if (posted_md == 0) {
217                                 desc->bd_md_count = 0;
218                                 RETURN(-ENOMEM);
219                         }
220                         break;
221                 }
222                 percpu_ref_get(&ptlrpc_pending);
223
224                 /* sanity.sh 224c: lets skip last md */
225                 if (posted_md == desc->bd_md_max_brw - 1)
226                         OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
227                                              CFS_FAIL_PTLRPC_OST_BULK_CB2);
228
229                 /* Network is about to get at the memory */
230                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
231                         rc = LNetPut(&self_nid, desc->bd_mds[posted_md],
232                                      LNET_ACK_REQ, &peer_id,
233                                      desc->bd_portal, mbits, 0, 0);
234                 else
235                         rc = LNetGet(&self_nid, desc->bd_mds[posted_md],
236                                      &peer_id, desc->bd_portal,
237                                      mbits, 0, false);
238
239                 posted_md++;
240                 if (rc != 0) {
241                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
242                                "rc = %d\n", exp->exp_obd->obd_name,
243                                libcfs_idstr(&peer_id), desc->bd_portal,
244                                mbits, rc);
245                         break;
246                 }
247         }
248
249         if (rc != 0) {
250                 /* Can't send, so we unlink the MD bound above.  The UNLINK
251                  * event this creates will signal completion with failure,
252                  * so we return SUCCESS here! */
253                 spin_lock(&desc->bd_lock);
254                 desc->bd_refs -= total_md - posted_md;
255                 spin_unlock(&desc->bd_lock);
256                 LASSERT(desc->bd_refs >= 0);
257
258                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
259                 RETURN(0);
260         }
261
262         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
263                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
264                desc->bd_nob, desc->bd_portal, libcfs_idstr(&peer_id),
265                mbits - posted_md, mbits - 1);
266
267         RETURN(0);
268 }
269
270 /**
271  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
272  * serialises with completion callback)
273  */
274 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
275 {
276         LASSERT(!in_interrupt());           /* might sleep */
277
278         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
279                 return;                         /* never started */
280
281         /* We used to poison the pages with 0xab here because we did not want to
282          * send any meaningful data over the wire for evicted clients (bug 9297)
283          * However, this is no longer safe now that we use the page cache on the
284          * OSS (bug 20560) */
285
286         /* The unlink ensures the callback happens ASAP and is the last
287          * one.  If it fails, it must be because completion just happened,
288          * but we must still wait_event_idle_timeout() in this case, to give
289          * us a chance to run server_bulk_callback()
290          */
291         __mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw, true);
292
293         for (;;) {
294                 /* Network access will complete in finite time but the HUGE
295                  * timeout lets us CWARN for visibility of sluggish NALs */
296                 int seconds = PTLRPC_REQ_LONG_UNLINK;
297
298                 while (seconds > 0 &&
299                        wait_event_idle_timeout(desc->bd_waitq,
300                                                !ptlrpc_server_bulk_active(desc),
301                                                cfs_time_seconds(1)) == 0)
302                         seconds -= 1;
303                 if (seconds > 0)
304                         return;
305
306                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
307         }
308 }
309 #endif /* HAVE_SERVER_SUPPORT */
310
311 /**
312  * Register bulk at the sender for later transfer.
313  * Returns 0 on success or error code.
314  */
315 int ptlrpc_register_bulk(struct ptlrpc_request *req)
316 {
317         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
318         struct lnet_processid peer;
319         int rc = 0;
320         int posted_md;
321         int total_md;
322         __u64 mbits;
323         struct lnet_me *me;
324         struct lnet_md md;
325         ENTRY;
326
327         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
328                 RETURN(0);
329
330         /* NB no locking required until desc is on the network */
331         LASSERT(desc->bd_nob > 0);
332         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
333         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
334         LASSERT(desc->bd_req != NULL);
335         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
336
337         /* cleanup the state of the bulk for it will be reused */
338         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
339                 desc->bd_nob_transferred = 0;
340         else if (desc->bd_nob_transferred != 0)
341                 /* If the network failed after an RPC was sent, this condition
342                  * could happen.  Rather than assert (was here before), return
343                  * an EIO error. */
344                 RETURN(-EIO);
345
346         desc->bd_failure = 0;
347
348         peer = desc->bd_import->imp_connection->c_peer;
349
350         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
351         LASSERT(desc->bd_cbid.cbid_arg == desc);
352
353         total_md = desc->bd_md_count;
354         /* rq_mbits is matchbits of the final bulk */
355         mbits = req->rq_mbits - desc->bd_md_count + 1;
356
357         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
358                  "first mbits = x%llu, last mbits = x%llu\n",
359                  mbits, req->rq_mbits);
360         LASSERTF(!(desc->bd_registered &&
361                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
362                  mbits != desc->bd_last_mbits,
363                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
364                  desc->bd_registered, mbits, desc->bd_last_mbits);
365
366         desc->bd_registered = 1;
367         desc->bd_last_mbits = mbits;
368         desc->bd_refs = total_md;
369         md.user_ptr = &desc->bd_cbid;
370         md.handler = ptlrpc_handler;
371         md.threshold = 1;                       /* PUT or GET */
372
373         for (posted_md = 0; posted_md < desc->bd_md_count;
374              posted_md++, mbits++) {
375                 md.options = PTLRPC_MD_OPTIONS |
376                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
377                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
378                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
379
380                 if (posted_md > 0 && posted_md + 1 == desc->bd_md_count &&
381                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
382                         rc = -ENOMEM;
383                 } else {
384                         me = LNetMEAttach(desc->bd_portal, &peer, mbits, 0,
385                                   LNET_UNLINK, LNET_INS_AFTER);
386                         rc = PTR_ERR_OR_ZERO(me);
387                 }
388                 if (rc != 0) {
389                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
390                                desc->bd_import->imp_obd->obd_name, mbits,
391                                posted_md, rc);
392                         break;
393                 }
394                 percpu_ref_get(&ptlrpc_pending);
395
396                 /* About to let the network at it... */
397                 rc = LNetMDAttach(me, &md, LNET_UNLINK,
398                                   &desc->bd_mds[posted_md]);
399                 if (rc != 0) {
400                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
401                                desc->bd_import->imp_obd->obd_name, mbits,
402                                posted_md, rc);
403                         break;
404                 }
405         }
406
407         if (rc != 0) {
408                 LASSERT(rc == -ENOMEM);
409                 spin_lock(&desc->bd_lock);
410                 desc->bd_refs -= total_md - posted_md;
411                 spin_unlock(&desc->bd_lock);
412                 LASSERT(desc->bd_refs >= 0);
413                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
414                 req->rq_status = -ENOMEM;
415                 desc->bd_registered = 0;
416                 RETURN(-ENOMEM);
417         }
418
419         spin_lock(&desc->bd_lock);
420         /* Holler if peer manages to touch buffers before he knows the mbits */
421         if (desc->bd_refs != total_md)
422                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
423                       desc->bd_import->imp_obd->obd_name, libcfs_idstr(&peer),
424                       total_md - desc->bd_refs);
425         spin_unlock(&desc->bd_lock);
426
427         CDEBUG(D_NET,
428                "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
429                desc->bd_refs,
430                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
431                desc->bd_iov_count, desc->bd_nob,
432                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
433
434         RETURN(0);
435 }
436
437 /**
438  * Disconnect a bulk desc from the network. Idempotent. Not
439  * thread-safe (i.e. only interlocks with completion callback).
440  * Returns 1 on success or 0 if network unregistration failed for whatever
441  * reason.
442  */
443 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
444 {
445         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
446         ENTRY;
447
448         LASSERT(!in_interrupt());     /* might sleep */
449
450         if (desc)
451                 desc->bd_registered = 0;
452
453         /* Let's setup deadline for reply unlink. */
454         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
455             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
456                 req->rq_bulk_deadline = ktime_get_real_seconds() +
457                                         PTLRPC_REQ_LONG_UNLINK;
458
459         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
460                 RETURN(1);                              /* never registered */
461
462         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
463
464         /* the unlink ensures the callback happens ASAP and is the last
465          * one.  If it fails, it must be because completion just happened,
466          * but we must still wait_event_idle_timeout() in this case to give
467          * us a chance to run client_bulk_callback()
468          */
469         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
470
471         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
472                 RETURN(1);                              /* never registered */
473
474         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
475         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
476
477         /* Do not wait for unlink to finish. */
478         if (async)
479                 RETURN(0);
480
481         for (;;) {
482                 /* The wq argument is ignored by user-space wait_event macros */
483                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
484                                         &req->rq_set->set_waitq :
485                                         &req->rq_reply_waitq;
486                 /*
487                  * Network access will complete in finite time but the HUGE
488                  * timeout lets us CWARN for visibility of sluggish NALs.
489                  */
490                 int seconds = PTLRPC_REQ_LONG_UNLINK;
491
492                 while (seconds > 0 &&
493                        wait_event_idle_timeout(*wq,
494                                                !ptlrpc_client_bulk_active(req),
495                                                cfs_time_seconds(1)) == 0)
496                         seconds -= 1;
497                 if (seconds > 0) {
498                         ptlrpc_rqphase_move(req, req->rq_next_phase);
499                         RETURN(1);
500                 }
501
502                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
503                           desc);
504         }
505         RETURN(0);
506 }
507
508 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
509 {
510         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
511         struct ptlrpc_service           *svc = svcpt->scp_service;
512         timeout_t service_timeout;
513
514         service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
515                                              req->rq_arrival_time.tv_sec, 1,
516                                   (AT_OFF ? obd_timeout * 3 / 2 : at_max));
517         if (!(flags & PTLRPC_REPLY_EARLY) &&
518             (req->rq_type != PTL_RPC_MSG_ERR) &&
519             (req->rq_reqmsg != NULL) &&
520             !(lustre_msg_get_flags(req->rq_reqmsg) &
521               (MSG_RESENT | MSG_REPLAY |
522                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
523                 /* early replies, errors and recovery requests don't count
524                  * toward our service time estimate
525                  */
526                 timeout_t oldse = at_measured(&svcpt->scp_at_estimate,
527                                               service_timeout);
528
529                 if (oldse != 0) {
530                         DEBUG_REQ(D_ADAPTTO, req,
531                                   "svc %s changed estimate from %d to %d",
532                                   svc->srv_name, oldse,
533                                   at_get(&svcpt->scp_at_estimate));
534                 }
535         }
536         /* Report actual service time for client latency calc */
537         lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
538         /* Report service time estimate for future client reqs, but report 0
539          * (to be ignored by client) if it's an error reply during recovery.
540          * b=15815
541          */
542         if (req->rq_type == PTL_RPC_MSG_ERR &&
543             (req->rq_export == NULL ||
544              req->rq_export->exp_obd->obd_recovering)) {
545                 lustre_msg_set_timeout(req->rq_repmsg, 0);
546         } else {
547                 timeout_t timeout;
548
549                 if (req->rq_export && req->rq_reqmsg != NULL &&
550                     (flags & PTLRPC_REPLY_EARLY) &&
551                     lustre_msg_get_flags(req->rq_reqmsg) &
552                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
553                         struct obd_device *exp_obd = req->rq_export->exp_obd;
554
555                         timeout = ktime_get_real_seconds() -
556                                   req->rq_arrival_time.tv_sec +
557                                   min_t(timeout_t, at_extra,
558                                         exp_obd->obd_recovery_timeout / 4);
559                 } else {
560                         timeout = at_get(&svcpt->scp_at_estimate);
561                 }
562                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
563         }
564
565         if (req->rq_reqmsg &&
566             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
567                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
568                        "req_flags=%#x magic=%x/%x len=%d\n",
569                        flags, lustre_msg_get_flags(req->rq_reqmsg),
570                        lustre_msg_get_magic(req->rq_reqmsg),
571                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
572         }
573 }
574
575 /**
576  * Send request reply from request \a req reply buffer.
577  * \a flags defines reply types
578  * Returns 0 on success or error code
579  */
580 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
581 {
582         struct ptlrpc_reply_state *rs = req->rq_reply_state;
583         struct ptlrpc_connection  *conn;
584         int                        rc;
585
586         /* We must already have a reply buffer (only ptlrpc_error() may be
587          * called without one). The reply generated by sptlrpc layer (e.g.
588          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
589          * have a request buffer which is either the actual (swabbed) incoming
590          * request, or a saved copy if this is a req saved in
591          * target_queue_final_reply().
592          */
593         LASSERT (req->rq_no_reply == 0);
594         LASSERT (req->rq_reqbuf != NULL);
595         LASSERT (rs != NULL);
596         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
597         LASSERT (req->rq_repmsg != NULL);
598         LASSERT (req->rq_repmsg == rs->rs_msg);
599         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
600         LASSERT (rs->rs_cb_id.cbid_arg == rs);
601
602         /* There may be no rq_export during failover */
603
604         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
605                      req->rq_export->exp_obd->obd_fail)) {
606                 /* Failed obd's only send ENODEV */
607                 req->rq_type = PTL_RPC_MSG_ERR;
608                 req->rq_status = -ENODEV;
609                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
610                        req->rq_export->exp_obd->obd_minor);
611         }
612
613         if (req->rq_type != PTL_RPC_MSG_ERR)
614                 req->rq_type = PTL_RPC_MSG_REPLY;
615
616         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
617         lustre_msg_set_status(req->rq_repmsg,
618                               ptlrpc_status_hton(req->rq_status));
619         lustre_msg_set_opc(req->rq_repmsg,
620                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
621
622         target_pack_pool_reply(req);
623
624         ptlrpc_at_set_reply(req, flags);
625
626         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
627                 conn = ptlrpc_connection_get(&req->rq_peer, &req->rq_self,
628                                              NULL);
629         else
630                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
631
632         if (unlikely(conn == NULL)) {
633                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
634                 return -ENOTCONN;
635         }
636         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
637
638         rc = sptlrpc_svc_wrap_reply(req);
639         if (unlikely(rc))
640                 goto out;
641
642         req->rq_sent = ktime_get_real_seconds();
643
644         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
645                           (rs->rs_difficult && !rs->rs_no_ack) ?
646                           LNET_ACK_REQ : LNET_NOACK_REQ,
647                           &rs->rs_cb_id, &req->rq_self,
648                           &req->rq_source,
649                           ptlrpc_req2svc(req)->srv_rep_portal,
650                           req->rq_rep_mbits ? req->rq_rep_mbits : req->rq_xid,
651                           req->rq_reply_off, NULL);
652 out:
653         if (unlikely(rc != 0))
654                 ptlrpc_req_drop_rs(req);
655         ptlrpc_connection_put(conn);
656         return rc;
657 }
658
659 int ptlrpc_reply (struct ptlrpc_request *req)
660 {
661         if (req->rq_no_reply)
662                 return 0;
663         else
664                 return (ptlrpc_send_reply(req, 0));
665 }
666
667 /**
668  * For request \a req send an error reply back. Create empty
669  * reply buffers if necessary.
670  */
671 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
672 {
673         int rc;
674         ENTRY;
675
676         if (req->rq_no_reply)
677                 RETURN(0);
678
679         if (!req->rq_repmsg) {
680                 rc = lustre_pack_reply(req, 1, NULL, NULL);
681                 if (rc)
682                         RETURN(rc);
683         }
684
685         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
686             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
687             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT &&
688             req->rq_status != -EROFS)
689                 req->rq_type = PTL_RPC_MSG_ERR;
690
691         rc = ptlrpc_send_reply(req, may_be_difficult);
692         RETURN(rc);
693 }
694
695 int ptlrpc_error(struct ptlrpc_request *req)
696 {
697         return ptlrpc_send_error(req, 0);
698 }
699
700 /**
701  * Send request \a request.
702  * if \a noreply is set, don't expect any reply back and don't set up
703  * reply buffers.
704  * Returns 0 on success or error code.
705  */
706 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
707 {
708         int rc;
709         __u32 opc;
710         int mpflag = 0;
711         bool rep_mbits = false;
712         struct lnet_handle_md bulk_cookie;
713         struct lnet_processid peer;
714         struct ptlrpc_connection *connection;
715         struct lnet_me *reply_me = NULL;
716         struct lnet_md reply_md;
717         struct obd_import *imp = request->rq_import;
718         struct obd_device *obd = imp->imp_obd;
719         ENTRY;
720
721         LNetInvalidateMDHandle(&bulk_cookie);
722
723         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
724                 RETURN(0);
725
726         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DELAY_RECOV) &&
727                      lustre_msg_get_opc(request->rq_reqmsg) == MDS_CONNECT &&
728                      strcmp(obd->obd_type->typ_name, LUSTRE_OSP_NAME) == 0)) {
729                 RETURN(0);
730         }
731
732         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
733         LASSERT(request->rq_wait_ctx == 0);
734
735         /* If this is a re-transmit, we're required to have disengaged
736          * cleanly from the previous attempt */
737         LASSERT(!request->rq_receiving_reply);
738         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
739                   (imp->imp_state == LUSTRE_IMP_FULL)));
740
741         if (unlikely(obd != NULL && obd->obd_fail)) {
742                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
743                        obd->obd_name);
744                 /* this prevents us from waiting in ptlrpc_queue_wait */
745                 spin_lock(&request->rq_lock);
746                 request->rq_err = 1;
747                 spin_unlock(&request->rq_lock);
748                 request->rq_status = -ENODEV;
749                 RETURN(-ENODEV);
750         }
751
752         connection = imp->imp_connection;
753
754         lustre_msg_set_handle(request->rq_reqmsg,
755                               &imp->imp_remote_handle);
756         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
757         lustre_msg_set_conn_cnt(request->rq_reqmsg,
758                                 imp->imp_conn_cnt);
759         lustre_msghdr_set_flags(request->rq_reqmsg,
760                                 imp->imp_msghdr_flags);
761
762         /* If it's the first time to resend the request for EINPROGRESS,
763          * we need to allocate a new XID (see after_reply()), it's different
764          * from the resend for reply timeout. */
765         if (request->rq_nr_resend != 0 &&
766             list_empty(&request->rq_unreplied_list)) {
767                 __u64 min_xid = 0;
768                 /* resend for EINPROGRESS, allocate new xid to avoid reply
769                  * reconstruction */
770                 spin_lock(&imp->imp_lock);
771                 ptlrpc_assign_next_xid_nolock(request);
772                 min_xid = ptlrpc_known_replied_xid(imp);
773                 spin_unlock(&imp->imp_lock);
774
775                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
776                 DEBUG_REQ(D_RPCTRACE, request,
777                           "Allocating new XID for resend on EINPROGRESS");
778         }
779
780         opc = lustre_msg_get_opc(request->rq_reqmsg);
781         if (opc != OST_CONNECT && opc != MDS_CONNECT &&
782             opc != MGS_CONNECT && OCD_HAS_FLAG(&imp->imp_connect_data, FLAGS2))
783                 rep_mbits = imp->imp_connect_data.ocd_connect_flags2 &
784                         OBD_CONNECT2_REP_MBITS;
785
786         if ((request->rq_bulk != NULL) || rep_mbits) {
787                 ptlrpc_set_mbits(request);
788                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
789         }
790
791         if (list_empty(&request->rq_unreplied_list) ||
792             request->rq_xid <= imp->imp_known_replied_xid) {
793                 DEBUG_REQ(D_ERROR, request,
794                           "xid=%llu, replied=%llu, list_empty=%d",
795                           request->rq_xid, imp->imp_known_replied_xid,
796                           list_empty(&request->rq_unreplied_list));
797                 LBUG();
798         }
799
800         /**
801          * For enabled AT all request should have AT_SUPPORT in the
802          * FULL import state when OBD_CONNECT_AT is set.
803          * This check has a race with ptlrpc_connect_import_locked()
804          * with low chance, don't panic, only report.
805          */
806         if (!(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
807             (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
808             !(imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_AT))) {
809                 DEBUG_REQ(D_HA, request, "Wrong state of import detected, AT=%d, imp=%d, msghdr=%d, conn=%d\n",
810                           AT_OFF, imp->imp_state != LUSTRE_IMP_FULL,
811                           (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT),
812                           !(imp->imp_connect_data.ocd_connect_flags &
813                             OBD_CONNECT_AT));
814         }
815         if (request->rq_resend) {
816                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
817                 if (request->rq_resend_cb != NULL)
818                         request->rq_resend_cb(request, &request->rq_async_args);
819         }
820         if (request->rq_memalloc)
821                 mpflag = memalloc_noreclaim_save();
822
823         rc = sptlrpc_cli_wrap_request(request);
824         if (rc)
825                 GOTO(out, rc);
826
827         /* bulk register should be done after wrap_request() */
828         if (request->rq_bulk != NULL) {
829                 rc = ptlrpc_register_bulk (request);
830                 if (rc != 0)
831                         GOTO(cleanup_bulk, rc);
832                 /*
833                  * All the mds in the request will have the same cpt
834                  * encoded in the cookie. So we can just get the first
835                  * one.
836                  */
837                 bulk_cookie = request->rq_bulk->bd_mds[0];
838         }
839
840         if (!noreply) {
841                 LASSERT (request->rq_replen != 0);
842                 if (request->rq_repbuf == NULL) {
843                         LASSERT(request->rq_repdata == NULL);
844                         LASSERT(request->rq_repmsg == NULL);
845                         rc = sptlrpc_cli_alloc_repbuf(request,
846                                                       request->rq_replen);
847                         if (rc) {
848                                 /* this prevents us from looping in
849                                  * ptlrpc_queue_wait */
850                                 spin_lock(&request->rq_lock);
851                                 request->rq_err = 1;
852                                 spin_unlock(&request->rq_lock);
853                                 request->rq_status = rc;
854                                 GOTO(cleanup_bulk, rc);
855                         }
856                 } else {
857                         request->rq_repdata = NULL;
858                         request->rq_repmsg = NULL;
859                 }
860
861                 peer = connection->c_peer;
862                 if (request->rq_bulk &&
863                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
864                         reply_me = ERR_PTR(-ENOMEM);
865                 } else {
866                         reply_me = LNetMEAttach(request->rq_reply_portal,
867                                                 &peer,
868                                                 rep_mbits ? request->rq_mbits :
869                                                 request->rq_xid,
870                                                 0, LNET_UNLINK, LNET_INS_AFTER);
871                 }
872
873                 if (IS_ERR(reply_me)) {
874                         rc = PTR_ERR(reply_me);
875                         CERROR("LNetMEAttach failed: %d\n", rc);
876                         LASSERT(rc == -ENOMEM);
877                         GOTO(cleanup_bulk, rc = -ENOMEM);
878                 }
879         }
880
881         spin_lock(&request->rq_lock);
882         /* We are responsible for unlinking the reply buffer */
883         request->rq_reply_unlinked = noreply;
884         request->rq_receiving_reply = !noreply;
885         /* Clear any flags that may be present from previous sends. */
886         request->rq_req_unlinked = 0;
887         request->rq_replied = 0;
888         request->rq_err = 0;
889         request->rq_timedout = 0;
890         request->rq_net_err = 0;
891         request->rq_resend = 0;
892         request->rq_restart = 0;
893         request->rq_reply_truncated = 0;
894         spin_unlock(&request->rq_lock);
895
896         if (!noreply) {
897                 reply_md.start     = request->rq_repbuf;
898                 reply_md.length    = request->rq_repbuf_len;
899                 /* Allow multiple early replies */
900                 reply_md.threshold = LNET_MD_THRESH_INF;
901                 /* Manage remote for early replies */
902                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
903                         LNET_MD_MANAGE_REMOTE |
904                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
905                 reply_md.user_ptr  = &request->rq_reply_cbid;
906                 reply_md.handler = ptlrpc_handler;
907
908                 /* We must see the unlink callback to set rq_reply_unlinked,
909                  * so we can't auto-unlink */
910                 rc = LNetMDAttach(reply_me, &reply_md, LNET_RETAIN,
911                                   &request->rq_reply_md_h);
912                 if (rc != 0) {
913                         CERROR("LNetMDAttach failed: %d\n", rc);
914                         LASSERT(rc == -ENOMEM);
915                         spin_lock(&request->rq_lock);
916                         /* ...but the MD attach didn't succeed... */
917                         request->rq_receiving_reply = 0;
918                         spin_unlock(&request->rq_lock);
919                         GOTO(cleanup_bulk, rc = -ENOMEM);
920                 }
921                 percpu_ref_get(&ptlrpc_pending);
922
923                 CDEBUG(D_NET,
924                        "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
925                        request->rq_repbuf_len, request->rq_xid,
926                        request->rq_reply_portal);
927         }
928
929         /* add references on request for request_out_callback */
930         ptlrpc_request_addref(request);
931         if (obd != NULL && obd->obd_svc_stats != NULL)
932                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
933                                     atomic_read(&imp->imp_inflight));
934
935         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
936
937         request->rq_sent_ns = ktime_get_real();
938         request->rq_sent = ktime_get_real_seconds();
939         /* We give the server rq_timeout secs to process the req, and
940          * add the network latency for our local timeout.
941          */
942         request->rq_deadline = request->rq_sent + request->rq_timeout +
943                 ptlrpc_at_get_net_latency(request);
944
945         DEBUG_REQ(D_INFO, request, "send flags=%x",
946                   lustre_msg_get_flags(request->rq_reqmsg));
947         rc = ptl_send_buf(&request->rq_req_md_h,
948                           request->rq_reqbuf, request->rq_reqdata_len,
949                           LNET_NOACK_REQ, &request->rq_req_cbid,
950                           NULL,
951                           &connection->c_peer,
952                           request->rq_request_portal,
953                           request->rq_xid, 0, &bulk_cookie);
954         if (likely(rc == 0))
955                 GOTO(out, rc);
956
957         request->rq_req_unlinked = 1;
958         ptlrpc_req_finished(request);
959         if (noreply)
960                 GOTO(out, rc);
961
962         LNetMDUnlink(request->rq_reply_md_h);
963
964         /* UNLINKED callback called synchronously */
965         LASSERT(!request->rq_receiving_reply);
966
967  cleanup_bulk:
968         /* We do sync unlink here as there was no real transfer here so
969          * the chance to have long unlink to sluggish net is smaller here. */
970         ptlrpc_unregister_bulk(request, 0);
971  out:
972         if (rc == -ENOMEM) {
973                 /* set rq_sent so that this request is treated
974                  * as a delayed send in the upper layers */
975                 request->rq_sent = ktime_get_real_seconds();
976         }
977
978         if (request->rq_memalloc)
979                 memalloc_noreclaim_restore(mpflag);
980
981         return rc;
982 }
983 EXPORT_SYMBOL(ptl_send_rpc);
984
985 /**
986  * Register request buffer descriptor for request receiving.
987  */
988 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
989 {
990         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
991         static struct lnet_processid match_id = {
992                 .nid = LNET_ANY_NID,
993                 .pid = LNET_PID_ANY
994         };
995         int rc;
996         struct lnet_md md;
997         struct lnet_me *me;
998
999         CDEBUG(D_NET, "%s: registering portal %d\n", service->srv_name,
1000                service->srv_req_portal);
1001
1002         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
1003                 return -ENOMEM;
1004
1005         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
1006          * which means buffer can only be attached on local CPT, and LND
1007          * threads can find it by grabbing a local lock */
1008         me = LNetMEAttach(service->srv_req_portal,
1009                           &match_id, 0, ~0, LNET_UNLINK,
1010                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
1011                           LNET_INS_LOCAL : LNET_INS_AFTER);
1012         if (IS_ERR(me)) {
1013                 CERROR("%s: LNetMEAttach failed: rc = %ld\n",
1014                        service->srv_name, PTR_ERR(me));
1015                 return PTR_ERR(me);
1016         }
1017
1018         LASSERT(rqbd->rqbd_refcount == 0);
1019         rqbd->rqbd_refcount = 1;
1020
1021         md.start     = rqbd->rqbd_buffer;
1022         md.length    = service->srv_buf_size;
1023         md.max_size  = service->srv_max_req_size;
1024         md.threshold = LNET_MD_THRESH_INF;
1025         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
1026         md.user_ptr  = &rqbd->rqbd_cbid;
1027         md.handler   = ptlrpc_handler;
1028
1029         rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h);
1030         if (rc == 0) {
1031                 percpu_ref_get(&ptlrpc_pending);
1032                 return 0;
1033         }
1034
1035         CERROR("%s: LNetMDAttach failed: rc = %d\n", service->srv_name, rc);
1036         LASSERT(rc == -ENOMEM);
1037         rqbd->rqbd_refcount = 0;
1038
1039         return rc;
1040 }